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_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/sched.h>
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * single value that denotes runtime == period, ie unlimited time.
120 #define RUNTIME_INF ((u64)~0ULL)
122 static inline int rt_policy(int policy
)
124 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
129 static inline int task_has_rt_policy(struct task_struct
*p
)
131 return rt_policy(p
->policy
);
135 * This is the priority-queue data structure of the RT scheduling class:
137 struct rt_prio_array
{
138 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
139 struct list_head queue
[MAX_RT_PRIO
];
142 struct rt_bandwidth
{
143 /* nests inside the rq lock: */
144 raw_spinlock_t rt_runtime_lock
;
147 struct hrtimer rt_period_timer
;
150 static struct rt_bandwidth def_rt_bandwidth
;
152 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
154 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
156 struct rt_bandwidth
*rt_b
=
157 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
163 now
= hrtimer_cb_get_time(timer
);
164 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
169 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
172 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
176 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
178 rt_b
->rt_period
= ns_to_ktime(period
);
179 rt_b
->rt_runtime
= runtime
;
181 raw_spin_lock_init(&rt_b
->rt_runtime_lock
);
183 hrtimer_init(&rt_b
->rt_period_timer
,
184 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
185 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
188 static inline int rt_bandwidth_enabled(void)
190 return sysctl_sched_rt_runtime
>= 0;
193 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
197 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
200 if (hrtimer_active(&rt_b
->rt_period_timer
))
203 raw_spin_lock(&rt_b
->rt_runtime_lock
);
208 if (hrtimer_active(&rt_b
->rt_period_timer
))
211 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
212 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
214 soft
= hrtimer_get_softexpires(&rt_b
->rt_period_timer
);
215 hard
= hrtimer_get_expires(&rt_b
->rt_period_timer
);
216 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
217 __hrtimer_start_range_ns(&rt_b
->rt_period_timer
, soft
, delta
,
218 HRTIMER_MODE_ABS_PINNED
, 0);
220 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
223 #ifdef CONFIG_RT_GROUP_SCHED
224 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
226 hrtimer_cancel(&rt_b
->rt_period_timer
);
231 * sched_domains_mutex serializes calls to arch_init_sched_domains,
232 * detach_destroy_domains and partition_sched_domains.
234 static DEFINE_MUTEX(sched_domains_mutex
);
236 #ifdef CONFIG_GROUP_SCHED
238 #include <linux/cgroup.h>
242 static LIST_HEAD(task_groups
);
244 /* task group related information */
246 #ifdef CONFIG_CGROUP_SCHED
247 struct cgroup_subsys_state css
;
250 #ifdef CONFIG_USER_SCHED
254 #ifdef CONFIG_FAIR_GROUP_SCHED
255 /* schedulable entities of this group on each cpu */
256 struct sched_entity
**se
;
257 /* runqueue "owned" by this group on each cpu */
258 struct cfs_rq
**cfs_rq
;
259 unsigned long shares
;
262 #ifdef CONFIG_RT_GROUP_SCHED
263 struct sched_rt_entity
**rt_se
;
264 struct rt_rq
**rt_rq
;
266 struct rt_bandwidth rt_bandwidth
;
270 struct list_head list
;
272 struct task_group
*parent
;
273 struct list_head siblings
;
274 struct list_head children
;
277 #ifdef CONFIG_USER_SCHED
279 /* Helper function to pass uid information to create_sched_user() */
280 void set_tg_uid(struct user_struct
*user
)
282 user
->tg
->uid
= user
->uid
;
287 * Every UID task group (including init_task_group aka UID-0) will
288 * be a child to this group.
290 struct task_group root_task_group
;
292 #ifdef CONFIG_FAIR_GROUP_SCHED
293 /* Default task group's sched entity on each cpu */
294 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
295 /* Default task group's cfs_rq on each cpu */
296 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cfs_rq
, init_tg_cfs_rq
);
297 #endif /* CONFIG_FAIR_GROUP_SCHED */
299 #ifdef CONFIG_RT_GROUP_SCHED
300 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
301 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rt_rq
, init_rt_rq_var
);
302 #endif /* CONFIG_RT_GROUP_SCHED */
303 #else /* !CONFIG_USER_SCHED */
304 #define root_task_group init_task_group
305 #endif /* CONFIG_USER_SCHED */
307 /* task_group_lock serializes add/remove of task groups and also changes to
308 * a task group's cpu shares.
310 static DEFINE_SPINLOCK(task_group_lock
);
312 #ifdef CONFIG_FAIR_GROUP_SCHED
315 static int root_task_group_empty(void)
317 return list_empty(&root_task_group
.children
);
321 #ifdef CONFIG_USER_SCHED
322 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
323 #else /* !CONFIG_USER_SCHED */
324 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
325 #endif /* CONFIG_USER_SCHED */
328 * A weight of 0 or 1 can cause arithmetics problems.
329 * A weight of a cfs_rq is the sum of weights of which entities
330 * are queued on this cfs_rq, so a weight of a entity should not be
331 * too large, so as the shares value of a task group.
332 * (The default weight is 1024 - so there's no practical
333 * limitation from this.)
336 #define MAX_SHARES (1UL << 18)
338 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
341 /* Default task group.
342 * Every task in system belong to this group at bootup.
344 struct task_group init_task_group
;
346 /* return group to which a task belongs */
347 static inline struct task_group
*task_group(struct task_struct
*p
)
349 struct task_group
*tg
;
351 #ifdef CONFIG_USER_SCHED
353 tg
= __task_cred(p
)->user
->tg
;
355 #elif defined(CONFIG_CGROUP_SCHED)
356 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
357 struct task_group
, css
);
359 tg
= &init_task_group
;
364 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
365 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
367 #ifdef CONFIG_FAIR_GROUP_SCHED
368 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
369 p
->se
.parent
= task_group(p
)->se
[cpu
];
372 #ifdef CONFIG_RT_GROUP_SCHED
373 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
374 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
380 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
381 static inline struct task_group
*task_group(struct task_struct
*p
)
386 #endif /* CONFIG_GROUP_SCHED */
388 /* CFS-related fields in a runqueue */
390 struct load_weight load
;
391 unsigned long nr_running
;
396 struct rb_root tasks_timeline
;
397 struct rb_node
*rb_leftmost
;
399 struct list_head tasks
;
400 struct list_head
*balance_iterator
;
403 * 'curr' points to currently running entity on this cfs_rq.
404 * It is set to NULL otherwise (i.e when none are currently running).
406 struct sched_entity
*curr
, *next
, *last
;
408 unsigned int nr_spread_over
;
410 #ifdef CONFIG_FAIR_GROUP_SCHED
411 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
414 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
415 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
416 * (like users, containers etc.)
418 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
419 * list is used during load balance.
421 struct list_head leaf_cfs_rq_list
;
422 struct task_group
*tg
; /* group that "owns" this runqueue */
426 * the part of load.weight contributed by tasks
428 unsigned long task_weight
;
431 * h_load = weight * f(tg)
433 * Where f(tg) is the recursive weight fraction assigned to
436 unsigned long h_load
;
439 * this cpu's part of tg->shares
441 unsigned long shares
;
444 * load.weight at the time we set shares
446 unsigned long rq_weight
;
451 /* Real-Time classes' related field in a runqueue: */
453 struct rt_prio_array active
;
454 unsigned long rt_nr_running
;
455 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
457 int curr
; /* highest queued rt task prio */
459 int next
; /* next highest */
464 unsigned long rt_nr_migratory
;
465 unsigned long rt_nr_total
;
467 struct plist_head pushable_tasks
;
472 /* Nests inside the rq lock: */
473 raw_spinlock_t rt_runtime_lock
;
475 #ifdef CONFIG_RT_GROUP_SCHED
476 unsigned long rt_nr_boosted
;
479 struct list_head leaf_rt_rq_list
;
480 struct task_group
*tg
;
481 struct sched_rt_entity
*rt_se
;
488 * We add the notion of a root-domain which will be used to define per-domain
489 * variables. Each exclusive cpuset essentially defines an island domain by
490 * fully partitioning the member cpus from any other cpuset. Whenever a new
491 * exclusive cpuset is created, we also create and attach a new root-domain
498 cpumask_var_t online
;
501 * The "RT overload" flag: it gets set if a CPU has more than
502 * one runnable RT task.
504 cpumask_var_t rto_mask
;
507 struct cpupri cpupri
;
512 * By default the system creates a single root-domain with all cpus as
513 * members (mimicking the global state we have today).
515 static struct root_domain def_root_domain
;
520 * This is the main, per-CPU runqueue data structure.
522 * Locking rule: those places that want to lock multiple runqueues
523 * (such as the load balancing or the thread migration code), lock
524 * acquire operations must be ordered by ascending &runqueue.
531 * nr_running and cpu_load should be in the same cacheline because
532 * remote CPUs use both these fields when doing load calculation.
534 unsigned long nr_running
;
535 #define CPU_LOAD_IDX_MAX 5
536 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
538 unsigned char in_nohz_recently
;
540 /* capture load from *all* tasks on this cpu: */
541 struct load_weight load
;
542 unsigned long nr_load_updates
;
548 #ifdef CONFIG_FAIR_GROUP_SCHED
549 /* list of leaf cfs_rq on this cpu: */
550 struct list_head leaf_cfs_rq_list
;
552 #ifdef CONFIG_RT_GROUP_SCHED
553 struct list_head leaf_rt_rq_list
;
557 * This is part of a global counter where only the total sum
558 * over all CPUs matters. A task can increase this counter on
559 * one CPU and if it got migrated afterwards it may decrease
560 * it on another CPU. Always updated under the runqueue lock:
562 unsigned long nr_uninterruptible
;
564 struct task_struct
*curr
, *idle
;
565 unsigned long next_balance
;
566 struct mm_struct
*prev_mm
;
573 struct root_domain
*rd
;
574 struct sched_domain
*sd
;
576 unsigned char idle_at_tick
;
577 /* For active balancing */
581 /* cpu of this runqueue: */
585 unsigned long avg_load_per_task
;
587 struct task_struct
*migration_thread
;
588 struct list_head migration_queue
;
596 /* calc_load related fields */
597 unsigned long calc_load_update
;
598 long calc_load_active
;
600 #ifdef CONFIG_SCHED_HRTICK
602 int hrtick_csd_pending
;
603 struct call_single_data hrtick_csd
;
605 struct hrtimer hrtick_timer
;
608 #ifdef CONFIG_SCHEDSTATS
610 struct sched_info rq_sched_info
;
611 unsigned long long rq_cpu_time
;
612 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
614 /* sys_sched_yield() stats */
615 unsigned int yld_count
;
617 /* schedule() stats */
618 unsigned int sched_switch
;
619 unsigned int sched_count
;
620 unsigned int sched_goidle
;
622 /* try_to_wake_up() stats */
623 unsigned int ttwu_count
;
624 unsigned int ttwu_local
;
627 unsigned int bkl_count
;
631 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
634 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
636 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
639 static inline int cpu_of(struct rq
*rq
)
649 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
650 * See detach_destroy_domains: synchronize_sched for details.
652 * The domain tree of any CPU may only be accessed from within
653 * preempt-disabled sections.
655 #define for_each_domain(cpu, __sd) \
656 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
658 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
659 #define this_rq() (&__get_cpu_var(runqueues))
660 #define task_rq(p) cpu_rq(task_cpu(p))
661 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
662 #define raw_rq() (&__raw_get_cpu_var(runqueues))
664 inline void update_rq_clock(struct rq
*rq
)
666 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
670 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
672 #ifdef CONFIG_SCHED_DEBUG
673 # define const_debug __read_mostly
675 # define const_debug static const
680 * @cpu: the processor in question.
682 * Returns true if the current cpu runqueue is locked.
683 * This interface allows printk to be called with the runqueue lock
684 * held and know whether or not it is OK to wake up the klogd.
686 int runqueue_is_locked(int cpu
)
688 return raw_spin_is_locked(&cpu_rq(cpu
)->lock
);
692 * Debugging: various feature bits
695 #define SCHED_FEAT(name, enabled) \
696 __SCHED_FEAT_##name ,
699 #include "sched_features.h"
704 #define SCHED_FEAT(name, enabled) \
705 (1UL << __SCHED_FEAT_##name) * enabled |
707 const_debug
unsigned int sysctl_sched_features
=
708 #include "sched_features.h"
713 #ifdef CONFIG_SCHED_DEBUG
714 #define SCHED_FEAT(name, enabled) \
717 static __read_mostly
char *sched_feat_names
[] = {
718 #include "sched_features.h"
724 static int sched_feat_show(struct seq_file
*m
, void *v
)
728 for (i
= 0; sched_feat_names
[i
]; i
++) {
729 if (!(sysctl_sched_features
& (1UL << i
)))
731 seq_printf(m
, "%s ", sched_feat_names
[i
]);
739 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
740 size_t cnt
, loff_t
*ppos
)
750 if (copy_from_user(&buf
, ubuf
, cnt
))
755 if (strncmp(buf
, "NO_", 3) == 0) {
760 for (i
= 0; sched_feat_names
[i
]; i
++) {
761 int len
= strlen(sched_feat_names
[i
]);
763 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
765 sysctl_sched_features
&= ~(1UL << i
);
767 sysctl_sched_features
|= (1UL << i
);
772 if (!sched_feat_names
[i
])
780 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
782 return single_open(filp
, sched_feat_show
, NULL
);
785 static const struct file_operations sched_feat_fops
= {
786 .open
= sched_feat_open
,
787 .write
= sched_feat_write
,
790 .release
= single_release
,
793 static __init
int sched_init_debug(void)
795 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
800 late_initcall(sched_init_debug
);
804 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
807 * Number of tasks to iterate in a single balance run.
808 * Limited because this is done with IRQs disabled.
810 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
813 * ratelimit for updating the group shares.
816 unsigned int sysctl_sched_shares_ratelimit
= 250000;
817 unsigned int normalized_sysctl_sched_shares_ratelimit
= 250000;
820 * Inject some fuzzyness into changing the per-cpu group shares
821 * this avoids remote rq-locks at the expense of fairness.
824 unsigned int sysctl_sched_shares_thresh
= 4;
827 * period over which we average the RT time consumption, measured
832 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
835 * period over which we measure -rt task cpu usage in us.
838 unsigned int sysctl_sched_rt_period
= 1000000;
840 static __read_mostly
int scheduler_running
;
843 * part of the period that we allow rt tasks to run in us.
846 int sysctl_sched_rt_runtime
= 950000;
848 static inline u64
global_rt_period(void)
850 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
853 static inline u64
global_rt_runtime(void)
855 if (sysctl_sched_rt_runtime
< 0)
858 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
861 #ifndef prepare_arch_switch
862 # define prepare_arch_switch(next) do { } while (0)
864 #ifndef finish_arch_switch
865 # define finish_arch_switch(prev) do { } while (0)
868 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
870 return rq
->curr
== p
;
873 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
874 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
876 return task_current(rq
, p
);
879 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
883 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
885 #ifdef CONFIG_DEBUG_SPINLOCK
886 /* this is a valid case when another task releases the spinlock */
887 rq
->lock
.owner
= current
;
890 * If we are tracking spinlock dependencies then we have to
891 * fix up the runqueue lock - which gets 'carried over' from
894 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
896 raw_spin_unlock_irq(&rq
->lock
);
899 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
900 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
905 return task_current(rq
, p
);
909 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
913 * We can optimise this out completely for !SMP, because the
914 * SMP rebalancing from interrupt is the only thing that cares
919 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
920 raw_spin_unlock_irq(&rq
->lock
);
922 raw_spin_unlock(&rq
->lock
);
926 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
930 * After ->oncpu is cleared, the task can be moved to a different CPU.
931 * We must ensure this doesn't happen until the switch is completely
937 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
941 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
944 * __task_rq_lock - lock the runqueue a given task resides on.
945 * Must be called interrupts disabled.
947 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
951 struct rq
*rq
= task_rq(p
);
952 raw_spin_lock(&rq
->lock
);
953 if (likely(rq
== task_rq(p
)))
955 raw_spin_unlock(&rq
->lock
);
960 * task_rq_lock - lock the runqueue a given task resides on and disable
961 * interrupts. Note the ordering: we can safely lookup the task_rq without
962 * explicitly disabling preemption.
964 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
970 local_irq_save(*flags
);
972 raw_spin_lock(&rq
->lock
);
973 if (likely(rq
== task_rq(p
)))
975 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
979 void task_rq_unlock_wait(struct task_struct
*p
)
981 struct rq
*rq
= task_rq(p
);
983 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
984 raw_spin_unlock_wait(&rq
->lock
);
987 static void __task_rq_unlock(struct rq
*rq
)
990 raw_spin_unlock(&rq
->lock
);
993 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
996 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
1000 * this_rq_lock - lock this runqueue and disable interrupts.
1002 static struct rq
*this_rq_lock(void)
1003 __acquires(rq
->lock
)
1007 local_irq_disable();
1009 raw_spin_lock(&rq
->lock
);
1014 #ifdef CONFIG_SCHED_HRTICK
1016 * Use HR-timers to deliver accurate preemption points.
1018 * Its all a bit involved since we cannot program an hrt while holding the
1019 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1022 * When we get rescheduled we reprogram the hrtick_timer outside of the
1028 * - enabled by features
1029 * - hrtimer is actually high res
1031 static inline int hrtick_enabled(struct rq
*rq
)
1033 if (!sched_feat(HRTICK
))
1035 if (!cpu_active(cpu_of(rq
)))
1037 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1040 static void hrtick_clear(struct rq
*rq
)
1042 if (hrtimer_active(&rq
->hrtick_timer
))
1043 hrtimer_cancel(&rq
->hrtick_timer
);
1047 * High-resolution timer tick.
1048 * Runs from hardirq context with interrupts disabled.
1050 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1052 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1054 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1056 raw_spin_lock(&rq
->lock
);
1057 update_rq_clock(rq
);
1058 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1059 raw_spin_unlock(&rq
->lock
);
1061 return HRTIMER_NORESTART
;
1066 * called from hardirq (IPI) context
1068 static void __hrtick_start(void *arg
)
1070 struct rq
*rq
= arg
;
1072 raw_spin_lock(&rq
->lock
);
1073 hrtimer_restart(&rq
->hrtick_timer
);
1074 rq
->hrtick_csd_pending
= 0;
1075 raw_spin_unlock(&rq
->lock
);
1079 * Called to set the hrtick timer state.
1081 * called with rq->lock held and irqs disabled
1083 static void hrtick_start(struct rq
*rq
, u64 delay
)
1085 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1086 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1088 hrtimer_set_expires(timer
, time
);
1090 if (rq
== this_rq()) {
1091 hrtimer_restart(timer
);
1092 } else if (!rq
->hrtick_csd_pending
) {
1093 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1094 rq
->hrtick_csd_pending
= 1;
1099 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1101 int cpu
= (int)(long)hcpu
;
1104 case CPU_UP_CANCELED
:
1105 case CPU_UP_CANCELED_FROZEN
:
1106 case CPU_DOWN_PREPARE
:
1107 case CPU_DOWN_PREPARE_FROZEN
:
1109 case CPU_DEAD_FROZEN
:
1110 hrtick_clear(cpu_rq(cpu
));
1117 static __init
void init_hrtick(void)
1119 hotcpu_notifier(hotplug_hrtick
, 0);
1123 * Called to set the hrtick timer state.
1125 * called with rq->lock held and irqs disabled
1127 static void hrtick_start(struct rq
*rq
, u64 delay
)
1129 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1130 HRTIMER_MODE_REL_PINNED
, 0);
1133 static inline void init_hrtick(void)
1136 #endif /* CONFIG_SMP */
1138 static void init_rq_hrtick(struct rq
*rq
)
1141 rq
->hrtick_csd_pending
= 0;
1143 rq
->hrtick_csd
.flags
= 0;
1144 rq
->hrtick_csd
.func
= __hrtick_start
;
1145 rq
->hrtick_csd
.info
= rq
;
1148 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1149 rq
->hrtick_timer
.function
= hrtick
;
1151 #else /* CONFIG_SCHED_HRTICK */
1152 static inline void hrtick_clear(struct rq
*rq
)
1156 static inline void init_rq_hrtick(struct rq
*rq
)
1160 static inline void init_hrtick(void)
1163 #endif /* CONFIG_SCHED_HRTICK */
1166 * resched_task - mark a task 'to be rescheduled now'.
1168 * On UP this means the setting of the need_resched flag, on SMP it
1169 * might also involve a cross-CPU call to trigger the scheduler on
1174 #ifndef tsk_is_polling
1175 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1178 static void resched_task(struct task_struct
*p
)
1182 assert_raw_spin_locked(&task_rq(p
)->lock
);
1184 if (test_tsk_need_resched(p
))
1187 set_tsk_need_resched(p
);
1190 if (cpu
== smp_processor_id())
1193 /* NEED_RESCHED must be visible before we test polling */
1195 if (!tsk_is_polling(p
))
1196 smp_send_reschedule(cpu
);
1199 static void resched_cpu(int cpu
)
1201 struct rq
*rq
= cpu_rq(cpu
);
1202 unsigned long flags
;
1204 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
1206 resched_task(cpu_curr(cpu
));
1207 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1212 * When add_timer_on() enqueues a timer into the timer wheel of an
1213 * idle CPU then this timer might expire before the next timer event
1214 * which is scheduled to wake up that CPU. In case of a completely
1215 * idle system the next event might even be infinite time into the
1216 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1217 * leaves the inner idle loop so the newly added timer is taken into
1218 * account when the CPU goes back to idle and evaluates the timer
1219 * wheel for the next timer event.
1221 void wake_up_idle_cpu(int cpu
)
1223 struct rq
*rq
= cpu_rq(cpu
);
1225 if (cpu
== smp_processor_id())
1229 * This is safe, as this function is called with the timer
1230 * wheel base lock of (cpu) held. When the CPU is on the way
1231 * to idle and has not yet set rq->curr to idle then it will
1232 * be serialized on the timer wheel base lock and take the new
1233 * timer into account automatically.
1235 if (rq
->curr
!= rq
->idle
)
1239 * We can set TIF_RESCHED on the idle task of the other CPU
1240 * lockless. The worst case is that the other CPU runs the
1241 * idle task through an additional NOOP schedule()
1243 set_tsk_need_resched(rq
->idle
);
1245 /* NEED_RESCHED must be visible before we test polling */
1247 if (!tsk_is_polling(rq
->idle
))
1248 smp_send_reschedule(cpu
);
1250 #endif /* CONFIG_NO_HZ */
1252 static u64
sched_avg_period(void)
1254 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1257 static void sched_avg_update(struct rq
*rq
)
1259 s64 period
= sched_avg_period();
1261 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1262 rq
->age_stamp
+= period
;
1267 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1269 rq
->rt_avg
+= rt_delta
;
1270 sched_avg_update(rq
);
1273 #else /* !CONFIG_SMP */
1274 static void resched_task(struct task_struct
*p
)
1276 assert_raw_spin_locked(&task_rq(p
)->lock
);
1277 set_tsk_need_resched(p
);
1280 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1283 #endif /* CONFIG_SMP */
1285 #if BITS_PER_LONG == 32
1286 # define WMULT_CONST (~0UL)
1288 # define WMULT_CONST (1UL << 32)
1291 #define WMULT_SHIFT 32
1294 * Shift right and round:
1296 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1299 * delta *= weight / lw
1301 static unsigned long
1302 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1303 struct load_weight
*lw
)
1307 if (!lw
->inv_weight
) {
1308 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1311 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1315 tmp
= (u64
)delta_exec
* weight
;
1317 * Check whether we'd overflow the 64-bit multiplication:
1319 if (unlikely(tmp
> WMULT_CONST
))
1320 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1323 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1325 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1328 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1334 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1341 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1342 * of tasks with abnormal "nice" values across CPUs the contribution that
1343 * each task makes to its run queue's load is weighted according to its
1344 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1345 * scaled version of the new time slice allocation that they receive on time
1349 #define WEIGHT_IDLEPRIO 3
1350 #define WMULT_IDLEPRIO 1431655765
1353 * Nice levels are multiplicative, with a gentle 10% change for every
1354 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1355 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1356 * that remained on nice 0.
1358 * The "10% effect" is relative and cumulative: from _any_ nice level,
1359 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1360 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1361 * If a task goes up by ~10% and another task goes down by ~10% then
1362 * the relative distance between them is ~25%.)
1364 static const int prio_to_weight
[40] = {
1365 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1366 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1367 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1368 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1369 /* 0 */ 1024, 820, 655, 526, 423,
1370 /* 5 */ 335, 272, 215, 172, 137,
1371 /* 10 */ 110, 87, 70, 56, 45,
1372 /* 15 */ 36, 29, 23, 18, 15,
1376 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1378 * In cases where the weight does not change often, we can use the
1379 * precalculated inverse to speed up arithmetics by turning divisions
1380 * into multiplications:
1382 static const u32 prio_to_wmult
[40] = {
1383 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1384 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1385 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1386 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1387 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1388 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1389 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1390 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1393 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1396 * runqueue iterator, to support SMP load-balancing between different
1397 * scheduling classes, without having to expose their internal data
1398 * structures to the load-balancing proper:
1400 struct rq_iterator
{
1402 struct task_struct
*(*start
)(void *);
1403 struct task_struct
*(*next
)(void *);
1407 static unsigned long
1408 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1409 unsigned long max_load_move
, struct sched_domain
*sd
,
1410 enum cpu_idle_type idle
, int *all_pinned
,
1411 int *this_best_prio
, struct rq_iterator
*iterator
);
1414 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1415 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1416 struct rq_iterator
*iterator
);
1419 /* Time spent by the tasks of the cpu accounting group executing in ... */
1420 enum cpuacct_stat_index
{
1421 CPUACCT_STAT_USER
, /* ... user mode */
1422 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1424 CPUACCT_STAT_NSTATS
,
1427 #ifdef CONFIG_CGROUP_CPUACCT
1428 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1429 static void cpuacct_update_stats(struct task_struct
*tsk
,
1430 enum cpuacct_stat_index idx
, cputime_t val
);
1432 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1433 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1434 enum cpuacct_stat_index idx
, cputime_t val
) {}
1437 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1439 update_load_add(&rq
->load
, load
);
1442 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1444 update_load_sub(&rq
->load
, load
);
1447 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1448 typedef int (*tg_visitor
)(struct task_group
*, void *);
1451 * Iterate the full tree, calling @down when first entering a node and @up when
1452 * leaving it for the final time.
1454 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1456 struct task_group
*parent
, *child
;
1460 parent
= &root_task_group
;
1462 ret
= (*down
)(parent
, data
);
1465 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1472 ret
= (*up
)(parent
, data
);
1477 parent
= parent
->parent
;
1486 static int tg_nop(struct task_group
*tg
, void *data
)
1493 /* Used instead of source_load when we know the type == 0 */
1494 static unsigned long weighted_cpuload(const int cpu
)
1496 return cpu_rq(cpu
)->load
.weight
;
1500 * Return a low guess at the load of a migration-source cpu weighted
1501 * according to the scheduling class and "nice" value.
1503 * We want to under-estimate the load of migration sources, to
1504 * balance conservatively.
1506 static unsigned long source_load(int cpu
, int type
)
1508 struct rq
*rq
= cpu_rq(cpu
);
1509 unsigned long total
= weighted_cpuload(cpu
);
1511 if (type
== 0 || !sched_feat(LB_BIAS
))
1514 return min(rq
->cpu_load
[type
-1], total
);
1518 * Return a high guess at the load of a migration-target cpu weighted
1519 * according to the scheduling class and "nice" value.
1521 static unsigned long target_load(int cpu
, int type
)
1523 struct rq
*rq
= cpu_rq(cpu
);
1524 unsigned long total
= weighted_cpuload(cpu
);
1526 if (type
== 0 || !sched_feat(LB_BIAS
))
1529 return max(rq
->cpu_load
[type
-1], total
);
1532 static struct sched_group
*group_of(int cpu
)
1534 struct sched_domain
*sd
= rcu_dereference(cpu_rq(cpu
)->sd
);
1542 static unsigned long power_of(int cpu
)
1544 struct sched_group
*group
= group_of(cpu
);
1547 return SCHED_LOAD_SCALE
;
1549 return group
->cpu_power
;
1552 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1554 static unsigned long cpu_avg_load_per_task(int cpu
)
1556 struct rq
*rq
= cpu_rq(cpu
);
1557 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1560 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1562 rq
->avg_load_per_task
= 0;
1564 return rq
->avg_load_per_task
;
1567 #ifdef CONFIG_FAIR_GROUP_SCHED
1569 static __read_mostly
unsigned long *update_shares_data
;
1571 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1574 * Calculate and set the cpu's group shares.
1576 static void update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1577 unsigned long sd_shares
,
1578 unsigned long sd_rq_weight
,
1579 unsigned long *usd_rq_weight
)
1581 unsigned long shares
, rq_weight
;
1584 rq_weight
= usd_rq_weight
[cpu
];
1587 rq_weight
= NICE_0_LOAD
;
1591 * \Sum_j shares_j * rq_weight_i
1592 * shares_i = -----------------------------
1593 * \Sum_j rq_weight_j
1595 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1596 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1598 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1599 sysctl_sched_shares_thresh
) {
1600 struct rq
*rq
= cpu_rq(cpu
);
1601 unsigned long flags
;
1603 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1604 tg
->cfs_rq
[cpu
]->rq_weight
= boost
? 0 : rq_weight
;
1605 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1606 __set_se_shares(tg
->se
[cpu
], shares
);
1607 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1612 * Re-compute the task group their per cpu shares over the given domain.
1613 * This needs to be done in a bottom-up fashion because the rq weight of a
1614 * parent group depends on the shares of its child groups.
1616 static int tg_shares_up(struct task_group
*tg
, void *data
)
1618 unsigned long weight
, rq_weight
= 0, sum_weight
= 0, shares
= 0;
1619 unsigned long *usd_rq_weight
;
1620 struct sched_domain
*sd
= data
;
1621 unsigned long flags
;
1627 local_irq_save(flags
);
1628 usd_rq_weight
= per_cpu_ptr(update_shares_data
, smp_processor_id());
1630 for_each_cpu(i
, sched_domain_span(sd
)) {
1631 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1632 usd_rq_weight
[i
] = weight
;
1634 rq_weight
+= weight
;
1636 * If there are currently no tasks on the cpu pretend there
1637 * is one of average load so that when a new task gets to
1638 * run here it will not get delayed by group starvation.
1641 weight
= NICE_0_LOAD
;
1643 sum_weight
+= weight
;
1644 shares
+= tg
->cfs_rq
[i
]->shares
;
1648 rq_weight
= sum_weight
;
1650 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1651 shares
= tg
->shares
;
1653 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1654 shares
= tg
->shares
;
1656 for_each_cpu(i
, sched_domain_span(sd
))
1657 update_group_shares_cpu(tg
, i
, shares
, rq_weight
, usd_rq_weight
);
1659 local_irq_restore(flags
);
1665 * Compute the cpu's hierarchical load factor for each task group.
1666 * This needs to be done in a top-down fashion because the load of a child
1667 * group is a fraction of its parents load.
1669 static int tg_load_down(struct task_group
*tg
, void *data
)
1672 long cpu
= (long)data
;
1675 load
= cpu_rq(cpu
)->load
.weight
;
1677 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1678 load
*= tg
->cfs_rq
[cpu
]->shares
;
1679 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1682 tg
->cfs_rq
[cpu
]->h_load
= load
;
1687 static void update_shares(struct sched_domain
*sd
)
1692 if (root_task_group_empty())
1695 now
= cpu_clock(raw_smp_processor_id());
1696 elapsed
= now
- sd
->last_update
;
1698 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1699 sd
->last_update
= now
;
1700 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1704 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1706 if (root_task_group_empty())
1709 raw_spin_unlock(&rq
->lock
);
1711 raw_spin_lock(&rq
->lock
);
1714 static void update_h_load(long cpu
)
1716 if (root_task_group_empty())
1719 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1724 static inline void update_shares(struct sched_domain
*sd
)
1728 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1734 #ifdef CONFIG_PREEMPT
1736 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1739 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1740 * way at the expense of forcing extra atomic operations in all
1741 * invocations. This assures that the double_lock is acquired using the
1742 * same underlying policy as the spinlock_t on this architecture, which
1743 * reduces latency compared to the unfair variant below. However, it
1744 * also adds more overhead and therefore may reduce throughput.
1746 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1747 __releases(this_rq
->lock
)
1748 __acquires(busiest
->lock
)
1749 __acquires(this_rq
->lock
)
1751 raw_spin_unlock(&this_rq
->lock
);
1752 double_rq_lock(this_rq
, busiest
);
1759 * Unfair double_lock_balance: Optimizes throughput at the expense of
1760 * latency by eliminating extra atomic operations when the locks are
1761 * already in proper order on entry. This favors lower cpu-ids and will
1762 * grant the double lock to lower cpus over higher ids under contention,
1763 * regardless of entry order into the function.
1765 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1766 __releases(this_rq
->lock
)
1767 __acquires(busiest
->lock
)
1768 __acquires(this_rq
->lock
)
1772 if (unlikely(!raw_spin_trylock(&busiest
->lock
))) {
1773 if (busiest
< this_rq
) {
1774 raw_spin_unlock(&this_rq
->lock
);
1775 raw_spin_lock(&busiest
->lock
);
1776 raw_spin_lock_nested(&this_rq
->lock
,
1777 SINGLE_DEPTH_NESTING
);
1780 raw_spin_lock_nested(&busiest
->lock
,
1781 SINGLE_DEPTH_NESTING
);
1786 #endif /* CONFIG_PREEMPT */
1789 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1791 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1793 if (unlikely(!irqs_disabled())) {
1794 /* printk() doesn't work good under rq->lock */
1795 raw_spin_unlock(&this_rq
->lock
);
1799 return _double_lock_balance(this_rq
, busiest
);
1802 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1803 __releases(busiest
->lock
)
1805 raw_spin_unlock(&busiest
->lock
);
1806 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1810 #ifdef CONFIG_FAIR_GROUP_SCHED
1811 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1814 cfs_rq
->shares
= shares
;
1819 static void calc_load_account_active(struct rq
*this_rq
);
1820 static void update_sysctl(void);
1821 static int get_update_sysctl_factor(void);
1823 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1825 set_task_rq(p
, cpu
);
1828 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1829 * successfuly executed on another CPU. We must ensure that updates of
1830 * per-task data have been completed by this moment.
1833 task_thread_info(p
)->cpu
= cpu
;
1837 #include "sched_stats.h"
1838 #include "sched_idletask.c"
1839 #include "sched_fair.c"
1840 #include "sched_rt.c"
1841 #ifdef CONFIG_SCHED_DEBUG
1842 # include "sched_debug.c"
1845 #define sched_class_highest (&rt_sched_class)
1846 #define for_each_class(class) \
1847 for (class = sched_class_highest; class; class = class->next)
1849 static void inc_nr_running(struct rq
*rq
)
1854 static void dec_nr_running(struct rq
*rq
)
1859 static void set_load_weight(struct task_struct
*p
)
1861 if (task_has_rt_policy(p
)) {
1862 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1863 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1868 * SCHED_IDLE tasks get minimal weight:
1870 if (p
->policy
== SCHED_IDLE
) {
1871 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1872 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1876 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1877 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1880 static void update_avg(u64
*avg
, u64 sample
)
1882 s64 diff
= sample
- *avg
;
1886 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1889 p
->se
.start_runtime
= p
->se
.sum_exec_runtime
;
1891 sched_info_queued(p
);
1892 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1896 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1899 if (p
->se
.last_wakeup
) {
1900 update_avg(&p
->se
.avg_overlap
,
1901 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1902 p
->se
.last_wakeup
= 0;
1904 update_avg(&p
->se
.avg_wakeup
,
1905 sysctl_sched_wakeup_granularity
);
1909 sched_info_dequeued(p
);
1910 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1915 * __normal_prio - return the priority that is based on the static prio
1917 static inline int __normal_prio(struct task_struct
*p
)
1919 return p
->static_prio
;
1923 * Calculate the expected normal priority: i.e. priority
1924 * without taking RT-inheritance into account. Might be
1925 * boosted by interactivity modifiers. Changes upon fork,
1926 * setprio syscalls, and whenever the interactivity
1927 * estimator recalculates.
1929 static inline int normal_prio(struct task_struct
*p
)
1933 if (task_has_rt_policy(p
))
1934 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1936 prio
= __normal_prio(p
);
1941 * Calculate the current priority, i.e. the priority
1942 * taken into account by the scheduler. This value might
1943 * be boosted by RT tasks, or might be boosted by
1944 * interactivity modifiers. Will be RT if the task got
1945 * RT-boosted. If not then it returns p->normal_prio.
1947 static int effective_prio(struct task_struct
*p
)
1949 p
->normal_prio
= normal_prio(p
);
1951 * If we are RT tasks or we were boosted to RT priority,
1952 * keep the priority unchanged. Otherwise, update priority
1953 * to the normal priority:
1955 if (!rt_prio(p
->prio
))
1956 return p
->normal_prio
;
1961 * activate_task - move a task to the runqueue.
1963 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1965 if (task_contributes_to_load(p
))
1966 rq
->nr_uninterruptible
--;
1968 enqueue_task(rq
, p
, wakeup
);
1973 * deactivate_task - remove a task from the runqueue.
1975 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1977 if (task_contributes_to_load(p
))
1978 rq
->nr_uninterruptible
++;
1980 dequeue_task(rq
, p
, sleep
);
1985 * task_curr - is this task currently executing on a CPU?
1986 * @p: the task in question.
1988 inline int task_curr(const struct task_struct
*p
)
1990 return cpu_curr(task_cpu(p
)) == p
;
1993 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1994 const struct sched_class
*prev_class
,
1995 int oldprio
, int running
)
1997 if (prev_class
!= p
->sched_class
) {
1998 if (prev_class
->switched_from
)
1999 prev_class
->switched_from(rq
, p
, running
);
2000 p
->sched_class
->switched_to(rq
, p
, running
);
2002 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
2007 * Is this task likely cache-hot:
2010 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2014 if (p
->sched_class
!= &fair_sched_class
)
2018 * Buddy candidates are cache hot:
2020 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
2021 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2022 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2025 if (sysctl_sched_migration_cost
== -1)
2027 if (sysctl_sched_migration_cost
== 0)
2030 delta
= now
- p
->se
.exec_start
;
2032 return delta
< (s64
)sysctl_sched_migration_cost
;
2035 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2037 #ifdef CONFIG_SCHED_DEBUG
2039 * We should never call set_task_cpu() on a blocked task,
2040 * ttwu() will sort out the placement.
2042 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
2043 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
2046 trace_sched_migrate_task(p
, new_cpu
);
2048 if (task_cpu(p
) != new_cpu
) {
2049 p
->se
.nr_migrations
++;
2050 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, 1, NULL
, 0);
2053 __set_task_cpu(p
, new_cpu
);
2056 struct migration_req
{
2057 struct list_head list
;
2059 struct task_struct
*task
;
2062 struct completion done
;
2066 * The task's runqueue lock must be held.
2067 * Returns true if you have to wait for migration thread.
2070 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
2072 struct rq
*rq
= task_rq(p
);
2075 * If the task is not on a runqueue (and not running), then
2076 * the next wake-up will properly place the task.
2078 if (!p
->se
.on_rq
&& !task_running(rq
, p
))
2081 init_completion(&req
->done
);
2083 req
->dest_cpu
= dest_cpu
;
2084 list_add(&req
->list
, &rq
->migration_queue
);
2090 * wait_task_context_switch - wait for a thread to complete at least one
2093 * @p must not be current.
2095 void wait_task_context_switch(struct task_struct
*p
)
2097 unsigned long nvcsw
, nivcsw
, flags
;
2105 * The runqueue is assigned before the actual context
2106 * switch. We need to take the runqueue lock.
2108 * We could check initially without the lock but it is
2109 * very likely that we need to take the lock in every
2112 rq
= task_rq_lock(p
, &flags
);
2113 running
= task_running(rq
, p
);
2114 task_rq_unlock(rq
, &flags
);
2116 if (likely(!running
))
2119 * The switch count is incremented before the actual
2120 * context switch. We thus wait for two switches to be
2121 * sure at least one completed.
2123 if ((p
->nvcsw
- nvcsw
) > 1)
2125 if ((p
->nivcsw
- nivcsw
) > 1)
2133 * wait_task_inactive - wait for a thread to unschedule.
2135 * If @match_state is nonzero, it's the @p->state value just checked and
2136 * not expected to change. If it changes, i.e. @p might have woken up,
2137 * then return zero. When we succeed in waiting for @p to be off its CPU,
2138 * we return a positive number (its total switch count). If a second call
2139 * a short while later returns the same number, the caller can be sure that
2140 * @p has remained unscheduled the whole time.
2142 * The caller must ensure that the task *will* unschedule sometime soon,
2143 * else this function might spin for a *long* time. This function can't
2144 * be called with interrupts off, or it may introduce deadlock with
2145 * smp_call_function() if an IPI is sent by the same process we are
2146 * waiting to become inactive.
2148 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2150 unsigned long flags
;
2157 * We do the initial early heuristics without holding
2158 * any task-queue locks at all. We'll only try to get
2159 * the runqueue lock when things look like they will
2165 * If the task is actively running on another CPU
2166 * still, just relax and busy-wait without holding
2169 * NOTE! Since we don't hold any locks, it's not
2170 * even sure that "rq" stays as the right runqueue!
2171 * But we don't care, since "task_running()" will
2172 * return false if the runqueue has changed and p
2173 * is actually now running somewhere else!
2175 while (task_running(rq
, p
)) {
2176 if (match_state
&& unlikely(p
->state
!= match_state
))
2182 * Ok, time to look more closely! We need the rq
2183 * lock now, to be *sure*. If we're wrong, we'll
2184 * just go back and repeat.
2186 rq
= task_rq_lock(p
, &flags
);
2187 trace_sched_wait_task(rq
, p
);
2188 running
= task_running(rq
, p
);
2189 on_rq
= p
->se
.on_rq
;
2191 if (!match_state
|| p
->state
== match_state
)
2192 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2193 task_rq_unlock(rq
, &flags
);
2196 * If it changed from the expected state, bail out now.
2198 if (unlikely(!ncsw
))
2202 * Was it really running after all now that we
2203 * checked with the proper locks actually held?
2205 * Oops. Go back and try again..
2207 if (unlikely(running
)) {
2213 * It's not enough that it's not actively running,
2214 * it must be off the runqueue _entirely_, and not
2217 * So if it was still runnable (but just not actively
2218 * running right now), it's preempted, and we should
2219 * yield - it could be a while.
2221 if (unlikely(on_rq
)) {
2222 schedule_timeout_uninterruptible(1);
2227 * Ahh, all good. It wasn't running, and it wasn't
2228 * runnable, which means that it will never become
2229 * running in the future either. We're all done!
2238 * kick_process - kick a running thread to enter/exit the kernel
2239 * @p: the to-be-kicked thread
2241 * Cause a process which is running on another CPU to enter
2242 * kernel-mode, without any delay. (to get signals handled.)
2244 * NOTE: this function doesnt have to take the runqueue lock,
2245 * because all it wants to ensure is that the remote task enters
2246 * the kernel. If the IPI races and the task has been migrated
2247 * to another CPU then no harm is done and the purpose has been
2250 void kick_process(struct task_struct
*p
)
2256 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2257 smp_send_reschedule(cpu
);
2260 EXPORT_SYMBOL_GPL(kick_process
);
2261 #endif /* CONFIG_SMP */
2264 * task_oncpu_function_call - call a function on the cpu on which a task runs
2265 * @p: the task to evaluate
2266 * @func: the function to be called
2267 * @info: the function call argument
2269 * Calls the function @func when the task is currently running. This might
2270 * be on the current CPU, which just calls the function directly
2272 void task_oncpu_function_call(struct task_struct
*p
,
2273 void (*func
) (void *info
), void *info
)
2280 smp_call_function_single(cpu
, func
, info
, 1);
2285 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2288 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
2290 /* Look for allowed, online CPU in same node. */
2291 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
2292 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
2295 /* Any allowed, online CPU? */
2296 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_active_mask
);
2297 if (dest_cpu
< nr_cpu_ids
)
2300 /* No more Mr. Nice Guy. */
2301 if (dest_cpu
>= nr_cpu_ids
) {
2303 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
2305 dest_cpu
= cpumask_any_and(cpu_active_mask
, &p
->cpus_allowed
);
2308 * Don't tell them about moving exiting tasks or
2309 * kernel threads (both mm NULL), since they never
2312 if (p
->mm
&& printk_ratelimit()) {
2313 printk(KERN_INFO
"process %d (%s) no "
2314 "longer affine to cpu%d\n",
2315 task_pid_nr(p
), p
->comm
, cpu
);
2323 * Gets called from 3 sites (exec, fork, wakeup), since it is called without
2324 * holding rq->lock we need to ensure ->cpus_allowed is stable, this is done
2327 * exec: is unstable, retry loop
2328 * fork & wake-up: serialize ->cpus_allowed against TASK_WAKING
2331 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
2333 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
2336 * In order not to call set_task_cpu() on a blocking task we need
2337 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2340 * Since this is common to all placement strategies, this lives here.
2342 * [ this allows ->select_task() to simply return task_cpu(p) and
2343 * not worry about this generic constraint ]
2345 if (unlikely(!cpumask_test_cpu(cpu
, &p
->cpus_allowed
) ||
2347 cpu
= select_fallback_rq(task_cpu(p
), p
);
2354 * try_to_wake_up - wake up a thread
2355 * @p: the to-be-woken-up thread
2356 * @state: the mask of task states that can be woken
2357 * @sync: do a synchronous wakeup?
2359 * Put it on the run-queue if it's not already there. The "current"
2360 * thread is always on the run-queue (except when the actual
2361 * re-schedule is in progress), and as such you're allowed to do
2362 * the simpler "current->state = TASK_RUNNING" to mark yourself
2363 * runnable without the overhead of this.
2365 * returns failure only if the task is already active.
2367 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
,
2370 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2371 unsigned long flags
;
2372 struct rq
*rq
, *orig_rq
;
2374 if (!sched_feat(SYNC_WAKEUPS
))
2375 wake_flags
&= ~WF_SYNC
;
2377 this_cpu
= get_cpu();
2380 rq
= orig_rq
= task_rq_lock(p
, &flags
);
2381 update_rq_clock(rq
);
2382 if (!(p
->state
& state
))
2392 if (unlikely(task_running(rq
, p
)))
2396 * In order to handle concurrent wakeups and release the rq->lock
2397 * we put the task in TASK_WAKING state.
2399 * First fix up the nr_uninterruptible count:
2401 if (task_contributes_to_load(p
))
2402 rq
->nr_uninterruptible
--;
2403 p
->state
= TASK_WAKING
;
2405 if (p
->sched_class
->task_waking
)
2406 p
->sched_class
->task_waking(rq
, p
);
2408 __task_rq_unlock(rq
);
2410 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
2411 if (cpu
!= orig_cpu
)
2412 set_task_cpu(p
, cpu
);
2414 rq
= __task_rq_lock(p
);
2415 update_rq_clock(rq
);
2417 WARN_ON(p
->state
!= TASK_WAKING
);
2420 #ifdef CONFIG_SCHEDSTATS
2421 schedstat_inc(rq
, ttwu_count
);
2422 if (cpu
== this_cpu
)
2423 schedstat_inc(rq
, ttwu_local
);
2425 struct sched_domain
*sd
;
2426 for_each_domain(this_cpu
, sd
) {
2427 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2428 schedstat_inc(sd
, ttwu_wake_remote
);
2433 #endif /* CONFIG_SCHEDSTATS */
2436 #endif /* CONFIG_SMP */
2437 schedstat_inc(p
, se
.nr_wakeups
);
2438 if (wake_flags
& WF_SYNC
)
2439 schedstat_inc(p
, se
.nr_wakeups_sync
);
2440 if (orig_cpu
!= cpu
)
2441 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2442 if (cpu
== this_cpu
)
2443 schedstat_inc(p
, se
.nr_wakeups_local
);
2445 schedstat_inc(p
, se
.nr_wakeups_remote
);
2446 activate_task(rq
, p
, 1);
2450 * Only attribute actual wakeups done by this task.
2452 if (!in_interrupt()) {
2453 struct sched_entity
*se
= ¤t
->se
;
2454 u64 sample
= se
->sum_exec_runtime
;
2456 if (se
->last_wakeup
)
2457 sample
-= se
->last_wakeup
;
2459 sample
-= se
->start_runtime
;
2460 update_avg(&se
->avg_wakeup
, sample
);
2462 se
->last_wakeup
= se
->sum_exec_runtime
;
2466 trace_sched_wakeup(rq
, p
, success
);
2467 check_preempt_curr(rq
, p
, wake_flags
);
2469 p
->state
= TASK_RUNNING
;
2471 if (p
->sched_class
->task_woken
)
2472 p
->sched_class
->task_woken(rq
, p
);
2474 if (unlikely(rq
->idle_stamp
)) {
2475 u64 delta
= rq
->clock
- rq
->idle_stamp
;
2476 u64 max
= 2*sysctl_sched_migration_cost
;
2481 update_avg(&rq
->avg_idle
, delta
);
2486 task_rq_unlock(rq
, &flags
);
2493 * wake_up_process - Wake up a specific process
2494 * @p: The process to be woken up.
2496 * Attempt to wake up the nominated process and move it to the set of runnable
2497 * processes. Returns 1 if the process was woken up, 0 if it was already
2500 * It may be assumed that this function implies a write memory barrier before
2501 * changing the task state if and only if any tasks are woken up.
2503 int wake_up_process(struct task_struct
*p
)
2505 return try_to_wake_up(p
, TASK_ALL
, 0);
2507 EXPORT_SYMBOL(wake_up_process
);
2509 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2511 return try_to_wake_up(p
, state
, 0);
2515 * Perform scheduler related setup for a newly forked process p.
2516 * p is forked by current.
2518 * __sched_fork() is basic setup used by init_idle() too:
2520 static void __sched_fork(struct task_struct
*p
)
2522 p
->se
.exec_start
= 0;
2523 p
->se
.sum_exec_runtime
= 0;
2524 p
->se
.prev_sum_exec_runtime
= 0;
2525 p
->se
.nr_migrations
= 0;
2526 p
->se
.last_wakeup
= 0;
2527 p
->se
.avg_overlap
= 0;
2528 p
->se
.start_runtime
= 0;
2529 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
2531 #ifdef CONFIG_SCHEDSTATS
2532 p
->se
.wait_start
= 0;
2534 p
->se
.wait_count
= 0;
2537 p
->se
.sleep_start
= 0;
2538 p
->se
.sleep_max
= 0;
2539 p
->se
.sum_sleep_runtime
= 0;
2541 p
->se
.block_start
= 0;
2542 p
->se
.block_max
= 0;
2544 p
->se
.slice_max
= 0;
2546 p
->se
.nr_migrations_cold
= 0;
2547 p
->se
.nr_failed_migrations_affine
= 0;
2548 p
->se
.nr_failed_migrations_running
= 0;
2549 p
->se
.nr_failed_migrations_hot
= 0;
2550 p
->se
.nr_forced_migrations
= 0;
2552 p
->se
.nr_wakeups
= 0;
2553 p
->se
.nr_wakeups_sync
= 0;
2554 p
->se
.nr_wakeups_migrate
= 0;
2555 p
->se
.nr_wakeups_local
= 0;
2556 p
->se
.nr_wakeups_remote
= 0;
2557 p
->se
.nr_wakeups_affine
= 0;
2558 p
->se
.nr_wakeups_affine_attempts
= 0;
2559 p
->se
.nr_wakeups_passive
= 0;
2560 p
->se
.nr_wakeups_idle
= 0;
2564 INIT_LIST_HEAD(&p
->rt
.run_list
);
2566 INIT_LIST_HEAD(&p
->se
.group_node
);
2568 #ifdef CONFIG_PREEMPT_NOTIFIERS
2569 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2574 * fork()/clone()-time setup:
2576 void sched_fork(struct task_struct
*p
, int clone_flags
)
2578 int cpu
= get_cpu();
2582 * We mark the process as waking here. This guarantees that
2583 * nobody will actually run it, and a signal or other external
2584 * event cannot wake it up and insert it on the runqueue either.
2586 p
->state
= TASK_WAKING
;
2589 * Revert to default priority/policy on fork if requested.
2591 if (unlikely(p
->sched_reset_on_fork
)) {
2592 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2593 p
->policy
= SCHED_NORMAL
;
2594 p
->normal_prio
= p
->static_prio
;
2597 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2598 p
->static_prio
= NICE_TO_PRIO(0);
2599 p
->normal_prio
= p
->static_prio
;
2604 * We don't need the reset flag anymore after the fork. It has
2605 * fulfilled its duty:
2607 p
->sched_reset_on_fork
= 0;
2611 * Make sure we do not leak PI boosting priority to the child.
2613 p
->prio
= current
->normal_prio
;
2615 if (!rt_prio(p
->prio
))
2616 p
->sched_class
= &fair_sched_class
;
2618 if (p
->sched_class
->task_fork
)
2619 p
->sched_class
->task_fork(p
);
2621 set_task_cpu(p
, cpu
);
2623 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2624 if (likely(sched_info_on()))
2625 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2627 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2630 #ifdef CONFIG_PREEMPT
2631 /* Want to start with kernel preemption disabled. */
2632 task_thread_info(p
)->preempt_count
= 1;
2634 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2640 * wake_up_new_task - wake up a newly created task for the first time.
2642 * This function will do some initial scheduler statistics housekeeping
2643 * that must be done for every newly created context, then puts the task
2644 * on the runqueue and wakes it.
2646 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2648 unsigned long flags
;
2650 int cpu
= get_cpu();
2654 * Fork balancing, do it here and not earlier because:
2655 * - cpus_allowed can change in the fork path
2656 * - any previously selected cpu might disappear through hotplug
2658 * We still have TASK_WAKING but PF_STARTING is gone now, meaning
2659 * ->cpus_allowed is stable, we have preemption disabled, meaning
2660 * cpu_online_mask is stable.
2662 cpu
= select_task_rq(p
, SD_BALANCE_FORK
, 0);
2663 set_task_cpu(p
, cpu
);
2666 rq
= task_rq_lock(p
, &flags
);
2667 BUG_ON(p
->state
!= TASK_WAKING
);
2668 p
->state
= TASK_RUNNING
;
2669 update_rq_clock(rq
);
2670 activate_task(rq
, p
, 0);
2671 trace_sched_wakeup_new(rq
, p
, 1);
2672 check_preempt_curr(rq
, p
, WF_FORK
);
2674 if (p
->sched_class
->task_woken
)
2675 p
->sched_class
->task_woken(rq
, p
);
2677 task_rq_unlock(rq
, &flags
);
2681 #ifdef CONFIG_PREEMPT_NOTIFIERS
2684 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2685 * @notifier: notifier struct to register
2687 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2689 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2691 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2694 * preempt_notifier_unregister - no longer interested in preemption notifications
2695 * @notifier: notifier struct to unregister
2697 * This is safe to call from within a preemption notifier.
2699 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2701 hlist_del(¬ifier
->link
);
2703 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2705 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2707 struct preempt_notifier
*notifier
;
2708 struct hlist_node
*node
;
2710 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2711 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2715 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2716 struct task_struct
*next
)
2718 struct preempt_notifier
*notifier
;
2719 struct hlist_node
*node
;
2721 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2722 notifier
->ops
->sched_out(notifier
, next
);
2725 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2727 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2732 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2733 struct task_struct
*next
)
2737 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2740 * prepare_task_switch - prepare to switch tasks
2741 * @rq: the runqueue preparing to switch
2742 * @prev: the current task that is being switched out
2743 * @next: the task we are going to switch to.
2745 * This is called with the rq lock held and interrupts off. It must
2746 * be paired with a subsequent finish_task_switch after the context
2749 * prepare_task_switch sets up locking and calls architecture specific
2753 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2754 struct task_struct
*next
)
2756 fire_sched_out_preempt_notifiers(prev
, next
);
2757 prepare_lock_switch(rq
, next
);
2758 prepare_arch_switch(next
);
2762 * finish_task_switch - clean up after a task-switch
2763 * @rq: runqueue associated with task-switch
2764 * @prev: the thread we just switched away from.
2766 * finish_task_switch must be called after the context switch, paired
2767 * with a prepare_task_switch call before the context switch.
2768 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2769 * and do any other architecture-specific cleanup actions.
2771 * Note that we may have delayed dropping an mm in context_switch(). If
2772 * so, we finish that here outside of the runqueue lock. (Doing it
2773 * with the lock held can cause deadlocks; see schedule() for
2776 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2777 __releases(rq
->lock
)
2779 struct mm_struct
*mm
= rq
->prev_mm
;
2785 * A task struct has one reference for the use as "current".
2786 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2787 * schedule one last time. The schedule call will never return, and
2788 * the scheduled task must drop that reference.
2789 * The test for TASK_DEAD must occur while the runqueue locks are
2790 * still held, otherwise prev could be scheduled on another cpu, die
2791 * there before we look at prev->state, and then the reference would
2793 * Manfred Spraul <manfred@colorfullife.com>
2795 prev_state
= prev
->state
;
2796 finish_arch_switch(prev
);
2797 perf_event_task_sched_in(current
, cpu_of(rq
));
2798 finish_lock_switch(rq
, prev
);
2800 fire_sched_in_preempt_notifiers(current
);
2803 if (unlikely(prev_state
== TASK_DEAD
)) {
2805 * Remove function-return probe instances associated with this
2806 * task and put them back on the free list.
2808 kprobe_flush_task(prev
);
2809 put_task_struct(prev
);
2815 /* assumes rq->lock is held */
2816 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2818 if (prev
->sched_class
->pre_schedule
)
2819 prev
->sched_class
->pre_schedule(rq
, prev
);
2822 /* rq->lock is NOT held, but preemption is disabled */
2823 static inline void post_schedule(struct rq
*rq
)
2825 if (rq
->post_schedule
) {
2826 unsigned long flags
;
2828 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2829 if (rq
->curr
->sched_class
->post_schedule
)
2830 rq
->curr
->sched_class
->post_schedule(rq
);
2831 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2833 rq
->post_schedule
= 0;
2839 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2843 static inline void post_schedule(struct rq
*rq
)
2850 * schedule_tail - first thing a freshly forked thread must call.
2851 * @prev: the thread we just switched away from.
2853 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2854 __releases(rq
->lock
)
2856 struct rq
*rq
= this_rq();
2858 finish_task_switch(rq
, prev
);
2861 * FIXME: do we need to worry about rq being invalidated by the
2866 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2867 /* In this case, finish_task_switch does not reenable preemption */
2870 if (current
->set_child_tid
)
2871 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2875 * context_switch - switch to the new MM and the new
2876 * thread's register state.
2879 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2880 struct task_struct
*next
)
2882 struct mm_struct
*mm
, *oldmm
;
2884 prepare_task_switch(rq
, prev
, next
);
2885 trace_sched_switch(rq
, prev
, next
);
2887 oldmm
= prev
->active_mm
;
2889 * For paravirt, this is coupled with an exit in switch_to to
2890 * combine the page table reload and the switch backend into
2893 arch_start_context_switch(prev
);
2896 next
->active_mm
= oldmm
;
2897 atomic_inc(&oldmm
->mm_count
);
2898 enter_lazy_tlb(oldmm
, next
);
2900 switch_mm(oldmm
, mm
, next
);
2902 if (likely(!prev
->mm
)) {
2903 prev
->active_mm
= NULL
;
2904 rq
->prev_mm
= oldmm
;
2907 * Since the runqueue lock will be released by the next
2908 * task (which is an invalid locking op but in the case
2909 * of the scheduler it's an obvious special-case), so we
2910 * do an early lockdep release here:
2912 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2913 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2916 /* Here we just switch the register state and the stack. */
2917 switch_to(prev
, next
, prev
);
2921 * this_rq must be evaluated again because prev may have moved
2922 * CPUs since it called schedule(), thus the 'rq' on its stack
2923 * frame will be invalid.
2925 finish_task_switch(this_rq(), prev
);
2929 * nr_running, nr_uninterruptible and nr_context_switches:
2931 * externally visible scheduler statistics: current number of runnable
2932 * threads, current number of uninterruptible-sleeping threads, total
2933 * number of context switches performed since bootup.
2935 unsigned long nr_running(void)
2937 unsigned long i
, sum
= 0;
2939 for_each_online_cpu(i
)
2940 sum
+= cpu_rq(i
)->nr_running
;
2945 unsigned long nr_uninterruptible(void)
2947 unsigned long i
, sum
= 0;
2949 for_each_possible_cpu(i
)
2950 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2953 * Since we read the counters lockless, it might be slightly
2954 * inaccurate. Do not allow it to go below zero though:
2956 if (unlikely((long)sum
< 0))
2962 unsigned long long nr_context_switches(void)
2965 unsigned long long sum
= 0;
2967 for_each_possible_cpu(i
)
2968 sum
+= cpu_rq(i
)->nr_switches
;
2973 unsigned long nr_iowait(void)
2975 unsigned long i
, sum
= 0;
2977 for_each_possible_cpu(i
)
2978 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2983 unsigned long nr_iowait_cpu(void)
2985 struct rq
*this = this_rq();
2986 return atomic_read(&this->nr_iowait
);
2989 unsigned long this_cpu_load(void)
2991 struct rq
*this = this_rq();
2992 return this->cpu_load
[0];
2996 /* Variables and functions for calc_load */
2997 static atomic_long_t calc_load_tasks
;
2998 static unsigned long calc_load_update
;
2999 unsigned long avenrun
[3];
3000 EXPORT_SYMBOL(avenrun
);
3003 * get_avenrun - get the load average array
3004 * @loads: pointer to dest load array
3005 * @offset: offset to add
3006 * @shift: shift count to shift the result left
3008 * These values are estimates at best, so no need for locking.
3010 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
3012 loads
[0] = (avenrun
[0] + offset
) << shift
;
3013 loads
[1] = (avenrun
[1] + offset
) << shift
;
3014 loads
[2] = (avenrun
[2] + offset
) << shift
;
3017 static unsigned long
3018 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
3021 load
+= active
* (FIXED_1
- exp
);
3022 return load
>> FSHIFT
;
3026 * calc_load - update the avenrun load estimates 10 ticks after the
3027 * CPUs have updated calc_load_tasks.
3029 void calc_global_load(void)
3031 unsigned long upd
= calc_load_update
+ 10;
3034 if (time_before(jiffies
, upd
))
3037 active
= atomic_long_read(&calc_load_tasks
);
3038 active
= active
> 0 ? active
* FIXED_1
: 0;
3040 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3041 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3042 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3044 calc_load_update
+= LOAD_FREQ
;
3048 * Either called from update_cpu_load() or from a cpu going idle
3050 static void calc_load_account_active(struct rq
*this_rq
)
3052 long nr_active
, delta
;
3054 nr_active
= this_rq
->nr_running
;
3055 nr_active
+= (long) this_rq
->nr_uninterruptible
;
3057 if (nr_active
!= this_rq
->calc_load_active
) {
3058 delta
= nr_active
- this_rq
->calc_load_active
;
3059 this_rq
->calc_load_active
= nr_active
;
3060 atomic_long_add(delta
, &calc_load_tasks
);
3065 * Update rq->cpu_load[] statistics. This function is usually called every
3066 * scheduler tick (TICK_NSEC).
3068 static void update_cpu_load(struct rq
*this_rq
)
3070 unsigned long this_load
= this_rq
->load
.weight
;
3073 this_rq
->nr_load_updates
++;
3075 /* Update our load: */
3076 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3077 unsigned long old_load
, new_load
;
3079 /* scale is effectively 1 << i now, and >> i divides by scale */
3081 old_load
= this_rq
->cpu_load
[i
];
3082 new_load
= this_load
;
3084 * Round up the averaging division if load is increasing. This
3085 * prevents us from getting stuck on 9 if the load is 10, for
3088 if (new_load
> old_load
)
3089 new_load
+= scale
-1;
3090 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3093 if (time_after_eq(jiffies
, this_rq
->calc_load_update
)) {
3094 this_rq
->calc_load_update
+= LOAD_FREQ
;
3095 calc_load_account_active(this_rq
);
3102 * double_rq_lock - safely lock two runqueues
3104 * Note this does not disable interrupts like task_rq_lock,
3105 * you need to do so manually before calling.
3107 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
3108 __acquires(rq1
->lock
)
3109 __acquires(rq2
->lock
)
3111 BUG_ON(!irqs_disabled());
3113 raw_spin_lock(&rq1
->lock
);
3114 __acquire(rq2
->lock
); /* Fake it out ;) */
3117 raw_spin_lock(&rq1
->lock
);
3118 raw_spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
3120 raw_spin_lock(&rq2
->lock
);
3121 raw_spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
3124 update_rq_clock(rq1
);
3125 update_rq_clock(rq2
);
3129 * double_rq_unlock - safely unlock two runqueues
3131 * Note this does not restore interrupts like task_rq_unlock,
3132 * you need to do so manually after calling.
3134 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
3135 __releases(rq1
->lock
)
3136 __releases(rq2
->lock
)
3138 raw_spin_unlock(&rq1
->lock
);
3140 raw_spin_unlock(&rq2
->lock
);
3142 __release(rq2
->lock
);
3146 * sched_exec - execve() is a valuable balancing opportunity, because at
3147 * this point the task has the smallest effective memory and cache footprint.
3149 void sched_exec(void)
3151 struct task_struct
*p
= current
;
3152 struct migration_req req
;
3153 int dest_cpu
, this_cpu
;
3154 unsigned long flags
;
3158 this_cpu
= get_cpu();
3159 dest_cpu
= select_task_rq(p
, SD_BALANCE_EXEC
, 0);
3160 if (dest_cpu
== this_cpu
) {
3165 rq
= task_rq_lock(p
, &flags
);
3169 * select_task_rq() can race against ->cpus_allowed
3171 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
3172 || unlikely(!cpu_active(dest_cpu
))) {
3173 task_rq_unlock(rq
, &flags
);
3177 /* force the process onto the specified CPU */
3178 if (migrate_task(p
, dest_cpu
, &req
)) {
3179 /* Need to wait for migration thread (might exit: take ref). */
3180 struct task_struct
*mt
= rq
->migration_thread
;
3182 get_task_struct(mt
);
3183 task_rq_unlock(rq
, &flags
);
3184 wake_up_process(mt
);
3185 put_task_struct(mt
);
3186 wait_for_completion(&req
.done
);
3190 task_rq_unlock(rq
, &flags
);
3194 * pull_task - move a task from a remote runqueue to the local runqueue.
3195 * Both runqueues must be locked.
3197 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
3198 struct rq
*this_rq
, int this_cpu
)
3200 deactivate_task(src_rq
, p
, 0);
3201 set_task_cpu(p
, this_cpu
);
3202 activate_task(this_rq
, p
, 0);
3203 check_preempt_curr(this_rq
, p
, 0);
3207 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3210 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
3211 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3214 int tsk_cache_hot
= 0;
3216 * We do not migrate tasks that are:
3217 * 1) running (obviously), or
3218 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3219 * 3) are cache-hot on their current CPU.
3221 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
3222 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
3227 if (task_running(rq
, p
)) {
3228 schedstat_inc(p
, se
.nr_failed_migrations_running
);
3233 * Aggressive migration if:
3234 * 1) task is cache cold, or
3235 * 2) too many balance attempts have failed.
3238 tsk_cache_hot
= task_hot(p
, rq
->clock
, sd
);
3239 if (!tsk_cache_hot
||
3240 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
3241 #ifdef CONFIG_SCHEDSTATS
3242 if (tsk_cache_hot
) {
3243 schedstat_inc(sd
, lb_hot_gained
[idle
]);
3244 schedstat_inc(p
, se
.nr_forced_migrations
);
3250 if (tsk_cache_hot
) {
3251 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
3257 static unsigned long
3258 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3259 unsigned long max_load_move
, struct sched_domain
*sd
,
3260 enum cpu_idle_type idle
, int *all_pinned
,
3261 int *this_best_prio
, struct rq_iterator
*iterator
)
3263 int loops
= 0, pulled
= 0, pinned
= 0;
3264 struct task_struct
*p
;
3265 long rem_load_move
= max_load_move
;
3267 if (max_load_move
== 0)
3273 * Start the load-balancing iterator:
3275 p
= iterator
->start(iterator
->arg
);
3277 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3280 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
3281 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3282 p
= iterator
->next(iterator
->arg
);
3286 pull_task(busiest
, p
, this_rq
, this_cpu
);
3288 rem_load_move
-= p
->se
.load
.weight
;
3290 #ifdef CONFIG_PREEMPT
3292 * NEWIDLE balancing is a source of latency, so preemptible kernels
3293 * will stop after the first task is pulled to minimize the critical
3296 if (idle
== CPU_NEWLY_IDLE
)
3301 * We only want to steal up to the prescribed amount of weighted load.
3303 if (rem_load_move
> 0) {
3304 if (p
->prio
< *this_best_prio
)
3305 *this_best_prio
= p
->prio
;
3306 p
= iterator
->next(iterator
->arg
);
3311 * Right now, this is one of only two places pull_task() is called,
3312 * so we can safely collect pull_task() stats here rather than
3313 * inside pull_task().
3315 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3318 *all_pinned
= pinned
;
3320 return max_load_move
- rem_load_move
;
3324 * move_tasks tries to move up to max_load_move weighted load from busiest to
3325 * this_rq, as part of a balancing operation within domain "sd".
3326 * Returns 1 if successful and 0 otherwise.
3328 * Called with both runqueues locked.
3330 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3331 unsigned long max_load_move
,
3332 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3335 const struct sched_class
*class = sched_class_highest
;
3336 unsigned long total_load_moved
= 0;
3337 int this_best_prio
= this_rq
->curr
->prio
;
3341 class->load_balance(this_rq
, this_cpu
, busiest
,
3342 max_load_move
- total_load_moved
,
3343 sd
, idle
, all_pinned
, &this_best_prio
);
3344 class = class->next
;
3346 #ifdef CONFIG_PREEMPT
3348 * NEWIDLE balancing is a source of latency, so preemptible
3349 * kernels will stop after the first task is pulled to minimize
3350 * the critical section.
3352 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3355 } while (class && max_load_move
> total_load_moved
);
3357 return total_load_moved
> 0;
3361 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3362 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3363 struct rq_iterator
*iterator
)
3365 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3369 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3370 pull_task(busiest
, p
, this_rq
, this_cpu
);
3372 * Right now, this is only the second place pull_task()
3373 * is called, so we can safely collect pull_task()
3374 * stats here rather than inside pull_task().
3376 schedstat_inc(sd
, lb_gained
[idle
]);
3380 p
= iterator
->next(iterator
->arg
);
3387 * move_one_task tries to move exactly one task from busiest to this_rq, as
3388 * part of active balancing operations within "domain".
3389 * Returns 1 if successful and 0 otherwise.
3391 * Called with both runqueues locked.
3393 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3394 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3396 const struct sched_class
*class;
3398 for_each_class(class) {
3399 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3405 /********** Helpers for find_busiest_group ************************/
3407 * sd_lb_stats - Structure to store the statistics of a sched_domain
3408 * during load balancing.
3410 struct sd_lb_stats
{
3411 struct sched_group
*busiest
; /* Busiest group in this sd */
3412 struct sched_group
*this; /* Local group in this sd */
3413 unsigned long total_load
; /* Total load of all groups in sd */
3414 unsigned long total_pwr
; /* Total power of all groups in sd */
3415 unsigned long avg_load
; /* Average load across all groups in sd */
3417 /** Statistics of this group */
3418 unsigned long this_load
;
3419 unsigned long this_load_per_task
;
3420 unsigned long this_nr_running
;
3422 /* Statistics of the busiest group */
3423 unsigned long max_load
;
3424 unsigned long busiest_load_per_task
;
3425 unsigned long busiest_nr_running
;
3426 unsigned long busiest_group_capacity
;
3428 int group_imb
; /* Is there imbalance in this sd */
3429 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3430 int power_savings_balance
; /* Is powersave balance needed for this sd */
3431 struct sched_group
*group_min
; /* Least loaded group in sd */
3432 struct sched_group
*group_leader
; /* Group which relieves group_min */
3433 unsigned long min_load_per_task
; /* load_per_task in group_min */
3434 unsigned long leader_nr_running
; /* Nr running of group_leader */
3435 unsigned long min_nr_running
; /* Nr running of group_min */
3440 * sg_lb_stats - stats of a sched_group required for load_balancing
3442 struct sg_lb_stats
{
3443 unsigned long avg_load
; /*Avg load across the CPUs of the group */
3444 unsigned long group_load
; /* Total load over the CPUs of the group */
3445 unsigned long sum_nr_running
; /* Nr tasks running in the group */
3446 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
3447 unsigned long group_capacity
;
3448 int group_imb
; /* Is there an imbalance in the group ? */
3452 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3453 * @group: The group whose first cpu is to be returned.
3455 static inline unsigned int group_first_cpu(struct sched_group
*group
)
3457 return cpumask_first(sched_group_cpus(group
));
3461 * get_sd_load_idx - Obtain the load index for a given sched domain.
3462 * @sd: The sched_domain whose load_idx is to be obtained.
3463 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3465 static inline int get_sd_load_idx(struct sched_domain
*sd
,
3466 enum cpu_idle_type idle
)
3472 load_idx
= sd
->busy_idx
;
3475 case CPU_NEWLY_IDLE
:
3476 load_idx
= sd
->newidle_idx
;
3479 load_idx
= sd
->idle_idx
;
3487 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3489 * init_sd_power_savings_stats - Initialize power savings statistics for
3490 * the given sched_domain, during load balancing.
3492 * @sd: Sched domain whose power-savings statistics are to be initialized.
3493 * @sds: Variable containing the statistics for sd.
3494 * @idle: Idle status of the CPU at which we're performing load-balancing.
3496 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3497 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3500 * Busy processors will not participate in power savings
3503 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3504 sds
->power_savings_balance
= 0;
3506 sds
->power_savings_balance
= 1;
3507 sds
->min_nr_running
= ULONG_MAX
;
3508 sds
->leader_nr_running
= 0;
3513 * update_sd_power_savings_stats - Update the power saving stats for a
3514 * sched_domain while performing load balancing.
3516 * @group: sched_group belonging to the sched_domain under consideration.
3517 * @sds: Variable containing the statistics of the sched_domain
3518 * @local_group: Does group contain the CPU for which we're performing
3520 * @sgs: Variable containing the statistics of the group.
3522 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3523 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3526 if (!sds
->power_savings_balance
)
3530 * If the local group is idle or completely loaded
3531 * no need to do power savings balance at this domain
3533 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
3534 !sds
->this_nr_running
))
3535 sds
->power_savings_balance
= 0;
3538 * If a group is already running at full capacity or idle,
3539 * don't include that group in power savings calculations
3541 if (!sds
->power_savings_balance
||
3542 sgs
->sum_nr_running
>= sgs
->group_capacity
||
3543 !sgs
->sum_nr_running
)
3547 * Calculate the group which has the least non-idle load.
3548 * This is the group from where we need to pick up the load
3551 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
3552 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
3553 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
3554 sds
->group_min
= group
;
3555 sds
->min_nr_running
= sgs
->sum_nr_running
;
3556 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
3557 sgs
->sum_nr_running
;
3561 * Calculate the group which is almost near its
3562 * capacity but still has some space to pick up some load
3563 * from other group and save more power
3565 if (sgs
->sum_nr_running
+ 1 > sgs
->group_capacity
)
3568 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
3569 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
3570 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
3571 sds
->group_leader
= group
;
3572 sds
->leader_nr_running
= sgs
->sum_nr_running
;
3577 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3578 * @sds: Variable containing the statistics of the sched_domain
3579 * under consideration.
3580 * @this_cpu: Cpu at which we're currently performing load-balancing.
3581 * @imbalance: Variable to store the imbalance.
3584 * Check if we have potential to perform some power-savings balance.
3585 * If yes, set the busiest group to be the least loaded group in the
3586 * sched_domain, so that it's CPUs can be put to idle.
3588 * Returns 1 if there is potential to perform power-savings balance.
3591 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3592 int this_cpu
, unsigned long *imbalance
)
3594 if (!sds
->power_savings_balance
)
3597 if (sds
->this != sds
->group_leader
||
3598 sds
->group_leader
== sds
->group_min
)
3601 *imbalance
= sds
->min_load_per_task
;
3602 sds
->busiest
= sds
->group_min
;
3607 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3608 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3609 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3614 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3615 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3620 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3621 int this_cpu
, unsigned long *imbalance
)
3625 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3628 unsigned long default_scale_freq_power(struct sched_domain
*sd
, int cpu
)
3630 return SCHED_LOAD_SCALE
;
3633 unsigned long __weak
arch_scale_freq_power(struct sched_domain
*sd
, int cpu
)
3635 return default_scale_freq_power(sd
, cpu
);
3638 unsigned long default_scale_smt_power(struct sched_domain
*sd
, int cpu
)
3640 unsigned long weight
= cpumask_weight(sched_domain_span(sd
));
3641 unsigned long smt_gain
= sd
->smt_gain
;
3648 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
3650 return default_scale_smt_power(sd
, cpu
);
3653 unsigned long scale_rt_power(int cpu
)
3655 struct rq
*rq
= cpu_rq(cpu
);
3656 u64 total
, available
;
3658 sched_avg_update(rq
);
3660 total
= sched_avg_period() + (rq
->clock
- rq
->age_stamp
);
3661 available
= total
- rq
->rt_avg
;
3663 if (unlikely((s64
)total
< SCHED_LOAD_SCALE
))
3664 total
= SCHED_LOAD_SCALE
;
3666 total
>>= SCHED_LOAD_SHIFT
;
3668 return div_u64(available
, total
);
3671 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
3673 unsigned long weight
= cpumask_weight(sched_domain_span(sd
));
3674 unsigned long power
= SCHED_LOAD_SCALE
;
3675 struct sched_group
*sdg
= sd
->groups
;
3677 if (sched_feat(ARCH_POWER
))
3678 power
*= arch_scale_freq_power(sd
, cpu
);
3680 power
*= default_scale_freq_power(sd
, cpu
);
3682 power
>>= SCHED_LOAD_SHIFT
;
3684 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
3685 if (sched_feat(ARCH_POWER
))
3686 power
*= arch_scale_smt_power(sd
, cpu
);
3688 power
*= default_scale_smt_power(sd
, cpu
);
3690 power
>>= SCHED_LOAD_SHIFT
;
3693 power
*= scale_rt_power(cpu
);
3694 power
>>= SCHED_LOAD_SHIFT
;
3699 sdg
->cpu_power
= power
;
3702 static void update_group_power(struct sched_domain
*sd
, int cpu
)
3704 struct sched_domain
*child
= sd
->child
;
3705 struct sched_group
*group
, *sdg
= sd
->groups
;
3706 unsigned long power
;
3709 update_cpu_power(sd
, cpu
);
3715 group
= child
->groups
;
3717 power
+= group
->cpu_power
;
3718 group
= group
->next
;
3719 } while (group
!= child
->groups
);
3721 sdg
->cpu_power
= power
;
3725 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3726 * @sd: The sched_domain whose statistics are to be updated.
3727 * @group: sched_group whose statistics are to be updated.
3728 * @this_cpu: Cpu for which load balance is currently performed.
3729 * @idle: Idle status of this_cpu
3730 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3731 * @sd_idle: Idle status of the sched_domain containing group.
3732 * @local_group: Does group contain this_cpu.
3733 * @cpus: Set of cpus considered for load balancing.
3734 * @balance: Should we balance.
3735 * @sgs: variable to hold the statistics for this group.
3737 static inline void update_sg_lb_stats(struct sched_domain
*sd
,
3738 struct sched_group
*group
, int this_cpu
,
3739 enum cpu_idle_type idle
, int load_idx
, int *sd_idle
,
3740 int local_group
, const struct cpumask
*cpus
,
3741 int *balance
, struct sg_lb_stats
*sgs
)
3743 unsigned long load
, max_cpu_load
, min_cpu_load
;
3745 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3746 unsigned long avg_load_per_task
= 0;
3749 balance_cpu
= group_first_cpu(group
);
3750 if (balance_cpu
== this_cpu
)
3751 update_group_power(sd
, this_cpu
);
3754 /* Tally up the load of all CPUs in the group */
3756 min_cpu_load
= ~0UL;
3758 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3759 struct rq
*rq
= cpu_rq(i
);
3761 if (*sd_idle
&& rq
->nr_running
)
3764 /* Bias balancing toward cpus of our domain */
3766 if (idle_cpu(i
) && !first_idle_cpu
) {
3771 load
= target_load(i
, load_idx
);
3773 load
= source_load(i
, load_idx
);
3774 if (load
> max_cpu_load
)
3775 max_cpu_load
= load
;
3776 if (min_cpu_load
> load
)
3777 min_cpu_load
= load
;
3780 sgs
->group_load
+= load
;
3781 sgs
->sum_nr_running
+= rq
->nr_running
;
3782 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
3787 * First idle cpu or the first cpu(busiest) in this sched group
3788 * is eligible for doing load balancing at this and above
3789 * domains. In the newly idle case, we will allow all the cpu's
3790 * to do the newly idle load balance.
3792 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3793 balance_cpu
!= this_cpu
&& balance
) {
3798 /* Adjust by relative CPU power of the group */
3799 sgs
->avg_load
= (sgs
->group_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
3802 * Consider the group unbalanced when the imbalance is larger
3803 * than the average weight of two tasks.
3805 * APZ: with cgroup the avg task weight can vary wildly and
3806 * might not be a suitable number - should we keep a
3807 * normalized nr_running number somewhere that negates
3810 if (sgs
->sum_nr_running
)
3811 avg_load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
3813 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3816 sgs
->group_capacity
=
3817 DIV_ROUND_CLOSEST(group
->cpu_power
, SCHED_LOAD_SCALE
);
3821 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3822 * @sd: sched_domain whose statistics are to be updated.
3823 * @this_cpu: Cpu for which load balance is currently performed.
3824 * @idle: Idle status of this_cpu
3825 * @sd_idle: Idle status of the sched_domain containing group.
3826 * @cpus: Set of cpus considered for load balancing.
3827 * @balance: Should we balance.
3828 * @sds: variable to hold the statistics for this sched_domain.
3830 static inline void update_sd_lb_stats(struct sched_domain
*sd
, int this_cpu
,
3831 enum cpu_idle_type idle
, int *sd_idle
,
3832 const struct cpumask
*cpus
, int *balance
,
3833 struct sd_lb_stats
*sds
)
3835 struct sched_domain
*child
= sd
->child
;
3836 struct sched_group
*group
= sd
->groups
;
3837 struct sg_lb_stats sgs
;
3838 int load_idx
, prefer_sibling
= 0;
3840 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
3843 init_sd_power_savings_stats(sd
, sds
, idle
);
3844 load_idx
= get_sd_load_idx(sd
, idle
);
3849 local_group
= cpumask_test_cpu(this_cpu
,
3850 sched_group_cpus(group
));
3851 memset(&sgs
, 0, sizeof(sgs
));
3852 update_sg_lb_stats(sd
, group
, this_cpu
, idle
, load_idx
, sd_idle
,
3853 local_group
, cpus
, balance
, &sgs
);
3855 if (local_group
&& balance
&& !(*balance
))
3858 sds
->total_load
+= sgs
.group_load
;
3859 sds
->total_pwr
+= group
->cpu_power
;
3862 * In case the child domain prefers tasks go to siblings
3863 * first, lower the group capacity to one so that we'll try
3864 * and move all the excess tasks away.
3867 sgs
.group_capacity
= min(sgs
.group_capacity
, 1UL);
3870 sds
->this_load
= sgs
.avg_load
;
3872 sds
->this_nr_running
= sgs
.sum_nr_running
;
3873 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
3874 } else if (sgs
.avg_load
> sds
->max_load
&&
3875 (sgs
.sum_nr_running
> sgs
.group_capacity
||
3877 sds
->max_load
= sgs
.avg_load
;
3878 sds
->busiest
= group
;
3879 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
3880 sds
->busiest_group_capacity
= sgs
.group_capacity
;
3881 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
3882 sds
->group_imb
= sgs
.group_imb
;
3885 update_sd_power_savings_stats(group
, sds
, local_group
, &sgs
);
3886 group
= group
->next
;
3887 } while (group
!= sd
->groups
);
3891 * fix_small_imbalance - Calculate the minor imbalance that exists
3892 * amongst the groups of a sched_domain, during
3894 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3895 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3896 * @imbalance: Variable to store the imbalance.
3898 static inline void fix_small_imbalance(struct sd_lb_stats
*sds
,
3899 int this_cpu
, unsigned long *imbalance
)
3901 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
3902 unsigned int imbn
= 2;
3903 unsigned long scaled_busy_load_per_task
;
3905 if (sds
->this_nr_running
) {
3906 sds
->this_load_per_task
/= sds
->this_nr_running
;
3907 if (sds
->busiest_load_per_task
>
3908 sds
->this_load_per_task
)
3911 sds
->this_load_per_task
=
3912 cpu_avg_load_per_task(this_cpu
);
3914 scaled_busy_load_per_task
= sds
->busiest_load_per_task
3916 scaled_busy_load_per_task
/= sds
->busiest
->cpu_power
;
3918 if (sds
->max_load
- sds
->this_load
+ scaled_busy_load_per_task
>=
3919 (scaled_busy_load_per_task
* imbn
)) {
3920 *imbalance
= sds
->busiest_load_per_task
;
3925 * OK, we don't have enough imbalance to justify moving tasks,
3926 * however we may be able to increase total CPU power used by
3930 pwr_now
+= sds
->busiest
->cpu_power
*
3931 min(sds
->busiest_load_per_task
, sds
->max_load
);
3932 pwr_now
+= sds
->this->cpu_power
*
3933 min(sds
->this_load_per_task
, sds
->this_load
);
3934 pwr_now
/= SCHED_LOAD_SCALE
;
3936 /* Amount of load we'd subtract */
3937 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
3938 sds
->busiest
->cpu_power
;
3939 if (sds
->max_load
> tmp
)
3940 pwr_move
+= sds
->busiest
->cpu_power
*
3941 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
3943 /* Amount of load we'd add */
3944 if (sds
->max_load
* sds
->busiest
->cpu_power
<
3945 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
)
3946 tmp
= (sds
->max_load
* sds
->busiest
->cpu_power
) /
3947 sds
->this->cpu_power
;
3949 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
3950 sds
->this->cpu_power
;
3951 pwr_move
+= sds
->this->cpu_power
*
3952 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
3953 pwr_move
/= SCHED_LOAD_SCALE
;
3955 /* Move if we gain throughput */
3956 if (pwr_move
> pwr_now
)
3957 *imbalance
= sds
->busiest_load_per_task
;
3961 * calculate_imbalance - Calculate the amount of imbalance present within the
3962 * groups of a given sched_domain during load balance.
3963 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3964 * @this_cpu: Cpu for which currently load balance is being performed.
3965 * @imbalance: The variable to store the imbalance.
3967 static inline void calculate_imbalance(struct sd_lb_stats
*sds
, int this_cpu
,
3968 unsigned long *imbalance
)
3970 unsigned long max_pull
, load_above_capacity
= ~0UL;
3972 sds
->busiest_load_per_task
/= sds
->busiest_nr_running
;
3973 if (sds
->group_imb
) {
3974 sds
->busiest_load_per_task
=
3975 min(sds
->busiest_load_per_task
, sds
->avg_load
);
3979 * In the presence of smp nice balancing, certain scenarios can have
3980 * max load less than avg load(as we skip the groups at or below
3981 * its cpu_power, while calculating max_load..)
3983 if (sds
->max_load
< sds
->avg_load
) {
3985 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3988 if (!sds
->group_imb
) {
3990 * Don't want to pull so many tasks that a group would go idle.
3992 load_above_capacity
= (sds
->busiest_nr_running
-
3993 sds
->busiest_group_capacity
);
3995 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_LOAD_SCALE
);
3997 load_above_capacity
/= sds
->busiest
->cpu_power
;
4001 * We're trying to get all the cpus to the average_load, so we don't
4002 * want to push ourselves above the average load, nor do we wish to
4003 * reduce the max loaded cpu below the average load. At the same time,
4004 * we also don't want to reduce the group load below the group capacity
4005 * (so that we can implement power-savings policies etc). Thus we look
4006 * for the minimum possible imbalance.
4007 * Be careful of negative numbers as they'll appear as very large values
4008 * with unsigned longs.
4010 max_pull
= min(sds
->max_load
- sds
->avg_load
, load_above_capacity
);
4012 /* How much load to actually move to equalise the imbalance */
4013 *imbalance
= min(max_pull
* sds
->busiest
->cpu_power
,
4014 (sds
->avg_load
- sds
->this_load
) * sds
->this->cpu_power
)
4018 * if *imbalance is less than the average load per runnable task
4019 * there is no gaurantee that any tasks will be moved so we'll have
4020 * a think about bumping its value to force at least one task to be
4023 if (*imbalance
< sds
->busiest_load_per_task
)
4024 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
4027 /******* find_busiest_group() helpers end here *********************/
4030 * find_busiest_group - Returns the busiest group within the sched_domain
4031 * if there is an imbalance. If there isn't an imbalance, and
4032 * the user has opted for power-savings, it returns a group whose
4033 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4034 * such a group exists.
4036 * Also calculates the amount of weighted load which should be moved
4037 * to restore balance.
4039 * @sd: The sched_domain whose busiest group is to be returned.
4040 * @this_cpu: The cpu for which load balancing is currently being performed.
4041 * @imbalance: Variable which stores amount of weighted load which should
4042 * be moved to restore balance/put a group to idle.
4043 * @idle: The idle status of this_cpu.
4044 * @sd_idle: The idleness of sd
4045 * @cpus: The set of CPUs under consideration for load-balancing.
4046 * @balance: Pointer to a variable indicating if this_cpu
4047 * is the appropriate cpu to perform load balancing at this_level.
4049 * Returns: - the busiest group if imbalance exists.
4050 * - If no imbalance and user has opted for power-savings balance,
4051 * return the least loaded group whose CPUs can be
4052 * put to idle by rebalancing its tasks onto our group.
4054 static struct sched_group
*
4055 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
4056 unsigned long *imbalance
, enum cpu_idle_type idle
,
4057 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
4059 struct sd_lb_stats sds
;
4061 memset(&sds
, 0, sizeof(sds
));
4064 * Compute the various statistics relavent for load balancing at
4067 update_sd_lb_stats(sd
, this_cpu
, idle
, sd_idle
, cpus
,
4070 /* Cases where imbalance does not exist from POV of this_cpu */
4071 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4073 * 2) There is no busy sibling group to pull from.
4074 * 3) This group is the busiest group.
4075 * 4) This group is more busy than the avg busieness at this
4077 * 5) The imbalance is within the specified limit.
4079 if (balance
&& !(*balance
))
4082 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
4085 if (sds
.this_load
>= sds
.max_load
)
4088 sds
.avg_load
= (SCHED_LOAD_SCALE
* sds
.total_load
) / sds
.total_pwr
;
4090 if (sds
.this_load
>= sds
.avg_load
)
4093 if (100 * sds
.max_load
<= sd
->imbalance_pct
* sds
.this_load
)
4096 /* Looks like there is an imbalance. Compute it */
4097 calculate_imbalance(&sds
, this_cpu
, imbalance
);
4102 * There is no obvious imbalance. But check if we can do some balancing
4105 if (check_power_save_busiest_group(&sds
, this_cpu
, imbalance
))
4113 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4116 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
4117 unsigned long imbalance
, const struct cpumask
*cpus
)
4119 struct rq
*busiest
= NULL
, *rq
;
4120 unsigned long max_load
= 0;
4123 for_each_cpu(i
, sched_group_cpus(group
)) {
4124 unsigned long power
= power_of(i
);
4125 unsigned long capacity
= DIV_ROUND_CLOSEST(power
, SCHED_LOAD_SCALE
);
4128 if (!cpumask_test_cpu(i
, cpus
))
4132 wl
= weighted_cpuload(i
);
4135 * When comparing with imbalance, use weighted_cpuload()
4136 * which is not scaled with the cpu power.
4138 if (capacity
&& rq
->nr_running
== 1 && wl
> imbalance
)
4142 * For the load comparisons with the other cpu's, consider
4143 * the weighted_cpuload() scaled with the cpu power, so that
4144 * the load can be moved away from the cpu that is potentially
4145 * running at a lower capacity.
4147 wl
= (wl
* SCHED_LOAD_SCALE
) / power
;
4149 if (wl
> max_load
) {
4159 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4160 * so long as it is large enough.
4162 #define MAX_PINNED_INTERVAL 512
4164 /* Working cpumask for load_balance and load_balance_newidle. */
4165 static DEFINE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
4168 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4169 * tasks if there is an imbalance.
4171 static int load_balance(int this_cpu
, struct rq
*this_rq
,
4172 struct sched_domain
*sd
, enum cpu_idle_type idle
,
4175 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
4176 struct sched_group
*group
;
4177 unsigned long imbalance
;
4179 unsigned long flags
;
4180 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4182 cpumask_copy(cpus
, cpu_active_mask
);
4185 * When power savings policy is enabled for the parent domain, idle
4186 * sibling can pick up load irrespective of busy siblings. In this case,
4187 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4188 * portraying it as CPU_NOT_IDLE.
4190 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4191 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4194 schedstat_inc(sd
, lb_count
[idle
]);
4198 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
4205 schedstat_inc(sd
, lb_nobusyg
[idle
]);
4209 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
4211 schedstat_inc(sd
, lb_nobusyq
[idle
]);
4215 BUG_ON(busiest
== this_rq
);
4217 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
4220 if (busiest
->nr_running
> 1) {
4222 * Attempt to move tasks. If find_busiest_group has found
4223 * an imbalance but busiest->nr_running <= 1, the group is
4224 * still unbalanced. ld_moved simply stays zero, so it is
4225 * correctly treated as an imbalance.
4227 local_irq_save(flags
);
4228 double_rq_lock(this_rq
, busiest
);
4229 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4230 imbalance
, sd
, idle
, &all_pinned
);
4231 double_rq_unlock(this_rq
, busiest
);
4232 local_irq_restore(flags
);
4235 * some other cpu did the load balance for us.
4237 if (ld_moved
&& this_cpu
!= smp_processor_id())
4238 resched_cpu(this_cpu
);
4240 /* All tasks on this runqueue were pinned by CPU affinity */
4241 if (unlikely(all_pinned
)) {
4242 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4243 if (!cpumask_empty(cpus
))
4250 schedstat_inc(sd
, lb_failed
[idle
]);
4251 sd
->nr_balance_failed
++;
4253 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
4255 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
4257 /* don't kick the migration_thread, if the curr
4258 * task on busiest cpu can't be moved to this_cpu
4260 if (!cpumask_test_cpu(this_cpu
,
4261 &busiest
->curr
->cpus_allowed
)) {
4262 raw_spin_unlock_irqrestore(&busiest
->lock
,
4265 goto out_one_pinned
;
4268 if (!busiest
->active_balance
) {
4269 busiest
->active_balance
= 1;
4270 busiest
->push_cpu
= this_cpu
;
4273 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
4275 wake_up_process(busiest
->migration_thread
);
4278 * We've kicked active balancing, reset the failure
4281 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
4284 sd
->nr_balance_failed
= 0;
4286 if (likely(!active_balance
)) {
4287 /* We were unbalanced, so reset the balancing interval */
4288 sd
->balance_interval
= sd
->min_interval
;
4291 * If we've begun active balancing, start to back off. This
4292 * case may not be covered by the all_pinned logic if there
4293 * is only 1 task on the busy runqueue (because we don't call
4296 if (sd
->balance_interval
< sd
->max_interval
)
4297 sd
->balance_interval
*= 2;
4300 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4301 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4307 schedstat_inc(sd
, lb_balanced
[idle
]);
4309 sd
->nr_balance_failed
= 0;
4312 /* tune up the balancing interval */
4313 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
4314 (sd
->balance_interval
< sd
->max_interval
))
4315 sd
->balance_interval
*= 2;
4317 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4318 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4329 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4330 * tasks if there is an imbalance.
4332 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4333 * this_rq is locked.
4336 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
4338 struct sched_group
*group
;
4339 struct rq
*busiest
= NULL
;
4340 unsigned long imbalance
;
4344 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4346 cpumask_copy(cpus
, cpu_active_mask
);
4349 * When power savings policy is enabled for the parent domain, idle
4350 * sibling can pick up load irrespective of busy siblings. In this case,
4351 * let the state of idle sibling percolate up as IDLE, instead of
4352 * portraying it as CPU_NOT_IDLE.
4354 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
4355 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4358 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
4360 update_shares_locked(this_rq
, sd
);
4361 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
4362 &sd_idle
, cpus
, NULL
);
4364 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
4368 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
4370 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
4374 BUG_ON(busiest
== this_rq
);
4376 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
4379 if (busiest
->nr_running
> 1) {
4380 /* Attempt to move tasks */
4381 double_lock_balance(this_rq
, busiest
);
4382 /* this_rq->clock is already updated */
4383 update_rq_clock(busiest
);
4384 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4385 imbalance
, sd
, CPU_NEWLY_IDLE
,
4387 double_unlock_balance(this_rq
, busiest
);
4389 if (unlikely(all_pinned
)) {
4390 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4391 if (!cpumask_empty(cpus
))
4397 int active_balance
= 0;
4399 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
4400 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4401 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4404 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
4407 if (sd
->nr_balance_failed
++ < 2)
4411 * The only task running in a non-idle cpu can be moved to this
4412 * cpu in an attempt to completely freeup the other CPU
4413 * package. The same method used to move task in load_balance()
4414 * have been extended for load_balance_newidle() to speedup
4415 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4417 * The package power saving logic comes from
4418 * find_busiest_group(). If there are no imbalance, then
4419 * f_b_g() will return NULL. However when sched_mc={1,2} then
4420 * f_b_g() will select a group from which a running task may be
4421 * pulled to this cpu in order to make the other package idle.
4422 * If there is no opportunity to make a package idle and if
4423 * there are no imbalance, then f_b_g() will return NULL and no
4424 * action will be taken in load_balance_newidle().
4426 * Under normal task pull operation due to imbalance, there
4427 * will be more than one task in the source run queue and
4428 * move_tasks() will succeed. ld_moved will be true and this
4429 * active balance code will not be triggered.
4432 /* Lock busiest in correct order while this_rq is held */
4433 double_lock_balance(this_rq
, busiest
);
4436 * don't kick the migration_thread, if the curr
4437 * task on busiest cpu can't be moved to this_cpu
4439 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
4440 double_unlock_balance(this_rq
, busiest
);
4445 if (!busiest
->active_balance
) {
4446 busiest
->active_balance
= 1;
4447 busiest
->push_cpu
= this_cpu
;
4451 double_unlock_balance(this_rq
, busiest
);
4453 * Should not call ttwu while holding a rq->lock
4455 raw_spin_unlock(&this_rq
->lock
);
4457 wake_up_process(busiest
->migration_thread
);
4458 raw_spin_lock(&this_rq
->lock
);
4461 sd
->nr_balance_failed
= 0;
4463 update_shares_locked(this_rq
, sd
);
4467 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
4468 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4469 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4471 sd
->nr_balance_failed
= 0;
4477 * idle_balance is called by schedule() if this_cpu is about to become
4478 * idle. Attempts to pull tasks from other CPUs.
4480 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
4482 struct sched_domain
*sd
;
4483 int pulled_task
= 0;
4484 unsigned long next_balance
= jiffies
+ HZ
;
4486 this_rq
->idle_stamp
= this_rq
->clock
;
4488 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
4491 for_each_domain(this_cpu
, sd
) {
4492 unsigned long interval
;
4494 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4497 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
4498 /* If we've pulled tasks over stop searching: */
4499 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
4502 interval
= msecs_to_jiffies(sd
->balance_interval
);
4503 if (time_after(next_balance
, sd
->last_balance
+ interval
))
4504 next_balance
= sd
->last_balance
+ interval
;
4506 this_rq
->idle_stamp
= 0;
4510 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
4512 * We are going idle. next_balance may be set based on
4513 * a busy processor. So reset next_balance.
4515 this_rq
->next_balance
= next_balance
;
4520 * active_load_balance is run by migration threads. It pushes running tasks
4521 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4522 * running on each physical CPU where possible, and avoids physical /
4523 * logical imbalances.
4525 * Called with busiest_rq locked.
4527 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
4529 int target_cpu
= busiest_rq
->push_cpu
;
4530 struct sched_domain
*sd
;
4531 struct rq
*target_rq
;
4533 /* Is there any task to move? */
4534 if (busiest_rq
->nr_running
<= 1)
4537 target_rq
= cpu_rq(target_cpu
);
4540 * This condition is "impossible", if it occurs
4541 * we need to fix it. Originally reported by
4542 * Bjorn Helgaas on a 128-cpu setup.
4544 BUG_ON(busiest_rq
== target_rq
);
4546 /* move a task from busiest_rq to target_rq */
4547 double_lock_balance(busiest_rq
, target_rq
);
4548 update_rq_clock(busiest_rq
);
4549 update_rq_clock(target_rq
);
4551 /* Search for an sd spanning us and the target CPU. */
4552 for_each_domain(target_cpu
, sd
) {
4553 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
4554 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
4559 schedstat_inc(sd
, alb_count
);
4561 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
4563 schedstat_inc(sd
, alb_pushed
);
4565 schedstat_inc(sd
, alb_failed
);
4567 double_unlock_balance(busiest_rq
, target_rq
);
4572 atomic_t load_balancer
;
4573 cpumask_var_t cpu_mask
;
4574 cpumask_var_t ilb_grp_nohz_mask
;
4575 } nohz ____cacheline_aligned
= {
4576 .load_balancer
= ATOMIC_INIT(-1),
4579 int get_nohz_load_balancer(void)
4581 return atomic_read(&nohz
.load_balancer
);
4584 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4586 * lowest_flag_domain - Return lowest sched_domain containing flag.
4587 * @cpu: The cpu whose lowest level of sched domain is to
4589 * @flag: The flag to check for the lowest sched_domain
4590 * for the given cpu.
4592 * Returns the lowest sched_domain of a cpu which contains the given flag.
4594 static inline struct sched_domain
*lowest_flag_domain(int cpu
, int flag
)
4596 struct sched_domain
*sd
;
4598 for_each_domain(cpu
, sd
)
4599 if (sd
&& (sd
->flags
& flag
))
4606 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4607 * @cpu: The cpu whose domains we're iterating over.
4608 * @sd: variable holding the value of the power_savings_sd
4610 * @flag: The flag to filter the sched_domains to be iterated.
4612 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4613 * set, starting from the lowest sched_domain to the highest.
4615 #define for_each_flag_domain(cpu, sd, flag) \
4616 for (sd = lowest_flag_domain(cpu, flag); \
4617 (sd && (sd->flags & flag)); sd = sd->parent)
4620 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4621 * @ilb_group: group to be checked for semi-idleness
4623 * Returns: 1 if the group is semi-idle. 0 otherwise.
4625 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4626 * and atleast one non-idle CPU. This helper function checks if the given
4627 * sched_group is semi-idle or not.
4629 static inline int is_semi_idle_group(struct sched_group
*ilb_group
)
4631 cpumask_and(nohz
.ilb_grp_nohz_mask
, nohz
.cpu_mask
,
4632 sched_group_cpus(ilb_group
));
4635 * A sched_group is semi-idle when it has atleast one busy cpu
4636 * and atleast one idle cpu.
4638 if (cpumask_empty(nohz
.ilb_grp_nohz_mask
))
4641 if (cpumask_equal(nohz
.ilb_grp_nohz_mask
, sched_group_cpus(ilb_group
)))
4647 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4648 * @cpu: The cpu which is nominating a new idle_load_balancer.
4650 * Returns: Returns the id of the idle load balancer if it exists,
4651 * Else, returns >= nr_cpu_ids.
4653 * This algorithm picks the idle load balancer such that it belongs to a
4654 * semi-idle powersavings sched_domain. The idea is to try and avoid
4655 * completely idle packages/cores just for the purpose of idle load balancing
4656 * when there are other idle cpu's which are better suited for that job.
4658 static int find_new_ilb(int cpu
)
4660 struct sched_domain
*sd
;
4661 struct sched_group
*ilb_group
;
4664 * Have idle load balancer selection from semi-idle packages only
4665 * when power-aware load balancing is enabled
4667 if (!(sched_smt_power_savings
|| sched_mc_power_savings
))
4671 * Optimize for the case when we have no idle CPUs or only one
4672 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4674 if (cpumask_weight(nohz
.cpu_mask
) < 2)
4677 for_each_flag_domain(cpu
, sd
, SD_POWERSAVINGS_BALANCE
) {
4678 ilb_group
= sd
->groups
;
4681 if (is_semi_idle_group(ilb_group
))
4682 return cpumask_first(nohz
.ilb_grp_nohz_mask
);
4684 ilb_group
= ilb_group
->next
;
4686 } while (ilb_group
!= sd
->groups
);
4690 return cpumask_first(nohz
.cpu_mask
);
4692 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4693 static inline int find_new_ilb(int call_cpu
)
4695 return cpumask_first(nohz
.cpu_mask
);
4700 * This routine will try to nominate the ilb (idle load balancing)
4701 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4702 * load balancing on behalf of all those cpus. If all the cpus in the system
4703 * go into this tickless mode, then there will be no ilb owner (as there is
4704 * no need for one) and all the cpus will sleep till the next wakeup event
4707 * For the ilb owner, tick is not stopped. And this tick will be used
4708 * for idle load balancing. ilb owner will still be part of
4711 * While stopping the tick, this cpu will become the ilb owner if there
4712 * is no other owner. And will be the owner till that cpu becomes busy
4713 * or if all cpus in the system stop their ticks at which point
4714 * there is no need for ilb owner.
4716 * When the ilb owner becomes busy, it nominates another owner, during the
4717 * next busy scheduler_tick()
4719 int select_nohz_load_balancer(int stop_tick
)
4721 int cpu
= smp_processor_id();
4724 cpu_rq(cpu
)->in_nohz_recently
= 1;
4726 if (!cpu_active(cpu
)) {
4727 if (atomic_read(&nohz
.load_balancer
) != cpu
)
4731 * If we are going offline and still the leader,
4734 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4740 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
4742 /* time for ilb owner also to sleep */
4743 if (cpumask_weight(nohz
.cpu_mask
) == num_active_cpus()) {
4744 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4745 atomic_set(&nohz
.load_balancer
, -1);
4749 if (atomic_read(&nohz
.load_balancer
) == -1) {
4750 /* make me the ilb owner */
4751 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
4753 } else if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4756 if (!(sched_smt_power_savings
||
4757 sched_mc_power_savings
))
4760 * Check to see if there is a more power-efficient
4763 new_ilb
= find_new_ilb(cpu
);
4764 if (new_ilb
< nr_cpu_ids
&& new_ilb
!= cpu
) {
4765 atomic_set(&nohz
.load_balancer
, -1);
4766 resched_cpu(new_ilb
);
4772 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4775 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4777 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4778 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4785 static DEFINE_SPINLOCK(balancing
);
4788 * It checks each scheduling domain to see if it is due to be balanced,
4789 * and initiates a balancing operation if so.
4791 * Balancing parameters are set up in arch_init_sched_domains.
4793 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
4796 struct rq
*rq
= cpu_rq(cpu
);
4797 unsigned long interval
;
4798 struct sched_domain
*sd
;
4799 /* Earliest time when we have to do rebalance again */
4800 unsigned long next_balance
= jiffies
+ 60*HZ
;
4801 int update_next_balance
= 0;
4804 for_each_domain(cpu
, sd
) {
4805 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4808 interval
= sd
->balance_interval
;
4809 if (idle
!= CPU_IDLE
)
4810 interval
*= sd
->busy_factor
;
4812 /* scale ms to jiffies */
4813 interval
= msecs_to_jiffies(interval
);
4814 if (unlikely(!interval
))
4816 if (interval
> HZ
*NR_CPUS
/10)
4817 interval
= HZ
*NR_CPUS
/10;
4819 need_serialize
= sd
->flags
& SD_SERIALIZE
;
4821 if (need_serialize
) {
4822 if (!spin_trylock(&balancing
))
4826 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4827 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
4829 * We've pulled tasks over so either we're no
4830 * longer idle, or one of our SMT siblings is
4833 idle
= CPU_NOT_IDLE
;
4835 sd
->last_balance
= jiffies
;
4838 spin_unlock(&balancing
);
4840 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4841 next_balance
= sd
->last_balance
+ interval
;
4842 update_next_balance
= 1;
4846 * Stop the load balance at this level. There is another
4847 * CPU in our sched group which is doing load balancing more
4855 * next_balance will be updated only when there is a need.
4856 * When the cpu is attached to null domain for ex, it will not be
4859 if (likely(update_next_balance
))
4860 rq
->next_balance
= next_balance
;
4864 * run_rebalance_domains is triggered when needed from the scheduler tick.
4865 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4866 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4868 static void run_rebalance_domains(struct softirq_action
*h
)
4870 int this_cpu
= smp_processor_id();
4871 struct rq
*this_rq
= cpu_rq(this_cpu
);
4872 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4873 CPU_IDLE
: CPU_NOT_IDLE
;
4875 rebalance_domains(this_cpu
, idle
);
4879 * If this cpu is the owner for idle load balancing, then do the
4880 * balancing on behalf of the other idle cpus whose ticks are
4883 if (this_rq
->idle_at_tick
&&
4884 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4888 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
4889 if (balance_cpu
== this_cpu
)
4893 * If this cpu gets work to do, stop the load balancing
4894 * work being done for other cpus. Next load
4895 * balancing owner will pick it up.
4900 rebalance_domains(balance_cpu
, CPU_IDLE
);
4902 rq
= cpu_rq(balance_cpu
);
4903 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4904 this_rq
->next_balance
= rq
->next_balance
;
4910 static inline int on_null_domain(int cpu
)
4912 return !rcu_dereference(cpu_rq(cpu
)->sd
);
4916 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4918 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4919 * idle load balancing owner or decide to stop the periodic load balancing,
4920 * if the whole system is idle.
4922 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4926 * If we were in the nohz mode recently and busy at the current
4927 * scheduler tick, then check if we need to nominate new idle
4930 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4931 rq
->in_nohz_recently
= 0;
4933 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4934 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4935 atomic_set(&nohz
.load_balancer
, -1);
4938 if (atomic_read(&nohz
.load_balancer
) == -1) {
4939 int ilb
= find_new_ilb(cpu
);
4941 if (ilb
< nr_cpu_ids
)
4947 * If this cpu is idle and doing idle load balancing for all the
4948 * cpus with ticks stopped, is it time for that to stop?
4950 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4951 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4957 * If this cpu is idle and the idle load balancing is done by
4958 * someone else, then no need raise the SCHED_SOFTIRQ
4960 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4961 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4964 /* Don't need to rebalance while attached to NULL domain */
4965 if (time_after_eq(jiffies
, rq
->next_balance
) &&
4966 likely(!on_null_domain(cpu
)))
4967 raise_softirq(SCHED_SOFTIRQ
);
4970 #else /* CONFIG_SMP */
4973 * on UP we do not need to balance between CPUs:
4975 static inline void idle_balance(int cpu
, struct rq
*rq
)
4981 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4983 EXPORT_PER_CPU_SYMBOL(kstat
);
4986 * Return any ns on the sched_clock that have not yet been accounted in
4987 * @p in case that task is currently running.
4989 * Called with task_rq_lock() held on @rq.
4991 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
4995 if (task_current(rq
, p
)) {
4996 update_rq_clock(rq
);
4997 ns
= rq
->clock
- p
->se
.exec_start
;
5005 unsigned long long task_delta_exec(struct task_struct
*p
)
5007 unsigned long flags
;
5011 rq
= task_rq_lock(p
, &flags
);
5012 ns
= do_task_delta_exec(p
, rq
);
5013 task_rq_unlock(rq
, &flags
);
5019 * Return accounted runtime for the task.
5020 * In case the task is currently running, return the runtime plus current's
5021 * pending runtime that have not been accounted yet.
5023 unsigned long long task_sched_runtime(struct task_struct
*p
)
5025 unsigned long flags
;
5029 rq
= task_rq_lock(p
, &flags
);
5030 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
5031 task_rq_unlock(rq
, &flags
);
5037 * Return sum_exec_runtime for the thread group.
5038 * In case the task is currently running, return the sum plus current's
5039 * pending runtime that have not been accounted yet.
5041 * Note that the thread group might have other running tasks as well,
5042 * so the return value not includes other pending runtime that other
5043 * running tasks might have.
5045 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
5047 struct task_cputime totals
;
5048 unsigned long flags
;
5052 rq
= task_rq_lock(p
, &flags
);
5053 thread_group_cputime(p
, &totals
);
5054 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
5055 task_rq_unlock(rq
, &flags
);
5061 * Account user cpu time to a process.
5062 * @p: the process that the cpu time gets accounted to
5063 * @cputime: the cpu time spent in user space since the last update
5064 * @cputime_scaled: cputime scaled by cpu frequency
5066 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
5067 cputime_t cputime_scaled
)
5069 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5072 /* Add user time to process. */
5073 p
->utime
= cputime_add(p
->utime
, cputime
);
5074 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
5075 account_group_user_time(p
, cputime
);
5077 /* Add user time to cpustat. */
5078 tmp
= cputime_to_cputime64(cputime
);
5079 if (TASK_NICE(p
) > 0)
5080 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
5082 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
5084 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
5085 /* Account for user time used */
5086 acct_update_integrals(p
);
5090 * Account guest cpu time to a process.
5091 * @p: the process that the cpu time gets accounted to
5092 * @cputime: the cpu time spent in virtual machine since the last update
5093 * @cputime_scaled: cputime scaled by cpu frequency
5095 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
5096 cputime_t cputime_scaled
)
5099 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5101 tmp
= cputime_to_cputime64(cputime
);
5103 /* Add guest time to process. */
5104 p
->utime
= cputime_add(p
->utime
, cputime
);
5105 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
5106 account_group_user_time(p
, cputime
);
5107 p
->gtime
= cputime_add(p
->gtime
, cputime
);
5109 /* Add guest time to cpustat. */
5110 if (TASK_NICE(p
) > 0) {
5111 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
5112 cpustat
->guest_nice
= cputime64_add(cpustat
->guest_nice
, tmp
);
5114 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
5115 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
5120 * Account system cpu time to a process.
5121 * @p: the process that the cpu time gets accounted to
5122 * @hardirq_offset: the offset to subtract from hardirq_count()
5123 * @cputime: the cpu time spent in kernel space since the last update
5124 * @cputime_scaled: cputime scaled by cpu frequency
5126 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
5127 cputime_t cputime
, cputime_t cputime_scaled
)
5129 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5132 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
5133 account_guest_time(p
, cputime
, cputime_scaled
);
5137 /* Add system time to process. */
5138 p
->stime
= cputime_add(p
->stime
, cputime
);
5139 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
5140 account_group_system_time(p
, cputime
);
5142 /* Add system time to cpustat. */
5143 tmp
= cputime_to_cputime64(cputime
);
5144 if (hardirq_count() - hardirq_offset
)
5145 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
5146 else if (softirq_count())
5147 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
5149 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
5151 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
5153 /* Account for system time used */
5154 acct_update_integrals(p
);
5158 * Account for involuntary wait time.
5159 * @steal: the cpu time spent in involuntary wait
5161 void account_steal_time(cputime_t cputime
)
5163 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5164 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
5166 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
5170 * Account for idle time.
5171 * @cputime: the cpu time spent in idle wait
5173 void account_idle_time(cputime_t cputime
)
5175 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5176 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
5177 struct rq
*rq
= this_rq();
5179 if (atomic_read(&rq
->nr_iowait
) > 0)
5180 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
5182 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
5185 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5188 * Account a single tick of cpu time.
5189 * @p: the process that the cpu time gets accounted to
5190 * @user_tick: indicates if the tick is a user or a system tick
5192 void account_process_tick(struct task_struct
*p
, int user_tick
)
5194 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
5195 struct rq
*rq
= this_rq();
5198 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
5199 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
5200 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
5203 account_idle_time(cputime_one_jiffy
);
5207 * Account multiple ticks of steal time.
5208 * @p: the process from which the cpu time has been stolen
5209 * @ticks: number of stolen ticks
5211 void account_steal_ticks(unsigned long ticks
)
5213 account_steal_time(jiffies_to_cputime(ticks
));
5217 * Account multiple ticks of idle time.
5218 * @ticks: number of stolen ticks
5220 void account_idle_ticks(unsigned long ticks
)
5222 account_idle_time(jiffies_to_cputime(ticks
));
5228 * Use precise platform statistics if available:
5230 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5231 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
5237 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
5239 struct task_cputime cputime
;
5241 thread_group_cputime(p
, &cputime
);
5243 *ut
= cputime
.utime
;
5244 *st
= cputime
.stime
;
5248 #ifndef nsecs_to_cputime
5249 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
5252 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
5254 cputime_t rtime
, utime
= p
->utime
, total
= cputime_add(utime
, p
->stime
);
5257 * Use CFS's precise accounting:
5259 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
5264 temp
= (u64
)(rtime
* utime
);
5265 do_div(temp
, total
);
5266 utime
= (cputime_t
)temp
;
5271 * Compare with previous values, to keep monotonicity:
5273 p
->prev_utime
= max(p
->prev_utime
, utime
);
5274 p
->prev_stime
= max(p
->prev_stime
, cputime_sub(rtime
, p
->prev_utime
));
5276 *ut
= p
->prev_utime
;
5277 *st
= p
->prev_stime
;
5281 * Must be called with siglock held.
5283 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
5285 struct signal_struct
*sig
= p
->signal
;
5286 struct task_cputime cputime
;
5287 cputime_t rtime
, utime
, total
;
5289 thread_group_cputime(p
, &cputime
);
5291 total
= cputime_add(cputime
.utime
, cputime
.stime
);
5292 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
5297 temp
= (u64
)(rtime
* cputime
.utime
);
5298 do_div(temp
, total
);
5299 utime
= (cputime_t
)temp
;
5303 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
5304 sig
->prev_stime
= max(sig
->prev_stime
,
5305 cputime_sub(rtime
, sig
->prev_utime
));
5307 *ut
= sig
->prev_utime
;
5308 *st
= sig
->prev_stime
;
5313 * This function gets called by the timer code, with HZ frequency.
5314 * We call it with interrupts disabled.
5316 * It also gets called by the fork code, when changing the parent's
5319 void scheduler_tick(void)
5321 int cpu
= smp_processor_id();
5322 struct rq
*rq
= cpu_rq(cpu
);
5323 struct task_struct
*curr
= rq
->curr
;
5327 raw_spin_lock(&rq
->lock
);
5328 update_rq_clock(rq
);
5329 update_cpu_load(rq
);
5330 curr
->sched_class
->task_tick(rq
, curr
, 0);
5331 raw_spin_unlock(&rq
->lock
);
5333 perf_event_task_tick(curr
, cpu
);
5336 rq
->idle_at_tick
= idle_cpu(cpu
);
5337 trigger_load_balance(rq
, cpu
);
5341 notrace
unsigned long get_parent_ip(unsigned long addr
)
5343 if (in_lock_functions(addr
)) {
5344 addr
= CALLER_ADDR2
;
5345 if (in_lock_functions(addr
))
5346 addr
= CALLER_ADDR3
;
5351 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5352 defined(CONFIG_PREEMPT_TRACER))
5354 void __kprobes
add_preempt_count(int val
)
5356 #ifdef CONFIG_DEBUG_PREEMPT
5360 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5363 preempt_count() += val
;
5364 #ifdef CONFIG_DEBUG_PREEMPT
5366 * Spinlock count overflowing soon?
5368 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
5371 if (preempt_count() == val
)
5372 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5374 EXPORT_SYMBOL(add_preempt_count
);
5376 void __kprobes
sub_preempt_count(int val
)
5378 #ifdef CONFIG_DEBUG_PREEMPT
5382 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
5385 * Is the spinlock portion underflowing?
5387 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
5388 !(preempt_count() & PREEMPT_MASK
)))
5392 if (preempt_count() == val
)
5393 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5394 preempt_count() -= val
;
5396 EXPORT_SYMBOL(sub_preempt_count
);
5401 * Print scheduling while atomic bug:
5403 static noinline
void __schedule_bug(struct task_struct
*prev
)
5405 struct pt_regs
*regs
= get_irq_regs();
5407 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
5408 prev
->comm
, prev
->pid
, preempt_count());
5410 debug_show_held_locks(prev
);
5412 if (irqs_disabled())
5413 print_irqtrace_events(prev
);
5422 * Various schedule()-time debugging checks and statistics:
5424 static inline void schedule_debug(struct task_struct
*prev
)
5427 * Test if we are atomic. Since do_exit() needs to call into
5428 * schedule() atomically, we ignore that path for now.
5429 * Otherwise, whine if we are scheduling when we should not be.
5431 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
5432 __schedule_bug(prev
);
5434 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
5436 schedstat_inc(this_rq(), sched_count
);
5437 #ifdef CONFIG_SCHEDSTATS
5438 if (unlikely(prev
->lock_depth
>= 0)) {
5439 schedstat_inc(this_rq(), bkl_count
);
5440 schedstat_inc(prev
, sched_info
.bkl_count
);
5445 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
5447 if (prev
->state
== TASK_RUNNING
) {
5448 u64 runtime
= prev
->se
.sum_exec_runtime
;
5450 runtime
-= prev
->se
.prev_sum_exec_runtime
;
5451 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
5454 * In order to avoid avg_overlap growing stale when we are
5455 * indeed overlapping and hence not getting put to sleep, grow
5456 * the avg_overlap on preemption.
5458 * We use the average preemption runtime because that
5459 * correlates to the amount of cache footprint a task can
5462 update_avg(&prev
->se
.avg_overlap
, runtime
);
5464 prev
->sched_class
->put_prev_task(rq
, prev
);
5468 * Pick up the highest-prio task:
5470 static inline struct task_struct
*
5471 pick_next_task(struct rq
*rq
)
5473 const struct sched_class
*class;
5474 struct task_struct
*p
;
5477 * Optimization: we know that if all tasks are in
5478 * the fair class we can call that function directly:
5480 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
5481 p
= fair_sched_class
.pick_next_task(rq
);
5486 class = sched_class_highest
;
5488 p
= class->pick_next_task(rq
);
5492 * Will never be NULL as the idle class always
5493 * returns a non-NULL p:
5495 class = class->next
;
5500 * schedule() is the main scheduler function.
5502 asmlinkage
void __sched
schedule(void)
5504 struct task_struct
*prev
, *next
;
5505 unsigned long *switch_count
;
5511 cpu
= smp_processor_id();
5515 switch_count
= &prev
->nivcsw
;
5517 release_kernel_lock(prev
);
5518 need_resched_nonpreemptible
:
5520 schedule_debug(prev
);
5522 if (sched_feat(HRTICK
))
5525 raw_spin_lock_irq(&rq
->lock
);
5526 update_rq_clock(rq
);
5527 clear_tsk_need_resched(prev
);
5529 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
5530 if (unlikely(signal_pending_state(prev
->state
, prev
)))
5531 prev
->state
= TASK_RUNNING
;
5533 deactivate_task(rq
, prev
, 1);
5534 switch_count
= &prev
->nvcsw
;
5537 pre_schedule(rq
, prev
);
5539 if (unlikely(!rq
->nr_running
))
5540 idle_balance(cpu
, rq
);
5542 put_prev_task(rq
, prev
);
5543 next
= pick_next_task(rq
);
5545 if (likely(prev
!= next
)) {
5546 sched_info_switch(prev
, next
);
5547 perf_event_task_sched_out(prev
, next
, cpu
);
5553 context_switch(rq
, prev
, next
); /* unlocks the rq */
5555 * the context switch might have flipped the stack from under
5556 * us, hence refresh the local variables.
5558 cpu
= smp_processor_id();
5561 raw_spin_unlock_irq(&rq
->lock
);
5565 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
5567 switch_count
= &prev
->nivcsw
;
5568 goto need_resched_nonpreemptible
;
5571 preempt_enable_no_resched();
5575 EXPORT_SYMBOL(schedule
);
5577 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
5579 * Look out! "owner" is an entirely speculative pointer
5580 * access and not reliable.
5582 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
5587 if (!sched_feat(OWNER_SPIN
))
5590 #ifdef CONFIG_DEBUG_PAGEALLOC
5592 * Need to access the cpu field knowing that
5593 * DEBUG_PAGEALLOC could have unmapped it if
5594 * the mutex owner just released it and exited.
5596 if (probe_kernel_address(&owner
->cpu
, cpu
))
5603 * Even if the access succeeded (likely case),
5604 * the cpu field may no longer be valid.
5606 if (cpu
>= nr_cpumask_bits
)
5610 * We need to validate that we can do a
5611 * get_cpu() and that we have the percpu area.
5613 if (!cpu_online(cpu
))
5620 * Owner changed, break to re-assess state.
5622 if (lock
->owner
!= owner
)
5626 * Is that owner really running on that cpu?
5628 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
5638 #ifdef CONFIG_PREEMPT
5640 * this is the entry point to schedule() from in-kernel preemption
5641 * off of preempt_enable. Kernel preemptions off return from interrupt
5642 * occur there and call schedule directly.
5644 asmlinkage
void __sched
preempt_schedule(void)
5646 struct thread_info
*ti
= current_thread_info();
5649 * If there is a non-zero preempt_count or interrupts are disabled,
5650 * we do not want to preempt the current task. Just return..
5652 if (likely(ti
->preempt_count
|| irqs_disabled()))
5656 add_preempt_count(PREEMPT_ACTIVE
);
5658 sub_preempt_count(PREEMPT_ACTIVE
);
5661 * Check again in case we missed a preemption opportunity
5662 * between schedule and now.
5665 } while (need_resched());
5667 EXPORT_SYMBOL(preempt_schedule
);
5670 * this is the entry point to schedule() from kernel preemption
5671 * off of irq context.
5672 * Note, that this is called and return with irqs disabled. This will
5673 * protect us against recursive calling from irq.
5675 asmlinkage
void __sched
preempt_schedule_irq(void)
5677 struct thread_info
*ti
= current_thread_info();
5679 /* Catch callers which need to be fixed */
5680 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
5683 add_preempt_count(PREEMPT_ACTIVE
);
5686 local_irq_disable();
5687 sub_preempt_count(PREEMPT_ACTIVE
);
5690 * Check again in case we missed a preemption opportunity
5691 * between schedule and now.
5694 } while (need_resched());
5697 #endif /* CONFIG_PREEMPT */
5699 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
5702 return try_to_wake_up(curr
->private, mode
, wake_flags
);
5704 EXPORT_SYMBOL(default_wake_function
);
5707 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5708 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5709 * number) then we wake all the non-exclusive tasks and one exclusive task.
5711 * There are circumstances in which we can try to wake a task which has already
5712 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5713 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5715 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
5716 int nr_exclusive
, int wake_flags
, void *key
)
5718 wait_queue_t
*curr
, *next
;
5720 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
5721 unsigned flags
= curr
->flags
;
5723 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
5724 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
5730 * __wake_up - wake up threads blocked on a waitqueue.
5732 * @mode: which threads
5733 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5734 * @key: is directly passed to the wakeup function
5736 * It may be assumed that this function implies a write memory barrier before
5737 * changing the task state if and only if any tasks are woken up.
5739 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
5740 int nr_exclusive
, void *key
)
5742 unsigned long flags
;
5744 spin_lock_irqsave(&q
->lock
, flags
);
5745 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
5746 spin_unlock_irqrestore(&q
->lock
, flags
);
5748 EXPORT_SYMBOL(__wake_up
);
5751 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5753 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
5755 __wake_up_common(q
, mode
, 1, 0, NULL
);
5758 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
5760 __wake_up_common(q
, mode
, 1, 0, key
);
5764 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5766 * @mode: which threads
5767 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5768 * @key: opaque value to be passed to wakeup targets
5770 * The sync wakeup differs that the waker knows that it will schedule
5771 * away soon, so while the target thread will be woken up, it will not
5772 * be migrated to another CPU - ie. the two threads are 'synchronized'
5773 * with each other. This can prevent needless bouncing between CPUs.
5775 * On UP it can prevent extra preemption.
5777 * It may be assumed that this function implies a write memory barrier before
5778 * changing the task state if and only if any tasks are woken up.
5780 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
5781 int nr_exclusive
, void *key
)
5783 unsigned long flags
;
5784 int wake_flags
= WF_SYNC
;
5789 if (unlikely(!nr_exclusive
))
5792 spin_lock_irqsave(&q
->lock
, flags
);
5793 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
5794 spin_unlock_irqrestore(&q
->lock
, flags
);
5796 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
5799 * __wake_up_sync - see __wake_up_sync_key()
5801 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
5803 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
5805 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
5808 * complete: - signals a single thread waiting on this completion
5809 * @x: holds the state of this particular completion
5811 * This will wake up a single thread waiting on this completion. Threads will be
5812 * awakened in the same order in which they were queued.
5814 * See also complete_all(), wait_for_completion() and related routines.
5816 * It may be assumed that this function implies a write memory barrier before
5817 * changing the task state if and only if any tasks are woken up.
5819 void complete(struct completion
*x
)
5821 unsigned long flags
;
5823 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5825 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
5826 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5828 EXPORT_SYMBOL(complete
);
5831 * complete_all: - signals all threads waiting on this completion
5832 * @x: holds the state of this particular completion
5834 * This will wake up all threads waiting on this particular completion event.
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_all(struct completion
*x
)
5841 unsigned long flags
;
5843 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5844 x
->done
+= UINT_MAX
/2;
5845 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
5846 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5848 EXPORT_SYMBOL(complete_all
);
5850 static inline long __sched
5851 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
5854 DECLARE_WAITQUEUE(wait
, current
);
5856 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
5857 __add_wait_queue_tail(&x
->wait
, &wait
);
5859 if (signal_pending_state(state
, current
)) {
5860 timeout
= -ERESTARTSYS
;
5863 __set_current_state(state
);
5864 spin_unlock_irq(&x
->wait
.lock
);
5865 timeout
= schedule_timeout(timeout
);
5866 spin_lock_irq(&x
->wait
.lock
);
5867 } while (!x
->done
&& timeout
);
5868 __remove_wait_queue(&x
->wait
, &wait
);
5873 return timeout
?: 1;
5877 wait_for_common(struct completion
*x
, long timeout
, int state
)
5881 spin_lock_irq(&x
->wait
.lock
);
5882 timeout
= do_wait_for_common(x
, timeout
, state
);
5883 spin_unlock_irq(&x
->wait
.lock
);
5888 * wait_for_completion: - waits for completion of a task
5889 * @x: holds the state of this particular completion
5891 * This waits to be signaled for completion of a specific task. It is NOT
5892 * interruptible and there is no timeout.
5894 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5895 * and interrupt capability. Also see complete().
5897 void __sched
wait_for_completion(struct completion
*x
)
5899 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
5901 EXPORT_SYMBOL(wait_for_completion
);
5904 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5905 * @x: holds the state of this particular completion
5906 * @timeout: timeout value in jiffies
5908 * This waits for either a completion of a specific task to be signaled or for a
5909 * specified timeout to expire. The timeout is in jiffies. It is not
5912 unsigned long __sched
5913 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
5915 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
5917 EXPORT_SYMBOL(wait_for_completion_timeout
);
5920 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5921 * @x: holds the state of this particular completion
5923 * This waits for completion of a specific task to be signaled. It is
5926 int __sched
wait_for_completion_interruptible(struct completion
*x
)
5928 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
5929 if (t
== -ERESTARTSYS
)
5933 EXPORT_SYMBOL(wait_for_completion_interruptible
);
5936 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5937 * @x: holds the state of this particular completion
5938 * @timeout: timeout value in jiffies
5940 * This waits for either a completion of a specific task to be signaled or for a
5941 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5943 unsigned long __sched
5944 wait_for_completion_interruptible_timeout(struct completion
*x
,
5945 unsigned long timeout
)
5947 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
5949 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
5952 * wait_for_completion_killable: - waits for completion of a task (killable)
5953 * @x: holds the state of this particular completion
5955 * This waits to be signaled for completion of a specific task. It can be
5956 * interrupted by a kill signal.
5958 int __sched
wait_for_completion_killable(struct completion
*x
)
5960 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
5961 if (t
== -ERESTARTSYS
)
5965 EXPORT_SYMBOL(wait_for_completion_killable
);
5968 * try_wait_for_completion - try to decrement a completion without blocking
5969 * @x: completion structure
5971 * Returns: 0 if a decrement cannot be done without blocking
5972 * 1 if a decrement succeeded.
5974 * If a completion is being used as a counting completion,
5975 * attempt to decrement the counter without blocking. This
5976 * enables us to avoid waiting if the resource the completion
5977 * is protecting is not available.
5979 bool try_wait_for_completion(struct completion
*x
)
5981 unsigned long flags
;
5984 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5989 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5992 EXPORT_SYMBOL(try_wait_for_completion
);
5995 * completion_done - Test to see if a completion has any waiters
5996 * @x: completion structure
5998 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5999 * 1 if there are no waiters.
6002 bool completion_done(struct completion
*x
)
6004 unsigned long flags
;
6007 spin_lock_irqsave(&x
->wait
.lock
, flags
);
6010 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
6013 EXPORT_SYMBOL(completion_done
);
6016 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
6018 unsigned long flags
;
6021 init_waitqueue_entry(&wait
, current
);
6023 __set_current_state(state
);
6025 spin_lock_irqsave(&q
->lock
, flags
);
6026 __add_wait_queue(q
, &wait
);
6027 spin_unlock(&q
->lock
);
6028 timeout
= schedule_timeout(timeout
);
6029 spin_lock_irq(&q
->lock
);
6030 __remove_wait_queue(q
, &wait
);
6031 spin_unlock_irqrestore(&q
->lock
, flags
);
6036 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
6038 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
6040 EXPORT_SYMBOL(interruptible_sleep_on
);
6043 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
6045 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
6047 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
6049 void __sched
sleep_on(wait_queue_head_t
*q
)
6051 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
6053 EXPORT_SYMBOL(sleep_on
);
6055 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
6057 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
6059 EXPORT_SYMBOL(sleep_on_timeout
);
6061 #ifdef CONFIG_RT_MUTEXES
6064 * rt_mutex_setprio - set the current priority of a task
6066 * @prio: prio value (kernel-internal form)
6068 * This function changes the 'effective' priority of a task. It does
6069 * not touch ->normal_prio like __setscheduler().
6071 * Used by the rt_mutex code to implement priority inheritance logic.
6073 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
6075 unsigned long flags
;
6076 int oldprio
, on_rq
, running
;
6078 const struct sched_class
*prev_class
;
6080 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
6082 rq
= task_rq_lock(p
, &flags
);
6083 update_rq_clock(rq
);
6086 prev_class
= p
->sched_class
;
6087 on_rq
= p
->se
.on_rq
;
6088 running
= task_current(rq
, p
);
6090 dequeue_task(rq
, p
, 0);
6092 p
->sched_class
->put_prev_task(rq
, p
);
6095 p
->sched_class
= &rt_sched_class
;
6097 p
->sched_class
= &fair_sched_class
;
6102 p
->sched_class
->set_curr_task(rq
);
6104 enqueue_task(rq
, p
, 0);
6106 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
6108 task_rq_unlock(rq
, &flags
);
6113 void set_user_nice(struct task_struct
*p
, long nice
)
6115 int old_prio
, delta
, on_rq
;
6116 unsigned long flags
;
6119 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
6122 * We have to be careful, if called from sys_setpriority(),
6123 * the task might be in the middle of scheduling on another CPU.
6125 rq
= task_rq_lock(p
, &flags
);
6126 update_rq_clock(rq
);
6128 * The RT priorities are set via sched_setscheduler(), but we still
6129 * allow the 'normal' nice value to be set - but as expected
6130 * it wont have any effect on scheduling until the task is
6131 * SCHED_FIFO/SCHED_RR:
6133 if (task_has_rt_policy(p
)) {
6134 p
->static_prio
= NICE_TO_PRIO(nice
);
6137 on_rq
= p
->se
.on_rq
;
6139 dequeue_task(rq
, p
, 0);
6141 p
->static_prio
= NICE_TO_PRIO(nice
);
6144 p
->prio
= effective_prio(p
);
6145 delta
= p
->prio
- old_prio
;
6148 enqueue_task(rq
, p
, 0);
6150 * If the task increased its priority or is running and
6151 * lowered its priority, then reschedule its CPU:
6153 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
6154 resched_task(rq
->curr
);
6157 task_rq_unlock(rq
, &flags
);
6159 EXPORT_SYMBOL(set_user_nice
);
6162 * can_nice - check if a task can reduce its nice value
6166 int can_nice(const struct task_struct
*p
, const int nice
)
6168 /* convert nice value [19,-20] to rlimit style value [1,40] */
6169 int nice_rlim
= 20 - nice
;
6171 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
6172 capable(CAP_SYS_NICE
));
6175 #ifdef __ARCH_WANT_SYS_NICE
6178 * sys_nice - change the priority of the current process.
6179 * @increment: priority increment
6181 * sys_setpriority is a more generic, but much slower function that
6182 * does similar things.
6184 SYSCALL_DEFINE1(nice
, int, increment
)
6189 * Setpriority might change our priority at the same moment.
6190 * We don't have to worry. Conceptually one call occurs first
6191 * and we have a single winner.
6193 if (increment
< -40)
6198 nice
= TASK_NICE(current
) + increment
;
6204 if (increment
< 0 && !can_nice(current
, nice
))
6207 retval
= security_task_setnice(current
, nice
);
6211 set_user_nice(current
, nice
);
6218 * task_prio - return the priority value of a given task.
6219 * @p: the task in question.
6221 * This is the priority value as seen by users in /proc.
6222 * RT tasks are offset by -200. Normal tasks are centered
6223 * around 0, value goes from -16 to +15.
6225 int task_prio(const struct task_struct
*p
)
6227 return p
->prio
- MAX_RT_PRIO
;
6231 * task_nice - return the nice value of a given task.
6232 * @p: the task in question.
6234 int task_nice(const struct task_struct
*p
)
6236 return TASK_NICE(p
);
6238 EXPORT_SYMBOL(task_nice
);
6241 * idle_cpu - is a given cpu idle currently?
6242 * @cpu: the processor in question.
6244 int idle_cpu(int cpu
)
6246 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
6250 * idle_task - return the idle task for a given cpu.
6251 * @cpu: the processor in question.
6253 struct task_struct
*idle_task(int cpu
)
6255 return cpu_rq(cpu
)->idle
;
6259 * find_process_by_pid - find a process with a matching PID value.
6260 * @pid: the pid in question.
6262 static struct task_struct
*find_process_by_pid(pid_t pid
)
6264 return pid
? find_task_by_vpid(pid
) : current
;
6267 /* Actually do priority change: must hold rq lock. */
6269 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
6271 BUG_ON(p
->se
.on_rq
);
6274 p
->rt_priority
= prio
;
6275 p
->normal_prio
= normal_prio(p
);
6276 /* we are holding p->pi_lock already */
6277 p
->prio
= rt_mutex_getprio(p
);
6278 if (rt_prio(p
->prio
))
6279 p
->sched_class
= &rt_sched_class
;
6281 p
->sched_class
= &fair_sched_class
;
6286 * check the target process has a UID that matches the current process's
6288 static bool check_same_owner(struct task_struct
*p
)
6290 const struct cred
*cred
= current_cred(), *pcred
;
6294 pcred
= __task_cred(p
);
6295 match
= (cred
->euid
== pcred
->euid
||
6296 cred
->euid
== pcred
->uid
);
6301 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
6302 struct sched_param
*param
, bool user
)
6304 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
6305 unsigned long flags
;
6306 const struct sched_class
*prev_class
;
6310 /* may grab non-irq protected spin_locks */
6311 BUG_ON(in_interrupt());
6313 /* double check policy once rq lock held */
6315 reset_on_fork
= p
->sched_reset_on_fork
;
6316 policy
= oldpolicy
= p
->policy
;
6318 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
6319 policy
&= ~SCHED_RESET_ON_FORK
;
6321 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
6322 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
6323 policy
!= SCHED_IDLE
)
6328 * Valid priorities for SCHED_FIFO and SCHED_RR are
6329 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6330 * SCHED_BATCH and SCHED_IDLE is 0.
6332 if (param
->sched_priority
< 0 ||
6333 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
6334 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
6336 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
6340 * Allow unprivileged RT tasks to decrease priority:
6342 if (user
&& !capable(CAP_SYS_NICE
)) {
6343 if (rt_policy(policy
)) {
6344 unsigned long rlim_rtprio
;
6346 if (!lock_task_sighand(p
, &flags
))
6348 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
6349 unlock_task_sighand(p
, &flags
);
6351 /* can't set/change the rt policy */
6352 if (policy
!= p
->policy
&& !rlim_rtprio
)
6355 /* can't increase priority */
6356 if (param
->sched_priority
> p
->rt_priority
&&
6357 param
->sched_priority
> rlim_rtprio
)
6361 * Like positive nice levels, dont allow tasks to
6362 * move out of SCHED_IDLE either:
6364 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
6367 /* can't change other user's priorities */
6368 if (!check_same_owner(p
))
6371 /* Normal users shall not reset the sched_reset_on_fork flag */
6372 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
6377 #ifdef CONFIG_RT_GROUP_SCHED
6379 * Do not allow realtime tasks into groups that have no runtime
6382 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
6383 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
6387 retval
= security_task_setscheduler(p
, policy
, param
);
6393 * make sure no PI-waiters arrive (or leave) while we are
6394 * changing the priority of the task:
6396 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
6398 * To be able to change p->policy safely, the apropriate
6399 * runqueue lock must be held.
6401 rq
= __task_rq_lock(p
);
6402 /* recheck policy now with rq lock held */
6403 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
6404 policy
= oldpolicy
= -1;
6405 __task_rq_unlock(rq
);
6406 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6409 update_rq_clock(rq
);
6410 on_rq
= p
->se
.on_rq
;
6411 running
= task_current(rq
, p
);
6413 deactivate_task(rq
, p
, 0);
6415 p
->sched_class
->put_prev_task(rq
, p
);
6417 p
->sched_reset_on_fork
= reset_on_fork
;
6420 prev_class
= p
->sched_class
;
6421 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
6424 p
->sched_class
->set_curr_task(rq
);
6426 activate_task(rq
, p
, 0);
6428 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
6430 __task_rq_unlock(rq
);
6431 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6433 rt_mutex_adjust_pi(p
);
6439 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6440 * @p: the task in question.
6441 * @policy: new policy.
6442 * @param: structure containing the new RT priority.
6444 * NOTE that the task may be already dead.
6446 int sched_setscheduler(struct task_struct
*p
, int policy
,
6447 struct sched_param
*param
)
6449 return __sched_setscheduler(p
, policy
, param
, true);
6451 EXPORT_SYMBOL_GPL(sched_setscheduler
);
6454 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6455 * @p: the task in question.
6456 * @policy: new policy.
6457 * @param: structure containing the new RT priority.
6459 * Just like sched_setscheduler, only don't bother checking if the
6460 * current context has permission. For example, this is needed in
6461 * stop_machine(): we create temporary high priority worker threads,
6462 * but our caller might not have that capability.
6464 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
6465 struct sched_param
*param
)
6467 return __sched_setscheduler(p
, policy
, param
, false);
6471 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
6473 struct sched_param lparam
;
6474 struct task_struct
*p
;
6477 if (!param
|| pid
< 0)
6479 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
6484 p
= find_process_by_pid(pid
);
6486 retval
= sched_setscheduler(p
, policy
, &lparam
);
6493 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6494 * @pid: the pid in question.
6495 * @policy: new policy.
6496 * @param: structure containing the new RT priority.
6498 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
6499 struct sched_param __user
*, param
)
6501 /* negative values for policy are not valid */
6505 return do_sched_setscheduler(pid
, policy
, param
);
6509 * sys_sched_setparam - set/change the RT priority of a thread
6510 * @pid: the pid in question.
6511 * @param: structure containing the new RT priority.
6513 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6515 return do_sched_setscheduler(pid
, -1, param
);
6519 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6520 * @pid: the pid in question.
6522 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
6524 struct task_struct
*p
;
6532 p
= find_process_by_pid(pid
);
6534 retval
= security_task_getscheduler(p
);
6537 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
6544 * sys_sched_getparam - get the RT priority of a thread
6545 * @pid: the pid in question.
6546 * @param: structure containing the RT priority.
6548 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6550 struct sched_param lp
;
6551 struct task_struct
*p
;
6554 if (!param
|| pid
< 0)
6558 p
= find_process_by_pid(pid
);
6563 retval
= security_task_getscheduler(p
);
6567 lp
.sched_priority
= p
->rt_priority
;
6571 * This one might sleep, we cannot do it with a spinlock held ...
6573 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
6582 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
6584 cpumask_var_t cpus_allowed
, new_mask
;
6585 struct task_struct
*p
;
6591 p
= find_process_by_pid(pid
);
6598 /* Prevent p going away */
6602 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
6606 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
6608 goto out_free_cpus_allowed
;
6611 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
6614 retval
= security_task_setscheduler(p
, 0, NULL
);
6618 cpuset_cpus_allowed(p
, cpus_allowed
);
6619 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
6621 retval
= set_cpus_allowed_ptr(p
, new_mask
);
6624 cpuset_cpus_allowed(p
, cpus_allowed
);
6625 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
6627 * We must have raced with a concurrent cpuset
6628 * update. Just reset the cpus_allowed to the
6629 * cpuset's cpus_allowed
6631 cpumask_copy(new_mask
, cpus_allowed
);
6636 free_cpumask_var(new_mask
);
6637 out_free_cpus_allowed
:
6638 free_cpumask_var(cpus_allowed
);
6645 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
6646 struct cpumask
*new_mask
)
6648 if (len
< cpumask_size())
6649 cpumask_clear(new_mask
);
6650 else if (len
> cpumask_size())
6651 len
= cpumask_size();
6653 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
6657 * sys_sched_setaffinity - set the cpu affinity of a process
6658 * @pid: pid of the process
6659 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6660 * @user_mask_ptr: user-space pointer to the new cpu mask
6662 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
6663 unsigned long __user
*, user_mask_ptr
)
6665 cpumask_var_t new_mask
;
6668 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
6671 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
6673 retval
= sched_setaffinity(pid
, new_mask
);
6674 free_cpumask_var(new_mask
);
6678 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
6680 struct task_struct
*p
;
6681 unsigned long flags
;
6689 p
= find_process_by_pid(pid
);
6693 retval
= security_task_getscheduler(p
);
6697 rq
= task_rq_lock(p
, &flags
);
6698 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
6699 task_rq_unlock(rq
, &flags
);
6709 * sys_sched_getaffinity - get the cpu affinity of a process
6710 * @pid: pid of the process
6711 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6712 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6714 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
6715 unsigned long __user
*, user_mask_ptr
)
6720 if (len
< cpumask_size())
6723 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
6726 ret
= sched_getaffinity(pid
, mask
);
6728 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
6731 ret
= cpumask_size();
6733 free_cpumask_var(mask
);
6739 * sys_sched_yield - yield the current processor to other threads.
6741 * This function yields the current CPU to other tasks. If there are no
6742 * other threads running on this CPU then this function will return.
6744 SYSCALL_DEFINE0(sched_yield
)
6746 struct rq
*rq
= this_rq_lock();
6748 schedstat_inc(rq
, yld_count
);
6749 current
->sched_class
->yield_task(rq
);
6752 * Since we are going to call schedule() anyway, there's
6753 * no need to preempt or enable interrupts:
6755 __release(rq
->lock
);
6756 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
6757 do_raw_spin_unlock(&rq
->lock
);
6758 preempt_enable_no_resched();
6765 static inline int should_resched(void)
6767 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
6770 static void __cond_resched(void)
6772 add_preempt_count(PREEMPT_ACTIVE
);
6774 sub_preempt_count(PREEMPT_ACTIVE
);
6777 int __sched
_cond_resched(void)
6779 if (should_resched()) {
6785 EXPORT_SYMBOL(_cond_resched
);
6788 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6789 * call schedule, and on return reacquire the lock.
6791 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6792 * operations here to prevent schedule() from being called twice (once via
6793 * spin_unlock(), once by hand).
6795 int __cond_resched_lock(spinlock_t
*lock
)
6797 int resched
= should_resched();
6800 lockdep_assert_held(lock
);
6802 if (spin_needbreak(lock
) || resched
) {
6813 EXPORT_SYMBOL(__cond_resched_lock
);
6815 int __sched
__cond_resched_softirq(void)
6817 BUG_ON(!in_softirq());
6819 if (should_resched()) {
6827 EXPORT_SYMBOL(__cond_resched_softirq
);
6830 * yield - yield the current processor to other threads.
6832 * This is a shortcut for kernel-space yielding - it marks the
6833 * thread runnable and calls sys_sched_yield().
6835 void __sched
yield(void)
6837 set_current_state(TASK_RUNNING
);
6840 EXPORT_SYMBOL(yield
);
6843 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6844 * that process accounting knows that this is a task in IO wait state.
6846 void __sched
io_schedule(void)
6848 struct rq
*rq
= raw_rq();
6850 delayacct_blkio_start();
6851 atomic_inc(&rq
->nr_iowait
);
6852 current
->in_iowait
= 1;
6854 current
->in_iowait
= 0;
6855 atomic_dec(&rq
->nr_iowait
);
6856 delayacct_blkio_end();
6858 EXPORT_SYMBOL(io_schedule
);
6860 long __sched
io_schedule_timeout(long timeout
)
6862 struct rq
*rq
= raw_rq();
6865 delayacct_blkio_start();
6866 atomic_inc(&rq
->nr_iowait
);
6867 current
->in_iowait
= 1;
6868 ret
= schedule_timeout(timeout
);
6869 current
->in_iowait
= 0;
6870 atomic_dec(&rq
->nr_iowait
);
6871 delayacct_blkio_end();
6876 * sys_sched_get_priority_max - return maximum RT priority.
6877 * @policy: scheduling class.
6879 * this syscall returns the maximum rt_priority that can be used
6880 * by a given scheduling class.
6882 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
6889 ret
= MAX_USER_RT_PRIO
-1;
6901 * sys_sched_get_priority_min - return minimum RT priority.
6902 * @policy: scheduling class.
6904 * this syscall returns the minimum rt_priority that can be used
6905 * by a given scheduling class.
6907 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
6925 * sys_sched_rr_get_interval - return the default timeslice of a process.
6926 * @pid: pid of the process.
6927 * @interval: userspace pointer to the timeslice value.
6929 * this syscall writes the default timeslice value of a given process
6930 * into the user-space timespec buffer. A value of '0' means infinity.
6932 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
6933 struct timespec __user
*, interval
)
6935 struct task_struct
*p
;
6936 unsigned int time_slice
;
6937 unsigned long flags
;
6947 p
= find_process_by_pid(pid
);
6951 retval
= security_task_getscheduler(p
);
6955 rq
= task_rq_lock(p
, &flags
);
6956 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
6957 task_rq_unlock(rq
, &flags
);
6960 jiffies_to_timespec(time_slice
, &t
);
6961 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
6969 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
6971 void sched_show_task(struct task_struct
*p
)
6973 unsigned long free
= 0;
6976 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
6977 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
6978 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
6979 #if BITS_PER_LONG == 32
6980 if (state
== TASK_RUNNING
)
6981 printk(KERN_CONT
" running ");
6983 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
6985 if (state
== TASK_RUNNING
)
6986 printk(KERN_CONT
" running task ");
6988 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
6990 #ifdef CONFIG_DEBUG_STACK_USAGE
6991 free
= stack_not_used(p
);
6993 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
6994 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
6995 (unsigned long)task_thread_info(p
)->flags
);
6997 show_stack(p
, NULL
);
7000 void show_state_filter(unsigned long state_filter
)
7002 struct task_struct
*g
, *p
;
7004 #if BITS_PER_LONG == 32
7006 " task PC stack pid father\n");
7009 " task PC stack pid father\n");
7011 read_lock(&tasklist_lock
);
7012 do_each_thread(g
, p
) {
7014 * reset the NMI-timeout, listing all files on a slow
7015 * console might take alot of time:
7017 touch_nmi_watchdog();
7018 if (!state_filter
|| (p
->state
& state_filter
))
7020 } while_each_thread(g
, p
);
7022 touch_all_softlockup_watchdogs();
7024 #ifdef CONFIG_SCHED_DEBUG
7025 sysrq_sched_debug_show();
7027 read_unlock(&tasklist_lock
);
7029 * Only show locks if all tasks are dumped:
7032 debug_show_all_locks();
7035 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
7037 idle
->sched_class
= &idle_sched_class
;
7041 * init_idle - set up an idle thread for a given CPU
7042 * @idle: task in question
7043 * @cpu: cpu the idle task belongs to
7045 * NOTE: this function does not set the idle thread's NEED_RESCHED
7046 * flag, to make booting more robust.
7048 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
7050 struct rq
*rq
= cpu_rq(cpu
);
7051 unsigned long flags
;
7053 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7056 idle
->state
= TASK_RUNNING
;
7057 idle
->se
.exec_start
= sched_clock();
7059 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
7060 __set_task_cpu(idle
, cpu
);
7062 rq
->curr
= rq
->idle
= idle
;
7063 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7066 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7068 /* Set the preempt count _outside_ the spinlocks! */
7069 #if defined(CONFIG_PREEMPT)
7070 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
7072 task_thread_info(idle
)->preempt_count
= 0;
7075 * The idle tasks have their own, simple scheduling class:
7077 idle
->sched_class
= &idle_sched_class
;
7078 ftrace_graph_init_task(idle
);
7082 * In a system that switches off the HZ timer nohz_cpu_mask
7083 * indicates which cpus entered this state. This is used
7084 * in the rcu update to wait only for active cpus. For system
7085 * which do not switch off the HZ timer nohz_cpu_mask should
7086 * always be CPU_BITS_NONE.
7088 cpumask_var_t nohz_cpu_mask
;
7091 * Increase the granularity value when there are more CPUs,
7092 * because with more CPUs the 'effective latency' as visible
7093 * to users decreases. But the relationship is not linear,
7094 * so pick a second-best guess by going with the log2 of the
7097 * This idea comes from the SD scheduler of Con Kolivas:
7099 static int get_update_sysctl_factor(void)
7101 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
7102 unsigned int factor
;
7104 switch (sysctl_sched_tunable_scaling
) {
7105 case SCHED_TUNABLESCALING_NONE
:
7108 case SCHED_TUNABLESCALING_LINEAR
:
7111 case SCHED_TUNABLESCALING_LOG
:
7113 factor
= 1 + ilog2(cpus
);
7120 static void update_sysctl(void)
7122 unsigned int factor
= get_update_sysctl_factor();
7124 #define SET_SYSCTL(name) \
7125 (sysctl_##name = (factor) * normalized_sysctl_##name)
7126 SET_SYSCTL(sched_min_granularity
);
7127 SET_SYSCTL(sched_latency
);
7128 SET_SYSCTL(sched_wakeup_granularity
);
7129 SET_SYSCTL(sched_shares_ratelimit
);
7133 static inline void sched_init_granularity(void)
7140 * This is how migration works:
7142 * 1) we queue a struct migration_req structure in the source CPU's
7143 * runqueue and wake up that CPU's migration thread.
7144 * 2) we down() the locked semaphore => thread blocks.
7145 * 3) migration thread wakes up (implicitly it forces the migrated
7146 * thread off the CPU)
7147 * 4) it gets the migration request and checks whether the migrated
7148 * task is still in the wrong runqueue.
7149 * 5) if it's in the wrong runqueue then the migration thread removes
7150 * it and puts it into the right queue.
7151 * 6) migration thread up()s the semaphore.
7152 * 7) we wake up and the migration is done.
7156 * Change a given task's CPU affinity. Migrate the thread to a
7157 * proper CPU and schedule it away if the CPU it's executing on
7158 * is removed from the allowed bitmask.
7160 * NOTE: the caller must have a valid reference to the task, the
7161 * task must not exit() & deallocate itself prematurely. The
7162 * call is not atomic; no spinlocks may be held.
7164 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
7166 struct migration_req req
;
7167 unsigned long flags
;
7172 * Since we rely on wake-ups to migrate sleeping tasks, don't change
7173 * the ->cpus_allowed mask from under waking tasks, which would be
7174 * possible when we change rq->lock in ttwu(), so synchronize against
7175 * TASK_WAKING to avoid that.
7177 * Make an exception for freshly cloned tasks, since cpuset namespaces
7178 * might move the task about, we have to validate the target in
7179 * wake_up_new_task() anyway since the cpu might have gone away.
7182 while (p
->state
== TASK_WAKING
&& !(p
->flags
& PF_STARTING
))
7185 rq
= task_rq_lock(p
, &flags
);
7187 if (p
->state
== TASK_WAKING
&& !(p
->flags
& PF_STARTING
)) {
7188 task_rq_unlock(rq
, &flags
);
7192 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
7197 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
7198 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
7203 if (p
->sched_class
->set_cpus_allowed
)
7204 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
7206 cpumask_copy(&p
->cpus_allowed
, new_mask
);
7207 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
7210 /* Can the task run on the task's current CPU? If so, we're done */
7211 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
7214 if (migrate_task(p
, cpumask_any_and(cpu_active_mask
, new_mask
), &req
)) {
7215 /* Need help from migration thread: drop lock and wait. */
7216 struct task_struct
*mt
= rq
->migration_thread
;
7218 get_task_struct(mt
);
7219 task_rq_unlock(rq
, &flags
);
7220 wake_up_process(rq
->migration_thread
);
7221 put_task_struct(mt
);
7222 wait_for_completion(&req
.done
);
7223 tlb_migrate_finish(p
->mm
);
7227 task_rq_unlock(rq
, &flags
);
7231 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
7234 * Move (not current) task off this cpu, onto dest cpu. We're doing
7235 * this because either it can't run here any more (set_cpus_allowed()
7236 * away from this CPU, or CPU going down), or because we're
7237 * attempting to rebalance this task on exec (sched_exec).
7239 * So we race with normal scheduler movements, but that's OK, as long
7240 * as the task is no longer on this CPU.
7242 * Returns non-zero if task was successfully migrated.
7244 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7246 struct rq
*rq_dest
, *rq_src
;
7249 if (unlikely(!cpu_active(dest_cpu
)))
7252 rq_src
= cpu_rq(src_cpu
);
7253 rq_dest
= cpu_rq(dest_cpu
);
7255 double_rq_lock(rq_src
, rq_dest
);
7256 /* Already moved. */
7257 if (task_cpu(p
) != src_cpu
)
7259 /* Affinity changed (again). */
7260 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7264 * If we're not on a rq, the next wake-up will ensure we're
7268 deactivate_task(rq_src
, p
, 0);
7269 set_task_cpu(p
, dest_cpu
);
7270 activate_task(rq_dest
, p
, 0);
7271 check_preempt_curr(rq_dest
, p
, 0);
7276 double_rq_unlock(rq_src
, rq_dest
);
7280 #define RCU_MIGRATION_IDLE 0
7281 #define RCU_MIGRATION_NEED_QS 1
7282 #define RCU_MIGRATION_GOT_QS 2
7283 #define RCU_MIGRATION_MUST_SYNC 3
7286 * migration_thread - this is a highprio system thread that performs
7287 * thread migration by bumping thread off CPU then 'pushing' onto
7290 static int migration_thread(void *data
)
7293 int cpu
= (long)data
;
7297 BUG_ON(rq
->migration_thread
!= current
);
7299 set_current_state(TASK_INTERRUPTIBLE
);
7300 while (!kthread_should_stop()) {
7301 struct migration_req
*req
;
7302 struct list_head
*head
;
7304 raw_spin_lock_irq(&rq
->lock
);
7306 if (cpu_is_offline(cpu
)) {
7307 raw_spin_unlock_irq(&rq
->lock
);
7311 if (rq
->active_balance
) {
7312 active_load_balance(rq
, cpu
);
7313 rq
->active_balance
= 0;
7316 head
= &rq
->migration_queue
;
7318 if (list_empty(head
)) {
7319 raw_spin_unlock_irq(&rq
->lock
);
7321 set_current_state(TASK_INTERRUPTIBLE
);
7324 req
= list_entry(head
->next
, struct migration_req
, list
);
7325 list_del_init(head
->next
);
7327 if (req
->task
!= NULL
) {
7328 raw_spin_unlock(&rq
->lock
);
7329 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
7330 } else if (likely(cpu
== (badcpu
= smp_processor_id()))) {
7331 req
->dest_cpu
= RCU_MIGRATION_GOT_QS
;
7332 raw_spin_unlock(&rq
->lock
);
7334 req
->dest_cpu
= RCU_MIGRATION_MUST_SYNC
;
7335 raw_spin_unlock(&rq
->lock
);
7336 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu
, cpu
);
7340 complete(&req
->done
);
7342 __set_current_state(TASK_RUNNING
);
7347 #ifdef CONFIG_HOTPLUG_CPU
7349 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7353 local_irq_disable();
7354 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
7360 * Figure out where task on dead CPU should go, use force if necessary.
7362 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
7367 dest_cpu
= select_fallback_rq(dead_cpu
, p
);
7369 /* It can have affinity changed while we were choosing. */
7370 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
7375 * While a dead CPU has no uninterruptible tasks queued at this point,
7376 * it might still have a nonzero ->nr_uninterruptible counter, because
7377 * for performance reasons the counter is not stricly tracking tasks to
7378 * their home CPUs. So we just add the counter to another CPU's counter,
7379 * to keep the global sum constant after CPU-down:
7381 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
7383 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
7384 unsigned long flags
;
7386 local_irq_save(flags
);
7387 double_rq_lock(rq_src
, rq_dest
);
7388 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
7389 rq_src
->nr_uninterruptible
= 0;
7390 double_rq_unlock(rq_src
, rq_dest
);
7391 local_irq_restore(flags
);
7394 /* Run through task list and migrate tasks from the dead cpu. */
7395 static void migrate_live_tasks(int src_cpu
)
7397 struct task_struct
*p
, *t
;
7399 read_lock(&tasklist_lock
);
7401 do_each_thread(t
, p
) {
7405 if (task_cpu(p
) == src_cpu
)
7406 move_task_off_dead_cpu(src_cpu
, p
);
7407 } while_each_thread(t
, p
);
7409 read_unlock(&tasklist_lock
);
7413 * Schedules idle task to be the next runnable task on current CPU.
7414 * It does so by boosting its priority to highest possible.
7415 * Used by CPU offline code.
7417 void sched_idle_next(void)
7419 int this_cpu
= smp_processor_id();
7420 struct rq
*rq
= cpu_rq(this_cpu
);
7421 struct task_struct
*p
= rq
->idle
;
7422 unsigned long flags
;
7424 /* cpu has to be offline */
7425 BUG_ON(cpu_online(this_cpu
));
7428 * Strictly not necessary since rest of the CPUs are stopped by now
7429 * and interrupts disabled on the current cpu.
7431 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7433 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7435 update_rq_clock(rq
);
7436 activate_task(rq
, p
, 0);
7438 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7442 * Ensures that the idle task is using init_mm right before its cpu goes
7445 void idle_task_exit(void)
7447 struct mm_struct
*mm
= current
->active_mm
;
7449 BUG_ON(cpu_online(smp_processor_id()));
7452 switch_mm(mm
, &init_mm
, current
);
7456 /* called under rq->lock with disabled interrupts */
7457 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
7459 struct rq
*rq
= cpu_rq(dead_cpu
);
7461 /* Must be exiting, otherwise would be on tasklist. */
7462 BUG_ON(!p
->exit_state
);
7464 /* Cannot have done final schedule yet: would have vanished. */
7465 BUG_ON(p
->state
== TASK_DEAD
);
7470 * Drop lock around migration; if someone else moves it,
7471 * that's OK. No task can be added to this CPU, so iteration is
7474 raw_spin_unlock_irq(&rq
->lock
);
7475 move_task_off_dead_cpu(dead_cpu
, p
);
7476 raw_spin_lock_irq(&rq
->lock
);
7481 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7482 static void migrate_dead_tasks(unsigned int dead_cpu
)
7484 struct rq
*rq
= cpu_rq(dead_cpu
);
7485 struct task_struct
*next
;
7488 if (!rq
->nr_running
)
7490 update_rq_clock(rq
);
7491 next
= pick_next_task(rq
);
7494 next
->sched_class
->put_prev_task(rq
, next
);
7495 migrate_dead(dead_cpu
, next
);
7501 * remove the tasks which were accounted by rq from calc_load_tasks.
7503 static void calc_global_load_remove(struct rq
*rq
)
7505 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
7506 rq
->calc_load_active
= 0;
7508 #endif /* CONFIG_HOTPLUG_CPU */
7510 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7512 static struct ctl_table sd_ctl_dir
[] = {
7514 .procname
= "sched_domain",
7520 static struct ctl_table sd_ctl_root
[] = {
7522 .procname
= "kernel",
7524 .child
= sd_ctl_dir
,
7529 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
7531 struct ctl_table
*entry
=
7532 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
7537 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
7539 struct ctl_table
*entry
;
7542 * In the intermediate directories, both the child directory and
7543 * procname are dynamically allocated and could fail but the mode
7544 * will always be set. In the lowest directory the names are
7545 * static strings and all have proc handlers.
7547 for (entry
= *tablep
; entry
->mode
; entry
++) {
7549 sd_free_ctl_entry(&entry
->child
);
7550 if (entry
->proc_handler
== NULL
)
7551 kfree(entry
->procname
);
7559 set_table_entry(struct ctl_table
*entry
,
7560 const char *procname
, void *data
, int maxlen
,
7561 mode_t mode
, proc_handler
*proc_handler
)
7563 entry
->procname
= procname
;
7565 entry
->maxlen
= maxlen
;
7567 entry
->proc_handler
= proc_handler
;
7570 static struct ctl_table
*
7571 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
7573 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
7578 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
7579 sizeof(long), 0644, proc_doulongvec_minmax
);
7580 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
7581 sizeof(long), 0644, proc_doulongvec_minmax
);
7582 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
7583 sizeof(int), 0644, proc_dointvec_minmax
);
7584 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
7585 sizeof(int), 0644, proc_dointvec_minmax
);
7586 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
7587 sizeof(int), 0644, proc_dointvec_minmax
);
7588 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
7589 sizeof(int), 0644, proc_dointvec_minmax
);
7590 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
7591 sizeof(int), 0644, proc_dointvec_minmax
);
7592 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
7593 sizeof(int), 0644, proc_dointvec_minmax
);
7594 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
7595 sizeof(int), 0644, proc_dointvec_minmax
);
7596 set_table_entry(&table
[9], "cache_nice_tries",
7597 &sd
->cache_nice_tries
,
7598 sizeof(int), 0644, proc_dointvec_minmax
);
7599 set_table_entry(&table
[10], "flags", &sd
->flags
,
7600 sizeof(int), 0644, proc_dointvec_minmax
);
7601 set_table_entry(&table
[11], "name", sd
->name
,
7602 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
7603 /* &table[12] is terminator */
7608 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
7610 struct ctl_table
*entry
, *table
;
7611 struct sched_domain
*sd
;
7612 int domain_num
= 0, i
;
7615 for_each_domain(cpu
, sd
)
7617 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
7622 for_each_domain(cpu
, sd
) {
7623 snprintf(buf
, 32, "domain%d", i
);
7624 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7626 entry
->child
= sd_alloc_ctl_domain_table(sd
);
7633 static struct ctl_table_header
*sd_sysctl_header
;
7634 static void register_sched_domain_sysctl(void)
7636 int i
, cpu_num
= num_possible_cpus();
7637 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
7640 WARN_ON(sd_ctl_dir
[0].child
);
7641 sd_ctl_dir
[0].child
= entry
;
7646 for_each_possible_cpu(i
) {
7647 snprintf(buf
, 32, "cpu%d", i
);
7648 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7650 entry
->child
= sd_alloc_ctl_cpu_table(i
);
7654 WARN_ON(sd_sysctl_header
);
7655 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
7658 /* may be called multiple times per register */
7659 static void unregister_sched_domain_sysctl(void)
7661 if (sd_sysctl_header
)
7662 unregister_sysctl_table(sd_sysctl_header
);
7663 sd_sysctl_header
= NULL
;
7664 if (sd_ctl_dir
[0].child
)
7665 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
7668 static void register_sched_domain_sysctl(void)
7671 static void unregister_sched_domain_sysctl(void)
7676 static void set_rq_online(struct rq
*rq
)
7679 const struct sched_class
*class;
7681 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
7684 for_each_class(class) {
7685 if (class->rq_online
)
7686 class->rq_online(rq
);
7691 static void set_rq_offline(struct rq
*rq
)
7694 const struct sched_class
*class;
7696 for_each_class(class) {
7697 if (class->rq_offline
)
7698 class->rq_offline(rq
);
7701 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
7707 * migration_call - callback that gets triggered when a CPU is added.
7708 * Here we can start up the necessary migration thread for the new CPU.
7710 static int __cpuinit
7711 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
7713 struct task_struct
*p
;
7714 int cpu
= (long)hcpu
;
7715 unsigned long flags
;
7720 case CPU_UP_PREPARE
:
7721 case CPU_UP_PREPARE_FROZEN
:
7722 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
7725 kthread_bind(p
, cpu
);
7726 /* Must be high prio: stop_machine expects to yield to it. */
7727 rq
= task_rq_lock(p
, &flags
);
7728 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7729 task_rq_unlock(rq
, &flags
);
7731 cpu_rq(cpu
)->migration_thread
= p
;
7732 rq
->calc_load_update
= calc_load_update
;
7736 case CPU_ONLINE_FROZEN
:
7737 /* Strictly unnecessary, as first user will wake it. */
7738 wake_up_process(cpu_rq(cpu
)->migration_thread
);
7740 /* Update our root-domain */
7742 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7744 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7748 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7751 #ifdef CONFIG_HOTPLUG_CPU
7752 case CPU_UP_CANCELED
:
7753 case CPU_UP_CANCELED_FROZEN
:
7754 if (!cpu_rq(cpu
)->migration_thread
)
7756 /* Unbind it from offline cpu so it can run. Fall thru. */
7757 kthread_bind(cpu_rq(cpu
)->migration_thread
,
7758 cpumask_any(cpu_online_mask
));
7759 kthread_stop(cpu_rq(cpu
)->migration_thread
);
7760 put_task_struct(cpu_rq(cpu
)->migration_thread
);
7761 cpu_rq(cpu
)->migration_thread
= NULL
;
7765 case CPU_DEAD_FROZEN
:
7766 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7767 migrate_live_tasks(cpu
);
7769 kthread_stop(rq
->migration_thread
);
7770 put_task_struct(rq
->migration_thread
);
7771 rq
->migration_thread
= NULL
;
7772 /* Idle task back to normal (off runqueue, low prio) */
7773 raw_spin_lock_irq(&rq
->lock
);
7774 update_rq_clock(rq
);
7775 deactivate_task(rq
, rq
->idle
, 0);
7776 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
7777 rq
->idle
->sched_class
= &idle_sched_class
;
7778 migrate_dead_tasks(cpu
);
7779 raw_spin_unlock_irq(&rq
->lock
);
7781 migrate_nr_uninterruptible(rq
);
7782 BUG_ON(rq
->nr_running
!= 0);
7783 calc_global_load_remove(rq
);
7785 * No need to migrate the tasks: it was best-effort if
7786 * they didn't take sched_hotcpu_mutex. Just wake up
7789 raw_spin_lock_irq(&rq
->lock
);
7790 while (!list_empty(&rq
->migration_queue
)) {
7791 struct migration_req
*req
;
7793 req
= list_entry(rq
->migration_queue
.next
,
7794 struct migration_req
, list
);
7795 list_del_init(&req
->list
);
7796 raw_spin_unlock_irq(&rq
->lock
);
7797 complete(&req
->done
);
7798 raw_spin_lock_irq(&rq
->lock
);
7800 raw_spin_unlock_irq(&rq
->lock
);
7804 case CPU_DYING_FROZEN
:
7805 /* Update our root-domain */
7807 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7809 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7812 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7820 * Register at high priority so that task migration (migrate_all_tasks)
7821 * happens before everything else. This has to be lower priority than
7822 * the notifier in the perf_event subsystem, though.
7824 static struct notifier_block __cpuinitdata migration_notifier
= {
7825 .notifier_call
= migration_call
,
7829 static int __init
migration_init(void)
7831 void *cpu
= (void *)(long)smp_processor_id();
7834 /* Start one for the boot CPU: */
7835 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
7836 BUG_ON(err
== NOTIFY_BAD
);
7837 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
7838 register_cpu_notifier(&migration_notifier
);
7842 early_initcall(migration_init
);
7847 #ifdef CONFIG_SCHED_DEBUG
7849 static __read_mostly
int sched_domain_debug_enabled
;
7851 static int __init
sched_domain_debug_setup(char *str
)
7853 sched_domain_debug_enabled
= 1;
7857 early_param("sched_debug", sched_domain_debug_setup
);
7859 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
7860 struct cpumask
*groupmask
)
7862 struct sched_group
*group
= sd
->groups
;
7865 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
7866 cpumask_clear(groupmask
);
7868 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
7870 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
7871 printk("does not load-balance\n");
7873 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
7878 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
7880 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
7881 printk(KERN_ERR
"ERROR: domain->span does not contain "
7884 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
7885 printk(KERN_ERR
"ERROR: domain->groups does not contain"
7889 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
7893 printk(KERN_ERR
"ERROR: group is NULL\n");
7897 if (!group
->cpu_power
) {
7898 printk(KERN_CONT
"\n");
7899 printk(KERN_ERR
"ERROR: domain->cpu_power not "
7904 if (!cpumask_weight(sched_group_cpus(group
))) {
7905 printk(KERN_CONT
"\n");
7906 printk(KERN_ERR
"ERROR: empty group\n");
7910 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
7911 printk(KERN_CONT
"\n");
7912 printk(KERN_ERR
"ERROR: repeated CPUs\n");
7916 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
7918 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
7920 printk(KERN_CONT
" %s", str
);
7921 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
7922 printk(KERN_CONT
" (cpu_power = %d)",
7926 group
= group
->next
;
7927 } while (group
!= sd
->groups
);
7928 printk(KERN_CONT
"\n");
7930 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
7931 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
7934 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
7935 printk(KERN_ERR
"ERROR: parent span is not a superset "
7936 "of domain->span\n");
7940 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
7942 cpumask_var_t groupmask
;
7945 if (!sched_domain_debug_enabled
)
7949 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
7953 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
7955 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
7956 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
7961 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
7968 free_cpumask_var(groupmask
);
7970 #else /* !CONFIG_SCHED_DEBUG */
7971 # define sched_domain_debug(sd, cpu) do { } while (0)
7972 #endif /* CONFIG_SCHED_DEBUG */
7974 static int sd_degenerate(struct sched_domain
*sd
)
7976 if (cpumask_weight(sched_domain_span(sd
)) == 1)
7979 /* Following flags need at least 2 groups */
7980 if (sd
->flags
& (SD_LOAD_BALANCE
|
7981 SD_BALANCE_NEWIDLE
|
7985 SD_SHARE_PKG_RESOURCES
)) {
7986 if (sd
->groups
!= sd
->groups
->next
)
7990 /* Following flags don't use groups */
7991 if (sd
->flags
& (SD_WAKE_AFFINE
))
7998 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
8000 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
8002 if (sd_degenerate(parent
))
8005 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
8008 /* Flags needing groups don't count if only 1 group in parent */
8009 if (parent
->groups
== parent
->groups
->next
) {
8010 pflags
&= ~(SD_LOAD_BALANCE
|
8011 SD_BALANCE_NEWIDLE
|
8015 SD_SHARE_PKG_RESOURCES
);
8016 if (nr_node_ids
== 1)
8017 pflags
&= ~SD_SERIALIZE
;
8019 if (~cflags
& pflags
)
8025 static void free_rootdomain(struct root_domain
*rd
)
8027 synchronize_sched();
8029 cpupri_cleanup(&rd
->cpupri
);
8031 free_cpumask_var(rd
->rto_mask
);
8032 free_cpumask_var(rd
->online
);
8033 free_cpumask_var(rd
->span
);
8037 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
8039 struct root_domain
*old_rd
= NULL
;
8040 unsigned long flags
;
8042 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8047 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
8050 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
8053 * If we dont want to free the old_rt yet then
8054 * set old_rd to NULL to skip the freeing later
8057 if (!atomic_dec_and_test(&old_rd
->refcount
))
8061 atomic_inc(&rd
->refcount
);
8064 cpumask_set_cpu(rq
->cpu
, rd
->span
);
8065 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
8068 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8071 free_rootdomain(old_rd
);
8074 static int init_rootdomain(struct root_domain
*rd
, bool bootmem
)
8076 gfp_t gfp
= GFP_KERNEL
;
8078 memset(rd
, 0, sizeof(*rd
));
8083 if (!alloc_cpumask_var(&rd
->span
, gfp
))
8085 if (!alloc_cpumask_var(&rd
->online
, gfp
))
8087 if (!alloc_cpumask_var(&rd
->rto_mask
, gfp
))
8090 if (cpupri_init(&rd
->cpupri
, bootmem
) != 0)
8095 free_cpumask_var(rd
->rto_mask
);
8097 free_cpumask_var(rd
->online
);
8099 free_cpumask_var(rd
->span
);
8104 static void init_defrootdomain(void)
8106 init_rootdomain(&def_root_domain
, true);
8108 atomic_set(&def_root_domain
.refcount
, 1);
8111 static struct root_domain
*alloc_rootdomain(void)
8113 struct root_domain
*rd
;
8115 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
8119 if (init_rootdomain(rd
, false) != 0) {
8128 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8129 * hold the hotplug lock.
8132 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
8134 struct rq
*rq
= cpu_rq(cpu
);
8135 struct sched_domain
*tmp
;
8137 /* Remove the sched domains which do not contribute to scheduling. */
8138 for (tmp
= sd
; tmp
; ) {
8139 struct sched_domain
*parent
= tmp
->parent
;
8143 if (sd_parent_degenerate(tmp
, parent
)) {
8144 tmp
->parent
= parent
->parent
;
8146 parent
->parent
->child
= tmp
;
8151 if (sd
&& sd_degenerate(sd
)) {
8157 sched_domain_debug(sd
, cpu
);
8159 rq_attach_root(rq
, rd
);
8160 rcu_assign_pointer(rq
->sd
, sd
);
8163 /* cpus with isolated domains */
8164 static cpumask_var_t cpu_isolated_map
;
8166 /* Setup the mask of cpus configured for isolated domains */
8167 static int __init
isolated_cpu_setup(char *str
)
8169 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
8170 cpulist_parse(str
, cpu_isolated_map
);
8174 __setup("isolcpus=", isolated_cpu_setup
);
8177 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8178 * to a function which identifies what group(along with sched group) a CPU
8179 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8180 * (due to the fact that we keep track of groups covered with a struct cpumask).
8182 * init_sched_build_groups will build a circular linked list of the groups
8183 * covered by the given span, and will set each group's ->cpumask correctly,
8184 * and ->cpu_power to 0.
8187 init_sched_build_groups(const struct cpumask
*span
,
8188 const struct cpumask
*cpu_map
,
8189 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
8190 struct sched_group
**sg
,
8191 struct cpumask
*tmpmask
),
8192 struct cpumask
*covered
, struct cpumask
*tmpmask
)
8194 struct sched_group
*first
= NULL
, *last
= NULL
;
8197 cpumask_clear(covered
);
8199 for_each_cpu(i
, span
) {
8200 struct sched_group
*sg
;
8201 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
8204 if (cpumask_test_cpu(i
, covered
))
8207 cpumask_clear(sched_group_cpus(sg
));
8210 for_each_cpu(j
, span
) {
8211 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
8214 cpumask_set_cpu(j
, covered
);
8215 cpumask_set_cpu(j
, sched_group_cpus(sg
));
8226 #define SD_NODES_PER_DOMAIN 16
8231 * find_next_best_node - find the next node to include in a sched_domain
8232 * @node: node whose sched_domain we're building
8233 * @used_nodes: nodes already in the sched_domain
8235 * Find the next node to include in a given scheduling domain. Simply
8236 * finds the closest node not already in the @used_nodes map.
8238 * Should use nodemask_t.
8240 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
8242 int i
, n
, val
, min_val
, best_node
= 0;
8246 for (i
= 0; i
< nr_node_ids
; i
++) {
8247 /* Start at @node */
8248 n
= (node
+ i
) % nr_node_ids
;
8250 if (!nr_cpus_node(n
))
8253 /* Skip already used nodes */
8254 if (node_isset(n
, *used_nodes
))
8257 /* Simple min distance search */
8258 val
= node_distance(node
, n
);
8260 if (val
< min_val
) {
8266 node_set(best_node
, *used_nodes
);
8271 * sched_domain_node_span - get a cpumask for a node's sched_domain
8272 * @node: node whose cpumask we're constructing
8273 * @span: resulting cpumask
8275 * Given a node, construct a good cpumask for its sched_domain to span. It
8276 * should be one that prevents unnecessary balancing, but also spreads tasks
8279 static void sched_domain_node_span(int node
, struct cpumask
*span
)
8281 nodemask_t used_nodes
;
8284 cpumask_clear(span
);
8285 nodes_clear(used_nodes
);
8287 cpumask_or(span
, span
, cpumask_of_node(node
));
8288 node_set(node
, used_nodes
);
8290 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
8291 int next_node
= find_next_best_node(node
, &used_nodes
);
8293 cpumask_or(span
, span
, cpumask_of_node(next_node
));
8296 #endif /* CONFIG_NUMA */
8298 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
8301 * The cpus mask in sched_group and sched_domain hangs off the end.
8303 * ( See the the comments in include/linux/sched.h:struct sched_group
8304 * and struct sched_domain. )
8306 struct static_sched_group
{
8307 struct sched_group sg
;
8308 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
8311 struct static_sched_domain
{
8312 struct sched_domain sd
;
8313 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
8319 cpumask_var_t domainspan
;
8320 cpumask_var_t covered
;
8321 cpumask_var_t notcovered
;
8323 cpumask_var_t nodemask
;
8324 cpumask_var_t this_sibling_map
;
8325 cpumask_var_t this_core_map
;
8326 cpumask_var_t send_covered
;
8327 cpumask_var_t tmpmask
;
8328 struct sched_group
**sched_group_nodes
;
8329 struct root_domain
*rd
;
8333 sa_sched_groups
= 0,
8338 sa_this_sibling_map
,
8340 sa_sched_group_nodes
,
8350 * SMT sched-domains:
8352 #ifdef CONFIG_SCHED_SMT
8353 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
8354 static DEFINE_PER_CPU(struct static_sched_group
, sched_groups
);
8357 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
8358 struct sched_group
**sg
, struct cpumask
*unused
)
8361 *sg
= &per_cpu(sched_groups
, cpu
).sg
;
8364 #endif /* CONFIG_SCHED_SMT */
8367 * multi-core sched-domains:
8369 #ifdef CONFIG_SCHED_MC
8370 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
8371 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
8372 #endif /* CONFIG_SCHED_MC */
8374 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8376 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8377 struct sched_group
**sg
, struct cpumask
*mask
)
8381 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8382 group
= cpumask_first(mask
);
8384 *sg
= &per_cpu(sched_group_core
, group
).sg
;
8387 #elif defined(CONFIG_SCHED_MC)
8389 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8390 struct sched_group
**sg
, struct cpumask
*unused
)
8393 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
8398 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
8399 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
8402 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
8403 struct sched_group
**sg
, struct cpumask
*mask
)
8406 #ifdef CONFIG_SCHED_MC
8407 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
8408 group
= cpumask_first(mask
);
8409 #elif defined(CONFIG_SCHED_SMT)
8410 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8411 group
= cpumask_first(mask
);
8416 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
8422 * The init_sched_build_groups can't handle what we want to do with node
8423 * groups, so roll our own. Now each node has its own list of groups which
8424 * gets dynamically allocated.
8426 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
8427 static struct sched_group
***sched_group_nodes_bycpu
;
8429 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
8430 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
8432 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
8433 struct sched_group
**sg
,
8434 struct cpumask
*nodemask
)
8438 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
8439 group
= cpumask_first(nodemask
);
8442 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
8446 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
8448 struct sched_group
*sg
= group_head
;
8454 for_each_cpu(j
, sched_group_cpus(sg
)) {
8455 struct sched_domain
*sd
;
8457 sd
= &per_cpu(phys_domains
, j
).sd
;
8458 if (j
!= group_first_cpu(sd
->groups
)) {
8460 * Only add "power" once for each
8466 sg
->cpu_power
+= sd
->groups
->cpu_power
;
8469 } while (sg
!= group_head
);
8472 static int build_numa_sched_groups(struct s_data
*d
,
8473 const struct cpumask
*cpu_map
, int num
)
8475 struct sched_domain
*sd
;
8476 struct sched_group
*sg
, *prev
;
8479 cpumask_clear(d
->covered
);
8480 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
8481 if (cpumask_empty(d
->nodemask
)) {
8482 d
->sched_group_nodes
[num
] = NULL
;
8486 sched_domain_node_span(num
, d
->domainspan
);
8487 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
8489 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8492 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
8496 d
->sched_group_nodes
[num
] = sg
;
8498 for_each_cpu(j
, d
->nodemask
) {
8499 sd
= &per_cpu(node_domains
, j
).sd
;
8504 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
8506 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
8509 for (j
= 0; j
< nr_node_ids
; j
++) {
8510 n
= (num
+ j
) % nr_node_ids
;
8511 cpumask_complement(d
->notcovered
, d
->covered
);
8512 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
8513 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
8514 if (cpumask_empty(d
->tmpmask
))
8516 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
8517 if (cpumask_empty(d
->tmpmask
))
8519 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8523 "Can not alloc domain group for node %d\n", j
);
8527 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
8528 sg
->next
= prev
->next
;
8529 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
8536 #endif /* CONFIG_NUMA */
8539 /* Free memory allocated for various sched_group structures */
8540 static void free_sched_groups(const struct cpumask
*cpu_map
,
8541 struct cpumask
*nodemask
)
8545 for_each_cpu(cpu
, cpu_map
) {
8546 struct sched_group
**sched_group_nodes
8547 = sched_group_nodes_bycpu
[cpu
];
8549 if (!sched_group_nodes
)
8552 for (i
= 0; i
< nr_node_ids
; i
++) {
8553 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
8555 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8556 if (cpumask_empty(nodemask
))
8566 if (oldsg
!= sched_group_nodes
[i
])
8569 kfree(sched_group_nodes
);
8570 sched_group_nodes_bycpu
[cpu
] = NULL
;
8573 #else /* !CONFIG_NUMA */
8574 static void free_sched_groups(const struct cpumask
*cpu_map
,
8575 struct cpumask
*nodemask
)
8578 #endif /* CONFIG_NUMA */
8581 * Initialize sched groups cpu_power.
8583 * cpu_power indicates the capacity of sched group, which is used while
8584 * distributing the load between different sched groups in a sched domain.
8585 * Typically cpu_power for all the groups in a sched domain will be same unless
8586 * there are asymmetries in the topology. If there are asymmetries, group
8587 * having more cpu_power will pickup more load compared to the group having
8590 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
8592 struct sched_domain
*child
;
8593 struct sched_group
*group
;
8597 WARN_ON(!sd
|| !sd
->groups
);
8599 if (cpu
!= group_first_cpu(sd
->groups
))
8604 sd
->groups
->cpu_power
= 0;
8607 power
= SCHED_LOAD_SCALE
;
8608 weight
= cpumask_weight(sched_domain_span(sd
));
8610 * SMT siblings share the power of a single core.
8611 * Usually multiple threads get a better yield out of
8612 * that one core than a single thread would have,
8613 * reflect that in sd->smt_gain.
8615 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
8616 power
*= sd
->smt_gain
;
8618 power
>>= SCHED_LOAD_SHIFT
;
8620 sd
->groups
->cpu_power
+= power
;
8625 * Add cpu_power of each child group to this groups cpu_power.
8627 group
= child
->groups
;
8629 sd
->groups
->cpu_power
+= group
->cpu_power
;
8630 group
= group
->next
;
8631 } while (group
!= child
->groups
);
8635 * Initializers for schedule domains
8636 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8639 #ifdef CONFIG_SCHED_DEBUG
8640 # define SD_INIT_NAME(sd, type) sd->name = #type
8642 # define SD_INIT_NAME(sd, type) do { } while (0)
8645 #define SD_INIT(sd, type) sd_init_##type(sd)
8647 #define SD_INIT_FUNC(type) \
8648 static noinline void sd_init_##type(struct sched_domain *sd) \
8650 memset(sd, 0, sizeof(*sd)); \
8651 *sd = SD_##type##_INIT; \
8652 sd->level = SD_LV_##type; \
8653 SD_INIT_NAME(sd, type); \
8658 SD_INIT_FUNC(ALLNODES
)
8661 #ifdef CONFIG_SCHED_SMT
8662 SD_INIT_FUNC(SIBLING
)
8664 #ifdef CONFIG_SCHED_MC
8668 static int default_relax_domain_level
= -1;
8670 static int __init
setup_relax_domain_level(char *str
)
8674 val
= simple_strtoul(str
, NULL
, 0);
8675 if (val
< SD_LV_MAX
)
8676 default_relax_domain_level
= val
;
8680 __setup("relax_domain_level=", setup_relax_domain_level
);
8682 static void set_domain_attribute(struct sched_domain
*sd
,
8683 struct sched_domain_attr
*attr
)
8687 if (!attr
|| attr
->relax_domain_level
< 0) {
8688 if (default_relax_domain_level
< 0)
8691 request
= default_relax_domain_level
;
8693 request
= attr
->relax_domain_level
;
8694 if (request
< sd
->level
) {
8695 /* turn off idle balance on this domain */
8696 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
8698 /* turn on idle balance on this domain */
8699 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
8703 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
8704 const struct cpumask
*cpu_map
)
8707 case sa_sched_groups
:
8708 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
8709 d
->sched_group_nodes
= NULL
;
8711 free_rootdomain(d
->rd
); /* fall through */
8713 free_cpumask_var(d
->tmpmask
); /* fall through */
8714 case sa_send_covered
:
8715 free_cpumask_var(d
->send_covered
); /* fall through */
8716 case sa_this_core_map
:
8717 free_cpumask_var(d
->this_core_map
); /* fall through */
8718 case sa_this_sibling_map
:
8719 free_cpumask_var(d
->this_sibling_map
); /* fall through */
8721 free_cpumask_var(d
->nodemask
); /* fall through */
8722 case sa_sched_group_nodes
:
8724 kfree(d
->sched_group_nodes
); /* fall through */
8726 free_cpumask_var(d
->notcovered
); /* fall through */
8728 free_cpumask_var(d
->covered
); /* fall through */
8730 free_cpumask_var(d
->domainspan
); /* fall through */
8737 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
8738 const struct cpumask
*cpu_map
)
8741 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
8743 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
8744 return sa_domainspan
;
8745 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
8747 /* Allocate the per-node list of sched groups */
8748 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
8749 sizeof(struct sched_group
*), GFP_KERNEL
);
8750 if (!d
->sched_group_nodes
) {
8751 printk(KERN_WARNING
"Can not alloc sched group node list\n");
8752 return sa_notcovered
;
8754 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
8756 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
8757 return sa_sched_group_nodes
;
8758 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
8760 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
8761 return sa_this_sibling_map
;
8762 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
8763 return sa_this_core_map
;
8764 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
8765 return sa_send_covered
;
8766 d
->rd
= alloc_rootdomain();
8768 printk(KERN_WARNING
"Cannot alloc root domain\n");
8771 return sa_rootdomain
;
8774 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
8775 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
8777 struct sched_domain
*sd
= NULL
;
8779 struct sched_domain
*parent
;
8782 if (cpumask_weight(cpu_map
) >
8783 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
8784 sd
= &per_cpu(allnodes_domains
, i
).sd
;
8785 SD_INIT(sd
, ALLNODES
);
8786 set_domain_attribute(sd
, attr
);
8787 cpumask_copy(sched_domain_span(sd
), cpu_map
);
8788 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8793 sd
= &per_cpu(node_domains
, i
).sd
;
8795 set_domain_attribute(sd
, attr
);
8796 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
8797 sd
->parent
= parent
;
8800 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
8805 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
8806 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8807 struct sched_domain
*parent
, int i
)
8809 struct sched_domain
*sd
;
8810 sd
= &per_cpu(phys_domains
, i
).sd
;
8812 set_domain_attribute(sd
, attr
);
8813 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
8814 sd
->parent
= parent
;
8817 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8821 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
8822 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8823 struct sched_domain
*parent
, int i
)
8825 struct sched_domain
*sd
= parent
;
8826 #ifdef CONFIG_SCHED_MC
8827 sd
= &per_cpu(core_domains
, i
).sd
;
8829 set_domain_attribute(sd
, attr
);
8830 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
8831 sd
->parent
= parent
;
8833 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8838 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
8839 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8840 struct sched_domain
*parent
, int i
)
8842 struct sched_domain
*sd
= parent
;
8843 #ifdef CONFIG_SCHED_SMT
8844 sd
= &per_cpu(cpu_domains
, i
).sd
;
8845 SD_INIT(sd
, SIBLING
);
8846 set_domain_attribute(sd
, attr
);
8847 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
8848 sd
->parent
= parent
;
8850 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8855 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
8856 const struct cpumask
*cpu_map
, int cpu
)
8859 #ifdef CONFIG_SCHED_SMT
8860 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
8861 cpumask_and(d
->this_sibling_map
, cpu_map
,
8862 topology_thread_cpumask(cpu
));
8863 if (cpu
== cpumask_first(d
->this_sibling_map
))
8864 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
8866 d
->send_covered
, d
->tmpmask
);
8869 #ifdef CONFIG_SCHED_MC
8870 case SD_LV_MC
: /* set up multi-core groups */
8871 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
8872 if (cpu
== cpumask_first(d
->this_core_map
))
8873 init_sched_build_groups(d
->this_core_map
, cpu_map
,
8875 d
->send_covered
, d
->tmpmask
);
8878 case SD_LV_CPU
: /* set up physical groups */
8879 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
8880 if (!cpumask_empty(d
->nodemask
))
8881 init_sched_build_groups(d
->nodemask
, cpu_map
,
8883 d
->send_covered
, d
->tmpmask
);
8886 case SD_LV_ALLNODES
:
8887 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
8888 d
->send_covered
, d
->tmpmask
);
8897 * Build sched domains for a given set of cpus and attach the sched domains
8898 * to the individual cpus
8900 static int __build_sched_domains(const struct cpumask
*cpu_map
,
8901 struct sched_domain_attr
*attr
)
8903 enum s_alloc alloc_state
= sa_none
;
8905 struct sched_domain
*sd
;
8911 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
8912 if (alloc_state
!= sa_rootdomain
)
8914 alloc_state
= sa_sched_groups
;
8917 * Set up domains for cpus specified by the cpu_map.
8919 for_each_cpu(i
, cpu_map
) {
8920 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
8923 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
8924 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8925 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8926 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8929 for_each_cpu(i
, cpu_map
) {
8930 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
8931 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
8934 /* Set up physical groups */
8935 for (i
= 0; i
< nr_node_ids
; i
++)
8936 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
8939 /* Set up node groups */
8941 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
8943 for (i
= 0; i
< nr_node_ids
; i
++)
8944 if (build_numa_sched_groups(&d
, cpu_map
, i
))
8948 /* Calculate CPU power for physical packages and nodes */
8949 #ifdef CONFIG_SCHED_SMT
8950 for_each_cpu(i
, cpu_map
) {
8951 sd
= &per_cpu(cpu_domains
, i
).sd
;
8952 init_sched_groups_power(i
, sd
);
8955 #ifdef CONFIG_SCHED_MC
8956 for_each_cpu(i
, cpu_map
) {
8957 sd
= &per_cpu(core_domains
, i
).sd
;
8958 init_sched_groups_power(i
, sd
);
8962 for_each_cpu(i
, cpu_map
) {
8963 sd
= &per_cpu(phys_domains
, i
).sd
;
8964 init_sched_groups_power(i
, sd
);
8968 for (i
= 0; i
< nr_node_ids
; i
++)
8969 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
8971 if (d
.sd_allnodes
) {
8972 struct sched_group
*sg
;
8974 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
8976 init_numa_sched_groups_power(sg
);
8980 /* Attach the domains */
8981 for_each_cpu(i
, cpu_map
) {
8982 #ifdef CONFIG_SCHED_SMT
8983 sd
= &per_cpu(cpu_domains
, i
).sd
;
8984 #elif defined(CONFIG_SCHED_MC)
8985 sd
= &per_cpu(core_domains
, i
).sd
;
8987 sd
= &per_cpu(phys_domains
, i
).sd
;
8989 cpu_attach_domain(sd
, d
.rd
, i
);
8992 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
8993 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
8997 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
9001 static int build_sched_domains(const struct cpumask
*cpu_map
)
9003 return __build_sched_domains(cpu_map
, NULL
);
9006 static cpumask_var_t
*doms_cur
; /* current sched domains */
9007 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
9008 static struct sched_domain_attr
*dattr_cur
;
9009 /* attribues of custom domains in 'doms_cur' */
9012 * Special case: If a kmalloc of a doms_cur partition (array of
9013 * cpumask) fails, then fallback to a single sched domain,
9014 * as determined by the single cpumask fallback_doms.
9016 static cpumask_var_t fallback_doms
;
9019 * arch_update_cpu_topology lets virtualized architectures update the
9020 * cpu core maps. It is supposed to return 1 if the topology changed
9021 * or 0 if it stayed the same.
9023 int __attribute__((weak
)) arch_update_cpu_topology(void)
9028 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
9031 cpumask_var_t
*doms
;
9033 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
9036 for (i
= 0; i
< ndoms
; i
++) {
9037 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
9038 free_sched_domains(doms
, i
);
9045 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
9048 for (i
= 0; i
< ndoms
; i
++)
9049 free_cpumask_var(doms
[i
]);
9054 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
9055 * For now this just excludes isolated cpus, but could be used to
9056 * exclude other special cases in the future.
9058 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
9062 arch_update_cpu_topology();
9064 doms_cur
= alloc_sched_domains(ndoms_cur
);
9066 doms_cur
= &fallback_doms
;
9067 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
9069 err
= build_sched_domains(doms_cur
[0]);
9070 register_sched_domain_sysctl();
9075 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
9076 struct cpumask
*tmpmask
)
9078 free_sched_groups(cpu_map
, tmpmask
);
9082 * Detach sched domains from a group of cpus specified in cpu_map
9083 * These cpus will now be attached to the NULL domain
9085 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
9087 /* Save because hotplug lock held. */
9088 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
9091 for_each_cpu(i
, cpu_map
)
9092 cpu_attach_domain(NULL
, &def_root_domain
, i
);
9093 synchronize_sched();
9094 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
9097 /* handle null as "default" */
9098 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
9099 struct sched_domain_attr
*new, int idx_new
)
9101 struct sched_domain_attr tmp
;
9108 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
9109 new ? (new + idx_new
) : &tmp
,
9110 sizeof(struct sched_domain_attr
));
9114 * Partition sched domains as specified by the 'ndoms_new'
9115 * cpumasks in the array doms_new[] of cpumasks. This compares
9116 * doms_new[] to the current sched domain partitioning, doms_cur[].
9117 * It destroys each deleted domain and builds each new domain.
9119 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
9120 * The masks don't intersect (don't overlap.) We should setup one
9121 * sched domain for each mask. CPUs not in any of the cpumasks will
9122 * not be load balanced. If the same cpumask appears both in the
9123 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9126 * The passed in 'doms_new' should be allocated using
9127 * alloc_sched_domains. This routine takes ownership of it and will
9128 * free_sched_domains it when done with it. If the caller failed the
9129 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
9130 * and partition_sched_domains() will fallback to the single partition
9131 * 'fallback_doms', it also forces the domains to be rebuilt.
9133 * If doms_new == NULL it will be replaced with cpu_online_mask.
9134 * ndoms_new == 0 is a special case for destroying existing domains,
9135 * and it will not create the default domain.
9137 * Call with hotplug lock held
9139 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
9140 struct sched_domain_attr
*dattr_new
)
9145 mutex_lock(&sched_domains_mutex
);
9147 /* always unregister in case we don't destroy any domains */
9148 unregister_sched_domain_sysctl();
9150 /* Let architecture update cpu core mappings. */
9151 new_topology
= arch_update_cpu_topology();
9153 n
= doms_new
? ndoms_new
: 0;
9155 /* Destroy deleted domains */
9156 for (i
= 0; i
< ndoms_cur
; i
++) {
9157 for (j
= 0; j
< n
&& !new_topology
; j
++) {
9158 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
9159 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
9162 /* no match - a current sched domain not in new doms_new[] */
9163 detach_destroy_domains(doms_cur
[i
]);
9168 if (doms_new
== NULL
) {
9170 doms_new
= &fallback_doms
;
9171 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
9172 WARN_ON_ONCE(dattr_new
);
9175 /* Build new domains */
9176 for (i
= 0; i
< ndoms_new
; i
++) {
9177 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
9178 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
9179 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
9182 /* no match - add a new doms_new */
9183 __build_sched_domains(doms_new
[i
],
9184 dattr_new
? dattr_new
+ i
: NULL
);
9189 /* Remember the new sched domains */
9190 if (doms_cur
!= &fallback_doms
)
9191 free_sched_domains(doms_cur
, ndoms_cur
);
9192 kfree(dattr_cur
); /* kfree(NULL) is safe */
9193 doms_cur
= doms_new
;
9194 dattr_cur
= dattr_new
;
9195 ndoms_cur
= ndoms_new
;
9197 register_sched_domain_sysctl();
9199 mutex_unlock(&sched_domains_mutex
);
9202 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9203 static void arch_reinit_sched_domains(void)
9207 /* Destroy domains first to force the rebuild */
9208 partition_sched_domains(0, NULL
, NULL
);
9210 rebuild_sched_domains();
9214 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
9216 unsigned int level
= 0;
9218 if (sscanf(buf
, "%u", &level
) != 1)
9222 * level is always be positive so don't check for
9223 * level < POWERSAVINGS_BALANCE_NONE which is 0
9224 * What happens on 0 or 1 byte write,
9225 * need to check for count as well?
9228 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
9232 sched_smt_power_savings
= level
;
9234 sched_mc_power_savings
= level
;
9236 arch_reinit_sched_domains();
9241 #ifdef CONFIG_SCHED_MC
9242 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
9245 return sprintf(page
, "%u\n", sched_mc_power_savings
);
9247 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
9248 const char *buf
, size_t count
)
9250 return sched_power_savings_store(buf
, count
, 0);
9252 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
9253 sched_mc_power_savings_show
,
9254 sched_mc_power_savings_store
);
9257 #ifdef CONFIG_SCHED_SMT
9258 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
9261 return sprintf(page
, "%u\n", sched_smt_power_savings
);
9263 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
9264 const char *buf
, size_t count
)
9266 return sched_power_savings_store(buf
, count
, 1);
9268 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
9269 sched_smt_power_savings_show
,
9270 sched_smt_power_savings_store
);
9273 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
9277 #ifdef CONFIG_SCHED_SMT
9279 err
= sysfs_create_file(&cls
->kset
.kobj
,
9280 &attr_sched_smt_power_savings
.attr
);
9282 #ifdef CONFIG_SCHED_MC
9283 if (!err
&& mc_capable())
9284 err
= sysfs_create_file(&cls
->kset
.kobj
,
9285 &attr_sched_mc_power_savings
.attr
);
9289 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9291 #ifndef CONFIG_CPUSETS
9293 * Add online and remove offline CPUs from the scheduler domains.
9294 * When cpusets are enabled they take over this function.
9296 static int update_sched_domains(struct notifier_block
*nfb
,
9297 unsigned long action
, void *hcpu
)
9301 case CPU_ONLINE_FROZEN
:
9302 case CPU_DOWN_PREPARE
:
9303 case CPU_DOWN_PREPARE_FROZEN
:
9304 case CPU_DOWN_FAILED
:
9305 case CPU_DOWN_FAILED_FROZEN
:
9306 partition_sched_domains(1, NULL
, NULL
);
9315 static int update_runtime(struct notifier_block
*nfb
,
9316 unsigned long action
, void *hcpu
)
9318 int cpu
= (int)(long)hcpu
;
9321 case CPU_DOWN_PREPARE
:
9322 case CPU_DOWN_PREPARE_FROZEN
:
9323 disable_runtime(cpu_rq(cpu
));
9326 case CPU_DOWN_FAILED
:
9327 case CPU_DOWN_FAILED_FROZEN
:
9329 case CPU_ONLINE_FROZEN
:
9330 enable_runtime(cpu_rq(cpu
));
9338 void __init
sched_init_smp(void)
9340 cpumask_var_t non_isolated_cpus
;
9342 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
9343 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
9345 #if defined(CONFIG_NUMA)
9346 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
9348 BUG_ON(sched_group_nodes_bycpu
== NULL
);
9351 mutex_lock(&sched_domains_mutex
);
9352 arch_init_sched_domains(cpu_active_mask
);
9353 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
9354 if (cpumask_empty(non_isolated_cpus
))
9355 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
9356 mutex_unlock(&sched_domains_mutex
);
9359 #ifndef CONFIG_CPUSETS
9360 /* XXX: Theoretical race here - CPU may be hotplugged now */
9361 hotcpu_notifier(update_sched_domains
, 0);
9364 /* RT runtime code needs to handle some hotplug events */
9365 hotcpu_notifier(update_runtime
, 0);
9369 /* Move init over to a non-isolated CPU */
9370 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
9372 sched_init_granularity();
9373 free_cpumask_var(non_isolated_cpus
);
9375 init_sched_rt_class();
9378 void __init
sched_init_smp(void)
9380 sched_init_granularity();
9382 #endif /* CONFIG_SMP */
9384 const_debug
unsigned int sysctl_timer_migration
= 1;
9386 int in_sched_functions(unsigned long addr
)
9388 return in_lock_functions(addr
) ||
9389 (addr
>= (unsigned long)__sched_text_start
9390 && addr
< (unsigned long)__sched_text_end
);
9393 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
9395 cfs_rq
->tasks_timeline
= RB_ROOT
;
9396 INIT_LIST_HEAD(&cfs_rq
->tasks
);
9397 #ifdef CONFIG_FAIR_GROUP_SCHED
9400 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
9403 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
9405 struct rt_prio_array
*array
;
9408 array
= &rt_rq
->active
;
9409 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
9410 INIT_LIST_HEAD(array
->queue
+ i
);
9411 __clear_bit(i
, array
->bitmap
);
9413 /* delimiter for bitsearch: */
9414 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
9416 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9417 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
9419 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
9423 rt_rq
->rt_nr_migratory
= 0;
9424 rt_rq
->overloaded
= 0;
9425 plist_head_init_raw(&rt_rq
->pushable_tasks
, &rq
->lock
);
9429 rt_rq
->rt_throttled
= 0;
9430 rt_rq
->rt_runtime
= 0;
9431 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
9433 #ifdef CONFIG_RT_GROUP_SCHED
9434 rt_rq
->rt_nr_boosted
= 0;
9439 #ifdef CONFIG_FAIR_GROUP_SCHED
9440 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
9441 struct sched_entity
*se
, int cpu
, int add
,
9442 struct sched_entity
*parent
)
9444 struct rq
*rq
= cpu_rq(cpu
);
9445 tg
->cfs_rq
[cpu
] = cfs_rq
;
9446 init_cfs_rq(cfs_rq
, rq
);
9449 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
9452 /* se could be NULL for init_task_group */
9457 se
->cfs_rq
= &rq
->cfs
;
9459 se
->cfs_rq
= parent
->my_q
;
9462 se
->load
.weight
= tg
->shares
;
9463 se
->load
.inv_weight
= 0;
9464 se
->parent
= parent
;
9468 #ifdef CONFIG_RT_GROUP_SCHED
9469 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
9470 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
9471 struct sched_rt_entity
*parent
)
9473 struct rq
*rq
= cpu_rq(cpu
);
9475 tg
->rt_rq
[cpu
] = rt_rq
;
9476 init_rt_rq(rt_rq
, rq
);
9478 rt_rq
->rt_se
= rt_se
;
9479 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9481 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
9483 tg
->rt_se
[cpu
] = rt_se
;
9488 rt_se
->rt_rq
= &rq
->rt
;
9490 rt_se
->rt_rq
= parent
->my_q
;
9492 rt_se
->my_q
= rt_rq
;
9493 rt_se
->parent
= parent
;
9494 INIT_LIST_HEAD(&rt_se
->run_list
);
9498 void __init
sched_init(void)
9501 unsigned long alloc_size
= 0, ptr
;
9503 #ifdef CONFIG_FAIR_GROUP_SCHED
9504 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9506 #ifdef CONFIG_RT_GROUP_SCHED
9507 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9509 #ifdef CONFIG_USER_SCHED
9512 #ifdef CONFIG_CPUMASK_OFFSTACK
9513 alloc_size
+= num_possible_cpus() * cpumask_size();
9516 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
9518 #ifdef CONFIG_FAIR_GROUP_SCHED
9519 init_task_group
.se
= (struct sched_entity
**)ptr
;
9520 ptr
+= nr_cpu_ids
* sizeof(void **);
9522 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9523 ptr
+= nr_cpu_ids
* sizeof(void **);
9525 #ifdef CONFIG_USER_SCHED
9526 root_task_group
.se
= (struct sched_entity
**)ptr
;
9527 ptr
+= nr_cpu_ids
* sizeof(void **);
9529 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9530 ptr
+= nr_cpu_ids
* sizeof(void **);
9531 #endif /* CONFIG_USER_SCHED */
9532 #endif /* CONFIG_FAIR_GROUP_SCHED */
9533 #ifdef CONFIG_RT_GROUP_SCHED
9534 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9535 ptr
+= nr_cpu_ids
* sizeof(void **);
9537 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9538 ptr
+= nr_cpu_ids
* sizeof(void **);
9540 #ifdef CONFIG_USER_SCHED
9541 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9542 ptr
+= nr_cpu_ids
* sizeof(void **);
9544 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9545 ptr
+= nr_cpu_ids
* sizeof(void **);
9546 #endif /* CONFIG_USER_SCHED */
9547 #endif /* CONFIG_RT_GROUP_SCHED */
9548 #ifdef CONFIG_CPUMASK_OFFSTACK
9549 for_each_possible_cpu(i
) {
9550 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
9551 ptr
+= cpumask_size();
9553 #endif /* CONFIG_CPUMASK_OFFSTACK */
9557 init_defrootdomain();
9560 init_rt_bandwidth(&def_rt_bandwidth
,
9561 global_rt_period(), global_rt_runtime());
9563 #ifdef CONFIG_RT_GROUP_SCHED
9564 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
9565 global_rt_period(), global_rt_runtime());
9566 #ifdef CONFIG_USER_SCHED
9567 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
9568 global_rt_period(), RUNTIME_INF
);
9569 #endif /* CONFIG_USER_SCHED */
9570 #endif /* CONFIG_RT_GROUP_SCHED */
9572 #ifdef CONFIG_GROUP_SCHED
9573 list_add(&init_task_group
.list
, &task_groups
);
9574 INIT_LIST_HEAD(&init_task_group
.children
);
9576 #ifdef CONFIG_USER_SCHED
9577 INIT_LIST_HEAD(&root_task_group
.children
);
9578 init_task_group
.parent
= &root_task_group
;
9579 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
9580 #endif /* CONFIG_USER_SCHED */
9581 #endif /* CONFIG_GROUP_SCHED */
9583 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9584 update_shares_data
= __alloc_percpu(nr_cpu_ids
* sizeof(unsigned long),
9585 __alignof__(unsigned long));
9587 for_each_possible_cpu(i
) {
9591 raw_spin_lock_init(&rq
->lock
);
9593 rq
->calc_load_active
= 0;
9594 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
9595 init_cfs_rq(&rq
->cfs
, rq
);
9596 init_rt_rq(&rq
->rt
, rq
);
9597 #ifdef CONFIG_FAIR_GROUP_SCHED
9598 init_task_group
.shares
= init_task_group_load
;
9599 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
9600 #ifdef CONFIG_CGROUP_SCHED
9602 * How much cpu bandwidth does init_task_group get?
9604 * In case of task-groups formed thr' the cgroup filesystem, it
9605 * gets 100% of the cpu resources in the system. This overall
9606 * system cpu resource is divided among the tasks of
9607 * init_task_group and its child task-groups in a fair manner,
9608 * based on each entity's (task or task-group's) weight
9609 * (se->load.weight).
9611 * In other words, if init_task_group has 10 tasks of weight
9612 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9613 * then A0's share of the cpu resource is:
9615 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9617 * We achieve this by letting init_task_group's tasks sit
9618 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9620 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
9621 #elif defined CONFIG_USER_SCHED
9622 root_task_group
.shares
= NICE_0_LOAD
;
9623 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
9625 * In case of task-groups formed thr' the user id of tasks,
9626 * init_task_group represents tasks belonging to root user.
9627 * Hence it forms a sibling of all subsequent groups formed.
9628 * In this case, init_task_group gets only a fraction of overall
9629 * system cpu resource, based on the weight assigned to root
9630 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9631 * by letting tasks of init_task_group sit in a separate cfs_rq
9632 * (init_tg_cfs_rq) and having one entity represent this group of
9633 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9635 init_tg_cfs_entry(&init_task_group
,
9636 &per_cpu(init_tg_cfs_rq
, i
),
9637 &per_cpu(init_sched_entity
, i
), i
, 1,
9638 root_task_group
.se
[i
]);
9641 #endif /* CONFIG_FAIR_GROUP_SCHED */
9643 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
9644 #ifdef CONFIG_RT_GROUP_SCHED
9645 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
9646 #ifdef CONFIG_CGROUP_SCHED
9647 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
9648 #elif defined CONFIG_USER_SCHED
9649 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
9650 init_tg_rt_entry(&init_task_group
,
9651 &per_cpu(init_rt_rq_var
, i
),
9652 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
9653 root_task_group
.rt_se
[i
]);
9657 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
9658 rq
->cpu_load
[j
] = 0;
9662 rq
->post_schedule
= 0;
9663 rq
->active_balance
= 0;
9664 rq
->next_balance
= jiffies
;
9668 rq
->migration_thread
= NULL
;
9670 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
9671 INIT_LIST_HEAD(&rq
->migration_queue
);
9672 rq_attach_root(rq
, &def_root_domain
);
9675 atomic_set(&rq
->nr_iowait
, 0);
9678 set_load_weight(&init_task
);
9680 #ifdef CONFIG_PREEMPT_NOTIFIERS
9681 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
9685 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
9688 #ifdef CONFIG_RT_MUTEXES
9689 plist_head_init_raw(&init_task
.pi_waiters
, &init_task
.pi_lock
);
9693 * The boot idle thread does lazy MMU switching as well:
9695 atomic_inc(&init_mm
.mm_count
);
9696 enter_lazy_tlb(&init_mm
, current
);
9699 * Make us the idle thread. Technically, schedule() should not be
9700 * called from this thread, however somewhere below it might be,
9701 * but because we are the idle thread, we just pick up running again
9702 * when this runqueue becomes "idle".
9704 init_idle(current
, smp_processor_id());
9706 calc_load_update
= jiffies
+ LOAD_FREQ
;
9709 * During early bootup we pretend to be a normal task:
9711 current
->sched_class
= &fair_sched_class
;
9713 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9714 zalloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
9717 zalloc_cpumask_var(&nohz
.cpu_mask
, GFP_NOWAIT
);
9718 alloc_cpumask_var(&nohz
.ilb_grp_nohz_mask
, GFP_NOWAIT
);
9720 /* May be allocated at isolcpus cmdline parse time */
9721 if (cpu_isolated_map
== NULL
)
9722 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
9727 scheduler_running
= 1;
9730 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9731 static inline int preempt_count_equals(int preempt_offset
)
9733 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
9735 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
9738 void __might_sleep(char *file
, int line
, int preempt_offset
)
9741 static unsigned long prev_jiffy
; /* ratelimiting */
9743 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
9744 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
9746 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
9748 prev_jiffy
= jiffies
;
9751 "BUG: sleeping function called from invalid context at %s:%d\n",
9754 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9755 in_atomic(), irqs_disabled(),
9756 current
->pid
, current
->comm
);
9758 debug_show_held_locks(current
);
9759 if (irqs_disabled())
9760 print_irqtrace_events(current
);
9764 EXPORT_SYMBOL(__might_sleep
);
9767 #ifdef CONFIG_MAGIC_SYSRQ
9768 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
9772 update_rq_clock(rq
);
9773 on_rq
= p
->se
.on_rq
;
9775 deactivate_task(rq
, p
, 0);
9776 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
9778 activate_task(rq
, p
, 0);
9779 resched_task(rq
->curr
);
9783 void normalize_rt_tasks(void)
9785 struct task_struct
*g
, *p
;
9786 unsigned long flags
;
9789 read_lock_irqsave(&tasklist_lock
, flags
);
9790 do_each_thread(g
, p
) {
9792 * Only normalize user tasks:
9797 p
->se
.exec_start
= 0;
9798 #ifdef CONFIG_SCHEDSTATS
9799 p
->se
.wait_start
= 0;
9800 p
->se
.sleep_start
= 0;
9801 p
->se
.block_start
= 0;
9806 * Renice negative nice level userspace
9809 if (TASK_NICE(p
) < 0 && p
->mm
)
9810 set_user_nice(p
, 0);
9814 raw_spin_lock(&p
->pi_lock
);
9815 rq
= __task_rq_lock(p
);
9817 normalize_task(rq
, p
);
9819 __task_rq_unlock(rq
);
9820 raw_spin_unlock(&p
->pi_lock
);
9821 } while_each_thread(g
, p
);
9823 read_unlock_irqrestore(&tasklist_lock
, flags
);
9826 #endif /* CONFIG_MAGIC_SYSRQ */
9830 * These functions are only useful for the IA64 MCA handling.
9832 * They can only be called when the whole system has been
9833 * stopped - every CPU needs to be quiescent, and no scheduling
9834 * activity can take place. Using them for anything else would
9835 * be a serious bug, and as a result, they aren't even visible
9836 * under any other configuration.
9840 * curr_task - return the current task for a given cpu.
9841 * @cpu: the processor in question.
9843 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9845 struct task_struct
*curr_task(int cpu
)
9847 return cpu_curr(cpu
);
9851 * set_curr_task - set the current task for a given cpu.
9852 * @cpu: the processor in question.
9853 * @p: the task pointer to set.
9855 * Description: This function must only be used when non-maskable interrupts
9856 * are serviced on a separate stack. It allows the architecture to switch the
9857 * notion of the current task on a cpu in a non-blocking manner. This function
9858 * must be called with all CPU's synchronized, and interrupts disabled, the
9859 * and caller must save the original value of the current task (see
9860 * curr_task() above) and restore that value before reenabling interrupts and
9861 * re-starting the system.
9863 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9865 void set_curr_task(int cpu
, struct task_struct
*p
)
9872 #ifdef CONFIG_FAIR_GROUP_SCHED
9873 static void free_fair_sched_group(struct task_group
*tg
)
9877 for_each_possible_cpu(i
) {
9879 kfree(tg
->cfs_rq
[i
]);
9889 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9891 struct cfs_rq
*cfs_rq
;
9892 struct sched_entity
*se
;
9896 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9899 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
9903 tg
->shares
= NICE_0_LOAD
;
9905 for_each_possible_cpu(i
) {
9908 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
9909 GFP_KERNEL
, cpu_to_node(i
));
9913 se
= kzalloc_node(sizeof(struct sched_entity
),
9914 GFP_KERNEL
, cpu_to_node(i
));
9918 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
9929 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9931 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
9932 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
9935 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9937 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
9939 #else /* !CONFG_FAIR_GROUP_SCHED */
9940 static inline void free_fair_sched_group(struct task_group
*tg
)
9945 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9950 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9954 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9957 #endif /* CONFIG_FAIR_GROUP_SCHED */
9959 #ifdef CONFIG_RT_GROUP_SCHED
9960 static void free_rt_sched_group(struct task_group
*tg
)
9964 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
9966 for_each_possible_cpu(i
) {
9968 kfree(tg
->rt_rq
[i
]);
9970 kfree(tg
->rt_se
[i
]);
9978 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9980 struct rt_rq
*rt_rq
;
9981 struct sched_rt_entity
*rt_se
;
9985 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9988 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
9992 init_rt_bandwidth(&tg
->rt_bandwidth
,
9993 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
9995 for_each_possible_cpu(i
) {
9998 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
9999 GFP_KERNEL
, cpu_to_node(i
));
10003 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
10004 GFP_KERNEL
, cpu_to_node(i
));
10008 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
10019 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
10021 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
10022 &cpu_rq(cpu
)->leaf_rt_rq_list
);
10025 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
10027 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
10029 #else /* !CONFIG_RT_GROUP_SCHED */
10030 static inline void free_rt_sched_group(struct task_group
*tg
)
10035 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
10040 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
10044 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
10047 #endif /* CONFIG_RT_GROUP_SCHED */
10049 #ifdef CONFIG_GROUP_SCHED
10050 static void free_sched_group(struct task_group
*tg
)
10052 free_fair_sched_group(tg
);
10053 free_rt_sched_group(tg
);
10057 /* allocate runqueue etc for a new task group */
10058 struct task_group
*sched_create_group(struct task_group
*parent
)
10060 struct task_group
*tg
;
10061 unsigned long flags
;
10064 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
10066 return ERR_PTR(-ENOMEM
);
10068 if (!alloc_fair_sched_group(tg
, parent
))
10071 if (!alloc_rt_sched_group(tg
, parent
))
10074 spin_lock_irqsave(&task_group_lock
, flags
);
10075 for_each_possible_cpu(i
) {
10076 register_fair_sched_group(tg
, i
);
10077 register_rt_sched_group(tg
, i
);
10079 list_add_rcu(&tg
->list
, &task_groups
);
10081 WARN_ON(!parent
); /* root should already exist */
10083 tg
->parent
= parent
;
10084 INIT_LIST_HEAD(&tg
->children
);
10085 list_add_rcu(&tg
->siblings
, &parent
->children
);
10086 spin_unlock_irqrestore(&task_group_lock
, flags
);
10091 free_sched_group(tg
);
10092 return ERR_PTR(-ENOMEM
);
10095 /* rcu callback to free various structures associated with a task group */
10096 static void free_sched_group_rcu(struct rcu_head
*rhp
)
10098 /* now it should be safe to free those cfs_rqs */
10099 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
10102 /* Destroy runqueue etc associated with a task group */
10103 void sched_destroy_group(struct task_group
*tg
)
10105 unsigned long flags
;
10108 spin_lock_irqsave(&task_group_lock
, flags
);
10109 for_each_possible_cpu(i
) {
10110 unregister_fair_sched_group(tg
, i
);
10111 unregister_rt_sched_group(tg
, i
);
10113 list_del_rcu(&tg
->list
);
10114 list_del_rcu(&tg
->siblings
);
10115 spin_unlock_irqrestore(&task_group_lock
, flags
);
10117 /* wait for possible concurrent references to cfs_rqs complete */
10118 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
10121 /* change task's runqueue when it moves between groups.
10122 * The caller of this function should have put the task in its new group
10123 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10124 * reflect its new group.
10126 void sched_move_task(struct task_struct
*tsk
)
10128 int on_rq
, running
;
10129 unsigned long flags
;
10132 rq
= task_rq_lock(tsk
, &flags
);
10134 update_rq_clock(rq
);
10136 running
= task_current(rq
, tsk
);
10137 on_rq
= tsk
->se
.on_rq
;
10140 dequeue_task(rq
, tsk
, 0);
10141 if (unlikely(running
))
10142 tsk
->sched_class
->put_prev_task(rq
, tsk
);
10144 set_task_rq(tsk
, task_cpu(tsk
));
10146 #ifdef CONFIG_FAIR_GROUP_SCHED
10147 if (tsk
->sched_class
->moved_group
)
10148 tsk
->sched_class
->moved_group(tsk
, on_rq
);
10151 if (unlikely(running
))
10152 tsk
->sched_class
->set_curr_task(rq
);
10154 enqueue_task(rq
, tsk
, 0);
10156 task_rq_unlock(rq
, &flags
);
10158 #endif /* CONFIG_GROUP_SCHED */
10160 #ifdef CONFIG_FAIR_GROUP_SCHED
10161 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
10163 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
10168 dequeue_entity(cfs_rq
, se
, 0);
10170 se
->load
.weight
= shares
;
10171 se
->load
.inv_weight
= 0;
10174 enqueue_entity(cfs_rq
, se
, 0);
10177 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
10179 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
10180 struct rq
*rq
= cfs_rq
->rq
;
10181 unsigned long flags
;
10183 raw_spin_lock_irqsave(&rq
->lock
, flags
);
10184 __set_se_shares(se
, shares
);
10185 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
10188 static DEFINE_MUTEX(shares_mutex
);
10190 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
10193 unsigned long flags
;
10196 * We can't change the weight of the root cgroup.
10201 if (shares
< MIN_SHARES
)
10202 shares
= MIN_SHARES
;
10203 else if (shares
> MAX_SHARES
)
10204 shares
= MAX_SHARES
;
10206 mutex_lock(&shares_mutex
);
10207 if (tg
->shares
== shares
)
10210 spin_lock_irqsave(&task_group_lock
, flags
);
10211 for_each_possible_cpu(i
)
10212 unregister_fair_sched_group(tg
, i
);
10213 list_del_rcu(&tg
->siblings
);
10214 spin_unlock_irqrestore(&task_group_lock
, flags
);
10216 /* wait for any ongoing reference to this group to finish */
10217 synchronize_sched();
10220 * Now we are free to modify the group's share on each cpu
10221 * w/o tripping rebalance_share or load_balance_fair.
10223 tg
->shares
= shares
;
10224 for_each_possible_cpu(i
) {
10226 * force a rebalance
10228 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
10229 set_se_shares(tg
->se
[i
], shares
);
10233 * Enable load balance activity on this group, by inserting it back on
10234 * each cpu's rq->leaf_cfs_rq_list.
10236 spin_lock_irqsave(&task_group_lock
, flags
);
10237 for_each_possible_cpu(i
)
10238 register_fair_sched_group(tg
, i
);
10239 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
10240 spin_unlock_irqrestore(&task_group_lock
, flags
);
10242 mutex_unlock(&shares_mutex
);
10246 unsigned long sched_group_shares(struct task_group
*tg
)
10252 #ifdef CONFIG_RT_GROUP_SCHED
10254 * Ensure that the real time constraints are schedulable.
10256 static DEFINE_MUTEX(rt_constraints_mutex
);
10258 static unsigned long to_ratio(u64 period
, u64 runtime
)
10260 if (runtime
== RUNTIME_INF
)
10263 return div64_u64(runtime
<< 20, period
);
10266 /* Must be called with tasklist_lock held */
10267 static inline int tg_has_rt_tasks(struct task_group
*tg
)
10269 struct task_struct
*g
, *p
;
10271 do_each_thread(g
, p
) {
10272 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
10274 } while_each_thread(g
, p
);
10279 struct rt_schedulable_data
{
10280 struct task_group
*tg
;
10285 static int tg_schedulable(struct task_group
*tg
, void *data
)
10287 struct rt_schedulable_data
*d
= data
;
10288 struct task_group
*child
;
10289 unsigned long total
, sum
= 0;
10290 u64 period
, runtime
;
10292 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10293 runtime
= tg
->rt_bandwidth
.rt_runtime
;
10296 period
= d
->rt_period
;
10297 runtime
= d
->rt_runtime
;
10300 #ifdef CONFIG_USER_SCHED
10301 if (tg
== &root_task_group
) {
10302 period
= global_rt_period();
10303 runtime
= global_rt_runtime();
10308 * Cannot have more runtime than the period.
10310 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10314 * Ensure we don't starve existing RT tasks.
10316 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
10319 total
= to_ratio(period
, runtime
);
10322 * Nobody can have more than the global setting allows.
10324 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
10328 * The sum of our children's runtime should not exceed our own.
10330 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
10331 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
10332 runtime
= child
->rt_bandwidth
.rt_runtime
;
10334 if (child
== d
->tg
) {
10335 period
= d
->rt_period
;
10336 runtime
= d
->rt_runtime
;
10339 sum
+= to_ratio(period
, runtime
);
10348 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
10350 struct rt_schedulable_data data
= {
10352 .rt_period
= period
,
10353 .rt_runtime
= runtime
,
10356 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
10359 static int tg_set_bandwidth(struct task_group
*tg
,
10360 u64 rt_period
, u64 rt_runtime
)
10364 mutex_lock(&rt_constraints_mutex
);
10365 read_lock(&tasklist_lock
);
10366 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
10370 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10371 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
10372 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
10374 for_each_possible_cpu(i
) {
10375 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
10377 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
10378 rt_rq
->rt_runtime
= rt_runtime
;
10379 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
10381 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10383 read_unlock(&tasklist_lock
);
10384 mutex_unlock(&rt_constraints_mutex
);
10389 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
10391 u64 rt_runtime
, rt_period
;
10393 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10394 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
10395 if (rt_runtime_us
< 0)
10396 rt_runtime
= RUNTIME_INF
;
10398 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10401 long sched_group_rt_runtime(struct task_group
*tg
)
10405 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
10408 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
10409 do_div(rt_runtime_us
, NSEC_PER_USEC
);
10410 return rt_runtime_us
;
10413 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
10415 u64 rt_runtime
, rt_period
;
10417 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
10418 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
10420 if (rt_period
== 0)
10423 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10426 long sched_group_rt_period(struct task_group
*tg
)
10430 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10431 do_div(rt_period_us
, NSEC_PER_USEC
);
10432 return rt_period_us
;
10435 static int sched_rt_global_constraints(void)
10437 u64 runtime
, period
;
10440 if (sysctl_sched_rt_period
<= 0)
10443 runtime
= global_rt_runtime();
10444 period
= global_rt_period();
10447 * Sanity check on the sysctl variables.
10449 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10452 mutex_lock(&rt_constraints_mutex
);
10453 read_lock(&tasklist_lock
);
10454 ret
= __rt_schedulable(NULL
, 0, 0);
10455 read_unlock(&tasklist_lock
);
10456 mutex_unlock(&rt_constraints_mutex
);
10461 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
10463 /* Don't accept realtime tasks when there is no way for them to run */
10464 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
10470 #else /* !CONFIG_RT_GROUP_SCHED */
10471 static int sched_rt_global_constraints(void)
10473 unsigned long flags
;
10476 if (sysctl_sched_rt_period
<= 0)
10480 * There's always some RT tasks in the root group
10481 * -- migration, kstopmachine etc..
10483 if (sysctl_sched_rt_runtime
== 0)
10486 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10487 for_each_possible_cpu(i
) {
10488 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
10490 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
10491 rt_rq
->rt_runtime
= global_rt_runtime();
10492 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
10494 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10498 #endif /* CONFIG_RT_GROUP_SCHED */
10500 int sched_rt_handler(struct ctl_table
*table
, int write
,
10501 void __user
*buffer
, size_t *lenp
,
10505 int old_period
, old_runtime
;
10506 static DEFINE_MUTEX(mutex
);
10508 mutex_lock(&mutex
);
10509 old_period
= sysctl_sched_rt_period
;
10510 old_runtime
= sysctl_sched_rt_runtime
;
10512 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
10514 if (!ret
&& write
) {
10515 ret
= sched_rt_global_constraints();
10517 sysctl_sched_rt_period
= old_period
;
10518 sysctl_sched_rt_runtime
= old_runtime
;
10520 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
10521 def_rt_bandwidth
.rt_period
=
10522 ns_to_ktime(global_rt_period());
10525 mutex_unlock(&mutex
);
10530 #ifdef CONFIG_CGROUP_SCHED
10532 /* return corresponding task_group object of a cgroup */
10533 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
10535 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
10536 struct task_group
, css
);
10539 static struct cgroup_subsys_state
*
10540 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10542 struct task_group
*tg
, *parent
;
10544 if (!cgrp
->parent
) {
10545 /* This is early initialization for the top cgroup */
10546 return &init_task_group
.css
;
10549 parent
= cgroup_tg(cgrp
->parent
);
10550 tg
= sched_create_group(parent
);
10552 return ERR_PTR(-ENOMEM
);
10558 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10560 struct task_group
*tg
= cgroup_tg(cgrp
);
10562 sched_destroy_group(tg
);
10566 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
10568 #ifdef CONFIG_RT_GROUP_SCHED
10569 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
10572 /* We don't support RT-tasks being in separate groups */
10573 if (tsk
->sched_class
!= &fair_sched_class
)
10580 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10581 struct task_struct
*tsk
, bool threadgroup
)
10583 int retval
= cpu_cgroup_can_attach_task(cgrp
, tsk
);
10587 struct task_struct
*c
;
10589 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
10590 retval
= cpu_cgroup_can_attach_task(cgrp
, c
);
10602 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10603 struct cgroup
*old_cont
, struct task_struct
*tsk
,
10606 sched_move_task(tsk
);
10608 struct task_struct
*c
;
10610 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
10611 sched_move_task(c
);
10617 #ifdef CONFIG_FAIR_GROUP_SCHED
10618 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
10621 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
10624 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
10626 struct task_group
*tg
= cgroup_tg(cgrp
);
10628 return (u64
) tg
->shares
;
10630 #endif /* CONFIG_FAIR_GROUP_SCHED */
10632 #ifdef CONFIG_RT_GROUP_SCHED
10633 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
10636 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
10639 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10641 return sched_group_rt_runtime(cgroup_tg(cgrp
));
10644 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
10647 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
10650 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
10652 return sched_group_rt_period(cgroup_tg(cgrp
));
10654 #endif /* CONFIG_RT_GROUP_SCHED */
10656 static struct cftype cpu_files
[] = {
10657 #ifdef CONFIG_FAIR_GROUP_SCHED
10660 .read_u64
= cpu_shares_read_u64
,
10661 .write_u64
= cpu_shares_write_u64
,
10664 #ifdef CONFIG_RT_GROUP_SCHED
10666 .name
= "rt_runtime_us",
10667 .read_s64
= cpu_rt_runtime_read
,
10668 .write_s64
= cpu_rt_runtime_write
,
10671 .name
= "rt_period_us",
10672 .read_u64
= cpu_rt_period_read_uint
,
10673 .write_u64
= cpu_rt_period_write_uint
,
10678 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
10680 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
10683 struct cgroup_subsys cpu_cgroup_subsys
= {
10685 .create
= cpu_cgroup_create
,
10686 .destroy
= cpu_cgroup_destroy
,
10687 .can_attach
= cpu_cgroup_can_attach
,
10688 .attach
= cpu_cgroup_attach
,
10689 .populate
= cpu_cgroup_populate
,
10690 .subsys_id
= cpu_cgroup_subsys_id
,
10694 #endif /* CONFIG_CGROUP_SCHED */
10696 #ifdef CONFIG_CGROUP_CPUACCT
10699 * CPU accounting code for task groups.
10701 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10702 * (balbir@in.ibm.com).
10705 /* track cpu usage of a group of tasks and its child groups */
10707 struct cgroup_subsys_state css
;
10708 /* cpuusage holds pointer to a u64-type object on every cpu */
10710 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
10711 struct cpuacct
*parent
;
10714 struct cgroup_subsys cpuacct_subsys
;
10716 /* return cpu accounting group corresponding to this container */
10717 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
10719 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
10720 struct cpuacct
, css
);
10723 /* return cpu accounting group to which this task belongs */
10724 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
10726 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
10727 struct cpuacct
, css
);
10730 /* create a new cpu accounting group */
10731 static struct cgroup_subsys_state
*cpuacct_create(
10732 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10734 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
10740 ca
->cpuusage
= alloc_percpu(u64
);
10744 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10745 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
10746 goto out_free_counters
;
10749 ca
->parent
= cgroup_ca(cgrp
->parent
);
10755 percpu_counter_destroy(&ca
->cpustat
[i
]);
10756 free_percpu(ca
->cpuusage
);
10760 return ERR_PTR(-ENOMEM
);
10763 /* destroy an existing cpu accounting group */
10765 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10767 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10770 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10771 percpu_counter_destroy(&ca
->cpustat
[i
]);
10772 free_percpu(ca
->cpuusage
);
10776 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
10778 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10781 #ifndef CONFIG_64BIT
10783 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10785 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
10787 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10795 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
10797 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10799 #ifndef CONFIG_64BIT
10801 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10803 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
10805 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10811 /* return total cpu usage (in nanoseconds) of a group */
10812 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10814 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10815 u64 totalcpuusage
= 0;
10818 for_each_present_cpu(i
)
10819 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
10821 return totalcpuusage
;
10824 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
10827 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10836 for_each_present_cpu(i
)
10837 cpuacct_cpuusage_write(ca
, i
, 0);
10843 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
10844 struct seq_file
*m
)
10846 struct cpuacct
*ca
= cgroup_ca(cgroup
);
10850 for_each_present_cpu(i
) {
10851 percpu
= cpuacct_cpuusage_read(ca
, i
);
10852 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
10854 seq_printf(m
, "\n");
10858 static const char *cpuacct_stat_desc
[] = {
10859 [CPUACCT_STAT_USER
] = "user",
10860 [CPUACCT_STAT_SYSTEM
] = "system",
10863 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
10864 struct cgroup_map_cb
*cb
)
10866 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10869 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
10870 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
10871 val
= cputime64_to_clock_t(val
);
10872 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
10877 static struct cftype files
[] = {
10880 .read_u64
= cpuusage_read
,
10881 .write_u64
= cpuusage_write
,
10884 .name
= "usage_percpu",
10885 .read_seq_string
= cpuacct_percpu_seq_read
,
10889 .read_map
= cpuacct_stats_show
,
10893 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10895 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
10899 * charge this task's execution time to its accounting group.
10901 * called with rq->lock held.
10903 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
10905 struct cpuacct
*ca
;
10908 if (unlikely(!cpuacct_subsys
.active
))
10911 cpu
= task_cpu(tsk
);
10917 for (; ca
; ca
= ca
->parent
) {
10918 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10919 *cpuusage
+= cputime
;
10926 * Charge the system/user time to the task's accounting group.
10928 static void cpuacct_update_stats(struct task_struct
*tsk
,
10929 enum cpuacct_stat_index idx
, cputime_t val
)
10931 struct cpuacct
*ca
;
10933 if (unlikely(!cpuacct_subsys
.active
))
10940 percpu_counter_add(&ca
->cpustat
[idx
], val
);
10946 struct cgroup_subsys cpuacct_subsys
= {
10948 .create
= cpuacct_create
,
10949 .destroy
= cpuacct_destroy
,
10950 .populate
= cpuacct_populate
,
10951 .subsys_id
= cpuacct_subsys_id
,
10953 #endif /* CONFIG_CGROUP_CPUACCT */
10957 int rcu_expedited_torture_stats(char *page
)
10961 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
10963 void synchronize_sched_expedited(void)
10966 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
10968 #else /* #ifndef CONFIG_SMP */
10970 static DEFINE_PER_CPU(struct migration_req
, rcu_migration_req
);
10971 static DEFINE_MUTEX(rcu_sched_expedited_mutex
);
10973 #define RCU_EXPEDITED_STATE_POST -2
10974 #define RCU_EXPEDITED_STATE_IDLE -1
10976 static int rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
10978 int rcu_expedited_torture_stats(char *page
)
10983 cnt
+= sprintf(&page
[cnt
], "state: %d /", rcu_expedited_state
);
10984 for_each_online_cpu(cpu
) {
10985 cnt
+= sprintf(&page
[cnt
], " %d:%d",
10986 cpu
, per_cpu(rcu_migration_req
, cpu
).dest_cpu
);
10988 cnt
+= sprintf(&page
[cnt
], "\n");
10991 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
10993 static long synchronize_sched_expedited_count
;
10996 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10997 * approach to force grace period to end quickly. This consumes
10998 * significant time on all CPUs, and is thus not recommended for
10999 * any sort of common-case code.
11001 * Note that it is illegal to call this function while holding any
11002 * lock that is acquired by a CPU-hotplug notifier. Failing to
11003 * observe this restriction will result in deadlock.
11005 void synchronize_sched_expedited(void)
11008 unsigned long flags
;
11009 bool need_full_sync
= 0;
11011 struct migration_req
*req
;
11015 smp_mb(); /* ensure prior mod happens before capturing snap. */
11016 snap
= ACCESS_ONCE(synchronize_sched_expedited_count
) + 1;
11018 while (!mutex_trylock(&rcu_sched_expedited_mutex
)) {
11020 if (trycount
++ < 10)
11021 udelay(trycount
* num_online_cpus());
11023 synchronize_sched();
11026 if (ACCESS_ONCE(synchronize_sched_expedited_count
) - snap
> 0) {
11027 smp_mb(); /* ensure test happens before caller kfree */
11032 rcu_expedited_state
= RCU_EXPEDITED_STATE_POST
;
11033 for_each_online_cpu(cpu
) {
11035 req
= &per_cpu(rcu_migration_req
, cpu
);
11036 init_completion(&req
->done
);
11038 req
->dest_cpu
= RCU_MIGRATION_NEED_QS
;
11039 raw_spin_lock_irqsave(&rq
->lock
, flags
);
11040 list_add(&req
->list
, &rq
->migration_queue
);
11041 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
11042 wake_up_process(rq
->migration_thread
);
11044 for_each_online_cpu(cpu
) {
11045 rcu_expedited_state
= cpu
;
11046 req
= &per_cpu(rcu_migration_req
, cpu
);
11048 wait_for_completion(&req
->done
);
11049 raw_spin_lock_irqsave(&rq
->lock
, flags
);
11050 if (unlikely(req
->dest_cpu
== RCU_MIGRATION_MUST_SYNC
))
11051 need_full_sync
= 1;
11052 req
->dest_cpu
= RCU_MIGRATION_IDLE
;
11053 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
11055 rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
11056 synchronize_sched_expedited_count
++;
11057 mutex_unlock(&rcu_sched_expedited_mutex
);
11059 if (need_full_sync
)
11060 synchronize_sched();
11062 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
11064 #endif /* #else #ifndef CONFIG_SMP */