4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/reciprocal_div.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/bootmem.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
75 #include <trace/sched.h>
78 #include <asm/irq_regs.h>
80 #include "sched_cpupri.h"
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
87 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
96 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
101 * Helpers for converting nanosecond timing to jiffy resolution
103 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * single value that denotes runtime == period, ie unlimited time.
119 #define RUNTIME_INF ((u64)~0ULL)
121 DEFINE_TRACE(sched_wait_task
);
122 DEFINE_TRACE(sched_wakeup
);
123 DEFINE_TRACE(sched_wakeup_new
);
124 DEFINE_TRACE(sched_switch
);
125 DEFINE_TRACE(sched_migrate_task
);
129 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
132 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
133 * Since cpu_power is a 'constant', we can use a reciprocal divide.
135 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
137 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
141 * Each time a sched group cpu_power is changed,
142 * we must compute its reciprocal value
144 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
146 sg
->__cpu_power
+= val
;
147 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
151 static inline int rt_policy(int policy
)
153 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
158 static inline int task_has_rt_policy(struct task_struct
*p
)
160 return rt_policy(p
->policy
);
164 * This is the priority-queue data structure of the RT scheduling class:
166 struct rt_prio_array
{
167 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
168 struct list_head queue
[MAX_RT_PRIO
];
171 struct rt_bandwidth
{
172 /* nests inside the rq lock: */
173 spinlock_t rt_runtime_lock
;
176 struct hrtimer rt_period_timer
;
179 static struct rt_bandwidth def_rt_bandwidth
;
181 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
183 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
185 struct rt_bandwidth
*rt_b
=
186 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
192 now
= hrtimer_cb_get_time(timer
);
193 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
198 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
201 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
205 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
207 rt_b
->rt_period
= ns_to_ktime(period
);
208 rt_b
->rt_runtime
= runtime
;
210 spin_lock_init(&rt_b
->rt_runtime_lock
);
212 hrtimer_init(&rt_b
->rt_period_timer
,
213 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
214 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
217 static inline int rt_bandwidth_enabled(void)
219 return sysctl_sched_rt_runtime
>= 0;
222 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
226 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
229 if (hrtimer_active(&rt_b
->rt_period_timer
))
232 spin_lock(&rt_b
->rt_runtime_lock
);
234 if (hrtimer_active(&rt_b
->rt_period_timer
))
237 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
238 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
239 hrtimer_start_expires(&rt_b
->rt_period_timer
,
242 spin_unlock(&rt_b
->rt_runtime_lock
);
245 #ifdef CONFIG_RT_GROUP_SCHED
246 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
248 hrtimer_cancel(&rt_b
->rt_period_timer
);
253 * sched_domains_mutex serializes calls to arch_init_sched_domains,
254 * detach_destroy_domains and partition_sched_domains.
256 static DEFINE_MUTEX(sched_domains_mutex
);
258 #ifdef CONFIG_GROUP_SCHED
260 #include <linux/cgroup.h>
264 static LIST_HEAD(task_groups
);
266 /* task group related information */
268 #ifdef CONFIG_CGROUP_SCHED
269 struct cgroup_subsys_state css
;
272 #ifdef CONFIG_USER_SCHED
276 #ifdef CONFIG_FAIR_GROUP_SCHED
277 /* schedulable entities of this group on each cpu */
278 struct sched_entity
**se
;
279 /* runqueue "owned" by this group on each cpu */
280 struct cfs_rq
**cfs_rq
;
281 unsigned long shares
;
284 #ifdef CONFIG_RT_GROUP_SCHED
285 struct sched_rt_entity
**rt_se
;
286 struct rt_rq
**rt_rq
;
288 struct rt_bandwidth rt_bandwidth
;
292 struct list_head list
;
294 struct task_group
*parent
;
295 struct list_head siblings
;
296 struct list_head children
;
299 #ifdef CONFIG_USER_SCHED
301 /* Helper function to pass uid information to create_sched_user() */
302 void set_tg_uid(struct user_struct
*user
)
304 user
->tg
->uid
= user
->uid
;
309 * Every UID task group (including init_task_group aka UID-0) will
310 * be a child to this group.
312 struct task_group root_task_group
;
314 #ifdef CONFIG_FAIR_GROUP_SCHED
315 /* Default task group's sched entity on each cpu */
316 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
317 /* Default task group's cfs_rq on each cpu */
318 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
319 #endif /* CONFIG_FAIR_GROUP_SCHED */
321 #ifdef CONFIG_RT_GROUP_SCHED
322 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
323 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
324 #endif /* CONFIG_RT_GROUP_SCHED */
325 #else /* !CONFIG_USER_SCHED */
326 #define root_task_group init_task_group
327 #endif /* CONFIG_USER_SCHED */
329 /* task_group_lock serializes add/remove of task groups and also changes to
330 * a task group's cpu shares.
332 static DEFINE_SPINLOCK(task_group_lock
);
335 static int root_task_group_empty(void)
337 return list_empty(&root_task_group
.children
);
341 #ifdef CONFIG_FAIR_GROUP_SCHED
342 #ifdef CONFIG_USER_SCHED
343 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
344 #else /* !CONFIG_USER_SCHED */
345 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
346 #endif /* CONFIG_USER_SCHED */
349 * A weight of 0 or 1 can cause arithmetics problems.
350 * A weight of a cfs_rq is the sum of weights of which entities
351 * are queued on this cfs_rq, so a weight of a entity should not be
352 * too large, so as the shares value of a task group.
353 * (The default weight is 1024 - so there's no practical
354 * limitation from this.)
357 #define MAX_SHARES (1UL << 18)
359 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
362 /* Default task group.
363 * Every task in system belong to this group at bootup.
365 struct task_group init_task_group
;
367 /* return group to which a task belongs */
368 static inline struct task_group
*task_group(struct task_struct
*p
)
370 struct task_group
*tg
;
372 #ifdef CONFIG_USER_SCHED
374 tg
= __task_cred(p
)->user
->tg
;
376 #elif defined(CONFIG_CGROUP_SCHED)
377 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
378 struct task_group
, css
);
380 tg
= &init_task_group
;
385 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
386 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
388 #ifdef CONFIG_FAIR_GROUP_SCHED
389 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
390 p
->se
.parent
= task_group(p
)->se
[cpu
];
393 #ifdef CONFIG_RT_GROUP_SCHED
394 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
395 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
402 static int root_task_group_empty(void)
408 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
409 static inline struct task_group
*task_group(struct task_struct
*p
)
414 #endif /* CONFIG_GROUP_SCHED */
416 /* CFS-related fields in a runqueue */
418 struct load_weight load
;
419 unsigned long nr_running
;
424 struct rb_root tasks_timeline
;
425 struct rb_node
*rb_leftmost
;
427 struct list_head tasks
;
428 struct list_head
*balance_iterator
;
431 * 'curr' points to currently running entity on this cfs_rq.
432 * It is set to NULL otherwise (i.e when none are currently running).
434 struct sched_entity
*curr
, *next
, *last
;
436 unsigned int nr_spread_over
;
438 #ifdef CONFIG_FAIR_GROUP_SCHED
439 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
442 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
443 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
444 * (like users, containers etc.)
446 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
447 * list is used during load balance.
449 struct list_head leaf_cfs_rq_list
;
450 struct task_group
*tg
; /* group that "owns" this runqueue */
454 * the part of load.weight contributed by tasks
456 unsigned long task_weight
;
459 * h_load = weight * f(tg)
461 * Where f(tg) is the recursive weight fraction assigned to
464 unsigned long h_load
;
467 * this cpu's part of tg->shares
469 unsigned long shares
;
472 * load.weight at the time we set shares
474 unsigned long rq_weight
;
479 /* Real-Time classes' related field in a runqueue: */
481 struct rt_prio_array active
;
482 unsigned long rt_nr_running
;
483 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
485 int curr
; /* highest queued rt task prio */
487 int next
; /* next highest */
492 unsigned long rt_nr_migratory
;
494 struct plist_head pushable_tasks
;
499 /* Nests inside the rq lock: */
500 spinlock_t rt_runtime_lock
;
502 #ifdef CONFIG_RT_GROUP_SCHED
503 unsigned long rt_nr_boosted
;
506 struct list_head leaf_rt_rq_list
;
507 struct task_group
*tg
;
508 struct sched_rt_entity
*rt_se
;
515 * We add the notion of a root-domain which will be used to define per-domain
516 * variables. Each exclusive cpuset essentially defines an island domain by
517 * fully partitioning the member cpus from any other cpuset. Whenever a new
518 * exclusive cpuset is created, we also create and attach a new root-domain
525 cpumask_var_t online
;
528 * The "RT overload" flag: it gets set if a CPU has more than
529 * one runnable RT task.
531 cpumask_var_t rto_mask
;
534 struct cpupri cpupri
;
536 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
538 * Preferred wake up cpu nominated by sched_mc balance that will be
539 * used when most cpus are idle in the system indicating overall very
540 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
542 unsigned int sched_mc_preferred_wakeup_cpu
;
547 * By default the system creates a single root-domain with all cpus as
548 * members (mimicking the global state we have today).
550 static struct root_domain def_root_domain
;
555 * This is the main, per-CPU runqueue data structure.
557 * Locking rule: those places that want to lock multiple runqueues
558 * (such as the load balancing or the thread migration code), lock
559 * acquire operations must be ordered by ascending &runqueue.
566 * nr_running and cpu_load should be in the same cacheline because
567 * remote CPUs use both these fields when doing load calculation.
569 unsigned long nr_running
;
570 #define CPU_LOAD_IDX_MAX 5
571 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
573 unsigned long last_tick_seen
;
574 unsigned char in_nohz_recently
;
576 /* capture load from *all* tasks on this cpu: */
577 struct load_weight load
;
578 unsigned long nr_load_updates
;
584 #ifdef CONFIG_FAIR_GROUP_SCHED
585 /* list of leaf cfs_rq on this cpu: */
586 struct list_head leaf_cfs_rq_list
;
588 #ifdef CONFIG_RT_GROUP_SCHED
589 struct list_head leaf_rt_rq_list
;
593 * This is part of a global counter where only the total sum
594 * over all CPUs matters. A task can increase this counter on
595 * one CPU and if it got migrated afterwards it may decrease
596 * it on another CPU. Always updated under the runqueue lock:
598 unsigned long nr_uninterruptible
;
600 struct task_struct
*curr
, *idle
;
601 unsigned long next_balance
;
602 struct mm_struct
*prev_mm
;
609 struct root_domain
*rd
;
610 struct sched_domain
*sd
;
612 unsigned char idle_at_tick
;
613 /* For active balancing */
616 /* cpu of this runqueue: */
620 unsigned long avg_load_per_task
;
622 struct task_struct
*migration_thread
;
623 struct list_head migration_queue
;
626 #ifdef CONFIG_SCHED_HRTICK
628 int hrtick_csd_pending
;
629 struct call_single_data hrtick_csd
;
631 struct hrtimer hrtick_timer
;
634 #ifdef CONFIG_SCHEDSTATS
636 struct sched_info rq_sched_info
;
637 unsigned long long rq_cpu_time
;
638 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
640 /* sys_sched_yield() stats */
641 unsigned int yld_count
;
643 /* schedule() stats */
644 unsigned int sched_switch
;
645 unsigned int sched_count
;
646 unsigned int sched_goidle
;
648 /* try_to_wake_up() stats */
649 unsigned int ttwu_count
;
650 unsigned int ttwu_local
;
653 unsigned int bkl_count
;
657 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
659 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int sync
)
661 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, sync
);
664 static inline int cpu_of(struct rq
*rq
)
674 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
675 * See detach_destroy_domains: synchronize_sched for details.
677 * The domain tree of any CPU may only be accessed from within
678 * preempt-disabled sections.
680 #define for_each_domain(cpu, __sd) \
681 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
683 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
684 #define this_rq() (&__get_cpu_var(runqueues))
685 #define task_rq(p) cpu_rq(task_cpu(p))
686 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
688 static inline void update_rq_clock(struct rq
*rq
)
690 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
694 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
696 #ifdef CONFIG_SCHED_DEBUG
697 # define const_debug __read_mostly
699 # define const_debug static const
705 * Returns true if the current cpu runqueue is locked.
706 * This interface allows printk to be called with the runqueue lock
707 * held and know whether or not it is OK to wake up the klogd.
709 int runqueue_is_locked(void)
712 struct rq
*rq
= cpu_rq(cpu
);
715 ret
= spin_is_locked(&rq
->lock
);
721 * Debugging: various feature bits
724 #define SCHED_FEAT(name, enabled) \
725 __SCHED_FEAT_##name ,
728 #include "sched_features.h"
733 #define SCHED_FEAT(name, enabled) \
734 (1UL << __SCHED_FEAT_##name) * enabled |
736 const_debug
unsigned int sysctl_sched_features
=
737 #include "sched_features.h"
742 #ifdef CONFIG_SCHED_DEBUG
743 #define SCHED_FEAT(name, enabled) \
746 static __read_mostly
char *sched_feat_names
[] = {
747 #include "sched_features.h"
753 static int sched_feat_show(struct seq_file
*m
, void *v
)
757 for (i
= 0; sched_feat_names
[i
]; i
++) {
758 if (!(sysctl_sched_features
& (1UL << i
)))
760 seq_printf(m
, "%s ", sched_feat_names
[i
]);
768 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
769 size_t cnt
, loff_t
*ppos
)
779 if (copy_from_user(&buf
, ubuf
, cnt
))
784 if (strncmp(buf
, "NO_", 3) == 0) {
789 for (i
= 0; sched_feat_names
[i
]; i
++) {
790 int len
= strlen(sched_feat_names
[i
]);
792 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
794 sysctl_sched_features
&= ~(1UL << i
);
796 sysctl_sched_features
|= (1UL << i
);
801 if (!sched_feat_names
[i
])
809 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
811 return single_open(filp
, sched_feat_show
, NULL
);
814 static struct file_operations sched_feat_fops
= {
815 .open
= sched_feat_open
,
816 .write
= sched_feat_write
,
819 .release
= single_release
,
822 static __init
int sched_init_debug(void)
824 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
829 late_initcall(sched_init_debug
);
833 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
836 * Number of tasks to iterate in a single balance run.
837 * Limited because this is done with IRQs disabled.
839 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
842 * ratelimit for updating the group shares.
845 unsigned int sysctl_sched_shares_ratelimit
= 250000;
848 * Inject some fuzzyness into changing the per-cpu group shares
849 * this avoids remote rq-locks at the expense of fairness.
852 unsigned int sysctl_sched_shares_thresh
= 4;
855 * period over which we measure -rt task cpu usage in us.
858 unsigned int sysctl_sched_rt_period
= 1000000;
860 static __read_mostly
int scheduler_running
;
863 * part of the period that we allow rt tasks to run in us.
866 int sysctl_sched_rt_runtime
= 950000;
868 static inline u64
global_rt_period(void)
870 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
873 static inline u64
global_rt_runtime(void)
875 if (sysctl_sched_rt_runtime
< 0)
878 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
881 #ifndef prepare_arch_switch
882 # define prepare_arch_switch(next) do { } while (0)
884 #ifndef finish_arch_switch
885 # define finish_arch_switch(prev) do { } while (0)
888 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
890 return rq
->curr
== p
;
893 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
894 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
896 return task_current(rq
, p
);
899 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
903 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
905 #ifdef CONFIG_DEBUG_SPINLOCK
906 /* this is a valid case when another task releases the spinlock */
907 rq
->lock
.owner
= current
;
910 * If we are tracking spinlock dependencies then we have to
911 * fix up the runqueue lock - which gets 'carried over' from
914 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
916 spin_unlock_irq(&rq
->lock
);
919 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
920 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
925 return task_current(rq
, p
);
929 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
933 * We can optimise this out completely for !SMP, because the
934 * SMP rebalancing from interrupt is the only thing that cares
939 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
940 spin_unlock_irq(&rq
->lock
);
942 spin_unlock(&rq
->lock
);
946 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
950 * After ->oncpu is cleared, the task can be moved to a different CPU.
951 * We must ensure this doesn't happen until the switch is completely
957 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
961 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
964 * __task_rq_lock - lock the runqueue a given task resides on.
965 * Must be called interrupts disabled.
967 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
971 struct rq
*rq
= task_rq(p
);
972 spin_lock(&rq
->lock
);
973 if (likely(rq
== task_rq(p
)))
975 spin_unlock(&rq
->lock
);
980 * task_rq_lock - lock the runqueue a given task resides on and disable
981 * interrupts. Note the ordering: we can safely lookup the task_rq without
982 * explicitly disabling preemption.
984 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
990 local_irq_save(*flags
);
992 spin_lock(&rq
->lock
);
993 if (likely(rq
== task_rq(p
)))
995 spin_unlock_irqrestore(&rq
->lock
, *flags
);
999 void task_rq_unlock_wait(struct task_struct
*p
)
1001 struct rq
*rq
= task_rq(p
);
1003 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1004 spin_unlock_wait(&rq
->lock
);
1007 static void __task_rq_unlock(struct rq
*rq
)
1008 __releases(rq
->lock
)
1010 spin_unlock(&rq
->lock
);
1013 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
1014 __releases(rq
->lock
)
1016 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1020 * this_rq_lock - lock this runqueue and disable interrupts.
1022 static struct rq
*this_rq_lock(void)
1023 __acquires(rq
->lock
)
1027 local_irq_disable();
1029 spin_lock(&rq
->lock
);
1034 #ifdef CONFIG_SCHED_HRTICK
1036 * Use HR-timers to deliver accurate preemption points.
1038 * Its all a bit involved since we cannot program an hrt while holding the
1039 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1042 * When we get rescheduled we reprogram the hrtick_timer outside of the
1048 * - enabled by features
1049 * - hrtimer is actually high res
1051 static inline int hrtick_enabled(struct rq
*rq
)
1053 if (!sched_feat(HRTICK
))
1055 if (!cpu_active(cpu_of(rq
)))
1057 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1060 static void hrtick_clear(struct rq
*rq
)
1062 if (hrtimer_active(&rq
->hrtick_timer
))
1063 hrtimer_cancel(&rq
->hrtick_timer
);
1067 * High-resolution timer tick.
1068 * Runs from hardirq context with interrupts disabled.
1070 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1072 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1074 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1076 spin_lock(&rq
->lock
);
1077 update_rq_clock(rq
);
1078 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1079 spin_unlock(&rq
->lock
);
1081 return HRTIMER_NORESTART
;
1086 * called from hardirq (IPI) context
1088 static void __hrtick_start(void *arg
)
1090 struct rq
*rq
= arg
;
1092 spin_lock(&rq
->lock
);
1093 hrtimer_restart(&rq
->hrtick_timer
);
1094 rq
->hrtick_csd_pending
= 0;
1095 spin_unlock(&rq
->lock
);
1099 * Called to set the hrtick timer state.
1101 * called with rq->lock held and irqs disabled
1103 static void hrtick_start(struct rq
*rq
, u64 delay
)
1105 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1106 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1108 hrtimer_set_expires(timer
, time
);
1110 if (rq
== this_rq()) {
1111 hrtimer_restart(timer
);
1112 } else if (!rq
->hrtick_csd_pending
) {
1113 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
);
1114 rq
->hrtick_csd_pending
= 1;
1119 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1121 int cpu
= (int)(long)hcpu
;
1124 case CPU_UP_CANCELED
:
1125 case CPU_UP_CANCELED_FROZEN
:
1126 case CPU_DOWN_PREPARE
:
1127 case CPU_DOWN_PREPARE_FROZEN
:
1129 case CPU_DEAD_FROZEN
:
1130 hrtick_clear(cpu_rq(cpu
));
1137 static __init
void init_hrtick(void)
1139 hotcpu_notifier(hotplug_hrtick
, 0);
1143 * Called to set the hrtick timer state.
1145 * called with rq->lock held and irqs disabled
1147 static void hrtick_start(struct rq
*rq
, u64 delay
)
1149 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
), HRTIMER_MODE_REL
);
1152 static inline void init_hrtick(void)
1155 #endif /* CONFIG_SMP */
1157 static void init_rq_hrtick(struct rq
*rq
)
1160 rq
->hrtick_csd_pending
= 0;
1162 rq
->hrtick_csd
.flags
= 0;
1163 rq
->hrtick_csd
.func
= __hrtick_start
;
1164 rq
->hrtick_csd
.info
= rq
;
1167 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1168 rq
->hrtick_timer
.function
= hrtick
;
1170 #else /* CONFIG_SCHED_HRTICK */
1171 static inline void hrtick_clear(struct rq
*rq
)
1175 static inline void init_rq_hrtick(struct rq
*rq
)
1179 static inline void init_hrtick(void)
1182 #endif /* CONFIG_SCHED_HRTICK */
1185 * resched_task - mark a task 'to be rescheduled now'.
1187 * On UP this means the setting of the need_resched flag, on SMP it
1188 * might also involve a cross-CPU call to trigger the scheduler on
1193 #ifndef tsk_is_polling
1194 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1197 static void resched_task(struct task_struct
*p
)
1201 assert_spin_locked(&task_rq(p
)->lock
);
1203 if (test_tsk_need_resched(p
))
1206 set_tsk_need_resched(p
);
1209 if (cpu
== smp_processor_id())
1212 /* NEED_RESCHED must be visible before we test polling */
1214 if (!tsk_is_polling(p
))
1215 smp_send_reschedule(cpu
);
1218 static void resched_cpu(int cpu
)
1220 struct rq
*rq
= cpu_rq(cpu
);
1221 unsigned long flags
;
1223 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1225 resched_task(cpu_curr(cpu
));
1226 spin_unlock_irqrestore(&rq
->lock
, flags
);
1231 * When add_timer_on() enqueues a timer into the timer wheel of an
1232 * idle CPU then this timer might expire before the next timer event
1233 * which is scheduled to wake up that CPU. In case of a completely
1234 * idle system the next event might even be infinite time into the
1235 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1236 * leaves the inner idle loop so the newly added timer is taken into
1237 * account when the CPU goes back to idle and evaluates the timer
1238 * wheel for the next timer event.
1240 void wake_up_idle_cpu(int cpu
)
1242 struct rq
*rq
= cpu_rq(cpu
);
1244 if (cpu
== smp_processor_id())
1248 * This is safe, as this function is called with the timer
1249 * wheel base lock of (cpu) held. When the CPU is on the way
1250 * to idle and has not yet set rq->curr to idle then it will
1251 * be serialized on the timer wheel base lock and take the new
1252 * timer into account automatically.
1254 if (rq
->curr
!= rq
->idle
)
1258 * We can set TIF_RESCHED on the idle task of the other CPU
1259 * lockless. The worst case is that the other CPU runs the
1260 * idle task through an additional NOOP schedule()
1262 set_tsk_need_resched(rq
->idle
);
1264 /* NEED_RESCHED must be visible before we test polling */
1266 if (!tsk_is_polling(rq
->idle
))
1267 smp_send_reschedule(cpu
);
1269 #endif /* CONFIG_NO_HZ */
1271 #else /* !CONFIG_SMP */
1272 static void resched_task(struct task_struct
*p
)
1274 assert_spin_locked(&task_rq(p
)->lock
);
1275 set_tsk_need_resched(p
);
1277 #endif /* CONFIG_SMP */
1279 #if BITS_PER_LONG == 32
1280 # define WMULT_CONST (~0UL)
1282 # define WMULT_CONST (1UL << 32)
1285 #define WMULT_SHIFT 32
1288 * Shift right and round:
1290 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1293 * delta *= weight / lw
1295 static unsigned long
1296 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1297 struct load_weight
*lw
)
1301 if (!lw
->inv_weight
) {
1302 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1305 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1309 tmp
= (u64
)delta_exec
* weight
;
1311 * Check whether we'd overflow the 64-bit multiplication:
1313 if (unlikely(tmp
> WMULT_CONST
))
1314 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1317 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1319 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1322 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1328 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1335 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1336 * of tasks with abnormal "nice" values across CPUs the contribution that
1337 * each task makes to its run queue's load is weighted according to its
1338 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1339 * scaled version of the new time slice allocation that they receive on time
1343 #define WEIGHT_IDLEPRIO 3
1344 #define WMULT_IDLEPRIO 1431655765
1347 * Nice levels are multiplicative, with a gentle 10% change for every
1348 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1349 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1350 * that remained on nice 0.
1352 * The "10% effect" is relative and cumulative: from _any_ nice level,
1353 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1354 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1355 * If a task goes up by ~10% and another task goes down by ~10% then
1356 * the relative distance between them is ~25%.)
1358 static const int prio_to_weight
[40] = {
1359 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1360 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1361 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1362 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1363 /* 0 */ 1024, 820, 655, 526, 423,
1364 /* 5 */ 335, 272, 215, 172, 137,
1365 /* 10 */ 110, 87, 70, 56, 45,
1366 /* 15 */ 36, 29, 23, 18, 15,
1370 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1372 * In cases where the weight does not change often, we can use the
1373 * precalculated inverse to speed up arithmetics by turning divisions
1374 * into multiplications:
1376 static const u32 prio_to_wmult
[40] = {
1377 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1378 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1379 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1380 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1381 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1382 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1383 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1384 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1387 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1390 * runqueue iterator, to support SMP load-balancing between different
1391 * scheduling classes, without having to expose their internal data
1392 * structures to the load-balancing proper:
1394 struct rq_iterator
{
1396 struct task_struct
*(*start
)(void *);
1397 struct task_struct
*(*next
)(void *);
1401 static unsigned long
1402 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1403 unsigned long max_load_move
, struct sched_domain
*sd
,
1404 enum cpu_idle_type idle
, int *all_pinned
,
1405 int *this_best_prio
, struct rq_iterator
*iterator
);
1408 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1409 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1410 struct rq_iterator
*iterator
);
1413 #ifdef CONFIG_CGROUP_CPUACCT
1414 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1416 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1419 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1421 update_load_add(&rq
->load
, load
);
1424 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1426 update_load_sub(&rq
->load
, load
);
1429 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1430 typedef int (*tg_visitor
)(struct task_group
*, void *);
1433 * Iterate the full tree, calling @down when first entering a node and @up when
1434 * leaving it for the final time.
1436 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1438 struct task_group
*parent
, *child
;
1442 parent
= &root_task_group
;
1444 ret
= (*down
)(parent
, data
);
1447 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1454 ret
= (*up
)(parent
, data
);
1459 parent
= parent
->parent
;
1468 static int tg_nop(struct task_group
*tg
, void *data
)
1475 static unsigned long source_load(int cpu
, int type
);
1476 static unsigned long target_load(int cpu
, int type
);
1477 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1479 static unsigned long cpu_avg_load_per_task(int cpu
)
1481 struct rq
*rq
= cpu_rq(cpu
);
1482 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1485 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1487 rq
->avg_load_per_task
= 0;
1489 return rq
->avg_load_per_task
;
1492 #ifdef CONFIG_FAIR_GROUP_SCHED
1494 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1497 * Calculate and set the cpu's group shares.
1500 update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1501 unsigned long sd_shares
, unsigned long sd_rq_weight
)
1503 unsigned long shares
;
1504 unsigned long rq_weight
;
1509 rq_weight
= tg
->cfs_rq
[cpu
]->rq_weight
;
1512 * \Sum shares * rq_weight
1513 * shares = -----------------------
1517 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1518 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1520 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1521 sysctl_sched_shares_thresh
) {
1522 struct rq
*rq
= cpu_rq(cpu
);
1523 unsigned long flags
;
1525 spin_lock_irqsave(&rq
->lock
, flags
);
1526 tg
->cfs_rq
[cpu
]->shares
= shares
;
1528 __set_se_shares(tg
->se
[cpu
], shares
);
1529 spin_unlock_irqrestore(&rq
->lock
, flags
);
1534 * Re-compute the task group their per cpu shares over the given domain.
1535 * This needs to be done in a bottom-up fashion because the rq weight of a
1536 * parent group depends on the shares of its child groups.
1538 static int tg_shares_up(struct task_group
*tg
, void *data
)
1540 unsigned long weight
, rq_weight
= 0;
1541 unsigned long shares
= 0;
1542 struct sched_domain
*sd
= data
;
1545 for_each_cpu(i
, sched_domain_span(sd
)) {
1547 * If there are currently no tasks on the cpu pretend there
1548 * is one of average load so that when a new task gets to
1549 * run here it will not get delayed by group starvation.
1551 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1553 weight
= NICE_0_LOAD
;
1555 tg
->cfs_rq
[i
]->rq_weight
= weight
;
1556 rq_weight
+= weight
;
1557 shares
+= tg
->cfs_rq
[i
]->shares
;
1560 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1561 shares
= tg
->shares
;
1563 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1564 shares
= tg
->shares
;
1566 for_each_cpu(i
, sched_domain_span(sd
))
1567 update_group_shares_cpu(tg
, i
, shares
, rq_weight
);
1573 * Compute the cpu's hierarchical load factor for each task group.
1574 * This needs to be done in a top-down fashion because the load of a child
1575 * group is a fraction of its parents load.
1577 static int tg_load_down(struct task_group
*tg
, void *data
)
1580 long cpu
= (long)data
;
1583 load
= cpu_rq(cpu
)->load
.weight
;
1585 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1586 load
*= tg
->cfs_rq
[cpu
]->shares
;
1587 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1590 tg
->cfs_rq
[cpu
]->h_load
= load
;
1595 static void update_shares(struct sched_domain
*sd
)
1597 u64 now
= cpu_clock(raw_smp_processor_id());
1598 s64 elapsed
= now
- sd
->last_update
;
1600 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1601 sd
->last_update
= now
;
1602 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1606 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1608 spin_unlock(&rq
->lock
);
1610 spin_lock(&rq
->lock
);
1613 static void update_h_load(long cpu
)
1615 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1620 static inline void update_shares(struct sched_domain
*sd
)
1624 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1630 #ifdef CONFIG_PREEMPT
1633 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1634 * way at the expense of forcing extra atomic operations in all
1635 * invocations. This assures that the double_lock is acquired using the
1636 * same underlying policy as the spinlock_t on this architecture, which
1637 * reduces latency compared to the unfair variant below. However, it
1638 * also adds more overhead and therefore may reduce throughput.
1640 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1641 __releases(this_rq
->lock
)
1642 __acquires(busiest
->lock
)
1643 __acquires(this_rq
->lock
)
1645 spin_unlock(&this_rq
->lock
);
1646 double_rq_lock(this_rq
, busiest
);
1653 * Unfair double_lock_balance: Optimizes throughput at the expense of
1654 * latency by eliminating extra atomic operations when the locks are
1655 * already in proper order on entry. This favors lower cpu-ids and will
1656 * grant the double lock to lower cpus over higher ids under contention,
1657 * regardless of entry order into the function.
1659 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1660 __releases(this_rq
->lock
)
1661 __acquires(busiest
->lock
)
1662 __acquires(this_rq
->lock
)
1666 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1667 if (busiest
< this_rq
) {
1668 spin_unlock(&this_rq
->lock
);
1669 spin_lock(&busiest
->lock
);
1670 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
1673 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
1678 #endif /* CONFIG_PREEMPT */
1681 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1683 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1685 if (unlikely(!irqs_disabled())) {
1686 /* printk() doesn't work good under rq->lock */
1687 spin_unlock(&this_rq
->lock
);
1691 return _double_lock_balance(this_rq
, busiest
);
1694 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1695 __releases(busiest
->lock
)
1697 spin_unlock(&busiest
->lock
);
1698 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1702 #ifdef CONFIG_FAIR_GROUP_SCHED
1703 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1706 cfs_rq
->shares
= shares
;
1711 #include "sched_stats.h"
1712 #include "sched_idletask.c"
1713 #include "sched_fair.c"
1714 #include "sched_rt.c"
1715 #ifdef CONFIG_SCHED_DEBUG
1716 # include "sched_debug.c"
1719 #define sched_class_highest (&rt_sched_class)
1720 #define for_each_class(class) \
1721 for (class = sched_class_highest; class; class = class->next)
1723 static void inc_nr_running(struct rq
*rq
)
1728 static void dec_nr_running(struct rq
*rq
)
1733 static void set_load_weight(struct task_struct
*p
)
1735 if (task_has_rt_policy(p
)) {
1736 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1737 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1742 * SCHED_IDLE tasks get minimal weight:
1744 if (p
->policy
== SCHED_IDLE
) {
1745 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1746 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1750 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1751 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1754 static void update_avg(u64
*avg
, u64 sample
)
1756 s64 diff
= sample
- *avg
;
1760 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1763 p
->se
.start_runtime
= p
->se
.sum_exec_runtime
;
1765 sched_info_queued(p
);
1766 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1770 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1773 if (p
->se
.last_wakeup
) {
1774 update_avg(&p
->se
.avg_overlap
,
1775 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1776 p
->se
.last_wakeup
= 0;
1778 update_avg(&p
->se
.avg_wakeup
,
1779 sysctl_sched_wakeup_granularity
);
1783 sched_info_dequeued(p
);
1784 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1789 * __normal_prio - return the priority that is based on the static prio
1791 static inline int __normal_prio(struct task_struct
*p
)
1793 return p
->static_prio
;
1797 * Calculate the expected normal priority: i.e. priority
1798 * without taking RT-inheritance into account. Might be
1799 * boosted by interactivity modifiers. Changes upon fork,
1800 * setprio syscalls, and whenever the interactivity
1801 * estimator recalculates.
1803 static inline int normal_prio(struct task_struct
*p
)
1807 if (task_has_rt_policy(p
))
1808 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1810 prio
= __normal_prio(p
);
1815 * Calculate the current priority, i.e. the priority
1816 * taken into account by the scheduler. This value might
1817 * be boosted by RT tasks, or might be boosted by
1818 * interactivity modifiers. Will be RT if the task got
1819 * RT-boosted. If not then it returns p->normal_prio.
1821 static int effective_prio(struct task_struct
*p
)
1823 p
->normal_prio
= normal_prio(p
);
1825 * If we are RT tasks or we were boosted to RT priority,
1826 * keep the priority unchanged. Otherwise, update priority
1827 * to the normal priority:
1829 if (!rt_prio(p
->prio
))
1830 return p
->normal_prio
;
1835 * activate_task - move a task to the runqueue.
1837 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1839 if (task_contributes_to_load(p
))
1840 rq
->nr_uninterruptible
--;
1842 enqueue_task(rq
, p
, wakeup
);
1847 * deactivate_task - remove a task from the runqueue.
1849 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1851 if (task_contributes_to_load(p
))
1852 rq
->nr_uninterruptible
++;
1854 dequeue_task(rq
, p
, sleep
);
1859 * task_curr - is this task currently executing on a CPU?
1860 * @p: the task in question.
1862 inline int task_curr(const struct task_struct
*p
)
1864 return cpu_curr(task_cpu(p
)) == p
;
1867 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1869 set_task_rq(p
, cpu
);
1872 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1873 * successfuly executed on another CPU. We must ensure that updates of
1874 * per-task data have been completed by this moment.
1877 task_thread_info(p
)->cpu
= cpu
;
1881 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1882 const struct sched_class
*prev_class
,
1883 int oldprio
, int running
)
1885 if (prev_class
!= p
->sched_class
) {
1886 if (prev_class
->switched_from
)
1887 prev_class
->switched_from(rq
, p
, running
);
1888 p
->sched_class
->switched_to(rq
, p
, running
);
1890 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1895 /* Used instead of source_load when we know the type == 0 */
1896 static unsigned long weighted_cpuload(const int cpu
)
1898 return cpu_rq(cpu
)->load
.weight
;
1902 * Is this task likely cache-hot:
1905 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1910 * Buddy candidates are cache hot:
1912 if (sched_feat(CACHE_HOT_BUDDY
) &&
1913 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
1914 &p
->se
== cfs_rq_of(&p
->se
)->last
))
1917 if (p
->sched_class
!= &fair_sched_class
)
1920 if (sysctl_sched_migration_cost
== -1)
1922 if (sysctl_sched_migration_cost
== 0)
1925 delta
= now
- p
->se
.exec_start
;
1927 return delta
< (s64
)sysctl_sched_migration_cost
;
1931 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1933 int old_cpu
= task_cpu(p
);
1934 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1935 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1936 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1939 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1941 trace_sched_migrate_task(p
, task_cpu(p
), new_cpu
);
1943 #ifdef CONFIG_SCHEDSTATS
1944 if (p
->se
.wait_start
)
1945 p
->se
.wait_start
-= clock_offset
;
1946 if (p
->se
.sleep_start
)
1947 p
->se
.sleep_start
-= clock_offset
;
1948 if (p
->se
.block_start
)
1949 p
->se
.block_start
-= clock_offset
;
1950 if (old_cpu
!= new_cpu
) {
1951 schedstat_inc(p
, se
.nr_migrations
);
1952 if (task_hot(p
, old_rq
->clock
, NULL
))
1953 schedstat_inc(p
, se
.nr_forced2_migrations
);
1956 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1957 new_cfsrq
->min_vruntime
;
1959 __set_task_cpu(p
, new_cpu
);
1962 struct migration_req
{
1963 struct list_head list
;
1965 struct task_struct
*task
;
1968 struct completion done
;
1972 * The task's runqueue lock must be held.
1973 * Returns true if you have to wait for migration thread.
1976 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1978 struct rq
*rq
= task_rq(p
);
1981 * If the task is not on a runqueue (and not running), then
1982 * it is sufficient to simply update the task's cpu field.
1984 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1985 set_task_cpu(p
, dest_cpu
);
1989 init_completion(&req
->done
);
1991 req
->dest_cpu
= dest_cpu
;
1992 list_add(&req
->list
, &rq
->migration_queue
);
1998 * wait_task_inactive - wait for a thread to unschedule.
2000 * If @match_state is nonzero, it's the @p->state value just checked and
2001 * not expected to change. If it changes, i.e. @p might have woken up,
2002 * then return zero. When we succeed in waiting for @p to be off its CPU,
2003 * we return a positive number (its total switch count). If a second call
2004 * a short while later returns the same number, the caller can be sure that
2005 * @p has remained unscheduled the whole time.
2007 * The caller must ensure that the task *will* unschedule sometime soon,
2008 * else this function might spin for a *long* time. This function can't
2009 * be called with interrupts off, or it may introduce deadlock with
2010 * smp_call_function() if an IPI is sent by the same process we are
2011 * waiting to become inactive.
2013 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2015 unsigned long flags
;
2022 * We do the initial early heuristics without holding
2023 * any task-queue locks at all. We'll only try to get
2024 * the runqueue lock when things look like they will
2030 * If the task is actively running on another CPU
2031 * still, just relax and busy-wait without holding
2034 * NOTE! Since we don't hold any locks, it's not
2035 * even sure that "rq" stays as the right runqueue!
2036 * But we don't care, since "task_running()" will
2037 * return false if the runqueue has changed and p
2038 * is actually now running somewhere else!
2040 while (task_running(rq
, p
)) {
2041 if (match_state
&& unlikely(p
->state
!= match_state
))
2047 * Ok, time to look more closely! We need the rq
2048 * lock now, to be *sure*. If we're wrong, we'll
2049 * just go back and repeat.
2051 rq
= task_rq_lock(p
, &flags
);
2052 trace_sched_wait_task(rq
, p
);
2053 running
= task_running(rq
, p
);
2054 on_rq
= p
->se
.on_rq
;
2056 if (!match_state
|| p
->state
== match_state
)
2057 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2058 task_rq_unlock(rq
, &flags
);
2061 * If it changed from the expected state, bail out now.
2063 if (unlikely(!ncsw
))
2067 * Was it really running after all now that we
2068 * checked with the proper locks actually held?
2070 * Oops. Go back and try again..
2072 if (unlikely(running
)) {
2078 * It's not enough that it's not actively running,
2079 * it must be off the runqueue _entirely_, and not
2082 * So if it was still runnable (but just not actively
2083 * running right now), it's preempted, and we should
2084 * yield - it could be a while.
2086 if (unlikely(on_rq
)) {
2087 schedule_timeout_uninterruptible(1);
2092 * Ahh, all good. It wasn't running, and it wasn't
2093 * runnable, which means that it will never become
2094 * running in the future either. We're all done!
2103 * kick_process - kick a running thread to enter/exit the kernel
2104 * @p: the to-be-kicked thread
2106 * Cause a process which is running on another CPU to enter
2107 * kernel-mode, without any delay. (to get signals handled.)
2109 * NOTE: this function doesnt have to take the runqueue lock,
2110 * because all it wants to ensure is that the remote task enters
2111 * the kernel. If the IPI races and the task has been migrated
2112 * to another CPU then no harm is done and the purpose has been
2115 void kick_process(struct task_struct
*p
)
2121 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2122 smp_send_reschedule(cpu
);
2127 * Return a low guess at the load of a migration-source cpu weighted
2128 * according to the scheduling class and "nice" value.
2130 * We want to under-estimate the load of migration sources, to
2131 * balance conservatively.
2133 static unsigned long source_load(int cpu
, int type
)
2135 struct rq
*rq
= cpu_rq(cpu
);
2136 unsigned long total
= weighted_cpuload(cpu
);
2138 if (type
== 0 || !sched_feat(LB_BIAS
))
2141 return min(rq
->cpu_load
[type
-1], total
);
2145 * Return a high guess at the load of a migration-target cpu weighted
2146 * according to the scheduling class and "nice" value.
2148 static unsigned long target_load(int cpu
, int type
)
2150 struct rq
*rq
= cpu_rq(cpu
);
2151 unsigned long total
= weighted_cpuload(cpu
);
2153 if (type
== 0 || !sched_feat(LB_BIAS
))
2156 return max(rq
->cpu_load
[type
-1], total
);
2160 * find_idlest_group finds and returns the least busy CPU group within the
2163 static struct sched_group
*
2164 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2166 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2167 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2168 int load_idx
= sd
->forkexec_idx
;
2169 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2172 unsigned long load
, avg_load
;
2176 /* Skip over this group if it has no CPUs allowed */
2177 if (!cpumask_intersects(sched_group_cpus(group
),
2181 local_group
= cpumask_test_cpu(this_cpu
,
2182 sched_group_cpus(group
));
2184 /* Tally up the load of all CPUs in the group */
2187 for_each_cpu(i
, sched_group_cpus(group
)) {
2188 /* Bias balancing toward cpus of our domain */
2190 load
= source_load(i
, load_idx
);
2192 load
= target_load(i
, load_idx
);
2197 /* Adjust by relative CPU power of the group */
2198 avg_load
= sg_div_cpu_power(group
,
2199 avg_load
* SCHED_LOAD_SCALE
);
2202 this_load
= avg_load
;
2204 } else if (avg_load
< min_load
) {
2205 min_load
= avg_load
;
2208 } while (group
= group
->next
, group
!= sd
->groups
);
2210 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2216 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2219 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
2221 unsigned long load
, min_load
= ULONG_MAX
;
2225 /* Traverse only the allowed CPUs */
2226 for_each_cpu_and(i
, sched_group_cpus(group
), &p
->cpus_allowed
) {
2227 load
= weighted_cpuload(i
);
2229 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2239 * sched_balance_self: balance the current task (running on cpu) in domains
2240 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2243 * Balance, ie. select the least loaded group.
2245 * Returns the target CPU number, or the same CPU if no balancing is needed.
2247 * preempt must be disabled.
2249 static int sched_balance_self(int cpu
, int flag
)
2251 struct task_struct
*t
= current
;
2252 struct sched_domain
*tmp
, *sd
= NULL
;
2254 for_each_domain(cpu
, tmp
) {
2256 * If power savings logic is enabled for a domain, stop there.
2258 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2260 if (tmp
->flags
& flag
)
2268 struct sched_group
*group
;
2269 int new_cpu
, weight
;
2271 if (!(sd
->flags
& flag
)) {
2276 group
= find_idlest_group(sd
, t
, cpu
);
2282 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
2283 if (new_cpu
== -1 || new_cpu
== cpu
) {
2284 /* Now try balancing at a lower domain level of cpu */
2289 /* Now try balancing at a lower domain level of new_cpu */
2291 weight
= cpumask_weight(sched_domain_span(sd
));
2293 for_each_domain(cpu
, tmp
) {
2294 if (weight
<= cpumask_weight(sched_domain_span(tmp
)))
2296 if (tmp
->flags
& flag
)
2299 /* while loop will break here if sd == NULL */
2305 #endif /* CONFIG_SMP */
2308 * try_to_wake_up - wake up a thread
2309 * @p: the to-be-woken-up thread
2310 * @state: the mask of task states that can be woken
2311 * @sync: do a synchronous wakeup?
2313 * Put it on the run-queue if it's not already there. The "current"
2314 * thread is always on the run-queue (except when the actual
2315 * re-schedule is in progress), and as such you're allowed to do
2316 * the simpler "current->state = TASK_RUNNING" to mark yourself
2317 * runnable without the overhead of this.
2319 * returns failure only if the task is already active.
2321 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2323 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2324 unsigned long flags
;
2328 if (!sched_feat(SYNC_WAKEUPS
))
2332 if (sched_feat(LB_WAKEUP_UPDATE
) && !root_task_group_empty()) {
2333 struct sched_domain
*sd
;
2335 this_cpu
= raw_smp_processor_id();
2338 for_each_domain(this_cpu
, sd
) {
2339 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2348 rq
= task_rq_lock(p
, &flags
);
2349 update_rq_clock(rq
);
2350 old_state
= p
->state
;
2351 if (!(old_state
& state
))
2359 this_cpu
= smp_processor_id();
2362 if (unlikely(task_running(rq
, p
)))
2365 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2366 if (cpu
!= orig_cpu
) {
2367 set_task_cpu(p
, cpu
);
2368 task_rq_unlock(rq
, &flags
);
2369 /* might preempt at this point */
2370 rq
= task_rq_lock(p
, &flags
);
2371 old_state
= p
->state
;
2372 if (!(old_state
& state
))
2377 this_cpu
= smp_processor_id();
2381 #ifdef CONFIG_SCHEDSTATS
2382 schedstat_inc(rq
, ttwu_count
);
2383 if (cpu
== this_cpu
)
2384 schedstat_inc(rq
, ttwu_local
);
2386 struct sched_domain
*sd
;
2387 for_each_domain(this_cpu
, sd
) {
2388 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2389 schedstat_inc(sd
, ttwu_wake_remote
);
2394 #endif /* CONFIG_SCHEDSTATS */
2397 #endif /* CONFIG_SMP */
2398 schedstat_inc(p
, se
.nr_wakeups
);
2400 schedstat_inc(p
, se
.nr_wakeups_sync
);
2401 if (orig_cpu
!= cpu
)
2402 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2403 if (cpu
== this_cpu
)
2404 schedstat_inc(p
, se
.nr_wakeups_local
);
2406 schedstat_inc(p
, se
.nr_wakeups_remote
);
2407 activate_task(rq
, p
, 1);
2411 * Only attribute actual wakeups done by this task.
2413 if (!in_interrupt()) {
2414 struct sched_entity
*se
= ¤t
->se
;
2415 u64 sample
= se
->sum_exec_runtime
;
2417 if (se
->last_wakeup
)
2418 sample
-= se
->last_wakeup
;
2420 sample
-= se
->start_runtime
;
2421 update_avg(&se
->avg_wakeup
, sample
);
2423 se
->last_wakeup
= se
->sum_exec_runtime
;
2427 trace_sched_wakeup(rq
, p
, success
);
2428 check_preempt_curr(rq
, p
, sync
);
2430 p
->state
= TASK_RUNNING
;
2432 if (p
->sched_class
->task_wake_up
)
2433 p
->sched_class
->task_wake_up(rq
, p
);
2436 task_rq_unlock(rq
, &flags
);
2441 int wake_up_process(struct task_struct
*p
)
2443 return try_to_wake_up(p
, TASK_ALL
, 0);
2445 EXPORT_SYMBOL(wake_up_process
);
2447 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2449 return try_to_wake_up(p
, state
, 0);
2453 * Perform scheduler related setup for a newly forked process p.
2454 * p is forked by current.
2456 * __sched_fork() is basic setup used by init_idle() too:
2458 static void __sched_fork(struct task_struct
*p
)
2460 p
->se
.exec_start
= 0;
2461 p
->se
.sum_exec_runtime
= 0;
2462 p
->se
.prev_sum_exec_runtime
= 0;
2463 p
->se
.last_wakeup
= 0;
2464 p
->se
.avg_overlap
= 0;
2465 p
->se
.start_runtime
= 0;
2466 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
2468 #ifdef CONFIG_SCHEDSTATS
2469 p
->se
.wait_start
= 0;
2470 p
->se
.sum_sleep_runtime
= 0;
2471 p
->se
.sleep_start
= 0;
2472 p
->se
.block_start
= 0;
2473 p
->se
.sleep_max
= 0;
2474 p
->se
.block_max
= 0;
2476 p
->se
.slice_max
= 0;
2480 INIT_LIST_HEAD(&p
->rt
.run_list
);
2482 INIT_LIST_HEAD(&p
->se
.group_node
);
2484 #ifdef CONFIG_PREEMPT_NOTIFIERS
2485 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2489 * We mark the process as running here, but have not actually
2490 * inserted it onto the runqueue yet. This guarantees that
2491 * nobody will actually run it, and a signal or other external
2492 * event cannot wake it up and insert it on the runqueue either.
2494 p
->state
= TASK_RUNNING
;
2498 * fork()/clone()-time setup:
2500 void sched_fork(struct task_struct
*p
, int clone_flags
)
2502 int cpu
= get_cpu();
2507 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2509 set_task_cpu(p
, cpu
);
2512 * Make sure we do not leak PI boosting priority to the child:
2514 p
->prio
= current
->normal_prio
;
2515 if (!rt_prio(p
->prio
))
2516 p
->sched_class
= &fair_sched_class
;
2518 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2519 if (likely(sched_info_on()))
2520 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2522 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2525 #ifdef CONFIG_PREEMPT
2526 /* Want to start with kernel preemption disabled. */
2527 task_thread_info(p
)->preempt_count
= 1;
2529 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2535 * wake_up_new_task - wake up a newly created task for the first time.
2537 * This function will do some initial scheduler statistics housekeeping
2538 * that must be done for every newly created context, then puts the task
2539 * on the runqueue and wakes it.
2541 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2543 unsigned long flags
;
2546 rq
= task_rq_lock(p
, &flags
);
2547 BUG_ON(p
->state
!= TASK_RUNNING
);
2548 update_rq_clock(rq
);
2550 p
->prio
= effective_prio(p
);
2552 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2553 activate_task(rq
, p
, 0);
2556 * Let the scheduling class do new task startup
2557 * management (if any):
2559 p
->sched_class
->task_new(rq
, p
);
2562 trace_sched_wakeup_new(rq
, p
, 1);
2563 check_preempt_curr(rq
, p
, 0);
2565 if (p
->sched_class
->task_wake_up
)
2566 p
->sched_class
->task_wake_up(rq
, p
);
2568 task_rq_unlock(rq
, &flags
);
2571 #ifdef CONFIG_PREEMPT_NOTIFIERS
2574 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2575 * @notifier: notifier struct to register
2577 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2579 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2581 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2584 * preempt_notifier_unregister - no longer interested in preemption notifications
2585 * @notifier: notifier struct to unregister
2587 * This is safe to call from within a preemption notifier.
2589 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2591 hlist_del(¬ifier
->link
);
2593 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2595 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2597 struct preempt_notifier
*notifier
;
2598 struct hlist_node
*node
;
2600 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2601 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2605 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2606 struct task_struct
*next
)
2608 struct preempt_notifier
*notifier
;
2609 struct hlist_node
*node
;
2611 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2612 notifier
->ops
->sched_out(notifier
, next
);
2615 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2617 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2622 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2623 struct task_struct
*next
)
2627 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2630 * prepare_task_switch - prepare to switch tasks
2631 * @rq: the runqueue preparing to switch
2632 * @prev: the current task that is being switched out
2633 * @next: the task we are going to switch to.
2635 * This is called with the rq lock held and interrupts off. It must
2636 * be paired with a subsequent finish_task_switch after the context
2639 * prepare_task_switch sets up locking and calls architecture specific
2643 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2644 struct task_struct
*next
)
2646 fire_sched_out_preempt_notifiers(prev
, next
);
2647 prepare_lock_switch(rq
, next
);
2648 prepare_arch_switch(next
);
2652 * finish_task_switch - clean up after a task-switch
2653 * @rq: runqueue associated with task-switch
2654 * @prev: the thread we just switched away from.
2656 * finish_task_switch must be called after the context switch, paired
2657 * with a prepare_task_switch call before the context switch.
2658 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2659 * and do any other architecture-specific cleanup actions.
2661 * Note that we may have delayed dropping an mm in context_switch(). If
2662 * so, we finish that here outside of the runqueue lock. (Doing it
2663 * with the lock held can cause deadlocks; see schedule() for
2666 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2667 __releases(rq
->lock
)
2669 struct mm_struct
*mm
= rq
->prev_mm
;
2672 int post_schedule
= 0;
2674 if (current
->sched_class
->needs_post_schedule
)
2675 post_schedule
= current
->sched_class
->needs_post_schedule(rq
);
2681 * A task struct has one reference for the use as "current".
2682 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2683 * schedule one last time. The schedule call will never return, and
2684 * the scheduled task must drop that reference.
2685 * The test for TASK_DEAD must occur while the runqueue locks are
2686 * still held, otherwise prev could be scheduled on another cpu, die
2687 * there before we look at prev->state, and then the reference would
2689 * Manfred Spraul <manfred@colorfullife.com>
2691 prev_state
= prev
->state
;
2692 finish_arch_switch(prev
);
2693 finish_lock_switch(rq
, prev
);
2696 current
->sched_class
->post_schedule(rq
);
2699 fire_sched_in_preempt_notifiers(current
);
2702 if (unlikely(prev_state
== TASK_DEAD
)) {
2704 * Remove function-return probe instances associated with this
2705 * task and put them back on the free list.
2707 kprobe_flush_task(prev
);
2708 put_task_struct(prev
);
2713 * schedule_tail - first thing a freshly forked thread must call.
2714 * @prev: the thread we just switched away from.
2716 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2717 __releases(rq
->lock
)
2719 struct rq
*rq
= this_rq();
2721 finish_task_switch(rq
, prev
);
2722 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2723 /* In this case, finish_task_switch does not reenable preemption */
2726 if (current
->set_child_tid
)
2727 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2731 * context_switch - switch to the new MM and the new
2732 * thread's register state.
2735 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2736 struct task_struct
*next
)
2738 struct mm_struct
*mm
, *oldmm
;
2740 prepare_task_switch(rq
, prev
, next
);
2741 trace_sched_switch(rq
, prev
, next
);
2743 oldmm
= prev
->active_mm
;
2745 * For paravirt, this is coupled with an exit in switch_to to
2746 * combine the page table reload and the switch backend into
2749 arch_enter_lazy_cpu_mode();
2751 if (unlikely(!mm
)) {
2752 next
->active_mm
= oldmm
;
2753 atomic_inc(&oldmm
->mm_count
);
2754 enter_lazy_tlb(oldmm
, next
);
2756 switch_mm(oldmm
, mm
, next
);
2758 if (unlikely(!prev
->mm
)) {
2759 prev
->active_mm
= NULL
;
2760 rq
->prev_mm
= oldmm
;
2763 * Since the runqueue lock will be released by the next
2764 * task (which is an invalid locking op but in the case
2765 * of the scheduler it's an obvious special-case), so we
2766 * do an early lockdep release here:
2768 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2769 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2772 /* Here we just switch the register state and the stack. */
2773 switch_to(prev
, next
, prev
);
2777 * this_rq must be evaluated again because prev may have moved
2778 * CPUs since it called schedule(), thus the 'rq' on its stack
2779 * frame will be invalid.
2781 finish_task_switch(this_rq(), prev
);
2785 * nr_running, nr_uninterruptible and nr_context_switches:
2787 * externally visible scheduler statistics: current number of runnable
2788 * threads, current number of uninterruptible-sleeping threads, total
2789 * number of context switches performed since bootup.
2791 unsigned long nr_running(void)
2793 unsigned long i
, sum
= 0;
2795 for_each_online_cpu(i
)
2796 sum
+= cpu_rq(i
)->nr_running
;
2801 unsigned long nr_uninterruptible(void)
2803 unsigned long i
, sum
= 0;
2805 for_each_possible_cpu(i
)
2806 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2809 * Since we read the counters lockless, it might be slightly
2810 * inaccurate. Do not allow it to go below zero though:
2812 if (unlikely((long)sum
< 0))
2818 unsigned long long nr_context_switches(void)
2821 unsigned long long sum
= 0;
2823 for_each_possible_cpu(i
)
2824 sum
+= cpu_rq(i
)->nr_switches
;
2829 unsigned long nr_iowait(void)
2831 unsigned long i
, sum
= 0;
2833 for_each_possible_cpu(i
)
2834 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2839 unsigned long nr_active(void)
2841 unsigned long i
, running
= 0, uninterruptible
= 0;
2843 for_each_online_cpu(i
) {
2844 running
+= cpu_rq(i
)->nr_running
;
2845 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2848 if (unlikely((long)uninterruptible
< 0))
2849 uninterruptible
= 0;
2851 return running
+ uninterruptible
;
2855 * Update rq->cpu_load[] statistics. This function is usually called every
2856 * scheduler tick (TICK_NSEC).
2858 static void update_cpu_load(struct rq
*this_rq
)
2860 unsigned long this_load
= this_rq
->load
.weight
;
2863 this_rq
->nr_load_updates
++;
2865 /* Update our load: */
2866 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2867 unsigned long old_load
, new_load
;
2869 /* scale is effectively 1 << i now, and >> i divides by scale */
2871 old_load
= this_rq
->cpu_load
[i
];
2872 new_load
= this_load
;
2874 * Round up the averaging division if load is increasing. This
2875 * prevents us from getting stuck on 9 if the load is 10, for
2878 if (new_load
> old_load
)
2879 new_load
+= scale
-1;
2880 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2887 * double_rq_lock - safely lock two runqueues
2889 * Note this does not disable interrupts like task_rq_lock,
2890 * you need to do so manually before calling.
2892 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2893 __acquires(rq1
->lock
)
2894 __acquires(rq2
->lock
)
2896 BUG_ON(!irqs_disabled());
2898 spin_lock(&rq1
->lock
);
2899 __acquire(rq2
->lock
); /* Fake it out ;) */
2902 spin_lock(&rq1
->lock
);
2903 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
2905 spin_lock(&rq2
->lock
);
2906 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
2909 update_rq_clock(rq1
);
2910 update_rq_clock(rq2
);
2914 * double_rq_unlock - safely unlock two runqueues
2916 * Note this does not restore interrupts like task_rq_unlock,
2917 * you need to do so manually after calling.
2919 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2920 __releases(rq1
->lock
)
2921 __releases(rq2
->lock
)
2923 spin_unlock(&rq1
->lock
);
2925 spin_unlock(&rq2
->lock
);
2927 __release(rq2
->lock
);
2931 * If dest_cpu is allowed for this process, migrate the task to it.
2932 * This is accomplished by forcing the cpu_allowed mask to only
2933 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2934 * the cpu_allowed mask is restored.
2936 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2938 struct migration_req req
;
2939 unsigned long flags
;
2942 rq
= task_rq_lock(p
, &flags
);
2943 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
2944 || unlikely(!cpu_active(dest_cpu
)))
2947 /* force the process onto the specified CPU */
2948 if (migrate_task(p
, dest_cpu
, &req
)) {
2949 /* Need to wait for migration thread (might exit: take ref). */
2950 struct task_struct
*mt
= rq
->migration_thread
;
2952 get_task_struct(mt
);
2953 task_rq_unlock(rq
, &flags
);
2954 wake_up_process(mt
);
2955 put_task_struct(mt
);
2956 wait_for_completion(&req
.done
);
2961 task_rq_unlock(rq
, &flags
);
2965 * sched_exec - execve() is a valuable balancing opportunity, because at
2966 * this point the task has the smallest effective memory and cache footprint.
2968 void sched_exec(void)
2970 int new_cpu
, this_cpu
= get_cpu();
2971 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2973 if (new_cpu
!= this_cpu
)
2974 sched_migrate_task(current
, new_cpu
);
2978 * pull_task - move a task from a remote runqueue to the local runqueue.
2979 * Both runqueues must be locked.
2981 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2982 struct rq
*this_rq
, int this_cpu
)
2984 deactivate_task(src_rq
, p
, 0);
2985 set_task_cpu(p
, this_cpu
);
2986 activate_task(this_rq
, p
, 0);
2988 * Note that idle threads have a prio of MAX_PRIO, for this test
2989 * to be always true for them.
2991 check_preempt_curr(this_rq
, p
, 0);
2995 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2998 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2999 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3002 int tsk_cache_hot
= 0;
3004 * We do not migrate tasks that are:
3005 * 1) running (obviously), or
3006 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3007 * 3) are cache-hot on their current CPU.
3009 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
3010 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
3015 if (task_running(rq
, p
)) {
3016 schedstat_inc(p
, se
.nr_failed_migrations_running
);
3021 * Aggressive migration if:
3022 * 1) task is cache cold, or
3023 * 2) too many balance attempts have failed.
3026 tsk_cache_hot
= task_hot(p
, rq
->clock
, sd
);
3027 if (!tsk_cache_hot
||
3028 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
3029 #ifdef CONFIG_SCHEDSTATS
3030 if (tsk_cache_hot
) {
3031 schedstat_inc(sd
, lb_hot_gained
[idle
]);
3032 schedstat_inc(p
, se
.nr_forced_migrations
);
3038 if (tsk_cache_hot
) {
3039 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
3045 static unsigned long
3046 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3047 unsigned long max_load_move
, struct sched_domain
*sd
,
3048 enum cpu_idle_type idle
, int *all_pinned
,
3049 int *this_best_prio
, struct rq_iterator
*iterator
)
3051 int loops
= 0, pulled
= 0, pinned
= 0;
3052 struct task_struct
*p
;
3053 long rem_load_move
= max_load_move
;
3055 if (max_load_move
== 0)
3061 * Start the load-balancing iterator:
3063 p
= iterator
->start(iterator
->arg
);
3065 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3068 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
3069 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3070 p
= iterator
->next(iterator
->arg
);
3074 pull_task(busiest
, p
, this_rq
, this_cpu
);
3076 rem_load_move
-= p
->se
.load
.weight
;
3078 #ifdef CONFIG_PREEMPT
3080 * NEWIDLE balancing is a source of latency, so preemptible kernels
3081 * will stop after the first task is pulled to minimize the critical
3084 if (idle
== CPU_NEWLY_IDLE
)
3089 * We only want to steal up to the prescribed amount of weighted load.
3091 if (rem_load_move
> 0) {
3092 if (p
->prio
< *this_best_prio
)
3093 *this_best_prio
= p
->prio
;
3094 p
= iterator
->next(iterator
->arg
);
3099 * Right now, this is one of only two places pull_task() is called,
3100 * so we can safely collect pull_task() stats here rather than
3101 * inside pull_task().
3103 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3106 *all_pinned
= pinned
;
3108 return max_load_move
- rem_load_move
;
3112 * move_tasks tries to move up to max_load_move weighted load from busiest to
3113 * this_rq, as part of a balancing operation within domain "sd".
3114 * Returns 1 if successful and 0 otherwise.
3116 * Called with both runqueues locked.
3118 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3119 unsigned long max_load_move
,
3120 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3123 const struct sched_class
*class = sched_class_highest
;
3124 unsigned long total_load_moved
= 0;
3125 int this_best_prio
= this_rq
->curr
->prio
;
3129 class->load_balance(this_rq
, this_cpu
, busiest
,
3130 max_load_move
- total_load_moved
,
3131 sd
, idle
, all_pinned
, &this_best_prio
);
3132 class = class->next
;
3134 #ifdef CONFIG_PREEMPT
3136 * NEWIDLE balancing is a source of latency, so preemptible
3137 * kernels will stop after the first task is pulled to minimize
3138 * the critical section.
3140 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3143 } while (class && max_load_move
> total_load_moved
);
3145 return total_load_moved
> 0;
3149 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3150 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3151 struct rq_iterator
*iterator
)
3153 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3157 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3158 pull_task(busiest
, p
, this_rq
, this_cpu
);
3160 * Right now, this is only the second place pull_task()
3161 * is called, so we can safely collect pull_task()
3162 * stats here rather than inside pull_task().
3164 schedstat_inc(sd
, lb_gained
[idle
]);
3168 p
= iterator
->next(iterator
->arg
);
3175 * move_one_task tries to move exactly one task from busiest to this_rq, as
3176 * part of active balancing operations within "domain".
3177 * Returns 1 if successful and 0 otherwise.
3179 * Called with both runqueues locked.
3181 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3182 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3184 const struct sched_class
*class;
3186 for (class = sched_class_highest
; class; class = class->next
)
3187 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3192 /********** Helpers for find_busiest_group ************************/
3194 * sd_lb_stats - Structure to store the statistics of a sched_domain
3195 * during load balancing.
3197 struct sd_lb_stats
{
3198 struct sched_group
*busiest
; /* Busiest group in this sd */
3199 struct sched_group
*this; /* Local group in this sd */
3200 unsigned long total_load
; /* Total load of all groups in sd */
3201 unsigned long total_pwr
; /* Total power of all groups in sd */
3202 unsigned long avg_load
; /* Average load across all groups in sd */
3204 /** Statistics of this group */
3205 unsigned long this_load
;
3206 unsigned long this_load_per_task
;
3207 unsigned long this_nr_running
;
3209 /* Statistics of the busiest group */
3210 unsigned long max_load
;
3211 unsigned long busiest_load_per_task
;
3212 unsigned long busiest_nr_running
;
3214 int group_imb
; /* Is there imbalance in this sd */
3215 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3216 int power_savings_balance
; /* Is powersave balance needed for this sd */
3217 struct sched_group
*group_min
; /* Least loaded group in sd */
3218 struct sched_group
*group_leader
; /* Group which relieves group_min */
3219 unsigned long min_load_per_task
; /* load_per_task in group_min */
3220 unsigned long leader_nr_running
; /* Nr running of group_leader */
3221 unsigned long min_nr_running
; /* Nr running of group_min */
3226 * sg_lb_stats - stats of a sched_group required for load_balancing
3228 struct sg_lb_stats
{
3229 unsigned long avg_load
; /*Avg load across the CPUs of the group */
3230 unsigned long group_load
; /* Total load over the CPUs of the group */
3231 unsigned long sum_nr_running
; /* Nr tasks running in the group */
3232 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
3233 unsigned long group_capacity
;
3234 int group_imb
; /* Is there an imbalance in the group ? */
3238 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3239 * @group: The group whose first cpu is to be returned.
3241 static inline unsigned int group_first_cpu(struct sched_group
*group
)
3243 return cpumask_first(sched_group_cpus(group
));
3247 * get_sd_load_idx - Obtain the load index for a given sched domain.
3248 * @sd: The sched_domain whose load_idx is to be obtained.
3249 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3251 static inline int get_sd_load_idx(struct sched_domain
*sd
,
3252 enum cpu_idle_type idle
)
3258 load_idx
= sd
->busy_idx
;
3261 case CPU_NEWLY_IDLE
:
3262 load_idx
= sd
->newidle_idx
;
3265 load_idx
= sd
->idle_idx
;
3273 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3275 * init_sd_power_savings_stats - Initialize power savings statistics for
3276 * the given sched_domain, during load balancing.
3278 * @sd: Sched domain whose power-savings statistics are to be initialized.
3279 * @sds: Variable containing the statistics for sd.
3280 * @idle: Idle status of the CPU at which we're performing load-balancing.
3282 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3283 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3286 * Busy processors will not participate in power savings
3289 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3290 sds
->power_savings_balance
= 0;
3292 sds
->power_savings_balance
= 1;
3293 sds
->min_nr_running
= ULONG_MAX
;
3294 sds
->leader_nr_running
= 0;
3299 * update_sd_power_savings_stats - Update the power saving stats for a
3300 * sched_domain while performing load balancing.
3302 * @group: sched_group belonging to the sched_domain under consideration.
3303 * @sds: Variable containing the statistics of the sched_domain
3304 * @local_group: Does group contain the CPU for which we're performing
3306 * @sgs: Variable containing the statistics of the group.
3308 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3309 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3312 if (!sds
->power_savings_balance
)
3316 * If the local group is idle or completely loaded
3317 * no need to do power savings balance at this domain
3319 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
3320 !sds
->this_nr_running
))
3321 sds
->power_savings_balance
= 0;
3324 * If a group is already running at full capacity or idle,
3325 * don't include that group in power savings calculations
3327 if (!sds
->power_savings_balance
||
3328 sgs
->sum_nr_running
>= sgs
->group_capacity
||
3329 !sgs
->sum_nr_running
)
3333 * Calculate the group which has the least non-idle load.
3334 * This is the group from where we need to pick up the load
3337 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
3338 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
3339 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
3340 sds
->group_min
= group
;
3341 sds
->min_nr_running
= sgs
->sum_nr_running
;
3342 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
3343 sgs
->sum_nr_running
;
3347 * Calculate the group which is almost near its
3348 * capacity but still has some space to pick up some load
3349 * from other group and save more power
3351 if (sgs
->sum_nr_running
> sgs
->group_capacity
- 1)
3354 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
3355 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
3356 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
3357 sds
->group_leader
= group
;
3358 sds
->leader_nr_running
= sgs
->sum_nr_running
;
3363 * check_power_save_busiest_group - Check if we have potential to perform
3364 * some power-savings balance. If yes, set the busiest group to be
3365 * the least loaded group in the sched_domain, so that it's CPUs can
3368 * @sds: Variable containing the statistics of the sched_domain
3369 * under consideration.
3370 * @this_cpu: Cpu at which we're currently performing load-balancing.
3371 * @imbalance: Variable to store the imbalance.
3373 * Returns 1 if there is potential to perform power-savings balance.
3376 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3377 int this_cpu
, unsigned long *imbalance
)
3379 if (!sds
->power_savings_balance
)
3382 if (sds
->this != sds
->group_leader
||
3383 sds
->group_leader
== sds
->group_min
)
3386 *imbalance
= sds
->min_load_per_task
;
3387 sds
->busiest
= sds
->group_min
;
3389 if (sched_mc_power_savings
>= POWERSAVINGS_BALANCE_WAKEUP
) {
3390 cpu_rq(this_cpu
)->rd
->sched_mc_preferred_wakeup_cpu
=
3391 group_first_cpu(sds
->group_leader
);
3397 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3398 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3399 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3404 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3405 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3410 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3411 int this_cpu
, unsigned long *imbalance
)
3415 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3419 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3420 * @group: sched_group whose statistics are to be updated.
3421 * @this_cpu: Cpu for which load balance is currently performed.
3422 * @idle: Idle status of this_cpu
3423 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3424 * @sd_idle: Idle status of the sched_domain containing group.
3425 * @local_group: Does group contain this_cpu.
3426 * @cpus: Set of cpus considered for load balancing.
3427 * @balance: Should we balance.
3428 * @sgs: variable to hold the statistics for this group.
3430 static inline void update_sg_lb_stats(struct sched_group
*group
, int this_cpu
,
3431 enum cpu_idle_type idle
, int load_idx
, int *sd_idle
,
3432 int local_group
, const struct cpumask
*cpus
,
3433 int *balance
, struct sg_lb_stats
*sgs
)
3435 unsigned long load
, max_cpu_load
, min_cpu_load
;
3437 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3438 unsigned long sum_avg_load_per_task
;
3439 unsigned long avg_load_per_task
;
3442 balance_cpu
= group_first_cpu(group
);
3444 /* Tally up the load of all CPUs in the group */
3445 sum_avg_load_per_task
= avg_load_per_task
= 0;
3447 min_cpu_load
= ~0UL;
3449 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3450 struct rq
*rq
= cpu_rq(i
);
3452 if (*sd_idle
&& rq
->nr_running
)
3455 /* Bias balancing toward cpus of our domain */
3457 if (idle_cpu(i
) && !first_idle_cpu
) {
3462 load
= target_load(i
, load_idx
);
3464 load
= source_load(i
, load_idx
);
3465 if (load
> max_cpu_load
)
3466 max_cpu_load
= load
;
3467 if (min_cpu_load
> load
)
3468 min_cpu_load
= load
;
3471 sgs
->group_load
+= load
;
3472 sgs
->sum_nr_running
+= rq
->nr_running
;
3473 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
3475 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3479 * First idle cpu or the first cpu(busiest) in this sched group
3480 * is eligible for doing load balancing at this and above
3481 * domains. In the newly idle case, we will allow all the cpu's
3482 * to do the newly idle load balance.
3484 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3485 balance_cpu
!= this_cpu
&& balance
) {
3490 /* Adjust by relative CPU power of the group */
3491 sgs
->avg_load
= sg_div_cpu_power(group
,
3492 sgs
->group_load
* SCHED_LOAD_SCALE
);
3496 * Consider the group unbalanced when the imbalance is larger
3497 * than the average weight of two tasks.
3499 * APZ: with cgroup the avg task weight can vary wildly and
3500 * might not be a suitable number - should we keep a
3501 * normalized nr_running number somewhere that negates
3504 avg_load_per_task
= sg_div_cpu_power(group
,
3505 sum_avg_load_per_task
* SCHED_LOAD_SCALE
);
3507 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3510 sgs
->group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3515 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3516 * @sd: sched_domain whose statistics are to be updated.
3517 * @this_cpu: Cpu for which load balance is currently performed.
3518 * @idle: Idle status of this_cpu
3519 * @sd_idle: Idle status of the sched_domain containing group.
3520 * @cpus: Set of cpus considered for load balancing.
3521 * @balance: Should we balance.
3522 * @sds: variable to hold the statistics for this sched_domain.
3524 static inline void update_sd_lb_stats(struct sched_domain
*sd
, int this_cpu
,
3525 enum cpu_idle_type idle
, int *sd_idle
,
3526 const struct cpumask
*cpus
, int *balance
,
3527 struct sd_lb_stats
*sds
)
3529 struct sched_group
*group
= sd
->groups
;
3530 struct sg_lb_stats sgs
;
3533 init_sd_power_savings_stats(sd
, sds
, idle
);
3534 load_idx
= get_sd_load_idx(sd
, idle
);
3539 local_group
= cpumask_test_cpu(this_cpu
,
3540 sched_group_cpus(group
));
3541 memset(&sgs
, 0, sizeof(sgs
));
3542 update_sg_lb_stats(group
, this_cpu
, idle
, load_idx
, sd_idle
,
3543 local_group
, cpus
, balance
, &sgs
);
3545 if (local_group
&& balance
&& !(*balance
))
3548 sds
->total_load
+= sgs
.group_load
;
3549 sds
->total_pwr
+= group
->__cpu_power
;
3552 sds
->this_load
= sgs
.avg_load
;
3554 sds
->this_nr_running
= sgs
.sum_nr_running
;
3555 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
3556 } else if (sgs
.avg_load
> sds
->max_load
&&
3557 (sgs
.sum_nr_running
> sgs
.group_capacity
||
3559 sds
->max_load
= sgs
.avg_load
;
3560 sds
->busiest
= group
;
3561 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
3562 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
3563 sds
->group_imb
= sgs
.group_imb
;
3566 update_sd_power_savings_stats(group
, sds
, local_group
, &sgs
);
3567 group
= group
->next
;
3568 } while (group
!= sd
->groups
);
3573 * fix_small_imbalance - Calculate the minor imbalance that exists
3574 * amongst the groups of a sched_domain, during
3576 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3577 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3578 * @imbalance: Variable to store the imbalance.
3580 static inline void fix_small_imbalance(struct sd_lb_stats
*sds
,
3581 int this_cpu
, unsigned long *imbalance
)
3583 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
3584 unsigned int imbn
= 2;
3586 if (sds
->this_nr_running
) {
3587 sds
->this_load_per_task
/= sds
->this_nr_running
;
3588 if (sds
->busiest_load_per_task
>
3589 sds
->this_load_per_task
)
3592 sds
->this_load_per_task
=
3593 cpu_avg_load_per_task(this_cpu
);
3595 if (sds
->max_load
- sds
->this_load
+ sds
->busiest_load_per_task
>=
3596 sds
->busiest_load_per_task
* imbn
) {
3597 *imbalance
= sds
->busiest_load_per_task
;
3602 * OK, we don't have enough imbalance to justify moving tasks,
3603 * however we may be able to increase total CPU power used by
3607 pwr_now
+= sds
->busiest
->__cpu_power
*
3608 min(sds
->busiest_load_per_task
, sds
->max_load
);
3609 pwr_now
+= sds
->this->__cpu_power
*
3610 min(sds
->this_load_per_task
, sds
->this_load
);
3611 pwr_now
/= SCHED_LOAD_SCALE
;
3613 /* Amount of load we'd subtract */
3614 tmp
= sg_div_cpu_power(sds
->busiest
,
3615 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
);
3616 if (sds
->max_load
> tmp
)
3617 pwr_move
+= sds
->busiest
->__cpu_power
*
3618 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
3620 /* Amount of load we'd add */
3621 if (sds
->max_load
* sds
->busiest
->__cpu_power
<
3622 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
)
3623 tmp
= sg_div_cpu_power(sds
->this,
3624 sds
->max_load
* sds
->busiest
->__cpu_power
);
3626 tmp
= sg_div_cpu_power(sds
->this,
3627 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
);
3628 pwr_move
+= sds
->this->__cpu_power
*
3629 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
3630 pwr_move
/= SCHED_LOAD_SCALE
;
3632 /* Move if we gain throughput */
3633 if (pwr_move
> pwr_now
)
3634 *imbalance
= sds
->busiest_load_per_task
;
3638 * calculate_imbalance - Calculate the amount of imbalance present within the
3639 * groups of a given sched_domain during load balance.
3640 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3641 * @this_cpu: Cpu for which currently load balance is being performed.
3642 * @imbalance: The variable to store the imbalance.
3644 static inline void calculate_imbalance(struct sd_lb_stats
*sds
, int this_cpu
,
3645 unsigned long *imbalance
)
3647 unsigned long max_pull
;
3649 * In the presence of smp nice balancing, certain scenarios can have
3650 * max load less than avg load(as we skip the groups at or below
3651 * its cpu_power, while calculating max_load..)
3653 if (sds
->max_load
< sds
->avg_load
) {
3655 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3658 /* Don't want to pull so many tasks that a group would go idle */
3659 max_pull
= min(sds
->max_load
- sds
->avg_load
,
3660 sds
->max_load
- sds
->busiest_load_per_task
);
3662 /* How much load to actually move to equalise the imbalance */
3663 *imbalance
= min(max_pull
* sds
->busiest
->__cpu_power
,
3664 (sds
->avg_load
- sds
->this_load
) * sds
->this->__cpu_power
)
3668 * if *imbalance is less than the average load per runnable task
3669 * there is no gaurantee that any tasks will be moved so we'll have
3670 * a think about bumping its value to force at least one task to be
3673 if (*imbalance
< sds
->busiest_load_per_task
)
3674 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3677 /******* find_busiest_group() helpers end here *********************/
3680 * find_busiest_group - Returns the busiest group within the sched_domain
3681 * if there is an imbalance. If there isn't an imbalance, and
3682 * the user has opted for power-savings, it returns a group whose
3683 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3684 * such a group exists.
3686 * Also calculates the amount of weighted load which should be moved
3687 * to restore balance.
3689 * @sd: The sched_domain whose busiest group is to be returned.
3690 * @this_cpu: The cpu for which load balancing is currently being performed.
3691 * @imbalance: Variable which stores amount of weighted load which should
3692 * be moved to restore balance/put a group to idle.
3693 * @idle: The idle status of this_cpu.
3694 * @sd_idle: The idleness of sd
3695 * @cpus: The set of CPUs under consideration for load-balancing.
3696 * @balance: Pointer to a variable indicating if this_cpu
3697 * is the appropriate cpu to perform load balancing at this_level.
3699 * Returns: - the busiest group if imbalance exists.
3700 * - If no imbalance and user has opted for power-savings balance,
3701 * return the least loaded group whose CPUs can be
3702 * put to idle by rebalancing its tasks onto our group.
3704 static struct sched_group
*
3705 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3706 unsigned long *imbalance
, enum cpu_idle_type idle
,
3707 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
3709 struct sd_lb_stats sds
;
3711 memset(&sds
, 0, sizeof(sds
));
3714 * Compute the various statistics relavent for load balancing at
3717 update_sd_lb_stats(sd
, this_cpu
, idle
, sd_idle
, cpus
,
3720 /* Cases where imbalance does not exist from POV of this_cpu */
3721 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3723 * 2) There is no busy sibling group to pull from.
3724 * 3) This group is the busiest group.
3725 * 4) This group is more busy than the avg busieness at this
3727 * 5) The imbalance is within the specified limit.
3728 * 6) Any rebalance would lead to ping-pong
3730 if (balance
&& !(*balance
))
3733 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
3736 if (sds
.this_load
>= sds
.max_load
)
3739 sds
.avg_load
= (SCHED_LOAD_SCALE
* sds
.total_load
) / sds
.total_pwr
;
3741 if (sds
.this_load
>= sds
.avg_load
)
3744 if (100 * sds
.max_load
<= sd
->imbalance_pct
* sds
.this_load
)
3747 sds
.busiest_load_per_task
/= sds
.busiest_nr_running
;
3749 sds
.busiest_load_per_task
=
3750 min(sds
.busiest_load_per_task
, sds
.avg_load
);
3753 * We're trying to get all the cpus to the average_load, so we don't
3754 * want to push ourselves above the average load, nor do we wish to
3755 * reduce the max loaded cpu below the average load, as either of these
3756 * actions would just result in more rebalancing later, and ping-pong
3757 * tasks around. Thus we look for the minimum possible imbalance.
3758 * Negative imbalances (*we* are more loaded than anyone else) will
3759 * be counted as no imbalance for these purposes -- we can't fix that
3760 * by pulling tasks to us. Be careful of negative numbers as they'll
3761 * appear as very large values with unsigned longs.
3763 if (sds
.max_load
<= sds
.busiest_load_per_task
)
3766 /* Looks like there is an imbalance. Compute it */
3767 calculate_imbalance(&sds
, this_cpu
, imbalance
);
3772 * There is no obvious imbalance. But check if we can do some balancing
3775 if (check_power_save_busiest_group(&sds
, this_cpu
, imbalance
))
3783 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3786 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3787 unsigned long imbalance
, const struct cpumask
*cpus
)
3789 struct rq
*busiest
= NULL
, *rq
;
3790 unsigned long max_load
= 0;
3793 for_each_cpu(i
, sched_group_cpus(group
)) {
3796 if (!cpumask_test_cpu(i
, cpus
))
3800 wl
= weighted_cpuload(i
);
3802 if (rq
->nr_running
== 1 && wl
> imbalance
)
3805 if (wl
> max_load
) {
3815 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3816 * so long as it is large enough.
3818 #define MAX_PINNED_INTERVAL 512
3821 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3822 * tasks if there is an imbalance.
3824 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3825 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3826 int *balance
, struct cpumask
*cpus
)
3828 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3829 struct sched_group
*group
;
3830 unsigned long imbalance
;
3832 unsigned long flags
;
3834 cpumask_setall(cpus
);
3837 * When power savings policy is enabled for the parent domain, idle
3838 * sibling can pick up load irrespective of busy siblings. In this case,
3839 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3840 * portraying it as CPU_NOT_IDLE.
3842 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3843 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3846 schedstat_inc(sd
, lb_count
[idle
]);
3850 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3857 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3861 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3863 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3867 BUG_ON(busiest
== this_rq
);
3869 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3872 if (busiest
->nr_running
> 1) {
3874 * Attempt to move tasks. If find_busiest_group has found
3875 * an imbalance but busiest->nr_running <= 1, the group is
3876 * still unbalanced. ld_moved simply stays zero, so it is
3877 * correctly treated as an imbalance.
3879 local_irq_save(flags
);
3880 double_rq_lock(this_rq
, busiest
);
3881 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3882 imbalance
, sd
, idle
, &all_pinned
);
3883 double_rq_unlock(this_rq
, busiest
);
3884 local_irq_restore(flags
);
3887 * some other cpu did the load balance for us.
3889 if (ld_moved
&& this_cpu
!= smp_processor_id())
3890 resched_cpu(this_cpu
);
3892 /* All tasks on this runqueue were pinned by CPU affinity */
3893 if (unlikely(all_pinned
)) {
3894 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
3895 if (!cpumask_empty(cpus
))
3902 schedstat_inc(sd
, lb_failed
[idle
]);
3903 sd
->nr_balance_failed
++;
3905 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3907 spin_lock_irqsave(&busiest
->lock
, flags
);
3909 /* don't kick the migration_thread, if the curr
3910 * task on busiest cpu can't be moved to this_cpu
3912 if (!cpumask_test_cpu(this_cpu
,
3913 &busiest
->curr
->cpus_allowed
)) {
3914 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3916 goto out_one_pinned
;
3919 if (!busiest
->active_balance
) {
3920 busiest
->active_balance
= 1;
3921 busiest
->push_cpu
= this_cpu
;
3924 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3926 wake_up_process(busiest
->migration_thread
);
3929 * We've kicked active balancing, reset the failure
3932 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3935 sd
->nr_balance_failed
= 0;
3937 if (likely(!active_balance
)) {
3938 /* We were unbalanced, so reset the balancing interval */
3939 sd
->balance_interval
= sd
->min_interval
;
3942 * If we've begun active balancing, start to back off. This
3943 * case may not be covered by the all_pinned logic if there
3944 * is only 1 task on the busy runqueue (because we don't call
3947 if (sd
->balance_interval
< sd
->max_interval
)
3948 sd
->balance_interval
*= 2;
3951 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3952 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3958 schedstat_inc(sd
, lb_balanced
[idle
]);
3960 sd
->nr_balance_failed
= 0;
3963 /* tune up the balancing interval */
3964 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3965 (sd
->balance_interval
< sd
->max_interval
))
3966 sd
->balance_interval
*= 2;
3968 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3969 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3980 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3981 * tasks if there is an imbalance.
3983 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3984 * this_rq is locked.
3987 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
,
3988 struct cpumask
*cpus
)
3990 struct sched_group
*group
;
3991 struct rq
*busiest
= NULL
;
3992 unsigned long imbalance
;
3997 cpumask_setall(cpus
);
4000 * When power savings policy is enabled for the parent domain, idle
4001 * sibling can pick up load irrespective of busy siblings. In this case,
4002 * let the state of idle sibling percolate up as IDLE, instead of
4003 * portraying it as CPU_NOT_IDLE.
4005 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
4006 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4009 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
4011 update_shares_locked(this_rq
, sd
);
4012 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
4013 &sd_idle
, cpus
, NULL
);
4015 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
4019 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
4021 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
4025 BUG_ON(busiest
== this_rq
);
4027 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
4030 if (busiest
->nr_running
> 1) {
4031 /* Attempt to move tasks */
4032 double_lock_balance(this_rq
, busiest
);
4033 /* this_rq->clock is already updated */
4034 update_rq_clock(busiest
);
4035 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4036 imbalance
, sd
, CPU_NEWLY_IDLE
,
4038 double_unlock_balance(this_rq
, busiest
);
4040 if (unlikely(all_pinned
)) {
4041 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4042 if (!cpumask_empty(cpus
))
4048 int active_balance
= 0;
4050 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
4051 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4052 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4055 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
4058 if (sd
->nr_balance_failed
++ < 2)
4062 * The only task running in a non-idle cpu can be moved to this
4063 * cpu in an attempt to completely freeup the other CPU
4064 * package. The same method used to move task in load_balance()
4065 * have been extended for load_balance_newidle() to speedup
4066 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4068 * The package power saving logic comes from
4069 * find_busiest_group(). If there are no imbalance, then
4070 * f_b_g() will return NULL. However when sched_mc={1,2} then
4071 * f_b_g() will select a group from which a running task may be
4072 * pulled to this cpu in order to make the other package idle.
4073 * If there is no opportunity to make a package idle and if
4074 * there are no imbalance, then f_b_g() will return NULL and no
4075 * action will be taken in load_balance_newidle().
4077 * Under normal task pull operation due to imbalance, there
4078 * will be more than one task in the source run queue and
4079 * move_tasks() will succeed. ld_moved will be true and this
4080 * active balance code will not be triggered.
4083 /* Lock busiest in correct order while this_rq is held */
4084 double_lock_balance(this_rq
, busiest
);
4087 * don't kick the migration_thread, if the curr
4088 * task on busiest cpu can't be moved to this_cpu
4090 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
4091 double_unlock_balance(this_rq
, busiest
);
4096 if (!busiest
->active_balance
) {
4097 busiest
->active_balance
= 1;
4098 busiest
->push_cpu
= this_cpu
;
4102 double_unlock_balance(this_rq
, busiest
);
4104 * Should not call ttwu while holding a rq->lock
4106 spin_unlock(&this_rq
->lock
);
4108 wake_up_process(busiest
->migration_thread
);
4109 spin_lock(&this_rq
->lock
);
4112 sd
->nr_balance_failed
= 0;
4114 update_shares_locked(this_rq
, sd
);
4118 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
4119 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4120 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4122 sd
->nr_balance_failed
= 0;
4128 * idle_balance is called by schedule() if this_cpu is about to become
4129 * idle. Attempts to pull tasks from other CPUs.
4131 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
4133 struct sched_domain
*sd
;
4134 int pulled_task
= 0;
4135 unsigned long next_balance
= jiffies
+ HZ
;
4136 cpumask_var_t tmpmask
;
4138 if (!alloc_cpumask_var(&tmpmask
, GFP_ATOMIC
))
4141 for_each_domain(this_cpu
, sd
) {
4142 unsigned long interval
;
4144 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4147 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
4148 /* If we've pulled tasks over stop searching: */
4149 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
4152 interval
= msecs_to_jiffies(sd
->balance_interval
);
4153 if (time_after(next_balance
, sd
->last_balance
+ interval
))
4154 next_balance
= sd
->last_balance
+ interval
;
4158 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
4160 * We are going idle. next_balance may be set based on
4161 * a busy processor. So reset next_balance.
4163 this_rq
->next_balance
= next_balance
;
4165 free_cpumask_var(tmpmask
);
4169 * active_load_balance is run by migration threads. It pushes running tasks
4170 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4171 * running on each physical CPU where possible, and avoids physical /
4172 * logical imbalances.
4174 * Called with busiest_rq locked.
4176 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
4178 int target_cpu
= busiest_rq
->push_cpu
;
4179 struct sched_domain
*sd
;
4180 struct rq
*target_rq
;
4182 /* Is there any task to move? */
4183 if (busiest_rq
->nr_running
<= 1)
4186 target_rq
= cpu_rq(target_cpu
);
4189 * This condition is "impossible", if it occurs
4190 * we need to fix it. Originally reported by
4191 * Bjorn Helgaas on a 128-cpu setup.
4193 BUG_ON(busiest_rq
== target_rq
);
4195 /* move a task from busiest_rq to target_rq */
4196 double_lock_balance(busiest_rq
, target_rq
);
4197 update_rq_clock(busiest_rq
);
4198 update_rq_clock(target_rq
);
4200 /* Search for an sd spanning us and the target CPU. */
4201 for_each_domain(target_cpu
, sd
) {
4202 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
4203 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
4208 schedstat_inc(sd
, alb_count
);
4210 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
4212 schedstat_inc(sd
, alb_pushed
);
4214 schedstat_inc(sd
, alb_failed
);
4216 double_unlock_balance(busiest_rq
, target_rq
);
4221 atomic_t load_balancer
;
4222 cpumask_var_t cpu_mask
;
4223 } nohz ____cacheline_aligned
= {
4224 .load_balancer
= ATOMIC_INIT(-1),
4228 * This routine will try to nominate the ilb (idle load balancing)
4229 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4230 * load balancing on behalf of all those cpus. If all the cpus in the system
4231 * go into this tickless mode, then there will be no ilb owner (as there is
4232 * no need for one) and all the cpus will sleep till the next wakeup event
4235 * For the ilb owner, tick is not stopped. And this tick will be used
4236 * for idle load balancing. ilb owner will still be part of
4239 * While stopping the tick, this cpu will become the ilb owner if there
4240 * is no other owner. And will be the owner till that cpu becomes busy
4241 * or if all cpus in the system stop their ticks at which point
4242 * there is no need for ilb owner.
4244 * When the ilb owner becomes busy, it nominates another owner, during the
4245 * next busy scheduler_tick()
4247 int select_nohz_load_balancer(int stop_tick
)
4249 int cpu
= smp_processor_id();
4252 cpu_rq(cpu
)->in_nohz_recently
= 1;
4254 if (!cpu_active(cpu
)) {
4255 if (atomic_read(&nohz
.load_balancer
) != cpu
)
4259 * If we are going offline and still the leader,
4262 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4268 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
4270 /* time for ilb owner also to sleep */
4271 if (cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4272 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4273 atomic_set(&nohz
.load_balancer
, -1);
4277 if (atomic_read(&nohz
.load_balancer
) == -1) {
4278 /* make me the ilb owner */
4279 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
4281 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
4284 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4287 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4289 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4290 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4297 static DEFINE_SPINLOCK(balancing
);
4300 * It checks each scheduling domain to see if it is due to be balanced,
4301 * and initiates a balancing operation if so.
4303 * Balancing parameters are set up in arch_init_sched_domains.
4305 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
4308 struct rq
*rq
= cpu_rq(cpu
);
4309 unsigned long interval
;
4310 struct sched_domain
*sd
;
4311 /* Earliest time when we have to do rebalance again */
4312 unsigned long next_balance
= jiffies
+ 60*HZ
;
4313 int update_next_balance
= 0;
4317 /* Fails alloc? Rebalancing probably not a priority right now. */
4318 if (!alloc_cpumask_var(&tmp
, GFP_ATOMIC
))
4321 for_each_domain(cpu
, sd
) {
4322 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4325 interval
= sd
->balance_interval
;
4326 if (idle
!= CPU_IDLE
)
4327 interval
*= sd
->busy_factor
;
4329 /* scale ms to jiffies */
4330 interval
= msecs_to_jiffies(interval
);
4331 if (unlikely(!interval
))
4333 if (interval
> HZ
*NR_CPUS
/10)
4334 interval
= HZ
*NR_CPUS
/10;
4336 need_serialize
= sd
->flags
& SD_SERIALIZE
;
4338 if (need_serialize
) {
4339 if (!spin_trylock(&balancing
))
4343 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4344 if (load_balance(cpu
, rq
, sd
, idle
, &balance
, tmp
)) {
4346 * We've pulled tasks over so either we're no
4347 * longer idle, or one of our SMT siblings is
4350 idle
= CPU_NOT_IDLE
;
4352 sd
->last_balance
= jiffies
;
4355 spin_unlock(&balancing
);
4357 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4358 next_balance
= sd
->last_balance
+ interval
;
4359 update_next_balance
= 1;
4363 * Stop the load balance at this level. There is another
4364 * CPU in our sched group which is doing load balancing more
4372 * next_balance will be updated only when there is a need.
4373 * When the cpu is attached to null domain for ex, it will not be
4376 if (likely(update_next_balance
))
4377 rq
->next_balance
= next_balance
;
4379 free_cpumask_var(tmp
);
4383 * run_rebalance_domains is triggered when needed from the scheduler tick.
4384 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4385 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4387 static void run_rebalance_domains(struct softirq_action
*h
)
4389 int this_cpu
= smp_processor_id();
4390 struct rq
*this_rq
= cpu_rq(this_cpu
);
4391 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4392 CPU_IDLE
: CPU_NOT_IDLE
;
4394 rebalance_domains(this_cpu
, idle
);
4398 * If this cpu is the owner for idle load balancing, then do the
4399 * balancing on behalf of the other idle cpus whose ticks are
4402 if (this_rq
->idle_at_tick
&&
4403 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4407 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
4408 if (balance_cpu
== this_cpu
)
4412 * If this cpu gets work to do, stop the load balancing
4413 * work being done for other cpus. Next load
4414 * balancing owner will pick it up.
4419 rebalance_domains(balance_cpu
, CPU_IDLE
);
4421 rq
= cpu_rq(balance_cpu
);
4422 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4423 this_rq
->next_balance
= rq
->next_balance
;
4429 static inline int on_null_domain(int cpu
)
4431 return !rcu_dereference(cpu_rq(cpu
)->sd
);
4435 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4437 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4438 * idle load balancing owner or decide to stop the periodic load balancing,
4439 * if the whole system is idle.
4441 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4445 * If we were in the nohz mode recently and busy at the current
4446 * scheduler tick, then check if we need to nominate new idle
4449 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4450 rq
->in_nohz_recently
= 0;
4452 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4453 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4454 atomic_set(&nohz
.load_balancer
, -1);
4457 if (atomic_read(&nohz
.load_balancer
) == -1) {
4459 * simple selection for now: Nominate the
4460 * first cpu in the nohz list to be the next
4463 * TBD: Traverse the sched domains and nominate
4464 * the nearest cpu in the nohz.cpu_mask.
4466 int ilb
= cpumask_first(nohz
.cpu_mask
);
4468 if (ilb
< nr_cpu_ids
)
4474 * If this cpu is idle and doing idle load balancing for all the
4475 * cpus with ticks stopped, is it time for that to stop?
4477 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4478 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4484 * If this cpu is idle and the idle load balancing is done by
4485 * someone else, then no need raise the SCHED_SOFTIRQ
4487 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4488 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4491 /* Don't need to rebalance while attached to NULL domain */
4492 if (time_after_eq(jiffies
, rq
->next_balance
) &&
4493 likely(!on_null_domain(cpu
)))
4494 raise_softirq(SCHED_SOFTIRQ
);
4497 #else /* CONFIG_SMP */
4500 * on UP we do not need to balance between CPUs:
4502 static inline void idle_balance(int cpu
, struct rq
*rq
)
4508 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4510 EXPORT_PER_CPU_SYMBOL(kstat
);
4513 * Return any ns on the sched_clock that have not yet been banked in
4514 * @p in case that task is currently running.
4516 unsigned long long task_delta_exec(struct task_struct
*p
)
4518 unsigned long flags
;
4522 rq
= task_rq_lock(p
, &flags
);
4524 if (task_current(rq
, p
)) {
4527 update_rq_clock(rq
);
4528 delta_exec
= rq
->clock
- p
->se
.exec_start
;
4529 if ((s64
)delta_exec
> 0)
4533 task_rq_unlock(rq
, &flags
);
4539 * Account user cpu time to a process.
4540 * @p: the process that the cpu time gets accounted to
4541 * @cputime: the cpu time spent in user space since the last update
4542 * @cputime_scaled: cputime scaled by cpu frequency
4544 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
4545 cputime_t cputime_scaled
)
4547 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4550 /* Add user time to process. */
4551 p
->utime
= cputime_add(p
->utime
, cputime
);
4552 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4553 account_group_user_time(p
, cputime
);
4555 /* Add user time to cpustat. */
4556 tmp
= cputime_to_cputime64(cputime
);
4557 if (TASK_NICE(p
) > 0)
4558 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4560 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4561 /* Account for user time used */
4562 acct_update_integrals(p
);
4566 * Account guest cpu time to a process.
4567 * @p: the process that the cpu time gets accounted to
4568 * @cputime: the cpu time spent in virtual machine since the last update
4569 * @cputime_scaled: cputime scaled by cpu frequency
4571 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
4572 cputime_t cputime_scaled
)
4575 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4577 tmp
= cputime_to_cputime64(cputime
);
4579 /* Add guest time to process. */
4580 p
->utime
= cputime_add(p
->utime
, cputime
);
4581 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4582 account_group_user_time(p
, cputime
);
4583 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4585 /* Add guest time to cpustat. */
4586 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4587 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4591 * Account system cpu time to a process.
4592 * @p: the process that the cpu time gets accounted to
4593 * @hardirq_offset: the offset to subtract from hardirq_count()
4594 * @cputime: the cpu time spent in kernel space since the last update
4595 * @cputime_scaled: cputime scaled by cpu frequency
4597 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4598 cputime_t cputime
, cputime_t cputime_scaled
)
4600 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4603 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4604 account_guest_time(p
, cputime
, cputime_scaled
);
4608 /* Add system time to process. */
4609 p
->stime
= cputime_add(p
->stime
, cputime
);
4610 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
4611 account_group_system_time(p
, cputime
);
4613 /* Add system time to cpustat. */
4614 tmp
= cputime_to_cputime64(cputime
);
4615 if (hardirq_count() - hardirq_offset
)
4616 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4617 else if (softirq_count())
4618 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4620 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4622 /* Account for system time used */
4623 acct_update_integrals(p
);
4627 * Account for involuntary wait time.
4628 * @steal: the cpu time spent in involuntary wait
4630 void account_steal_time(cputime_t cputime
)
4632 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4633 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
4635 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
4639 * Account for idle time.
4640 * @cputime: the cpu time spent in idle wait
4642 void account_idle_time(cputime_t cputime
)
4644 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4645 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
4646 struct rq
*rq
= this_rq();
4648 if (atomic_read(&rq
->nr_iowait
) > 0)
4649 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
4651 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
4654 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4657 * Account a single tick of cpu time.
4658 * @p: the process that the cpu time gets accounted to
4659 * @user_tick: indicates if the tick is a user or a system tick
4661 void account_process_tick(struct task_struct
*p
, int user_tick
)
4663 cputime_t one_jiffy
= jiffies_to_cputime(1);
4664 cputime_t one_jiffy_scaled
= cputime_to_scaled(one_jiffy
);
4665 struct rq
*rq
= this_rq();
4668 account_user_time(p
, one_jiffy
, one_jiffy_scaled
);
4669 else if (p
!= rq
->idle
)
4670 account_system_time(p
, HARDIRQ_OFFSET
, one_jiffy
,
4673 account_idle_time(one_jiffy
);
4677 * Account multiple ticks of steal time.
4678 * @p: the process from which the cpu time has been stolen
4679 * @ticks: number of stolen ticks
4681 void account_steal_ticks(unsigned long ticks
)
4683 account_steal_time(jiffies_to_cputime(ticks
));
4687 * Account multiple ticks of idle time.
4688 * @ticks: number of stolen ticks
4690 void account_idle_ticks(unsigned long ticks
)
4692 account_idle_time(jiffies_to_cputime(ticks
));
4698 * Use precise platform statistics if available:
4700 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4701 cputime_t
task_utime(struct task_struct
*p
)
4706 cputime_t
task_stime(struct task_struct
*p
)
4711 cputime_t
task_utime(struct task_struct
*p
)
4713 clock_t utime
= cputime_to_clock_t(p
->utime
),
4714 total
= utime
+ cputime_to_clock_t(p
->stime
);
4718 * Use CFS's precise accounting:
4720 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
4724 do_div(temp
, total
);
4726 utime
= (clock_t)temp
;
4728 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
4729 return p
->prev_utime
;
4732 cputime_t
task_stime(struct task_struct
*p
)
4737 * Use CFS's precise accounting. (we subtract utime from
4738 * the total, to make sure the total observed by userspace
4739 * grows monotonically - apps rely on that):
4741 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
4742 cputime_to_clock_t(task_utime(p
));
4745 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
4747 return p
->prev_stime
;
4751 inline cputime_t
task_gtime(struct task_struct
*p
)
4757 * This function gets called by the timer code, with HZ frequency.
4758 * We call it with interrupts disabled.
4760 * It also gets called by the fork code, when changing the parent's
4763 void scheduler_tick(void)
4765 int cpu
= smp_processor_id();
4766 struct rq
*rq
= cpu_rq(cpu
);
4767 struct task_struct
*curr
= rq
->curr
;
4771 spin_lock(&rq
->lock
);
4772 update_rq_clock(rq
);
4773 update_cpu_load(rq
);
4774 curr
->sched_class
->task_tick(rq
, curr
, 0);
4775 spin_unlock(&rq
->lock
);
4778 rq
->idle_at_tick
= idle_cpu(cpu
);
4779 trigger_load_balance(rq
, cpu
);
4783 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4784 defined(CONFIG_PREEMPT_TRACER))
4786 static inline unsigned long get_parent_ip(unsigned long addr
)
4788 if (in_lock_functions(addr
)) {
4789 addr
= CALLER_ADDR2
;
4790 if (in_lock_functions(addr
))
4791 addr
= CALLER_ADDR3
;
4796 void __kprobes
add_preempt_count(int val
)
4798 #ifdef CONFIG_DEBUG_PREEMPT
4802 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4805 preempt_count() += val
;
4806 #ifdef CONFIG_DEBUG_PREEMPT
4808 * Spinlock count overflowing soon?
4810 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4813 if (preempt_count() == val
)
4814 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4816 EXPORT_SYMBOL(add_preempt_count
);
4818 void __kprobes
sub_preempt_count(int val
)
4820 #ifdef CONFIG_DEBUG_PREEMPT
4824 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4827 * Is the spinlock portion underflowing?
4829 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4830 !(preempt_count() & PREEMPT_MASK
)))
4834 if (preempt_count() == val
)
4835 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4836 preempt_count() -= val
;
4838 EXPORT_SYMBOL(sub_preempt_count
);
4843 * Print scheduling while atomic bug:
4845 static noinline
void __schedule_bug(struct task_struct
*prev
)
4847 struct pt_regs
*regs
= get_irq_regs();
4849 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4850 prev
->comm
, prev
->pid
, preempt_count());
4852 debug_show_held_locks(prev
);
4854 if (irqs_disabled())
4855 print_irqtrace_events(prev
);
4864 * Various schedule()-time debugging checks and statistics:
4866 static inline void schedule_debug(struct task_struct
*prev
)
4869 * Test if we are atomic. Since do_exit() needs to call into
4870 * schedule() atomically, we ignore that path for now.
4871 * Otherwise, whine if we are scheduling when we should not be.
4873 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4874 __schedule_bug(prev
);
4876 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4878 schedstat_inc(this_rq(), sched_count
);
4879 #ifdef CONFIG_SCHEDSTATS
4880 if (unlikely(prev
->lock_depth
>= 0)) {
4881 schedstat_inc(this_rq(), bkl_count
);
4882 schedstat_inc(prev
, sched_info
.bkl_count
);
4887 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
4889 if (prev
->state
== TASK_RUNNING
) {
4890 u64 runtime
= prev
->se
.sum_exec_runtime
;
4892 runtime
-= prev
->se
.prev_sum_exec_runtime
;
4893 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
4896 * In order to avoid avg_overlap growing stale when we are
4897 * indeed overlapping and hence not getting put to sleep, grow
4898 * the avg_overlap on preemption.
4900 * We use the average preemption runtime because that
4901 * correlates to the amount of cache footprint a task can
4904 update_avg(&prev
->se
.avg_overlap
, runtime
);
4906 prev
->sched_class
->put_prev_task(rq
, prev
);
4910 * Pick up the highest-prio task:
4912 static inline struct task_struct
*
4913 pick_next_task(struct rq
*rq
)
4915 const struct sched_class
*class;
4916 struct task_struct
*p
;
4919 * Optimization: we know that if all tasks are in
4920 * the fair class we can call that function directly:
4922 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4923 p
= fair_sched_class
.pick_next_task(rq
);
4928 class = sched_class_highest
;
4930 p
= class->pick_next_task(rq
);
4934 * Will never be NULL as the idle class always
4935 * returns a non-NULL p:
4937 class = class->next
;
4942 * schedule() is the main scheduler function.
4944 asmlinkage
void __sched
schedule(void)
4946 struct task_struct
*prev
, *next
;
4947 unsigned long *switch_count
;
4953 cpu
= smp_processor_id();
4957 switch_count
= &prev
->nivcsw
;
4959 release_kernel_lock(prev
);
4960 need_resched_nonpreemptible
:
4962 schedule_debug(prev
);
4964 if (sched_feat(HRTICK
))
4967 spin_lock_irq(&rq
->lock
);
4968 update_rq_clock(rq
);
4969 clear_tsk_need_resched(prev
);
4971 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4972 if (unlikely(signal_pending_state(prev
->state
, prev
)))
4973 prev
->state
= TASK_RUNNING
;
4975 deactivate_task(rq
, prev
, 1);
4976 switch_count
= &prev
->nvcsw
;
4980 if (prev
->sched_class
->pre_schedule
)
4981 prev
->sched_class
->pre_schedule(rq
, prev
);
4984 if (unlikely(!rq
->nr_running
))
4985 idle_balance(cpu
, rq
);
4987 put_prev_task(rq
, prev
);
4988 next
= pick_next_task(rq
);
4990 if (likely(prev
!= next
)) {
4991 sched_info_switch(prev
, next
);
4997 context_switch(rq
, prev
, next
); /* unlocks the rq */
4999 * the context switch might have flipped the stack from under
5000 * us, hence refresh the local variables.
5002 cpu
= smp_processor_id();
5005 spin_unlock_irq(&rq
->lock
);
5007 if (unlikely(reacquire_kernel_lock(current
) < 0))
5008 goto need_resched_nonpreemptible
;
5010 preempt_enable_no_resched();
5011 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
5014 EXPORT_SYMBOL(schedule
);
5016 #ifdef CONFIG_PREEMPT
5018 * this is the entry point to schedule() from in-kernel preemption
5019 * off of preempt_enable. Kernel preemptions off return from interrupt
5020 * occur there and call schedule directly.
5022 asmlinkage
void __sched
preempt_schedule(void)
5024 struct thread_info
*ti
= current_thread_info();
5027 * If there is a non-zero preempt_count or interrupts are disabled,
5028 * we do not want to preempt the current task. Just return..
5030 if (likely(ti
->preempt_count
|| irqs_disabled()))
5034 add_preempt_count(PREEMPT_ACTIVE
);
5036 sub_preempt_count(PREEMPT_ACTIVE
);
5039 * Check again in case we missed a preemption opportunity
5040 * between schedule and now.
5043 } while (need_resched());
5045 EXPORT_SYMBOL(preempt_schedule
);
5048 * this is the entry point to schedule() from kernel preemption
5049 * off of irq context.
5050 * Note, that this is called and return with irqs disabled. This will
5051 * protect us against recursive calling from irq.
5053 asmlinkage
void __sched
preempt_schedule_irq(void)
5055 struct thread_info
*ti
= current_thread_info();
5057 /* Catch callers which need to be fixed */
5058 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
5061 add_preempt_count(PREEMPT_ACTIVE
);
5064 local_irq_disable();
5065 sub_preempt_count(PREEMPT_ACTIVE
);
5068 * Check again in case we missed a preemption opportunity
5069 * between schedule and now.
5072 } while (need_resched());
5075 #endif /* CONFIG_PREEMPT */
5077 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
5080 return try_to_wake_up(curr
->private, mode
, sync
);
5082 EXPORT_SYMBOL(default_wake_function
);
5085 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5086 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5087 * number) then we wake all the non-exclusive tasks and one exclusive task.
5089 * There are circumstances in which we can try to wake a task which has already
5090 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5091 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5093 void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
5094 int nr_exclusive
, int sync
, void *key
)
5096 wait_queue_t
*curr
, *next
;
5098 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
5099 unsigned flags
= curr
->flags
;
5101 if (curr
->func(curr
, mode
, sync
, key
) &&
5102 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
5108 * __wake_up - wake up threads blocked on a waitqueue.
5110 * @mode: which threads
5111 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5112 * @key: is directly passed to the wakeup function
5114 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
5115 int nr_exclusive
, void *key
)
5117 unsigned long flags
;
5119 spin_lock_irqsave(&q
->lock
, flags
);
5120 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
5121 spin_unlock_irqrestore(&q
->lock
, flags
);
5123 EXPORT_SYMBOL(__wake_up
);
5126 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5128 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
5130 __wake_up_common(q
, mode
, 1, 0, NULL
);
5134 * __wake_up_sync - wake up threads blocked on a waitqueue.
5136 * @mode: which threads
5137 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5139 * The sync wakeup differs that the waker knows that it will schedule
5140 * away soon, so while the target thread will be woken up, it will not
5141 * be migrated to another CPU - ie. the two threads are 'synchronized'
5142 * with each other. This can prevent needless bouncing between CPUs.
5144 * On UP it can prevent extra preemption.
5147 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
5149 unsigned long flags
;
5155 if (unlikely(!nr_exclusive
))
5158 spin_lock_irqsave(&q
->lock
, flags
);
5159 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
5160 spin_unlock_irqrestore(&q
->lock
, flags
);
5162 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
5165 * complete: - signals a single thread waiting on this completion
5166 * @x: holds the state of this particular completion
5168 * This will wake up a single thread waiting on this completion. Threads will be
5169 * awakened in the same order in which they were queued.
5171 * See also complete_all(), wait_for_completion() and related routines.
5173 void complete(struct completion
*x
)
5175 unsigned long flags
;
5177 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5179 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
5180 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5182 EXPORT_SYMBOL(complete
);
5185 * complete_all: - signals all threads waiting on this completion
5186 * @x: holds the state of this particular completion
5188 * This will wake up all threads waiting on this particular completion event.
5190 void complete_all(struct completion
*x
)
5192 unsigned long flags
;
5194 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5195 x
->done
+= UINT_MAX
/2;
5196 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
5197 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5199 EXPORT_SYMBOL(complete_all
);
5201 static inline long __sched
5202 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
5205 DECLARE_WAITQUEUE(wait
, current
);
5207 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
5208 __add_wait_queue_tail(&x
->wait
, &wait
);
5210 if (signal_pending_state(state
, current
)) {
5211 timeout
= -ERESTARTSYS
;
5214 __set_current_state(state
);
5215 spin_unlock_irq(&x
->wait
.lock
);
5216 timeout
= schedule_timeout(timeout
);
5217 spin_lock_irq(&x
->wait
.lock
);
5218 } while (!x
->done
&& timeout
);
5219 __remove_wait_queue(&x
->wait
, &wait
);
5224 return timeout
?: 1;
5228 wait_for_common(struct completion
*x
, long timeout
, int state
)
5232 spin_lock_irq(&x
->wait
.lock
);
5233 timeout
= do_wait_for_common(x
, timeout
, state
);
5234 spin_unlock_irq(&x
->wait
.lock
);
5239 * wait_for_completion: - waits for completion of a task
5240 * @x: holds the state of this particular completion
5242 * This waits to be signaled for completion of a specific task. It is NOT
5243 * interruptible and there is no timeout.
5245 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5246 * and interrupt capability. Also see complete().
5248 void __sched
wait_for_completion(struct completion
*x
)
5250 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
5252 EXPORT_SYMBOL(wait_for_completion
);
5255 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5256 * @x: holds the state of this particular completion
5257 * @timeout: timeout value in jiffies
5259 * This waits for either a completion of a specific task to be signaled or for a
5260 * specified timeout to expire. The timeout is in jiffies. It is not
5263 unsigned long __sched
5264 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
5266 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
5268 EXPORT_SYMBOL(wait_for_completion_timeout
);
5271 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5272 * @x: holds the state of this particular completion
5274 * This waits for completion of a specific task to be signaled. It is
5277 int __sched
wait_for_completion_interruptible(struct completion
*x
)
5279 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
5280 if (t
== -ERESTARTSYS
)
5284 EXPORT_SYMBOL(wait_for_completion_interruptible
);
5287 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5288 * @x: holds the state of this particular completion
5289 * @timeout: timeout value in jiffies
5291 * This waits for either a completion of a specific task to be signaled or for a
5292 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5294 unsigned long __sched
5295 wait_for_completion_interruptible_timeout(struct completion
*x
,
5296 unsigned long timeout
)
5298 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
5300 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
5303 * wait_for_completion_killable: - waits for completion of a task (killable)
5304 * @x: holds the state of this particular completion
5306 * This waits to be signaled for completion of a specific task. It can be
5307 * interrupted by a kill signal.
5309 int __sched
wait_for_completion_killable(struct completion
*x
)
5311 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
5312 if (t
== -ERESTARTSYS
)
5316 EXPORT_SYMBOL(wait_for_completion_killable
);
5319 * try_wait_for_completion - try to decrement a completion without blocking
5320 * @x: completion structure
5322 * Returns: 0 if a decrement cannot be done without blocking
5323 * 1 if a decrement succeeded.
5325 * If a completion is being used as a counting completion,
5326 * attempt to decrement the counter without blocking. This
5327 * enables us to avoid waiting if the resource the completion
5328 * is protecting is not available.
5330 bool try_wait_for_completion(struct completion
*x
)
5334 spin_lock_irq(&x
->wait
.lock
);
5339 spin_unlock_irq(&x
->wait
.lock
);
5342 EXPORT_SYMBOL(try_wait_for_completion
);
5345 * completion_done - Test to see if a completion has any waiters
5346 * @x: completion structure
5348 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5349 * 1 if there are no waiters.
5352 bool completion_done(struct completion
*x
)
5356 spin_lock_irq(&x
->wait
.lock
);
5359 spin_unlock_irq(&x
->wait
.lock
);
5362 EXPORT_SYMBOL(completion_done
);
5365 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
5367 unsigned long flags
;
5370 init_waitqueue_entry(&wait
, current
);
5372 __set_current_state(state
);
5374 spin_lock_irqsave(&q
->lock
, flags
);
5375 __add_wait_queue(q
, &wait
);
5376 spin_unlock(&q
->lock
);
5377 timeout
= schedule_timeout(timeout
);
5378 spin_lock_irq(&q
->lock
);
5379 __remove_wait_queue(q
, &wait
);
5380 spin_unlock_irqrestore(&q
->lock
, flags
);
5385 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
5387 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5389 EXPORT_SYMBOL(interruptible_sleep_on
);
5392 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5394 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
5396 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
5398 void __sched
sleep_on(wait_queue_head_t
*q
)
5400 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5402 EXPORT_SYMBOL(sleep_on
);
5404 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5406 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
5408 EXPORT_SYMBOL(sleep_on_timeout
);
5410 #ifdef CONFIG_RT_MUTEXES
5413 * rt_mutex_setprio - set the current priority of a task
5415 * @prio: prio value (kernel-internal form)
5417 * This function changes the 'effective' priority of a task. It does
5418 * not touch ->normal_prio like __setscheduler().
5420 * Used by the rt_mutex code to implement priority inheritance logic.
5422 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
5424 unsigned long flags
;
5425 int oldprio
, on_rq
, running
;
5427 const struct sched_class
*prev_class
= p
->sched_class
;
5429 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
5431 rq
= task_rq_lock(p
, &flags
);
5432 update_rq_clock(rq
);
5435 on_rq
= p
->se
.on_rq
;
5436 running
= task_current(rq
, p
);
5438 dequeue_task(rq
, p
, 0);
5440 p
->sched_class
->put_prev_task(rq
, p
);
5443 p
->sched_class
= &rt_sched_class
;
5445 p
->sched_class
= &fair_sched_class
;
5450 p
->sched_class
->set_curr_task(rq
);
5452 enqueue_task(rq
, p
, 0);
5454 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5456 task_rq_unlock(rq
, &flags
);
5461 void set_user_nice(struct task_struct
*p
, long nice
)
5463 int old_prio
, delta
, on_rq
;
5464 unsigned long flags
;
5467 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
5470 * We have to be careful, if called from sys_setpriority(),
5471 * the task might be in the middle of scheduling on another CPU.
5473 rq
= task_rq_lock(p
, &flags
);
5474 update_rq_clock(rq
);
5476 * The RT priorities are set via sched_setscheduler(), but we still
5477 * allow the 'normal' nice value to be set - but as expected
5478 * it wont have any effect on scheduling until the task is
5479 * SCHED_FIFO/SCHED_RR:
5481 if (task_has_rt_policy(p
)) {
5482 p
->static_prio
= NICE_TO_PRIO(nice
);
5485 on_rq
= p
->se
.on_rq
;
5487 dequeue_task(rq
, p
, 0);
5489 p
->static_prio
= NICE_TO_PRIO(nice
);
5492 p
->prio
= effective_prio(p
);
5493 delta
= p
->prio
- old_prio
;
5496 enqueue_task(rq
, p
, 0);
5498 * If the task increased its priority or is running and
5499 * lowered its priority, then reschedule its CPU:
5501 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
5502 resched_task(rq
->curr
);
5505 task_rq_unlock(rq
, &flags
);
5507 EXPORT_SYMBOL(set_user_nice
);
5510 * can_nice - check if a task can reduce its nice value
5514 int can_nice(const struct task_struct
*p
, const int nice
)
5516 /* convert nice value [19,-20] to rlimit style value [1,40] */
5517 int nice_rlim
= 20 - nice
;
5519 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
5520 capable(CAP_SYS_NICE
));
5523 #ifdef __ARCH_WANT_SYS_NICE
5526 * sys_nice - change the priority of the current process.
5527 * @increment: priority increment
5529 * sys_setpriority is a more generic, but much slower function that
5530 * does similar things.
5532 SYSCALL_DEFINE1(nice
, int, increment
)
5537 * Setpriority might change our priority at the same moment.
5538 * We don't have to worry. Conceptually one call occurs first
5539 * and we have a single winner.
5541 if (increment
< -40)
5546 nice
= TASK_NICE(current
) + increment
;
5552 if (increment
< 0 && !can_nice(current
, nice
))
5555 retval
= security_task_setnice(current
, nice
);
5559 set_user_nice(current
, nice
);
5566 * task_prio - return the priority value of a given task.
5567 * @p: the task in question.
5569 * This is the priority value as seen by users in /proc.
5570 * RT tasks are offset by -200. Normal tasks are centered
5571 * around 0, value goes from -16 to +15.
5573 int task_prio(const struct task_struct
*p
)
5575 return p
->prio
- MAX_RT_PRIO
;
5579 * task_nice - return the nice value of a given task.
5580 * @p: the task in question.
5582 int task_nice(const struct task_struct
*p
)
5584 return TASK_NICE(p
);
5586 EXPORT_SYMBOL(task_nice
);
5589 * idle_cpu - is a given cpu idle currently?
5590 * @cpu: the processor in question.
5592 int idle_cpu(int cpu
)
5594 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
5598 * idle_task - return the idle task for a given cpu.
5599 * @cpu: the processor in question.
5601 struct task_struct
*idle_task(int cpu
)
5603 return cpu_rq(cpu
)->idle
;
5607 * find_process_by_pid - find a process with a matching PID value.
5608 * @pid: the pid in question.
5610 static struct task_struct
*find_process_by_pid(pid_t pid
)
5612 return pid
? find_task_by_vpid(pid
) : current
;
5615 /* Actually do priority change: must hold rq lock. */
5617 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
5619 BUG_ON(p
->se
.on_rq
);
5622 switch (p
->policy
) {
5626 p
->sched_class
= &fair_sched_class
;
5630 p
->sched_class
= &rt_sched_class
;
5634 p
->rt_priority
= prio
;
5635 p
->normal_prio
= normal_prio(p
);
5636 /* we are holding p->pi_lock already */
5637 p
->prio
= rt_mutex_getprio(p
);
5642 * check the target process has a UID that matches the current process's
5644 static bool check_same_owner(struct task_struct
*p
)
5646 const struct cred
*cred
= current_cred(), *pcred
;
5650 pcred
= __task_cred(p
);
5651 match
= (cred
->euid
== pcred
->euid
||
5652 cred
->euid
== pcred
->uid
);
5657 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
5658 struct sched_param
*param
, bool user
)
5660 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
5661 unsigned long flags
;
5662 const struct sched_class
*prev_class
= p
->sched_class
;
5665 /* may grab non-irq protected spin_locks */
5666 BUG_ON(in_interrupt());
5668 /* double check policy once rq lock held */
5670 policy
= oldpolicy
= p
->policy
;
5671 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
5672 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
5673 policy
!= SCHED_IDLE
)
5676 * Valid priorities for SCHED_FIFO and SCHED_RR are
5677 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5678 * SCHED_BATCH and SCHED_IDLE is 0.
5680 if (param
->sched_priority
< 0 ||
5681 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5682 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
5684 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
5688 * Allow unprivileged RT tasks to decrease priority:
5690 if (user
&& !capable(CAP_SYS_NICE
)) {
5691 if (rt_policy(policy
)) {
5692 unsigned long rlim_rtprio
;
5694 if (!lock_task_sighand(p
, &flags
))
5696 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
5697 unlock_task_sighand(p
, &flags
);
5699 /* can't set/change the rt policy */
5700 if (policy
!= p
->policy
&& !rlim_rtprio
)
5703 /* can't increase priority */
5704 if (param
->sched_priority
> p
->rt_priority
&&
5705 param
->sched_priority
> rlim_rtprio
)
5709 * Like positive nice levels, dont allow tasks to
5710 * move out of SCHED_IDLE either:
5712 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
5715 /* can't change other user's priorities */
5716 if (!check_same_owner(p
))
5721 #ifdef CONFIG_RT_GROUP_SCHED
5723 * Do not allow realtime tasks into groups that have no runtime
5726 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
5727 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
5731 retval
= security_task_setscheduler(p
, policy
, param
);
5737 * make sure no PI-waiters arrive (or leave) while we are
5738 * changing the priority of the task:
5740 spin_lock_irqsave(&p
->pi_lock
, flags
);
5742 * To be able to change p->policy safely, the apropriate
5743 * runqueue lock must be held.
5745 rq
= __task_rq_lock(p
);
5746 /* recheck policy now with rq lock held */
5747 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5748 policy
= oldpolicy
= -1;
5749 __task_rq_unlock(rq
);
5750 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5753 update_rq_clock(rq
);
5754 on_rq
= p
->se
.on_rq
;
5755 running
= task_current(rq
, p
);
5757 deactivate_task(rq
, p
, 0);
5759 p
->sched_class
->put_prev_task(rq
, p
);
5762 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5765 p
->sched_class
->set_curr_task(rq
);
5767 activate_task(rq
, p
, 0);
5769 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5771 __task_rq_unlock(rq
);
5772 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5774 rt_mutex_adjust_pi(p
);
5780 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5781 * @p: the task in question.
5782 * @policy: new policy.
5783 * @param: structure containing the new RT priority.
5785 * NOTE that the task may be already dead.
5787 int sched_setscheduler(struct task_struct
*p
, int policy
,
5788 struct sched_param
*param
)
5790 return __sched_setscheduler(p
, policy
, param
, true);
5792 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5795 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5796 * @p: the task in question.
5797 * @policy: new policy.
5798 * @param: structure containing the new RT priority.
5800 * Just like sched_setscheduler, only don't bother checking if the
5801 * current context has permission. For example, this is needed in
5802 * stop_machine(): we create temporary high priority worker threads,
5803 * but our caller might not have that capability.
5805 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5806 struct sched_param
*param
)
5808 return __sched_setscheduler(p
, policy
, param
, false);
5812 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5814 struct sched_param lparam
;
5815 struct task_struct
*p
;
5818 if (!param
|| pid
< 0)
5820 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5825 p
= find_process_by_pid(pid
);
5827 retval
= sched_setscheduler(p
, policy
, &lparam
);
5834 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5835 * @pid: the pid in question.
5836 * @policy: new policy.
5837 * @param: structure containing the new RT priority.
5839 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
5840 struct sched_param __user
*, param
)
5842 /* negative values for policy are not valid */
5846 return do_sched_setscheduler(pid
, policy
, param
);
5850 * sys_sched_setparam - set/change the RT priority of a thread
5851 * @pid: the pid in question.
5852 * @param: structure containing the new RT priority.
5854 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5856 return do_sched_setscheduler(pid
, -1, param
);
5860 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5861 * @pid: the pid in question.
5863 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
5865 struct task_struct
*p
;
5872 read_lock(&tasklist_lock
);
5873 p
= find_process_by_pid(pid
);
5875 retval
= security_task_getscheduler(p
);
5879 read_unlock(&tasklist_lock
);
5884 * sys_sched_getscheduler - get the RT priority of a thread
5885 * @pid: the pid in question.
5886 * @param: structure containing the RT priority.
5888 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5890 struct sched_param lp
;
5891 struct task_struct
*p
;
5894 if (!param
|| pid
< 0)
5897 read_lock(&tasklist_lock
);
5898 p
= find_process_by_pid(pid
);
5903 retval
= security_task_getscheduler(p
);
5907 lp
.sched_priority
= p
->rt_priority
;
5908 read_unlock(&tasklist_lock
);
5911 * This one might sleep, we cannot do it with a spinlock held ...
5913 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5918 read_unlock(&tasklist_lock
);
5922 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
5924 cpumask_var_t cpus_allowed
, new_mask
;
5925 struct task_struct
*p
;
5929 read_lock(&tasklist_lock
);
5931 p
= find_process_by_pid(pid
);
5933 read_unlock(&tasklist_lock
);
5939 * It is not safe to call set_cpus_allowed with the
5940 * tasklist_lock held. We will bump the task_struct's
5941 * usage count and then drop tasklist_lock.
5944 read_unlock(&tasklist_lock
);
5946 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
5950 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
5952 goto out_free_cpus_allowed
;
5955 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
5958 retval
= security_task_setscheduler(p
, 0, NULL
);
5962 cpuset_cpus_allowed(p
, cpus_allowed
);
5963 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
5965 retval
= set_cpus_allowed_ptr(p
, new_mask
);
5968 cpuset_cpus_allowed(p
, cpus_allowed
);
5969 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
5971 * We must have raced with a concurrent cpuset
5972 * update. Just reset the cpus_allowed to the
5973 * cpuset's cpus_allowed
5975 cpumask_copy(new_mask
, cpus_allowed
);
5980 free_cpumask_var(new_mask
);
5981 out_free_cpus_allowed
:
5982 free_cpumask_var(cpus_allowed
);
5989 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5990 struct cpumask
*new_mask
)
5992 if (len
< cpumask_size())
5993 cpumask_clear(new_mask
);
5994 else if (len
> cpumask_size())
5995 len
= cpumask_size();
5997 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
6001 * sys_sched_setaffinity - set the cpu affinity of a process
6002 * @pid: pid of the process
6003 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6004 * @user_mask_ptr: user-space pointer to the new cpu mask
6006 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
6007 unsigned long __user
*, user_mask_ptr
)
6009 cpumask_var_t new_mask
;
6012 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
6015 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
6017 retval
= sched_setaffinity(pid
, new_mask
);
6018 free_cpumask_var(new_mask
);
6022 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
6024 struct task_struct
*p
;
6028 read_lock(&tasklist_lock
);
6031 p
= find_process_by_pid(pid
);
6035 retval
= security_task_getscheduler(p
);
6039 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
6042 read_unlock(&tasklist_lock
);
6049 * sys_sched_getaffinity - get the cpu affinity of a process
6050 * @pid: pid of the process
6051 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6052 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6054 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
6055 unsigned long __user
*, user_mask_ptr
)
6060 if (len
< cpumask_size())
6063 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
6066 ret
= sched_getaffinity(pid
, mask
);
6068 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
6071 ret
= cpumask_size();
6073 free_cpumask_var(mask
);
6079 * sys_sched_yield - yield the current processor to other threads.
6081 * This function yields the current CPU to other tasks. If there are no
6082 * other threads running on this CPU then this function will return.
6084 SYSCALL_DEFINE0(sched_yield
)
6086 struct rq
*rq
= this_rq_lock();
6088 schedstat_inc(rq
, yld_count
);
6089 current
->sched_class
->yield_task(rq
);
6092 * Since we are going to call schedule() anyway, there's
6093 * no need to preempt or enable interrupts:
6095 __release(rq
->lock
);
6096 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
6097 _raw_spin_unlock(&rq
->lock
);
6098 preempt_enable_no_resched();
6105 static void __cond_resched(void)
6107 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6108 __might_sleep(__FILE__
, __LINE__
);
6111 * The BKS might be reacquired before we have dropped
6112 * PREEMPT_ACTIVE, which could trigger a second
6113 * cond_resched() call.
6116 add_preempt_count(PREEMPT_ACTIVE
);
6118 sub_preempt_count(PREEMPT_ACTIVE
);
6119 } while (need_resched());
6122 int __sched
_cond_resched(void)
6124 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
6125 system_state
== SYSTEM_RUNNING
) {
6131 EXPORT_SYMBOL(_cond_resched
);
6134 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
6135 * call schedule, and on return reacquire the lock.
6137 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6138 * operations here to prevent schedule() from being called twice (once via
6139 * spin_unlock(), once by hand).
6141 int cond_resched_lock(spinlock_t
*lock
)
6143 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
6146 if (spin_needbreak(lock
) || resched
) {
6148 if (resched
&& need_resched())
6157 EXPORT_SYMBOL(cond_resched_lock
);
6159 int __sched
cond_resched_softirq(void)
6161 BUG_ON(!in_softirq());
6163 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
6171 EXPORT_SYMBOL(cond_resched_softirq
);
6174 * yield - yield the current processor to other threads.
6176 * This is a shortcut for kernel-space yielding - it marks the
6177 * thread runnable and calls sys_sched_yield().
6179 void __sched
yield(void)
6181 set_current_state(TASK_RUNNING
);
6184 EXPORT_SYMBOL(yield
);
6187 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6188 * that process accounting knows that this is a task in IO wait state.
6190 * But don't do that if it is a deliberate, throttling IO wait (this task
6191 * has set its backing_dev_info: the queue against which it should throttle)
6193 void __sched
io_schedule(void)
6195 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
6197 delayacct_blkio_start();
6198 atomic_inc(&rq
->nr_iowait
);
6200 atomic_dec(&rq
->nr_iowait
);
6201 delayacct_blkio_end();
6203 EXPORT_SYMBOL(io_schedule
);
6205 long __sched
io_schedule_timeout(long timeout
)
6207 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
6210 delayacct_blkio_start();
6211 atomic_inc(&rq
->nr_iowait
);
6212 ret
= schedule_timeout(timeout
);
6213 atomic_dec(&rq
->nr_iowait
);
6214 delayacct_blkio_end();
6219 * sys_sched_get_priority_max - return maximum RT priority.
6220 * @policy: scheduling class.
6222 * this syscall returns the maximum rt_priority that can be used
6223 * by a given scheduling class.
6225 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
6232 ret
= MAX_USER_RT_PRIO
-1;
6244 * sys_sched_get_priority_min - return minimum RT priority.
6245 * @policy: scheduling class.
6247 * this syscall returns the minimum rt_priority that can be used
6248 * by a given scheduling class.
6250 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
6268 * sys_sched_rr_get_interval - return the default timeslice of a process.
6269 * @pid: pid of the process.
6270 * @interval: userspace pointer to the timeslice value.
6272 * this syscall writes the default timeslice value of a given process
6273 * into the user-space timespec buffer. A value of '0' means infinity.
6275 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
6276 struct timespec __user
*, interval
)
6278 struct task_struct
*p
;
6279 unsigned int time_slice
;
6287 read_lock(&tasklist_lock
);
6288 p
= find_process_by_pid(pid
);
6292 retval
= security_task_getscheduler(p
);
6297 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6298 * tasks that are on an otherwise idle runqueue:
6301 if (p
->policy
== SCHED_RR
) {
6302 time_slice
= DEF_TIMESLICE
;
6303 } else if (p
->policy
!= SCHED_FIFO
) {
6304 struct sched_entity
*se
= &p
->se
;
6305 unsigned long flags
;
6308 rq
= task_rq_lock(p
, &flags
);
6309 if (rq
->cfs
.load
.weight
)
6310 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
6311 task_rq_unlock(rq
, &flags
);
6313 read_unlock(&tasklist_lock
);
6314 jiffies_to_timespec(time_slice
, &t
);
6315 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
6319 read_unlock(&tasklist_lock
);
6323 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
6325 void sched_show_task(struct task_struct
*p
)
6327 unsigned long free
= 0;
6330 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
6331 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
6332 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
6333 #if BITS_PER_LONG == 32
6334 if (state
== TASK_RUNNING
)
6335 printk(KERN_CONT
" running ");
6337 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
6339 if (state
== TASK_RUNNING
)
6340 printk(KERN_CONT
" running task ");
6342 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
6344 #ifdef CONFIG_DEBUG_STACK_USAGE
6346 unsigned long *n
= end_of_stack(p
);
6349 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
6352 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
6353 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
6355 show_stack(p
, NULL
);
6358 void show_state_filter(unsigned long state_filter
)
6360 struct task_struct
*g
, *p
;
6362 #if BITS_PER_LONG == 32
6364 " task PC stack pid father\n");
6367 " task PC stack pid father\n");
6369 read_lock(&tasklist_lock
);
6370 do_each_thread(g
, p
) {
6372 * reset the NMI-timeout, listing all files on a slow
6373 * console might take alot of time:
6375 touch_nmi_watchdog();
6376 if (!state_filter
|| (p
->state
& state_filter
))
6378 } while_each_thread(g
, p
);
6380 touch_all_softlockup_watchdogs();
6382 #ifdef CONFIG_SCHED_DEBUG
6383 sysrq_sched_debug_show();
6385 read_unlock(&tasklist_lock
);
6387 * Only show locks if all tasks are dumped:
6389 if (state_filter
== -1)
6390 debug_show_all_locks();
6393 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
6395 idle
->sched_class
= &idle_sched_class
;
6399 * init_idle - set up an idle thread for a given CPU
6400 * @idle: task in question
6401 * @cpu: cpu the idle task belongs to
6403 * NOTE: this function does not set the idle thread's NEED_RESCHED
6404 * flag, to make booting more robust.
6406 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
6408 struct rq
*rq
= cpu_rq(cpu
);
6409 unsigned long flags
;
6411 spin_lock_irqsave(&rq
->lock
, flags
);
6414 idle
->se
.exec_start
= sched_clock();
6416 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
6417 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
6418 __set_task_cpu(idle
, cpu
);
6420 rq
->curr
= rq
->idle
= idle
;
6421 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6424 spin_unlock_irqrestore(&rq
->lock
, flags
);
6426 /* Set the preempt count _outside_ the spinlocks! */
6427 #if defined(CONFIG_PREEMPT)
6428 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
6430 task_thread_info(idle
)->preempt_count
= 0;
6433 * The idle tasks have their own, simple scheduling class:
6435 idle
->sched_class
= &idle_sched_class
;
6436 ftrace_graph_init_task(idle
);
6440 * In a system that switches off the HZ timer nohz_cpu_mask
6441 * indicates which cpus entered this state. This is used
6442 * in the rcu update to wait only for active cpus. For system
6443 * which do not switch off the HZ timer nohz_cpu_mask should
6444 * always be CPU_BITS_NONE.
6446 cpumask_var_t nohz_cpu_mask
;
6449 * Increase the granularity value when there are more CPUs,
6450 * because with more CPUs the 'effective latency' as visible
6451 * to users decreases. But the relationship is not linear,
6452 * so pick a second-best guess by going with the log2 of the
6455 * This idea comes from the SD scheduler of Con Kolivas:
6457 static inline void sched_init_granularity(void)
6459 unsigned int factor
= 1 + ilog2(num_online_cpus());
6460 const unsigned long limit
= 200000000;
6462 sysctl_sched_min_granularity
*= factor
;
6463 if (sysctl_sched_min_granularity
> limit
)
6464 sysctl_sched_min_granularity
= limit
;
6466 sysctl_sched_latency
*= factor
;
6467 if (sysctl_sched_latency
> limit
)
6468 sysctl_sched_latency
= limit
;
6470 sysctl_sched_wakeup_granularity
*= factor
;
6472 sysctl_sched_shares_ratelimit
*= factor
;
6477 * This is how migration works:
6479 * 1) we queue a struct migration_req structure in the source CPU's
6480 * runqueue and wake up that CPU's migration thread.
6481 * 2) we down() the locked semaphore => thread blocks.
6482 * 3) migration thread wakes up (implicitly it forces the migrated
6483 * thread off the CPU)
6484 * 4) it gets the migration request and checks whether the migrated
6485 * task is still in the wrong runqueue.
6486 * 5) if it's in the wrong runqueue then the migration thread removes
6487 * it and puts it into the right queue.
6488 * 6) migration thread up()s the semaphore.
6489 * 7) we wake up and the migration is done.
6493 * Change a given task's CPU affinity. Migrate the thread to a
6494 * proper CPU and schedule it away if the CPU it's executing on
6495 * is removed from the allowed bitmask.
6497 * NOTE: the caller must have a valid reference to the task, the
6498 * task must not exit() & deallocate itself prematurely. The
6499 * call is not atomic; no spinlocks may be held.
6501 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
6503 struct migration_req req
;
6504 unsigned long flags
;
6508 rq
= task_rq_lock(p
, &flags
);
6509 if (!cpumask_intersects(new_mask
, cpu_online_mask
)) {
6514 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
6515 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
6520 if (p
->sched_class
->set_cpus_allowed
)
6521 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
6523 cpumask_copy(&p
->cpus_allowed
, new_mask
);
6524 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
6527 /* Can the task run on the task's current CPU? If so, we're done */
6528 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
6531 if (migrate_task(p
, cpumask_any_and(cpu_online_mask
, new_mask
), &req
)) {
6532 /* Need help from migration thread: drop lock and wait. */
6533 task_rq_unlock(rq
, &flags
);
6534 wake_up_process(rq
->migration_thread
);
6535 wait_for_completion(&req
.done
);
6536 tlb_migrate_finish(p
->mm
);
6540 task_rq_unlock(rq
, &flags
);
6544 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
6547 * Move (not current) task off this cpu, onto dest cpu. We're doing
6548 * this because either it can't run here any more (set_cpus_allowed()
6549 * away from this CPU, or CPU going down), or because we're
6550 * attempting to rebalance this task on exec (sched_exec).
6552 * So we race with normal scheduler movements, but that's OK, as long
6553 * as the task is no longer on this CPU.
6555 * Returns non-zero if task was successfully migrated.
6557 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6559 struct rq
*rq_dest
, *rq_src
;
6562 if (unlikely(!cpu_active(dest_cpu
)))
6565 rq_src
= cpu_rq(src_cpu
);
6566 rq_dest
= cpu_rq(dest_cpu
);
6568 double_rq_lock(rq_src
, rq_dest
);
6569 /* Already moved. */
6570 if (task_cpu(p
) != src_cpu
)
6572 /* Affinity changed (again). */
6573 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
6576 on_rq
= p
->se
.on_rq
;
6578 deactivate_task(rq_src
, p
, 0);
6580 set_task_cpu(p
, dest_cpu
);
6582 activate_task(rq_dest
, p
, 0);
6583 check_preempt_curr(rq_dest
, p
, 0);
6588 double_rq_unlock(rq_src
, rq_dest
);
6593 * migration_thread - this is a highprio system thread that performs
6594 * thread migration by bumping thread off CPU then 'pushing' onto
6597 static int migration_thread(void *data
)
6599 int cpu
= (long)data
;
6603 BUG_ON(rq
->migration_thread
!= current
);
6605 set_current_state(TASK_INTERRUPTIBLE
);
6606 while (!kthread_should_stop()) {
6607 struct migration_req
*req
;
6608 struct list_head
*head
;
6610 spin_lock_irq(&rq
->lock
);
6612 if (cpu_is_offline(cpu
)) {
6613 spin_unlock_irq(&rq
->lock
);
6617 if (rq
->active_balance
) {
6618 active_load_balance(rq
, cpu
);
6619 rq
->active_balance
= 0;
6622 head
= &rq
->migration_queue
;
6624 if (list_empty(head
)) {
6625 spin_unlock_irq(&rq
->lock
);
6627 set_current_state(TASK_INTERRUPTIBLE
);
6630 req
= list_entry(head
->next
, struct migration_req
, list
);
6631 list_del_init(head
->next
);
6633 spin_unlock(&rq
->lock
);
6634 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
6637 complete(&req
->done
);
6639 __set_current_state(TASK_RUNNING
);
6643 /* Wait for kthread_stop */
6644 set_current_state(TASK_INTERRUPTIBLE
);
6645 while (!kthread_should_stop()) {
6647 set_current_state(TASK_INTERRUPTIBLE
);
6649 __set_current_state(TASK_RUNNING
);
6653 #ifdef CONFIG_HOTPLUG_CPU
6655 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6659 local_irq_disable();
6660 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
6666 * Figure out where task on dead CPU should go, use force if necessary.
6668 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
6671 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(dead_cpu
));
6674 /* Look for allowed, online CPU in same node. */
6675 for_each_cpu_and(dest_cpu
, nodemask
, cpu_online_mask
)
6676 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
6679 /* Any allowed, online CPU? */
6680 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_online_mask
);
6681 if (dest_cpu
< nr_cpu_ids
)
6684 /* No more Mr. Nice Guy. */
6685 if (dest_cpu
>= nr_cpu_ids
) {
6686 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
6687 dest_cpu
= cpumask_any_and(cpu_online_mask
, &p
->cpus_allowed
);
6690 * Don't tell them about moving exiting tasks or
6691 * kernel threads (both mm NULL), since they never
6694 if (p
->mm
&& printk_ratelimit()) {
6695 printk(KERN_INFO
"process %d (%s) no "
6696 "longer affine to cpu%d\n",
6697 task_pid_nr(p
), p
->comm
, dead_cpu
);
6702 /* It can have affinity changed while we were choosing. */
6703 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
6708 * While a dead CPU has no uninterruptible tasks queued at this point,
6709 * it might still have a nonzero ->nr_uninterruptible counter, because
6710 * for performance reasons the counter is not stricly tracking tasks to
6711 * their home CPUs. So we just add the counter to another CPU's counter,
6712 * to keep the global sum constant after CPU-down:
6714 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
6716 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_online_mask
));
6717 unsigned long flags
;
6719 local_irq_save(flags
);
6720 double_rq_lock(rq_src
, rq_dest
);
6721 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
6722 rq_src
->nr_uninterruptible
= 0;
6723 double_rq_unlock(rq_src
, rq_dest
);
6724 local_irq_restore(flags
);
6727 /* Run through task list and migrate tasks from the dead cpu. */
6728 static void migrate_live_tasks(int src_cpu
)
6730 struct task_struct
*p
, *t
;
6732 read_lock(&tasklist_lock
);
6734 do_each_thread(t
, p
) {
6738 if (task_cpu(p
) == src_cpu
)
6739 move_task_off_dead_cpu(src_cpu
, p
);
6740 } while_each_thread(t
, p
);
6742 read_unlock(&tasklist_lock
);
6746 * Schedules idle task to be the next runnable task on current CPU.
6747 * It does so by boosting its priority to highest possible.
6748 * Used by CPU offline code.
6750 void sched_idle_next(void)
6752 int this_cpu
= smp_processor_id();
6753 struct rq
*rq
= cpu_rq(this_cpu
);
6754 struct task_struct
*p
= rq
->idle
;
6755 unsigned long flags
;
6757 /* cpu has to be offline */
6758 BUG_ON(cpu_online(this_cpu
));
6761 * Strictly not necessary since rest of the CPUs are stopped by now
6762 * and interrupts disabled on the current cpu.
6764 spin_lock_irqsave(&rq
->lock
, flags
);
6766 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6768 update_rq_clock(rq
);
6769 activate_task(rq
, p
, 0);
6771 spin_unlock_irqrestore(&rq
->lock
, flags
);
6775 * Ensures that the idle task is using init_mm right before its cpu goes
6778 void idle_task_exit(void)
6780 struct mm_struct
*mm
= current
->active_mm
;
6782 BUG_ON(cpu_online(smp_processor_id()));
6785 switch_mm(mm
, &init_mm
, current
);
6789 /* called under rq->lock with disabled interrupts */
6790 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
6792 struct rq
*rq
= cpu_rq(dead_cpu
);
6794 /* Must be exiting, otherwise would be on tasklist. */
6795 BUG_ON(!p
->exit_state
);
6797 /* Cannot have done final schedule yet: would have vanished. */
6798 BUG_ON(p
->state
== TASK_DEAD
);
6803 * Drop lock around migration; if someone else moves it,
6804 * that's OK. No task can be added to this CPU, so iteration is
6807 spin_unlock_irq(&rq
->lock
);
6808 move_task_off_dead_cpu(dead_cpu
, p
);
6809 spin_lock_irq(&rq
->lock
);
6814 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6815 static void migrate_dead_tasks(unsigned int dead_cpu
)
6817 struct rq
*rq
= cpu_rq(dead_cpu
);
6818 struct task_struct
*next
;
6821 if (!rq
->nr_running
)
6823 update_rq_clock(rq
);
6824 next
= pick_next_task(rq
);
6827 next
->sched_class
->put_prev_task(rq
, next
);
6828 migrate_dead(dead_cpu
, next
);
6832 #endif /* CONFIG_HOTPLUG_CPU */
6834 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6836 static struct ctl_table sd_ctl_dir
[] = {
6838 .procname
= "sched_domain",
6844 static struct ctl_table sd_ctl_root
[] = {
6846 .ctl_name
= CTL_KERN
,
6847 .procname
= "kernel",
6849 .child
= sd_ctl_dir
,
6854 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6856 struct ctl_table
*entry
=
6857 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6862 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6864 struct ctl_table
*entry
;
6867 * In the intermediate directories, both the child directory and
6868 * procname are dynamically allocated and could fail but the mode
6869 * will always be set. In the lowest directory the names are
6870 * static strings and all have proc handlers.
6872 for (entry
= *tablep
; entry
->mode
; entry
++) {
6874 sd_free_ctl_entry(&entry
->child
);
6875 if (entry
->proc_handler
== NULL
)
6876 kfree(entry
->procname
);
6884 set_table_entry(struct ctl_table
*entry
,
6885 const char *procname
, void *data
, int maxlen
,
6886 mode_t mode
, proc_handler
*proc_handler
)
6888 entry
->procname
= procname
;
6890 entry
->maxlen
= maxlen
;
6892 entry
->proc_handler
= proc_handler
;
6895 static struct ctl_table
*
6896 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6898 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
6903 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6904 sizeof(long), 0644, proc_doulongvec_minmax
);
6905 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6906 sizeof(long), 0644, proc_doulongvec_minmax
);
6907 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6908 sizeof(int), 0644, proc_dointvec_minmax
);
6909 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6910 sizeof(int), 0644, proc_dointvec_minmax
);
6911 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6912 sizeof(int), 0644, proc_dointvec_minmax
);
6913 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6914 sizeof(int), 0644, proc_dointvec_minmax
);
6915 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6916 sizeof(int), 0644, proc_dointvec_minmax
);
6917 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6918 sizeof(int), 0644, proc_dointvec_minmax
);
6919 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6920 sizeof(int), 0644, proc_dointvec_minmax
);
6921 set_table_entry(&table
[9], "cache_nice_tries",
6922 &sd
->cache_nice_tries
,
6923 sizeof(int), 0644, proc_dointvec_minmax
);
6924 set_table_entry(&table
[10], "flags", &sd
->flags
,
6925 sizeof(int), 0644, proc_dointvec_minmax
);
6926 set_table_entry(&table
[11], "name", sd
->name
,
6927 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
6928 /* &table[12] is terminator */
6933 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6935 struct ctl_table
*entry
, *table
;
6936 struct sched_domain
*sd
;
6937 int domain_num
= 0, i
;
6940 for_each_domain(cpu
, sd
)
6942 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6947 for_each_domain(cpu
, sd
) {
6948 snprintf(buf
, 32, "domain%d", i
);
6949 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6951 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6958 static struct ctl_table_header
*sd_sysctl_header
;
6959 static void register_sched_domain_sysctl(void)
6961 int i
, cpu_num
= num_online_cpus();
6962 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6965 WARN_ON(sd_ctl_dir
[0].child
);
6966 sd_ctl_dir
[0].child
= entry
;
6971 for_each_online_cpu(i
) {
6972 snprintf(buf
, 32, "cpu%d", i
);
6973 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6975 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6979 WARN_ON(sd_sysctl_header
);
6980 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6983 /* may be called multiple times per register */
6984 static void unregister_sched_domain_sysctl(void)
6986 if (sd_sysctl_header
)
6987 unregister_sysctl_table(sd_sysctl_header
);
6988 sd_sysctl_header
= NULL
;
6989 if (sd_ctl_dir
[0].child
)
6990 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6993 static void register_sched_domain_sysctl(void)
6996 static void unregister_sched_domain_sysctl(void)
7001 static void set_rq_online(struct rq
*rq
)
7004 const struct sched_class
*class;
7006 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
7009 for_each_class(class) {
7010 if (class->rq_online
)
7011 class->rq_online(rq
);
7016 static void set_rq_offline(struct rq
*rq
)
7019 const struct sched_class
*class;
7021 for_each_class(class) {
7022 if (class->rq_offline
)
7023 class->rq_offline(rq
);
7026 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
7032 * migration_call - callback that gets triggered when a CPU is added.
7033 * Here we can start up the necessary migration thread for the new CPU.
7035 static int __cpuinit
7036 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
7038 struct task_struct
*p
;
7039 int cpu
= (long)hcpu
;
7040 unsigned long flags
;
7045 case CPU_UP_PREPARE
:
7046 case CPU_UP_PREPARE_FROZEN
:
7047 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
7050 kthread_bind(p
, cpu
);
7051 /* Must be high prio: stop_machine expects to yield to it. */
7052 rq
= task_rq_lock(p
, &flags
);
7053 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7054 task_rq_unlock(rq
, &flags
);
7055 cpu_rq(cpu
)->migration_thread
= p
;
7059 case CPU_ONLINE_FROZEN
:
7060 /* Strictly unnecessary, as first user will wake it. */
7061 wake_up_process(cpu_rq(cpu
)->migration_thread
);
7063 /* Update our root-domain */
7065 spin_lock_irqsave(&rq
->lock
, flags
);
7067 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7071 spin_unlock_irqrestore(&rq
->lock
, flags
);
7074 #ifdef CONFIG_HOTPLUG_CPU
7075 case CPU_UP_CANCELED
:
7076 case CPU_UP_CANCELED_FROZEN
:
7077 if (!cpu_rq(cpu
)->migration_thread
)
7079 /* Unbind it from offline cpu so it can run. Fall thru. */
7080 kthread_bind(cpu_rq(cpu
)->migration_thread
,
7081 cpumask_any(cpu_online_mask
));
7082 kthread_stop(cpu_rq(cpu
)->migration_thread
);
7083 cpu_rq(cpu
)->migration_thread
= NULL
;
7087 case CPU_DEAD_FROZEN
:
7088 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7089 migrate_live_tasks(cpu
);
7091 kthread_stop(rq
->migration_thread
);
7092 rq
->migration_thread
= NULL
;
7093 /* Idle task back to normal (off runqueue, low prio) */
7094 spin_lock_irq(&rq
->lock
);
7095 update_rq_clock(rq
);
7096 deactivate_task(rq
, rq
->idle
, 0);
7097 rq
->idle
->static_prio
= MAX_PRIO
;
7098 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
7099 rq
->idle
->sched_class
= &idle_sched_class
;
7100 migrate_dead_tasks(cpu
);
7101 spin_unlock_irq(&rq
->lock
);
7103 migrate_nr_uninterruptible(rq
);
7104 BUG_ON(rq
->nr_running
!= 0);
7107 * No need to migrate the tasks: it was best-effort if
7108 * they didn't take sched_hotcpu_mutex. Just wake up
7111 spin_lock_irq(&rq
->lock
);
7112 while (!list_empty(&rq
->migration_queue
)) {
7113 struct migration_req
*req
;
7115 req
= list_entry(rq
->migration_queue
.next
,
7116 struct migration_req
, list
);
7117 list_del_init(&req
->list
);
7118 spin_unlock_irq(&rq
->lock
);
7119 complete(&req
->done
);
7120 spin_lock_irq(&rq
->lock
);
7122 spin_unlock_irq(&rq
->lock
);
7126 case CPU_DYING_FROZEN
:
7127 /* Update our root-domain */
7129 spin_lock_irqsave(&rq
->lock
, flags
);
7131 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7134 spin_unlock_irqrestore(&rq
->lock
, flags
);
7141 /* Register at highest priority so that task migration (migrate_all_tasks)
7142 * happens before everything else.
7144 static struct notifier_block __cpuinitdata migration_notifier
= {
7145 .notifier_call
= migration_call
,
7149 static int __init
migration_init(void)
7151 void *cpu
= (void *)(long)smp_processor_id();
7154 /* Start one for the boot CPU: */
7155 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
7156 BUG_ON(err
== NOTIFY_BAD
);
7157 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
7158 register_cpu_notifier(&migration_notifier
);
7162 early_initcall(migration_init
);
7167 #ifdef CONFIG_SCHED_DEBUG
7169 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
7170 struct cpumask
*groupmask
)
7172 struct sched_group
*group
= sd
->groups
;
7175 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
7176 cpumask_clear(groupmask
);
7178 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
7180 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
7181 printk("does not load-balance\n");
7183 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
7188 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
7190 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
7191 printk(KERN_ERR
"ERROR: domain->span does not contain "
7194 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
7195 printk(KERN_ERR
"ERROR: domain->groups does not contain"
7199 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
7203 printk(KERN_ERR
"ERROR: group is NULL\n");
7207 if (!group
->__cpu_power
) {
7208 printk(KERN_CONT
"\n");
7209 printk(KERN_ERR
"ERROR: domain->cpu_power not "
7214 if (!cpumask_weight(sched_group_cpus(group
))) {
7215 printk(KERN_CONT
"\n");
7216 printk(KERN_ERR
"ERROR: empty group\n");
7220 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
7221 printk(KERN_CONT
"\n");
7222 printk(KERN_ERR
"ERROR: repeated CPUs\n");
7226 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
7228 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
7229 printk(KERN_CONT
" %s", str
);
7231 group
= group
->next
;
7232 } while (group
!= sd
->groups
);
7233 printk(KERN_CONT
"\n");
7235 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
7236 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
7239 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
7240 printk(KERN_ERR
"ERROR: parent span is not a superset "
7241 "of domain->span\n");
7245 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
7247 cpumask_var_t groupmask
;
7251 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
7255 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
7257 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
7258 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
7263 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
7270 free_cpumask_var(groupmask
);
7272 #else /* !CONFIG_SCHED_DEBUG */
7273 # define sched_domain_debug(sd, cpu) do { } while (0)
7274 #endif /* CONFIG_SCHED_DEBUG */
7276 static int sd_degenerate(struct sched_domain
*sd
)
7278 if (cpumask_weight(sched_domain_span(sd
)) == 1)
7281 /* Following flags need at least 2 groups */
7282 if (sd
->flags
& (SD_LOAD_BALANCE
|
7283 SD_BALANCE_NEWIDLE
|
7287 SD_SHARE_PKG_RESOURCES
)) {
7288 if (sd
->groups
!= sd
->groups
->next
)
7292 /* Following flags don't use groups */
7293 if (sd
->flags
& (SD_WAKE_IDLE
|
7302 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
7304 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
7306 if (sd_degenerate(parent
))
7309 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
7312 /* Does parent contain flags not in child? */
7313 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7314 if (cflags
& SD_WAKE_AFFINE
)
7315 pflags
&= ~SD_WAKE_BALANCE
;
7316 /* Flags needing groups don't count if only 1 group in parent */
7317 if (parent
->groups
== parent
->groups
->next
) {
7318 pflags
&= ~(SD_LOAD_BALANCE
|
7319 SD_BALANCE_NEWIDLE
|
7323 SD_SHARE_PKG_RESOURCES
);
7324 if (nr_node_ids
== 1)
7325 pflags
&= ~SD_SERIALIZE
;
7327 if (~cflags
& pflags
)
7333 static void free_rootdomain(struct root_domain
*rd
)
7335 cpupri_cleanup(&rd
->cpupri
);
7337 free_cpumask_var(rd
->rto_mask
);
7338 free_cpumask_var(rd
->online
);
7339 free_cpumask_var(rd
->span
);
7343 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
7345 struct root_domain
*old_rd
= NULL
;
7346 unsigned long flags
;
7348 spin_lock_irqsave(&rq
->lock
, flags
);
7353 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
7356 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
7359 * If we dont want to free the old_rt yet then
7360 * set old_rd to NULL to skip the freeing later
7363 if (!atomic_dec_and_test(&old_rd
->refcount
))
7367 atomic_inc(&rd
->refcount
);
7370 cpumask_set_cpu(rq
->cpu
, rd
->span
);
7371 if (cpumask_test_cpu(rq
->cpu
, cpu_online_mask
))
7374 spin_unlock_irqrestore(&rq
->lock
, flags
);
7377 free_rootdomain(old_rd
);
7380 static int __init_refok
init_rootdomain(struct root_domain
*rd
, bool bootmem
)
7382 memset(rd
, 0, sizeof(*rd
));
7385 alloc_bootmem_cpumask_var(&def_root_domain
.span
);
7386 alloc_bootmem_cpumask_var(&def_root_domain
.online
);
7387 alloc_bootmem_cpumask_var(&def_root_domain
.rto_mask
);
7388 cpupri_init(&rd
->cpupri
, true);
7392 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
7394 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
7396 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
7399 if (cpupri_init(&rd
->cpupri
, false) != 0)
7404 free_cpumask_var(rd
->rto_mask
);
7406 free_cpumask_var(rd
->online
);
7408 free_cpumask_var(rd
->span
);
7413 static void init_defrootdomain(void)
7415 init_rootdomain(&def_root_domain
, true);
7417 atomic_set(&def_root_domain
.refcount
, 1);
7420 static struct root_domain
*alloc_rootdomain(void)
7422 struct root_domain
*rd
;
7424 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
7428 if (init_rootdomain(rd
, false) != 0) {
7437 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7438 * hold the hotplug lock.
7441 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
7443 struct rq
*rq
= cpu_rq(cpu
);
7444 struct sched_domain
*tmp
;
7446 /* Remove the sched domains which do not contribute to scheduling. */
7447 for (tmp
= sd
; tmp
; ) {
7448 struct sched_domain
*parent
= tmp
->parent
;
7452 if (sd_parent_degenerate(tmp
, parent
)) {
7453 tmp
->parent
= parent
->parent
;
7455 parent
->parent
->child
= tmp
;
7460 if (sd
&& sd_degenerate(sd
)) {
7466 sched_domain_debug(sd
, cpu
);
7468 rq_attach_root(rq
, rd
);
7469 rcu_assign_pointer(rq
->sd
, sd
);
7472 /* cpus with isolated domains */
7473 static cpumask_var_t cpu_isolated_map
;
7475 /* Setup the mask of cpus configured for isolated domains */
7476 static int __init
isolated_cpu_setup(char *str
)
7478 cpulist_parse(str
, cpu_isolated_map
);
7482 __setup("isolcpus=", isolated_cpu_setup
);
7485 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7486 * to a function which identifies what group(along with sched group) a CPU
7487 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7488 * (due to the fact that we keep track of groups covered with a struct cpumask).
7490 * init_sched_build_groups will build a circular linked list of the groups
7491 * covered by the given span, and will set each group's ->cpumask correctly,
7492 * and ->cpu_power to 0.
7495 init_sched_build_groups(const struct cpumask
*span
,
7496 const struct cpumask
*cpu_map
,
7497 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
7498 struct sched_group
**sg
,
7499 struct cpumask
*tmpmask
),
7500 struct cpumask
*covered
, struct cpumask
*tmpmask
)
7502 struct sched_group
*first
= NULL
, *last
= NULL
;
7505 cpumask_clear(covered
);
7507 for_each_cpu(i
, span
) {
7508 struct sched_group
*sg
;
7509 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
7512 if (cpumask_test_cpu(i
, covered
))
7515 cpumask_clear(sched_group_cpus(sg
));
7516 sg
->__cpu_power
= 0;
7518 for_each_cpu(j
, span
) {
7519 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
7522 cpumask_set_cpu(j
, covered
);
7523 cpumask_set_cpu(j
, sched_group_cpus(sg
));
7534 #define SD_NODES_PER_DOMAIN 16
7539 * find_next_best_node - find the next node to include in a sched_domain
7540 * @node: node whose sched_domain we're building
7541 * @used_nodes: nodes already in the sched_domain
7543 * Find the next node to include in a given scheduling domain. Simply
7544 * finds the closest node not already in the @used_nodes map.
7546 * Should use nodemask_t.
7548 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
7550 int i
, n
, val
, min_val
, best_node
= 0;
7554 for (i
= 0; i
< nr_node_ids
; i
++) {
7555 /* Start at @node */
7556 n
= (node
+ i
) % nr_node_ids
;
7558 if (!nr_cpus_node(n
))
7561 /* Skip already used nodes */
7562 if (node_isset(n
, *used_nodes
))
7565 /* Simple min distance search */
7566 val
= node_distance(node
, n
);
7568 if (val
< min_val
) {
7574 node_set(best_node
, *used_nodes
);
7579 * sched_domain_node_span - get a cpumask for a node's sched_domain
7580 * @node: node whose cpumask we're constructing
7581 * @span: resulting cpumask
7583 * Given a node, construct a good cpumask for its sched_domain to span. It
7584 * should be one that prevents unnecessary balancing, but also spreads tasks
7587 static void sched_domain_node_span(int node
, struct cpumask
*span
)
7589 nodemask_t used_nodes
;
7592 cpumask_clear(span
);
7593 nodes_clear(used_nodes
);
7595 cpumask_or(span
, span
, cpumask_of_node(node
));
7596 node_set(node
, used_nodes
);
7598 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
7599 int next_node
= find_next_best_node(node
, &used_nodes
);
7601 cpumask_or(span
, span
, cpumask_of_node(next_node
));
7604 #endif /* CONFIG_NUMA */
7606 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
7609 * The cpus mask in sched_group and sched_domain hangs off the end.
7610 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7611 * for nr_cpu_ids < CONFIG_NR_CPUS.
7613 struct static_sched_group
{
7614 struct sched_group sg
;
7615 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
7618 struct static_sched_domain
{
7619 struct sched_domain sd
;
7620 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
7624 * SMT sched-domains:
7626 #ifdef CONFIG_SCHED_SMT
7627 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
7628 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_cpus
);
7631 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
7632 struct sched_group
**sg
, struct cpumask
*unused
)
7635 *sg
= &per_cpu(sched_group_cpus
, cpu
).sg
;
7638 #endif /* CONFIG_SCHED_SMT */
7641 * multi-core sched-domains:
7643 #ifdef CONFIG_SCHED_MC
7644 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
7645 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
7646 #endif /* CONFIG_SCHED_MC */
7648 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7650 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
7651 struct sched_group
**sg
, struct cpumask
*mask
)
7655 cpumask_and(mask
, &per_cpu(cpu_sibling_map
, cpu
), cpu_map
);
7656 group
= cpumask_first(mask
);
7658 *sg
= &per_cpu(sched_group_core
, group
).sg
;
7661 #elif defined(CONFIG_SCHED_MC)
7663 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
7664 struct sched_group
**sg
, struct cpumask
*unused
)
7667 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
7672 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
7673 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
7676 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
7677 struct sched_group
**sg
, struct cpumask
*mask
)
7680 #ifdef CONFIG_SCHED_MC
7681 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
7682 group
= cpumask_first(mask
);
7683 #elif defined(CONFIG_SCHED_SMT)
7684 cpumask_and(mask
, &per_cpu(cpu_sibling_map
, cpu
), cpu_map
);
7685 group
= cpumask_first(mask
);
7690 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
7696 * The init_sched_build_groups can't handle what we want to do with node
7697 * groups, so roll our own. Now each node has its own list of groups which
7698 * gets dynamically allocated.
7700 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
7701 static struct sched_group
***sched_group_nodes_bycpu
;
7703 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
7704 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
7706 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
7707 struct sched_group
**sg
,
7708 struct cpumask
*nodemask
)
7712 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
7713 group
= cpumask_first(nodemask
);
7716 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
7720 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
7722 struct sched_group
*sg
= group_head
;
7728 for_each_cpu(j
, sched_group_cpus(sg
)) {
7729 struct sched_domain
*sd
;
7731 sd
= &per_cpu(phys_domains
, j
).sd
;
7732 if (j
!= cpumask_first(sched_group_cpus(sd
->groups
))) {
7734 * Only add "power" once for each
7740 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
7743 } while (sg
!= group_head
);
7745 #endif /* CONFIG_NUMA */
7748 /* Free memory allocated for various sched_group structures */
7749 static void free_sched_groups(const struct cpumask
*cpu_map
,
7750 struct cpumask
*nodemask
)
7754 for_each_cpu(cpu
, cpu_map
) {
7755 struct sched_group
**sched_group_nodes
7756 = sched_group_nodes_bycpu
[cpu
];
7758 if (!sched_group_nodes
)
7761 for (i
= 0; i
< nr_node_ids
; i
++) {
7762 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
7764 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
7765 if (cpumask_empty(nodemask
))
7775 if (oldsg
!= sched_group_nodes
[i
])
7778 kfree(sched_group_nodes
);
7779 sched_group_nodes_bycpu
[cpu
] = NULL
;
7782 #else /* !CONFIG_NUMA */
7783 static void free_sched_groups(const struct cpumask
*cpu_map
,
7784 struct cpumask
*nodemask
)
7787 #endif /* CONFIG_NUMA */
7790 * Initialize sched groups cpu_power.
7792 * cpu_power indicates the capacity of sched group, which is used while
7793 * distributing the load between different sched groups in a sched domain.
7794 * Typically cpu_power for all the groups in a sched domain will be same unless
7795 * there are asymmetries in the topology. If there are asymmetries, group
7796 * having more cpu_power will pickup more load compared to the group having
7799 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7800 * the maximum number of tasks a group can handle in the presence of other idle
7801 * or lightly loaded groups in the same sched domain.
7803 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7805 struct sched_domain
*child
;
7806 struct sched_group
*group
;
7808 WARN_ON(!sd
|| !sd
->groups
);
7810 if (cpu
!= cpumask_first(sched_group_cpus(sd
->groups
)))
7815 sd
->groups
->__cpu_power
= 0;
7818 * For perf policy, if the groups in child domain share resources
7819 * (for example cores sharing some portions of the cache hierarchy
7820 * or SMT), then set this domain groups cpu_power such that each group
7821 * can handle only one task, when there are other idle groups in the
7822 * same sched domain.
7824 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
7826 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
7827 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
7832 * add cpu_power of each child group to this groups cpu_power
7834 group
= child
->groups
;
7836 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
7837 group
= group
->next
;
7838 } while (group
!= child
->groups
);
7842 * Initializers for schedule domains
7843 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7846 #ifdef CONFIG_SCHED_DEBUG
7847 # define SD_INIT_NAME(sd, type) sd->name = #type
7849 # define SD_INIT_NAME(sd, type) do { } while (0)
7852 #define SD_INIT(sd, type) sd_init_##type(sd)
7854 #define SD_INIT_FUNC(type) \
7855 static noinline void sd_init_##type(struct sched_domain *sd) \
7857 memset(sd, 0, sizeof(*sd)); \
7858 *sd = SD_##type##_INIT; \
7859 sd->level = SD_LV_##type; \
7860 SD_INIT_NAME(sd, type); \
7865 SD_INIT_FUNC(ALLNODES
)
7868 #ifdef CONFIG_SCHED_SMT
7869 SD_INIT_FUNC(SIBLING
)
7871 #ifdef CONFIG_SCHED_MC
7875 static int default_relax_domain_level
= -1;
7877 static int __init
setup_relax_domain_level(char *str
)
7881 val
= simple_strtoul(str
, NULL
, 0);
7882 if (val
< SD_LV_MAX
)
7883 default_relax_domain_level
= val
;
7887 __setup("relax_domain_level=", setup_relax_domain_level
);
7889 static void set_domain_attribute(struct sched_domain
*sd
,
7890 struct sched_domain_attr
*attr
)
7894 if (!attr
|| attr
->relax_domain_level
< 0) {
7895 if (default_relax_domain_level
< 0)
7898 request
= default_relax_domain_level
;
7900 request
= attr
->relax_domain_level
;
7901 if (request
< sd
->level
) {
7902 /* turn off idle balance on this domain */
7903 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
7905 /* turn on idle balance on this domain */
7906 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
7911 * Build sched domains for a given set of cpus and attach the sched domains
7912 * to the individual cpus
7914 static int __build_sched_domains(const struct cpumask
*cpu_map
,
7915 struct sched_domain_attr
*attr
)
7917 int i
, err
= -ENOMEM
;
7918 struct root_domain
*rd
;
7919 cpumask_var_t nodemask
, this_sibling_map
, this_core_map
, send_covered
,
7922 cpumask_var_t domainspan
, covered
, notcovered
;
7923 struct sched_group
**sched_group_nodes
= NULL
;
7924 int sd_allnodes
= 0;
7926 if (!alloc_cpumask_var(&domainspan
, GFP_KERNEL
))
7928 if (!alloc_cpumask_var(&covered
, GFP_KERNEL
))
7929 goto free_domainspan
;
7930 if (!alloc_cpumask_var(¬covered
, GFP_KERNEL
))
7934 if (!alloc_cpumask_var(&nodemask
, GFP_KERNEL
))
7935 goto free_notcovered
;
7936 if (!alloc_cpumask_var(&this_sibling_map
, GFP_KERNEL
))
7938 if (!alloc_cpumask_var(&this_core_map
, GFP_KERNEL
))
7939 goto free_this_sibling_map
;
7940 if (!alloc_cpumask_var(&send_covered
, GFP_KERNEL
))
7941 goto free_this_core_map
;
7942 if (!alloc_cpumask_var(&tmpmask
, GFP_KERNEL
))
7943 goto free_send_covered
;
7947 * Allocate the per-node list of sched groups
7949 sched_group_nodes
= kcalloc(nr_node_ids
, sizeof(struct sched_group
*),
7951 if (!sched_group_nodes
) {
7952 printk(KERN_WARNING
"Can not alloc sched group node list\n");
7957 rd
= alloc_rootdomain();
7959 printk(KERN_WARNING
"Cannot alloc root domain\n");
7960 goto free_sched_groups
;
7964 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = sched_group_nodes
;
7968 * Set up domains for cpus specified by the cpu_map.
7970 for_each_cpu(i
, cpu_map
) {
7971 struct sched_domain
*sd
= NULL
, *p
;
7973 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(i
)), cpu_map
);
7976 if (cpumask_weight(cpu_map
) >
7977 SD_NODES_PER_DOMAIN
*cpumask_weight(nodemask
)) {
7978 sd
= &per_cpu(allnodes_domains
, i
).sd
;
7979 SD_INIT(sd
, ALLNODES
);
7980 set_domain_attribute(sd
, attr
);
7981 cpumask_copy(sched_domain_span(sd
), cpu_map
);
7982 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7988 sd
= &per_cpu(node_domains
, i
).sd
;
7990 set_domain_attribute(sd
, attr
);
7991 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
7995 cpumask_and(sched_domain_span(sd
),
7996 sched_domain_span(sd
), cpu_map
);
8000 sd
= &per_cpu(phys_domains
, i
).sd
;
8002 set_domain_attribute(sd
, attr
);
8003 cpumask_copy(sched_domain_span(sd
), nodemask
);
8007 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8009 #ifdef CONFIG_SCHED_MC
8011 sd
= &per_cpu(core_domains
, i
).sd
;
8013 set_domain_attribute(sd
, attr
);
8014 cpumask_and(sched_domain_span(sd
), cpu_map
,
8015 cpu_coregroup_mask(i
));
8018 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8021 #ifdef CONFIG_SCHED_SMT
8023 sd
= &per_cpu(cpu_domains
, i
).sd
;
8024 SD_INIT(sd
, SIBLING
);
8025 set_domain_attribute(sd
, attr
);
8026 cpumask_and(sched_domain_span(sd
),
8027 &per_cpu(cpu_sibling_map
, i
), cpu_map
);
8030 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8034 #ifdef CONFIG_SCHED_SMT
8035 /* Set up CPU (sibling) groups */
8036 for_each_cpu(i
, cpu_map
) {
8037 cpumask_and(this_sibling_map
,
8038 &per_cpu(cpu_sibling_map
, i
), cpu_map
);
8039 if (i
!= cpumask_first(this_sibling_map
))
8042 init_sched_build_groups(this_sibling_map
, cpu_map
,
8044 send_covered
, tmpmask
);
8048 #ifdef CONFIG_SCHED_MC
8049 /* Set up multi-core groups */
8050 for_each_cpu(i
, cpu_map
) {
8051 cpumask_and(this_core_map
, cpu_coregroup_mask(i
), cpu_map
);
8052 if (i
!= cpumask_first(this_core_map
))
8055 init_sched_build_groups(this_core_map
, cpu_map
,
8057 send_covered
, tmpmask
);
8061 /* Set up physical groups */
8062 for (i
= 0; i
< nr_node_ids
; i
++) {
8063 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8064 if (cpumask_empty(nodemask
))
8067 init_sched_build_groups(nodemask
, cpu_map
,
8069 send_covered
, tmpmask
);
8073 /* Set up node groups */
8075 init_sched_build_groups(cpu_map
, cpu_map
,
8076 &cpu_to_allnodes_group
,
8077 send_covered
, tmpmask
);
8080 for (i
= 0; i
< nr_node_ids
; i
++) {
8081 /* Set up node groups */
8082 struct sched_group
*sg
, *prev
;
8085 cpumask_clear(covered
);
8086 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8087 if (cpumask_empty(nodemask
)) {
8088 sched_group_nodes
[i
] = NULL
;
8092 sched_domain_node_span(i
, domainspan
);
8093 cpumask_and(domainspan
, domainspan
, cpu_map
);
8095 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8098 printk(KERN_WARNING
"Can not alloc domain group for "
8102 sched_group_nodes
[i
] = sg
;
8103 for_each_cpu(j
, nodemask
) {
8104 struct sched_domain
*sd
;
8106 sd
= &per_cpu(node_domains
, j
).sd
;
8109 sg
->__cpu_power
= 0;
8110 cpumask_copy(sched_group_cpus(sg
), nodemask
);
8112 cpumask_or(covered
, covered
, nodemask
);
8115 for (j
= 0; j
< nr_node_ids
; j
++) {
8116 int n
= (i
+ j
) % nr_node_ids
;
8118 cpumask_complement(notcovered
, covered
);
8119 cpumask_and(tmpmask
, notcovered
, cpu_map
);
8120 cpumask_and(tmpmask
, tmpmask
, domainspan
);
8121 if (cpumask_empty(tmpmask
))
8124 cpumask_and(tmpmask
, tmpmask
, cpumask_of_node(n
));
8125 if (cpumask_empty(tmpmask
))
8128 sg
= kmalloc_node(sizeof(struct sched_group
) +
8133 "Can not alloc domain group for node %d\n", j
);
8136 sg
->__cpu_power
= 0;
8137 cpumask_copy(sched_group_cpus(sg
), tmpmask
);
8138 sg
->next
= prev
->next
;
8139 cpumask_or(covered
, covered
, tmpmask
);
8146 /* Calculate CPU power for physical packages and nodes */
8147 #ifdef CONFIG_SCHED_SMT
8148 for_each_cpu(i
, cpu_map
) {
8149 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
).sd
;
8151 init_sched_groups_power(i
, sd
);
8154 #ifdef CONFIG_SCHED_MC
8155 for_each_cpu(i
, cpu_map
) {
8156 struct sched_domain
*sd
= &per_cpu(core_domains
, i
).sd
;
8158 init_sched_groups_power(i
, sd
);
8162 for_each_cpu(i
, cpu_map
) {
8163 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
).sd
;
8165 init_sched_groups_power(i
, sd
);
8169 for (i
= 0; i
< nr_node_ids
; i
++)
8170 init_numa_sched_groups_power(sched_group_nodes
[i
]);
8173 struct sched_group
*sg
;
8175 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
8177 init_numa_sched_groups_power(sg
);
8181 /* Attach the domains */
8182 for_each_cpu(i
, cpu_map
) {
8183 struct sched_domain
*sd
;
8184 #ifdef CONFIG_SCHED_SMT
8185 sd
= &per_cpu(cpu_domains
, i
).sd
;
8186 #elif defined(CONFIG_SCHED_MC)
8187 sd
= &per_cpu(core_domains
, i
).sd
;
8189 sd
= &per_cpu(phys_domains
, i
).sd
;
8191 cpu_attach_domain(sd
, rd
, i
);
8197 free_cpumask_var(tmpmask
);
8199 free_cpumask_var(send_covered
);
8201 free_cpumask_var(this_core_map
);
8202 free_this_sibling_map
:
8203 free_cpumask_var(this_sibling_map
);
8205 free_cpumask_var(nodemask
);
8208 free_cpumask_var(notcovered
);
8210 free_cpumask_var(covered
);
8212 free_cpumask_var(domainspan
);
8219 kfree(sched_group_nodes
);
8225 free_sched_groups(cpu_map
, tmpmask
);
8226 free_rootdomain(rd
);
8231 static int build_sched_domains(const struct cpumask
*cpu_map
)
8233 return __build_sched_domains(cpu_map
, NULL
);
8236 static struct cpumask
*doms_cur
; /* current sched domains */
8237 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
8238 static struct sched_domain_attr
*dattr_cur
;
8239 /* attribues of custom domains in 'doms_cur' */
8242 * Special case: If a kmalloc of a doms_cur partition (array of
8243 * cpumask) fails, then fallback to a single sched domain,
8244 * as determined by the single cpumask fallback_doms.
8246 static cpumask_var_t fallback_doms
;
8249 * arch_update_cpu_topology lets virtualized architectures update the
8250 * cpu core maps. It is supposed to return 1 if the topology changed
8251 * or 0 if it stayed the same.
8253 int __attribute__((weak
)) arch_update_cpu_topology(void)
8259 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8260 * For now this just excludes isolated cpus, but could be used to
8261 * exclude other special cases in the future.
8263 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
8267 arch_update_cpu_topology();
8269 doms_cur
= kmalloc(cpumask_size(), GFP_KERNEL
);
8271 doms_cur
= fallback_doms
;
8272 cpumask_andnot(doms_cur
, cpu_map
, cpu_isolated_map
);
8274 err
= build_sched_domains(doms_cur
);
8275 register_sched_domain_sysctl();
8280 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
8281 struct cpumask
*tmpmask
)
8283 free_sched_groups(cpu_map
, tmpmask
);
8287 * Detach sched domains from a group of cpus specified in cpu_map
8288 * These cpus will now be attached to the NULL domain
8290 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
8292 /* Save because hotplug lock held. */
8293 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
8296 for_each_cpu(i
, cpu_map
)
8297 cpu_attach_domain(NULL
, &def_root_domain
, i
);
8298 synchronize_sched();
8299 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
8302 /* handle null as "default" */
8303 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
8304 struct sched_domain_attr
*new, int idx_new
)
8306 struct sched_domain_attr tmp
;
8313 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
8314 new ? (new + idx_new
) : &tmp
,
8315 sizeof(struct sched_domain_attr
));
8319 * Partition sched domains as specified by the 'ndoms_new'
8320 * cpumasks in the array doms_new[] of cpumasks. This compares
8321 * doms_new[] to the current sched domain partitioning, doms_cur[].
8322 * It destroys each deleted domain and builds each new domain.
8324 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8325 * The masks don't intersect (don't overlap.) We should setup one
8326 * sched domain for each mask. CPUs not in any of the cpumasks will
8327 * not be load balanced. If the same cpumask appears both in the
8328 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8331 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8332 * ownership of it and will kfree it when done with it. If the caller
8333 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8334 * ndoms_new == 1, and partition_sched_domains() will fallback to
8335 * the single partition 'fallback_doms', it also forces the domains
8338 * If doms_new == NULL it will be replaced with cpu_online_mask.
8339 * ndoms_new == 0 is a special case for destroying existing domains,
8340 * and it will not create the default domain.
8342 * Call with hotplug lock held
8344 /* FIXME: Change to struct cpumask *doms_new[] */
8345 void partition_sched_domains(int ndoms_new
, struct cpumask
*doms_new
,
8346 struct sched_domain_attr
*dattr_new
)
8351 mutex_lock(&sched_domains_mutex
);
8353 /* always unregister in case we don't destroy any domains */
8354 unregister_sched_domain_sysctl();
8356 /* Let architecture update cpu core mappings. */
8357 new_topology
= arch_update_cpu_topology();
8359 n
= doms_new
? ndoms_new
: 0;
8361 /* Destroy deleted domains */
8362 for (i
= 0; i
< ndoms_cur
; i
++) {
8363 for (j
= 0; j
< n
&& !new_topology
; j
++) {
8364 if (cpumask_equal(&doms_cur
[i
], &doms_new
[j
])
8365 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
8368 /* no match - a current sched domain not in new doms_new[] */
8369 detach_destroy_domains(doms_cur
+ i
);
8374 if (doms_new
== NULL
) {
8376 doms_new
= fallback_doms
;
8377 cpumask_andnot(&doms_new
[0], cpu_online_mask
, cpu_isolated_map
);
8378 WARN_ON_ONCE(dattr_new
);
8381 /* Build new domains */
8382 for (i
= 0; i
< ndoms_new
; i
++) {
8383 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
8384 if (cpumask_equal(&doms_new
[i
], &doms_cur
[j
])
8385 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
8388 /* no match - add a new doms_new */
8389 __build_sched_domains(doms_new
+ i
,
8390 dattr_new
? dattr_new
+ i
: NULL
);
8395 /* Remember the new sched domains */
8396 if (doms_cur
!= fallback_doms
)
8398 kfree(dattr_cur
); /* kfree(NULL) is safe */
8399 doms_cur
= doms_new
;
8400 dattr_cur
= dattr_new
;
8401 ndoms_cur
= ndoms_new
;
8403 register_sched_domain_sysctl();
8405 mutex_unlock(&sched_domains_mutex
);
8408 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8409 static void arch_reinit_sched_domains(void)
8413 /* Destroy domains first to force the rebuild */
8414 partition_sched_domains(0, NULL
, NULL
);
8416 rebuild_sched_domains();
8420 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
8422 unsigned int level
= 0;
8424 if (sscanf(buf
, "%u", &level
) != 1)
8428 * level is always be positive so don't check for
8429 * level < POWERSAVINGS_BALANCE_NONE which is 0
8430 * What happens on 0 or 1 byte write,
8431 * need to check for count as well?
8434 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
8438 sched_smt_power_savings
= level
;
8440 sched_mc_power_savings
= level
;
8442 arch_reinit_sched_domains();
8447 #ifdef CONFIG_SCHED_MC
8448 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
8451 return sprintf(page
, "%u\n", sched_mc_power_savings
);
8453 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
8454 const char *buf
, size_t count
)
8456 return sched_power_savings_store(buf
, count
, 0);
8458 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
8459 sched_mc_power_savings_show
,
8460 sched_mc_power_savings_store
);
8463 #ifdef CONFIG_SCHED_SMT
8464 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
8467 return sprintf(page
, "%u\n", sched_smt_power_savings
);
8469 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
8470 const char *buf
, size_t count
)
8472 return sched_power_savings_store(buf
, count
, 1);
8474 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
8475 sched_smt_power_savings_show
,
8476 sched_smt_power_savings_store
);
8479 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
8483 #ifdef CONFIG_SCHED_SMT
8485 err
= sysfs_create_file(&cls
->kset
.kobj
,
8486 &attr_sched_smt_power_savings
.attr
);
8488 #ifdef CONFIG_SCHED_MC
8489 if (!err
&& mc_capable())
8490 err
= sysfs_create_file(&cls
->kset
.kobj
,
8491 &attr_sched_mc_power_savings
.attr
);
8495 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8497 #ifndef CONFIG_CPUSETS
8499 * Add online and remove offline CPUs from the scheduler domains.
8500 * When cpusets are enabled they take over this function.
8502 static int update_sched_domains(struct notifier_block
*nfb
,
8503 unsigned long action
, void *hcpu
)
8507 case CPU_ONLINE_FROZEN
:
8509 case CPU_DEAD_FROZEN
:
8510 partition_sched_domains(1, NULL
, NULL
);
8519 static int update_runtime(struct notifier_block
*nfb
,
8520 unsigned long action
, void *hcpu
)
8522 int cpu
= (int)(long)hcpu
;
8525 case CPU_DOWN_PREPARE
:
8526 case CPU_DOWN_PREPARE_FROZEN
:
8527 disable_runtime(cpu_rq(cpu
));
8530 case CPU_DOWN_FAILED
:
8531 case CPU_DOWN_FAILED_FROZEN
:
8533 case CPU_ONLINE_FROZEN
:
8534 enable_runtime(cpu_rq(cpu
));
8542 void __init
sched_init_smp(void)
8544 cpumask_var_t non_isolated_cpus
;
8546 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
8548 #if defined(CONFIG_NUMA)
8549 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
8551 BUG_ON(sched_group_nodes_bycpu
== NULL
);
8554 mutex_lock(&sched_domains_mutex
);
8555 arch_init_sched_domains(cpu_online_mask
);
8556 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
8557 if (cpumask_empty(non_isolated_cpus
))
8558 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
8559 mutex_unlock(&sched_domains_mutex
);
8562 #ifndef CONFIG_CPUSETS
8563 /* XXX: Theoretical race here - CPU may be hotplugged now */
8564 hotcpu_notifier(update_sched_domains
, 0);
8567 /* RT runtime code needs to handle some hotplug events */
8568 hotcpu_notifier(update_runtime
, 0);
8572 /* Move init over to a non-isolated CPU */
8573 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
8575 sched_init_granularity();
8576 free_cpumask_var(non_isolated_cpus
);
8578 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
8579 init_sched_rt_class();
8582 void __init
sched_init_smp(void)
8584 sched_init_granularity();
8586 #endif /* CONFIG_SMP */
8588 int in_sched_functions(unsigned long addr
)
8590 return in_lock_functions(addr
) ||
8591 (addr
>= (unsigned long)__sched_text_start
8592 && addr
< (unsigned long)__sched_text_end
);
8595 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
8597 cfs_rq
->tasks_timeline
= RB_ROOT
;
8598 INIT_LIST_HEAD(&cfs_rq
->tasks
);
8599 #ifdef CONFIG_FAIR_GROUP_SCHED
8602 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
8605 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
8607 struct rt_prio_array
*array
;
8610 array
= &rt_rq
->active
;
8611 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
8612 INIT_LIST_HEAD(array
->queue
+ i
);
8613 __clear_bit(i
, array
->bitmap
);
8615 /* delimiter for bitsearch: */
8616 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
8618 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8619 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
8621 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
8625 rt_rq
->rt_nr_migratory
= 0;
8626 rt_rq
->overloaded
= 0;
8627 plist_head_init(&rq
->rt
.pushable_tasks
, &rq
->lock
);
8631 rt_rq
->rt_throttled
= 0;
8632 rt_rq
->rt_runtime
= 0;
8633 spin_lock_init(&rt_rq
->rt_runtime_lock
);
8635 #ifdef CONFIG_RT_GROUP_SCHED
8636 rt_rq
->rt_nr_boosted
= 0;
8641 #ifdef CONFIG_FAIR_GROUP_SCHED
8642 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8643 struct sched_entity
*se
, int cpu
, int add
,
8644 struct sched_entity
*parent
)
8646 struct rq
*rq
= cpu_rq(cpu
);
8647 tg
->cfs_rq
[cpu
] = cfs_rq
;
8648 init_cfs_rq(cfs_rq
, rq
);
8651 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
8654 /* se could be NULL for init_task_group */
8659 se
->cfs_rq
= &rq
->cfs
;
8661 se
->cfs_rq
= parent
->my_q
;
8664 se
->load
.weight
= tg
->shares
;
8665 se
->load
.inv_weight
= 0;
8666 se
->parent
= parent
;
8670 #ifdef CONFIG_RT_GROUP_SCHED
8671 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
8672 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
8673 struct sched_rt_entity
*parent
)
8675 struct rq
*rq
= cpu_rq(cpu
);
8677 tg
->rt_rq
[cpu
] = rt_rq
;
8678 init_rt_rq(rt_rq
, rq
);
8680 rt_rq
->rt_se
= rt_se
;
8681 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8683 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
8685 tg
->rt_se
[cpu
] = rt_se
;
8690 rt_se
->rt_rq
= &rq
->rt
;
8692 rt_se
->rt_rq
= parent
->my_q
;
8694 rt_se
->my_q
= rt_rq
;
8695 rt_se
->parent
= parent
;
8696 INIT_LIST_HEAD(&rt_se
->run_list
);
8700 void __init
sched_init(void)
8703 unsigned long alloc_size
= 0, ptr
;
8705 #ifdef CONFIG_FAIR_GROUP_SCHED
8706 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8708 #ifdef CONFIG_RT_GROUP_SCHED
8709 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8711 #ifdef CONFIG_USER_SCHED
8715 * As sched_init() is called before page_alloc is setup,
8716 * we use alloc_bootmem().
8719 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
8721 #ifdef CONFIG_FAIR_GROUP_SCHED
8722 init_task_group
.se
= (struct sched_entity
**)ptr
;
8723 ptr
+= nr_cpu_ids
* sizeof(void **);
8725 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8726 ptr
+= nr_cpu_ids
* sizeof(void **);
8728 #ifdef CONFIG_USER_SCHED
8729 root_task_group
.se
= (struct sched_entity
**)ptr
;
8730 ptr
+= nr_cpu_ids
* sizeof(void **);
8732 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8733 ptr
+= nr_cpu_ids
* sizeof(void **);
8734 #endif /* CONFIG_USER_SCHED */
8735 #endif /* CONFIG_FAIR_GROUP_SCHED */
8736 #ifdef CONFIG_RT_GROUP_SCHED
8737 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8738 ptr
+= nr_cpu_ids
* sizeof(void **);
8740 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8741 ptr
+= nr_cpu_ids
* sizeof(void **);
8743 #ifdef CONFIG_USER_SCHED
8744 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8745 ptr
+= nr_cpu_ids
* sizeof(void **);
8747 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8748 ptr
+= nr_cpu_ids
* sizeof(void **);
8749 #endif /* CONFIG_USER_SCHED */
8750 #endif /* CONFIG_RT_GROUP_SCHED */
8754 init_defrootdomain();
8757 init_rt_bandwidth(&def_rt_bandwidth
,
8758 global_rt_period(), global_rt_runtime());
8760 #ifdef CONFIG_RT_GROUP_SCHED
8761 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
8762 global_rt_period(), global_rt_runtime());
8763 #ifdef CONFIG_USER_SCHED
8764 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8765 global_rt_period(), RUNTIME_INF
);
8766 #endif /* CONFIG_USER_SCHED */
8767 #endif /* CONFIG_RT_GROUP_SCHED */
8769 #ifdef CONFIG_GROUP_SCHED
8770 list_add(&init_task_group
.list
, &task_groups
);
8771 INIT_LIST_HEAD(&init_task_group
.children
);
8773 #ifdef CONFIG_USER_SCHED
8774 INIT_LIST_HEAD(&root_task_group
.children
);
8775 init_task_group
.parent
= &root_task_group
;
8776 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
8777 #endif /* CONFIG_USER_SCHED */
8778 #endif /* CONFIG_GROUP_SCHED */
8780 for_each_possible_cpu(i
) {
8784 spin_lock_init(&rq
->lock
);
8786 init_cfs_rq(&rq
->cfs
, rq
);
8787 init_rt_rq(&rq
->rt
, rq
);
8788 #ifdef CONFIG_FAIR_GROUP_SCHED
8789 init_task_group
.shares
= init_task_group_load
;
8790 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8791 #ifdef CONFIG_CGROUP_SCHED
8793 * How much cpu bandwidth does init_task_group get?
8795 * In case of task-groups formed thr' the cgroup filesystem, it
8796 * gets 100% of the cpu resources in the system. This overall
8797 * system cpu resource is divided among the tasks of
8798 * init_task_group and its child task-groups in a fair manner,
8799 * based on each entity's (task or task-group's) weight
8800 * (se->load.weight).
8802 * In other words, if init_task_group has 10 tasks of weight
8803 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8804 * then A0's share of the cpu resource is:
8806 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8808 * We achieve this by letting init_task_group's tasks sit
8809 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8811 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
8812 #elif defined CONFIG_USER_SCHED
8813 root_task_group
.shares
= NICE_0_LOAD
;
8814 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
8816 * In case of task-groups formed thr' the user id of tasks,
8817 * init_task_group represents tasks belonging to root user.
8818 * Hence it forms a sibling of all subsequent groups formed.
8819 * In this case, init_task_group gets only a fraction of overall
8820 * system cpu resource, based on the weight assigned to root
8821 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8822 * by letting tasks of init_task_group sit in a separate cfs_rq
8823 * (init_cfs_rq) and having one entity represent this group of
8824 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8826 init_tg_cfs_entry(&init_task_group
,
8827 &per_cpu(init_cfs_rq
, i
),
8828 &per_cpu(init_sched_entity
, i
), i
, 1,
8829 root_task_group
.se
[i
]);
8832 #endif /* CONFIG_FAIR_GROUP_SCHED */
8834 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8835 #ifdef CONFIG_RT_GROUP_SCHED
8836 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
8837 #ifdef CONFIG_CGROUP_SCHED
8838 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
8839 #elif defined CONFIG_USER_SCHED
8840 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
8841 init_tg_rt_entry(&init_task_group
,
8842 &per_cpu(init_rt_rq
, i
),
8843 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
8844 root_task_group
.rt_se
[i
]);
8848 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
8849 rq
->cpu_load
[j
] = 0;
8853 rq
->active_balance
= 0;
8854 rq
->next_balance
= jiffies
;
8858 rq
->migration_thread
= NULL
;
8859 INIT_LIST_HEAD(&rq
->migration_queue
);
8860 rq_attach_root(rq
, &def_root_domain
);
8863 atomic_set(&rq
->nr_iowait
, 0);
8866 set_load_weight(&init_task
);
8868 #ifdef CONFIG_PREEMPT_NOTIFIERS
8869 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
8873 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8876 #ifdef CONFIG_RT_MUTEXES
8877 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
8881 * The boot idle thread does lazy MMU switching as well:
8883 atomic_inc(&init_mm
.mm_count
);
8884 enter_lazy_tlb(&init_mm
, current
);
8887 * Make us the idle thread. Technically, schedule() should not be
8888 * called from this thread, however somewhere below it might be,
8889 * but because we are the idle thread, we just pick up running again
8890 * when this runqueue becomes "idle".
8892 init_idle(current
, smp_processor_id());
8894 * During early bootup we pretend to be a normal task:
8896 current
->sched_class
= &fair_sched_class
;
8898 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8899 alloc_bootmem_cpumask_var(&nohz_cpu_mask
);
8902 alloc_bootmem_cpumask_var(&nohz
.cpu_mask
);
8904 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
8907 scheduler_running
= 1;
8910 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8911 void __might_sleep(char *file
, int line
)
8914 static unsigned long prev_jiffy
; /* ratelimiting */
8916 if ((!in_atomic() && !irqs_disabled()) ||
8917 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
8919 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8921 prev_jiffy
= jiffies
;
8924 "BUG: sleeping function called from invalid context at %s:%d\n",
8927 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8928 in_atomic(), irqs_disabled(),
8929 current
->pid
, current
->comm
);
8931 debug_show_held_locks(current
);
8932 if (irqs_disabled())
8933 print_irqtrace_events(current
);
8937 EXPORT_SYMBOL(__might_sleep
);
8940 #ifdef CONFIG_MAGIC_SYSRQ
8941 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8945 update_rq_clock(rq
);
8946 on_rq
= p
->se
.on_rq
;
8948 deactivate_task(rq
, p
, 0);
8949 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8951 activate_task(rq
, p
, 0);
8952 resched_task(rq
->curr
);
8956 void normalize_rt_tasks(void)
8958 struct task_struct
*g
, *p
;
8959 unsigned long flags
;
8962 read_lock_irqsave(&tasklist_lock
, flags
);
8963 do_each_thread(g
, p
) {
8965 * Only normalize user tasks:
8970 p
->se
.exec_start
= 0;
8971 #ifdef CONFIG_SCHEDSTATS
8972 p
->se
.wait_start
= 0;
8973 p
->se
.sleep_start
= 0;
8974 p
->se
.block_start
= 0;
8979 * Renice negative nice level userspace
8982 if (TASK_NICE(p
) < 0 && p
->mm
)
8983 set_user_nice(p
, 0);
8987 spin_lock(&p
->pi_lock
);
8988 rq
= __task_rq_lock(p
);
8990 normalize_task(rq
, p
);
8992 __task_rq_unlock(rq
);
8993 spin_unlock(&p
->pi_lock
);
8994 } while_each_thread(g
, p
);
8996 read_unlock_irqrestore(&tasklist_lock
, flags
);
8999 #endif /* CONFIG_MAGIC_SYSRQ */
9003 * These functions are only useful for the IA64 MCA handling.
9005 * They can only be called when the whole system has been
9006 * stopped - every CPU needs to be quiescent, and no scheduling
9007 * activity can take place. Using them for anything else would
9008 * be a serious bug, and as a result, they aren't even visible
9009 * under any other configuration.
9013 * curr_task - return the current task for a given cpu.
9014 * @cpu: the processor in question.
9016 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9018 struct task_struct
*curr_task(int cpu
)
9020 return cpu_curr(cpu
);
9024 * set_curr_task - set the current task for a given cpu.
9025 * @cpu: the processor in question.
9026 * @p: the task pointer to set.
9028 * Description: This function must only be used when non-maskable interrupts
9029 * are serviced on a separate stack. It allows the architecture to switch the
9030 * notion of the current task on a cpu in a non-blocking manner. This function
9031 * must be called with all CPU's synchronized, and interrupts disabled, the
9032 * and caller must save the original value of the current task (see
9033 * curr_task() above) and restore that value before reenabling interrupts and
9034 * re-starting the system.
9036 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9038 void set_curr_task(int cpu
, struct task_struct
*p
)
9045 #ifdef CONFIG_FAIR_GROUP_SCHED
9046 static void free_fair_sched_group(struct task_group
*tg
)
9050 for_each_possible_cpu(i
) {
9052 kfree(tg
->cfs_rq
[i
]);
9062 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9064 struct cfs_rq
*cfs_rq
;
9065 struct sched_entity
*se
;
9069 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9072 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
9076 tg
->shares
= NICE_0_LOAD
;
9078 for_each_possible_cpu(i
) {
9081 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
9082 GFP_KERNEL
, cpu_to_node(i
));
9086 se
= kzalloc_node(sizeof(struct sched_entity
),
9087 GFP_KERNEL
, cpu_to_node(i
));
9091 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
9100 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9102 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
9103 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
9106 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9108 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
9110 #else /* !CONFG_FAIR_GROUP_SCHED */
9111 static inline void free_fair_sched_group(struct task_group
*tg
)
9116 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9121 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9125 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9128 #endif /* CONFIG_FAIR_GROUP_SCHED */
9130 #ifdef CONFIG_RT_GROUP_SCHED
9131 static void free_rt_sched_group(struct task_group
*tg
)
9135 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
9137 for_each_possible_cpu(i
) {
9139 kfree(tg
->rt_rq
[i
]);
9141 kfree(tg
->rt_se
[i
]);
9149 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9151 struct rt_rq
*rt_rq
;
9152 struct sched_rt_entity
*rt_se
;
9156 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9159 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
9163 init_rt_bandwidth(&tg
->rt_bandwidth
,
9164 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
9166 for_each_possible_cpu(i
) {
9169 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
9170 GFP_KERNEL
, cpu_to_node(i
));
9174 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
9175 GFP_KERNEL
, cpu_to_node(i
));
9179 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
9188 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9190 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
9191 &cpu_rq(cpu
)->leaf_rt_rq_list
);
9194 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9196 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
9198 #else /* !CONFIG_RT_GROUP_SCHED */
9199 static inline void free_rt_sched_group(struct task_group
*tg
)
9204 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9209 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9213 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9216 #endif /* CONFIG_RT_GROUP_SCHED */
9218 #ifdef CONFIG_GROUP_SCHED
9219 static void free_sched_group(struct task_group
*tg
)
9221 free_fair_sched_group(tg
);
9222 free_rt_sched_group(tg
);
9226 /* allocate runqueue etc for a new task group */
9227 struct task_group
*sched_create_group(struct task_group
*parent
)
9229 struct task_group
*tg
;
9230 unsigned long flags
;
9233 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
9235 return ERR_PTR(-ENOMEM
);
9237 if (!alloc_fair_sched_group(tg
, parent
))
9240 if (!alloc_rt_sched_group(tg
, parent
))
9243 spin_lock_irqsave(&task_group_lock
, flags
);
9244 for_each_possible_cpu(i
) {
9245 register_fair_sched_group(tg
, i
);
9246 register_rt_sched_group(tg
, i
);
9248 list_add_rcu(&tg
->list
, &task_groups
);
9250 WARN_ON(!parent
); /* root should already exist */
9252 tg
->parent
= parent
;
9253 INIT_LIST_HEAD(&tg
->children
);
9254 list_add_rcu(&tg
->siblings
, &parent
->children
);
9255 spin_unlock_irqrestore(&task_group_lock
, flags
);
9260 free_sched_group(tg
);
9261 return ERR_PTR(-ENOMEM
);
9264 /* rcu callback to free various structures associated with a task group */
9265 static void free_sched_group_rcu(struct rcu_head
*rhp
)
9267 /* now it should be safe to free those cfs_rqs */
9268 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
9271 /* Destroy runqueue etc associated with a task group */
9272 void sched_destroy_group(struct task_group
*tg
)
9274 unsigned long flags
;
9277 spin_lock_irqsave(&task_group_lock
, flags
);
9278 for_each_possible_cpu(i
) {
9279 unregister_fair_sched_group(tg
, i
);
9280 unregister_rt_sched_group(tg
, i
);
9282 list_del_rcu(&tg
->list
);
9283 list_del_rcu(&tg
->siblings
);
9284 spin_unlock_irqrestore(&task_group_lock
, flags
);
9286 /* wait for possible concurrent references to cfs_rqs complete */
9287 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
9290 /* change task's runqueue when it moves between groups.
9291 * The caller of this function should have put the task in its new group
9292 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9293 * reflect its new group.
9295 void sched_move_task(struct task_struct
*tsk
)
9298 unsigned long flags
;
9301 rq
= task_rq_lock(tsk
, &flags
);
9303 update_rq_clock(rq
);
9305 running
= task_current(rq
, tsk
);
9306 on_rq
= tsk
->se
.on_rq
;
9309 dequeue_task(rq
, tsk
, 0);
9310 if (unlikely(running
))
9311 tsk
->sched_class
->put_prev_task(rq
, tsk
);
9313 set_task_rq(tsk
, task_cpu(tsk
));
9315 #ifdef CONFIG_FAIR_GROUP_SCHED
9316 if (tsk
->sched_class
->moved_group
)
9317 tsk
->sched_class
->moved_group(tsk
);
9320 if (unlikely(running
))
9321 tsk
->sched_class
->set_curr_task(rq
);
9323 enqueue_task(rq
, tsk
, 0);
9325 task_rq_unlock(rq
, &flags
);
9327 #endif /* CONFIG_GROUP_SCHED */
9329 #ifdef CONFIG_FAIR_GROUP_SCHED
9330 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
9332 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
9337 dequeue_entity(cfs_rq
, se
, 0);
9339 se
->load
.weight
= shares
;
9340 se
->load
.inv_weight
= 0;
9343 enqueue_entity(cfs_rq
, se
, 0);
9346 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
9348 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
9349 struct rq
*rq
= cfs_rq
->rq
;
9350 unsigned long flags
;
9352 spin_lock_irqsave(&rq
->lock
, flags
);
9353 __set_se_shares(se
, shares
);
9354 spin_unlock_irqrestore(&rq
->lock
, flags
);
9357 static DEFINE_MUTEX(shares_mutex
);
9359 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
9362 unsigned long flags
;
9365 * We can't change the weight of the root cgroup.
9370 if (shares
< MIN_SHARES
)
9371 shares
= MIN_SHARES
;
9372 else if (shares
> MAX_SHARES
)
9373 shares
= MAX_SHARES
;
9375 mutex_lock(&shares_mutex
);
9376 if (tg
->shares
== shares
)
9379 spin_lock_irqsave(&task_group_lock
, flags
);
9380 for_each_possible_cpu(i
)
9381 unregister_fair_sched_group(tg
, i
);
9382 list_del_rcu(&tg
->siblings
);
9383 spin_unlock_irqrestore(&task_group_lock
, flags
);
9385 /* wait for any ongoing reference to this group to finish */
9386 synchronize_sched();
9389 * Now we are free to modify the group's share on each cpu
9390 * w/o tripping rebalance_share or load_balance_fair.
9392 tg
->shares
= shares
;
9393 for_each_possible_cpu(i
) {
9397 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
9398 set_se_shares(tg
->se
[i
], shares
);
9402 * Enable load balance activity on this group, by inserting it back on
9403 * each cpu's rq->leaf_cfs_rq_list.
9405 spin_lock_irqsave(&task_group_lock
, flags
);
9406 for_each_possible_cpu(i
)
9407 register_fair_sched_group(tg
, i
);
9408 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
9409 spin_unlock_irqrestore(&task_group_lock
, flags
);
9411 mutex_unlock(&shares_mutex
);
9415 unsigned long sched_group_shares(struct task_group
*tg
)
9421 #ifdef CONFIG_RT_GROUP_SCHED
9423 * Ensure that the real time constraints are schedulable.
9425 static DEFINE_MUTEX(rt_constraints_mutex
);
9427 static unsigned long to_ratio(u64 period
, u64 runtime
)
9429 if (runtime
== RUNTIME_INF
)
9432 return div64_u64(runtime
<< 20, period
);
9435 /* Must be called with tasklist_lock held */
9436 static inline int tg_has_rt_tasks(struct task_group
*tg
)
9438 struct task_struct
*g
, *p
;
9440 do_each_thread(g
, p
) {
9441 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
9443 } while_each_thread(g
, p
);
9448 struct rt_schedulable_data
{
9449 struct task_group
*tg
;
9454 static int tg_schedulable(struct task_group
*tg
, void *data
)
9456 struct rt_schedulable_data
*d
= data
;
9457 struct task_group
*child
;
9458 unsigned long total
, sum
= 0;
9459 u64 period
, runtime
;
9461 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9462 runtime
= tg
->rt_bandwidth
.rt_runtime
;
9465 period
= d
->rt_period
;
9466 runtime
= d
->rt_runtime
;
9469 #ifdef CONFIG_USER_SCHED
9470 if (tg
== &root_task_group
) {
9471 period
= global_rt_period();
9472 runtime
= global_rt_runtime();
9477 * Cannot have more runtime than the period.
9479 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9483 * Ensure we don't starve existing RT tasks.
9485 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
9488 total
= to_ratio(period
, runtime
);
9491 * Nobody can have more than the global setting allows.
9493 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
9497 * The sum of our children's runtime should not exceed our own.
9499 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
9500 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
9501 runtime
= child
->rt_bandwidth
.rt_runtime
;
9503 if (child
== d
->tg
) {
9504 period
= d
->rt_period
;
9505 runtime
= d
->rt_runtime
;
9508 sum
+= to_ratio(period
, runtime
);
9517 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
9519 struct rt_schedulable_data data
= {
9521 .rt_period
= period
,
9522 .rt_runtime
= runtime
,
9525 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
9528 static int tg_set_bandwidth(struct task_group
*tg
,
9529 u64 rt_period
, u64 rt_runtime
)
9533 mutex_lock(&rt_constraints_mutex
);
9534 read_lock(&tasklist_lock
);
9535 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
9539 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
9540 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
9541 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
9543 for_each_possible_cpu(i
) {
9544 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
9546 spin_lock(&rt_rq
->rt_runtime_lock
);
9547 rt_rq
->rt_runtime
= rt_runtime
;
9548 spin_unlock(&rt_rq
->rt_runtime_lock
);
9550 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
9552 read_unlock(&tasklist_lock
);
9553 mutex_unlock(&rt_constraints_mutex
);
9558 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
9560 u64 rt_runtime
, rt_period
;
9562 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9563 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
9564 if (rt_runtime_us
< 0)
9565 rt_runtime
= RUNTIME_INF
;
9567 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9570 long sched_group_rt_runtime(struct task_group
*tg
)
9574 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
9577 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
9578 do_div(rt_runtime_us
, NSEC_PER_USEC
);
9579 return rt_runtime_us
;
9582 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
9584 u64 rt_runtime
, rt_period
;
9586 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
9587 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9592 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9595 long sched_group_rt_period(struct task_group
*tg
)
9599 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9600 do_div(rt_period_us
, NSEC_PER_USEC
);
9601 return rt_period_us
;
9604 static int sched_rt_global_constraints(void)
9606 u64 runtime
, period
;
9609 if (sysctl_sched_rt_period
<= 0)
9612 runtime
= global_rt_runtime();
9613 period
= global_rt_period();
9616 * Sanity check on the sysctl variables.
9618 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9621 mutex_lock(&rt_constraints_mutex
);
9622 read_lock(&tasklist_lock
);
9623 ret
= __rt_schedulable(NULL
, 0, 0);
9624 read_unlock(&tasklist_lock
);
9625 mutex_unlock(&rt_constraints_mutex
);
9630 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
9632 /* Don't accept realtime tasks when there is no way for them to run */
9633 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
9639 #else /* !CONFIG_RT_GROUP_SCHED */
9640 static int sched_rt_global_constraints(void)
9642 unsigned long flags
;
9645 if (sysctl_sched_rt_period
<= 0)
9648 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9649 for_each_possible_cpu(i
) {
9650 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
9652 spin_lock(&rt_rq
->rt_runtime_lock
);
9653 rt_rq
->rt_runtime
= global_rt_runtime();
9654 spin_unlock(&rt_rq
->rt_runtime_lock
);
9656 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9660 #endif /* CONFIG_RT_GROUP_SCHED */
9662 int sched_rt_handler(struct ctl_table
*table
, int write
,
9663 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
9667 int old_period
, old_runtime
;
9668 static DEFINE_MUTEX(mutex
);
9671 old_period
= sysctl_sched_rt_period
;
9672 old_runtime
= sysctl_sched_rt_runtime
;
9674 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
9676 if (!ret
&& write
) {
9677 ret
= sched_rt_global_constraints();
9679 sysctl_sched_rt_period
= old_period
;
9680 sysctl_sched_rt_runtime
= old_runtime
;
9682 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
9683 def_rt_bandwidth
.rt_period
=
9684 ns_to_ktime(global_rt_period());
9687 mutex_unlock(&mutex
);
9692 #ifdef CONFIG_CGROUP_SCHED
9694 /* return corresponding task_group object of a cgroup */
9695 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
9697 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
9698 struct task_group
, css
);
9701 static struct cgroup_subsys_state
*
9702 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9704 struct task_group
*tg
, *parent
;
9706 if (!cgrp
->parent
) {
9707 /* This is early initialization for the top cgroup */
9708 return &init_task_group
.css
;
9711 parent
= cgroup_tg(cgrp
->parent
);
9712 tg
= sched_create_group(parent
);
9714 return ERR_PTR(-ENOMEM
);
9720 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9722 struct task_group
*tg
= cgroup_tg(cgrp
);
9724 sched_destroy_group(tg
);
9728 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9729 struct task_struct
*tsk
)
9731 #ifdef CONFIG_RT_GROUP_SCHED
9732 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
9735 /* We don't support RT-tasks being in separate groups */
9736 if (tsk
->sched_class
!= &fair_sched_class
)
9744 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9745 struct cgroup
*old_cont
, struct task_struct
*tsk
)
9747 sched_move_task(tsk
);
9750 #ifdef CONFIG_FAIR_GROUP_SCHED
9751 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9754 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
9757 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9759 struct task_group
*tg
= cgroup_tg(cgrp
);
9761 return (u64
) tg
->shares
;
9763 #endif /* CONFIG_FAIR_GROUP_SCHED */
9765 #ifdef CONFIG_RT_GROUP_SCHED
9766 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9769 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9772 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9774 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9777 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9780 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9783 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9785 return sched_group_rt_period(cgroup_tg(cgrp
));
9787 #endif /* CONFIG_RT_GROUP_SCHED */
9789 static struct cftype cpu_files
[] = {
9790 #ifdef CONFIG_FAIR_GROUP_SCHED
9793 .read_u64
= cpu_shares_read_u64
,
9794 .write_u64
= cpu_shares_write_u64
,
9797 #ifdef CONFIG_RT_GROUP_SCHED
9799 .name
= "rt_runtime_us",
9800 .read_s64
= cpu_rt_runtime_read
,
9801 .write_s64
= cpu_rt_runtime_write
,
9804 .name
= "rt_period_us",
9805 .read_u64
= cpu_rt_period_read_uint
,
9806 .write_u64
= cpu_rt_period_write_uint
,
9811 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
9813 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
9816 struct cgroup_subsys cpu_cgroup_subsys
= {
9818 .create
= cpu_cgroup_create
,
9819 .destroy
= cpu_cgroup_destroy
,
9820 .can_attach
= cpu_cgroup_can_attach
,
9821 .attach
= cpu_cgroup_attach
,
9822 .populate
= cpu_cgroup_populate
,
9823 .subsys_id
= cpu_cgroup_subsys_id
,
9827 #endif /* CONFIG_CGROUP_SCHED */
9829 #ifdef CONFIG_CGROUP_CPUACCT
9832 * CPU accounting code for task groups.
9834 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9835 * (balbir@in.ibm.com).
9838 /* track cpu usage of a group of tasks and its child groups */
9840 struct cgroup_subsys_state css
;
9841 /* cpuusage holds pointer to a u64-type object on every cpu */
9843 struct cpuacct
*parent
;
9846 struct cgroup_subsys cpuacct_subsys
;
9848 /* return cpu accounting group corresponding to this container */
9849 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
9851 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
9852 struct cpuacct
, css
);
9855 /* return cpu accounting group to which this task belongs */
9856 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
9858 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
9859 struct cpuacct
, css
);
9862 /* create a new cpu accounting group */
9863 static struct cgroup_subsys_state
*cpuacct_create(
9864 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9866 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
9869 return ERR_PTR(-ENOMEM
);
9871 ca
->cpuusage
= alloc_percpu(u64
);
9872 if (!ca
->cpuusage
) {
9874 return ERR_PTR(-ENOMEM
);
9878 ca
->parent
= cgroup_ca(cgrp
->parent
);
9883 /* destroy an existing cpu accounting group */
9885 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9887 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9889 free_percpu(ca
->cpuusage
);
9893 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
9895 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, cpu
);
9898 #ifndef CONFIG_64BIT
9900 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9902 spin_lock_irq(&cpu_rq(cpu
)->lock
);
9904 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9912 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
9914 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, cpu
);
9916 #ifndef CONFIG_64BIT
9918 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9920 spin_lock_irq(&cpu_rq(cpu
)->lock
);
9922 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9928 /* return total cpu usage (in nanoseconds) of a group */
9929 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9931 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9932 u64 totalcpuusage
= 0;
9935 for_each_present_cpu(i
)
9936 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
9938 return totalcpuusage
;
9941 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9944 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9953 for_each_present_cpu(i
)
9954 cpuacct_cpuusage_write(ca
, i
, 0);
9960 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
9963 struct cpuacct
*ca
= cgroup_ca(cgroup
);
9967 for_each_present_cpu(i
) {
9968 percpu
= cpuacct_cpuusage_read(ca
, i
);
9969 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
9971 seq_printf(m
, "\n");
9975 static struct cftype files
[] = {
9978 .read_u64
= cpuusage_read
,
9979 .write_u64
= cpuusage_write
,
9982 .name
= "usage_percpu",
9983 .read_seq_string
= cpuacct_percpu_seq_read
,
9988 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9990 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9994 * charge this task's execution time to its accounting group.
9996 * called with rq->lock held.
9998 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
10000 struct cpuacct
*ca
;
10003 if (unlikely(!cpuacct_subsys
.active
))
10006 cpu
= task_cpu(tsk
);
10009 for (; ca
; ca
= ca
->parent
) {
10010 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, cpu
);
10011 *cpuusage
+= cputime
;
10015 struct cgroup_subsys cpuacct_subsys
= {
10017 .create
= cpuacct_create
,
10018 .destroy
= cpuacct_destroy
,
10019 .populate
= cpuacct_populate
,
10020 .subsys_id
= cpuacct_subsys_id
,
10022 #endif /* CONFIG_CGROUP_CPUACCT */