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
);
334 #ifdef CONFIG_FAIR_GROUP_SCHED
335 #ifdef CONFIG_USER_SCHED
336 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
337 #else /* !CONFIG_USER_SCHED */
338 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
339 #endif /* CONFIG_USER_SCHED */
342 * A weight of 0 or 1 can cause arithmetics problems.
343 * A weight of a cfs_rq is the sum of weights of which entities
344 * are queued on this cfs_rq, so a weight of a entity should not be
345 * too large, so as the shares value of a task group.
346 * (The default weight is 1024 - so there's no practical
347 * limitation from this.)
350 #define MAX_SHARES (1UL << 18)
352 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
355 /* Default task group.
356 * Every task in system belong to this group at bootup.
358 struct task_group init_task_group
;
360 /* return group to which a task belongs */
361 static inline struct task_group
*task_group(struct task_struct
*p
)
363 struct task_group
*tg
;
365 #ifdef CONFIG_USER_SCHED
367 tg
= __task_cred(p
)->user
->tg
;
369 #elif defined(CONFIG_CGROUP_SCHED)
370 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
371 struct task_group
, css
);
373 tg
= &init_task_group
;
378 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
379 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
381 #ifdef CONFIG_FAIR_GROUP_SCHED
382 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
383 p
->se
.parent
= task_group(p
)->se
[cpu
];
386 #ifdef CONFIG_RT_GROUP_SCHED
387 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
388 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
394 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
395 static inline struct task_group
*task_group(struct task_struct
*p
)
400 #endif /* CONFIG_GROUP_SCHED */
402 /* CFS-related fields in a runqueue */
404 struct load_weight load
;
405 unsigned long nr_running
;
410 struct rb_root tasks_timeline
;
411 struct rb_node
*rb_leftmost
;
413 struct list_head tasks
;
414 struct list_head
*balance_iterator
;
417 * 'curr' points to currently running entity on this cfs_rq.
418 * It is set to NULL otherwise (i.e when none are currently running).
420 struct sched_entity
*curr
, *next
, *last
;
422 unsigned int nr_spread_over
;
424 #ifdef CONFIG_FAIR_GROUP_SCHED
425 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
428 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
429 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
430 * (like users, containers etc.)
432 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
433 * list is used during load balance.
435 struct list_head leaf_cfs_rq_list
;
436 struct task_group
*tg
; /* group that "owns" this runqueue */
440 * the part of load.weight contributed by tasks
442 unsigned long task_weight
;
445 * h_load = weight * f(tg)
447 * Where f(tg) is the recursive weight fraction assigned to
450 unsigned long h_load
;
453 * this cpu's part of tg->shares
455 unsigned long shares
;
458 * load.weight at the time we set shares
460 unsigned long rq_weight
;
465 /* Real-Time classes' related field in a runqueue: */
467 struct rt_prio_array active
;
468 unsigned long rt_nr_running
;
469 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
470 int highest_prio
; /* highest queued rt task prio */
473 unsigned long rt_nr_migratory
;
479 /* Nests inside the rq lock: */
480 spinlock_t rt_runtime_lock
;
482 #ifdef CONFIG_RT_GROUP_SCHED
483 unsigned long rt_nr_boosted
;
486 struct list_head leaf_rt_rq_list
;
487 struct task_group
*tg
;
488 struct sched_rt_entity
*rt_se
;
495 * We add the notion of a root-domain which will be used to define per-domain
496 * variables. Each exclusive cpuset essentially defines an island domain by
497 * fully partitioning the member cpus from any other cpuset. Whenever a new
498 * exclusive cpuset is created, we also create and attach a new root-domain
505 cpumask_var_t online
;
508 * The "RT overload" flag: it gets set if a CPU has more than
509 * one runnable RT task.
511 cpumask_var_t rto_mask
;
514 struct cpupri cpupri
;
516 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
518 * Preferred wake up cpu nominated by sched_mc balance that will be
519 * used when most cpus are idle in the system indicating overall very
520 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
522 unsigned int sched_mc_preferred_wakeup_cpu
;
527 * By default the system creates a single root-domain with all cpus as
528 * members (mimicking the global state we have today).
530 static struct root_domain def_root_domain
;
535 * This is the main, per-CPU runqueue data structure.
537 * Locking rule: those places that want to lock multiple runqueues
538 * (such as the load balancing or the thread migration code), lock
539 * acquire operations must be ordered by ascending &runqueue.
546 * nr_running and cpu_load should be in the same cacheline because
547 * remote CPUs use both these fields when doing load calculation.
549 unsigned long nr_running
;
550 #define CPU_LOAD_IDX_MAX 5
551 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
552 unsigned char idle_at_tick
;
554 unsigned long last_tick_seen
;
555 unsigned char in_nohz_recently
;
557 /* capture load from *all* tasks on this cpu: */
558 struct load_weight load
;
559 unsigned long nr_load_updates
;
561 u64 nr_migrations_in
;
566 #ifdef CONFIG_FAIR_GROUP_SCHED
567 /* list of leaf cfs_rq on this cpu: */
568 struct list_head leaf_cfs_rq_list
;
570 #ifdef CONFIG_RT_GROUP_SCHED
571 struct list_head leaf_rt_rq_list
;
575 * This is part of a global counter where only the total sum
576 * over all CPUs matters. A task can increase this counter on
577 * one CPU and if it got migrated afterwards it may decrease
578 * it on another CPU. Always updated under the runqueue lock:
580 unsigned long nr_uninterruptible
;
582 struct task_struct
*curr
, *idle
;
583 unsigned long next_balance
;
584 struct mm_struct
*prev_mm
;
591 struct root_domain
*rd
;
592 struct sched_domain
*sd
;
594 /* For active balancing */
597 /* cpu of this runqueue: */
601 unsigned long avg_load_per_task
;
603 struct task_struct
*migration_thread
;
604 struct list_head migration_queue
;
607 #ifdef CONFIG_SCHED_HRTICK
609 int hrtick_csd_pending
;
610 struct call_single_data hrtick_csd
;
612 struct hrtimer hrtick_timer
;
615 #ifdef CONFIG_SCHEDSTATS
617 struct sched_info rq_sched_info
;
618 unsigned long long rq_cpu_time
;
619 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
621 /* sys_sched_yield() stats */
622 unsigned int yld_exp_empty
;
623 unsigned int yld_act_empty
;
624 unsigned int yld_both_empty
;
625 unsigned int yld_count
;
627 /* schedule() stats */
628 unsigned int sched_switch
;
629 unsigned int sched_count
;
630 unsigned int sched_goidle
;
632 /* try_to_wake_up() stats */
633 unsigned int ttwu_count
;
634 unsigned int ttwu_local
;
637 unsigned int bkl_count
;
641 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
643 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int sync
)
645 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, sync
);
648 static inline int cpu_of(struct rq
*rq
)
658 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
659 * See detach_destroy_domains: synchronize_sched for details.
661 * The domain tree of any CPU may only be accessed from within
662 * preempt-disabled sections.
664 #define for_each_domain(cpu, __sd) \
665 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
667 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
668 #define this_rq() (&__get_cpu_var(runqueues))
669 #define task_rq(p) cpu_rq(task_cpu(p))
670 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
672 inline void update_rq_clock(struct rq
*rq
)
674 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
678 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
680 #ifdef CONFIG_SCHED_DEBUG
681 # define const_debug __read_mostly
683 # define const_debug static const
689 * Returns true if the current cpu runqueue is locked.
690 * This interface allows printk to be called with the runqueue lock
691 * held and know whether or not it is OK to wake up the klogd.
693 int runqueue_is_locked(void)
696 struct rq
*rq
= cpu_rq(cpu
);
699 ret
= spin_is_locked(&rq
->lock
);
705 * Debugging: various feature bits
708 #define SCHED_FEAT(name, enabled) \
709 __SCHED_FEAT_##name ,
712 #include "sched_features.h"
717 #define SCHED_FEAT(name, enabled) \
718 (1UL << __SCHED_FEAT_##name) * enabled |
720 const_debug
unsigned int sysctl_sched_features
=
721 #include "sched_features.h"
726 #ifdef CONFIG_SCHED_DEBUG
727 #define SCHED_FEAT(name, enabled) \
730 static __read_mostly
char *sched_feat_names
[] = {
731 #include "sched_features.h"
737 static int sched_feat_show(struct seq_file
*m
, void *v
)
741 for (i
= 0; sched_feat_names
[i
]; i
++) {
742 if (!(sysctl_sched_features
& (1UL << i
)))
744 seq_printf(m
, "%s ", sched_feat_names
[i
]);
752 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
753 size_t cnt
, loff_t
*ppos
)
763 if (copy_from_user(&buf
, ubuf
, cnt
))
768 if (strncmp(buf
, "NO_", 3) == 0) {
773 for (i
= 0; sched_feat_names
[i
]; i
++) {
774 int len
= strlen(sched_feat_names
[i
]);
776 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
778 sysctl_sched_features
&= ~(1UL << i
);
780 sysctl_sched_features
|= (1UL << i
);
785 if (!sched_feat_names
[i
])
793 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
795 return single_open(filp
, sched_feat_show
, NULL
);
798 static struct file_operations sched_feat_fops
= {
799 .open
= sched_feat_open
,
800 .write
= sched_feat_write
,
803 .release
= single_release
,
806 static __init
int sched_init_debug(void)
808 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
813 late_initcall(sched_init_debug
);
817 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
820 * Number of tasks to iterate in a single balance run.
821 * Limited because this is done with IRQs disabled.
823 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
826 * ratelimit for updating the group shares.
829 unsigned int sysctl_sched_shares_ratelimit
= 250000;
832 * Inject some fuzzyness into changing the per-cpu group shares
833 * this avoids remote rq-locks at the expense of fairness.
836 unsigned int sysctl_sched_shares_thresh
= 4;
839 * period over which we measure -rt task cpu usage in us.
842 unsigned int sysctl_sched_rt_period
= 1000000;
844 static __read_mostly
int scheduler_running
;
847 * part of the period that we allow rt tasks to run in us.
850 int sysctl_sched_rt_runtime
= 950000;
852 static inline u64
global_rt_period(void)
854 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
857 static inline u64
global_rt_runtime(void)
859 if (sysctl_sched_rt_runtime
< 0)
862 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
865 #ifndef prepare_arch_switch
866 # define prepare_arch_switch(next) do { } while (0)
868 #ifndef finish_arch_switch
869 # define finish_arch_switch(prev) do { } while (0)
872 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
874 return rq
->curr
== p
;
877 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
878 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
880 return task_current(rq
, p
);
883 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
887 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
889 #ifdef CONFIG_DEBUG_SPINLOCK
890 /* this is a valid case when another task releases the spinlock */
891 rq
->lock
.owner
= current
;
894 * If we are tracking spinlock dependencies then we have to
895 * fix up the runqueue lock - which gets 'carried over' from
898 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
900 spin_unlock_irq(&rq
->lock
);
903 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
904 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
909 return task_current(rq
, p
);
913 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
917 * We can optimise this out completely for !SMP, because the
918 * SMP rebalancing from interrupt is the only thing that cares
923 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
924 spin_unlock_irq(&rq
->lock
);
926 spin_unlock(&rq
->lock
);
930 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
934 * After ->oncpu is cleared, the task can be moved to a different CPU.
935 * We must ensure this doesn't happen until the switch is completely
941 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
945 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
948 * __task_rq_lock - lock the runqueue a given task resides on.
949 * Must be called interrupts disabled.
951 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
955 struct rq
*rq
= task_rq(p
);
956 spin_lock(&rq
->lock
);
957 if (likely(rq
== task_rq(p
)))
959 spin_unlock(&rq
->lock
);
964 * task_rq_lock - lock the runqueue a given task resides on and disable
965 * interrupts. Note the ordering: we can safely lookup the task_rq without
966 * explicitly disabling preemption.
968 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
974 local_irq_save(*flags
);
976 spin_lock(&rq
->lock
);
977 if (likely(rq
== task_rq(p
)))
979 spin_unlock_irqrestore(&rq
->lock
, *flags
);
983 void curr_rq_lock_irq_save(unsigned long *flags
)
988 local_irq_save(*flags
);
989 rq
= cpu_rq(smp_processor_id());
990 spin_lock(&rq
->lock
);
993 void curr_rq_unlock_irq_restore(unsigned long *flags
)
998 rq
= cpu_rq(smp_processor_id());
999 spin_unlock(&rq
->lock
);
1000 local_irq_restore(*flags
);
1003 void task_rq_unlock_wait(struct task_struct
*p
)
1005 struct rq
*rq
= task_rq(p
);
1007 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1008 spin_unlock_wait(&rq
->lock
);
1011 static void __task_rq_unlock(struct rq
*rq
)
1012 __releases(rq
->lock
)
1014 spin_unlock(&rq
->lock
);
1017 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
1018 __releases(rq
->lock
)
1020 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1024 * this_rq_lock - lock this runqueue and disable interrupts.
1026 static struct rq
*this_rq_lock(void)
1027 __acquires(rq
->lock
)
1031 local_irq_disable();
1033 spin_lock(&rq
->lock
);
1038 #ifdef CONFIG_SCHED_HRTICK
1040 * Use HR-timers to deliver accurate preemption points.
1042 * Its all a bit involved since we cannot program an hrt while holding the
1043 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1046 * When we get rescheduled we reprogram the hrtick_timer outside of the
1052 * - enabled by features
1053 * - hrtimer is actually high res
1055 static inline int hrtick_enabled(struct rq
*rq
)
1057 if (!sched_feat(HRTICK
))
1059 if (!cpu_active(cpu_of(rq
)))
1061 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1064 static void hrtick_clear(struct rq
*rq
)
1066 if (hrtimer_active(&rq
->hrtick_timer
))
1067 hrtimer_cancel(&rq
->hrtick_timer
);
1071 * High-resolution timer tick.
1072 * Runs from hardirq context with interrupts disabled.
1074 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1076 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1078 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1080 spin_lock(&rq
->lock
);
1081 update_rq_clock(rq
);
1082 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1083 spin_unlock(&rq
->lock
);
1085 return HRTIMER_NORESTART
;
1090 * called from hardirq (IPI) context
1092 static void __hrtick_start(void *arg
)
1094 struct rq
*rq
= arg
;
1096 spin_lock(&rq
->lock
);
1097 hrtimer_restart(&rq
->hrtick_timer
);
1098 rq
->hrtick_csd_pending
= 0;
1099 spin_unlock(&rq
->lock
);
1103 * Called to set the hrtick timer state.
1105 * called with rq->lock held and irqs disabled
1107 static void hrtick_start(struct rq
*rq
, u64 delay
)
1109 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1110 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1112 hrtimer_set_expires(timer
, time
);
1114 if (rq
== this_rq()) {
1115 hrtimer_restart(timer
);
1116 } else if (!rq
->hrtick_csd_pending
) {
1117 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
);
1118 rq
->hrtick_csd_pending
= 1;
1123 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1125 int cpu
= (int)(long)hcpu
;
1128 case CPU_UP_CANCELED
:
1129 case CPU_UP_CANCELED_FROZEN
:
1130 case CPU_DOWN_PREPARE
:
1131 case CPU_DOWN_PREPARE_FROZEN
:
1133 case CPU_DEAD_FROZEN
:
1134 hrtick_clear(cpu_rq(cpu
));
1141 static __init
void init_hrtick(void)
1143 hotcpu_notifier(hotplug_hrtick
, 0);
1147 * Called to set the hrtick timer state.
1149 * called with rq->lock held and irqs disabled
1151 static void hrtick_start(struct rq
*rq
, u64 delay
)
1153 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
), HRTIMER_MODE_REL
);
1156 static inline void init_hrtick(void)
1159 #endif /* CONFIG_SMP */
1161 static void init_rq_hrtick(struct rq
*rq
)
1164 rq
->hrtick_csd_pending
= 0;
1166 rq
->hrtick_csd
.flags
= 0;
1167 rq
->hrtick_csd
.func
= __hrtick_start
;
1168 rq
->hrtick_csd
.info
= rq
;
1171 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1172 rq
->hrtick_timer
.function
= hrtick
;
1174 #else /* CONFIG_SCHED_HRTICK */
1175 static inline void hrtick_clear(struct rq
*rq
)
1179 static inline void init_rq_hrtick(struct rq
*rq
)
1183 static inline void init_hrtick(void)
1186 #endif /* CONFIG_SCHED_HRTICK */
1189 * resched_task - mark a task 'to be rescheduled now'.
1191 * On UP this means the setting of the need_resched flag, on SMP it
1192 * might also involve a cross-CPU call to trigger the scheduler on
1197 #ifndef tsk_is_polling
1198 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1201 static void resched_task(struct task_struct
*p
)
1205 assert_spin_locked(&task_rq(p
)->lock
);
1207 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
1210 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
1213 if (cpu
== smp_processor_id())
1216 /* NEED_RESCHED must be visible before we test polling */
1218 if (!tsk_is_polling(p
))
1219 smp_send_reschedule(cpu
);
1222 static void resched_cpu(int cpu
)
1224 struct rq
*rq
= cpu_rq(cpu
);
1225 unsigned long flags
;
1227 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1229 resched_task(cpu_curr(cpu
));
1230 spin_unlock_irqrestore(&rq
->lock
, flags
);
1235 * When add_timer_on() enqueues a timer into the timer wheel of an
1236 * idle CPU then this timer might expire before the next timer event
1237 * which is scheduled to wake up that CPU. In case of a completely
1238 * idle system the next event might even be infinite time into the
1239 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1240 * leaves the inner idle loop so the newly added timer is taken into
1241 * account when the CPU goes back to idle and evaluates the timer
1242 * wheel for the next timer event.
1244 void wake_up_idle_cpu(int cpu
)
1246 struct rq
*rq
= cpu_rq(cpu
);
1248 if (cpu
== smp_processor_id())
1252 * This is safe, as this function is called with the timer
1253 * wheel base lock of (cpu) held. When the CPU is on the way
1254 * to idle and has not yet set rq->curr to idle then it will
1255 * be serialized on the timer wheel base lock and take the new
1256 * timer into account automatically.
1258 if (rq
->curr
!= rq
->idle
)
1262 * We can set TIF_RESCHED on the idle task of the other CPU
1263 * lockless. The worst case is that the other CPU runs the
1264 * idle task through an additional NOOP schedule()
1266 set_tsk_thread_flag(rq
->idle
, TIF_NEED_RESCHED
);
1268 /* NEED_RESCHED must be visible before we test polling */
1270 if (!tsk_is_polling(rq
->idle
))
1271 smp_send_reschedule(cpu
);
1273 #endif /* CONFIG_NO_HZ */
1275 #else /* !CONFIG_SMP */
1276 static void resched_task(struct task_struct
*p
)
1278 assert_spin_locked(&task_rq(p
)->lock
);
1279 set_tsk_need_resched(p
);
1281 #endif /* CONFIG_SMP */
1283 #if BITS_PER_LONG == 32
1284 # define WMULT_CONST (~0UL)
1286 # define WMULT_CONST (1UL << 32)
1289 #define WMULT_SHIFT 32
1292 * Shift right and round:
1294 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1297 * delta *= weight / lw
1299 static unsigned long
1300 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1301 struct load_weight
*lw
)
1305 if (!lw
->inv_weight
) {
1306 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1309 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1313 tmp
= (u64
)delta_exec
* weight
;
1315 * Check whether we'd overflow the 64-bit multiplication:
1317 if (unlikely(tmp
> WMULT_CONST
))
1318 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1321 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1323 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1326 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1332 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1339 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1340 * of tasks with abnormal "nice" values across CPUs the contribution that
1341 * each task makes to its run queue's load is weighted according to its
1342 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1343 * scaled version of the new time slice allocation that they receive on time
1347 #define WEIGHT_IDLEPRIO 3
1348 #define WMULT_IDLEPRIO 1431655765
1351 * Nice levels are multiplicative, with a gentle 10% change for every
1352 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1353 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1354 * that remained on nice 0.
1356 * The "10% effect" is relative and cumulative: from _any_ nice level,
1357 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1358 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1359 * If a task goes up by ~10% and another task goes down by ~10% then
1360 * the relative distance between them is ~25%.)
1362 static const int prio_to_weight
[40] = {
1363 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1364 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1365 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1366 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1367 /* 0 */ 1024, 820, 655, 526, 423,
1368 /* 5 */ 335, 272, 215, 172, 137,
1369 /* 10 */ 110, 87, 70, 56, 45,
1370 /* 15 */ 36, 29, 23, 18, 15,
1374 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1376 * In cases where the weight does not change often, we can use the
1377 * precalculated inverse to speed up arithmetics by turning divisions
1378 * into multiplications:
1380 static const u32 prio_to_wmult
[40] = {
1381 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1382 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1383 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1384 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1385 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1386 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1387 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1388 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1391 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1394 * runqueue iterator, to support SMP load-balancing between different
1395 * scheduling classes, without having to expose their internal data
1396 * structures to the load-balancing proper:
1398 struct rq_iterator
{
1400 struct task_struct
*(*start
)(void *);
1401 struct task_struct
*(*next
)(void *);
1405 static unsigned long
1406 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1407 unsigned long max_load_move
, struct sched_domain
*sd
,
1408 enum cpu_idle_type idle
, int *all_pinned
,
1409 int *this_best_prio
, struct rq_iterator
*iterator
);
1412 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1413 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1414 struct rq_iterator
*iterator
);
1417 #ifdef CONFIG_CGROUP_CPUACCT
1418 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1420 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1423 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1425 update_load_add(&rq
->load
, load
);
1428 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1430 update_load_sub(&rq
->load
, load
);
1433 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1434 typedef int (*tg_visitor
)(struct task_group
*, void *);
1437 * Iterate the full tree, calling @down when first entering a node and @up when
1438 * leaving it for the final time.
1440 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1442 struct task_group
*parent
, *child
;
1446 parent
= &root_task_group
;
1448 ret
= (*down
)(parent
, data
);
1451 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1458 ret
= (*up
)(parent
, data
);
1463 parent
= parent
->parent
;
1472 static int tg_nop(struct task_group
*tg
, void *data
)
1479 static unsigned long source_load(int cpu
, int type
);
1480 static unsigned long target_load(int cpu
, int type
);
1481 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1483 static unsigned long cpu_avg_load_per_task(int cpu
)
1485 struct rq
*rq
= cpu_rq(cpu
);
1486 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1489 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1491 rq
->avg_load_per_task
= 0;
1493 return rq
->avg_load_per_task
;
1496 #ifdef CONFIG_FAIR_GROUP_SCHED
1498 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1501 * Calculate and set the cpu's group shares.
1504 update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1505 unsigned long sd_shares
, unsigned long sd_rq_weight
)
1507 unsigned long shares
;
1508 unsigned long rq_weight
;
1513 rq_weight
= tg
->cfs_rq
[cpu
]->rq_weight
;
1516 * \Sum shares * rq_weight
1517 * shares = -----------------------
1521 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1522 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1524 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1525 sysctl_sched_shares_thresh
) {
1526 struct rq
*rq
= cpu_rq(cpu
);
1527 unsigned long flags
;
1529 spin_lock_irqsave(&rq
->lock
, flags
);
1530 tg
->cfs_rq
[cpu
]->shares
= shares
;
1532 __set_se_shares(tg
->se
[cpu
], shares
);
1533 spin_unlock_irqrestore(&rq
->lock
, flags
);
1538 * Re-compute the task group their per cpu shares over the given domain.
1539 * This needs to be done in a bottom-up fashion because the rq weight of a
1540 * parent group depends on the shares of its child groups.
1542 static int tg_shares_up(struct task_group
*tg
, void *data
)
1544 unsigned long weight
, rq_weight
= 0;
1545 unsigned long shares
= 0;
1546 struct sched_domain
*sd
= data
;
1549 for_each_cpu(i
, sched_domain_span(sd
)) {
1551 * If there are currently no tasks on the cpu pretend there
1552 * is one of average load so that when a new task gets to
1553 * run here it will not get delayed by group starvation.
1555 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1557 weight
= NICE_0_LOAD
;
1559 tg
->cfs_rq
[i
]->rq_weight
= weight
;
1560 rq_weight
+= weight
;
1561 shares
+= tg
->cfs_rq
[i
]->shares
;
1564 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1565 shares
= tg
->shares
;
1567 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1568 shares
= tg
->shares
;
1570 for_each_cpu(i
, sched_domain_span(sd
))
1571 update_group_shares_cpu(tg
, i
, shares
, rq_weight
);
1577 * Compute the cpu's hierarchical load factor for each task group.
1578 * This needs to be done in a top-down fashion because the load of a child
1579 * group is a fraction of its parents load.
1581 static int tg_load_down(struct task_group
*tg
, void *data
)
1584 long cpu
= (long)data
;
1587 load
= cpu_rq(cpu
)->load
.weight
;
1589 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1590 load
*= tg
->cfs_rq
[cpu
]->shares
;
1591 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1594 tg
->cfs_rq
[cpu
]->h_load
= load
;
1599 static void update_shares(struct sched_domain
*sd
)
1601 u64 now
= cpu_clock(raw_smp_processor_id());
1602 s64 elapsed
= now
- sd
->last_update
;
1604 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1605 sd
->last_update
= now
;
1606 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1610 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1612 spin_unlock(&rq
->lock
);
1614 spin_lock(&rq
->lock
);
1617 static void update_h_load(long cpu
)
1619 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1624 static inline void update_shares(struct sched_domain
*sd
)
1628 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1635 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1637 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1638 __releases(this_rq
->lock
)
1639 __acquires(busiest
->lock
)
1640 __acquires(this_rq
->lock
)
1644 if (unlikely(!irqs_disabled())) {
1645 /* printk() doesn't work good under rq->lock */
1646 spin_unlock(&this_rq
->lock
);
1649 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1650 if (busiest
< this_rq
) {
1651 spin_unlock(&this_rq
->lock
);
1652 spin_lock(&busiest
->lock
);
1653 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
1656 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
1661 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1662 __releases(busiest
->lock
)
1664 spin_unlock(&busiest
->lock
);
1665 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1669 #ifdef CONFIG_FAIR_GROUP_SCHED
1670 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1673 cfs_rq
->shares
= shares
;
1678 #include "sched_stats.h"
1679 #include "sched_idletask.c"
1680 #include "sched_fair.c"
1681 #include "sched_rt.c"
1682 #ifdef CONFIG_SCHED_DEBUG
1683 # include "sched_debug.c"
1686 #define sched_class_highest (&rt_sched_class)
1687 #define for_each_class(class) \
1688 for (class = sched_class_highest; class; class = class->next)
1690 static void inc_nr_running(struct rq
*rq
)
1695 static void dec_nr_running(struct rq
*rq
)
1700 static void set_load_weight(struct task_struct
*p
)
1702 if (task_has_rt_policy(p
)) {
1703 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1704 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1709 * SCHED_IDLE tasks get minimal weight:
1711 if (p
->policy
== SCHED_IDLE
) {
1712 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1713 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1717 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1718 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1721 static void update_avg(u64
*avg
, u64 sample
)
1723 s64 diff
= sample
- *avg
;
1727 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1729 sched_info_queued(p
);
1730 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1734 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1736 if (sleep
&& p
->se
.last_wakeup
) {
1737 update_avg(&p
->se
.avg_overlap
,
1738 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1739 p
->se
.last_wakeup
= 0;
1742 sched_info_dequeued(p
);
1743 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1748 * __normal_prio - return the priority that is based on the static prio
1750 static inline int __normal_prio(struct task_struct
*p
)
1752 return p
->static_prio
;
1756 * Calculate the expected normal priority: i.e. priority
1757 * without taking RT-inheritance into account. Might be
1758 * boosted by interactivity modifiers. Changes upon fork,
1759 * setprio syscalls, and whenever the interactivity
1760 * estimator recalculates.
1762 static inline int normal_prio(struct task_struct
*p
)
1766 if (task_has_rt_policy(p
))
1767 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1769 prio
= __normal_prio(p
);
1774 * Calculate the current priority, i.e. the priority
1775 * taken into account by the scheduler. This value might
1776 * be boosted by RT tasks, or might be boosted by
1777 * interactivity modifiers. Will be RT if the task got
1778 * RT-boosted. If not then it returns p->normal_prio.
1780 static int effective_prio(struct task_struct
*p
)
1782 p
->normal_prio
= normal_prio(p
);
1784 * If we are RT tasks or we were boosted to RT priority,
1785 * keep the priority unchanged. Otherwise, update priority
1786 * to the normal priority:
1788 if (!rt_prio(p
->prio
))
1789 return p
->normal_prio
;
1794 * activate_task - move a task to the runqueue.
1796 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1798 if (task_contributes_to_load(p
))
1799 rq
->nr_uninterruptible
--;
1801 enqueue_task(rq
, p
, wakeup
);
1806 * deactivate_task - remove a task from the runqueue.
1808 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1810 if (task_contributes_to_load(p
))
1811 rq
->nr_uninterruptible
++;
1813 dequeue_task(rq
, p
, sleep
);
1818 * task_curr - is this task currently executing on a CPU?
1819 * @p: the task in question.
1821 inline int task_curr(const struct task_struct
*p
)
1823 return cpu_curr(task_cpu(p
)) == p
;
1826 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1828 set_task_rq(p
, cpu
);
1831 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1832 * successfuly executed on another CPU. We must ensure that updates of
1833 * per-task data have been completed by this moment.
1836 task_thread_info(p
)->cpu
= cpu
;
1840 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1841 const struct sched_class
*prev_class
,
1842 int oldprio
, int running
)
1844 if (prev_class
!= p
->sched_class
) {
1845 if (prev_class
->switched_from
)
1846 prev_class
->switched_from(rq
, p
, running
);
1847 p
->sched_class
->switched_to(rq
, p
, running
);
1849 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1854 /* Used instead of source_load when we know the type == 0 */
1855 static unsigned long weighted_cpuload(const int cpu
)
1857 return cpu_rq(cpu
)->load
.weight
;
1861 * Is this task likely cache-hot:
1864 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1869 * Buddy candidates are cache hot:
1871 if (sched_feat(CACHE_HOT_BUDDY
) &&
1872 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
1873 &p
->se
== cfs_rq_of(&p
->se
)->last
))
1876 if (p
->sched_class
!= &fair_sched_class
)
1879 if (sysctl_sched_migration_cost
== -1)
1881 if (sysctl_sched_migration_cost
== 0)
1884 delta
= now
- p
->se
.exec_start
;
1886 return delta
< (s64
)sysctl_sched_migration_cost
;
1890 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1892 int old_cpu
= task_cpu(p
);
1893 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1894 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1895 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1898 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1900 trace_sched_migrate_task(p
, task_cpu(p
), new_cpu
);
1902 #ifdef CONFIG_SCHEDSTATS
1903 if (p
->se
.wait_start
)
1904 p
->se
.wait_start
-= clock_offset
;
1905 if (p
->se
.sleep_start
)
1906 p
->se
.sleep_start
-= clock_offset
;
1907 if (p
->se
.block_start
)
1908 p
->se
.block_start
-= clock_offset
;
1910 if (old_cpu
!= new_cpu
) {
1911 p
->se
.nr_migrations
++;
1912 new_rq
->nr_migrations_in
++;
1913 #ifdef CONFIG_SCHEDSTATS
1914 if (task_hot(p
, old_rq
->clock
, NULL
))
1915 schedstat_inc(p
, se
.nr_forced2_migrations
);
1918 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1919 new_cfsrq
->min_vruntime
;
1921 __set_task_cpu(p
, new_cpu
);
1924 struct migration_req
{
1925 struct list_head list
;
1927 struct task_struct
*task
;
1930 struct completion done
;
1934 * The task's runqueue lock must be held.
1935 * Returns true if you have to wait for migration thread.
1938 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1940 struct rq
*rq
= task_rq(p
);
1943 * If the task is not on a runqueue (and not running), then
1944 * it is sufficient to simply update the task's cpu field.
1946 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1947 set_task_cpu(p
, dest_cpu
);
1951 init_completion(&req
->done
);
1953 req
->dest_cpu
= dest_cpu
;
1954 list_add(&req
->list
, &rq
->migration_queue
);
1960 * wait_task_inactive - wait for a thread to unschedule.
1962 * If @match_state is nonzero, it's the @p->state value just checked and
1963 * not expected to change. If it changes, i.e. @p might have woken up,
1964 * then return zero. When we succeed in waiting for @p to be off its CPU,
1965 * we return a positive number (its total switch count). If a second call
1966 * a short while later returns the same number, the caller can be sure that
1967 * @p has remained unscheduled the whole time.
1969 * The caller must ensure that the task *will* unschedule sometime soon,
1970 * else this function might spin for a *long* time. This function can't
1971 * be called with interrupts off, or it may introduce deadlock with
1972 * smp_call_function() if an IPI is sent by the same process we are
1973 * waiting to become inactive.
1975 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1977 unsigned long flags
;
1984 * We do the initial early heuristics without holding
1985 * any task-queue locks at all. We'll only try to get
1986 * the runqueue lock when things look like they will
1992 * If the task is actively running on another CPU
1993 * still, just relax and busy-wait without holding
1996 * NOTE! Since we don't hold any locks, it's not
1997 * even sure that "rq" stays as the right runqueue!
1998 * But we don't care, since "task_running()" will
1999 * return false if the runqueue has changed and p
2000 * is actually now running somewhere else!
2002 while (task_running(rq
, p
)) {
2003 if (match_state
&& unlikely(p
->state
!= match_state
))
2009 * Ok, time to look more closely! We need the rq
2010 * lock now, to be *sure*. If we're wrong, we'll
2011 * just go back and repeat.
2013 rq
= task_rq_lock(p
, &flags
);
2014 trace_sched_wait_task(rq
, p
);
2015 running
= task_running(rq
, p
);
2016 on_rq
= p
->se
.on_rq
;
2018 if (!match_state
|| p
->state
== match_state
)
2019 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2020 task_rq_unlock(rq
, &flags
);
2023 * If it changed from the expected state, bail out now.
2025 if (unlikely(!ncsw
))
2029 * Was it really running after all now that we
2030 * checked with the proper locks actually held?
2032 * Oops. Go back and try again..
2034 if (unlikely(running
)) {
2040 * It's not enough that it's not actively running,
2041 * it must be off the runqueue _entirely_, and not
2044 * So if it wa still runnable (but just not actively
2045 * running right now), it's preempted, and we should
2046 * yield - it could be a while.
2048 if (unlikely(on_rq
)) {
2049 schedule_timeout_uninterruptible(1);
2054 * Ahh, all good. It wasn't running, and it wasn't
2055 * runnable, which means that it will never become
2056 * running in the future either. We're all done!
2065 * kick_process - kick a running thread to enter/exit the kernel
2066 * @p: the to-be-kicked thread
2068 * Cause a process which is running on another CPU to enter
2069 * kernel-mode, without any delay. (to get signals handled.)
2071 * NOTE: this function doesnt have to take the runqueue lock,
2072 * because all it wants to ensure is that the remote task enters
2073 * the kernel. If the IPI races and the task has been migrated
2074 * to another CPU then no harm is done and the purpose has been
2077 void kick_process(struct task_struct
*p
)
2083 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2084 smp_send_reschedule(cpu
);
2089 * Return a low guess at the load of a migration-source cpu weighted
2090 * according to the scheduling class and "nice" value.
2092 * We want to under-estimate the load of migration sources, to
2093 * balance conservatively.
2095 static unsigned long source_load(int cpu
, int type
)
2097 struct rq
*rq
= cpu_rq(cpu
);
2098 unsigned long total
= weighted_cpuload(cpu
);
2100 if (type
== 0 || !sched_feat(LB_BIAS
))
2103 return min(rq
->cpu_load
[type
-1], total
);
2107 * Return a high guess at the load of a migration-target cpu weighted
2108 * according to the scheduling class and "nice" value.
2110 static unsigned long target_load(int cpu
, int type
)
2112 struct rq
*rq
= cpu_rq(cpu
);
2113 unsigned long total
= weighted_cpuload(cpu
);
2115 if (type
== 0 || !sched_feat(LB_BIAS
))
2118 return max(rq
->cpu_load
[type
-1], total
);
2122 * find_idlest_group finds and returns the least busy CPU group within the
2125 static struct sched_group
*
2126 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2128 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2129 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2130 int load_idx
= sd
->forkexec_idx
;
2131 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2134 unsigned long load
, avg_load
;
2138 /* Skip over this group if it has no CPUs allowed */
2139 if (!cpumask_intersects(sched_group_cpus(group
),
2143 local_group
= cpumask_test_cpu(this_cpu
,
2144 sched_group_cpus(group
));
2146 /* Tally up the load of all CPUs in the group */
2149 for_each_cpu(i
, sched_group_cpus(group
)) {
2150 /* Bias balancing toward cpus of our domain */
2152 load
= source_load(i
, load_idx
);
2154 load
= target_load(i
, load_idx
);
2159 /* Adjust by relative CPU power of the group */
2160 avg_load
= sg_div_cpu_power(group
,
2161 avg_load
* SCHED_LOAD_SCALE
);
2164 this_load
= avg_load
;
2166 } else if (avg_load
< min_load
) {
2167 min_load
= avg_load
;
2170 } while (group
= group
->next
, group
!= sd
->groups
);
2172 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2178 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2181 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
2183 unsigned long load
, min_load
= ULONG_MAX
;
2187 /* Traverse only the allowed CPUs */
2188 for_each_cpu_and(i
, sched_group_cpus(group
), &p
->cpus_allowed
) {
2189 load
= weighted_cpuload(i
);
2191 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2201 * sched_balance_self: balance the current task (running on cpu) in domains
2202 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2205 * Balance, ie. select the least loaded group.
2207 * Returns the target CPU number, or the same CPU if no balancing is needed.
2209 * preempt must be disabled.
2211 static int sched_balance_self(int cpu
, int flag
)
2213 struct task_struct
*t
= current
;
2214 struct sched_domain
*tmp
, *sd
= NULL
;
2216 for_each_domain(cpu
, tmp
) {
2218 * If power savings logic is enabled for a domain, stop there.
2220 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2222 if (tmp
->flags
& flag
)
2230 struct sched_group
*group
;
2231 int new_cpu
, weight
;
2233 if (!(sd
->flags
& flag
)) {
2238 group
= find_idlest_group(sd
, t
, cpu
);
2244 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
2245 if (new_cpu
== -1 || new_cpu
== cpu
) {
2246 /* Now try balancing at a lower domain level of cpu */
2251 /* Now try balancing at a lower domain level of new_cpu */
2253 weight
= cpumask_weight(sched_domain_span(sd
));
2255 for_each_domain(cpu
, tmp
) {
2256 if (weight
<= cpumask_weight(sched_domain_span(tmp
)))
2258 if (tmp
->flags
& flag
)
2261 /* while loop will break here if sd == NULL */
2267 #endif /* CONFIG_SMP */
2270 * task_oncpu_function_call - call a function on the cpu on which a task runs
2271 * @p: the task to evaluate
2272 * @func: the function to be called
2273 * @info: the function call argument
2275 * Calls the function @func when the task is currently running. This might
2276 * be on the current CPU, which just calls the function directly
2278 void task_oncpu_function_call(struct task_struct
*p
,
2279 void (*func
) (void *info
), void *info
)
2286 smp_call_function_single(cpu
, func
, info
, 1);
2291 * try_to_wake_up - wake up a thread
2292 * @p: the to-be-woken-up thread
2293 * @state: the mask of task states that can be woken
2294 * @sync: do a synchronous wakeup?
2296 * Put it on the run-queue if it's not already there. The "current"
2297 * thread is always on the run-queue (except when the actual
2298 * re-schedule is in progress), and as such you're allowed to do
2299 * the simpler "current->state = TASK_RUNNING" to mark yourself
2300 * runnable without the overhead of this.
2302 * returns failure only if the task is already active.
2304 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2306 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2307 unsigned long flags
;
2311 if (!sched_feat(SYNC_WAKEUPS
))
2315 if (sched_feat(LB_WAKEUP_UPDATE
)) {
2316 struct sched_domain
*sd
;
2318 this_cpu
= raw_smp_processor_id();
2321 for_each_domain(this_cpu
, sd
) {
2322 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2331 rq
= task_rq_lock(p
, &flags
);
2332 update_rq_clock(rq
);
2333 old_state
= p
->state
;
2334 if (!(old_state
& state
))
2342 this_cpu
= smp_processor_id();
2345 if (unlikely(task_running(rq
, p
)))
2348 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2349 if (cpu
!= orig_cpu
) {
2350 set_task_cpu(p
, cpu
);
2351 task_rq_unlock(rq
, &flags
);
2352 /* might preempt at this point */
2353 rq
= task_rq_lock(p
, &flags
);
2354 old_state
= p
->state
;
2355 if (!(old_state
& state
))
2360 this_cpu
= smp_processor_id();
2364 #ifdef CONFIG_SCHEDSTATS
2365 schedstat_inc(rq
, ttwu_count
);
2366 if (cpu
== this_cpu
)
2367 schedstat_inc(rq
, ttwu_local
);
2369 struct sched_domain
*sd
;
2370 for_each_domain(this_cpu
, sd
) {
2371 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2372 schedstat_inc(sd
, ttwu_wake_remote
);
2377 #endif /* CONFIG_SCHEDSTATS */
2380 #endif /* CONFIG_SMP */
2381 schedstat_inc(p
, se
.nr_wakeups
);
2383 schedstat_inc(p
, se
.nr_wakeups_sync
);
2384 if (orig_cpu
!= cpu
)
2385 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2386 if (cpu
== this_cpu
)
2387 schedstat_inc(p
, se
.nr_wakeups_local
);
2389 schedstat_inc(p
, se
.nr_wakeups_remote
);
2390 activate_task(rq
, p
, 1);
2394 trace_sched_wakeup(rq
, p
, success
);
2395 check_preempt_curr(rq
, p
, sync
);
2397 p
->state
= TASK_RUNNING
;
2399 if (p
->sched_class
->task_wake_up
)
2400 p
->sched_class
->task_wake_up(rq
, p
);
2403 current
->se
.last_wakeup
= current
->se
.sum_exec_runtime
;
2405 task_rq_unlock(rq
, &flags
);
2410 int wake_up_process(struct task_struct
*p
)
2412 return try_to_wake_up(p
, TASK_ALL
, 0);
2414 EXPORT_SYMBOL(wake_up_process
);
2416 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2418 return try_to_wake_up(p
, state
, 0);
2422 * Perform scheduler related setup for a newly forked process p.
2423 * p is forked by current.
2425 * __sched_fork() is basic setup used by init_idle() too:
2427 static void __sched_fork(struct task_struct
*p
)
2429 p
->se
.exec_start
= 0;
2430 p
->se
.sum_exec_runtime
= 0;
2431 p
->se
.prev_sum_exec_runtime
= 0;
2432 p
->se
.nr_migrations
= 0;
2433 p
->se
.last_wakeup
= 0;
2434 p
->se
.avg_overlap
= 0;
2436 #ifdef CONFIG_SCHEDSTATS
2437 p
->se
.wait_start
= 0;
2438 p
->se
.sum_sleep_runtime
= 0;
2439 p
->se
.sleep_start
= 0;
2440 p
->se
.block_start
= 0;
2441 p
->se
.sleep_max
= 0;
2442 p
->se
.block_max
= 0;
2444 p
->se
.slice_max
= 0;
2448 INIT_LIST_HEAD(&p
->rt
.run_list
);
2450 INIT_LIST_HEAD(&p
->se
.group_node
);
2452 #ifdef CONFIG_PREEMPT_NOTIFIERS
2453 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2457 * We mark the process as running here, but have not actually
2458 * inserted it onto the runqueue yet. This guarantees that
2459 * nobody will actually run it, and a signal or other external
2460 * event cannot wake it up and insert it on the runqueue either.
2462 p
->state
= TASK_RUNNING
;
2466 * fork()/clone()-time setup:
2468 void sched_fork(struct task_struct
*p
, int clone_flags
)
2470 int cpu
= get_cpu();
2475 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2477 set_task_cpu(p
, cpu
);
2480 * Make sure we do not leak PI boosting priority to the child:
2482 p
->prio
= current
->normal_prio
;
2483 if (!rt_prio(p
->prio
))
2484 p
->sched_class
= &fair_sched_class
;
2486 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2487 if (likely(sched_info_on()))
2488 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2490 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2493 #ifdef CONFIG_PREEMPT
2494 /* Want to start with kernel preemption disabled. */
2495 task_thread_info(p
)->preempt_count
= 1;
2501 * wake_up_new_task - wake up a newly created task for the first time.
2503 * This function will do some initial scheduler statistics housekeeping
2504 * that must be done for every newly created context, then puts the task
2505 * on the runqueue and wakes it.
2507 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2509 unsigned long flags
;
2512 rq
= task_rq_lock(p
, &flags
);
2513 BUG_ON(p
->state
!= TASK_RUNNING
);
2514 update_rq_clock(rq
);
2516 p
->prio
= effective_prio(p
);
2518 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2519 activate_task(rq
, p
, 0);
2522 * Let the scheduling class do new task startup
2523 * management (if any):
2525 p
->sched_class
->task_new(rq
, p
);
2528 trace_sched_wakeup_new(rq
, p
, 1);
2529 check_preempt_curr(rq
, p
, 0);
2531 if (p
->sched_class
->task_wake_up
)
2532 p
->sched_class
->task_wake_up(rq
, p
);
2534 task_rq_unlock(rq
, &flags
);
2537 #ifdef CONFIG_PREEMPT_NOTIFIERS
2540 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2541 * @notifier: notifier struct to register
2543 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2545 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2547 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2550 * preempt_notifier_unregister - no longer interested in preemption notifications
2551 * @notifier: notifier struct to unregister
2553 * This is safe to call from within a preemption notifier.
2555 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2557 hlist_del(¬ifier
->link
);
2559 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2561 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2563 struct preempt_notifier
*notifier
;
2564 struct hlist_node
*node
;
2566 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2567 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2571 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2572 struct task_struct
*next
)
2574 struct preempt_notifier
*notifier
;
2575 struct hlist_node
*node
;
2577 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2578 notifier
->ops
->sched_out(notifier
, next
);
2581 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2583 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2588 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2589 struct task_struct
*next
)
2593 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2596 * prepare_task_switch - prepare to switch tasks
2597 * @rq: the runqueue preparing to switch
2598 * @prev: the current task that is being switched out
2599 * @next: the task we are going to switch to.
2601 * This is called with the rq lock held and interrupts off. It must
2602 * be paired with a subsequent finish_task_switch after the context
2605 * prepare_task_switch sets up locking and calls architecture specific
2609 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2610 struct task_struct
*next
)
2612 fire_sched_out_preempt_notifiers(prev
, next
);
2613 prepare_lock_switch(rq
, next
);
2614 prepare_arch_switch(next
);
2618 * finish_task_switch - clean up after a task-switch
2619 * @rq: runqueue associated with task-switch
2620 * @prev: the thread we just switched away from.
2622 * finish_task_switch must be called after the context switch, paired
2623 * with a prepare_task_switch call before the context switch.
2624 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2625 * and do any other architecture-specific cleanup actions.
2627 * Note that we may have delayed dropping an mm in context_switch(). If
2628 * so, we finish that here outside of the runqueue lock. (Doing it
2629 * with the lock held can cause deadlocks; see schedule() for
2632 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2633 __releases(rq
->lock
)
2635 struct mm_struct
*mm
= rq
->prev_mm
;
2641 * A task struct has one reference for the use as "current".
2642 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2643 * schedule one last time. The schedule call will never return, and
2644 * the scheduled task must drop that reference.
2645 * The test for TASK_DEAD must occur while the runqueue locks are
2646 * still held, otherwise prev could be scheduled on another cpu, die
2647 * there before we look at prev->state, and then the reference would
2649 * Manfred Spraul <manfred@colorfullife.com>
2651 prev_state
= prev
->state
;
2652 finish_arch_switch(prev
);
2653 perf_counter_task_sched_in(current
, cpu_of(rq
));
2654 finish_lock_switch(rq
, prev
);
2656 if (current
->sched_class
->post_schedule
)
2657 current
->sched_class
->post_schedule(rq
);
2660 fire_sched_in_preempt_notifiers(current
);
2663 if (unlikely(prev_state
== TASK_DEAD
)) {
2665 * Remove function-return probe instances associated with this
2666 * task and put them back on the free list.
2668 kprobe_flush_task(prev
);
2669 put_task_struct(prev
);
2674 * schedule_tail - first thing a freshly forked thread must call.
2675 * @prev: the thread we just switched away from.
2677 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2678 __releases(rq
->lock
)
2680 struct rq
*rq
= this_rq();
2682 finish_task_switch(rq
, prev
);
2683 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2684 /* In this case, finish_task_switch does not reenable preemption */
2687 if (current
->set_child_tid
)
2688 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2692 * context_switch - switch to the new MM and the new
2693 * thread's register state.
2696 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2697 struct task_struct
*next
)
2699 struct mm_struct
*mm
, *oldmm
;
2701 prepare_task_switch(rq
, prev
, next
);
2702 trace_sched_switch(rq
, prev
, next
);
2704 oldmm
= prev
->active_mm
;
2706 * For paravirt, this is coupled with an exit in switch_to to
2707 * combine the page table reload and the switch backend into
2710 arch_enter_lazy_cpu_mode();
2712 if (unlikely(!mm
)) {
2713 next
->active_mm
= oldmm
;
2714 atomic_inc(&oldmm
->mm_count
);
2715 enter_lazy_tlb(oldmm
, next
);
2717 switch_mm(oldmm
, mm
, next
);
2719 if (unlikely(!prev
->mm
)) {
2720 prev
->active_mm
= NULL
;
2721 rq
->prev_mm
= oldmm
;
2724 * Since the runqueue lock will be released by the next
2725 * task (which is an invalid locking op but in the case
2726 * of the scheduler it's an obvious special-case), so we
2727 * do an early lockdep release here:
2729 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2730 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2733 /* Here we just switch the register state and the stack. */
2734 switch_to(prev
, next
, prev
);
2738 * this_rq must be evaluated again because prev may have moved
2739 * CPUs since it called schedule(), thus the 'rq' on its stack
2740 * frame will be invalid.
2742 finish_task_switch(this_rq(), prev
);
2746 * nr_running, nr_uninterruptible and nr_context_switches:
2748 * externally visible scheduler statistics: current number of runnable
2749 * threads, current number of uninterruptible-sleeping threads, total
2750 * number of context switches performed since bootup.
2752 unsigned long nr_running(void)
2754 unsigned long i
, sum
= 0;
2756 for_each_online_cpu(i
)
2757 sum
+= cpu_rq(i
)->nr_running
;
2762 unsigned long nr_uninterruptible(void)
2764 unsigned long i
, sum
= 0;
2766 for_each_possible_cpu(i
)
2767 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2770 * Since we read the counters lockless, it might be slightly
2771 * inaccurate. Do not allow it to go below zero though:
2773 if (unlikely((long)sum
< 0))
2779 unsigned long long nr_context_switches(void)
2782 unsigned long long sum
= 0;
2784 for_each_possible_cpu(i
)
2785 sum
+= cpu_rq(i
)->nr_switches
;
2790 unsigned long nr_iowait(void)
2792 unsigned long i
, sum
= 0;
2794 for_each_possible_cpu(i
)
2795 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2800 unsigned long nr_active(void)
2802 unsigned long i
, running
= 0, uninterruptible
= 0;
2804 for_each_online_cpu(i
) {
2805 running
+= cpu_rq(i
)->nr_running
;
2806 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2809 if (unlikely((long)uninterruptible
< 0))
2810 uninterruptible
= 0;
2812 return running
+ uninterruptible
;
2816 * Externally visible per-cpu scheduler statistics:
2817 * cpu_nr_switches(cpu) - number of context switches on that cpu
2818 * cpu_nr_migrations(cpu) - number of migrations into that cpu
2820 u64
cpu_nr_switches(int cpu
)
2822 return cpu_rq(cpu
)->nr_switches
;
2825 u64
cpu_nr_migrations(int cpu
)
2827 return cpu_rq(cpu
)->nr_migrations_in
;
2831 * Update rq->cpu_load[] statistics. This function is usually called every
2832 * scheduler tick (TICK_NSEC).
2834 static void update_cpu_load(struct rq
*this_rq
)
2836 unsigned long this_load
= this_rq
->load
.weight
;
2839 this_rq
->nr_load_updates
++;
2841 /* Update our load: */
2842 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2843 unsigned long old_load
, new_load
;
2845 /* scale is effectively 1 << i now, and >> i divides by scale */
2847 old_load
= this_rq
->cpu_load
[i
];
2848 new_load
= this_load
;
2850 * Round up the averaging division if load is increasing. This
2851 * prevents us from getting stuck on 9 if the load is 10, for
2854 if (new_load
> old_load
)
2855 new_load
+= scale
-1;
2856 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2863 * double_rq_lock - safely lock two runqueues
2865 * Note this does not disable interrupts like task_rq_lock,
2866 * you need to do so manually before calling.
2868 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2869 __acquires(rq1
->lock
)
2870 __acquires(rq2
->lock
)
2872 BUG_ON(!irqs_disabled());
2874 spin_lock(&rq1
->lock
);
2875 __acquire(rq2
->lock
); /* Fake it out ;) */
2878 spin_lock(&rq1
->lock
);
2879 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
2881 spin_lock(&rq2
->lock
);
2882 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
2885 update_rq_clock(rq1
);
2886 update_rq_clock(rq2
);
2890 * double_rq_unlock - safely unlock two runqueues
2892 * Note this does not restore interrupts like task_rq_unlock,
2893 * you need to do so manually after calling.
2895 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2896 __releases(rq1
->lock
)
2897 __releases(rq2
->lock
)
2899 spin_unlock(&rq1
->lock
);
2901 spin_unlock(&rq2
->lock
);
2903 __release(rq2
->lock
);
2907 * If dest_cpu is allowed for this process, migrate the task to it.
2908 * This is accomplished by forcing the cpu_allowed mask to only
2909 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2910 * the cpu_allowed mask is restored.
2912 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2914 struct migration_req req
;
2915 unsigned long flags
;
2918 rq
= task_rq_lock(p
, &flags
);
2919 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
2920 || unlikely(!cpu_active(dest_cpu
)))
2923 /* force the process onto the specified CPU */
2924 if (migrate_task(p
, dest_cpu
, &req
)) {
2925 /* Need to wait for migration thread (might exit: take ref). */
2926 struct task_struct
*mt
= rq
->migration_thread
;
2928 get_task_struct(mt
);
2929 task_rq_unlock(rq
, &flags
);
2930 wake_up_process(mt
);
2931 put_task_struct(mt
);
2932 wait_for_completion(&req
.done
);
2937 task_rq_unlock(rq
, &flags
);
2941 * sched_exec - execve() is a valuable balancing opportunity, because at
2942 * this point the task has the smallest effective memory and cache footprint.
2944 void sched_exec(void)
2946 int new_cpu
, this_cpu
= get_cpu();
2947 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2949 if (new_cpu
!= this_cpu
)
2950 sched_migrate_task(current
, new_cpu
);
2954 * pull_task - move a task from a remote runqueue to the local runqueue.
2955 * Both runqueues must be locked.
2957 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2958 struct rq
*this_rq
, int this_cpu
)
2960 deactivate_task(src_rq
, p
, 0);
2961 set_task_cpu(p
, this_cpu
);
2962 activate_task(this_rq
, p
, 0);
2964 * Note that idle threads have a prio of MAX_PRIO, for this test
2965 * to be always true for them.
2967 check_preempt_curr(this_rq
, p
, 0);
2971 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2974 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2975 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2979 * We do not migrate tasks that are:
2980 * 1) running (obviously), or
2981 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2982 * 3) are cache-hot on their current CPU.
2984 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
2985 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2990 if (task_running(rq
, p
)) {
2991 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2996 * Aggressive migration if:
2997 * 1) task is cache cold, or
2998 * 2) too many balance attempts have failed.
3001 if (!task_hot(p
, rq
->clock
, sd
) ||
3002 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
3003 #ifdef CONFIG_SCHEDSTATS
3004 if (task_hot(p
, rq
->clock
, sd
)) {
3005 schedstat_inc(sd
, lb_hot_gained
[idle
]);
3006 schedstat_inc(p
, se
.nr_forced_migrations
);
3012 if (task_hot(p
, rq
->clock
, sd
)) {
3013 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
3019 static unsigned long
3020 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3021 unsigned long max_load_move
, struct sched_domain
*sd
,
3022 enum cpu_idle_type idle
, int *all_pinned
,
3023 int *this_best_prio
, struct rq_iterator
*iterator
)
3025 int loops
= 0, pulled
= 0, pinned
= 0;
3026 struct task_struct
*p
;
3027 long rem_load_move
= max_load_move
;
3029 if (max_load_move
== 0)
3035 * Start the load-balancing iterator:
3037 p
= iterator
->start(iterator
->arg
);
3039 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3042 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
3043 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3044 p
= iterator
->next(iterator
->arg
);
3048 pull_task(busiest
, p
, this_rq
, this_cpu
);
3050 rem_load_move
-= p
->se
.load
.weight
;
3053 * We only want to steal up to the prescribed amount of weighted load.
3055 if (rem_load_move
> 0) {
3056 if (p
->prio
< *this_best_prio
)
3057 *this_best_prio
= p
->prio
;
3058 p
= iterator
->next(iterator
->arg
);
3063 * Right now, this is one of only two places pull_task() is called,
3064 * so we can safely collect pull_task() stats here rather than
3065 * inside pull_task().
3067 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3070 *all_pinned
= pinned
;
3072 return max_load_move
- rem_load_move
;
3076 * move_tasks tries to move up to max_load_move weighted load from busiest to
3077 * this_rq, as part of a balancing operation within domain "sd".
3078 * Returns 1 if successful and 0 otherwise.
3080 * Called with both runqueues locked.
3082 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3083 unsigned long max_load_move
,
3084 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3087 const struct sched_class
*class = sched_class_highest
;
3088 unsigned long total_load_moved
= 0;
3089 int this_best_prio
= this_rq
->curr
->prio
;
3093 class->load_balance(this_rq
, this_cpu
, busiest
,
3094 max_load_move
- total_load_moved
,
3095 sd
, idle
, all_pinned
, &this_best_prio
);
3096 class = class->next
;
3098 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3101 } while (class && max_load_move
> total_load_moved
);
3103 return total_load_moved
> 0;
3107 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3108 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3109 struct rq_iterator
*iterator
)
3111 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3115 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3116 pull_task(busiest
, p
, this_rq
, this_cpu
);
3118 * Right now, this is only the second place pull_task()
3119 * is called, so we can safely collect pull_task()
3120 * stats here rather than inside pull_task().
3122 schedstat_inc(sd
, lb_gained
[idle
]);
3126 p
= iterator
->next(iterator
->arg
);
3133 * move_one_task tries to move exactly one task from busiest to this_rq, as
3134 * part of active balancing operations within "domain".
3135 * Returns 1 if successful and 0 otherwise.
3137 * Called with both runqueues locked.
3139 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3140 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3142 const struct sched_class
*class;
3144 for (class = sched_class_highest
; class; class = class->next
)
3145 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3152 * find_busiest_group finds and returns the busiest CPU group within the
3153 * domain. It calculates and returns the amount of weighted load which
3154 * should be moved to restore balance via the imbalance parameter.
3156 static struct sched_group
*
3157 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3158 unsigned long *imbalance
, enum cpu_idle_type idle
,
3159 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
3161 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
3162 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
3163 unsigned long max_pull
;
3164 unsigned long busiest_load_per_task
, busiest_nr_running
;
3165 unsigned long this_load_per_task
, this_nr_running
;
3166 int load_idx
, group_imb
= 0;
3167 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3168 int power_savings_balance
= 1;
3169 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
3170 unsigned long min_nr_running
= ULONG_MAX
;
3171 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
3174 max_load
= this_load
= total_load
= total_pwr
= 0;
3175 busiest_load_per_task
= busiest_nr_running
= 0;
3176 this_load_per_task
= this_nr_running
= 0;
3178 if (idle
== CPU_NOT_IDLE
)
3179 load_idx
= sd
->busy_idx
;
3180 else if (idle
== CPU_NEWLY_IDLE
)
3181 load_idx
= sd
->newidle_idx
;
3183 load_idx
= sd
->idle_idx
;
3186 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
3189 int __group_imb
= 0;
3190 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3191 unsigned long sum_nr_running
, sum_weighted_load
;
3192 unsigned long sum_avg_load_per_task
;
3193 unsigned long avg_load_per_task
;
3195 local_group
= cpumask_test_cpu(this_cpu
,
3196 sched_group_cpus(group
));
3199 balance_cpu
= cpumask_first(sched_group_cpus(group
));
3201 /* Tally up the load of all CPUs in the group */
3202 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
3203 sum_avg_load_per_task
= avg_load_per_task
= 0;
3206 min_cpu_load
= ~0UL;
3208 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3209 struct rq
*rq
= cpu_rq(i
);
3211 if (*sd_idle
&& rq
->nr_running
)
3214 /* Bias balancing toward cpus of our domain */
3216 if (idle_cpu(i
) && !first_idle_cpu
) {
3221 load
= target_load(i
, load_idx
);
3223 load
= source_load(i
, load_idx
);
3224 if (load
> max_cpu_load
)
3225 max_cpu_load
= load
;
3226 if (min_cpu_load
> load
)
3227 min_cpu_load
= load
;
3231 sum_nr_running
+= rq
->nr_running
;
3232 sum_weighted_load
+= weighted_cpuload(i
);
3234 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3238 * First idle cpu or the first cpu(busiest) in this sched group
3239 * is eligible for doing load balancing at this and above
3240 * domains. In the newly idle case, we will allow all the cpu's
3241 * to do the newly idle load balance.
3243 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3244 balance_cpu
!= this_cpu
&& balance
) {
3249 total_load
+= avg_load
;
3250 total_pwr
+= group
->__cpu_power
;
3252 /* Adjust by relative CPU power of the group */
3253 avg_load
= sg_div_cpu_power(group
,
3254 avg_load
* SCHED_LOAD_SCALE
);
3258 * Consider the group unbalanced when the imbalance is larger
3259 * than the average weight of two tasks.
3261 * APZ: with cgroup the avg task weight can vary wildly and
3262 * might not be a suitable number - should we keep a
3263 * normalized nr_running number somewhere that negates
3266 avg_load_per_task
= sg_div_cpu_power(group
,
3267 sum_avg_load_per_task
* SCHED_LOAD_SCALE
);
3269 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3272 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3275 this_load
= avg_load
;
3277 this_nr_running
= sum_nr_running
;
3278 this_load_per_task
= sum_weighted_load
;
3279 } else if (avg_load
> max_load
&&
3280 (sum_nr_running
> group_capacity
|| __group_imb
)) {
3281 max_load
= avg_load
;
3283 busiest_nr_running
= sum_nr_running
;
3284 busiest_load_per_task
= sum_weighted_load
;
3285 group_imb
= __group_imb
;
3288 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3290 * Busy processors will not participate in power savings
3293 if (idle
== CPU_NOT_IDLE
||
3294 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3298 * If the local group is idle or completely loaded
3299 * no need to do power savings balance at this domain
3301 if (local_group
&& (this_nr_running
>= group_capacity
||
3303 power_savings_balance
= 0;
3306 * If a group is already running at full capacity or idle,
3307 * don't include that group in power savings calculations
3309 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
3314 * Calculate the group which has the least non-idle load.
3315 * This is the group from where we need to pick up the load
3318 if ((sum_nr_running
< min_nr_running
) ||
3319 (sum_nr_running
== min_nr_running
&&
3320 cpumask_first(sched_group_cpus(group
)) >
3321 cpumask_first(sched_group_cpus(group_min
)))) {
3323 min_nr_running
= sum_nr_running
;
3324 min_load_per_task
= sum_weighted_load
/
3329 * Calculate the group which is almost near its
3330 * capacity but still has some space to pick up some load
3331 * from other group and save more power
3333 if (sum_nr_running
<= group_capacity
- 1) {
3334 if (sum_nr_running
> leader_nr_running
||
3335 (sum_nr_running
== leader_nr_running
&&
3336 cpumask_first(sched_group_cpus(group
)) <
3337 cpumask_first(sched_group_cpus(group_leader
)))) {
3338 group_leader
= group
;
3339 leader_nr_running
= sum_nr_running
;
3344 group
= group
->next
;
3345 } while (group
!= sd
->groups
);
3347 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
3350 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
3352 if (this_load
>= avg_load
||
3353 100*max_load
<= sd
->imbalance_pct
*this_load
)
3356 busiest_load_per_task
/= busiest_nr_running
;
3358 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
3361 * We're trying to get all the cpus to the average_load, so we don't
3362 * want to push ourselves above the average load, nor do we wish to
3363 * reduce the max loaded cpu below the average load, as either of these
3364 * actions would just result in more rebalancing later, and ping-pong
3365 * tasks around. Thus we look for the minimum possible imbalance.
3366 * Negative imbalances (*we* are more loaded than anyone else) will
3367 * be counted as no imbalance for these purposes -- we can't fix that
3368 * by pulling tasks to us. Be careful of negative numbers as they'll
3369 * appear as very large values with unsigned longs.
3371 if (max_load
<= busiest_load_per_task
)
3375 * In the presence of smp nice balancing, certain scenarios can have
3376 * max load less than avg load(as we skip the groups at or below
3377 * its cpu_power, while calculating max_load..)
3379 if (max_load
< avg_load
) {
3381 goto small_imbalance
;
3384 /* Don't want to pull so many tasks that a group would go idle */
3385 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
3387 /* How much load to actually move to equalise the imbalance */
3388 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
3389 (avg_load
- this_load
) * this->__cpu_power
)
3393 * if *imbalance is less than the average load per runnable task
3394 * there is no gaurantee that any tasks will be moved so we'll have
3395 * a think about bumping its value to force at least one task to be
3398 if (*imbalance
< busiest_load_per_task
) {
3399 unsigned long tmp
, pwr_now
, pwr_move
;
3403 pwr_move
= pwr_now
= 0;
3405 if (this_nr_running
) {
3406 this_load_per_task
/= this_nr_running
;
3407 if (busiest_load_per_task
> this_load_per_task
)
3410 this_load_per_task
= cpu_avg_load_per_task(this_cpu
);
3412 if (max_load
- this_load
+ busiest_load_per_task
>=
3413 busiest_load_per_task
* imbn
) {
3414 *imbalance
= busiest_load_per_task
;
3419 * OK, we don't have enough imbalance to justify moving tasks,
3420 * however we may be able to increase total CPU power used by
3424 pwr_now
+= busiest
->__cpu_power
*
3425 min(busiest_load_per_task
, max_load
);
3426 pwr_now
+= this->__cpu_power
*
3427 min(this_load_per_task
, this_load
);
3428 pwr_now
/= SCHED_LOAD_SCALE
;
3430 /* Amount of load we'd subtract */
3431 tmp
= sg_div_cpu_power(busiest
,
3432 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3434 pwr_move
+= busiest
->__cpu_power
*
3435 min(busiest_load_per_task
, max_load
- tmp
);
3437 /* Amount of load we'd add */
3438 if (max_load
* busiest
->__cpu_power
<
3439 busiest_load_per_task
* SCHED_LOAD_SCALE
)
3440 tmp
= sg_div_cpu_power(this,
3441 max_load
* busiest
->__cpu_power
);
3443 tmp
= sg_div_cpu_power(this,
3444 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3445 pwr_move
+= this->__cpu_power
*
3446 min(this_load_per_task
, this_load
+ tmp
);
3447 pwr_move
/= SCHED_LOAD_SCALE
;
3449 /* Move if we gain throughput */
3450 if (pwr_move
> pwr_now
)
3451 *imbalance
= busiest_load_per_task
;
3457 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3458 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3461 if (this == group_leader
&& group_leader
!= group_min
) {
3462 *imbalance
= min_load_per_task
;
3463 if (sched_mc_power_savings
>= POWERSAVINGS_BALANCE_WAKEUP
) {
3464 cpu_rq(this_cpu
)->rd
->sched_mc_preferred_wakeup_cpu
=
3465 cpumask_first(sched_group_cpus(group_leader
));
3476 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3479 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3480 unsigned long imbalance
, const struct cpumask
*cpus
)
3482 struct rq
*busiest
= NULL
, *rq
;
3483 unsigned long max_load
= 0;
3486 for_each_cpu(i
, sched_group_cpus(group
)) {
3489 if (!cpumask_test_cpu(i
, cpus
))
3493 wl
= weighted_cpuload(i
);
3495 if (rq
->nr_running
== 1 && wl
> imbalance
)
3498 if (wl
> max_load
) {
3508 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3509 * so long as it is large enough.
3511 #define MAX_PINNED_INTERVAL 512
3514 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3515 * tasks if there is an imbalance.
3517 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3518 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3519 int *balance
, struct cpumask
*cpus
)
3521 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3522 struct sched_group
*group
;
3523 unsigned long imbalance
;
3525 unsigned long flags
;
3527 cpumask_setall(cpus
);
3530 * When power savings policy is enabled for the parent domain, idle
3531 * sibling can pick up load irrespective of busy siblings. In this case,
3532 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3533 * portraying it as CPU_NOT_IDLE.
3535 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3536 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3539 schedstat_inc(sd
, lb_count
[idle
]);
3543 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3550 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3554 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3556 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3560 BUG_ON(busiest
== this_rq
);
3562 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3565 if (busiest
->nr_running
> 1) {
3567 * Attempt to move tasks. If find_busiest_group has found
3568 * an imbalance but busiest->nr_running <= 1, the group is
3569 * still unbalanced. ld_moved simply stays zero, so it is
3570 * correctly treated as an imbalance.
3572 local_irq_save(flags
);
3573 double_rq_lock(this_rq
, busiest
);
3574 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3575 imbalance
, sd
, idle
, &all_pinned
);
3576 double_rq_unlock(this_rq
, busiest
);
3577 local_irq_restore(flags
);
3580 * some other cpu did the load balance for us.
3582 if (ld_moved
&& this_cpu
!= smp_processor_id())
3583 resched_cpu(this_cpu
);
3585 /* All tasks on this runqueue were pinned by CPU affinity */
3586 if (unlikely(all_pinned
)) {
3587 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
3588 if (!cpumask_empty(cpus
))
3595 schedstat_inc(sd
, lb_failed
[idle
]);
3596 sd
->nr_balance_failed
++;
3598 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3600 spin_lock_irqsave(&busiest
->lock
, flags
);
3602 /* don't kick the migration_thread, if the curr
3603 * task on busiest cpu can't be moved to this_cpu
3605 if (!cpumask_test_cpu(this_cpu
,
3606 &busiest
->curr
->cpus_allowed
)) {
3607 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3609 goto out_one_pinned
;
3612 if (!busiest
->active_balance
) {
3613 busiest
->active_balance
= 1;
3614 busiest
->push_cpu
= this_cpu
;
3617 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3619 wake_up_process(busiest
->migration_thread
);
3622 * We've kicked active balancing, reset the failure
3625 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3628 sd
->nr_balance_failed
= 0;
3630 if (likely(!active_balance
)) {
3631 /* We were unbalanced, so reset the balancing interval */
3632 sd
->balance_interval
= sd
->min_interval
;
3635 * If we've begun active balancing, start to back off. This
3636 * case may not be covered by the all_pinned logic if there
3637 * is only 1 task on the busy runqueue (because we don't call
3640 if (sd
->balance_interval
< sd
->max_interval
)
3641 sd
->balance_interval
*= 2;
3644 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3645 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3651 schedstat_inc(sd
, lb_balanced
[idle
]);
3653 sd
->nr_balance_failed
= 0;
3656 /* tune up the balancing interval */
3657 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3658 (sd
->balance_interval
< sd
->max_interval
))
3659 sd
->balance_interval
*= 2;
3661 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3662 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3673 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3674 * tasks if there is an imbalance.
3676 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3677 * this_rq is locked.
3680 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
,
3681 struct cpumask
*cpus
)
3683 struct sched_group
*group
;
3684 struct rq
*busiest
= NULL
;
3685 unsigned long imbalance
;
3690 cpumask_setall(cpus
);
3693 * When power savings policy is enabled for the parent domain, idle
3694 * sibling can pick up load irrespective of busy siblings. In this case,
3695 * let the state of idle sibling percolate up as IDLE, instead of
3696 * portraying it as CPU_NOT_IDLE.
3698 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3699 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3702 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3704 update_shares_locked(this_rq
, sd
);
3705 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3706 &sd_idle
, cpus
, NULL
);
3708 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3712 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
3714 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3718 BUG_ON(busiest
== this_rq
);
3720 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3723 if (busiest
->nr_running
> 1) {
3724 /* Attempt to move tasks */
3725 double_lock_balance(this_rq
, busiest
);
3726 /* this_rq->clock is already updated */
3727 update_rq_clock(busiest
);
3728 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3729 imbalance
, sd
, CPU_NEWLY_IDLE
,
3731 double_unlock_balance(this_rq
, busiest
);
3733 if (unlikely(all_pinned
)) {
3734 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
3735 if (!cpumask_empty(cpus
))
3741 int active_balance
= 0;
3743 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3744 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3745 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3748 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
3751 if (sd
->nr_balance_failed
++ < 2)
3755 * The only task running in a non-idle cpu can be moved to this
3756 * cpu in an attempt to completely freeup the other CPU
3757 * package. The same method used to move task in load_balance()
3758 * have been extended for load_balance_newidle() to speedup
3759 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
3761 * The package power saving logic comes from
3762 * find_busiest_group(). If there are no imbalance, then
3763 * f_b_g() will return NULL. However when sched_mc={1,2} then
3764 * f_b_g() will select a group from which a running task may be
3765 * pulled to this cpu in order to make the other package idle.
3766 * If there is no opportunity to make a package idle and if
3767 * there are no imbalance, then f_b_g() will return NULL and no
3768 * action will be taken in load_balance_newidle().
3770 * Under normal task pull operation due to imbalance, there
3771 * will be more than one task in the source run queue and
3772 * move_tasks() will succeed. ld_moved will be true and this
3773 * active balance code will not be triggered.
3776 /* Lock busiest in correct order while this_rq is held */
3777 double_lock_balance(this_rq
, busiest
);
3780 * don't kick the migration_thread, if the curr
3781 * task on busiest cpu can't be moved to this_cpu
3783 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
3784 double_unlock_balance(this_rq
, busiest
);
3789 if (!busiest
->active_balance
) {
3790 busiest
->active_balance
= 1;
3791 busiest
->push_cpu
= this_cpu
;
3795 double_unlock_balance(this_rq
, busiest
);
3797 * Should not call ttwu while holding a rq->lock
3799 spin_unlock(&this_rq
->lock
);
3801 wake_up_process(busiest
->migration_thread
);
3802 spin_lock(&this_rq
->lock
);
3805 sd
->nr_balance_failed
= 0;
3807 update_shares_locked(this_rq
, sd
);
3811 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3812 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3813 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3815 sd
->nr_balance_failed
= 0;
3821 * idle_balance is called by schedule() if this_cpu is about to become
3822 * idle. Attempts to pull tasks from other CPUs.
3824 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3826 struct sched_domain
*sd
;
3827 int pulled_task
= 0;
3828 unsigned long next_balance
= jiffies
+ HZ
;
3829 cpumask_var_t tmpmask
;
3831 if (!alloc_cpumask_var(&tmpmask
, GFP_ATOMIC
))
3834 for_each_domain(this_cpu
, sd
) {
3835 unsigned long interval
;
3837 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3840 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3841 /* If we've pulled tasks over stop searching: */
3842 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
3845 interval
= msecs_to_jiffies(sd
->balance_interval
);
3846 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3847 next_balance
= sd
->last_balance
+ interval
;
3851 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3853 * We are going idle. next_balance may be set based on
3854 * a busy processor. So reset next_balance.
3856 this_rq
->next_balance
= next_balance
;
3858 free_cpumask_var(tmpmask
);
3862 * active_load_balance is run by migration threads. It pushes running tasks
3863 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3864 * running on each physical CPU where possible, and avoids physical /
3865 * logical imbalances.
3867 * Called with busiest_rq locked.
3869 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3871 int target_cpu
= busiest_rq
->push_cpu
;
3872 struct sched_domain
*sd
;
3873 struct rq
*target_rq
;
3875 /* Is there any task to move? */
3876 if (busiest_rq
->nr_running
<= 1)
3879 target_rq
= cpu_rq(target_cpu
);
3882 * This condition is "impossible", if it occurs
3883 * we need to fix it. Originally reported by
3884 * Bjorn Helgaas on a 128-cpu setup.
3886 BUG_ON(busiest_rq
== target_rq
);
3888 /* move a task from busiest_rq to target_rq */
3889 double_lock_balance(busiest_rq
, target_rq
);
3890 update_rq_clock(busiest_rq
);
3891 update_rq_clock(target_rq
);
3893 /* Search for an sd spanning us and the target CPU. */
3894 for_each_domain(target_cpu
, sd
) {
3895 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3896 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
3901 schedstat_inc(sd
, alb_count
);
3903 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3905 schedstat_inc(sd
, alb_pushed
);
3907 schedstat_inc(sd
, alb_failed
);
3909 double_unlock_balance(busiest_rq
, target_rq
);
3914 atomic_t load_balancer
;
3915 cpumask_var_t cpu_mask
;
3916 } nohz ____cacheline_aligned
= {
3917 .load_balancer
= ATOMIC_INIT(-1),
3921 * This routine will try to nominate the ilb (idle load balancing)
3922 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3923 * load balancing on behalf of all those cpus. If all the cpus in the system
3924 * go into this tickless mode, then there will be no ilb owner (as there is
3925 * no need for one) and all the cpus will sleep till the next wakeup event
3928 * For the ilb owner, tick is not stopped. And this tick will be used
3929 * for idle load balancing. ilb owner will still be part of
3932 * While stopping the tick, this cpu will become the ilb owner if there
3933 * is no other owner. And will be the owner till that cpu becomes busy
3934 * or if all cpus in the system stop their ticks at which point
3935 * there is no need for ilb owner.
3937 * When the ilb owner becomes busy, it nominates another owner, during the
3938 * next busy scheduler_tick()
3940 int select_nohz_load_balancer(int stop_tick
)
3942 int cpu
= smp_processor_id();
3945 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
3946 cpu_rq(cpu
)->in_nohz_recently
= 1;
3949 * If we are going offline and still the leader, give up!
3951 if (!cpu_active(cpu
) &&
3952 atomic_read(&nohz
.load_balancer
) == cpu
) {
3953 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3958 /* time for ilb owner also to sleep */
3959 if (cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3960 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3961 atomic_set(&nohz
.load_balancer
, -1);
3965 if (atomic_read(&nohz
.load_balancer
) == -1) {
3966 /* make me the ilb owner */
3967 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3969 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3972 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
3975 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
3977 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3978 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3985 static DEFINE_SPINLOCK(balancing
);
3988 * It checks each scheduling domain to see if it is due to be balanced,
3989 * and initiates a balancing operation if so.
3991 * Balancing parameters are set up in arch_init_sched_domains.
3993 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3996 struct rq
*rq
= cpu_rq(cpu
);
3997 unsigned long interval
;
3998 struct sched_domain
*sd
;
3999 /* Earliest time when we have to do rebalance again */
4000 unsigned long next_balance
= jiffies
+ 60*HZ
;
4001 int update_next_balance
= 0;
4005 /* Fails alloc? Rebalancing probably not a priority right now. */
4006 if (!alloc_cpumask_var(&tmp
, GFP_ATOMIC
))
4009 for_each_domain(cpu
, sd
) {
4010 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4013 interval
= sd
->balance_interval
;
4014 if (idle
!= CPU_IDLE
)
4015 interval
*= sd
->busy_factor
;
4017 /* scale ms to jiffies */
4018 interval
= msecs_to_jiffies(interval
);
4019 if (unlikely(!interval
))
4021 if (interval
> HZ
*NR_CPUS
/10)
4022 interval
= HZ
*NR_CPUS
/10;
4024 need_serialize
= sd
->flags
& SD_SERIALIZE
;
4026 if (need_serialize
) {
4027 if (!spin_trylock(&balancing
))
4031 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4032 if (load_balance(cpu
, rq
, sd
, idle
, &balance
, tmp
)) {
4034 * We've pulled tasks over so either we're no
4035 * longer idle, or one of our SMT siblings is
4038 idle
= CPU_NOT_IDLE
;
4040 sd
->last_balance
= jiffies
;
4043 spin_unlock(&balancing
);
4045 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4046 next_balance
= sd
->last_balance
+ interval
;
4047 update_next_balance
= 1;
4051 * Stop the load balance at this level. There is another
4052 * CPU in our sched group which is doing load balancing more
4060 * next_balance will be updated only when there is a need.
4061 * When the cpu is attached to null domain for ex, it will not be
4064 if (likely(update_next_balance
))
4065 rq
->next_balance
= next_balance
;
4067 free_cpumask_var(tmp
);
4071 * run_rebalance_domains is triggered when needed from the scheduler tick.
4072 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4073 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4075 static void run_rebalance_domains(struct softirq_action
*h
)
4077 int this_cpu
= smp_processor_id();
4078 struct rq
*this_rq
= cpu_rq(this_cpu
);
4079 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4080 CPU_IDLE
: CPU_NOT_IDLE
;
4082 rebalance_domains(this_cpu
, idle
);
4086 * If this cpu is the owner for idle load balancing, then do the
4087 * balancing on behalf of the other idle cpus whose ticks are
4090 if (this_rq
->idle_at_tick
&&
4091 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4095 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
4096 if (balance_cpu
== this_cpu
)
4100 * If this cpu gets work to do, stop the load balancing
4101 * work being done for other cpus. Next load
4102 * balancing owner will pick it up.
4107 rebalance_domains(balance_cpu
, CPU_IDLE
);
4109 rq
= cpu_rq(balance_cpu
);
4110 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4111 this_rq
->next_balance
= rq
->next_balance
;
4118 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4120 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4121 * idle load balancing owner or decide to stop the periodic load balancing,
4122 * if the whole system is idle.
4124 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4128 * If we were in the nohz mode recently and busy at the current
4129 * scheduler tick, then check if we need to nominate new idle
4132 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4133 rq
->in_nohz_recently
= 0;
4135 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4136 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4137 atomic_set(&nohz
.load_balancer
, -1);
4140 if (atomic_read(&nohz
.load_balancer
) == -1) {
4142 * simple selection for now: Nominate the
4143 * first cpu in the nohz list to be the next
4146 * TBD: Traverse the sched domains and nominate
4147 * the nearest cpu in the nohz.cpu_mask.
4149 int ilb
= cpumask_first(nohz
.cpu_mask
);
4151 if (ilb
< nr_cpu_ids
)
4157 * If this cpu is idle and doing idle load balancing for all the
4158 * cpus with ticks stopped, is it time for that to stop?
4160 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4161 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4167 * If this cpu is idle and the idle load balancing is done by
4168 * someone else, then no need raise the SCHED_SOFTIRQ
4170 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4171 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4174 if (time_after_eq(jiffies
, rq
->next_balance
))
4175 raise_softirq(SCHED_SOFTIRQ
);
4178 #else /* CONFIG_SMP */
4181 * on UP we do not need to balance between CPUs:
4183 static inline void idle_balance(int cpu
, struct rq
*rq
)
4189 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4191 EXPORT_PER_CPU_SYMBOL(kstat
);
4194 * Return any ns on the sched_clock that have not yet been banked in
4195 * @p in case that task is currently running.
4197 unsigned long long __task_delta_exec(struct task_struct
*p
, int update
)
4203 WARN_ON_ONCE(!runqueue_is_locked());
4204 WARN_ON_ONCE(!task_current(rq
, p
));
4207 update_rq_clock(rq
);
4209 delta_exec
= rq
->clock
- p
->se
.exec_start
;
4211 WARN_ON_ONCE(delta_exec
< 0);
4217 * Return any ns on the sched_clock that have not yet been banked in
4218 * @p in case that task is currently running.
4220 unsigned long long task_delta_exec(struct task_struct
*p
)
4222 unsigned long flags
;
4226 rq
= task_rq_lock(p
, &flags
);
4228 if (task_current(rq
, p
)) {
4231 update_rq_clock(rq
);
4232 delta_exec
= rq
->clock
- p
->se
.exec_start
;
4233 if ((s64
)delta_exec
> 0)
4237 task_rq_unlock(rq
, &flags
);
4243 * Account user cpu time to a process.
4244 * @p: the process that the cpu time gets accounted to
4245 * @cputime: the cpu time spent in user space since the last update
4246 * @cputime_scaled: cputime scaled by cpu frequency
4248 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
4249 cputime_t cputime_scaled
)
4251 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4254 /* Add user time to process. */
4255 p
->utime
= cputime_add(p
->utime
, cputime
);
4256 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4257 account_group_user_time(p
, cputime
);
4259 /* Add user time to cpustat. */
4260 tmp
= cputime_to_cputime64(cputime
);
4261 if (TASK_NICE(p
) > 0)
4262 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4264 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4265 /* Account for user time used */
4266 acct_update_integrals(p
);
4270 * Account guest cpu time to a process.
4271 * @p: the process that the cpu time gets accounted to
4272 * @cputime: the cpu time spent in virtual machine since the last update
4273 * @cputime_scaled: cputime scaled by cpu frequency
4275 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
4276 cputime_t cputime_scaled
)
4279 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4281 tmp
= cputime_to_cputime64(cputime
);
4283 /* Add guest time to process. */
4284 p
->utime
= cputime_add(p
->utime
, cputime
);
4285 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4286 account_group_user_time(p
, cputime
);
4287 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4289 /* Add guest time to cpustat. */
4290 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4291 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4295 * Account system cpu time to a process.
4296 * @p: the process that the cpu time gets accounted to
4297 * @hardirq_offset: the offset to subtract from hardirq_count()
4298 * @cputime: the cpu time spent in kernel space since the last update
4299 * @cputime_scaled: cputime scaled by cpu frequency
4301 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4302 cputime_t cputime
, cputime_t cputime_scaled
)
4304 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4307 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4308 account_guest_time(p
, cputime
, cputime_scaled
);
4312 /* Add system time to process. */
4313 p
->stime
= cputime_add(p
->stime
, cputime
);
4314 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
4315 account_group_system_time(p
, cputime
);
4317 /* Add system time to cpustat. */
4318 tmp
= cputime_to_cputime64(cputime
);
4319 if (hardirq_count() - hardirq_offset
)
4320 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4321 else if (softirq_count())
4322 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4324 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4326 /* Account for system time used */
4327 acct_update_integrals(p
);
4331 * Account for involuntary wait time.
4332 * @steal: the cpu time spent in involuntary wait
4334 void account_steal_time(cputime_t cputime
)
4336 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4337 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
4339 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
4343 * Account for idle time.
4344 * @cputime: the cpu time spent in idle wait
4346 void account_idle_time(cputime_t cputime
)
4348 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4349 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
4350 struct rq
*rq
= this_rq();
4352 if (atomic_read(&rq
->nr_iowait
) > 0)
4353 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
4355 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
4358 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4361 * Account a single tick of cpu time.
4362 * @p: the process that the cpu time gets accounted to
4363 * @user_tick: indicates if the tick is a user or a system tick
4365 void account_process_tick(struct task_struct
*p
, int user_tick
)
4367 cputime_t one_jiffy
= jiffies_to_cputime(1);
4368 cputime_t one_jiffy_scaled
= cputime_to_scaled(one_jiffy
);
4369 struct rq
*rq
= this_rq();
4372 account_user_time(p
, one_jiffy
, one_jiffy_scaled
);
4373 else if (p
!= rq
->idle
)
4374 account_system_time(p
, HARDIRQ_OFFSET
, one_jiffy
,
4377 account_idle_time(one_jiffy
);
4381 * Account multiple ticks of steal time.
4382 * @p: the process from which the cpu time has been stolen
4383 * @ticks: number of stolen ticks
4385 void account_steal_ticks(unsigned long ticks
)
4387 account_steal_time(jiffies_to_cputime(ticks
));
4391 * Account multiple ticks of idle time.
4392 * @ticks: number of stolen ticks
4394 void account_idle_ticks(unsigned long ticks
)
4396 account_idle_time(jiffies_to_cputime(ticks
));
4402 * Use precise platform statistics if available:
4404 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4405 cputime_t
task_utime(struct task_struct
*p
)
4410 cputime_t
task_stime(struct task_struct
*p
)
4415 cputime_t
task_utime(struct task_struct
*p
)
4417 clock_t utime
= cputime_to_clock_t(p
->utime
),
4418 total
= utime
+ cputime_to_clock_t(p
->stime
);
4422 * Use CFS's precise accounting:
4424 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
4428 do_div(temp
, total
);
4430 utime
= (clock_t)temp
;
4432 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
4433 return p
->prev_utime
;
4436 cputime_t
task_stime(struct task_struct
*p
)
4441 * Use CFS's precise accounting. (we subtract utime from
4442 * the total, to make sure the total observed by userspace
4443 * grows monotonically - apps rely on that):
4445 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
4446 cputime_to_clock_t(task_utime(p
));
4449 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
4451 return p
->prev_stime
;
4455 inline cputime_t
task_gtime(struct task_struct
*p
)
4461 * This function gets called by the timer code, with HZ frequency.
4462 * We call it with interrupts disabled.
4464 * It also gets called by the fork code, when changing the parent's
4467 void scheduler_tick(void)
4469 int cpu
= smp_processor_id();
4470 struct rq
*rq
= cpu_rq(cpu
);
4471 struct task_struct
*curr
= rq
->curr
;
4475 spin_lock(&rq
->lock
);
4476 update_rq_clock(rq
);
4477 update_cpu_load(rq
);
4478 curr
->sched_class
->task_tick(rq
, curr
, 0);
4479 perf_counter_task_tick(curr
, cpu
);
4480 spin_unlock(&rq
->lock
);
4483 rq
->idle_at_tick
= idle_cpu(cpu
);
4484 trigger_load_balance(rq
, cpu
);
4488 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4489 defined(CONFIG_PREEMPT_TRACER))
4491 static inline unsigned long get_parent_ip(unsigned long addr
)
4493 if (in_lock_functions(addr
)) {
4494 addr
= CALLER_ADDR2
;
4495 if (in_lock_functions(addr
))
4496 addr
= CALLER_ADDR3
;
4501 void __kprobes
add_preempt_count(int val
)
4503 #ifdef CONFIG_DEBUG_PREEMPT
4507 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4510 preempt_count() += val
;
4511 #ifdef CONFIG_DEBUG_PREEMPT
4513 * Spinlock count overflowing soon?
4515 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4518 if (preempt_count() == val
)
4519 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4521 EXPORT_SYMBOL(add_preempt_count
);
4523 void __kprobes
sub_preempt_count(int val
)
4525 #ifdef CONFIG_DEBUG_PREEMPT
4529 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4532 * Is the spinlock portion underflowing?
4534 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4535 !(preempt_count() & PREEMPT_MASK
)))
4539 if (preempt_count() == val
)
4540 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4541 preempt_count() -= val
;
4543 EXPORT_SYMBOL(sub_preempt_count
);
4548 * Print scheduling while atomic bug:
4550 static noinline
void __schedule_bug(struct task_struct
*prev
)
4552 struct pt_regs
*regs
= get_irq_regs();
4554 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4555 prev
->comm
, prev
->pid
, preempt_count());
4557 debug_show_held_locks(prev
);
4559 if (irqs_disabled())
4560 print_irqtrace_events(prev
);
4569 * Various schedule()-time debugging checks and statistics:
4571 static inline void schedule_debug(struct task_struct
*prev
)
4574 * Test if we are atomic. Since do_exit() needs to call into
4575 * schedule() atomically, we ignore that path for now.
4576 * Otherwise, whine if we are scheduling when we should not be.
4578 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4579 __schedule_bug(prev
);
4581 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4583 schedstat_inc(this_rq(), sched_count
);
4584 #ifdef CONFIG_SCHEDSTATS
4585 if (unlikely(prev
->lock_depth
>= 0)) {
4586 schedstat_inc(this_rq(), bkl_count
);
4587 schedstat_inc(prev
, sched_info
.bkl_count
);
4593 * Pick up the highest-prio task:
4595 static inline struct task_struct
*
4596 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
4598 const struct sched_class
*class;
4599 struct task_struct
*p
;
4602 * Optimization: we know that if all tasks are in
4603 * the fair class we can call that function directly:
4605 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4606 p
= fair_sched_class
.pick_next_task(rq
);
4611 class = sched_class_highest
;
4613 p
= class->pick_next_task(rq
);
4617 * Will never be NULL as the idle class always
4618 * returns a non-NULL p:
4620 class = class->next
;
4625 * schedule() is the main scheduler function.
4627 asmlinkage
void __sched
schedule(void)
4629 struct task_struct
*prev
, *next
;
4630 unsigned long *switch_count
;
4636 cpu
= smp_processor_id();
4640 switch_count
= &prev
->nivcsw
;
4642 release_kernel_lock(prev
);
4643 need_resched_nonpreemptible
:
4645 schedule_debug(prev
);
4647 if (sched_feat(HRTICK
))
4650 spin_lock_irq(&rq
->lock
);
4651 update_rq_clock(rq
);
4652 clear_tsk_need_resched(prev
);
4654 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4655 if (unlikely(signal_pending_state(prev
->state
, prev
)))
4656 prev
->state
= TASK_RUNNING
;
4658 deactivate_task(rq
, prev
, 1);
4659 switch_count
= &prev
->nvcsw
;
4663 if (prev
->sched_class
->pre_schedule
)
4664 prev
->sched_class
->pre_schedule(rq
, prev
);
4667 if (unlikely(!rq
->nr_running
))
4668 idle_balance(cpu
, rq
);
4670 prev
->sched_class
->put_prev_task(rq
, prev
);
4671 next
= pick_next_task(rq
, prev
);
4673 if (likely(prev
!= next
)) {
4674 sched_info_switch(prev
, next
);
4675 perf_counter_task_sched_out(prev
, cpu
);
4681 context_switch(rq
, prev
, next
); /* unlocks the rq */
4683 * the context switch might have flipped the stack from under
4684 * us, hence refresh the local variables.
4686 cpu
= smp_processor_id();
4689 spin_unlock_irq(&rq
->lock
);
4691 if (unlikely(reacquire_kernel_lock(current
) < 0))
4692 goto need_resched_nonpreemptible
;
4694 preempt_enable_no_resched();
4695 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
4698 EXPORT_SYMBOL(schedule
);
4700 #ifdef CONFIG_PREEMPT
4702 * this is the entry point to schedule() from in-kernel preemption
4703 * off of preempt_enable. Kernel preemptions off return from interrupt
4704 * occur there and call schedule directly.
4706 asmlinkage
void __sched
preempt_schedule(void)
4708 struct thread_info
*ti
= current_thread_info();
4711 * If there is a non-zero preempt_count or interrupts are disabled,
4712 * we do not want to preempt the current task. Just return..
4714 if (likely(ti
->preempt_count
|| irqs_disabled()))
4718 add_preempt_count(PREEMPT_ACTIVE
);
4720 sub_preempt_count(PREEMPT_ACTIVE
);
4723 * Check again in case we missed a preemption opportunity
4724 * between schedule and now.
4727 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4729 EXPORT_SYMBOL(preempt_schedule
);
4732 * this is the entry point to schedule() from kernel preemption
4733 * off of irq context.
4734 * Note, that this is called and return with irqs disabled. This will
4735 * protect us against recursive calling from irq.
4737 asmlinkage
void __sched
preempt_schedule_irq(void)
4739 struct thread_info
*ti
= current_thread_info();
4741 /* Catch callers which need to be fixed */
4742 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4745 add_preempt_count(PREEMPT_ACTIVE
);
4748 local_irq_disable();
4749 sub_preempt_count(PREEMPT_ACTIVE
);
4752 * Check again in case we missed a preemption opportunity
4753 * between schedule and now.
4756 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4759 #endif /* CONFIG_PREEMPT */
4761 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4764 return try_to_wake_up(curr
->private, mode
, sync
);
4766 EXPORT_SYMBOL(default_wake_function
);
4769 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4770 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4771 * number) then we wake all the non-exclusive tasks and one exclusive task.
4773 * There are circumstances in which we can try to wake a task which has already
4774 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4775 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4777 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4778 int nr_exclusive
, int sync
, void *key
)
4780 wait_queue_t
*curr
, *next
;
4782 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4783 unsigned flags
= curr
->flags
;
4785 if (curr
->func(curr
, mode
, sync
, key
) &&
4786 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4792 * __wake_up - wake up threads blocked on a waitqueue.
4794 * @mode: which threads
4795 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4796 * @key: is directly passed to the wakeup function
4798 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4799 int nr_exclusive
, void *key
)
4801 unsigned long flags
;
4803 spin_lock_irqsave(&q
->lock
, flags
);
4804 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4805 spin_unlock_irqrestore(&q
->lock
, flags
);
4807 EXPORT_SYMBOL(__wake_up
);
4810 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4812 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4814 __wake_up_common(q
, mode
, 1, 0, NULL
);
4818 * __wake_up_sync - wake up threads blocked on a waitqueue.
4820 * @mode: which threads
4821 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4823 * The sync wakeup differs that the waker knows that it will schedule
4824 * away soon, so while the target thread will be woken up, it will not
4825 * be migrated to another CPU - ie. the two threads are 'synchronized'
4826 * with each other. This can prevent needless bouncing between CPUs.
4828 * On UP it can prevent extra preemption.
4831 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4833 unsigned long flags
;
4839 if (unlikely(!nr_exclusive
))
4842 spin_lock_irqsave(&q
->lock
, flags
);
4843 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4844 spin_unlock_irqrestore(&q
->lock
, flags
);
4846 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4849 * complete: - signals a single thread waiting on this completion
4850 * @x: holds the state of this particular completion
4852 * This will wake up a single thread waiting on this completion. Threads will be
4853 * awakened in the same order in which they were queued.
4855 * See also complete_all(), wait_for_completion() and related routines.
4857 void complete(struct completion
*x
)
4859 unsigned long flags
;
4861 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4863 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4864 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4866 EXPORT_SYMBOL(complete
);
4869 * complete_all: - signals all threads waiting on this completion
4870 * @x: holds the state of this particular completion
4872 * This will wake up all threads waiting on this particular completion event.
4874 void complete_all(struct completion
*x
)
4876 unsigned long flags
;
4878 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4879 x
->done
+= UINT_MAX
/2;
4880 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4881 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4883 EXPORT_SYMBOL(complete_all
);
4885 static inline long __sched
4886 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4889 DECLARE_WAITQUEUE(wait
, current
);
4891 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4892 __add_wait_queue_tail(&x
->wait
, &wait
);
4894 if (signal_pending_state(state
, current
)) {
4895 timeout
= -ERESTARTSYS
;
4898 __set_current_state(state
);
4899 spin_unlock_irq(&x
->wait
.lock
);
4900 timeout
= schedule_timeout(timeout
);
4901 spin_lock_irq(&x
->wait
.lock
);
4902 } while (!x
->done
&& timeout
);
4903 __remove_wait_queue(&x
->wait
, &wait
);
4908 return timeout
?: 1;
4912 wait_for_common(struct completion
*x
, long timeout
, int state
)
4916 spin_lock_irq(&x
->wait
.lock
);
4917 timeout
= do_wait_for_common(x
, timeout
, state
);
4918 spin_unlock_irq(&x
->wait
.lock
);
4923 * wait_for_completion: - waits for completion of a task
4924 * @x: holds the state of this particular completion
4926 * This waits to be signaled for completion of a specific task. It is NOT
4927 * interruptible and there is no timeout.
4929 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4930 * and interrupt capability. Also see complete().
4932 void __sched
wait_for_completion(struct completion
*x
)
4934 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4936 EXPORT_SYMBOL(wait_for_completion
);
4939 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4940 * @x: holds the state of this particular completion
4941 * @timeout: timeout value in jiffies
4943 * This waits for either a completion of a specific task to be signaled or for a
4944 * specified timeout to expire. The timeout is in jiffies. It is not
4947 unsigned long __sched
4948 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4950 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4952 EXPORT_SYMBOL(wait_for_completion_timeout
);
4955 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4956 * @x: holds the state of this particular completion
4958 * This waits for completion of a specific task to be signaled. It is
4961 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4963 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4964 if (t
== -ERESTARTSYS
)
4968 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4971 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4972 * @x: holds the state of this particular completion
4973 * @timeout: timeout value in jiffies
4975 * This waits for either a completion of a specific task to be signaled or for a
4976 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4978 unsigned long __sched
4979 wait_for_completion_interruptible_timeout(struct completion
*x
,
4980 unsigned long timeout
)
4982 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4984 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4987 * wait_for_completion_killable: - waits for completion of a task (killable)
4988 * @x: holds the state of this particular completion
4990 * This waits to be signaled for completion of a specific task. It can be
4991 * interrupted by a kill signal.
4993 int __sched
wait_for_completion_killable(struct completion
*x
)
4995 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4996 if (t
== -ERESTARTSYS
)
5000 EXPORT_SYMBOL(wait_for_completion_killable
);
5003 * try_wait_for_completion - try to decrement a completion without blocking
5004 * @x: completion structure
5006 * Returns: 0 if a decrement cannot be done without blocking
5007 * 1 if a decrement succeeded.
5009 * If a completion is being used as a counting completion,
5010 * attempt to decrement the counter without blocking. This
5011 * enables us to avoid waiting if the resource the completion
5012 * is protecting is not available.
5014 bool try_wait_for_completion(struct completion
*x
)
5018 spin_lock_irq(&x
->wait
.lock
);
5023 spin_unlock_irq(&x
->wait
.lock
);
5026 EXPORT_SYMBOL(try_wait_for_completion
);
5029 * completion_done - Test to see if a completion has any waiters
5030 * @x: completion structure
5032 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5033 * 1 if there are no waiters.
5036 bool completion_done(struct completion
*x
)
5040 spin_lock_irq(&x
->wait
.lock
);
5043 spin_unlock_irq(&x
->wait
.lock
);
5046 EXPORT_SYMBOL(completion_done
);
5049 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
5051 unsigned long flags
;
5054 init_waitqueue_entry(&wait
, current
);
5056 __set_current_state(state
);
5058 spin_lock_irqsave(&q
->lock
, flags
);
5059 __add_wait_queue(q
, &wait
);
5060 spin_unlock(&q
->lock
);
5061 timeout
= schedule_timeout(timeout
);
5062 spin_lock_irq(&q
->lock
);
5063 __remove_wait_queue(q
, &wait
);
5064 spin_unlock_irqrestore(&q
->lock
, flags
);
5069 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
5071 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5073 EXPORT_SYMBOL(interruptible_sleep_on
);
5076 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5078 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
5080 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
5082 void __sched
sleep_on(wait_queue_head_t
*q
)
5084 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5086 EXPORT_SYMBOL(sleep_on
);
5088 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5090 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
5092 EXPORT_SYMBOL(sleep_on_timeout
);
5094 #ifdef CONFIG_RT_MUTEXES
5097 * rt_mutex_setprio - set the current priority of a task
5099 * @prio: prio value (kernel-internal form)
5101 * This function changes the 'effective' priority of a task. It does
5102 * not touch ->normal_prio like __setscheduler().
5104 * Used by the rt_mutex code to implement priority inheritance logic.
5106 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
5108 unsigned long flags
;
5109 int oldprio
, on_rq
, running
;
5111 const struct sched_class
*prev_class
= p
->sched_class
;
5113 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
5115 rq
= task_rq_lock(p
, &flags
);
5116 update_rq_clock(rq
);
5119 on_rq
= p
->se
.on_rq
;
5120 running
= task_current(rq
, p
);
5122 dequeue_task(rq
, p
, 0);
5124 p
->sched_class
->put_prev_task(rq
, p
);
5127 p
->sched_class
= &rt_sched_class
;
5129 p
->sched_class
= &fair_sched_class
;
5134 p
->sched_class
->set_curr_task(rq
);
5136 enqueue_task(rq
, p
, 0);
5138 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5140 task_rq_unlock(rq
, &flags
);
5145 void set_user_nice(struct task_struct
*p
, long nice
)
5147 int old_prio
, delta
, on_rq
;
5148 unsigned long flags
;
5151 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
5154 * We have to be careful, if called from sys_setpriority(),
5155 * the task might be in the middle of scheduling on another CPU.
5157 rq
= task_rq_lock(p
, &flags
);
5158 update_rq_clock(rq
);
5160 * The RT priorities are set via sched_setscheduler(), but we still
5161 * allow the 'normal' nice value to be set - but as expected
5162 * it wont have any effect on scheduling until the task is
5163 * SCHED_FIFO/SCHED_RR:
5165 if (task_has_rt_policy(p
)) {
5166 p
->static_prio
= NICE_TO_PRIO(nice
);
5169 on_rq
= p
->se
.on_rq
;
5171 dequeue_task(rq
, p
, 0);
5173 p
->static_prio
= NICE_TO_PRIO(nice
);
5176 p
->prio
= effective_prio(p
);
5177 delta
= p
->prio
- old_prio
;
5180 enqueue_task(rq
, p
, 0);
5182 * If the task increased its priority or is running and
5183 * lowered its priority, then reschedule its CPU:
5185 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
5186 resched_task(rq
->curr
);
5189 task_rq_unlock(rq
, &flags
);
5191 EXPORT_SYMBOL(set_user_nice
);
5194 * can_nice - check if a task can reduce its nice value
5198 int can_nice(const struct task_struct
*p
, const int nice
)
5200 /* convert nice value [19,-20] to rlimit style value [1,40] */
5201 int nice_rlim
= 20 - nice
;
5203 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
5204 capable(CAP_SYS_NICE
));
5207 #ifdef __ARCH_WANT_SYS_NICE
5210 * sys_nice - change the priority of the current process.
5211 * @increment: priority increment
5213 * sys_setpriority is a more generic, but much slower function that
5214 * does similar things.
5216 SYSCALL_DEFINE1(nice
, int, increment
)
5221 * Setpriority might change our priority at the same moment.
5222 * We don't have to worry. Conceptually one call occurs first
5223 * and we have a single winner.
5225 if (increment
< -40)
5230 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
5236 if (increment
< 0 && !can_nice(current
, nice
))
5239 retval
= security_task_setnice(current
, nice
);
5243 set_user_nice(current
, nice
);
5250 * task_prio - return the priority value of a given task.
5251 * @p: the task in question.
5253 * This is the priority value as seen by users in /proc.
5254 * RT tasks are offset by -200. Normal tasks are centered
5255 * around 0, value goes from -16 to +15.
5257 int task_prio(const struct task_struct
*p
)
5259 return p
->prio
- MAX_RT_PRIO
;
5263 * task_nice - return the nice value of a given task.
5264 * @p: the task in question.
5266 int task_nice(const struct task_struct
*p
)
5268 return TASK_NICE(p
);
5270 EXPORT_SYMBOL(task_nice
);
5273 * idle_cpu - is a given cpu idle currently?
5274 * @cpu: the processor in question.
5276 int idle_cpu(int cpu
)
5278 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
5282 * idle_task - return the idle task for a given cpu.
5283 * @cpu: the processor in question.
5285 struct task_struct
*idle_task(int cpu
)
5287 return cpu_rq(cpu
)->idle
;
5291 * find_process_by_pid - find a process with a matching PID value.
5292 * @pid: the pid in question.
5294 static struct task_struct
*find_process_by_pid(pid_t pid
)
5296 return pid
? find_task_by_vpid(pid
) : current
;
5299 /* Actually do priority change: must hold rq lock. */
5301 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
5303 BUG_ON(p
->se
.on_rq
);
5306 switch (p
->policy
) {
5310 p
->sched_class
= &fair_sched_class
;
5314 p
->sched_class
= &rt_sched_class
;
5318 p
->rt_priority
= prio
;
5319 p
->normal_prio
= normal_prio(p
);
5320 /* we are holding p->pi_lock already */
5321 p
->prio
= rt_mutex_getprio(p
);
5326 * check the target process has a UID that matches the current process's
5328 static bool check_same_owner(struct task_struct
*p
)
5330 const struct cred
*cred
= current_cred(), *pcred
;
5334 pcred
= __task_cred(p
);
5335 match
= (cred
->euid
== pcred
->euid
||
5336 cred
->euid
== pcred
->uid
);
5341 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
5342 struct sched_param
*param
, bool user
)
5344 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
5345 unsigned long flags
;
5346 const struct sched_class
*prev_class
= p
->sched_class
;
5349 /* may grab non-irq protected spin_locks */
5350 BUG_ON(in_interrupt());
5352 /* double check policy once rq lock held */
5354 policy
= oldpolicy
= p
->policy
;
5355 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
5356 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
5357 policy
!= SCHED_IDLE
)
5360 * Valid priorities for SCHED_FIFO and SCHED_RR are
5361 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5362 * SCHED_BATCH and SCHED_IDLE is 0.
5364 if (param
->sched_priority
< 0 ||
5365 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5366 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
5368 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
5372 * Allow unprivileged RT tasks to decrease priority:
5374 if (user
&& !capable(CAP_SYS_NICE
)) {
5375 if (rt_policy(policy
)) {
5376 unsigned long rlim_rtprio
;
5378 if (!lock_task_sighand(p
, &flags
))
5380 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
5381 unlock_task_sighand(p
, &flags
);
5383 /* can't set/change the rt policy */
5384 if (policy
!= p
->policy
&& !rlim_rtprio
)
5387 /* can't increase priority */
5388 if (param
->sched_priority
> p
->rt_priority
&&
5389 param
->sched_priority
> rlim_rtprio
)
5393 * Like positive nice levels, dont allow tasks to
5394 * move out of SCHED_IDLE either:
5396 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
5399 /* can't change other user's priorities */
5400 if (!check_same_owner(p
))
5405 #ifdef CONFIG_RT_GROUP_SCHED
5407 * Do not allow realtime tasks into groups that have no runtime
5410 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
5411 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
5415 retval
= security_task_setscheduler(p
, policy
, param
);
5421 * make sure no PI-waiters arrive (or leave) while we are
5422 * changing the priority of the task:
5424 spin_lock_irqsave(&p
->pi_lock
, flags
);
5426 * To be able to change p->policy safely, the apropriate
5427 * runqueue lock must be held.
5429 rq
= __task_rq_lock(p
);
5430 /* recheck policy now with rq lock held */
5431 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5432 policy
= oldpolicy
= -1;
5433 __task_rq_unlock(rq
);
5434 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5437 update_rq_clock(rq
);
5438 on_rq
= p
->se
.on_rq
;
5439 running
= task_current(rq
, p
);
5441 deactivate_task(rq
, p
, 0);
5443 p
->sched_class
->put_prev_task(rq
, p
);
5446 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5449 p
->sched_class
->set_curr_task(rq
);
5451 activate_task(rq
, p
, 0);
5453 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5455 __task_rq_unlock(rq
);
5456 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5458 rt_mutex_adjust_pi(p
);
5464 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5465 * @p: the task in question.
5466 * @policy: new policy.
5467 * @param: structure containing the new RT priority.
5469 * NOTE that the task may be already dead.
5471 int sched_setscheduler(struct task_struct
*p
, int policy
,
5472 struct sched_param
*param
)
5474 return __sched_setscheduler(p
, policy
, param
, true);
5476 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5479 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5480 * @p: the task in question.
5481 * @policy: new policy.
5482 * @param: structure containing the new RT priority.
5484 * Just like sched_setscheduler, only don't bother checking if the
5485 * current context has permission. For example, this is needed in
5486 * stop_machine(): we create temporary high priority worker threads,
5487 * but our caller might not have that capability.
5489 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5490 struct sched_param
*param
)
5492 return __sched_setscheduler(p
, policy
, param
, false);
5496 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5498 struct sched_param lparam
;
5499 struct task_struct
*p
;
5502 if (!param
|| pid
< 0)
5504 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5509 p
= find_process_by_pid(pid
);
5511 retval
= sched_setscheduler(p
, policy
, &lparam
);
5518 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5519 * @pid: the pid in question.
5520 * @policy: new policy.
5521 * @param: structure containing the new RT priority.
5523 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
5524 struct sched_param __user
*, param
)
5526 /* negative values for policy are not valid */
5530 return do_sched_setscheduler(pid
, policy
, param
);
5534 * sys_sched_setparam - set/change the RT priority of a thread
5535 * @pid: the pid in question.
5536 * @param: structure containing the new RT priority.
5538 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5540 return do_sched_setscheduler(pid
, -1, param
);
5544 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5545 * @pid: the pid in question.
5547 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
5549 struct task_struct
*p
;
5556 read_lock(&tasklist_lock
);
5557 p
= find_process_by_pid(pid
);
5559 retval
= security_task_getscheduler(p
);
5563 read_unlock(&tasklist_lock
);
5568 * sys_sched_getscheduler - get the RT priority of a thread
5569 * @pid: the pid in question.
5570 * @param: structure containing the RT priority.
5572 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5574 struct sched_param lp
;
5575 struct task_struct
*p
;
5578 if (!param
|| pid
< 0)
5581 read_lock(&tasklist_lock
);
5582 p
= find_process_by_pid(pid
);
5587 retval
= security_task_getscheduler(p
);
5591 lp
.sched_priority
= p
->rt_priority
;
5592 read_unlock(&tasklist_lock
);
5595 * This one might sleep, we cannot do it with a spinlock held ...
5597 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5602 read_unlock(&tasklist_lock
);
5606 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
5608 cpumask_var_t cpus_allowed
, new_mask
;
5609 struct task_struct
*p
;
5613 read_lock(&tasklist_lock
);
5615 p
= find_process_by_pid(pid
);
5617 read_unlock(&tasklist_lock
);
5623 * It is not safe to call set_cpus_allowed with the
5624 * tasklist_lock held. We will bump the task_struct's
5625 * usage count and then drop tasklist_lock.
5628 read_unlock(&tasklist_lock
);
5630 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
5634 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
5636 goto out_free_cpus_allowed
;
5639 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
5642 retval
= security_task_setscheduler(p
, 0, NULL
);
5646 cpuset_cpus_allowed(p
, cpus_allowed
);
5647 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
5649 retval
= set_cpus_allowed_ptr(p
, new_mask
);
5652 cpuset_cpus_allowed(p
, cpus_allowed
);
5653 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
5655 * We must have raced with a concurrent cpuset
5656 * update. Just reset the cpus_allowed to the
5657 * cpuset's cpus_allowed
5659 cpumask_copy(new_mask
, cpus_allowed
);
5664 free_cpumask_var(new_mask
);
5665 out_free_cpus_allowed
:
5666 free_cpumask_var(cpus_allowed
);
5673 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5674 struct cpumask
*new_mask
)
5676 if (len
< cpumask_size())
5677 cpumask_clear(new_mask
);
5678 else if (len
> cpumask_size())
5679 len
= cpumask_size();
5681 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5685 * sys_sched_setaffinity - set the cpu affinity of a process
5686 * @pid: pid of the process
5687 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5688 * @user_mask_ptr: user-space pointer to the new cpu mask
5690 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
5691 unsigned long __user
*, user_mask_ptr
)
5693 cpumask_var_t new_mask
;
5696 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
5699 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
5701 retval
= sched_setaffinity(pid
, new_mask
);
5702 free_cpumask_var(new_mask
);
5706 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
5708 struct task_struct
*p
;
5712 read_lock(&tasklist_lock
);
5715 p
= find_process_by_pid(pid
);
5719 retval
= security_task_getscheduler(p
);
5723 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
5726 read_unlock(&tasklist_lock
);
5733 * sys_sched_getaffinity - get the cpu affinity of a process
5734 * @pid: pid of the process
5735 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5736 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5738 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
5739 unsigned long __user
*, user_mask_ptr
)
5744 if (len
< cpumask_size())
5747 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
5750 ret
= sched_getaffinity(pid
, mask
);
5752 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
5755 ret
= cpumask_size();
5757 free_cpumask_var(mask
);
5763 * sys_sched_yield - yield the current processor to other threads.
5765 * This function yields the current CPU to other tasks. If there are no
5766 * other threads running on this CPU then this function will return.
5768 SYSCALL_DEFINE0(sched_yield
)
5770 struct rq
*rq
= this_rq_lock();
5772 schedstat_inc(rq
, yld_count
);
5773 current
->sched_class
->yield_task(rq
);
5776 * Since we are going to call schedule() anyway, there's
5777 * no need to preempt or enable interrupts:
5779 __release(rq
->lock
);
5780 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5781 _raw_spin_unlock(&rq
->lock
);
5782 preempt_enable_no_resched();
5789 static void __cond_resched(void)
5791 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5792 __might_sleep(__FILE__
, __LINE__
);
5795 * The BKS might be reacquired before we have dropped
5796 * PREEMPT_ACTIVE, which could trigger a second
5797 * cond_resched() call.
5800 add_preempt_count(PREEMPT_ACTIVE
);
5802 sub_preempt_count(PREEMPT_ACTIVE
);
5803 } while (need_resched());
5806 int __sched
_cond_resched(void)
5808 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
5809 system_state
== SYSTEM_RUNNING
) {
5815 EXPORT_SYMBOL(_cond_resched
);
5818 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5819 * call schedule, and on return reacquire the lock.
5821 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5822 * operations here to prevent schedule() from being called twice (once via
5823 * spin_unlock(), once by hand).
5825 int cond_resched_lock(spinlock_t
*lock
)
5827 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
5830 if (spin_needbreak(lock
) || resched
) {
5832 if (resched
&& need_resched())
5841 EXPORT_SYMBOL(cond_resched_lock
);
5843 int __sched
cond_resched_softirq(void)
5845 BUG_ON(!in_softirq());
5847 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
5855 EXPORT_SYMBOL(cond_resched_softirq
);
5858 * yield - yield the current processor to other threads.
5860 * This is a shortcut for kernel-space yielding - it marks the
5861 * thread runnable and calls sys_sched_yield().
5863 void __sched
yield(void)
5865 set_current_state(TASK_RUNNING
);
5868 EXPORT_SYMBOL(yield
);
5871 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5872 * that process accounting knows that this is a task in IO wait state.
5874 * But don't do that if it is a deliberate, throttling IO wait (this task
5875 * has set its backing_dev_info: the queue against which it should throttle)
5877 void __sched
io_schedule(void)
5879 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5881 delayacct_blkio_start();
5882 atomic_inc(&rq
->nr_iowait
);
5884 atomic_dec(&rq
->nr_iowait
);
5885 delayacct_blkio_end();
5887 EXPORT_SYMBOL(io_schedule
);
5889 long __sched
io_schedule_timeout(long timeout
)
5891 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5894 delayacct_blkio_start();
5895 atomic_inc(&rq
->nr_iowait
);
5896 ret
= schedule_timeout(timeout
);
5897 atomic_dec(&rq
->nr_iowait
);
5898 delayacct_blkio_end();
5903 * sys_sched_get_priority_max - return maximum RT priority.
5904 * @policy: scheduling class.
5906 * this syscall returns the maximum rt_priority that can be used
5907 * by a given scheduling class.
5909 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5916 ret
= MAX_USER_RT_PRIO
-1;
5928 * sys_sched_get_priority_min - return minimum RT priority.
5929 * @policy: scheduling class.
5931 * this syscall returns the minimum rt_priority that can be used
5932 * by a given scheduling class.
5934 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5952 * sys_sched_rr_get_interval - return the default timeslice of a process.
5953 * @pid: pid of the process.
5954 * @interval: userspace pointer to the timeslice value.
5956 * this syscall writes the default timeslice value of a given process
5957 * into the user-space timespec buffer. A value of '0' means infinity.
5959 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5960 struct timespec __user
*, interval
)
5962 struct task_struct
*p
;
5963 unsigned int time_slice
;
5971 read_lock(&tasklist_lock
);
5972 p
= find_process_by_pid(pid
);
5976 retval
= security_task_getscheduler(p
);
5981 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5982 * tasks that are on an otherwise idle runqueue:
5985 if (p
->policy
== SCHED_RR
) {
5986 time_slice
= DEF_TIMESLICE
;
5987 } else if (p
->policy
!= SCHED_FIFO
) {
5988 struct sched_entity
*se
= &p
->se
;
5989 unsigned long flags
;
5992 rq
= task_rq_lock(p
, &flags
);
5993 if (rq
->cfs
.load
.weight
)
5994 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5995 task_rq_unlock(rq
, &flags
);
5997 read_unlock(&tasklist_lock
);
5998 jiffies_to_timespec(time_slice
, &t
);
5999 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
6003 read_unlock(&tasklist_lock
);
6007 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
6009 void sched_show_task(struct task_struct
*p
)
6011 unsigned long free
= 0;
6014 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
6015 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
6016 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
6017 #if BITS_PER_LONG == 32
6018 if (state
== TASK_RUNNING
)
6019 printk(KERN_CONT
" running ");
6021 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
6023 if (state
== TASK_RUNNING
)
6024 printk(KERN_CONT
" running task ");
6026 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
6028 #ifdef CONFIG_DEBUG_STACK_USAGE
6029 free
= stack_not_used(p
);
6031 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
6032 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
6034 show_stack(p
, NULL
);
6037 void show_state_filter(unsigned long state_filter
)
6039 struct task_struct
*g
, *p
;
6041 #if BITS_PER_LONG == 32
6043 " task PC stack pid father\n");
6046 " task PC stack pid father\n");
6048 read_lock(&tasklist_lock
);
6049 do_each_thread(g
, p
) {
6051 * reset the NMI-timeout, listing all files on a slow
6052 * console might take alot of time:
6054 touch_nmi_watchdog();
6055 if (!state_filter
|| (p
->state
& state_filter
))
6057 } while_each_thread(g
, p
);
6059 touch_all_softlockup_watchdogs();
6061 #ifdef CONFIG_SCHED_DEBUG
6062 sysrq_sched_debug_show();
6064 read_unlock(&tasklist_lock
);
6066 * Only show locks if all tasks are dumped:
6068 if (state_filter
== -1)
6069 debug_show_all_locks();
6072 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
6074 idle
->sched_class
= &idle_sched_class
;
6078 * init_idle - set up an idle thread for a given CPU
6079 * @idle: task in question
6080 * @cpu: cpu the idle task belongs to
6082 * NOTE: this function does not set the idle thread's NEED_RESCHED
6083 * flag, to make booting more robust.
6085 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
6087 struct rq
*rq
= cpu_rq(cpu
);
6088 unsigned long flags
;
6090 spin_lock_irqsave(&rq
->lock
, flags
);
6093 idle
->se
.exec_start
= sched_clock();
6095 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
6096 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
6097 __set_task_cpu(idle
, cpu
);
6099 rq
->curr
= rq
->idle
= idle
;
6100 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6103 spin_unlock_irqrestore(&rq
->lock
, flags
);
6105 /* Set the preempt count _outside_ the spinlocks! */
6106 #if defined(CONFIG_PREEMPT)
6107 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
6109 task_thread_info(idle
)->preempt_count
= 0;
6112 * The idle tasks have their own, simple scheduling class:
6114 idle
->sched_class
= &idle_sched_class
;
6115 ftrace_graph_init_task(idle
);
6119 * In a system that switches off the HZ timer nohz_cpu_mask
6120 * indicates which cpus entered this state. This is used
6121 * in the rcu update to wait only for active cpus. For system
6122 * which do not switch off the HZ timer nohz_cpu_mask should
6123 * always be CPU_BITS_NONE.
6125 cpumask_var_t nohz_cpu_mask
;
6128 * Increase the granularity value when there are more CPUs,
6129 * because with more CPUs the 'effective latency' as visible
6130 * to users decreases. But the relationship is not linear,
6131 * so pick a second-best guess by going with the log2 of the
6134 * This idea comes from the SD scheduler of Con Kolivas:
6136 static inline void sched_init_granularity(void)
6138 unsigned int factor
= 1 + ilog2(num_online_cpus());
6139 const unsigned long limit
= 200000000;
6141 sysctl_sched_min_granularity
*= factor
;
6142 if (sysctl_sched_min_granularity
> limit
)
6143 sysctl_sched_min_granularity
= limit
;
6145 sysctl_sched_latency
*= factor
;
6146 if (sysctl_sched_latency
> limit
)
6147 sysctl_sched_latency
= limit
;
6149 sysctl_sched_wakeup_granularity
*= factor
;
6151 sysctl_sched_shares_ratelimit
*= factor
;
6156 * This is how migration works:
6158 * 1) we queue a struct migration_req structure in the source CPU's
6159 * runqueue and wake up that CPU's migration thread.
6160 * 2) we down() the locked semaphore => thread blocks.
6161 * 3) migration thread wakes up (implicitly it forces the migrated
6162 * thread off the CPU)
6163 * 4) it gets the migration request and checks whether the migrated
6164 * task is still in the wrong runqueue.
6165 * 5) if it's in the wrong runqueue then the migration thread removes
6166 * it and puts it into the right queue.
6167 * 6) migration thread up()s the semaphore.
6168 * 7) we wake up and the migration is done.
6172 * Change a given task's CPU affinity. Migrate the thread to a
6173 * proper CPU and schedule it away if the CPU it's executing on
6174 * is removed from the allowed bitmask.
6176 * NOTE: the caller must have a valid reference to the task, the
6177 * task must not exit() & deallocate itself prematurely. The
6178 * call is not atomic; no spinlocks may be held.
6180 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
6182 struct migration_req req
;
6183 unsigned long flags
;
6187 rq
= task_rq_lock(p
, &flags
);
6188 if (!cpumask_intersects(new_mask
, cpu_online_mask
)) {
6193 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
6194 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
6199 if (p
->sched_class
->set_cpus_allowed
)
6200 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
6202 cpumask_copy(&p
->cpus_allowed
, new_mask
);
6203 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
6206 /* Can the task run on the task's current CPU? If so, we're done */
6207 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
6210 if (migrate_task(p
, cpumask_any_and(cpu_online_mask
, new_mask
), &req
)) {
6211 /* Need help from migration thread: drop lock and wait. */
6212 task_rq_unlock(rq
, &flags
);
6213 wake_up_process(rq
->migration_thread
);
6214 wait_for_completion(&req
.done
);
6215 tlb_migrate_finish(p
->mm
);
6219 task_rq_unlock(rq
, &flags
);
6223 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
6226 * Move (not current) task off this cpu, onto dest cpu. We're doing
6227 * this because either it can't run here any more (set_cpus_allowed()
6228 * away from this CPU, or CPU going down), or because we're
6229 * attempting to rebalance this task on exec (sched_exec).
6231 * So we race with normal scheduler movements, but that's OK, as long
6232 * as the task is no longer on this CPU.
6234 * Returns non-zero if task was successfully migrated.
6236 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6238 struct rq
*rq_dest
, *rq_src
;
6241 if (unlikely(!cpu_active(dest_cpu
)))
6244 rq_src
= cpu_rq(src_cpu
);
6245 rq_dest
= cpu_rq(dest_cpu
);
6247 double_rq_lock(rq_src
, rq_dest
);
6248 /* Already moved. */
6249 if (task_cpu(p
) != src_cpu
)
6251 /* Affinity changed (again). */
6252 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
6255 on_rq
= p
->se
.on_rq
;
6257 deactivate_task(rq_src
, p
, 0);
6259 set_task_cpu(p
, dest_cpu
);
6261 activate_task(rq_dest
, p
, 0);
6262 check_preempt_curr(rq_dest
, p
, 0);
6267 double_rq_unlock(rq_src
, rq_dest
);
6272 * migration_thread - this is a highprio system thread that performs
6273 * thread migration by bumping thread off CPU then 'pushing' onto
6276 static int migration_thread(void *data
)
6278 int cpu
= (long)data
;
6282 BUG_ON(rq
->migration_thread
!= current
);
6284 set_current_state(TASK_INTERRUPTIBLE
);
6285 while (!kthread_should_stop()) {
6286 struct migration_req
*req
;
6287 struct list_head
*head
;
6289 spin_lock_irq(&rq
->lock
);
6291 if (cpu_is_offline(cpu
)) {
6292 spin_unlock_irq(&rq
->lock
);
6296 if (rq
->active_balance
) {
6297 active_load_balance(rq
, cpu
);
6298 rq
->active_balance
= 0;
6301 head
= &rq
->migration_queue
;
6303 if (list_empty(head
)) {
6304 spin_unlock_irq(&rq
->lock
);
6306 set_current_state(TASK_INTERRUPTIBLE
);
6309 req
= list_entry(head
->next
, struct migration_req
, list
);
6310 list_del_init(head
->next
);
6312 spin_unlock(&rq
->lock
);
6313 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
6316 complete(&req
->done
);
6318 __set_current_state(TASK_RUNNING
);
6322 /* Wait for kthread_stop */
6323 set_current_state(TASK_INTERRUPTIBLE
);
6324 while (!kthread_should_stop()) {
6326 set_current_state(TASK_INTERRUPTIBLE
);
6328 __set_current_state(TASK_RUNNING
);
6332 #ifdef CONFIG_HOTPLUG_CPU
6334 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6338 local_irq_disable();
6339 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
6345 * Figure out where task on dead CPU should go, use force if necessary.
6347 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
6350 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(dead_cpu
));
6353 /* Look for allowed, online CPU in same node. */
6354 for_each_cpu_and(dest_cpu
, nodemask
, cpu_online_mask
)
6355 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
6358 /* Any allowed, online CPU? */
6359 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_online_mask
);
6360 if (dest_cpu
< nr_cpu_ids
)
6363 /* No more Mr. Nice Guy. */
6364 if (dest_cpu
>= nr_cpu_ids
) {
6365 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
6366 dest_cpu
= cpumask_any_and(cpu_online_mask
, &p
->cpus_allowed
);
6369 * Don't tell them about moving exiting tasks or
6370 * kernel threads (both mm NULL), since they never
6373 if (p
->mm
&& printk_ratelimit()) {
6374 printk(KERN_INFO
"process %d (%s) no "
6375 "longer affine to cpu%d\n",
6376 task_pid_nr(p
), p
->comm
, dead_cpu
);
6381 /* It can have affinity changed while we were choosing. */
6382 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
6387 * While a dead CPU has no uninterruptible tasks queued at this point,
6388 * it might still have a nonzero ->nr_uninterruptible counter, because
6389 * for performance reasons the counter is not stricly tracking tasks to
6390 * their home CPUs. So we just add the counter to another CPU's counter,
6391 * to keep the global sum constant after CPU-down:
6393 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
6395 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_online_mask
));
6396 unsigned long flags
;
6398 local_irq_save(flags
);
6399 double_rq_lock(rq_src
, rq_dest
);
6400 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
6401 rq_src
->nr_uninterruptible
= 0;
6402 double_rq_unlock(rq_src
, rq_dest
);
6403 local_irq_restore(flags
);
6406 /* Run through task list and migrate tasks from the dead cpu. */
6407 static void migrate_live_tasks(int src_cpu
)
6409 struct task_struct
*p
, *t
;
6411 read_lock(&tasklist_lock
);
6413 do_each_thread(t
, p
) {
6417 if (task_cpu(p
) == src_cpu
)
6418 move_task_off_dead_cpu(src_cpu
, p
);
6419 } while_each_thread(t
, p
);
6421 read_unlock(&tasklist_lock
);
6425 * Schedules idle task to be the next runnable task on current CPU.
6426 * It does so by boosting its priority to highest possible.
6427 * Used by CPU offline code.
6429 void sched_idle_next(void)
6431 int this_cpu
= smp_processor_id();
6432 struct rq
*rq
= cpu_rq(this_cpu
);
6433 struct task_struct
*p
= rq
->idle
;
6434 unsigned long flags
;
6436 /* cpu has to be offline */
6437 BUG_ON(cpu_online(this_cpu
));
6440 * Strictly not necessary since rest of the CPUs are stopped by now
6441 * and interrupts disabled on the current cpu.
6443 spin_lock_irqsave(&rq
->lock
, flags
);
6445 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6447 update_rq_clock(rq
);
6448 activate_task(rq
, p
, 0);
6450 spin_unlock_irqrestore(&rq
->lock
, flags
);
6454 * Ensures that the idle task is using init_mm right before its cpu goes
6457 void idle_task_exit(void)
6459 struct mm_struct
*mm
= current
->active_mm
;
6461 BUG_ON(cpu_online(smp_processor_id()));
6464 switch_mm(mm
, &init_mm
, current
);
6468 /* called under rq->lock with disabled interrupts */
6469 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
6471 struct rq
*rq
= cpu_rq(dead_cpu
);
6473 /* Must be exiting, otherwise would be on tasklist. */
6474 BUG_ON(!p
->exit_state
);
6476 /* Cannot have done final schedule yet: would have vanished. */
6477 BUG_ON(p
->state
== TASK_DEAD
);
6482 * Drop lock around migration; if someone else moves it,
6483 * that's OK. No task can be added to this CPU, so iteration is
6486 spin_unlock_irq(&rq
->lock
);
6487 move_task_off_dead_cpu(dead_cpu
, p
);
6488 spin_lock_irq(&rq
->lock
);
6493 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6494 static void migrate_dead_tasks(unsigned int dead_cpu
)
6496 struct rq
*rq
= cpu_rq(dead_cpu
);
6497 struct task_struct
*next
;
6500 if (!rq
->nr_running
)
6502 update_rq_clock(rq
);
6503 next
= pick_next_task(rq
, rq
->curr
);
6506 next
->sched_class
->put_prev_task(rq
, next
);
6507 migrate_dead(dead_cpu
, next
);
6511 #endif /* CONFIG_HOTPLUG_CPU */
6513 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6515 static struct ctl_table sd_ctl_dir
[] = {
6517 .procname
= "sched_domain",
6523 static struct ctl_table sd_ctl_root
[] = {
6525 .ctl_name
= CTL_KERN
,
6526 .procname
= "kernel",
6528 .child
= sd_ctl_dir
,
6533 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6535 struct ctl_table
*entry
=
6536 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6541 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6543 struct ctl_table
*entry
;
6546 * In the intermediate directories, both the child directory and
6547 * procname are dynamically allocated and could fail but the mode
6548 * will always be set. In the lowest directory the names are
6549 * static strings and all have proc handlers.
6551 for (entry
= *tablep
; entry
->mode
; entry
++) {
6553 sd_free_ctl_entry(&entry
->child
);
6554 if (entry
->proc_handler
== NULL
)
6555 kfree(entry
->procname
);
6563 set_table_entry(struct ctl_table
*entry
,
6564 const char *procname
, void *data
, int maxlen
,
6565 mode_t mode
, proc_handler
*proc_handler
)
6567 entry
->procname
= procname
;
6569 entry
->maxlen
= maxlen
;
6571 entry
->proc_handler
= proc_handler
;
6574 static struct ctl_table
*
6575 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6577 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
6582 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6583 sizeof(long), 0644, proc_doulongvec_minmax
);
6584 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6585 sizeof(long), 0644, proc_doulongvec_minmax
);
6586 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6587 sizeof(int), 0644, proc_dointvec_minmax
);
6588 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6589 sizeof(int), 0644, proc_dointvec_minmax
);
6590 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6591 sizeof(int), 0644, proc_dointvec_minmax
);
6592 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6593 sizeof(int), 0644, proc_dointvec_minmax
);
6594 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6595 sizeof(int), 0644, proc_dointvec_minmax
);
6596 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6597 sizeof(int), 0644, proc_dointvec_minmax
);
6598 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6599 sizeof(int), 0644, proc_dointvec_minmax
);
6600 set_table_entry(&table
[9], "cache_nice_tries",
6601 &sd
->cache_nice_tries
,
6602 sizeof(int), 0644, proc_dointvec_minmax
);
6603 set_table_entry(&table
[10], "flags", &sd
->flags
,
6604 sizeof(int), 0644, proc_dointvec_minmax
);
6605 set_table_entry(&table
[11], "name", sd
->name
,
6606 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
6607 /* &table[12] is terminator */
6612 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6614 struct ctl_table
*entry
, *table
;
6615 struct sched_domain
*sd
;
6616 int domain_num
= 0, i
;
6619 for_each_domain(cpu
, sd
)
6621 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6626 for_each_domain(cpu
, sd
) {
6627 snprintf(buf
, 32, "domain%d", i
);
6628 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6630 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6637 static struct ctl_table_header
*sd_sysctl_header
;
6638 static void register_sched_domain_sysctl(void)
6640 int i
, cpu_num
= num_online_cpus();
6641 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6644 WARN_ON(sd_ctl_dir
[0].child
);
6645 sd_ctl_dir
[0].child
= entry
;
6650 for_each_online_cpu(i
) {
6651 snprintf(buf
, 32, "cpu%d", i
);
6652 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6654 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6658 WARN_ON(sd_sysctl_header
);
6659 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6662 /* may be called multiple times per register */
6663 static void unregister_sched_domain_sysctl(void)
6665 if (sd_sysctl_header
)
6666 unregister_sysctl_table(sd_sysctl_header
);
6667 sd_sysctl_header
= NULL
;
6668 if (sd_ctl_dir
[0].child
)
6669 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6672 static void register_sched_domain_sysctl(void)
6675 static void unregister_sched_domain_sysctl(void)
6680 static void set_rq_online(struct rq
*rq
)
6683 const struct sched_class
*class;
6685 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
6688 for_each_class(class) {
6689 if (class->rq_online
)
6690 class->rq_online(rq
);
6695 static void set_rq_offline(struct rq
*rq
)
6698 const struct sched_class
*class;
6700 for_each_class(class) {
6701 if (class->rq_offline
)
6702 class->rq_offline(rq
);
6705 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
6711 * migration_call - callback that gets triggered when a CPU is added.
6712 * Here we can start up the necessary migration thread for the new CPU.
6714 static int __cpuinit
6715 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6717 struct task_struct
*p
;
6718 int cpu
= (long)hcpu
;
6719 unsigned long flags
;
6724 case CPU_UP_PREPARE
:
6725 case CPU_UP_PREPARE_FROZEN
:
6726 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
6729 kthread_bind(p
, cpu
);
6730 /* Must be high prio: stop_machine expects to yield to it. */
6731 rq
= task_rq_lock(p
, &flags
);
6732 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6733 task_rq_unlock(rq
, &flags
);
6734 cpu_rq(cpu
)->migration_thread
= p
;
6738 case CPU_ONLINE_FROZEN
:
6739 /* Strictly unnecessary, as first user will wake it. */
6740 wake_up_process(cpu_rq(cpu
)->migration_thread
);
6742 /* Update our root-domain */
6744 spin_lock_irqsave(&rq
->lock
, flags
);
6746 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6750 spin_unlock_irqrestore(&rq
->lock
, flags
);
6753 #ifdef CONFIG_HOTPLUG_CPU
6754 case CPU_UP_CANCELED
:
6755 case CPU_UP_CANCELED_FROZEN
:
6756 if (!cpu_rq(cpu
)->migration_thread
)
6758 /* Unbind it from offline cpu so it can run. Fall thru. */
6759 kthread_bind(cpu_rq(cpu
)->migration_thread
,
6760 cpumask_any(cpu_online_mask
));
6761 kthread_stop(cpu_rq(cpu
)->migration_thread
);
6762 cpu_rq(cpu
)->migration_thread
= NULL
;
6766 case CPU_DEAD_FROZEN
:
6767 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6768 migrate_live_tasks(cpu
);
6770 kthread_stop(rq
->migration_thread
);
6771 rq
->migration_thread
= NULL
;
6772 /* Idle task back to normal (off runqueue, low prio) */
6773 spin_lock_irq(&rq
->lock
);
6774 update_rq_clock(rq
);
6775 deactivate_task(rq
, rq
->idle
, 0);
6776 rq
->idle
->static_prio
= MAX_PRIO
;
6777 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
6778 rq
->idle
->sched_class
= &idle_sched_class
;
6779 migrate_dead_tasks(cpu
);
6780 spin_unlock_irq(&rq
->lock
);
6782 migrate_nr_uninterruptible(rq
);
6783 BUG_ON(rq
->nr_running
!= 0);
6786 * No need to migrate the tasks: it was best-effort if
6787 * they didn't take sched_hotcpu_mutex. Just wake up
6790 spin_lock_irq(&rq
->lock
);
6791 while (!list_empty(&rq
->migration_queue
)) {
6792 struct migration_req
*req
;
6794 req
= list_entry(rq
->migration_queue
.next
,
6795 struct migration_req
, list
);
6796 list_del_init(&req
->list
);
6797 spin_unlock_irq(&rq
->lock
);
6798 complete(&req
->done
);
6799 spin_lock_irq(&rq
->lock
);
6801 spin_unlock_irq(&rq
->lock
);
6805 case CPU_DYING_FROZEN
:
6806 /* Update our root-domain */
6808 spin_lock_irqsave(&rq
->lock
, flags
);
6810 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6813 spin_unlock_irqrestore(&rq
->lock
, flags
);
6820 /* Register at highest priority so that task migration (migrate_all_tasks)
6821 * happens before everything else.
6823 static struct notifier_block __cpuinitdata migration_notifier
= {
6824 .notifier_call
= migration_call
,
6828 static int __init
migration_init(void)
6830 void *cpu
= (void *)(long)smp_processor_id();
6833 /* Start one for the boot CPU: */
6834 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6835 BUG_ON(err
== NOTIFY_BAD
);
6836 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6837 register_cpu_notifier(&migration_notifier
);
6841 early_initcall(migration_init
);
6846 #ifdef CONFIG_SCHED_DEBUG
6848 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6849 struct cpumask
*groupmask
)
6851 struct sched_group
*group
= sd
->groups
;
6854 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
6855 cpumask_clear(groupmask
);
6857 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6859 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6860 printk("does not load-balance\n");
6862 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6867 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
6869 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
6870 printk(KERN_ERR
"ERROR: domain->span does not contain "
6873 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
6874 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6878 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6882 printk(KERN_ERR
"ERROR: group is NULL\n");
6886 if (!group
->__cpu_power
) {
6887 printk(KERN_CONT
"\n");
6888 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6893 if (!cpumask_weight(sched_group_cpus(group
))) {
6894 printk(KERN_CONT
"\n");
6895 printk(KERN_ERR
"ERROR: empty group\n");
6899 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
6900 printk(KERN_CONT
"\n");
6901 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6905 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
6907 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
6908 printk(KERN_CONT
" %s", str
);
6910 group
= group
->next
;
6911 } while (group
!= sd
->groups
);
6912 printk(KERN_CONT
"\n");
6914 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
6915 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6918 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
6919 printk(KERN_ERR
"ERROR: parent span is not a superset "
6920 "of domain->span\n");
6924 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6926 cpumask_var_t groupmask
;
6930 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6934 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6936 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
6937 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6942 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6949 free_cpumask_var(groupmask
);
6951 #else /* !CONFIG_SCHED_DEBUG */
6952 # define sched_domain_debug(sd, cpu) do { } while (0)
6953 #endif /* CONFIG_SCHED_DEBUG */
6955 static int sd_degenerate(struct sched_domain
*sd
)
6957 if (cpumask_weight(sched_domain_span(sd
)) == 1)
6960 /* Following flags need at least 2 groups */
6961 if (sd
->flags
& (SD_LOAD_BALANCE
|
6962 SD_BALANCE_NEWIDLE
|
6966 SD_SHARE_PKG_RESOURCES
)) {
6967 if (sd
->groups
!= sd
->groups
->next
)
6971 /* Following flags don't use groups */
6972 if (sd
->flags
& (SD_WAKE_IDLE
|
6981 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6983 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6985 if (sd_degenerate(parent
))
6988 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
6991 /* Does parent contain flags not in child? */
6992 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6993 if (cflags
& SD_WAKE_AFFINE
)
6994 pflags
&= ~SD_WAKE_BALANCE
;
6995 /* Flags needing groups don't count if only 1 group in parent */
6996 if (parent
->groups
== parent
->groups
->next
) {
6997 pflags
&= ~(SD_LOAD_BALANCE
|
6998 SD_BALANCE_NEWIDLE
|
7002 SD_SHARE_PKG_RESOURCES
);
7003 if (nr_node_ids
== 1)
7004 pflags
&= ~SD_SERIALIZE
;
7006 if (~cflags
& pflags
)
7012 static void free_rootdomain(struct root_domain
*rd
)
7014 cpupri_cleanup(&rd
->cpupri
);
7016 free_cpumask_var(rd
->rto_mask
);
7017 free_cpumask_var(rd
->online
);
7018 free_cpumask_var(rd
->span
);
7022 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
7024 unsigned long flags
;
7026 spin_lock_irqsave(&rq
->lock
, flags
);
7029 struct root_domain
*old_rd
= rq
->rd
;
7031 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
7034 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
7036 if (atomic_dec_and_test(&old_rd
->refcount
))
7037 free_rootdomain(old_rd
);
7040 atomic_inc(&rd
->refcount
);
7043 cpumask_set_cpu(rq
->cpu
, rd
->span
);
7044 if (cpumask_test_cpu(rq
->cpu
, cpu_online_mask
))
7047 spin_unlock_irqrestore(&rq
->lock
, flags
);
7050 static int __init_refok
init_rootdomain(struct root_domain
*rd
, bool bootmem
)
7052 memset(rd
, 0, sizeof(*rd
));
7055 alloc_bootmem_cpumask_var(&def_root_domain
.span
);
7056 alloc_bootmem_cpumask_var(&def_root_domain
.online
);
7057 alloc_bootmem_cpumask_var(&def_root_domain
.rto_mask
);
7058 cpupri_init(&rd
->cpupri
, true);
7062 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
7064 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
7066 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
7069 if (cpupri_init(&rd
->cpupri
, false) != 0)
7074 free_cpumask_var(rd
->rto_mask
);
7076 free_cpumask_var(rd
->online
);
7078 free_cpumask_var(rd
->span
);
7083 static void init_defrootdomain(void)
7085 init_rootdomain(&def_root_domain
, true);
7087 atomic_set(&def_root_domain
.refcount
, 1);
7090 static struct root_domain
*alloc_rootdomain(void)
7092 struct root_domain
*rd
;
7094 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
7098 if (init_rootdomain(rd
, false) != 0) {
7107 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7108 * hold the hotplug lock.
7111 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
7113 struct rq
*rq
= cpu_rq(cpu
);
7114 struct sched_domain
*tmp
;
7116 /* Remove the sched domains which do not contribute to scheduling. */
7117 for (tmp
= sd
; tmp
; ) {
7118 struct sched_domain
*parent
= tmp
->parent
;
7122 if (sd_parent_degenerate(tmp
, parent
)) {
7123 tmp
->parent
= parent
->parent
;
7125 parent
->parent
->child
= tmp
;
7130 if (sd
&& sd_degenerate(sd
)) {
7136 sched_domain_debug(sd
, cpu
);
7138 rq_attach_root(rq
, rd
);
7139 rcu_assign_pointer(rq
->sd
, sd
);
7142 /* cpus with isolated domains */
7143 static cpumask_var_t cpu_isolated_map
;
7145 /* Setup the mask of cpus configured for isolated domains */
7146 static int __init
isolated_cpu_setup(char *str
)
7148 cpulist_parse(str
, cpu_isolated_map
);
7152 __setup("isolcpus=", isolated_cpu_setup
);
7155 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7156 * to a function which identifies what group(along with sched group) a CPU
7157 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7158 * (due to the fact that we keep track of groups covered with a struct cpumask).
7160 * init_sched_build_groups will build a circular linked list of the groups
7161 * covered by the given span, and will set each group's ->cpumask correctly,
7162 * and ->cpu_power to 0.
7165 init_sched_build_groups(const struct cpumask
*span
,
7166 const struct cpumask
*cpu_map
,
7167 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
7168 struct sched_group
**sg
,
7169 struct cpumask
*tmpmask
),
7170 struct cpumask
*covered
, struct cpumask
*tmpmask
)
7172 struct sched_group
*first
= NULL
, *last
= NULL
;
7175 cpumask_clear(covered
);
7177 for_each_cpu(i
, span
) {
7178 struct sched_group
*sg
;
7179 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
7182 if (cpumask_test_cpu(i
, covered
))
7185 cpumask_clear(sched_group_cpus(sg
));
7186 sg
->__cpu_power
= 0;
7188 for_each_cpu(j
, span
) {
7189 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
7192 cpumask_set_cpu(j
, covered
);
7193 cpumask_set_cpu(j
, sched_group_cpus(sg
));
7204 #define SD_NODES_PER_DOMAIN 16
7209 * find_next_best_node - find the next node to include in a sched_domain
7210 * @node: node whose sched_domain we're building
7211 * @used_nodes: nodes already in the sched_domain
7213 * Find the next node to include in a given scheduling domain. Simply
7214 * finds the closest node not already in the @used_nodes map.
7216 * Should use nodemask_t.
7218 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
7220 int i
, n
, val
, min_val
, best_node
= 0;
7224 for (i
= 0; i
< nr_node_ids
; i
++) {
7225 /* Start at @node */
7226 n
= (node
+ i
) % nr_node_ids
;
7228 if (!nr_cpus_node(n
))
7231 /* Skip already used nodes */
7232 if (node_isset(n
, *used_nodes
))
7235 /* Simple min distance search */
7236 val
= node_distance(node
, n
);
7238 if (val
< min_val
) {
7244 node_set(best_node
, *used_nodes
);
7249 * sched_domain_node_span - get a cpumask for a node's sched_domain
7250 * @node: node whose cpumask we're constructing
7251 * @span: resulting cpumask
7253 * Given a node, construct a good cpumask for its sched_domain to span. It
7254 * should be one that prevents unnecessary balancing, but also spreads tasks
7257 static void sched_domain_node_span(int node
, struct cpumask
*span
)
7259 nodemask_t used_nodes
;
7262 cpumask_clear(span
);
7263 nodes_clear(used_nodes
);
7265 cpumask_or(span
, span
, cpumask_of_node(node
));
7266 node_set(node
, used_nodes
);
7268 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
7269 int next_node
= find_next_best_node(node
, &used_nodes
);
7271 cpumask_or(span
, span
, cpumask_of_node(next_node
));
7274 #endif /* CONFIG_NUMA */
7276 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
7279 * The cpus mask in sched_group and sched_domain hangs off the end.
7280 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7281 * for nr_cpu_ids < CONFIG_NR_CPUS.
7283 struct static_sched_group
{
7284 struct sched_group sg
;
7285 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
7288 struct static_sched_domain
{
7289 struct sched_domain sd
;
7290 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
7294 * SMT sched-domains:
7296 #ifdef CONFIG_SCHED_SMT
7297 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
7298 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_cpus
);
7301 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
7302 struct sched_group
**sg
, struct cpumask
*unused
)
7305 *sg
= &per_cpu(sched_group_cpus
, cpu
).sg
;
7308 #endif /* CONFIG_SCHED_SMT */
7311 * multi-core sched-domains:
7313 #ifdef CONFIG_SCHED_MC
7314 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
7315 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
7316 #endif /* CONFIG_SCHED_MC */
7318 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7320 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
7321 struct sched_group
**sg
, struct cpumask
*mask
)
7325 cpumask_and(mask
, &per_cpu(cpu_sibling_map
, cpu
), cpu_map
);
7326 group
= cpumask_first(mask
);
7328 *sg
= &per_cpu(sched_group_core
, group
).sg
;
7331 #elif defined(CONFIG_SCHED_MC)
7333 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
7334 struct sched_group
**sg
, struct cpumask
*unused
)
7337 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
7342 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
7343 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
7346 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
7347 struct sched_group
**sg
, struct cpumask
*mask
)
7350 #ifdef CONFIG_SCHED_MC
7351 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
7352 group
= cpumask_first(mask
);
7353 #elif defined(CONFIG_SCHED_SMT)
7354 cpumask_and(mask
, &per_cpu(cpu_sibling_map
, cpu
), cpu_map
);
7355 group
= cpumask_first(mask
);
7360 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
7366 * The init_sched_build_groups can't handle what we want to do with node
7367 * groups, so roll our own. Now each node has its own list of groups which
7368 * gets dynamically allocated.
7370 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
7371 static struct sched_group
***sched_group_nodes_bycpu
;
7373 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
7374 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
7376 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
7377 struct sched_group
**sg
,
7378 struct cpumask
*nodemask
)
7382 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
7383 group
= cpumask_first(nodemask
);
7386 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
7390 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
7392 struct sched_group
*sg
= group_head
;
7398 for_each_cpu(j
, sched_group_cpus(sg
)) {
7399 struct sched_domain
*sd
;
7401 sd
= &per_cpu(phys_domains
, j
).sd
;
7402 if (j
!= cpumask_first(sched_group_cpus(sd
->groups
))) {
7404 * Only add "power" once for each
7410 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
7413 } while (sg
!= group_head
);
7415 #endif /* CONFIG_NUMA */
7418 /* Free memory allocated for various sched_group structures */
7419 static void free_sched_groups(const struct cpumask
*cpu_map
,
7420 struct cpumask
*nodemask
)
7424 for_each_cpu(cpu
, cpu_map
) {
7425 struct sched_group
**sched_group_nodes
7426 = sched_group_nodes_bycpu
[cpu
];
7428 if (!sched_group_nodes
)
7431 for (i
= 0; i
< nr_node_ids
; i
++) {
7432 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
7434 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
7435 if (cpumask_empty(nodemask
))
7445 if (oldsg
!= sched_group_nodes
[i
])
7448 kfree(sched_group_nodes
);
7449 sched_group_nodes_bycpu
[cpu
] = NULL
;
7452 #else /* !CONFIG_NUMA */
7453 static void free_sched_groups(const struct cpumask
*cpu_map
,
7454 struct cpumask
*nodemask
)
7457 #endif /* CONFIG_NUMA */
7460 * Initialize sched groups cpu_power.
7462 * cpu_power indicates the capacity of sched group, which is used while
7463 * distributing the load between different sched groups in a sched domain.
7464 * Typically cpu_power for all the groups in a sched domain will be same unless
7465 * there are asymmetries in the topology. If there are asymmetries, group
7466 * having more cpu_power will pickup more load compared to the group having
7469 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7470 * the maximum number of tasks a group can handle in the presence of other idle
7471 * or lightly loaded groups in the same sched domain.
7473 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7475 struct sched_domain
*child
;
7476 struct sched_group
*group
;
7478 WARN_ON(!sd
|| !sd
->groups
);
7480 if (cpu
!= cpumask_first(sched_group_cpus(sd
->groups
)))
7485 sd
->groups
->__cpu_power
= 0;
7488 * For perf policy, if the groups in child domain share resources
7489 * (for example cores sharing some portions of the cache hierarchy
7490 * or SMT), then set this domain groups cpu_power such that each group
7491 * can handle only one task, when there are other idle groups in the
7492 * same sched domain.
7494 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
7496 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
7497 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
7502 * add cpu_power of each child group to this groups cpu_power
7504 group
= child
->groups
;
7506 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
7507 group
= group
->next
;
7508 } while (group
!= child
->groups
);
7512 * Initializers for schedule domains
7513 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7516 #ifdef CONFIG_SCHED_DEBUG
7517 # define SD_INIT_NAME(sd, type) sd->name = #type
7519 # define SD_INIT_NAME(sd, type) do { } while (0)
7522 #define SD_INIT(sd, type) sd_init_##type(sd)
7524 #define SD_INIT_FUNC(type) \
7525 static noinline void sd_init_##type(struct sched_domain *sd) \
7527 memset(sd, 0, sizeof(*sd)); \
7528 *sd = SD_##type##_INIT; \
7529 sd->level = SD_LV_##type; \
7530 SD_INIT_NAME(sd, type); \
7535 SD_INIT_FUNC(ALLNODES
)
7538 #ifdef CONFIG_SCHED_SMT
7539 SD_INIT_FUNC(SIBLING
)
7541 #ifdef CONFIG_SCHED_MC
7545 static int default_relax_domain_level
= -1;
7547 static int __init
setup_relax_domain_level(char *str
)
7551 val
= simple_strtoul(str
, NULL
, 0);
7552 if (val
< SD_LV_MAX
)
7553 default_relax_domain_level
= val
;
7557 __setup("relax_domain_level=", setup_relax_domain_level
);
7559 static void set_domain_attribute(struct sched_domain
*sd
,
7560 struct sched_domain_attr
*attr
)
7564 if (!attr
|| attr
->relax_domain_level
< 0) {
7565 if (default_relax_domain_level
< 0)
7568 request
= default_relax_domain_level
;
7570 request
= attr
->relax_domain_level
;
7571 if (request
< sd
->level
) {
7572 /* turn off idle balance on this domain */
7573 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
7575 /* turn on idle balance on this domain */
7576 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
7581 * Build sched domains for a given set of cpus and attach the sched domains
7582 * to the individual cpus
7584 static int __build_sched_domains(const struct cpumask
*cpu_map
,
7585 struct sched_domain_attr
*attr
)
7587 int i
, err
= -ENOMEM
;
7588 struct root_domain
*rd
;
7589 cpumask_var_t nodemask
, this_sibling_map
, this_core_map
, send_covered
,
7592 cpumask_var_t domainspan
, covered
, notcovered
;
7593 struct sched_group
**sched_group_nodes
= NULL
;
7594 int sd_allnodes
= 0;
7596 if (!alloc_cpumask_var(&domainspan
, GFP_KERNEL
))
7598 if (!alloc_cpumask_var(&covered
, GFP_KERNEL
))
7599 goto free_domainspan
;
7600 if (!alloc_cpumask_var(¬covered
, GFP_KERNEL
))
7604 if (!alloc_cpumask_var(&nodemask
, GFP_KERNEL
))
7605 goto free_notcovered
;
7606 if (!alloc_cpumask_var(&this_sibling_map
, GFP_KERNEL
))
7608 if (!alloc_cpumask_var(&this_core_map
, GFP_KERNEL
))
7609 goto free_this_sibling_map
;
7610 if (!alloc_cpumask_var(&send_covered
, GFP_KERNEL
))
7611 goto free_this_core_map
;
7612 if (!alloc_cpumask_var(&tmpmask
, GFP_KERNEL
))
7613 goto free_send_covered
;
7617 * Allocate the per-node list of sched groups
7619 sched_group_nodes
= kcalloc(nr_node_ids
, sizeof(struct sched_group
*),
7621 if (!sched_group_nodes
) {
7622 printk(KERN_WARNING
"Can not alloc sched group node list\n");
7627 rd
= alloc_rootdomain();
7629 printk(KERN_WARNING
"Cannot alloc root domain\n");
7630 goto free_sched_groups
;
7634 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = sched_group_nodes
;
7638 * Set up domains for cpus specified by the cpu_map.
7640 for_each_cpu(i
, cpu_map
) {
7641 struct sched_domain
*sd
= NULL
, *p
;
7643 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(i
)), cpu_map
);
7646 if (cpumask_weight(cpu_map
) >
7647 SD_NODES_PER_DOMAIN
*cpumask_weight(nodemask
)) {
7648 sd
= &per_cpu(allnodes_domains
, i
).sd
;
7649 SD_INIT(sd
, ALLNODES
);
7650 set_domain_attribute(sd
, attr
);
7651 cpumask_copy(sched_domain_span(sd
), cpu_map
);
7652 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7658 sd
= &per_cpu(node_domains
, i
).sd
;
7660 set_domain_attribute(sd
, attr
);
7661 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
7665 cpumask_and(sched_domain_span(sd
),
7666 sched_domain_span(sd
), cpu_map
);
7670 sd
= &per_cpu(phys_domains
, i
).sd
;
7672 set_domain_attribute(sd
, attr
);
7673 cpumask_copy(sched_domain_span(sd
), nodemask
);
7677 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7679 #ifdef CONFIG_SCHED_MC
7681 sd
= &per_cpu(core_domains
, i
).sd
;
7683 set_domain_attribute(sd
, attr
);
7684 cpumask_and(sched_domain_span(sd
), cpu_map
,
7685 cpu_coregroup_mask(i
));
7688 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7691 #ifdef CONFIG_SCHED_SMT
7693 sd
= &per_cpu(cpu_domains
, i
).sd
;
7694 SD_INIT(sd
, SIBLING
);
7695 set_domain_attribute(sd
, attr
);
7696 cpumask_and(sched_domain_span(sd
),
7697 &per_cpu(cpu_sibling_map
, i
), cpu_map
);
7700 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7704 #ifdef CONFIG_SCHED_SMT
7705 /* Set up CPU (sibling) groups */
7706 for_each_cpu(i
, cpu_map
) {
7707 cpumask_and(this_sibling_map
,
7708 &per_cpu(cpu_sibling_map
, i
), cpu_map
);
7709 if (i
!= cpumask_first(this_sibling_map
))
7712 init_sched_build_groups(this_sibling_map
, cpu_map
,
7714 send_covered
, tmpmask
);
7718 #ifdef CONFIG_SCHED_MC
7719 /* Set up multi-core groups */
7720 for_each_cpu(i
, cpu_map
) {
7721 cpumask_and(this_core_map
, cpu_coregroup_mask(i
), cpu_map
);
7722 if (i
!= cpumask_first(this_core_map
))
7725 init_sched_build_groups(this_core_map
, cpu_map
,
7727 send_covered
, tmpmask
);
7731 /* Set up physical groups */
7732 for (i
= 0; i
< nr_node_ids
; i
++) {
7733 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
7734 if (cpumask_empty(nodemask
))
7737 init_sched_build_groups(nodemask
, cpu_map
,
7739 send_covered
, tmpmask
);
7743 /* Set up node groups */
7745 init_sched_build_groups(cpu_map
, cpu_map
,
7746 &cpu_to_allnodes_group
,
7747 send_covered
, tmpmask
);
7750 for (i
= 0; i
< nr_node_ids
; i
++) {
7751 /* Set up node groups */
7752 struct sched_group
*sg
, *prev
;
7755 cpumask_clear(covered
);
7756 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
7757 if (cpumask_empty(nodemask
)) {
7758 sched_group_nodes
[i
] = NULL
;
7762 sched_domain_node_span(i
, domainspan
);
7763 cpumask_and(domainspan
, domainspan
, cpu_map
);
7765 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
7768 printk(KERN_WARNING
"Can not alloc domain group for "
7772 sched_group_nodes
[i
] = sg
;
7773 for_each_cpu(j
, nodemask
) {
7774 struct sched_domain
*sd
;
7776 sd
= &per_cpu(node_domains
, j
).sd
;
7779 sg
->__cpu_power
= 0;
7780 cpumask_copy(sched_group_cpus(sg
), nodemask
);
7782 cpumask_or(covered
, covered
, nodemask
);
7785 for (j
= 0; j
< nr_node_ids
; j
++) {
7786 int n
= (i
+ j
) % nr_node_ids
;
7788 cpumask_complement(notcovered
, covered
);
7789 cpumask_and(tmpmask
, notcovered
, cpu_map
);
7790 cpumask_and(tmpmask
, tmpmask
, domainspan
);
7791 if (cpumask_empty(tmpmask
))
7794 cpumask_and(tmpmask
, tmpmask
, cpumask_of_node(n
));
7795 if (cpumask_empty(tmpmask
))
7798 sg
= kmalloc_node(sizeof(struct sched_group
) +
7803 "Can not alloc domain group for node %d\n", j
);
7806 sg
->__cpu_power
= 0;
7807 cpumask_copy(sched_group_cpus(sg
), tmpmask
);
7808 sg
->next
= prev
->next
;
7809 cpumask_or(covered
, covered
, tmpmask
);
7816 /* Calculate CPU power for physical packages and nodes */
7817 #ifdef CONFIG_SCHED_SMT
7818 for_each_cpu(i
, cpu_map
) {
7819 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
).sd
;
7821 init_sched_groups_power(i
, sd
);
7824 #ifdef CONFIG_SCHED_MC
7825 for_each_cpu(i
, cpu_map
) {
7826 struct sched_domain
*sd
= &per_cpu(core_domains
, i
).sd
;
7828 init_sched_groups_power(i
, sd
);
7832 for_each_cpu(i
, cpu_map
) {
7833 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
).sd
;
7835 init_sched_groups_power(i
, sd
);
7839 for (i
= 0; i
< nr_node_ids
; i
++)
7840 init_numa_sched_groups_power(sched_group_nodes
[i
]);
7843 struct sched_group
*sg
;
7845 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
7847 init_numa_sched_groups_power(sg
);
7851 /* Attach the domains */
7852 for_each_cpu(i
, cpu_map
) {
7853 struct sched_domain
*sd
;
7854 #ifdef CONFIG_SCHED_SMT
7855 sd
= &per_cpu(cpu_domains
, i
).sd
;
7856 #elif defined(CONFIG_SCHED_MC)
7857 sd
= &per_cpu(core_domains
, i
).sd
;
7859 sd
= &per_cpu(phys_domains
, i
).sd
;
7861 cpu_attach_domain(sd
, rd
, i
);
7867 free_cpumask_var(tmpmask
);
7869 free_cpumask_var(send_covered
);
7871 free_cpumask_var(this_core_map
);
7872 free_this_sibling_map
:
7873 free_cpumask_var(this_sibling_map
);
7875 free_cpumask_var(nodemask
);
7878 free_cpumask_var(notcovered
);
7880 free_cpumask_var(covered
);
7882 free_cpumask_var(domainspan
);
7889 kfree(sched_group_nodes
);
7895 free_sched_groups(cpu_map
, tmpmask
);
7896 free_rootdomain(rd
);
7901 static int build_sched_domains(const struct cpumask
*cpu_map
)
7903 return __build_sched_domains(cpu_map
, NULL
);
7906 static struct cpumask
*doms_cur
; /* current sched domains */
7907 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7908 static struct sched_domain_attr
*dattr_cur
;
7909 /* attribues of custom domains in 'doms_cur' */
7912 * Special case: If a kmalloc of a doms_cur partition (array of
7913 * cpumask) fails, then fallback to a single sched domain,
7914 * as determined by the single cpumask fallback_doms.
7916 static cpumask_var_t fallback_doms
;
7919 * arch_update_cpu_topology lets virtualized architectures update the
7920 * cpu core maps. It is supposed to return 1 if the topology changed
7921 * or 0 if it stayed the same.
7923 int __attribute__((weak
)) arch_update_cpu_topology(void)
7929 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7930 * For now this just excludes isolated cpus, but could be used to
7931 * exclude other special cases in the future.
7933 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
7937 arch_update_cpu_topology();
7939 doms_cur
= kmalloc(cpumask_size(), GFP_KERNEL
);
7941 doms_cur
= fallback_doms
;
7942 cpumask_andnot(doms_cur
, cpu_map
, cpu_isolated_map
);
7944 err
= build_sched_domains(doms_cur
);
7945 register_sched_domain_sysctl();
7950 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
7951 struct cpumask
*tmpmask
)
7953 free_sched_groups(cpu_map
, tmpmask
);
7957 * Detach sched domains from a group of cpus specified in cpu_map
7958 * These cpus will now be attached to the NULL domain
7960 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7962 /* Save because hotplug lock held. */
7963 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
7966 for_each_cpu(i
, cpu_map
)
7967 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7968 synchronize_sched();
7969 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
7972 /* handle null as "default" */
7973 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7974 struct sched_domain_attr
*new, int idx_new
)
7976 struct sched_domain_attr tmp
;
7983 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7984 new ? (new + idx_new
) : &tmp
,
7985 sizeof(struct sched_domain_attr
));
7989 * Partition sched domains as specified by the 'ndoms_new'
7990 * cpumasks in the array doms_new[] of cpumasks. This compares
7991 * doms_new[] to the current sched domain partitioning, doms_cur[].
7992 * It destroys each deleted domain and builds each new domain.
7994 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
7995 * The masks don't intersect (don't overlap.) We should setup one
7996 * sched domain for each mask. CPUs not in any of the cpumasks will
7997 * not be load balanced. If the same cpumask appears both in the
7998 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8001 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8002 * ownership of it and will kfree it when done with it. If the caller
8003 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8004 * ndoms_new == 1, and partition_sched_domains() will fallback to
8005 * the single partition 'fallback_doms', it also forces the domains
8008 * If doms_new == NULL it will be replaced with cpu_online_mask.
8009 * ndoms_new == 0 is a special case for destroying existing domains,
8010 * and it will not create the default domain.
8012 * Call with hotplug lock held
8014 /* FIXME: Change to struct cpumask *doms_new[] */
8015 void partition_sched_domains(int ndoms_new
, struct cpumask
*doms_new
,
8016 struct sched_domain_attr
*dattr_new
)
8021 mutex_lock(&sched_domains_mutex
);
8023 /* always unregister in case we don't destroy any domains */
8024 unregister_sched_domain_sysctl();
8026 /* Let architecture update cpu core mappings. */
8027 new_topology
= arch_update_cpu_topology();
8029 n
= doms_new
? ndoms_new
: 0;
8031 /* Destroy deleted domains */
8032 for (i
= 0; i
< ndoms_cur
; i
++) {
8033 for (j
= 0; j
< n
&& !new_topology
; j
++) {
8034 if (cpumask_equal(&doms_cur
[i
], &doms_new
[j
])
8035 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
8038 /* no match - a current sched domain not in new doms_new[] */
8039 detach_destroy_domains(doms_cur
+ i
);
8044 if (doms_new
== NULL
) {
8046 doms_new
= fallback_doms
;
8047 cpumask_andnot(&doms_new
[0], cpu_online_mask
, cpu_isolated_map
);
8048 WARN_ON_ONCE(dattr_new
);
8051 /* Build new domains */
8052 for (i
= 0; i
< ndoms_new
; i
++) {
8053 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
8054 if (cpumask_equal(&doms_new
[i
], &doms_cur
[j
])
8055 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
8058 /* no match - add a new doms_new */
8059 __build_sched_domains(doms_new
+ i
,
8060 dattr_new
? dattr_new
+ i
: NULL
);
8065 /* Remember the new sched domains */
8066 if (doms_cur
!= fallback_doms
)
8068 kfree(dattr_cur
); /* kfree(NULL) is safe */
8069 doms_cur
= doms_new
;
8070 dattr_cur
= dattr_new
;
8071 ndoms_cur
= ndoms_new
;
8073 register_sched_domain_sysctl();
8075 mutex_unlock(&sched_domains_mutex
);
8078 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8079 static void arch_reinit_sched_domains(void)
8083 /* Destroy domains first to force the rebuild */
8084 partition_sched_domains(0, NULL
, NULL
);
8086 rebuild_sched_domains();
8090 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
8092 unsigned int level
= 0;
8094 if (sscanf(buf
, "%u", &level
) != 1)
8098 * level is always be positive so don't check for
8099 * level < POWERSAVINGS_BALANCE_NONE which is 0
8100 * What happens on 0 or 1 byte write,
8101 * need to check for count as well?
8104 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
8108 sched_smt_power_savings
= level
;
8110 sched_mc_power_savings
= level
;
8112 arch_reinit_sched_domains();
8117 #ifdef CONFIG_SCHED_MC
8118 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
8121 return sprintf(page
, "%u\n", sched_mc_power_savings
);
8123 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
8124 const char *buf
, size_t count
)
8126 return sched_power_savings_store(buf
, count
, 0);
8128 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
8129 sched_mc_power_savings_show
,
8130 sched_mc_power_savings_store
);
8133 #ifdef CONFIG_SCHED_SMT
8134 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
8137 return sprintf(page
, "%u\n", sched_smt_power_savings
);
8139 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
8140 const char *buf
, size_t count
)
8142 return sched_power_savings_store(buf
, count
, 1);
8144 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
8145 sched_smt_power_savings_show
,
8146 sched_smt_power_savings_store
);
8149 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
8153 #ifdef CONFIG_SCHED_SMT
8155 err
= sysfs_create_file(&cls
->kset
.kobj
,
8156 &attr_sched_smt_power_savings
.attr
);
8158 #ifdef CONFIG_SCHED_MC
8159 if (!err
&& mc_capable())
8160 err
= sysfs_create_file(&cls
->kset
.kobj
,
8161 &attr_sched_mc_power_savings
.attr
);
8165 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8167 #ifndef CONFIG_CPUSETS
8169 * Add online and remove offline CPUs from the scheduler domains.
8170 * When cpusets are enabled they take over this function.
8172 static int update_sched_domains(struct notifier_block
*nfb
,
8173 unsigned long action
, void *hcpu
)
8177 case CPU_ONLINE_FROZEN
:
8179 case CPU_DEAD_FROZEN
:
8180 partition_sched_domains(1, NULL
, NULL
);
8189 static int update_runtime(struct notifier_block
*nfb
,
8190 unsigned long action
, void *hcpu
)
8192 int cpu
= (int)(long)hcpu
;
8195 case CPU_DOWN_PREPARE
:
8196 case CPU_DOWN_PREPARE_FROZEN
:
8197 disable_runtime(cpu_rq(cpu
));
8200 case CPU_DOWN_FAILED
:
8201 case CPU_DOWN_FAILED_FROZEN
:
8203 case CPU_ONLINE_FROZEN
:
8204 enable_runtime(cpu_rq(cpu
));
8212 void __init
sched_init_smp(void)
8214 cpumask_var_t non_isolated_cpus
;
8216 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
8218 #if defined(CONFIG_NUMA)
8219 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
8221 BUG_ON(sched_group_nodes_bycpu
== NULL
);
8224 mutex_lock(&sched_domains_mutex
);
8225 arch_init_sched_domains(cpu_online_mask
);
8226 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
8227 if (cpumask_empty(non_isolated_cpus
))
8228 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
8229 mutex_unlock(&sched_domains_mutex
);
8232 #ifndef CONFIG_CPUSETS
8233 /* XXX: Theoretical race here - CPU may be hotplugged now */
8234 hotcpu_notifier(update_sched_domains
, 0);
8237 /* RT runtime code needs to handle some hotplug events */
8238 hotcpu_notifier(update_runtime
, 0);
8242 /* Move init over to a non-isolated CPU */
8243 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
8245 sched_init_granularity();
8246 free_cpumask_var(non_isolated_cpus
);
8248 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
8249 init_sched_rt_class();
8252 void __init
sched_init_smp(void)
8254 sched_init_granularity();
8256 #endif /* CONFIG_SMP */
8258 int in_sched_functions(unsigned long addr
)
8260 return in_lock_functions(addr
) ||
8261 (addr
>= (unsigned long)__sched_text_start
8262 && addr
< (unsigned long)__sched_text_end
);
8265 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
8267 cfs_rq
->tasks_timeline
= RB_ROOT
;
8268 INIT_LIST_HEAD(&cfs_rq
->tasks
);
8269 #ifdef CONFIG_FAIR_GROUP_SCHED
8272 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
8275 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
8277 struct rt_prio_array
*array
;
8280 array
= &rt_rq
->active
;
8281 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
8282 INIT_LIST_HEAD(array
->queue
+ i
);
8283 __clear_bit(i
, array
->bitmap
);
8285 /* delimiter for bitsearch: */
8286 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
8288 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8289 rt_rq
->highest_prio
= MAX_RT_PRIO
;
8292 rt_rq
->rt_nr_migratory
= 0;
8293 rt_rq
->overloaded
= 0;
8297 rt_rq
->rt_throttled
= 0;
8298 rt_rq
->rt_runtime
= 0;
8299 spin_lock_init(&rt_rq
->rt_runtime_lock
);
8301 #ifdef CONFIG_RT_GROUP_SCHED
8302 rt_rq
->rt_nr_boosted
= 0;
8307 #ifdef CONFIG_FAIR_GROUP_SCHED
8308 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8309 struct sched_entity
*se
, int cpu
, int add
,
8310 struct sched_entity
*parent
)
8312 struct rq
*rq
= cpu_rq(cpu
);
8313 tg
->cfs_rq
[cpu
] = cfs_rq
;
8314 init_cfs_rq(cfs_rq
, rq
);
8317 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
8320 /* se could be NULL for init_task_group */
8325 se
->cfs_rq
= &rq
->cfs
;
8327 se
->cfs_rq
= parent
->my_q
;
8330 se
->load
.weight
= tg
->shares
;
8331 se
->load
.inv_weight
= 0;
8332 se
->parent
= parent
;
8336 #ifdef CONFIG_RT_GROUP_SCHED
8337 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
8338 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
8339 struct sched_rt_entity
*parent
)
8341 struct rq
*rq
= cpu_rq(cpu
);
8343 tg
->rt_rq
[cpu
] = rt_rq
;
8344 init_rt_rq(rt_rq
, rq
);
8346 rt_rq
->rt_se
= rt_se
;
8347 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8349 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
8351 tg
->rt_se
[cpu
] = rt_se
;
8356 rt_se
->rt_rq
= &rq
->rt
;
8358 rt_se
->rt_rq
= parent
->my_q
;
8360 rt_se
->my_q
= rt_rq
;
8361 rt_se
->parent
= parent
;
8362 INIT_LIST_HEAD(&rt_se
->run_list
);
8366 void __init
sched_init(void)
8369 unsigned long alloc_size
= 0, ptr
;
8371 #ifdef CONFIG_FAIR_GROUP_SCHED
8372 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8374 #ifdef CONFIG_RT_GROUP_SCHED
8375 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8377 #ifdef CONFIG_USER_SCHED
8381 * As sched_init() is called before page_alloc is setup,
8382 * we use alloc_bootmem().
8385 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
8387 #ifdef CONFIG_FAIR_GROUP_SCHED
8388 init_task_group
.se
= (struct sched_entity
**)ptr
;
8389 ptr
+= nr_cpu_ids
* sizeof(void **);
8391 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8392 ptr
+= nr_cpu_ids
* sizeof(void **);
8394 #ifdef CONFIG_USER_SCHED
8395 root_task_group
.se
= (struct sched_entity
**)ptr
;
8396 ptr
+= nr_cpu_ids
* sizeof(void **);
8398 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8399 ptr
+= nr_cpu_ids
* sizeof(void **);
8400 #endif /* CONFIG_USER_SCHED */
8401 #endif /* CONFIG_FAIR_GROUP_SCHED */
8402 #ifdef CONFIG_RT_GROUP_SCHED
8403 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8404 ptr
+= nr_cpu_ids
* sizeof(void **);
8406 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8407 ptr
+= nr_cpu_ids
* sizeof(void **);
8409 #ifdef CONFIG_USER_SCHED
8410 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8411 ptr
+= nr_cpu_ids
* sizeof(void **);
8413 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8414 ptr
+= nr_cpu_ids
* sizeof(void **);
8415 #endif /* CONFIG_USER_SCHED */
8416 #endif /* CONFIG_RT_GROUP_SCHED */
8420 init_defrootdomain();
8423 init_rt_bandwidth(&def_rt_bandwidth
,
8424 global_rt_period(), global_rt_runtime());
8426 #ifdef CONFIG_RT_GROUP_SCHED
8427 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
8428 global_rt_period(), global_rt_runtime());
8429 #ifdef CONFIG_USER_SCHED
8430 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8431 global_rt_period(), RUNTIME_INF
);
8432 #endif /* CONFIG_USER_SCHED */
8433 #endif /* CONFIG_RT_GROUP_SCHED */
8435 #ifdef CONFIG_GROUP_SCHED
8436 list_add(&init_task_group
.list
, &task_groups
);
8437 INIT_LIST_HEAD(&init_task_group
.children
);
8439 #ifdef CONFIG_USER_SCHED
8440 INIT_LIST_HEAD(&root_task_group
.children
);
8441 init_task_group
.parent
= &root_task_group
;
8442 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
8443 #endif /* CONFIG_USER_SCHED */
8444 #endif /* CONFIG_GROUP_SCHED */
8446 for_each_possible_cpu(i
) {
8450 spin_lock_init(&rq
->lock
);
8452 init_cfs_rq(&rq
->cfs
, rq
);
8453 init_rt_rq(&rq
->rt
, rq
);
8454 #ifdef CONFIG_FAIR_GROUP_SCHED
8455 init_task_group
.shares
= init_task_group_load
;
8456 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8457 #ifdef CONFIG_CGROUP_SCHED
8459 * How much cpu bandwidth does init_task_group get?
8461 * In case of task-groups formed thr' the cgroup filesystem, it
8462 * gets 100% of the cpu resources in the system. This overall
8463 * system cpu resource is divided among the tasks of
8464 * init_task_group and its child task-groups in a fair manner,
8465 * based on each entity's (task or task-group's) weight
8466 * (se->load.weight).
8468 * In other words, if init_task_group has 10 tasks of weight
8469 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8470 * then A0's share of the cpu resource is:
8472 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8474 * We achieve this by letting init_task_group's tasks sit
8475 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8477 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
8478 #elif defined CONFIG_USER_SCHED
8479 root_task_group
.shares
= NICE_0_LOAD
;
8480 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
8482 * In case of task-groups formed thr' the user id of tasks,
8483 * init_task_group represents tasks belonging to root user.
8484 * Hence it forms a sibling of all subsequent groups formed.
8485 * In this case, init_task_group gets only a fraction of overall
8486 * system cpu resource, based on the weight assigned to root
8487 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8488 * by letting tasks of init_task_group sit in a separate cfs_rq
8489 * (init_cfs_rq) and having one entity represent this group of
8490 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8492 init_tg_cfs_entry(&init_task_group
,
8493 &per_cpu(init_cfs_rq
, i
),
8494 &per_cpu(init_sched_entity
, i
), i
, 1,
8495 root_task_group
.se
[i
]);
8498 #endif /* CONFIG_FAIR_GROUP_SCHED */
8500 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8501 #ifdef CONFIG_RT_GROUP_SCHED
8502 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
8503 #ifdef CONFIG_CGROUP_SCHED
8504 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
8505 #elif defined CONFIG_USER_SCHED
8506 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
8507 init_tg_rt_entry(&init_task_group
,
8508 &per_cpu(init_rt_rq
, i
),
8509 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
8510 root_task_group
.rt_se
[i
]);
8514 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
8515 rq
->cpu_load
[j
] = 0;
8519 rq
->active_balance
= 0;
8520 rq
->next_balance
= jiffies
;
8524 rq
->migration_thread
= NULL
;
8525 INIT_LIST_HEAD(&rq
->migration_queue
);
8526 rq_attach_root(rq
, &def_root_domain
);
8529 atomic_set(&rq
->nr_iowait
, 0);
8532 set_load_weight(&init_task
);
8534 #ifdef CONFIG_PREEMPT_NOTIFIERS
8535 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
8539 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8542 #ifdef CONFIG_RT_MUTEXES
8543 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
8547 * The boot idle thread does lazy MMU switching as well:
8549 atomic_inc(&init_mm
.mm_count
);
8550 enter_lazy_tlb(&init_mm
, current
);
8553 * Make us the idle thread. Technically, schedule() should not be
8554 * called from this thread, however somewhere below it might be,
8555 * but because we are the idle thread, we just pick up running again
8556 * when this runqueue becomes "idle".
8558 init_idle(current
, smp_processor_id());
8560 * During early bootup we pretend to be a normal task:
8562 current
->sched_class
= &fair_sched_class
;
8564 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8565 alloc_bootmem_cpumask_var(&nohz_cpu_mask
);
8568 alloc_bootmem_cpumask_var(&nohz
.cpu_mask
);
8570 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
8573 scheduler_running
= 1;
8576 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8577 void __might_sleep(char *file
, int line
)
8580 static unsigned long prev_jiffy
; /* ratelimiting */
8582 if ((!in_atomic() && !irqs_disabled()) ||
8583 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
8585 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8587 prev_jiffy
= jiffies
;
8590 "BUG: sleeping function called from invalid context at %s:%d\n",
8593 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8594 in_atomic(), irqs_disabled(),
8595 current
->pid
, current
->comm
);
8597 debug_show_held_locks(current
);
8598 if (irqs_disabled())
8599 print_irqtrace_events(current
);
8603 EXPORT_SYMBOL(__might_sleep
);
8606 #ifdef CONFIG_MAGIC_SYSRQ
8607 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8611 update_rq_clock(rq
);
8612 on_rq
= p
->se
.on_rq
;
8614 deactivate_task(rq
, p
, 0);
8615 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8617 activate_task(rq
, p
, 0);
8618 resched_task(rq
->curr
);
8622 void normalize_rt_tasks(void)
8624 struct task_struct
*g
, *p
;
8625 unsigned long flags
;
8628 read_lock_irqsave(&tasklist_lock
, flags
);
8629 do_each_thread(g
, p
) {
8631 * Only normalize user tasks:
8636 p
->se
.exec_start
= 0;
8637 #ifdef CONFIG_SCHEDSTATS
8638 p
->se
.wait_start
= 0;
8639 p
->se
.sleep_start
= 0;
8640 p
->se
.block_start
= 0;
8645 * Renice negative nice level userspace
8648 if (TASK_NICE(p
) < 0 && p
->mm
)
8649 set_user_nice(p
, 0);
8653 spin_lock(&p
->pi_lock
);
8654 rq
= __task_rq_lock(p
);
8656 normalize_task(rq
, p
);
8658 __task_rq_unlock(rq
);
8659 spin_unlock(&p
->pi_lock
);
8660 } while_each_thread(g
, p
);
8662 read_unlock_irqrestore(&tasklist_lock
, flags
);
8665 #endif /* CONFIG_MAGIC_SYSRQ */
8669 * These functions are only useful for the IA64 MCA handling.
8671 * They can only be called when the whole system has been
8672 * stopped - every CPU needs to be quiescent, and no scheduling
8673 * activity can take place. Using them for anything else would
8674 * be a serious bug, and as a result, they aren't even visible
8675 * under any other configuration.
8679 * curr_task - return the current task for a given cpu.
8680 * @cpu: the processor in question.
8682 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8684 struct task_struct
*curr_task(int cpu
)
8686 return cpu_curr(cpu
);
8690 * set_curr_task - set the current task for a given cpu.
8691 * @cpu: the processor in question.
8692 * @p: the task pointer to set.
8694 * Description: This function must only be used when non-maskable interrupts
8695 * are serviced on a separate stack. It allows the architecture to switch the
8696 * notion of the current task on a cpu in a non-blocking manner. This function
8697 * must be called with all CPU's synchronized, and interrupts disabled, the
8698 * and caller must save the original value of the current task (see
8699 * curr_task() above) and restore that value before reenabling interrupts and
8700 * re-starting the system.
8702 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8704 void set_curr_task(int cpu
, struct task_struct
*p
)
8711 #ifdef CONFIG_FAIR_GROUP_SCHED
8712 static void free_fair_sched_group(struct task_group
*tg
)
8716 for_each_possible_cpu(i
) {
8718 kfree(tg
->cfs_rq
[i
]);
8728 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8730 struct cfs_rq
*cfs_rq
;
8731 struct sched_entity
*se
;
8735 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8738 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8742 tg
->shares
= NICE_0_LOAD
;
8744 for_each_possible_cpu(i
) {
8747 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
8748 GFP_KERNEL
, cpu_to_node(i
));
8752 se
= kzalloc_node(sizeof(struct sched_entity
),
8753 GFP_KERNEL
, cpu_to_node(i
));
8757 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
8766 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8768 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8769 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8772 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8774 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8776 #else /* !CONFG_FAIR_GROUP_SCHED */
8777 static inline void free_fair_sched_group(struct task_group
*tg
)
8782 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8787 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8791 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8794 #endif /* CONFIG_FAIR_GROUP_SCHED */
8796 #ifdef CONFIG_RT_GROUP_SCHED
8797 static void free_rt_sched_group(struct task_group
*tg
)
8801 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8803 for_each_possible_cpu(i
) {
8805 kfree(tg
->rt_rq
[i
]);
8807 kfree(tg
->rt_se
[i
]);
8815 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8817 struct rt_rq
*rt_rq
;
8818 struct sched_rt_entity
*rt_se
;
8822 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8825 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8829 init_rt_bandwidth(&tg
->rt_bandwidth
,
8830 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8832 for_each_possible_cpu(i
) {
8835 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
8836 GFP_KERNEL
, cpu_to_node(i
));
8840 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
8841 GFP_KERNEL
, cpu_to_node(i
));
8845 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
8854 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8856 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8857 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8860 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8862 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8864 #else /* !CONFIG_RT_GROUP_SCHED */
8865 static inline void free_rt_sched_group(struct task_group
*tg
)
8870 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8875 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8879 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8882 #endif /* CONFIG_RT_GROUP_SCHED */
8884 #ifdef CONFIG_GROUP_SCHED
8885 static void free_sched_group(struct task_group
*tg
)
8887 free_fair_sched_group(tg
);
8888 free_rt_sched_group(tg
);
8892 /* allocate runqueue etc for a new task group */
8893 struct task_group
*sched_create_group(struct task_group
*parent
)
8895 struct task_group
*tg
;
8896 unsigned long flags
;
8899 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8901 return ERR_PTR(-ENOMEM
);
8903 if (!alloc_fair_sched_group(tg
, parent
))
8906 if (!alloc_rt_sched_group(tg
, parent
))
8909 spin_lock_irqsave(&task_group_lock
, flags
);
8910 for_each_possible_cpu(i
) {
8911 register_fair_sched_group(tg
, i
);
8912 register_rt_sched_group(tg
, i
);
8914 list_add_rcu(&tg
->list
, &task_groups
);
8916 WARN_ON(!parent
); /* root should already exist */
8918 tg
->parent
= parent
;
8919 INIT_LIST_HEAD(&tg
->children
);
8920 list_add_rcu(&tg
->siblings
, &parent
->children
);
8921 spin_unlock_irqrestore(&task_group_lock
, flags
);
8926 free_sched_group(tg
);
8927 return ERR_PTR(-ENOMEM
);
8930 /* rcu callback to free various structures associated with a task group */
8931 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8933 /* now it should be safe to free those cfs_rqs */
8934 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8937 /* Destroy runqueue etc associated with a task group */
8938 void sched_destroy_group(struct task_group
*tg
)
8940 unsigned long flags
;
8943 spin_lock_irqsave(&task_group_lock
, flags
);
8944 for_each_possible_cpu(i
) {
8945 unregister_fair_sched_group(tg
, i
);
8946 unregister_rt_sched_group(tg
, i
);
8948 list_del_rcu(&tg
->list
);
8949 list_del_rcu(&tg
->siblings
);
8950 spin_unlock_irqrestore(&task_group_lock
, flags
);
8952 /* wait for possible concurrent references to cfs_rqs complete */
8953 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8956 /* change task's runqueue when it moves between groups.
8957 * The caller of this function should have put the task in its new group
8958 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8959 * reflect its new group.
8961 void sched_move_task(struct task_struct
*tsk
)
8964 unsigned long flags
;
8967 rq
= task_rq_lock(tsk
, &flags
);
8969 update_rq_clock(rq
);
8971 running
= task_current(rq
, tsk
);
8972 on_rq
= tsk
->se
.on_rq
;
8975 dequeue_task(rq
, tsk
, 0);
8976 if (unlikely(running
))
8977 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8979 set_task_rq(tsk
, task_cpu(tsk
));
8981 #ifdef CONFIG_FAIR_GROUP_SCHED
8982 if (tsk
->sched_class
->moved_group
)
8983 tsk
->sched_class
->moved_group(tsk
);
8986 if (unlikely(running
))
8987 tsk
->sched_class
->set_curr_task(rq
);
8989 enqueue_task(rq
, tsk
, 0);
8991 task_rq_unlock(rq
, &flags
);
8993 #endif /* CONFIG_GROUP_SCHED */
8995 #ifdef CONFIG_FAIR_GROUP_SCHED
8996 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8998 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
9003 dequeue_entity(cfs_rq
, se
, 0);
9005 se
->load
.weight
= shares
;
9006 se
->load
.inv_weight
= 0;
9009 enqueue_entity(cfs_rq
, se
, 0);
9012 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
9014 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
9015 struct rq
*rq
= cfs_rq
->rq
;
9016 unsigned long flags
;
9018 spin_lock_irqsave(&rq
->lock
, flags
);
9019 __set_se_shares(se
, shares
);
9020 spin_unlock_irqrestore(&rq
->lock
, flags
);
9023 static DEFINE_MUTEX(shares_mutex
);
9025 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
9028 unsigned long flags
;
9031 * We can't change the weight of the root cgroup.
9036 if (shares
< MIN_SHARES
)
9037 shares
= MIN_SHARES
;
9038 else if (shares
> MAX_SHARES
)
9039 shares
= MAX_SHARES
;
9041 mutex_lock(&shares_mutex
);
9042 if (tg
->shares
== shares
)
9045 spin_lock_irqsave(&task_group_lock
, flags
);
9046 for_each_possible_cpu(i
)
9047 unregister_fair_sched_group(tg
, i
);
9048 list_del_rcu(&tg
->siblings
);
9049 spin_unlock_irqrestore(&task_group_lock
, flags
);
9051 /* wait for any ongoing reference to this group to finish */
9052 synchronize_sched();
9055 * Now we are free to modify the group's share on each cpu
9056 * w/o tripping rebalance_share or load_balance_fair.
9058 tg
->shares
= shares
;
9059 for_each_possible_cpu(i
) {
9063 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
9064 set_se_shares(tg
->se
[i
], shares
);
9068 * Enable load balance activity on this group, by inserting it back on
9069 * each cpu's rq->leaf_cfs_rq_list.
9071 spin_lock_irqsave(&task_group_lock
, flags
);
9072 for_each_possible_cpu(i
)
9073 register_fair_sched_group(tg
, i
);
9074 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
9075 spin_unlock_irqrestore(&task_group_lock
, flags
);
9077 mutex_unlock(&shares_mutex
);
9081 unsigned long sched_group_shares(struct task_group
*tg
)
9087 #ifdef CONFIG_RT_GROUP_SCHED
9089 * Ensure that the real time constraints are schedulable.
9091 static DEFINE_MUTEX(rt_constraints_mutex
);
9093 static unsigned long to_ratio(u64 period
, u64 runtime
)
9095 if (runtime
== RUNTIME_INF
)
9098 return div64_u64(runtime
<< 20, period
);
9101 /* Must be called with tasklist_lock held */
9102 static inline int tg_has_rt_tasks(struct task_group
*tg
)
9104 struct task_struct
*g
, *p
;
9106 do_each_thread(g
, p
) {
9107 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
9109 } while_each_thread(g
, p
);
9114 struct rt_schedulable_data
{
9115 struct task_group
*tg
;
9120 static int tg_schedulable(struct task_group
*tg
, void *data
)
9122 struct rt_schedulable_data
*d
= data
;
9123 struct task_group
*child
;
9124 unsigned long total
, sum
= 0;
9125 u64 period
, runtime
;
9127 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9128 runtime
= tg
->rt_bandwidth
.rt_runtime
;
9131 period
= d
->rt_period
;
9132 runtime
= d
->rt_runtime
;
9135 #ifdef CONFIG_USER_SCHED
9136 if (tg
== &root_task_group
) {
9137 period
= global_rt_period();
9138 runtime
= global_rt_runtime();
9143 * Cannot have more runtime than the period.
9145 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9149 * Ensure we don't starve existing RT tasks.
9151 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
9154 total
= to_ratio(period
, runtime
);
9157 * Nobody can have more than the global setting allows.
9159 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
9163 * The sum of our children's runtime should not exceed our own.
9165 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
9166 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
9167 runtime
= child
->rt_bandwidth
.rt_runtime
;
9169 if (child
== d
->tg
) {
9170 period
= d
->rt_period
;
9171 runtime
= d
->rt_runtime
;
9174 sum
+= to_ratio(period
, runtime
);
9183 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
9185 struct rt_schedulable_data data
= {
9187 .rt_period
= period
,
9188 .rt_runtime
= runtime
,
9191 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
9194 static int tg_set_bandwidth(struct task_group
*tg
,
9195 u64 rt_period
, u64 rt_runtime
)
9199 mutex_lock(&rt_constraints_mutex
);
9200 read_lock(&tasklist_lock
);
9201 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
9205 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
9206 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
9207 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
9209 for_each_possible_cpu(i
) {
9210 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
9212 spin_lock(&rt_rq
->rt_runtime_lock
);
9213 rt_rq
->rt_runtime
= rt_runtime
;
9214 spin_unlock(&rt_rq
->rt_runtime_lock
);
9216 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
9218 read_unlock(&tasklist_lock
);
9219 mutex_unlock(&rt_constraints_mutex
);
9224 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
9226 u64 rt_runtime
, rt_period
;
9228 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9229 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
9230 if (rt_runtime_us
< 0)
9231 rt_runtime
= RUNTIME_INF
;
9233 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9236 long sched_group_rt_runtime(struct task_group
*tg
)
9240 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
9243 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
9244 do_div(rt_runtime_us
, NSEC_PER_USEC
);
9245 return rt_runtime_us
;
9248 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
9250 u64 rt_runtime
, rt_period
;
9252 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
9253 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9258 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9261 long sched_group_rt_period(struct task_group
*tg
)
9265 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9266 do_div(rt_period_us
, NSEC_PER_USEC
);
9267 return rt_period_us
;
9270 static int sched_rt_global_constraints(void)
9272 u64 runtime
, period
;
9275 if (sysctl_sched_rt_period
<= 0)
9278 runtime
= global_rt_runtime();
9279 period
= global_rt_period();
9282 * Sanity check on the sysctl variables.
9284 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9287 mutex_lock(&rt_constraints_mutex
);
9288 read_lock(&tasklist_lock
);
9289 ret
= __rt_schedulable(NULL
, 0, 0);
9290 read_unlock(&tasklist_lock
);
9291 mutex_unlock(&rt_constraints_mutex
);
9295 #else /* !CONFIG_RT_GROUP_SCHED */
9296 static int sched_rt_global_constraints(void)
9298 unsigned long flags
;
9301 if (sysctl_sched_rt_period
<= 0)
9304 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9305 for_each_possible_cpu(i
) {
9306 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
9308 spin_lock(&rt_rq
->rt_runtime_lock
);
9309 rt_rq
->rt_runtime
= global_rt_runtime();
9310 spin_unlock(&rt_rq
->rt_runtime_lock
);
9312 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9316 #endif /* CONFIG_RT_GROUP_SCHED */
9318 int sched_rt_handler(struct ctl_table
*table
, int write
,
9319 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
9323 int old_period
, old_runtime
;
9324 static DEFINE_MUTEX(mutex
);
9327 old_period
= sysctl_sched_rt_period
;
9328 old_runtime
= sysctl_sched_rt_runtime
;
9330 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
9332 if (!ret
&& write
) {
9333 ret
= sched_rt_global_constraints();
9335 sysctl_sched_rt_period
= old_period
;
9336 sysctl_sched_rt_runtime
= old_runtime
;
9338 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
9339 def_rt_bandwidth
.rt_period
=
9340 ns_to_ktime(global_rt_period());
9343 mutex_unlock(&mutex
);
9348 #ifdef CONFIG_CGROUP_SCHED
9350 /* return corresponding task_group object of a cgroup */
9351 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
9353 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
9354 struct task_group
, css
);
9357 static struct cgroup_subsys_state
*
9358 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9360 struct task_group
*tg
, *parent
;
9362 if (!cgrp
->parent
) {
9363 /* This is early initialization for the top cgroup */
9364 return &init_task_group
.css
;
9367 parent
= cgroup_tg(cgrp
->parent
);
9368 tg
= sched_create_group(parent
);
9370 return ERR_PTR(-ENOMEM
);
9376 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9378 struct task_group
*tg
= cgroup_tg(cgrp
);
9380 sched_destroy_group(tg
);
9384 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9385 struct task_struct
*tsk
)
9387 #ifdef CONFIG_RT_GROUP_SCHED
9388 /* Don't accept realtime tasks when there is no way for them to run */
9389 if (rt_task(tsk
) && cgroup_tg(cgrp
)->rt_bandwidth
.rt_runtime
== 0)
9392 /* We don't support RT-tasks being in separate groups */
9393 if (tsk
->sched_class
!= &fair_sched_class
)
9401 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9402 struct cgroup
*old_cont
, struct task_struct
*tsk
)
9404 sched_move_task(tsk
);
9407 #ifdef CONFIG_FAIR_GROUP_SCHED
9408 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9411 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
9414 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9416 struct task_group
*tg
= cgroup_tg(cgrp
);
9418 return (u64
) tg
->shares
;
9420 #endif /* CONFIG_FAIR_GROUP_SCHED */
9422 #ifdef CONFIG_RT_GROUP_SCHED
9423 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9426 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9429 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9431 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9434 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9437 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9440 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9442 return sched_group_rt_period(cgroup_tg(cgrp
));
9444 #endif /* CONFIG_RT_GROUP_SCHED */
9446 static struct cftype cpu_files
[] = {
9447 #ifdef CONFIG_FAIR_GROUP_SCHED
9450 .read_u64
= cpu_shares_read_u64
,
9451 .write_u64
= cpu_shares_write_u64
,
9454 #ifdef CONFIG_RT_GROUP_SCHED
9456 .name
= "rt_runtime_us",
9457 .read_s64
= cpu_rt_runtime_read
,
9458 .write_s64
= cpu_rt_runtime_write
,
9461 .name
= "rt_period_us",
9462 .read_u64
= cpu_rt_period_read_uint
,
9463 .write_u64
= cpu_rt_period_write_uint
,
9468 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
9470 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
9473 struct cgroup_subsys cpu_cgroup_subsys
= {
9475 .create
= cpu_cgroup_create
,
9476 .destroy
= cpu_cgroup_destroy
,
9477 .can_attach
= cpu_cgroup_can_attach
,
9478 .attach
= cpu_cgroup_attach
,
9479 .populate
= cpu_cgroup_populate
,
9480 .subsys_id
= cpu_cgroup_subsys_id
,
9484 #endif /* CONFIG_CGROUP_SCHED */
9486 #ifdef CONFIG_CGROUP_CPUACCT
9489 * CPU accounting code for task groups.
9491 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9492 * (balbir@in.ibm.com).
9495 /* track cpu usage of a group of tasks and its child groups */
9497 struct cgroup_subsys_state css
;
9498 /* cpuusage holds pointer to a u64-type object on every cpu */
9500 struct cpuacct
*parent
;
9503 struct cgroup_subsys cpuacct_subsys
;
9505 /* return cpu accounting group corresponding to this container */
9506 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
9508 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
9509 struct cpuacct
, css
);
9512 /* return cpu accounting group to which this task belongs */
9513 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
9515 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
9516 struct cpuacct
, css
);
9519 /* create a new cpu accounting group */
9520 static struct cgroup_subsys_state
*cpuacct_create(
9521 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9523 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
9526 return ERR_PTR(-ENOMEM
);
9528 ca
->cpuusage
= alloc_percpu(u64
);
9529 if (!ca
->cpuusage
) {
9531 return ERR_PTR(-ENOMEM
);
9535 ca
->parent
= cgroup_ca(cgrp
->parent
);
9540 /* destroy an existing cpu accounting group */
9542 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9544 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9546 free_percpu(ca
->cpuusage
);
9550 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
9552 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, cpu
);
9555 #ifndef CONFIG_64BIT
9557 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9559 spin_lock_irq(&cpu_rq(cpu
)->lock
);
9561 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9569 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
9571 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, cpu
);
9573 #ifndef CONFIG_64BIT
9575 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9577 spin_lock_irq(&cpu_rq(cpu
)->lock
);
9579 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9585 /* return total cpu usage (in nanoseconds) of a group */
9586 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9588 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9589 u64 totalcpuusage
= 0;
9592 for_each_present_cpu(i
)
9593 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
9595 return totalcpuusage
;
9598 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9601 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9610 for_each_present_cpu(i
)
9611 cpuacct_cpuusage_write(ca
, i
, 0);
9617 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
9620 struct cpuacct
*ca
= cgroup_ca(cgroup
);
9624 for_each_present_cpu(i
) {
9625 percpu
= cpuacct_cpuusage_read(ca
, i
);
9626 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
9628 seq_printf(m
, "\n");
9632 static struct cftype files
[] = {
9635 .read_u64
= cpuusage_read
,
9636 .write_u64
= cpuusage_write
,
9639 .name
= "usage_percpu",
9640 .read_seq_string
= cpuacct_percpu_seq_read
,
9645 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9647 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9651 * charge this task's execution time to its accounting group.
9653 * called with rq->lock held.
9655 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9660 if (!cpuacct_subsys
.active
)
9663 cpu
= task_cpu(tsk
);
9666 for (; ca
; ca
= ca
->parent
) {
9667 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, cpu
);
9668 *cpuusage
+= cputime
;
9672 struct cgroup_subsys cpuacct_subsys
= {
9674 .create
= cpuacct_create
,
9675 .destroy
= cpuacct_destroy
,
9676 .populate
= cpuacct_populate
,
9677 .subsys_id
= cpuacct_subsys_id
,
9679 #endif /* CONFIG_CGROUP_CPUACCT */