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
;
565 #ifdef CONFIG_FAIR_GROUP_SCHED
566 /* list of leaf cfs_rq on this cpu: */
567 struct list_head leaf_cfs_rq_list
;
569 #ifdef CONFIG_RT_GROUP_SCHED
570 struct list_head leaf_rt_rq_list
;
574 * This is part of a global counter where only the total sum
575 * over all CPUs matters. A task can increase this counter on
576 * one CPU and if it got migrated afterwards it may decrease
577 * it on another CPU. Always updated under the runqueue lock:
579 unsigned long nr_uninterruptible
;
581 struct task_struct
*curr
, *idle
;
582 unsigned long next_balance
;
583 struct mm_struct
*prev_mm
;
590 struct root_domain
*rd
;
591 struct sched_domain
*sd
;
593 /* For active balancing */
596 /* cpu of this runqueue: */
600 unsigned long avg_load_per_task
;
602 struct task_struct
*migration_thread
;
603 struct list_head migration_queue
;
606 #ifdef CONFIG_SCHED_HRTICK
608 int hrtick_csd_pending
;
609 struct call_single_data hrtick_csd
;
611 struct hrtimer hrtick_timer
;
614 #ifdef CONFIG_SCHEDSTATS
616 struct sched_info rq_sched_info
;
617 unsigned long long rq_cpu_time
;
618 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
620 /* sys_sched_yield() stats */
621 unsigned int yld_exp_empty
;
622 unsigned int yld_act_empty
;
623 unsigned int yld_both_empty
;
624 unsigned int yld_count
;
626 /* schedule() stats */
627 unsigned int sched_switch
;
628 unsigned int sched_count
;
629 unsigned int sched_goidle
;
631 /* try_to_wake_up() stats */
632 unsigned int ttwu_count
;
633 unsigned int ttwu_local
;
636 unsigned int bkl_count
;
640 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
642 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int sync
)
644 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, sync
);
647 static inline int cpu_of(struct rq
*rq
)
657 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
658 * See detach_destroy_domains: synchronize_sched for details.
660 * The domain tree of any CPU may only be accessed from within
661 * preempt-disabled sections.
663 #define for_each_domain(cpu, __sd) \
664 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
666 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
667 #define this_rq() (&__get_cpu_var(runqueues))
668 #define task_rq(p) cpu_rq(task_cpu(p))
669 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
671 inline void update_rq_clock(struct rq
*rq
)
673 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
677 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
679 #ifdef CONFIG_SCHED_DEBUG
680 # define const_debug __read_mostly
682 # define const_debug static const
688 * Returns true if the current cpu runqueue is locked.
689 * This interface allows printk to be called with the runqueue lock
690 * held and know whether or not it is OK to wake up the klogd.
692 int runqueue_is_locked(void)
695 struct rq
*rq
= cpu_rq(cpu
);
698 ret
= spin_is_locked(&rq
->lock
);
704 * Debugging: various feature bits
707 #define SCHED_FEAT(name, enabled) \
708 __SCHED_FEAT_##name ,
711 #include "sched_features.h"
716 #define SCHED_FEAT(name, enabled) \
717 (1UL << __SCHED_FEAT_##name) * enabled |
719 const_debug
unsigned int sysctl_sched_features
=
720 #include "sched_features.h"
725 #ifdef CONFIG_SCHED_DEBUG
726 #define SCHED_FEAT(name, enabled) \
729 static __read_mostly
char *sched_feat_names
[] = {
730 #include "sched_features.h"
736 static int sched_feat_show(struct seq_file
*m
, void *v
)
740 for (i
= 0; sched_feat_names
[i
]; i
++) {
741 if (!(sysctl_sched_features
& (1UL << i
)))
743 seq_printf(m
, "%s ", sched_feat_names
[i
]);
751 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
752 size_t cnt
, loff_t
*ppos
)
762 if (copy_from_user(&buf
, ubuf
, cnt
))
767 if (strncmp(buf
, "NO_", 3) == 0) {
772 for (i
= 0; sched_feat_names
[i
]; i
++) {
773 int len
= strlen(sched_feat_names
[i
]);
775 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
777 sysctl_sched_features
&= ~(1UL << i
);
779 sysctl_sched_features
|= (1UL << i
);
784 if (!sched_feat_names
[i
])
792 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
794 return single_open(filp
, sched_feat_show
, NULL
);
797 static struct file_operations sched_feat_fops
= {
798 .open
= sched_feat_open
,
799 .write
= sched_feat_write
,
802 .release
= single_release
,
805 static __init
int sched_init_debug(void)
807 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
812 late_initcall(sched_init_debug
);
816 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
819 * Number of tasks to iterate in a single balance run.
820 * Limited because this is done with IRQs disabled.
822 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
825 * ratelimit for updating the group shares.
828 unsigned int sysctl_sched_shares_ratelimit
= 250000;
831 * Inject some fuzzyness into changing the per-cpu group shares
832 * this avoids remote rq-locks at the expense of fairness.
835 unsigned int sysctl_sched_shares_thresh
= 4;
838 * period over which we measure -rt task cpu usage in us.
841 unsigned int sysctl_sched_rt_period
= 1000000;
843 static __read_mostly
int scheduler_running
;
846 * part of the period that we allow rt tasks to run in us.
849 int sysctl_sched_rt_runtime
= 950000;
851 static inline u64
global_rt_period(void)
853 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
856 static inline u64
global_rt_runtime(void)
858 if (sysctl_sched_rt_runtime
< 0)
861 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
864 #ifndef prepare_arch_switch
865 # define prepare_arch_switch(next) do { } while (0)
867 #ifndef finish_arch_switch
868 # define finish_arch_switch(prev) do { } while (0)
871 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
873 return rq
->curr
== p
;
876 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
877 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
879 return task_current(rq
, p
);
882 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
886 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
888 #ifdef CONFIG_DEBUG_SPINLOCK
889 /* this is a valid case when another task releases the spinlock */
890 rq
->lock
.owner
= current
;
893 * If we are tracking spinlock dependencies then we have to
894 * fix up the runqueue lock - which gets 'carried over' from
897 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
899 spin_unlock_irq(&rq
->lock
);
902 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
903 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
908 return task_current(rq
, p
);
912 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
916 * We can optimise this out completely for !SMP, because the
917 * SMP rebalancing from interrupt is the only thing that cares
922 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
923 spin_unlock_irq(&rq
->lock
);
925 spin_unlock(&rq
->lock
);
929 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
933 * After ->oncpu is cleared, the task can be moved to a different CPU.
934 * We must ensure this doesn't happen until the switch is completely
940 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
944 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
947 * __task_rq_lock - lock the runqueue a given task resides on.
948 * Must be called interrupts disabled.
950 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
954 struct rq
*rq
= task_rq(p
);
955 spin_lock(&rq
->lock
);
956 if (likely(rq
== task_rq(p
)))
958 spin_unlock(&rq
->lock
);
963 * task_rq_lock - lock the runqueue a given task resides on and disable
964 * interrupts. Note the ordering: we can safely lookup the task_rq without
965 * explicitly disabling preemption.
967 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
973 local_irq_save(*flags
);
975 spin_lock(&rq
->lock
);
976 if (likely(rq
== task_rq(p
)))
978 spin_unlock_irqrestore(&rq
->lock
, *flags
);
982 void curr_rq_lock_irq_save(unsigned long *flags
)
987 local_irq_save(*flags
);
988 rq
= cpu_rq(smp_processor_id());
989 spin_lock(&rq
->lock
);
992 void curr_rq_unlock_irq_restore(unsigned long *flags
)
997 rq
= cpu_rq(smp_processor_id());
998 spin_unlock(&rq
->lock
);
999 local_irq_restore(*flags
);
1002 void task_rq_unlock_wait(struct task_struct
*p
)
1004 struct rq
*rq
= task_rq(p
);
1006 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1007 spin_unlock_wait(&rq
->lock
);
1010 static void __task_rq_unlock(struct rq
*rq
)
1011 __releases(rq
->lock
)
1013 spin_unlock(&rq
->lock
);
1016 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
1017 __releases(rq
->lock
)
1019 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1023 * this_rq_lock - lock this runqueue and disable interrupts.
1025 static struct rq
*this_rq_lock(void)
1026 __acquires(rq
->lock
)
1030 local_irq_disable();
1032 spin_lock(&rq
->lock
);
1037 #ifdef CONFIG_SCHED_HRTICK
1039 * Use HR-timers to deliver accurate preemption points.
1041 * Its all a bit involved since we cannot program an hrt while holding the
1042 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1045 * When we get rescheduled we reprogram the hrtick_timer outside of the
1051 * - enabled by features
1052 * - hrtimer is actually high res
1054 static inline int hrtick_enabled(struct rq
*rq
)
1056 if (!sched_feat(HRTICK
))
1058 if (!cpu_active(cpu_of(rq
)))
1060 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1063 static void hrtick_clear(struct rq
*rq
)
1065 if (hrtimer_active(&rq
->hrtick_timer
))
1066 hrtimer_cancel(&rq
->hrtick_timer
);
1070 * High-resolution timer tick.
1071 * Runs from hardirq context with interrupts disabled.
1073 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1075 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1077 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1079 spin_lock(&rq
->lock
);
1080 update_rq_clock(rq
);
1081 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1082 spin_unlock(&rq
->lock
);
1084 return HRTIMER_NORESTART
;
1089 * called from hardirq (IPI) context
1091 static void __hrtick_start(void *arg
)
1093 struct rq
*rq
= arg
;
1095 spin_lock(&rq
->lock
);
1096 hrtimer_restart(&rq
->hrtick_timer
);
1097 rq
->hrtick_csd_pending
= 0;
1098 spin_unlock(&rq
->lock
);
1102 * Called to set the hrtick timer state.
1104 * called with rq->lock held and irqs disabled
1106 static void hrtick_start(struct rq
*rq
, u64 delay
)
1108 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1109 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1111 hrtimer_set_expires(timer
, time
);
1113 if (rq
== this_rq()) {
1114 hrtimer_restart(timer
);
1115 } else if (!rq
->hrtick_csd_pending
) {
1116 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
);
1117 rq
->hrtick_csd_pending
= 1;
1122 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1124 int cpu
= (int)(long)hcpu
;
1127 case CPU_UP_CANCELED
:
1128 case CPU_UP_CANCELED_FROZEN
:
1129 case CPU_DOWN_PREPARE
:
1130 case CPU_DOWN_PREPARE_FROZEN
:
1132 case CPU_DEAD_FROZEN
:
1133 hrtick_clear(cpu_rq(cpu
));
1140 static __init
void init_hrtick(void)
1142 hotcpu_notifier(hotplug_hrtick
, 0);
1146 * Called to set the hrtick timer state.
1148 * called with rq->lock held and irqs disabled
1150 static void hrtick_start(struct rq
*rq
, u64 delay
)
1152 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
), HRTIMER_MODE_REL
);
1155 static inline void init_hrtick(void)
1158 #endif /* CONFIG_SMP */
1160 static void init_rq_hrtick(struct rq
*rq
)
1163 rq
->hrtick_csd_pending
= 0;
1165 rq
->hrtick_csd
.flags
= 0;
1166 rq
->hrtick_csd
.func
= __hrtick_start
;
1167 rq
->hrtick_csd
.info
= rq
;
1170 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1171 rq
->hrtick_timer
.function
= hrtick
;
1173 #else /* CONFIG_SCHED_HRTICK */
1174 static inline void hrtick_clear(struct rq
*rq
)
1178 static inline void init_rq_hrtick(struct rq
*rq
)
1182 static inline void init_hrtick(void)
1185 #endif /* CONFIG_SCHED_HRTICK */
1188 * resched_task - mark a task 'to be rescheduled now'.
1190 * On UP this means the setting of the need_resched flag, on SMP it
1191 * might also involve a cross-CPU call to trigger the scheduler on
1196 #ifndef tsk_is_polling
1197 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1200 static void resched_task(struct task_struct
*p
)
1204 assert_spin_locked(&task_rq(p
)->lock
);
1206 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
1209 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
1212 if (cpu
== smp_processor_id())
1215 /* NEED_RESCHED must be visible before we test polling */
1217 if (!tsk_is_polling(p
))
1218 smp_send_reschedule(cpu
);
1221 static void resched_cpu(int cpu
)
1223 struct rq
*rq
= cpu_rq(cpu
);
1224 unsigned long flags
;
1226 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1228 resched_task(cpu_curr(cpu
));
1229 spin_unlock_irqrestore(&rq
->lock
, flags
);
1234 * When add_timer_on() enqueues a timer into the timer wheel of an
1235 * idle CPU then this timer might expire before the next timer event
1236 * which is scheduled to wake up that CPU. In case of a completely
1237 * idle system the next event might even be infinite time into the
1238 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1239 * leaves the inner idle loop so the newly added timer is taken into
1240 * account when the CPU goes back to idle and evaluates the timer
1241 * wheel for the next timer event.
1243 void wake_up_idle_cpu(int cpu
)
1245 struct rq
*rq
= cpu_rq(cpu
);
1247 if (cpu
== smp_processor_id())
1251 * This is safe, as this function is called with the timer
1252 * wheel base lock of (cpu) held. When the CPU is on the way
1253 * to idle and has not yet set rq->curr to idle then it will
1254 * be serialized on the timer wheel base lock and take the new
1255 * timer into account automatically.
1257 if (rq
->curr
!= rq
->idle
)
1261 * We can set TIF_RESCHED on the idle task of the other CPU
1262 * lockless. The worst case is that the other CPU runs the
1263 * idle task through an additional NOOP schedule()
1265 set_tsk_thread_flag(rq
->idle
, TIF_NEED_RESCHED
);
1267 /* NEED_RESCHED must be visible before we test polling */
1269 if (!tsk_is_polling(rq
->idle
))
1270 smp_send_reschedule(cpu
);
1272 #endif /* CONFIG_NO_HZ */
1274 #else /* !CONFIG_SMP */
1275 static void resched_task(struct task_struct
*p
)
1277 assert_spin_locked(&task_rq(p
)->lock
);
1278 set_tsk_need_resched(p
);
1280 #endif /* CONFIG_SMP */
1282 #if BITS_PER_LONG == 32
1283 # define WMULT_CONST (~0UL)
1285 # define WMULT_CONST (1UL << 32)
1288 #define WMULT_SHIFT 32
1291 * Shift right and round:
1293 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1296 * delta *= weight / lw
1298 static unsigned long
1299 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1300 struct load_weight
*lw
)
1304 if (!lw
->inv_weight
) {
1305 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1308 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1312 tmp
= (u64
)delta_exec
* weight
;
1314 * Check whether we'd overflow the 64-bit multiplication:
1316 if (unlikely(tmp
> WMULT_CONST
))
1317 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1320 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1322 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1325 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1331 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1338 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1339 * of tasks with abnormal "nice" values across CPUs the contribution that
1340 * each task makes to its run queue's load is weighted according to its
1341 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1342 * scaled version of the new time slice allocation that they receive on time
1346 #define WEIGHT_IDLEPRIO 3
1347 #define WMULT_IDLEPRIO 1431655765
1350 * Nice levels are multiplicative, with a gentle 10% change for every
1351 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1352 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1353 * that remained on nice 0.
1355 * The "10% effect" is relative and cumulative: from _any_ nice level,
1356 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1357 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1358 * If a task goes up by ~10% and another task goes down by ~10% then
1359 * the relative distance between them is ~25%.)
1361 static const int prio_to_weight
[40] = {
1362 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1363 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1364 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1365 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1366 /* 0 */ 1024, 820, 655, 526, 423,
1367 /* 5 */ 335, 272, 215, 172, 137,
1368 /* 10 */ 110, 87, 70, 56, 45,
1369 /* 15 */ 36, 29, 23, 18, 15,
1373 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1375 * In cases where the weight does not change often, we can use the
1376 * precalculated inverse to speed up arithmetics by turning divisions
1377 * into multiplications:
1379 static const u32 prio_to_wmult
[40] = {
1380 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1381 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1382 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1383 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1384 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1385 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1386 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1387 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1390 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1393 * runqueue iterator, to support SMP load-balancing between different
1394 * scheduling classes, without having to expose their internal data
1395 * structures to the load-balancing proper:
1397 struct rq_iterator
{
1399 struct task_struct
*(*start
)(void *);
1400 struct task_struct
*(*next
)(void *);
1404 static unsigned long
1405 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1406 unsigned long max_load_move
, struct sched_domain
*sd
,
1407 enum cpu_idle_type idle
, int *all_pinned
,
1408 int *this_best_prio
, struct rq_iterator
*iterator
);
1411 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1412 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1413 struct rq_iterator
*iterator
);
1416 #ifdef CONFIG_CGROUP_CPUACCT
1417 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1419 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1422 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1424 update_load_add(&rq
->load
, load
);
1427 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1429 update_load_sub(&rq
->load
, load
);
1432 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1433 typedef int (*tg_visitor
)(struct task_group
*, void *);
1436 * Iterate the full tree, calling @down when first entering a node and @up when
1437 * leaving it for the final time.
1439 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1441 struct task_group
*parent
, *child
;
1445 parent
= &root_task_group
;
1447 ret
= (*down
)(parent
, data
);
1450 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1457 ret
= (*up
)(parent
, data
);
1462 parent
= parent
->parent
;
1471 static int tg_nop(struct task_group
*tg
, void *data
)
1478 static unsigned long source_load(int cpu
, int type
);
1479 static unsigned long target_load(int cpu
, int type
);
1480 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1482 static unsigned long cpu_avg_load_per_task(int cpu
)
1484 struct rq
*rq
= cpu_rq(cpu
);
1485 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1488 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1490 rq
->avg_load_per_task
= 0;
1492 return rq
->avg_load_per_task
;
1495 #ifdef CONFIG_FAIR_GROUP_SCHED
1497 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1500 * Calculate and set the cpu's group shares.
1503 update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1504 unsigned long sd_shares
, unsigned long sd_rq_weight
)
1506 unsigned long shares
;
1507 unsigned long rq_weight
;
1512 rq_weight
= tg
->cfs_rq
[cpu
]->rq_weight
;
1515 * \Sum shares * rq_weight
1516 * shares = -----------------------
1520 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1521 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1523 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1524 sysctl_sched_shares_thresh
) {
1525 struct rq
*rq
= cpu_rq(cpu
);
1526 unsigned long flags
;
1528 spin_lock_irqsave(&rq
->lock
, flags
);
1529 tg
->cfs_rq
[cpu
]->shares
= shares
;
1531 __set_se_shares(tg
->se
[cpu
], shares
);
1532 spin_unlock_irqrestore(&rq
->lock
, flags
);
1537 * Re-compute the task group their per cpu shares over the given domain.
1538 * This needs to be done in a bottom-up fashion because the rq weight of a
1539 * parent group depends on the shares of its child groups.
1541 static int tg_shares_up(struct task_group
*tg
, void *data
)
1543 unsigned long weight
, rq_weight
= 0;
1544 unsigned long shares
= 0;
1545 struct sched_domain
*sd
= data
;
1548 for_each_cpu(i
, sched_domain_span(sd
)) {
1550 * If there are currently no tasks on the cpu pretend there
1551 * is one of average load so that when a new task gets to
1552 * run here it will not get delayed by group starvation.
1554 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1556 weight
= NICE_0_LOAD
;
1558 tg
->cfs_rq
[i
]->rq_weight
= weight
;
1559 rq_weight
+= weight
;
1560 shares
+= tg
->cfs_rq
[i
]->shares
;
1563 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1564 shares
= tg
->shares
;
1566 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1567 shares
= tg
->shares
;
1569 for_each_cpu(i
, sched_domain_span(sd
))
1570 update_group_shares_cpu(tg
, i
, shares
, rq_weight
);
1576 * Compute the cpu's hierarchical load factor for each task group.
1577 * This needs to be done in a top-down fashion because the load of a child
1578 * group is a fraction of its parents load.
1580 static int tg_load_down(struct task_group
*tg
, void *data
)
1583 long cpu
= (long)data
;
1586 load
= cpu_rq(cpu
)->load
.weight
;
1588 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1589 load
*= tg
->cfs_rq
[cpu
]->shares
;
1590 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1593 tg
->cfs_rq
[cpu
]->h_load
= load
;
1598 static void update_shares(struct sched_domain
*sd
)
1600 u64 now
= cpu_clock(raw_smp_processor_id());
1601 s64 elapsed
= now
- sd
->last_update
;
1603 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1604 sd
->last_update
= now
;
1605 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1609 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1611 spin_unlock(&rq
->lock
);
1613 spin_lock(&rq
->lock
);
1616 static void update_h_load(long cpu
)
1618 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1623 static inline void update_shares(struct sched_domain
*sd
)
1627 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1634 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1636 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1637 __releases(this_rq
->lock
)
1638 __acquires(busiest
->lock
)
1639 __acquires(this_rq
->lock
)
1643 if (unlikely(!irqs_disabled())) {
1644 /* printk() doesn't work good under rq->lock */
1645 spin_unlock(&this_rq
->lock
);
1648 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1649 if (busiest
< this_rq
) {
1650 spin_unlock(&this_rq
->lock
);
1651 spin_lock(&busiest
->lock
);
1652 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
1655 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
1660 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1661 __releases(busiest
->lock
)
1663 spin_unlock(&busiest
->lock
);
1664 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1668 #ifdef CONFIG_FAIR_GROUP_SCHED
1669 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1672 cfs_rq
->shares
= shares
;
1677 #include "sched_stats.h"
1678 #include "sched_idletask.c"
1679 #include "sched_fair.c"
1680 #include "sched_rt.c"
1681 #ifdef CONFIG_SCHED_DEBUG
1682 # include "sched_debug.c"
1685 #define sched_class_highest (&rt_sched_class)
1686 #define for_each_class(class) \
1687 for (class = sched_class_highest; class; class = class->next)
1689 static void inc_nr_running(struct rq
*rq
)
1694 static void dec_nr_running(struct rq
*rq
)
1699 static void set_load_weight(struct task_struct
*p
)
1701 if (task_has_rt_policy(p
)) {
1702 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1703 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1708 * SCHED_IDLE tasks get minimal weight:
1710 if (p
->policy
== SCHED_IDLE
) {
1711 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1712 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1716 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1717 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1720 static void update_avg(u64
*avg
, u64 sample
)
1722 s64 diff
= sample
- *avg
;
1726 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1728 sched_info_queued(p
);
1729 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1733 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1735 if (sleep
&& p
->se
.last_wakeup
) {
1736 update_avg(&p
->se
.avg_overlap
,
1737 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1738 p
->se
.last_wakeup
= 0;
1741 sched_info_dequeued(p
);
1742 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1747 * __normal_prio - return the priority that is based on the static prio
1749 static inline int __normal_prio(struct task_struct
*p
)
1751 return p
->static_prio
;
1755 * Calculate the expected normal priority: i.e. priority
1756 * without taking RT-inheritance into account. Might be
1757 * boosted by interactivity modifiers. Changes upon fork,
1758 * setprio syscalls, and whenever the interactivity
1759 * estimator recalculates.
1761 static inline int normal_prio(struct task_struct
*p
)
1765 if (task_has_rt_policy(p
))
1766 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1768 prio
= __normal_prio(p
);
1773 * Calculate the current priority, i.e. the priority
1774 * taken into account by the scheduler. This value might
1775 * be boosted by RT tasks, or might be boosted by
1776 * interactivity modifiers. Will be RT if the task got
1777 * RT-boosted. If not then it returns p->normal_prio.
1779 static int effective_prio(struct task_struct
*p
)
1781 p
->normal_prio
= normal_prio(p
);
1783 * If we are RT tasks or we were boosted to RT priority,
1784 * keep the priority unchanged. Otherwise, update priority
1785 * to the normal priority:
1787 if (!rt_prio(p
->prio
))
1788 return p
->normal_prio
;
1793 * activate_task - move a task to the runqueue.
1795 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1797 if (task_contributes_to_load(p
))
1798 rq
->nr_uninterruptible
--;
1800 enqueue_task(rq
, p
, wakeup
);
1805 * deactivate_task - remove a task from the runqueue.
1807 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1809 if (task_contributes_to_load(p
))
1810 rq
->nr_uninterruptible
++;
1812 dequeue_task(rq
, p
, sleep
);
1817 * task_curr - is this task currently executing on a CPU?
1818 * @p: the task in question.
1820 inline int task_curr(const struct task_struct
*p
)
1822 return cpu_curr(task_cpu(p
)) == p
;
1825 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1827 set_task_rq(p
, cpu
);
1830 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1831 * successfuly executed on another CPU. We must ensure that updates of
1832 * per-task data have been completed by this moment.
1835 task_thread_info(p
)->cpu
= cpu
;
1839 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1840 const struct sched_class
*prev_class
,
1841 int oldprio
, int running
)
1843 if (prev_class
!= p
->sched_class
) {
1844 if (prev_class
->switched_from
)
1845 prev_class
->switched_from(rq
, p
, running
);
1846 p
->sched_class
->switched_to(rq
, p
, running
);
1848 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1853 /* Used instead of source_load when we know the type == 0 */
1854 static unsigned long weighted_cpuload(const int cpu
)
1856 return cpu_rq(cpu
)->load
.weight
;
1860 * Is this task likely cache-hot:
1863 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1868 * Buddy candidates are cache hot:
1870 if (sched_feat(CACHE_HOT_BUDDY
) &&
1871 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
1872 &p
->se
== cfs_rq_of(&p
->se
)->last
))
1875 if (p
->sched_class
!= &fair_sched_class
)
1878 if (sysctl_sched_migration_cost
== -1)
1880 if (sysctl_sched_migration_cost
== 0)
1883 delta
= now
- p
->se
.exec_start
;
1885 return delta
< (s64
)sysctl_sched_migration_cost
;
1889 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1891 int old_cpu
= task_cpu(p
);
1892 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1893 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1894 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1897 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1899 trace_sched_migrate_task(p
, task_cpu(p
), new_cpu
);
1901 #ifdef CONFIG_SCHEDSTATS
1902 if (p
->se
.wait_start
)
1903 p
->se
.wait_start
-= clock_offset
;
1904 if (p
->se
.sleep_start
)
1905 p
->se
.sleep_start
-= clock_offset
;
1906 if (p
->se
.block_start
)
1907 p
->se
.block_start
-= clock_offset
;
1909 if (old_cpu
!= new_cpu
) {
1910 p
->se
.nr_migrations
++;
1911 #ifdef CONFIG_SCHEDSTATS
1912 if (task_hot(p
, old_rq
->clock
, NULL
))
1913 schedstat_inc(p
, se
.nr_forced2_migrations
);
1916 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1917 new_cfsrq
->min_vruntime
;
1919 __set_task_cpu(p
, new_cpu
);
1922 struct migration_req
{
1923 struct list_head list
;
1925 struct task_struct
*task
;
1928 struct completion done
;
1932 * The task's runqueue lock must be held.
1933 * Returns true if you have to wait for migration thread.
1936 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1938 struct rq
*rq
= task_rq(p
);
1941 * If the task is not on a runqueue (and not running), then
1942 * it is sufficient to simply update the task's cpu field.
1944 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1945 set_task_cpu(p
, dest_cpu
);
1949 init_completion(&req
->done
);
1951 req
->dest_cpu
= dest_cpu
;
1952 list_add(&req
->list
, &rq
->migration_queue
);
1958 * wait_task_inactive - wait for a thread to unschedule.
1960 * If @match_state is nonzero, it's the @p->state value just checked and
1961 * not expected to change. If it changes, i.e. @p might have woken up,
1962 * then return zero. When we succeed in waiting for @p to be off its CPU,
1963 * we return a positive number (its total switch count). If a second call
1964 * a short while later returns the same number, the caller can be sure that
1965 * @p has remained unscheduled the whole time.
1967 * The caller must ensure that the task *will* unschedule sometime soon,
1968 * else this function might spin for a *long* time. This function can't
1969 * be called with interrupts off, or it may introduce deadlock with
1970 * smp_call_function() if an IPI is sent by the same process we are
1971 * waiting to become inactive.
1973 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1975 unsigned long flags
;
1982 * We do the initial early heuristics without holding
1983 * any task-queue locks at all. We'll only try to get
1984 * the runqueue lock when things look like they will
1990 * If the task is actively running on another CPU
1991 * still, just relax and busy-wait without holding
1994 * NOTE! Since we don't hold any locks, it's not
1995 * even sure that "rq" stays as the right runqueue!
1996 * But we don't care, since "task_running()" will
1997 * return false if the runqueue has changed and p
1998 * is actually now running somewhere else!
2000 while (task_running(rq
, p
)) {
2001 if (match_state
&& unlikely(p
->state
!= match_state
))
2007 * Ok, time to look more closely! We need the rq
2008 * lock now, to be *sure*. If we're wrong, we'll
2009 * just go back and repeat.
2011 rq
= task_rq_lock(p
, &flags
);
2012 trace_sched_wait_task(rq
, p
);
2013 running
= task_running(rq
, p
);
2014 on_rq
= p
->se
.on_rq
;
2016 if (!match_state
|| p
->state
== match_state
)
2017 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2018 task_rq_unlock(rq
, &flags
);
2021 * If it changed from the expected state, bail out now.
2023 if (unlikely(!ncsw
))
2027 * Was it really running after all now that we
2028 * checked with the proper locks actually held?
2030 * Oops. Go back and try again..
2032 if (unlikely(running
)) {
2038 * It's not enough that it's not actively running,
2039 * it must be off the runqueue _entirely_, and not
2042 * So if it wa still runnable (but just not actively
2043 * running right now), it's preempted, and we should
2044 * yield - it could be a while.
2046 if (unlikely(on_rq
)) {
2047 schedule_timeout_uninterruptible(1);
2052 * Ahh, all good. It wasn't running, and it wasn't
2053 * runnable, which means that it will never become
2054 * running in the future either. We're all done!
2063 * kick_process - kick a running thread to enter/exit the kernel
2064 * @p: the to-be-kicked thread
2066 * Cause a process which is running on another CPU to enter
2067 * kernel-mode, without any delay. (to get signals handled.)
2069 * NOTE: this function doesnt have to take the runqueue lock,
2070 * because all it wants to ensure is that the remote task enters
2071 * the kernel. If the IPI races and the task has been migrated
2072 * to another CPU then no harm is done and the purpose has been
2075 void kick_process(struct task_struct
*p
)
2081 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2082 smp_send_reschedule(cpu
);
2087 * Return a low guess at the load of a migration-source cpu weighted
2088 * according to the scheduling class and "nice" value.
2090 * We want to under-estimate the load of migration sources, to
2091 * balance conservatively.
2093 static unsigned long source_load(int cpu
, int type
)
2095 struct rq
*rq
= cpu_rq(cpu
);
2096 unsigned long total
= weighted_cpuload(cpu
);
2098 if (type
== 0 || !sched_feat(LB_BIAS
))
2101 return min(rq
->cpu_load
[type
-1], total
);
2105 * Return a high guess at the load of a migration-target cpu weighted
2106 * according to the scheduling class and "nice" value.
2108 static unsigned long target_load(int cpu
, int type
)
2110 struct rq
*rq
= cpu_rq(cpu
);
2111 unsigned long total
= weighted_cpuload(cpu
);
2113 if (type
== 0 || !sched_feat(LB_BIAS
))
2116 return max(rq
->cpu_load
[type
-1], total
);
2120 * find_idlest_group finds and returns the least busy CPU group within the
2123 static struct sched_group
*
2124 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2126 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2127 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2128 int load_idx
= sd
->forkexec_idx
;
2129 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2132 unsigned long load
, avg_load
;
2136 /* Skip over this group if it has no CPUs allowed */
2137 if (!cpumask_intersects(sched_group_cpus(group
),
2141 local_group
= cpumask_test_cpu(this_cpu
,
2142 sched_group_cpus(group
));
2144 /* Tally up the load of all CPUs in the group */
2147 for_each_cpu(i
, sched_group_cpus(group
)) {
2148 /* Bias balancing toward cpus of our domain */
2150 load
= source_load(i
, load_idx
);
2152 load
= target_load(i
, load_idx
);
2157 /* Adjust by relative CPU power of the group */
2158 avg_load
= sg_div_cpu_power(group
,
2159 avg_load
* SCHED_LOAD_SCALE
);
2162 this_load
= avg_load
;
2164 } else if (avg_load
< min_load
) {
2165 min_load
= avg_load
;
2168 } while (group
= group
->next
, group
!= sd
->groups
);
2170 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2176 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2179 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
2181 unsigned long load
, min_load
= ULONG_MAX
;
2185 /* Traverse only the allowed CPUs */
2186 for_each_cpu_and(i
, sched_group_cpus(group
), &p
->cpus_allowed
) {
2187 load
= weighted_cpuload(i
);
2189 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2199 * sched_balance_self: balance the current task (running on cpu) in domains
2200 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2203 * Balance, ie. select the least loaded group.
2205 * Returns the target CPU number, or the same CPU if no balancing is needed.
2207 * preempt must be disabled.
2209 static int sched_balance_self(int cpu
, int flag
)
2211 struct task_struct
*t
= current
;
2212 struct sched_domain
*tmp
, *sd
= NULL
;
2214 for_each_domain(cpu
, tmp
) {
2216 * If power savings logic is enabled for a domain, stop there.
2218 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2220 if (tmp
->flags
& flag
)
2228 struct sched_group
*group
;
2229 int new_cpu
, weight
;
2231 if (!(sd
->flags
& flag
)) {
2236 group
= find_idlest_group(sd
, t
, cpu
);
2242 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
2243 if (new_cpu
== -1 || new_cpu
== cpu
) {
2244 /* Now try balancing at a lower domain level of cpu */
2249 /* Now try balancing at a lower domain level of new_cpu */
2251 weight
= cpumask_weight(sched_domain_span(sd
));
2253 for_each_domain(cpu
, tmp
) {
2254 if (weight
<= cpumask_weight(sched_domain_span(tmp
)))
2256 if (tmp
->flags
& flag
)
2259 /* while loop will break here if sd == NULL */
2265 #endif /* CONFIG_SMP */
2268 * task_oncpu_function_call - call a function on the cpu on which a task runs
2269 * @p: the task to evaluate
2270 * @func: the function to be called
2271 * @info: the function call argument
2273 * Calls the function @func when the task is currently running. This might
2274 * be on the current CPU, which just calls the function directly
2276 void task_oncpu_function_call(struct task_struct
*p
,
2277 void (*func
) (void *info
), void *info
)
2284 smp_call_function_single(cpu
, func
, info
, 1);
2289 * try_to_wake_up - wake up a thread
2290 * @p: the to-be-woken-up thread
2291 * @state: the mask of task states that can be woken
2292 * @sync: do a synchronous wakeup?
2294 * Put it on the run-queue if it's not already there. The "current"
2295 * thread is always on the run-queue (except when the actual
2296 * re-schedule is in progress), and as such you're allowed to do
2297 * the simpler "current->state = TASK_RUNNING" to mark yourself
2298 * runnable without the overhead of this.
2300 * returns failure only if the task is already active.
2302 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2304 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2305 unsigned long flags
;
2309 if (!sched_feat(SYNC_WAKEUPS
))
2313 if (sched_feat(LB_WAKEUP_UPDATE
)) {
2314 struct sched_domain
*sd
;
2316 this_cpu
= raw_smp_processor_id();
2319 for_each_domain(this_cpu
, sd
) {
2320 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2329 rq
= task_rq_lock(p
, &flags
);
2330 update_rq_clock(rq
);
2331 old_state
= p
->state
;
2332 if (!(old_state
& state
))
2340 this_cpu
= smp_processor_id();
2343 if (unlikely(task_running(rq
, p
)))
2346 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2347 if (cpu
!= orig_cpu
) {
2348 set_task_cpu(p
, cpu
);
2349 task_rq_unlock(rq
, &flags
);
2350 /* might preempt at this point */
2351 rq
= task_rq_lock(p
, &flags
);
2352 old_state
= p
->state
;
2353 if (!(old_state
& state
))
2358 this_cpu
= smp_processor_id();
2362 #ifdef CONFIG_SCHEDSTATS
2363 schedstat_inc(rq
, ttwu_count
);
2364 if (cpu
== this_cpu
)
2365 schedstat_inc(rq
, ttwu_local
);
2367 struct sched_domain
*sd
;
2368 for_each_domain(this_cpu
, sd
) {
2369 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2370 schedstat_inc(sd
, ttwu_wake_remote
);
2375 #endif /* CONFIG_SCHEDSTATS */
2378 #endif /* CONFIG_SMP */
2379 schedstat_inc(p
, se
.nr_wakeups
);
2381 schedstat_inc(p
, se
.nr_wakeups_sync
);
2382 if (orig_cpu
!= cpu
)
2383 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2384 if (cpu
== this_cpu
)
2385 schedstat_inc(p
, se
.nr_wakeups_local
);
2387 schedstat_inc(p
, se
.nr_wakeups_remote
);
2388 activate_task(rq
, p
, 1);
2392 trace_sched_wakeup(rq
, p
, success
);
2393 check_preempt_curr(rq
, p
, sync
);
2395 p
->state
= TASK_RUNNING
;
2397 if (p
->sched_class
->task_wake_up
)
2398 p
->sched_class
->task_wake_up(rq
, p
);
2401 current
->se
.last_wakeup
= current
->se
.sum_exec_runtime
;
2403 task_rq_unlock(rq
, &flags
);
2408 int wake_up_process(struct task_struct
*p
)
2410 return try_to_wake_up(p
, TASK_ALL
, 0);
2412 EXPORT_SYMBOL(wake_up_process
);
2414 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2416 return try_to_wake_up(p
, state
, 0);
2420 * Perform scheduler related setup for a newly forked process p.
2421 * p is forked by current.
2423 * __sched_fork() is basic setup used by init_idle() too:
2425 static void __sched_fork(struct task_struct
*p
)
2427 p
->se
.exec_start
= 0;
2428 p
->se
.sum_exec_runtime
= 0;
2429 p
->se
.prev_sum_exec_runtime
= 0;
2430 p
->se
.nr_migrations
= 0;
2431 p
->se
.last_wakeup
= 0;
2432 p
->se
.avg_overlap
= 0;
2434 #ifdef CONFIG_SCHEDSTATS
2435 p
->se
.wait_start
= 0;
2436 p
->se
.sum_sleep_runtime
= 0;
2437 p
->se
.sleep_start
= 0;
2438 p
->se
.block_start
= 0;
2439 p
->se
.sleep_max
= 0;
2440 p
->se
.block_max
= 0;
2442 p
->se
.slice_max
= 0;
2446 INIT_LIST_HEAD(&p
->rt
.run_list
);
2448 INIT_LIST_HEAD(&p
->se
.group_node
);
2450 #ifdef CONFIG_PREEMPT_NOTIFIERS
2451 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2455 * We mark the process as running here, but have not actually
2456 * inserted it onto the runqueue yet. This guarantees that
2457 * nobody will actually run it, and a signal or other external
2458 * event cannot wake it up and insert it on the runqueue either.
2460 p
->state
= TASK_RUNNING
;
2464 * fork()/clone()-time setup:
2466 void sched_fork(struct task_struct
*p
, int clone_flags
)
2468 int cpu
= get_cpu();
2473 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2475 set_task_cpu(p
, cpu
);
2478 * Make sure we do not leak PI boosting priority to the child:
2480 p
->prio
= current
->normal_prio
;
2481 if (!rt_prio(p
->prio
))
2482 p
->sched_class
= &fair_sched_class
;
2484 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2485 if (likely(sched_info_on()))
2486 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2488 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2491 #ifdef CONFIG_PREEMPT
2492 /* Want to start with kernel preemption disabled. */
2493 task_thread_info(p
)->preempt_count
= 1;
2499 * wake_up_new_task - wake up a newly created task for the first time.
2501 * This function will do some initial scheduler statistics housekeeping
2502 * that must be done for every newly created context, then puts the task
2503 * on the runqueue and wakes it.
2505 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2507 unsigned long flags
;
2510 rq
= task_rq_lock(p
, &flags
);
2511 BUG_ON(p
->state
!= TASK_RUNNING
);
2512 update_rq_clock(rq
);
2514 p
->prio
= effective_prio(p
);
2516 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2517 activate_task(rq
, p
, 0);
2520 * Let the scheduling class do new task startup
2521 * management (if any):
2523 p
->sched_class
->task_new(rq
, p
);
2526 trace_sched_wakeup_new(rq
, p
, 1);
2527 check_preempt_curr(rq
, p
, 0);
2529 if (p
->sched_class
->task_wake_up
)
2530 p
->sched_class
->task_wake_up(rq
, p
);
2532 task_rq_unlock(rq
, &flags
);
2535 #ifdef CONFIG_PREEMPT_NOTIFIERS
2538 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2539 * @notifier: notifier struct to register
2541 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2543 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2545 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2548 * preempt_notifier_unregister - no longer interested in preemption notifications
2549 * @notifier: notifier struct to unregister
2551 * This is safe to call from within a preemption notifier.
2553 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2555 hlist_del(¬ifier
->link
);
2557 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2559 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2561 struct preempt_notifier
*notifier
;
2562 struct hlist_node
*node
;
2564 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2565 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2569 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2570 struct task_struct
*next
)
2572 struct preempt_notifier
*notifier
;
2573 struct hlist_node
*node
;
2575 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2576 notifier
->ops
->sched_out(notifier
, next
);
2579 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2581 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2586 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2587 struct task_struct
*next
)
2591 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2594 * prepare_task_switch - prepare to switch tasks
2595 * @rq: the runqueue preparing to switch
2596 * @prev: the current task that is being switched out
2597 * @next: the task we are going to switch to.
2599 * This is called with the rq lock held and interrupts off. It must
2600 * be paired with a subsequent finish_task_switch after the context
2603 * prepare_task_switch sets up locking and calls architecture specific
2607 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2608 struct task_struct
*next
)
2610 fire_sched_out_preempt_notifiers(prev
, next
);
2611 prepare_lock_switch(rq
, next
);
2612 prepare_arch_switch(next
);
2616 * finish_task_switch - clean up after a task-switch
2617 * @rq: runqueue associated with task-switch
2618 * @prev: the thread we just switched away from.
2620 * finish_task_switch must be called after the context switch, paired
2621 * with a prepare_task_switch call before the context switch.
2622 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2623 * and do any other architecture-specific cleanup actions.
2625 * Note that we may have delayed dropping an mm in context_switch(). If
2626 * so, we finish that here outside of the runqueue lock. (Doing it
2627 * with the lock held can cause deadlocks; see schedule() for
2630 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2631 __releases(rq
->lock
)
2633 struct mm_struct
*mm
= rq
->prev_mm
;
2639 * A task struct has one reference for the use as "current".
2640 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2641 * schedule one last time. The schedule call will never return, and
2642 * the scheduled task must drop that reference.
2643 * The test for TASK_DEAD must occur while the runqueue locks are
2644 * still held, otherwise prev could be scheduled on another cpu, die
2645 * there before we look at prev->state, and then the reference would
2647 * Manfred Spraul <manfred@colorfullife.com>
2649 prev_state
= prev
->state
;
2650 finish_arch_switch(prev
);
2651 perf_counter_task_sched_in(current
, cpu_of(rq
));
2652 finish_lock_switch(rq
, prev
);
2654 if (current
->sched_class
->post_schedule
)
2655 current
->sched_class
->post_schedule(rq
);
2658 fire_sched_in_preempt_notifiers(current
);
2661 if (unlikely(prev_state
== TASK_DEAD
)) {
2663 * Remove function-return probe instances associated with this
2664 * task and put them back on the free list.
2666 kprobe_flush_task(prev
);
2667 put_task_struct(prev
);
2672 * schedule_tail - first thing a freshly forked thread must call.
2673 * @prev: the thread we just switched away from.
2675 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2676 __releases(rq
->lock
)
2678 struct rq
*rq
= this_rq();
2680 finish_task_switch(rq
, prev
);
2681 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2682 /* In this case, finish_task_switch does not reenable preemption */
2685 if (current
->set_child_tid
)
2686 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2690 * context_switch - switch to the new MM and the new
2691 * thread's register state.
2694 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2695 struct task_struct
*next
)
2697 struct mm_struct
*mm
, *oldmm
;
2699 prepare_task_switch(rq
, prev
, next
);
2700 trace_sched_switch(rq
, prev
, next
);
2702 oldmm
= prev
->active_mm
;
2704 * For paravirt, this is coupled with an exit in switch_to to
2705 * combine the page table reload and the switch backend into
2708 arch_enter_lazy_cpu_mode();
2710 if (unlikely(!mm
)) {
2711 next
->active_mm
= oldmm
;
2712 atomic_inc(&oldmm
->mm_count
);
2713 enter_lazy_tlb(oldmm
, next
);
2715 switch_mm(oldmm
, mm
, next
);
2717 if (unlikely(!prev
->mm
)) {
2718 prev
->active_mm
= NULL
;
2719 rq
->prev_mm
= oldmm
;
2722 * Since the runqueue lock will be released by the next
2723 * task (which is an invalid locking op but in the case
2724 * of the scheduler it's an obvious special-case), so we
2725 * do an early lockdep release here:
2727 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2728 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2731 /* Here we just switch the register state and the stack. */
2732 switch_to(prev
, next
, prev
);
2736 * this_rq must be evaluated again because prev may have moved
2737 * CPUs since it called schedule(), thus the 'rq' on its stack
2738 * frame will be invalid.
2740 finish_task_switch(this_rq(), prev
);
2744 * nr_running, nr_uninterruptible and nr_context_switches:
2746 * externally visible scheduler statistics: current number of runnable
2747 * threads, current number of uninterruptible-sleeping threads, total
2748 * number of context switches performed since bootup.
2750 unsigned long nr_running(void)
2752 unsigned long i
, sum
= 0;
2754 for_each_online_cpu(i
)
2755 sum
+= cpu_rq(i
)->nr_running
;
2760 unsigned long nr_uninterruptible(void)
2762 unsigned long i
, sum
= 0;
2764 for_each_possible_cpu(i
)
2765 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2768 * Since we read the counters lockless, it might be slightly
2769 * inaccurate. Do not allow it to go below zero though:
2771 if (unlikely((long)sum
< 0))
2777 unsigned long long nr_context_switches(void)
2780 unsigned long long sum
= 0;
2782 for_each_possible_cpu(i
)
2783 sum
+= cpu_rq(i
)->nr_switches
;
2788 unsigned long nr_iowait(void)
2790 unsigned long i
, sum
= 0;
2792 for_each_possible_cpu(i
)
2793 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2798 unsigned long nr_active(void)
2800 unsigned long i
, running
= 0, uninterruptible
= 0;
2802 for_each_online_cpu(i
) {
2803 running
+= cpu_rq(i
)->nr_running
;
2804 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2807 if (unlikely((long)uninterruptible
< 0))
2808 uninterruptible
= 0;
2810 return running
+ uninterruptible
;
2814 * Update rq->cpu_load[] statistics. This function is usually called every
2815 * scheduler tick (TICK_NSEC).
2817 static void update_cpu_load(struct rq
*this_rq
)
2819 unsigned long this_load
= this_rq
->load
.weight
;
2822 this_rq
->nr_load_updates
++;
2824 /* Update our load: */
2825 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2826 unsigned long old_load
, new_load
;
2828 /* scale is effectively 1 << i now, and >> i divides by scale */
2830 old_load
= this_rq
->cpu_load
[i
];
2831 new_load
= this_load
;
2833 * Round up the averaging division if load is increasing. This
2834 * prevents us from getting stuck on 9 if the load is 10, for
2837 if (new_load
> old_load
)
2838 new_load
+= scale
-1;
2839 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2846 * double_rq_lock - safely lock two runqueues
2848 * Note this does not disable interrupts like task_rq_lock,
2849 * you need to do so manually before calling.
2851 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2852 __acquires(rq1
->lock
)
2853 __acquires(rq2
->lock
)
2855 BUG_ON(!irqs_disabled());
2857 spin_lock(&rq1
->lock
);
2858 __acquire(rq2
->lock
); /* Fake it out ;) */
2861 spin_lock(&rq1
->lock
);
2862 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
2864 spin_lock(&rq2
->lock
);
2865 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
2868 update_rq_clock(rq1
);
2869 update_rq_clock(rq2
);
2873 * double_rq_unlock - safely unlock two runqueues
2875 * Note this does not restore interrupts like task_rq_unlock,
2876 * you need to do so manually after calling.
2878 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2879 __releases(rq1
->lock
)
2880 __releases(rq2
->lock
)
2882 spin_unlock(&rq1
->lock
);
2884 spin_unlock(&rq2
->lock
);
2886 __release(rq2
->lock
);
2890 * If dest_cpu is allowed for this process, migrate the task to it.
2891 * This is accomplished by forcing the cpu_allowed mask to only
2892 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2893 * the cpu_allowed mask is restored.
2895 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2897 struct migration_req req
;
2898 unsigned long flags
;
2901 rq
= task_rq_lock(p
, &flags
);
2902 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
2903 || unlikely(!cpu_active(dest_cpu
)))
2906 /* force the process onto the specified CPU */
2907 if (migrate_task(p
, dest_cpu
, &req
)) {
2908 /* Need to wait for migration thread (might exit: take ref). */
2909 struct task_struct
*mt
= rq
->migration_thread
;
2911 get_task_struct(mt
);
2912 task_rq_unlock(rq
, &flags
);
2913 wake_up_process(mt
);
2914 put_task_struct(mt
);
2915 wait_for_completion(&req
.done
);
2920 task_rq_unlock(rq
, &flags
);
2924 * sched_exec - execve() is a valuable balancing opportunity, because at
2925 * this point the task has the smallest effective memory and cache footprint.
2927 void sched_exec(void)
2929 int new_cpu
, this_cpu
= get_cpu();
2930 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2932 if (new_cpu
!= this_cpu
)
2933 sched_migrate_task(current
, new_cpu
);
2937 * pull_task - move a task from a remote runqueue to the local runqueue.
2938 * Both runqueues must be locked.
2940 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2941 struct rq
*this_rq
, int this_cpu
)
2943 deactivate_task(src_rq
, p
, 0);
2944 set_task_cpu(p
, this_cpu
);
2945 activate_task(this_rq
, p
, 0);
2947 * Note that idle threads have a prio of MAX_PRIO, for this test
2948 * to be always true for them.
2950 check_preempt_curr(this_rq
, p
, 0);
2954 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2957 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2958 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2962 * We do not migrate tasks that are:
2963 * 1) running (obviously), or
2964 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2965 * 3) are cache-hot on their current CPU.
2967 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
2968 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2973 if (task_running(rq
, p
)) {
2974 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2979 * Aggressive migration if:
2980 * 1) task is cache cold, or
2981 * 2) too many balance attempts have failed.
2984 if (!task_hot(p
, rq
->clock
, sd
) ||
2985 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2986 #ifdef CONFIG_SCHEDSTATS
2987 if (task_hot(p
, rq
->clock
, sd
)) {
2988 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2989 schedstat_inc(p
, se
.nr_forced_migrations
);
2995 if (task_hot(p
, rq
->clock
, sd
)) {
2996 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
3002 static unsigned long
3003 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3004 unsigned long max_load_move
, struct sched_domain
*sd
,
3005 enum cpu_idle_type idle
, int *all_pinned
,
3006 int *this_best_prio
, struct rq_iterator
*iterator
)
3008 int loops
= 0, pulled
= 0, pinned
= 0;
3009 struct task_struct
*p
;
3010 long rem_load_move
= max_load_move
;
3012 if (max_load_move
== 0)
3018 * Start the load-balancing iterator:
3020 p
= iterator
->start(iterator
->arg
);
3022 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3025 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
3026 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3027 p
= iterator
->next(iterator
->arg
);
3031 pull_task(busiest
, p
, this_rq
, this_cpu
);
3033 rem_load_move
-= p
->se
.load
.weight
;
3036 * We only want to steal up to the prescribed amount of weighted load.
3038 if (rem_load_move
> 0) {
3039 if (p
->prio
< *this_best_prio
)
3040 *this_best_prio
= p
->prio
;
3041 p
= iterator
->next(iterator
->arg
);
3046 * Right now, this is one of only two places pull_task() is called,
3047 * so we can safely collect pull_task() stats here rather than
3048 * inside pull_task().
3050 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3053 *all_pinned
= pinned
;
3055 return max_load_move
- rem_load_move
;
3059 * move_tasks tries to move up to max_load_move weighted load from busiest to
3060 * this_rq, as part of a balancing operation within domain "sd".
3061 * Returns 1 if successful and 0 otherwise.
3063 * Called with both runqueues locked.
3065 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3066 unsigned long max_load_move
,
3067 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3070 const struct sched_class
*class = sched_class_highest
;
3071 unsigned long total_load_moved
= 0;
3072 int this_best_prio
= this_rq
->curr
->prio
;
3076 class->load_balance(this_rq
, this_cpu
, busiest
,
3077 max_load_move
- total_load_moved
,
3078 sd
, idle
, all_pinned
, &this_best_prio
);
3079 class = class->next
;
3081 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3084 } while (class && max_load_move
> total_load_moved
);
3086 return total_load_moved
> 0;
3090 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3091 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3092 struct rq_iterator
*iterator
)
3094 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3098 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3099 pull_task(busiest
, p
, this_rq
, this_cpu
);
3101 * Right now, this is only the second place pull_task()
3102 * is called, so we can safely collect pull_task()
3103 * stats here rather than inside pull_task().
3105 schedstat_inc(sd
, lb_gained
[idle
]);
3109 p
= iterator
->next(iterator
->arg
);
3116 * move_one_task tries to move exactly one task from busiest to this_rq, as
3117 * part of active balancing operations within "domain".
3118 * Returns 1 if successful and 0 otherwise.
3120 * Called with both runqueues locked.
3122 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3123 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3125 const struct sched_class
*class;
3127 for (class = sched_class_highest
; class; class = class->next
)
3128 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3135 * find_busiest_group finds and returns the busiest CPU group within the
3136 * domain. It calculates and returns the amount of weighted load which
3137 * should be moved to restore balance via the imbalance parameter.
3139 static struct sched_group
*
3140 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3141 unsigned long *imbalance
, enum cpu_idle_type idle
,
3142 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
3144 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
3145 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
3146 unsigned long max_pull
;
3147 unsigned long busiest_load_per_task
, busiest_nr_running
;
3148 unsigned long this_load_per_task
, this_nr_running
;
3149 int load_idx
, group_imb
= 0;
3150 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3151 int power_savings_balance
= 1;
3152 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
3153 unsigned long min_nr_running
= ULONG_MAX
;
3154 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
3157 max_load
= this_load
= total_load
= total_pwr
= 0;
3158 busiest_load_per_task
= busiest_nr_running
= 0;
3159 this_load_per_task
= this_nr_running
= 0;
3161 if (idle
== CPU_NOT_IDLE
)
3162 load_idx
= sd
->busy_idx
;
3163 else if (idle
== CPU_NEWLY_IDLE
)
3164 load_idx
= sd
->newidle_idx
;
3166 load_idx
= sd
->idle_idx
;
3169 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
3172 int __group_imb
= 0;
3173 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3174 unsigned long sum_nr_running
, sum_weighted_load
;
3175 unsigned long sum_avg_load_per_task
;
3176 unsigned long avg_load_per_task
;
3178 local_group
= cpumask_test_cpu(this_cpu
,
3179 sched_group_cpus(group
));
3182 balance_cpu
= cpumask_first(sched_group_cpus(group
));
3184 /* Tally up the load of all CPUs in the group */
3185 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
3186 sum_avg_load_per_task
= avg_load_per_task
= 0;
3189 min_cpu_load
= ~0UL;
3191 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3192 struct rq
*rq
= cpu_rq(i
);
3194 if (*sd_idle
&& rq
->nr_running
)
3197 /* Bias balancing toward cpus of our domain */
3199 if (idle_cpu(i
) && !first_idle_cpu
) {
3204 load
= target_load(i
, load_idx
);
3206 load
= source_load(i
, load_idx
);
3207 if (load
> max_cpu_load
)
3208 max_cpu_load
= load
;
3209 if (min_cpu_load
> load
)
3210 min_cpu_load
= load
;
3214 sum_nr_running
+= rq
->nr_running
;
3215 sum_weighted_load
+= weighted_cpuload(i
);
3217 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3221 * First idle cpu or the first cpu(busiest) in this sched group
3222 * is eligible for doing load balancing at this and above
3223 * domains. In the newly idle case, we will allow all the cpu's
3224 * to do the newly idle load balance.
3226 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3227 balance_cpu
!= this_cpu
&& balance
) {
3232 total_load
+= avg_load
;
3233 total_pwr
+= group
->__cpu_power
;
3235 /* Adjust by relative CPU power of the group */
3236 avg_load
= sg_div_cpu_power(group
,
3237 avg_load
* SCHED_LOAD_SCALE
);
3241 * Consider the group unbalanced when the imbalance is larger
3242 * than the average weight of two tasks.
3244 * APZ: with cgroup the avg task weight can vary wildly and
3245 * might not be a suitable number - should we keep a
3246 * normalized nr_running number somewhere that negates
3249 avg_load_per_task
= sg_div_cpu_power(group
,
3250 sum_avg_load_per_task
* SCHED_LOAD_SCALE
);
3252 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3255 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3258 this_load
= avg_load
;
3260 this_nr_running
= sum_nr_running
;
3261 this_load_per_task
= sum_weighted_load
;
3262 } else if (avg_load
> max_load
&&
3263 (sum_nr_running
> group_capacity
|| __group_imb
)) {
3264 max_load
= avg_load
;
3266 busiest_nr_running
= sum_nr_running
;
3267 busiest_load_per_task
= sum_weighted_load
;
3268 group_imb
= __group_imb
;
3271 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3273 * Busy processors will not participate in power savings
3276 if (idle
== CPU_NOT_IDLE
||
3277 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3281 * If the local group is idle or completely loaded
3282 * no need to do power savings balance at this domain
3284 if (local_group
&& (this_nr_running
>= group_capacity
||
3286 power_savings_balance
= 0;
3289 * If a group is already running at full capacity or idle,
3290 * don't include that group in power savings calculations
3292 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
3297 * Calculate the group which has the least non-idle load.
3298 * This is the group from where we need to pick up the load
3301 if ((sum_nr_running
< min_nr_running
) ||
3302 (sum_nr_running
== min_nr_running
&&
3303 cpumask_first(sched_group_cpus(group
)) >
3304 cpumask_first(sched_group_cpus(group_min
)))) {
3306 min_nr_running
= sum_nr_running
;
3307 min_load_per_task
= sum_weighted_load
/
3312 * Calculate the group which is almost near its
3313 * capacity but still has some space to pick up some load
3314 * from other group and save more power
3316 if (sum_nr_running
<= group_capacity
- 1) {
3317 if (sum_nr_running
> leader_nr_running
||
3318 (sum_nr_running
== leader_nr_running
&&
3319 cpumask_first(sched_group_cpus(group
)) <
3320 cpumask_first(sched_group_cpus(group_leader
)))) {
3321 group_leader
= group
;
3322 leader_nr_running
= sum_nr_running
;
3327 group
= group
->next
;
3328 } while (group
!= sd
->groups
);
3330 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
3333 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
3335 if (this_load
>= avg_load
||
3336 100*max_load
<= sd
->imbalance_pct
*this_load
)
3339 busiest_load_per_task
/= busiest_nr_running
;
3341 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
3344 * We're trying to get all the cpus to the average_load, so we don't
3345 * want to push ourselves above the average load, nor do we wish to
3346 * reduce the max loaded cpu below the average load, as either of these
3347 * actions would just result in more rebalancing later, and ping-pong
3348 * tasks around. Thus we look for the minimum possible imbalance.
3349 * Negative imbalances (*we* are more loaded than anyone else) will
3350 * be counted as no imbalance for these purposes -- we can't fix that
3351 * by pulling tasks to us. Be careful of negative numbers as they'll
3352 * appear as very large values with unsigned longs.
3354 if (max_load
<= busiest_load_per_task
)
3358 * In the presence of smp nice balancing, certain scenarios can have
3359 * max load less than avg load(as we skip the groups at or below
3360 * its cpu_power, while calculating max_load..)
3362 if (max_load
< avg_load
) {
3364 goto small_imbalance
;
3367 /* Don't want to pull so many tasks that a group would go idle */
3368 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
3370 /* How much load to actually move to equalise the imbalance */
3371 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
3372 (avg_load
- this_load
) * this->__cpu_power
)
3376 * if *imbalance is less than the average load per runnable task
3377 * there is no gaurantee that any tasks will be moved so we'll have
3378 * a think about bumping its value to force at least one task to be
3381 if (*imbalance
< busiest_load_per_task
) {
3382 unsigned long tmp
, pwr_now
, pwr_move
;
3386 pwr_move
= pwr_now
= 0;
3388 if (this_nr_running
) {
3389 this_load_per_task
/= this_nr_running
;
3390 if (busiest_load_per_task
> this_load_per_task
)
3393 this_load_per_task
= cpu_avg_load_per_task(this_cpu
);
3395 if (max_load
- this_load
+ busiest_load_per_task
>=
3396 busiest_load_per_task
* imbn
) {
3397 *imbalance
= busiest_load_per_task
;
3402 * OK, we don't have enough imbalance to justify moving tasks,
3403 * however we may be able to increase total CPU power used by
3407 pwr_now
+= busiest
->__cpu_power
*
3408 min(busiest_load_per_task
, max_load
);
3409 pwr_now
+= this->__cpu_power
*
3410 min(this_load_per_task
, this_load
);
3411 pwr_now
/= SCHED_LOAD_SCALE
;
3413 /* Amount of load we'd subtract */
3414 tmp
= sg_div_cpu_power(busiest
,
3415 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3417 pwr_move
+= busiest
->__cpu_power
*
3418 min(busiest_load_per_task
, max_load
- tmp
);
3420 /* Amount of load we'd add */
3421 if (max_load
* busiest
->__cpu_power
<
3422 busiest_load_per_task
* SCHED_LOAD_SCALE
)
3423 tmp
= sg_div_cpu_power(this,
3424 max_load
* busiest
->__cpu_power
);
3426 tmp
= sg_div_cpu_power(this,
3427 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3428 pwr_move
+= this->__cpu_power
*
3429 min(this_load_per_task
, this_load
+ tmp
);
3430 pwr_move
/= SCHED_LOAD_SCALE
;
3432 /* Move if we gain throughput */
3433 if (pwr_move
> pwr_now
)
3434 *imbalance
= busiest_load_per_task
;
3440 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3441 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3444 if (this == group_leader
&& group_leader
!= group_min
) {
3445 *imbalance
= min_load_per_task
;
3446 if (sched_mc_power_savings
>= POWERSAVINGS_BALANCE_WAKEUP
) {
3447 cpu_rq(this_cpu
)->rd
->sched_mc_preferred_wakeup_cpu
=
3448 cpumask_first(sched_group_cpus(group_leader
));
3459 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3462 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3463 unsigned long imbalance
, const struct cpumask
*cpus
)
3465 struct rq
*busiest
= NULL
, *rq
;
3466 unsigned long max_load
= 0;
3469 for_each_cpu(i
, sched_group_cpus(group
)) {
3472 if (!cpumask_test_cpu(i
, cpus
))
3476 wl
= weighted_cpuload(i
);
3478 if (rq
->nr_running
== 1 && wl
> imbalance
)
3481 if (wl
> max_load
) {
3491 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3492 * so long as it is large enough.
3494 #define MAX_PINNED_INTERVAL 512
3497 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3498 * tasks if there is an imbalance.
3500 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3501 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3502 int *balance
, struct cpumask
*cpus
)
3504 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3505 struct sched_group
*group
;
3506 unsigned long imbalance
;
3508 unsigned long flags
;
3510 cpumask_setall(cpus
);
3513 * When power savings policy is enabled for the parent domain, idle
3514 * sibling can pick up load irrespective of busy siblings. In this case,
3515 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3516 * portraying it as CPU_NOT_IDLE.
3518 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3519 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3522 schedstat_inc(sd
, lb_count
[idle
]);
3526 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3533 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3537 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3539 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3543 BUG_ON(busiest
== this_rq
);
3545 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3548 if (busiest
->nr_running
> 1) {
3550 * Attempt to move tasks. If find_busiest_group has found
3551 * an imbalance but busiest->nr_running <= 1, the group is
3552 * still unbalanced. ld_moved simply stays zero, so it is
3553 * correctly treated as an imbalance.
3555 local_irq_save(flags
);
3556 double_rq_lock(this_rq
, busiest
);
3557 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3558 imbalance
, sd
, idle
, &all_pinned
);
3559 double_rq_unlock(this_rq
, busiest
);
3560 local_irq_restore(flags
);
3563 * some other cpu did the load balance for us.
3565 if (ld_moved
&& this_cpu
!= smp_processor_id())
3566 resched_cpu(this_cpu
);
3568 /* All tasks on this runqueue were pinned by CPU affinity */
3569 if (unlikely(all_pinned
)) {
3570 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
3571 if (!cpumask_empty(cpus
))
3578 schedstat_inc(sd
, lb_failed
[idle
]);
3579 sd
->nr_balance_failed
++;
3581 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3583 spin_lock_irqsave(&busiest
->lock
, flags
);
3585 /* don't kick the migration_thread, if the curr
3586 * task on busiest cpu can't be moved to this_cpu
3588 if (!cpumask_test_cpu(this_cpu
,
3589 &busiest
->curr
->cpus_allowed
)) {
3590 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3592 goto out_one_pinned
;
3595 if (!busiest
->active_balance
) {
3596 busiest
->active_balance
= 1;
3597 busiest
->push_cpu
= this_cpu
;
3600 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3602 wake_up_process(busiest
->migration_thread
);
3605 * We've kicked active balancing, reset the failure
3608 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3611 sd
->nr_balance_failed
= 0;
3613 if (likely(!active_balance
)) {
3614 /* We were unbalanced, so reset the balancing interval */
3615 sd
->balance_interval
= sd
->min_interval
;
3618 * If we've begun active balancing, start to back off. This
3619 * case may not be covered by the all_pinned logic if there
3620 * is only 1 task on the busy runqueue (because we don't call
3623 if (sd
->balance_interval
< sd
->max_interval
)
3624 sd
->balance_interval
*= 2;
3627 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3628 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3634 schedstat_inc(sd
, lb_balanced
[idle
]);
3636 sd
->nr_balance_failed
= 0;
3639 /* tune up the balancing interval */
3640 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3641 (sd
->balance_interval
< sd
->max_interval
))
3642 sd
->balance_interval
*= 2;
3644 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3645 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3656 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3657 * tasks if there is an imbalance.
3659 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3660 * this_rq is locked.
3663 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
,
3664 struct cpumask
*cpus
)
3666 struct sched_group
*group
;
3667 struct rq
*busiest
= NULL
;
3668 unsigned long imbalance
;
3673 cpumask_setall(cpus
);
3676 * When power savings policy is enabled for the parent domain, idle
3677 * sibling can pick up load irrespective of busy siblings. In this case,
3678 * let the state of idle sibling percolate up as IDLE, instead of
3679 * portraying it as CPU_NOT_IDLE.
3681 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3682 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3685 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3687 update_shares_locked(this_rq
, sd
);
3688 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3689 &sd_idle
, cpus
, NULL
);
3691 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3695 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
3697 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3701 BUG_ON(busiest
== this_rq
);
3703 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3706 if (busiest
->nr_running
> 1) {
3707 /* Attempt to move tasks */
3708 double_lock_balance(this_rq
, busiest
);
3709 /* this_rq->clock is already updated */
3710 update_rq_clock(busiest
);
3711 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3712 imbalance
, sd
, CPU_NEWLY_IDLE
,
3714 double_unlock_balance(this_rq
, busiest
);
3716 if (unlikely(all_pinned
)) {
3717 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
3718 if (!cpumask_empty(cpus
))
3724 int active_balance
= 0;
3726 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3727 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3728 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3731 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
3734 if (sd
->nr_balance_failed
++ < 2)
3738 * The only task running in a non-idle cpu can be moved to this
3739 * cpu in an attempt to completely freeup the other CPU
3740 * package. The same method used to move task in load_balance()
3741 * have been extended for load_balance_newidle() to speedup
3742 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
3744 * The package power saving logic comes from
3745 * find_busiest_group(). If there are no imbalance, then
3746 * f_b_g() will return NULL. However when sched_mc={1,2} then
3747 * f_b_g() will select a group from which a running task may be
3748 * pulled to this cpu in order to make the other package idle.
3749 * If there is no opportunity to make a package idle and if
3750 * there are no imbalance, then f_b_g() will return NULL and no
3751 * action will be taken in load_balance_newidle().
3753 * Under normal task pull operation due to imbalance, there
3754 * will be more than one task in the source run queue and
3755 * move_tasks() will succeed. ld_moved will be true and this
3756 * active balance code will not be triggered.
3759 /* Lock busiest in correct order while this_rq is held */
3760 double_lock_balance(this_rq
, busiest
);
3763 * don't kick the migration_thread, if the curr
3764 * task on busiest cpu can't be moved to this_cpu
3766 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
3767 double_unlock_balance(this_rq
, busiest
);
3772 if (!busiest
->active_balance
) {
3773 busiest
->active_balance
= 1;
3774 busiest
->push_cpu
= this_cpu
;
3778 double_unlock_balance(this_rq
, busiest
);
3780 * Should not call ttwu while holding a rq->lock
3782 spin_unlock(&this_rq
->lock
);
3784 wake_up_process(busiest
->migration_thread
);
3785 spin_lock(&this_rq
->lock
);
3788 sd
->nr_balance_failed
= 0;
3790 update_shares_locked(this_rq
, sd
);
3794 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3795 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3796 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3798 sd
->nr_balance_failed
= 0;
3804 * idle_balance is called by schedule() if this_cpu is about to become
3805 * idle. Attempts to pull tasks from other CPUs.
3807 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3809 struct sched_domain
*sd
;
3810 int pulled_task
= 0;
3811 unsigned long next_balance
= jiffies
+ HZ
;
3812 cpumask_var_t tmpmask
;
3814 if (!alloc_cpumask_var(&tmpmask
, GFP_ATOMIC
))
3817 for_each_domain(this_cpu
, sd
) {
3818 unsigned long interval
;
3820 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3823 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3824 /* If we've pulled tasks over stop searching: */
3825 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
3828 interval
= msecs_to_jiffies(sd
->balance_interval
);
3829 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3830 next_balance
= sd
->last_balance
+ interval
;
3834 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3836 * We are going idle. next_balance may be set based on
3837 * a busy processor. So reset next_balance.
3839 this_rq
->next_balance
= next_balance
;
3841 free_cpumask_var(tmpmask
);
3845 * active_load_balance is run by migration threads. It pushes running tasks
3846 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3847 * running on each physical CPU where possible, and avoids physical /
3848 * logical imbalances.
3850 * Called with busiest_rq locked.
3852 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3854 int target_cpu
= busiest_rq
->push_cpu
;
3855 struct sched_domain
*sd
;
3856 struct rq
*target_rq
;
3858 /* Is there any task to move? */
3859 if (busiest_rq
->nr_running
<= 1)
3862 target_rq
= cpu_rq(target_cpu
);
3865 * This condition is "impossible", if it occurs
3866 * we need to fix it. Originally reported by
3867 * Bjorn Helgaas on a 128-cpu setup.
3869 BUG_ON(busiest_rq
== target_rq
);
3871 /* move a task from busiest_rq to target_rq */
3872 double_lock_balance(busiest_rq
, target_rq
);
3873 update_rq_clock(busiest_rq
);
3874 update_rq_clock(target_rq
);
3876 /* Search for an sd spanning us and the target CPU. */
3877 for_each_domain(target_cpu
, sd
) {
3878 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3879 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
3884 schedstat_inc(sd
, alb_count
);
3886 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3888 schedstat_inc(sd
, alb_pushed
);
3890 schedstat_inc(sd
, alb_failed
);
3892 double_unlock_balance(busiest_rq
, target_rq
);
3897 atomic_t load_balancer
;
3898 cpumask_var_t cpu_mask
;
3899 } nohz ____cacheline_aligned
= {
3900 .load_balancer
= ATOMIC_INIT(-1),
3904 * This routine will try to nominate the ilb (idle load balancing)
3905 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3906 * load balancing on behalf of all those cpus. If all the cpus in the system
3907 * go into this tickless mode, then there will be no ilb owner (as there is
3908 * no need for one) and all the cpus will sleep till the next wakeup event
3911 * For the ilb owner, tick is not stopped. And this tick will be used
3912 * for idle load balancing. ilb owner will still be part of
3915 * While stopping the tick, this cpu will become the ilb owner if there
3916 * is no other owner. And will be the owner till that cpu becomes busy
3917 * or if all cpus in the system stop their ticks at which point
3918 * there is no need for ilb owner.
3920 * When the ilb owner becomes busy, it nominates another owner, during the
3921 * next busy scheduler_tick()
3923 int select_nohz_load_balancer(int stop_tick
)
3925 int cpu
= smp_processor_id();
3928 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
3929 cpu_rq(cpu
)->in_nohz_recently
= 1;
3932 * If we are going offline and still the leader, give up!
3934 if (!cpu_active(cpu
) &&
3935 atomic_read(&nohz
.load_balancer
) == cpu
) {
3936 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3941 /* time for ilb owner also to sleep */
3942 if (cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3943 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3944 atomic_set(&nohz
.load_balancer
, -1);
3948 if (atomic_read(&nohz
.load_balancer
) == -1) {
3949 /* make me the ilb owner */
3950 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3952 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3955 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
3958 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
3960 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3961 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3968 static DEFINE_SPINLOCK(balancing
);
3971 * It checks each scheduling domain to see if it is due to be balanced,
3972 * and initiates a balancing operation if so.
3974 * Balancing parameters are set up in arch_init_sched_domains.
3976 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3979 struct rq
*rq
= cpu_rq(cpu
);
3980 unsigned long interval
;
3981 struct sched_domain
*sd
;
3982 /* Earliest time when we have to do rebalance again */
3983 unsigned long next_balance
= jiffies
+ 60*HZ
;
3984 int update_next_balance
= 0;
3988 /* Fails alloc? Rebalancing probably not a priority right now. */
3989 if (!alloc_cpumask_var(&tmp
, GFP_ATOMIC
))
3992 for_each_domain(cpu
, sd
) {
3993 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3996 interval
= sd
->balance_interval
;
3997 if (idle
!= CPU_IDLE
)
3998 interval
*= sd
->busy_factor
;
4000 /* scale ms to jiffies */
4001 interval
= msecs_to_jiffies(interval
);
4002 if (unlikely(!interval
))
4004 if (interval
> HZ
*NR_CPUS
/10)
4005 interval
= HZ
*NR_CPUS
/10;
4007 need_serialize
= sd
->flags
& SD_SERIALIZE
;
4009 if (need_serialize
) {
4010 if (!spin_trylock(&balancing
))
4014 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4015 if (load_balance(cpu
, rq
, sd
, idle
, &balance
, tmp
)) {
4017 * We've pulled tasks over so either we're no
4018 * longer idle, or one of our SMT siblings is
4021 idle
= CPU_NOT_IDLE
;
4023 sd
->last_balance
= jiffies
;
4026 spin_unlock(&balancing
);
4028 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4029 next_balance
= sd
->last_balance
+ interval
;
4030 update_next_balance
= 1;
4034 * Stop the load balance at this level. There is another
4035 * CPU in our sched group which is doing load balancing more
4043 * next_balance will be updated only when there is a need.
4044 * When the cpu is attached to null domain for ex, it will not be
4047 if (likely(update_next_balance
))
4048 rq
->next_balance
= next_balance
;
4050 free_cpumask_var(tmp
);
4054 * run_rebalance_domains is triggered when needed from the scheduler tick.
4055 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4056 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4058 static void run_rebalance_domains(struct softirq_action
*h
)
4060 int this_cpu
= smp_processor_id();
4061 struct rq
*this_rq
= cpu_rq(this_cpu
);
4062 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4063 CPU_IDLE
: CPU_NOT_IDLE
;
4065 rebalance_domains(this_cpu
, idle
);
4069 * If this cpu is the owner for idle load balancing, then do the
4070 * balancing on behalf of the other idle cpus whose ticks are
4073 if (this_rq
->idle_at_tick
&&
4074 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4078 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
4079 if (balance_cpu
== this_cpu
)
4083 * If this cpu gets work to do, stop the load balancing
4084 * work being done for other cpus. Next load
4085 * balancing owner will pick it up.
4090 rebalance_domains(balance_cpu
, CPU_IDLE
);
4092 rq
= cpu_rq(balance_cpu
);
4093 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4094 this_rq
->next_balance
= rq
->next_balance
;
4101 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4103 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4104 * idle load balancing owner or decide to stop the periodic load balancing,
4105 * if the whole system is idle.
4107 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4111 * If we were in the nohz mode recently and busy at the current
4112 * scheduler tick, then check if we need to nominate new idle
4115 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4116 rq
->in_nohz_recently
= 0;
4118 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4119 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4120 atomic_set(&nohz
.load_balancer
, -1);
4123 if (atomic_read(&nohz
.load_balancer
) == -1) {
4125 * simple selection for now: Nominate the
4126 * first cpu in the nohz list to be the next
4129 * TBD: Traverse the sched domains and nominate
4130 * the nearest cpu in the nohz.cpu_mask.
4132 int ilb
= cpumask_first(nohz
.cpu_mask
);
4134 if (ilb
< nr_cpu_ids
)
4140 * If this cpu is idle and doing idle load balancing for all the
4141 * cpus with ticks stopped, is it time for that to stop?
4143 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4144 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4150 * If this cpu is idle and the idle load balancing is done by
4151 * someone else, then no need raise the SCHED_SOFTIRQ
4153 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4154 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4157 if (time_after_eq(jiffies
, rq
->next_balance
))
4158 raise_softirq(SCHED_SOFTIRQ
);
4161 #else /* CONFIG_SMP */
4164 * on UP we do not need to balance between CPUs:
4166 static inline void idle_balance(int cpu
, struct rq
*rq
)
4172 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4174 EXPORT_PER_CPU_SYMBOL(kstat
);
4177 * Return any ns on the sched_clock that have not yet been banked in
4178 * @p in case that task is currently running.
4180 unsigned long long __task_delta_exec(struct task_struct
*p
, int update
)
4186 WARN_ON_ONCE(!runqueue_is_locked());
4187 WARN_ON_ONCE(!task_current(rq
, p
));
4190 update_rq_clock(rq
);
4192 delta_exec
= rq
->clock
- p
->se
.exec_start
;
4194 WARN_ON_ONCE(delta_exec
< 0);
4200 * Return any ns on the sched_clock that have not yet been banked in
4201 * @p in case that task is currently running.
4203 unsigned long long task_delta_exec(struct task_struct
*p
)
4205 unsigned long flags
;
4209 rq
= task_rq_lock(p
, &flags
);
4211 if (task_current(rq
, p
)) {
4214 update_rq_clock(rq
);
4215 delta_exec
= rq
->clock
- p
->se
.exec_start
;
4216 if ((s64
)delta_exec
> 0)
4220 task_rq_unlock(rq
, &flags
);
4226 * Account user cpu time to a process.
4227 * @p: the process that the cpu time gets accounted to
4228 * @cputime: the cpu time spent in user space since the last update
4229 * @cputime_scaled: cputime scaled by cpu frequency
4231 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
4232 cputime_t cputime_scaled
)
4234 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4237 /* Add user time to process. */
4238 p
->utime
= cputime_add(p
->utime
, cputime
);
4239 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4240 account_group_user_time(p
, cputime
);
4242 /* Add user time to cpustat. */
4243 tmp
= cputime_to_cputime64(cputime
);
4244 if (TASK_NICE(p
) > 0)
4245 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4247 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4248 /* Account for user time used */
4249 acct_update_integrals(p
);
4253 * Account guest cpu time to a process.
4254 * @p: the process that the cpu time gets accounted to
4255 * @cputime: the cpu time spent in virtual machine since the last update
4256 * @cputime_scaled: cputime scaled by cpu frequency
4258 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
4259 cputime_t cputime_scaled
)
4262 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4264 tmp
= cputime_to_cputime64(cputime
);
4266 /* Add guest time to process. */
4267 p
->utime
= cputime_add(p
->utime
, cputime
);
4268 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4269 account_group_user_time(p
, cputime
);
4270 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4272 /* Add guest time to cpustat. */
4273 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4274 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4278 * Account system cpu time to a process.
4279 * @p: the process that the cpu time gets accounted to
4280 * @hardirq_offset: the offset to subtract from hardirq_count()
4281 * @cputime: the cpu time spent in kernel space since the last update
4282 * @cputime_scaled: cputime scaled by cpu frequency
4284 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4285 cputime_t cputime
, cputime_t cputime_scaled
)
4287 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4290 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4291 account_guest_time(p
, cputime
, cputime_scaled
);
4295 /* Add system time to process. */
4296 p
->stime
= cputime_add(p
->stime
, cputime
);
4297 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
4298 account_group_system_time(p
, cputime
);
4300 /* Add system time to cpustat. */
4301 tmp
= cputime_to_cputime64(cputime
);
4302 if (hardirq_count() - hardirq_offset
)
4303 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4304 else if (softirq_count())
4305 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4307 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4309 /* Account for system time used */
4310 acct_update_integrals(p
);
4314 * Account for involuntary wait time.
4315 * @steal: the cpu time spent in involuntary wait
4317 void account_steal_time(cputime_t cputime
)
4319 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4320 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
4322 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
4326 * Account for idle time.
4327 * @cputime: the cpu time spent in idle wait
4329 void account_idle_time(cputime_t cputime
)
4331 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4332 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
4333 struct rq
*rq
= this_rq();
4335 if (atomic_read(&rq
->nr_iowait
) > 0)
4336 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
4338 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
4341 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4344 * Account a single tick of cpu time.
4345 * @p: the process that the cpu time gets accounted to
4346 * @user_tick: indicates if the tick is a user or a system tick
4348 void account_process_tick(struct task_struct
*p
, int user_tick
)
4350 cputime_t one_jiffy
= jiffies_to_cputime(1);
4351 cputime_t one_jiffy_scaled
= cputime_to_scaled(one_jiffy
);
4352 struct rq
*rq
= this_rq();
4355 account_user_time(p
, one_jiffy
, one_jiffy_scaled
);
4356 else if (p
!= rq
->idle
)
4357 account_system_time(p
, HARDIRQ_OFFSET
, one_jiffy
,
4360 account_idle_time(one_jiffy
);
4364 * Account multiple ticks of steal time.
4365 * @p: the process from which the cpu time has been stolen
4366 * @ticks: number of stolen ticks
4368 void account_steal_ticks(unsigned long ticks
)
4370 account_steal_time(jiffies_to_cputime(ticks
));
4374 * Account multiple ticks of idle time.
4375 * @ticks: number of stolen ticks
4377 void account_idle_ticks(unsigned long ticks
)
4379 account_idle_time(jiffies_to_cputime(ticks
));
4385 * Use precise platform statistics if available:
4387 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4388 cputime_t
task_utime(struct task_struct
*p
)
4393 cputime_t
task_stime(struct task_struct
*p
)
4398 cputime_t
task_utime(struct task_struct
*p
)
4400 clock_t utime
= cputime_to_clock_t(p
->utime
),
4401 total
= utime
+ cputime_to_clock_t(p
->stime
);
4405 * Use CFS's precise accounting:
4407 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
4411 do_div(temp
, total
);
4413 utime
= (clock_t)temp
;
4415 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
4416 return p
->prev_utime
;
4419 cputime_t
task_stime(struct task_struct
*p
)
4424 * Use CFS's precise accounting. (we subtract utime from
4425 * the total, to make sure the total observed by userspace
4426 * grows monotonically - apps rely on that):
4428 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
4429 cputime_to_clock_t(task_utime(p
));
4432 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
4434 return p
->prev_stime
;
4438 inline cputime_t
task_gtime(struct task_struct
*p
)
4444 * This function gets called by the timer code, with HZ frequency.
4445 * We call it with interrupts disabled.
4447 * It also gets called by the fork code, when changing the parent's
4450 void scheduler_tick(void)
4452 int cpu
= smp_processor_id();
4453 struct rq
*rq
= cpu_rq(cpu
);
4454 struct task_struct
*curr
= rq
->curr
;
4458 spin_lock(&rq
->lock
);
4459 update_rq_clock(rq
);
4460 update_cpu_load(rq
);
4461 curr
->sched_class
->task_tick(rq
, curr
, 0);
4462 perf_counter_task_tick(curr
, cpu
);
4463 spin_unlock(&rq
->lock
);
4466 rq
->idle_at_tick
= idle_cpu(cpu
);
4467 trigger_load_balance(rq
, cpu
);
4471 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4472 defined(CONFIG_PREEMPT_TRACER))
4474 static inline unsigned long get_parent_ip(unsigned long addr
)
4476 if (in_lock_functions(addr
)) {
4477 addr
= CALLER_ADDR2
;
4478 if (in_lock_functions(addr
))
4479 addr
= CALLER_ADDR3
;
4484 void __kprobes
add_preempt_count(int val
)
4486 #ifdef CONFIG_DEBUG_PREEMPT
4490 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4493 preempt_count() += val
;
4494 #ifdef CONFIG_DEBUG_PREEMPT
4496 * Spinlock count overflowing soon?
4498 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4501 if (preempt_count() == val
)
4502 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4504 EXPORT_SYMBOL(add_preempt_count
);
4506 void __kprobes
sub_preempt_count(int val
)
4508 #ifdef CONFIG_DEBUG_PREEMPT
4512 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4515 * Is the spinlock portion underflowing?
4517 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4518 !(preempt_count() & PREEMPT_MASK
)))
4522 if (preempt_count() == val
)
4523 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4524 preempt_count() -= val
;
4526 EXPORT_SYMBOL(sub_preempt_count
);
4531 * Print scheduling while atomic bug:
4533 static noinline
void __schedule_bug(struct task_struct
*prev
)
4535 struct pt_regs
*regs
= get_irq_regs();
4537 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4538 prev
->comm
, prev
->pid
, preempt_count());
4540 debug_show_held_locks(prev
);
4542 if (irqs_disabled())
4543 print_irqtrace_events(prev
);
4552 * Various schedule()-time debugging checks and statistics:
4554 static inline void schedule_debug(struct task_struct
*prev
)
4557 * Test if we are atomic. Since do_exit() needs to call into
4558 * schedule() atomically, we ignore that path for now.
4559 * Otherwise, whine if we are scheduling when we should not be.
4561 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4562 __schedule_bug(prev
);
4564 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4566 schedstat_inc(this_rq(), sched_count
);
4567 #ifdef CONFIG_SCHEDSTATS
4568 if (unlikely(prev
->lock_depth
>= 0)) {
4569 schedstat_inc(this_rq(), bkl_count
);
4570 schedstat_inc(prev
, sched_info
.bkl_count
);
4576 * Pick up the highest-prio task:
4578 static inline struct task_struct
*
4579 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
4581 const struct sched_class
*class;
4582 struct task_struct
*p
;
4585 * Optimization: we know that if all tasks are in
4586 * the fair class we can call that function directly:
4588 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4589 p
= fair_sched_class
.pick_next_task(rq
);
4594 class = sched_class_highest
;
4596 p
= class->pick_next_task(rq
);
4600 * Will never be NULL as the idle class always
4601 * returns a non-NULL p:
4603 class = class->next
;
4608 * schedule() is the main scheduler function.
4610 asmlinkage
void __sched
schedule(void)
4612 struct task_struct
*prev
, *next
;
4613 unsigned long *switch_count
;
4619 cpu
= smp_processor_id();
4623 switch_count
= &prev
->nivcsw
;
4625 release_kernel_lock(prev
);
4626 need_resched_nonpreemptible
:
4628 schedule_debug(prev
);
4630 if (sched_feat(HRTICK
))
4633 spin_lock_irq(&rq
->lock
);
4634 update_rq_clock(rq
);
4635 clear_tsk_need_resched(prev
);
4637 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4638 if (unlikely(signal_pending_state(prev
->state
, prev
)))
4639 prev
->state
= TASK_RUNNING
;
4641 deactivate_task(rq
, prev
, 1);
4642 switch_count
= &prev
->nvcsw
;
4646 if (prev
->sched_class
->pre_schedule
)
4647 prev
->sched_class
->pre_schedule(rq
, prev
);
4650 if (unlikely(!rq
->nr_running
))
4651 idle_balance(cpu
, rq
);
4653 prev
->sched_class
->put_prev_task(rq
, prev
);
4654 next
= pick_next_task(rq
, prev
);
4656 if (likely(prev
!= next
)) {
4657 sched_info_switch(prev
, next
);
4658 perf_counter_task_sched_out(prev
, cpu
);
4664 context_switch(rq
, prev
, next
); /* unlocks the rq */
4666 * the context switch might have flipped the stack from under
4667 * us, hence refresh the local variables.
4669 cpu
= smp_processor_id();
4672 spin_unlock_irq(&rq
->lock
);
4674 if (unlikely(reacquire_kernel_lock(current
) < 0))
4675 goto need_resched_nonpreemptible
;
4677 preempt_enable_no_resched();
4678 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
4681 EXPORT_SYMBOL(schedule
);
4683 #ifdef CONFIG_PREEMPT
4685 * this is the entry point to schedule() from in-kernel preemption
4686 * off of preempt_enable. Kernel preemptions off return from interrupt
4687 * occur there and call schedule directly.
4689 asmlinkage
void __sched
preempt_schedule(void)
4691 struct thread_info
*ti
= current_thread_info();
4694 * If there is a non-zero preempt_count or interrupts are disabled,
4695 * we do not want to preempt the current task. Just return..
4697 if (likely(ti
->preempt_count
|| irqs_disabled()))
4701 add_preempt_count(PREEMPT_ACTIVE
);
4703 sub_preempt_count(PREEMPT_ACTIVE
);
4706 * Check again in case we missed a preemption opportunity
4707 * between schedule and now.
4710 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4712 EXPORT_SYMBOL(preempt_schedule
);
4715 * this is the entry point to schedule() from kernel preemption
4716 * off of irq context.
4717 * Note, that this is called and return with irqs disabled. This will
4718 * protect us against recursive calling from irq.
4720 asmlinkage
void __sched
preempt_schedule_irq(void)
4722 struct thread_info
*ti
= current_thread_info();
4724 /* Catch callers which need to be fixed */
4725 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4728 add_preempt_count(PREEMPT_ACTIVE
);
4731 local_irq_disable();
4732 sub_preempt_count(PREEMPT_ACTIVE
);
4735 * Check again in case we missed a preemption opportunity
4736 * between schedule and now.
4739 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4742 #endif /* CONFIG_PREEMPT */
4744 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4747 return try_to_wake_up(curr
->private, mode
, sync
);
4749 EXPORT_SYMBOL(default_wake_function
);
4752 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4753 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4754 * number) then we wake all the non-exclusive tasks and one exclusive task.
4756 * There are circumstances in which we can try to wake a task which has already
4757 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4758 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4760 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4761 int nr_exclusive
, int sync
, void *key
)
4763 wait_queue_t
*curr
, *next
;
4765 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4766 unsigned flags
= curr
->flags
;
4768 if (curr
->func(curr
, mode
, sync
, key
) &&
4769 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4775 * __wake_up - wake up threads blocked on a waitqueue.
4777 * @mode: which threads
4778 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4779 * @key: is directly passed to the wakeup function
4781 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4782 int nr_exclusive
, void *key
)
4784 unsigned long flags
;
4786 spin_lock_irqsave(&q
->lock
, flags
);
4787 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4788 spin_unlock_irqrestore(&q
->lock
, flags
);
4790 EXPORT_SYMBOL(__wake_up
);
4793 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4795 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4797 __wake_up_common(q
, mode
, 1, 0, NULL
);
4801 * __wake_up_sync - wake up threads blocked on a waitqueue.
4803 * @mode: which threads
4804 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4806 * The sync wakeup differs that the waker knows that it will schedule
4807 * away soon, so while the target thread will be woken up, it will not
4808 * be migrated to another CPU - ie. the two threads are 'synchronized'
4809 * with each other. This can prevent needless bouncing between CPUs.
4811 * On UP it can prevent extra preemption.
4814 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4816 unsigned long flags
;
4822 if (unlikely(!nr_exclusive
))
4825 spin_lock_irqsave(&q
->lock
, flags
);
4826 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4827 spin_unlock_irqrestore(&q
->lock
, flags
);
4829 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4832 * complete: - signals a single thread waiting on this completion
4833 * @x: holds the state of this particular completion
4835 * This will wake up a single thread waiting on this completion. Threads will be
4836 * awakened in the same order in which they were queued.
4838 * See also complete_all(), wait_for_completion() and related routines.
4840 void complete(struct completion
*x
)
4842 unsigned long flags
;
4844 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4846 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4847 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4849 EXPORT_SYMBOL(complete
);
4852 * complete_all: - signals all threads waiting on this completion
4853 * @x: holds the state of this particular completion
4855 * This will wake up all threads waiting on this particular completion event.
4857 void complete_all(struct completion
*x
)
4859 unsigned long flags
;
4861 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4862 x
->done
+= UINT_MAX
/2;
4863 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4864 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4866 EXPORT_SYMBOL(complete_all
);
4868 static inline long __sched
4869 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4872 DECLARE_WAITQUEUE(wait
, current
);
4874 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4875 __add_wait_queue_tail(&x
->wait
, &wait
);
4877 if (signal_pending_state(state
, current
)) {
4878 timeout
= -ERESTARTSYS
;
4881 __set_current_state(state
);
4882 spin_unlock_irq(&x
->wait
.lock
);
4883 timeout
= schedule_timeout(timeout
);
4884 spin_lock_irq(&x
->wait
.lock
);
4885 } while (!x
->done
&& timeout
);
4886 __remove_wait_queue(&x
->wait
, &wait
);
4891 return timeout
?: 1;
4895 wait_for_common(struct completion
*x
, long timeout
, int state
)
4899 spin_lock_irq(&x
->wait
.lock
);
4900 timeout
= do_wait_for_common(x
, timeout
, state
);
4901 spin_unlock_irq(&x
->wait
.lock
);
4906 * wait_for_completion: - waits for completion of a task
4907 * @x: holds the state of this particular completion
4909 * This waits to be signaled for completion of a specific task. It is NOT
4910 * interruptible and there is no timeout.
4912 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4913 * and interrupt capability. Also see complete().
4915 void __sched
wait_for_completion(struct completion
*x
)
4917 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4919 EXPORT_SYMBOL(wait_for_completion
);
4922 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4923 * @x: holds the state of this particular completion
4924 * @timeout: timeout value in jiffies
4926 * This waits for either a completion of a specific task to be signaled or for a
4927 * specified timeout to expire. The timeout is in jiffies. It is not
4930 unsigned long __sched
4931 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4933 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4935 EXPORT_SYMBOL(wait_for_completion_timeout
);
4938 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4939 * @x: holds the state of this particular completion
4941 * This waits for completion of a specific task to be signaled. It is
4944 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4946 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4947 if (t
== -ERESTARTSYS
)
4951 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4954 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4955 * @x: holds the state of this particular completion
4956 * @timeout: timeout value in jiffies
4958 * This waits for either a completion of a specific task to be signaled or for a
4959 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4961 unsigned long __sched
4962 wait_for_completion_interruptible_timeout(struct completion
*x
,
4963 unsigned long timeout
)
4965 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4967 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4970 * wait_for_completion_killable: - waits for completion of a task (killable)
4971 * @x: holds the state of this particular completion
4973 * This waits to be signaled for completion of a specific task. It can be
4974 * interrupted by a kill signal.
4976 int __sched
wait_for_completion_killable(struct completion
*x
)
4978 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4979 if (t
== -ERESTARTSYS
)
4983 EXPORT_SYMBOL(wait_for_completion_killable
);
4986 * try_wait_for_completion - try to decrement a completion without blocking
4987 * @x: completion structure
4989 * Returns: 0 if a decrement cannot be done without blocking
4990 * 1 if a decrement succeeded.
4992 * If a completion is being used as a counting completion,
4993 * attempt to decrement the counter without blocking. This
4994 * enables us to avoid waiting if the resource the completion
4995 * is protecting is not available.
4997 bool try_wait_for_completion(struct completion
*x
)
5001 spin_lock_irq(&x
->wait
.lock
);
5006 spin_unlock_irq(&x
->wait
.lock
);
5009 EXPORT_SYMBOL(try_wait_for_completion
);
5012 * completion_done - Test to see if a completion has any waiters
5013 * @x: completion structure
5015 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5016 * 1 if there are no waiters.
5019 bool completion_done(struct completion
*x
)
5023 spin_lock_irq(&x
->wait
.lock
);
5026 spin_unlock_irq(&x
->wait
.lock
);
5029 EXPORT_SYMBOL(completion_done
);
5032 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
5034 unsigned long flags
;
5037 init_waitqueue_entry(&wait
, current
);
5039 __set_current_state(state
);
5041 spin_lock_irqsave(&q
->lock
, flags
);
5042 __add_wait_queue(q
, &wait
);
5043 spin_unlock(&q
->lock
);
5044 timeout
= schedule_timeout(timeout
);
5045 spin_lock_irq(&q
->lock
);
5046 __remove_wait_queue(q
, &wait
);
5047 spin_unlock_irqrestore(&q
->lock
, flags
);
5052 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
5054 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5056 EXPORT_SYMBOL(interruptible_sleep_on
);
5059 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5061 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
5063 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
5065 void __sched
sleep_on(wait_queue_head_t
*q
)
5067 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5069 EXPORT_SYMBOL(sleep_on
);
5071 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5073 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
5075 EXPORT_SYMBOL(sleep_on_timeout
);
5077 #ifdef CONFIG_RT_MUTEXES
5080 * rt_mutex_setprio - set the current priority of a task
5082 * @prio: prio value (kernel-internal form)
5084 * This function changes the 'effective' priority of a task. It does
5085 * not touch ->normal_prio like __setscheduler().
5087 * Used by the rt_mutex code to implement priority inheritance logic.
5089 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
5091 unsigned long flags
;
5092 int oldprio
, on_rq
, running
;
5094 const struct sched_class
*prev_class
= p
->sched_class
;
5096 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
5098 rq
= task_rq_lock(p
, &flags
);
5099 update_rq_clock(rq
);
5102 on_rq
= p
->se
.on_rq
;
5103 running
= task_current(rq
, p
);
5105 dequeue_task(rq
, p
, 0);
5107 p
->sched_class
->put_prev_task(rq
, p
);
5110 p
->sched_class
= &rt_sched_class
;
5112 p
->sched_class
= &fair_sched_class
;
5117 p
->sched_class
->set_curr_task(rq
);
5119 enqueue_task(rq
, p
, 0);
5121 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5123 task_rq_unlock(rq
, &flags
);
5128 void set_user_nice(struct task_struct
*p
, long nice
)
5130 int old_prio
, delta
, on_rq
;
5131 unsigned long flags
;
5134 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
5137 * We have to be careful, if called from sys_setpriority(),
5138 * the task might be in the middle of scheduling on another CPU.
5140 rq
= task_rq_lock(p
, &flags
);
5141 update_rq_clock(rq
);
5143 * The RT priorities are set via sched_setscheduler(), but we still
5144 * allow the 'normal' nice value to be set - but as expected
5145 * it wont have any effect on scheduling until the task is
5146 * SCHED_FIFO/SCHED_RR:
5148 if (task_has_rt_policy(p
)) {
5149 p
->static_prio
= NICE_TO_PRIO(nice
);
5152 on_rq
= p
->se
.on_rq
;
5154 dequeue_task(rq
, p
, 0);
5156 p
->static_prio
= NICE_TO_PRIO(nice
);
5159 p
->prio
= effective_prio(p
);
5160 delta
= p
->prio
- old_prio
;
5163 enqueue_task(rq
, p
, 0);
5165 * If the task increased its priority or is running and
5166 * lowered its priority, then reschedule its CPU:
5168 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
5169 resched_task(rq
->curr
);
5172 task_rq_unlock(rq
, &flags
);
5174 EXPORT_SYMBOL(set_user_nice
);
5177 * can_nice - check if a task can reduce its nice value
5181 int can_nice(const struct task_struct
*p
, const int nice
)
5183 /* convert nice value [19,-20] to rlimit style value [1,40] */
5184 int nice_rlim
= 20 - nice
;
5186 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
5187 capable(CAP_SYS_NICE
));
5190 #ifdef __ARCH_WANT_SYS_NICE
5193 * sys_nice - change the priority of the current process.
5194 * @increment: priority increment
5196 * sys_setpriority is a more generic, but much slower function that
5197 * does similar things.
5199 SYSCALL_DEFINE1(nice
, int, increment
)
5204 * Setpriority might change our priority at the same moment.
5205 * We don't have to worry. Conceptually one call occurs first
5206 * and we have a single winner.
5208 if (increment
< -40)
5213 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
5219 if (increment
< 0 && !can_nice(current
, nice
))
5222 retval
= security_task_setnice(current
, nice
);
5226 set_user_nice(current
, nice
);
5233 * task_prio - return the priority value of a given task.
5234 * @p: the task in question.
5236 * This is the priority value as seen by users in /proc.
5237 * RT tasks are offset by -200. Normal tasks are centered
5238 * around 0, value goes from -16 to +15.
5240 int task_prio(const struct task_struct
*p
)
5242 return p
->prio
- MAX_RT_PRIO
;
5246 * task_nice - return the nice value of a given task.
5247 * @p: the task in question.
5249 int task_nice(const struct task_struct
*p
)
5251 return TASK_NICE(p
);
5253 EXPORT_SYMBOL(task_nice
);
5256 * idle_cpu - is a given cpu idle currently?
5257 * @cpu: the processor in question.
5259 int idle_cpu(int cpu
)
5261 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
5265 * idle_task - return the idle task for a given cpu.
5266 * @cpu: the processor in question.
5268 struct task_struct
*idle_task(int cpu
)
5270 return cpu_rq(cpu
)->idle
;
5274 * find_process_by_pid - find a process with a matching PID value.
5275 * @pid: the pid in question.
5277 static struct task_struct
*find_process_by_pid(pid_t pid
)
5279 return pid
? find_task_by_vpid(pid
) : current
;
5282 /* Actually do priority change: must hold rq lock. */
5284 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
5286 BUG_ON(p
->se
.on_rq
);
5289 switch (p
->policy
) {
5293 p
->sched_class
= &fair_sched_class
;
5297 p
->sched_class
= &rt_sched_class
;
5301 p
->rt_priority
= prio
;
5302 p
->normal_prio
= normal_prio(p
);
5303 /* we are holding p->pi_lock already */
5304 p
->prio
= rt_mutex_getprio(p
);
5309 * check the target process has a UID that matches the current process's
5311 static bool check_same_owner(struct task_struct
*p
)
5313 const struct cred
*cred
= current_cred(), *pcred
;
5317 pcred
= __task_cred(p
);
5318 match
= (cred
->euid
== pcred
->euid
||
5319 cred
->euid
== pcred
->uid
);
5324 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
5325 struct sched_param
*param
, bool user
)
5327 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
5328 unsigned long flags
;
5329 const struct sched_class
*prev_class
= p
->sched_class
;
5332 /* may grab non-irq protected spin_locks */
5333 BUG_ON(in_interrupt());
5335 /* double check policy once rq lock held */
5337 policy
= oldpolicy
= p
->policy
;
5338 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
5339 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
5340 policy
!= SCHED_IDLE
)
5343 * Valid priorities for SCHED_FIFO and SCHED_RR are
5344 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5345 * SCHED_BATCH and SCHED_IDLE is 0.
5347 if (param
->sched_priority
< 0 ||
5348 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5349 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
5351 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
5355 * Allow unprivileged RT tasks to decrease priority:
5357 if (user
&& !capable(CAP_SYS_NICE
)) {
5358 if (rt_policy(policy
)) {
5359 unsigned long rlim_rtprio
;
5361 if (!lock_task_sighand(p
, &flags
))
5363 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
5364 unlock_task_sighand(p
, &flags
);
5366 /* can't set/change the rt policy */
5367 if (policy
!= p
->policy
&& !rlim_rtprio
)
5370 /* can't increase priority */
5371 if (param
->sched_priority
> p
->rt_priority
&&
5372 param
->sched_priority
> rlim_rtprio
)
5376 * Like positive nice levels, dont allow tasks to
5377 * move out of SCHED_IDLE either:
5379 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
5382 /* can't change other user's priorities */
5383 if (!check_same_owner(p
))
5388 #ifdef CONFIG_RT_GROUP_SCHED
5390 * Do not allow realtime tasks into groups that have no runtime
5393 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
5394 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
5398 retval
= security_task_setscheduler(p
, policy
, param
);
5404 * make sure no PI-waiters arrive (or leave) while we are
5405 * changing the priority of the task:
5407 spin_lock_irqsave(&p
->pi_lock
, flags
);
5409 * To be able to change p->policy safely, the apropriate
5410 * runqueue lock must be held.
5412 rq
= __task_rq_lock(p
);
5413 /* recheck policy now with rq lock held */
5414 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5415 policy
= oldpolicy
= -1;
5416 __task_rq_unlock(rq
);
5417 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5420 update_rq_clock(rq
);
5421 on_rq
= p
->se
.on_rq
;
5422 running
= task_current(rq
, p
);
5424 deactivate_task(rq
, p
, 0);
5426 p
->sched_class
->put_prev_task(rq
, p
);
5429 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5432 p
->sched_class
->set_curr_task(rq
);
5434 activate_task(rq
, p
, 0);
5436 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5438 __task_rq_unlock(rq
);
5439 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5441 rt_mutex_adjust_pi(p
);
5447 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5448 * @p: the task in question.
5449 * @policy: new policy.
5450 * @param: structure containing the new RT priority.
5452 * NOTE that the task may be already dead.
5454 int sched_setscheduler(struct task_struct
*p
, int policy
,
5455 struct sched_param
*param
)
5457 return __sched_setscheduler(p
, policy
, param
, true);
5459 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5462 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5463 * @p: the task in question.
5464 * @policy: new policy.
5465 * @param: structure containing the new RT priority.
5467 * Just like sched_setscheduler, only don't bother checking if the
5468 * current context has permission. For example, this is needed in
5469 * stop_machine(): we create temporary high priority worker threads,
5470 * but our caller might not have that capability.
5472 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5473 struct sched_param
*param
)
5475 return __sched_setscheduler(p
, policy
, param
, false);
5479 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5481 struct sched_param lparam
;
5482 struct task_struct
*p
;
5485 if (!param
|| pid
< 0)
5487 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5492 p
= find_process_by_pid(pid
);
5494 retval
= sched_setscheduler(p
, policy
, &lparam
);
5501 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5502 * @pid: the pid in question.
5503 * @policy: new policy.
5504 * @param: structure containing the new RT priority.
5506 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
5507 struct sched_param __user
*, param
)
5509 /* negative values for policy are not valid */
5513 return do_sched_setscheduler(pid
, policy
, param
);
5517 * sys_sched_setparam - set/change the RT priority of a thread
5518 * @pid: the pid in question.
5519 * @param: structure containing the new RT priority.
5521 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5523 return do_sched_setscheduler(pid
, -1, param
);
5527 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5528 * @pid: the pid in question.
5530 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
5532 struct task_struct
*p
;
5539 read_lock(&tasklist_lock
);
5540 p
= find_process_by_pid(pid
);
5542 retval
= security_task_getscheduler(p
);
5546 read_unlock(&tasklist_lock
);
5551 * sys_sched_getscheduler - get the RT priority of a thread
5552 * @pid: the pid in question.
5553 * @param: structure containing the RT priority.
5555 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5557 struct sched_param lp
;
5558 struct task_struct
*p
;
5561 if (!param
|| pid
< 0)
5564 read_lock(&tasklist_lock
);
5565 p
= find_process_by_pid(pid
);
5570 retval
= security_task_getscheduler(p
);
5574 lp
.sched_priority
= p
->rt_priority
;
5575 read_unlock(&tasklist_lock
);
5578 * This one might sleep, we cannot do it with a spinlock held ...
5580 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5585 read_unlock(&tasklist_lock
);
5589 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
5591 cpumask_var_t cpus_allowed
, new_mask
;
5592 struct task_struct
*p
;
5596 read_lock(&tasklist_lock
);
5598 p
= find_process_by_pid(pid
);
5600 read_unlock(&tasklist_lock
);
5606 * It is not safe to call set_cpus_allowed with the
5607 * tasklist_lock held. We will bump the task_struct's
5608 * usage count and then drop tasklist_lock.
5611 read_unlock(&tasklist_lock
);
5613 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
5617 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
5619 goto out_free_cpus_allowed
;
5622 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
5625 retval
= security_task_setscheduler(p
, 0, NULL
);
5629 cpuset_cpus_allowed(p
, cpus_allowed
);
5630 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
5632 retval
= set_cpus_allowed_ptr(p
, new_mask
);
5635 cpuset_cpus_allowed(p
, cpus_allowed
);
5636 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
5638 * We must have raced with a concurrent cpuset
5639 * update. Just reset the cpus_allowed to the
5640 * cpuset's cpus_allowed
5642 cpumask_copy(new_mask
, cpus_allowed
);
5647 free_cpumask_var(new_mask
);
5648 out_free_cpus_allowed
:
5649 free_cpumask_var(cpus_allowed
);
5656 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5657 struct cpumask
*new_mask
)
5659 if (len
< cpumask_size())
5660 cpumask_clear(new_mask
);
5661 else if (len
> cpumask_size())
5662 len
= cpumask_size();
5664 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5668 * sys_sched_setaffinity - set the cpu affinity of a process
5669 * @pid: pid of the process
5670 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5671 * @user_mask_ptr: user-space pointer to the new cpu mask
5673 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
5674 unsigned long __user
*, user_mask_ptr
)
5676 cpumask_var_t new_mask
;
5679 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
5682 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
5684 retval
= sched_setaffinity(pid
, new_mask
);
5685 free_cpumask_var(new_mask
);
5689 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
5691 struct task_struct
*p
;
5695 read_lock(&tasklist_lock
);
5698 p
= find_process_by_pid(pid
);
5702 retval
= security_task_getscheduler(p
);
5706 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
5709 read_unlock(&tasklist_lock
);
5716 * sys_sched_getaffinity - get the cpu affinity of a process
5717 * @pid: pid of the process
5718 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5719 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5721 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
5722 unsigned long __user
*, user_mask_ptr
)
5727 if (len
< cpumask_size())
5730 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
5733 ret
= sched_getaffinity(pid
, mask
);
5735 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
5738 ret
= cpumask_size();
5740 free_cpumask_var(mask
);
5746 * sys_sched_yield - yield the current processor to other threads.
5748 * This function yields the current CPU to other tasks. If there are no
5749 * other threads running on this CPU then this function will return.
5751 SYSCALL_DEFINE0(sched_yield
)
5753 struct rq
*rq
= this_rq_lock();
5755 schedstat_inc(rq
, yld_count
);
5756 current
->sched_class
->yield_task(rq
);
5759 * Since we are going to call schedule() anyway, there's
5760 * no need to preempt or enable interrupts:
5762 __release(rq
->lock
);
5763 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5764 _raw_spin_unlock(&rq
->lock
);
5765 preempt_enable_no_resched();
5772 static void __cond_resched(void)
5774 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5775 __might_sleep(__FILE__
, __LINE__
);
5778 * The BKS might be reacquired before we have dropped
5779 * PREEMPT_ACTIVE, which could trigger a second
5780 * cond_resched() call.
5783 add_preempt_count(PREEMPT_ACTIVE
);
5785 sub_preempt_count(PREEMPT_ACTIVE
);
5786 } while (need_resched());
5789 int __sched
_cond_resched(void)
5791 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
5792 system_state
== SYSTEM_RUNNING
) {
5798 EXPORT_SYMBOL(_cond_resched
);
5801 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5802 * call schedule, and on return reacquire the lock.
5804 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5805 * operations here to prevent schedule() from being called twice (once via
5806 * spin_unlock(), once by hand).
5808 int cond_resched_lock(spinlock_t
*lock
)
5810 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
5813 if (spin_needbreak(lock
) || resched
) {
5815 if (resched
&& need_resched())
5824 EXPORT_SYMBOL(cond_resched_lock
);
5826 int __sched
cond_resched_softirq(void)
5828 BUG_ON(!in_softirq());
5830 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
5838 EXPORT_SYMBOL(cond_resched_softirq
);
5841 * yield - yield the current processor to other threads.
5843 * This is a shortcut for kernel-space yielding - it marks the
5844 * thread runnable and calls sys_sched_yield().
5846 void __sched
yield(void)
5848 set_current_state(TASK_RUNNING
);
5851 EXPORT_SYMBOL(yield
);
5854 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5855 * that process accounting knows that this is a task in IO wait state.
5857 * But don't do that if it is a deliberate, throttling IO wait (this task
5858 * has set its backing_dev_info: the queue against which it should throttle)
5860 void __sched
io_schedule(void)
5862 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5864 delayacct_blkio_start();
5865 atomic_inc(&rq
->nr_iowait
);
5867 atomic_dec(&rq
->nr_iowait
);
5868 delayacct_blkio_end();
5870 EXPORT_SYMBOL(io_schedule
);
5872 long __sched
io_schedule_timeout(long timeout
)
5874 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5877 delayacct_blkio_start();
5878 atomic_inc(&rq
->nr_iowait
);
5879 ret
= schedule_timeout(timeout
);
5880 atomic_dec(&rq
->nr_iowait
);
5881 delayacct_blkio_end();
5886 * sys_sched_get_priority_max - return maximum RT priority.
5887 * @policy: scheduling class.
5889 * this syscall returns the maximum rt_priority that can be used
5890 * by a given scheduling class.
5892 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5899 ret
= MAX_USER_RT_PRIO
-1;
5911 * sys_sched_get_priority_min - return minimum RT priority.
5912 * @policy: scheduling class.
5914 * this syscall returns the minimum rt_priority that can be used
5915 * by a given scheduling class.
5917 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5935 * sys_sched_rr_get_interval - return the default timeslice of a process.
5936 * @pid: pid of the process.
5937 * @interval: userspace pointer to the timeslice value.
5939 * this syscall writes the default timeslice value of a given process
5940 * into the user-space timespec buffer. A value of '0' means infinity.
5942 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5943 struct timespec __user
*, interval
)
5945 struct task_struct
*p
;
5946 unsigned int time_slice
;
5954 read_lock(&tasklist_lock
);
5955 p
= find_process_by_pid(pid
);
5959 retval
= security_task_getscheduler(p
);
5964 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5965 * tasks that are on an otherwise idle runqueue:
5968 if (p
->policy
== SCHED_RR
) {
5969 time_slice
= DEF_TIMESLICE
;
5970 } else if (p
->policy
!= SCHED_FIFO
) {
5971 struct sched_entity
*se
= &p
->se
;
5972 unsigned long flags
;
5975 rq
= task_rq_lock(p
, &flags
);
5976 if (rq
->cfs
.load
.weight
)
5977 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5978 task_rq_unlock(rq
, &flags
);
5980 read_unlock(&tasklist_lock
);
5981 jiffies_to_timespec(time_slice
, &t
);
5982 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5986 read_unlock(&tasklist_lock
);
5990 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5992 void sched_show_task(struct task_struct
*p
)
5994 unsigned long free
= 0;
5997 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5998 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5999 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
6000 #if BITS_PER_LONG == 32
6001 if (state
== TASK_RUNNING
)
6002 printk(KERN_CONT
" running ");
6004 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
6006 if (state
== TASK_RUNNING
)
6007 printk(KERN_CONT
" running task ");
6009 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
6011 #ifdef CONFIG_DEBUG_STACK_USAGE
6012 free
= stack_not_used(p
);
6014 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
6015 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
6017 show_stack(p
, NULL
);
6020 void show_state_filter(unsigned long state_filter
)
6022 struct task_struct
*g
, *p
;
6024 #if BITS_PER_LONG == 32
6026 " task PC stack pid father\n");
6029 " task PC stack pid father\n");
6031 read_lock(&tasklist_lock
);
6032 do_each_thread(g
, p
) {
6034 * reset the NMI-timeout, listing all files on a slow
6035 * console might take alot of time:
6037 touch_nmi_watchdog();
6038 if (!state_filter
|| (p
->state
& state_filter
))
6040 } while_each_thread(g
, p
);
6042 touch_all_softlockup_watchdogs();
6044 #ifdef CONFIG_SCHED_DEBUG
6045 sysrq_sched_debug_show();
6047 read_unlock(&tasklist_lock
);
6049 * Only show locks if all tasks are dumped:
6051 if (state_filter
== -1)
6052 debug_show_all_locks();
6055 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
6057 idle
->sched_class
= &idle_sched_class
;
6061 * init_idle - set up an idle thread for a given CPU
6062 * @idle: task in question
6063 * @cpu: cpu the idle task belongs to
6065 * NOTE: this function does not set the idle thread's NEED_RESCHED
6066 * flag, to make booting more robust.
6068 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
6070 struct rq
*rq
= cpu_rq(cpu
);
6071 unsigned long flags
;
6073 spin_lock_irqsave(&rq
->lock
, flags
);
6076 idle
->se
.exec_start
= sched_clock();
6078 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
6079 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
6080 __set_task_cpu(idle
, cpu
);
6082 rq
->curr
= rq
->idle
= idle
;
6083 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6086 spin_unlock_irqrestore(&rq
->lock
, flags
);
6088 /* Set the preempt count _outside_ the spinlocks! */
6089 #if defined(CONFIG_PREEMPT)
6090 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
6092 task_thread_info(idle
)->preempt_count
= 0;
6095 * The idle tasks have their own, simple scheduling class:
6097 idle
->sched_class
= &idle_sched_class
;
6098 ftrace_graph_init_task(idle
);
6102 * In a system that switches off the HZ timer nohz_cpu_mask
6103 * indicates which cpus entered this state. This is used
6104 * in the rcu update to wait only for active cpus. For system
6105 * which do not switch off the HZ timer nohz_cpu_mask should
6106 * always be CPU_BITS_NONE.
6108 cpumask_var_t nohz_cpu_mask
;
6111 * Increase the granularity value when there are more CPUs,
6112 * because with more CPUs the 'effective latency' as visible
6113 * to users decreases. But the relationship is not linear,
6114 * so pick a second-best guess by going with the log2 of the
6117 * This idea comes from the SD scheduler of Con Kolivas:
6119 static inline void sched_init_granularity(void)
6121 unsigned int factor
= 1 + ilog2(num_online_cpus());
6122 const unsigned long limit
= 200000000;
6124 sysctl_sched_min_granularity
*= factor
;
6125 if (sysctl_sched_min_granularity
> limit
)
6126 sysctl_sched_min_granularity
= limit
;
6128 sysctl_sched_latency
*= factor
;
6129 if (sysctl_sched_latency
> limit
)
6130 sysctl_sched_latency
= limit
;
6132 sysctl_sched_wakeup_granularity
*= factor
;
6134 sysctl_sched_shares_ratelimit
*= factor
;
6139 * This is how migration works:
6141 * 1) we queue a struct migration_req structure in the source CPU's
6142 * runqueue and wake up that CPU's migration thread.
6143 * 2) we down() the locked semaphore => thread blocks.
6144 * 3) migration thread wakes up (implicitly it forces the migrated
6145 * thread off the CPU)
6146 * 4) it gets the migration request and checks whether the migrated
6147 * task is still in the wrong runqueue.
6148 * 5) if it's in the wrong runqueue then the migration thread removes
6149 * it and puts it into the right queue.
6150 * 6) migration thread up()s the semaphore.
6151 * 7) we wake up and the migration is done.
6155 * Change a given task's CPU affinity. Migrate the thread to a
6156 * proper CPU and schedule it away if the CPU it's executing on
6157 * is removed from the allowed bitmask.
6159 * NOTE: the caller must have a valid reference to the task, the
6160 * task must not exit() & deallocate itself prematurely. The
6161 * call is not atomic; no spinlocks may be held.
6163 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
6165 struct migration_req req
;
6166 unsigned long flags
;
6170 rq
= task_rq_lock(p
, &flags
);
6171 if (!cpumask_intersects(new_mask
, cpu_online_mask
)) {
6176 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
6177 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
6182 if (p
->sched_class
->set_cpus_allowed
)
6183 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
6185 cpumask_copy(&p
->cpus_allowed
, new_mask
);
6186 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
6189 /* Can the task run on the task's current CPU? If so, we're done */
6190 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
6193 if (migrate_task(p
, cpumask_any_and(cpu_online_mask
, new_mask
), &req
)) {
6194 /* Need help from migration thread: drop lock and wait. */
6195 task_rq_unlock(rq
, &flags
);
6196 wake_up_process(rq
->migration_thread
);
6197 wait_for_completion(&req
.done
);
6198 tlb_migrate_finish(p
->mm
);
6202 task_rq_unlock(rq
, &flags
);
6206 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
6209 * Move (not current) task off this cpu, onto dest cpu. We're doing
6210 * this because either it can't run here any more (set_cpus_allowed()
6211 * away from this CPU, or CPU going down), or because we're
6212 * attempting to rebalance this task on exec (sched_exec).
6214 * So we race with normal scheduler movements, but that's OK, as long
6215 * as the task is no longer on this CPU.
6217 * Returns non-zero if task was successfully migrated.
6219 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6221 struct rq
*rq_dest
, *rq_src
;
6224 if (unlikely(!cpu_active(dest_cpu
)))
6227 rq_src
= cpu_rq(src_cpu
);
6228 rq_dest
= cpu_rq(dest_cpu
);
6230 double_rq_lock(rq_src
, rq_dest
);
6231 /* Already moved. */
6232 if (task_cpu(p
) != src_cpu
)
6234 /* Affinity changed (again). */
6235 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
6238 on_rq
= p
->se
.on_rq
;
6240 deactivate_task(rq_src
, p
, 0);
6242 set_task_cpu(p
, dest_cpu
);
6244 activate_task(rq_dest
, p
, 0);
6245 check_preempt_curr(rq_dest
, p
, 0);
6250 double_rq_unlock(rq_src
, rq_dest
);
6255 * migration_thread - this is a highprio system thread that performs
6256 * thread migration by bumping thread off CPU then 'pushing' onto
6259 static int migration_thread(void *data
)
6261 int cpu
= (long)data
;
6265 BUG_ON(rq
->migration_thread
!= current
);
6267 set_current_state(TASK_INTERRUPTIBLE
);
6268 while (!kthread_should_stop()) {
6269 struct migration_req
*req
;
6270 struct list_head
*head
;
6272 spin_lock_irq(&rq
->lock
);
6274 if (cpu_is_offline(cpu
)) {
6275 spin_unlock_irq(&rq
->lock
);
6279 if (rq
->active_balance
) {
6280 active_load_balance(rq
, cpu
);
6281 rq
->active_balance
= 0;
6284 head
= &rq
->migration_queue
;
6286 if (list_empty(head
)) {
6287 spin_unlock_irq(&rq
->lock
);
6289 set_current_state(TASK_INTERRUPTIBLE
);
6292 req
= list_entry(head
->next
, struct migration_req
, list
);
6293 list_del_init(head
->next
);
6295 spin_unlock(&rq
->lock
);
6296 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
6299 complete(&req
->done
);
6301 __set_current_state(TASK_RUNNING
);
6305 /* Wait for kthread_stop */
6306 set_current_state(TASK_INTERRUPTIBLE
);
6307 while (!kthread_should_stop()) {
6309 set_current_state(TASK_INTERRUPTIBLE
);
6311 __set_current_state(TASK_RUNNING
);
6315 #ifdef CONFIG_HOTPLUG_CPU
6317 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6321 local_irq_disable();
6322 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
6328 * Figure out where task on dead CPU should go, use force if necessary.
6330 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
6333 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(dead_cpu
));
6336 /* Look for allowed, online CPU in same node. */
6337 for_each_cpu_and(dest_cpu
, nodemask
, cpu_online_mask
)
6338 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
6341 /* Any allowed, online CPU? */
6342 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_online_mask
);
6343 if (dest_cpu
< nr_cpu_ids
)
6346 /* No more Mr. Nice Guy. */
6347 if (dest_cpu
>= nr_cpu_ids
) {
6348 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
6349 dest_cpu
= cpumask_any_and(cpu_online_mask
, &p
->cpus_allowed
);
6352 * Don't tell them about moving exiting tasks or
6353 * kernel threads (both mm NULL), since they never
6356 if (p
->mm
&& printk_ratelimit()) {
6357 printk(KERN_INFO
"process %d (%s) no "
6358 "longer affine to cpu%d\n",
6359 task_pid_nr(p
), p
->comm
, dead_cpu
);
6364 /* It can have affinity changed while we were choosing. */
6365 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
6370 * While a dead CPU has no uninterruptible tasks queued at this point,
6371 * it might still have a nonzero ->nr_uninterruptible counter, because
6372 * for performance reasons the counter is not stricly tracking tasks to
6373 * their home CPUs. So we just add the counter to another CPU's counter,
6374 * to keep the global sum constant after CPU-down:
6376 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
6378 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_online_mask
));
6379 unsigned long flags
;
6381 local_irq_save(flags
);
6382 double_rq_lock(rq_src
, rq_dest
);
6383 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
6384 rq_src
->nr_uninterruptible
= 0;
6385 double_rq_unlock(rq_src
, rq_dest
);
6386 local_irq_restore(flags
);
6389 /* Run through task list and migrate tasks from the dead cpu. */
6390 static void migrate_live_tasks(int src_cpu
)
6392 struct task_struct
*p
, *t
;
6394 read_lock(&tasklist_lock
);
6396 do_each_thread(t
, p
) {
6400 if (task_cpu(p
) == src_cpu
)
6401 move_task_off_dead_cpu(src_cpu
, p
);
6402 } while_each_thread(t
, p
);
6404 read_unlock(&tasklist_lock
);
6408 * Schedules idle task to be the next runnable task on current CPU.
6409 * It does so by boosting its priority to highest possible.
6410 * Used by CPU offline code.
6412 void sched_idle_next(void)
6414 int this_cpu
= smp_processor_id();
6415 struct rq
*rq
= cpu_rq(this_cpu
);
6416 struct task_struct
*p
= rq
->idle
;
6417 unsigned long flags
;
6419 /* cpu has to be offline */
6420 BUG_ON(cpu_online(this_cpu
));
6423 * Strictly not necessary since rest of the CPUs are stopped by now
6424 * and interrupts disabled on the current cpu.
6426 spin_lock_irqsave(&rq
->lock
, flags
);
6428 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6430 update_rq_clock(rq
);
6431 activate_task(rq
, p
, 0);
6433 spin_unlock_irqrestore(&rq
->lock
, flags
);
6437 * Ensures that the idle task is using init_mm right before its cpu goes
6440 void idle_task_exit(void)
6442 struct mm_struct
*mm
= current
->active_mm
;
6444 BUG_ON(cpu_online(smp_processor_id()));
6447 switch_mm(mm
, &init_mm
, current
);
6451 /* called under rq->lock with disabled interrupts */
6452 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
6454 struct rq
*rq
= cpu_rq(dead_cpu
);
6456 /* Must be exiting, otherwise would be on tasklist. */
6457 BUG_ON(!p
->exit_state
);
6459 /* Cannot have done final schedule yet: would have vanished. */
6460 BUG_ON(p
->state
== TASK_DEAD
);
6465 * Drop lock around migration; if someone else moves it,
6466 * that's OK. No task can be added to this CPU, so iteration is
6469 spin_unlock_irq(&rq
->lock
);
6470 move_task_off_dead_cpu(dead_cpu
, p
);
6471 spin_lock_irq(&rq
->lock
);
6476 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6477 static void migrate_dead_tasks(unsigned int dead_cpu
)
6479 struct rq
*rq
= cpu_rq(dead_cpu
);
6480 struct task_struct
*next
;
6483 if (!rq
->nr_running
)
6485 update_rq_clock(rq
);
6486 next
= pick_next_task(rq
, rq
->curr
);
6489 next
->sched_class
->put_prev_task(rq
, next
);
6490 migrate_dead(dead_cpu
, next
);
6494 #endif /* CONFIG_HOTPLUG_CPU */
6496 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6498 static struct ctl_table sd_ctl_dir
[] = {
6500 .procname
= "sched_domain",
6506 static struct ctl_table sd_ctl_root
[] = {
6508 .ctl_name
= CTL_KERN
,
6509 .procname
= "kernel",
6511 .child
= sd_ctl_dir
,
6516 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6518 struct ctl_table
*entry
=
6519 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6524 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6526 struct ctl_table
*entry
;
6529 * In the intermediate directories, both the child directory and
6530 * procname are dynamically allocated and could fail but the mode
6531 * will always be set. In the lowest directory the names are
6532 * static strings and all have proc handlers.
6534 for (entry
= *tablep
; entry
->mode
; entry
++) {
6536 sd_free_ctl_entry(&entry
->child
);
6537 if (entry
->proc_handler
== NULL
)
6538 kfree(entry
->procname
);
6546 set_table_entry(struct ctl_table
*entry
,
6547 const char *procname
, void *data
, int maxlen
,
6548 mode_t mode
, proc_handler
*proc_handler
)
6550 entry
->procname
= procname
;
6552 entry
->maxlen
= maxlen
;
6554 entry
->proc_handler
= proc_handler
;
6557 static struct ctl_table
*
6558 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6560 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
6565 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6566 sizeof(long), 0644, proc_doulongvec_minmax
);
6567 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6568 sizeof(long), 0644, proc_doulongvec_minmax
);
6569 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6570 sizeof(int), 0644, proc_dointvec_minmax
);
6571 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6572 sizeof(int), 0644, proc_dointvec_minmax
);
6573 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6574 sizeof(int), 0644, proc_dointvec_minmax
);
6575 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6576 sizeof(int), 0644, proc_dointvec_minmax
);
6577 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6578 sizeof(int), 0644, proc_dointvec_minmax
);
6579 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6580 sizeof(int), 0644, proc_dointvec_minmax
);
6581 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6582 sizeof(int), 0644, proc_dointvec_minmax
);
6583 set_table_entry(&table
[9], "cache_nice_tries",
6584 &sd
->cache_nice_tries
,
6585 sizeof(int), 0644, proc_dointvec_minmax
);
6586 set_table_entry(&table
[10], "flags", &sd
->flags
,
6587 sizeof(int), 0644, proc_dointvec_minmax
);
6588 set_table_entry(&table
[11], "name", sd
->name
,
6589 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
6590 /* &table[12] is terminator */
6595 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6597 struct ctl_table
*entry
, *table
;
6598 struct sched_domain
*sd
;
6599 int domain_num
= 0, i
;
6602 for_each_domain(cpu
, sd
)
6604 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6609 for_each_domain(cpu
, sd
) {
6610 snprintf(buf
, 32, "domain%d", i
);
6611 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6613 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6620 static struct ctl_table_header
*sd_sysctl_header
;
6621 static void register_sched_domain_sysctl(void)
6623 int i
, cpu_num
= num_online_cpus();
6624 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6627 WARN_ON(sd_ctl_dir
[0].child
);
6628 sd_ctl_dir
[0].child
= entry
;
6633 for_each_online_cpu(i
) {
6634 snprintf(buf
, 32, "cpu%d", i
);
6635 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6637 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6641 WARN_ON(sd_sysctl_header
);
6642 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6645 /* may be called multiple times per register */
6646 static void unregister_sched_domain_sysctl(void)
6648 if (sd_sysctl_header
)
6649 unregister_sysctl_table(sd_sysctl_header
);
6650 sd_sysctl_header
= NULL
;
6651 if (sd_ctl_dir
[0].child
)
6652 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6655 static void register_sched_domain_sysctl(void)
6658 static void unregister_sched_domain_sysctl(void)
6663 static void set_rq_online(struct rq
*rq
)
6666 const struct sched_class
*class;
6668 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
6671 for_each_class(class) {
6672 if (class->rq_online
)
6673 class->rq_online(rq
);
6678 static void set_rq_offline(struct rq
*rq
)
6681 const struct sched_class
*class;
6683 for_each_class(class) {
6684 if (class->rq_offline
)
6685 class->rq_offline(rq
);
6688 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
6694 * migration_call - callback that gets triggered when a CPU is added.
6695 * Here we can start up the necessary migration thread for the new CPU.
6697 static int __cpuinit
6698 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6700 struct task_struct
*p
;
6701 int cpu
= (long)hcpu
;
6702 unsigned long flags
;
6707 case CPU_UP_PREPARE
:
6708 case CPU_UP_PREPARE_FROZEN
:
6709 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
6712 kthread_bind(p
, cpu
);
6713 /* Must be high prio: stop_machine expects to yield to it. */
6714 rq
= task_rq_lock(p
, &flags
);
6715 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6716 task_rq_unlock(rq
, &flags
);
6717 cpu_rq(cpu
)->migration_thread
= p
;
6721 case CPU_ONLINE_FROZEN
:
6722 /* Strictly unnecessary, as first user will wake it. */
6723 wake_up_process(cpu_rq(cpu
)->migration_thread
);
6725 /* Update our root-domain */
6727 spin_lock_irqsave(&rq
->lock
, flags
);
6729 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6733 spin_unlock_irqrestore(&rq
->lock
, flags
);
6736 #ifdef CONFIG_HOTPLUG_CPU
6737 case CPU_UP_CANCELED
:
6738 case CPU_UP_CANCELED_FROZEN
:
6739 if (!cpu_rq(cpu
)->migration_thread
)
6741 /* Unbind it from offline cpu so it can run. Fall thru. */
6742 kthread_bind(cpu_rq(cpu
)->migration_thread
,
6743 cpumask_any(cpu_online_mask
));
6744 kthread_stop(cpu_rq(cpu
)->migration_thread
);
6745 cpu_rq(cpu
)->migration_thread
= NULL
;
6749 case CPU_DEAD_FROZEN
:
6750 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6751 migrate_live_tasks(cpu
);
6753 kthread_stop(rq
->migration_thread
);
6754 rq
->migration_thread
= NULL
;
6755 /* Idle task back to normal (off runqueue, low prio) */
6756 spin_lock_irq(&rq
->lock
);
6757 update_rq_clock(rq
);
6758 deactivate_task(rq
, rq
->idle
, 0);
6759 rq
->idle
->static_prio
= MAX_PRIO
;
6760 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
6761 rq
->idle
->sched_class
= &idle_sched_class
;
6762 migrate_dead_tasks(cpu
);
6763 spin_unlock_irq(&rq
->lock
);
6765 migrate_nr_uninterruptible(rq
);
6766 BUG_ON(rq
->nr_running
!= 0);
6769 * No need to migrate the tasks: it was best-effort if
6770 * they didn't take sched_hotcpu_mutex. Just wake up
6773 spin_lock_irq(&rq
->lock
);
6774 while (!list_empty(&rq
->migration_queue
)) {
6775 struct migration_req
*req
;
6777 req
= list_entry(rq
->migration_queue
.next
,
6778 struct migration_req
, list
);
6779 list_del_init(&req
->list
);
6780 spin_unlock_irq(&rq
->lock
);
6781 complete(&req
->done
);
6782 spin_lock_irq(&rq
->lock
);
6784 spin_unlock_irq(&rq
->lock
);
6788 case CPU_DYING_FROZEN
:
6789 /* Update our root-domain */
6791 spin_lock_irqsave(&rq
->lock
, flags
);
6793 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6796 spin_unlock_irqrestore(&rq
->lock
, flags
);
6803 /* Register at highest priority so that task migration (migrate_all_tasks)
6804 * happens before everything else.
6806 static struct notifier_block __cpuinitdata migration_notifier
= {
6807 .notifier_call
= migration_call
,
6811 static int __init
migration_init(void)
6813 void *cpu
= (void *)(long)smp_processor_id();
6816 /* Start one for the boot CPU: */
6817 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6818 BUG_ON(err
== NOTIFY_BAD
);
6819 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6820 register_cpu_notifier(&migration_notifier
);
6824 early_initcall(migration_init
);
6829 #ifdef CONFIG_SCHED_DEBUG
6831 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6832 struct cpumask
*groupmask
)
6834 struct sched_group
*group
= sd
->groups
;
6837 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
6838 cpumask_clear(groupmask
);
6840 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6842 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6843 printk("does not load-balance\n");
6845 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6850 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
6852 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
6853 printk(KERN_ERR
"ERROR: domain->span does not contain "
6856 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
6857 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6861 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6865 printk(KERN_ERR
"ERROR: group is NULL\n");
6869 if (!group
->__cpu_power
) {
6870 printk(KERN_CONT
"\n");
6871 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6876 if (!cpumask_weight(sched_group_cpus(group
))) {
6877 printk(KERN_CONT
"\n");
6878 printk(KERN_ERR
"ERROR: empty group\n");
6882 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
6883 printk(KERN_CONT
"\n");
6884 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6888 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
6890 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
6891 printk(KERN_CONT
" %s", str
);
6893 group
= group
->next
;
6894 } while (group
!= sd
->groups
);
6895 printk(KERN_CONT
"\n");
6897 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
6898 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6901 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
6902 printk(KERN_ERR
"ERROR: parent span is not a superset "
6903 "of domain->span\n");
6907 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6909 cpumask_var_t groupmask
;
6913 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6917 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6919 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
6920 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6925 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6932 free_cpumask_var(groupmask
);
6934 #else /* !CONFIG_SCHED_DEBUG */
6935 # define sched_domain_debug(sd, cpu) do { } while (0)
6936 #endif /* CONFIG_SCHED_DEBUG */
6938 static int sd_degenerate(struct sched_domain
*sd
)
6940 if (cpumask_weight(sched_domain_span(sd
)) == 1)
6943 /* Following flags need at least 2 groups */
6944 if (sd
->flags
& (SD_LOAD_BALANCE
|
6945 SD_BALANCE_NEWIDLE
|
6949 SD_SHARE_PKG_RESOURCES
)) {
6950 if (sd
->groups
!= sd
->groups
->next
)
6954 /* Following flags don't use groups */
6955 if (sd
->flags
& (SD_WAKE_IDLE
|
6964 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6966 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6968 if (sd_degenerate(parent
))
6971 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
6974 /* Does parent contain flags not in child? */
6975 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6976 if (cflags
& SD_WAKE_AFFINE
)
6977 pflags
&= ~SD_WAKE_BALANCE
;
6978 /* Flags needing groups don't count if only 1 group in parent */
6979 if (parent
->groups
== parent
->groups
->next
) {
6980 pflags
&= ~(SD_LOAD_BALANCE
|
6981 SD_BALANCE_NEWIDLE
|
6985 SD_SHARE_PKG_RESOURCES
);
6986 if (nr_node_ids
== 1)
6987 pflags
&= ~SD_SERIALIZE
;
6989 if (~cflags
& pflags
)
6995 static void free_rootdomain(struct root_domain
*rd
)
6997 cpupri_cleanup(&rd
->cpupri
);
6999 free_cpumask_var(rd
->rto_mask
);
7000 free_cpumask_var(rd
->online
);
7001 free_cpumask_var(rd
->span
);
7005 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
7007 unsigned long flags
;
7009 spin_lock_irqsave(&rq
->lock
, flags
);
7012 struct root_domain
*old_rd
= rq
->rd
;
7014 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
7017 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
7019 if (atomic_dec_and_test(&old_rd
->refcount
))
7020 free_rootdomain(old_rd
);
7023 atomic_inc(&rd
->refcount
);
7026 cpumask_set_cpu(rq
->cpu
, rd
->span
);
7027 if (cpumask_test_cpu(rq
->cpu
, cpu_online_mask
))
7030 spin_unlock_irqrestore(&rq
->lock
, flags
);
7033 static int __init_refok
init_rootdomain(struct root_domain
*rd
, bool bootmem
)
7035 memset(rd
, 0, sizeof(*rd
));
7038 alloc_bootmem_cpumask_var(&def_root_domain
.span
);
7039 alloc_bootmem_cpumask_var(&def_root_domain
.online
);
7040 alloc_bootmem_cpumask_var(&def_root_domain
.rto_mask
);
7041 cpupri_init(&rd
->cpupri
, true);
7045 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
7047 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
7049 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
7052 if (cpupri_init(&rd
->cpupri
, false) != 0)
7057 free_cpumask_var(rd
->rto_mask
);
7059 free_cpumask_var(rd
->online
);
7061 free_cpumask_var(rd
->span
);
7066 static void init_defrootdomain(void)
7068 init_rootdomain(&def_root_domain
, true);
7070 atomic_set(&def_root_domain
.refcount
, 1);
7073 static struct root_domain
*alloc_rootdomain(void)
7075 struct root_domain
*rd
;
7077 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
7081 if (init_rootdomain(rd
, false) != 0) {
7090 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7091 * hold the hotplug lock.
7094 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
7096 struct rq
*rq
= cpu_rq(cpu
);
7097 struct sched_domain
*tmp
;
7099 /* Remove the sched domains which do not contribute to scheduling. */
7100 for (tmp
= sd
; tmp
; ) {
7101 struct sched_domain
*parent
= tmp
->parent
;
7105 if (sd_parent_degenerate(tmp
, parent
)) {
7106 tmp
->parent
= parent
->parent
;
7108 parent
->parent
->child
= tmp
;
7113 if (sd
&& sd_degenerate(sd
)) {
7119 sched_domain_debug(sd
, cpu
);
7121 rq_attach_root(rq
, rd
);
7122 rcu_assign_pointer(rq
->sd
, sd
);
7125 /* cpus with isolated domains */
7126 static cpumask_var_t cpu_isolated_map
;
7128 /* Setup the mask of cpus configured for isolated domains */
7129 static int __init
isolated_cpu_setup(char *str
)
7131 cpulist_parse(str
, cpu_isolated_map
);
7135 __setup("isolcpus=", isolated_cpu_setup
);
7138 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7139 * to a function which identifies what group(along with sched group) a CPU
7140 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7141 * (due to the fact that we keep track of groups covered with a struct cpumask).
7143 * init_sched_build_groups will build a circular linked list of the groups
7144 * covered by the given span, and will set each group's ->cpumask correctly,
7145 * and ->cpu_power to 0.
7148 init_sched_build_groups(const struct cpumask
*span
,
7149 const struct cpumask
*cpu_map
,
7150 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
7151 struct sched_group
**sg
,
7152 struct cpumask
*tmpmask
),
7153 struct cpumask
*covered
, struct cpumask
*tmpmask
)
7155 struct sched_group
*first
= NULL
, *last
= NULL
;
7158 cpumask_clear(covered
);
7160 for_each_cpu(i
, span
) {
7161 struct sched_group
*sg
;
7162 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
7165 if (cpumask_test_cpu(i
, covered
))
7168 cpumask_clear(sched_group_cpus(sg
));
7169 sg
->__cpu_power
= 0;
7171 for_each_cpu(j
, span
) {
7172 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
7175 cpumask_set_cpu(j
, covered
);
7176 cpumask_set_cpu(j
, sched_group_cpus(sg
));
7187 #define SD_NODES_PER_DOMAIN 16
7192 * find_next_best_node - find the next node to include in a sched_domain
7193 * @node: node whose sched_domain we're building
7194 * @used_nodes: nodes already in the sched_domain
7196 * Find the next node to include in a given scheduling domain. Simply
7197 * finds the closest node not already in the @used_nodes map.
7199 * Should use nodemask_t.
7201 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
7203 int i
, n
, val
, min_val
, best_node
= 0;
7207 for (i
= 0; i
< nr_node_ids
; i
++) {
7208 /* Start at @node */
7209 n
= (node
+ i
) % nr_node_ids
;
7211 if (!nr_cpus_node(n
))
7214 /* Skip already used nodes */
7215 if (node_isset(n
, *used_nodes
))
7218 /* Simple min distance search */
7219 val
= node_distance(node
, n
);
7221 if (val
< min_val
) {
7227 node_set(best_node
, *used_nodes
);
7232 * sched_domain_node_span - get a cpumask for a node's sched_domain
7233 * @node: node whose cpumask we're constructing
7234 * @span: resulting cpumask
7236 * Given a node, construct a good cpumask for its sched_domain to span. It
7237 * should be one that prevents unnecessary balancing, but also spreads tasks
7240 static void sched_domain_node_span(int node
, struct cpumask
*span
)
7242 nodemask_t used_nodes
;
7245 cpumask_clear(span
);
7246 nodes_clear(used_nodes
);
7248 cpumask_or(span
, span
, cpumask_of_node(node
));
7249 node_set(node
, used_nodes
);
7251 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
7252 int next_node
= find_next_best_node(node
, &used_nodes
);
7254 cpumask_or(span
, span
, cpumask_of_node(next_node
));
7257 #endif /* CONFIG_NUMA */
7259 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
7262 * The cpus mask in sched_group and sched_domain hangs off the end.
7263 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7264 * for nr_cpu_ids < CONFIG_NR_CPUS.
7266 struct static_sched_group
{
7267 struct sched_group sg
;
7268 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
7271 struct static_sched_domain
{
7272 struct sched_domain sd
;
7273 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
7277 * SMT sched-domains:
7279 #ifdef CONFIG_SCHED_SMT
7280 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
7281 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_cpus
);
7284 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
7285 struct sched_group
**sg
, struct cpumask
*unused
)
7288 *sg
= &per_cpu(sched_group_cpus
, cpu
).sg
;
7291 #endif /* CONFIG_SCHED_SMT */
7294 * multi-core sched-domains:
7296 #ifdef CONFIG_SCHED_MC
7297 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
7298 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
7299 #endif /* CONFIG_SCHED_MC */
7301 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7303 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
7304 struct sched_group
**sg
, struct cpumask
*mask
)
7308 cpumask_and(mask
, &per_cpu(cpu_sibling_map
, cpu
), cpu_map
);
7309 group
= cpumask_first(mask
);
7311 *sg
= &per_cpu(sched_group_core
, group
).sg
;
7314 #elif defined(CONFIG_SCHED_MC)
7316 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
7317 struct sched_group
**sg
, struct cpumask
*unused
)
7320 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
7325 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
7326 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
7329 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
7330 struct sched_group
**sg
, struct cpumask
*mask
)
7333 #ifdef CONFIG_SCHED_MC
7334 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
7335 group
= cpumask_first(mask
);
7336 #elif defined(CONFIG_SCHED_SMT)
7337 cpumask_and(mask
, &per_cpu(cpu_sibling_map
, cpu
), cpu_map
);
7338 group
= cpumask_first(mask
);
7343 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
7349 * The init_sched_build_groups can't handle what we want to do with node
7350 * groups, so roll our own. Now each node has its own list of groups which
7351 * gets dynamically allocated.
7353 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
7354 static struct sched_group
***sched_group_nodes_bycpu
;
7356 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
7357 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
7359 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
7360 struct sched_group
**sg
,
7361 struct cpumask
*nodemask
)
7365 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
7366 group
= cpumask_first(nodemask
);
7369 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
7373 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
7375 struct sched_group
*sg
= group_head
;
7381 for_each_cpu(j
, sched_group_cpus(sg
)) {
7382 struct sched_domain
*sd
;
7384 sd
= &per_cpu(phys_domains
, j
).sd
;
7385 if (j
!= cpumask_first(sched_group_cpus(sd
->groups
))) {
7387 * Only add "power" once for each
7393 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
7396 } while (sg
!= group_head
);
7398 #endif /* CONFIG_NUMA */
7401 /* Free memory allocated for various sched_group structures */
7402 static void free_sched_groups(const struct cpumask
*cpu_map
,
7403 struct cpumask
*nodemask
)
7407 for_each_cpu(cpu
, cpu_map
) {
7408 struct sched_group
**sched_group_nodes
7409 = sched_group_nodes_bycpu
[cpu
];
7411 if (!sched_group_nodes
)
7414 for (i
= 0; i
< nr_node_ids
; i
++) {
7415 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
7417 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
7418 if (cpumask_empty(nodemask
))
7428 if (oldsg
!= sched_group_nodes
[i
])
7431 kfree(sched_group_nodes
);
7432 sched_group_nodes_bycpu
[cpu
] = NULL
;
7435 #else /* !CONFIG_NUMA */
7436 static void free_sched_groups(const struct cpumask
*cpu_map
,
7437 struct cpumask
*nodemask
)
7440 #endif /* CONFIG_NUMA */
7443 * Initialize sched groups cpu_power.
7445 * cpu_power indicates the capacity of sched group, which is used while
7446 * distributing the load between different sched groups in a sched domain.
7447 * Typically cpu_power for all the groups in a sched domain will be same unless
7448 * there are asymmetries in the topology. If there are asymmetries, group
7449 * having more cpu_power will pickup more load compared to the group having
7452 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7453 * the maximum number of tasks a group can handle in the presence of other idle
7454 * or lightly loaded groups in the same sched domain.
7456 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7458 struct sched_domain
*child
;
7459 struct sched_group
*group
;
7461 WARN_ON(!sd
|| !sd
->groups
);
7463 if (cpu
!= cpumask_first(sched_group_cpus(sd
->groups
)))
7468 sd
->groups
->__cpu_power
= 0;
7471 * For perf policy, if the groups in child domain share resources
7472 * (for example cores sharing some portions of the cache hierarchy
7473 * or SMT), then set this domain groups cpu_power such that each group
7474 * can handle only one task, when there are other idle groups in the
7475 * same sched domain.
7477 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
7479 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
7480 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
7485 * add cpu_power of each child group to this groups cpu_power
7487 group
= child
->groups
;
7489 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
7490 group
= group
->next
;
7491 } while (group
!= child
->groups
);
7495 * Initializers for schedule domains
7496 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7499 #ifdef CONFIG_SCHED_DEBUG
7500 # define SD_INIT_NAME(sd, type) sd->name = #type
7502 # define SD_INIT_NAME(sd, type) do { } while (0)
7505 #define SD_INIT(sd, type) sd_init_##type(sd)
7507 #define SD_INIT_FUNC(type) \
7508 static noinline void sd_init_##type(struct sched_domain *sd) \
7510 memset(sd, 0, sizeof(*sd)); \
7511 *sd = SD_##type##_INIT; \
7512 sd->level = SD_LV_##type; \
7513 SD_INIT_NAME(sd, type); \
7518 SD_INIT_FUNC(ALLNODES
)
7521 #ifdef CONFIG_SCHED_SMT
7522 SD_INIT_FUNC(SIBLING
)
7524 #ifdef CONFIG_SCHED_MC
7528 static int default_relax_domain_level
= -1;
7530 static int __init
setup_relax_domain_level(char *str
)
7534 val
= simple_strtoul(str
, NULL
, 0);
7535 if (val
< SD_LV_MAX
)
7536 default_relax_domain_level
= val
;
7540 __setup("relax_domain_level=", setup_relax_domain_level
);
7542 static void set_domain_attribute(struct sched_domain
*sd
,
7543 struct sched_domain_attr
*attr
)
7547 if (!attr
|| attr
->relax_domain_level
< 0) {
7548 if (default_relax_domain_level
< 0)
7551 request
= default_relax_domain_level
;
7553 request
= attr
->relax_domain_level
;
7554 if (request
< sd
->level
) {
7555 /* turn off idle balance on this domain */
7556 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
7558 /* turn on idle balance on this domain */
7559 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
7564 * Build sched domains for a given set of cpus and attach the sched domains
7565 * to the individual cpus
7567 static int __build_sched_domains(const struct cpumask
*cpu_map
,
7568 struct sched_domain_attr
*attr
)
7570 int i
, err
= -ENOMEM
;
7571 struct root_domain
*rd
;
7572 cpumask_var_t nodemask
, this_sibling_map
, this_core_map
, send_covered
,
7575 cpumask_var_t domainspan
, covered
, notcovered
;
7576 struct sched_group
**sched_group_nodes
= NULL
;
7577 int sd_allnodes
= 0;
7579 if (!alloc_cpumask_var(&domainspan
, GFP_KERNEL
))
7581 if (!alloc_cpumask_var(&covered
, GFP_KERNEL
))
7582 goto free_domainspan
;
7583 if (!alloc_cpumask_var(¬covered
, GFP_KERNEL
))
7587 if (!alloc_cpumask_var(&nodemask
, GFP_KERNEL
))
7588 goto free_notcovered
;
7589 if (!alloc_cpumask_var(&this_sibling_map
, GFP_KERNEL
))
7591 if (!alloc_cpumask_var(&this_core_map
, GFP_KERNEL
))
7592 goto free_this_sibling_map
;
7593 if (!alloc_cpumask_var(&send_covered
, GFP_KERNEL
))
7594 goto free_this_core_map
;
7595 if (!alloc_cpumask_var(&tmpmask
, GFP_KERNEL
))
7596 goto free_send_covered
;
7600 * Allocate the per-node list of sched groups
7602 sched_group_nodes
= kcalloc(nr_node_ids
, sizeof(struct sched_group
*),
7604 if (!sched_group_nodes
) {
7605 printk(KERN_WARNING
"Can not alloc sched group node list\n");
7610 rd
= alloc_rootdomain();
7612 printk(KERN_WARNING
"Cannot alloc root domain\n");
7613 goto free_sched_groups
;
7617 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = sched_group_nodes
;
7621 * Set up domains for cpus specified by the cpu_map.
7623 for_each_cpu(i
, cpu_map
) {
7624 struct sched_domain
*sd
= NULL
, *p
;
7626 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(i
)), cpu_map
);
7629 if (cpumask_weight(cpu_map
) >
7630 SD_NODES_PER_DOMAIN
*cpumask_weight(nodemask
)) {
7631 sd
= &per_cpu(allnodes_domains
, i
).sd
;
7632 SD_INIT(sd
, ALLNODES
);
7633 set_domain_attribute(sd
, attr
);
7634 cpumask_copy(sched_domain_span(sd
), cpu_map
);
7635 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7641 sd
= &per_cpu(node_domains
, i
).sd
;
7643 set_domain_attribute(sd
, attr
);
7644 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
7648 cpumask_and(sched_domain_span(sd
),
7649 sched_domain_span(sd
), cpu_map
);
7653 sd
= &per_cpu(phys_domains
, i
).sd
;
7655 set_domain_attribute(sd
, attr
);
7656 cpumask_copy(sched_domain_span(sd
), nodemask
);
7660 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7662 #ifdef CONFIG_SCHED_MC
7664 sd
= &per_cpu(core_domains
, i
).sd
;
7666 set_domain_attribute(sd
, attr
);
7667 cpumask_and(sched_domain_span(sd
), cpu_map
,
7668 cpu_coregroup_mask(i
));
7671 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7674 #ifdef CONFIG_SCHED_SMT
7676 sd
= &per_cpu(cpu_domains
, i
).sd
;
7677 SD_INIT(sd
, SIBLING
);
7678 set_domain_attribute(sd
, attr
);
7679 cpumask_and(sched_domain_span(sd
),
7680 &per_cpu(cpu_sibling_map
, i
), cpu_map
);
7683 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7687 #ifdef CONFIG_SCHED_SMT
7688 /* Set up CPU (sibling) groups */
7689 for_each_cpu(i
, cpu_map
) {
7690 cpumask_and(this_sibling_map
,
7691 &per_cpu(cpu_sibling_map
, i
), cpu_map
);
7692 if (i
!= cpumask_first(this_sibling_map
))
7695 init_sched_build_groups(this_sibling_map
, cpu_map
,
7697 send_covered
, tmpmask
);
7701 #ifdef CONFIG_SCHED_MC
7702 /* Set up multi-core groups */
7703 for_each_cpu(i
, cpu_map
) {
7704 cpumask_and(this_core_map
, cpu_coregroup_mask(i
), cpu_map
);
7705 if (i
!= cpumask_first(this_core_map
))
7708 init_sched_build_groups(this_core_map
, cpu_map
,
7710 send_covered
, tmpmask
);
7714 /* Set up physical groups */
7715 for (i
= 0; i
< nr_node_ids
; i
++) {
7716 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
7717 if (cpumask_empty(nodemask
))
7720 init_sched_build_groups(nodemask
, cpu_map
,
7722 send_covered
, tmpmask
);
7726 /* Set up node groups */
7728 init_sched_build_groups(cpu_map
, cpu_map
,
7729 &cpu_to_allnodes_group
,
7730 send_covered
, tmpmask
);
7733 for (i
= 0; i
< nr_node_ids
; i
++) {
7734 /* Set up node groups */
7735 struct sched_group
*sg
, *prev
;
7738 cpumask_clear(covered
);
7739 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
7740 if (cpumask_empty(nodemask
)) {
7741 sched_group_nodes
[i
] = NULL
;
7745 sched_domain_node_span(i
, domainspan
);
7746 cpumask_and(domainspan
, domainspan
, cpu_map
);
7748 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
7751 printk(KERN_WARNING
"Can not alloc domain group for "
7755 sched_group_nodes
[i
] = sg
;
7756 for_each_cpu(j
, nodemask
) {
7757 struct sched_domain
*sd
;
7759 sd
= &per_cpu(node_domains
, j
).sd
;
7762 sg
->__cpu_power
= 0;
7763 cpumask_copy(sched_group_cpus(sg
), nodemask
);
7765 cpumask_or(covered
, covered
, nodemask
);
7768 for (j
= 0; j
< nr_node_ids
; j
++) {
7769 int n
= (i
+ j
) % nr_node_ids
;
7771 cpumask_complement(notcovered
, covered
);
7772 cpumask_and(tmpmask
, notcovered
, cpu_map
);
7773 cpumask_and(tmpmask
, tmpmask
, domainspan
);
7774 if (cpumask_empty(tmpmask
))
7777 cpumask_and(tmpmask
, tmpmask
, cpumask_of_node(n
));
7778 if (cpumask_empty(tmpmask
))
7781 sg
= kmalloc_node(sizeof(struct sched_group
) +
7786 "Can not alloc domain group for node %d\n", j
);
7789 sg
->__cpu_power
= 0;
7790 cpumask_copy(sched_group_cpus(sg
), tmpmask
);
7791 sg
->next
= prev
->next
;
7792 cpumask_or(covered
, covered
, tmpmask
);
7799 /* Calculate CPU power for physical packages and nodes */
7800 #ifdef CONFIG_SCHED_SMT
7801 for_each_cpu(i
, cpu_map
) {
7802 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
).sd
;
7804 init_sched_groups_power(i
, sd
);
7807 #ifdef CONFIG_SCHED_MC
7808 for_each_cpu(i
, cpu_map
) {
7809 struct sched_domain
*sd
= &per_cpu(core_domains
, i
).sd
;
7811 init_sched_groups_power(i
, sd
);
7815 for_each_cpu(i
, cpu_map
) {
7816 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
).sd
;
7818 init_sched_groups_power(i
, sd
);
7822 for (i
= 0; i
< nr_node_ids
; i
++)
7823 init_numa_sched_groups_power(sched_group_nodes
[i
]);
7826 struct sched_group
*sg
;
7828 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
7830 init_numa_sched_groups_power(sg
);
7834 /* Attach the domains */
7835 for_each_cpu(i
, cpu_map
) {
7836 struct sched_domain
*sd
;
7837 #ifdef CONFIG_SCHED_SMT
7838 sd
= &per_cpu(cpu_domains
, i
).sd
;
7839 #elif defined(CONFIG_SCHED_MC)
7840 sd
= &per_cpu(core_domains
, i
).sd
;
7842 sd
= &per_cpu(phys_domains
, i
).sd
;
7844 cpu_attach_domain(sd
, rd
, i
);
7850 free_cpumask_var(tmpmask
);
7852 free_cpumask_var(send_covered
);
7854 free_cpumask_var(this_core_map
);
7855 free_this_sibling_map
:
7856 free_cpumask_var(this_sibling_map
);
7858 free_cpumask_var(nodemask
);
7861 free_cpumask_var(notcovered
);
7863 free_cpumask_var(covered
);
7865 free_cpumask_var(domainspan
);
7872 kfree(sched_group_nodes
);
7878 free_sched_groups(cpu_map
, tmpmask
);
7879 free_rootdomain(rd
);
7884 static int build_sched_domains(const struct cpumask
*cpu_map
)
7886 return __build_sched_domains(cpu_map
, NULL
);
7889 static struct cpumask
*doms_cur
; /* current sched domains */
7890 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7891 static struct sched_domain_attr
*dattr_cur
;
7892 /* attribues of custom domains in 'doms_cur' */
7895 * Special case: If a kmalloc of a doms_cur partition (array of
7896 * cpumask) fails, then fallback to a single sched domain,
7897 * as determined by the single cpumask fallback_doms.
7899 static cpumask_var_t fallback_doms
;
7902 * arch_update_cpu_topology lets virtualized architectures update the
7903 * cpu core maps. It is supposed to return 1 if the topology changed
7904 * or 0 if it stayed the same.
7906 int __attribute__((weak
)) arch_update_cpu_topology(void)
7912 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7913 * For now this just excludes isolated cpus, but could be used to
7914 * exclude other special cases in the future.
7916 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
7920 arch_update_cpu_topology();
7922 doms_cur
= kmalloc(cpumask_size(), GFP_KERNEL
);
7924 doms_cur
= fallback_doms
;
7925 cpumask_andnot(doms_cur
, cpu_map
, cpu_isolated_map
);
7927 err
= build_sched_domains(doms_cur
);
7928 register_sched_domain_sysctl();
7933 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
7934 struct cpumask
*tmpmask
)
7936 free_sched_groups(cpu_map
, tmpmask
);
7940 * Detach sched domains from a group of cpus specified in cpu_map
7941 * These cpus will now be attached to the NULL domain
7943 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7945 /* Save because hotplug lock held. */
7946 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
7949 for_each_cpu(i
, cpu_map
)
7950 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7951 synchronize_sched();
7952 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
7955 /* handle null as "default" */
7956 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7957 struct sched_domain_attr
*new, int idx_new
)
7959 struct sched_domain_attr tmp
;
7966 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7967 new ? (new + idx_new
) : &tmp
,
7968 sizeof(struct sched_domain_attr
));
7972 * Partition sched domains as specified by the 'ndoms_new'
7973 * cpumasks in the array doms_new[] of cpumasks. This compares
7974 * doms_new[] to the current sched domain partitioning, doms_cur[].
7975 * It destroys each deleted domain and builds each new domain.
7977 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
7978 * The masks don't intersect (don't overlap.) We should setup one
7979 * sched domain for each mask. CPUs not in any of the cpumasks will
7980 * not be load balanced. If the same cpumask appears both in the
7981 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7984 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7985 * ownership of it and will kfree it when done with it. If the caller
7986 * failed the kmalloc call, then it can pass in doms_new == NULL &&
7987 * ndoms_new == 1, and partition_sched_domains() will fallback to
7988 * the single partition 'fallback_doms', it also forces the domains
7991 * If doms_new == NULL it will be replaced with cpu_online_mask.
7992 * ndoms_new == 0 is a special case for destroying existing domains,
7993 * and it will not create the default domain.
7995 * Call with hotplug lock held
7997 /* FIXME: Change to struct cpumask *doms_new[] */
7998 void partition_sched_domains(int ndoms_new
, struct cpumask
*doms_new
,
7999 struct sched_domain_attr
*dattr_new
)
8004 mutex_lock(&sched_domains_mutex
);
8006 /* always unregister in case we don't destroy any domains */
8007 unregister_sched_domain_sysctl();
8009 /* Let architecture update cpu core mappings. */
8010 new_topology
= arch_update_cpu_topology();
8012 n
= doms_new
? ndoms_new
: 0;
8014 /* Destroy deleted domains */
8015 for (i
= 0; i
< ndoms_cur
; i
++) {
8016 for (j
= 0; j
< n
&& !new_topology
; j
++) {
8017 if (cpumask_equal(&doms_cur
[i
], &doms_new
[j
])
8018 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
8021 /* no match - a current sched domain not in new doms_new[] */
8022 detach_destroy_domains(doms_cur
+ i
);
8027 if (doms_new
== NULL
) {
8029 doms_new
= fallback_doms
;
8030 cpumask_andnot(&doms_new
[0], cpu_online_mask
, cpu_isolated_map
);
8031 WARN_ON_ONCE(dattr_new
);
8034 /* Build new domains */
8035 for (i
= 0; i
< ndoms_new
; i
++) {
8036 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
8037 if (cpumask_equal(&doms_new
[i
], &doms_cur
[j
])
8038 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
8041 /* no match - add a new doms_new */
8042 __build_sched_domains(doms_new
+ i
,
8043 dattr_new
? dattr_new
+ i
: NULL
);
8048 /* Remember the new sched domains */
8049 if (doms_cur
!= fallback_doms
)
8051 kfree(dattr_cur
); /* kfree(NULL) is safe */
8052 doms_cur
= doms_new
;
8053 dattr_cur
= dattr_new
;
8054 ndoms_cur
= ndoms_new
;
8056 register_sched_domain_sysctl();
8058 mutex_unlock(&sched_domains_mutex
);
8061 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8062 static void arch_reinit_sched_domains(void)
8066 /* Destroy domains first to force the rebuild */
8067 partition_sched_domains(0, NULL
, NULL
);
8069 rebuild_sched_domains();
8073 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
8075 unsigned int level
= 0;
8077 if (sscanf(buf
, "%u", &level
) != 1)
8081 * level is always be positive so don't check for
8082 * level < POWERSAVINGS_BALANCE_NONE which is 0
8083 * What happens on 0 or 1 byte write,
8084 * need to check for count as well?
8087 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
8091 sched_smt_power_savings
= level
;
8093 sched_mc_power_savings
= level
;
8095 arch_reinit_sched_domains();
8100 #ifdef CONFIG_SCHED_MC
8101 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
8104 return sprintf(page
, "%u\n", sched_mc_power_savings
);
8106 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
8107 const char *buf
, size_t count
)
8109 return sched_power_savings_store(buf
, count
, 0);
8111 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
8112 sched_mc_power_savings_show
,
8113 sched_mc_power_savings_store
);
8116 #ifdef CONFIG_SCHED_SMT
8117 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
8120 return sprintf(page
, "%u\n", sched_smt_power_savings
);
8122 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
8123 const char *buf
, size_t count
)
8125 return sched_power_savings_store(buf
, count
, 1);
8127 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
8128 sched_smt_power_savings_show
,
8129 sched_smt_power_savings_store
);
8132 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
8136 #ifdef CONFIG_SCHED_SMT
8138 err
= sysfs_create_file(&cls
->kset
.kobj
,
8139 &attr_sched_smt_power_savings
.attr
);
8141 #ifdef CONFIG_SCHED_MC
8142 if (!err
&& mc_capable())
8143 err
= sysfs_create_file(&cls
->kset
.kobj
,
8144 &attr_sched_mc_power_savings
.attr
);
8148 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8150 #ifndef CONFIG_CPUSETS
8152 * Add online and remove offline CPUs from the scheduler domains.
8153 * When cpusets are enabled they take over this function.
8155 static int update_sched_domains(struct notifier_block
*nfb
,
8156 unsigned long action
, void *hcpu
)
8160 case CPU_ONLINE_FROZEN
:
8162 case CPU_DEAD_FROZEN
:
8163 partition_sched_domains(1, NULL
, NULL
);
8172 static int update_runtime(struct notifier_block
*nfb
,
8173 unsigned long action
, void *hcpu
)
8175 int cpu
= (int)(long)hcpu
;
8178 case CPU_DOWN_PREPARE
:
8179 case CPU_DOWN_PREPARE_FROZEN
:
8180 disable_runtime(cpu_rq(cpu
));
8183 case CPU_DOWN_FAILED
:
8184 case CPU_DOWN_FAILED_FROZEN
:
8186 case CPU_ONLINE_FROZEN
:
8187 enable_runtime(cpu_rq(cpu
));
8195 void __init
sched_init_smp(void)
8197 cpumask_var_t non_isolated_cpus
;
8199 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
8201 #if defined(CONFIG_NUMA)
8202 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
8204 BUG_ON(sched_group_nodes_bycpu
== NULL
);
8207 mutex_lock(&sched_domains_mutex
);
8208 arch_init_sched_domains(cpu_online_mask
);
8209 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
8210 if (cpumask_empty(non_isolated_cpus
))
8211 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
8212 mutex_unlock(&sched_domains_mutex
);
8215 #ifndef CONFIG_CPUSETS
8216 /* XXX: Theoretical race here - CPU may be hotplugged now */
8217 hotcpu_notifier(update_sched_domains
, 0);
8220 /* RT runtime code needs to handle some hotplug events */
8221 hotcpu_notifier(update_runtime
, 0);
8225 /* Move init over to a non-isolated CPU */
8226 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
8228 sched_init_granularity();
8229 free_cpumask_var(non_isolated_cpus
);
8231 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
8232 init_sched_rt_class();
8235 void __init
sched_init_smp(void)
8237 sched_init_granularity();
8239 #endif /* CONFIG_SMP */
8241 int in_sched_functions(unsigned long addr
)
8243 return in_lock_functions(addr
) ||
8244 (addr
>= (unsigned long)__sched_text_start
8245 && addr
< (unsigned long)__sched_text_end
);
8248 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
8250 cfs_rq
->tasks_timeline
= RB_ROOT
;
8251 INIT_LIST_HEAD(&cfs_rq
->tasks
);
8252 #ifdef CONFIG_FAIR_GROUP_SCHED
8255 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
8258 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
8260 struct rt_prio_array
*array
;
8263 array
= &rt_rq
->active
;
8264 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
8265 INIT_LIST_HEAD(array
->queue
+ i
);
8266 __clear_bit(i
, array
->bitmap
);
8268 /* delimiter for bitsearch: */
8269 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
8271 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8272 rt_rq
->highest_prio
= MAX_RT_PRIO
;
8275 rt_rq
->rt_nr_migratory
= 0;
8276 rt_rq
->overloaded
= 0;
8280 rt_rq
->rt_throttled
= 0;
8281 rt_rq
->rt_runtime
= 0;
8282 spin_lock_init(&rt_rq
->rt_runtime_lock
);
8284 #ifdef CONFIG_RT_GROUP_SCHED
8285 rt_rq
->rt_nr_boosted
= 0;
8290 #ifdef CONFIG_FAIR_GROUP_SCHED
8291 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8292 struct sched_entity
*se
, int cpu
, int add
,
8293 struct sched_entity
*parent
)
8295 struct rq
*rq
= cpu_rq(cpu
);
8296 tg
->cfs_rq
[cpu
] = cfs_rq
;
8297 init_cfs_rq(cfs_rq
, rq
);
8300 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
8303 /* se could be NULL for init_task_group */
8308 se
->cfs_rq
= &rq
->cfs
;
8310 se
->cfs_rq
= parent
->my_q
;
8313 se
->load
.weight
= tg
->shares
;
8314 se
->load
.inv_weight
= 0;
8315 se
->parent
= parent
;
8319 #ifdef CONFIG_RT_GROUP_SCHED
8320 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
8321 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
8322 struct sched_rt_entity
*parent
)
8324 struct rq
*rq
= cpu_rq(cpu
);
8326 tg
->rt_rq
[cpu
] = rt_rq
;
8327 init_rt_rq(rt_rq
, rq
);
8329 rt_rq
->rt_se
= rt_se
;
8330 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8332 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
8334 tg
->rt_se
[cpu
] = rt_se
;
8339 rt_se
->rt_rq
= &rq
->rt
;
8341 rt_se
->rt_rq
= parent
->my_q
;
8343 rt_se
->my_q
= rt_rq
;
8344 rt_se
->parent
= parent
;
8345 INIT_LIST_HEAD(&rt_se
->run_list
);
8349 void __init
sched_init(void)
8352 unsigned long alloc_size
= 0, ptr
;
8354 #ifdef CONFIG_FAIR_GROUP_SCHED
8355 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8357 #ifdef CONFIG_RT_GROUP_SCHED
8358 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8360 #ifdef CONFIG_USER_SCHED
8364 * As sched_init() is called before page_alloc is setup,
8365 * we use alloc_bootmem().
8368 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
8370 #ifdef CONFIG_FAIR_GROUP_SCHED
8371 init_task_group
.se
= (struct sched_entity
**)ptr
;
8372 ptr
+= nr_cpu_ids
* sizeof(void **);
8374 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8375 ptr
+= nr_cpu_ids
* sizeof(void **);
8377 #ifdef CONFIG_USER_SCHED
8378 root_task_group
.se
= (struct sched_entity
**)ptr
;
8379 ptr
+= nr_cpu_ids
* sizeof(void **);
8381 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8382 ptr
+= nr_cpu_ids
* sizeof(void **);
8383 #endif /* CONFIG_USER_SCHED */
8384 #endif /* CONFIG_FAIR_GROUP_SCHED */
8385 #ifdef CONFIG_RT_GROUP_SCHED
8386 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8387 ptr
+= nr_cpu_ids
* sizeof(void **);
8389 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8390 ptr
+= nr_cpu_ids
* sizeof(void **);
8392 #ifdef CONFIG_USER_SCHED
8393 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8394 ptr
+= nr_cpu_ids
* sizeof(void **);
8396 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8397 ptr
+= nr_cpu_ids
* sizeof(void **);
8398 #endif /* CONFIG_USER_SCHED */
8399 #endif /* CONFIG_RT_GROUP_SCHED */
8403 init_defrootdomain();
8406 init_rt_bandwidth(&def_rt_bandwidth
,
8407 global_rt_period(), global_rt_runtime());
8409 #ifdef CONFIG_RT_GROUP_SCHED
8410 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
8411 global_rt_period(), global_rt_runtime());
8412 #ifdef CONFIG_USER_SCHED
8413 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8414 global_rt_period(), RUNTIME_INF
);
8415 #endif /* CONFIG_USER_SCHED */
8416 #endif /* CONFIG_RT_GROUP_SCHED */
8418 #ifdef CONFIG_GROUP_SCHED
8419 list_add(&init_task_group
.list
, &task_groups
);
8420 INIT_LIST_HEAD(&init_task_group
.children
);
8422 #ifdef CONFIG_USER_SCHED
8423 INIT_LIST_HEAD(&root_task_group
.children
);
8424 init_task_group
.parent
= &root_task_group
;
8425 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
8426 #endif /* CONFIG_USER_SCHED */
8427 #endif /* CONFIG_GROUP_SCHED */
8429 for_each_possible_cpu(i
) {
8433 spin_lock_init(&rq
->lock
);
8435 init_cfs_rq(&rq
->cfs
, rq
);
8436 init_rt_rq(&rq
->rt
, rq
);
8437 #ifdef CONFIG_FAIR_GROUP_SCHED
8438 init_task_group
.shares
= init_task_group_load
;
8439 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8440 #ifdef CONFIG_CGROUP_SCHED
8442 * How much cpu bandwidth does init_task_group get?
8444 * In case of task-groups formed thr' the cgroup filesystem, it
8445 * gets 100% of the cpu resources in the system. This overall
8446 * system cpu resource is divided among the tasks of
8447 * init_task_group and its child task-groups in a fair manner,
8448 * based on each entity's (task or task-group's) weight
8449 * (se->load.weight).
8451 * In other words, if init_task_group has 10 tasks of weight
8452 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8453 * then A0's share of the cpu resource is:
8455 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8457 * We achieve this by letting init_task_group's tasks sit
8458 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8460 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
8461 #elif defined CONFIG_USER_SCHED
8462 root_task_group
.shares
= NICE_0_LOAD
;
8463 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
8465 * In case of task-groups formed thr' the user id of tasks,
8466 * init_task_group represents tasks belonging to root user.
8467 * Hence it forms a sibling of all subsequent groups formed.
8468 * In this case, init_task_group gets only a fraction of overall
8469 * system cpu resource, based on the weight assigned to root
8470 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8471 * by letting tasks of init_task_group sit in a separate cfs_rq
8472 * (init_cfs_rq) and having one entity represent this group of
8473 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8475 init_tg_cfs_entry(&init_task_group
,
8476 &per_cpu(init_cfs_rq
, i
),
8477 &per_cpu(init_sched_entity
, i
), i
, 1,
8478 root_task_group
.se
[i
]);
8481 #endif /* CONFIG_FAIR_GROUP_SCHED */
8483 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8484 #ifdef CONFIG_RT_GROUP_SCHED
8485 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
8486 #ifdef CONFIG_CGROUP_SCHED
8487 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
8488 #elif defined CONFIG_USER_SCHED
8489 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
8490 init_tg_rt_entry(&init_task_group
,
8491 &per_cpu(init_rt_rq
, i
),
8492 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
8493 root_task_group
.rt_se
[i
]);
8497 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
8498 rq
->cpu_load
[j
] = 0;
8502 rq
->active_balance
= 0;
8503 rq
->next_balance
= jiffies
;
8507 rq
->migration_thread
= NULL
;
8508 INIT_LIST_HEAD(&rq
->migration_queue
);
8509 rq_attach_root(rq
, &def_root_domain
);
8512 atomic_set(&rq
->nr_iowait
, 0);
8515 set_load_weight(&init_task
);
8517 #ifdef CONFIG_PREEMPT_NOTIFIERS
8518 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
8522 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8525 #ifdef CONFIG_RT_MUTEXES
8526 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
8530 * The boot idle thread does lazy MMU switching as well:
8532 atomic_inc(&init_mm
.mm_count
);
8533 enter_lazy_tlb(&init_mm
, current
);
8536 * Make us the idle thread. Technically, schedule() should not be
8537 * called from this thread, however somewhere below it might be,
8538 * but because we are the idle thread, we just pick up running again
8539 * when this runqueue becomes "idle".
8541 init_idle(current
, smp_processor_id());
8543 * During early bootup we pretend to be a normal task:
8545 current
->sched_class
= &fair_sched_class
;
8547 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8548 alloc_bootmem_cpumask_var(&nohz_cpu_mask
);
8551 alloc_bootmem_cpumask_var(&nohz
.cpu_mask
);
8553 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
8556 scheduler_running
= 1;
8559 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8560 void __might_sleep(char *file
, int line
)
8563 static unsigned long prev_jiffy
; /* ratelimiting */
8565 if ((!in_atomic() && !irqs_disabled()) ||
8566 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
8568 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8570 prev_jiffy
= jiffies
;
8573 "BUG: sleeping function called from invalid context at %s:%d\n",
8576 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8577 in_atomic(), irqs_disabled(),
8578 current
->pid
, current
->comm
);
8580 debug_show_held_locks(current
);
8581 if (irqs_disabled())
8582 print_irqtrace_events(current
);
8586 EXPORT_SYMBOL(__might_sleep
);
8589 #ifdef CONFIG_MAGIC_SYSRQ
8590 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8594 update_rq_clock(rq
);
8595 on_rq
= p
->se
.on_rq
;
8597 deactivate_task(rq
, p
, 0);
8598 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8600 activate_task(rq
, p
, 0);
8601 resched_task(rq
->curr
);
8605 void normalize_rt_tasks(void)
8607 struct task_struct
*g
, *p
;
8608 unsigned long flags
;
8611 read_lock_irqsave(&tasklist_lock
, flags
);
8612 do_each_thread(g
, p
) {
8614 * Only normalize user tasks:
8619 p
->se
.exec_start
= 0;
8620 #ifdef CONFIG_SCHEDSTATS
8621 p
->se
.wait_start
= 0;
8622 p
->se
.sleep_start
= 0;
8623 p
->se
.block_start
= 0;
8628 * Renice negative nice level userspace
8631 if (TASK_NICE(p
) < 0 && p
->mm
)
8632 set_user_nice(p
, 0);
8636 spin_lock(&p
->pi_lock
);
8637 rq
= __task_rq_lock(p
);
8639 normalize_task(rq
, p
);
8641 __task_rq_unlock(rq
);
8642 spin_unlock(&p
->pi_lock
);
8643 } while_each_thread(g
, p
);
8645 read_unlock_irqrestore(&tasklist_lock
, flags
);
8648 #endif /* CONFIG_MAGIC_SYSRQ */
8652 * These functions are only useful for the IA64 MCA handling.
8654 * They can only be called when the whole system has been
8655 * stopped - every CPU needs to be quiescent, and no scheduling
8656 * activity can take place. Using them for anything else would
8657 * be a serious bug, and as a result, they aren't even visible
8658 * under any other configuration.
8662 * curr_task - return the current task for a given cpu.
8663 * @cpu: the processor in question.
8665 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8667 struct task_struct
*curr_task(int cpu
)
8669 return cpu_curr(cpu
);
8673 * set_curr_task - set the current task for a given cpu.
8674 * @cpu: the processor in question.
8675 * @p: the task pointer to set.
8677 * Description: This function must only be used when non-maskable interrupts
8678 * are serviced on a separate stack. It allows the architecture to switch the
8679 * notion of the current task on a cpu in a non-blocking manner. This function
8680 * must be called with all CPU's synchronized, and interrupts disabled, the
8681 * and caller must save the original value of the current task (see
8682 * curr_task() above) and restore that value before reenabling interrupts and
8683 * re-starting the system.
8685 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8687 void set_curr_task(int cpu
, struct task_struct
*p
)
8694 #ifdef CONFIG_FAIR_GROUP_SCHED
8695 static void free_fair_sched_group(struct task_group
*tg
)
8699 for_each_possible_cpu(i
) {
8701 kfree(tg
->cfs_rq
[i
]);
8711 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8713 struct cfs_rq
*cfs_rq
;
8714 struct sched_entity
*se
;
8718 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8721 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8725 tg
->shares
= NICE_0_LOAD
;
8727 for_each_possible_cpu(i
) {
8730 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
8731 GFP_KERNEL
, cpu_to_node(i
));
8735 se
= kzalloc_node(sizeof(struct sched_entity
),
8736 GFP_KERNEL
, cpu_to_node(i
));
8740 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
8749 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8751 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8752 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8755 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8757 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8759 #else /* !CONFG_FAIR_GROUP_SCHED */
8760 static inline void free_fair_sched_group(struct task_group
*tg
)
8765 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8770 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8774 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8777 #endif /* CONFIG_FAIR_GROUP_SCHED */
8779 #ifdef CONFIG_RT_GROUP_SCHED
8780 static void free_rt_sched_group(struct task_group
*tg
)
8784 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8786 for_each_possible_cpu(i
) {
8788 kfree(tg
->rt_rq
[i
]);
8790 kfree(tg
->rt_se
[i
]);
8798 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8800 struct rt_rq
*rt_rq
;
8801 struct sched_rt_entity
*rt_se
;
8805 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8808 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8812 init_rt_bandwidth(&tg
->rt_bandwidth
,
8813 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8815 for_each_possible_cpu(i
) {
8818 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
8819 GFP_KERNEL
, cpu_to_node(i
));
8823 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
8824 GFP_KERNEL
, cpu_to_node(i
));
8828 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
8837 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8839 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8840 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8843 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8845 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8847 #else /* !CONFIG_RT_GROUP_SCHED */
8848 static inline void free_rt_sched_group(struct task_group
*tg
)
8853 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8858 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8862 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8865 #endif /* CONFIG_RT_GROUP_SCHED */
8867 #ifdef CONFIG_GROUP_SCHED
8868 static void free_sched_group(struct task_group
*tg
)
8870 free_fair_sched_group(tg
);
8871 free_rt_sched_group(tg
);
8875 /* allocate runqueue etc for a new task group */
8876 struct task_group
*sched_create_group(struct task_group
*parent
)
8878 struct task_group
*tg
;
8879 unsigned long flags
;
8882 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8884 return ERR_PTR(-ENOMEM
);
8886 if (!alloc_fair_sched_group(tg
, parent
))
8889 if (!alloc_rt_sched_group(tg
, parent
))
8892 spin_lock_irqsave(&task_group_lock
, flags
);
8893 for_each_possible_cpu(i
) {
8894 register_fair_sched_group(tg
, i
);
8895 register_rt_sched_group(tg
, i
);
8897 list_add_rcu(&tg
->list
, &task_groups
);
8899 WARN_ON(!parent
); /* root should already exist */
8901 tg
->parent
= parent
;
8902 INIT_LIST_HEAD(&tg
->children
);
8903 list_add_rcu(&tg
->siblings
, &parent
->children
);
8904 spin_unlock_irqrestore(&task_group_lock
, flags
);
8909 free_sched_group(tg
);
8910 return ERR_PTR(-ENOMEM
);
8913 /* rcu callback to free various structures associated with a task group */
8914 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8916 /* now it should be safe to free those cfs_rqs */
8917 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8920 /* Destroy runqueue etc associated with a task group */
8921 void sched_destroy_group(struct task_group
*tg
)
8923 unsigned long flags
;
8926 spin_lock_irqsave(&task_group_lock
, flags
);
8927 for_each_possible_cpu(i
) {
8928 unregister_fair_sched_group(tg
, i
);
8929 unregister_rt_sched_group(tg
, i
);
8931 list_del_rcu(&tg
->list
);
8932 list_del_rcu(&tg
->siblings
);
8933 spin_unlock_irqrestore(&task_group_lock
, flags
);
8935 /* wait for possible concurrent references to cfs_rqs complete */
8936 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8939 /* change task's runqueue when it moves between groups.
8940 * The caller of this function should have put the task in its new group
8941 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8942 * reflect its new group.
8944 void sched_move_task(struct task_struct
*tsk
)
8947 unsigned long flags
;
8950 rq
= task_rq_lock(tsk
, &flags
);
8952 update_rq_clock(rq
);
8954 running
= task_current(rq
, tsk
);
8955 on_rq
= tsk
->se
.on_rq
;
8958 dequeue_task(rq
, tsk
, 0);
8959 if (unlikely(running
))
8960 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8962 set_task_rq(tsk
, task_cpu(tsk
));
8964 #ifdef CONFIG_FAIR_GROUP_SCHED
8965 if (tsk
->sched_class
->moved_group
)
8966 tsk
->sched_class
->moved_group(tsk
);
8969 if (unlikely(running
))
8970 tsk
->sched_class
->set_curr_task(rq
);
8972 enqueue_task(rq
, tsk
, 0);
8974 task_rq_unlock(rq
, &flags
);
8976 #endif /* CONFIG_GROUP_SCHED */
8978 #ifdef CONFIG_FAIR_GROUP_SCHED
8979 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8981 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8986 dequeue_entity(cfs_rq
, se
, 0);
8988 se
->load
.weight
= shares
;
8989 se
->load
.inv_weight
= 0;
8992 enqueue_entity(cfs_rq
, se
, 0);
8995 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8997 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8998 struct rq
*rq
= cfs_rq
->rq
;
8999 unsigned long flags
;
9001 spin_lock_irqsave(&rq
->lock
, flags
);
9002 __set_se_shares(se
, shares
);
9003 spin_unlock_irqrestore(&rq
->lock
, flags
);
9006 static DEFINE_MUTEX(shares_mutex
);
9008 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
9011 unsigned long flags
;
9014 * We can't change the weight of the root cgroup.
9019 if (shares
< MIN_SHARES
)
9020 shares
= MIN_SHARES
;
9021 else if (shares
> MAX_SHARES
)
9022 shares
= MAX_SHARES
;
9024 mutex_lock(&shares_mutex
);
9025 if (tg
->shares
== shares
)
9028 spin_lock_irqsave(&task_group_lock
, flags
);
9029 for_each_possible_cpu(i
)
9030 unregister_fair_sched_group(tg
, i
);
9031 list_del_rcu(&tg
->siblings
);
9032 spin_unlock_irqrestore(&task_group_lock
, flags
);
9034 /* wait for any ongoing reference to this group to finish */
9035 synchronize_sched();
9038 * Now we are free to modify the group's share on each cpu
9039 * w/o tripping rebalance_share or load_balance_fair.
9041 tg
->shares
= shares
;
9042 for_each_possible_cpu(i
) {
9046 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
9047 set_se_shares(tg
->se
[i
], shares
);
9051 * Enable load balance activity on this group, by inserting it back on
9052 * each cpu's rq->leaf_cfs_rq_list.
9054 spin_lock_irqsave(&task_group_lock
, flags
);
9055 for_each_possible_cpu(i
)
9056 register_fair_sched_group(tg
, i
);
9057 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
9058 spin_unlock_irqrestore(&task_group_lock
, flags
);
9060 mutex_unlock(&shares_mutex
);
9064 unsigned long sched_group_shares(struct task_group
*tg
)
9070 #ifdef CONFIG_RT_GROUP_SCHED
9072 * Ensure that the real time constraints are schedulable.
9074 static DEFINE_MUTEX(rt_constraints_mutex
);
9076 static unsigned long to_ratio(u64 period
, u64 runtime
)
9078 if (runtime
== RUNTIME_INF
)
9081 return div64_u64(runtime
<< 20, period
);
9084 /* Must be called with tasklist_lock held */
9085 static inline int tg_has_rt_tasks(struct task_group
*tg
)
9087 struct task_struct
*g
, *p
;
9089 do_each_thread(g
, p
) {
9090 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
9092 } while_each_thread(g
, p
);
9097 struct rt_schedulable_data
{
9098 struct task_group
*tg
;
9103 static int tg_schedulable(struct task_group
*tg
, void *data
)
9105 struct rt_schedulable_data
*d
= data
;
9106 struct task_group
*child
;
9107 unsigned long total
, sum
= 0;
9108 u64 period
, runtime
;
9110 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9111 runtime
= tg
->rt_bandwidth
.rt_runtime
;
9114 period
= d
->rt_period
;
9115 runtime
= d
->rt_runtime
;
9118 #ifdef CONFIG_USER_SCHED
9119 if (tg
== &root_task_group
) {
9120 period
= global_rt_period();
9121 runtime
= global_rt_runtime();
9126 * Cannot have more runtime than the period.
9128 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9132 * Ensure we don't starve existing RT tasks.
9134 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
9137 total
= to_ratio(period
, runtime
);
9140 * Nobody can have more than the global setting allows.
9142 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
9146 * The sum of our children's runtime should not exceed our own.
9148 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
9149 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
9150 runtime
= child
->rt_bandwidth
.rt_runtime
;
9152 if (child
== d
->tg
) {
9153 period
= d
->rt_period
;
9154 runtime
= d
->rt_runtime
;
9157 sum
+= to_ratio(period
, runtime
);
9166 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
9168 struct rt_schedulable_data data
= {
9170 .rt_period
= period
,
9171 .rt_runtime
= runtime
,
9174 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
9177 static int tg_set_bandwidth(struct task_group
*tg
,
9178 u64 rt_period
, u64 rt_runtime
)
9182 mutex_lock(&rt_constraints_mutex
);
9183 read_lock(&tasklist_lock
);
9184 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
9188 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
9189 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
9190 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
9192 for_each_possible_cpu(i
) {
9193 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
9195 spin_lock(&rt_rq
->rt_runtime_lock
);
9196 rt_rq
->rt_runtime
= rt_runtime
;
9197 spin_unlock(&rt_rq
->rt_runtime_lock
);
9199 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
9201 read_unlock(&tasklist_lock
);
9202 mutex_unlock(&rt_constraints_mutex
);
9207 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
9209 u64 rt_runtime
, rt_period
;
9211 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9212 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
9213 if (rt_runtime_us
< 0)
9214 rt_runtime
= RUNTIME_INF
;
9216 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9219 long sched_group_rt_runtime(struct task_group
*tg
)
9223 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
9226 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
9227 do_div(rt_runtime_us
, NSEC_PER_USEC
);
9228 return rt_runtime_us
;
9231 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
9233 u64 rt_runtime
, rt_period
;
9235 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
9236 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9241 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9244 long sched_group_rt_period(struct task_group
*tg
)
9248 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9249 do_div(rt_period_us
, NSEC_PER_USEC
);
9250 return rt_period_us
;
9253 static int sched_rt_global_constraints(void)
9255 u64 runtime
, period
;
9258 if (sysctl_sched_rt_period
<= 0)
9261 runtime
= global_rt_runtime();
9262 period
= global_rt_period();
9265 * Sanity check on the sysctl variables.
9267 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9270 mutex_lock(&rt_constraints_mutex
);
9271 read_lock(&tasklist_lock
);
9272 ret
= __rt_schedulable(NULL
, 0, 0);
9273 read_unlock(&tasklist_lock
);
9274 mutex_unlock(&rt_constraints_mutex
);
9278 #else /* !CONFIG_RT_GROUP_SCHED */
9279 static int sched_rt_global_constraints(void)
9281 unsigned long flags
;
9284 if (sysctl_sched_rt_period
<= 0)
9287 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9288 for_each_possible_cpu(i
) {
9289 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
9291 spin_lock(&rt_rq
->rt_runtime_lock
);
9292 rt_rq
->rt_runtime
= global_rt_runtime();
9293 spin_unlock(&rt_rq
->rt_runtime_lock
);
9295 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9299 #endif /* CONFIG_RT_GROUP_SCHED */
9301 int sched_rt_handler(struct ctl_table
*table
, int write
,
9302 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
9306 int old_period
, old_runtime
;
9307 static DEFINE_MUTEX(mutex
);
9310 old_period
= sysctl_sched_rt_period
;
9311 old_runtime
= sysctl_sched_rt_runtime
;
9313 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
9315 if (!ret
&& write
) {
9316 ret
= sched_rt_global_constraints();
9318 sysctl_sched_rt_period
= old_period
;
9319 sysctl_sched_rt_runtime
= old_runtime
;
9321 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
9322 def_rt_bandwidth
.rt_period
=
9323 ns_to_ktime(global_rt_period());
9326 mutex_unlock(&mutex
);
9331 #ifdef CONFIG_CGROUP_SCHED
9333 /* return corresponding task_group object of a cgroup */
9334 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
9336 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
9337 struct task_group
, css
);
9340 static struct cgroup_subsys_state
*
9341 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9343 struct task_group
*tg
, *parent
;
9345 if (!cgrp
->parent
) {
9346 /* This is early initialization for the top cgroup */
9347 return &init_task_group
.css
;
9350 parent
= cgroup_tg(cgrp
->parent
);
9351 tg
= sched_create_group(parent
);
9353 return ERR_PTR(-ENOMEM
);
9359 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9361 struct task_group
*tg
= cgroup_tg(cgrp
);
9363 sched_destroy_group(tg
);
9367 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9368 struct task_struct
*tsk
)
9370 #ifdef CONFIG_RT_GROUP_SCHED
9371 /* Don't accept realtime tasks when there is no way for them to run */
9372 if (rt_task(tsk
) && cgroup_tg(cgrp
)->rt_bandwidth
.rt_runtime
== 0)
9375 /* We don't support RT-tasks being in separate groups */
9376 if (tsk
->sched_class
!= &fair_sched_class
)
9384 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9385 struct cgroup
*old_cont
, struct task_struct
*tsk
)
9387 sched_move_task(tsk
);
9390 #ifdef CONFIG_FAIR_GROUP_SCHED
9391 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9394 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
9397 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9399 struct task_group
*tg
= cgroup_tg(cgrp
);
9401 return (u64
) tg
->shares
;
9403 #endif /* CONFIG_FAIR_GROUP_SCHED */
9405 #ifdef CONFIG_RT_GROUP_SCHED
9406 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9409 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9412 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9414 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9417 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9420 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9423 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9425 return sched_group_rt_period(cgroup_tg(cgrp
));
9427 #endif /* CONFIG_RT_GROUP_SCHED */
9429 static struct cftype cpu_files
[] = {
9430 #ifdef CONFIG_FAIR_GROUP_SCHED
9433 .read_u64
= cpu_shares_read_u64
,
9434 .write_u64
= cpu_shares_write_u64
,
9437 #ifdef CONFIG_RT_GROUP_SCHED
9439 .name
= "rt_runtime_us",
9440 .read_s64
= cpu_rt_runtime_read
,
9441 .write_s64
= cpu_rt_runtime_write
,
9444 .name
= "rt_period_us",
9445 .read_u64
= cpu_rt_period_read_uint
,
9446 .write_u64
= cpu_rt_period_write_uint
,
9451 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
9453 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
9456 struct cgroup_subsys cpu_cgroup_subsys
= {
9458 .create
= cpu_cgroup_create
,
9459 .destroy
= cpu_cgroup_destroy
,
9460 .can_attach
= cpu_cgroup_can_attach
,
9461 .attach
= cpu_cgroup_attach
,
9462 .populate
= cpu_cgroup_populate
,
9463 .subsys_id
= cpu_cgroup_subsys_id
,
9467 #endif /* CONFIG_CGROUP_SCHED */
9469 #ifdef CONFIG_CGROUP_CPUACCT
9472 * CPU accounting code for task groups.
9474 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9475 * (balbir@in.ibm.com).
9478 /* track cpu usage of a group of tasks and its child groups */
9480 struct cgroup_subsys_state css
;
9481 /* cpuusage holds pointer to a u64-type object on every cpu */
9483 struct cpuacct
*parent
;
9486 struct cgroup_subsys cpuacct_subsys
;
9488 /* return cpu accounting group corresponding to this container */
9489 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
9491 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
9492 struct cpuacct
, css
);
9495 /* return cpu accounting group to which this task belongs */
9496 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
9498 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
9499 struct cpuacct
, css
);
9502 /* create a new cpu accounting group */
9503 static struct cgroup_subsys_state
*cpuacct_create(
9504 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9506 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
9509 return ERR_PTR(-ENOMEM
);
9511 ca
->cpuusage
= alloc_percpu(u64
);
9512 if (!ca
->cpuusage
) {
9514 return ERR_PTR(-ENOMEM
);
9518 ca
->parent
= cgroup_ca(cgrp
->parent
);
9523 /* destroy an existing cpu accounting group */
9525 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9527 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9529 free_percpu(ca
->cpuusage
);
9533 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
9535 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, cpu
);
9538 #ifndef CONFIG_64BIT
9540 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9542 spin_lock_irq(&cpu_rq(cpu
)->lock
);
9544 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9552 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
9554 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, cpu
);
9556 #ifndef CONFIG_64BIT
9558 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9560 spin_lock_irq(&cpu_rq(cpu
)->lock
);
9562 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9568 /* return total cpu usage (in nanoseconds) of a group */
9569 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9571 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9572 u64 totalcpuusage
= 0;
9575 for_each_present_cpu(i
)
9576 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
9578 return totalcpuusage
;
9581 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9584 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9593 for_each_present_cpu(i
)
9594 cpuacct_cpuusage_write(ca
, i
, 0);
9600 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
9603 struct cpuacct
*ca
= cgroup_ca(cgroup
);
9607 for_each_present_cpu(i
) {
9608 percpu
= cpuacct_cpuusage_read(ca
, i
);
9609 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
9611 seq_printf(m
, "\n");
9615 static struct cftype files
[] = {
9618 .read_u64
= cpuusage_read
,
9619 .write_u64
= cpuusage_write
,
9622 .name
= "usage_percpu",
9623 .read_seq_string
= cpuacct_percpu_seq_read
,
9628 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9630 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9634 * charge this task's execution time to its accounting group.
9636 * called with rq->lock held.
9638 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9643 if (!cpuacct_subsys
.active
)
9646 cpu
= task_cpu(tsk
);
9649 for (; ca
; ca
= ca
->parent
) {
9650 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, cpu
);
9651 *cpuusage
+= cputime
;
9655 struct cgroup_subsys cpuacct_subsys
= {
9657 .create
= cpuacct_create
,
9658 .destroy
= cpuacct_destroy
,
9659 .populate
= cpuacct_populate
,
9660 .subsys_id
= cpuacct_subsys_id
,
9662 #endif /* CONFIG_CGROUP_CPUACCT */