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 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
130 * Since cpu_power is a 'constant', we can use a reciprocal divide.
132 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
134 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
138 * Each time a sched group cpu_power is changed,
139 * we must compute its reciprocal value
141 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
143 sg
->__cpu_power
+= val
;
144 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
148 static inline int rt_policy(int policy
)
150 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
155 static inline int task_has_rt_policy(struct task_struct
*p
)
157 return rt_policy(p
->policy
);
161 * This is the priority-queue data structure of the RT scheduling class:
163 struct rt_prio_array
{
164 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
165 struct list_head queue
[MAX_RT_PRIO
];
168 struct rt_bandwidth
{
169 /* nests inside the rq lock: */
170 spinlock_t rt_runtime_lock
;
173 struct hrtimer rt_period_timer
;
176 static struct rt_bandwidth def_rt_bandwidth
;
178 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
180 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
182 struct rt_bandwidth
*rt_b
=
183 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
189 now
= hrtimer_cb_get_time(timer
);
190 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
195 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
198 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
202 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
204 rt_b
->rt_period
= ns_to_ktime(period
);
205 rt_b
->rt_runtime
= runtime
;
207 spin_lock_init(&rt_b
->rt_runtime_lock
);
209 hrtimer_init(&rt_b
->rt_period_timer
,
210 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
211 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
212 rt_b
->rt_period_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_UNLOCKED
;
215 static inline int rt_bandwidth_enabled(void)
217 return sysctl_sched_rt_runtime
>= 0;
220 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
224 if (rt_bandwidth_enabled() && rt_b
->rt_runtime
== RUNTIME_INF
)
227 if (hrtimer_active(&rt_b
->rt_period_timer
))
230 spin_lock(&rt_b
->rt_runtime_lock
);
232 if (hrtimer_active(&rt_b
->rt_period_timer
))
235 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
236 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
237 hrtimer_start_expires(&rt_b
->rt_period_timer
,
240 spin_unlock(&rt_b
->rt_runtime_lock
);
243 #ifdef CONFIG_RT_GROUP_SCHED
244 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
246 hrtimer_cancel(&rt_b
->rt_period_timer
);
251 * sched_domains_mutex serializes calls to arch_init_sched_domains,
252 * detach_destroy_domains and partition_sched_domains.
254 static DEFINE_MUTEX(sched_domains_mutex
);
256 #ifdef CONFIG_GROUP_SCHED
258 #include <linux/cgroup.h>
262 static LIST_HEAD(task_groups
);
264 /* task group related information */
266 #ifdef CONFIG_CGROUP_SCHED
267 struct cgroup_subsys_state css
;
270 #ifdef CONFIG_USER_SCHED
274 #ifdef CONFIG_FAIR_GROUP_SCHED
275 /* schedulable entities of this group on each cpu */
276 struct sched_entity
**se
;
277 /* runqueue "owned" by this group on each cpu */
278 struct cfs_rq
**cfs_rq
;
279 unsigned long shares
;
282 #ifdef CONFIG_RT_GROUP_SCHED
283 struct sched_rt_entity
**rt_se
;
284 struct rt_rq
**rt_rq
;
286 struct rt_bandwidth rt_bandwidth
;
290 struct list_head list
;
292 struct task_group
*parent
;
293 struct list_head siblings
;
294 struct list_head children
;
297 #ifdef CONFIG_USER_SCHED
299 /* Helper function to pass uid information to create_sched_user() */
300 void set_tg_uid(struct user_struct
*user
)
302 user
->tg
->uid
= user
->uid
;
307 * Every UID task group (including init_task_group aka UID-0) will
308 * be a child to this group.
310 struct task_group root_task_group
;
312 #ifdef CONFIG_FAIR_GROUP_SCHED
313 /* Default task group's sched entity on each cpu */
314 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
315 /* Default task group's cfs_rq on each cpu */
316 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
317 #endif /* CONFIG_FAIR_GROUP_SCHED */
319 #ifdef CONFIG_RT_GROUP_SCHED
320 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
321 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
322 #endif /* CONFIG_RT_GROUP_SCHED */
323 #else /* !CONFIG_USER_SCHED */
324 #define root_task_group init_task_group
325 #endif /* CONFIG_USER_SCHED */
327 /* task_group_lock serializes add/remove of task groups and also changes to
328 * a task group's cpu shares.
330 static DEFINE_SPINLOCK(task_group_lock
);
332 #ifdef CONFIG_FAIR_GROUP_SCHED
333 #ifdef CONFIG_USER_SCHED
334 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
335 #else /* !CONFIG_USER_SCHED */
336 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
337 #endif /* CONFIG_USER_SCHED */
340 * A weight of 0 or 1 can cause arithmetics problems.
341 * A weight of a cfs_rq is the sum of weights of which entities
342 * are queued on this cfs_rq, so a weight of a entity should not be
343 * too large, so as the shares value of a task group.
344 * (The default weight is 1024 - so there's no practical
345 * limitation from this.)
348 #define MAX_SHARES (1UL << 18)
350 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
353 /* Default task group.
354 * Every task in system belong to this group at bootup.
356 struct task_group init_task_group
;
358 /* return group to which a task belongs */
359 static inline struct task_group
*task_group(struct task_struct
*p
)
361 struct task_group
*tg
;
363 #ifdef CONFIG_USER_SCHED
365 #elif defined(CONFIG_CGROUP_SCHED)
366 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
367 struct task_group
, css
);
369 tg
= &init_task_group
;
374 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
375 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
377 #ifdef CONFIG_FAIR_GROUP_SCHED
378 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
379 p
->se
.parent
= task_group(p
)->se
[cpu
];
382 #ifdef CONFIG_RT_GROUP_SCHED
383 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
384 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
390 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
391 static inline struct task_group
*task_group(struct task_struct
*p
)
396 #endif /* CONFIG_GROUP_SCHED */
398 /* CFS-related fields in a runqueue */
400 struct load_weight load
;
401 unsigned long nr_running
;
406 struct rb_root tasks_timeline
;
407 struct rb_node
*rb_leftmost
;
409 struct list_head tasks
;
410 struct list_head
*balance_iterator
;
413 * 'curr' points to currently running entity on this cfs_rq.
414 * It is set to NULL otherwise (i.e when none are currently running).
416 struct sched_entity
*curr
, *next
, *last
;
418 unsigned int nr_spread_over
;
420 #ifdef CONFIG_FAIR_GROUP_SCHED
421 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
424 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
425 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
426 * (like users, containers etc.)
428 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
429 * list is used during load balance.
431 struct list_head leaf_cfs_rq_list
;
432 struct task_group
*tg
; /* group that "owns" this runqueue */
436 * the part of load.weight contributed by tasks
438 unsigned long task_weight
;
441 * h_load = weight * f(tg)
443 * Where f(tg) is the recursive weight fraction assigned to
446 unsigned long h_load
;
449 * this cpu's part of tg->shares
451 unsigned long shares
;
454 * load.weight at the time we set shares
456 unsigned long rq_weight
;
461 /* Real-Time classes' related field in a runqueue: */
463 struct rt_prio_array active
;
464 unsigned long rt_nr_running
;
465 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
466 int highest_prio
; /* highest queued rt task prio */
469 unsigned long rt_nr_migratory
;
475 /* Nests inside the rq lock: */
476 spinlock_t rt_runtime_lock
;
478 #ifdef CONFIG_RT_GROUP_SCHED
479 unsigned long rt_nr_boosted
;
482 struct list_head leaf_rt_rq_list
;
483 struct task_group
*tg
;
484 struct sched_rt_entity
*rt_se
;
491 * We add the notion of a root-domain which will be used to define per-domain
492 * variables. Each exclusive cpuset essentially defines an island domain by
493 * fully partitioning the member cpus from any other cpuset. Whenever a new
494 * exclusive cpuset is created, we also create and attach a new root-domain
501 cpumask_var_t online
;
504 * The "RT overload" flag: it gets set if a CPU has more than
505 * one runnable RT task.
507 cpumask_var_t rto_mask
;
510 struct cpupri cpupri
;
515 * By default the system creates a single root-domain with all cpus as
516 * members (mimicking the global state we have today).
518 static struct root_domain def_root_domain
;
523 * This is the main, per-CPU runqueue data structure.
525 * Locking rule: those places that want to lock multiple runqueues
526 * (such as the load balancing or the thread migration code), lock
527 * acquire operations must be ordered by ascending &runqueue.
534 * nr_running and cpu_load should be in the same cacheline because
535 * remote CPUs use both these fields when doing load calculation.
537 unsigned long nr_running
;
538 #define CPU_LOAD_IDX_MAX 5
539 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
540 unsigned char idle_at_tick
;
542 unsigned long last_tick_seen
;
543 unsigned char in_nohz_recently
;
545 /* capture load from *all* tasks on this cpu: */
546 struct load_weight load
;
547 unsigned long nr_load_updates
;
553 #ifdef CONFIG_FAIR_GROUP_SCHED
554 /* list of leaf cfs_rq on this cpu: */
555 struct list_head leaf_cfs_rq_list
;
557 #ifdef CONFIG_RT_GROUP_SCHED
558 struct list_head leaf_rt_rq_list
;
562 * This is part of a global counter where only the total sum
563 * over all CPUs matters. A task can increase this counter on
564 * one CPU and if it got migrated afterwards it may decrease
565 * it on another CPU. Always updated under the runqueue lock:
567 unsigned long nr_uninterruptible
;
569 struct task_struct
*curr
, *idle
;
570 unsigned long next_balance
;
571 struct mm_struct
*prev_mm
;
578 struct root_domain
*rd
;
579 struct sched_domain
*sd
;
581 /* For active balancing */
584 /* cpu of this runqueue: */
588 unsigned long avg_load_per_task
;
590 struct task_struct
*migration_thread
;
591 struct list_head migration_queue
;
594 #ifdef CONFIG_SCHED_HRTICK
596 int hrtick_csd_pending
;
597 struct call_single_data hrtick_csd
;
599 struct hrtimer hrtick_timer
;
602 #ifdef CONFIG_SCHEDSTATS
604 struct sched_info rq_sched_info
;
606 /* sys_sched_yield() stats */
607 unsigned int yld_exp_empty
;
608 unsigned int yld_act_empty
;
609 unsigned int yld_both_empty
;
610 unsigned int yld_count
;
612 /* schedule() stats */
613 unsigned int sched_switch
;
614 unsigned int sched_count
;
615 unsigned int sched_goidle
;
617 /* try_to_wake_up() stats */
618 unsigned int ttwu_count
;
619 unsigned int ttwu_local
;
622 unsigned int bkl_count
;
626 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
628 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int sync
)
630 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, sync
);
633 static inline int cpu_of(struct rq
*rq
)
643 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
644 * See detach_destroy_domains: synchronize_sched for details.
646 * The domain tree of any CPU may only be accessed from within
647 * preempt-disabled sections.
649 #define for_each_domain(cpu, __sd) \
650 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
652 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
653 #define this_rq() (&__get_cpu_var(runqueues))
654 #define task_rq(p) cpu_rq(task_cpu(p))
655 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
657 static inline void update_rq_clock(struct rq
*rq
)
659 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
663 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
665 #ifdef CONFIG_SCHED_DEBUG
666 # define const_debug __read_mostly
668 # define const_debug static const
674 * Returns true if the current cpu runqueue is locked.
675 * This interface allows printk to be called with the runqueue lock
676 * held and know whether or not it is OK to wake up the klogd.
678 int runqueue_is_locked(void)
681 struct rq
*rq
= cpu_rq(cpu
);
684 ret
= spin_is_locked(&rq
->lock
);
690 * Debugging: various feature bits
693 #define SCHED_FEAT(name, enabled) \
694 __SCHED_FEAT_##name ,
697 #include "sched_features.h"
702 #define SCHED_FEAT(name, enabled) \
703 (1UL << __SCHED_FEAT_##name) * enabled |
705 const_debug
unsigned int sysctl_sched_features
=
706 #include "sched_features.h"
711 #ifdef CONFIG_SCHED_DEBUG
712 #define SCHED_FEAT(name, enabled) \
715 static __read_mostly
char *sched_feat_names
[] = {
716 #include "sched_features.h"
722 static int sched_feat_show(struct seq_file
*m
, void *v
)
726 for (i
= 0; sched_feat_names
[i
]; i
++) {
727 if (!(sysctl_sched_features
& (1UL << i
)))
729 seq_printf(m
, "%s ", sched_feat_names
[i
]);
737 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
738 size_t cnt
, loff_t
*ppos
)
748 if (copy_from_user(&buf
, ubuf
, cnt
))
753 if (strncmp(buf
, "NO_", 3) == 0) {
758 for (i
= 0; sched_feat_names
[i
]; i
++) {
759 int len
= strlen(sched_feat_names
[i
]);
761 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
763 sysctl_sched_features
&= ~(1UL << i
);
765 sysctl_sched_features
|= (1UL << i
);
770 if (!sched_feat_names
[i
])
778 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
780 return single_open(filp
, sched_feat_show
, NULL
);
783 static struct file_operations sched_feat_fops
= {
784 .open
= sched_feat_open
,
785 .write
= sched_feat_write
,
788 .release
= single_release
,
791 static __init
int sched_init_debug(void)
793 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
798 late_initcall(sched_init_debug
);
802 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
805 * Number of tasks to iterate in a single balance run.
806 * Limited because this is done with IRQs disabled.
808 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
811 * ratelimit for updating the group shares.
814 unsigned int sysctl_sched_shares_ratelimit
= 250000;
817 * Inject some fuzzyness into changing the per-cpu group shares
818 * this avoids remote rq-locks at the expense of fairness.
821 unsigned int sysctl_sched_shares_thresh
= 4;
824 * period over which we measure -rt task cpu usage in us.
827 unsigned int sysctl_sched_rt_period
= 1000000;
829 static __read_mostly
int scheduler_running
;
832 * part of the period that we allow rt tasks to run in us.
835 int sysctl_sched_rt_runtime
= 950000;
837 static inline u64
global_rt_period(void)
839 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
842 static inline u64
global_rt_runtime(void)
844 if (sysctl_sched_rt_runtime
< 0)
847 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
850 #ifndef prepare_arch_switch
851 # define prepare_arch_switch(next) do { } while (0)
853 #ifndef finish_arch_switch
854 # define finish_arch_switch(prev) do { } while (0)
857 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
859 return rq
->curr
== p
;
862 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
863 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
865 return task_current(rq
, p
);
868 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
872 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
874 #ifdef CONFIG_DEBUG_SPINLOCK
875 /* this is a valid case when another task releases the spinlock */
876 rq
->lock
.owner
= current
;
879 * If we are tracking spinlock dependencies then we have to
880 * fix up the runqueue lock - which gets 'carried over' from
883 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
885 spin_unlock_irq(&rq
->lock
);
888 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
889 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
894 return task_current(rq
, p
);
898 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
902 * We can optimise this out completely for !SMP, because the
903 * SMP rebalancing from interrupt is the only thing that cares
908 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
909 spin_unlock_irq(&rq
->lock
);
911 spin_unlock(&rq
->lock
);
915 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
919 * After ->oncpu is cleared, the task can be moved to a different CPU.
920 * We must ensure this doesn't happen until the switch is completely
926 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
930 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
933 * __task_rq_lock - lock the runqueue a given task resides on.
934 * Must be called interrupts disabled.
936 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
940 struct rq
*rq
= task_rq(p
);
941 spin_lock(&rq
->lock
);
942 if (likely(rq
== task_rq(p
)))
944 spin_unlock(&rq
->lock
);
949 * task_rq_lock - lock the runqueue a given task resides on and disable
950 * interrupts. Note the ordering: we can safely lookup the task_rq without
951 * explicitly disabling preemption.
953 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
959 local_irq_save(*flags
);
961 spin_lock(&rq
->lock
);
962 if (likely(rq
== task_rq(p
)))
964 spin_unlock_irqrestore(&rq
->lock
, *flags
);
968 void task_rq_unlock_wait(struct task_struct
*p
)
970 struct rq
*rq
= task_rq(p
);
972 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
973 spin_unlock_wait(&rq
->lock
);
976 static void __task_rq_unlock(struct rq
*rq
)
979 spin_unlock(&rq
->lock
);
982 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
985 spin_unlock_irqrestore(&rq
->lock
, *flags
);
989 * this_rq_lock - lock this runqueue and disable interrupts.
991 static struct rq
*this_rq_lock(void)
998 spin_lock(&rq
->lock
);
1003 #ifdef CONFIG_SCHED_HRTICK
1005 * Use HR-timers to deliver accurate preemption points.
1007 * Its all a bit involved since we cannot program an hrt while holding the
1008 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1011 * When we get rescheduled we reprogram the hrtick_timer outside of the
1017 * - enabled by features
1018 * - hrtimer is actually high res
1020 static inline int hrtick_enabled(struct rq
*rq
)
1022 if (!sched_feat(HRTICK
))
1024 if (!cpu_active(cpu_of(rq
)))
1026 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1029 static void hrtick_clear(struct rq
*rq
)
1031 if (hrtimer_active(&rq
->hrtick_timer
))
1032 hrtimer_cancel(&rq
->hrtick_timer
);
1036 * High-resolution timer tick.
1037 * Runs from hardirq context with interrupts disabled.
1039 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1041 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1043 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1045 spin_lock(&rq
->lock
);
1046 update_rq_clock(rq
);
1047 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1048 spin_unlock(&rq
->lock
);
1050 return HRTIMER_NORESTART
;
1055 * called from hardirq (IPI) context
1057 static void __hrtick_start(void *arg
)
1059 struct rq
*rq
= arg
;
1061 spin_lock(&rq
->lock
);
1062 hrtimer_restart(&rq
->hrtick_timer
);
1063 rq
->hrtick_csd_pending
= 0;
1064 spin_unlock(&rq
->lock
);
1068 * Called to set the hrtick timer state.
1070 * called with rq->lock held and irqs disabled
1072 static void hrtick_start(struct rq
*rq
, u64 delay
)
1074 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1075 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1077 hrtimer_set_expires(timer
, time
);
1079 if (rq
== this_rq()) {
1080 hrtimer_restart(timer
);
1081 } else if (!rq
->hrtick_csd_pending
) {
1082 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
);
1083 rq
->hrtick_csd_pending
= 1;
1088 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1090 int cpu
= (int)(long)hcpu
;
1093 case CPU_UP_CANCELED
:
1094 case CPU_UP_CANCELED_FROZEN
:
1095 case CPU_DOWN_PREPARE
:
1096 case CPU_DOWN_PREPARE_FROZEN
:
1098 case CPU_DEAD_FROZEN
:
1099 hrtick_clear(cpu_rq(cpu
));
1106 static __init
void init_hrtick(void)
1108 hotcpu_notifier(hotplug_hrtick
, 0);
1112 * Called to set the hrtick timer state.
1114 * called with rq->lock held and irqs disabled
1116 static void hrtick_start(struct rq
*rq
, u64 delay
)
1118 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
), HRTIMER_MODE_REL
);
1121 static inline void init_hrtick(void)
1124 #endif /* CONFIG_SMP */
1126 static void init_rq_hrtick(struct rq
*rq
)
1129 rq
->hrtick_csd_pending
= 0;
1131 rq
->hrtick_csd
.flags
= 0;
1132 rq
->hrtick_csd
.func
= __hrtick_start
;
1133 rq
->hrtick_csd
.info
= rq
;
1136 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1137 rq
->hrtick_timer
.function
= hrtick
;
1138 rq
->hrtick_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_PERCPU
;
1140 #else /* CONFIG_SCHED_HRTICK */
1141 static inline void hrtick_clear(struct rq
*rq
)
1145 static inline void init_rq_hrtick(struct rq
*rq
)
1149 static inline void init_hrtick(void)
1152 #endif /* CONFIG_SCHED_HRTICK */
1155 * resched_task - mark a task 'to be rescheduled now'.
1157 * On UP this means the setting of the need_resched flag, on SMP it
1158 * might also involve a cross-CPU call to trigger the scheduler on
1163 #ifndef tsk_is_polling
1164 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1167 static void resched_task(struct task_struct
*p
)
1171 assert_spin_locked(&task_rq(p
)->lock
);
1173 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
1176 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
1179 if (cpu
== smp_processor_id())
1182 /* NEED_RESCHED must be visible before we test polling */
1184 if (!tsk_is_polling(p
))
1185 smp_send_reschedule(cpu
);
1188 static void resched_cpu(int cpu
)
1190 struct rq
*rq
= cpu_rq(cpu
);
1191 unsigned long flags
;
1193 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1195 resched_task(cpu_curr(cpu
));
1196 spin_unlock_irqrestore(&rq
->lock
, flags
);
1201 * When add_timer_on() enqueues a timer into the timer wheel of an
1202 * idle CPU then this timer might expire before the next timer event
1203 * which is scheduled to wake up that CPU. In case of a completely
1204 * idle system the next event might even be infinite time into the
1205 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1206 * leaves the inner idle loop so the newly added timer is taken into
1207 * account when the CPU goes back to idle and evaluates the timer
1208 * wheel for the next timer event.
1210 void wake_up_idle_cpu(int cpu
)
1212 struct rq
*rq
= cpu_rq(cpu
);
1214 if (cpu
== smp_processor_id())
1218 * This is safe, as this function is called with the timer
1219 * wheel base lock of (cpu) held. When the CPU is on the way
1220 * to idle and has not yet set rq->curr to idle then it will
1221 * be serialized on the timer wheel base lock and take the new
1222 * timer into account automatically.
1224 if (rq
->curr
!= rq
->idle
)
1228 * We can set TIF_RESCHED on the idle task of the other CPU
1229 * lockless. The worst case is that the other CPU runs the
1230 * idle task through an additional NOOP schedule()
1232 set_tsk_thread_flag(rq
->idle
, TIF_NEED_RESCHED
);
1234 /* NEED_RESCHED must be visible before we test polling */
1236 if (!tsk_is_polling(rq
->idle
))
1237 smp_send_reschedule(cpu
);
1239 #endif /* CONFIG_NO_HZ */
1241 #else /* !CONFIG_SMP */
1242 static void resched_task(struct task_struct
*p
)
1244 assert_spin_locked(&task_rq(p
)->lock
);
1245 set_tsk_need_resched(p
);
1247 #endif /* CONFIG_SMP */
1249 #if BITS_PER_LONG == 32
1250 # define WMULT_CONST (~0UL)
1252 # define WMULT_CONST (1UL << 32)
1255 #define WMULT_SHIFT 32
1258 * Shift right and round:
1260 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1263 * delta *= weight / lw
1265 static unsigned long
1266 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1267 struct load_weight
*lw
)
1271 if (!lw
->inv_weight
) {
1272 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1275 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1279 tmp
= (u64
)delta_exec
* weight
;
1281 * Check whether we'd overflow the 64-bit multiplication:
1283 if (unlikely(tmp
> WMULT_CONST
))
1284 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1287 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1289 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1292 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1298 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1305 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1306 * of tasks with abnormal "nice" values across CPUs the contribution that
1307 * each task makes to its run queue's load is weighted according to its
1308 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1309 * scaled version of the new time slice allocation that they receive on time
1313 #define WEIGHT_IDLEPRIO 2
1314 #define WMULT_IDLEPRIO (1 << 31)
1317 * Nice levels are multiplicative, with a gentle 10% change for every
1318 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1319 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1320 * that remained on nice 0.
1322 * The "10% effect" is relative and cumulative: from _any_ nice level,
1323 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1324 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1325 * If a task goes up by ~10% and another task goes down by ~10% then
1326 * the relative distance between them is ~25%.)
1328 static const int prio_to_weight
[40] = {
1329 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1330 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1331 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1332 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1333 /* 0 */ 1024, 820, 655, 526, 423,
1334 /* 5 */ 335, 272, 215, 172, 137,
1335 /* 10 */ 110, 87, 70, 56, 45,
1336 /* 15 */ 36, 29, 23, 18, 15,
1340 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1342 * In cases where the weight does not change often, we can use the
1343 * precalculated inverse to speed up arithmetics by turning divisions
1344 * into multiplications:
1346 static const u32 prio_to_wmult
[40] = {
1347 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1348 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1349 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1350 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1351 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1352 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1353 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1354 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1357 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1360 * runqueue iterator, to support SMP load-balancing between different
1361 * scheduling classes, without having to expose their internal data
1362 * structures to the load-balancing proper:
1364 struct rq_iterator
{
1366 struct task_struct
*(*start
)(void *);
1367 struct task_struct
*(*next
)(void *);
1371 static unsigned long
1372 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1373 unsigned long max_load_move
, struct sched_domain
*sd
,
1374 enum cpu_idle_type idle
, int *all_pinned
,
1375 int *this_best_prio
, struct rq_iterator
*iterator
);
1378 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1379 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1380 struct rq_iterator
*iterator
);
1383 #ifdef CONFIG_CGROUP_CPUACCT
1384 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1386 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1389 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1391 update_load_add(&rq
->load
, load
);
1394 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1396 update_load_sub(&rq
->load
, load
);
1399 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1400 typedef int (*tg_visitor
)(struct task_group
*, void *);
1403 * Iterate the full tree, calling @down when first entering a node and @up when
1404 * leaving it for the final time.
1406 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1408 struct task_group
*parent
, *child
;
1412 parent
= &root_task_group
;
1414 ret
= (*down
)(parent
, data
);
1417 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1424 ret
= (*up
)(parent
, data
);
1429 parent
= parent
->parent
;
1438 static int tg_nop(struct task_group
*tg
, void *data
)
1445 static unsigned long source_load(int cpu
, int type
);
1446 static unsigned long target_load(int cpu
, int type
);
1447 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1449 static unsigned long cpu_avg_load_per_task(int cpu
)
1451 struct rq
*rq
= cpu_rq(cpu
);
1452 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1455 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1457 rq
->avg_load_per_task
= 0;
1459 return rq
->avg_load_per_task
;
1462 #ifdef CONFIG_FAIR_GROUP_SCHED
1464 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1467 * Calculate and set the cpu's group shares.
1470 update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1471 unsigned long sd_shares
, unsigned long sd_rq_weight
)
1473 unsigned long shares
;
1474 unsigned long rq_weight
;
1479 rq_weight
= tg
->cfs_rq
[cpu
]->rq_weight
;
1482 * \Sum shares * rq_weight
1483 * shares = -----------------------
1487 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1488 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1490 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1491 sysctl_sched_shares_thresh
) {
1492 struct rq
*rq
= cpu_rq(cpu
);
1493 unsigned long flags
;
1495 spin_lock_irqsave(&rq
->lock
, flags
);
1496 tg
->cfs_rq
[cpu
]->shares
= shares
;
1498 __set_se_shares(tg
->se
[cpu
], shares
);
1499 spin_unlock_irqrestore(&rq
->lock
, flags
);
1504 * Re-compute the task group their per cpu shares over the given domain.
1505 * This needs to be done in a bottom-up fashion because the rq weight of a
1506 * parent group depends on the shares of its child groups.
1508 static int tg_shares_up(struct task_group
*tg
, void *data
)
1510 unsigned long weight
, rq_weight
= 0;
1511 unsigned long shares
= 0;
1512 struct sched_domain
*sd
= data
;
1515 for_each_cpu(i
, sched_domain_span(sd
)) {
1517 * If there are currently no tasks on the cpu pretend there
1518 * is one of average load so that when a new task gets to
1519 * run here it will not get delayed by group starvation.
1521 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1523 weight
= NICE_0_LOAD
;
1525 tg
->cfs_rq
[i
]->rq_weight
= weight
;
1526 rq_weight
+= weight
;
1527 shares
+= tg
->cfs_rq
[i
]->shares
;
1530 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1531 shares
= tg
->shares
;
1533 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1534 shares
= tg
->shares
;
1536 for_each_cpu(i
, sched_domain_span(sd
))
1537 update_group_shares_cpu(tg
, i
, shares
, rq_weight
);
1543 * Compute the cpu's hierarchical load factor for each task group.
1544 * This needs to be done in a top-down fashion because the load of a child
1545 * group is a fraction of its parents load.
1547 static int tg_load_down(struct task_group
*tg
, void *data
)
1550 long cpu
= (long)data
;
1553 load
= cpu_rq(cpu
)->load
.weight
;
1555 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1556 load
*= tg
->cfs_rq
[cpu
]->shares
;
1557 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1560 tg
->cfs_rq
[cpu
]->h_load
= load
;
1565 static void update_shares(struct sched_domain
*sd
)
1567 u64 now
= cpu_clock(raw_smp_processor_id());
1568 s64 elapsed
= now
- sd
->last_update
;
1570 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1571 sd
->last_update
= now
;
1572 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1576 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1578 spin_unlock(&rq
->lock
);
1580 spin_lock(&rq
->lock
);
1583 static void update_h_load(long cpu
)
1585 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1590 static inline void update_shares(struct sched_domain
*sd
)
1594 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1601 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1603 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1604 __releases(this_rq
->lock
)
1605 __acquires(busiest
->lock
)
1606 __acquires(this_rq
->lock
)
1610 if (unlikely(!irqs_disabled())) {
1611 /* printk() doesn't work good under rq->lock */
1612 spin_unlock(&this_rq
->lock
);
1615 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1616 if (busiest
< this_rq
) {
1617 spin_unlock(&this_rq
->lock
);
1618 spin_lock(&busiest
->lock
);
1619 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
1622 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
1627 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1628 __releases(busiest
->lock
)
1630 spin_unlock(&busiest
->lock
);
1631 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1635 #ifdef CONFIG_FAIR_GROUP_SCHED
1636 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1639 cfs_rq
->shares
= shares
;
1644 #include "sched_stats.h"
1645 #include "sched_idletask.c"
1646 #include "sched_fair.c"
1647 #include "sched_rt.c"
1648 #ifdef CONFIG_SCHED_DEBUG
1649 # include "sched_debug.c"
1652 #define sched_class_highest (&rt_sched_class)
1653 #define for_each_class(class) \
1654 for (class = sched_class_highest; class; class = class->next)
1656 static void inc_nr_running(struct rq
*rq
)
1661 static void dec_nr_running(struct rq
*rq
)
1666 static void set_load_weight(struct task_struct
*p
)
1668 if (task_has_rt_policy(p
)) {
1669 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1670 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1675 * SCHED_IDLE tasks get minimal weight:
1677 if (p
->policy
== SCHED_IDLE
) {
1678 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1679 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1683 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1684 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1687 static void update_avg(u64
*avg
, u64 sample
)
1689 s64 diff
= sample
- *avg
;
1693 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1695 sched_info_queued(p
);
1696 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1700 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1702 if (sleep
&& p
->se
.last_wakeup
) {
1703 update_avg(&p
->se
.avg_overlap
,
1704 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1705 p
->se
.last_wakeup
= 0;
1708 sched_info_dequeued(p
);
1709 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1714 * __normal_prio - return the priority that is based on the static prio
1716 static inline int __normal_prio(struct task_struct
*p
)
1718 return p
->static_prio
;
1722 * Calculate the expected normal priority: i.e. priority
1723 * without taking RT-inheritance into account. Might be
1724 * boosted by interactivity modifiers. Changes upon fork,
1725 * setprio syscalls, and whenever the interactivity
1726 * estimator recalculates.
1728 static inline int normal_prio(struct task_struct
*p
)
1732 if (task_has_rt_policy(p
))
1733 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1735 prio
= __normal_prio(p
);
1740 * Calculate the current priority, i.e. the priority
1741 * taken into account by the scheduler. This value might
1742 * be boosted by RT tasks, or might be boosted by
1743 * interactivity modifiers. Will be RT if the task got
1744 * RT-boosted. If not then it returns p->normal_prio.
1746 static int effective_prio(struct task_struct
*p
)
1748 p
->normal_prio
= normal_prio(p
);
1750 * If we are RT tasks or we were boosted to RT priority,
1751 * keep the priority unchanged. Otherwise, update priority
1752 * to the normal priority:
1754 if (!rt_prio(p
->prio
))
1755 return p
->normal_prio
;
1760 * activate_task - move a task to the runqueue.
1762 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1764 if (task_contributes_to_load(p
))
1765 rq
->nr_uninterruptible
--;
1767 enqueue_task(rq
, p
, wakeup
);
1772 * deactivate_task - remove a task from the runqueue.
1774 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1776 if (task_contributes_to_load(p
))
1777 rq
->nr_uninterruptible
++;
1779 dequeue_task(rq
, p
, sleep
);
1784 * task_curr - is this task currently executing on a CPU?
1785 * @p: the task in question.
1787 inline int task_curr(const struct task_struct
*p
)
1789 return cpu_curr(task_cpu(p
)) == p
;
1792 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1794 set_task_rq(p
, cpu
);
1797 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1798 * successfuly executed on another CPU. We must ensure that updates of
1799 * per-task data have been completed by this moment.
1802 task_thread_info(p
)->cpu
= cpu
;
1806 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1807 const struct sched_class
*prev_class
,
1808 int oldprio
, int running
)
1810 if (prev_class
!= p
->sched_class
) {
1811 if (prev_class
->switched_from
)
1812 prev_class
->switched_from(rq
, p
, running
);
1813 p
->sched_class
->switched_to(rq
, p
, running
);
1815 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1820 /* Used instead of source_load when we know the type == 0 */
1821 static unsigned long weighted_cpuload(const int cpu
)
1823 return cpu_rq(cpu
)->load
.weight
;
1827 * Is this task likely cache-hot:
1830 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1835 * Buddy candidates are cache hot:
1837 if (sched_feat(CACHE_HOT_BUDDY
) &&
1838 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
1839 &p
->se
== cfs_rq_of(&p
->se
)->last
))
1842 if (p
->sched_class
!= &fair_sched_class
)
1845 if (sysctl_sched_migration_cost
== -1)
1847 if (sysctl_sched_migration_cost
== 0)
1850 delta
= now
- p
->se
.exec_start
;
1852 return delta
< (s64
)sysctl_sched_migration_cost
;
1856 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1858 int old_cpu
= task_cpu(p
);
1859 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1860 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1861 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1864 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1866 #ifdef CONFIG_SCHEDSTATS
1867 if (p
->se
.wait_start
)
1868 p
->se
.wait_start
-= clock_offset
;
1869 if (p
->se
.sleep_start
)
1870 p
->se
.sleep_start
-= clock_offset
;
1871 if (p
->se
.block_start
)
1872 p
->se
.block_start
-= clock_offset
;
1873 if (old_cpu
!= new_cpu
) {
1874 schedstat_inc(p
, se
.nr_migrations
);
1875 if (task_hot(p
, old_rq
->clock
, NULL
))
1876 schedstat_inc(p
, se
.nr_forced2_migrations
);
1879 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1880 new_cfsrq
->min_vruntime
;
1882 __set_task_cpu(p
, new_cpu
);
1885 struct migration_req
{
1886 struct list_head list
;
1888 struct task_struct
*task
;
1891 struct completion done
;
1895 * The task's runqueue lock must be held.
1896 * Returns true if you have to wait for migration thread.
1899 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1901 struct rq
*rq
= task_rq(p
);
1904 * If the task is not on a runqueue (and not running), then
1905 * it is sufficient to simply update the task's cpu field.
1907 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1908 set_task_cpu(p
, dest_cpu
);
1912 init_completion(&req
->done
);
1914 req
->dest_cpu
= dest_cpu
;
1915 list_add(&req
->list
, &rq
->migration_queue
);
1921 * wait_task_inactive - wait for a thread to unschedule.
1923 * If @match_state is nonzero, it's the @p->state value just checked and
1924 * not expected to change. If it changes, i.e. @p might have woken up,
1925 * then return zero. When we succeed in waiting for @p to be off its CPU,
1926 * we return a positive number (its total switch count). If a second call
1927 * a short while later returns the same number, the caller can be sure that
1928 * @p has remained unscheduled the whole time.
1930 * The caller must ensure that the task *will* unschedule sometime soon,
1931 * else this function might spin for a *long* time. This function can't
1932 * be called with interrupts off, or it may introduce deadlock with
1933 * smp_call_function() if an IPI is sent by the same process we are
1934 * waiting to become inactive.
1936 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1938 unsigned long flags
;
1945 * We do the initial early heuristics without holding
1946 * any task-queue locks at all. We'll only try to get
1947 * the runqueue lock when things look like they will
1953 * If the task is actively running on another CPU
1954 * still, just relax and busy-wait without holding
1957 * NOTE! Since we don't hold any locks, it's not
1958 * even sure that "rq" stays as the right runqueue!
1959 * But we don't care, since "task_running()" will
1960 * return false if the runqueue has changed and p
1961 * is actually now running somewhere else!
1963 while (task_running(rq
, p
)) {
1964 if (match_state
&& unlikely(p
->state
!= match_state
))
1970 * Ok, time to look more closely! We need the rq
1971 * lock now, to be *sure*. If we're wrong, we'll
1972 * just go back and repeat.
1974 rq
= task_rq_lock(p
, &flags
);
1975 trace_sched_wait_task(rq
, p
);
1976 running
= task_running(rq
, p
);
1977 on_rq
= p
->se
.on_rq
;
1979 if (!match_state
|| p
->state
== match_state
)
1980 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1981 task_rq_unlock(rq
, &flags
);
1984 * If it changed from the expected state, bail out now.
1986 if (unlikely(!ncsw
))
1990 * Was it really running after all now that we
1991 * checked with the proper locks actually held?
1993 * Oops. Go back and try again..
1995 if (unlikely(running
)) {
2001 * It's not enough that it's not actively running,
2002 * it must be off the runqueue _entirely_, and not
2005 * So if it wa still runnable (but just not actively
2006 * running right now), it's preempted, and we should
2007 * yield - it could be a while.
2009 if (unlikely(on_rq
)) {
2010 schedule_timeout_uninterruptible(1);
2015 * Ahh, all good. It wasn't running, and it wasn't
2016 * runnable, which means that it will never become
2017 * running in the future either. We're all done!
2026 * kick_process - kick a running thread to enter/exit the kernel
2027 * @p: the to-be-kicked thread
2029 * Cause a process which is running on another CPU to enter
2030 * kernel-mode, without any delay. (to get signals handled.)
2032 * NOTE: this function doesnt have to take the runqueue lock,
2033 * because all it wants to ensure is that the remote task enters
2034 * the kernel. If the IPI races and the task has been migrated
2035 * to another CPU then no harm is done and the purpose has been
2038 void kick_process(struct task_struct
*p
)
2044 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2045 smp_send_reschedule(cpu
);
2050 * Return a low guess at the load of a migration-source cpu weighted
2051 * according to the scheduling class and "nice" value.
2053 * We want to under-estimate the load of migration sources, to
2054 * balance conservatively.
2056 static unsigned long source_load(int cpu
, int type
)
2058 struct rq
*rq
= cpu_rq(cpu
);
2059 unsigned long total
= weighted_cpuload(cpu
);
2061 if (type
== 0 || !sched_feat(LB_BIAS
))
2064 return min(rq
->cpu_load
[type
-1], total
);
2068 * Return a high guess at the load of a migration-target cpu weighted
2069 * according to the scheduling class and "nice" value.
2071 static unsigned long target_load(int cpu
, int type
)
2073 struct rq
*rq
= cpu_rq(cpu
);
2074 unsigned long total
= weighted_cpuload(cpu
);
2076 if (type
== 0 || !sched_feat(LB_BIAS
))
2079 return max(rq
->cpu_load
[type
-1], total
);
2083 * find_idlest_group finds and returns the least busy CPU group within the
2086 static struct sched_group
*
2087 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2089 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2090 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2091 int load_idx
= sd
->forkexec_idx
;
2092 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2095 unsigned long load
, avg_load
;
2099 /* Skip over this group if it has no CPUs allowed */
2100 if (!cpumask_intersects(sched_group_cpus(group
),
2104 local_group
= cpumask_test_cpu(this_cpu
,
2105 sched_group_cpus(group
));
2107 /* Tally up the load of all CPUs in the group */
2110 for_each_cpu(i
, sched_group_cpus(group
)) {
2111 /* Bias balancing toward cpus of our domain */
2113 load
= source_load(i
, load_idx
);
2115 load
= target_load(i
, load_idx
);
2120 /* Adjust by relative CPU power of the group */
2121 avg_load
= sg_div_cpu_power(group
,
2122 avg_load
* SCHED_LOAD_SCALE
);
2125 this_load
= avg_load
;
2127 } else if (avg_load
< min_load
) {
2128 min_load
= avg_load
;
2131 } while (group
= group
->next
, group
!= sd
->groups
);
2133 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2139 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2142 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
2144 unsigned long load
, min_load
= ULONG_MAX
;
2148 /* Traverse only the allowed CPUs */
2149 for_each_cpu_and(i
, sched_group_cpus(group
), &p
->cpus_allowed
) {
2150 load
= weighted_cpuload(i
);
2152 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2162 * sched_balance_self: balance the current task (running on cpu) in domains
2163 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2166 * Balance, ie. select the least loaded group.
2168 * Returns the target CPU number, or the same CPU if no balancing is needed.
2170 * preempt must be disabled.
2172 static int sched_balance_self(int cpu
, int flag
)
2174 struct task_struct
*t
= current
;
2175 struct sched_domain
*tmp
, *sd
= NULL
;
2177 for_each_domain(cpu
, tmp
) {
2179 * If power savings logic is enabled for a domain, stop there.
2181 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2183 if (tmp
->flags
& flag
)
2191 struct sched_group
*group
;
2192 int new_cpu
, weight
;
2194 if (!(sd
->flags
& flag
)) {
2199 group
= find_idlest_group(sd
, t
, cpu
);
2205 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
2206 if (new_cpu
== -1 || new_cpu
== cpu
) {
2207 /* Now try balancing at a lower domain level of cpu */
2212 /* Now try balancing at a lower domain level of new_cpu */
2214 weight
= cpumask_weight(sched_domain_span(sd
));
2216 for_each_domain(cpu
, tmp
) {
2217 if (weight
<= cpumask_weight(sched_domain_span(tmp
)))
2219 if (tmp
->flags
& flag
)
2222 /* while loop will break here if sd == NULL */
2228 #endif /* CONFIG_SMP */
2231 * try_to_wake_up - wake up a thread
2232 * @p: the to-be-woken-up thread
2233 * @state: the mask of task states that can be woken
2234 * @sync: do a synchronous wakeup?
2236 * Put it on the run-queue if it's not already there. The "current"
2237 * thread is always on the run-queue (except when the actual
2238 * re-schedule is in progress), and as such you're allowed to do
2239 * the simpler "current->state = TASK_RUNNING" to mark yourself
2240 * runnable without the overhead of this.
2242 * returns failure only if the task is already active.
2244 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2246 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2247 unsigned long flags
;
2251 if (!sched_feat(SYNC_WAKEUPS
))
2255 if (sched_feat(LB_WAKEUP_UPDATE
)) {
2256 struct sched_domain
*sd
;
2258 this_cpu
= raw_smp_processor_id();
2261 for_each_domain(this_cpu
, sd
) {
2262 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2271 rq
= task_rq_lock(p
, &flags
);
2272 old_state
= p
->state
;
2273 if (!(old_state
& state
))
2281 this_cpu
= smp_processor_id();
2284 if (unlikely(task_running(rq
, p
)))
2287 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2288 if (cpu
!= orig_cpu
) {
2289 set_task_cpu(p
, cpu
);
2290 task_rq_unlock(rq
, &flags
);
2291 /* might preempt at this point */
2292 rq
= task_rq_lock(p
, &flags
);
2293 old_state
= p
->state
;
2294 if (!(old_state
& state
))
2299 this_cpu
= smp_processor_id();
2303 #ifdef CONFIG_SCHEDSTATS
2304 schedstat_inc(rq
, ttwu_count
);
2305 if (cpu
== this_cpu
)
2306 schedstat_inc(rq
, ttwu_local
);
2308 struct sched_domain
*sd
;
2309 for_each_domain(this_cpu
, sd
) {
2310 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2311 schedstat_inc(sd
, ttwu_wake_remote
);
2316 #endif /* CONFIG_SCHEDSTATS */
2319 #endif /* CONFIG_SMP */
2320 schedstat_inc(p
, se
.nr_wakeups
);
2322 schedstat_inc(p
, se
.nr_wakeups_sync
);
2323 if (orig_cpu
!= cpu
)
2324 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2325 if (cpu
== this_cpu
)
2326 schedstat_inc(p
, se
.nr_wakeups_local
);
2328 schedstat_inc(p
, se
.nr_wakeups_remote
);
2329 update_rq_clock(rq
);
2330 activate_task(rq
, p
, 1);
2334 trace_sched_wakeup(rq
, p
);
2335 check_preempt_curr(rq
, p
, sync
);
2337 p
->state
= TASK_RUNNING
;
2339 if (p
->sched_class
->task_wake_up
)
2340 p
->sched_class
->task_wake_up(rq
, p
);
2343 current
->se
.last_wakeup
= current
->se
.sum_exec_runtime
;
2345 task_rq_unlock(rq
, &flags
);
2350 int wake_up_process(struct task_struct
*p
)
2352 return try_to_wake_up(p
, TASK_ALL
, 0);
2354 EXPORT_SYMBOL(wake_up_process
);
2356 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2358 return try_to_wake_up(p
, state
, 0);
2362 * Perform scheduler related setup for a newly forked process p.
2363 * p is forked by current.
2365 * __sched_fork() is basic setup used by init_idle() too:
2367 static void __sched_fork(struct task_struct
*p
)
2369 p
->se
.exec_start
= 0;
2370 p
->se
.sum_exec_runtime
= 0;
2371 p
->se
.prev_sum_exec_runtime
= 0;
2372 p
->se
.last_wakeup
= 0;
2373 p
->se
.avg_overlap
= 0;
2375 #ifdef CONFIG_SCHEDSTATS
2376 p
->se
.wait_start
= 0;
2377 p
->se
.sum_sleep_runtime
= 0;
2378 p
->se
.sleep_start
= 0;
2379 p
->se
.block_start
= 0;
2380 p
->se
.sleep_max
= 0;
2381 p
->se
.block_max
= 0;
2383 p
->se
.slice_max
= 0;
2387 INIT_LIST_HEAD(&p
->rt
.run_list
);
2389 INIT_LIST_HEAD(&p
->se
.group_node
);
2391 #ifdef CONFIG_PREEMPT_NOTIFIERS
2392 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2396 * We mark the process as running here, but have not actually
2397 * inserted it onto the runqueue yet. This guarantees that
2398 * nobody will actually run it, and a signal or other external
2399 * event cannot wake it up and insert it on the runqueue either.
2401 p
->state
= TASK_RUNNING
;
2405 * fork()/clone()-time setup:
2407 void sched_fork(struct task_struct
*p
, int clone_flags
)
2409 int cpu
= get_cpu();
2414 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2416 set_task_cpu(p
, cpu
);
2419 * Make sure we do not leak PI boosting priority to the child:
2421 p
->prio
= current
->normal_prio
;
2422 if (!rt_prio(p
->prio
))
2423 p
->sched_class
= &fair_sched_class
;
2425 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2426 if (likely(sched_info_on()))
2427 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2429 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2432 #ifdef CONFIG_PREEMPT
2433 /* Want to start with kernel preemption disabled. */
2434 task_thread_info(p
)->preempt_count
= 1;
2440 * wake_up_new_task - wake up a newly created task for the first time.
2442 * This function will do some initial scheduler statistics housekeeping
2443 * that must be done for every newly created context, then puts the task
2444 * on the runqueue and wakes it.
2446 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2448 unsigned long flags
;
2451 rq
= task_rq_lock(p
, &flags
);
2452 BUG_ON(p
->state
!= TASK_RUNNING
);
2453 update_rq_clock(rq
);
2455 p
->prio
= effective_prio(p
);
2457 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2458 activate_task(rq
, p
, 0);
2461 * Let the scheduling class do new task startup
2462 * management (if any):
2464 p
->sched_class
->task_new(rq
, p
);
2467 trace_sched_wakeup_new(rq
, p
);
2468 check_preempt_curr(rq
, p
, 0);
2470 if (p
->sched_class
->task_wake_up
)
2471 p
->sched_class
->task_wake_up(rq
, p
);
2473 task_rq_unlock(rq
, &flags
);
2476 #ifdef CONFIG_PREEMPT_NOTIFIERS
2479 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2480 * @notifier: notifier struct to register
2482 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2484 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2486 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2489 * preempt_notifier_unregister - no longer interested in preemption notifications
2490 * @notifier: notifier struct to unregister
2492 * This is safe to call from within a preemption notifier.
2494 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2496 hlist_del(¬ifier
->link
);
2498 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2500 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2502 struct preempt_notifier
*notifier
;
2503 struct hlist_node
*node
;
2505 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2506 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2510 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2511 struct task_struct
*next
)
2513 struct preempt_notifier
*notifier
;
2514 struct hlist_node
*node
;
2516 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2517 notifier
->ops
->sched_out(notifier
, next
);
2520 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2522 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2527 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2528 struct task_struct
*next
)
2532 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2535 * prepare_task_switch - prepare to switch tasks
2536 * @rq: the runqueue preparing to switch
2537 * @prev: the current task that is being switched out
2538 * @next: the task we are going to switch to.
2540 * This is called with the rq lock held and interrupts off. It must
2541 * be paired with a subsequent finish_task_switch after the context
2544 * prepare_task_switch sets up locking and calls architecture specific
2548 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2549 struct task_struct
*next
)
2551 fire_sched_out_preempt_notifiers(prev
, next
);
2552 prepare_lock_switch(rq
, next
);
2553 prepare_arch_switch(next
);
2557 * finish_task_switch - clean up after a task-switch
2558 * @rq: runqueue associated with task-switch
2559 * @prev: the thread we just switched away from.
2561 * finish_task_switch must be called after the context switch, paired
2562 * with a prepare_task_switch call before the context switch.
2563 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2564 * and do any other architecture-specific cleanup actions.
2566 * Note that we may have delayed dropping an mm in context_switch(). If
2567 * so, we finish that here outside of the runqueue lock. (Doing it
2568 * with the lock held can cause deadlocks; see schedule() for
2571 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2572 __releases(rq
->lock
)
2574 struct mm_struct
*mm
= rq
->prev_mm
;
2580 * A task struct has one reference for the use as "current".
2581 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2582 * schedule one last time. The schedule call will never return, and
2583 * the scheduled task must drop that reference.
2584 * The test for TASK_DEAD must occur while the runqueue locks are
2585 * still held, otherwise prev could be scheduled on another cpu, die
2586 * there before we look at prev->state, and then the reference would
2588 * Manfred Spraul <manfred@colorfullife.com>
2590 prev_state
= prev
->state
;
2591 finish_arch_switch(prev
);
2592 finish_lock_switch(rq
, prev
);
2594 if (current
->sched_class
->post_schedule
)
2595 current
->sched_class
->post_schedule(rq
);
2598 fire_sched_in_preempt_notifiers(current
);
2601 if (unlikely(prev_state
== TASK_DEAD
)) {
2603 * Remove function-return probe instances associated with this
2604 * task and put them back on the free list.
2606 kprobe_flush_task(prev
);
2607 put_task_struct(prev
);
2612 * schedule_tail - first thing a freshly forked thread must call.
2613 * @prev: the thread we just switched away from.
2615 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2616 __releases(rq
->lock
)
2618 struct rq
*rq
= this_rq();
2620 finish_task_switch(rq
, prev
);
2621 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2622 /* In this case, finish_task_switch does not reenable preemption */
2625 if (current
->set_child_tid
)
2626 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2630 * context_switch - switch to the new MM and the new
2631 * thread's register state.
2634 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2635 struct task_struct
*next
)
2637 struct mm_struct
*mm
, *oldmm
;
2639 prepare_task_switch(rq
, prev
, next
);
2640 trace_sched_switch(rq
, prev
, next
);
2642 oldmm
= prev
->active_mm
;
2644 * For paravirt, this is coupled with an exit in switch_to to
2645 * combine the page table reload and the switch backend into
2648 arch_enter_lazy_cpu_mode();
2650 if (unlikely(!mm
)) {
2651 next
->active_mm
= oldmm
;
2652 atomic_inc(&oldmm
->mm_count
);
2653 enter_lazy_tlb(oldmm
, next
);
2655 switch_mm(oldmm
, mm
, next
);
2657 if (unlikely(!prev
->mm
)) {
2658 prev
->active_mm
= NULL
;
2659 rq
->prev_mm
= oldmm
;
2662 * Since the runqueue lock will be released by the next
2663 * task (which is an invalid locking op but in the case
2664 * of the scheduler it's an obvious special-case), so we
2665 * do an early lockdep release here:
2667 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2668 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2671 /* Here we just switch the register state and the stack. */
2672 switch_to(prev
, next
, prev
);
2676 * this_rq must be evaluated again because prev may have moved
2677 * CPUs since it called schedule(), thus the 'rq' on its stack
2678 * frame will be invalid.
2680 finish_task_switch(this_rq(), prev
);
2684 * nr_running, nr_uninterruptible and nr_context_switches:
2686 * externally visible scheduler statistics: current number of runnable
2687 * threads, current number of uninterruptible-sleeping threads, total
2688 * number of context switches performed since bootup.
2690 unsigned long nr_running(void)
2692 unsigned long i
, sum
= 0;
2694 for_each_online_cpu(i
)
2695 sum
+= cpu_rq(i
)->nr_running
;
2700 unsigned long nr_uninterruptible(void)
2702 unsigned long i
, sum
= 0;
2704 for_each_possible_cpu(i
)
2705 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2708 * Since we read the counters lockless, it might be slightly
2709 * inaccurate. Do not allow it to go below zero though:
2711 if (unlikely((long)sum
< 0))
2717 unsigned long long nr_context_switches(void)
2720 unsigned long long sum
= 0;
2722 for_each_possible_cpu(i
)
2723 sum
+= cpu_rq(i
)->nr_switches
;
2728 unsigned long nr_iowait(void)
2730 unsigned long i
, sum
= 0;
2732 for_each_possible_cpu(i
)
2733 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2738 unsigned long nr_active(void)
2740 unsigned long i
, running
= 0, uninterruptible
= 0;
2742 for_each_online_cpu(i
) {
2743 running
+= cpu_rq(i
)->nr_running
;
2744 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2747 if (unlikely((long)uninterruptible
< 0))
2748 uninterruptible
= 0;
2750 return running
+ uninterruptible
;
2754 * Update rq->cpu_load[] statistics. This function is usually called every
2755 * scheduler tick (TICK_NSEC).
2757 static void update_cpu_load(struct rq
*this_rq
)
2759 unsigned long this_load
= this_rq
->load
.weight
;
2762 this_rq
->nr_load_updates
++;
2764 /* Update our load: */
2765 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2766 unsigned long old_load
, new_load
;
2768 /* scale is effectively 1 << i now, and >> i divides by scale */
2770 old_load
= this_rq
->cpu_load
[i
];
2771 new_load
= this_load
;
2773 * Round up the averaging division if load is increasing. This
2774 * prevents us from getting stuck on 9 if the load is 10, for
2777 if (new_load
> old_load
)
2778 new_load
+= scale
-1;
2779 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2786 * double_rq_lock - safely lock two runqueues
2788 * Note this does not disable interrupts like task_rq_lock,
2789 * you need to do so manually before calling.
2791 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2792 __acquires(rq1
->lock
)
2793 __acquires(rq2
->lock
)
2795 BUG_ON(!irqs_disabled());
2797 spin_lock(&rq1
->lock
);
2798 __acquire(rq2
->lock
); /* Fake it out ;) */
2801 spin_lock(&rq1
->lock
);
2802 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
2804 spin_lock(&rq2
->lock
);
2805 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
2808 update_rq_clock(rq1
);
2809 update_rq_clock(rq2
);
2813 * double_rq_unlock - safely unlock two runqueues
2815 * Note this does not restore interrupts like task_rq_unlock,
2816 * you need to do so manually after calling.
2818 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2819 __releases(rq1
->lock
)
2820 __releases(rq2
->lock
)
2822 spin_unlock(&rq1
->lock
);
2824 spin_unlock(&rq2
->lock
);
2826 __release(rq2
->lock
);
2830 * If dest_cpu is allowed for this process, migrate the task to it.
2831 * This is accomplished by forcing the cpu_allowed mask to only
2832 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2833 * the cpu_allowed mask is restored.
2835 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2837 struct migration_req req
;
2838 unsigned long flags
;
2841 rq
= task_rq_lock(p
, &flags
);
2842 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
2843 || unlikely(!cpu_active(dest_cpu
)))
2846 trace_sched_migrate_task(rq
, p
, dest_cpu
);
2847 /* force the process onto the specified CPU */
2848 if (migrate_task(p
, dest_cpu
, &req
)) {
2849 /* Need to wait for migration thread (might exit: take ref). */
2850 struct task_struct
*mt
= rq
->migration_thread
;
2852 get_task_struct(mt
);
2853 task_rq_unlock(rq
, &flags
);
2854 wake_up_process(mt
);
2855 put_task_struct(mt
);
2856 wait_for_completion(&req
.done
);
2861 task_rq_unlock(rq
, &flags
);
2865 * sched_exec - execve() is a valuable balancing opportunity, because at
2866 * this point the task has the smallest effective memory and cache footprint.
2868 void sched_exec(void)
2870 int new_cpu
, this_cpu
= get_cpu();
2871 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2873 if (new_cpu
!= this_cpu
)
2874 sched_migrate_task(current
, new_cpu
);
2878 * pull_task - move a task from a remote runqueue to the local runqueue.
2879 * Both runqueues must be locked.
2881 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2882 struct rq
*this_rq
, int this_cpu
)
2884 deactivate_task(src_rq
, p
, 0);
2885 set_task_cpu(p
, this_cpu
);
2886 activate_task(this_rq
, p
, 0);
2888 * Note that idle threads have a prio of MAX_PRIO, for this test
2889 * to be always true for them.
2891 check_preempt_curr(this_rq
, p
, 0);
2895 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2898 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2899 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2903 * We do not migrate tasks that are:
2904 * 1) running (obviously), or
2905 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2906 * 3) are cache-hot on their current CPU.
2908 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
2909 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2914 if (task_running(rq
, p
)) {
2915 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2920 * Aggressive migration if:
2921 * 1) task is cache cold, or
2922 * 2) too many balance attempts have failed.
2925 if (!task_hot(p
, rq
->clock
, sd
) ||
2926 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2927 #ifdef CONFIG_SCHEDSTATS
2928 if (task_hot(p
, rq
->clock
, sd
)) {
2929 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2930 schedstat_inc(p
, se
.nr_forced_migrations
);
2936 if (task_hot(p
, rq
->clock
, sd
)) {
2937 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2943 static unsigned long
2944 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2945 unsigned long max_load_move
, struct sched_domain
*sd
,
2946 enum cpu_idle_type idle
, int *all_pinned
,
2947 int *this_best_prio
, struct rq_iterator
*iterator
)
2949 int loops
= 0, pulled
= 0, pinned
= 0;
2950 struct task_struct
*p
;
2951 long rem_load_move
= max_load_move
;
2953 if (max_load_move
== 0)
2959 * Start the load-balancing iterator:
2961 p
= iterator
->start(iterator
->arg
);
2963 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2966 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
2967 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2968 p
= iterator
->next(iterator
->arg
);
2972 pull_task(busiest
, p
, this_rq
, this_cpu
);
2974 rem_load_move
-= p
->se
.load
.weight
;
2977 * We only want to steal up to the prescribed amount of weighted load.
2979 if (rem_load_move
> 0) {
2980 if (p
->prio
< *this_best_prio
)
2981 *this_best_prio
= p
->prio
;
2982 p
= iterator
->next(iterator
->arg
);
2987 * Right now, this is one of only two places pull_task() is called,
2988 * so we can safely collect pull_task() stats here rather than
2989 * inside pull_task().
2991 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2994 *all_pinned
= pinned
;
2996 return max_load_move
- rem_load_move
;
3000 * move_tasks tries to move up to max_load_move weighted load from busiest to
3001 * this_rq, as part of a balancing operation within domain "sd".
3002 * Returns 1 if successful and 0 otherwise.
3004 * Called with both runqueues locked.
3006 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3007 unsigned long max_load_move
,
3008 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3011 const struct sched_class
*class = sched_class_highest
;
3012 unsigned long total_load_moved
= 0;
3013 int this_best_prio
= this_rq
->curr
->prio
;
3017 class->load_balance(this_rq
, this_cpu
, busiest
,
3018 max_load_move
- total_load_moved
,
3019 sd
, idle
, all_pinned
, &this_best_prio
);
3020 class = class->next
;
3022 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3025 } while (class && max_load_move
> total_load_moved
);
3027 return total_load_moved
> 0;
3031 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3032 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3033 struct rq_iterator
*iterator
)
3035 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3039 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3040 pull_task(busiest
, p
, this_rq
, this_cpu
);
3042 * Right now, this is only the second place pull_task()
3043 * is called, so we can safely collect pull_task()
3044 * stats here rather than inside pull_task().
3046 schedstat_inc(sd
, lb_gained
[idle
]);
3050 p
= iterator
->next(iterator
->arg
);
3057 * move_one_task tries to move exactly one task from busiest to this_rq, as
3058 * part of active balancing operations within "domain".
3059 * Returns 1 if successful and 0 otherwise.
3061 * Called with both runqueues locked.
3063 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3064 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3066 const struct sched_class
*class;
3068 for (class = sched_class_highest
; class; class = class->next
)
3069 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3076 * find_busiest_group finds and returns the busiest CPU group within the
3077 * domain. It calculates and returns the amount of weighted load which
3078 * should be moved to restore balance via the imbalance parameter.
3080 static struct sched_group
*
3081 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3082 unsigned long *imbalance
, enum cpu_idle_type idle
,
3083 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
3085 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
3086 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
3087 unsigned long max_pull
;
3088 unsigned long busiest_load_per_task
, busiest_nr_running
;
3089 unsigned long this_load_per_task
, this_nr_running
;
3090 int load_idx
, group_imb
= 0;
3091 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3092 int power_savings_balance
= 1;
3093 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
3094 unsigned long min_nr_running
= ULONG_MAX
;
3095 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
3098 max_load
= this_load
= total_load
= total_pwr
= 0;
3099 busiest_load_per_task
= busiest_nr_running
= 0;
3100 this_load_per_task
= this_nr_running
= 0;
3102 if (idle
== CPU_NOT_IDLE
)
3103 load_idx
= sd
->busy_idx
;
3104 else if (idle
== CPU_NEWLY_IDLE
)
3105 load_idx
= sd
->newidle_idx
;
3107 load_idx
= sd
->idle_idx
;
3110 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
3113 int __group_imb
= 0;
3114 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3115 unsigned long sum_nr_running
, sum_weighted_load
;
3116 unsigned long sum_avg_load_per_task
;
3117 unsigned long avg_load_per_task
;
3119 local_group
= cpumask_test_cpu(this_cpu
,
3120 sched_group_cpus(group
));
3123 balance_cpu
= cpumask_first(sched_group_cpus(group
));
3125 /* Tally up the load of all CPUs in the group */
3126 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
3127 sum_avg_load_per_task
= avg_load_per_task
= 0;
3130 min_cpu_load
= ~0UL;
3132 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3133 struct rq
*rq
= cpu_rq(i
);
3135 if (*sd_idle
&& rq
->nr_running
)
3138 /* Bias balancing toward cpus of our domain */
3140 if (idle_cpu(i
) && !first_idle_cpu
) {
3145 load
= target_load(i
, load_idx
);
3147 load
= source_load(i
, load_idx
);
3148 if (load
> max_cpu_load
)
3149 max_cpu_load
= load
;
3150 if (min_cpu_load
> load
)
3151 min_cpu_load
= load
;
3155 sum_nr_running
+= rq
->nr_running
;
3156 sum_weighted_load
+= weighted_cpuload(i
);
3158 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3162 * First idle cpu or the first cpu(busiest) in this sched group
3163 * is eligible for doing load balancing at this and above
3164 * domains. In the newly idle case, we will allow all the cpu's
3165 * to do the newly idle load balance.
3167 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3168 balance_cpu
!= this_cpu
&& balance
) {
3173 total_load
+= avg_load
;
3174 total_pwr
+= group
->__cpu_power
;
3176 /* Adjust by relative CPU power of the group */
3177 avg_load
= sg_div_cpu_power(group
,
3178 avg_load
* SCHED_LOAD_SCALE
);
3182 * Consider the group unbalanced when the imbalance is larger
3183 * than the average weight of two tasks.
3185 * APZ: with cgroup the avg task weight can vary wildly and
3186 * might not be a suitable number - should we keep a
3187 * normalized nr_running number somewhere that negates
3190 avg_load_per_task
= sg_div_cpu_power(group
,
3191 sum_avg_load_per_task
* SCHED_LOAD_SCALE
);
3193 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3196 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3199 this_load
= avg_load
;
3201 this_nr_running
= sum_nr_running
;
3202 this_load_per_task
= sum_weighted_load
;
3203 } else if (avg_load
> max_load
&&
3204 (sum_nr_running
> group_capacity
|| __group_imb
)) {
3205 max_load
= avg_load
;
3207 busiest_nr_running
= sum_nr_running
;
3208 busiest_load_per_task
= sum_weighted_load
;
3209 group_imb
= __group_imb
;
3212 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3214 * Busy processors will not participate in power savings
3217 if (idle
== CPU_NOT_IDLE
||
3218 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3222 * If the local group is idle or completely loaded
3223 * no need to do power savings balance at this domain
3225 if (local_group
&& (this_nr_running
>= group_capacity
||
3227 power_savings_balance
= 0;
3230 * If a group is already running at full capacity or idle,
3231 * don't include that group in power savings calculations
3233 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
3238 * Calculate the group which has the least non-idle load.
3239 * This is the group from where we need to pick up the load
3242 if ((sum_nr_running
< min_nr_running
) ||
3243 (sum_nr_running
== min_nr_running
&&
3244 cpumask_first(sched_group_cpus(group
)) <
3245 cpumask_first(sched_group_cpus(group_min
)))) {
3247 min_nr_running
= sum_nr_running
;
3248 min_load_per_task
= sum_weighted_load
/
3253 * Calculate the group which is almost near its
3254 * capacity but still has some space to pick up some load
3255 * from other group and save more power
3257 if (sum_nr_running
<= group_capacity
- 1) {
3258 if (sum_nr_running
> leader_nr_running
||
3259 (sum_nr_running
== leader_nr_running
&&
3260 cpumask_first(sched_group_cpus(group
)) >
3261 cpumask_first(sched_group_cpus(group_leader
)))) {
3262 group_leader
= group
;
3263 leader_nr_running
= sum_nr_running
;
3268 group
= group
->next
;
3269 } while (group
!= sd
->groups
);
3271 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
3274 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
3276 if (this_load
>= avg_load
||
3277 100*max_load
<= sd
->imbalance_pct
*this_load
)
3280 busiest_load_per_task
/= busiest_nr_running
;
3282 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
3285 * We're trying to get all the cpus to the average_load, so we don't
3286 * want to push ourselves above the average load, nor do we wish to
3287 * reduce the max loaded cpu below the average load, as either of these
3288 * actions would just result in more rebalancing later, and ping-pong
3289 * tasks around. Thus we look for the minimum possible imbalance.
3290 * Negative imbalances (*we* are more loaded than anyone else) will
3291 * be counted as no imbalance for these purposes -- we can't fix that
3292 * by pulling tasks to us. Be careful of negative numbers as they'll
3293 * appear as very large values with unsigned longs.
3295 if (max_load
<= busiest_load_per_task
)
3299 * In the presence of smp nice balancing, certain scenarios can have
3300 * max load less than avg load(as we skip the groups at or below
3301 * its cpu_power, while calculating max_load..)
3303 if (max_load
< avg_load
) {
3305 goto small_imbalance
;
3308 /* Don't want to pull so many tasks that a group would go idle */
3309 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
3311 /* How much load to actually move to equalise the imbalance */
3312 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
3313 (avg_load
- this_load
) * this->__cpu_power
)
3317 * if *imbalance is less than the average load per runnable task
3318 * there is no gaurantee that any tasks will be moved so we'll have
3319 * a think about bumping its value to force at least one task to be
3322 if (*imbalance
< busiest_load_per_task
) {
3323 unsigned long tmp
, pwr_now
, pwr_move
;
3327 pwr_move
= pwr_now
= 0;
3329 if (this_nr_running
) {
3330 this_load_per_task
/= this_nr_running
;
3331 if (busiest_load_per_task
> this_load_per_task
)
3334 this_load_per_task
= cpu_avg_load_per_task(this_cpu
);
3336 if (max_load
- this_load
+ busiest_load_per_task
>=
3337 busiest_load_per_task
* imbn
) {
3338 *imbalance
= busiest_load_per_task
;
3343 * OK, we don't have enough imbalance to justify moving tasks,
3344 * however we may be able to increase total CPU power used by
3348 pwr_now
+= busiest
->__cpu_power
*
3349 min(busiest_load_per_task
, max_load
);
3350 pwr_now
+= this->__cpu_power
*
3351 min(this_load_per_task
, this_load
);
3352 pwr_now
/= SCHED_LOAD_SCALE
;
3354 /* Amount of load we'd subtract */
3355 tmp
= sg_div_cpu_power(busiest
,
3356 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3358 pwr_move
+= busiest
->__cpu_power
*
3359 min(busiest_load_per_task
, max_load
- tmp
);
3361 /* Amount of load we'd add */
3362 if (max_load
* busiest
->__cpu_power
<
3363 busiest_load_per_task
* SCHED_LOAD_SCALE
)
3364 tmp
= sg_div_cpu_power(this,
3365 max_load
* busiest
->__cpu_power
);
3367 tmp
= sg_div_cpu_power(this,
3368 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3369 pwr_move
+= this->__cpu_power
*
3370 min(this_load_per_task
, this_load
+ tmp
);
3371 pwr_move
/= SCHED_LOAD_SCALE
;
3373 /* Move if we gain throughput */
3374 if (pwr_move
> pwr_now
)
3375 *imbalance
= busiest_load_per_task
;
3381 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3382 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3385 if (this == group_leader
&& group_leader
!= group_min
) {
3386 *imbalance
= min_load_per_task
;
3396 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3399 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3400 unsigned long imbalance
, const struct cpumask
*cpus
)
3402 struct rq
*busiest
= NULL
, *rq
;
3403 unsigned long max_load
= 0;
3406 for_each_cpu(i
, sched_group_cpus(group
)) {
3409 if (!cpumask_test_cpu(i
, cpus
))
3413 wl
= weighted_cpuload(i
);
3415 if (rq
->nr_running
== 1 && wl
> imbalance
)
3418 if (wl
> max_load
) {
3428 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3429 * so long as it is large enough.
3431 #define MAX_PINNED_INTERVAL 512
3434 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3435 * tasks if there is an imbalance.
3437 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3438 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3439 int *balance
, struct cpumask
*cpus
)
3441 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3442 struct sched_group
*group
;
3443 unsigned long imbalance
;
3445 unsigned long flags
;
3447 cpumask_setall(cpus
);
3450 * When power savings policy is enabled for the parent domain, idle
3451 * sibling can pick up load irrespective of busy siblings. In this case,
3452 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3453 * portraying it as CPU_NOT_IDLE.
3455 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3456 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3459 schedstat_inc(sd
, lb_count
[idle
]);
3463 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3470 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3474 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3476 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3480 BUG_ON(busiest
== this_rq
);
3482 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3485 if (busiest
->nr_running
> 1) {
3487 * Attempt to move tasks. If find_busiest_group has found
3488 * an imbalance but busiest->nr_running <= 1, the group is
3489 * still unbalanced. ld_moved simply stays zero, so it is
3490 * correctly treated as an imbalance.
3492 local_irq_save(flags
);
3493 double_rq_lock(this_rq
, busiest
);
3494 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3495 imbalance
, sd
, idle
, &all_pinned
);
3496 double_rq_unlock(this_rq
, busiest
);
3497 local_irq_restore(flags
);
3500 * some other cpu did the load balance for us.
3502 if (ld_moved
&& this_cpu
!= smp_processor_id())
3503 resched_cpu(this_cpu
);
3505 /* All tasks on this runqueue were pinned by CPU affinity */
3506 if (unlikely(all_pinned
)) {
3507 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
3508 if (!cpumask_empty(cpus
))
3515 schedstat_inc(sd
, lb_failed
[idle
]);
3516 sd
->nr_balance_failed
++;
3518 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3520 spin_lock_irqsave(&busiest
->lock
, flags
);
3522 /* don't kick the migration_thread, if the curr
3523 * task on busiest cpu can't be moved to this_cpu
3525 if (!cpumask_test_cpu(this_cpu
,
3526 &busiest
->curr
->cpus_allowed
)) {
3527 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3529 goto out_one_pinned
;
3532 if (!busiest
->active_balance
) {
3533 busiest
->active_balance
= 1;
3534 busiest
->push_cpu
= this_cpu
;
3537 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3539 wake_up_process(busiest
->migration_thread
);
3542 * We've kicked active balancing, reset the failure
3545 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3548 sd
->nr_balance_failed
= 0;
3550 if (likely(!active_balance
)) {
3551 /* We were unbalanced, so reset the balancing interval */
3552 sd
->balance_interval
= sd
->min_interval
;
3555 * If we've begun active balancing, start to back off. This
3556 * case may not be covered by the all_pinned logic if there
3557 * is only 1 task on the busy runqueue (because we don't call
3560 if (sd
->balance_interval
< sd
->max_interval
)
3561 sd
->balance_interval
*= 2;
3564 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3565 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3571 schedstat_inc(sd
, lb_balanced
[idle
]);
3573 sd
->nr_balance_failed
= 0;
3576 /* tune up the balancing interval */
3577 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3578 (sd
->balance_interval
< sd
->max_interval
))
3579 sd
->balance_interval
*= 2;
3581 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3582 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3593 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3594 * tasks if there is an imbalance.
3596 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3597 * this_rq is locked.
3600 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
,
3601 struct cpumask
*cpus
)
3603 struct sched_group
*group
;
3604 struct rq
*busiest
= NULL
;
3605 unsigned long imbalance
;
3610 cpumask_setall(cpus
);
3613 * When power savings policy is enabled for the parent domain, idle
3614 * sibling can pick up load irrespective of busy siblings. In this case,
3615 * let the state of idle sibling percolate up as IDLE, instead of
3616 * portraying it as CPU_NOT_IDLE.
3618 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3619 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3622 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3624 update_shares_locked(this_rq
, sd
);
3625 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3626 &sd_idle
, cpus
, NULL
);
3628 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3632 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
3634 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3638 BUG_ON(busiest
== this_rq
);
3640 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3643 if (busiest
->nr_running
> 1) {
3644 /* Attempt to move tasks */
3645 double_lock_balance(this_rq
, busiest
);
3646 /* this_rq->clock is already updated */
3647 update_rq_clock(busiest
);
3648 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3649 imbalance
, sd
, CPU_NEWLY_IDLE
,
3651 double_unlock_balance(this_rq
, busiest
);
3653 if (unlikely(all_pinned
)) {
3654 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
3655 if (!cpumask_empty(cpus
))
3661 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3662 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3663 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3666 sd
->nr_balance_failed
= 0;
3668 update_shares_locked(this_rq
, sd
);
3672 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3673 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3674 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3676 sd
->nr_balance_failed
= 0;
3682 * idle_balance is called by schedule() if this_cpu is about to become
3683 * idle. Attempts to pull tasks from other CPUs.
3685 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3687 struct sched_domain
*sd
;
3688 int pulled_task
= 0;
3689 unsigned long next_balance
= jiffies
+ HZ
;
3690 cpumask_var_t tmpmask
;
3692 if (!alloc_cpumask_var(&tmpmask
, GFP_ATOMIC
))
3695 for_each_domain(this_cpu
, sd
) {
3696 unsigned long interval
;
3698 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3701 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3702 /* If we've pulled tasks over stop searching: */
3703 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
3706 interval
= msecs_to_jiffies(sd
->balance_interval
);
3707 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3708 next_balance
= sd
->last_balance
+ interval
;
3712 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3714 * We are going idle. next_balance may be set based on
3715 * a busy processor. So reset next_balance.
3717 this_rq
->next_balance
= next_balance
;
3719 free_cpumask_var(tmpmask
);
3723 * active_load_balance is run by migration threads. It pushes running tasks
3724 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3725 * running on each physical CPU where possible, and avoids physical /
3726 * logical imbalances.
3728 * Called with busiest_rq locked.
3730 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3732 int target_cpu
= busiest_rq
->push_cpu
;
3733 struct sched_domain
*sd
;
3734 struct rq
*target_rq
;
3736 /* Is there any task to move? */
3737 if (busiest_rq
->nr_running
<= 1)
3740 target_rq
= cpu_rq(target_cpu
);
3743 * This condition is "impossible", if it occurs
3744 * we need to fix it. Originally reported by
3745 * Bjorn Helgaas on a 128-cpu setup.
3747 BUG_ON(busiest_rq
== target_rq
);
3749 /* move a task from busiest_rq to target_rq */
3750 double_lock_balance(busiest_rq
, target_rq
);
3751 update_rq_clock(busiest_rq
);
3752 update_rq_clock(target_rq
);
3754 /* Search for an sd spanning us and the target CPU. */
3755 for_each_domain(target_cpu
, sd
) {
3756 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3757 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
3762 schedstat_inc(sd
, alb_count
);
3764 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3766 schedstat_inc(sd
, alb_pushed
);
3768 schedstat_inc(sd
, alb_failed
);
3770 double_unlock_balance(busiest_rq
, target_rq
);
3775 atomic_t load_balancer
;
3776 cpumask_var_t cpu_mask
;
3777 } nohz ____cacheline_aligned
= {
3778 .load_balancer
= ATOMIC_INIT(-1),
3782 * This routine will try to nominate the ilb (idle load balancing)
3783 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3784 * load balancing on behalf of all those cpus. If all the cpus in the system
3785 * go into this tickless mode, then there will be no ilb owner (as there is
3786 * no need for one) and all the cpus will sleep till the next wakeup event
3789 * For the ilb owner, tick is not stopped. And this tick will be used
3790 * for idle load balancing. ilb owner will still be part of
3793 * While stopping the tick, this cpu will become the ilb owner if there
3794 * is no other owner. And will be the owner till that cpu becomes busy
3795 * or if all cpus in the system stop their ticks at which point
3796 * there is no need for ilb owner.
3798 * When the ilb owner becomes busy, it nominates another owner, during the
3799 * next busy scheduler_tick()
3801 int select_nohz_load_balancer(int stop_tick
)
3803 int cpu
= smp_processor_id();
3806 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
3807 cpu_rq(cpu
)->in_nohz_recently
= 1;
3810 * If we are going offline and still the leader, give up!
3812 if (!cpu_active(cpu
) &&
3813 atomic_read(&nohz
.load_balancer
) == cpu
) {
3814 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3819 /* time for ilb owner also to sleep */
3820 if (cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3821 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3822 atomic_set(&nohz
.load_balancer
, -1);
3826 if (atomic_read(&nohz
.load_balancer
) == -1) {
3827 /* make me the ilb owner */
3828 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3830 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3833 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
3836 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
3838 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3839 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3846 static DEFINE_SPINLOCK(balancing
);
3849 * It checks each scheduling domain to see if it is due to be balanced,
3850 * and initiates a balancing operation if so.
3852 * Balancing parameters are set up in arch_init_sched_domains.
3854 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3857 struct rq
*rq
= cpu_rq(cpu
);
3858 unsigned long interval
;
3859 struct sched_domain
*sd
;
3860 /* Earliest time when we have to do rebalance again */
3861 unsigned long next_balance
= jiffies
+ 60*HZ
;
3862 int update_next_balance
= 0;
3866 /* Fails alloc? Rebalancing probably not a priority right now. */
3867 if (!alloc_cpumask_var(&tmp
, GFP_ATOMIC
))
3870 for_each_domain(cpu
, sd
) {
3871 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3874 interval
= sd
->balance_interval
;
3875 if (idle
!= CPU_IDLE
)
3876 interval
*= sd
->busy_factor
;
3878 /* scale ms to jiffies */
3879 interval
= msecs_to_jiffies(interval
);
3880 if (unlikely(!interval
))
3882 if (interval
> HZ
*NR_CPUS
/10)
3883 interval
= HZ
*NR_CPUS
/10;
3885 need_serialize
= sd
->flags
& SD_SERIALIZE
;
3887 if (need_serialize
) {
3888 if (!spin_trylock(&balancing
))
3892 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3893 if (load_balance(cpu
, rq
, sd
, idle
, &balance
, tmp
)) {
3895 * We've pulled tasks over so either we're no
3896 * longer idle, or one of our SMT siblings is
3899 idle
= CPU_NOT_IDLE
;
3901 sd
->last_balance
= jiffies
;
3904 spin_unlock(&balancing
);
3906 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3907 next_balance
= sd
->last_balance
+ interval
;
3908 update_next_balance
= 1;
3912 * Stop the load balance at this level. There is another
3913 * CPU in our sched group which is doing load balancing more
3921 * next_balance will be updated only when there is a need.
3922 * When the cpu is attached to null domain for ex, it will not be
3925 if (likely(update_next_balance
))
3926 rq
->next_balance
= next_balance
;
3928 free_cpumask_var(tmp
);
3932 * run_rebalance_domains is triggered when needed from the scheduler tick.
3933 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3934 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3936 static void run_rebalance_domains(struct softirq_action
*h
)
3938 int this_cpu
= smp_processor_id();
3939 struct rq
*this_rq
= cpu_rq(this_cpu
);
3940 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3941 CPU_IDLE
: CPU_NOT_IDLE
;
3943 rebalance_domains(this_cpu
, idle
);
3947 * If this cpu is the owner for idle load balancing, then do the
3948 * balancing on behalf of the other idle cpus whose ticks are
3951 if (this_rq
->idle_at_tick
&&
3952 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3956 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
3957 if (balance_cpu
== this_cpu
)
3961 * If this cpu gets work to do, stop the load balancing
3962 * work being done for other cpus. Next load
3963 * balancing owner will pick it up.
3968 rebalance_domains(balance_cpu
, CPU_IDLE
);
3970 rq
= cpu_rq(balance_cpu
);
3971 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3972 this_rq
->next_balance
= rq
->next_balance
;
3979 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3981 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3982 * idle load balancing owner or decide to stop the periodic load balancing,
3983 * if the whole system is idle.
3985 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3989 * If we were in the nohz mode recently and busy at the current
3990 * scheduler tick, then check if we need to nominate new idle
3993 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3994 rq
->in_nohz_recently
= 0;
3996 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3997 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
3998 atomic_set(&nohz
.load_balancer
, -1);
4001 if (atomic_read(&nohz
.load_balancer
) == -1) {
4003 * simple selection for now: Nominate the
4004 * first cpu in the nohz list to be the next
4007 * TBD: Traverse the sched domains and nominate
4008 * the nearest cpu in the nohz.cpu_mask.
4010 int ilb
= cpumask_first(nohz
.cpu_mask
);
4012 if (ilb
< nr_cpu_ids
)
4018 * If this cpu is idle and doing idle load balancing for all the
4019 * cpus with ticks stopped, is it time for that to stop?
4021 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4022 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4028 * If this cpu is idle and the idle load balancing is done by
4029 * someone else, then no need raise the SCHED_SOFTIRQ
4031 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4032 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4035 if (time_after_eq(jiffies
, rq
->next_balance
))
4036 raise_softirq(SCHED_SOFTIRQ
);
4039 #else /* CONFIG_SMP */
4042 * on UP we do not need to balance between CPUs:
4044 static inline void idle_balance(int cpu
, struct rq
*rq
)
4050 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4052 EXPORT_PER_CPU_SYMBOL(kstat
);
4055 * Return any ns on the sched_clock that have not yet been banked in
4056 * @p in case that task is currently running.
4058 unsigned long long task_delta_exec(struct task_struct
*p
)
4060 unsigned long flags
;
4064 rq
= task_rq_lock(p
, &flags
);
4066 if (task_current(rq
, p
)) {
4069 update_rq_clock(rq
);
4070 delta_exec
= rq
->clock
- p
->se
.exec_start
;
4071 if ((s64
)delta_exec
> 0)
4075 task_rq_unlock(rq
, &flags
);
4081 * Account user cpu time to a process.
4082 * @p: the process that the cpu time gets accounted to
4083 * @cputime: the cpu time spent in user space since the last update
4085 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
4087 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4090 p
->utime
= cputime_add(p
->utime
, cputime
);
4091 account_group_user_time(p
, cputime
);
4093 /* Add user time to cpustat. */
4094 tmp
= cputime_to_cputime64(cputime
);
4095 if (TASK_NICE(p
) > 0)
4096 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4098 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4099 /* Account for user time used */
4100 acct_update_integrals(p
);
4104 * Account guest cpu time to a process.
4105 * @p: the process that the cpu time gets accounted to
4106 * @cputime: the cpu time spent in virtual machine since the last update
4108 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
4111 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4113 tmp
= cputime_to_cputime64(cputime
);
4115 p
->utime
= cputime_add(p
->utime
, cputime
);
4116 account_group_user_time(p
, cputime
);
4117 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4119 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4120 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4124 * Account scaled user cpu time to a process.
4125 * @p: the process that the cpu time gets accounted to
4126 * @cputime: the cpu time spent in user space since the last update
4128 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4130 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
4134 * Account system cpu time to a process.
4135 * @p: the process that the cpu time gets accounted to
4136 * @hardirq_offset: the offset to subtract from hardirq_count()
4137 * @cputime: the cpu time spent in kernel space since the last update
4139 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4142 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4143 struct rq
*rq
= this_rq();
4146 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4147 account_guest_time(p
, cputime
);
4151 p
->stime
= cputime_add(p
->stime
, cputime
);
4152 account_group_system_time(p
, cputime
);
4154 /* Add system time to cpustat. */
4155 tmp
= cputime_to_cputime64(cputime
);
4156 if (hardirq_count() - hardirq_offset
)
4157 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4158 else if (softirq_count())
4159 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4160 else if (p
!= rq
->idle
)
4161 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4162 else if (atomic_read(&rq
->nr_iowait
) > 0)
4163 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4165 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4166 /* Account for system time used */
4167 acct_update_integrals(p
);
4171 * Account scaled system cpu time to a process.
4172 * @p: the process that the cpu time gets accounted to
4173 * @hardirq_offset: the offset to subtract from hardirq_count()
4174 * @cputime: the cpu time spent in kernel space since the last update
4176 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4178 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
4182 * Account for involuntary wait time.
4183 * @p: the process from which the cpu time has been stolen
4184 * @steal: the cpu time spent in involuntary wait
4186 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
4188 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4189 cputime64_t tmp
= cputime_to_cputime64(steal
);
4190 struct rq
*rq
= this_rq();
4192 if (p
== rq
->idle
) {
4193 p
->stime
= cputime_add(p
->stime
, steal
);
4194 if (atomic_read(&rq
->nr_iowait
) > 0)
4195 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4197 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4199 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
4203 * Use precise platform statistics if available:
4205 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4206 cputime_t
task_utime(struct task_struct
*p
)
4211 cputime_t
task_stime(struct task_struct
*p
)
4216 cputime_t
task_utime(struct task_struct
*p
)
4218 clock_t utime
= cputime_to_clock_t(p
->utime
),
4219 total
= utime
+ cputime_to_clock_t(p
->stime
);
4223 * Use CFS's precise accounting:
4225 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
4229 do_div(temp
, total
);
4231 utime
= (clock_t)temp
;
4233 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
4234 return p
->prev_utime
;
4237 cputime_t
task_stime(struct task_struct
*p
)
4242 * Use CFS's precise accounting. (we subtract utime from
4243 * the total, to make sure the total observed by userspace
4244 * grows monotonically - apps rely on that):
4246 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
4247 cputime_to_clock_t(task_utime(p
));
4250 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
4252 return p
->prev_stime
;
4256 inline cputime_t
task_gtime(struct task_struct
*p
)
4262 * This function gets called by the timer code, with HZ frequency.
4263 * We call it with interrupts disabled.
4265 * It also gets called by the fork code, when changing the parent's
4268 void scheduler_tick(void)
4270 int cpu
= smp_processor_id();
4271 struct rq
*rq
= cpu_rq(cpu
);
4272 struct task_struct
*curr
= rq
->curr
;
4276 spin_lock(&rq
->lock
);
4277 update_rq_clock(rq
);
4278 update_cpu_load(rq
);
4279 curr
->sched_class
->task_tick(rq
, curr
, 0);
4280 spin_unlock(&rq
->lock
);
4283 rq
->idle_at_tick
= idle_cpu(cpu
);
4284 trigger_load_balance(rq
, cpu
);
4288 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4289 defined(CONFIG_PREEMPT_TRACER))
4291 static inline unsigned long get_parent_ip(unsigned long addr
)
4293 if (in_lock_functions(addr
)) {
4294 addr
= CALLER_ADDR2
;
4295 if (in_lock_functions(addr
))
4296 addr
= CALLER_ADDR3
;
4301 void __kprobes
add_preempt_count(int val
)
4303 #ifdef CONFIG_DEBUG_PREEMPT
4307 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4310 preempt_count() += val
;
4311 #ifdef CONFIG_DEBUG_PREEMPT
4313 * Spinlock count overflowing soon?
4315 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4318 if (preempt_count() == val
)
4319 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4321 EXPORT_SYMBOL(add_preempt_count
);
4323 void __kprobes
sub_preempt_count(int val
)
4325 #ifdef CONFIG_DEBUG_PREEMPT
4329 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count() - (!!kernel_locked())))
4332 * Is the spinlock portion underflowing?
4334 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4335 !(preempt_count() & PREEMPT_MASK
)))
4339 if (preempt_count() == val
)
4340 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4341 preempt_count() -= val
;
4343 EXPORT_SYMBOL(sub_preempt_count
);
4348 * Print scheduling while atomic bug:
4350 static noinline
void __schedule_bug(struct task_struct
*prev
)
4352 struct pt_regs
*regs
= get_irq_regs();
4354 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4355 prev
->comm
, prev
->pid
, preempt_count());
4357 debug_show_held_locks(prev
);
4359 if (irqs_disabled())
4360 print_irqtrace_events(prev
);
4369 * Various schedule()-time debugging checks and statistics:
4371 static inline void schedule_debug(struct task_struct
*prev
)
4374 * Test if we are atomic. Since do_exit() needs to call into
4375 * schedule() atomically, we ignore that path for now.
4376 * Otherwise, whine if we are scheduling when we should not be.
4378 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4379 __schedule_bug(prev
);
4381 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4383 schedstat_inc(this_rq(), sched_count
);
4384 #ifdef CONFIG_SCHEDSTATS
4385 if (unlikely(prev
->lock_depth
>= 0)) {
4386 schedstat_inc(this_rq(), bkl_count
);
4387 schedstat_inc(prev
, sched_info
.bkl_count
);
4393 * Pick up the highest-prio task:
4395 static inline struct task_struct
*
4396 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
4398 const struct sched_class
*class;
4399 struct task_struct
*p
;
4402 * Optimization: we know that if all tasks are in
4403 * the fair class we can call that function directly:
4405 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4406 p
= fair_sched_class
.pick_next_task(rq
);
4411 class = sched_class_highest
;
4413 p
= class->pick_next_task(rq
);
4417 * Will never be NULL as the idle class always
4418 * returns a non-NULL p:
4420 class = class->next
;
4425 * schedule() is the main scheduler function.
4427 asmlinkage
void __sched
schedule(void)
4429 struct task_struct
*prev
, *next
;
4430 unsigned long *switch_count
;
4436 cpu
= smp_processor_id();
4440 switch_count
= &prev
->nivcsw
;
4442 release_kernel_lock(prev
);
4443 need_resched_nonpreemptible
:
4445 schedule_debug(prev
);
4447 if (sched_feat(HRTICK
))
4450 spin_lock_irq(&rq
->lock
);
4451 update_rq_clock(rq
);
4452 clear_tsk_need_resched(prev
);
4454 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4455 if (unlikely(signal_pending_state(prev
->state
, prev
)))
4456 prev
->state
= TASK_RUNNING
;
4458 deactivate_task(rq
, prev
, 1);
4459 switch_count
= &prev
->nvcsw
;
4463 if (prev
->sched_class
->pre_schedule
)
4464 prev
->sched_class
->pre_schedule(rq
, prev
);
4467 if (unlikely(!rq
->nr_running
))
4468 idle_balance(cpu
, rq
);
4470 prev
->sched_class
->put_prev_task(rq
, prev
);
4471 next
= pick_next_task(rq
, prev
);
4473 if (likely(prev
!= next
)) {
4474 sched_info_switch(prev
, next
);
4480 context_switch(rq
, prev
, next
); /* unlocks the rq */
4482 * the context switch might have flipped the stack from under
4483 * us, hence refresh the local variables.
4485 cpu
= smp_processor_id();
4488 spin_unlock_irq(&rq
->lock
);
4490 if (unlikely(reacquire_kernel_lock(current
) < 0))
4491 goto need_resched_nonpreemptible
;
4493 preempt_enable_no_resched();
4494 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
4497 EXPORT_SYMBOL(schedule
);
4499 #ifdef CONFIG_PREEMPT
4501 * this is the entry point to schedule() from in-kernel preemption
4502 * off of preempt_enable. Kernel preemptions off return from interrupt
4503 * occur there and call schedule directly.
4505 asmlinkage
void __sched
preempt_schedule(void)
4507 struct thread_info
*ti
= current_thread_info();
4510 * If there is a non-zero preempt_count or interrupts are disabled,
4511 * we do not want to preempt the current task. Just return..
4513 if (likely(ti
->preempt_count
|| irqs_disabled()))
4517 add_preempt_count(PREEMPT_ACTIVE
);
4519 sub_preempt_count(PREEMPT_ACTIVE
);
4522 * Check again in case we missed a preemption opportunity
4523 * between schedule and now.
4526 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4528 EXPORT_SYMBOL(preempt_schedule
);
4531 * this is the entry point to schedule() from kernel preemption
4532 * off of irq context.
4533 * Note, that this is called and return with irqs disabled. This will
4534 * protect us against recursive calling from irq.
4536 asmlinkage
void __sched
preempt_schedule_irq(void)
4538 struct thread_info
*ti
= current_thread_info();
4540 /* Catch callers which need to be fixed */
4541 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4544 add_preempt_count(PREEMPT_ACTIVE
);
4547 local_irq_disable();
4548 sub_preempt_count(PREEMPT_ACTIVE
);
4551 * Check again in case we missed a preemption opportunity
4552 * between schedule and now.
4555 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4558 #endif /* CONFIG_PREEMPT */
4560 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4563 return try_to_wake_up(curr
->private, mode
, sync
);
4565 EXPORT_SYMBOL(default_wake_function
);
4568 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4569 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4570 * number) then we wake all the non-exclusive tasks and one exclusive task.
4572 * There are circumstances in which we can try to wake a task which has already
4573 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4574 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4576 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4577 int nr_exclusive
, int sync
, void *key
)
4579 wait_queue_t
*curr
, *next
;
4581 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4582 unsigned flags
= curr
->flags
;
4584 if (curr
->func(curr
, mode
, sync
, key
) &&
4585 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4591 * __wake_up - wake up threads blocked on a waitqueue.
4593 * @mode: which threads
4594 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4595 * @key: is directly passed to the wakeup function
4597 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4598 int nr_exclusive
, void *key
)
4600 unsigned long flags
;
4602 spin_lock_irqsave(&q
->lock
, flags
);
4603 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4604 spin_unlock_irqrestore(&q
->lock
, flags
);
4606 EXPORT_SYMBOL(__wake_up
);
4609 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4611 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4613 __wake_up_common(q
, mode
, 1, 0, NULL
);
4617 * __wake_up_sync - wake up threads blocked on a waitqueue.
4619 * @mode: which threads
4620 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4622 * The sync wakeup differs that the waker knows that it will schedule
4623 * away soon, so while the target thread will be woken up, it will not
4624 * be migrated to another CPU - ie. the two threads are 'synchronized'
4625 * with each other. This can prevent needless bouncing between CPUs.
4627 * On UP it can prevent extra preemption.
4630 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4632 unsigned long flags
;
4638 if (unlikely(!nr_exclusive
))
4641 spin_lock_irqsave(&q
->lock
, flags
);
4642 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4643 spin_unlock_irqrestore(&q
->lock
, flags
);
4645 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4648 * complete: - signals a single thread waiting on this completion
4649 * @x: holds the state of this particular completion
4651 * This will wake up a single thread waiting on this completion. Threads will be
4652 * awakened in the same order in which they were queued.
4654 * See also complete_all(), wait_for_completion() and related routines.
4656 void complete(struct completion
*x
)
4658 unsigned long flags
;
4660 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4662 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4663 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4665 EXPORT_SYMBOL(complete
);
4668 * complete_all: - signals all threads waiting on this completion
4669 * @x: holds the state of this particular completion
4671 * This will wake up all threads waiting on this particular completion event.
4673 void complete_all(struct completion
*x
)
4675 unsigned long flags
;
4677 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4678 x
->done
+= UINT_MAX
/2;
4679 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4680 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4682 EXPORT_SYMBOL(complete_all
);
4684 static inline long __sched
4685 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4688 DECLARE_WAITQUEUE(wait
, current
);
4690 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4691 __add_wait_queue_tail(&x
->wait
, &wait
);
4693 if (signal_pending_state(state
, current
)) {
4694 timeout
= -ERESTARTSYS
;
4697 __set_current_state(state
);
4698 spin_unlock_irq(&x
->wait
.lock
);
4699 timeout
= schedule_timeout(timeout
);
4700 spin_lock_irq(&x
->wait
.lock
);
4701 } while (!x
->done
&& timeout
);
4702 __remove_wait_queue(&x
->wait
, &wait
);
4707 return timeout
?: 1;
4711 wait_for_common(struct completion
*x
, long timeout
, int state
)
4715 spin_lock_irq(&x
->wait
.lock
);
4716 timeout
= do_wait_for_common(x
, timeout
, state
);
4717 spin_unlock_irq(&x
->wait
.lock
);
4722 * wait_for_completion: - waits for completion of a task
4723 * @x: holds the state of this particular completion
4725 * This waits to be signaled for completion of a specific task. It is NOT
4726 * interruptible and there is no timeout.
4728 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4729 * and interrupt capability. Also see complete().
4731 void __sched
wait_for_completion(struct completion
*x
)
4733 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4735 EXPORT_SYMBOL(wait_for_completion
);
4738 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4739 * @x: holds the state of this particular completion
4740 * @timeout: timeout value in jiffies
4742 * This waits for either a completion of a specific task to be signaled or for a
4743 * specified timeout to expire. The timeout is in jiffies. It is not
4746 unsigned long __sched
4747 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4749 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4751 EXPORT_SYMBOL(wait_for_completion_timeout
);
4754 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4755 * @x: holds the state of this particular completion
4757 * This waits for completion of a specific task to be signaled. It is
4760 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4762 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4763 if (t
== -ERESTARTSYS
)
4767 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4770 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4771 * @x: holds the state of this particular completion
4772 * @timeout: timeout value in jiffies
4774 * This waits for either a completion of a specific task to be signaled or for a
4775 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4777 unsigned long __sched
4778 wait_for_completion_interruptible_timeout(struct completion
*x
,
4779 unsigned long timeout
)
4781 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4783 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4786 * wait_for_completion_killable: - waits for completion of a task (killable)
4787 * @x: holds the state of this particular completion
4789 * This waits to be signaled for completion of a specific task. It can be
4790 * interrupted by a kill signal.
4792 int __sched
wait_for_completion_killable(struct completion
*x
)
4794 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4795 if (t
== -ERESTARTSYS
)
4799 EXPORT_SYMBOL(wait_for_completion_killable
);
4802 * try_wait_for_completion - try to decrement a completion without blocking
4803 * @x: completion structure
4805 * Returns: 0 if a decrement cannot be done without blocking
4806 * 1 if a decrement succeeded.
4808 * If a completion is being used as a counting completion,
4809 * attempt to decrement the counter without blocking. This
4810 * enables us to avoid waiting if the resource the completion
4811 * is protecting is not available.
4813 bool try_wait_for_completion(struct completion
*x
)
4817 spin_lock_irq(&x
->wait
.lock
);
4822 spin_unlock_irq(&x
->wait
.lock
);
4825 EXPORT_SYMBOL(try_wait_for_completion
);
4828 * completion_done - Test to see if a completion has any waiters
4829 * @x: completion structure
4831 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4832 * 1 if there are no waiters.
4835 bool completion_done(struct completion
*x
)
4839 spin_lock_irq(&x
->wait
.lock
);
4842 spin_unlock_irq(&x
->wait
.lock
);
4845 EXPORT_SYMBOL(completion_done
);
4848 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4850 unsigned long flags
;
4853 init_waitqueue_entry(&wait
, current
);
4855 __set_current_state(state
);
4857 spin_lock_irqsave(&q
->lock
, flags
);
4858 __add_wait_queue(q
, &wait
);
4859 spin_unlock(&q
->lock
);
4860 timeout
= schedule_timeout(timeout
);
4861 spin_lock_irq(&q
->lock
);
4862 __remove_wait_queue(q
, &wait
);
4863 spin_unlock_irqrestore(&q
->lock
, flags
);
4868 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4870 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4872 EXPORT_SYMBOL(interruptible_sleep_on
);
4875 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4877 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4879 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4881 void __sched
sleep_on(wait_queue_head_t
*q
)
4883 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4885 EXPORT_SYMBOL(sleep_on
);
4887 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4889 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4891 EXPORT_SYMBOL(sleep_on_timeout
);
4893 #ifdef CONFIG_RT_MUTEXES
4896 * rt_mutex_setprio - set the current priority of a task
4898 * @prio: prio value (kernel-internal form)
4900 * This function changes the 'effective' priority of a task. It does
4901 * not touch ->normal_prio like __setscheduler().
4903 * Used by the rt_mutex code to implement priority inheritance logic.
4905 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4907 unsigned long flags
;
4908 int oldprio
, on_rq
, running
;
4910 const struct sched_class
*prev_class
= p
->sched_class
;
4912 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4914 rq
= task_rq_lock(p
, &flags
);
4915 update_rq_clock(rq
);
4918 on_rq
= p
->se
.on_rq
;
4919 running
= task_current(rq
, p
);
4921 dequeue_task(rq
, p
, 0);
4923 p
->sched_class
->put_prev_task(rq
, p
);
4926 p
->sched_class
= &rt_sched_class
;
4928 p
->sched_class
= &fair_sched_class
;
4933 p
->sched_class
->set_curr_task(rq
);
4935 enqueue_task(rq
, p
, 0);
4937 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4939 task_rq_unlock(rq
, &flags
);
4944 void set_user_nice(struct task_struct
*p
, long nice
)
4946 int old_prio
, delta
, on_rq
;
4947 unsigned long flags
;
4950 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4953 * We have to be careful, if called from sys_setpriority(),
4954 * the task might be in the middle of scheduling on another CPU.
4956 rq
= task_rq_lock(p
, &flags
);
4957 update_rq_clock(rq
);
4959 * The RT priorities are set via sched_setscheduler(), but we still
4960 * allow the 'normal' nice value to be set - but as expected
4961 * it wont have any effect on scheduling until the task is
4962 * SCHED_FIFO/SCHED_RR:
4964 if (task_has_rt_policy(p
)) {
4965 p
->static_prio
= NICE_TO_PRIO(nice
);
4968 on_rq
= p
->se
.on_rq
;
4970 dequeue_task(rq
, p
, 0);
4972 p
->static_prio
= NICE_TO_PRIO(nice
);
4975 p
->prio
= effective_prio(p
);
4976 delta
= p
->prio
- old_prio
;
4979 enqueue_task(rq
, p
, 0);
4981 * If the task increased its priority or is running and
4982 * lowered its priority, then reschedule its CPU:
4984 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4985 resched_task(rq
->curr
);
4988 task_rq_unlock(rq
, &flags
);
4990 EXPORT_SYMBOL(set_user_nice
);
4993 * can_nice - check if a task can reduce its nice value
4997 int can_nice(const struct task_struct
*p
, const int nice
)
4999 /* convert nice value [19,-20] to rlimit style value [1,40] */
5000 int nice_rlim
= 20 - nice
;
5002 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
5003 capable(CAP_SYS_NICE
));
5006 #ifdef __ARCH_WANT_SYS_NICE
5009 * sys_nice - change the priority of the current process.
5010 * @increment: priority increment
5012 * sys_setpriority is a more generic, but much slower function that
5013 * does similar things.
5015 asmlinkage
long sys_nice(int increment
)
5020 * Setpriority might change our priority at the same moment.
5021 * We don't have to worry. Conceptually one call occurs first
5022 * and we have a single winner.
5024 if (increment
< -40)
5029 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
5035 if (increment
< 0 && !can_nice(current
, nice
))
5038 retval
= security_task_setnice(current
, nice
);
5042 set_user_nice(current
, nice
);
5049 * task_prio - return the priority value of a given task.
5050 * @p: the task in question.
5052 * This is the priority value as seen by users in /proc.
5053 * RT tasks are offset by -200. Normal tasks are centered
5054 * around 0, value goes from -16 to +15.
5056 int task_prio(const struct task_struct
*p
)
5058 return p
->prio
- MAX_RT_PRIO
;
5062 * task_nice - return the nice value of a given task.
5063 * @p: the task in question.
5065 int task_nice(const struct task_struct
*p
)
5067 return TASK_NICE(p
);
5069 EXPORT_SYMBOL(task_nice
);
5072 * idle_cpu - is a given cpu idle currently?
5073 * @cpu: the processor in question.
5075 int idle_cpu(int cpu
)
5077 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
5081 * idle_task - return the idle task for a given cpu.
5082 * @cpu: the processor in question.
5084 struct task_struct
*idle_task(int cpu
)
5086 return cpu_rq(cpu
)->idle
;
5090 * find_process_by_pid - find a process with a matching PID value.
5091 * @pid: the pid in question.
5093 static struct task_struct
*find_process_by_pid(pid_t pid
)
5095 return pid
? find_task_by_vpid(pid
) : current
;
5098 /* Actually do priority change: must hold rq lock. */
5100 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
5102 BUG_ON(p
->se
.on_rq
);
5105 switch (p
->policy
) {
5109 p
->sched_class
= &fair_sched_class
;
5113 p
->sched_class
= &rt_sched_class
;
5117 p
->rt_priority
= prio
;
5118 p
->normal_prio
= normal_prio(p
);
5119 /* we are holding p->pi_lock already */
5120 p
->prio
= rt_mutex_getprio(p
);
5124 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
5125 struct sched_param
*param
, bool user
)
5127 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
5128 unsigned long flags
;
5129 const struct sched_class
*prev_class
= p
->sched_class
;
5132 /* may grab non-irq protected spin_locks */
5133 BUG_ON(in_interrupt());
5135 /* double check policy once rq lock held */
5137 policy
= oldpolicy
= p
->policy
;
5138 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
5139 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
5140 policy
!= SCHED_IDLE
)
5143 * Valid priorities for SCHED_FIFO and SCHED_RR are
5144 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5145 * SCHED_BATCH and SCHED_IDLE is 0.
5147 if (param
->sched_priority
< 0 ||
5148 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5149 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
5151 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
5155 * Allow unprivileged RT tasks to decrease priority:
5157 if (user
&& !capable(CAP_SYS_NICE
)) {
5158 if (rt_policy(policy
)) {
5159 unsigned long rlim_rtprio
;
5161 if (!lock_task_sighand(p
, &flags
))
5163 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
5164 unlock_task_sighand(p
, &flags
);
5166 /* can't set/change the rt policy */
5167 if (policy
!= p
->policy
&& !rlim_rtprio
)
5170 /* can't increase priority */
5171 if (param
->sched_priority
> p
->rt_priority
&&
5172 param
->sched_priority
> rlim_rtprio
)
5176 * Like positive nice levels, dont allow tasks to
5177 * move out of SCHED_IDLE either:
5179 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
5182 /* can't change other user's priorities */
5183 if ((current
->euid
!= p
->euid
) &&
5184 (current
->euid
!= p
->uid
))
5189 #ifdef CONFIG_RT_GROUP_SCHED
5191 * Do not allow realtime tasks into groups that have no runtime
5194 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
5195 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
5199 retval
= security_task_setscheduler(p
, policy
, param
);
5205 * make sure no PI-waiters arrive (or leave) while we are
5206 * changing the priority of the task:
5208 spin_lock_irqsave(&p
->pi_lock
, flags
);
5210 * To be able to change p->policy safely, the apropriate
5211 * runqueue lock must be held.
5213 rq
= __task_rq_lock(p
);
5214 /* recheck policy now with rq lock held */
5215 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5216 policy
= oldpolicy
= -1;
5217 __task_rq_unlock(rq
);
5218 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5221 update_rq_clock(rq
);
5222 on_rq
= p
->se
.on_rq
;
5223 running
= task_current(rq
, p
);
5225 deactivate_task(rq
, p
, 0);
5227 p
->sched_class
->put_prev_task(rq
, p
);
5230 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5233 p
->sched_class
->set_curr_task(rq
);
5235 activate_task(rq
, p
, 0);
5237 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5239 __task_rq_unlock(rq
);
5240 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5242 rt_mutex_adjust_pi(p
);
5248 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5249 * @p: the task in question.
5250 * @policy: new policy.
5251 * @param: structure containing the new RT priority.
5253 * NOTE that the task may be already dead.
5255 int sched_setscheduler(struct task_struct
*p
, int policy
,
5256 struct sched_param
*param
)
5258 return __sched_setscheduler(p
, policy
, param
, true);
5260 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5263 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5264 * @p: the task in question.
5265 * @policy: new policy.
5266 * @param: structure containing the new RT priority.
5268 * Just like sched_setscheduler, only don't bother checking if the
5269 * current context has permission. For example, this is needed in
5270 * stop_machine(): we create temporary high priority worker threads,
5271 * but our caller might not have that capability.
5273 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5274 struct sched_param
*param
)
5276 return __sched_setscheduler(p
, policy
, param
, false);
5280 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5282 struct sched_param lparam
;
5283 struct task_struct
*p
;
5286 if (!param
|| pid
< 0)
5288 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5293 p
= find_process_by_pid(pid
);
5295 retval
= sched_setscheduler(p
, policy
, &lparam
);
5302 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5303 * @pid: the pid in question.
5304 * @policy: new policy.
5305 * @param: structure containing the new RT priority.
5308 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5310 /* negative values for policy are not valid */
5314 return do_sched_setscheduler(pid
, policy
, param
);
5318 * sys_sched_setparam - set/change the RT priority of a thread
5319 * @pid: the pid in question.
5320 * @param: structure containing the new RT priority.
5322 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
5324 return do_sched_setscheduler(pid
, -1, param
);
5328 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5329 * @pid: the pid in question.
5331 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
5333 struct task_struct
*p
;
5340 read_lock(&tasklist_lock
);
5341 p
= find_process_by_pid(pid
);
5343 retval
= security_task_getscheduler(p
);
5347 read_unlock(&tasklist_lock
);
5352 * sys_sched_getscheduler - get the RT priority of a thread
5353 * @pid: the pid in question.
5354 * @param: structure containing the RT priority.
5356 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
5358 struct sched_param lp
;
5359 struct task_struct
*p
;
5362 if (!param
|| pid
< 0)
5365 read_lock(&tasklist_lock
);
5366 p
= find_process_by_pid(pid
);
5371 retval
= security_task_getscheduler(p
);
5375 lp
.sched_priority
= p
->rt_priority
;
5376 read_unlock(&tasklist_lock
);
5379 * This one might sleep, we cannot do it with a spinlock held ...
5381 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5386 read_unlock(&tasklist_lock
);
5390 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
5392 cpumask_var_t cpus_allowed
, new_mask
;
5393 struct task_struct
*p
;
5397 read_lock(&tasklist_lock
);
5399 p
= find_process_by_pid(pid
);
5401 read_unlock(&tasklist_lock
);
5407 * It is not safe to call set_cpus_allowed with the
5408 * tasklist_lock held. We will bump the task_struct's
5409 * usage count and then drop tasklist_lock.
5412 read_unlock(&tasklist_lock
);
5414 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
5418 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
5420 goto out_free_cpus_allowed
;
5423 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
5424 !capable(CAP_SYS_NICE
))
5427 retval
= security_task_setscheduler(p
, 0, NULL
);
5431 cpuset_cpus_allowed(p
, cpus_allowed
);
5432 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
5434 retval
= set_cpus_allowed_ptr(p
, new_mask
);
5437 cpuset_cpus_allowed(p
, cpus_allowed
);
5438 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
5440 * We must have raced with a concurrent cpuset
5441 * update. Just reset the cpus_allowed to the
5442 * cpuset's cpus_allowed
5444 cpumask_copy(new_mask
, cpus_allowed
);
5449 free_cpumask_var(new_mask
);
5450 out_free_cpus_allowed
:
5451 free_cpumask_var(cpus_allowed
);
5458 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5459 struct cpumask
*new_mask
)
5461 if (len
< cpumask_size())
5462 cpumask_clear(new_mask
);
5463 else if (len
> cpumask_size())
5464 len
= cpumask_size();
5466 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5470 * sys_sched_setaffinity - set the cpu affinity of a process
5471 * @pid: pid of the process
5472 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5473 * @user_mask_ptr: user-space pointer to the new cpu mask
5475 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
5476 unsigned long __user
*user_mask_ptr
)
5478 cpumask_var_t new_mask
;
5481 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
5484 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
5486 retval
= sched_setaffinity(pid
, new_mask
);
5487 free_cpumask_var(new_mask
);
5491 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
5493 struct task_struct
*p
;
5497 read_lock(&tasklist_lock
);
5500 p
= find_process_by_pid(pid
);
5504 retval
= security_task_getscheduler(p
);
5508 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
5511 read_unlock(&tasklist_lock
);
5518 * sys_sched_getaffinity - get the cpu affinity of a process
5519 * @pid: pid of the process
5520 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5521 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5523 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
5524 unsigned long __user
*user_mask_ptr
)
5529 if (len
< cpumask_size())
5532 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
5535 ret
= sched_getaffinity(pid
, mask
);
5537 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
5540 ret
= cpumask_size();
5542 free_cpumask_var(mask
);
5548 * sys_sched_yield - yield the current processor to other threads.
5550 * This function yields the current CPU to other tasks. If there are no
5551 * other threads running on this CPU then this function will return.
5553 asmlinkage
long sys_sched_yield(void)
5555 struct rq
*rq
= this_rq_lock();
5557 schedstat_inc(rq
, yld_count
);
5558 current
->sched_class
->yield_task(rq
);
5561 * Since we are going to call schedule() anyway, there's
5562 * no need to preempt or enable interrupts:
5564 __release(rq
->lock
);
5565 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5566 _raw_spin_unlock(&rq
->lock
);
5567 preempt_enable_no_resched();
5574 static void __cond_resched(void)
5576 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5577 __might_sleep(__FILE__
, __LINE__
);
5580 * The BKS might be reacquired before we have dropped
5581 * PREEMPT_ACTIVE, which could trigger a second
5582 * cond_resched() call.
5585 add_preempt_count(PREEMPT_ACTIVE
);
5587 sub_preempt_count(PREEMPT_ACTIVE
);
5588 } while (need_resched());
5591 int __sched
_cond_resched(void)
5593 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
5594 system_state
== SYSTEM_RUNNING
) {
5600 EXPORT_SYMBOL(_cond_resched
);
5603 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5604 * call schedule, and on return reacquire the lock.
5606 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5607 * operations here to prevent schedule() from being called twice (once via
5608 * spin_unlock(), once by hand).
5610 int cond_resched_lock(spinlock_t
*lock
)
5612 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
5615 if (spin_needbreak(lock
) || resched
) {
5617 if (resched
&& need_resched())
5626 EXPORT_SYMBOL(cond_resched_lock
);
5628 int __sched
cond_resched_softirq(void)
5630 BUG_ON(!in_softirq());
5632 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
5640 EXPORT_SYMBOL(cond_resched_softirq
);
5643 * yield - yield the current processor to other threads.
5645 * This is a shortcut for kernel-space yielding - it marks the
5646 * thread runnable and calls sys_sched_yield().
5648 void __sched
yield(void)
5650 set_current_state(TASK_RUNNING
);
5653 EXPORT_SYMBOL(yield
);
5656 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5657 * that process accounting knows that this is a task in IO wait state.
5659 * But don't do that if it is a deliberate, throttling IO wait (this task
5660 * has set its backing_dev_info: the queue against which it should throttle)
5662 void __sched
io_schedule(void)
5664 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5666 delayacct_blkio_start();
5667 atomic_inc(&rq
->nr_iowait
);
5669 atomic_dec(&rq
->nr_iowait
);
5670 delayacct_blkio_end();
5672 EXPORT_SYMBOL(io_schedule
);
5674 long __sched
io_schedule_timeout(long timeout
)
5676 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5679 delayacct_blkio_start();
5680 atomic_inc(&rq
->nr_iowait
);
5681 ret
= schedule_timeout(timeout
);
5682 atomic_dec(&rq
->nr_iowait
);
5683 delayacct_blkio_end();
5688 * sys_sched_get_priority_max - return maximum RT priority.
5689 * @policy: scheduling class.
5691 * this syscall returns the maximum rt_priority that can be used
5692 * by a given scheduling class.
5694 asmlinkage
long sys_sched_get_priority_max(int policy
)
5701 ret
= MAX_USER_RT_PRIO
-1;
5713 * sys_sched_get_priority_min - return minimum RT priority.
5714 * @policy: scheduling class.
5716 * this syscall returns the minimum rt_priority that can be used
5717 * by a given scheduling class.
5719 asmlinkage
long sys_sched_get_priority_min(int policy
)
5737 * sys_sched_rr_get_interval - return the default timeslice of a process.
5738 * @pid: pid of the process.
5739 * @interval: userspace pointer to the timeslice value.
5741 * this syscall writes the default timeslice value of a given process
5742 * into the user-space timespec buffer. A value of '0' means infinity.
5745 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
5747 struct task_struct
*p
;
5748 unsigned int time_slice
;
5756 read_lock(&tasklist_lock
);
5757 p
= find_process_by_pid(pid
);
5761 retval
= security_task_getscheduler(p
);
5766 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5767 * tasks that are on an otherwise idle runqueue:
5770 if (p
->policy
== SCHED_RR
) {
5771 time_slice
= DEF_TIMESLICE
;
5772 } else if (p
->policy
!= SCHED_FIFO
) {
5773 struct sched_entity
*se
= &p
->se
;
5774 unsigned long flags
;
5777 rq
= task_rq_lock(p
, &flags
);
5778 if (rq
->cfs
.load
.weight
)
5779 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5780 task_rq_unlock(rq
, &flags
);
5782 read_unlock(&tasklist_lock
);
5783 jiffies_to_timespec(time_slice
, &t
);
5784 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5788 read_unlock(&tasklist_lock
);
5792 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5794 void sched_show_task(struct task_struct
*p
)
5796 unsigned long free
= 0;
5799 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5800 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5801 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5802 #if BITS_PER_LONG == 32
5803 if (state
== TASK_RUNNING
)
5804 printk(KERN_CONT
" running ");
5806 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5808 if (state
== TASK_RUNNING
)
5809 printk(KERN_CONT
" running task ");
5811 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5813 #ifdef CONFIG_DEBUG_STACK_USAGE
5815 unsigned long *n
= end_of_stack(p
);
5818 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5821 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5822 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5824 show_stack(p
, NULL
);
5827 void show_state_filter(unsigned long state_filter
)
5829 struct task_struct
*g
, *p
;
5831 #if BITS_PER_LONG == 32
5833 " task PC stack pid father\n");
5836 " task PC stack pid father\n");
5838 read_lock(&tasklist_lock
);
5839 do_each_thread(g
, p
) {
5841 * reset the NMI-timeout, listing all files on a slow
5842 * console might take alot of time:
5844 touch_nmi_watchdog();
5845 if (!state_filter
|| (p
->state
& state_filter
))
5847 } while_each_thread(g
, p
);
5849 touch_all_softlockup_watchdogs();
5851 #ifdef CONFIG_SCHED_DEBUG
5852 sysrq_sched_debug_show();
5854 read_unlock(&tasklist_lock
);
5856 * Only show locks if all tasks are dumped:
5858 if (state_filter
== -1)
5859 debug_show_all_locks();
5862 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5864 idle
->sched_class
= &idle_sched_class
;
5868 * init_idle - set up an idle thread for a given CPU
5869 * @idle: task in question
5870 * @cpu: cpu the idle task belongs to
5872 * NOTE: this function does not set the idle thread's NEED_RESCHED
5873 * flag, to make booting more robust.
5875 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5877 struct rq
*rq
= cpu_rq(cpu
);
5878 unsigned long flags
;
5880 spin_lock_irqsave(&rq
->lock
, flags
);
5883 idle
->se
.exec_start
= sched_clock();
5885 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5886 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
5887 __set_task_cpu(idle
, cpu
);
5889 rq
->curr
= rq
->idle
= idle
;
5890 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5893 spin_unlock_irqrestore(&rq
->lock
, flags
);
5895 /* Set the preempt count _outside_ the spinlocks! */
5896 #if defined(CONFIG_PREEMPT)
5897 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5899 task_thread_info(idle
)->preempt_count
= 0;
5902 * The idle tasks have their own, simple scheduling class:
5904 idle
->sched_class
= &idle_sched_class
;
5905 ftrace_graph_init_task(idle
);
5909 * In a system that switches off the HZ timer nohz_cpu_mask
5910 * indicates which cpus entered this state. This is used
5911 * in the rcu update to wait only for active cpus. For system
5912 * which do not switch off the HZ timer nohz_cpu_mask should
5913 * always be CPU_BITS_NONE.
5915 cpumask_var_t nohz_cpu_mask
;
5918 * Increase the granularity value when there are more CPUs,
5919 * because with more CPUs the 'effective latency' as visible
5920 * to users decreases. But the relationship is not linear,
5921 * so pick a second-best guess by going with the log2 of the
5924 * This idea comes from the SD scheduler of Con Kolivas:
5926 static inline void sched_init_granularity(void)
5928 unsigned int factor
= 1 + ilog2(num_online_cpus());
5929 const unsigned long limit
= 200000000;
5931 sysctl_sched_min_granularity
*= factor
;
5932 if (sysctl_sched_min_granularity
> limit
)
5933 sysctl_sched_min_granularity
= limit
;
5935 sysctl_sched_latency
*= factor
;
5936 if (sysctl_sched_latency
> limit
)
5937 sysctl_sched_latency
= limit
;
5939 sysctl_sched_wakeup_granularity
*= factor
;
5941 sysctl_sched_shares_ratelimit
*= factor
;
5946 * This is how migration works:
5948 * 1) we queue a struct migration_req structure in the source CPU's
5949 * runqueue and wake up that CPU's migration thread.
5950 * 2) we down() the locked semaphore => thread blocks.
5951 * 3) migration thread wakes up (implicitly it forces the migrated
5952 * thread off the CPU)
5953 * 4) it gets the migration request and checks whether the migrated
5954 * task is still in the wrong runqueue.
5955 * 5) if it's in the wrong runqueue then the migration thread removes
5956 * it and puts it into the right queue.
5957 * 6) migration thread up()s the semaphore.
5958 * 7) we wake up and the migration is done.
5962 * Change a given task's CPU affinity. Migrate the thread to a
5963 * proper CPU and schedule it away if the CPU it's executing on
5964 * is removed from the allowed bitmask.
5966 * NOTE: the caller must have a valid reference to the task, the
5967 * task must not exit() & deallocate itself prematurely. The
5968 * call is not atomic; no spinlocks may be held.
5970 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
5972 struct migration_req req
;
5973 unsigned long flags
;
5977 rq
= task_rq_lock(p
, &flags
);
5978 if (!cpumask_intersects(new_mask
, cpu_online_mask
)) {
5983 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5984 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
5989 if (p
->sched_class
->set_cpus_allowed
)
5990 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5992 cpumask_copy(&p
->cpus_allowed
, new_mask
);
5993 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
5996 /* Can the task run on the task's current CPU? If so, we're done */
5997 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
6000 if (migrate_task(p
, cpumask_any_and(cpu_online_mask
, new_mask
), &req
)) {
6001 /* Need help from migration thread: drop lock and wait. */
6002 task_rq_unlock(rq
, &flags
);
6003 wake_up_process(rq
->migration_thread
);
6004 wait_for_completion(&req
.done
);
6005 tlb_migrate_finish(p
->mm
);
6009 task_rq_unlock(rq
, &flags
);
6013 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
6016 * Move (not current) task off this cpu, onto dest cpu. We're doing
6017 * this because either it can't run here any more (set_cpus_allowed()
6018 * away from this CPU, or CPU going down), or because we're
6019 * attempting to rebalance this task on exec (sched_exec).
6021 * So we race with normal scheduler movements, but that's OK, as long
6022 * as the task is no longer on this CPU.
6024 * Returns non-zero if task was successfully migrated.
6026 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6028 struct rq
*rq_dest
, *rq_src
;
6031 if (unlikely(!cpu_active(dest_cpu
)))
6034 rq_src
= cpu_rq(src_cpu
);
6035 rq_dest
= cpu_rq(dest_cpu
);
6037 double_rq_lock(rq_src
, rq_dest
);
6038 /* Already moved. */
6039 if (task_cpu(p
) != src_cpu
)
6041 /* Affinity changed (again). */
6042 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
6045 on_rq
= p
->se
.on_rq
;
6047 deactivate_task(rq_src
, p
, 0);
6049 set_task_cpu(p
, dest_cpu
);
6051 activate_task(rq_dest
, p
, 0);
6052 check_preempt_curr(rq_dest
, p
, 0);
6057 double_rq_unlock(rq_src
, rq_dest
);
6062 * migration_thread - this is a highprio system thread that performs
6063 * thread migration by bumping thread off CPU then 'pushing' onto
6066 static int migration_thread(void *data
)
6068 int cpu
= (long)data
;
6072 BUG_ON(rq
->migration_thread
!= current
);
6074 set_current_state(TASK_INTERRUPTIBLE
);
6075 while (!kthread_should_stop()) {
6076 struct migration_req
*req
;
6077 struct list_head
*head
;
6079 spin_lock_irq(&rq
->lock
);
6081 if (cpu_is_offline(cpu
)) {
6082 spin_unlock_irq(&rq
->lock
);
6086 if (rq
->active_balance
) {
6087 active_load_balance(rq
, cpu
);
6088 rq
->active_balance
= 0;
6091 head
= &rq
->migration_queue
;
6093 if (list_empty(head
)) {
6094 spin_unlock_irq(&rq
->lock
);
6096 set_current_state(TASK_INTERRUPTIBLE
);
6099 req
= list_entry(head
->next
, struct migration_req
, list
);
6100 list_del_init(head
->next
);
6102 spin_unlock(&rq
->lock
);
6103 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
6106 complete(&req
->done
);
6108 __set_current_state(TASK_RUNNING
);
6112 /* Wait for kthread_stop */
6113 set_current_state(TASK_INTERRUPTIBLE
);
6114 while (!kthread_should_stop()) {
6116 set_current_state(TASK_INTERRUPTIBLE
);
6118 __set_current_state(TASK_RUNNING
);
6122 #ifdef CONFIG_HOTPLUG_CPU
6124 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6128 local_irq_disable();
6129 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
6135 * Figure out where task on dead CPU should go, use force if necessary.
6137 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
6140 /* FIXME: Use cpumask_of_node here. */
6141 cpumask_t _nodemask
= node_to_cpumask(cpu_to_node(dead_cpu
));
6142 const struct cpumask
*nodemask
= &_nodemask
;
6145 /* Look for allowed, online CPU in same node. */
6146 for_each_cpu_and(dest_cpu
, nodemask
, cpu_online_mask
)
6147 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
6150 /* Any allowed, online CPU? */
6151 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_online_mask
);
6152 if (dest_cpu
< nr_cpu_ids
)
6155 /* No more Mr. Nice Guy. */
6156 if (dest_cpu
>= nr_cpu_ids
) {
6157 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
6158 dest_cpu
= cpumask_any_and(cpu_online_mask
, &p
->cpus_allowed
);
6161 * Don't tell them about moving exiting tasks or
6162 * kernel threads (both mm NULL), since they never
6165 if (p
->mm
&& printk_ratelimit()) {
6166 printk(KERN_INFO
"process %d (%s) no "
6167 "longer affine to cpu%d\n",
6168 task_pid_nr(p
), p
->comm
, dead_cpu
);
6173 /* It can have affinity changed while we were choosing. */
6174 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
6179 * While a dead CPU has no uninterruptible tasks queued at this point,
6180 * it might still have a nonzero ->nr_uninterruptible counter, because
6181 * for performance reasons the counter is not stricly tracking tasks to
6182 * their home CPUs. So we just add the counter to another CPU's counter,
6183 * to keep the global sum constant after CPU-down:
6185 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
6187 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_online_mask
));
6188 unsigned long flags
;
6190 local_irq_save(flags
);
6191 double_rq_lock(rq_src
, rq_dest
);
6192 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
6193 rq_src
->nr_uninterruptible
= 0;
6194 double_rq_unlock(rq_src
, rq_dest
);
6195 local_irq_restore(flags
);
6198 /* Run through task list and migrate tasks from the dead cpu. */
6199 static void migrate_live_tasks(int src_cpu
)
6201 struct task_struct
*p
, *t
;
6203 read_lock(&tasklist_lock
);
6205 do_each_thread(t
, p
) {
6209 if (task_cpu(p
) == src_cpu
)
6210 move_task_off_dead_cpu(src_cpu
, p
);
6211 } while_each_thread(t
, p
);
6213 read_unlock(&tasklist_lock
);
6217 * Schedules idle task to be the next runnable task on current CPU.
6218 * It does so by boosting its priority to highest possible.
6219 * Used by CPU offline code.
6221 void sched_idle_next(void)
6223 int this_cpu
= smp_processor_id();
6224 struct rq
*rq
= cpu_rq(this_cpu
);
6225 struct task_struct
*p
= rq
->idle
;
6226 unsigned long flags
;
6228 /* cpu has to be offline */
6229 BUG_ON(cpu_online(this_cpu
));
6232 * Strictly not necessary since rest of the CPUs are stopped by now
6233 * and interrupts disabled on the current cpu.
6235 spin_lock_irqsave(&rq
->lock
, flags
);
6237 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6239 update_rq_clock(rq
);
6240 activate_task(rq
, p
, 0);
6242 spin_unlock_irqrestore(&rq
->lock
, flags
);
6246 * Ensures that the idle task is using init_mm right before its cpu goes
6249 void idle_task_exit(void)
6251 struct mm_struct
*mm
= current
->active_mm
;
6253 BUG_ON(cpu_online(smp_processor_id()));
6256 switch_mm(mm
, &init_mm
, current
);
6260 /* called under rq->lock with disabled interrupts */
6261 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
6263 struct rq
*rq
= cpu_rq(dead_cpu
);
6265 /* Must be exiting, otherwise would be on tasklist. */
6266 BUG_ON(!p
->exit_state
);
6268 /* Cannot have done final schedule yet: would have vanished. */
6269 BUG_ON(p
->state
== TASK_DEAD
);
6274 * Drop lock around migration; if someone else moves it,
6275 * that's OK. No task can be added to this CPU, so iteration is
6278 spin_unlock_irq(&rq
->lock
);
6279 move_task_off_dead_cpu(dead_cpu
, p
);
6280 spin_lock_irq(&rq
->lock
);
6285 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6286 static void migrate_dead_tasks(unsigned int dead_cpu
)
6288 struct rq
*rq
= cpu_rq(dead_cpu
);
6289 struct task_struct
*next
;
6292 if (!rq
->nr_running
)
6294 update_rq_clock(rq
);
6295 next
= pick_next_task(rq
, rq
->curr
);
6298 next
->sched_class
->put_prev_task(rq
, next
);
6299 migrate_dead(dead_cpu
, next
);
6303 #endif /* CONFIG_HOTPLUG_CPU */
6305 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6307 static struct ctl_table sd_ctl_dir
[] = {
6309 .procname
= "sched_domain",
6315 static struct ctl_table sd_ctl_root
[] = {
6317 .ctl_name
= CTL_KERN
,
6318 .procname
= "kernel",
6320 .child
= sd_ctl_dir
,
6325 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6327 struct ctl_table
*entry
=
6328 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6333 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6335 struct ctl_table
*entry
;
6338 * In the intermediate directories, both the child directory and
6339 * procname are dynamically allocated and could fail but the mode
6340 * will always be set. In the lowest directory the names are
6341 * static strings and all have proc handlers.
6343 for (entry
= *tablep
; entry
->mode
; entry
++) {
6345 sd_free_ctl_entry(&entry
->child
);
6346 if (entry
->proc_handler
== NULL
)
6347 kfree(entry
->procname
);
6355 set_table_entry(struct ctl_table
*entry
,
6356 const char *procname
, void *data
, int maxlen
,
6357 mode_t mode
, proc_handler
*proc_handler
)
6359 entry
->procname
= procname
;
6361 entry
->maxlen
= maxlen
;
6363 entry
->proc_handler
= proc_handler
;
6366 static struct ctl_table
*
6367 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6369 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
6374 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6375 sizeof(long), 0644, proc_doulongvec_minmax
);
6376 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6377 sizeof(long), 0644, proc_doulongvec_minmax
);
6378 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6379 sizeof(int), 0644, proc_dointvec_minmax
);
6380 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6381 sizeof(int), 0644, proc_dointvec_minmax
);
6382 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6383 sizeof(int), 0644, proc_dointvec_minmax
);
6384 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6385 sizeof(int), 0644, proc_dointvec_minmax
);
6386 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6387 sizeof(int), 0644, proc_dointvec_minmax
);
6388 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6389 sizeof(int), 0644, proc_dointvec_minmax
);
6390 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6391 sizeof(int), 0644, proc_dointvec_minmax
);
6392 set_table_entry(&table
[9], "cache_nice_tries",
6393 &sd
->cache_nice_tries
,
6394 sizeof(int), 0644, proc_dointvec_minmax
);
6395 set_table_entry(&table
[10], "flags", &sd
->flags
,
6396 sizeof(int), 0644, proc_dointvec_minmax
);
6397 set_table_entry(&table
[11], "name", sd
->name
,
6398 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
6399 /* &table[12] is terminator */
6404 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6406 struct ctl_table
*entry
, *table
;
6407 struct sched_domain
*sd
;
6408 int domain_num
= 0, i
;
6411 for_each_domain(cpu
, sd
)
6413 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6418 for_each_domain(cpu
, sd
) {
6419 snprintf(buf
, 32, "domain%d", i
);
6420 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6422 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6429 static struct ctl_table_header
*sd_sysctl_header
;
6430 static void register_sched_domain_sysctl(void)
6432 int i
, cpu_num
= num_online_cpus();
6433 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6436 WARN_ON(sd_ctl_dir
[0].child
);
6437 sd_ctl_dir
[0].child
= entry
;
6442 for_each_online_cpu(i
) {
6443 snprintf(buf
, 32, "cpu%d", i
);
6444 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6446 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6450 WARN_ON(sd_sysctl_header
);
6451 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6454 /* may be called multiple times per register */
6455 static void unregister_sched_domain_sysctl(void)
6457 if (sd_sysctl_header
)
6458 unregister_sysctl_table(sd_sysctl_header
);
6459 sd_sysctl_header
= NULL
;
6460 if (sd_ctl_dir
[0].child
)
6461 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6464 static void register_sched_domain_sysctl(void)
6467 static void unregister_sched_domain_sysctl(void)
6472 static void set_rq_online(struct rq
*rq
)
6475 const struct sched_class
*class;
6477 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
6480 for_each_class(class) {
6481 if (class->rq_online
)
6482 class->rq_online(rq
);
6487 static void set_rq_offline(struct rq
*rq
)
6490 const struct sched_class
*class;
6492 for_each_class(class) {
6493 if (class->rq_offline
)
6494 class->rq_offline(rq
);
6497 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
6503 * migration_call - callback that gets triggered when a CPU is added.
6504 * Here we can start up the necessary migration thread for the new CPU.
6506 static int __cpuinit
6507 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6509 struct task_struct
*p
;
6510 int cpu
= (long)hcpu
;
6511 unsigned long flags
;
6516 case CPU_UP_PREPARE
:
6517 case CPU_UP_PREPARE_FROZEN
:
6518 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
6521 kthread_bind(p
, cpu
);
6522 /* Must be high prio: stop_machine expects to yield to it. */
6523 rq
= task_rq_lock(p
, &flags
);
6524 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6525 task_rq_unlock(rq
, &flags
);
6526 cpu_rq(cpu
)->migration_thread
= p
;
6530 case CPU_ONLINE_FROZEN
:
6531 /* Strictly unnecessary, as first user will wake it. */
6532 wake_up_process(cpu_rq(cpu
)->migration_thread
);
6534 /* Update our root-domain */
6536 spin_lock_irqsave(&rq
->lock
, flags
);
6538 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6542 spin_unlock_irqrestore(&rq
->lock
, flags
);
6545 #ifdef CONFIG_HOTPLUG_CPU
6546 case CPU_UP_CANCELED
:
6547 case CPU_UP_CANCELED_FROZEN
:
6548 if (!cpu_rq(cpu
)->migration_thread
)
6550 /* Unbind it from offline cpu so it can run. Fall thru. */
6551 kthread_bind(cpu_rq(cpu
)->migration_thread
,
6552 cpumask_any(cpu_online_mask
));
6553 kthread_stop(cpu_rq(cpu
)->migration_thread
);
6554 cpu_rq(cpu
)->migration_thread
= NULL
;
6558 case CPU_DEAD_FROZEN
:
6559 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6560 migrate_live_tasks(cpu
);
6562 kthread_stop(rq
->migration_thread
);
6563 rq
->migration_thread
= NULL
;
6564 /* Idle task back to normal (off runqueue, low prio) */
6565 spin_lock_irq(&rq
->lock
);
6566 update_rq_clock(rq
);
6567 deactivate_task(rq
, rq
->idle
, 0);
6568 rq
->idle
->static_prio
= MAX_PRIO
;
6569 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
6570 rq
->idle
->sched_class
= &idle_sched_class
;
6571 migrate_dead_tasks(cpu
);
6572 spin_unlock_irq(&rq
->lock
);
6574 migrate_nr_uninterruptible(rq
);
6575 BUG_ON(rq
->nr_running
!= 0);
6578 * No need to migrate the tasks: it was best-effort if
6579 * they didn't take sched_hotcpu_mutex. Just wake up
6582 spin_lock_irq(&rq
->lock
);
6583 while (!list_empty(&rq
->migration_queue
)) {
6584 struct migration_req
*req
;
6586 req
= list_entry(rq
->migration_queue
.next
,
6587 struct migration_req
, list
);
6588 list_del_init(&req
->list
);
6589 spin_unlock_irq(&rq
->lock
);
6590 complete(&req
->done
);
6591 spin_lock_irq(&rq
->lock
);
6593 spin_unlock_irq(&rq
->lock
);
6597 case CPU_DYING_FROZEN
:
6598 /* Update our root-domain */
6600 spin_lock_irqsave(&rq
->lock
, flags
);
6602 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6605 spin_unlock_irqrestore(&rq
->lock
, flags
);
6612 /* Register at highest priority so that task migration (migrate_all_tasks)
6613 * happens before everything else.
6615 static struct notifier_block __cpuinitdata migration_notifier
= {
6616 .notifier_call
= migration_call
,
6620 static int __init
migration_init(void)
6622 void *cpu
= (void *)(long)smp_processor_id();
6625 /* Start one for the boot CPU: */
6626 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6627 BUG_ON(err
== NOTIFY_BAD
);
6628 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6629 register_cpu_notifier(&migration_notifier
);
6633 early_initcall(migration_init
);
6638 #ifdef CONFIG_SCHED_DEBUG
6640 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6641 struct cpumask
*groupmask
)
6643 struct sched_group
*group
= sd
->groups
;
6646 cpulist_scnprintf(str
, sizeof(str
), *sched_domain_span(sd
));
6647 cpumask_clear(groupmask
);
6649 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6651 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6652 printk("does not load-balance\n");
6654 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6659 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
6661 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
6662 printk(KERN_ERR
"ERROR: domain->span does not contain "
6665 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
6666 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6670 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6674 printk(KERN_ERR
"ERROR: group is NULL\n");
6678 if (!group
->__cpu_power
) {
6679 printk(KERN_CONT
"\n");
6680 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6685 if (!cpumask_weight(sched_group_cpus(group
))) {
6686 printk(KERN_CONT
"\n");
6687 printk(KERN_ERR
"ERROR: empty group\n");
6691 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
6692 printk(KERN_CONT
"\n");
6693 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6697 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
6699 cpulist_scnprintf(str
, sizeof(str
), *sched_group_cpus(group
));
6700 printk(KERN_CONT
" %s", str
);
6702 group
= group
->next
;
6703 } while (group
!= sd
->groups
);
6704 printk(KERN_CONT
"\n");
6706 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
6707 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6710 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
6711 printk(KERN_ERR
"ERROR: parent span is not a superset "
6712 "of domain->span\n");
6716 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6718 cpumask_var_t groupmask
;
6722 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6726 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6728 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
6729 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6734 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6741 free_cpumask_var(groupmask
);
6743 #else /* !CONFIG_SCHED_DEBUG */
6744 # define sched_domain_debug(sd, cpu) do { } while (0)
6745 #endif /* CONFIG_SCHED_DEBUG */
6747 static int sd_degenerate(struct sched_domain
*sd
)
6749 if (cpumask_weight(sched_domain_span(sd
)) == 1)
6752 /* Following flags need at least 2 groups */
6753 if (sd
->flags
& (SD_LOAD_BALANCE
|
6754 SD_BALANCE_NEWIDLE
|
6758 SD_SHARE_PKG_RESOURCES
)) {
6759 if (sd
->groups
!= sd
->groups
->next
)
6763 /* Following flags don't use groups */
6764 if (sd
->flags
& (SD_WAKE_IDLE
|
6773 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6775 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6777 if (sd_degenerate(parent
))
6780 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
6783 /* Does parent contain flags not in child? */
6784 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6785 if (cflags
& SD_WAKE_AFFINE
)
6786 pflags
&= ~SD_WAKE_BALANCE
;
6787 /* Flags needing groups don't count if only 1 group in parent */
6788 if (parent
->groups
== parent
->groups
->next
) {
6789 pflags
&= ~(SD_LOAD_BALANCE
|
6790 SD_BALANCE_NEWIDLE
|
6794 SD_SHARE_PKG_RESOURCES
);
6795 if (nr_node_ids
== 1)
6796 pflags
&= ~SD_SERIALIZE
;
6798 if (~cflags
& pflags
)
6804 static void free_rootdomain(struct root_domain
*rd
)
6806 cpupri_cleanup(&rd
->cpupri
);
6808 free_cpumask_var(rd
->rto_mask
);
6809 free_cpumask_var(rd
->online
);
6810 free_cpumask_var(rd
->span
);
6814 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6816 unsigned long flags
;
6818 spin_lock_irqsave(&rq
->lock
, flags
);
6821 struct root_domain
*old_rd
= rq
->rd
;
6823 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
6826 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
6828 if (atomic_dec_and_test(&old_rd
->refcount
))
6829 free_rootdomain(old_rd
);
6832 atomic_inc(&rd
->refcount
);
6835 cpumask_set_cpu(rq
->cpu
, rd
->span
);
6836 if (cpumask_test_cpu(rq
->cpu
, cpu_online_mask
))
6839 spin_unlock_irqrestore(&rq
->lock
, flags
);
6842 static int init_rootdomain(struct root_domain
*rd
, bool bootmem
)
6844 memset(rd
, 0, sizeof(*rd
));
6847 alloc_bootmem_cpumask_var(&def_root_domain
.span
);
6848 alloc_bootmem_cpumask_var(&def_root_domain
.online
);
6849 alloc_bootmem_cpumask_var(&def_root_domain
.rto_mask
);
6850 cpupri_init(&rd
->cpupri
, true);
6854 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
6856 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
6858 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
6861 if (cpupri_init(&rd
->cpupri
, false) != 0)
6866 free_cpumask_var(rd
->rto_mask
);
6868 free_cpumask_var(rd
->online
);
6870 free_cpumask_var(rd
->span
);
6876 static void init_defrootdomain(void)
6878 init_rootdomain(&def_root_domain
, true);
6880 atomic_set(&def_root_domain
.refcount
, 1);
6883 static struct root_domain
*alloc_rootdomain(void)
6885 struct root_domain
*rd
;
6887 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6891 if (init_rootdomain(rd
, false) != 0) {
6900 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6901 * hold the hotplug lock.
6904 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6906 struct rq
*rq
= cpu_rq(cpu
);
6907 struct sched_domain
*tmp
;
6909 /* Remove the sched domains which do not contribute to scheduling. */
6910 for (tmp
= sd
; tmp
; ) {
6911 struct sched_domain
*parent
= tmp
->parent
;
6915 if (sd_parent_degenerate(tmp
, parent
)) {
6916 tmp
->parent
= parent
->parent
;
6918 parent
->parent
->child
= tmp
;
6923 if (sd
&& sd_degenerate(sd
)) {
6929 sched_domain_debug(sd
, cpu
);
6931 rq_attach_root(rq
, rd
);
6932 rcu_assign_pointer(rq
->sd
, sd
);
6935 /* cpus with isolated domains */
6936 static cpumask_var_t cpu_isolated_map
;
6938 /* Setup the mask of cpus configured for isolated domains */
6939 static int __init
isolated_cpu_setup(char *str
)
6941 cpulist_parse(str
, *cpu_isolated_map
);
6945 __setup("isolcpus=", isolated_cpu_setup
);
6948 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6949 * to a function which identifies what group(along with sched group) a CPU
6950 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6951 * (due to the fact that we keep track of groups covered with a struct cpumask).
6953 * init_sched_build_groups will build a circular linked list of the groups
6954 * covered by the given span, and will set each group's ->cpumask correctly,
6955 * and ->cpu_power to 0.
6958 init_sched_build_groups(const struct cpumask
*span
,
6959 const struct cpumask
*cpu_map
,
6960 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
6961 struct sched_group
**sg
,
6962 struct cpumask
*tmpmask
),
6963 struct cpumask
*covered
, struct cpumask
*tmpmask
)
6965 struct sched_group
*first
= NULL
, *last
= NULL
;
6968 cpumask_clear(covered
);
6970 for_each_cpu(i
, span
) {
6971 struct sched_group
*sg
;
6972 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6975 if (cpumask_test_cpu(i
, covered
))
6978 cpumask_clear(sched_group_cpus(sg
));
6979 sg
->__cpu_power
= 0;
6981 for_each_cpu(j
, span
) {
6982 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6985 cpumask_set_cpu(j
, covered
);
6986 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6997 #define SD_NODES_PER_DOMAIN 16
7002 * find_next_best_node - find the next node to include in a sched_domain
7003 * @node: node whose sched_domain we're building
7004 * @used_nodes: nodes already in the sched_domain
7006 * Find the next node to include in a given scheduling domain. Simply
7007 * finds the closest node not already in the @used_nodes map.
7009 * Should use nodemask_t.
7011 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
7013 int i
, n
, val
, min_val
, best_node
= 0;
7017 for (i
= 0; i
< nr_node_ids
; i
++) {
7018 /* Start at @node */
7019 n
= (node
+ i
) % nr_node_ids
;
7021 if (!nr_cpus_node(n
))
7024 /* Skip already used nodes */
7025 if (node_isset(n
, *used_nodes
))
7028 /* Simple min distance search */
7029 val
= node_distance(node
, n
);
7031 if (val
< min_val
) {
7037 node_set(best_node
, *used_nodes
);
7042 * sched_domain_node_span - get a cpumask for a node's sched_domain
7043 * @node: node whose cpumask we're constructing
7044 * @span: resulting cpumask
7046 * Given a node, construct a good cpumask for its sched_domain to span. It
7047 * should be one that prevents unnecessary balancing, but also spreads tasks
7050 static void sched_domain_node_span(int node
, struct cpumask
*span
)
7052 nodemask_t used_nodes
;
7053 /* FIXME: use cpumask_of_node() */
7054 node_to_cpumask_ptr(nodemask
, node
);
7058 nodes_clear(used_nodes
);
7060 cpus_or(*span
, *span
, *nodemask
);
7061 node_set(node
, used_nodes
);
7063 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
7064 int next_node
= find_next_best_node(node
, &used_nodes
);
7066 node_to_cpumask_ptr_next(nodemask
, next_node
);
7067 cpus_or(*span
, *span
, *nodemask
);
7070 #endif /* CONFIG_NUMA */
7072 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
7075 * The cpus mask in sched_group and sched_domain hangs off the end.
7076 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7077 * for nr_cpu_ids < CONFIG_NR_CPUS.
7079 struct static_sched_group
{
7080 struct sched_group sg
;
7081 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
7084 struct static_sched_domain
{
7085 struct sched_domain sd
;
7086 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
7090 * SMT sched-domains:
7092 #ifdef CONFIG_SCHED_SMT
7093 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
7094 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_cpus
);
7097 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
7098 struct sched_group
**sg
, struct cpumask
*unused
)
7101 *sg
= &per_cpu(sched_group_cpus
, cpu
).sg
;
7104 #endif /* CONFIG_SCHED_SMT */
7107 * multi-core sched-domains:
7109 #ifdef CONFIG_SCHED_MC
7110 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
7111 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
7112 #endif /* CONFIG_SCHED_MC */
7114 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7116 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
7117 struct sched_group
**sg
, struct cpumask
*mask
)
7121 cpumask_and(mask
, &per_cpu(cpu_sibling_map
, cpu
), cpu_map
);
7122 group
= cpumask_first(mask
);
7124 *sg
= &per_cpu(sched_group_core
, group
).sg
;
7127 #elif defined(CONFIG_SCHED_MC)
7129 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
7130 struct sched_group
**sg
, struct cpumask
*unused
)
7133 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
7138 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
7139 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
7142 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
7143 struct sched_group
**sg
, struct cpumask
*mask
)
7146 #ifdef CONFIG_SCHED_MC
7147 /* FIXME: Use cpu_coregroup_mask. */
7148 *mask
= cpu_coregroup_map(cpu
);
7149 cpus_and(*mask
, *mask
, *cpu_map
);
7150 group
= cpumask_first(mask
);
7151 #elif defined(CONFIG_SCHED_SMT)
7152 cpumask_and(mask
, &per_cpu(cpu_sibling_map
, cpu
), cpu_map
);
7153 group
= cpumask_first(mask
);
7158 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
7164 * The init_sched_build_groups can't handle what we want to do with node
7165 * groups, so roll our own. Now each node has its own list of groups which
7166 * gets dynamically allocated.
7168 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
7169 static struct sched_group
***sched_group_nodes_bycpu
;
7171 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
7172 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
7174 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
7175 struct sched_group
**sg
,
7176 struct cpumask
*nodemask
)
7179 /* FIXME: use cpumask_of_node */
7180 node_to_cpumask_ptr(pnodemask
, cpu_to_node(cpu
));
7182 cpumask_and(nodemask
, pnodemask
, cpu_map
);
7183 group
= cpumask_first(nodemask
);
7186 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
7190 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
7192 struct sched_group
*sg
= group_head
;
7198 for_each_cpu(j
, sched_group_cpus(sg
)) {
7199 struct sched_domain
*sd
;
7201 sd
= &per_cpu(phys_domains
, j
).sd
;
7202 if (j
!= cpumask_first(sched_group_cpus(sd
->groups
))) {
7204 * Only add "power" once for each
7210 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
7213 } while (sg
!= group_head
);
7215 #endif /* CONFIG_NUMA */
7218 /* Free memory allocated for various sched_group structures */
7219 static void free_sched_groups(const struct cpumask
*cpu_map
,
7220 struct cpumask
*nodemask
)
7224 for_each_cpu(cpu
, cpu_map
) {
7225 struct sched_group
**sched_group_nodes
7226 = sched_group_nodes_bycpu
[cpu
];
7228 if (!sched_group_nodes
)
7231 for (i
= 0; i
< nr_node_ids
; i
++) {
7232 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
7233 /* FIXME: Use cpumask_of_node */
7234 node_to_cpumask_ptr(pnodemask
, i
);
7236 cpus_and(*nodemask
, *pnodemask
, *cpu_map
);
7237 if (cpumask_empty(nodemask
))
7247 if (oldsg
!= sched_group_nodes
[i
])
7250 kfree(sched_group_nodes
);
7251 sched_group_nodes_bycpu
[cpu
] = NULL
;
7254 #else /* !CONFIG_NUMA */
7255 static void free_sched_groups(const struct cpumask
*cpu_map
,
7256 struct cpumask
*nodemask
)
7259 #endif /* CONFIG_NUMA */
7262 * Initialize sched groups cpu_power.
7264 * cpu_power indicates the capacity of sched group, which is used while
7265 * distributing the load between different sched groups in a sched domain.
7266 * Typically cpu_power for all the groups in a sched domain will be same unless
7267 * there are asymmetries in the topology. If there are asymmetries, group
7268 * having more cpu_power will pickup more load compared to the group having
7271 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7272 * the maximum number of tasks a group can handle in the presence of other idle
7273 * or lightly loaded groups in the same sched domain.
7275 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7277 struct sched_domain
*child
;
7278 struct sched_group
*group
;
7280 WARN_ON(!sd
|| !sd
->groups
);
7282 if (cpu
!= cpumask_first(sched_group_cpus(sd
->groups
)))
7287 sd
->groups
->__cpu_power
= 0;
7290 * For perf policy, if the groups in child domain share resources
7291 * (for example cores sharing some portions of the cache hierarchy
7292 * or SMT), then set this domain groups cpu_power such that each group
7293 * can handle only one task, when there are other idle groups in the
7294 * same sched domain.
7296 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
7298 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
7299 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
7304 * add cpu_power of each child group to this groups cpu_power
7306 group
= child
->groups
;
7308 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
7309 group
= group
->next
;
7310 } while (group
!= child
->groups
);
7314 * Initializers for schedule domains
7315 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7318 #ifdef CONFIG_SCHED_DEBUG
7319 # define SD_INIT_NAME(sd, type) sd->name = #type
7321 # define SD_INIT_NAME(sd, type) do { } while (0)
7324 #define SD_INIT(sd, type) sd_init_##type(sd)
7326 #define SD_INIT_FUNC(type) \
7327 static noinline void sd_init_##type(struct sched_domain *sd) \
7329 memset(sd, 0, sizeof(*sd)); \
7330 *sd = SD_##type##_INIT; \
7331 sd->level = SD_LV_##type; \
7332 SD_INIT_NAME(sd, type); \
7337 SD_INIT_FUNC(ALLNODES
)
7340 #ifdef CONFIG_SCHED_SMT
7341 SD_INIT_FUNC(SIBLING
)
7343 #ifdef CONFIG_SCHED_MC
7347 static int default_relax_domain_level
= -1;
7349 static int __init
setup_relax_domain_level(char *str
)
7353 val
= simple_strtoul(str
, NULL
, 0);
7354 if (val
< SD_LV_MAX
)
7355 default_relax_domain_level
= val
;
7359 __setup("relax_domain_level=", setup_relax_domain_level
);
7361 static void set_domain_attribute(struct sched_domain
*sd
,
7362 struct sched_domain_attr
*attr
)
7366 if (!attr
|| attr
->relax_domain_level
< 0) {
7367 if (default_relax_domain_level
< 0)
7370 request
= default_relax_domain_level
;
7372 request
= attr
->relax_domain_level
;
7373 if (request
< sd
->level
) {
7374 /* turn off idle balance on this domain */
7375 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
7377 /* turn on idle balance on this domain */
7378 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
7383 * Build sched domains for a given set of cpus and attach the sched domains
7384 * to the individual cpus
7386 static int __build_sched_domains(const struct cpumask
*cpu_map
,
7387 struct sched_domain_attr
*attr
)
7389 int i
, err
= -ENOMEM
;
7390 struct root_domain
*rd
;
7391 cpumask_var_t nodemask
, this_sibling_map
, this_core_map
, send_covered
,
7394 cpumask_var_t domainspan
, covered
, notcovered
;
7395 struct sched_group
**sched_group_nodes
= NULL
;
7396 int sd_allnodes
= 0;
7398 if (!alloc_cpumask_var(&domainspan
, GFP_KERNEL
))
7400 if (!alloc_cpumask_var(&covered
, GFP_KERNEL
))
7401 goto free_domainspan
;
7402 if (!alloc_cpumask_var(¬covered
, GFP_KERNEL
))
7406 if (!alloc_cpumask_var(&nodemask
, GFP_KERNEL
))
7407 goto free_notcovered
;
7408 if (!alloc_cpumask_var(&this_sibling_map
, GFP_KERNEL
))
7410 if (!alloc_cpumask_var(&this_core_map
, GFP_KERNEL
))
7411 goto free_this_sibling_map
;
7412 if (!alloc_cpumask_var(&send_covered
, GFP_KERNEL
))
7413 goto free_this_core_map
;
7414 if (!alloc_cpumask_var(&tmpmask
, GFP_KERNEL
))
7415 goto free_send_covered
;
7419 * Allocate the per-node list of sched groups
7421 sched_group_nodes
= kcalloc(nr_node_ids
, sizeof(struct sched_group
*),
7423 if (!sched_group_nodes
) {
7424 printk(KERN_WARNING
"Can not alloc sched group node list\n");
7429 rd
= alloc_rootdomain();
7431 printk(KERN_WARNING
"Cannot alloc root domain\n");
7432 goto free_sched_groups
;
7436 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = sched_group_nodes
;
7440 * Set up domains for cpus specified by the cpu_map.
7442 for_each_cpu(i
, cpu_map
) {
7443 struct sched_domain
*sd
= NULL
, *p
;
7445 /* FIXME: use cpumask_of_node */
7446 *nodemask
= node_to_cpumask(cpu_to_node(i
));
7447 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7450 if (cpumask_weight(cpu_map
) >
7451 SD_NODES_PER_DOMAIN
*cpumask_weight(nodemask
)) {
7452 sd
= &per_cpu(allnodes_domains
, i
);
7453 SD_INIT(sd
, ALLNODES
);
7454 set_domain_attribute(sd
, attr
);
7455 cpumask_copy(sched_domain_span(sd
), cpu_map
);
7456 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7462 sd
= &per_cpu(node_domains
, i
);
7464 set_domain_attribute(sd
, attr
);
7465 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
7469 cpumask_and(sched_domain_span(sd
),
7470 sched_domain_span(sd
), cpu_map
);
7474 sd
= &per_cpu(phys_domains
, i
).sd
;
7476 set_domain_attribute(sd
, attr
);
7477 cpumask_copy(sched_domain_span(sd
), nodemask
);
7481 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7483 #ifdef CONFIG_SCHED_MC
7485 sd
= &per_cpu(core_domains
, i
).sd
;
7487 set_domain_attribute(sd
, attr
);
7488 *sched_domain_span(sd
) = cpu_coregroup_map(i
);
7489 cpumask_and(sched_domain_span(sd
),
7490 sched_domain_span(sd
), cpu_map
);
7493 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7496 #ifdef CONFIG_SCHED_SMT
7498 sd
= &per_cpu(cpu_domains
, i
).sd
;
7499 SD_INIT(sd
, SIBLING
);
7500 set_domain_attribute(sd
, attr
);
7501 cpumask_and(sched_domain_span(sd
),
7502 &per_cpu(cpu_sibling_map
, i
), cpu_map
);
7505 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7509 #ifdef CONFIG_SCHED_SMT
7510 /* Set up CPU (sibling) groups */
7511 for_each_cpu(i
, cpu_map
) {
7512 cpumask_and(this_sibling_map
,
7513 &per_cpu(cpu_sibling_map
, i
), cpu_map
);
7514 if (i
!= cpumask_first(this_sibling_map
))
7517 init_sched_build_groups(this_sibling_map
, cpu_map
,
7519 send_covered
, tmpmask
);
7523 #ifdef CONFIG_SCHED_MC
7524 /* Set up multi-core groups */
7525 for_each_cpu(i
, cpu_map
) {
7526 /* FIXME: Use cpu_coregroup_mask */
7527 *this_core_map
= cpu_coregroup_map(i
);
7528 cpus_and(*this_core_map
, *this_core_map
, *cpu_map
);
7529 if (i
!= cpumask_first(this_core_map
))
7532 init_sched_build_groups(this_core_map
, cpu_map
,
7534 send_covered
, tmpmask
);
7538 /* Set up physical groups */
7539 for (i
= 0; i
< nr_node_ids
; i
++) {
7540 /* FIXME: Use cpumask_of_node */
7541 *nodemask
= node_to_cpumask(i
);
7542 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7543 if (cpumask_empty(nodemask
))
7546 init_sched_build_groups(nodemask
, cpu_map
,
7548 send_covered
, tmpmask
);
7552 /* Set up node groups */
7554 init_sched_build_groups(cpu_map
, cpu_map
,
7555 &cpu_to_allnodes_group
,
7556 send_covered
, tmpmask
);
7559 for (i
= 0; i
< nr_node_ids
; i
++) {
7560 /* Set up node groups */
7561 struct sched_group
*sg
, *prev
;
7564 /* FIXME: Use cpumask_of_node */
7565 *nodemask
= node_to_cpumask(i
);
7566 cpumask_clear(covered
);
7568 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7569 if (cpumask_empty(nodemask
)) {
7570 sched_group_nodes
[i
] = NULL
;
7574 sched_domain_node_span(i
, domainspan
);
7575 cpumask_and(domainspan
, domainspan
, cpu_map
);
7577 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
7580 printk(KERN_WARNING
"Can not alloc domain group for "
7584 sched_group_nodes
[i
] = sg
;
7585 for_each_cpu(j
, nodemask
) {
7586 struct sched_domain
*sd
;
7588 sd
= &per_cpu(node_domains
, j
);
7591 sg
->__cpu_power
= 0;
7592 cpumask_copy(sched_group_cpus(sg
), nodemask
);
7594 cpumask_or(covered
, covered
, nodemask
);
7597 for (j
= 0; j
< nr_node_ids
; j
++) {
7598 int n
= (i
+ j
) % nr_node_ids
;
7599 /* FIXME: Use cpumask_of_node */
7600 node_to_cpumask_ptr(pnodemask
, n
);
7602 cpumask_complement(notcovered
, covered
);
7603 cpumask_and(tmpmask
, notcovered
, cpu_map
);
7604 cpumask_and(tmpmask
, tmpmask
, domainspan
);
7605 if (cpumask_empty(tmpmask
))
7608 cpumask_and(tmpmask
, tmpmask
, pnodemask
);
7609 if (cpumask_empty(tmpmask
))
7612 sg
= kmalloc_node(sizeof(struct sched_group
) +
7617 "Can not alloc domain group for node %d\n", j
);
7620 sg
->__cpu_power
= 0;
7621 cpumask_copy(sched_group_cpus(sg
), tmpmask
);
7622 sg
->next
= prev
->next
;
7623 cpumask_or(covered
, covered
, tmpmask
);
7630 /* Calculate CPU power for physical packages and nodes */
7631 #ifdef CONFIG_SCHED_SMT
7632 for_each_cpu(i
, cpu_map
) {
7633 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
).sd
;
7635 init_sched_groups_power(i
, sd
);
7638 #ifdef CONFIG_SCHED_MC
7639 for_each_cpu(i
, cpu_map
) {
7640 struct sched_domain
*sd
= &per_cpu(core_domains
, i
).sd
;
7642 init_sched_groups_power(i
, sd
);
7646 for_each_cpu(i
, cpu_map
) {
7647 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
).sd
;
7649 init_sched_groups_power(i
, sd
);
7653 for (i
= 0; i
< nr_node_ids
; i
++)
7654 init_numa_sched_groups_power(sched_group_nodes
[i
]);
7657 struct sched_group
*sg
;
7659 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
7661 init_numa_sched_groups_power(sg
);
7665 /* Attach the domains */
7666 for_each_cpu(i
, cpu_map
) {
7667 struct sched_domain
*sd
;
7668 #ifdef CONFIG_SCHED_SMT
7669 sd
= &per_cpu(cpu_domains
, i
).sd
;
7670 #elif defined(CONFIG_SCHED_MC)
7671 sd
= &per_cpu(core_domains
, i
).sd
;
7673 sd
= &per_cpu(phys_domains
, i
).sd
;
7675 cpu_attach_domain(sd
, rd
, i
);
7681 free_cpumask_var(tmpmask
);
7683 free_cpumask_var(send_covered
);
7685 free_cpumask_var(this_core_map
);
7686 free_this_sibling_map
:
7687 free_cpumask_var(this_sibling_map
);
7689 free_cpumask_var(nodemask
);
7692 free_cpumask_var(notcovered
);
7694 free_cpumask_var(covered
);
7696 free_cpumask_var(domainspan
);
7703 kfree(sched_group_nodes
);
7709 free_sched_groups(cpu_map
, tmpmask
);
7710 free_rootdomain(rd
);
7715 static int build_sched_domains(const struct cpumask
*cpu_map
)
7717 return __build_sched_domains(cpu_map
, NULL
);
7720 static struct cpumask
*doms_cur
; /* current sched domains */
7721 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7722 static struct sched_domain_attr
*dattr_cur
;
7723 /* attribues of custom domains in 'doms_cur' */
7726 * Special case: If a kmalloc of a doms_cur partition (array of
7727 * cpumask) fails, then fallback to a single sched domain,
7728 * as determined by the single cpumask fallback_doms.
7730 static cpumask_var_t fallback_doms
;
7733 * arch_update_cpu_topology lets virtualized architectures update the
7734 * cpu core maps. It is supposed to return 1 if the topology changed
7735 * or 0 if it stayed the same.
7737 int __attribute__((weak
)) arch_update_cpu_topology(void)
7743 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7744 * For now this just excludes isolated cpus, but could be used to
7745 * exclude other special cases in the future.
7747 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
7751 arch_update_cpu_topology();
7753 doms_cur
= kmalloc(cpumask_size(), GFP_KERNEL
);
7755 doms_cur
= fallback_doms
;
7756 cpumask_andnot(doms_cur
, cpu_map
, cpu_isolated_map
);
7758 err
= build_sched_domains(doms_cur
);
7759 register_sched_domain_sysctl();
7764 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
7765 struct cpumask
*tmpmask
)
7767 free_sched_groups(cpu_map
, tmpmask
);
7771 * Detach sched domains from a group of cpus specified in cpu_map
7772 * These cpus will now be attached to the NULL domain
7774 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7776 /* Save because hotplug lock held. */
7777 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
7780 for_each_cpu(i
, cpu_map
)
7781 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7782 synchronize_sched();
7783 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
7786 /* handle null as "default" */
7787 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7788 struct sched_domain_attr
*new, int idx_new
)
7790 struct sched_domain_attr tmp
;
7797 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7798 new ? (new + idx_new
) : &tmp
,
7799 sizeof(struct sched_domain_attr
));
7803 * Partition sched domains as specified by the 'ndoms_new'
7804 * cpumasks in the array doms_new[] of cpumasks. This compares
7805 * doms_new[] to the current sched domain partitioning, doms_cur[].
7806 * It destroys each deleted domain and builds each new domain.
7808 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
7809 * The masks don't intersect (don't overlap.) We should setup one
7810 * sched domain for each mask. CPUs not in any of the cpumasks will
7811 * not be load balanced. If the same cpumask appears both in the
7812 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7815 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7816 * ownership of it and will kfree it when done with it. If the caller
7817 * failed the kmalloc call, then it can pass in doms_new == NULL &&
7818 * ndoms_new == 1, and partition_sched_domains() will fallback to
7819 * the single partition 'fallback_doms', it also forces the domains
7822 * If doms_new == NULL it will be replaced with cpu_online_mask.
7823 * ndoms_new == 0 is a special case for destroying existing domains,
7824 * and it will not create the default domain.
7826 * Call with hotplug lock held
7828 /* FIXME: Change to struct cpumask *doms_new[] */
7829 void partition_sched_domains(int ndoms_new
, struct cpumask
*doms_new
,
7830 struct sched_domain_attr
*dattr_new
)
7835 mutex_lock(&sched_domains_mutex
);
7837 /* always unregister in case we don't destroy any domains */
7838 unregister_sched_domain_sysctl();
7840 /* Let architecture update cpu core mappings. */
7841 new_topology
= arch_update_cpu_topology();
7843 n
= doms_new
? ndoms_new
: 0;
7845 /* Destroy deleted domains */
7846 for (i
= 0; i
< ndoms_cur
; i
++) {
7847 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7848 if (cpumask_equal(&doms_cur
[i
], &doms_new
[j
])
7849 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7852 /* no match - a current sched domain not in new doms_new[] */
7853 detach_destroy_domains(doms_cur
+ i
);
7858 if (doms_new
== NULL
) {
7860 doms_new
= fallback_doms
;
7861 cpumask_andnot(&doms_new
[0], cpu_online_mask
, cpu_isolated_map
);
7862 WARN_ON_ONCE(dattr_new
);
7865 /* Build new domains */
7866 for (i
= 0; i
< ndoms_new
; i
++) {
7867 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
7868 if (cpumask_equal(&doms_new
[i
], &doms_cur
[j
])
7869 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7872 /* no match - add a new doms_new */
7873 __build_sched_domains(doms_new
+ i
,
7874 dattr_new
? dattr_new
+ i
: NULL
);
7879 /* Remember the new sched domains */
7880 if (doms_cur
!= fallback_doms
)
7882 kfree(dattr_cur
); /* kfree(NULL) is safe */
7883 doms_cur
= doms_new
;
7884 dattr_cur
= dattr_new
;
7885 ndoms_cur
= ndoms_new
;
7887 register_sched_domain_sysctl();
7889 mutex_unlock(&sched_domains_mutex
);
7892 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7893 int arch_reinit_sched_domains(void)
7897 /* Destroy domains first to force the rebuild */
7898 partition_sched_domains(0, NULL
, NULL
);
7900 rebuild_sched_domains();
7906 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7910 if (buf
[0] != '0' && buf
[0] != '1')
7914 sched_smt_power_savings
= (buf
[0] == '1');
7916 sched_mc_power_savings
= (buf
[0] == '1');
7918 ret
= arch_reinit_sched_domains();
7920 return ret
? ret
: count
;
7923 #ifdef CONFIG_SCHED_MC
7924 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7927 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7929 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7930 const char *buf
, size_t count
)
7932 return sched_power_savings_store(buf
, count
, 0);
7934 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7935 sched_mc_power_savings_show
,
7936 sched_mc_power_savings_store
);
7939 #ifdef CONFIG_SCHED_SMT
7940 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7943 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7945 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7946 const char *buf
, size_t count
)
7948 return sched_power_savings_store(buf
, count
, 1);
7950 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7951 sched_smt_power_savings_show
,
7952 sched_smt_power_savings_store
);
7955 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7959 #ifdef CONFIG_SCHED_SMT
7961 err
= sysfs_create_file(&cls
->kset
.kobj
,
7962 &attr_sched_smt_power_savings
.attr
);
7964 #ifdef CONFIG_SCHED_MC
7965 if (!err
&& mc_capable())
7966 err
= sysfs_create_file(&cls
->kset
.kobj
,
7967 &attr_sched_mc_power_savings
.attr
);
7971 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7973 #ifndef CONFIG_CPUSETS
7975 * Add online and remove offline CPUs from the scheduler domains.
7976 * When cpusets are enabled they take over this function.
7978 static int update_sched_domains(struct notifier_block
*nfb
,
7979 unsigned long action
, void *hcpu
)
7983 case CPU_ONLINE_FROZEN
:
7985 case CPU_DEAD_FROZEN
:
7986 partition_sched_domains(1, NULL
, NULL
);
7995 static int update_runtime(struct notifier_block
*nfb
,
7996 unsigned long action
, void *hcpu
)
7998 int cpu
= (int)(long)hcpu
;
8001 case CPU_DOWN_PREPARE
:
8002 case CPU_DOWN_PREPARE_FROZEN
:
8003 disable_runtime(cpu_rq(cpu
));
8006 case CPU_DOWN_FAILED
:
8007 case CPU_DOWN_FAILED_FROZEN
:
8009 case CPU_ONLINE_FROZEN
:
8010 enable_runtime(cpu_rq(cpu
));
8018 void __init
sched_init_smp(void)
8020 cpumask_var_t non_isolated_cpus
;
8022 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
8024 #if defined(CONFIG_NUMA)
8025 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
8027 BUG_ON(sched_group_nodes_bycpu
== NULL
);
8030 mutex_lock(&sched_domains_mutex
);
8031 arch_init_sched_domains(cpu_online_mask
);
8032 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
8033 if (cpumask_empty(non_isolated_cpus
))
8034 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
8035 mutex_unlock(&sched_domains_mutex
);
8038 #ifndef CONFIG_CPUSETS
8039 /* XXX: Theoretical race here - CPU may be hotplugged now */
8040 hotcpu_notifier(update_sched_domains
, 0);
8043 /* RT runtime code needs to handle some hotplug events */
8044 hotcpu_notifier(update_runtime
, 0);
8048 /* Move init over to a non-isolated CPU */
8049 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
8051 sched_init_granularity();
8052 free_cpumask_var(non_isolated_cpus
);
8054 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
8055 init_sched_rt_class();
8058 void __init
sched_init_smp(void)
8060 sched_init_granularity();
8062 #endif /* CONFIG_SMP */
8064 int in_sched_functions(unsigned long addr
)
8066 return in_lock_functions(addr
) ||
8067 (addr
>= (unsigned long)__sched_text_start
8068 && addr
< (unsigned long)__sched_text_end
);
8071 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
8073 cfs_rq
->tasks_timeline
= RB_ROOT
;
8074 INIT_LIST_HEAD(&cfs_rq
->tasks
);
8075 #ifdef CONFIG_FAIR_GROUP_SCHED
8078 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
8081 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
8083 struct rt_prio_array
*array
;
8086 array
= &rt_rq
->active
;
8087 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
8088 INIT_LIST_HEAD(array
->queue
+ i
);
8089 __clear_bit(i
, array
->bitmap
);
8091 /* delimiter for bitsearch: */
8092 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
8094 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8095 rt_rq
->highest_prio
= MAX_RT_PRIO
;
8098 rt_rq
->rt_nr_migratory
= 0;
8099 rt_rq
->overloaded
= 0;
8103 rt_rq
->rt_throttled
= 0;
8104 rt_rq
->rt_runtime
= 0;
8105 spin_lock_init(&rt_rq
->rt_runtime_lock
);
8107 #ifdef CONFIG_RT_GROUP_SCHED
8108 rt_rq
->rt_nr_boosted
= 0;
8113 #ifdef CONFIG_FAIR_GROUP_SCHED
8114 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8115 struct sched_entity
*se
, int cpu
, int add
,
8116 struct sched_entity
*parent
)
8118 struct rq
*rq
= cpu_rq(cpu
);
8119 tg
->cfs_rq
[cpu
] = cfs_rq
;
8120 init_cfs_rq(cfs_rq
, rq
);
8123 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
8126 /* se could be NULL for init_task_group */
8131 se
->cfs_rq
= &rq
->cfs
;
8133 se
->cfs_rq
= parent
->my_q
;
8136 se
->load
.weight
= tg
->shares
;
8137 se
->load
.inv_weight
= 0;
8138 se
->parent
= parent
;
8142 #ifdef CONFIG_RT_GROUP_SCHED
8143 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
8144 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
8145 struct sched_rt_entity
*parent
)
8147 struct rq
*rq
= cpu_rq(cpu
);
8149 tg
->rt_rq
[cpu
] = rt_rq
;
8150 init_rt_rq(rt_rq
, rq
);
8152 rt_rq
->rt_se
= rt_se
;
8153 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8155 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
8157 tg
->rt_se
[cpu
] = rt_se
;
8162 rt_se
->rt_rq
= &rq
->rt
;
8164 rt_se
->rt_rq
= parent
->my_q
;
8166 rt_se
->my_q
= rt_rq
;
8167 rt_se
->parent
= parent
;
8168 INIT_LIST_HEAD(&rt_se
->run_list
);
8172 void __init
sched_init(void)
8175 unsigned long alloc_size
= 0, ptr
;
8177 #ifdef CONFIG_FAIR_GROUP_SCHED
8178 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8180 #ifdef CONFIG_RT_GROUP_SCHED
8181 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8183 #ifdef CONFIG_USER_SCHED
8187 * As sched_init() is called before page_alloc is setup,
8188 * we use alloc_bootmem().
8191 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
8193 #ifdef CONFIG_FAIR_GROUP_SCHED
8194 init_task_group
.se
= (struct sched_entity
**)ptr
;
8195 ptr
+= nr_cpu_ids
* sizeof(void **);
8197 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8198 ptr
+= nr_cpu_ids
* sizeof(void **);
8200 #ifdef CONFIG_USER_SCHED
8201 root_task_group
.se
= (struct sched_entity
**)ptr
;
8202 ptr
+= nr_cpu_ids
* sizeof(void **);
8204 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8205 ptr
+= nr_cpu_ids
* sizeof(void **);
8206 #endif /* CONFIG_USER_SCHED */
8207 #endif /* CONFIG_FAIR_GROUP_SCHED */
8208 #ifdef CONFIG_RT_GROUP_SCHED
8209 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8210 ptr
+= nr_cpu_ids
* sizeof(void **);
8212 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8213 ptr
+= nr_cpu_ids
* sizeof(void **);
8215 #ifdef CONFIG_USER_SCHED
8216 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8217 ptr
+= nr_cpu_ids
* sizeof(void **);
8219 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8220 ptr
+= nr_cpu_ids
* sizeof(void **);
8221 #endif /* CONFIG_USER_SCHED */
8222 #endif /* CONFIG_RT_GROUP_SCHED */
8226 init_defrootdomain();
8229 init_rt_bandwidth(&def_rt_bandwidth
,
8230 global_rt_period(), global_rt_runtime());
8232 #ifdef CONFIG_RT_GROUP_SCHED
8233 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
8234 global_rt_period(), global_rt_runtime());
8235 #ifdef CONFIG_USER_SCHED
8236 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8237 global_rt_period(), RUNTIME_INF
);
8238 #endif /* CONFIG_USER_SCHED */
8239 #endif /* CONFIG_RT_GROUP_SCHED */
8241 #ifdef CONFIG_GROUP_SCHED
8242 list_add(&init_task_group
.list
, &task_groups
);
8243 INIT_LIST_HEAD(&init_task_group
.children
);
8245 #ifdef CONFIG_USER_SCHED
8246 INIT_LIST_HEAD(&root_task_group
.children
);
8247 init_task_group
.parent
= &root_task_group
;
8248 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
8249 #endif /* CONFIG_USER_SCHED */
8250 #endif /* CONFIG_GROUP_SCHED */
8252 for_each_possible_cpu(i
) {
8256 spin_lock_init(&rq
->lock
);
8258 init_cfs_rq(&rq
->cfs
, rq
);
8259 init_rt_rq(&rq
->rt
, rq
);
8260 #ifdef CONFIG_FAIR_GROUP_SCHED
8261 init_task_group
.shares
= init_task_group_load
;
8262 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8263 #ifdef CONFIG_CGROUP_SCHED
8265 * How much cpu bandwidth does init_task_group get?
8267 * In case of task-groups formed thr' the cgroup filesystem, it
8268 * gets 100% of the cpu resources in the system. This overall
8269 * system cpu resource is divided among the tasks of
8270 * init_task_group and its child task-groups in a fair manner,
8271 * based on each entity's (task or task-group's) weight
8272 * (se->load.weight).
8274 * In other words, if init_task_group has 10 tasks of weight
8275 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8276 * then A0's share of the cpu resource is:
8278 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8280 * We achieve this by letting init_task_group's tasks sit
8281 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8283 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
8284 #elif defined CONFIG_USER_SCHED
8285 root_task_group
.shares
= NICE_0_LOAD
;
8286 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
8288 * In case of task-groups formed thr' the user id of tasks,
8289 * init_task_group represents tasks belonging to root user.
8290 * Hence it forms a sibling of all subsequent groups formed.
8291 * In this case, init_task_group gets only a fraction of overall
8292 * system cpu resource, based on the weight assigned to root
8293 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8294 * by letting tasks of init_task_group sit in a separate cfs_rq
8295 * (init_cfs_rq) and having one entity represent this group of
8296 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8298 init_tg_cfs_entry(&init_task_group
,
8299 &per_cpu(init_cfs_rq
, i
),
8300 &per_cpu(init_sched_entity
, i
), i
, 1,
8301 root_task_group
.se
[i
]);
8304 #endif /* CONFIG_FAIR_GROUP_SCHED */
8306 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8307 #ifdef CONFIG_RT_GROUP_SCHED
8308 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
8309 #ifdef CONFIG_CGROUP_SCHED
8310 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
8311 #elif defined CONFIG_USER_SCHED
8312 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
8313 init_tg_rt_entry(&init_task_group
,
8314 &per_cpu(init_rt_rq
, i
),
8315 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
8316 root_task_group
.rt_se
[i
]);
8320 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
8321 rq
->cpu_load
[j
] = 0;
8325 rq
->active_balance
= 0;
8326 rq
->next_balance
= jiffies
;
8330 rq
->migration_thread
= NULL
;
8331 INIT_LIST_HEAD(&rq
->migration_queue
);
8332 rq_attach_root(rq
, &def_root_domain
);
8335 atomic_set(&rq
->nr_iowait
, 0);
8338 set_load_weight(&init_task
);
8340 #ifdef CONFIG_PREEMPT_NOTIFIERS
8341 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
8345 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8348 #ifdef CONFIG_RT_MUTEXES
8349 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
8353 * The boot idle thread does lazy MMU switching as well:
8355 atomic_inc(&init_mm
.mm_count
);
8356 enter_lazy_tlb(&init_mm
, current
);
8359 * Make us the idle thread. Technically, schedule() should not be
8360 * called from this thread, however somewhere below it might be,
8361 * but because we are the idle thread, we just pick up running again
8362 * when this runqueue becomes "idle".
8364 init_idle(current
, smp_processor_id());
8366 * During early bootup we pretend to be a normal task:
8368 current
->sched_class
= &fair_sched_class
;
8370 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8371 alloc_bootmem_cpumask_var(&nohz_cpu_mask
);
8374 alloc_bootmem_cpumask_var(&nohz
.cpu_mask
);
8376 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
8379 scheduler_running
= 1;
8382 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8383 void __might_sleep(char *file
, int line
)
8386 static unsigned long prev_jiffy
; /* ratelimiting */
8388 if ((!in_atomic() && !irqs_disabled()) ||
8389 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
8391 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8393 prev_jiffy
= jiffies
;
8396 "BUG: sleeping function called from invalid context at %s:%d\n",
8399 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8400 in_atomic(), irqs_disabled(),
8401 current
->pid
, current
->comm
);
8403 debug_show_held_locks(current
);
8404 if (irqs_disabled())
8405 print_irqtrace_events(current
);
8409 EXPORT_SYMBOL(__might_sleep
);
8412 #ifdef CONFIG_MAGIC_SYSRQ
8413 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8417 update_rq_clock(rq
);
8418 on_rq
= p
->se
.on_rq
;
8420 deactivate_task(rq
, p
, 0);
8421 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8423 activate_task(rq
, p
, 0);
8424 resched_task(rq
->curr
);
8428 void normalize_rt_tasks(void)
8430 struct task_struct
*g
, *p
;
8431 unsigned long flags
;
8434 read_lock_irqsave(&tasklist_lock
, flags
);
8435 do_each_thread(g
, p
) {
8437 * Only normalize user tasks:
8442 p
->se
.exec_start
= 0;
8443 #ifdef CONFIG_SCHEDSTATS
8444 p
->se
.wait_start
= 0;
8445 p
->se
.sleep_start
= 0;
8446 p
->se
.block_start
= 0;
8451 * Renice negative nice level userspace
8454 if (TASK_NICE(p
) < 0 && p
->mm
)
8455 set_user_nice(p
, 0);
8459 spin_lock(&p
->pi_lock
);
8460 rq
= __task_rq_lock(p
);
8462 normalize_task(rq
, p
);
8464 __task_rq_unlock(rq
);
8465 spin_unlock(&p
->pi_lock
);
8466 } while_each_thread(g
, p
);
8468 read_unlock_irqrestore(&tasklist_lock
, flags
);
8471 #endif /* CONFIG_MAGIC_SYSRQ */
8475 * These functions are only useful for the IA64 MCA handling.
8477 * They can only be called when the whole system has been
8478 * stopped - every CPU needs to be quiescent, and no scheduling
8479 * activity can take place. Using them for anything else would
8480 * be a serious bug, and as a result, they aren't even visible
8481 * under any other configuration.
8485 * curr_task - return the current task for a given cpu.
8486 * @cpu: the processor in question.
8488 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8490 struct task_struct
*curr_task(int cpu
)
8492 return cpu_curr(cpu
);
8496 * set_curr_task - set the current task for a given cpu.
8497 * @cpu: the processor in question.
8498 * @p: the task pointer to set.
8500 * Description: This function must only be used when non-maskable interrupts
8501 * are serviced on a separate stack. It allows the architecture to switch the
8502 * notion of the current task on a cpu in a non-blocking manner. This function
8503 * must be called with all CPU's synchronized, and interrupts disabled, the
8504 * and caller must save the original value of the current task (see
8505 * curr_task() above) and restore that value before reenabling interrupts and
8506 * re-starting the system.
8508 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8510 void set_curr_task(int cpu
, struct task_struct
*p
)
8517 #ifdef CONFIG_FAIR_GROUP_SCHED
8518 static void free_fair_sched_group(struct task_group
*tg
)
8522 for_each_possible_cpu(i
) {
8524 kfree(tg
->cfs_rq
[i
]);
8534 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8536 struct cfs_rq
*cfs_rq
;
8537 struct sched_entity
*se
;
8541 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8544 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8548 tg
->shares
= NICE_0_LOAD
;
8550 for_each_possible_cpu(i
) {
8553 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
8554 GFP_KERNEL
, cpu_to_node(i
));
8558 se
= kzalloc_node(sizeof(struct sched_entity
),
8559 GFP_KERNEL
, cpu_to_node(i
));
8563 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
8572 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8574 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8575 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8578 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8580 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8582 #else /* !CONFG_FAIR_GROUP_SCHED */
8583 static inline void free_fair_sched_group(struct task_group
*tg
)
8588 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8593 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8597 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8600 #endif /* CONFIG_FAIR_GROUP_SCHED */
8602 #ifdef CONFIG_RT_GROUP_SCHED
8603 static void free_rt_sched_group(struct task_group
*tg
)
8607 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8609 for_each_possible_cpu(i
) {
8611 kfree(tg
->rt_rq
[i
]);
8613 kfree(tg
->rt_se
[i
]);
8621 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8623 struct rt_rq
*rt_rq
;
8624 struct sched_rt_entity
*rt_se
;
8628 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8631 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8635 init_rt_bandwidth(&tg
->rt_bandwidth
,
8636 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8638 for_each_possible_cpu(i
) {
8641 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
8642 GFP_KERNEL
, cpu_to_node(i
));
8646 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
8647 GFP_KERNEL
, cpu_to_node(i
));
8651 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
8660 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8662 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8663 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8666 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8668 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8670 #else /* !CONFIG_RT_GROUP_SCHED */
8671 static inline void free_rt_sched_group(struct task_group
*tg
)
8676 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8681 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8685 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8688 #endif /* CONFIG_RT_GROUP_SCHED */
8690 #ifdef CONFIG_GROUP_SCHED
8691 static void free_sched_group(struct task_group
*tg
)
8693 free_fair_sched_group(tg
);
8694 free_rt_sched_group(tg
);
8698 /* allocate runqueue etc for a new task group */
8699 struct task_group
*sched_create_group(struct task_group
*parent
)
8701 struct task_group
*tg
;
8702 unsigned long flags
;
8705 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8707 return ERR_PTR(-ENOMEM
);
8709 if (!alloc_fair_sched_group(tg
, parent
))
8712 if (!alloc_rt_sched_group(tg
, parent
))
8715 spin_lock_irqsave(&task_group_lock
, flags
);
8716 for_each_possible_cpu(i
) {
8717 register_fair_sched_group(tg
, i
);
8718 register_rt_sched_group(tg
, i
);
8720 list_add_rcu(&tg
->list
, &task_groups
);
8722 WARN_ON(!parent
); /* root should already exist */
8724 tg
->parent
= parent
;
8725 INIT_LIST_HEAD(&tg
->children
);
8726 list_add_rcu(&tg
->siblings
, &parent
->children
);
8727 spin_unlock_irqrestore(&task_group_lock
, flags
);
8732 free_sched_group(tg
);
8733 return ERR_PTR(-ENOMEM
);
8736 /* rcu callback to free various structures associated with a task group */
8737 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8739 /* now it should be safe to free those cfs_rqs */
8740 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8743 /* Destroy runqueue etc associated with a task group */
8744 void sched_destroy_group(struct task_group
*tg
)
8746 unsigned long flags
;
8749 spin_lock_irqsave(&task_group_lock
, flags
);
8750 for_each_possible_cpu(i
) {
8751 unregister_fair_sched_group(tg
, i
);
8752 unregister_rt_sched_group(tg
, i
);
8754 list_del_rcu(&tg
->list
);
8755 list_del_rcu(&tg
->siblings
);
8756 spin_unlock_irqrestore(&task_group_lock
, flags
);
8758 /* wait for possible concurrent references to cfs_rqs complete */
8759 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8762 /* change task's runqueue when it moves between groups.
8763 * The caller of this function should have put the task in its new group
8764 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8765 * reflect its new group.
8767 void sched_move_task(struct task_struct
*tsk
)
8770 unsigned long flags
;
8773 rq
= task_rq_lock(tsk
, &flags
);
8775 update_rq_clock(rq
);
8777 running
= task_current(rq
, tsk
);
8778 on_rq
= tsk
->se
.on_rq
;
8781 dequeue_task(rq
, tsk
, 0);
8782 if (unlikely(running
))
8783 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8785 set_task_rq(tsk
, task_cpu(tsk
));
8787 #ifdef CONFIG_FAIR_GROUP_SCHED
8788 if (tsk
->sched_class
->moved_group
)
8789 tsk
->sched_class
->moved_group(tsk
);
8792 if (unlikely(running
))
8793 tsk
->sched_class
->set_curr_task(rq
);
8795 enqueue_task(rq
, tsk
, 0);
8797 task_rq_unlock(rq
, &flags
);
8799 #endif /* CONFIG_GROUP_SCHED */
8801 #ifdef CONFIG_FAIR_GROUP_SCHED
8802 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8804 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8809 dequeue_entity(cfs_rq
, se
, 0);
8811 se
->load
.weight
= shares
;
8812 se
->load
.inv_weight
= 0;
8815 enqueue_entity(cfs_rq
, se
, 0);
8818 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8820 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8821 struct rq
*rq
= cfs_rq
->rq
;
8822 unsigned long flags
;
8824 spin_lock_irqsave(&rq
->lock
, flags
);
8825 __set_se_shares(se
, shares
);
8826 spin_unlock_irqrestore(&rq
->lock
, flags
);
8829 static DEFINE_MUTEX(shares_mutex
);
8831 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8834 unsigned long flags
;
8837 * We can't change the weight of the root cgroup.
8842 if (shares
< MIN_SHARES
)
8843 shares
= MIN_SHARES
;
8844 else if (shares
> MAX_SHARES
)
8845 shares
= MAX_SHARES
;
8847 mutex_lock(&shares_mutex
);
8848 if (tg
->shares
== shares
)
8851 spin_lock_irqsave(&task_group_lock
, flags
);
8852 for_each_possible_cpu(i
)
8853 unregister_fair_sched_group(tg
, i
);
8854 list_del_rcu(&tg
->siblings
);
8855 spin_unlock_irqrestore(&task_group_lock
, flags
);
8857 /* wait for any ongoing reference to this group to finish */
8858 synchronize_sched();
8861 * Now we are free to modify the group's share on each cpu
8862 * w/o tripping rebalance_share or load_balance_fair.
8864 tg
->shares
= shares
;
8865 for_each_possible_cpu(i
) {
8869 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8870 set_se_shares(tg
->se
[i
], shares
);
8874 * Enable load balance activity on this group, by inserting it back on
8875 * each cpu's rq->leaf_cfs_rq_list.
8877 spin_lock_irqsave(&task_group_lock
, flags
);
8878 for_each_possible_cpu(i
)
8879 register_fair_sched_group(tg
, i
);
8880 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8881 spin_unlock_irqrestore(&task_group_lock
, flags
);
8883 mutex_unlock(&shares_mutex
);
8887 unsigned long sched_group_shares(struct task_group
*tg
)
8893 #ifdef CONFIG_RT_GROUP_SCHED
8895 * Ensure that the real time constraints are schedulable.
8897 static DEFINE_MUTEX(rt_constraints_mutex
);
8899 static unsigned long to_ratio(u64 period
, u64 runtime
)
8901 if (runtime
== RUNTIME_INF
)
8904 return div64_u64(runtime
<< 20, period
);
8907 /* Must be called with tasklist_lock held */
8908 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8910 struct task_struct
*g
, *p
;
8912 do_each_thread(g
, p
) {
8913 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8915 } while_each_thread(g
, p
);
8920 struct rt_schedulable_data
{
8921 struct task_group
*tg
;
8926 static int tg_schedulable(struct task_group
*tg
, void *data
)
8928 struct rt_schedulable_data
*d
= data
;
8929 struct task_group
*child
;
8930 unsigned long total
, sum
= 0;
8931 u64 period
, runtime
;
8933 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8934 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8937 period
= d
->rt_period
;
8938 runtime
= d
->rt_runtime
;
8942 * Cannot have more runtime than the period.
8944 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8948 * Ensure we don't starve existing RT tasks.
8950 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8953 total
= to_ratio(period
, runtime
);
8956 * Nobody can have more than the global setting allows.
8958 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8962 * The sum of our children's runtime should not exceed our own.
8964 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8965 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8966 runtime
= child
->rt_bandwidth
.rt_runtime
;
8968 if (child
== d
->tg
) {
8969 period
= d
->rt_period
;
8970 runtime
= d
->rt_runtime
;
8973 sum
+= to_ratio(period
, runtime
);
8982 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8984 struct rt_schedulable_data data
= {
8986 .rt_period
= period
,
8987 .rt_runtime
= runtime
,
8990 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8993 static int tg_set_bandwidth(struct task_group
*tg
,
8994 u64 rt_period
, u64 rt_runtime
)
8998 mutex_lock(&rt_constraints_mutex
);
8999 read_lock(&tasklist_lock
);
9000 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
9004 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
9005 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
9006 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
9008 for_each_possible_cpu(i
) {
9009 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
9011 spin_lock(&rt_rq
->rt_runtime_lock
);
9012 rt_rq
->rt_runtime
= rt_runtime
;
9013 spin_unlock(&rt_rq
->rt_runtime_lock
);
9015 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
9017 read_unlock(&tasklist_lock
);
9018 mutex_unlock(&rt_constraints_mutex
);
9023 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
9025 u64 rt_runtime
, rt_period
;
9027 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9028 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
9029 if (rt_runtime_us
< 0)
9030 rt_runtime
= RUNTIME_INF
;
9032 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9035 long sched_group_rt_runtime(struct task_group
*tg
)
9039 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
9042 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
9043 do_div(rt_runtime_us
, NSEC_PER_USEC
);
9044 return rt_runtime_us
;
9047 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
9049 u64 rt_runtime
, rt_period
;
9051 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
9052 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9057 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9060 long sched_group_rt_period(struct task_group
*tg
)
9064 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9065 do_div(rt_period_us
, NSEC_PER_USEC
);
9066 return rt_period_us
;
9069 static int sched_rt_global_constraints(void)
9071 u64 runtime
, period
;
9074 if (sysctl_sched_rt_period
<= 0)
9077 runtime
= global_rt_runtime();
9078 period
= global_rt_period();
9081 * Sanity check on the sysctl variables.
9083 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9086 mutex_lock(&rt_constraints_mutex
);
9087 read_lock(&tasklist_lock
);
9088 ret
= __rt_schedulable(NULL
, 0, 0);
9089 read_unlock(&tasklist_lock
);
9090 mutex_unlock(&rt_constraints_mutex
);
9094 #else /* !CONFIG_RT_GROUP_SCHED */
9095 static int sched_rt_global_constraints(void)
9097 unsigned long flags
;
9100 if (sysctl_sched_rt_period
<= 0)
9103 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9104 for_each_possible_cpu(i
) {
9105 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
9107 spin_lock(&rt_rq
->rt_runtime_lock
);
9108 rt_rq
->rt_runtime
= global_rt_runtime();
9109 spin_unlock(&rt_rq
->rt_runtime_lock
);
9111 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9115 #endif /* CONFIG_RT_GROUP_SCHED */
9117 int sched_rt_handler(struct ctl_table
*table
, int write
,
9118 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
9122 int old_period
, old_runtime
;
9123 static DEFINE_MUTEX(mutex
);
9126 old_period
= sysctl_sched_rt_period
;
9127 old_runtime
= sysctl_sched_rt_runtime
;
9129 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
9131 if (!ret
&& write
) {
9132 ret
= sched_rt_global_constraints();
9134 sysctl_sched_rt_period
= old_period
;
9135 sysctl_sched_rt_runtime
= old_runtime
;
9137 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
9138 def_rt_bandwidth
.rt_period
=
9139 ns_to_ktime(global_rt_period());
9142 mutex_unlock(&mutex
);
9147 #ifdef CONFIG_CGROUP_SCHED
9149 /* return corresponding task_group object of a cgroup */
9150 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
9152 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
9153 struct task_group
, css
);
9156 static struct cgroup_subsys_state
*
9157 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9159 struct task_group
*tg
, *parent
;
9161 if (!cgrp
->parent
) {
9162 /* This is early initialization for the top cgroup */
9163 return &init_task_group
.css
;
9166 parent
= cgroup_tg(cgrp
->parent
);
9167 tg
= sched_create_group(parent
);
9169 return ERR_PTR(-ENOMEM
);
9175 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9177 struct task_group
*tg
= cgroup_tg(cgrp
);
9179 sched_destroy_group(tg
);
9183 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9184 struct task_struct
*tsk
)
9186 #ifdef CONFIG_RT_GROUP_SCHED
9187 /* Don't accept realtime tasks when there is no way for them to run */
9188 if (rt_task(tsk
) && cgroup_tg(cgrp
)->rt_bandwidth
.rt_runtime
== 0)
9191 /* We don't support RT-tasks being in separate groups */
9192 if (tsk
->sched_class
!= &fair_sched_class
)
9200 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9201 struct cgroup
*old_cont
, struct task_struct
*tsk
)
9203 sched_move_task(tsk
);
9206 #ifdef CONFIG_FAIR_GROUP_SCHED
9207 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9210 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
9213 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9215 struct task_group
*tg
= cgroup_tg(cgrp
);
9217 return (u64
) tg
->shares
;
9219 #endif /* CONFIG_FAIR_GROUP_SCHED */
9221 #ifdef CONFIG_RT_GROUP_SCHED
9222 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9225 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9228 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9230 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9233 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9236 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9239 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9241 return sched_group_rt_period(cgroup_tg(cgrp
));
9243 #endif /* CONFIG_RT_GROUP_SCHED */
9245 static struct cftype cpu_files
[] = {
9246 #ifdef CONFIG_FAIR_GROUP_SCHED
9249 .read_u64
= cpu_shares_read_u64
,
9250 .write_u64
= cpu_shares_write_u64
,
9253 #ifdef CONFIG_RT_GROUP_SCHED
9255 .name
= "rt_runtime_us",
9256 .read_s64
= cpu_rt_runtime_read
,
9257 .write_s64
= cpu_rt_runtime_write
,
9260 .name
= "rt_period_us",
9261 .read_u64
= cpu_rt_period_read_uint
,
9262 .write_u64
= cpu_rt_period_write_uint
,
9267 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
9269 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
9272 struct cgroup_subsys cpu_cgroup_subsys
= {
9274 .create
= cpu_cgroup_create
,
9275 .destroy
= cpu_cgroup_destroy
,
9276 .can_attach
= cpu_cgroup_can_attach
,
9277 .attach
= cpu_cgroup_attach
,
9278 .populate
= cpu_cgroup_populate
,
9279 .subsys_id
= cpu_cgroup_subsys_id
,
9283 #endif /* CONFIG_CGROUP_SCHED */
9285 #ifdef CONFIG_CGROUP_CPUACCT
9288 * CPU accounting code for task groups.
9290 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9291 * (balbir@in.ibm.com).
9294 /* track cpu usage of a group of tasks and its child groups */
9296 struct cgroup_subsys_state css
;
9297 /* cpuusage holds pointer to a u64-type object on every cpu */
9299 struct cpuacct
*parent
;
9302 struct cgroup_subsys cpuacct_subsys
;
9304 /* return cpu accounting group corresponding to this container */
9305 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
9307 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
9308 struct cpuacct
, css
);
9311 /* return cpu accounting group to which this task belongs */
9312 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
9314 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
9315 struct cpuacct
, css
);
9318 /* create a new cpu accounting group */
9319 static struct cgroup_subsys_state
*cpuacct_create(
9320 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9322 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
9325 return ERR_PTR(-ENOMEM
);
9327 ca
->cpuusage
= alloc_percpu(u64
);
9328 if (!ca
->cpuusage
) {
9330 return ERR_PTR(-ENOMEM
);
9334 ca
->parent
= cgroup_ca(cgrp
->parent
);
9339 /* destroy an existing cpu accounting group */
9341 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9343 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9345 free_percpu(ca
->cpuusage
);
9349 /* return total cpu usage (in nanoseconds) of a group */
9350 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9352 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9353 u64 totalcpuusage
= 0;
9356 for_each_possible_cpu(i
) {
9357 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9360 * Take rq->lock to make 64-bit addition safe on 32-bit
9363 spin_lock_irq(&cpu_rq(i
)->lock
);
9364 totalcpuusage
+= *cpuusage
;
9365 spin_unlock_irq(&cpu_rq(i
)->lock
);
9368 return totalcpuusage
;
9371 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9374 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9383 for_each_possible_cpu(i
) {
9384 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9386 spin_lock_irq(&cpu_rq(i
)->lock
);
9388 spin_unlock_irq(&cpu_rq(i
)->lock
);
9394 static struct cftype files
[] = {
9397 .read_u64
= cpuusage_read
,
9398 .write_u64
= cpuusage_write
,
9402 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9404 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9408 * charge this task's execution time to its accounting group.
9410 * called with rq->lock held.
9412 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9417 if (!cpuacct_subsys
.active
)
9420 cpu
= task_cpu(tsk
);
9423 for (; ca
; ca
= ca
->parent
) {
9424 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, cpu
);
9425 *cpuusage
+= cputime
;
9429 struct cgroup_subsys cpuacct_subsys
= {
9431 .create
= cpuacct_create
,
9432 .destroy
= cpuacct_destroy
,
9433 .populate
= cpuacct_populate
,
9434 .subsys_id
= cpuacct_subsys_id
,
9436 #endif /* CONFIG_CGROUP_CPUACCT */