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/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/reciprocal_div.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/bootmem.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
81 * Convert user-nice values [ -20 ... 0 ... 19 ]
82 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
85 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
86 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
87 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
90 * 'User priority' is the nice value converted to something we
91 * can work with better when scaling various scheduler parameters,
92 * it's a [ 0 ... 39 ] range.
94 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
95 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
96 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
99 * Helpers for converting nanosecond timing to jiffy resolution
101 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
103 #define NICE_0_LOAD SCHED_LOAD_SCALE
104 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
107 * These are the 'tuning knobs' of the scheduler:
109 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
110 * Timeslices get refilled after they expire.
112 #define DEF_TIMESLICE (100 * HZ / 1000)
115 * single value that denotes runtime == period, ie unlimited time.
117 #define RUNTIME_INF ((u64)~0ULL)
121 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
122 * Since cpu_power is a 'constant', we can use a reciprocal divide.
124 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
126 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
130 * Each time a sched group cpu_power is changed,
131 * we must compute its reciprocal value
133 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
135 sg
->__cpu_power
+= val
;
136 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
140 static inline int rt_policy(int policy
)
142 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
147 static inline int task_has_rt_policy(struct task_struct
*p
)
149 return rt_policy(p
->policy
);
153 * This is the priority-queue data structure of the RT scheduling class:
155 struct rt_prio_array
{
156 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
157 struct list_head queue
[MAX_RT_PRIO
];
160 struct rt_bandwidth
{
161 /* nests inside the rq lock: */
162 spinlock_t rt_runtime_lock
;
165 struct hrtimer rt_period_timer
;
168 static struct rt_bandwidth def_rt_bandwidth
;
170 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
172 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
174 struct rt_bandwidth
*rt_b
=
175 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
181 now
= hrtimer_cb_get_time(timer
);
182 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
187 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
190 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
194 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
196 rt_b
->rt_period
= ns_to_ktime(period
);
197 rt_b
->rt_runtime
= runtime
;
199 spin_lock_init(&rt_b
->rt_runtime_lock
);
201 hrtimer_init(&rt_b
->rt_period_timer
,
202 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
203 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
204 rt_b
->rt_period_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_NO_SOFTIRQ
;
207 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
211 if (rt_b
->rt_runtime
== RUNTIME_INF
)
214 if (hrtimer_active(&rt_b
->rt_period_timer
))
217 spin_lock(&rt_b
->rt_runtime_lock
);
219 if (hrtimer_active(&rt_b
->rt_period_timer
))
222 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
223 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
224 hrtimer_start(&rt_b
->rt_period_timer
,
225 rt_b
->rt_period_timer
.expires
,
228 spin_unlock(&rt_b
->rt_runtime_lock
);
231 #ifdef CONFIG_RT_GROUP_SCHED
232 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
234 hrtimer_cancel(&rt_b
->rt_period_timer
);
239 * sched_domains_mutex serializes calls to arch_init_sched_domains,
240 * detach_destroy_domains and partition_sched_domains.
242 static DEFINE_MUTEX(sched_domains_mutex
);
244 #ifdef CONFIG_GROUP_SCHED
246 #include <linux/cgroup.h>
250 static LIST_HEAD(task_groups
);
252 /* task group related information */
254 #ifdef CONFIG_CGROUP_SCHED
255 struct cgroup_subsys_state css
;
258 #ifdef CONFIG_FAIR_GROUP_SCHED
259 /* schedulable entities of this group on each cpu */
260 struct sched_entity
**se
;
261 /* runqueue "owned" by this group on each cpu */
262 struct cfs_rq
**cfs_rq
;
263 unsigned long shares
;
266 #ifdef CONFIG_RT_GROUP_SCHED
267 struct sched_rt_entity
**rt_se
;
268 struct rt_rq
**rt_rq
;
270 struct rt_bandwidth rt_bandwidth
;
274 struct list_head list
;
276 struct task_group
*parent
;
277 struct list_head siblings
;
278 struct list_head children
;
281 #ifdef CONFIG_USER_SCHED
285 * Every UID task group (including init_task_group aka UID-0) will
286 * be a child to this group.
288 struct task_group root_task_group
;
290 #ifdef CONFIG_FAIR_GROUP_SCHED
291 /* Default task group's sched entity on each cpu */
292 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
293 /* Default task group's cfs_rq on each cpu */
294 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
295 #endif /* CONFIG_FAIR_GROUP_SCHED */
297 #ifdef CONFIG_RT_GROUP_SCHED
298 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
299 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
300 #endif /* CONFIG_RT_GROUP_SCHED */
301 #else /* !CONFIG_FAIR_GROUP_SCHED */
302 #define root_task_group init_task_group
303 #endif /* CONFIG_FAIR_GROUP_SCHED */
305 /* task_group_lock serializes add/remove of task groups and also changes to
306 * a task group's cpu shares.
308 static DEFINE_SPINLOCK(task_group_lock
);
310 #ifdef CONFIG_FAIR_GROUP_SCHED
311 #ifdef CONFIG_USER_SCHED
312 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
313 #else /* !CONFIG_USER_SCHED */
314 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
315 #endif /* CONFIG_USER_SCHED */
318 * A weight of 0 or 1 can cause arithmetics problems.
319 * A weight of a cfs_rq is the sum of weights of which entities
320 * are queued on this cfs_rq, so a weight of a entity should not be
321 * too large, so as the shares value of a task group.
322 * (The default weight is 1024 - so there's no practical
323 * limitation from this.)
326 #define MAX_SHARES (1UL << 18)
328 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
331 /* Default task group.
332 * Every task in system belong to this group at bootup.
334 struct task_group init_task_group
;
336 /* return group to which a task belongs */
337 static inline struct task_group
*task_group(struct task_struct
*p
)
339 struct task_group
*tg
;
341 #ifdef CONFIG_USER_SCHED
343 #elif defined(CONFIG_CGROUP_SCHED)
344 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
345 struct task_group
, css
);
347 tg
= &init_task_group
;
352 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
353 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
355 #ifdef CONFIG_FAIR_GROUP_SCHED
356 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
357 p
->se
.parent
= task_group(p
)->se
[cpu
];
360 #ifdef CONFIG_RT_GROUP_SCHED
361 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
362 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
368 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
369 static inline struct task_group
*task_group(struct task_struct
*p
)
374 #endif /* CONFIG_GROUP_SCHED */
376 /* CFS-related fields in a runqueue */
378 struct load_weight load
;
379 unsigned long nr_running
;
385 struct rb_root tasks_timeline
;
386 struct rb_node
*rb_leftmost
;
388 struct list_head tasks
;
389 struct list_head
*balance_iterator
;
392 * 'curr' points to currently running entity on this cfs_rq.
393 * It is set to NULL otherwise (i.e when none are currently running).
395 struct sched_entity
*curr
, *next
;
397 unsigned long nr_spread_over
;
399 #ifdef CONFIG_FAIR_GROUP_SCHED
400 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
403 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
404 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
405 * (like users, containers etc.)
407 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
408 * list is used during load balance.
410 struct list_head leaf_cfs_rq_list
;
411 struct task_group
*tg
; /* group that "owns" this runqueue */
415 * the part of load.weight contributed by tasks
417 unsigned long task_weight
;
420 * h_load = weight * f(tg)
422 * Where f(tg) is the recursive weight fraction assigned to
425 unsigned long h_load
;
428 * this cpu's part of tg->shares
430 unsigned long shares
;
433 * load.weight at the time we set shares
435 unsigned long rq_weight
;
440 /* Real-Time classes' related field in a runqueue: */
442 struct rt_prio_array active
;
443 unsigned long rt_nr_running
;
444 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
445 int highest_prio
; /* highest queued rt task prio */
448 unsigned long rt_nr_migratory
;
454 /* Nests inside the rq lock: */
455 spinlock_t rt_runtime_lock
;
457 #ifdef CONFIG_RT_GROUP_SCHED
458 unsigned long rt_nr_boosted
;
461 struct list_head leaf_rt_rq_list
;
462 struct task_group
*tg
;
463 struct sched_rt_entity
*rt_se
;
470 * We add the notion of a root-domain which will be used to define per-domain
471 * variables. Each exclusive cpuset essentially defines an island domain by
472 * fully partitioning the member cpus from any other cpuset. Whenever a new
473 * exclusive cpuset is created, we also create and attach a new root-domain
483 * The "RT overload" flag: it gets set if a CPU has more than
484 * one runnable RT task.
489 struct cpupri cpupri
;
494 * By default the system creates a single root-domain with all cpus as
495 * members (mimicking the global state we have today).
497 static struct root_domain def_root_domain
;
502 * This is the main, per-CPU runqueue data structure.
504 * Locking rule: those places that want to lock multiple runqueues
505 * (such as the load balancing or the thread migration code), lock
506 * acquire operations must be ordered by ascending &runqueue.
513 * nr_running and cpu_load should be in the same cacheline because
514 * remote CPUs use both these fields when doing load calculation.
516 unsigned long nr_running
;
517 #define CPU_LOAD_IDX_MAX 5
518 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
519 unsigned char idle_at_tick
;
521 unsigned long last_tick_seen
;
522 unsigned char in_nohz_recently
;
524 /* capture load from *all* tasks on this cpu: */
525 struct load_weight load
;
526 unsigned long nr_load_updates
;
532 #ifdef CONFIG_FAIR_GROUP_SCHED
533 /* list of leaf cfs_rq on this cpu: */
534 struct list_head leaf_cfs_rq_list
;
536 #ifdef CONFIG_RT_GROUP_SCHED
537 struct list_head leaf_rt_rq_list
;
541 * This is part of a global counter where only the total sum
542 * over all CPUs matters. A task can increase this counter on
543 * one CPU and if it got migrated afterwards it may decrease
544 * it on another CPU. Always updated under the runqueue lock:
546 unsigned long nr_uninterruptible
;
548 struct task_struct
*curr
, *idle
;
549 unsigned long next_balance
;
550 struct mm_struct
*prev_mm
;
557 struct root_domain
*rd
;
558 struct sched_domain
*sd
;
560 /* For active balancing */
563 /* cpu of this runqueue: */
567 unsigned long avg_load_per_task
;
569 struct task_struct
*migration_thread
;
570 struct list_head migration_queue
;
573 #ifdef CONFIG_SCHED_HRTICK
575 int hrtick_csd_pending
;
576 struct call_single_data hrtick_csd
;
578 struct hrtimer hrtick_timer
;
581 #ifdef CONFIG_SCHEDSTATS
583 struct sched_info rq_sched_info
;
585 /* sys_sched_yield() stats */
586 unsigned int yld_exp_empty
;
587 unsigned int yld_act_empty
;
588 unsigned int yld_both_empty
;
589 unsigned int yld_count
;
591 /* schedule() stats */
592 unsigned int sched_switch
;
593 unsigned int sched_count
;
594 unsigned int sched_goidle
;
596 /* try_to_wake_up() stats */
597 unsigned int ttwu_count
;
598 unsigned int ttwu_local
;
601 unsigned int bkl_count
;
605 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
607 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int sync
)
609 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, sync
);
612 static inline int cpu_of(struct rq
*rq
)
622 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
623 * See detach_destroy_domains: synchronize_sched for details.
625 * The domain tree of any CPU may only be accessed from within
626 * preempt-disabled sections.
628 #define for_each_domain(cpu, __sd) \
629 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
631 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
632 #define this_rq() (&__get_cpu_var(runqueues))
633 #define task_rq(p) cpu_rq(task_cpu(p))
634 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
636 static inline void update_rq_clock(struct rq
*rq
)
638 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
642 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
644 #ifdef CONFIG_SCHED_DEBUG
645 # define const_debug __read_mostly
647 # define const_debug static const
653 * Returns true if the current cpu runqueue is locked.
654 * This interface allows printk to be called with the runqueue lock
655 * held and know whether or not it is OK to wake up the klogd.
657 int runqueue_is_locked(void)
660 struct rq
*rq
= cpu_rq(cpu
);
663 ret
= spin_is_locked(&rq
->lock
);
669 * Debugging: various feature bits
672 #define SCHED_FEAT(name, enabled) \
673 __SCHED_FEAT_##name ,
676 #include "sched_features.h"
681 #define SCHED_FEAT(name, enabled) \
682 (1UL << __SCHED_FEAT_##name) * enabled |
684 const_debug
unsigned int sysctl_sched_features
=
685 #include "sched_features.h"
690 #ifdef CONFIG_SCHED_DEBUG
691 #define SCHED_FEAT(name, enabled) \
694 static __read_mostly
char *sched_feat_names
[] = {
695 #include "sched_features.h"
701 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
703 filp
->private_data
= inode
->i_private
;
708 sched_feat_read(struct file
*filp
, char __user
*ubuf
,
709 size_t cnt
, loff_t
*ppos
)
716 for (i
= 0; sched_feat_names
[i
]; i
++) {
717 len
+= strlen(sched_feat_names
[i
]);
721 buf
= kmalloc(len
+ 2, GFP_KERNEL
);
725 for (i
= 0; sched_feat_names
[i
]; i
++) {
726 if (sysctl_sched_features
& (1UL << i
))
727 r
+= sprintf(buf
+ r
, "%s ", sched_feat_names
[i
]);
729 r
+= sprintf(buf
+ r
, "NO_%s ", sched_feat_names
[i
]);
732 r
+= sprintf(buf
+ r
, "\n");
733 WARN_ON(r
>= len
+ 2);
735 r
= simple_read_from_buffer(ubuf
, cnt
, ppos
, buf
, r
);
743 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
744 size_t cnt
, loff_t
*ppos
)
754 if (copy_from_user(&buf
, ubuf
, cnt
))
759 if (strncmp(buf
, "NO_", 3) == 0) {
764 for (i
= 0; sched_feat_names
[i
]; i
++) {
765 int len
= strlen(sched_feat_names
[i
]);
767 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
769 sysctl_sched_features
&= ~(1UL << i
);
771 sysctl_sched_features
|= (1UL << i
);
776 if (!sched_feat_names
[i
])
784 static struct file_operations sched_feat_fops
= {
785 .open
= sched_feat_open
,
786 .read
= sched_feat_read
,
787 .write
= sched_feat_write
,
790 static __init
int sched_init_debug(void)
792 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
797 late_initcall(sched_init_debug
);
801 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
804 * Number of tasks to iterate in a single balance run.
805 * Limited because this is done with IRQs disabled.
807 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
810 * ratelimit for updating the group shares.
813 unsigned int sysctl_sched_shares_ratelimit
= 250000;
816 * period over which we measure -rt task cpu usage in us.
819 unsigned int sysctl_sched_rt_period
= 1000000;
821 static __read_mostly
int scheduler_running
;
824 * part of the period that we allow rt tasks to run in us.
827 int sysctl_sched_rt_runtime
= 950000;
829 static inline u64
global_rt_period(void)
831 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
834 static inline u64
global_rt_runtime(void)
836 if (sysctl_sched_rt_runtime
< 0)
839 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
842 #ifndef prepare_arch_switch
843 # define prepare_arch_switch(next) do { } while (0)
845 #ifndef finish_arch_switch
846 # define finish_arch_switch(prev) do { } while (0)
849 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
851 return rq
->curr
== p
;
854 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
855 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
857 return task_current(rq
, p
);
860 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
864 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
866 #ifdef CONFIG_DEBUG_SPINLOCK
867 /* this is a valid case when another task releases the spinlock */
868 rq
->lock
.owner
= current
;
871 * If we are tracking spinlock dependencies then we have to
872 * fix up the runqueue lock - which gets 'carried over' from
875 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
877 spin_unlock_irq(&rq
->lock
);
880 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
881 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
886 return task_current(rq
, p
);
890 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
894 * We can optimise this out completely for !SMP, because the
895 * SMP rebalancing from interrupt is the only thing that cares
900 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
901 spin_unlock_irq(&rq
->lock
);
903 spin_unlock(&rq
->lock
);
907 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
911 * After ->oncpu is cleared, the task can be moved to a different CPU.
912 * We must ensure this doesn't happen until the switch is completely
918 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
922 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
925 * __task_rq_lock - lock the runqueue a given task resides on.
926 * Must be called interrupts disabled.
928 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
932 struct rq
*rq
= task_rq(p
);
933 spin_lock(&rq
->lock
);
934 if (likely(rq
== task_rq(p
)))
936 spin_unlock(&rq
->lock
);
941 * task_rq_lock - lock the runqueue a given task resides on and disable
942 * interrupts. Note the ordering: we can safely lookup the task_rq without
943 * explicitly disabling preemption.
945 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
951 local_irq_save(*flags
);
953 spin_lock(&rq
->lock
);
954 if (likely(rq
== task_rq(p
)))
956 spin_unlock_irqrestore(&rq
->lock
, *flags
);
960 static void __task_rq_unlock(struct rq
*rq
)
963 spin_unlock(&rq
->lock
);
966 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
969 spin_unlock_irqrestore(&rq
->lock
, *flags
);
973 * this_rq_lock - lock this runqueue and disable interrupts.
975 static struct rq
*this_rq_lock(void)
982 spin_lock(&rq
->lock
);
987 #ifdef CONFIG_SCHED_HRTICK
989 * Use HR-timers to deliver accurate preemption points.
991 * Its all a bit involved since we cannot program an hrt while holding the
992 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
995 * When we get rescheduled we reprogram the hrtick_timer outside of the
1001 * - enabled by features
1002 * - hrtimer is actually high res
1004 static inline int hrtick_enabled(struct rq
*rq
)
1006 if (!sched_feat(HRTICK
))
1008 if (!cpu_active(cpu_of(rq
)))
1010 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1013 static void hrtick_clear(struct rq
*rq
)
1015 if (hrtimer_active(&rq
->hrtick_timer
))
1016 hrtimer_cancel(&rq
->hrtick_timer
);
1020 * High-resolution timer tick.
1021 * Runs from hardirq context with interrupts disabled.
1023 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1025 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1027 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1029 spin_lock(&rq
->lock
);
1030 update_rq_clock(rq
);
1031 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1032 spin_unlock(&rq
->lock
);
1034 return HRTIMER_NORESTART
;
1039 * called from hardirq (IPI) context
1041 static void __hrtick_start(void *arg
)
1043 struct rq
*rq
= arg
;
1045 spin_lock(&rq
->lock
);
1046 hrtimer_restart(&rq
->hrtick_timer
);
1047 rq
->hrtick_csd_pending
= 0;
1048 spin_unlock(&rq
->lock
);
1052 * Called to set the hrtick timer state.
1054 * called with rq->lock held and irqs disabled
1056 static void hrtick_start(struct rq
*rq
, u64 delay
)
1058 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1059 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1061 timer
->expires
= time
;
1063 if (rq
== this_rq()) {
1064 hrtimer_restart(timer
);
1065 } else if (!rq
->hrtick_csd_pending
) {
1066 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
);
1067 rq
->hrtick_csd_pending
= 1;
1072 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1074 int cpu
= (int)(long)hcpu
;
1077 case CPU_UP_CANCELED
:
1078 case CPU_UP_CANCELED_FROZEN
:
1079 case CPU_DOWN_PREPARE
:
1080 case CPU_DOWN_PREPARE_FROZEN
:
1082 case CPU_DEAD_FROZEN
:
1083 hrtick_clear(cpu_rq(cpu
));
1090 static void init_hrtick(void)
1092 hotcpu_notifier(hotplug_hrtick
, 0);
1096 * Called to set the hrtick timer state.
1098 * called with rq->lock held and irqs disabled
1100 static void hrtick_start(struct rq
*rq
, u64 delay
)
1102 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
), HRTIMER_MODE_REL
);
1105 static inline void init_hrtick(void)
1108 #endif /* CONFIG_SMP */
1110 static void init_rq_hrtick(struct rq
*rq
)
1113 rq
->hrtick_csd_pending
= 0;
1115 rq
->hrtick_csd
.flags
= 0;
1116 rq
->hrtick_csd
.func
= __hrtick_start
;
1117 rq
->hrtick_csd
.info
= rq
;
1120 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1121 rq
->hrtick_timer
.function
= hrtick
;
1122 rq
->hrtick_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_NO_SOFTIRQ
;
1124 #else /* CONFIG_SCHED_HRTICK */
1125 static inline void hrtick_clear(struct rq
*rq
)
1129 static inline void init_rq_hrtick(struct rq
*rq
)
1133 static inline void init_hrtick(void)
1136 #endif /* CONFIG_SCHED_HRTICK */
1139 * resched_task - mark a task 'to be rescheduled now'.
1141 * On UP this means the setting of the need_resched flag, on SMP it
1142 * might also involve a cross-CPU call to trigger the scheduler on
1147 #ifndef tsk_is_polling
1148 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1151 static void resched_task(struct task_struct
*p
)
1155 assert_spin_locked(&task_rq(p
)->lock
);
1157 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
1160 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
1163 if (cpu
== smp_processor_id())
1166 /* NEED_RESCHED must be visible before we test polling */
1168 if (!tsk_is_polling(p
))
1169 smp_send_reschedule(cpu
);
1172 static void resched_cpu(int cpu
)
1174 struct rq
*rq
= cpu_rq(cpu
);
1175 unsigned long flags
;
1177 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1179 resched_task(cpu_curr(cpu
));
1180 spin_unlock_irqrestore(&rq
->lock
, flags
);
1185 * When add_timer_on() enqueues a timer into the timer wheel of an
1186 * idle CPU then this timer might expire before the next timer event
1187 * which is scheduled to wake up that CPU. In case of a completely
1188 * idle system the next event might even be infinite time into the
1189 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1190 * leaves the inner idle loop so the newly added timer is taken into
1191 * account when the CPU goes back to idle and evaluates the timer
1192 * wheel for the next timer event.
1194 void wake_up_idle_cpu(int cpu
)
1196 struct rq
*rq
= cpu_rq(cpu
);
1198 if (cpu
== smp_processor_id())
1202 * This is safe, as this function is called with the timer
1203 * wheel base lock of (cpu) held. When the CPU is on the way
1204 * to idle and has not yet set rq->curr to idle then it will
1205 * be serialized on the timer wheel base lock and take the new
1206 * timer into account automatically.
1208 if (rq
->curr
!= rq
->idle
)
1212 * We can set TIF_RESCHED on the idle task of the other CPU
1213 * lockless. The worst case is that the other CPU runs the
1214 * idle task through an additional NOOP schedule()
1216 set_tsk_thread_flag(rq
->idle
, TIF_NEED_RESCHED
);
1218 /* NEED_RESCHED must be visible before we test polling */
1220 if (!tsk_is_polling(rq
->idle
))
1221 smp_send_reschedule(cpu
);
1223 #endif /* CONFIG_NO_HZ */
1225 #else /* !CONFIG_SMP */
1226 static void resched_task(struct task_struct
*p
)
1228 assert_spin_locked(&task_rq(p
)->lock
);
1229 set_tsk_need_resched(p
);
1231 #endif /* CONFIG_SMP */
1233 #if BITS_PER_LONG == 32
1234 # define WMULT_CONST (~0UL)
1236 # define WMULT_CONST (1UL << 32)
1239 #define WMULT_SHIFT 32
1242 * Shift right and round:
1244 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1247 * delta *= weight / lw
1249 static unsigned long
1250 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1251 struct load_weight
*lw
)
1255 if (!lw
->inv_weight
) {
1256 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1259 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1263 tmp
= (u64
)delta_exec
* weight
;
1265 * Check whether we'd overflow the 64-bit multiplication:
1267 if (unlikely(tmp
> WMULT_CONST
))
1268 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1271 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1273 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1276 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1282 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1289 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1290 * of tasks with abnormal "nice" values across CPUs the contribution that
1291 * each task makes to its run queue's load is weighted according to its
1292 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1293 * scaled version of the new time slice allocation that they receive on time
1297 #define WEIGHT_IDLEPRIO 2
1298 #define WMULT_IDLEPRIO (1 << 31)
1301 * Nice levels are multiplicative, with a gentle 10% change for every
1302 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1303 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1304 * that remained on nice 0.
1306 * The "10% effect" is relative and cumulative: from _any_ nice level,
1307 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1308 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1309 * If a task goes up by ~10% and another task goes down by ~10% then
1310 * the relative distance between them is ~25%.)
1312 static const int prio_to_weight
[40] = {
1313 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1314 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1315 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1316 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1317 /* 0 */ 1024, 820, 655, 526, 423,
1318 /* 5 */ 335, 272, 215, 172, 137,
1319 /* 10 */ 110, 87, 70, 56, 45,
1320 /* 15 */ 36, 29, 23, 18, 15,
1324 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1326 * In cases where the weight does not change often, we can use the
1327 * precalculated inverse to speed up arithmetics by turning divisions
1328 * into multiplications:
1330 static const u32 prio_to_wmult
[40] = {
1331 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1332 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1333 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1334 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1335 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1336 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1337 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1338 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1341 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1344 * runqueue iterator, to support SMP load-balancing between different
1345 * scheduling classes, without having to expose their internal data
1346 * structures to the load-balancing proper:
1348 struct rq_iterator
{
1350 struct task_struct
*(*start
)(void *);
1351 struct task_struct
*(*next
)(void *);
1355 static unsigned long
1356 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1357 unsigned long max_load_move
, struct sched_domain
*sd
,
1358 enum cpu_idle_type idle
, int *all_pinned
,
1359 int *this_best_prio
, struct rq_iterator
*iterator
);
1362 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1363 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1364 struct rq_iterator
*iterator
);
1367 #ifdef CONFIG_CGROUP_CPUACCT
1368 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1370 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1373 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1375 update_load_add(&rq
->load
, load
);
1378 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1380 update_load_sub(&rq
->load
, load
);
1384 static unsigned long source_load(int cpu
, int type
);
1385 static unsigned long target_load(int cpu
, int type
);
1386 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1388 static unsigned long cpu_avg_load_per_task(int cpu
)
1390 struct rq
*rq
= cpu_rq(cpu
);
1393 rq
->avg_load_per_task
= rq
->load
.weight
/ rq
->nr_running
;
1395 return rq
->avg_load_per_task
;
1398 #ifdef CONFIG_FAIR_GROUP_SCHED
1400 typedef void (*tg_visitor
)(struct task_group
*, int, struct sched_domain
*);
1403 * Iterate the full tree, calling @down when first entering a node and @up when
1404 * leaving it for the final time.
1407 walk_tg_tree(tg_visitor down
, tg_visitor up
, int cpu
, struct sched_domain
*sd
)
1409 struct task_group
*parent
, *child
;
1412 parent
= &root_task_group
;
1414 (*down
)(parent
, cpu
, sd
);
1415 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1422 (*up
)(parent
, cpu
, sd
);
1425 parent
= parent
->parent
;
1431 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1434 * Calculate and set the cpu's group shares.
1437 __update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1438 unsigned long sd_shares
, unsigned long sd_rq_weight
)
1441 unsigned long shares
;
1442 unsigned long rq_weight
;
1447 rq_weight
= tg
->cfs_rq
[cpu
]->load
.weight
;
1450 * If there are currently no tasks on the cpu pretend there is one of
1451 * average load so that when a new task gets to run here it will not
1452 * get delayed by group starvation.
1456 rq_weight
= NICE_0_LOAD
;
1459 if (unlikely(rq_weight
> sd_rq_weight
))
1460 rq_weight
= sd_rq_weight
;
1463 * \Sum shares * rq_weight
1464 * shares = -----------------------
1468 shares
= (sd_shares
* rq_weight
) / (sd_rq_weight
+ 1);
1471 * record the actual number of shares, not the boosted amount.
1473 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1474 tg
->cfs_rq
[cpu
]->rq_weight
= rq_weight
;
1476 if (shares
< MIN_SHARES
)
1477 shares
= MIN_SHARES
;
1478 else if (shares
> MAX_SHARES
)
1479 shares
= MAX_SHARES
;
1481 __set_se_shares(tg
->se
[cpu
], shares
);
1485 * Re-compute the task group their per cpu shares over the given domain.
1486 * This needs to be done in a bottom-up fashion because the rq weight of a
1487 * parent group depends on the shares of its child groups.
1490 tg_shares_up(struct task_group
*tg
, int cpu
, struct sched_domain
*sd
)
1492 unsigned long rq_weight
= 0;
1493 unsigned long shares
= 0;
1496 for_each_cpu_mask(i
, sd
->span
) {
1497 rq_weight
+= tg
->cfs_rq
[i
]->load
.weight
;
1498 shares
+= tg
->cfs_rq
[i
]->shares
;
1501 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1502 shares
= tg
->shares
;
1504 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1505 shares
= tg
->shares
;
1508 rq_weight
= cpus_weight(sd
->span
) * NICE_0_LOAD
;
1510 for_each_cpu_mask(i
, sd
->span
) {
1511 struct rq
*rq
= cpu_rq(i
);
1512 unsigned long flags
;
1514 spin_lock_irqsave(&rq
->lock
, flags
);
1515 __update_group_shares_cpu(tg
, i
, shares
, rq_weight
);
1516 spin_unlock_irqrestore(&rq
->lock
, flags
);
1521 * Compute the cpu's hierarchical load factor for each task group.
1522 * This needs to be done in a top-down fashion because the load of a child
1523 * group is a fraction of its parents load.
1526 tg_load_down(struct task_group
*tg
, int cpu
, struct sched_domain
*sd
)
1531 load
= cpu_rq(cpu
)->load
.weight
;
1533 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1534 load
*= tg
->cfs_rq
[cpu
]->shares
;
1535 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1538 tg
->cfs_rq
[cpu
]->h_load
= load
;
1542 tg_nop(struct task_group
*tg
, int cpu
, struct sched_domain
*sd
)
1546 static void update_shares(struct sched_domain
*sd
)
1548 u64 now
= cpu_clock(raw_smp_processor_id());
1549 s64 elapsed
= now
- sd
->last_update
;
1551 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1552 sd
->last_update
= now
;
1553 walk_tg_tree(tg_nop
, tg_shares_up
, 0, sd
);
1557 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1559 spin_unlock(&rq
->lock
);
1561 spin_lock(&rq
->lock
);
1564 static void update_h_load(int cpu
)
1566 walk_tg_tree(tg_load_down
, tg_nop
, cpu
, NULL
);
1571 static inline void update_shares(struct sched_domain
*sd
)
1575 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1583 #ifdef CONFIG_FAIR_GROUP_SCHED
1584 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1587 cfs_rq
->shares
= shares
;
1592 #include "sched_stats.h"
1593 #include "sched_idletask.c"
1594 #include "sched_fair.c"
1595 #include "sched_rt.c"
1596 #ifdef CONFIG_SCHED_DEBUG
1597 # include "sched_debug.c"
1600 #define sched_class_highest (&rt_sched_class)
1601 #define for_each_class(class) \
1602 for (class = sched_class_highest; class; class = class->next)
1604 static void inc_nr_running(struct rq
*rq
)
1609 static void dec_nr_running(struct rq
*rq
)
1614 static void set_load_weight(struct task_struct
*p
)
1616 if (task_has_rt_policy(p
)) {
1617 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1618 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1623 * SCHED_IDLE tasks get minimal weight:
1625 if (p
->policy
== SCHED_IDLE
) {
1626 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1627 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1631 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1632 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1635 static void update_avg(u64
*avg
, u64 sample
)
1637 s64 diff
= sample
- *avg
;
1641 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1643 sched_info_queued(p
);
1644 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1648 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1650 if (sleep
&& p
->se
.last_wakeup
) {
1651 update_avg(&p
->se
.avg_overlap
,
1652 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1653 p
->se
.last_wakeup
= 0;
1656 sched_info_dequeued(p
);
1657 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1662 * __normal_prio - return the priority that is based on the static prio
1664 static inline int __normal_prio(struct task_struct
*p
)
1666 return p
->static_prio
;
1670 * Calculate the expected normal priority: i.e. priority
1671 * without taking RT-inheritance into account. Might be
1672 * boosted by interactivity modifiers. Changes upon fork,
1673 * setprio syscalls, and whenever the interactivity
1674 * estimator recalculates.
1676 static inline int normal_prio(struct task_struct
*p
)
1680 if (task_has_rt_policy(p
))
1681 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1683 prio
= __normal_prio(p
);
1688 * Calculate the current priority, i.e. the priority
1689 * taken into account by the scheduler. This value might
1690 * be boosted by RT tasks, or might be boosted by
1691 * interactivity modifiers. Will be RT if the task got
1692 * RT-boosted. If not then it returns p->normal_prio.
1694 static int effective_prio(struct task_struct
*p
)
1696 p
->normal_prio
= normal_prio(p
);
1698 * If we are RT tasks or we were boosted to RT priority,
1699 * keep the priority unchanged. Otherwise, update priority
1700 * to the normal priority:
1702 if (!rt_prio(p
->prio
))
1703 return p
->normal_prio
;
1708 * activate_task - move a task to the runqueue.
1710 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1712 if (task_contributes_to_load(p
))
1713 rq
->nr_uninterruptible
--;
1715 enqueue_task(rq
, p
, wakeup
);
1720 * deactivate_task - remove a task from the runqueue.
1722 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1724 if (task_contributes_to_load(p
))
1725 rq
->nr_uninterruptible
++;
1727 dequeue_task(rq
, p
, sleep
);
1732 * task_curr - is this task currently executing on a CPU?
1733 * @p: the task in question.
1735 inline int task_curr(const struct task_struct
*p
)
1737 return cpu_curr(task_cpu(p
)) == p
;
1740 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1742 set_task_rq(p
, cpu
);
1745 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1746 * successfuly executed on another CPU. We must ensure that updates of
1747 * per-task data have been completed by this moment.
1750 task_thread_info(p
)->cpu
= cpu
;
1754 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1755 const struct sched_class
*prev_class
,
1756 int oldprio
, int running
)
1758 if (prev_class
!= p
->sched_class
) {
1759 if (prev_class
->switched_from
)
1760 prev_class
->switched_from(rq
, p
, running
);
1761 p
->sched_class
->switched_to(rq
, p
, running
);
1763 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1768 /* Used instead of source_load when we know the type == 0 */
1769 static unsigned long weighted_cpuload(const int cpu
)
1771 return cpu_rq(cpu
)->load
.weight
;
1775 * Is this task likely cache-hot:
1778 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1783 * Buddy candidates are cache hot:
1785 if (sched_feat(CACHE_HOT_BUDDY
) && (&p
->se
== cfs_rq_of(&p
->se
)->next
))
1788 if (p
->sched_class
!= &fair_sched_class
)
1791 if (sysctl_sched_migration_cost
== -1)
1793 if (sysctl_sched_migration_cost
== 0)
1796 delta
= now
- p
->se
.exec_start
;
1798 return delta
< (s64
)sysctl_sched_migration_cost
;
1802 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1804 int old_cpu
= task_cpu(p
);
1805 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1806 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1807 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1810 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1812 #ifdef CONFIG_SCHEDSTATS
1813 if (p
->se
.wait_start
)
1814 p
->se
.wait_start
-= clock_offset
;
1815 if (p
->se
.sleep_start
)
1816 p
->se
.sleep_start
-= clock_offset
;
1817 if (p
->se
.block_start
)
1818 p
->se
.block_start
-= clock_offset
;
1819 if (old_cpu
!= new_cpu
) {
1820 schedstat_inc(p
, se
.nr_migrations
);
1821 if (task_hot(p
, old_rq
->clock
, NULL
))
1822 schedstat_inc(p
, se
.nr_forced2_migrations
);
1825 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1826 new_cfsrq
->min_vruntime
;
1828 __set_task_cpu(p
, new_cpu
);
1831 struct migration_req
{
1832 struct list_head list
;
1834 struct task_struct
*task
;
1837 struct completion done
;
1841 * The task's runqueue lock must be held.
1842 * Returns true if you have to wait for migration thread.
1845 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1847 struct rq
*rq
= task_rq(p
);
1850 * If the task is not on a runqueue (and not running), then
1851 * it is sufficient to simply update the task's cpu field.
1853 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1854 set_task_cpu(p
, dest_cpu
);
1858 init_completion(&req
->done
);
1860 req
->dest_cpu
= dest_cpu
;
1861 list_add(&req
->list
, &rq
->migration_queue
);
1867 * wait_task_inactive - wait for a thread to unschedule.
1869 * If @match_state is nonzero, it's the @p->state value just checked and
1870 * not expected to change. If it changes, i.e. @p might have woken up,
1871 * then return zero. When we succeed in waiting for @p to be off its CPU,
1872 * we return a positive number (its total switch count). If a second call
1873 * a short while later returns the same number, the caller can be sure that
1874 * @p has remained unscheduled the whole time.
1876 * The caller must ensure that the task *will* unschedule sometime soon,
1877 * else this function might spin for a *long* time. This function can't
1878 * be called with interrupts off, or it may introduce deadlock with
1879 * smp_call_function() if an IPI is sent by the same process we are
1880 * waiting to become inactive.
1882 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1884 unsigned long flags
;
1891 * We do the initial early heuristics without holding
1892 * any task-queue locks at all. We'll only try to get
1893 * the runqueue lock when things look like they will
1899 * If the task is actively running on another CPU
1900 * still, just relax and busy-wait without holding
1903 * NOTE! Since we don't hold any locks, it's not
1904 * even sure that "rq" stays as the right runqueue!
1905 * But we don't care, since "task_running()" will
1906 * return false if the runqueue has changed and p
1907 * is actually now running somewhere else!
1909 while (task_running(rq
, p
)) {
1910 if (match_state
&& unlikely(p
->state
!= match_state
))
1916 * Ok, time to look more closely! We need the rq
1917 * lock now, to be *sure*. If we're wrong, we'll
1918 * just go back and repeat.
1920 rq
= task_rq_lock(p
, &flags
);
1921 running
= task_running(rq
, p
);
1922 on_rq
= p
->se
.on_rq
;
1924 if (!match_state
|| p
->state
== match_state
)
1925 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1926 task_rq_unlock(rq
, &flags
);
1929 * If it changed from the expected state, bail out now.
1931 if (unlikely(!ncsw
))
1935 * Was it really running after all now that we
1936 * checked with the proper locks actually held?
1938 * Oops. Go back and try again..
1940 if (unlikely(running
)) {
1946 * It's not enough that it's not actively running,
1947 * it must be off the runqueue _entirely_, and not
1950 * So if it wa still runnable (but just not actively
1951 * running right now), it's preempted, and we should
1952 * yield - it could be a while.
1954 if (unlikely(on_rq
)) {
1955 schedule_timeout_uninterruptible(1);
1960 * Ahh, all good. It wasn't running, and it wasn't
1961 * runnable, which means that it will never become
1962 * running in the future either. We're all done!
1971 * kick_process - kick a running thread to enter/exit the kernel
1972 * @p: the to-be-kicked thread
1974 * Cause a process which is running on another CPU to enter
1975 * kernel-mode, without any delay. (to get signals handled.)
1977 * NOTE: this function doesnt have to take the runqueue lock,
1978 * because all it wants to ensure is that the remote task enters
1979 * the kernel. If the IPI races and the task has been migrated
1980 * to another CPU then no harm is done and the purpose has been
1983 void kick_process(struct task_struct
*p
)
1989 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1990 smp_send_reschedule(cpu
);
1995 * Return a low guess at the load of a migration-source cpu weighted
1996 * according to the scheduling class and "nice" value.
1998 * We want to under-estimate the load of migration sources, to
1999 * balance conservatively.
2001 static unsigned long source_load(int cpu
, int type
)
2003 struct rq
*rq
= cpu_rq(cpu
);
2004 unsigned long total
= weighted_cpuload(cpu
);
2006 if (type
== 0 || !sched_feat(LB_BIAS
))
2009 return min(rq
->cpu_load
[type
-1], total
);
2013 * Return a high guess at the load of a migration-target cpu weighted
2014 * according to the scheduling class and "nice" value.
2016 static unsigned long target_load(int cpu
, int type
)
2018 struct rq
*rq
= cpu_rq(cpu
);
2019 unsigned long total
= weighted_cpuload(cpu
);
2021 if (type
== 0 || !sched_feat(LB_BIAS
))
2024 return max(rq
->cpu_load
[type
-1], total
);
2028 * find_idlest_group finds and returns the least busy CPU group within the
2031 static struct sched_group
*
2032 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2034 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2035 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2036 int load_idx
= sd
->forkexec_idx
;
2037 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2040 unsigned long load
, avg_load
;
2044 /* Skip over this group if it has no CPUs allowed */
2045 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
2048 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2050 /* Tally up the load of all CPUs in the group */
2053 for_each_cpu_mask_nr(i
, group
->cpumask
) {
2054 /* Bias balancing toward cpus of our domain */
2056 load
= source_load(i
, load_idx
);
2058 load
= target_load(i
, load_idx
);
2063 /* Adjust by relative CPU power of the group */
2064 avg_load
= sg_div_cpu_power(group
,
2065 avg_load
* SCHED_LOAD_SCALE
);
2068 this_load
= avg_load
;
2070 } else if (avg_load
< min_load
) {
2071 min_load
= avg_load
;
2074 } while (group
= group
->next
, group
!= sd
->groups
);
2076 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2082 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2085 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
,
2088 unsigned long load
, min_load
= ULONG_MAX
;
2092 /* Traverse only the allowed CPUs */
2093 cpus_and(*tmp
, group
->cpumask
, p
->cpus_allowed
);
2095 for_each_cpu_mask_nr(i
, *tmp
) {
2096 load
= weighted_cpuload(i
);
2098 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2108 * sched_balance_self: balance the current task (running on cpu) in domains
2109 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2112 * Balance, ie. select the least loaded group.
2114 * Returns the target CPU number, or the same CPU if no balancing is needed.
2116 * preempt must be disabled.
2118 static int sched_balance_self(int cpu
, int flag
)
2120 struct task_struct
*t
= current
;
2121 struct sched_domain
*tmp
, *sd
= NULL
;
2123 for_each_domain(cpu
, tmp
) {
2125 * If power savings logic is enabled for a domain, stop there.
2127 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2129 if (tmp
->flags
& flag
)
2137 cpumask_t span
, tmpmask
;
2138 struct sched_group
*group
;
2139 int new_cpu
, weight
;
2141 if (!(sd
->flags
& flag
)) {
2147 group
= find_idlest_group(sd
, t
, cpu
);
2153 new_cpu
= find_idlest_cpu(group
, t
, cpu
, &tmpmask
);
2154 if (new_cpu
== -1 || new_cpu
== cpu
) {
2155 /* Now try balancing at a lower domain level of cpu */
2160 /* Now try balancing at a lower domain level of new_cpu */
2163 weight
= cpus_weight(span
);
2164 for_each_domain(cpu
, tmp
) {
2165 if (weight
<= cpus_weight(tmp
->span
))
2167 if (tmp
->flags
& flag
)
2170 /* while loop will break here if sd == NULL */
2176 #endif /* CONFIG_SMP */
2179 * try_to_wake_up - wake up a thread
2180 * @p: the to-be-woken-up thread
2181 * @state: the mask of task states that can be woken
2182 * @sync: do a synchronous wakeup?
2184 * Put it on the run-queue if it's not already there. The "current"
2185 * thread is always on the run-queue (except when the actual
2186 * re-schedule is in progress), and as such you're allowed to do
2187 * the simpler "current->state = TASK_RUNNING" to mark yourself
2188 * runnable without the overhead of this.
2190 * returns failure only if the task is already active.
2192 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2194 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2195 unsigned long flags
;
2199 if (!sched_feat(SYNC_WAKEUPS
))
2203 if (sched_feat(LB_WAKEUP_UPDATE
)) {
2204 struct sched_domain
*sd
;
2206 this_cpu
= raw_smp_processor_id();
2209 for_each_domain(this_cpu
, sd
) {
2210 if (cpu_isset(cpu
, sd
->span
)) {
2219 rq
= task_rq_lock(p
, &flags
);
2220 old_state
= p
->state
;
2221 if (!(old_state
& state
))
2229 this_cpu
= smp_processor_id();
2232 if (unlikely(task_running(rq
, p
)))
2235 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2236 if (cpu
!= orig_cpu
) {
2237 set_task_cpu(p
, cpu
);
2238 task_rq_unlock(rq
, &flags
);
2239 /* might preempt at this point */
2240 rq
= task_rq_lock(p
, &flags
);
2241 old_state
= p
->state
;
2242 if (!(old_state
& state
))
2247 this_cpu
= smp_processor_id();
2251 #ifdef CONFIG_SCHEDSTATS
2252 schedstat_inc(rq
, ttwu_count
);
2253 if (cpu
== this_cpu
)
2254 schedstat_inc(rq
, ttwu_local
);
2256 struct sched_domain
*sd
;
2257 for_each_domain(this_cpu
, sd
) {
2258 if (cpu_isset(cpu
, sd
->span
)) {
2259 schedstat_inc(sd
, ttwu_wake_remote
);
2264 #endif /* CONFIG_SCHEDSTATS */
2267 #endif /* CONFIG_SMP */
2268 schedstat_inc(p
, se
.nr_wakeups
);
2270 schedstat_inc(p
, se
.nr_wakeups_sync
);
2271 if (orig_cpu
!= cpu
)
2272 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2273 if (cpu
== this_cpu
)
2274 schedstat_inc(p
, se
.nr_wakeups_local
);
2276 schedstat_inc(p
, se
.nr_wakeups_remote
);
2277 update_rq_clock(rq
);
2278 activate_task(rq
, p
, 1);
2282 trace_mark(kernel_sched_wakeup
,
2283 "pid %d state %ld ## rq %p task %p rq->curr %p",
2284 p
->pid
, p
->state
, rq
, p
, rq
->curr
);
2285 check_preempt_curr(rq
, p
, sync
);
2287 p
->state
= TASK_RUNNING
;
2289 if (p
->sched_class
->task_wake_up
)
2290 p
->sched_class
->task_wake_up(rq
, p
);
2293 current
->se
.last_wakeup
= current
->se
.sum_exec_runtime
;
2295 task_rq_unlock(rq
, &flags
);
2300 int wake_up_process(struct task_struct
*p
)
2302 return try_to_wake_up(p
, TASK_ALL
, 0);
2304 EXPORT_SYMBOL(wake_up_process
);
2306 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2308 return try_to_wake_up(p
, state
, 0);
2312 * Perform scheduler related setup for a newly forked process p.
2313 * p is forked by current.
2315 * __sched_fork() is basic setup used by init_idle() too:
2317 static void __sched_fork(struct task_struct
*p
)
2319 p
->se
.exec_start
= 0;
2320 p
->se
.sum_exec_runtime
= 0;
2321 p
->se
.prev_sum_exec_runtime
= 0;
2322 p
->se
.last_wakeup
= 0;
2323 p
->se
.avg_overlap
= 0;
2325 #ifdef CONFIG_SCHEDSTATS
2326 p
->se
.wait_start
= 0;
2327 p
->se
.sum_sleep_runtime
= 0;
2328 p
->se
.sleep_start
= 0;
2329 p
->se
.block_start
= 0;
2330 p
->se
.sleep_max
= 0;
2331 p
->se
.block_max
= 0;
2333 p
->se
.slice_max
= 0;
2337 INIT_LIST_HEAD(&p
->rt
.run_list
);
2339 INIT_LIST_HEAD(&p
->se
.group_node
);
2341 #ifdef CONFIG_PREEMPT_NOTIFIERS
2342 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2346 * We mark the process as running here, but have not actually
2347 * inserted it onto the runqueue yet. This guarantees that
2348 * nobody will actually run it, and a signal or other external
2349 * event cannot wake it up and insert it on the runqueue either.
2351 p
->state
= TASK_RUNNING
;
2355 * fork()/clone()-time setup:
2357 void sched_fork(struct task_struct
*p
, int clone_flags
)
2359 int cpu
= get_cpu();
2364 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2366 set_task_cpu(p
, cpu
);
2369 * Make sure we do not leak PI boosting priority to the child:
2371 p
->prio
= current
->normal_prio
;
2372 if (!rt_prio(p
->prio
))
2373 p
->sched_class
= &fair_sched_class
;
2375 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2376 if (likely(sched_info_on()))
2377 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2379 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2382 #ifdef CONFIG_PREEMPT
2383 /* Want to start with kernel preemption disabled. */
2384 task_thread_info(p
)->preempt_count
= 1;
2390 * wake_up_new_task - wake up a newly created task for the first time.
2392 * This function will do some initial scheduler statistics housekeeping
2393 * that must be done for every newly created context, then puts the task
2394 * on the runqueue and wakes it.
2396 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2398 unsigned long flags
;
2401 rq
= task_rq_lock(p
, &flags
);
2402 BUG_ON(p
->state
!= TASK_RUNNING
);
2403 update_rq_clock(rq
);
2405 p
->prio
= effective_prio(p
);
2407 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2408 activate_task(rq
, p
, 0);
2411 * Let the scheduling class do new task startup
2412 * management (if any):
2414 p
->sched_class
->task_new(rq
, p
);
2417 trace_mark(kernel_sched_wakeup_new
,
2418 "pid %d state %ld ## rq %p task %p rq->curr %p",
2419 p
->pid
, p
->state
, rq
, p
, rq
->curr
);
2420 check_preempt_curr(rq
, p
, 0);
2422 if (p
->sched_class
->task_wake_up
)
2423 p
->sched_class
->task_wake_up(rq
, p
);
2425 task_rq_unlock(rq
, &flags
);
2428 #ifdef CONFIG_PREEMPT_NOTIFIERS
2431 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2432 * @notifier: notifier struct to register
2434 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2436 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2438 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2441 * preempt_notifier_unregister - no longer interested in preemption notifications
2442 * @notifier: notifier struct to unregister
2444 * This is safe to call from within a preemption notifier.
2446 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2448 hlist_del(¬ifier
->link
);
2450 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2452 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2454 struct preempt_notifier
*notifier
;
2455 struct hlist_node
*node
;
2457 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2458 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2462 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2463 struct task_struct
*next
)
2465 struct preempt_notifier
*notifier
;
2466 struct hlist_node
*node
;
2468 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2469 notifier
->ops
->sched_out(notifier
, next
);
2472 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2474 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2479 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2480 struct task_struct
*next
)
2484 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2487 * prepare_task_switch - prepare to switch tasks
2488 * @rq: the runqueue preparing to switch
2489 * @prev: the current task that is being switched out
2490 * @next: the task we are going to switch to.
2492 * This is called with the rq lock held and interrupts off. It must
2493 * be paired with a subsequent finish_task_switch after the context
2496 * prepare_task_switch sets up locking and calls architecture specific
2500 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2501 struct task_struct
*next
)
2503 fire_sched_out_preempt_notifiers(prev
, next
);
2504 prepare_lock_switch(rq
, next
);
2505 prepare_arch_switch(next
);
2509 * finish_task_switch - clean up after a task-switch
2510 * @rq: runqueue associated with task-switch
2511 * @prev: the thread we just switched away from.
2513 * finish_task_switch must be called after the context switch, paired
2514 * with a prepare_task_switch call before the context switch.
2515 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2516 * and do any other architecture-specific cleanup actions.
2518 * Note that we may have delayed dropping an mm in context_switch(). If
2519 * so, we finish that here outside of the runqueue lock. (Doing it
2520 * with the lock held can cause deadlocks; see schedule() for
2523 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2524 __releases(rq
->lock
)
2526 struct mm_struct
*mm
= rq
->prev_mm
;
2532 * A task struct has one reference for the use as "current".
2533 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2534 * schedule one last time. The schedule call will never return, and
2535 * the scheduled task must drop that reference.
2536 * The test for TASK_DEAD must occur while the runqueue locks are
2537 * still held, otherwise prev could be scheduled on another cpu, die
2538 * there before we look at prev->state, and then the reference would
2540 * Manfred Spraul <manfred@colorfullife.com>
2542 prev_state
= prev
->state
;
2543 finish_arch_switch(prev
);
2544 finish_lock_switch(rq
, prev
);
2546 if (current
->sched_class
->post_schedule
)
2547 current
->sched_class
->post_schedule(rq
);
2550 fire_sched_in_preempt_notifiers(current
);
2553 if (unlikely(prev_state
== TASK_DEAD
)) {
2555 * Remove function-return probe instances associated with this
2556 * task and put them back on the free list.
2558 kprobe_flush_task(prev
);
2559 put_task_struct(prev
);
2564 * schedule_tail - first thing a freshly forked thread must call.
2565 * @prev: the thread we just switched away from.
2567 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2568 __releases(rq
->lock
)
2570 struct rq
*rq
= this_rq();
2572 finish_task_switch(rq
, prev
);
2573 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2574 /* In this case, finish_task_switch does not reenable preemption */
2577 if (current
->set_child_tid
)
2578 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2582 * context_switch - switch to the new MM and the new
2583 * thread's register state.
2586 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2587 struct task_struct
*next
)
2589 struct mm_struct
*mm
, *oldmm
;
2591 prepare_task_switch(rq
, prev
, next
);
2592 trace_mark(kernel_sched_schedule
,
2593 "prev_pid %d next_pid %d prev_state %ld "
2594 "## rq %p prev %p next %p",
2595 prev
->pid
, next
->pid
, prev
->state
,
2598 oldmm
= prev
->active_mm
;
2600 * For paravirt, this is coupled with an exit in switch_to to
2601 * combine the page table reload and the switch backend into
2604 arch_enter_lazy_cpu_mode();
2606 if (unlikely(!mm
)) {
2607 next
->active_mm
= oldmm
;
2608 atomic_inc(&oldmm
->mm_count
);
2609 enter_lazy_tlb(oldmm
, next
);
2611 switch_mm(oldmm
, mm
, next
);
2613 if (unlikely(!prev
->mm
)) {
2614 prev
->active_mm
= NULL
;
2615 rq
->prev_mm
= oldmm
;
2618 * Since the runqueue lock will be released by the next
2619 * task (which is an invalid locking op but in the case
2620 * of the scheduler it's an obvious special-case), so we
2621 * do an early lockdep release here:
2623 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2624 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2627 /* Here we just switch the register state and the stack. */
2628 switch_to(prev
, next
, prev
);
2632 * this_rq must be evaluated again because prev may have moved
2633 * CPUs since it called schedule(), thus the 'rq' on its stack
2634 * frame will be invalid.
2636 finish_task_switch(this_rq(), prev
);
2640 * nr_running, nr_uninterruptible and nr_context_switches:
2642 * externally visible scheduler statistics: current number of runnable
2643 * threads, current number of uninterruptible-sleeping threads, total
2644 * number of context switches performed since bootup.
2646 unsigned long nr_running(void)
2648 unsigned long i
, sum
= 0;
2650 for_each_online_cpu(i
)
2651 sum
+= cpu_rq(i
)->nr_running
;
2656 unsigned long nr_uninterruptible(void)
2658 unsigned long i
, sum
= 0;
2660 for_each_possible_cpu(i
)
2661 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2664 * Since we read the counters lockless, it might be slightly
2665 * inaccurate. Do not allow it to go below zero though:
2667 if (unlikely((long)sum
< 0))
2673 unsigned long long nr_context_switches(void)
2676 unsigned long long sum
= 0;
2678 for_each_possible_cpu(i
)
2679 sum
+= cpu_rq(i
)->nr_switches
;
2684 unsigned long nr_iowait(void)
2686 unsigned long i
, sum
= 0;
2688 for_each_possible_cpu(i
)
2689 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2694 unsigned long nr_active(void)
2696 unsigned long i
, running
= 0, uninterruptible
= 0;
2698 for_each_online_cpu(i
) {
2699 running
+= cpu_rq(i
)->nr_running
;
2700 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2703 if (unlikely((long)uninterruptible
< 0))
2704 uninterruptible
= 0;
2706 return running
+ uninterruptible
;
2710 * Update rq->cpu_load[] statistics. This function is usually called every
2711 * scheduler tick (TICK_NSEC).
2713 static void update_cpu_load(struct rq
*this_rq
)
2715 unsigned long this_load
= this_rq
->load
.weight
;
2718 this_rq
->nr_load_updates
++;
2720 /* Update our load: */
2721 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2722 unsigned long old_load
, new_load
;
2724 /* scale is effectively 1 << i now, and >> i divides by scale */
2726 old_load
= this_rq
->cpu_load
[i
];
2727 new_load
= this_load
;
2729 * Round up the averaging division if load is increasing. This
2730 * prevents us from getting stuck on 9 if the load is 10, for
2733 if (new_load
> old_load
)
2734 new_load
+= scale
-1;
2735 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2742 * double_rq_lock - safely lock two runqueues
2744 * Note this does not disable interrupts like task_rq_lock,
2745 * you need to do so manually before calling.
2747 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2748 __acquires(rq1
->lock
)
2749 __acquires(rq2
->lock
)
2751 BUG_ON(!irqs_disabled());
2753 spin_lock(&rq1
->lock
);
2754 __acquire(rq2
->lock
); /* Fake it out ;) */
2757 spin_lock(&rq1
->lock
);
2758 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
2760 spin_lock(&rq2
->lock
);
2761 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
2764 update_rq_clock(rq1
);
2765 update_rq_clock(rq2
);
2769 * double_rq_unlock - safely unlock two runqueues
2771 * Note this does not restore interrupts like task_rq_unlock,
2772 * you need to do so manually after calling.
2774 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2775 __releases(rq1
->lock
)
2776 __releases(rq2
->lock
)
2778 spin_unlock(&rq1
->lock
);
2780 spin_unlock(&rq2
->lock
);
2782 __release(rq2
->lock
);
2786 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2788 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2789 __releases(this_rq
->lock
)
2790 __acquires(busiest
->lock
)
2791 __acquires(this_rq
->lock
)
2795 if (unlikely(!irqs_disabled())) {
2796 /* printk() doesn't work good under rq->lock */
2797 spin_unlock(&this_rq
->lock
);
2800 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2801 if (busiest
< this_rq
) {
2802 spin_unlock(&this_rq
->lock
);
2803 spin_lock(&busiest
->lock
);
2804 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
2807 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
2812 static void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2813 __releases(busiest
->lock
)
2815 spin_unlock(&busiest
->lock
);
2816 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
2820 * If dest_cpu is allowed for this process, migrate the task to it.
2821 * This is accomplished by forcing the cpu_allowed mask to only
2822 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2823 * the cpu_allowed mask is restored.
2825 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2827 struct migration_req req
;
2828 unsigned long flags
;
2831 rq
= task_rq_lock(p
, &flags
);
2832 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2833 || unlikely(!cpu_active(dest_cpu
)))
2836 /* force the process onto the specified CPU */
2837 if (migrate_task(p
, dest_cpu
, &req
)) {
2838 /* Need to wait for migration thread (might exit: take ref). */
2839 struct task_struct
*mt
= rq
->migration_thread
;
2841 get_task_struct(mt
);
2842 task_rq_unlock(rq
, &flags
);
2843 wake_up_process(mt
);
2844 put_task_struct(mt
);
2845 wait_for_completion(&req
.done
);
2850 task_rq_unlock(rq
, &flags
);
2854 * sched_exec - execve() is a valuable balancing opportunity, because at
2855 * this point the task has the smallest effective memory and cache footprint.
2857 void sched_exec(void)
2859 int new_cpu
, this_cpu
= get_cpu();
2860 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2862 if (new_cpu
!= this_cpu
)
2863 sched_migrate_task(current
, new_cpu
);
2867 * pull_task - move a task from a remote runqueue to the local runqueue.
2868 * Both runqueues must be locked.
2870 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2871 struct rq
*this_rq
, int this_cpu
)
2873 deactivate_task(src_rq
, p
, 0);
2874 set_task_cpu(p
, this_cpu
);
2875 activate_task(this_rq
, p
, 0);
2877 * Note that idle threads have a prio of MAX_PRIO, for this test
2878 * to be always true for them.
2880 check_preempt_curr(this_rq
, p
, 0);
2884 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2887 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2888 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2892 * We do not migrate tasks that are:
2893 * 1) running (obviously), or
2894 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2895 * 3) are cache-hot on their current CPU.
2897 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2898 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2903 if (task_running(rq
, p
)) {
2904 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2909 * Aggressive migration if:
2910 * 1) task is cache cold, or
2911 * 2) too many balance attempts have failed.
2914 if (!task_hot(p
, rq
->clock
, sd
) ||
2915 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2916 #ifdef CONFIG_SCHEDSTATS
2917 if (task_hot(p
, rq
->clock
, sd
)) {
2918 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2919 schedstat_inc(p
, se
.nr_forced_migrations
);
2925 if (task_hot(p
, rq
->clock
, sd
)) {
2926 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2932 static unsigned long
2933 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2934 unsigned long max_load_move
, struct sched_domain
*sd
,
2935 enum cpu_idle_type idle
, int *all_pinned
,
2936 int *this_best_prio
, struct rq_iterator
*iterator
)
2938 int loops
= 0, pulled
= 0, pinned
= 0;
2939 struct task_struct
*p
;
2940 long rem_load_move
= max_load_move
;
2942 if (max_load_move
== 0)
2948 * Start the load-balancing iterator:
2950 p
= iterator
->start(iterator
->arg
);
2952 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2955 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
2956 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2957 p
= iterator
->next(iterator
->arg
);
2961 pull_task(busiest
, p
, this_rq
, this_cpu
);
2963 rem_load_move
-= p
->se
.load
.weight
;
2966 * We only want to steal up to the prescribed amount of weighted load.
2968 if (rem_load_move
> 0) {
2969 if (p
->prio
< *this_best_prio
)
2970 *this_best_prio
= p
->prio
;
2971 p
= iterator
->next(iterator
->arg
);
2976 * Right now, this is one of only two places pull_task() is called,
2977 * so we can safely collect pull_task() stats here rather than
2978 * inside pull_task().
2980 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2983 *all_pinned
= pinned
;
2985 return max_load_move
- rem_load_move
;
2989 * move_tasks tries to move up to max_load_move weighted load from busiest to
2990 * this_rq, as part of a balancing operation within domain "sd".
2991 * Returns 1 if successful and 0 otherwise.
2993 * Called with both runqueues locked.
2995 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2996 unsigned long max_load_move
,
2997 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3000 const struct sched_class
*class = sched_class_highest
;
3001 unsigned long total_load_moved
= 0;
3002 int this_best_prio
= this_rq
->curr
->prio
;
3006 class->load_balance(this_rq
, this_cpu
, busiest
,
3007 max_load_move
- total_load_moved
,
3008 sd
, idle
, all_pinned
, &this_best_prio
);
3009 class = class->next
;
3011 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3014 } while (class && max_load_move
> total_load_moved
);
3016 return total_load_moved
> 0;
3020 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3021 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3022 struct rq_iterator
*iterator
)
3024 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3028 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3029 pull_task(busiest
, p
, this_rq
, this_cpu
);
3031 * Right now, this is only the second place pull_task()
3032 * is called, so we can safely collect pull_task()
3033 * stats here rather than inside pull_task().
3035 schedstat_inc(sd
, lb_gained
[idle
]);
3039 p
= iterator
->next(iterator
->arg
);
3046 * move_one_task tries to move exactly one task from busiest to this_rq, as
3047 * part of active balancing operations within "domain".
3048 * Returns 1 if successful and 0 otherwise.
3050 * Called with both runqueues locked.
3052 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3053 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3055 const struct sched_class
*class;
3057 for (class = sched_class_highest
; class; class = class->next
)
3058 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3065 * find_busiest_group finds and returns the busiest CPU group within the
3066 * domain. It calculates and returns the amount of weighted load which
3067 * should be moved to restore balance via the imbalance parameter.
3069 static struct sched_group
*
3070 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3071 unsigned long *imbalance
, enum cpu_idle_type idle
,
3072 int *sd_idle
, const cpumask_t
*cpus
, int *balance
)
3074 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
3075 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
3076 unsigned long max_pull
;
3077 unsigned long busiest_load_per_task
, busiest_nr_running
;
3078 unsigned long this_load_per_task
, this_nr_running
;
3079 int load_idx
, group_imb
= 0;
3080 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3081 int power_savings_balance
= 1;
3082 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
3083 unsigned long min_nr_running
= ULONG_MAX
;
3084 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
3087 max_load
= this_load
= total_load
= total_pwr
= 0;
3088 busiest_load_per_task
= busiest_nr_running
= 0;
3089 this_load_per_task
= this_nr_running
= 0;
3091 if (idle
== CPU_NOT_IDLE
)
3092 load_idx
= sd
->busy_idx
;
3093 else if (idle
== CPU_NEWLY_IDLE
)
3094 load_idx
= sd
->newidle_idx
;
3096 load_idx
= sd
->idle_idx
;
3099 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
3102 int __group_imb
= 0;
3103 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3104 unsigned long sum_nr_running
, sum_weighted_load
;
3105 unsigned long sum_avg_load_per_task
;
3106 unsigned long avg_load_per_task
;
3108 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
3111 balance_cpu
= first_cpu(group
->cpumask
);
3113 /* Tally up the load of all CPUs in the group */
3114 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
3115 sum_avg_load_per_task
= avg_load_per_task
= 0;
3118 min_cpu_load
= ~0UL;
3120 for_each_cpu_mask_nr(i
, group
->cpumask
) {
3123 if (!cpu_isset(i
, *cpus
))
3128 if (*sd_idle
&& rq
->nr_running
)
3131 /* Bias balancing toward cpus of our domain */
3133 if (idle_cpu(i
) && !first_idle_cpu
) {
3138 load
= target_load(i
, load_idx
);
3140 load
= source_load(i
, load_idx
);
3141 if (load
> max_cpu_load
)
3142 max_cpu_load
= load
;
3143 if (min_cpu_load
> load
)
3144 min_cpu_load
= load
;
3148 sum_nr_running
+= rq
->nr_running
;
3149 sum_weighted_load
+= weighted_cpuload(i
);
3151 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3155 * First idle cpu or the first cpu(busiest) in this sched group
3156 * is eligible for doing load balancing at this and above
3157 * domains. In the newly idle case, we will allow all the cpu's
3158 * to do the newly idle load balance.
3160 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3161 balance_cpu
!= this_cpu
&& balance
) {
3166 total_load
+= avg_load
;
3167 total_pwr
+= group
->__cpu_power
;
3169 /* Adjust by relative CPU power of the group */
3170 avg_load
= sg_div_cpu_power(group
,
3171 avg_load
* SCHED_LOAD_SCALE
);
3175 * Consider the group unbalanced when the imbalance is larger
3176 * than the average weight of two tasks.
3178 * APZ: with cgroup the avg task weight can vary wildly and
3179 * might not be a suitable number - should we keep a
3180 * normalized nr_running number somewhere that negates
3183 avg_load_per_task
= sg_div_cpu_power(group
,
3184 sum_avg_load_per_task
* SCHED_LOAD_SCALE
);
3186 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3189 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3192 this_load
= avg_load
;
3194 this_nr_running
= sum_nr_running
;
3195 this_load_per_task
= sum_weighted_load
;
3196 } else if (avg_load
> max_load
&&
3197 (sum_nr_running
> group_capacity
|| __group_imb
)) {
3198 max_load
= avg_load
;
3200 busiest_nr_running
= sum_nr_running
;
3201 busiest_load_per_task
= sum_weighted_load
;
3202 group_imb
= __group_imb
;
3205 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3207 * Busy processors will not participate in power savings
3210 if (idle
== CPU_NOT_IDLE
||
3211 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3215 * If the local group is idle or completely loaded
3216 * no need to do power savings balance at this domain
3218 if (local_group
&& (this_nr_running
>= group_capacity
||
3220 power_savings_balance
= 0;
3223 * If a group is already running at full capacity or idle,
3224 * don't include that group in power savings calculations
3226 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
3231 * Calculate the group which has the least non-idle load.
3232 * This is the group from where we need to pick up the load
3235 if ((sum_nr_running
< min_nr_running
) ||
3236 (sum_nr_running
== min_nr_running
&&
3237 first_cpu(group
->cpumask
) <
3238 first_cpu(group_min
->cpumask
))) {
3240 min_nr_running
= sum_nr_running
;
3241 min_load_per_task
= sum_weighted_load
/
3246 * Calculate the group which is almost near its
3247 * capacity but still has some space to pick up some load
3248 * from other group and save more power
3250 if (sum_nr_running
<= group_capacity
- 1) {
3251 if (sum_nr_running
> leader_nr_running
||
3252 (sum_nr_running
== leader_nr_running
&&
3253 first_cpu(group
->cpumask
) >
3254 first_cpu(group_leader
->cpumask
))) {
3255 group_leader
= group
;
3256 leader_nr_running
= sum_nr_running
;
3261 group
= group
->next
;
3262 } while (group
!= sd
->groups
);
3264 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
3267 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
3269 if (this_load
>= avg_load
||
3270 100*max_load
<= sd
->imbalance_pct
*this_load
)
3273 busiest_load_per_task
/= busiest_nr_running
;
3275 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
3278 * We're trying to get all the cpus to the average_load, so we don't
3279 * want to push ourselves above the average load, nor do we wish to
3280 * reduce the max loaded cpu below the average load, as either of these
3281 * actions would just result in more rebalancing later, and ping-pong
3282 * tasks around. Thus we look for the minimum possible imbalance.
3283 * Negative imbalances (*we* are more loaded than anyone else) will
3284 * be counted as no imbalance for these purposes -- we can't fix that
3285 * by pulling tasks to us. Be careful of negative numbers as they'll
3286 * appear as very large values with unsigned longs.
3288 if (max_load
<= busiest_load_per_task
)
3292 * In the presence of smp nice balancing, certain scenarios can have
3293 * max load less than avg load(as we skip the groups at or below
3294 * its cpu_power, while calculating max_load..)
3296 if (max_load
< avg_load
) {
3298 goto small_imbalance
;
3301 /* Don't want to pull so many tasks that a group would go idle */
3302 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
3304 /* How much load to actually move to equalise the imbalance */
3305 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
3306 (avg_load
- this_load
) * this->__cpu_power
)
3310 * if *imbalance is less than the average load per runnable task
3311 * there is no gaurantee that any tasks will be moved so we'll have
3312 * a think about bumping its value to force at least one task to be
3315 if (*imbalance
< busiest_load_per_task
) {
3316 unsigned long tmp
, pwr_now
, pwr_move
;
3320 pwr_move
= pwr_now
= 0;
3322 if (this_nr_running
) {
3323 this_load_per_task
/= this_nr_running
;
3324 if (busiest_load_per_task
> this_load_per_task
)
3327 this_load_per_task
= cpu_avg_load_per_task(this_cpu
);
3329 if (max_load
- this_load
+ 2*busiest_load_per_task
>=
3330 busiest_load_per_task
* imbn
) {
3331 *imbalance
= busiest_load_per_task
;
3336 * OK, we don't have enough imbalance to justify moving tasks,
3337 * however we may be able to increase total CPU power used by
3341 pwr_now
+= busiest
->__cpu_power
*
3342 min(busiest_load_per_task
, max_load
);
3343 pwr_now
+= this->__cpu_power
*
3344 min(this_load_per_task
, this_load
);
3345 pwr_now
/= SCHED_LOAD_SCALE
;
3347 /* Amount of load we'd subtract */
3348 tmp
= sg_div_cpu_power(busiest
,
3349 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3351 pwr_move
+= busiest
->__cpu_power
*
3352 min(busiest_load_per_task
, max_load
- tmp
);
3354 /* Amount of load we'd add */
3355 if (max_load
* busiest
->__cpu_power
<
3356 busiest_load_per_task
* SCHED_LOAD_SCALE
)
3357 tmp
= sg_div_cpu_power(this,
3358 max_load
* busiest
->__cpu_power
);
3360 tmp
= sg_div_cpu_power(this,
3361 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3362 pwr_move
+= this->__cpu_power
*
3363 min(this_load_per_task
, this_load
+ tmp
);
3364 pwr_move
/= SCHED_LOAD_SCALE
;
3366 /* Move if we gain throughput */
3367 if (pwr_move
> pwr_now
)
3368 *imbalance
= busiest_load_per_task
;
3374 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3375 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3378 if (this == group_leader
&& group_leader
!= group_min
) {
3379 *imbalance
= min_load_per_task
;
3389 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3392 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3393 unsigned long imbalance
, const cpumask_t
*cpus
)
3395 struct rq
*busiest
= NULL
, *rq
;
3396 unsigned long max_load
= 0;
3399 for_each_cpu_mask_nr(i
, group
->cpumask
) {
3402 if (!cpu_isset(i
, *cpus
))
3406 wl
= weighted_cpuload(i
);
3408 if (rq
->nr_running
== 1 && wl
> imbalance
)
3411 if (wl
> max_load
) {
3421 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3422 * so long as it is large enough.
3424 #define MAX_PINNED_INTERVAL 512
3427 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3428 * tasks if there is an imbalance.
3430 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3431 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3432 int *balance
, cpumask_t
*cpus
)
3434 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3435 struct sched_group
*group
;
3436 unsigned long imbalance
;
3438 unsigned long flags
;
3443 * When power savings policy is enabled for the parent domain, idle
3444 * sibling can pick up load irrespective of busy siblings. In this case,
3445 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3446 * portraying it as CPU_NOT_IDLE.
3448 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3449 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3452 schedstat_inc(sd
, lb_count
[idle
]);
3456 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3463 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3467 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3469 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3473 BUG_ON(busiest
== this_rq
);
3475 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3478 if (busiest
->nr_running
> 1) {
3480 * Attempt to move tasks. If find_busiest_group has found
3481 * an imbalance but busiest->nr_running <= 1, the group is
3482 * still unbalanced. ld_moved simply stays zero, so it is
3483 * correctly treated as an imbalance.
3485 local_irq_save(flags
);
3486 double_rq_lock(this_rq
, busiest
);
3487 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3488 imbalance
, sd
, idle
, &all_pinned
);
3489 double_rq_unlock(this_rq
, busiest
);
3490 local_irq_restore(flags
);
3493 * some other cpu did the load balance for us.
3495 if (ld_moved
&& this_cpu
!= smp_processor_id())
3496 resched_cpu(this_cpu
);
3498 /* All tasks on this runqueue were pinned by CPU affinity */
3499 if (unlikely(all_pinned
)) {
3500 cpu_clear(cpu_of(busiest
), *cpus
);
3501 if (!cpus_empty(*cpus
))
3508 schedstat_inc(sd
, lb_failed
[idle
]);
3509 sd
->nr_balance_failed
++;
3511 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3513 spin_lock_irqsave(&busiest
->lock
, flags
);
3515 /* don't kick the migration_thread, if the curr
3516 * task on busiest cpu can't be moved to this_cpu
3518 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
3519 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3521 goto out_one_pinned
;
3524 if (!busiest
->active_balance
) {
3525 busiest
->active_balance
= 1;
3526 busiest
->push_cpu
= this_cpu
;
3529 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3531 wake_up_process(busiest
->migration_thread
);
3534 * We've kicked active balancing, reset the failure
3537 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3540 sd
->nr_balance_failed
= 0;
3542 if (likely(!active_balance
)) {
3543 /* We were unbalanced, so reset the balancing interval */
3544 sd
->balance_interval
= sd
->min_interval
;
3547 * If we've begun active balancing, start to back off. This
3548 * case may not be covered by the all_pinned logic if there
3549 * is only 1 task on the busy runqueue (because we don't call
3552 if (sd
->balance_interval
< sd
->max_interval
)
3553 sd
->balance_interval
*= 2;
3556 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3557 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3563 schedstat_inc(sd
, lb_balanced
[idle
]);
3565 sd
->nr_balance_failed
= 0;
3568 /* tune up the balancing interval */
3569 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3570 (sd
->balance_interval
< sd
->max_interval
))
3571 sd
->balance_interval
*= 2;
3573 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3574 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3585 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3586 * tasks if there is an imbalance.
3588 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3589 * this_rq is locked.
3592 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
,
3595 struct sched_group
*group
;
3596 struct rq
*busiest
= NULL
;
3597 unsigned long imbalance
;
3605 * When power savings policy is enabled for the parent domain, idle
3606 * sibling can pick up load irrespective of busy siblings. In this case,
3607 * let the state of idle sibling percolate up as IDLE, instead of
3608 * portraying it as CPU_NOT_IDLE.
3610 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3611 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3614 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3616 update_shares_locked(this_rq
, sd
);
3617 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3618 &sd_idle
, cpus
, NULL
);
3620 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3624 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
3626 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3630 BUG_ON(busiest
== this_rq
);
3632 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3635 if (busiest
->nr_running
> 1) {
3636 /* Attempt to move tasks */
3637 double_lock_balance(this_rq
, busiest
);
3638 /* this_rq->clock is already updated */
3639 update_rq_clock(busiest
);
3640 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3641 imbalance
, sd
, CPU_NEWLY_IDLE
,
3643 double_unlock_balance(this_rq
, busiest
);
3645 if (unlikely(all_pinned
)) {
3646 cpu_clear(cpu_of(busiest
), *cpus
);
3647 if (!cpus_empty(*cpus
))
3653 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3654 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3655 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3658 sd
->nr_balance_failed
= 0;
3660 update_shares_locked(this_rq
, sd
);
3664 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3665 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3666 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3668 sd
->nr_balance_failed
= 0;
3674 * idle_balance is called by schedule() if this_cpu is about to become
3675 * idle. Attempts to pull tasks from other CPUs.
3677 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3679 struct sched_domain
*sd
;
3680 int pulled_task
= -1;
3681 unsigned long next_balance
= jiffies
+ HZ
;
3684 for_each_domain(this_cpu
, sd
) {
3685 unsigned long interval
;
3687 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3690 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3691 /* If we've pulled tasks over stop searching: */
3692 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
3695 interval
= msecs_to_jiffies(sd
->balance_interval
);
3696 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3697 next_balance
= sd
->last_balance
+ interval
;
3701 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3703 * We are going idle. next_balance may be set based on
3704 * a busy processor. So reset next_balance.
3706 this_rq
->next_balance
= next_balance
;
3711 * active_load_balance is run by migration threads. It pushes running tasks
3712 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3713 * running on each physical CPU where possible, and avoids physical /
3714 * logical imbalances.
3716 * Called with busiest_rq locked.
3718 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3720 int target_cpu
= busiest_rq
->push_cpu
;
3721 struct sched_domain
*sd
;
3722 struct rq
*target_rq
;
3724 /* Is there any task to move? */
3725 if (busiest_rq
->nr_running
<= 1)
3728 target_rq
= cpu_rq(target_cpu
);
3731 * This condition is "impossible", if it occurs
3732 * we need to fix it. Originally reported by
3733 * Bjorn Helgaas on a 128-cpu setup.
3735 BUG_ON(busiest_rq
== target_rq
);
3737 /* move a task from busiest_rq to target_rq */
3738 double_lock_balance(busiest_rq
, target_rq
);
3739 update_rq_clock(busiest_rq
);
3740 update_rq_clock(target_rq
);
3742 /* Search for an sd spanning us and the target CPU. */
3743 for_each_domain(target_cpu
, sd
) {
3744 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3745 cpu_isset(busiest_cpu
, sd
->span
))
3750 schedstat_inc(sd
, alb_count
);
3752 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3754 schedstat_inc(sd
, alb_pushed
);
3756 schedstat_inc(sd
, alb_failed
);
3758 double_unlock_balance(busiest_rq
, target_rq
);
3763 atomic_t load_balancer
;
3765 } nohz ____cacheline_aligned
= {
3766 .load_balancer
= ATOMIC_INIT(-1),
3767 .cpu_mask
= CPU_MASK_NONE
,
3771 * This routine will try to nominate the ilb (idle load balancing)
3772 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3773 * load balancing on behalf of all those cpus. If all the cpus in the system
3774 * go into this tickless mode, then there will be no ilb owner (as there is
3775 * no need for one) and all the cpus will sleep till the next wakeup event
3778 * For the ilb owner, tick is not stopped. And this tick will be used
3779 * for idle load balancing. ilb owner will still be part of
3782 * While stopping the tick, this cpu will become the ilb owner if there
3783 * is no other owner. And will be the owner till that cpu becomes busy
3784 * or if all cpus in the system stop their ticks at which point
3785 * there is no need for ilb owner.
3787 * When the ilb owner becomes busy, it nominates another owner, during the
3788 * next busy scheduler_tick()
3790 int select_nohz_load_balancer(int stop_tick
)
3792 int cpu
= smp_processor_id();
3795 cpu_set(cpu
, nohz
.cpu_mask
);
3796 cpu_rq(cpu
)->in_nohz_recently
= 1;
3799 * If we are going offline and still the leader, give up!
3801 if (!cpu_active(cpu
) &&
3802 atomic_read(&nohz
.load_balancer
) == cpu
) {
3803 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3808 /* time for ilb owner also to sleep */
3809 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3810 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3811 atomic_set(&nohz
.load_balancer
, -1);
3815 if (atomic_read(&nohz
.load_balancer
) == -1) {
3816 /* make me the ilb owner */
3817 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3819 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3822 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3825 cpu_clear(cpu
, nohz
.cpu_mask
);
3827 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3828 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3835 static DEFINE_SPINLOCK(balancing
);
3838 * It checks each scheduling domain to see if it is due to be balanced,
3839 * and initiates a balancing operation if so.
3841 * Balancing parameters are set up in arch_init_sched_domains.
3843 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3846 struct rq
*rq
= cpu_rq(cpu
);
3847 unsigned long interval
;
3848 struct sched_domain
*sd
;
3849 /* Earliest time when we have to do rebalance again */
3850 unsigned long next_balance
= jiffies
+ 60*HZ
;
3851 int update_next_balance
= 0;
3855 for_each_domain(cpu
, sd
) {
3856 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3859 interval
= sd
->balance_interval
;
3860 if (idle
!= CPU_IDLE
)
3861 interval
*= sd
->busy_factor
;
3863 /* scale ms to jiffies */
3864 interval
= msecs_to_jiffies(interval
);
3865 if (unlikely(!interval
))
3867 if (interval
> HZ
*NR_CPUS
/10)
3868 interval
= HZ
*NR_CPUS
/10;
3870 need_serialize
= sd
->flags
& SD_SERIALIZE
;
3872 if (need_serialize
) {
3873 if (!spin_trylock(&balancing
))
3877 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3878 if (load_balance(cpu
, rq
, sd
, idle
, &balance
, &tmp
)) {
3880 * We've pulled tasks over so either we're no
3881 * longer idle, or one of our SMT siblings is
3884 idle
= CPU_NOT_IDLE
;
3886 sd
->last_balance
= jiffies
;
3889 spin_unlock(&balancing
);
3891 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3892 next_balance
= sd
->last_balance
+ interval
;
3893 update_next_balance
= 1;
3897 * Stop the load balance at this level. There is another
3898 * CPU in our sched group which is doing load balancing more
3906 * next_balance will be updated only when there is a need.
3907 * When the cpu is attached to null domain for ex, it will not be
3910 if (likely(update_next_balance
))
3911 rq
->next_balance
= next_balance
;
3915 * run_rebalance_domains is triggered when needed from the scheduler tick.
3916 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3917 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3919 static void run_rebalance_domains(struct softirq_action
*h
)
3921 int this_cpu
= smp_processor_id();
3922 struct rq
*this_rq
= cpu_rq(this_cpu
);
3923 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3924 CPU_IDLE
: CPU_NOT_IDLE
;
3926 rebalance_domains(this_cpu
, idle
);
3930 * If this cpu is the owner for idle load balancing, then do the
3931 * balancing on behalf of the other idle cpus whose ticks are
3934 if (this_rq
->idle_at_tick
&&
3935 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3936 cpumask_t cpus
= nohz
.cpu_mask
;
3940 cpu_clear(this_cpu
, cpus
);
3941 for_each_cpu_mask_nr(balance_cpu
, cpus
) {
3943 * If this cpu gets work to do, stop the load balancing
3944 * work being done for other cpus. Next load
3945 * balancing owner will pick it up.
3950 rebalance_domains(balance_cpu
, CPU_IDLE
);
3952 rq
= cpu_rq(balance_cpu
);
3953 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3954 this_rq
->next_balance
= rq
->next_balance
;
3961 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3963 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3964 * idle load balancing owner or decide to stop the periodic load balancing,
3965 * if the whole system is idle.
3967 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3971 * If we were in the nohz mode recently and busy at the current
3972 * scheduler tick, then check if we need to nominate new idle
3975 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3976 rq
->in_nohz_recently
= 0;
3978 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3979 cpu_clear(cpu
, nohz
.cpu_mask
);
3980 atomic_set(&nohz
.load_balancer
, -1);
3983 if (atomic_read(&nohz
.load_balancer
) == -1) {
3985 * simple selection for now: Nominate the
3986 * first cpu in the nohz list to be the next
3989 * TBD: Traverse the sched domains and nominate
3990 * the nearest cpu in the nohz.cpu_mask.
3992 int ilb
= first_cpu(nohz
.cpu_mask
);
3994 if (ilb
< nr_cpu_ids
)
4000 * If this cpu is idle and doing idle load balancing for all the
4001 * cpus with ticks stopped, is it time for that to stop?
4003 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4004 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4010 * If this cpu is idle and the idle load balancing is done by
4011 * someone else, then no need raise the SCHED_SOFTIRQ
4013 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4014 cpu_isset(cpu
, nohz
.cpu_mask
))
4017 if (time_after_eq(jiffies
, rq
->next_balance
))
4018 raise_softirq(SCHED_SOFTIRQ
);
4021 #else /* CONFIG_SMP */
4024 * on UP we do not need to balance between CPUs:
4026 static inline void idle_balance(int cpu
, struct rq
*rq
)
4032 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4034 EXPORT_PER_CPU_SYMBOL(kstat
);
4037 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4038 * that have not yet been banked in case the task is currently running.
4040 unsigned long long task_sched_runtime(struct task_struct
*p
)
4042 unsigned long flags
;
4046 rq
= task_rq_lock(p
, &flags
);
4047 ns
= p
->se
.sum_exec_runtime
;
4048 if (task_current(rq
, p
)) {
4049 update_rq_clock(rq
);
4050 delta_exec
= rq
->clock
- p
->se
.exec_start
;
4051 if ((s64
)delta_exec
> 0)
4054 task_rq_unlock(rq
, &flags
);
4060 * Account user cpu time to a process.
4061 * @p: the process that the cpu time gets accounted to
4062 * @cputime: the cpu time spent in user space since the last update
4064 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
4066 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4069 p
->utime
= cputime_add(p
->utime
, cputime
);
4071 /* Add user time to cpustat. */
4072 tmp
= cputime_to_cputime64(cputime
);
4073 if (TASK_NICE(p
) > 0)
4074 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4076 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4077 /* Account for user time used */
4078 acct_update_integrals(p
);
4082 * Account guest cpu time to a process.
4083 * @p: the process that the cpu time gets accounted to
4084 * @cputime: the cpu time spent in virtual machine since the last update
4086 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
4089 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4091 tmp
= cputime_to_cputime64(cputime
);
4093 p
->utime
= cputime_add(p
->utime
, cputime
);
4094 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4096 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4097 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4101 * Account scaled user cpu time to a process.
4102 * @p: the process that the cpu time gets accounted to
4103 * @cputime: the cpu time spent in user space since the last update
4105 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4107 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
4111 * Account system cpu time to a process.
4112 * @p: the process that the cpu time gets accounted to
4113 * @hardirq_offset: the offset to subtract from hardirq_count()
4114 * @cputime: the cpu time spent in kernel space since the last update
4116 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4119 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4120 struct rq
*rq
= this_rq();
4123 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4124 account_guest_time(p
, cputime
);
4128 p
->stime
= cputime_add(p
->stime
, cputime
);
4130 /* Add system time to cpustat. */
4131 tmp
= cputime_to_cputime64(cputime
);
4132 if (hardirq_count() - hardirq_offset
)
4133 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4134 else if (softirq_count())
4135 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4136 else if (p
!= rq
->idle
)
4137 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4138 else if (atomic_read(&rq
->nr_iowait
) > 0)
4139 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4141 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4142 /* Account for system time used */
4143 acct_update_integrals(p
);
4147 * Account scaled system cpu time to a process.
4148 * @p: the process that the cpu time gets accounted to
4149 * @hardirq_offset: the offset to subtract from hardirq_count()
4150 * @cputime: the cpu time spent in kernel space since the last update
4152 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4154 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
4158 * Account for involuntary wait time.
4159 * @p: the process from which the cpu time has been stolen
4160 * @steal: the cpu time spent in involuntary wait
4162 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
4164 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4165 cputime64_t tmp
= cputime_to_cputime64(steal
);
4166 struct rq
*rq
= this_rq();
4168 if (p
== rq
->idle
) {
4169 p
->stime
= cputime_add(p
->stime
, steal
);
4170 if (atomic_read(&rq
->nr_iowait
) > 0)
4171 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4173 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4175 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
4179 * Use precise platform statistics if available:
4181 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4182 cputime_t
task_utime(struct task_struct
*p
)
4187 cputime_t
task_stime(struct task_struct
*p
)
4192 cputime_t
task_utime(struct task_struct
*p
)
4194 clock_t utime
= cputime_to_clock_t(p
->utime
),
4195 total
= utime
+ cputime_to_clock_t(p
->stime
);
4199 * Use CFS's precise accounting:
4201 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
4205 do_div(temp
, total
);
4207 utime
= (clock_t)temp
;
4209 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
4210 return p
->prev_utime
;
4213 cputime_t
task_stime(struct task_struct
*p
)
4218 * Use CFS's precise accounting. (we subtract utime from
4219 * the total, to make sure the total observed by userspace
4220 * grows monotonically - apps rely on that):
4222 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
4223 cputime_to_clock_t(task_utime(p
));
4226 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
4228 return p
->prev_stime
;
4232 inline cputime_t
task_gtime(struct task_struct
*p
)
4238 * This function gets called by the timer code, with HZ frequency.
4239 * We call it with interrupts disabled.
4241 * It also gets called by the fork code, when changing the parent's
4244 void scheduler_tick(void)
4246 int cpu
= smp_processor_id();
4247 struct rq
*rq
= cpu_rq(cpu
);
4248 struct task_struct
*curr
= rq
->curr
;
4252 spin_lock(&rq
->lock
);
4253 update_rq_clock(rq
);
4254 update_cpu_load(rq
);
4255 curr
->sched_class
->task_tick(rq
, curr
, 0);
4256 spin_unlock(&rq
->lock
);
4259 rq
->idle_at_tick
= idle_cpu(cpu
);
4260 trigger_load_balance(rq
, cpu
);
4264 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4265 defined(CONFIG_PREEMPT_TRACER))
4267 static inline unsigned long get_parent_ip(unsigned long addr
)
4269 if (in_lock_functions(addr
)) {
4270 addr
= CALLER_ADDR2
;
4271 if (in_lock_functions(addr
))
4272 addr
= CALLER_ADDR3
;
4277 void __kprobes
add_preempt_count(int val
)
4279 #ifdef CONFIG_DEBUG_PREEMPT
4283 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4286 preempt_count() += val
;
4287 #ifdef CONFIG_DEBUG_PREEMPT
4289 * Spinlock count overflowing soon?
4291 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4294 if (preempt_count() == val
)
4295 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4297 EXPORT_SYMBOL(add_preempt_count
);
4299 void __kprobes
sub_preempt_count(int val
)
4301 #ifdef CONFIG_DEBUG_PREEMPT
4305 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4308 * Is the spinlock portion underflowing?
4310 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4311 !(preempt_count() & PREEMPT_MASK
)))
4315 if (preempt_count() == val
)
4316 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4317 preempt_count() -= val
;
4319 EXPORT_SYMBOL(sub_preempt_count
);
4324 * Print scheduling while atomic bug:
4326 static noinline
void __schedule_bug(struct task_struct
*prev
)
4328 struct pt_regs
*regs
= get_irq_regs();
4330 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4331 prev
->comm
, prev
->pid
, preempt_count());
4333 debug_show_held_locks(prev
);
4335 if (irqs_disabled())
4336 print_irqtrace_events(prev
);
4345 * Various schedule()-time debugging checks and statistics:
4347 static inline void schedule_debug(struct task_struct
*prev
)
4350 * Test if we are atomic. Since do_exit() needs to call into
4351 * schedule() atomically, we ignore that path for now.
4352 * Otherwise, whine if we are scheduling when we should not be.
4354 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4355 __schedule_bug(prev
);
4357 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4359 schedstat_inc(this_rq(), sched_count
);
4360 #ifdef CONFIG_SCHEDSTATS
4361 if (unlikely(prev
->lock_depth
>= 0)) {
4362 schedstat_inc(this_rq(), bkl_count
);
4363 schedstat_inc(prev
, sched_info
.bkl_count
);
4369 * Pick up the highest-prio task:
4371 static inline struct task_struct
*
4372 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
4374 const struct sched_class
*class;
4375 struct task_struct
*p
;
4378 * Optimization: we know that if all tasks are in
4379 * the fair class we can call that function directly:
4381 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4382 p
= fair_sched_class
.pick_next_task(rq
);
4387 class = sched_class_highest
;
4389 p
= class->pick_next_task(rq
);
4393 * Will never be NULL as the idle class always
4394 * returns a non-NULL p:
4396 class = class->next
;
4401 * schedule() is the main scheduler function.
4403 asmlinkage
void __sched
schedule(void)
4405 struct task_struct
*prev
, *next
;
4406 unsigned long *switch_count
;
4412 cpu
= smp_processor_id();
4416 switch_count
= &prev
->nivcsw
;
4418 release_kernel_lock(prev
);
4419 need_resched_nonpreemptible
:
4421 schedule_debug(prev
);
4423 if (sched_feat(HRTICK
))
4427 * Do the rq-clock update outside the rq lock:
4429 local_irq_disable();
4430 update_rq_clock(rq
);
4431 spin_lock(&rq
->lock
);
4432 clear_tsk_need_resched(prev
);
4434 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4435 if (unlikely(signal_pending_state(prev
->state
, prev
)))
4436 prev
->state
= TASK_RUNNING
;
4438 deactivate_task(rq
, prev
, 1);
4439 switch_count
= &prev
->nvcsw
;
4443 if (prev
->sched_class
->pre_schedule
)
4444 prev
->sched_class
->pre_schedule(rq
, prev
);
4447 if (unlikely(!rq
->nr_running
))
4448 idle_balance(cpu
, rq
);
4450 prev
->sched_class
->put_prev_task(rq
, prev
);
4451 next
= pick_next_task(rq
, prev
);
4453 if (likely(prev
!= next
)) {
4454 sched_info_switch(prev
, next
);
4460 context_switch(rq
, prev
, next
); /* unlocks the rq */
4462 * the context switch might have flipped the stack from under
4463 * us, hence refresh the local variables.
4465 cpu
= smp_processor_id();
4468 spin_unlock_irq(&rq
->lock
);
4470 if (unlikely(reacquire_kernel_lock(current
) < 0))
4471 goto need_resched_nonpreemptible
;
4473 preempt_enable_no_resched();
4474 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
4477 EXPORT_SYMBOL(schedule
);
4479 #ifdef CONFIG_PREEMPT
4481 * this is the entry point to schedule() from in-kernel preemption
4482 * off of preempt_enable. Kernel preemptions off return from interrupt
4483 * occur there and call schedule directly.
4485 asmlinkage
void __sched
preempt_schedule(void)
4487 struct thread_info
*ti
= current_thread_info();
4490 * If there is a non-zero preempt_count or interrupts are disabled,
4491 * we do not want to preempt the current task. Just return..
4493 if (likely(ti
->preempt_count
|| irqs_disabled()))
4497 add_preempt_count(PREEMPT_ACTIVE
);
4499 sub_preempt_count(PREEMPT_ACTIVE
);
4502 * Check again in case we missed a preemption opportunity
4503 * between schedule and now.
4506 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4508 EXPORT_SYMBOL(preempt_schedule
);
4511 * this is the entry point to schedule() from kernel preemption
4512 * off of irq context.
4513 * Note, that this is called and return with irqs disabled. This will
4514 * protect us against recursive calling from irq.
4516 asmlinkage
void __sched
preempt_schedule_irq(void)
4518 struct thread_info
*ti
= current_thread_info();
4520 /* Catch callers which need to be fixed */
4521 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4524 add_preempt_count(PREEMPT_ACTIVE
);
4527 local_irq_disable();
4528 sub_preempt_count(PREEMPT_ACTIVE
);
4531 * Check again in case we missed a preemption opportunity
4532 * between schedule and now.
4535 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4538 #endif /* CONFIG_PREEMPT */
4540 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4543 return try_to_wake_up(curr
->private, mode
, sync
);
4545 EXPORT_SYMBOL(default_wake_function
);
4548 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4549 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4550 * number) then we wake all the non-exclusive tasks and one exclusive task.
4552 * There are circumstances in which we can try to wake a task which has already
4553 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4554 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4556 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4557 int nr_exclusive
, int sync
, void *key
)
4559 wait_queue_t
*curr
, *next
;
4561 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4562 unsigned flags
= curr
->flags
;
4564 if (curr
->func(curr
, mode
, sync
, key
) &&
4565 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4571 * __wake_up - wake up threads blocked on a waitqueue.
4573 * @mode: which threads
4574 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4575 * @key: is directly passed to the wakeup function
4577 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4578 int nr_exclusive
, void *key
)
4580 unsigned long flags
;
4582 spin_lock_irqsave(&q
->lock
, flags
);
4583 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4584 spin_unlock_irqrestore(&q
->lock
, flags
);
4586 EXPORT_SYMBOL(__wake_up
);
4589 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4591 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4593 __wake_up_common(q
, mode
, 1, 0, NULL
);
4597 * __wake_up_sync - wake up threads blocked on a waitqueue.
4599 * @mode: which threads
4600 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4602 * The sync wakeup differs that the waker knows that it will schedule
4603 * away soon, so while the target thread will be woken up, it will not
4604 * be migrated to another CPU - ie. the two threads are 'synchronized'
4605 * with each other. This can prevent needless bouncing between CPUs.
4607 * On UP it can prevent extra preemption.
4610 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4612 unsigned long flags
;
4618 if (unlikely(!nr_exclusive
))
4621 spin_lock_irqsave(&q
->lock
, flags
);
4622 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4623 spin_unlock_irqrestore(&q
->lock
, flags
);
4625 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4628 * complete: - signals a single thread waiting on this completion
4629 * @x: holds the state of this particular completion
4631 * This will wake up a single thread waiting on this completion. Threads will be
4632 * awakened in the same order in which they were queued.
4634 * See also complete_all(), wait_for_completion() and related routines.
4636 void complete(struct completion
*x
)
4638 unsigned long flags
;
4640 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4642 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4643 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4645 EXPORT_SYMBOL(complete
);
4648 * complete_all: - signals all threads waiting on this completion
4649 * @x: holds the state of this particular completion
4651 * This will wake up all threads waiting on this particular completion event.
4653 void complete_all(struct completion
*x
)
4655 unsigned long flags
;
4657 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4658 x
->done
+= UINT_MAX
/2;
4659 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4660 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4662 EXPORT_SYMBOL(complete_all
);
4664 static inline long __sched
4665 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4668 DECLARE_WAITQUEUE(wait
, current
);
4670 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4671 __add_wait_queue_tail(&x
->wait
, &wait
);
4673 if (signal_pending_state(state
, current
)) {
4674 timeout
= -ERESTARTSYS
;
4677 __set_current_state(state
);
4678 spin_unlock_irq(&x
->wait
.lock
);
4679 timeout
= schedule_timeout(timeout
);
4680 spin_lock_irq(&x
->wait
.lock
);
4681 } while (!x
->done
&& timeout
);
4682 __remove_wait_queue(&x
->wait
, &wait
);
4687 return timeout
?: 1;
4691 wait_for_common(struct completion
*x
, long timeout
, int state
)
4695 spin_lock_irq(&x
->wait
.lock
);
4696 timeout
= do_wait_for_common(x
, timeout
, state
);
4697 spin_unlock_irq(&x
->wait
.lock
);
4702 * wait_for_completion: - waits for completion of a task
4703 * @x: holds the state of this particular completion
4705 * This waits to be signaled for completion of a specific task. It is NOT
4706 * interruptible and there is no timeout.
4708 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4709 * and interrupt capability. Also see complete().
4711 void __sched
wait_for_completion(struct completion
*x
)
4713 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4715 EXPORT_SYMBOL(wait_for_completion
);
4718 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4719 * @x: holds the state of this particular completion
4720 * @timeout: timeout value in jiffies
4722 * This waits for either a completion of a specific task to be signaled or for a
4723 * specified timeout to expire. The timeout is in jiffies. It is not
4726 unsigned long __sched
4727 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4729 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4731 EXPORT_SYMBOL(wait_for_completion_timeout
);
4734 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4735 * @x: holds the state of this particular completion
4737 * This waits for completion of a specific task to be signaled. It is
4740 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4742 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4743 if (t
== -ERESTARTSYS
)
4747 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4750 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4751 * @x: holds the state of this particular completion
4752 * @timeout: timeout value in jiffies
4754 * This waits for either a completion of a specific task to be signaled or for a
4755 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4757 unsigned long __sched
4758 wait_for_completion_interruptible_timeout(struct completion
*x
,
4759 unsigned long timeout
)
4761 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4763 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4766 * wait_for_completion_killable: - waits for completion of a task (killable)
4767 * @x: holds the state of this particular completion
4769 * This waits to be signaled for completion of a specific task. It can be
4770 * interrupted by a kill signal.
4772 int __sched
wait_for_completion_killable(struct completion
*x
)
4774 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4775 if (t
== -ERESTARTSYS
)
4779 EXPORT_SYMBOL(wait_for_completion_killable
);
4782 * try_wait_for_completion - try to decrement a completion without blocking
4783 * @x: completion structure
4785 * Returns: 0 if a decrement cannot be done without blocking
4786 * 1 if a decrement succeeded.
4788 * If a completion is being used as a counting completion,
4789 * attempt to decrement the counter without blocking. This
4790 * enables us to avoid waiting if the resource the completion
4791 * is protecting is not available.
4793 bool try_wait_for_completion(struct completion
*x
)
4797 spin_lock_irq(&x
->wait
.lock
);
4802 spin_unlock_irq(&x
->wait
.lock
);
4805 EXPORT_SYMBOL(try_wait_for_completion
);
4808 * completion_done - Test to see if a completion has any waiters
4809 * @x: completion structure
4811 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4812 * 1 if there are no waiters.
4815 bool completion_done(struct completion
*x
)
4819 spin_lock_irq(&x
->wait
.lock
);
4822 spin_unlock_irq(&x
->wait
.lock
);
4825 EXPORT_SYMBOL(completion_done
);
4828 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4830 unsigned long flags
;
4833 init_waitqueue_entry(&wait
, current
);
4835 __set_current_state(state
);
4837 spin_lock_irqsave(&q
->lock
, flags
);
4838 __add_wait_queue(q
, &wait
);
4839 spin_unlock(&q
->lock
);
4840 timeout
= schedule_timeout(timeout
);
4841 spin_lock_irq(&q
->lock
);
4842 __remove_wait_queue(q
, &wait
);
4843 spin_unlock_irqrestore(&q
->lock
, flags
);
4848 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4850 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4852 EXPORT_SYMBOL(interruptible_sleep_on
);
4855 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4857 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4859 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4861 void __sched
sleep_on(wait_queue_head_t
*q
)
4863 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4865 EXPORT_SYMBOL(sleep_on
);
4867 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4869 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4871 EXPORT_SYMBOL(sleep_on_timeout
);
4873 #ifdef CONFIG_RT_MUTEXES
4876 * rt_mutex_setprio - set the current priority of a task
4878 * @prio: prio value (kernel-internal form)
4880 * This function changes the 'effective' priority of a task. It does
4881 * not touch ->normal_prio like __setscheduler().
4883 * Used by the rt_mutex code to implement priority inheritance logic.
4885 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4887 unsigned long flags
;
4888 int oldprio
, on_rq
, running
;
4890 const struct sched_class
*prev_class
= p
->sched_class
;
4892 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4894 rq
= task_rq_lock(p
, &flags
);
4895 update_rq_clock(rq
);
4898 on_rq
= p
->se
.on_rq
;
4899 running
= task_current(rq
, p
);
4901 dequeue_task(rq
, p
, 0);
4903 p
->sched_class
->put_prev_task(rq
, p
);
4906 p
->sched_class
= &rt_sched_class
;
4908 p
->sched_class
= &fair_sched_class
;
4913 p
->sched_class
->set_curr_task(rq
);
4915 enqueue_task(rq
, p
, 0);
4917 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4919 task_rq_unlock(rq
, &flags
);
4924 void set_user_nice(struct task_struct
*p
, long nice
)
4926 int old_prio
, delta
, on_rq
;
4927 unsigned long flags
;
4930 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4933 * We have to be careful, if called from sys_setpriority(),
4934 * the task might be in the middle of scheduling on another CPU.
4936 rq
= task_rq_lock(p
, &flags
);
4937 update_rq_clock(rq
);
4939 * The RT priorities are set via sched_setscheduler(), but we still
4940 * allow the 'normal' nice value to be set - but as expected
4941 * it wont have any effect on scheduling until the task is
4942 * SCHED_FIFO/SCHED_RR:
4944 if (task_has_rt_policy(p
)) {
4945 p
->static_prio
= NICE_TO_PRIO(nice
);
4948 on_rq
= p
->se
.on_rq
;
4950 dequeue_task(rq
, p
, 0);
4952 p
->static_prio
= NICE_TO_PRIO(nice
);
4955 p
->prio
= effective_prio(p
);
4956 delta
= p
->prio
- old_prio
;
4959 enqueue_task(rq
, p
, 0);
4961 * If the task increased its priority or is running and
4962 * lowered its priority, then reschedule its CPU:
4964 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4965 resched_task(rq
->curr
);
4968 task_rq_unlock(rq
, &flags
);
4970 EXPORT_SYMBOL(set_user_nice
);
4973 * can_nice - check if a task can reduce its nice value
4977 int can_nice(const struct task_struct
*p
, const int nice
)
4979 /* convert nice value [19,-20] to rlimit style value [1,40] */
4980 int nice_rlim
= 20 - nice
;
4982 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4983 capable(CAP_SYS_NICE
));
4986 #ifdef __ARCH_WANT_SYS_NICE
4989 * sys_nice - change the priority of the current process.
4990 * @increment: priority increment
4992 * sys_setpriority is a more generic, but much slower function that
4993 * does similar things.
4995 asmlinkage
long sys_nice(int increment
)
5000 * Setpriority might change our priority at the same moment.
5001 * We don't have to worry. Conceptually one call occurs first
5002 * and we have a single winner.
5004 if (increment
< -40)
5009 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
5015 if (increment
< 0 && !can_nice(current
, nice
))
5018 retval
= security_task_setnice(current
, nice
);
5022 set_user_nice(current
, nice
);
5029 * task_prio - return the priority value of a given task.
5030 * @p: the task in question.
5032 * This is the priority value as seen by users in /proc.
5033 * RT tasks are offset by -200. Normal tasks are centered
5034 * around 0, value goes from -16 to +15.
5036 int task_prio(const struct task_struct
*p
)
5038 return p
->prio
- MAX_RT_PRIO
;
5042 * task_nice - return the nice value of a given task.
5043 * @p: the task in question.
5045 int task_nice(const struct task_struct
*p
)
5047 return TASK_NICE(p
);
5049 EXPORT_SYMBOL(task_nice
);
5052 * idle_cpu - is a given cpu idle currently?
5053 * @cpu: the processor in question.
5055 int idle_cpu(int cpu
)
5057 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
5061 * idle_task - return the idle task for a given cpu.
5062 * @cpu: the processor in question.
5064 struct task_struct
*idle_task(int cpu
)
5066 return cpu_rq(cpu
)->idle
;
5070 * find_process_by_pid - find a process with a matching PID value.
5071 * @pid: the pid in question.
5073 static struct task_struct
*find_process_by_pid(pid_t pid
)
5075 return pid
? find_task_by_vpid(pid
) : current
;
5078 /* Actually do priority change: must hold rq lock. */
5080 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
5082 BUG_ON(p
->se
.on_rq
);
5085 switch (p
->policy
) {
5089 p
->sched_class
= &fair_sched_class
;
5093 p
->sched_class
= &rt_sched_class
;
5097 p
->rt_priority
= prio
;
5098 p
->normal_prio
= normal_prio(p
);
5099 /* we are holding p->pi_lock already */
5100 p
->prio
= rt_mutex_getprio(p
);
5104 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
5105 struct sched_param
*param
, bool user
)
5107 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
5108 unsigned long flags
;
5109 const struct sched_class
*prev_class
= p
->sched_class
;
5112 /* may grab non-irq protected spin_locks */
5113 BUG_ON(in_interrupt());
5115 /* double check policy once rq lock held */
5117 policy
= oldpolicy
= p
->policy
;
5118 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
5119 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
5120 policy
!= SCHED_IDLE
)
5123 * Valid priorities for SCHED_FIFO and SCHED_RR are
5124 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5125 * SCHED_BATCH and SCHED_IDLE is 0.
5127 if (param
->sched_priority
< 0 ||
5128 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5129 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
5131 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
5135 * Allow unprivileged RT tasks to decrease priority:
5137 if (user
&& !capable(CAP_SYS_NICE
)) {
5138 if (rt_policy(policy
)) {
5139 unsigned long rlim_rtprio
;
5141 if (!lock_task_sighand(p
, &flags
))
5143 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
5144 unlock_task_sighand(p
, &flags
);
5146 /* can't set/change the rt policy */
5147 if (policy
!= p
->policy
&& !rlim_rtprio
)
5150 /* can't increase priority */
5151 if (param
->sched_priority
> p
->rt_priority
&&
5152 param
->sched_priority
> rlim_rtprio
)
5156 * Like positive nice levels, dont allow tasks to
5157 * move out of SCHED_IDLE either:
5159 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
5162 /* can't change other user's priorities */
5163 if ((current
->euid
!= p
->euid
) &&
5164 (current
->euid
!= p
->uid
))
5169 #ifdef CONFIG_RT_GROUP_SCHED
5171 * Do not allow realtime tasks into groups that have no runtime
5174 if (rt_policy(policy
) && task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
5178 retval
= security_task_setscheduler(p
, policy
, param
);
5184 * make sure no PI-waiters arrive (or leave) while we are
5185 * changing the priority of the task:
5187 spin_lock_irqsave(&p
->pi_lock
, flags
);
5189 * To be able to change p->policy safely, the apropriate
5190 * runqueue lock must be held.
5192 rq
= __task_rq_lock(p
);
5193 /* recheck policy now with rq lock held */
5194 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5195 policy
= oldpolicy
= -1;
5196 __task_rq_unlock(rq
);
5197 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5200 update_rq_clock(rq
);
5201 on_rq
= p
->se
.on_rq
;
5202 running
= task_current(rq
, p
);
5204 deactivate_task(rq
, p
, 0);
5206 p
->sched_class
->put_prev_task(rq
, p
);
5209 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5212 p
->sched_class
->set_curr_task(rq
);
5214 activate_task(rq
, p
, 0);
5216 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5218 __task_rq_unlock(rq
);
5219 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5221 rt_mutex_adjust_pi(p
);
5227 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5228 * @p: the task in question.
5229 * @policy: new policy.
5230 * @param: structure containing the new RT priority.
5232 * NOTE that the task may be already dead.
5234 int sched_setscheduler(struct task_struct
*p
, int policy
,
5235 struct sched_param
*param
)
5237 return __sched_setscheduler(p
, policy
, param
, true);
5239 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5242 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5243 * @p: the task in question.
5244 * @policy: new policy.
5245 * @param: structure containing the new RT priority.
5247 * Just like sched_setscheduler, only don't bother checking if the
5248 * current context has permission. For example, this is needed in
5249 * stop_machine(): we create temporary high priority worker threads,
5250 * but our caller might not have that capability.
5252 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5253 struct sched_param
*param
)
5255 return __sched_setscheduler(p
, policy
, param
, false);
5259 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5261 struct sched_param lparam
;
5262 struct task_struct
*p
;
5265 if (!param
|| pid
< 0)
5267 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5272 p
= find_process_by_pid(pid
);
5274 retval
= sched_setscheduler(p
, policy
, &lparam
);
5281 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5282 * @pid: the pid in question.
5283 * @policy: new policy.
5284 * @param: structure containing the new RT priority.
5287 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5289 /* negative values for policy are not valid */
5293 return do_sched_setscheduler(pid
, policy
, param
);
5297 * sys_sched_setparam - set/change the RT priority of a thread
5298 * @pid: the pid in question.
5299 * @param: structure containing the new RT priority.
5301 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
5303 return do_sched_setscheduler(pid
, -1, param
);
5307 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5308 * @pid: the pid in question.
5310 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
5312 struct task_struct
*p
;
5319 read_lock(&tasklist_lock
);
5320 p
= find_process_by_pid(pid
);
5322 retval
= security_task_getscheduler(p
);
5326 read_unlock(&tasklist_lock
);
5331 * sys_sched_getscheduler - get the RT priority of a thread
5332 * @pid: the pid in question.
5333 * @param: structure containing the RT priority.
5335 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
5337 struct sched_param lp
;
5338 struct task_struct
*p
;
5341 if (!param
|| pid
< 0)
5344 read_lock(&tasklist_lock
);
5345 p
= find_process_by_pid(pid
);
5350 retval
= security_task_getscheduler(p
);
5354 lp
.sched_priority
= p
->rt_priority
;
5355 read_unlock(&tasklist_lock
);
5358 * This one might sleep, we cannot do it with a spinlock held ...
5360 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5365 read_unlock(&tasklist_lock
);
5369 long sched_setaffinity(pid_t pid
, const cpumask_t
*in_mask
)
5371 cpumask_t cpus_allowed
;
5372 cpumask_t new_mask
= *in_mask
;
5373 struct task_struct
*p
;
5377 read_lock(&tasklist_lock
);
5379 p
= find_process_by_pid(pid
);
5381 read_unlock(&tasklist_lock
);
5387 * It is not safe to call set_cpus_allowed with the
5388 * tasklist_lock held. We will bump the task_struct's
5389 * usage count and then drop tasklist_lock.
5392 read_unlock(&tasklist_lock
);
5395 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
5396 !capable(CAP_SYS_NICE
))
5399 retval
= security_task_setscheduler(p
, 0, NULL
);
5403 cpuset_cpus_allowed(p
, &cpus_allowed
);
5404 cpus_and(new_mask
, new_mask
, cpus_allowed
);
5406 retval
= set_cpus_allowed_ptr(p
, &new_mask
);
5409 cpuset_cpus_allowed(p
, &cpus_allowed
);
5410 if (!cpus_subset(new_mask
, cpus_allowed
)) {
5412 * We must have raced with a concurrent cpuset
5413 * update. Just reset the cpus_allowed to the
5414 * cpuset's cpus_allowed
5416 new_mask
= cpus_allowed
;
5426 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5427 cpumask_t
*new_mask
)
5429 if (len
< sizeof(cpumask_t
)) {
5430 memset(new_mask
, 0, sizeof(cpumask_t
));
5431 } else if (len
> sizeof(cpumask_t
)) {
5432 len
= sizeof(cpumask_t
);
5434 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5438 * sys_sched_setaffinity - set the cpu affinity of a process
5439 * @pid: pid of the process
5440 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5441 * @user_mask_ptr: user-space pointer to the new cpu mask
5443 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
5444 unsigned long __user
*user_mask_ptr
)
5449 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
5453 return sched_setaffinity(pid
, &new_mask
);
5456 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
5458 struct task_struct
*p
;
5462 read_lock(&tasklist_lock
);
5465 p
= find_process_by_pid(pid
);
5469 retval
= security_task_getscheduler(p
);
5473 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
5476 read_unlock(&tasklist_lock
);
5483 * sys_sched_getaffinity - get the cpu affinity of a process
5484 * @pid: pid of the process
5485 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5486 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5488 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
5489 unsigned long __user
*user_mask_ptr
)
5494 if (len
< sizeof(cpumask_t
))
5497 ret
= sched_getaffinity(pid
, &mask
);
5501 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
5504 return sizeof(cpumask_t
);
5508 * sys_sched_yield - yield the current processor to other threads.
5510 * This function yields the current CPU to other tasks. If there are no
5511 * other threads running on this CPU then this function will return.
5513 asmlinkage
long sys_sched_yield(void)
5515 struct rq
*rq
= this_rq_lock();
5517 schedstat_inc(rq
, yld_count
);
5518 current
->sched_class
->yield_task(rq
);
5521 * Since we are going to call schedule() anyway, there's
5522 * no need to preempt or enable interrupts:
5524 __release(rq
->lock
);
5525 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5526 _raw_spin_unlock(&rq
->lock
);
5527 preempt_enable_no_resched();
5534 static void __cond_resched(void)
5536 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5537 __might_sleep(__FILE__
, __LINE__
);
5540 * The BKS might be reacquired before we have dropped
5541 * PREEMPT_ACTIVE, which could trigger a second
5542 * cond_resched() call.
5545 add_preempt_count(PREEMPT_ACTIVE
);
5547 sub_preempt_count(PREEMPT_ACTIVE
);
5548 } while (need_resched());
5551 int __sched
_cond_resched(void)
5553 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
5554 system_state
== SYSTEM_RUNNING
) {
5560 EXPORT_SYMBOL(_cond_resched
);
5563 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5564 * call schedule, and on return reacquire the lock.
5566 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5567 * operations here to prevent schedule() from being called twice (once via
5568 * spin_unlock(), once by hand).
5570 int cond_resched_lock(spinlock_t
*lock
)
5572 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
5575 if (spin_needbreak(lock
) || resched
) {
5577 if (resched
&& need_resched())
5586 EXPORT_SYMBOL(cond_resched_lock
);
5588 int __sched
cond_resched_softirq(void)
5590 BUG_ON(!in_softirq());
5592 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
5600 EXPORT_SYMBOL(cond_resched_softirq
);
5603 * yield - yield the current processor to other threads.
5605 * This is a shortcut for kernel-space yielding - it marks the
5606 * thread runnable and calls sys_sched_yield().
5608 void __sched
yield(void)
5610 set_current_state(TASK_RUNNING
);
5613 EXPORT_SYMBOL(yield
);
5616 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5617 * that process accounting knows that this is a task in IO wait state.
5619 * But don't do that if it is a deliberate, throttling IO wait (this task
5620 * has set its backing_dev_info: the queue against which it should throttle)
5622 void __sched
io_schedule(void)
5624 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5626 delayacct_blkio_start();
5627 atomic_inc(&rq
->nr_iowait
);
5629 atomic_dec(&rq
->nr_iowait
);
5630 delayacct_blkio_end();
5632 EXPORT_SYMBOL(io_schedule
);
5634 long __sched
io_schedule_timeout(long timeout
)
5636 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5639 delayacct_blkio_start();
5640 atomic_inc(&rq
->nr_iowait
);
5641 ret
= schedule_timeout(timeout
);
5642 atomic_dec(&rq
->nr_iowait
);
5643 delayacct_blkio_end();
5648 * sys_sched_get_priority_max - return maximum RT priority.
5649 * @policy: scheduling class.
5651 * this syscall returns the maximum rt_priority that can be used
5652 * by a given scheduling class.
5654 asmlinkage
long sys_sched_get_priority_max(int policy
)
5661 ret
= MAX_USER_RT_PRIO
-1;
5673 * sys_sched_get_priority_min - return minimum RT priority.
5674 * @policy: scheduling class.
5676 * this syscall returns the minimum rt_priority that can be used
5677 * by a given scheduling class.
5679 asmlinkage
long sys_sched_get_priority_min(int policy
)
5697 * sys_sched_rr_get_interval - return the default timeslice of a process.
5698 * @pid: pid of the process.
5699 * @interval: userspace pointer to the timeslice value.
5701 * this syscall writes the default timeslice value of a given process
5702 * into the user-space timespec buffer. A value of '0' means infinity.
5705 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
5707 struct task_struct
*p
;
5708 unsigned int time_slice
;
5716 read_lock(&tasklist_lock
);
5717 p
= find_process_by_pid(pid
);
5721 retval
= security_task_getscheduler(p
);
5726 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5727 * tasks that are on an otherwise idle runqueue:
5730 if (p
->policy
== SCHED_RR
) {
5731 time_slice
= DEF_TIMESLICE
;
5732 } else if (p
->policy
!= SCHED_FIFO
) {
5733 struct sched_entity
*se
= &p
->se
;
5734 unsigned long flags
;
5737 rq
= task_rq_lock(p
, &flags
);
5738 if (rq
->cfs
.load
.weight
)
5739 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5740 task_rq_unlock(rq
, &flags
);
5742 read_unlock(&tasklist_lock
);
5743 jiffies_to_timespec(time_slice
, &t
);
5744 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5748 read_unlock(&tasklist_lock
);
5752 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5754 void sched_show_task(struct task_struct
*p
)
5756 unsigned long free
= 0;
5759 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5760 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5761 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5762 #if BITS_PER_LONG == 32
5763 if (state
== TASK_RUNNING
)
5764 printk(KERN_CONT
" running ");
5766 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5768 if (state
== TASK_RUNNING
)
5769 printk(KERN_CONT
" running task ");
5771 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5773 #ifdef CONFIG_DEBUG_STACK_USAGE
5775 unsigned long *n
= end_of_stack(p
);
5778 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5781 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5782 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5784 show_stack(p
, NULL
);
5787 void show_state_filter(unsigned long state_filter
)
5789 struct task_struct
*g
, *p
;
5791 #if BITS_PER_LONG == 32
5793 " task PC stack pid father\n");
5796 " task PC stack pid father\n");
5798 read_lock(&tasklist_lock
);
5799 do_each_thread(g
, p
) {
5801 * reset the NMI-timeout, listing all files on a slow
5802 * console might take alot of time:
5804 touch_nmi_watchdog();
5805 if (!state_filter
|| (p
->state
& state_filter
))
5807 } while_each_thread(g
, p
);
5809 touch_all_softlockup_watchdogs();
5811 #ifdef CONFIG_SCHED_DEBUG
5812 sysrq_sched_debug_show();
5814 read_unlock(&tasklist_lock
);
5816 * Only show locks if all tasks are dumped:
5818 if (state_filter
== -1)
5819 debug_show_all_locks();
5822 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5824 idle
->sched_class
= &idle_sched_class
;
5828 * init_idle - set up an idle thread for a given CPU
5829 * @idle: task in question
5830 * @cpu: cpu the idle task belongs to
5832 * NOTE: this function does not set the idle thread's NEED_RESCHED
5833 * flag, to make booting more robust.
5835 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5837 struct rq
*rq
= cpu_rq(cpu
);
5838 unsigned long flags
;
5841 idle
->se
.exec_start
= sched_clock();
5843 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5844 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5845 __set_task_cpu(idle
, cpu
);
5847 spin_lock_irqsave(&rq
->lock
, flags
);
5848 rq
->curr
= rq
->idle
= idle
;
5849 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5852 spin_unlock_irqrestore(&rq
->lock
, flags
);
5854 /* Set the preempt count _outside_ the spinlocks! */
5855 #if defined(CONFIG_PREEMPT)
5856 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5858 task_thread_info(idle
)->preempt_count
= 0;
5861 * The idle tasks have their own, simple scheduling class:
5863 idle
->sched_class
= &idle_sched_class
;
5867 * In a system that switches off the HZ timer nohz_cpu_mask
5868 * indicates which cpus entered this state. This is used
5869 * in the rcu update to wait only for active cpus. For system
5870 * which do not switch off the HZ timer nohz_cpu_mask should
5871 * always be CPU_MASK_NONE.
5873 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5876 * Increase the granularity value when there are more CPUs,
5877 * because with more CPUs the 'effective latency' as visible
5878 * to users decreases. But the relationship is not linear,
5879 * so pick a second-best guess by going with the log2 of the
5882 * This idea comes from the SD scheduler of Con Kolivas:
5884 static inline void sched_init_granularity(void)
5886 unsigned int factor
= 1 + ilog2(num_online_cpus());
5887 const unsigned long limit
= 200000000;
5889 sysctl_sched_min_granularity
*= factor
;
5890 if (sysctl_sched_min_granularity
> limit
)
5891 sysctl_sched_min_granularity
= limit
;
5893 sysctl_sched_latency
*= factor
;
5894 if (sysctl_sched_latency
> limit
)
5895 sysctl_sched_latency
= limit
;
5897 sysctl_sched_wakeup_granularity
*= factor
;
5899 sysctl_sched_shares_ratelimit
*= factor
;
5904 * This is how migration works:
5906 * 1) we queue a struct migration_req structure in the source CPU's
5907 * runqueue and wake up that CPU's migration thread.
5908 * 2) we down() the locked semaphore => thread blocks.
5909 * 3) migration thread wakes up (implicitly it forces the migrated
5910 * thread off the CPU)
5911 * 4) it gets the migration request and checks whether the migrated
5912 * task is still in the wrong runqueue.
5913 * 5) if it's in the wrong runqueue then the migration thread removes
5914 * it and puts it into the right queue.
5915 * 6) migration thread up()s the semaphore.
5916 * 7) we wake up and the migration is done.
5920 * Change a given task's CPU affinity. Migrate the thread to a
5921 * proper CPU and schedule it away if the CPU it's executing on
5922 * is removed from the allowed bitmask.
5924 * NOTE: the caller must have a valid reference to the task, the
5925 * task must not exit() & deallocate itself prematurely. The
5926 * call is not atomic; no spinlocks may be held.
5928 int set_cpus_allowed_ptr(struct task_struct
*p
, const cpumask_t
*new_mask
)
5930 struct migration_req req
;
5931 unsigned long flags
;
5935 rq
= task_rq_lock(p
, &flags
);
5936 if (!cpus_intersects(*new_mask
, cpu_online_map
)) {
5941 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5942 !cpus_equal(p
->cpus_allowed
, *new_mask
))) {
5947 if (p
->sched_class
->set_cpus_allowed
)
5948 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5950 p
->cpus_allowed
= *new_mask
;
5951 p
->rt
.nr_cpus_allowed
= cpus_weight(*new_mask
);
5954 /* Can the task run on the task's current CPU? If so, we're done */
5955 if (cpu_isset(task_cpu(p
), *new_mask
))
5958 if (migrate_task(p
, any_online_cpu(*new_mask
), &req
)) {
5959 /* Need help from migration thread: drop lock and wait. */
5960 task_rq_unlock(rq
, &flags
);
5961 wake_up_process(rq
->migration_thread
);
5962 wait_for_completion(&req
.done
);
5963 tlb_migrate_finish(p
->mm
);
5967 task_rq_unlock(rq
, &flags
);
5971 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5974 * Move (not current) task off this cpu, onto dest cpu. We're doing
5975 * this because either it can't run here any more (set_cpus_allowed()
5976 * away from this CPU, or CPU going down), or because we're
5977 * attempting to rebalance this task on exec (sched_exec).
5979 * So we race with normal scheduler movements, but that's OK, as long
5980 * as the task is no longer on this CPU.
5982 * Returns non-zero if task was successfully migrated.
5984 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5986 struct rq
*rq_dest
, *rq_src
;
5989 if (unlikely(!cpu_active(dest_cpu
)))
5992 rq_src
= cpu_rq(src_cpu
);
5993 rq_dest
= cpu_rq(dest_cpu
);
5995 double_rq_lock(rq_src
, rq_dest
);
5996 /* Already moved. */
5997 if (task_cpu(p
) != src_cpu
)
5999 /* Affinity changed (again). */
6000 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
6003 on_rq
= p
->se
.on_rq
;
6005 deactivate_task(rq_src
, p
, 0);
6007 set_task_cpu(p
, dest_cpu
);
6009 activate_task(rq_dest
, p
, 0);
6010 check_preempt_curr(rq_dest
, p
, 0);
6015 double_rq_unlock(rq_src
, rq_dest
);
6020 * migration_thread - this is a highprio system thread that performs
6021 * thread migration by bumping thread off CPU then 'pushing' onto
6024 static int migration_thread(void *data
)
6026 int cpu
= (long)data
;
6030 BUG_ON(rq
->migration_thread
!= current
);
6032 set_current_state(TASK_INTERRUPTIBLE
);
6033 while (!kthread_should_stop()) {
6034 struct migration_req
*req
;
6035 struct list_head
*head
;
6037 spin_lock_irq(&rq
->lock
);
6039 if (cpu_is_offline(cpu
)) {
6040 spin_unlock_irq(&rq
->lock
);
6044 if (rq
->active_balance
) {
6045 active_load_balance(rq
, cpu
);
6046 rq
->active_balance
= 0;
6049 head
= &rq
->migration_queue
;
6051 if (list_empty(head
)) {
6052 spin_unlock_irq(&rq
->lock
);
6054 set_current_state(TASK_INTERRUPTIBLE
);
6057 req
= list_entry(head
->next
, struct migration_req
, list
);
6058 list_del_init(head
->next
);
6060 spin_unlock(&rq
->lock
);
6061 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
6064 complete(&req
->done
);
6066 __set_current_state(TASK_RUNNING
);
6070 /* Wait for kthread_stop */
6071 set_current_state(TASK_INTERRUPTIBLE
);
6072 while (!kthread_should_stop()) {
6074 set_current_state(TASK_INTERRUPTIBLE
);
6076 __set_current_state(TASK_RUNNING
);
6080 #ifdef CONFIG_HOTPLUG_CPU
6082 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6086 local_irq_disable();
6087 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
6093 * Figure out where task on dead CPU should go, use force if necessary.
6094 * NOTE: interrupts should be disabled by the caller
6096 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
6098 unsigned long flags
;
6105 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
6106 cpus_and(mask
, mask
, p
->cpus_allowed
);
6107 dest_cpu
= any_online_cpu(mask
);
6109 /* On any allowed CPU? */
6110 if (dest_cpu
>= nr_cpu_ids
)
6111 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
6113 /* No more Mr. Nice Guy. */
6114 if (dest_cpu
>= nr_cpu_ids
) {
6115 cpumask_t cpus_allowed
;
6117 cpuset_cpus_allowed_locked(p
, &cpus_allowed
);
6119 * Try to stay on the same cpuset, where the
6120 * current cpuset may be a subset of all cpus.
6121 * The cpuset_cpus_allowed_locked() variant of
6122 * cpuset_cpus_allowed() will not block. It must be
6123 * called within calls to cpuset_lock/cpuset_unlock.
6125 rq
= task_rq_lock(p
, &flags
);
6126 p
->cpus_allowed
= cpus_allowed
;
6127 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
6128 task_rq_unlock(rq
, &flags
);
6131 * Don't tell them about moving exiting tasks or
6132 * kernel threads (both mm NULL), since they never
6135 if (p
->mm
&& printk_ratelimit()) {
6136 printk(KERN_INFO
"process %d (%s) no "
6137 "longer affine to cpu%d\n",
6138 task_pid_nr(p
), p
->comm
, dead_cpu
);
6141 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
6145 * While a dead CPU has no uninterruptible tasks queued at this point,
6146 * it might still have a nonzero ->nr_uninterruptible counter, because
6147 * for performance reasons the counter is not stricly tracking tasks to
6148 * their home CPUs. So we just add the counter to another CPU's counter,
6149 * to keep the global sum constant after CPU-down:
6151 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
6153 struct rq
*rq_dest
= cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR
));
6154 unsigned long flags
;
6156 local_irq_save(flags
);
6157 double_rq_lock(rq_src
, rq_dest
);
6158 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
6159 rq_src
->nr_uninterruptible
= 0;
6160 double_rq_unlock(rq_src
, rq_dest
);
6161 local_irq_restore(flags
);
6164 /* Run through task list and migrate tasks from the dead cpu. */
6165 static void migrate_live_tasks(int src_cpu
)
6167 struct task_struct
*p
, *t
;
6169 read_lock(&tasklist_lock
);
6171 do_each_thread(t
, p
) {
6175 if (task_cpu(p
) == src_cpu
)
6176 move_task_off_dead_cpu(src_cpu
, p
);
6177 } while_each_thread(t
, p
);
6179 read_unlock(&tasklist_lock
);
6183 * Schedules idle task to be the next runnable task on current CPU.
6184 * It does so by boosting its priority to highest possible.
6185 * Used by CPU offline code.
6187 void sched_idle_next(void)
6189 int this_cpu
= smp_processor_id();
6190 struct rq
*rq
= cpu_rq(this_cpu
);
6191 struct task_struct
*p
= rq
->idle
;
6192 unsigned long flags
;
6194 /* cpu has to be offline */
6195 BUG_ON(cpu_online(this_cpu
));
6198 * Strictly not necessary since rest of the CPUs are stopped by now
6199 * and interrupts disabled on the current cpu.
6201 spin_lock_irqsave(&rq
->lock
, flags
);
6203 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6205 update_rq_clock(rq
);
6206 activate_task(rq
, p
, 0);
6208 spin_unlock_irqrestore(&rq
->lock
, flags
);
6212 * Ensures that the idle task is using init_mm right before its cpu goes
6215 void idle_task_exit(void)
6217 struct mm_struct
*mm
= current
->active_mm
;
6219 BUG_ON(cpu_online(smp_processor_id()));
6222 switch_mm(mm
, &init_mm
, current
);
6226 /* called under rq->lock with disabled interrupts */
6227 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
6229 struct rq
*rq
= cpu_rq(dead_cpu
);
6231 /* Must be exiting, otherwise would be on tasklist. */
6232 BUG_ON(!p
->exit_state
);
6234 /* Cannot have done final schedule yet: would have vanished. */
6235 BUG_ON(p
->state
== TASK_DEAD
);
6240 * Drop lock around migration; if someone else moves it,
6241 * that's OK. No task can be added to this CPU, so iteration is
6244 spin_unlock_irq(&rq
->lock
);
6245 move_task_off_dead_cpu(dead_cpu
, p
);
6246 spin_lock_irq(&rq
->lock
);
6251 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6252 static void migrate_dead_tasks(unsigned int dead_cpu
)
6254 struct rq
*rq
= cpu_rq(dead_cpu
);
6255 struct task_struct
*next
;
6258 if (!rq
->nr_running
)
6260 update_rq_clock(rq
);
6261 next
= pick_next_task(rq
, rq
->curr
);
6264 next
->sched_class
->put_prev_task(rq
, next
);
6265 migrate_dead(dead_cpu
, next
);
6269 #endif /* CONFIG_HOTPLUG_CPU */
6271 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6273 static struct ctl_table sd_ctl_dir
[] = {
6275 .procname
= "sched_domain",
6281 static struct ctl_table sd_ctl_root
[] = {
6283 .ctl_name
= CTL_KERN
,
6284 .procname
= "kernel",
6286 .child
= sd_ctl_dir
,
6291 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6293 struct ctl_table
*entry
=
6294 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6299 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6301 struct ctl_table
*entry
;
6304 * In the intermediate directories, both the child directory and
6305 * procname are dynamically allocated and could fail but the mode
6306 * will always be set. In the lowest directory the names are
6307 * static strings and all have proc handlers.
6309 for (entry
= *tablep
; entry
->mode
; entry
++) {
6311 sd_free_ctl_entry(&entry
->child
);
6312 if (entry
->proc_handler
== NULL
)
6313 kfree(entry
->procname
);
6321 set_table_entry(struct ctl_table
*entry
,
6322 const char *procname
, void *data
, int maxlen
,
6323 mode_t mode
, proc_handler
*proc_handler
)
6325 entry
->procname
= procname
;
6327 entry
->maxlen
= maxlen
;
6329 entry
->proc_handler
= proc_handler
;
6332 static struct ctl_table
*
6333 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6335 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
6340 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6341 sizeof(long), 0644, proc_doulongvec_minmax
);
6342 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6343 sizeof(long), 0644, proc_doulongvec_minmax
);
6344 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6345 sizeof(int), 0644, proc_dointvec_minmax
);
6346 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6347 sizeof(int), 0644, proc_dointvec_minmax
);
6348 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6349 sizeof(int), 0644, proc_dointvec_minmax
);
6350 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6351 sizeof(int), 0644, proc_dointvec_minmax
);
6352 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6353 sizeof(int), 0644, proc_dointvec_minmax
);
6354 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6355 sizeof(int), 0644, proc_dointvec_minmax
);
6356 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6357 sizeof(int), 0644, proc_dointvec_minmax
);
6358 set_table_entry(&table
[9], "cache_nice_tries",
6359 &sd
->cache_nice_tries
,
6360 sizeof(int), 0644, proc_dointvec_minmax
);
6361 set_table_entry(&table
[10], "flags", &sd
->flags
,
6362 sizeof(int), 0644, proc_dointvec_minmax
);
6363 /* &table[11] is terminator */
6368 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6370 struct ctl_table
*entry
, *table
;
6371 struct sched_domain
*sd
;
6372 int domain_num
= 0, i
;
6375 for_each_domain(cpu
, sd
)
6377 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6382 for_each_domain(cpu
, sd
) {
6383 snprintf(buf
, 32, "domain%d", i
);
6384 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6386 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6393 static struct ctl_table_header
*sd_sysctl_header
;
6394 static void register_sched_domain_sysctl(void)
6396 int i
, cpu_num
= num_online_cpus();
6397 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6400 WARN_ON(sd_ctl_dir
[0].child
);
6401 sd_ctl_dir
[0].child
= entry
;
6406 for_each_online_cpu(i
) {
6407 snprintf(buf
, 32, "cpu%d", i
);
6408 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6410 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6414 WARN_ON(sd_sysctl_header
);
6415 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6418 /* may be called multiple times per register */
6419 static void unregister_sched_domain_sysctl(void)
6421 if (sd_sysctl_header
)
6422 unregister_sysctl_table(sd_sysctl_header
);
6423 sd_sysctl_header
= NULL
;
6424 if (sd_ctl_dir
[0].child
)
6425 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6428 static void register_sched_domain_sysctl(void)
6431 static void unregister_sched_domain_sysctl(void)
6436 static void set_rq_online(struct rq
*rq
)
6439 const struct sched_class
*class;
6441 cpu_set(rq
->cpu
, rq
->rd
->online
);
6444 for_each_class(class) {
6445 if (class->rq_online
)
6446 class->rq_online(rq
);
6451 static void set_rq_offline(struct rq
*rq
)
6454 const struct sched_class
*class;
6456 for_each_class(class) {
6457 if (class->rq_offline
)
6458 class->rq_offline(rq
);
6461 cpu_clear(rq
->cpu
, rq
->rd
->online
);
6467 * migration_call - callback that gets triggered when a CPU is added.
6468 * Here we can start up the necessary migration thread for the new CPU.
6470 static int __cpuinit
6471 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6473 struct task_struct
*p
;
6474 int cpu
= (long)hcpu
;
6475 unsigned long flags
;
6480 case CPU_UP_PREPARE
:
6481 case CPU_UP_PREPARE_FROZEN
:
6482 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
6485 kthread_bind(p
, cpu
);
6486 /* Must be high prio: stop_machine expects to yield to it. */
6487 rq
= task_rq_lock(p
, &flags
);
6488 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6489 task_rq_unlock(rq
, &flags
);
6490 cpu_rq(cpu
)->migration_thread
= p
;
6494 case CPU_ONLINE_FROZEN
:
6495 /* Strictly unnecessary, as first user will wake it. */
6496 wake_up_process(cpu_rq(cpu
)->migration_thread
);
6498 /* Update our root-domain */
6500 spin_lock_irqsave(&rq
->lock
, flags
);
6502 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6506 spin_unlock_irqrestore(&rq
->lock
, flags
);
6509 #ifdef CONFIG_HOTPLUG_CPU
6510 case CPU_UP_CANCELED
:
6511 case CPU_UP_CANCELED_FROZEN
:
6512 if (!cpu_rq(cpu
)->migration_thread
)
6514 /* Unbind it from offline cpu so it can run. Fall thru. */
6515 kthread_bind(cpu_rq(cpu
)->migration_thread
,
6516 any_online_cpu(cpu_online_map
));
6517 kthread_stop(cpu_rq(cpu
)->migration_thread
);
6518 cpu_rq(cpu
)->migration_thread
= NULL
;
6522 case CPU_DEAD_FROZEN
:
6523 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6524 migrate_live_tasks(cpu
);
6526 kthread_stop(rq
->migration_thread
);
6527 rq
->migration_thread
= NULL
;
6528 /* Idle task back to normal (off runqueue, low prio) */
6529 spin_lock_irq(&rq
->lock
);
6530 update_rq_clock(rq
);
6531 deactivate_task(rq
, rq
->idle
, 0);
6532 rq
->idle
->static_prio
= MAX_PRIO
;
6533 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
6534 rq
->idle
->sched_class
= &idle_sched_class
;
6535 migrate_dead_tasks(cpu
);
6536 spin_unlock_irq(&rq
->lock
);
6538 migrate_nr_uninterruptible(rq
);
6539 BUG_ON(rq
->nr_running
!= 0);
6542 * No need to migrate the tasks: it was best-effort if
6543 * they didn't take sched_hotcpu_mutex. Just wake up
6546 spin_lock_irq(&rq
->lock
);
6547 while (!list_empty(&rq
->migration_queue
)) {
6548 struct migration_req
*req
;
6550 req
= list_entry(rq
->migration_queue
.next
,
6551 struct migration_req
, list
);
6552 list_del_init(&req
->list
);
6553 complete(&req
->done
);
6555 spin_unlock_irq(&rq
->lock
);
6559 case CPU_DYING_FROZEN
:
6560 /* Update our root-domain */
6562 spin_lock_irqsave(&rq
->lock
, flags
);
6564 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6567 spin_unlock_irqrestore(&rq
->lock
, flags
);
6574 /* Register at highest priority so that task migration (migrate_all_tasks)
6575 * happens before everything else.
6577 static struct notifier_block __cpuinitdata migration_notifier
= {
6578 .notifier_call
= migration_call
,
6582 static int __init
migration_init(void)
6584 void *cpu
= (void *)(long)smp_processor_id();
6587 /* Start one for the boot CPU: */
6588 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6589 BUG_ON(err
== NOTIFY_BAD
);
6590 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6591 register_cpu_notifier(&migration_notifier
);
6595 early_initcall(migration_init
);
6600 #ifdef CONFIG_SCHED_DEBUG
6602 static inline const char *sd_level_to_string(enum sched_domain_level lvl
)
6615 case SD_LV_ALLNODES
:
6624 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6625 cpumask_t
*groupmask
)
6627 struct sched_group
*group
= sd
->groups
;
6630 cpulist_scnprintf(str
, sizeof(str
), sd
->span
);
6631 cpus_clear(*groupmask
);
6633 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6635 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6636 printk("does not load-balance\n");
6638 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6643 printk(KERN_CONT
"span %s level %s\n",
6644 str
, sd_level_to_string(sd
->level
));
6646 if (!cpu_isset(cpu
, sd
->span
)) {
6647 printk(KERN_ERR
"ERROR: domain->span does not contain "
6650 if (!cpu_isset(cpu
, group
->cpumask
)) {
6651 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6655 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6659 printk(KERN_ERR
"ERROR: group is NULL\n");
6663 if (!group
->__cpu_power
) {
6664 printk(KERN_CONT
"\n");
6665 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6670 if (!cpus_weight(group
->cpumask
)) {
6671 printk(KERN_CONT
"\n");
6672 printk(KERN_ERR
"ERROR: empty group\n");
6676 if (cpus_intersects(*groupmask
, group
->cpumask
)) {
6677 printk(KERN_CONT
"\n");
6678 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6682 cpus_or(*groupmask
, *groupmask
, group
->cpumask
);
6684 cpulist_scnprintf(str
, sizeof(str
), group
->cpumask
);
6685 printk(KERN_CONT
" %s", str
);
6687 group
= group
->next
;
6688 } while (group
!= sd
->groups
);
6689 printk(KERN_CONT
"\n");
6691 if (!cpus_equal(sd
->span
, *groupmask
))
6692 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6694 if (sd
->parent
&& !cpus_subset(*groupmask
, sd
->parent
->span
))
6695 printk(KERN_ERR
"ERROR: parent span is not a superset "
6696 "of domain->span\n");
6700 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6702 cpumask_t
*groupmask
;
6706 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6710 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6712 groupmask
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6714 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6719 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6728 #else /* !CONFIG_SCHED_DEBUG */
6729 # define sched_domain_debug(sd, cpu) do { } while (0)
6730 #endif /* CONFIG_SCHED_DEBUG */
6732 static int sd_degenerate(struct sched_domain
*sd
)
6734 if (cpus_weight(sd
->span
) == 1)
6737 /* Following flags need at least 2 groups */
6738 if (sd
->flags
& (SD_LOAD_BALANCE
|
6739 SD_BALANCE_NEWIDLE
|
6743 SD_SHARE_PKG_RESOURCES
)) {
6744 if (sd
->groups
!= sd
->groups
->next
)
6748 /* Following flags don't use groups */
6749 if (sd
->flags
& (SD_WAKE_IDLE
|
6758 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6760 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6762 if (sd_degenerate(parent
))
6765 if (!cpus_equal(sd
->span
, parent
->span
))
6768 /* Does parent contain flags not in child? */
6769 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6770 if (cflags
& SD_WAKE_AFFINE
)
6771 pflags
&= ~SD_WAKE_BALANCE
;
6772 /* Flags needing groups don't count if only 1 group in parent */
6773 if (parent
->groups
== parent
->groups
->next
) {
6774 pflags
&= ~(SD_LOAD_BALANCE
|
6775 SD_BALANCE_NEWIDLE
|
6779 SD_SHARE_PKG_RESOURCES
);
6781 if (~cflags
& pflags
)
6787 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6789 unsigned long flags
;
6791 spin_lock_irqsave(&rq
->lock
, flags
);
6794 struct root_domain
*old_rd
= rq
->rd
;
6796 if (cpu_isset(rq
->cpu
, old_rd
->online
))
6799 cpu_clear(rq
->cpu
, old_rd
->span
);
6801 if (atomic_dec_and_test(&old_rd
->refcount
))
6805 atomic_inc(&rd
->refcount
);
6808 cpu_set(rq
->cpu
, rd
->span
);
6809 if (cpu_isset(rq
->cpu
, cpu_online_map
))
6812 spin_unlock_irqrestore(&rq
->lock
, flags
);
6815 static void init_rootdomain(struct root_domain
*rd
)
6817 memset(rd
, 0, sizeof(*rd
));
6819 cpus_clear(rd
->span
);
6820 cpus_clear(rd
->online
);
6822 cpupri_init(&rd
->cpupri
);
6825 static void init_defrootdomain(void)
6827 init_rootdomain(&def_root_domain
);
6828 atomic_set(&def_root_domain
.refcount
, 1);
6831 static struct root_domain
*alloc_rootdomain(void)
6833 struct root_domain
*rd
;
6835 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6839 init_rootdomain(rd
);
6845 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6846 * hold the hotplug lock.
6849 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6851 struct rq
*rq
= cpu_rq(cpu
);
6852 struct sched_domain
*tmp
;
6854 /* Remove the sched domains which do not contribute to scheduling. */
6855 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
6856 struct sched_domain
*parent
= tmp
->parent
;
6859 if (sd_parent_degenerate(tmp
, parent
)) {
6860 tmp
->parent
= parent
->parent
;
6862 parent
->parent
->child
= tmp
;
6866 if (sd
&& sd_degenerate(sd
)) {
6872 sched_domain_debug(sd
, cpu
);
6874 rq_attach_root(rq
, rd
);
6875 rcu_assign_pointer(rq
->sd
, sd
);
6878 /* cpus with isolated domains */
6879 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
6881 /* Setup the mask of cpus configured for isolated domains */
6882 static int __init
isolated_cpu_setup(char *str
)
6884 static int __initdata ints
[NR_CPUS
];
6887 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
6888 cpus_clear(cpu_isolated_map
);
6889 for (i
= 1; i
<= ints
[0]; i
++)
6890 if (ints
[i
] < NR_CPUS
)
6891 cpu_set(ints
[i
], cpu_isolated_map
);
6895 __setup("isolcpus=", isolated_cpu_setup
);
6898 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6899 * to a function which identifies what group(along with sched group) a CPU
6900 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6901 * (due to the fact that we keep track of groups covered with a cpumask_t).
6903 * init_sched_build_groups will build a circular linked list of the groups
6904 * covered by the given span, and will set each group's ->cpumask correctly,
6905 * and ->cpu_power to 0.
6908 init_sched_build_groups(const cpumask_t
*span
, const cpumask_t
*cpu_map
,
6909 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
6910 struct sched_group
**sg
,
6911 cpumask_t
*tmpmask
),
6912 cpumask_t
*covered
, cpumask_t
*tmpmask
)
6914 struct sched_group
*first
= NULL
, *last
= NULL
;
6917 cpus_clear(*covered
);
6919 for_each_cpu_mask_nr(i
, *span
) {
6920 struct sched_group
*sg
;
6921 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6924 if (cpu_isset(i
, *covered
))
6927 cpus_clear(sg
->cpumask
);
6928 sg
->__cpu_power
= 0;
6930 for_each_cpu_mask_nr(j
, *span
) {
6931 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6934 cpu_set(j
, *covered
);
6935 cpu_set(j
, sg
->cpumask
);
6946 #define SD_NODES_PER_DOMAIN 16
6951 * find_next_best_node - find the next node to include in a sched_domain
6952 * @node: node whose sched_domain we're building
6953 * @used_nodes: nodes already in the sched_domain
6955 * Find the next node to include in a given scheduling domain. Simply
6956 * finds the closest node not already in the @used_nodes map.
6958 * Should use nodemask_t.
6960 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6962 int i
, n
, val
, min_val
, best_node
= 0;
6966 for (i
= 0; i
< nr_node_ids
; i
++) {
6967 /* Start at @node */
6968 n
= (node
+ i
) % nr_node_ids
;
6970 if (!nr_cpus_node(n
))
6973 /* Skip already used nodes */
6974 if (node_isset(n
, *used_nodes
))
6977 /* Simple min distance search */
6978 val
= node_distance(node
, n
);
6980 if (val
< min_val
) {
6986 node_set(best_node
, *used_nodes
);
6991 * sched_domain_node_span - get a cpumask for a node's sched_domain
6992 * @node: node whose cpumask we're constructing
6993 * @span: resulting cpumask
6995 * Given a node, construct a good cpumask for its sched_domain to span. It
6996 * should be one that prevents unnecessary balancing, but also spreads tasks
6999 static void sched_domain_node_span(int node
, cpumask_t
*span
)
7001 nodemask_t used_nodes
;
7002 node_to_cpumask_ptr(nodemask
, node
);
7006 nodes_clear(used_nodes
);
7008 cpus_or(*span
, *span
, *nodemask
);
7009 node_set(node
, used_nodes
);
7011 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
7012 int next_node
= find_next_best_node(node
, &used_nodes
);
7014 node_to_cpumask_ptr_next(nodemask
, next_node
);
7015 cpus_or(*span
, *span
, *nodemask
);
7018 #endif /* CONFIG_NUMA */
7020 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
7023 * SMT sched-domains:
7025 #ifdef CONFIG_SCHED_SMT
7026 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
7027 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
7030 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7034 *sg
= &per_cpu(sched_group_cpus
, cpu
);
7037 #endif /* CONFIG_SCHED_SMT */
7040 * multi-core sched-domains:
7042 #ifdef CONFIG_SCHED_MC
7043 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
7044 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
7045 #endif /* CONFIG_SCHED_MC */
7047 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7049 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7054 *mask
= per_cpu(cpu_sibling_map
, cpu
);
7055 cpus_and(*mask
, *mask
, *cpu_map
);
7056 group
= first_cpu(*mask
);
7058 *sg
= &per_cpu(sched_group_core
, group
);
7061 #elif defined(CONFIG_SCHED_MC)
7063 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7067 *sg
= &per_cpu(sched_group_core
, cpu
);
7072 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
7073 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
7076 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7080 #ifdef CONFIG_SCHED_MC
7081 *mask
= cpu_coregroup_map(cpu
);
7082 cpus_and(*mask
, *mask
, *cpu_map
);
7083 group
= first_cpu(*mask
);
7084 #elif defined(CONFIG_SCHED_SMT)
7085 *mask
= per_cpu(cpu_sibling_map
, cpu
);
7086 cpus_and(*mask
, *mask
, *cpu_map
);
7087 group
= first_cpu(*mask
);
7092 *sg
= &per_cpu(sched_group_phys
, group
);
7098 * The init_sched_build_groups can't handle what we want to do with node
7099 * groups, so roll our own. Now each node has its own list of groups which
7100 * gets dynamically allocated.
7102 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
7103 static struct sched_group
***sched_group_nodes_bycpu
;
7105 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
7106 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
7108 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
7109 struct sched_group
**sg
, cpumask_t
*nodemask
)
7113 *nodemask
= node_to_cpumask(cpu_to_node(cpu
));
7114 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7115 group
= first_cpu(*nodemask
);
7118 *sg
= &per_cpu(sched_group_allnodes
, group
);
7122 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
7124 struct sched_group
*sg
= group_head
;
7130 for_each_cpu_mask_nr(j
, sg
->cpumask
) {
7131 struct sched_domain
*sd
;
7133 sd
= &per_cpu(phys_domains
, j
);
7134 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
7136 * Only add "power" once for each
7142 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
7145 } while (sg
!= group_head
);
7147 #endif /* CONFIG_NUMA */
7150 /* Free memory allocated for various sched_group structures */
7151 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
7155 for_each_cpu_mask_nr(cpu
, *cpu_map
) {
7156 struct sched_group
**sched_group_nodes
7157 = sched_group_nodes_bycpu
[cpu
];
7159 if (!sched_group_nodes
)
7162 for (i
= 0; i
< nr_node_ids
; i
++) {
7163 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
7165 *nodemask
= node_to_cpumask(i
);
7166 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7167 if (cpus_empty(*nodemask
))
7177 if (oldsg
!= sched_group_nodes
[i
])
7180 kfree(sched_group_nodes
);
7181 sched_group_nodes_bycpu
[cpu
] = NULL
;
7184 #else /* !CONFIG_NUMA */
7185 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
7188 #endif /* CONFIG_NUMA */
7191 * Initialize sched groups cpu_power.
7193 * cpu_power indicates the capacity of sched group, which is used while
7194 * distributing the load between different sched groups in a sched domain.
7195 * Typically cpu_power for all the groups in a sched domain will be same unless
7196 * there are asymmetries in the topology. If there are asymmetries, group
7197 * having more cpu_power will pickup more load compared to the group having
7200 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7201 * the maximum number of tasks a group can handle in the presence of other idle
7202 * or lightly loaded groups in the same sched domain.
7204 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7206 struct sched_domain
*child
;
7207 struct sched_group
*group
;
7209 WARN_ON(!sd
|| !sd
->groups
);
7211 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
7216 sd
->groups
->__cpu_power
= 0;
7219 * For perf policy, if the groups in child domain share resources
7220 * (for example cores sharing some portions of the cache hierarchy
7221 * or SMT), then set this domain groups cpu_power such that each group
7222 * can handle only one task, when there are other idle groups in the
7223 * same sched domain.
7225 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
7227 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
7228 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
7233 * add cpu_power of each child group to this groups cpu_power
7235 group
= child
->groups
;
7237 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
7238 group
= group
->next
;
7239 } while (group
!= child
->groups
);
7243 * Initializers for schedule domains
7244 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7247 #define SD_INIT(sd, type) sd_init_##type(sd)
7248 #define SD_INIT_FUNC(type) \
7249 static noinline void sd_init_##type(struct sched_domain *sd) \
7251 memset(sd, 0, sizeof(*sd)); \
7252 *sd = SD_##type##_INIT; \
7253 sd->level = SD_LV_##type; \
7258 SD_INIT_FUNC(ALLNODES
)
7261 #ifdef CONFIG_SCHED_SMT
7262 SD_INIT_FUNC(SIBLING
)
7264 #ifdef CONFIG_SCHED_MC
7269 * To minimize stack usage kmalloc room for cpumasks and share the
7270 * space as the usage in build_sched_domains() dictates. Used only
7271 * if the amount of space is significant.
7274 cpumask_t tmpmask
; /* make this one first */
7277 cpumask_t this_sibling_map
;
7278 cpumask_t this_core_map
;
7280 cpumask_t send_covered
;
7283 cpumask_t domainspan
;
7285 cpumask_t notcovered
;
7290 #define SCHED_CPUMASK_ALLOC 1
7291 #define SCHED_CPUMASK_FREE(v) kfree(v)
7292 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7294 #define SCHED_CPUMASK_ALLOC 0
7295 #define SCHED_CPUMASK_FREE(v)
7296 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7299 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7300 ((unsigned long)(a) + offsetof(struct allmasks, v))
7302 static int default_relax_domain_level
= -1;
7304 static int __init
setup_relax_domain_level(char *str
)
7308 val
= simple_strtoul(str
, NULL
, 0);
7309 if (val
< SD_LV_MAX
)
7310 default_relax_domain_level
= val
;
7314 __setup("relax_domain_level=", setup_relax_domain_level
);
7316 static void set_domain_attribute(struct sched_domain
*sd
,
7317 struct sched_domain_attr
*attr
)
7321 if (!attr
|| attr
->relax_domain_level
< 0) {
7322 if (default_relax_domain_level
< 0)
7325 request
= default_relax_domain_level
;
7327 request
= attr
->relax_domain_level
;
7328 if (request
< sd
->level
) {
7329 /* turn off idle balance on this domain */
7330 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
7332 /* turn on idle balance on this domain */
7333 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
7338 * Build sched domains for a given set of cpus and attach the sched domains
7339 * to the individual cpus
7341 static int __build_sched_domains(const cpumask_t
*cpu_map
,
7342 struct sched_domain_attr
*attr
)
7345 struct root_domain
*rd
;
7346 SCHED_CPUMASK_DECLARE(allmasks
);
7349 struct sched_group
**sched_group_nodes
= NULL
;
7350 int sd_allnodes
= 0;
7353 * Allocate the per-node list of sched groups
7355 sched_group_nodes
= kcalloc(nr_node_ids
, sizeof(struct sched_group
*),
7357 if (!sched_group_nodes
) {
7358 printk(KERN_WARNING
"Can not alloc sched group node list\n");
7363 rd
= alloc_rootdomain();
7365 printk(KERN_WARNING
"Cannot alloc root domain\n");
7367 kfree(sched_group_nodes
);
7372 #if SCHED_CPUMASK_ALLOC
7373 /* get space for all scratch cpumask variables */
7374 allmasks
= kmalloc(sizeof(*allmasks
), GFP_KERNEL
);
7376 printk(KERN_WARNING
"Cannot alloc cpumask array\n");
7379 kfree(sched_group_nodes
);
7384 tmpmask
= (cpumask_t
*)allmasks
;
7388 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
7392 * Set up domains for cpus specified by the cpu_map.
7394 for_each_cpu_mask_nr(i
, *cpu_map
) {
7395 struct sched_domain
*sd
= NULL
, *p
;
7396 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7398 *nodemask
= node_to_cpumask(cpu_to_node(i
));
7399 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7402 if (cpus_weight(*cpu_map
) >
7403 SD_NODES_PER_DOMAIN
*cpus_weight(*nodemask
)) {
7404 sd
= &per_cpu(allnodes_domains
, i
);
7405 SD_INIT(sd
, ALLNODES
);
7406 set_domain_attribute(sd
, attr
);
7407 sd
->span
= *cpu_map
;
7408 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7414 sd
= &per_cpu(node_domains
, i
);
7416 set_domain_attribute(sd
, attr
);
7417 sched_domain_node_span(cpu_to_node(i
), &sd
->span
);
7421 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7425 sd
= &per_cpu(phys_domains
, i
);
7427 set_domain_attribute(sd
, attr
);
7428 sd
->span
= *nodemask
;
7432 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7434 #ifdef CONFIG_SCHED_MC
7436 sd
= &per_cpu(core_domains
, i
);
7438 set_domain_attribute(sd
, attr
);
7439 sd
->span
= cpu_coregroup_map(i
);
7440 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7443 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7446 #ifdef CONFIG_SCHED_SMT
7448 sd
= &per_cpu(cpu_domains
, i
);
7449 SD_INIT(sd
, SIBLING
);
7450 set_domain_attribute(sd
, attr
);
7451 sd
->span
= per_cpu(cpu_sibling_map
, i
);
7452 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7455 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7459 #ifdef CONFIG_SCHED_SMT
7460 /* Set up CPU (sibling) groups */
7461 for_each_cpu_mask_nr(i
, *cpu_map
) {
7462 SCHED_CPUMASK_VAR(this_sibling_map
, allmasks
);
7463 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7465 *this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
7466 cpus_and(*this_sibling_map
, *this_sibling_map
, *cpu_map
);
7467 if (i
!= first_cpu(*this_sibling_map
))
7470 init_sched_build_groups(this_sibling_map
, cpu_map
,
7472 send_covered
, tmpmask
);
7476 #ifdef CONFIG_SCHED_MC
7477 /* Set up multi-core groups */
7478 for_each_cpu_mask_nr(i
, *cpu_map
) {
7479 SCHED_CPUMASK_VAR(this_core_map
, allmasks
);
7480 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7482 *this_core_map
= cpu_coregroup_map(i
);
7483 cpus_and(*this_core_map
, *this_core_map
, *cpu_map
);
7484 if (i
!= first_cpu(*this_core_map
))
7487 init_sched_build_groups(this_core_map
, cpu_map
,
7489 send_covered
, tmpmask
);
7493 /* Set up physical groups */
7494 for (i
= 0; i
< nr_node_ids
; i
++) {
7495 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7496 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7498 *nodemask
= node_to_cpumask(i
);
7499 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7500 if (cpus_empty(*nodemask
))
7503 init_sched_build_groups(nodemask
, cpu_map
,
7505 send_covered
, tmpmask
);
7509 /* Set up node groups */
7511 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7513 init_sched_build_groups(cpu_map
, cpu_map
,
7514 &cpu_to_allnodes_group
,
7515 send_covered
, tmpmask
);
7518 for (i
= 0; i
< nr_node_ids
; i
++) {
7519 /* Set up node groups */
7520 struct sched_group
*sg
, *prev
;
7521 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7522 SCHED_CPUMASK_VAR(domainspan
, allmasks
);
7523 SCHED_CPUMASK_VAR(covered
, allmasks
);
7526 *nodemask
= node_to_cpumask(i
);
7527 cpus_clear(*covered
);
7529 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7530 if (cpus_empty(*nodemask
)) {
7531 sched_group_nodes
[i
] = NULL
;
7535 sched_domain_node_span(i
, domainspan
);
7536 cpus_and(*domainspan
, *domainspan
, *cpu_map
);
7538 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
7540 printk(KERN_WARNING
"Can not alloc domain group for "
7544 sched_group_nodes
[i
] = sg
;
7545 for_each_cpu_mask_nr(j
, *nodemask
) {
7546 struct sched_domain
*sd
;
7548 sd
= &per_cpu(node_domains
, j
);
7551 sg
->__cpu_power
= 0;
7552 sg
->cpumask
= *nodemask
;
7554 cpus_or(*covered
, *covered
, *nodemask
);
7557 for (j
= 0; j
< nr_node_ids
; j
++) {
7558 SCHED_CPUMASK_VAR(notcovered
, allmasks
);
7559 int n
= (i
+ j
) % nr_node_ids
;
7560 node_to_cpumask_ptr(pnodemask
, n
);
7562 cpus_complement(*notcovered
, *covered
);
7563 cpus_and(*tmpmask
, *notcovered
, *cpu_map
);
7564 cpus_and(*tmpmask
, *tmpmask
, *domainspan
);
7565 if (cpus_empty(*tmpmask
))
7568 cpus_and(*tmpmask
, *tmpmask
, *pnodemask
);
7569 if (cpus_empty(*tmpmask
))
7572 sg
= kmalloc_node(sizeof(struct sched_group
),
7576 "Can not alloc domain group for node %d\n", j
);
7579 sg
->__cpu_power
= 0;
7580 sg
->cpumask
= *tmpmask
;
7581 sg
->next
= prev
->next
;
7582 cpus_or(*covered
, *covered
, *tmpmask
);
7589 /* Calculate CPU power for physical packages and nodes */
7590 #ifdef CONFIG_SCHED_SMT
7591 for_each_cpu_mask_nr(i
, *cpu_map
) {
7592 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
7594 init_sched_groups_power(i
, sd
);
7597 #ifdef CONFIG_SCHED_MC
7598 for_each_cpu_mask_nr(i
, *cpu_map
) {
7599 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
7601 init_sched_groups_power(i
, sd
);
7605 for_each_cpu_mask_nr(i
, *cpu_map
) {
7606 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
7608 init_sched_groups_power(i
, sd
);
7612 for (i
= 0; i
< nr_node_ids
; i
++)
7613 init_numa_sched_groups_power(sched_group_nodes
[i
]);
7616 struct sched_group
*sg
;
7618 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
,
7620 init_numa_sched_groups_power(sg
);
7624 /* Attach the domains */
7625 for_each_cpu_mask_nr(i
, *cpu_map
) {
7626 struct sched_domain
*sd
;
7627 #ifdef CONFIG_SCHED_SMT
7628 sd
= &per_cpu(cpu_domains
, i
);
7629 #elif defined(CONFIG_SCHED_MC)
7630 sd
= &per_cpu(core_domains
, i
);
7632 sd
= &per_cpu(phys_domains
, i
);
7634 cpu_attach_domain(sd
, rd
, i
);
7637 SCHED_CPUMASK_FREE((void *)allmasks
);
7642 free_sched_groups(cpu_map
, tmpmask
);
7643 SCHED_CPUMASK_FREE((void *)allmasks
);
7648 static int build_sched_domains(const cpumask_t
*cpu_map
)
7650 return __build_sched_domains(cpu_map
, NULL
);
7653 static cpumask_t
*doms_cur
; /* current sched domains */
7654 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7655 static struct sched_domain_attr
*dattr_cur
;
7656 /* attribues of custom domains in 'doms_cur' */
7659 * Special case: If a kmalloc of a doms_cur partition (array of
7660 * cpumask_t) fails, then fallback to a single sched domain,
7661 * as determined by the single cpumask_t fallback_doms.
7663 static cpumask_t fallback_doms
;
7665 void __attribute__((weak
)) arch_update_cpu_topology(void)
7670 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7671 * For now this just excludes isolated cpus, but could be used to
7672 * exclude other special cases in the future.
7674 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
7678 arch_update_cpu_topology();
7680 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
7682 doms_cur
= &fallback_doms
;
7683 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
7685 err
= build_sched_domains(doms_cur
);
7686 register_sched_domain_sysctl();
7691 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
,
7694 free_sched_groups(cpu_map
, tmpmask
);
7698 * Detach sched domains from a group of cpus specified in cpu_map
7699 * These cpus will now be attached to the NULL domain
7701 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
7706 unregister_sched_domain_sysctl();
7708 for_each_cpu_mask_nr(i
, *cpu_map
)
7709 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7710 synchronize_sched();
7711 arch_destroy_sched_domains(cpu_map
, &tmpmask
);
7714 /* handle null as "default" */
7715 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7716 struct sched_domain_attr
*new, int idx_new
)
7718 struct sched_domain_attr tmp
;
7725 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7726 new ? (new + idx_new
) : &tmp
,
7727 sizeof(struct sched_domain_attr
));
7731 * Partition sched domains as specified by the 'ndoms_new'
7732 * cpumasks in the array doms_new[] of cpumasks. This compares
7733 * doms_new[] to the current sched domain partitioning, doms_cur[].
7734 * It destroys each deleted domain and builds each new domain.
7736 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7737 * The masks don't intersect (don't overlap.) We should setup one
7738 * sched domain for each mask. CPUs not in any of the cpumasks will
7739 * not be load balanced. If the same cpumask appears both in the
7740 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7743 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7744 * ownership of it and will kfree it when done with it. If the caller
7745 * failed the kmalloc call, then it can pass in doms_new == NULL,
7746 * and partition_sched_domains() will fallback to the single partition
7747 * 'fallback_doms', it also forces the domains to be rebuilt.
7749 * If doms_new==NULL it will be replaced with cpu_online_map.
7750 * ndoms_new==0 is a special case for destroying existing domains.
7751 * It will not create the default domain.
7753 * Call with hotplug lock held
7755 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
,
7756 struct sched_domain_attr
*dattr_new
)
7760 mutex_lock(&sched_domains_mutex
);
7762 /* always unregister in case we don't destroy any domains */
7763 unregister_sched_domain_sysctl();
7765 n
= doms_new
? ndoms_new
: 0;
7767 /* Destroy deleted domains */
7768 for (i
= 0; i
< ndoms_cur
; i
++) {
7769 for (j
= 0; j
< n
; j
++) {
7770 if (cpus_equal(doms_cur
[i
], doms_new
[j
])
7771 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7774 /* no match - a current sched domain not in new doms_new[] */
7775 detach_destroy_domains(doms_cur
+ i
);
7780 if (doms_new
== NULL
) {
7782 doms_new
= &fallback_doms
;
7783 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
7787 /* Build new domains */
7788 for (i
= 0; i
< ndoms_new
; i
++) {
7789 for (j
= 0; j
< ndoms_cur
; j
++) {
7790 if (cpus_equal(doms_new
[i
], doms_cur
[j
])
7791 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7794 /* no match - add a new doms_new */
7795 __build_sched_domains(doms_new
+ i
,
7796 dattr_new
? dattr_new
+ i
: NULL
);
7801 /* Remember the new sched domains */
7802 if (doms_cur
!= &fallback_doms
)
7804 kfree(dattr_cur
); /* kfree(NULL) is safe */
7805 doms_cur
= doms_new
;
7806 dattr_cur
= dattr_new
;
7807 ndoms_cur
= ndoms_new
;
7809 register_sched_domain_sysctl();
7811 mutex_unlock(&sched_domains_mutex
);
7814 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7815 int arch_reinit_sched_domains(void)
7819 /* Destroy domains first to force the rebuild */
7820 partition_sched_domains(0, NULL
, NULL
);
7822 rebuild_sched_domains();
7828 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7832 if (buf
[0] != '0' && buf
[0] != '1')
7836 sched_smt_power_savings
= (buf
[0] == '1');
7838 sched_mc_power_savings
= (buf
[0] == '1');
7840 ret
= arch_reinit_sched_domains();
7842 return ret
? ret
: count
;
7845 #ifdef CONFIG_SCHED_MC
7846 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7849 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7851 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7852 const char *buf
, size_t count
)
7854 return sched_power_savings_store(buf
, count
, 0);
7856 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7857 sched_mc_power_savings_show
,
7858 sched_mc_power_savings_store
);
7861 #ifdef CONFIG_SCHED_SMT
7862 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7865 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7867 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7868 const char *buf
, size_t count
)
7870 return sched_power_savings_store(buf
, count
, 1);
7872 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7873 sched_smt_power_savings_show
,
7874 sched_smt_power_savings_store
);
7877 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7881 #ifdef CONFIG_SCHED_SMT
7883 err
= sysfs_create_file(&cls
->kset
.kobj
,
7884 &attr_sched_smt_power_savings
.attr
);
7886 #ifdef CONFIG_SCHED_MC
7887 if (!err
&& mc_capable())
7888 err
= sysfs_create_file(&cls
->kset
.kobj
,
7889 &attr_sched_mc_power_savings
.attr
);
7893 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7895 #ifndef CONFIG_CPUSETS
7897 * Add online and remove offline CPUs from the scheduler domains.
7898 * When cpusets are enabled they take over this function.
7900 static int update_sched_domains(struct notifier_block
*nfb
,
7901 unsigned long action
, void *hcpu
)
7905 case CPU_ONLINE_FROZEN
:
7907 case CPU_DEAD_FROZEN
:
7908 partition_sched_domains(1, NULL
, NULL
);
7917 static int update_runtime(struct notifier_block
*nfb
,
7918 unsigned long action
, void *hcpu
)
7920 int cpu
= (int)(long)hcpu
;
7923 case CPU_DOWN_PREPARE
:
7924 case CPU_DOWN_PREPARE_FROZEN
:
7925 disable_runtime(cpu_rq(cpu
));
7928 case CPU_DOWN_FAILED
:
7929 case CPU_DOWN_FAILED_FROZEN
:
7931 case CPU_ONLINE_FROZEN
:
7932 enable_runtime(cpu_rq(cpu
));
7940 void __init
sched_init_smp(void)
7942 cpumask_t non_isolated_cpus
;
7944 #if defined(CONFIG_NUMA)
7945 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7947 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7950 mutex_lock(&sched_domains_mutex
);
7951 arch_init_sched_domains(&cpu_online_map
);
7952 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
7953 if (cpus_empty(non_isolated_cpus
))
7954 cpu_set(smp_processor_id(), non_isolated_cpus
);
7955 mutex_unlock(&sched_domains_mutex
);
7958 #ifndef CONFIG_CPUSETS
7959 /* XXX: Theoretical race here - CPU may be hotplugged now */
7960 hotcpu_notifier(update_sched_domains
, 0);
7963 /* RT runtime code needs to handle some hotplug events */
7964 hotcpu_notifier(update_runtime
, 0);
7968 /* Move init over to a non-isolated CPU */
7969 if (set_cpus_allowed_ptr(current
, &non_isolated_cpus
) < 0)
7971 sched_init_granularity();
7974 void __init
sched_init_smp(void)
7976 sched_init_granularity();
7978 #endif /* CONFIG_SMP */
7980 int in_sched_functions(unsigned long addr
)
7982 return in_lock_functions(addr
) ||
7983 (addr
>= (unsigned long)__sched_text_start
7984 && addr
< (unsigned long)__sched_text_end
);
7987 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7989 cfs_rq
->tasks_timeline
= RB_ROOT
;
7990 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7991 #ifdef CONFIG_FAIR_GROUP_SCHED
7994 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7997 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7999 struct rt_prio_array
*array
;
8002 array
= &rt_rq
->active
;
8003 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
8004 INIT_LIST_HEAD(array
->queue
+ i
);
8005 __clear_bit(i
, array
->bitmap
);
8007 /* delimiter for bitsearch: */
8008 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
8010 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8011 rt_rq
->highest_prio
= MAX_RT_PRIO
;
8014 rt_rq
->rt_nr_migratory
= 0;
8015 rt_rq
->overloaded
= 0;
8019 rt_rq
->rt_throttled
= 0;
8020 rt_rq
->rt_runtime
= 0;
8021 spin_lock_init(&rt_rq
->rt_runtime_lock
);
8023 #ifdef CONFIG_RT_GROUP_SCHED
8024 rt_rq
->rt_nr_boosted
= 0;
8029 #ifdef CONFIG_FAIR_GROUP_SCHED
8030 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8031 struct sched_entity
*se
, int cpu
, int add
,
8032 struct sched_entity
*parent
)
8034 struct rq
*rq
= cpu_rq(cpu
);
8035 tg
->cfs_rq
[cpu
] = cfs_rq
;
8036 init_cfs_rq(cfs_rq
, rq
);
8039 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
8042 /* se could be NULL for init_task_group */
8047 se
->cfs_rq
= &rq
->cfs
;
8049 se
->cfs_rq
= parent
->my_q
;
8052 se
->load
.weight
= tg
->shares
;
8053 se
->load
.inv_weight
= 0;
8054 se
->parent
= parent
;
8058 #ifdef CONFIG_RT_GROUP_SCHED
8059 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
8060 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
8061 struct sched_rt_entity
*parent
)
8063 struct rq
*rq
= cpu_rq(cpu
);
8065 tg
->rt_rq
[cpu
] = rt_rq
;
8066 init_rt_rq(rt_rq
, rq
);
8068 rt_rq
->rt_se
= rt_se
;
8069 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8071 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
8073 tg
->rt_se
[cpu
] = rt_se
;
8078 rt_se
->rt_rq
= &rq
->rt
;
8080 rt_se
->rt_rq
= parent
->my_q
;
8082 rt_se
->my_q
= rt_rq
;
8083 rt_se
->parent
= parent
;
8084 INIT_LIST_HEAD(&rt_se
->run_list
);
8088 void __init
sched_init(void)
8091 unsigned long alloc_size
= 0, ptr
;
8093 #ifdef CONFIG_FAIR_GROUP_SCHED
8094 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8096 #ifdef CONFIG_RT_GROUP_SCHED
8097 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8099 #ifdef CONFIG_USER_SCHED
8103 * As sched_init() is called before page_alloc is setup,
8104 * we use alloc_bootmem().
8107 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
8109 #ifdef CONFIG_FAIR_GROUP_SCHED
8110 init_task_group
.se
= (struct sched_entity
**)ptr
;
8111 ptr
+= nr_cpu_ids
* sizeof(void **);
8113 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8114 ptr
+= nr_cpu_ids
* sizeof(void **);
8116 #ifdef CONFIG_USER_SCHED
8117 root_task_group
.se
= (struct sched_entity
**)ptr
;
8118 ptr
+= nr_cpu_ids
* sizeof(void **);
8120 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8121 ptr
+= nr_cpu_ids
* sizeof(void **);
8122 #endif /* CONFIG_USER_SCHED */
8123 #endif /* CONFIG_FAIR_GROUP_SCHED */
8124 #ifdef CONFIG_RT_GROUP_SCHED
8125 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8126 ptr
+= nr_cpu_ids
* sizeof(void **);
8128 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8129 ptr
+= nr_cpu_ids
* sizeof(void **);
8131 #ifdef CONFIG_USER_SCHED
8132 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8133 ptr
+= nr_cpu_ids
* sizeof(void **);
8135 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8136 ptr
+= nr_cpu_ids
* sizeof(void **);
8137 #endif /* CONFIG_USER_SCHED */
8138 #endif /* CONFIG_RT_GROUP_SCHED */
8142 init_defrootdomain();
8145 init_rt_bandwidth(&def_rt_bandwidth
,
8146 global_rt_period(), global_rt_runtime());
8148 #ifdef CONFIG_RT_GROUP_SCHED
8149 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
8150 global_rt_period(), global_rt_runtime());
8151 #ifdef CONFIG_USER_SCHED
8152 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8153 global_rt_period(), RUNTIME_INF
);
8154 #endif /* CONFIG_USER_SCHED */
8155 #endif /* CONFIG_RT_GROUP_SCHED */
8157 #ifdef CONFIG_GROUP_SCHED
8158 list_add(&init_task_group
.list
, &task_groups
);
8159 INIT_LIST_HEAD(&init_task_group
.children
);
8161 #ifdef CONFIG_USER_SCHED
8162 INIT_LIST_HEAD(&root_task_group
.children
);
8163 init_task_group
.parent
= &root_task_group
;
8164 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
8165 #endif /* CONFIG_USER_SCHED */
8166 #endif /* CONFIG_GROUP_SCHED */
8168 for_each_possible_cpu(i
) {
8172 spin_lock_init(&rq
->lock
);
8174 init_cfs_rq(&rq
->cfs
, rq
);
8175 init_rt_rq(&rq
->rt
, rq
);
8176 #ifdef CONFIG_FAIR_GROUP_SCHED
8177 init_task_group
.shares
= init_task_group_load
;
8178 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8179 #ifdef CONFIG_CGROUP_SCHED
8181 * How much cpu bandwidth does init_task_group get?
8183 * In case of task-groups formed thr' the cgroup filesystem, it
8184 * gets 100% of the cpu resources in the system. This overall
8185 * system cpu resource is divided among the tasks of
8186 * init_task_group and its child task-groups in a fair manner,
8187 * based on each entity's (task or task-group's) weight
8188 * (se->load.weight).
8190 * In other words, if init_task_group has 10 tasks of weight
8191 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8192 * then A0's share of the cpu resource is:
8194 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8196 * We achieve this by letting init_task_group's tasks sit
8197 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8199 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
8200 #elif defined CONFIG_USER_SCHED
8201 root_task_group
.shares
= NICE_0_LOAD
;
8202 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
8204 * In case of task-groups formed thr' the user id of tasks,
8205 * init_task_group represents tasks belonging to root user.
8206 * Hence it forms a sibling of all subsequent groups formed.
8207 * In this case, init_task_group gets only a fraction of overall
8208 * system cpu resource, based on the weight assigned to root
8209 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8210 * by letting tasks of init_task_group sit in a separate cfs_rq
8211 * (init_cfs_rq) and having one entity represent this group of
8212 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8214 init_tg_cfs_entry(&init_task_group
,
8215 &per_cpu(init_cfs_rq
, i
),
8216 &per_cpu(init_sched_entity
, i
), i
, 1,
8217 root_task_group
.se
[i
]);
8220 #endif /* CONFIG_FAIR_GROUP_SCHED */
8222 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8223 #ifdef CONFIG_RT_GROUP_SCHED
8224 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
8225 #ifdef CONFIG_CGROUP_SCHED
8226 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
8227 #elif defined CONFIG_USER_SCHED
8228 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
8229 init_tg_rt_entry(&init_task_group
,
8230 &per_cpu(init_rt_rq
, i
),
8231 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
8232 root_task_group
.rt_se
[i
]);
8236 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
8237 rq
->cpu_load
[j
] = 0;
8241 rq
->active_balance
= 0;
8242 rq
->next_balance
= jiffies
;
8246 rq
->migration_thread
= NULL
;
8247 INIT_LIST_HEAD(&rq
->migration_queue
);
8248 rq_attach_root(rq
, &def_root_domain
);
8251 atomic_set(&rq
->nr_iowait
, 0);
8254 set_load_weight(&init_task
);
8256 #ifdef CONFIG_PREEMPT_NOTIFIERS
8257 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
8261 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8264 #ifdef CONFIG_RT_MUTEXES
8265 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
8269 * The boot idle thread does lazy MMU switching as well:
8271 atomic_inc(&init_mm
.mm_count
);
8272 enter_lazy_tlb(&init_mm
, current
);
8275 * Make us the idle thread. Technically, schedule() should not be
8276 * called from this thread, however somewhere below it might be,
8277 * but because we are the idle thread, we just pick up running again
8278 * when this runqueue becomes "idle".
8280 init_idle(current
, smp_processor_id());
8282 * During early bootup we pretend to be a normal task:
8284 current
->sched_class
= &fair_sched_class
;
8286 scheduler_running
= 1;
8289 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8290 void __might_sleep(char *file
, int line
)
8293 static unsigned long prev_jiffy
; /* ratelimiting */
8295 if ((!in_atomic() && !irqs_disabled()) ||
8296 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
8298 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8300 prev_jiffy
= jiffies
;
8303 "BUG: sleeping function called from invalid context at %s:%d\n",
8306 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8307 in_atomic(), irqs_disabled(),
8308 current
->pid
, current
->comm
);
8310 debug_show_held_locks(current
);
8311 if (irqs_disabled())
8312 print_irqtrace_events(current
);
8316 EXPORT_SYMBOL(__might_sleep
);
8319 #ifdef CONFIG_MAGIC_SYSRQ
8320 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8324 update_rq_clock(rq
);
8325 on_rq
= p
->se
.on_rq
;
8327 deactivate_task(rq
, p
, 0);
8328 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8330 activate_task(rq
, p
, 0);
8331 resched_task(rq
->curr
);
8335 void normalize_rt_tasks(void)
8337 struct task_struct
*g
, *p
;
8338 unsigned long flags
;
8341 read_lock_irqsave(&tasklist_lock
, flags
);
8342 do_each_thread(g
, p
) {
8344 * Only normalize user tasks:
8349 p
->se
.exec_start
= 0;
8350 #ifdef CONFIG_SCHEDSTATS
8351 p
->se
.wait_start
= 0;
8352 p
->se
.sleep_start
= 0;
8353 p
->se
.block_start
= 0;
8358 * Renice negative nice level userspace
8361 if (TASK_NICE(p
) < 0 && p
->mm
)
8362 set_user_nice(p
, 0);
8366 spin_lock(&p
->pi_lock
);
8367 rq
= __task_rq_lock(p
);
8369 normalize_task(rq
, p
);
8371 __task_rq_unlock(rq
);
8372 spin_unlock(&p
->pi_lock
);
8373 } while_each_thread(g
, p
);
8375 read_unlock_irqrestore(&tasklist_lock
, flags
);
8378 #endif /* CONFIG_MAGIC_SYSRQ */
8382 * These functions are only useful for the IA64 MCA handling.
8384 * They can only be called when the whole system has been
8385 * stopped - every CPU needs to be quiescent, and no scheduling
8386 * activity can take place. Using them for anything else would
8387 * be a serious bug, and as a result, they aren't even visible
8388 * under any other configuration.
8392 * curr_task - return the current task for a given cpu.
8393 * @cpu: the processor in question.
8395 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8397 struct task_struct
*curr_task(int cpu
)
8399 return cpu_curr(cpu
);
8403 * set_curr_task - set the current task for a given cpu.
8404 * @cpu: the processor in question.
8405 * @p: the task pointer to set.
8407 * Description: This function must only be used when non-maskable interrupts
8408 * are serviced on a separate stack. It allows the architecture to switch the
8409 * notion of the current task on a cpu in a non-blocking manner. This function
8410 * must be called with all CPU's synchronized, and interrupts disabled, the
8411 * and caller must save the original value of the current task (see
8412 * curr_task() above) and restore that value before reenabling interrupts and
8413 * re-starting the system.
8415 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8417 void set_curr_task(int cpu
, struct task_struct
*p
)
8424 #ifdef CONFIG_FAIR_GROUP_SCHED
8425 static void free_fair_sched_group(struct task_group
*tg
)
8429 for_each_possible_cpu(i
) {
8431 kfree(tg
->cfs_rq
[i
]);
8441 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8443 struct cfs_rq
*cfs_rq
;
8444 struct sched_entity
*se
, *parent_se
;
8448 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8451 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8455 tg
->shares
= NICE_0_LOAD
;
8457 for_each_possible_cpu(i
) {
8460 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
),
8461 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8465 se
= kmalloc_node(sizeof(struct sched_entity
),
8466 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8470 parent_se
= parent
? parent
->se
[i
] : NULL
;
8471 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent_se
);
8480 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8482 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8483 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8486 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8488 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8490 #else /* !CONFG_FAIR_GROUP_SCHED */
8491 static inline void free_fair_sched_group(struct task_group
*tg
)
8496 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8501 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8505 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8508 #endif /* CONFIG_FAIR_GROUP_SCHED */
8510 #ifdef CONFIG_RT_GROUP_SCHED
8511 static void free_rt_sched_group(struct task_group
*tg
)
8515 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8517 for_each_possible_cpu(i
) {
8519 kfree(tg
->rt_rq
[i
]);
8521 kfree(tg
->rt_se
[i
]);
8529 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8531 struct rt_rq
*rt_rq
;
8532 struct sched_rt_entity
*rt_se
, *parent_se
;
8536 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8539 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8543 init_rt_bandwidth(&tg
->rt_bandwidth
,
8544 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8546 for_each_possible_cpu(i
) {
8549 rt_rq
= kmalloc_node(sizeof(struct rt_rq
),
8550 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8554 rt_se
= kmalloc_node(sizeof(struct sched_rt_entity
),
8555 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8559 parent_se
= parent
? parent
->rt_se
[i
] : NULL
;
8560 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent_se
);
8569 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8571 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8572 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8575 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8577 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8579 #else /* !CONFIG_RT_GROUP_SCHED */
8580 static inline void free_rt_sched_group(struct task_group
*tg
)
8585 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8590 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8594 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8597 #endif /* CONFIG_RT_GROUP_SCHED */
8599 #ifdef CONFIG_GROUP_SCHED
8600 static void free_sched_group(struct task_group
*tg
)
8602 free_fair_sched_group(tg
);
8603 free_rt_sched_group(tg
);
8607 /* allocate runqueue etc for a new task group */
8608 struct task_group
*sched_create_group(struct task_group
*parent
)
8610 struct task_group
*tg
;
8611 unsigned long flags
;
8614 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8616 return ERR_PTR(-ENOMEM
);
8618 if (!alloc_fair_sched_group(tg
, parent
))
8621 if (!alloc_rt_sched_group(tg
, parent
))
8624 spin_lock_irqsave(&task_group_lock
, flags
);
8625 for_each_possible_cpu(i
) {
8626 register_fair_sched_group(tg
, i
);
8627 register_rt_sched_group(tg
, i
);
8629 list_add_rcu(&tg
->list
, &task_groups
);
8631 WARN_ON(!parent
); /* root should already exist */
8633 tg
->parent
= parent
;
8634 INIT_LIST_HEAD(&tg
->children
);
8635 list_add_rcu(&tg
->siblings
, &parent
->children
);
8636 spin_unlock_irqrestore(&task_group_lock
, flags
);
8641 free_sched_group(tg
);
8642 return ERR_PTR(-ENOMEM
);
8645 /* rcu callback to free various structures associated with a task group */
8646 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8648 /* now it should be safe to free those cfs_rqs */
8649 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8652 /* Destroy runqueue etc associated with a task group */
8653 void sched_destroy_group(struct task_group
*tg
)
8655 unsigned long flags
;
8658 spin_lock_irqsave(&task_group_lock
, flags
);
8659 for_each_possible_cpu(i
) {
8660 unregister_fair_sched_group(tg
, i
);
8661 unregister_rt_sched_group(tg
, i
);
8663 list_del_rcu(&tg
->list
);
8664 list_del_rcu(&tg
->siblings
);
8665 spin_unlock_irqrestore(&task_group_lock
, flags
);
8667 /* wait for possible concurrent references to cfs_rqs complete */
8668 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8671 /* change task's runqueue when it moves between groups.
8672 * The caller of this function should have put the task in its new group
8673 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8674 * reflect its new group.
8676 void sched_move_task(struct task_struct
*tsk
)
8679 unsigned long flags
;
8682 rq
= task_rq_lock(tsk
, &flags
);
8684 update_rq_clock(rq
);
8686 running
= task_current(rq
, tsk
);
8687 on_rq
= tsk
->se
.on_rq
;
8690 dequeue_task(rq
, tsk
, 0);
8691 if (unlikely(running
))
8692 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8694 set_task_rq(tsk
, task_cpu(tsk
));
8696 #ifdef CONFIG_FAIR_GROUP_SCHED
8697 if (tsk
->sched_class
->moved_group
)
8698 tsk
->sched_class
->moved_group(tsk
);
8701 if (unlikely(running
))
8702 tsk
->sched_class
->set_curr_task(rq
);
8704 enqueue_task(rq
, tsk
, 0);
8706 task_rq_unlock(rq
, &flags
);
8708 #endif /* CONFIG_GROUP_SCHED */
8710 #ifdef CONFIG_FAIR_GROUP_SCHED
8711 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8713 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8718 dequeue_entity(cfs_rq
, se
, 0);
8720 se
->load
.weight
= shares
;
8721 se
->load
.inv_weight
= 0;
8724 enqueue_entity(cfs_rq
, se
, 0);
8727 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8729 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8730 struct rq
*rq
= cfs_rq
->rq
;
8731 unsigned long flags
;
8733 spin_lock_irqsave(&rq
->lock
, flags
);
8734 __set_se_shares(se
, shares
);
8735 spin_unlock_irqrestore(&rq
->lock
, flags
);
8738 static DEFINE_MUTEX(shares_mutex
);
8740 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8743 unsigned long flags
;
8746 * We can't change the weight of the root cgroup.
8751 if (shares
< MIN_SHARES
)
8752 shares
= MIN_SHARES
;
8753 else if (shares
> MAX_SHARES
)
8754 shares
= MAX_SHARES
;
8756 mutex_lock(&shares_mutex
);
8757 if (tg
->shares
== shares
)
8760 spin_lock_irqsave(&task_group_lock
, flags
);
8761 for_each_possible_cpu(i
)
8762 unregister_fair_sched_group(tg
, i
);
8763 list_del_rcu(&tg
->siblings
);
8764 spin_unlock_irqrestore(&task_group_lock
, flags
);
8766 /* wait for any ongoing reference to this group to finish */
8767 synchronize_sched();
8770 * Now we are free to modify the group's share on each cpu
8771 * w/o tripping rebalance_share or load_balance_fair.
8773 tg
->shares
= shares
;
8774 for_each_possible_cpu(i
) {
8778 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8779 set_se_shares(tg
->se
[i
], shares
);
8783 * Enable load balance activity on this group, by inserting it back on
8784 * each cpu's rq->leaf_cfs_rq_list.
8786 spin_lock_irqsave(&task_group_lock
, flags
);
8787 for_each_possible_cpu(i
)
8788 register_fair_sched_group(tg
, i
);
8789 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8790 spin_unlock_irqrestore(&task_group_lock
, flags
);
8792 mutex_unlock(&shares_mutex
);
8796 unsigned long sched_group_shares(struct task_group
*tg
)
8802 #ifdef CONFIG_RT_GROUP_SCHED
8804 * Ensure that the real time constraints are schedulable.
8806 static DEFINE_MUTEX(rt_constraints_mutex
);
8808 static unsigned long to_ratio(u64 period
, u64 runtime
)
8810 if (runtime
== RUNTIME_INF
)
8813 return div64_u64(runtime
<< 16, period
);
8816 #ifdef CONFIG_CGROUP_SCHED
8817 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8819 struct task_group
*tgi
, *parent
= tg
->parent
;
8820 unsigned long total
= 0;
8823 if (global_rt_period() < period
)
8826 return to_ratio(period
, runtime
) <
8827 to_ratio(global_rt_period(), global_rt_runtime());
8830 if (ktime_to_ns(parent
->rt_bandwidth
.rt_period
) < period
)
8834 list_for_each_entry_rcu(tgi
, &parent
->children
, siblings
) {
8838 total
+= to_ratio(ktime_to_ns(tgi
->rt_bandwidth
.rt_period
),
8839 tgi
->rt_bandwidth
.rt_runtime
);
8843 return total
+ to_ratio(period
, runtime
) <=
8844 to_ratio(ktime_to_ns(parent
->rt_bandwidth
.rt_period
),
8845 parent
->rt_bandwidth
.rt_runtime
);
8847 #elif defined CONFIG_USER_SCHED
8848 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8850 struct task_group
*tgi
;
8851 unsigned long total
= 0;
8852 unsigned long global_ratio
=
8853 to_ratio(global_rt_period(), global_rt_runtime());
8856 list_for_each_entry_rcu(tgi
, &task_groups
, list
) {
8860 total
+= to_ratio(ktime_to_ns(tgi
->rt_bandwidth
.rt_period
),
8861 tgi
->rt_bandwidth
.rt_runtime
);
8865 return total
+ to_ratio(period
, runtime
) < global_ratio
;
8869 /* Must be called with tasklist_lock held */
8870 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8872 struct task_struct
*g
, *p
;
8873 do_each_thread(g
, p
) {
8874 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8876 } while_each_thread(g
, p
);
8880 static int tg_set_bandwidth(struct task_group
*tg
,
8881 u64 rt_period
, u64 rt_runtime
)
8885 mutex_lock(&rt_constraints_mutex
);
8886 read_lock(&tasklist_lock
);
8887 if (rt_runtime
== 0 && tg_has_rt_tasks(tg
)) {
8891 if (!__rt_schedulable(tg
, rt_period
, rt_runtime
)) {
8896 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8897 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8898 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8900 for_each_possible_cpu(i
) {
8901 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8903 spin_lock(&rt_rq
->rt_runtime_lock
);
8904 rt_rq
->rt_runtime
= rt_runtime
;
8905 spin_unlock(&rt_rq
->rt_runtime_lock
);
8907 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8909 read_unlock(&tasklist_lock
);
8910 mutex_unlock(&rt_constraints_mutex
);
8915 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8917 u64 rt_runtime
, rt_period
;
8919 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8920 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8921 if (rt_runtime_us
< 0)
8922 rt_runtime
= RUNTIME_INF
;
8924 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8927 long sched_group_rt_runtime(struct task_group
*tg
)
8931 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8934 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8935 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8936 return rt_runtime_us
;
8939 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8941 u64 rt_runtime
, rt_period
;
8943 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8944 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8949 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8952 long sched_group_rt_period(struct task_group
*tg
)
8956 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8957 do_div(rt_period_us
, NSEC_PER_USEC
);
8958 return rt_period_us
;
8961 static int sched_rt_global_constraints(void)
8963 struct task_group
*tg
= &root_task_group
;
8964 u64 rt_runtime
, rt_period
;
8967 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8968 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8970 mutex_lock(&rt_constraints_mutex
);
8971 if (!__rt_schedulable(tg
, rt_period
, rt_runtime
))
8973 mutex_unlock(&rt_constraints_mutex
);
8977 #else /* !CONFIG_RT_GROUP_SCHED */
8978 static int sched_rt_global_constraints(void)
8980 unsigned long flags
;
8983 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8984 for_each_possible_cpu(i
) {
8985 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8987 spin_lock(&rt_rq
->rt_runtime_lock
);
8988 rt_rq
->rt_runtime
= global_rt_runtime();
8989 spin_unlock(&rt_rq
->rt_runtime_lock
);
8991 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8995 #endif /* CONFIG_RT_GROUP_SCHED */
8997 int sched_rt_handler(struct ctl_table
*table
, int write
,
8998 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
9002 int old_period
, old_runtime
;
9003 static DEFINE_MUTEX(mutex
);
9006 old_period
= sysctl_sched_rt_period
;
9007 old_runtime
= sysctl_sched_rt_runtime
;
9009 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
9011 if (!ret
&& write
) {
9012 ret
= sched_rt_global_constraints();
9014 sysctl_sched_rt_period
= old_period
;
9015 sysctl_sched_rt_runtime
= old_runtime
;
9017 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
9018 def_rt_bandwidth
.rt_period
=
9019 ns_to_ktime(global_rt_period());
9022 mutex_unlock(&mutex
);
9027 #ifdef CONFIG_CGROUP_SCHED
9029 /* return corresponding task_group object of a cgroup */
9030 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
9032 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
9033 struct task_group
, css
);
9036 static struct cgroup_subsys_state
*
9037 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9039 struct task_group
*tg
, *parent
;
9041 if (!cgrp
->parent
) {
9042 /* This is early initialization for the top cgroup */
9043 init_task_group
.css
.cgroup
= cgrp
;
9044 return &init_task_group
.css
;
9047 parent
= cgroup_tg(cgrp
->parent
);
9048 tg
= sched_create_group(parent
);
9050 return ERR_PTR(-ENOMEM
);
9052 /* Bind the cgroup to task_group object we just created */
9053 tg
->css
.cgroup
= cgrp
;
9059 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9061 struct task_group
*tg
= cgroup_tg(cgrp
);
9063 sched_destroy_group(tg
);
9067 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9068 struct task_struct
*tsk
)
9070 #ifdef CONFIG_RT_GROUP_SCHED
9071 /* Don't accept realtime tasks when there is no way for them to run */
9072 if (rt_task(tsk
) && cgroup_tg(cgrp
)->rt_bandwidth
.rt_runtime
== 0)
9075 /* We don't support RT-tasks being in separate groups */
9076 if (tsk
->sched_class
!= &fair_sched_class
)
9084 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9085 struct cgroup
*old_cont
, struct task_struct
*tsk
)
9087 sched_move_task(tsk
);
9090 #ifdef CONFIG_FAIR_GROUP_SCHED
9091 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9094 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
9097 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9099 struct task_group
*tg
= cgroup_tg(cgrp
);
9101 return (u64
) tg
->shares
;
9103 #endif /* CONFIG_FAIR_GROUP_SCHED */
9105 #ifdef CONFIG_RT_GROUP_SCHED
9106 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9109 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9112 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9114 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9117 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9120 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9123 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9125 return sched_group_rt_period(cgroup_tg(cgrp
));
9127 #endif /* CONFIG_RT_GROUP_SCHED */
9129 static struct cftype cpu_files
[] = {
9130 #ifdef CONFIG_FAIR_GROUP_SCHED
9133 .read_u64
= cpu_shares_read_u64
,
9134 .write_u64
= cpu_shares_write_u64
,
9137 #ifdef CONFIG_RT_GROUP_SCHED
9139 .name
= "rt_runtime_us",
9140 .read_s64
= cpu_rt_runtime_read
,
9141 .write_s64
= cpu_rt_runtime_write
,
9144 .name
= "rt_period_us",
9145 .read_u64
= cpu_rt_period_read_uint
,
9146 .write_u64
= cpu_rt_period_write_uint
,
9151 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
9153 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
9156 struct cgroup_subsys cpu_cgroup_subsys
= {
9158 .create
= cpu_cgroup_create
,
9159 .destroy
= cpu_cgroup_destroy
,
9160 .can_attach
= cpu_cgroup_can_attach
,
9161 .attach
= cpu_cgroup_attach
,
9162 .populate
= cpu_cgroup_populate
,
9163 .subsys_id
= cpu_cgroup_subsys_id
,
9167 #endif /* CONFIG_CGROUP_SCHED */
9169 #ifdef CONFIG_CGROUP_CPUACCT
9172 * CPU accounting code for task groups.
9174 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9175 * (balbir@in.ibm.com).
9178 /* track cpu usage of a group of tasks */
9180 struct cgroup_subsys_state css
;
9181 /* cpuusage holds pointer to a u64-type object on every cpu */
9185 struct cgroup_subsys cpuacct_subsys
;
9187 /* return cpu accounting group corresponding to this container */
9188 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
9190 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
9191 struct cpuacct
, css
);
9194 /* return cpu accounting group to which this task belongs */
9195 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
9197 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
9198 struct cpuacct
, css
);
9201 /* create a new cpu accounting group */
9202 static struct cgroup_subsys_state
*cpuacct_create(
9203 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9205 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
9208 return ERR_PTR(-ENOMEM
);
9210 ca
->cpuusage
= alloc_percpu(u64
);
9211 if (!ca
->cpuusage
) {
9213 return ERR_PTR(-ENOMEM
);
9219 /* destroy an existing cpu accounting group */
9221 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9223 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9225 free_percpu(ca
->cpuusage
);
9229 /* return total cpu usage (in nanoseconds) of a group */
9230 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9232 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9233 u64 totalcpuusage
= 0;
9236 for_each_possible_cpu(i
) {
9237 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9240 * Take rq->lock to make 64-bit addition safe on 32-bit
9243 spin_lock_irq(&cpu_rq(i
)->lock
);
9244 totalcpuusage
+= *cpuusage
;
9245 spin_unlock_irq(&cpu_rq(i
)->lock
);
9248 return totalcpuusage
;
9251 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9254 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9263 for_each_possible_cpu(i
) {
9264 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9266 spin_lock_irq(&cpu_rq(i
)->lock
);
9268 spin_unlock_irq(&cpu_rq(i
)->lock
);
9274 static struct cftype files
[] = {
9277 .read_u64
= cpuusage_read
,
9278 .write_u64
= cpuusage_write
,
9282 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9284 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9288 * charge this task's execution time to its accounting group.
9290 * called with rq->lock held.
9292 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9296 if (!cpuacct_subsys
.active
)
9301 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
9303 *cpuusage
+= cputime
;
9307 struct cgroup_subsys cpuacct_subsys
= {
9309 .create
= cpuacct_create
,
9310 .destroy
= cpuacct_destroy
,
9311 .populate
= cpuacct_populate
,
9312 .subsys_id
= cpuacct_subsys_id
,
9314 #endif /* CONFIG_CGROUP_CPUACCT */