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_UNLOCKED
;
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
))
760 if (strncmp(buf
, "NO_", 3) == 0) {
765 for (i
= 0; sched_feat_names
[i
]; i
++) {
766 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
768 sysctl_sched_features
&= ~(1UL << i
);
770 sysctl_sched_features
|= (1UL << i
);
775 if (!sched_feat_names
[i
])
783 static struct file_operations sched_feat_fops
= {
784 .open
= sched_feat_open
,
785 .read
= sched_feat_read
,
786 .write
= sched_feat_write
,
789 static __init
int sched_init_debug(void)
791 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
796 late_initcall(sched_init_debug
);
800 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
803 * Number of tasks to iterate in a single balance run.
804 * Limited because this is done with IRQs disabled.
806 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
809 * ratelimit for updating the group shares.
812 unsigned int sysctl_sched_shares_ratelimit
= 250000;
815 * period over which we measure -rt task cpu usage in us.
818 unsigned int sysctl_sched_rt_period
= 1000000;
820 static __read_mostly
int scheduler_running
;
823 * part of the period that we allow rt tasks to run in us.
826 int sysctl_sched_rt_runtime
= 950000;
828 static inline u64
global_rt_period(void)
830 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
833 static inline u64
global_rt_runtime(void)
835 if (sysctl_sched_rt_runtime
< 0)
838 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
841 #ifndef prepare_arch_switch
842 # define prepare_arch_switch(next) do { } while (0)
844 #ifndef finish_arch_switch
845 # define finish_arch_switch(prev) do { } while (0)
848 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
850 return rq
->curr
== p
;
853 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
854 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
856 return task_current(rq
, p
);
859 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
863 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
865 #ifdef CONFIG_DEBUG_SPINLOCK
866 /* this is a valid case when another task releases the spinlock */
867 rq
->lock
.owner
= current
;
870 * If we are tracking spinlock dependencies then we have to
871 * fix up the runqueue lock - which gets 'carried over' from
874 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
876 spin_unlock_irq(&rq
->lock
);
879 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
880 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
885 return task_current(rq
, p
);
889 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
893 * We can optimise this out completely for !SMP, because the
894 * SMP rebalancing from interrupt is the only thing that cares
899 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
900 spin_unlock_irq(&rq
->lock
);
902 spin_unlock(&rq
->lock
);
906 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
910 * After ->oncpu is cleared, the task can be moved to a different CPU.
911 * We must ensure this doesn't happen until the switch is completely
917 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
921 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
924 * __task_rq_lock - lock the runqueue a given task resides on.
925 * Must be called interrupts disabled.
927 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
931 struct rq
*rq
= task_rq(p
);
932 spin_lock(&rq
->lock
);
933 if (likely(rq
== task_rq(p
)))
935 spin_unlock(&rq
->lock
);
940 * task_rq_lock - lock the runqueue a given task resides on and disable
941 * interrupts. Note the ordering: we can safely lookup the task_rq without
942 * explicitly disabling preemption.
944 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
950 local_irq_save(*flags
);
952 spin_lock(&rq
->lock
);
953 if (likely(rq
== task_rq(p
)))
955 spin_unlock_irqrestore(&rq
->lock
, *flags
);
959 static void __task_rq_unlock(struct rq
*rq
)
962 spin_unlock(&rq
->lock
);
965 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
968 spin_unlock_irqrestore(&rq
->lock
, *flags
);
972 * this_rq_lock - lock this runqueue and disable interrupts.
974 static struct rq
*this_rq_lock(void)
981 spin_lock(&rq
->lock
);
986 #ifdef CONFIG_SCHED_HRTICK
988 * Use HR-timers to deliver accurate preemption points.
990 * Its all a bit involved since we cannot program an hrt while holding the
991 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
994 * When we get rescheduled we reprogram the hrtick_timer outside of the
1000 * - enabled by features
1001 * - hrtimer is actually high res
1003 static inline int hrtick_enabled(struct rq
*rq
)
1005 if (!sched_feat(HRTICK
))
1007 if (!cpu_active(cpu_of(rq
)))
1009 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1012 static void hrtick_clear(struct rq
*rq
)
1014 if (hrtimer_active(&rq
->hrtick_timer
))
1015 hrtimer_cancel(&rq
->hrtick_timer
);
1019 * High-resolution timer tick.
1020 * Runs from hardirq context with interrupts disabled.
1022 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1024 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1026 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1028 spin_lock(&rq
->lock
);
1029 update_rq_clock(rq
);
1030 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1031 spin_unlock(&rq
->lock
);
1033 return HRTIMER_NORESTART
;
1038 * called from hardirq (IPI) context
1040 static void __hrtick_start(void *arg
)
1042 struct rq
*rq
= arg
;
1044 spin_lock(&rq
->lock
);
1045 hrtimer_restart(&rq
->hrtick_timer
);
1046 rq
->hrtick_csd_pending
= 0;
1047 spin_unlock(&rq
->lock
);
1051 * Called to set the hrtick timer state.
1053 * called with rq->lock held and irqs disabled
1055 static void hrtick_start(struct rq
*rq
, u64 delay
)
1057 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1058 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1060 timer
->expires
= time
;
1062 if (rq
== this_rq()) {
1063 hrtimer_restart(timer
);
1064 } else if (!rq
->hrtick_csd_pending
) {
1065 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
);
1066 rq
->hrtick_csd_pending
= 1;
1071 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1073 int cpu
= (int)(long)hcpu
;
1076 case CPU_UP_CANCELED
:
1077 case CPU_UP_CANCELED_FROZEN
:
1078 case CPU_DOWN_PREPARE
:
1079 case CPU_DOWN_PREPARE_FROZEN
:
1081 case CPU_DEAD_FROZEN
:
1082 hrtick_clear(cpu_rq(cpu
));
1089 static __init
void init_hrtick(void)
1091 hotcpu_notifier(hotplug_hrtick
, 0);
1095 * Called to set the hrtick timer state.
1097 * called with rq->lock held and irqs disabled
1099 static void hrtick_start(struct rq
*rq
, u64 delay
)
1101 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
), HRTIMER_MODE_REL
);
1104 static void init_hrtick(void)
1107 #endif /* CONFIG_SMP */
1109 static void init_rq_hrtick(struct rq
*rq
)
1112 rq
->hrtick_csd_pending
= 0;
1114 rq
->hrtick_csd
.flags
= 0;
1115 rq
->hrtick_csd
.func
= __hrtick_start
;
1116 rq
->hrtick_csd
.info
= rq
;
1119 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1120 rq
->hrtick_timer
.function
= hrtick
;
1121 rq
->hrtick_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_PERCPU
;
1124 static inline void hrtick_clear(struct rq
*rq
)
1128 static inline void init_rq_hrtick(struct rq
*rq
)
1132 static inline void init_hrtick(void)
1138 * resched_task - mark a task 'to be rescheduled now'.
1140 * On UP this means the setting of the need_resched flag, on SMP it
1141 * might also involve a cross-CPU call to trigger the scheduler on
1146 #ifndef tsk_is_polling
1147 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1150 static void resched_task(struct task_struct
*p
)
1154 assert_spin_locked(&task_rq(p
)->lock
);
1156 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
1159 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
1162 if (cpu
== smp_processor_id())
1165 /* NEED_RESCHED must be visible before we test polling */
1167 if (!tsk_is_polling(p
))
1168 smp_send_reschedule(cpu
);
1171 static void resched_cpu(int cpu
)
1173 struct rq
*rq
= cpu_rq(cpu
);
1174 unsigned long flags
;
1176 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1178 resched_task(cpu_curr(cpu
));
1179 spin_unlock_irqrestore(&rq
->lock
, flags
);
1184 * When add_timer_on() enqueues a timer into the timer wheel of an
1185 * idle CPU then this timer might expire before the next timer event
1186 * which is scheduled to wake up that CPU. In case of a completely
1187 * idle system the next event might even be infinite time into the
1188 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1189 * leaves the inner idle loop so the newly added timer is taken into
1190 * account when the CPU goes back to idle and evaluates the timer
1191 * wheel for the next timer event.
1193 void wake_up_idle_cpu(int cpu
)
1195 struct rq
*rq
= cpu_rq(cpu
);
1197 if (cpu
== smp_processor_id())
1201 * This is safe, as this function is called with the timer
1202 * wheel base lock of (cpu) held. When the CPU is on the way
1203 * to idle and has not yet set rq->curr to idle then it will
1204 * be serialized on the timer wheel base lock and take the new
1205 * timer into account automatically.
1207 if (rq
->curr
!= rq
->idle
)
1211 * We can set TIF_RESCHED on the idle task of the other CPU
1212 * lockless. The worst case is that the other CPU runs the
1213 * idle task through an additional NOOP schedule()
1215 set_tsk_thread_flag(rq
->idle
, TIF_NEED_RESCHED
);
1217 /* NEED_RESCHED must be visible before we test polling */
1219 if (!tsk_is_polling(rq
->idle
))
1220 smp_send_reschedule(cpu
);
1222 #endif /* CONFIG_NO_HZ */
1224 #else /* !CONFIG_SMP */
1225 static void resched_task(struct task_struct
*p
)
1227 assert_spin_locked(&task_rq(p
)->lock
);
1228 set_tsk_need_resched(p
);
1230 #endif /* CONFIG_SMP */
1232 #if BITS_PER_LONG == 32
1233 # define WMULT_CONST (~0UL)
1235 # define WMULT_CONST (1UL << 32)
1238 #define WMULT_SHIFT 32
1241 * Shift right and round:
1243 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1246 * delta *= weight / lw
1248 static unsigned long
1249 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1250 struct load_weight
*lw
)
1254 if (!lw
->inv_weight
) {
1255 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1258 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1262 tmp
= (u64
)delta_exec
* weight
;
1264 * Check whether we'd overflow the 64-bit multiplication:
1266 if (unlikely(tmp
> WMULT_CONST
))
1267 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1270 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1272 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1275 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1281 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1288 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1289 * of tasks with abnormal "nice" values across CPUs the contribution that
1290 * each task makes to its run queue's load is weighted according to its
1291 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1292 * scaled version of the new time slice allocation that they receive on time
1296 #define WEIGHT_IDLEPRIO 2
1297 #define WMULT_IDLEPRIO (1 << 31)
1300 * Nice levels are multiplicative, with a gentle 10% change for every
1301 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1302 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1303 * that remained on nice 0.
1305 * The "10% effect" is relative and cumulative: from _any_ nice level,
1306 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1307 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1308 * If a task goes up by ~10% and another task goes down by ~10% then
1309 * the relative distance between them is ~25%.)
1311 static const int prio_to_weight
[40] = {
1312 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1313 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1314 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1315 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1316 /* 0 */ 1024, 820, 655, 526, 423,
1317 /* 5 */ 335, 272, 215, 172, 137,
1318 /* 10 */ 110, 87, 70, 56, 45,
1319 /* 15 */ 36, 29, 23, 18, 15,
1323 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1325 * In cases where the weight does not change often, we can use the
1326 * precalculated inverse to speed up arithmetics by turning divisions
1327 * into multiplications:
1329 static const u32 prio_to_wmult
[40] = {
1330 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1331 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1332 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1333 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1334 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1335 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1336 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1337 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1340 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1343 * runqueue iterator, to support SMP load-balancing between different
1344 * scheduling classes, without having to expose their internal data
1345 * structures to the load-balancing proper:
1347 struct rq_iterator
{
1349 struct task_struct
*(*start
)(void *);
1350 struct task_struct
*(*next
)(void *);
1354 static unsigned long
1355 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1356 unsigned long max_load_move
, struct sched_domain
*sd
,
1357 enum cpu_idle_type idle
, int *all_pinned
,
1358 int *this_best_prio
, struct rq_iterator
*iterator
);
1361 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1362 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1363 struct rq_iterator
*iterator
);
1366 #ifdef CONFIG_CGROUP_CPUACCT
1367 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1369 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1372 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1374 update_load_add(&rq
->load
, load
);
1377 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1379 update_load_sub(&rq
->load
, load
);
1383 static unsigned long source_load(int cpu
, int type
);
1384 static unsigned long target_load(int cpu
, int type
);
1385 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1387 static unsigned long cpu_avg_load_per_task(int cpu
)
1389 struct rq
*rq
= cpu_rq(cpu
);
1392 rq
->avg_load_per_task
= rq
->load
.weight
/ rq
->nr_running
;
1394 return rq
->avg_load_per_task
;
1397 #ifdef CONFIG_FAIR_GROUP_SCHED
1399 typedef void (*tg_visitor
)(struct task_group
*, int, struct sched_domain
*);
1402 * Iterate the full tree, calling @down when first entering a node and @up when
1403 * leaving it for the final time.
1406 walk_tg_tree(tg_visitor down
, tg_visitor up
, int cpu
, struct sched_domain
*sd
)
1408 struct task_group
*parent
, *child
;
1411 parent
= &root_task_group
;
1413 (*down
)(parent
, cpu
, sd
);
1414 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1421 (*up
)(parent
, cpu
, sd
);
1424 parent
= parent
->parent
;
1430 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1433 * Calculate and set the cpu's group shares.
1436 __update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1437 unsigned long sd_shares
, unsigned long sd_rq_weight
)
1440 unsigned long shares
;
1441 unsigned long rq_weight
;
1446 rq_weight
= tg
->cfs_rq
[cpu
]->load
.weight
;
1449 * If there are currently no tasks on the cpu pretend there is one of
1450 * average load so that when a new task gets to run here it will not
1451 * get delayed by group starvation.
1455 rq_weight
= NICE_0_LOAD
;
1458 if (unlikely(rq_weight
> sd_rq_weight
))
1459 rq_weight
= sd_rq_weight
;
1462 * \Sum shares * rq_weight
1463 * shares = -----------------------
1467 shares
= (sd_shares
* rq_weight
) / (sd_rq_weight
+ 1);
1470 * record the actual number of shares, not the boosted amount.
1472 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1473 tg
->cfs_rq
[cpu
]->rq_weight
= rq_weight
;
1475 if (shares
< MIN_SHARES
)
1476 shares
= MIN_SHARES
;
1477 else if (shares
> MAX_SHARES
)
1478 shares
= MAX_SHARES
;
1480 __set_se_shares(tg
->se
[cpu
], shares
);
1484 * Re-compute the task group their per cpu shares over the given domain.
1485 * This needs to be done in a bottom-up fashion because the rq weight of a
1486 * parent group depends on the shares of its child groups.
1489 tg_shares_up(struct task_group
*tg
, int cpu
, struct sched_domain
*sd
)
1491 unsigned long rq_weight
= 0;
1492 unsigned long shares
= 0;
1495 for_each_cpu_mask(i
, sd
->span
) {
1496 rq_weight
+= tg
->cfs_rq
[i
]->load
.weight
;
1497 shares
+= tg
->cfs_rq
[i
]->shares
;
1500 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1501 shares
= tg
->shares
;
1503 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1504 shares
= tg
->shares
;
1507 rq_weight
= cpus_weight(sd
->span
) * NICE_0_LOAD
;
1509 for_each_cpu_mask(i
, sd
->span
) {
1510 struct rq
*rq
= cpu_rq(i
);
1511 unsigned long flags
;
1513 spin_lock_irqsave(&rq
->lock
, flags
);
1514 __update_group_shares_cpu(tg
, i
, shares
, rq_weight
);
1515 spin_unlock_irqrestore(&rq
->lock
, flags
);
1520 * Compute the cpu's hierarchical load factor for each task group.
1521 * This needs to be done in a top-down fashion because the load of a child
1522 * group is a fraction of its parents load.
1525 tg_load_down(struct task_group
*tg
, int cpu
, struct sched_domain
*sd
)
1530 load
= cpu_rq(cpu
)->load
.weight
;
1532 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1533 load
*= tg
->cfs_rq
[cpu
]->shares
;
1534 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1537 tg
->cfs_rq
[cpu
]->h_load
= load
;
1541 tg_nop(struct task_group
*tg
, int cpu
, struct sched_domain
*sd
)
1545 static void update_shares(struct sched_domain
*sd
)
1547 u64 now
= cpu_clock(raw_smp_processor_id());
1548 s64 elapsed
= now
- sd
->last_update
;
1550 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1551 sd
->last_update
= now
;
1552 walk_tg_tree(tg_nop
, tg_shares_up
, 0, sd
);
1556 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1558 spin_unlock(&rq
->lock
);
1560 spin_lock(&rq
->lock
);
1563 static void update_h_load(int cpu
)
1565 walk_tg_tree(tg_load_down
, tg_nop
, cpu
, NULL
);
1570 static inline void update_shares(struct sched_domain
*sd
)
1574 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1582 #ifdef CONFIG_FAIR_GROUP_SCHED
1583 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1586 cfs_rq
->shares
= shares
;
1591 #include "sched_stats.h"
1592 #include "sched_idletask.c"
1593 #include "sched_fair.c"
1594 #include "sched_rt.c"
1595 #ifdef CONFIG_SCHED_DEBUG
1596 # include "sched_debug.c"
1599 #define sched_class_highest (&rt_sched_class)
1600 #define for_each_class(class) \
1601 for (class = sched_class_highest; class; class = class->next)
1603 static void inc_nr_running(struct rq
*rq
)
1608 static void dec_nr_running(struct rq
*rq
)
1613 static void set_load_weight(struct task_struct
*p
)
1615 if (task_has_rt_policy(p
)) {
1616 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1617 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1622 * SCHED_IDLE tasks get minimal weight:
1624 if (p
->policy
== SCHED_IDLE
) {
1625 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1626 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1630 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1631 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1634 static void update_avg(u64
*avg
, u64 sample
)
1636 s64 diff
= sample
- *avg
;
1640 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1642 sched_info_queued(p
);
1643 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1647 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1649 if (sleep
&& p
->se
.last_wakeup
) {
1650 update_avg(&p
->se
.avg_overlap
,
1651 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1652 p
->se
.last_wakeup
= 0;
1655 sched_info_dequeued(p
);
1656 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1661 * __normal_prio - return the priority that is based on the static prio
1663 static inline int __normal_prio(struct task_struct
*p
)
1665 return p
->static_prio
;
1669 * Calculate the expected normal priority: i.e. priority
1670 * without taking RT-inheritance into account. Might be
1671 * boosted by interactivity modifiers. Changes upon fork,
1672 * setprio syscalls, and whenever the interactivity
1673 * estimator recalculates.
1675 static inline int normal_prio(struct task_struct
*p
)
1679 if (task_has_rt_policy(p
))
1680 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1682 prio
= __normal_prio(p
);
1687 * Calculate the current priority, i.e. the priority
1688 * taken into account by the scheduler. This value might
1689 * be boosted by RT tasks, or might be boosted by
1690 * interactivity modifiers. Will be RT if the task got
1691 * RT-boosted. If not then it returns p->normal_prio.
1693 static int effective_prio(struct task_struct
*p
)
1695 p
->normal_prio
= normal_prio(p
);
1697 * If we are RT tasks or we were boosted to RT priority,
1698 * keep the priority unchanged. Otherwise, update priority
1699 * to the normal priority:
1701 if (!rt_prio(p
->prio
))
1702 return p
->normal_prio
;
1707 * activate_task - move a task to the runqueue.
1709 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1711 if (task_contributes_to_load(p
))
1712 rq
->nr_uninterruptible
--;
1714 enqueue_task(rq
, p
, wakeup
);
1719 * deactivate_task - remove a task from the runqueue.
1721 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1723 if (task_contributes_to_load(p
))
1724 rq
->nr_uninterruptible
++;
1726 dequeue_task(rq
, p
, sleep
);
1731 * task_curr - is this task currently executing on a CPU?
1732 * @p: the task in question.
1734 inline int task_curr(const struct task_struct
*p
)
1736 return cpu_curr(task_cpu(p
)) == p
;
1739 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1741 set_task_rq(p
, cpu
);
1744 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1745 * successfuly executed on another CPU. We must ensure that updates of
1746 * per-task data have been completed by this moment.
1749 task_thread_info(p
)->cpu
= cpu
;
1753 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1754 const struct sched_class
*prev_class
,
1755 int oldprio
, int running
)
1757 if (prev_class
!= p
->sched_class
) {
1758 if (prev_class
->switched_from
)
1759 prev_class
->switched_from(rq
, p
, running
);
1760 p
->sched_class
->switched_to(rq
, p
, running
);
1762 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1767 /* Used instead of source_load when we know the type == 0 */
1768 static unsigned long weighted_cpuload(const int cpu
)
1770 return cpu_rq(cpu
)->load
.weight
;
1774 * Is this task likely cache-hot:
1777 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1782 * Buddy candidates are cache hot:
1784 if (sched_feat(CACHE_HOT_BUDDY
) && (&p
->se
== cfs_rq_of(&p
->se
)->next
))
1787 if (p
->sched_class
!= &fair_sched_class
)
1790 if (sysctl_sched_migration_cost
== -1)
1792 if (sysctl_sched_migration_cost
== 0)
1795 delta
= now
- p
->se
.exec_start
;
1797 return delta
< (s64
)sysctl_sched_migration_cost
;
1801 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1803 int old_cpu
= task_cpu(p
);
1804 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1805 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1806 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1809 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1811 #ifdef CONFIG_SCHEDSTATS
1812 if (p
->se
.wait_start
)
1813 p
->se
.wait_start
-= clock_offset
;
1814 if (p
->se
.sleep_start
)
1815 p
->se
.sleep_start
-= clock_offset
;
1816 if (p
->se
.block_start
)
1817 p
->se
.block_start
-= clock_offset
;
1818 if (old_cpu
!= new_cpu
) {
1819 schedstat_inc(p
, se
.nr_migrations
);
1820 if (task_hot(p
, old_rq
->clock
, NULL
))
1821 schedstat_inc(p
, se
.nr_forced2_migrations
);
1824 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1825 new_cfsrq
->min_vruntime
;
1827 __set_task_cpu(p
, new_cpu
);
1830 struct migration_req
{
1831 struct list_head list
;
1833 struct task_struct
*task
;
1836 struct completion done
;
1840 * The task's runqueue lock must be held.
1841 * Returns true if you have to wait for migration thread.
1844 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1846 struct rq
*rq
= task_rq(p
);
1849 * If the task is not on a runqueue (and not running), then
1850 * it is sufficient to simply update the task's cpu field.
1852 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1853 set_task_cpu(p
, dest_cpu
);
1857 init_completion(&req
->done
);
1859 req
->dest_cpu
= dest_cpu
;
1860 list_add(&req
->list
, &rq
->migration_queue
);
1866 * wait_task_inactive - wait for a thread to unschedule.
1868 * If @match_state is nonzero, it's the @p->state value just checked and
1869 * not expected to change. If it changes, i.e. @p might have woken up,
1870 * then return zero. When we succeed in waiting for @p to be off its CPU,
1871 * we return a positive number (its total switch count). If a second call
1872 * a short while later returns the same number, the caller can be sure that
1873 * @p has remained unscheduled the whole time.
1875 * The caller must ensure that the task *will* unschedule sometime soon,
1876 * else this function might spin for a *long* time. This function can't
1877 * be called with interrupts off, or it may introduce deadlock with
1878 * smp_call_function() if an IPI is sent by the same process we are
1879 * waiting to become inactive.
1881 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1883 unsigned long flags
;
1890 * We do the initial early heuristics without holding
1891 * any task-queue locks at all. We'll only try to get
1892 * the runqueue lock when things look like they will
1898 * If the task is actively running on another CPU
1899 * still, just relax and busy-wait without holding
1902 * NOTE! Since we don't hold any locks, it's not
1903 * even sure that "rq" stays as the right runqueue!
1904 * But we don't care, since "task_running()" will
1905 * return false if the runqueue has changed and p
1906 * is actually now running somewhere else!
1908 while (task_running(rq
, p
)) {
1909 if (match_state
&& unlikely(p
->state
!= match_state
))
1915 * Ok, time to look more closely! We need the rq
1916 * lock now, to be *sure*. If we're wrong, we'll
1917 * just go back and repeat.
1919 rq
= task_rq_lock(p
, &flags
);
1920 running
= task_running(rq
, p
);
1921 on_rq
= p
->se
.on_rq
;
1923 if (!match_state
|| p
->state
== match_state
) {
1924 ncsw
= p
->nivcsw
+ p
->nvcsw
;
1925 if (unlikely(!ncsw
))
1928 task_rq_unlock(rq
, &flags
);
1931 * If it changed from the expected state, bail out now.
1933 if (unlikely(!ncsw
))
1937 * Was it really running after all now that we
1938 * checked with the proper locks actually held?
1940 * Oops. Go back and try again..
1942 if (unlikely(running
)) {
1948 * It's not enough that it's not actively running,
1949 * it must be off the runqueue _entirely_, and not
1952 * So if it wa still runnable (but just not actively
1953 * running right now), it's preempted, and we should
1954 * yield - it could be a while.
1956 if (unlikely(on_rq
)) {
1957 schedule_timeout_uninterruptible(1);
1962 * Ahh, all good. It wasn't running, and it wasn't
1963 * runnable, which means that it will never become
1964 * running in the future either. We're all done!
1973 * kick_process - kick a running thread to enter/exit the kernel
1974 * @p: the to-be-kicked thread
1976 * Cause a process which is running on another CPU to enter
1977 * kernel-mode, without any delay. (to get signals handled.)
1979 * NOTE: this function doesnt have to take the runqueue lock,
1980 * because all it wants to ensure is that the remote task enters
1981 * the kernel. If the IPI races and the task has been migrated
1982 * to another CPU then no harm is done and the purpose has been
1985 void kick_process(struct task_struct
*p
)
1991 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1992 smp_send_reschedule(cpu
);
1997 * Return a low guess at the load of a migration-source cpu weighted
1998 * according to the scheduling class and "nice" value.
2000 * We want to under-estimate the load of migration sources, to
2001 * balance conservatively.
2003 static unsigned long source_load(int cpu
, int type
)
2005 struct rq
*rq
= cpu_rq(cpu
);
2006 unsigned long total
= weighted_cpuload(cpu
);
2008 if (type
== 0 || !sched_feat(LB_BIAS
))
2011 return min(rq
->cpu_load
[type
-1], total
);
2015 * Return a high guess at the load of a migration-target cpu weighted
2016 * according to the scheduling class and "nice" value.
2018 static unsigned long target_load(int cpu
, int type
)
2020 struct rq
*rq
= cpu_rq(cpu
);
2021 unsigned long total
= weighted_cpuload(cpu
);
2023 if (type
== 0 || !sched_feat(LB_BIAS
))
2026 return max(rq
->cpu_load
[type
-1], total
);
2030 * find_idlest_group finds and returns the least busy CPU group within the
2033 static struct sched_group
*
2034 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2036 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2037 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2038 int load_idx
= sd
->forkexec_idx
;
2039 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2042 unsigned long load
, avg_load
;
2046 /* Skip over this group if it has no CPUs allowed */
2047 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
2050 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2052 /* Tally up the load of all CPUs in the group */
2055 for_each_cpu_mask_nr(i
, group
->cpumask
) {
2056 /* Bias balancing toward cpus of our domain */
2058 load
= source_load(i
, load_idx
);
2060 load
= target_load(i
, load_idx
);
2065 /* Adjust by relative CPU power of the group */
2066 avg_load
= sg_div_cpu_power(group
,
2067 avg_load
* SCHED_LOAD_SCALE
);
2070 this_load
= avg_load
;
2072 } else if (avg_load
< min_load
) {
2073 min_load
= avg_load
;
2076 } while (group
= group
->next
, group
!= sd
->groups
);
2078 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2084 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2087 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
,
2090 unsigned long load
, min_load
= ULONG_MAX
;
2094 /* Traverse only the allowed CPUs */
2095 cpus_and(*tmp
, group
->cpumask
, p
->cpus_allowed
);
2097 for_each_cpu_mask_nr(i
, *tmp
) {
2098 load
= weighted_cpuload(i
);
2100 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2110 * sched_balance_self: balance the current task (running on cpu) in domains
2111 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2114 * Balance, ie. select the least loaded group.
2116 * Returns the target CPU number, or the same CPU if no balancing is needed.
2118 * preempt must be disabled.
2120 static int sched_balance_self(int cpu
, int flag
)
2122 struct task_struct
*t
= current
;
2123 struct sched_domain
*tmp
, *sd
= NULL
;
2125 for_each_domain(cpu
, tmp
) {
2127 * If power savings logic is enabled for a domain, stop there.
2129 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2131 if (tmp
->flags
& flag
)
2139 cpumask_t span
, tmpmask
;
2140 struct sched_group
*group
;
2141 int new_cpu
, weight
;
2143 if (!(sd
->flags
& flag
)) {
2149 group
= find_idlest_group(sd
, t
, cpu
);
2155 new_cpu
= find_idlest_cpu(group
, t
, cpu
, &tmpmask
);
2156 if (new_cpu
== -1 || new_cpu
== cpu
) {
2157 /* Now try balancing at a lower domain level of cpu */
2162 /* Now try balancing at a lower domain level of new_cpu */
2165 weight
= cpus_weight(span
);
2166 for_each_domain(cpu
, tmp
) {
2167 if (weight
<= cpus_weight(tmp
->span
))
2169 if (tmp
->flags
& flag
)
2172 /* while loop will break here if sd == NULL */
2178 #endif /* CONFIG_SMP */
2181 * try_to_wake_up - wake up a thread
2182 * @p: the to-be-woken-up thread
2183 * @state: the mask of task states that can be woken
2184 * @sync: do a synchronous wakeup?
2186 * Put it on the run-queue if it's not already there. The "current"
2187 * thread is always on the run-queue (except when the actual
2188 * re-schedule is in progress), and as such you're allowed to do
2189 * the simpler "current->state = TASK_RUNNING" to mark yourself
2190 * runnable without the overhead of this.
2192 * returns failure only if the task is already active.
2194 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2196 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2197 unsigned long flags
;
2201 if (!sched_feat(SYNC_WAKEUPS
))
2205 if (sched_feat(LB_WAKEUP_UPDATE
)) {
2206 struct sched_domain
*sd
;
2208 this_cpu
= raw_smp_processor_id();
2211 for_each_domain(this_cpu
, sd
) {
2212 if (cpu_isset(cpu
, sd
->span
)) {
2221 rq
= task_rq_lock(p
, &flags
);
2222 old_state
= p
->state
;
2223 if (!(old_state
& state
))
2231 this_cpu
= smp_processor_id();
2234 if (unlikely(task_running(rq
, p
)))
2237 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2238 if (cpu
!= orig_cpu
) {
2239 set_task_cpu(p
, cpu
);
2240 task_rq_unlock(rq
, &flags
);
2241 /* might preempt at this point */
2242 rq
= task_rq_lock(p
, &flags
);
2243 old_state
= p
->state
;
2244 if (!(old_state
& state
))
2249 this_cpu
= smp_processor_id();
2253 #ifdef CONFIG_SCHEDSTATS
2254 schedstat_inc(rq
, ttwu_count
);
2255 if (cpu
== this_cpu
)
2256 schedstat_inc(rq
, ttwu_local
);
2258 struct sched_domain
*sd
;
2259 for_each_domain(this_cpu
, sd
) {
2260 if (cpu_isset(cpu
, sd
->span
)) {
2261 schedstat_inc(sd
, ttwu_wake_remote
);
2266 #endif /* CONFIG_SCHEDSTATS */
2269 #endif /* CONFIG_SMP */
2270 schedstat_inc(p
, se
.nr_wakeups
);
2272 schedstat_inc(p
, se
.nr_wakeups_sync
);
2273 if (orig_cpu
!= cpu
)
2274 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2275 if (cpu
== this_cpu
)
2276 schedstat_inc(p
, se
.nr_wakeups_local
);
2278 schedstat_inc(p
, se
.nr_wakeups_remote
);
2279 update_rq_clock(rq
);
2280 activate_task(rq
, p
, 1);
2284 trace_mark(kernel_sched_wakeup
,
2285 "pid %d state %ld ## rq %p task %p rq->curr %p",
2286 p
->pid
, p
->state
, rq
, p
, rq
->curr
);
2287 check_preempt_curr(rq
, p
, sync
);
2289 p
->state
= TASK_RUNNING
;
2291 if (p
->sched_class
->task_wake_up
)
2292 p
->sched_class
->task_wake_up(rq
, p
);
2295 current
->se
.last_wakeup
= current
->se
.sum_exec_runtime
;
2297 task_rq_unlock(rq
, &flags
);
2302 int wake_up_process(struct task_struct
*p
)
2304 return try_to_wake_up(p
, TASK_ALL
, 0);
2306 EXPORT_SYMBOL(wake_up_process
);
2308 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2310 return try_to_wake_up(p
, state
, 0);
2314 * Perform scheduler related setup for a newly forked process p.
2315 * p is forked by current.
2317 * __sched_fork() is basic setup used by init_idle() too:
2319 static void __sched_fork(struct task_struct
*p
)
2321 p
->se
.exec_start
= 0;
2322 p
->se
.sum_exec_runtime
= 0;
2323 p
->se
.prev_sum_exec_runtime
= 0;
2324 p
->se
.last_wakeup
= 0;
2325 p
->se
.avg_overlap
= 0;
2327 #ifdef CONFIG_SCHEDSTATS
2328 p
->se
.wait_start
= 0;
2329 p
->se
.sum_sleep_runtime
= 0;
2330 p
->se
.sleep_start
= 0;
2331 p
->se
.block_start
= 0;
2332 p
->se
.sleep_max
= 0;
2333 p
->se
.block_max
= 0;
2335 p
->se
.slice_max
= 0;
2339 INIT_LIST_HEAD(&p
->rt
.run_list
);
2341 INIT_LIST_HEAD(&p
->se
.group_node
);
2343 #ifdef CONFIG_PREEMPT_NOTIFIERS
2344 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2348 * We mark the process as running here, but have not actually
2349 * inserted it onto the runqueue yet. This guarantees that
2350 * nobody will actually run it, and a signal or other external
2351 * event cannot wake it up and insert it on the runqueue either.
2353 p
->state
= TASK_RUNNING
;
2357 * fork()/clone()-time setup:
2359 void sched_fork(struct task_struct
*p
, int clone_flags
)
2361 int cpu
= get_cpu();
2366 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2368 set_task_cpu(p
, cpu
);
2371 * Make sure we do not leak PI boosting priority to the child:
2373 p
->prio
= current
->normal_prio
;
2374 if (!rt_prio(p
->prio
))
2375 p
->sched_class
= &fair_sched_class
;
2377 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2378 if (likely(sched_info_on()))
2379 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2381 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2384 #ifdef CONFIG_PREEMPT
2385 /* Want to start with kernel preemption disabled. */
2386 task_thread_info(p
)->preempt_count
= 1;
2392 * wake_up_new_task - wake up a newly created task for the first time.
2394 * This function will do some initial scheduler statistics housekeeping
2395 * that must be done for every newly created context, then puts the task
2396 * on the runqueue and wakes it.
2398 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2400 unsigned long flags
;
2403 rq
= task_rq_lock(p
, &flags
);
2404 BUG_ON(p
->state
!= TASK_RUNNING
);
2405 update_rq_clock(rq
);
2407 p
->prio
= effective_prio(p
);
2409 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2410 activate_task(rq
, p
, 0);
2413 * Let the scheduling class do new task startup
2414 * management (if any):
2416 p
->sched_class
->task_new(rq
, p
);
2419 trace_mark(kernel_sched_wakeup_new
,
2420 "pid %d state %ld ## rq %p task %p rq->curr %p",
2421 p
->pid
, p
->state
, rq
, p
, rq
->curr
);
2422 check_preempt_curr(rq
, p
, 0);
2424 if (p
->sched_class
->task_wake_up
)
2425 p
->sched_class
->task_wake_up(rq
, p
);
2427 task_rq_unlock(rq
, &flags
);
2430 #ifdef CONFIG_PREEMPT_NOTIFIERS
2433 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2434 * @notifier: notifier struct to register
2436 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2438 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2440 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2443 * preempt_notifier_unregister - no longer interested in preemption notifications
2444 * @notifier: notifier struct to unregister
2446 * This is safe to call from within a preemption notifier.
2448 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2450 hlist_del(¬ifier
->link
);
2452 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2454 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2456 struct preempt_notifier
*notifier
;
2457 struct hlist_node
*node
;
2459 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2460 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2464 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2465 struct task_struct
*next
)
2467 struct preempt_notifier
*notifier
;
2468 struct hlist_node
*node
;
2470 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2471 notifier
->ops
->sched_out(notifier
, next
);
2474 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2476 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2481 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2482 struct task_struct
*next
)
2486 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2489 * prepare_task_switch - prepare to switch tasks
2490 * @rq: the runqueue preparing to switch
2491 * @prev: the current task that is being switched out
2492 * @next: the task we are going to switch to.
2494 * This is called with the rq lock held and interrupts off. It must
2495 * be paired with a subsequent finish_task_switch after the context
2498 * prepare_task_switch sets up locking and calls architecture specific
2502 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2503 struct task_struct
*next
)
2505 fire_sched_out_preempt_notifiers(prev
, next
);
2506 prepare_lock_switch(rq
, next
);
2507 prepare_arch_switch(next
);
2511 * finish_task_switch - clean up after a task-switch
2512 * @rq: runqueue associated with task-switch
2513 * @prev: the thread we just switched away from.
2515 * finish_task_switch must be called after the context switch, paired
2516 * with a prepare_task_switch call before the context switch.
2517 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2518 * and do any other architecture-specific cleanup actions.
2520 * Note that we may have delayed dropping an mm in context_switch(). If
2521 * so, we finish that here outside of the runqueue lock. (Doing it
2522 * with the lock held can cause deadlocks; see schedule() for
2525 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2526 __releases(rq
->lock
)
2528 struct mm_struct
*mm
= rq
->prev_mm
;
2534 * A task struct has one reference for the use as "current".
2535 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2536 * schedule one last time. The schedule call will never return, and
2537 * the scheduled task must drop that reference.
2538 * The test for TASK_DEAD must occur while the runqueue locks are
2539 * still held, otherwise prev could be scheduled on another cpu, die
2540 * there before we look at prev->state, and then the reference would
2542 * Manfred Spraul <manfred@colorfullife.com>
2544 prev_state
= prev
->state
;
2545 finish_arch_switch(prev
);
2546 finish_lock_switch(rq
, prev
);
2548 if (current
->sched_class
->post_schedule
)
2549 current
->sched_class
->post_schedule(rq
);
2552 fire_sched_in_preempt_notifiers(current
);
2555 if (unlikely(prev_state
== TASK_DEAD
)) {
2557 * Remove function-return probe instances associated with this
2558 * task and put them back on the free list.
2560 kprobe_flush_task(prev
);
2561 put_task_struct(prev
);
2566 * schedule_tail - first thing a freshly forked thread must call.
2567 * @prev: the thread we just switched away from.
2569 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2570 __releases(rq
->lock
)
2572 struct rq
*rq
= this_rq();
2574 finish_task_switch(rq
, prev
);
2575 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2576 /* In this case, finish_task_switch does not reenable preemption */
2579 if (current
->set_child_tid
)
2580 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2584 * context_switch - switch to the new MM and the new
2585 * thread's register state.
2588 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2589 struct task_struct
*next
)
2591 struct mm_struct
*mm
, *oldmm
;
2593 prepare_task_switch(rq
, prev
, next
);
2594 trace_mark(kernel_sched_schedule
,
2595 "prev_pid %d next_pid %d prev_state %ld "
2596 "## rq %p prev %p next %p",
2597 prev
->pid
, next
->pid
, prev
->state
,
2600 oldmm
= prev
->active_mm
;
2602 * For paravirt, this is coupled with an exit in switch_to to
2603 * combine the page table reload and the switch backend into
2606 arch_enter_lazy_cpu_mode();
2608 if (unlikely(!mm
)) {
2609 next
->active_mm
= oldmm
;
2610 atomic_inc(&oldmm
->mm_count
);
2611 enter_lazy_tlb(oldmm
, next
);
2613 switch_mm(oldmm
, mm
, next
);
2615 if (unlikely(!prev
->mm
)) {
2616 prev
->active_mm
= NULL
;
2617 rq
->prev_mm
= oldmm
;
2620 * Since the runqueue lock will be released by the next
2621 * task (which is an invalid locking op but in the case
2622 * of the scheduler it's an obvious special-case), so we
2623 * do an early lockdep release here:
2625 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2626 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2629 /* Here we just switch the register state and the stack. */
2630 switch_to(prev
, next
, prev
);
2634 * this_rq must be evaluated again because prev may have moved
2635 * CPUs since it called schedule(), thus the 'rq' on its stack
2636 * frame will be invalid.
2638 finish_task_switch(this_rq(), prev
);
2642 * nr_running, nr_uninterruptible and nr_context_switches:
2644 * externally visible scheduler statistics: current number of runnable
2645 * threads, current number of uninterruptible-sleeping threads, total
2646 * number of context switches performed since bootup.
2648 unsigned long nr_running(void)
2650 unsigned long i
, sum
= 0;
2652 for_each_online_cpu(i
)
2653 sum
+= cpu_rq(i
)->nr_running
;
2658 unsigned long nr_uninterruptible(void)
2660 unsigned long i
, sum
= 0;
2662 for_each_possible_cpu(i
)
2663 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2666 * Since we read the counters lockless, it might be slightly
2667 * inaccurate. Do not allow it to go below zero though:
2669 if (unlikely((long)sum
< 0))
2675 unsigned long long nr_context_switches(void)
2678 unsigned long long sum
= 0;
2680 for_each_possible_cpu(i
)
2681 sum
+= cpu_rq(i
)->nr_switches
;
2686 unsigned long nr_iowait(void)
2688 unsigned long i
, sum
= 0;
2690 for_each_possible_cpu(i
)
2691 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2696 unsigned long nr_active(void)
2698 unsigned long i
, running
= 0, uninterruptible
= 0;
2700 for_each_online_cpu(i
) {
2701 running
+= cpu_rq(i
)->nr_running
;
2702 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2705 if (unlikely((long)uninterruptible
< 0))
2706 uninterruptible
= 0;
2708 return running
+ uninterruptible
;
2712 * Update rq->cpu_load[] statistics. This function is usually called every
2713 * scheduler tick (TICK_NSEC).
2715 static void update_cpu_load(struct rq
*this_rq
)
2717 unsigned long this_load
= this_rq
->load
.weight
;
2720 this_rq
->nr_load_updates
++;
2722 /* Update our load: */
2723 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2724 unsigned long old_load
, new_load
;
2726 /* scale is effectively 1 << i now, and >> i divides by scale */
2728 old_load
= this_rq
->cpu_load
[i
];
2729 new_load
= this_load
;
2731 * Round up the averaging division if load is increasing. This
2732 * prevents us from getting stuck on 9 if the load is 10, for
2735 if (new_load
> old_load
)
2736 new_load
+= scale
-1;
2737 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2744 * double_rq_lock - safely lock two runqueues
2746 * Note this does not disable interrupts like task_rq_lock,
2747 * you need to do so manually before calling.
2749 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2750 __acquires(rq1
->lock
)
2751 __acquires(rq2
->lock
)
2753 BUG_ON(!irqs_disabled());
2755 spin_lock(&rq1
->lock
);
2756 __acquire(rq2
->lock
); /* Fake it out ;) */
2759 spin_lock(&rq1
->lock
);
2760 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
2762 spin_lock(&rq2
->lock
);
2763 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
2766 update_rq_clock(rq1
);
2767 update_rq_clock(rq2
);
2771 * double_rq_unlock - safely unlock two runqueues
2773 * Note this does not restore interrupts like task_rq_unlock,
2774 * you need to do so manually after calling.
2776 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2777 __releases(rq1
->lock
)
2778 __releases(rq2
->lock
)
2780 spin_unlock(&rq1
->lock
);
2782 spin_unlock(&rq2
->lock
);
2784 __release(rq2
->lock
);
2788 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2790 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2791 __releases(this_rq
->lock
)
2792 __acquires(busiest
->lock
)
2793 __acquires(this_rq
->lock
)
2797 if (unlikely(!irqs_disabled())) {
2798 /* printk() doesn't work good under rq->lock */
2799 spin_unlock(&this_rq
->lock
);
2802 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2803 if (busiest
< this_rq
) {
2804 spin_unlock(&this_rq
->lock
);
2805 spin_lock(&busiest
->lock
);
2806 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
2809 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
2814 static void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2815 __releases(busiest
->lock
)
2817 spin_unlock(&busiest
->lock
);
2818 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
2822 * If dest_cpu is allowed for this process, migrate the task to it.
2823 * This is accomplished by forcing the cpu_allowed mask to only
2824 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2825 * the cpu_allowed mask is restored.
2827 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2829 struct migration_req req
;
2830 unsigned long flags
;
2833 rq
= task_rq_lock(p
, &flags
);
2834 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2835 || unlikely(!cpu_active(dest_cpu
)))
2838 /* force the process onto the specified CPU */
2839 if (migrate_task(p
, dest_cpu
, &req
)) {
2840 /* Need to wait for migration thread (might exit: take ref). */
2841 struct task_struct
*mt
= rq
->migration_thread
;
2843 get_task_struct(mt
);
2844 task_rq_unlock(rq
, &flags
);
2845 wake_up_process(mt
);
2846 put_task_struct(mt
);
2847 wait_for_completion(&req
.done
);
2852 task_rq_unlock(rq
, &flags
);
2856 * sched_exec - execve() is a valuable balancing opportunity, because at
2857 * this point the task has the smallest effective memory and cache footprint.
2859 void sched_exec(void)
2861 int new_cpu
, this_cpu
= get_cpu();
2862 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2864 if (new_cpu
!= this_cpu
)
2865 sched_migrate_task(current
, new_cpu
);
2869 * pull_task - move a task from a remote runqueue to the local runqueue.
2870 * Both runqueues must be locked.
2872 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2873 struct rq
*this_rq
, int this_cpu
)
2875 deactivate_task(src_rq
, p
, 0);
2876 set_task_cpu(p
, this_cpu
);
2877 activate_task(this_rq
, p
, 0);
2879 * Note that idle threads have a prio of MAX_PRIO, for this test
2880 * to be always true for them.
2882 check_preempt_curr(this_rq
, p
, 0);
2886 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2889 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2890 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2894 * We do not migrate tasks that are:
2895 * 1) running (obviously), or
2896 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2897 * 3) are cache-hot on their current CPU.
2899 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2900 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2905 if (task_running(rq
, p
)) {
2906 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2911 * Aggressive migration if:
2912 * 1) task is cache cold, or
2913 * 2) too many balance attempts have failed.
2916 if (!task_hot(p
, rq
->clock
, sd
) ||
2917 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2918 #ifdef CONFIG_SCHEDSTATS
2919 if (task_hot(p
, rq
->clock
, sd
)) {
2920 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2921 schedstat_inc(p
, se
.nr_forced_migrations
);
2927 if (task_hot(p
, rq
->clock
, sd
)) {
2928 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2934 static unsigned long
2935 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2936 unsigned long max_load_move
, struct sched_domain
*sd
,
2937 enum cpu_idle_type idle
, int *all_pinned
,
2938 int *this_best_prio
, struct rq_iterator
*iterator
)
2940 int loops
= 0, pulled
= 0, pinned
= 0;
2941 struct task_struct
*p
;
2942 long rem_load_move
= max_load_move
;
2944 if (max_load_move
== 0)
2950 * Start the load-balancing iterator:
2952 p
= iterator
->start(iterator
->arg
);
2954 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2957 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
2958 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2959 p
= iterator
->next(iterator
->arg
);
2963 pull_task(busiest
, p
, this_rq
, this_cpu
);
2965 rem_load_move
-= p
->se
.load
.weight
;
2968 * We only want to steal up to the prescribed amount of weighted load.
2970 if (rem_load_move
> 0) {
2971 if (p
->prio
< *this_best_prio
)
2972 *this_best_prio
= p
->prio
;
2973 p
= iterator
->next(iterator
->arg
);
2978 * Right now, this is one of only two places pull_task() is called,
2979 * so we can safely collect pull_task() stats here rather than
2980 * inside pull_task().
2982 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2985 *all_pinned
= pinned
;
2987 return max_load_move
- rem_load_move
;
2991 * move_tasks tries to move up to max_load_move weighted load from busiest to
2992 * this_rq, as part of a balancing operation within domain "sd".
2993 * Returns 1 if successful and 0 otherwise.
2995 * Called with both runqueues locked.
2997 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2998 unsigned long max_load_move
,
2999 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3002 const struct sched_class
*class = sched_class_highest
;
3003 unsigned long total_load_moved
= 0;
3004 int this_best_prio
= this_rq
->curr
->prio
;
3008 class->load_balance(this_rq
, this_cpu
, busiest
,
3009 max_load_move
- total_load_moved
,
3010 sd
, idle
, all_pinned
, &this_best_prio
);
3011 class = class->next
;
3013 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3016 } while (class && max_load_move
> total_load_moved
);
3018 return total_load_moved
> 0;
3022 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3023 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3024 struct rq_iterator
*iterator
)
3026 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3030 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3031 pull_task(busiest
, p
, this_rq
, this_cpu
);
3033 * Right now, this is only the second place pull_task()
3034 * is called, so we can safely collect pull_task()
3035 * stats here rather than inside pull_task().
3037 schedstat_inc(sd
, lb_gained
[idle
]);
3041 p
= iterator
->next(iterator
->arg
);
3048 * move_one_task tries to move exactly one task from busiest to this_rq, as
3049 * part of active balancing operations within "domain".
3050 * Returns 1 if successful and 0 otherwise.
3052 * Called with both runqueues locked.
3054 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3055 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3057 const struct sched_class
*class;
3059 for (class = sched_class_highest
; class; class = class->next
)
3060 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3067 * find_busiest_group finds and returns the busiest CPU group within the
3068 * domain. It calculates and returns the amount of weighted load which
3069 * should be moved to restore balance via the imbalance parameter.
3071 static struct sched_group
*
3072 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3073 unsigned long *imbalance
, enum cpu_idle_type idle
,
3074 int *sd_idle
, const cpumask_t
*cpus
, int *balance
)
3076 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
3077 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
3078 unsigned long max_pull
;
3079 unsigned long busiest_load_per_task
, busiest_nr_running
;
3080 unsigned long this_load_per_task
, this_nr_running
;
3081 int load_idx
, group_imb
= 0;
3082 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3083 int power_savings_balance
= 1;
3084 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
3085 unsigned long min_nr_running
= ULONG_MAX
;
3086 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
3089 max_load
= this_load
= total_load
= total_pwr
= 0;
3090 busiest_load_per_task
= busiest_nr_running
= 0;
3091 this_load_per_task
= this_nr_running
= 0;
3093 if (idle
== CPU_NOT_IDLE
)
3094 load_idx
= sd
->busy_idx
;
3095 else if (idle
== CPU_NEWLY_IDLE
)
3096 load_idx
= sd
->newidle_idx
;
3098 load_idx
= sd
->idle_idx
;
3101 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
3104 int __group_imb
= 0;
3105 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3106 unsigned long sum_nr_running
, sum_weighted_load
;
3107 unsigned long sum_avg_load_per_task
;
3108 unsigned long avg_load_per_task
;
3110 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
3113 balance_cpu
= first_cpu(group
->cpumask
);
3115 /* Tally up the load of all CPUs in the group */
3116 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
3117 sum_avg_load_per_task
= avg_load_per_task
= 0;
3120 min_cpu_load
= ~0UL;
3122 for_each_cpu_mask_nr(i
, group
->cpumask
) {
3125 if (!cpu_isset(i
, *cpus
))
3130 if (*sd_idle
&& rq
->nr_running
)
3133 /* Bias balancing toward cpus of our domain */
3135 if (idle_cpu(i
) && !first_idle_cpu
) {
3140 load
= target_load(i
, load_idx
);
3142 load
= source_load(i
, load_idx
);
3143 if (load
> max_cpu_load
)
3144 max_cpu_load
= load
;
3145 if (min_cpu_load
> load
)
3146 min_cpu_load
= load
;
3150 sum_nr_running
+= rq
->nr_running
;
3151 sum_weighted_load
+= weighted_cpuload(i
);
3153 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3157 * First idle cpu or the first cpu(busiest) in this sched group
3158 * is eligible for doing load balancing at this and above
3159 * domains. In the newly idle case, we will allow all the cpu's
3160 * to do the newly idle load balance.
3162 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3163 balance_cpu
!= this_cpu
&& balance
) {
3168 total_load
+= avg_load
;
3169 total_pwr
+= group
->__cpu_power
;
3171 /* Adjust by relative CPU power of the group */
3172 avg_load
= sg_div_cpu_power(group
,
3173 avg_load
* SCHED_LOAD_SCALE
);
3177 * Consider the group unbalanced when the imbalance is larger
3178 * than the average weight of two tasks.
3180 * APZ: with cgroup the avg task weight can vary wildly and
3181 * might not be a suitable number - should we keep a
3182 * normalized nr_running number somewhere that negates
3185 avg_load_per_task
= sg_div_cpu_power(group
,
3186 sum_avg_load_per_task
* SCHED_LOAD_SCALE
);
3188 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3191 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3194 this_load
= avg_load
;
3196 this_nr_running
= sum_nr_running
;
3197 this_load_per_task
= sum_weighted_load
;
3198 } else if (avg_load
> max_load
&&
3199 (sum_nr_running
> group_capacity
|| __group_imb
)) {
3200 max_load
= avg_load
;
3202 busiest_nr_running
= sum_nr_running
;
3203 busiest_load_per_task
= sum_weighted_load
;
3204 group_imb
= __group_imb
;
3207 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3209 * Busy processors will not participate in power savings
3212 if (idle
== CPU_NOT_IDLE
||
3213 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3217 * If the local group is idle or completely loaded
3218 * no need to do power savings balance at this domain
3220 if (local_group
&& (this_nr_running
>= group_capacity
||
3222 power_savings_balance
= 0;
3225 * If a group is already running at full capacity or idle,
3226 * don't include that group in power savings calculations
3228 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
3233 * Calculate the group which has the least non-idle load.
3234 * This is the group from where we need to pick up the load
3237 if ((sum_nr_running
< min_nr_running
) ||
3238 (sum_nr_running
== min_nr_running
&&
3239 first_cpu(group
->cpumask
) <
3240 first_cpu(group_min
->cpumask
))) {
3242 min_nr_running
= sum_nr_running
;
3243 min_load_per_task
= sum_weighted_load
/
3248 * Calculate the group which is almost near its
3249 * capacity but still has some space to pick up some load
3250 * from other group and save more power
3252 if (sum_nr_running
<= group_capacity
- 1) {
3253 if (sum_nr_running
> leader_nr_running
||
3254 (sum_nr_running
== leader_nr_running
&&
3255 first_cpu(group
->cpumask
) >
3256 first_cpu(group_leader
->cpumask
))) {
3257 group_leader
= group
;
3258 leader_nr_running
= sum_nr_running
;
3263 group
= group
->next
;
3264 } while (group
!= sd
->groups
);
3266 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
3269 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
3271 if (this_load
>= avg_load
||
3272 100*max_load
<= sd
->imbalance_pct
*this_load
)
3275 busiest_load_per_task
/= busiest_nr_running
;
3277 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
3280 * We're trying to get all the cpus to the average_load, so we don't
3281 * want to push ourselves above the average load, nor do we wish to
3282 * reduce the max loaded cpu below the average load, as either of these
3283 * actions would just result in more rebalancing later, and ping-pong
3284 * tasks around. Thus we look for the minimum possible imbalance.
3285 * Negative imbalances (*we* are more loaded than anyone else) will
3286 * be counted as no imbalance for these purposes -- we can't fix that
3287 * by pulling tasks to us. Be careful of negative numbers as they'll
3288 * appear as very large values with unsigned longs.
3290 if (max_load
<= busiest_load_per_task
)
3294 * In the presence of smp nice balancing, certain scenarios can have
3295 * max load less than avg load(as we skip the groups at or below
3296 * its cpu_power, while calculating max_load..)
3298 if (max_load
< avg_load
) {
3300 goto small_imbalance
;
3303 /* Don't want to pull so many tasks that a group would go idle */
3304 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
3306 /* How much load to actually move to equalise the imbalance */
3307 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
3308 (avg_load
- this_load
) * this->__cpu_power
)
3312 * if *imbalance is less than the average load per runnable task
3313 * there is no gaurantee that any tasks will be moved so we'll have
3314 * a think about bumping its value to force at least one task to be
3317 if (*imbalance
< busiest_load_per_task
) {
3318 unsigned long tmp
, pwr_now
, pwr_move
;
3322 pwr_move
= pwr_now
= 0;
3324 if (this_nr_running
) {
3325 this_load_per_task
/= this_nr_running
;
3326 if (busiest_load_per_task
> this_load_per_task
)
3329 this_load_per_task
= cpu_avg_load_per_task(this_cpu
);
3331 if (max_load
- this_load
+ 2*busiest_load_per_task
>=
3332 busiest_load_per_task
* imbn
) {
3333 *imbalance
= busiest_load_per_task
;
3338 * OK, we don't have enough imbalance to justify moving tasks,
3339 * however we may be able to increase total CPU power used by
3343 pwr_now
+= busiest
->__cpu_power
*
3344 min(busiest_load_per_task
, max_load
);
3345 pwr_now
+= this->__cpu_power
*
3346 min(this_load_per_task
, this_load
);
3347 pwr_now
/= SCHED_LOAD_SCALE
;
3349 /* Amount of load we'd subtract */
3350 tmp
= sg_div_cpu_power(busiest
,
3351 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3353 pwr_move
+= busiest
->__cpu_power
*
3354 min(busiest_load_per_task
, max_load
- tmp
);
3356 /* Amount of load we'd add */
3357 if (max_load
* busiest
->__cpu_power
<
3358 busiest_load_per_task
* SCHED_LOAD_SCALE
)
3359 tmp
= sg_div_cpu_power(this,
3360 max_load
* busiest
->__cpu_power
);
3362 tmp
= sg_div_cpu_power(this,
3363 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3364 pwr_move
+= this->__cpu_power
*
3365 min(this_load_per_task
, this_load
+ tmp
);
3366 pwr_move
/= SCHED_LOAD_SCALE
;
3368 /* Move if we gain throughput */
3369 if (pwr_move
> pwr_now
)
3370 *imbalance
= busiest_load_per_task
;
3376 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3377 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3380 if (this == group_leader
&& group_leader
!= group_min
) {
3381 *imbalance
= min_load_per_task
;
3391 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3394 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3395 unsigned long imbalance
, const cpumask_t
*cpus
)
3397 struct rq
*busiest
= NULL
, *rq
;
3398 unsigned long max_load
= 0;
3401 for_each_cpu_mask_nr(i
, group
->cpumask
) {
3404 if (!cpu_isset(i
, *cpus
))
3408 wl
= weighted_cpuload(i
);
3410 if (rq
->nr_running
== 1 && wl
> imbalance
)
3413 if (wl
> max_load
) {
3423 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3424 * so long as it is large enough.
3426 #define MAX_PINNED_INTERVAL 512
3429 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3430 * tasks if there is an imbalance.
3432 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3433 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3434 int *balance
, cpumask_t
*cpus
)
3436 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3437 struct sched_group
*group
;
3438 unsigned long imbalance
;
3440 unsigned long flags
;
3445 * When power savings policy is enabled for the parent domain, idle
3446 * sibling can pick up load irrespective of busy siblings. In this case,
3447 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3448 * portraying it as CPU_NOT_IDLE.
3450 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3451 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3454 schedstat_inc(sd
, lb_count
[idle
]);
3458 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3465 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3469 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3471 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3475 BUG_ON(busiest
== this_rq
);
3477 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3480 if (busiest
->nr_running
> 1) {
3482 * Attempt to move tasks. If find_busiest_group has found
3483 * an imbalance but busiest->nr_running <= 1, the group is
3484 * still unbalanced. ld_moved simply stays zero, so it is
3485 * correctly treated as an imbalance.
3487 local_irq_save(flags
);
3488 double_rq_lock(this_rq
, busiest
);
3489 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3490 imbalance
, sd
, idle
, &all_pinned
);
3491 double_rq_unlock(this_rq
, busiest
);
3492 local_irq_restore(flags
);
3495 * some other cpu did the load balance for us.
3497 if (ld_moved
&& this_cpu
!= smp_processor_id())
3498 resched_cpu(this_cpu
);
3500 /* All tasks on this runqueue were pinned by CPU affinity */
3501 if (unlikely(all_pinned
)) {
3502 cpu_clear(cpu_of(busiest
), *cpus
);
3503 if (!cpus_empty(*cpus
))
3510 schedstat_inc(sd
, lb_failed
[idle
]);
3511 sd
->nr_balance_failed
++;
3513 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3515 spin_lock_irqsave(&busiest
->lock
, flags
);
3517 /* don't kick the migration_thread, if the curr
3518 * task on busiest cpu can't be moved to this_cpu
3520 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
3521 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3523 goto out_one_pinned
;
3526 if (!busiest
->active_balance
) {
3527 busiest
->active_balance
= 1;
3528 busiest
->push_cpu
= this_cpu
;
3531 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3533 wake_up_process(busiest
->migration_thread
);
3536 * We've kicked active balancing, reset the failure
3539 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3542 sd
->nr_balance_failed
= 0;
3544 if (likely(!active_balance
)) {
3545 /* We were unbalanced, so reset the balancing interval */
3546 sd
->balance_interval
= sd
->min_interval
;
3549 * If we've begun active balancing, start to back off. This
3550 * case may not be covered by the all_pinned logic if there
3551 * is only 1 task on the busy runqueue (because we don't call
3554 if (sd
->balance_interval
< sd
->max_interval
)
3555 sd
->balance_interval
*= 2;
3558 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3559 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3565 schedstat_inc(sd
, lb_balanced
[idle
]);
3567 sd
->nr_balance_failed
= 0;
3570 /* tune up the balancing interval */
3571 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3572 (sd
->balance_interval
< sd
->max_interval
))
3573 sd
->balance_interval
*= 2;
3575 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3576 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3587 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3588 * tasks if there is an imbalance.
3590 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3591 * this_rq is locked.
3594 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
,
3597 struct sched_group
*group
;
3598 struct rq
*busiest
= NULL
;
3599 unsigned long imbalance
;
3607 * When power savings policy is enabled for the parent domain, idle
3608 * sibling can pick up load irrespective of busy siblings. In this case,
3609 * let the state of idle sibling percolate up as IDLE, instead of
3610 * portraying it as CPU_NOT_IDLE.
3612 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3613 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3616 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3618 update_shares_locked(this_rq
, sd
);
3619 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3620 &sd_idle
, cpus
, NULL
);
3622 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3626 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
3628 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3632 BUG_ON(busiest
== this_rq
);
3634 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3637 if (busiest
->nr_running
> 1) {
3638 /* Attempt to move tasks */
3639 double_lock_balance(this_rq
, busiest
);
3640 /* this_rq->clock is already updated */
3641 update_rq_clock(busiest
);
3642 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3643 imbalance
, sd
, CPU_NEWLY_IDLE
,
3645 double_unlock_balance(this_rq
, busiest
);
3647 if (unlikely(all_pinned
)) {
3648 cpu_clear(cpu_of(busiest
), *cpus
);
3649 if (!cpus_empty(*cpus
))
3655 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3656 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3657 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3660 sd
->nr_balance_failed
= 0;
3662 update_shares_locked(this_rq
, sd
);
3666 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3667 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3668 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3670 sd
->nr_balance_failed
= 0;
3676 * idle_balance is called by schedule() if this_cpu is about to become
3677 * idle. Attempts to pull tasks from other CPUs.
3679 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3681 struct sched_domain
*sd
;
3682 int pulled_task
= -1;
3683 unsigned long next_balance
= jiffies
+ HZ
;
3686 for_each_domain(this_cpu
, sd
) {
3687 unsigned long interval
;
3689 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3692 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3693 /* If we've pulled tasks over stop searching: */
3694 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
3697 interval
= msecs_to_jiffies(sd
->balance_interval
);
3698 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3699 next_balance
= sd
->last_balance
+ interval
;
3703 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3705 * We are going idle. next_balance may be set based on
3706 * a busy processor. So reset next_balance.
3708 this_rq
->next_balance
= next_balance
;
3713 * active_load_balance is run by migration threads. It pushes running tasks
3714 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3715 * running on each physical CPU where possible, and avoids physical /
3716 * logical imbalances.
3718 * Called with busiest_rq locked.
3720 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3722 int target_cpu
= busiest_rq
->push_cpu
;
3723 struct sched_domain
*sd
;
3724 struct rq
*target_rq
;
3726 /* Is there any task to move? */
3727 if (busiest_rq
->nr_running
<= 1)
3730 target_rq
= cpu_rq(target_cpu
);
3733 * This condition is "impossible", if it occurs
3734 * we need to fix it. Originally reported by
3735 * Bjorn Helgaas on a 128-cpu setup.
3737 BUG_ON(busiest_rq
== target_rq
);
3739 /* move a task from busiest_rq to target_rq */
3740 double_lock_balance(busiest_rq
, target_rq
);
3741 update_rq_clock(busiest_rq
);
3742 update_rq_clock(target_rq
);
3744 /* Search for an sd spanning us and the target CPU. */
3745 for_each_domain(target_cpu
, sd
) {
3746 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3747 cpu_isset(busiest_cpu
, sd
->span
))
3752 schedstat_inc(sd
, alb_count
);
3754 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3756 schedstat_inc(sd
, alb_pushed
);
3758 schedstat_inc(sd
, alb_failed
);
3760 double_unlock_balance(busiest_rq
, target_rq
);
3765 atomic_t load_balancer
;
3767 } nohz ____cacheline_aligned
= {
3768 .load_balancer
= ATOMIC_INIT(-1),
3769 .cpu_mask
= CPU_MASK_NONE
,
3773 * This routine will try to nominate the ilb (idle load balancing)
3774 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3775 * load balancing on behalf of all those cpus. If all the cpus in the system
3776 * go into this tickless mode, then there will be no ilb owner (as there is
3777 * no need for one) and all the cpus will sleep till the next wakeup event
3780 * For the ilb owner, tick is not stopped. And this tick will be used
3781 * for idle load balancing. ilb owner will still be part of
3784 * While stopping the tick, this cpu will become the ilb owner if there
3785 * is no other owner. And will be the owner till that cpu becomes busy
3786 * or if all cpus in the system stop their ticks at which point
3787 * there is no need for ilb owner.
3789 * When the ilb owner becomes busy, it nominates another owner, during the
3790 * next busy scheduler_tick()
3792 int select_nohz_load_balancer(int stop_tick
)
3794 int cpu
= smp_processor_id();
3797 cpu_set(cpu
, nohz
.cpu_mask
);
3798 cpu_rq(cpu
)->in_nohz_recently
= 1;
3801 * If we are going offline and still the leader, give up!
3803 if (!cpu_active(cpu
) &&
3804 atomic_read(&nohz
.load_balancer
) == cpu
) {
3805 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3810 /* time for ilb owner also to sleep */
3811 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3812 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3813 atomic_set(&nohz
.load_balancer
, -1);
3817 if (atomic_read(&nohz
.load_balancer
) == -1) {
3818 /* make me the ilb owner */
3819 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3821 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3824 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3827 cpu_clear(cpu
, nohz
.cpu_mask
);
3829 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3830 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3837 static DEFINE_SPINLOCK(balancing
);
3840 * It checks each scheduling domain to see if it is due to be balanced,
3841 * and initiates a balancing operation if so.
3843 * Balancing parameters are set up in arch_init_sched_domains.
3845 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3848 struct rq
*rq
= cpu_rq(cpu
);
3849 unsigned long interval
;
3850 struct sched_domain
*sd
;
3851 /* Earliest time when we have to do rebalance again */
3852 unsigned long next_balance
= jiffies
+ 60*HZ
;
3853 int update_next_balance
= 0;
3857 for_each_domain(cpu
, sd
) {
3858 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3861 interval
= sd
->balance_interval
;
3862 if (idle
!= CPU_IDLE
)
3863 interval
*= sd
->busy_factor
;
3865 /* scale ms to jiffies */
3866 interval
= msecs_to_jiffies(interval
);
3867 if (unlikely(!interval
))
3869 if (interval
> HZ
*NR_CPUS
/10)
3870 interval
= HZ
*NR_CPUS
/10;
3872 need_serialize
= sd
->flags
& SD_SERIALIZE
;
3874 if (need_serialize
) {
3875 if (!spin_trylock(&balancing
))
3879 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3880 if (load_balance(cpu
, rq
, sd
, idle
, &balance
, &tmp
)) {
3882 * We've pulled tasks over so either we're no
3883 * longer idle, or one of our SMT siblings is
3886 idle
= CPU_NOT_IDLE
;
3888 sd
->last_balance
= jiffies
;
3891 spin_unlock(&balancing
);
3893 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3894 next_balance
= sd
->last_balance
+ interval
;
3895 update_next_balance
= 1;
3899 * Stop the load balance at this level. There is another
3900 * CPU in our sched group which is doing load balancing more
3908 * next_balance will be updated only when there is a need.
3909 * When the cpu is attached to null domain for ex, it will not be
3912 if (likely(update_next_balance
))
3913 rq
->next_balance
= next_balance
;
3917 * run_rebalance_domains is triggered when needed from the scheduler tick.
3918 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3919 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3921 static void run_rebalance_domains(struct softirq_action
*h
)
3923 int this_cpu
= smp_processor_id();
3924 struct rq
*this_rq
= cpu_rq(this_cpu
);
3925 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3926 CPU_IDLE
: CPU_NOT_IDLE
;
3928 rebalance_domains(this_cpu
, idle
);
3932 * If this cpu is the owner for idle load balancing, then do the
3933 * balancing on behalf of the other idle cpus whose ticks are
3936 if (this_rq
->idle_at_tick
&&
3937 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3938 cpumask_t cpus
= nohz
.cpu_mask
;
3942 cpu_clear(this_cpu
, cpus
);
3943 for_each_cpu_mask_nr(balance_cpu
, cpus
) {
3945 * If this cpu gets work to do, stop the load balancing
3946 * work being done for other cpus. Next load
3947 * balancing owner will pick it up.
3952 rebalance_domains(balance_cpu
, CPU_IDLE
);
3954 rq
= cpu_rq(balance_cpu
);
3955 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3956 this_rq
->next_balance
= rq
->next_balance
;
3963 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3965 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3966 * idle load balancing owner or decide to stop the periodic load balancing,
3967 * if the whole system is idle.
3969 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3973 * If we were in the nohz mode recently and busy at the current
3974 * scheduler tick, then check if we need to nominate new idle
3977 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3978 rq
->in_nohz_recently
= 0;
3980 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3981 cpu_clear(cpu
, nohz
.cpu_mask
);
3982 atomic_set(&nohz
.load_balancer
, -1);
3985 if (atomic_read(&nohz
.load_balancer
) == -1) {
3987 * simple selection for now: Nominate the
3988 * first cpu in the nohz list to be the next
3991 * TBD: Traverse the sched domains and nominate
3992 * the nearest cpu in the nohz.cpu_mask.
3994 int ilb
= first_cpu(nohz
.cpu_mask
);
3996 if (ilb
< nr_cpu_ids
)
4002 * If this cpu is idle and doing idle load balancing for all the
4003 * cpus with ticks stopped, is it time for that to stop?
4005 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4006 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4012 * If this cpu is idle and the idle load balancing is done by
4013 * someone else, then no need raise the SCHED_SOFTIRQ
4015 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4016 cpu_isset(cpu
, nohz
.cpu_mask
))
4019 if (time_after_eq(jiffies
, rq
->next_balance
))
4020 raise_softirq(SCHED_SOFTIRQ
);
4023 #else /* CONFIG_SMP */
4026 * on UP we do not need to balance between CPUs:
4028 static inline void idle_balance(int cpu
, struct rq
*rq
)
4034 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4036 EXPORT_PER_CPU_SYMBOL(kstat
);
4039 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4040 * that have not yet been banked in case the task is currently running.
4042 unsigned long long task_sched_runtime(struct task_struct
*p
)
4044 unsigned long flags
;
4048 rq
= task_rq_lock(p
, &flags
);
4049 ns
= p
->se
.sum_exec_runtime
;
4050 if (task_current(rq
, p
)) {
4051 update_rq_clock(rq
);
4052 delta_exec
= rq
->clock
- p
->se
.exec_start
;
4053 if ((s64
)delta_exec
> 0)
4056 task_rq_unlock(rq
, &flags
);
4062 * Account user cpu time to a process.
4063 * @p: the process that the cpu time gets accounted to
4064 * @cputime: the cpu time spent in user space since the last update
4066 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
4068 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4071 p
->utime
= cputime_add(p
->utime
, cputime
);
4073 /* Add user time to cpustat. */
4074 tmp
= cputime_to_cputime64(cputime
);
4075 if (TASK_NICE(p
) > 0)
4076 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4078 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4079 /* Account for user time used */
4080 acct_update_integrals(p
);
4084 * Account guest cpu time to a process.
4085 * @p: the process that the cpu time gets accounted to
4086 * @cputime: the cpu time spent in virtual machine since the last update
4088 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
4091 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4093 tmp
= cputime_to_cputime64(cputime
);
4095 p
->utime
= cputime_add(p
->utime
, cputime
);
4096 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4098 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4099 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4103 * Account scaled user cpu time to a process.
4104 * @p: the process that the cpu time gets accounted to
4105 * @cputime: the cpu time spent in user space since the last update
4107 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4109 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
4113 * Account system cpu time to a process.
4114 * @p: the process that the cpu time gets accounted to
4115 * @hardirq_offset: the offset to subtract from hardirq_count()
4116 * @cputime: the cpu time spent in kernel space since the last update
4118 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4121 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4122 struct rq
*rq
= this_rq();
4125 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4126 account_guest_time(p
, cputime
);
4130 p
->stime
= cputime_add(p
->stime
, cputime
);
4132 /* Add system time to cpustat. */
4133 tmp
= cputime_to_cputime64(cputime
);
4134 if (hardirq_count() - hardirq_offset
)
4135 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4136 else if (softirq_count())
4137 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4138 else if (p
!= rq
->idle
)
4139 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4140 else if (atomic_read(&rq
->nr_iowait
) > 0)
4141 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4143 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4144 /* Account for system time used */
4145 acct_update_integrals(p
);
4149 * Account scaled system cpu time to a process.
4150 * @p: the process that the cpu time gets accounted to
4151 * @hardirq_offset: the offset to subtract from hardirq_count()
4152 * @cputime: the cpu time spent in kernel space since the last update
4154 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4156 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
4160 * Account for involuntary wait time.
4161 * @p: the process from which the cpu time has been stolen
4162 * @steal: the cpu time spent in involuntary wait
4164 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
4166 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4167 cputime64_t tmp
= cputime_to_cputime64(steal
);
4168 struct rq
*rq
= this_rq();
4170 if (p
== rq
->idle
) {
4171 p
->stime
= cputime_add(p
->stime
, steal
);
4172 if (atomic_read(&rq
->nr_iowait
) > 0)
4173 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4175 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4177 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
4181 * Use precise platform statistics if available:
4183 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4184 cputime_t
task_utime(struct task_struct
*p
)
4189 cputime_t
task_stime(struct task_struct
*p
)
4194 cputime_t
task_utime(struct task_struct
*p
)
4196 clock_t utime
= cputime_to_clock_t(p
->utime
),
4197 total
= utime
+ cputime_to_clock_t(p
->stime
);
4201 * Use CFS's precise accounting:
4203 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
4207 do_div(temp
, total
);
4209 utime
= (clock_t)temp
;
4211 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
4212 return p
->prev_utime
;
4215 cputime_t
task_stime(struct task_struct
*p
)
4220 * Use CFS's precise accounting. (we subtract utime from
4221 * the total, to make sure the total observed by userspace
4222 * grows monotonically - apps rely on that):
4224 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
4225 cputime_to_clock_t(task_utime(p
));
4228 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
4230 return p
->prev_stime
;
4234 inline cputime_t
task_gtime(struct task_struct
*p
)
4240 * This function gets called by the timer code, with HZ frequency.
4241 * We call it with interrupts disabled.
4243 * It also gets called by the fork code, when changing the parent's
4246 void scheduler_tick(void)
4248 int cpu
= smp_processor_id();
4249 struct rq
*rq
= cpu_rq(cpu
);
4250 struct task_struct
*curr
= rq
->curr
;
4254 spin_lock(&rq
->lock
);
4255 update_rq_clock(rq
);
4256 update_cpu_load(rq
);
4257 curr
->sched_class
->task_tick(rq
, curr
, 0);
4258 spin_unlock(&rq
->lock
);
4261 rq
->idle_at_tick
= idle_cpu(cpu
);
4262 trigger_load_balance(rq
, cpu
);
4266 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4267 defined(CONFIG_PREEMPT_TRACER))
4269 static inline unsigned long get_parent_ip(unsigned long addr
)
4271 if (in_lock_functions(addr
)) {
4272 addr
= CALLER_ADDR2
;
4273 if (in_lock_functions(addr
))
4274 addr
= CALLER_ADDR3
;
4279 void __kprobes
add_preempt_count(int val
)
4281 #ifdef CONFIG_DEBUG_PREEMPT
4285 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4288 preempt_count() += val
;
4289 #ifdef CONFIG_DEBUG_PREEMPT
4291 * Spinlock count overflowing soon?
4293 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4296 if (preempt_count() == val
)
4297 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4299 EXPORT_SYMBOL(add_preempt_count
);
4301 void __kprobes
sub_preempt_count(int val
)
4303 #ifdef CONFIG_DEBUG_PREEMPT
4307 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4310 * Is the spinlock portion underflowing?
4312 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4313 !(preempt_count() & PREEMPT_MASK
)))
4317 if (preempt_count() == val
)
4318 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4319 preempt_count() -= val
;
4321 EXPORT_SYMBOL(sub_preempt_count
);
4326 * Print scheduling while atomic bug:
4328 static noinline
void __schedule_bug(struct task_struct
*prev
)
4330 struct pt_regs
*regs
= get_irq_regs();
4332 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4333 prev
->comm
, prev
->pid
, preempt_count());
4335 debug_show_held_locks(prev
);
4337 if (irqs_disabled())
4338 print_irqtrace_events(prev
);
4347 * Various schedule()-time debugging checks and statistics:
4349 static inline void schedule_debug(struct task_struct
*prev
)
4352 * Test if we are atomic. Since do_exit() needs to call into
4353 * schedule() atomically, we ignore that path for now.
4354 * Otherwise, whine if we are scheduling when we should not be.
4356 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4357 __schedule_bug(prev
);
4359 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4361 schedstat_inc(this_rq(), sched_count
);
4362 #ifdef CONFIG_SCHEDSTATS
4363 if (unlikely(prev
->lock_depth
>= 0)) {
4364 schedstat_inc(this_rq(), bkl_count
);
4365 schedstat_inc(prev
, sched_info
.bkl_count
);
4371 * Pick up the highest-prio task:
4373 static inline struct task_struct
*
4374 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
4376 const struct sched_class
*class;
4377 struct task_struct
*p
;
4380 * Optimization: we know that if all tasks are in
4381 * the fair class we can call that function directly:
4383 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4384 p
= fair_sched_class
.pick_next_task(rq
);
4389 class = sched_class_highest
;
4391 p
= class->pick_next_task(rq
);
4395 * Will never be NULL as the idle class always
4396 * returns a non-NULL p:
4398 class = class->next
;
4403 * schedule() is the main scheduler function.
4405 asmlinkage
void __sched
schedule(void)
4407 struct task_struct
*prev
, *next
;
4408 unsigned long *switch_count
;
4414 cpu
= smp_processor_id();
4418 switch_count
= &prev
->nivcsw
;
4420 release_kernel_lock(prev
);
4421 need_resched_nonpreemptible
:
4423 schedule_debug(prev
);
4425 if (sched_feat(HRTICK
))
4429 * Do the rq-clock update outside the rq lock:
4431 local_irq_disable();
4432 update_rq_clock(rq
);
4433 spin_lock(&rq
->lock
);
4434 clear_tsk_need_resched(prev
);
4436 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4437 if (unlikely(signal_pending_state(prev
->state
, prev
)))
4438 prev
->state
= TASK_RUNNING
;
4440 deactivate_task(rq
, prev
, 1);
4441 switch_count
= &prev
->nvcsw
;
4445 if (prev
->sched_class
->pre_schedule
)
4446 prev
->sched_class
->pre_schedule(rq
, prev
);
4449 if (unlikely(!rq
->nr_running
))
4450 idle_balance(cpu
, rq
);
4452 prev
->sched_class
->put_prev_task(rq
, prev
);
4453 next
= pick_next_task(rq
, prev
);
4455 if (likely(prev
!= next
)) {
4456 sched_info_switch(prev
, next
);
4462 context_switch(rq
, prev
, next
); /* unlocks the rq */
4464 * the context switch might have flipped the stack from under
4465 * us, hence refresh the local variables.
4467 cpu
= smp_processor_id();
4470 spin_unlock_irq(&rq
->lock
);
4472 if (unlikely(reacquire_kernel_lock(current
) < 0))
4473 goto need_resched_nonpreemptible
;
4475 preempt_enable_no_resched();
4476 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
4479 EXPORT_SYMBOL(schedule
);
4481 #ifdef CONFIG_PREEMPT
4483 * this is the entry point to schedule() from in-kernel preemption
4484 * off of preempt_enable. Kernel preemptions off return from interrupt
4485 * occur there and call schedule directly.
4487 asmlinkage
void __sched
preempt_schedule(void)
4489 struct thread_info
*ti
= current_thread_info();
4492 * If there is a non-zero preempt_count or interrupts are disabled,
4493 * we do not want to preempt the current task. Just return..
4495 if (likely(ti
->preempt_count
|| irqs_disabled()))
4499 add_preempt_count(PREEMPT_ACTIVE
);
4501 sub_preempt_count(PREEMPT_ACTIVE
);
4504 * Check again in case we missed a preemption opportunity
4505 * between schedule and now.
4508 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4510 EXPORT_SYMBOL(preempt_schedule
);
4513 * this is the entry point to schedule() from kernel preemption
4514 * off of irq context.
4515 * Note, that this is called and return with irqs disabled. This will
4516 * protect us against recursive calling from irq.
4518 asmlinkage
void __sched
preempt_schedule_irq(void)
4520 struct thread_info
*ti
= current_thread_info();
4522 /* Catch callers which need to be fixed */
4523 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4526 add_preempt_count(PREEMPT_ACTIVE
);
4529 local_irq_disable();
4530 sub_preempt_count(PREEMPT_ACTIVE
);
4533 * Check again in case we missed a preemption opportunity
4534 * between schedule and now.
4537 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4540 #endif /* CONFIG_PREEMPT */
4542 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4545 return try_to_wake_up(curr
->private, mode
, sync
);
4547 EXPORT_SYMBOL(default_wake_function
);
4550 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4551 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4552 * number) then we wake all the non-exclusive tasks and one exclusive task.
4554 * There are circumstances in which we can try to wake a task which has already
4555 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4556 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4558 void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4559 int nr_exclusive
, int sync
, void *key
)
4561 wait_queue_t
*curr
, *next
;
4563 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4564 unsigned flags
= curr
->flags
;
4566 if (curr
->func(curr
, mode
, sync
, key
) &&
4567 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4573 * __wake_up - wake up threads blocked on a waitqueue.
4575 * @mode: which threads
4576 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4577 * @key: is directly passed to the wakeup function
4579 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4580 int nr_exclusive
, void *key
)
4582 unsigned long flags
;
4584 spin_lock_irqsave(&q
->lock
, flags
);
4585 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4586 spin_unlock_irqrestore(&q
->lock
, flags
);
4588 EXPORT_SYMBOL(__wake_up
);
4591 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4593 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4595 __wake_up_common(q
, mode
, 1, 0, NULL
);
4599 * __wake_up_sync - wake up threads blocked on a waitqueue.
4601 * @mode: which threads
4602 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4604 * The sync wakeup differs that the waker knows that it will schedule
4605 * away soon, so while the target thread will be woken up, it will not
4606 * be migrated to another CPU - ie. the two threads are 'synchronized'
4607 * with each other. This can prevent needless bouncing between CPUs.
4609 * On UP it can prevent extra preemption.
4612 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4614 unsigned long flags
;
4620 if (unlikely(!nr_exclusive
))
4623 spin_lock_irqsave(&q
->lock
, flags
);
4624 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4625 spin_unlock_irqrestore(&q
->lock
, flags
);
4627 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4629 void complete(struct completion
*x
)
4631 unsigned long flags
;
4633 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4635 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4636 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4638 EXPORT_SYMBOL(complete
);
4640 void complete_all(struct completion
*x
)
4642 unsigned long flags
;
4644 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4645 x
->done
+= UINT_MAX
/2;
4646 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4647 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4649 EXPORT_SYMBOL(complete_all
);
4651 static inline long __sched
4652 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4655 DECLARE_WAITQUEUE(wait
, current
);
4657 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4658 __add_wait_queue_tail(&x
->wait
, &wait
);
4660 if ((state
== TASK_INTERRUPTIBLE
&&
4661 signal_pending(current
)) ||
4662 (state
== TASK_KILLABLE
&&
4663 fatal_signal_pending(current
))) {
4664 timeout
= -ERESTARTSYS
;
4667 __set_current_state(state
);
4668 spin_unlock_irq(&x
->wait
.lock
);
4669 timeout
= schedule_timeout(timeout
);
4670 spin_lock_irq(&x
->wait
.lock
);
4671 } while (!x
->done
&& timeout
);
4672 __remove_wait_queue(&x
->wait
, &wait
);
4677 return timeout
?: 1;
4681 wait_for_common(struct completion
*x
, long timeout
, int state
)
4685 spin_lock_irq(&x
->wait
.lock
);
4686 timeout
= do_wait_for_common(x
, timeout
, state
);
4687 spin_unlock_irq(&x
->wait
.lock
);
4691 void __sched
wait_for_completion(struct completion
*x
)
4693 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4695 EXPORT_SYMBOL(wait_for_completion
);
4697 unsigned long __sched
4698 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4700 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4702 EXPORT_SYMBOL(wait_for_completion_timeout
);
4704 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4706 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4707 if (t
== -ERESTARTSYS
)
4711 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4713 unsigned long __sched
4714 wait_for_completion_interruptible_timeout(struct completion
*x
,
4715 unsigned long timeout
)
4717 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4719 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4721 int __sched
wait_for_completion_killable(struct completion
*x
)
4723 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4724 if (t
== -ERESTARTSYS
)
4728 EXPORT_SYMBOL(wait_for_completion_killable
);
4731 * try_wait_for_completion - try to decrement a completion without blocking
4732 * @x: completion structure
4734 * Returns: 0 if a decrement cannot be done without blocking
4735 * 1 if a decrement succeeded.
4737 * If a completion is being used as a counting completion,
4738 * attempt to decrement the counter without blocking. This
4739 * enables us to avoid waiting if the resource the completion
4740 * is protecting is not available.
4742 bool try_wait_for_completion(struct completion
*x
)
4746 spin_lock_irq(&x
->wait
.lock
);
4751 spin_unlock_irq(&x
->wait
.lock
);
4754 EXPORT_SYMBOL(try_wait_for_completion
);
4757 * completion_done - Test to see if a completion has any waiters
4758 * @x: completion structure
4760 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4761 * 1 if there are no waiters.
4764 bool completion_done(struct completion
*x
)
4768 spin_lock_irq(&x
->wait
.lock
);
4771 spin_unlock_irq(&x
->wait
.lock
);
4774 EXPORT_SYMBOL(completion_done
);
4777 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4779 unsigned long flags
;
4782 init_waitqueue_entry(&wait
, current
);
4784 __set_current_state(state
);
4786 spin_lock_irqsave(&q
->lock
, flags
);
4787 __add_wait_queue(q
, &wait
);
4788 spin_unlock(&q
->lock
);
4789 timeout
= schedule_timeout(timeout
);
4790 spin_lock_irq(&q
->lock
);
4791 __remove_wait_queue(q
, &wait
);
4792 spin_unlock_irqrestore(&q
->lock
, flags
);
4797 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4799 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4801 EXPORT_SYMBOL(interruptible_sleep_on
);
4804 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4806 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4808 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4810 void __sched
sleep_on(wait_queue_head_t
*q
)
4812 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4814 EXPORT_SYMBOL(sleep_on
);
4816 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4818 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4820 EXPORT_SYMBOL(sleep_on_timeout
);
4822 #ifdef CONFIG_RT_MUTEXES
4825 * rt_mutex_setprio - set the current priority of a task
4827 * @prio: prio value (kernel-internal form)
4829 * This function changes the 'effective' priority of a task. It does
4830 * not touch ->normal_prio like __setscheduler().
4832 * Used by the rt_mutex code to implement priority inheritance logic.
4834 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4836 unsigned long flags
;
4837 int oldprio
, on_rq
, running
;
4839 const struct sched_class
*prev_class
= p
->sched_class
;
4841 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4843 rq
= task_rq_lock(p
, &flags
);
4844 update_rq_clock(rq
);
4847 on_rq
= p
->se
.on_rq
;
4848 running
= task_current(rq
, p
);
4850 dequeue_task(rq
, p
, 0);
4852 p
->sched_class
->put_prev_task(rq
, p
);
4855 p
->sched_class
= &rt_sched_class
;
4857 p
->sched_class
= &fair_sched_class
;
4862 p
->sched_class
->set_curr_task(rq
);
4864 enqueue_task(rq
, p
, 0);
4866 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4868 task_rq_unlock(rq
, &flags
);
4873 void set_user_nice(struct task_struct
*p
, long nice
)
4875 int old_prio
, delta
, on_rq
;
4876 unsigned long flags
;
4879 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4882 * We have to be careful, if called from sys_setpriority(),
4883 * the task might be in the middle of scheduling on another CPU.
4885 rq
= task_rq_lock(p
, &flags
);
4886 update_rq_clock(rq
);
4888 * The RT priorities are set via sched_setscheduler(), but we still
4889 * allow the 'normal' nice value to be set - but as expected
4890 * it wont have any effect on scheduling until the task is
4891 * SCHED_FIFO/SCHED_RR:
4893 if (task_has_rt_policy(p
)) {
4894 p
->static_prio
= NICE_TO_PRIO(nice
);
4897 on_rq
= p
->se
.on_rq
;
4899 dequeue_task(rq
, p
, 0);
4901 p
->static_prio
= NICE_TO_PRIO(nice
);
4904 p
->prio
= effective_prio(p
);
4905 delta
= p
->prio
- old_prio
;
4908 enqueue_task(rq
, p
, 0);
4910 * If the task increased its priority or is running and
4911 * lowered its priority, then reschedule its CPU:
4913 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4914 resched_task(rq
->curr
);
4917 task_rq_unlock(rq
, &flags
);
4919 EXPORT_SYMBOL(set_user_nice
);
4922 * can_nice - check if a task can reduce its nice value
4926 int can_nice(const struct task_struct
*p
, const int nice
)
4928 /* convert nice value [19,-20] to rlimit style value [1,40] */
4929 int nice_rlim
= 20 - nice
;
4931 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4932 capable(CAP_SYS_NICE
));
4935 #ifdef __ARCH_WANT_SYS_NICE
4938 * sys_nice - change the priority of the current process.
4939 * @increment: priority increment
4941 * sys_setpriority is a more generic, but much slower function that
4942 * does similar things.
4944 SYSCALL_DEFINE1(nice
, int, increment
)
4949 * Setpriority might change our priority at the same moment.
4950 * We don't have to worry. Conceptually one call occurs first
4951 * and we have a single winner.
4953 if (increment
< -40)
4958 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4964 if (increment
< 0 && !can_nice(current
, nice
))
4967 retval
= security_task_setnice(current
, nice
);
4971 set_user_nice(current
, nice
);
4978 * task_prio - return the priority value of a given task.
4979 * @p: the task in question.
4981 * This is the priority value as seen by users in /proc.
4982 * RT tasks are offset by -200. Normal tasks are centered
4983 * around 0, value goes from -16 to +15.
4985 int task_prio(const struct task_struct
*p
)
4987 return p
->prio
- MAX_RT_PRIO
;
4991 * task_nice - return the nice value of a given task.
4992 * @p: the task in question.
4994 int task_nice(const struct task_struct
*p
)
4996 return TASK_NICE(p
);
4998 EXPORT_SYMBOL(task_nice
);
5001 * idle_cpu - is a given cpu idle currently?
5002 * @cpu: the processor in question.
5004 int idle_cpu(int cpu
)
5006 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
5010 * idle_task - return the idle task for a given cpu.
5011 * @cpu: the processor in question.
5013 struct task_struct
*idle_task(int cpu
)
5015 return cpu_rq(cpu
)->idle
;
5019 * find_process_by_pid - find a process with a matching PID value.
5020 * @pid: the pid in question.
5022 static struct task_struct
*find_process_by_pid(pid_t pid
)
5024 return pid
? find_task_by_vpid(pid
) : current
;
5027 /* Actually do priority change: must hold rq lock. */
5029 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
5031 BUG_ON(p
->se
.on_rq
);
5034 switch (p
->policy
) {
5038 p
->sched_class
= &fair_sched_class
;
5042 p
->sched_class
= &rt_sched_class
;
5046 p
->rt_priority
= prio
;
5047 p
->normal_prio
= normal_prio(p
);
5048 /* we are holding p->pi_lock already */
5049 p
->prio
= rt_mutex_getprio(p
);
5053 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
5054 struct sched_param
*param
, bool user
)
5056 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
5057 unsigned long flags
;
5058 const struct sched_class
*prev_class
= p
->sched_class
;
5061 /* may grab non-irq protected spin_locks */
5062 BUG_ON(in_interrupt());
5064 /* double check policy once rq lock held */
5066 policy
= oldpolicy
= p
->policy
;
5067 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
5068 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
5069 policy
!= SCHED_IDLE
)
5072 * Valid priorities for SCHED_FIFO and SCHED_RR are
5073 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5074 * SCHED_BATCH and SCHED_IDLE is 0.
5076 if (param
->sched_priority
< 0 ||
5077 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5078 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
5080 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
5084 * Allow unprivileged RT tasks to decrease priority:
5086 if (user
&& !capable(CAP_SYS_NICE
)) {
5087 if (rt_policy(policy
)) {
5088 unsigned long rlim_rtprio
;
5090 if (!lock_task_sighand(p
, &flags
))
5092 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
5093 unlock_task_sighand(p
, &flags
);
5095 /* can't set/change the rt policy */
5096 if (policy
!= p
->policy
&& !rlim_rtprio
)
5099 /* can't increase priority */
5100 if (param
->sched_priority
> p
->rt_priority
&&
5101 param
->sched_priority
> rlim_rtprio
)
5105 * Like positive nice levels, dont allow tasks to
5106 * move out of SCHED_IDLE either:
5108 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
5111 /* can't change other user's priorities */
5112 if ((current
->euid
!= p
->euid
) &&
5113 (current
->euid
!= p
->uid
))
5118 #ifdef CONFIG_RT_GROUP_SCHED
5120 * Do not allow realtime tasks into groups that have no runtime
5123 if (rt_policy(policy
) && task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
5127 retval
= security_task_setscheduler(p
, policy
, param
);
5133 * make sure no PI-waiters arrive (or leave) while we are
5134 * changing the priority of the task:
5136 spin_lock_irqsave(&p
->pi_lock
, flags
);
5138 * To be able to change p->policy safely, the apropriate
5139 * runqueue lock must be held.
5141 rq
= __task_rq_lock(p
);
5142 /* recheck policy now with rq lock held */
5143 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5144 policy
= oldpolicy
= -1;
5145 __task_rq_unlock(rq
);
5146 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5149 update_rq_clock(rq
);
5150 on_rq
= p
->se
.on_rq
;
5151 running
= task_current(rq
, p
);
5153 deactivate_task(rq
, p
, 0);
5155 p
->sched_class
->put_prev_task(rq
, p
);
5158 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5161 p
->sched_class
->set_curr_task(rq
);
5163 activate_task(rq
, p
, 0);
5165 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5167 __task_rq_unlock(rq
);
5168 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5170 rt_mutex_adjust_pi(p
);
5176 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5177 * @p: the task in question.
5178 * @policy: new policy.
5179 * @param: structure containing the new RT priority.
5181 * NOTE that the task may be already dead.
5183 int sched_setscheduler(struct task_struct
*p
, int policy
,
5184 struct sched_param
*param
)
5186 return __sched_setscheduler(p
, policy
, param
, true);
5188 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5191 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5192 * @p: the task in question.
5193 * @policy: new policy.
5194 * @param: structure containing the new RT priority.
5196 * Just like sched_setscheduler, only don't bother checking if the
5197 * current context has permission. For example, this is needed in
5198 * stop_machine(): we create temporary high priority worker threads,
5199 * but our caller might not have that capability.
5201 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5202 struct sched_param
*param
)
5204 return __sched_setscheduler(p
, policy
, param
, false);
5208 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5210 struct sched_param lparam
;
5211 struct task_struct
*p
;
5214 if (!param
|| pid
< 0)
5216 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5221 p
= find_process_by_pid(pid
);
5223 retval
= sched_setscheduler(p
, policy
, &lparam
);
5230 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5231 * @pid: the pid in question.
5232 * @policy: new policy.
5233 * @param: structure containing the new RT priority.
5235 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
5236 struct sched_param __user
*, param
)
5238 /* negative values for policy are not valid */
5242 return do_sched_setscheduler(pid
, policy
, param
);
5246 * sys_sched_setparam - set/change the RT priority of a thread
5247 * @pid: the pid in question.
5248 * @param: structure containing the new RT priority.
5250 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5252 return do_sched_setscheduler(pid
, -1, param
);
5256 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5257 * @pid: the pid in question.
5259 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
5261 struct task_struct
*p
;
5268 read_lock(&tasklist_lock
);
5269 p
= find_process_by_pid(pid
);
5271 retval
= security_task_getscheduler(p
);
5275 read_unlock(&tasklist_lock
);
5280 * sys_sched_getscheduler - get the RT priority of a thread
5281 * @pid: the pid in question.
5282 * @param: structure containing the RT priority.
5284 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5286 struct sched_param lp
;
5287 struct task_struct
*p
;
5290 if (!param
|| pid
< 0)
5293 read_lock(&tasklist_lock
);
5294 p
= find_process_by_pid(pid
);
5299 retval
= security_task_getscheduler(p
);
5303 lp
.sched_priority
= p
->rt_priority
;
5304 read_unlock(&tasklist_lock
);
5307 * This one might sleep, we cannot do it with a spinlock held ...
5309 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5314 read_unlock(&tasklist_lock
);
5318 long sched_setaffinity(pid_t pid
, const cpumask_t
*in_mask
)
5320 cpumask_t cpus_allowed
;
5321 cpumask_t new_mask
= *in_mask
;
5322 struct task_struct
*p
;
5326 read_lock(&tasklist_lock
);
5328 p
= find_process_by_pid(pid
);
5330 read_unlock(&tasklist_lock
);
5336 * It is not safe to call set_cpus_allowed with the
5337 * tasklist_lock held. We will bump the task_struct's
5338 * usage count and then drop tasklist_lock.
5341 read_unlock(&tasklist_lock
);
5344 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
5345 !capable(CAP_SYS_NICE
))
5348 retval
= security_task_setscheduler(p
, 0, NULL
);
5352 cpuset_cpus_allowed(p
, &cpus_allowed
);
5353 cpus_and(new_mask
, new_mask
, cpus_allowed
);
5355 retval
= set_cpus_allowed_ptr(p
, &new_mask
);
5358 cpuset_cpus_allowed(p
, &cpus_allowed
);
5359 if (!cpus_subset(new_mask
, cpus_allowed
)) {
5361 * We must have raced with a concurrent cpuset
5362 * update. Just reset the cpus_allowed to the
5363 * cpuset's cpus_allowed
5365 new_mask
= cpus_allowed
;
5375 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5376 cpumask_t
*new_mask
)
5378 if (len
< sizeof(cpumask_t
)) {
5379 memset(new_mask
, 0, sizeof(cpumask_t
));
5380 } else if (len
> sizeof(cpumask_t
)) {
5381 len
= sizeof(cpumask_t
);
5383 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5387 * sys_sched_setaffinity - set the cpu affinity of a process
5388 * @pid: pid of the process
5389 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5390 * @user_mask_ptr: user-space pointer to the new cpu mask
5392 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
5393 unsigned long __user
*, user_mask_ptr
)
5398 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
5402 return sched_setaffinity(pid
, &new_mask
);
5405 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
5407 struct task_struct
*p
;
5411 read_lock(&tasklist_lock
);
5414 p
= find_process_by_pid(pid
);
5418 retval
= security_task_getscheduler(p
);
5422 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
5425 read_unlock(&tasklist_lock
);
5432 * sys_sched_getaffinity - get the cpu affinity of a process
5433 * @pid: pid of the process
5434 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5435 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5437 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
5438 unsigned long __user
*, user_mask_ptr
)
5443 if (len
< sizeof(cpumask_t
))
5446 ret
= sched_getaffinity(pid
, &mask
);
5450 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
5453 return sizeof(cpumask_t
);
5457 * sys_sched_yield - yield the current processor to other threads.
5459 * This function yields the current CPU to other tasks. If there are no
5460 * other threads running on this CPU then this function will return.
5462 SYSCALL_DEFINE0(sched_yield
)
5464 struct rq
*rq
= this_rq_lock();
5466 schedstat_inc(rq
, yld_count
);
5467 current
->sched_class
->yield_task(rq
);
5470 * Since we are going to call schedule() anyway, there's
5471 * no need to preempt or enable interrupts:
5473 __release(rq
->lock
);
5474 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5475 _raw_spin_unlock(&rq
->lock
);
5476 preempt_enable_no_resched();
5483 static void __cond_resched(void)
5485 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5486 __might_sleep(__FILE__
, __LINE__
);
5489 * The BKS might be reacquired before we have dropped
5490 * PREEMPT_ACTIVE, which could trigger a second
5491 * cond_resched() call.
5494 add_preempt_count(PREEMPT_ACTIVE
);
5496 sub_preempt_count(PREEMPT_ACTIVE
);
5497 } while (need_resched());
5500 int __sched
_cond_resched(void)
5502 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
5503 system_state
== SYSTEM_RUNNING
) {
5509 EXPORT_SYMBOL(_cond_resched
);
5512 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5513 * call schedule, and on return reacquire the lock.
5515 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5516 * operations here to prevent schedule() from being called twice (once via
5517 * spin_unlock(), once by hand).
5519 int cond_resched_lock(spinlock_t
*lock
)
5521 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
5524 if (spin_needbreak(lock
) || resched
) {
5526 if (resched
&& need_resched())
5535 EXPORT_SYMBOL(cond_resched_lock
);
5537 int __sched
cond_resched_softirq(void)
5539 BUG_ON(!in_softirq());
5541 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
5549 EXPORT_SYMBOL(cond_resched_softirq
);
5552 * yield - yield the current processor to other threads.
5554 * This is a shortcut for kernel-space yielding - it marks the
5555 * thread runnable and calls sys_sched_yield().
5557 void __sched
yield(void)
5559 set_current_state(TASK_RUNNING
);
5562 EXPORT_SYMBOL(yield
);
5565 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5566 * that process accounting knows that this is a task in IO wait state.
5568 * But don't do that if it is a deliberate, throttling IO wait (this task
5569 * has set its backing_dev_info: the queue against which it should throttle)
5571 void __sched
io_schedule(void)
5573 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5575 delayacct_blkio_start();
5576 atomic_inc(&rq
->nr_iowait
);
5578 atomic_dec(&rq
->nr_iowait
);
5579 delayacct_blkio_end();
5581 EXPORT_SYMBOL(io_schedule
);
5583 long __sched
io_schedule_timeout(long timeout
)
5585 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5588 delayacct_blkio_start();
5589 atomic_inc(&rq
->nr_iowait
);
5590 ret
= schedule_timeout(timeout
);
5591 atomic_dec(&rq
->nr_iowait
);
5592 delayacct_blkio_end();
5597 * sys_sched_get_priority_max - return maximum RT priority.
5598 * @policy: scheduling class.
5600 * this syscall returns the maximum rt_priority that can be used
5601 * by a given scheduling class.
5603 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5610 ret
= MAX_USER_RT_PRIO
-1;
5622 * sys_sched_get_priority_min - return minimum RT priority.
5623 * @policy: scheduling class.
5625 * this syscall returns the minimum rt_priority that can be used
5626 * by a given scheduling class.
5628 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5646 * sys_sched_rr_get_interval - return the default timeslice of a process.
5647 * @pid: pid of the process.
5648 * @interval: userspace pointer to the timeslice value.
5650 * this syscall writes the default timeslice value of a given process
5651 * into the user-space timespec buffer. A value of '0' means infinity.
5653 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5654 struct timespec __user
*, interval
)
5656 struct task_struct
*p
;
5657 unsigned int time_slice
;
5665 read_lock(&tasklist_lock
);
5666 p
= find_process_by_pid(pid
);
5670 retval
= security_task_getscheduler(p
);
5675 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5676 * tasks that are on an otherwise idle runqueue:
5679 if (p
->policy
== SCHED_RR
) {
5680 time_slice
= DEF_TIMESLICE
;
5681 } else if (p
->policy
!= SCHED_FIFO
) {
5682 struct sched_entity
*se
= &p
->se
;
5683 unsigned long flags
;
5686 rq
= task_rq_lock(p
, &flags
);
5687 if (rq
->cfs
.load
.weight
)
5688 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5689 task_rq_unlock(rq
, &flags
);
5691 read_unlock(&tasklist_lock
);
5692 jiffies_to_timespec(time_slice
, &t
);
5693 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5697 read_unlock(&tasklist_lock
);
5701 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5703 void sched_show_task(struct task_struct
*p
)
5705 unsigned long free
= 0;
5708 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5709 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5710 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5711 #if BITS_PER_LONG == 32
5712 if (state
== TASK_RUNNING
)
5713 printk(KERN_CONT
" running ");
5715 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5717 if (state
== TASK_RUNNING
)
5718 printk(KERN_CONT
" running task ");
5720 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5722 #ifdef CONFIG_DEBUG_STACK_USAGE
5724 unsigned long *n
= end_of_stack(p
);
5727 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5730 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5731 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5733 show_stack(p
, NULL
);
5736 void show_state_filter(unsigned long state_filter
)
5738 struct task_struct
*g
, *p
;
5740 #if BITS_PER_LONG == 32
5742 " task PC stack pid father\n");
5745 " task PC stack pid father\n");
5747 read_lock(&tasklist_lock
);
5748 do_each_thread(g
, p
) {
5750 * reset the NMI-timeout, listing all files on a slow
5751 * console might take alot of time:
5753 touch_nmi_watchdog();
5754 if (!state_filter
|| (p
->state
& state_filter
))
5756 } while_each_thread(g
, p
);
5758 touch_all_softlockup_watchdogs();
5760 #ifdef CONFIG_SCHED_DEBUG
5761 sysrq_sched_debug_show();
5763 read_unlock(&tasklist_lock
);
5765 * Only show locks if all tasks are dumped:
5767 if (state_filter
== -1)
5768 debug_show_all_locks();
5771 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5773 idle
->sched_class
= &idle_sched_class
;
5777 * init_idle - set up an idle thread for a given CPU
5778 * @idle: task in question
5779 * @cpu: cpu the idle task belongs to
5781 * NOTE: this function does not set the idle thread's NEED_RESCHED
5782 * flag, to make booting more robust.
5784 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5786 struct rq
*rq
= cpu_rq(cpu
);
5787 unsigned long flags
;
5790 idle
->se
.exec_start
= sched_clock();
5792 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5793 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5794 __set_task_cpu(idle
, cpu
);
5796 spin_lock_irqsave(&rq
->lock
, flags
);
5797 rq
->curr
= rq
->idle
= idle
;
5798 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5801 spin_unlock_irqrestore(&rq
->lock
, flags
);
5803 /* Set the preempt count _outside_ the spinlocks! */
5804 #if defined(CONFIG_PREEMPT)
5805 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5807 task_thread_info(idle
)->preempt_count
= 0;
5810 * The idle tasks have their own, simple scheduling class:
5812 idle
->sched_class
= &idle_sched_class
;
5816 * In a system that switches off the HZ timer nohz_cpu_mask
5817 * indicates which cpus entered this state. This is used
5818 * in the rcu update to wait only for active cpus. For system
5819 * which do not switch off the HZ timer nohz_cpu_mask should
5820 * always be CPU_MASK_NONE.
5822 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5825 * Increase the granularity value when there are more CPUs,
5826 * because with more CPUs the 'effective latency' as visible
5827 * to users decreases. But the relationship is not linear,
5828 * so pick a second-best guess by going with the log2 of the
5831 * This idea comes from the SD scheduler of Con Kolivas:
5833 static inline void sched_init_granularity(void)
5835 unsigned int factor
= 1 + ilog2(num_online_cpus());
5836 const unsigned long limit
= 200000000;
5838 sysctl_sched_min_granularity
*= factor
;
5839 if (sysctl_sched_min_granularity
> limit
)
5840 sysctl_sched_min_granularity
= limit
;
5842 sysctl_sched_latency
*= factor
;
5843 if (sysctl_sched_latency
> limit
)
5844 sysctl_sched_latency
= limit
;
5846 sysctl_sched_wakeup_granularity
*= factor
;
5848 sysctl_sched_shares_ratelimit
*= factor
;
5853 * This is how migration works:
5855 * 1) we queue a struct migration_req structure in the source CPU's
5856 * runqueue and wake up that CPU's migration thread.
5857 * 2) we down() the locked semaphore => thread blocks.
5858 * 3) migration thread wakes up (implicitly it forces the migrated
5859 * thread off the CPU)
5860 * 4) it gets the migration request and checks whether the migrated
5861 * task is still in the wrong runqueue.
5862 * 5) if it's in the wrong runqueue then the migration thread removes
5863 * it and puts it into the right queue.
5864 * 6) migration thread up()s the semaphore.
5865 * 7) we wake up and the migration is done.
5869 * Change a given task's CPU affinity. Migrate the thread to a
5870 * proper CPU and schedule it away if the CPU it's executing on
5871 * is removed from the allowed bitmask.
5873 * NOTE: the caller must have a valid reference to the task, the
5874 * task must not exit() & deallocate itself prematurely. The
5875 * call is not atomic; no spinlocks may be held.
5877 int set_cpus_allowed_ptr(struct task_struct
*p
, const cpumask_t
*new_mask
)
5879 struct migration_req req
;
5880 unsigned long flags
;
5884 rq
= task_rq_lock(p
, &flags
);
5885 if (!cpus_intersects(*new_mask
, cpu_online_map
)) {
5890 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5891 !cpus_equal(p
->cpus_allowed
, *new_mask
))) {
5896 if (p
->sched_class
->set_cpus_allowed
)
5897 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5899 p
->cpus_allowed
= *new_mask
;
5900 p
->rt
.nr_cpus_allowed
= cpus_weight(*new_mask
);
5903 /* Can the task run on the task's current CPU? If so, we're done */
5904 if (cpu_isset(task_cpu(p
), *new_mask
))
5907 if (migrate_task(p
, any_online_cpu(*new_mask
), &req
)) {
5908 /* Need help from migration thread: drop lock and wait. */
5909 task_rq_unlock(rq
, &flags
);
5910 wake_up_process(rq
->migration_thread
);
5911 wait_for_completion(&req
.done
);
5912 tlb_migrate_finish(p
->mm
);
5916 task_rq_unlock(rq
, &flags
);
5920 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5923 * Move (not current) task off this cpu, onto dest cpu. We're doing
5924 * this because either it can't run here any more (set_cpus_allowed()
5925 * away from this CPU, or CPU going down), or because we're
5926 * attempting to rebalance this task on exec (sched_exec).
5928 * So we race with normal scheduler movements, but that's OK, as long
5929 * as the task is no longer on this CPU.
5931 * Returns non-zero if task was successfully migrated.
5933 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5935 struct rq
*rq_dest
, *rq_src
;
5938 if (unlikely(!cpu_active(dest_cpu
)))
5941 rq_src
= cpu_rq(src_cpu
);
5942 rq_dest
= cpu_rq(dest_cpu
);
5944 double_rq_lock(rq_src
, rq_dest
);
5945 /* Already moved. */
5946 if (task_cpu(p
) != src_cpu
)
5948 /* Affinity changed (again). */
5949 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5952 on_rq
= p
->se
.on_rq
;
5954 deactivate_task(rq_src
, p
, 0);
5956 set_task_cpu(p
, dest_cpu
);
5958 activate_task(rq_dest
, p
, 0);
5959 check_preempt_curr(rq_dest
, p
, 0);
5964 double_rq_unlock(rq_src
, rq_dest
);
5969 * migration_thread - this is a highprio system thread that performs
5970 * thread migration by bumping thread off CPU then 'pushing' onto
5973 static int migration_thread(void *data
)
5975 int cpu
= (long)data
;
5979 BUG_ON(rq
->migration_thread
!= current
);
5981 set_current_state(TASK_INTERRUPTIBLE
);
5982 while (!kthread_should_stop()) {
5983 struct migration_req
*req
;
5984 struct list_head
*head
;
5986 spin_lock_irq(&rq
->lock
);
5988 if (cpu_is_offline(cpu
)) {
5989 spin_unlock_irq(&rq
->lock
);
5993 if (rq
->active_balance
) {
5994 active_load_balance(rq
, cpu
);
5995 rq
->active_balance
= 0;
5998 head
= &rq
->migration_queue
;
6000 if (list_empty(head
)) {
6001 spin_unlock_irq(&rq
->lock
);
6003 set_current_state(TASK_INTERRUPTIBLE
);
6006 req
= list_entry(head
->next
, struct migration_req
, list
);
6007 list_del_init(head
->next
);
6009 spin_unlock(&rq
->lock
);
6010 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
6013 complete(&req
->done
);
6015 __set_current_state(TASK_RUNNING
);
6019 /* Wait for kthread_stop */
6020 set_current_state(TASK_INTERRUPTIBLE
);
6021 while (!kthread_should_stop()) {
6023 set_current_state(TASK_INTERRUPTIBLE
);
6025 __set_current_state(TASK_RUNNING
);
6029 #ifdef CONFIG_HOTPLUG_CPU
6031 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6035 local_irq_disable();
6036 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
6042 * Figure out where task on dead CPU should go, use force if necessary.
6043 * NOTE: interrupts should be disabled by the caller
6045 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
6047 unsigned long flags
;
6054 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
6055 cpus_and(mask
, mask
, p
->cpus_allowed
);
6056 dest_cpu
= any_online_cpu(mask
);
6058 /* On any allowed CPU? */
6059 if (dest_cpu
>= nr_cpu_ids
)
6060 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
6062 /* No more Mr. Nice Guy. */
6063 if (dest_cpu
>= nr_cpu_ids
) {
6064 cpumask_t cpus_allowed
;
6066 cpuset_cpus_allowed_locked(p
, &cpus_allowed
);
6068 * Try to stay on the same cpuset, where the
6069 * current cpuset may be a subset of all cpus.
6070 * The cpuset_cpus_allowed_locked() variant of
6071 * cpuset_cpus_allowed() will not block. It must be
6072 * called within calls to cpuset_lock/cpuset_unlock.
6074 rq
= task_rq_lock(p
, &flags
);
6075 p
->cpus_allowed
= cpus_allowed
;
6076 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
6077 task_rq_unlock(rq
, &flags
);
6080 * Don't tell them about moving exiting tasks or
6081 * kernel threads (both mm NULL), since they never
6084 if (p
->mm
&& printk_ratelimit()) {
6085 printk(KERN_INFO
"process %d (%s) no "
6086 "longer affine to cpu%d\n",
6087 task_pid_nr(p
), p
->comm
, dead_cpu
);
6090 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
6094 * While a dead CPU has no uninterruptible tasks queued at this point,
6095 * it might still have a nonzero ->nr_uninterruptible counter, because
6096 * for performance reasons the counter is not stricly tracking tasks to
6097 * their home CPUs. So we just add the counter to another CPU's counter,
6098 * to keep the global sum constant after CPU-down:
6100 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
6102 struct rq
*rq_dest
= cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR
));
6103 unsigned long flags
;
6105 local_irq_save(flags
);
6106 double_rq_lock(rq_src
, rq_dest
);
6107 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
6108 rq_src
->nr_uninterruptible
= 0;
6109 double_rq_unlock(rq_src
, rq_dest
);
6110 local_irq_restore(flags
);
6113 /* Run through task list and migrate tasks from the dead cpu. */
6114 static void migrate_live_tasks(int src_cpu
)
6116 struct task_struct
*p
, *t
;
6118 read_lock(&tasklist_lock
);
6120 do_each_thread(t
, p
) {
6124 if (task_cpu(p
) == src_cpu
)
6125 move_task_off_dead_cpu(src_cpu
, p
);
6126 } while_each_thread(t
, p
);
6128 read_unlock(&tasklist_lock
);
6132 * Schedules idle task to be the next runnable task on current CPU.
6133 * It does so by boosting its priority to highest possible.
6134 * Used by CPU offline code.
6136 void sched_idle_next(void)
6138 int this_cpu
= smp_processor_id();
6139 struct rq
*rq
= cpu_rq(this_cpu
);
6140 struct task_struct
*p
= rq
->idle
;
6141 unsigned long flags
;
6143 /* cpu has to be offline */
6144 BUG_ON(cpu_online(this_cpu
));
6147 * Strictly not necessary since rest of the CPUs are stopped by now
6148 * and interrupts disabled on the current cpu.
6150 spin_lock_irqsave(&rq
->lock
, flags
);
6152 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6154 update_rq_clock(rq
);
6155 activate_task(rq
, p
, 0);
6157 spin_unlock_irqrestore(&rq
->lock
, flags
);
6161 * Ensures that the idle task is using init_mm right before its cpu goes
6164 void idle_task_exit(void)
6166 struct mm_struct
*mm
= current
->active_mm
;
6168 BUG_ON(cpu_online(smp_processor_id()));
6171 switch_mm(mm
, &init_mm
, current
);
6175 /* called under rq->lock with disabled interrupts */
6176 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
6178 struct rq
*rq
= cpu_rq(dead_cpu
);
6180 /* Must be exiting, otherwise would be on tasklist. */
6181 BUG_ON(!p
->exit_state
);
6183 /* Cannot have done final schedule yet: would have vanished. */
6184 BUG_ON(p
->state
== TASK_DEAD
);
6189 * Drop lock around migration; if someone else moves it,
6190 * that's OK. No task can be added to this CPU, so iteration is
6193 spin_unlock_irq(&rq
->lock
);
6194 move_task_off_dead_cpu(dead_cpu
, p
);
6195 spin_lock_irq(&rq
->lock
);
6200 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6201 static void migrate_dead_tasks(unsigned int dead_cpu
)
6203 struct rq
*rq
= cpu_rq(dead_cpu
);
6204 struct task_struct
*next
;
6207 if (!rq
->nr_running
)
6209 update_rq_clock(rq
);
6210 next
= pick_next_task(rq
, rq
->curr
);
6213 next
->sched_class
->put_prev_task(rq
, next
);
6214 migrate_dead(dead_cpu
, next
);
6218 #endif /* CONFIG_HOTPLUG_CPU */
6220 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6222 static struct ctl_table sd_ctl_dir
[] = {
6224 .procname
= "sched_domain",
6230 static struct ctl_table sd_ctl_root
[] = {
6232 .ctl_name
= CTL_KERN
,
6233 .procname
= "kernel",
6235 .child
= sd_ctl_dir
,
6240 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6242 struct ctl_table
*entry
=
6243 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6248 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6250 struct ctl_table
*entry
;
6253 * In the intermediate directories, both the child directory and
6254 * procname are dynamically allocated and could fail but the mode
6255 * will always be set. In the lowest directory the names are
6256 * static strings and all have proc handlers.
6258 for (entry
= *tablep
; entry
->mode
; entry
++) {
6260 sd_free_ctl_entry(&entry
->child
);
6261 if (entry
->proc_handler
== NULL
)
6262 kfree(entry
->procname
);
6270 set_table_entry(struct ctl_table
*entry
,
6271 const char *procname
, void *data
, int maxlen
,
6272 mode_t mode
, proc_handler
*proc_handler
)
6274 entry
->procname
= procname
;
6276 entry
->maxlen
= maxlen
;
6278 entry
->proc_handler
= proc_handler
;
6281 static struct ctl_table
*
6282 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6284 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
6289 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6290 sizeof(long), 0644, proc_doulongvec_minmax
);
6291 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6292 sizeof(long), 0644, proc_doulongvec_minmax
);
6293 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6294 sizeof(int), 0644, proc_dointvec_minmax
);
6295 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6296 sizeof(int), 0644, proc_dointvec_minmax
);
6297 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6298 sizeof(int), 0644, proc_dointvec_minmax
);
6299 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6300 sizeof(int), 0644, proc_dointvec_minmax
);
6301 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6302 sizeof(int), 0644, proc_dointvec_minmax
);
6303 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6304 sizeof(int), 0644, proc_dointvec_minmax
);
6305 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6306 sizeof(int), 0644, proc_dointvec_minmax
);
6307 set_table_entry(&table
[9], "cache_nice_tries",
6308 &sd
->cache_nice_tries
,
6309 sizeof(int), 0644, proc_dointvec_minmax
);
6310 set_table_entry(&table
[10], "flags", &sd
->flags
,
6311 sizeof(int), 0644, proc_dointvec_minmax
);
6312 /* &table[11] is terminator */
6317 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6319 struct ctl_table
*entry
, *table
;
6320 struct sched_domain
*sd
;
6321 int domain_num
= 0, i
;
6324 for_each_domain(cpu
, sd
)
6326 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6331 for_each_domain(cpu
, sd
) {
6332 snprintf(buf
, 32, "domain%d", i
);
6333 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6335 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6342 static struct ctl_table_header
*sd_sysctl_header
;
6343 static void register_sched_domain_sysctl(void)
6345 int i
, cpu_num
= num_online_cpus();
6346 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6349 WARN_ON(sd_ctl_dir
[0].child
);
6350 sd_ctl_dir
[0].child
= entry
;
6355 for_each_online_cpu(i
) {
6356 snprintf(buf
, 32, "cpu%d", i
);
6357 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6359 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6363 WARN_ON(sd_sysctl_header
);
6364 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6367 /* may be called multiple times per register */
6368 static void unregister_sched_domain_sysctl(void)
6370 if (sd_sysctl_header
)
6371 unregister_sysctl_table(sd_sysctl_header
);
6372 sd_sysctl_header
= NULL
;
6373 if (sd_ctl_dir
[0].child
)
6374 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6377 static void register_sched_domain_sysctl(void)
6380 static void unregister_sched_domain_sysctl(void)
6385 static void set_rq_online(struct rq
*rq
)
6388 const struct sched_class
*class;
6390 cpu_set(rq
->cpu
, rq
->rd
->online
);
6393 for_each_class(class) {
6394 if (class->rq_online
)
6395 class->rq_online(rq
);
6400 static void set_rq_offline(struct rq
*rq
)
6403 const struct sched_class
*class;
6405 for_each_class(class) {
6406 if (class->rq_offline
)
6407 class->rq_offline(rq
);
6410 cpu_clear(rq
->cpu
, rq
->rd
->online
);
6416 * migration_call - callback that gets triggered when a CPU is added.
6417 * Here we can start up the necessary migration thread for the new CPU.
6419 static int __cpuinit
6420 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6422 struct task_struct
*p
;
6423 int cpu
= (long)hcpu
;
6424 unsigned long flags
;
6429 case CPU_UP_PREPARE
:
6430 case CPU_UP_PREPARE_FROZEN
:
6431 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
6434 kthread_bind(p
, cpu
);
6435 /* Must be high prio: stop_machine expects to yield to it. */
6436 rq
= task_rq_lock(p
, &flags
);
6437 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6438 task_rq_unlock(rq
, &flags
);
6439 cpu_rq(cpu
)->migration_thread
= p
;
6443 case CPU_ONLINE_FROZEN
:
6444 /* Strictly unnecessary, as first user will wake it. */
6445 wake_up_process(cpu_rq(cpu
)->migration_thread
);
6447 /* Update our root-domain */
6449 spin_lock_irqsave(&rq
->lock
, flags
);
6451 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6455 spin_unlock_irqrestore(&rq
->lock
, flags
);
6458 #ifdef CONFIG_HOTPLUG_CPU
6459 case CPU_UP_CANCELED
:
6460 case CPU_UP_CANCELED_FROZEN
:
6461 if (!cpu_rq(cpu
)->migration_thread
)
6463 /* Unbind it from offline cpu so it can run. Fall thru. */
6464 kthread_bind(cpu_rq(cpu
)->migration_thread
,
6465 any_online_cpu(cpu_online_map
));
6466 kthread_stop(cpu_rq(cpu
)->migration_thread
);
6467 cpu_rq(cpu
)->migration_thread
= NULL
;
6471 case CPU_DEAD_FROZEN
:
6472 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6473 migrate_live_tasks(cpu
);
6475 kthread_stop(rq
->migration_thread
);
6476 rq
->migration_thread
= NULL
;
6477 /* Idle task back to normal (off runqueue, low prio) */
6478 spin_lock_irq(&rq
->lock
);
6479 update_rq_clock(rq
);
6480 deactivate_task(rq
, rq
->idle
, 0);
6481 rq
->idle
->static_prio
= MAX_PRIO
;
6482 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
6483 rq
->idle
->sched_class
= &idle_sched_class
;
6484 migrate_dead_tasks(cpu
);
6485 spin_unlock_irq(&rq
->lock
);
6487 migrate_nr_uninterruptible(rq
);
6488 BUG_ON(rq
->nr_running
!= 0);
6491 * No need to migrate the tasks: it was best-effort if
6492 * they didn't take sched_hotcpu_mutex. Just wake up
6495 spin_lock_irq(&rq
->lock
);
6496 while (!list_empty(&rq
->migration_queue
)) {
6497 struct migration_req
*req
;
6499 req
= list_entry(rq
->migration_queue
.next
,
6500 struct migration_req
, list
);
6501 list_del_init(&req
->list
);
6502 spin_unlock_irq(&rq
->lock
);
6503 complete(&req
->done
);
6504 spin_lock_irq(&rq
->lock
);
6506 spin_unlock_irq(&rq
->lock
);
6510 case CPU_DYING_FROZEN
:
6511 /* Update our root-domain */
6513 spin_lock_irqsave(&rq
->lock
, flags
);
6515 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6518 spin_unlock_irqrestore(&rq
->lock
, flags
);
6525 /* Register at highest priority so that task migration (migrate_all_tasks)
6526 * happens before everything else.
6528 static struct notifier_block __cpuinitdata migration_notifier
= {
6529 .notifier_call
= migration_call
,
6533 static int __init
migration_init(void)
6535 void *cpu
= (void *)(long)smp_processor_id();
6538 /* Start one for the boot CPU: */
6539 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6540 BUG_ON(err
== NOTIFY_BAD
);
6541 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6542 register_cpu_notifier(&migration_notifier
);
6546 early_initcall(migration_init
);
6551 #ifdef CONFIG_SCHED_DEBUG
6553 static inline const char *sd_level_to_string(enum sched_domain_level lvl
)
6566 case SD_LV_ALLNODES
:
6575 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6576 cpumask_t
*groupmask
)
6578 struct sched_group
*group
= sd
->groups
;
6581 cpulist_scnprintf(str
, sizeof(str
), sd
->span
);
6582 cpus_clear(*groupmask
);
6584 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6586 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6587 printk("does not load-balance\n");
6589 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6594 printk(KERN_CONT
"span %s level %s\n",
6595 str
, sd_level_to_string(sd
->level
));
6597 if (!cpu_isset(cpu
, sd
->span
)) {
6598 printk(KERN_ERR
"ERROR: domain->span does not contain "
6601 if (!cpu_isset(cpu
, group
->cpumask
)) {
6602 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6606 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6610 printk(KERN_ERR
"ERROR: group is NULL\n");
6614 if (!group
->__cpu_power
) {
6615 printk(KERN_CONT
"\n");
6616 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6621 if (!cpus_weight(group
->cpumask
)) {
6622 printk(KERN_CONT
"\n");
6623 printk(KERN_ERR
"ERROR: empty group\n");
6627 if (cpus_intersects(*groupmask
, group
->cpumask
)) {
6628 printk(KERN_CONT
"\n");
6629 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6633 cpus_or(*groupmask
, *groupmask
, group
->cpumask
);
6635 cpulist_scnprintf(str
, sizeof(str
), group
->cpumask
);
6636 printk(KERN_CONT
" %s", str
);
6638 group
= group
->next
;
6639 } while (group
!= sd
->groups
);
6640 printk(KERN_CONT
"\n");
6642 if (!cpus_equal(sd
->span
, *groupmask
))
6643 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6645 if (sd
->parent
&& !cpus_subset(*groupmask
, sd
->parent
->span
))
6646 printk(KERN_ERR
"ERROR: parent span is not a superset "
6647 "of domain->span\n");
6651 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6653 cpumask_t
*groupmask
;
6657 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6661 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6663 groupmask
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6665 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6670 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6679 #else /* !CONFIG_SCHED_DEBUG */
6680 # define sched_domain_debug(sd, cpu) do { } while (0)
6681 #endif /* CONFIG_SCHED_DEBUG */
6683 static int sd_degenerate(struct sched_domain
*sd
)
6685 if (cpus_weight(sd
->span
) == 1)
6688 /* Following flags need at least 2 groups */
6689 if (sd
->flags
& (SD_LOAD_BALANCE
|
6690 SD_BALANCE_NEWIDLE
|
6694 SD_SHARE_PKG_RESOURCES
)) {
6695 if (sd
->groups
!= sd
->groups
->next
)
6699 /* Following flags don't use groups */
6700 if (sd
->flags
& (SD_WAKE_IDLE
|
6709 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6711 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6713 if (sd_degenerate(parent
))
6716 if (!cpus_equal(sd
->span
, parent
->span
))
6719 /* Does parent contain flags not in child? */
6720 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6721 if (cflags
& SD_WAKE_AFFINE
)
6722 pflags
&= ~SD_WAKE_BALANCE
;
6723 /* Flags needing groups don't count if only 1 group in parent */
6724 if (parent
->groups
== parent
->groups
->next
) {
6725 pflags
&= ~(SD_LOAD_BALANCE
|
6726 SD_BALANCE_NEWIDLE
|
6730 SD_SHARE_PKG_RESOURCES
);
6732 if (~cflags
& pflags
)
6738 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6740 unsigned long flags
;
6742 spin_lock_irqsave(&rq
->lock
, flags
);
6745 struct root_domain
*old_rd
= rq
->rd
;
6747 if (cpu_isset(rq
->cpu
, old_rd
->online
))
6750 cpu_clear(rq
->cpu
, old_rd
->span
);
6752 if (atomic_dec_and_test(&old_rd
->refcount
))
6756 atomic_inc(&rd
->refcount
);
6759 cpu_set(rq
->cpu
, rd
->span
);
6760 if (cpu_isset(rq
->cpu
, cpu_online_map
))
6763 spin_unlock_irqrestore(&rq
->lock
, flags
);
6766 static void init_rootdomain(struct root_domain
*rd
)
6768 memset(rd
, 0, sizeof(*rd
));
6770 cpus_clear(rd
->span
);
6771 cpus_clear(rd
->online
);
6773 cpupri_init(&rd
->cpupri
);
6776 static void init_defrootdomain(void)
6778 init_rootdomain(&def_root_domain
);
6779 atomic_set(&def_root_domain
.refcount
, 1);
6782 static struct root_domain
*alloc_rootdomain(void)
6784 struct root_domain
*rd
;
6786 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6790 init_rootdomain(rd
);
6796 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6797 * hold the hotplug lock.
6800 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6802 struct rq
*rq
= cpu_rq(cpu
);
6803 struct sched_domain
*tmp
;
6805 /* Remove the sched domains which do not contribute to scheduling. */
6806 for (tmp
= sd
; tmp
; ) {
6807 struct sched_domain
*parent
= tmp
->parent
;
6811 if (sd_parent_degenerate(tmp
, parent
)) {
6812 tmp
->parent
= parent
->parent
;
6814 parent
->parent
->child
= tmp
;
6819 if (sd
&& sd_degenerate(sd
)) {
6825 sched_domain_debug(sd
, cpu
);
6827 rq_attach_root(rq
, rd
);
6828 rcu_assign_pointer(rq
->sd
, sd
);
6831 /* cpus with isolated domains */
6832 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
6834 /* Setup the mask of cpus configured for isolated domains */
6835 static int __init
isolated_cpu_setup(char *str
)
6837 static int __initdata ints
[NR_CPUS
];
6840 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
6841 cpus_clear(cpu_isolated_map
);
6842 for (i
= 1; i
<= ints
[0]; i
++)
6843 if (ints
[i
] < NR_CPUS
)
6844 cpu_set(ints
[i
], cpu_isolated_map
);
6848 __setup("isolcpus=", isolated_cpu_setup
);
6851 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6852 * to a function which identifies what group(along with sched group) a CPU
6853 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6854 * (due to the fact that we keep track of groups covered with a cpumask_t).
6856 * init_sched_build_groups will build a circular linked list of the groups
6857 * covered by the given span, and will set each group's ->cpumask correctly,
6858 * and ->cpu_power to 0.
6861 init_sched_build_groups(const cpumask_t
*span
, const cpumask_t
*cpu_map
,
6862 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
6863 struct sched_group
**sg
,
6864 cpumask_t
*tmpmask
),
6865 cpumask_t
*covered
, cpumask_t
*tmpmask
)
6867 struct sched_group
*first
= NULL
, *last
= NULL
;
6870 cpus_clear(*covered
);
6872 for_each_cpu_mask_nr(i
, *span
) {
6873 struct sched_group
*sg
;
6874 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6877 if (cpu_isset(i
, *covered
))
6880 cpus_clear(sg
->cpumask
);
6881 sg
->__cpu_power
= 0;
6883 for_each_cpu_mask_nr(j
, *span
) {
6884 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6887 cpu_set(j
, *covered
);
6888 cpu_set(j
, sg
->cpumask
);
6899 #define SD_NODES_PER_DOMAIN 16
6904 * find_next_best_node - find the next node to include in a sched_domain
6905 * @node: node whose sched_domain we're building
6906 * @used_nodes: nodes already in the sched_domain
6908 * Find the next node to include in a given scheduling domain. Simply
6909 * finds the closest node not already in the @used_nodes map.
6911 * Should use nodemask_t.
6913 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6915 int i
, n
, val
, min_val
, best_node
= 0;
6919 for (i
= 0; i
< nr_node_ids
; i
++) {
6920 /* Start at @node */
6921 n
= (node
+ i
) % nr_node_ids
;
6923 if (!nr_cpus_node(n
))
6926 /* Skip already used nodes */
6927 if (node_isset(n
, *used_nodes
))
6930 /* Simple min distance search */
6931 val
= node_distance(node
, n
);
6933 if (val
< min_val
) {
6939 node_set(best_node
, *used_nodes
);
6944 * sched_domain_node_span - get a cpumask for a node's sched_domain
6945 * @node: node whose cpumask we're constructing
6946 * @span: resulting cpumask
6948 * Given a node, construct a good cpumask for its sched_domain to span. It
6949 * should be one that prevents unnecessary balancing, but also spreads tasks
6952 static void sched_domain_node_span(int node
, cpumask_t
*span
)
6954 nodemask_t used_nodes
;
6955 node_to_cpumask_ptr(nodemask
, node
);
6959 nodes_clear(used_nodes
);
6961 cpus_or(*span
, *span
, *nodemask
);
6962 node_set(node
, used_nodes
);
6964 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6965 int next_node
= find_next_best_node(node
, &used_nodes
);
6967 node_to_cpumask_ptr_next(nodemask
, next_node
);
6968 cpus_or(*span
, *span
, *nodemask
);
6971 #endif /* CONFIG_NUMA */
6973 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6976 * SMT sched-domains:
6978 #ifdef CONFIG_SCHED_SMT
6979 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6980 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
6983 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6987 *sg
= &per_cpu(sched_group_cpus
, cpu
);
6990 #endif /* CONFIG_SCHED_SMT */
6993 * multi-core sched-domains:
6995 #ifdef CONFIG_SCHED_MC
6996 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6997 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
6998 #endif /* CONFIG_SCHED_MC */
7000 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7002 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7007 *mask
= per_cpu(cpu_sibling_map
, cpu
);
7008 cpus_and(*mask
, *mask
, *cpu_map
);
7009 group
= first_cpu(*mask
);
7011 *sg
= &per_cpu(sched_group_core
, group
);
7014 #elif defined(CONFIG_SCHED_MC)
7016 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7020 *sg
= &per_cpu(sched_group_core
, cpu
);
7025 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
7026 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
7029 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7033 #ifdef CONFIG_SCHED_MC
7034 *mask
= cpu_coregroup_map(cpu
);
7035 cpus_and(*mask
, *mask
, *cpu_map
);
7036 group
= first_cpu(*mask
);
7037 #elif defined(CONFIG_SCHED_SMT)
7038 *mask
= per_cpu(cpu_sibling_map
, cpu
);
7039 cpus_and(*mask
, *mask
, *cpu_map
);
7040 group
= first_cpu(*mask
);
7045 *sg
= &per_cpu(sched_group_phys
, group
);
7051 * The init_sched_build_groups can't handle what we want to do with node
7052 * groups, so roll our own. Now each node has its own list of groups which
7053 * gets dynamically allocated.
7055 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
7056 static struct sched_group
***sched_group_nodes_bycpu
;
7058 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
7059 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
7061 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
7062 struct sched_group
**sg
, cpumask_t
*nodemask
)
7066 *nodemask
= node_to_cpumask(cpu_to_node(cpu
));
7067 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7068 group
= first_cpu(*nodemask
);
7071 *sg
= &per_cpu(sched_group_allnodes
, group
);
7075 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
7077 struct sched_group
*sg
= group_head
;
7083 for_each_cpu_mask_nr(j
, sg
->cpumask
) {
7084 struct sched_domain
*sd
;
7086 sd
= &per_cpu(phys_domains
, j
);
7087 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
7089 * Only add "power" once for each
7095 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
7098 } while (sg
!= group_head
);
7100 #endif /* CONFIG_NUMA */
7103 /* Free memory allocated for various sched_group structures */
7104 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
7108 for_each_cpu_mask_nr(cpu
, *cpu_map
) {
7109 struct sched_group
**sched_group_nodes
7110 = sched_group_nodes_bycpu
[cpu
];
7112 if (!sched_group_nodes
)
7115 for (i
= 0; i
< nr_node_ids
; i
++) {
7116 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
7118 *nodemask
= node_to_cpumask(i
);
7119 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7120 if (cpus_empty(*nodemask
))
7130 if (oldsg
!= sched_group_nodes
[i
])
7133 kfree(sched_group_nodes
);
7134 sched_group_nodes_bycpu
[cpu
] = NULL
;
7137 #else /* !CONFIG_NUMA */
7138 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
7141 #endif /* CONFIG_NUMA */
7144 * Initialize sched groups cpu_power.
7146 * cpu_power indicates the capacity of sched group, which is used while
7147 * distributing the load between different sched groups in a sched domain.
7148 * Typically cpu_power for all the groups in a sched domain will be same unless
7149 * there are asymmetries in the topology. If there are asymmetries, group
7150 * having more cpu_power will pickup more load compared to the group having
7153 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7154 * the maximum number of tasks a group can handle in the presence of other idle
7155 * or lightly loaded groups in the same sched domain.
7157 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7159 struct sched_domain
*child
;
7160 struct sched_group
*group
;
7162 WARN_ON(!sd
|| !sd
->groups
);
7164 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
7169 sd
->groups
->__cpu_power
= 0;
7172 * For perf policy, if the groups in child domain share resources
7173 * (for example cores sharing some portions of the cache hierarchy
7174 * or SMT), then set this domain groups cpu_power such that each group
7175 * can handle only one task, when there are other idle groups in the
7176 * same sched domain.
7178 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
7180 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
7181 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
7186 * add cpu_power of each child group to this groups cpu_power
7188 group
= child
->groups
;
7190 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
7191 group
= group
->next
;
7192 } while (group
!= child
->groups
);
7196 * Initializers for schedule domains
7197 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7200 #define SD_INIT(sd, type) sd_init_##type(sd)
7201 #define SD_INIT_FUNC(type) \
7202 static noinline void sd_init_##type(struct sched_domain *sd) \
7204 memset(sd, 0, sizeof(*sd)); \
7205 *sd = SD_##type##_INIT; \
7206 sd->level = SD_LV_##type; \
7211 SD_INIT_FUNC(ALLNODES
)
7214 #ifdef CONFIG_SCHED_SMT
7215 SD_INIT_FUNC(SIBLING
)
7217 #ifdef CONFIG_SCHED_MC
7222 * To minimize stack usage kmalloc room for cpumasks and share the
7223 * space as the usage in build_sched_domains() dictates. Used only
7224 * if the amount of space is significant.
7227 cpumask_t tmpmask
; /* make this one first */
7230 cpumask_t this_sibling_map
;
7231 cpumask_t this_core_map
;
7233 cpumask_t send_covered
;
7236 cpumask_t domainspan
;
7238 cpumask_t notcovered
;
7243 #define SCHED_CPUMASK_ALLOC 1
7244 #define SCHED_CPUMASK_FREE(v) kfree(v)
7245 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7247 #define SCHED_CPUMASK_ALLOC 0
7248 #define SCHED_CPUMASK_FREE(v)
7249 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7252 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7253 ((unsigned long)(a) + offsetof(struct allmasks, v))
7255 static int default_relax_domain_level
= -1;
7257 static int __init
setup_relax_domain_level(char *str
)
7261 val
= simple_strtoul(str
, NULL
, 0);
7262 if (val
< SD_LV_MAX
)
7263 default_relax_domain_level
= val
;
7267 __setup("relax_domain_level=", setup_relax_domain_level
);
7269 static void set_domain_attribute(struct sched_domain
*sd
,
7270 struct sched_domain_attr
*attr
)
7274 if (!attr
|| attr
->relax_domain_level
< 0) {
7275 if (default_relax_domain_level
< 0)
7278 request
= default_relax_domain_level
;
7280 request
= attr
->relax_domain_level
;
7281 if (request
< sd
->level
) {
7282 /* turn off idle balance on this domain */
7283 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
7285 /* turn on idle balance on this domain */
7286 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
7291 * Build sched domains for a given set of cpus and attach the sched domains
7292 * to the individual cpus
7294 static int __build_sched_domains(const cpumask_t
*cpu_map
,
7295 struct sched_domain_attr
*attr
)
7298 struct root_domain
*rd
;
7299 SCHED_CPUMASK_DECLARE(allmasks
);
7302 struct sched_group
**sched_group_nodes
= NULL
;
7303 int sd_allnodes
= 0;
7306 * Allocate the per-node list of sched groups
7308 sched_group_nodes
= kcalloc(nr_node_ids
, sizeof(struct sched_group
*),
7310 if (!sched_group_nodes
) {
7311 printk(KERN_WARNING
"Can not alloc sched group node list\n");
7316 rd
= alloc_rootdomain();
7318 printk(KERN_WARNING
"Cannot alloc root domain\n");
7320 kfree(sched_group_nodes
);
7325 #if SCHED_CPUMASK_ALLOC
7326 /* get space for all scratch cpumask variables */
7327 allmasks
= kmalloc(sizeof(*allmasks
), GFP_KERNEL
);
7329 printk(KERN_WARNING
"Cannot alloc cpumask array\n");
7332 kfree(sched_group_nodes
);
7337 tmpmask
= (cpumask_t
*)allmasks
;
7341 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
7345 * Set up domains for cpus specified by the cpu_map.
7347 for_each_cpu_mask_nr(i
, *cpu_map
) {
7348 struct sched_domain
*sd
= NULL
, *p
;
7349 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7351 *nodemask
= node_to_cpumask(cpu_to_node(i
));
7352 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7355 if (cpus_weight(*cpu_map
) >
7356 SD_NODES_PER_DOMAIN
*cpus_weight(*nodemask
)) {
7357 sd
= &per_cpu(allnodes_domains
, i
);
7358 SD_INIT(sd
, ALLNODES
);
7359 set_domain_attribute(sd
, attr
);
7360 sd
->span
= *cpu_map
;
7361 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7367 sd
= &per_cpu(node_domains
, i
);
7369 set_domain_attribute(sd
, attr
);
7370 sched_domain_node_span(cpu_to_node(i
), &sd
->span
);
7374 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7378 sd
= &per_cpu(phys_domains
, i
);
7380 set_domain_attribute(sd
, attr
);
7381 sd
->span
= *nodemask
;
7385 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7387 #ifdef CONFIG_SCHED_MC
7389 sd
= &per_cpu(core_domains
, i
);
7391 set_domain_attribute(sd
, attr
);
7392 sd
->span
= cpu_coregroup_map(i
);
7393 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7396 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7399 #ifdef CONFIG_SCHED_SMT
7401 sd
= &per_cpu(cpu_domains
, i
);
7402 SD_INIT(sd
, SIBLING
);
7403 set_domain_attribute(sd
, attr
);
7404 sd
->span
= per_cpu(cpu_sibling_map
, i
);
7405 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7408 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7412 #ifdef CONFIG_SCHED_SMT
7413 /* Set up CPU (sibling) groups */
7414 for_each_cpu_mask_nr(i
, *cpu_map
) {
7415 SCHED_CPUMASK_VAR(this_sibling_map
, allmasks
);
7416 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7418 *this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
7419 cpus_and(*this_sibling_map
, *this_sibling_map
, *cpu_map
);
7420 if (i
!= first_cpu(*this_sibling_map
))
7423 init_sched_build_groups(this_sibling_map
, cpu_map
,
7425 send_covered
, tmpmask
);
7429 #ifdef CONFIG_SCHED_MC
7430 /* Set up multi-core groups */
7431 for_each_cpu_mask_nr(i
, *cpu_map
) {
7432 SCHED_CPUMASK_VAR(this_core_map
, allmasks
);
7433 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7435 *this_core_map
= cpu_coregroup_map(i
);
7436 cpus_and(*this_core_map
, *this_core_map
, *cpu_map
);
7437 if (i
!= first_cpu(*this_core_map
))
7440 init_sched_build_groups(this_core_map
, cpu_map
,
7442 send_covered
, tmpmask
);
7446 /* Set up physical groups */
7447 for (i
= 0; i
< nr_node_ids
; i
++) {
7448 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7449 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7451 *nodemask
= node_to_cpumask(i
);
7452 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7453 if (cpus_empty(*nodemask
))
7456 init_sched_build_groups(nodemask
, cpu_map
,
7458 send_covered
, tmpmask
);
7462 /* Set up node groups */
7464 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7466 init_sched_build_groups(cpu_map
, cpu_map
,
7467 &cpu_to_allnodes_group
,
7468 send_covered
, tmpmask
);
7471 for (i
= 0; i
< nr_node_ids
; i
++) {
7472 /* Set up node groups */
7473 struct sched_group
*sg
, *prev
;
7474 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7475 SCHED_CPUMASK_VAR(domainspan
, allmasks
);
7476 SCHED_CPUMASK_VAR(covered
, allmasks
);
7479 *nodemask
= node_to_cpumask(i
);
7480 cpus_clear(*covered
);
7482 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7483 if (cpus_empty(*nodemask
)) {
7484 sched_group_nodes
[i
] = NULL
;
7488 sched_domain_node_span(i
, domainspan
);
7489 cpus_and(*domainspan
, *domainspan
, *cpu_map
);
7491 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
7493 printk(KERN_WARNING
"Can not alloc domain group for "
7497 sched_group_nodes
[i
] = sg
;
7498 for_each_cpu_mask_nr(j
, *nodemask
) {
7499 struct sched_domain
*sd
;
7501 sd
= &per_cpu(node_domains
, j
);
7504 sg
->__cpu_power
= 0;
7505 sg
->cpumask
= *nodemask
;
7507 cpus_or(*covered
, *covered
, *nodemask
);
7510 for (j
= 0; j
< nr_node_ids
; j
++) {
7511 SCHED_CPUMASK_VAR(notcovered
, allmasks
);
7512 int n
= (i
+ j
) % nr_node_ids
;
7513 node_to_cpumask_ptr(pnodemask
, n
);
7515 cpus_complement(*notcovered
, *covered
);
7516 cpus_and(*tmpmask
, *notcovered
, *cpu_map
);
7517 cpus_and(*tmpmask
, *tmpmask
, *domainspan
);
7518 if (cpus_empty(*tmpmask
))
7521 cpus_and(*tmpmask
, *tmpmask
, *pnodemask
);
7522 if (cpus_empty(*tmpmask
))
7525 sg
= kmalloc_node(sizeof(struct sched_group
),
7529 "Can not alloc domain group for node %d\n", j
);
7532 sg
->__cpu_power
= 0;
7533 sg
->cpumask
= *tmpmask
;
7534 sg
->next
= prev
->next
;
7535 cpus_or(*covered
, *covered
, *tmpmask
);
7542 /* Calculate CPU power for physical packages and nodes */
7543 #ifdef CONFIG_SCHED_SMT
7544 for_each_cpu_mask_nr(i
, *cpu_map
) {
7545 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
7547 init_sched_groups_power(i
, sd
);
7550 #ifdef CONFIG_SCHED_MC
7551 for_each_cpu_mask_nr(i
, *cpu_map
) {
7552 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
7554 init_sched_groups_power(i
, sd
);
7558 for_each_cpu_mask_nr(i
, *cpu_map
) {
7559 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
7561 init_sched_groups_power(i
, sd
);
7565 for (i
= 0; i
< nr_node_ids
; i
++)
7566 init_numa_sched_groups_power(sched_group_nodes
[i
]);
7569 struct sched_group
*sg
;
7571 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
,
7573 init_numa_sched_groups_power(sg
);
7577 /* Attach the domains */
7578 for_each_cpu_mask_nr(i
, *cpu_map
) {
7579 struct sched_domain
*sd
;
7580 #ifdef CONFIG_SCHED_SMT
7581 sd
= &per_cpu(cpu_domains
, i
);
7582 #elif defined(CONFIG_SCHED_MC)
7583 sd
= &per_cpu(core_domains
, i
);
7585 sd
= &per_cpu(phys_domains
, i
);
7587 cpu_attach_domain(sd
, rd
, i
);
7590 SCHED_CPUMASK_FREE((void *)allmasks
);
7595 free_sched_groups(cpu_map
, tmpmask
);
7596 SCHED_CPUMASK_FREE((void *)allmasks
);
7601 static int build_sched_domains(const cpumask_t
*cpu_map
)
7603 return __build_sched_domains(cpu_map
, NULL
);
7606 static cpumask_t
*doms_cur
; /* current sched domains */
7607 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7608 static struct sched_domain_attr
*dattr_cur
;
7609 /* attribues of custom domains in 'doms_cur' */
7612 * Special case: If a kmalloc of a doms_cur partition (array of
7613 * cpumask_t) fails, then fallback to a single sched domain,
7614 * as determined by the single cpumask_t fallback_doms.
7616 static cpumask_t fallback_doms
;
7618 void __attribute__((weak
)) arch_update_cpu_topology(void)
7623 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7624 * For now this just excludes isolated cpus, but could be used to
7625 * exclude other special cases in the future.
7627 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
7631 arch_update_cpu_topology();
7633 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
7635 doms_cur
= &fallback_doms
;
7636 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
7638 err
= build_sched_domains(doms_cur
);
7639 register_sched_domain_sysctl();
7644 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
,
7647 free_sched_groups(cpu_map
, tmpmask
);
7651 * Detach sched domains from a group of cpus specified in cpu_map
7652 * These cpus will now be attached to the NULL domain
7654 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
7659 unregister_sched_domain_sysctl();
7661 for_each_cpu_mask_nr(i
, *cpu_map
)
7662 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7663 synchronize_sched();
7664 arch_destroy_sched_domains(cpu_map
, &tmpmask
);
7667 /* handle null as "default" */
7668 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7669 struct sched_domain_attr
*new, int idx_new
)
7671 struct sched_domain_attr tmp
;
7678 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7679 new ? (new + idx_new
) : &tmp
,
7680 sizeof(struct sched_domain_attr
));
7684 * Partition sched domains as specified by the 'ndoms_new'
7685 * cpumasks in the array doms_new[] of cpumasks. This compares
7686 * doms_new[] to the current sched domain partitioning, doms_cur[].
7687 * It destroys each deleted domain and builds each new domain.
7689 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7690 * The masks don't intersect (don't overlap.) We should setup one
7691 * sched domain for each mask. CPUs not in any of the cpumasks will
7692 * not be load balanced. If the same cpumask appears both in the
7693 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7696 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7697 * ownership of it and will kfree it when done with it. If the caller
7698 * failed the kmalloc call, then it can pass in doms_new == NULL &&
7699 * ndoms_new == 1, and partition_sched_domains() will fallback to
7700 * the single partition 'fallback_doms', it also forces the domains
7703 * If doms_new == NULL it will be replaced with cpu_online_map.
7704 * ndoms_new == 0 is a special case for destroying existing domains,
7705 * and it will not create the default domain.
7707 * Call with hotplug lock held
7709 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
,
7710 struct sched_domain_attr
*dattr_new
)
7714 mutex_lock(&sched_domains_mutex
);
7716 /* always unregister in case we don't destroy any domains */
7717 unregister_sched_domain_sysctl();
7719 n
= doms_new
? ndoms_new
: 0;
7721 /* Destroy deleted domains */
7722 for (i
= 0; i
< ndoms_cur
; i
++) {
7723 for (j
= 0; j
< n
; j
++) {
7724 if (cpus_equal(doms_cur
[i
], doms_new
[j
])
7725 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7728 /* no match - a current sched domain not in new doms_new[] */
7729 detach_destroy_domains(doms_cur
+ i
);
7734 if (doms_new
== NULL
) {
7736 doms_new
= &fallback_doms
;
7737 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
7741 /* Build new domains */
7742 for (i
= 0; i
< ndoms_new
; i
++) {
7743 for (j
= 0; j
< ndoms_cur
; j
++) {
7744 if (cpus_equal(doms_new
[i
], doms_cur
[j
])
7745 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7748 /* no match - add a new doms_new */
7749 __build_sched_domains(doms_new
+ i
,
7750 dattr_new
? dattr_new
+ i
: NULL
);
7755 /* Remember the new sched domains */
7756 if (doms_cur
!= &fallback_doms
)
7758 kfree(dattr_cur
); /* kfree(NULL) is safe */
7759 doms_cur
= doms_new
;
7760 dattr_cur
= dattr_new
;
7761 ndoms_cur
= ndoms_new
;
7763 register_sched_domain_sysctl();
7765 mutex_unlock(&sched_domains_mutex
);
7768 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7769 int arch_reinit_sched_domains(void)
7773 /* Destroy domains first to force the rebuild */
7774 partition_sched_domains(0, NULL
, NULL
);
7776 rebuild_sched_domains();
7782 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7786 if (buf
[0] != '0' && buf
[0] != '1')
7790 sched_smt_power_savings
= (buf
[0] == '1');
7792 sched_mc_power_savings
= (buf
[0] == '1');
7794 ret
= arch_reinit_sched_domains();
7796 return ret
? ret
: count
;
7799 #ifdef CONFIG_SCHED_MC
7800 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7803 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7805 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7806 const char *buf
, size_t count
)
7808 return sched_power_savings_store(buf
, count
, 0);
7810 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7811 sched_mc_power_savings_show
,
7812 sched_mc_power_savings_store
);
7815 #ifdef CONFIG_SCHED_SMT
7816 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7819 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7821 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7822 const char *buf
, size_t count
)
7824 return sched_power_savings_store(buf
, count
, 1);
7826 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7827 sched_smt_power_savings_show
,
7828 sched_smt_power_savings_store
);
7831 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7835 #ifdef CONFIG_SCHED_SMT
7837 err
= sysfs_create_file(&cls
->kset
.kobj
,
7838 &attr_sched_smt_power_savings
.attr
);
7840 #ifdef CONFIG_SCHED_MC
7841 if (!err
&& mc_capable())
7842 err
= sysfs_create_file(&cls
->kset
.kobj
,
7843 &attr_sched_mc_power_savings
.attr
);
7847 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7849 #ifndef CONFIG_CPUSETS
7851 * Add online and remove offline CPUs from the scheduler domains.
7852 * When cpusets are enabled they take over this function.
7854 static int update_sched_domains(struct notifier_block
*nfb
,
7855 unsigned long action
, void *hcpu
)
7859 case CPU_ONLINE_FROZEN
:
7861 case CPU_DEAD_FROZEN
:
7862 partition_sched_domains(1, NULL
, NULL
);
7871 static int update_runtime(struct notifier_block
*nfb
,
7872 unsigned long action
, void *hcpu
)
7874 int cpu
= (int)(long)hcpu
;
7877 case CPU_DOWN_PREPARE
:
7878 case CPU_DOWN_PREPARE_FROZEN
:
7879 disable_runtime(cpu_rq(cpu
));
7882 case CPU_DOWN_FAILED
:
7883 case CPU_DOWN_FAILED_FROZEN
:
7885 case CPU_ONLINE_FROZEN
:
7886 enable_runtime(cpu_rq(cpu
));
7894 void __init
sched_init_smp(void)
7896 cpumask_t non_isolated_cpus
;
7898 #if defined(CONFIG_NUMA)
7899 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7901 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7904 mutex_lock(&sched_domains_mutex
);
7905 arch_init_sched_domains(&cpu_online_map
);
7906 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
7907 if (cpus_empty(non_isolated_cpus
))
7908 cpu_set(smp_processor_id(), non_isolated_cpus
);
7909 mutex_unlock(&sched_domains_mutex
);
7912 #ifndef CONFIG_CPUSETS
7913 /* XXX: Theoretical race here - CPU may be hotplugged now */
7914 hotcpu_notifier(update_sched_domains
, 0);
7917 /* RT runtime code needs to handle some hotplug events */
7918 hotcpu_notifier(update_runtime
, 0);
7922 /* Move init over to a non-isolated CPU */
7923 if (set_cpus_allowed_ptr(current
, &non_isolated_cpus
) < 0)
7925 sched_init_granularity();
7928 void __init
sched_init_smp(void)
7930 sched_init_granularity();
7932 #endif /* CONFIG_SMP */
7934 int in_sched_functions(unsigned long addr
)
7936 return in_lock_functions(addr
) ||
7937 (addr
>= (unsigned long)__sched_text_start
7938 && addr
< (unsigned long)__sched_text_end
);
7941 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7943 cfs_rq
->tasks_timeline
= RB_ROOT
;
7944 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7945 #ifdef CONFIG_FAIR_GROUP_SCHED
7948 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7951 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7953 struct rt_prio_array
*array
;
7956 array
= &rt_rq
->active
;
7957 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7958 INIT_LIST_HEAD(array
->queue
+ i
);
7959 __clear_bit(i
, array
->bitmap
);
7961 /* delimiter for bitsearch: */
7962 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7964 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7965 rt_rq
->highest_prio
= MAX_RT_PRIO
;
7968 rt_rq
->rt_nr_migratory
= 0;
7969 rt_rq
->overloaded
= 0;
7973 rt_rq
->rt_throttled
= 0;
7974 rt_rq
->rt_runtime
= 0;
7975 spin_lock_init(&rt_rq
->rt_runtime_lock
);
7977 #ifdef CONFIG_RT_GROUP_SCHED
7978 rt_rq
->rt_nr_boosted
= 0;
7983 #ifdef CONFIG_FAIR_GROUP_SCHED
7984 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7985 struct sched_entity
*se
, int cpu
, int add
,
7986 struct sched_entity
*parent
)
7988 struct rq
*rq
= cpu_rq(cpu
);
7989 tg
->cfs_rq
[cpu
] = cfs_rq
;
7990 init_cfs_rq(cfs_rq
, rq
);
7993 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7996 /* se could be NULL for init_task_group */
8001 se
->cfs_rq
= &rq
->cfs
;
8003 se
->cfs_rq
= parent
->my_q
;
8006 se
->load
.weight
= tg
->shares
;
8007 se
->load
.inv_weight
= 0;
8008 se
->parent
= parent
;
8012 #ifdef CONFIG_RT_GROUP_SCHED
8013 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
8014 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
8015 struct sched_rt_entity
*parent
)
8017 struct rq
*rq
= cpu_rq(cpu
);
8019 tg
->rt_rq
[cpu
] = rt_rq
;
8020 init_rt_rq(rt_rq
, rq
);
8022 rt_rq
->rt_se
= rt_se
;
8023 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8025 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
8027 tg
->rt_se
[cpu
] = rt_se
;
8032 rt_se
->rt_rq
= &rq
->rt
;
8034 rt_se
->rt_rq
= parent
->my_q
;
8036 rt_se
->my_q
= rt_rq
;
8037 rt_se
->parent
= parent
;
8038 INIT_LIST_HEAD(&rt_se
->run_list
);
8042 void __init
sched_init(void)
8045 unsigned long alloc_size
= 0, ptr
;
8047 #ifdef CONFIG_FAIR_GROUP_SCHED
8048 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8050 #ifdef CONFIG_RT_GROUP_SCHED
8051 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8053 #ifdef CONFIG_USER_SCHED
8057 * As sched_init() is called before page_alloc is setup,
8058 * we use alloc_bootmem().
8061 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
8063 #ifdef CONFIG_FAIR_GROUP_SCHED
8064 init_task_group
.se
= (struct sched_entity
**)ptr
;
8065 ptr
+= nr_cpu_ids
* sizeof(void **);
8067 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8068 ptr
+= nr_cpu_ids
* sizeof(void **);
8070 #ifdef CONFIG_USER_SCHED
8071 root_task_group
.se
= (struct sched_entity
**)ptr
;
8072 ptr
+= nr_cpu_ids
* sizeof(void **);
8074 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8075 ptr
+= nr_cpu_ids
* sizeof(void **);
8076 #endif /* CONFIG_USER_SCHED */
8077 #endif /* CONFIG_FAIR_GROUP_SCHED */
8078 #ifdef CONFIG_RT_GROUP_SCHED
8079 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8080 ptr
+= nr_cpu_ids
* sizeof(void **);
8082 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8083 ptr
+= nr_cpu_ids
* sizeof(void **);
8085 #ifdef CONFIG_USER_SCHED
8086 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8087 ptr
+= nr_cpu_ids
* sizeof(void **);
8089 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8090 ptr
+= nr_cpu_ids
* sizeof(void **);
8091 #endif /* CONFIG_USER_SCHED */
8092 #endif /* CONFIG_RT_GROUP_SCHED */
8096 init_defrootdomain();
8099 init_rt_bandwidth(&def_rt_bandwidth
,
8100 global_rt_period(), global_rt_runtime());
8102 #ifdef CONFIG_RT_GROUP_SCHED
8103 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
8104 global_rt_period(), global_rt_runtime());
8105 #ifdef CONFIG_USER_SCHED
8106 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8107 global_rt_period(), RUNTIME_INF
);
8108 #endif /* CONFIG_USER_SCHED */
8109 #endif /* CONFIG_RT_GROUP_SCHED */
8111 #ifdef CONFIG_GROUP_SCHED
8112 list_add(&init_task_group
.list
, &task_groups
);
8113 INIT_LIST_HEAD(&init_task_group
.children
);
8115 #ifdef CONFIG_USER_SCHED
8116 INIT_LIST_HEAD(&root_task_group
.children
);
8117 init_task_group
.parent
= &root_task_group
;
8118 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
8119 #endif /* CONFIG_USER_SCHED */
8120 #endif /* CONFIG_GROUP_SCHED */
8122 for_each_possible_cpu(i
) {
8126 spin_lock_init(&rq
->lock
);
8128 init_cfs_rq(&rq
->cfs
, rq
);
8129 init_rt_rq(&rq
->rt
, rq
);
8130 #ifdef CONFIG_FAIR_GROUP_SCHED
8131 init_task_group
.shares
= init_task_group_load
;
8132 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8133 #ifdef CONFIG_CGROUP_SCHED
8135 * How much cpu bandwidth does init_task_group get?
8137 * In case of task-groups formed thr' the cgroup filesystem, it
8138 * gets 100% of the cpu resources in the system. This overall
8139 * system cpu resource is divided among the tasks of
8140 * init_task_group and its child task-groups in a fair manner,
8141 * based on each entity's (task or task-group's) weight
8142 * (se->load.weight).
8144 * In other words, if init_task_group has 10 tasks of weight
8145 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8146 * then A0's share of the cpu resource is:
8148 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8150 * We achieve this by letting init_task_group's tasks sit
8151 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8153 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
8154 #elif defined CONFIG_USER_SCHED
8155 root_task_group
.shares
= NICE_0_LOAD
;
8156 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
8158 * In case of task-groups formed thr' the user id of tasks,
8159 * init_task_group represents tasks belonging to root user.
8160 * Hence it forms a sibling of all subsequent groups formed.
8161 * In this case, init_task_group gets only a fraction of overall
8162 * system cpu resource, based on the weight assigned to root
8163 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8164 * by letting tasks of init_task_group sit in a separate cfs_rq
8165 * (init_cfs_rq) and having one entity represent this group of
8166 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8168 init_tg_cfs_entry(&init_task_group
,
8169 &per_cpu(init_cfs_rq
, i
),
8170 &per_cpu(init_sched_entity
, i
), i
, 1,
8171 root_task_group
.se
[i
]);
8174 #endif /* CONFIG_FAIR_GROUP_SCHED */
8176 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8177 #ifdef CONFIG_RT_GROUP_SCHED
8178 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
8179 #ifdef CONFIG_CGROUP_SCHED
8180 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
8181 #elif defined CONFIG_USER_SCHED
8182 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
8183 init_tg_rt_entry(&init_task_group
,
8184 &per_cpu(init_rt_rq
, i
),
8185 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
8186 root_task_group
.rt_se
[i
]);
8190 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
8191 rq
->cpu_load
[j
] = 0;
8195 rq
->active_balance
= 0;
8196 rq
->next_balance
= jiffies
;
8200 rq
->migration_thread
= NULL
;
8201 INIT_LIST_HEAD(&rq
->migration_queue
);
8202 rq_attach_root(rq
, &def_root_domain
);
8205 atomic_set(&rq
->nr_iowait
, 0);
8208 set_load_weight(&init_task
);
8210 #ifdef CONFIG_PREEMPT_NOTIFIERS
8211 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
8215 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8218 #ifdef CONFIG_RT_MUTEXES
8219 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
8223 * The boot idle thread does lazy MMU switching as well:
8225 atomic_inc(&init_mm
.mm_count
);
8226 enter_lazy_tlb(&init_mm
, current
);
8229 * Make us the idle thread. Technically, schedule() should not be
8230 * called from this thread, however somewhere below it might be,
8231 * but because we are the idle thread, we just pick up running again
8232 * when this runqueue becomes "idle".
8234 init_idle(current
, smp_processor_id());
8236 * During early bootup we pretend to be a normal task:
8238 current
->sched_class
= &fair_sched_class
;
8240 scheduler_running
= 1;
8243 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8244 void __might_sleep(char *file
, int line
)
8247 static unsigned long prev_jiffy
; /* ratelimiting */
8249 if ((in_atomic() || irqs_disabled()) &&
8250 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
8251 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8253 prev_jiffy
= jiffies
;
8254 printk(KERN_ERR
"BUG: sleeping function called from invalid"
8255 " context at %s:%d\n", file
, line
);
8256 printk("in_atomic():%d, irqs_disabled():%d\n",
8257 in_atomic(), irqs_disabled());
8258 debug_show_held_locks(current
);
8259 if (irqs_disabled())
8260 print_irqtrace_events(current
);
8265 EXPORT_SYMBOL(__might_sleep
);
8268 #ifdef CONFIG_MAGIC_SYSRQ
8269 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8273 update_rq_clock(rq
);
8274 on_rq
= p
->se
.on_rq
;
8276 deactivate_task(rq
, p
, 0);
8277 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8279 activate_task(rq
, p
, 0);
8280 resched_task(rq
->curr
);
8284 void normalize_rt_tasks(void)
8286 struct task_struct
*g
, *p
;
8287 unsigned long flags
;
8290 read_lock_irqsave(&tasklist_lock
, flags
);
8291 do_each_thread(g
, p
) {
8293 * Only normalize user tasks:
8298 p
->se
.exec_start
= 0;
8299 #ifdef CONFIG_SCHEDSTATS
8300 p
->se
.wait_start
= 0;
8301 p
->se
.sleep_start
= 0;
8302 p
->se
.block_start
= 0;
8307 * Renice negative nice level userspace
8310 if (TASK_NICE(p
) < 0 && p
->mm
)
8311 set_user_nice(p
, 0);
8315 spin_lock(&p
->pi_lock
);
8316 rq
= __task_rq_lock(p
);
8318 normalize_task(rq
, p
);
8320 __task_rq_unlock(rq
);
8321 spin_unlock(&p
->pi_lock
);
8322 } while_each_thread(g
, p
);
8324 read_unlock_irqrestore(&tasklist_lock
, flags
);
8327 #endif /* CONFIG_MAGIC_SYSRQ */
8331 * These functions are only useful for the IA64 MCA handling.
8333 * They can only be called when the whole system has been
8334 * stopped - every CPU needs to be quiescent, and no scheduling
8335 * activity can take place. Using them for anything else would
8336 * be a serious bug, and as a result, they aren't even visible
8337 * under any other configuration.
8341 * curr_task - return the current task for a given cpu.
8342 * @cpu: the processor in question.
8344 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8346 struct task_struct
*curr_task(int cpu
)
8348 return cpu_curr(cpu
);
8352 * set_curr_task - set the current task for a given cpu.
8353 * @cpu: the processor in question.
8354 * @p: the task pointer to set.
8356 * Description: This function must only be used when non-maskable interrupts
8357 * are serviced on a separate stack. It allows the architecture to switch the
8358 * notion of the current task on a cpu in a non-blocking manner. This function
8359 * must be called with all CPU's synchronized, and interrupts disabled, the
8360 * and caller must save the original value of the current task (see
8361 * curr_task() above) and restore that value before reenabling interrupts and
8362 * re-starting the system.
8364 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8366 void set_curr_task(int cpu
, struct task_struct
*p
)
8373 #ifdef CONFIG_FAIR_GROUP_SCHED
8374 static void free_fair_sched_group(struct task_group
*tg
)
8378 for_each_possible_cpu(i
) {
8380 kfree(tg
->cfs_rq
[i
]);
8390 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8392 struct cfs_rq
*cfs_rq
;
8393 struct sched_entity
*se
, *parent_se
;
8397 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8400 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8404 tg
->shares
= NICE_0_LOAD
;
8406 for_each_possible_cpu(i
) {
8409 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
),
8410 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8414 se
= kmalloc_node(sizeof(struct sched_entity
),
8415 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8419 parent_se
= parent
? parent
->se
[i
] : NULL
;
8420 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent_se
);
8429 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8431 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8432 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8435 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8437 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8439 #else /* !CONFG_FAIR_GROUP_SCHED */
8440 static inline void free_fair_sched_group(struct task_group
*tg
)
8445 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8450 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8454 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8457 #endif /* CONFIG_FAIR_GROUP_SCHED */
8459 #ifdef CONFIG_RT_GROUP_SCHED
8460 static void free_rt_sched_group(struct task_group
*tg
)
8464 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8466 for_each_possible_cpu(i
) {
8468 kfree(tg
->rt_rq
[i
]);
8470 kfree(tg
->rt_se
[i
]);
8478 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8480 struct rt_rq
*rt_rq
;
8481 struct sched_rt_entity
*rt_se
, *parent_se
;
8485 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8488 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8492 init_rt_bandwidth(&tg
->rt_bandwidth
,
8493 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8495 for_each_possible_cpu(i
) {
8498 rt_rq
= kmalloc_node(sizeof(struct rt_rq
),
8499 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8503 rt_se
= kmalloc_node(sizeof(struct sched_rt_entity
),
8504 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8508 parent_se
= parent
? parent
->rt_se
[i
] : NULL
;
8509 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent_se
);
8518 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8520 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8521 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8524 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8526 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8528 #else /* !CONFIG_RT_GROUP_SCHED */
8529 static inline void free_rt_sched_group(struct task_group
*tg
)
8534 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8539 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8543 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8546 #endif /* CONFIG_RT_GROUP_SCHED */
8548 #ifdef CONFIG_GROUP_SCHED
8549 static void free_sched_group(struct task_group
*tg
)
8551 free_fair_sched_group(tg
);
8552 free_rt_sched_group(tg
);
8556 /* allocate runqueue etc for a new task group */
8557 struct task_group
*sched_create_group(struct task_group
*parent
)
8559 struct task_group
*tg
;
8560 unsigned long flags
;
8563 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8565 return ERR_PTR(-ENOMEM
);
8567 if (!alloc_fair_sched_group(tg
, parent
))
8570 if (!alloc_rt_sched_group(tg
, parent
))
8573 spin_lock_irqsave(&task_group_lock
, flags
);
8574 for_each_possible_cpu(i
) {
8575 register_fair_sched_group(tg
, i
);
8576 register_rt_sched_group(tg
, i
);
8578 list_add_rcu(&tg
->list
, &task_groups
);
8580 WARN_ON(!parent
); /* root should already exist */
8582 tg
->parent
= parent
;
8583 INIT_LIST_HEAD(&tg
->children
);
8584 list_add_rcu(&tg
->siblings
, &parent
->children
);
8585 spin_unlock_irqrestore(&task_group_lock
, flags
);
8590 free_sched_group(tg
);
8591 return ERR_PTR(-ENOMEM
);
8594 /* rcu callback to free various structures associated with a task group */
8595 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8597 /* now it should be safe to free those cfs_rqs */
8598 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8601 /* Destroy runqueue etc associated with a task group */
8602 void sched_destroy_group(struct task_group
*tg
)
8604 unsigned long flags
;
8607 spin_lock_irqsave(&task_group_lock
, flags
);
8608 for_each_possible_cpu(i
) {
8609 unregister_fair_sched_group(tg
, i
);
8610 unregister_rt_sched_group(tg
, i
);
8612 list_del_rcu(&tg
->list
);
8613 list_del_rcu(&tg
->siblings
);
8614 spin_unlock_irqrestore(&task_group_lock
, flags
);
8616 /* wait for possible concurrent references to cfs_rqs complete */
8617 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8620 /* change task's runqueue when it moves between groups.
8621 * The caller of this function should have put the task in its new group
8622 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8623 * reflect its new group.
8625 void sched_move_task(struct task_struct
*tsk
)
8628 unsigned long flags
;
8631 rq
= task_rq_lock(tsk
, &flags
);
8633 update_rq_clock(rq
);
8635 running
= task_current(rq
, tsk
);
8636 on_rq
= tsk
->se
.on_rq
;
8639 dequeue_task(rq
, tsk
, 0);
8640 if (unlikely(running
))
8641 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8643 set_task_rq(tsk
, task_cpu(tsk
));
8645 #ifdef CONFIG_FAIR_GROUP_SCHED
8646 if (tsk
->sched_class
->moved_group
)
8647 tsk
->sched_class
->moved_group(tsk
);
8650 if (unlikely(running
))
8651 tsk
->sched_class
->set_curr_task(rq
);
8653 enqueue_task(rq
, tsk
, 0);
8655 task_rq_unlock(rq
, &flags
);
8657 #endif /* CONFIG_GROUP_SCHED */
8659 #ifdef CONFIG_FAIR_GROUP_SCHED
8660 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8662 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8667 dequeue_entity(cfs_rq
, se
, 0);
8669 se
->load
.weight
= shares
;
8670 se
->load
.inv_weight
= 0;
8673 enqueue_entity(cfs_rq
, se
, 0);
8676 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8678 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8679 struct rq
*rq
= cfs_rq
->rq
;
8680 unsigned long flags
;
8682 spin_lock_irqsave(&rq
->lock
, flags
);
8683 __set_se_shares(se
, shares
);
8684 spin_unlock_irqrestore(&rq
->lock
, flags
);
8687 static DEFINE_MUTEX(shares_mutex
);
8689 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8692 unsigned long flags
;
8695 * We can't change the weight of the root cgroup.
8700 if (shares
< MIN_SHARES
)
8701 shares
= MIN_SHARES
;
8702 else if (shares
> MAX_SHARES
)
8703 shares
= MAX_SHARES
;
8705 mutex_lock(&shares_mutex
);
8706 if (tg
->shares
== shares
)
8709 spin_lock_irqsave(&task_group_lock
, flags
);
8710 for_each_possible_cpu(i
)
8711 unregister_fair_sched_group(tg
, i
);
8712 list_del_rcu(&tg
->siblings
);
8713 spin_unlock_irqrestore(&task_group_lock
, flags
);
8715 /* wait for any ongoing reference to this group to finish */
8716 synchronize_sched();
8719 * Now we are free to modify the group's share on each cpu
8720 * w/o tripping rebalance_share or load_balance_fair.
8722 tg
->shares
= shares
;
8723 for_each_possible_cpu(i
) {
8727 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8728 set_se_shares(tg
->se
[i
], shares
);
8732 * Enable load balance activity on this group, by inserting it back on
8733 * each cpu's rq->leaf_cfs_rq_list.
8735 spin_lock_irqsave(&task_group_lock
, flags
);
8736 for_each_possible_cpu(i
)
8737 register_fair_sched_group(tg
, i
);
8738 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8739 spin_unlock_irqrestore(&task_group_lock
, flags
);
8741 mutex_unlock(&shares_mutex
);
8745 unsigned long sched_group_shares(struct task_group
*tg
)
8751 #ifdef CONFIG_RT_GROUP_SCHED
8753 * Ensure that the real time constraints are schedulable.
8755 static DEFINE_MUTEX(rt_constraints_mutex
);
8757 static unsigned long to_ratio(u64 period
, u64 runtime
)
8759 if (runtime
== RUNTIME_INF
)
8762 return div64_u64(runtime
<< 16, period
);
8765 #ifdef CONFIG_CGROUP_SCHED
8766 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8768 struct task_group
*tgi
, *parent
= tg
->parent
;
8769 unsigned long total
= 0;
8772 if (global_rt_period() < period
)
8775 return to_ratio(period
, runtime
) <
8776 to_ratio(global_rt_period(), global_rt_runtime());
8779 if (ktime_to_ns(parent
->rt_bandwidth
.rt_period
) < period
)
8783 list_for_each_entry_rcu(tgi
, &parent
->children
, siblings
) {
8787 total
+= to_ratio(ktime_to_ns(tgi
->rt_bandwidth
.rt_period
),
8788 tgi
->rt_bandwidth
.rt_runtime
);
8792 return total
+ to_ratio(period
, runtime
) <=
8793 to_ratio(ktime_to_ns(parent
->rt_bandwidth
.rt_period
),
8794 parent
->rt_bandwidth
.rt_runtime
);
8796 #elif defined CONFIG_USER_SCHED
8797 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8799 struct task_group
*tgi
;
8800 unsigned long total
= 0;
8801 unsigned long global_ratio
=
8802 to_ratio(global_rt_period(), global_rt_runtime());
8805 list_for_each_entry_rcu(tgi
, &task_groups
, list
) {
8809 total
+= to_ratio(ktime_to_ns(tgi
->rt_bandwidth
.rt_period
),
8810 tgi
->rt_bandwidth
.rt_runtime
);
8814 return total
+ to_ratio(period
, runtime
) < global_ratio
;
8818 /* Must be called with tasklist_lock held */
8819 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8821 struct task_struct
*g
, *p
;
8822 do_each_thread(g
, p
) {
8823 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8825 } while_each_thread(g
, p
);
8829 static int tg_set_bandwidth(struct task_group
*tg
,
8830 u64 rt_period
, u64 rt_runtime
)
8834 mutex_lock(&rt_constraints_mutex
);
8835 read_lock(&tasklist_lock
);
8836 if (rt_runtime
== 0 && tg_has_rt_tasks(tg
)) {
8840 if (!__rt_schedulable(tg
, rt_period
, rt_runtime
)) {
8845 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8846 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8847 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8849 for_each_possible_cpu(i
) {
8850 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8852 spin_lock(&rt_rq
->rt_runtime_lock
);
8853 rt_rq
->rt_runtime
= rt_runtime
;
8854 spin_unlock(&rt_rq
->rt_runtime_lock
);
8856 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8858 read_unlock(&tasklist_lock
);
8859 mutex_unlock(&rt_constraints_mutex
);
8864 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8866 u64 rt_runtime
, rt_period
;
8868 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8869 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8870 if (rt_runtime_us
< 0)
8871 rt_runtime
= RUNTIME_INF
;
8873 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8876 long sched_group_rt_runtime(struct task_group
*tg
)
8880 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8883 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8884 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8885 return rt_runtime_us
;
8888 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8890 u64 rt_runtime
, rt_period
;
8892 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8893 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8898 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8901 long sched_group_rt_period(struct task_group
*tg
)
8905 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8906 do_div(rt_period_us
, NSEC_PER_USEC
);
8907 return rt_period_us
;
8910 static int sched_rt_global_constraints(void)
8912 struct task_group
*tg
= &root_task_group
;
8913 u64 rt_runtime
, rt_period
;
8916 if (sysctl_sched_rt_period
<= 0)
8919 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8920 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8922 mutex_lock(&rt_constraints_mutex
);
8923 if (!__rt_schedulable(tg
, rt_period
, rt_runtime
))
8925 mutex_unlock(&rt_constraints_mutex
);
8929 #else /* !CONFIG_RT_GROUP_SCHED */
8930 static int sched_rt_global_constraints(void)
8932 unsigned long flags
;
8935 if (sysctl_sched_rt_period
<= 0)
8938 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8939 for_each_possible_cpu(i
) {
8940 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8942 spin_lock(&rt_rq
->rt_runtime_lock
);
8943 rt_rq
->rt_runtime
= global_rt_runtime();
8944 spin_unlock(&rt_rq
->rt_runtime_lock
);
8946 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8950 #endif /* CONFIG_RT_GROUP_SCHED */
8952 int sched_rt_handler(struct ctl_table
*table
, int write
,
8953 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
8957 int old_period
, old_runtime
;
8958 static DEFINE_MUTEX(mutex
);
8961 old_period
= sysctl_sched_rt_period
;
8962 old_runtime
= sysctl_sched_rt_runtime
;
8964 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
8966 if (!ret
&& write
) {
8967 ret
= sched_rt_global_constraints();
8969 sysctl_sched_rt_period
= old_period
;
8970 sysctl_sched_rt_runtime
= old_runtime
;
8972 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8973 def_rt_bandwidth
.rt_period
=
8974 ns_to_ktime(global_rt_period());
8977 mutex_unlock(&mutex
);
8982 #ifdef CONFIG_CGROUP_SCHED
8984 /* return corresponding task_group object of a cgroup */
8985 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
8987 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
8988 struct task_group
, css
);
8991 static struct cgroup_subsys_state
*
8992 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8994 struct task_group
*tg
, *parent
;
8996 if (!cgrp
->parent
) {
8997 /* This is early initialization for the top cgroup */
8998 init_task_group
.css
.cgroup
= cgrp
;
8999 return &init_task_group
.css
;
9002 parent
= cgroup_tg(cgrp
->parent
);
9003 tg
= sched_create_group(parent
);
9005 return ERR_PTR(-ENOMEM
);
9007 /* Bind the cgroup to task_group object we just created */
9008 tg
->css
.cgroup
= cgrp
;
9014 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9016 struct task_group
*tg
= cgroup_tg(cgrp
);
9018 sched_destroy_group(tg
);
9022 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9023 struct task_struct
*tsk
)
9025 #ifdef CONFIG_RT_GROUP_SCHED
9026 /* Don't accept realtime tasks when there is no way for them to run */
9027 if (rt_task(tsk
) && cgroup_tg(cgrp
)->rt_bandwidth
.rt_runtime
== 0)
9030 /* We don't support RT-tasks being in separate groups */
9031 if (tsk
->sched_class
!= &fair_sched_class
)
9039 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9040 struct cgroup
*old_cont
, struct task_struct
*tsk
)
9042 sched_move_task(tsk
);
9045 #ifdef CONFIG_FAIR_GROUP_SCHED
9046 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9049 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
9052 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9054 struct task_group
*tg
= cgroup_tg(cgrp
);
9056 return (u64
) tg
->shares
;
9058 #endif /* CONFIG_FAIR_GROUP_SCHED */
9060 #ifdef CONFIG_RT_GROUP_SCHED
9061 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9064 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9067 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9069 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9072 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9075 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9078 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9080 return sched_group_rt_period(cgroup_tg(cgrp
));
9082 #endif /* CONFIG_RT_GROUP_SCHED */
9084 static struct cftype cpu_files
[] = {
9085 #ifdef CONFIG_FAIR_GROUP_SCHED
9088 .read_u64
= cpu_shares_read_u64
,
9089 .write_u64
= cpu_shares_write_u64
,
9092 #ifdef CONFIG_RT_GROUP_SCHED
9094 .name
= "rt_runtime_us",
9095 .read_s64
= cpu_rt_runtime_read
,
9096 .write_s64
= cpu_rt_runtime_write
,
9099 .name
= "rt_period_us",
9100 .read_u64
= cpu_rt_period_read_uint
,
9101 .write_u64
= cpu_rt_period_write_uint
,
9106 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
9108 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
9111 struct cgroup_subsys cpu_cgroup_subsys
= {
9113 .create
= cpu_cgroup_create
,
9114 .destroy
= cpu_cgroup_destroy
,
9115 .can_attach
= cpu_cgroup_can_attach
,
9116 .attach
= cpu_cgroup_attach
,
9117 .populate
= cpu_cgroup_populate
,
9118 .subsys_id
= cpu_cgroup_subsys_id
,
9122 #endif /* CONFIG_CGROUP_SCHED */
9124 #ifdef CONFIG_CGROUP_CPUACCT
9127 * CPU accounting code for task groups.
9129 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9130 * (balbir@in.ibm.com).
9133 /* track cpu usage of a group of tasks */
9135 struct cgroup_subsys_state css
;
9136 /* cpuusage holds pointer to a u64-type object on every cpu */
9140 struct cgroup_subsys cpuacct_subsys
;
9142 /* return cpu accounting group corresponding to this container */
9143 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
9145 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
9146 struct cpuacct
, css
);
9149 /* return cpu accounting group to which this task belongs */
9150 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
9152 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
9153 struct cpuacct
, css
);
9156 /* create a new cpu accounting group */
9157 static struct cgroup_subsys_state
*cpuacct_create(
9158 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9160 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
9163 return ERR_PTR(-ENOMEM
);
9165 ca
->cpuusage
= alloc_percpu(u64
);
9166 if (!ca
->cpuusage
) {
9168 return ERR_PTR(-ENOMEM
);
9174 /* destroy an existing cpu accounting group */
9176 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9178 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9180 free_percpu(ca
->cpuusage
);
9184 /* return total cpu usage (in nanoseconds) of a group */
9185 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9187 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9188 u64 totalcpuusage
= 0;
9191 for_each_possible_cpu(i
) {
9192 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9195 * Take rq->lock to make 64-bit addition safe on 32-bit
9198 spin_lock_irq(&cpu_rq(i
)->lock
);
9199 totalcpuusage
+= *cpuusage
;
9200 spin_unlock_irq(&cpu_rq(i
)->lock
);
9203 return totalcpuusage
;
9206 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9209 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9218 for_each_possible_cpu(i
) {
9219 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9221 spin_lock_irq(&cpu_rq(i
)->lock
);
9223 spin_unlock_irq(&cpu_rq(i
)->lock
);
9229 static struct cftype files
[] = {
9232 .read_u64
= cpuusage_read
,
9233 .write_u64
= cpuusage_write
,
9237 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9239 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9243 * charge this task's execution time to its accounting group.
9245 * called with rq->lock held.
9247 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9251 if (!cpuacct_subsys
.active
)
9256 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
9258 *cpuusage
+= cputime
;
9262 struct cgroup_subsys cpuacct_subsys
= {
9264 .create
= cpuacct_create
,
9265 .destroy
= cpuacct_destroy
,
9266 .populate
= cpuacct_populate
,
9267 .subsys_id
= cpuacct_subsys_id
,
9269 #endif /* CONFIG_CGROUP_CPUACCT */