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>
75 #include <asm/irq_regs.h>
77 #include "sched_cpupri.h"
80 * Convert user-nice values [ -20 ... 0 ... 19 ]
81 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
84 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
85 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
86 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
89 * 'User priority' is the nice value converted to something we
90 * can work with better when scaling various scheduler parameters,
91 * it's a [ 0 ... 39 ] range.
93 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
94 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
95 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
98 * Helpers for converting nanosecond timing to jiffy resolution
100 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
102 #define NICE_0_LOAD SCHED_LOAD_SCALE
103 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
106 * These are the 'tuning knobs' of the scheduler:
108 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
109 * Timeslices get refilled after they expire.
111 #define DEF_TIMESLICE (100 * HZ / 1000)
114 * single value that denotes runtime == period, ie unlimited time.
116 #define RUNTIME_INF ((u64)~0ULL)
120 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
121 * Since cpu_power is a 'constant', we can use a reciprocal divide.
123 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
125 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
129 * Each time a sched group cpu_power is changed,
130 * we must compute its reciprocal value
132 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
134 sg
->__cpu_power
+= val
;
135 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
139 static inline int rt_policy(int policy
)
141 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
146 static inline int task_has_rt_policy(struct task_struct
*p
)
148 return rt_policy(p
->policy
);
152 * This is the priority-queue data structure of the RT scheduling class:
154 struct rt_prio_array
{
155 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
156 struct list_head xqueue
[MAX_RT_PRIO
]; /* exclusive queue */
157 struct list_head squeue
[MAX_RT_PRIO
]; /* shared queue */
160 struct rt_bandwidth
{
161 /* nests inside the rq lock: */
162 spinlock_t rt_runtime_lock
;
165 struct hrtimer rt_period_timer
;
168 static struct rt_bandwidth def_rt_bandwidth
;
170 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
172 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
174 struct rt_bandwidth
*rt_b
=
175 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
181 now
= hrtimer_cb_get_time(timer
);
182 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
187 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
190 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
194 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
196 rt_b
->rt_period
= ns_to_ktime(period
);
197 rt_b
->rt_runtime
= runtime
;
199 spin_lock_init(&rt_b
->rt_runtime_lock
);
201 hrtimer_init(&rt_b
->rt_period_timer
,
202 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
203 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
204 rt_b
->rt_period_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_NO_SOFTIRQ
;
207 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
211 if (rt_b
->rt_runtime
== RUNTIME_INF
)
214 if (hrtimer_active(&rt_b
->rt_period_timer
))
217 spin_lock(&rt_b
->rt_runtime_lock
);
219 if (hrtimer_active(&rt_b
->rt_period_timer
))
222 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
223 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
224 hrtimer_start(&rt_b
->rt_period_timer
,
225 rt_b
->rt_period_timer
.expires
,
228 spin_unlock(&rt_b
->rt_runtime_lock
);
231 #ifdef CONFIG_RT_GROUP_SCHED
232 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
234 hrtimer_cancel(&rt_b
->rt_period_timer
);
239 * sched_domains_mutex serializes calls to arch_init_sched_domains,
240 * detach_destroy_domains and partition_sched_domains.
242 static DEFINE_MUTEX(sched_domains_mutex
);
244 #ifdef CONFIG_GROUP_SCHED
246 #include <linux/cgroup.h>
250 static LIST_HEAD(task_groups
);
252 /* task group related information */
254 #ifdef CONFIG_CGROUP_SCHED
255 struct cgroup_subsys_state css
;
258 #ifdef CONFIG_FAIR_GROUP_SCHED
259 /* schedulable entities of this group on each cpu */
260 struct sched_entity
**se
;
261 /* runqueue "owned" by this group on each cpu */
262 struct cfs_rq
**cfs_rq
;
263 unsigned long shares
;
266 #ifdef CONFIG_RT_GROUP_SCHED
267 struct sched_rt_entity
**rt_se
;
268 struct rt_rq
**rt_rq
;
270 struct rt_bandwidth rt_bandwidth
;
274 struct list_head list
;
276 struct task_group
*parent
;
277 struct list_head siblings
;
278 struct list_head children
;
281 #ifdef CONFIG_USER_SCHED
285 * Every UID task group (including init_task_group aka UID-0) will
286 * be a child to this group.
288 struct task_group root_task_group
;
290 #ifdef CONFIG_FAIR_GROUP_SCHED
291 /* Default task group's sched entity on each cpu */
292 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
293 /* Default task group's cfs_rq on each cpu */
294 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
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
;
302 #define root_task_group init_task_group
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)
314 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
318 * A weight of 0, 1 or ULONG_MAX can cause arithmetics problems.
319 * (The default weight is 1024 - so there's no practical
320 * limitation from this.)
323 #define MAX_SHARES (ULONG_MAX - 1)
325 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
328 /* Default task group.
329 * Every task in system belong to this group at bootup.
331 struct task_group init_task_group
;
333 /* return group to which a task belongs */
334 static inline struct task_group
*task_group(struct task_struct
*p
)
336 struct task_group
*tg
;
338 #ifdef CONFIG_USER_SCHED
340 #elif defined(CONFIG_CGROUP_SCHED)
341 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
342 struct task_group
, css
);
344 tg
= &init_task_group
;
349 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
350 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
352 #ifdef CONFIG_FAIR_GROUP_SCHED
353 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
354 p
->se
.parent
= task_group(p
)->se
[cpu
];
357 #ifdef CONFIG_RT_GROUP_SCHED
358 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
359 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
365 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
367 #endif /* CONFIG_GROUP_SCHED */
369 /* CFS-related fields in a runqueue */
371 struct load_weight load
;
372 unsigned long nr_running
;
377 struct rb_root tasks_timeline
;
378 struct rb_node
*rb_leftmost
;
380 struct list_head tasks
;
381 struct list_head
*balance_iterator
;
384 * 'curr' points to currently running entity on this cfs_rq.
385 * It is set to NULL otherwise (i.e when none are currently running).
387 struct sched_entity
*curr
, *next
;
389 unsigned long nr_spread_over
;
391 #ifdef CONFIG_FAIR_GROUP_SCHED
392 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
395 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
396 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
397 * (like users, containers etc.)
399 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
400 * list is used during load balance.
402 struct list_head leaf_cfs_rq_list
;
403 struct task_group
*tg
; /* group that "owns" this runqueue */
407 /* Real-Time classes' related field in a runqueue: */
409 struct rt_prio_array active
;
410 unsigned long rt_nr_running
;
411 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
412 int highest_prio
; /* highest queued rt task prio */
415 unsigned long rt_nr_migratory
;
421 /* Nests inside the rq lock: */
422 spinlock_t rt_runtime_lock
;
424 #ifdef CONFIG_RT_GROUP_SCHED
425 unsigned long rt_nr_boosted
;
428 struct list_head leaf_rt_rq_list
;
429 struct task_group
*tg
;
430 struct sched_rt_entity
*rt_se
;
437 * We add the notion of a root-domain which will be used to define per-domain
438 * variables. Each exclusive cpuset essentially defines an island domain by
439 * fully partitioning the member cpus from any other cpuset. Whenever a new
440 * exclusive cpuset is created, we also create and attach a new root-domain
450 * The "RT overload" flag: it gets set if a CPU has more than
451 * one runnable RT task.
456 struct cpupri cpupri
;
461 * By default the system creates a single root-domain with all cpus as
462 * members (mimicking the global state we have today).
464 static struct root_domain def_root_domain
;
469 * This is the main, per-CPU runqueue data structure.
471 * Locking rule: those places that want to lock multiple runqueues
472 * (such as the load balancing or the thread migration code), lock
473 * acquire operations must be ordered by ascending &runqueue.
480 * nr_running and cpu_load should be in the same cacheline because
481 * remote CPUs use both these fields when doing load calculation.
483 unsigned long nr_running
;
484 #define CPU_LOAD_IDX_MAX 5
485 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
486 unsigned char idle_at_tick
;
488 unsigned long last_tick_seen
;
489 unsigned char in_nohz_recently
;
491 /* capture load from *all* tasks on this cpu: */
492 struct load_weight load
;
493 unsigned long nr_load_updates
;
499 #ifdef CONFIG_FAIR_GROUP_SCHED
500 /* list of leaf cfs_rq on this cpu: */
501 struct list_head leaf_cfs_rq_list
;
503 #ifdef CONFIG_RT_GROUP_SCHED
504 struct list_head leaf_rt_rq_list
;
508 * This is part of a global counter where only the total sum
509 * over all CPUs matters. A task can increase this counter on
510 * one CPU and if it got migrated afterwards it may decrease
511 * it on another CPU. Always updated under the runqueue lock:
513 unsigned long nr_uninterruptible
;
515 struct task_struct
*curr
, *idle
;
516 unsigned long next_balance
;
517 struct mm_struct
*prev_mm
;
524 struct root_domain
*rd
;
525 struct sched_domain
*sd
;
527 /* For active balancing */
530 /* cpu of this runqueue: */
533 struct task_struct
*migration_thread
;
534 struct list_head migration_queue
;
537 #ifdef CONFIG_SCHED_HRTICK
538 unsigned long hrtick_flags
;
539 ktime_t hrtick_expire
;
540 struct hrtimer hrtick_timer
;
543 #ifdef CONFIG_SCHEDSTATS
545 struct sched_info rq_sched_info
;
547 /* sys_sched_yield() stats */
548 unsigned int yld_exp_empty
;
549 unsigned int yld_act_empty
;
550 unsigned int yld_both_empty
;
551 unsigned int yld_count
;
553 /* schedule() stats */
554 unsigned int sched_switch
;
555 unsigned int sched_count
;
556 unsigned int sched_goidle
;
558 /* try_to_wake_up() stats */
559 unsigned int ttwu_count
;
560 unsigned int ttwu_local
;
563 unsigned int bkl_count
;
565 struct lock_class_key rq_lock_key
;
568 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
570 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
572 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
575 static inline int cpu_of(struct rq
*rq
)
585 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
586 * See detach_destroy_domains: synchronize_sched for details.
588 * The domain tree of any CPU may only be accessed from within
589 * preempt-disabled sections.
591 #define for_each_domain(cpu, __sd) \
592 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
594 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
595 #define this_rq() (&__get_cpu_var(runqueues))
596 #define task_rq(p) cpu_rq(task_cpu(p))
597 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
599 static inline void update_rq_clock(struct rq
*rq
)
601 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
605 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
607 #ifdef CONFIG_SCHED_DEBUG
608 # define const_debug __read_mostly
610 # define const_debug static const
614 * Debugging: various feature bits
617 #define SCHED_FEAT(name, enabled) \
618 __SCHED_FEAT_##name ,
621 #include "sched_features.h"
626 #define SCHED_FEAT(name, enabled) \
627 (1UL << __SCHED_FEAT_##name) * enabled |
629 const_debug
unsigned int sysctl_sched_features
=
630 #include "sched_features.h"
635 #ifdef CONFIG_SCHED_DEBUG
636 #define SCHED_FEAT(name, enabled) \
639 static __read_mostly
char *sched_feat_names
[] = {
640 #include "sched_features.h"
646 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
648 filp
->private_data
= inode
->i_private
;
653 sched_feat_read(struct file
*filp
, char __user
*ubuf
,
654 size_t cnt
, loff_t
*ppos
)
661 for (i
= 0; sched_feat_names
[i
]; i
++) {
662 len
+= strlen(sched_feat_names
[i
]);
666 buf
= kmalloc(len
+ 2, GFP_KERNEL
);
670 for (i
= 0; sched_feat_names
[i
]; i
++) {
671 if (sysctl_sched_features
& (1UL << i
))
672 r
+= sprintf(buf
+ r
, "%s ", sched_feat_names
[i
]);
674 r
+= sprintf(buf
+ r
, "NO_%s ", sched_feat_names
[i
]);
677 r
+= sprintf(buf
+ r
, "\n");
678 WARN_ON(r
>= len
+ 2);
680 r
= simple_read_from_buffer(ubuf
, cnt
, ppos
, buf
, r
);
688 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
689 size_t cnt
, loff_t
*ppos
)
699 if (copy_from_user(&buf
, ubuf
, cnt
))
704 if (strncmp(buf
, "NO_", 3) == 0) {
709 for (i
= 0; sched_feat_names
[i
]; i
++) {
710 int len
= strlen(sched_feat_names
[i
]);
712 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
714 sysctl_sched_features
&= ~(1UL << i
);
716 sysctl_sched_features
|= (1UL << i
);
721 if (!sched_feat_names
[i
])
729 static struct file_operations sched_feat_fops
= {
730 .open
= sched_feat_open
,
731 .read
= sched_feat_read
,
732 .write
= sched_feat_write
,
735 static __init
int sched_init_debug(void)
737 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
742 late_initcall(sched_init_debug
);
746 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
749 * Number of tasks to iterate in a single balance run.
750 * Limited because this is done with IRQs disabled.
752 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
755 * period over which we measure -rt task cpu usage in us.
758 unsigned int sysctl_sched_rt_period
= 1000000;
760 static __read_mostly
int scheduler_running
;
763 * part of the period that we allow rt tasks to run in us.
766 int sysctl_sched_rt_runtime
= 950000;
768 static inline u64
global_rt_period(void)
770 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
773 static inline u64
global_rt_runtime(void)
775 if (sysctl_sched_rt_period
< 0)
778 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
781 unsigned long long time_sync_thresh
= 100000;
783 static DEFINE_PER_CPU(unsigned long long, time_offset
);
784 static DEFINE_PER_CPU(unsigned long long, prev_cpu_time
);
787 * Global lock which we take every now and then to synchronize
788 * the CPUs time. This method is not warp-safe, but it's good
789 * enough to synchronize slowly diverging time sources and thus
790 * it's good enough for tracing:
792 static DEFINE_SPINLOCK(time_sync_lock
);
793 static unsigned long long prev_global_time
;
795 static unsigned long long __sync_cpu_clock(unsigned long long time
, int cpu
)
798 * We want this inlined, to not get tracer function calls
799 * in this critical section:
801 spin_acquire(&time_sync_lock
.dep_map
, 0, 0, _THIS_IP_
);
802 __raw_spin_lock(&time_sync_lock
.raw_lock
);
804 if (time
< prev_global_time
) {
805 per_cpu(time_offset
, cpu
) += prev_global_time
- time
;
806 time
= prev_global_time
;
808 prev_global_time
= time
;
811 __raw_spin_unlock(&time_sync_lock
.raw_lock
);
812 spin_release(&time_sync_lock
.dep_map
, 1, _THIS_IP_
);
817 static unsigned long long __cpu_clock(int cpu
)
819 unsigned long long now
;
822 * Only call sched_clock() if the scheduler has already been
823 * initialized (some code might call cpu_clock() very early):
825 if (unlikely(!scheduler_running
))
828 now
= sched_clock_cpu(cpu
);
834 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
835 * clock constructed from sched_clock():
837 unsigned long long cpu_clock(int cpu
)
839 unsigned long long prev_cpu_time
, time
, delta_time
;
842 local_irq_save(flags
);
843 prev_cpu_time
= per_cpu(prev_cpu_time
, cpu
);
844 time
= __cpu_clock(cpu
) + per_cpu(time_offset
, cpu
);
845 delta_time
= time
-prev_cpu_time
;
847 if (unlikely(delta_time
> time_sync_thresh
)) {
848 time
= __sync_cpu_clock(time
, cpu
);
849 per_cpu(prev_cpu_time
, cpu
) = time
;
851 local_irq_restore(flags
);
855 EXPORT_SYMBOL_GPL(cpu_clock
);
857 #ifndef prepare_arch_switch
858 # define prepare_arch_switch(next) do { } while (0)
860 #ifndef finish_arch_switch
861 # define finish_arch_switch(prev) do { } while (0)
864 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
866 return rq
->curr
== p
;
869 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
870 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
872 return task_current(rq
, p
);
875 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
879 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
881 #ifdef CONFIG_DEBUG_SPINLOCK
882 /* this is a valid case when another task releases the spinlock */
883 rq
->lock
.owner
= current
;
886 * If we are tracking spinlock dependencies then we have to
887 * fix up the runqueue lock - which gets 'carried over' from
890 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
892 spin_unlock_irq(&rq
->lock
);
895 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
896 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
901 return task_current(rq
, p
);
905 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
909 * We can optimise this out completely for !SMP, because the
910 * SMP rebalancing from interrupt is the only thing that cares
915 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
916 spin_unlock_irq(&rq
->lock
);
918 spin_unlock(&rq
->lock
);
922 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
926 * After ->oncpu is cleared, the task can be moved to a different CPU.
927 * We must ensure this doesn't happen until the switch is completely
933 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
937 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
940 * __task_rq_lock - lock the runqueue a given task resides on.
941 * Must be called interrupts disabled.
943 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
947 struct rq
*rq
= task_rq(p
);
948 spin_lock(&rq
->lock
);
949 if (likely(rq
== task_rq(p
)))
951 spin_unlock(&rq
->lock
);
956 * task_rq_lock - lock the runqueue a given task resides on and disable
957 * interrupts. Note the ordering: we can safely lookup the task_rq without
958 * explicitly disabling preemption.
960 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
966 local_irq_save(*flags
);
968 spin_lock(&rq
->lock
);
969 if (likely(rq
== task_rq(p
)))
971 spin_unlock_irqrestore(&rq
->lock
, *flags
);
975 static void __task_rq_unlock(struct rq
*rq
)
978 spin_unlock(&rq
->lock
);
981 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
984 spin_unlock_irqrestore(&rq
->lock
, *flags
);
988 * this_rq_lock - lock this runqueue and disable interrupts.
990 static struct rq
*this_rq_lock(void)
997 spin_lock(&rq
->lock
);
1002 static void __resched_task(struct task_struct
*p
, int tif_bit
);
1004 static inline void resched_task(struct task_struct
*p
)
1006 __resched_task(p
, TIF_NEED_RESCHED
);
1009 #ifdef CONFIG_SCHED_HRTICK
1011 * Use HR-timers to deliver accurate preemption points.
1013 * Its all a bit involved since we cannot program an hrt while holding the
1014 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1017 * When we get rescheduled we reprogram the hrtick_timer outside of the
1020 static inline void resched_hrt(struct task_struct
*p
)
1022 __resched_task(p
, TIF_HRTICK_RESCHED
);
1025 static inline void resched_rq(struct rq
*rq
)
1027 unsigned long flags
;
1029 spin_lock_irqsave(&rq
->lock
, flags
);
1030 resched_task(rq
->curr
);
1031 spin_unlock_irqrestore(&rq
->lock
, flags
);
1035 HRTICK_SET
, /* re-programm hrtick_timer */
1036 HRTICK_RESET
, /* not a new slice */
1037 HRTICK_BLOCK
, /* stop hrtick operations */
1042 * - enabled by features
1043 * - hrtimer is actually high res
1045 static inline int hrtick_enabled(struct rq
*rq
)
1047 if (!sched_feat(HRTICK
))
1049 if (unlikely(test_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
)))
1051 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1055 * Called to set the hrtick timer state.
1057 * called with rq->lock held and irqs disabled
1059 static void hrtick_start(struct rq
*rq
, u64 delay
, int reset
)
1061 assert_spin_locked(&rq
->lock
);
1064 * preempt at: now + delay
1067 ktime_add_ns(rq
->hrtick_timer
.base
->get_time(), delay
);
1069 * indicate we need to program the timer
1071 __set_bit(HRTICK_SET
, &rq
->hrtick_flags
);
1073 __set_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
1076 * New slices are called from the schedule path and don't need a
1077 * forced reschedule.
1080 resched_hrt(rq
->curr
);
1083 static void hrtick_clear(struct rq
*rq
)
1085 if (hrtimer_active(&rq
->hrtick_timer
))
1086 hrtimer_cancel(&rq
->hrtick_timer
);
1090 * Update the timer from the possible pending state.
1092 static void hrtick_set(struct rq
*rq
)
1096 unsigned long flags
;
1098 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1100 spin_lock_irqsave(&rq
->lock
, flags
);
1101 set
= __test_and_clear_bit(HRTICK_SET
, &rq
->hrtick_flags
);
1102 reset
= __test_and_clear_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
1103 time
= rq
->hrtick_expire
;
1104 clear_thread_flag(TIF_HRTICK_RESCHED
);
1105 spin_unlock_irqrestore(&rq
->lock
, flags
);
1108 hrtimer_start(&rq
->hrtick_timer
, time
, HRTIMER_MODE_ABS
);
1109 if (reset
&& !hrtimer_active(&rq
->hrtick_timer
))
1116 * High-resolution timer tick.
1117 * Runs from hardirq context with interrupts disabled.
1119 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1121 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1123 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1125 spin_lock(&rq
->lock
);
1126 update_rq_clock(rq
);
1127 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1128 spin_unlock(&rq
->lock
);
1130 return HRTIMER_NORESTART
;
1134 static void hotplug_hrtick_disable(int cpu
)
1136 struct rq
*rq
= cpu_rq(cpu
);
1137 unsigned long flags
;
1139 spin_lock_irqsave(&rq
->lock
, flags
);
1140 rq
->hrtick_flags
= 0;
1141 __set_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
);
1142 spin_unlock_irqrestore(&rq
->lock
, flags
);
1147 static void hotplug_hrtick_enable(int cpu
)
1149 struct rq
*rq
= cpu_rq(cpu
);
1150 unsigned long flags
;
1152 spin_lock_irqsave(&rq
->lock
, flags
);
1153 __clear_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
);
1154 spin_unlock_irqrestore(&rq
->lock
, flags
);
1158 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1160 int cpu
= (int)(long)hcpu
;
1163 case CPU_UP_CANCELED
:
1164 case CPU_UP_CANCELED_FROZEN
:
1165 case CPU_DOWN_PREPARE
:
1166 case CPU_DOWN_PREPARE_FROZEN
:
1168 case CPU_DEAD_FROZEN
:
1169 hotplug_hrtick_disable(cpu
);
1172 case CPU_UP_PREPARE
:
1173 case CPU_UP_PREPARE_FROZEN
:
1174 case CPU_DOWN_FAILED
:
1175 case CPU_DOWN_FAILED_FROZEN
:
1177 case CPU_ONLINE_FROZEN
:
1178 hotplug_hrtick_enable(cpu
);
1185 static void init_hrtick(void)
1187 hotcpu_notifier(hotplug_hrtick
, 0);
1189 #endif /* CONFIG_SMP */
1191 static void init_rq_hrtick(struct rq
*rq
)
1193 rq
->hrtick_flags
= 0;
1194 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1195 rq
->hrtick_timer
.function
= hrtick
;
1196 rq
->hrtick_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_NO_SOFTIRQ
;
1199 void hrtick_resched(void)
1202 unsigned long flags
;
1204 if (!test_thread_flag(TIF_HRTICK_RESCHED
))
1207 local_irq_save(flags
);
1208 rq
= cpu_rq(smp_processor_id());
1210 local_irq_restore(flags
);
1213 static inline void hrtick_clear(struct rq
*rq
)
1217 static inline void hrtick_set(struct rq
*rq
)
1221 static inline void init_rq_hrtick(struct rq
*rq
)
1225 void hrtick_resched(void)
1229 static inline void init_hrtick(void)
1235 * resched_task - mark a task 'to be rescheduled now'.
1237 * On UP this means the setting of the need_resched flag, on SMP it
1238 * might also involve a cross-CPU call to trigger the scheduler on
1243 #ifndef tsk_is_polling
1244 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1247 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1251 assert_spin_locked(&task_rq(p
)->lock
);
1253 if (unlikely(test_tsk_thread_flag(p
, tif_bit
)))
1256 set_tsk_thread_flag(p
, tif_bit
);
1259 if (cpu
== smp_processor_id())
1262 /* NEED_RESCHED must be visible before we test polling */
1264 if (!tsk_is_polling(p
))
1265 smp_send_reschedule(cpu
);
1268 static void resched_cpu(int cpu
)
1270 struct rq
*rq
= cpu_rq(cpu
);
1271 unsigned long flags
;
1273 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1275 resched_task(cpu_curr(cpu
));
1276 spin_unlock_irqrestore(&rq
->lock
, flags
);
1281 * When add_timer_on() enqueues a timer into the timer wheel of an
1282 * idle CPU then this timer might expire before the next timer event
1283 * which is scheduled to wake up that CPU. In case of a completely
1284 * idle system the next event might even be infinite time into the
1285 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1286 * leaves the inner idle loop so the newly added timer is taken into
1287 * account when the CPU goes back to idle and evaluates the timer
1288 * wheel for the next timer event.
1290 void wake_up_idle_cpu(int cpu
)
1292 struct rq
*rq
= cpu_rq(cpu
);
1294 if (cpu
== smp_processor_id())
1298 * This is safe, as this function is called with the timer
1299 * wheel base lock of (cpu) held. When the CPU is on the way
1300 * to idle and has not yet set rq->curr to idle then it will
1301 * be serialized on the timer wheel base lock and take the new
1302 * timer into account automatically.
1304 if (rq
->curr
!= rq
->idle
)
1308 * We can set TIF_RESCHED on the idle task of the other CPU
1309 * lockless. The worst case is that the other CPU runs the
1310 * idle task through an additional NOOP schedule()
1312 set_tsk_thread_flag(rq
->idle
, TIF_NEED_RESCHED
);
1314 /* NEED_RESCHED must be visible before we test polling */
1316 if (!tsk_is_polling(rq
->idle
))
1317 smp_send_reschedule(cpu
);
1322 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1324 assert_spin_locked(&task_rq(p
)->lock
);
1325 set_tsk_thread_flag(p
, tif_bit
);
1329 #if BITS_PER_LONG == 32
1330 # define WMULT_CONST (~0UL)
1332 # define WMULT_CONST (1UL << 32)
1335 #define WMULT_SHIFT 32
1338 * Shift right and round:
1340 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1342 static unsigned long
1343 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1344 struct load_weight
*lw
)
1348 if (!lw
->inv_weight
)
1349 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)/(lw
->weight
+1);
1351 tmp
= (u64
)delta_exec
* weight
;
1353 * Check whether we'd overflow the 64-bit multiplication:
1355 if (unlikely(tmp
> WMULT_CONST
))
1356 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1359 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1361 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1364 static inline unsigned long
1365 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
1367 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
1370 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1376 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1383 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1384 * of tasks with abnormal "nice" values across CPUs the contribution that
1385 * each task makes to its run queue's load is weighted according to its
1386 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1387 * scaled version of the new time slice allocation that they receive on time
1391 #define WEIGHT_IDLEPRIO 2
1392 #define WMULT_IDLEPRIO (1 << 31)
1395 * Nice levels are multiplicative, with a gentle 10% change for every
1396 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1397 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1398 * that remained on nice 0.
1400 * The "10% effect" is relative and cumulative: from _any_ nice level,
1401 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1402 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1403 * If a task goes up by ~10% and another task goes down by ~10% then
1404 * the relative distance between them is ~25%.)
1406 static const int prio_to_weight
[40] = {
1407 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1408 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1409 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1410 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1411 /* 0 */ 1024, 820, 655, 526, 423,
1412 /* 5 */ 335, 272, 215, 172, 137,
1413 /* 10 */ 110, 87, 70, 56, 45,
1414 /* 15 */ 36, 29, 23, 18, 15,
1418 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1420 * In cases where the weight does not change often, we can use the
1421 * precalculated inverse to speed up arithmetics by turning divisions
1422 * into multiplications:
1424 static const u32 prio_to_wmult
[40] = {
1425 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1426 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1427 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1428 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1429 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1430 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1431 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1432 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1435 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1438 * runqueue iterator, to support SMP load-balancing between different
1439 * scheduling classes, without having to expose their internal data
1440 * structures to the load-balancing proper:
1442 struct rq_iterator
{
1444 struct task_struct
*(*start
)(void *);
1445 struct task_struct
*(*next
)(void *);
1449 static unsigned long
1450 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1451 unsigned long max_load_move
, struct sched_domain
*sd
,
1452 enum cpu_idle_type idle
, int *all_pinned
,
1453 int *this_best_prio
, struct rq_iterator
*iterator
);
1456 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1457 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1458 struct rq_iterator
*iterator
);
1461 #ifdef CONFIG_CGROUP_CPUACCT
1462 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1464 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1467 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1469 update_load_add(&rq
->load
, load
);
1472 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1474 update_load_sub(&rq
->load
, load
);
1478 static unsigned long source_load(int cpu
, int type
);
1479 static unsigned long target_load(int cpu
, int type
);
1480 static unsigned long cpu_avg_load_per_task(int cpu
);
1481 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1482 #else /* CONFIG_SMP */
1484 #ifdef CONFIG_FAIR_GROUP_SCHED
1485 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1490 #endif /* CONFIG_SMP */
1492 #include "sched_stats.h"
1493 #include "sched_idletask.c"
1494 #include "sched_fair.c"
1495 #include "sched_rt.c"
1496 #ifdef CONFIG_SCHED_DEBUG
1497 # include "sched_debug.c"
1500 #define sched_class_highest (&rt_sched_class)
1502 static inline void inc_load(struct rq
*rq
, const struct task_struct
*p
)
1504 update_load_add(&rq
->load
, p
->se
.load
.weight
);
1507 static inline void dec_load(struct rq
*rq
, const struct task_struct
*p
)
1509 update_load_sub(&rq
->load
, p
->se
.load
.weight
);
1512 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
1518 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
1524 static void set_load_weight(struct task_struct
*p
)
1526 if (task_has_rt_policy(p
)) {
1527 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1528 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1533 * SCHED_IDLE tasks get minimal weight:
1535 if (p
->policy
== SCHED_IDLE
) {
1536 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1537 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1541 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1542 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1545 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1547 sched_info_queued(p
);
1548 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1552 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1554 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1559 * __normal_prio - return the priority that is based on the static prio
1561 static inline int __normal_prio(struct task_struct
*p
)
1563 return p
->static_prio
;
1567 * Calculate the expected normal priority: i.e. priority
1568 * without taking RT-inheritance into account. Might be
1569 * boosted by interactivity modifiers. Changes upon fork,
1570 * setprio syscalls, and whenever the interactivity
1571 * estimator recalculates.
1573 static inline int normal_prio(struct task_struct
*p
)
1577 if (task_has_rt_policy(p
))
1578 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1580 prio
= __normal_prio(p
);
1585 * Calculate the current priority, i.e. the priority
1586 * taken into account by the scheduler. This value might
1587 * be boosted by RT tasks, or might be boosted by
1588 * interactivity modifiers. Will be RT if the task got
1589 * RT-boosted. If not then it returns p->normal_prio.
1591 static int effective_prio(struct task_struct
*p
)
1593 p
->normal_prio
= normal_prio(p
);
1595 * If we are RT tasks or we were boosted to RT priority,
1596 * keep the priority unchanged. Otherwise, update priority
1597 * to the normal priority:
1599 if (!rt_prio(p
->prio
))
1600 return p
->normal_prio
;
1605 * activate_task - move a task to the runqueue.
1607 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1609 if (task_contributes_to_load(p
))
1610 rq
->nr_uninterruptible
--;
1612 enqueue_task(rq
, p
, wakeup
);
1613 inc_nr_running(p
, rq
);
1617 * deactivate_task - remove a task from the runqueue.
1619 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1621 if (task_contributes_to_load(p
))
1622 rq
->nr_uninterruptible
++;
1624 dequeue_task(rq
, p
, sleep
);
1625 dec_nr_running(p
, rq
);
1629 * task_curr - is this task currently executing on a CPU?
1630 * @p: the task in question.
1632 inline int task_curr(const struct task_struct
*p
)
1634 return cpu_curr(task_cpu(p
)) == p
;
1637 /* Used instead of source_load when we know the type == 0 */
1638 static unsigned long weighted_cpuload(const int cpu
)
1640 return cpu_rq(cpu
)->load
.weight
;
1643 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1645 set_task_rq(p
, cpu
);
1648 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1649 * successfuly executed on another CPU. We must ensure that updates of
1650 * per-task data have been completed by this moment.
1653 task_thread_info(p
)->cpu
= cpu
;
1657 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1658 const struct sched_class
*prev_class
,
1659 int oldprio
, int running
)
1661 if (prev_class
!= p
->sched_class
) {
1662 if (prev_class
->switched_from
)
1663 prev_class
->switched_from(rq
, p
, running
);
1664 p
->sched_class
->switched_to(rq
, p
, running
);
1666 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1672 * Is this task likely cache-hot:
1675 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1680 * Buddy candidates are cache hot:
1682 if (sched_feat(CACHE_HOT_BUDDY
) && (&p
->se
== cfs_rq_of(&p
->se
)->next
))
1685 if (p
->sched_class
!= &fair_sched_class
)
1688 if (sysctl_sched_migration_cost
== -1)
1690 if (sysctl_sched_migration_cost
== 0)
1693 delta
= now
- p
->se
.exec_start
;
1695 return delta
< (s64
)sysctl_sched_migration_cost
;
1699 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1701 int old_cpu
= task_cpu(p
);
1702 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1703 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1704 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1707 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1709 #ifdef CONFIG_SCHEDSTATS
1710 if (p
->se
.wait_start
)
1711 p
->se
.wait_start
-= clock_offset
;
1712 if (p
->se
.sleep_start
)
1713 p
->se
.sleep_start
-= clock_offset
;
1714 if (p
->se
.block_start
)
1715 p
->se
.block_start
-= clock_offset
;
1716 if (old_cpu
!= new_cpu
) {
1717 schedstat_inc(p
, se
.nr_migrations
);
1718 if (task_hot(p
, old_rq
->clock
, NULL
))
1719 schedstat_inc(p
, se
.nr_forced2_migrations
);
1722 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1723 new_cfsrq
->min_vruntime
;
1725 __set_task_cpu(p
, new_cpu
);
1728 struct migration_req
{
1729 struct list_head list
;
1731 struct task_struct
*task
;
1734 struct completion done
;
1738 * The task's runqueue lock must be held.
1739 * Returns true if you have to wait for migration thread.
1742 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1744 struct rq
*rq
= task_rq(p
);
1747 * If the task is not on a runqueue (and not running), then
1748 * it is sufficient to simply update the task's cpu field.
1750 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1751 set_task_cpu(p
, dest_cpu
);
1755 init_completion(&req
->done
);
1757 req
->dest_cpu
= dest_cpu
;
1758 list_add(&req
->list
, &rq
->migration_queue
);
1764 * wait_task_inactive - wait for a thread to unschedule.
1766 * The caller must ensure that the task *will* unschedule sometime soon,
1767 * else this function might spin for a *long* time. This function can't
1768 * be called with interrupts off, or it may introduce deadlock with
1769 * smp_call_function() if an IPI is sent by the same process we are
1770 * waiting to become inactive.
1772 void wait_task_inactive(struct task_struct
*p
)
1774 unsigned long flags
;
1780 * We do the initial early heuristics without holding
1781 * any task-queue locks at all. We'll only try to get
1782 * the runqueue lock when things look like they will
1788 * If the task is actively running on another CPU
1789 * still, just relax and busy-wait without holding
1792 * NOTE! Since we don't hold any locks, it's not
1793 * even sure that "rq" stays as the right runqueue!
1794 * But we don't care, since "task_running()" will
1795 * return false if the runqueue has changed and p
1796 * is actually now running somewhere else!
1798 while (task_running(rq
, p
))
1802 * Ok, time to look more closely! We need the rq
1803 * lock now, to be *sure*. If we're wrong, we'll
1804 * just go back and repeat.
1806 rq
= task_rq_lock(p
, &flags
);
1807 running
= task_running(rq
, p
);
1808 on_rq
= p
->se
.on_rq
;
1809 task_rq_unlock(rq
, &flags
);
1812 * Was it really running after all now that we
1813 * checked with the proper locks actually held?
1815 * Oops. Go back and try again..
1817 if (unlikely(running
)) {
1823 * It's not enough that it's not actively running,
1824 * it must be off the runqueue _entirely_, and not
1827 * So if it wa still runnable (but just not actively
1828 * running right now), it's preempted, and we should
1829 * yield - it could be a while.
1831 if (unlikely(on_rq
)) {
1832 schedule_timeout_uninterruptible(1);
1837 * Ahh, all good. It wasn't running, and it wasn't
1838 * runnable, which means that it will never become
1839 * running in the future either. We're all done!
1846 * kick_process - kick a running thread to enter/exit the kernel
1847 * @p: the to-be-kicked thread
1849 * Cause a process which is running on another CPU to enter
1850 * kernel-mode, without any delay. (to get signals handled.)
1852 * NOTE: this function doesnt have to take the runqueue lock,
1853 * because all it wants to ensure is that the remote task enters
1854 * the kernel. If the IPI races and the task has been migrated
1855 * to another CPU then no harm is done and the purpose has been
1858 void kick_process(struct task_struct
*p
)
1864 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1865 smp_send_reschedule(cpu
);
1870 * Return a low guess at the load of a migration-source cpu weighted
1871 * according to the scheduling class and "nice" value.
1873 * We want to under-estimate the load of migration sources, to
1874 * balance conservatively.
1876 static unsigned long source_load(int cpu
, int type
)
1878 struct rq
*rq
= cpu_rq(cpu
);
1879 unsigned long total
= weighted_cpuload(cpu
);
1884 return min(rq
->cpu_load
[type
-1], total
);
1888 * Return a high guess at the load of a migration-target cpu weighted
1889 * according to the scheduling class and "nice" value.
1891 static unsigned long target_load(int cpu
, int type
)
1893 struct rq
*rq
= cpu_rq(cpu
);
1894 unsigned long total
= weighted_cpuload(cpu
);
1899 return max(rq
->cpu_load
[type
-1], total
);
1903 * Return the average load per task on the cpu's run queue
1905 static unsigned long cpu_avg_load_per_task(int cpu
)
1907 struct rq
*rq
= cpu_rq(cpu
);
1908 unsigned long total
= weighted_cpuload(cpu
);
1909 unsigned long n
= rq
->nr_running
;
1911 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1915 * find_idlest_group finds and returns the least busy CPU group within the
1918 static struct sched_group
*
1919 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1921 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1922 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1923 int load_idx
= sd
->forkexec_idx
;
1924 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1927 unsigned long load
, avg_load
;
1931 /* Skip over this group if it has no CPUs allowed */
1932 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1935 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1937 /* Tally up the load of all CPUs in the group */
1940 for_each_cpu_mask(i
, group
->cpumask
) {
1941 /* Bias balancing toward cpus of our domain */
1943 load
= source_load(i
, load_idx
);
1945 load
= target_load(i
, load_idx
);
1950 /* Adjust by relative CPU power of the group */
1951 avg_load
= sg_div_cpu_power(group
,
1952 avg_load
* SCHED_LOAD_SCALE
);
1955 this_load
= avg_load
;
1957 } else if (avg_load
< min_load
) {
1958 min_load
= avg_load
;
1961 } while (group
= group
->next
, group
!= sd
->groups
);
1963 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1969 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1972 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
,
1975 unsigned long load
, min_load
= ULONG_MAX
;
1979 /* Traverse only the allowed CPUs */
1980 cpus_and(*tmp
, group
->cpumask
, p
->cpus_allowed
);
1982 for_each_cpu_mask(i
, *tmp
) {
1983 load
= weighted_cpuload(i
);
1985 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1995 * sched_balance_self: balance the current task (running on cpu) in domains
1996 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1999 * Balance, ie. select the least loaded group.
2001 * Returns the target CPU number, or the same CPU if no balancing is needed.
2003 * preempt must be disabled.
2005 static int sched_balance_self(int cpu
, int flag
)
2007 struct task_struct
*t
= current
;
2008 struct sched_domain
*tmp
, *sd
= NULL
;
2010 for_each_domain(cpu
, tmp
) {
2012 * If power savings logic is enabled for a domain, stop there.
2014 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2016 if (tmp
->flags
& flag
)
2021 cpumask_t span
, tmpmask
;
2022 struct sched_group
*group
;
2023 int new_cpu
, weight
;
2025 if (!(sd
->flags
& flag
)) {
2031 group
= find_idlest_group(sd
, t
, cpu
);
2037 new_cpu
= find_idlest_cpu(group
, t
, cpu
, &tmpmask
);
2038 if (new_cpu
== -1 || new_cpu
== cpu
) {
2039 /* Now try balancing at a lower domain level of cpu */
2044 /* Now try balancing at a lower domain level of new_cpu */
2047 weight
= cpus_weight(span
);
2048 for_each_domain(cpu
, tmp
) {
2049 if (weight
<= cpus_weight(tmp
->span
))
2051 if (tmp
->flags
& flag
)
2054 /* while loop will break here if sd == NULL */
2060 #endif /* CONFIG_SMP */
2063 * try_to_wake_up - wake up a thread
2064 * @p: the to-be-woken-up thread
2065 * @state: the mask of task states that can be woken
2066 * @sync: do a synchronous wakeup?
2068 * Put it on the run-queue if it's not already there. The "current"
2069 * thread is always on the run-queue (except when the actual
2070 * re-schedule is in progress), and as such you're allowed to do
2071 * the simpler "current->state = TASK_RUNNING" to mark yourself
2072 * runnable without the overhead of this.
2074 * returns failure only if the task is already active.
2076 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2078 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2079 unsigned long flags
;
2083 if (!sched_feat(SYNC_WAKEUPS
))
2087 rq
= task_rq_lock(p
, &flags
);
2088 old_state
= p
->state
;
2089 if (!(old_state
& state
))
2097 this_cpu
= smp_processor_id();
2100 if (unlikely(task_running(rq
, p
)))
2103 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2104 if (cpu
!= orig_cpu
) {
2105 set_task_cpu(p
, cpu
);
2106 task_rq_unlock(rq
, &flags
);
2107 /* might preempt at this point */
2108 rq
= task_rq_lock(p
, &flags
);
2109 old_state
= p
->state
;
2110 if (!(old_state
& state
))
2115 this_cpu
= smp_processor_id();
2119 #ifdef CONFIG_SCHEDSTATS
2120 schedstat_inc(rq
, ttwu_count
);
2121 if (cpu
== this_cpu
)
2122 schedstat_inc(rq
, ttwu_local
);
2124 struct sched_domain
*sd
;
2125 for_each_domain(this_cpu
, sd
) {
2126 if (cpu_isset(cpu
, sd
->span
)) {
2127 schedstat_inc(sd
, ttwu_wake_remote
);
2135 #endif /* CONFIG_SMP */
2136 schedstat_inc(p
, se
.nr_wakeups
);
2138 schedstat_inc(p
, se
.nr_wakeups_sync
);
2139 if (orig_cpu
!= cpu
)
2140 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2141 if (cpu
== this_cpu
)
2142 schedstat_inc(p
, se
.nr_wakeups_local
);
2144 schedstat_inc(p
, se
.nr_wakeups_remote
);
2145 update_rq_clock(rq
);
2146 activate_task(rq
, p
, 1);
2150 check_preempt_curr(rq
, p
);
2152 p
->state
= TASK_RUNNING
;
2154 if (p
->sched_class
->task_wake_up
)
2155 p
->sched_class
->task_wake_up(rq
, p
);
2158 task_rq_unlock(rq
, &flags
);
2163 int wake_up_process(struct task_struct
*p
)
2165 return try_to_wake_up(p
, TASK_ALL
, 0);
2167 EXPORT_SYMBOL(wake_up_process
);
2169 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2171 return try_to_wake_up(p
, state
, 0);
2175 * Perform scheduler related setup for a newly forked process p.
2176 * p is forked by current.
2178 * __sched_fork() is basic setup used by init_idle() too:
2180 static void __sched_fork(struct task_struct
*p
)
2182 p
->se
.exec_start
= 0;
2183 p
->se
.sum_exec_runtime
= 0;
2184 p
->se
.prev_sum_exec_runtime
= 0;
2185 p
->se
.last_wakeup
= 0;
2186 p
->se
.avg_overlap
= 0;
2188 #ifdef CONFIG_SCHEDSTATS
2189 p
->se
.wait_start
= 0;
2190 p
->se
.sum_sleep_runtime
= 0;
2191 p
->se
.sleep_start
= 0;
2192 p
->se
.block_start
= 0;
2193 p
->se
.sleep_max
= 0;
2194 p
->se
.block_max
= 0;
2196 p
->se
.slice_max
= 0;
2200 INIT_LIST_HEAD(&p
->rt
.run_list
);
2202 INIT_LIST_HEAD(&p
->se
.group_node
);
2204 #ifdef CONFIG_PREEMPT_NOTIFIERS
2205 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2209 * We mark the process as running here, but have not actually
2210 * inserted it onto the runqueue yet. This guarantees that
2211 * nobody will actually run it, and a signal or other external
2212 * event cannot wake it up and insert it on the runqueue either.
2214 p
->state
= TASK_RUNNING
;
2218 * fork()/clone()-time setup:
2220 void sched_fork(struct task_struct
*p
, int clone_flags
)
2222 int cpu
= get_cpu();
2227 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2229 set_task_cpu(p
, cpu
);
2232 * Make sure we do not leak PI boosting priority to the child:
2234 p
->prio
= current
->normal_prio
;
2235 if (!rt_prio(p
->prio
))
2236 p
->sched_class
= &fair_sched_class
;
2238 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2239 if (likely(sched_info_on()))
2240 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2242 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2245 #ifdef CONFIG_PREEMPT
2246 /* Want to start with kernel preemption disabled. */
2247 task_thread_info(p
)->preempt_count
= 1;
2253 * wake_up_new_task - wake up a newly created task for the first time.
2255 * This function will do some initial scheduler statistics housekeeping
2256 * that must be done for every newly created context, then puts the task
2257 * on the runqueue and wakes it.
2259 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2261 unsigned long flags
;
2264 rq
= task_rq_lock(p
, &flags
);
2265 BUG_ON(p
->state
!= TASK_RUNNING
);
2266 update_rq_clock(rq
);
2268 p
->prio
= effective_prio(p
);
2270 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2271 activate_task(rq
, p
, 0);
2274 * Let the scheduling class do new task startup
2275 * management (if any):
2277 p
->sched_class
->task_new(rq
, p
);
2278 inc_nr_running(p
, rq
);
2280 check_preempt_curr(rq
, p
);
2282 if (p
->sched_class
->task_wake_up
)
2283 p
->sched_class
->task_wake_up(rq
, p
);
2285 task_rq_unlock(rq
, &flags
);
2288 #ifdef CONFIG_PREEMPT_NOTIFIERS
2291 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2292 * @notifier: notifier struct to register
2294 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2296 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2298 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2301 * preempt_notifier_unregister - no longer interested in preemption notifications
2302 * @notifier: notifier struct to unregister
2304 * This is safe to call from within a preemption notifier.
2306 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2308 hlist_del(¬ifier
->link
);
2310 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2312 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2314 struct preempt_notifier
*notifier
;
2315 struct hlist_node
*node
;
2317 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2318 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2322 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2323 struct task_struct
*next
)
2325 struct preempt_notifier
*notifier
;
2326 struct hlist_node
*node
;
2328 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2329 notifier
->ops
->sched_out(notifier
, next
);
2334 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2339 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2340 struct task_struct
*next
)
2347 * prepare_task_switch - prepare to switch tasks
2348 * @rq: the runqueue preparing to switch
2349 * @prev: the current task that is being switched out
2350 * @next: the task we are going to switch to.
2352 * This is called with the rq lock held and interrupts off. It must
2353 * be paired with a subsequent finish_task_switch after the context
2356 * prepare_task_switch sets up locking and calls architecture specific
2360 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2361 struct task_struct
*next
)
2363 fire_sched_out_preempt_notifiers(prev
, next
);
2364 prepare_lock_switch(rq
, next
);
2365 prepare_arch_switch(next
);
2369 * finish_task_switch - clean up after a task-switch
2370 * @rq: runqueue associated with task-switch
2371 * @prev: the thread we just switched away from.
2373 * finish_task_switch must be called after the context switch, paired
2374 * with a prepare_task_switch call before the context switch.
2375 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2376 * and do any other architecture-specific cleanup actions.
2378 * Note that we may have delayed dropping an mm in context_switch(). If
2379 * so, we finish that here outside of the runqueue lock. (Doing it
2380 * with the lock held can cause deadlocks; see schedule() for
2383 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2384 __releases(rq
->lock
)
2386 struct mm_struct
*mm
= rq
->prev_mm
;
2392 * A task struct has one reference for the use as "current".
2393 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2394 * schedule one last time. The schedule call will never return, and
2395 * the scheduled task must drop that reference.
2396 * The test for TASK_DEAD must occur while the runqueue locks are
2397 * still held, otherwise prev could be scheduled on another cpu, die
2398 * there before we look at prev->state, and then the reference would
2400 * Manfred Spraul <manfred@colorfullife.com>
2402 prev_state
= prev
->state
;
2403 finish_arch_switch(prev
);
2404 finish_lock_switch(rq
, prev
);
2406 if (current
->sched_class
->post_schedule
)
2407 current
->sched_class
->post_schedule(rq
);
2410 fire_sched_in_preempt_notifiers(current
);
2413 if (unlikely(prev_state
== TASK_DEAD
)) {
2415 * Remove function-return probe instances associated with this
2416 * task and put them back on the free list.
2418 kprobe_flush_task(prev
);
2419 put_task_struct(prev
);
2424 * schedule_tail - first thing a freshly forked thread must call.
2425 * @prev: the thread we just switched away from.
2427 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2428 __releases(rq
->lock
)
2430 struct rq
*rq
= this_rq();
2432 finish_task_switch(rq
, prev
);
2433 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2434 /* In this case, finish_task_switch does not reenable preemption */
2437 if (current
->set_child_tid
)
2438 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2442 * context_switch - switch to the new MM and the new
2443 * thread's register state.
2446 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2447 struct task_struct
*next
)
2449 struct mm_struct
*mm
, *oldmm
;
2451 prepare_task_switch(rq
, prev
, next
);
2453 oldmm
= prev
->active_mm
;
2455 * For paravirt, this is coupled with an exit in switch_to to
2456 * combine the page table reload and the switch backend into
2459 arch_enter_lazy_cpu_mode();
2461 if (unlikely(!mm
)) {
2462 next
->active_mm
= oldmm
;
2463 atomic_inc(&oldmm
->mm_count
);
2464 enter_lazy_tlb(oldmm
, next
);
2466 switch_mm(oldmm
, mm
, next
);
2468 if (unlikely(!prev
->mm
)) {
2469 prev
->active_mm
= NULL
;
2470 rq
->prev_mm
= oldmm
;
2473 * Since the runqueue lock will be released by the next
2474 * task (which is an invalid locking op but in the case
2475 * of the scheduler it's an obvious special-case), so we
2476 * do an early lockdep release here:
2478 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2479 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2482 /* Here we just switch the register state and the stack. */
2483 switch_to(prev
, next
, prev
);
2487 * this_rq must be evaluated again because prev may have moved
2488 * CPUs since it called schedule(), thus the 'rq' on its stack
2489 * frame will be invalid.
2491 finish_task_switch(this_rq(), prev
);
2495 * nr_running, nr_uninterruptible and nr_context_switches:
2497 * externally visible scheduler statistics: current number of runnable
2498 * threads, current number of uninterruptible-sleeping threads, total
2499 * number of context switches performed since bootup.
2501 unsigned long nr_running(void)
2503 unsigned long i
, sum
= 0;
2505 for_each_online_cpu(i
)
2506 sum
+= cpu_rq(i
)->nr_running
;
2511 unsigned long nr_uninterruptible(void)
2513 unsigned long i
, sum
= 0;
2515 for_each_possible_cpu(i
)
2516 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2519 * Since we read the counters lockless, it might be slightly
2520 * inaccurate. Do not allow it to go below zero though:
2522 if (unlikely((long)sum
< 0))
2528 unsigned long long nr_context_switches(void)
2531 unsigned long long sum
= 0;
2533 for_each_possible_cpu(i
)
2534 sum
+= cpu_rq(i
)->nr_switches
;
2539 unsigned long nr_iowait(void)
2541 unsigned long i
, sum
= 0;
2543 for_each_possible_cpu(i
)
2544 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2549 unsigned long nr_active(void)
2551 unsigned long i
, running
= 0, uninterruptible
= 0;
2553 for_each_online_cpu(i
) {
2554 running
+= cpu_rq(i
)->nr_running
;
2555 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2558 if (unlikely((long)uninterruptible
< 0))
2559 uninterruptible
= 0;
2561 return running
+ uninterruptible
;
2565 * Update rq->cpu_load[] statistics. This function is usually called every
2566 * scheduler tick (TICK_NSEC).
2568 static void update_cpu_load(struct rq
*this_rq
)
2570 unsigned long this_load
= this_rq
->load
.weight
;
2573 this_rq
->nr_load_updates
++;
2575 /* Update our load: */
2576 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2577 unsigned long old_load
, new_load
;
2579 /* scale is effectively 1 << i now, and >> i divides by scale */
2581 old_load
= this_rq
->cpu_load
[i
];
2582 new_load
= this_load
;
2584 * Round up the averaging division if load is increasing. This
2585 * prevents us from getting stuck on 9 if the load is 10, for
2588 if (new_load
> old_load
)
2589 new_load
+= scale
-1;
2590 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2597 * double_rq_lock - safely lock two runqueues
2599 * Note this does not disable interrupts like task_rq_lock,
2600 * you need to do so manually before calling.
2602 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2603 __acquires(rq1
->lock
)
2604 __acquires(rq2
->lock
)
2606 BUG_ON(!irqs_disabled());
2608 spin_lock(&rq1
->lock
);
2609 __acquire(rq2
->lock
); /* Fake it out ;) */
2612 spin_lock(&rq1
->lock
);
2613 spin_lock(&rq2
->lock
);
2615 spin_lock(&rq2
->lock
);
2616 spin_lock(&rq1
->lock
);
2619 update_rq_clock(rq1
);
2620 update_rq_clock(rq2
);
2624 * double_rq_unlock - safely unlock two runqueues
2626 * Note this does not restore interrupts like task_rq_unlock,
2627 * you need to do so manually after calling.
2629 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2630 __releases(rq1
->lock
)
2631 __releases(rq2
->lock
)
2633 spin_unlock(&rq1
->lock
);
2635 spin_unlock(&rq2
->lock
);
2637 __release(rq2
->lock
);
2641 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2643 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2644 __releases(this_rq
->lock
)
2645 __acquires(busiest
->lock
)
2646 __acquires(this_rq
->lock
)
2650 if (unlikely(!irqs_disabled())) {
2651 /* printk() doesn't work good under rq->lock */
2652 spin_unlock(&this_rq
->lock
);
2655 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2656 if (busiest
< this_rq
) {
2657 spin_unlock(&this_rq
->lock
);
2658 spin_lock(&busiest
->lock
);
2659 spin_lock(&this_rq
->lock
);
2662 spin_lock(&busiest
->lock
);
2668 * If dest_cpu is allowed for this process, migrate the task to it.
2669 * This is accomplished by forcing the cpu_allowed mask to only
2670 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2671 * the cpu_allowed mask is restored.
2673 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2675 struct migration_req req
;
2676 unsigned long flags
;
2679 rq
= task_rq_lock(p
, &flags
);
2680 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2681 || unlikely(cpu_is_offline(dest_cpu
)))
2684 /* force the process onto the specified CPU */
2685 if (migrate_task(p
, dest_cpu
, &req
)) {
2686 /* Need to wait for migration thread (might exit: take ref). */
2687 struct task_struct
*mt
= rq
->migration_thread
;
2689 get_task_struct(mt
);
2690 task_rq_unlock(rq
, &flags
);
2691 wake_up_process(mt
);
2692 put_task_struct(mt
);
2693 wait_for_completion(&req
.done
);
2698 task_rq_unlock(rq
, &flags
);
2702 * sched_exec - execve() is a valuable balancing opportunity, because at
2703 * this point the task has the smallest effective memory and cache footprint.
2705 void sched_exec(void)
2707 int new_cpu
, this_cpu
= get_cpu();
2708 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2710 if (new_cpu
!= this_cpu
)
2711 sched_migrate_task(current
, new_cpu
);
2715 * pull_task - move a task from a remote runqueue to the local runqueue.
2716 * Both runqueues must be locked.
2718 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2719 struct rq
*this_rq
, int this_cpu
)
2721 deactivate_task(src_rq
, p
, 0);
2722 set_task_cpu(p
, this_cpu
);
2723 activate_task(this_rq
, p
, 0);
2725 * Note that idle threads have a prio of MAX_PRIO, for this test
2726 * to be always true for them.
2728 check_preempt_curr(this_rq
, p
);
2732 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2735 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2736 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2740 * We do not migrate tasks that are:
2741 * 1) running (obviously), or
2742 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2743 * 3) are cache-hot on their current CPU.
2745 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2746 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2751 if (task_running(rq
, p
)) {
2752 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2757 * Aggressive migration if:
2758 * 1) task is cache cold, or
2759 * 2) too many balance attempts have failed.
2762 if (!task_hot(p
, rq
->clock
, sd
) ||
2763 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2764 #ifdef CONFIG_SCHEDSTATS
2765 if (task_hot(p
, rq
->clock
, sd
)) {
2766 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2767 schedstat_inc(p
, se
.nr_forced_migrations
);
2773 if (task_hot(p
, rq
->clock
, sd
)) {
2774 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2780 static unsigned long
2781 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2782 unsigned long max_load_move
, struct sched_domain
*sd
,
2783 enum cpu_idle_type idle
, int *all_pinned
,
2784 int *this_best_prio
, struct rq_iterator
*iterator
)
2786 int loops
= 0, pulled
= 0, pinned
= 0, skip_for_load
;
2787 struct task_struct
*p
;
2788 long rem_load_move
= max_load_move
;
2790 if (max_load_move
== 0)
2796 * Start the load-balancing iterator:
2798 p
= iterator
->start(iterator
->arg
);
2800 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2803 * To help distribute high priority tasks across CPUs we don't
2804 * skip a task if it will be the highest priority task (i.e. smallest
2805 * prio value) on its new queue regardless of its load weight
2807 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2808 SCHED_LOAD_SCALE_FUZZ
;
2809 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2810 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2811 p
= iterator
->next(iterator
->arg
);
2815 pull_task(busiest
, p
, this_rq
, this_cpu
);
2817 rem_load_move
-= p
->se
.load
.weight
;
2820 * We only want to steal up to the prescribed amount of weighted load.
2822 if (rem_load_move
> 0) {
2823 if (p
->prio
< *this_best_prio
)
2824 *this_best_prio
= p
->prio
;
2825 p
= iterator
->next(iterator
->arg
);
2830 * Right now, this is one of only two places pull_task() is called,
2831 * so we can safely collect pull_task() stats here rather than
2832 * inside pull_task().
2834 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2837 *all_pinned
= pinned
;
2839 return max_load_move
- rem_load_move
;
2843 * move_tasks tries to move up to max_load_move weighted load from busiest to
2844 * this_rq, as part of a balancing operation within domain "sd".
2845 * Returns 1 if successful and 0 otherwise.
2847 * Called with both runqueues locked.
2849 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2850 unsigned long max_load_move
,
2851 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2854 const struct sched_class
*class = sched_class_highest
;
2855 unsigned long total_load_moved
= 0;
2856 int this_best_prio
= this_rq
->curr
->prio
;
2860 class->load_balance(this_rq
, this_cpu
, busiest
,
2861 max_load_move
- total_load_moved
,
2862 sd
, idle
, all_pinned
, &this_best_prio
);
2863 class = class->next
;
2864 } while (class && max_load_move
> total_load_moved
);
2866 return total_load_moved
> 0;
2870 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2871 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2872 struct rq_iterator
*iterator
)
2874 struct task_struct
*p
= iterator
->start(iterator
->arg
);
2878 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2879 pull_task(busiest
, p
, this_rq
, this_cpu
);
2881 * Right now, this is only the second place pull_task()
2882 * is called, so we can safely collect pull_task()
2883 * stats here rather than inside pull_task().
2885 schedstat_inc(sd
, lb_gained
[idle
]);
2889 p
= iterator
->next(iterator
->arg
);
2896 * move_one_task tries to move exactly one task from busiest to this_rq, as
2897 * part of active balancing operations within "domain".
2898 * Returns 1 if successful and 0 otherwise.
2900 * Called with both runqueues locked.
2902 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2903 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2905 const struct sched_class
*class;
2907 for (class = sched_class_highest
; class; class = class->next
)
2908 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
2915 * find_busiest_group finds and returns the busiest CPU group within the
2916 * domain. It calculates and returns the amount of weighted load which
2917 * should be moved to restore balance via the imbalance parameter.
2919 static struct sched_group
*
2920 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2921 unsigned long *imbalance
, enum cpu_idle_type idle
,
2922 int *sd_idle
, const cpumask_t
*cpus
, int *balance
)
2924 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2925 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2926 unsigned long max_pull
;
2927 unsigned long busiest_load_per_task
, busiest_nr_running
;
2928 unsigned long this_load_per_task
, this_nr_running
;
2929 int load_idx
, group_imb
= 0;
2930 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2931 int power_savings_balance
= 1;
2932 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2933 unsigned long min_nr_running
= ULONG_MAX
;
2934 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2937 max_load
= this_load
= total_load
= total_pwr
= 0;
2938 busiest_load_per_task
= busiest_nr_running
= 0;
2939 this_load_per_task
= this_nr_running
= 0;
2940 if (idle
== CPU_NOT_IDLE
)
2941 load_idx
= sd
->busy_idx
;
2942 else if (idle
== CPU_NEWLY_IDLE
)
2943 load_idx
= sd
->newidle_idx
;
2945 load_idx
= sd
->idle_idx
;
2948 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
2951 int __group_imb
= 0;
2952 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2953 unsigned long sum_nr_running
, sum_weighted_load
;
2955 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2958 balance_cpu
= first_cpu(group
->cpumask
);
2960 /* Tally up the load of all CPUs in the group */
2961 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2963 min_cpu_load
= ~0UL;
2965 for_each_cpu_mask(i
, group
->cpumask
) {
2968 if (!cpu_isset(i
, *cpus
))
2973 if (*sd_idle
&& rq
->nr_running
)
2976 /* Bias balancing toward cpus of our domain */
2978 if (idle_cpu(i
) && !first_idle_cpu
) {
2983 load
= target_load(i
, load_idx
);
2985 load
= source_load(i
, load_idx
);
2986 if (load
> max_cpu_load
)
2987 max_cpu_load
= load
;
2988 if (min_cpu_load
> load
)
2989 min_cpu_load
= load
;
2993 sum_nr_running
+= rq
->nr_running
;
2994 sum_weighted_load
+= weighted_cpuload(i
);
2998 * First idle cpu or the first cpu(busiest) in this sched group
2999 * is eligible for doing load balancing at this and above
3000 * domains. In the newly idle case, we will allow all the cpu's
3001 * to do the newly idle load balance.
3003 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3004 balance_cpu
!= this_cpu
&& balance
) {
3009 total_load
+= avg_load
;
3010 total_pwr
+= group
->__cpu_power
;
3012 /* Adjust by relative CPU power of the group */
3013 avg_load
= sg_div_cpu_power(group
,
3014 avg_load
* SCHED_LOAD_SCALE
);
3016 if ((max_cpu_load
- min_cpu_load
) > SCHED_LOAD_SCALE
)
3019 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3022 this_load
= avg_load
;
3024 this_nr_running
= sum_nr_running
;
3025 this_load_per_task
= sum_weighted_load
;
3026 } else if (avg_load
> max_load
&&
3027 (sum_nr_running
> group_capacity
|| __group_imb
)) {
3028 max_load
= avg_load
;
3030 busiest_nr_running
= sum_nr_running
;
3031 busiest_load_per_task
= sum_weighted_load
;
3032 group_imb
= __group_imb
;
3035 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3037 * Busy processors will not participate in power savings
3040 if (idle
== CPU_NOT_IDLE
||
3041 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3045 * If the local group is idle or completely loaded
3046 * no need to do power savings balance at this domain
3048 if (local_group
&& (this_nr_running
>= group_capacity
||
3050 power_savings_balance
= 0;
3053 * If a group is already running at full capacity or idle,
3054 * don't include that group in power savings calculations
3056 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
3061 * Calculate the group which has the least non-idle load.
3062 * This is the group from where we need to pick up the load
3065 if ((sum_nr_running
< min_nr_running
) ||
3066 (sum_nr_running
== min_nr_running
&&
3067 first_cpu(group
->cpumask
) <
3068 first_cpu(group_min
->cpumask
))) {
3070 min_nr_running
= sum_nr_running
;
3071 min_load_per_task
= sum_weighted_load
/
3076 * Calculate the group which is almost near its
3077 * capacity but still has some space to pick up some load
3078 * from other group and save more power
3080 if (sum_nr_running
<= group_capacity
- 1) {
3081 if (sum_nr_running
> leader_nr_running
||
3082 (sum_nr_running
== leader_nr_running
&&
3083 first_cpu(group
->cpumask
) >
3084 first_cpu(group_leader
->cpumask
))) {
3085 group_leader
= group
;
3086 leader_nr_running
= sum_nr_running
;
3091 group
= group
->next
;
3092 } while (group
!= sd
->groups
);
3094 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
3097 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
3099 if (this_load
>= avg_load
||
3100 100*max_load
<= sd
->imbalance_pct
*this_load
)
3103 busiest_load_per_task
/= busiest_nr_running
;
3105 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
3108 * We're trying to get all the cpus to the average_load, so we don't
3109 * want to push ourselves above the average load, nor do we wish to
3110 * reduce the max loaded cpu below the average load, as either of these
3111 * actions would just result in more rebalancing later, and ping-pong
3112 * tasks around. Thus we look for the minimum possible imbalance.
3113 * Negative imbalances (*we* are more loaded than anyone else) will
3114 * be counted as no imbalance for these purposes -- we can't fix that
3115 * by pulling tasks to us. Be careful of negative numbers as they'll
3116 * appear as very large values with unsigned longs.
3118 if (max_load
<= busiest_load_per_task
)
3122 * In the presence of smp nice balancing, certain scenarios can have
3123 * max load less than avg load(as we skip the groups at or below
3124 * its cpu_power, while calculating max_load..)
3126 if (max_load
< avg_load
) {
3128 goto small_imbalance
;
3131 /* Don't want to pull so many tasks that a group would go idle */
3132 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
3134 /* How much load to actually move to equalise the imbalance */
3135 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
3136 (avg_load
- this_load
) * this->__cpu_power
)
3140 * if *imbalance is less than the average load per runnable task
3141 * there is no gaurantee that any tasks will be moved so we'll have
3142 * a think about bumping its value to force at least one task to be
3145 if (*imbalance
< busiest_load_per_task
) {
3146 unsigned long tmp
, pwr_now
, pwr_move
;
3150 pwr_move
= pwr_now
= 0;
3152 if (this_nr_running
) {
3153 this_load_per_task
/= this_nr_running
;
3154 if (busiest_load_per_task
> this_load_per_task
)
3157 this_load_per_task
= SCHED_LOAD_SCALE
;
3159 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
3160 busiest_load_per_task
* imbn
) {
3161 *imbalance
= busiest_load_per_task
;
3166 * OK, we don't have enough imbalance to justify moving tasks,
3167 * however we may be able to increase total CPU power used by
3171 pwr_now
+= busiest
->__cpu_power
*
3172 min(busiest_load_per_task
, max_load
);
3173 pwr_now
+= this->__cpu_power
*
3174 min(this_load_per_task
, this_load
);
3175 pwr_now
/= SCHED_LOAD_SCALE
;
3177 /* Amount of load we'd subtract */
3178 tmp
= sg_div_cpu_power(busiest
,
3179 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3181 pwr_move
+= busiest
->__cpu_power
*
3182 min(busiest_load_per_task
, max_load
- tmp
);
3184 /* Amount of load we'd add */
3185 if (max_load
* busiest
->__cpu_power
<
3186 busiest_load_per_task
* SCHED_LOAD_SCALE
)
3187 tmp
= sg_div_cpu_power(this,
3188 max_load
* busiest
->__cpu_power
);
3190 tmp
= sg_div_cpu_power(this,
3191 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3192 pwr_move
+= this->__cpu_power
*
3193 min(this_load_per_task
, this_load
+ tmp
);
3194 pwr_move
/= SCHED_LOAD_SCALE
;
3196 /* Move if we gain throughput */
3197 if (pwr_move
> pwr_now
)
3198 *imbalance
= busiest_load_per_task
;
3204 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3205 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3208 if (this == group_leader
&& group_leader
!= group_min
) {
3209 *imbalance
= min_load_per_task
;
3219 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3222 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3223 unsigned long imbalance
, const cpumask_t
*cpus
)
3225 struct rq
*busiest
= NULL
, *rq
;
3226 unsigned long max_load
= 0;
3229 for_each_cpu_mask(i
, group
->cpumask
) {
3232 if (!cpu_isset(i
, *cpus
))
3236 wl
= weighted_cpuload(i
);
3238 if (rq
->nr_running
== 1 && wl
> imbalance
)
3241 if (wl
> max_load
) {
3251 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3252 * so long as it is large enough.
3254 #define MAX_PINNED_INTERVAL 512
3257 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3258 * tasks if there is an imbalance.
3260 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3261 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3262 int *balance
, cpumask_t
*cpus
)
3264 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3265 struct sched_group
*group
;
3266 unsigned long imbalance
;
3268 unsigned long flags
;
3273 * When power savings policy is enabled for the parent domain, idle
3274 * sibling can pick up load irrespective of busy siblings. In this case,
3275 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3276 * portraying it as CPU_NOT_IDLE.
3278 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3279 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3282 schedstat_inc(sd
, lb_count
[idle
]);
3285 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3292 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3296 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3298 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3302 BUG_ON(busiest
== this_rq
);
3304 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3307 if (busiest
->nr_running
> 1) {
3309 * Attempt to move tasks. If find_busiest_group has found
3310 * an imbalance but busiest->nr_running <= 1, the group is
3311 * still unbalanced. ld_moved simply stays zero, so it is
3312 * correctly treated as an imbalance.
3314 local_irq_save(flags
);
3315 double_rq_lock(this_rq
, busiest
);
3316 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3317 imbalance
, sd
, idle
, &all_pinned
);
3318 double_rq_unlock(this_rq
, busiest
);
3319 local_irq_restore(flags
);
3322 * some other cpu did the load balance for us.
3324 if (ld_moved
&& this_cpu
!= smp_processor_id())
3325 resched_cpu(this_cpu
);
3327 /* All tasks on this runqueue were pinned by CPU affinity */
3328 if (unlikely(all_pinned
)) {
3329 cpu_clear(cpu_of(busiest
), *cpus
);
3330 if (!cpus_empty(*cpus
))
3337 schedstat_inc(sd
, lb_failed
[idle
]);
3338 sd
->nr_balance_failed
++;
3340 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3342 spin_lock_irqsave(&busiest
->lock
, flags
);
3344 /* don't kick the migration_thread, if the curr
3345 * task on busiest cpu can't be moved to this_cpu
3347 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
3348 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3350 goto out_one_pinned
;
3353 if (!busiest
->active_balance
) {
3354 busiest
->active_balance
= 1;
3355 busiest
->push_cpu
= this_cpu
;
3358 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3360 wake_up_process(busiest
->migration_thread
);
3363 * We've kicked active balancing, reset the failure
3366 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3369 sd
->nr_balance_failed
= 0;
3371 if (likely(!active_balance
)) {
3372 /* We were unbalanced, so reset the balancing interval */
3373 sd
->balance_interval
= sd
->min_interval
;
3376 * If we've begun active balancing, start to back off. This
3377 * case may not be covered by the all_pinned logic if there
3378 * is only 1 task on the busy runqueue (because we don't call
3381 if (sd
->balance_interval
< sd
->max_interval
)
3382 sd
->balance_interval
*= 2;
3385 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3386 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3391 schedstat_inc(sd
, lb_balanced
[idle
]);
3393 sd
->nr_balance_failed
= 0;
3396 /* tune up the balancing interval */
3397 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3398 (sd
->balance_interval
< sd
->max_interval
))
3399 sd
->balance_interval
*= 2;
3401 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3402 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3408 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3409 * tasks if there is an imbalance.
3411 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3412 * this_rq is locked.
3415 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
,
3418 struct sched_group
*group
;
3419 struct rq
*busiest
= NULL
;
3420 unsigned long imbalance
;
3428 * When power savings policy is enabled for the parent domain, idle
3429 * sibling can pick up load irrespective of busy siblings. In this case,
3430 * let the state of idle sibling percolate up as IDLE, instead of
3431 * portraying it as CPU_NOT_IDLE.
3433 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3434 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3437 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3439 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3440 &sd_idle
, cpus
, NULL
);
3442 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3446 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
3448 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3452 BUG_ON(busiest
== this_rq
);
3454 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3457 if (busiest
->nr_running
> 1) {
3458 /* Attempt to move tasks */
3459 double_lock_balance(this_rq
, busiest
);
3460 /* this_rq->clock is already updated */
3461 update_rq_clock(busiest
);
3462 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3463 imbalance
, sd
, CPU_NEWLY_IDLE
,
3465 spin_unlock(&busiest
->lock
);
3467 if (unlikely(all_pinned
)) {
3468 cpu_clear(cpu_of(busiest
), *cpus
);
3469 if (!cpus_empty(*cpus
))
3475 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3476 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3477 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3480 sd
->nr_balance_failed
= 0;
3485 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3486 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3487 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3489 sd
->nr_balance_failed
= 0;
3495 * idle_balance is called by schedule() if this_cpu is about to become
3496 * idle. Attempts to pull tasks from other CPUs.
3498 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3500 struct sched_domain
*sd
;
3501 int pulled_task
= -1;
3502 unsigned long next_balance
= jiffies
+ HZ
;
3505 for_each_domain(this_cpu
, sd
) {
3506 unsigned long interval
;
3508 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3511 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3512 /* If we've pulled tasks over stop searching: */
3513 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
3516 interval
= msecs_to_jiffies(sd
->balance_interval
);
3517 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3518 next_balance
= sd
->last_balance
+ interval
;
3522 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3524 * We are going idle. next_balance may be set based on
3525 * a busy processor. So reset next_balance.
3527 this_rq
->next_balance
= next_balance
;
3532 * active_load_balance is run by migration threads. It pushes running tasks
3533 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3534 * running on each physical CPU where possible, and avoids physical /
3535 * logical imbalances.
3537 * Called with busiest_rq locked.
3539 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3541 int target_cpu
= busiest_rq
->push_cpu
;
3542 struct sched_domain
*sd
;
3543 struct rq
*target_rq
;
3545 /* Is there any task to move? */
3546 if (busiest_rq
->nr_running
<= 1)
3549 target_rq
= cpu_rq(target_cpu
);
3552 * This condition is "impossible", if it occurs
3553 * we need to fix it. Originally reported by
3554 * Bjorn Helgaas on a 128-cpu setup.
3556 BUG_ON(busiest_rq
== target_rq
);
3558 /* move a task from busiest_rq to target_rq */
3559 double_lock_balance(busiest_rq
, target_rq
);
3560 update_rq_clock(busiest_rq
);
3561 update_rq_clock(target_rq
);
3563 /* Search for an sd spanning us and the target CPU. */
3564 for_each_domain(target_cpu
, sd
) {
3565 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3566 cpu_isset(busiest_cpu
, sd
->span
))
3571 schedstat_inc(sd
, alb_count
);
3573 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3575 schedstat_inc(sd
, alb_pushed
);
3577 schedstat_inc(sd
, alb_failed
);
3579 spin_unlock(&target_rq
->lock
);
3584 atomic_t load_balancer
;
3586 } nohz ____cacheline_aligned
= {
3587 .load_balancer
= ATOMIC_INIT(-1),
3588 .cpu_mask
= CPU_MASK_NONE
,
3592 * This routine will try to nominate the ilb (idle load balancing)
3593 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3594 * load balancing on behalf of all those cpus. If all the cpus in the system
3595 * go into this tickless mode, then there will be no ilb owner (as there is
3596 * no need for one) and all the cpus will sleep till the next wakeup event
3599 * For the ilb owner, tick is not stopped. And this tick will be used
3600 * for idle load balancing. ilb owner will still be part of
3603 * While stopping the tick, this cpu will become the ilb owner if there
3604 * is no other owner. And will be the owner till that cpu becomes busy
3605 * or if all cpus in the system stop their ticks at which point
3606 * there is no need for ilb owner.
3608 * When the ilb owner becomes busy, it nominates another owner, during the
3609 * next busy scheduler_tick()
3611 int select_nohz_load_balancer(int stop_tick
)
3613 int cpu
= smp_processor_id();
3616 cpu_set(cpu
, nohz
.cpu_mask
);
3617 cpu_rq(cpu
)->in_nohz_recently
= 1;
3620 * If we are going offline and still the leader, give up!
3622 if (cpu_is_offline(cpu
) &&
3623 atomic_read(&nohz
.load_balancer
) == cpu
) {
3624 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3629 /* time for ilb owner also to sleep */
3630 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3631 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3632 atomic_set(&nohz
.load_balancer
, -1);
3636 if (atomic_read(&nohz
.load_balancer
) == -1) {
3637 /* make me the ilb owner */
3638 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3640 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3643 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3646 cpu_clear(cpu
, nohz
.cpu_mask
);
3648 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3649 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3656 static DEFINE_SPINLOCK(balancing
);
3659 * It checks each scheduling domain to see if it is due to be balanced,
3660 * and initiates a balancing operation if so.
3662 * Balancing parameters are set up in arch_init_sched_domains.
3664 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3667 struct rq
*rq
= cpu_rq(cpu
);
3668 unsigned long interval
;
3669 struct sched_domain
*sd
;
3670 /* Earliest time when we have to do rebalance again */
3671 unsigned long next_balance
= jiffies
+ 60*HZ
;
3672 int update_next_balance
= 0;
3676 for_each_domain(cpu
, sd
) {
3677 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3680 interval
= sd
->balance_interval
;
3681 if (idle
!= CPU_IDLE
)
3682 interval
*= sd
->busy_factor
;
3684 /* scale ms to jiffies */
3685 interval
= msecs_to_jiffies(interval
);
3686 if (unlikely(!interval
))
3688 if (interval
> HZ
*NR_CPUS
/10)
3689 interval
= HZ
*NR_CPUS
/10;
3691 need_serialize
= sd
->flags
& SD_SERIALIZE
;
3693 if (need_serialize
) {
3694 if (!spin_trylock(&balancing
))
3698 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3699 if (load_balance(cpu
, rq
, sd
, idle
, &balance
, &tmp
)) {
3701 * We've pulled tasks over so either we're no
3702 * longer idle, or one of our SMT siblings is
3705 idle
= CPU_NOT_IDLE
;
3707 sd
->last_balance
= jiffies
;
3710 spin_unlock(&balancing
);
3712 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3713 next_balance
= sd
->last_balance
+ interval
;
3714 update_next_balance
= 1;
3718 * Stop the load balance at this level. There is another
3719 * CPU in our sched group which is doing load balancing more
3727 * next_balance will be updated only when there is a need.
3728 * When the cpu is attached to null domain for ex, it will not be
3731 if (likely(update_next_balance
))
3732 rq
->next_balance
= next_balance
;
3736 * run_rebalance_domains is triggered when needed from the scheduler tick.
3737 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3738 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3740 static void run_rebalance_domains(struct softirq_action
*h
)
3742 int this_cpu
= smp_processor_id();
3743 struct rq
*this_rq
= cpu_rq(this_cpu
);
3744 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3745 CPU_IDLE
: CPU_NOT_IDLE
;
3747 rebalance_domains(this_cpu
, idle
);
3751 * If this cpu is the owner for idle load balancing, then do the
3752 * balancing on behalf of the other idle cpus whose ticks are
3755 if (this_rq
->idle_at_tick
&&
3756 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3757 cpumask_t cpus
= nohz
.cpu_mask
;
3761 cpu_clear(this_cpu
, cpus
);
3762 for_each_cpu_mask(balance_cpu
, cpus
) {
3764 * If this cpu gets work to do, stop the load balancing
3765 * work being done for other cpus. Next load
3766 * balancing owner will pick it up.
3771 rebalance_domains(balance_cpu
, CPU_IDLE
);
3773 rq
= cpu_rq(balance_cpu
);
3774 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3775 this_rq
->next_balance
= rq
->next_balance
;
3782 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3784 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3785 * idle load balancing owner or decide to stop the periodic load balancing,
3786 * if the whole system is idle.
3788 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3792 * If we were in the nohz mode recently and busy at the current
3793 * scheduler tick, then check if we need to nominate new idle
3796 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3797 rq
->in_nohz_recently
= 0;
3799 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3800 cpu_clear(cpu
, nohz
.cpu_mask
);
3801 atomic_set(&nohz
.load_balancer
, -1);
3804 if (atomic_read(&nohz
.load_balancer
) == -1) {
3806 * simple selection for now: Nominate the
3807 * first cpu in the nohz list to be the next
3810 * TBD: Traverse the sched domains and nominate
3811 * the nearest cpu in the nohz.cpu_mask.
3813 int ilb
= first_cpu(nohz
.cpu_mask
);
3815 if (ilb
< nr_cpu_ids
)
3821 * If this cpu is idle and doing idle load balancing for all the
3822 * cpus with ticks stopped, is it time for that to stop?
3824 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3825 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3831 * If this cpu is idle and the idle load balancing is done by
3832 * someone else, then no need raise the SCHED_SOFTIRQ
3834 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3835 cpu_isset(cpu
, nohz
.cpu_mask
))
3838 if (time_after_eq(jiffies
, rq
->next_balance
))
3839 raise_softirq(SCHED_SOFTIRQ
);
3842 #else /* CONFIG_SMP */
3845 * on UP we do not need to balance between CPUs:
3847 static inline void idle_balance(int cpu
, struct rq
*rq
)
3853 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3855 EXPORT_PER_CPU_SYMBOL(kstat
);
3858 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3859 * that have not yet been banked in case the task is currently running.
3861 unsigned long long task_sched_runtime(struct task_struct
*p
)
3863 unsigned long flags
;
3867 rq
= task_rq_lock(p
, &flags
);
3868 ns
= p
->se
.sum_exec_runtime
;
3869 if (task_current(rq
, p
)) {
3870 update_rq_clock(rq
);
3871 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3872 if ((s64
)delta_exec
> 0)
3875 task_rq_unlock(rq
, &flags
);
3881 * Account user cpu time to a process.
3882 * @p: the process that the cpu time gets accounted to
3883 * @cputime: the cpu time spent in user space since the last update
3885 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3887 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3890 p
->utime
= cputime_add(p
->utime
, cputime
);
3892 /* Add user time to cpustat. */
3893 tmp
= cputime_to_cputime64(cputime
);
3894 if (TASK_NICE(p
) > 0)
3895 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3897 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3901 * Account guest cpu time to a process.
3902 * @p: the process that the cpu time gets accounted to
3903 * @cputime: the cpu time spent in virtual machine since the last update
3905 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
3908 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3910 tmp
= cputime_to_cputime64(cputime
);
3912 p
->utime
= cputime_add(p
->utime
, cputime
);
3913 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3915 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3916 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3920 * Account scaled user cpu time to a process.
3921 * @p: the process that the cpu time gets accounted to
3922 * @cputime: the cpu time spent in user space since the last update
3924 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3926 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
3930 * Account system cpu time to a process.
3931 * @p: the process that the cpu time gets accounted to
3932 * @hardirq_offset: the offset to subtract from hardirq_count()
3933 * @cputime: the cpu time spent in kernel space since the last update
3935 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3938 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3939 struct rq
*rq
= this_rq();
3942 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
3943 account_guest_time(p
, cputime
);
3947 p
->stime
= cputime_add(p
->stime
, cputime
);
3949 /* Add system time to cpustat. */
3950 tmp
= cputime_to_cputime64(cputime
);
3951 if (hardirq_count() - hardirq_offset
)
3952 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3953 else if (softirq_count())
3954 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3955 else if (p
!= rq
->idle
)
3956 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3957 else if (atomic_read(&rq
->nr_iowait
) > 0)
3958 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3960 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3961 /* Account for system time used */
3962 acct_update_integrals(p
);
3966 * Account scaled system cpu time to a process.
3967 * @p: the process that the cpu time gets accounted to
3968 * @hardirq_offset: the offset to subtract from hardirq_count()
3969 * @cputime: the cpu time spent in kernel space since the last update
3971 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3973 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
3977 * Account for involuntary wait time.
3978 * @p: the process from which the cpu time has been stolen
3979 * @steal: the cpu time spent in involuntary wait
3981 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3983 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3984 cputime64_t tmp
= cputime_to_cputime64(steal
);
3985 struct rq
*rq
= this_rq();
3987 if (p
== rq
->idle
) {
3988 p
->stime
= cputime_add(p
->stime
, steal
);
3989 if (atomic_read(&rq
->nr_iowait
) > 0)
3990 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3992 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3994 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3998 * This function gets called by the timer code, with HZ frequency.
3999 * We call it with interrupts disabled.
4001 * It also gets called by the fork code, when changing the parent's
4004 void scheduler_tick(void)
4006 int cpu
= smp_processor_id();
4007 struct rq
*rq
= cpu_rq(cpu
);
4008 struct task_struct
*curr
= rq
->curr
;
4012 spin_lock(&rq
->lock
);
4013 update_rq_clock(rq
);
4014 update_cpu_load(rq
);
4015 curr
->sched_class
->task_tick(rq
, curr
, 0);
4016 spin_unlock(&rq
->lock
);
4019 rq
->idle_at_tick
= idle_cpu(cpu
);
4020 trigger_load_balance(rq
, cpu
);
4024 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4026 void __kprobes
add_preempt_count(int val
)
4031 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4033 preempt_count() += val
;
4035 * Spinlock count overflowing soon?
4037 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4040 EXPORT_SYMBOL(add_preempt_count
);
4042 void __kprobes
sub_preempt_count(int val
)
4047 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4050 * Is the spinlock portion underflowing?
4052 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4053 !(preempt_count() & PREEMPT_MASK
)))
4056 preempt_count() -= val
;
4058 EXPORT_SYMBOL(sub_preempt_count
);
4063 * Print scheduling while atomic bug:
4065 static noinline
void __schedule_bug(struct task_struct
*prev
)
4067 struct pt_regs
*regs
= get_irq_regs();
4069 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4070 prev
->comm
, prev
->pid
, preempt_count());
4072 debug_show_held_locks(prev
);
4073 if (irqs_disabled())
4074 print_irqtrace_events(prev
);
4083 * Various schedule()-time debugging checks and statistics:
4085 static inline void schedule_debug(struct task_struct
*prev
)
4088 * Test if we are atomic. Since do_exit() needs to call into
4089 * schedule() atomically, we ignore that path for now.
4090 * Otherwise, whine if we are scheduling when we should not be.
4092 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4093 __schedule_bug(prev
);
4095 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4097 schedstat_inc(this_rq(), sched_count
);
4098 #ifdef CONFIG_SCHEDSTATS
4099 if (unlikely(prev
->lock_depth
>= 0)) {
4100 schedstat_inc(this_rq(), bkl_count
);
4101 schedstat_inc(prev
, sched_info
.bkl_count
);
4107 * Pick up the highest-prio task:
4109 static inline struct task_struct
*
4110 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
4112 const struct sched_class
*class;
4113 struct task_struct
*p
;
4116 * Optimization: we know that if all tasks are in
4117 * the fair class we can call that function directly:
4119 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4120 p
= fair_sched_class
.pick_next_task(rq
);
4125 class = sched_class_highest
;
4127 p
= class->pick_next_task(rq
);
4131 * Will never be NULL as the idle class always
4132 * returns a non-NULL p:
4134 class = class->next
;
4139 * schedule() is the main scheduler function.
4141 asmlinkage
void __sched
schedule(void)
4143 struct task_struct
*prev
, *next
;
4144 unsigned long *switch_count
;
4146 int cpu
, hrtick
= sched_feat(HRTICK
);
4150 cpu
= smp_processor_id();
4154 switch_count
= &prev
->nivcsw
;
4156 release_kernel_lock(prev
);
4157 need_resched_nonpreemptible
:
4159 schedule_debug(prev
);
4165 * Do the rq-clock update outside the rq lock:
4167 local_irq_disable();
4168 update_rq_clock(rq
);
4169 spin_lock(&rq
->lock
);
4170 clear_tsk_need_resched(prev
);
4172 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4173 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
4174 signal_pending(prev
))) {
4175 prev
->state
= TASK_RUNNING
;
4177 deactivate_task(rq
, prev
, 1);
4179 switch_count
= &prev
->nvcsw
;
4183 if (prev
->sched_class
->pre_schedule
)
4184 prev
->sched_class
->pre_schedule(rq
, prev
);
4187 if (unlikely(!rq
->nr_running
))
4188 idle_balance(cpu
, rq
);
4190 prev
->sched_class
->put_prev_task(rq
, prev
);
4191 next
= pick_next_task(rq
, prev
);
4193 if (likely(prev
!= next
)) {
4194 sched_info_switch(prev
, next
);
4200 context_switch(rq
, prev
, next
); /* unlocks the rq */
4202 * the context switch might have flipped the stack from under
4203 * us, hence refresh the local variables.
4205 cpu
= smp_processor_id();
4208 spin_unlock_irq(&rq
->lock
);
4213 if (unlikely(reacquire_kernel_lock(current
) < 0))
4214 goto need_resched_nonpreemptible
;
4216 preempt_enable_no_resched();
4217 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
4220 EXPORT_SYMBOL(schedule
);
4222 #ifdef CONFIG_PREEMPT
4224 * this is the entry point to schedule() from in-kernel preemption
4225 * off of preempt_enable. Kernel preemptions off return from interrupt
4226 * occur there and call schedule directly.
4228 asmlinkage
void __sched
preempt_schedule(void)
4230 struct thread_info
*ti
= current_thread_info();
4233 * If there is a non-zero preempt_count or interrupts are disabled,
4234 * we do not want to preempt the current task. Just return..
4236 if (likely(ti
->preempt_count
|| irqs_disabled()))
4240 add_preempt_count(PREEMPT_ACTIVE
);
4242 sub_preempt_count(PREEMPT_ACTIVE
);
4245 * Check again in case we missed a preemption opportunity
4246 * between schedule and now.
4249 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4251 EXPORT_SYMBOL(preempt_schedule
);
4254 * this is the entry point to schedule() from kernel preemption
4255 * off of irq context.
4256 * Note, that this is called and return with irqs disabled. This will
4257 * protect us against recursive calling from irq.
4259 asmlinkage
void __sched
preempt_schedule_irq(void)
4261 struct thread_info
*ti
= current_thread_info();
4263 /* Catch callers which need to be fixed */
4264 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4267 add_preempt_count(PREEMPT_ACTIVE
);
4270 local_irq_disable();
4271 sub_preempt_count(PREEMPT_ACTIVE
);
4274 * Check again in case we missed a preemption opportunity
4275 * between schedule and now.
4278 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4281 #endif /* CONFIG_PREEMPT */
4283 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4286 return try_to_wake_up(curr
->private, mode
, sync
);
4288 EXPORT_SYMBOL(default_wake_function
);
4291 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4292 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4293 * number) then we wake all the non-exclusive tasks and one exclusive task.
4295 * There are circumstances in which we can try to wake a task which has already
4296 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4297 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4299 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4300 int nr_exclusive
, int sync
, void *key
)
4302 wait_queue_t
*curr
, *next
;
4304 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4305 unsigned flags
= curr
->flags
;
4307 if (curr
->func(curr
, mode
, sync
, key
) &&
4308 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4314 * __wake_up - wake up threads blocked on a waitqueue.
4316 * @mode: which threads
4317 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4318 * @key: is directly passed to the wakeup function
4320 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4321 int nr_exclusive
, void *key
)
4323 unsigned long flags
;
4325 spin_lock_irqsave(&q
->lock
, flags
);
4326 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4327 spin_unlock_irqrestore(&q
->lock
, flags
);
4329 EXPORT_SYMBOL(__wake_up
);
4332 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4334 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4336 __wake_up_common(q
, mode
, 1, 0, NULL
);
4340 * __wake_up_sync - wake up threads blocked on a waitqueue.
4342 * @mode: which threads
4343 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4345 * The sync wakeup differs that the waker knows that it will schedule
4346 * away soon, so while the target thread will be woken up, it will not
4347 * be migrated to another CPU - ie. the two threads are 'synchronized'
4348 * with each other. This can prevent needless bouncing between CPUs.
4350 * On UP it can prevent extra preemption.
4353 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4355 unsigned long flags
;
4361 if (unlikely(!nr_exclusive
))
4364 spin_lock_irqsave(&q
->lock
, flags
);
4365 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4366 spin_unlock_irqrestore(&q
->lock
, flags
);
4368 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4370 void complete(struct completion
*x
)
4372 unsigned long flags
;
4374 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4376 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4377 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4379 EXPORT_SYMBOL(complete
);
4381 void complete_all(struct completion
*x
)
4383 unsigned long flags
;
4385 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4386 x
->done
+= UINT_MAX
/2;
4387 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4388 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4390 EXPORT_SYMBOL(complete_all
);
4392 static inline long __sched
4393 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4396 DECLARE_WAITQUEUE(wait
, current
);
4398 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4399 __add_wait_queue_tail(&x
->wait
, &wait
);
4401 if ((state
== TASK_INTERRUPTIBLE
&&
4402 signal_pending(current
)) ||
4403 (state
== TASK_KILLABLE
&&
4404 fatal_signal_pending(current
))) {
4405 __remove_wait_queue(&x
->wait
, &wait
);
4406 return -ERESTARTSYS
;
4408 __set_current_state(state
);
4409 spin_unlock_irq(&x
->wait
.lock
);
4410 timeout
= schedule_timeout(timeout
);
4411 spin_lock_irq(&x
->wait
.lock
);
4413 __remove_wait_queue(&x
->wait
, &wait
);
4417 __remove_wait_queue(&x
->wait
, &wait
);
4424 wait_for_common(struct completion
*x
, long timeout
, int state
)
4428 spin_lock_irq(&x
->wait
.lock
);
4429 timeout
= do_wait_for_common(x
, timeout
, state
);
4430 spin_unlock_irq(&x
->wait
.lock
);
4434 void __sched
wait_for_completion(struct completion
*x
)
4436 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4438 EXPORT_SYMBOL(wait_for_completion
);
4440 unsigned long __sched
4441 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4443 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4445 EXPORT_SYMBOL(wait_for_completion_timeout
);
4447 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4449 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4450 if (t
== -ERESTARTSYS
)
4454 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4456 unsigned long __sched
4457 wait_for_completion_interruptible_timeout(struct completion
*x
,
4458 unsigned long timeout
)
4460 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4462 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4464 int __sched
wait_for_completion_killable(struct completion
*x
)
4466 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4467 if (t
== -ERESTARTSYS
)
4471 EXPORT_SYMBOL(wait_for_completion_killable
);
4474 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4476 unsigned long flags
;
4479 init_waitqueue_entry(&wait
, current
);
4481 __set_current_state(state
);
4483 spin_lock_irqsave(&q
->lock
, flags
);
4484 __add_wait_queue(q
, &wait
);
4485 spin_unlock(&q
->lock
);
4486 timeout
= schedule_timeout(timeout
);
4487 spin_lock_irq(&q
->lock
);
4488 __remove_wait_queue(q
, &wait
);
4489 spin_unlock_irqrestore(&q
->lock
, flags
);
4494 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4496 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4498 EXPORT_SYMBOL(interruptible_sleep_on
);
4501 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4503 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4505 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4507 void __sched
sleep_on(wait_queue_head_t
*q
)
4509 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4511 EXPORT_SYMBOL(sleep_on
);
4513 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4515 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4517 EXPORT_SYMBOL(sleep_on_timeout
);
4519 #ifdef CONFIG_RT_MUTEXES
4522 * rt_mutex_setprio - set the current priority of a task
4524 * @prio: prio value (kernel-internal form)
4526 * This function changes the 'effective' priority of a task. It does
4527 * not touch ->normal_prio like __setscheduler().
4529 * Used by the rt_mutex code to implement priority inheritance logic.
4531 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4533 unsigned long flags
;
4534 int oldprio
, on_rq
, running
;
4536 const struct sched_class
*prev_class
= p
->sched_class
;
4538 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4540 rq
= task_rq_lock(p
, &flags
);
4541 update_rq_clock(rq
);
4544 on_rq
= p
->se
.on_rq
;
4545 running
= task_current(rq
, p
);
4547 dequeue_task(rq
, p
, 0);
4549 p
->sched_class
->put_prev_task(rq
, p
);
4552 p
->sched_class
= &rt_sched_class
;
4554 p
->sched_class
= &fair_sched_class
;
4559 p
->sched_class
->set_curr_task(rq
);
4561 enqueue_task(rq
, p
, 0);
4563 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4565 task_rq_unlock(rq
, &flags
);
4570 void set_user_nice(struct task_struct
*p
, long nice
)
4572 int old_prio
, delta
, on_rq
;
4573 unsigned long flags
;
4576 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4579 * We have to be careful, if called from sys_setpriority(),
4580 * the task might be in the middle of scheduling on another CPU.
4582 rq
= task_rq_lock(p
, &flags
);
4583 update_rq_clock(rq
);
4585 * The RT priorities are set via sched_setscheduler(), but we still
4586 * allow the 'normal' nice value to be set - but as expected
4587 * it wont have any effect on scheduling until the task is
4588 * SCHED_FIFO/SCHED_RR:
4590 if (task_has_rt_policy(p
)) {
4591 p
->static_prio
= NICE_TO_PRIO(nice
);
4594 on_rq
= p
->se
.on_rq
;
4596 dequeue_task(rq
, p
, 0);
4600 p
->static_prio
= NICE_TO_PRIO(nice
);
4603 p
->prio
= effective_prio(p
);
4604 delta
= p
->prio
- old_prio
;
4607 enqueue_task(rq
, p
, 0);
4610 * If the task increased its priority or is running and
4611 * lowered its priority, then reschedule its CPU:
4613 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4614 resched_task(rq
->curr
);
4617 task_rq_unlock(rq
, &flags
);
4619 EXPORT_SYMBOL(set_user_nice
);
4622 * can_nice - check if a task can reduce its nice value
4626 int can_nice(const struct task_struct
*p
, const int nice
)
4628 /* convert nice value [19,-20] to rlimit style value [1,40] */
4629 int nice_rlim
= 20 - nice
;
4631 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4632 capable(CAP_SYS_NICE
));
4635 #ifdef __ARCH_WANT_SYS_NICE
4638 * sys_nice - change the priority of the current process.
4639 * @increment: priority increment
4641 * sys_setpriority is a more generic, but much slower function that
4642 * does similar things.
4644 asmlinkage
long sys_nice(int increment
)
4649 * Setpriority might change our priority at the same moment.
4650 * We don't have to worry. Conceptually one call occurs first
4651 * and we have a single winner.
4653 if (increment
< -40)
4658 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4664 if (increment
< 0 && !can_nice(current
, nice
))
4667 retval
= security_task_setnice(current
, nice
);
4671 set_user_nice(current
, nice
);
4678 * task_prio - return the priority value of a given task.
4679 * @p: the task in question.
4681 * This is the priority value as seen by users in /proc.
4682 * RT tasks are offset by -200. Normal tasks are centered
4683 * around 0, value goes from -16 to +15.
4685 int task_prio(const struct task_struct
*p
)
4687 return p
->prio
- MAX_RT_PRIO
;
4691 * task_nice - return the nice value of a given task.
4692 * @p: the task in question.
4694 int task_nice(const struct task_struct
*p
)
4696 return TASK_NICE(p
);
4698 EXPORT_SYMBOL(task_nice
);
4701 * idle_cpu - is a given cpu idle currently?
4702 * @cpu: the processor in question.
4704 int idle_cpu(int cpu
)
4706 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4710 * idle_task - return the idle task for a given cpu.
4711 * @cpu: the processor in question.
4713 struct task_struct
*idle_task(int cpu
)
4715 return cpu_rq(cpu
)->idle
;
4719 * find_process_by_pid - find a process with a matching PID value.
4720 * @pid: the pid in question.
4722 static struct task_struct
*find_process_by_pid(pid_t pid
)
4724 return pid
? find_task_by_vpid(pid
) : current
;
4727 /* Actually do priority change: must hold rq lock. */
4729 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4731 BUG_ON(p
->se
.on_rq
);
4734 switch (p
->policy
) {
4738 p
->sched_class
= &fair_sched_class
;
4742 p
->sched_class
= &rt_sched_class
;
4746 p
->rt_priority
= prio
;
4747 p
->normal_prio
= normal_prio(p
);
4748 /* we are holding p->pi_lock already */
4749 p
->prio
= rt_mutex_getprio(p
);
4754 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4755 * @p: the task in question.
4756 * @policy: new policy.
4757 * @param: structure containing the new RT priority.
4759 * NOTE that the task may be already dead.
4761 int sched_setscheduler(struct task_struct
*p
, int policy
,
4762 struct sched_param
*param
)
4764 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4765 unsigned long flags
;
4766 const struct sched_class
*prev_class
= p
->sched_class
;
4769 /* may grab non-irq protected spin_locks */
4770 BUG_ON(in_interrupt());
4772 /* double check policy once rq lock held */
4774 policy
= oldpolicy
= p
->policy
;
4775 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4776 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4777 policy
!= SCHED_IDLE
)
4780 * Valid priorities for SCHED_FIFO and SCHED_RR are
4781 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4782 * SCHED_BATCH and SCHED_IDLE is 0.
4784 if (param
->sched_priority
< 0 ||
4785 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4786 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4788 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4792 * Allow unprivileged RT tasks to decrease priority:
4794 if (!capable(CAP_SYS_NICE
)) {
4795 if (rt_policy(policy
)) {
4796 unsigned long rlim_rtprio
;
4798 if (!lock_task_sighand(p
, &flags
))
4800 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4801 unlock_task_sighand(p
, &flags
);
4803 /* can't set/change the rt policy */
4804 if (policy
!= p
->policy
&& !rlim_rtprio
)
4807 /* can't increase priority */
4808 if (param
->sched_priority
> p
->rt_priority
&&
4809 param
->sched_priority
> rlim_rtprio
)
4813 * Like positive nice levels, dont allow tasks to
4814 * move out of SCHED_IDLE either:
4816 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4819 /* can't change other user's priorities */
4820 if ((current
->euid
!= p
->euid
) &&
4821 (current
->euid
!= p
->uid
))
4825 #ifdef CONFIG_RT_GROUP_SCHED
4827 * Do not allow realtime tasks into groups that have no runtime
4830 if (rt_policy(policy
) && task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
4834 retval
= security_task_setscheduler(p
, policy
, param
);
4838 * make sure no PI-waiters arrive (or leave) while we are
4839 * changing the priority of the task:
4841 spin_lock_irqsave(&p
->pi_lock
, flags
);
4843 * To be able to change p->policy safely, the apropriate
4844 * runqueue lock must be held.
4846 rq
= __task_rq_lock(p
);
4847 /* recheck policy now with rq lock held */
4848 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4849 policy
= oldpolicy
= -1;
4850 __task_rq_unlock(rq
);
4851 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4854 update_rq_clock(rq
);
4855 on_rq
= p
->se
.on_rq
;
4856 running
= task_current(rq
, p
);
4858 deactivate_task(rq
, p
, 0);
4860 p
->sched_class
->put_prev_task(rq
, p
);
4863 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4866 p
->sched_class
->set_curr_task(rq
);
4868 activate_task(rq
, p
, 0);
4870 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4872 __task_rq_unlock(rq
);
4873 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4875 rt_mutex_adjust_pi(p
);
4879 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4882 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4884 struct sched_param lparam
;
4885 struct task_struct
*p
;
4888 if (!param
|| pid
< 0)
4890 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4895 p
= find_process_by_pid(pid
);
4897 retval
= sched_setscheduler(p
, policy
, &lparam
);
4904 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4905 * @pid: the pid in question.
4906 * @policy: new policy.
4907 * @param: structure containing the new RT priority.
4910 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4912 /* negative values for policy are not valid */
4916 return do_sched_setscheduler(pid
, policy
, param
);
4920 * sys_sched_setparam - set/change the RT priority of a thread
4921 * @pid: the pid in question.
4922 * @param: structure containing the new RT priority.
4924 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4926 return do_sched_setscheduler(pid
, -1, param
);
4930 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4931 * @pid: the pid in question.
4933 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4935 struct task_struct
*p
;
4942 read_lock(&tasklist_lock
);
4943 p
= find_process_by_pid(pid
);
4945 retval
= security_task_getscheduler(p
);
4949 read_unlock(&tasklist_lock
);
4954 * sys_sched_getscheduler - get the RT priority of a thread
4955 * @pid: the pid in question.
4956 * @param: structure containing the RT priority.
4958 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4960 struct sched_param lp
;
4961 struct task_struct
*p
;
4964 if (!param
|| pid
< 0)
4967 read_lock(&tasklist_lock
);
4968 p
= find_process_by_pid(pid
);
4973 retval
= security_task_getscheduler(p
);
4977 lp
.sched_priority
= p
->rt_priority
;
4978 read_unlock(&tasklist_lock
);
4981 * This one might sleep, we cannot do it with a spinlock held ...
4983 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4988 read_unlock(&tasklist_lock
);
4992 long sched_setaffinity(pid_t pid
, const cpumask_t
*in_mask
)
4994 cpumask_t cpus_allowed
;
4995 cpumask_t new_mask
= *in_mask
;
4996 struct task_struct
*p
;
5000 read_lock(&tasklist_lock
);
5002 p
= find_process_by_pid(pid
);
5004 read_unlock(&tasklist_lock
);
5010 * It is not safe to call set_cpus_allowed with the
5011 * tasklist_lock held. We will bump the task_struct's
5012 * usage count and then drop tasklist_lock.
5015 read_unlock(&tasklist_lock
);
5018 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
5019 !capable(CAP_SYS_NICE
))
5022 retval
= security_task_setscheduler(p
, 0, NULL
);
5026 cpuset_cpus_allowed(p
, &cpus_allowed
);
5027 cpus_and(new_mask
, new_mask
, cpus_allowed
);
5029 retval
= set_cpus_allowed_ptr(p
, &new_mask
);
5032 cpuset_cpus_allowed(p
, &cpus_allowed
);
5033 if (!cpus_subset(new_mask
, cpus_allowed
)) {
5035 * We must have raced with a concurrent cpuset
5036 * update. Just reset the cpus_allowed to the
5037 * cpuset's cpus_allowed
5039 new_mask
= cpus_allowed
;
5049 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5050 cpumask_t
*new_mask
)
5052 if (len
< sizeof(cpumask_t
)) {
5053 memset(new_mask
, 0, sizeof(cpumask_t
));
5054 } else if (len
> sizeof(cpumask_t
)) {
5055 len
= sizeof(cpumask_t
);
5057 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5061 * sys_sched_setaffinity - set the cpu affinity of a process
5062 * @pid: pid of the process
5063 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5064 * @user_mask_ptr: user-space pointer to the new cpu mask
5066 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
5067 unsigned long __user
*user_mask_ptr
)
5072 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
5076 return sched_setaffinity(pid
, &new_mask
);
5080 * Represents all cpu's present in the system
5081 * In systems capable of hotplug, this map could dynamically grow
5082 * as new cpu's are detected in the system via any platform specific
5083 * method, such as ACPI for e.g.
5086 cpumask_t cpu_present_map __read_mostly
;
5087 EXPORT_SYMBOL(cpu_present_map
);
5090 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
5091 EXPORT_SYMBOL(cpu_online_map
);
5093 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
5094 EXPORT_SYMBOL(cpu_possible_map
);
5097 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
5099 struct task_struct
*p
;
5103 read_lock(&tasklist_lock
);
5106 p
= find_process_by_pid(pid
);
5110 retval
= security_task_getscheduler(p
);
5114 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
5117 read_unlock(&tasklist_lock
);
5124 * sys_sched_getaffinity - get the cpu affinity of a process
5125 * @pid: pid of the process
5126 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5127 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5129 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
5130 unsigned long __user
*user_mask_ptr
)
5135 if (len
< sizeof(cpumask_t
))
5138 ret
= sched_getaffinity(pid
, &mask
);
5142 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
5145 return sizeof(cpumask_t
);
5149 * sys_sched_yield - yield the current processor to other threads.
5151 * This function yields the current CPU to other tasks. If there are no
5152 * other threads running on this CPU then this function will return.
5154 asmlinkage
long sys_sched_yield(void)
5156 struct rq
*rq
= this_rq_lock();
5158 schedstat_inc(rq
, yld_count
);
5159 current
->sched_class
->yield_task(rq
);
5162 * Since we are going to call schedule() anyway, there's
5163 * no need to preempt or enable interrupts:
5165 __release(rq
->lock
);
5166 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5167 _raw_spin_unlock(&rq
->lock
);
5168 preempt_enable_no_resched();
5175 static void __cond_resched(void)
5177 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5178 __might_sleep(__FILE__
, __LINE__
);
5181 * The BKS might be reacquired before we have dropped
5182 * PREEMPT_ACTIVE, which could trigger a second
5183 * cond_resched() call.
5186 add_preempt_count(PREEMPT_ACTIVE
);
5188 sub_preempt_count(PREEMPT_ACTIVE
);
5189 } while (need_resched());
5192 int __sched
_cond_resched(void)
5194 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
5195 system_state
== SYSTEM_RUNNING
) {
5201 EXPORT_SYMBOL(_cond_resched
);
5204 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5205 * call schedule, and on return reacquire the lock.
5207 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5208 * operations here to prevent schedule() from being called twice (once via
5209 * spin_unlock(), once by hand).
5211 int cond_resched_lock(spinlock_t
*lock
)
5213 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
5216 if (spin_needbreak(lock
) || resched
) {
5218 if (resched
&& need_resched())
5227 EXPORT_SYMBOL(cond_resched_lock
);
5229 int __sched
cond_resched_softirq(void)
5231 BUG_ON(!in_softirq());
5233 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
5241 EXPORT_SYMBOL(cond_resched_softirq
);
5244 * yield - yield the current processor to other threads.
5246 * This is a shortcut for kernel-space yielding - it marks the
5247 * thread runnable and calls sys_sched_yield().
5249 void __sched
yield(void)
5251 set_current_state(TASK_RUNNING
);
5254 EXPORT_SYMBOL(yield
);
5257 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5258 * that process accounting knows that this is a task in IO wait state.
5260 * But don't do that if it is a deliberate, throttling IO wait (this task
5261 * has set its backing_dev_info: the queue against which it should throttle)
5263 void __sched
io_schedule(void)
5265 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5267 delayacct_blkio_start();
5268 atomic_inc(&rq
->nr_iowait
);
5270 atomic_dec(&rq
->nr_iowait
);
5271 delayacct_blkio_end();
5273 EXPORT_SYMBOL(io_schedule
);
5275 long __sched
io_schedule_timeout(long timeout
)
5277 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5280 delayacct_blkio_start();
5281 atomic_inc(&rq
->nr_iowait
);
5282 ret
= schedule_timeout(timeout
);
5283 atomic_dec(&rq
->nr_iowait
);
5284 delayacct_blkio_end();
5289 * sys_sched_get_priority_max - return maximum RT priority.
5290 * @policy: scheduling class.
5292 * this syscall returns the maximum rt_priority that can be used
5293 * by a given scheduling class.
5295 asmlinkage
long sys_sched_get_priority_max(int policy
)
5302 ret
= MAX_USER_RT_PRIO
-1;
5314 * sys_sched_get_priority_min - return minimum RT priority.
5315 * @policy: scheduling class.
5317 * this syscall returns the minimum rt_priority that can be used
5318 * by a given scheduling class.
5320 asmlinkage
long sys_sched_get_priority_min(int policy
)
5338 * sys_sched_rr_get_interval - return the default timeslice of a process.
5339 * @pid: pid of the process.
5340 * @interval: userspace pointer to the timeslice value.
5342 * this syscall writes the default timeslice value of a given process
5343 * into the user-space timespec buffer. A value of '0' means infinity.
5346 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
5348 struct task_struct
*p
;
5349 unsigned int time_slice
;
5357 read_lock(&tasklist_lock
);
5358 p
= find_process_by_pid(pid
);
5362 retval
= security_task_getscheduler(p
);
5367 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5368 * tasks that are on an otherwise idle runqueue:
5371 if (p
->policy
== SCHED_RR
) {
5372 time_slice
= DEF_TIMESLICE
;
5373 } else if (p
->policy
!= SCHED_FIFO
) {
5374 struct sched_entity
*se
= &p
->se
;
5375 unsigned long flags
;
5378 rq
= task_rq_lock(p
, &flags
);
5379 if (rq
->cfs
.load
.weight
)
5380 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5381 task_rq_unlock(rq
, &flags
);
5383 read_unlock(&tasklist_lock
);
5384 jiffies_to_timespec(time_slice
, &t
);
5385 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5389 read_unlock(&tasklist_lock
);
5393 static const char stat_nam
[] = "RSDTtZX";
5395 void sched_show_task(struct task_struct
*p
)
5397 unsigned long free
= 0;
5400 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5401 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5402 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5403 #if BITS_PER_LONG == 32
5404 if (state
== TASK_RUNNING
)
5405 printk(KERN_CONT
" running ");
5407 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5409 if (state
== TASK_RUNNING
)
5410 printk(KERN_CONT
" running task ");
5412 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5414 #ifdef CONFIG_DEBUG_STACK_USAGE
5416 unsigned long *n
= end_of_stack(p
);
5419 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5422 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5423 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5425 show_stack(p
, NULL
);
5428 void show_state_filter(unsigned long state_filter
)
5430 struct task_struct
*g
, *p
;
5432 #if BITS_PER_LONG == 32
5434 " task PC stack pid father\n");
5437 " task PC stack pid father\n");
5439 read_lock(&tasklist_lock
);
5440 do_each_thread(g
, p
) {
5442 * reset the NMI-timeout, listing all files on a slow
5443 * console might take alot of time:
5445 touch_nmi_watchdog();
5446 if (!state_filter
|| (p
->state
& state_filter
))
5448 } while_each_thread(g
, p
);
5450 touch_all_softlockup_watchdogs();
5452 #ifdef CONFIG_SCHED_DEBUG
5453 sysrq_sched_debug_show();
5455 read_unlock(&tasklist_lock
);
5457 * Only show locks if all tasks are dumped:
5459 if (state_filter
== -1)
5460 debug_show_all_locks();
5463 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5465 idle
->sched_class
= &idle_sched_class
;
5469 * init_idle - set up an idle thread for a given CPU
5470 * @idle: task in question
5471 * @cpu: cpu the idle task belongs to
5473 * NOTE: this function does not set the idle thread's NEED_RESCHED
5474 * flag, to make booting more robust.
5476 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5478 struct rq
*rq
= cpu_rq(cpu
);
5479 unsigned long flags
;
5482 idle
->se
.exec_start
= sched_clock();
5484 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5485 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5486 __set_task_cpu(idle
, cpu
);
5488 spin_lock_irqsave(&rq
->lock
, flags
);
5489 rq
->curr
= rq
->idle
= idle
;
5490 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5493 spin_unlock_irqrestore(&rq
->lock
, flags
);
5495 /* Set the preempt count _outside_ the spinlocks! */
5496 #if defined(CONFIG_PREEMPT)
5497 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5499 task_thread_info(idle
)->preempt_count
= 0;
5502 * The idle tasks have their own, simple scheduling class:
5504 idle
->sched_class
= &idle_sched_class
;
5508 * In a system that switches off the HZ timer nohz_cpu_mask
5509 * indicates which cpus entered this state. This is used
5510 * in the rcu update to wait only for active cpus. For system
5511 * which do not switch off the HZ timer nohz_cpu_mask should
5512 * always be CPU_MASK_NONE.
5514 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5517 * Increase the granularity value when there are more CPUs,
5518 * because with more CPUs the 'effective latency' as visible
5519 * to users decreases. But the relationship is not linear,
5520 * so pick a second-best guess by going with the log2 of the
5523 * This idea comes from the SD scheduler of Con Kolivas:
5525 static inline void sched_init_granularity(void)
5527 unsigned int factor
= 1 + ilog2(num_online_cpus());
5528 const unsigned long limit
= 200000000;
5530 sysctl_sched_min_granularity
*= factor
;
5531 if (sysctl_sched_min_granularity
> limit
)
5532 sysctl_sched_min_granularity
= limit
;
5534 sysctl_sched_latency
*= factor
;
5535 if (sysctl_sched_latency
> limit
)
5536 sysctl_sched_latency
= limit
;
5538 sysctl_sched_wakeup_granularity
*= factor
;
5543 * This is how migration works:
5545 * 1) we queue a struct migration_req structure in the source CPU's
5546 * runqueue and wake up that CPU's migration thread.
5547 * 2) we down() the locked semaphore => thread blocks.
5548 * 3) migration thread wakes up (implicitly it forces the migrated
5549 * thread off the CPU)
5550 * 4) it gets the migration request and checks whether the migrated
5551 * task is still in the wrong runqueue.
5552 * 5) if it's in the wrong runqueue then the migration thread removes
5553 * it and puts it into the right queue.
5554 * 6) migration thread up()s the semaphore.
5555 * 7) we wake up and the migration is done.
5559 * Change a given task's CPU affinity. Migrate the thread to a
5560 * proper CPU and schedule it away if the CPU it's executing on
5561 * is removed from the allowed bitmask.
5563 * NOTE: the caller must have a valid reference to the task, the
5564 * task must not exit() & deallocate itself prematurely. The
5565 * call is not atomic; no spinlocks may be held.
5567 int set_cpus_allowed_ptr(struct task_struct
*p
, const cpumask_t
*new_mask
)
5569 struct migration_req req
;
5570 unsigned long flags
;
5574 rq
= task_rq_lock(p
, &flags
);
5575 if (!cpus_intersects(*new_mask
, cpu_online_map
)) {
5580 if (p
->sched_class
->set_cpus_allowed
)
5581 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5583 p
->cpus_allowed
= *new_mask
;
5584 p
->rt
.nr_cpus_allowed
= cpus_weight(*new_mask
);
5587 /* Can the task run on the task's current CPU? If so, we're done */
5588 if (cpu_isset(task_cpu(p
), *new_mask
))
5591 if (migrate_task(p
, any_online_cpu(*new_mask
), &req
)) {
5592 /* Need help from migration thread: drop lock and wait. */
5593 task_rq_unlock(rq
, &flags
);
5594 wake_up_process(rq
->migration_thread
);
5595 wait_for_completion(&req
.done
);
5596 tlb_migrate_finish(p
->mm
);
5600 task_rq_unlock(rq
, &flags
);
5604 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5607 * Move (not current) task off this cpu, onto dest cpu. We're doing
5608 * this because either it can't run here any more (set_cpus_allowed()
5609 * away from this CPU, or CPU going down), or because we're
5610 * attempting to rebalance this task on exec (sched_exec).
5612 * So we race with normal scheduler movements, but that's OK, as long
5613 * as the task is no longer on this CPU.
5615 * Returns non-zero if task was successfully migrated.
5617 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5619 struct rq
*rq_dest
, *rq_src
;
5622 if (unlikely(cpu_is_offline(dest_cpu
)))
5625 rq_src
= cpu_rq(src_cpu
);
5626 rq_dest
= cpu_rq(dest_cpu
);
5628 double_rq_lock(rq_src
, rq_dest
);
5629 /* Already moved. */
5630 if (task_cpu(p
) != src_cpu
)
5632 /* Affinity changed (again). */
5633 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5636 on_rq
= p
->se
.on_rq
;
5638 deactivate_task(rq_src
, p
, 0);
5640 set_task_cpu(p
, dest_cpu
);
5642 activate_task(rq_dest
, p
, 0);
5643 check_preempt_curr(rq_dest
, p
);
5647 double_rq_unlock(rq_src
, rq_dest
);
5652 * migration_thread - this is a highprio system thread that performs
5653 * thread migration by bumping thread off CPU then 'pushing' onto
5656 static int migration_thread(void *data
)
5658 int cpu
= (long)data
;
5662 BUG_ON(rq
->migration_thread
!= current
);
5664 set_current_state(TASK_INTERRUPTIBLE
);
5665 while (!kthread_should_stop()) {
5666 struct migration_req
*req
;
5667 struct list_head
*head
;
5669 spin_lock_irq(&rq
->lock
);
5671 if (cpu_is_offline(cpu
)) {
5672 spin_unlock_irq(&rq
->lock
);
5676 if (rq
->active_balance
) {
5677 active_load_balance(rq
, cpu
);
5678 rq
->active_balance
= 0;
5681 head
= &rq
->migration_queue
;
5683 if (list_empty(head
)) {
5684 spin_unlock_irq(&rq
->lock
);
5686 set_current_state(TASK_INTERRUPTIBLE
);
5689 req
= list_entry(head
->next
, struct migration_req
, list
);
5690 list_del_init(head
->next
);
5692 spin_unlock(&rq
->lock
);
5693 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5696 complete(&req
->done
);
5698 __set_current_state(TASK_RUNNING
);
5702 /* Wait for kthread_stop */
5703 set_current_state(TASK_INTERRUPTIBLE
);
5704 while (!kthread_should_stop()) {
5706 set_current_state(TASK_INTERRUPTIBLE
);
5708 __set_current_state(TASK_RUNNING
);
5712 #ifdef CONFIG_HOTPLUG_CPU
5714 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5718 local_irq_disable();
5719 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
5725 * Figure out where task on dead CPU should go, use force if necessary.
5726 * NOTE: interrupts should be disabled by the caller
5728 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5730 unsigned long flags
;
5737 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5738 cpus_and(mask
, mask
, p
->cpus_allowed
);
5739 dest_cpu
= any_online_cpu(mask
);
5741 /* On any allowed CPU? */
5742 if (dest_cpu
>= nr_cpu_ids
)
5743 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5745 /* No more Mr. Nice Guy. */
5746 if (dest_cpu
>= nr_cpu_ids
) {
5747 cpumask_t cpus_allowed
;
5749 cpuset_cpus_allowed_locked(p
, &cpus_allowed
);
5751 * Try to stay on the same cpuset, where the
5752 * current cpuset may be a subset of all cpus.
5753 * The cpuset_cpus_allowed_locked() variant of
5754 * cpuset_cpus_allowed() will not block. It must be
5755 * called within calls to cpuset_lock/cpuset_unlock.
5757 rq
= task_rq_lock(p
, &flags
);
5758 p
->cpus_allowed
= cpus_allowed
;
5759 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5760 task_rq_unlock(rq
, &flags
);
5763 * Don't tell them about moving exiting tasks or
5764 * kernel threads (both mm NULL), since they never
5767 if (p
->mm
&& printk_ratelimit()) {
5768 printk(KERN_INFO
"process %d (%s) no "
5769 "longer affine to cpu%d\n",
5770 task_pid_nr(p
), p
->comm
, dead_cpu
);
5773 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
5777 * While a dead CPU has no uninterruptible tasks queued at this point,
5778 * it might still have a nonzero ->nr_uninterruptible counter, because
5779 * for performance reasons the counter is not stricly tracking tasks to
5780 * their home CPUs. So we just add the counter to another CPU's counter,
5781 * to keep the global sum constant after CPU-down:
5783 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5785 struct rq
*rq_dest
= cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR
));
5786 unsigned long flags
;
5788 local_irq_save(flags
);
5789 double_rq_lock(rq_src
, rq_dest
);
5790 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5791 rq_src
->nr_uninterruptible
= 0;
5792 double_rq_unlock(rq_src
, rq_dest
);
5793 local_irq_restore(flags
);
5796 /* Run through task list and migrate tasks from the dead cpu. */
5797 static void migrate_live_tasks(int src_cpu
)
5799 struct task_struct
*p
, *t
;
5801 read_lock(&tasklist_lock
);
5803 do_each_thread(t
, p
) {
5807 if (task_cpu(p
) == src_cpu
)
5808 move_task_off_dead_cpu(src_cpu
, p
);
5809 } while_each_thread(t
, p
);
5811 read_unlock(&tasklist_lock
);
5815 * Schedules idle task to be the next runnable task on current CPU.
5816 * It does so by boosting its priority to highest possible.
5817 * Used by CPU offline code.
5819 void sched_idle_next(void)
5821 int this_cpu
= smp_processor_id();
5822 struct rq
*rq
= cpu_rq(this_cpu
);
5823 struct task_struct
*p
= rq
->idle
;
5824 unsigned long flags
;
5826 /* cpu has to be offline */
5827 BUG_ON(cpu_online(this_cpu
));
5830 * Strictly not necessary since rest of the CPUs are stopped by now
5831 * and interrupts disabled on the current cpu.
5833 spin_lock_irqsave(&rq
->lock
, flags
);
5835 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5837 update_rq_clock(rq
);
5838 activate_task(rq
, p
, 0);
5840 spin_unlock_irqrestore(&rq
->lock
, flags
);
5844 * Ensures that the idle task is using init_mm right before its cpu goes
5847 void idle_task_exit(void)
5849 struct mm_struct
*mm
= current
->active_mm
;
5851 BUG_ON(cpu_online(smp_processor_id()));
5854 switch_mm(mm
, &init_mm
, current
);
5858 /* called under rq->lock with disabled interrupts */
5859 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5861 struct rq
*rq
= cpu_rq(dead_cpu
);
5863 /* Must be exiting, otherwise would be on tasklist. */
5864 BUG_ON(!p
->exit_state
);
5866 /* Cannot have done final schedule yet: would have vanished. */
5867 BUG_ON(p
->state
== TASK_DEAD
);
5872 * Drop lock around migration; if someone else moves it,
5873 * that's OK. No task can be added to this CPU, so iteration is
5876 spin_unlock_irq(&rq
->lock
);
5877 move_task_off_dead_cpu(dead_cpu
, p
);
5878 spin_lock_irq(&rq
->lock
);
5883 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5884 static void migrate_dead_tasks(unsigned int dead_cpu
)
5886 struct rq
*rq
= cpu_rq(dead_cpu
);
5887 struct task_struct
*next
;
5890 if (!rq
->nr_running
)
5892 update_rq_clock(rq
);
5893 next
= pick_next_task(rq
, rq
->curr
);
5896 migrate_dead(dead_cpu
, next
);
5900 #endif /* CONFIG_HOTPLUG_CPU */
5902 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5904 static struct ctl_table sd_ctl_dir
[] = {
5906 .procname
= "sched_domain",
5912 static struct ctl_table sd_ctl_root
[] = {
5914 .ctl_name
= CTL_KERN
,
5915 .procname
= "kernel",
5917 .child
= sd_ctl_dir
,
5922 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5924 struct ctl_table
*entry
=
5925 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5930 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5932 struct ctl_table
*entry
;
5935 * In the intermediate directories, both the child directory and
5936 * procname are dynamically allocated and could fail but the mode
5937 * will always be set. In the lowest directory the names are
5938 * static strings and all have proc handlers.
5940 for (entry
= *tablep
; entry
->mode
; entry
++) {
5942 sd_free_ctl_entry(&entry
->child
);
5943 if (entry
->proc_handler
== NULL
)
5944 kfree(entry
->procname
);
5952 set_table_entry(struct ctl_table
*entry
,
5953 const char *procname
, void *data
, int maxlen
,
5954 mode_t mode
, proc_handler
*proc_handler
)
5956 entry
->procname
= procname
;
5958 entry
->maxlen
= maxlen
;
5960 entry
->proc_handler
= proc_handler
;
5963 static struct ctl_table
*
5964 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5966 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
5971 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5972 sizeof(long), 0644, proc_doulongvec_minmax
);
5973 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5974 sizeof(long), 0644, proc_doulongvec_minmax
);
5975 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5976 sizeof(int), 0644, proc_dointvec_minmax
);
5977 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5978 sizeof(int), 0644, proc_dointvec_minmax
);
5979 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5980 sizeof(int), 0644, proc_dointvec_minmax
);
5981 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5982 sizeof(int), 0644, proc_dointvec_minmax
);
5983 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5984 sizeof(int), 0644, proc_dointvec_minmax
);
5985 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5986 sizeof(int), 0644, proc_dointvec_minmax
);
5987 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5988 sizeof(int), 0644, proc_dointvec_minmax
);
5989 set_table_entry(&table
[9], "cache_nice_tries",
5990 &sd
->cache_nice_tries
,
5991 sizeof(int), 0644, proc_dointvec_minmax
);
5992 set_table_entry(&table
[10], "flags", &sd
->flags
,
5993 sizeof(int), 0644, proc_dointvec_minmax
);
5994 /* &table[11] is terminator */
5999 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6001 struct ctl_table
*entry
, *table
;
6002 struct sched_domain
*sd
;
6003 int domain_num
= 0, i
;
6006 for_each_domain(cpu
, sd
)
6008 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6013 for_each_domain(cpu
, sd
) {
6014 snprintf(buf
, 32, "domain%d", i
);
6015 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6017 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6024 static struct ctl_table_header
*sd_sysctl_header
;
6025 static void register_sched_domain_sysctl(void)
6027 int i
, cpu_num
= num_online_cpus();
6028 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6031 WARN_ON(sd_ctl_dir
[0].child
);
6032 sd_ctl_dir
[0].child
= entry
;
6037 for_each_online_cpu(i
) {
6038 snprintf(buf
, 32, "cpu%d", i
);
6039 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6041 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6045 WARN_ON(sd_sysctl_header
);
6046 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6049 /* may be called multiple times per register */
6050 static void unregister_sched_domain_sysctl(void)
6052 if (sd_sysctl_header
)
6053 unregister_sysctl_table(sd_sysctl_header
);
6054 sd_sysctl_header
= NULL
;
6055 if (sd_ctl_dir
[0].child
)
6056 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6059 static void register_sched_domain_sysctl(void)
6062 static void unregister_sched_domain_sysctl(void)
6068 * migration_call - callback that gets triggered when a CPU is added.
6069 * Here we can start up the necessary migration thread for the new CPU.
6071 static int __cpuinit
6072 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6074 struct task_struct
*p
;
6075 int cpu
= (long)hcpu
;
6076 unsigned long flags
;
6081 case CPU_UP_PREPARE
:
6082 case CPU_UP_PREPARE_FROZEN
:
6083 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
6086 kthread_bind(p
, cpu
);
6087 /* Must be high prio: stop_machine expects to yield to it. */
6088 rq
= task_rq_lock(p
, &flags
);
6089 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6090 task_rq_unlock(rq
, &flags
);
6091 cpu_rq(cpu
)->migration_thread
= p
;
6095 case CPU_ONLINE_FROZEN
:
6096 /* Strictly unnecessary, as first user will wake it. */
6097 wake_up_process(cpu_rq(cpu
)->migration_thread
);
6099 /* Update our root-domain */
6101 spin_lock_irqsave(&rq
->lock
, flags
);
6103 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6104 cpu_set(cpu
, rq
->rd
->online
);
6106 spin_unlock_irqrestore(&rq
->lock
, flags
);
6109 #ifdef CONFIG_HOTPLUG_CPU
6110 case CPU_UP_CANCELED
:
6111 case CPU_UP_CANCELED_FROZEN
:
6112 if (!cpu_rq(cpu
)->migration_thread
)
6114 /* Unbind it from offline cpu so it can run. Fall thru. */
6115 kthread_bind(cpu_rq(cpu
)->migration_thread
,
6116 any_online_cpu(cpu_online_map
));
6117 kthread_stop(cpu_rq(cpu
)->migration_thread
);
6118 cpu_rq(cpu
)->migration_thread
= NULL
;
6122 case CPU_DEAD_FROZEN
:
6123 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6124 migrate_live_tasks(cpu
);
6126 kthread_stop(rq
->migration_thread
);
6127 rq
->migration_thread
= NULL
;
6128 /* Idle task back to normal (off runqueue, low prio) */
6129 spin_lock_irq(&rq
->lock
);
6130 update_rq_clock(rq
);
6131 deactivate_task(rq
, rq
->idle
, 0);
6132 rq
->idle
->static_prio
= MAX_PRIO
;
6133 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
6134 rq
->idle
->sched_class
= &idle_sched_class
;
6135 migrate_dead_tasks(cpu
);
6136 spin_unlock_irq(&rq
->lock
);
6138 migrate_nr_uninterruptible(rq
);
6139 BUG_ON(rq
->nr_running
!= 0);
6142 * No need to migrate the tasks: it was best-effort if
6143 * they didn't take sched_hotcpu_mutex. Just wake up
6146 spin_lock_irq(&rq
->lock
);
6147 while (!list_empty(&rq
->migration_queue
)) {
6148 struct migration_req
*req
;
6150 req
= list_entry(rq
->migration_queue
.next
,
6151 struct migration_req
, list
);
6152 list_del_init(&req
->list
);
6153 complete(&req
->done
);
6155 spin_unlock_irq(&rq
->lock
);
6159 case CPU_DYING_FROZEN
:
6160 /* Update our root-domain */
6162 spin_lock_irqsave(&rq
->lock
, flags
);
6164 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6165 cpu_clear(cpu
, rq
->rd
->online
);
6167 spin_unlock_irqrestore(&rq
->lock
, flags
);
6174 /* Register at highest priority so that task migration (migrate_all_tasks)
6175 * happens before everything else.
6177 static struct notifier_block __cpuinitdata migration_notifier
= {
6178 .notifier_call
= migration_call
,
6182 void __init
migration_init(void)
6184 void *cpu
= (void *)(long)smp_processor_id();
6187 /* Start one for the boot CPU: */
6188 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6189 BUG_ON(err
== NOTIFY_BAD
);
6190 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6191 register_cpu_notifier(&migration_notifier
);
6197 #ifdef CONFIG_SCHED_DEBUG
6199 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6200 cpumask_t
*groupmask
)
6202 struct sched_group
*group
= sd
->groups
;
6205 cpulist_scnprintf(str
, sizeof(str
), sd
->span
);
6206 cpus_clear(*groupmask
);
6208 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6210 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6211 printk("does not load-balance\n");
6213 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6218 printk(KERN_CONT
"span %s\n", str
);
6220 if (!cpu_isset(cpu
, sd
->span
)) {
6221 printk(KERN_ERR
"ERROR: domain->span does not contain "
6224 if (!cpu_isset(cpu
, group
->cpumask
)) {
6225 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6229 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6233 printk(KERN_ERR
"ERROR: group is NULL\n");
6237 if (!group
->__cpu_power
) {
6238 printk(KERN_CONT
"\n");
6239 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6244 if (!cpus_weight(group
->cpumask
)) {
6245 printk(KERN_CONT
"\n");
6246 printk(KERN_ERR
"ERROR: empty group\n");
6250 if (cpus_intersects(*groupmask
, group
->cpumask
)) {
6251 printk(KERN_CONT
"\n");
6252 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6256 cpus_or(*groupmask
, *groupmask
, group
->cpumask
);
6258 cpulist_scnprintf(str
, sizeof(str
), group
->cpumask
);
6259 printk(KERN_CONT
" %s", str
);
6261 group
= group
->next
;
6262 } while (group
!= sd
->groups
);
6263 printk(KERN_CONT
"\n");
6265 if (!cpus_equal(sd
->span
, *groupmask
))
6266 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6268 if (sd
->parent
&& !cpus_subset(*groupmask
, sd
->parent
->span
))
6269 printk(KERN_ERR
"ERROR: parent span is not a superset "
6270 "of domain->span\n");
6274 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6276 cpumask_t
*groupmask
;
6280 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6284 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6286 groupmask
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6288 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6293 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6303 # define sched_domain_debug(sd, cpu) do { } while (0)
6306 static int sd_degenerate(struct sched_domain
*sd
)
6308 if (cpus_weight(sd
->span
) == 1)
6311 /* Following flags need at least 2 groups */
6312 if (sd
->flags
& (SD_LOAD_BALANCE
|
6313 SD_BALANCE_NEWIDLE
|
6317 SD_SHARE_PKG_RESOURCES
)) {
6318 if (sd
->groups
!= sd
->groups
->next
)
6322 /* Following flags don't use groups */
6323 if (sd
->flags
& (SD_WAKE_IDLE
|
6332 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6334 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6336 if (sd_degenerate(parent
))
6339 if (!cpus_equal(sd
->span
, parent
->span
))
6342 /* Does parent contain flags not in child? */
6343 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6344 if (cflags
& SD_WAKE_AFFINE
)
6345 pflags
&= ~SD_WAKE_BALANCE
;
6346 /* Flags needing groups don't count if only 1 group in parent */
6347 if (parent
->groups
== parent
->groups
->next
) {
6348 pflags
&= ~(SD_LOAD_BALANCE
|
6349 SD_BALANCE_NEWIDLE
|
6353 SD_SHARE_PKG_RESOURCES
);
6355 if (~cflags
& pflags
)
6361 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6363 unsigned long flags
;
6364 const struct sched_class
*class;
6366 spin_lock_irqsave(&rq
->lock
, flags
);
6369 struct root_domain
*old_rd
= rq
->rd
;
6371 for (class = sched_class_highest
; class; class = class->next
) {
6372 if (class->leave_domain
)
6373 class->leave_domain(rq
);
6376 cpu_clear(rq
->cpu
, old_rd
->span
);
6377 cpu_clear(rq
->cpu
, old_rd
->online
);
6379 if (atomic_dec_and_test(&old_rd
->refcount
))
6383 atomic_inc(&rd
->refcount
);
6386 cpu_set(rq
->cpu
, rd
->span
);
6387 if (cpu_isset(rq
->cpu
, cpu_online_map
))
6388 cpu_set(rq
->cpu
, rd
->online
);
6390 for (class = sched_class_highest
; class; class = class->next
) {
6391 if (class->join_domain
)
6392 class->join_domain(rq
);
6395 spin_unlock_irqrestore(&rq
->lock
, flags
);
6398 static void init_rootdomain(struct root_domain
*rd
)
6400 memset(rd
, 0, sizeof(*rd
));
6402 cpus_clear(rd
->span
);
6403 cpus_clear(rd
->online
);
6405 cpupri_init(&rd
->cpupri
);
6408 static void init_defrootdomain(void)
6410 init_rootdomain(&def_root_domain
);
6411 atomic_set(&def_root_domain
.refcount
, 1);
6414 static struct root_domain
*alloc_rootdomain(void)
6416 struct root_domain
*rd
;
6418 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6422 init_rootdomain(rd
);
6428 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6429 * hold the hotplug lock.
6432 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6434 struct rq
*rq
= cpu_rq(cpu
);
6435 struct sched_domain
*tmp
;
6437 /* Remove the sched domains which do not contribute to scheduling. */
6438 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
6439 struct sched_domain
*parent
= tmp
->parent
;
6442 if (sd_parent_degenerate(tmp
, parent
)) {
6443 tmp
->parent
= parent
->parent
;
6445 parent
->parent
->child
= tmp
;
6449 if (sd
&& sd_degenerate(sd
)) {
6455 sched_domain_debug(sd
, cpu
);
6457 rq_attach_root(rq
, rd
);
6458 rcu_assign_pointer(rq
->sd
, sd
);
6461 /* cpus with isolated domains */
6462 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
6464 /* Setup the mask of cpus configured for isolated domains */
6465 static int __init
isolated_cpu_setup(char *str
)
6467 int ints
[NR_CPUS
], i
;
6469 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
6470 cpus_clear(cpu_isolated_map
);
6471 for (i
= 1; i
<= ints
[0]; i
++)
6472 if (ints
[i
] < NR_CPUS
)
6473 cpu_set(ints
[i
], cpu_isolated_map
);
6477 __setup("isolcpus=", isolated_cpu_setup
);
6480 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6481 * to a function which identifies what group(along with sched group) a CPU
6482 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6483 * (due to the fact that we keep track of groups covered with a cpumask_t).
6485 * init_sched_build_groups will build a circular linked list of the groups
6486 * covered by the given span, and will set each group's ->cpumask correctly,
6487 * and ->cpu_power to 0.
6490 init_sched_build_groups(const cpumask_t
*span
, const cpumask_t
*cpu_map
,
6491 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
6492 struct sched_group
**sg
,
6493 cpumask_t
*tmpmask
),
6494 cpumask_t
*covered
, cpumask_t
*tmpmask
)
6496 struct sched_group
*first
= NULL
, *last
= NULL
;
6499 cpus_clear(*covered
);
6501 for_each_cpu_mask(i
, *span
) {
6502 struct sched_group
*sg
;
6503 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6506 if (cpu_isset(i
, *covered
))
6509 cpus_clear(sg
->cpumask
);
6510 sg
->__cpu_power
= 0;
6512 for_each_cpu_mask(j
, *span
) {
6513 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6516 cpu_set(j
, *covered
);
6517 cpu_set(j
, sg
->cpumask
);
6528 #define SD_NODES_PER_DOMAIN 16
6533 * find_next_best_node - find the next node to include in a sched_domain
6534 * @node: node whose sched_domain we're building
6535 * @used_nodes: nodes already in the sched_domain
6537 * Find the next node to include in a given scheduling domain. Simply
6538 * finds the closest node not already in the @used_nodes map.
6540 * Should use nodemask_t.
6542 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6544 int i
, n
, val
, min_val
, best_node
= 0;
6548 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6549 /* Start at @node */
6550 n
= (node
+ i
) % MAX_NUMNODES
;
6552 if (!nr_cpus_node(n
))
6555 /* Skip already used nodes */
6556 if (node_isset(n
, *used_nodes
))
6559 /* Simple min distance search */
6560 val
= node_distance(node
, n
);
6562 if (val
< min_val
) {
6568 node_set(best_node
, *used_nodes
);
6573 * sched_domain_node_span - get a cpumask for a node's sched_domain
6574 * @node: node whose cpumask we're constructing
6575 * @span: resulting cpumask
6577 * Given a node, construct a good cpumask for its sched_domain to span. It
6578 * should be one that prevents unnecessary balancing, but also spreads tasks
6581 static void sched_domain_node_span(int node
, cpumask_t
*span
)
6583 nodemask_t used_nodes
;
6584 node_to_cpumask_ptr(nodemask
, node
);
6588 nodes_clear(used_nodes
);
6590 cpus_or(*span
, *span
, *nodemask
);
6591 node_set(node
, used_nodes
);
6593 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6594 int next_node
= find_next_best_node(node
, &used_nodes
);
6596 node_to_cpumask_ptr_next(nodemask
, next_node
);
6597 cpus_or(*span
, *span
, *nodemask
);
6602 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6605 * SMT sched-domains:
6607 #ifdef CONFIG_SCHED_SMT
6608 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6609 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
6612 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6616 *sg
= &per_cpu(sched_group_cpus
, cpu
);
6622 * multi-core sched-domains:
6624 #ifdef CONFIG_SCHED_MC
6625 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6626 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
6629 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6631 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6636 *mask
= per_cpu(cpu_sibling_map
, cpu
);
6637 cpus_and(*mask
, *mask
, *cpu_map
);
6638 group
= first_cpu(*mask
);
6640 *sg
= &per_cpu(sched_group_core
, group
);
6643 #elif defined(CONFIG_SCHED_MC)
6645 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6649 *sg
= &per_cpu(sched_group_core
, cpu
);
6654 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6655 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
6658 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6662 #ifdef CONFIG_SCHED_MC
6663 *mask
= cpu_coregroup_map(cpu
);
6664 cpus_and(*mask
, *mask
, *cpu_map
);
6665 group
= first_cpu(*mask
);
6666 #elif defined(CONFIG_SCHED_SMT)
6667 *mask
= per_cpu(cpu_sibling_map
, cpu
);
6668 cpus_and(*mask
, *mask
, *cpu_map
);
6669 group
= first_cpu(*mask
);
6674 *sg
= &per_cpu(sched_group_phys
, group
);
6680 * The init_sched_build_groups can't handle what we want to do with node
6681 * groups, so roll our own. Now each node has its own list of groups which
6682 * gets dynamically allocated.
6684 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6685 static struct sched_group
***sched_group_nodes_bycpu
;
6687 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6688 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
6690 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
6691 struct sched_group
**sg
, cpumask_t
*nodemask
)
6695 *nodemask
= node_to_cpumask(cpu_to_node(cpu
));
6696 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
6697 group
= first_cpu(*nodemask
);
6700 *sg
= &per_cpu(sched_group_allnodes
, group
);
6704 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6706 struct sched_group
*sg
= group_head
;
6712 for_each_cpu_mask(j
, sg
->cpumask
) {
6713 struct sched_domain
*sd
;
6715 sd
= &per_cpu(phys_domains
, j
);
6716 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6718 * Only add "power" once for each
6724 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
6727 } while (sg
!= group_head
);
6732 /* Free memory allocated for various sched_group structures */
6733 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
6737 for_each_cpu_mask(cpu
, *cpu_map
) {
6738 struct sched_group
**sched_group_nodes
6739 = sched_group_nodes_bycpu
[cpu
];
6741 if (!sched_group_nodes
)
6744 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6745 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6747 *nodemask
= node_to_cpumask(i
);
6748 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
6749 if (cpus_empty(*nodemask
))
6759 if (oldsg
!= sched_group_nodes
[i
])
6762 kfree(sched_group_nodes
);
6763 sched_group_nodes_bycpu
[cpu
] = NULL
;
6767 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
6773 * Initialize sched groups cpu_power.
6775 * cpu_power indicates the capacity of sched group, which is used while
6776 * distributing the load between different sched groups in a sched domain.
6777 * Typically cpu_power for all the groups in a sched domain will be same unless
6778 * there are asymmetries in the topology. If there are asymmetries, group
6779 * having more cpu_power will pickup more load compared to the group having
6782 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6783 * the maximum number of tasks a group can handle in the presence of other idle
6784 * or lightly loaded groups in the same sched domain.
6786 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6788 struct sched_domain
*child
;
6789 struct sched_group
*group
;
6791 WARN_ON(!sd
|| !sd
->groups
);
6793 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6798 sd
->groups
->__cpu_power
= 0;
6801 * For perf policy, if the groups in child domain share resources
6802 * (for example cores sharing some portions of the cache hierarchy
6803 * or SMT), then set this domain groups cpu_power such that each group
6804 * can handle only one task, when there are other idle groups in the
6805 * same sched domain.
6807 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6809 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6810 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6815 * add cpu_power of each child group to this groups cpu_power
6817 group
= child
->groups
;
6819 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6820 group
= group
->next
;
6821 } while (group
!= child
->groups
);
6825 * Initializers for schedule domains
6826 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6829 #define SD_INIT(sd, type) sd_init_##type(sd)
6830 #define SD_INIT_FUNC(type) \
6831 static noinline void sd_init_##type(struct sched_domain *sd) \
6833 memset(sd, 0, sizeof(*sd)); \
6834 *sd = SD_##type##_INIT; \
6835 sd->level = SD_LV_##type; \
6840 SD_INIT_FUNC(ALLNODES
)
6843 #ifdef CONFIG_SCHED_SMT
6844 SD_INIT_FUNC(SIBLING
)
6846 #ifdef CONFIG_SCHED_MC
6851 * To minimize stack usage kmalloc room for cpumasks and share the
6852 * space as the usage in build_sched_domains() dictates. Used only
6853 * if the amount of space is significant.
6856 cpumask_t tmpmask
; /* make this one first */
6859 cpumask_t this_sibling_map
;
6860 cpumask_t this_core_map
;
6862 cpumask_t send_covered
;
6865 cpumask_t domainspan
;
6867 cpumask_t notcovered
;
6872 #define SCHED_CPUMASK_ALLOC 1
6873 #define SCHED_CPUMASK_FREE(v) kfree(v)
6874 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
6876 #define SCHED_CPUMASK_ALLOC 0
6877 #define SCHED_CPUMASK_FREE(v)
6878 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
6881 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
6882 ((unsigned long)(a) + offsetof(struct allmasks, v))
6884 static int default_relax_domain_level
= -1;
6886 static int __init
setup_relax_domain_level(char *str
)
6888 default_relax_domain_level
= simple_strtoul(str
, NULL
, 0);
6891 __setup("relax_domain_level=", setup_relax_domain_level
);
6893 static void set_domain_attribute(struct sched_domain
*sd
,
6894 struct sched_domain_attr
*attr
)
6898 if (!attr
|| attr
->relax_domain_level
< 0) {
6899 if (default_relax_domain_level
< 0)
6902 request
= default_relax_domain_level
;
6904 request
= attr
->relax_domain_level
;
6905 if (request
< sd
->level
) {
6906 /* turn off idle balance on this domain */
6907 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
6909 /* turn on idle balance on this domain */
6910 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
6915 * Build sched domains for a given set of cpus and attach the sched domains
6916 * to the individual cpus
6918 static int __build_sched_domains(const cpumask_t
*cpu_map
,
6919 struct sched_domain_attr
*attr
)
6922 struct root_domain
*rd
;
6923 SCHED_CPUMASK_DECLARE(allmasks
);
6926 struct sched_group
**sched_group_nodes
= NULL
;
6927 int sd_allnodes
= 0;
6930 * Allocate the per-node list of sched groups
6932 sched_group_nodes
= kcalloc(MAX_NUMNODES
, sizeof(struct sched_group
*),
6934 if (!sched_group_nodes
) {
6935 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6940 rd
= alloc_rootdomain();
6942 printk(KERN_WARNING
"Cannot alloc root domain\n");
6944 kfree(sched_group_nodes
);
6949 #if SCHED_CPUMASK_ALLOC
6950 /* get space for all scratch cpumask variables */
6951 allmasks
= kmalloc(sizeof(*allmasks
), GFP_KERNEL
);
6953 printk(KERN_WARNING
"Cannot alloc cpumask array\n");
6956 kfree(sched_group_nodes
);
6961 tmpmask
= (cpumask_t
*)allmasks
;
6965 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6969 * Set up domains for cpus specified by the cpu_map.
6971 for_each_cpu_mask(i
, *cpu_map
) {
6972 struct sched_domain
*sd
= NULL
, *p
;
6973 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
6975 *nodemask
= node_to_cpumask(cpu_to_node(i
));
6976 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
6979 if (cpus_weight(*cpu_map
) >
6980 SD_NODES_PER_DOMAIN
*cpus_weight(*nodemask
)) {
6981 sd
= &per_cpu(allnodes_domains
, i
);
6982 SD_INIT(sd
, ALLNODES
);
6983 set_domain_attribute(sd
, attr
);
6984 sd
->span
= *cpu_map
;
6985 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
6991 sd
= &per_cpu(node_domains
, i
);
6993 set_domain_attribute(sd
, attr
);
6994 sched_domain_node_span(cpu_to_node(i
), &sd
->span
);
6998 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7002 sd
= &per_cpu(phys_domains
, i
);
7004 set_domain_attribute(sd
, attr
);
7005 sd
->span
= *nodemask
;
7009 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7011 #ifdef CONFIG_SCHED_MC
7013 sd
= &per_cpu(core_domains
, i
);
7015 set_domain_attribute(sd
, attr
);
7016 sd
->span
= cpu_coregroup_map(i
);
7017 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7020 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7023 #ifdef CONFIG_SCHED_SMT
7025 sd
= &per_cpu(cpu_domains
, i
);
7026 SD_INIT(sd
, SIBLING
);
7027 set_domain_attribute(sd
, attr
);
7028 sd
->span
= per_cpu(cpu_sibling_map
, i
);
7029 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7032 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7036 #ifdef CONFIG_SCHED_SMT
7037 /* Set up CPU (sibling) groups */
7038 for_each_cpu_mask(i
, *cpu_map
) {
7039 SCHED_CPUMASK_VAR(this_sibling_map
, allmasks
);
7040 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7042 *this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
7043 cpus_and(*this_sibling_map
, *this_sibling_map
, *cpu_map
);
7044 if (i
!= first_cpu(*this_sibling_map
))
7047 init_sched_build_groups(this_sibling_map
, cpu_map
,
7049 send_covered
, tmpmask
);
7053 #ifdef CONFIG_SCHED_MC
7054 /* Set up multi-core groups */
7055 for_each_cpu_mask(i
, *cpu_map
) {
7056 SCHED_CPUMASK_VAR(this_core_map
, allmasks
);
7057 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7059 *this_core_map
= cpu_coregroup_map(i
);
7060 cpus_and(*this_core_map
, *this_core_map
, *cpu_map
);
7061 if (i
!= first_cpu(*this_core_map
))
7064 init_sched_build_groups(this_core_map
, cpu_map
,
7066 send_covered
, tmpmask
);
7070 /* Set up physical groups */
7071 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7072 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7073 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7075 *nodemask
= node_to_cpumask(i
);
7076 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7077 if (cpus_empty(*nodemask
))
7080 init_sched_build_groups(nodemask
, cpu_map
,
7082 send_covered
, tmpmask
);
7086 /* Set up node groups */
7088 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7090 init_sched_build_groups(cpu_map
, cpu_map
,
7091 &cpu_to_allnodes_group
,
7092 send_covered
, tmpmask
);
7095 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7096 /* Set up node groups */
7097 struct sched_group
*sg
, *prev
;
7098 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7099 SCHED_CPUMASK_VAR(domainspan
, allmasks
);
7100 SCHED_CPUMASK_VAR(covered
, allmasks
);
7103 *nodemask
= node_to_cpumask(i
);
7104 cpus_clear(*covered
);
7106 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7107 if (cpus_empty(*nodemask
)) {
7108 sched_group_nodes
[i
] = NULL
;
7112 sched_domain_node_span(i
, domainspan
);
7113 cpus_and(*domainspan
, *domainspan
, *cpu_map
);
7115 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
7117 printk(KERN_WARNING
"Can not alloc domain group for "
7121 sched_group_nodes
[i
] = sg
;
7122 for_each_cpu_mask(j
, *nodemask
) {
7123 struct sched_domain
*sd
;
7125 sd
= &per_cpu(node_domains
, j
);
7128 sg
->__cpu_power
= 0;
7129 sg
->cpumask
= *nodemask
;
7131 cpus_or(*covered
, *covered
, *nodemask
);
7134 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
7135 SCHED_CPUMASK_VAR(notcovered
, allmasks
);
7136 int n
= (i
+ j
) % MAX_NUMNODES
;
7137 node_to_cpumask_ptr(pnodemask
, n
);
7139 cpus_complement(*notcovered
, *covered
);
7140 cpus_and(*tmpmask
, *notcovered
, *cpu_map
);
7141 cpus_and(*tmpmask
, *tmpmask
, *domainspan
);
7142 if (cpus_empty(*tmpmask
))
7145 cpus_and(*tmpmask
, *tmpmask
, *pnodemask
);
7146 if (cpus_empty(*tmpmask
))
7149 sg
= kmalloc_node(sizeof(struct sched_group
),
7153 "Can not alloc domain group for node %d\n", j
);
7156 sg
->__cpu_power
= 0;
7157 sg
->cpumask
= *tmpmask
;
7158 sg
->next
= prev
->next
;
7159 cpus_or(*covered
, *covered
, *tmpmask
);
7166 /* Calculate CPU power for physical packages and nodes */
7167 #ifdef CONFIG_SCHED_SMT
7168 for_each_cpu_mask(i
, *cpu_map
) {
7169 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
7171 init_sched_groups_power(i
, sd
);
7174 #ifdef CONFIG_SCHED_MC
7175 for_each_cpu_mask(i
, *cpu_map
) {
7176 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
7178 init_sched_groups_power(i
, sd
);
7182 for_each_cpu_mask(i
, *cpu_map
) {
7183 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
7185 init_sched_groups_power(i
, sd
);
7189 for (i
= 0; i
< MAX_NUMNODES
; i
++)
7190 init_numa_sched_groups_power(sched_group_nodes
[i
]);
7193 struct sched_group
*sg
;
7195 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
,
7197 init_numa_sched_groups_power(sg
);
7201 /* Attach the domains */
7202 for_each_cpu_mask(i
, *cpu_map
) {
7203 struct sched_domain
*sd
;
7204 #ifdef CONFIG_SCHED_SMT
7205 sd
= &per_cpu(cpu_domains
, i
);
7206 #elif defined(CONFIG_SCHED_MC)
7207 sd
= &per_cpu(core_domains
, i
);
7209 sd
= &per_cpu(phys_domains
, i
);
7211 cpu_attach_domain(sd
, rd
, i
);
7214 SCHED_CPUMASK_FREE((void *)allmasks
);
7219 free_sched_groups(cpu_map
, tmpmask
);
7220 SCHED_CPUMASK_FREE((void *)allmasks
);
7225 static int build_sched_domains(const cpumask_t
*cpu_map
)
7227 return __build_sched_domains(cpu_map
, NULL
);
7230 static cpumask_t
*doms_cur
; /* current sched domains */
7231 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7232 static struct sched_domain_attr
*dattr_cur
;
7233 /* attribues of custom domains in 'doms_cur' */
7236 * Special case: If a kmalloc of a doms_cur partition (array of
7237 * cpumask_t) fails, then fallback to a single sched domain,
7238 * as determined by the single cpumask_t fallback_doms.
7240 static cpumask_t fallback_doms
;
7242 void __attribute__((weak
)) arch_update_cpu_topology(void)
7247 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7248 * For now this just excludes isolated cpus, but could be used to
7249 * exclude other special cases in the future.
7251 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
7255 arch_update_cpu_topology();
7257 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
7259 doms_cur
= &fallback_doms
;
7260 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
7262 err
= build_sched_domains(doms_cur
);
7263 register_sched_domain_sysctl();
7268 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
,
7271 free_sched_groups(cpu_map
, tmpmask
);
7275 * Detach sched domains from a group of cpus specified in cpu_map
7276 * These cpus will now be attached to the NULL domain
7278 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
7283 unregister_sched_domain_sysctl();
7285 for_each_cpu_mask(i
, *cpu_map
)
7286 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7287 synchronize_sched();
7288 arch_destroy_sched_domains(cpu_map
, &tmpmask
);
7291 /* handle null as "default" */
7292 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7293 struct sched_domain_attr
*new, int idx_new
)
7295 struct sched_domain_attr tmp
;
7302 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7303 new ? (new + idx_new
) : &tmp
,
7304 sizeof(struct sched_domain_attr
));
7308 * Partition sched domains as specified by the 'ndoms_new'
7309 * cpumasks in the array doms_new[] of cpumasks. This compares
7310 * doms_new[] to the current sched domain partitioning, doms_cur[].
7311 * It destroys each deleted domain and builds each new domain.
7313 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7314 * The masks don't intersect (don't overlap.) We should setup one
7315 * sched domain for each mask. CPUs not in any of the cpumasks will
7316 * not be load balanced. If the same cpumask appears both in the
7317 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7320 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7321 * ownership of it and will kfree it when done with it. If the caller
7322 * failed the kmalloc call, then it can pass in doms_new == NULL,
7323 * and partition_sched_domains() will fallback to the single partition
7326 * Call with hotplug lock held
7328 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
,
7329 struct sched_domain_attr
*dattr_new
)
7333 mutex_lock(&sched_domains_mutex
);
7335 /* always unregister in case we don't destroy any domains */
7336 unregister_sched_domain_sysctl();
7338 if (doms_new
== NULL
) {
7340 doms_new
= &fallback_doms
;
7341 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
7345 /* Destroy deleted domains */
7346 for (i
= 0; i
< ndoms_cur
; i
++) {
7347 for (j
= 0; j
< ndoms_new
; j
++) {
7348 if (cpus_equal(doms_cur
[i
], doms_new
[j
])
7349 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7352 /* no match - a current sched domain not in new doms_new[] */
7353 detach_destroy_domains(doms_cur
+ i
);
7358 /* Build new domains */
7359 for (i
= 0; i
< ndoms_new
; i
++) {
7360 for (j
= 0; j
< ndoms_cur
; j
++) {
7361 if (cpus_equal(doms_new
[i
], doms_cur
[j
])
7362 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7365 /* no match - add a new doms_new */
7366 __build_sched_domains(doms_new
+ i
,
7367 dattr_new
? dattr_new
+ i
: NULL
);
7372 /* Remember the new sched domains */
7373 if (doms_cur
!= &fallback_doms
)
7375 kfree(dattr_cur
); /* kfree(NULL) is safe */
7376 doms_cur
= doms_new
;
7377 dattr_cur
= dattr_new
;
7378 ndoms_cur
= ndoms_new
;
7380 register_sched_domain_sysctl();
7382 mutex_unlock(&sched_domains_mutex
);
7385 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7386 int arch_reinit_sched_domains(void)
7391 mutex_lock(&sched_domains_mutex
);
7392 detach_destroy_domains(&cpu_online_map
);
7393 err
= arch_init_sched_domains(&cpu_online_map
);
7394 mutex_unlock(&sched_domains_mutex
);
7400 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7404 if (buf
[0] != '0' && buf
[0] != '1')
7408 sched_smt_power_savings
= (buf
[0] == '1');
7410 sched_mc_power_savings
= (buf
[0] == '1');
7412 ret
= arch_reinit_sched_domains();
7414 return ret
? ret
: count
;
7417 #ifdef CONFIG_SCHED_MC
7418 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
7420 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7422 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
7423 const char *buf
, size_t count
)
7425 return sched_power_savings_store(buf
, count
, 0);
7427 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
7428 sched_mc_power_savings_store
);
7431 #ifdef CONFIG_SCHED_SMT
7432 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
7434 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7436 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
7437 const char *buf
, size_t count
)
7439 return sched_power_savings_store(buf
, count
, 1);
7441 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
7442 sched_smt_power_savings_store
);
7445 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7449 #ifdef CONFIG_SCHED_SMT
7451 err
= sysfs_create_file(&cls
->kset
.kobj
,
7452 &attr_sched_smt_power_savings
.attr
);
7454 #ifdef CONFIG_SCHED_MC
7455 if (!err
&& mc_capable())
7456 err
= sysfs_create_file(&cls
->kset
.kobj
,
7457 &attr_sched_mc_power_savings
.attr
);
7464 * Force a reinitialization of the sched domains hierarchy. The domains
7465 * and groups cannot be updated in place without racing with the balancing
7466 * code, so we temporarily attach all running cpus to the NULL domain
7467 * which will prevent rebalancing while the sched domains are recalculated.
7469 static int update_sched_domains(struct notifier_block
*nfb
,
7470 unsigned long action
, void *hcpu
)
7473 case CPU_UP_PREPARE
:
7474 case CPU_UP_PREPARE_FROZEN
:
7475 case CPU_DOWN_PREPARE
:
7476 case CPU_DOWN_PREPARE_FROZEN
:
7477 detach_destroy_domains(&cpu_online_map
);
7480 case CPU_UP_CANCELED
:
7481 case CPU_UP_CANCELED_FROZEN
:
7482 case CPU_DOWN_FAILED
:
7483 case CPU_DOWN_FAILED_FROZEN
:
7485 case CPU_ONLINE_FROZEN
:
7487 case CPU_DEAD_FROZEN
:
7489 * Fall through and re-initialise the domains.
7496 /* The hotplug lock is already held by cpu_up/cpu_down */
7497 arch_init_sched_domains(&cpu_online_map
);
7502 void __init
sched_init_smp(void)
7504 cpumask_t non_isolated_cpus
;
7506 #if defined(CONFIG_NUMA)
7507 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7509 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7512 mutex_lock(&sched_domains_mutex
);
7513 arch_init_sched_domains(&cpu_online_map
);
7514 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
7515 if (cpus_empty(non_isolated_cpus
))
7516 cpu_set(smp_processor_id(), non_isolated_cpus
);
7517 mutex_unlock(&sched_domains_mutex
);
7519 /* XXX: Theoretical race here - CPU may be hotplugged now */
7520 hotcpu_notifier(update_sched_domains
, 0);
7523 /* Move init over to a non-isolated CPU */
7524 if (set_cpus_allowed_ptr(current
, &non_isolated_cpus
) < 0)
7526 sched_init_granularity();
7529 void __init
sched_init_smp(void)
7531 sched_init_granularity();
7533 #endif /* CONFIG_SMP */
7535 int in_sched_functions(unsigned long addr
)
7537 return in_lock_functions(addr
) ||
7538 (addr
>= (unsigned long)__sched_text_start
7539 && addr
< (unsigned long)__sched_text_end
);
7542 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7544 cfs_rq
->tasks_timeline
= RB_ROOT
;
7545 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7546 #ifdef CONFIG_FAIR_GROUP_SCHED
7549 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7552 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7554 struct rt_prio_array
*array
;
7557 array
= &rt_rq
->active
;
7558 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7559 INIT_LIST_HEAD(array
->xqueue
+ i
);
7560 INIT_LIST_HEAD(array
->squeue
+ i
);
7561 __clear_bit(i
, array
->bitmap
);
7563 /* delimiter for bitsearch: */
7564 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7566 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7567 rt_rq
->highest_prio
= MAX_RT_PRIO
;
7570 rt_rq
->rt_nr_migratory
= 0;
7571 rt_rq
->overloaded
= 0;
7575 rt_rq
->rt_throttled
= 0;
7576 rt_rq
->rt_runtime
= 0;
7577 spin_lock_init(&rt_rq
->rt_runtime_lock
);
7579 #ifdef CONFIG_RT_GROUP_SCHED
7580 rt_rq
->rt_nr_boosted
= 0;
7585 #ifdef CONFIG_FAIR_GROUP_SCHED
7586 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7587 struct sched_entity
*se
, int cpu
, int add
,
7588 struct sched_entity
*parent
)
7590 struct rq
*rq
= cpu_rq(cpu
);
7591 tg
->cfs_rq
[cpu
] = cfs_rq
;
7592 init_cfs_rq(cfs_rq
, rq
);
7595 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7598 /* se could be NULL for init_task_group */
7603 se
->cfs_rq
= &rq
->cfs
;
7605 se
->cfs_rq
= parent
->my_q
;
7608 se
->load
.weight
= tg
->shares
;
7609 se
->load
.inv_weight
= 0;
7610 se
->parent
= parent
;
7614 #ifdef CONFIG_RT_GROUP_SCHED
7615 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
7616 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
7617 struct sched_rt_entity
*parent
)
7619 struct rq
*rq
= cpu_rq(cpu
);
7621 tg
->rt_rq
[cpu
] = rt_rq
;
7622 init_rt_rq(rt_rq
, rq
);
7624 rt_rq
->rt_se
= rt_se
;
7625 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7627 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7629 tg
->rt_se
[cpu
] = rt_se
;
7634 rt_se
->rt_rq
= &rq
->rt
;
7636 rt_se
->rt_rq
= parent
->my_q
;
7638 rt_se
->rt_rq
= &rq
->rt
;
7639 rt_se
->my_q
= rt_rq
;
7640 rt_se
->parent
= parent
;
7641 INIT_LIST_HEAD(&rt_se
->run_list
);
7645 void __init
sched_init(void)
7648 unsigned long alloc_size
= 0, ptr
;
7650 #ifdef CONFIG_FAIR_GROUP_SCHED
7651 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7653 #ifdef CONFIG_RT_GROUP_SCHED
7654 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7656 #ifdef CONFIG_USER_SCHED
7660 * As sched_init() is called before page_alloc is setup,
7661 * we use alloc_bootmem().
7664 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
7666 #ifdef CONFIG_FAIR_GROUP_SCHED
7667 init_task_group
.se
= (struct sched_entity
**)ptr
;
7668 ptr
+= nr_cpu_ids
* sizeof(void **);
7670 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7671 ptr
+= nr_cpu_ids
* sizeof(void **);
7673 #ifdef CONFIG_USER_SCHED
7674 root_task_group
.se
= (struct sched_entity
**)ptr
;
7675 ptr
+= nr_cpu_ids
* sizeof(void **);
7677 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7678 ptr
+= nr_cpu_ids
* sizeof(void **);
7681 #ifdef CONFIG_RT_GROUP_SCHED
7682 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7683 ptr
+= nr_cpu_ids
* sizeof(void **);
7685 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7686 ptr
+= nr_cpu_ids
* sizeof(void **);
7688 #ifdef CONFIG_USER_SCHED
7689 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7690 ptr
+= nr_cpu_ids
* sizeof(void **);
7692 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7693 ptr
+= nr_cpu_ids
* sizeof(void **);
7699 init_defrootdomain();
7702 init_rt_bandwidth(&def_rt_bandwidth
,
7703 global_rt_period(), global_rt_runtime());
7705 #ifdef CONFIG_RT_GROUP_SCHED
7706 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
7707 global_rt_period(), global_rt_runtime());
7708 #ifdef CONFIG_USER_SCHED
7709 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7710 global_rt_period(), RUNTIME_INF
);
7714 #ifdef CONFIG_GROUP_SCHED
7715 list_add(&init_task_group
.list
, &task_groups
);
7716 INIT_LIST_HEAD(&init_task_group
.children
);
7718 #ifdef CONFIG_USER_SCHED
7719 INIT_LIST_HEAD(&root_task_group
.children
);
7720 init_task_group
.parent
= &root_task_group
;
7721 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
7725 for_each_possible_cpu(i
) {
7729 spin_lock_init(&rq
->lock
);
7730 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
7732 init_cfs_rq(&rq
->cfs
, rq
);
7733 init_rt_rq(&rq
->rt
, rq
);
7734 #ifdef CONFIG_FAIR_GROUP_SCHED
7735 init_task_group
.shares
= init_task_group_load
;
7736 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7737 #ifdef CONFIG_CGROUP_SCHED
7739 * How much cpu bandwidth does init_task_group get?
7741 * In case of task-groups formed thr' the cgroup filesystem, it
7742 * gets 100% of the cpu resources in the system. This overall
7743 * system cpu resource is divided among the tasks of
7744 * init_task_group and its child task-groups in a fair manner,
7745 * based on each entity's (task or task-group's) weight
7746 * (se->load.weight).
7748 * In other words, if init_task_group has 10 tasks of weight
7749 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7750 * then A0's share of the cpu resource is:
7752 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7754 * We achieve this by letting init_task_group's tasks sit
7755 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7757 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
7758 #elif defined CONFIG_USER_SCHED
7759 root_task_group
.shares
= NICE_0_LOAD
;
7760 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
7762 * In case of task-groups formed thr' the user id of tasks,
7763 * init_task_group represents tasks belonging to root user.
7764 * Hence it forms a sibling of all subsequent groups formed.
7765 * In this case, init_task_group gets only a fraction of overall
7766 * system cpu resource, based on the weight assigned to root
7767 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
7768 * by letting tasks of init_task_group sit in a separate cfs_rq
7769 * (init_cfs_rq) and having one entity represent this group of
7770 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
7772 init_tg_cfs_entry(&init_task_group
,
7773 &per_cpu(init_cfs_rq
, i
),
7774 &per_cpu(init_sched_entity
, i
), i
, 1,
7775 root_task_group
.se
[i
]);
7778 #endif /* CONFIG_FAIR_GROUP_SCHED */
7780 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7781 #ifdef CONFIG_RT_GROUP_SCHED
7782 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7783 #ifdef CONFIG_CGROUP_SCHED
7784 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
7785 #elif defined CONFIG_USER_SCHED
7786 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
7787 init_tg_rt_entry(&init_task_group
,
7788 &per_cpu(init_rt_rq
, i
),
7789 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
7790 root_task_group
.rt_se
[i
]);
7794 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7795 rq
->cpu_load
[j
] = 0;
7799 rq
->active_balance
= 0;
7800 rq
->next_balance
= jiffies
;
7803 rq
->migration_thread
= NULL
;
7804 INIT_LIST_HEAD(&rq
->migration_queue
);
7805 rq_attach_root(rq
, &def_root_domain
);
7808 atomic_set(&rq
->nr_iowait
, 0);
7811 set_load_weight(&init_task
);
7813 #ifdef CONFIG_PREEMPT_NOTIFIERS
7814 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7818 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
7821 #ifdef CONFIG_RT_MUTEXES
7822 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7826 * The boot idle thread does lazy MMU switching as well:
7828 atomic_inc(&init_mm
.mm_count
);
7829 enter_lazy_tlb(&init_mm
, current
);
7832 * Make us the idle thread. Technically, schedule() should not be
7833 * called from this thread, however somewhere below it might be,
7834 * but because we are the idle thread, we just pick up running again
7835 * when this runqueue becomes "idle".
7837 init_idle(current
, smp_processor_id());
7839 * During early bootup we pretend to be a normal task:
7841 current
->sched_class
= &fair_sched_class
;
7843 scheduler_running
= 1;
7846 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7847 void __might_sleep(char *file
, int line
)
7850 static unsigned long prev_jiffy
; /* ratelimiting */
7852 if ((in_atomic() || irqs_disabled()) &&
7853 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
7854 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7856 prev_jiffy
= jiffies
;
7857 printk(KERN_ERR
"BUG: sleeping function called from invalid"
7858 " context at %s:%d\n", file
, line
);
7859 printk("in_atomic():%d, irqs_disabled():%d\n",
7860 in_atomic(), irqs_disabled());
7861 debug_show_held_locks(current
);
7862 if (irqs_disabled())
7863 print_irqtrace_events(current
);
7868 EXPORT_SYMBOL(__might_sleep
);
7871 #ifdef CONFIG_MAGIC_SYSRQ
7872 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7876 update_rq_clock(rq
);
7877 on_rq
= p
->se
.on_rq
;
7879 deactivate_task(rq
, p
, 0);
7880 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7882 activate_task(rq
, p
, 0);
7883 resched_task(rq
->curr
);
7887 void normalize_rt_tasks(void)
7889 struct task_struct
*g
, *p
;
7890 unsigned long flags
;
7893 read_lock_irqsave(&tasklist_lock
, flags
);
7894 do_each_thread(g
, p
) {
7896 * Only normalize user tasks:
7901 p
->se
.exec_start
= 0;
7902 #ifdef CONFIG_SCHEDSTATS
7903 p
->se
.wait_start
= 0;
7904 p
->se
.sleep_start
= 0;
7905 p
->se
.block_start
= 0;
7910 * Renice negative nice level userspace
7913 if (TASK_NICE(p
) < 0 && p
->mm
)
7914 set_user_nice(p
, 0);
7918 spin_lock(&p
->pi_lock
);
7919 rq
= __task_rq_lock(p
);
7921 normalize_task(rq
, p
);
7923 __task_rq_unlock(rq
);
7924 spin_unlock(&p
->pi_lock
);
7925 } while_each_thread(g
, p
);
7927 read_unlock_irqrestore(&tasklist_lock
, flags
);
7930 #endif /* CONFIG_MAGIC_SYSRQ */
7934 * These functions are only useful for the IA64 MCA handling.
7936 * They can only be called when the whole system has been
7937 * stopped - every CPU needs to be quiescent, and no scheduling
7938 * activity can take place. Using them for anything else would
7939 * be a serious bug, and as a result, they aren't even visible
7940 * under any other configuration.
7944 * curr_task - return the current task for a given cpu.
7945 * @cpu: the processor in question.
7947 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7949 struct task_struct
*curr_task(int cpu
)
7951 return cpu_curr(cpu
);
7955 * set_curr_task - set the current task for a given cpu.
7956 * @cpu: the processor in question.
7957 * @p: the task pointer to set.
7959 * Description: This function must only be used when non-maskable interrupts
7960 * are serviced on a separate stack. It allows the architecture to switch the
7961 * notion of the current task on a cpu in a non-blocking manner. This function
7962 * must be called with all CPU's synchronized, and interrupts disabled, the
7963 * and caller must save the original value of the current task (see
7964 * curr_task() above) and restore that value before reenabling interrupts and
7965 * re-starting the system.
7967 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7969 void set_curr_task(int cpu
, struct task_struct
*p
)
7976 #ifdef CONFIG_FAIR_GROUP_SCHED
7977 static void free_fair_sched_group(struct task_group
*tg
)
7981 for_each_possible_cpu(i
) {
7983 kfree(tg
->cfs_rq
[i
]);
7993 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7995 struct cfs_rq
*cfs_rq
;
7996 struct sched_entity
*se
, *parent_se
;
8000 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8003 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8007 tg
->shares
= NICE_0_LOAD
;
8009 for_each_possible_cpu(i
) {
8012 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
),
8013 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8017 se
= kmalloc_node(sizeof(struct sched_entity
),
8018 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8022 parent_se
= parent
? parent
->se
[i
] : NULL
;
8023 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent_se
);
8032 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8034 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8035 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8038 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8040 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8043 static inline void free_fair_sched_group(struct task_group
*tg
)
8048 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8053 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8057 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8062 #ifdef CONFIG_RT_GROUP_SCHED
8063 static void free_rt_sched_group(struct task_group
*tg
)
8067 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8069 for_each_possible_cpu(i
) {
8071 kfree(tg
->rt_rq
[i
]);
8073 kfree(tg
->rt_se
[i
]);
8081 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8083 struct rt_rq
*rt_rq
;
8084 struct sched_rt_entity
*rt_se
, *parent_se
;
8088 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8091 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8095 init_rt_bandwidth(&tg
->rt_bandwidth
,
8096 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8098 for_each_possible_cpu(i
) {
8101 rt_rq
= kmalloc_node(sizeof(struct rt_rq
),
8102 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8106 rt_se
= kmalloc_node(sizeof(struct sched_rt_entity
),
8107 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8111 parent_se
= parent
? parent
->rt_se
[i
] : NULL
;
8112 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent_se
);
8121 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8123 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8124 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8127 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8129 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8132 static inline void free_rt_sched_group(struct task_group
*tg
)
8137 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8142 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8146 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8151 #ifdef CONFIG_GROUP_SCHED
8152 static void free_sched_group(struct task_group
*tg
)
8154 free_fair_sched_group(tg
);
8155 free_rt_sched_group(tg
);
8159 /* allocate runqueue etc for a new task group */
8160 struct task_group
*sched_create_group(struct task_group
*parent
)
8162 struct task_group
*tg
;
8163 unsigned long flags
;
8166 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8168 return ERR_PTR(-ENOMEM
);
8170 if (!alloc_fair_sched_group(tg
, parent
))
8173 if (!alloc_rt_sched_group(tg
, parent
))
8176 spin_lock_irqsave(&task_group_lock
, flags
);
8177 for_each_possible_cpu(i
) {
8178 register_fair_sched_group(tg
, i
);
8179 register_rt_sched_group(tg
, i
);
8181 list_add_rcu(&tg
->list
, &task_groups
);
8183 WARN_ON(!parent
); /* root should already exist */
8185 tg
->parent
= parent
;
8186 list_add_rcu(&tg
->siblings
, &parent
->children
);
8187 INIT_LIST_HEAD(&tg
->children
);
8188 spin_unlock_irqrestore(&task_group_lock
, flags
);
8193 free_sched_group(tg
);
8194 return ERR_PTR(-ENOMEM
);
8197 /* rcu callback to free various structures associated with a task group */
8198 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8200 /* now it should be safe to free those cfs_rqs */
8201 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8204 /* Destroy runqueue etc associated with a task group */
8205 void sched_destroy_group(struct task_group
*tg
)
8207 unsigned long flags
;
8210 spin_lock_irqsave(&task_group_lock
, flags
);
8211 for_each_possible_cpu(i
) {
8212 unregister_fair_sched_group(tg
, i
);
8213 unregister_rt_sched_group(tg
, i
);
8215 list_del_rcu(&tg
->list
);
8216 list_del_rcu(&tg
->siblings
);
8217 spin_unlock_irqrestore(&task_group_lock
, flags
);
8219 /* wait for possible concurrent references to cfs_rqs complete */
8220 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8223 /* change task's runqueue when it moves between groups.
8224 * The caller of this function should have put the task in its new group
8225 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8226 * reflect its new group.
8228 void sched_move_task(struct task_struct
*tsk
)
8231 unsigned long flags
;
8234 rq
= task_rq_lock(tsk
, &flags
);
8236 update_rq_clock(rq
);
8238 running
= task_current(rq
, tsk
);
8239 on_rq
= tsk
->se
.on_rq
;
8242 dequeue_task(rq
, tsk
, 0);
8243 if (unlikely(running
))
8244 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8246 set_task_rq(tsk
, task_cpu(tsk
));
8248 #ifdef CONFIG_FAIR_GROUP_SCHED
8249 if (tsk
->sched_class
->moved_group
)
8250 tsk
->sched_class
->moved_group(tsk
);
8253 if (unlikely(running
))
8254 tsk
->sched_class
->set_curr_task(rq
);
8256 enqueue_task(rq
, tsk
, 0);
8258 task_rq_unlock(rq
, &flags
);
8262 #ifdef CONFIG_FAIR_GROUP_SCHED
8263 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8265 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8266 struct rq
*rq
= cfs_rq
->rq
;
8269 spin_lock_irq(&rq
->lock
);
8273 dequeue_entity(cfs_rq
, se
, 0);
8275 se
->load
.weight
= shares
;
8276 se
->load
.inv_weight
= 0;
8279 enqueue_entity(cfs_rq
, se
, 0);
8281 spin_unlock_irq(&rq
->lock
);
8284 static DEFINE_MUTEX(shares_mutex
);
8286 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8289 unsigned long flags
;
8292 * We can't change the weight of the root cgroup.
8297 if (shares
< MIN_SHARES
)
8298 shares
= MIN_SHARES
;
8299 else if (shares
> MAX_SHARES
)
8300 shares
= MAX_SHARES
;
8302 mutex_lock(&shares_mutex
);
8303 if (tg
->shares
== shares
)
8306 spin_lock_irqsave(&task_group_lock
, flags
);
8307 for_each_possible_cpu(i
)
8308 unregister_fair_sched_group(tg
, i
);
8309 list_del_rcu(&tg
->siblings
);
8310 spin_unlock_irqrestore(&task_group_lock
, flags
);
8312 /* wait for any ongoing reference to this group to finish */
8313 synchronize_sched();
8316 * Now we are free to modify the group's share on each cpu
8317 * w/o tripping rebalance_share or load_balance_fair.
8319 tg
->shares
= shares
;
8320 for_each_possible_cpu(i
)
8321 set_se_shares(tg
->se
[i
], shares
);
8324 * Enable load balance activity on this group, by inserting it back on
8325 * each cpu's rq->leaf_cfs_rq_list.
8327 spin_lock_irqsave(&task_group_lock
, flags
);
8328 for_each_possible_cpu(i
)
8329 register_fair_sched_group(tg
, i
);
8330 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8331 spin_unlock_irqrestore(&task_group_lock
, flags
);
8333 mutex_unlock(&shares_mutex
);
8337 unsigned long sched_group_shares(struct task_group
*tg
)
8343 #ifdef CONFIG_RT_GROUP_SCHED
8345 * Ensure that the real time constraints are schedulable.
8347 static DEFINE_MUTEX(rt_constraints_mutex
);
8349 static unsigned long to_ratio(u64 period
, u64 runtime
)
8351 if (runtime
== RUNTIME_INF
)
8354 return div64_u64(runtime
<< 16, period
);
8357 #ifdef CONFIG_CGROUP_SCHED
8358 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8360 struct task_group
*tgi
, *parent
= tg
->parent
;
8361 unsigned long total
= 0;
8364 if (global_rt_period() < period
)
8367 return to_ratio(period
, runtime
) <
8368 to_ratio(global_rt_period(), global_rt_runtime());
8371 if (ktime_to_ns(parent
->rt_bandwidth
.rt_period
) < period
)
8375 list_for_each_entry_rcu(tgi
, &parent
->children
, siblings
) {
8379 total
+= to_ratio(ktime_to_ns(tgi
->rt_bandwidth
.rt_period
),
8380 tgi
->rt_bandwidth
.rt_runtime
);
8384 return total
+ to_ratio(period
, runtime
) <
8385 to_ratio(ktime_to_ns(parent
->rt_bandwidth
.rt_period
),
8386 parent
->rt_bandwidth
.rt_runtime
);
8388 #elif defined CONFIG_USER_SCHED
8389 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8391 struct task_group
*tgi
;
8392 unsigned long total
= 0;
8393 unsigned long global_ratio
=
8394 to_ratio(global_rt_period(), global_rt_runtime());
8397 list_for_each_entry_rcu(tgi
, &task_groups
, list
) {
8401 total
+= to_ratio(ktime_to_ns(tgi
->rt_bandwidth
.rt_period
),
8402 tgi
->rt_bandwidth
.rt_runtime
);
8406 return total
+ to_ratio(period
, runtime
) < global_ratio
;
8410 /* Must be called with tasklist_lock held */
8411 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8413 struct task_struct
*g
, *p
;
8414 do_each_thread(g
, p
) {
8415 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8417 } while_each_thread(g
, p
);
8421 static int tg_set_bandwidth(struct task_group
*tg
,
8422 u64 rt_period
, u64 rt_runtime
)
8426 mutex_lock(&rt_constraints_mutex
);
8427 read_lock(&tasklist_lock
);
8428 if (rt_runtime
== 0 && tg_has_rt_tasks(tg
)) {
8432 if (!__rt_schedulable(tg
, rt_period
, rt_runtime
)) {
8437 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8438 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8439 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8441 for_each_possible_cpu(i
) {
8442 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8444 spin_lock(&rt_rq
->rt_runtime_lock
);
8445 rt_rq
->rt_runtime
= rt_runtime
;
8446 spin_unlock(&rt_rq
->rt_runtime_lock
);
8448 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8450 read_unlock(&tasklist_lock
);
8451 mutex_unlock(&rt_constraints_mutex
);
8456 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8458 u64 rt_runtime
, rt_period
;
8460 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8461 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8462 if (rt_runtime_us
< 0)
8463 rt_runtime
= RUNTIME_INF
;
8465 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8468 long sched_group_rt_runtime(struct task_group
*tg
)
8472 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8475 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8476 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8477 return rt_runtime_us
;
8480 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8482 u64 rt_runtime
, rt_period
;
8484 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8485 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8487 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8490 long sched_group_rt_period(struct task_group
*tg
)
8494 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8495 do_div(rt_period_us
, NSEC_PER_USEC
);
8496 return rt_period_us
;
8499 static int sched_rt_global_constraints(void)
8503 mutex_lock(&rt_constraints_mutex
);
8504 if (!__rt_schedulable(NULL
, 1, 0))
8506 mutex_unlock(&rt_constraints_mutex
);
8511 static int sched_rt_global_constraints(void)
8513 unsigned long flags
;
8516 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8517 for_each_possible_cpu(i
) {
8518 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8520 spin_lock(&rt_rq
->rt_runtime_lock
);
8521 rt_rq
->rt_runtime
= global_rt_runtime();
8522 spin_unlock(&rt_rq
->rt_runtime_lock
);
8524 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8530 int sched_rt_handler(struct ctl_table
*table
, int write
,
8531 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
8535 int old_period
, old_runtime
;
8536 static DEFINE_MUTEX(mutex
);
8539 old_period
= sysctl_sched_rt_period
;
8540 old_runtime
= sysctl_sched_rt_runtime
;
8542 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
8544 if (!ret
&& write
) {
8545 ret
= sched_rt_global_constraints();
8547 sysctl_sched_rt_period
= old_period
;
8548 sysctl_sched_rt_runtime
= old_runtime
;
8550 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8551 def_rt_bandwidth
.rt_period
=
8552 ns_to_ktime(global_rt_period());
8555 mutex_unlock(&mutex
);
8560 #ifdef CONFIG_CGROUP_SCHED
8562 /* return corresponding task_group object of a cgroup */
8563 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
8565 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
8566 struct task_group
, css
);
8569 static struct cgroup_subsys_state
*
8570 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8572 struct task_group
*tg
, *parent
;
8574 if (!cgrp
->parent
) {
8575 /* This is early initialization for the top cgroup */
8576 init_task_group
.css
.cgroup
= cgrp
;
8577 return &init_task_group
.css
;
8580 parent
= cgroup_tg(cgrp
->parent
);
8581 tg
= sched_create_group(parent
);
8583 return ERR_PTR(-ENOMEM
);
8585 /* Bind the cgroup to task_group object we just created */
8586 tg
->css
.cgroup
= cgrp
;
8592 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8594 struct task_group
*tg
= cgroup_tg(cgrp
);
8596 sched_destroy_group(tg
);
8600 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8601 struct task_struct
*tsk
)
8603 #ifdef CONFIG_RT_GROUP_SCHED
8604 /* Don't accept realtime tasks when there is no way for them to run */
8605 if (rt_task(tsk
) && cgroup_tg(cgrp
)->rt_bandwidth
.rt_runtime
== 0)
8608 /* We don't support RT-tasks being in separate groups */
8609 if (tsk
->sched_class
!= &fair_sched_class
)
8617 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8618 struct cgroup
*old_cont
, struct task_struct
*tsk
)
8620 sched_move_task(tsk
);
8623 #ifdef CONFIG_FAIR_GROUP_SCHED
8624 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
8627 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
8630 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
8632 struct task_group
*tg
= cgroup_tg(cgrp
);
8634 return (u64
) tg
->shares
;
8638 #ifdef CONFIG_RT_GROUP_SCHED
8639 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8642 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8645 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8647 return sched_group_rt_runtime(cgroup_tg(cgrp
));
8650 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8653 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8656 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8658 return sched_group_rt_period(cgroup_tg(cgrp
));
8662 static struct cftype cpu_files
[] = {
8663 #ifdef CONFIG_FAIR_GROUP_SCHED
8666 .read_u64
= cpu_shares_read_u64
,
8667 .write_u64
= cpu_shares_write_u64
,
8670 #ifdef CONFIG_RT_GROUP_SCHED
8672 .name
= "rt_runtime_us",
8673 .read_s64
= cpu_rt_runtime_read
,
8674 .write_s64
= cpu_rt_runtime_write
,
8677 .name
= "rt_period_us",
8678 .read_u64
= cpu_rt_period_read_uint
,
8679 .write_u64
= cpu_rt_period_write_uint
,
8684 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8686 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
8689 struct cgroup_subsys cpu_cgroup_subsys
= {
8691 .create
= cpu_cgroup_create
,
8692 .destroy
= cpu_cgroup_destroy
,
8693 .can_attach
= cpu_cgroup_can_attach
,
8694 .attach
= cpu_cgroup_attach
,
8695 .populate
= cpu_cgroup_populate
,
8696 .subsys_id
= cpu_cgroup_subsys_id
,
8700 #endif /* CONFIG_CGROUP_SCHED */
8702 #ifdef CONFIG_CGROUP_CPUACCT
8705 * CPU accounting code for task groups.
8707 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8708 * (balbir@in.ibm.com).
8711 /* track cpu usage of a group of tasks */
8713 struct cgroup_subsys_state css
;
8714 /* cpuusage holds pointer to a u64-type object on every cpu */
8718 struct cgroup_subsys cpuacct_subsys
;
8720 /* return cpu accounting group corresponding to this container */
8721 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
8723 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
8724 struct cpuacct
, css
);
8727 /* return cpu accounting group to which this task belongs */
8728 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
8730 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
8731 struct cpuacct
, css
);
8734 /* create a new cpu accounting group */
8735 static struct cgroup_subsys_state
*cpuacct_create(
8736 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8738 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8741 return ERR_PTR(-ENOMEM
);
8743 ca
->cpuusage
= alloc_percpu(u64
);
8744 if (!ca
->cpuusage
) {
8746 return ERR_PTR(-ENOMEM
);
8752 /* destroy an existing cpu accounting group */
8754 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8756 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8758 free_percpu(ca
->cpuusage
);
8762 /* return total cpu usage (in nanoseconds) of a group */
8763 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8765 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8766 u64 totalcpuusage
= 0;
8769 for_each_possible_cpu(i
) {
8770 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
8773 * Take rq->lock to make 64-bit addition safe on 32-bit
8776 spin_lock_irq(&cpu_rq(i
)->lock
);
8777 totalcpuusage
+= *cpuusage
;
8778 spin_unlock_irq(&cpu_rq(i
)->lock
);
8781 return totalcpuusage
;
8784 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8787 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8796 for_each_possible_cpu(i
) {
8797 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
8799 spin_lock_irq(&cpu_rq(i
)->lock
);
8801 spin_unlock_irq(&cpu_rq(i
)->lock
);
8807 static struct cftype files
[] = {
8810 .read_u64
= cpuusage_read
,
8811 .write_u64
= cpuusage_write
,
8815 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8817 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
8821 * charge this task's execution time to its accounting group.
8823 * called with rq->lock held.
8825 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8829 if (!cpuacct_subsys
.active
)
8834 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
8836 *cpuusage
+= cputime
;
8840 struct cgroup_subsys cpuacct_subsys
= {
8842 .create
= cpuacct_create
,
8843 .destroy
= cpuacct_destroy
,
8844 .populate
= cpuacct_populate
,
8845 .subsys_id
= cpuacct_subsys_id
,
8847 #endif /* CONFIG_CGROUP_CPUACCT */