4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/reciprocal_div.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/bootmem.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
75 #include <trace/sched.h>
78 #include <asm/irq_regs.h>
80 #include "sched_cpupri.h"
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
87 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
96 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
101 * Helpers for converting nanosecond timing to jiffy resolution
103 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * single value that denotes runtime == period, ie unlimited time.
119 #define RUNTIME_INF ((u64)~0ULL)
121 DEFINE_TRACE(sched_wait_task
);
122 DEFINE_TRACE(sched_wakeup
);
123 DEFINE_TRACE(sched_wakeup_new
);
124 DEFINE_TRACE(sched_switch
);
125 DEFINE_TRACE(sched_migrate_task
);
129 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
132 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
133 * Since cpu_power is a 'constant', we can use a reciprocal divide.
135 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
137 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
141 * Each time a sched group cpu_power is changed,
142 * we must compute its reciprocal value
144 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
146 sg
->__cpu_power
+= val
;
147 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
151 static inline int rt_policy(int policy
)
153 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
158 static inline int task_has_rt_policy(struct task_struct
*p
)
160 return rt_policy(p
->policy
);
164 * This is the priority-queue data structure of the RT scheduling class:
166 struct rt_prio_array
{
167 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
168 struct list_head queue
[MAX_RT_PRIO
];
171 struct rt_bandwidth
{
172 /* nests inside the rq lock: */
173 spinlock_t rt_runtime_lock
;
176 struct hrtimer rt_period_timer
;
179 static struct rt_bandwidth def_rt_bandwidth
;
181 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
183 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
185 struct rt_bandwidth
*rt_b
=
186 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
192 now
= hrtimer_cb_get_time(timer
);
193 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
198 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
201 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
205 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
207 rt_b
->rt_period
= ns_to_ktime(period
);
208 rt_b
->rt_runtime
= runtime
;
210 spin_lock_init(&rt_b
->rt_runtime_lock
);
212 hrtimer_init(&rt_b
->rt_period_timer
,
213 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
214 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
217 static inline int rt_bandwidth_enabled(void)
219 return sysctl_sched_rt_runtime
>= 0;
222 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
226 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
229 if (hrtimer_active(&rt_b
->rt_period_timer
))
232 spin_lock(&rt_b
->rt_runtime_lock
);
237 if (hrtimer_active(&rt_b
->rt_period_timer
))
240 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
241 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
243 soft
= hrtimer_get_softexpires(&rt_b
->rt_period_timer
);
244 hard
= hrtimer_get_expires(&rt_b
->rt_period_timer
);
245 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
246 __hrtimer_start_range_ns(&rt_b
->rt_period_timer
, soft
, delta
,
247 HRTIMER_MODE_ABS
, 0);
249 spin_unlock(&rt_b
->rt_runtime_lock
);
252 #ifdef CONFIG_RT_GROUP_SCHED
253 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
255 hrtimer_cancel(&rt_b
->rt_period_timer
);
260 * sched_domains_mutex serializes calls to arch_init_sched_domains,
261 * detach_destroy_domains and partition_sched_domains.
263 static DEFINE_MUTEX(sched_domains_mutex
);
265 #ifdef CONFIG_GROUP_SCHED
267 #include <linux/cgroup.h>
271 static LIST_HEAD(task_groups
);
273 /* task group related information */
275 #ifdef CONFIG_CGROUP_SCHED
276 struct cgroup_subsys_state css
;
279 #ifdef CONFIG_USER_SCHED
283 #ifdef CONFIG_FAIR_GROUP_SCHED
284 /* schedulable entities of this group on each cpu */
285 struct sched_entity
**se
;
286 /* runqueue "owned" by this group on each cpu */
287 struct cfs_rq
**cfs_rq
;
288 unsigned long shares
;
291 #ifdef CONFIG_RT_GROUP_SCHED
292 struct sched_rt_entity
**rt_se
;
293 struct rt_rq
**rt_rq
;
295 struct rt_bandwidth rt_bandwidth
;
299 struct list_head list
;
301 struct task_group
*parent
;
302 struct list_head siblings
;
303 struct list_head children
;
306 #ifdef CONFIG_USER_SCHED
308 /* Helper function to pass uid information to create_sched_user() */
309 void set_tg_uid(struct user_struct
*user
)
311 user
->tg
->uid
= user
->uid
;
316 * Every UID task group (including init_task_group aka UID-0) will
317 * be a child to this group.
319 struct task_group root_task_group
;
321 #ifdef CONFIG_FAIR_GROUP_SCHED
322 /* Default task group's sched entity on each cpu */
323 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
324 /* Default task group's cfs_rq on each cpu */
325 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
326 #endif /* CONFIG_FAIR_GROUP_SCHED */
328 #ifdef CONFIG_RT_GROUP_SCHED
329 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
330 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
331 #endif /* CONFIG_RT_GROUP_SCHED */
332 #else /* !CONFIG_USER_SCHED */
333 #define root_task_group init_task_group
334 #endif /* CONFIG_USER_SCHED */
336 /* task_group_lock serializes add/remove of task groups and also changes to
337 * a task group's cpu shares.
339 static DEFINE_SPINLOCK(task_group_lock
);
342 static int root_task_group_empty(void)
344 return list_empty(&root_task_group
.children
);
348 #ifdef CONFIG_FAIR_GROUP_SCHED
349 #ifdef CONFIG_USER_SCHED
350 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
351 #else /* !CONFIG_USER_SCHED */
352 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
353 #endif /* CONFIG_USER_SCHED */
356 * A weight of 0 or 1 can cause arithmetics problems.
357 * A weight of a cfs_rq is the sum of weights of which entities
358 * are queued on this cfs_rq, so a weight of a entity should not be
359 * too large, so as the shares value of a task group.
360 * (The default weight is 1024 - so there's no practical
361 * limitation from this.)
364 #define MAX_SHARES (1UL << 18)
366 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
369 /* Default task group.
370 * Every task in system belong to this group at bootup.
372 struct task_group init_task_group
;
374 /* return group to which a task belongs */
375 static inline struct task_group
*task_group(struct task_struct
*p
)
377 struct task_group
*tg
;
379 #ifdef CONFIG_USER_SCHED
381 tg
= __task_cred(p
)->user
->tg
;
383 #elif defined(CONFIG_CGROUP_SCHED)
384 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
385 struct task_group
, css
);
387 tg
= &init_task_group
;
392 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
393 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
395 #ifdef CONFIG_FAIR_GROUP_SCHED
396 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
397 p
->se
.parent
= task_group(p
)->se
[cpu
];
400 #ifdef CONFIG_RT_GROUP_SCHED
401 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
402 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
409 static int root_task_group_empty(void)
415 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
416 static inline struct task_group
*task_group(struct task_struct
*p
)
421 #endif /* CONFIG_GROUP_SCHED */
423 /* CFS-related fields in a runqueue */
425 struct load_weight load
;
426 unsigned long nr_running
;
431 struct rb_root tasks_timeline
;
432 struct rb_node
*rb_leftmost
;
434 struct list_head tasks
;
435 struct list_head
*balance_iterator
;
438 * 'curr' points to currently running entity on this cfs_rq.
439 * It is set to NULL otherwise (i.e when none are currently running).
441 struct sched_entity
*curr
, *next
, *last
;
443 unsigned int nr_spread_over
;
445 #ifdef CONFIG_FAIR_GROUP_SCHED
446 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
449 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
450 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
451 * (like users, containers etc.)
453 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
454 * list is used during load balance.
456 struct list_head leaf_cfs_rq_list
;
457 struct task_group
*tg
; /* group that "owns" this runqueue */
461 * the part of load.weight contributed by tasks
463 unsigned long task_weight
;
466 * h_load = weight * f(tg)
468 * Where f(tg) is the recursive weight fraction assigned to
471 unsigned long h_load
;
474 * this cpu's part of tg->shares
476 unsigned long shares
;
479 * load.weight at the time we set shares
481 unsigned long rq_weight
;
486 /* Real-Time classes' related field in a runqueue: */
488 struct rt_prio_array active
;
489 unsigned long rt_nr_running
;
490 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
492 int curr
; /* highest queued rt task prio */
494 int next
; /* next highest */
499 unsigned long rt_nr_migratory
;
501 struct plist_head pushable_tasks
;
506 /* Nests inside the rq lock: */
507 spinlock_t rt_runtime_lock
;
509 #ifdef CONFIG_RT_GROUP_SCHED
510 unsigned long rt_nr_boosted
;
513 struct list_head leaf_rt_rq_list
;
514 struct task_group
*tg
;
515 struct sched_rt_entity
*rt_se
;
522 * We add the notion of a root-domain which will be used to define per-domain
523 * variables. Each exclusive cpuset essentially defines an island domain by
524 * fully partitioning the member cpus from any other cpuset. Whenever a new
525 * exclusive cpuset is created, we also create and attach a new root-domain
532 cpumask_var_t online
;
535 * The "RT overload" flag: it gets set if a CPU has more than
536 * one runnable RT task.
538 cpumask_var_t rto_mask
;
541 struct cpupri cpupri
;
543 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
545 * Preferred wake up cpu nominated by sched_mc balance that will be
546 * used when most cpus are idle in the system indicating overall very
547 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
549 unsigned int sched_mc_preferred_wakeup_cpu
;
554 * By default the system creates a single root-domain with all cpus as
555 * members (mimicking the global state we have today).
557 static struct root_domain def_root_domain
;
562 * This is the main, per-CPU runqueue data structure.
564 * Locking rule: those places that want to lock multiple runqueues
565 * (such as the load balancing or the thread migration code), lock
566 * acquire operations must be ordered by ascending &runqueue.
573 * nr_running and cpu_load should be in the same cacheline because
574 * remote CPUs use both these fields when doing load calculation.
576 unsigned long nr_running
;
577 #define CPU_LOAD_IDX_MAX 5
578 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
580 unsigned long last_tick_seen
;
581 unsigned char in_nohz_recently
;
583 /* capture load from *all* tasks on this cpu: */
584 struct load_weight load
;
585 unsigned long nr_load_updates
;
591 #ifdef CONFIG_FAIR_GROUP_SCHED
592 /* list of leaf cfs_rq on this cpu: */
593 struct list_head leaf_cfs_rq_list
;
595 #ifdef CONFIG_RT_GROUP_SCHED
596 struct list_head leaf_rt_rq_list
;
600 * This is part of a global counter where only the total sum
601 * over all CPUs matters. A task can increase this counter on
602 * one CPU and if it got migrated afterwards it may decrease
603 * it on another CPU. Always updated under the runqueue lock:
605 unsigned long nr_uninterruptible
;
607 struct task_struct
*curr
, *idle
;
608 unsigned long next_balance
;
609 struct mm_struct
*prev_mm
;
616 struct root_domain
*rd
;
617 struct sched_domain
*sd
;
619 unsigned char idle_at_tick
;
620 /* For active balancing */
623 /* cpu of this runqueue: */
627 unsigned long avg_load_per_task
;
629 struct task_struct
*migration_thread
;
630 struct list_head migration_queue
;
633 #ifdef CONFIG_SCHED_HRTICK
635 int hrtick_csd_pending
;
636 struct call_single_data hrtick_csd
;
638 struct hrtimer hrtick_timer
;
641 #ifdef CONFIG_SCHEDSTATS
643 struct sched_info rq_sched_info
;
644 unsigned long long rq_cpu_time
;
645 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
647 /* sys_sched_yield() stats */
648 unsigned int yld_count
;
650 /* schedule() stats */
651 unsigned int sched_switch
;
652 unsigned int sched_count
;
653 unsigned int sched_goidle
;
655 /* try_to_wake_up() stats */
656 unsigned int ttwu_count
;
657 unsigned int ttwu_local
;
660 unsigned int bkl_count
;
664 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
666 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int sync
)
668 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, sync
);
671 static inline int cpu_of(struct rq
*rq
)
681 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
682 * See detach_destroy_domains: synchronize_sched for details.
684 * The domain tree of any CPU may only be accessed from within
685 * preempt-disabled sections.
687 #define for_each_domain(cpu, __sd) \
688 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
690 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
691 #define this_rq() (&__get_cpu_var(runqueues))
692 #define task_rq(p) cpu_rq(task_cpu(p))
693 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
695 static inline void update_rq_clock(struct rq
*rq
)
697 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
701 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
703 #ifdef CONFIG_SCHED_DEBUG
704 # define const_debug __read_mostly
706 # define const_debug static const
712 * Returns true if the current cpu runqueue is locked.
713 * This interface allows printk to be called with the runqueue lock
714 * held and know whether or not it is OK to wake up the klogd.
716 int runqueue_is_locked(void)
719 struct rq
*rq
= cpu_rq(cpu
);
722 ret
= spin_is_locked(&rq
->lock
);
728 * Debugging: various feature bits
731 #define SCHED_FEAT(name, enabled) \
732 __SCHED_FEAT_##name ,
735 #include "sched_features.h"
740 #define SCHED_FEAT(name, enabled) \
741 (1UL << __SCHED_FEAT_##name) * enabled |
743 const_debug
unsigned int sysctl_sched_features
=
744 #include "sched_features.h"
749 #ifdef CONFIG_SCHED_DEBUG
750 #define SCHED_FEAT(name, enabled) \
753 static __read_mostly
char *sched_feat_names
[] = {
754 #include "sched_features.h"
760 static int sched_feat_show(struct seq_file
*m
, void *v
)
764 for (i
= 0; sched_feat_names
[i
]; i
++) {
765 if (!(sysctl_sched_features
& (1UL << i
)))
767 seq_printf(m
, "%s ", sched_feat_names
[i
]);
775 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
776 size_t cnt
, loff_t
*ppos
)
786 if (copy_from_user(&buf
, ubuf
, cnt
))
791 if (strncmp(buf
, "NO_", 3) == 0) {
796 for (i
= 0; sched_feat_names
[i
]; i
++) {
797 int len
= strlen(sched_feat_names
[i
]);
799 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
801 sysctl_sched_features
&= ~(1UL << i
);
803 sysctl_sched_features
|= (1UL << i
);
808 if (!sched_feat_names
[i
])
816 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
818 return single_open(filp
, sched_feat_show
, NULL
);
821 static struct file_operations sched_feat_fops
= {
822 .open
= sched_feat_open
,
823 .write
= sched_feat_write
,
826 .release
= single_release
,
829 static __init
int sched_init_debug(void)
831 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
836 late_initcall(sched_init_debug
);
840 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
843 * Number of tasks to iterate in a single balance run.
844 * Limited because this is done with IRQs disabled.
846 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
849 * ratelimit for updating the group shares.
852 unsigned int sysctl_sched_shares_ratelimit
= 250000;
855 * Inject some fuzzyness into changing the per-cpu group shares
856 * this avoids remote rq-locks at the expense of fairness.
859 unsigned int sysctl_sched_shares_thresh
= 4;
862 * period over which we measure -rt task cpu usage in us.
865 unsigned int sysctl_sched_rt_period
= 1000000;
867 static __read_mostly
int scheduler_running
;
870 * part of the period that we allow rt tasks to run in us.
873 int sysctl_sched_rt_runtime
= 950000;
875 static inline u64
global_rt_period(void)
877 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
880 static inline u64
global_rt_runtime(void)
882 if (sysctl_sched_rt_runtime
< 0)
885 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
888 #ifndef prepare_arch_switch
889 # define prepare_arch_switch(next) do { } while (0)
891 #ifndef finish_arch_switch
892 # define finish_arch_switch(prev) do { } while (0)
895 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
897 return rq
->curr
== p
;
900 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
901 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
903 return task_current(rq
, p
);
906 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
910 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
912 #ifdef CONFIG_DEBUG_SPINLOCK
913 /* this is a valid case when another task releases the spinlock */
914 rq
->lock
.owner
= current
;
917 * If we are tracking spinlock dependencies then we have to
918 * fix up the runqueue lock - which gets 'carried over' from
921 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
923 spin_unlock_irq(&rq
->lock
);
926 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
927 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
932 return task_current(rq
, p
);
936 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
940 * We can optimise this out completely for !SMP, because the
941 * SMP rebalancing from interrupt is the only thing that cares
946 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
947 spin_unlock_irq(&rq
->lock
);
949 spin_unlock(&rq
->lock
);
953 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
957 * After ->oncpu is cleared, the task can be moved to a different CPU.
958 * We must ensure this doesn't happen until the switch is completely
964 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
968 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
971 * __task_rq_lock - lock the runqueue a given task resides on.
972 * Must be called interrupts disabled.
974 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
978 struct rq
*rq
= task_rq(p
);
979 spin_lock(&rq
->lock
);
980 if (likely(rq
== task_rq(p
)))
982 spin_unlock(&rq
->lock
);
987 * task_rq_lock - lock the runqueue a given task resides on and disable
988 * interrupts. Note the ordering: we can safely lookup the task_rq without
989 * explicitly disabling preemption.
991 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
997 local_irq_save(*flags
);
999 spin_lock(&rq
->lock
);
1000 if (likely(rq
== task_rq(p
)))
1002 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1006 void task_rq_unlock_wait(struct task_struct
*p
)
1008 struct rq
*rq
= task_rq(p
);
1010 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1011 spin_unlock_wait(&rq
->lock
);
1014 static void __task_rq_unlock(struct rq
*rq
)
1015 __releases(rq
->lock
)
1017 spin_unlock(&rq
->lock
);
1020 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
1021 __releases(rq
->lock
)
1023 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1027 * this_rq_lock - lock this runqueue and disable interrupts.
1029 static struct rq
*this_rq_lock(void)
1030 __acquires(rq
->lock
)
1034 local_irq_disable();
1036 spin_lock(&rq
->lock
);
1041 #ifdef CONFIG_SCHED_HRTICK
1043 * Use HR-timers to deliver accurate preemption points.
1045 * Its all a bit involved since we cannot program an hrt while holding the
1046 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1049 * When we get rescheduled we reprogram the hrtick_timer outside of the
1055 * - enabled by features
1056 * - hrtimer is actually high res
1058 static inline int hrtick_enabled(struct rq
*rq
)
1060 if (!sched_feat(HRTICK
))
1062 if (!cpu_active(cpu_of(rq
)))
1064 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1067 static void hrtick_clear(struct rq
*rq
)
1069 if (hrtimer_active(&rq
->hrtick_timer
))
1070 hrtimer_cancel(&rq
->hrtick_timer
);
1074 * High-resolution timer tick.
1075 * Runs from hardirq context with interrupts disabled.
1077 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1079 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1081 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1083 spin_lock(&rq
->lock
);
1084 update_rq_clock(rq
);
1085 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1086 spin_unlock(&rq
->lock
);
1088 return HRTIMER_NORESTART
;
1093 * called from hardirq (IPI) context
1095 static void __hrtick_start(void *arg
)
1097 struct rq
*rq
= arg
;
1099 spin_lock(&rq
->lock
);
1100 hrtimer_restart(&rq
->hrtick_timer
);
1101 rq
->hrtick_csd_pending
= 0;
1102 spin_unlock(&rq
->lock
);
1106 * Called to set the hrtick timer state.
1108 * called with rq->lock held and irqs disabled
1110 static void hrtick_start(struct rq
*rq
, u64 delay
)
1112 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1113 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1115 hrtimer_set_expires(timer
, time
);
1117 if (rq
== this_rq()) {
1118 hrtimer_restart(timer
);
1119 } else if (!rq
->hrtick_csd_pending
) {
1120 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
);
1121 rq
->hrtick_csd_pending
= 1;
1126 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1128 int cpu
= (int)(long)hcpu
;
1131 case CPU_UP_CANCELED
:
1132 case CPU_UP_CANCELED_FROZEN
:
1133 case CPU_DOWN_PREPARE
:
1134 case CPU_DOWN_PREPARE_FROZEN
:
1136 case CPU_DEAD_FROZEN
:
1137 hrtick_clear(cpu_rq(cpu
));
1144 static __init
void init_hrtick(void)
1146 hotcpu_notifier(hotplug_hrtick
, 0);
1150 * Called to set the hrtick timer state.
1152 * called with rq->lock held and irqs disabled
1154 static void hrtick_start(struct rq
*rq
, u64 delay
)
1156 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1157 HRTIMER_MODE_REL
, 0);
1160 static inline void init_hrtick(void)
1163 #endif /* CONFIG_SMP */
1165 static void init_rq_hrtick(struct rq
*rq
)
1168 rq
->hrtick_csd_pending
= 0;
1170 rq
->hrtick_csd
.flags
= 0;
1171 rq
->hrtick_csd
.func
= __hrtick_start
;
1172 rq
->hrtick_csd
.info
= rq
;
1175 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1176 rq
->hrtick_timer
.function
= hrtick
;
1178 #else /* CONFIG_SCHED_HRTICK */
1179 static inline void hrtick_clear(struct rq
*rq
)
1183 static inline void init_rq_hrtick(struct rq
*rq
)
1187 static inline void init_hrtick(void)
1190 #endif /* CONFIG_SCHED_HRTICK */
1193 * resched_task - mark a task 'to be rescheduled now'.
1195 * On UP this means the setting of the need_resched flag, on SMP it
1196 * might also involve a cross-CPU call to trigger the scheduler on
1201 #ifndef tsk_is_polling
1202 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1205 static void resched_task(struct task_struct
*p
)
1209 assert_spin_locked(&task_rq(p
)->lock
);
1211 if (test_tsk_need_resched(p
))
1214 set_tsk_need_resched(p
);
1217 if (cpu
== smp_processor_id())
1220 /* NEED_RESCHED must be visible before we test polling */
1222 if (!tsk_is_polling(p
))
1223 smp_send_reschedule(cpu
);
1226 static void resched_cpu(int cpu
)
1228 struct rq
*rq
= cpu_rq(cpu
);
1229 unsigned long flags
;
1231 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1233 resched_task(cpu_curr(cpu
));
1234 spin_unlock_irqrestore(&rq
->lock
, flags
);
1239 * When add_timer_on() enqueues a timer into the timer wheel of an
1240 * idle CPU then this timer might expire before the next timer event
1241 * which is scheduled to wake up that CPU. In case of a completely
1242 * idle system the next event might even be infinite time into the
1243 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1244 * leaves the inner idle loop so the newly added timer is taken into
1245 * account when the CPU goes back to idle and evaluates the timer
1246 * wheel for the next timer event.
1248 void wake_up_idle_cpu(int cpu
)
1250 struct rq
*rq
= cpu_rq(cpu
);
1252 if (cpu
== smp_processor_id())
1256 * This is safe, as this function is called with the timer
1257 * wheel base lock of (cpu) held. When the CPU is on the way
1258 * to idle and has not yet set rq->curr to idle then it will
1259 * be serialized on the timer wheel base lock and take the new
1260 * timer into account automatically.
1262 if (rq
->curr
!= rq
->idle
)
1266 * We can set TIF_RESCHED on the idle task of the other CPU
1267 * lockless. The worst case is that the other CPU runs the
1268 * idle task through an additional NOOP schedule()
1270 set_tsk_need_resched(rq
->idle
);
1272 /* NEED_RESCHED must be visible before we test polling */
1274 if (!tsk_is_polling(rq
->idle
))
1275 smp_send_reschedule(cpu
);
1277 #endif /* CONFIG_NO_HZ */
1279 #else /* !CONFIG_SMP */
1280 static void resched_task(struct task_struct
*p
)
1282 assert_spin_locked(&task_rq(p
)->lock
);
1283 set_tsk_need_resched(p
);
1285 #endif /* CONFIG_SMP */
1287 #if BITS_PER_LONG == 32
1288 # define WMULT_CONST (~0UL)
1290 # define WMULT_CONST (1UL << 32)
1293 #define WMULT_SHIFT 32
1296 * Shift right and round:
1298 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1301 * delta *= weight / lw
1303 static unsigned long
1304 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1305 struct load_weight
*lw
)
1309 if (!lw
->inv_weight
) {
1310 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1313 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1317 tmp
= (u64
)delta_exec
* weight
;
1319 * Check whether we'd overflow the 64-bit multiplication:
1321 if (unlikely(tmp
> WMULT_CONST
))
1322 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1325 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1327 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1330 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1336 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1343 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1344 * of tasks with abnormal "nice" values across CPUs the contribution that
1345 * each task makes to its run queue's load is weighted according to its
1346 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1347 * scaled version of the new time slice allocation that they receive on time
1351 #define WEIGHT_IDLEPRIO 3
1352 #define WMULT_IDLEPRIO 1431655765
1355 * Nice levels are multiplicative, with a gentle 10% change for every
1356 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1357 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1358 * that remained on nice 0.
1360 * The "10% effect" is relative and cumulative: from _any_ nice level,
1361 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1362 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1363 * If a task goes up by ~10% and another task goes down by ~10% then
1364 * the relative distance between them is ~25%.)
1366 static const int prio_to_weight
[40] = {
1367 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1368 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1369 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1370 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1371 /* 0 */ 1024, 820, 655, 526, 423,
1372 /* 5 */ 335, 272, 215, 172, 137,
1373 /* 10 */ 110, 87, 70, 56, 45,
1374 /* 15 */ 36, 29, 23, 18, 15,
1378 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1380 * In cases where the weight does not change often, we can use the
1381 * precalculated inverse to speed up arithmetics by turning divisions
1382 * into multiplications:
1384 static const u32 prio_to_wmult
[40] = {
1385 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1386 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1387 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1388 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1389 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1390 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1391 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1392 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1395 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1398 * runqueue iterator, to support SMP load-balancing between different
1399 * scheduling classes, without having to expose their internal data
1400 * structures to the load-balancing proper:
1402 struct rq_iterator
{
1404 struct task_struct
*(*start
)(void *);
1405 struct task_struct
*(*next
)(void *);
1409 static unsigned long
1410 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1411 unsigned long max_load_move
, struct sched_domain
*sd
,
1412 enum cpu_idle_type idle
, int *all_pinned
,
1413 int *this_best_prio
, struct rq_iterator
*iterator
);
1416 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1417 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1418 struct rq_iterator
*iterator
);
1421 #ifdef CONFIG_CGROUP_CPUACCT
1422 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1424 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1427 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1429 update_load_add(&rq
->load
, load
);
1432 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1434 update_load_sub(&rq
->load
, load
);
1437 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1438 typedef int (*tg_visitor
)(struct task_group
*, void *);
1441 * Iterate the full tree, calling @down when first entering a node and @up when
1442 * leaving it for the final time.
1444 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1446 struct task_group
*parent
, *child
;
1450 parent
= &root_task_group
;
1452 ret
= (*down
)(parent
, data
);
1455 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1462 ret
= (*up
)(parent
, data
);
1467 parent
= parent
->parent
;
1476 static int tg_nop(struct task_group
*tg
, void *data
)
1483 static unsigned long source_load(int cpu
, int type
);
1484 static unsigned long target_load(int cpu
, int type
);
1485 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1487 static unsigned long cpu_avg_load_per_task(int cpu
)
1489 struct rq
*rq
= cpu_rq(cpu
);
1490 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1493 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1495 rq
->avg_load_per_task
= 0;
1497 return rq
->avg_load_per_task
;
1500 #ifdef CONFIG_FAIR_GROUP_SCHED
1502 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1505 * Calculate and set the cpu's group shares.
1508 update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1509 unsigned long sd_shares
, unsigned long sd_rq_weight
)
1511 unsigned long shares
;
1512 unsigned long rq_weight
;
1517 rq_weight
= tg
->cfs_rq
[cpu
]->rq_weight
;
1520 * \Sum shares * rq_weight
1521 * shares = -----------------------
1525 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1526 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1528 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1529 sysctl_sched_shares_thresh
) {
1530 struct rq
*rq
= cpu_rq(cpu
);
1531 unsigned long flags
;
1533 spin_lock_irqsave(&rq
->lock
, flags
);
1534 tg
->cfs_rq
[cpu
]->shares
= shares
;
1536 __set_se_shares(tg
->se
[cpu
], shares
);
1537 spin_unlock_irqrestore(&rq
->lock
, flags
);
1542 * Re-compute the task group their per cpu shares over the given domain.
1543 * This needs to be done in a bottom-up fashion because the rq weight of a
1544 * parent group depends on the shares of its child groups.
1546 static int tg_shares_up(struct task_group
*tg
, void *data
)
1548 unsigned long weight
, rq_weight
= 0;
1549 unsigned long shares
= 0;
1550 struct sched_domain
*sd
= data
;
1553 for_each_cpu(i
, sched_domain_span(sd
)) {
1555 * If there are currently no tasks on the cpu pretend there
1556 * is one of average load so that when a new task gets to
1557 * run here it will not get delayed by group starvation.
1559 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1561 weight
= NICE_0_LOAD
;
1563 tg
->cfs_rq
[i
]->rq_weight
= weight
;
1564 rq_weight
+= weight
;
1565 shares
+= tg
->cfs_rq
[i
]->shares
;
1568 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1569 shares
= tg
->shares
;
1571 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1572 shares
= tg
->shares
;
1574 for_each_cpu(i
, sched_domain_span(sd
))
1575 update_group_shares_cpu(tg
, i
, shares
, rq_weight
);
1581 * Compute the cpu's hierarchical load factor for each task group.
1582 * This needs to be done in a top-down fashion because the load of a child
1583 * group is a fraction of its parents load.
1585 static int tg_load_down(struct task_group
*tg
, void *data
)
1588 long cpu
= (long)data
;
1591 load
= cpu_rq(cpu
)->load
.weight
;
1593 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1594 load
*= tg
->cfs_rq
[cpu
]->shares
;
1595 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1598 tg
->cfs_rq
[cpu
]->h_load
= load
;
1603 static void update_shares(struct sched_domain
*sd
)
1605 u64 now
= cpu_clock(raw_smp_processor_id());
1606 s64 elapsed
= now
- sd
->last_update
;
1608 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1609 sd
->last_update
= now
;
1610 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1614 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1616 spin_unlock(&rq
->lock
);
1618 spin_lock(&rq
->lock
);
1621 static void update_h_load(long cpu
)
1623 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1628 static inline void update_shares(struct sched_domain
*sd
)
1632 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1638 #ifdef CONFIG_PREEMPT
1641 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1642 * way at the expense of forcing extra atomic operations in all
1643 * invocations. This assures that the double_lock is acquired using the
1644 * same underlying policy as the spinlock_t on this architecture, which
1645 * reduces latency compared to the unfair variant below. However, it
1646 * also adds more overhead and therefore may reduce throughput.
1648 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1649 __releases(this_rq
->lock
)
1650 __acquires(busiest
->lock
)
1651 __acquires(this_rq
->lock
)
1653 spin_unlock(&this_rq
->lock
);
1654 double_rq_lock(this_rq
, busiest
);
1661 * Unfair double_lock_balance: Optimizes throughput at the expense of
1662 * latency by eliminating extra atomic operations when the locks are
1663 * already in proper order on entry. This favors lower cpu-ids and will
1664 * grant the double lock to lower cpus over higher ids under contention,
1665 * regardless of entry order into the function.
1667 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1668 __releases(this_rq
->lock
)
1669 __acquires(busiest
->lock
)
1670 __acquires(this_rq
->lock
)
1674 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1675 if (busiest
< this_rq
) {
1676 spin_unlock(&this_rq
->lock
);
1677 spin_lock(&busiest
->lock
);
1678 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
1681 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
1686 #endif /* CONFIG_PREEMPT */
1689 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1691 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1693 if (unlikely(!irqs_disabled())) {
1694 /* printk() doesn't work good under rq->lock */
1695 spin_unlock(&this_rq
->lock
);
1699 return _double_lock_balance(this_rq
, busiest
);
1702 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1703 __releases(busiest
->lock
)
1705 spin_unlock(&busiest
->lock
);
1706 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1710 #ifdef CONFIG_FAIR_GROUP_SCHED
1711 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1714 cfs_rq
->shares
= shares
;
1719 #include "sched_stats.h"
1720 #include "sched_idletask.c"
1721 #include "sched_fair.c"
1722 #include "sched_rt.c"
1723 #ifdef CONFIG_SCHED_DEBUG
1724 # include "sched_debug.c"
1727 #define sched_class_highest (&rt_sched_class)
1728 #define for_each_class(class) \
1729 for (class = sched_class_highest; class; class = class->next)
1731 static void inc_nr_running(struct rq
*rq
)
1736 static void dec_nr_running(struct rq
*rq
)
1741 static void set_load_weight(struct task_struct
*p
)
1743 if (task_has_rt_policy(p
)) {
1744 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1745 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1750 * SCHED_IDLE tasks get minimal weight:
1752 if (p
->policy
== SCHED_IDLE
) {
1753 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1754 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1758 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1759 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1762 static void update_avg(u64
*avg
, u64 sample
)
1764 s64 diff
= sample
- *avg
;
1768 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1771 p
->se
.start_runtime
= p
->se
.sum_exec_runtime
;
1773 sched_info_queued(p
);
1774 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1778 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1781 if (p
->se
.last_wakeup
) {
1782 update_avg(&p
->se
.avg_overlap
,
1783 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1784 p
->se
.last_wakeup
= 0;
1786 update_avg(&p
->se
.avg_wakeup
,
1787 sysctl_sched_wakeup_granularity
);
1791 sched_info_dequeued(p
);
1792 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1797 * __normal_prio - return the priority that is based on the static prio
1799 static inline int __normal_prio(struct task_struct
*p
)
1801 return p
->static_prio
;
1805 * Calculate the expected normal priority: i.e. priority
1806 * without taking RT-inheritance into account. Might be
1807 * boosted by interactivity modifiers. Changes upon fork,
1808 * setprio syscalls, and whenever the interactivity
1809 * estimator recalculates.
1811 static inline int normal_prio(struct task_struct
*p
)
1815 if (task_has_rt_policy(p
))
1816 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1818 prio
= __normal_prio(p
);
1823 * Calculate the current priority, i.e. the priority
1824 * taken into account by the scheduler. This value might
1825 * be boosted by RT tasks, or might be boosted by
1826 * interactivity modifiers. Will be RT if the task got
1827 * RT-boosted. If not then it returns p->normal_prio.
1829 static int effective_prio(struct task_struct
*p
)
1831 p
->normal_prio
= normal_prio(p
);
1833 * If we are RT tasks or we were boosted to RT priority,
1834 * keep the priority unchanged. Otherwise, update priority
1835 * to the normal priority:
1837 if (!rt_prio(p
->prio
))
1838 return p
->normal_prio
;
1843 * activate_task - move a task to the runqueue.
1845 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1847 if (task_contributes_to_load(p
))
1848 rq
->nr_uninterruptible
--;
1850 enqueue_task(rq
, p
, wakeup
);
1855 * deactivate_task - remove a task from the runqueue.
1857 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1859 if (task_contributes_to_load(p
))
1860 rq
->nr_uninterruptible
++;
1862 dequeue_task(rq
, p
, sleep
);
1867 * task_curr - is this task currently executing on a CPU?
1868 * @p: the task in question.
1870 inline int task_curr(const struct task_struct
*p
)
1872 return cpu_curr(task_cpu(p
)) == p
;
1875 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1877 set_task_rq(p
, cpu
);
1880 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1881 * successfuly executed on another CPU. We must ensure that updates of
1882 * per-task data have been completed by this moment.
1885 task_thread_info(p
)->cpu
= cpu
;
1889 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1890 const struct sched_class
*prev_class
,
1891 int oldprio
, int running
)
1893 if (prev_class
!= p
->sched_class
) {
1894 if (prev_class
->switched_from
)
1895 prev_class
->switched_from(rq
, p
, running
);
1896 p
->sched_class
->switched_to(rq
, p
, running
);
1898 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1903 /* Used instead of source_load when we know the type == 0 */
1904 static unsigned long weighted_cpuload(const int cpu
)
1906 return cpu_rq(cpu
)->load
.weight
;
1910 * Is this task likely cache-hot:
1913 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1918 * Buddy candidates are cache hot:
1920 if (sched_feat(CACHE_HOT_BUDDY
) &&
1921 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
1922 &p
->se
== cfs_rq_of(&p
->se
)->last
))
1925 if (p
->sched_class
!= &fair_sched_class
)
1928 if (sysctl_sched_migration_cost
== -1)
1930 if (sysctl_sched_migration_cost
== 0)
1933 delta
= now
- p
->se
.exec_start
;
1935 return delta
< (s64
)sysctl_sched_migration_cost
;
1939 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1941 int old_cpu
= task_cpu(p
);
1942 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1943 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1944 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1947 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1949 trace_sched_migrate_task(p
, task_cpu(p
), new_cpu
);
1951 #ifdef CONFIG_SCHEDSTATS
1952 if (p
->se
.wait_start
)
1953 p
->se
.wait_start
-= clock_offset
;
1954 if (p
->se
.sleep_start
)
1955 p
->se
.sleep_start
-= clock_offset
;
1956 if (p
->se
.block_start
)
1957 p
->se
.block_start
-= clock_offset
;
1958 if (old_cpu
!= new_cpu
) {
1959 schedstat_inc(p
, se
.nr_migrations
);
1960 if (task_hot(p
, old_rq
->clock
, NULL
))
1961 schedstat_inc(p
, se
.nr_forced2_migrations
);
1964 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1965 new_cfsrq
->min_vruntime
;
1967 __set_task_cpu(p
, new_cpu
);
1970 struct migration_req
{
1971 struct list_head list
;
1973 struct task_struct
*task
;
1976 struct completion done
;
1980 * The task's runqueue lock must be held.
1981 * Returns true if you have to wait for migration thread.
1984 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1986 struct rq
*rq
= task_rq(p
);
1989 * If the task is not on a runqueue (and not running), then
1990 * it is sufficient to simply update the task's cpu field.
1992 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1993 set_task_cpu(p
, dest_cpu
);
1997 init_completion(&req
->done
);
1999 req
->dest_cpu
= dest_cpu
;
2000 list_add(&req
->list
, &rq
->migration_queue
);
2006 * wait_task_inactive - wait for a thread to unschedule.
2008 * If @match_state is nonzero, it's the @p->state value just checked and
2009 * not expected to change. If it changes, i.e. @p might have woken up,
2010 * then return zero. When we succeed in waiting for @p to be off its CPU,
2011 * we return a positive number (its total switch count). If a second call
2012 * a short while later returns the same number, the caller can be sure that
2013 * @p has remained unscheduled the whole time.
2015 * The caller must ensure that the task *will* unschedule sometime soon,
2016 * else this function might spin for a *long* time. This function can't
2017 * be called with interrupts off, or it may introduce deadlock with
2018 * smp_call_function() if an IPI is sent by the same process we are
2019 * waiting to become inactive.
2021 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2023 unsigned long flags
;
2030 * We do the initial early heuristics without holding
2031 * any task-queue locks at all. We'll only try to get
2032 * the runqueue lock when things look like they will
2038 * If the task is actively running on another CPU
2039 * still, just relax and busy-wait without holding
2042 * NOTE! Since we don't hold any locks, it's not
2043 * even sure that "rq" stays as the right runqueue!
2044 * But we don't care, since "task_running()" will
2045 * return false if the runqueue has changed and p
2046 * is actually now running somewhere else!
2048 while (task_running(rq
, p
)) {
2049 if (match_state
&& unlikely(p
->state
!= match_state
))
2055 * Ok, time to look more closely! We need the rq
2056 * lock now, to be *sure*. If we're wrong, we'll
2057 * just go back and repeat.
2059 rq
= task_rq_lock(p
, &flags
);
2060 trace_sched_wait_task(rq
, p
);
2061 running
= task_running(rq
, p
);
2062 on_rq
= p
->se
.on_rq
;
2064 if (!match_state
|| p
->state
== match_state
)
2065 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2066 task_rq_unlock(rq
, &flags
);
2069 * If it changed from the expected state, bail out now.
2071 if (unlikely(!ncsw
))
2075 * Was it really running after all now that we
2076 * checked with the proper locks actually held?
2078 * Oops. Go back and try again..
2080 if (unlikely(running
)) {
2086 * It's not enough that it's not actively running,
2087 * it must be off the runqueue _entirely_, and not
2090 * So if it was still runnable (but just not actively
2091 * running right now), it's preempted, and we should
2092 * yield - it could be a while.
2094 if (unlikely(on_rq
)) {
2095 schedule_timeout_uninterruptible(1);
2100 * Ahh, all good. It wasn't running, and it wasn't
2101 * runnable, which means that it will never become
2102 * running in the future either. We're all done!
2111 * kick_process - kick a running thread to enter/exit the kernel
2112 * @p: the to-be-kicked thread
2114 * Cause a process which is running on another CPU to enter
2115 * kernel-mode, without any delay. (to get signals handled.)
2117 * NOTE: this function doesnt have to take the runqueue lock,
2118 * because all it wants to ensure is that the remote task enters
2119 * the kernel. If the IPI races and the task has been migrated
2120 * to another CPU then no harm is done and the purpose has been
2123 void kick_process(struct task_struct
*p
)
2129 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2130 smp_send_reschedule(cpu
);
2135 * Return a low guess at the load of a migration-source cpu weighted
2136 * according to the scheduling class and "nice" value.
2138 * We want to under-estimate the load of migration sources, to
2139 * balance conservatively.
2141 static unsigned long source_load(int cpu
, int type
)
2143 struct rq
*rq
= cpu_rq(cpu
);
2144 unsigned long total
= weighted_cpuload(cpu
);
2146 if (type
== 0 || !sched_feat(LB_BIAS
))
2149 return min(rq
->cpu_load
[type
-1], total
);
2153 * Return a high guess at the load of a migration-target cpu weighted
2154 * according to the scheduling class and "nice" value.
2156 static unsigned long target_load(int cpu
, int type
)
2158 struct rq
*rq
= cpu_rq(cpu
);
2159 unsigned long total
= weighted_cpuload(cpu
);
2161 if (type
== 0 || !sched_feat(LB_BIAS
))
2164 return max(rq
->cpu_load
[type
-1], total
);
2168 * find_idlest_group finds and returns the least busy CPU group within the
2171 static struct sched_group
*
2172 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2174 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2175 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2176 int load_idx
= sd
->forkexec_idx
;
2177 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2180 unsigned long load
, avg_load
;
2184 /* Skip over this group if it has no CPUs allowed */
2185 if (!cpumask_intersects(sched_group_cpus(group
),
2189 local_group
= cpumask_test_cpu(this_cpu
,
2190 sched_group_cpus(group
));
2192 /* Tally up the load of all CPUs in the group */
2195 for_each_cpu(i
, sched_group_cpus(group
)) {
2196 /* Bias balancing toward cpus of our domain */
2198 load
= source_load(i
, load_idx
);
2200 load
= target_load(i
, load_idx
);
2205 /* Adjust by relative CPU power of the group */
2206 avg_load
= sg_div_cpu_power(group
,
2207 avg_load
* SCHED_LOAD_SCALE
);
2210 this_load
= avg_load
;
2212 } else if (avg_load
< min_load
) {
2213 min_load
= avg_load
;
2216 } while (group
= group
->next
, group
!= sd
->groups
);
2218 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2224 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2227 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
2229 unsigned long load
, min_load
= ULONG_MAX
;
2233 /* Traverse only the allowed CPUs */
2234 for_each_cpu_and(i
, sched_group_cpus(group
), &p
->cpus_allowed
) {
2235 load
= weighted_cpuload(i
);
2237 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2247 * sched_balance_self: balance the current task (running on cpu) in domains
2248 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2251 * Balance, ie. select the least loaded group.
2253 * Returns the target CPU number, or the same CPU if no balancing is needed.
2255 * preempt must be disabled.
2257 static int sched_balance_self(int cpu
, int flag
)
2259 struct task_struct
*t
= current
;
2260 struct sched_domain
*tmp
, *sd
= NULL
;
2262 for_each_domain(cpu
, tmp
) {
2264 * If power savings logic is enabled for a domain, stop there.
2266 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2268 if (tmp
->flags
& flag
)
2276 struct sched_group
*group
;
2277 int new_cpu
, weight
;
2279 if (!(sd
->flags
& flag
)) {
2284 group
= find_idlest_group(sd
, t
, cpu
);
2290 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
2291 if (new_cpu
== -1 || new_cpu
== cpu
) {
2292 /* Now try balancing at a lower domain level of cpu */
2297 /* Now try balancing at a lower domain level of new_cpu */
2299 weight
= cpumask_weight(sched_domain_span(sd
));
2301 for_each_domain(cpu
, tmp
) {
2302 if (weight
<= cpumask_weight(sched_domain_span(tmp
)))
2304 if (tmp
->flags
& flag
)
2307 /* while loop will break here if sd == NULL */
2313 #endif /* CONFIG_SMP */
2316 * try_to_wake_up - wake up a thread
2317 * @p: the to-be-woken-up thread
2318 * @state: the mask of task states that can be woken
2319 * @sync: do a synchronous wakeup?
2321 * Put it on the run-queue if it's not already there. The "current"
2322 * thread is always on the run-queue (except when the actual
2323 * re-schedule is in progress), and as such you're allowed to do
2324 * the simpler "current->state = TASK_RUNNING" to mark yourself
2325 * runnable without the overhead of this.
2327 * returns failure only if the task is already active.
2329 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2331 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2332 unsigned long flags
;
2336 if (!sched_feat(SYNC_WAKEUPS
))
2340 if (sched_feat(LB_WAKEUP_UPDATE
) && !root_task_group_empty()) {
2341 struct sched_domain
*sd
;
2343 this_cpu
= raw_smp_processor_id();
2346 for_each_domain(this_cpu
, sd
) {
2347 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2356 rq
= task_rq_lock(p
, &flags
);
2357 update_rq_clock(rq
);
2358 old_state
= p
->state
;
2359 if (!(old_state
& state
))
2367 this_cpu
= smp_processor_id();
2370 if (unlikely(task_running(rq
, p
)))
2373 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2374 if (cpu
!= orig_cpu
) {
2375 set_task_cpu(p
, cpu
);
2376 task_rq_unlock(rq
, &flags
);
2377 /* might preempt at this point */
2378 rq
= task_rq_lock(p
, &flags
);
2379 old_state
= p
->state
;
2380 if (!(old_state
& state
))
2385 this_cpu
= smp_processor_id();
2389 #ifdef CONFIG_SCHEDSTATS
2390 schedstat_inc(rq
, ttwu_count
);
2391 if (cpu
== this_cpu
)
2392 schedstat_inc(rq
, ttwu_local
);
2394 struct sched_domain
*sd
;
2395 for_each_domain(this_cpu
, sd
) {
2396 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2397 schedstat_inc(sd
, ttwu_wake_remote
);
2402 #endif /* CONFIG_SCHEDSTATS */
2405 #endif /* CONFIG_SMP */
2406 schedstat_inc(p
, se
.nr_wakeups
);
2408 schedstat_inc(p
, se
.nr_wakeups_sync
);
2409 if (orig_cpu
!= cpu
)
2410 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2411 if (cpu
== this_cpu
)
2412 schedstat_inc(p
, se
.nr_wakeups_local
);
2414 schedstat_inc(p
, se
.nr_wakeups_remote
);
2415 activate_task(rq
, p
, 1);
2419 * Only attribute actual wakeups done by this task.
2421 if (!in_interrupt()) {
2422 struct sched_entity
*se
= ¤t
->se
;
2423 u64 sample
= se
->sum_exec_runtime
;
2425 if (se
->last_wakeup
)
2426 sample
-= se
->last_wakeup
;
2428 sample
-= se
->start_runtime
;
2429 update_avg(&se
->avg_wakeup
, sample
);
2431 se
->last_wakeup
= se
->sum_exec_runtime
;
2435 trace_sched_wakeup(rq
, p
, success
);
2436 check_preempt_curr(rq
, p
, sync
);
2438 p
->state
= TASK_RUNNING
;
2440 if (p
->sched_class
->task_wake_up
)
2441 p
->sched_class
->task_wake_up(rq
, p
);
2444 task_rq_unlock(rq
, &flags
);
2449 int wake_up_process(struct task_struct
*p
)
2451 return try_to_wake_up(p
, TASK_ALL
, 0);
2453 EXPORT_SYMBOL(wake_up_process
);
2455 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2457 return try_to_wake_up(p
, state
, 0);
2461 * Perform scheduler related setup for a newly forked process p.
2462 * p is forked by current.
2464 * __sched_fork() is basic setup used by init_idle() too:
2466 static void __sched_fork(struct task_struct
*p
)
2468 p
->se
.exec_start
= 0;
2469 p
->se
.sum_exec_runtime
= 0;
2470 p
->se
.prev_sum_exec_runtime
= 0;
2471 p
->se
.last_wakeup
= 0;
2472 p
->se
.avg_overlap
= 0;
2473 p
->se
.start_runtime
= 0;
2474 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
2476 #ifdef CONFIG_SCHEDSTATS
2477 p
->se
.wait_start
= 0;
2478 p
->se
.sum_sleep_runtime
= 0;
2479 p
->se
.sleep_start
= 0;
2480 p
->se
.block_start
= 0;
2481 p
->se
.sleep_max
= 0;
2482 p
->se
.block_max
= 0;
2484 p
->se
.slice_max
= 0;
2488 INIT_LIST_HEAD(&p
->rt
.run_list
);
2490 INIT_LIST_HEAD(&p
->se
.group_node
);
2492 #ifdef CONFIG_PREEMPT_NOTIFIERS
2493 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2497 * We mark the process as running here, but have not actually
2498 * inserted it onto the runqueue yet. This guarantees that
2499 * nobody will actually run it, and a signal or other external
2500 * event cannot wake it up and insert it on the runqueue either.
2502 p
->state
= TASK_RUNNING
;
2506 * fork()/clone()-time setup:
2508 void sched_fork(struct task_struct
*p
, int clone_flags
)
2510 int cpu
= get_cpu();
2515 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2517 set_task_cpu(p
, cpu
);
2520 * Make sure we do not leak PI boosting priority to the child:
2522 p
->prio
= current
->normal_prio
;
2523 if (!rt_prio(p
->prio
))
2524 p
->sched_class
= &fair_sched_class
;
2526 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2527 if (likely(sched_info_on()))
2528 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2530 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2533 #ifdef CONFIG_PREEMPT
2534 /* Want to start with kernel preemption disabled. */
2535 task_thread_info(p
)->preempt_count
= 1;
2537 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2543 * wake_up_new_task - wake up a newly created task for the first time.
2545 * This function will do some initial scheduler statistics housekeeping
2546 * that must be done for every newly created context, then puts the task
2547 * on the runqueue and wakes it.
2549 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2551 unsigned long flags
;
2554 rq
= task_rq_lock(p
, &flags
);
2555 BUG_ON(p
->state
!= TASK_RUNNING
);
2556 update_rq_clock(rq
);
2558 p
->prio
= effective_prio(p
);
2560 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2561 activate_task(rq
, p
, 0);
2564 * Let the scheduling class do new task startup
2565 * management (if any):
2567 p
->sched_class
->task_new(rq
, p
);
2570 trace_sched_wakeup_new(rq
, p
, 1);
2571 check_preempt_curr(rq
, p
, 0);
2573 if (p
->sched_class
->task_wake_up
)
2574 p
->sched_class
->task_wake_up(rq
, p
);
2576 task_rq_unlock(rq
, &flags
);
2579 #ifdef CONFIG_PREEMPT_NOTIFIERS
2582 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2583 * @notifier: notifier struct to register
2585 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2587 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2589 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2592 * preempt_notifier_unregister - no longer interested in preemption notifications
2593 * @notifier: notifier struct to unregister
2595 * This is safe to call from within a preemption notifier.
2597 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2599 hlist_del(¬ifier
->link
);
2601 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2603 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2605 struct preempt_notifier
*notifier
;
2606 struct hlist_node
*node
;
2608 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2609 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2613 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2614 struct task_struct
*next
)
2616 struct preempt_notifier
*notifier
;
2617 struct hlist_node
*node
;
2619 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2620 notifier
->ops
->sched_out(notifier
, next
);
2623 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2625 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2630 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2631 struct task_struct
*next
)
2635 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2638 * prepare_task_switch - prepare to switch tasks
2639 * @rq: the runqueue preparing to switch
2640 * @prev: the current task that is being switched out
2641 * @next: the task we are going to switch to.
2643 * This is called with the rq lock held and interrupts off. It must
2644 * be paired with a subsequent finish_task_switch after the context
2647 * prepare_task_switch sets up locking and calls architecture specific
2651 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2652 struct task_struct
*next
)
2654 fire_sched_out_preempt_notifiers(prev
, next
);
2655 prepare_lock_switch(rq
, next
);
2656 prepare_arch_switch(next
);
2660 * finish_task_switch - clean up after a task-switch
2661 * @rq: runqueue associated with task-switch
2662 * @prev: the thread we just switched away from.
2664 * finish_task_switch must be called after the context switch, paired
2665 * with a prepare_task_switch call before the context switch.
2666 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2667 * and do any other architecture-specific cleanup actions.
2669 * Note that we may have delayed dropping an mm in context_switch(). If
2670 * so, we finish that here outside of the runqueue lock. (Doing it
2671 * with the lock held can cause deadlocks; see schedule() for
2674 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2675 __releases(rq
->lock
)
2677 struct mm_struct
*mm
= rq
->prev_mm
;
2680 int post_schedule
= 0;
2682 if (current
->sched_class
->needs_post_schedule
)
2683 post_schedule
= current
->sched_class
->needs_post_schedule(rq
);
2689 * A task struct has one reference for the use as "current".
2690 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2691 * schedule one last time. The schedule call will never return, and
2692 * the scheduled task must drop that reference.
2693 * The test for TASK_DEAD must occur while the runqueue locks are
2694 * still held, otherwise prev could be scheduled on another cpu, die
2695 * there before we look at prev->state, and then the reference would
2697 * Manfred Spraul <manfred@colorfullife.com>
2699 prev_state
= prev
->state
;
2700 finish_arch_switch(prev
);
2701 finish_lock_switch(rq
, prev
);
2704 current
->sched_class
->post_schedule(rq
);
2707 fire_sched_in_preempt_notifiers(current
);
2710 if (unlikely(prev_state
== TASK_DEAD
)) {
2712 * Remove function-return probe instances associated with this
2713 * task and put them back on the free list.
2715 kprobe_flush_task(prev
);
2716 put_task_struct(prev
);
2721 * schedule_tail - first thing a freshly forked thread must call.
2722 * @prev: the thread we just switched away from.
2724 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2725 __releases(rq
->lock
)
2727 struct rq
*rq
= this_rq();
2729 finish_task_switch(rq
, prev
);
2730 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2731 /* In this case, finish_task_switch does not reenable preemption */
2734 if (current
->set_child_tid
)
2735 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2739 * context_switch - switch to the new MM and the new
2740 * thread's register state.
2743 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2744 struct task_struct
*next
)
2746 struct mm_struct
*mm
, *oldmm
;
2748 prepare_task_switch(rq
, prev
, next
);
2749 trace_sched_switch(rq
, prev
, next
);
2751 oldmm
= prev
->active_mm
;
2753 * For paravirt, this is coupled with an exit in switch_to to
2754 * combine the page table reload and the switch backend into
2757 arch_enter_lazy_cpu_mode();
2759 if (unlikely(!mm
)) {
2760 next
->active_mm
= oldmm
;
2761 atomic_inc(&oldmm
->mm_count
);
2762 enter_lazy_tlb(oldmm
, next
);
2764 switch_mm(oldmm
, mm
, next
);
2766 if (unlikely(!prev
->mm
)) {
2767 prev
->active_mm
= NULL
;
2768 rq
->prev_mm
= oldmm
;
2771 * Since the runqueue lock will be released by the next
2772 * task (which is an invalid locking op but in the case
2773 * of the scheduler it's an obvious special-case), so we
2774 * do an early lockdep release here:
2776 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2777 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2780 /* Here we just switch the register state and the stack. */
2781 switch_to(prev
, next
, prev
);
2785 * this_rq must be evaluated again because prev may have moved
2786 * CPUs since it called schedule(), thus the 'rq' on its stack
2787 * frame will be invalid.
2789 finish_task_switch(this_rq(), prev
);
2793 * nr_running, nr_uninterruptible and nr_context_switches:
2795 * externally visible scheduler statistics: current number of runnable
2796 * threads, current number of uninterruptible-sleeping threads, total
2797 * number of context switches performed since bootup.
2799 unsigned long nr_running(void)
2801 unsigned long i
, sum
= 0;
2803 for_each_online_cpu(i
)
2804 sum
+= cpu_rq(i
)->nr_running
;
2809 unsigned long nr_uninterruptible(void)
2811 unsigned long i
, sum
= 0;
2813 for_each_possible_cpu(i
)
2814 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2817 * Since we read the counters lockless, it might be slightly
2818 * inaccurate. Do not allow it to go below zero though:
2820 if (unlikely((long)sum
< 0))
2826 unsigned long long nr_context_switches(void)
2829 unsigned long long sum
= 0;
2831 for_each_possible_cpu(i
)
2832 sum
+= cpu_rq(i
)->nr_switches
;
2837 unsigned long nr_iowait(void)
2839 unsigned long i
, sum
= 0;
2841 for_each_possible_cpu(i
)
2842 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2847 unsigned long nr_active(void)
2849 unsigned long i
, running
= 0, uninterruptible
= 0;
2851 for_each_online_cpu(i
) {
2852 running
+= cpu_rq(i
)->nr_running
;
2853 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2856 if (unlikely((long)uninterruptible
< 0))
2857 uninterruptible
= 0;
2859 return running
+ uninterruptible
;
2863 * Update rq->cpu_load[] statistics. This function is usually called every
2864 * scheduler tick (TICK_NSEC).
2866 static void update_cpu_load(struct rq
*this_rq
)
2868 unsigned long this_load
= this_rq
->load
.weight
;
2871 this_rq
->nr_load_updates
++;
2873 /* Update our load: */
2874 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2875 unsigned long old_load
, new_load
;
2877 /* scale is effectively 1 << i now, and >> i divides by scale */
2879 old_load
= this_rq
->cpu_load
[i
];
2880 new_load
= this_load
;
2882 * Round up the averaging division if load is increasing. This
2883 * prevents us from getting stuck on 9 if the load is 10, for
2886 if (new_load
> old_load
)
2887 new_load
+= scale
-1;
2888 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2895 * double_rq_lock - safely lock two runqueues
2897 * Note this does not disable interrupts like task_rq_lock,
2898 * you need to do so manually before calling.
2900 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2901 __acquires(rq1
->lock
)
2902 __acquires(rq2
->lock
)
2904 BUG_ON(!irqs_disabled());
2906 spin_lock(&rq1
->lock
);
2907 __acquire(rq2
->lock
); /* Fake it out ;) */
2910 spin_lock(&rq1
->lock
);
2911 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
2913 spin_lock(&rq2
->lock
);
2914 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
2917 update_rq_clock(rq1
);
2918 update_rq_clock(rq2
);
2922 * double_rq_unlock - safely unlock two runqueues
2924 * Note this does not restore interrupts like task_rq_unlock,
2925 * you need to do so manually after calling.
2927 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2928 __releases(rq1
->lock
)
2929 __releases(rq2
->lock
)
2931 spin_unlock(&rq1
->lock
);
2933 spin_unlock(&rq2
->lock
);
2935 __release(rq2
->lock
);
2939 * If dest_cpu is allowed for this process, migrate the task to it.
2940 * This is accomplished by forcing the cpu_allowed mask to only
2941 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2942 * the cpu_allowed mask is restored.
2944 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2946 struct migration_req req
;
2947 unsigned long flags
;
2950 rq
= task_rq_lock(p
, &flags
);
2951 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
2952 || unlikely(!cpu_active(dest_cpu
)))
2955 /* force the process onto the specified CPU */
2956 if (migrate_task(p
, dest_cpu
, &req
)) {
2957 /* Need to wait for migration thread (might exit: take ref). */
2958 struct task_struct
*mt
= rq
->migration_thread
;
2960 get_task_struct(mt
);
2961 task_rq_unlock(rq
, &flags
);
2962 wake_up_process(mt
);
2963 put_task_struct(mt
);
2964 wait_for_completion(&req
.done
);
2969 task_rq_unlock(rq
, &flags
);
2973 * sched_exec - execve() is a valuable balancing opportunity, because at
2974 * this point the task has the smallest effective memory and cache footprint.
2976 void sched_exec(void)
2978 int new_cpu
, this_cpu
= get_cpu();
2979 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2981 if (new_cpu
!= this_cpu
)
2982 sched_migrate_task(current
, new_cpu
);
2986 * pull_task - move a task from a remote runqueue to the local runqueue.
2987 * Both runqueues must be locked.
2989 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2990 struct rq
*this_rq
, int this_cpu
)
2992 deactivate_task(src_rq
, p
, 0);
2993 set_task_cpu(p
, this_cpu
);
2994 activate_task(this_rq
, p
, 0);
2996 * Note that idle threads have a prio of MAX_PRIO, for this test
2997 * to be always true for them.
2999 check_preempt_curr(this_rq
, p
, 0);
3003 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3006 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
3007 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3010 int tsk_cache_hot
= 0;
3012 * We do not migrate tasks that are:
3013 * 1) running (obviously), or
3014 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3015 * 3) are cache-hot on their current CPU.
3017 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
3018 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
3023 if (task_running(rq
, p
)) {
3024 schedstat_inc(p
, se
.nr_failed_migrations_running
);
3029 * Aggressive migration if:
3030 * 1) task is cache cold, or
3031 * 2) too many balance attempts have failed.
3034 tsk_cache_hot
= task_hot(p
, rq
->clock
, sd
);
3035 if (!tsk_cache_hot
||
3036 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
3037 #ifdef CONFIG_SCHEDSTATS
3038 if (tsk_cache_hot
) {
3039 schedstat_inc(sd
, lb_hot_gained
[idle
]);
3040 schedstat_inc(p
, se
.nr_forced_migrations
);
3046 if (tsk_cache_hot
) {
3047 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
3053 static unsigned long
3054 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3055 unsigned long max_load_move
, struct sched_domain
*sd
,
3056 enum cpu_idle_type idle
, int *all_pinned
,
3057 int *this_best_prio
, struct rq_iterator
*iterator
)
3059 int loops
= 0, pulled
= 0, pinned
= 0;
3060 struct task_struct
*p
;
3061 long rem_load_move
= max_load_move
;
3063 if (max_load_move
== 0)
3069 * Start the load-balancing iterator:
3071 p
= iterator
->start(iterator
->arg
);
3073 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3076 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
3077 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3078 p
= iterator
->next(iterator
->arg
);
3082 pull_task(busiest
, p
, this_rq
, this_cpu
);
3084 rem_load_move
-= p
->se
.load
.weight
;
3086 #ifdef CONFIG_PREEMPT
3088 * NEWIDLE balancing is a source of latency, so preemptible kernels
3089 * will stop after the first task is pulled to minimize the critical
3092 if (idle
== CPU_NEWLY_IDLE
)
3097 * We only want to steal up to the prescribed amount of weighted load.
3099 if (rem_load_move
> 0) {
3100 if (p
->prio
< *this_best_prio
)
3101 *this_best_prio
= p
->prio
;
3102 p
= iterator
->next(iterator
->arg
);
3107 * Right now, this is one of only two places pull_task() is called,
3108 * so we can safely collect pull_task() stats here rather than
3109 * inside pull_task().
3111 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3114 *all_pinned
= pinned
;
3116 return max_load_move
- rem_load_move
;
3120 * move_tasks tries to move up to max_load_move weighted load from busiest to
3121 * this_rq, as part of a balancing operation within domain "sd".
3122 * Returns 1 if successful and 0 otherwise.
3124 * Called with both runqueues locked.
3126 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3127 unsigned long max_load_move
,
3128 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3131 const struct sched_class
*class = sched_class_highest
;
3132 unsigned long total_load_moved
= 0;
3133 int this_best_prio
= this_rq
->curr
->prio
;
3137 class->load_balance(this_rq
, this_cpu
, busiest
,
3138 max_load_move
- total_load_moved
,
3139 sd
, idle
, all_pinned
, &this_best_prio
);
3140 class = class->next
;
3142 #ifdef CONFIG_PREEMPT
3144 * NEWIDLE balancing is a source of latency, so preemptible
3145 * kernels will stop after the first task is pulled to minimize
3146 * the critical section.
3148 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3151 } while (class && max_load_move
> total_load_moved
);
3153 return total_load_moved
> 0;
3157 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3158 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3159 struct rq_iterator
*iterator
)
3161 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3165 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3166 pull_task(busiest
, p
, this_rq
, this_cpu
);
3168 * Right now, this is only the second place pull_task()
3169 * is called, so we can safely collect pull_task()
3170 * stats here rather than inside pull_task().
3172 schedstat_inc(sd
, lb_gained
[idle
]);
3176 p
= iterator
->next(iterator
->arg
);
3183 * move_one_task tries to move exactly one task from busiest to this_rq, as
3184 * part of active balancing operations within "domain".
3185 * Returns 1 if successful and 0 otherwise.
3187 * Called with both runqueues locked.
3189 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3190 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3192 const struct sched_class
*class;
3194 for (class = sched_class_highest
; class; class = class->next
)
3195 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3200 /********** Helpers for find_busiest_group ************************/
3202 * sd_lb_stats - Structure to store the statistics of a sched_domain
3203 * during load balancing.
3205 struct sd_lb_stats
{
3206 struct sched_group
*busiest
; /* Busiest group in this sd */
3207 struct sched_group
*this; /* Local group in this sd */
3208 unsigned long total_load
; /* Total load of all groups in sd */
3209 unsigned long total_pwr
; /* Total power of all groups in sd */
3210 unsigned long avg_load
; /* Average load across all groups in sd */
3212 /** Statistics of this group */
3213 unsigned long this_load
;
3214 unsigned long this_load_per_task
;
3215 unsigned long this_nr_running
;
3217 /* Statistics of the busiest group */
3218 unsigned long max_load
;
3219 unsigned long busiest_load_per_task
;
3220 unsigned long busiest_nr_running
;
3222 int group_imb
; /* Is there imbalance in this sd */
3223 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3224 int power_savings_balance
; /* Is powersave balance needed for this sd */
3225 struct sched_group
*group_min
; /* Least loaded group in sd */
3226 struct sched_group
*group_leader
; /* Group which relieves group_min */
3227 unsigned long min_load_per_task
; /* load_per_task in group_min */
3228 unsigned long leader_nr_running
; /* Nr running of group_leader */
3229 unsigned long min_nr_running
; /* Nr running of group_min */
3234 * sg_lb_stats - stats of a sched_group required for load_balancing
3236 struct sg_lb_stats
{
3237 unsigned long avg_load
; /*Avg load across the CPUs of the group */
3238 unsigned long group_load
; /* Total load over the CPUs of the group */
3239 unsigned long sum_nr_running
; /* Nr tasks running in the group */
3240 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
3241 unsigned long group_capacity
;
3242 int group_imb
; /* Is there an imbalance in the group ? */
3246 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3247 * @group: The group whose first cpu is to be returned.
3249 static inline unsigned int group_first_cpu(struct sched_group
*group
)
3251 return cpumask_first(sched_group_cpus(group
));
3255 * get_sd_load_idx - Obtain the load index for a given sched domain.
3256 * @sd: The sched_domain whose load_idx is to be obtained.
3257 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3259 static inline int get_sd_load_idx(struct sched_domain
*sd
,
3260 enum cpu_idle_type idle
)
3266 load_idx
= sd
->busy_idx
;
3269 case CPU_NEWLY_IDLE
:
3270 load_idx
= sd
->newidle_idx
;
3273 load_idx
= sd
->idle_idx
;
3281 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3283 * init_sd_power_savings_stats - Initialize power savings statistics for
3284 * the given sched_domain, during load balancing.
3286 * @sd: Sched domain whose power-savings statistics are to be initialized.
3287 * @sds: Variable containing the statistics for sd.
3288 * @idle: Idle status of the CPU at which we're performing load-balancing.
3290 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3291 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3294 * Busy processors will not participate in power savings
3297 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3298 sds
->power_savings_balance
= 0;
3300 sds
->power_savings_balance
= 1;
3301 sds
->min_nr_running
= ULONG_MAX
;
3302 sds
->leader_nr_running
= 0;
3307 * update_sd_power_savings_stats - Update the power saving stats for a
3308 * sched_domain while performing load balancing.
3310 * @group: sched_group belonging to the sched_domain under consideration.
3311 * @sds: Variable containing the statistics of the sched_domain
3312 * @local_group: Does group contain the CPU for which we're performing
3314 * @sgs: Variable containing the statistics of the group.
3316 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3317 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3320 if (!sds
->power_savings_balance
)
3324 * If the local group is idle or completely loaded
3325 * no need to do power savings balance at this domain
3327 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
3328 !sds
->this_nr_running
))
3329 sds
->power_savings_balance
= 0;
3332 * If a group is already running at full capacity or idle,
3333 * don't include that group in power savings calculations
3335 if (!sds
->power_savings_balance
||
3336 sgs
->sum_nr_running
>= sgs
->group_capacity
||
3337 !sgs
->sum_nr_running
)
3341 * Calculate the group which has the least non-idle load.
3342 * This is the group from where we need to pick up the load
3345 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
3346 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
3347 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
3348 sds
->group_min
= group
;
3349 sds
->min_nr_running
= sgs
->sum_nr_running
;
3350 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
3351 sgs
->sum_nr_running
;
3355 * Calculate the group which is almost near its
3356 * capacity but still has some space to pick up some load
3357 * from other group and save more power
3359 if (sgs
->sum_nr_running
> sgs
->group_capacity
- 1)
3362 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
3363 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
3364 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
3365 sds
->group_leader
= group
;
3366 sds
->leader_nr_running
= sgs
->sum_nr_running
;
3371 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3372 * @sds: Variable containing the statistics of the sched_domain
3373 * under consideration.
3374 * @this_cpu: Cpu at which we're currently performing load-balancing.
3375 * @imbalance: Variable to store the imbalance.
3378 * Check if we have potential to perform some power-savings balance.
3379 * If yes, set the busiest group to be the least loaded group in the
3380 * sched_domain, so that it's CPUs can be put to idle.
3382 * Returns 1 if there is potential to perform power-savings balance.
3385 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3386 int this_cpu
, unsigned long *imbalance
)
3388 if (!sds
->power_savings_balance
)
3391 if (sds
->this != sds
->group_leader
||
3392 sds
->group_leader
== sds
->group_min
)
3395 *imbalance
= sds
->min_load_per_task
;
3396 sds
->busiest
= sds
->group_min
;
3398 if (sched_mc_power_savings
>= POWERSAVINGS_BALANCE_WAKEUP
) {
3399 cpu_rq(this_cpu
)->rd
->sched_mc_preferred_wakeup_cpu
=
3400 group_first_cpu(sds
->group_leader
);
3406 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3407 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3408 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3413 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3414 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3419 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3420 int this_cpu
, unsigned long *imbalance
)
3424 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3428 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3429 * @group: sched_group whose statistics are to be updated.
3430 * @this_cpu: Cpu for which load balance is currently performed.
3431 * @idle: Idle status of this_cpu
3432 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3433 * @sd_idle: Idle status of the sched_domain containing group.
3434 * @local_group: Does group contain this_cpu.
3435 * @cpus: Set of cpus considered for load balancing.
3436 * @balance: Should we balance.
3437 * @sgs: variable to hold the statistics for this group.
3439 static inline void update_sg_lb_stats(struct sched_group
*group
, int this_cpu
,
3440 enum cpu_idle_type idle
, int load_idx
, int *sd_idle
,
3441 int local_group
, const struct cpumask
*cpus
,
3442 int *balance
, struct sg_lb_stats
*sgs
)
3444 unsigned long load
, max_cpu_load
, min_cpu_load
;
3446 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3447 unsigned long sum_avg_load_per_task
;
3448 unsigned long avg_load_per_task
;
3451 balance_cpu
= group_first_cpu(group
);
3453 /* Tally up the load of all CPUs in the group */
3454 sum_avg_load_per_task
= avg_load_per_task
= 0;
3456 min_cpu_load
= ~0UL;
3458 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3459 struct rq
*rq
= cpu_rq(i
);
3461 if (*sd_idle
&& rq
->nr_running
)
3464 /* Bias balancing toward cpus of our domain */
3466 if (idle_cpu(i
) && !first_idle_cpu
) {
3471 load
= target_load(i
, load_idx
);
3473 load
= source_load(i
, load_idx
);
3474 if (load
> max_cpu_load
)
3475 max_cpu_load
= load
;
3476 if (min_cpu_load
> load
)
3477 min_cpu_load
= load
;
3480 sgs
->group_load
+= load
;
3481 sgs
->sum_nr_running
+= rq
->nr_running
;
3482 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
3484 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3488 * First idle cpu or the first cpu(busiest) in this sched group
3489 * is eligible for doing load balancing at this and above
3490 * domains. In the newly idle case, we will allow all the cpu's
3491 * to do the newly idle load balance.
3493 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3494 balance_cpu
!= this_cpu
&& balance
) {
3499 /* Adjust by relative CPU power of the group */
3500 sgs
->avg_load
= sg_div_cpu_power(group
,
3501 sgs
->group_load
* SCHED_LOAD_SCALE
);
3505 * Consider the group unbalanced when the imbalance is larger
3506 * than the average weight of two tasks.
3508 * APZ: with cgroup the avg task weight can vary wildly and
3509 * might not be a suitable number - should we keep a
3510 * normalized nr_running number somewhere that negates
3513 avg_load_per_task
= sg_div_cpu_power(group
,
3514 sum_avg_load_per_task
* SCHED_LOAD_SCALE
);
3516 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3519 sgs
->group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3524 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3525 * @sd: sched_domain whose statistics are to be updated.
3526 * @this_cpu: Cpu for which load balance is currently performed.
3527 * @idle: Idle status of this_cpu
3528 * @sd_idle: Idle status of the sched_domain containing group.
3529 * @cpus: Set of cpus considered for load balancing.
3530 * @balance: Should we balance.
3531 * @sds: variable to hold the statistics for this sched_domain.
3533 static inline void update_sd_lb_stats(struct sched_domain
*sd
, int this_cpu
,
3534 enum cpu_idle_type idle
, int *sd_idle
,
3535 const struct cpumask
*cpus
, int *balance
,
3536 struct sd_lb_stats
*sds
)
3538 struct sched_group
*group
= sd
->groups
;
3539 struct sg_lb_stats sgs
;
3542 init_sd_power_savings_stats(sd
, sds
, idle
);
3543 load_idx
= get_sd_load_idx(sd
, idle
);
3548 local_group
= cpumask_test_cpu(this_cpu
,
3549 sched_group_cpus(group
));
3550 memset(&sgs
, 0, sizeof(sgs
));
3551 update_sg_lb_stats(group
, this_cpu
, idle
, load_idx
, sd_idle
,
3552 local_group
, cpus
, balance
, &sgs
);
3554 if (local_group
&& balance
&& !(*balance
))
3557 sds
->total_load
+= sgs
.group_load
;
3558 sds
->total_pwr
+= group
->__cpu_power
;
3561 sds
->this_load
= sgs
.avg_load
;
3563 sds
->this_nr_running
= sgs
.sum_nr_running
;
3564 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
3565 } else if (sgs
.avg_load
> sds
->max_load
&&
3566 (sgs
.sum_nr_running
> sgs
.group_capacity
||
3568 sds
->max_load
= sgs
.avg_load
;
3569 sds
->busiest
= group
;
3570 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
3571 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
3572 sds
->group_imb
= sgs
.group_imb
;
3575 update_sd_power_savings_stats(group
, sds
, local_group
, &sgs
);
3576 group
= group
->next
;
3577 } while (group
!= sd
->groups
);
3582 * fix_small_imbalance - Calculate the minor imbalance that exists
3583 * amongst the groups of a sched_domain, during
3585 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3586 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3587 * @imbalance: Variable to store the imbalance.
3589 static inline void fix_small_imbalance(struct sd_lb_stats
*sds
,
3590 int this_cpu
, unsigned long *imbalance
)
3592 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
3593 unsigned int imbn
= 2;
3595 if (sds
->this_nr_running
) {
3596 sds
->this_load_per_task
/= sds
->this_nr_running
;
3597 if (sds
->busiest_load_per_task
>
3598 sds
->this_load_per_task
)
3601 sds
->this_load_per_task
=
3602 cpu_avg_load_per_task(this_cpu
);
3604 if (sds
->max_load
- sds
->this_load
+ sds
->busiest_load_per_task
>=
3605 sds
->busiest_load_per_task
* imbn
) {
3606 *imbalance
= sds
->busiest_load_per_task
;
3611 * OK, we don't have enough imbalance to justify moving tasks,
3612 * however we may be able to increase total CPU power used by
3616 pwr_now
+= sds
->busiest
->__cpu_power
*
3617 min(sds
->busiest_load_per_task
, sds
->max_load
);
3618 pwr_now
+= sds
->this->__cpu_power
*
3619 min(sds
->this_load_per_task
, sds
->this_load
);
3620 pwr_now
/= SCHED_LOAD_SCALE
;
3622 /* Amount of load we'd subtract */
3623 tmp
= sg_div_cpu_power(sds
->busiest
,
3624 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
);
3625 if (sds
->max_load
> tmp
)
3626 pwr_move
+= sds
->busiest
->__cpu_power
*
3627 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
3629 /* Amount of load we'd add */
3630 if (sds
->max_load
* sds
->busiest
->__cpu_power
<
3631 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
)
3632 tmp
= sg_div_cpu_power(sds
->this,
3633 sds
->max_load
* sds
->busiest
->__cpu_power
);
3635 tmp
= sg_div_cpu_power(sds
->this,
3636 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
);
3637 pwr_move
+= sds
->this->__cpu_power
*
3638 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
3639 pwr_move
/= SCHED_LOAD_SCALE
;
3641 /* Move if we gain throughput */
3642 if (pwr_move
> pwr_now
)
3643 *imbalance
= sds
->busiest_load_per_task
;
3647 * calculate_imbalance - Calculate the amount of imbalance present within the
3648 * groups of a given sched_domain during load balance.
3649 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3650 * @this_cpu: Cpu for which currently load balance is being performed.
3651 * @imbalance: The variable to store the imbalance.
3653 static inline void calculate_imbalance(struct sd_lb_stats
*sds
, int this_cpu
,
3654 unsigned long *imbalance
)
3656 unsigned long max_pull
;
3658 * In the presence of smp nice balancing, certain scenarios can have
3659 * max load less than avg load(as we skip the groups at or below
3660 * its cpu_power, while calculating max_load..)
3662 if (sds
->max_load
< sds
->avg_load
) {
3664 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3667 /* Don't want to pull so many tasks that a group would go idle */
3668 max_pull
= min(sds
->max_load
- sds
->avg_load
,
3669 sds
->max_load
- sds
->busiest_load_per_task
);
3671 /* How much load to actually move to equalise the imbalance */
3672 *imbalance
= min(max_pull
* sds
->busiest
->__cpu_power
,
3673 (sds
->avg_load
- sds
->this_load
) * sds
->this->__cpu_power
)
3677 * if *imbalance is less than the average load per runnable task
3678 * there is no gaurantee that any tasks will be moved so we'll have
3679 * a think about bumping its value to force at least one task to be
3682 if (*imbalance
< sds
->busiest_load_per_task
)
3683 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3686 /******* find_busiest_group() helpers end here *********************/
3689 * find_busiest_group - Returns the busiest group within the sched_domain
3690 * if there is an imbalance. If there isn't an imbalance, and
3691 * the user has opted for power-savings, it returns a group whose
3692 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3693 * such a group exists.
3695 * Also calculates the amount of weighted load which should be moved
3696 * to restore balance.
3698 * @sd: The sched_domain whose busiest group is to be returned.
3699 * @this_cpu: The cpu for which load balancing is currently being performed.
3700 * @imbalance: Variable which stores amount of weighted load which should
3701 * be moved to restore balance/put a group to idle.
3702 * @idle: The idle status of this_cpu.
3703 * @sd_idle: The idleness of sd
3704 * @cpus: The set of CPUs under consideration for load-balancing.
3705 * @balance: Pointer to a variable indicating if this_cpu
3706 * is the appropriate cpu to perform load balancing at this_level.
3708 * Returns: - the busiest group if imbalance exists.
3709 * - If no imbalance and user has opted for power-savings balance,
3710 * return the least loaded group whose CPUs can be
3711 * put to idle by rebalancing its tasks onto our group.
3713 static struct sched_group
*
3714 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3715 unsigned long *imbalance
, enum cpu_idle_type idle
,
3716 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
3718 struct sd_lb_stats sds
;
3720 memset(&sds
, 0, sizeof(sds
));
3723 * Compute the various statistics relavent for load balancing at
3726 update_sd_lb_stats(sd
, this_cpu
, idle
, sd_idle
, cpus
,
3729 /* Cases where imbalance does not exist from POV of this_cpu */
3730 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3732 * 2) There is no busy sibling group to pull from.
3733 * 3) This group is the busiest group.
3734 * 4) This group is more busy than the avg busieness at this
3736 * 5) The imbalance is within the specified limit.
3737 * 6) Any rebalance would lead to ping-pong
3739 if (balance
&& !(*balance
))
3742 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
3745 if (sds
.this_load
>= sds
.max_load
)
3748 sds
.avg_load
= (SCHED_LOAD_SCALE
* sds
.total_load
) / sds
.total_pwr
;
3750 if (sds
.this_load
>= sds
.avg_load
)
3753 if (100 * sds
.max_load
<= sd
->imbalance_pct
* sds
.this_load
)
3756 sds
.busiest_load_per_task
/= sds
.busiest_nr_running
;
3758 sds
.busiest_load_per_task
=
3759 min(sds
.busiest_load_per_task
, sds
.avg_load
);
3762 * We're trying to get all the cpus to the average_load, so we don't
3763 * want to push ourselves above the average load, nor do we wish to
3764 * reduce the max loaded cpu below the average load, as either of these
3765 * actions would just result in more rebalancing later, and ping-pong
3766 * tasks around. Thus we look for the minimum possible imbalance.
3767 * Negative imbalances (*we* are more loaded than anyone else) will
3768 * be counted as no imbalance for these purposes -- we can't fix that
3769 * by pulling tasks to us. Be careful of negative numbers as they'll
3770 * appear as very large values with unsigned longs.
3772 if (sds
.max_load
<= sds
.busiest_load_per_task
)
3775 /* Looks like there is an imbalance. Compute it */
3776 calculate_imbalance(&sds
, this_cpu
, imbalance
);
3781 * There is no obvious imbalance. But check if we can do some balancing
3784 if (check_power_save_busiest_group(&sds
, this_cpu
, imbalance
))
3792 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3795 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3796 unsigned long imbalance
, const struct cpumask
*cpus
)
3798 struct rq
*busiest
= NULL
, *rq
;
3799 unsigned long max_load
= 0;
3802 for_each_cpu(i
, sched_group_cpus(group
)) {
3805 if (!cpumask_test_cpu(i
, cpus
))
3809 wl
= weighted_cpuload(i
);
3811 if (rq
->nr_running
== 1 && wl
> imbalance
)
3814 if (wl
> max_load
) {
3824 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3825 * so long as it is large enough.
3827 #define MAX_PINNED_INTERVAL 512
3830 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3831 * tasks if there is an imbalance.
3833 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3834 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3835 int *balance
, struct cpumask
*cpus
)
3837 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3838 struct sched_group
*group
;
3839 unsigned long imbalance
;
3841 unsigned long flags
;
3843 cpumask_setall(cpus
);
3846 * When power savings policy is enabled for the parent domain, idle
3847 * sibling can pick up load irrespective of busy siblings. In this case,
3848 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3849 * portraying it as CPU_NOT_IDLE.
3851 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3852 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3855 schedstat_inc(sd
, lb_count
[idle
]);
3859 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3866 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3870 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3872 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3876 BUG_ON(busiest
== this_rq
);
3878 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3881 if (busiest
->nr_running
> 1) {
3883 * Attempt to move tasks. If find_busiest_group has found
3884 * an imbalance but busiest->nr_running <= 1, the group is
3885 * still unbalanced. ld_moved simply stays zero, so it is
3886 * correctly treated as an imbalance.
3888 local_irq_save(flags
);
3889 double_rq_lock(this_rq
, busiest
);
3890 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3891 imbalance
, sd
, idle
, &all_pinned
);
3892 double_rq_unlock(this_rq
, busiest
);
3893 local_irq_restore(flags
);
3896 * some other cpu did the load balance for us.
3898 if (ld_moved
&& this_cpu
!= smp_processor_id())
3899 resched_cpu(this_cpu
);
3901 /* All tasks on this runqueue were pinned by CPU affinity */
3902 if (unlikely(all_pinned
)) {
3903 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
3904 if (!cpumask_empty(cpus
))
3911 schedstat_inc(sd
, lb_failed
[idle
]);
3912 sd
->nr_balance_failed
++;
3914 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3916 spin_lock_irqsave(&busiest
->lock
, flags
);
3918 /* don't kick the migration_thread, if the curr
3919 * task on busiest cpu can't be moved to this_cpu
3921 if (!cpumask_test_cpu(this_cpu
,
3922 &busiest
->curr
->cpus_allowed
)) {
3923 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3925 goto out_one_pinned
;
3928 if (!busiest
->active_balance
) {
3929 busiest
->active_balance
= 1;
3930 busiest
->push_cpu
= this_cpu
;
3933 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3935 wake_up_process(busiest
->migration_thread
);
3938 * We've kicked active balancing, reset the failure
3941 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3944 sd
->nr_balance_failed
= 0;
3946 if (likely(!active_balance
)) {
3947 /* We were unbalanced, so reset the balancing interval */
3948 sd
->balance_interval
= sd
->min_interval
;
3951 * If we've begun active balancing, start to back off. This
3952 * case may not be covered by the all_pinned logic if there
3953 * is only 1 task on the busy runqueue (because we don't call
3956 if (sd
->balance_interval
< sd
->max_interval
)
3957 sd
->balance_interval
*= 2;
3960 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3961 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3967 schedstat_inc(sd
, lb_balanced
[idle
]);
3969 sd
->nr_balance_failed
= 0;
3972 /* tune up the balancing interval */
3973 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3974 (sd
->balance_interval
< sd
->max_interval
))
3975 sd
->balance_interval
*= 2;
3977 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3978 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3989 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3990 * tasks if there is an imbalance.
3992 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3993 * this_rq is locked.
3996 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
,
3997 struct cpumask
*cpus
)
3999 struct sched_group
*group
;
4000 struct rq
*busiest
= NULL
;
4001 unsigned long imbalance
;
4006 cpumask_setall(cpus
);
4009 * When power savings policy is enabled for the parent domain, idle
4010 * sibling can pick up load irrespective of busy siblings. In this case,
4011 * let the state of idle sibling percolate up as IDLE, instead of
4012 * portraying it as CPU_NOT_IDLE.
4014 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
4015 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4018 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
4020 update_shares_locked(this_rq
, sd
);
4021 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
4022 &sd_idle
, cpus
, NULL
);
4024 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
4028 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
4030 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
4034 BUG_ON(busiest
== this_rq
);
4036 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
4039 if (busiest
->nr_running
> 1) {
4040 /* Attempt to move tasks */
4041 double_lock_balance(this_rq
, busiest
);
4042 /* this_rq->clock is already updated */
4043 update_rq_clock(busiest
);
4044 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4045 imbalance
, sd
, CPU_NEWLY_IDLE
,
4047 double_unlock_balance(this_rq
, busiest
);
4049 if (unlikely(all_pinned
)) {
4050 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4051 if (!cpumask_empty(cpus
))
4057 int active_balance
= 0;
4059 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
4060 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4061 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4064 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
4067 if (sd
->nr_balance_failed
++ < 2)
4071 * The only task running in a non-idle cpu can be moved to this
4072 * cpu in an attempt to completely freeup the other CPU
4073 * package. The same method used to move task in load_balance()
4074 * have been extended for load_balance_newidle() to speedup
4075 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4077 * The package power saving logic comes from
4078 * find_busiest_group(). If there are no imbalance, then
4079 * f_b_g() will return NULL. However when sched_mc={1,2} then
4080 * f_b_g() will select a group from which a running task may be
4081 * pulled to this cpu in order to make the other package idle.
4082 * If there is no opportunity to make a package idle and if
4083 * there are no imbalance, then f_b_g() will return NULL and no
4084 * action will be taken in load_balance_newidle().
4086 * Under normal task pull operation due to imbalance, there
4087 * will be more than one task in the source run queue and
4088 * move_tasks() will succeed. ld_moved will be true and this
4089 * active balance code will not be triggered.
4092 /* Lock busiest in correct order while this_rq is held */
4093 double_lock_balance(this_rq
, busiest
);
4096 * don't kick the migration_thread, if the curr
4097 * task on busiest cpu can't be moved to this_cpu
4099 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
4100 double_unlock_balance(this_rq
, busiest
);
4105 if (!busiest
->active_balance
) {
4106 busiest
->active_balance
= 1;
4107 busiest
->push_cpu
= this_cpu
;
4111 double_unlock_balance(this_rq
, busiest
);
4113 * Should not call ttwu while holding a rq->lock
4115 spin_unlock(&this_rq
->lock
);
4117 wake_up_process(busiest
->migration_thread
);
4118 spin_lock(&this_rq
->lock
);
4121 sd
->nr_balance_failed
= 0;
4123 update_shares_locked(this_rq
, sd
);
4127 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
4128 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4129 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4131 sd
->nr_balance_failed
= 0;
4137 * idle_balance is called by schedule() if this_cpu is about to become
4138 * idle. Attempts to pull tasks from other CPUs.
4140 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
4142 struct sched_domain
*sd
;
4143 int pulled_task
= 0;
4144 unsigned long next_balance
= jiffies
+ HZ
;
4145 cpumask_var_t tmpmask
;
4147 if (!alloc_cpumask_var(&tmpmask
, GFP_ATOMIC
))
4150 for_each_domain(this_cpu
, sd
) {
4151 unsigned long interval
;
4153 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4156 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
4157 /* If we've pulled tasks over stop searching: */
4158 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
4161 interval
= msecs_to_jiffies(sd
->balance_interval
);
4162 if (time_after(next_balance
, sd
->last_balance
+ interval
))
4163 next_balance
= sd
->last_balance
+ interval
;
4167 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
4169 * We are going idle. next_balance may be set based on
4170 * a busy processor. So reset next_balance.
4172 this_rq
->next_balance
= next_balance
;
4174 free_cpumask_var(tmpmask
);
4178 * active_load_balance is run by migration threads. It pushes running tasks
4179 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4180 * running on each physical CPU where possible, and avoids physical /
4181 * logical imbalances.
4183 * Called with busiest_rq locked.
4185 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
4187 int target_cpu
= busiest_rq
->push_cpu
;
4188 struct sched_domain
*sd
;
4189 struct rq
*target_rq
;
4191 /* Is there any task to move? */
4192 if (busiest_rq
->nr_running
<= 1)
4195 target_rq
= cpu_rq(target_cpu
);
4198 * This condition is "impossible", if it occurs
4199 * we need to fix it. Originally reported by
4200 * Bjorn Helgaas on a 128-cpu setup.
4202 BUG_ON(busiest_rq
== target_rq
);
4204 /* move a task from busiest_rq to target_rq */
4205 double_lock_balance(busiest_rq
, target_rq
);
4206 update_rq_clock(busiest_rq
);
4207 update_rq_clock(target_rq
);
4209 /* Search for an sd spanning us and the target CPU. */
4210 for_each_domain(target_cpu
, sd
) {
4211 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
4212 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
4217 schedstat_inc(sd
, alb_count
);
4219 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
4221 schedstat_inc(sd
, alb_pushed
);
4223 schedstat_inc(sd
, alb_failed
);
4225 double_unlock_balance(busiest_rq
, target_rq
);
4230 atomic_t load_balancer
;
4231 cpumask_var_t cpu_mask
;
4232 } nohz ____cacheline_aligned
= {
4233 .load_balancer
= ATOMIC_INIT(-1),
4237 * This routine will try to nominate the ilb (idle load balancing)
4238 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4239 * load balancing on behalf of all those cpus. If all the cpus in the system
4240 * go into this tickless mode, then there will be no ilb owner (as there is
4241 * no need for one) and all the cpus will sleep till the next wakeup event
4244 * For the ilb owner, tick is not stopped. And this tick will be used
4245 * for idle load balancing. ilb owner will still be part of
4248 * While stopping the tick, this cpu will become the ilb owner if there
4249 * is no other owner. And will be the owner till that cpu becomes busy
4250 * or if all cpus in the system stop their ticks at which point
4251 * there is no need for ilb owner.
4253 * When the ilb owner becomes busy, it nominates another owner, during the
4254 * next busy scheduler_tick()
4256 int select_nohz_load_balancer(int stop_tick
)
4258 int cpu
= smp_processor_id();
4261 cpu_rq(cpu
)->in_nohz_recently
= 1;
4263 if (!cpu_active(cpu
)) {
4264 if (atomic_read(&nohz
.load_balancer
) != cpu
)
4268 * If we are going offline and still the leader,
4271 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4277 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
4279 /* time for ilb owner also to sleep */
4280 if (cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4281 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4282 atomic_set(&nohz
.load_balancer
, -1);
4286 if (atomic_read(&nohz
.load_balancer
) == -1) {
4287 /* make me the ilb owner */
4288 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
4290 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
4293 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4296 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4298 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4299 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4306 static DEFINE_SPINLOCK(balancing
);
4309 * It checks each scheduling domain to see if it is due to be balanced,
4310 * and initiates a balancing operation if so.
4312 * Balancing parameters are set up in arch_init_sched_domains.
4314 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
4317 struct rq
*rq
= cpu_rq(cpu
);
4318 unsigned long interval
;
4319 struct sched_domain
*sd
;
4320 /* Earliest time when we have to do rebalance again */
4321 unsigned long next_balance
= jiffies
+ 60*HZ
;
4322 int update_next_balance
= 0;
4326 /* Fails alloc? Rebalancing probably not a priority right now. */
4327 if (!alloc_cpumask_var(&tmp
, GFP_ATOMIC
))
4330 for_each_domain(cpu
, sd
) {
4331 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4334 interval
= sd
->balance_interval
;
4335 if (idle
!= CPU_IDLE
)
4336 interval
*= sd
->busy_factor
;
4338 /* scale ms to jiffies */
4339 interval
= msecs_to_jiffies(interval
);
4340 if (unlikely(!interval
))
4342 if (interval
> HZ
*NR_CPUS
/10)
4343 interval
= HZ
*NR_CPUS
/10;
4345 need_serialize
= sd
->flags
& SD_SERIALIZE
;
4347 if (need_serialize
) {
4348 if (!spin_trylock(&balancing
))
4352 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4353 if (load_balance(cpu
, rq
, sd
, idle
, &balance
, tmp
)) {
4355 * We've pulled tasks over so either we're no
4356 * longer idle, or one of our SMT siblings is
4359 idle
= CPU_NOT_IDLE
;
4361 sd
->last_balance
= jiffies
;
4364 spin_unlock(&balancing
);
4366 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4367 next_balance
= sd
->last_balance
+ interval
;
4368 update_next_balance
= 1;
4372 * Stop the load balance at this level. There is another
4373 * CPU in our sched group which is doing load balancing more
4381 * next_balance will be updated only when there is a need.
4382 * When the cpu is attached to null domain for ex, it will not be
4385 if (likely(update_next_balance
))
4386 rq
->next_balance
= next_balance
;
4388 free_cpumask_var(tmp
);
4392 * run_rebalance_domains is triggered when needed from the scheduler tick.
4393 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4394 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4396 static void run_rebalance_domains(struct softirq_action
*h
)
4398 int this_cpu
= smp_processor_id();
4399 struct rq
*this_rq
= cpu_rq(this_cpu
);
4400 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4401 CPU_IDLE
: CPU_NOT_IDLE
;
4403 rebalance_domains(this_cpu
, idle
);
4407 * If this cpu is the owner for idle load balancing, then do the
4408 * balancing on behalf of the other idle cpus whose ticks are
4411 if (this_rq
->idle_at_tick
&&
4412 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4416 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
4417 if (balance_cpu
== this_cpu
)
4421 * If this cpu gets work to do, stop the load balancing
4422 * work being done for other cpus. Next load
4423 * balancing owner will pick it up.
4428 rebalance_domains(balance_cpu
, CPU_IDLE
);
4430 rq
= cpu_rq(balance_cpu
);
4431 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4432 this_rq
->next_balance
= rq
->next_balance
;
4438 static inline int on_null_domain(int cpu
)
4440 return !rcu_dereference(cpu_rq(cpu
)->sd
);
4444 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4446 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4447 * idle load balancing owner or decide to stop the periodic load balancing,
4448 * if the whole system is idle.
4450 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4454 * If we were in the nohz mode recently and busy at the current
4455 * scheduler tick, then check if we need to nominate new idle
4458 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4459 rq
->in_nohz_recently
= 0;
4461 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4462 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4463 atomic_set(&nohz
.load_balancer
, -1);
4466 if (atomic_read(&nohz
.load_balancer
) == -1) {
4468 * simple selection for now: Nominate the
4469 * first cpu in the nohz list to be the next
4472 * TBD: Traverse the sched domains and nominate
4473 * the nearest cpu in the nohz.cpu_mask.
4475 int ilb
= cpumask_first(nohz
.cpu_mask
);
4477 if (ilb
< nr_cpu_ids
)
4483 * If this cpu is idle and doing idle load balancing for all the
4484 * cpus with ticks stopped, is it time for that to stop?
4486 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4487 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4493 * If this cpu is idle and the idle load balancing is done by
4494 * someone else, then no need raise the SCHED_SOFTIRQ
4496 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4497 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4500 /* Don't need to rebalance while attached to NULL domain */
4501 if (time_after_eq(jiffies
, rq
->next_balance
) &&
4502 likely(!on_null_domain(cpu
)))
4503 raise_softirq(SCHED_SOFTIRQ
);
4506 #else /* CONFIG_SMP */
4509 * on UP we do not need to balance between CPUs:
4511 static inline void idle_balance(int cpu
, struct rq
*rq
)
4517 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4519 EXPORT_PER_CPU_SYMBOL(kstat
);
4522 * Return any ns on the sched_clock that have not yet been banked in
4523 * @p in case that task is currently running.
4525 unsigned long long task_delta_exec(struct task_struct
*p
)
4527 unsigned long flags
;
4531 rq
= task_rq_lock(p
, &flags
);
4533 if (task_current(rq
, p
)) {
4536 update_rq_clock(rq
);
4537 delta_exec
= rq
->clock
- p
->se
.exec_start
;
4538 if ((s64
)delta_exec
> 0)
4542 task_rq_unlock(rq
, &flags
);
4548 * Account user cpu time to a process.
4549 * @p: the process that the cpu time gets accounted to
4550 * @cputime: the cpu time spent in user space since the last update
4551 * @cputime_scaled: cputime scaled by cpu frequency
4553 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
4554 cputime_t cputime_scaled
)
4556 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4559 /* Add user time to process. */
4560 p
->utime
= cputime_add(p
->utime
, cputime
);
4561 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4562 account_group_user_time(p
, cputime
);
4564 /* Add user time to cpustat. */
4565 tmp
= cputime_to_cputime64(cputime
);
4566 if (TASK_NICE(p
) > 0)
4567 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4569 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4570 /* Account for user time used */
4571 acct_update_integrals(p
);
4575 * Account guest cpu time to a process.
4576 * @p: the process that the cpu time gets accounted to
4577 * @cputime: the cpu time spent in virtual machine since the last update
4578 * @cputime_scaled: cputime scaled by cpu frequency
4580 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
4581 cputime_t cputime_scaled
)
4584 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4586 tmp
= cputime_to_cputime64(cputime
);
4588 /* Add guest time to process. */
4589 p
->utime
= cputime_add(p
->utime
, cputime
);
4590 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4591 account_group_user_time(p
, cputime
);
4592 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4594 /* Add guest time to cpustat. */
4595 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4596 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4600 * Account system cpu time to a process.
4601 * @p: the process that the cpu time gets accounted to
4602 * @hardirq_offset: the offset to subtract from hardirq_count()
4603 * @cputime: the cpu time spent in kernel space since the last update
4604 * @cputime_scaled: cputime scaled by cpu frequency
4606 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4607 cputime_t cputime
, cputime_t cputime_scaled
)
4609 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4612 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4613 account_guest_time(p
, cputime
, cputime_scaled
);
4617 /* Add system time to process. */
4618 p
->stime
= cputime_add(p
->stime
, cputime
);
4619 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
4620 account_group_system_time(p
, cputime
);
4622 /* Add system time to cpustat. */
4623 tmp
= cputime_to_cputime64(cputime
);
4624 if (hardirq_count() - hardirq_offset
)
4625 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4626 else if (softirq_count())
4627 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4629 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4631 /* Account for system time used */
4632 acct_update_integrals(p
);
4636 * Account for involuntary wait time.
4637 * @steal: the cpu time spent in involuntary wait
4639 void account_steal_time(cputime_t cputime
)
4641 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4642 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
4644 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
4648 * Account for idle time.
4649 * @cputime: the cpu time spent in idle wait
4651 void account_idle_time(cputime_t cputime
)
4653 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4654 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
4655 struct rq
*rq
= this_rq();
4657 if (atomic_read(&rq
->nr_iowait
) > 0)
4658 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
4660 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
4663 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4666 * Account a single tick of cpu time.
4667 * @p: the process that the cpu time gets accounted to
4668 * @user_tick: indicates if the tick is a user or a system tick
4670 void account_process_tick(struct task_struct
*p
, int user_tick
)
4672 cputime_t one_jiffy
= jiffies_to_cputime(1);
4673 cputime_t one_jiffy_scaled
= cputime_to_scaled(one_jiffy
);
4674 struct rq
*rq
= this_rq();
4677 account_user_time(p
, one_jiffy
, one_jiffy_scaled
);
4678 else if (p
!= rq
->idle
)
4679 account_system_time(p
, HARDIRQ_OFFSET
, one_jiffy
,
4682 account_idle_time(one_jiffy
);
4686 * Account multiple ticks of steal time.
4687 * @p: the process from which the cpu time has been stolen
4688 * @ticks: number of stolen ticks
4690 void account_steal_ticks(unsigned long ticks
)
4692 account_steal_time(jiffies_to_cputime(ticks
));
4696 * Account multiple ticks of idle time.
4697 * @ticks: number of stolen ticks
4699 void account_idle_ticks(unsigned long ticks
)
4701 account_idle_time(jiffies_to_cputime(ticks
));
4707 * Use precise platform statistics if available:
4709 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4710 cputime_t
task_utime(struct task_struct
*p
)
4715 cputime_t
task_stime(struct task_struct
*p
)
4720 cputime_t
task_utime(struct task_struct
*p
)
4722 clock_t utime
= cputime_to_clock_t(p
->utime
),
4723 total
= utime
+ cputime_to_clock_t(p
->stime
);
4727 * Use CFS's precise accounting:
4729 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
4733 do_div(temp
, total
);
4735 utime
= (clock_t)temp
;
4737 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
4738 return p
->prev_utime
;
4741 cputime_t
task_stime(struct task_struct
*p
)
4746 * Use CFS's precise accounting. (we subtract utime from
4747 * the total, to make sure the total observed by userspace
4748 * grows monotonically - apps rely on that):
4750 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
4751 cputime_to_clock_t(task_utime(p
));
4754 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
4756 return p
->prev_stime
;
4760 inline cputime_t
task_gtime(struct task_struct
*p
)
4766 * This function gets called by the timer code, with HZ frequency.
4767 * We call it with interrupts disabled.
4769 * It also gets called by the fork code, when changing the parent's
4772 void scheduler_tick(void)
4774 int cpu
= smp_processor_id();
4775 struct rq
*rq
= cpu_rq(cpu
);
4776 struct task_struct
*curr
= rq
->curr
;
4780 spin_lock(&rq
->lock
);
4781 update_rq_clock(rq
);
4782 update_cpu_load(rq
);
4783 curr
->sched_class
->task_tick(rq
, curr
, 0);
4784 spin_unlock(&rq
->lock
);
4787 rq
->idle_at_tick
= idle_cpu(cpu
);
4788 trigger_load_balance(rq
, cpu
);
4792 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4793 defined(CONFIG_PREEMPT_TRACER))
4795 static inline unsigned long get_parent_ip(unsigned long addr
)
4797 if (in_lock_functions(addr
)) {
4798 addr
= CALLER_ADDR2
;
4799 if (in_lock_functions(addr
))
4800 addr
= CALLER_ADDR3
;
4805 void __kprobes
add_preempt_count(int val
)
4807 #ifdef CONFIG_DEBUG_PREEMPT
4811 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4814 preempt_count() += val
;
4815 #ifdef CONFIG_DEBUG_PREEMPT
4817 * Spinlock count overflowing soon?
4819 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4822 if (preempt_count() == val
)
4823 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4825 EXPORT_SYMBOL(add_preempt_count
);
4827 void __kprobes
sub_preempt_count(int val
)
4829 #ifdef CONFIG_DEBUG_PREEMPT
4833 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4836 * Is the spinlock portion underflowing?
4838 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4839 !(preempt_count() & PREEMPT_MASK
)))
4843 if (preempt_count() == val
)
4844 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4845 preempt_count() -= val
;
4847 EXPORT_SYMBOL(sub_preempt_count
);
4852 * Print scheduling while atomic bug:
4854 static noinline
void __schedule_bug(struct task_struct
*prev
)
4856 struct pt_regs
*regs
= get_irq_regs();
4858 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4859 prev
->comm
, prev
->pid
, preempt_count());
4861 debug_show_held_locks(prev
);
4863 if (irqs_disabled())
4864 print_irqtrace_events(prev
);
4873 * Various schedule()-time debugging checks and statistics:
4875 static inline void schedule_debug(struct task_struct
*prev
)
4878 * Test if we are atomic. Since do_exit() needs to call into
4879 * schedule() atomically, we ignore that path for now.
4880 * Otherwise, whine if we are scheduling when we should not be.
4882 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4883 __schedule_bug(prev
);
4885 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4887 schedstat_inc(this_rq(), sched_count
);
4888 #ifdef CONFIG_SCHEDSTATS
4889 if (unlikely(prev
->lock_depth
>= 0)) {
4890 schedstat_inc(this_rq(), bkl_count
);
4891 schedstat_inc(prev
, sched_info
.bkl_count
);
4896 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
4898 if (prev
->state
== TASK_RUNNING
) {
4899 u64 runtime
= prev
->se
.sum_exec_runtime
;
4901 runtime
-= prev
->se
.prev_sum_exec_runtime
;
4902 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
4905 * In order to avoid avg_overlap growing stale when we are
4906 * indeed overlapping and hence not getting put to sleep, grow
4907 * the avg_overlap on preemption.
4909 * We use the average preemption runtime because that
4910 * correlates to the amount of cache footprint a task can
4913 update_avg(&prev
->se
.avg_overlap
, runtime
);
4915 prev
->sched_class
->put_prev_task(rq
, prev
);
4919 * Pick up the highest-prio task:
4921 static inline struct task_struct
*
4922 pick_next_task(struct rq
*rq
)
4924 const struct sched_class
*class;
4925 struct task_struct
*p
;
4928 * Optimization: we know that if all tasks are in
4929 * the fair class we can call that function directly:
4931 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4932 p
= fair_sched_class
.pick_next_task(rq
);
4937 class = sched_class_highest
;
4939 p
= class->pick_next_task(rq
);
4943 * Will never be NULL as the idle class always
4944 * returns a non-NULL p:
4946 class = class->next
;
4951 * schedule() is the main scheduler function.
4953 asmlinkage
void __sched
__schedule(void)
4955 struct task_struct
*prev
, *next
;
4956 unsigned long *switch_count
;
4960 cpu
= smp_processor_id();
4964 switch_count
= &prev
->nivcsw
;
4966 release_kernel_lock(prev
);
4967 need_resched_nonpreemptible
:
4969 schedule_debug(prev
);
4971 if (sched_feat(HRTICK
))
4974 spin_lock_irq(&rq
->lock
);
4975 update_rq_clock(rq
);
4976 clear_tsk_need_resched(prev
);
4978 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4979 if (unlikely(signal_pending_state(prev
->state
, prev
)))
4980 prev
->state
= TASK_RUNNING
;
4982 deactivate_task(rq
, prev
, 1);
4983 switch_count
= &prev
->nvcsw
;
4987 if (prev
->sched_class
->pre_schedule
)
4988 prev
->sched_class
->pre_schedule(rq
, prev
);
4991 if (unlikely(!rq
->nr_running
))
4992 idle_balance(cpu
, rq
);
4994 put_prev_task(rq
, prev
);
4995 next
= pick_next_task(rq
);
4997 if (likely(prev
!= next
)) {
4998 sched_info_switch(prev
, next
);
5004 context_switch(rq
, prev
, next
); /* unlocks the rq */
5006 * the context switch might have flipped the stack from under
5007 * us, hence refresh the local variables.
5009 cpu
= smp_processor_id();
5012 spin_unlock_irq(&rq
->lock
);
5014 if (unlikely(reacquire_kernel_lock(current
) < 0))
5015 goto need_resched_nonpreemptible
;
5018 asmlinkage
void __sched
schedule(void)
5023 preempt_enable_no_resched();
5024 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
5027 EXPORT_SYMBOL(schedule
);
5031 * Look out! "owner" is an entirely speculative pointer
5032 * access and not reliable.
5034 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
5039 if (!sched_feat(OWNER_SPIN
))
5042 #ifdef CONFIG_DEBUG_PAGEALLOC
5044 * Need to access the cpu field knowing that
5045 * DEBUG_PAGEALLOC could have unmapped it if
5046 * the mutex owner just released it and exited.
5048 if (probe_kernel_address(&owner
->cpu
, cpu
))
5055 * Even if the access succeeded (likely case),
5056 * the cpu field may no longer be valid.
5058 if (cpu
>= nr_cpumask_bits
)
5062 * We need to validate that we can do a
5063 * get_cpu() and that we have the percpu area.
5065 if (!cpu_online(cpu
))
5072 * Owner changed, break to re-assess state.
5074 if (lock
->owner
!= owner
)
5078 * Is that owner really running on that cpu?
5080 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
5090 #ifdef CONFIG_PREEMPT
5092 * this is the entry point to schedule() from in-kernel preemption
5093 * off of preempt_enable. Kernel preemptions off return from interrupt
5094 * occur there and call schedule directly.
5096 asmlinkage
void __sched
preempt_schedule(void)
5098 struct thread_info
*ti
= current_thread_info();
5101 * If there is a non-zero preempt_count or interrupts are disabled,
5102 * we do not want to preempt the current task. Just return..
5104 if (likely(ti
->preempt_count
|| irqs_disabled()))
5108 add_preempt_count(PREEMPT_ACTIVE
);
5110 sub_preempt_count(PREEMPT_ACTIVE
);
5113 * Check again in case we missed a preemption opportunity
5114 * between schedule and now.
5117 } while (need_resched());
5119 EXPORT_SYMBOL(preempt_schedule
);
5122 * this is the entry point to schedule() from kernel preemption
5123 * off of irq context.
5124 * Note, that this is called and return with irqs disabled. This will
5125 * protect us against recursive calling from irq.
5127 asmlinkage
void __sched
preempt_schedule_irq(void)
5129 struct thread_info
*ti
= current_thread_info();
5131 /* Catch callers which need to be fixed */
5132 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
5135 add_preempt_count(PREEMPT_ACTIVE
);
5138 local_irq_disable();
5139 sub_preempt_count(PREEMPT_ACTIVE
);
5142 * Check again in case we missed a preemption opportunity
5143 * between schedule and now.
5146 } while (need_resched());
5149 #endif /* CONFIG_PREEMPT */
5151 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
5154 return try_to_wake_up(curr
->private, mode
, sync
);
5156 EXPORT_SYMBOL(default_wake_function
);
5159 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5160 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5161 * number) then we wake all the non-exclusive tasks and one exclusive task.
5163 * There are circumstances in which we can try to wake a task which has already
5164 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5165 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5167 void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
5168 int nr_exclusive
, int sync
, void *key
)
5170 wait_queue_t
*curr
, *next
;
5172 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
5173 unsigned flags
= curr
->flags
;
5175 if (curr
->func(curr
, mode
, sync
, key
) &&
5176 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
5182 * __wake_up - wake up threads blocked on a waitqueue.
5184 * @mode: which threads
5185 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5186 * @key: is directly passed to the wakeup function
5188 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
5189 int nr_exclusive
, void *key
)
5191 unsigned long flags
;
5193 spin_lock_irqsave(&q
->lock
, flags
);
5194 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
5195 spin_unlock_irqrestore(&q
->lock
, flags
);
5197 EXPORT_SYMBOL(__wake_up
);
5200 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5202 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
5204 __wake_up_common(q
, mode
, 1, 0, NULL
);
5208 * __wake_up_sync - wake up threads blocked on a waitqueue.
5210 * @mode: which threads
5211 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5213 * The sync wakeup differs that the waker knows that it will schedule
5214 * away soon, so while the target thread will be woken up, it will not
5215 * be migrated to another CPU - ie. the two threads are 'synchronized'
5216 * with each other. This can prevent needless bouncing between CPUs.
5218 * On UP it can prevent extra preemption.
5221 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
5223 unsigned long flags
;
5229 if (unlikely(!nr_exclusive
))
5232 spin_lock_irqsave(&q
->lock
, flags
);
5233 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
5234 spin_unlock_irqrestore(&q
->lock
, flags
);
5236 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
5239 * complete: - signals a single thread waiting on this completion
5240 * @x: holds the state of this particular completion
5242 * This will wake up a single thread waiting on this completion. Threads will be
5243 * awakened in the same order in which they were queued.
5245 * See also complete_all(), wait_for_completion() and related routines.
5247 void complete(struct completion
*x
)
5249 unsigned long flags
;
5251 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5253 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
5254 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5256 EXPORT_SYMBOL(complete
);
5259 * complete_all: - signals all threads waiting on this completion
5260 * @x: holds the state of this particular completion
5262 * This will wake up all threads waiting on this particular completion event.
5264 void complete_all(struct completion
*x
)
5266 unsigned long flags
;
5268 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5269 x
->done
+= UINT_MAX
/2;
5270 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
5271 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5273 EXPORT_SYMBOL(complete_all
);
5275 static inline long __sched
5276 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
5279 DECLARE_WAITQUEUE(wait
, current
);
5281 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
5282 __add_wait_queue_tail(&x
->wait
, &wait
);
5284 if (signal_pending_state(state
, current
)) {
5285 timeout
= -ERESTARTSYS
;
5288 __set_current_state(state
);
5289 spin_unlock_irq(&x
->wait
.lock
);
5290 timeout
= schedule_timeout(timeout
);
5291 spin_lock_irq(&x
->wait
.lock
);
5292 } while (!x
->done
&& timeout
);
5293 __remove_wait_queue(&x
->wait
, &wait
);
5298 return timeout
?: 1;
5302 wait_for_common(struct completion
*x
, long timeout
, int state
)
5306 spin_lock_irq(&x
->wait
.lock
);
5307 timeout
= do_wait_for_common(x
, timeout
, state
);
5308 spin_unlock_irq(&x
->wait
.lock
);
5313 * wait_for_completion: - waits for completion of a task
5314 * @x: holds the state of this particular completion
5316 * This waits to be signaled for completion of a specific task. It is NOT
5317 * interruptible and there is no timeout.
5319 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5320 * and interrupt capability. Also see complete().
5322 void __sched
wait_for_completion(struct completion
*x
)
5324 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
5326 EXPORT_SYMBOL(wait_for_completion
);
5329 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5330 * @x: holds the state of this particular completion
5331 * @timeout: timeout value in jiffies
5333 * This waits for either a completion of a specific task to be signaled or for a
5334 * specified timeout to expire. The timeout is in jiffies. It is not
5337 unsigned long __sched
5338 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
5340 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
5342 EXPORT_SYMBOL(wait_for_completion_timeout
);
5345 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5346 * @x: holds the state of this particular completion
5348 * This waits for completion of a specific task to be signaled. It is
5351 int __sched
wait_for_completion_interruptible(struct completion
*x
)
5353 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
5354 if (t
== -ERESTARTSYS
)
5358 EXPORT_SYMBOL(wait_for_completion_interruptible
);
5361 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5362 * @x: holds the state of this particular completion
5363 * @timeout: timeout value in jiffies
5365 * This waits for either a completion of a specific task to be signaled or for a
5366 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5368 unsigned long __sched
5369 wait_for_completion_interruptible_timeout(struct completion
*x
,
5370 unsigned long timeout
)
5372 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
5374 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
5377 * wait_for_completion_killable: - waits for completion of a task (killable)
5378 * @x: holds the state of this particular completion
5380 * This waits to be signaled for completion of a specific task. It can be
5381 * interrupted by a kill signal.
5383 int __sched
wait_for_completion_killable(struct completion
*x
)
5385 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
5386 if (t
== -ERESTARTSYS
)
5390 EXPORT_SYMBOL(wait_for_completion_killable
);
5393 * try_wait_for_completion - try to decrement a completion without blocking
5394 * @x: completion structure
5396 * Returns: 0 if a decrement cannot be done without blocking
5397 * 1 if a decrement succeeded.
5399 * If a completion is being used as a counting completion,
5400 * attempt to decrement the counter without blocking. This
5401 * enables us to avoid waiting if the resource the completion
5402 * is protecting is not available.
5404 bool try_wait_for_completion(struct completion
*x
)
5408 spin_lock_irq(&x
->wait
.lock
);
5413 spin_unlock_irq(&x
->wait
.lock
);
5416 EXPORT_SYMBOL(try_wait_for_completion
);
5419 * completion_done - Test to see if a completion has any waiters
5420 * @x: completion structure
5422 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5423 * 1 if there are no waiters.
5426 bool completion_done(struct completion
*x
)
5430 spin_lock_irq(&x
->wait
.lock
);
5433 spin_unlock_irq(&x
->wait
.lock
);
5436 EXPORT_SYMBOL(completion_done
);
5439 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
5441 unsigned long flags
;
5444 init_waitqueue_entry(&wait
, current
);
5446 __set_current_state(state
);
5448 spin_lock_irqsave(&q
->lock
, flags
);
5449 __add_wait_queue(q
, &wait
);
5450 spin_unlock(&q
->lock
);
5451 timeout
= schedule_timeout(timeout
);
5452 spin_lock_irq(&q
->lock
);
5453 __remove_wait_queue(q
, &wait
);
5454 spin_unlock_irqrestore(&q
->lock
, flags
);
5459 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
5461 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5463 EXPORT_SYMBOL(interruptible_sleep_on
);
5466 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5468 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
5470 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
5472 void __sched
sleep_on(wait_queue_head_t
*q
)
5474 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5476 EXPORT_SYMBOL(sleep_on
);
5478 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5480 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
5482 EXPORT_SYMBOL(sleep_on_timeout
);
5484 #ifdef CONFIG_RT_MUTEXES
5487 * rt_mutex_setprio - set the current priority of a task
5489 * @prio: prio value (kernel-internal form)
5491 * This function changes the 'effective' priority of a task. It does
5492 * not touch ->normal_prio like __setscheduler().
5494 * Used by the rt_mutex code to implement priority inheritance logic.
5496 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
5498 unsigned long flags
;
5499 int oldprio
, on_rq
, running
;
5501 const struct sched_class
*prev_class
= p
->sched_class
;
5503 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
5505 rq
= task_rq_lock(p
, &flags
);
5506 update_rq_clock(rq
);
5509 on_rq
= p
->se
.on_rq
;
5510 running
= task_current(rq
, p
);
5512 dequeue_task(rq
, p
, 0);
5514 p
->sched_class
->put_prev_task(rq
, p
);
5517 p
->sched_class
= &rt_sched_class
;
5519 p
->sched_class
= &fair_sched_class
;
5524 p
->sched_class
->set_curr_task(rq
);
5526 enqueue_task(rq
, p
, 0);
5528 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5530 task_rq_unlock(rq
, &flags
);
5535 void set_user_nice(struct task_struct
*p
, long nice
)
5537 int old_prio
, delta
, on_rq
;
5538 unsigned long flags
;
5541 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
5544 * We have to be careful, if called from sys_setpriority(),
5545 * the task might be in the middle of scheduling on another CPU.
5547 rq
= task_rq_lock(p
, &flags
);
5548 update_rq_clock(rq
);
5550 * The RT priorities are set via sched_setscheduler(), but we still
5551 * allow the 'normal' nice value to be set - but as expected
5552 * it wont have any effect on scheduling until the task is
5553 * SCHED_FIFO/SCHED_RR:
5555 if (task_has_rt_policy(p
)) {
5556 p
->static_prio
= NICE_TO_PRIO(nice
);
5559 on_rq
= p
->se
.on_rq
;
5561 dequeue_task(rq
, p
, 0);
5563 p
->static_prio
= NICE_TO_PRIO(nice
);
5566 p
->prio
= effective_prio(p
);
5567 delta
= p
->prio
- old_prio
;
5570 enqueue_task(rq
, p
, 0);
5572 * If the task increased its priority or is running and
5573 * lowered its priority, then reschedule its CPU:
5575 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
5576 resched_task(rq
->curr
);
5579 task_rq_unlock(rq
, &flags
);
5581 EXPORT_SYMBOL(set_user_nice
);
5584 * can_nice - check if a task can reduce its nice value
5588 int can_nice(const struct task_struct
*p
, const int nice
)
5590 /* convert nice value [19,-20] to rlimit style value [1,40] */
5591 int nice_rlim
= 20 - nice
;
5593 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
5594 capable(CAP_SYS_NICE
));
5597 #ifdef __ARCH_WANT_SYS_NICE
5600 * sys_nice - change the priority of the current process.
5601 * @increment: priority increment
5603 * sys_setpriority is a more generic, but much slower function that
5604 * does similar things.
5606 SYSCALL_DEFINE1(nice
, int, increment
)
5611 * Setpriority might change our priority at the same moment.
5612 * We don't have to worry. Conceptually one call occurs first
5613 * and we have a single winner.
5615 if (increment
< -40)
5620 nice
= TASK_NICE(current
) + increment
;
5626 if (increment
< 0 && !can_nice(current
, nice
))
5629 retval
= security_task_setnice(current
, nice
);
5633 set_user_nice(current
, nice
);
5640 * task_prio - return the priority value of a given task.
5641 * @p: the task in question.
5643 * This is the priority value as seen by users in /proc.
5644 * RT tasks are offset by -200. Normal tasks are centered
5645 * around 0, value goes from -16 to +15.
5647 int task_prio(const struct task_struct
*p
)
5649 return p
->prio
- MAX_RT_PRIO
;
5653 * task_nice - return the nice value of a given task.
5654 * @p: the task in question.
5656 int task_nice(const struct task_struct
*p
)
5658 return TASK_NICE(p
);
5660 EXPORT_SYMBOL(task_nice
);
5663 * idle_cpu - is a given cpu idle currently?
5664 * @cpu: the processor in question.
5666 int idle_cpu(int cpu
)
5668 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
5672 * idle_task - return the idle task for a given cpu.
5673 * @cpu: the processor in question.
5675 struct task_struct
*idle_task(int cpu
)
5677 return cpu_rq(cpu
)->idle
;
5681 * find_process_by_pid - find a process with a matching PID value.
5682 * @pid: the pid in question.
5684 static struct task_struct
*find_process_by_pid(pid_t pid
)
5686 return pid
? find_task_by_vpid(pid
) : current
;
5689 /* Actually do priority change: must hold rq lock. */
5691 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
5693 BUG_ON(p
->se
.on_rq
);
5696 switch (p
->policy
) {
5700 p
->sched_class
= &fair_sched_class
;
5704 p
->sched_class
= &rt_sched_class
;
5708 p
->rt_priority
= prio
;
5709 p
->normal_prio
= normal_prio(p
);
5710 /* we are holding p->pi_lock already */
5711 p
->prio
= rt_mutex_getprio(p
);
5716 * check the target process has a UID that matches the current process's
5718 static bool check_same_owner(struct task_struct
*p
)
5720 const struct cred
*cred
= current_cred(), *pcred
;
5724 pcred
= __task_cred(p
);
5725 match
= (cred
->euid
== pcred
->euid
||
5726 cred
->euid
== pcred
->uid
);
5731 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
5732 struct sched_param
*param
, bool user
)
5734 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
5735 unsigned long flags
;
5736 const struct sched_class
*prev_class
= p
->sched_class
;
5739 /* may grab non-irq protected spin_locks */
5740 BUG_ON(in_interrupt());
5742 /* double check policy once rq lock held */
5744 policy
= oldpolicy
= p
->policy
;
5745 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
5746 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
5747 policy
!= SCHED_IDLE
)
5750 * Valid priorities for SCHED_FIFO and SCHED_RR are
5751 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5752 * SCHED_BATCH and SCHED_IDLE is 0.
5754 if (param
->sched_priority
< 0 ||
5755 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5756 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
5758 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
5762 * Allow unprivileged RT tasks to decrease priority:
5764 if (user
&& !capable(CAP_SYS_NICE
)) {
5765 if (rt_policy(policy
)) {
5766 unsigned long rlim_rtprio
;
5768 if (!lock_task_sighand(p
, &flags
))
5770 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
5771 unlock_task_sighand(p
, &flags
);
5773 /* can't set/change the rt policy */
5774 if (policy
!= p
->policy
&& !rlim_rtprio
)
5777 /* can't increase priority */
5778 if (param
->sched_priority
> p
->rt_priority
&&
5779 param
->sched_priority
> rlim_rtprio
)
5783 * Like positive nice levels, dont allow tasks to
5784 * move out of SCHED_IDLE either:
5786 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
5789 /* can't change other user's priorities */
5790 if (!check_same_owner(p
))
5795 #ifdef CONFIG_RT_GROUP_SCHED
5797 * Do not allow realtime tasks into groups that have no runtime
5800 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
5801 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
5805 retval
= security_task_setscheduler(p
, policy
, param
);
5811 * make sure no PI-waiters arrive (or leave) while we are
5812 * changing the priority of the task:
5814 spin_lock_irqsave(&p
->pi_lock
, flags
);
5816 * To be able to change p->policy safely, the apropriate
5817 * runqueue lock must be held.
5819 rq
= __task_rq_lock(p
);
5820 /* recheck policy now with rq lock held */
5821 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5822 policy
= oldpolicy
= -1;
5823 __task_rq_unlock(rq
);
5824 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5827 update_rq_clock(rq
);
5828 on_rq
= p
->se
.on_rq
;
5829 running
= task_current(rq
, p
);
5831 deactivate_task(rq
, p
, 0);
5833 p
->sched_class
->put_prev_task(rq
, p
);
5836 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5839 p
->sched_class
->set_curr_task(rq
);
5841 activate_task(rq
, p
, 0);
5843 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5845 __task_rq_unlock(rq
);
5846 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5848 rt_mutex_adjust_pi(p
);
5854 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5855 * @p: the task in question.
5856 * @policy: new policy.
5857 * @param: structure containing the new RT priority.
5859 * NOTE that the task may be already dead.
5861 int sched_setscheduler(struct task_struct
*p
, int policy
,
5862 struct sched_param
*param
)
5864 return __sched_setscheduler(p
, policy
, param
, true);
5866 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5869 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5870 * @p: the task in question.
5871 * @policy: new policy.
5872 * @param: structure containing the new RT priority.
5874 * Just like sched_setscheduler, only don't bother checking if the
5875 * current context has permission. For example, this is needed in
5876 * stop_machine(): we create temporary high priority worker threads,
5877 * but our caller might not have that capability.
5879 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5880 struct sched_param
*param
)
5882 return __sched_setscheduler(p
, policy
, param
, false);
5886 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5888 struct sched_param lparam
;
5889 struct task_struct
*p
;
5892 if (!param
|| pid
< 0)
5894 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5899 p
= find_process_by_pid(pid
);
5901 retval
= sched_setscheduler(p
, policy
, &lparam
);
5908 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5909 * @pid: the pid in question.
5910 * @policy: new policy.
5911 * @param: structure containing the new RT priority.
5913 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
5914 struct sched_param __user
*, param
)
5916 /* negative values for policy are not valid */
5920 return do_sched_setscheduler(pid
, policy
, param
);
5924 * sys_sched_setparam - set/change the RT priority of a thread
5925 * @pid: the pid in question.
5926 * @param: structure containing the new RT priority.
5928 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5930 return do_sched_setscheduler(pid
, -1, param
);
5934 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5935 * @pid: the pid in question.
5937 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
5939 struct task_struct
*p
;
5946 read_lock(&tasklist_lock
);
5947 p
= find_process_by_pid(pid
);
5949 retval
= security_task_getscheduler(p
);
5953 read_unlock(&tasklist_lock
);
5958 * sys_sched_getscheduler - get the RT priority of a thread
5959 * @pid: the pid in question.
5960 * @param: structure containing the RT priority.
5962 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5964 struct sched_param lp
;
5965 struct task_struct
*p
;
5968 if (!param
|| pid
< 0)
5971 read_lock(&tasklist_lock
);
5972 p
= find_process_by_pid(pid
);
5977 retval
= security_task_getscheduler(p
);
5981 lp
.sched_priority
= p
->rt_priority
;
5982 read_unlock(&tasklist_lock
);
5985 * This one might sleep, we cannot do it with a spinlock held ...
5987 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5992 read_unlock(&tasklist_lock
);
5996 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
5998 cpumask_var_t cpus_allowed
, new_mask
;
5999 struct task_struct
*p
;
6003 read_lock(&tasklist_lock
);
6005 p
= find_process_by_pid(pid
);
6007 read_unlock(&tasklist_lock
);
6013 * It is not safe to call set_cpus_allowed with the
6014 * tasklist_lock held. We will bump the task_struct's
6015 * usage count and then drop tasklist_lock.
6018 read_unlock(&tasklist_lock
);
6020 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
6024 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
6026 goto out_free_cpus_allowed
;
6029 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
6032 retval
= security_task_setscheduler(p
, 0, NULL
);
6036 cpuset_cpus_allowed(p
, cpus_allowed
);
6037 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
6039 retval
= set_cpus_allowed_ptr(p
, new_mask
);
6042 cpuset_cpus_allowed(p
, cpus_allowed
);
6043 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
6045 * We must have raced with a concurrent cpuset
6046 * update. Just reset the cpus_allowed to the
6047 * cpuset's cpus_allowed
6049 cpumask_copy(new_mask
, cpus_allowed
);
6054 free_cpumask_var(new_mask
);
6055 out_free_cpus_allowed
:
6056 free_cpumask_var(cpus_allowed
);
6063 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
6064 struct cpumask
*new_mask
)
6066 if (len
< cpumask_size())
6067 cpumask_clear(new_mask
);
6068 else if (len
> cpumask_size())
6069 len
= cpumask_size();
6071 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
6075 * sys_sched_setaffinity - set the cpu affinity of a process
6076 * @pid: pid of the process
6077 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6078 * @user_mask_ptr: user-space pointer to the new cpu mask
6080 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
6081 unsigned long __user
*, user_mask_ptr
)
6083 cpumask_var_t new_mask
;
6086 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
6089 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
6091 retval
= sched_setaffinity(pid
, new_mask
);
6092 free_cpumask_var(new_mask
);
6096 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
6098 struct task_struct
*p
;
6102 read_lock(&tasklist_lock
);
6105 p
= find_process_by_pid(pid
);
6109 retval
= security_task_getscheduler(p
);
6113 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
6116 read_unlock(&tasklist_lock
);
6123 * sys_sched_getaffinity - get the cpu affinity of a process
6124 * @pid: pid of the process
6125 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6126 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6128 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
6129 unsigned long __user
*, user_mask_ptr
)
6134 if (len
< cpumask_size())
6137 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
6140 ret
= sched_getaffinity(pid
, mask
);
6142 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
6145 ret
= cpumask_size();
6147 free_cpumask_var(mask
);
6153 * sys_sched_yield - yield the current processor to other threads.
6155 * This function yields the current CPU to other tasks. If there are no
6156 * other threads running on this CPU then this function will return.
6158 SYSCALL_DEFINE0(sched_yield
)
6160 struct rq
*rq
= this_rq_lock();
6162 schedstat_inc(rq
, yld_count
);
6163 current
->sched_class
->yield_task(rq
);
6166 * Since we are going to call schedule() anyway, there's
6167 * no need to preempt or enable interrupts:
6169 __release(rq
->lock
);
6170 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
6171 _raw_spin_unlock(&rq
->lock
);
6172 preempt_enable_no_resched();
6179 static void __cond_resched(void)
6181 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6182 __might_sleep(__FILE__
, __LINE__
);
6185 * The BKS might be reacquired before we have dropped
6186 * PREEMPT_ACTIVE, which could trigger a second
6187 * cond_resched() call.
6190 add_preempt_count(PREEMPT_ACTIVE
);
6192 sub_preempt_count(PREEMPT_ACTIVE
);
6193 } while (need_resched());
6196 int __sched
_cond_resched(void)
6198 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
6199 system_state
== SYSTEM_RUNNING
) {
6205 EXPORT_SYMBOL(_cond_resched
);
6208 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
6209 * call schedule, and on return reacquire the lock.
6211 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6212 * operations here to prevent schedule() from being called twice (once via
6213 * spin_unlock(), once by hand).
6215 int cond_resched_lock(spinlock_t
*lock
)
6217 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
6220 if (spin_needbreak(lock
) || resched
) {
6222 if (resched
&& need_resched())
6231 EXPORT_SYMBOL(cond_resched_lock
);
6233 int __sched
cond_resched_softirq(void)
6235 BUG_ON(!in_softirq());
6237 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
6245 EXPORT_SYMBOL(cond_resched_softirq
);
6248 * yield - yield the current processor to other threads.
6250 * This is a shortcut for kernel-space yielding - it marks the
6251 * thread runnable and calls sys_sched_yield().
6253 void __sched
yield(void)
6255 set_current_state(TASK_RUNNING
);
6258 EXPORT_SYMBOL(yield
);
6261 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6262 * that process accounting knows that this is a task in IO wait state.
6264 * But don't do that if it is a deliberate, throttling IO wait (this task
6265 * has set its backing_dev_info: the queue against which it should throttle)
6267 void __sched
io_schedule(void)
6269 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
6271 delayacct_blkio_start();
6272 atomic_inc(&rq
->nr_iowait
);
6274 atomic_dec(&rq
->nr_iowait
);
6275 delayacct_blkio_end();
6277 EXPORT_SYMBOL(io_schedule
);
6279 long __sched
io_schedule_timeout(long timeout
)
6281 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
6284 delayacct_blkio_start();
6285 atomic_inc(&rq
->nr_iowait
);
6286 ret
= schedule_timeout(timeout
);
6287 atomic_dec(&rq
->nr_iowait
);
6288 delayacct_blkio_end();
6293 * sys_sched_get_priority_max - return maximum RT priority.
6294 * @policy: scheduling class.
6296 * this syscall returns the maximum rt_priority that can be used
6297 * by a given scheduling class.
6299 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
6306 ret
= MAX_USER_RT_PRIO
-1;
6318 * sys_sched_get_priority_min - return minimum RT priority.
6319 * @policy: scheduling class.
6321 * this syscall returns the minimum rt_priority that can be used
6322 * by a given scheduling class.
6324 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
6342 * sys_sched_rr_get_interval - return the default timeslice of a process.
6343 * @pid: pid of the process.
6344 * @interval: userspace pointer to the timeslice value.
6346 * this syscall writes the default timeslice value of a given process
6347 * into the user-space timespec buffer. A value of '0' means infinity.
6349 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
6350 struct timespec __user
*, interval
)
6352 struct task_struct
*p
;
6353 unsigned int time_slice
;
6361 read_lock(&tasklist_lock
);
6362 p
= find_process_by_pid(pid
);
6366 retval
= security_task_getscheduler(p
);
6371 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6372 * tasks that are on an otherwise idle runqueue:
6375 if (p
->policy
== SCHED_RR
) {
6376 time_slice
= DEF_TIMESLICE
;
6377 } else if (p
->policy
!= SCHED_FIFO
) {
6378 struct sched_entity
*se
= &p
->se
;
6379 unsigned long flags
;
6382 rq
= task_rq_lock(p
, &flags
);
6383 if (rq
->cfs
.load
.weight
)
6384 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
6385 task_rq_unlock(rq
, &flags
);
6387 read_unlock(&tasklist_lock
);
6388 jiffies_to_timespec(time_slice
, &t
);
6389 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
6393 read_unlock(&tasklist_lock
);
6397 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
6399 void sched_show_task(struct task_struct
*p
)
6401 unsigned long free
= 0;
6404 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
6405 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
6406 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
6407 #if BITS_PER_LONG == 32
6408 if (state
== TASK_RUNNING
)
6409 printk(KERN_CONT
" running ");
6411 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
6413 if (state
== TASK_RUNNING
)
6414 printk(KERN_CONT
" running task ");
6416 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
6418 #ifdef CONFIG_DEBUG_STACK_USAGE
6419 free
= stack_not_used(p
);
6421 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
6422 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
6424 show_stack(p
, NULL
);
6427 void show_state_filter(unsigned long state_filter
)
6429 struct task_struct
*g
, *p
;
6431 #if BITS_PER_LONG == 32
6433 " task PC stack pid father\n");
6436 " task PC stack pid father\n");
6438 read_lock(&tasklist_lock
);
6439 do_each_thread(g
, p
) {
6441 * reset the NMI-timeout, listing all files on a slow
6442 * console might take alot of time:
6444 touch_nmi_watchdog();
6445 if (!state_filter
|| (p
->state
& state_filter
))
6447 } while_each_thread(g
, p
);
6449 touch_all_softlockup_watchdogs();
6451 #ifdef CONFIG_SCHED_DEBUG
6452 sysrq_sched_debug_show();
6454 read_unlock(&tasklist_lock
);
6456 * Only show locks if all tasks are dumped:
6458 if (state_filter
== -1)
6459 debug_show_all_locks();
6462 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
6464 idle
->sched_class
= &idle_sched_class
;
6468 * init_idle - set up an idle thread for a given CPU
6469 * @idle: task in question
6470 * @cpu: cpu the idle task belongs to
6472 * NOTE: this function does not set the idle thread's NEED_RESCHED
6473 * flag, to make booting more robust.
6475 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
6477 struct rq
*rq
= cpu_rq(cpu
);
6478 unsigned long flags
;
6480 spin_lock_irqsave(&rq
->lock
, flags
);
6483 idle
->se
.exec_start
= sched_clock();
6485 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
6486 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
6487 __set_task_cpu(idle
, cpu
);
6489 rq
->curr
= rq
->idle
= idle
;
6490 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6493 spin_unlock_irqrestore(&rq
->lock
, flags
);
6495 /* Set the preempt count _outside_ the spinlocks! */
6496 #if defined(CONFIG_PREEMPT)
6497 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
6499 task_thread_info(idle
)->preempt_count
= 0;
6502 * The idle tasks have their own, simple scheduling class:
6504 idle
->sched_class
= &idle_sched_class
;
6505 ftrace_graph_init_task(idle
);
6509 * In a system that switches off the HZ timer nohz_cpu_mask
6510 * indicates which cpus entered this state. This is used
6511 * in the rcu update to wait only for active cpus. For system
6512 * which do not switch off the HZ timer nohz_cpu_mask should
6513 * always be CPU_BITS_NONE.
6515 cpumask_var_t nohz_cpu_mask
;
6518 * Increase the granularity value when there are more CPUs,
6519 * because with more CPUs the 'effective latency' as visible
6520 * to users decreases. But the relationship is not linear,
6521 * so pick a second-best guess by going with the log2 of the
6524 * This idea comes from the SD scheduler of Con Kolivas:
6526 static inline void sched_init_granularity(void)
6528 unsigned int factor
= 1 + ilog2(num_online_cpus());
6529 const unsigned long limit
= 200000000;
6531 sysctl_sched_min_granularity
*= factor
;
6532 if (sysctl_sched_min_granularity
> limit
)
6533 sysctl_sched_min_granularity
= limit
;
6535 sysctl_sched_latency
*= factor
;
6536 if (sysctl_sched_latency
> limit
)
6537 sysctl_sched_latency
= limit
;
6539 sysctl_sched_wakeup_granularity
*= factor
;
6541 sysctl_sched_shares_ratelimit
*= factor
;
6546 * This is how migration works:
6548 * 1) we queue a struct migration_req structure in the source CPU's
6549 * runqueue and wake up that CPU's migration thread.
6550 * 2) we down() the locked semaphore => thread blocks.
6551 * 3) migration thread wakes up (implicitly it forces the migrated
6552 * thread off the CPU)
6553 * 4) it gets the migration request and checks whether the migrated
6554 * task is still in the wrong runqueue.
6555 * 5) if it's in the wrong runqueue then the migration thread removes
6556 * it and puts it into the right queue.
6557 * 6) migration thread up()s the semaphore.
6558 * 7) we wake up and the migration is done.
6562 * Change a given task's CPU affinity. Migrate the thread to a
6563 * proper CPU and schedule it away if the CPU it's executing on
6564 * is removed from the allowed bitmask.
6566 * NOTE: the caller must have a valid reference to the task, the
6567 * task must not exit() & deallocate itself prematurely. The
6568 * call is not atomic; no spinlocks may be held.
6570 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
6572 struct migration_req req
;
6573 unsigned long flags
;
6577 rq
= task_rq_lock(p
, &flags
);
6578 if (!cpumask_intersects(new_mask
, cpu_online_mask
)) {
6583 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
6584 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
6589 if (p
->sched_class
->set_cpus_allowed
)
6590 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
6592 cpumask_copy(&p
->cpus_allowed
, new_mask
);
6593 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
6596 /* Can the task run on the task's current CPU? If so, we're done */
6597 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
6600 if (migrate_task(p
, cpumask_any_and(cpu_online_mask
, new_mask
), &req
)) {
6601 /* Need help from migration thread: drop lock and wait. */
6602 task_rq_unlock(rq
, &flags
);
6603 wake_up_process(rq
->migration_thread
);
6604 wait_for_completion(&req
.done
);
6605 tlb_migrate_finish(p
->mm
);
6609 task_rq_unlock(rq
, &flags
);
6613 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
6616 * Move (not current) task off this cpu, onto dest cpu. We're doing
6617 * this because either it can't run here any more (set_cpus_allowed()
6618 * away from this CPU, or CPU going down), or because we're
6619 * attempting to rebalance this task on exec (sched_exec).
6621 * So we race with normal scheduler movements, but that's OK, as long
6622 * as the task is no longer on this CPU.
6624 * Returns non-zero if task was successfully migrated.
6626 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6628 struct rq
*rq_dest
, *rq_src
;
6631 if (unlikely(!cpu_active(dest_cpu
)))
6634 rq_src
= cpu_rq(src_cpu
);
6635 rq_dest
= cpu_rq(dest_cpu
);
6637 double_rq_lock(rq_src
, rq_dest
);
6638 /* Already moved. */
6639 if (task_cpu(p
) != src_cpu
)
6641 /* Affinity changed (again). */
6642 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
6645 on_rq
= p
->se
.on_rq
;
6647 deactivate_task(rq_src
, p
, 0);
6649 set_task_cpu(p
, dest_cpu
);
6651 activate_task(rq_dest
, p
, 0);
6652 check_preempt_curr(rq_dest
, p
, 0);
6657 double_rq_unlock(rq_src
, rq_dest
);
6662 * migration_thread - this is a highprio system thread that performs
6663 * thread migration by bumping thread off CPU then 'pushing' onto
6666 static int migration_thread(void *data
)
6668 int cpu
= (long)data
;
6672 BUG_ON(rq
->migration_thread
!= current
);
6674 set_current_state(TASK_INTERRUPTIBLE
);
6675 while (!kthread_should_stop()) {
6676 struct migration_req
*req
;
6677 struct list_head
*head
;
6679 spin_lock_irq(&rq
->lock
);
6681 if (cpu_is_offline(cpu
)) {
6682 spin_unlock_irq(&rq
->lock
);
6686 if (rq
->active_balance
) {
6687 active_load_balance(rq
, cpu
);
6688 rq
->active_balance
= 0;
6691 head
= &rq
->migration_queue
;
6693 if (list_empty(head
)) {
6694 spin_unlock_irq(&rq
->lock
);
6696 set_current_state(TASK_INTERRUPTIBLE
);
6699 req
= list_entry(head
->next
, struct migration_req
, list
);
6700 list_del_init(head
->next
);
6702 spin_unlock(&rq
->lock
);
6703 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
6706 complete(&req
->done
);
6708 __set_current_state(TASK_RUNNING
);
6712 /* Wait for kthread_stop */
6713 set_current_state(TASK_INTERRUPTIBLE
);
6714 while (!kthread_should_stop()) {
6716 set_current_state(TASK_INTERRUPTIBLE
);
6718 __set_current_state(TASK_RUNNING
);
6722 #ifdef CONFIG_HOTPLUG_CPU
6724 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6728 local_irq_disable();
6729 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
6735 * Figure out where task on dead CPU should go, use force if necessary.
6737 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
6740 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(dead_cpu
));
6743 /* Look for allowed, online CPU in same node. */
6744 for_each_cpu_and(dest_cpu
, nodemask
, cpu_online_mask
)
6745 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
6748 /* Any allowed, online CPU? */
6749 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_online_mask
);
6750 if (dest_cpu
< nr_cpu_ids
)
6753 /* No more Mr. Nice Guy. */
6754 if (dest_cpu
>= nr_cpu_ids
) {
6755 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
6756 dest_cpu
= cpumask_any_and(cpu_online_mask
, &p
->cpus_allowed
);
6759 * Don't tell them about moving exiting tasks or
6760 * kernel threads (both mm NULL), since they never
6763 if (p
->mm
&& printk_ratelimit()) {
6764 printk(KERN_INFO
"process %d (%s) no "
6765 "longer affine to cpu%d\n",
6766 task_pid_nr(p
), p
->comm
, dead_cpu
);
6771 /* It can have affinity changed while we were choosing. */
6772 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
6777 * While a dead CPU has no uninterruptible tasks queued at this point,
6778 * it might still have a nonzero ->nr_uninterruptible counter, because
6779 * for performance reasons the counter is not stricly tracking tasks to
6780 * their home CPUs. So we just add the counter to another CPU's counter,
6781 * to keep the global sum constant after CPU-down:
6783 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
6785 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_online_mask
));
6786 unsigned long flags
;
6788 local_irq_save(flags
);
6789 double_rq_lock(rq_src
, rq_dest
);
6790 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
6791 rq_src
->nr_uninterruptible
= 0;
6792 double_rq_unlock(rq_src
, rq_dest
);
6793 local_irq_restore(flags
);
6796 /* Run through task list and migrate tasks from the dead cpu. */
6797 static void migrate_live_tasks(int src_cpu
)
6799 struct task_struct
*p
, *t
;
6801 read_lock(&tasklist_lock
);
6803 do_each_thread(t
, p
) {
6807 if (task_cpu(p
) == src_cpu
)
6808 move_task_off_dead_cpu(src_cpu
, p
);
6809 } while_each_thread(t
, p
);
6811 read_unlock(&tasklist_lock
);
6815 * Schedules idle task to be the next runnable task on current CPU.
6816 * It does so by boosting its priority to highest possible.
6817 * Used by CPU offline code.
6819 void sched_idle_next(void)
6821 int this_cpu
= smp_processor_id();
6822 struct rq
*rq
= cpu_rq(this_cpu
);
6823 struct task_struct
*p
= rq
->idle
;
6824 unsigned long flags
;
6826 /* cpu has to be offline */
6827 BUG_ON(cpu_online(this_cpu
));
6830 * Strictly not necessary since rest of the CPUs are stopped by now
6831 * and interrupts disabled on the current cpu.
6833 spin_lock_irqsave(&rq
->lock
, flags
);
6835 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6837 update_rq_clock(rq
);
6838 activate_task(rq
, p
, 0);
6840 spin_unlock_irqrestore(&rq
->lock
, flags
);
6844 * Ensures that the idle task is using init_mm right before its cpu goes
6847 void idle_task_exit(void)
6849 struct mm_struct
*mm
= current
->active_mm
;
6851 BUG_ON(cpu_online(smp_processor_id()));
6854 switch_mm(mm
, &init_mm
, current
);
6858 /* called under rq->lock with disabled interrupts */
6859 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
6861 struct rq
*rq
= cpu_rq(dead_cpu
);
6863 /* Must be exiting, otherwise would be on tasklist. */
6864 BUG_ON(!p
->exit_state
);
6866 /* Cannot have done final schedule yet: would have vanished. */
6867 BUG_ON(p
->state
== TASK_DEAD
);
6872 * Drop lock around migration; if someone else moves it,
6873 * that's OK. No task can be added to this CPU, so iteration is
6876 spin_unlock_irq(&rq
->lock
);
6877 move_task_off_dead_cpu(dead_cpu
, p
);
6878 spin_lock_irq(&rq
->lock
);
6883 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6884 static void migrate_dead_tasks(unsigned int dead_cpu
)
6886 struct rq
*rq
= cpu_rq(dead_cpu
);
6887 struct task_struct
*next
;
6890 if (!rq
->nr_running
)
6892 update_rq_clock(rq
);
6893 next
= pick_next_task(rq
);
6896 next
->sched_class
->put_prev_task(rq
, next
);
6897 migrate_dead(dead_cpu
, next
);
6901 #endif /* CONFIG_HOTPLUG_CPU */
6903 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6905 static struct ctl_table sd_ctl_dir
[] = {
6907 .procname
= "sched_domain",
6913 static struct ctl_table sd_ctl_root
[] = {
6915 .ctl_name
= CTL_KERN
,
6916 .procname
= "kernel",
6918 .child
= sd_ctl_dir
,
6923 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6925 struct ctl_table
*entry
=
6926 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6931 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6933 struct ctl_table
*entry
;
6936 * In the intermediate directories, both the child directory and
6937 * procname are dynamically allocated and could fail but the mode
6938 * will always be set. In the lowest directory the names are
6939 * static strings and all have proc handlers.
6941 for (entry
= *tablep
; entry
->mode
; entry
++) {
6943 sd_free_ctl_entry(&entry
->child
);
6944 if (entry
->proc_handler
== NULL
)
6945 kfree(entry
->procname
);
6953 set_table_entry(struct ctl_table
*entry
,
6954 const char *procname
, void *data
, int maxlen
,
6955 mode_t mode
, proc_handler
*proc_handler
)
6957 entry
->procname
= procname
;
6959 entry
->maxlen
= maxlen
;
6961 entry
->proc_handler
= proc_handler
;
6964 static struct ctl_table
*
6965 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6967 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
6972 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6973 sizeof(long), 0644, proc_doulongvec_minmax
);
6974 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6975 sizeof(long), 0644, proc_doulongvec_minmax
);
6976 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6977 sizeof(int), 0644, proc_dointvec_minmax
);
6978 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6979 sizeof(int), 0644, proc_dointvec_minmax
);
6980 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6981 sizeof(int), 0644, proc_dointvec_minmax
);
6982 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6983 sizeof(int), 0644, proc_dointvec_minmax
);
6984 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6985 sizeof(int), 0644, proc_dointvec_minmax
);
6986 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6987 sizeof(int), 0644, proc_dointvec_minmax
);
6988 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6989 sizeof(int), 0644, proc_dointvec_minmax
);
6990 set_table_entry(&table
[9], "cache_nice_tries",
6991 &sd
->cache_nice_tries
,
6992 sizeof(int), 0644, proc_dointvec_minmax
);
6993 set_table_entry(&table
[10], "flags", &sd
->flags
,
6994 sizeof(int), 0644, proc_dointvec_minmax
);
6995 set_table_entry(&table
[11], "name", sd
->name
,
6996 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
6997 /* &table[12] is terminator */
7002 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
7004 struct ctl_table
*entry
, *table
;
7005 struct sched_domain
*sd
;
7006 int domain_num
= 0, i
;
7009 for_each_domain(cpu
, sd
)
7011 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
7016 for_each_domain(cpu
, sd
) {
7017 snprintf(buf
, 32, "domain%d", i
);
7018 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7020 entry
->child
= sd_alloc_ctl_domain_table(sd
);
7027 static struct ctl_table_header
*sd_sysctl_header
;
7028 static void register_sched_domain_sysctl(void)
7030 int i
, cpu_num
= num_online_cpus();
7031 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
7034 WARN_ON(sd_ctl_dir
[0].child
);
7035 sd_ctl_dir
[0].child
= entry
;
7040 for_each_online_cpu(i
) {
7041 snprintf(buf
, 32, "cpu%d", i
);
7042 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7044 entry
->child
= sd_alloc_ctl_cpu_table(i
);
7048 WARN_ON(sd_sysctl_header
);
7049 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
7052 /* may be called multiple times per register */
7053 static void unregister_sched_domain_sysctl(void)
7055 if (sd_sysctl_header
)
7056 unregister_sysctl_table(sd_sysctl_header
);
7057 sd_sysctl_header
= NULL
;
7058 if (sd_ctl_dir
[0].child
)
7059 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
7062 static void register_sched_domain_sysctl(void)
7065 static void unregister_sched_domain_sysctl(void)
7070 static void set_rq_online(struct rq
*rq
)
7073 const struct sched_class
*class;
7075 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
7078 for_each_class(class) {
7079 if (class->rq_online
)
7080 class->rq_online(rq
);
7085 static void set_rq_offline(struct rq
*rq
)
7088 const struct sched_class
*class;
7090 for_each_class(class) {
7091 if (class->rq_offline
)
7092 class->rq_offline(rq
);
7095 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
7101 * migration_call - callback that gets triggered when a CPU is added.
7102 * Here we can start up the necessary migration thread for the new CPU.
7104 static int __cpuinit
7105 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
7107 struct task_struct
*p
;
7108 int cpu
= (long)hcpu
;
7109 unsigned long flags
;
7114 case CPU_UP_PREPARE
:
7115 case CPU_UP_PREPARE_FROZEN
:
7116 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
7119 kthread_bind(p
, cpu
);
7120 /* Must be high prio: stop_machine expects to yield to it. */
7121 rq
= task_rq_lock(p
, &flags
);
7122 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7123 task_rq_unlock(rq
, &flags
);
7124 cpu_rq(cpu
)->migration_thread
= p
;
7128 case CPU_ONLINE_FROZEN
:
7129 /* Strictly unnecessary, as first user will wake it. */
7130 wake_up_process(cpu_rq(cpu
)->migration_thread
);
7132 /* Update our root-domain */
7134 spin_lock_irqsave(&rq
->lock
, flags
);
7136 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7140 spin_unlock_irqrestore(&rq
->lock
, flags
);
7143 #ifdef CONFIG_HOTPLUG_CPU
7144 case CPU_UP_CANCELED
:
7145 case CPU_UP_CANCELED_FROZEN
:
7146 if (!cpu_rq(cpu
)->migration_thread
)
7148 /* Unbind it from offline cpu so it can run. Fall thru. */
7149 kthread_bind(cpu_rq(cpu
)->migration_thread
,
7150 cpumask_any(cpu_online_mask
));
7151 kthread_stop(cpu_rq(cpu
)->migration_thread
);
7152 cpu_rq(cpu
)->migration_thread
= NULL
;
7156 case CPU_DEAD_FROZEN
:
7157 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7158 migrate_live_tasks(cpu
);
7160 kthread_stop(rq
->migration_thread
);
7161 rq
->migration_thread
= NULL
;
7162 /* Idle task back to normal (off runqueue, low prio) */
7163 spin_lock_irq(&rq
->lock
);
7164 update_rq_clock(rq
);
7165 deactivate_task(rq
, rq
->idle
, 0);
7166 rq
->idle
->static_prio
= MAX_PRIO
;
7167 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
7168 rq
->idle
->sched_class
= &idle_sched_class
;
7169 migrate_dead_tasks(cpu
);
7170 spin_unlock_irq(&rq
->lock
);
7172 migrate_nr_uninterruptible(rq
);
7173 BUG_ON(rq
->nr_running
!= 0);
7176 * No need to migrate the tasks: it was best-effort if
7177 * they didn't take sched_hotcpu_mutex. Just wake up
7180 spin_lock_irq(&rq
->lock
);
7181 while (!list_empty(&rq
->migration_queue
)) {
7182 struct migration_req
*req
;
7184 req
= list_entry(rq
->migration_queue
.next
,
7185 struct migration_req
, list
);
7186 list_del_init(&req
->list
);
7187 spin_unlock_irq(&rq
->lock
);
7188 complete(&req
->done
);
7189 spin_lock_irq(&rq
->lock
);
7191 spin_unlock_irq(&rq
->lock
);
7195 case CPU_DYING_FROZEN
:
7196 /* Update our root-domain */
7198 spin_lock_irqsave(&rq
->lock
, flags
);
7200 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7203 spin_unlock_irqrestore(&rq
->lock
, flags
);
7210 /* Register at highest priority so that task migration (migrate_all_tasks)
7211 * happens before everything else.
7213 static struct notifier_block __cpuinitdata migration_notifier
= {
7214 .notifier_call
= migration_call
,
7218 static int __init
migration_init(void)
7220 void *cpu
= (void *)(long)smp_processor_id();
7223 /* Start one for the boot CPU: */
7224 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
7225 BUG_ON(err
== NOTIFY_BAD
);
7226 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
7227 register_cpu_notifier(&migration_notifier
);
7231 early_initcall(migration_init
);
7236 #ifdef CONFIG_SCHED_DEBUG
7238 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
7239 struct cpumask
*groupmask
)
7241 struct sched_group
*group
= sd
->groups
;
7244 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
7245 cpumask_clear(groupmask
);
7247 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
7249 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
7250 printk("does not load-balance\n");
7252 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
7257 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
7259 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
7260 printk(KERN_ERR
"ERROR: domain->span does not contain "
7263 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
7264 printk(KERN_ERR
"ERROR: domain->groups does not contain"
7268 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
7272 printk(KERN_ERR
"ERROR: group is NULL\n");
7276 if (!group
->__cpu_power
) {
7277 printk(KERN_CONT
"\n");
7278 printk(KERN_ERR
"ERROR: domain->cpu_power not "
7283 if (!cpumask_weight(sched_group_cpus(group
))) {
7284 printk(KERN_CONT
"\n");
7285 printk(KERN_ERR
"ERROR: empty group\n");
7289 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
7290 printk(KERN_CONT
"\n");
7291 printk(KERN_ERR
"ERROR: repeated CPUs\n");
7295 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
7297 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
7298 printk(KERN_CONT
" %s", str
);
7300 group
= group
->next
;
7301 } while (group
!= sd
->groups
);
7302 printk(KERN_CONT
"\n");
7304 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
7305 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
7308 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
7309 printk(KERN_ERR
"ERROR: parent span is not a superset "
7310 "of domain->span\n");
7314 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
7316 cpumask_var_t groupmask
;
7320 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
7324 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
7326 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
7327 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
7332 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
7339 free_cpumask_var(groupmask
);
7341 #else /* !CONFIG_SCHED_DEBUG */
7342 # define sched_domain_debug(sd, cpu) do { } while (0)
7343 #endif /* CONFIG_SCHED_DEBUG */
7345 static int sd_degenerate(struct sched_domain
*sd
)
7347 if (cpumask_weight(sched_domain_span(sd
)) == 1)
7350 /* Following flags need at least 2 groups */
7351 if (sd
->flags
& (SD_LOAD_BALANCE
|
7352 SD_BALANCE_NEWIDLE
|
7356 SD_SHARE_PKG_RESOURCES
)) {
7357 if (sd
->groups
!= sd
->groups
->next
)
7361 /* Following flags don't use groups */
7362 if (sd
->flags
& (SD_WAKE_IDLE
|
7371 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
7373 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
7375 if (sd_degenerate(parent
))
7378 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
7381 /* Does parent contain flags not in child? */
7382 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7383 if (cflags
& SD_WAKE_AFFINE
)
7384 pflags
&= ~SD_WAKE_BALANCE
;
7385 /* Flags needing groups don't count if only 1 group in parent */
7386 if (parent
->groups
== parent
->groups
->next
) {
7387 pflags
&= ~(SD_LOAD_BALANCE
|
7388 SD_BALANCE_NEWIDLE
|
7392 SD_SHARE_PKG_RESOURCES
);
7393 if (nr_node_ids
== 1)
7394 pflags
&= ~SD_SERIALIZE
;
7396 if (~cflags
& pflags
)
7402 static void free_rootdomain(struct root_domain
*rd
)
7404 cpupri_cleanup(&rd
->cpupri
);
7406 free_cpumask_var(rd
->rto_mask
);
7407 free_cpumask_var(rd
->online
);
7408 free_cpumask_var(rd
->span
);
7412 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
7414 struct root_domain
*old_rd
= NULL
;
7415 unsigned long flags
;
7417 spin_lock_irqsave(&rq
->lock
, flags
);
7422 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
7425 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
7428 * If we dont want to free the old_rt yet then
7429 * set old_rd to NULL to skip the freeing later
7432 if (!atomic_dec_and_test(&old_rd
->refcount
))
7436 atomic_inc(&rd
->refcount
);
7439 cpumask_set_cpu(rq
->cpu
, rd
->span
);
7440 if (cpumask_test_cpu(rq
->cpu
, cpu_online_mask
))
7443 spin_unlock_irqrestore(&rq
->lock
, flags
);
7446 free_rootdomain(old_rd
);
7449 static int __init_refok
init_rootdomain(struct root_domain
*rd
, bool bootmem
)
7451 memset(rd
, 0, sizeof(*rd
));
7454 alloc_bootmem_cpumask_var(&def_root_domain
.span
);
7455 alloc_bootmem_cpumask_var(&def_root_domain
.online
);
7456 alloc_bootmem_cpumask_var(&def_root_domain
.rto_mask
);
7457 cpupri_init(&rd
->cpupri
, true);
7461 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
7463 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
7465 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
7468 if (cpupri_init(&rd
->cpupri
, false) != 0)
7473 free_cpumask_var(rd
->rto_mask
);
7475 free_cpumask_var(rd
->online
);
7477 free_cpumask_var(rd
->span
);
7482 static void init_defrootdomain(void)
7484 init_rootdomain(&def_root_domain
, true);
7486 atomic_set(&def_root_domain
.refcount
, 1);
7489 static struct root_domain
*alloc_rootdomain(void)
7491 struct root_domain
*rd
;
7493 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
7497 if (init_rootdomain(rd
, false) != 0) {
7506 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7507 * hold the hotplug lock.
7510 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
7512 struct rq
*rq
= cpu_rq(cpu
);
7513 struct sched_domain
*tmp
;
7515 /* Remove the sched domains which do not contribute to scheduling. */
7516 for (tmp
= sd
; tmp
; ) {
7517 struct sched_domain
*parent
= tmp
->parent
;
7521 if (sd_parent_degenerate(tmp
, parent
)) {
7522 tmp
->parent
= parent
->parent
;
7524 parent
->parent
->child
= tmp
;
7529 if (sd
&& sd_degenerate(sd
)) {
7535 sched_domain_debug(sd
, cpu
);
7537 rq_attach_root(rq
, rd
);
7538 rcu_assign_pointer(rq
->sd
, sd
);
7541 /* cpus with isolated domains */
7542 static cpumask_var_t cpu_isolated_map
;
7544 /* Setup the mask of cpus configured for isolated domains */
7545 static int __init
isolated_cpu_setup(char *str
)
7547 cpulist_parse(str
, cpu_isolated_map
);
7551 __setup("isolcpus=", isolated_cpu_setup
);
7554 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7555 * to a function which identifies what group(along with sched group) a CPU
7556 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7557 * (due to the fact that we keep track of groups covered with a struct cpumask).
7559 * init_sched_build_groups will build a circular linked list of the groups
7560 * covered by the given span, and will set each group's ->cpumask correctly,
7561 * and ->cpu_power to 0.
7564 init_sched_build_groups(const struct cpumask
*span
,
7565 const struct cpumask
*cpu_map
,
7566 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
7567 struct sched_group
**sg
,
7568 struct cpumask
*tmpmask
),
7569 struct cpumask
*covered
, struct cpumask
*tmpmask
)
7571 struct sched_group
*first
= NULL
, *last
= NULL
;
7574 cpumask_clear(covered
);
7576 for_each_cpu(i
, span
) {
7577 struct sched_group
*sg
;
7578 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
7581 if (cpumask_test_cpu(i
, covered
))
7584 cpumask_clear(sched_group_cpus(sg
));
7585 sg
->__cpu_power
= 0;
7587 for_each_cpu(j
, span
) {
7588 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
7591 cpumask_set_cpu(j
, covered
);
7592 cpumask_set_cpu(j
, sched_group_cpus(sg
));
7603 #define SD_NODES_PER_DOMAIN 16
7608 * find_next_best_node - find the next node to include in a sched_domain
7609 * @node: node whose sched_domain we're building
7610 * @used_nodes: nodes already in the sched_domain
7612 * Find the next node to include in a given scheduling domain. Simply
7613 * finds the closest node not already in the @used_nodes map.
7615 * Should use nodemask_t.
7617 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
7619 int i
, n
, val
, min_val
, best_node
= 0;
7623 for (i
= 0; i
< nr_node_ids
; i
++) {
7624 /* Start at @node */
7625 n
= (node
+ i
) % nr_node_ids
;
7627 if (!nr_cpus_node(n
))
7630 /* Skip already used nodes */
7631 if (node_isset(n
, *used_nodes
))
7634 /* Simple min distance search */
7635 val
= node_distance(node
, n
);
7637 if (val
< min_val
) {
7643 node_set(best_node
, *used_nodes
);
7648 * sched_domain_node_span - get a cpumask for a node's sched_domain
7649 * @node: node whose cpumask we're constructing
7650 * @span: resulting cpumask
7652 * Given a node, construct a good cpumask for its sched_domain to span. It
7653 * should be one that prevents unnecessary balancing, but also spreads tasks
7656 static void sched_domain_node_span(int node
, struct cpumask
*span
)
7658 nodemask_t used_nodes
;
7661 cpumask_clear(span
);
7662 nodes_clear(used_nodes
);
7664 cpumask_or(span
, span
, cpumask_of_node(node
));
7665 node_set(node
, used_nodes
);
7667 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
7668 int next_node
= find_next_best_node(node
, &used_nodes
);
7670 cpumask_or(span
, span
, cpumask_of_node(next_node
));
7673 #endif /* CONFIG_NUMA */
7675 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
7678 * The cpus mask in sched_group and sched_domain hangs off the end.
7679 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7680 * for nr_cpu_ids < CONFIG_NR_CPUS.
7682 struct static_sched_group
{
7683 struct sched_group sg
;
7684 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
7687 struct static_sched_domain
{
7688 struct sched_domain sd
;
7689 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
7693 * SMT sched-domains:
7695 #ifdef CONFIG_SCHED_SMT
7696 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
7697 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_cpus
);
7700 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
7701 struct sched_group
**sg
, struct cpumask
*unused
)
7704 *sg
= &per_cpu(sched_group_cpus
, cpu
).sg
;
7707 #endif /* CONFIG_SCHED_SMT */
7710 * multi-core sched-domains:
7712 #ifdef CONFIG_SCHED_MC
7713 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
7714 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
7715 #endif /* CONFIG_SCHED_MC */
7717 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7719 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
7720 struct sched_group
**sg
, struct cpumask
*mask
)
7724 cpumask_and(mask
, &per_cpu(cpu_sibling_map
, cpu
), cpu_map
);
7725 group
= cpumask_first(mask
);
7727 *sg
= &per_cpu(sched_group_core
, group
).sg
;
7730 #elif defined(CONFIG_SCHED_MC)
7732 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
7733 struct sched_group
**sg
, struct cpumask
*unused
)
7736 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
7741 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
7742 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
7745 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
7746 struct sched_group
**sg
, struct cpumask
*mask
)
7749 #ifdef CONFIG_SCHED_MC
7750 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
7751 group
= cpumask_first(mask
);
7752 #elif defined(CONFIG_SCHED_SMT)
7753 cpumask_and(mask
, &per_cpu(cpu_sibling_map
, cpu
), cpu_map
);
7754 group
= cpumask_first(mask
);
7759 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
7765 * The init_sched_build_groups can't handle what we want to do with node
7766 * groups, so roll our own. Now each node has its own list of groups which
7767 * gets dynamically allocated.
7769 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
7770 static struct sched_group
***sched_group_nodes_bycpu
;
7772 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
7773 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
7775 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
7776 struct sched_group
**sg
,
7777 struct cpumask
*nodemask
)
7781 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
7782 group
= cpumask_first(nodemask
);
7785 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
7789 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
7791 struct sched_group
*sg
= group_head
;
7797 for_each_cpu(j
, sched_group_cpus(sg
)) {
7798 struct sched_domain
*sd
;
7800 sd
= &per_cpu(phys_domains
, j
).sd
;
7801 if (j
!= cpumask_first(sched_group_cpus(sd
->groups
))) {
7803 * Only add "power" once for each
7809 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
7812 } while (sg
!= group_head
);
7814 #endif /* CONFIG_NUMA */
7817 /* Free memory allocated for various sched_group structures */
7818 static void free_sched_groups(const struct cpumask
*cpu_map
,
7819 struct cpumask
*nodemask
)
7823 for_each_cpu(cpu
, cpu_map
) {
7824 struct sched_group
**sched_group_nodes
7825 = sched_group_nodes_bycpu
[cpu
];
7827 if (!sched_group_nodes
)
7830 for (i
= 0; i
< nr_node_ids
; i
++) {
7831 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
7833 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
7834 if (cpumask_empty(nodemask
))
7844 if (oldsg
!= sched_group_nodes
[i
])
7847 kfree(sched_group_nodes
);
7848 sched_group_nodes_bycpu
[cpu
] = NULL
;
7851 #else /* !CONFIG_NUMA */
7852 static void free_sched_groups(const struct cpumask
*cpu_map
,
7853 struct cpumask
*nodemask
)
7856 #endif /* CONFIG_NUMA */
7859 * Initialize sched groups cpu_power.
7861 * cpu_power indicates the capacity of sched group, which is used while
7862 * distributing the load between different sched groups in a sched domain.
7863 * Typically cpu_power for all the groups in a sched domain will be same unless
7864 * there are asymmetries in the topology. If there are asymmetries, group
7865 * having more cpu_power will pickup more load compared to the group having
7868 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7869 * the maximum number of tasks a group can handle in the presence of other idle
7870 * or lightly loaded groups in the same sched domain.
7872 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7874 struct sched_domain
*child
;
7875 struct sched_group
*group
;
7877 WARN_ON(!sd
|| !sd
->groups
);
7879 if (cpu
!= cpumask_first(sched_group_cpus(sd
->groups
)))
7884 sd
->groups
->__cpu_power
= 0;
7887 * For perf policy, if the groups in child domain share resources
7888 * (for example cores sharing some portions of the cache hierarchy
7889 * or SMT), then set this domain groups cpu_power such that each group
7890 * can handle only one task, when there are other idle groups in the
7891 * same sched domain.
7893 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
7895 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
7896 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
7901 * add cpu_power of each child group to this groups cpu_power
7903 group
= child
->groups
;
7905 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
7906 group
= group
->next
;
7907 } while (group
!= child
->groups
);
7911 * Initializers for schedule domains
7912 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7915 #ifdef CONFIG_SCHED_DEBUG
7916 # define SD_INIT_NAME(sd, type) sd->name = #type
7918 # define SD_INIT_NAME(sd, type) do { } while (0)
7921 #define SD_INIT(sd, type) sd_init_##type(sd)
7923 #define SD_INIT_FUNC(type) \
7924 static noinline void sd_init_##type(struct sched_domain *sd) \
7926 memset(sd, 0, sizeof(*sd)); \
7927 *sd = SD_##type##_INIT; \
7928 sd->level = SD_LV_##type; \
7929 SD_INIT_NAME(sd, type); \
7934 SD_INIT_FUNC(ALLNODES
)
7937 #ifdef CONFIG_SCHED_SMT
7938 SD_INIT_FUNC(SIBLING
)
7940 #ifdef CONFIG_SCHED_MC
7944 static int default_relax_domain_level
= -1;
7946 static int __init
setup_relax_domain_level(char *str
)
7950 val
= simple_strtoul(str
, NULL
, 0);
7951 if (val
< SD_LV_MAX
)
7952 default_relax_domain_level
= val
;
7956 __setup("relax_domain_level=", setup_relax_domain_level
);
7958 static void set_domain_attribute(struct sched_domain
*sd
,
7959 struct sched_domain_attr
*attr
)
7963 if (!attr
|| attr
->relax_domain_level
< 0) {
7964 if (default_relax_domain_level
< 0)
7967 request
= default_relax_domain_level
;
7969 request
= attr
->relax_domain_level
;
7970 if (request
< sd
->level
) {
7971 /* turn off idle balance on this domain */
7972 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
7974 /* turn on idle balance on this domain */
7975 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
7980 * Build sched domains for a given set of cpus and attach the sched domains
7981 * to the individual cpus
7983 static int __build_sched_domains(const struct cpumask
*cpu_map
,
7984 struct sched_domain_attr
*attr
)
7986 int i
, err
= -ENOMEM
;
7987 struct root_domain
*rd
;
7988 cpumask_var_t nodemask
, this_sibling_map
, this_core_map
, send_covered
,
7991 cpumask_var_t domainspan
, covered
, notcovered
;
7992 struct sched_group
**sched_group_nodes
= NULL
;
7993 int sd_allnodes
= 0;
7995 if (!alloc_cpumask_var(&domainspan
, GFP_KERNEL
))
7997 if (!alloc_cpumask_var(&covered
, GFP_KERNEL
))
7998 goto free_domainspan
;
7999 if (!alloc_cpumask_var(¬covered
, GFP_KERNEL
))
8003 if (!alloc_cpumask_var(&nodemask
, GFP_KERNEL
))
8004 goto free_notcovered
;
8005 if (!alloc_cpumask_var(&this_sibling_map
, GFP_KERNEL
))
8007 if (!alloc_cpumask_var(&this_core_map
, GFP_KERNEL
))
8008 goto free_this_sibling_map
;
8009 if (!alloc_cpumask_var(&send_covered
, GFP_KERNEL
))
8010 goto free_this_core_map
;
8011 if (!alloc_cpumask_var(&tmpmask
, GFP_KERNEL
))
8012 goto free_send_covered
;
8016 * Allocate the per-node list of sched groups
8018 sched_group_nodes
= kcalloc(nr_node_ids
, sizeof(struct sched_group
*),
8020 if (!sched_group_nodes
) {
8021 printk(KERN_WARNING
"Can not alloc sched group node list\n");
8026 rd
= alloc_rootdomain();
8028 printk(KERN_WARNING
"Cannot alloc root domain\n");
8029 goto free_sched_groups
;
8033 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = sched_group_nodes
;
8037 * Set up domains for cpus specified by the cpu_map.
8039 for_each_cpu(i
, cpu_map
) {
8040 struct sched_domain
*sd
= NULL
, *p
;
8042 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(i
)), cpu_map
);
8045 if (cpumask_weight(cpu_map
) >
8046 SD_NODES_PER_DOMAIN
*cpumask_weight(nodemask
)) {
8047 sd
= &per_cpu(allnodes_domains
, i
).sd
;
8048 SD_INIT(sd
, ALLNODES
);
8049 set_domain_attribute(sd
, attr
);
8050 cpumask_copy(sched_domain_span(sd
), cpu_map
);
8051 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8057 sd
= &per_cpu(node_domains
, i
).sd
;
8059 set_domain_attribute(sd
, attr
);
8060 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
8064 cpumask_and(sched_domain_span(sd
),
8065 sched_domain_span(sd
), cpu_map
);
8069 sd
= &per_cpu(phys_domains
, i
).sd
;
8071 set_domain_attribute(sd
, attr
);
8072 cpumask_copy(sched_domain_span(sd
), nodemask
);
8076 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8078 #ifdef CONFIG_SCHED_MC
8080 sd
= &per_cpu(core_domains
, i
).sd
;
8082 set_domain_attribute(sd
, attr
);
8083 cpumask_and(sched_domain_span(sd
), cpu_map
,
8084 cpu_coregroup_mask(i
));
8087 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8090 #ifdef CONFIG_SCHED_SMT
8092 sd
= &per_cpu(cpu_domains
, i
).sd
;
8093 SD_INIT(sd
, SIBLING
);
8094 set_domain_attribute(sd
, attr
);
8095 cpumask_and(sched_domain_span(sd
),
8096 &per_cpu(cpu_sibling_map
, i
), cpu_map
);
8099 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8103 #ifdef CONFIG_SCHED_SMT
8104 /* Set up CPU (sibling) groups */
8105 for_each_cpu(i
, cpu_map
) {
8106 cpumask_and(this_sibling_map
,
8107 &per_cpu(cpu_sibling_map
, i
), cpu_map
);
8108 if (i
!= cpumask_first(this_sibling_map
))
8111 init_sched_build_groups(this_sibling_map
, cpu_map
,
8113 send_covered
, tmpmask
);
8117 #ifdef CONFIG_SCHED_MC
8118 /* Set up multi-core groups */
8119 for_each_cpu(i
, cpu_map
) {
8120 cpumask_and(this_core_map
, cpu_coregroup_mask(i
), cpu_map
);
8121 if (i
!= cpumask_first(this_core_map
))
8124 init_sched_build_groups(this_core_map
, cpu_map
,
8126 send_covered
, tmpmask
);
8130 /* Set up physical groups */
8131 for (i
= 0; i
< nr_node_ids
; i
++) {
8132 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8133 if (cpumask_empty(nodemask
))
8136 init_sched_build_groups(nodemask
, cpu_map
,
8138 send_covered
, tmpmask
);
8142 /* Set up node groups */
8144 init_sched_build_groups(cpu_map
, cpu_map
,
8145 &cpu_to_allnodes_group
,
8146 send_covered
, tmpmask
);
8149 for (i
= 0; i
< nr_node_ids
; i
++) {
8150 /* Set up node groups */
8151 struct sched_group
*sg
, *prev
;
8154 cpumask_clear(covered
);
8155 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8156 if (cpumask_empty(nodemask
)) {
8157 sched_group_nodes
[i
] = NULL
;
8161 sched_domain_node_span(i
, domainspan
);
8162 cpumask_and(domainspan
, domainspan
, cpu_map
);
8164 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8167 printk(KERN_WARNING
"Can not alloc domain group for "
8171 sched_group_nodes
[i
] = sg
;
8172 for_each_cpu(j
, nodemask
) {
8173 struct sched_domain
*sd
;
8175 sd
= &per_cpu(node_domains
, j
).sd
;
8178 sg
->__cpu_power
= 0;
8179 cpumask_copy(sched_group_cpus(sg
), nodemask
);
8181 cpumask_or(covered
, covered
, nodemask
);
8184 for (j
= 0; j
< nr_node_ids
; j
++) {
8185 int n
= (i
+ j
) % nr_node_ids
;
8187 cpumask_complement(notcovered
, covered
);
8188 cpumask_and(tmpmask
, notcovered
, cpu_map
);
8189 cpumask_and(tmpmask
, tmpmask
, domainspan
);
8190 if (cpumask_empty(tmpmask
))
8193 cpumask_and(tmpmask
, tmpmask
, cpumask_of_node(n
));
8194 if (cpumask_empty(tmpmask
))
8197 sg
= kmalloc_node(sizeof(struct sched_group
) +
8202 "Can not alloc domain group for node %d\n", j
);
8205 sg
->__cpu_power
= 0;
8206 cpumask_copy(sched_group_cpus(sg
), tmpmask
);
8207 sg
->next
= prev
->next
;
8208 cpumask_or(covered
, covered
, tmpmask
);
8215 /* Calculate CPU power for physical packages and nodes */
8216 #ifdef CONFIG_SCHED_SMT
8217 for_each_cpu(i
, cpu_map
) {
8218 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
).sd
;
8220 init_sched_groups_power(i
, sd
);
8223 #ifdef CONFIG_SCHED_MC
8224 for_each_cpu(i
, cpu_map
) {
8225 struct sched_domain
*sd
= &per_cpu(core_domains
, i
).sd
;
8227 init_sched_groups_power(i
, sd
);
8231 for_each_cpu(i
, cpu_map
) {
8232 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
).sd
;
8234 init_sched_groups_power(i
, sd
);
8238 for (i
= 0; i
< nr_node_ids
; i
++)
8239 init_numa_sched_groups_power(sched_group_nodes
[i
]);
8242 struct sched_group
*sg
;
8244 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
8246 init_numa_sched_groups_power(sg
);
8250 /* Attach the domains */
8251 for_each_cpu(i
, cpu_map
) {
8252 struct sched_domain
*sd
;
8253 #ifdef CONFIG_SCHED_SMT
8254 sd
= &per_cpu(cpu_domains
, i
).sd
;
8255 #elif defined(CONFIG_SCHED_MC)
8256 sd
= &per_cpu(core_domains
, i
).sd
;
8258 sd
= &per_cpu(phys_domains
, i
).sd
;
8260 cpu_attach_domain(sd
, rd
, i
);
8266 free_cpumask_var(tmpmask
);
8268 free_cpumask_var(send_covered
);
8270 free_cpumask_var(this_core_map
);
8271 free_this_sibling_map
:
8272 free_cpumask_var(this_sibling_map
);
8274 free_cpumask_var(nodemask
);
8277 free_cpumask_var(notcovered
);
8279 free_cpumask_var(covered
);
8281 free_cpumask_var(domainspan
);
8288 kfree(sched_group_nodes
);
8294 free_sched_groups(cpu_map
, tmpmask
);
8295 free_rootdomain(rd
);
8300 static int build_sched_domains(const struct cpumask
*cpu_map
)
8302 return __build_sched_domains(cpu_map
, NULL
);
8305 static struct cpumask
*doms_cur
; /* current sched domains */
8306 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
8307 static struct sched_domain_attr
*dattr_cur
;
8308 /* attribues of custom domains in 'doms_cur' */
8311 * Special case: If a kmalloc of a doms_cur partition (array of
8312 * cpumask) fails, then fallback to a single sched domain,
8313 * as determined by the single cpumask fallback_doms.
8315 static cpumask_var_t fallback_doms
;
8318 * arch_update_cpu_topology lets virtualized architectures update the
8319 * cpu core maps. It is supposed to return 1 if the topology changed
8320 * or 0 if it stayed the same.
8322 int __attribute__((weak
)) arch_update_cpu_topology(void)
8328 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8329 * For now this just excludes isolated cpus, but could be used to
8330 * exclude other special cases in the future.
8332 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
8336 arch_update_cpu_topology();
8338 doms_cur
= kmalloc(cpumask_size(), GFP_KERNEL
);
8340 doms_cur
= fallback_doms
;
8341 cpumask_andnot(doms_cur
, cpu_map
, cpu_isolated_map
);
8343 err
= build_sched_domains(doms_cur
);
8344 register_sched_domain_sysctl();
8349 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
8350 struct cpumask
*tmpmask
)
8352 free_sched_groups(cpu_map
, tmpmask
);
8356 * Detach sched domains from a group of cpus specified in cpu_map
8357 * These cpus will now be attached to the NULL domain
8359 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
8361 /* Save because hotplug lock held. */
8362 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
8365 for_each_cpu(i
, cpu_map
)
8366 cpu_attach_domain(NULL
, &def_root_domain
, i
);
8367 synchronize_sched();
8368 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
8371 /* handle null as "default" */
8372 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
8373 struct sched_domain_attr
*new, int idx_new
)
8375 struct sched_domain_attr tmp
;
8382 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
8383 new ? (new + idx_new
) : &tmp
,
8384 sizeof(struct sched_domain_attr
));
8388 * Partition sched domains as specified by the 'ndoms_new'
8389 * cpumasks in the array doms_new[] of cpumasks. This compares
8390 * doms_new[] to the current sched domain partitioning, doms_cur[].
8391 * It destroys each deleted domain and builds each new domain.
8393 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8394 * The masks don't intersect (don't overlap.) We should setup one
8395 * sched domain for each mask. CPUs not in any of the cpumasks will
8396 * not be load balanced. If the same cpumask appears both in the
8397 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8400 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8401 * ownership of it and will kfree it when done with it. If the caller
8402 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8403 * ndoms_new == 1, and partition_sched_domains() will fallback to
8404 * the single partition 'fallback_doms', it also forces the domains
8407 * If doms_new == NULL it will be replaced with cpu_online_mask.
8408 * ndoms_new == 0 is a special case for destroying existing domains,
8409 * and it will not create the default domain.
8411 * Call with hotplug lock held
8413 /* FIXME: Change to struct cpumask *doms_new[] */
8414 void partition_sched_domains(int ndoms_new
, struct cpumask
*doms_new
,
8415 struct sched_domain_attr
*dattr_new
)
8420 mutex_lock(&sched_domains_mutex
);
8422 /* always unregister in case we don't destroy any domains */
8423 unregister_sched_domain_sysctl();
8425 /* Let architecture update cpu core mappings. */
8426 new_topology
= arch_update_cpu_topology();
8428 n
= doms_new
? ndoms_new
: 0;
8430 /* Destroy deleted domains */
8431 for (i
= 0; i
< ndoms_cur
; i
++) {
8432 for (j
= 0; j
< n
&& !new_topology
; j
++) {
8433 if (cpumask_equal(&doms_cur
[i
], &doms_new
[j
])
8434 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
8437 /* no match - a current sched domain not in new doms_new[] */
8438 detach_destroy_domains(doms_cur
+ i
);
8443 if (doms_new
== NULL
) {
8445 doms_new
= fallback_doms
;
8446 cpumask_andnot(&doms_new
[0], cpu_online_mask
, cpu_isolated_map
);
8447 WARN_ON_ONCE(dattr_new
);
8450 /* Build new domains */
8451 for (i
= 0; i
< ndoms_new
; i
++) {
8452 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
8453 if (cpumask_equal(&doms_new
[i
], &doms_cur
[j
])
8454 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
8457 /* no match - add a new doms_new */
8458 __build_sched_domains(doms_new
+ i
,
8459 dattr_new
? dattr_new
+ i
: NULL
);
8464 /* Remember the new sched domains */
8465 if (doms_cur
!= fallback_doms
)
8467 kfree(dattr_cur
); /* kfree(NULL) is safe */
8468 doms_cur
= doms_new
;
8469 dattr_cur
= dattr_new
;
8470 ndoms_cur
= ndoms_new
;
8472 register_sched_domain_sysctl();
8474 mutex_unlock(&sched_domains_mutex
);
8477 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8478 static void arch_reinit_sched_domains(void)
8482 /* Destroy domains first to force the rebuild */
8483 partition_sched_domains(0, NULL
, NULL
);
8485 rebuild_sched_domains();
8489 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
8491 unsigned int level
= 0;
8493 if (sscanf(buf
, "%u", &level
) != 1)
8497 * level is always be positive so don't check for
8498 * level < POWERSAVINGS_BALANCE_NONE which is 0
8499 * What happens on 0 or 1 byte write,
8500 * need to check for count as well?
8503 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
8507 sched_smt_power_savings
= level
;
8509 sched_mc_power_savings
= level
;
8511 arch_reinit_sched_domains();
8516 #ifdef CONFIG_SCHED_MC
8517 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
8520 return sprintf(page
, "%u\n", sched_mc_power_savings
);
8522 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
8523 const char *buf
, size_t count
)
8525 return sched_power_savings_store(buf
, count
, 0);
8527 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
8528 sched_mc_power_savings_show
,
8529 sched_mc_power_savings_store
);
8532 #ifdef CONFIG_SCHED_SMT
8533 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
8536 return sprintf(page
, "%u\n", sched_smt_power_savings
);
8538 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
8539 const char *buf
, size_t count
)
8541 return sched_power_savings_store(buf
, count
, 1);
8543 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
8544 sched_smt_power_savings_show
,
8545 sched_smt_power_savings_store
);
8548 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
8552 #ifdef CONFIG_SCHED_SMT
8554 err
= sysfs_create_file(&cls
->kset
.kobj
,
8555 &attr_sched_smt_power_savings
.attr
);
8557 #ifdef CONFIG_SCHED_MC
8558 if (!err
&& mc_capable())
8559 err
= sysfs_create_file(&cls
->kset
.kobj
,
8560 &attr_sched_mc_power_savings
.attr
);
8564 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8566 #ifndef CONFIG_CPUSETS
8568 * Add online and remove offline CPUs from the scheduler domains.
8569 * When cpusets are enabled they take over this function.
8571 static int update_sched_domains(struct notifier_block
*nfb
,
8572 unsigned long action
, void *hcpu
)
8576 case CPU_ONLINE_FROZEN
:
8578 case CPU_DEAD_FROZEN
:
8579 partition_sched_domains(1, NULL
, NULL
);
8588 static int update_runtime(struct notifier_block
*nfb
,
8589 unsigned long action
, void *hcpu
)
8591 int cpu
= (int)(long)hcpu
;
8594 case CPU_DOWN_PREPARE
:
8595 case CPU_DOWN_PREPARE_FROZEN
:
8596 disable_runtime(cpu_rq(cpu
));
8599 case CPU_DOWN_FAILED
:
8600 case CPU_DOWN_FAILED_FROZEN
:
8602 case CPU_ONLINE_FROZEN
:
8603 enable_runtime(cpu_rq(cpu
));
8611 void __init
sched_init_smp(void)
8613 cpumask_var_t non_isolated_cpus
;
8615 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
8617 #if defined(CONFIG_NUMA)
8618 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
8620 BUG_ON(sched_group_nodes_bycpu
== NULL
);
8623 mutex_lock(&sched_domains_mutex
);
8624 arch_init_sched_domains(cpu_online_mask
);
8625 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
8626 if (cpumask_empty(non_isolated_cpus
))
8627 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
8628 mutex_unlock(&sched_domains_mutex
);
8631 #ifndef CONFIG_CPUSETS
8632 /* XXX: Theoretical race here - CPU may be hotplugged now */
8633 hotcpu_notifier(update_sched_domains
, 0);
8636 /* RT runtime code needs to handle some hotplug events */
8637 hotcpu_notifier(update_runtime
, 0);
8641 /* Move init over to a non-isolated CPU */
8642 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
8644 sched_init_granularity();
8645 free_cpumask_var(non_isolated_cpus
);
8647 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
8648 init_sched_rt_class();
8651 void __init
sched_init_smp(void)
8653 sched_init_granularity();
8655 #endif /* CONFIG_SMP */
8657 int in_sched_functions(unsigned long addr
)
8659 return in_lock_functions(addr
) ||
8660 (addr
>= (unsigned long)__sched_text_start
8661 && addr
< (unsigned long)__sched_text_end
);
8664 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
8666 cfs_rq
->tasks_timeline
= RB_ROOT
;
8667 INIT_LIST_HEAD(&cfs_rq
->tasks
);
8668 #ifdef CONFIG_FAIR_GROUP_SCHED
8671 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
8674 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
8676 struct rt_prio_array
*array
;
8679 array
= &rt_rq
->active
;
8680 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
8681 INIT_LIST_HEAD(array
->queue
+ i
);
8682 __clear_bit(i
, array
->bitmap
);
8684 /* delimiter for bitsearch: */
8685 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
8687 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8688 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
8690 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
8694 rt_rq
->rt_nr_migratory
= 0;
8695 rt_rq
->overloaded
= 0;
8696 plist_head_init(&rq
->rt
.pushable_tasks
, &rq
->lock
);
8700 rt_rq
->rt_throttled
= 0;
8701 rt_rq
->rt_runtime
= 0;
8702 spin_lock_init(&rt_rq
->rt_runtime_lock
);
8704 #ifdef CONFIG_RT_GROUP_SCHED
8705 rt_rq
->rt_nr_boosted
= 0;
8710 #ifdef CONFIG_FAIR_GROUP_SCHED
8711 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8712 struct sched_entity
*se
, int cpu
, int add
,
8713 struct sched_entity
*parent
)
8715 struct rq
*rq
= cpu_rq(cpu
);
8716 tg
->cfs_rq
[cpu
] = cfs_rq
;
8717 init_cfs_rq(cfs_rq
, rq
);
8720 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
8723 /* se could be NULL for init_task_group */
8728 se
->cfs_rq
= &rq
->cfs
;
8730 se
->cfs_rq
= parent
->my_q
;
8733 se
->load
.weight
= tg
->shares
;
8734 se
->load
.inv_weight
= 0;
8735 se
->parent
= parent
;
8739 #ifdef CONFIG_RT_GROUP_SCHED
8740 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
8741 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
8742 struct sched_rt_entity
*parent
)
8744 struct rq
*rq
= cpu_rq(cpu
);
8746 tg
->rt_rq
[cpu
] = rt_rq
;
8747 init_rt_rq(rt_rq
, rq
);
8749 rt_rq
->rt_se
= rt_se
;
8750 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8752 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
8754 tg
->rt_se
[cpu
] = rt_se
;
8759 rt_se
->rt_rq
= &rq
->rt
;
8761 rt_se
->rt_rq
= parent
->my_q
;
8763 rt_se
->my_q
= rt_rq
;
8764 rt_se
->parent
= parent
;
8765 INIT_LIST_HEAD(&rt_se
->run_list
);
8769 void __init
sched_init(void)
8772 unsigned long alloc_size
= 0, ptr
;
8774 #ifdef CONFIG_FAIR_GROUP_SCHED
8775 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8777 #ifdef CONFIG_RT_GROUP_SCHED
8778 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8780 #ifdef CONFIG_USER_SCHED
8784 * As sched_init() is called before page_alloc is setup,
8785 * we use alloc_bootmem().
8788 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
8790 #ifdef CONFIG_FAIR_GROUP_SCHED
8791 init_task_group
.se
= (struct sched_entity
**)ptr
;
8792 ptr
+= nr_cpu_ids
* sizeof(void **);
8794 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8795 ptr
+= nr_cpu_ids
* sizeof(void **);
8797 #ifdef CONFIG_USER_SCHED
8798 root_task_group
.se
= (struct sched_entity
**)ptr
;
8799 ptr
+= nr_cpu_ids
* sizeof(void **);
8801 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8802 ptr
+= nr_cpu_ids
* sizeof(void **);
8803 #endif /* CONFIG_USER_SCHED */
8804 #endif /* CONFIG_FAIR_GROUP_SCHED */
8805 #ifdef CONFIG_RT_GROUP_SCHED
8806 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8807 ptr
+= nr_cpu_ids
* sizeof(void **);
8809 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8810 ptr
+= nr_cpu_ids
* sizeof(void **);
8812 #ifdef CONFIG_USER_SCHED
8813 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8814 ptr
+= nr_cpu_ids
* sizeof(void **);
8816 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8817 ptr
+= nr_cpu_ids
* sizeof(void **);
8818 #endif /* CONFIG_USER_SCHED */
8819 #endif /* CONFIG_RT_GROUP_SCHED */
8823 init_defrootdomain();
8826 init_rt_bandwidth(&def_rt_bandwidth
,
8827 global_rt_period(), global_rt_runtime());
8829 #ifdef CONFIG_RT_GROUP_SCHED
8830 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
8831 global_rt_period(), global_rt_runtime());
8832 #ifdef CONFIG_USER_SCHED
8833 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8834 global_rt_period(), RUNTIME_INF
);
8835 #endif /* CONFIG_USER_SCHED */
8836 #endif /* CONFIG_RT_GROUP_SCHED */
8838 #ifdef CONFIG_GROUP_SCHED
8839 list_add(&init_task_group
.list
, &task_groups
);
8840 INIT_LIST_HEAD(&init_task_group
.children
);
8842 #ifdef CONFIG_USER_SCHED
8843 INIT_LIST_HEAD(&root_task_group
.children
);
8844 init_task_group
.parent
= &root_task_group
;
8845 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
8846 #endif /* CONFIG_USER_SCHED */
8847 #endif /* CONFIG_GROUP_SCHED */
8849 for_each_possible_cpu(i
) {
8853 spin_lock_init(&rq
->lock
);
8855 init_cfs_rq(&rq
->cfs
, rq
);
8856 init_rt_rq(&rq
->rt
, rq
);
8857 #ifdef CONFIG_FAIR_GROUP_SCHED
8858 init_task_group
.shares
= init_task_group_load
;
8859 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8860 #ifdef CONFIG_CGROUP_SCHED
8862 * How much cpu bandwidth does init_task_group get?
8864 * In case of task-groups formed thr' the cgroup filesystem, it
8865 * gets 100% of the cpu resources in the system. This overall
8866 * system cpu resource is divided among the tasks of
8867 * init_task_group and its child task-groups in a fair manner,
8868 * based on each entity's (task or task-group's) weight
8869 * (se->load.weight).
8871 * In other words, if init_task_group has 10 tasks of weight
8872 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8873 * then A0's share of the cpu resource is:
8875 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8877 * We achieve this by letting init_task_group's tasks sit
8878 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8880 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
8881 #elif defined CONFIG_USER_SCHED
8882 root_task_group
.shares
= NICE_0_LOAD
;
8883 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
8885 * In case of task-groups formed thr' the user id of tasks,
8886 * init_task_group represents tasks belonging to root user.
8887 * Hence it forms a sibling of all subsequent groups formed.
8888 * In this case, init_task_group gets only a fraction of overall
8889 * system cpu resource, based on the weight assigned to root
8890 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8891 * by letting tasks of init_task_group sit in a separate cfs_rq
8892 * (init_cfs_rq) and having one entity represent this group of
8893 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8895 init_tg_cfs_entry(&init_task_group
,
8896 &per_cpu(init_cfs_rq
, i
),
8897 &per_cpu(init_sched_entity
, i
), i
, 1,
8898 root_task_group
.se
[i
]);
8901 #endif /* CONFIG_FAIR_GROUP_SCHED */
8903 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8904 #ifdef CONFIG_RT_GROUP_SCHED
8905 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
8906 #ifdef CONFIG_CGROUP_SCHED
8907 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
8908 #elif defined CONFIG_USER_SCHED
8909 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
8910 init_tg_rt_entry(&init_task_group
,
8911 &per_cpu(init_rt_rq
, i
),
8912 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
8913 root_task_group
.rt_se
[i
]);
8917 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
8918 rq
->cpu_load
[j
] = 0;
8922 rq
->active_balance
= 0;
8923 rq
->next_balance
= jiffies
;
8927 rq
->migration_thread
= NULL
;
8928 INIT_LIST_HEAD(&rq
->migration_queue
);
8929 rq_attach_root(rq
, &def_root_domain
);
8932 atomic_set(&rq
->nr_iowait
, 0);
8935 set_load_weight(&init_task
);
8937 #ifdef CONFIG_PREEMPT_NOTIFIERS
8938 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
8942 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8945 #ifdef CONFIG_RT_MUTEXES
8946 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
8950 * The boot idle thread does lazy MMU switching as well:
8952 atomic_inc(&init_mm
.mm_count
);
8953 enter_lazy_tlb(&init_mm
, current
);
8956 * Make us the idle thread. Technically, schedule() should not be
8957 * called from this thread, however somewhere below it might be,
8958 * but because we are the idle thread, we just pick up running again
8959 * when this runqueue becomes "idle".
8961 init_idle(current
, smp_processor_id());
8963 * During early bootup we pretend to be a normal task:
8965 current
->sched_class
= &fair_sched_class
;
8967 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8968 alloc_bootmem_cpumask_var(&nohz_cpu_mask
);
8971 alloc_bootmem_cpumask_var(&nohz
.cpu_mask
);
8973 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
8976 scheduler_running
= 1;
8979 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8980 void __might_sleep(char *file
, int line
)
8983 static unsigned long prev_jiffy
; /* ratelimiting */
8985 if ((!in_atomic() && !irqs_disabled()) ||
8986 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
8988 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8990 prev_jiffy
= jiffies
;
8993 "BUG: sleeping function called from invalid context at %s:%d\n",
8996 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8997 in_atomic(), irqs_disabled(),
8998 current
->pid
, current
->comm
);
9000 debug_show_held_locks(current
);
9001 if (irqs_disabled())
9002 print_irqtrace_events(current
);
9006 EXPORT_SYMBOL(__might_sleep
);
9009 #ifdef CONFIG_MAGIC_SYSRQ
9010 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
9014 update_rq_clock(rq
);
9015 on_rq
= p
->se
.on_rq
;
9017 deactivate_task(rq
, p
, 0);
9018 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
9020 activate_task(rq
, p
, 0);
9021 resched_task(rq
->curr
);
9025 void normalize_rt_tasks(void)
9027 struct task_struct
*g
, *p
;
9028 unsigned long flags
;
9031 read_lock_irqsave(&tasklist_lock
, flags
);
9032 do_each_thread(g
, p
) {
9034 * Only normalize user tasks:
9039 p
->se
.exec_start
= 0;
9040 #ifdef CONFIG_SCHEDSTATS
9041 p
->se
.wait_start
= 0;
9042 p
->se
.sleep_start
= 0;
9043 p
->se
.block_start
= 0;
9048 * Renice negative nice level userspace
9051 if (TASK_NICE(p
) < 0 && p
->mm
)
9052 set_user_nice(p
, 0);
9056 spin_lock(&p
->pi_lock
);
9057 rq
= __task_rq_lock(p
);
9059 normalize_task(rq
, p
);
9061 __task_rq_unlock(rq
);
9062 spin_unlock(&p
->pi_lock
);
9063 } while_each_thread(g
, p
);
9065 read_unlock_irqrestore(&tasklist_lock
, flags
);
9068 #endif /* CONFIG_MAGIC_SYSRQ */
9072 * These functions are only useful for the IA64 MCA handling.
9074 * They can only be called when the whole system has been
9075 * stopped - every CPU needs to be quiescent, and no scheduling
9076 * activity can take place. Using them for anything else would
9077 * be a serious bug, and as a result, they aren't even visible
9078 * under any other configuration.
9082 * curr_task - return the current task for a given cpu.
9083 * @cpu: the processor in question.
9085 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9087 struct task_struct
*curr_task(int cpu
)
9089 return cpu_curr(cpu
);
9093 * set_curr_task - set the current task for a given cpu.
9094 * @cpu: the processor in question.
9095 * @p: the task pointer to set.
9097 * Description: This function must only be used when non-maskable interrupts
9098 * are serviced on a separate stack. It allows the architecture to switch the
9099 * notion of the current task on a cpu in a non-blocking manner. This function
9100 * must be called with all CPU's synchronized, and interrupts disabled, the
9101 * and caller must save the original value of the current task (see
9102 * curr_task() above) and restore that value before reenabling interrupts and
9103 * re-starting the system.
9105 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9107 void set_curr_task(int cpu
, struct task_struct
*p
)
9114 #ifdef CONFIG_FAIR_GROUP_SCHED
9115 static void free_fair_sched_group(struct task_group
*tg
)
9119 for_each_possible_cpu(i
) {
9121 kfree(tg
->cfs_rq
[i
]);
9131 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9133 struct cfs_rq
*cfs_rq
;
9134 struct sched_entity
*se
;
9138 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9141 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
9145 tg
->shares
= NICE_0_LOAD
;
9147 for_each_possible_cpu(i
) {
9150 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
9151 GFP_KERNEL
, cpu_to_node(i
));
9155 se
= kzalloc_node(sizeof(struct sched_entity
),
9156 GFP_KERNEL
, cpu_to_node(i
));
9160 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
9169 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9171 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
9172 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
9175 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9177 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
9179 #else /* !CONFG_FAIR_GROUP_SCHED */
9180 static inline void free_fair_sched_group(struct task_group
*tg
)
9185 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9190 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9194 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9197 #endif /* CONFIG_FAIR_GROUP_SCHED */
9199 #ifdef CONFIG_RT_GROUP_SCHED
9200 static void free_rt_sched_group(struct task_group
*tg
)
9204 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
9206 for_each_possible_cpu(i
) {
9208 kfree(tg
->rt_rq
[i
]);
9210 kfree(tg
->rt_se
[i
]);
9218 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9220 struct rt_rq
*rt_rq
;
9221 struct sched_rt_entity
*rt_se
;
9225 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9228 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
9232 init_rt_bandwidth(&tg
->rt_bandwidth
,
9233 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
9235 for_each_possible_cpu(i
) {
9238 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
9239 GFP_KERNEL
, cpu_to_node(i
));
9243 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
9244 GFP_KERNEL
, cpu_to_node(i
));
9248 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
9257 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9259 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
9260 &cpu_rq(cpu
)->leaf_rt_rq_list
);
9263 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9265 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
9267 #else /* !CONFIG_RT_GROUP_SCHED */
9268 static inline void free_rt_sched_group(struct task_group
*tg
)
9273 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9278 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9282 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9285 #endif /* CONFIG_RT_GROUP_SCHED */
9287 #ifdef CONFIG_GROUP_SCHED
9288 static void free_sched_group(struct task_group
*tg
)
9290 free_fair_sched_group(tg
);
9291 free_rt_sched_group(tg
);
9295 /* allocate runqueue etc for a new task group */
9296 struct task_group
*sched_create_group(struct task_group
*parent
)
9298 struct task_group
*tg
;
9299 unsigned long flags
;
9302 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
9304 return ERR_PTR(-ENOMEM
);
9306 if (!alloc_fair_sched_group(tg
, parent
))
9309 if (!alloc_rt_sched_group(tg
, parent
))
9312 spin_lock_irqsave(&task_group_lock
, flags
);
9313 for_each_possible_cpu(i
) {
9314 register_fair_sched_group(tg
, i
);
9315 register_rt_sched_group(tg
, i
);
9317 list_add_rcu(&tg
->list
, &task_groups
);
9319 WARN_ON(!parent
); /* root should already exist */
9321 tg
->parent
= parent
;
9322 INIT_LIST_HEAD(&tg
->children
);
9323 list_add_rcu(&tg
->siblings
, &parent
->children
);
9324 spin_unlock_irqrestore(&task_group_lock
, flags
);
9329 free_sched_group(tg
);
9330 return ERR_PTR(-ENOMEM
);
9333 /* rcu callback to free various structures associated with a task group */
9334 static void free_sched_group_rcu(struct rcu_head
*rhp
)
9336 /* now it should be safe to free those cfs_rqs */
9337 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
9340 /* Destroy runqueue etc associated with a task group */
9341 void sched_destroy_group(struct task_group
*tg
)
9343 unsigned long flags
;
9346 spin_lock_irqsave(&task_group_lock
, flags
);
9347 for_each_possible_cpu(i
) {
9348 unregister_fair_sched_group(tg
, i
);
9349 unregister_rt_sched_group(tg
, i
);
9351 list_del_rcu(&tg
->list
);
9352 list_del_rcu(&tg
->siblings
);
9353 spin_unlock_irqrestore(&task_group_lock
, flags
);
9355 /* wait for possible concurrent references to cfs_rqs complete */
9356 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
9359 /* change task's runqueue when it moves between groups.
9360 * The caller of this function should have put the task in its new group
9361 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9362 * reflect its new group.
9364 void sched_move_task(struct task_struct
*tsk
)
9367 unsigned long flags
;
9370 rq
= task_rq_lock(tsk
, &flags
);
9372 update_rq_clock(rq
);
9374 running
= task_current(rq
, tsk
);
9375 on_rq
= tsk
->se
.on_rq
;
9378 dequeue_task(rq
, tsk
, 0);
9379 if (unlikely(running
))
9380 tsk
->sched_class
->put_prev_task(rq
, tsk
);
9382 set_task_rq(tsk
, task_cpu(tsk
));
9384 #ifdef CONFIG_FAIR_GROUP_SCHED
9385 if (tsk
->sched_class
->moved_group
)
9386 tsk
->sched_class
->moved_group(tsk
);
9389 if (unlikely(running
))
9390 tsk
->sched_class
->set_curr_task(rq
);
9392 enqueue_task(rq
, tsk
, 0);
9394 task_rq_unlock(rq
, &flags
);
9396 #endif /* CONFIG_GROUP_SCHED */
9398 #ifdef CONFIG_FAIR_GROUP_SCHED
9399 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
9401 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
9406 dequeue_entity(cfs_rq
, se
, 0);
9408 se
->load
.weight
= shares
;
9409 se
->load
.inv_weight
= 0;
9412 enqueue_entity(cfs_rq
, se
, 0);
9415 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
9417 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
9418 struct rq
*rq
= cfs_rq
->rq
;
9419 unsigned long flags
;
9421 spin_lock_irqsave(&rq
->lock
, flags
);
9422 __set_se_shares(se
, shares
);
9423 spin_unlock_irqrestore(&rq
->lock
, flags
);
9426 static DEFINE_MUTEX(shares_mutex
);
9428 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
9431 unsigned long flags
;
9434 * We can't change the weight of the root cgroup.
9439 if (shares
< MIN_SHARES
)
9440 shares
= MIN_SHARES
;
9441 else if (shares
> MAX_SHARES
)
9442 shares
= MAX_SHARES
;
9444 mutex_lock(&shares_mutex
);
9445 if (tg
->shares
== shares
)
9448 spin_lock_irqsave(&task_group_lock
, flags
);
9449 for_each_possible_cpu(i
)
9450 unregister_fair_sched_group(tg
, i
);
9451 list_del_rcu(&tg
->siblings
);
9452 spin_unlock_irqrestore(&task_group_lock
, flags
);
9454 /* wait for any ongoing reference to this group to finish */
9455 synchronize_sched();
9458 * Now we are free to modify the group's share on each cpu
9459 * w/o tripping rebalance_share or load_balance_fair.
9461 tg
->shares
= shares
;
9462 for_each_possible_cpu(i
) {
9466 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
9467 set_se_shares(tg
->se
[i
], shares
);
9471 * Enable load balance activity on this group, by inserting it back on
9472 * each cpu's rq->leaf_cfs_rq_list.
9474 spin_lock_irqsave(&task_group_lock
, flags
);
9475 for_each_possible_cpu(i
)
9476 register_fair_sched_group(tg
, i
);
9477 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
9478 spin_unlock_irqrestore(&task_group_lock
, flags
);
9480 mutex_unlock(&shares_mutex
);
9484 unsigned long sched_group_shares(struct task_group
*tg
)
9490 #ifdef CONFIG_RT_GROUP_SCHED
9492 * Ensure that the real time constraints are schedulable.
9494 static DEFINE_MUTEX(rt_constraints_mutex
);
9496 static unsigned long to_ratio(u64 period
, u64 runtime
)
9498 if (runtime
== RUNTIME_INF
)
9501 return div64_u64(runtime
<< 20, period
);
9504 /* Must be called with tasklist_lock held */
9505 static inline int tg_has_rt_tasks(struct task_group
*tg
)
9507 struct task_struct
*g
, *p
;
9509 do_each_thread(g
, p
) {
9510 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
9512 } while_each_thread(g
, p
);
9517 struct rt_schedulable_data
{
9518 struct task_group
*tg
;
9523 static int tg_schedulable(struct task_group
*tg
, void *data
)
9525 struct rt_schedulable_data
*d
= data
;
9526 struct task_group
*child
;
9527 unsigned long total
, sum
= 0;
9528 u64 period
, runtime
;
9530 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9531 runtime
= tg
->rt_bandwidth
.rt_runtime
;
9534 period
= d
->rt_period
;
9535 runtime
= d
->rt_runtime
;
9538 #ifdef CONFIG_USER_SCHED
9539 if (tg
== &root_task_group
) {
9540 period
= global_rt_period();
9541 runtime
= global_rt_runtime();
9546 * Cannot have more runtime than the period.
9548 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9552 * Ensure we don't starve existing RT tasks.
9554 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
9557 total
= to_ratio(period
, runtime
);
9560 * Nobody can have more than the global setting allows.
9562 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
9566 * The sum of our children's runtime should not exceed our own.
9568 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
9569 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
9570 runtime
= child
->rt_bandwidth
.rt_runtime
;
9572 if (child
== d
->tg
) {
9573 period
= d
->rt_period
;
9574 runtime
= d
->rt_runtime
;
9577 sum
+= to_ratio(period
, runtime
);
9586 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
9588 struct rt_schedulable_data data
= {
9590 .rt_period
= period
,
9591 .rt_runtime
= runtime
,
9594 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
9597 static int tg_set_bandwidth(struct task_group
*tg
,
9598 u64 rt_period
, u64 rt_runtime
)
9602 mutex_lock(&rt_constraints_mutex
);
9603 read_lock(&tasklist_lock
);
9604 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
9608 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
9609 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
9610 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
9612 for_each_possible_cpu(i
) {
9613 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
9615 spin_lock(&rt_rq
->rt_runtime_lock
);
9616 rt_rq
->rt_runtime
= rt_runtime
;
9617 spin_unlock(&rt_rq
->rt_runtime_lock
);
9619 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
9621 read_unlock(&tasklist_lock
);
9622 mutex_unlock(&rt_constraints_mutex
);
9627 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
9629 u64 rt_runtime
, rt_period
;
9631 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9632 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
9633 if (rt_runtime_us
< 0)
9634 rt_runtime
= RUNTIME_INF
;
9636 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9639 long sched_group_rt_runtime(struct task_group
*tg
)
9643 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
9646 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
9647 do_div(rt_runtime_us
, NSEC_PER_USEC
);
9648 return rt_runtime_us
;
9651 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
9653 u64 rt_runtime
, rt_period
;
9655 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
9656 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9661 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9664 long sched_group_rt_period(struct task_group
*tg
)
9668 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9669 do_div(rt_period_us
, NSEC_PER_USEC
);
9670 return rt_period_us
;
9673 static int sched_rt_global_constraints(void)
9675 u64 runtime
, period
;
9678 if (sysctl_sched_rt_period
<= 0)
9681 runtime
= global_rt_runtime();
9682 period
= global_rt_period();
9685 * Sanity check on the sysctl variables.
9687 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9690 mutex_lock(&rt_constraints_mutex
);
9691 read_lock(&tasklist_lock
);
9692 ret
= __rt_schedulable(NULL
, 0, 0);
9693 read_unlock(&tasklist_lock
);
9694 mutex_unlock(&rt_constraints_mutex
);
9699 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
9701 /* Don't accept realtime tasks when there is no way for them to run */
9702 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
9708 #else /* !CONFIG_RT_GROUP_SCHED */
9709 static int sched_rt_global_constraints(void)
9711 unsigned long flags
;
9714 if (sysctl_sched_rt_period
<= 0)
9717 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9718 for_each_possible_cpu(i
) {
9719 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
9721 spin_lock(&rt_rq
->rt_runtime_lock
);
9722 rt_rq
->rt_runtime
= global_rt_runtime();
9723 spin_unlock(&rt_rq
->rt_runtime_lock
);
9725 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9729 #endif /* CONFIG_RT_GROUP_SCHED */
9731 int sched_rt_handler(struct ctl_table
*table
, int write
,
9732 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
9736 int old_period
, old_runtime
;
9737 static DEFINE_MUTEX(mutex
);
9740 old_period
= sysctl_sched_rt_period
;
9741 old_runtime
= sysctl_sched_rt_runtime
;
9743 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
9745 if (!ret
&& write
) {
9746 ret
= sched_rt_global_constraints();
9748 sysctl_sched_rt_period
= old_period
;
9749 sysctl_sched_rt_runtime
= old_runtime
;
9751 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
9752 def_rt_bandwidth
.rt_period
=
9753 ns_to_ktime(global_rt_period());
9756 mutex_unlock(&mutex
);
9761 #ifdef CONFIG_CGROUP_SCHED
9763 /* return corresponding task_group object of a cgroup */
9764 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
9766 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
9767 struct task_group
, css
);
9770 static struct cgroup_subsys_state
*
9771 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9773 struct task_group
*tg
, *parent
;
9775 if (!cgrp
->parent
) {
9776 /* This is early initialization for the top cgroup */
9777 return &init_task_group
.css
;
9780 parent
= cgroup_tg(cgrp
->parent
);
9781 tg
= sched_create_group(parent
);
9783 return ERR_PTR(-ENOMEM
);
9789 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9791 struct task_group
*tg
= cgroup_tg(cgrp
);
9793 sched_destroy_group(tg
);
9797 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9798 struct task_struct
*tsk
)
9800 #ifdef CONFIG_RT_GROUP_SCHED
9801 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
9804 /* We don't support RT-tasks being in separate groups */
9805 if (tsk
->sched_class
!= &fair_sched_class
)
9813 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9814 struct cgroup
*old_cont
, struct task_struct
*tsk
)
9816 sched_move_task(tsk
);
9819 #ifdef CONFIG_FAIR_GROUP_SCHED
9820 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9823 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
9826 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9828 struct task_group
*tg
= cgroup_tg(cgrp
);
9830 return (u64
) tg
->shares
;
9832 #endif /* CONFIG_FAIR_GROUP_SCHED */
9834 #ifdef CONFIG_RT_GROUP_SCHED
9835 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9838 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9841 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9843 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9846 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9849 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9852 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9854 return sched_group_rt_period(cgroup_tg(cgrp
));
9856 #endif /* CONFIG_RT_GROUP_SCHED */
9858 static struct cftype cpu_files
[] = {
9859 #ifdef CONFIG_FAIR_GROUP_SCHED
9862 .read_u64
= cpu_shares_read_u64
,
9863 .write_u64
= cpu_shares_write_u64
,
9866 #ifdef CONFIG_RT_GROUP_SCHED
9868 .name
= "rt_runtime_us",
9869 .read_s64
= cpu_rt_runtime_read
,
9870 .write_s64
= cpu_rt_runtime_write
,
9873 .name
= "rt_period_us",
9874 .read_u64
= cpu_rt_period_read_uint
,
9875 .write_u64
= cpu_rt_period_write_uint
,
9880 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
9882 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
9885 struct cgroup_subsys cpu_cgroup_subsys
= {
9887 .create
= cpu_cgroup_create
,
9888 .destroy
= cpu_cgroup_destroy
,
9889 .can_attach
= cpu_cgroup_can_attach
,
9890 .attach
= cpu_cgroup_attach
,
9891 .populate
= cpu_cgroup_populate
,
9892 .subsys_id
= cpu_cgroup_subsys_id
,
9896 #endif /* CONFIG_CGROUP_SCHED */
9898 #ifdef CONFIG_CGROUP_CPUACCT
9901 * CPU accounting code for task groups.
9903 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9904 * (balbir@in.ibm.com).
9907 /* track cpu usage of a group of tasks and its child groups */
9909 struct cgroup_subsys_state css
;
9910 /* cpuusage holds pointer to a u64-type object on every cpu */
9912 struct cpuacct
*parent
;
9915 struct cgroup_subsys cpuacct_subsys
;
9917 /* return cpu accounting group corresponding to this container */
9918 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
9920 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
9921 struct cpuacct
, css
);
9924 /* return cpu accounting group to which this task belongs */
9925 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
9927 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
9928 struct cpuacct
, css
);
9931 /* create a new cpu accounting group */
9932 static struct cgroup_subsys_state
*cpuacct_create(
9933 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9935 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
9938 return ERR_PTR(-ENOMEM
);
9940 ca
->cpuusage
= alloc_percpu(u64
);
9941 if (!ca
->cpuusage
) {
9943 return ERR_PTR(-ENOMEM
);
9947 ca
->parent
= cgroup_ca(cgrp
->parent
);
9952 /* destroy an existing cpu accounting group */
9954 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9956 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9958 free_percpu(ca
->cpuusage
);
9962 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
9964 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9967 #ifndef CONFIG_64BIT
9969 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9971 spin_lock_irq(&cpu_rq(cpu
)->lock
);
9973 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9981 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
9983 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9985 #ifndef CONFIG_64BIT
9987 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9989 spin_lock_irq(&cpu_rq(cpu
)->lock
);
9991 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9997 /* return total cpu usage (in nanoseconds) of a group */
9998 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10000 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10001 u64 totalcpuusage
= 0;
10004 for_each_present_cpu(i
)
10005 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
10007 return totalcpuusage
;
10010 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
10013 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10022 for_each_present_cpu(i
)
10023 cpuacct_cpuusage_write(ca
, i
, 0);
10029 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
10030 struct seq_file
*m
)
10032 struct cpuacct
*ca
= cgroup_ca(cgroup
);
10036 for_each_present_cpu(i
) {
10037 percpu
= cpuacct_cpuusage_read(ca
, i
);
10038 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
10040 seq_printf(m
, "\n");
10044 static struct cftype files
[] = {
10047 .read_u64
= cpuusage_read
,
10048 .write_u64
= cpuusage_write
,
10051 .name
= "usage_percpu",
10052 .read_seq_string
= cpuacct_percpu_seq_read
,
10057 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10059 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
10063 * charge this task's execution time to its accounting group.
10065 * called with rq->lock held.
10067 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
10069 struct cpuacct
*ca
;
10072 if (unlikely(!cpuacct_subsys
.active
))
10075 cpu
= task_cpu(tsk
);
10078 for (; ca
; ca
= ca
->parent
) {
10079 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10080 *cpuusage
+= cputime
;
10084 struct cgroup_subsys cpuacct_subsys
= {
10086 .create
= cpuacct_create
,
10087 .destroy
= cpuacct_destroy
,
10088 .populate
= cpuacct_populate
,
10089 .subsys_id
= cpuacct_subsys_id
,
10091 #endif /* CONFIG_CGROUP_CPUACCT */