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 /* calc_load related fields */
634 unsigned long calc_load_update
;
635 long calc_load_active
;
637 #ifdef CONFIG_SCHED_HRTICK
639 int hrtick_csd_pending
;
640 struct call_single_data hrtick_csd
;
642 struct hrtimer hrtick_timer
;
645 #ifdef CONFIG_SCHEDSTATS
647 struct sched_info rq_sched_info
;
648 unsigned long long rq_cpu_time
;
649 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
651 /* sys_sched_yield() stats */
652 unsigned int yld_count
;
654 /* schedule() stats */
655 unsigned int sched_switch
;
656 unsigned int sched_count
;
657 unsigned int sched_goidle
;
659 /* try_to_wake_up() stats */
660 unsigned int ttwu_count
;
661 unsigned int ttwu_local
;
664 unsigned int bkl_count
;
668 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
670 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int sync
)
672 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, sync
);
675 static inline int cpu_of(struct rq
*rq
)
685 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
686 * See detach_destroy_domains: synchronize_sched for details.
688 * The domain tree of any CPU may only be accessed from within
689 * preempt-disabled sections.
691 #define for_each_domain(cpu, __sd) \
692 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
694 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
695 #define this_rq() (&__get_cpu_var(runqueues))
696 #define task_rq(p) cpu_rq(task_cpu(p))
697 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
699 static inline void update_rq_clock(struct rq
*rq
)
701 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
705 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
707 #ifdef CONFIG_SCHED_DEBUG
708 # define const_debug __read_mostly
710 # define const_debug static const
716 * Returns true if the current cpu runqueue is locked.
717 * This interface allows printk to be called with the runqueue lock
718 * held and know whether or not it is OK to wake up the klogd.
720 int runqueue_is_locked(void)
723 struct rq
*rq
= cpu_rq(cpu
);
726 ret
= spin_is_locked(&rq
->lock
);
732 * Debugging: various feature bits
735 #define SCHED_FEAT(name, enabled) \
736 __SCHED_FEAT_##name ,
739 #include "sched_features.h"
744 #define SCHED_FEAT(name, enabled) \
745 (1UL << __SCHED_FEAT_##name) * enabled |
747 const_debug
unsigned int sysctl_sched_features
=
748 #include "sched_features.h"
753 #ifdef CONFIG_SCHED_DEBUG
754 #define SCHED_FEAT(name, enabled) \
757 static __read_mostly
char *sched_feat_names
[] = {
758 #include "sched_features.h"
764 static int sched_feat_show(struct seq_file
*m
, void *v
)
768 for (i
= 0; sched_feat_names
[i
]; i
++) {
769 if (!(sysctl_sched_features
& (1UL << i
)))
771 seq_printf(m
, "%s ", sched_feat_names
[i
]);
779 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
780 size_t cnt
, loff_t
*ppos
)
790 if (copy_from_user(&buf
, ubuf
, cnt
))
795 if (strncmp(buf
, "NO_", 3) == 0) {
800 for (i
= 0; sched_feat_names
[i
]; i
++) {
801 int len
= strlen(sched_feat_names
[i
]);
803 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
805 sysctl_sched_features
&= ~(1UL << i
);
807 sysctl_sched_features
|= (1UL << i
);
812 if (!sched_feat_names
[i
])
820 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
822 return single_open(filp
, sched_feat_show
, NULL
);
825 static struct file_operations sched_feat_fops
= {
826 .open
= sched_feat_open
,
827 .write
= sched_feat_write
,
830 .release
= single_release
,
833 static __init
int sched_init_debug(void)
835 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
840 late_initcall(sched_init_debug
);
844 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
847 * Number of tasks to iterate in a single balance run.
848 * Limited because this is done with IRQs disabled.
850 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
853 * ratelimit for updating the group shares.
856 unsigned int sysctl_sched_shares_ratelimit
= 250000;
859 * Inject some fuzzyness into changing the per-cpu group shares
860 * this avoids remote rq-locks at the expense of fairness.
863 unsigned int sysctl_sched_shares_thresh
= 4;
866 * period over which we measure -rt task cpu usage in us.
869 unsigned int sysctl_sched_rt_period
= 1000000;
871 static __read_mostly
int scheduler_running
;
874 * part of the period that we allow rt tasks to run in us.
877 int sysctl_sched_rt_runtime
= 950000;
879 static inline u64
global_rt_period(void)
881 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
884 static inline u64
global_rt_runtime(void)
886 if (sysctl_sched_rt_runtime
< 0)
889 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
892 #ifndef prepare_arch_switch
893 # define prepare_arch_switch(next) do { } while (0)
895 #ifndef finish_arch_switch
896 # define finish_arch_switch(prev) do { } while (0)
899 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
901 return rq
->curr
== p
;
904 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
905 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
907 return task_current(rq
, p
);
910 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
914 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
916 #ifdef CONFIG_DEBUG_SPINLOCK
917 /* this is a valid case when another task releases the spinlock */
918 rq
->lock
.owner
= current
;
921 * If we are tracking spinlock dependencies then we have to
922 * fix up the runqueue lock - which gets 'carried over' from
925 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
927 spin_unlock_irq(&rq
->lock
);
930 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
931 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
936 return task_current(rq
, p
);
940 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
944 * We can optimise this out completely for !SMP, because the
945 * SMP rebalancing from interrupt is the only thing that cares
950 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
951 spin_unlock_irq(&rq
->lock
);
953 spin_unlock(&rq
->lock
);
957 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
961 * After ->oncpu is cleared, the task can be moved to a different CPU.
962 * We must ensure this doesn't happen until the switch is completely
968 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
972 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
975 * __task_rq_lock - lock the runqueue a given task resides on.
976 * Must be called interrupts disabled.
978 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
982 struct rq
*rq
= task_rq(p
);
983 spin_lock(&rq
->lock
);
984 if (likely(rq
== task_rq(p
)))
986 spin_unlock(&rq
->lock
);
991 * task_rq_lock - lock the runqueue a given task resides on and disable
992 * interrupts. Note the ordering: we can safely lookup the task_rq without
993 * explicitly disabling preemption.
995 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
1001 local_irq_save(*flags
);
1003 spin_lock(&rq
->lock
);
1004 if (likely(rq
== task_rq(p
)))
1006 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1010 void task_rq_unlock_wait(struct task_struct
*p
)
1012 struct rq
*rq
= task_rq(p
);
1014 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1015 spin_unlock_wait(&rq
->lock
);
1018 static void __task_rq_unlock(struct rq
*rq
)
1019 __releases(rq
->lock
)
1021 spin_unlock(&rq
->lock
);
1024 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
1025 __releases(rq
->lock
)
1027 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1031 * this_rq_lock - lock this runqueue and disable interrupts.
1033 static struct rq
*this_rq_lock(void)
1034 __acquires(rq
->lock
)
1038 local_irq_disable();
1040 spin_lock(&rq
->lock
);
1045 #ifdef CONFIG_SCHED_HRTICK
1047 * Use HR-timers to deliver accurate preemption points.
1049 * Its all a bit involved since we cannot program an hrt while holding the
1050 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1053 * When we get rescheduled we reprogram the hrtick_timer outside of the
1059 * - enabled by features
1060 * - hrtimer is actually high res
1062 static inline int hrtick_enabled(struct rq
*rq
)
1064 if (!sched_feat(HRTICK
))
1066 if (!cpu_active(cpu_of(rq
)))
1068 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1071 static void hrtick_clear(struct rq
*rq
)
1073 if (hrtimer_active(&rq
->hrtick_timer
))
1074 hrtimer_cancel(&rq
->hrtick_timer
);
1078 * High-resolution timer tick.
1079 * Runs from hardirq context with interrupts disabled.
1081 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1083 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1085 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1087 spin_lock(&rq
->lock
);
1088 update_rq_clock(rq
);
1089 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1090 spin_unlock(&rq
->lock
);
1092 return HRTIMER_NORESTART
;
1097 * called from hardirq (IPI) context
1099 static void __hrtick_start(void *arg
)
1101 struct rq
*rq
= arg
;
1103 spin_lock(&rq
->lock
);
1104 hrtimer_restart(&rq
->hrtick_timer
);
1105 rq
->hrtick_csd_pending
= 0;
1106 spin_unlock(&rq
->lock
);
1110 * Called to set the hrtick timer state.
1112 * called with rq->lock held and irqs disabled
1114 static void hrtick_start(struct rq
*rq
, u64 delay
)
1116 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1117 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1119 hrtimer_set_expires(timer
, time
);
1121 if (rq
== this_rq()) {
1122 hrtimer_restart(timer
);
1123 } else if (!rq
->hrtick_csd_pending
) {
1124 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1125 rq
->hrtick_csd_pending
= 1;
1130 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1132 int cpu
= (int)(long)hcpu
;
1135 case CPU_UP_CANCELED
:
1136 case CPU_UP_CANCELED_FROZEN
:
1137 case CPU_DOWN_PREPARE
:
1138 case CPU_DOWN_PREPARE_FROZEN
:
1140 case CPU_DEAD_FROZEN
:
1141 hrtick_clear(cpu_rq(cpu
));
1148 static __init
void init_hrtick(void)
1150 hotcpu_notifier(hotplug_hrtick
, 0);
1154 * Called to set the hrtick timer state.
1156 * called with rq->lock held and irqs disabled
1158 static void hrtick_start(struct rq
*rq
, u64 delay
)
1160 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1161 HRTIMER_MODE_REL
, 0);
1164 static inline void init_hrtick(void)
1167 #endif /* CONFIG_SMP */
1169 static void init_rq_hrtick(struct rq
*rq
)
1172 rq
->hrtick_csd_pending
= 0;
1174 rq
->hrtick_csd
.flags
= 0;
1175 rq
->hrtick_csd
.func
= __hrtick_start
;
1176 rq
->hrtick_csd
.info
= rq
;
1179 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1180 rq
->hrtick_timer
.function
= hrtick
;
1182 #else /* CONFIG_SCHED_HRTICK */
1183 static inline void hrtick_clear(struct rq
*rq
)
1187 static inline void init_rq_hrtick(struct rq
*rq
)
1191 static inline void init_hrtick(void)
1194 #endif /* CONFIG_SCHED_HRTICK */
1197 * resched_task - mark a task 'to be rescheduled now'.
1199 * On UP this means the setting of the need_resched flag, on SMP it
1200 * might also involve a cross-CPU call to trigger the scheduler on
1205 #ifndef tsk_is_polling
1206 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1209 static void resched_task(struct task_struct
*p
)
1213 assert_spin_locked(&task_rq(p
)->lock
);
1215 if (test_tsk_need_resched(p
))
1218 set_tsk_need_resched(p
);
1221 if (cpu
== smp_processor_id())
1224 /* NEED_RESCHED must be visible before we test polling */
1226 if (!tsk_is_polling(p
))
1227 smp_send_reschedule(cpu
);
1230 static void resched_cpu(int cpu
)
1232 struct rq
*rq
= cpu_rq(cpu
);
1233 unsigned long flags
;
1235 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1237 resched_task(cpu_curr(cpu
));
1238 spin_unlock_irqrestore(&rq
->lock
, flags
);
1243 * When add_timer_on() enqueues a timer into the timer wheel of an
1244 * idle CPU then this timer might expire before the next timer event
1245 * which is scheduled to wake up that CPU. In case of a completely
1246 * idle system the next event might even be infinite time into the
1247 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1248 * leaves the inner idle loop so the newly added timer is taken into
1249 * account when the CPU goes back to idle and evaluates the timer
1250 * wheel for the next timer event.
1252 void wake_up_idle_cpu(int cpu
)
1254 struct rq
*rq
= cpu_rq(cpu
);
1256 if (cpu
== smp_processor_id())
1260 * This is safe, as this function is called with the timer
1261 * wheel base lock of (cpu) held. When the CPU is on the way
1262 * to idle and has not yet set rq->curr to idle then it will
1263 * be serialized on the timer wheel base lock and take the new
1264 * timer into account automatically.
1266 if (rq
->curr
!= rq
->idle
)
1270 * We can set TIF_RESCHED on the idle task of the other CPU
1271 * lockless. The worst case is that the other CPU runs the
1272 * idle task through an additional NOOP schedule()
1274 set_tsk_need_resched(rq
->idle
);
1276 /* NEED_RESCHED must be visible before we test polling */
1278 if (!tsk_is_polling(rq
->idle
))
1279 smp_send_reschedule(cpu
);
1281 #endif /* CONFIG_NO_HZ */
1283 #else /* !CONFIG_SMP */
1284 static void resched_task(struct task_struct
*p
)
1286 assert_spin_locked(&task_rq(p
)->lock
);
1287 set_tsk_need_resched(p
);
1289 #endif /* CONFIG_SMP */
1291 #if BITS_PER_LONG == 32
1292 # define WMULT_CONST (~0UL)
1294 # define WMULT_CONST (1UL << 32)
1297 #define WMULT_SHIFT 32
1300 * Shift right and round:
1302 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1305 * delta *= weight / lw
1307 static unsigned long
1308 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1309 struct load_weight
*lw
)
1313 if (!lw
->inv_weight
) {
1314 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1317 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1321 tmp
= (u64
)delta_exec
* weight
;
1323 * Check whether we'd overflow the 64-bit multiplication:
1325 if (unlikely(tmp
> WMULT_CONST
))
1326 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1329 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1331 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1334 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1340 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1347 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1348 * of tasks with abnormal "nice" values across CPUs the contribution that
1349 * each task makes to its run queue's load is weighted according to its
1350 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1351 * scaled version of the new time slice allocation that they receive on time
1355 #define WEIGHT_IDLEPRIO 3
1356 #define WMULT_IDLEPRIO 1431655765
1359 * Nice levels are multiplicative, with a gentle 10% change for every
1360 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1361 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1362 * that remained on nice 0.
1364 * The "10% effect" is relative and cumulative: from _any_ nice level,
1365 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1366 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1367 * If a task goes up by ~10% and another task goes down by ~10% then
1368 * the relative distance between them is ~25%.)
1370 static const int prio_to_weight
[40] = {
1371 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1372 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1373 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1374 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1375 /* 0 */ 1024, 820, 655, 526, 423,
1376 /* 5 */ 335, 272, 215, 172, 137,
1377 /* 10 */ 110, 87, 70, 56, 45,
1378 /* 15 */ 36, 29, 23, 18, 15,
1382 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1384 * In cases where the weight does not change often, we can use the
1385 * precalculated inverse to speed up arithmetics by turning divisions
1386 * into multiplications:
1388 static const u32 prio_to_wmult
[40] = {
1389 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1390 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1391 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1392 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1393 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1394 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1395 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1396 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1399 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1402 * runqueue iterator, to support SMP load-balancing between different
1403 * scheduling classes, without having to expose their internal data
1404 * structures to the load-balancing proper:
1406 struct rq_iterator
{
1408 struct task_struct
*(*start
)(void *);
1409 struct task_struct
*(*next
)(void *);
1413 static unsigned long
1414 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1415 unsigned long max_load_move
, struct sched_domain
*sd
,
1416 enum cpu_idle_type idle
, int *all_pinned
,
1417 int *this_best_prio
, struct rq_iterator
*iterator
);
1420 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1421 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1422 struct rq_iterator
*iterator
);
1425 /* Time spent by the tasks of the cpu accounting group executing in ... */
1426 enum cpuacct_stat_index
{
1427 CPUACCT_STAT_USER
, /* ... user mode */
1428 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1430 CPUACCT_STAT_NSTATS
,
1433 #ifdef CONFIG_CGROUP_CPUACCT
1434 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1435 static void cpuacct_update_stats(struct task_struct
*tsk
,
1436 enum cpuacct_stat_index idx
, cputime_t val
);
1438 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1439 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1440 enum cpuacct_stat_index idx
, cputime_t val
) {}
1443 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1445 update_load_add(&rq
->load
, load
);
1448 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1450 update_load_sub(&rq
->load
, load
);
1453 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1454 typedef int (*tg_visitor
)(struct task_group
*, void *);
1457 * Iterate the full tree, calling @down when first entering a node and @up when
1458 * leaving it for the final time.
1460 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1462 struct task_group
*parent
, *child
;
1466 parent
= &root_task_group
;
1468 ret
= (*down
)(parent
, data
);
1471 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1478 ret
= (*up
)(parent
, data
);
1483 parent
= parent
->parent
;
1492 static int tg_nop(struct task_group
*tg
, void *data
)
1499 static unsigned long source_load(int cpu
, int type
);
1500 static unsigned long target_load(int cpu
, int type
);
1501 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1503 static unsigned long cpu_avg_load_per_task(int cpu
)
1505 struct rq
*rq
= cpu_rq(cpu
);
1506 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1509 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1511 rq
->avg_load_per_task
= 0;
1513 return rq
->avg_load_per_task
;
1516 #ifdef CONFIG_FAIR_GROUP_SCHED
1518 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1521 * Calculate and set the cpu's group shares.
1524 update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1525 unsigned long sd_shares
, unsigned long sd_rq_weight
)
1527 unsigned long shares
;
1528 unsigned long rq_weight
;
1533 rq_weight
= tg
->cfs_rq
[cpu
]->rq_weight
;
1536 * \Sum shares * rq_weight
1537 * shares = -----------------------
1541 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1542 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1544 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1545 sysctl_sched_shares_thresh
) {
1546 struct rq
*rq
= cpu_rq(cpu
);
1547 unsigned long flags
;
1549 spin_lock_irqsave(&rq
->lock
, flags
);
1550 tg
->cfs_rq
[cpu
]->shares
= shares
;
1552 __set_se_shares(tg
->se
[cpu
], shares
);
1553 spin_unlock_irqrestore(&rq
->lock
, flags
);
1558 * Re-compute the task group their per cpu shares over the given domain.
1559 * This needs to be done in a bottom-up fashion because the rq weight of a
1560 * parent group depends on the shares of its child groups.
1562 static int tg_shares_up(struct task_group
*tg
, void *data
)
1564 unsigned long weight
, rq_weight
= 0;
1565 unsigned long shares
= 0;
1566 struct sched_domain
*sd
= data
;
1569 for_each_cpu(i
, sched_domain_span(sd
)) {
1571 * If there are currently no tasks on the cpu pretend there
1572 * is one of average load so that when a new task gets to
1573 * run here it will not get delayed by group starvation.
1575 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1577 weight
= NICE_0_LOAD
;
1579 tg
->cfs_rq
[i
]->rq_weight
= weight
;
1580 rq_weight
+= weight
;
1581 shares
+= tg
->cfs_rq
[i
]->shares
;
1584 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1585 shares
= tg
->shares
;
1587 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1588 shares
= tg
->shares
;
1590 for_each_cpu(i
, sched_domain_span(sd
))
1591 update_group_shares_cpu(tg
, i
, shares
, rq_weight
);
1597 * Compute the cpu's hierarchical load factor for each task group.
1598 * This needs to be done in a top-down fashion because the load of a child
1599 * group is a fraction of its parents load.
1601 static int tg_load_down(struct task_group
*tg
, void *data
)
1604 long cpu
= (long)data
;
1607 load
= cpu_rq(cpu
)->load
.weight
;
1609 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1610 load
*= tg
->cfs_rq
[cpu
]->shares
;
1611 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1614 tg
->cfs_rq
[cpu
]->h_load
= load
;
1619 static void update_shares(struct sched_domain
*sd
)
1621 u64 now
= cpu_clock(raw_smp_processor_id());
1622 s64 elapsed
= now
- sd
->last_update
;
1624 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1625 sd
->last_update
= now
;
1626 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1630 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1632 spin_unlock(&rq
->lock
);
1634 spin_lock(&rq
->lock
);
1637 static void update_h_load(long cpu
)
1639 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1644 static inline void update_shares(struct sched_domain
*sd
)
1648 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1654 #ifdef CONFIG_PREEMPT
1657 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1658 * way at the expense of forcing extra atomic operations in all
1659 * invocations. This assures that the double_lock is acquired using the
1660 * same underlying policy as the spinlock_t on this architecture, which
1661 * reduces latency compared to the unfair variant below. However, it
1662 * also adds more overhead and therefore may reduce throughput.
1664 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1665 __releases(this_rq
->lock
)
1666 __acquires(busiest
->lock
)
1667 __acquires(this_rq
->lock
)
1669 spin_unlock(&this_rq
->lock
);
1670 double_rq_lock(this_rq
, busiest
);
1677 * Unfair double_lock_balance: Optimizes throughput at the expense of
1678 * latency by eliminating extra atomic operations when the locks are
1679 * already in proper order on entry. This favors lower cpu-ids and will
1680 * grant the double lock to lower cpus over higher ids under contention,
1681 * regardless of entry order into the function.
1683 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1684 __releases(this_rq
->lock
)
1685 __acquires(busiest
->lock
)
1686 __acquires(this_rq
->lock
)
1690 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1691 if (busiest
< this_rq
) {
1692 spin_unlock(&this_rq
->lock
);
1693 spin_lock(&busiest
->lock
);
1694 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
1697 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
1702 #endif /* CONFIG_PREEMPT */
1705 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1707 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1709 if (unlikely(!irqs_disabled())) {
1710 /* printk() doesn't work good under rq->lock */
1711 spin_unlock(&this_rq
->lock
);
1715 return _double_lock_balance(this_rq
, busiest
);
1718 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1719 __releases(busiest
->lock
)
1721 spin_unlock(&busiest
->lock
);
1722 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1726 #ifdef CONFIG_FAIR_GROUP_SCHED
1727 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1730 cfs_rq
->shares
= shares
;
1735 static void calc_load_account_active(struct rq
*this_rq
);
1737 #include "sched_stats.h"
1738 #include "sched_idletask.c"
1739 #include "sched_fair.c"
1740 #include "sched_rt.c"
1741 #ifdef CONFIG_SCHED_DEBUG
1742 # include "sched_debug.c"
1745 #define sched_class_highest (&rt_sched_class)
1746 #define for_each_class(class) \
1747 for (class = sched_class_highest; class; class = class->next)
1749 static void inc_nr_running(struct rq
*rq
)
1754 static void dec_nr_running(struct rq
*rq
)
1759 static void set_load_weight(struct task_struct
*p
)
1761 if (task_has_rt_policy(p
)) {
1762 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1763 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1768 * SCHED_IDLE tasks get minimal weight:
1770 if (p
->policy
== SCHED_IDLE
) {
1771 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1772 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1776 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1777 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1780 static void update_avg(u64
*avg
, u64 sample
)
1782 s64 diff
= sample
- *avg
;
1786 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1789 p
->se
.start_runtime
= p
->se
.sum_exec_runtime
;
1791 sched_info_queued(p
);
1792 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1796 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1799 if (p
->se
.last_wakeup
) {
1800 update_avg(&p
->se
.avg_overlap
,
1801 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1802 p
->se
.last_wakeup
= 0;
1804 update_avg(&p
->se
.avg_wakeup
,
1805 sysctl_sched_wakeup_granularity
);
1809 sched_info_dequeued(p
);
1810 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1815 * __normal_prio - return the priority that is based on the static prio
1817 static inline int __normal_prio(struct task_struct
*p
)
1819 return p
->static_prio
;
1823 * Calculate the expected normal priority: i.e. priority
1824 * without taking RT-inheritance into account. Might be
1825 * boosted by interactivity modifiers. Changes upon fork,
1826 * setprio syscalls, and whenever the interactivity
1827 * estimator recalculates.
1829 static inline int normal_prio(struct task_struct
*p
)
1833 if (task_has_rt_policy(p
))
1834 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1836 prio
= __normal_prio(p
);
1841 * Calculate the current priority, i.e. the priority
1842 * taken into account by the scheduler. This value might
1843 * be boosted by RT tasks, or might be boosted by
1844 * interactivity modifiers. Will be RT if the task got
1845 * RT-boosted. If not then it returns p->normal_prio.
1847 static int effective_prio(struct task_struct
*p
)
1849 p
->normal_prio
= normal_prio(p
);
1851 * If we are RT tasks or we were boosted to RT priority,
1852 * keep the priority unchanged. Otherwise, update priority
1853 * to the normal priority:
1855 if (!rt_prio(p
->prio
))
1856 return p
->normal_prio
;
1861 * activate_task - move a task to the runqueue.
1863 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1865 if (task_contributes_to_load(p
))
1866 rq
->nr_uninterruptible
--;
1868 enqueue_task(rq
, p
, wakeup
);
1873 * deactivate_task - remove a task from the runqueue.
1875 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1877 if (task_contributes_to_load(p
))
1878 rq
->nr_uninterruptible
++;
1880 dequeue_task(rq
, p
, sleep
);
1885 * task_curr - is this task currently executing on a CPU?
1886 * @p: the task in question.
1888 inline int task_curr(const struct task_struct
*p
)
1890 return cpu_curr(task_cpu(p
)) == p
;
1893 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1895 set_task_rq(p
, cpu
);
1898 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1899 * successfuly executed on another CPU. We must ensure that updates of
1900 * per-task data have been completed by this moment.
1903 task_thread_info(p
)->cpu
= cpu
;
1907 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1908 const struct sched_class
*prev_class
,
1909 int oldprio
, int running
)
1911 if (prev_class
!= p
->sched_class
) {
1912 if (prev_class
->switched_from
)
1913 prev_class
->switched_from(rq
, p
, running
);
1914 p
->sched_class
->switched_to(rq
, p
, running
);
1916 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1921 /* Used instead of source_load when we know the type == 0 */
1922 static unsigned long weighted_cpuload(const int cpu
)
1924 return cpu_rq(cpu
)->load
.weight
;
1928 * Is this task likely cache-hot:
1931 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1936 * Buddy candidates are cache hot:
1938 if (sched_feat(CACHE_HOT_BUDDY
) &&
1939 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
1940 &p
->se
== cfs_rq_of(&p
->se
)->last
))
1943 if (p
->sched_class
!= &fair_sched_class
)
1946 if (sysctl_sched_migration_cost
== -1)
1948 if (sysctl_sched_migration_cost
== 0)
1951 delta
= now
- p
->se
.exec_start
;
1953 return delta
< (s64
)sysctl_sched_migration_cost
;
1957 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1959 int old_cpu
= task_cpu(p
);
1960 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1961 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1962 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1965 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1967 trace_sched_migrate_task(p
, task_cpu(p
), new_cpu
);
1969 #ifdef CONFIG_SCHEDSTATS
1970 if (p
->se
.wait_start
)
1971 p
->se
.wait_start
-= clock_offset
;
1972 if (p
->se
.sleep_start
)
1973 p
->se
.sleep_start
-= clock_offset
;
1974 if (p
->se
.block_start
)
1975 p
->se
.block_start
-= clock_offset
;
1976 if (old_cpu
!= new_cpu
) {
1977 schedstat_inc(p
, se
.nr_migrations
);
1978 if (task_hot(p
, old_rq
->clock
, NULL
))
1979 schedstat_inc(p
, se
.nr_forced2_migrations
);
1982 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1983 new_cfsrq
->min_vruntime
;
1985 __set_task_cpu(p
, new_cpu
);
1988 struct migration_req
{
1989 struct list_head list
;
1991 struct task_struct
*task
;
1994 struct completion done
;
1998 * The task's runqueue lock must be held.
1999 * Returns true if you have to wait for migration thread.
2002 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
2004 struct rq
*rq
= task_rq(p
);
2007 * If the task is not on a runqueue (and not running), then
2008 * it is sufficient to simply update the task's cpu field.
2010 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
2011 set_task_cpu(p
, dest_cpu
);
2015 init_completion(&req
->done
);
2017 req
->dest_cpu
= dest_cpu
;
2018 list_add(&req
->list
, &rq
->migration_queue
);
2024 * wait_task_inactive - wait for a thread to unschedule.
2026 * If @match_state is nonzero, it's the @p->state value just checked and
2027 * not expected to change. If it changes, i.e. @p might have woken up,
2028 * then return zero. When we succeed in waiting for @p to be off its CPU,
2029 * we return a positive number (its total switch count). If a second call
2030 * a short while later returns the same number, the caller can be sure that
2031 * @p has remained unscheduled the whole time.
2033 * The caller must ensure that the task *will* unschedule sometime soon,
2034 * else this function might spin for a *long* time. This function can't
2035 * be called with interrupts off, or it may introduce deadlock with
2036 * smp_call_function() if an IPI is sent by the same process we are
2037 * waiting to become inactive.
2039 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2041 unsigned long flags
;
2048 * We do the initial early heuristics without holding
2049 * any task-queue locks at all. We'll only try to get
2050 * the runqueue lock when things look like they will
2056 * If the task is actively running on another CPU
2057 * still, just relax and busy-wait without holding
2060 * NOTE! Since we don't hold any locks, it's not
2061 * even sure that "rq" stays as the right runqueue!
2062 * But we don't care, since "task_running()" will
2063 * return false if the runqueue has changed and p
2064 * is actually now running somewhere else!
2066 while (task_running(rq
, p
)) {
2067 if (match_state
&& unlikely(p
->state
!= match_state
))
2073 * Ok, time to look more closely! We need the rq
2074 * lock now, to be *sure*. If we're wrong, we'll
2075 * just go back and repeat.
2077 rq
= task_rq_lock(p
, &flags
);
2078 trace_sched_wait_task(rq
, p
);
2079 running
= task_running(rq
, p
);
2080 on_rq
= p
->se
.on_rq
;
2082 if (!match_state
|| p
->state
== match_state
)
2083 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2084 task_rq_unlock(rq
, &flags
);
2087 * If it changed from the expected state, bail out now.
2089 if (unlikely(!ncsw
))
2093 * Was it really running after all now that we
2094 * checked with the proper locks actually held?
2096 * Oops. Go back and try again..
2098 if (unlikely(running
)) {
2104 * It's not enough that it's not actively running,
2105 * it must be off the runqueue _entirely_, and not
2108 * So if it was still runnable (but just not actively
2109 * running right now), it's preempted, and we should
2110 * yield - it could be a while.
2112 if (unlikely(on_rq
)) {
2113 schedule_timeout_uninterruptible(1);
2118 * Ahh, all good. It wasn't running, and it wasn't
2119 * runnable, which means that it will never become
2120 * running in the future either. We're all done!
2129 * kick_process - kick a running thread to enter/exit the kernel
2130 * @p: the to-be-kicked thread
2132 * Cause a process which is running on another CPU to enter
2133 * kernel-mode, without any delay. (to get signals handled.)
2135 * NOTE: this function doesnt have to take the runqueue lock,
2136 * because all it wants to ensure is that the remote task enters
2137 * the kernel. If the IPI races and the task has been migrated
2138 * to another CPU then no harm is done and the purpose has been
2141 void kick_process(struct task_struct
*p
)
2147 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2148 smp_send_reschedule(cpu
);
2153 * Return a low guess at the load of a migration-source cpu weighted
2154 * according to the scheduling class and "nice" value.
2156 * We want to under-estimate the load of migration sources, to
2157 * balance conservatively.
2159 static unsigned long source_load(int cpu
, int type
)
2161 struct rq
*rq
= cpu_rq(cpu
);
2162 unsigned long total
= weighted_cpuload(cpu
);
2164 if (type
== 0 || !sched_feat(LB_BIAS
))
2167 return min(rq
->cpu_load
[type
-1], total
);
2171 * Return a high guess at the load of a migration-target cpu weighted
2172 * according to the scheduling class and "nice" value.
2174 static unsigned long target_load(int cpu
, int type
)
2176 struct rq
*rq
= cpu_rq(cpu
);
2177 unsigned long total
= weighted_cpuload(cpu
);
2179 if (type
== 0 || !sched_feat(LB_BIAS
))
2182 return max(rq
->cpu_load
[type
-1], total
);
2186 * find_idlest_group finds and returns the least busy CPU group within the
2189 static struct sched_group
*
2190 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2192 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2193 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2194 int load_idx
= sd
->forkexec_idx
;
2195 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2198 unsigned long load
, avg_load
;
2202 /* Skip over this group if it has no CPUs allowed */
2203 if (!cpumask_intersects(sched_group_cpus(group
),
2207 local_group
= cpumask_test_cpu(this_cpu
,
2208 sched_group_cpus(group
));
2210 /* Tally up the load of all CPUs in the group */
2213 for_each_cpu(i
, sched_group_cpus(group
)) {
2214 /* Bias balancing toward cpus of our domain */
2216 load
= source_load(i
, load_idx
);
2218 load
= target_load(i
, load_idx
);
2223 /* Adjust by relative CPU power of the group */
2224 avg_load
= sg_div_cpu_power(group
,
2225 avg_load
* SCHED_LOAD_SCALE
);
2228 this_load
= avg_load
;
2230 } else if (avg_load
< min_load
) {
2231 min_load
= avg_load
;
2234 } while (group
= group
->next
, group
!= sd
->groups
);
2236 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2242 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2245 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
2247 unsigned long load
, min_load
= ULONG_MAX
;
2251 /* Traverse only the allowed CPUs */
2252 for_each_cpu_and(i
, sched_group_cpus(group
), &p
->cpus_allowed
) {
2253 load
= weighted_cpuload(i
);
2255 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2265 * sched_balance_self: balance the current task (running on cpu) in domains
2266 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2269 * Balance, ie. select the least loaded group.
2271 * Returns the target CPU number, or the same CPU if no balancing is needed.
2273 * preempt must be disabled.
2275 static int sched_balance_self(int cpu
, int flag
)
2277 struct task_struct
*t
= current
;
2278 struct sched_domain
*tmp
, *sd
= NULL
;
2280 for_each_domain(cpu
, tmp
) {
2282 * If power savings logic is enabled for a domain, stop there.
2284 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2286 if (tmp
->flags
& flag
)
2294 struct sched_group
*group
;
2295 int new_cpu
, weight
;
2297 if (!(sd
->flags
& flag
)) {
2302 group
= find_idlest_group(sd
, t
, cpu
);
2308 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
2309 if (new_cpu
== -1 || new_cpu
== cpu
) {
2310 /* Now try balancing at a lower domain level of cpu */
2315 /* Now try balancing at a lower domain level of new_cpu */
2317 weight
= cpumask_weight(sched_domain_span(sd
));
2319 for_each_domain(cpu
, tmp
) {
2320 if (weight
<= cpumask_weight(sched_domain_span(tmp
)))
2322 if (tmp
->flags
& flag
)
2325 /* while loop will break here if sd == NULL */
2331 #endif /* CONFIG_SMP */
2334 * try_to_wake_up - wake up a thread
2335 * @p: the to-be-woken-up thread
2336 * @state: the mask of task states that can be woken
2337 * @sync: do a synchronous wakeup?
2339 * Put it on the run-queue if it's not already there. The "current"
2340 * thread is always on the run-queue (except when the actual
2341 * re-schedule is in progress), and as such you're allowed to do
2342 * the simpler "current->state = TASK_RUNNING" to mark yourself
2343 * runnable without the overhead of this.
2345 * returns failure only if the task is already active.
2347 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2349 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2350 unsigned long flags
;
2354 if (!sched_feat(SYNC_WAKEUPS
))
2358 if (sched_feat(LB_WAKEUP_UPDATE
) && !root_task_group_empty()) {
2359 struct sched_domain
*sd
;
2361 this_cpu
= raw_smp_processor_id();
2364 for_each_domain(this_cpu
, sd
) {
2365 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2374 rq
= task_rq_lock(p
, &flags
);
2375 update_rq_clock(rq
);
2376 old_state
= p
->state
;
2377 if (!(old_state
& state
))
2385 this_cpu
= smp_processor_id();
2388 if (unlikely(task_running(rq
, p
)))
2391 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2392 if (cpu
!= orig_cpu
) {
2393 set_task_cpu(p
, cpu
);
2394 task_rq_unlock(rq
, &flags
);
2395 /* might preempt at this point */
2396 rq
= task_rq_lock(p
, &flags
);
2397 old_state
= p
->state
;
2398 if (!(old_state
& state
))
2403 this_cpu
= smp_processor_id();
2407 #ifdef CONFIG_SCHEDSTATS
2408 schedstat_inc(rq
, ttwu_count
);
2409 if (cpu
== this_cpu
)
2410 schedstat_inc(rq
, ttwu_local
);
2412 struct sched_domain
*sd
;
2413 for_each_domain(this_cpu
, sd
) {
2414 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2415 schedstat_inc(sd
, ttwu_wake_remote
);
2420 #endif /* CONFIG_SCHEDSTATS */
2423 #endif /* CONFIG_SMP */
2424 schedstat_inc(p
, se
.nr_wakeups
);
2426 schedstat_inc(p
, se
.nr_wakeups_sync
);
2427 if (orig_cpu
!= cpu
)
2428 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2429 if (cpu
== this_cpu
)
2430 schedstat_inc(p
, se
.nr_wakeups_local
);
2432 schedstat_inc(p
, se
.nr_wakeups_remote
);
2433 activate_task(rq
, p
, 1);
2437 * Only attribute actual wakeups done by this task.
2439 if (!in_interrupt()) {
2440 struct sched_entity
*se
= ¤t
->se
;
2441 u64 sample
= se
->sum_exec_runtime
;
2443 if (se
->last_wakeup
)
2444 sample
-= se
->last_wakeup
;
2446 sample
-= se
->start_runtime
;
2447 update_avg(&se
->avg_wakeup
, sample
);
2449 se
->last_wakeup
= se
->sum_exec_runtime
;
2453 trace_sched_wakeup(rq
, p
, success
);
2454 check_preempt_curr(rq
, p
, sync
);
2456 p
->state
= TASK_RUNNING
;
2458 if (p
->sched_class
->task_wake_up
)
2459 p
->sched_class
->task_wake_up(rq
, p
);
2462 task_rq_unlock(rq
, &flags
);
2468 * wake_up_process - Wake up a specific process
2469 * @p: The process to be woken up.
2471 * Attempt to wake up the nominated process and move it to the set of runnable
2472 * processes. Returns 1 if the process was woken up, 0 if it was already
2475 * It may be assumed that this function implies a write memory barrier before
2476 * changing the task state if and only if any tasks are woken up.
2478 int wake_up_process(struct task_struct
*p
)
2480 return try_to_wake_up(p
, TASK_ALL
, 0);
2482 EXPORT_SYMBOL(wake_up_process
);
2484 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2486 return try_to_wake_up(p
, state
, 0);
2490 * Perform scheduler related setup for a newly forked process p.
2491 * p is forked by current.
2493 * __sched_fork() is basic setup used by init_idle() too:
2495 static void __sched_fork(struct task_struct
*p
)
2497 p
->se
.exec_start
= 0;
2498 p
->se
.sum_exec_runtime
= 0;
2499 p
->se
.prev_sum_exec_runtime
= 0;
2500 p
->se
.last_wakeup
= 0;
2501 p
->se
.avg_overlap
= 0;
2502 p
->se
.start_runtime
= 0;
2503 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
2505 #ifdef CONFIG_SCHEDSTATS
2506 p
->se
.wait_start
= 0;
2507 p
->se
.sum_sleep_runtime
= 0;
2508 p
->se
.sleep_start
= 0;
2509 p
->se
.block_start
= 0;
2510 p
->se
.sleep_max
= 0;
2511 p
->se
.block_max
= 0;
2513 p
->se
.slice_max
= 0;
2517 INIT_LIST_HEAD(&p
->rt
.run_list
);
2519 INIT_LIST_HEAD(&p
->se
.group_node
);
2521 #ifdef CONFIG_PREEMPT_NOTIFIERS
2522 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2526 * We mark the process as running here, but have not actually
2527 * inserted it onto the runqueue yet. This guarantees that
2528 * nobody will actually run it, and a signal or other external
2529 * event cannot wake it up and insert it on the runqueue either.
2531 p
->state
= TASK_RUNNING
;
2535 * fork()/clone()-time setup:
2537 void sched_fork(struct task_struct
*p
, int clone_flags
)
2539 int cpu
= get_cpu();
2544 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2546 set_task_cpu(p
, cpu
);
2549 * Make sure we do not leak PI boosting priority to the child:
2551 p
->prio
= current
->normal_prio
;
2552 if (!rt_prio(p
->prio
))
2553 p
->sched_class
= &fair_sched_class
;
2555 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2556 if (likely(sched_info_on()))
2557 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2559 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2562 #ifdef CONFIG_PREEMPT
2563 /* Want to start with kernel preemption disabled. */
2564 task_thread_info(p
)->preempt_count
= 1;
2566 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2572 * wake_up_new_task - wake up a newly created task for the first time.
2574 * This function will do some initial scheduler statistics housekeeping
2575 * that must be done for every newly created context, then puts the task
2576 * on the runqueue and wakes it.
2578 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2580 unsigned long flags
;
2583 rq
= task_rq_lock(p
, &flags
);
2584 BUG_ON(p
->state
!= TASK_RUNNING
);
2585 update_rq_clock(rq
);
2587 p
->prio
= effective_prio(p
);
2589 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2590 activate_task(rq
, p
, 0);
2593 * Let the scheduling class do new task startup
2594 * management (if any):
2596 p
->sched_class
->task_new(rq
, p
);
2599 trace_sched_wakeup_new(rq
, p
, 1);
2600 check_preempt_curr(rq
, p
, 0);
2602 if (p
->sched_class
->task_wake_up
)
2603 p
->sched_class
->task_wake_up(rq
, p
);
2605 task_rq_unlock(rq
, &flags
);
2608 #ifdef CONFIG_PREEMPT_NOTIFIERS
2611 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2612 * @notifier: notifier struct to register
2614 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2616 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2618 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2621 * preempt_notifier_unregister - no longer interested in preemption notifications
2622 * @notifier: notifier struct to unregister
2624 * This is safe to call from within a preemption notifier.
2626 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2628 hlist_del(¬ifier
->link
);
2630 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2632 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2634 struct preempt_notifier
*notifier
;
2635 struct hlist_node
*node
;
2637 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2638 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2642 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2643 struct task_struct
*next
)
2645 struct preempt_notifier
*notifier
;
2646 struct hlist_node
*node
;
2648 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2649 notifier
->ops
->sched_out(notifier
, next
);
2652 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2654 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2659 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2660 struct task_struct
*next
)
2664 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2667 * prepare_task_switch - prepare to switch tasks
2668 * @rq: the runqueue preparing to switch
2669 * @prev: the current task that is being switched out
2670 * @next: the task we are going to switch to.
2672 * This is called with the rq lock held and interrupts off. It must
2673 * be paired with a subsequent finish_task_switch after the context
2676 * prepare_task_switch sets up locking and calls architecture specific
2680 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2681 struct task_struct
*next
)
2683 fire_sched_out_preempt_notifiers(prev
, next
);
2684 prepare_lock_switch(rq
, next
);
2685 prepare_arch_switch(next
);
2689 * finish_task_switch - clean up after a task-switch
2690 * @rq: runqueue associated with task-switch
2691 * @prev: the thread we just switched away from.
2693 * finish_task_switch must be called after the context switch, paired
2694 * with a prepare_task_switch call before the context switch.
2695 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2696 * and do any other architecture-specific cleanup actions.
2698 * Note that we may have delayed dropping an mm in context_switch(). If
2699 * so, we finish that here outside of the runqueue lock. (Doing it
2700 * with the lock held can cause deadlocks; see schedule() for
2703 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2704 __releases(rq
->lock
)
2706 struct mm_struct
*mm
= rq
->prev_mm
;
2709 int post_schedule
= 0;
2711 if (current
->sched_class
->needs_post_schedule
)
2712 post_schedule
= current
->sched_class
->needs_post_schedule(rq
);
2718 * A task struct has one reference for the use as "current".
2719 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2720 * schedule one last time. The schedule call will never return, and
2721 * the scheduled task must drop that reference.
2722 * The test for TASK_DEAD must occur while the runqueue locks are
2723 * still held, otherwise prev could be scheduled on another cpu, die
2724 * there before we look at prev->state, and then the reference would
2726 * Manfred Spraul <manfred@colorfullife.com>
2728 prev_state
= prev
->state
;
2729 finish_arch_switch(prev
);
2730 finish_lock_switch(rq
, prev
);
2733 current
->sched_class
->post_schedule(rq
);
2736 fire_sched_in_preempt_notifiers(current
);
2739 if (unlikely(prev_state
== TASK_DEAD
)) {
2741 * Remove function-return probe instances associated with this
2742 * task and put them back on the free list.
2744 kprobe_flush_task(prev
);
2745 put_task_struct(prev
);
2750 * schedule_tail - first thing a freshly forked thread must call.
2751 * @prev: the thread we just switched away from.
2753 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2754 __releases(rq
->lock
)
2756 struct rq
*rq
= this_rq();
2758 finish_task_switch(rq
, prev
);
2759 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2760 /* In this case, finish_task_switch does not reenable preemption */
2763 if (current
->set_child_tid
)
2764 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2768 * context_switch - switch to the new MM and the new
2769 * thread's register state.
2772 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2773 struct task_struct
*next
)
2775 struct mm_struct
*mm
, *oldmm
;
2777 prepare_task_switch(rq
, prev
, next
);
2778 trace_sched_switch(rq
, prev
, next
);
2780 oldmm
= prev
->active_mm
;
2782 * For paravirt, this is coupled with an exit in switch_to to
2783 * combine the page table reload and the switch backend into
2786 arch_start_context_switch(prev
);
2788 if (unlikely(!mm
)) {
2789 next
->active_mm
= oldmm
;
2790 atomic_inc(&oldmm
->mm_count
);
2791 enter_lazy_tlb(oldmm
, next
);
2793 switch_mm(oldmm
, mm
, next
);
2795 if (unlikely(!prev
->mm
)) {
2796 prev
->active_mm
= NULL
;
2797 rq
->prev_mm
= oldmm
;
2800 * Since the runqueue lock will be released by the next
2801 * task (which is an invalid locking op but in the case
2802 * of the scheduler it's an obvious special-case), so we
2803 * do an early lockdep release here:
2805 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2806 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2809 /* Here we just switch the register state and the stack. */
2810 switch_to(prev
, next
, prev
);
2814 * this_rq must be evaluated again because prev may have moved
2815 * CPUs since it called schedule(), thus the 'rq' on its stack
2816 * frame will be invalid.
2818 finish_task_switch(this_rq(), prev
);
2822 * nr_running, nr_uninterruptible and nr_context_switches:
2824 * externally visible scheduler statistics: current number of runnable
2825 * threads, current number of uninterruptible-sleeping threads, total
2826 * number of context switches performed since bootup.
2828 unsigned long nr_running(void)
2830 unsigned long i
, sum
= 0;
2832 for_each_online_cpu(i
)
2833 sum
+= cpu_rq(i
)->nr_running
;
2838 unsigned long nr_uninterruptible(void)
2840 unsigned long i
, sum
= 0;
2842 for_each_possible_cpu(i
)
2843 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2846 * Since we read the counters lockless, it might be slightly
2847 * inaccurate. Do not allow it to go below zero though:
2849 if (unlikely((long)sum
< 0))
2855 unsigned long long nr_context_switches(void)
2858 unsigned long long sum
= 0;
2860 for_each_possible_cpu(i
)
2861 sum
+= cpu_rq(i
)->nr_switches
;
2866 unsigned long nr_iowait(void)
2868 unsigned long i
, sum
= 0;
2870 for_each_possible_cpu(i
)
2871 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2876 /* Variables and functions for calc_load */
2877 static atomic_long_t calc_load_tasks
;
2878 static unsigned long calc_load_update
;
2879 unsigned long avenrun
[3];
2880 EXPORT_SYMBOL(avenrun
);
2883 * get_avenrun - get the load average array
2884 * @loads: pointer to dest load array
2885 * @offset: offset to add
2886 * @shift: shift count to shift the result left
2888 * These values are estimates at best, so no need for locking.
2890 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2892 loads
[0] = (avenrun
[0] + offset
) << shift
;
2893 loads
[1] = (avenrun
[1] + offset
) << shift
;
2894 loads
[2] = (avenrun
[2] + offset
) << shift
;
2897 static unsigned long
2898 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2901 load
+= active
* (FIXED_1
- exp
);
2902 return load
>> FSHIFT
;
2906 * calc_load - update the avenrun load estimates 10 ticks after the
2907 * CPUs have updated calc_load_tasks.
2909 void calc_global_load(void)
2911 unsigned long upd
= calc_load_update
+ 10;
2914 if (time_before(jiffies
, upd
))
2917 active
= atomic_long_read(&calc_load_tasks
);
2918 active
= active
> 0 ? active
* FIXED_1
: 0;
2920 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
2921 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
2922 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
2924 calc_load_update
+= LOAD_FREQ
;
2928 * Either called from update_cpu_load() or from a cpu going idle
2930 static void calc_load_account_active(struct rq
*this_rq
)
2932 long nr_active
, delta
;
2934 nr_active
= this_rq
->nr_running
;
2935 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2937 if (nr_active
!= this_rq
->calc_load_active
) {
2938 delta
= nr_active
- this_rq
->calc_load_active
;
2939 this_rq
->calc_load_active
= nr_active
;
2940 atomic_long_add(delta
, &calc_load_tasks
);
2945 * Update rq->cpu_load[] statistics. This function is usually called every
2946 * scheduler tick (TICK_NSEC).
2948 static void update_cpu_load(struct rq
*this_rq
)
2950 unsigned long this_load
= this_rq
->load
.weight
;
2953 this_rq
->nr_load_updates
++;
2955 /* Update our load: */
2956 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2957 unsigned long old_load
, new_load
;
2959 /* scale is effectively 1 << i now, and >> i divides by scale */
2961 old_load
= this_rq
->cpu_load
[i
];
2962 new_load
= this_load
;
2964 * Round up the averaging division if load is increasing. This
2965 * prevents us from getting stuck on 9 if the load is 10, for
2968 if (new_load
> old_load
)
2969 new_load
+= scale
-1;
2970 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2973 if (time_after_eq(jiffies
, this_rq
->calc_load_update
)) {
2974 this_rq
->calc_load_update
+= LOAD_FREQ
;
2975 calc_load_account_active(this_rq
);
2982 * double_rq_lock - safely lock two runqueues
2984 * Note this does not disable interrupts like task_rq_lock,
2985 * you need to do so manually before calling.
2987 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2988 __acquires(rq1
->lock
)
2989 __acquires(rq2
->lock
)
2991 BUG_ON(!irqs_disabled());
2993 spin_lock(&rq1
->lock
);
2994 __acquire(rq2
->lock
); /* Fake it out ;) */
2997 spin_lock(&rq1
->lock
);
2998 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
3000 spin_lock(&rq2
->lock
);
3001 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
3004 update_rq_clock(rq1
);
3005 update_rq_clock(rq2
);
3009 * double_rq_unlock - safely unlock two runqueues
3011 * Note this does not restore interrupts like task_rq_unlock,
3012 * you need to do so manually after calling.
3014 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
3015 __releases(rq1
->lock
)
3016 __releases(rq2
->lock
)
3018 spin_unlock(&rq1
->lock
);
3020 spin_unlock(&rq2
->lock
);
3022 __release(rq2
->lock
);
3026 * If dest_cpu is allowed for this process, migrate the task to it.
3027 * This is accomplished by forcing the cpu_allowed mask to only
3028 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3029 * the cpu_allowed mask is restored.
3031 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
3033 struct migration_req req
;
3034 unsigned long flags
;
3037 rq
= task_rq_lock(p
, &flags
);
3038 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
3039 || unlikely(!cpu_active(dest_cpu
)))
3042 /* force the process onto the specified CPU */
3043 if (migrate_task(p
, dest_cpu
, &req
)) {
3044 /* Need to wait for migration thread (might exit: take ref). */
3045 struct task_struct
*mt
= rq
->migration_thread
;
3047 get_task_struct(mt
);
3048 task_rq_unlock(rq
, &flags
);
3049 wake_up_process(mt
);
3050 put_task_struct(mt
);
3051 wait_for_completion(&req
.done
);
3056 task_rq_unlock(rq
, &flags
);
3060 * sched_exec - execve() is a valuable balancing opportunity, because at
3061 * this point the task has the smallest effective memory and cache footprint.
3063 void sched_exec(void)
3065 int new_cpu
, this_cpu
= get_cpu();
3066 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
3068 if (new_cpu
!= this_cpu
)
3069 sched_migrate_task(current
, new_cpu
);
3073 * pull_task - move a task from a remote runqueue to the local runqueue.
3074 * Both runqueues must be locked.
3076 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
3077 struct rq
*this_rq
, int this_cpu
)
3079 deactivate_task(src_rq
, p
, 0);
3080 set_task_cpu(p
, this_cpu
);
3081 activate_task(this_rq
, p
, 0);
3083 * Note that idle threads have a prio of MAX_PRIO, for this test
3084 * to be always true for them.
3086 check_preempt_curr(this_rq
, p
, 0);
3090 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3093 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
3094 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3097 int tsk_cache_hot
= 0;
3099 * We do not migrate tasks that are:
3100 * 1) running (obviously), or
3101 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3102 * 3) are cache-hot on their current CPU.
3104 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
3105 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
3110 if (task_running(rq
, p
)) {
3111 schedstat_inc(p
, se
.nr_failed_migrations_running
);
3116 * Aggressive migration if:
3117 * 1) task is cache cold, or
3118 * 2) too many balance attempts have failed.
3121 tsk_cache_hot
= task_hot(p
, rq
->clock
, sd
);
3122 if (!tsk_cache_hot
||
3123 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
3124 #ifdef CONFIG_SCHEDSTATS
3125 if (tsk_cache_hot
) {
3126 schedstat_inc(sd
, lb_hot_gained
[idle
]);
3127 schedstat_inc(p
, se
.nr_forced_migrations
);
3133 if (tsk_cache_hot
) {
3134 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
3140 static unsigned long
3141 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3142 unsigned long max_load_move
, struct sched_domain
*sd
,
3143 enum cpu_idle_type idle
, int *all_pinned
,
3144 int *this_best_prio
, struct rq_iterator
*iterator
)
3146 int loops
= 0, pulled
= 0, pinned
= 0;
3147 struct task_struct
*p
;
3148 long rem_load_move
= max_load_move
;
3150 if (max_load_move
== 0)
3156 * Start the load-balancing iterator:
3158 p
= iterator
->start(iterator
->arg
);
3160 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3163 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
3164 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3165 p
= iterator
->next(iterator
->arg
);
3169 pull_task(busiest
, p
, this_rq
, this_cpu
);
3171 rem_load_move
-= p
->se
.load
.weight
;
3173 #ifdef CONFIG_PREEMPT
3175 * NEWIDLE balancing is a source of latency, so preemptible kernels
3176 * will stop after the first task is pulled to minimize the critical
3179 if (idle
== CPU_NEWLY_IDLE
)
3184 * We only want to steal up to the prescribed amount of weighted load.
3186 if (rem_load_move
> 0) {
3187 if (p
->prio
< *this_best_prio
)
3188 *this_best_prio
= p
->prio
;
3189 p
= iterator
->next(iterator
->arg
);
3194 * Right now, this is one of only two places pull_task() is called,
3195 * so we can safely collect pull_task() stats here rather than
3196 * inside pull_task().
3198 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3201 *all_pinned
= pinned
;
3203 return max_load_move
- rem_load_move
;
3207 * move_tasks tries to move up to max_load_move weighted load from busiest to
3208 * this_rq, as part of a balancing operation within domain "sd".
3209 * Returns 1 if successful and 0 otherwise.
3211 * Called with both runqueues locked.
3213 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3214 unsigned long max_load_move
,
3215 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3218 const struct sched_class
*class = sched_class_highest
;
3219 unsigned long total_load_moved
= 0;
3220 int this_best_prio
= this_rq
->curr
->prio
;
3224 class->load_balance(this_rq
, this_cpu
, busiest
,
3225 max_load_move
- total_load_moved
,
3226 sd
, idle
, all_pinned
, &this_best_prio
);
3227 class = class->next
;
3229 #ifdef CONFIG_PREEMPT
3231 * NEWIDLE balancing is a source of latency, so preemptible
3232 * kernels will stop after the first task is pulled to minimize
3233 * the critical section.
3235 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3238 } while (class && max_load_move
> total_load_moved
);
3240 return total_load_moved
> 0;
3244 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3245 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3246 struct rq_iterator
*iterator
)
3248 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3252 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3253 pull_task(busiest
, p
, this_rq
, this_cpu
);
3255 * Right now, this is only the second place pull_task()
3256 * is called, so we can safely collect pull_task()
3257 * stats here rather than inside pull_task().
3259 schedstat_inc(sd
, lb_gained
[idle
]);
3263 p
= iterator
->next(iterator
->arg
);
3270 * move_one_task tries to move exactly one task from busiest to this_rq, as
3271 * part of active balancing operations within "domain".
3272 * Returns 1 if successful and 0 otherwise.
3274 * Called with both runqueues locked.
3276 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3277 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3279 const struct sched_class
*class;
3281 for (class = sched_class_highest
; class; class = class->next
)
3282 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3287 /********** Helpers for find_busiest_group ************************/
3289 * sd_lb_stats - Structure to store the statistics of a sched_domain
3290 * during load balancing.
3292 struct sd_lb_stats
{
3293 struct sched_group
*busiest
; /* Busiest group in this sd */
3294 struct sched_group
*this; /* Local group in this sd */
3295 unsigned long total_load
; /* Total load of all groups in sd */
3296 unsigned long total_pwr
; /* Total power of all groups in sd */
3297 unsigned long avg_load
; /* Average load across all groups in sd */
3299 /** Statistics of this group */
3300 unsigned long this_load
;
3301 unsigned long this_load_per_task
;
3302 unsigned long this_nr_running
;
3304 /* Statistics of the busiest group */
3305 unsigned long max_load
;
3306 unsigned long busiest_load_per_task
;
3307 unsigned long busiest_nr_running
;
3309 int group_imb
; /* Is there imbalance in this sd */
3310 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3311 int power_savings_balance
; /* Is powersave balance needed for this sd */
3312 struct sched_group
*group_min
; /* Least loaded group in sd */
3313 struct sched_group
*group_leader
; /* Group which relieves group_min */
3314 unsigned long min_load_per_task
; /* load_per_task in group_min */
3315 unsigned long leader_nr_running
; /* Nr running of group_leader */
3316 unsigned long min_nr_running
; /* Nr running of group_min */
3321 * sg_lb_stats - stats of a sched_group required for load_balancing
3323 struct sg_lb_stats
{
3324 unsigned long avg_load
; /*Avg load across the CPUs of the group */
3325 unsigned long group_load
; /* Total load over the CPUs of the group */
3326 unsigned long sum_nr_running
; /* Nr tasks running in the group */
3327 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
3328 unsigned long group_capacity
;
3329 int group_imb
; /* Is there an imbalance in the group ? */
3333 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3334 * @group: The group whose first cpu is to be returned.
3336 static inline unsigned int group_first_cpu(struct sched_group
*group
)
3338 return cpumask_first(sched_group_cpus(group
));
3342 * get_sd_load_idx - Obtain the load index for a given sched domain.
3343 * @sd: The sched_domain whose load_idx is to be obtained.
3344 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3346 static inline int get_sd_load_idx(struct sched_domain
*sd
,
3347 enum cpu_idle_type idle
)
3353 load_idx
= sd
->busy_idx
;
3356 case CPU_NEWLY_IDLE
:
3357 load_idx
= sd
->newidle_idx
;
3360 load_idx
= sd
->idle_idx
;
3368 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3370 * init_sd_power_savings_stats - Initialize power savings statistics for
3371 * the given sched_domain, during load balancing.
3373 * @sd: Sched domain whose power-savings statistics are to be initialized.
3374 * @sds: Variable containing the statistics for sd.
3375 * @idle: Idle status of the CPU at which we're performing load-balancing.
3377 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3378 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3381 * Busy processors will not participate in power savings
3384 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3385 sds
->power_savings_balance
= 0;
3387 sds
->power_savings_balance
= 1;
3388 sds
->min_nr_running
= ULONG_MAX
;
3389 sds
->leader_nr_running
= 0;
3394 * update_sd_power_savings_stats - Update the power saving stats for a
3395 * sched_domain while performing load balancing.
3397 * @group: sched_group belonging to the sched_domain under consideration.
3398 * @sds: Variable containing the statistics of the sched_domain
3399 * @local_group: Does group contain the CPU for which we're performing
3401 * @sgs: Variable containing the statistics of the group.
3403 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3404 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3407 if (!sds
->power_savings_balance
)
3411 * If the local group is idle or completely loaded
3412 * no need to do power savings balance at this domain
3414 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
3415 !sds
->this_nr_running
))
3416 sds
->power_savings_balance
= 0;
3419 * If a group is already running at full capacity or idle,
3420 * don't include that group in power savings calculations
3422 if (!sds
->power_savings_balance
||
3423 sgs
->sum_nr_running
>= sgs
->group_capacity
||
3424 !sgs
->sum_nr_running
)
3428 * Calculate the group which has the least non-idle load.
3429 * This is the group from where we need to pick up the load
3432 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
3433 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
3434 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
3435 sds
->group_min
= group
;
3436 sds
->min_nr_running
= sgs
->sum_nr_running
;
3437 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
3438 sgs
->sum_nr_running
;
3442 * Calculate the group which is almost near its
3443 * capacity but still has some space to pick up some load
3444 * from other group and save more power
3446 if (sgs
->sum_nr_running
> sgs
->group_capacity
- 1)
3449 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
3450 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
3451 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
3452 sds
->group_leader
= group
;
3453 sds
->leader_nr_running
= sgs
->sum_nr_running
;
3458 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3459 * @sds: Variable containing the statistics of the sched_domain
3460 * under consideration.
3461 * @this_cpu: Cpu at which we're currently performing load-balancing.
3462 * @imbalance: Variable to store the imbalance.
3465 * Check if we have potential to perform some power-savings balance.
3466 * If yes, set the busiest group to be the least loaded group in the
3467 * sched_domain, so that it's CPUs can be put to idle.
3469 * Returns 1 if there is potential to perform power-savings balance.
3472 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3473 int this_cpu
, unsigned long *imbalance
)
3475 if (!sds
->power_savings_balance
)
3478 if (sds
->this != sds
->group_leader
||
3479 sds
->group_leader
== sds
->group_min
)
3482 *imbalance
= sds
->min_load_per_task
;
3483 sds
->busiest
= sds
->group_min
;
3485 if (sched_mc_power_savings
>= POWERSAVINGS_BALANCE_WAKEUP
) {
3486 cpu_rq(this_cpu
)->rd
->sched_mc_preferred_wakeup_cpu
=
3487 group_first_cpu(sds
->group_leader
);
3493 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3494 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3495 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3500 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3501 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3506 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3507 int this_cpu
, unsigned long *imbalance
)
3511 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3515 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3516 * @group: sched_group whose statistics are to be updated.
3517 * @this_cpu: Cpu for which load balance is currently performed.
3518 * @idle: Idle status of this_cpu
3519 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3520 * @sd_idle: Idle status of the sched_domain containing group.
3521 * @local_group: Does group contain this_cpu.
3522 * @cpus: Set of cpus considered for load balancing.
3523 * @balance: Should we balance.
3524 * @sgs: variable to hold the statistics for this group.
3526 static inline void update_sg_lb_stats(struct sched_group
*group
, int this_cpu
,
3527 enum cpu_idle_type idle
, int load_idx
, int *sd_idle
,
3528 int local_group
, const struct cpumask
*cpus
,
3529 int *balance
, struct sg_lb_stats
*sgs
)
3531 unsigned long load
, max_cpu_load
, min_cpu_load
;
3533 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3534 unsigned long sum_avg_load_per_task
;
3535 unsigned long avg_load_per_task
;
3538 balance_cpu
= group_first_cpu(group
);
3540 /* Tally up the load of all CPUs in the group */
3541 sum_avg_load_per_task
= avg_load_per_task
= 0;
3543 min_cpu_load
= ~0UL;
3545 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3546 struct rq
*rq
= cpu_rq(i
);
3548 if (*sd_idle
&& rq
->nr_running
)
3551 /* Bias balancing toward cpus of our domain */
3553 if (idle_cpu(i
) && !first_idle_cpu
) {
3558 load
= target_load(i
, load_idx
);
3560 load
= source_load(i
, load_idx
);
3561 if (load
> max_cpu_load
)
3562 max_cpu_load
= load
;
3563 if (min_cpu_load
> load
)
3564 min_cpu_load
= load
;
3567 sgs
->group_load
+= load
;
3568 sgs
->sum_nr_running
+= rq
->nr_running
;
3569 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
3571 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3575 * First idle cpu or the first cpu(busiest) in this sched group
3576 * is eligible for doing load balancing at this and above
3577 * domains. In the newly idle case, we will allow all the cpu's
3578 * to do the newly idle load balance.
3580 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3581 balance_cpu
!= this_cpu
&& balance
) {
3586 /* Adjust by relative CPU power of the group */
3587 sgs
->avg_load
= sg_div_cpu_power(group
,
3588 sgs
->group_load
* SCHED_LOAD_SCALE
);
3592 * Consider the group unbalanced when the imbalance is larger
3593 * than the average weight of two tasks.
3595 * APZ: with cgroup the avg task weight can vary wildly and
3596 * might not be a suitable number - should we keep a
3597 * normalized nr_running number somewhere that negates
3600 avg_load_per_task
= sg_div_cpu_power(group
,
3601 sum_avg_load_per_task
* SCHED_LOAD_SCALE
);
3603 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3606 sgs
->group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3611 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3612 * @sd: sched_domain whose statistics are to be updated.
3613 * @this_cpu: Cpu for which load balance is currently performed.
3614 * @idle: Idle status of this_cpu
3615 * @sd_idle: Idle status of the sched_domain containing group.
3616 * @cpus: Set of cpus considered for load balancing.
3617 * @balance: Should we balance.
3618 * @sds: variable to hold the statistics for this sched_domain.
3620 static inline void update_sd_lb_stats(struct sched_domain
*sd
, int this_cpu
,
3621 enum cpu_idle_type idle
, int *sd_idle
,
3622 const struct cpumask
*cpus
, int *balance
,
3623 struct sd_lb_stats
*sds
)
3625 struct sched_group
*group
= sd
->groups
;
3626 struct sg_lb_stats sgs
;
3629 init_sd_power_savings_stats(sd
, sds
, idle
);
3630 load_idx
= get_sd_load_idx(sd
, idle
);
3635 local_group
= cpumask_test_cpu(this_cpu
,
3636 sched_group_cpus(group
));
3637 memset(&sgs
, 0, sizeof(sgs
));
3638 update_sg_lb_stats(group
, this_cpu
, idle
, load_idx
, sd_idle
,
3639 local_group
, cpus
, balance
, &sgs
);
3641 if (local_group
&& balance
&& !(*balance
))
3644 sds
->total_load
+= sgs
.group_load
;
3645 sds
->total_pwr
+= group
->__cpu_power
;
3648 sds
->this_load
= sgs
.avg_load
;
3650 sds
->this_nr_running
= sgs
.sum_nr_running
;
3651 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
3652 } else if (sgs
.avg_load
> sds
->max_load
&&
3653 (sgs
.sum_nr_running
> sgs
.group_capacity
||
3655 sds
->max_load
= sgs
.avg_load
;
3656 sds
->busiest
= group
;
3657 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
3658 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
3659 sds
->group_imb
= sgs
.group_imb
;
3662 update_sd_power_savings_stats(group
, sds
, local_group
, &sgs
);
3663 group
= group
->next
;
3664 } while (group
!= sd
->groups
);
3669 * fix_small_imbalance - Calculate the minor imbalance that exists
3670 * amongst the groups of a sched_domain, during
3672 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3673 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3674 * @imbalance: Variable to store the imbalance.
3676 static inline void fix_small_imbalance(struct sd_lb_stats
*sds
,
3677 int this_cpu
, unsigned long *imbalance
)
3679 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
3680 unsigned int imbn
= 2;
3682 if (sds
->this_nr_running
) {
3683 sds
->this_load_per_task
/= sds
->this_nr_running
;
3684 if (sds
->busiest_load_per_task
>
3685 sds
->this_load_per_task
)
3688 sds
->this_load_per_task
=
3689 cpu_avg_load_per_task(this_cpu
);
3691 if (sds
->max_load
- sds
->this_load
+ sds
->busiest_load_per_task
>=
3692 sds
->busiest_load_per_task
* imbn
) {
3693 *imbalance
= sds
->busiest_load_per_task
;
3698 * OK, we don't have enough imbalance to justify moving tasks,
3699 * however we may be able to increase total CPU power used by
3703 pwr_now
+= sds
->busiest
->__cpu_power
*
3704 min(sds
->busiest_load_per_task
, sds
->max_load
);
3705 pwr_now
+= sds
->this->__cpu_power
*
3706 min(sds
->this_load_per_task
, sds
->this_load
);
3707 pwr_now
/= SCHED_LOAD_SCALE
;
3709 /* Amount of load we'd subtract */
3710 tmp
= sg_div_cpu_power(sds
->busiest
,
3711 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
);
3712 if (sds
->max_load
> tmp
)
3713 pwr_move
+= sds
->busiest
->__cpu_power
*
3714 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
3716 /* Amount of load we'd add */
3717 if (sds
->max_load
* sds
->busiest
->__cpu_power
<
3718 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
)
3719 tmp
= sg_div_cpu_power(sds
->this,
3720 sds
->max_load
* sds
->busiest
->__cpu_power
);
3722 tmp
= sg_div_cpu_power(sds
->this,
3723 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
);
3724 pwr_move
+= sds
->this->__cpu_power
*
3725 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
3726 pwr_move
/= SCHED_LOAD_SCALE
;
3728 /* Move if we gain throughput */
3729 if (pwr_move
> pwr_now
)
3730 *imbalance
= sds
->busiest_load_per_task
;
3734 * calculate_imbalance - Calculate the amount of imbalance present within the
3735 * groups of a given sched_domain during load balance.
3736 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3737 * @this_cpu: Cpu for which currently load balance is being performed.
3738 * @imbalance: The variable to store the imbalance.
3740 static inline void calculate_imbalance(struct sd_lb_stats
*sds
, int this_cpu
,
3741 unsigned long *imbalance
)
3743 unsigned long max_pull
;
3745 * In the presence of smp nice balancing, certain scenarios can have
3746 * max load less than avg load(as we skip the groups at or below
3747 * its cpu_power, while calculating max_load..)
3749 if (sds
->max_load
< sds
->avg_load
) {
3751 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3754 /* Don't want to pull so many tasks that a group would go idle */
3755 max_pull
= min(sds
->max_load
- sds
->avg_load
,
3756 sds
->max_load
- sds
->busiest_load_per_task
);
3758 /* How much load to actually move to equalise the imbalance */
3759 *imbalance
= min(max_pull
* sds
->busiest
->__cpu_power
,
3760 (sds
->avg_load
- sds
->this_load
) * sds
->this->__cpu_power
)
3764 * if *imbalance is less than the average load per runnable task
3765 * there is no gaurantee that any tasks will be moved so we'll have
3766 * a think about bumping its value to force at least one task to be
3769 if (*imbalance
< sds
->busiest_load_per_task
)
3770 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3773 /******* find_busiest_group() helpers end here *********************/
3776 * find_busiest_group - Returns the busiest group within the sched_domain
3777 * if there is an imbalance. If there isn't an imbalance, and
3778 * the user has opted for power-savings, it returns a group whose
3779 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3780 * such a group exists.
3782 * Also calculates the amount of weighted load which should be moved
3783 * to restore balance.
3785 * @sd: The sched_domain whose busiest group is to be returned.
3786 * @this_cpu: The cpu for which load balancing is currently being performed.
3787 * @imbalance: Variable which stores amount of weighted load which should
3788 * be moved to restore balance/put a group to idle.
3789 * @idle: The idle status of this_cpu.
3790 * @sd_idle: The idleness of sd
3791 * @cpus: The set of CPUs under consideration for load-balancing.
3792 * @balance: Pointer to a variable indicating if this_cpu
3793 * is the appropriate cpu to perform load balancing at this_level.
3795 * Returns: - the busiest group if imbalance exists.
3796 * - If no imbalance and user has opted for power-savings balance,
3797 * return the least loaded group whose CPUs can be
3798 * put to idle by rebalancing its tasks onto our group.
3800 static struct sched_group
*
3801 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3802 unsigned long *imbalance
, enum cpu_idle_type idle
,
3803 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
3805 struct sd_lb_stats sds
;
3807 memset(&sds
, 0, sizeof(sds
));
3810 * Compute the various statistics relavent for load balancing at
3813 update_sd_lb_stats(sd
, this_cpu
, idle
, sd_idle
, cpus
,
3816 /* Cases where imbalance does not exist from POV of this_cpu */
3817 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3819 * 2) There is no busy sibling group to pull from.
3820 * 3) This group is the busiest group.
3821 * 4) This group is more busy than the avg busieness at this
3823 * 5) The imbalance is within the specified limit.
3824 * 6) Any rebalance would lead to ping-pong
3826 if (balance
&& !(*balance
))
3829 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
3832 if (sds
.this_load
>= sds
.max_load
)
3835 sds
.avg_load
= (SCHED_LOAD_SCALE
* sds
.total_load
) / sds
.total_pwr
;
3837 if (sds
.this_load
>= sds
.avg_load
)
3840 if (100 * sds
.max_load
<= sd
->imbalance_pct
* sds
.this_load
)
3843 sds
.busiest_load_per_task
/= sds
.busiest_nr_running
;
3845 sds
.busiest_load_per_task
=
3846 min(sds
.busiest_load_per_task
, sds
.avg_load
);
3849 * We're trying to get all the cpus to the average_load, so we don't
3850 * want to push ourselves above the average load, nor do we wish to
3851 * reduce the max loaded cpu below the average load, as either of these
3852 * actions would just result in more rebalancing later, and ping-pong
3853 * tasks around. Thus we look for the minimum possible imbalance.
3854 * Negative imbalances (*we* are more loaded than anyone else) will
3855 * be counted as no imbalance for these purposes -- we can't fix that
3856 * by pulling tasks to us. Be careful of negative numbers as they'll
3857 * appear as very large values with unsigned longs.
3859 if (sds
.max_load
<= sds
.busiest_load_per_task
)
3862 /* Looks like there is an imbalance. Compute it */
3863 calculate_imbalance(&sds
, this_cpu
, imbalance
);
3868 * There is no obvious imbalance. But check if we can do some balancing
3871 if (check_power_save_busiest_group(&sds
, this_cpu
, imbalance
))
3879 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3882 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3883 unsigned long imbalance
, const struct cpumask
*cpus
)
3885 struct rq
*busiest
= NULL
, *rq
;
3886 unsigned long max_load
= 0;
3889 for_each_cpu(i
, sched_group_cpus(group
)) {
3892 if (!cpumask_test_cpu(i
, cpus
))
3896 wl
= weighted_cpuload(i
);
3898 if (rq
->nr_running
== 1 && wl
> imbalance
)
3901 if (wl
> max_load
) {
3911 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3912 * so long as it is large enough.
3914 #define MAX_PINNED_INTERVAL 512
3916 /* Working cpumask for load_balance and load_balance_newidle. */
3917 static DEFINE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
3920 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3921 * tasks if there is an imbalance.
3923 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3924 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3927 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3928 struct sched_group
*group
;
3929 unsigned long imbalance
;
3931 unsigned long flags
;
3932 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
3934 cpumask_setall(cpus
);
3937 * When power savings policy is enabled for the parent domain, idle
3938 * sibling can pick up load irrespective of busy siblings. In this case,
3939 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3940 * portraying it as CPU_NOT_IDLE.
3942 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3943 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3946 schedstat_inc(sd
, lb_count
[idle
]);
3950 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3957 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3961 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3963 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3967 BUG_ON(busiest
== this_rq
);
3969 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3972 if (busiest
->nr_running
> 1) {
3974 * Attempt to move tasks. If find_busiest_group has found
3975 * an imbalance but busiest->nr_running <= 1, the group is
3976 * still unbalanced. ld_moved simply stays zero, so it is
3977 * correctly treated as an imbalance.
3979 local_irq_save(flags
);
3980 double_rq_lock(this_rq
, busiest
);
3981 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3982 imbalance
, sd
, idle
, &all_pinned
);
3983 double_rq_unlock(this_rq
, busiest
);
3984 local_irq_restore(flags
);
3987 * some other cpu did the load balance for us.
3989 if (ld_moved
&& this_cpu
!= smp_processor_id())
3990 resched_cpu(this_cpu
);
3992 /* All tasks on this runqueue were pinned by CPU affinity */
3993 if (unlikely(all_pinned
)) {
3994 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
3995 if (!cpumask_empty(cpus
))
4002 schedstat_inc(sd
, lb_failed
[idle
]);
4003 sd
->nr_balance_failed
++;
4005 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
4007 spin_lock_irqsave(&busiest
->lock
, flags
);
4009 /* don't kick the migration_thread, if the curr
4010 * task on busiest cpu can't be moved to this_cpu
4012 if (!cpumask_test_cpu(this_cpu
,
4013 &busiest
->curr
->cpus_allowed
)) {
4014 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4016 goto out_one_pinned
;
4019 if (!busiest
->active_balance
) {
4020 busiest
->active_balance
= 1;
4021 busiest
->push_cpu
= this_cpu
;
4024 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4026 wake_up_process(busiest
->migration_thread
);
4029 * We've kicked active balancing, reset the failure
4032 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
4035 sd
->nr_balance_failed
= 0;
4037 if (likely(!active_balance
)) {
4038 /* We were unbalanced, so reset the balancing interval */
4039 sd
->balance_interval
= sd
->min_interval
;
4042 * If we've begun active balancing, start to back off. This
4043 * case may not be covered by the all_pinned logic if there
4044 * is only 1 task on the busy runqueue (because we don't call
4047 if (sd
->balance_interval
< sd
->max_interval
)
4048 sd
->balance_interval
*= 2;
4051 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4052 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4058 schedstat_inc(sd
, lb_balanced
[idle
]);
4060 sd
->nr_balance_failed
= 0;
4063 /* tune up the balancing interval */
4064 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
4065 (sd
->balance_interval
< sd
->max_interval
))
4066 sd
->balance_interval
*= 2;
4068 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4069 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4080 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4081 * tasks if there is an imbalance.
4083 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4084 * this_rq is locked.
4087 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
4089 struct sched_group
*group
;
4090 struct rq
*busiest
= NULL
;
4091 unsigned long imbalance
;
4095 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4097 cpumask_setall(cpus
);
4100 * When power savings policy is enabled for the parent domain, idle
4101 * sibling can pick up load irrespective of busy siblings. In this case,
4102 * let the state of idle sibling percolate up as IDLE, instead of
4103 * portraying it as CPU_NOT_IDLE.
4105 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
4106 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4109 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
4111 update_shares_locked(this_rq
, sd
);
4112 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
4113 &sd_idle
, cpus
, NULL
);
4115 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
4119 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
4121 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
4125 BUG_ON(busiest
== this_rq
);
4127 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
4130 if (busiest
->nr_running
> 1) {
4131 /* Attempt to move tasks */
4132 double_lock_balance(this_rq
, busiest
);
4133 /* this_rq->clock is already updated */
4134 update_rq_clock(busiest
);
4135 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4136 imbalance
, sd
, CPU_NEWLY_IDLE
,
4138 double_unlock_balance(this_rq
, busiest
);
4140 if (unlikely(all_pinned
)) {
4141 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4142 if (!cpumask_empty(cpus
))
4148 int active_balance
= 0;
4150 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
4151 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4152 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4155 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
4158 if (sd
->nr_balance_failed
++ < 2)
4162 * The only task running in a non-idle cpu can be moved to this
4163 * cpu in an attempt to completely freeup the other CPU
4164 * package. The same method used to move task in load_balance()
4165 * have been extended for load_balance_newidle() to speedup
4166 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4168 * The package power saving logic comes from
4169 * find_busiest_group(). If there are no imbalance, then
4170 * f_b_g() will return NULL. However when sched_mc={1,2} then
4171 * f_b_g() will select a group from which a running task may be
4172 * pulled to this cpu in order to make the other package idle.
4173 * If there is no opportunity to make a package idle and if
4174 * there are no imbalance, then f_b_g() will return NULL and no
4175 * action will be taken in load_balance_newidle().
4177 * Under normal task pull operation due to imbalance, there
4178 * will be more than one task in the source run queue and
4179 * move_tasks() will succeed. ld_moved will be true and this
4180 * active balance code will not be triggered.
4183 /* Lock busiest in correct order while this_rq is held */
4184 double_lock_balance(this_rq
, busiest
);
4187 * don't kick the migration_thread, if the curr
4188 * task on busiest cpu can't be moved to this_cpu
4190 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
4191 double_unlock_balance(this_rq
, busiest
);
4196 if (!busiest
->active_balance
) {
4197 busiest
->active_balance
= 1;
4198 busiest
->push_cpu
= this_cpu
;
4202 double_unlock_balance(this_rq
, busiest
);
4204 * Should not call ttwu while holding a rq->lock
4206 spin_unlock(&this_rq
->lock
);
4208 wake_up_process(busiest
->migration_thread
);
4209 spin_lock(&this_rq
->lock
);
4212 sd
->nr_balance_failed
= 0;
4214 update_shares_locked(this_rq
, sd
);
4218 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
4219 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4220 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4222 sd
->nr_balance_failed
= 0;
4228 * idle_balance is called by schedule() if this_cpu is about to become
4229 * idle. Attempts to pull tasks from other CPUs.
4231 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
4233 struct sched_domain
*sd
;
4234 int pulled_task
= 0;
4235 unsigned long next_balance
= jiffies
+ HZ
;
4237 for_each_domain(this_cpu
, sd
) {
4238 unsigned long interval
;
4240 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4243 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
4244 /* If we've pulled tasks over stop searching: */
4245 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
4248 interval
= msecs_to_jiffies(sd
->balance_interval
);
4249 if (time_after(next_balance
, sd
->last_balance
+ interval
))
4250 next_balance
= sd
->last_balance
+ interval
;
4254 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
4256 * We are going idle. next_balance may be set based on
4257 * a busy processor. So reset next_balance.
4259 this_rq
->next_balance
= next_balance
;
4264 * active_load_balance is run by migration threads. It pushes running tasks
4265 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4266 * running on each physical CPU where possible, and avoids physical /
4267 * logical imbalances.
4269 * Called with busiest_rq locked.
4271 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
4273 int target_cpu
= busiest_rq
->push_cpu
;
4274 struct sched_domain
*sd
;
4275 struct rq
*target_rq
;
4277 /* Is there any task to move? */
4278 if (busiest_rq
->nr_running
<= 1)
4281 target_rq
= cpu_rq(target_cpu
);
4284 * This condition is "impossible", if it occurs
4285 * we need to fix it. Originally reported by
4286 * Bjorn Helgaas on a 128-cpu setup.
4288 BUG_ON(busiest_rq
== target_rq
);
4290 /* move a task from busiest_rq to target_rq */
4291 double_lock_balance(busiest_rq
, target_rq
);
4292 update_rq_clock(busiest_rq
);
4293 update_rq_clock(target_rq
);
4295 /* Search for an sd spanning us and the target CPU. */
4296 for_each_domain(target_cpu
, sd
) {
4297 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
4298 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
4303 schedstat_inc(sd
, alb_count
);
4305 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
4307 schedstat_inc(sd
, alb_pushed
);
4309 schedstat_inc(sd
, alb_failed
);
4311 double_unlock_balance(busiest_rq
, target_rq
);
4316 atomic_t load_balancer
;
4317 cpumask_var_t cpu_mask
;
4318 cpumask_var_t ilb_grp_nohz_mask
;
4319 } nohz ____cacheline_aligned
= {
4320 .load_balancer
= ATOMIC_INIT(-1),
4323 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4325 * lowest_flag_domain - Return lowest sched_domain containing flag.
4326 * @cpu: The cpu whose lowest level of sched domain is to
4328 * @flag: The flag to check for the lowest sched_domain
4329 * for the given cpu.
4331 * Returns the lowest sched_domain of a cpu which contains the given flag.
4333 static inline struct sched_domain
*lowest_flag_domain(int cpu
, int flag
)
4335 struct sched_domain
*sd
;
4337 for_each_domain(cpu
, sd
)
4338 if (sd
&& (sd
->flags
& flag
))
4345 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4346 * @cpu: The cpu whose domains we're iterating over.
4347 * @sd: variable holding the value of the power_savings_sd
4349 * @flag: The flag to filter the sched_domains to be iterated.
4351 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4352 * set, starting from the lowest sched_domain to the highest.
4354 #define for_each_flag_domain(cpu, sd, flag) \
4355 for (sd = lowest_flag_domain(cpu, flag); \
4356 (sd && (sd->flags & flag)); sd = sd->parent)
4359 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4360 * @ilb_group: group to be checked for semi-idleness
4362 * Returns: 1 if the group is semi-idle. 0 otherwise.
4364 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4365 * and atleast one non-idle CPU. This helper function checks if the given
4366 * sched_group is semi-idle or not.
4368 static inline int is_semi_idle_group(struct sched_group
*ilb_group
)
4370 cpumask_and(nohz
.ilb_grp_nohz_mask
, nohz
.cpu_mask
,
4371 sched_group_cpus(ilb_group
));
4374 * A sched_group is semi-idle when it has atleast one busy cpu
4375 * and atleast one idle cpu.
4377 if (cpumask_empty(nohz
.ilb_grp_nohz_mask
))
4380 if (cpumask_equal(nohz
.ilb_grp_nohz_mask
, sched_group_cpus(ilb_group
)))
4386 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4387 * @cpu: The cpu which is nominating a new idle_load_balancer.
4389 * Returns: Returns the id of the idle load balancer if it exists,
4390 * Else, returns >= nr_cpu_ids.
4392 * This algorithm picks the idle load balancer such that it belongs to a
4393 * semi-idle powersavings sched_domain. The idea is to try and avoid
4394 * completely idle packages/cores just for the purpose of idle load balancing
4395 * when there are other idle cpu's which are better suited for that job.
4397 static int find_new_ilb(int cpu
)
4399 struct sched_domain
*sd
;
4400 struct sched_group
*ilb_group
;
4403 * Have idle load balancer selection from semi-idle packages only
4404 * when power-aware load balancing is enabled
4406 if (!(sched_smt_power_savings
|| sched_mc_power_savings
))
4410 * Optimize for the case when we have no idle CPUs or only one
4411 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4413 if (cpumask_weight(nohz
.cpu_mask
) < 2)
4416 for_each_flag_domain(cpu
, sd
, SD_POWERSAVINGS_BALANCE
) {
4417 ilb_group
= sd
->groups
;
4420 if (is_semi_idle_group(ilb_group
))
4421 return cpumask_first(nohz
.ilb_grp_nohz_mask
);
4423 ilb_group
= ilb_group
->next
;
4425 } while (ilb_group
!= sd
->groups
);
4429 return cpumask_first(nohz
.cpu_mask
);
4431 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4432 static inline int find_new_ilb(int call_cpu
)
4434 return cpumask_first(nohz
.cpu_mask
);
4439 * This routine will try to nominate the ilb (idle load balancing)
4440 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4441 * load balancing on behalf of all those cpus. If all the cpus in the system
4442 * go into this tickless mode, then there will be no ilb owner (as there is
4443 * no need for one) and all the cpus will sleep till the next wakeup event
4446 * For the ilb owner, tick is not stopped. And this tick will be used
4447 * for idle load balancing. ilb owner will still be part of
4450 * While stopping the tick, this cpu will become the ilb owner if there
4451 * is no other owner. And will be the owner till that cpu becomes busy
4452 * or if all cpus in the system stop their ticks at which point
4453 * there is no need for ilb owner.
4455 * When the ilb owner becomes busy, it nominates another owner, during the
4456 * next busy scheduler_tick()
4458 int select_nohz_load_balancer(int stop_tick
)
4460 int cpu
= smp_processor_id();
4463 cpu_rq(cpu
)->in_nohz_recently
= 1;
4465 if (!cpu_active(cpu
)) {
4466 if (atomic_read(&nohz
.load_balancer
) != cpu
)
4470 * If we are going offline and still the leader,
4473 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4479 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
4481 /* time for ilb owner also to sleep */
4482 if (cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4483 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4484 atomic_set(&nohz
.load_balancer
, -1);
4488 if (atomic_read(&nohz
.load_balancer
) == -1) {
4489 /* make me the ilb owner */
4490 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
4492 } else if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4495 if (!(sched_smt_power_savings
||
4496 sched_mc_power_savings
))
4499 * Check to see if there is a more power-efficient
4502 new_ilb
= find_new_ilb(cpu
);
4503 if (new_ilb
< nr_cpu_ids
&& new_ilb
!= cpu
) {
4504 atomic_set(&nohz
.load_balancer
, -1);
4505 resched_cpu(new_ilb
);
4511 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4514 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4516 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4517 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4524 static DEFINE_SPINLOCK(balancing
);
4527 * It checks each scheduling domain to see if it is due to be balanced,
4528 * and initiates a balancing operation if so.
4530 * Balancing parameters are set up in arch_init_sched_domains.
4532 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
4535 struct rq
*rq
= cpu_rq(cpu
);
4536 unsigned long interval
;
4537 struct sched_domain
*sd
;
4538 /* Earliest time when we have to do rebalance again */
4539 unsigned long next_balance
= jiffies
+ 60*HZ
;
4540 int update_next_balance
= 0;
4543 for_each_domain(cpu
, sd
) {
4544 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4547 interval
= sd
->balance_interval
;
4548 if (idle
!= CPU_IDLE
)
4549 interval
*= sd
->busy_factor
;
4551 /* scale ms to jiffies */
4552 interval
= msecs_to_jiffies(interval
);
4553 if (unlikely(!interval
))
4555 if (interval
> HZ
*NR_CPUS
/10)
4556 interval
= HZ
*NR_CPUS
/10;
4558 need_serialize
= sd
->flags
& SD_SERIALIZE
;
4560 if (need_serialize
) {
4561 if (!spin_trylock(&balancing
))
4565 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4566 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
4568 * We've pulled tasks over so either we're no
4569 * longer idle, or one of our SMT siblings is
4572 idle
= CPU_NOT_IDLE
;
4574 sd
->last_balance
= jiffies
;
4577 spin_unlock(&balancing
);
4579 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4580 next_balance
= sd
->last_balance
+ interval
;
4581 update_next_balance
= 1;
4585 * Stop the load balance at this level. There is another
4586 * CPU in our sched group which is doing load balancing more
4594 * next_balance will be updated only when there is a need.
4595 * When the cpu is attached to null domain for ex, it will not be
4598 if (likely(update_next_balance
))
4599 rq
->next_balance
= next_balance
;
4603 * run_rebalance_domains is triggered when needed from the scheduler tick.
4604 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4605 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4607 static void run_rebalance_domains(struct softirq_action
*h
)
4609 int this_cpu
= smp_processor_id();
4610 struct rq
*this_rq
= cpu_rq(this_cpu
);
4611 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4612 CPU_IDLE
: CPU_NOT_IDLE
;
4614 rebalance_domains(this_cpu
, idle
);
4618 * If this cpu is the owner for idle load balancing, then do the
4619 * balancing on behalf of the other idle cpus whose ticks are
4622 if (this_rq
->idle_at_tick
&&
4623 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4627 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
4628 if (balance_cpu
== this_cpu
)
4632 * If this cpu gets work to do, stop the load balancing
4633 * work being done for other cpus. Next load
4634 * balancing owner will pick it up.
4639 rebalance_domains(balance_cpu
, CPU_IDLE
);
4641 rq
= cpu_rq(balance_cpu
);
4642 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4643 this_rq
->next_balance
= rq
->next_balance
;
4649 static inline int on_null_domain(int cpu
)
4651 return !rcu_dereference(cpu_rq(cpu
)->sd
);
4655 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4657 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4658 * idle load balancing owner or decide to stop the periodic load balancing,
4659 * if the whole system is idle.
4661 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4665 * If we were in the nohz mode recently and busy at the current
4666 * scheduler tick, then check if we need to nominate new idle
4669 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4670 rq
->in_nohz_recently
= 0;
4672 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4673 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4674 atomic_set(&nohz
.load_balancer
, -1);
4677 if (atomic_read(&nohz
.load_balancer
) == -1) {
4678 int ilb
= find_new_ilb(cpu
);
4680 if (ilb
< nr_cpu_ids
)
4686 * If this cpu is idle and doing idle load balancing for all the
4687 * cpus with ticks stopped, is it time for that to stop?
4689 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4690 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4696 * If this cpu is idle and the idle load balancing is done by
4697 * someone else, then no need raise the SCHED_SOFTIRQ
4699 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4700 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4703 /* Don't need to rebalance while attached to NULL domain */
4704 if (time_after_eq(jiffies
, rq
->next_balance
) &&
4705 likely(!on_null_domain(cpu
)))
4706 raise_softirq(SCHED_SOFTIRQ
);
4709 #else /* CONFIG_SMP */
4712 * on UP we do not need to balance between CPUs:
4714 static inline void idle_balance(int cpu
, struct rq
*rq
)
4720 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4722 EXPORT_PER_CPU_SYMBOL(kstat
);
4725 * Return any ns on the sched_clock that have not yet been accounted in
4726 * @p in case that task is currently running.
4728 * Called with task_rq_lock() held on @rq.
4730 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
4734 if (task_current(rq
, p
)) {
4735 update_rq_clock(rq
);
4736 ns
= rq
->clock
- p
->se
.exec_start
;
4744 unsigned long long task_delta_exec(struct task_struct
*p
)
4746 unsigned long flags
;
4750 rq
= task_rq_lock(p
, &flags
);
4751 ns
= do_task_delta_exec(p
, rq
);
4752 task_rq_unlock(rq
, &flags
);
4758 * Return accounted runtime for the task.
4759 * In case the task is currently running, return the runtime plus current's
4760 * pending runtime that have not been accounted yet.
4762 unsigned long long task_sched_runtime(struct task_struct
*p
)
4764 unsigned long flags
;
4768 rq
= task_rq_lock(p
, &flags
);
4769 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4770 task_rq_unlock(rq
, &flags
);
4776 * Return sum_exec_runtime for the thread group.
4777 * In case the task is currently running, return the sum plus current's
4778 * pending runtime that have not been accounted yet.
4780 * Note that the thread group might have other running tasks as well,
4781 * so the return value not includes other pending runtime that other
4782 * running tasks might have.
4784 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
4786 struct task_cputime totals
;
4787 unsigned long flags
;
4791 rq
= task_rq_lock(p
, &flags
);
4792 thread_group_cputime(p
, &totals
);
4793 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4794 task_rq_unlock(rq
, &flags
);
4800 * Account user cpu time to a process.
4801 * @p: the process that the cpu time gets accounted to
4802 * @cputime: the cpu time spent in user space since the last update
4803 * @cputime_scaled: cputime scaled by cpu frequency
4805 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
4806 cputime_t cputime_scaled
)
4808 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4811 /* Add user time to process. */
4812 p
->utime
= cputime_add(p
->utime
, cputime
);
4813 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4814 account_group_user_time(p
, cputime
);
4816 /* Add user time to cpustat. */
4817 tmp
= cputime_to_cputime64(cputime
);
4818 if (TASK_NICE(p
) > 0)
4819 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4821 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4823 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
4824 /* Account for user time used */
4825 acct_update_integrals(p
);
4829 * Account guest cpu time to a process.
4830 * @p: the process that the cpu time gets accounted to
4831 * @cputime: the cpu time spent in virtual machine since the last update
4832 * @cputime_scaled: cputime scaled by cpu frequency
4834 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
4835 cputime_t cputime_scaled
)
4838 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4840 tmp
= cputime_to_cputime64(cputime
);
4842 /* Add guest time to process. */
4843 p
->utime
= cputime_add(p
->utime
, cputime
);
4844 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4845 account_group_user_time(p
, cputime
);
4846 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4848 /* Add guest time to cpustat. */
4849 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4850 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4854 * Account system cpu time to a process.
4855 * @p: the process that the cpu time gets accounted to
4856 * @hardirq_offset: the offset to subtract from hardirq_count()
4857 * @cputime: the cpu time spent in kernel space since the last update
4858 * @cputime_scaled: cputime scaled by cpu frequency
4860 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4861 cputime_t cputime
, cputime_t cputime_scaled
)
4863 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4866 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4867 account_guest_time(p
, cputime
, cputime_scaled
);
4871 /* Add system time to process. */
4872 p
->stime
= cputime_add(p
->stime
, cputime
);
4873 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
4874 account_group_system_time(p
, cputime
);
4876 /* Add system time to cpustat. */
4877 tmp
= cputime_to_cputime64(cputime
);
4878 if (hardirq_count() - hardirq_offset
)
4879 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4880 else if (softirq_count())
4881 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4883 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4885 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
4887 /* Account for system time used */
4888 acct_update_integrals(p
);
4892 * Account for involuntary wait time.
4893 * @steal: the cpu time spent in involuntary wait
4895 void account_steal_time(cputime_t cputime
)
4897 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4898 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
4900 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
4904 * Account for idle time.
4905 * @cputime: the cpu time spent in idle wait
4907 void account_idle_time(cputime_t cputime
)
4909 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4910 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
4911 struct rq
*rq
= this_rq();
4913 if (atomic_read(&rq
->nr_iowait
) > 0)
4914 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
4916 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
4919 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4922 * Account a single tick of cpu time.
4923 * @p: the process that the cpu time gets accounted to
4924 * @user_tick: indicates if the tick is a user or a system tick
4926 void account_process_tick(struct task_struct
*p
, int user_tick
)
4928 cputime_t one_jiffy
= jiffies_to_cputime(1);
4929 cputime_t one_jiffy_scaled
= cputime_to_scaled(one_jiffy
);
4930 struct rq
*rq
= this_rq();
4933 account_user_time(p
, one_jiffy
, one_jiffy_scaled
);
4934 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
4935 account_system_time(p
, HARDIRQ_OFFSET
, one_jiffy
,
4938 account_idle_time(one_jiffy
);
4942 * Account multiple ticks of steal time.
4943 * @p: the process from which the cpu time has been stolen
4944 * @ticks: number of stolen ticks
4946 void account_steal_ticks(unsigned long ticks
)
4948 account_steal_time(jiffies_to_cputime(ticks
));
4952 * Account multiple ticks of idle time.
4953 * @ticks: number of stolen ticks
4955 void account_idle_ticks(unsigned long ticks
)
4957 account_idle_time(jiffies_to_cputime(ticks
));
4963 * Use precise platform statistics if available:
4965 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4966 cputime_t
task_utime(struct task_struct
*p
)
4971 cputime_t
task_stime(struct task_struct
*p
)
4976 cputime_t
task_utime(struct task_struct
*p
)
4978 clock_t utime
= cputime_to_clock_t(p
->utime
),
4979 total
= utime
+ cputime_to_clock_t(p
->stime
);
4983 * Use CFS's precise accounting:
4985 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
4989 do_div(temp
, total
);
4991 utime
= (clock_t)temp
;
4993 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
4994 return p
->prev_utime
;
4997 cputime_t
task_stime(struct task_struct
*p
)
5002 * Use CFS's precise accounting. (we subtract utime from
5003 * the total, to make sure the total observed by userspace
5004 * grows monotonically - apps rely on that):
5006 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
5007 cputime_to_clock_t(task_utime(p
));
5010 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
5012 return p
->prev_stime
;
5016 inline cputime_t
task_gtime(struct task_struct
*p
)
5022 * This function gets called by the timer code, with HZ frequency.
5023 * We call it with interrupts disabled.
5025 * It also gets called by the fork code, when changing the parent's
5028 void scheduler_tick(void)
5030 int cpu
= smp_processor_id();
5031 struct rq
*rq
= cpu_rq(cpu
);
5032 struct task_struct
*curr
= rq
->curr
;
5036 spin_lock(&rq
->lock
);
5037 update_rq_clock(rq
);
5038 update_cpu_load(rq
);
5039 curr
->sched_class
->task_tick(rq
, curr
, 0);
5040 spin_unlock(&rq
->lock
);
5043 rq
->idle_at_tick
= idle_cpu(cpu
);
5044 trigger_load_balance(rq
, cpu
);
5048 notrace
unsigned long get_parent_ip(unsigned long addr
)
5050 if (in_lock_functions(addr
)) {
5051 addr
= CALLER_ADDR2
;
5052 if (in_lock_functions(addr
))
5053 addr
= CALLER_ADDR3
;
5058 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5059 defined(CONFIG_PREEMPT_TRACER))
5061 void __kprobes
add_preempt_count(int val
)
5063 #ifdef CONFIG_DEBUG_PREEMPT
5067 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5070 preempt_count() += val
;
5071 #ifdef CONFIG_DEBUG_PREEMPT
5073 * Spinlock count overflowing soon?
5075 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
5078 if (preempt_count() == val
)
5079 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5081 EXPORT_SYMBOL(add_preempt_count
);
5083 void __kprobes
sub_preempt_count(int val
)
5085 #ifdef CONFIG_DEBUG_PREEMPT
5089 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
5092 * Is the spinlock portion underflowing?
5094 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
5095 !(preempt_count() & PREEMPT_MASK
)))
5099 if (preempt_count() == val
)
5100 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5101 preempt_count() -= val
;
5103 EXPORT_SYMBOL(sub_preempt_count
);
5108 * Print scheduling while atomic bug:
5110 static noinline
void __schedule_bug(struct task_struct
*prev
)
5112 struct pt_regs
*regs
= get_irq_regs();
5114 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
5115 prev
->comm
, prev
->pid
, preempt_count());
5117 debug_show_held_locks(prev
);
5119 if (irqs_disabled())
5120 print_irqtrace_events(prev
);
5129 * Various schedule()-time debugging checks and statistics:
5131 static inline void schedule_debug(struct task_struct
*prev
)
5134 * Test if we are atomic. Since do_exit() needs to call into
5135 * schedule() atomically, we ignore that path for now.
5136 * Otherwise, whine if we are scheduling when we should not be.
5138 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
5139 __schedule_bug(prev
);
5141 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
5143 schedstat_inc(this_rq(), sched_count
);
5144 #ifdef CONFIG_SCHEDSTATS
5145 if (unlikely(prev
->lock_depth
>= 0)) {
5146 schedstat_inc(this_rq(), bkl_count
);
5147 schedstat_inc(prev
, sched_info
.bkl_count
);
5152 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
5154 if (prev
->state
== TASK_RUNNING
) {
5155 u64 runtime
= prev
->se
.sum_exec_runtime
;
5157 runtime
-= prev
->se
.prev_sum_exec_runtime
;
5158 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
5161 * In order to avoid avg_overlap growing stale when we are
5162 * indeed overlapping and hence not getting put to sleep, grow
5163 * the avg_overlap on preemption.
5165 * We use the average preemption runtime because that
5166 * correlates to the amount of cache footprint a task can
5169 update_avg(&prev
->se
.avg_overlap
, runtime
);
5171 prev
->sched_class
->put_prev_task(rq
, prev
);
5175 * Pick up the highest-prio task:
5177 static inline struct task_struct
*
5178 pick_next_task(struct rq
*rq
)
5180 const struct sched_class
*class;
5181 struct task_struct
*p
;
5184 * Optimization: we know that if all tasks are in
5185 * the fair class we can call that function directly:
5187 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
5188 p
= fair_sched_class
.pick_next_task(rq
);
5193 class = sched_class_highest
;
5195 p
= class->pick_next_task(rq
);
5199 * Will never be NULL as the idle class always
5200 * returns a non-NULL p:
5202 class = class->next
;
5207 * schedule() is the main scheduler function.
5209 asmlinkage
void __sched
schedule(void)
5211 struct task_struct
*prev
, *next
;
5212 unsigned long *switch_count
;
5218 cpu
= smp_processor_id();
5222 switch_count
= &prev
->nivcsw
;
5224 release_kernel_lock(prev
);
5225 need_resched_nonpreemptible
:
5227 schedule_debug(prev
);
5229 if (sched_feat(HRTICK
))
5232 spin_lock_irq(&rq
->lock
);
5233 update_rq_clock(rq
);
5234 clear_tsk_need_resched(prev
);
5236 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
5237 if (unlikely(signal_pending_state(prev
->state
, prev
)))
5238 prev
->state
= TASK_RUNNING
;
5240 deactivate_task(rq
, prev
, 1);
5241 switch_count
= &prev
->nvcsw
;
5245 if (prev
->sched_class
->pre_schedule
)
5246 prev
->sched_class
->pre_schedule(rq
, prev
);
5249 if (unlikely(!rq
->nr_running
))
5250 idle_balance(cpu
, rq
);
5252 put_prev_task(rq
, prev
);
5253 next
= pick_next_task(rq
);
5255 if (likely(prev
!= next
)) {
5256 sched_info_switch(prev
, next
);
5262 context_switch(rq
, prev
, next
); /* unlocks the rq */
5264 * the context switch might have flipped the stack from under
5265 * us, hence refresh the local variables.
5267 cpu
= smp_processor_id();
5270 spin_unlock_irq(&rq
->lock
);
5272 if (unlikely(reacquire_kernel_lock(current
) < 0))
5273 goto need_resched_nonpreemptible
;
5275 preempt_enable_no_resched();
5279 EXPORT_SYMBOL(schedule
);
5283 * Look out! "owner" is an entirely speculative pointer
5284 * access and not reliable.
5286 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
5291 if (!sched_feat(OWNER_SPIN
))
5294 #ifdef CONFIG_DEBUG_PAGEALLOC
5296 * Need to access the cpu field knowing that
5297 * DEBUG_PAGEALLOC could have unmapped it if
5298 * the mutex owner just released it and exited.
5300 if (probe_kernel_address(&owner
->cpu
, cpu
))
5307 * Even if the access succeeded (likely case),
5308 * the cpu field may no longer be valid.
5310 if (cpu
>= nr_cpumask_bits
)
5314 * We need to validate that we can do a
5315 * get_cpu() and that we have the percpu area.
5317 if (!cpu_online(cpu
))
5324 * Owner changed, break to re-assess state.
5326 if (lock
->owner
!= owner
)
5330 * Is that owner really running on that cpu?
5332 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
5342 #ifdef CONFIG_PREEMPT
5344 * this is the entry point to schedule() from in-kernel preemption
5345 * off of preempt_enable. Kernel preemptions off return from interrupt
5346 * occur there and call schedule directly.
5348 asmlinkage
void __sched
preempt_schedule(void)
5350 struct thread_info
*ti
= current_thread_info();
5353 * If there is a non-zero preempt_count or interrupts are disabled,
5354 * we do not want to preempt the current task. Just return..
5356 if (likely(ti
->preempt_count
|| irqs_disabled()))
5360 add_preempt_count(PREEMPT_ACTIVE
);
5362 sub_preempt_count(PREEMPT_ACTIVE
);
5365 * Check again in case we missed a preemption opportunity
5366 * between schedule and now.
5369 } while (need_resched());
5371 EXPORT_SYMBOL(preempt_schedule
);
5374 * this is the entry point to schedule() from kernel preemption
5375 * off of irq context.
5376 * Note, that this is called and return with irqs disabled. This will
5377 * protect us against recursive calling from irq.
5379 asmlinkage
void __sched
preempt_schedule_irq(void)
5381 struct thread_info
*ti
= current_thread_info();
5383 /* Catch callers which need to be fixed */
5384 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
5387 add_preempt_count(PREEMPT_ACTIVE
);
5390 local_irq_disable();
5391 sub_preempt_count(PREEMPT_ACTIVE
);
5394 * Check again in case we missed a preemption opportunity
5395 * between schedule and now.
5398 } while (need_resched());
5401 #endif /* CONFIG_PREEMPT */
5403 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
5406 return try_to_wake_up(curr
->private, mode
, sync
);
5408 EXPORT_SYMBOL(default_wake_function
);
5411 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5412 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5413 * number) then we wake all the non-exclusive tasks and one exclusive task.
5415 * There are circumstances in which we can try to wake a task which has already
5416 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5417 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5419 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
5420 int nr_exclusive
, int sync
, void *key
)
5422 wait_queue_t
*curr
, *next
;
5424 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
5425 unsigned flags
= curr
->flags
;
5427 if (curr
->func(curr
, mode
, sync
, key
) &&
5428 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
5434 * __wake_up - wake up threads blocked on a waitqueue.
5436 * @mode: which threads
5437 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5438 * @key: is directly passed to the wakeup function
5440 * It may be assumed that this function implies a write memory barrier before
5441 * changing the task state if and only if any tasks are woken up.
5443 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
5444 int nr_exclusive
, void *key
)
5446 unsigned long flags
;
5448 spin_lock_irqsave(&q
->lock
, flags
);
5449 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
5450 spin_unlock_irqrestore(&q
->lock
, flags
);
5452 EXPORT_SYMBOL(__wake_up
);
5455 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5457 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
5459 __wake_up_common(q
, mode
, 1, 0, NULL
);
5462 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
5464 __wake_up_common(q
, mode
, 1, 0, key
);
5468 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5470 * @mode: which threads
5471 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5472 * @key: opaque value to be passed to wakeup targets
5474 * The sync wakeup differs that the waker knows that it will schedule
5475 * away soon, so while the target thread will be woken up, it will not
5476 * be migrated to another CPU - ie. the two threads are 'synchronized'
5477 * with each other. This can prevent needless bouncing between CPUs.
5479 * On UP it can prevent extra preemption.
5481 * It may be assumed that this function implies a write memory barrier before
5482 * changing the task state if and only if any tasks are woken up.
5484 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
5485 int nr_exclusive
, void *key
)
5487 unsigned long flags
;
5493 if (unlikely(!nr_exclusive
))
5496 spin_lock_irqsave(&q
->lock
, flags
);
5497 __wake_up_common(q
, mode
, nr_exclusive
, sync
, key
);
5498 spin_unlock_irqrestore(&q
->lock
, flags
);
5500 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
5503 * __wake_up_sync - see __wake_up_sync_key()
5505 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
5507 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
5509 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
5512 * complete: - signals a single thread waiting on this completion
5513 * @x: holds the state of this particular completion
5515 * This will wake up a single thread waiting on this completion. Threads will be
5516 * awakened in the same order in which they were queued.
5518 * See also complete_all(), wait_for_completion() and related routines.
5520 * It may be assumed that this function implies a write memory barrier before
5521 * changing the task state if and only if any tasks are woken up.
5523 void complete(struct completion
*x
)
5525 unsigned long flags
;
5527 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5529 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
5530 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5532 EXPORT_SYMBOL(complete
);
5535 * complete_all: - signals all threads waiting on this completion
5536 * @x: holds the state of this particular completion
5538 * This will wake up all threads waiting on this particular completion event.
5540 * It may be assumed that this function implies a write memory barrier before
5541 * changing the task state if and only if any tasks are woken up.
5543 void complete_all(struct completion
*x
)
5545 unsigned long flags
;
5547 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5548 x
->done
+= UINT_MAX
/2;
5549 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
5550 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5552 EXPORT_SYMBOL(complete_all
);
5554 static inline long __sched
5555 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
5558 DECLARE_WAITQUEUE(wait
, current
);
5560 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
5561 __add_wait_queue_tail(&x
->wait
, &wait
);
5563 if (signal_pending_state(state
, current
)) {
5564 timeout
= -ERESTARTSYS
;
5567 __set_current_state(state
);
5568 spin_unlock_irq(&x
->wait
.lock
);
5569 timeout
= schedule_timeout(timeout
);
5570 spin_lock_irq(&x
->wait
.lock
);
5571 } while (!x
->done
&& timeout
);
5572 __remove_wait_queue(&x
->wait
, &wait
);
5577 return timeout
?: 1;
5581 wait_for_common(struct completion
*x
, long timeout
, int state
)
5585 spin_lock_irq(&x
->wait
.lock
);
5586 timeout
= do_wait_for_common(x
, timeout
, state
);
5587 spin_unlock_irq(&x
->wait
.lock
);
5592 * wait_for_completion: - waits for completion of a task
5593 * @x: holds the state of this particular completion
5595 * This waits to be signaled for completion of a specific task. It is NOT
5596 * interruptible and there is no timeout.
5598 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5599 * and interrupt capability. Also see complete().
5601 void __sched
wait_for_completion(struct completion
*x
)
5603 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
5605 EXPORT_SYMBOL(wait_for_completion
);
5608 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5609 * @x: holds the state of this particular completion
5610 * @timeout: timeout value in jiffies
5612 * This waits for either a completion of a specific task to be signaled or for a
5613 * specified timeout to expire. The timeout is in jiffies. It is not
5616 unsigned long __sched
5617 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
5619 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
5621 EXPORT_SYMBOL(wait_for_completion_timeout
);
5624 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5625 * @x: holds the state of this particular completion
5627 * This waits for completion of a specific task to be signaled. It is
5630 int __sched
wait_for_completion_interruptible(struct completion
*x
)
5632 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
5633 if (t
== -ERESTARTSYS
)
5637 EXPORT_SYMBOL(wait_for_completion_interruptible
);
5640 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5641 * @x: holds the state of this particular completion
5642 * @timeout: timeout value in jiffies
5644 * This waits for either a completion of a specific task to be signaled or for a
5645 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5647 unsigned long __sched
5648 wait_for_completion_interruptible_timeout(struct completion
*x
,
5649 unsigned long timeout
)
5651 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
5653 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
5656 * wait_for_completion_killable: - waits for completion of a task (killable)
5657 * @x: holds the state of this particular completion
5659 * This waits to be signaled for completion of a specific task. It can be
5660 * interrupted by a kill signal.
5662 int __sched
wait_for_completion_killable(struct completion
*x
)
5664 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
5665 if (t
== -ERESTARTSYS
)
5669 EXPORT_SYMBOL(wait_for_completion_killable
);
5672 * try_wait_for_completion - try to decrement a completion without blocking
5673 * @x: completion structure
5675 * Returns: 0 if a decrement cannot be done without blocking
5676 * 1 if a decrement succeeded.
5678 * If a completion is being used as a counting completion,
5679 * attempt to decrement the counter without blocking. This
5680 * enables us to avoid waiting if the resource the completion
5681 * is protecting is not available.
5683 bool try_wait_for_completion(struct completion
*x
)
5687 spin_lock_irq(&x
->wait
.lock
);
5692 spin_unlock_irq(&x
->wait
.lock
);
5695 EXPORT_SYMBOL(try_wait_for_completion
);
5698 * completion_done - Test to see if a completion has any waiters
5699 * @x: completion structure
5701 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5702 * 1 if there are no waiters.
5705 bool completion_done(struct completion
*x
)
5709 spin_lock_irq(&x
->wait
.lock
);
5712 spin_unlock_irq(&x
->wait
.lock
);
5715 EXPORT_SYMBOL(completion_done
);
5718 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
5720 unsigned long flags
;
5723 init_waitqueue_entry(&wait
, current
);
5725 __set_current_state(state
);
5727 spin_lock_irqsave(&q
->lock
, flags
);
5728 __add_wait_queue(q
, &wait
);
5729 spin_unlock(&q
->lock
);
5730 timeout
= schedule_timeout(timeout
);
5731 spin_lock_irq(&q
->lock
);
5732 __remove_wait_queue(q
, &wait
);
5733 spin_unlock_irqrestore(&q
->lock
, flags
);
5738 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
5740 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5742 EXPORT_SYMBOL(interruptible_sleep_on
);
5745 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5747 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
5749 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
5751 void __sched
sleep_on(wait_queue_head_t
*q
)
5753 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5755 EXPORT_SYMBOL(sleep_on
);
5757 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5759 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
5761 EXPORT_SYMBOL(sleep_on_timeout
);
5763 #ifdef CONFIG_RT_MUTEXES
5766 * rt_mutex_setprio - set the current priority of a task
5768 * @prio: prio value (kernel-internal form)
5770 * This function changes the 'effective' priority of a task. It does
5771 * not touch ->normal_prio like __setscheduler().
5773 * Used by the rt_mutex code to implement priority inheritance logic.
5775 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
5777 unsigned long flags
;
5778 int oldprio
, on_rq
, running
;
5780 const struct sched_class
*prev_class
= p
->sched_class
;
5782 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
5784 rq
= task_rq_lock(p
, &flags
);
5785 update_rq_clock(rq
);
5788 on_rq
= p
->se
.on_rq
;
5789 running
= task_current(rq
, p
);
5791 dequeue_task(rq
, p
, 0);
5793 p
->sched_class
->put_prev_task(rq
, p
);
5796 p
->sched_class
= &rt_sched_class
;
5798 p
->sched_class
= &fair_sched_class
;
5803 p
->sched_class
->set_curr_task(rq
);
5805 enqueue_task(rq
, p
, 0);
5807 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5809 task_rq_unlock(rq
, &flags
);
5814 void set_user_nice(struct task_struct
*p
, long nice
)
5816 int old_prio
, delta
, on_rq
;
5817 unsigned long flags
;
5820 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
5823 * We have to be careful, if called from sys_setpriority(),
5824 * the task might be in the middle of scheduling on another CPU.
5826 rq
= task_rq_lock(p
, &flags
);
5827 update_rq_clock(rq
);
5829 * The RT priorities are set via sched_setscheduler(), but we still
5830 * allow the 'normal' nice value to be set - but as expected
5831 * it wont have any effect on scheduling until the task is
5832 * SCHED_FIFO/SCHED_RR:
5834 if (task_has_rt_policy(p
)) {
5835 p
->static_prio
= NICE_TO_PRIO(nice
);
5838 on_rq
= p
->se
.on_rq
;
5840 dequeue_task(rq
, p
, 0);
5842 p
->static_prio
= NICE_TO_PRIO(nice
);
5845 p
->prio
= effective_prio(p
);
5846 delta
= p
->prio
- old_prio
;
5849 enqueue_task(rq
, p
, 0);
5851 * If the task increased its priority or is running and
5852 * lowered its priority, then reschedule its CPU:
5854 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
5855 resched_task(rq
->curr
);
5858 task_rq_unlock(rq
, &flags
);
5860 EXPORT_SYMBOL(set_user_nice
);
5863 * can_nice - check if a task can reduce its nice value
5867 int can_nice(const struct task_struct
*p
, const int nice
)
5869 /* convert nice value [19,-20] to rlimit style value [1,40] */
5870 int nice_rlim
= 20 - nice
;
5872 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
5873 capable(CAP_SYS_NICE
));
5876 #ifdef __ARCH_WANT_SYS_NICE
5879 * sys_nice - change the priority of the current process.
5880 * @increment: priority increment
5882 * sys_setpriority is a more generic, but much slower function that
5883 * does similar things.
5885 SYSCALL_DEFINE1(nice
, int, increment
)
5890 * Setpriority might change our priority at the same moment.
5891 * We don't have to worry. Conceptually one call occurs first
5892 * and we have a single winner.
5894 if (increment
< -40)
5899 nice
= TASK_NICE(current
) + increment
;
5905 if (increment
< 0 && !can_nice(current
, nice
))
5908 retval
= security_task_setnice(current
, nice
);
5912 set_user_nice(current
, nice
);
5919 * task_prio - return the priority value of a given task.
5920 * @p: the task in question.
5922 * This is the priority value as seen by users in /proc.
5923 * RT tasks are offset by -200. Normal tasks are centered
5924 * around 0, value goes from -16 to +15.
5926 int task_prio(const struct task_struct
*p
)
5928 return p
->prio
- MAX_RT_PRIO
;
5932 * task_nice - return the nice value of a given task.
5933 * @p: the task in question.
5935 int task_nice(const struct task_struct
*p
)
5937 return TASK_NICE(p
);
5939 EXPORT_SYMBOL(task_nice
);
5942 * idle_cpu - is a given cpu idle currently?
5943 * @cpu: the processor in question.
5945 int idle_cpu(int cpu
)
5947 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
5951 * idle_task - return the idle task for a given cpu.
5952 * @cpu: the processor in question.
5954 struct task_struct
*idle_task(int cpu
)
5956 return cpu_rq(cpu
)->idle
;
5960 * find_process_by_pid - find a process with a matching PID value.
5961 * @pid: the pid in question.
5963 static struct task_struct
*find_process_by_pid(pid_t pid
)
5965 return pid
? find_task_by_vpid(pid
) : current
;
5968 /* Actually do priority change: must hold rq lock. */
5970 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
5972 BUG_ON(p
->se
.on_rq
);
5975 switch (p
->policy
) {
5979 p
->sched_class
= &fair_sched_class
;
5983 p
->sched_class
= &rt_sched_class
;
5987 p
->rt_priority
= prio
;
5988 p
->normal_prio
= normal_prio(p
);
5989 /* we are holding p->pi_lock already */
5990 p
->prio
= rt_mutex_getprio(p
);
5995 * check the target process has a UID that matches the current process's
5997 static bool check_same_owner(struct task_struct
*p
)
5999 const struct cred
*cred
= current_cred(), *pcred
;
6003 pcred
= __task_cred(p
);
6004 match
= (cred
->euid
== pcred
->euid
||
6005 cred
->euid
== pcred
->uid
);
6010 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
6011 struct sched_param
*param
, bool user
)
6013 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
6014 unsigned long flags
;
6015 const struct sched_class
*prev_class
= p
->sched_class
;
6018 /* may grab non-irq protected spin_locks */
6019 BUG_ON(in_interrupt());
6021 /* double check policy once rq lock held */
6023 policy
= oldpolicy
= p
->policy
;
6024 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
6025 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
6026 policy
!= SCHED_IDLE
)
6029 * Valid priorities for SCHED_FIFO and SCHED_RR are
6030 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6031 * SCHED_BATCH and SCHED_IDLE is 0.
6033 if (param
->sched_priority
< 0 ||
6034 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
6035 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
6037 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
6041 * Allow unprivileged RT tasks to decrease priority:
6043 if (user
&& !capable(CAP_SYS_NICE
)) {
6044 if (rt_policy(policy
)) {
6045 unsigned long rlim_rtprio
;
6047 if (!lock_task_sighand(p
, &flags
))
6049 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
6050 unlock_task_sighand(p
, &flags
);
6052 /* can't set/change the rt policy */
6053 if (policy
!= p
->policy
&& !rlim_rtprio
)
6056 /* can't increase priority */
6057 if (param
->sched_priority
> p
->rt_priority
&&
6058 param
->sched_priority
> rlim_rtprio
)
6062 * Like positive nice levels, dont allow tasks to
6063 * move out of SCHED_IDLE either:
6065 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
6068 /* can't change other user's priorities */
6069 if (!check_same_owner(p
))
6074 #ifdef CONFIG_RT_GROUP_SCHED
6076 * Do not allow realtime tasks into groups that have no runtime
6079 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
6080 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
6084 retval
= security_task_setscheduler(p
, policy
, param
);
6090 * make sure no PI-waiters arrive (or leave) while we are
6091 * changing the priority of the task:
6093 spin_lock_irqsave(&p
->pi_lock
, flags
);
6095 * To be able to change p->policy safely, the apropriate
6096 * runqueue lock must be held.
6098 rq
= __task_rq_lock(p
);
6099 /* recheck policy now with rq lock held */
6100 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
6101 policy
= oldpolicy
= -1;
6102 __task_rq_unlock(rq
);
6103 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6106 update_rq_clock(rq
);
6107 on_rq
= p
->se
.on_rq
;
6108 running
= task_current(rq
, p
);
6110 deactivate_task(rq
, p
, 0);
6112 p
->sched_class
->put_prev_task(rq
, p
);
6115 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
6118 p
->sched_class
->set_curr_task(rq
);
6120 activate_task(rq
, p
, 0);
6122 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
6124 __task_rq_unlock(rq
);
6125 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6127 rt_mutex_adjust_pi(p
);
6133 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6134 * @p: the task in question.
6135 * @policy: new policy.
6136 * @param: structure containing the new RT priority.
6138 * NOTE that the task may be already dead.
6140 int sched_setscheduler(struct task_struct
*p
, int policy
,
6141 struct sched_param
*param
)
6143 return __sched_setscheduler(p
, policy
, param
, true);
6145 EXPORT_SYMBOL_GPL(sched_setscheduler
);
6148 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6149 * @p: the task in question.
6150 * @policy: new policy.
6151 * @param: structure containing the new RT priority.
6153 * Just like sched_setscheduler, only don't bother checking if the
6154 * current context has permission. For example, this is needed in
6155 * stop_machine(): we create temporary high priority worker threads,
6156 * but our caller might not have that capability.
6158 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
6159 struct sched_param
*param
)
6161 return __sched_setscheduler(p
, policy
, param
, false);
6165 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
6167 struct sched_param lparam
;
6168 struct task_struct
*p
;
6171 if (!param
|| pid
< 0)
6173 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
6178 p
= find_process_by_pid(pid
);
6180 retval
= sched_setscheduler(p
, policy
, &lparam
);
6187 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6188 * @pid: the pid in question.
6189 * @policy: new policy.
6190 * @param: structure containing the new RT priority.
6192 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
6193 struct sched_param __user
*, param
)
6195 /* negative values for policy are not valid */
6199 return do_sched_setscheduler(pid
, policy
, param
);
6203 * sys_sched_setparam - set/change the RT priority of a thread
6204 * @pid: the pid in question.
6205 * @param: structure containing the new RT priority.
6207 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6209 return do_sched_setscheduler(pid
, -1, param
);
6213 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6214 * @pid: the pid in question.
6216 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
6218 struct task_struct
*p
;
6225 read_lock(&tasklist_lock
);
6226 p
= find_process_by_pid(pid
);
6228 retval
= security_task_getscheduler(p
);
6232 read_unlock(&tasklist_lock
);
6237 * sys_sched_getscheduler - get the RT priority of a thread
6238 * @pid: the pid in question.
6239 * @param: structure containing the RT priority.
6241 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6243 struct sched_param lp
;
6244 struct task_struct
*p
;
6247 if (!param
|| pid
< 0)
6250 read_lock(&tasklist_lock
);
6251 p
= find_process_by_pid(pid
);
6256 retval
= security_task_getscheduler(p
);
6260 lp
.sched_priority
= p
->rt_priority
;
6261 read_unlock(&tasklist_lock
);
6264 * This one might sleep, we cannot do it with a spinlock held ...
6266 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
6271 read_unlock(&tasklist_lock
);
6275 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
6277 cpumask_var_t cpus_allowed
, new_mask
;
6278 struct task_struct
*p
;
6282 read_lock(&tasklist_lock
);
6284 p
= find_process_by_pid(pid
);
6286 read_unlock(&tasklist_lock
);
6292 * It is not safe to call set_cpus_allowed with the
6293 * tasklist_lock held. We will bump the task_struct's
6294 * usage count and then drop tasklist_lock.
6297 read_unlock(&tasklist_lock
);
6299 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
6303 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
6305 goto out_free_cpus_allowed
;
6308 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
6311 retval
= security_task_setscheduler(p
, 0, NULL
);
6315 cpuset_cpus_allowed(p
, cpus_allowed
);
6316 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
6318 retval
= set_cpus_allowed_ptr(p
, new_mask
);
6321 cpuset_cpus_allowed(p
, cpus_allowed
);
6322 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
6324 * We must have raced with a concurrent cpuset
6325 * update. Just reset the cpus_allowed to the
6326 * cpuset's cpus_allowed
6328 cpumask_copy(new_mask
, cpus_allowed
);
6333 free_cpumask_var(new_mask
);
6334 out_free_cpus_allowed
:
6335 free_cpumask_var(cpus_allowed
);
6342 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
6343 struct cpumask
*new_mask
)
6345 if (len
< cpumask_size())
6346 cpumask_clear(new_mask
);
6347 else if (len
> cpumask_size())
6348 len
= cpumask_size();
6350 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
6354 * sys_sched_setaffinity - set the cpu affinity of a process
6355 * @pid: pid of the process
6356 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6357 * @user_mask_ptr: user-space pointer to the new cpu mask
6359 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
6360 unsigned long __user
*, user_mask_ptr
)
6362 cpumask_var_t new_mask
;
6365 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
6368 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
6370 retval
= sched_setaffinity(pid
, new_mask
);
6371 free_cpumask_var(new_mask
);
6375 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
6377 struct task_struct
*p
;
6381 read_lock(&tasklist_lock
);
6384 p
= find_process_by_pid(pid
);
6388 retval
= security_task_getscheduler(p
);
6392 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
6395 read_unlock(&tasklist_lock
);
6402 * sys_sched_getaffinity - get the cpu affinity of a process
6403 * @pid: pid of the process
6404 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6405 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6407 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
6408 unsigned long __user
*, user_mask_ptr
)
6413 if (len
< cpumask_size())
6416 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
6419 ret
= sched_getaffinity(pid
, mask
);
6421 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
6424 ret
= cpumask_size();
6426 free_cpumask_var(mask
);
6432 * sys_sched_yield - yield the current processor to other threads.
6434 * This function yields the current CPU to other tasks. If there are no
6435 * other threads running on this CPU then this function will return.
6437 SYSCALL_DEFINE0(sched_yield
)
6439 struct rq
*rq
= this_rq_lock();
6441 schedstat_inc(rq
, yld_count
);
6442 current
->sched_class
->yield_task(rq
);
6445 * Since we are going to call schedule() anyway, there's
6446 * no need to preempt or enable interrupts:
6448 __release(rq
->lock
);
6449 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
6450 _raw_spin_unlock(&rq
->lock
);
6451 preempt_enable_no_resched();
6458 static void __cond_resched(void)
6460 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6461 __might_sleep(__FILE__
, __LINE__
);
6464 * The BKS might be reacquired before we have dropped
6465 * PREEMPT_ACTIVE, which could trigger a second
6466 * cond_resched() call.
6469 add_preempt_count(PREEMPT_ACTIVE
);
6471 sub_preempt_count(PREEMPT_ACTIVE
);
6472 } while (need_resched());
6475 int __sched
_cond_resched(void)
6477 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
6478 system_state
== SYSTEM_RUNNING
) {
6484 EXPORT_SYMBOL(_cond_resched
);
6487 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
6488 * call schedule, and on return reacquire the lock.
6490 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6491 * operations here to prevent schedule() from being called twice (once via
6492 * spin_unlock(), once by hand).
6494 int cond_resched_lock(spinlock_t
*lock
)
6496 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
6499 if (spin_needbreak(lock
) || resched
) {
6501 if (resched
&& need_resched())
6510 EXPORT_SYMBOL(cond_resched_lock
);
6512 int __sched
cond_resched_softirq(void)
6514 BUG_ON(!in_softirq());
6516 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
6524 EXPORT_SYMBOL(cond_resched_softirq
);
6527 * yield - yield the current processor to other threads.
6529 * This is a shortcut for kernel-space yielding - it marks the
6530 * thread runnable and calls sys_sched_yield().
6532 void __sched
yield(void)
6534 set_current_state(TASK_RUNNING
);
6537 EXPORT_SYMBOL(yield
);
6540 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6541 * that process accounting knows that this is a task in IO wait state.
6543 * But don't do that if it is a deliberate, throttling IO wait (this task
6544 * has set its backing_dev_info: the queue against which it should throttle)
6546 void __sched
io_schedule(void)
6548 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
6550 delayacct_blkio_start();
6551 atomic_inc(&rq
->nr_iowait
);
6553 atomic_dec(&rq
->nr_iowait
);
6554 delayacct_blkio_end();
6556 EXPORT_SYMBOL(io_schedule
);
6558 long __sched
io_schedule_timeout(long timeout
)
6560 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
6563 delayacct_blkio_start();
6564 atomic_inc(&rq
->nr_iowait
);
6565 ret
= schedule_timeout(timeout
);
6566 atomic_dec(&rq
->nr_iowait
);
6567 delayacct_blkio_end();
6572 * sys_sched_get_priority_max - return maximum RT priority.
6573 * @policy: scheduling class.
6575 * this syscall returns the maximum rt_priority that can be used
6576 * by a given scheduling class.
6578 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
6585 ret
= MAX_USER_RT_PRIO
-1;
6597 * sys_sched_get_priority_min - return minimum RT priority.
6598 * @policy: scheduling class.
6600 * this syscall returns the minimum rt_priority that can be used
6601 * by a given scheduling class.
6603 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
6621 * sys_sched_rr_get_interval - return the default timeslice of a process.
6622 * @pid: pid of the process.
6623 * @interval: userspace pointer to the timeslice value.
6625 * this syscall writes the default timeslice value of a given process
6626 * into the user-space timespec buffer. A value of '0' means infinity.
6628 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
6629 struct timespec __user
*, interval
)
6631 struct task_struct
*p
;
6632 unsigned int time_slice
;
6640 read_lock(&tasklist_lock
);
6641 p
= find_process_by_pid(pid
);
6645 retval
= security_task_getscheduler(p
);
6650 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6651 * tasks that are on an otherwise idle runqueue:
6654 if (p
->policy
== SCHED_RR
) {
6655 time_slice
= DEF_TIMESLICE
;
6656 } else if (p
->policy
!= SCHED_FIFO
) {
6657 struct sched_entity
*se
= &p
->se
;
6658 unsigned long flags
;
6661 rq
= task_rq_lock(p
, &flags
);
6662 if (rq
->cfs
.load
.weight
)
6663 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
6664 task_rq_unlock(rq
, &flags
);
6666 read_unlock(&tasklist_lock
);
6667 jiffies_to_timespec(time_slice
, &t
);
6668 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
6672 read_unlock(&tasklist_lock
);
6676 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
6678 void sched_show_task(struct task_struct
*p
)
6680 unsigned long free
= 0;
6683 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
6684 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
6685 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
6686 #if BITS_PER_LONG == 32
6687 if (state
== TASK_RUNNING
)
6688 printk(KERN_CONT
" running ");
6690 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
6692 if (state
== TASK_RUNNING
)
6693 printk(KERN_CONT
" running task ");
6695 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
6697 #ifdef CONFIG_DEBUG_STACK_USAGE
6698 free
= stack_not_used(p
);
6700 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
6701 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
6702 (unsigned long)task_thread_info(p
)->flags
);
6704 show_stack(p
, NULL
);
6707 void show_state_filter(unsigned long state_filter
)
6709 struct task_struct
*g
, *p
;
6711 #if BITS_PER_LONG == 32
6713 " task PC stack pid father\n");
6716 " task PC stack pid father\n");
6718 read_lock(&tasklist_lock
);
6719 do_each_thread(g
, p
) {
6721 * reset the NMI-timeout, listing all files on a slow
6722 * console might take alot of time:
6724 touch_nmi_watchdog();
6725 if (!state_filter
|| (p
->state
& state_filter
))
6727 } while_each_thread(g
, p
);
6729 touch_all_softlockup_watchdogs();
6731 #ifdef CONFIG_SCHED_DEBUG
6732 sysrq_sched_debug_show();
6734 read_unlock(&tasklist_lock
);
6736 * Only show locks if all tasks are dumped:
6738 if (state_filter
== -1)
6739 debug_show_all_locks();
6742 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
6744 idle
->sched_class
= &idle_sched_class
;
6748 * init_idle - set up an idle thread for a given CPU
6749 * @idle: task in question
6750 * @cpu: cpu the idle task belongs to
6752 * NOTE: this function does not set the idle thread's NEED_RESCHED
6753 * flag, to make booting more robust.
6755 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
6757 struct rq
*rq
= cpu_rq(cpu
);
6758 unsigned long flags
;
6760 spin_lock_irqsave(&rq
->lock
, flags
);
6763 idle
->se
.exec_start
= sched_clock();
6765 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
6766 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
6767 __set_task_cpu(idle
, cpu
);
6769 rq
->curr
= rq
->idle
= idle
;
6770 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6773 spin_unlock_irqrestore(&rq
->lock
, flags
);
6775 /* Set the preempt count _outside_ the spinlocks! */
6776 #if defined(CONFIG_PREEMPT)
6777 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
6779 task_thread_info(idle
)->preempt_count
= 0;
6782 * The idle tasks have their own, simple scheduling class:
6784 idle
->sched_class
= &idle_sched_class
;
6785 ftrace_graph_init_task(idle
);
6789 * In a system that switches off the HZ timer nohz_cpu_mask
6790 * indicates which cpus entered this state. This is used
6791 * in the rcu update to wait only for active cpus. For system
6792 * which do not switch off the HZ timer nohz_cpu_mask should
6793 * always be CPU_BITS_NONE.
6795 cpumask_var_t nohz_cpu_mask
;
6798 * Increase the granularity value when there are more CPUs,
6799 * because with more CPUs the 'effective latency' as visible
6800 * to users decreases. But the relationship is not linear,
6801 * so pick a second-best guess by going with the log2 of the
6804 * This idea comes from the SD scheduler of Con Kolivas:
6806 static inline void sched_init_granularity(void)
6808 unsigned int factor
= 1 + ilog2(num_online_cpus());
6809 const unsigned long limit
= 200000000;
6811 sysctl_sched_min_granularity
*= factor
;
6812 if (sysctl_sched_min_granularity
> limit
)
6813 sysctl_sched_min_granularity
= limit
;
6815 sysctl_sched_latency
*= factor
;
6816 if (sysctl_sched_latency
> limit
)
6817 sysctl_sched_latency
= limit
;
6819 sysctl_sched_wakeup_granularity
*= factor
;
6821 sysctl_sched_shares_ratelimit
*= factor
;
6826 * This is how migration works:
6828 * 1) we queue a struct migration_req structure in the source CPU's
6829 * runqueue and wake up that CPU's migration thread.
6830 * 2) we down() the locked semaphore => thread blocks.
6831 * 3) migration thread wakes up (implicitly it forces the migrated
6832 * thread off the CPU)
6833 * 4) it gets the migration request and checks whether the migrated
6834 * task is still in the wrong runqueue.
6835 * 5) if it's in the wrong runqueue then the migration thread removes
6836 * it and puts it into the right queue.
6837 * 6) migration thread up()s the semaphore.
6838 * 7) we wake up and the migration is done.
6842 * Change a given task's CPU affinity. Migrate the thread to a
6843 * proper CPU and schedule it away if the CPU it's executing on
6844 * is removed from the allowed bitmask.
6846 * NOTE: the caller must have a valid reference to the task, the
6847 * task must not exit() & deallocate itself prematurely. The
6848 * call is not atomic; no spinlocks may be held.
6850 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
6852 struct migration_req req
;
6853 unsigned long flags
;
6857 rq
= task_rq_lock(p
, &flags
);
6858 if (!cpumask_intersects(new_mask
, cpu_online_mask
)) {
6863 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
6864 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
6869 if (p
->sched_class
->set_cpus_allowed
)
6870 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
6872 cpumask_copy(&p
->cpus_allowed
, new_mask
);
6873 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
6876 /* Can the task run on the task's current CPU? If so, we're done */
6877 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
6880 if (migrate_task(p
, cpumask_any_and(cpu_online_mask
, new_mask
), &req
)) {
6881 /* Need help from migration thread: drop lock and wait. */
6882 task_rq_unlock(rq
, &flags
);
6883 wake_up_process(rq
->migration_thread
);
6884 wait_for_completion(&req
.done
);
6885 tlb_migrate_finish(p
->mm
);
6889 task_rq_unlock(rq
, &flags
);
6893 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
6896 * Move (not current) task off this cpu, onto dest cpu. We're doing
6897 * this because either it can't run here any more (set_cpus_allowed()
6898 * away from this CPU, or CPU going down), or because we're
6899 * attempting to rebalance this task on exec (sched_exec).
6901 * So we race with normal scheduler movements, but that's OK, as long
6902 * as the task is no longer on this CPU.
6904 * Returns non-zero if task was successfully migrated.
6906 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6908 struct rq
*rq_dest
, *rq_src
;
6911 if (unlikely(!cpu_active(dest_cpu
)))
6914 rq_src
= cpu_rq(src_cpu
);
6915 rq_dest
= cpu_rq(dest_cpu
);
6917 double_rq_lock(rq_src
, rq_dest
);
6918 /* Already moved. */
6919 if (task_cpu(p
) != src_cpu
)
6921 /* Affinity changed (again). */
6922 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
6925 on_rq
= p
->se
.on_rq
;
6927 deactivate_task(rq_src
, p
, 0);
6929 set_task_cpu(p
, dest_cpu
);
6931 activate_task(rq_dest
, p
, 0);
6932 check_preempt_curr(rq_dest
, p
, 0);
6937 double_rq_unlock(rq_src
, rq_dest
);
6942 * migration_thread - this is a highprio system thread that performs
6943 * thread migration by bumping thread off CPU then 'pushing' onto
6946 static int migration_thread(void *data
)
6948 int cpu
= (long)data
;
6952 BUG_ON(rq
->migration_thread
!= current
);
6954 set_current_state(TASK_INTERRUPTIBLE
);
6955 while (!kthread_should_stop()) {
6956 struct migration_req
*req
;
6957 struct list_head
*head
;
6959 spin_lock_irq(&rq
->lock
);
6961 if (cpu_is_offline(cpu
)) {
6962 spin_unlock_irq(&rq
->lock
);
6966 if (rq
->active_balance
) {
6967 active_load_balance(rq
, cpu
);
6968 rq
->active_balance
= 0;
6971 head
= &rq
->migration_queue
;
6973 if (list_empty(head
)) {
6974 spin_unlock_irq(&rq
->lock
);
6976 set_current_state(TASK_INTERRUPTIBLE
);
6979 req
= list_entry(head
->next
, struct migration_req
, list
);
6980 list_del_init(head
->next
);
6982 spin_unlock(&rq
->lock
);
6983 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
6986 complete(&req
->done
);
6988 __set_current_state(TASK_RUNNING
);
6992 /* Wait for kthread_stop */
6993 set_current_state(TASK_INTERRUPTIBLE
);
6994 while (!kthread_should_stop()) {
6996 set_current_state(TASK_INTERRUPTIBLE
);
6998 __set_current_state(TASK_RUNNING
);
7002 #ifdef CONFIG_HOTPLUG_CPU
7004 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7008 local_irq_disable();
7009 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
7015 * Figure out where task on dead CPU should go, use force if necessary.
7017 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
7020 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(dead_cpu
));
7023 /* Look for allowed, online CPU in same node. */
7024 for_each_cpu_and(dest_cpu
, nodemask
, cpu_online_mask
)
7025 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7028 /* Any allowed, online CPU? */
7029 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_online_mask
);
7030 if (dest_cpu
< nr_cpu_ids
)
7033 /* No more Mr. Nice Guy. */
7034 if (dest_cpu
>= nr_cpu_ids
) {
7035 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
7036 dest_cpu
= cpumask_any_and(cpu_online_mask
, &p
->cpus_allowed
);
7039 * Don't tell them about moving exiting tasks or
7040 * kernel threads (both mm NULL), since they never
7043 if (p
->mm
&& printk_ratelimit()) {
7044 printk(KERN_INFO
"process %d (%s) no "
7045 "longer affine to cpu%d\n",
7046 task_pid_nr(p
), p
->comm
, dead_cpu
);
7051 /* It can have affinity changed while we were choosing. */
7052 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
7057 * While a dead CPU has no uninterruptible tasks queued at this point,
7058 * it might still have a nonzero ->nr_uninterruptible counter, because
7059 * for performance reasons the counter is not stricly tracking tasks to
7060 * their home CPUs. So we just add the counter to another CPU's counter,
7061 * to keep the global sum constant after CPU-down:
7063 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
7065 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_online_mask
));
7066 unsigned long flags
;
7068 local_irq_save(flags
);
7069 double_rq_lock(rq_src
, rq_dest
);
7070 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
7071 rq_src
->nr_uninterruptible
= 0;
7072 double_rq_unlock(rq_src
, rq_dest
);
7073 local_irq_restore(flags
);
7076 /* Run through task list and migrate tasks from the dead cpu. */
7077 static void migrate_live_tasks(int src_cpu
)
7079 struct task_struct
*p
, *t
;
7081 read_lock(&tasklist_lock
);
7083 do_each_thread(t
, p
) {
7087 if (task_cpu(p
) == src_cpu
)
7088 move_task_off_dead_cpu(src_cpu
, p
);
7089 } while_each_thread(t
, p
);
7091 read_unlock(&tasklist_lock
);
7095 * Schedules idle task to be the next runnable task on current CPU.
7096 * It does so by boosting its priority to highest possible.
7097 * Used by CPU offline code.
7099 void sched_idle_next(void)
7101 int this_cpu
= smp_processor_id();
7102 struct rq
*rq
= cpu_rq(this_cpu
);
7103 struct task_struct
*p
= rq
->idle
;
7104 unsigned long flags
;
7106 /* cpu has to be offline */
7107 BUG_ON(cpu_online(this_cpu
));
7110 * Strictly not necessary since rest of the CPUs are stopped by now
7111 * and interrupts disabled on the current cpu.
7113 spin_lock_irqsave(&rq
->lock
, flags
);
7115 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7117 update_rq_clock(rq
);
7118 activate_task(rq
, p
, 0);
7120 spin_unlock_irqrestore(&rq
->lock
, flags
);
7124 * Ensures that the idle task is using init_mm right before its cpu goes
7127 void idle_task_exit(void)
7129 struct mm_struct
*mm
= current
->active_mm
;
7131 BUG_ON(cpu_online(smp_processor_id()));
7134 switch_mm(mm
, &init_mm
, current
);
7138 /* called under rq->lock with disabled interrupts */
7139 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
7141 struct rq
*rq
= cpu_rq(dead_cpu
);
7143 /* Must be exiting, otherwise would be on tasklist. */
7144 BUG_ON(!p
->exit_state
);
7146 /* Cannot have done final schedule yet: would have vanished. */
7147 BUG_ON(p
->state
== TASK_DEAD
);
7152 * Drop lock around migration; if someone else moves it,
7153 * that's OK. No task can be added to this CPU, so iteration is
7156 spin_unlock_irq(&rq
->lock
);
7157 move_task_off_dead_cpu(dead_cpu
, p
);
7158 spin_lock_irq(&rq
->lock
);
7163 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7164 static void migrate_dead_tasks(unsigned int dead_cpu
)
7166 struct rq
*rq
= cpu_rq(dead_cpu
);
7167 struct task_struct
*next
;
7170 if (!rq
->nr_running
)
7172 update_rq_clock(rq
);
7173 next
= pick_next_task(rq
);
7176 next
->sched_class
->put_prev_task(rq
, next
);
7177 migrate_dead(dead_cpu
, next
);
7183 * remove the tasks which were accounted by rq from calc_load_tasks.
7185 static void calc_global_load_remove(struct rq
*rq
)
7187 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
7189 #endif /* CONFIG_HOTPLUG_CPU */
7191 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7193 static struct ctl_table sd_ctl_dir
[] = {
7195 .procname
= "sched_domain",
7201 static struct ctl_table sd_ctl_root
[] = {
7203 .ctl_name
= CTL_KERN
,
7204 .procname
= "kernel",
7206 .child
= sd_ctl_dir
,
7211 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
7213 struct ctl_table
*entry
=
7214 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
7219 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
7221 struct ctl_table
*entry
;
7224 * In the intermediate directories, both the child directory and
7225 * procname are dynamically allocated and could fail but the mode
7226 * will always be set. In the lowest directory the names are
7227 * static strings and all have proc handlers.
7229 for (entry
= *tablep
; entry
->mode
; entry
++) {
7231 sd_free_ctl_entry(&entry
->child
);
7232 if (entry
->proc_handler
== NULL
)
7233 kfree(entry
->procname
);
7241 set_table_entry(struct ctl_table
*entry
,
7242 const char *procname
, void *data
, int maxlen
,
7243 mode_t mode
, proc_handler
*proc_handler
)
7245 entry
->procname
= procname
;
7247 entry
->maxlen
= maxlen
;
7249 entry
->proc_handler
= proc_handler
;
7252 static struct ctl_table
*
7253 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
7255 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
7260 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
7261 sizeof(long), 0644, proc_doulongvec_minmax
);
7262 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
7263 sizeof(long), 0644, proc_doulongvec_minmax
);
7264 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
7265 sizeof(int), 0644, proc_dointvec_minmax
);
7266 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
7267 sizeof(int), 0644, proc_dointvec_minmax
);
7268 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
7269 sizeof(int), 0644, proc_dointvec_minmax
);
7270 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
7271 sizeof(int), 0644, proc_dointvec_minmax
);
7272 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
7273 sizeof(int), 0644, proc_dointvec_minmax
);
7274 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
7275 sizeof(int), 0644, proc_dointvec_minmax
);
7276 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
7277 sizeof(int), 0644, proc_dointvec_minmax
);
7278 set_table_entry(&table
[9], "cache_nice_tries",
7279 &sd
->cache_nice_tries
,
7280 sizeof(int), 0644, proc_dointvec_minmax
);
7281 set_table_entry(&table
[10], "flags", &sd
->flags
,
7282 sizeof(int), 0644, proc_dointvec_minmax
);
7283 set_table_entry(&table
[11], "name", sd
->name
,
7284 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
7285 /* &table[12] is terminator */
7290 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
7292 struct ctl_table
*entry
, *table
;
7293 struct sched_domain
*sd
;
7294 int domain_num
= 0, i
;
7297 for_each_domain(cpu
, sd
)
7299 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
7304 for_each_domain(cpu
, sd
) {
7305 snprintf(buf
, 32, "domain%d", i
);
7306 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7308 entry
->child
= sd_alloc_ctl_domain_table(sd
);
7315 static struct ctl_table_header
*sd_sysctl_header
;
7316 static void register_sched_domain_sysctl(void)
7318 int i
, cpu_num
= num_online_cpus();
7319 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
7322 WARN_ON(sd_ctl_dir
[0].child
);
7323 sd_ctl_dir
[0].child
= entry
;
7328 for_each_online_cpu(i
) {
7329 snprintf(buf
, 32, "cpu%d", i
);
7330 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7332 entry
->child
= sd_alloc_ctl_cpu_table(i
);
7336 WARN_ON(sd_sysctl_header
);
7337 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
7340 /* may be called multiple times per register */
7341 static void unregister_sched_domain_sysctl(void)
7343 if (sd_sysctl_header
)
7344 unregister_sysctl_table(sd_sysctl_header
);
7345 sd_sysctl_header
= NULL
;
7346 if (sd_ctl_dir
[0].child
)
7347 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
7350 static void register_sched_domain_sysctl(void)
7353 static void unregister_sched_domain_sysctl(void)
7358 static void set_rq_online(struct rq
*rq
)
7361 const struct sched_class
*class;
7363 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
7366 for_each_class(class) {
7367 if (class->rq_online
)
7368 class->rq_online(rq
);
7373 static void set_rq_offline(struct rq
*rq
)
7376 const struct sched_class
*class;
7378 for_each_class(class) {
7379 if (class->rq_offline
)
7380 class->rq_offline(rq
);
7383 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
7389 * migration_call - callback that gets triggered when a CPU is added.
7390 * Here we can start up the necessary migration thread for the new CPU.
7392 static int __cpuinit
7393 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
7395 struct task_struct
*p
;
7396 int cpu
= (long)hcpu
;
7397 unsigned long flags
;
7402 case CPU_UP_PREPARE
:
7403 case CPU_UP_PREPARE_FROZEN
:
7404 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
7407 kthread_bind(p
, cpu
);
7408 /* Must be high prio: stop_machine expects to yield to it. */
7409 rq
= task_rq_lock(p
, &flags
);
7410 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7411 task_rq_unlock(rq
, &flags
);
7412 cpu_rq(cpu
)->migration_thread
= p
;
7416 case CPU_ONLINE_FROZEN
:
7417 /* Strictly unnecessary, as first user will wake it. */
7418 wake_up_process(cpu_rq(cpu
)->migration_thread
);
7420 /* Update our root-domain */
7422 spin_lock_irqsave(&rq
->lock
, flags
);
7423 rq
->calc_load_update
= calc_load_update
;
7424 rq
->calc_load_active
= 0;
7426 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7430 spin_unlock_irqrestore(&rq
->lock
, flags
);
7433 #ifdef CONFIG_HOTPLUG_CPU
7434 case CPU_UP_CANCELED
:
7435 case CPU_UP_CANCELED_FROZEN
:
7436 if (!cpu_rq(cpu
)->migration_thread
)
7438 /* Unbind it from offline cpu so it can run. Fall thru. */
7439 kthread_bind(cpu_rq(cpu
)->migration_thread
,
7440 cpumask_any(cpu_online_mask
));
7441 kthread_stop(cpu_rq(cpu
)->migration_thread
);
7442 cpu_rq(cpu
)->migration_thread
= NULL
;
7446 case CPU_DEAD_FROZEN
:
7447 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7448 migrate_live_tasks(cpu
);
7450 kthread_stop(rq
->migration_thread
);
7451 rq
->migration_thread
= NULL
;
7452 /* Idle task back to normal (off runqueue, low prio) */
7453 spin_lock_irq(&rq
->lock
);
7454 update_rq_clock(rq
);
7455 deactivate_task(rq
, rq
->idle
, 0);
7456 rq
->idle
->static_prio
= MAX_PRIO
;
7457 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
7458 rq
->idle
->sched_class
= &idle_sched_class
;
7459 migrate_dead_tasks(cpu
);
7460 spin_unlock_irq(&rq
->lock
);
7462 migrate_nr_uninterruptible(rq
);
7463 BUG_ON(rq
->nr_running
!= 0);
7464 calc_global_load_remove(rq
);
7466 * No need to migrate the tasks: it was best-effort if
7467 * they didn't take sched_hotcpu_mutex. Just wake up
7470 spin_lock_irq(&rq
->lock
);
7471 while (!list_empty(&rq
->migration_queue
)) {
7472 struct migration_req
*req
;
7474 req
= list_entry(rq
->migration_queue
.next
,
7475 struct migration_req
, list
);
7476 list_del_init(&req
->list
);
7477 spin_unlock_irq(&rq
->lock
);
7478 complete(&req
->done
);
7479 spin_lock_irq(&rq
->lock
);
7481 spin_unlock_irq(&rq
->lock
);
7485 case CPU_DYING_FROZEN
:
7486 /* Update our root-domain */
7488 spin_lock_irqsave(&rq
->lock
, flags
);
7490 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7493 spin_unlock_irqrestore(&rq
->lock
, flags
);
7500 /* Register at highest priority so that task migration (migrate_all_tasks)
7501 * happens before everything else.
7503 static struct notifier_block __cpuinitdata migration_notifier
= {
7504 .notifier_call
= migration_call
,
7508 static int __init
migration_init(void)
7510 void *cpu
= (void *)(long)smp_processor_id();
7513 /* Start one for the boot CPU: */
7514 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
7515 BUG_ON(err
== NOTIFY_BAD
);
7516 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
7517 register_cpu_notifier(&migration_notifier
);
7521 early_initcall(migration_init
);
7526 #ifdef CONFIG_SCHED_DEBUG
7528 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
7529 struct cpumask
*groupmask
)
7531 struct sched_group
*group
= sd
->groups
;
7534 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
7535 cpumask_clear(groupmask
);
7537 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
7539 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
7540 printk("does not load-balance\n");
7542 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
7547 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
7549 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
7550 printk(KERN_ERR
"ERROR: domain->span does not contain "
7553 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
7554 printk(KERN_ERR
"ERROR: domain->groups does not contain"
7558 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
7562 printk(KERN_ERR
"ERROR: group is NULL\n");
7566 if (!group
->__cpu_power
) {
7567 printk(KERN_CONT
"\n");
7568 printk(KERN_ERR
"ERROR: domain->cpu_power not "
7573 if (!cpumask_weight(sched_group_cpus(group
))) {
7574 printk(KERN_CONT
"\n");
7575 printk(KERN_ERR
"ERROR: empty group\n");
7579 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
7580 printk(KERN_CONT
"\n");
7581 printk(KERN_ERR
"ERROR: repeated CPUs\n");
7585 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
7587 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
7589 printk(KERN_CONT
" %s", str
);
7590 if (group
->__cpu_power
!= SCHED_LOAD_SCALE
) {
7591 printk(KERN_CONT
" (__cpu_power = %d)",
7592 group
->__cpu_power
);
7595 group
= group
->next
;
7596 } while (group
!= sd
->groups
);
7597 printk(KERN_CONT
"\n");
7599 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
7600 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
7603 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
7604 printk(KERN_ERR
"ERROR: parent span is not a superset "
7605 "of domain->span\n");
7609 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
7611 cpumask_var_t groupmask
;
7615 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
7619 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
7621 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
7622 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
7627 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
7634 free_cpumask_var(groupmask
);
7636 #else /* !CONFIG_SCHED_DEBUG */
7637 # define sched_domain_debug(sd, cpu) do { } while (0)
7638 #endif /* CONFIG_SCHED_DEBUG */
7640 static int sd_degenerate(struct sched_domain
*sd
)
7642 if (cpumask_weight(sched_domain_span(sd
)) == 1)
7645 /* Following flags need at least 2 groups */
7646 if (sd
->flags
& (SD_LOAD_BALANCE
|
7647 SD_BALANCE_NEWIDLE
|
7651 SD_SHARE_PKG_RESOURCES
)) {
7652 if (sd
->groups
!= sd
->groups
->next
)
7656 /* Following flags don't use groups */
7657 if (sd
->flags
& (SD_WAKE_IDLE
|
7666 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
7668 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
7670 if (sd_degenerate(parent
))
7673 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
7676 /* Does parent contain flags not in child? */
7677 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7678 if (cflags
& SD_WAKE_AFFINE
)
7679 pflags
&= ~SD_WAKE_BALANCE
;
7680 /* Flags needing groups don't count if only 1 group in parent */
7681 if (parent
->groups
== parent
->groups
->next
) {
7682 pflags
&= ~(SD_LOAD_BALANCE
|
7683 SD_BALANCE_NEWIDLE
|
7687 SD_SHARE_PKG_RESOURCES
);
7688 if (nr_node_ids
== 1)
7689 pflags
&= ~SD_SERIALIZE
;
7691 if (~cflags
& pflags
)
7697 static void free_rootdomain(struct root_domain
*rd
)
7699 cpupri_cleanup(&rd
->cpupri
);
7701 free_cpumask_var(rd
->rto_mask
);
7702 free_cpumask_var(rd
->online
);
7703 free_cpumask_var(rd
->span
);
7707 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
7709 struct root_domain
*old_rd
= NULL
;
7710 unsigned long flags
;
7712 spin_lock_irqsave(&rq
->lock
, flags
);
7717 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
7720 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
7723 * If we dont want to free the old_rt yet then
7724 * set old_rd to NULL to skip the freeing later
7727 if (!atomic_dec_and_test(&old_rd
->refcount
))
7731 atomic_inc(&rd
->refcount
);
7734 cpumask_set_cpu(rq
->cpu
, rd
->span
);
7735 if (cpumask_test_cpu(rq
->cpu
, cpu_online_mask
))
7738 spin_unlock_irqrestore(&rq
->lock
, flags
);
7741 free_rootdomain(old_rd
);
7744 static int __init_refok
init_rootdomain(struct root_domain
*rd
, bool bootmem
)
7746 memset(rd
, 0, sizeof(*rd
));
7749 alloc_bootmem_cpumask_var(&def_root_domain
.span
);
7750 alloc_bootmem_cpumask_var(&def_root_domain
.online
);
7751 alloc_bootmem_cpumask_var(&def_root_domain
.rto_mask
);
7752 cpupri_init(&rd
->cpupri
, true);
7756 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
7758 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
7760 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
7763 if (cpupri_init(&rd
->cpupri
, false) != 0)
7768 free_cpumask_var(rd
->rto_mask
);
7770 free_cpumask_var(rd
->online
);
7772 free_cpumask_var(rd
->span
);
7777 static void init_defrootdomain(void)
7779 init_rootdomain(&def_root_domain
, true);
7781 atomic_set(&def_root_domain
.refcount
, 1);
7784 static struct root_domain
*alloc_rootdomain(void)
7786 struct root_domain
*rd
;
7788 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
7792 if (init_rootdomain(rd
, false) != 0) {
7801 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7802 * hold the hotplug lock.
7805 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
7807 struct rq
*rq
= cpu_rq(cpu
);
7808 struct sched_domain
*tmp
;
7810 /* Remove the sched domains which do not contribute to scheduling. */
7811 for (tmp
= sd
; tmp
; ) {
7812 struct sched_domain
*parent
= tmp
->parent
;
7816 if (sd_parent_degenerate(tmp
, parent
)) {
7817 tmp
->parent
= parent
->parent
;
7819 parent
->parent
->child
= tmp
;
7824 if (sd
&& sd_degenerate(sd
)) {
7830 sched_domain_debug(sd
, cpu
);
7832 rq_attach_root(rq
, rd
);
7833 rcu_assign_pointer(rq
->sd
, sd
);
7836 /* cpus with isolated domains */
7837 static cpumask_var_t cpu_isolated_map
;
7839 /* Setup the mask of cpus configured for isolated domains */
7840 static int __init
isolated_cpu_setup(char *str
)
7842 cpulist_parse(str
, cpu_isolated_map
);
7846 __setup("isolcpus=", isolated_cpu_setup
);
7849 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7850 * to a function which identifies what group(along with sched group) a CPU
7851 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7852 * (due to the fact that we keep track of groups covered with a struct cpumask).
7854 * init_sched_build_groups will build a circular linked list of the groups
7855 * covered by the given span, and will set each group's ->cpumask correctly,
7856 * and ->cpu_power to 0.
7859 init_sched_build_groups(const struct cpumask
*span
,
7860 const struct cpumask
*cpu_map
,
7861 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
7862 struct sched_group
**sg
,
7863 struct cpumask
*tmpmask
),
7864 struct cpumask
*covered
, struct cpumask
*tmpmask
)
7866 struct sched_group
*first
= NULL
, *last
= NULL
;
7869 cpumask_clear(covered
);
7871 for_each_cpu(i
, span
) {
7872 struct sched_group
*sg
;
7873 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
7876 if (cpumask_test_cpu(i
, covered
))
7879 cpumask_clear(sched_group_cpus(sg
));
7880 sg
->__cpu_power
= 0;
7882 for_each_cpu(j
, span
) {
7883 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
7886 cpumask_set_cpu(j
, covered
);
7887 cpumask_set_cpu(j
, sched_group_cpus(sg
));
7898 #define SD_NODES_PER_DOMAIN 16
7903 * find_next_best_node - find the next node to include in a sched_domain
7904 * @node: node whose sched_domain we're building
7905 * @used_nodes: nodes already in the sched_domain
7907 * Find the next node to include in a given scheduling domain. Simply
7908 * finds the closest node not already in the @used_nodes map.
7910 * Should use nodemask_t.
7912 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
7914 int i
, n
, val
, min_val
, best_node
= 0;
7918 for (i
= 0; i
< nr_node_ids
; i
++) {
7919 /* Start at @node */
7920 n
= (node
+ i
) % nr_node_ids
;
7922 if (!nr_cpus_node(n
))
7925 /* Skip already used nodes */
7926 if (node_isset(n
, *used_nodes
))
7929 /* Simple min distance search */
7930 val
= node_distance(node
, n
);
7932 if (val
< min_val
) {
7938 node_set(best_node
, *used_nodes
);
7943 * sched_domain_node_span - get a cpumask for a node's sched_domain
7944 * @node: node whose cpumask we're constructing
7945 * @span: resulting cpumask
7947 * Given a node, construct a good cpumask for its sched_domain to span. It
7948 * should be one that prevents unnecessary balancing, but also spreads tasks
7951 static void sched_domain_node_span(int node
, struct cpumask
*span
)
7953 nodemask_t used_nodes
;
7956 cpumask_clear(span
);
7957 nodes_clear(used_nodes
);
7959 cpumask_or(span
, span
, cpumask_of_node(node
));
7960 node_set(node
, used_nodes
);
7962 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
7963 int next_node
= find_next_best_node(node
, &used_nodes
);
7965 cpumask_or(span
, span
, cpumask_of_node(next_node
));
7968 #endif /* CONFIG_NUMA */
7970 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
7973 * The cpus mask in sched_group and sched_domain hangs off the end.
7975 * ( See the the comments in include/linux/sched.h:struct sched_group
7976 * and struct sched_domain. )
7978 struct static_sched_group
{
7979 struct sched_group sg
;
7980 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
7983 struct static_sched_domain
{
7984 struct sched_domain sd
;
7985 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
7989 * SMT sched-domains:
7991 #ifdef CONFIG_SCHED_SMT
7992 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
7993 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_cpus
);
7996 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
7997 struct sched_group
**sg
, struct cpumask
*unused
)
8000 *sg
= &per_cpu(sched_group_cpus
, cpu
).sg
;
8003 #endif /* CONFIG_SCHED_SMT */
8006 * multi-core sched-domains:
8008 #ifdef CONFIG_SCHED_MC
8009 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
8010 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
8011 #endif /* CONFIG_SCHED_MC */
8013 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8015 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8016 struct sched_group
**sg
, struct cpumask
*mask
)
8020 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8021 group
= cpumask_first(mask
);
8023 *sg
= &per_cpu(sched_group_core
, group
).sg
;
8026 #elif defined(CONFIG_SCHED_MC)
8028 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8029 struct sched_group
**sg
, struct cpumask
*unused
)
8032 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
8037 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
8038 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
8041 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
8042 struct sched_group
**sg
, struct cpumask
*mask
)
8045 #ifdef CONFIG_SCHED_MC
8046 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
8047 group
= cpumask_first(mask
);
8048 #elif defined(CONFIG_SCHED_SMT)
8049 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8050 group
= cpumask_first(mask
);
8055 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
8061 * The init_sched_build_groups can't handle what we want to do with node
8062 * groups, so roll our own. Now each node has its own list of groups which
8063 * gets dynamically allocated.
8065 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
8066 static struct sched_group
***sched_group_nodes_bycpu
;
8068 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
8069 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
8071 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
8072 struct sched_group
**sg
,
8073 struct cpumask
*nodemask
)
8077 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
8078 group
= cpumask_first(nodemask
);
8081 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
8085 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
8087 struct sched_group
*sg
= group_head
;
8093 for_each_cpu(j
, sched_group_cpus(sg
)) {
8094 struct sched_domain
*sd
;
8096 sd
= &per_cpu(phys_domains
, j
).sd
;
8097 if (j
!= group_first_cpu(sd
->groups
)) {
8099 * Only add "power" once for each
8105 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
8108 } while (sg
!= group_head
);
8110 #endif /* CONFIG_NUMA */
8113 /* Free memory allocated for various sched_group structures */
8114 static void free_sched_groups(const struct cpumask
*cpu_map
,
8115 struct cpumask
*nodemask
)
8119 for_each_cpu(cpu
, cpu_map
) {
8120 struct sched_group
**sched_group_nodes
8121 = sched_group_nodes_bycpu
[cpu
];
8123 if (!sched_group_nodes
)
8126 for (i
= 0; i
< nr_node_ids
; i
++) {
8127 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
8129 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8130 if (cpumask_empty(nodemask
))
8140 if (oldsg
!= sched_group_nodes
[i
])
8143 kfree(sched_group_nodes
);
8144 sched_group_nodes_bycpu
[cpu
] = NULL
;
8147 #else /* !CONFIG_NUMA */
8148 static void free_sched_groups(const struct cpumask
*cpu_map
,
8149 struct cpumask
*nodemask
)
8152 #endif /* CONFIG_NUMA */
8155 * Initialize sched groups cpu_power.
8157 * cpu_power indicates the capacity of sched group, which is used while
8158 * distributing the load between different sched groups in a sched domain.
8159 * Typically cpu_power for all the groups in a sched domain will be same unless
8160 * there are asymmetries in the topology. If there are asymmetries, group
8161 * having more cpu_power will pickup more load compared to the group having
8164 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
8165 * the maximum number of tasks a group can handle in the presence of other idle
8166 * or lightly loaded groups in the same sched domain.
8168 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
8170 struct sched_domain
*child
;
8171 struct sched_group
*group
;
8173 WARN_ON(!sd
|| !sd
->groups
);
8175 if (cpu
!= group_first_cpu(sd
->groups
))
8180 sd
->groups
->__cpu_power
= 0;
8183 * For perf policy, if the groups in child domain share resources
8184 * (for example cores sharing some portions of the cache hierarchy
8185 * or SMT), then set this domain groups cpu_power such that each group
8186 * can handle only one task, when there are other idle groups in the
8187 * same sched domain.
8189 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
8191 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
8192 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
8197 * add cpu_power of each child group to this groups cpu_power
8199 group
= child
->groups
;
8201 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
8202 group
= group
->next
;
8203 } while (group
!= child
->groups
);
8207 * Initializers for schedule domains
8208 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8211 #ifdef CONFIG_SCHED_DEBUG
8212 # define SD_INIT_NAME(sd, type) sd->name = #type
8214 # define SD_INIT_NAME(sd, type) do { } while (0)
8217 #define SD_INIT(sd, type) sd_init_##type(sd)
8219 #define SD_INIT_FUNC(type) \
8220 static noinline void sd_init_##type(struct sched_domain *sd) \
8222 memset(sd, 0, sizeof(*sd)); \
8223 *sd = SD_##type##_INIT; \
8224 sd->level = SD_LV_##type; \
8225 SD_INIT_NAME(sd, type); \
8230 SD_INIT_FUNC(ALLNODES
)
8233 #ifdef CONFIG_SCHED_SMT
8234 SD_INIT_FUNC(SIBLING
)
8236 #ifdef CONFIG_SCHED_MC
8240 static int default_relax_domain_level
= -1;
8242 static int __init
setup_relax_domain_level(char *str
)
8246 val
= simple_strtoul(str
, NULL
, 0);
8247 if (val
< SD_LV_MAX
)
8248 default_relax_domain_level
= val
;
8252 __setup("relax_domain_level=", setup_relax_domain_level
);
8254 static void set_domain_attribute(struct sched_domain
*sd
,
8255 struct sched_domain_attr
*attr
)
8259 if (!attr
|| attr
->relax_domain_level
< 0) {
8260 if (default_relax_domain_level
< 0)
8263 request
= default_relax_domain_level
;
8265 request
= attr
->relax_domain_level
;
8266 if (request
< sd
->level
) {
8267 /* turn off idle balance on this domain */
8268 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
8270 /* turn on idle balance on this domain */
8271 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
8276 * Build sched domains for a given set of cpus and attach the sched domains
8277 * to the individual cpus
8279 static int __build_sched_domains(const struct cpumask
*cpu_map
,
8280 struct sched_domain_attr
*attr
)
8282 int i
, err
= -ENOMEM
;
8283 struct root_domain
*rd
;
8284 cpumask_var_t nodemask
, this_sibling_map
, this_core_map
, send_covered
,
8287 cpumask_var_t domainspan
, covered
, notcovered
;
8288 struct sched_group
**sched_group_nodes
= NULL
;
8289 int sd_allnodes
= 0;
8291 if (!alloc_cpumask_var(&domainspan
, GFP_KERNEL
))
8293 if (!alloc_cpumask_var(&covered
, GFP_KERNEL
))
8294 goto free_domainspan
;
8295 if (!alloc_cpumask_var(¬covered
, GFP_KERNEL
))
8299 if (!alloc_cpumask_var(&nodemask
, GFP_KERNEL
))
8300 goto free_notcovered
;
8301 if (!alloc_cpumask_var(&this_sibling_map
, GFP_KERNEL
))
8303 if (!alloc_cpumask_var(&this_core_map
, GFP_KERNEL
))
8304 goto free_this_sibling_map
;
8305 if (!alloc_cpumask_var(&send_covered
, GFP_KERNEL
))
8306 goto free_this_core_map
;
8307 if (!alloc_cpumask_var(&tmpmask
, GFP_KERNEL
))
8308 goto free_send_covered
;
8312 * Allocate the per-node list of sched groups
8314 sched_group_nodes
= kcalloc(nr_node_ids
, sizeof(struct sched_group
*),
8316 if (!sched_group_nodes
) {
8317 printk(KERN_WARNING
"Can not alloc sched group node list\n");
8322 rd
= alloc_rootdomain();
8324 printk(KERN_WARNING
"Cannot alloc root domain\n");
8325 goto free_sched_groups
;
8329 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = sched_group_nodes
;
8333 * Set up domains for cpus specified by the cpu_map.
8335 for_each_cpu(i
, cpu_map
) {
8336 struct sched_domain
*sd
= NULL
, *p
;
8338 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(i
)), cpu_map
);
8341 if (cpumask_weight(cpu_map
) >
8342 SD_NODES_PER_DOMAIN
*cpumask_weight(nodemask
)) {
8343 sd
= &per_cpu(allnodes_domains
, i
).sd
;
8344 SD_INIT(sd
, ALLNODES
);
8345 set_domain_attribute(sd
, attr
);
8346 cpumask_copy(sched_domain_span(sd
), cpu_map
);
8347 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8353 sd
= &per_cpu(node_domains
, i
).sd
;
8355 set_domain_attribute(sd
, attr
);
8356 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
8360 cpumask_and(sched_domain_span(sd
),
8361 sched_domain_span(sd
), cpu_map
);
8365 sd
= &per_cpu(phys_domains
, i
).sd
;
8367 set_domain_attribute(sd
, attr
);
8368 cpumask_copy(sched_domain_span(sd
), nodemask
);
8372 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8374 #ifdef CONFIG_SCHED_MC
8376 sd
= &per_cpu(core_domains
, i
).sd
;
8378 set_domain_attribute(sd
, attr
);
8379 cpumask_and(sched_domain_span(sd
), cpu_map
,
8380 cpu_coregroup_mask(i
));
8383 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8386 #ifdef CONFIG_SCHED_SMT
8388 sd
= &per_cpu(cpu_domains
, i
).sd
;
8389 SD_INIT(sd
, SIBLING
);
8390 set_domain_attribute(sd
, attr
);
8391 cpumask_and(sched_domain_span(sd
),
8392 topology_thread_cpumask(i
), cpu_map
);
8395 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8399 #ifdef CONFIG_SCHED_SMT
8400 /* Set up CPU (sibling) groups */
8401 for_each_cpu(i
, cpu_map
) {
8402 cpumask_and(this_sibling_map
,
8403 topology_thread_cpumask(i
), cpu_map
);
8404 if (i
!= cpumask_first(this_sibling_map
))
8407 init_sched_build_groups(this_sibling_map
, cpu_map
,
8409 send_covered
, tmpmask
);
8413 #ifdef CONFIG_SCHED_MC
8414 /* Set up multi-core groups */
8415 for_each_cpu(i
, cpu_map
) {
8416 cpumask_and(this_core_map
, cpu_coregroup_mask(i
), cpu_map
);
8417 if (i
!= cpumask_first(this_core_map
))
8420 init_sched_build_groups(this_core_map
, cpu_map
,
8422 send_covered
, tmpmask
);
8426 /* Set up physical groups */
8427 for (i
= 0; i
< nr_node_ids
; i
++) {
8428 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8429 if (cpumask_empty(nodemask
))
8432 init_sched_build_groups(nodemask
, cpu_map
,
8434 send_covered
, tmpmask
);
8438 /* Set up node groups */
8440 init_sched_build_groups(cpu_map
, cpu_map
,
8441 &cpu_to_allnodes_group
,
8442 send_covered
, tmpmask
);
8445 for (i
= 0; i
< nr_node_ids
; i
++) {
8446 /* Set up node groups */
8447 struct sched_group
*sg
, *prev
;
8450 cpumask_clear(covered
);
8451 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8452 if (cpumask_empty(nodemask
)) {
8453 sched_group_nodes
[i
] = NULL
;
8457 sched_domain_node_span(i
, domainspan
);
8458 cpumask_and(domainspan
, domainspan
, cpu_map
);
8460 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8463 printk(KERN_WARNING
"Can not alloc domain group for "
8467 sched_group_nodes
[i
] = sg
;
8468 for_each_cpu(j
, nodemask
) {
8469 struct sched_domain
*sd
;
8471 sd
= &per_cpu(node_domains
, j
).sd
;
8474 sg
->__cpu_power
= 0;
8475 cpumask_copy(sched_group_cpus(sg
), nodemask
);
8477 cpumask_or(covered
, covered
, nodemask
);
8480 for (j
= 0; j
< nr_node_ids
; j
++) {
8481 int n
= (i
+ j
) % nr_node_ids
;
8483 cpumask_complement(notcovered
, covered
);
8484 cpumask_and(tmpmask
, notcovered
, cpu_map
);
8485 cpumask_and(tmpmask
, tmpmask
, domainspan
);
8486 if (cpumask_empty(tmpmask
))
8489 cpumask_and(tmpmask
, tmpmask
, cpumask_of_node(n
));
8490 if (cpumask_empty(tmpmask
))
8493 sg
= kmalloc_node(sizeof(struct sched_group
) +
8498 "Can not alloc domain group for node %d\n", j
);
8501 sg
->__cpu_power
= 0;
8502 cpumask_copy(sched_group_cpus(sg
), tmpmask
);
8503 sg
->next
= prev
->next
;
8504 cpumask_or(covered
, covered
, tmpmask
);
8511 /* Calculate CPU power for physical packages and nodes */
8512 #ifdef CONFIG_SCHED_SMT
8513 for_each_cpu(i
, cpu_map
) {
8514 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
).sd
;
8516 init_sched_groups_power(i
, sd
);
8519 #ifdef CONFIG_SCHED_MC
8520 for_each_cpu(i
, cpu_map
) {
8521 struct sched_domain
*sd
= &per_cpu(core_domains
, i
).sd
;
8523 init_sched_groups_power(i
, sd
);
8527 for_each_cpu(i
, cpu_map
) {
8528 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
).sd
;
8530 init_sched_groups_power(i
, sd
);
8534 for (i
= 0; i
< nr_node_ids
; i
++)
8535 init_numa_sched_groups_power(sched_group_nodes
[i
]);
8538 struct sched_group
*sg
;
8540 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
8542 init_numa_sched_groups_power(sg
);
8546 /* Attach the domains */
8547 for_each_cpu(i
, cpu_map
) {
8548 struct sched_domain
*sd
;
8549 #ifdef CONFIG_SCHED_SMT
8550 sd
= &per_cpu(cpu_domains
, i
).sd
;
8551 #elif defined(CONFIG_SCHED_MC)
8552 sd
= &per_cpu(core_domains
, i
).sd
;
8554 sd
= &per_cpu(phys_domains
, i
).sd
;
8556 cpu_attach_domain(sd
, rd
, i
);
8562 free_cpumask_var(tmpmask
);
8564 free_cpumask_var(send_covered
);
8566 free_cpumask_var(this_core_map
);
8567 free_this_sibling_map
:
8568 free_cpumask_var(this_sibling_map
);
8570 free_cpumask_var(nodemask
);
8573 free_cpumask_var(notcovered
);
8575 free_cpumask_var(covered
);
8577 free_cpumask_var(domainspan
);
8584 kfree(sched_group_nodes
);
8590 free_sched_groups(cpu_map
, tmpmask
);
8591 free_rootdomain(rd
);
8596 static int build_sched_domains(const struct cpumask
*cpu_map
)
8598 return __build_sched_domains(cpu_map
, NULL
);
8601 static struct cpumask
*doms_cur
; /* current sched domains */
8602 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
8603 static struct sched_domain_attr
*dattr_cur
;
8604 /* attribues of custom domains in 'doms_cur' */
8607 * Special case: If a kmalloc of a doms_cur partition (array of
8608 * cpumask) fails, then fallback to a single sched domain,
8609 * as determined by the single cpumask fallback_doms.
8611 static cpumask_var_t fallback_doms
;
8614 * arch_update_cpu_topology lets virtualized architectures update the
8615 * cpu core maps. It is supposed to return 1 if the topology changed
8616 * or 0 if it stayed the same.
8618 int __attribute__((weak
)) arch_update_cpu_topology(void)
8624 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8625 * For now this just excludes isolated cpus, but could be used to
8626 * exclude other special cases in the future.
8628 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
8632 arch_update_cpu_topology();
8634 doms_cur
= kmalloc(cpumask_size(), GFP_KERNEL
);
8636 doms_cur
= fallback_doms
;
8637 cpumask_andnot(doms_cur
, cpu_map
, cpu_isolated_map
);
8639 err
= build_sched_domains(doms_cur
);
8640 register_sched_domain_sysctl();
8645 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
8646 struct cpumask
*tmpmask
)
8648 free_sched_groups(cpu_map
, tmpmask
);
8652 * Detach sched domains from a group of cpus specified in cpu_map
8653 * These cpus will now be attached to the NULL domain
8655 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
8657 /* Save because hotplug lock held. */
8658 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
8661 for_each_cpu(i
, cpu_map
)
8662 cpu_attach_domain(NULL
, &def_root_domain
, i
);
8663 synchronize_sched();
8664 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
8667 /* handle null as "default" */
8668 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
8669 struct sched_domain_attr
*new, int idx_new
)
8671 struct sched_domain_attr tmp
;
8678 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
8679 new ? (new + idx_new
) : &tmp
,
8680 sizeof(struct sched_domain_attr
));
8684 * Partition sched domains as specified by the 'ndoms_new'
8685 * cpumasks in the array doms_new[] of cpumasks. This compares
8686 * doms_new[] to the current sched domain partitioning, doms_cur[].
8687 * It destroys each deleted domain and builds each new domain.
8689 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8690 * The masks don't intersect (don't overlap.) We should setup one
8691 * sched domain for each mask. CPUs not in any of the cpumasks will
8692 * not be load balanced. If the same cpumask appears both in the
8693 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8696 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8697 * ownership of it and will kfree it when done with it. If the caller
8698 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8699 * ndoms_new == 1, and partition_sched_domains() will fallback to
8700 * the single partition 'fallback_doms', it also forces the domains
8703 * If doms_new == NULL it will be replaced with cpu_online_mask.
8704 * ndoms_new == 0 is a special case for destroying existing domains,
8705 * and it will not create the default domain.
8707 * Call with hotplug lock held
8709 /* FIXME: Change to struct cpumask *doms_new[] */
8710 void partition_sched_domains(int ndoms_new
, struct cpumask
*doms_new
,
8711 struct sched_domain_attr
*dattr_new
)
8716 mutex_lock(&sched_domains_mutex
);
8718 /* always unregister in case we don't destroy any domains */
8719 unregister_sched_domain_sysctl();
8721 /* Let architecture update cpu core mappings. */
8722 new_topology
= arch_update_cpu_topology();
8724 n
= doms_new
? ndoms_new
: 0;
8726 /* Destroy deleted domains */
8727 for (i
= 0; i
< ndoms_cur
; i
++) {
8728 for (j
= 0; j
< n
&& !new_topology
; j
++) {
8729 if (cpumask_equal(&doms_cur
[i
], &doms_new
[j
])
8730 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
8733 /* no match - a current sched domain not in new doms_new[] */
8734 detach_destroy_domains(doms_cur
+ i
);
8739 if (doms_new
== NULL
) {
8741 doms_new
= fallback_doms
;
8742 cpumask_andnot(&doms_new
[0], cpu_online_mask
, cpu_isolated_map
);
8743 WARN_ON_ONCE(dattr_new
);
8746 /* Build new domains */
8747 for (i
= 0; i
< ndoms_new
; i
++) {
8748 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
8749 if (cpumask_equal(&doms_new
[i
], &doms_cur
[j
])
8750 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
8753 /* no match - add a new doms_new */
8754 __build_sched_domains(doms_new
+ i
,
8755 dattr_new
? dattr_new
+ i
: NULL
);
8760 /* Remember the new sched domains */
8761 if (doms_cur
!= fallback_doms
)
8763 kfree(dattr_cur
); /* kfree(NULL) is safe */
8764 doms_cur
= doms_new
;
8765 dattr_cur
= dattr_new
;
8766 ndoms_cur
= ndoms_new
;
8768 register_sched_domain_sysctl();
8770 mutex_unlock(&sched_domains_mutex
);
8773 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8774 static void arch_reinit_sched_domains(void)
8778 /* Destroy domains first to force the rebuild */
8779 partition_sched_domains(0, NULL
, NULL
);
8781 rebuild_sched_domains();
8785 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
8787 unsigned int level
= 0;
8789 if (sscanf(buf
, "%u", &level
) != 1)
8793 * level is always be positive so don't check for
8794 * level < POWERSAVINGS_BALANCE_NONE which is 0
8795 * What happens on 0 or 1 byte write,
8796 * need to check for count as well?
8799 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
8803 sched_smt_power_savings
= level
;
8805 sched_mc_power_savings
= level
;
8807 arch_reinit_sched_domains();
8812 #ifdef CONFIG_SCHED_MC
8813 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
8816 return sprintf(page
, "%u\n", sched_mc_power_savings
);
8818 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
8819 const char *buf
, size_t count
)
8821 return sched_power_savings_store(buf
, count
, 0);
8823 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
8824 sched_mc_power_savings_show
,
8825 sched_mc_power_savings_store
);
8828 #ifdef CONFIG_SCHED_SMT
8829 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
8832 return sprintf(page
, "%u\n", sched_smt_power_savings
);
8834 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
8835 const char *buf
, size_t count
)
8837 return sched_power_savings_store(buf
, count
, 1);
8839 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
8840 sched_smt_power_savings_show
,
8841 sched_smt_power_savings_store
);
8844 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
8848 #ifdef CONFIG_SCHED_SMT
8850 err
= sysfs_create_file(&cls
->kset
.kobj
,
8851 &attr_sched_smt_power_savings
.attr
);
8853 #ifdef CONFIG_SCHED_MC
8854 if (!err
&& mc_capable())
8855 err
= sysfs_create_file(&cls
->kset
.kobj
,
8856 &attr_sched_mc_power_savings
.attr
);
8860 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8862 #ifndef CONFIG_CPUSETS
8864 * Add online and remove offline CPUs from the scheduler domains.
8865 * When cpusets are enabled they take over this function.
8867 static int update_sched_domains(struct notifier_block
*nfb
,
8868 unsigned long action
, void *hcpu
)
8872 case CPU_ONLINE_FROZEN
:
8874 case CPU_DEAD_FROZEN
:
8875 partition_sched_domains(1, NULL
, NULL
);
8884 static int update_runtime(struct notifier_block
*nfb
,
8885 unsigned long action
, void *hcpu
)
8887 int cpu
= (int)(long)hcpu
;
8890 case CPU_DOWN_PREPARE
:
8891 case CPU_DOWN_PREPARE_FROZEN
:
8892 disable_runtime(cpu_rq(cpu
));
8895 case CPU_DOWN_FAILED
:
8896 case CPU_DOWN_FAILED_FROZEN
:
8898 case CPU_ONLINE_FROZEN
:
8899 enable_runtime(cpu_rq(cpu
));
8907 void __init
sched_init_smp(void)
8909 cpumask_var_t non_isolated_cpus
;
8911 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
8913 #if defined(CONFIG_NUMA)
8914 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
8916 BUG_ON(sched_group_nodes_bycpu
== NULL
);
8919 mutex_lock(&sched_domains_mutex
);
8920 arch_init_sched_domains(cpu_online_mask
);
8921 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
8922 if (cpumask_empty(non_isolated_cpus
))
8923 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
8924 mutex_unlock(&sched_domains_mutex
);
8927 #ifndef CONFIG_CPUSETS
8928 /* XXX: Theoretical race here - CPU may be hotplugged now */
8929 hotcpu_notifier(update_sched_domains
, 0);
8932 /* RT runtime code needs to handle some hotplug events */
8933 hotcpu_notifier(update_runtime
, 0);
8937 /* Move init over to a non-isolated CPU */
8938 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
8940 sched_init_granularity();
8941 free_cpumask_var(non_isolated_cpus
);
8943 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
8944 init_sched_rt_class();
8947 void __init
sched_init_smp(void)
8949 sched_init_granularity();
8951 #endif /* CONFIG_SMP */
8953 int in_sched_functions(unsigned long addr
)
8955 return in_lock_functions(addr
) ||
8956 (addr
>= (unsigned long)__sched_text_start
8957 && addr
< (unsigned long)__sched_text_end
);
8960 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
8962 cfs_rq
->tasks_timeline
= RB_ROOT
;
8963 INIT_LIST_HEAD(&cfs_rq
->tasks
);
8964 #ifdef CONFIG_FAIR_GROUP_SCHED
8967 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
8970 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
8972 struct rt_prio_array
*array
;
8975 array
= &rt_rq
->active
;
8976 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
8977 INIT_LIST_HEAD(array
->queue
+ i
);
8978 __clear_bit(i
, array
->bitmap
);
8980 /* delimiter for bitsearch: */
8981 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
8983 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8984 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
8986 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
8990 rt_rq
->rt_nr_migratory
= 0;
8991 rt_rq
->overloaded
= 0;
8992 plist_head_init(&rq
->rt
.pushable_tasks
, &rq
->lock
);
8996 rt_rq
->rt_throttled
= 0;
8997 rt_rq
->rt_runtime
= 0;
8998 spin_lock_init(&rt_rq
->rt_runtime_lock
);
9000 #ifdef CONFIG_RT_GROUP_SCHED
9001 rt_rq
->rt_nr_boosted
= 0;
9006 #ifdef CONFIG_FAIR_GROUP_SCHED
9007 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
9008 struct sched_entity
*se
, int cpu
, int add
,
9009 struct sched_entity
*parent
)
9011 struct rq
*rq
= cpu_rq(cpu
);
9012 tg
->cfs_rq
[cpu
] = cfs_rq
;
9013 init_cfs_rq(cfs_rq
, rq
);
9016 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
9019 /* se could be NULL for init_task_group */
9024 se
->cfs_rq
= &rq
->cfs
;
9026 se
->cfs_rq
= parent
->my_q
;
9029 se
->load
.weight
= tg
->shares
;
9030 se
->load
.inv_weight
= 0;
9031 se
->parent
= parent
;
9035 #ifdef CONFIG_RT_GROUP_SCHED
9036 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
9037 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
9038 struct sched_rt_entity
*parent
)
9040 struct rq
*rq
= cpu_rq(cpu
);
9042 tg
->rt_rq
[cpu
] = rt_rq
;
9043 init_rt_rq(rt_rq
, rq
);
9045 rt_rq
->rt_se
= rt_se
;
9046 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9048 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
9050 tg
->rt_se
[cpu
] = rt_se
;
9055 rt_se
->rt_rq
= &rq
->rt
;
9057 rt_se
->rt_rq
= parent
->my_q
;
9059 rt_se
->my_q
= rt_rq
;
9060 rt_se
->parent
= parent
;
9061 INIT_LIST_HEAD(&rt_se
->run_list
);
9065 void __init
sched_init(void)
9068 unsigned long alloc_size
= 0, ptr
;
9070 #ifdef CONFIG_FAIR_GROUP_SCHED
9071 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9073 #ifdef CONFIG_RT_GROUP_SCHED
9074 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9076 #ifdef CONFIG_USER_SCHED
9079 #ifdef CONFIG_CPUMASK_OFFSTACK
9080 alloc_size
+= num_possible_cpus() * cpumask_size();
9083 * As sched_init() is called before page_alloc is setup,
9084 * we use alloc_bootmem().
9087 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
9089 #ifdef CONFIG_FAIR_GROUP_SCHED
9090 init_task_group
.se
= (struct sched_entity
**)ptr
;
9091 ptr
+= nr_cpu_ids
* sizeof(void **);
9093 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9094 ptr
+= nr_cpu_ids
* sizeof(void **);
9096 #ifdef CONFIG_USER_SCHED
9097 root_task_group
.se
= (struct sched_entity
**)ptr
;
9098 ptr
+= nr_cpu_ids
* sizeof(void **);
9100 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9101 ptr
+= nr_cpu_ids
* sizeof(void **);
9102 #endif /* CONFIG_USER_SCHED */
9103 #endif /* CONFIG_FAIR_GROUP_SCHED */
9104 #ifdef CONFIG_RT_GROUP_SCHED
9105 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9106 ptr
+= nr_cpu_ids
* sizeof(void **);
9108 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9109 ptr
+= nr_cpu_ids
* sizeof(void **);
9111 #ifdef CONFIG_USER_SCHED
9112 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9113 ptr
+= nr_cpu_ids
* sizeof(void **);
9115 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9116 ptr
+= nr_cpu_ids
* sizeof(void **);
9117 #endif /* CONFIG_USER_SCHED */
9118 #endif /* CONFIG_RT_GROUP_SCHED */
9119 #ifdef CONFIG_CPUMASK_OFFSTACK
9120 for_each_possible_cpu(i
) {
9121 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
9122 ptr
+= cpumask_size();
9124 #endif /* CONFIG_CPUMASK_OFFSTACK */
9128 init_defrootdomain();
9131 init_rt_bandwidth(&def_rt_bandwidth
,
9132 global_rt_period(), global_rt_runtime());
9134 #ifdef CONFIG_RT_GROUP_SCHED
9135 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
9136 global_rt_period(), global_rt_runtime());
9137 #ifdef CONFIG_USER_SCHED
9138 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
9139 global_rt_period(), RUNTIME_INF
);
9140 #endif /* CONFIG_USER_SCHED */
9141 #endif /* CONFIG_RT_GROUP_SCHED */
9143 #ifdef CONFIG_GROUP_SCHED
9144 list_add(&init_task_group
.list
, &task_groups
);
9145 INIT_LIST_HEAD(&init_task_group
.children
);
9147 #ifdef CONFIG_USER_SCHED
9148 INIT_LIST_HEAD(&root_task_group
.children
);
9149 init_task_group
.parent
= &root_task_group
;
9150 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
9151 #endif /* CONFIG_USER_SCHED */
9152 #endif /* CONFIG_GROUP_SCHED */
9154 for_each_possible_cpu(i
) {
9158 spin_lock_init(&rq
->lock
);
9160 rq
->calc_load_active
= 0;
9161 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
9162 init_cfs_rq(&rq
->cfs
, rq
);
9163 init_rt_rq(&rq
->rt
, rq
);
9164 #ifdef CONFIG_FAIR_GROUP_SCHED
9165 init_task_group
.shares
= init_task_group_load
;
9166 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
9167 #ifdef CONFIG_CGROUP_SCHED
9169 * How much cpu bandwidth does init_task_group get?
9171 * In case of task-groups formed thr' the cgroup filesystem, it
9172 * gets 100% of the cpu resources in the system. This overall
9173 * system cpu resource is divided among the tasks of
9174 * init_task_group and its child task-groups in a fair manner,
9175 * based on each entity's (task or task-group's) weight
9176 * (se->load.weight).
9178 * In other words, if init_task_group has 10 tasks of weight
9179 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9180 * then A0's share of the cpu resource is:
9182 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9184 * We achieve this by letting init_task_group's tasks sit
9185 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9187 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
9188 #elif defined CONFIG_USER_SCHED
9189 root_task_group
.shares
= NICE_0_LOAD
;
9190 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
9192 * In case of task-groups formed thr' the user id of tasks,
9193 * init_task_group represents tasks belonging to root user.
9194 * Hence it forms a sibling of all subsequent groups formed.
9195 * In this case, init_task_group gets only a fraction of overall
9196 * system cpu resource, based on the weight assigned to root
9197 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9198 * by letting tasks of init_task_group sit in a separate cfs_rq
9199 * (init_cfs_rq) and having one entity represent this group of
9200 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9202 init_tg_cfs_entry(&init_task_group
,
9203 &per_cpu(init_cfs_rq
, i
),
9204 &per_cpu(init_sched_entity
, i
), i
, 1,
9205 root_task_group
.se
[i
]);
9208 #endif /* CONFIG_FAIR_GROUP_SCHED */
9210 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
9211 #ifdef CONFIG_RT_GROUP_SCHED
9212 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
9213 #ifdef CONFIG_CGROUP_SCHED
9214 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
9215 #elif defined CONFIG_USER_SCHED
9216 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
9217 init_tg_rt_entry(&init_task_group
,
9218 &per_cpu(init_rt_rq
, i
),
9219 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
9220 root_task_group
.rt_se
[i
]);
9224 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
9225 rq
->cpu_load
[j
] = 0;
9229 rq
->active_balance
= 0;
9230 rq
->next_balance
= jiffies
;
9234 rq
->migration_thread
= NULL
;
9235 INIT_LIST_HEAD(&rq
->migration_queue
);
9236 rq_attach_root(rq
, &def_root_domain
);
9239 atomic_set(&rq
->nr_iowait
, 0);
9242 set_load_weight(&init_task
);
9244 #ifdef CONFIG_PREEMPT_NOTIFIERS
9245 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
9249 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
9252 #ifdef CONFIG_RT_MUTEXES
9253 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
9257 * The boot idle thread does lazy MMU switching as well:
9259 atomic_inc(&init_mm
.mm_count
);
9260 enter_lazy_tlb(&init_mm
, current
);
9263 * Make us the idle thread. Technically, schedule() should not be
9264 * called from this thread, however somewhere below it might be,
9265 * but because we are the idle thread, we just pick up running again
9266 * when this runqueue becomes "idle".
9268 init_idle(current
, smp_processor_id());
9270 calc_load_update
= jiffies
+ LOAD_FREQ
;
9273 * During early bootup we pretend to be a normal task:
9275 current
->sched_class
= &fair_sched_class
;
9277 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9278 alloc_bootmem_cpumask_var(&nohz_cpu_mask
);
9281 alloc_bootmem_cpumask_var(&nohz
.cpu_mask
);
9282 alloc_bootmem_cpumask_var(&nohz
.ilb_grp_nohz_mask
);
9284 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
9287 scheduler_running
= 1;
9290 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9291 void __might_sleep(char *file
, int line
)
9294 static unsigned long prev_jiffy
; /* ratelimiting */
9296 if ((!in_atomic() && !irqs_disabled()) ||
9297 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
9299 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
9301 prev_jiffy
= jiffies
;
9304 "BUG: sleeping function called from invalid context at %s:%d\n",
9307 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9308 in_atomic(), irqs_disabled(),
9309 current
->pid
, current
->comm
);
9311 debug_show_held_locks(current
);
9312 if (irqs_disabled())
9313 print_irqtrace_events(current
);
9317 EXPORT_SYMBOL(__might_sleep
);
9320 #ifdef CONFIG_MAGIC_SYSRQ
9321 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
9325 update_rq_clock(rq
);
9326 on_rq
= p
->se
.on_rq
;
9328 deactivate_task(rq
, p
, 0);
9329 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
9331 activate_task(rq
, p
, 0);
9332 resched_task(rq
->curr
);
9336 void normalize_rt_tasks(void)
9338 struct task_struct
*g
, *p
;
9339 unsigned long flags
;
9342 read_lock_irqsave(&tasklist_lock
, flags
);
9343 do_each_thread(g
, p
) {
9345 * Only normalize user tasks:
9350 p
->se
.exec_start
= 0;
9351 #ifdef CONFIG_SCHEDSTATS
9352 p
->se
.wait_start
= 0;
9353 p
->se
.sleep_start
= 0;
9354 p
->se
.block_start
= 0;
9359 * Renice negative nice level userspace
9362 if (TASK_NICE(p
) < 0 && p
->mm
)
9363 set_user_nice(p
, 0);
9367 spin_lock(&p
->pi_lock
);
9368 rq
= __task_rq_lock(p
);
9370 normalize_task(rq
, p
);
9372 __task_rq_unlock(rq
);
9373 spin_unlock(&p
->pi_lock
);
9374 } while_each_thread(g
, p
);
9376 read_unlock_irqrestore(&tasklist_lock
, flags
);
9379 #endif /* CONFIG_MAGIC_SYSRQ */
9383 * These functions are only useful for the IA64 MCA handling.
9385 * They can only be called when the whole system has been
9386 * stopped - every CPU needs to be quiescent, and no scheduling
9387 * activity can take place. Using them for anything else would
9388 * be a serious bug, and as a result, they aren't even visible
9389 * under any other configuration.
9393 * curr_task - return the current task for a given cpu.
9394 * @cpu: the processor in question.
9396 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9398 struct task_struct
*curr_task(int cpu
)
9400 return cpu_curr(cpu
);
9404 * set_curr_task - set the current task for a given cpu.
9405 * @cpu: the processor in question.
9406 * @p: the task pointer to set.
9408 * Description: This function must only be used when non-maskable interrupts
9409 * are serviced on a separate stack. It allows the architecture to switch the
9410 * notion of the current task on a cpu in a non-blocking manner. This function
9411 * must be called with all CPU's synchronized, and interrupts disabled, the
9412 * and caller must save the original value of the current task (see
9413 * curr_task() above) and restore that value before reenabling interrupts and
9414 * re-starting the system.
9416 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9418 void set_curr_task(int cpu
, struct task_struct
*p
)
9425 #ifdef CONFIG_FAIR_GROUP_SCHED
9426 static void free_fair_sched_group(struct task_group
*tg
)
9430 for_each_possible_cpu(i
) {
9432 kfree(tg
->cfs_rq
[i
]);
9442 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9444 struct cfs_rq
*cfs_rq
;
9445 struct sched_entity
*se
;
9449 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9452 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
9456 tg
->shares
= NICE_0_LOAD
;
9458 for_each_possible_cpu(i
) {
9461 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
9462 GFP_KERNEL
, cpu_to_node(i
));
9466 se
= kzalloc_node(sizeof(struct sched_entity
),
9467 GFP_KERNEL
, cpu_to_node(i
));
9471 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
9480 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9482 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
9483 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
9486 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9488 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
9490 #else /* !CONFG_FAIR_GROUP_SCHED */
9491 static inline void free_fair_sched_group(struct task_group
*tg
)
9496 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9501 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9505 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9508 #endif /* CONFIG_FAIR_GROUP_SCHED */
9510 #ifdef CONFIG_RT_GROUP_SCHED
9511 static void free_rt_sched_group(struct task_group
*tg
)
9515 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
9517 for_each_possible_cpu(i
) {
9519 kfree(tg
->rt_rq
[i
]);
9521 kfree(tg
->rt_se
[i
]);
9529 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9531 struct rt_rq
*rt_rq
;
9532 struct sched_rt_entity
*rt_se
;
9536 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9539 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
9543 init_rt_bandwidth(&tg
->rt_bandwidth
,
9544 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
9546 for_each_possible_cpu(i
) {
9549 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
9550 GFP_KERNEL
, cpu_to_node(i
));
9554 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
9555 GFP_KERNEL
, cpu_to_node(i
));
9559 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
9568 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9570 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
9571 &cpu_rq(cpu
)->leaf_rt_rq_list
);
9574 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9576 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
9578 #else /* !CONFIG_RT_GROUP_SCHED */
9579 static inline void free_rt_sched_group(struct task_group
*tg
)
9584 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9589 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9593 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9596 #endif /* CONFIG_RT_GROUP_SCHED */
9598 #ifdef CONFIG_GROUP_SCHED
9599 static void free_sched_group(struct task_group
*tg
)
9601 free_fair_sched_group(tg
);
9602 free_rt_sched_group(tg
);
9606 /* allocate runqueue etc for a new task group */
9607 struct task_group
*sched_create_group(struct task_group
*parent
)
9609 struct task_group
*tg
;
9610 unsigned long flags
;
9613 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
9615 return ERR_PTR(-ENOMEM
);
9617 if (!alloc_fair_sched_group(tg
, parent
))
9620 if (!alloc_rt_sched_group(tg
, parent
))
9623 spin_lock_irqsave(&task_group_lock
, flags
);
9624 for_each_possible_cpu(i
) {
9625 register_fair_sched_group(tg
, i
);
9626 register_rt_sched_group(tg
, i
);
9628 list_add_rcu(&tg
->list
, &task_groups
);
9630 WARN_ON(!parent
); /* root should already exist */
9632 tg
->parent
= parent
;
9633 INIT_LIST_HEAD(&tg
->children
);
9634 list_add_rcu(&tg
->siblings
, &parent
->children
);
9635 spin_unlock_irqrestore(&task_group_lock
, flags
);
9640 free_sched_group(tg
);
9641 return ERR_PTR(-ENOMEM
);
9644 /* rcu callback to free various structures associated with a task group */
9645 static void free_sched_group_rcu(struct rcu_head
*rhp
)
9647 /* now it should be safe to free those cfs_rqs */
9648 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
9651 /* Destroy runqueue etc associated with a task group */
9652 void sched_destroy_group(struct task_group
*tg
)
9654 unsigned long flags
;
9657 spin_lock_irqsave(&task_group_lock
, flags
);
9658 for_each_possible_cpu(i
) {
9659 unregister_fair_sched_group(tg
, i
);
9660 unregister_rt_sched_group(tg
, i
);
9662 list_del_rcu(&tg
->list
);
9663 list_del_rcu(&tg
->siblings
);
9664 spin_unlock_irqrestore(&task_group_lock
, flags
);
9666 /* wait for possible concurrent references to cfs_rqs complete */
9667 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
9670 /* change task's runqueue when it moves between groups.
9671 * The caller of this function should have put the task in its new group
9672 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9673 * reflect its new group.
9675 void sched_move_task(struct task_struct
*tsk
)
9678 unsigned long flags
;
9681 rq
= task_rq_lock(tsk
, &flags
);
9683 update_rq_clock(rq
);
9685 running
= task_current(rq
, tsk
);
9686 on_rq
= tsk
->se
.on_rq
;
9689 dequeue_task(rq
, tsk
, 0);
9690 if (unlikely(running
))
9691 tsk
->sched_class
->put_prev_task(rq
, tsk
);
9693 set_task_rq(tsk
, task_cpu(tsk
));
9695 #ifdef CONFIG_FAIR_GROUP_SCHED
9696 if (tsk
->sched_class
->moved_group
)
9697 tsk
->sched_class
->moved_group(tsk
);
9700 if (unlikely(running
))
9701 tsk
->sched_class
->set_curr_task(rq
);
9703 enqueue_task(rq
, tsk
, 0);
9705 task_rq_unlock(rq
, &flags
);
9707 #endif /* CONFIG_GROUP_SCHED */
9709 #ifdef CONFIG_FAIR_GROUP_SCHED
9710 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
9712 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
9717 dequeue_entity(cfs_rq
, se
, 0);
9719 se
->load
.weight
= shares
;
9720 se
->load
.inv_weight
= 0;
9723 enqueue_entity(cfs_rq
, se
, 0);
9726 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
9728 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
9729 struct rq
*rq
= cfs_rq
->rq
;
9730 unsigned long flags
;
9732 spin_lock_irqsave(&rq
->lock
, flags
);
9733 __set_se_shares(se
, shares
);
9734 spin_unlock_irqrestore(&rq
->lock
, flags
);
9737 static DEFINE_MUTEX(shares_mutex
);
9739 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
9742 unsigned long flags
;
9745 * We can't change the weight of the root cgroup.
9750 if (shares
< MIN_SHARES
)
9751 shares
= MIN_SHARES
;
9752 else if (shares
> MAX_SHARES
)
9753 shares
= MAX_SHARES
;
9755 mutex_lock(&shares_mutex
);
9756 if (tg
->shares
== shares
)
9759 spin_lock_irqsave(&task_group_lock
, flags
);
9760 for_each_possible_cpu(i
)
9761 unregister_fair_sched_group(tg
, i
);
9762 list_del_rcu(&tg
->siblings
);
9763 spin_unlock_irqrestore(&task_group_lock
, flags
);
9765 /* wait for any ongoing reference to this group to finish */
9766 synchronize_sched();
9769 * Now we are free to modify the group's share on each cpu
9770 * w/o tripping rebalance_share or load_balance_fair.
9772 tg
->shares
= shares
;
9773 for_each_possible_cpu(i
) {
9777 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
9778 set_se_shares(tg
->se
[i
], shares
);
9782 * Enable load balance activity on this group, by inserting it back on
9783 * each cpu's rq->leaf_cfs_rq_list.
9785 spin_lock_irqsave(&task_group_lock
, flags
);
9786 for_each_possible_cpu(i
)
9787 register_fair_sched_group(tg
, i
);
9788 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
9789 spin_unlock_irqrestore(&task_group_lock
, flags
);
9791 mutex_unlock(&shares_mutex
);
9795 unsigned long sched_group_shares(struct task_group
*tg
)
9801 #ifdef CONFIG_RT_GROUP_SCHED
9803 * Ensure that the real time constraints are schedulable.
9805 static DEFINE_MUTEX(rt_constraints_mutex
);
9807 static unsigned long to_ratio(u64 period
, u64 runtime
)
9809 if (runtime
== RUNTIME_INF
)
9812 return div64_u64(runtime
<< 20, period
);
9815 /* Must be called with tasklist_lock held */
9816 static inline int tg_has_rt_tasks(struct task_group
*tg
)
9818 struct task_struct
*g
, *p
;
9820 do_each_thread(g
, p
) {
9821 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
9823 } while_each_thread(g
, p
);
9828 struct rt_schedulable_data
{
9829 struct task_group
*tg
;
9834 static int tg_schedulable(struct task_group
*tg
, void *data
)
9836 struct rt_schedulable_data
*d
= data
;
9837 struct task_group
*child
;
9838 unsigned long total
, sum
= 0;
9839 u64 period
, runtime
;
9841 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9842 runtime
= tg
->rt_bandwidth
.rt_runtime
;
9845 period
= d
->rt_period
;
9846 runtime
= d
->rt_runtime
;
9849 #ifdef CONFIG_USER_SCHED
9850 if (tg
== &root_task_group
) {
9851 period
= global_rt_period();
9852 runtime
= global_rt_runtime();
9857 * Cannot have more runtime than the period.
9859 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9863 * Ensure we don't starve existing RT tasks.
9865 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
9868 total
= to_ratio(period
, runtime
);
9871 * Nobody can have more than the global setting allows.
9873 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
9877 * The sum of our children's runtime should not exceed our own.
9879 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
9880 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
9881 runtime
= child
->rt_bandwidth
.rt_runtime
;
9883 if (child
== d
->tg
) {
9884 period
= d
->rt_period
;
9885 runtime
= d
->rt_runtime
;
9888 sum
+= to_ratio(period
, runtime
);
9897 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
9899 struct rt_schedulable_data data
= {
9901 .rt_period
= period
,
9902 .rt_runtime
= runtime
,
9905 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
9908 static int tg_set_bandwidth(struct task_group
*tg
,
9909 u64 rt_period
, u64 rt_runtime
)
9913 mutex_lock(&rt_constraints_mutex
);
9914 read_lock(&tasklist_lock
);
9915 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
9919 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
9920 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
9921 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
9923 for_each_possible_cpu(i
) {
9924 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
9926 spin_lock(&rt_rq
->rt_runtime_lock
);
9927 rt_rq
->rt_runtime
= rt_runtime
;
9928 spin_unlock(&rt_rq
->rt_runtime_lock
);
9930 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
9932 read_unlock(&tasklist_lock
);
9933 mutex_unlock(&rt_constraints_mutex
);
9938 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
9940 u64 rt_runtime
, rt_period
;
9942 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9943 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
9944 if (rt_runtime_us
< 0)
9945 rt_runtime
= RUNTIME_INF
;
9947 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9950 long sched_group_rt_runtime(struct task_group
*tg
)
9954 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
9957 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
9958 do_div(rt_runtime_us
, NSEC_PER_USEC
);
9959 return rt_runtime_us
;
9962 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
9964 u64 rt_runtime
, rt_period
;
9966 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
9967 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9972 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9975 long sched_group_rt_period(struct task_group
*tg
)
9979 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9980 do_div(rt_period_us
, NSEC_PER_USEC
);
9981 return rt_period_us
;
9984 static int sched_rt_global_constraints(void)
9986 u64 runtime
, period
;
9989 if (sysctl_sched_rt_period
<= 0)
9992 runtime
= global_rt_runtime();
9993 period
= global_rt_period();
9996 * Sanity check on the sysctl variables.
9998 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10001 mutex_lock(&rt_constraints_mutex
);
10002 read_lock(&tasklist_lock
);
10003 ret
= __rt_schedulable(NULL
, 0, 0);
10004 read_unlock(&tasklist_lock
);
10005 mutex_unlock(&rt_constraints_mutex
);
10010 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
10012 /* Don't accept realtime tasks when there is no way for them to run */
10013 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
10019 #else /* !CONFIG_RT_GROUP_SCHED */
10020 static int sched_rt_global_constraints(void)
10022 unsigned long flags
;
10025 if (sysctl_sched_rt_period
<= 0)
10029 * There's always some RT tasks in the root group
10030 * -- migration, kstopmachine etc..
10032 if (sysctl_sched_rt_runtime
== 0)
10035 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10036 for_each_possible_cpu(i
) {
10037 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
10039 spin_lock(&rt_rq
->rt_runtime_lock
);
10040 rt_rq
->rt_runtime
= global_rt_runtime();
10041 spin_unlock(&rt_rq
->rt_runtime_lock
);
10043 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10047 #endif /* CONFIG_RT_GROUP_SCHED */
10049 int sched_rt_handler(struct ctl_table
*table
, int write
,
10050 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
10054 int old_period
, old_runtime
;
10055 static DEFINE_MUTEX(mutex
);
10057 mutex_lock(&mutex
);
10058 old_period
= sysctl_sched_rt_period
;
10059 old_runtime
= sysctl_sched_rt_runtime
;
10061 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
10063 if (!ret
&& write
) {
10064 ret
= sched_rt_global_constraints();
10066 sysctl_sched_rt_period
= old_period
;
10067 sysctl_sched_rt_runtime
= old_runtime
;
10069 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
10070 def_rt_bandwidth
.rt_period
=
10071 ns_to_ktime(global_rt_period());
10074 mutex_unlock(&mutex
);
10079 #ifdef CONFIG_CGROUP_SCHED
10081 /* return corresponding task_group object of a cgroup */
10082 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
10084 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
10085 struct task_group
, css
);
10088 static struct cgroup_subsys_state
*
10089 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10091 struct task_group
*tg
, *parent
;
10093 if (!cgrp
->parent
) {
10094 /* This is early initialization for the top cgroup */
10095 return &init_task_group
.css
;
10098 parent
= cgroup_tg(cgrp
->parent
);
10099 tg
= sched_create_group(parent
);
10101 return ERR_PTR(-ENOMEM
);
10107 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10109 struct task_group
*tg
= cgroup_tg(cgrp
);
10111 sched_destroy_group(tg
);
10115 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10116 struct task_struct
*tsk
)
10118 #ifdef CONFIG_RT_GROUP_SCHED
10119 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
10122 /* We don't support RT-tasks being in separate groups */
10123 if (tsk
->sched_class
!= &fair_sched_class
)
10131 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10132 struct cgroup
*old_cont
, struct task_struct
*tsk
)
10134 sched_move_task(tsk
);
10137 #ifdef CONFIG_FAIR_GROUP_SCHED
10138 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
10141 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
10144 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
10146 struct task_group
*tg
= cgroup_tg(cgrp
);
10148 return (u64
) tg
->shares
;
10150 #endif /* CONFIG_FAIR_GROUP_SCHED */
10152 #ifdef CONFIG_RT_GROUP_SCHED
10153 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
10156 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
10159 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10161 return sched_group_rt_runtime(cgroup_tg(cgrp
));
10164 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
10167 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
10170 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
10172 return sched_group_rt_period(cgroup_tg(cgrp
));
10174 #endif /* CONFIG_RT_GROUP_SCHED */
10176 static struct cftype cpu_files
[] = {
10177 #ifdef CONFIG_FAIR_GROUP_SCHED
10180 .read_u64
= cpu_shares_read_u64
,
10181 .write_u64
= cpu_shares_write_u64
,
10184 #ifdef CONFIG_RT_GROUP_SCHED
10186 .name
= "rt_runtime_us",
10187 .read_s64
= cpu_rt_runtime_read
,
10188 .write_s64
= cpu_rt_runtime_write
,
10191 .name
= "rt_period_us",
10192 .read_u64
= cpu_rt_period_read_uint
,
10193 .write_u64
= cpu_rt_period_write_uint
,
10198 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
10200 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
10203 struct cgroup_subsys cpu_cgroup_subsys
= {
10205 .create
= cpu_cgroup_create
,
10206 .destroy
= cpu_cgroup_destroy
,
10207 .can_attach
= cpu_cgroup_can_attach
,
10208 .attach
= cpu_cgroup_attach
,
10209 .populate
= cpu_cgroup_populate
,
10210 .subsys_id
= cpu_cgroup_subsys_id
,
10214 #endif /* CONFIG_CGROUP_SCHED */
10216 #ifdef CONFIG_CGROUP_CPUACCT
10219 * CPU accounting code for task groups.
10221 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10222 * (balbir@in.ibm.com).
10225 /* track cpu usage of a group of tasks and its child groups */
10227 struct cgroup_subsys_state css
;
10228 /* cpuusage holds pointer to a u64-type object on every cpu */
10230 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
10231 struct cpuacct
*parent
;
10234 struct cgroup_subsys cpuacct_subsys
;
10236 /* return cpu accounting group corresponding to this container */
10237 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
10239 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
10240 struct cpuacct
, css
);
10243 /* return cpu accounting group to which this task belongs */
10244 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
10246 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
10247 struct cpuacct
, css
);
10250 /* create a new cpu accounting group */
10251 static struct cgroup_subsys_state
*cpuacct_create(
10252 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10254 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
10260 ca
->cpuusage
= alloc_percpu(u64
);
10264 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10265 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
10266 goto out_free_counters
;
10269 ca
->parent
= cgroup_ca(cgrp
->parent
);
10275 percpu_counter_destroy(&ca
->cpustat
[i
]);
10276 free_percpu(ca
->cpuusage
);
10280 return ERR_PTR(-ENOMEM
);
10283 /* destroy an existing cpu accounting group */
10285 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10287 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10290 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10291 percpu_counter_destroy(&ca
->cpustat
[i
]);
10292 free_percpu(ca
->cpuusage
);
10296 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
10298 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10301 #ifndef CONFIG_64BIT
10303 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10305 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10307 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10315 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
10317 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10319 #ifndef CONFIG_64BIT
10321 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10323 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10325 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10331 /* return total cpu usage (in nanoseconds) of a group */
10332 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10334 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10335 u64 totalcpuusage
= 0;
10338 for_each_present_cpu(i
)
10339 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
10341 return totalcpuusage
;
10344 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
10347 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10356 for_each_present_cpu(i
)
10357 cpuacct_cpuusage_write(ca
, i
, 0);
10363 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
10364 struct seq_file
*m
)
10366 struct cpuacct
*ca
= cgroup_ca(cgroup
);
10370 for_each_present_cpu(i
) {
10371 percpu
= cpuacct_cpuusage_read(ca
, i
);
10372 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
10374 seq_printf(m
, "\n");
10378 static const char *cpuacct_stat_desc
[] = {
10379 [CPUACCT_STAT_USER
] = "user",
10380 [CPUACCT_STAT_SYSTEM
] = "system",
10383 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
10384 struct cgroup_map_cb
*cb
)
10386 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10389 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
10390 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
10391 val
= cputime64_to_clock_t(val
);
10392 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
10397 static struct cftype files
[] = {
10400 .read_u64
= cpuusage_read
,
10401 .write_u64
= cpuusage_write
,
10404 .name
= "usage_percpu",
10405 .read_seq_string
= cpuacct_percpu_seq_read
,
10409 .read_map
= cpuacct_stats_show
,
10413 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10415 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
10419 * charge this task's execution time to its accounting group.
10421 * called with rq->lock held.
10423 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
10425 struct cpuacct
*ca
;
10428 if (unlikely(!cpuacct_subsys
.active
))
10431 cpu
= task_cpu(tsk
);
10437 for (; ca
; ca
= ca
->parent
) {
10438 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10439 *cpuusage
+= cputime
;
10446 * Charge the system/user time to the task's accounting group.
10448 static void cpuacct_update_stats(struct task_struct
*tsk
,
10449 enum cpuacct_stat_index idx
, cputime_t val
)
10451 struct cpuacct
*ca
;
10453 if (unlikely(!cpuacct_subsys
.active
))
10460 percpu_counter_add(&ca
->cpustat
[idx
], val
);
10466 struct cgroup_subsys cpuacct_subsys
= {
10468 .create
= cpuacct_create
,
10469 .destroy
= cpuacct_destroy
,
10470 .populate
= cpuacct_populate
,
10471 .subsys_id
= cpuacct_subsys_id
,
10473 #endif /* CONFIG_CGROUP_CPUACCT */