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
);
234 if (hrtimer_active(&rt_b
->rt_period_timer
))
237 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
238 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
239 hrtimer_start_expires(&rt_b
->rt_period_timer
,
242 spin_unlock(&rt_b
->rt_runtime_lock
);
245 #ifdef CONFIG_RT_GROUP_SCHED
246 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
248 hrtimer_cancel(&rt_b
->rt_period_timer
);
253 * sched_domains_mutex serializes calls to arch_init_sched_domains,
254 * detach_destroy_domains and partition_sched_domains.
256 static DEFINE_MUTEX(sched_domains_mutex
);
258 #ifdef CONFIG_GROUP_SCHED
260 #include <linux/cgroup.h>
264 static LIST_HEAD(task_groups
);
266 /* task group related information */
268 #ifdef CONFIG_CGROUP_SCHED
269 struct cgroup_subsys_state css
;
272 #ifdef CONFIG_USER_SCHED
276 #ifdef CONFIG_FAIR_GROUP_SCHED
277 /* schedulable entities of this group on each cpu */
278 struct sched_entity
**se
;
279 /* runqueue "owned" by this group on each cpu */
280 struct cfs_rq
**cfs_rq
;
281 unsigned long shares
;
284 #ifdef CONFIG_RT_GROUP_SCHED
285 struct sched_rt_entity
**rt_se
;
286 struct rt_rq
**rt_rq
;
288 struct rt_bandwidth rt_bandwidth
;
292 struct list_head list
;
294 struct task_group
*parent
;
295 struct list_head siblings
;
296 struct list_head children
;
299 #ifdef CONFIG_USER_SCHED
301 /* Helper function to pass uid information to create_sched_user() */
302 void set_tg_uid(struct user_struct
*user
)
304 user
->tg
->uid
= user
->uid
;
309 * Every UID task group (including init_task_group aka UID-0) will
310 * be a child to this group.
312 struct task_group root_task_group
;
314 #ifdef CONFIG_FAIR_GROUP_SCHED
315 /* Default task group's sched entity on each cpu */
316 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
317 /* Default task group's cfs_rq on each cpu */
318 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
319 #endif /* CONFIG_FAIR_GROUP_SCHED */
321 #ifdef CONFIG_RT_GROUP_SCHED
322 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
323 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
324 #endif /* CONFIG_RT_GROUP_SCHED */
325 #else /* !CONFIG_USER_SCHED */
326 #define root_task_group init_task_group
327 #endif /* CONFIG_USER_SCHED */
329 /* task_group_lock serializes add/remove of task groups and also changes to
330 * a task group's cpu shares.
332 static DEFINE_SPINLOCK(task_group_lock
);
334 #ifdef CONFIG_FAIR_GROUP_SCHED
335 #ifdef CONFIG_USER_SCHED
336 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
337 #else /* !CONFIG_USER_SCHED */
338 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
339 #endif /* CONFIG_USER_SCHED */
342 * A weight of 0 or 1 can cause arithmetics problems.
343 * A weight of a cfs_rq is the sum of weights of which entities
344 * are queued on this cfs_rq, so a weight of a entity should not be
345 * too large, so as the shares value of a task group.
346 * (The default weight is 1024 - so there's no practical
347 * limitation from this.)
350 #define MAX_SHARES (1UL << 18)
352 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
355 /* Default task group.
356 * Every task in system belong to this group at bootup.
358 struct task_group init_task_group
;
360 /* return group to which a task belongs */
361 static inline struct task_group
*task_group(struct task_struct
*p
)
363 struct task_group
*tg
;
365 #ifdef CONFIG_USER_SCHED
367 tg
= __task_cred(p
)->user
->tg
;
369 #elif defined(CONFIG_CGROUP_SCHED)
370 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
371 struct task_group
, css
);
373 tg
= &init_task_group
;
378 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
379 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
381 #ifdef CONFIG_FAIR_GROUP_SCHED
382 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
383 p
->se
.parent
= task_group(p
)->se
[cpu
];
386 #ifdef CONFIG_RT_GROUP_SCHED
387 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
388 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
394 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
395 static inline struct task_group
*task_group(struct task_struct
*p
)
400 #endif /* CONFIG_GROUP_SCHED */
402 /* CFS-related fields in a runqueue */
404 struct load_weight load
;
405 unsigned long nr_running
;
410 struct rb_root tasks_timeline
;
411 struct rb_node
*rb_leftmost
;
413 struct list_head tasks
;
414 struct list_head
*balance_iterator
;
417 * 'curr' points to currently running entity on this cfs_rq.
418 * It is set to NULL otherwise (i.e when none are currently running).
420 struct sched_entity
*curr
, *next
, *last
;
422 unsigned int nr_spread_over
;
424 #ifdef CONFIG_FAIR_GROUP_SCHED
425 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
428 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
429 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
430 * (like users, containers etc.)
432 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
433 * list is used during load balance.
435 struct list_head leaf_cfs_rq_list
;
436 struct task_group
*tg
; /* group that "owns" this runqueue */
440 * the part of load.weight contributed by tasks
442 unsigned long task_weight
;
445 * h_load = weight * f(tg)
447 * Where f(tg) is the recursive weight fraction assigned to
450 unsigned long h_load
;
453 * this cpu's part of tg->shares
455 unsigned long shares
;
458 * load.weight at the time we set shares
460 unsigned long rq_weight
;
465 /* Real-Time classes' related field in a runqueue: */
467 struct rt_prio_array active
;
468 unsigned long rt_nr_running
;
469 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
470 int highest_prio
; /* highest queued rt task prio */
473 unsigned long rt_nr_migratory
;
479 /* Nests inside the rq lock: */
480 spinlock_t rt_runtime_lock
;
482 #ifdef CONFIG_RT_GROUP_SCHED
483 unsigned long rt_nr_boosted
;
486 struct list_head leaf_rt_rq_list
;
487 struct task_group
*tg
;
488 struct sched_rt_entity
*rt_se
;
495 * We add the notion of a root-domain which will be used to define per-domain
496 * variables. Each exclusive cpuset essentially defines an island domain by
497 * fully partitioning the member cpus from any other cpuset. Whenever a new
498 * exclusive cpuset is created, we also create and attach a new root-domain
505 cpumask_var_t online
;
508 * The "RT overload" flag: it gets set if a CPU has more than
509 * one runnable RT task.
511 cpumask_var_t rto_mask
;
514 struct cpupri cpupri
;
516 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
518 * Preferred wake up cpu nominated by sched_mc balance that will be
519 * used when most cpus are idle in the system indicating overall very
520 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
522 unsigned int sched_mc_preferred_wakeup_cpu
;
527 * By default the system creates a single root-domain with all cpus as
528 * members (mimicking the global state we have today).
530 static struct root_domain def_root_domain
;
535 * This is the main, per-CPU runqueue data structure.
537 * Locking rule: those places that want to lock multiple runqueues
538 * (such as the load balancing or the thread migration code), lock
539 * acquire operations must be ordered by ascending &runqueue.
546 * nr_running and cpu_load should be in the same cacheline because
547 * remote CPUs use both these fields when doing load calculation.
549 unsigned long nr_running
;
550 #define CPU_LOAD_IDX_MAX 5
551 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
552 unsigned char idle_at_tick
;
554 unsigned long last_tick_seen
;
555 unsigned char in_nohz_recently
;
557 /* capture load from *all* tasks on this cpu: */
558 struct load_weight load
;
559 unsigned long nr_load_updates
;
565 #ifdef CONFIG_FAIR_GROUP_SCHED
566 /* list of leaf cfs_rq on this cpu: */
567 struct list_head leaf_cfs_rq_list
;
569 #ifdef CONFIG_RT_GROUP_SCHED
570 struct list_head leaf_rt_rq_list
;
574 * This is part of a global counter where only the total sum
575 * over all CPUs matters. A task can increase this counter on
576 * one CPU and if it got migrated afterwards it may decrease
577 * it on another CPU. Always updated under the runqueue lock:
579 unsigned long nr_uninterruptible
;
581 struct task_struct
*curr
, *idle
;
582 unsigned long next_balance
;
583 struct mm_struct
*prev_mm
;
590 struct root_domain
*rd
;
591 struct sched_domain
*sd
;
593 /* For active balancing */
596 /* cpu of this runqueue: */
600 unsigned long avg_load_per_task
;
602 struct task_struct
*migration_thread
;
603 struct list_head migration_queue
;
606 #ifdef CONFIG_SCHED_HRTICK
608 int hrtick_csd_pending
;
609 struct call_single_data hrtick_csd
;
611 struct hrtimer hrtick_timer
;
614 #ifdef CONFIG_SCHEDSTATS
616 struct sched_info rq_sched_info
;
617 unsigned long long rq_cpu_time
;
618 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
620 /* sys_sched_yield() stats */
621 unsigned int yld_exp_empty
;
622 unsigned int yld_act_empty
;
623 unsigned int yld_both_empty
;
624 unsigned int yld_count
;
626 /* schedule() stats */
627 unsigned int sched_switch
;
628 unsigned int sched_count
;
629 unsigned int sched_goidle
;
631 /* try_to_wake_up() stats */
632 unsigned int ttwu_count
;
633 unsigned int ttwu_local
;
636 unsigned int bkl_count
;
640 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
642 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int sync
)
644 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, sync
);
647 static inline int cpu_of(struct rq
*rq
)
657 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
658 * See detach_destroy_domains: synchronize_sched for details.
660 * The domain tree of any CPU may only be accessed from within
661 * preempt-disabled sections.
663 #define for_each_domain(cpu, __sd) \
664 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
666 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
667 #define this_rq() (&__get_cpu_var(runqueues))
668 #define task_rq(p) cpu_rq(task_cpu(p))
669 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
671 static inline void update_rq_clock(struct rq
*rq
)
673 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
677 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
679 #ifdef CONFIG_SCHED_DEBUG
680 # define const_debug __read_mostly
682 # define const_debug static const
688 * Returns true if the current cpu runqueue is locked.
689 * This interface allows printk to be called with the runqueue lock
690 * held and know whether or not it is OK to wake up the klogd.
692 int runqueue_is_locked(void)
695 struct rq
*rq
= cpu_rq(cpu
);
698 ret
= spin_is_locked(&rq
->lock
);
704 * Debugging: various feature bits
707 #define SCHED_FEAT(name, enabled) \
708 __SCHED_FEAT_##name ,
711 #include "sched_features.h"
716 #define SCHED_FEAT(name, enabled) \
717 (1UL << __SCHED_FEAT_##name) * enabled |
719 const_debug
unsigned int sysctl_sched_features
=
720 #include "sched_features.h"
725 #ifdef CONFIG_SCHED_DEBUG
726 #define SCHED_FEAT(name, enabled) \
729 static __read_mostly
char *sched_feat_names
[] = {
730 #include "sched_features.h"
736 static int sched_feat_show(struct seq_file
*m
, void *v
)
740 for (i
= 0; sched_feat_names
[i
]; i
++) {
741 if (!(sysctl_sched_features
& (1UL << i
)))
743 seq_printf(m
, "%s ", sched_feat_names
[i
]);
751 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
752 size_t cnt
, loff_t
*ppos
)
762 if (copy_from_user(&buf
, ubuf
, cnt
))
767 if (strncmp(buf
, "NO_", 3) == 0) {
772 for (i
= 0; sched_feat_names
[i
]; i
++) {
773 int len
= strlen(sched_feat_names
[i
]);
775 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
777 sysctl_sched_features
&= ~(1UL << i
);
779 sysctl_sched_features
|= (1UL << i
);
784 if (!sched_feat_names
[i
])
792 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
794 return single_open(filp
, sched_feat_show
, NULL
);
797 static struct file_operations sched_feat_fops
= {
798 .open
= sched_feat_open
,
799 .write
= sched_feat_write
,
802 .release
= single_release
,
805 static __init
int sched_init_debug(void)
807 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
812 late_initcall(sched_init_debug
);
816 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
819 * Number of tasks to iterate in a single balance run.
820 * Limited because this is done with IRQs disabled.
822 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
825 * ratelimit for updating the group shares.
828 unsigned int sysctl_sched_shares_ratelimit
= 250000;
831 * Inject some fuzzyness into changing the per-cpu group shares
832 * this avoids remote rq-locks at the expense of fairness.
835 unsigned int sysctl_sched_shares_thresh
= 4;
838 * period over which we measure -rt task cpu usage in us.
841 unsigned int sysctl_sched_rt_period
= 1000000;
843 static __read_mostly
int scheduler_running
;
846 * part of the period that we allow rt tasks to run in us.
849 int sysctl_sched_rt_runtime
= 950000;
851 static inline u64
global_rt_period(void)
853 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
856 static inline u64
global_rt_runtime(void)
858 if (sysctl_sched_rt_runtime
< 0)
861 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
864 #ifndef prepare_arch_switch
865 # define prepare_arch_switch(next) do { } while (0)
867 #ifndef finish_arch_switch
868 # define finish_arch_switch(prev) do { } while (0)
871 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
873 return rq
->curr
== p
;
876 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
877 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
879 return task_current(rq
, p
);
882 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
886 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
888 #ifdef CONFIG_DEBUG_SPINLOCK
889 /* this is a valid case when another task releases the spinlock */
890 rq
->lock
.owner
= current
;
893 * If we are tracking spinlock dependencies then we have to
894 * fix up the runqueue lock - which gets 'carried over' from
897 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
899 spin_unlock_irq(&rq
->lock
);
902 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
903 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
908 return task_current(rq
, p
);
912 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
916 * We can optimise this out completely for !SMP, because the
917 * SMP rebalancing from interrupt is the only thing that cares
922 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
923 spin_unlock_irq(&rq
->lock
);
925 spin_unlock(&rq
->lock
);
929 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
933 * After ->oncpu is cleared, the task can be moved to a different CPU.
934 * We must ensure this doesn't happen until the switch is completely
940 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
944 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
947 * __task_rq_lock - lock the runqueue a given task resides on.
948 * Must be called interrupts disabled.
950 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
954 struct rq
*rq
= task_rq(p
);
955 spin_lock(&rq
->lock
);
956 if (likely(rq
== task_rq(p
)))
958 spin_unlock(&rq
->lock
);
963 * task_rq_lock - lock the runqueue a given task resides on and disable
964 * interrupts. Note the ordering: we can safely lookup the task_rq without
965 * explicitly disabling preemption.
967 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
973 local_irq_save(*flags
);
975 spin_lock(&rq
->lock
);
976 if (likely(rq
== task_rq(p
)))
978 spin_unlock_irqrestore(&rq
->lock
, *flags
);
982 void task_rq_unlock_wait(struct task_struct
*p
)
984 struct rq
*rq
= task_rq(p
);
986 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
987 spin_unlock_wait(&rq
->lock
);
990 static void __task_rq_unlock(struct rq
*rq
)
993 spin_unlock(&rq
->lock
);
996 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
999 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1003 * this_rq_lock - lock this runqueue and disable interrupts.
1005 static struct rq
*this_rq_lock(void)
1006 __acquires(rq
->lock
)
1010 local_irq_disable();
1012 spin_lock(&rq
->lock
);
1017 #ifdef CONFIG_SCHED_HRTICK
1019 * Use HR-timers to deliver accurate preemption points.
1021 * Its all a bit involved since we cannot program an hrt while holding the
1022 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1025 * When we get rescheduled we reprogram the hrtick_timer outside of the
1031 * - enabled by features
1032 * - hrtimer is actually high res
1034 static inline int hrtick_enabled(struct rq
*rq
)
1036 if (!sched_feat(HRTICK
))
1038 if (!cpu_active(cpu_of(rq
)))
1040 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1043 static void hrtick_clear(struct rq
*rq
)
1045 if (hrtimer_active(&rq
->hrtick_timer
))
1046 hrtimer_cancel(&rq
->hrtick_timer
);
1050 * High-resolution timer tick.
1051 * Runs from hardirq context with interrupts disabled.
1053 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1055 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1057 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1059 spin_lock(&rq
->lock
);
1060 update_rq_clock(rq
);
1061 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1062 spin_unlock(&rq
->lock
);
1064 return HRTIMER_NORESTART
;
1069 * called from hardirq (IPI) context
1071 static void __hrtick_start(void *arg
)
1073 struct rq
*rq
= arg
;
1075 spin_lock(&rq
->lock
);
1076 hrtimer_restart(&rq
->hrtick_timer
);
1077 rq
->hrtick_csd_pending
= 0;
1078 spin_unlock(&rq
->lock
);
1082 * Called to set the hrtick timer state.
1084 * called with rq->lock held and irqs disabled
1086 static void hrtick_start(struct rq
*rq
, u64 delay
)
1088 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1089 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1091 hrtimer_set_expires(timer
, time
);
1093 if (rq
== this_rq()) {
1094 hrtimer_restart(timer
);
1095 } else if (!rq
->hrtick_csd_pending
) {
1096 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
);
1097 rq
->hrtick_csd_pending
= 1;
1102 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1104 int cpu
= (int)(long)hcpu
;
1107 case CPU_UP_CANCELED
:
1108 case CPU_UP_CANCELED_FROZEN
:
1109 case CPU_DOWN_PREPARE
:
1110 case CPU_DOWN_PREPARE_FROZEN
:
1112 case CPU_DEAD_FROZEN
:
1113 hrtick_clear(cpu_rq(cpu
));
1120 static __init
void init_hrtick(void)
1122 hotcpu_notifier(hotplug_hrtick
, 0);
1126 * Called to set the hrtick timer state.
1128 * called with rq->lock held and irqs disabled
1130 static void hrtick_start(struct rq
*rq
, u64 delay
)
1132 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
), HRTIMER_MODE_REL
);
1135 static inline void init_hrtick(void)
1138 #endif /* CONFIG_SMP */
1140 static void init_rq_hrtick(struct rq
*rq
)
1143 rq
->hrtick_csd_pending
= 0;
1145 rq
->hrtick_csd
.flags
= 0;
1146 rq
->hrtick_csd
.func
= __hrtick_start
;
1147 rq
->hrtick_csd
.info
= rq
;
1150 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1151 rq
->hrtick_timer
.function
= hrtick
;
1153 #else /* CONFIG_SCHED_HRTICK */
1154 static inline void hrtick_clear(struct rq
*rq
)
1158 static inline void init_rq_hrtick(struct rq
*rq
)
1162 static inline void init_hrtick(void)
1165 #endif /* CONFIG_SCHED_HRTICK */
1168 * resched_task - mark a task 'to be rescheduled now'.
1170 * On UP this means the setting of the need_resched flag, on SMP it
1171 * might also involve a cross-CPU call to trigger the scheduler on
1176 #ifndef tsk_is_polling
1177 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1180 static void resched_task(struct task_struct
*p
)
1184 assert_spin_locked(&task_rq(p
)->lock
);
1186 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
1189 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
1192 if (cpu
== smp_processor_id())
1195 /* NEED_RESCHED must be visible before we test polling */
1197 if (!tsk_is_polling(p
))
1198 smp_send_reschedule(cpu
);
1201 static void resched_cpu(int cpu
)
1203 struct rq
*rq
= cpu_rq(cpu
);
1204 unsigned long flags
;
1206 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1208 resched_task(cpu_curr(cpu
));
1209 spin_unlock_irqrestore(&rq
->lock
, flags
);
1214 * When add_timer_on() enqueues a timer into the timer wheel of an
1215 * idle CPU then this timer might expire before the next timer event
1216 * which is scheduled to wake up that CPU. In case of a completely
1217 * idle system the next event might even be infinite time into the
1218 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1219 * leaves the inner idle loop so the newly added timer is taken into
1220 * account when the CPU goes back to idle and evaluates the timer
1221 * wheel for the next timer event.
1223 void wake_up_idle_cpu(int cpu
)
1225 struct rq
*rq
= cpu_rq(cpu
);
1227 if (cpu
== smp_processor_id())
1231 * This is safe, as this function is called with the timer
1232 * wheel base lock of (cpu) held. When the CPU is on the way
1233 * to idle and has not yet set rq->curr to idle then it will
1234 * be serialized on the timer wheel base lock and take the new
1235 * timer into account automatically.
1237 if (rq
->curr
!= rq
->idle
)
1241 * We can set TIF_RESCHED on the idle task of the other CPU
1242 * lockless. The worst case is that the other CPU runs the
1243 * idle task through an additional NOOP schedule()
1245 set_tsk_thread_flag(rq
->idle
, TIF_NEED_RESCHED
);
1247 /* NEED_RESCHED must be visible before we test polling */
1249 if (!tsk_is_polling(rq
->idle
))
1250 smp_send_reschedule(cpu
);
1252 #endif /* CONFIG_NO_HZ */
1254 #else /* !CONFIG_SMP */
1255 static void resched_task(struct task_struct
*p
)
1257 assert_spin_locked(&task_rq(p
)->lock
);
1258 set_tsk_need_resched(p
);
1260 #endif /* CONFIG_SMP */
1262 #if BITS_PER_LONG == 32
1263 # define WMULT_CONST (~0UL)
1265 # define WMULT_CONST (1UL << 32)
1268 #define WMULT_SHIFT 32
1271 * Shift right and round:
1273 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1276 * delta *= weight / lw
1278 static unsigned long
1279 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1280 struct load_weight
*lw
)
1284 if (!lw
->inv_weight
) {
1285 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1288 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1292 tmp
= (u64
)delta_exec
* weight
;
1294 * Check whether we'd overflow the 64-bit multiplication:
1296 if (unlikely(tmp
> WMULT_CONST
))
1297 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1300 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1302 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1305 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1311 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1318 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1319 * of tasks with abnormal "nice" values across CPUs the contribution that
1320 * each task makes to its run queue's load is weighted according to its
1321 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1322 * scaled version of the new time slice allocation that they receive on time
1326 #define WEIGHT_IDLEPRIO 2
1327 #define WMULT_IDLEPRIO (1 << 31)
1330 * Nice levels are multiplicative, with a gentle 10% change for every
1331 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1332 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1333 * that remained on nice 0.
1335 * The "10% effect" is relative and cumulative: from _any_ nice level,
1336 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1337 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1338 * If a task goes up by ~10% and another task goes down by ~10% then
1339 * the relative distance between them is ~25%.)
1341 static const int prio_to_weight
[40] = {
1342 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1343 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1344 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1345 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1346 /* 0 */ 1024, 820, 655, 526, 423,
1347 /* 5 */ 335, 272, 215, 172, 137,
1348 /* 10 */ 110, 87, 70, 56, 45,
1349 /* 15 */ 36, 29, 23, 18, 15,
1353 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1355 * In cases where the weight does not change often, we can use the
1356 * precalculated inverse to speed up arithmetics by turning divisions
1357 * into multiplications:
1359 static const u32 prio_to_wmult
[40] = {
1360 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1361 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1362 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1363 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1364 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1365 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1366 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1367 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1370 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1373 * runqueue iterator, to support SMP load-balancing between different
1374 * scheduling classes, without having to expose their internal data
1375 * structures to the load-balancing proper:
1377 struct rq_iterator
{
1379 struct task_struct
*(*start
)(void *);
1380 struct task_struct
*(*next
)(void *);
1384 static unsigned long
1385 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1386 unsigned long max_load_move
, struct sched_domain
*sd
,
1387 enum cpu_idle_type idle
, int *all_pinned
,
1388 int *this_best_prio
, struct rq_iterator
*iterator
);
1391 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1392 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1393 struct rq_iterator
*iterator
);
1396 #ifdef CONFIG_CGROUP_CPUACCT
1397 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1399 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1402 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1404 update_load_add(&rq
->load
, load
);
1407 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1409 update_load_sub(&rq
->load
, load
);
1412 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1413 typedef int (*tg_visitor
)(struct task_group
*, void *);
1416 * Iterate the full tree, calling @down when first entering a node and @up when
1417 * leaving it for the final time.
1419 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1421 struct task_group
*parent
, *child
;
1425 parent
= &root_task_group
;
1427 ret
= (*down
)(parent
, data
);
1430 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1437 ret
= (*up
)(parent
, data
);
1442 parent
= parent
->parent
;
1451 static int tg_nop(struct task_group
*tg
, void *data
)
1458 static unsigned long source_load(int cpu
, int type
);
1459 static unsigned long target_load(int cpu
, int type
);
1460 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1462 static unsigned long cpu_avg_load_per_task(int cpu
)
1464 struct rq
*rq
= cpu_rq(cpu
);
1465 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1468 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1470 rq
->avg_load_per_task
= 0;
1472 return rq
->avg_load_per_task
;
1475 #ifdef CONFIG_FAIR_GROUP_SCHED
1477 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1480 * Calculate and set the cpu's group shares.
1483 update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1484 unsigned long sd_shares
, unsigned long sd_rq_weight
)
1486 unsigned long shares
;
1487 unsigned long rq_weight
;
1492 rq_weight
= tg
->cfs_rq
[cpu
]->rq_weight
;
1495 * \Sum shares * rq_weight
1496 * shares = -----------------------
1500 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1501 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1503 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1504 sysctl_sched_shares_thresh
) {
1505 struct rq
*rq
= cpu_rq(cpu
);
1506 unsigned long flags
;
1508 spin_lock_irqsave(&rq
->lock
, flags
);
1509 tg
->cfs_rq
[cpu
]->shares
= shares
;
1511 __set_se_shares(tg
->se
[cpu
], shares
);
1512 spin_unlock_irqrestore(&rq
->lock
, flags
);
1517 * Re-compute the task group their per cpu shares over the given domain.
1518 * This needs to be done in a bottom-up fashion because the rq weight of a
1519 * parent group depends on the shares of its child groups.
1521 static int tg_shares_up(struct task_group
*tg
, void *data
)
1523 unsigned long weight
, rq_weight
= 0;
1524 unsigned long shares
= 0;
1525 struct sched_domain
*sd
= data
;
1528 for_each_cpu(i
, sched_domain_span(sd
)) {
1530 * If there are currently no tasks on the cpu pretend there
1531 * is one of average load so that when a new task gets to
1532 * run here it will not get delayed by group starvation.
1534 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1536 weight
= NICE_0_LOAD
;
1538 tg
->cfs_rq
[i
]->rq_weight
= weight
;
1539 rq_weight
+= weight
;
1540 shares
+= tg
->cfs_rq
[i
]->shares
;
1543 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1544 shares
= tg
->shares
;
1546 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1547 shares
= tg
->shares
;
1549 for_each_cpu(i
, sched_domain_span(sd
))
1550 update_group_shares_cpu(tg
, i
, shares
, rq_weight
);
1556 * Compute the cpu's hierarchical load factor for each task group.
1557 * This needs to be done in a top-down fashion because the load of a child
1558 * group is a fraction of its parents load.
1560 static int tg_load_down(struct task_group
*tg
, void *data
)
1563 long cpu
= (long)data
;
1566 load
= cpu_rq(cpu
)->load
.weight
;
1568 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1569 load
*= tg
->cfs_rq
[cpu
]->shares
;
1570 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1573 tg
->cfs_rq
[cpu
]->h_load
= load
;
1578 static void update_shares(struct sched_domain
*sd
)
1580 u64 now
= cpu_clock(raw_smp_processor_id());
1581 s64 elapsed
= now
- sd
->last_update
;
1583 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1584 sd
->last_update
= now
;
1585 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1589 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1591 spin_unlock(&rq
->lock
);
1593 spin_lock(&rq
->lock
);
1596 static void update_h_load(long cpu
)
1598 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1603 static inline void update_shares(struct sched_domain
*sd
)
1607 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1614 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1616 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1617 __releases(this_rq
->lock
)
1618 __acquires(busiest
->lock
)
1619 __acquires(this_rq
->lock
)
1623 if (unlikely(!irqs_disabled())) {
1624 /* printk() doesn't work good under rq->lock */
1625 spin_unlock(&this_rq
->lock
);
1628 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1629 if (busiest
< this_rq
) {
1630 spin_unlock(&this_rq
->lock
);
1631 spin_lock(&busiest
->lock
);
1632 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
1635 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
1640 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1641 __releases(busiest
->lock
)
1643 spin_unlock(&busiest
->lock
);
1644 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1648 #ifdef CONFIG_FAIR_GROUP_SCHED
1649 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1652 cfs_rq
->shares
= shares
;
1657 #include "sched_stats.h"
1658 #include "sched_idletask.c"
1659 #include "sched_fair.c"
1660 #include "sched_rt.c"
1661 #ifdef CONFIG_SCHED_DEBUG
1662 # include "sched_debug.c"
1665 #define sched_class_highest (&rt_sched_class)
1666 #define for_each_class(class) \
1667 for (class = sched_class_highest; class; class = class->next)
1669 static void inc_nr_running(struct rq
*rq
)
1674 static void dec_nr_running(struct rq
*rq
)
1679 static void set_load_weight(struct task_struct
*p
)
1681 if (task_has_rt_policy(p
)) {
1682 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1683 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1688 * SCHED_IDLE tasks get minimal weight:
1690 if (p
->policy
== SCHED_IDLE
) {
1691 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1692 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1696 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1697 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1700 static void update_avg(u64
*avg
, u64 sample
)
1702 s64 diff
= sample
- *avg
;
1706 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1709 p
->se
.start_runtime
= p
->se
.sum_exec_runtime
;
1711 sched_info_queued(p
);
1712 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1716 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1719 if (p
->se
.last_wakeup
) {
1720 update_avg(&p
->se
.avg_overlap
,
1721 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1722 p
->se
.last_wakeup
= 0;
1724 update_avg(&p
->se
.avg_wakeup
,
1725 sysctl_sched_wakeup_granularity
);
1729 sched_info_dequeued(p
);
1730 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1735 * __normal_prio - return the priority that is based on the static prio
1737 static inline int __normal_prio(struct task_struct
*p
)
1739 return p
->static_prio
;
1743 * Calculate the expected normal priority: i.e. priority
1744 * without taking RT-inheritance into account. Might be
1745 * boosted by interactivity modifiers. Changes upon fork,
1746 * setprio syscalls, and whenever the interactivity
1747 * estimator recalculates.
1749 static inline int normal_prio(struct task_struct
*p
)
1753 if (task_has_rt_policy(p
))
1754 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1756 prio
= __normal_prio(p
);
1761 * Calculate the current priority, i.e. the priority
1762 * taken into account by the scheduler. This value might
1763 * be boosted by RT tasks, or might be boosted by
1764 * interactivity modifiers. Will be RT if the task got
1765 * RT-boosted. If not then it returns p->normal_prio.
1767 static int effective_prio(struct task_struct
*p
)
1769 p
->normal_prio
= normal_prio(p
);
1771 * If we are RT tasks or we were boosted to RT priority,
1772 * keep the priority unchanged. Otherwise, update priority
1773 * to the normal priority:
1775 if (!rt_prio(p
->prio
))
1776 return p
->normal_prio
;
1781 * activate_task - move a task to the runqueue.
1783 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1785 if (task_contributes_to_load(p
))
1786 rq
->nr_uninterruptible
--;
1788 enqueue_task(rq
, p
, wakeup
);
1793 * deactivate_task - remove a task from the runqueue.
1795 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1797 if (task_contributes_to_load(p
))
1798 rq
->nr_uninterruptible
++;
1800 dequeue_task(rq
, p
, sleep
);
1805 * task_curr - is this task currently executing on a CPU?
1806 * @p: the task in question.
1808 inline int task_curr(const struct task_struct
*p
)
1810 return cpu_curr(task_cpu(p
)) == p
;
1813 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1815 set_task_rq(p
, cpu
);
1818 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1819 * successfuly executed on another CPU. We must ensure that updates of
1820 * per-task data have been completed by this moment.
1823 task_thread_info(p
)->cpu
= cpu
;
1827 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1828 const struct sched_class
*prev_class
,
1829 int oldprio
, int running
)
1831 if (prev_class
!= p
->sched_class
) {
1832 if (prev_class
->switched_from
)
1833 prev_class
->switched_from(rq
, p
, running
);
1834 p
->sched_class
->switched_to(rq
, p
, running
);
1836 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1841 /* Used instead of source_load when we know the type == 0 */
1842 static unsigned long weighted_cpuload(const int cpu
)
1844 return cpu_rq(cpu
)->load
.weight
;
1848 * Is this task likely cache-hot:
1851 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1856 * Buddy candidates are cache hot:
1858 if (sched_feat(CACHE_HOT_BUDDY
) &&
1859 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
1860 &p
->se
== cfs_rq_of(&p
->se
)->last
))
1863 if (p
->sched_class
!= &fair_sched_class
)
1866 if (sysctl_sched_migration_cost
== -1)
1868 if (sysctl_sched_migration_cost
== 0)
1871 delta
= now
- p
->se
.exec_start
;
1873 return delta
< (s64
)sysctl_sched_migration_cost
;
1877 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1879 int old_cpu
= task_cpu(p
);
1880 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1881 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1882 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1885 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1887 trace_sched_migrate_task(p
, task_cpu(p
), new_cpu
);
1889 #ifdef CONFIG_SCHEDSTATS
1890 if (p
->se
.wait_start
)
1891 p
->se
.wait_start
-= clock_offset
;
1892 if (p
->se
.sleep_start
)
1893 p
->se
.sleep_start
-= clock_offset
;
1894 if (p
->se
.block_start
)
1895 p
->se
.block_start
-= clock_offset
;
1896 if (old_cpu
!= new_cpu
) {
1897 schedstat_inc(p
, se
.nr_migrations
);
1898 if (task_hot(p
, old_rq
->clock
, NULL
))
1899 schedstat_inc(p
, se
.nr_forced2_migrations
);
1902 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1903 new_cfsrq
->min_vruntime
;
1905 __set_task_cpu(p
, new_cpu
);
1908 struct migration_req
{
1909 struct list_head list
;
1911 struct task_struct
*task
;
1914 struct completion done
;
1918 * The task's runqueue lock must be held.
1919 * Returns true if you have to wait for migration thread.
1922 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1924 struct rq
*rq
= task_rq(p
);
1927 * If the task is not on a runqueue (and not running), then
1928 * it is sufficient to simply update the task's cpu field.
1930 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1931 set_task_cpu(p
, dest_cpu
);
1935 init_completion(&req
->done
);
1937 req
->dest_cpu
= dest_cpu
;
1938 list_add(&req
->list
, &rq
->migration_queue
);
1944 * wait_task_inactive - wait for a thread to unschedule.
1946 * If @match_state is nonzero, it's the @p->state value just checked and
1947 * not expected to change. If it changes, i.e. @p might have woken up,
1948 * then return zero. When we succeed in waiting for @p to be off its CPU,
1949 * we return a positive number (its total switch count). If a second call
1950 * a short while later returns the same number, the caller can be sure that
1951 * @p has remained unscheduled the whole time.
1953 * The caller must ensure that the task *will* unschedule sometime soon,
1954 * else this function might spin for a *long* time. This function can't
1955 * be called with interrupts off, or it may introduce deadlock with
1956 * smp_call_function() if an IPI is sent by the same process we are
1957 * waiting to become inactive.
1959 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1961 unsigned long flags
;
1968 * We do the initial early heuristics without holding
1969 * any task-queue locks at all. We'll only try to get
1970 * the runqueue lock when things look like they will
1976 * If the task is actively running on another CPU
1977 * still, just relax and busy-wait without holding
1980 * NOTE! Since we don't hold any locks, it's not
1981 * even sure that "rq" stays as the right runqueue!
1982 * But we don't care, since "task_running()" will
1983 * return false if the runqueue has changed and p
1984 * is actually now running somewhere else!
1986 while (task_running(rq
, p
)) {
1987 if (match_state
&& unlikely(p
->state
!= match_state
))
1993 * Ok, time to look more closely! We need the rq
1994 * lock now, to be *sure*. If we're wrong, we'll
1995 * just go back and repeat.
1997 rq
= task_rq_lock(p
, &flags
);
1998 trace_sched_wait_task(rq
, p
);
1999 running
= task_running(rq
, p
);
2000 on_rq
= p
->se
.on_rq
;
2002 if (!match_state
|| p
->state
== match_state
)
2003 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2004 task_rq_unlock(rq
, &flags
);
2007 * If it changed from the expected state, bail out now.
2009 if (unlikely(!ncsw
))
2013 * Was it really running after all now that we
2014 * checked with the proper locks actually held?
2016 * Oops. Go back and try again..
2018 if (unlikely(running
)) {
2024 * It's not enough that it's not actively running,
2025 * it must be off the runqueue _entirely_, and not
2028 * So if it wa still runnable (but just not actively
2029 * running right now), it's preempted, and we should
2030 * yield - it could be a while.
2032 if (unlikely(on_rq
)) {
2033 schedule_timeout_uninterruptible(1);
2038 * Ahh, all good. It wasn't running, and it wasn't
2039 * runnable, which means that it will never become
2040 * running in the future either. We're all done!
2049 * kick_process - kick a running thread to enter/exit the kernel
2050 * @p: the to-be-kicked thread
2052 * Cause a process which is running on another CPU to enter
2053 * kernel-mode, without any delay. (to get signals handled.)
2055 * NOTE: this function doesnt have to take the runqueue lock,
2056 * because all it wants to ensure is that the remote task enters
2057 * the kernel. If the IPI races and the task has been migrated
2058 * to another CPU then no harm is done and the purpose has been
2061 void kick_process(struct task_struct
*p
)
2067 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2068 smp_send_reschedule(cpu
);
2073 * Return a low guess at the load of a migration-source cpu weighted
2074 * according to the scheduling class and "nice" value.
2076 * We want to under-estimate the load of migration sources, to
2077 * balance conservatively.
2079 static unsigned long source_load(int cpu
, int type
)
2081 struct rq
*rq
= cpu_rq(cpu
);
2082 unsigned long total
= weighted_cpuload(cpu
);
2084 if (type
== 0 || !sched_feat(LB_BIAS
))
2087 return min(rq
->cpu_load
[type
-1], total
);
2091 * Return a high guess at the load of a migration-target cpu weighted
2092 * according to the scheduling class and "nice" value.
2094 static unsigned long target_load(int cpu
, int type
)
2096 struct rq
*rq
= cpu_rq(cpu
);
2097 unsigned long total
= weighted_cpuload(cpu
);
2099 if (type
== 0 || !sched_feat(LB_BIAS
))
2102 return max(rq
->cpu_load
[type
-1], total
);
2106 * find_idlest_group finds and returns the least busy CPU group within the
2109 static struct sched_group
*
2110 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2112 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2113 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2114 int load_idx
= sd
->forkexec_idx
;
2115 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2118 unsigned long load
, avg_load
;
2122 /* Skip over this group if it has no CPUs allowed */
2123 if (!cpumask_intersects(sched_group_cpus(group
),
2127 local_group
= cpumask_test_cpu(this_cpu
,
2128 sched_group_cpus(group
));
2130 /* Tally up the load of all CPUs in the group */
2133 for_each_cpu(i
, sched_group_cpus(group
)) {
2134 /* Bias balancing toward cpus of our domain */
2136 load
= source_load(i
, load_idx
);
2138 load
= target_load(i
, load_idx
);
2143 /* Adjust by relative CPU power of the group */
2144 avg_load
= sg_div_cpu_power(group
,
2145 avg_load
* SCHED_LOAD_SCALE
);
2148 this_load
= avg_load
;
2150 } else if (avg_load
< min_load
) {
2151 min_load
= avg_load
;
2154 } while (group
= group
->next
, group
!= sd
->groups
);
2156 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2162 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2165 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
2167 unsigned long load
, min_load
= ULONG_MAX
;
2171 /* Traverse only the allowed CPUs */
2172 for_each_cpu_and(i
, sched_group_cpus(group
), &p
->cpus_allowed
) {
2173 load
= weighted_cpuload(i
);
2175 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2185 * sched_balance_self: balance the current task (running on cpu) in domains
2186 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2189 * Balance, ie. select the least loaded group.
2191 * Returns the target CPU number, or the same CPU if no balancing is needed.
2193 * preempt must be disabled.
2195 static int sched_balance_self(int cpu
, int flag
)
2197 struct task_struct
*t
= current
;
2198 struct sched_domain
*tmp
, *sd
= NULL
;
2200 for_each_domain(cpu
, tmp
) {
2202 * If power savings logic is enabled for a domain, stop there.
2204 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2206 if (tmp
->flags
& flag
)
2214 struct sched_group
*group
;
2215 int new_cpu
, weight
;
2217 if (!(sd
->flags
& flag
)) {
2222 group
= find_idlest_group(sd
, t
, cpu
);
2228 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
2229 if (new_cpu
== -1 || new_cpu
== cpu
) {
2230 /* Now try balancing at a lower domain level of cpu */
2235 /* Now try balancing at a lower domain level of new_cpu */
2237 weight
= cpumask_weight(sched_domain_span(sd
));
2239 for_each_domain(cpu
, tmp
) {
2240 if (weight
<= cpumask_weight(sched_domain_span(tmp
)))
2242 if (tmp
->flags
& flag
)
2245 /* while loop will break here if sd == NULL */
2251 #endif /* CONFIG_SMP */
2254 * try_to_wake_up - wake up a thread
2255 * @p: the to-be-woken-up thread
2256 * @state: the mask of task states that can be woken
2257 * @sync: do a synchronous wakeup?
2259 * Put it on the run-queue if it's not already there. The "current"
2260 * thread is always on the run-queue (except when the actual
2261 * re-schedule is in progress), and as such you're allowed to do
2262 * the simpler "current->state = TASK_RUNNING" to mark yourself
2263 * runnable without the overhead of this.
2265 * returns failure only if the task is already active.
2267 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2269 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2270 unsigned long flags
;
2274 if (!sched_feat(SYNC_WAKEUPS
))
2278 if (sched_feat(LB_WAKEUP_UPDATE
)) {
2279 struct sched_domain
*sd
;
2281 this_cpu
= raw_smp_processor_id();
2284 for_each_domain(this_cpu
, sd
) {
2285 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2294 rq
= task_rq_lock(p
, &flags
);
2295 update_rq_clock(rq
);
2296 old_state
= p
->state
;
2297 if (!(old_state
& state
))
2305 this_cpu
= smp_processor_id();
2308 if (unlikely(task_running(rq
, p
)))
2311 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2312 if (cpu
!= orig_cpu
) {
2313 set_task_cpu(p
, cpu
);
2314 task_rq_unlock(rq
, &flags
);
2315 /* might preempt at this point */
2316 rq
= task_rq_lock(p
, &flags
);
2317 old_state
= p
->state
;
2318 if (!(old_state
& state
))
2323 this_cpu
= smp_processor_id();
2327 #ifdef CONFIG_SCHEDSTATS
2328 schedstat_inc(rq
, ttwu_count
);
2329 if (cpu
== this_cpu
)
2330 schedstat_inc(rq
, ttwu_local
);
2332 struct sched_domain
*sd
;
2333 for_each_domain(this_cpu
, sd
) {
2334 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2335 schedstat_inc(sd
, ttwu_wake_remote
);
2340 #endif /* CONFIG_SCHEDSTATS */
2343 #endif /* CONFIG_SMP */
2344 schedstat_inc(p
, se
.nr_wakeups
);
2346 schedstat_inc(p
, se
.nr_wakeups_sync
);
2347 if (orig_cpu
!= cpu
)
2348 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2349 if (cpu
== this_cpu
)
2350 schedstat_inc(p
, se
.nr_wakeups_local
);
2352 schedstat_inc(p
, se
.nr_wakeups_remote
);
2353 activate_task(rq
, p
, 1);
2357 * Only attribute actual wakeups done by this task.
2359 if (!in_interrupt()) {
2360 struct sched_entity
*se
= ¤t
->se
;
2361 u64 sample
= se
->sum_exec_runtime
;
2363 if (se
->last_wakeup
)
2364 sample
-= se
->last_wakeup
;
2366 sample
-= se
->start_runtime
;
2367 update_avg(&se
->avg_wakeup
, sample
);
2369 se
->last_wakeup
= se
->sum_exec_runtime
;
2373 trace_sched_wakeup(rq
, p
, success
);
2374 check_preempt_curr(rq
, p
, sync
);
2376 p
->state
= TASK_RUNNING
;
2378 if (p
->sched_class
->task_wake_up
)
2379 p
->sched_class
->task_wake_up(rq
, p
);
2382 task_rq_unlock(rq
, &flags
);
2387 int wake_up_process(struct task_struct
*p
)
2389 return try_to_wake_up(p
, TASK_ALL
, 0);
2391 EXPORT_SYMBOL(wake_up_process
);
2393 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2395 return try_to_wake_up(p
, state
, 0);
2399 * Perform scheduler related setup for a newly forked process p.
2400 * p is forked by current.
2402 * __sched_fork() is basic setup used by init_idle() too:
2404 static void __sched_fork(struct task_struct
*p
)
2406 p
->se
.exec_start
= 0;
2407 p
->se
.sum_exec_runtime
= 0;
2408 p
->se
.prev_sum_exec_runtime
= 0;
2409 p
->se
.last_wakeup
= 0;
2410 p
->se
.avg_overlap
= 0;
2411 p
->se
.start_runtime
= 0;
2412 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
2414 #ifdef CONFIG_SCHEDSTATS
2415 p
->se
.wait_start
= 0;
2416 p
->se
.sum_sleep_runtime
= 0;
2417 p
->se
.sleep_start
= 0;
2418 p
->se
.block_start
= 0;
2419 p
->se
.sleep_max
= 0;
2420 p
->se
.block_max
= 0;
2422 p
->se
.slice_max
= 0;
2426 INIT_LIST_HEAD(&p
->rt
.run_list
);
2428 INIT_LIST_HEAD(&p
->se
.group_node
);
2430 #ifdef CONFIG_PREEMPT_NOTIFIERS
2431 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2435 * We mark the process as running here, but have not actually
2436 * inserted it onto the runqueue yet. This guarantees that
2437 * nobody will actually run it, and a signal or other external
2438 * event cannot wake it up and insert it on the runqueue either.
2440 p
->state
= TASK_RUNNING
;
2444 * fork()/clone()-time setup:
2446 void sched_fork(struct task_struct
*p
, int clone_flags
)
2448 int cpu
= get_cpu();
2453 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2455 set_task_cpu(p
, cpu
);
2458 * Make sure we do not leak PI boosting priority to the child:
2460 p
->prio
= current
->normal_prio
;
2461 if (!rt_prio(p
->prio
))
2462 p
->sched_class
= &fair_sched_class
;
2464 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2465 if (likely(sched_info_on()))
2466 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2468 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2471 #ifdef CONFIG_PREEMPT
2472 /* Want to start with kernel preemption disabled. */
2473 task_thread_info(p
)->preempt_count
= 1;
2479 * wake_up_new_task - wake up a newly created task for the first time.
2481 * This function will do some initial scheduler statistics housekeeping
2482 * that must be done for every newly created context, then puts the task
2483 * on the runqueue and wakes it.
2485 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2487 unsigned long flags
;
2490 rq
= task_rq_lock(p
, &flags
);
2491 BUG_ON(p
->state
!= TASK_RUNNING
);
2492 update_rq_clock(rq
);
2494 p
->prio
= effective_prio(p
);
2496 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2497 activate_task(rq
, p
, 0);
2500 * Let the scheduling class do new task startup
2501 * management (if any):
2503 p
->sched_class
->task_new(rq
, p
);
2506 trace_sched_wakeup_new(rq
, p
, 1);
2507 check_preempt_curr(rq
, p
, 0);
2509 if (p
->sched_class
->task_wake_up
)
2510 p
->sched_class
->task_wake_up(rq
, p
);
2512 task_rq_unlock(rq
, &flags
);
2515 #ifdef CONFIG_PREEMPT_NOTIFIERS
2518 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2519 * @notifier: notifier struct to register
2521 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2523 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2525 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2528 * preempt_notifier_unregister - no longer interested in preemption notifications
2529 * @notifier: notifier struct to unregister
2531 * This is safe to call from within a preemption notifier.
2533 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2535 hlist_del(¬ifier
->link
);
2537 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2539 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2541 struct preempt_notifier
*notifier
;
2542 struct hlist_node
*node
;
2544 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2545 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2549 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2550 struct task_struct
*next
)
2552 struct preempt_notifier
*notifier
;
2553 struct hlist_node
*node
;
2555 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2556 notifier
->ops
->sched_out(notifier
, next
);
2559 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2561 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2566 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2567 struct task_struct
*next
)
2571 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2574 * prepare_task_switch - prepare to switch tasks
2575 * @rq: the runqueue preparing to switch
2576 * @prev: the current task that is being switched out
2577 * @next: the task we are going to switch to.
2579 * This is called with the rq lock held and interrupts off. It must
2580 * be paired with a subsequent finish_task_switch after the context
2583 * prepare_task_switch sets up locking and calls architecture specific
2587 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2588 struct task_struct
*next
)
2590 fire_sched_out_preempt_notifiers(prev
, next
);
2591 prepare_lock_switch(rq
, next
);
2592 prepare_arch_switch(next
);
2596 * finish_task_switch - clean up after a task-switch
2597 * @rq: runqueue associated with task-switch
2598 * @prev: the thread we just switched away from.
2600 * finish_task_switch must be called after the context switch, paired
2601 * with a prepare_task_switch call before the context switch.
2602 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2603 * and do any other architecture-specific cleanup actions.
2605 * Note that we may have delayed dropping an mm in context_switch(). If
2606 * so, we finish that here outside of the runqueue lock. (Doing it
2607 * with the lock held can cause deadlocks; see schedule() for
2610 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2611 __releases(rq
->lock
)
2613 struct mm_struct
*mm
= rq
->prev_mm
;
2619 * A task struct has one reference for the use as "current".
2620 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2621 * schedule one last time. The schedule call will never return, and
2622 * the scheduled task must drop that reference.
2623 * The test for TASK_DEAD must occur while the runqueue locks are
2624 * still held, otherwise prev could be scheduled on another cpu, die
2625 * there before we look at prev->state, and then the reference would
2627 * Manfred Spraul <manfred@colorfullife.com>
2629 prev_state
= prev
->state
;
2630 finish_arch_switch(prev
);
2631 finish_lock_switch(rq
, prev
);
2633 if (current
->sched_class
->post_schedule
)
2634 current
->sched_class
->post_schedule(rq
);
2637 fire_sched_in_preempt_notifiers(current
);
2640 if (unlikely(prev_state
== TASK_DEAD
)) {
2642 * Remove function-return probe instances associated with this
2643 * task and put them back on the free list.
2645 kprobe_flush_task(prev
);
2646 put_task_struct(prev
);
2651 * schedule_tail - first thing a freshly forked thread must call.
2652 * @prev: the thread we just switched away from.
2654 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2655 __releases(rq
->lock
)
2657 struct rq
*rq
= this_rq();
2659 finish_task_switch(rq
, prev
);
2660 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2661 /* In this case, finish_task_switch does not reenable preemption */
2664 if (current
->set_child_tid
)
2665 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2669 * context_switch - switch to the new MM and the new
2670 * thread's register state.
2673 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2674 struct task_struct
*next
)
2676 struct mm_struct
*mm
, *oldmm
;
2678 prepare_task_switch(rq
, prev
, next
);
2679 trace_sched_switch(rq
, prev
, next
);
2681 oldmm
= prev
->active_mm
;
2683 * For paravirt, this is coupled with an exit in switch_to to
2684 * combine the page table reload and the switch backend into
2687 arch_enter_lazy_cpu_mode();
2689 if (unlikely(!mm
)) {
2690 next
->active_mm
= oldmm
;
2691 atomic_inc(&oldmm
->mm_count
);
2692 enter_lazy_tlb(oldmm
, next
);
2694 switch_mm(oldmm
, mm
, next
);
2696 if (unlikely(!prev
->mm
)) {
2697 prev
->active_mm
= NULL
;
2698 rq
->prev_mm
= oldmm
;
2701 * Since the runqueue lock will be released by the next
2702 * task (which is an invalid locking op but in the case
2703 * of the scheduler it's an obvious special-case), so we
2704 * do an early lockdep release here:
2706 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2707 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2710 /* Here we just switch the register state and the stack. */
2711 switch_to(prev
, next
, prev
);
2715 * this_rq must be evaluated again because prev may have moved
2716 * CPUs since it called schedule(), thus the 'rq' on its stack
2717 * frame will be invalid.
2719 finish_task_switch(this_rq(), prev
);
2723 * nr_running, nr_uninterruptible and nr_context_switches:
2725 * externally visible scheduler statistics: current number of runnable
2726 * threads, current number of uninterruptible-sleeping threads, total
2727 * number of context switches performed since bootup.
2729 unsigned long nr_running(void)
2731 unsigned long i
, sum
= 0;
2733 for_each_online_cpu(i
)
2734 sum
+= cpu_rq(i
)->nr_running
;
2739 unsigned long nr_uninterruptible(void)
2741 unsigned long i
, sum
= 0;
2743 for_each_possible_cpu(i
)
2744 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2747 * Since we read the counters lockless, it might be slightly
2748 * inaccurate. Do not allow it to go below zero though:
2750 if (unlikely((long)sum
< 0))
2756 unsigned long long nr_context_switches(void)
2759 unsigned long long sum
= 0;
2761 for_each_possible_cpu(i
)
2762 sum
+= cpu_rq(i
)->nr_switches
;
2767 unsigned long nr_iowait(void)
2769 unsigned long i
, sum
= 0;
2771 for_each_possible_cpu(i
)
2772 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2777 unsigned long nr_active(void)
2779 unsigned long i
, running
= 0, uninterruptible
= 0;
2781 for_each_online_cpu(i
) {
2782 running
+= cpu_rq(i
)->nr_running
;
2783 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2786 if (unlikely((long)uninterruptible
< 0))
2787 uninterruptible
= 0;
2789 return running
+ uninterruptible
;
2793 * Update rq->cpu_load[] statistics. This function is usually called every
2794 * scheduler tick (TICK_NSEC).
2796 static void update_cpu_load(struct rq
*this_rq
)
2798 unsigned long this_load
= this_rq
->load
.weight
;
2801 this_rq
->nr_load_updates
++;
2803 /* Update our load: */
2804 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2805 unsigned long old_load
, new_load
;
2807 /* scale is effectively 1 << i now, and >> i divides by scale */
2809 old_load
= this_rq
->cpu_load
[i
];
2810 new_load
= this_load
;
2812 * Round up the averaging division if load is increasing. This
2813 * prevents us from getting stuck on 9 if the load is 10, for
2816 if (new_load
> old_load
)
2817 new_load
+= scale
-1;
2818 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2825 * double_rq_lock - safely lock two runqueues
2827 * Note this does not disable interrupts like task_rq_lock,
2828 * you need to do so manually before calling.
2830 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2831 __acquires(rq1
->lock
)
2832 __acquires(rq2
->lock
)
2834 BUG_ON(!irqs_disabled());
2836 spin_lock(&rq1
->lock
);
2837 __acquire(rq2
->lock
); /* Fake it out ;) */
2840 spin_lock(&rq1
->lock
);
2841 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
2843 spin_lock(&rq2
->lock
);
2844 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
2847 update_rq_clock(rq1
);
2848 update_rq_clock(rq2
);
2852 * double_rq_unlock - safely unlock two runqueues
2854 * Note this does not restore interrupts like task_rq_unlock,
2855 * you need to do so manually after calling.
2857 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2858 __releases(rq1
->lock
)
2859 __releases(rq2
->lock
)
2861 spin_unlock(&rq1
->lock
);
2863 spin_unlock(&rq2
->lock
);
2865 __release(rq2
->lock
);
2869 * If dest_cpu is allowed for this process, migrate the task to it.
2870 * This is accomplished by forcing the cpu_allowed mask to only
2871 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2872 * the cpu_allowed mask is restored.
2874 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2876 struct migration_req req
;
2877 unsigned long flags
;
2880 rq
= task_rq_lock(p
, &flags
);
2881 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
2882 || unlikely(!cpu_active(dest_cpu
)))
2885 /* force the process onto the specified CPU */
2886 if (migrate_task(p
, dest_cpu
, &req
)) {
2887 /* Need to wait for migration thread (might exit: take ref). */
2888 struct task_struct
*mt
= rq
->migration_thread
;
2890 get_task_struct(mt
);
2891 task_rq_unlock(rq
, &flags
);
2892 wake_up_process(mt
);
2893 put_task_struct(mt
);
2894 wait_for_completion(&req
.done
);
2899 task_rq_unlock(rq
, &flags
);
2903 * sched_exec - execve() is a valuable balancing opportunity, because at
2904 * this point the task has the smallest effective memory and cache footprint.
2906 void sched_exec(void)
2908 int new_cpu
, this_cpu
= get_cpu();
2909 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2911 if (new_cpu
!= this_cpu
)
2912 sched_migrate_task(current
, new_cpu
);
2916 * pull_task - move a task from a remote runqueue to the local runqueue.
2917 * Both runqueues must be locked.
2919 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2920 struct rq
*this_rq
, int this_cpu
)
2922 deactivate_task(src_rq
, p
, 0);
2923 set_task_cpu(p
, this_cpu
);
2924 activate_task(this_rq
, p
, 0);
2926 * Note that idle threads have a prio of MAX_PRIO, for this test
2927 * to be always true for them.
2929 check_preempt_curr(this_rq
, p
, 0);
2933 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2936 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2937 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2941 * We do not migrate tasks that are:
2942 * 1) running (obviously), or
2943 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2944 * 3) are cache-hot on their current CPU.
2946 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
2947 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2952 if (task_running(rq
, p
)) {
2953 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2958 * Aggressive migration if:
2959 * 1) task is cache cold, or
2960 * 2) too many balance attempts have failed.
2963 if (!task_hot(p
, rq
->clock
, sd
) ||
2964 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2965 #ifdef CONFIG_SCHEDSTATS
2966 if (task_hot(p
, rq
->clock
, sd
)) {
2967 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2968 schedstat_inc(p
, se
.nr_forced_migrations
);
2974 if (task_hot(p
, rq
->clock
, sd
)) {
2975 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2981 static unsigned long
2982 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2983 unsigned long max_load_move
, struct sched_domain
*sd
,
2984 enum cpu_idle_type idle
, int *all_pinned
,
2985 int *this_best_prio
, struct rq_iterator
*iterator
)
2987 int loops
= 0, pulled
= 0, pinned
= 0;
2988 struct task_struct
*p
;
2989 long rem_load_move
= max_load_move
;
2991 if (max_load_move
== 0)
2997 * Start the load-balancing iterator:
2999 p
= iterator
->start(iterator
->arg
);
3001 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3004 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
3005 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3006 p
= iterator
->next(iterator
->arg
);
3010 pull_task(busiest
, p
, this_rq
, this_cpu
);
3012 rem_load_move
-= p
->se
.load
.weight
;
3015 * We only want to steal up to the prescribed amount of weighted load.
3017 if (rem_load_move
> 0) {
3018 if (p
->prio
< *this_best_prio
)
3019 *this_best_prio
= p
->prio
;
3020 p
= iterator
->next(iterator
->arg
);
3025 * Right now, this is one of only two places pull_task() is called,
3026 * so we can safely collect pull_task() stats here rather than
3027 * inside pull_task().
3029 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3032 *all_pinned
= pinned
;
3034 return max_load_move
- rem_load_move
;
3038 * move_tasks tries to move up to max_load_move weighted load from busiest to
3039 * this_rq, as part of a balancing operation within domain "sd".
3040 * Returns 1 if successful and 0 otherwise.
3042 * Called with both runqueues locked.
3044 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3045 unsigned long max_load_move
,
3046 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3049 const struct sched_class
*class = sched_class_highest
;
3050 unsigned long total_load_moved
= 0;
3051 int this_best_prio
= this_rq
->curr
->prio
;
3055 class->load_balance(this_rq
, this_cpu
, busiest
,
3056 max_load_move
- total_load_moved
,
3057 sd
, idle
, all_pinned
, &this_best_prio
);
3058 class = class->next
;
3060 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3063 } while (class && max_load_move
> total_load_moved
);
3065 return total_load_moved
> 0;
3069 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3070 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3071 struct rq_iterator
*iterator
)
3073 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3077 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3078 pull_task(busiest
, p
, this_rq
, this_cpu
);
3080 * Right now, this is only the second place pull_task()
3081 * is called, so we can safely collect pull_task()
3082 * stats here rather than inside pull_task().
3084 schedstat_inc(sd
, lb_gained
[idle
]);
3088 p
= iterator
->next(iterator
->arg
);
3095 * move_one_task tries to move exactly one task from busiest to this_rq, as
3096 * part of active balancing operations within "domain".
3097 * Returns 1 if successful and 0 otherwise.
3099 * Called with both runqueues locked.
3101 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3102 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3104 const struct sched_class
*class;
3106 for (class = sched_class_highest
; class; class = class->next
)
3107 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3114 * find_busiest_group finds and returns the busiest CPU group within the
3115 * domain. It calculates and returns the amount of weighted load which
3116 * should be moved to restore balance via the imbalance parameter.
3118 static struct sched_group
*
3119 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3120 unsigned long *imbalance
, enum cpu_idle_type idle
,
3121 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
3123 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
3124 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
3125 unsigned long max_pull
;
3126 unsigned long busiest_load_per_task
, busiest_nr_running
;
3127 unsigned long this_load_per_task
, this_nr_running
;
3128 int load_idx
, group_imb
= 0;
3129 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3130 int power_savings_balance
= 1;
3131 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
3132 unsigned long min_nr_running
= ULONG_MAX
;
3133 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
3136 max_load
= this_load
= total_load
= total_pwr
= 0;
3137 busiest_load_per_task
= busiest_nr_running
= 0;
3138 this_load_per_task
= this_nr_running
= 0;
3140 if (idle
== CPU_NOT_IDLE
)
3141 load_idx
= sd
->busy_idx
;
3142 else if (idle
== CPU_NEWLY_IDLE
)
3143 load_idx
= sd
->newidle_idx
;
3145 load_idx
= sd
->idle_idx
;
3148 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
3151 int __group_imb
= 0;
3152 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3153 unsigned long sum_nr_running
, sum_weighted_load
;
3154 unsigned long sum_avg_load_per_task
;
3155 unsigned long avg_load_per_task
;
3157 local_group
= cpumask_test_cpu(this_cpu
,
3158 sched_group_cpus(group
));
3161 balance_cpu
= cpumask_first(sched_group_cpus(group
));
3163 /* Tally up the load of all CPUs in the group */
3164 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
3165 sum_avg_load_per_task
= avg_load_per_task
= 0;
3168 min_cpu_load
= ~0UL;
3170 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3171 struct rq
*rq
= cpu_rq(i
);
3173 if (*sd_idle
&& rq
->nr_running
)
3176 /* Bias balancing toward cpus of our domain */
3178 if (idle_cpu(i
) && !first_idle_cpu
) {
3183 load
= target_load(i
, load_idx
);
3185 load
= source_load(i
, load_idx
);
3186 if (load
> max_cpu_load
)
3187 max_cpu_load
= load
;
3188 if (min_cpu_load
> load
)
3189 min_cpu_load
= load
;
3193 sum_nr_running
+= rq
->nr_running
;
3194 sum_weighted_load
+= weighted_cpuload(i
);
3196 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3200 * First idle cpu or the first cpu(busiest) in this sched group
3201 * is eligible for doing load balancing at this and above
3202 * domains. In the newly idle case, we will allow all the cpu's
3203 * to do the newly idle load balance.
3205 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3206 balance_cpu
!= this_cpu
&& balance
) {
3211 total_load
+= avg_load
;
3212 total_pwr
+= group
->__cpu_power
;
3214 /* Adjust by relative CPU power of the group */
3215 avg_load
= sg_div_cpu_power(group
,
3216 avg_load
* SCHED_LOAD_SCALE
);
3220 * Consider the group unbalanced when the imbalance is larger
3221 * than the average weight of two tasks.
3223 * APZ: with cgroup the avg task weight can vary wildly and
3224 * might not be a suitable number - should we keep a
3225 * normalized nr_running number somewhere that negates
3228 avg_load_per_task
= sg_div_cpu_power(group
,
3229 sum_avg_load_per_task
* SCHED_LOAD_SCALE
);
3231 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3234 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3237 this_load
= avg_load
;
3239 this_nr_running
= sum_nr_running
;
3240 this_load_per_task
= sum_weighted_load
;
3241 } else if (avg_load
> max_load
&&
3242 (sum_nr_running
> group_capacity
|| __group_imb
)) {
3243 max_load
= avg_load
;
3245 busiest_nr_running
= sum_nr_running
;
3246 busiest_load_per_task
= sum_weighted_load
;
3247 group_imb
= __group_imb
;
3250 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3252 * Busy processors will not participate in power savings
3255 if (idle
== CPU_NOT_IDLE
||
3256 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3260 * If the local group is idle or completely loaded
3261 * no need to do power savings balance at this domain
3263 if (local_group
&& (this_nr_running
>= group_capacity
||
3265 power_savings_balance
= 0;
3268 * If a group is already running at full capacity or idle,
3269 * don't include that group in power savings calculations
3271 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
3276 * Calculate the group which has the least non-idle load.
3277 * This is the group from where we need to pick up the load
3280 if ((sum_nr_running
< min_nr_running
) ||
3281 (sum_nr_running
== min_nr_running
&&
3282 cpumask_first(sched_group_cpus(group
)) >
3283 cpumask_first(sched_group_cpus(group_min
)))) {
3285 min_nr_running
= sum_nr_running
;
3286 min_load_per_task
= sum_weighted_load
/
3291 * Calculate the group which is almost near its
3292 * capacity but still has some space to pick up some load
3293 * from other group and save more power
3295 if (sum_nr_running
<= group_capacity
- 1) {
3296 if (sum_nr_running
> leader_nr_running
||
3297 (sum_nr_running
== leader_nr_running
&&
3298 cpumask_first(sched_group_cpus(group
)) <
3299 cpumask_first(sched_group_cpus(group_leader
)))) {
3300 group_leader
= group
;
3301 leader_nr_running
= sum_nr_running
;
3306 group
= group
->next
;
3307 } while (group
!= sd
->groups
);
3309 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
3312 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
3314 if (this_load
>= avg_load
||
3315 100*max_load
<= sd
->imbalance_pct
*this_load
)
3318 busiest_load_per_task
/= busiest_nr_running
;
3320 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
3323 * We're trying to get all the cpus to the average_load, so we don't
3324 * want to push ourselves above the average load, nor do we wish to
3325 * reduce the max loaded cpu below the average load, as either of these
3326 * actions would just result in more rebalancing later, and ping-pong
3327 * tasks around. Thus we look for the minimum possible imbalance.
3328 * Negative imbalances (*we* are more loaded than anyone else) will
3329 * be counted as no imbalance for these purposes -- we can't fix that
3330 * by pulling tasks to us. Be careful of negative numbers as they'll
3331 * appear as very large values with unsigned longs.
3333 if (max_load
<= busiest_load_per_task
)
3337 * In the presence of smp nice balancing, certain scenarios can have
3338 * max load less than avg load(as we skip the groups at or below
3339 * its cpu_power, while calculating max_load..)
3341 if (max_load
< avg_load
) {
3343 goto small_imbalance
;
3346 /* Don't want to pull so many tasks that a group would go idle */
3347 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
3349 /* How much load to actually move to equalise the imbalance */
3350 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
3351 (avg_load
- this_load
) * this->__cpu_power
)
3355 * if *imbalance is less than the average load per runnable task
3356 * there is no gaurantee that any tasks will be moved so we'll have
3357 * a think about bumping its value to force at least one task to be
3360 if (*imbalance
< busiest_load_per_task
) {
3361 unsigned long tmp
, pwr_now
, pwr_move
;
3365 pwr_move
= pwr_now
= 0;
3367 if (this_nr_running
) {
3368 this_load_per_task
/= this_nr_running
;
3369 if (busiest_load_per_task
> this_load_per_task
)
3372 this_load_per_task
= cpu_avg_load_per_task(this_cpu
);
3374 if (max_load
- this_load
+ busiest_load_per_task
>=
3375 busiest_load_per_task
* imbn
) {
3376 *imbalance
= busiest_load_per_task
;
3381 * OK, we don't have enough imbalance to justify moving tasks,
3382 * however we may be able to increase total CPU power used by
3386 pwr_now
+= busiest
->__cpu_power
*
3387 min(busiest_load_per_task
, max_load
);
3388 pwr_now
+= this->__cpu_power
*
3389 min(this_load_per_task
, this_load
);
3390 pwr_now
/= SCHED_LOAD_SCALE
;
3392 /* Amount of load we'd subtract */
3393 tmp
= sg_div_cpu_power(busiest
,
3394 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3396 pwr_move
+= busiest
->__cpu_power
*
3397 min(busiest_load_per_task
, max_load
- tmp
);
3399 /* Amount of load we'd add */
3400 if (max_load
* busiest
->__cpu_power
<
3401 busiest_load_per_task
* SCHED_LOAD_SCALE
)
3402 tmp
= sg_div_cpu_power(this,
3403 max_load
* busiest
->__cpu_power
);
3405 tmp
= sg_div_cpu_power(this,
3406 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3407 pwr_move
+= this->__cpu_power
*
3408 min(this_load_per_task
, this_load
+ tmp
);
3409 pwr_move
/= SCHED_LOAD_SCALE
;
3411 /* Move if we gain throughput */
3412 if (pwr_move
> pwr_now
)
3413 *imbalance
= busiest_load_per_task
;
3419 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3420 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3423 if (this == group_leader
&& group_leader
!= group_min
) {
3424 *imbalance
= min_load_per_task
;
3425 if (sched_mc_power_savings
>= POWERSAVINGS_BALANCE_WAKEUP
) {
3426 cpu_rq(this_cpu
)->rd
->sched_mc_preferred_wakeup_cpu
=
3427 cpumask_first(sched_group_cpus(group_leader
));
3438 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3441 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3442 unsigned long imbalance
, const struct cpumask
*cpus
)
3444 struct rq
*busiest
= NULL
, *rq
;
3445 unsigned long max_load
= 0;
3448 for_each_cpu(i
, sched_group_cpus(group
)) {
3451 if (!cpumask_test_cpu(i
, cpus
))
3455 wl
= weighted_cpuload(i
);
3457 if (rq
->nr_running
== 1 && wl
> imbalance
)
3460 if (wl
> max_load
) {
3470 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3471 * so long as it is large enough.
3473 #define MAX_PINNED_INTERVAL 512
3476 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3477 * tasks if there is an imbalance.
3479 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3480 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3481 int *balance
, struct cpumask
*cpus
)
3483 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3484 struct sched_group
*group
;
3485 unsigned long imbalance
;
3487 unsigned long flags
;
3489 cpumask_setall(cpus
);
3492 * When power savings policy is enabled for the parent domain, idle
3493 * sibling can pick up load irrespective of busy siblings. In this case,
3494 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3495 * portraying it as CPU_NOT_IDLE.
3497 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3498 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3501 schedstat_inc(sd
, lb_count
[idle
]);
3505 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3512 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3516 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3518 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3522 BUG_ON(busiest
== this_rq
);
3524 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3527 if (busiest
->nr_running
> 1) {
3529 * Attempt to move tasks. If find_busiest_group has found
3530 * an imbalance but busiest->nr_running <= 1, the group is
3531 * still unbalanced. ld_moved simply stays zero, so it is
3532 * correctly treated as an imbalance.
3534 local_irq_save(flags
);
3535 double_rq_lock(this_rq
, busiest
);
3536 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3537 imbalance
, sd
, idle
, &all_pinned
);
3538 double_rq_unlock(this_rq
, busiest
);
3539 local_irq_restore(flags
);
3542 * some other cpu did the load balance for us.
3544 if (ld_moved
&& this_cpu
!= smp_processor_id())
3545 resched_cpu(this_cpu
);
3547 /* All tasks on this runqueue were pinned by CPU affinity */
3548 if (unlikely(all_pinned
)) {
3549 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
3550 if (!cpumask_empty(cpus
))
3557 schedstat_inc(sd
, lb_failed
[idle
]);
3558 sd
->nr_balance_failed
++;
3560 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3562 spin_lock_irqsave(&busiest
->lock
, flags
);
3564 /* don't kick the migration_thread, if the curr
3565 * task on busiest cpu can't be moved to this_cpu
3567 if (!cpumask_test_cpu(this_cpu
,
3568 &busiest
->curr
->cpus_allowed
)) {
3569 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3571 goto out_one_pinned
;
3574 if (!busiest
->active_balance
) {
3575 busiest
->active_balance
= 1;
3576 busiest
->push_cpu
= this_cpu
;
3579 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3581 wake_up_process(busiest
->migration_thread
);
3584 * We've kicked active balancing, reset the failure
3587 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3590 sd
->nr_balance_failed
= 0;
3592 if (likely(!active_balance
)) {
3593 /* We were unbalanced, so reset the balancing interval */
3594 sd
->balance_interval
= sd
->min_interval
;
3597 * If we've begun active balancing, start to back off. This
3598 * case may not be covered by the all_pinned logic if there
3599 * is only 1 task on the busy runqueue (because we don't call
3602 if (sd
->balance_interval
< sd
->max_interval
)
3603 sd
->balance_interval
*= 2;
3606 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3607 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3613 schedstat_inc(sd
, lb_balanced
[idle
]);
3615 sd
->nr_balance_failed
= 0;
3618 /* tune up the balancing interval */
3619 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3620 (sd
->balance_interval
< sd
->max_interval
))
3621 sd
->balance_interval
*= 2;
3623 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3624 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3635 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3636 * tasks if there is an imbalance.
3638 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3639 * this_rq is locked.
3642 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
,
3643 struct cpumask
*cpus
)
3645 struct sched_group
*group
;
3646 struct rq
*busiest
= NULL
;
3647 unsigned long imbalance
;
3652 cpumask_setall(cpus
);
3655 * When power savings policy is enabled for the parent domain, idle
3656 * sibling can pick up load irrespective of busy siblings. In this case,
3657 * let the state of idle sibling percolate up as IDLE, instead of
3658 * portraying it as CPU_NOT_IDLE.
3660 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3661 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3664 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3666 update_shares_locked(this_rq
, sd
);
3667 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3668 &sd_idle
, cpus
, NULL
);
3670 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3674 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
3676 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3680 BUG_ON(busiest
== this_rq
);
3682 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3685 if (busiest
->nr_running
> 1) {
3686 /* Attempt to move tasks */
3687 double_lock_balance(this_rq
, busiest
);
3688 /* this_rq->clock is already updated */
3689 update_rq_clock(busiest
);
3690 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3691 imbalance
, sd
, CPU_NEWLY_IDLE
,
3693 double_unlock_balance(this_rq
, busiest
);
3695 if (unlikely(all_pinned
)) {
3696 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
3697 if (!cpumask_empty(cpus
))
3703 int active_balance
= 0;
3705 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3706 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3707 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3710 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
3713 if (sd
->nr_balance_failed
++ < 2)
3717 * The only task running in a non-idle cpu can be moved to this
3718 * cpu in an attempt to completely freeup the other CPU
3719 * package. The same method used to move task in load_balance()
3720 * have been extended for load_balance_newidle() to speedup
3721 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
3723 * The package power saving logic comes from
3724 * find_busiest_group(). If there are no imbalance, then
3725 * f_b_g() will return NULL. However when sched_mc={1,2} then
3726 * f_b_g() will select a group from which a running task may be
3727 * pulled to this cpu in order to make the other package idle.
3728 * If there is no opportunity to make a package idle and if
3729 * there are no imbalance, then f_b_g() will return NULL and no
3730 * action will be taken in load_balance_newidle().
3732 * Under normal task pull operation due to imbalance, there
3733 * will be more than one task in the source run queue and
3734 * move_tasks() will succeed. ld_moved will be true and this
3735 * active balance code will not be triggered.
3738 /* Lock busiest in correct order while this_rq is held */
3739 double_lock_balance(this_rq
, busiest
);
3742 * don't kick the migration_thread, if the curr
3743 * task on busiest cpu can't be moved to this_cpu
3745 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
3746 double_unlock_balance(this_rq
, busiest
);
3751 if (!busiest
->active_balance
) {
3752 busiest
->active_balance
= 1;
3753 busiest
->push_cpu
= this_cpu
;
3757 double_unlock_balance(this_rq
, busiest
);
3759 * Should not call ttwu while holding a rq->lock
3761 spin_unlock(&this_rq
->lock
);
3763 wake_up_process(busiest
->migration_thread
);
3764 spin_lock(&this_rq
->lock
);
3767 sd
->nr_balance_failed
= 0;
3769 update_shares_locked(this_rq
, sd
);
3773 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3774 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3775 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3777 sd
->nr_balance_failed
= 0;
3783 * idle_balance is called by schedule() if this_cpu is about to become
3784 * idle. Attempts to pull tasks from other CPUs.
3786 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3788 struct sched_domain
*sd
;
3789 int pulled_task
= 0;
3790 unsigned long next_balance
= jiffies
+ HZ
;
3791 cpumask_var_t tmpmask
;
3793 if (!alloc_cpumask_var(&tmpmask
, GFP_ATOMIC
))
3796 for_each_domain(this_cpu
, sd
) {
3797 unsigned long interval
;
3799 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3802 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3803 /* If we've pulled tasks over stop searching: */
3804 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
3807 interval
= msecs_to_jiffies(sd
->balance_interval
);
3808 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3809 next_balance
= sd
->last_balance
+ interval
;
3813 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3815 * We are going idle. next_balance may be set based on
3816 * a busy processor. So reset next_balance.
3818 this_rq
->next_balance
= next_balance
;
3820 free_cpumask_var(tmpmask
);
3824 * active_load_balance is run by migration threads. It pushes running tasks
3825 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3826 * running on each physical CPU where possible, and avoids physical /
3827 * logical imbalances.
3829 * Called with busiest_rq locked.
3831 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3833 int target_cpu
= busiest_rq
->push_cpu
;
3834 struct sched_domain
*sd
;
3835 struct rq
*target_rq
;
3837 /* Is there any task to move? */
3838 if (busiest_rq
->nr_running
<= 1)
3841 target_rq
= cpu_rq(target_cpu
);
3844 * This condition is "impossible", if it occurs
3845 * we need to fix it. Originally reported by
3846 * Bjorn Helgaas on a 128-cpu setup.
3848 BUG_ON(busiest_rq
== target_rq
);
3850 /* move a task from busiest_rq to target_rq */
3851 double_lock_balance(busiest_rq
, target_rq
);
3852 update_rq_clock(busiest_rq
);
3853 update_rq_clock(target_rq
);
3855 /* Search for an sd spanning us and the target CPU. */
3856 for_each_domain(target_cpu
, sd
) {
3857 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3858 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
3863 schedstat_inc(sd
, alb_count
);
3865 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3867 schedstat_inc(sd
, alb_pushed
);
3869 schedstat_inc(sd
, alb_failed
);
3871 double_unlock_balance(busiest_rq
, target_rq
);
3876 atomic_t load_balancer
;
3877 cpumask_var_t cpu_mask
;
3878 } nohz ____cacheline_aligned
= {
3879 .load_balancer
= ATOMIC_INIT(-1),
3883 * This routine will try to nominate the ilb (idle load balancing)
3884 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3885 * load balancing on behalf of all those cpus. If all the cpus in the system
3886 * go into this tickless mode, then there will be no ilb owner (as there is
3887 * no need for one) and all the cpus will sleep till the next wakeup event
3890 * For the ilb owner, tick is not stopped. And this tick will be used
3891 * for idle load balancing. ilb owner will still be part of
3894 * While stopping the tick, this cpu will become the ilb owner if there
3895 * is no other owner. And will be the owner till that cpu becomes busy
3896 * or if all cpus in the system stop their ticks at which point
3897 * there is no need for ilb owner.
3899 * When the ilb owner becomes busy, it nominates another owner, during the
3900 * next busy scheduler_tick()
3902 int select_nohz_load_balancer(int stop_tick
)
3904 int cpu
= smp_processor_id();
3907 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
3908 cpu_rq(cpu
)->in_nohz_recently
= 1;
3911 * If we are going offline and still the leader, give up!
3913 if (!cpu_active(cpu
) &&
3914 atomic_read(&nohz
.load_balancer
) == cpu
) {
3915 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3920 /* time for ilb owner also to sleep */
3921 if (cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3922 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3923 atomic_set(&nohz
.load_balancer
, -1);
3927 if (atomic_read(&nohz
.load_balancer
) == -1) {
3928 /* make me the ilb owner */
3929 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3931 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3934 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
3937 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
3939 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3940 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3947 static DEFINE_SPINLOCK(balancing
);
3950 * It checks each scheduling domain to see if it is due to be balanced,
3951 * and initiates a balancing operation if so.
3953 * Balancing parameters are set up in arch_init_sched_domains.
3955 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3958 struct rq
*rq
= cpu_rq(cpu
);
3959 unsigned long interval
;
3960 struct sched_domain
*sd
;
3961 /* Earliest time when we have to do rebalance again */
3962 unsigned long next_balance
= jiffies
+ 60*HZ
;
3963 int update_next_balance
= 0;
3967 /* Fails alloc? Rebalancing probably not a priority right now. */
3968 if (!alloc_cpumask_var(&tmp
, GFP_ATOMIC
))
3971 for_each_domain(cpu
, sd
) {
3972 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3975 interval
= sd
->balance_interval
;
3976 if (idle
!= CPU_IDLE
)
3977 interval
*= sd
->busy_factor
;
3979 /* scale ms to jiffies */
3980 interval
= msecs_to_jiffies(interval
);
3981 if (unlikely(!interval
))
3983 if (interval
> HZ
*NR_CPUS
/10)
3984 interval
= HZ
*NR_CPUS
/10;
3986 need_serialize
= sd
->flags
& SD_SERIALIZE
;
3988 if (need_serialize
) {
3989 if (!spin_trylock(&balancing
))
3993 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3994 if (load_balance(cpu
, rq
, sd
, idle
, &balance
, tmp
)) {
3996 * We've pulled tasks over so either we're no
3997 * longer idle, or one of our SMT siblings is
4000 idle
= CPU_NOT_IDLE
;
4002 sd
->last_balance
= jiffies
;
4005 spin_unlock(&balancing
);
4007 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4008 next_balance
= sd
->last_balance
+ interval
;
4009 update_next_balance
= 1;
4013 * Stop the load balance at this level. There is another
4014 * CPU in our sched group which is doing load balancing more
4022 * next_balance will be updated only when there is a need.
4023 * When the cpu is attached to null domain for ex, it will not be
4026 if (likely(update_next_balance
))
4027 rq
->next_balance
= next_balance
;
4029 free_cpumask_var(tmp
);
4033 * run_rebalance_domains is triggered when needed from the scheduler tick.
4034 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4035 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4037 static void run_rebalance_domains(struct softirq_action
*h
)
4039 int this_cpu
= smp_processor_id();
4040 struct rq
*this_rq
= cpu_rq(this_cpu
);
4041 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4042 CPU_IDLE
: CPU_NOT_IDLE
;
4044 rebalance_domains(this_cpu
, idle
);
4048 * If this cpu is the owner for idle load balancing, then do the
4049 * balancing on behalf of the other idle cpus whose ticks are
4052 if (this_rq
->idle_at_tick
&&
4053 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4057 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
4058 if (balance_cpu
== this_cpu
)
4062 * If this cpu gets work to do, stop the load balancing
4063 * work being done for other cpus. Next load
4064 * balancing owner will pick it up.
4069 rebalance_domains(balance_cpu
, CPU_IDLE
);
4071 rq
= cpu_rq(balance_cpu
);
4072 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4073 this_rq
->next_balance
= rq
->next_balance
;
4080 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4082 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4083 * idle load balancing owner or decide to stop the periodic load balancing,
4084 * if the whole system is idle.
4086 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4090 * If we were in the nohz mode recently and busy at the current
4091 * scheduler tick, then check if we need to nominate new idle
4094 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4095 rq
->in_nohz_recently
= 0;
4097 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4098 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4099 atomic_set(&nohz
.load_balancer
, -1);
4102 if (atomic_read(&nohz
.load_balancer
) == -1) {
4104 * simple selection for now: Nominate the
4105 * first cpu in the nohz list to be the next
4108 * TBD: Traverse the sched domains and nominate
4109 * the nearest cpu in the nohz.cpu_mask.
4111 int ilb
= cpumask_first(nohz
.cpu_mask
);
4113 if (ilb
< nr_cpu_ids
)
4119 * If this cpu is idle and doing idle load balancing for all the
4120 * cpus with ticks stopped, is it time for that to stop?
4122 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4123 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4129 * If this cpu is idle and the idle load balancing is done by
4130 * someone else, then no need raise the SCHED_SOFTIRQ
4132 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4133 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4136 if (time_after_eq(jiffies
, rq
->next_balance
))
4137 raise_softirq(SCHED_SOFTIRQ
);
4140 #else /* CONFIG_SMP */
4143 * on UP we do not need to balance between CPUs:
4145 static inline void idle_balance(int cpu
, struct rq
*rq
)
4151 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4153 EXPORT_PER_CPU_SYMBOL(kstat
);
4156 * Return any ns on the sched_clock that have not yet been banked in
4157 * @p in case that task is currently running.
4159 unsigned long long task_delta_exec(struct task_struct
*p
)
4161 unsigned long flags
;
4165 rq
= task_rq_lock(p
, &flags
);
4167 if (task_current(rq
, p
)) {
4170 update_rq_clock(rq
);
4171 delta_exec
= rq
->clock
- p
->se
.exec_start
;
4172 if ((s64
)delta_exec
> 0)
4176 task_rq_unlock(rq
, &flags
);
4182 * Account user cpu time to a process.
4183 * @p: the process that the cpu time gets accounted to
4184 * @cputime: the cpu time spent in user space since the last update
4185 * @cputime_scaled: cputime scaled by cpu frequency
4187 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
4188 cputime_t cputime_scaled
)
4190 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4193 /* Add user time to process. */
4194 p
->utime
= cputime_add(p
->utime
, cputime
);
4195 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4196 account_group_user_time(p
, cputime
);
4198 /* Add user time to cpustat. */
4199 tmp
= cputime_to_cputime64(cputime
);
4200 if (TASK_NICE(p
) > 0)
4201 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4203 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4204 /* Account for user time used */
4205 acct_update_integrals(p
);
4209 * Account guest cpu time to a process.
4210 * @p: the process that the cpu time gets accounted to
4211 * @cputime: the cpu time spent in virtual machine since the last update
4212 * @cputime_scaled: cputime scaled by cpu frequency
4214 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
4215 cputime_t cputime_scaled
)
4218 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4220 tmp
= cputime_to_cputime64(cputime
);
4222 /* Add guest time to process. */
4223 p
->utime
= cputime_add(p
->utime
, cputime
);
4224 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4225 account_group_user_time(p
, cputime
);
4226 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4228 /* Add guest time to cpustat. */
4229 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4230 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4234 * Account system cpu time to a process.
4235 * @p: the process that the cpu time gets accounted to
4236 * @hardirq_offset: the offset to subtract from hardirq_count()
4237 * @cputime: the cpu time spent in kernel space since the last update
4238 * @cputime_scaled: cputime scaled by cpu frequency
4240 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4241 cputime_t cputime
, cputime_t cputime_scaled
)
4243 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4246 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4247 account_guest_time(p
, cputime
, cputime_scaled
);
4251 /* Add system time to process. */
4252 p
->stime
= cputime_add(p
->stime
, cputime
);
4253 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
4254 account_group_system_time(p
, cputime
);
4256 /* Add system time to cpustat. */
4257 tmp
= cputime_to_cputime64(cputime
);
4258 if (hardirq_count() - hardirq_offset
)
4259 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4260 else if (softirq_count())
4261 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4263 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4265 /* Account for system time used */
4266 acct_update_integrals(p
);
4270 * Account for involuntary wait time.
4271 * @steal: the cpu time spent in involuntary wait
4273 void account_steal_time(cputime_t cputime
)
4275 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4276 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
4278 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
4282 * Account for idle time.
4283 * @cputime: the cpu time spent in idle wait
4285 void account_idle_time(cputime_t cputime
)
4287 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4288 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
4289 struct rq
*rq
= this_rq();
4291 if (atomic_read(&rq
->nr_iowait
) > 0)
4292 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
4294 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
4297 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4300 * Account a single tick of cpu time.
4301 * @p: the process that the cpu time gets accounted to
4302 * @user_tick: indicates if the tick is a user or a system tick
4304 void account_process_tick(struct task_struct
*p
, int user_tick
)
4306 cputime_t one_jiffy
= jiffies_to_cputime(1);
4307 cputime_t one_jiffy_scaled
= cputime_to_scaled(one_jiffy
);
4308 struct rq
*rq
= this_rq();
4311 account_user_time(p
, one_jiffy
, one_jiffy_scaled
);
4312 else if (p
!= rq
->idle
)
4313 account_system_time(p
, HARDIRQ_OFFSET
, one_jiffy
,
4316 account_idle_time(one_jiffy
);
4320 * Account multiple ticks of steal time.
4321 * @p: the process from which the cpu time has been stolen
4322 * @ticks: number of stolen ticks
4324 void account_steal_ticks(unsigned long ticks
)
4326 account_steal_time(jiffies_to_cputime(ticks
));
4330 * Account multiple ticks of idle time.
4331 * @ticks: number of stolen ticks
4333 void account_idle_ticks(unsigned long ticks
)
4335 account_idle_time(jiffies_to_cputime(ticks
));
4341 * Use precise platform statistics if available:
4343 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4344 cputime_t
task_utime(struct task_struct
*p
)
4349 cputime_t
task_stime(struct task_struct
*p
)
4354 cputime_t
task_utime(struct task_struct
*p
)
4356 clock_t utime
= cputime_to_clock_t(p
->utime
),
4357 total
= utime
+ cputime_to_clock_t(p
->stime
);
4361 * Use CFS's precise accounting:
4363 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
4367 do_div(temp
, total
);
4369 utime
= (clock_t)temp
;
4371 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
4372 return p
->prev_utime
;
4375 cputime_t
task_stime(struct task_struct
*p
)
4380 * Use CFS's precise accounting. (we subtract utime from
4381 * the total, to make sure the total observed by userspace
4382 * grows monotonically - apps rely on that):
4384 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
4385 cputime_to_clock_t(task_utime(p
));
4388 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
4390 return p
->prev_stime
;
4394 inline cputime_t
task_gtime(struct task_struct
*p
)
4400 * This function gets called by the timer code, with HZ frequency.
4401 * We call it with interrupts disabled.
4403 * It also gets called by the fork code, when changing the parent's
4406 void scheduler_tick(void)
4408 int cpu
= smp_processor_id();
4409 struct rq
*rq
= cpu_rq(cpu
);
4410 struct task_struct
*curr
= rq
->curr
;
4414 spin_lock(&rq
->lock
);
4415 update_rq_clock(rq
);
4416 update_cpu_load(rq
);
4417 curr
->sched_class
->task_tick(rq
, curr
, 0);
4418 spin_unlock(&rq
->lock
);
4421 rq
->idle_at_tick
= idle_cpu(cpu
);
4422 trigger_load_balance(rq
, cpu
);
4426 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4427 defined(CONFIG_PREEMPT_TRACER))
4429 static inline unsigned long get_parent_ip(unsigned long addr
)
4431 if (in_lock_functions(addr
)) {
4432 addr
= CALLER_ADDR2
;
4433 if (in_lock_functions(addr
))
4434 addr
= CALLER_ADDR3
;
4439 void __kprobes
add_preempt_count(int val
)
4441 #ifdef CONFIG_DEBUG_PREEMPT
4445 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4448 preempt_count() += val
;
4449 #ifdef CONFIG_DEBUG_PREEMPT
4451 * Spinlock count overflowing soon?
4453 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4456 if (preempt_count() == val
)
4457 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4459 EXPORT_SYMBOL(add_preempt_count
);
4461 void __kprobes
sub_preempt_count(int val
)
4463 #ifdef CONFIG_DEBUG_PREEMPT
4467 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count() - (!!kernel_locked())))
4470 * Is the spinlock portion underflowing?
4472 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4473 !(preempt_count() & PREEMPT_MASK
)))
4477 if (preempt_count() == val
)
4478 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4479 preempt_count() -= val
;
4481 EXPORT_SYMBOL(sub_preempt_count
);
4486 * Print scheduling while atomic bug:
4488 static noinline
void __schedule_bug(struct task_struct
*prev
)
4490 struct pt_regs
*regs
= get_irq_regs();
4492 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4493 prev
->comm
, prev
->pid
, preempt_count());
4495 debug_show_held_locks(prev
);
4497 if (irqs_disabled())
4498 print_irqtrace_events(prev
);
4507 * Various schedule()-time debugging checks and statistics:
4509 static inline void schedule_debug(struct task_struct
*prev
)
4512 * Test if we are atomic. Since do_exit() needs to call into
4513 * schedule() atomically, we ignore that path for now.
4514 * Otherwise, whine if we are scheduling when we should not be.
4516 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4517 __schedule_bug(prev
);
4519 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4521 schedstat_inc(this_rq(), sched_count
);
4522 #ifdef CONFIG_SCHEDSTATS
4523 if (unlikely(prev
->lock_depth
>= 0)) {
4524 schedstat_inc(this_rq(), bkl_count
);
4525 schedstat_inc(prev
, sched_info
.bkl_count
);
4531 * Pick up the highest-prio task:
4533 static inline struct task_struct
*
4534 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
4536 const struct sched_class
*class;
4537 struct task_struct
*p
;
4540 * Optimization: we know that if all tasks are in
4541 * the fair class we can call that function directly:
4543 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4544 p
= fair_sched_class
.pick_next_task(rq
);
4549 class = sched_class_highest
;
4551 p
= class->pick_next_task(rq
);
4555 * Will never be NULL as the idle class always
4556 * returns a non-NULL p:
4558 class = class->next
;
4563 * schedule() is the main scheduler function.
4565 asmlinkage
void __sched
schedule(void)
4567 struct task_struct
*prev
, *next
;
4568 unsigned long *switch_count
;
4574 cpu
= smp_processor_id();
4578 switch_count
= &prev
->nivcsw
;
4580 release_kernel_lock(prev
);
4581 need_resched_nonpreemptible
:
4583 schedule_debug(prev
);
4585 if (sched_feat(HRTICK
))
4588 spin_lock_irq(&rq
->lock
);
4589 update_rq_clock(rq
);
4590 clear_tsk_need_resched(prev
);
4592 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4593 if (unlikely(signal_pending_state(prev
->state
, prev
)))
4594 prev
->state
= TASK_RUNNING
;
4596 deactivate_task(rq
, prev
, 1);
4597 switch_count
= &prev
->nvcsw
;
4601 if (prev
->sched_class
->pre_schedule
)
4602 prev
->sched_class
->pre_schedule(rq
, prev
);
4605 if (unlikely(!rq
->nr_running
))
4606 idle_balance(cpu
, rq
);
4608 prev
->sched_class
->put_prev_task(rq
, prev
);
4609 next
= pick_next_task(rq
, prev
);
4611 if (likely(prev
!= next
)) {
4612 sched_info_switch(prev
, next
);
4618 context_switch(rq
, prev
, next
); /* unlocks the rq */
4620 * the context switch might have flipped the stack from under
4621 * us, hence refresh the local variables.
4623 cpu
= smp_processor_id();
4626 spin_unlock_irq(&rq
->lock
);
4628 if (unlikely(reacquire_kernel_lock(current
) < 0))
4629 goto need_resched_nonpreemptible
;
4631 preempt_enable_no_resched();
4632 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
4635 EXPORT_SYMBOL(schedule
);
4637 #ifdef CONFIG_PREEMPT
4639 * this is the entry point to schedule() from in-kernel preemption
4640 * off of preempt_enable. Kernel preemptions off return from interrupt
4641 * occur there and call schedule directly.
4643 asmlinkage
void __sched
preempt_schedule(void)
4645 struct thread_info
*ti
= current_thread_info();
4648 * If there is a non-zero preempt_count or interrupts are disabled,
4649 * we do not want to preempt the current task. Just return..
4651 if (likely(ti
->preempt_count
|| irqs_disabled()))
4655 add_preempt_count(PREEMPT_ACTIVE
);
4657 sub_preempt_count(PREEMPT_ACTIVE
);
4660 * Check again in case we missed a preemption opportunity
4661 * between schedule and now.
4664 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4666 EXPORT_SYMBOL(preempt_schedule
);
4669 * this is the entry point to schedule() from kernel preemption
4670 * off of irq context.
4671 * Note, that this is called and return with irqs disabled. This will
4672 * protect us against recursive calling from irq.
4674 asmlinkage
void __sched
preempt_schedule_irq(void)
4676 struct thread_info
*ti
= current_thread_info();
4678 /* Catch callers which need to be fixed */
4679 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4682 add_preempt_count(PREEMPT_ACTIVE
);
4685 local_irq_disable();
4686 sub_preempt_count(PREEMPT_ACTIVE
);
4689 * Check again in case we missed a preemption opportunity
4690 * between schedule and now.
4693 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4696 #endif /* CONFIG_PREEMPT */
4698 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4701 return try_to_wake_up(curr
->private, mode
, sync
);
4703 EXPORT_SYMBOL(default_wake_function
);
4706 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4707 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4708 * number) then we wake all the non-exclusive tasks and one exclusive task.
4710 * There are circumstances in which we can try to wake a task which has already
4711 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4712 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4714 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4715 int nr_exclusive
, int sync
, void *key
)
4717 wait_queue_t
*curr
, *next
;
4719 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4720 unsigned flags
= curr
->flags
;
4722 if (curr
->func(curr
, mode
, sync
, key
) &&
4723 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4729 * __wake_up - wake up threads blocked on a waitqueue.
4731 * @mode: which threads
4732 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4733 * @key: is directly passed to the wakeup function
4735 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4736 int nr_exclusive
, void *key
)
4738 unsigned long flags
;
4740 spin_lock_irqsave(&q
->lock
, flags
);
4741 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4742 spin_unlock_irqrestore(&q
->lock
, flags
);
4744 EXPORT_SYMBOL(__wake_up
);
4747 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4749 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4751 __wake_up_common(q
, mode
, 1, 0, NULL
);
4755 * __wake_up_sync - wake up threads blocked on a waitqueue.
4757 * @mode: which threads
4758 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4760 * The sync wakeup differs that the waker knows that it will schedule
4761 * away soon, so while the target thread will be woken up, it will not
4762 * be migrated to another CPU - ie. the two threads are 'synchronized'
4763 * with each other. This can prevent needless bouncing between CPUs.
4765 * On UP it can prevent extra preemption.
4768 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4770 unsigned long flags
;
4776 if (unlikely(!nr_exclusive
))
4779 spin_lock_irqsave(&q
->lock
, flags
);
4780 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4781 spin_unlock_irqrestore(&q
->lock
, flags
);
4783 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4786 * complete: - signals a single thread waiting on this completion
4787 * @x: holds the state of this particular completion
4789 * This will wake up a single thread waiting on this completion. Threads will be
4790 * awakened in the same order in which they were queued.
4792 * See also complete_all(), wait_for_completion() and related routines.
4794 void complete(struct completion
*x
)
4796 unsigned long flags
;
4798 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4800 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4801 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4803 EXPORT_SYMBOL(complete
);
4806 * complete_all: - signals all threads waiting on this completion
4807 * @x: holds the state of this particular completion
4809 * This will wake up all threads waiting on this particular completion event.
4811 void complete_all(struct completion
*x
)
4813 unsigned long flags
;
4815 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4816 x
->done
+= UINT_MAX
/2;
4817 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4818 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4820 EXPORT_SYMBOL(complete_all
);
4822 static inline long __sched
4823 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4826 DECLARE_WAITQUEUE(wait
, current
);
4828 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4829 __add_wait_queue_tail(&x
->wait
, &wait
);
4831 if (signal_pending_state(state
, current
)) {
4832 timeout
= -ERESTARTSYS
;
4835 __set_current_state(state
);
4836 spin_unlock_irq(&x
->wait
.lock
);
4837 timeout
= schedule_timeout(timeout
);
4838 spin_lock_irq(&x
->wait
.lock
);
4839 } while (!x
->done
&& timeout
);
4840 __remove_wait_queue(&x
->wait
, &wait
);
4845 return timeout
?: 1;
4849 wait_for_common(struct completion
*x
, long timeout
, int state
)
4853 spin_lock_irq(&x
->wait
.lock
);
4854 timeout
= do_wait_for_common(x
, timeout
, state
);
4855 spin_unlock_irq(&x
->wait
.lock
);
4860 * wait_for_completion: - waits for completion of a task
4861 * @x: holds the state of this particular completion
4863 * This waits to be signaled for completion of a specific task. It is NOT
4864 * interruptible and there is no timeout.
4866 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4867 * and interrupt capability. Also see complete().
4869 void __sched
wait_for_completion(struct completion
*x
)
4871 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4873 EXPORT_SYMBOL(wait_for_completion
);
4876 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4877 * @x: holds the state of this particular completion
4878 * @timeout: timeout value in jiffies
4880 * This waits for either a completion of a specific task to be signaled or for a
4881 * specified timeout to expire. The timeout is in jiffies. It is not
4884 unsigned long __sched
4885 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4887 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4889 EXPORT_SYMBOL(wait_for_completion_timeout
);
4892 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4893 * @x: holds the state of this particular completion
4895 * This waits for completion of a specific task to be signaled. It is
4898 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4900 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4901 if (t
== -ERESTARTSYS
)
4905 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4908 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4909 * @x: holds the state of this particular completion
4910 * @timeout: timeout value in jiffies
4912 * This waits for either a completion of a specific task to be signaled or for a
4913 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4915 unsigned long __sched
4916 wait_for_completion_interruptible_timeout(struct completion
*x
,
4917 unsigned long timeout
)
4919 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4921 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4924 * wait_for_completion_killable: - waits for completion of a task (killable)
4925 * @x: holds the state of this particular completion
4927 * This waits to be signaled for completion of a specific task. It can be
4928 * interrupted by a kill signal.
4930 int __sched
wait_for_completion_killable(struct completion
*x
)
4932 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4933 if (t
== -ERESTARTSYS
)
4937 EXPORT_SYMBOL(wait_for_completion_killable
);
4940 * try_wait_for_completion - try to decrement a completion without blocking
4941 * @x: completion structure
4943 * Returns: 0 if a decrement cannot be done without blocking
4944 * 1 if a decrement succeeded.
4946 * If a completion is being used as a counting completion,
4947 * attempt to decrement the counter without blocking. This
4948 * enables us to avoid waiting if the resource the completion
4949 * is protecting is not available.
4951 bool try_wait_for_completion(struct completion
*x
)
4955 spin_lock_irq(&x
->wait
.lock
);
4960 spin_unlock_irq(&x
->wait
.lock
);
4963 EXPORT_SYMBOL(try_wait_for_completion
);
4966 * completion_done - Test to see if a completion has any waiters
4967 * @x: completion structure
4969 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4970 * 1 if there are no waiters.
4973 bool completion_done(struct completion
*x
)
4977 spin_lock_irq(&x
->wait
.lock
);
4980 spin_unlock_irq(&x
->wait
.lock
);
4983 EXPORT_SYMBOL(completion_done
);
4986 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4988 unsigned long flags
;
4991 init_waitqueue_entry(&wait
, current
);
4993 __set_current_state(state
);
4995 spin_lock_irqsave(&q
->lock
, flags
);
4996 __add_wait_queue(q
, &wait
);
4997 spin_unlock(&q
->lock
);
4998 timeout
= schedule_timeout(timeout
);
4999 spin_lock_irq(&q
->lock
);
5000 __remove_wait_queue(q
, &wait
);
5001 spin_unlock_irqrestore(&q
->lock
, flags
);
5006 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
5008 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5010 EXPORT_SYMBOL(interruptible_sleep_on
);
5013 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5015 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
5017 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
5019 void __sched
sleep_on(wait_queue_head_t
*q
)
5021 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5023 EXPORT_SYMBOL(sleep_on
);
5025 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5027 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
5029 EXPORT_SYMBOL(sleep_on_timeout
);
5031 #ifdef CONFIG_RT_MUTEXES
5034 * rt_mutex_setprio - set the current priority of a task
5036 * @prio: prio value (kernel-internal form)
5038 * This function changes the 'effective' priority of a task. It does
5039 * not touch ->normal_prio like __setscheduler().
5041 * Used by the rt_mutex code to implement priority inheritance logic.
5043 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
5045 unsigned long flags
;
5046 int oldprio
, on_rq
, running
;
5048 const struct sched_class
*prev_class
= p
->sched_class
;
5050 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
5052 rq
= task_rq_lock(p
, &flags
);
5053 update_rq_clock(rq
);
5056 on_rq
= p
->se
.on_rq
;
5057 running
= task_current(rq
, p
);
5059 dequeue_task(rq
, p
, 0);
5061 p
->sched_class
->put_prev_task(rq
, p
);
5064 p
->sched_class
= &rt_sched_class
;
5066 p
->sched_class
= &fair_sched_class
;
5071 p
->sched_class
->set_curr_task(rq
);
5073 enqueue_task(rq
, p
, 0);
5075 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5077 task_rq_unlock(rq
, &flags
);
5082 void set_user_nice(struct task_struct
*p
, long nice
)
5084 int old_prio
, delta
, on_rq
;
5085 unsigned long flags
;
5088 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
5091 * We have to be careful, if called from sys_setpriority(),
5092 * the task might be in the middle of scheduling on another CPU.
5094 rq
= task_rq_lock(p
, &flags
);
5095 update_rq_clock(rq
);
5097 * The RT priorities are set via sched_setscheduler(), but we still
5098 * allow the 'normal' nice value to be set - but as expected
5099 * it wont have any effect on scheduling until the task is
5100 * SCHED_FIFO/SCHED_RR:
5102 if (task_has_rt_policy(p
)) {
5103 p
->static_prio
= NICE_TO_PRIO(nice
);
5106 on_rq
= p
->se
.on_rq
;
5108 dequeue_task(rq
, p
, 0);
5110 p
->static_prio
= NICE_TO_PRIO(nice
);
5113 p
->prio
= effective_prio(p
);
5114 delta
= p
->prio
- old_prio
;
5117 enqueue_task(rq
, p
, 0);
5119 * If the task increased its priority or is running and
5120 * lowered its priority, then reschedule its CPU:
5122 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
5123 resched_task(rq
->curr
);
5126 task_rq_unlock(rq
, &flags
);
5128 EXPORT_SYMBOL(set_user_nice
);
5131 * can_nice - check if a task can reduce its nice value
5135 int can_nice(const struct task_struct
*p
, const int nice
)
5137 /* convert nice value [19,-20] to rlimit style value [1,40] */
5138 int nice_rlim
= 20 - nice
;
5140 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
5141 capable(CAP_SYS_NICE
));
5144 #ifdef __ARCH_WANT_SYS_NICE
5147 * sys_nice - change the priority of the current process.
5148 * @increment: priority increment
5150 * sys_setpriority is a more generic, but much slower function that
5151 * does similar things.
5153 asmlinkage
long sys_nice(int increment
)
5158 * Setpriority might change our priority at the same moment.
5159 * We don't have to worry. Conceptually one call occurs first
5160 * and we have a single winner.
5162 if (increment
< -40)
5167 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
5173 if (increment
< 0 && !can_nice(current
, nice
))
5176 retval
= security_task_setnice(current
, nice
);
5180 set_user_nice(current
, nice
);
5187 * task_prio - return the priority value of a given task.
5188 * @p: the task in question.
5190 * This is the priority value as seen by users in /proc.
5191 * RT tasks are offset by -200. Normal tasks are centered
5192 * around 0, value goes from -16 to +15.
5194 int task_prio(const struct task_struct
*p
)
5196 return p
->prio
- MAX_RT_PRIO
;
5200 * task_nice - return the nice value of a given task.
5201 * @p: the task in question.
5203 int task_nice(const struct task_struct
*p
)
5205 return TASK_NICE(p
);
5207 EXPORT_SYMBOL(task_nice
);
5210 * idle_cpu - is a given cpu idle currently?
5211 * @cpu: the processor in question.
5213 int idle_cpu(int cpu
)
5215 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
5219 * idle_task - return the idle task for a given cpu.
5220 * @cpu: the processor in question.
5222 struct task_struct
*idle_task(int cpu
)
5224 return cpu_rq(cpu
)->idle
;
5228 * find_process_by_pid - find a process with a matching PID value.
5229 * @pid: the pid in question.
5231 static struct task_struct
*find_process_by_pid(pid_t pid
)
5233 return pid
? find_task_by_vpid(pid
) : current
;
5236 /* Actually do priority change: must hold rq lock. */
5238 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
5240 BUG_ON(p
->se
.on_rq
);
5243 switch (p
->policy
) {
5247 p
->sched_class
= &fair_sched_class
;
5251 p
->sched_class
= &rt_sched_class
;
5255 p
->rt_priority
= prio
;
5256 p
->normal_prio
= normal_prio(p
);
5257 /* we are holding p->pi_lock already */
5258 p
->prio
= rt_mutex_getprio(p
);
5263 * check the target process has a UID that matches the current process's
5265 static bool check_same_owner(struct task_struct
*p
)
5267 const struct cred
*cred
= current_cred(), *pcred
;
5271 pcred
= __task_cred(p
);
5272 match
= (cred
->euid
== pcred
->euid
||
5273 cred
->euid
== pcred
->uid
);
5278 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
5279 struct sched_param
*param
, bool user
)
5281 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
5282 unsigned long flags
;
5283 const struct sched_class
*prev_class
= p
->sched_class
;
5286 /* may grab non-irq protected spin_locks */
5287 BUG_ON(in_interrupt());
5289 /* double check policy once rq lock held */
5291 policy
= oldpolicy
= p
->policy
;
5292 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
5293 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
5294 policy
!= SCHED_IDLE
)
5297 * Valid priorities for SCHED_FIFO and SCHED_RR are
5298 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5299 * SCHED_BATCH and SCHED_IDLE is 0.
5301 if (param
->sched_priority
< 0 ||
5302 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5303 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
5305 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
5309 * Allow unprivileged RT tasks to decrease priority:
5311 if (user
&& !capable(CAP_SYS_NICE
)) {
5312 if (rt_policy(policy
)) {
5313 unsigned long rlim_rtprio
;
5315 if (!lock_task_sighand(p
, &flags
))
5317 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
5318 unlock_task_sighand(p
, &flags
);
5320 /* can't set/change the rt policy */
5321 if (policy
!= p
->policy
&& !rlim_rtprio
)
5324 /* can't increase priority */
5325 if (param
->sched_priority
> p
->rt_priority
&&
5326 param
->sched_priority
> rlim_rtprio
)
5330 * Like positive nice levels, dont allow tasks to
5331 * move out of SCHED_IDLE either:
5333 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
5336 /* can't change other user's priorities */
5337 if (!check_same_owner(p
))
5342 #ifdef CONFIG_RT_GROUP_SCHED
5344 * Do not allow realtime tasks into groups that have no runtime
5347 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
5348 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
5352 retval
= security_task_setscheduler(p
, policy
, param
);
5358 * make sure no PI-waiters arrive (or leave) while we are
5359 * changing the priority of the task:
5361 spin_lock_irqsave(&p
->pi_lock
, flags
);
5363 * To be able to change p->policy safely, the apropriate
5364 * runqueue lock must be held.
5366 rq
= __task_rq_lock(p
);
5367 /* recheck policy now with rq lock held */
5368 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5369 policy
= oldpolicy
= -1;
5370 __task_rq_unlock(rq
);
5371 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5374 update_rq_clock(rq
);
5375 on_rq
= p
->se
.on_rq
;
5376 running
= task_current(rq
, p
);
5378 deactivate_task(rq
, p
, 0);
5380 p
->sched_class
->put_prev_task(rq
, p
);
5383 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5386 p
->sched_class
->set_curr_task(rq
);
5388 activate_task(rq
, p
, 0);
5390 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5392 __task_rq_unlock(rq
);
5393 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5395 rt_mutex_adjust_pi(p
);
5401 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5402 * @p: the task in question.
5403 * @policy: new policy.
5404 * @param: structure containing the new RT priority.
5406 * NOTE that the task may be already dead.
5408 int sched_setscheduler(struct task_struct
*p
, int policy
,
5409 struct sched_param
*param
)
5411 return __sched_setscheduler(p
, policy
, param
, true);
5413 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5416 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5417 * @p: the task in question.
5418 * @policy: new policy.
5419 * @param: structure containing the new RT priority.
5421 * Just like sched_setscheduler, only don't bother checking if the
5422 * current context has permission. For example, this is needed in
5423 * stop_machine(): we create temporary high priority worker threads,
5424 * but our caller might not have that capability.
5426 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5427 struct sched_param
*param
)
5429 return __sched_setscheduler(p
, policy
, param
, false);
5433 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5435 struct sched_param lparam
;
5436 struct task_struct
*p
;
5439 if (!param
|| pid
< 0)
5441 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5446 p
= find_process_by_pid(pid
);
5448 retval
= sched_setscheduler(p
, policy
, &lparam
);
5455 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5456 * @pid: the pid in question.
5457 * @policy: new policy.
5458 * @param: structure containing the new RT priority.
5461 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5463 /* negative values for policy are not valid */
5467 return do_sched_setscheduler(pid
, policy
, param
);
5471 * sys_sched_setparam - set/change the RT priority of a thread
5472 * @pid: the pid in question.
5473 * @param: structure containing the new RT priority.
5475 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
5477 return do_sched_setscheduler(pid
, -1, param
);
5481 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5482 * @pid: the pid in question.
5484 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
5486 struct task_struct
*p
;
5493 read_lock(&tasklist_lock
);
5494 p
= find_process_by_pid(pid
);
5496 retval
= security_task_getscheduler(p
);
5500 read_unlock(&tasklist_lock
);
5505 * sys_sched_getscheduler - get the RT priority of a thread
5506 * @pid: the pid in question.
5507 * @param: structure containing the RT priority.
5509 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
5511 struct sched_param lp
;
5512 struct task_struct
*p
;
5515 if (!param
|| pid
< 0)
5518 read_lock(&tasklist_lock
);
5519 p
= find_process_by_pid(pid
);
5524 retval
= security_task_getscheduler(p
);
5528 lp
.sched_priority
= p
->rt_priority
;
5529 read_unlock(&tasklist_lock
);
5532 * This one might sleep, we cannot do it with a spinlock held ...
5534 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5539 read_unlock(&tasklist_lock
);
5543 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
5545 cpumask_var_t cpus_allowed
, new_mask
;
5546 struct task_struct
*p
;
5550 read_lock(&tasklist_lock
);
5552 p
= find_process_by_pid(pid
);
5554 read_unlock(&tasklist_lock
);
5560 * It is not safe to call set_cpus_allowed with the
5561 * tasklist_lock held. We will bump the task_struct's
5562 * usage count and then drop tasklist_lock.
5565 read_unlock(&tasklist_lock
);
5567 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
5571 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
5573 goto out_free_cpus_allowed
;
5576 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
5579 retval
= security_task_setscheduler(p
, 0, NULL
);
5583 cpuset_cpus_allowed(p
, cpus_allowed
);
5584 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
5586 retval
= set_cpus_allowed_ptr(p
, new_mask
);
5589 cpuset_cpus_allowed(p
, cpus_allowed
);
5590 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
5592 * We must have raced with a concurrent cpuset
5593 * update. Just reset the cpus_allowed to the
5594 * cpuset's cpus_allowed
5596 cpumask_copy(new_mask
, cpus_allowed
);
5601 free_cpumask_var(new_mask
);
5602 out_free_cpus_allowed
:
5603 free_cpumask_var(cpus_allowed
);
5610 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5611 struct cpumask
*new_mask
)
5613 if (len
< cpumask_size())
5614 cpumask_clear(new_mask
);
5615 else if (len
> cpumask_size())
5616 len
= cpumask_size();
5618 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5622 * sys_sched_setaffinity - set the cpu affinity of a process
5623 * @pid: pid of the process
5624 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5625 * @user_mask_ptr: user-space pointer to the new cpu mask
5627 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
5628 unsigned long __user
*user_mask_ptr
)
5630 cpumask_var_t new_mask
;
5633 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
5636 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
5638 retval
= sched_setaffinity(pid
, new_mask
);
5639 free_cpumask_var(new_mask
);
5643 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
5645 struct task_struct
*p
;
5649 read_lock(&tasklist_lock
);
5652 p
= find_process_by_pid(pid
);
5656 retval
= security_task_getscheduler(p
);
5660 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
5663 read_unlock(&tasklist_lock
);
5670 * sys_sched_getaffinity - get the cpu affinity of a process
5671 * @pid: pid of the process
5672 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5673 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5675 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
5676 unsigned long __user
*user_mask_ptr
)
5681 if (len
< cpumask_size())
5684 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
5687 ret
= sched_getaffinity(pid
, mask
);
5689 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
5692 ret
= cpumask_size();
5694 free_cpumask_var(mask
);
5700 * sys_sched_yield - yield the current processor to other threads.
5702 * This function yields the current CPU to other tasks. If there are no
5703 * other threads running on this CPU then this function will return.
5705 asmlinkage
long sys_sched_yield(void)
5707 struct rq
*rq
= this_rq_lock();
5709 schedstat_inc(rq
, yld_count
);
5710 current
->sched_class
->yield_task(rq
);
5713 * Since we are going to call schedule() anyway, there's
5714 * no need to preempt or enable interrupts:
5716 __release(rq
->lock
);
5717 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5718 _raw_spin_unlock(&rq
->lock
);
5719 preempt_enable_no_resched();
5726 static void __cond_resched(void)
5728 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5729 __might_sleep(__FILE__
, __LINE__
);
5732 * The BKS might be reacquired before we have dropped
5733 * PREEMPT_ACTIVE, which could trigger a second
5734 * cond_resched() call.
5737 add_preempt_count(PREEMPT_ACTIVE
);
5739 sub_preempt_count(PREEMPT_ACTIVE
);
5740 } while (need_resched());
5743 int __sched
_cond_resched(void)
5745 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
5746 system_state
== SYSTEM_RUNNING
) {
5752 EXPORT_SYMBOL(_cond_resched
);
5755 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5756 * call schedule, and on return reacquire the lock.
5758 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5759 * operations here to prevent schedule() from being called twice (once via
5760 * spin_unlock(), once by hand).
5762 int cond_resched_lock(spinlock_t
*lock
)
5764 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
5767 if (spin_needbreak(lock
) || resched
) {
5769 if (resched
&& need_resched())
5778 EXPORT_SYMBOL(cond_resched_lock
);
5780 int __sched
cond_resched_softirq(void)
5782 BUG_ON(!in_softirq());
5784 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
5792 EXPORT_SYMBOL(cond_resched_softirq
);
5795 * yield - yield the current processor to other threads.
5797 * This is a shortcut for kernel-space yielding - it marks the
5798 * thread runnable and calls sys_sched_yield().
5800 void __sched
yield(void)
5802 set_current_state(TASK_RUNNING
);
5805 EXPORT_SYMBOL(yield
);
5808 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5809 * that process accounting knows that this is a task in IO wait state.
5811 * But don't do that if it is a deliberate, throttling IO wait (this task
5812 * has set its backing_dev_info: the queue against which it should throttle)
5814 void __sched
io_schedule(void)
5816 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5818 delayacct_blkio_start();
5819 atomic_inc(&rq
->nr_iowait
);
5821 atomic_dec(&rq
->nr_iowait
);
5822 delayacct_blkio_end();
5824 EXPORT_SYMBOL(io_schedule
);
5826 long __sched
io_schedule_timeout(long timeout
)
5828 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5831 delayacct_blkio_start();
5832 atomic_inc(&rq
->nr_iowait
);
5833 ret
= schedule_timeout(timeout
);
5834 atomic_dec(&rq
->nr_iowait
);
5835 delayacct_blkio_end();
5840 * sys_sched_get_priority_max - return maximum RT priority.
5841 * @policy: scheduling class.
5843 * this syscall returns the maximum rt_priority that can be used
5844 * by a given scheduling class.
5846 asmlinkage
long sys_sched_get_priority_max(int policy
)
5853 ret
= MAX_USER_RT_PRIO
-1;
5865 * sys_sched_get_priority_min - return minimum RT priority.
5866 * @policy: scheduling class.
5868 * this syscall returns the minimum rt_priority that can be used
5869 * by a given scheduling class.
5871 asmlinkage
long sys_sched_get_priority_min(int policy
)
5889 * sys_sched_rr_get_interval - return the default timeslice of a process.
5890 * @pid: pid of the process.
5891 * @interval: userspace pointer to the timeslice value.
5893 * this syscall writes the default timeslice value of a given process
5894 * into the user-space timespec buffer. A value of '0' means infinity.
5897 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
5899 struct task_struct
*p
;
5900 unsigned int time_slice
;
5908 read_lock(&tasklist_lock
);
5909 p
= find_process_by_pid(pid
);
5913 retval
= security_task_getscheduler(p
);
5918 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5919 * tasks that are on an otherwise idle runqueue:
5922 if (p
->policy
== SCHED_RR
) {
5923 time_slice
= DEF_TIMESLICE
;
5924 } else if (p
->policy
!= SCHED_FIFO
) {
5925 struct sched_entity
*se
= &p
->se
;
5926 unsigned long flags
;
5929 rq
= task_rq_lock(p
, &flags
);
5930 if (rq
->cfs
.load
.weight
)
5931 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5932 task_rq_unlock(rq
, &flags
);
5934 read_unlock(&tasklist_lock
);
5935 jiffies_to_timespec(time_slice
, &t
);
5936 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5940 read_unlock(&tasklist_lock
);
5944 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5946 void sched_show_task(struct task_struct
*p
)
5948 unsigned long free
= 0;
5951 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5952 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5953 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5954 #if BITS_PER_LONG == 32
5955 if (state
== TASK_RUNNING
)
5956 printk(KERN_CONT
" running ");
5958 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5960 if (state
== TASK_RUNNING
)
5961 printk(KERN_CONT
" running task ");
5963 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5965 #ifdef CONFIG_DEBUG_STACK_USAGE
5967 unsigned long *n
= end_of_stack(p
);
5970 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5973 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5974 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5976 show_stack(p
, NULL
);
5979 void show_state_filter(unsigned long state_filter
)
5981 struct task_struct
*g
, *p
;
5983 #if BITS_PER_LONG == 32
5985 " task PC stack pid father\n");
5988 " task PC stack pid father\n");
5990 read_lock(&tasklist_lock
);
5991 do_each_thread(g
, p
) {
5993 * reset the NMI-timeout, listing all files on a slow
5994 * console might take alot of time:
5996 touch_nmi_watchdog();
5997 if (!state_filter
|| (p
->state
& state_filter
))
5999 } while_each_thread(g
, p
);
6001 touch_all_softlockup_watchdogs();
6003 #ifdef CONFIG_SCHED_DEBUG
6004 sysrq_sched_debug_show();
6006 read_unlock(&tasklist_lock
);
6008 * Only show locks if all tasks are dumped:
6010 if (state_filter
== -1)
6011 debug_show_all_locks();
6014 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
6016 idle
->sched_class
= &idle_sched_class
;
6020 * init_idle - set up an idle thread for a given CPU
6021 * @idle: task in question
6022 * @cpu: cpu the idle task belongs to
6024 * NOTE: this function does not set the idle thread's NEED_RESCHED
6025 * flag, to make booting more robust.
6027 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
6029 struct rq
*rq
= cpu_rq(cpu
);
6030 unsigned long flags
;
6032 spin_lock_irqsave(&rq
->lock
, flags
);
6035 idle
->se
.exec_start
= sched_clock();
6037 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
6038 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
6039 __set_task_cpu(idle
, cpu
);
6041 rq
->curr
= rq
->idle
= idle
;
6042 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6045 spin_unlock_irqrestore(&rq
->lock
, flags
);
6047 /* Set the preempt count _outside_ the spinlocks! */
6048 #if defined(CONFIG_PREEMPT)
6049 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
6051 task_thread_info(idle
)->preempt_count
= 0;
6054 * The idle tasks have their own, simple scheduling class:
6056 idle
->sched_class
= &idle_sched_class
;
6057 ftrace_graph_init_task(idle
);
6061 * In a system that switches off the HZ timer nohz_cpu_mask
6062 * indicates which cpus entered this state. This is used
6063 * in the rcu update to wait only for active cpus. For system
6064 * which do not switch off the HZ timer nohz_cpu_mask should
6065 * always be CPU_BITS_NONE.
6067 cpumask_var_t nohz_cpu_mask
;
6070 * Increase the granularity value when there are more CPUs,
6071 * because with more CPUs the 'effective latency' as visible
6072 * to users decreases. But the relationship is not linear,
6073 * so pick a second-best guess by going with the log2 of the
6076 * This idea comes from the SD scheduler of Con Kolivas:
6078 static inline void sched_init_granularity(void)
6080 unsigned int factor
= 1 + ilog2(num_online_cpus());
6081 const unsigned long limit
= 200000000;
6083 sysctl_sched_min_granularity
*= factor
;
6084 if (sysctl_sched_min_granularity
> limit
)
6085 sysctl_sched_min_granularity
= limit
;
6087 sysctl_sched_latency
*= factor
;
6088 if (sysctl_sched_latency
> limit
)
6089 sysctl_sched_latency
= limit
;
6091 sysctl_sched_wakeup_granularity
*= factor
;
6093 sysctl_sched_shares_ratelimit
*= factor
;
6098 * This is how migration works:
6100 * 1) we queue a struct migration_req structure in the source CPU's
6101 * runqueue and wake up that CPU's migration thread.
6102 * 2) we down() the locked semaphore => thread blocks.
6103 * 3) migration thread wakes up (implicitly it forces the migrated
6104 * thread off the CPU)
6105 * 4) it gets the migration request and checks whether the migrated
6106 * task is still in the wrong runqueue.
6107 * 5) if it's in the wrong runqueue then the migration thread removes
6108 * it and puts it into the right queue.
6109 * 6) migration thread up()s the semaphore.
6110 * 7) we wake up and the migration is done.
6114 * Change a given task's CPU affinity. Migrate the thread to a
6115 * proper CPU and schedule it away if the CPU it's executing on
6116 * is removed from the allowed bitmask.
6118 * NOTE: the caller must have a valid reference to the task, the
6119 * task must not exit() & deallocate itself prematurely. The
6120 * call is not atomic; no spinlocks may be held.
6122 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
6124 struct migration_req req
;
6125 unsigned long flags
;
6129 rq
= task_rq_lock(p
, &flags
);
6130 if (!cpumask_intersects(new_mask
, cpu_online_mask
)) {
6135 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
6136 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
6141 if (p
->sched_class
->set_cpus_allowed
)
6142 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
6144 cpumask_copy(&p
->cpus_allowed
, new_mask
);
6145 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
6148 /* Can the task run on the task's current CPU? If so, we're done */
6149 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
6152 if (migrate_task(p
, cpumask_any_and(cpu_online_mask
, new_mask
), &req
)) {
6153 /* Need help from migration thread: drop lock and wait. */
6154 task_rq_unlock(rq
, &flags
);
6155 wake_up_process(rq
->migration_thread
);
6156 wait_for_completion(&req
.done
);
6157 tlb_migrate_finish(p
->mm
);
6161 task_rq_unlock(rq
, &flags
);
6165 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
6168 * Move (not current) task off this cpu, onto dest cpu. We're doing
6169 * this because either it can't run here any more (set_cpus_allowed()
6170 * away from this CPU, or CPU going down), or because we're
6171 * attempting to rebalance this task on exec (sched_exec).
6173 * So we race with normal scheduler movements, but that's OK, as long
6174 * as the task is no longer on this CPU.
6176 * Returns non-zero if task was successfully migrated.
6178 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6180 struct rq
*rq_dest
, *rq_src
;
6183 if (unlikely(!cpu_active(dest_cpu
)))
6186 rq_src
= cpu_rq(src_cpu
);
6187 rq_dest
= cpu_rq(dest_cpu
);
6189 double_rq_lock(rq_src
, rq_dest
);
6190 /* Already moved. */
6191 if (task_cpu(p
) != src_cpu
)
6193 /* Affinity changed (again). */
6194 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
6197 on_rq
= p
->se
.on_rq
;
6199 deactivate_task(rq_src
, p
, 0);
6201 set_task_cpu(p
, dest_cpu
);
6203 activate_task(rq_dest
, p
, 0);
6204 check_preempt_curr(rq_dest
, p
, 0);
6209 double_rq_unlock(rq_src
, rq_dest
);
6214 * migration_thread - this is a highprio system thread that performs
6215 * thread migration by bumping thread off CPU then 'pushing' onto
6218 static int migration_thread(void *data
)
6220 int cpu
= (long)data
;
6224 BUG_ON(rq
->migration_thread
!= current
);
6226 set_current_state(TASK_INTERRUPTIBLE
);
6227 while (!kthread_should_stop()) {
6228 struct migration_req
*req
;
6229 struct list_head
*head
;
6231 spin_lock_irq(&rq
->lock
);
6233 if (cpu_is_offline(cpu
)) {
6234 spin_unlock_irq(&rq
->lock
);
6238 if (rq
->active_balance
) {
6239 active_load_balance(rq
, cpu
);
6240 rq
->active_balance
= 0;
6243 head
= &rq
->migration_queue
;
6245 if (list_empty(head
)) {
6246 spin_unlock_irq(&rq
->lock
);
6248 set_current_state(TASK_INTERRUPTIBLE
);
6251 req
= list_entry(head
->next
, struct migration_req
, list
);
6252 list_del_init(head
->next
);
6254 spin_unlock(&rq
->lock
);
6255 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
6258 complete(&req
->done
);
6260 __set_current_state(TASK_RUNNING
);
6264 /* Wait for kthread_stop */
6265 set_current_state(TASK_INTERRUPTIBLE
);
6266 while (!kthread_should_stop()) {
6268 set_current_state(TASK_INTERRUPTIBLE
);
6270 __set_current_state(TASK_RUNNING
);
6274 #ifdef CONFIG_HOTPLUG_CPU
6276 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6280 local_irq_disable();
6281 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
6287 * Figure out where task on dead CPU should go, use force if necessary.
6289 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
6292 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(dead_cpu
));
6295 /* Look for allowed, online CPU in same node. */
6296 for_each_cpu_and(dest_cpu
, nodemask
, cpu_online_mask
)
6297 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
6300 /* Any allowed, online CPU? */
6301 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_online_mask
);
6302 if (dest_cpu
< nr_cpu_ids
)
6305 /* No more Mr. Nice Guy. */
6306 if (dest_cpu
>= nr_cpu_ids
) {
6307 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
6308 dest_cpu
= cpumask_any_and(cpu_online_mask
, &p
->cpus_allowed
);
6311 * Don't tell them about moving exiting tasks or
6312 * kernel threads (both mm NULL), since they never
6315 if (p
->mm
&& printk_ratelimit()) {
6316 printk(KERN_INFO
"process %d (%s) no "
6317 "longer affine to cpu%d\n",
6318 task_pid_nr(p
), p
->comm
, dead_cpu
);
6323 /* It can have affinity changed while we were choosing. */
6324 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
6329 * While a dead CPU has no uninterruptible tasks queued at this point,
6330 * it might still have a nonzero ->nr_uninterruptible counter, because
6331 * for performance reasons the counter is not stricly tracking tasks to
6332 * their home CPUs. So we just add the counter to another CPU's counter,
6333 * to keep the global sum constant after CPU-down:
6335 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
6337 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_online_mask
));
6338 unsigned long flags
;
6340 local_irq_save(flags
);
6341 double_rq_lock(rq_src
, rq_dest
);
6342 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
6343 rq_src
->nr_uninterruptible
= 0;
6344 double_rq_unlock(rq_src
, rq_dest
);
6345 local_irq_restore(flags
);
6348 /* Run through task list and migrate tasks from the dead cpu. */
6349 static void migrate_live_tasks(int src_cpu
)
6351 struct task_struct
*p
, *t
;
6353 read_lock(&tasklist_lock
);
6355 do_each_thread(t
, p
) {
6359 if (task_cpu(p
) == src_cpu
)
6360 move_task_off_dead_cpu(src_cpu
, p
);
6361 } while_each_thread(t
, p
);
6363 read_unlock(&tasklist_lock
);
6367 * Schedules idle task to be the next runnable task on current CPU.
6368 * It does so by boosting its priority to highest possible.
6369 * Used by CPU offline code.
6371 void sched_idle_next(void)
6373 int this_cpu
= smp_processor_id();
6374 struct rq
*rq
= cpu_rq(this_cpu
);
6375 struct task_struct
*p
= rq
->idle
;
6376 unsigned long flags
;
6378 /* cpu has to be offline */
6379 BUG_ON(cpu_online(this_cpu
));
6382 * Strictly not necessary since rest of the CPUs are stopped by now
6383 * and interrupts disabled on the current cpu.
6385 spin_lock_irqsave(&rq
->lock
, flags
);
6387 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6389 update_rq_clock(rq
);
6390 activate_task(rq
, p
, 0);
6392 spin_unlock_irqrestore(&rq
->lock
, flags
);
6396 * Ensures that the idle task is using init_mm right before its cpu goes
6399 void idle_task_exit(void)
6401 struct mm_struct
*mm
= current
->active_mm
;
6403 BUG_ON(cpu_online(smp_processor_id()));
6406 switch_mm(mm
, &init_mm
, current
);
6410 /* called under rq->lock with disabled interrupts */
6411 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
6413 struct rq
*rq
= cpu_rq(dead_cpu
);
6415 /* Must be exiting, otherwise would be on tasklist. */
6416 BUG_ON(!p
->exit_state
);
6418 /* Cannot have done final schedule yet: would have vanished. */
6419 BUG_ON(p
->state
== TASK_DEAD
);
6424 * Drop lock around migration; if someone else moves it,
6425 * that's OK. No task can be added to this CPU, so iteration is
6428 spin_unlock_irq(&rq
->lock
);
6429 move_task_off_dead_cpu(dead_cpu
, p
);
6430 spin_lock_irq(&rq
->lock
);
6435 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6436 static void migrate_dead_tasks(unsigned int dead_cpu
)
6438 struct rq
*rq
= cpu_rq(dead_cpu
);
6439 struct task_struct
*next
;
6442 if (!rq
->nr_running
)
6444 update_rq_clock(rq
);
6445 next
= pick_next_task(rq
, rq
->curr
);
6448 next
->sched_class
->put_prev_task(rq
, next
);
6449 migrate_dead(dead_cpu
, next
);
6453 #endif /* CONFIG_HOTPLUG_CPU */
6455 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6457 static struct ctl_table sd_ctl_dir
[] = {
6459 .procname
= "sched_domain",
6465 static struct ctl_table sd_ctl_root
[] = {
6467 .ctl_name
= CTL_KERN
,
6468 .procname
= "kernel",
6470 .child
= sd_ctl_dir
,
6475 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6477 struct ctl_table
*entry
=
6478 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6483 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6485 struct ctl_table
*entry
;
6488 * In the intermediate directories, both the child directory and
6489 * procname are dynamically allocated and could fail but the mode
6490 * will always be set. In the lowest directory the names are
6491 * static strings and all have proc handlers.
6493 for (entry
= *tablep
; entry
->mode
; entry
++) {
6495 sd_free_ctl_entry(&entry
->child
);
6496 if (entry
->proc_handler
== NULL
)
6497 kfree(entry
->procname
);
6505 set_table_entry(struct ctl_table
*entry
,
6506 const char *procname
, void *data
, int maxlen
,
6507 mode_t mode
, proc_handler
*proc_handler
)
6509 entry
->procname
= procname
;
6511 entry
->maxlen
= maxlen
;
6513 entry
->proc_handler
= proc_handler
;
6516 static struct ctl_table
*
6517 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6519 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
6524 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6525 sizeof(long), 0644, proc_doulongvec_minmax
);
6526 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6527 sizeof(long), 0644, proc_doulongvec_minmax
);
6528 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6529 sizeof(int), 0644, proc_dointvec_minmax
);
6530 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6531 sizeof(int), 0644, proc_dointvec_minmax
);
6532 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6533 sizeof(int), 0644, proc_dointvec_minmax
);
6534 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6535 sizeof(int), 0644, proc_dointvec_minmax
);
6536 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6537 sizeof(int), 0644, proc_dointvec_minmax
);
6538 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6539 sizeof(int), 0644, proc_dointvec_minmax
);
6540 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6541 sizeof(int), 0644, proc_dointvec_minmax
);
6542 set_table_entry(&table
[9], "cache_nice_tries",
6543 &sd
->cache_nice_tries
,
6544 sizeof(int), 0644, proc_dointvec_minmax
);
6545 set_table_entry(&table
[10], "flags", &sd
->flags
,
6546 sizeof(int), 0644, proc_dointvec_minmax
);
6547 set_table_entry(&table
[11], "name", sd
->name
,
6548 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
6549 /* &table[12] is terminator */
6554 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6556 struct ctl_table
*entry
, *table
;
6557 struct sched_domain
*sd
;
6558 int domain_num
= 0, i
;
6561 for_each_domain(cpu
, sd
)
6563 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6568 for_each_domain(cpu
, sd
) {
6569 snprintf(buf
, 32, "domain%d", i
);
6570 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6572 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6579 static struct ctl_table_header
*sd_sysctl_header
;
6580 static void register_sched_domain_sysctl(void)
6582 int i
, cpu_num
= num_online_cpus();
6583 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6586 WARN_ON(sd_ctl_dir
[0].child
);
6587 sd_ctl_dir
[0].child
= entry
;
6592 for_each_online_cpu(i
) {
6593 snprintf(buf
, 32, "cpu%d", i
);
6594 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6596 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6600 WARN_ON(sd_sysctl_header
);
6601 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6604 /* may be called multiple times per register */
6605 static void unregister_sched_domain_sysctl(void)
6607 if (sd_sysctl_header
)
6608 unregister_sysctl_table(sd_sysctl_header
);
6609 sd_sysctl_header
= NULL
;
6610 if (sd_ctl_dir
[0].child
)
6611 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6614 static void register_sched_domain_sysctl(void)
6617 static void unregister_sched_domain_sysctl(void)
6622 static void set_rq_online(struct rq
*rq
)
6625 const struct sched_class
*class;
6627 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
6630 for_each_class(class) {
6631 if (class->rq_online
)
6632 class->rq_online(rq
);
6637 static void set_rq_offline(struct rq
*rq
)
6640 const struct sched_class
*class;
6642 for_each_class(class) {
6643 if (class->rq_offline
)
6644 class->rq_offline(rq
);
6647 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
6653 * migration_call - callback that gets triggered when a CPU is added.
6654 * Here we can start up the necessary migration thread for the new CPU.
6656 static int __cpuinit
6657 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6659 struct task_struct
*p
;
6660 int cpu
= (long)hcpu
;
6661 unsigned long flags
;
6666 case CPU_UP_PREPARE
:
6667 case CPU_UP_PREPARE_FROZEN
:
6668 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
6671 kthread_bind(p
, cpu
);
6672 /* Must be high prio: stop_machine expects to yield to it. */
6673 rq
= task_rq_lock(p
, &flags
);
6674 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6675 task_rq_unlock(rq
, &flags
);
6676 cpu_rq(cpu
)->migration_thread
= p
;
6680 case CPU_ONLINE_FROZEN
:
6681 /* Strictly unnecessary, as first user will wake it. */
6682 wake_up_process(cpu_rq(cpu
)->migration_thread
);
6684 /* Update our root-domain */
6686 spin_lock_irqsave(&rq
->lock
, flags
);
6688 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6692 spin_unlock_irqrestore(&rq
->lock
, flags
);
6695 #ifdef CONFIG_HOTPLUG_CPU
6696 case CPU_UP_CANCELED
:
6697 case CPU_UP_CANCELED_FROZEN
:
6698 if (!cpu_rq(cpu
)->migration_thread
)
6700 /* Unbind it from offline cpu so it can run. Fall thru. */
6701 kthread_bind(cpu_rq(cpu
)->migration_thread
,
6702 cpumask_any(cpu_online_mask
));
6703 kthread_stop(cpu_rq(cpu
)->migration_thread
);
6704 cpu_rq(cpu
)->migration_thread
= NULL
;
6708 case CPU_DEAD_FROZEN
:
6709 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6710 migrate_live_tasks(cpu
);
6712 kthread_stop(rq
->migration_thread
);
6713 rq
->migration_thread
= NULL
;
6714 /* Idle task back to normal (off runqueue, low prio) */
6715 spin_lock_irq(&rq
->lock
);
6716 update_rq_clock(rq
);
6717 deactivate_task(rq
, rq
->idle
, 0);
6718 rq
->idle
->static_prio
= MAX_PRIO
;
6719 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
6720 rq
->idle
->sched_class
= &idle_sched_class
;
6721 migrate_dead_tasks(cpu
);
6722 spin_unlock_irq(&rq
->lock
);
6724 migrate_nr_uninterruptible(rq
);
6725 BUG_ON(rq
->nr_running
!= 0);
6728 * No need to migrate the tasks: it was best-effort if
6729 * they didn't take sched_hotcpu_mutex. Just wake up
6732 spin_lock_irq(&rq
->lock
);
6733 while (!list_empty(&rq
->migration_queue
)) {
6734 struct migration_req
*req
;
6736 req
= list_entry(rq
->migration_queue
.next
,
6737 struct migration_req
, list
);
6738 list_del_init(&req
->list
);
6739 spin_unlock_irq(&rq
->lock
);
6740 complete(&req
->done
);
6741 spin_lock_irq(&rq
->lock
);
6743 spin_unlock_irq(&rq
->lock
);
6747 case CPU_DYING_FROZEN
:
6748 /* Update our root-domain */
6750 spin_lock_irqsave(&rq
->lock
, flags
);
6752 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6755 spin_unlock_irqrestore(&rq
->lock
, flags
);
6762 /* Register at highest priority so that task migration (migrate_all_tasks)
6763 * happens before everything else.
6765 static struct notifier_block __cpuinitdata migration_notifier
= {
6766 .notifier_call
= migration_call
,
6770 static int __init
migration_init(void)
6772 void *cpu
= (void *)(long)smp_processor_id();
6775 /* Start one for the boot CPU: */
6776 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6777 BUG_ON(err
== NOTIFY_BAD
);
6778 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6779 register_cpu_notifier(&migration_notifier
);
6783 early_initcall(migration_init
);
6788 #ifdef CONFIG_SCHED_DEBUG
6790 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6791 struct cpumask
*groupmask
)
6793 struct sched_group
*group
= sd
->groups
;
6796 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
6797 cpumask_clear(groupmask
);
6799 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6801 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6802 printk("does not load-balance\n");
6804 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6809 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
6811 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
6812 printk(KERN_ERR
"ERROR: domain->span does not contain "
6815 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
6816 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6820 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6824 printk(KERN_ERR
"ERROR: group is NULL\n");
6828 if (!group
->__cpu_power
) {
6829 printk(KERN_CONT
"\n");
6830 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6835 if (!cpumask_weight(sched_group_cpus(group
))) {
6836 printk(KERN_CONT
"\n");
6837 printk(KERN_ERR
"ERROR: empty group\n");
6841 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
6842 printk(KERN_CONT
"\n");
6843 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6847 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
6849 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
6850 printk(KERN_CONT
" %s", str
);
6852 group
= group
->next
;
6853 } while (group
!= sd
->groups
);
6854 printk(KERN_CONT
"\n");
6856 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
6857 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6860 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
6861 printk(KERN_ERR
"ERROR: parent span is not a superset "
6862 "of domain->span\n");
6866 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6868 cpumask_var_t groupmask
;
6872 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6876 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6878 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
6879 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6884 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6891 free_cpumask_var(groupmask
);
6893 #else /* !CONFIG_SCHED_DEBUG */
6894 # define sched_domain_debug(sd, cpu) do { } while (0)
6895 #endif /* CONFIG_SCHED_DEBUG */
6897 static int sd_degenerate(struct sched_domain
*sd
)
6899 if (cpumask_weight(sched_domain_span(sd
)) == 1)
6902 /* Following flags need at least 2 groups */
6903 if (sd
->flags
& (SD_LOAD_BALANCE
|
6904 SD_BALANCE_NEWIDLE
|
6908 SD_SHARE_PKG_RESOURCES
)) {
6909 if (sd
->groups
!= sd
->groups
->next
)
6913 /* Following flags don't use groups */
6914 if (sd
->flags
& (SD_WAKE_IDLE
|
6923 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6925 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6927 if (sd_degenerate(parent
))
6930 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
6933 /* Does parent contain flags not in child? */
6934 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6935 if (cflags
& SD_WAKE_AFFINE
)
6936 pflags
&= ~SD_WAKE_BALANCE
;
6937 /* Flags needing groups don't count if only 1 group in parent */
6938 if (parent
->groups
== parent
->groups
->next
) {
6939 pflags
&= ~(SD_LOAD_BALANCE
|
6940 SD_BALANCE_NEWIDLE
|
6944 SD_SHARE_PKG_RESOURCES
);
6945 if (nr_node_ids
== 1)
6946 pflags
&= ~SD_SERIALIZE
;
6948 if (~cflags
& pflags
)
6954 static void free_rootdomain(struct root_domain
*rd
)
6956 cpupri_cleanup(&rd
->cpupri
);
6958 free_cpumask_var(rd
->rto_mask
);
6959 free_cpumask_var(rd
->online
);
6960 free_cpumask_var(rd
->span
);
6964 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6966 unsigned long flags
;
6968 spin_lock_irqsave(&rq
->lock
, flags
);
6971 struct root_domain
*old_rd
= rq
->rd
;
6973 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
6976 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
6978 if (atomic_dec_and_test(&old_rd
->refcount
))
6979 free_rootdomain(old_rd
);
6982 atomic_inc(&rd
->refcount
);
6985 cpumask_set_cpu(rq
->cpu
, rd
->span
);
6986 if (cpumask_test_cpu(rq
->cpu
, cpu_online_mask
))
6989 spin_unlock_irqrestore(&rq
->lock
, flags
);
6992 static int __init_refok
init_rootdomain(struct root_domain
*rd
, bool bootmem
)
6994 memset(rd
, 0, sizeof(*rd
));
6997 alloc_bootmem_cpumask_var(&def_root_domain
.span
);
6998 alloc_bootmem_cpumask_var(&def_root_domain
.online
);
6999 alloc_bootmem_cpumask_var(&def_root_domain
.rto_mask
);
7000 cpupri_init(&rd
->cpupri
, true);
7004 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
7006 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
7008 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
7011 if (cpupri_init(&rd
->cpupri
, false) != 0)
7016 free_cpumask_var(rd
->rto_mask
);
7018 free_cpumask_var(rd
->online
);
7020 free_cpumask_var(rd
->span
);
7025 static void init_defrootdomain(void)
7027 init_rootdomain(&def_root_domain
, true);
7029 atomic_set(&def_root_domain
.refcount
, 1);
7032 static struct root_domain
*alloc_rootdomain(void)
7034 struct root_domain
*rd
;
7036 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
7040 if (init_rootdomain(rd
, false) != 0) {
7049 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7050 * hold the hotplug lock.
7053 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
7055 struct rq
*rq
= cpu_rq(cpu
);
7056 struct sched_domain
*tmp
;
7058 /* Remove the sched domains which do not contribute to scheduling. */
7059 for (tmp
= sd
; tmp
; ) {
7060 struct sched_domain
*parent
= tmp
->parent
;
7064 if (sd_parent_degenerate(tmp
, parent
)) {
7065 tmp
->parent
= parent
->parent
;
7067 parent
->parent
->child
= tmp
;
7072 if (sd
&& sd_degenerate(sd
)) {
7078 sched_domain_debug(sd
, cpu
);
7080 rq_attach_root(rq
, rd
);
7081 rcu_assign_pointer(rq
->sd
, sd
);
7084 /* cpus with isolated domains */
7085 static cpumask_var_t cpu_isolated_map
;
7087 /* Setup the mask of cpus configured for isolated domains */
7088 static int __init
isolated_cpu_setup(char *str
)
7090 cpulist_parse(str
, cpu_isolated_map
);
7094 __setup("isolcpus=", isolated_cpu_setup
);
7097 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7098 * to a function which identifies what group(along with sched group) a CPU
7099 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7100 * (due to the fact that we keep track of groups covered with a struct cpumask).
7102 * init_sched_build_groups will build a circular linked list of the groups
7103 * covered by the given span, and will set each group's ->cpumask correctly,
7104 * and ->cpu_power to 0.
7107 init_sched_build_groups(const struct cpumask
*span
,
7108 const struct cpumask
*cpu_map
,
7109 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
7110 struct sched_group
**sg
,
7111 struct cpumask
*tmpmask
),
7112 struct cpumask
*covered
, struct cpumask
*tmpmask
)
7114 struct sched_group
*first
= NULL
, *last
= NULL
;
7117 cpumask_clear(covered
);
7119 for_each_cpu(i
, span
) {
7120 struct sched_group
*sg
;
7121 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
7124 if (cpumask_test_cpu(i
, covered
))
7127 cpumask_clear(sched_group_cpus(sg
));
7128 sg
->__cpu_power
= 0;
7130 for_each_cpu(j
, span
) {
7131 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
7134 cpumask_set_cpu(j
, covered
);
7135 cpumask_set_cpu(j
, sched_group_cpus(sg
));
7146 #define SD_NODES_PER_DOMAIN 16
7151 * find_next_best_node - find the next node to include in a sched_domain
7152 * @node: node whose sched_domain we're building
7153 * @used_nodes: nodes already in the sched_domain
7155 * Find the next node to include in a given scheduling domain. Simply
7156 * finds the closest node not already in the @used_nodes map.
7158 * Should use nodemask_t.
7160 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
7162 int i
, n
, val
, min_val
, best_node
= 0;
7166 for (i
= 0; i
< nr_node_ids
; i
++) {
7167 /* Start at @node */
7168 n
= (node
+ i
) % nr_node_ids
;
7170 if (!nr_cpus_node(n
))
7173 /* Skip already used nodes */
7174 if (node_isset(n
, *used_nodes
))
7177 /* Simple min distance search */
7178 val
= node_distance(node
, n
);
7180 if (val
< min_val
) {
7186 node_set(best_node
, *used_nodes
);
7191 * sched_domain_node_span - get a cpumask for a node's sched_domain
7192 * @node: node whose cpumask we're constructing
7193 * @span: resulting cpumask
7195 * Given a node, construct a good cpumask for its sched_domain to span. It
7196 * should be one that prevents unnecessary balancing, but also spreads tasks
7199 static void sched_domain_node_span(int node
, struct cpumask
*span
)
7201 nodemask_t used_nodes
;
7204 cpumask_clear(span
);
7205 nodes_clear(used_nodes
);
7207 cpumask_or(span
, span
, cpumask_of_node(node
));
7208 node_set(node
, used_nodes
);
7210 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
7211 int next_node
= find_next_best_node(node
, &used_nodes
);
7213 cpumask_or(span
, span
, cpumask_of_node(next_node
));
7216 #endif /* CONFIG_NUMA */
7218 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
7221 * The cpus mask in sched_group and sched_domain hangs off the end.
7222 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7223 * for nr_cpu_ids < CONFIG_NR_CPUS.
7225 struct static_sched_group
{
7226 struct sched_group sg
;
7227 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
7230 struct static_sched_domain
{
7231 struct sched_domain sd
;
7232 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
7236 * SMT sched-domains:
7238 #ifdef CONFIG_SCHED_SMT
7239 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
7240 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_cpus
);
7243 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
7244 struct sched_group
**sg
, struct cpumask
*unused
)
7247 *sg
= &per_cpu(sched_group_cpus
, cpu
).sg
;
7250 #endif /* CONFIG_SCHED_SMT */
7253 * multi-core sched-domains:
7255 #ifdef CONFIG_SCHED_MC
7256 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
7257 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
7258 #endif /* CONFIG_SCHED_MC */
7260 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7262 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
7263 struct sched_group
**sg
, struct cpumask
*mask
)
7267 cpumask_and(mask
, &per_cpu(cpu_sibling_map
, cpu
), cpu_map
);
7268 group
= cpumask_first(mask
);
7270 *sg
= &per_cpu(sched_group_core
, group
).sg
;
7273 #elif defined(CONFIG_SCHED_MC)
7275 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
7276 struct sched_group
**sg
, struct cpumask
*unused
)
7279 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
7284 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
7285 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
7288 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
7289 struct sched_group
**sg
, struct cpumask
*mask
)
7292 #ifdef CONFIG_SCHED_MC
7293 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
7294 group
= cpumask_first(mask
);
7295 #elif defined(CONFIG_SCHED_SMT)
7296 cpumask_and(mask
, &per_cpu(cpu_sibling_map
, cpu
), cpu_map
);
7297 group
= cpumask_first(mask
);
7302 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
7308 * The init_sched_build_groups can't handle what we want to do with node
7309 * groups, so roll our own. Now each node has its own list of groups which
7310 * gets dynamically allocated.
7312 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
7313 static struct sched_group
***sched_group_nodes_bycpu
;
7315 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
7316 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
7318 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
7319 struct sched_group
**sg
,
7320 struct cpumask
*nodemask
)
7324 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
7325 group
= cpumask_first(nodemask
);
7328 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
7332 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
7334 struct sched_group
*sg
= group_head
;
7340 for_each_cpu(j
, sched_group_cpus(sg
)) {
7341 struct sched_domain
*sd
;
7343 sd
= &per_cpu(phys_domains
, j
).sd
;
7344 if (j
!= cpumask_first(sched_group_cpus(sd
->groups
))) {
7346 * Only add "power" once for each
7352 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
7355 } while (sg
!= group_head
);
7357 #endif /* CONFIG_NUMA */
7360 /* Free memory allocated for various sched_group structures */
7361 static void free_sched_groups(const struct cpumask
*cpu_map
,
7362 struct cpumask
*nodemask
)
7366 for_each_cpu(cpu
, cpu_map
) {
7367 struct sched_group
**sched_group_nodes
7368 = sched_group_nodes_bycpu
[cpu
];
7370 if (!sched_group_nodes
)
7373 for (i
= 0; i
< nr_node_ids
; i
++) {
7374 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
7376 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
7377 if (cpumask_empty(nodemask
))
7387 if (oldsg
!= sched_group_nodes
[i
])
7390 kfree(sched_group_nodes
);
7391 sched_group_nodes_bycpu
[cpu
] = NULL
;
7394 #else /* !CONFIG_NUMA */
7395 static void free_sched_groups(const struct cpumask
*cpu_map
,
7396 struct cpumask
*nodemask
)
7399 #endif /* CONFIG_NUMA */
7402 * Initialize sched groups cpu_power.
7404 * cpu_power indicates the capacity of sched group, which is used while
7405 * distributing the load between different sched groups in a sched domain.
7406 * Typically cpu_power for all the groups in a sched domain will be same unless
7407 * there are asymmetries in the topology. If there are asymmetries, group
7408 * having more cpu_power will pickup more load compared to the group having
7411 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7412 * the maximum number of tasks a group can handle in the presence of other idle
7413 * or lightly loaded groups in the same sched domain.
7415 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7417 struct sched_domain
*child
;
7418 struct sched_group
*group
;
7420 WARN_ON(!sd
|| !sd
->groups
);
7422 if (cpu
!= cpumask_first(sched_group_cpus(sd
->groups
)))
7427 sd
->groups
->__cpu_power
= 0;
7430 * For perf policy, if the groups in child domain share resources
7431 * (for example cores sharing some portions of the cache hierarchy
7432 * or SMT), then set this domain groups cpu_power such that each group
7433 * can handle only one task, when there are other idle groups in the
7434 * same sched domain.
7436 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
7438 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
7439 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
7444 * add cpu_power of each child group to this groups cpu_power
7446 group
= child
->groups
;
7448 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
7449 group
= group
->next
;
7450 } while (group
!= child
->groups
);
7454 * Initializers for schedule domains
7455 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7458 #ifdef CONFIG_SCHED_DEBUG
7459 # define SD_INIT_NAME(sd, type) sd->name = #type
7461 # define SD_INIT_NAME(sd, type) do { } while (0)
7464 #define SD_INIT(sd, type) sd_init_##type(sd)
7466 #define SD_INIT_FUNC(type) \
7467 static noinline void sd_init_##type(struct sched_domain *sd) \
7469 memset(sd, 0, sizeof(*sd)); \
7470 *sd = SD_##type##_INIT; \
7471 sd->level = SD_LV_##type; \
7472 SD_INIT_NAME(sd, type); \
7477 SD_INIT_FUNC(ALLNODES
)
7480 #ifdef CONFIG_SCHED_SMT
7481 SD_INIT_FUNC(SIBLING
)
7483 #ifdef CONFIG_SCHED_MC
7487 static int default_relax_domain_level
= -1;
7489 static int __init
setup_relax_domain_level(char *str
)
7493 val
= simple_strtoul(str
, NULL
, 0);
7494 if (val
< SD_LV_MAX
)
7495 default_relax_domain_level
= val
;
7499 __setup("relax_domain_level=", setup_relax_domain_level
);
7501 static void set_domain_attribute(struct sched_domain
*sd
,
7502 struct sched_domain_attr
*attr
)
7506 if (!attr
|| attr
->relax_domain_level
< 0) {
7507 if (default_relax_domain_level
< 0)
7510 request
= default_relax_domain_level
;
7512 request
= attr
->relax_domain_level
;
7513 if (request
< sd
->level
) {
7514 /* turn off idle balance on this domain */
7515 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
7517 /* turn on idle balance on this domain */
7518 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
7523 * Build sched domains for a given set of cpus and attach the sched domains
7524 * to the individual cpus
7526 static int __build_sched_domains(const struct cpumask
*cpu_map
,
7527 struct sched_domain_attr
*attr
)
7529 int i
, err
= -ENOMEM
;
7530 struct root_domain
*rd
;
7531 cpumask_var_t nodemask
, this_sibling_map
, this_core_map
, send_covered
,
7534 cpumask_var_t domainspan
, covered
, notcovered
;
7535 struct sched_group
**sched_group_nodes
= NULL
;
7536 int sd_allnodes
= 0;
7538 if (!alloc_cpumask_var(&domainspan
, GFP_KERNEL
))
7540 if (!alloc_cpumask_var(&covered
, GFP_KERNEL
))
7541 goto free_domainspan
;
7542 if (!alloc_cpumask_var(¬covered
, GFP_KERNEL
))
7546 if (!alloc_cpumask_var(&nodemask
, GFP_KERNEL
))
7547 goto free_notcovered
;
7548 if (!alloc_cpumask_var(&this_sibling_map
, GFP_KERNEL
))
7550 if (!alloc_cpumask_var(&this_core_map
, GFP_KERNEL
))
7551 goto free_this_sibling_map
;
7552 if (!alloc_cpumask_var(&send_covered
, GFP_KERNEL
))
7553 goto free_this_core_map
;
7554 if (!alloc_cpumask_var(&tmpmask
, GFP_KERNEL
))
7555 goto free_send_covered
;
7559 * Allocate the per-node list of sched groups
7561 sched_group_nodes
= kcalloc(nr_node_ids
, sizeof(struct sched_group
*),
7563 if (!sched_group_nodes
) {
7564 printk(KERN_WARNING
"Can not alloc sched group node list\n");
7569 rd
= alloc_rootdomain();
7571 printk(KERN_WARNING
"Cannot alloc root domain\n");
7572 goto free_sched_groups
;
7576 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = sched_group_nodes
;
7580 * Set up domains for cpus specified by the cpu_map.
7582 for_each_cpu(i
, cpu_map
) {
7583 struct sched_domain
*sd
= NULL
, *p
;
7585 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(i
)), cpu_map
);
7588 if (cpumask_weight(cpu_map
) >
7589 SD_NODES_PER_DOMAIN
*cpumask_weight(nodemask
)) {
7590 sd
= &per_cpu(allnodes_domains
, i
).sd
;
7591 SD_INIT(sd
, ALLNODES
);
7592 set_domain_attribute(sd
, attr
);
7593 cpumask_copy(sched_domain_span(sd
), cpu_map
);
7594 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7600 sd
= &per_cpu(node_domains
, i
).sd
;
7602 set_domain_attribute(sd
, attr
);
7603 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
7607 cpumask_and(sched_domain_span(sd
),
7608 sched_domain_span(sd
), cpu_map
);
7612 sd
= &per_cpu(phys_domains
, i
).sd
;
7614 set_domain_attribute(sd
, attr
);
7615 cpumask_copy(sched_domain_span(sd
), nodemask
);
7619 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7621 #ifdef CONFIG_SCHED_MC
7623 sd
= &per_cpu(core_domains
, i
).sd
;
7625 set_domain_attribute(sd
, attr
);
7626 cpumask_and(sched_domain_span(sd
), cpu_map
,
7627 cpu_coregroup_mask(i
));
7630 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7633 #ifdef CONFIG_SCHED_SMT
7635 sd
= &per_cpu(cpu_domains
, i
).sd
;
7636 SD_INIT(sd
, SIBLING
);
7637 set_domain_attribute(sd
, attr
);
7638 cpumask_and(sched_domain_span(sd
),
7639 &per_cpu(cpu_sibling_map
, i
), cpu_map
);
7642 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7646 #ifdef CONFIG_SCHED_SMT
7647 /* Set up CPU (sibling) groups */
7648 for_each_cpu(i
, cpu_map
) {
7649 cpumask_and(this_sibling_map
,
7650 &per_cpu(cpu_sibling_map
, i
), cpu_map
);
7651 if (i
!= cpumask_first(this_sibling_map
))
7654 init_sched_build_groups(this_sibling_map
, cpu_map
,
7656 send_covered
, tmpmask
);
7660 #ifdef CONFIG_SCHED_MC
7661 /* Set up multi-core groups */
7662 for_each_cpu(i
, cpu_map
) {
7663 cpumask_and(this_core_map
, cpu_coregroup_mask(i
), cpu_map
);
7664 if (i
!= cpumask_first(this_core_map
))
7667 init_sched_build_groups(this_core_map
, cpu_map
,
7669 send_covered
, tmpmask
);
7673 /* Set up physical groups */
7674 for (i
= 0; i
< nr_node_ids
; i
++) {
7675 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
7676 if (cpumask_empty(nodemask
))
7679 init_sched_build_groups(nodemask
, cpu_map
,
7681 send_covered
, tmpmask
);
7685 /* Set up node groups */
7687 init_sched_build_groups(cpu_map
, cpu_map
,
7688 &cpu_to_allnodes_group
,
7689 send_covered
, tmpmask
);
7692 for (i
= 0; i
< nr_node_ids
; i
++) {
7693 /* Set up node groups */
7694 struct sched_group
*sg
, *prev
;
7697 cpumask_clear(covered
);
7698 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
7699 if (cpumask_empty(nodemask
)) {
7700 sched_group_nodes
[i
] = NULL
;
7704 sched_domain_node_span(i
, domainspan
);
7705 cpumask_and(domainspan
, domainspan
, cpu_map
);
7707 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
7710 printk(KERN_WARNING
"Can not alloc domain group for "
7714 sched_group_nodes
[i
] = sg
;
7715 for_each_cpu(j
, nodemask
) {
7716 struct sched_domain
*sd
;
7718 sd
= &per_cpu(node_domains
, j
).sd
;
7721 sg
->__cpu_power
= 0;
7722 cpumask_copy(sched_group_cpus(sg
), nodemask
);
7724 cpumask_or(covered
, covered
, nodemask
);
7727 for (j
= 0; j
< nr_node_ids
; j
++) {
7728 int n
= (i
+ j
) % nr_node_ids
;
7730 cpumask_complement(notcovered
, covered
);
7731 cpumask_and(tmpmask
, notcovered
, cpu_map
);
7732 cpumask_and(tmpmask
, tmpmask
, domainspan
);
7733 if (cpumask_empty(tmpmask
))
7736 cpumask_and(tmpmask
, tmpmask
, cpumask_of_node(n
));
7737 if (cpumask_empty(tmpmask
))
7740 sg
= kmalloc_node(sizeof(struct sched_group
) +
7745 "Can not alloc domain group for node %d\n", j
);
7748 sg
->__cpu_power
= 0;
7749 cpumask_copy(sched_group_cpus(sg
), tmpmask
);
7750 sg
->next
= prev
->next
;
7751 cpumask_or(covered
, covered
, tmpmask
);
7758 /* Calculate CPU power for physical packages and nodes */
7759 #ifdef CONFIG_SCHED_SMT
7760 for_each_cpu(i
, cpu_map
) {
7761 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
).sd
;
7763 init_sched_groups_power(i
, sd
);
7766 #ifdef CONFIG_SCHED_MC
7767 for_each_cpu(i
, cpu_map
) {
7768 struct sched_domain
*sd
= &per_cpu(core_domains
, i
).sd
;
7770 init_sched_groups_power(i
, sd
);
7774 for_each_cpu(i
, cpu_map
) {
7775 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
).sd
;
7777 init_sched_groups_power(i
, sd
);
7781 for (i
= 0; i
< nr_node_ids
; i
++)
7782 init_numa_sched_groups_power(sched_group_nodes
[i
]);
7785 struct sched_group
*sg
;
7787 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
7789 init_numa_sched_groups_power(sg
);
7793 /* Attach the domains */
7794 for_each_cpu(i
, cpu_map
) {
7795 struct sched_domain
*sd
;
7796 #ifdef CONFIG_SCHED_SMT
7797 sd
= &per_cpu(cpu_domains
, i
).sd
;
7798 #elif defined(CONFIG_SCHED_MC)
7799 sd
= &per_cpu(core_domains
, i
).sd
;
7801 sd
= &per_cpu(phys_domains
, i
).sd
;
7803 cpu_attach_domain(sd
, rd
, i
);
7809 free_cpumask_var(tmpmask
);
7811 free_cpumask_var(send_covered
);
7813 free_cpumask_var(this_core_map
);
7814 free_this_sibling_map
:
7815 free_cpumask_var(this_sibling_map
);
7817 free_cpumask_var(nodemask
);
7820 free_cpumask_var(notcovered
);
7822 free_cpumask_var(covered
);
7824 free_cpumask_var(domainspan
);
7831 kfree(sched_group_nodes
);
7837 free_sched_groups(cpu_map
, tmpmask
);
7838 free_rootdomain(rd
);
7843 static int build_sched_domains(const struct cpumask
*cpu_map
)
7845 return __build_sched_domains(cpu_map
, NULL
);
7848 static struct cpumask
*doms_cur
; /* current sched domains */
7849 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7850 static struct sched_domain_attr
*dattr_cur
;
7851 /* attribues of custom domains in 'doms_cur' */
7854 * Special case: If a kmalloc of a doms_cur partition (array of
7855 * cpumask) fails, then fallback to a single sched domain,
7856 * as determined by the single cpumask fallback_doms.
7858 static cpumask_var_t fallback_doms
;
7861 * arch_update_cpu_topology lets virtualized architectures update the
7862 * cpu core maps. It is supposed to return 1 if the topology changed
7863 * or 0 if it stayed the same.
7865 int __attribute__((weak
)) arch_update_cpu_topology(void)
7871 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7872 * For now this just excludes isolated cpus, but could be used to
7873 * exclude other special cases in the future.
7875 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
7879 arch_update_cpu_topology();
7881 doms_cur
= kmalloc(cpumask_size(), GFP_KERNEL
);
7883 doms_cur
= fallback_doms
;
7884 cpumask_andnot(doms_cur
, cpu_map
, cpu_isolated_map
);
7886 err
= build_sched_domains(doms_cur
);
7887 register_sched_domain_sysctl();
7892 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
7893 struct cpumask
*tmpmask
)
7895 free_sched_groups(cpu_map
, tmpmask
);
7899 * Detach sched domains from a group of cpus specified in cpu_map
7900 * These cpus will now be attached to the NULL domain
7902 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7904 /* Save because hotplug lock held. */
7905 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
7908 for_each_cpu(i
, cpu_map
)
7909 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7910 synchronize_sched();
7911 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
7914 /* handle null as "default" */
7915 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7916 struct sched_domain_attr
*new, int idx_new
)
7918 struct sched_domain_attr tmp
;
7925 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7926 new ? (new + idx_new
) : &tmp
,
7927 sizeof(struct sched_domain_attr
));
7931 * Partition sched domains as specified by the 'ndoms_new'
7932 * cpumasks in the array doms_new[] of cpumasks. This compares
7933 * doms_new[] to the current sched domain partitioning, doms_cur[].
7934 * It destroys each deleted domain and builds each new domain.
7936 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
7937 * The masks don't intersect (don't overlap.) We should setup one
7938 * sched domain for each mask. CPUs not in any of the cpumasks will
7939 * not be load balanced. If the same cpumask appears both in the
7940 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7943 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7944 * ownership of it and will kfree it when done with it. If the caller
7945 * failed the kmalloc call, then it can pass in doms_new == NULL &&
7946 * ndoms_new == 1, and partition_sched_domains() will fallback to
7947 * the single partition 'fallback_doms', it also forces the domains
7950 * If doms_new == NULL it will be replaced with cpu_online_mask.
7951 * ndoms_new == 0 is a special case for destroying existing domains,
7952 * and it will not create the default domain.
7954 * Call with hotplug lock held
7956 /* FIXME: Change to struct cpumask *doms_new[] */
7957 void partition_sched_domains(int ndoms_new
, struct cpumask
*doms_new
,
7958 struct sched_domain_attr
*dattr_new
)
7963 mutex_lock(&sched_domains_mutex
);
7965 /* always unregister in case we don't destroy any domains */
7966 unregister_sched_domain_sysctl();
7968 /* Let architecture update cpu core mappings. */
7969 new_topology
= arch_update_cpu_topology();
7971 n
= doms_new
? ndoms_new
: 0;
7973 /* Destroy deleted domains */
7974 for (i
= 0; i
< ndoms_cur
; i
++) {
7975 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7976 if (cpumask_equal(&doms_cur
[i
], &doms_new
[j
])
7977 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7980 /* no match - a current sched domain not in new doms_new[] */
7981 detach_destroy_domains(doms_cur
+ i
);
7986 if (doms_new
== NULL
) {
7988 doms_new
= fallback_doms
;
7989 cpumask_andnot(&doms_new
[0], cpu_online_mask
, cpu_isolated_map
);
7990 WARN_ON_ONCE(dattr_new
);
7993 /* Build new domains */
7994 for (i
= 0; i
< ndoms_new
; i
++) {
7995 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
7996 if (cpumask_equal(&doms_new
[i
], &doms_cur
[j
])
7997 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
8000 /* no match - add a new doms_new */
8001 __build_sched_domains(doms_new
+ i
,
8002 dattr_new
? dattr_new
+ i
: NULL
);
8007 /* Remember the new sched domains */
8008 if (doms_cur
!= fallback_doms
)
8010 kfree(dattr_cur
); /* kfree(NULL) is safe */
8011 doms_cur
= doms_new
;
8012 dattr_cur
= dattr_new
;
8013 ndoms_cur
= ndoms_new
;
8015 register_sched_domain_sysctl();
8017 mutex_unlock(&sched_domains_mutex
);
8020 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8021 static void arch_reinit_sched_domains(void)
8025 /* Destroy domains first to force the rebuild */
8026 partition_sched_domains(0, NULL
, NULL
);
8028 rebuild_sched_domains();
8032 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
8034 unsigned int level
= 0;
8036 if (sscanf(buf
, "%u", &level
) != 1)
8040 * level is always be positive so don't check for
8041 * level < POWERSAVINGS_BALANCE_NONE which is 0
8042 * What happens on 0 or 1 byte write,
8043 * need to check for count as well?
8046 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
8050 sched_smt_power_savings
= level
;
8052 sched_mc_power_savings
= level
;
8054 arch_reinit_sched_domains();
8059 #ifdef CONFIG_SCHED_MC
8060 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
8063 return sprintf(page
, "%u\n", sched_mc_power_savings
);
8065 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
8066 const char *buf
, size_t count
)
8068 return sched_power_savings_store(buf
, count
, 0);
8070 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
8071 sched_mc_power_savings_show
,
8072 sched_mc_power_savings_store
);
8075 #ifdef CONFIG_SCHED_SMT
8076 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
8079 return sprintf(page
, "%u\n", sched_smt_power_savings
);
8081 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
8082 const char *buf
, size_t count
)
8084 return sched_power_savings_store(buf
, count
, 1);
8086 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
8087 sched_smt_power_savings_show
,
8088 sched_smt_power_savings_store
);
8091 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
8095 #ifdef CONFIG_SCHED_SMT
8097 err
= sysfs_create_file(&cls
->kset
.kobj
,
8098 &attr_sched_smt_power_savings
.attr
);
8100 #ifdef CONFIG_SCHED_MC
8101 if (!err
&& mc_capable())
8102 err
= sysfs_create_file(&cls
->kset
.kobj
,
8103 &attr_sched_mc_power_savings
.attr
);
8107 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8109 #ifndef CONFIG_CPUSETS
8111 * Add online and remove offline CPUs from the scheduler domains.
8112 * When cpusets are enabled they take over this function.
8114 static int update_sched_domains(struct notifier_block
*nfb
,
8115 unsigned long action
, void *hcpu
)
8119 case CPU_ONLINE_FROZEN
:
8121 case CPU_DEAD_FROZEN
:
8122 partition_sched_domains(1, NULL
, NULL
);
8131 static int update_runtime(struct notifier_block
*nfb
,
8132 unsigned long action
, void *hcpu
)
8134 int cpu
= (int)(long)hcpu
;
8137 case CPU_DOWN_PREPARE
:
8138 case CPU_DOWN_PREPARE_FROZEN
:
8139 disable_runtime(cpu_rq(cpu
));
8142 case CPU_DOWN_FAILED
:
8143 case CPU_DOWN_FAILED_FROZEN
:
8145 case CPU_ONLINE_FROZEN
:
8146 enable_runtime(cpu_rq(cpu
));
8154 void __init
sched_init_smp(void)
8156 cpumask_var_t non_isolated_cpus
;
8158 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
8160 #if defined(CONFIG_NUMA)
8161 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
8163 BUG_ON(sched_group_nodes_bycpu
== NULL
);
8166 mutex_lock(&sched_domains_mutex
);
8167 arch_init_sched_domains(cpu_online_mask
);
8168 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
8169 if (cpumask_empty(non_isolated_cpus
))
8170 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
8171 mutex_unlock(&sched_domains_mutex
);
8174 #ifndef CONFIG_CPUSETS
8175 /* XXX: Theoretical race here - CPU may be hotplugged now */
8176 hotcpu_notifier(update_sched_domains
, 0);
8179 /* RT runtime code needs to handle some hotplug events */
8180 hotcpu_notifier(update_runtime
, 0);
8184 /* Move init over to a non-isolated CPU */
8185 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
8187 sched_init_granularity();
8188 free_cpumask_var(non_isolated_cpus
);
8190 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
8191 init_sched_rt_class();
8194 void __init
sched_init_smp(void)
8196 sched_init_granularity();
8198 #endif /* CONFIG_SMP */
8200 int in_sched_functions(unsigned long addr
)
8202 return in_lock_functions(addr
) ||
8203 (addr
>= (unsigned long)__sched_text_start
8204 && addr
< (unsigned long)__sched_text_end
);
8207 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
8209 cfs_rq
->tasks_timeline
= RB_ROOT
;
8210 INIT_LIST_HEAD(&cfs_rq
->tasks
);
8211 #ifdef CONFIG_FAIR_GROUP_SCHED
8214 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
8217 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
8219 struct rt_prio_array
*array
;
8222 array
= &rt_rq
->active
;
8223 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
8224 INIT_LIST_HEAD(array
->queue
+ i
);
8225 __clear_bit(i
, array
->bitmap
);
8227 /* delimiter for bitsearch: */
8228 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
8230 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8231 rt_rq
->highest_prio
= MAX_RT_PRIO
;
8234 rt_rq
->rt_nr_migratory
= 0;
8235 rt_rq
->overloaded
= 0;
8239 rt_rq
->rt_throttled
= 0;
8240 rt_rq
->rt_runtime
= 0;
8241 spin_lock_init(&rt_rq
->rt_runtime_lock
);
8243 #ifdef CONFIG_RT_GROUP_SCHED
8244 rt_rq
->rt_nr_boosted
= 0;
8249 #ifdef CONFIG_FAIR_GROUP_SCHED
8250 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8251 struct sched_entity
*se
, int cpu
, int add
,
8252 struct sched_entity
*parent
)
8254 struct rq
*rq
= cpu_rq(cpu
);
8255 tg
->cfs_rq
[cpu
] = cfs_rq
;
8256 init_cfs_rq(cfs_rq
, rq
);
8259 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
8262 /* se could be NULL for init_task_group */
8267 se
->cfs_rq
= &rq
->cfs
;
8269 se
->cfs_rq
= parent
->my_q
;
8272 se
->load
.weight
= tg
->shares
;
8273 se
->load
.inv_weight
= 0;
8274 se
->parent
= parent
;
8278 #ifdef CONFIG_RT_GROUP_SCHED
8279 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
8280 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
8281 struct sched_rt_entity
*parent
)
8283 struct rq
*rq
= cpu_rq(cpu
);
8285 tg
->rt_rq
[cpu
] = rt_rq
;
8286 init_rt_rq(rt_rq
, rq
);
8288 rt_rq
->rt_se
= rt_se
;
8289 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8291 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
8293 tg
->rt_se
[cpu
] = rt_se
;
8298 rt_se
->rt_rq
= &rq
->rt
;
8300 rt_se
->rt_rq
= parent
->my_q
;
8302 rt_se
->my_q
= rt_rq
;
8303 rt_se
->parent
= parent
;
8304 INIT_LIST_HEAD(&rt_se
->run_list
);
8308 void __init
sched_init(void)
8311 unsigned long alloc_size
= 0, ptr
;
8313 #ifdef CONFIG_FAIR_GROUP_SCHED
8314 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8316 #ifdef CONFIG_RT_GROUP_SCHED
8317 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8319 #ifdef CONFIG_USER_SCHED
8323 * As sched_init() is called before page_alloc is setup,
8324 * we use alloc_bootmem().
8327 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
8329 #ifdef CONFIG_FAIR_GROUP_SCHED
8330 init_task_group
.se
= (struct sched_entity
**)ptr
;
8331 ptr
+= nr_cpu_ids
* sizeof(void **);
8333 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8334 ptr
+= nr_cpu_ids
* sizeof(void **);
8336 #ifdef CONFIG_USER_SCHED
8337 root_task_group
.se
= (struct sched_entity
**)ptr
;
8338 ptr
+= nr_cpu_ids
* sizeof(void **);
8340 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8341 ptr
+= nr_cpu_ids
* sizeof(void **);
8342 #endif /* CONFIG_USER_SCHED */
8343 #endif /* CONFIG_FAIR_GROUP_SCHED */
8344 #ifdef CONFIG_RT_GROUP_SCHED
8345 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8346 ptr
+= nr_cpu_ids
* sizeof(void **);
8348 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8349 ptr
+= nr_cpu_ids
* sizeof(void **);
8351 #ifdef CONFIG_USER_SCHED
8352 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8353 ptr
+= nr_cpu_ids
* sizeof(void **);
8355 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8356 ptr
+= nr_cpu_ids
* sizeof(void **);
8357 #endif /* CONFIG_USER_SCHED */
8358 #endif /* CONFIG_RT_GROUP_SCHED */
8362 init_defrootdomain();
8365 init_rt_bandwidth(&def_rt_bandwidth
,
8366 global_rt_period(), global_rt_runtime());
8368 #ifdef CONFIG_RT_GROUP_SCHED
8369 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
8370 global_rt_period(), global_rt_runtime());
8371 #ifdef CONFIG_USER_SCHED
8372 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8373 global_rt_period(), RUNTIME_INF
);
8374 #endif /* CONFIG_USER_SCHED */
8375 #endif /* CONFIG_RT_GROUP_SCHED */
8377 #ifdef CONFIG_GROUP_SCHED
8378 list_add(&init_task_group
.list
, &task_groups
);
8379 INIT_LIST_HEAD(&init_task_group
.children
);
8381 #ifdef CONFIG_USER_SCHED
8382 INIT_LIST_HEAD(&root_task_group
.children
);
8383 init_task_group
.parent
= &root_task_group
;
8384 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
8385 #endif /* CONFIG_USER_SCHED */
8386 #endif /* CONFIG_GROUP_SCHED */
8388 for_each_possible_cpu(i
) {
8392 spin_lock_init(&rq
->lock
);
8394 init_cfs_rq(&rq
->cfs
, rq
);
8395 init_rt_rq(&rq
->rt
, rq
);
8396 #ifdef CONFIG_FAIR_GROUP_SCHED
8397 init_task_group
.shares
= init_task_group_load
;
8398 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8399 #ifdef CONFIG_CGROUP_SCHED
8401 * How much cpu bandwidth does init_task_group get?
8403 * In case of task-groups formed thr' the cgroup filesystem, it
8404 * gets 100% of the cpu resources in the system. This overall
8405 * system cpu resource is divided among the tasks of
8406 * init_task_group and its child task-groups in a fair manner,
8407 * based on each entity's (task or task-group's) weight
8408 * (se->load.weight).
8410 * In other words, if init_task_group has 10 tasks of weight
8411 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8412 * then A0's share of the cpu resource is:
8414 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8416 * We achieve this by letting init_task_group's tasks sit
8417 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8419 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
8420 #elif defined CONFIG_USER_SCHED
8421 root_task_group
.shares
= NICE_0_LOAD
;
8422 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
8424 * In case of task-groups formed thr' the user id of tasks,
8425 * init_task_group represents tasks belonging to root user.
8426 * Hence it forms a sibling of all subsequent groups formed.
8427 * In this case, init_task_group gets only a fraction of overall
8428 * system cpu resource, based on the weight assigned to root
8429 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8430 * by letting tasks of init_task_group sit in a separate cfs_rq
8431 * (init_cfs_rq) and having one entity represent this group of
8432 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8434 init_tg_cfs_entry(&init_task_group
,
8435 &per_cpu(init_cfs_rq
, i
),
8436 &per_cpu(init_sched_entity
, i
), i
, 1,
8437 root_task_group
.se
[i
]);
8440 #endif /* CONFIG_FAIR_GROUP_SCHED */
8442 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8443 #ifdef CONFIG_RT_GROUP_SCHED
8444 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
8445 #ifdef CONFIG_CGROUP_SCHED
8446 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
8447 #elif defined CONFIG_USER_SCHED
8448 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
8449 init_tg_rt_entry(&init_task_group
,
8450 &per_cpu(init_rt_rq
, i
),
8451 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
8452 root_task_group
.rt_se
[i
]);
8456 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
8457 rq
->cpu_load
[j
] = 0;
8461 rq
->active_balance
= 0;
8462 rq
->next_balance
= jiffies
;
8466 rq
->migration_thread
= NULL
;
8467 INIT_LIST_HEAD(&rq
->migration_queue
);
8468 rq_attach_root(rq
, &def_root_domain
);
8471 atomic_set(&rq
->nr_iowait
, 0);
8474 set_load_weight(&init_task
);
8476 #ifdef CONFIG_PREEMPT_NOTIFIERS
8477 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
8481 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8484 #ifdef CONFIG_RT_MUTEXES
8485 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
8489 * The boot idle thread does lazy MMU switching as well:
8491 atomic_inc(&init_mm
.mm_count
);
8492 enter_lazy_tlb(&init_mm
, current
);
8495 * Make us the idle thread. Technically, schedule() should not be
8496 * called from this thread, however somewhere below it might be,
8497 * but because we are the idle thread, we just pick up running again
8498 * when this runqueue becomes "idle".
8500 init_idle(current
, smp_processor_id());
8502 * During early bootup we pretend to be a normal task:
8504 current
->sched_class
= &fair_sched_class
;
8506 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8507 alloc_bootmem_cpumask_var(&nohz_cpu_mask
);
8510 alloc_bootmem_cpumask_var(&nohz
.cpu_mask
);
8512 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
8515 scheduler_running
= 1;
8518 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8519 void __might_sleep(char *file
, int line
)
8522 static unsigned long prev_jiffy
; /* ratelimiting */
8524 if ((!in_atomic() && !irqs_disabled()) ||
8525 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
8527 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8529 prev_jiffy
= jiffies
;
8532 "BUG: sleeping function called from invalid context at %s:%d\n",
8535 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8536 in_atomic(), irqs_disabled(),
8537 current
->pid
, current
->comm
);
8539 debug_show_held_locks(current
);
8540 if (irqs_disabled())
8541 print_irqtrace_events(current
);
8545 EXPORT_SYMBOL(__might_sleep
);
8548 #ifdef CONFIG_MAGIC_SYSRQ
8549 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8553 update_rq_clock(rq
);
8554 on_rq
= p
->se
.on_rq
;
8556 deactivate_task(rq
, p
, 0);
8557 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8559 activate_task(rq
, p
, 0);
8560 resched_task(rq
->curr
);
8564 void normalize_rt_tasks(void)
8566 struct task_struct
*g
, *p
;
8567 unsigned long flags
;
8570 read_lock_irqsave(&tasklist_lock
, flags
);
8571 do_each_thread(g
, p
) {
8573 * Only normalize user tasks:
8578 p
->se
.exec_start
= 0;
8579 #ifdef CONFIG_SCHEDSTATS
8580 p
->se
.wait_start
= 0;
8581 p
->se
.sleep_start
= 0;
8582 p
->se
.block_start
= 0;
8587 * Renice negative nice level userspace
8590 if (TASK_NICE(p
) < 0 && p
->mm
)
8591 set_user_nice(p
, 0);
8595 spin_lock(&p
->pi_lock
);
8596 rq
= __task_rq_lock(p
);
8598 normalize_task(rq
, p
);
8600 __task_rq_unlock(rq
);
8601 spin_unlock(&p
->pi_lock
);
8602 } while_each_thread(g
, p
);
8604 read_unlock_irqrestore(&tasklist_lock
, flags
);
8607 #endif /* CONFIG_MAGIC_SYSRQ */
8611 * These functions are only useful for the IA64 MCA handling.
8613 * They can only be called when the whole system has been
8614 * stopped - every CPU needs to be quiescent, and no scheduling
8615 * activity can take place. Using them for anything else would
8616 * be a serious bug, and as a result, they aren't even visible
8617 * under any other configuration.
8621 * curr_task - return the current task for a given cpu.
8622 * @cpu: the processor in question.
8624 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8626 struct task_struct
*curr_task(int cpu
)
8628 return cpu_curr(cpu
);
8632 * set_curr_task - set the current task for a given cpu.
8633 * @cpu: the processor in question.
8634 * @p: the task pointer to set.
8636 * Description: This function must only be used when non-maskable interrupts
8637 * are serviced on a separate stack. It allows the architecture to switch the
8638 * notion of the current task on a cpu in a non-blocking manner. This function
8639 * must be called with all CPU's synchronized, and interrupts disabled, the
8640 * and caller must save the original value of the current task (see
8641 * curr_task() above) and restore that value before reenabling interrupts and
8642 * re-starting the system.
8644 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8646 void set_curr_task(int cpu
, struct task_struct
*p
)
8653 #ifdef CONFIG_FAIR_GROUP_SCHED
8654 static void free_fair_sched_group(struct task_group
*tg
)
8658 for_each_possible_cpu(i
) {
8660 kfree(tg
->cfs_rq
[i
]);
8670 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8672 struct cfs_rq
*cfs_rq
;
8673 struct sched_entity
*se
;
8677 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8680 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8684 tg
->shares
= NICE_0_LOAD
;
8686 for_each_possible_cpu(i
) {
8689 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
8690 GFP_KERNEL
, cpu_to_node(i
));
8694 se
= kzalloc_node(sizeof(struct sched_entity
),
8695 GFP_KERNEL
, cpu_to_node(i
));
8699 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
8708 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8710 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8711 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8714 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8716 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8718 #else /* !CONFG_FAIR_GROUP_SCHED */
8719 static inline void free_fair_sched_group(struct task_group
*tg
)
8724 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8729 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8733 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8736 #endif /* CONFIG_FAIR_GROUP_SCHED */
8738 #ifdef CONFIG_RT_GROUP_SCHED
8739 static void free_rt_sched_group(struct task_group
*tg
)
8743 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8745 for_each_possible_cpu(i
) {
8747 kfree(tg
->rt_rq
[i
]);
8749 kfree(tg
->rt_se
[i
]);
8757 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8759 struct rt_rq
*rt_rq
;
8760 struct sched_rt_entity
*rt_se
;
8764 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8767 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8771 init_rt_bandwidth(&tg
->rt_bandwidth
,
8772 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8774 for_each_possible_cpu(i
) {
8777 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
8778 GFP_KERNEL
, cpu_to_node(i
));
8782 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
8783 GFP_KERNEL
, cpu_to_node(i
));
8787 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
8796 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8798 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8799 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8802 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8804 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8806 #else /* !CONFIG_RT_GROUP_SCHED */
8807 static inline void free_rt_sched_group(struct task_group
*tg
)
8812 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8817 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8821 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8824 #endif /* CONFIG_RT_GROUP_SCHED */
8826 #ifdef CONFIG_GROUP_SCHED
8827 static void free_sched_group(struct task_group
*tg
)
8829 free_fair_sched_group(tg
);
8830 free_rt_sched_group(tg
);
8834 /* allocate runqueue etc for a new task group */
8835 struct task_group
*sched_create_group(struct task_group
*parent
)
8837 struct task_group
*tg
;
8838 unsigned long flags
;
8841 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8843 return ERR_PTR(-ENOMEM
);
8845 if (!alloc_fair_sched_group(tg
, parent
))
8848 if (!alloc_rt_sched_group(tg
, parent
))
8851 spin_lock_irqsave(&task_group_lock
, flags
);
8852 for_each_possible_cpu(i
) {
8853 register_fair_sched_group(tg
, i
);
8854 register_rt_sched_group(tg
, i
);
8856 list_add_rcu(&tg
->list
, &task_groups
);
8858 WARN_ON(!parent
); /* root should already exist */
8860 tg
->parent
= parent
;
8861 INIT_LIST_HEAD(&tg
->children
);
8862 list_add_rcu(&tg
->siblings
, &parent
->children
);
8863 spin_unlock_irqrestore(&task_group_lock
, flags
);
8868 free_sched_group(tg
);
8869 return ERR_PTR(-ENOMEM
);
8872 /* rcu callback to free various structures associated with a task group */
8873 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8875 /* now it should be safe to free those cfs_rqs */
8876 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8879 /* Destroy runqueue etc associated with a task group */
8880 void sched_destroy_group(struct task_group
*tg
)
8882 unsigned long flags
;
8885 spin_lock_irqsave(&task_group_lock
, flags
);
8886 for_each_possible_cpu(i
) {
8887 unregister_fair_sched_group(tg
, i
);
8888 unregister_rt_sched_group(tg
, i
);
8890 list_del_rcu(&tg
->list
);
8891 list_del_rcu(&tg
->siblings
);
8892 spin_unlock_irqrestore(&task_group_lock
, flags
);
8894 /* wait for possible concurrent references to cfs_rqs complete */
8895 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8898 /* change task's runqueue when it moves between groups.
8899 * The caller of this function should have put the task in its new group
8900 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8901 * reflect its new group.
8903 void sched_move_task(struct task_struct
*tsk
)
8906 unsigned long flags
;
8909 rq
= task_rq_lock(tsk
, &flags
);
8911 update_rq_clock(rq
);
8913 running
= task_current(rq
, tsk
);
8914 on_rq
= tsk
->se
.on_rq
;
8917 dequeue_task(rq
, tsk
, 0);
8918 if (unlikely(running
))
8919 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8921 set_task_rq(tsk
, task_cpu(tsk
));
8923 #ifdef CONFIG_FAIR_GROUP_SCHED
8924 if (tsk
->sched_class
->moved_group
)
8925 tsk
->sched_class
->moved_group(tsk
);
8928 if (unlikely(running
))
8929 tsk
->sched_class
->set_curr_task(rq
);
8931 enqueue_task(rq
, tsk
, 0);
8933 task_rq_unlock(rq
, &flags
);
8935 #endif /* CONFIG_GROUP_SCHED */
8937 #ifdef CONFIG_FAIR_GROUP_SCHED
8938 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8940 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8945 dequeue_entity(cfs_rq
, se
, 0);
8947 se
->load
.weight
= shares
;
8948 se
->load
.inv_weight
= 0;
8951 enqueue_entity(cfs_rq
, se
, 0);
8954 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8956 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8957 struct rq
*rq
= cfs_rq
->rq
;
8958 unsigned long flags
;
8960 spin_lock_irqsave(&rq
->lock
, flags
);
8961 __set_se_shares(se
, shares
);
8962 spin_unlock_irqrestore(&rq
->lock
, flags
);
8965 static DEFINE_MUTEX(shares_mutex
);
8967 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8970 unsigned long flags
;
8973 * We can't change the weight of the root cgroup.
8978 if (shares
< MIN_SHARES
)
8979 shares
= MIN_SHARES
;
8980 else if (shares
> MAX_SHARES
)
8981 shares
= MAX_SHARES
;
8983 mutex_lock(&shares_mutex
);
8984 if (tg
->shares
== shares
)
8987 spin_lock_irqsave(&task_group_lock
, flags
);
8988 for_each_possible_cpu(i
)
8989 unregister_fair_sched_group(tg
, i
);
8990 list_del_rcu(&tg
->siblings
);
8991 spin_unlock_irqrestore(&task_group_lock
, flags
);
8993 /* wait for any ongoing reference to this group to finish */
8994 synchronize_sched();
8997 * Now we are free to modify the group's share on each cpu
8998 * w/o tripping rebalance_share or load_balance_fair.
9000 tg
->shares
= shares
;
9001 for_each_possible_cpu(i
) {
9005 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
9006 set_se_shares(tg
->se
[i
], shares
);
9010 * Enable load balance activity on this group, by inserting it back on
9011 * each cpu's rq->leaf_cfs_rq_list.
9013 spin_lock_irqsave(&task_group_lock
, flags
);
9014 for_each_possible_cpu(i
)
9015 register_fair_sched_group(tg
, i
);
9016 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
9017 spin_unlock_irqrestore(&task_group_lock
, flags
);
9019 mutex_unlock(&shares_mutex
);
9023 unsigned long sched_group_shares(struct task_group
*tg
)
9029 #ifdef CONFIG_RT_GROUP_SCHED
9031 * Ensure that the real time constraints are schedulable.
9033 static DEFINE_MUTEX(rt_constraints_mutex
);
9035 static unsigned long to_ratio(u64 period
, u64 runtime
)
9037 if (runtime
== RUNTIME_INF
)
9040 return div64_u64(runtime
<< 20, period
);
9043 /* Must be called with tasklist_lock held */
9044 static inline int tg_has_rt_tasks(struct task_group
*tg
)
9046 struct task_struct
*g
, *p
;
9048 do_each_thread(g
, p
) {
9049 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
9051 } while_each_thread(g
, p
);
9056 struct rt_schedulable_data
{
9057 struct task_group
*tg
;
9062 static int tg_schedulable(struct task_group
*tg
, void *data
)
9064 struct rt_schedulable_data
*d
= data
;
9065 struct task_group
*child
;
9066 unsigned long total
, sum
= 0;
9067 u64 period
, runtime
;
9069 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9070 runtime
= tg
->rt_bandwidth
.rt_runtime
;
9073 period
= d
->rt_period
;
9074 runtime
= d
->rt_runtime
;
9078 * Cannot have more runtime than the period.
9080 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9084 * Ensure we don't starve existing RT tasks.
9086 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
9089 total
= to_ratio(period
, runtime
);
9092 * Nobody can have more than the global setting allows.
9094 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
9098 * The sum of our children's runtime should not exceed our own.
9100 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
9101 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
9102 runtime
= child
->rt_bandwidth
.rt_runtime
;
9104 if (child
== d
->tg
) {
9105 period
= d
->rt_period
;
9106 runtime
= d
->rt_runtime
;
9109 sum
+= to_ratio(period
, runtime
);
9118 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
9120 struct rt_schedulable_data data
= {
9122 .rt_period
= period
,
9123 .rt_runtime
= runtime
,
9126 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
9129 static int tg_set_bandwidth(struct task_group
*tg
,
9130 u64 rt_period
, u64 rt_runtime
)
9134 mutex_lock(&rt_constraints_mutex
);
9135 read_lock(&tasklist_lock
);
9136 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
9140 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
9141 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
9142 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
9144 for_each_possible_cpu(i
) {
9145 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
9147 spin_lock(&rt_rq
->rt_runtime_lock
);
9148 rt_rq
->rt_runtime
= rt_runtime
;
9149 spin_unlock(&rt_rq
->rt_runtime_lock
);
9151 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
9153 read_unlock(&tasklist_lock
);
9154 mutex_unlock(&rt_constraints_mutex
);
9159 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
9161 u64 rt_runtime
, rt_period
;
9163 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9164 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
9165 if (rt_runtime_us
< 0)
9166 rt_runtime
= RUNTIME_INF
;
9168 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9171 long sched_group_rt_runtime(struct task_group
*tg
)
9175 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
9178 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
9179 do_div(rt_runtime_us
, NSEC_PER_USEC
);
9180 return rt_runtime_us
;
9183 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
9185 u64 rt_runtime
, rt_period
;
9187 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
9188 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9193 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9196 long sched_group_rt_period(struct task_group
*tg
)
9200 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9201 do_div(rt_period_us
, NSEC_PER_USEC
);
9202 return rt_period_us
;
9205 static int sched_rt_global_constraints(void)
9207 u64 runtime
, period
;
9210 if (sysctl_sched_rt_period
<= 0)
9213 runtime
= global_rt_runtime();
9214 period
= global_rt_period();
9217 * Sanity check on the sysctl variables.
9219 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9222 mutex_lock(&rt_constraints_mutex
);
9223 read_lock(&tasklist_lock
);
9224 ret
= __rt_schedulable(NULL
, 0, 0);
9225 read_unlock(&tasklist_lock
);
9226 mutex_unlock(&rt_constraints_mutex
);
9230 #else /* !CONFIG_RT_GROUP_SCHED */
9231 static int sched_rt_global_constraints(void)
9233 unsigned long flags
;
9236 if (sysctl_sched_rt_period
<= 0)
9239 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9240 for_each_possible_cpu(i
) {
9241 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
9243 spin_lock(&rt_rq
->rt_runtime_lock
);
9244 rt_rq
->rt_runtime
= global_rt_runtime();
9245 spin_unlock(&rt_rq
->rt_runtime_lock
);
9247 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9251 #endif /* CONFIG_RT_GROUP_SCHED */
9253 int sched_rt_handler(struct ctl_table
*table
, int write
,
9254 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
9258 int old_period
, old_runtime
;
9259 static DEFINE_MUTEX(mutex
);
9262 old_period
= sysctl_sched_rt_period
;
9263 old_runtime
= sysctl_sched_rt_runtime
;
9265 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
9267 if (!ret
&& write
) {
9268 ret
= sched_rt_global_constraints();
9270 sysctl_sched_rt_period
= old_period
;
9271 sysctl_sched_rt_runtime
= old_runtime
;
9273 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
9274 def_rt_bandwidth
.rt_period
=
9275 ns_to_ktime(global_rt_period());
9278 mutex_unlock(&mutex
);
9283 #ifdef CONFIG_CGROUP_SCHED
9285 /* return corresponding task_group object of a cgroup */
9286 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
9288 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
9289 struct task_group
, css
);
9292 static struct cgroup_subsys_state
*
9293 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9295 struct task_group
*tg
, *parent
;
9297 if (!cgrp
->parent
) {
9298 /* This is early initialization for the top cgroup */
9299 return &init_task_group
.css
;
9302 parent
= cgroup_tg(cgrp
->parent
);
9303 tg
= sched_create_group(parent
);
9305 return ERR_PTR(-ENOMEM
);
9311 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9313 struct task_group
*tg
= cgroup_tg(cgrp
);
9315 sched_destroy_group(tg
);
9319 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9320 struct task_struct
*tsk
)
9322 #ifdef CONFIG_RT_GROUP_SCHED
9323 /* Don't accept realtime tasks when there is no way for them to run */
9324 if (rt_task(tsk
) && cgroup_tg(cgrp
)->rt_bandwidth
.rt_runtime
== 0)
9327 /* We don't support RT-tasks being in separate groups */
9328 if (tsk
->sched_class
!= &fair_sched_class
)
9336 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9337 struct cgroup
*old_cont
, struct task_struct
*tsk
)
9339 sched_move_task(tsk
);
9342 #ifdef CONFIG_FAIR_GROUP_SCHED
9343 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9346 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
9349 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9351 struct task_group
*tg
= cgroup_tg(cgrp
);
9353 return (u64
) tg
->shares
;
9355 #endif /* CONFIG_FAIR_GROUP_SCHED */
9357 #ifdef CONFIG_RT_GROUP_SCHED
9358 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9361 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9364 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9366 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9369 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9372 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9375 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9377 return sched_group_rt_period(cgroup_tg(cgrp
));
9379 #endif /* CONFIG_RT_GROUP_SCHED */
9381 static struct cftype cpu_files
[] = {
9382 #ifdef CONFIG_FAIR_GROUP_SCHED
9385 .read_u64
= cpu_shares_read_u64
,
9386 .write_u64
= cpu_shares_write_u64
,
9389 #ifdef CONFIG_RT_GROUP_SCHED
9391 .name
= "rt_runtime_us",
9392 .read_s64
= cpu_rt_runtime_read
,
9393 .write_s64
= cpu_rt_runtime_write
,
9396 .name
= "rt_period_us",
9397 .read_u64
= cpu_rt_period_read_uint
,
9398 .write_u64
= cpu_rt_period_write_uint
,
9403 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
9405 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
9408 struct cgroup_subsys cpu_cgroup_subsys
= {
9410 .create
= cpu_cgroup_create
,
9411 .destroy
= cpu_cgroup_destroy
,
9412 .can_attach
= cpu_cgroup_can_attach
,
9413 .attach
= cpu_cgroup_attach
,
9414 .populate
= cpu_cgroup_populate
,
9415 .subsys_id
= cpu_cgroup_subsys_id
,
9419 #endif /* CONFIG_CGROUP_SCHED */
9421 #ifdef CONFIG_CGROUP_CPUACCT
9424 * CPU accounting code for task groups.
9426 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9427 * (balbir@in.ibm.com).
9430 /* track cpu usage of a group of tasks and its child groups */
9432 struct cgroup_subsys_state css
;
9433 /* cpuusage holds pointer to a u64-type object on every cpu */
9435 struct cpuacct
*parent
;
9438 struct cgroup_subsys cpuacct_subsys
;
9440 /* return cpu accounting group corresponding to this container */
9441 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
9443 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
9444 struct cpuacct
, css
);
9447 /* return cpu accounting group to which this task belongs */
9448 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
9450 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
9451 struct cpuacct
, css
);
9454 /* create a new cpu accounting group */
9455 static struct cgroup_subsys_state
*cpuacct_create(
9456 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9458 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
9461 return ERR_PTR(-ENOMEM
);
9463 ca
->cpuusage
= alloc_percpu(u64
);
9464 if (!ca
->cpuusage
) {
9466 return ERR_PTR(-ENOMEM
);
9470 ca
->parent
= cgroup_ca(cgrp
->parent
);
9475 /* destroy an existing cpu accounting group */
9477 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9479 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9481 free_percpu(ca
->cpuusage
);
9485 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
9487 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, cpu
);
9490 #ifndef CONFIG_64BIT
9492 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9494 spin_lock_irq(&cpu_rq(cpu
)->lock
);
9496 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9504 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
9506 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, cpu
);
9508 #ifndef CONFIG_64BIT
9510 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9512 spin_lock_irq(&cpu_rq(cpu
)->lock
);
9514 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9520 /* return total cpu usage (in nanoseconds) of a group */
9521 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9523 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9524 u64 totalcpuusage
= 0;
9527 for_each_present_cpu(i
)
9528 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
9530 return totalcpuusage
;
9533 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9536 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9545 for_each_present_cpu(i
)
9546 cpuacct_cpuusage_write(ca
, i
, 0);
9552 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
9555 struct cpuacct
*ca
= cgroup_ca(cgroup
);
9559 for_each_present_cpu(i
) {
9560 percpu
= cpuacct_cpuusage_read(ca
, i
);
9561 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
9563 seq_printf(m
, "\n");
9567 static struct cftype files
[] = {
9570 .read_u64
= cpuusage_read
,
9571 .write_u64
= cpuusage_write
,
9574 .name
= "usage_percpu",
9575 .read_seq_string
= cpuacct_percpu_seq_read
,
9580 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9582 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9586 * charge this task's execution time to its accounting group.
9588 * called with rq->lock held.
9590 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9595 if (!cpuacct_subsys
.active
)
9598 cpu
= task_cpu(tsk
);
9601 for (; ca
; ca
= ca
->parent
) {
9602 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, cpu
);
9603 *cpuusage
+= cputime
;
9607 struct cgroup_subsys cpuacct_subsys
= {
9609 .create
= cpuacct_create
,
9610 .destroy
= cpuacct_destroy
,
9611 .populate
= cpuacct_populate
,
9612 .subsys_id
= cpuacct_subsys_id
,
9614 #endif /* CONFIG_CGROUP_CPUACCT */