blob: 0d7544c3dba7d539a6032a9d48f11728d522f0ce [file] [log] [blame]
Peter Zijlstra029632f2011-10-25 10:00:11 +02001
2#include <linux/sched.h>
Clark Williamscf4aebc22013-02-07 09:46:59 -06003#include <linux/sched/sysctl.h>
Clark Williams8bd75c72013-02-07 09:47:07 -06004#include <linux/sched/rt.h>
Peter Zijlstra029632f2011-10-25 10:00:11 +02005#include <linux/mutex.h>
6#include <linux/spinlock.h>
7#include <linux/stop_machine.h>
Frederic Weisbecker9f3660c2013-04-20 14:35:09 +02008#include <linux/tick.h>
Peter Zijlstra029632f2011-10-25 10:00:11 +02009
Peter Zijlstra391e43d2011-11-15 17:14:39 +010010#include "cpupri.h"
Li Zefan60fed782013-03-29 14:36:43 +080011#include "cpuacct.h"
Peter Zijlstra029632f2011-10-25 10:00:11 +020012
Paul Gortmaker45ceebf2013-04-19 15:10:49 -040013struct rq;
14
Peter Zijlstra029632f2011-10-25 10:00:11 +020015extern __read_mostly int scheduler_running;
16
Paul Gortmaker45ceebf2013-04-19 15:10:49 -040017extern unsigned long calc_load_update;
18extern atomic_long_t calc_load_tasks;
19
20extern long calc_load_fold_active(struct rq *this_rq);
21extern void update_cpu_load_active(struct rq *this_rq);
22
Peter Zijlstra029632f2011-10-25 10:00:11 +020023/*
24 * Convert user-nice values [ -20 ... 0 ... 19 ]
25 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
26 * and back.
27 */
28#define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
29#define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
30#define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
31
32/*
33 * 'User priority' is the nice value converted to something we
34 * can work with better when scaling various scheduler parameters,
35 * it's a [ 0 ... 39 ] range.
36 */
37#define USER_PRIO(p) ((p)-MAX_RT_PRIO)
38#define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
39#define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
40
41/*
42 * Helpers for converting nanosecond timing to jiffy resolution
43 */
44#define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
45
Li Zefancc1f4b12013-03-05 16:06:09 +080046/*
47 * Increase resolution of nice-level calculations for 64-bit architectures.
48 * The extra resolution improves shares distribution and load balancing of
49 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup
50 * hierarchies, especially on larger systems. This is not a user-visible change
51 * and does not change the user-interface for setting shares/weights.
52 *
53 * We increase resolution only if we have enough bits to allow this increased
54 * resolution (i.e. BITS_PER_LONG > 32). The costs for increasing resolution
55 * when BITS_PER_LONG <= 32 are pretty high and the returns do not justify the
56 * increased costs.
57 */
58#if 0 /* BITS_PER_LONG > 32 -- currently broken: it increases power usage under light load */
59# define SCHED_LOAD_RESOLUTION 10
60# define scale_load(w) ((w) << SCHED_LOAD_RESOLUTION)
61# define scale_load_down(w) ((w) >> SCHED_LOAD_RESOLUTION)
62#else
63# define SCHED_LOAD_RESOLUTION 0
64# define scale_load(w) (w)
65# define scale_load_down(w) (w)
66#endif
67
68#define SCHED_LOAD_SHIFT (10 + SCHED_LOAD_RESOLUTION)
69#define SCHED_LOAD_SCALE (1L << SCHED_LOAD_SHIFT)
70
Peter Zijlstra029632f2011-10-25 10:00:11 +020071#define NICE_0_LOAD SCHED_LOAD_SCALE
72#define NICE_0_SHIFT SCHED_LOAD_SHIFT
73
74/*
75 * These are the 'tuning knobs' of the scheduler:
Peter Zijlstra029632f2011-10-25 10:00:11 +020076 */
Peter Zijlstra029632f2011-10-25 10:00:11 +020077
78/*
79 * single value that denotes runtime == period, ie unlimited time.
80 */
81#define RUNTIME_INF ((u64)~0ULL)
82
83static inline int rt_policy(int policy)
84{
85 if (policy == SCHED_FIFO || policy == SCHED_RR)
86 return 1;
87 return 0;
88}
89
90static inline int task_has_rt_policy(struct task_struct *p)
91{
92 return rt_policy(p->policy);
93}
94
95/*
96 * This is the priority-queue data structure of the RT scheduling class:
97 */
98struct rt_prio_array {
99 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
100 struct list_head queue[MAX_RT_PRIO];
101};
102
103struct rt_bandwidth {
104 /* nests inside the rq lock: */
105 raw_spinlock_t rt_runtime_lock;
106 ktime_t rt_period;
107 u64 rt_runtime;
108 struct hrtimer rt_period_timer;
109};
110
111extern struct mutex sched_domains_mutex;
112
113#ifdef CONFIG_CGROUP_SCHED
114
115#include <linux/cgroup.h>
116
117struct cfs_rq;
118struct rt_rq;
119
Mike Galbraith35cf4e52012-08-07 05:00:13 +0200120extern struct list_head task_groups;
Peter Zijlstra029632f2011-10-25 10:00:11 +0200121
122struct cfs_bandwidth {
123#ifdef CONFIG_CFS_BANDWIDTH
124 raw_spinlock_t lock;
125 ktime_t period;
126 u64 quota, runtime;
127 s64 hierarchal_quota;
128 u64 runtime_expires;
129
130 int idle, timer_active;
131 struct hrtimer period_timer, slack_timer;
132 struct list_head throttled_cfs_rq;
133
134 /* statistics */
135 int nr_periods, nr_throttled;
136 u64 throttled_time;
137#endif
138};
139
140/* task group related information */
141struct task_group {
142 struct cgroup_subsys_state css;
143
144#ifdef CONFIG_FAIR_GROUP_SCHED
145 /* schedulable entities of this group on each cpu */
146 struct sched_entity **se;
147 /* runqueue "owned" by this group on each cpu */
148 struct cfs_rq **cfs_rq;
149 unsigned long shares;
150
Alex Shifa6bdde2013-06-20 10:18:46 +0800151#ifdef CONFIG_SMP
Alex Shibf5b9862013-06-20 10:18:54 +0800152 atomic_long_t load_avg;
Paul Turnerbb17f652012-10-04 13:18:31 +0200153 atomic_t runnable_avg;
Peter Zijlstra029632f2011-10-25 10:00:11 +0200154#endif
Alex Shifa6bdde2013-06-20 10:18:46 +0800155#endif
Peter Zijlstra029632f2011-10-25 10:00:11 +0200156
157#ifdef CONFIG_RT_GROUP_SCHED
158 struct sched_rt_entity **rt_se;
159 struct rt_rq **rt_rq;
160
161 struct rt_bandwidth rt_bandwidth;
162#endif
163
164 struct rcu_head rcu;
165 struct list_head list;
166
167 struct task_group *parent;
168 struct list_head siblings;
169 struct list_head children;
170
171#ifdef CONFIG_SCHED_AUTOGROUP
172 struct autogroup *autogroup;
173#endif
174
175 struct cfs_bandwidth cfs_bandwidth;
176};
177
178#ifdef CONFIG_FAIR_GROUP_SCHED
179#define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
180
181/*
182 * A weight of 0 or 1 can cause arithmetics problems.
183 * A weight of a cfs_rq is the sum of weights of which entities
184 * are queued on this cfs_rq, so a weight of a entity should not be
185 * too large, so as the shares value of a task group.
186 * (The default weight is 1024 - so there's no practical
187 * limitation from this.)
188 */
189#define MIN_SHARES (1UL << 1)
190#define MAX_SHARES (1UL << 18)
191#endif
192
Peter Zijlstra029632f2011-10-25 10:00:11 +0200193typedef int (*tg_visitor)(struct task_group *, void *);
194
195extern int walk_tg_tree_from(struct task_group *from,
196 tg_visitor down, tg_visitor up, void *data);
197
198/*
199 * Iterate the full tree, calling @down when first entering a node and @up when
200 * leaving it for the final time.
201 *
202 * Caller must hold rcu_lock or sufficient equivalent.
203 */
204static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
205{
206 return walk_tg_tree_from(&root_task_group, down, up, data);
207}
208
209extern int tg_nop(struct task_group *tg, void *data);
210
211extern void free_fair_sched_group(struct task_group *tg);
212extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent);
213extern void unregister_fair_sched_group(struct task_group *tg, int cpu);
214extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
215 struct sched_entity *se, int cpu,
216 struct sched_entity *parent);
217extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
218extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
219
220extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b);
221extern void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
222extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq);
223
224extern void free_rt_sched_group(struct task_group *tg);
225extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent);
226extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
227 struct sched_rt_entity *rt_se, int cpu,
228 struct sched_rt_entity *parent);
229
Li Zefan25cc7da2013-03-05 16:07:33 +0800230extern struct task_group *sched_create_group(struct task_group *parent);
231extern void sched_online_group(struct task_group *tg,
232 struct task_group *parent);
233extern void sched_destroy_group(struct task_group *tg);
234extern void sched_offline_group(struct task_group *tg);
235
236extern void sched_move_task(struct task_struct *tsk);
237
238#ifdef CONFIG_FAIR_GROUP_SCHED
239extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
240#endif
241
Peter Zijlstra029632f2011-10-25 10:00:11 +0200242#else /* CONFIG_CGROUP_SCHED */
243
244struct cfs_bandwidth { };
245
246#endif /* CONFIG_CGROUP_SCHED */
247
248/* CFS-related fields in a runqueue */
249struct cfs_rq {
250 struct load_weight load;
Peter Zijlstrac82513e2012-04-26 13:12:27 +0200251 unsigned int nr_running, h_nr_running;
Peter Zijlstra029632f2011-10-25 10:00:11 +0200252
253 u64 exec_clock;
254 u64 min_vruntime;
255#ifndef CONFIG_64BIT
256 u64 min_vruntime_copy;
257#endif
258
259 struct rb_root tasks_timeline;
260 struct rb_node *rb_leftmost;
261
Peter Zijlstra029632f2011-10-25 10:00:11 +0200262 /*
263 * 'curr' points to currently running entity on this cfs_rq.
264 * It is set to NULL otherwise (i.e when none are currently running).
265 */
266 struct sched_entity *curr, *next, *last, *skip;
267
268#ifdef CONFIG_SCHED_DEBUG
269 unsigned int nr_spread_over;
270#endif
271
Paul Turner2dac7542012-10-04 13:18:30 +0200272#ifdef CONFIG_SMP
273 /*
274 * CFS Load tracking
275 * Under CFS, load is tracked on a per-entity basis and aggregated up.
276 * This allows for the description of both thread and group usage (in
277 * the FAIR_GROUP_SCHED case).
278 */
Alex Shi72a4cf22013-06-20 10:18:53 +0800279 unsigned long runnable_load_avg, blocked_load_avg;
Alex Shi25099402013-06-20 10:18:55 +0800280 atomic64_t decay_counter;
Paul Turner9ee474f2012-10-04 13:18:30 +0200281 u64 last_decay;
Alex Shi25099402013-06-20 10:18:55 +0800282 atomic_long_t removed_load;
Alex Shi141965c2013-06-26 13:05:39 +0800283
Paul Turnerc566e8e2012-10-04 13:18:30 +0200284#ifdef CONFIG_FAIR_GROUP_SCHED
Alex Shi141965c2013-06-26 13:05:39 +0800285 /* Required to track per-cpu representation of a task_group */
Paul Turnerbb17f652012-10-04 13:18:31 +0200286 u32 tg_runnable_contrib;
Alex Shibf5b9862013-06-20 10:18:54 +0800287 unsigned long tg_load_contrib;
Paul Turner82958362012-10-04 13:18:31 +0200288
289 /*
290 * h_load = weight * f(tg)
291 *
292 * Where f(tg) is the recursive weight fraction assigned to
293 * this group.
294 */
295 unsigned long h_load;
Vladimir Davydov68520792013-07-15 17:49:19 +0400296 u64 last_h_load_update;
297 struct sched_entity *h_load_next;
298#endif /* CONFIG_FAIR_GROUP_SCHED */
Paul Turner82958362012-10-04 13:18:31 +0200299#endif /* CONFIG_SMP */
300
Peter Zijlstra029632f2011-10-25 10:00:11 +0200301#ifdef CONFIG_FAIR_GROUP_SCHED
302 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
303
304 /*
305 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
306 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
307 * (like users, containers etc.)
308 *
309 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
310 * list is used during load balance.
311 */
312 int on_list;
313 struct list_head leaf_cfs_rq_list;
314 struct task_group *tg; /* group that "owns" this runqueue */
315
Peter Zijlstra029632f2011-10-25 10:00:11 +0200316#ifdef CONFIG_CFS_BANDWIDTH
317 int runtime_enabled;
318 u64 runtime_expires;
319 s64 runtime_remaining;
320
Paul Turnerf1b17282012-10-04 13:18:31 +0200321 u64 throttled_clock, throttled_clock_task;
322 u64 throttled_clock_task_time;
Peter Zijlstra029632f2011-10-25 10:00:11 +0200323 int throttled, throttle_count;
324 struct list_head throttled_list;
325#endif /* CONFIG_CFS_BANDWIDTH */
326#endif /* CONFIG_FAIR_GROUP_SCHED */
327};
328
329static inline int rt_bandwidth_enabled(void)
330{
331 return sysctl_sched_rt_runtime >= 0;
332}
333
334/* Real-Time classes' related field in a runqueue: */
335struct rt_rq {
336 struct rt_prio_array active;
Peter Zijlstrac82513e2012-04-26 13:12:27 +0200337 unsigned int rt_nr_running;
Peter Zijlstra029632f2011-10-25 10:00:11 +0200338#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
339 struct {
340 int curr; /* highest queued rt task prio */
341#ifdef CONFIG_SMP
342 int next; /* next highest */
343#endif
344 } highest_prio;
345#endif
346#ifdef CONFIG_SMP
347 unsigned long rt_nr_migratory;
348 unsigned long rt_nr_total;
349 int overloaded;
350 struct plist_head pushable_tasks;
351#endif
352 int rt_throttled;
353 u64 rt_time;
354 u64 rt_runtime;
355 /* Nests inside the rq lock: */
356 raw_spinlock_t rt_runtime_lock;
357
358#ifdef CONFIG_RT_GROUP_SCHED
359 unsigned long rt_nr_boosted;
360
361 struct rq *rq;
Peter Zijlstra029632f2011-10-25 10:00:11 +0200362 struct task_group *tg;
363#endif
364};
365
366#ifdef CONFIG_SMP
367
368/*
369 * We add the notion of a root-domain which will be used to define per-domain
370 * variables. Each exclusive cpuset essentially defines an island domain by
371 * fully partitioning the member cpus from any other cpuset. Whenever a new
372 * exclusive cpuset is created, we also create and attach a new root-domain
373 * object.
374 *
375 */
376struct root_domain {
377 atomic_t refcount;
378 atomic_t rto_count;
379 struct rcu_head rcu;
380 cpumask_var_t span;
381 cpumask_var_t online;
382
383 /*
384 * The "RT overload" flag: it gets set if a CPU has more than
385 * one runnable RT task.
386 */
387 cpumask_var_t rto_mask;
388 struct cpupri cpupri;
389};
390
391extern struct root_domain def_root_domain;
392
393#endif /* CONFIG_SMP */
394
395/*
396 * This is the main, per-CPU runqueue data structure.
397 *
398 * Locking rule: those places that want to lock multiple runqueues
399 * (such as the load balancing or the thread migration code), lock
400 * acquire operations must be ordered by ascending &runqueue.
401 */
402struct rq {
403 /* runqueue lock: */
404 raw_spinlock_t lock;
405
406 /*
407 * nr_running and cpu_load should be in the same cacheline because
408 * remote CPUs use both these fields when doing load calculation.
409 */
Peter Zijlstrac82513e2012-04-26 13:12:27 +0200410 unsigned int nr_running;
Peter Zijlstra029632f2011-10-25 10:00:11 +0200411 #define CPU_LOAD_IDX_MAX 5
412 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
413 unsigned long last_load_update_tick;
Frederic Weisbecker3451d022011-08-10 23:21:01 +0200414#ifdef CONFIG_NO_HZ_COMMON
Peter Zijlstra029632f2011-10-25 10:00:11 +0200415 u64 nohz_stamp;
Suresh Siddha1c792db2011-12-01 17:07:32 -0800416 unsigned long nohz_flags;
Peter Zijlstra029632f2011-10-25 10:00:11 +0200417#endif
Frederic Weisbecker265f22a2013-05-03 03:39:05 +0200418#ifdef CONFIG_NO_HZ_FULL
419 unsigned long last_sched_tick;
420#endif
Peter Zijlstra029632f2011-10-25 10:00:11 +0200421 int skip_clock_update;
422
423 /* capture load from *all* tasks on this cpu: */
424 struct load_weight load;
425 unsigned long nr_load_updates;
426 u64 nr_switches;
427
428 struct cfs_rq cfs;
429 struct rt_rq rt;
430
431#ifdef CONFIG_FAIR_GROUP_SCHED
432 /* list of leaf cfs_rq on this cpu: */
433 struct list_head leaf_cfs_rq_list;
Peter Zijlstraa35b6462012-08-08 21:46:40 +0200434#endif /* CONFIG_FAIR_GROUP_SCHED */
435
Peter Zijlstra029632f2011-10-25 10:00:11 +0200436#ifdef CONFIG_RT_GROUP_SCHED
437 struct list_head leaf_rt_rq_list;
438#endif
439
440 /*
441 * This is part of a global counter where only the total sum
442 * over all CPUs matters. A task can increase this counter on
443 * one CPU and if it got migrated afterwards it may decrease
444 * it on another CPU. Always updated under the runqueue lock:
445 */
446 unsigned long nr_uninterruptible;
447
448 struct task_struct *curr, *idle, *stop;
449 unsigned long next_balance;
450 struct mm_struct *prev_mm;
451
452 u64 clock;
453 u64 clock_task;
454
455 atomic_t nr_iowait;
456
457#ifdef CONFIG_SMP
458 struct root_domain *rd;
459 struct sched_domain *sd;
460
461 unsigned long cpu_power;
462
463 unsigned char idle_balance;
464 /* For active balancing */
465 int post_schedule;
466 int active_balance;
467 int push_cpu;
468 struct cpu_stop_work active_balance_work;
469 /* cpu of this runqueue: */
470 int cpu;
471 int online;
472
Peter Zijlstra367456c2012-02-20 21:49:09 +0100473 struct list_head cfs_tasks;
474
Peter Zijlstra029632f2011-10-25 10:00:11 +0200475 u64 rt_avg;
476 u64 age_stamp;
477 u64 idle_stamp;
478 u64 avg_idle;
479#endif
480
481#ifdef CONFIG_IRQ_TIME_ACCOUNTING
482 u64 prev_irq_time;
483#endif
484#ifdef CONFIG_PARAVIRT
485 u64 prev_steal_time;
486#endif
487#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
488 u64 prev_steal_time_rq;
489#endif
490
491 /* calc_load related fields */
492 unsigned long calc_load_update;
493 long calc_load_active;
494
495#ifdef CONFIG_SCHED_HRTICK
496#ifdef CONFIG_SMP
497 int hrtick_csd_pending;
498 struct call_single_data hrtick_csd;
499#endif
500 struct hrtimer hrtick_timer;
501#endif
502
503#ifdef CONFIG_SCHEDSTATS
504 /* latency stats */
505 struct sched_info rq_sched_info;
506 unsigned long long rq_cpu_time;
507 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
508
509 /* sys_sched_yield() stats */
510 unsigned int yld_count;
511
512 /* schedule() stats */
Peter Zijlstra029632f2011-10-25 10:00:11 +0200513 unsigned int sched_count;
514 unsigned int sched_goidle;
515
516 /* try_to_wake_up() stats */
517 unsigned int ttwu_count;
518 unsigned int ttwu_local;
519#endif
520
521#ifdef CONFIG_SMP
522 struct llist_head wake_list;
523#endif
Ben Segall18bf2802012-10-04 12:51:20 +0200524
525 struct sched_avg avg;
Peter Zijlstra029632f2011-10-25 10:00:11 +0200526};
527
528static inline int cpu_of(struct rq *rq)
529{
530#ifdef CONFIG_SMP
531 return rq->cpu;
532#else
533 return 0;
534#endif
535}
536
537DECLARE_PER_CPU(struct rq, runqueues);
538
Peter Zijlstra518cd622011-12-07 15:07:31 +0100539#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
540#define this_rq() (&__get_cpu_var(runqueues))
541#define task_rq(p) cpu_rq(task_cpu(p))
542#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
543#define raw_rq() (&__raw_get_cpu_var(runqueues))
544
Frederic Weisbecker78becc22013-04-12 01:51:02 +0200545static inline u64 rq_clock(struct rq *rq)
546{
547 return rq->clock;
548}
549
550static inline u64 rq_clock_task(struct rq *rq)
551{
552 return rq->clock_task;
553}
554
Peter Zijlstra518cd622011-12-07 15:07:31 +0100555#ifdef CONFIG_SMP
556
Peter Zijlstra029632f2011-10-25 10:00:11 +0200557#define rcu_dereference_check_sched_domain(p) \
558 rcu_dereference_check((p), \
559 lockdep_is_held(&sched_domains_mutex))
560
561/*
562 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
563 * See detach_destroy_domains: synchronize_sched for details.
564 *
565 * The domain tree of any CPU may only be accessed from within
566 * preempt-disabled sections.
567 */
568#define for_each_domain(cpu, __sd) \
Peter Zijlstra518cd622011-12-07 15:07:31 +0100569 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
570 __sd; __sd = __sd->parent)
Peter Zijlstra029632f2011-10-25 10:00:11 +0200571
Suresh Siddha77e81362011-11-17 11:08:23 -0800572#define for_each_lower_domain(sd) for (; sd; sd = sd->child)
573
Peter Zijlstra518cd622011-12-07 15:07:31 +0100574/**
575 * highest_flag_domain - Return highest sched_domain containing flag.
576 * @cpu: The cpu whose highest level of sched domain is to
577 * be returned.
578 * @flag: The flag to check for the highest sched_domain
579 * for the given cpu.
580 *
581 * Returns the highest sched_domain of a cpu which contains the given flag.
582 */
583static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
584{
585 struct sched_domain *sd, *hsd = NULL;
586
587 for_each_domain(cpu, sd) {
588 if (!(sd->flags & flag))
589 break;
590 hsd = sd;
591 }
592
593 return hsd;
594}
595
596DECLARE_PER_CPU(struct sched_domain *, sd_llc);
Peter Zijlstra7d9ffa82013-07-04 12:56:46 +0800597DECLARE_PER_CPU(int, sd_llc_size);
Peter Zijlstra518cd622011-12-07 15:07:31 +0100598DECLARE_PER_CPU(int, sd_llc_id);
599
Li Zefan5e6521e2013-03-05 16:06:23 +0800600struct sched_group_power {
601 atomic_t ref;
602 /*
603 * CPU power of this group, SCHED_LOAD_SCALE being max power for a
604 * single CPU.
605 */
606 unsigned int power, power_orig;
607 unsigned long next_update;
Peter Zijlstra62633222013-08-19 12:41:09 +0200608 int imbalance; /* XXX unrelated to power but shared group state */
Li Zefan5e6521e2013-03-05 16:06:23 +0800609 /*
610 * Number of busy cpus in this group.
611 */
612 atomic_t nr_busy_cpus;
613
614 unsigned long cpumask[0]; /* iteration mask */
615};
616
617struct sched_group {
618 struct sched_group *next; /* Must be a circular list */
619 atomic_t ref;
620
621 unsigned int group_weight;
622 struct sched_group_power *sgp;
623
624 /*
625 * The CPUs this group covers.
626 *
627 * NOTE: this field is variable length. (Allocated dynamically
628 * by attaching extra space to the end of the structure,
629 * depending on how many CPUs the kernel has booted up with)
630 */
631 unsigned long cpumask[0];
632};
633
634static inline struct cpumask *sched_group_cpus(struct sched_group *sg)
635{
636 return to_cpumask(sg->cpumask);
637}
638
639/*
640 * cpumask masking which cpus in the group are allowed to iterate up the domain
641 * tree.
642 */
643static inline struct cpumask *sched_group_mask(struct sched_group *sg)
644{
645 return to_cpumask(sg->sgp->cpumask);
646}
647
648/**
649 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
650 * @group: The group whose first cpu is to be returned.
651 */
652static inline unsigned int group_first_cpu(struct sched_group *group)
653{
654 return cpumask_first(sched_group_cpus(group));
655}
656
Peter Zijlstrac1174872012-05-31 14:47:33 +0200657extern int group_balance_cpu(struct sched_group *sg);
658
Peter Zijlstra518cd622011-12-07 15:07:31 +0100659#endif /* CONFIG_SMP */
Peter Zijlstra029632f2011-10-25 10:00:11 +0200660
Peter Zijlstra391e43d2011-11-15 17:14:39 +0100661#include "stats.h"
662#include "auto_group.h"
Peter Zijlstra029632f2011-10-25 10:00:11 +0200663
664#ifdef CONFIG_CGROUP_SCHED
665
666/*
667 * Return the group to which this tasks belongs.
668 *
Tejun Heo8af01f52013-08-08 20:11:22 -0400669 * We cannot use task_css() and friends because the cgroup subsystem
670 * changes that value before the cgroup_subsys::attach() method is called,
671 * therefore we cannot pin it and might observe the wrong value.
Peter Zijlstra8323f262012-06-22 13:36:05 +0200672 *
673 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
674 * core changes this before calling sched_move_task().
675 *
676 * Instead we use a 'copy' which is updated from sched_move_task() while
677 * holding both task_struct::pi_lock and rq::lock.
Peter Zijlstra029632f2011-10-25 10:00:11 +0200678 */
679static inline struct task_group *task_group(struct task_struct *p)
680{
Peter Zijlstra8323f262012-06-22 13:36:05 +0200681 return p->sched_task_group;
Peter Zijlstra029632f2011-10-25 10:00:11 +0200682}
683
684/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
685static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
686{
687#if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED)
688 struct task_group *tg = task_group(p);
689#endif
690
691#ifdef CONFIG_FAIR_GROUP_SCHED
692 p->se.cfs_rq = tg->cfs_rq[cpu];
693 p->se.parent = tg->se[cpu];
694#endif
695
696#ifdef CONFIG_RT_GROUP_SCHED
697 p->rt.rt_rq = tg->rt_rq[cpu];
698 p->rt.parent = tg->rt_se[cpu];
699#endif
700}
701
702#else /* CONFIG_CGROUP_SCHED */
703
704static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
705static inline struct task_group *task_group(struct task_struct *p)
706{
707 return NULL;
708}
709
710#endif /* CONFIG_CGROUP_SCHED */
711
712static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
713{
714 set_task_rq(p, cpu);
715#ifdef CONFIG_SMP
716 /*
717 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
718 * successfuly executed on another CPU. We must ensure that updates of
719 * per-task data have been completed by this moment.
720 */
721 smp_wmb();
722 task_thread_info(p)->cpu = cpu;
723#endif
724}
725
726/*
727 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
728 */
729#ifdef CONFIG_SCHED_DEBUG
Ingo Molnarc5905af2012-02-24 08:31:31 +0100730# include <linux/static_key.h>
Peter Zijlstra029632f2011-10-25 10:00:11 +0200731# define const_debug __read_mostly
732#else
733# define const_debug const
734#endif
735
736extern const_debug unsigned int sysctl_sched_features;
737
738#define SCHED_FEAT(name, enabled) \
739 __SCHED_FEAT_##name ,
740
741enum {
Peter Zijlstra391e43d2011-11-15 17:14:39 +0100742#include "features.h"
Peter Zijlstraf8b6d1c2011-07-06 14:20:14 +0200743 __SCHED_FEAT_NR,
Peter Zijlstra029632f2011-10-25 10:00:11 +0200744};
745
746#undef SCHED_FEAT
747
Peter Zijlstraf8b6d1c2011-07-06 14:20:14 +0200748#if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
Ingo Molnarc5905af2012-02-24 08:31:31 +0100749static __always_inline bool static_branch__true(struct static_key *key)
Peter Zijlstraf8b6d1c2011-07-06 14:20:14 +0200750{
Ingo Molnarc5905af2012-02-24 08:31:31 +0100751 return static_key_true(key); /* Not out of line branch. */
Peter Zijlstraf8b6d1c2011-07-06 14:20:14 +0200752}
753
Ingo Molnarc5905af2012-02-24 08:31:31 +0100754static __always_inline bool static_branch__false(struct static_key *key)
Peter Zijlstraf8b6d1c2011-07-06 14:20:14 +0200755{
Ingo Molnarc5905af2012-02-24 08:31:31 +0100756 return static_key_false(key); /* Out of line branch. */
Peter Zijlstraf8b6d1c2011-07-06 14:20:14 +0200757}
758
759#define SCHED_FEAT(name, enabled) \
Ingo Molnarc5905af2012-02-24 08:31:31 +0100760static __always_inline bool static_branch_##name(struct static_key *key) \
Peter Zijlstraf8b6d1c2011-07-06 14:20:14 +0200761{ \
762 return static_branch__##enabled(key); \
763}
764
765#include "features.h"
766
767#undef SCHED_FEAT
768
Ingo Molnarc5905af2012-02-24 08:31:31 +0100769extern struct static_key sched_feat_keys[__SCHED_FEAT_NR];
Peter Zijlstraf8b6d1c2011-07-06 14:20:14 +0200770#define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x]))
771#else /* !(SCHED_DEBUG && HAVE_JUMP_LABEL) */
Peter Zijlstra029632f2011-10-25 10:00:11 +0200772#define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
Peter Zijlstraf8b6d1c2011-07-06 14:20:14 +0200773#endif /* SCHED_DEBUG && HAVE_JUMP_LABEL */
Peter Zijlstra029632f2011-10-25 10:00:11 +0200774
Peter Zijlstracbee9f82012-10-25 14:16:43 +0200775#ifdef CONFIG_NUMA_BALANCING
776#define sched_feat_numa(x) sched_feat(x)
Mel Gorman3105b862012-11-23 11:23:49 +0000777#ifdef CONFIG_SCHED_DEBUG
778#define numabalancing_enabled sched_feat_numa(NUMA)
779#else
780extern bool numabalancing_enabled;
781#endif /* CONFIG_SCHED_DEBUG */
Peter Zijlstracbee9f82012-10-25 14:16:43 +0200782#else
783#define sched_feat_numa(x) (0)
Mel Gorman3105b862012-11-23 11:23:49 +0000784#define numabalancing_enabled (0)
785#endif /* CONFIG_NUMA_BALANCING */
Peter Zijlstracbee9f82012-10-25 14:16:43 +0200786
Peter Zijlstra029632f2011-10-25 10:00:11 +0200787static inline u64 global_rt_period(void)
788{
789 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
790}
791
792static inline u64 global_rt_runtime(void)
793{
794 if (sysctl_sched_rt_runtime < 0)
795 return RUNTIME_INF;
796
797 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
798}
799
800
801
802static inline int task_current(struct rq *rq, struct task_struct *p)
803{
804 return rq->curr == p;
805}
806
807static inline int task_running(struct rq *rq, struct task_struct *p)
808{
809#ifdef CONFIG_SMP
810 return p->on_cpu;
811#else
812 return task_current(rq, p);
813#endif
814}
815
816
817#ifndef prepare_arch_switch
818# define prepare_arch_switch(next) do { } while (0)
819#endif
820#ifndef finish_arch_switch
821# define finish_arch_switch(prev) do { } while (0)
822#endif
Catalin Marinas01f23e12011-11-27 21:43:10 +0000823#ifndef finish_arch_post_lock_switch
824# define finish_arch_post_lock_switch() do { } while (0)
825#endif
Peter Zijlstra029632f2011-10-25 10:00:11 +0200826
827#ifndef __ARCH_WANT_UNLOCKED_CTXSW
828static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
829{
830#ifdef CONFIG_SMP
831 /*
832 * We can optimise this out completely for !SMP, because the
833 * SMP rebalancing from interrupt is the only thing that cares
834 * here.
835 */
836 next->on_cpu = 1;
837#endif
838}
839
840static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
841{
842#ifdef CONFIG_SMP
843 /*
844 * After ->on_cpu is cleared, the task can be moved to a different CPU.
845 * We must ensure this doesn't happen until the switch is completely
846 * finished.
847 */
848 smp_wmb();
849 prev->on_cpu = 0;
850#endif
851#ifdef CONFIG_DEBUG_SPINLOCK
852 /* this is a valid case when another task releases the spinlock */
853 rq->lock.owner = current;
854#endif
855 /*
856 * If we are tracking spinlock dependencies then we have to
857 * fix up the runqueue lock - which gets 'carried over' from
858 * prev into current:
859 */
860 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
861
862 raw_spin_unlock_irq(&rq->lock);
863}
864
865#else /* __ARCH_WANT_UNLOCKED_CTXSW */
866static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
867{
868#ifdef CONFIG_SMP
869 /*
870 * We can optimise this out completely for !SMP, because the
871 * SMP rebalancing from interrupt is the only thing that cares
872 * here.
873 */
874 next->on_cpu = 1;
875#endif
Peter Zijlstra029632f2011-10-25 10:00:11 +0200876 raw_spin_unlock(&rq->lock);
Peter Zijlstra029632f2011-10-25 10:00:11 +0200877}
878
879static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
880{
881#ifdef CONFIG_SMP
882 /*
883 * After ->on_cpu is cleared, the task can be moved to a different CPU.
884 * We must ensure this doesn't happen until the switch is completely
885 * finished.
886 */
887 smp_wmb();
888 prev->on_cpu = 0;
889#endif
Peter Zijlstra029632f2011-10-25 10:00:11 +0200890 local_irq_enable();
Peter Zijlstra029632f2011-10-25 10:00:11 +0200891}
892#endif /* __ARCH_WANT_UNLOCKED_CTXSW */
893
Li Zefanb13095f2013-03-05 16:06:38 +0800894/*
895 * wake flags
896 */
897#define WF_SYNC 0x01 /* waker goes to sleep after wakeup */
898#define WF_FORK 0x02 /* child wakeup after fork */
899#define WF_MIGRATED 0x4 /* internal use, task got migrated */
900
Peter Zijlstra029632f2011-10-25 10:00:11 +0200901/*
902 * To aid in avoiding the subversion of "niceness" due to uneven distribution
903 * of tasks with abnormal "nice" values across CPUs the contribution that
904 * each task makes to its run queue's load is weighted according to its
905 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
906 * scaled version of the new time slice allocation that they receive on time
907 * slice expiry etc.
908 */
909
910#define WEIGHT_IDLEPRIO 3
911#define WMULT_IDLEPRIO 1431655765
912
913/*
914 * Nice levels are multiplicative, with a gentle 10% change for every
915 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
916 * nice 1, it will get ~10% less CPU time than another CPU-bound task
917 * that remained on nice 0.
918 *
919 * The "10% effect" is relative and cumulative: from _any_ nice level,
920 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
921 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
922 * If a task goes up by ~10% and another task goes down by ~10% then
923 * the relative distance between them is ~25%.)
924 */
925static const int prio_to_weight[40] = {
926 /* -20 */ 88761, 71755, 56483, 46273, 36291,
927 /* -15 */ 29154, 23254, 18705, 14949, 11916,
928 /* -10 */ 9548, 7620, 6100, 4904, 3906,
929 /* -5 */ 3121, 2501, 1991, 1586, 1277,
930 /* 0 */ 1024, 820, 655, 526, 423,
931 /* 5 */ 335, 272, 215, 172, 137,
932 /* 10 */ 110, 87, 70, 56, 45,
933 /* 15 */ 36, 29, 23, 18, 15,
934};
935
936/*
937 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
938 *
939 * In cases where the weight does not change often, we can use the
940 * precalculated inverse to speed up arithmetics by turning divisions
941 * into multiplications:
942 */
943static const u32 prio_to_wmult[40] = {
944 /* -20 */ 48388, 59856, 76040, 92818, 118348,
945 /* -15 */ 147320, 184698, 229616, 287308, 360437,
946 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
947 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
948 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
949 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
950 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
951 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
952};
953
Li Zefanc82ba9f2013-03-05 16:06:55 +0800954#define ENQUEUE_WAKEUP 1
955#define ENQUEUE_HEAD 2
956#ifdef CONFIG_SMP
957#define ENQUEUE_WAKING 4 /* sched_class::task_waking was called */
958#else
959#define ENQUEUE_WAKING 0
960#endif
961
962#define DEQUEUE_SLEEP 1
963
964struct sched_class {
965 const struct sched_class *next;
966
967 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags);
968 void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags);
969 void (*yield_task) (struct rq *rq);
970 bool (*yield_to_task) (struct rq *rq, struct task_struct *p, bool preempt);
971
972 void (*check_preempt_curr) (struct rq *rq, struct task_struct *p, int flags);
973
974 struct task_struct * (*pick_next_task) (struct rq *rq);
975 void (*put_prev_task) (struct rq *rq, struct task_struct *p);
976
977#ifdef CONFIG_SMP
978 int (*select_task_rq)(struct task_struct *p, int sd_flag, int flags);
979 void (*migrate_task_rq)(struct task_struct *p, int next_cpu);
980
981 void (*pre_schedule) (struct rq *this_rq, struct task_struct *task);
982 void (*post_schedule) (struct rq *this_rq);
983 void (*task_waking) (struct task_struct *task);
984 void (*task_woken) (struct rq *this_rq, struct task_struct *task);
985
986 void (*set_cpus_allowed)(struct task_struct *p,
987 const struct cpumask *newmask);
988
989 void (*rq_online)(struct rq *rq);
990 void (*rq_offline)(struct rq *rq);
991#endif
992
993 void (*set_curr_task) (struct rq *rq);
994 void (*task_tick) (struct rq *rq, struct task_struct *p, int queued);
995 void (*task_fork) (struct task_struct *p);
996
997 void (*switched_from) (struct rq *this_rq, struct task_struct *task);
998 void (*switched_to) (struct rq *this_rq, struct task_struct *task);
999 void (*prio_changed) (struct rq *this_rq, struct task_struct *task,
1000 int oldprio);
1001
1002 unsigned int (*get_rr_interval) (struct rq *rq,
1003 struct task_struct *task);
1004
1005#ifdef CONFIG_FAIR_GROUP_SCHED
1006 void (*task_move_group) (struct task_struct *p, int on_rq);
1007#endif
1008};
Peter Zijlstra029632f2011-10-25 10:00:11 +02001009
1010#define sched_class_highest (&stop_sched_class)
1011#define for_each_class(class) \
1012 for (class = sched_class_highest; class; class = class->next)
1013
1014extern const struct sched_class stop_sched_class;
1015extern const struct sched_class rt_sched_class;
1016extern const struct sched_class fair_sched_class;
1017extern const struct sched_class idle_sched_class;
1018
1019
1020#ifdef CONFIG_SMP
1021
Li Zefanb7192032013-03-07 10:00:26 +08001022extern void update_group_power(struct sched_domain *sd, int cpu);
1023
Peter Zijlstra029632f2011-10-25 10:00:11 +02001024extern void trigger_load_balance(struct rq *rq, int cpu);
1025extern void idle_balance(int this_cpu, struct rq *this_rq);
1026
Vincent Guittot642dbc32013-04-18 18:34:26 +02001027extern void idle_enter_fair(struct rq *this_rq);
1028extern void idle_exit_fair(struct rq *this_rq);
Vincent Guittot642dbc32013-04-18 18:34:26 +02001029
Peter Zijlstra029632f2011-10-25 10:00:11 +02001030#else /* CONFIG_SMP */
1031
1032static inline void idle_balance(int cpu, struct rq *rq)
1033{
1034}
1035
1036#endif
1037
1038extern void sysrq_sched_debug_show(void);
1039extern void sched_init_granularity(void);
1040extern void update_max_interval(void);
Peter Zijlstra029632f2011-10-25 10:00:11 +02001041extern void init_sched_rt_class(void);
1042extern void init_sched_fair_class(void);
1043
1044extern void resched_task(struct task_struct *p);
1045extern void resched_cpu(int cpu);
1046
1047extern struct rt_bandwidth def_rt_bandwidth;
1048extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime);
1049
Peter Zijlstra556061b2012-05-11 17:31:26 +02001050extern void update_idle_cpu_load(struct rq *this_rq);
Peter Zijlstra029632f2011-10-25 10:00:11 +02001051
Alex Shia75cdaa2013-06-20 10:18:47 +08001052extern void init_task_runnable_average(struct task_struct *p);
1053
Frederic Weisbecker73fbec62012-06-16 15:57:37 +02001054#ifdef CONFIG_PARAVIRT
1055static inline u64 steal_ticks(u64 steal)
1056{
1057 if (unlikely(steal > NSEC_PER_SEC))
1058 return div_u64(steal, TICK_NSEC);
1059
1060 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
1061}
1062#endif
1063
Peter Zijlstra029632f2011-10-25 10:00:11 +02001064static inline void inc_nr_running(struct rq *rq)
1065{
1066 rq->nr_running++;
Frederic Weisbecker9f3660c2013-04-20 14:35:09 +02001067
1068#ifdef CONFIG_NO_HZ_FULL
1069 if (rq->nr_running == 2) {
1070 if (tick_nohz_full_cpu(rq->cpu)) {
1071 /* Order rq->nr_running write against the IPI */
1072 smp_wmb();
1073 smp_send_reschedule(rq->cpu);
1074 }
1075 }
1076#endif
Peter Zijlstra029632f2011-10-25 10:00:11 +02001077}
1078
1079static inline void dec_nr_running(struct rq *rq)
1080{
1081 rq->nr_running--;
1082}
1083
Frederic Weisbecker265f22a2013-05-03 03:39:05 +02001084static inline void rq_last_tick_reset(struct rq *rq)
1085{
1086#ifdef CONFIG_NO_HZ_FULL
1087 rq->last_sched_tick = jiffies;
1088#endif
1089}
1090
Peter Zijlstra029632f2011-10-25 10:00:11 +02001091extern void update_rq_clock(struct rq *rq);
1092
1093extern void activate_task(struct rq *rq, struct task_struct *p, int flags);
1094extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags);
1095
1096extern void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
1097
1098extern const_debug unsigned int sysctl_sched_time_avg;
1099extern const_debug unsigned int sysctl_sched_nr_migrate;
1100extern const_debug unsigned int sysctl_sched_migration_cost;
1101
1102static inline u64 sched_avg_period(void)
1103{
1104 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1105}
1106
Peter Zijlstra029632f2011-10-25 10:00:11 +02001107#ifdef CONFIG_SCHED_HRTICK
1108
1109/*
1110 * Use hrtick when:
1111 * - enabled by features
1112 * - hrtimer is actually high res
1113 */
1114static inline int hrtick_enabled(struct rq *rq)
1115{
1116 if (!sched_feat(HRTICK))
1117 return 0;
1118 if (!cpu_active(cpu_of(rq)))
1119 return 0;
1120 return hrtimer_is_hres_active(&rq->hrtick_timer);
1121}
1122
1123void hrtick_start(struct rq *rq, u64 delay);
1124
Mike Galbraithb39e66e2011-11-22 15:20:07 +01001125#else
1126
1127static inline int hrtick_enabled(struct rq *rq)
1128{
1129 return 0;
1130}
1131
Peter Zijlstra029632f2011-10-25 10:00:11 +02001132#endif /* CONFIG_SCHED_HRTICK */
1133
1134#ifdef CONFIG_SMP
1135extern void sched_avg_update(struct rq *rq);
1136static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1137{
1138 rq->rt_avg += rt_delta;
1139 sched_avg_update(rq);
1140}
1141#else
1142static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta) { }
1143static inline void sched_avg_update(struct rq *rq) { }
1144#endif
1145
1146extern void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period);
1147
1148#ifdef CONFIG_SMP
1149#ifdef CONFIG_PREEMPT
1150
1151static inline void double_rq_lock(struct rq *rq1, struct rq *rq2);
1152
1153/*
1154 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1155 * way at the expense of forcing extra atomic operations in all
1156 * invocations. This assures that the double_lock is acquired using the
1157 * same underlying policy as the spinlock_t on this architecture, which
1158 * reduces latency compared to the unfair variant below. However, it
1159 * also adds more overhead and therefore may reduce throughput.
1160 */
1161static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1162 __releases(this_rq->lock)
1163 __acquires(busiest->lock)
1164 __acquires(this_rq->lock)
1165{
1166 raw_spin_unlock(&this_rq->lock);
1167 double_rq_lock(this_rq, busiest);
1168
1169 return 1;
1170}
1171
1172#else
1173/*
1174 * Unfair double_lock_balance: Optimizes throughput at the expense of
1175 * latency by eliminating extra atomic operations when the locks are
1176 * already in proper order on entry. This favors lower cpu-ids and will
1177 * grant the double lock to lower cpus over higher ids under contention,
1178 * regardless of entry order into the function.
1179 */
1180static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1181 __releases(this_rq->lock)
1182 __acquires(busiest->lock)
1183 __acquires(this_rq->lock)
1184{
1185 int ret = 0;
1186
1187 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1188 if (busiest < this_rq) {
1189 raw_spin_unlock(&this_rq->lock);
1190 raw_spin_lock(&busiest->lock);
1191 raw_spin_lock_nested(&this_rq->lock,
1192 SINGLE_DEPTH_NESTING);
1193 ret = 1;
1194 } else
1195 raw_spin_lock_nested(&busiest->lock,
1196 SINGLE_DEPTH_NESTING);
1197 }
1198 return ret;
1199}
1200
1201#endif /* CONFIG_PREEMPT */
1202
1203/*
1204 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1205 */
1206static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1207{
1208 if (unlikely(!irqs_disabled())) {
1209 /* printk() doesn't work good under rq->lock */
1210 raw_spin_unlock(&this_rq->lock);
1211 BUG_ON(1);
1212 }
1213
1214 return _double_lock_balance(this_rq, busiest);
1215}
1216
1217static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1218 __releases(busiest->lock)
1219{
1220 raw_spin_unlock(&busiest->lock);
1221 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1222}
1223
1224/*
1225 * double_rq_lock - safely lock two runqueues
1226 *
1227 * Note this does not disable interrupts like task_rq_lock,
1228 * you need to do so manually before calling.
1229 */
1230static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1231 __acquires(rq1->lock)
1232 __acquires(rq2->lock)
1233{
1234 BUG_ON(!irqs_disabled());
1235 if (rq1 == rq2) {
1236 raw_spin_lock(&rq1->lock);
1237 __acquire(rq2->lock); /* Fake it out ;) */
1238 } else {
1239 if (rq1 < rq2) {
1240 raw_spin_lock(&rq1->lock);
1241 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1242 } else {
1243 raw_spin_lock(&rq2->lock);
1244 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1245 }
1246 }
1247}
1248
1249/*
1250 * double_rq_unlock - safely unlock two runqueues
1251 *
1252 * Note this does not restore interrupts like task_rq_unlock,
1253 * you need to do so manually after calling.
1254 */
1255static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1256 __releases(rq1->lock)
1257 __releases(rq2->lock)
1258{
1259 raw_spin_unlock(&rq1->lock);
1260 if (rq1 != rq2)
1261 raw_spin_unlock(&rq2->lock);
1262 else
1263 __release(rq2->lock);
1264}
1265
1266#else /* CONFIG_SMP */
1267
1268/*
1269 * double_rq_lock - safely lock two runqueues
1270 *
1271 * Note this does not disable interrupts like task_rq_lock,
1272 * you need to do so manually before calling.
1273 */
1274static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1275 __acquires(rq1->lock)
1276 __acquires(rq2->lock)
1277{
1278 BUG_ON(!irqs_disabled());
1279 BUG_ON(rq1 != rq2);
1280 raw_spin_lock(&rq1->lock);
1281 __acquire(rq2->lock); /* Fake it out ;) */
1282}
1283
1284/*
1285 * double_rq_unlock - safely unlock two runqueues
1286 *
1287 * Note this does not restore interrupts like task_rq_unlock,
1288 * you need to do so manually after calling.
1289 */
1290static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1291 __releases(rq1->lock)
1292 __releases(rq2->lock)
1293{
1294 BUG_ON(rq1 != rq2);
1295 raw_spin_unlock(&rq1->lock);
1296 __release(rq2->lock);
1297}
1298
1299#endif
1300
1301extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq);
1302extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq);
1303extern void print_cfs_stats(struct seq_file *m, int cpu);
1304extern void print_rt_stats(struct seq_file *m, int cpu);
1305
1306extern void init_cfs_rq(struct cfs_rq *cfs_rq);
1307extern void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq);
Peter Zijlstra029632f2011-10-25 10:00:11 +02001308
1309extern void account_cfs_bandwidth_used(int enabled, int was_enabled);
Suresh Siddha1c792db2011-12-01 17:07:32 -08001310
Frederic Weisbecker3451d022011-08-10 23:21:01 +02001311#ifdef CONFIG_NO_HZ_COMMON
Suresh Siddha1c792db2011-12-01 17:07:32 -08001312enum rq_nohz_flag_bits {
1313 NOHZ_TICK_STOPPED,
1314 NOHZ_BALANCE_KICK,
1315};
1316
1317#define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags)
1318#endif
Frederic Weisbecker73fbec62012-06-16 15:57:37 +02001319
1320#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1321
1322DECLARE_PER_CPU(u64, cpu_hardirq_time);
1323DECLARE_PER_CPU(u64, cpu_softirq_time);
1324
1325#ifndef CONFIG_64BIT
1326DECLARE_PER_CPU(seqcount_t, irq_time_seq);
1327
1328static inline void irq_time_write_begin(void)
1329{
1330 __this_cpu_inc(irq_time_seq.sequence);
1331 smp_wmb();
1332}
1333
1334static inline void irq_time_write_end(void)
1335{
1336 smp_wmb();
1337 __this_cpu_inc(irq_time_seq.sequence);
1338}
1339
1340static inline u64 irq_time_read(int cpu)
1341{
1342 u64 irq_time;
1343 unsigned seq;
1344
1345 do {
1346 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1347 irq_time = per_cpu(cpu_softirq_time, cpu) +
1348 per_cpu(cpu_hardirq_time, cpu);
1349 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1350
1351 return irq_time;
1352}
1353#else /* CONFIG_64BIT */
1354static inline void irq_time_write_begin(void)
1355{
1356}
1357
1358static inline void irq_time_write_end(void)
1359{
1360}
1361
1362static inline u64 irq_time_read(int cpu)
1363{
1364 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1365}
1366#endif /* CONFIG_64BIT */
1367#endif /* CONFIG_IRQ_TIME_ACCOUNTING */