]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * kernel/sched.c | |
3 | * | |
4 | * Kernel scheduler and related syscalls | |
5 | * | |
6 | * Copyright (C) 1991-2002 Linus Torvalds | |
7 | * | |
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 | |
11 | * by Andrea Arcangeli | |
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 | |
22 | * by Peter Williams | |
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 | |
27 | */ | |
28 | ||
29 | #include <linux/mm.h> | |
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/perf_event.h> | |
43 | #include <linux/security.h> | |
44 | #include <linux/notifier.h> | |
45 | #include <linux/profile.h> | |
46 | #include <linux/freezer.h> | |
47 | #include <linux/vmalloc.h> | |
48 | #include <linux/blkdev.h> | |
49 | #include <linux/delay.h> | |
50 | #include <linux/pid_namespace.h> | |
51 | #include <linux/smp.h> | |
52 | #include <linux/threads.h> | |
53 | #include <linux/timer.h> | |
54 | #include <linux/rcupdate.h> | |
55 | #include <linux/cpu.h> | |
56 | #include <linux/cpuset.h> | |
57 | #include <linux/percpu.h> | |
58 | #include <linux/kthread.h> | |
59 | #include <linux/proc_fs.h> | |
60 | #include <linux/seq_file.h> | |
61 | #include <linux/sysctl.h> | |
62 | #include <linux/syscalls.h> | |
63 | #include <linux/times.h> | |
64 | #include <linux/tsacct_kern.h> | |
65 | #include <linux/kprobes.h> | |
66 | #include <linux/delayacct.h> | |
67 | #include <linux/unistd.h> | |
68 | #include <linux/pagemap.h> | |
69 | #include <linux/hrtimer.h> | |
70 | #include <linux/tick.h> | |
71 | #include <linux/debugfs.h> | |
72 | #include <linux/ctype.h> | |
73 | #include <linux/ftrace.h> | |
74 | ||
75 | #include <asm/tlb.h> | |
76 | #include <asm/irq_regs.h> | |
77 | ||
78 | #include "sched_cpupri.h" | |
79 | ||
80 | #define CREATE_TRACE_POINTS | |
81 | #include <trace/events/sched.h> | |
82 | ||
83 | /* | |
84 | * Convert user-nice values [ -20 ... 0 ... 19 ] | |
85 | * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ], | |
86 | * and back. | |
87 | */ | |
88 | #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20) | |
89 | #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20) | |
90 | #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio) | |
91 | ||
92 | /* | |
93 | * 'User priority' is the nice value converted to something we | |
94 | * can work with better when scaling various scheduler parameters, | |
95 | * it's a [ 0 ... 39 ] range. | |
96 | */ | |
97 | #define USER_PRIO(p) ((p)-MAX_RT_PRIO) | |
98 | #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio) | |
99 | #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO)) | |
100 | ||
101 | /* | |
102 | * Helpers for converting nanosecond timing to jiffy resolution | |
103 | */ | |
104 | #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ)) | |
105 | ||
106 | #define NICE_0_LOAD SCHED_LOAD_SCALE | |
107 | #define NICE_0_SHIFT SCHED_LOAD_SHIFT | |
108 | ||
109 | /* | |
110 | * These are the 'tuning knobs' of the scheduler: | |
111 | * | |
112 | * default timeslice is 100 msecs (used only for SCHED_RR tasks). | |
113 | * Timeslices get refilled after they expire. | |
114 | */ | |
115 | #define DEF_TIMESLICE (100 * HZ / 1000) | |
116 | ||
117 | /* | |
118 | * single value that denotes runtime == period, ie unlimited time. | |
119 | */ | |
120 | #define RUNTIME_INF ((u64)~0ULL) | |
121 | ||
122 | static inline int rt_policy(int policy) | |
123 | { | |
124 | if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR)) | |
125 | return 1; | |
126 | return 0; | |
127 | } | |
128 | ||
129 | static inline int task_has_rt_policy(struct task_struct *p) | |
130 | { | |
131 | return rt_policy(p->policy); | |
132 | } | |
133 | ||
134 | /* | |
135 | * This is the priority-queue data structure of the RT scheduling class: | |
136 | */ | |
137 | struct rt_prio_array { | |
138 | DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */ | |
139 | struct list_head queue[MAX_RT_PRIO]; | |
140 | }; | |
141 | ||
142 | struct rt_bandwidth { | |
143 | /* nests inside the rq lock: */ | |
144 | spinlock_t rt_runtime_lock; | |
145 | ktime_t rt_period; | |
146 | u64 rt_runtime; | |
147 | struct hrtimer rt_period_timer; | |
148 | }; | |
149 | ||
150 | static struct rt_bandwidth def_rt_bandwidth; | |
151 | ||
152 | static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun); | |
153 | ||
154 | static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer) | |
155 | { | |
156 | struct rt_bandwidth *rt_b = | |
157 | container_of(timer, struct rt_bandwidth, rt_period_timer); | |
158 | ktime_t now; | |
159 | int overrun; | |
160 | int idle = 0; | |
161 | ||
162 | for (;;) { | |
163 | now = hrtimer_cb_get_time(timer); | |
164 | overrun = hrtimer_forward(timer, now, rt_b->rt_period); | |
165 | ||
166 | if (!overrun) | |
167 | break; | |
168 | ||
169 | idle = do_sched_rt_period_timer(rt_b, overrun); | |
170 | } | |
171 | ||
172 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
173 | } | |
174 | ||
175 | static | |
176 | void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime) | |
177 | { | |
178 | rt_b->rt_period = ns_to_ktime(period); | |
179 | rt_b->rt_runtime = runtime; | |
180 | ||
181 | spin_lock_init(&rt_b->rt_runtime_lock); | |
182 | ||
183 | hrtimer_init(&rt_b->rt_period_timer, | |
184 | CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
185 | rt_b->rt_period_timer.function = sched_rt_period_timer; | |
186 | } | |
187 | ||
188 | static inline int rt_bandwidth_enabled(void) | |
189 | { | |
190 | return sysctl_sched_rt_runtime >= 0; | |
191 | } | |
192 | ||
193 | static void start_rt_bandwidth(struct rt_bandwidth *rt_b) | |
194 | { | |
195 | ktime_t now; | |
196 | ||
197 | if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF) | |
198 | return; | |
199 | ||
200 | if (hrtimer_active(&rt_b->rt_period_timer)) | |
201 | return; | |
202 | ||
203 | spin_lock(&rt_b->rt_runtime_lock); | |
204 | for (;;) { | |
205 | unsigned long delta; | |
206 | ktime_t soft, hard; | |
207 | ||
208 | if (hrtimer_active(&rt_b->rt_period_timer)) | |
209 | break; | |
210 | ||
211 | now = hrtimer_cb_get_time(&rt_b->rt_period_timer); | |
212 | hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period); | |
213 | ||
214 | soft = hrtimer_get_softexpires(&rt_b->rt_period_timer); | |
215 | hard = hrtimer_get_expires(&rt_b->rt_period_timer); | |
216 | delta = ktime_to_ns(ktime_sub(hard, soft)); | |
217 | __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta, | |
218 | HRTIMER_MODE_ABS_PINNED, 0); | |
219 | } | |
220 | spin_unlock(&rt_b->rt_runtime_lock); | |
221 | } | |
222 | ||
223 | #ifdef CONFIG_RT_GROUP_SCHED | |
224 | static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b) | |
225 | { | |
226 | hrtimer_cancel(&rt_b->rt_period_timer); | |
227 | } | |
228 | #endif | |
229 | ||
230 | /* | |
231 | * sched_domains_mutex serializes calls to arch_init_sched_domains, | |
232 | * detach_destroy_domains and partition_sched_domains. | |
233 | */ | |
234 | static DEFINE_MUTEX(sched_domains_mutex); | |
235 | ||
236 | #ifdef CONFIG_GROUP_SCHED | |
237 | ||
238 | #include <linux/cgroup.h> | |
239 | ||
240 | struct cfs_rq; | |
241 | ||
242 | static LIST_HEAD(task_groups); | |
243 | ||
244 | /* task group related information */ | |
245 | struct task_group { | |
246 | #ifdef CONFIG_CGROUP_SCHED | |
247 | struct cgroup_subsys_state css; | |
248 | #endif | |
249 | ||
250 | #ifdef CONFIG_USER_SCHED | |
251 | uid_t uid; | |
252 | #endif | |
253 | ||
254 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
255 | /* schedulable entities of this group on each cpu */ | |
256 | struct sched_entity **se; | |
257 | /* runqueue "owned" by this group on each cpu */ | |
258 | struct cfs_rq **cfs_rq; | |
259 | unsigned long shares; | |
260 | #endif | |
261 | ||
262 | #ifdef CONFIG_RT_GROUP_SCHED | |
263 | struct sched_rt_entity **rt_se; | |
264 | struct rt_rq **rt_rq; | |
265 | ||
266 | struct rt_bandwidth rt_bandwidth; | |
267 | #endif | |
268 | ||
269 | struct rcu_head rcu; | |
270 | struct list_head list; | |
271 | ||
272 | struct task_group *parent; | |
273 | struct list_head siblings; | |
274 | struct list_head children; | |
275 | }; | |
276 | ||
277 | #ifdef CONFIG_USER_SCHED | |
278 | ||
279 | /* Helper function to pass uid information to create_sched_user() */ | |
280 | void set_tg_uid(struct user_struct *user) | |
281 | { | |
282 | user->tg->uid = user->uid; | |
283 | } | |
284 | ||
285 | /* | |
286 | * Root task group. | |
287 | * Every UID task group (including init_task_group aka UID-0) will | |
288 | * be a child to this group. | |
289 | */ | |
290 | struct task_group root_task_group; | |
291 | ||
292 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
293 | /* Default task group's sched entity on each cpu */ | |
294 | static DEFINE_PER_CPU(struct sched_entity, init_sched_entity); | |
295 | /* Default task group's cfs_rq on each cpu */ | |
296 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct cfs_rq, init_tg_cfs_rq); | |
297 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
298 | ||
299 | #ifdef CONFIG_RT_GROUP_SCHED | |
300 | static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity); | |
301 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct rt_rq, init_rt_rq); | |
302 | #endif /* CONFIG_RT_GROUP_SCHED */ | |
303 | #else /* !CONFIG_USER_SCHED */ | |
304 | #define root_task_group init_task_group | |
305 | #endif /* CONFIG_USER_SCHED */ | |
306 | ||
307 | /* task_group_lock serializes add/remove of task groups and also changes to | |
308 | * a task group's cpu shares. | |
309 | */ | |
310 | static DEFINE_SPINLOCK(task_group_lock); | |
311 | ||
312 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
313 | ||
314 | #ifdef CONFIG_SMP | |
315 | static int root_task_group_empty(void) | |
316 | { | |
317 | return list_empty(&root_task_group.children); | |
318 | } | |
319 | #endif | |
320 | ||
321 | #ifdef CONFIG_USER_SCHED | |
322 | # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD) | |
323 | #else /* !CONFIG_USER_SCHED */ | |
324 | # define INIT_TASK_GROUP_LOAD NICE_0_LOAD | |
325 | #endif /* CONFIG_USER_SCHED */ | |
326 | ||
327 | /* | |
328 | * A weight of 0 or 1 can cause arithmetics problems. | |
329 | * A weight of a cfs_rq is the sum of weights of which entities | |
330 | * are queued on this cfs_rq, so a weight of a entity should not be | |
331 | * too large, so as the shares value of a task group. | |
332 | * (The default weight is 1024 - so there's no practical | |
333 | * limitation from this.) | |
334 | */ | |
335 | #define MIN_SHARES 2 | |
336 | #define MAX_SHARES (1UL << 18) | |
337 | ||
338 | static int init_task_group_load = INIT_TASK_GROUP_LOAD; | |
339 | #endif | |
340 | ||
341 | /* Default task group. | |
342 | * Every task in system belong to this group at bootup. | |
343 | */ | |
344 | struct task_group init_task_group; | |
345 | ||
346 | /* return group to which a task belongs */ | |
347 | static inline struct task_group *task_group(struct task_struct *p) | |
348 | { | |
349 | struct task_group *tg; | |
350 | ||
351 | #ifdef CONFIG_USER_SCHED | |
352 | rcu_read_lock(); | |
353 | tg = __task_cred(p)->user->tg; | |
354 | rcu_read_unlock(); | |
355 | #elif defined(CONFIG_CGROUP_SCHED) | |
356 | tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id), | |
357 | struct task_group, css); | |
358 | #else | |
359 | tg = &init_task_group; | |
360 | #endif | |
361 | return tg; | |
362 | } | |
363 | ||
364 | /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */ | |
365 | static inline void set_task_rq(struct task_struct *p, unsigned int cpu) | |
366 | { | |
367 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
368 | p->se.cfs_rq = task_group(p)->cfs_rq[cpu]; | |
369 | p->se.parent = task_group(p)->se[cpu]; | |
370 | #endif | |
371 | ||
372 | #ifdef CONFIG_RT_GROUP_SCHED | |
373 | p->rt.rt_rq = task_group(p)->rt_rq[cpu]; | |
374 | p->rt.parent = task_group(p)->rt_se[cpu]; | |
375 | #endif | |
376 | } | |
377 | ||
378 | #else | |
379 | ||
380 | static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { } | |
381 | static inline struct task_group *task_group(struct task_struct *p) | |
382 | { | |
383 | return NULL; | |
384 | } | |
385 | ||
386 | #endif /* CONFIG_GROUP_SCHED */ | |
387 | ||
388 | /* CFS-related fields in a runqueue */ | |
389 | struct cfs_rq { | |
390 | struct load_weight load; | |
391 | unsigned long nr_running; | |
392 | ||
393 | u64 exec_clock; | |
394 | u64 min_vruntime; | |
395 | ||
396 | struct rb_root tasks_timeline; | |
397 | struct rb_node *rb_leftmost; | |
398 | ||
399 | struct list_head tasks; | |
400 | struct list_head *balance_iterator; | |
401 | ||
402 | /* | |
403 | * 'curr' points to currently running entity on this cfs_rq. | |
404 | * It is set to NULL otherwise (i.e when none are currently running). | |
405 | */ | |
406 | struct sched_entity *curr, *next, *last; | |
407 | ||
408 | unsigned int nr_spread_over; | |
409 | ||
410 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
411 | struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */ | |
412 | ||
413 | /* | |
414 | * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in | |
415 | * a hierarchy). Non-leaf lrqs hold other higher schedulable entities | |
416 | * (like users, containers etc.) | |
417 | * | |
418 | * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This | |
419 | * list is used during load balance. | |
420 | */ | |
421 | struct list_head leaf_cfs_rq_list; | |
422 | struct task_group *tg; /* group that "owns" this runqueue */ | |
423 | ||
424 | #ifdef CONFIG_SMP | |
425 | /* | |
426 | * the part of load.weight contributed by tasks | |
427 | */ | |
428 | unsigned long task_weight; | |
429 | ||
430 | /* | |
431 | * h_load = weight * f(tg) | |
432 | * | |
433 | * Where f(tg) is the recursive weight fraction assigned to | |
434 | * this group. | |
435 | */ | |
436 | unsigned long h_load; | |
437 | ||
438 | /* | |
439 | * this cpu's part of tg->shares | |
440 | */ | |
441 | unsigned long shares; | |
442 | ||
443 | /* | |
444 | * load.weight at the time we set shares | |
445 | */ | |
446 | unsigned long rq_weight; | |
447 | #endif | |
448 | #endif | |
449 | }; | |
450 | ||
451 | /* Real-Time classes' related field in a runqueue: */ | |
452 | struct rt_rq { | |
453 | struct rt_prio_array active; | |
454 | unsigned long rt_nr_running; | |
455 | #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED | |
456 | struct { | |
457 | int curr; /* highest queued rt task prio */ | |
458 | #ifdef CONFIG_SMP | |
459 | int next; /* next highest */ | |
460 | #endif | |
461 | } highest_prio; | |
462 | #endif | |
463 | #ifdef CONFIG_SMP | |
464 | unsigned long rt_nr_migratory; | |
465 | unsigned long rt_nr_total; | |
466 | int overloaded; | |
467 | struct plist_head pushable_tasks; | |
468 | #endif | |
469 | int rt_throttled; | |
470 | u64 rt_time; | |
471 | u64 rt_runtime; | |
472 | /* Nests inside the rq lock: */ | |
473 | spinlock_t rt_runtime_lock; | |
474 | ||
475 | #ifdef CONFIG_RT_GROUP_SCHED | |
476 | unsigned long rt_nr_boosted; | |
477 | ||
478 | struct rq *rq; | |
479 | struct list_head leaf_rt_rq_list; | |
480 | struct task_group *tg; | |
481 | struct sched_rt_entity *rt_se; | |
482 | #endif | |
483 | }; | |
484 | ||
485 | #ifdef CONFIG_SMP | |
486 | ||
487 | /* | |
488 | * We add the notion of a root-domain which will be used to define per-domain | |
489 | * variables. Each exclusive cpuset essentially defines an island domain by | |
490 | * fully partitioning the member cpus from any other cpuset. Whenever a new | |
491 | * exclusive cpuset is created, we also create and attach a new root-domain | |
492 | * object. | |
493 | * | |
494 | */ | |
495 | struct root_domain { | |
496 | atomic_t refcount; | |
497 | cpumask_var_t span; | |
498 | cpumask_var_t online; | |
499 | ||
500 | /* | |
501 | * The "RT overload" flag: it gets set if a CPU has more than | |
502 | * one runnable RT task. | |
503 | */ | |
504 | cpumask_var_t rto_mask; | |
505 | atomic_t rto_count; | |
506 | #ifdef CONFIG_SMP | |
507 | struct cpupri cpupri; | |
508 | #endif | |
509 | }; | |
510 | ||
511 | /* | |
512 | * By default the system creates a single root-domain with all cpus as | |
513 | * members (mimicking the global state we have today). | |
514 | */ | |
515 | static struct root_domain def_root_domain; | |
516 | ||
517 | #endif | |
518 | ||
519 | /* | |
520 | * This is the main, per-CPU runqueue data structure. | |
521 | * | |
522 | * Locking rule: those places that want to lock multiple runqueues | |
523 | * (such as the load balancing or the thread migration code), lock | |
524 | * acquire operations must be ordered by ascending &runqueue. | |
525 | */ | |
526 | struct rq { | |
527 | /* runqueue lock: */ | |
528 | spinlock_t lock; | |
529 | ||
530 | /* | |
531 | * nr_running and cpu_load should be in the same cacheline because | |
532 | * remote CPUs use both these fields when doing load calculation. | |
533 | */ | |
534 | unsigned long nr_running; | |
535 | #define CPU_LOAD_IDX_MAX 5 | |
536 | unsigned long cpu_load[CPU_LOAD_IDX_MAX]; | |
537 | #ifdef CONFIG_NO_HZ | |
538 | unsigned long last_tick_seen; | |
539 | unsigned char in_nohz_recently; | |
540 | #endif | |
541 | /* capture load from *all* tasks on this cpu: */ | |
542 | struct load_weight load; | |
543 | unsigned long nr_load_updates; | |
544 | u64 nr_switches; | |
545 | u64 nr_migrations_in; | |
546 | ||
547 | struct cfs_rq cfs; | |
548 | struct rt_rq rt; | |
549 | ||
550 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
551 | /* list of leaf cfs_rq on this cpu: */ | |
552 | struct list_head leaf_cfs_rq_list; | |
553 | #endif | |
554 | #ifdef CONFIG_RT_GROUP_SCHED | |
555 | struct list_head leaf_rt_rq_list; | |
556 | #endif | |
557 | ||
558 | /* | |
559 | * This is part of a global counter where only the total sum | |
560 | * over all CPUs matters. A task can increase this counter on | |
561 | * one CPU and if it got migrated afterwards it may decrease | |
562 | * it on another CPU. Always updated under the runqueue lock: | |
563 | */ | |
564 | unsigned long nr_uninterruptible; | |
565 | ||
566 | struct task_struct *curr, *idle; | |
567 | unsigned long next_balance; | |
568 | struct mm_struct *prev_mm; | |
569 | ||
570 | u64 clock; | |
571 | ||
572 | atomic_t nr_iowait; | |
573 | ||
574 | #ifdef CONFIG_SMP | |
575 | struct root_domain *rd; | |
576 | struct sched_domain *sd; | |
577 | ||
578 | unsigned char idle_at_tick; | |
579 | /* For active balancing */ | |
580 | int post_schedule; | |
581 | int active_balance; | |
582 | int push_cpu; | |
583 | /* cpu of this runqueue: */ | |
584 | int cpu; | |
585 | int online; | |
586 | ||
587 | unsigned long avg_load_per_task; | |
588 | ||
589 | struct task_struct *migration_thread; | |
590 | struct list_head migration_queue; | |
591 | ||
592 | u64 rt_avg; | |
593 | u64 age_stamp; | |
594 | #endif | |
595 | ||
596 | /* calc_load related fields */ | |
597 | unsigned long calc_load_update; | |
598 | long calc_load_active; | |
599 | ||
600 | #ifdef CONFIG_SCHED_HRTICK | |
601 | #ifdef CONFIG_SMP | |
602 | int hrtick_csd_pending; | |
603 | struct call_single_data hrtick_csd; | |
604 | #endif | |
605 | struct hrtimer hrtick_timer; | |
606 | #endif | |
607 | ||
608 | #ifdef CONFIG_SCHEDSTATS | |
609 | /* latency stats */ | |
610 | struct sched_info rq_sched_info; | |
611 | unsigned long long rq_cpu_time; | |
612 | /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */ | |
613 | ||
614 | /* sys_sched_yield() stats */ | |
615 | unsigned int yld_count; | |
616 | ||
617 | /* schedule() stats */ | |
618 | unsigned int sched_switch; | |
619 | unsigned int sched_count; | |
620 | unsigned int sched_goidle; | |
621 | ||
622 | /* try_to_wake_up() stats */ | |
623 | unsigned int ttwu_count; | |
624 | unsigned int ttwu_local; | |
625 | ||
626 | /* BKL stats */ | |
627 | unsigned int bkl_count; | |
628 | #endif | |
629 | }; | |
630 | ||
631 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); | |
632 | ||
633 | static inline | |
634 | void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) | |
635 | { | |
636 | rq->curr->sched_class->check_preempt_curr(rq, p, flags); | |
637 | } | |
638 | ||
639 | static inline int cpu_of(struct rq *rq) | |
640 | { | |
641 | #ifdef CONFIG_SMP | |
642 | return rq->cpu; | |
643 | #else | |
644 | return 0; | |
645 | #endif | |
646 | } | |
647 | ||
648 | /* | |
649 | * The domain tree (rq->sd) is protected by RCU's quiescent state transition. | |
650 | * See detach_destroy_domains: synchronize_sched for details. | |
651 | * | |
652 | * The domain tree of any CPU may only be accessed from within | |
653 | * preempt-disabled sections. | |
654 | */ | |
655 | #define for_each_domain(cpu, __sd) \ | |
656 | for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent) | |
657 | ||
658 | #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) | |
659 | #define this_rq() (&__get_cpu_var(runqueues)) | |
660 | #define task_rq(p) cpu_rq(task_cpu(p)) | |
661 | #define cpu_curr(cpu) (cpu_rq(cpu)->curr) | |
662 | #define raw_rq() (&__raw_get_cpu_var(runqueues)) | |
663 | ||
664 | inline void update_rq_clock(struct rq *rq) | |
665 | { | |
666 | rq->clock = sched_clock_cpu(cpu_of(rq)); | |
667 | } | |
668 | ||
669 | /* | |
670 | * Tunables that become constants when CONFIG_SCHED_DEBUG is off: | |
671 | */ | |
672 | #ifdef CONFIG_SCHED_DEBUG | |
673 | # define const_debug __read_mostly | |
674 | #else | |
675 | # define const_debug static const | |
676 | #endif | |
677 | ||
678 | /** | |
679 | * runqueue_is_locked | |
680 | * @cpu: the processor in question. | |
681 | * | |
682 | * Returns true if the current cpu runqueue is locked. | |
683 | * This interface allows printk to be called with the runqueue lock | |
684 | * held and know whether or not it is OK to wake up the klogd. | |
685 | */ | |
686 | int runqueue_is_locked(int cpu) | |
687 | { | |
688 | return spin_is_locked(&cpu_rq(cpu)->lock); | |
689 | } | |
690 | ||
691 | /* | |
692 | * Debugging: various feature bits | |
693 | */ | |
694 | ||
695 | #define SCHED_FEAT(name, enabled) \ | |
696 | __SCHED_FEAT_##name , | |
697 | ||
698 | enum { | |
699 | #include "sched_features.h" | |
700 | }; | |
701 | ||
702 | #undef SCHED_FEAT | |
703 | ||
704 | #define SCHED_FEAT(name, enabled) \ | |
705 | (1UL << __SCHED_FEAT_##name) * enabled | | |
706 | ||
707 | const_debug unsigned int sysctl_sched_features = | |
708 | #include "sched_features.h" | |
709 | 0; | |
710 | ||
711 | #undef SCHED_FEAT | |
712 | ||
713 | #ifdef CONFIG_SCHED_DEBUG | |
714 | #define SCHED_FEAT(name, enabled) \ | |
715 | #name , | |
716 | ||
717 | static __read_mostly char *sched_feat_names[] = { | |
718 | #include "sched_features.h" | |
719 | NULL | |
720 | }; | |
721 | ||
722 | #undef SCHED_FEAT | |
723 | ||
724 | static int sched_feat_show(struct seq_file *m, void *v) | |
725 | { | |
726 | int i; | |
727 | ||
728 | for (i = 0; sched_feat_names[i]; i++) { | |
729 | if (!(sysctl_sched_features & (1UL << i))) | |
730 | seq_puts(m, "NO_"); | |
731 | seq_printf(m, "%s ", sched_feat_names[i]); | |
732 | } | |
733 | seq_puts(m, "\n"); | |
734 | ||
735 | return 0; | |
736 | } | |
737 | ||
738 | static ssize_t | |
739 | sched_feat_write(struct file *filp, const char __user *ubuf, | |
740 | size_t cnt, loff_t *ppos) | |
741 | { | |
742 | char buf[64]; | |
743 | char *cmp = buf; | |
744 | int neg = 0; | |
745 | int i; | |
746 | ||
747 | if (cnt > 63) | |
748 | cnt = 63; | |
749 | ||
750 | if (copy_from_user(&buf, ubuf, cnt)) | |
751 | return -EFAULT; | |
752 | ||
753 | buf[cnt] = 0; | |
754 | ||
755 | if (strncmp(buf, "NO_", 3) == 0) { | |
756 | neg = 1; | |
757 | cmp += 3; | |
758 | } | |
759 | ||
760 | for (i = 0; sched_feat_names[i]; i++) { | |
761 | int len = strlen(sched_feat_names[i]); | |
762 | ||
763 | if (strncmp(cmp, sched_feat_names[i], len) == 0) { | |
764 | if (neg) | |
765 | sysctl_sched_features &= ~(1UL << i); | |
766 | else | |
767 | sysctl_sched_features |= (1UL << i); | |
768 | break; | |
769 | } | |
770 | } | |
771 | ||
772 | if (!sched_feat_names[i]) | |
773 | return -EINVAL; | |
774 | ||
775 | filp->f_pos += cnt; | |
776 | ||
777 | return cnt; | |
778 | } | |
779 | ||
780 | static int sched_feat_open(struct inode *inode, struct file *filp) | |
781 | { | |
782 | return single_open(filp, sched_feat_show, NULL); | |
783 | } | |
784 | ||
785 | static const struct file_operations sched_feat_fops = { | |
786 | .open = sched_feat_open, | |
787 | .write = sched_feat_write, | |
788 | .read = seq_read, | |
789 | .llseek = seq_lseek, | |
790 | .release = single_release, | |
791 | }; | |
792 | ||
793 | static __init int sched_init_debug(void) | |
794 | { | |
795 | debugfs_create_file("sched_features", 0644, NULL, NULL, | |
796 | &sched_feat_fops); | |
797 | ||
798 | return 0; | |
799 | } | |
800 | late_initcall(sched_init_debug); | |
801 | ||
802 | #endif | |
803 | ||
804 | #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x)) | |
805 | ||
806 | /* | |
807 | * Number of tasks to iterate in a single balance run. | |
808 | * Limited because this is done with IRQs disabled. | |
809 | */ | |
810 | const_debug unsigned int sysctl_sched_nr_migrate = 32; | |
811 | ||
812 | /* | |
813 | * ratelimit for updating the group shares. | |
814 | * default: 0.25ms | |
815 | */ | |
816 | unsigned int sysctl_sched_shares_ratelimit = 250000; | |
817 | ||
818 | /* | |
819 | * Inject some fuzzyness into changing the per-cpu group shares | |
820 | * this avoids remote rq-locks at the expense of fairness. | |
821 | * default: 4 | |
822 | */ | |
823 | unsigned int sysctl_sched_shares_thresh = 4; | |
824 | ||
825 | /* | |
826 | * period over which we average the RT time consumption, measured | |
827 | * in ms. | |
828 | * | |
829 | * default: 1s | |
830 | */ | |
831 | const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; | |
832 | ||
833 | /* | |
834 | * period over which we measure -rt task cpu usage in us. | |
835 | * default: 1s | |
836 | */ | |
837 | unsigned int sysctl_sched_rt_period = 1000000; | |
838 | ||
839 | static __read_mostly int scheduler_running; | |
840 | ||
841 | /* | |
842 | * part of the period that we allow rt tasks to run in us. | |
843 | * default: 0.95s | |
844 | */ | |
845 | int sysctl_sched_rt_runtime = 950000; | |
846 | ||
847 | static inline u64 global_rt_period(void) | |
848 | { | |
849 | return (u64)sysctl_sched_rt_period * NSEC_PER_USEC; | |
850 | } | |
851 | ||
852 | static inline u64 global_rt_runtime(void) | |
853 | { | |
854 | if (sysctl_sched_rt_runtime < 0) | |
855 | return RUNTIME_INF; | |
856 | ||
857 | return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC; | |
858 | } | |
859 | ||
860 | #ifndef prepare_arch_switch | |
861 | # define prepare_arch_switch(next) do { } while (0) | |
862 | #endif | |
863 | #ifndef finish_arch_switch | |
864 | # define finish_arch_switch(prev) do { } while (0) | |
865 | #endif | |
866 | ||
867 | static inline int task_current(struct rq *rq, struct task_struct *p) | |
868 | { | |
869 | return rq->curr == p; | |
870 | } | |
871 | ||
872 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW | |
873 | static inline int task_running(struct rq *rq, struct task_struct *p) | |
874 | { | |
875 | return task_current(rq, p); | |
876 | } | |
877 | ||
878 | static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) | |
879 | { | |
880 | } | |
881 | ||
882 | static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) | |
883 | { | |
884 | #ifdef CONFIG_DEBUG_SPINLOCK | |
885 | /* this is a valid case when another task releases the spinlock */ | |
886 | rq->lock.owner = current; | |
887 | #endif | |
888 | /* | |
889 | * If we are tracking spinlock dependencies then we have to | |
890 | * fix up the runqueue lock - which gets 'carried over' from | |
891 | * prev into current: | |
892 | */ | |
893 | spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_); | |
894 | ||
895 | spin_unlock_irq(&rq->lock); | |
896 | } | |
897 | ||
898 | #else /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
899 | static inline int task_running(struct rq *rq, struct task_struct *p) | |
900 | { | |
901 | #ifdef CONFIG_SMP | |
902 | return p->oncpu; | |
903 | #else | |
904 | return task_current(rq, p); | |
905 | #endif | |
906 | } | |
907 | ||
908 | static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) | |
909 | { | |
910 | #ifdef CONFIG_SMP | |
911 | /* | |
912 | * We can optimise this out completely for !SMP, because the | |
913 | * SMP rebalancing from interrupt is the only thing that cares | |
914 | * here. | |
915 | */ | |
916 | next->oncpu = 1; | |
917 | #endif | |
918 | #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
919 | spin_unlock_irq(&rq->lock); | |
920 | #else | |
921 | spin_unlock(&rq->lock); | |
922 | #endif | |
923 | } | |
924 | ||
925 | static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) | |
926 | { | |
927 | #ifdef CONFIG_SMP | |
928 | /* | |
929 | * After ->oncpu is cleared, the task can be moved to a different CPU. | |
930 | * We must ensure this doesn't happen until the switch is completely | |
931 | * finished. | |
932 | */ | |
933 | smp_wmb(); | |
934 | prev->oncpu = 0; | |
935 | #endif | |
936 | #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
937 | local_irq_enable(); | |
938 | #endif | |
939 | } | |
940 | #endif /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
941 | ||
942 | /* | |
943 | * __task_rq_lock - lock the runqueue a given task resides on. | |
944 | * Must be called interrupts disabled. | |
945 | */ | |
946 | static inline struct rq *__task_rq_lock(struct task_struct *p) | |
947 | __acquires(rq->lock) | |
948 | { | |
949 | for (;;) { | |
950 | struct rq *rq = task_rq(p); | |
951 | spin_lock(&rq->lock); | |
952 | if (likely(rq == task_rq(p))) | |
953 | return rq; | |
954 | spin_unlock(&rq->lock); | |
955 | } | |
956 | } | |
957 | ||
958 | /* | |
959 | * task_rq_lock - lock the runqueue a given task resides on and disable | |
960 | * interrupts. Note the ordering: we can safely lookup the task_rq without | |
961 | * explicitly disabling preemption. | |
962 | */ | |
963 | static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) | |
964 | __acquires(rq->lock) | |
965 | { | |
966 | struct rq *rq; | |
967 | ||
968 | for (;;) { | |
969 | local_irq_save(*flags); | |
970 | rq = task_rq(p); | |
971 | spin_lock(&rq->lock); | |
972 | if (likely(rq == task_rq(p))) | |
973 | return rq; | |
974 | spin_unlock_irqrestore(&rq->lock, *flags); | |
975 | } | |
976 | } | |
977 | ||
978 | void task_rq_unlock_wait(struct task_struct *p) | |
979 | { | |
980 | struct rq *rq = task_rq(p); | |
981 | ||
982 | smp_mb(); /* spin-unlock-wait is not a full memory barrier */ | |
983 | spin_unlock_wait(&rq->lock); | |
984 | } | |
985 | ||
986 | static void __task_rq_unlock(struct rq *rq) | |
987 | __releases(rq->lock) | |
988 | { | |
989 | spin_unlock(&rq->lock); | |
990 | } | |
991 | ||
992 | static inline void task_rq_unlock(struct rq *rq, unsigned long *flags) | |
993 | __releases(rq->lock) | |
994 | { | |
995 | spin_unlock_irqrestore(&rq->lock, *flags); | |
996 | } | |
997 | ||
998 | /* | |
999 | * this_rq_lock - lock this runqueue and disable interrupts. | |
1000 | */ | |
1001 | static struct rq *this_rq_lock(void) | |
1002 | __acquires(rq->lock) | |
1003 | { | |
1004 | struct rq *rq; | |
1005 | ||
1006 | local_irq_disable(); | |
1007 | rq = this_rq(); | |
1008 | spin_lock(&rq->lock); | |
1009 | ||
1010 | return rq; | |
1011 | } | |
1012 | ||
1013 | #ifdef CONFIG_SCHED_HRTICK | |
1014 | /* | |
1015 | * Use HR-timers to deliver accurate preemption points. | |
1016 | * | |
1017 | * Its all a bit involved since we cannot program an hrt while holding the | |
1018 | * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a | |
1019 | * reschedule event. | |
1020 | * | |
1021 | * When we get rescheduled we reprogram the hrtick_timer outside of the | |
1022 | * rq->lock. | |
1023 | */ | |
1024 | ||
1025 | /* | |
1026 | * Use hrtick when: | |
1027 | * - enabled by features | |
1028 | * - hrtimer is actually high res | |
1029 | */ | |
1030 | static inline int hrtick_enabled(struct rq *rq) | |
1031 | { | |
1032 | if (!sched_feat(HRTICK)) | |
1033 | return 0; | |
1034 | if (!cpu_active(cpu_of(rq))) | |
1035 | return 0; | |
1036 | return hrtimer_is_hres_active(&rq->hrtick_timer); | |
1037 | } | |
1038 | ||
1039 | static void hrtick_clear(struct rq *rq) | |
1040 | { | |
1041 | if (hrtimer_active(&rq->hrtick_timer)) | |
1042 | hrtimer_cancel(&rq->hrtick_timer); | |
1043 | } | |
1044 | ||
1045 | /* | |
1046 | * High-resolution timer tick. | |
1047 | * Runs from hardirq context with interrupts disabled. | |
1048 | */ | |
1049 | static enum hrtimer_restart hrtick(struct hrtimer *timer) | |
1050 | { | |
1051 | struct rq *rq = container_of(timer, struct rq, hrtick_timer); | |
1052 | ||
1053 | WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); | |
1054 | ||
1055 | spin_lock(&rq->lock); | |
1056 | update_rq_clock(rq); | |
1057 | rq->curr->sched_class->task_tick(rq, rq->curr, 1); | |
1058 | spin_unlock(&rq->lock); | |
1059 | ||
1060 | return HRTIMER_NORESTART; | |
1061 | } | |
1062 | ||
1063 | #ifdef CONFIG_SMP | |
1064 | /* | |
1065 | * called from hardirq (IPI) context | |
1066 | */ | |
1067 | static void __hrtick_start(void *arg) | |
1068 | { | |
1069 | struct rq *rq = arg; | |
1070 | ||
1071 | spin_lock(&rq->lock); | |
1072 | hrtimer_restart(&rq->hrtick_timer); | |
1073 | rq->hrtick_csd_pending = 0; | |
1074 | spin_unlock(&rq->lock); | |
1075 | } | |
1076 | ||
1077 | /* | |
1078 | * Called to set the hrtick timer state. | |
1079 | * | |
1080 | * called with rq->lock held and irqs disabled | |
1081 | */ | |
1082 | static void hrtick_start(struct rq *rq, u64 delay) | |
1083 | { | |
1084 | struct hrtimer *timer = &rq->hrtick_timer; | |
1085 | ktime_t time = ktime_add_ns(timer->base->get_time(), delay); | |
1086 | ||
1087 | hrtimer_set_expires(timer, time); | |
1088 | ||
1089 | if (rq == this_rq()) { | |
1090 | hrtimer_restart(timer); | |
1091 | } else if (!rq->hrtick_csd_pending) { | |
1092 | __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0); | |
1093 | rq->hrtick_csd_pending = 1; | |
1094 | } | |
1095 | } | |
1096 | ||
1097 | static int | |
1098 | hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) | |
1099 | { | |
1100 | int cpu = (int)(long)hcpu; | |
1101 | ||
1102 | switch (action) { | |
1103 | case CPU_UP_CANCELED: | |
1104 | case CPU_UP_CANCELED_FROZEN: | |
1105 | case CPU_DOWN_PREPARE: | |
1106 | case CPU_DOWN_PREPARE_FROZEN: | |
1107 | case CPU_DEAD: | |
1108 | case CPU_DEAD_FROZEN: | |
1109 | hrtick_clear(cpu_rq(cpu)); | |
1110 | return NOTIFY_OK; | |
1111 | } | |
1112 | ||
1113 | return NOTIFY_DONE; | |
1114 | } | |
1115 | ||
1116 | static __init void init_hrtick(void) | |
1117 | { | |
1118 | hotcpu_notifier(hotplug_hrtick, 0); | |
1119 | } | |
1120 | #else | |
1121 | /* | |
1122 | * Called to set the hrtick timer state. | |
1123 | * | |
1124 | * called with rq->lock held and irqs disabled | |
1125 | */ | |
1126 | static void hrtick_start(struct rq *rq, u64 delay) | |
1127 | { | |
1128 | __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, | |
1129 | HRTIMER_MODE_REL_PINNED, 0); | |
1130 | } | |
1131 | ||
1132 | static inline void init_hrtick(void) | |
1133 | { | |
1134 | } | |
1135 | #endif /* CONFIG_SMP */ | |
1136 | ||
1137 | static void init_rq_hrtick(struct rq *rq) | |
1138 | { | |
1139 | #ifdef CONFIG_SMP | |
1140 | rq->hrtick_csd_pending = 0; | |
1141 | ||
1142 | rq->hrtick_csd.flags = 0; | |
1143 | rq->hrtick_csd.func = __hrtick_start; | |
1144 | rq->hrtick_csd.info = rq; | |
1145 | #endif | |
1146 | ||
1147 | hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
1148 | rq->hrtick_timer.function = hrtick; | |
1149 | } | |
1150 | #else /* CONFIG_SCHED_HRTICK */ | |
1151 | static inline void hrtick_clear(struct rq *rq) | |
1152 | { | |
1153 | } | |
1154 | ||
1155 | static inline void init_rq_hrtick(struct rq *rq) | |
1156 | { | |
1157 | } | |
1158 | ||
1159 | static inline void init_hrtick(void) | |
1160 | { | |
1161 | } | |
1162 | #endif /* CONFIG_SCHED_HRTICK */ | |
1163 | ||
1164 | /* | |
1165 | * resched_task - mark a task 'to be rescheduled now'. | |
1166 | * | |
1167 | * On UP this means the setting of the need_resched flag, on SMP it | |
1168 | * might also involve a cross-CPU call to trigger the scheduler on | |
1169 | * the target CPU. | |
1170 | */ | |
1171 | #ifdef CONFIG_SMP | |
1172 | ||
1173 | #ifndef tsk_is_polling | |
1174 | #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG) | |
1175 | #endif | |
1176 | ||
1177 | static void resched_task(struct task_struct *p) | |
1178 | { | |
1179 | int cpu; | |
1180 | ||
1181 | assert_spin_locked(&task_rq(p)->lock); | |
1182 | ||
1183 | if (test_tsk_need_resched(p)) | |
1184 | return; | |
1185 | ||
1186 | set_tsk_need_resched(p); | |
1187 | ||
1188 | cpu = task_cpu(p); | |
1189 | if (cpu == smp_processor_id()) | |
1190 | return; | |
1191 | ||
1192 | /* NEED_RESCHED must be visible before we test polling */ | |
1193 | smp_mb(); | |
1194 | if (!tsk_is_polling(p)) | |
1195 | smp_send_reschedule(cpu); | |
1196 | } | |
1197 | ||
1198 | static void resched_cpu(int cpu) | |
1199 | { | |
1200 | struct rq *rq = cpu_rq(cpu); | |
1201 | unsigned long flags; | |
1202 | ||
1203 | if (!spin_trylock_irqsave(&rq->lock, flags)) | |
1204 | return; | |
1205 | resched_task(cpu_curr(cpu)); | |
1206 | spin_unlock_irqrestore(&rq->lock, flags); | |
1207 | } | |
1208 | ||
1209 | #ifdef CONFIG_NO_HZ | |
1210 | /* | |
1211 | * When add_timer_on() enqueues a timer into the timer wheel of an | |
1212 | * idle CPU then this timer might expire before the next timer event | |
1213 | * which is scheduled to wake up that CPU. In case of a completely | |
1214 | * idle system the next event might even be infinite time into the | |
1215 | * future. wake_up_idle_cpu() ensures that the CPU is woken up and | |
1216 | * leaves the inner idle loop so the newly added timer is taken into | |
1217 | * account when the CPU goes back to idle and evaluates the timer | |
1218 | * wheel for the next timer event. | |
1219 | */ | |
1220 | void wake_up_idle_cpu(int cpu) | |
1221 | { | |
1222 | struct rq *rq = cpu_rq(cpu); | |
1223 | ||
1224 | if (cpu == smp_processor_id()) | |
1225 | return; | |
1226 | ||
1227 | /* | |
1228 | * This is safe, as this function is called with the timer | |
1229 | * wheel base lock of (cpu) held. When the CPU is on the way | |
1230 | * to idle and has not yet set rq->curr to idle then it will | |
1231 | * be serialized on the timer wheel base lock and take the new | |
1232 | * timer into account automatically. | |
1233 | */ | |
1234 | if (rq->curr != rq->idle) | |
1235 | return; | |
1236 | ||
1237 | /* | |
1238 | * We can set TIF_RESCHED on the idle task of the other CPU | |
1239 | * lockless. The worst case is that the other CPU runs the | |
1240 | * idle task through an additional NOOP schedule() | |
1241 | */ | |
1242 | set_tsk_need_resched(rq->idle); | |
1243 | ||
1244 | /* NEED_RESCHED must be visible before we test polling */ | |
1245 | smp_mb(); | |
1246 | if (!tsk_is_polling(rq->idle)) | |
1247 | smp_send_reschedule(cpu); | |
1248 | } | |
1249 | #endif /* CONFIG_NO_HZ */ | |
1250 | ||
1251 | static u64 sched_avg_period(void) | |
1252 | { | |
1253 | return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2; | |
1254 | } | |
1255 | ||
1256 | static void sched_avg_update(struct rq *rq) | |
1257 | { | |
1258 | s64 period = sched_avg_period(); | |
1259 | ||
1260 | while ((s64)(rq->clock - rq->age_stamp) > period) { | |
1261 | rq->age_stamp += period; | |
1262 | rq->rt_avg /= 2; | |
1263 | } | |
1264 | } | |
1265 | ||
1266 | static void sched_rt_avg_update(struct rq *rq, u64 rt_delta) | |
1267 | { | |
1268 | rq->rt_avg += rt_delta; | |
1269 | sched_avg_update(rq); | |
1270 | } | |
1271 | ||
1272 | #else /* !CONFIG_SMP */ | |
1273 | static void resched_task(struct task_struct *p) | |
1274 | { | |
1275 | assert_spin_locked(&task_rq(p)->lock); | |
1276 | set_tsk_need_resched(p); | |
1277 | } | |
1278 | ||
1279 | static void sched_rt_avg_update(struct rq *rq, u64 rt_delta) | |
1280 | { | |
1281 | } | |
1282 | #endif /* CONFIG_SMP */ | |
1283 | ||
1284 | #if BITS_PER_LONG == 32 | |
1285 | # define WMULT_CONST (~0UL) | |
1286 | #else | |
1287 | # define WMULT_CONST (1UL << 32) | |
1288 | #endif | |
1289 | ||
1290 | #define WMULT_SHIFT 32 | |
1291 | ||
1292 | /* | |
1293 | * Shift right and round: | |
1294 | */ | |
1295 | #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y)) | |
1296 | ||
1297 | /* | |
1298 | * delta *= weight / lw | |
1299 | */ | |
1300 | static unsigned long | |
1301 | calc_delta_mine(unsigned long delta_exec, unsigned long weight, | |
1302 | struct load_weight *lw) | |
1303 | { | |
1304 | u64 tmp; | |
1305 | ||
1306 | if (!lw->inv_weight) { | |
1307 | if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST)) | |
1308 | lw->inv_weight = 1; | |
1309 | else | |
1310 | lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2) | |
1311 | / (lw->weight+1); | |
1312 | } | |
1313 | ||
1314 | tmp = (u64)delta_exec * weight; | |
1315 | /* | |
1316 | * Check whether we'd overflow the 64-bit multiplication: | |
1317 | */ | |
1318 | if (unlikely(tmp > WMULT_CONST)) | |
1319 | tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight, | |
1320 | WMULT_SHIFT/2); | |
1321 | else | |
1322 | tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT); | |
1323 | ||
1324 | return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX); | |
1325 | } | |
1326 | ||
1327 | static inline void update_load_add(struct load_weight *lw, unsigned long inc) | |
1328 | { | |
1329 | lw->weight += inc; | |
1330 | lw->inv_weight = 0; | |
1331 | } | |
1332 | ||
1333 | static inline void update_load_sub(struct load_weight *lw, unsigned long dec) | |
1334 | { | |
1335 | lw->weight -= dec; | |
1336 | lw->inv_weight = 0; | |
1337 | } | |
1338 | ||
1339 | /* | |
1340 | * To aid in avoiding the subversion of "niceness" due to uneven distribution | |
1341 | * of tasks with abnormal "nice" values across CPUs the contribution that | |
1342 | * each task makes to its run queue's load is weighted according to its | |
1343 | * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a | |
1344 | * scaled version of the new time slice allocation that they receive on time | |
1345 | * slice expiry etc. | |
1346 | */ | |
1347 | ||
1348 | #define WEIGHT_IDLEPRIO 3 | |
1349 | #define WMULT_IDLEPRIO 1431655765 | |
1350 | ||
1351 | /* | |
1352 | * Nice levels are multiplicative, with a gentle 10% change for every | |
1353 | * nice level changed. I.e. when a CPU-bound task goes from nice 0 to | |
1354 | * nice 1, it will get ~10% less CPU time than another CPU-bound task | |
1355 | * that remained on nice 0. | |
1356 | * | |
1357 | * The "10% effect" is relative and cumulative: from _any_ nice level, | |
1358 | * if you go up 1 level, it's -10% CPU usage, if you go down 1 level | |
1359 | * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25. | |
1360 | * If a task goes up by ~10% and another task goes down by ~10% then | |
1361 | * the relative distance between them is ~25%.) | |
1362 | */ | |
1363 | static const int prio_to_weight[40] = { | |
1364 | /* -20 */ 88761, 71755, 56483, 46273, 36291, | |
1365 | /* -15 */ 29154, 23254, 18705, 14949, 11916, | |
1366 | /* -10 */ 9548, 7620, 6100, 4904, 3906, | |
1367 | /* -5 */ 3121, 2501, 1991, 1586, 1277, | |
1368 | /* 0 */ 1024, 820, 655, 526, 423, | |
1369 | /* 5 */ 335, 272, 215, 172, 137, | |
1370 | /* 10 */ 110, 87, 70, 56, 45, | |
1371 | /* 15 */ 36, 29, 23, 18, 15, | |
1372 | }; | |
1373 | ||
1374 | /* | |
1375 | * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated. | |
1376 | * | |
1377 | * In cases where the weight does not change often, we can use the | |
1378 | * precalculated inverse to speed up arithmetics by turning divisions | |
1379 | * into multiplications: | |
1380 | */ | |
1381 | static const u32 prio_to_wmult[40] = { | |
1382 | /* -20 */ 48388, 59856, 76040, 92818, 118348, | |
1383 | /* -15 */ 147320, 184698, 229616, 287308, 360437, | |
1384 | /* -10 */ 449829, 563644, 704093, 875809, 1099582, | |
1385 | /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326, | |
1386 | /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587, | |
1387 | /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126, | |
1388 | /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717, | |
1389 | /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153, | |
1390 | }; | |
1391 | ||
1392 | static void activate_task(struct rq *rq, struct task_struct *p, int wakeup); | |
1393 | ||
1394 | /* | |
1395 | * runqueue iterator, to support SMP load-balancing between different | |
1396 | * scheduling classes, without having to expose their internal data | |
1397 | * structures to the load-balancing proper: | |
1398 | */ | |
1399 | struct rq_iterator { | |
1400 | void *arg; | |
1401 | struct task_struct *(*start)(void *); | |
1402 | struct task_struct *(*next)(void *); | |
1403 | }; | |
1404 | ||
1405 | #ifdef CONFIG_SMP | |
1406 | static unsigned long | |
1407 | balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1408 | unsigned long max_load_move, struct sched_domain *sd, | |
1409 | enum cpu_idle_type idle, int *all_pinned, | |
1410 | int *this_best_prio, struct rq_iterator *iterator); | |
1411 | ||
1412 | static int | |
1413 | iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1414 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1415 | struct rq_iterator *iterator); | |
1416 | #endif | |
1417 | ||
1418 | /* Time spent by the tasks of the cpu accounting group executing in ... */ | |
1419 | enum cpuacct_stat_index { | |
1420 | CPUACCT_STAT_USER, /* ... user mode */ | |
1421 | CPUACCT_STAT_SYSTEM, /* ... kernel mode */ | |
1422 | ||
1423 | CPUACCT_STAT_NSTATS, | |
1424 | }; | |
1425 | ||
1426 | #ifdef CONFIG_CGROUP_CPUACCT | |
1427 | static void cpuacct_charge(struct task_struct *tsk, u64 cputime); | |
1428 | static void cpuacct_update_stats(struct task_struct *tsk, | |
1429 | enum cpuacct_stat_index idx, cputime_t val); | |
1430 | #else | |
1431 | static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {} | |
1432 | static inline void cpuacct_update_stats(struct task_struct *tsk, | |
1433 | enum cpuacct_stat_index idx, cputime_t val) {} | |
1434 | #endif | |
1435 | ||
1436 | static inline void inc_cpu_load(struct rq *rq, unsigned long load) | |
1437 | { | |
1438 | update_load_add(&rq->load, load); | |
1439 | } | |
1440 | ||
1441 | static inline void dec_cpu_load(struct rq *rq, unsigned long load) | |
1442 | { | |
1443 | update_load_sub(&rq->load, load); | |
1444 | } | |
1445 | ||
1446 | #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED) | |
1447 | typedef int (*tg_visitor)(struct task_group *, void *); | |
1448 | ||
1449 | /* | |
1450 | * Iterate the full tree, calling @down when first entering a node and @up when | |
1451 | * leaving it for the final time. | |
1452 | */ | |
1453 | static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data) | |
1454 | { | |
1455 | struct task_group *parent, *child; | |
1456 | int ret; | |
1457 | ||
1458 | rcu_read_lock(); | |
1459 | parent = &root_task_group; | |
1460 | down: | |
1461 | ret = (*down)(parent, data); | |
1462 | if (ret) | |
1463 | goto out_unlock; | |
1464 | list_for_each_entry_rcu(child, &parent->children, siblings) { | |
1465 | parent = child; | |
1466 | goto down; | |
1467 | ||
1468 | up: | |
1469 | continue; | |
1470 | } | |
1471 | ret = (*up)(parent, data); | |
1472 | if (ret) | |
1473 | goto out_unlock; | |
1474 | ||
1475 | child = parent; | |
1476 | parent = parent->parent; | |
1477 | if (parent) | |
1478 | goto up; | |
1479 | out_unlock: | |
1480 | rcu_read_unlock(); | |
1481 | ||
1482 | return ret; | |
1483 | } | |
1484 | ||
1485 | static int tg_nop(struct task_group *tg, void *data) | |
1486 | { | |
1487 | return 0; | |
1488 | } | |
1489 | #endif | |
1490 | ||
1491 | #ifdef CONFIG_SMP | |
1492 | /* Used instead of source_load when we know the type == 0 */ | |
1493 | static unsigned long weighted_cpuload(const int cpu) | |
1494 | { | |
1495 | return cpu_rq(cpu)->load.weight; | |
1496 | } | |
1497 | ||
1498 | /* | |
1499 | * Return a low guess at the load of a migration-source cpu weighted | |
1500 | * according to the scheduling class and "nice" value. | |
1501 | * | |
1502 | * We want to under-estimate the load of migration sources, to | |
1503 | * balance conservatively. | |
1504 | */ | |
1505 | static unsigned long source_load(int cpu, int type) | |
1506 | { | |
1507 | struct rq *rq = cpu_rq(cpu); | |
1508 | unsigned long total = weighted_cpuload(cpu); | |
1509 | ||
1510 | if (type == 0 || !sched_feat(LB_BIAS)) | |
1511 | return total; | |
1512 | ||
1513 | return min(rq->cpu_load[type-1], total); | |
1514 | } | |
1515 | ||
1516 | /* | |
1517 | * Return a high guess at the load of a migration-target cpu weighted | |
1518 | * according to the scheduling class and "nice" value. | |
1519 | */ | |
1520 | static unsigned long target_load(int cpu, int type) | |
1521 | { | |
1522 | struct rq *rq = cpu_rq(cpu); | |
1523 | unsigned long total = weighted_cpuload(cpu); | |
1524 | ||
1525 | if (type == 0 || !sched_feat(LB_BIAS)) | |
1526 | return total; | |
1527 | ||
1528 | return max(rq->cpu_load[type-1], total); | |
1529 | } | |
1530 | ||
1531 | static struct sched_group *group_of(int cpu) | |
1532 | { | |
1533 | struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd); | |
1534 | ||
1535 | if (!sd) | |
1536 | return NULL; | |
1537 | ||
1538 | return sd->groups; | |
1539 | } | |
1540 | ||
1541 | static unsigned long power_of(int cpu) | |
1542 | { | |
1543 | struct sched_group *group = group_of(cpu); | |
1544 | ||
1545 | if (!group) | |
1546 | return SCHED_LOAD_SCALE; | |
1547 | ||
1548 | return group->cpu_power; | |
1549 | } | |
1550 | ||
1551 | static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd); | |
1552 | ||
1553 | static unsigned long cpu_avg_load_per_task(int cpu) | |
1554 | { | |
1555 | struct rq *rq = cpu_rq(cpu); | |
1556 | unsigned long nr_running = ACCESS_ONCE(rq->nr_running); | |
1557 | ||
1558 | if (nr_running) | |
1559 | rq->avg_load_per_task = rq->load.weight / nr_running; | |
1560 | else | |
1561 | rq->avg_load_per_task = 0; | |
1562 | ||
1563 | return rq->avg_load_per_task; | |
1564 | } | |
1565 | ||
1566 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1567 | ||
1568 | static __read_mostly unsigned long *update_shares_data; | |
1569 | ||
1570 | static void __set_se_shares(struct sched_entity *se, unsigned long shares); | |
1571 | ||
1572 | /* | |
1573 | * Calculate and set the cpu's group shares. | |
1574 | */ | |
1575 | static void update_group_shares_cpu(struct task_group *tg, int cpu, | |
1576 | unsigned long sd_shares, | |
1577 | unsigned long sd_rq_weight, | |
1578 | unsigned long *usd_rq_weight) | |
1579 | { | |
1580 | unsigned long shares, rq_weight; | |
1581 | int boost = 0; | |
1582 | ||
1583 | rq_weight = usd_rq_weight[cpu]; | |
1584 | if (!rq_weight) { | |
1585 | boost = 1; | |
1586 | rq_weight = NICE_0_LOAD; | |
1587 | } | |
1588 | ||
1589 | /* | |
1590 | * \Sum_j shares_j * rq_weight_i | |
1591 | * shares_i = ----------------------------- | |
1592 | * \Sum_j rq_weight_j | |
1593 | */ | |
1594 | shares = (sd_shares * rq_weight) / sd_rq_weight; | |
1595 | shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES); | |
1596 | ||
1597 | if (abs(shares - tg->se[cpu]->load.weight) > | |
1598 | sysctl_sched_shares_thresh) { | |
1599 | struct rq *rq = cpu_rq(cpu); | |
1600 | unsigned long flags; | |
1601 | ||
1602 | spin_lock_irqsave(&rq->lock, flags); | |
1603 | tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight; | |
1604 | tg->cfs_rq[cpu]->shares = boost ? 0 : shares; | |
1605 | __set_se_shares(tg->se[cpu], shares); | |
1606 | spin_unlock_irqrestore(&rq->lock, flags); | |
1607 | } | |
1608 | } | |
1609 | ||
1610 | /* | |
1611 | * Re-compute the task group their per cpu shares over the given domain. | |
1612 | * This needs to be done in a bottom-up fashion because the rq weight of a | |
1613 | * parent group depends on the shares of its child groups. | |
1614 | */ | |
1615 | static int tg_shares_up(struct task_group *tg, void *data) | |
1616 | { | |
1617 | unsigned long weight, rq_weight = 0, shares = 0; | |
1618 | unsigned long *usd_rq_weight; | |
1619 | struct sched_domain *sd = data; | |
1620 | unsigned long flags; | |
1621 | int i; | |
1622 | ||
1623 | if (!tg->se[0]) | |
1624 | return 0; | |
1625 | ||
1626 | local_irq_save(flags); | |
1627 | usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id()); | |
1628 | ||
1629 | for_each_cpu(i, sched_domain_span(sd)) { | |
1630 | weight = tg->cfs_rq[i]->load.weight; | |
1631 | usd_rq_weight[i] = weight; | |
1632 | ||
1633 | /* | |
1634 | * If there are currently no tasks on the cpu pretend there | |
1635 | * is one of average load so that when a new task gets to | |
1636 | * run here it will not get delayed by group starvation. | |
1637 | */ | |
1638 | if (!weight) | |
1639 | weight = NICE_0_LOAD; | |
1640 | ||
1641 | rq_weight += weight; | |
1642 | shares += tg->cfs_rq[i]->shares; | |
1643 | } | |
1644 | ||
1645 | if ((!shares && rq_weight) || shares > tg->shares) | |
1646 | shares = tg->shares; | |
1647 | ||
1648 | if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE)) | |
1649 | shares = tg->shares; | |
1650 | ||
1651 | for_each_cpu(i, sched_domain_span(sd)) | |
1652 | update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight); | |
1653 | ||
1654 | local_irq_restore(flags); | |
1655 | ||
1656 | return 0; | |
1657 | } | |
1658 | ||
1659 | /* | |
1660 | * Compute the cpu's hierarchical load factor for each task group. | |
1661 | * This needs to be done in a top-down fashion because the load of a child | |
1662 | * group is a fraction of its parents load. | |
1663 | */ | |
1664 | static int tg_load_down(struct task_group *tg, void *data) | |
1665 | { | |
1666 | unsigned long load; | |
1667 | long cpu = (long)data; | |
1668 | ||
1669 | if (!tg->parent) { | |
1670 | load = cpu_rq(cpu)->load.weight; | |
1671 | } else { | |
1672 | load = tg->parent->cfs_rq[cpu]->h_load; | |
1673 | load *= tg->cfs_rq[cpu]->shares; | |
1674 | load /= tg->parent->cfs_rq[cpu]->load.weight + 1; | |
1675 | } | |
1676 | ||
1677 | tg->cfs_rq[cpu]->h_load = load; | |
1678 | ||
1679 | return 0; | |
1680 | } | |
1681 | ||
1682 | static void update_shares(struct sched_domain *sd) | |
1683 | { | |
1684 | s64 elapsed; | |
1685 | u64 now; | |
1686 | ||
1687 | if (root_task_group_empty()) | |
1688 | return; | |
1689 | ||
1690 | now = cpu_clock(raw_smp_processor_id()); | |
1691 | elapsed = now - sd->last_update; | |
1692 | ||
1693 | if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) { | |
1694 | sd->last_update = now; | |
1695 | walk_tg_tree(tg_nop, tg_shares_up, sd); | |
1696 | } | |
1697 | } | |
1698 | ||
1699 | static void update_shares_locked(struct rq *rq, struct sched_domain *sd) | |
1700 | { | |
1701 | if (root_task_group_empty()) | |
1702 | return; | |
1703 | ||
1704 | spin_unlock(&rq->lock); | |
1705 | update_shares(sd); | |
1706 | spin_lock(&rq->lock); | |
1707 | } | |
1708 | ||
1709 | static void update_h_load(long cpu) | |
1710 | { | |
1711 | if (root_task_group_empty()) | |
1712 | return; | |
1713 | ||
1714 | walk_tg_tree(tg_load_down, tg_nop, (void *)cpu); | |
1715 | } | |
1716 | ||
1717 | #else | |
1718 | ||
1719 | static inline void update_shares(struct sched_domain *sd) | |
1720 | { | |
1721 | } | |
1722 | ||
1723 | static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd) | |
1724 | { | |
1725 | } | |
1726 | ||
1727 | #endif | |
1728 | ||
1729 | #ifdef CONFIG_PREEMPT | |
1730 | ||
1731 | static void double_rq_lock(struct rq *rq1, struct rq *rq2); | |
1732 | ||
1733 | /* | |
1734 | * fair double_lock_balance: Safely acquires both rq->locks in a fair | |
1735 | * way at the expense of forcing extra atomic operations in all | |
1736 | * invocations. This assures that the double_lock is acquired using the | |
1737 | * same underlying policy as the spinlock_t on this architecture, which | |
1738 | * reduces latency compared to the unfair variant below. However, it | |
1739 | * also adds more overhead and therefore may reduce throughput. | |
1740 | */ | |
1741 | static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest) | |
1742 | __releases(this_rq->lock) | |
1743 | __acquires(busiest->lock) | |
1744 | __acquires(this_rq->lock) | |
1745 | { | |
1746 | spin_unlock(&this_rq->lock); | |
1747 | double_rq_lock(this_rq, busiest); | |
1748 | ||
1749 | return 1; | |
1750 | } | |
1751 | ||
1752 | #else | |
1753 | /* | |
1754 | * Unfair double_lock_balance: Optimizes throughput at the expense of | |
1755 | * latency by eliminating extra atomic operations when the locks are | |
1756 | * already in proper order on entry. This favors lower cpu-ids and will | |
1757 | * grant the double lock to lower cpus over higher ids under contention, | |
1758 | * regardless of entry order into the function. | |
1759 | */ | |
1760 | static int _double_lock_balance(struct rq *this_rq, struct rq *busiest) | |
1761 | __releases(this_rq->lock) | |
1762 | __acquires(busiest->lock) | |
1763 | __acquires(this_rq->lock) | |
1764 | { | |
1765 | int ret = 0; | |
1766 | ||
1767 | if (unlikely(!spin_trylock(&busiest->lock))) { | |
1768 | if (busiest < this_rq) { | |
1769 | spin_unlock(&this_rq->lock); | |
1770 | spin_lock(&busiest->lock); | |
1771 | spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING); | |
1772 | ret = 1; | |
1773 | } else | |
1774 | spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING); | |
1775 | } | |
1776 | return ret; | |
1777 | } | |
1778 | ||
1779 | #endif /* CONFIG_PREEMPT */ | |
1780 | ||
1781 | /* | |
1782 | * double_lock_balance - lock the busiest runqueue, this_rq is locked already. | |
1783 | */ | |
1784 | static int double_lock_balance(struct rq *this_rq, struct rq *busiest) | |
1785 | { | |
1786 | if (unlikely(!irqs_disabled())) { | |
1787 | /* printk() doesn't work good under rq->lock */ | |
1788 | spin_unlock(&this_rq->lock); | |
1789 | BUG_ON(1); | |
1790 | } | |
1791 | ||
1792 | return _double_lock_balance(this_rq, busiest); | |
1793 | } | |
1794 | ||
1795 | static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest) | |
1796 | __releases(busiest->lock) | |
1797 | { | |
1798 | spin_unlock(&busiest->lock); | |
1799 | lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_); | |
1800 | } | |
1801 | #endif | |
1802 | ||
1803 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1804 | static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares) | |
1805 | { | |
1806 | #ifdef CONFIG_SMP | |
1807 | cfs_rq->shares = shares; | |
1808 | #endif | |
1809 | } | |
1810 | #endif | |
1811 | ||
1812 | static void calc_load_account_active(struct rq *this_rq); | |
1813 | ||
1814 | #include "sched_stats.h" | |
1815 | #include "sched_idletask.c" | |
1816 | #include "sched_fair.c" | |
1817 | #include "sched_rt.c" | |
1818 | #ifdef CONFIG_SCHED_DEBUG | |
1819 | # include "sched_debug.c" | |
1820 | #endif | |
1821 | ||
1822 | #define sched_class_highest (&rt_sched_class) | |
1823 | #define for_each_class(class) \ | |
1824 | for (class = sched_class_highest; class; class = class->next) | |
1825 | ||
1826 | static void inc_nr_running(struct rq *rq) | |
1827 | { | |
1828 | rq->nr_running++; | |
1829 | } | |
1830 | ||
1831 | static void dec_nr_running(struct rq *rq) | |
1832 | { | |
1833 | rq->nr_running--; | |
1834 | } | |
1835 | ||
1836 | static void set_load_weight(struct task_struct *p) | |
1837 | { | |
1838 | if (task_has_rt_policy(p)) { | |
1839 | p->se.load.weight = prio_to_weight[0] * 2; | |
1840 | p->se.load.inv_weight = prio_to_wmult[0] >> 1; | |
1841 | return; | |
1842 | } | |
1843 | ||
1844 | /* | |
1845 | * SCHED_IDLE tasks get minimal weight: | |
1846 | */ | |
1847 | if (p->policy == SCHED_IDLE) { | |
1848 | p->se.load.weight = WEIGHT_IDLEPRIO; | |
1849 | p->se.load.inv_weight = WMULT_IDLEPRIO; | |
1850 | return; | |
1851 | } | |
1852 | ||
1853 | p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO]; | |
1854 | p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO]; | |
1855 | } | |
1856 | ||
1857 | static void update_avg(u64 *avg, u64 sample) | |
1858 | { | |
1859 | s64 diff = sample - *avg; | |
1860 | *avg += diff >> 3; | |
1861 | } | |
1862 | ||
1863 | static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup) | |
1864 | { | |
1865 | if (wakeup) | |
1866 | p->se.start_runtime = p->se.sum_exec_runtime; | |
1867 | ||
1868 | sched_info_queued(p); | |
1869 | p->sched_class->enqueue_task(rq, p, wakeup); | |
1870 | p->se.on_rq = 1; | |
1871 | } | |
1872 | ||
1873 | static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep) | |
1874 | { | |
1875 | if (sleep) { | |
1876 | if (p->se.last_wakeup) { | |
1877 | update_avg(&p->se.avg_overlap, | |
1878 | p->se.sum_exec_runtime - p->se.last_wakeup); | |
1879 | p->se.last_wakeup = 0; | |
1880 | } else { | |
1881 | update_avg(&p->se.avg_wakeup, | |
1882 | sysctl_sched_wakeup_granularity); | |
1883 | } | |
1884 | } | |
1885 | ||
1886 | sched_info_dequeued(p); | |
1887 | p->sched_class->dequeue_task(rq, p, sleep); | |
1888 | p->se.on_rq = 0; | |
1889 | } | |
1890 | ||
1891 | /* | |
1892 | * __normal_prio - return the priority that is based on the static prio | |
1893 | */ | |
1894 | static inline int __normal_prio(struct task_struct *p) | |
1895 | { | |
1896 | return p->static_prio; | |
1897 | } | |
1898 | ||
1899 | /* | |
1900 | * Calculate the expected normal priority: i.e. priority | |
1901 | * without taking RT-inheritance into account. Might be | |
1902 | * boosted by interactivity modifiers. Changes upon fork, | |
1903 | * setprio syscalls, and whenever the interactivity | |
1904 | * estimator recalculates. | |
1905 | */ | |
1906 | static inline int normal_prio(struct task_struct *p) | |
1907 | { | |
1908 | int prio; | |
1909 | ||
1910 | if (task_has_rt_policy(p)) | |
1911 | prio = MAX_RT_PRIO-1 - p->rt_priority; | |
1912 | else | |
1913 | prio = __normal_prio(p); | |
1914 | return prio; | |
1915 | } | |
1916 | ||
1917 | /* | |
1918 | * Calculate the current priority, i.e. the priority | |
1919 | * taken into account by the scheduler. This value might | |
1920 | * be boosted by RT tasks, or might be boosted by | |
1921 | * interactivity modifiers. Will be RT if the task got | |
1922 | * RT-boosted. If not then it returns p->normal_prio. | |
1923 | */ | |
1924 | static int effective_prio(struct task_struct *p) | |
1925 | { | |
1926 | p->normal_prio = normal_prio(p); | |
1927 | /* | |
1928 | * If we are RT tasks or we were boosted to RT priority, | |
1929 | * keep the priority unchanged. Otherwise, update priority | |
1930 | * to the normal priority: | |
1931 | */ | |
1932 | if (!rt_prio(p->prio)) | |
1933 | return p->normal_prio; | |
1934 | return p->prio; | |
1935 | } | |
1936 | ||
1937 | /* | |
1938 | * activate_task - move a task to the runqueue. | |
1939 | */ | |
1940 | static void activate_task(struct rq *rq, struct task_struct *p, int wakeup) | |
1941 | { | |
1942 | if (task_contributes_to_load(p)) | |
1943 | rq->nr_uninterruptible--; | |
1944 | ||
1945 | enqueue_task(rq, p, wakeup); | |
1946 | inc_nr_running(rq); | |
1947 | } | |
1948 | ||
1949 | /* | |
1950 | * deactivate_task - remove a task from the runqueue. | |
1951 | */ | |
1952 | static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep) | |
1953 | { | |
1954 | if (task_contributes_to_load(p)) | |
1955 | rq->nr_uninterruptible++; | |
1956 | ||
1957 | dequeue_task(rq, p, sleep); | |
1958 | dec_nr_running(rq); | |
1959 | } | |
1960 | ||
1961 | /** | |
1962 | * task_curr - is this task currently executing on a CPU? | |
1963 | * @p: the task in question. | |
1964 | */ | |
1965 | inline int task_curr(const struct task_struct *p) | |
1966 | { | |
1967 | return cpu_curr(task_cpu(p)) == p; | |
1968 | } | |
1969 | ||
1970 | static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu) | |
1971 | { | |
1972 | set_task_rq(p, cpu); | |
1973 | #ifdef CONFIG_SMP | |
1974 | /* | |
1975 | * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be | |
1976 | * successfuly executed on another CPU. We must ensure that updates of | |
1977 | * per-task data have been completed by this moment. | |
1978 | */ | |
1979 | smp_wmb(); | |
1980 | task_thread_info(p)->cpu = cpu; | |
1981 | #endif | |
1982 | } | |
1983 | ||
1984 | static inline void check_class_changed(struct rq *rq, struct task_struct *p, | |
1985 | const struct sched_class *prev_class, | |
1986 | int oldprio, int running) | |
1987 | { | |
1988 | if (prev_class != p->sched_class) { | |
1989 | if (prev_class->switched_from) | |
1990 | prev_class->switched_from(rq, p, running); | |
1991 | p->sched_class->switched_to(rq, p, running); | |
1992 | } else | |
1993 | p->sched_class->prio_changed(rq, p, oldprio, running); | |
1994 | } | |
1995 | ||
1996 | /** | |
1997 | * kthread_bind - bind a just-created kthread to a cpu. | |
1998 | * @p: thread created by kthread_create(). | |
1999 | * @cpu: cpu (might not be online, must be possible) for @k to run on. | |
2000 | * | |
2001 | * Description: This function is equivalent to set_cpus_allowed(), | |
2002 | * except that @cpu doesn't need to be online, and the thread must be | |
2003 | * stopped (i.e., just returned from kthread_create()). | |
2004 | * | |
2005 | * Function lives here instead of kthread.c because it messes with | |
2006 | * scheduler internals which require locking. | |
2007 | */ | |
2008 | void kthread_bind(struct task_struct *p, unsigned int cpu) | |
2009 | { | |
2010 | struct rq *rq = cpu_rq(cpu); | |
2011 | unsigned long flags; | |
2012 | ||
2013 | /* Must have done schedule() in kthread() before we set_task_cpu */ | |
2014 | if (!wait_task_inactive(p, TASK_UNINTERRUPTIBLE)) { | |
2015 | WARN_ON(1); | |
2016 | return; | |
2017 | } | |
2018 | ||
2019 | spin_lock_irqsave(&rq->lock, flags); | |
2020 | set_task_cpu(p, cpu); | |
2021 | p->cpus_allowed = cpumask_of_cpu(cpu); | |
2022 | p->rt.nr_cpus_allowed = 1; | |
2023 | p->flags |= PF_THREAD_BOUND; | |
2024 | spin_unlock_irqrestore(&rq->lock, flags); | |
2025 | } | |
2026 | EXPORT_SYMBOL(kthread_bind); | |
2027 | ||
2028 | #ifdef CONFIG_SMP | |
2029 | /* | |
2030 | * Is this task likely cache-hot: | |
2031 | */ | |
2032 | static int | |
2033 | task_hot(struct task_struct *p, u64 now, struct sched_domain *sd) | |
2034 | { | |
2035 | s64 delta; | |
2036 | ||
2037 | /* | |
2038 | * Buddy candidates are cache hot: | |
2039 | */ | |
2040 | if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running && | |
2041 | (&p->se == cfs_rq_of(&p->se)->next || | |
2042 | &p->se == cfs_rq_of(&p->se)->last)) | |
2043 | return 1; | |
2044 | ||
2045 | if (p->sched_class != &fair_sched_class) | |
2046 | return 0; | |
2047 | ||
2048 | if (sysctl_sched_migration_cost == -1) | |
2049 | return 1; | |
2050 | if (sysctl_sched_migration_cost == 0) | |
2051 | return 0; | |
2052 | ||
2053 | delta = now - p->se.exec_start; | |
2054 | ||
2055 | return delta < (s64)sysctl_sched_migration_cost; | |
2056 | } | |
2057 | ||
2058 | ||
2059 | void set_task_cpu(struct task_struct *p, unsigned int new_cpu) | |
2060 | { | |
2061 | int old_cpu = task_cpu(p); | |
2062 | struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu); | |
2063 | struct cfs_rq *old_cfsrq = task_cfs_rq(p), | |
2064 | *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu); | |
2065 | u64 clock_offset; | |
2066 | ||
2067 | clock_offset = old_rq->clock - new_rq->clock; | |
2068 | ||
2069 | trace_sched_migrate_task(p, new_cpu); | |
2070 | ||
2071 | #ifdef CONFIG_SCHEDSTATS | |
2072 | if (p->se.wait_start) | |
2073 | p->se.wait_start -= clock_offset; | |
2074 | if (p->se.sleep_start) | |
2075 | p->se.sleep_start -= clock_offset; | |
2076 | if (p->se.block_start) | |
2077 | p->se.block_start -= clock_offset; | |
2078 | #endif | |
2079 | if (old_cpu != new_cpu) { | |
2080 | p->se.nr_migrations++; | |
2081 | new_rq->nr_migrations_in++; | |
2082 | #ifdef CONFIG_SCHEDSTATS | |
2083 | if (task_hot(p, old_rq->clock, NULL)) | |
2084 | schedstat_inc(p, se.nr_forced2_migrations); | |
2085 | #endif | |
2086 | perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, | |
2087 | 1, 1, NULL, 0); | |
2088 | } | |
2089 | p->se.vruntime -= old_cfsrq->min_vruntime - | |
2090 | new_cfsrq->min_vruntime; | |
2091 | ||
2092 | __set_task_cpu(p, new_cpu); | |
2093 | } | |
2094 | ||
2095 | struct migration_req { | |
2096 | struct list_head list; | |
2097 | ||
2098 | struct task_struct *task; | |
2099 | int dest_cpu; | |
2100 | ||
2101 | struct completion done; | |
2102 | }; | |
2103 | ||
2104 | /* | |
2105 | * The task's runqueue lock must be held. | |
2106 | * Returns true if you have to wait for migration thread. | |
2107 | */ | |
2108 | static int | |
2109 | migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req) | |
2110 | { | |
2111 | struct rq *rq = task_rq(p); | |
2112 | ||
2113 | /* | |
2114 | * If the task is not on a runqueue (and not running), then | |
2115 | * it is sufficient to simply update the task's cpu field. | |
2116 | */ | |
2117 | if (!p->se.on_rq && !task_running(rq, p)) { | |
2118 | set_task_cpu(p, dest_cpu); | |
2119 | return 0; | |
2120 | } | |
2121 | ||
2122 | init_completion(&req->done); | |
2123 | req->task = p; | |
2124 | req->dest_cpu = dest_cpu; | |
2125 | list_add(&req->list, &rq->migration_queue); | |
2126 | ||
2127 | return 1; | |
2128 | } | |
2129 | ||
2130 | /* | |
2131 | * wait_task_context_switch - wait for a thread to complete at least one | |
2132 | * context switch. | |
2133 | * | |
2134 | * @p must not be current. | |
2135 | */ | |
2136 | void wait_task_context_switch(struct task_struct *p) | |
2137 | { | |
2138 | unsigned long nvcsw, nivcsw, flags; | |
2139 | int running; | |
2140 | struct rq *rq; | |
2141 | ||
2142 | nvcsw = p->nvcsw; | |
2143 | nivcsw = p->nivcsw; | |
2144 | for (;;) { | |
2145 | /* | |
2146 | * The runqueue is assigned before the actual context | |
2147 | * switch. We need to take the runqueue lock. | |
2148 | * | |
2149 | * We could check initially without the lock but it is | |
2150 | * very likely that we need to take the lock in every | |
2151 | * iteration. | |
2152 | */ | |
2153 | rq = task_rq_lock(p, &flags); | |
2154 | running = task_running(rq, p); | |
2155 | task_rq_unlock(rq, &flags); | |
2156 | ||
2157 | if (likely(!running)) | |
2158 | break; | |
2159 | /* | |
2160 | * The switch count is incremented before the actual | |
2161 | * context switch. We thus wait for two switches to be | |
2162 | * sure at least one completed. | |
2163 | */ | |
2164 | if ((p->nvcsw - nvcsw) > 1) | |
2165 | break; | |
2166 | if ((p->nivcsw - nivcsw) > 1) | |
2167 | break; | |
2168 | ||
2169 | cpu_relax(); | |
2170 | } | |
2171 | } | |
2172 | ||
2173 | /* | |
2174 | * wait_task_inactive - wait for a thread to unschedule. | |
2175 | * | |
2176 | * If @match_state is nonzero, it's the @p->state value just checked and | |
2177 | * not expected to change. If it changes, i.e. @p might have woken up, | |
2178 | * then return zero. When we succeed in waiting for @p to be off its CPU, | |
2179 | * we return a positive number (its total switch count). If a second call | |
2180 | * a short while later returns the same number, the caller can be sure that | |
2181 | * @p has remained unscheduled the whole time. | |
2182 | * | |
2183 | * The caller must ensure that the task *will* unschedule sometime soon, | |
2184 | * else this function might spin for a *long* time. This function can't | |
2185 | * be called with interrupts off, or it may introduce deadlock with | |
2186 | * smp_call_function() if an IPI is sent by the same process we are | |
2187 | * waiting to become inactive. | |
2188 | */ | |
2189 | unsigned long wait_task_inactive(struct task_struct *p, long match_state) | |
2190 | { | |
2191 | unsigned long flags; | |
2192 | int running, on_rq; | |
2193 | unsigned long ncsw; | |
2194 | struct rq *rq; | |
2195 | ||
2196 | for (;;) { | |
2197 | /* | |
2198 | * We do the initial early heuristics without holding | |
2199 | * any task-queue locks at all. We'll only try to get | |
2200 | * the runqueue lock when things look like they will | |
2201 | * work out! | |
2202 | */ | |
2203 | rq = task_rq(p); | |
2204 | ||
2205 | /* | |
2206 | * If the task is actively running on another CPU | |
2207 | * still, just relax and busy-wait without holding | |
2208 | * any locks. | |
2209 | * | |
2210 | * NOTE! Since we don't hold any locks, it's not | |
2211 | * even sure that "rq" stays as the right runqueue! | |
2212 | * But we don't care, since "task_running()" will | |
2213 | * return false if the runqueue has changed and p | |
2214 | * is actually now running somewhere else! | |
2215 | */ | |
2216 | while (task_running(rq, p)) { | |
2217 | if (match_state && unlikely(p->state != match_state)) | |
2218 | return 0; | |
2219 | cpu_relax(); | |
2220 | } | |
2221 | ||
2222 | /* | |
2223 | * Ok, time to look more closely! We need the rq | |
2224 | * lock now, to be *sure*. If we're wrong, we'll | |
2225 | * just go back and repeat. | |
2226 | */ | |
2227 | rq = task_rq_lock(p, &flags); | |
2228 | trace_sched_wait_task(rq, p); | |
2229 | running = task_running(rq, p); | |
2230 | on_rq = p->se.on_rq; | |
2231 | ncsw = 0; | |
2232 | if (!match_state || p->state == match_state) | |
2233 | ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ | |
2234 | task_rq_unlock(rq, &flags); | |
2235 | ||
2236 | /* | |
2237 | * If it changed from the expected state, bail out now. | |
2238 | */ | |
2239 | if (unlikely(!ncsw)) | |
2240 | break; | |
2241 | ||
2242 | /* | |
2243 | * Was it really running after all now that we | |
2244 | * checked with the proper locks actually held? | |
2245 | * | |
2246 | * Oops. Go back and try again.. | |
2247 | */ | |
2248 | if (unlikely(running)) { | |
2249 | cpu_relax(); | |
2250 | continue; | |
2251 | } | |
2252 | ||
2253 | /* | |
2254 | * It's not enough that it's not actively running, | |
2255 | * it must be off the runqueue _entirely_, and not | |
2256 | * preempted! | |
2257 | * | |
2258 | * So if it was still runnable (but just not actively | |
2259 | * running right now), it's preempted, and we should | |
2260 | * yield - it could be a while. | |
2261 | */ | |
2262 | if (unlikely(on_rq)) { | |
2263 | schedule_timeout_uninterruptible(1); | |
2264 | continue; | |
2265 | } | |
2266 | ||
2267 | /* | |
2268 | * Ahh, all good. It wasn't running, and it wasn't | |
2269 | * runnable, which means that it will never become | |
2270 | * running in the future either. We're all done! | |
2271 | */ | |
2272 | break; | |
2273 | } | |
2274 | ||
2275 | return ncsw; | |
2276 | } | |
2277 | ||
2278 | /*** | |
2279 | * kick_process - kick a running thread to enter/exit the kernel | |
2280 | * @p: the to-be-kicked thread | |
2281 | * | |
2282 | * Cause a process which is running on another CPU to enter | |
2283 | * kernel-mode, without any delay. (to get signals handled.) | |
2284 | * | |
2285 | * NOTE: this function doesnt have to take the runqueue lock, | |
2286 | * because all it wants to ensure is that the remote task enters | |
2287 | * the kernel. If the IPI races and the task has been migrated | |
2288 | * to another CPU then no harm is done and the purpose has been | |
2289 | * achieved as well. | |
2290 | */ | |
2291 | void kick_process(struct task_struct *p) | |
2292 | { | |
2293 | int cpu; | |
2294 | ||
2295 | preempt_disable(); | |
2296 | cpu = task_cpu(p); | |
2297 | if ((cpu != smp_processor_id()) && task_curr(p)) | |
2298 | smp_send_reschedule(cpu); | |
2299 | preempt_enable(); | |
2300 | } | |
2301 | EXPORT_SYMBOL_GPL(kick_process); | |
2302 | #endif /* CONFIG_SMP */ | |
2303 | ||
2304 | /** | |
2305 | * task_oncpu_function_call - call a function on the cpu on which a task runs | |
2306 | * @p: the task to evaluate | |
2307 | * @func: the function to be called | |
2308 | * @info: the function call argument | |
2309 | * | |
2310 | * Calls the function @func when the task is currently running. This might | |
2311 | * be on the current CPU, which just calls the function directly | |
2312 | */ | |
2313 | void task_oncpu_function_call(struct task_struct *p, | |
2314 | void (*func) (void *info), void *info) | |
2315 | { | |
2316 | int cpu; | |
2317 | ||
2318 | preempt_disable(); | |
2319 | cpu = task_cpu(p); | |
2320 | if (task_curr(p)) | |
2321 | smp_call_function_single(cpu, func, info, 1); | |
2322 | preempt_enable(); | |
2323 | } | |
2324 | ||
2325 | /*** | |
2326 | * try_to_wake_up - wake up a thread | |
2327 | * @p: the to-be-woken-up thread | |
2328 | * @state: the mask of task states that can be woken | |
2329 | * @sync: do a synchronous wakeup? | |
2330 | * | |
2331 | * Put it on the run-queue if it's not already there. The "current" | |
2332 | * thread is always on the run-queue (except when the actual | |
2333 | * re-schedule is in progress), and as such you're allowed to do | |
2334 | * the simpler "current->state = TASK_RUNNING" to mark yourself | |
2335 | * runnable without the overhead of this. | |
2336 | * | |
2337 | * returns failure only if the task is already active. | |
2338 | */ | |
2339 | static int try_to_wake_up(struct task_struct *p, unsigned int state, | |
2340 | int wake_flags) | |
2341 | { | |
2342 | int cpu, orig_cpu, this_cpu, success = 0; | |
2343 | unsigned long flags; | |
2344 | struct rq *rq, *orig_rq; | |
2345 | ||
2346 | if (!sched_feat(SYNC_WAKEUPS)) | |
2347 | wake_flags &= ~WF_SYNC; | |
2348 | ||
2349 | this_cpu = get_cpu(); | |
2350 | ||
2351 | smp_wmb(); | |
2352 | rq = orig_rq = task_rq_lock(p, &flags); | |
2353 | update_rq_clock(rq); | |
2354 | if (!(p->state & state)) | |
2355 | goto out; | |
2356 | ||
2357 | if (p->se.on_rq) | |
2358 | goto out_running; | |
2359 | ||
2360 | cpu = task_cpu(p); | |
2361 | orig_cpu = cpu; | |
2362 | ||
2363 | #ifdef CONFIG_SMP | |
2364 | if (unlikely(task_running(rq, p))) | |
2365 | goto out_activate; | |
2366 | ||
2367 | /* | |
2368 | * In order to handle concurrent wakeups and release the rq->lock | |
2369 | * we put the task in TASK_WAKING state. | |
2370 | * | |
2371 | * First fix up the nr_uninterruptible count: | |
2372 | */ | |
2373 | if (task_contributes_to_load(p)) | |
2374 | rq->nr_uninterruptible--; | |
2375 | p->state = TASK_WAKING; | |
2376 | task_rq_unlock(rq, &flags); | |
2377 | ||
2378 | cpu = p->sched_class->select_task_rq(p, SD_BALANCE_WAKE, wake_flags); | |
2379 | if (cpu != orig_cpu) | |
2380 | set_task_cpu(p, cpu); | |
2381 | ||
2382 | rq = task_rq_lock(p, &flags); | |
2383 | ||
2384 | if (rq != orig_rq) | |
2385 | update_rq_clock(rq); | |
2386 | ||
2387 | WARN_ON(p->state != TASK_WAKING); | |
2388 | cpu = task_cpu(p); | |
2389 | ||
2390 | #ifdef CONFIG_SCHEDSTATS | |
2391 | schedstat_inc(rq, ttwu_count); | |
2392 | if (cpu == this_cpu) | |
2393 | schedstat_inc(rq, ttwu_local); | |
2394 | else { | |
2395 | struct sched_domain *sd; | |
2396 | for_each_domain(this_cpu, sd) { | |
2397 | if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { | |
2398 | schedstat_inc(sd, ttwu_wake_remote); | |
2399 | break; | |
2400 | } | |
2401 | } | |
2402 | } | |
2403 | #endif /* CONFIG_SCHEDSTATS */ | |
2404 | ||
2405 | out_activate: | |
2406 | #endif /* CONFIG_SMP */ | |
2407 | schedstat_inc(p, se.nr_wakeups); | |
2408 | if (wake_flags & WF_SYNC) | |
2409 | schedstat_inc(p, se.nr_wakeups_sync); | |
2410 | if (orig_cpu != cpu) | |
2411 | schedstat_inc(p, se.nr_wakeups_migrate); | |
2412 | if (cpu == this_cpu) | |
2413 | schedstat_inc(p, se.nr_wakeups_local); | |
2414 | else | |
2415 | schedstat_inc(p, se.nr_wakeups_remote); | |
2416 | activate_task(rq, p, 1); | |
2417 | success = 1; | |
2418 | ||
2419 | /* | |
2420 | * Only attribute actual wakeups done by this task. | |
2421 | */ | |
2422 | if (!in_interrupt()) { | |
2423 | struct sched_entity *se = ¤t->se; | |
2424 | u64 sample = se->sum_exec_runtime; | |
2425 | ||
2426 | if (se->last_wakeup) | |
2427 | sample -= se->last_wakeup; | |
2428 | else | |
2429 | sample -= se->start_runtime; | |
2430 | update_avg(&se->avg_wakeup, sample); | |
2431 | ||
2432 | se->last_wakeup = se->sum_exec_runtime; | |
2433 | } | |
2434 | ||
2435 | out_running: | |
2436 | trace_sched_wakeup(rq, p, success); | |
2437 | check_preempt_curr(rq, p, wake_flags); | |
2438 | ||
2439 | p->state = TASK_RUNNING; | |
2440 | #ifdef CONFIG_SMP | |
2441 | if (p->sched_class->task_wake_up) | |
2442 | p->sched_class->task_wake_up(rq, p); | |
2443 | #endif | |
2444 | out: | |
2445 | task_rq_unlock(rq, &flags); | |
2446 | put_cpu(); | |
2447 | ||
2448 | return success; | |
2449 | } | |
2450 | ||
2451 | /** | |
2452 | * wake_up_process - Wake up a specific process | |
2453 | * @p: The process to be woken up. | |
2454 | * | |
2455 | * Attempt to wake up the nominated process and move it to the set of runnable | |
2456 | * processes. Returns 1 if the process was woken up, 0 if it was already | |
2457 | * running. | |
2458 | * | |
2459 | * It may be assumed that this function implies a write memory barrier before | |
2460 | * changing the task state if and only if any tasks are woken up. | |
2461 | */ | |
2462 | int wake_up_process(struct task_struct *p) | |
2463 | { | |
2464 | return try_to_wake_up(p, TASK_ALL, 0); | |
2465 | } | |
2466 | EXPORT_SYMBOL(wake_up_process); | |
2467 | ||
2468 | int wake_up_state(struct task_struct *p, unsigned int state) | |
2469 | { | |
2470 | return try_to_wake_up(p, state, 0); | |
2471 | } | |
2472 | ||
2473 | /* | |
2474 | * Perform scheduler related setup for a newly forked process p. | |
2475 | * p is forked by current. | |
2476 | * | |
2477 | * __sched_fork() is basic setup used by init_idle() too: | |
2478 | */ | |
2479 | static void __sched_fork(struct task_struct *p) | |
2480 | { | |
2481 | p->se.exec_start = 0; | |
2482 | p->se.sum_exec_runtime = 0; | |
2483 | p->se.prev_sum_exec_runtime = 0; | |
2484 | p->se.nr_migrations = 0; | |
2485 | p->se.last_wakeup = 0; | |
2486 | p->se.avg_overlap = 0; | |
2487 | p->se.start_runtime = 0; | |
2488 | p->se.avg_wakeup = sysctl_sched_wakeup_granularity; | |
2489 | p->se.avg_running = 0; | |
2490 | ||
2491 | #ifdef CONFIG_SCHEDSTATS | |
2492 | p->se.wait_start = 0; | |
2493 | p->se.wait_max = 0; | |
2494 | p->se.wait_count = 0; | |
2495 | p->se.wait_sum = 0; | |
2496 | ||
2497 | p->se.sleep_start = 0; | |
2498 | p->se.sleep_max = 0; | |
2499 | p->se.sum_sleep_runtime = 0; | |
2500 | ||
2501 | p->se.block_start = 0; | |
2502 | p->se.block_max = 0; | |
2503 | p->se.exec_max = 0; | |
2504 | p->se.slice_max = 0; | |
2505 | ||
2506 | p->se.nr_migrations_cold = 0; | |
2507 | p->se.nr_failed_migrations_affine = 0; | |
2508 | p->se.nr_failed_migrations_running = 0; | |
2509 | p->se.nr_failed_migrations_hot = 0; | |
2510 | p->se.nr_forced_migrations = 0; | |
2511 | p->se.nr_forced2_migrations = 0; | |
2512 | ||
2513 | p->se.nr_wakeups = 0; | |
2514 | p->se.nr_wakeups_sync = 0; | |
2515 | p->se.nr_wakeups_migrate = 0; | |
2516 | p->se.nr_wakeups_local = 0; | |
2517 | p->se.nr_wakeups_remote = 0; | |
2518 | p->se.nr_wakeups_affine = 0; | |
2519 | p->se.nr_wakeups_affine_attempts = 0; | |
2520 | p->se.nr_wakeups_passive = 0; | |
2521 | p->se.nr_wakeups_idle = 0; | |
2522 | ||
2523 | #endif | |
2524 | ||
2525 | INIT_LIST_HEAD(&p->rt.run_list); | |
2526 | p->se.on_rq = 0; | |
2527 | INIT_LIST_HEAD(&p->se.group_node); | |
2528 | ||
2529 | #ifdef CONFIG_PREEMPT_NOTIFIERS | |
2530 | INIT_HLIST_HEAD(&p->preempt_notifiers); | |
2531 | #endif | |
2532 | ||
2533 | /* | |
2534 | * We mark the process as running here, but have not actually | |
2535 | * inserted it onto the runqueue yet. This guarantees that | |
2536 | * nobody will actually run it, and a signal or other external | |
2537 | * event cannot wake it up and insert it on the runqueue either. | |
2538 | */ | |
2539 | p->state = TASK_RUNNING; | |
2540 | } | |
2541 | ||
2542 | /* | |
2543 | * fork()/clone()-time setup: | |
2544 | */ | |
2545 | void sched_fork(struct task_struct *p, int clone_flags) | |
2546 | { | |
2547 | int cpu = get_cpu(); | |
2548 | ||
2549 | __sched_fork(p); | |
2550 | ||
2551 | /* | |
2552 | * Revert to default priority/policy on fork if requested. | |
2553 | */ | |
2554 | if (unlikely(p->sched_reset_on_fork)) { | |
2555 | if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) { | |
2556 | p->policy = SCHED_NORMAL; | |
2557 | p->normal_prio = p->static_prio; | |
2558 | } | |
2559 | ||
2560 | if (PRIO_TO_NICE(p->static_prio) < 0) { | |
2561 | p->static_prio = NICE_TO_PRIO(0); | |
2562 | p->normal_prio = p->static_prio; | |
2563 | set_load_weight(p); | |
2564 | } | |
2565 | ||
2566 | /* | |
2567 | * We don't need the reset flag anymore after the fork. It has | |
2568 | * fulfilled its duty: | |
2569 | */ | |
2570 | p->sched_reset_on_fork = 0; | |
2571 | } | |
2572 | ||
2573 | /* | |
2574 | * Make sure we do not leak PI boosting priority to the child. | |
2575 | */ | |
2576 | p->prio = current->normal_prio; | |
2577 | ||
2578 | if (!rt_prio(p->prio)) | |
2579 | p->sched_class = &fair_sched_class; | |
2580 | ||
2581 | #ifdef CONFIG_SMP | |
2582 | cpu = p->sched_class->select_task_rq(p, SD_BALANCE_FORK, 0); | |
2583 | #endif | |
2584 | set_task_cpu(p, cpu); | |
2585 | ||
2586 | #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) | |
2587 | if (likely(sched_info_on())) | |
2588 | memset(&p->sched_info, 0, sizeof(p->sched_info)); | |
2589 | #endif | |
2590 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) | |
2591 | p->oncpu = 0; | |
2592 | #endif | |
2593 | #ifdef CONFIG_PREEMPT | |
2594 | /* Want to start with kernel preemption disabled. */ | |
2595 | task_thread_info(p)->preempt_count = 1; | |
2596 | #endif | |
2597 | plist_node_init(&p->pushable_tasks, MAX_PRIO); | |
2598 | ||
2599 | put_cpu(); | |
2600 | } | |
2601 | ||
2602 | /* | |
2603 | * wake_up_new_task - wake up a newly created task for the first time. | |
2604 | * | |
2605 | * This function will do some initial scheduler statistics housekeeping | |
2606 | * that must be done for every newly created context, then puts the task | |
2607 | * on the runqueue and wakes it. | |
2608 | */ | |
2609 | void wake_up_new_task(struct task_struct *p, unsigned long clone_flags) | |
2610 | { | |
2611 | unsigned long flags; | |
2612 | struct rq *rq; | |
2613 | ||
2614 | rq = task_rq_lock(p, &flags); | |
2615 | BUG_ON(p->state != TASK_RUNNING); | |
2616 | update_rq_clock(rq); | |
2617 | ||
2618 | if (!p->sched_class->task_new || !current->se.on_rq) { | |
2619 | activate_task(rq, p, 0); | |
2620 | } else { | |
2621 | /* | |
2622 | * Let the scheduling class do new task startup | |
2623 | * management (if any): | |
2624 | */ | |
2625 | p->sched_class->task_new(rq, p); | |
2626 | inc_nr_running(rq); | |
2627 | } | |
2628 | trace_sched_wakeup_new(rq, p, 1); | |
2629 | check_preempt_curr(rq, p, WF_FORK); | |
2630 | #ifdef CONFIG_SMP | |
2631 | if (p->sched_class->task_wake_up) | |
2632 | p->sched_class->task_wake_up(rq, p); | |
2633 | #endif | |
2634 | task_rq_unlock(rq, &flags); | |
2635 | } | |
2636 | ||
2637 | #ifdef CONFIG_PREEMPT_NOTIFIERS | |
2638 | ||
2639 | /** | |
2640 | * preempt_notifier_register - tell me when current is being preempted & rescheduled | |
2641 | * @notifier: notifier struct to register | |
2642 | */ | |
2643 | void preempt_notifier_register(struct preempt_notifier *notifier) | |
2644 | { | |
2645 | hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); | |
2646 | } | |
2647 | EXPORT_SYMBOL_GPL(preempt_notifier_register); | |
2648 | ||
2649 | /** | |
2650 | * preempt_notifier_unregister - no longer interested in preemption notifications | |
2651 | * @notifier: notifier struct to unregister | |
2652 | * | |
2653 | * This is safe to call from within a preemption notifier. | |
2654 | */ | |
2655 | void preempt_notifier_unregister(struct preempt_notifier *notifier) | |
2656 | { | |
2657 | hlist_del(¬ifier->link); | |
2658 | } | |
2659 | EXPORT_SYMBOL_GPL(preempt_notifier_unregister); | |
2660 | ||
2661 | static void fire_sched_in_preempt_notifiers(struct task_struct *curr) | |
2662 | { | |
2663 | struct preempt_notifier *notifier; | |
2664 | struct hlist_node *node; | |
2665 | ||
2666 | hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) | |
2667 | notifier->ops->sched_in(notifier, raw_smp_processor_id()); | |
2668 | } | |
2669 | ||
2670 | static void | |
2671 | fire_sched_out_preempt_notifiers(struct task_struct *curr, | |
2672 | struct task_struct *next) | |
2673 | { | |
2674 | struct preempt_notifier *notifier; | |
2675 | struct hlist_node *node; | |
2676 | ||
2677 | hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) | |
2678 | notifier->ops->sched_out(notifier, next); | |
2679 | } | |
2680 | ||
2681 | #else /* !CONFIG_PREEMPT_NOTIFIERS */ | |
2682 | ||
2683 | static void fire_sched_in_preempt_notifiers(struct task_struct *curr) | |
2684 | { | |
2685 | } | |
2686 | ||
2687 | static void | |
2688 | fire_sched_out_preempt_notifiers(struct task_struct *curr, | |
2689 | struct task_struct *next) | |
2690 | { | |
2691 | } | |
2692 | ||
2693 | #endif /* CONFIG_PREEMPT_NOTIFIERS */ | |
2694 | ||
2695 | /** | |
2696 | * prepare_task_switch - prepare to switch tasks | |
2697 | * @rq: the runqueue preparing to switch | |
2698 | * @prev: the current task that is being switched out | |
2699 | * @next: the task we are going to switch to. | |
2700 | * | |
2701 | * This is called with the rq lock held and interrupts off. It must | |
2702 | * be paired with a subsequent finish_task_switch after the context | |
2703 | * switch. | |
2704 | * | |
2705 | * prepare_task_switch sets up locking and calls architecture specific | |
2706 | * hooks. | |
2707 | */ | |
2708 | static inline void | |
2709 | prepare_task_switch(struct rq *rq, struct task_struct *prev, | |
2710 | struct task_struct *next) | |
2711 | { | |
2712 | fire_sched_out_preempt_notifiers(prev, next); | |
2713 | prepare_lock_switch(rq, next); | |
2714 | prepare_arch_switch(next); | |
2715 | } | |
2716 | ||
2717 | /** | |
2718 | * finish_task_switch - clean up after a task-switch | |
2719 | * @rq: runqueue associated with task-switch | |
2720 | * @prev: the thread we just switched away from. | |
2721 | * | |
2722 | * finish_task_switch must be called after the context switch, paired | |
2723 | * with a prepare_task_switch call before the context switch. | |
2724 | * finish_task_switch will reconcile locking set up by prepare_task_switch, | |
2725 | * and do any other architecture-specific cleanup actions. | |
2726 | * | |
2727 | * Note that we may have delayed dropping an mm in context_switch(). If | |
2728 | * so, we finish that here outside of the runqueue lock. (Doing it | |
2729 | * with the lock held can cause deadlocks; see schedule() for | |
2730 | * details.) | |
2731 | */ | |
2732 | static void finish_task_switch(struct rq *rq, struct task_struct *prev) | |
2733 | __releases(rq->lock) | |
2734 | { | |
2735 | struct mm_struct *mm = rq->prev_mm; | |
2736 | long prev_state; | |
2737 | ||
2738 | rq->prev_mm = NULL; | |
2739 | ||
2740 | /* | |
2741 | * A task struct has one reference for the use as "current". | |
2742 | * If a task dies, then it sets TASK_DEAD in tsk->state and calls | |
2743 | * schedule one last time. The schedule call will never return, and | |
2744 | * the scheduled task must drop that reference. | |
2745 | * The test for TASK_DEAD must occur while the runqueue locks are | |
2746 | * still held, otherwise prev could be scheduled on another cpu, die | |
2747 | * there before we look at prev->state, and then the reference would | |
2748 | * be dropped twice. | |
2749 | * Manfred Spraul <manfred@colorfullife.com> | |
2750 | */ | |
2751 | prev_state = prev->state; | |
2752 | finish_arch_switch(prev); | |
2753 | perf_event_task_sched_in(current, cpu_of(rq)); | |
2754 | finish_lock_switch(rq, prev); | |
2755 | ||
2756 | fire_sched_in_preempt_notifiers(current); | |
2757 | if (mm) | |
2758 | mmdrop(mm); | |
2759 | if (unlikely(prev_state == TASK_DEAD)) { | |
2760 | /* | |
2761 | * Remove function-return probe instances associated with this | |
2762 | * task and put them back on the free list. | |
2763 | */ | |
2764 | kprobe_flush_task(prev); | |
2765 | put_task_struct(prev); | |
2766 | } | |
2767 | } | |
2768 | ||
2769 | #ifdef CONFIG_SMP | |
2770 | ||
2771 | /* assumes rq->lock is held */ | |
2772 | static inline void pre_schedule(struct rq *rq, struct task_struct *prev) | |
2773 | { | |
2774 | if (prev->sched_class->pre_schedule) | |
2775 | prev->sched_class->pre_schedule(rq, prev); | |
2776 | } | |
2777 | ||
2778 | /* rq->lock is NOT held, but preemption is disabled */ | |
2779 | static inline void post_schedule(struct rq *rq) | |
2780 | { | |
2781 | if (rq->post_schedule) { | |
2782 | unsigned long flags; | |
2783 | ||
2784 | spin_lock_irqsave(&rq->lock, flags); | |
2785 | if (rq->curr->sched_class->post_schedule) | |
2786 | rq->curr->sched_class->post_schedule(rq); | |
2787 | spin_unlock_irqrestore(&rq->lock, flags); | |
2788 | ||
2789 | rq->post_schedule = 0; | |
2790 | } | |
2791 | } | |
2792 | ||
2793 | #else | |
2794 | ||
2795 | static inline void pre_schedule(struct rq *rq, struct task_struct *p) | |
2796 | { | |
2797 | } | |
2798 | ||
2799 | static inline void post_schedule(struct rq *rq) | |
2800 | { | |
2801 | } | |
2802 | ||
2803 | #endif | |
2804 | ||
2805 | /** | |
2806 | * schedule_tail - first thing a freshly forked thread must call. | |
2807 | * @prev: the thread we just switched away from. | |
2808 | */ | |
2809 | asmlinkage void schedule_tail(struct task_struct *prev) | |
2810 | __releases(rq->lock) | |
2811 | { | |
2812 | struct rq *rq = this_rq(); | |
2813 | ||
2814 | finish_task_switch(rq, prev); | |
2815 | ||
2816 | /* | |
2817 | * FIXME: do we need to worry about rq being invalidated by the | |
2818 | * task_switch? | |
2819 | */ | |
2820 | post_schedule(rq); | |
2821 | ||
2822 | #ifdef __ARCH_WANT_UNLOCKED_CTXSW | |
2823 | /* In this case, finish_task_switch does not reenable preemption */ | |
2824 | preempt_enable(); | |
2825 | #endif | |
2826 | if (current->set_child_tid) | |
2827 | put_user(task_pid_vnr(current), current->set_child_tid); | |
2828 | } | |
2829 | ||
2830 | /* | |
2831 | * context_switch - switch to the new MM and the new | |
2832 | * thread's register state. | |
2833 | */ | |
2834 | static inline void | |
2835 | context_switch(struct rq *rq, struct task_struct *prev, | |
2836 | struct task_struct *next) | |
2837 | { | |
2838 | struct mm_struct *mm, *oldmm; | |
2839 | ||
2840 | prepare_task_switch(rq, prev, next); | |
2841 | trace_sched_switch(rq, prev, next); | |
2842 | mm = next->mm; | |
2843 | oldmm = prev->active_mm; | |
2844 | /* | |
2845 | * For paravirt, this is coupled with an exit in switch_to to | |
2846 | * combine the page table reload and the switch backend into | |
2847 | * one hypercall. | |
2848 | */ | |
2849 | arch_start_context_switch(prev); | |
2850 | ||
2851 | if (unlikely(!mm)) { | |
2852 | next->active_mm = oldmm; | |
2853 | atomic_inc(&oldmm->mm_count); | |
2854 | enter_lazy_tlb(oldmm, next); | |
2855 | } else | |
2856 | switch_mm(oldmm, mm, next); | |
2857 | ||
2858 | if (unlikely(!prev->mm)) { | |
2859 | prev->active_mm = NULL; | |
2860 | rq->prev_mm = oldmm; | |
2861 | } | |
2862 | /* | |
2863 | * Since the runqueue lock will be released by the next | |
2864 | * task (which is an invalid locking op but in the case | |
2865 | * of the scheduler it's an obvious special-case), so we | |
2866 | * do an early lockdep release here: | |
2867 | */ | |
2868 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW | |
2869 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); | |
2870 | #endif | |
2871 | ||
2872 | /* Here we just switch the register state and the stack. */ | |
2873 | switch_to(prev, next, prev); | |
2874 | ||
2875 | barrier(); | |
2876 | /* | |
2877 | * this_rq must be evaluated again because prev may have moved | |
2878 | * CPUs since it called schedule(), thus the 'rq' on its stack | |
2879 | * frame will be invalid. | |
2880 | */ | |
2881 | finish_task_switch(this_rq(), prev); | |
2882 | } | |
2883 | ||
2884 | /* | |
2885 | * nr_running, nr_uninterruptible and nr_context_switches: | |
2886 | * | |
2887 | * externally visible scheduler statistics: current number of runnable | |
2888 | * threads, current number of uninterruptible-sleeping threads, total | |
2889 | * number of context switches performed since bootup. | |
2890 | */ | |
2891 | unsigned long nr_running(void) | |
2892 | { | |
2893 | unsigned long i, sum = 0; | |
2894 | ||
2895 | for_each_online_cpu(i) | |
2896 | sum += cpu_rq(i)->nr_running; | |
2897 | ||
2898 | return sum; | |
2899 | } | |
2900 | ||
2901 | unsigned long nr_uninterruptible(void) | |
2902 | { | |
2903 | unsigned long i, sum = 0; | |
2904 | ||
2905 | for_each_possible_cpu(i) | |
2906 | sum += cpu_rq(i)->nr_uninterruptible; | |
2907 | ||
2908 | /* | |
2909 | * Since we read the counters lockless, it might be slightly | |
2910 | * inaccurate. Do not allow it to go below zero though: | |
2911 | */ | |
2912 | if (unlikely((long)sum < 0)) | |
2913 | sum = 0; | |
2914 | ||
2915 | return sum; | |
2916 | } | |
2917 | ||
2918 | unsigned long long nr_context_switches(void) | |
2919 | { | |
2920 | int i; | |
2921 | unsigned long long sum = 0; | |
2922 | ||
2923 | for_each_possible_cpu(i) | |
2924 | sum += cpu_rq(i)->nr_switches; | |
2925 | ||
2926 | return sum; | |
2927 | } | |
2928 | ||
2929 | unsigned long nr_iowait(void) | |
2930 | { | |
2931 | unsigned long i, sum = 0; | |
2932 | ||
2933 | for_each_possible_cpu(i) | |
2934 | sum += atomic_read(&cpu_rq(i)->nr_iowait); | |
2935 | ||
2936 | return sum; | |
2937 | } | |
2938 | ||
2939 | unsigned long nr_iowait_cpu(void) | |
2940 | { | |
2941 | struct rq *this = this_rq(); | |
2942 | return atomic_read(&this->nr_iowait); | |
2943 | } | |
2944 | ||
2945 | unsigned long this_cpu_load(void) | |
2946 | { | |
2947 | struct rq *this = this_rq(); | |
2948 | return this->cpu_load[0]; | |
2949 | } | |
2950 | ||
2951 | ||
2952 | /* Variables and functions for calc_load */ | |
2953 | static atomic_long_t calc_load_tasks; | |
2954 | static unsigned long calc_load_update; | |
2955 | unsigned long avenrun[3]; | |
2956 | EXPORT_SYMBOL(avenrun); | |
2957 | ||
2958 | /** | |
2959 | * get_avenrun - get the load average array | |
2960 | * @loads: pointer to dest load array | |
2961 | * @offset: offset to add | |
2962 | * @shift: shift count to shift the result left | |
2963 | * | |
2964 | * These values are estimates at best, so no need for locking. | |
2965 | */ | |
2966 | void get_avenrun(unsigned long *loads, unsigned long offset, int shift) | |
2967 | { | |
2968 | loads[0] = (avenrun[0] + offset) << shift; | |
2969 | loads[1] = (avenrun[1] + offset) << shift; | |
2970 | loads[2] = (avenrun[2] + offset) << shift; | |
2971 | } | |
2972 | ||
2973 | static unsigned long | |
2974 | calc_load(unsigned long load, unsigned long exp, unsigned long active) | |
2975 | { | |
2976 | load *= exp; | |
2977 | load += active * (FIXED_1 - exp); | |
2978 | return load >> FSHIFT; | |
2979 | } | |
2980 | ||
2981 | /* | |
2982 | * calc_load - update the avenrun load estimates 10 ticks after the | |
2983 | * CPUs have updated calc_load_tasks. | |
2984 | */ | |
2985 | void calc_global_load(void) | |
2986 | { | |
2987 | unsigned long upd = calc_load_update + 10; | |
2988 | long active; | |
2989 | ||
2990 | if (time_before(jiffies, upd)) | |
2991 | return; | |
2992 | ||
2993 | active = atomic_long_read(&calc_load_tasks); | |
2994 | active = active > 0 ? active * FIXED_1 : 0; | |
2995 | ||
2996 | avenrun[0] = calc_load(avenrun[0], EXP_1, active); | |
2997 | avenrun[1] = calc_load(avenrun[1], EXP_5, active); | |
2998 | avenrun[2] = calc_load(avenrun[2], EXP_15, active); | |
2999 | ||
3000 | calc_load_update += LOAD_FREQ; | |
3001 | } | |
3002 | ||
3003 | /* | |
3004 | * Either called from update_cpu_load() or from a cpu going idle | |
3005 | */ | |
3006 | static void calc_load_account_active(struct rq *this_rq) | |
3007 | { | |
3008 | long nr_active, delta; | |
3009 | ||
3010 | nr_active = this_rq->nr_running; | |
3011 | nr_active += (long) this_rq->nr_uninterruptible; | |
3012 | ||
3013 | if (nr_active != this_rq->calc_load_active) { | |
3014 | delta = nr_active - this_rq->calc_load_active; | |
3015 | this_rq->calc_load_active = nr_active; | |
3016 | atomic_long_add(delta, &calc_load_tasks); | |
3017 | } | |
3018 | } | |
3019 | ||
3020 | /* | |
3021 | * Externally visible per-cpu scheduler statistics: | |
3022 | * cpu_nr_migrations(cpu) - number of migrations into that cpu | |
3023 | */ | |
3024 | u64 cpu_nr_migrations(int cpu) | |
3025 | { | |
3026 | return cpu_rq(cpu)->nr_migrations_in; | |
3027 | } | |
3028 | ||
3029 | /* | |
3030 | * Update rq->cpu_load[] statistics. This function is usually called every | |
3031 | * scheduler tick (TICK_NSEC). | |
3032 | */ | |
3033 | static void update_cpu_load(struct rq *this_rq) | |
3034 | { | |
3035 | unsigned long this_load = this_rq->load.weight; | |
3036 | int i, scale; | |
3037 | ||
3038 | this_rq->nr_load_updates++; | |
3039 | ||
3040 | /* Update our load: */ | |
3041 | for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { | |
3042 | unsigned long old_load, new_load; | |
3043 | ||
3044 | /* scale is effectively 1 << i now, and >> i divides by scale */ | |
3045 | ||
3046 | old_load = this_rq->cpu_load[i]; | |
3047 | new_load = this_load; | |
3048 | /* | |
3049 | * Round up the averaging division if load is increasing. This | |
3050 | * prevents us from getting stuck on 9 if the load is 10, for | |
3051 | * example. | |
3052 | */ | |
3053 | if (new_load > old_load) | |
3054 | new_load += scale-1; | |
3055 | this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i; | |
3056 | } | |
3057 | ||
3058 | if (time_after_eq(jiffies, this_rq->calc_load_update)) { | |
3059 | this_rq->calc_load_update += LOAD_FREQ; | |
3060 | calc_load_account_active(this_rq); | |
3061 | } | |
3062 | } | |
3063 | ||
3064 | #ifdef CONFIG_SMP | |
3065 | ||
3066 | /* | |
3067 | * double_rq_lock - safely lock two runqueues | |
3068 | * | |
3069 | * Note this does not disable interrupts like task_rq_lock, | |
3070 | * you need to do so manually before calling. | |
3071 | */ | |
3072 | static void double_rq_lock(struct rq *rq1, struct rq *rq2) | |
3073 | __acquires(rq1->lock) | |
3074 | __acquires(rq2->lock) | |
3075 | { | |
3076 | BUG_ON(!irqs_disabled()); | |
3077 | if (rq1 == rq2) { | |
3078 | spin_lock(&rq1->lock); | |
3079 | __acquire(rq2->lock); /* Fake it out ;) */ | |
3080 | } else { | |
3081 | if (rq1 < rq2) { | |
3082 | spin_lock(&rq1->lock); | |
3083 | spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING); | |
3084 | } else { | |
3085 | spin_lock(&rq2->lock); | |
3086 | spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING); | |
3087 | } | |
3088 | } | |
3089 | update_rq_clock(rq1); | |
3090 | update_rq_clock(rq2); | |
3091 | } | |
3092 | ||
3093 | /* | |
3094 | * double_rq_unlock - safely unlock two runqueues | |
3095 | * | |
3096 | * Note this does not restore interrupts like task_rq_unlock, | |
3097 | * you need to do so manually after calling. | |
3098 | */ | |
3099 | static void double_rq_unlock(struct rq *rq1, struct rq *rq2) | |
3100 | __releases(rq1->lock) | |
3101 | __releases(rq2->lock) | |
3102 | { | |
3103 | spin_unlock(&rq1->lock); | |
3104 | if (rq1 != rq2) | |
3105 | spin_unlock(&rq2->lock); | |
3106 | else | |
3107 | __release(rq2->lock); | |
3108 | } | |
3109 | ||
3110 | /* | |
3111 | * If dest_cpu is allowed for this process, migrate the task to it. | |
3112 | * This is accomplished by forcing the cpu_allowed mask to only | |
3113 | * allow dest_cpu, which will force the cpu onto dest_cpu. Then | |
3114 | * the cpu_allowed mask is restored. | |
3115 | */ | |
3116 | static void sched_migrate_task(struct task_struct *p, int dest_cpu) | |
3117 | { | |
3118 | struct migration_req req; | |
3119 | unsigned long flags; | |
3120 | struct rq *rq; | |
3121 | ||
3122 | rq = task_rq_lock(p, &flags); | |
3123 | if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed) | |
3124 | || unlikely(!cpu_active(dest_cpu))) | |
3125 | goto out; | |
3126 | ||
3127 | /* force the process onto the specified CPU */ | |
3128 | if (migrate_task(p, dest_cpu, &req)) { | |
3129 | /* Need to wait for migration thread (might exit: take ref). */ | |
3130 | struct task_struct *mt = rq->migration_thread; | |
3131 | ||
3132 | get_task_struct(mt); | |
3133 | task_rq_unlock(rq, &flags); | |
3134 | wake_up_process(mt); | |
3135 | put_task_struct(mt); | |
3136 | wait_for_completion(&req.done); | |
3137 | ||
3138 | return; | |
3139 | } | |
3140 | out: | |
3141 | task_rq_unlock(rq, &flags); | |
3142 | } | |
3143 | ||
3144 | /* | |
3145 | * sched_exec - execve() is a valuable balancing opportunity, because at | |
3146 | * this point the task has the smallest effective memory and cache footprint. | |
3147 | */ | |
3148 | void sched_exec(void) | |
3149 | { | |
3150 | int new_cpu, this_cpu = get_cpu(); | |
3151 | new_cpu = current->sched_class->select_task_rq(current, SD_BALANCE_EXEC, 0); | |
3152 | put_cpu(); | |
3153 | if (new_cpu != this_cpu) | |
3154 | sched_migrate_task(current, new_cpu); | |
3155 | } | |
3156 | ||
3157 | /* | |
3158 | * pull_task - move a task from a remote runqueue to the local runqueue. | |
3159 | * Both runqueues must be locked. | |
3160 | */ | |
3161 | static void pull_task(struct rq *src_rq, struct task_struct *p, | |
3162 | struct rq *this_rq, int this_cpu) | |
3163 | { | |
3164 | deactivate_task(src_rq, p, 0); | |
3165 | set_task_cpu(p, this_cpu); | |
3166 | activate_task(this_rq, p, 0); | |
3167 | /* | |
3168 | * Note that idle threads have a prio of MAX_PRIO, for this test | |
3169 | * to be always true for them. | |
3170 | */ | |
3171 | check_preempt_curr(this_rq, p, 0); | |
3172 | } | |
3173 | ||
3174 | /* | |
3175 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
3176 | */ | |
3177 | static | |
3178 | int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu, | |
3179 | struct sched_domain *sd, enum cpu_idle_type idle, | |
3180 | int *all_pinned) | |
3181 | { | |
3182 | int tsk_cache_hot = 0; | |
3183 | /* | |
3184 | * We do not migrate tasks that are: | |
3185 | * 1) running (obviously), or | |
3186 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
3187 | * 3) are cache-hot on their current CPU. | |
3188 | */ | |
3189 | if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) { | |
3190 | schedstat_inc(p, se.nr_failed_migrations_affine); | |
3191 | return 0; | |
3192 | } | |
3193 | *all_pinned = 0; | |
3194 | ||
3195 | if (task_running(rq, p)) { | |
3196 | schedstat_inc(p, se.nr_failed_migrations_running); | |
3197 | return 0; | |
3198 | } | |
3199 | ||
3200 | /* | |
3201 | * Aggressive migration if: | |
3202 | * 1) task is cache cold, or | |
3203 | * 2) too many balance attempts have failed. | |
3204 | */ | |
3205 | ||
3206 | tsk_cache_hot = task_hot(p, rq->clock, sd); | |
3207 | if (!tsk_cache_hot || | |
3208 | sd->nr_balance_failed > sd->cache_nice_tries) { | |
3209 | #ifdef CONFIG_SCHEDSTATS | |
3210 | if (tsk_cache_hot) { | |
3211 | schedstat_inc(sd, lb_hot_gained[idle]); | |
3212 | schedstat_inc(p, se.nr_forced_migrations); | |
3213 | } | |
3214 | #endif | |
3215 | return 1; | |
3216 | } | |
3217 | ||
3218 | if (tsk_cache_hot) { | |
3219 | schedstat_inc(p, se.nr_failed_migrations_hot); | |
3220 | return 0; | |
3221 | } | |
3222 | return 1; | |
3223 | } | |
3224 | ||
3225 | static unsigned long | |
3226 | balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
3227 | unsigned long max_load_move, struct sched_domain *sd, | |
3228 | enum cpu_idle_type idle, int *all_pinned, | |
3229 | int *this_best_prio, struct rq_iterator *iterator) | |
3230 | { | |
3231 | int loops = 0, pulled = 0, pinned = 0; | |
3232 | struct task_struct *p; | |
3233 | long rem_load_move = max_load_move; | |
3234 | ||
3235 | if (max_load_move == 0) | |
3236 | goto out; | |
3237 | ||
3238 | pinned = 1; | |
3239 | ||
3240 | /* | |
3241 | * Start the load-balancing iterator: | |
3242 | */ | |
3243 | p = iterator->start(iterator->arg); | |
3244 | next: | |
3245 | if (!p || loops++ > sysctl_sched_nr_migrate) | |
3246 | goto out; | |
3247 | ||
3248 | if ((p->se.load.weight >> 1) > rem_load_move || | |
3249 | !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) { | |
3250 | p = iterator->next(iterator->arg); | |
3251 | goto next; | |
3252 | } | |
3253 | ||
3254 | pull_task(busiest, p, this_rq, this_cpu); | |
3255 | pulled++; | |
3256 | rem_load_move -= p->se.load.weight; | |
3257 | ||
3258 | #ifdef CONFIG_PREEMPT | |
3259 | /* | |
3260 | * NEWIDLE balancing is a source of latency, so preemptible kernels | |
3261 | * will stop after the first task is pulled to minimize the critical | |
3262 | * section. | |
3263 | */ | |
3264 | if (idle == CPU_NEWLY_IDLE) | |
3265 | goto out; | |
3266 | #endif | |
3267 | ||
3268 | /* | |
3269 | * We only want to steal up to the prescribed amount of weighted load. | |
3270 | */ | |
3271 | if (rem_load_move > 0) { | |
3272 | if (p->prio < *this_best_prio) | |
3273 | *this_best_prio = p->prio; | |
3274 | p = iterator->next(iterator->arg); | |
3275 | goto next; | |
3276 | } | |
3277 | out: | |
3278 | /* | |
3279 | * Right now, this is one of only two places pull_task() is called, | |
3280 | * so we can safely collect pull_task() stats here rather than | |
3281 | * inside pull_task(). | |
3282 | */ | |
3283 | schedstat_add(sd, lb_gained[idle], pulled); | |
3284 | ||
3285 | if (all_pinned) | |
3286 | *all_pinned = pinned; | |
3287 | ||
3288 | return max_load_move - rem_load_move; | |
3289 | } | |
3290 | ||
3291 | /* | |
3292 | * move_tasks tries to move up to max_load_move weighted load from busiest to | |
3293 | * this_rq, as part of a balancing operation within domain "sd". | |
3294 | * Returns 1 if successful and 0 otherwise. | |
3295 | * | |
3296 | * Called with both runqueues locked. | |
3297 | */ | |
3298 | static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
3299 | unsigned long max_load_move, | |
3300 | struct sched_domain *sd, enum cpu_idle_type idle, | |
3301 | int *all_pinned) | |
3302 | { | |
3303 | const struct sched_class *class = sched_class_highest; | |
3304 | unsigned long total_load_moved = 0; | |
3305 | int this_best_prio = this_rq->curr->prio; | |
3306 | ||
3307 | do { | |
3308 | total_load_moved += | |
3309 | class->load_balance(this_rq, this_cpu, busiest, | |
3310 | max_load_move - total_load_moved, | |
3311 | sd, idle, all_pinned, &this_best_prio); | |
3312 | class = class->next; | |
3313 | ||
3314 | #ifdef CONFIG_PREEMPT | |
3315 | /* | |
3316 | * NEWIDLE balancing is a source of latency, so preemptible | |
3317 | * kernels will stop after the first task is pulled to minimize | |
3318 | * the critical section. | |
3319 | */ | |
3320 | if (idle == CPU_NEWLY_IDLE && this_rq->nr_running) | |
3321 | break; | |
3322 | #endif | |
3323 | } while (class && max_load_move > total_load_moved); | |
3324 | ||
3325 | return total_load_moved > 0; | |
3326 | } | |
3327 | ||
3328 | static int | |
3329 | iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
3330 | struct sched_domain *sd, enum cpu_idle_type idle, | |
3331 | struct rq_iterator *iterator) | |
3332 | { | |
3333 | struct task_struct *p = iterator->start(iterator->arg); | |
3334 | int pinned = 0; | |
3335 | ||
3336 | while (p) { | |
3337 | if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) { | |
3338 | pull_task(busiest, p, this_rq, this_cpu); | |
3339 | /* | |
3340 | * Right now, this is only the second place pull_task() | |
3341 | * is called, so we can safely collect pull_task() | |
3342 | * stats here rather than inside pull_task(). | |
3343 | */ | |
3344 | schedstat_inc(sd, lb_gained[idle]); | |
3345 | ||
3346 | return 1; | |
3347 | } | |
3348 | p = iterator->next(iterator->arg); | |
3349 | } | |
3350 | ||
3351 | return 0; | |
3352 | } | |
3353 | ||
3354 | /* | |
3355 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
3356 | * part of active balancing operations within "domain". | |
3357 | * Returns 1 if successful and 0 otherwise. | |
3358 | * | |
3359 | * Called with both runqueues locked. | |
3360 | */ | |
3361 | static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
3362 | struct sched_domain *sd, enum cpu_idle_type idle) | |
3363 | { | |
3364 | const struct sched_class *class; | |
3365 | ||
3366 | for_each_class(class) { | |
3367 | if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle)) | |
3368 | return 1; | |
3369 | } | |
3370 | ||
3371 | return 0; | |
3372 | } | |
3373 | /********** Helpers for find_busiest_group ************************/ | |
3374 | /* | |
3375 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
3376 | * during load balancing. | |
3377 | */ | |
3378 | struct sd_lb_stats { | |
3379 | struct sched_group *busiest; /* Busiest group in this sd */ | |
3380 | struct sched_group *this; /* Local group in this sd */ | |
3381 | unsigned long total_load; /* Total load of all groups in sd */ | |
3382 | unsigned long total_pwr; /* Total power of all groups in sd */ | |
3383 | unsigned long avg_load; /* Average load across all groups in sd */ | |
3384 | ||
3385 | /** Statistics of this group */ | |
3386 | unsigned long this_load; | |
3387 | unsigned long this_load_per_task; | |
3388 | unsigned long this_nr_running; | |
3389 | ||
3390 | /* Statistics of the busiest group */ | |
3391 | unsigned long max_load; | |
3392 | unsigned long busiest_load_per_task; | |
3393 | unsigned long busiest_nr_running; | |
3394 | ||
3395 | int group_imb; /* Is there imbalance in this sd */ | |
3396 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
3397 | int power_savings_balance; /* Is powersave balance needed for this sd */ | |
3398 | struct sched_group *group_min; /* Least loaded group in sd */ | |
3399 | struct sched_group *group_leader; /* Group which relieves group_min */ | |
3400 | unsigned long min_load_per_task; /* load_per_task in group_min */ | |
3401 | unsigned long leader_nr_running; /* Nr running of group_leader */ | |
3402 | unsigned long min_nr_running; /* Nr running of group_min */ | |
3403 | #endif | |
3404 | }; | |
3405 | ||
3406 | /* | |
3407 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
3408 | */ | |
3409 | struct sg_lb_stats { | |
3410 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
3411 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
3412 | unsigned long sum_nr_running; /* Nr tasks running in the group */ | |
3413 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ | |
3414 | unsigned long group_capacity; | |
3415 | int group_imb; /* Is there an imbalance in the group ? */ | |
3416 | }; | |
3417 | ||
3418 | /** | |
3419 | * group_first_cpu - Returns the first cpu in the cpumask of a sched_group. | |
3420 | * @group: The group whose first cpu is to be returned. | |
3421 | */ | |
3422 | static inline unsigned int group_first_cpu(struct sched_group *group) | |
3423 | { | |
3424 | return cpumask_first(sched_group_cpus(group)); | |
3425 | } | |
3426 | ||
3427 | /** | |
3428 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
3429 | * @sd: The sched_domain whose load_idx is to be obtained. | |
3430 | * @idle: The Idle status of the CPU for whose sd load_icx is obtained. | |
3431 | */ | |
3432 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
3433 | enum cpu_idle_type idle) | |
3434 | { | |
3435 | int load_idx; | |
3436 | ||
3437 | switch (idle) { | |
3438 | case CPU_NOT_IDLE: | |
3439 | load_idx = sd->busy_idx; | |
3440 | break; | |
3441 | ||
3442 | case CPU_NEWLY_IDLE: | |
3443 | load_idx = sd->newidle_idx; | |
3444 | break; | |
3445 | default: | |
3446 | load_idx = sd->idle_idx; | |
3447 | break; | |
3448 | } | |
3449 | ||
3450 | return load_idx; | |
3451 | } | |
3452 | ||
3453 | ||
3454 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
3455 | /** | |
3456 | * init_sd_power_savings_stats - Initialize power savings statistics for | |
3457 | * the given sched_domain, during load balancing. | |
3458 | * | |
3459 | * @sd: Sched domain whose power-savings statistics are to be initialized. | |
3460 | * @sds: Variable containing the statistics for sd. | |
3461 | * @idle: Idle status of the CPU at which we're performing load-balancing. | |
3462 | */ | |
3463 | static inline void init_sd_power_savings_stats(struct sched_domain *sd, | |
3464 | struct sd_lb_stats *sds, enum cpu_idle_type idle) | |
3465 | { | |
3466 | /* | |
3467 | * Busy processors will not participate in power savings | |
3468 | * balance. | |
3469 | */ | |
3470 | if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE)) | |
3471 | sds->power_savings_balance = 0; | |
3472 | else { | |
3473 | sds->power_savings_balance = 1; | |
3474 | sds->min_nr_running = ULONG_MAX; | |
3475 | sds->leader_nr_running = 0; | |
3476 | } | |
3477 | } | |
3478 | ||
3479 | /** | |
3480 | * update_sd_power_savings_stats - Update the power saving stats for a | |
3481 | * sched_domain while performing load balancing. | |
3482 | * | |
3483 | * @group: sched_group belonging to the sched_domain under consideration. | |
3484 | * @sds: Variable containing the statistics of the sched_domain | |
3485 | * @local_group: Does group contain the CPU for which we're performing | |
3486 | * load balancing ? | |
3487 | * @sgs: Variable containing the statistics of the group. | |
3488 | */ | |
3489 | static inline void update_sd_power_savings_stats(struct sched_group *group, | |
3490 | struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs) | |
3491 | { | |
3492 | ||
3493 | if (!sds->power_savings_balance) | |
3494 | return; | |
3495 | ||
3496 | /* | |
3497 | * If the local group is idle or completely loaded | |
3498 | * no need to do power savings balance at this domain | |
3499 | */ | |
3500 | if (local_group && (sds->this_nr_running >= sgs->group_capacity || | |
3501 | !sds->this_nr_running)) | |
3502 | sds->power_savings_balance = 0; | |
3503 | ||
3504 | /* | |
3505 | * If a group is already running at full capacity or idle, | |
3506 | * don't include that group in power savings calculations | |
3507 | */ | |
3508 | if (!sds->power_savings_balance || | |
3509 | sgs->sum_nr_running >= sgs->group_capacity || | |
3510 | !sgs->sum_nr_running) | |
3511 | return; | |
3512 | ||
3513 | /* | |
3514 | * Calculate the group which has the least non-idle load. | |
3515 | * This is the group from where we need to pick up the load | |
3516 | * for saving power | |
3517 | */ | |
3518 | if ((sgs->sum_nr_running < sds->min_nr_running) || | |
3519 | (sgs->sum_nr_running == sds->min_nr_running && | |
3520 | group_first_cpu(group) > group_first_cpu(sds->group_min))) { | |
3521 | sds->group_min = group; | |
3522 | sds->min_nr_running = sgs->sum_nr_running; | |
3523 | sds->min_load_per_task = sgs->sum_weighted_load / | |
3524 | sgs->sum_nr_running; | |
3525 | } | |
3526 | ||
3527 | /* | |
3528 | * Calculate the group which is almost near its | |
3529 | * capacity but still has some space to pick up some load | |
3530 | * from other group and save more power | |
3531 | */ | |
3532 | if (sgs->sum_nr_running + 1 > sgs->group_capacity) | |
3533 | return; | |
3534 | ||
3535 | if (sgs->sum_nr_running > sds->leader_nr_running || | |
3536 | (sgs->sum_nr_running == sds->leader_nr_running && | |
3537 | group_first_cpu(group) < group_first_cpu(sds->group_leader))) { | |
3538 | sds->group_leader = group; | |
3539 | sds->leader_nr_running = sgs->sum_nr_running; | |
3540 | } | |
3541 | } | |
3542 | ||
3543 | /** | |
3544 | * check_power_save_busiest_group - see if there is potential for some power-savings balance | |
3545 | * @sds: Variable containing the statistics of the sched_domain | |
3546 | * under consideration. | |
3547 | * @this_cpu: Cpu at which we're currently performing load-balancing. | |
3548 | * @imbalance: Variable to store the imbalance. | |
3549 | * | |
3550 | * Description: | |
3551 | * Check if we have potential to perform some power-savings balance. | |
3552 | * If yes, set the busiest group to be the least loaded group in the | |
3553 | * sched_domain, so that it's CPUs can be put to idle. | |
3554 | * | |
3555 | * Returns 1 if there is potential to perform power-savings balance. | |
3556 | * Else returns 0. | |
3557 | */ | |
3558 | static inline int check_power_save_busiest_group(struct sd_lb_stats *sds, | |
3559 | int this_cpu, unsigned long *imbalance) | |
3560 | { | |
3561 | if (!sds->power_savings_balance) | |
3562 | return 0; | |
3563 | ||
3564 | if (sds->this != sds->group_leader || | |
3565 | sds->group_leader == sds->group_min) | |
3566 | return 0; | |
3567 | ||
3568 | *imbalance = sds->min_load_per_task; | |
3569 | sds->busiest = sds->group_min; | |
3570 | ||
3571 | return 1; | |
3572 | ||
3573 | } | |
3574 | #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ | |
3575 | static inline void init_sd_power_savings_stats(struct sched_domain *sd, | |
3576 | struct sd_lb_stats *sds, enum cpu_idle_type idle) | |
3577 | { | |
3578 | return; | |
3579 | } | |
3580 | ||
3581 | static inline void update_sd_power_savings_stats(struct sched_group *group, | |
3582 | struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs) | |
3583 | { | |
3584 | return; | |
3585 | } | |
3586 | ||
3587 | static inline int check_power_save_busiest_group(struct sd_lb_stats *sds, | |
3588 | int this_cpu, unsigned long *imbalance) | |
3589 | { | |
3590 | return 0; | |
3591 | } | |
3592 | #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ | |
3593 | ||
3594 | ||
3595 | unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu) | |
3596 | { | |
3597 | return SCHED_LOAD_SCALE; | |
3598 | } | |
3599 | ||
3600 | unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu) | |
3601 | { | |
3602 | return default_scale_freq_power(sd, cpu); | |
3603 | } | |
3604 | ||
3605 | unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu) | |
3606 | { | |
3607 | unsigned long weight = cpumask_weight(sched_domain_span(sd)); | |
3608 | unsigned long smt_gain = sd->smt_gain; | |
3609 | ||
3610 | smt_gain /= weight; | |
3611 | ||
3612 | return smt_gain; | |
3613 | } | |
3614 | ||
3615 | unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu) | |
3616 | { | |
3617 | return default_scale_smt_power(sd, cpu); | |
3618 | } | |
3619 | ||
3620 | unsigned long scale_rt_power(int cpu) | |
3621 | { | |
3622 | struct rq *rq = cpu_rq(cpu); | |
3623 | u64 total, available; | |
3624 | ||
3625 | sched_avg_update(rq); | |
3626 | ||
3627 | total = sched_avg_period() + (rq->clock - rq->age_stamp); | |
3628 | available = total - rq->rt_avg; | |
3629 | ||
3630 | if (unlikely((s64)total < SCHED_LOAD_SCALE)) | |
3631 | total = SCHED_LOAD_SCALE; | |
3632 | ||
3633 | total >>= SCHED_LOAD_SHIFT; | |
3634 | ||
3635 | return div_u64(available, total); | |
3636 | } | |
3637 | ||
3638 | static void update_cpu_power(struct sched_domain *sd, int cpu) | |
3639 | { | |
3640 | unsigned long weight = cpumask_weight(sched_domain_span(sd)); | |
3641 | unsigned long power = SCHED_LOAD_SCALE; | |
3642 | struct sched_group *sdg = sd->groups; | |
3643 | ||
3644 | if (sched_feat(ARCH_POWER)) | |
3645 | power *= arch_scale_freq_power(sd, cpu); | |
3646 | else | |
3647 | power *= default_scale_freq_power(sd, cpu); | |
3648 | ||
3649 | power >>= SCHED_LOAD_SHIFT; | |
3650 | ||
3651 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { | |
3652 | if (sched_feat(ARCH_POWER)) | |
3653 | power *= arch_scale_smt_power(sd, cpu); | |
3654 | else | |
3655 | power *= default_scale_smt_power(sd, cpu); | |
3656 | ||
3657 | power >>= SCHED_LOAD_SHIFT; | |
3658 | } | |
3659 | ||
3660 | power *= scale_rt_power(cpu); | |
3661 | power >>= SCHED_LOAD_SHIFT; | |
3662 | ||
3663 | if (!power) | |
3664 | power = 1; | |
3665 | ||
3666 | sdg->cpu_power = power; | |
3667 | } | |
3668 | ||
3669 | static void update_group_power(struct sched_domain *sd, int cpu) | |
3670 | { | |
3671 | struct sched_domain *child = sd->child; | |
3672 | struct sched_group *group, *sdg = sd->groups; | |
3673 | unsigned long power; | |
3674 | ||
3675 | if (!child) { | |
3676 | update_cpu_power(sd, cpu); | |
3677 | return; | |
3678 | } | |
3679 | ||
3680 | power = 0; | |
3681 | ||
3682 | group = child->groups; | |
3683 | do { | |
3684 | power += group->cpu_power; | |
3685 | group = group->next; | |
3686 | } while (group != child->groups); | |
3687 | ||
3688 | sdg->cpu_power = power; | |
3689 | } | |
3690 | ||
3691 | /** | |
3692 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
3693 | * @sd: The sched_domain whose statistics are to be updated. | |
3694 | * @group: sched_group whose statistics are to be updated. | |
3695 | * @this_cpu: Cpu for which load balance is currently performed. | |
3696 | * @idle: Idle status of this_cpu | |
3697 | * @load_idx: Load index of sched_domain of this_cpu for load calc. | |
3698 | * @sd_idle: Idle status of the sched_domain containing group. | |
3699 | * @local_group: Does group contain this_cpu. | |
3700 | * @cpus: Set of cpus considered for load balancing. | |
3701 | * @balance: Should we balance. | |
3702 | * @sgs: variable to hold the statistics for this group. | |
3703 | */ | |
3704 | static inline void update_sg_lb_stats(struct sched_domain *sd, | |
3705 | struct sched_group *group, int this_cpu, | |
3706 | enum cpu_idle_type idle, int load_idx, int *sd_idle, | |
3707 | int local_group, const struct cpumask *cpus, | |
3708 | int *balance, struct sg_lb_stats *sgs) | |
3709 | { | |
3710 | unsigned long load, max_cpu_load, min_cpu_load; | |
3711 | int i; | |
3712 | unsigned int balance_cpu = -1, first_idle_cpu = 0; | |
3713 | unsigned long sum_avg_load_per_task; | |
3714 | unsigned long avg_load_per_task; | |
3715 | ||
3716 | if (local_group) { | |
3717 | balance_cpu = group_first_cpu(group); | |
3718 | if (balance_cpu == this_cpu) | |
3719 | update_group_power(sd, this_cpu); | |
3720 | } | |
3721 | ||
3722 | /* Tally up the load of all CPUs in the group */ | |
3723 | sum_avg_load_per_task = avg_load_per_task = 0; | |
3724 | max_cpu_load = 0; | |
3725 | min_cpu_load = ~0UL; | |
3726 | ||
3727 | for_each_cpu_and(i, sched_group_cpus(group), cpus) { | |
3728 | struct rq *rq = cpu_rq(i); | |
3729 | ||
3730 | if (*sd_idle && rq->nr_running) | |
3731 | *sd_idle = 0; | |
3732 | ||
3733 | /* Bias balancing toward cpus of our domain */ | |
3734 | if (local_group) { | |
3735 | if (idle_cpu(i) && !first_idle_cpu) { | |
3736 | first_idle_cpu = 1; | |
3737 | balance_cpu = i; | |
3738 | } | |
3739 | ||
3740 | load = target_load(i, load_idx); | |
3741 | } else { | |
3742 | load = source_load(i, load_idx); | |
3743 | if (load > max_cpu_load) | |
3744 | max_cpu_load = load; | |
3745 | if (min_cpu_load > load) | |
3746 | min_cpu_load = load; | |
3747 | } | |
3748 | ||
3749 | sgs->group_load += load; | |
3750 | sgs->sum_nr_running += rq->nr_running; | |
3751 | sgs->sum_weighted_load += weighted_cpuload(i); | |
3752 | ||
3753 | sum_avg_load_per_task += cpu_avg_load_per_task(i); | |
3754 | } | |
3755 | ||
3756 | /* | |
3757 | * First idle cpu or the first cpu(busiest) in this sched group | |
3758 | * is eligible for doing load balancing at this and above | |
3759 | * domains. In the newly idle case, we will allow all the cpu's | |
3760 | * to do the newly idle load balance. | |
3761 | */ | |
3762 | if (idle != CPU_NEWLY_IDLE && local_group && | |
3763 | balance_cpu != this_cpu && balance) { | |
3764 | *balance = 0; | |
3765 | return; | |
3766 | } | |
3767 | ||
3768 | /* Adjust by relative CPU power of the group */ | |
3769 | sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
3770 | ||
3771 | ||
3772 | /* | |
3773 | * Consider the group unbalanced when the imbalance is larger | |
3774 | * than the average weight of two tasks. | |
3775 | * | |
3776 | * APZ: with cgroup the avg task weight can vary wildly and | |
3777 | * might not be a suitable number - should we keep a | |
3778 | * normalized nr_running number somewhere that negates | |
3779 | * the hierarchy? | |
3780 | */ | |
3781 | avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) / | |
3782 | group->cpu_power; | |
3783 | ||
3784 | if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task) | |
3785 | sgs->group_imb = 1; | |
3786 | ||
3787 | sgs->group_capacity = | |
3788 | DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE); | |
3789 | } | |
3790 | ||
3791 | /** | |
3792 | * update_sd_lb_stats - Update sched_group's statistics for load balancing. | |
3793 | * @sd: sched_domain whose statistics are to be updated. | |
3794 | * @this_cpu: Cpu for which load balance is currently performed. | |
3795 | * @idle: Idle status of this_cpu | |
3796 | * @sd_idle: Idle status of the sched_domain containing group. | |
3797 | * @cpus: Set of cpus considered for load balancing. | |
3798 | * @balance: Should we balance. | |
3799 | * @sds: variable to hold the statistics for this sched_domain. | |
3800 | */ | |
3801 | static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu, | |
3802 | enum cpu_idle_type idle, int *sd_idle, | |
3803 | const struct cpumask *cpus, int *balance, | |
3804 | struct sd_lb_stats *sds) | |
3805 | { | |
3806 | struct sched_domain *child = sd->child; | |
3807 | struct sched_group *group = sd->groups; | |
3808 | struct sg_lb_stats sgs; | |
3809 | int load_idx, prefer_sibling = 0; | |
3810 | ||
3811 | if (child && child->flags & SD_PREFER_SIBLING) | |
3812 | prefer_sibling = 1; | |
3813 | ||
3814 | init_sd_power_savings_stats(sd, sds, idle); | |
3815 | load_idx = get_sd_load_idx(sd, idle); | |
3816 | ||
3817 | do { | |
3818 | int local_group; | |
3819 | ||
3820 | local_group = cpumask_test_cpu(this_cpu, | |
3821 | sched_group_cpus(group)); | |
3822 | memset(&sgs, 0, sizeof(sgs)); | |
3823 | update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle, | |
3824 | local_group, cpus, balance, &sgs); | |
3825 | ||
3826 | if (local_group && balance && !(*balance)) | |
3827 | return; | |
3828 | ||
3829 | sds->total_load += sgs.group_load; | |
3830 | sds->total_pwr += group->cpu_power; | |
3831 | ||
3832 | /* | |
3833 | * In case the child domain prefers tasks go to siblings | |
3834 | * first, lower the group capacity to one so that we'll try | |
3835 | * and move all the excess tasks away. | |
3836 | */ | |
3837 | if (prefer_sibling) | |
3838 | sgs.group_capacity = min(sgs.group_capacity, 1UL); | |
3839 | ||
3840 | if (local_group) { | |
3841 | sds->this_load = sgs.avg_load; | |
3842 | sds->this = group; | |
3843 | sds->this_nr_running = sgs.sum_nr_running; | |
3844 | sds->this_load_per_task = sgs.sum_weighted_load; | |
3845 | } else if (sgs.avg_load > sds->max_load && | |
3846 | (sgs.sum_nr_running > sgs.group_capacity || | |
3847 | sgs.group_imb)) { | |
3848 | sds->max_load = sgs.avg_load; | |
3849 | sds->busiest = group; | |
3850 | sds->busiest_nr_running = sgs.sum_nr_running; | |
3851 | sds->busiest_load_per_task = sgs.sum_weighted_load; | |
3852 | sds->group_imb = sgs.group_imb; | |
3853 | } | |
3854 | ||
3855 | update_sd_power_savings_stats(group, sds, local_group, &sgs); | |
3856 | group = group->next; | |
3857 | } while (group != sd->groups); | |
3858 | } | |
3859 | ||
3860 | /** | |
3861 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
3862 | * amongst the groups of a sched_domain, during | |
3863 | * load balancing. | |
3864 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. | |
3865 | * @this_cpu: The cpu at whose sched_domain we're performing load-balance. | |
3866 | * @imbalance: Variable to store the imbalance. | |
3867 | */ | |
3868 | static inline void fix_small_imbalance(struct sd_lb_stats *sds, | |
3869 | int this_cpu, unsigned long *imbalance) | |
3870 | { | |
3871 | unsigned long tmp, pwr_now = 0, pwr_move = 0; | |
3872 | unsigned int imbn = 2; | |
3873 | ||
3874 | if (sds->this_nr_running) { | |
3875 | sds->this_load_per_task /= sds->this_nr_running; | |
3876 | if (sds->busiest_load_per_task > | |
3877 | sds->this_load_per_task) | |
3878 | imbn = 1; | |
3879 | } else | |
3880 | sds->this_load_per_task = | |
3881 | cpu_avg_load_per_task(this_cpu); | |
3882 | ||
3883 | if (sds->max_load - sds->this_load + sds->busiest_load_per_task >= | |
3884 | sds->busiest_load_per_task * imbn) { | |
3885 | *imbalance = sds->busiest_load_per_task; | |
3886 | return; | |
3887 | } | |
3888 | ||
3889 | /* | |
3890 | * OK, we don't have enough imbalance to justify moving tasks, | |
3891 | * however we may be able to increase total CPU power used by | |
3892 | * moving them. | |
3893 | */ | |
3894 | ||
3895 | pwr_now += sds->busiest->cpu_power * | |
3896 | min(sds->busiest_load_per_task, sds->max_load); | |
3897 | pwr_now += sds->this->cpu_power * | |
3898 | min(sds->this_load_per_task, sds->this_load); | |
3899 | pwr_now /= SCHED_LOAD_SCALE; | |
3900 | ||
3901 | /* Amount of load we'd subtract */ | |
3902 | tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) / | |
3903 | sds->busiest->cpu_power; | |
3904 | if (sds->max_load > tmp) | |
3905 | pwr_move += sds->busiest->cpu_power * | |
3906 | min(sds->busiest_load_per_task, sds->max_load - tmp); | |
3907 | ||
3908 | /* Amount of load we'd add */ | |
3909 | if (sds->max_load * sds->busiest->cpu_power < | |
3910 | sds->busiest_load_per_task * SCHED_LOAD_SCALE) | |
3911 | tmp = (sds->max_load * sds->busiest->cpu_power) / | |
3912 | sds->this->cpu_power; | |
3913 | else | |
3914 | tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) / | |
3915 | sds->this->cpu_power; | |
3916 | pwr_move += sds->this->cpu_power * | |
3917 | min(sds->this_load_per_task, sds->this_load + tmp); | |
3918 | pwr_move /= SCHED_LOAD_SCALE; | |
3919 | ||
3920 | /* Move if we gain throughput */ | |
3921 | if (pwr_move > pwr_now) | |
3922 | *imbalance = sds->busiest_load_per_task; | |
3923 | } | |
3924 | ||
3925 | /** | |
3926 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
3927 | * groups of a given sched_domain during load balance. | |
3928 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. | |
3929 | * @this_cpu: Cpu for which currently load balance is being performed. | |
3930 | * @imbalance: The variable to store the imbalance. | |
3931 | */ | |
3932 | static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu, | |
3933 | unsigned long *imbalance) | |
3934 | { | |
3935 | unsigned long max_pull; | |
3936 | /* | |
3937 | * In the presence of smp nice balancing, certain scenarios can have | |
3938 | * max load less than avg load(as we skip the groups at or below | |
3939 | * its cpu_power, while calculating max_load..) | |
3940 | */ | |
3941 | if (sds->max_load < sds->avg_load) { | |
3942 | *imbalance = 0; | |
3943 | return fix_small_imbalance(sds, this_cpu, imbalance); | |
3944 | } | |
3945 | ||
3946 | /* Don't want to pull so many tasks that a group would go idle */ | |
3947 | max_pull = min(sds->max_load - sds->avg_load, | |
3948 | sds->max_load - sds->busiest_load_per_task); | |
3949 | ||
3950 | /* How much load to actually move to equalise the imbalance */ | |
3951 | *imbalance = min(max_pull * sds->busiest->cpu_power, | |
3952 | (sds->avg_load - sds->this_load) * sds->this->cpu_power) | |
3953 | / SCHED_LOAD_SCALE; | |
3954 | ||
3955 | /* | |
3956 | * if *imbalance is less than the average load per runnable task | |
3957 | * there is no gaurantee that any tasks will be moved so we'll have | |
3958 | * a think about bumping its value to force at least one task to be | |
3959 | * moved | |
3960 | */ | |
3961 | if (*imbalance < sds->busiest_load_per_task) | |
3962 | return fix_small_imbalance(sds, this_cpu, imbalance); | |
3963 | ||
3964 | } | |
3965 | /******* find_busiest_group() helpers end here *********************/ | |
3966 | ||
3967 | /** | |
3968 | * find_busiest_group - Returns the busiest group within the sched_domain | |
3969 | * if there is an imbalance. If there isn't an imbalance, and | |
3970 | * the user has opted for power-savings, it returns a group whose | |
3971 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | |
3972 | * such a group exists. | |
3973 | * | |
3974 | * Also calculates the amount of weighted load which should be moved | |
3975 | * to restore balance. | |
3976 | * | |
3977 | * @sd: The sched_domain whose busiest group is to be returned. | |
3978 | * @this_cpu: The cpu for which load balancing is currently being performed. | |
3979 | * @imbalance: Variable which stores amount of weighted load which should | |
3980 | * be moved to restore balance/put a group to idle. | |
3981 | * @idle: The idle status of this_cpu. | |
3982 | * @sd_idle: The idleness of sd | |
3983 | * @cpus: The set of CPUs under consideration for load-balancing. | |
3984 | * @balance: Pointer to a variable indicating if this_cpu | |
3985 | * is the appropriate cpu to perform load balancing at this_level. | |
3986 | * | |
3987 | * Returns: - the busiest group if imbalance exists. | |
3988 | * - If no imbalance and user has opted for power-savings balance, | |
3989 | * return the least loaded group whose CPUs can be | |
3990 | * put to idle by rebalancing its tasks onto our group. | |
3991 | */ | |
3992 | static struct sched_group * | |
3993 | find_busiest_group(struct sched_domain *sd, int this_cpu, | |
3994 | unsigned long *imbalance, enum cpu_idle_type idle, | |
3995 | int *sd_idle, const struct cpumask *cpus, int *balance) | |
3996 | { | |
3997 | struct sd_lb_stats sds; | |
3998 | ||
3999 | memset(&sds, 0, sizeof(sds)); | |
4000 | ||
4001 | /* | |
4002 | * Compute the various statistics relavent for load balancing at | |
4003 | * this level. | |
4004 | */ | |
4005 | update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus, | |
4006 | balance, &sds); | |
4007 | ||
4008 | /* Cases where imbalance does not exist from POV of this_cpu */ | |
4009 | /* 1) this_cpu is not the appropriate cpu to perform load balancing | |
4010 | * at this level. | |
4011 | * 2) There is no busy sibling group to pull from. | |
4012 | * 3) This group is the busiest group. | |
4013 | * 4) This group is more busy than the avg busieness at this | |
4014 | * sched_domain. | |
4015 | * 5) The imbalance is within the specified limit. | |
4016 | * 6) Any rebalance would lead to ping-pong | |
4017 | */ | |
4018 | if (balance && !(*balance)) | |
4019 | goto ret; | |
4020 | ||
4021 | if (!sds.busiest || sds.busiest_nr_running == 0) | |
4022 | goto out_balanced; | |
4023 | ||
4024 | if (sds.this_load >= sds.max_load) | |
4025 | goto out_balanced; | |
4026 | ||
4027 | sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr; | |
4028 | ||
4029 | if (sds.this_load >= sds.avg_load) | |
4030 | goto out_balanced; | |
4031 | ||
4032 | if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load) | |
4033 | goto out_balanced; | |
4034 | ||
4035 | sds.busiest_load_per_task /= sds.busiest_nr_running; | |
4036 | if (sds.group_imb) | |
4037 | sds.busiest_load_per_task = | |
4038 | min(sds.busiest_load_per_task, sds.avg_load); | |
4039 | ||
4040 | /* | |
4041 | * We're trying to get all the cpus to the average_load, so we don't | |
4042 | * want to push ourselves above the average load, nor do we wish to | |
4043 | * reduce the max loaded cpu below the average load, as either of these | |
4044 | * actions would just result in more rebalancing later, and ping-pong | |
4045 | * tasks around. Thus we look for the minimum possible imbalance. | |
4046 | * Negative imbalances (*we* are more loaded than anyone else) will | |
4047 | * be counted as no imbalance for these purposes -- we can't fix that | |
4048 | * by pulling tasks to us. Be careful of negative numbers as they'll | |
4049 | * appear as very large values with unsigned longs. | |
4050 | */ | |
4051 | if (sds.max_load <= sds.busiest_load_per_task) | |
4052 | goto out_balanced; | |
4053 | ||
4054 | /* Looks like there is an imbalance. Compute it */ | |
4055 | calculate_imbalance(&sds, this_cpu, imbalance); | |
4056 | return sds.busiest; | |
4057 | ||
4058 | out_balanced: | |
4059 | /* | |
4060 | * There is no obvious imbalance. But check if we can do some balancing | |
4061 | * to save power. | |
4062 | */ | |
4063 | if (check_power_save_busiest_group(&sds, this_cpu, imbalance)) | |
4064 | return sds.busiest; | |
4065 | ret: | |
4066 | *imbalance = 0; | |
4067 | return NULL; | |
4068 | } | |
4069 | ||
4070 | /* | |
4071 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
4072 | */ | |
4073 | static struct rq * | |
4074 | find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle, | |
4075 | unsigned long imbalance, const struct cpumask *cpus) | |
4076 | { | |
4077 | struct rq *busiest = NULL, *rq; | |
4078 | unsigned long max_load = 0; | |
4079 | int i; | |
4080 | ||
4081 | for_each_cpu(i, sched_group_cpus(group)) { | |
4082 | unsigned long power = power_of(i); | |
4083 | unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE); | |
4084 | unsigned long wl; | |
4085 | ||
4086 | if (!cpumask_test_cpu(i, cpus)) | |
4087 | continue; | |
4088 | ||
4089 | rq = cpu_rq(i); | |
4090 | wl = weighted_cpuload(i) * SCHED_LOAD_SCALE; | |
4091 | wl /= power; | |
4092 | ||
4093 | if (capacity && rq->nr_running == 1 && wl > imbalance) | |
4094 | continue; | |
4095 | ||
4096 | if (wl > max_load) { | |
4097 | max_load = wl; | |
4098 | busiest = rq; | |
4099 | } | |
4100 | } | |
4101 | ||
4102 | return busiest; | |
4103 | } | |
4104 | ||
4105 | /* | |
4106 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
4107 | * so long as it is large enough. | |
4108 | */ | |
4109 | #define MAX_PINNED_INTERVAL 512 | |
4110 | ||
4111 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
4112 | static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask); | |
4113 | ||
4114 | /* | |
4115 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
4116 | * tasks if there is an imbalance. | |
4117 | */ | |
4118 | static int load_balance(int this_cpu, struct rq *this_rq, | |
4119 | struct sched_domain *sd, enum cpu_idle_type idle, | |
4120 | int *balance) | |
4121 | { | |
4122 | int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0; | |
4123 | struct sched_group *group; | |
4124 | unsigned long imbalance; | |
4125 | struct rq *busiest; | |
4126 | unsigned long flags; | |
4127 | struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask); | |
4128 | ||
4129 | cpumask_setall(cpus); | |
4130 | ||
4131 | /* | |
4132 | * When power savings policy is enabled for the parent domain, idle | |
4133 | * sibling can pick up load irrespective of busy siblings. In this case, | |
4134 | * let the state of idle sibling percolate up as CPU_IDLE, instead of | |
4135 | * portraying it as CPU_NOT_IDLE. | |
4136 | */ | |
4137 | if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER && | |
4138 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
4139 | sd_idle = 1; | |
4140 | ||
4141 | schedstat_inc(sd, lb_count[idle]); | |
4142 | ||
4143 | redo: | |
4144 | update_shares(sd); | |
4145 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle, | |
4146 | cpus, balance); | |
4147 | ||
4148 | if (*balance == 0) | |
4149 | goto out_balanced; | |
4150 | ||
4151 | if (!group) { | |
4152 | schedstat_inc(sd, lb_nobusyg[idle]); | |
4153 | goto out_balanced; | |
4154 | } | |
4155 | ||
4156 | busiest = find_busiest_queue(group, idle, imbalance, cpus); | |
4157 | if (!busiest) { | |
4158 | schedstat_inc(sd, lb_nobusyq[idle]); | |
4159 | goto out_balanced; | |
4160 | } | |
4161 | ||
4162 | BUG_ON(busiest == this_rq); | |
4163 | ||
4164 | schedstat_add(sd, lb_imbalance[idle], imbalance); | |
4165 | ||
4166 | ld_moved = 0; | |
4167 | if (busiest->nr_running > 1) { | |
4168 | /* | |
4169 | * Attempt to move tasks. If find_busiest_group has found | |
4170 | * an imbalance but busiest->nr_running <= 1, the group is | |
4171 | * still unbalanced. ld_moved simply stays zero, so it is | |
4172 | * correctly treated as an imbalance. | |
4173 | */ | |
4174 | local_irq_save(flags); | |
4175 | double_rq_lock(this_rq, busiest); | |
4176 | ld_moved = move_tasks(this_rq, this_cpu, busiest, | |
4177 | imbalance, sd, idle, &all_pinned); | |
4178 | double_rq_unlock(this_rq, busiest); | |
4179 | local_irq_restore(flags); | |
4180 | ||
4181 | /* | |
4182 | * some other cpu did the load balance for us. | |
4183 | */ | |
4184 | if (ld_moved && this_cpu != smp_processor_id()) | |
4185 | resched_cpu(this_cpu); | |
4186 | ||
4187 | /* All tasks on this runqueue were pinned by CPU affinity */ | |
4188 | if (unlikely(all_pinned)) { | |
4189 | cpumask_clear_cpu(cpu_of(busiest), cpus); | |
4190 | if (!cpumask_empty(cpus)) | |
4191 | goto redo; | |
4192 | goto out_balanced; | |
4193 | } | |
4194 | } | |
4195 | ||
4196 | if (!ld_moved) { | |
4197 | schedstat_inc(sd, lb_failed[idle]); | |
4198 | sd->nr_balance_failed++; | |
4199 | ||
4200 | if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) { | |
4201 | ||
4202 | spin_lock_irqsave(&busiest->lock, flags); | |
4203 | ||
4204 | /* don't kick the migration_thread, if the curr | |
4205 | * task on busiest cpu can't be moved to this_cpu | |
4206 | */ | |
4207 | if (!cpumask_test_cpu(this_cpu, | |
4208 | &busiest->curr->cpus_allowed)) { | |
4209 | spin_unlock_irqrestore(&busiest->lock, flags); | |
4210 | all_pinned = 1; | |
4211 | goto out_one_pinned; | |
4212 | } | |
4213 | ||
4214 | if (!busiest->active_balance) { | |
4215 | busiest->active_balance = 1; | |
4216 | busiest->push_cpu = this_cpu; | |
4217 | active_balance = 1; | |
4218 | } | |
4219 | spin_unlock_irqrestore(&busiest->lock, flags); | |
4220 | if (active_balance) | |
4221 | wake_up_process(busiest->migration_thread); | |
4222 | ||
4223 | /* | |
4224 | * We've kicked active balancing, reset the failure | |
4225 | * counter. | |
4226 | */ | |
4227 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
4228 | } | |
4229 | } else | |
4230 | sd->nr_balance_failed = 0; | |
4231 | ||
4232 | if (likely(!active_balance)) { | |
4233 | /* We were unbalanced, so reset the balancing interval */ | |
4234 | sd->balance_interval = sd->min_interval; | |
4235 | } else { | |
4236 | /* | |
4237 | * If we've begun active balancing, start to back off. This | |
4238 | * case may not be covered by the all_pinned logic if there | |
4239 | * is only 1 task on the busy runqueue (because we don't call | |
4240 | * move_tasks). | |
4241 | */ | |
4242 | if (sd->balance_interval < sd->max_interval) | |
4243 | sd->balance_interval *= 2; | |
4244 | } | |
4245 | ||
4246 | if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER && | |
4247 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
4248 | ld_moved = -1; | |
4249 | ||
4250 | goto out; | |
4251 | ||
4252 | out_balanced: | |
4253 | schedstat_inc(sd, lb_balanced[idle]); | |
4254 | ||
4255 | sd->nr_balance_failed = 0; | |
4256 | ||
4257 | out_one_pinned: | |
4258 | /* tune up the balancing interval */ | |
4259 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || | |
4260 | (sd->balance_interval < sd->max_interval)) | |
4261 | sd->balance_interval *= 2; | |
4262 | ||
4263 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && | |
4264 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
4265 | ld_moved = -1; | |
4266 | else | |
4267 | ld_moved = 0; | |
4268 | out: | |
4269 | if (ld_moved) | |
4270 | update_shares(sd); | |
4271 | return ld_moved; | |
4272 | } | |
4273 | ||
4274 | /* | |
4275 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
4276 | * tasks if there is an imbalance. | |
4277 | * | |
4278 | * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE). | |
4279 | * this_rq is locked. | |
4280 | */ | |
4281 | static int | |
4282 | load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd) | |
4283 | { | |
4284 | struct sched_group *group; | |
4285 | struct rq *busiest = NULL; | |
4286 | unsigned long imbalance; | |
4287 | int ld_moved = 0; | |
4288 | int sd_idle = 0; | |
4289 | int all_pinned = 0; | |
4290 | struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask); | |
4291 | ||
4292 | cpumask_setall(cpus); | |
4293 | ||
4294 | /* | |
4295 | * When power savings policy is enabled for the parent domain, idle | |
4296 | * sibling can pick up load irrespective of busy siblings. In this case, | |
4297 | * let the state of idle sibling percolate up as IDLE, instead of | |
4298 | * portraying it as CPU_NOT_IDLE. | |
4299 | */ | |
4300 | if (sd->flags & SD_SHARE_CPUPOWER && | |
4301 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
4302 | sd_idle = 1; | |
4303 | ||
4304 | schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]); | |
4305 | redo: | |
4306 | update_shares_locked(this_rq, sd); | |
4307 | group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE, | |
4308 | &sd_idle, cpus, NULL); | |
4309 | if (!group) { | |
4310 | schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]); | |
4311 | goto out_balanced; | |
4312 | } | |
4313 | ||
4314 | busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus); | |
4315 | if (!busiest) { | |
4316 | schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]); | |
4317 | goto out_balanced; | |
4318 | } | |
4319 | ||
4320 | BUG_ON(busiest == this_rq); | |
4321 | ||
4322 | schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance); | |
4323 | ||
4324 | ld_moved = 0; | |
4325 | if (busiest->nr_running > 1) { | |
4326 | /* Attempt to move tasks */ | |
4327 | double_lock_balance(this_rq, busiest); | |
4328 | /* this_rq->clock is already updated */ | |
4329 | update_rq_clock(busiest); | |
4330 | ld_moved = move_tasks(this_rq, this_cpu, busiest, | |
4331 | imbalance, sd, CPU_NEWLY_IDLE, | |
4332 | &all_pinned); | |
4333 | double_unlock_balance(this_rq, busiest); | |
4334 | ||
4335 | if (unlikely(all_pinned)) { | |
4336 | cpumask_clear_cpu(cpu_of(busiest), cpus); | |
4337 | if (!cpumask_empty(cpus)) | |
4338 | goto redo; | |
4339 | } | |
4340 | } | |
4341 | ||
4342 | if (!ld_moved) { | |
4343 | int active_balance = 0; | |
4344 | ||
4345 | schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]); | |
4346 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && | |
4347 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
4348 | return -1; | |
4349 | ||
4350 | if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP) | |
4351 | return -1; | |
4352 | ||
4353 | if (sd->nr_balance_failed++ < 2) | |
4354 | return -1; | |
4355 | ||
4356 | /* | |
4357 | * The only task running in a non-idle cpu can be moved to this | |
4358 | * cpu in an attempt to completely freeup the other CPU | |
4359 | * package. The same method used to move task in load_balance() | |
4360 | * have been extended for load_balance_newidle() to speedup | |
4361 | * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2) | |
4362 | * | |
4363 | * The package power saving logic comes from | |
4364 | * find_busiest_group(). If there are no imbalance, then | |
4365 | * f_b_g() will return NULL. However when sched_mc={1,2} then | |
4366 | * f_b_g() will select a group from which a running task may be | |
4367 | * pulled to this cpu in order to make the other package idle. | |
4368 | * If there is no opportunity to make a package idle and if | |
4369 | * there are no imbalance, then f_b_g() will return NULL and no | |
4370 | * action will be taken in load_balance_newidle(). | |
4371 | * | |
4372 | * Under normal task pull operation due to imbalance, there | |
4373 | * will be more than one task in the source run queue and | |
4374 | * move_tasks() will succeed. ld_moved will be true and this | |
4375 | * active balance code will not be triggered. | |
4376 | */ | |
4377 | ||
4378 | /* Lock busiest in correct order while this_rq is held */ | |
4379 | double_lock_balance(this_rq, busiest); | |
4380 | ||
4381 | /* | |
4382 | * don't kick the migration_thread, if the curr | |
4383 | * task on busiest cpu can't be moved to this_cpu | |
4384 | */ | |
4385 | if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) { | |
4386 | double_unlock_balance(this_rq, busiest); | |
4387 | all_pinned = 1; | |
4388 | return ld_moved; | |
4389 | } | |
4390 | ||
4391 | if (!busiest->active_balance) { | |
4392 | busiest->active_balance = 1; | |
4393 | busiest->push_cpu = this_cpu; | |
4394 | active_balance = 1; | |
4395 | } | |
4396 | ||
4397 | double_unlock_balance(this_rq, busiest); | |
4398 | /* | |
4399 | * Should not call ttwu while holding a rq->lock | |
4400 | */ | |
4401 | spin_unlock(&this_rq->lock); | |
4402 | if (active_balance) | |
4403 | wake_up_process(busiest->migration_thread); | |
4404 | spin_lock(&this_rq->lock); | |
4405 | ||
4406 | } else | |
4407 | sd->nr_balance_failed = 0; | |
4408 | ||
4409 | update_shares_locked(this_rq, sd); | |
4410 | return ld_moved; | |
4411 | ||
4412 | out_balanced: | |
4413 | schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]); | |
4414 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && | |
4415 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
4416 | return -1; | |
4417 | sd->nr_balance_failed = 0; | |
4418 | ||
4419 | return 0; | |
4420 | } | |
4421 | ||
4422 | /* | |
4423 | * idle_balance is called by schedule() if this_cpu is about to become | |
4424 | * idle. Attempts to pull tasks from other CPUs. | |
4425 | */ | |
4426 | static void idle_balance(int this_cpu, struct rq *this_rq) | |
4427 | { | |
4428 | struct sched_domain *sd; | |
4429 | int pulled_task = 0; | |
4430 | unsigned long next_balance = jiffies + HZ; | |
4431 | ||
4432 | for_each_domain(this_cpu, sd) { | |
4433 | unsigned long interval; | |
4434 | ||
4435 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
4436 | continue; | |
4437 | ||
4438 | if (sd->flags & SD_BALANCE_NEWIDLE) | |
4439 | /* If we've pulled tasks over stop searching: */ | |
4440 | pulled_task = load_balance_newidle(this_cpu, this_rq, | |
4441 | sd); | |
4442 | ||
4443 | interval = msecs_to_jiffies(sd->balance_interval); | |
4444 | if (time_after(next_balance, sd->last_balance + interval)) | |
4445 | next_balance = sd->last_balance + interval; | |
4446 | if (pulled_task) | |
4447 | break; | |
4448 | } | |
4449 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { | |
4450 | /* | |
4451 | * We are going idle. next_balance may be set based on | |
4452 | * a busy processor. So reset next_balance. | |
4453 | */ | |
4454 | this_rq->next_balance = next_balance; | |
4455 | } | |
4456 | } | |
4457 | ||
4458 | /* | |
4459 | * active_load_balance is run by migration threads. It pushes running tasks | |
4460 | * off the busiest CPU onto idle CPUs. It requires at least 1 task to be | |
4461 | * running on each physical CPU where possible, and avoids physical / | |
4462 | * logical imbalances. | |
4463 | * | |
4464 | * Called with busiest_rq locked. | |
4465 | */ | |
4466 | static void active_load_balance(struct rq *busiest_rq, int busiest_cpu) | |
4467 | { | |
4468 | int target_cpu = busiest_rq->push_cpu; | |
4469 | struct sched_domain *sd; | |
4470 | struct rq *target_rq; | |
4471 | ||
4472 | /* Is there any task to move? */ | |
4473 | if (busiest_rq->nr_running <= 1) | |
4474 | return; | |
4475 | ||
4476 | target_rq = cpu_rq(target_cpu); | |
4477 | ||
4478 | /* | |
4479 | * This condition is "impossible", if it occurs | |
4480 | * we need to fix it. Originally reported by | |
4481 | * Bjorn Helgaas on a 128-cpu setup. | |
4482 | */ | |
4483 | BUG_ON(busiest_rq == target_rq); | |
4484 | ||
4485 | /* move a task from busiest_rq to target_rq */ | |
4486 | double_lock_balance(busiest_rq, target_rq); | |
4487 | update_rq_clock(busiest_rq); | |
4488 | update_rq_clock(target_rq); | |
4489 | ||
4490 | /* Search for an sd spanning us and the target CPU. */ | |
4491 | for_each_domain(target_cpu, sd) { | |
4492 | if ((sd->flags & SD_LOAD_BALANCE) && | |
4493 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
4494 | break; | |
4495 | } | |
4496 | ||
4497 | if (likely(sd)) { | |
4498 | schedstat_inc(sd, alb_count); | |
4499 | ||
4500 | if (move_one_task(target_rq, target_cpu, busiest_rq, | |
4501 | sd, CPU_IDLE)) | |
4502 | schedstat_inc(sd, alb_pushed); | |
4503 | else | |
4504 | schedstat_inc(sd, alb_failed); | |
4505 | } | |
4506 | double_unlock_balance(busiest_rq, target_rq); | |
4507 | } | |
4508 | ||
4509 | #ifdef CONFIG_NO_HZ | |
4510 | static struct { | |
4511 | atomic_t load_balancer; | |
4512 | cpumask_var_t cpu_mask; | |
4513 | cpumask_var_t ilb_grp_nohz_mask; | |
4514 | } nohz ____cacheline_aligned = { | |
4515 | .load_balancer = ATOMIC_INIT(-1), | |
4516 | }; | |
4517 | ||
4518 | int get_nohz_load_balancer(void) | |
4519 | { | |
4520 | return atomic_read(&nohz.load_balancer); | |
4521 | } | |
4522 | ||
4523 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
4524 | /** | |
4525 | * lowest_flag_domain - Return lowest sched_domain containing flag. | |
4526 | * @cpu: The cpu whose lowest level of sched domain is to | |
4527 | * be returned. | |
4528 | * @flag: The flag to check for the lowest sched_domain | |
4529 | * for the given cpu. | |
4530 | * | |
4531 | * Returns the lowest sched_domain of a cpu which contains the given flag. | |
4532 | */ | |
4533 | static inline struct sched_domain *lowest_flag_domain(int cpu, int flag) | |
4534 | { | |
4535 | struct sched_domain *sd; | |
4536 | ||
4537 | for_each_domain(cpu, sd) | |
4538 | if (sd && (sd->flags & flag)) | |
4539 | break; | |
4540 | ||
4541 | return sd; | |
4542 | } | |
4543 | ||
4544 | /** | |
4545 | * for_each_flag_domain - Iterates over sched_domains containing the flag. | |
4546 | * @cpu: The cpu whose domains we're iterating over. | |
4547 | * @sd: variable holding the value of the power_savings_sd | |
4548 | * for cpu. | |
4549 | * @flag: The flag to filter the sched_domains to be iterated. | |
4550 | * | |
4551 | * Iterates over all the scheduler domains for a given cpu that has the 'flag' | |
4552 | * set, starting from the lowest sched_domain to the highest. | |
4553 | */ | |
4554 | #define for_each_flag_domain(cpu, sd, flag) \ | |
4555 | for (sd = lowest_flag_domain(cpu, flag); \ | |
4556 | (sd && (sd->flags & flag)); sd = sd->parent) | |
4557 | ||
4558 | /** | |
4559 | * is_semi_idle_group - Checks if the given sched_group is semi-idle. | |
4560 | * @ilb_group: group to be checked for semi-idleness | |
4561 | * | |
4562 | * Returns: 1 if the group is semi-idle. 0 otherwise. | |
4563 | * | |
4564 | * We define a sched_group to be semi idle if it has atleast one idle-CPU | |
4565 | * and atleast one non-idle CPU. This helper function checks if the given | |
4566 | * sched_group is semi-idle or not. | |
4567 | */ | |
4568 | static inline int is_semi_idle_group(struct sched_group *ilb_group) | |
4569 | { | |
4570 | cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask, | |
4571 | sched_group_cpus(ilb_group)); | |
4572 | ||
4573 | /* | |
4574 | * A sched_group is semi-idle when it has atleast one busy cpu | |
4575 | * and atleast one idle cpu. | |
4576 | */ | |
4577 | if (cpumask_empty(nohz.ilb_grp_nohz_mask)) | |
4578 | return 0; | |
4579 | ||
4580 | if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group))) | |
4581 | return 0; | |
4582 | ||
4583 | return 1; | |
4584 | } | |
4585 | /** | |
4586 | * find_new_ilb - Finds the optimum idle load balancer for nomination. | |
4587 | * @cpu: The cpu which is nominating a new idle_load_balancer. | |
4588 | * | |
4589 | * Returns: Returns the id of the idle load balancer if it exists, | |
4590 | * Else, returns >= nr_cpu_ids. | |
4591 | * | |
4592 | * This algorithm picks the idle load balancer such that it belongs to a | |
4593 | * semi-idle powersavings sched_domain. The idea is to try and avoid | |
4594 | * completely idle packages/cores just for the purpose of idle load balancing | |
4595 | * when there are other idle cpu's which are better suited for that job. | |
4596 | */ | |
4597 | static int find_new_ilb(int cpu) | |
4598 | { | |
4599 | struct sched_domain *sd; | |
4600 | struct sched_group *ilb_group; | |
4601 | ||
4602 | /* | |
4603 | * Have idle load balancer selection from semi-idle packages only | |
4604 | * when power-aware load balancing is enabled | |
4605 | */ | |
4606 | if (!(sched_smt_power_savings || sched_mc_power_savings)) | |
4607 | goto out_done; | |
4608 | ||
4609 | /* | |
4610 | * Optimize for the case when we have no idle CPUs or only one | |
4611 | * idle CPU. Don't walk the sched_domain hierarchy in such cases | |
4612 | */ | |
4613 | if (cpumask_weight(nohz.cpu_mask) < 2) | |
4614 | goto out_done; | |
4615 | ||
4616 | for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) { | |
4617 | ilb_group = sd->groups; | |
4618 | ||
4619 | do { | |
4620 | if (is_semi_idle_group(ilb_group)) | |
4621 | return cpumask_first(nohz.ilb_grp_nohz_mask); | |
4622 | ||
4623 | ilb_group = ilb_group->next; | |
4624 | ||
4625 | } while (ilb_group != sd->groups); | |
4626 | } | |
4627 | ||
4628 | out_done: | |
4629 | return cpumask_first(nohz.cpu_mask); | |
4630 | } | |
4631 | #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */ | |
4632 | static inline int find_new_ilb(int call_cpu) | |
4633 | { | |
4634 | return cpumask_first(nohz.cpu_mask); | |
4635 | } | |
4636 | #endif | |
4637 | ||
4638 | /* | |
4639 | * This routine will try to nominate the ilb (idle load balancing) | |
4640 | * owner among the cpus whose ticks are stopped. ilb owner will do the idle | |
4641 | * load balancing on behalf of all those cpus. If all the cpus in the system | |
4642 | * go into this tickless mode, then there will be no ilb owner (as there is | |
4643 | * no need for one) and all the cpus will sleep till the next wakeup event | |
4644 | * arrives... | |
4645 | * | |
4646 | * For the ilb owner, tick is not stopped. And this tick will be used | |
4647 | * for idle load balancing. ilb owner will still be part of | |
4648 | * nohz.cpu_mask.. | |
4649 | * | |
4650 | * While stopping the tick, this cpu will become the ilb owner if there | |
4651 | * is no other owner. And will be the owner till that cpu becomes busy | |
4652 | * or if all cpus in the system stop their ticks at which point | |
4653 | * there is no need for ilb owner. | |
4654 | * | |
4655 | * When the ilb owner becomes busy, it nominates another owner, during the | |
4656 | * next busy scheduler_tick() | |
4657 | */ | |
4658 | int select_nohz_load_balancer(int stop_tick) | |
4659 | { | |
4660 | int cpu = smp_processor_id(); | |
4661 | ||
4662 | if (stop_tick) { | |
4663 | cpu_rq(cpu)->in_nohz_recently = 1; | |
4664 | ||
4665 | if (!cpu_active(cpu)) { | |
4666 | if (atomic_read(&nohz.load_balancer) != cpu) | |
4667 | return 0; | |
4668 | ||
4669 | /* | |
4670 | * If we are going offline and still the leader, | |
4671 | * give up! | |
4672 | */ | |
4673 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu) | |
4674 | BUG(); | |
4675 | ||
4676 | return 0; | |
4677 | } | |
4678 | ||
4679 | cpumask_set_cpu(cpu, nohz.cpu_mask); | |
4680 | ||
4681 | /* time for ilb owner also to sleep */ | |
4682 | if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) { | |
4683 | if (atomic_read(&nohz.load_balancer) == cpu) | |
4684 | atomic_set(&nohz.load_balancer, -1); | |
4685 | return 0; | |
4686 | } | |
4687 | ||
4688 | if (atomic_read(&nohz.load_balancer) == -1) { | |
4689 | /* make me the ilb owner */ | |
4690 | if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1) | |
4691 | return 1; | |
4692 | } else if (atomic_read(&nohz.load_balancer) == cpu) { | |
4693 | int new_ilb; | |
4694 | ||
4695 | if (!(sched_smt_power_savings || | |
4696 | sched_mc_power_savings)) | |
4697 | return 1; | |
4698 | /* | |
4699 | * Check to see if there is a more power-efficient | |
4700 | * ilb. | |
4701 | */ | |
4702 | new_ilb = find_new_ilb(cpu); | |
4703 | if (new_ilb < nr_cpu_ids && new_ilb != cpu) { | |
4704 | atomic_set(&nohz.load_balancer, -1); | |
4705 | resched_cpu(new_ilb); | |
4706 | return 0; | |
4707 | } | |
4708 | return 1; | |
4709 | } | |
4710 | } else { | |
4711 | if (!cpumask_test_cpu(cpu, nohz.cpu_mask)) | |
4712 | return 0; | |
4713 | ||
4714 | cpumask_clear_cpu(cpu, nohz.cpu_mask); | |
4715 | ||
4716 | if (atomic_read(&nohz.load_balancer) == cpu) | |
4717 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu) | |
4718 | BUG(); | |
4719 | } | |
4720 | return 0; | |
4721 | } | |
4722 | #endif | |
4723 | ||
4724 | static DEFINE_SPINLOCK(balancing); | |
4725 | ||
4726 | /* | |
4727 | * It checks each scheduling domain to see if it is due to be balanced, | |
4728 | * and initiates a balancing operation if so. | |
4729 | * | |
4730 | * Balancing parameters are set up in arch_init_sched_domains. | |
4731 | */ | |
4732 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) | |
4733 | { | |
4734 | int balance = 1; | |
4735 | struct rq *rq = cpu_rq(cpu); | |
4736 | unsigned long interval; | |
4737 | struct sched_domain *sd; | |
4738 | /* Earliest time when we have to do rebalance again */ | |
4739 | unsigned long next_balance = jiffies + 60*HZ; | |
4740 | int update_next_balance = 0; | |
4741 | int need_serialize; | |
4742 | ||
4743 | for_each_domain(cpu, sd) { | |
4744 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
4745 | continue; | |
4746 | ||
4747 | interval = sd->balance_interval; | |
4748 | if (idle != CPU_IDLE) | |
4749 | interval *= sd->busy_factor; | |
4750 | ||
4751 | /* scale ms to jiffies */ | |
4752 | interval = msecs_to_jiffies(interval); | |
4753 | if (unlikely(!interval)) | |
4754 | interval = 1; | |
4755 | if (interval > HZ*NR_CPUS/10) | |
4756 | interval = HZ*NR_CPUS/10; | |
4757 | ||
4758 | need_serialize = sd->flags & SD_SERIALIZE; | |
4759 | ||
4760 | if (need_serialize) { | |
4761 | if (!spin_trylock(&balancing)) | |
4762 | goto out; | |
4763 | } | |
4764 | ||
4765 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
4766 | if (load_balance(cpu, rq, sd, idle, &balance)) { | |
4767 | /* | |
4768 | * We've pulled tasks over so either we're no | |
4769 | * longer idle, or one of our SMT siblings is | |
4770 | * not idle. | |
4771 | */ | |
4772 | idle = CPU_NOT_IDLE; | |
4773 | } | |
4774 | sd->last_balance = jiffies; | |
4775 | } | |
4776 | if (need_serialize) | |
4777 | spin_unlock(&balancing); | |
4778 | out: | |
4779 | if (time_after(next_balance, sd->last_balance + interval)) { | |
4780 | next_balance = sd->last_balance + interval; | |
4781 | update_next_balance = 1; | |
4782 | } | |
4783 | ||
4784 | /* | |
4785 | * Stop the load balance at this level. There is another | |
4786 | * CPU in our sched group which is doing load balancing more | |
4787 | * actively. | |
4788 | */ | |
4789 | if (!balance) | |
4790 | break; | |
4791 | } | |
4792 | ||
4793 | /* | |
4794 | * next_balance will be updated only when there is a need. | |
4795 | * When the cpu is attached to null domain for ex, it will not be | |
4796 | * updated. | |
4797 | */ | |
4798 | if (likely(update_next_balance)) | |
4799 | rq->next_balance = next_balance; | |
4800 | } | |
4801 | ||
4802 | /* | |
4803 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
4804 | * In CONFIG_NO_HZ case, the idle load balance owner will do the | |
4805 | * rebalancing for all the cpus for whom scheduler ticks are stopped. | |
4806 | */ | |
4807 | static void run_rebalance_domains(struct softirq_action *h) | |
4808 | { | |
4809 | int this_cpu = smp_processor_id(); | |
4810 | struct rq *this_rq = cpu_rq(this_cpu); | |
4811 | enum cpu_idle_type idle = this_rq->idle_at_tick ? | |
4812 | CPU_IDLE : CPU_NOT_IDLE; | |
4813 | ||
4814 | rebalance_domains(this_cpu, idle); | |
4815 | ||
4816 | #ifdef CONFIG_NO_HZ | |
4817 | /* | |
4818 | * If this cpu is the owner for idle load balancing, then do the | |
4819 | * balancing on behalf of the other idle cpus whose ticks are | |
4820 | * stopped. | |
4821 | */ | |
4822 | if (this_rq->idle_at_tick && | |
4823 | atomic_read(&nohz.load_balancer) == this_cpu) { | |
4824 | struct rq *rq; | |
4825 | int balance_cpu; | |
4826 | ||
4827 | for_each_cpu(balance_cpu, nohz.cpu_mask) { | |
4828 | if (balance_cpu == this_cpu) | |
4829 | continue; | |
4830 | ||
4831 | /* | |
4832 | * If this cpu gets work to do, stop the load balancing | |
4833 | * work being done for other cpus. Next load | |
4834 | * balancing owner will pick it up. | |
4835 | */ | |
4836 | if (need_resched()) | |
4837 | break; | |
4838 | ||
4839 | rebalance_domains(balance_cpu, CPU_IDLE); | |
4840 | ||
4841 | rq = cpu_rq(balance_cpu); | |
4842 | if (time_after(this_rq->next_balance, rq->next_balance)) | |
4843 | this_rq->next_balance = rq->next_balance; | |
4844 | } | |
4845 | } | |
4846 | #endif | |
4847 | } | |
4848 | ||
4849 | static inline int on_null_domain(int cpu) | |
4850 | { | |
4851 | return !rcu_dereference(cpu_rq(cpu)->sd); | |
4852 | } | |
4853 | ||
4854 | /* | |
4855 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
4856 | * | |
4857 | * In case of CONFIG_NO_HZ, this is the place where we nominate a new | |
4858 | * idle load balancing owner or decide to stop the periodic load balancing, | |
4859 | * if the whole system is idle. | |
4860 | */ | |
4861 | static inline void trigger_load_balance(struct rq *rq, int cpu) | |
4862 | { | |
4863 | #ifdef CONFIG_NO_HZ | |
4864 | /* | |
4865 | * If we were in the nohz mode recently and busy at the current | |
4866 | * scheduler tick, then check if we need to nominate new idle | |
4867 | * load balancer. | |
4868 | */ | |
4869 | if (rq->in_nohz_recently && !rq->idle_at_tick) { | |
4870 | rq->in_nohz_recently = 0; | |
4871 | ||
4872 | if (atomic_read(&nohz.load_balancer) == cpu) { | |
4873 | cpumask_clear_cpu(cpu, nohz.cpu_mask); | |
4874 | atomic_set(&nohz.load_balancer, -1); | |
4875 | } | |
4876 | ||
4877 | if (atomic_read(&nohz.load_balancer) == -1) { | |
4878 | int ilb = find_new_ilb(cpu); | |
4879 | ||
4880 | if (ilb < nr_cpu_ids) | |
4881 | resched_cpu(ilb); | |
4882 | } | |
4883 | } | |
4884 | ||
4885 | /* | |
4886 | * If this cpu is idle and doing idle load balancing for all the | |
4887 | * cpus with ticks stopped, is it time for that to stop? | |
4888 | */ | |
4889 | if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu && | |
4890 | cpumask_weight(nohz.cpu_mask) == num_online_cpus()) { | |
4891 | resched_cpu(cpu); | |
4892 | return; | |
4893 | } | |
4894 | ||
4895 | /* | |
4896 | * If this cpu is idle and the idle load balancing is done by | |
4897 | * someone else, then no need raise the SCHED_SOFTIRQ | |
4898 | */ | |
4899 | if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu && | |
4900 | cpumask_test_cpu(cpu, nohz.cpu_mask)) | |
4901 | return; | |
4902 | #endif | |
4903 | /* Don't need to rebalance while attached to NULL domain */ | |
4904 | if (time_after_eq(jiffies, rq->next_balance) && | |
4905 | likely(!on_null_domain(cpu))) | |
4906 | raise_softirq(SCHED_SOFTIRQ); | |
4907 | } | |
4908 | ||
4909 | #else /* CONFIG_SMP */ | |
4910 | ||
4911 | /* | |
4912 | * on UP we do not need to balance between CPUs: | |
4913 | */ | |
4914 | static inline void idle_balance(int cpu, struct rq *rq) | |
4915 | { | |
4916 | } | |
4917 | ||
4918 | #endif | |
4919 | ||
4920 | DEFINE_PER_CPU(struct kernel_stat, kstat); | |
4921 | ||
4922 | EXPORT_PER_CPU_SYMBOL(kstat); | |
4923 | ||
4924 | /* | |
4925 | * Return any ns on the sched_clock that have not yet been accounted in | |
4926 | * @p in case that task is currently running. | |
4927 | * | |
4928 | * Called with task_rq_lock() held on @rq. | |
4929 | */ | |
4930 | static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) | |
4931 | { | |
4932 | u64 ns = 0; | |
4933 | ||
4934 | if (task_current(rq, p)) { | |
4935 | update_rq_clock(rq); | |
4936 | ns = rq->clock - p->se.exec_start; | |
4937 | if ((s64)ns < 0) | |
4938 | ns = 0; | |
4939 | } | |
4940 | ||
4941 | return ns; | |
4942 | } | |
4943 | ||
4944 | unsigned long long task_delta_exec(struct task_struct *p) | |
4945 | { | |
4946 | unsigned long flags; | |
4947 | struct rq *rq; | |
4948 | u64 ns = 0; | |
4949 | ||
4950 | rq = task_rq_lock(p, &flags); | |
4951 | ns = do_task_delta_exec(p, rq); | |
4952 | task_rq_unlock(rq, &flags); | |
4953 | ||
4954 | return ns; | |
4955 | } | |
4956 | ||
4957 | /* | |
4958 | * Return accounted runtime for the task. | |
4959 | * In case the task is currently running, return the runtime plus current's | |
4960 | * pending runtime that have not been accounted yet. | |
4961 | */ | |
4962 | unsigned long long task_sched_runtime(struct task_struct *p) | |
4963 | { | |
4964 | unsigned long flags; | |
4965 | struct rq *rq; | |
4966 | u64 ns = 0; | |
4967 | ||
4968 | rq = task_rq_lock(p, &flags); | |
4969 | ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); | |
4970 | task_rq_unlock(rq, &flags); | |
4971 | ||
4972 | return ns; | |
4973 | } | |
4974 | ||
4975 | /* | |
4976 | * Return sum_exec_runtime for the thread group. | |
4977 | * In case the task is currently running, return the sum plus current's | |
4978 | * pending runtime that have not been accounted yet. | |
4979 | * | |
4980 | * Note that the thread group might have other running tasks as well, | |
4981 | * so the return value not includes other pending runtime that other | |
4982 | * running tasks might have. | |
4983 | */ | |
4984 | unsigned long long thread_group_sched_runtime(struct task_struct *p) | |
4985 | { | |
4986 | struct task_cputime totals; | |
4987 | unsigned long flags; | |
4988 | struct rq *rq; | |
4989 | u64 ns; | |
4990 | ||
4991 | rq = task_rq_lock(p, &flags); | |
4992 | thread_group_cputime(p, &totals); | |
4993 | ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq); | |
4994 | task_rq_unlock(rq, &flags); | |
4995 | ||
4996 | return ns; | |
4997 | } | |
4998 | ||
4999 | /* | |
5000 | * Account user cpu time to a process. | |
5001 | * @p: the process that the cpu time gets accounted to | |
5002 | * @cputime: the cpu time spent in user space since the last update | |
5003 | * @cputime_scaled: cputime scaled by cpu frequency | |
5004 | */ | |
5005 | void account_user_time(struct task_struct *p, cputime_t cputime, | |
5006 | cputime_t cputime_scaled) | |
5007 | { | |
5008 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
5009 | cputime64_t tmp; | |
5010 | ||
5011 | /* Add user time to process. */ | |
5012 | p->utime = cputime_add(p->utime, cputime); | |
5013 | p->utimescaled = cputime_add(p->utimescaled, cputime_scaled); | |
5014 | account_group_user_time(p, cputime); | |
5015 | ||
5016 | /* Add user time to cpustat. */ | |
5017 | tmp = cputime_to_cputime64(cputime); | |
5018 | if (TASK_NICE(p) > 0) | |
5019 | cpustat->nice = cputime64_add(cpustat->nice, tmp); | |
5020 | else | |
5021 | cpustat->user = cputime64_add(cpustat->user, tmp); | |
5022 | ||
5023 | cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime); | |
5024 | /* Account for user time used */ | |
5025 | acct_update_integrals(p); | |
5026 | } | |
5027 | ||
5028 | /* | |
5029 | * Account guest cpu time to a process. | |
5030 | * @p: the process that the cpu time gets accounted to | |
5031 | * @cputime: the cpu time spent in virtual machine since the last update | |
5032 | * @cputime_scaled: cputime scaled by cpu frequency | |
5033 | */ | |
5034 | static void account_guest_time(struct task_struct *p, cputime_t cputime, | |
5035 | cputime_t cputime_scaled) | |
5036 | { | |
5037 | cputime64_t tmp; | |
5038 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
5039 | ||
5040 | tmp = cputime_to_cputime64(cputime); | |
5041 | ||
5042 | /* Add guest time to process. */ | |
5043 | p->utime = cputime_add(p->utime, cputime); | |
5044 | p->utimescaled = cputime_add(p->utimescaled, cputime_scaled); | |
5045 | account_group_user_time(p, cputime); | |
5046 | p->gtime = cputime_add(p->gtime, cputime); | |
5047 | ||
5048 | /* Add guest time to cpustat. */ | |
5049 | cpustat->user = cputime64_add(cpustat->user, tmp); | |
5050 | cpustat->guest = cputime64_add(cpustat->guest, tmp); | |
5051 | } | |
5052 | ||
5053 | /* | |
5054 | * Account system cpu time to a process. | |
5055 | * @p: the process that the cpu time gets accounted to | |
5056 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
5057 | * @cputime: the cpu time spent in kernel space since the last update | |
5058 | * @cputime_scaled: cputime scaled by cpu frequency | |
5059 | */ | |
5060 | void account_system_time(struct task_struct *p, int hardirq_offset, | |
5061 | cputime_t cputime, cputime_t cputime_scaled) | |
5062 | { | |
5063 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
5064 | cputime64_t tmp; | |
5065 | ||
5066 | if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) { | |
5067 | account_guest_time(p, cputime, cputime_scaled); | |
5068 | return; | |
5069 | } | |
5070 | ||
5071 | /* Add system time to process. */ | |
5072 | p->stime = cputime_add(p->stime, cputime); | |
5073 | p->stimescaled = cputime_add(p->stimescaled, cputime_scaled); | |
5074 | account_group_system_time(p, cputime); | |
5075 | ||
5076 | /* Add system time to cpustat. */ | |
5077 | tmp = cputime_to_cputime64(cputime); | |
5078 | if (hardirq_count() - hardirq_offset) | |
5079 | cpustat->irq = cputime64_add(cpustat->irq, tmp); | |
5080 | else if (softirq_count()) | |
5081 | cpustat->softirq = cputime64_add(cpustat->softirq, tmp); | |
5082 | else | |
5083 | cpustat->system = cputime64_add(cpustat->system, tmp); | |
5084 | ||
5085 | cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime); | |
5086 | ||
5087 | /* Account for system time used */ | |
5088 | acct_update_integrals(p); | |
5089 | } | |
5090 | ||
5091 | /* | |
5092 | * Account for involuntary wait time. | |
5093 | * @steal: the cpu time spent in involuntary wait | |
5094 | */ | |
5095 | void account_steal_time(cputime_t cputime) | |
5096 | { | |
5097 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
5098 | cputime64_t cputime64 = cputime_to_cputime64(cputime); | |
5099 | ||
5100 | cpustat->steal = cputime64_add(cpustat->steal, cputime64); | |
5101 | } | |
5102 | ||
5103 | /* | |
5104 | * Account for idle time. | |
5105 | * @cputime: the cpu time spent in idle wait | |
5106 | */ | |
5107 | void account_idle_time(cputime_t cputime) | |
5108 | { | |
5109 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
5110 | cputime64_t cputime64 = cputime_to_cputime64(cputime); | |
5111 | struct rq *rq = this_rq(); | |
5112 | ||
5113 | if (atomic_read(&rq->nr_iowait) > 0) | |
5114 | cpustat->iowait = cputime64_add(cpustat->iowait, cputime64); | |
5115 | else | |
5116 | cpustat->idle = cputime64_add(cpustat->idle, cputime64); | |
5117 | } | |
5118 | ||
5119 | #ifndef CONFIG_VIRT_CPU_ACCOUNTING | |
5120 | ||
5121 | /* | |
5122 | * Account a single tick of cpu time. | |
5123 | * @p: the process that the cpu time gets accounted to | |
5124 | * @user_tick: indicates if the tick is a user or a system tick | |
5125 | */ | |
5126 | void account_process_tick(struct task_struct *p, int user_tick) | |
5127 | { | |
5128 | cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy); | |
5129 | struct rq *rq = this_rq(); | |
5130 | ||
5131 | if (user_tick) | |
5132 | account_user_time(p, cputime_one_jiffy, one_jiffy_scaled); | |
5133 | else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET)) | |
5134 | account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy, | |
5135 | one_jiffy_scaled); | |
5136 | else | |
5137 | account_idle_time(cputime_one_jiffy); | |
5138 | } | |
5139 | ||
5140 | /* | |
5141 | * Account multiple ticks of steal time. | |
5142 | * @p: the process from which the cpu time has been stolen | |
5143 | * @ticks: number of stolen ticks | |
5144 | */ | |
5145 | void account_steal_ticks(unsigned long ticks) | |
5146 | { | |
5147 | account_steal_time(jiffies_to_cputime(ticks)); | |
5148 | } | |
5149 | ||
5150 | /* | |
5151 | * Account multiple ticks of idle time. | |
5152 | * @ticks: number of stolen ticks | |
5153 | */ | |
5154 | void account_idle_ticks(unsigned long ticks) | |
5155 | { | |
5156 | account_idle_time(jiffies_to_cputime(ticks)); | |
5157 | } | |
5158 | ||
5159 | #endif | |
5160 | ||
5161 | /* | |
5162 | * Use precise platform statistics if available: | |
5163 | */ | |
5164 | #ifdef CONFIG_VIRT_CPU_ACCOUNTING | |
5165 | cputime_t task_utime(struct task_struct *p) | |
5166 | { | |
5167 | return p->utime; | |
5168 | } | |
5169 | ||
5170 | cputime_t task_stime(struct task_struct *p) | |
5171 | { | |
5172 | return p->stime; | |
5173 | } | |
5174 | #else | |
5175 | cputime_t task_utime(struct task_struct *p) | |
5176 | { | |
5177 | clock_t utime = cputime_to_clock_t(p->utime), | |
5178 | total = utime + cputime_to_clock_t(p->stime); | |
5179 | u64 temp; | |
5180 | ||
5181 | /* | |
5182 | * Use CFS's precise accounting: | |
5183 | */ | |
5184 | temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime); | |
5185 | ||
5186 | if (total) { | |
5187 | temp *= utime; | |
5188 | do_div(temp, total); | |
5189 | } | |
5190 | utime = (clock_t)temp; | |
5191 | ||
5192 | p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime)); | |
5193 | return p->prev_utime; | |
5194 | } | |
5195 | ||
5196 | cputime_t task_stime(struct task_struct *p) | |
5197 | { | |
5198 | clock_t stime; | |
5199 | ||
5200 | /* | |
5201 | * Use CFS's precise accounting. (we subtract utime from | |
5202 | * the total, to make sure the total observed by userspace | |
5203 | * grows monotonically - apps rely on that): | |
5204 | */ | |
5205 | stime = nsec_to_clock_t(p->se.sum_exec_runtime) - | |
5206 | cputime_to_clock_t(task_utime(p)); | |
5207 | ||
5208 | if (stime >= 0) | |
5209 | p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime)); | |
5210 | ||
5211 | return p->prev_stime; | |
5212 | } | |
5213 | #endif | |
5214 | ||
5215 | inline cputime_t task_gtime(struct task_struct *p) | |
5216 | { | |
5217 | return p->gtime; | |
5218 | } | |
5219 | ||
5220 | /* | |
5221 | * This function gets called by the timer code, with HZ frequency. | |
5222 | * We call it with interrupts disabled. | |
5223 | * | |
5224 | * It also gets called by the fork code, when changing the parent's | |
5225 | * timeslices. | |
5226 | */ | |
5227 | void scheduler_tick(void) | |
5228 | { | |
5229 | int cpu = smp_processor_id(); | |
5230 | struct rq *rq = cpu_rq(cpu); | |
5231 | struct task_struct *curr = rq->curr; | |
5232 | ||
5233 | sched_clock_tick(); | |
5234 | ||
5235 | spin_lock(&rq->lock); | |
5236 | update_rq_clock(rq); | |
5237 | update_cpu_load(rq); | |
5238 | curr->sched_class->task_tick(rq, curr, 0); | |
5239 | spin_unlock(&rq->lock); | |
5240 | ||
5241 | perf_event_task_tick(curr, cpu); | |
5242 | ||
5243 | #ifdef CONFIG_SMP | |
5244 | rq->idle_at_tick = idle_cpu(cpu); | |
5245 | trigger_load_balance(rq, cpu); | |
5246 | #endif | |
5247 | } | |
5248 | ||
5249 | notrace unsigned long get_parent_ip(unsigned long addr) | |
5250 | { | |
5251 | if (in_lock_functions(addr)) { | |
5252 | addr = CALLER_ADDR2; | |
5253 | if (in_lock_functions(addr)) | |
5254 | addr = CALLER_ADDR3; | |
5255 | } | |
5256 | return addr; | |
5257 | } | |
5258 | ||
5259 | #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ | |
5260 | defined(CONFIG_PREEMPT_TRACER)) | |
5261 | ||
5262 | void __kprobes add_preempt_count(int val) | |
5263 | { | |
5264 | #ifdef CONFIG_DEBUG_PREEMPT | |
5265 | /* | |
5266 | * Underflow? | |
5267 | */ | |
5268 | if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) | |
5269 | return; | |
5270 | #endif | |
5271 | preempt_count() += val; | |
5272 | #ifdef CONFIG_DEBUG_PREEMPT | |
5273 | /* | |
5274 | * Spinlock count overflowing soon? | |
5275 | */ | |
5276 | DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= | |
5277 | PREEMPT_MASK - 10); | |
5278 | #endif | |
5279 | if (preempt_count() == val) | |
5280 | trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); | |
5281 | } | |
5282 | EXPORT_SYMBOL(add_preempt_count); | |
5283 | ||
5284 | void __kprobes sub_preempt_count(int val) | |
5285 | { | |
5286 | #ifdef CONFIG_DEBUG_PREEMPT | |
5287 | /* | |
5288 | * Underflow? | |
5289 | */ | |
5290 | if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) | |
5291 | return; | |
5292 | /* | |
5293 | * Is the spinlock portion underflowing? | |
5294 | */ | |
5295 | if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && | |
5296 | !(preempt_count() & PREEMPT_MASK))) | |
5297 | return; | |
5298 | #endif | |
5299 | ||
5300 | if (preempt_count() == val) | |
5301 | trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); | |
5302 | preempt_count() -= val; | |
5303 | } | |
5304 | EXPORT_SYMBOL(sub_preempt_count); | |
5305 | ||
5306 | #endif | |
5307 | ||
5308 | /* | |
5309 | * Print scheduling while atomic bug: | |
5310 | */ | |
5311 | static noinline void __schedule_bug(struct task_struct *prev) | |
5312 | { | |
5313 | struct pt_regs *regs = get_irq_regs(); | |
5314 | ||
5315 | printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", | |
5316 | prev->comm, prev->pid, preempt_count()); | |
5317 | ||
5318 | debug_show_held_locks(prev); | |
5319 | print_modules(); | |
5320 | if (irqs_disabled()) | |
5321 | print_irqtrace_events(prev); | |
5322 | ||
5323 | if (regs) | |
5324 | show_regs(regs); | |
5325 | else | |
5326 | dump_stack(); | |
5327 | } | |
5328 | ||
5329 | /* | |
5330 | * Various schedule()-time debugging checks and statistics: | |
5331 | */ | |
5332 | static inline void schedule_debug(struct task_struct *prev) | |
5333 | { | |
5334 | /* | |
5335 | * Test if we are atomic. Since do_exit() needs to call into | |
5336 | * schedule() atomically, we ignore that path for now. | |
5337 | * Otherwise, whine if we are scheduling when we should not be. | |
5338 | */ | |
5339 | if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) | |
5340 | __schedule_bug(prev); | |
5341 | ||
5342 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); | |
5343 | ||
5344 | schedstat_inc(this_rq(), sched_count); | |
5345 | #ifdef CONFIG_SCHEDSTATS | |
5346 | if (unlikely(prev->lock_depth >= 0)) { | |
5347 | schedstat_inc(this_rq(), bkl_count); | |
5348 | schedstat_inc(prev, sched_info.bkl_count); | |
5349 | } | |
5350 | #endif | |
5351 | } | |
5352 | ||
5353 | static void put_prev_task(struct rq *rq, struct task_struct *p) | |
5354 | { | |
5355 | u64 runtime = p->se.sum_exec_runtime - p->se.prev_sum_exec_runtime; | |
5356 | ||
5357 | update_avg(&p->se.avg_running, runtime); | |
5358 | ||
5359 | if (p->state == TASK_RUNNING) { | |
5360 | /* | |
5361 | * In order to avoid avg_overlap growing stale when we are | |
5362 | * indeed overlapping and hence not getting put to sleep, grow | |
5363 | * the avg_overlap on preemption. | |
5364 | * | |
5365 | * We use the average preemption runtime because that | |
5366 | * correlates to the amount of cache footprint a task can | |
5367 | * build up. | |
5368 | */ | |
5369 | runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost); | |
5370 | update_avg(&p->se.avg_overlap, runtime); | |
5371 | } else { | |
5372 | update_avg(&p->se.avg_running, 0); | |
5373 | } | |
5374 | p->sched_class->put_prev_task(rq, p); | |
5375 | } | |
5376 | ||
5377 | /* | |
5378 | * Pick up the highest-prio task: | |
5379 | */ | |
5380 | static inline struct task_struct * | |
5381 | pick_next_task(struct rq *rq) | |
5382 | { | |
5383 | const struct sched_class *class; | |
5384 | struct task_struct *p; | |
5385 | ||
5386 | /* | |
5387 | * Optimization: we know that if all tasks are in | |
5388 | * the fair class we can call that function directly: | |
5389 | */ | |
5390 | if (likely(rq->nr_running == rq->cfs.nr_running)) { | |
5391 | p = fair_sched_class.pick_next_task(rq); | |
5392 | if (likely(p)) | |
5393 | return p; | |
5394 | } | |
5395 | ||
5396 | class = sched_class_highest; | |
5397 | for ( ; ; ) { | |
5398 | p = class->pick_next_task(rq); | |
5399 | if (p) | |
5400 | return p; | |
5401 | /* | |
5402 | * Will never be NULL as the idle class always | |
5403 | * returns a non-NULL p: | |
5404 | */ | |
5405 | class = class->next; | |
5406 | } | |
5407 | } | |
5408 | ||
5409 | /* | |
5410 | * schedule() is the main scheduler function. | |
5411 | */ | |
5412 | asmlinkage void __sched schedule(void) | |
5413 | { | |
5414 | struct task_struct *prev, *next; | |
5415 | unsigned long *switch_count; | |
5416 | struct rq *rq; | |
5417 | int cpu; | |
5418 | ||
5419 | need_resched: | |
5420 | preempt_disable(); | |
5421 | cpu = smp_processor_id(); | |
5422 | rq = cpu_rq(cpu); | |
5423 | rcu_sched_qs(cpu); | |
5424 | prev = rq->curr; | |
5425 | switch_count = &prev->nivcsw; | |
5426 | ||
5427 | release_kernel_lock(prev); | |
5428 | need_resched_nonpreemptible: | |
5429 | ||
5430 | schedule_debug(prev); | |
5431 | ||
5432 | if (sched_feat(HRTICK)) | |
5433 | hrtick_clear(rq); | |
5434 | ||
5435 | spin_lock_irq(&rq->lock); | |
5436 | update_rq_clock(rq); | |
5437 | clear_tsk_need_resched(prev); | |
5438 | ||
5439 | if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { | |
5440 | if (unlikely(signal_pending_state(prev->state, prev))) | |
5441 | prev->state = TASK_RUNNING; | |
5442 | else | |
5443 | deactivate_task(rq, prev, 1); | |
5444 | switch_count = &prev->nvcsw; | |
5445 | } | |
5446 | ||
5447 | pre_schedule(rq, prev); | |
5448 | ||
5449 | if (unlikely(!rq->nr_running)) | |
5450 | idle_balance(cpu, rq); | |
5451 | ||
5452 | put_prev_task(rq, prev); | |
5453 | next = pick_next_task(rq); | |
5454 | ||
5455 | if (likely(prev != next)) { | |
5456 | sched_info_switch(prev, next); | |
5457 | perf_event_task_sched_out(prev, next, cpu); | |
5458 | ||
5459 | rq->nr_switches++; | |
5460 | rq->curr = next; | |
5461 | ++*switch_count; | |
5462 | ||
5463 | context_switch(rq, prev, next); /* unlocks the rq */ | |
5464 | /* | |
5465 | * the context switch might have flipped the stack from under | |
5466 | * us, hence refresh the local variables. | |
5467 | */ | |
5468 | cpu = smp_processor_id(); | |
5469 | rq = cpu_rq(cpu); | |
5470 | } else | |
5471 | spin_unlock_irq(&rq->lock); | |
5472 | ||
5473 | post_schedule(rq); | |
5474 | ||
5475 | if (unlikely(reacquire_kernel_lock(current) < 0)) | |
5476 | goto need_resched_nonpreemptible; | |
5477 | ||
5478 | preempt_enable_no_resched(); | |
5479 | if (need_resched()) | |
5480 | goto need_resched; | |
5481 | } | |
5482 | EXPORT_SYMBOL(schedule); | |
5483 | ||
5484 | #ifdef CONFIG_SMP | |
5485 | /* | |
5486 | * Look out! "owner" is an entirely speculative pointer | |
5487 | * access and not reliable. | |
5488 | */ | |
5489 | int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner) | |
5490 | { | |
5491 | unsigned int cpu; | |
5492 | struct rq *rq; | |
5493 | ||
5494 | if (!sched_feat(OWNER_SPIN)) | |
5495 | return 0; | |
5496 | ||
5497 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
5498 | /* | |
5499 | * Need to access the cpu field knowing that | |
5500 | * DEBUG_PAGEALLOC could have unmapped it if | |
5501 | * the mutex owner just released it and exited. | |
5502 | */ | |
5503 | if (probe_kernel_address(&owner->cpu, cpu)) | |
5504 | goto out; | |
5505 | #else | |
5506 | cpu = owner->cpu; | |
5507 | #endif | |
5508 | ||
5509 | /* | |
5510 | * Even if the access succeeded (likely case), | |
5511 | * the cpu field may no longer be valid. | |
5512 | */ | |
5513 | if (cpu >= nr_cpumask_bits) | |
5514 | goto out; | |
5515 | ||
5516 | /* | |
5517 | * We need to validate that we can do a | |
5518 | * get_cpu() and that we have the percpu area. | |
5519 | */ | |
5520 | if (!cpu_online(cpu)) | |
5521 | goto out; | |
5522 | ||
5523 | rq = cpu_rq(cpu); | |
5524 | ||
5525 | for (;;) { | |
5526 | /* | |
5527 | * Owner changed, break to re-assess state. | |
5528 | */ | |
5529 | if (lock->owner != owner) | |
5530 | break; | |
5531 | ||
5532 | /* | |
5533 | * Is that owner really running on that cpu? | |
5534 | */ | |
5535 | if (task_thread_info(rq->curr) != owner || need_resched()) | |
5536 | return 0; | |
5537 | ||
5538 | cpu_relax(); | |
5539 | } | |
5540 | out: | |
5541 | return 1; | |
5542 | } | |
5543 | #endif | |
5544 | ||
5545 | #ifdef CONFIG_PREEMPT | |
5546 | /* | |
5547 | * this is the entry point to schedule() from in-kernel preemption | |
5548 | * off of preempt_enable. Kernel preemptions off return from interrupt | |
5549 | * occur there and call schedule directly. | |
5550 | */ | |
5551 | asmlinkage void __sched preempt_schedule(void) | |
5552 | { | |
5553 | struct thread_info *ti = current_thread_info(); | |
5554 | ||
5555 | /* | |
5556 | * If there is a non-zero preempt_count or interrupts are disabled, | |
5557 | * we do not want to preempt the current task. Just return.. | |
5558 | */ | |
5559 | if (likely(ti->preempt_count || irqs_disabled())) | |
5560 | return; | |
5561 | ||
5562 | do { | |
5563 | add_preempt_count(PREEMPT_ACTIVE); | |
5564 | schedule(); | |
5565 | sub_preempt_count(PREEMPT_ACTIVE); | |
5566 | ||
5567 | /* | |
5568 | * Check again in case we missed a preemption opportunity | |
5569 | * between schedule and now. | |
5570 | */ | |
5571 | barrier(); | |
5572 | } while (need_resched()); | |
5573 | } | |
5574 | EXPORT_SYMBOL(preempt_schedule); | |
5575 | ||
5576 | /* | |
5577 | * this is the entry point to schedule() from kernel preemption | |
5578 | * off of irq context. | |
5579 | * Note, that this is called and return with irqs disabled. This will | |
5580 | * protect us against recursive calling from irq. | |
5581 | */ | |
5582 | asmlinkage void __sched preempt_schedule_irq(void) | |
5583 | { | |
5584 | struct thread_info *ti = current_thread_info(); | |
5585 | ||
5586 | /* Catch callers which need to be fixed */ | |
5587 | BUG_ON(ti->preempt_count || !irqs_disabled()); | |
5588 | ||
5589 | do { | |
5590 | add_preempt_count(PREEMPT_ACTIVE); | |
5591 | local_irq_enable(); | |
5592 | schedule(); | |
5593 | local_irq_disable(); | |
5594 | sub_preempt_count(PREEMPT_ACTIVE); | |
5595 | ||
5596 | /* | |
5597 | * Check again in case we missed a preemption opportunity | |
5598 | * between schedule and now. | |
5599 | */ | |
5600 | barrier(); | |
5601 | } while (need_resched()); | |
5602 | } | |
5603 | ||
5604 | #endif /* CONFIG_PREEMPT */ | |
5605 | ||
5606 | int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, | |
5607 | void *key) | |
5608 | { | |
5609 | return try_to_wake_up(curr->private, mode, wake_flags); | |
5610 | } | |
5611 | EXPORT_SYMBOL(default_wake_function); | |
5612 | ||
5613 | /* | |
5614 | * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just | |
5615 | * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve | |
5616 | * number) then we wake all the non-exclusive tasks and one exclusive task. | |
5617 | * | |
5618 | * There are circumstances in which we can try to wake a task which has already | |
5619 | * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns | |
5620 | * zero in this (rare) case, and we handle it by continuing to scan the queue. | |
5621 | */ | |
5622 | static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, | |
5623 | int nr_exclusive, int wake_flags, void *key) | |
5624 | { | |
5625 | wait_queue_t *curr, *next; | |
5626 | ||
5627 | list_for_each_entry_safe(curr, next, &q->task_list, task_list) { | |
5628 | unsigned flags = curr->flags; | |
5629 | ||
5630 | if (curr->func(curr, mode, wake_flags, key) && | |
5631 | (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) | |
5632 | break; | |
5633 | } | |
5634 | } | |
5635 | ||
5636 | /** | |
5637 | * __wake_up - wake up threads blocked on a waitqueue. | |
5638 | * @q: the waitqueue | |
5639 | * @mode: which threads | |
5640 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
5641 | * @key: is directly passed to the wakeup function | |
5642 | * | |
5643 | * It may be assumed that this function implies a write memory barrier before | |
5644 | * changing the task state if and only if any tasks are woken up. | |
5645 | */ | |
5646 | void __wake_up(wait_queue_head_t *q, unsigned int mode, | |
5647 | int nr_exclusive, void *key) | |
5648 | { | |
5649 | unsigned long flags; | |
5650 | ||
5651 | spin_lock_irqsave(&q->lock, flags); | |
5652 | __wake_up_common(q, mode, nr_exclusive, 0, key); | |
5653 | spin_unlock_irqrestore(&q->lock, flags); | |
5654 | } | |
5655 | EXPORT_SYMBOL(__wake_up); | |
5656 | ||
5657 | /* | |
5658 | * Same as __wake_up but called with the spinlock in wait_queue_head_t held. | |
5659 | */ | |
5660 | void __wake_up_locked(wait_queue_head_t *q, unsigned int mode) | |
5661 | { | |
5662 | __wake_up_common(q, mode, 1, 0, NULL); | |
5663 | } | |
5664 | ||
5665 | void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key) | |
5666 | { | |
5667 | __wake_up_common(q, mode, 1, 0, key); | |
5668 | } | |
5669 | ||
5670 | /** | |
5671 | * __wake_up_sync_key - wake up threads blocked on a waitqueue. | |
5672 | * @q: the waitqueue | |
5673 | * @mode: which threads | |
5674 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
5675 | * @key: opaque value to be passed to wakeup targets | |
5676 | * | |
5677 | * The sync wakeup differs that the waker knows that it will schedule | |
5678 | * away soon, so while the target thread will be woken up, it will not | |
5679 | * be migrated to another CPU - ie. the two threads are 'synchronized' | |
5680 | * with each other. This can prevent needless bouncing between CPUs. | |
5681 | * | |
5682 | * On UP it can prevent extra preemption. | |
5683 | * | |
5684 | * It may be assumed that this function implies a write memory barrier before | |
5685 | * changing the task state if and only if any tasks are woken up. | |
5686 | */ | |
5687 | void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode, | |
5688 | int nr_exclusive, void *key) | |
5689 | { | |
5690 | unsigned long flags; | |
5691 | int wake_flags = WF_SYNC; | |
5692 | ||
5693 | if (unlikely(!q)) | |
5694 | return; | |
5695 | ||
5696 | if (unlikely(!nr_exclusive)) | |
5697 | wake_flags = 0; | |
5698 | ||
5699 | spin_lock_irqsave(&q->lock, flags); | |
5700 | __wake_up_common(q, mode, nr_exclusive, wake_flags, key); | |
5701 | spin_unlock_irqrestore(&q->lock, flags); | |
5702 | } | |
5703 | EXPORT_SYMBOL_GPL(__wake_up_sync_key); | |
5704 | ||
5705 | /* | |
5706 | * __wake_up_sync - see __wake_up_sync_key() | |
5707 | */ | |
5708 | void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) | |
5709 | { | |
5710 | __wake_up_sync_key(q, mode, nr_exclusive, NULL); | |
5711 | } | |
5712 | EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ | |
5713 | ||
5714 | /** | |
5715 | * complete: - signals a single thread waiting on this completion | |
5716 | * @x: holds the state of this particular completion | |
5717 | * | |
5718 | * This will wake up a single thread waiting on this completion. Threads will be | |
5719 | * awakened in the same order in which they were queued. | |
5720 | * | |
5721 | * See also complete_all(), wait_for_completion() and related routines. | |
5722 | * | |
5723 | * It may be assumed that this function implies a write memory barrier before | |
5724 | * changing the task state if and only if any tasks are woken up. | |
5725 | */ | |
5726 | void complete(struct completion *x) | |
5727 | { | |
5728 | unsigned long flags; | |
5729 | ||
5730 | spin_lock_irqsave(&x->wait.lock, flags); | |
5731 | x->done++; | |
5732 | __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); | |
5733 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
5734 | } | |
5735 | EXPORT_SYMBOL(complete); | |
5736 | ||
5737 | /** | |
5738 | * complete_all: - signals all threads waiting on this completion | |
5739 | * @x: holds the state of this particular completion | |
5740 | * | |
5741 | * This will wake up all threads waiting on this particular completion event. | |
5742 | * | |
5743 | * It may be assumed that this function implies a write memory barrier before | |
5744 | * changing the task state if and only if any tasks are woken up. | |
5745 | */ | |
5746 | void complete_all(struct completion *x) | |
5747 | { | |
5748 | unsigned long flags; | |
5749 | ||
5750 | spin_lock_irqsave(&x->wait.lock, flags); | |
5751 | x->done += UINT_MAX/2; | |
5752 | __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); | |
5753 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
5754 | } | |
5755 | EXPORT_SYMBOL(complete_all); | |
5756 | ||
5757 | static inline long __sched | |
5758 | do_wait_for_common(struct completion *x, long timeout, int state) | |
5759 | { | |
5760 | if (!x->done) { | |
5761 | DECLARE_WAITQUEUE(wait, current); | |
5762 | ||
5763 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
5764 | __add_wait_queue_tail(&x->wait, &wait); | |
5765 | do { | |
5766 | if (signal_pending_state(state, current)) { | |
5767 | timeout = -ERESTARTSYS; | |
5768 | break; | |
5769 | } | |
5770 | __set_current_state(state); | |
5771 | spin_unlock_irq(&x->wait.lock); | |
5772 | timeout = schedule_timeout(timeout); | |
5773 | spin_lock_irq(&x->wait.lock); | |
5774 | } while (!x->done && timeout); | |
5775 | __remove_wait_queue(&x->wait, &wait); | |
5776 | if (!x->done) | |
5777 | return timeout; | |
5778 | } | |
5779 | x->done--; | |
5780 | return timeout ?: 1; | |
5781 | } | |
5782 | ||
5783 | static long __sched | |
5784 | wait_for_common(struct completion *x, long timeout, int state) | |
5785 | { | |
5786 | might_sleep(); | |
5787 | ||
5788 | spin_lock_irq(&x->wait.lock); | |
5789 | timeout = do_wait_for_common(x, timeout, state); | |
5790 | spin_unlock_irq(&x->wait.lock); | |
5791 | return timeout; | |
5792 | } | |
5793 | ||
5794 | /** | |
5795 | * wait_for_completion: - waits for completion of a task | |
5796 | * @x: holds the state of this particular completion | |
5797 | * | |
5798 | * This waits to be signaled for completion of a specific task. It is NOT | |
5799 | * interruptible and there is no timeout. | |
5800 | * | |
5801 | * See also similar routines (i.e. wait_for_completion_timeout()) with timeout | |
5802 | * and interrupt capability. Also see complete(). | |
5803 | */ | |
5804 | void __sched wait_for_completion(struct completion *x) | |
5805 | { | |
5806 | wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); | |
5807 | } | |
5808 | EXPORT_SYMBOL(wait_for_completion); | |
5809 | ||
5810 | /** | |
5811 | * wait_for_completion_timeout: - waits for completion of a task (w/timeout) | |
5812 | * @x: holds the state of this particular completion | |
5813 | * @timeout: timeout value in jiffies | |
5814 | * | |
5815 | * This waits for either a completion of a specific task to be signaled or for a | |
5816 | * specified timeout to expire. The timeout is in jiffies. It is not | |
5817 | * interruptible. | |
5818 | */ | |
5819 | unsigned long __sched | |
5820 | wait_for_completion_timeout(struct completion *x, unsigned long timeout) | |
5821 | { | |
5822 | return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); | |
5823 | } | |
5824 | EXPORT_SYMBOL(wait_for_completion_timeout); | |
5825 | ||
5826 | /** | |
5827 | * wait_for_completion_interruptible: - waits for completion of a task (w/intr) | |
5828 | * @x: holds the state of this particular completion | |
5829 | * | |
5830 | * This waits for completion of a specific task to be signaled. It is | |
5831 | * interruptible. | |
5832 | */ | |
5833 | int __sched wait_for_completion_interruptible(struct completion *x) | |
5834 | { | |
5835 | long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); | |
5836 | if (t == -ERESTARTSYS) | |
5837 | return t; | |
5838 | return 0; | |
5839 | } | |
5840 | EXPORT_SYMBOL(wait_for_completion_interruptible); | |
5841 | ||
5842 | /** | |
5843 | * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) | |
5844 | * @x: holds the state of this particular completion | |
5845 | * @timeout: timeout value in jiffies | |
5846 | * | |
5847 | * This waits for either a completion of a specific task to be signaled or for a | |
5848 | * specified timeout to expire. It is interruptible. The timeout is in jiffies. | |
5849 | */ | |
5850 | unsigned long __sched | |
5851 | wait_for_completion_interruptible_timeout(struct completion *x, | |
5852 | unsigned long timeout) | |
5853 | { | |
5854 | return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); | |
5855 | } | |
5856 | EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); | |
5857 | ||
5858 | /** | |
5859 | * wait_for_completion_killable: - waits for completion of a task (killable) | |
5860 | * @x: holds the state of this particular completion | |
5861 | * | |
5862 | * This waits to be signaled for completion of a specific task. It can be | |
5863 | * interrupted by a kill signal. | |
5864 | */ | |
5865 | int __sched wait_for_completion_killable(struct completion *x) | |
5866 | { | |
5867 | long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); | |
5868 | if (t == -ERESTARTSYS) | |
5869 | return t; | |
5870 | return 0; | |
5871 | } | |
5872 | EXPORT_SYMBOL(wait_for_completion_killable); | |
5873 | ||
5874 | /** | |
5875 | * try_wait_for_completion - try to decrement a completion without blocking | |
5876 | * @x: completion structure | |
5877 | * | |
5878 | * Returns: 0 if a decrement cannot be done without blocking | |
5879 | * 1 if a decrement succeeded. | |
5880 | * | |
5881 | * If a completion is being used as a counting completion, | |
5882 | * attempt to decrement the counter without blocking. This | |
5883 | * enables us to avoid waiting if the resource the completion | |
5884 | * is protecting is not available. | |
5885 | */ | |
5886 | bool try_wait_for_completion(struct completion *x) | |
5887 | { | |
5888 | int ret = 1; | |
5889 | ||
5890 | spin_lock_irq(&x->wait.lock); | |
5891 | if (!x->done) | |
5892 | ret = 0; | |
5893 | else | |
5894 | x->done--; | |
5895 | spin_unlock_irq(&x->wait.lock); | |
5896 | return ret; | |
5897 | } | |
5898 | EXPORT_SYMBOL(try_wait_for_completion); | |
5899 | ||
5900 | /** | |
5901 | * completion_done - Test to see if a completion has any waiters | |
5902 | * @x: completion structure | |
5903 | * | |
5904 | * Returns: 0 if there are waiters (wait_for_completion() in progress) | |
5905 | * 1 if there are no waiters. | |
5906 | * | |
5907 | */ | |
5908 | bool completion_done(struct completion *x) | |
5909 | { | |
5910 | int ret = 1; | |
5911 | ||
5912 | spin_lock_irq(&x->wait.lock); | |
5913 | if (!x->done) | |
5914 | ret = 0; | |
5915 | spin_unlock_irq(&x->wait.lock); | |
5916 | return ret; | |
5917 | } | |
5918 | EXPORT_SYMBOL(completion_done); | |
5919 | ||
5920 | static long __sched | |
5921 | sleep_on_common(wait_queue_head_t *q, int state, long timeout) | |
5922 | { | |
5923 | unsigned long flags; | |
5924 | wait_queue_t wait; | |
5925 | ||
5926 | init_waitqueue_entry(&wait, current); | |
5927 | ||
5928 | __set_current_state(state); | |
5929 | ||
5930 | spin_lock_irqsave(&q->lock, flags); | |
5931 | __add_wait_queue(q, &wait); | |
5932 | spin_unlock(&q->lock); | |
5933 | timeout = schedule_timeout(timeout); | |
5934 | spin_lock_irq(&q->lock); | |
5935 | __remove_wait_queue(q, &wait); | |
5936 | spin_unlock_irqrestore(&q->lock, flags); | |
5937 | ||
5938 | return timeout; | |
5939 | } | |
5940 | ||
5941 | void __sched interruptible_sleep_on(wait_queue_head_t *q) | |
5942 | { | |
5943 | sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); | |
5944 | } | |
5945 | EXPORT_SYMBOL(interruptible_sleep_on); | |
5946 | ||
5947 | long __sched | |
5948 | interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) | |
5949 | { | |
5950 | return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); | |
5951 | } | |
5952 | EXPORT_SYMBOL(interruptible_sleep_on_timeout); | |
5953 | ||
5954 | void __sched sleep_on(wait_queue_head_t *q) | |
5955 | { | |
5956 | sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); | |
5957 | } | |
5958 | EXPORT_SYMBOL(sleep_on); | |
5959 | ||
5960 | long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) | |
5961 | { | |
5962 | return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); | |
5963 | } | |
5964 | EXPORT_SYMBOL(sleep_on_timeout); | |
5965 | ||
5966 | #ifdef CONFIG_RT_MUTEXES | |
5967 | ||
5968 | /* | |
5969 | * rt_mutex_setprio - set the current priority of a task | |
5970 | * @p: task | |
5971 | * @prio: prio value (kernel-internal form) | |
5972 | * | |
5973 | * This function changes the 'effective' priority of a task. It does | |
5974 | * not touch ->normal_prio like __setscheduler(). | |
5975 | * | |
5976 | * Used by the rt_mutex code to implement priority inheritance logic. | |
5977 | */ | |
5978 | void rt_mutex_setprio(struct task_struct *p, int prio) | |
5979 | { | |
5980 | unsigned long flags; | |
5981 | int oldprio, on_rq, running; | |
5982 | struct rq *rq; | |
5983 | const struct sched_class *prev_class = p->sched_class; | |
5984 | ||
5985 | BUG_ON(prio < 0 || prio > MAX_PRIO); | |
5986 | ||
5987 | rq = task_rq_lock(p, &flags); | |
5988 | update_rq_clock(rq); | |
5989 | ||
5990 | oldprio = p->prio; | |
5991 | on_rq = p->se.on_rq; | |
5992 | running = task_current(rq, p); | |
5993 | if (on_rq) | |
5994 | dequeue_task(rq, p, 0); | |
5995 | if (running) | |
5996 | p->sched_class->put_prev_task(rq, p); | |
5997 | ||
5998 | if (rt_prio(prio)) | |
5999 | p->sched_class = &rt_sched_class; | |
6000 | else | |
6001 | p->sched_class = &fair_sched_class; | |
6002 | ||
6003 | p->prio = prio; | |
6004 | ||
6005 | if (running) | |
6006 | p->sched_class->set_curr_task(rq); | |
6007 | if (on_rq) { | |
6008 | enqueue_task(rq, p, 0); | |
6009 | ||
6010 | check_class_changed(rq, p, prev_class, oldprio, running); | |
6011 | } | |
6012 | task_rq_unlock(rq, &flags); | |
6013 | } | |
6014 | ||
6015 | #endif | |
6016 | ||
6017 | void set_user_nice(struct task_struct *p, long nice) | |
6018 | { | |
6019 | int old_prio, delta, on_rq; | |
6020 | unsigned long flags; | |
6021 | struct rq *rq; | |
6022 | ||
6023 | if (TASK_NICE(p) == nice || nice < -20 || nice > 19) | |
6024 | return; | |
6025 | /* | |
6026 | * We have to be careful, if called from sys_setpriority(), | |
6027 | * the task might be in the middle of scheduling on another CPU. | |
6028 | */ | |
6029 | rq = task_rq_lock(p, &flags); | |
6030 | update_rq_clock(rq); | |
6031 | /* | |
6032 | * The RT priorities are set via sched_setscheduler(), but we still | |
6033 | * allow the 'normal' nice value to be set - but as expected | |
6034 | * it wont have any effect on scheduling until the task is | |
6035 | * SCHED_FIFO/SCHED_RR: | |
6036 | */ | |
6037 | if (task_has_rt_policy(p)) { | |
6038 | p->static_prio = NICE_TO_PRIO(nice); | |
6039 | goto out_unlock; | |
6040 | } | |
6041 | on_rq = p->se.on_rq; | |
6042 | if (on_rq) | |
6043 | dequeue_task(rq, p, 0); | |
6044 | ||
6045 | p->static_prio = NICE_TO_PRIO(nice); | |
6046 | set_load_weight(p); | |
6047 | old_prio = p->prio; | |
6048 | p->prio = effective_prio(p); | |
6049 | delta = p->prio - old_prio; | |
6050 | ||
6051 | if (on_rq) { | |
6052 | enqueue_task(rq, p, 0); | |
6053 | /* | |
6054 | * If the task increased its priority or is running and | |
6055 | * lowered its priority, then reschedule its CPU: | |
6056 | */ | |
6057 | if (delta < 0 || (delta > 0 && task_running(rq, p))) | |
6058 | resched_task(rq->curr); | |
6059 | } | |
6060 | out_unlock: | |
6061 | task_rq_unlock(rq, &flags); | |
6062 | } | |
6063 | EXPORT_SYMBOL(set_user_nice); | |
6064 | ||
6065 | /* | |
6066 | * can_nice - check if a task can reduce its nice value | |
6067 | * @p: task | |
6068 | * @nice: nice value | |
6069 | */ | |
6070 | int can_nice(const struct task_struct *p, const int nice) | |
6071 | { | |
6072 | /* convert nice value [19,-20] to rlimit style value [1,40] */ | |
6073 | int nice_rlim = 20 - nice; | |
6074 | ||
6075 | return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur || | |
6076 | capable(CAP_SYS_NICE)); | |
6077 | } | |
6078 | ||
6079 | #ifdef __ARCH_WANT_SYS_NICE | |
6080 | ||
6081 | /* | |
6082 | * sys_nice - change the priority of the current process. | |
6083 | * @increment: priority increment | |
6084 | * | |
6085 | * sys_setpriority is a more generic, but much slower function that | |
6086 | * does similar things. | |
6087 | */ | |
6088 | SYSCALL_DEFINE1(nice, int, increment) | |
6089 | { | |
6090 | long nice, retval; | |
6091 | ||
6092 | /* | |
6093 | * Setpriority might change our priority at the same moment. | |
6094 | * We don't have to worry. Conceptually one call occurs first | |
6095 | * and we have a single winner. | |
6096 | */ | |
6097 | if (increment < -40) | |
6098 | increment = -40; | |
6099 | if (increment > 40) | |
6100 | increment = 40; | |
6101 | ||
6102 | nice = TASK_NICE(current) + increment; | |
6103 | if (nice < -20) | |
6104 | nice = -20; | |
6105 | if (nice > 19) | |
6106 | nice = 19; | |
6107 | ||
6108 | if (increment < 0 && !can_nice(current, nice)) | |
6109 | return -EPERM; | |
6110 | ||
6111 | retval = security_task_setnice(current, nice); | |
6112 | if (retval) | |
6113 | return retval; | |
6114 | ||
6115 | set_user_nice(current, nice); | |
6116 | return 0; | |
6117 | } | |
6118 | ||
6119 | #endif | |
6120 | ||
6121 | /** | |
6122 | * task_prio - return the priority value of a given task. | |
6123 | * @p: the task in question. | |
6124 | * | |
6125 | * This is the priority value as seen by users in /proc. | |
6126 | * RT tasks are offset by -200. Normal tasks are centered | |
6127 | * around 0, value goes from -16 to +15. | |
6128 | */ | |
6129 | int task_prio(const struct task_struct *p) | |
6130 | { | |
6131 | return p->prio - MAX_RT_PRIO; | |
6132 | } | |
6133 | ||
6134 | /** | |
6135 | * task_nice - return the nice value of a given task. | |
6136 | * @p: the task in question. | |
6137 | */ | |
6138 | int task_nice(const struct task_struct *p) | |
6139 | { | |
6140 | return TASK_NICE(p); | |
6141 | } | |
6142 | EXPORT_SYMBOL(task_nice); | |
6143 | ||
6144 | /** | |
6145 | * idle_cpu - is a given cpu idle currently? | |
6146 | * @cpu: the processor in question. | |
6147 | */ | |
6148 | int idle_cpu(int cpu) | |
6149 | { | |
6150 | return cpu_curr(cpu) == cpu_rq(cpu)->idle; | |
6151 | } | |
6152 | ||
6153 | /** | |
6154 | * idle_task - return the idle task for a given cpu. | |
6155 | * @cpu: the processor in question. | |
6156 | */ | |
6157 | struct task_struct *idle_task(int cpu) | |
6158 | { | |
6159 | return cpu_rq(cpu)->idle; | |
6160 | } | |
6161 | ||
6162 | /** | |
6163 | * find_process_by_pid - find a process with a matching PID value. | |
6164 | * @pid: the pid in question. | |
6165 | */ | |
6166 | static struct task_struct *find_process_by_pid(pid_t pid) | |
6167 | { | |
6168 | return pid ? find_task_by_vpid(pid) : current; | |
6169 | } | |
6170 | ||
6171 | /* Actually do priority change: must hold rq lock. */ | |
6172 | static void | |
6173 | __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) | |
6174 | { | |
6175 | BUG_ON(p->se.on_rq); | |
6176 | ||
6177 | p->policy = policy; | |
6178 | switch (p->policy) { | |
6179 | case SCHED_NORMAL: | |
6180 | case SCHED_BATCH: | |
6181 | case SCHED_IDLE: | |
6182 | p->sched_class = &fair_sched_class; | |
6183 | break; | |
6184 | case SCHED_FIFO: | |
6185 | case SCHED_RR: | |
6186 | p->sched_class = &rt_sched_class; | |
6187 | break; | |
6188 | } | |
6189 | ||
6190 | p->rt_priority = prio; | |
6191 | p->normal_prio = normal_prio(p); | |
6192 | /* we are holding p->pi_lock already */ | |
6193 | p->prio = rt_mutex_getprio(p); | |
6194 | set_load_weight(p); | |
6195 | } | |
6196 | ||
6197 | /* | |
6198 | * check the target process has a UID that matches the current process's | |
6199 | */ | |
6200 | static bool check_same_owner(struct task_struct *p) | |
6201 | { | |
6202 | const struct cred *cred = current_cred(), *pcred; | |
6203 | bool match; | |
6204 | ||
6205 | rcu_read_lock(); | |
6206 | pcred = __task_cred(p); | |
6207 | match = (cred->euid == pcred->euid || | |
6208 | cred->euid == pcred->uid); | |
6209 | rcu_read_unlock(); | |
6210 | return match; | |
6211 | } | |
6212 | ||
6213 | static int __sched_setscheduler(struct task_struct *p, int policy, | |
6214 | struct sched_param *param, bool user) | |
6215 | { | |
6216 | int retval, oldprio, oldpolicy = -1, on_rq, running; | |
6217 | unsigned long flags; | |
6218 | const struct sched_class *prev_class = p->sched_class; | |
6219 | struct rq *rq; | |
6220 | int reset_on_fork; | |
6221 | ||
6222 | /* may grab non-irq protected spin_locks */ | |
6223 | BUG_ON(in_interrupt()); | |
6224 | recheck: | |
6225 | /* double check policy once rq lock held */ | |
6226 | if (policy < 0) { | |
6227 | reset_on_fork = p->sched_reset_on_fork; | |
6228 | policy = oldpolicy = p->policy; | |
6229 | } else { | |
6230 | reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); | |
6231 | policy &= ~SCHED_RESET_ON_FORK; | |
6232 | ||
6233 | if (policy != SCHED_FIFO && policy != SCHED_RR && | |
6234 | policy != SCHED_NORMAL && policy != SCHED_BATCH && | |
6235 | policy != SCHED_IDLE) | |
6236 | return -EINVAL; | |
6237 | } | |
6238 | ||
6239 | /* | |
6240 | * Valid priorities for SCHED_FIFO and SCHED_RR are | |
6241 | * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, | |
6242 | * SCHED_BATCH and SCHED_IDLE is 0. | |
6243 | */ | |
6244 | if (param->sched_priority < 0 || | |
6245 | (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || | |
6246 | (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) | |
6247 | return -EINVAL; | |
6248 | if (rt_policy(policy) != (param->sched_priority != 0)) | |
6249 | return -EINVAL; | |
6250 | ||
6251 | /* | |
6252 | * Allow unprivileged RT tasks to decrease priority: | |
6253 | */ | |
6254 | if (user && !capable(CAP_SYS_NICE)) { | |
6255 | if (rt_policy(policy)) { | |
6256 | unsigned long rlim_rtprio; | |
6257 | ||
6258 | if (!lock_task_sighand(p, &flags)) | |
6259 | return -ESRCH; | |
6260 | rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur; | |
6261 | unlock_task_sighand(p, &flags); | |
6262 | ||
6263 | /* can't set/change the rt policy */ | |
6264 | if (policy != p->policy && !rlim_rtprio) | |
6265 | return -EPERM; | |
6266 | ||
6267 | /* can't increase priority */ | |
6268 | if (param->sched_priority > p->rt_priority && | |
6269 | param->sched_priority > rlim_rtprio) | |
6270 | return -EPERM; | |
6271 | } | |
6272 | /* | |
6273 | * Like positive nice levels, dont allow tasks to | |
6274 | * move out of SCHED_IDLE either: | |
6275 | */ | |
6276 | if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) | |
6277 | return -EPERM; | |
6278 | ||
6279 | /* can't change other user's priorities */ | |
6280 | if (!check_same_owner(p)) | |
6281 | return -EPERM; | |
6282 | ||
6283 | /* Normal users shall not reset the sched_reset_on_fork flag */ | |
6284 | if (p->sched_reset_on_fork && !reset_on_fork) | |
6285 | return -EPERM; | |
6286 | } | |
6287 | ||
6288 | if (user) { | |
6289 | #ifdef CONFIG_RT_GROUP_SCHED | |
6290 | /* | |
6291 | * Do not allow realtime tasks into groups that have no runtime | |
6292 | * assigned. | |
6293 | */ | |
6294 | if (rt_bandwidth_enabled() && rt_policy(policy) && | |
6295 | task_group(p)->rt_bandwidth.rt_runtime == 0) | |
6296 | return -EPERM; | |
6297 | #endif | |
6298 | ||
6299 | retval = security_task_setscheduler(p, policy, param); | |
6300 | if (retval) | |
6301 | return retval; | |
6302 | } | |
6303 | ||
6304 | /* | |
6305 | * make sure no PI-waiters arrive (or leave) while we are | |
6306 | * changing the priority of the task: | |
6307 | */ | |
6308 | spin_lock_irqsave(&p->pi_lock, flags); | |
6309 | /* | |
6310 | * To be able to change p->policy safely, the apropriate | |
6311 | * runqueue lock must be held. | |
6312 | */ | |
6313 | rq = __task_rq_lock(p); | |
6314 | /* recheck policy now with rq lock held */ | |
6315 | if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { | |
6316 | policy = oldpolicy = -1; | |
6317 | __task_rq_unlock(rq); | |
6318 | spin_unlock_irqrestore(&p->pi_lock, flags); | |
6319 | goto recheck; | |
6320 | } | |
6321 | update_rq_clock(rq); | |
6322 | on_rq = p->se.on_rq; | |
6323 | running = task_current(rq, p); | |
6324 | if (on_rq) | |
6325 | deactivate_task(rq, p, 0); | |
6326 | if (running) | |
6327 | p->sched_class->put_prev_task(rq, p); | |
6328 | ||
6329 | p->sched_reset_on_fork = reset_on_fork; | |
6330 | ||
6331 | oldprio = p->prio; | |
6332 | __setscheduler(rq, p, policy, param->sched_priority); | |
6333 | ||
6334 | if (running) | |
6335 | p->sched_class->set_curr_task(rq); | |
6336 | if (on_rq) { | |
6337 | activate_task(rq, p, 0); | |
6338 | ||
6339 | check_class_changed(rq, p, prev_class, oldprio, running); | |
6340 | } | |
6341 | __task_rq_unlock(rq); | |
6342 | spin_unlock_irqrestore(&p->pi_lock, flags); | |
6343 | ||
6344 | rt_mutex_adjust_pi(p); | |
6345 | ||
6346 | return 0; | |
6347 | } | |
6348 | ||
6349 | /** | |
6350 | * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. | |
6351 | * @p: the task in question. | |
6352 | * @policy: new policy. | |
6353 | * @param: structure containing the new RT priority. | |
6354 | * | |
6355 | * NOTE that the task may be already dead. | |
6356 | */ | |
6357 | int sched_setscheduler(struct task_struct *p, int policy, | |
6358 | struct sched_param *param) | |
6359 | { | |
6360 | return __sched_setscheduler(p, policy, param, true); | |
6361 | } | |
6362 | EXPORT_SYMBOL_GPL(sched_setscheduler); | |
6363 | ||
6364 | /** | |
6365 | * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. | |
6366 | * @p: the task in question. | |
6367 | * @policy: new policy. | |
6368 | * @param: structure containing the new RT priority. | |
6369 | * | |
6370 | * Just like sched_setscheduler, only don't bother checking if the | |
6371 | * current context has permission. For example, this is needed in | |
6372 | * stop_machine(): we create temporary high priority worker threads, | |
6373 | * but our caller might not have that capability. | |
6374 | */ | |
6375 | int sched_setscheduler_nocheck(struct task_struct *p, int policy, | |
6376 | struct sched_param *param) | |
6377 | { | |
6378 | return __sched_setscheduler(p, policy, param, false); | |
6379 | } | |
6380 | ||
6381 | static int | |
6382 | do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) | |
6383 | { | |
6384 | struct sched_param lparam; | |
6385 | struct task_struct *p; | |
6386 | int retval; | |
6387 | ||
6388 | if (!param || pid < 0) | |
6389 | return -EINVAL; | |
6390 | if (copy_from_user(&lparam, param, sizeof(struct sched_param))) | |
6391 | return -EFAULT; | |
6392 | ||
6393 | rcu_read_lock(); | |
6394 | retval = -ESRCH; | |
6395 | p = find_process_by_pid(pid); | |
6396 | if (p != NULL) | |
6397 | retval = sched_setscheduler(p, policy, &lparam); | |
6398 | rcu_read_unlock(); | |
6399 | ||
6400 | return retval; | |
6401 | } | |
6402 | ||
6403 | /** | |
6404 | * sys_sched_setscheduler - set/change the scheduler policy and RT priority | |
6405 | * @pid: the pid in question. | |
6406 | * @policy: new policy. | |
6407 | * @param: structure containing the new RT priority. | |
6408 | */ | |
6409 | SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, | |
6410 | struct sched_param __user *, param) | |
6411 | { | |
6412 | /* negative values for policy are not valid */ | |
6413 | if (policy < 0) | |
6414 | return -EINVAL; | |
6415 | ||
6416 | return do_sched_setscheduler(pid, policy, param); | |
6417 | } | |
6418 | ||
6419 | /** | |
6420 | * sys_sched_setparam - set/change the RT priority of a thread | |
6421 | * @pid: the pid in question. | |
6422 | * @param: structure containing the new RT priority. | |
6423 | */ | |
6424 | SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) | |
6425 | { | |
6426 | return do_sched_setscheduler(pid, -1, param); | |
6427 | } | |
6428 | ||
6429 | /** | |
6430 | * sys_sched_getscheduler - get the policy (scheduling class) of a thread | |
6431 | * @pid: the pid in question. | |
6432 | */ | |
6433 | SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) | |
6434 | { | |
6435 | struct task_struct *p; | |
6436 | int retval; | |
6437 | ||
6438 | if (pid < 0) | |
6439 | return -EINVAL; | |
6440 | ||
6441 | retval = -ESRCH; | |
6442 | read_lock(&tasklist_lock); | |
6443 | p = find_process_by_pid(pid); | |
6444 | if (p) { | |
6445 | retval = security_task_getscheduler(p); | |
6446 | if (!retval) | |
6447 | retval = p->policy | |
6448 | | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); | |
6449 | } | |
6450 | read_unlock(&tasklist_lock); | |
6451 | return retval; | |
6452 | } | |
6453 | ||
6454 | /** | |
6455 | * sys_sched_getparam - get the RT priority of a thread | |
6456 | * @pid: the pid in question. | |
6457 | * @param: structure containing the RT priority. | |
6458 | */ | |
6459 | SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) | |
6460 | { | |
6461 | struct sched_param lp; | |
6462 | struct task_struct *p; | |
6463 | int retval; | |
6464 | ||
6465 | if (!param || pid < 0) | |
6466 | return -EINVAL; | |
6467 | ||
6468 | read_lock(&tasklist_lock); | |
6469 | p = find_process_by_pid(pid); | |
6470 | retval = -ESRCH; | |
6471 | if (!p) | |
6472 | goto out_unlock; | |
6473 | ||
6474 | retval = security_task_getscheduler(p); | |
6475 | if (retval) | |
6476 | goto out_unlock; | |
6477 | ||
6478 | lp.sched_priority = p->rt_priority; | |
6479 | read_unlock(&tasklist_lock); | |
6480 | ||
6481 | /* | |
6482 | * This one might sleep, we cannot do it with a spinlock held ... | |
6483 | */ | |
6484 | retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; | |
6485 | ||
6486 | return retval; | |
6487 | ||
6488 | out_unlock: | |
6489 | read_unlock(&tasklist_lock); | |
6490 | return retval; | |
6491 | } | |
6492 | ||
6493 | long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) | |
6494 | { | |
6495 | cpumask_var_t cpus_allowed, new_mask; | |
6496 | struct task_struct *p; | |
6497 | int retval; | |
6498 | ||
6499 | get_online_cpus(); | |
6500 | read_lock(&tasklist_lock); | |
6501 | ||
6502 | p = find_process_by_pid(pid); | |
6503 | if (!p) { | |
6504 | read_unlock(&tasklist_lock); | |
6505 | put_online_cpus(); | |
6506 | return -ESRCH; | |
6507 | } | |
6508 | ||
6509 | /* | |
6510 | * It is not safe to call set_cpus_allowed with the | |
6511 | * tasklist_lock held. We will bump the task_struct's | |
6512 | * usage count and then drop tasklist_lock. | |
6513 | */ | |
6514 | get_task_struct(p); | |
6515 | read_unlock(&tasklist_lock); | |
6516 | ||
6517 | if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { | |
6518 | retval = -ENOMEM; | |
6519 | goto out_put_task; | |
6520 | } | |
6521 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { | |
6522 | retval = -ENOMEM; | |
6523 | goto out_free_cpus_allowed; | |
6524 | } | |
6525 | retval = -EPERM; | |
6526 | if (!check_same_owner(p) && !capable(CAP_SYS_NICE)) | |
6527 | goto out_unlock; | |
6528 | ||
6529 | retval = security_task_setscheduler(p, 0, NULL); | |
6530 | if (retval) | |
6531 | goto out_unlock; | |
6532 | ||
6533 | cpuset_cpus_allowed(p, cpus_allowed); | |
6534 | cpumask_and(new_mask, in_mask, cpus_allowed); | |
6535 | again: | |
6536 | retval = set_cpus_allowed_ptr(p, new_mask); | |
6537 | ||
6538 | if (!retval) { | |
6539 | cpuset_cpus_allowed(p, cpus_allowed); | |
6540 | if (!cpumask_subset(new_mask, cpus_allowed)) { | |
6541 | /* | |
6542 | * We must have raced with a concurrent cpuset | |
6543 | * update. Just reset the cpus_allowed to the | |
6544 | * cpuset's cpus_allowed | |
6545 | */ | |
6546 | cpumask_copy(new_mask, cpus_allowed); | |
6547 | goto again; | |
6548 | } | |
6549 | } | |
6550 | out_unlock: | |
6551 | free_cpumask_var(new_mask); | |
6552 | out_free_cpus_allowed: | |
6553 | free_cpumask_var(cpus_allowed); | |
6554 | out_put_task: | |
6555 | put_task_struct(p); | |
6556 | put_online_cpus(); | |
6557 | return retval; | |
6558 | } | |
6559 | ||
6560 | static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, | |
6561 | struct cpumask *new_mask) | |
6562 | { | |
6563 | if (len < cpumask_size()) | |
6564 | cpumask_clear(new_mask); | |
6565 | else if (len > cpumask_size()) | |
6566 | len = cpumask_size(); | |
6567 | ||
6568 | return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; | |
6569 | } | |
6570 | ||
6571 | /** | |
6572 | * sys_sched_setaffinity - set the cpu affinity of a process | |
6573 | * @pid: pid of the process | |
6574 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
6575 | * @user_mask_ptr: user-space pointer to the new cpu mask | |
6576 | */ | |
6577 | SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, | |
6578 | unsigned long __user *, user_mask_ptr) | |
6579 | { | |
6580 | cpumask_var_t new_mask; | |
6581 | int retval; | |
6582 | ||
6583 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) | |
6584 | return -ENOMEM; | |
6585 | ||
6586 | retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); | |
6587 | if (retval == 0) | |
6588 | retval = sched_setaffinity(pid, new_mask); | |
6589 | free_cpumask_var(new_mask); | |
6590 | return retval; | |
6591 | } | |
6592 | ||
6593 | long sched_getaffinity(pid_t pid, struct cpumask *mask) | |
6594 | { | |
6595 | struct task_struct *p; | |
6596 | int retval; | |
6597 | ||
6598 | get_online_cpus(); | |
6599 | read_lock(&tasklist_lock); | |
6600 | ||
6601 | retval = -ESRCH; | |
6602 | p = find_process_by_pid(pid); | |
6603 | if (!p) | |
6604 | goto out_unlock; | |
6605 | ||
6606 | retval = security_task_getscheduler(p); | |
6607 | if (retval) | |
6608 | goto out_unlock; | |
6609 | ||
6610 | cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); | |
6611 | ||
6612 | out_unlock: | |
6613 | read_unlock(&tasklist_lock); | |
6614 | put_online_cpus(); | |
6615 | ||
6616 | return retval; | |
6617 | } | |
6618 | ||
6619 | /** | |
6620 | * sys_sched_getaffinity - get the cpu affinity of a process | |
6621 | * @pid: pid of the process | |
6622 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
6623 | * @user_mask_ptr: user-space pointer to hold the current cpu mask | |
6624 | */ | |
6625 | SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, | |
6626 | unsigned long __user *, user_mask_ptr) | |
6627 | { | |
6628 | int ret; | |
6629 | cpumask_var_t mask; | |
6630 | ||
6631 | if (len < cpumask_size()) | |
6632 | return -EINVAL; | |
6633 | ||
6634 | if (!alloc_cpumask_var(&mask, GFP_KERNEL)) | |
6635 | return -ENOMEM; | |
6636 | ||
6637 | ret = sched_getaffinity(pid, mask); | |
6638 | if (ret == 0) { | |
6639 | if (copy_to_user(user_mask_ptr, mask, cpumask_size())) | |
6640 | ret = -EFAULT; | |
6641 | else | |
6642 | ret = cpumask_size(); | |
6643 | } | |
6644 | free_cpumask_var(mask); | |
6645 | ||
6646 | return ret; | |
6647 | } | |
6648 | ||
6649 | /** | |
6650 | * sys_sched_yield - yield the current processor to other threads. | |
6651 | * | |
6652 | * This function yields the current CPU to other tasks. If there are no | |
6653 | * other threads running on this CPU then this function will return. | |
6654 | */ | |
6655 | SYSCALL_DEFINE0(sched_yield) | |
6656 | { | |
6657 | struct rq *rq = this_rq_lock(); | |
6658 | ||
6659 | schedstat_inc(rq, yld_count); | |
6660 | current->sched_class->yield_task(rq); | |
6661 | ||
6662 | /* | |
6663 | * Since we are going to call schedule() anyway, there's | |
6664 | * no need to preempt or enable interrupts: | |
6665 | */ | |
6666 | __release(rq->lock); | |
6667 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); | |
6668 | _raw_spin_unlock(&rq->lock); | |
6669 | preempt_enable_no_resched(); | |
6670 | ||
6671 | schedule(); | |
6672 | ||
6673 | return 0; | |
6674 | } | |
6675 | ||
6676 | static inline int should_resched(void) | |
6677 | { | |
6678 | return need_resched() && !(preempt_count() & PREEMPT_ACTIVE); | |
6679 | } | |
6680 | ||
6681 | static void __cond_resched(void) | |
6682 | { | |
6683 | add_preempt_count(PREEMPT_ACTIVE); | |
6684 | schedule(); | |
6685 | sub_preempt_count(PREEMPT_ACTIVE); | |
6686 | } | |
6687 | ||
6688 | int __sched _cond_resched(void) | |
6689 | { | |
6690 | if (should_resched()) { | |
6691 | __cond_resched(); | |
6692 | return 1; | |
6693 | } | |
6694 | return 0; | |
6695 | } | |
6696 | EXPORT_SYMBOL(_cond_resched); | |
6697 | ||
6698 | /* | |
6699 | * __cond_resched_lock() - if a reschedule is pending, drop the given lock, | |
6700 | * call schedule, and on return reacquire the lock. | |
6701 | * | |
6702 | * This works OK both with and without CONFIG_PREEMPT. We do strange low-level | |
6703 | * operations here to prevent schedule() from being called twice (once via | |
6704 | * spin_unlock(), once by hand). | |
6705 | */ | |
6706 | int __cond_resched_lock(spinlock_t *lock) | |
6707 | { | |
6708 | int resched = should_resched(); | |
6709 | int ret = 0; | |
6710 | ||
6711 | lockdep_assert_held(lock); | |
6712 | ||
6713 | if (spin_needbreak(lock) || resched) { | |
6714 | spin_unlock(lock); | |
6715 | if (resched) | |
6716 | __cond_resched(); | |
6717 | else | |
6718 | cpu_relax(); | |
6719 | ret = 1; | |
6720 | spin_lock(lock); | |
6721 | } | |
6722 | return ret; | |
6723 | } | |
6724 | EXPORT_SYMBOL(__cond_resched_lock); | |
6725 | ||
6726 | int __sched __cond_resched_softirq(void) | |
6727 | { | |
6728 | BUG_ON(!in_softirq()); | |
6729 | ||
6730 | if (should_resched()) { | |
6731 | local_bh_enable(); | |
6732 | __cond_resched(); | |
6733 | local_bh_disable(); | |
6734 | return 1; | |
6735 | } | |
6736 | return 0; | |
6737 | } | |
6738 | EXPORT_SYMBOL(__cond_resched_softirq); | |
6739 | ||
6740 | /** | |
6741 | * yield - yield the current processor to other threads. | |
6742 | * | |
6743 | * This is a shortcut for kernel-space yielding - it marks the | |
6744 | * thread runnable and calls sys_sched_yield(). | |
6745 | */ | |
6746 | void __sched yield(void) | |
6747 | { | |
6748 | set_current_state(TASK_RUNNING); | |
6749 | sys_sched_yield(); | |
6750 | } | |
6751 | EXPORT_SYMBOL(yield); | |
6752 | ||
6753 | /* | |
6754 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so | |
6755 | * that process accounting knows that this is a task in IO wait state. | |
6756 | */ | |
6757 | void __sched io_schedule(void) | |
6758 | { | |
6759 | struct rq *rq = raw_rq(); | |
6760 | ||
6761 | delayacct_blkio_start(); | |
6762 | atomic_inc(&rq->nr_iowait); | |
6763 | current->in_iowait = 1; | |
6764 | schedule(); | |
6765 | current->in_iowait = 0; | |
6766 | atomic_dec(&rq->nr_iowait); | |
6767 | delayacct_blkio_end(); | |
6768 | } | |
6769 | EXPORT_SYMBOL(io_schedule); | |
6770 | ||
6771 | long __sched io_schedule_timeout(long timeout) | |
6772 | { | |
6773 | struct rq *rq = raw_rq(); | |
6774 | long ret; | |
6775 | ||
6776 | delayacct_blkio_start(); | |
6777 | atomic_inc(&rq->nr_iowait); | |
6778 | current->in_iowait = 1; | |
6779 | ret = schedule_timeout(timeout); | |
6780 | current->in_iowait = 0; | |
6781 | atomic_dec(&rq->nr_iowait); | |
6782 | delayacct_blkio_end(); | |
6783 | return ret; | |
6784 | } | |
6785 | ||
6786 | /** | |
6787 | * sys_sched_get_priority_max - return maximum RT priority. | |
6788 | * @policy: scheduling class. | |
6789 | * | |
6790 | * this syscall returns the maximum rt_priority that can be used | |
6791 | * by a given scheduling class. | |
6792 | */ | |
6793 | SYSCALL_DEFINE1(sched_get_priority_max, int, policy) | |
6794 | { | |
6795 | int ret = -EINVAL; | |
6796 | ||
6797 | switch (policy) { | |
6798 | case SCHED_FIFO: | |
6799 | case SCHED_RR: | |
6800 | ret = MAX_USER_RT_PRIO-1; | |
6801 | break; | |
6802 | case SCHED_NORMAL: | |
6803 | case SCHED_BATCH: | |
6804 | case SCHED_IDLE: | |
6805 | ret = 0; | |
6806 | break; | |
6807 | } | |
6808 | return ret; | |
6809 | } | |
6810 | ||
6811 | /** | |
6812 | * sys_sched_get_priority_min - return minimum RT priority. | |
6813 | * @policy: scheduling class. | |
6814 | * | |
6815 | * this syscall returns the minimum rt_priority that can be used | |
6816 | * by a given scheduling class. | |
6817 | */ | |
6818 | SYSCALL_DEFINE1(sched_get_priority_min, int, policy) | |
6819 | { | |
6820 | int ret = -EINVAL; | |
6821 | ||
6822 | switch (policy) { | |
6823 | case SCHED_FIFO: | |
6824 | case SCHED_RR: | |
6825 | ret = 1; | |
6826 | break; | |
6827 | case SCHED_NORMAL: | |
6828 | case SCHED_BATCH: | |
6829 | case SCHED_IDLE: | |
6830 | ret = 0; | |
6831 | } | |
6832 | return ret; | |
6833 | } | |
6834 | ||
6835 | /** | |
6836 | * sys_sched_rr_get_interval - return the default timeslice of a process. | |
6837 | * @pid: pid of the process. | |
6838 | * @interval: userspace pointer to the timeslice value. | |
6839 | * | |
6840 | * this syscall writes the default timeslice value of a given process | |
6841 | * into the user-space timespec buffer. A value of '0' means infinity. | |
6842 | */ | |
6843 | SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, | |
6844 | struct timespec __user *, interval) | |
6845 | { | |
6846 | struct task_struct *p; | |
6847 | unsigned int time_slice; | |
6848 | int retval; | |
6849 | struct timespec t; | |
6850 | ||
6851 | if (pid < 0) | |
6852 | return -EINVAL; | |
6853 | ||
6854 | retval = -ESRCH; | |
6855 | read_lock(&tasklist_lock); | |
6856 | p = find_process_by_pid(pid); | |
6857 | if (!p) | |
6858 | goto out_unlock; | |
6859 | ||
6860 | retval = security_task_getscheduler(p); | |
6861 | if (retval) | |
6862 | goto out_unlock; | |
6863 | ||
6864 | time_slice = p->sched_class->get_rr_interval(p); | |
6865 | ||
6866 | read_unlock(&tasklist_lock); | |
6867 | jiffies_to_timespec(time_slice, &t); | |
6868 | retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; | |
6869 | return retval; | |
6870 | ||
6871 | out_unlock: | |
6872 | read_unlock(&tasklist_lock); | |
6873 | return retval; | |
6874 | } | |
6875 | ||
6876 | static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; | |
6877 | ||
6878 | void sched_show_task(struct task_struct *p) | |
6879 | { | |
6880 | unsigned long free = 0; | |
6881 | unsigned state; | |
6882 | ||
6883 | state = p->state ? __ffs(p->state) + 1 : 0; | |
6884 | printk(KERN_INFO "%-13.13s %c", p->comm, | |
6885 | state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); | |
6886 | #if BITS_PER_LONG == 32 | |
6887 | if (state == TASK_RUNNING) | |
6888 | printk(KERN_CONT " running "); | |
6889 | else | |
6890 | printk(KERN_CONT " %08lx ", thread_saved_pc(p)); | |
6891 | #else | |
6892 | if (state == TASK_RUNNING) | |
6893 | printk(KERN_CONT " running task "); | |
6894 | else | |
6895 | printk(KERN_CONT " %016lx ", thread_saved_pc(p)); | |
6896 | #endif | |
6897 | #ifdef CONFIG_DEBUG_STACK_USAGE | |
6898 | free = stack_not_used(p); | |
6899 | #endif | |
6900 | printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, | |
6901 | task_pid_nr(p), task_pid_nr(p->real_parent), | |
6902 | (unsigned long)task_thread_info(p)->flags); | |
6903 | ||
6904 | show_stack(p, NULL); | |
6905 | } | |
6906 | ||
6907 | void show_state_filter(unsigned long state_filter) | |
6908 | { | |
6909 | struct task_struct *g, *p; | |
6910 | ||
6911 | #if BITS_PER_LONG == 32 | |
6912 | printk(KERN_INFO | |
6913 | " task PC stack pid father\n"); | |
6914 | #else | |
6915 | printk(KERN_INFO | |
6916 | " task PC stack pid father\n"); | |
6917 | #endif | |
6918 | read_lock(&tasklist_lock); | |
6919 | do_each_thread(g, p) { | |
6920 | /* | |
6921 | * reset the NMI-timeout, listing all files on a slow | |
6922 | * console might take alot of time: | |
6923 | */ | |
6924 | touch_nmi_watchdog(); | |
6925 | if (!state_filter || (p->state & state_filter)) | |
6926 | sched_show_task(p); | |
6927 | } while_each_thread(g, p); | |
6928 | ||
6929 | touch_all_softlockup_watchdogs(); | |
6930 | ||
6931 | #ifdef CONFIG_SCHED_DEBUG | |
6932 | sysrq_sched_debug_show(); | |
6933 | #endif | |
6934 | read_unlock(&tasklist_lock); | |
6935 | /* | |
6936 | * Only show locks if all tasks are dumped: | |
6937 | */ | |
6938 | if (state_filter == -1) | |
6939 | debug_show_all_locks(); | |
6940 | } | |
6941 | ||
6942 | void __cpuinit init_idle_bootup_task(struct task_struct *idle) | |
6943 | { | |
6944 | idle->sched_class = &idle_sched_class; | |
6945 | } | |
6946 | ||
6947 | /** | |
6948 | * init_idle - set up an idle thread for a given CPU | |
6949 | * @idle: task in question | |
6950 | * @cpu: cpu the idle task belongs to | |
6951 | * | |
6952 | * NOTE: this function does not set the idle thread's NEED_RESCHED | |
6953 | * flag, to make booting more robust. | |
6954 | */ | |
6955 | void __cpuinit init_idle(struct task_struct *idle, int cpu) | |
6956 | { | |
6957 | struct rq *rq = cpu_rq(cpu); | |
6958 | unsigned long flags; | |
6959 | ||
6960 | spin_lock_irqsave(&rq->lock, flags); | |
6961 | ||
6962 | __sched_fork(idle); | |
6963 | idle->se.exec_start = sched_clock(); | |
6964 | ||
6965 | idle->prio = idle->normal_prio = MAX_PRIO; | |
6966 | cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu)); | |
6967 | __set_task_cpu(idle, cpu); | |
6968 | ||
6969 | rq->curr = rq->idle = idle; | |
6970 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) | |
6971 | idle->oncpu = 1; | |
6972 | #endif | |
6973 | spin_unlock_irqrestore(&rq->lock, flags); | |
6974 | ||
6975 | /* Set the preempt count _outside_ the spinlocks! */ | |
6976 | #if defined(CONFIG_PREEMPT) | |
6977 | task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0); | |
6978 | #else | |
6979 | task_thread_info(idle)->preempt_count = 0; | |
6980 | #endif | |
6981 | /* | |
6982 | * The idle tasks have their own, simple scheduling class: | |
6983 | */ | |
6984 | idle->sched_class = &idle_sched_class; | |
6985 | ftrace_graph_init_task(idle); | |
6986 | } | |
6987 | ||
6988 | /* | |
6989 | * In a system that switches off the HZ timer nohz_cpu_mask | |
6990 | * indicates which cpus entered this state. This is used | |
6991 | * in the rcu update to wait only for active cpus. For system | |
6992 | * which do not switch off the HZ timer nohz_cpu_mask should | |
6993 | * always be CPU_BITS_NONE. | |
6994 | */ | |
6995 | cpumask_var_t nohz_cpu_mask; | |
6996 | ||
6997 | /* | |
6998 | * Increase the granularity value when there are more CPUs, | |
6999 | * because with more CPUs the 'effective latency' as visible | |
7000 | * to users decreases. But the relationship is not linear, | |
7001 | * so pick a second-best guess by going with the log2 of the | |
7002 | * number of CPUs. | |
7003 | * | |
7004 | * This idea comes from the SD scheduler of Con Kolivas: | |
7005 | */ | |
7006 | static inline void sched_init_granularity(void) | |
7007 | { | |
7008 | unsigned int factor = 1 + ilog2(num_online_cpus()); | |
7009 | const unsigned long limit = 200000000; | |
7010 | ||
7011 | sysctl_sched_min_granularity *= factor; | |
7012 | if (sysctl_sched_min_granularity > limit) | |
7013 | sysctl_sched_min_granularity = limit; | |
7014 | ||
7015 | sysctl_sched_latency *= factor; | |
7016 | if (sysctl_sched_latency > limit) | |
7017 | sysctl_sched_latency = limit; | |
7018 | ||
7019 | sysctl_sched_wakeup_granularity *= factor; | |
7020 | ||
7021 | sysctl_sched_shares_ratelimit *= factor; | |
7022 | } | |
7023 | ||
7024 | #ifdef CONFIG_SMP | |
7025 | /* | |
7026 | * This is how migration works: | |
7027 | * | |
7028 | * 1) we queue a struct migration_req structure in the source CPU's | |
7029 | * runqueue and wake up that CPU's migration thread. | |
7030 | * 2) we down() the locked semaphore => thread blocks. | |
7031 | * 3) migration thread wakes up (implicitly it forces the migrated | |
7032 | * thread off the CPU) | |
7033 | * 4) it gets the migration request and checks whether the migrated | |
7034 | * task is still in the wrong runqueue. | |
7035 | * 5) if it's in the wrong runqueue then the migration thread removes | |
7036 | * it and puts it into the right queue. | |
7037 | * 6) migration thread up()s the semaphore. | |
7038 | * 7) we wake up and the migration is done. | |
7039 | */ | |
7040 | ||
7041 | /* | |
7042 | * Change a given task's CPU affinity. Migrate the thread to a | |
7043 | * proper CPU and schedule it away if the CPU it's executing on | |
7044 | * is removed from the allowed bitmask. | |
7045 | * | |
7046 | * NOTE: the caller must have a valid reference to the task, the | |
7047 | * task must not exit() & deallocate itself prematurely. The | |
7048 | * call is not atomic; no spinlocks may be held. | |
7049 | */ | |
7050 | int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) | |
7051 | { | |
7052 | struct migration_req req; | |
7053 | unsigned long flags; | |
7054 | struct rq *rq; | |
7055 | int ret = 0; | |
7056 | ||
7057 | rq = task_rq_lock(p, &flags); | |
7058 | if (!cpumask_intersects(new_mask, cpu_online_mask)) { | |
7059 | ret = -EINVAL; | |
7060 | goto out; | |
7061 | } | |
7062 | ||
7063 | if (unlikely((p->flags & PF_THREAD_BOUND) && p != current && | |
7064 | !cpumask_equal(&p->cpus_allowed, new_mask))) { | |
7065 | ret = -EINVAL; | |
7066 | goto out; | |
7067 | } | |
7068 | ||
7069 | if (p->sched_class->set_cpus_allowed) | |
7070 | p->sched_class->set_cpus_allowed(p, new_mask); | |
7071 | else { | |
7072 | cpumask_copy(&p->cpus_allowed, new_mask); | |
7073 | p->rt.nr_cpus_allowed = cpumask_weight(new_mask); | |
7074 | } | |
7075 | ||
7076 | /* Can the task run on the task's current CPU? If so, we're done */ | |
7077 | if (cpumask_test_cpu(task_cpu(p), new_mask)) | |
7078 | goto out; | |
7079 | ||
7080 | if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) { | |
7081 | /* Need help from migration thread: drop lock and wait. */ | |
7082 | struct task_struct *mt = rq->migration_thread; | |
7083 | ||
7084 | get_task_struct(mt); | |
7085 | task_rq_unlock(rq, &flags); | |
7086 | wake_up_process(rq->migration_thread); | |
7087 | put_task_struct(mt); | |
7088 | wait_for_completion(&req.done); | |
7089 | tlb_migrate_finish(p->mm); | |
7090 | return 0; | |
7091 | } | |
7092 | out: | |
7093 | task_rq_unlock(rq, &flags); | |
7094 | ||
7095 | return ret; | |
7096 | } | |
7097 | EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); | |
7098 | ||
7099 | /* | |
7100 | * Move (not current) task off this cpu, onto dest cpu. We're doing | |
7101 | * this because either it can't run here any more (set_cpus_allowed() | |
7102 | * away from this CPU, or CPU going down), or because we're | |
7103 | * attempting to rebalance this task on exec (sched_exec). | |
7104 | * | |
7105 | * So we race with normal scheduler movements, but that's OK, as long | |
7106 | * as the task is no longer on this CPU. | |
7107 | * | |
7108 | * Returns non-zero if task was successfully migrated. | |
7109 | */ | |
7110 | static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) | |
7111 | { | |
7112 | struct rq *rq_dest, *rq_src; | |
7113 | int ret = 0, on_rq; | |
7114 | ||
7115 | if (unlikely(!cpu_active(dest_cpu))) | |
7116 | return ret; | |
7117 | ||
7118 | rq_src = cpu_rq(src_cpu); | |
7119 | rq_dest = cpu_rq(dest_cpu); | |
7120 | ||
7121 | double_rq_lock(rq_src, rq_dest); | |
7122 | /* Already moved. */ | |
7123 | if (task_cpu(p) != src_cpu) | |
7124 | goto done; | |
7125 | /* Affinity changed (again). */ | |
7126 | if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)) | |
7127 | goto fail; | |
7128 | ||
7129 | on_rq = p->se.on_rq; | |
7130 | if (on_rq) | |
7131 | deactivate_task(rq_src, p, 0); | |
7132 | ||
7133 | set_task_cpu(p, dest_cpu); | |
7134 | if (on_rq) { | |
7135 | activate_task(rq_dest, p, 0); | |
7136 | check_preempt_curr(rq_dest, p, 0); | |
7137 | } | |
7138 | done: | |
7139 | ret = 1; | |
7140 | fail: | |
7141 | double_rq_unlock(rq_src, rq_dest); | |
7142 | return ret; | |
7143 | } | |
7144 | ||
7145 | #define RCU_MIGRATION_IDLE 0 | |
7146 | #define RCU_MIGRATION_NEED_QS 1 | |
7147 | #define RCU_MIGRATION_GOT_QS 2 | |
7148 | #define RCU_MIGRATION_MUST_SYNC 3 | |
7149 | ||
7150 | /* | |
7151 | * migration_thread - this is a highprio system thread that performs | |
7152 | * thread migration by bumping thread off CPU then 'pushing' onto | |
7153 | * another runqueue. | |
7154 | */ | |
7155 | static int migration_thread(void *data) | |
7156 | { | |
7157 | int badcpu; | |
7158 | int cpu = (long)data; | |
7159 | struct rq *rq; | |
7160 | ||
7161 | rq = cpu_rq(cpu); | |
7162 | BUG_ON(rq->migration_thread != current); | |
7163 | ||
7164 | set_current_state(TASK_INTERRUPTIBLE); | |
7165 | while (!kthread_should_stop()) { | |
7166 | struct migration_req *req; | |
7167 | struct list_head *head; | |
7168 | ||
7169 | spin_lock_irq(&rq->lock); | |
7170 | ||
7171 | if (cpu_is_offline(cpu)) { | |
7172 | spin_unlock_irq(&rq->lock); | |
7173 | break; | |
7174 | } | |
7175 | ||
7176 | if (rq->active_balance) { | |
7177 | active_load_balance(rq, cpu); | |
7178 | rq->active_balance = 0; | |
7179 | } | |
7180 | ||
7181 | head = &rq->migration_queue; | |
7182 | ||
7183 | if (list_empty(head)) { | |
7184 | spin_unlock_irq(&rq->lock); | |
7185 | schedule(); | |
7186 | set_current_state(TASK_INTERRUPTIBLE); | |
7187 | continue; | |
7188 | } | |
7189 | req = list_entry(head->next, struct migration_req, list); | |
7190 | list_del_init(head->next); | |
7191 | ||
7192 | if (req->task != NULL) { | |
7193 | spin_unlock(&rq->lock); | |
7194 | __migrate_task(req->task, cpu, req->dest_cpu); | |
7195 | } else if (likely(cpu == (badcpu = smp_processor_id()))) { | |
7196 | req->dest_cpu = RCU_MIGRATION_GOT_QS; | |
7197 | spin_unlock(&rq->lock); | |
7198 | } else { | |
7199 | req->dest_cpu = RCU_MIGRATION_MUST_SYNC; | |
7200 | spin_unlock(&rq->lock); | |
7201 | WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu); | |
7202 | } | |
7203 | local_irq_enable(); | |
7204 | ||
7205 | complete(&req->done); | |
7206 | } | |
7207 | __set_current_state(TASK_RUNNING); | |
7208 | ||
7209 | return 0; | |
7210 | } | |
7211 | ||
7212 | #ifdef CONFIG_HOTPLUG_CPU | |
7213 | ||
7214 | static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu) | |
7215 | { | |
7216 | int ret; | |
7217 | ||
7218 | local_irq_disable(); | |
7219 | ret = __migrate_task(p, src_cpu, dest_cpu); | |
7220 | local_irq_enable(); | |
7221 | return ret; | |
7222 | } | |
7223 | ||
7224 | /* | |
7225 | * Figure out where task on dead CPU should go, use force if necessary. | |
7226 | */ | |
7227 | static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p) | |
7228 | { | |
7229 | int dest_cpu; | |
7230 | const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu)); | |
7231 | ||
7232 | again: | |
7233 | /* Look for allowed, online CPU in same node. */ | |
7234 | for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask) | |
7235 | if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed)) | |
7236 | goto move; | |
7237 | ||
7238 | /* Any allowed, online CPU? */ | |
7239 | dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask); | |
7240 | if (dest_cpu < nr_cpu_ids) | |
7241 | goto move; | |
7242 | ||
7243 | /* No more Mr. Nice Guy. */ | |
7244 | if (dest_cpu >= nr_cpu_ids) { | |
7245 | cpuset_cpus_allowed_locked(p, &p->cpus_allowed); | |
7246 | dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed); | |
7247 | ||
7248 | /* | |
7249 | * Don't tell them about moving exiting tasks or | |
7250 | * kernel threads (both mm NULL), since they never | |
7251 | * leave kernel. | |
7252 | */ | |
7253 | if (p->mm && printk_ratelimit()) { | |
7254 | printk(KERN_INFO "process %d (%s) no " | |
7255 | "longer affine to cpu%d\n", | |
7256 | task_pid_nr(p), p->comm, dead_cpu); | |
7257 | } | |
7258 | } | |
7259 | ||
7260 | move: | |
7261 | /* It can have affinity changed while we were choosing. */ | |
7262 | if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu))) | |
7263 | goto again; | |
7264 | } | |
7265 | ||
7266 | /* | |
7267 | * While a dead CPU has no uninterruptible tasks queued at this point, | |
7268 | * it might still have a nonzero ->nr_uninterruptible counter, because | |
7269 | * for performance reasons the counter is not stricly tracking tasks to | |
7270 | * their home CPUs. So we just add the counter to another CPU's counter, | |
7271 | * to keep the global sum constant after CPU-down: | |
7272 | */ | |
7273 | static void migrate_nr_uninterruptible(struct rq *rq_src) | |
7274 | { | |
7275 | struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask)); | |
7276 | unsigned long flags; | |
7277 | ||
7278 | local_irq_save(flags); | |
7279 | double_rq_lock(rq_src, rq_dest); | |
7280 | rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible; | |
7281 | rq_src->nr_uninterruptible = 0; | |
7282 | double_rq_unlock(rq_src, rq_dest); | |
7283 | local_irq_restore(flags); | |
7284 | } | |
7285 | ||
7286 | /* Run through task list and migrate tasks from the dead cpu. */ | |
7287 | static void migrate_live_tasks(int src_cpu) | |
7288 | { | |
7289 | struct task_struct *p, *t; | |
7290 | ||
7291 | read_lock(&tasklist_lock); | |
7292 | ||
7293 | do_each_thread(t, p) { | |
7294 | if (p == current) | |
7295 | continue; | |
7296 | ||
7297 | if (task_cpu(p) == src_cpu) | |
7298 | move_task_off_dead_cpu(src_cpu, p); | |
7299 | } while_each_thread(t, p); | |
7300 | ||
7301 | read_unlock(&tasklist_lock); | |
7302 | } | |
7303 | ||
7304 | /* | |
7305 | * Schedules idle task to be the next runnable task on current CPU. | |
7306 | * It does so by boosting its priority to highest possible. | |
7307 | * Used by CPU offline code. | |
7308 | */ | |
7309 | void sched_idle_next(void) | |
7310 | { | |
7311 | int this_cpu = smp_processor_id(); | |
7312 | struct rq *rq = cpu_rq(this_cpu); | |
7313 | struct task_struct *p = rq->idle; | |
7314 | unsigned long flags; | |
7315 | ||
7316 | /* cpu has to be offline */ | |
7317 | BUG_ON(cpu_online(this_cpu)); | |
7318 | ||
7319 | /* | |
7320 | * Strictly not necessary since rest of the CPUs are stopped by now | |
7321 | * and interrupts disabled on the current cpu. | |
7322 | */ | |
7323 | spin_lock_irqsave(&rq->lock, flags); | |
7324 | ||
7325 | __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1); | |
7326 | ||
7327 | update_rq_clock(rq); | |
7328 | activate_task(rq, p, 0); | |
7329 | ||
7330 | spin_unlock_irqrestore(&rq->lock, flags); | |
7331 | } | |
7332 | ||
7333 | /* | |
7334 | * Ensures that the idle task is using init_mm right before its cpu goes | |
7335 | * offline. | |
7336 | */ | |
7337 | void idle_task_exit(void) | |
7338 | { | |
7339 | struct mm_struct *mm = current->active_mm; | |
7340 | ||
7341 | BUG_ON(cpu_online(smp_processor_id())); | |
7342 | ||
7343 | if (mm != &init_mm) | |
7344 | switch_mm(mm, &init_mm, current); | |
7345 | mmdrop(mm); | |
7346 | } | |
7347 | ||
7348 | /* called under rq->lock with disabled interrupts */ | |
7349 | static void migrate_dead(unsigned int dead_cpu, struct task_struct *p) | |
7350 | { | |
7351 | struct rq *rq = cpu_rq(dead_cpu); | |
7352 | ||
7353 | /* Must be exiting, otherwise would be on tasklist. */ | |
7354 | BUG_ON(!p->exit_state); | |
7355 | ||
7356 | /* Cannot have done final schedule yet: would have vanished. */ | |
7357 | BUG_ON(p->state == TASK_DEAD); | |
7358 | ||
7359 | get_task_struct(p); | |
7360 | ||
7361 | /* | |
7362 | * Drop lock around migration; if someone else moves it, | |
7363 | * that's OK. No task can be added to this CPU, so iteration is | |
7364 | * fine. | |
7365 | */ | |
7366 | spin_unlock_irq(&rq->lock); | |
7367 | move_task_off_dead_cpu(dead_cpu, p); | |
7368 | spin_lock_irq(&rq->lock); | |
7369 | ||
7370 | put_task_struct(p); | |
7371 | } | |
7372 | ||
7373 | /* release_task() removes task from tasklist, so we won't find dead tasks. */ | |
7374 | static void migrate_dead_tasks(unsigned int dead_cpu) | |
7375 | { | |
7376 | struct rq *rq = cpu_rq(dead_cpu); | |
7377 | struct task_struct *next; | |
7378 | ||
7379 | for ( ; ; ) { | |
7380 | if (!rq->nr_running) | |
7381 | break; | |
7382 | update_rq_clock(rq); | |
7383 | next = pick_next_task(rq); | |
7384 | if (!next) | |
7385 | break; | |
7386 | next->sched_class->put_prev_task(rq, next); | |
7387 | migrate_dead(dead_cpu, next); | |
7388 | ||
7389 | } | |
7390 | } | |
7391 | ||
7392 | /* | |
7393 | * remove the tasks which were accounted by rq from calc_load_tasks. | |
7394 | */ | |
7395 | static void calc_global_load_remove(struct rq *rq) | |
7396 | { | |
7397 | atomic_long_sub(rq->calc_load_active, &calc_load_tasks); | |
7398 | rq->calc_load_active = 0; | |
7399 | } | |
7400 | #endif /* CONFIG_HOTPLUG_CPU */ | |
7401 | ||
7402 | #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) | |
7403 | ||
7404 | static struct ctl_table sd_ctl_dir[] = { | |
7405 | { | |
7406 | .procname = "sched_domain", | |
7407 | .mode = 0555, | |
7408 | }, | |
7409 | {0, }, | |
7410 | }; | |
7411 | ||
7412 | static struct ctl_table sd_ctl_root[] = { | |
7413 | { | |
7414 | .ctl_name = CTL_KERN, | |
7415 | .procname = "kernel", | |
7416 | .mode = 0555, | |
7417 | .child = sd_ctl_dir, | |
7418 | }, | |
7419 | {0, }, | |
7420 | }; | |
7421 | ||
7422 | static struct ctl_table *sd_alloc_ctl_entry(int n) | |
7423 | { | |
7424 | struct ctl_table *entry = | |
7425 | kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); | |
7426 | ||
7427 | return entry; | |
7428 | } | |
7429 | ||
7430 | static void sd_free_ctl_entry(struct ctl_table **tablep) | |
7431 | { | |
7432 | struct ctl_table *entry; | |
7433 | ||
7434 | /* | |
7435 | * In the intermediate directories, both the child directory and | |
7436 | * procname are dynamically allocated and could fail but the mode | |
7437 | * will always be set. In the lowest directory the names are | |
7438 | * static strings and all have proc handlers. | |
7439 | */ | |
7440 | for (entry = *tablep; entry->mode; entry++) { | |
7441 | if (entry->child) | |
7442 | sd_free_ctl_entry(&entry->child); | |
7443 | if (entry->proc_handler == NULL) | |
7444 | kfree(entry->procname); | |
7445 | } | |
7446 | ||
7447 | kfree(*tablep); | |
7448 | *tablep = NULL; | |
7449 | } | |
7450 | ||
7451 | static void | |
7452 | set_table_entry(struct ctl_table *entry, | |
7453 | const char *procname, void *data, int maxlen, | |
7454 | mode_t mode, proc_handler *proc_handler) | |
7455 | { | |
7456 | entry->procname = procname; | |
7457 | entry->data = data; | |
7458 | entry->maxlen = maxlen; | |
7459 | entry->mode = mode; | |
7460 | entry->proc_handler = proc_handler; | |
7461 | } | |
7462 | ||
7463 | static struct ctl_table * | |
7464 | sd_alloc_ctl_domain_table(struct sched_domain *sd) | |
7465 | { | |
7466 | struct ctl_table *table = sd_alloc_ctl_entry(13); | |
7467 | ||
7468 | if (table == NULL) | |
7469 | return NULL; | |
7470 | ||
7471 | set_table_entry(&table[0], "min_interval", &sd->min_interval, | |
7472 | sizeof(long), 0644, proc_doulongvec_minmax); | |
7473 | set_table_entry(&table[1], "max_interval", &sd->max_interval, | |
7474 | sizeof(long), 0644, proc_doulongvec_minmax); | |
7475 | set_table_entry(&table[2], "busy_idx", &sd->busy_idx, | |
7476 | sizeof(int), 0644, proc_dointvec_minmax); | |
7477 | set_table_entry(&table[3], "idle_idx", &sd->idle_idx, | |
7478 | sizeof(int), 0644, proc_dointvec_minmax); | |
7479 | set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, | |
7480 | sizeof(int), 0644, proc_dointvec_minmax); | |
7481 | set_table_entry(&table[5], "wake_idx", &sd->wake_idx, | |
7482 | sizeof(int), 0644, proc_dointvec_minmax); | |
7483 | set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, | |
7484 | sizeof(int), 0644, proc_dointvec_minmax); | |
7485 | set_table_entry(&table[7], "busy_factor", &sd->busy_factor, | |
7486 | sizeof(int), 0644, proc_dointvec_minmax); | |
7487 | set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, | |
7488 | sizeof(int), 0644, proc_dointvec_minmax); | |
7489 | set_table_entry(&table[9], "cache_nice_tries", | |
7490 | &sd->cache_nice_tries, | |
7491 | sizeof(int), 0644, proc_dointvec_minmax); | |
7492 | set_table_entry(&table[10], "flags", &sd->flags, | |
7493 | sizeof(int), 0644, proc_dointvec_minmax); | |
7494 | set_table_entry(&table[11], "name", sd->name, | |
7495 | CORENAME_MAX_SIZE, 0444, proc_dostring); | |
7496 | /* &table[12] is terminator */ | |
7497 | ||
7498 | return table; | |
7499 | } | |
7500 | ||
7501 | static ctl_table *sd_alloc_ctl_cpu_table(int cpu) | |
7502 | { | |
7503 | struct ctl_table *entry, *table; | |
7504 | struct sched_domain *sd; | |
7505 | int domain_num = 0, i; | |
7506 | char buf[32]; | |
7507 | ||
7508 | for_each_domain(cpu, sd) | |
7509 | domain_num++; | |
7510 | entry = table = sd_alloc_ctl_entry(domain_num + 1); | |
7511 | if (table == NULL) | |
7512 | return NULL; | |
7513 | ||
7514 | i = 0; | |
7515 | for_each_domain(cpu, sd) { | |
7516 | snprintf(buf, 32, "domain%d", i); | |
7517 | entry->procname = kstrdup(buf, GFP_KERNEL); | |
7518 | entry->mode = 0555; | |
7519 | entry->child = sd_alloc_ctl_domain_table(sd); | |
7520 | entry++; | |
7521 | i++; | |
7522 | } | |
7523 | return table; | |
7524 | } | |
7525 | ||
7526 | static struct ctl_table_header *sd_sysctl_header; | |
7527 | static void register_sched_domain_sysctl(void) | |
7528 | { | |
7529 | int i, cpu_num = num_online_cpus(); | |
7530 | struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); | |
7531 | char buf[32]; | |
7532 | ||
7533 | WARN_ON(sd_ctl_dir[0].child); | |
7534 | sd_ctl_dir[0].child = entry; | |
7535 | ||
7536 | if (entry == NULL) | |
7537 | return; | |
7538 | ||
7539 | for_each_online_cpu(i) { | |
7540 | snprintf(buf, 32, "cpu%d", i); | |
7541 | entry->procname = kstrdup(buf, GFP_KERNEL); | |
7542 | entry->mode = 0555; | |
7543 | entry->child = sd_alloc_ctl_cpu_table(i); | |
7544 | entry++; | |
7545 | } | |
7546 | ||
7547 | WARN_ON(sd_sysctl_header); | |
7548 | sd_sysctl_header = register_sysctl_table(sd_ctl_root); | |
7549 | } | |
7550 | ||
7551 | /* may be called multiple times per register */ | |
7552 | static void unregister_sched_domain_sysctl(void) | |
7553 | { | |
7554 | if (sd_sysctl_header) | |
7555 | unregister_sysctl_table(sd_sysctl_header); | |
7556 | sd_sysctl_header = NULL; | |
7557 | if (sd_ctl_dir[0].child) | |
7558 | sd_free_ctl_entry(&sd_ctl_dir[0].child); | |
7559 | } | |
7560 | #else | |
7561 | static void register_sched_domain_sysctl(void) | |
7562 | { | |
7563 | } | |
7564 | static void unregister_sched_domain_sysctl(void) | |
7565 | { | |
7566 | } | |
7567 | #endif | |
7568 | ||
7569 | static void set_rq_online(struct rq *rq) | |
7570 | { | |
7571 | if (!rq->online) { | |
7572 | const struct sched_class *class; | |
7573 | ||
7574 | cpumask_set_cpu(rq->cpu, rq->rd->online); | |
7575 | rq->online = 1; | |
7576 | ||
7577 | for_each_class(class) { | |
7578 | if (class->rq_online) | |
7579 | class->rq_online(rq); | |
7580 | } | |
7581 | } | |
7582 | } | |
7583 | ||
7584 | static void set_rq_offline(struct rq *rq) | |
7585 | { | |
7586 | if (rq->online) { | |
7587 | const struct sched_class *class; | |
7588 | ||
7589 | for_each_class(class) { | |
7590 | if (class->rq_offline) | |
7591 | class->rq_offline(rq); | |
7592 | } | |
7593 | ||
7594 | cpumask_clear_cpu(rq->cpu, rq->rd->online); | |
7595 | rq->online = 0; | |
7596 | } | |
7597 | } | |
7598 | ||
7599 | /* | |
7600 | * migration_call - callback that gets triggered when a CPU is added. | |
7601 | * Here we can start up the necessary migration thread for the new CPU. | |
7602 | */ | |
7603 | static int __cpuinit | |
7604 | migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) | |
7605 | { | |
7606 | struct task_struct *p; | |
7607 | int cpu = (long)hcpu; | |
7608 | unsigned long flags; | |
7609 | struct rq *rq; | |
7610 | ||
7611 | switch (action) { | |
7612 | ||
7613 | case CPU_UP_PREPARE: | |
7614 | case CPU_UP_PREPARE_FROZEN: | |
7615 | p = kthread_create(migration_thread, hcpu, "migration/%d", cpu); | |
7616 | if (IS_ERR(p)) | |
7617 | return NOTIFY_BAD; | |
7618 | kthread_bind(p, cpu); | |
7619 | /* Must be high prio: stop_machine expects to yield to it. */ | |
7620 | rq = task_rq_lock(p, &flags); | |
7621 | __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1); | |
7622 | task_rq_unlock(rq, &flags); | |
7623 | get_task_struct(p); | |
7624 | cpu_rq(cpu)->migration_thread = p; | |
7625 | rq->calc_load_update = calc_load_update; | |
7626 | break; | |
7627 | ||
7628 | case CPU_ONLINE: | |
7629 | case CPU_ONLINE_FROZEN: | |
7630 | /* Strictly unnecessary, as first user will wake it. */ | |
7631 | wake_up_process(cpu_rq(cpu)->migration_thread); | |
7632 | ||
7633 | /* Update our root-domain */ | |
7634 | rq = cpu_rq(cpu); | |
7635 | spin_lock_irqsave(&rq->lock, flags); | |
7636 | if (rq->rd) { | |
7637 | BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); | |
7638 | ||
7639 | set_rq_online(rq); | |
7640 | } | |
7641 | spin_unlock_irqrestore(&rq->lock, flags); | |
7642 | break; | |
7643 | ||
7644 | #ifdef CONFIG_HOTPLUG_CPU | |
7645 | case CPU_UP_CANCELED: | |
7646 | case CPU_UP_CANCELED_FROZEN: | |
7647 | if (!cpu_rq(cpu)->migration_thread) | |
7648 | break; | |
7649 | /* Unbind it from offline cpu so it can run. Fall thru. */ | |
7650 | kthread_bind(cpu_rq(cpu)->migration_thread, | |
7651 | cpumask_any(cpu_online_mask)); | |
7652 | kthread_stop(cpu_rq(cpu)->migration_thread); | |
7653 | put_task_struct(cpu_rq(cpu)->migration_thread); | |
7654 | cpu_rq(cpu)->migration_thread = NULL; | |
7655 | break; | |
7656 | ||
7657 | case CPU_DEAD: | |
7658 | case CPU_DEAD_FROZEN: | |
7659 | cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */ | |
7660 | migrate_live_tasks(cpu); | |
7661 | rq = cpu_rq(cpu); | |
7662 | kthread_stop(rq->migration_thread); | |
7663 | put_task_struct(rq->migration_thread); | |
7664 | rq->migration_thread = NULL; | |
7665 | /* Idle task back to normal (off runqueue, low prio) */ | |
7666 | spin_lock_irq(&rq->lock); | |
7667 | update_rq_clock(rq); | |
7668 | deactivate_task(rq, rq->idle, 0); | |
7669 | rq->idle->static_prio = MAX_PRIO; | |
7670 | __setscheduler(rq, rq->idle, SCHED_NORMAL, 0); | |
7671 | rq->idle->sched_class = &idle_sched_class; | |
7672 | migrate_dead_tasks(cpu); | |
7673 | spin_unlock_irq(&rq->lock); | |
7674 | cpuset_unlock(); | |
7675 | migrate_nr_uninterruptible(rq); | |
7676 | BUG_ON(rq->nr_running != 0); | |
7677 | calc_global_load_remove(rq); | |
7678 | /* | |
7679 | * No need to migrate the tasks: it was best-effort if | |
7680 | * they didn't take sched_hotcpu_mutex. Just wake up | |
7681 | * the requestors. | |
7682 | */ | |
7683 | spin_lock_irq(&rq->lock); | |
7684 | while (!list_empty(&rq->migration_queue)) { | |
7685 | struct migration_req *req; | |
7686 | ||
7687 | req = list_entry(rq->migration_queue.next, | |
7688 | struct migration_req, list); | |
7689 | list_del_init(&req->list); | |
7690 | spin_unlock_irq(&rq->lock); | |
7691 | complete(&req->done); | |
7692 | spin_lock_irq(&rq->lock); | |
7693 | } | |
7694 | spin_unlock_irq(&rq->lock); | |
7695 | break; | |
7696 | ||
7697 | case CPU_DYING: | |
7698 | case CPU_DYING_FROZEN: | |
7699 | /* Update our root-domain */ | |
7700 | rq = cpu_rq(cpu); | |
7701 | spin_lock_irqsave(&rq->lock, flags); | |
7702 | if (rq->rd) { | |
7703 | BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); | |
7704 | set_rq_offline(rq); | |
7705 | } | |
7706 | spin_unlock_irqrestore(&rq->lock, flags); | |
7707 | break; | |
7708 | #endif | |
7709 | } | |
7710 | return NOTIFY_OK; | |
7711 | } | |
7712 | ||
7713 | /* | |
7714 | * Register at high priority so that task migration (migrate_all_tasks) | |
7715 | * happens before everything else. This has to be lower priority than | |
7716 | * the notifier in the perf_event subsystem, though. | |
7717 | */ | |
7718 | static struct notifier_block __cpuinitdata migration_notifier = { | |
7719 | .notifier_call = migration_call, | |
7720 | .priority = 10 | |
7721 | }; | |
7722 | ||
7723 | static int __init migration_init(void) | |
7724 | { | |
7725 | void *cpu = (void *)(long)smp_processor_id(); | |
7726 | int err; | |
7727 | ||
7728 | /* Start one for the boot CPU: */ | |
7729 | err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); | |
7730 | BUG_ON(err == NOTIFY_BAD); | |
7731 | migration_call(&migration_notifier, CPU_ONLINE, cpu); | |
7732 | register_cpu_notifier(&migration_notifier); | |
7733 | ||
7734 | return 0; | |
7735 | } | |
7736 | early_initcall(migration_init); | |
7737 | #endif | |
7738 | ||
7739 | #ifdef CONFIG_SMP | |
7740 | ||
7741 | #ifdef CONFIG_SCHED_DEBUG | |
7742 | ||
7743 | static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, | |
7744 | struct cpumask *groupmask) | |
7745 | { | |
7746 | struct sched_group *group = sd->groups; | |
7747 | char str[256]; | |
7748 | ||
7749 | cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); | |
7750 | cpumask_clear(groupmask); | |
7751 | ||
7752 | printk(KERN_DEBUG "%*s domain %d: ", level, "", level); | |
7753 | ||
7754 | if (!(sd->flags & SD_LOAD_BALANCE)) { | |
7755 | printk("does not load-balance\n"); | |
7756 | if (sd->parent) | |
7757 | printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" | |
7758 | " has parent"); | |
7759 | return -1; | |
7760 | } | |
7761 | ||
7762 | printk(KERN_CONT "span %s level %s\n", str, sd->name); | |
7763 | ||
7764 | if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { | |
7765 | printk(KERN_ERR "ERROR: domain->span does not contain " | |
7766 | "CPU%d\n", cpu); | |
7767 | } | |
7768 | if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { | |
7769 | printk(KERN_ERR "ERROR: domain->groups does not contain" | |
7770 | " CPU%d\n", cpu); | |
7771 | } | |
7772 | ||
7773 | printk(KERN_DEBUG "%*s groups:", level + 1, ""); | |
7774 | do { | |
7775 | if (!group) { | |
7776 | printk("\n"); | |
7777 | printk(KERN_ERR "ERROR: group is NULL\n"); | |
7778 | break; | |
7779 | } | |
7780 | ||
7781 | if (!group->cpu_power) { | |
7782 | printk(KERN_CONT "\n"); | |
7783 | printk(KERN_ERR "ERROR: domain->cpu_power not " | |
7784 | "set\n"); | |
7785 | break; | |
7786 | } | |
7787 | ||
7788 | if (!cpumask_weight(sched_group_cpus(group))) { | |
7789 | printk(KERN_CONT "\n"); | |
7790 | printk(KERN_ERR "ERROR: empty group\n"); | |
7791 | break; | |
7792 | } | |
7793 | ||
7794 | if (cpumask_intersects(groupmask, sched_group_cpus(group))) { | |
7795 | printk(KERN_CONT "\n"); | |
7796 | printk(KERN_ERR "ERROR: repeated CPUs\n"); | |
7797 | break; | |
7798 | } | |
7799 | ||
7800 | cpumask_or(groupmask, groupmask, sched_group_cpus(group)); | |
7801 | ||
7802 | cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); | |
7803 | ||
7804 | printk(KERN_CONT " %s", str); | |
7805 | if (group->cpu_power != SCHED_LOAD_SCALE) { | |
7806 | printk(KERN_CONT " (cpu_power = %d)", | |
7807 | group->cpu_power); | |
7808 | } | |
7809 | ||
7810 | group = group->next; | |
7811 | } while (group != sd->groups); | |
7812 | printk(KERN_CONT "\n"); | |
7813 | ||
7814 | if (!cpumask_equal(sched_domain_span(sd), groupmask)) | |
7815 | printk(KERN_ERR "ERROR: groups don't span domain->span\n"); | |
7816 | ||
7817 | if (sd->parent && | |
7818 | !cpumask_subset(groupmask, sched_domain_span(sd->parent))) | |
7819 | printk(KERN_ERR "ERROR: parent span is not a superset " | |
7820 | "of domain->span\n"); | |
7821 | return 0; | |
7822 | } | |
7823 | ||
7824 | static void sched_domain_debug(struct sched_domain *sd, int cpu) | |
7825 | { | |
7826 | cpumask_var_t groupmask; | |
7827 | int level = 0; | |
7828 | ||
7829 | if (!sd) { | |
7830 | printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); | |
7831 | return; | |
7832 | } | |
7833 | ||
7834 | printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); | |
7835 | ||
7836 | if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) { | |
7837 | printk(KERN_DEBUG "Cannot load-balance (out of memory)\n"); | |
7838 | return; | |
7839 | } | |
7840 | ||
7841 | for (;;) { | |
7842 | if (sched_domain_debug_one(sd, cpu, level, groupmask)) | |
7843 | break; | |
7844 | level++; | |
7845 | sd = sd->parent; | |
7846 | if (!sd) | |
7847 | break; | |
7848 | } | |
7849 | free_cpumask_var(groupmask); | |
7850 | } | |
7851 | #else /* !CONFIG_SCHED_DEBUG */ | |
7852 | # define sched_domain_debug(sd, cpu) do { } while (0) | |
7853 | #endif /* CONFIG_SCHED_DEBUG */ | |
7854 | ||
7855 | static int sd_degenerate(struct sched_domain *sd) | |
7856 | { | |
7857 | if (cpumask_weight(sched_domain_span(sd)) == 1) | |
7858 | return 1; | |
7859 | ||
7860 | /* Following flags need at least 2 groups */ | |
7861 | if (sd->flags & (SD_LOAD_BALANCE | | |
7862 | SD_BALANCE_NEWIDLE | | |
7863 | SD_BALANCE_FORK | | |
7864 | SD_BALANCE_EXEC | | |
7865 | SD_SHARE_CPUPOWER | | |
7866 | SD_SHARE_PKG_RESOURCES)) { | |
7867 | if (sd->groups != sd->groups->next) | |
7868 | return 0; | |
7869 | } | |
7870 | ||
7871 | /* Following flags don't use groups */ | |
7872 | if (sd->flags & (SD_WAKE_AFFINE)) | |
7873 | return 0; | |
7874 | ||
7875 | return 1; | |
7876 | } | |
7877 | ||
7878 | static int | |
7879 | sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) | |
7880 | { | |
7881 | unsigned long cflags = sd->flags, pflags = parent->flags; | |
7882 | ||
7883 | if (sd_degenerate(parent)) | |
7884 | return 1; | |
7885 | ||
7886 | if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) | |
7887 | return 0; | |
7888 | ||
7889 | /* Flags needing groups don't count if only 1 group in parent */ | |
7890 | if (parent->groups == parent->groups->next) { | |
7891 | pflags &= ~(SD_LOAD_BALANCE | | |
7892 | SD_BALANCE_NEWIDLE | | |
7893 | SD_BALANCE_FORK | | |
7894 | SD_BALANCE_EXEC | | |
7895 | SD_SHARE_CPUPOWER | | |
7896 | SD_SHARE_PKG_RESOURCES); | |
7897 | if (nr_node_ids == 1) | |
7898 | pflags &= ~SD_SERIALIZE; | |
7899 | } | |
7900 | if (~cflags & pflags) | |
7901 | return 0; | |
7902 | ||
7903 | return 1; | |
7904 | } | |
7905 | ||
7906 | static void free_rootdomain(struct root_domain *rd) | |
7907 | { | |
7908 | cpupri_cleanup(&rd->cpupri); | |
7909 | ||
7910 | free_cpumask_var(rd->rto_mask); | |
7911 | free_cpumask_var(rd->online); | |
7912 | free_cpumask_var(rd->span); | |
7913 | kfree(rd); | |
7914 | } | |
7915 | ||
7916 | static void rq_attach_root(struct rq *rq, struct root_domain *rd) | |
7917 | { | |
7918 | struct root_domain *old_rd = NULL; | |
7919 | unsigned long flags; | |
7920 | ||
7921 | spin_lock_irqsave(&rq->lock, flags); | |
7922 | ||
7923 | if (rq->rd) { | |
7924 | old_rd = rq->rd; | |
7925 | ||
7926 | if (cpumask_test_cpu(rq->cpu, old_rd->online)) | |
7927 | set_rq_offline(rq); | |
7928 | ||
7929 | cpumask_clear_cpu(rq->cpu, old_rd->span); | |
7930 | ||
7931 | /* | |
7932 | * If we dont want to free the old_rt yet then | |
7933 | * set old_rd to NULL to skip the freeing later | |
7934 | * in this function: | |
7935 | */ | |
7936 | if (!atomic_dec_and_test(&old_rd->refcount)) | |
7937 | old_rd = NULL; | |
7938 | } | |
7939 | ||
7940 | atomic_inc(&rd->refcount); | |
7941 | rq->rd = rd; | |
7942 | ||
7943 | cpumask_set_cpu(rq->cpu, rd->span); | |
7944 | if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) | |
7945 | set_rq_online(rq); | |
7946 | ||
7947 | spin_unlock_irqrestore(&rq->lock, flags); | |
7948 | ||
7949 | if (old_rd) | |
7950 | free_rootdomain(old_rd); | |
7951 | } | |
7952 | ||
7953 | static int init_rootdomain(struct root_domain *rd, bool bootmem) | |
7954 | { | |
7955 | gfp_t gfp = GFP_KERNEL; | |
7956 | ||
7957 | memset(rd, 0, sizeof(*rd)); | |
7958 | ||
7959 | if (bootmem) | |
7960 | gfp = GFP_NOWAIT; | |
7961 | ||
7962 | if (!alloc_cpumask_var(&rd->span, gfp)) | |
7963 | goto out; | |
7964 | if (!alloc_cpumask_var(&rd->online, gfp)) | |
7965 | goto free_span; | |
7966 | if (!alloc_cpumask_var(&rd->rto_mask, gfp)) | |
7967 | goto free_online; | |
7968 | ||
7969 | if (cpupri_init(&rd->cpupri, bootmem) != 0) | |
7970 | goto free_rto_mask; | |
7971 | return 0; | |
7972 | ||
7973 | free_rto_mask: | |
7974 | free_cpumask_var(rd->rto_mask); | |
7975 | free_online: | |
7976 | free_cpumask_var(rd->online); | |
7977 | free_span: | |
7978 | free_cpumask_var(rd->span); | |
7979 | out: | |
7980 | return -ENOMEM; | |
7981 | } | |
7982 | ||
7983 | static void init_defrootdomain(void) | |
7984 | { | |
7985 | init_rootdomain(&def_root_domain, true); | |
7986 | ||
7987 | atomic_set(&def_root_domain.refcount, 1); | |
7988 | } | |
7989 | ||
7990 | static struct root_domain *alloc_rootdomain(void) | |
7991 | { | |
7992 | struct root_domain *rd; | |
7993 | ||
7994 | rd = kmalloc(sizeof(*rd), GFP_KERNEL); | |
7995 | if (!rd) | |
7996 | return NULL; | |
7997 | ||
7998 | if (init_rootdomain(rd, false) != 0) { | |
7999 | kfree(rd); | |
8000 | return NULL; | |
8001 | } | |
8002 | ||
8003 | return rd; | |
8004 | } | |
8005 | ||
8006 | /* | |
8007 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must | |
8008 | * hold the hotplug lock. | |
8009 | */ | |
8010 | static void | |
8011 | cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) | |
8012 | { | |
8013 | struct rq *rq = cpu_rq(cpu); | |
8014 | struct sched_domain *tmp; | |
8015 | ||
8016 | /* Remove the sched domains which do not contribute to scheduling. */ | |
8017 | for (tmp = sd; tmp; ) { | |
8018 | struct sched_domain *parent = tmp->parent; | |
8019 | if (!parent) | |
8020 | break; | |
8021 | ||
8022 | if (sd_parent_degenerate(tmp, parent)) { | |
8023 | tmp->parent = parent->parent; | |
8024 | if (parent->parent) | |
8025 | parent->parent->child = tmp; | |
8026 | } else | |
8027 | tmp = tmp->parent; | |
8028 | } | |
8029 | ||
8030 | if (sd && sd_degenerate(sd)) { | |
8031 | sd = sd->parent; | |
8032 | if (sd) | |
8033 | sd->child = NULL; | |
8034 | } | |
8035 | ||
8036 | sched_domain_debug(sd, cpu); | |
8037 | ||
8038 | rq_attach_root(rq, rd); | |
8039 | rcu_assign_pointer(rq->sd, sd); | |
8040 | } | |
8041 | ||
8042 | /* cpus with isolated domains */ | |
8043 | static cpumask_var_t cpu_isolated_map; | |
8044 | ||
8045 | /* Setup the mask of cpus configured for isolated domains */ | |
8046 | static int __init isolated_cpu_setup(char *str) | |
8047 | { | |
8048 | cpulist_parse(str, cpu_isolated_map); | |
8049 | return 1; | |
8050 | } | |
8051 | ||
8052 | __setup("isolcpus=", isolated_cpu_setup); | |
8053 | ||
8054 | /* | |
8055 | * init_sched_build_groups takes the cpumask we wish to span, and a pointer | |
8056 | * to a function which identifies what group(along with sched group) a CPU | |
8057 | * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids | |
8058 | * (due to the fact that we keep track of groups covered with a struct cpumask). | |
8059 | * | |
8060 | * init_sched_build_groups will build a circular linked list of the groups | |
8061 | * covered by the given span, and will set each group's ->cpumask correctly, | |
8062 | * and ->cpu_power to 0. | |
8063 | */ | |
8064 | static void | |
8065 | init_sched_build_groups(const struct cpumask *span, | |
8066 | const struct cpumask *cpu_map, | |
8067 | int (*group_fn)(int cpu, const struct cpumask *cpu_map, | |
8068 | struct sched_group **sg, | |
8069 | struct cpumask *tmpmask), | |
8070 | struct cpumask *covered, struct cpumask *tmpmask) | |
8071 | { | |
8072 | struct sched_group *first = NULL, *last = NULL; | |
8073 | int i; | |
8074 | ||
8075 | cpumask_clear(covered); | |
8076 | ||
8077 | for_each_cpu(i, span) { | |
8078 | struct sched_group *sg; | |
8079 | int group = group_fn(i, cpu_map, &sg, tmpmask); | |
8080 | int j; | |
8081 | ||
8082 | if (cpumask_test_cpu(i, covered)) | |
8083 | continue; | |
8084 | ||
8085 | cpumask_clear(sched_group_cpus(sg)); | |
8086 | sg->cpu_power = 0; | |
8087 | ||
8088 | for_each_cpu(j, span) { | |
8089 | if (group_fn(j, cpu_map, NULL, tmpmask) != group) | |
8090 | continue; | |
8091 | ||
8092 | cpumask_set_cpu(j, covered); | |
8093 | cpumask_set_cpu(j, sched_group_cpus(sg)); | |
8094 | } | |
8095 | if (!first) | |
8096 | first = sg; | |
8097 | if (last) | |
8098 | last->next = sg; | |
8099 | last = sg; | |
8100 | } | |
8101 | last->next = first; | |
8102 | } | |
8103 | ||
8104 | #define SD_NODES_PER_DOMAIN 16 | |
8105 | ||
8106 | #ifdef CONFIG_NUMA | |
8107 | ||
8108 | /** | |
8109 | * find_next_best_node - find the next node to include in a sched_domain | |
8110 | * @node: node whose sched_domain we're building | |
8111 | * @used_nodes: nodes already in the sched_domain | |
8112 | * | |
8113 | * Find the next node to include in a given scheduling domain. Simply | |
8114 | * finds the closest node not already in the @used_nodes map. | |
8115 | * | |
8116 | * Should use nodemask_t. | |
8117 | */ | |
8118 | static int find_next_best_node(int node, nodemask_t *used_nodes) | |
8119 | { | |
8120 | int i, n, val, min_val, best_node = 0; | |
8121 | ||
8122 | min_val = INT_MAX; | |
8123 | ||
8124 | for (i = 0; i < nr_node_ids; i++) { | |
8125 | /* Start at @node */ | |
8126 | n = (node + i) % nr_node_ids; | |
8127 | ||
8128 | if (!nr_cpus_node(n)) | |
8129 | continue; | |
8130 | ||
8131 | /* Skip already used nodes */ | |
8132 | if (node_isset(n, *used_nodes)) | |
8133 | continue; | |
8134 | ||
8135 | /* Simple min distance search */ | |
8136 | val = node_distance(node, n); | |
8137 | ||
8138 | if (val < min_val) { | |
8139 | min_val = val; | |
8140 | best_node = n; | |
8141 | } | |
8142 | } | |
8143 | ||
8144 | node_set(best_node, *used_nodes); | |
8145 | return best_node; | |
8146 | } | |
8147 | ||
8148 | /** | |
8149 | * sched_domain_node_span - get a cpumask for a node's sched_domain | |
8150 | * @node: node whose cpumask we're constructing | |
8151 | * @span: resulting cpumask | |
8152 | * | |
8153 | * Given a node, construct a good cpumask for its sched_domain to span. It | |
8154 | * should be one that prevents unnecessary balancing, but also spreads tasks | |
8155 | * out optimally. | |
8156 | */ | |
8157 | static void sched_domain_node_span(int node, struct cpumask *span) | |
8158 | { | |
8159 | nodemask_t used_nodes; | |
8160 | int i; | |
8161 | ||
8162 | cpumask_clear(span); | |
8163 | nodes_clear(used_nodes); | |
8164 | ||
8165 | cpumask_or(span, span, cpumask_of_node(node)); | |
8166 | node_set(node, used_nodes); | |
8167 | ||
8168 | for (i = 1; i < SD_NODES_PER_DOMAIN; i++) { | |
8169 | int next_node = find_next_best_node(node, &used_nodes); | |
8170 | ||
8171 | cpumask_or(span, span, cpumask_of_node(next_node)); | |
8172 | } | |
8173 | } | |
8174 | #endif /* CONFIG_NUMA */ | |
8175 | ||
8176 | int sched_smt_power_savings = 0, sched_mc_power_savings = 0; | |
8177 | ||
8178 | /* | |
8179 | * The cpus mask in sched_group and sched_domain hangs off the end. | |
8180 | * | |
8181 | * ( See the the comments in include/linux/sched.h:struct sched_group | |
8182 | * and struct sched_domain. ) | |
8183 | */ | |
8184 | struct static_sched_group { | |
8185 | struct sched_group sg; | |
8186 | DECLARE_BITMAP(cpus, CONFIG_NR_CPUS); | |
8187 | }; | |
8188 | ||
8189 | struct static_sched_domain { | |
8190 | struct sched_domain sd; | |
8191 | DECLARE_BITMAP(span, CONFIG_NR_CPUS); | |
8192 | }; | |
8193 | ||
8194 | struct s_data { | |
8195 | #ifdef CONFIG_NUMA | |
8196 | int sd_allnodes; | |
8197 | cpumask_var_t domainspan; | |
8198 | cpumask_var_t covered; | |
8199 | cpumask_var_t notcovered; | |
8200 | #endif | |
8201 | cpumask_var_t nodemask; | |
8202 | cpumask_var_t this_sibling_map; | |
8203 | cpumask_var_t this_core_map; | |
8204 | cpumask_var_t send_covered; | |
8205 | cpumask_var_t tmpmask; | |
8206 | struct sched_group **sched_group_nodes; | |
8207 | struct root_domain *rd; | |
8208 | }; | |
8209 | ||
8210 | enum s_alloc { | |
8211 | sa_sched_groups = 0, | |
8212 | sa_rootdomain, | |
8213 | sa_tmpmask, | |
8214 | sa_send_covered, | |
8215 | sa_this_core_map, | |
8216 | sa_this_sibling_map, | |
8217 | sa_nodemask, | |
8218 | sa_sched_group_nodes, | |
8219 | #ifdef CONFIG_NUMA | |
8220 | sa_notcovered, | |
8221 | sa_covered, | |
8222 | sa_domainspan, | |
8223 | #endif | |
8224 | sa_none, | |
8225 | }; | |
8226 | ||
8227 | /* | |
8228 | * SMT sched-domains: | |
8229 | */ | |
8230 | #ifdef CONFIG_SCHED_SMT | |
8231 | static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains); | |
8232 | static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus); | |
8233 | ||
8234 | static int | |
8235 | cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map, | |
8236 | struct sched_group **sg, struct cpumask *unused) | |
8237 | { | |
8238 | if (sg) | |
8239 | *sg = &per_cpu(sched_group_cpus, cpu).sg; | |
8240 | return cpu; | |
8241 | } | |
8242 | #endif /* CONFIG_SCHED_SMT */ | |
8243 | ||
8244 | /* | |
8245 | * multi-core sched-domains: | |
8246 | */ | |
8247 | #ifdef CONFIG_SCHED_MC | |
8248 | static DEFINE_PER_CPU(struct static_sched_domain, core_domains); | |
8249 | static DEFINE_PER_CPU(struct static_sched_group, sched_group_core); | |
8250 | #endif /* CONFIG_SCHED_MC */ | |
8251 | ||
8252 | #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT) | |
8253 | static int | |
8254 | cpu_to_core_group(int cpu, const struct cpumask *cpu_map, | |
8255 | struct sched_group **sg, struct cpumask *mask) | |
8256 | { | |
8257 | int group; | |
8258 | ||
8259 | cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map); | |
8260 | group = cpumask_first(mask); | |
8261 | if (sg) | |
8262 | *sg = &per_cpu(sched_group_core, group).sg; | |
8263 | return group; | |
8264 | } | |
8265 | #elif defined(CONFIG_SCHED_MC) | |
8266 | static int | |
8267 | cpu_to_core_group(int cpu, const struct cpumask *cpu_map, | |
8268 | struct sched_group **sg, struct cpumask *unused) | |
8269 | { | |
8270 | if (sg) | |
8271 | *sg = &per_cpu(sched_group_core, cpu).sg; | |
8272 | return cpu; | |
8273 | } | |
8274 | #endif | |
8275 | ||
8276 | static DEFINE_PER_CPU(struct static_sched_domain, phys_domains); | |
8277 | static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys); | |
8278 | ||
8279 | static int | |
8280 | cpu_to_phys_group(int cpu, const struct cpumask *cpu_map, | |
8281 | struct sched_group **sg, struct cpumask *mask) | |
8282 | { | |
8283 | int group; | |
8284 | #ifdef CONFIG_SCHED_MC | |
8285 | cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map); | |
8286 | group = cpumask_first(mask); | |
8287 | #elif defined(CONFIG_SCHED_SMT) | |
8288 | cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map); | |
8289 | group = cpumask_first(mask); | |
8290 | #else | |
8291 | group = cpu; | |
8292 | #endif | |
8293 | if (sg) | |
8294 | *sg = &per_cpu(sched_group_phys, group).sg; | |
8295 | return group; | |
8296 | } | |
8297 | ||
8298 | #ifdef CONFIG_NUMA | |
8299 | /* | |
8300 | * The init_sched_build_groups can't handle what we want to do with node | |
8301 | * groups, so roll our own. Now each node has its own list of groups which | |
8302 | * gets dynamically allocated. | |
8303 | */ | |
8304 | static DEFINE_PER_CPU(struct static_sched_domain, node_domains); | |
8305 | static struct sched_group ***sched_group_nodes_bycpu; | |
8306 | ||
8307 | static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains); | |
8308 | static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes); | |
8309 | ||
8310 | static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map, | |
8311 | struct sched_group **sg, | |
8312 | struct cpumask *nodemask) | |
8313 | { | |
8314 | int group; | |
8315 | ||
8316 | cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map); | |
8317 | group = cpumask_first(nodemask); | |
8318 | ||
8319 | if (sg) | |
8320 | *sg = &per_cpu(sched_group_allnodes, group).sg; | |
8321 | return group; | |
8322 | } | |
8323 | ||
8324 | static void init_numa_sched_groups_power(struct sched_group *group_head) | |
8325 | { | |
8326 | struct sched_group *sg = group_head; | |
8327 | int j; | |
8328 | ||
8329 | if (!sg) | |
8330 | return; | |
8331 | do { | |
8332 | for_each_cpu(j, sched_group_cpus(sg)) { | |
8333 | struct sched_domain *sd; | |
8334 | ||
8335 | sd = &per_cpu(phys_domains, j).sd; | |
8336 | if (j != group_first_cpu(sd->groups)) { | |
8337 | /* | |
8338 | * Only add "power" once for each | |
8339 | * physical package. | |
8340 | */ | |
8341 | continue; | |
8342 | } | |
8343 | ||
8344 | sg->cpu_power += sd->groups->cpu_power; | |
8345 | } | |
8346 | sg = sg->next; | |
8347 | } while (sg != group_head); | |
8348 | } | |
8349 | ||
8350 | static int build_numa_sched_groups(struct s_data *d, | |
8351 | const struct cpumask *cpu_map, int num) | |
8352 | { | |
8353 | struct sched_domain *sd; | |
8354 | struct sched_group *sg, *prev; | |
8355 | int n, j; | |
8356 | ||
8357 | cpumask_clear(d->covered); | |
8358 | cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map); | |
8359 | if (cpumask_empty(d->nodemask)) { | |
8360 | d->sched_group_nodes[num] = NULL; | |
8361 | goto out; | |
8362 | } | |
8363 | ||
8364 | sched_domain_node_span(num, d->domainspan); | |
8365 | cpumask_and(d->domainspan, d->domainspan, cpu_map); | |
8366 | ||
8367 | sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(), | |
8368 | GFP_KERNEL, num); | |
8369 | if (!sg) { | |
8370 | printk(KERN_WARNING "Can not alloc domain group for node %d\n", | |
8371 | num); | |
8372 | return -ENOMEM; | |
8373 | } | |
8374 | d->sched_group_nodes[num] = sg; | |
8375 | ||
8376 | for_each_cpu(j, d->nodemask) { | |
8377 | sd = &per_cpu(node_domains, j).sd; | |
8378 | sd->groups = sg; | |
8379 | } | |
8380 | ||
8381 | sg->cpu_power = 0; | |
8382 | cpumask_copy(sched_group_cpus(sg), d->nodemask); | |
8383 | sg->next = sg; | |
8384 | cpumask_or(d->covered, d->covered, d->nodemask); | |
8385 | ||
8386 | prev = sg; | |
8387 | for (j = 0; j < nr_node_ids; j++) { | |
8388 | n = (num + j) % nr_node_ids; | |
8389 | cpumask_complement(d->notcovered, d->covered); | |
8390 | cpumask_and(d->tmpmask, d->notcovered, cpu_map); | |
8391 | cpumask_and(d->tmpmask, d->tmpmask, d->domainspan); | |
8392 | if (cpumask_empty(d->tmpmask)) | |
8393 | break; | |
8394 | cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n)); | |
8395 | if (cpumask_empty(d->tmpmask)) | |
8396 | continue; | |
8397 | sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(), | |
8398 | GFP_KERNEL, num); | |
8399 | if (!sg) { | |
8400 | printk(KERN_WARNING | |
8401 | "Can not alloc domain group for node %d\n", j); | |
8402 | return -ENOMEM; | |
8403 | } | |
8404 | sg->cpu_power = 0; | |
8405 | cpumask_copy(sched_group_cpus(sg), d->tmpmask); | |
8406 | sg->next = prev->next; | |
8407 | cpumask_or(d->covered, d->covered, d->tmpmask); | |
8408 | prev->next = sg; | |
8409 | prev = sg; | |
8410 | } | |
8411 | out: | |
8412 | return 0; | |
8413 | } | |
8414 | #endif /* CONFIG_NUMA */ | |
8415 | ||
8416 | #ifdef CONFIG_NUMA | |
8417 | /* Free memory allocated for various sched_group structures */ | |
8418 | static void free_sched_groups(const struct cpumask *cpu_map, | |
8419 | struct cpumask *nodemask) | |
8420 | { | |
8421 | int cpu, i; | |
8422 | ||
8423 | for_each_cpu(cpu, cpu_map) { | |
8424 | struct sched_group **sched_group_nodes | |
8425 | = sched_group_nodes_bycpu[cpu]; | |
8426 | ||
8427 | if (!sched_group_nodes) | |
8428 | continue; | |
8429 | ||
8430 | for (i = 0; i < nr_node_ids; i++) { | |
8431 | struct sched_group *oldsg, *sg = sched_group_nodes[i]; | |
8432 | ||
8433 | cpumask_and(nodemask, cpumask_of_node(i), cpu_map); | |
8434 | if (cpumask_empty(nodemask)) | |
8435 | continue; | |
8436 | ||
8437 | if (sg == NULL) | |
8438 | continue; | |
8439 | sg = sg->next; | |
8440 | next_sg: | |
8441 | oldsg = sg; | |
8442 | sg = sg->next; | |
8443 | kfree(oldsg); | |
8444 | if (oldsg != sched_group_nodes[i]) | |
8445 | goto next_sg; | |
8446 | } | |
8447 | kfree(sched_group_nodes); | |
8448 | sched_group_nodes_bycpu[cpu] = NULL; | |
8449 | } | |
8450 | } | |
8451 | #else /* !CONFIG_NUMA */ | |
8452 | static void free_sched_groups(const struct cpumask *cpu_map, | |
8453 | struct cpumask *nodemask) | |
8454 | { | |
8455 | } | |
8456 | #endif /* CONFIG_NUMA */ | |
8457 | ||
8458 | /* | |
8459 | * Initialize sched groups cpu_power. | |
8460 | * | |
8461 | * cpu_power indicates the capacity of sched group, which is used while | |
8462 | * distributing the load between different sched groups in a sched domain. | |
8463 | * Typically cpu_power for all the groups in a sched domain will be same unless | |
8464 | * there are asymmetries in the topology. If there are asymmetries, group | |
8465 | * having more cpu_power will pickup more load compared to the group having | |
8466 | * less cpu_power. | |
8467 | */ | |
8468 | static void init_sched_groups_power(int cpu, struct sched_domain *sd) | |
8469 | { | |
8470 | struct sched_domain *child; | |
8471 | struct sched_group *group; | |
8472 | long power; | |
8473 | int weight; | |
8474 | ||
8475 | WARN_ON(!sd || !sd->groups); | |
8476 | ||
8477 | if (cpu != group_first_cpu(sd->groups)) | |
8478 | return; | |
8479 | ||
8480 | child = sd->child; | |
8481 | ||
8482 | sd->groups->cpu_power = 0; | |
8483 | ||
8484 | if (!child) { | |
8485 | power = SCHED_LOAD_SCALE; | |
8486 | weight = cpumask_weight(sched_domain_span(sd)); | |
8487 | /* | |
8488 | * SMT siblings share the power of a single core. | |
8489 | * Usually multiple threads get a better yield out of | |
8490 | * that one core than a single thread would have, | |
8491 | * reflect that in sd->smt_gain. | |
8492 | */ | |
8493 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { | |
8494 | power *= sd->smt_gain; | |
8495 | power /= weight; | |
8496 | power >>= SCHED_LOAD_SHIFT; | |
8497 | } | |
8498 | sd->groups->cpu_power += power; | |
8499 | return; | |
8500 | } | |
8501 | ||
8502 | /* | |
8503 | * Add cpu_power of each child group to this groups cpu_power. | |
8504 | */ | |
8505 | group = child->groups; | |
8506 | do { | |
8507 | sd->groups->cpu_power += group->cpu_power; | |
8508 | group = group->next; | |
8509 | } while (group != child->groups); | |
8510 | } | |
8511 | ||
8512 | /* | |
8513 | * Initializers for schedule domains | |
8514 | * Non-inlined to reduce accumulated stack pressure in build_sched_domains() | |
8515 | */ | |
8516 | ||
8517 | #ifdef CONFIG_SCHED_DEBUG | |
8518 | # define SD_INIT_NAME(sd, type) sd->name = #type | |
8519 | #else | |
8520 | # define SD_INIT_NAME(sd, type) do { } while (0) | |
8521 | #endif | |
8522 | ||
8523 | #define SD_INIT(sd, type) sd_init_##type(sd) | |
8524 | ||
8525 | #define SD_INIT_FUNC(type) \ | |
8526 | static noinline void sd_init_##type(struct sched_domain *sd) \ | |
8527 | { \ | |
8528 | memset(sd, 0, sizeof(*sd)); \ | |
8529 | *sd = SD_##type##_INIT; \ | |
8530 | sd->level = SD_LV_##type; \ | |
8531 | SD_INIT_NAME(sd, type); \ | |
8532 | } | |
8533 | ||
8534 | SD_INIT_FUNC(CPU) | |
8535 | #ifdef CONFIG_NUMA | |
8536 | SD_INIT_FUNC(ALLNODES) | |
8537 | SD_INIT_FUNC(NODE) | |
8538 | #endif | |
8539 | #ifdef CONFIG_SCHED_SMT | |
8540 | SD_INIT_FUNC(SIBLING) | |
8541 | #endif | |
8542 | #ifdef CONFIG_SCHED_MC | |
8543 | SD_INIT_FUNC(MC) | |
8544 | #endif | |
8545 | ||
8546 | static int default_relax_domain_level = -1; | |
8547 | ||
8548 | static int __init setup_relax_domain_level(char *str) | |
8549 | { | |
8550 | unsigned long val; | |
8551 | ||
8552 | val = simple_strtoul(str, NULL, 0); | |
8553 | if (val < SD_LV_MAX) | |
8554 | default_relax_domain_level = val; | |
8555 | ||
8556 | return 1; | |
8557 | } | |
8558 | __setup("relax_domain_level=", setup_relax_domain_level); | |
8559 | ||
8560 | static void set_domain_attribute(struct sched_domain *sd, | |
8561 | struct sched_domain_attr *attr) | |
8562 | { | |
8563 | int request; | |
8564 | ||
8565 | if (!attr || attr->relax_domain_level < 0) { | |
8566 | if (default_relax_domain_level < 0) | |
8567 | return; | |
8568 | else | |
8569 | request = default_relax_domain_level; | |
8570 | } else | |
8571 | request = attr->relax_domain_level; | |
8572 | if (request < sd->level) { | |
8573 | /* turn off idle balance on this domain */ | |
8574 | sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); | |
8575 | } else { | |
8576 | /* turn on idle balance on this domain */ | |
8577 | sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); | |
8578 | } | |
8579 | } | |
8580 | ||
8581 | static void __free_domain_allocs(struct s_data *d, enum s_alloc what, | |
8582 | const struct cpumask *cpu_map) | |
8583 | { | |
8584 | switch (what) { | |
8585 | case sa_sched_groups: | |
8586 | free_sched_groups(cpu_map, d->tmpmask); /* fall through */ | |
8587 | d->sched_group_nodes = NULL; | |
8588 | case sa_rootdomain: | |
8589 | free_rootdomain(d->rd); /* fall through */ | |
8590 | case sa_tmpmask: | |
8591 | free_cpumask_var(d->tmpmask); /* fall through */ | |
8592 | case sa_send_covered: | |
8593 | free_cpumask_var(d->send_covered); /* fall through */ | |
8594 | case sa_this_core_map: | |
8595 | free_cpumask_var(d->this_core_map); /* fall through */ | |
8596 | case sa_this_sibling_map: | |
8597 | free_cpumask_var(d->this_sibling_map); /* fall through */ | |
8598 | case sa_nodemask: | |
8599 | free_cpumask_var(d->nodemask); /* fall through */ | |
8600 | case sa_sched_group_nodes: | |
8601 | #ifdef CONFIG_NUMA | |
8602 | kfree(d->sched_group_nodes); /* fall through */ | |
8603 | case sa_notcovered: | |
8604 | free_cpumask_var(d->notcovered); /* fall through */ | |
8605 | case sa_covered: | |
8606 | free_cpumask_var(d->covered); /* fall through */ | |
8607 | case sa_domainspan: | |
8608 | free_cpumask_var(d->domainspan); /* fall through */ | |
8609 | #endif | |
8610 | case sa_none: | |
8611 | break; | |
8612 | } | |
8613 | } | |
8614 | ||
8615 | static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, | |
8616 | const struct cpumask *cpu_map) | |
8617 | { | |
8618 | #ifdef CONFIG_NUMA | |
8619 | if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL)) | |
8620 | return sa_none; | |
8621 | if (!alloc_cpumask_var(&d->covered, GFP_KERNEL)) | |
8622 | return sa_domainspan; | |
8623 | if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL)) | |
8624 | return sa_covered; | |
8625 | /* Allocate the per-node list of sched groups */ | |
8626 | d->sched_group_nodes = kcalloc(nr_node_ids, | |
8627 | sizeof(struct sched_group *), GFP_KERNEL); | |
8628 | if (!d->sched_group_nodes) { | |
8629 | printk(KERN_WARNING "Can not alloc sched group node list\n"); | |
8630 | return sa_notcovered; | |
8631 | } | |
8632 | sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes; | |
8633 | #endif | |
8634 | if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL)) | |
8635 | return sa_sched_group_nodes; | |
8636 | if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL)) | |
8637 | return sa_nodemask; | |
8638 | if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL)) | |
8639 | return sa_this_sibling_map; | |
8640 | if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL)) | |
8641 | return sa_this_core_map; | |
8642 | if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL)) | |
8643 | return sa_send_covered; | |
8644 | d->rd = alloc_rootdomain(); | |
8645 | if (!d->rd) { | |
8646 | printk(KERN_WARNING "Cannot alloc root domain\n"); | |
8647 | return sa_tmpmask; | |
8648 | } | |
8649 | return sa_rootdomain; | |
8650 | } | |
8651 | ||
8652 | static struct sched_domain *__build_numa_sched_domains(struct s_data *d, | |
8653 | const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i) | |
8654 | { | |
8655 | struct sched_domain *sd = NULL; | |
8656 | #ifdef CONFIG_NUMA | |
8657 | struct sched_domain *parent; | |
8658 | ||
8659 | d->sd_allnodes = 0; | |
8660 | if (cpumask_weight(cpu_map) > | |
8661 | SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) { | |
8662 | sd = &per_cpu(allnodes_domains, i).sd; | |
8663 | SD_INIT(sd, ALLNODES); | |
8664 | set_domain_attribute(sd, attr); | |
8665 | cpumask_copy(sched_domain_span(sd), cpu_map); | |
8666 | cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask); | |
8667 | d->sd_allnodes = 1; | |
8668 | } | |
8669 | parent = sd; | |
8670 | ||
8671 | sd = &per_cpu(node_domains, i).sd; | |
8672 | SD_INIT(sd, NODE); | |
8673 | set_domain_attribute(sd, attr); | |
8674 | sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd)); | |
8675 | sd->parent = parent; | |
8676 | if (parent) | |
8677 | parent->child = sd; | |
8678 | cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map); | |
8679 | #endif | |
8680 | return sd; | |
8681 | } | |
8682 | ||
8683 | static struct sched_domain *__build_cpu_sched_domain(struct s_data *d, | |
8684 | const struct cpumask *cpu_map, struct sched_domain_attr *attr, | |
8685 | struct sched_domain *parent, int i) | |
8686 | { | |
8687 | struct sched_domain *sd; | |
8688 | sd = &per_cpu(phys_domains, i).sd; | |
8689 | SD_INIT(sd, CPU); | |
8690 | set_domain_attribute(sd, attr); | |
8691 | cpumask_copy(sched_domain_span(sd), d->nodemask); | |
8692 | sd->parent = parent; | |
8693 | if (parent) | |
8694 | parent->child = sd; | |
8695 | cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask); | |
8696 | return sd; | |
8697 | } | |
8698 | ||
8699 | static struct sched_domain *__build_mc_sched_domain(struct s_data *d, | |
8700 | const struct cpumask *cpu_map, struct sched_domain_attr *attr, | |
8701 | struct sched_domain *parent, int i) | |
8702 | { | |
8703 | struct sched_domain *sd = parent; | |
8704 | #ifdef CONFIG_SCHED_MC | |
8705 | sd = &per_cpu(core_domains, i).sd; | |
8706 | SD_INIT(sd, MC); | |
8707 | set_domain_attribute(sd, attr); | |
8708 | cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i)); | |
8709 | sd->parent = parent; | |
8710 | parent->child = sd; | |
8711 | cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask); | |
8712 | #endif | |
8713 | return sd; | |
8714 | } | |
8715 | ||
8716 | static struct sched_domain *__build_smt_sched_domain(struct s_data *d, | |
8717 | const struct cpumask *cpu_map, struct sched_domain_attr *attr, | |
8718 | struct sched_domain *parent, int i) | |
8719 | { | |
8720 | struct sched_domain *sd = parent; | |
8721 | #ifdef CONFIG_SCHED_SMT | |
8722 | sd = &per_cpu(cpu_domains, i).sd; | |
8723 | SD_INIT(sd, SIBLING); | |
8724 | set_domain_attribute(sd, attr); | |
8725 | cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i)); | |
8726 | sd->parent = parent; | |
8727 | parent->child = sd; | |
8728 | cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask); | |
8729 | #endif | |
8730 | return sd; | |
8731 | } | |
8732 | ||
8733 | static void build_sched_groups(struct s_data *d, enum sched_domain_level l, | |
8734 | const struct cpumask *cpu_map, int cpu) | |
8735 | { | |
8736 | switch (l) { | |
8737 | #ifdef CONFIG_SCHED_SMT | |
8738 | case SD_LV_SIBLING: /* set up CPU (sibling) groups */ | |
8739 | cpumask_and(d->this_sibling_map, cpu_map, | |
8740 | topology_thread_cpumask(cpu)); | |
8741 | if (cpu == cpumask_first(d->this_sibling_map)) | |
8742 | init_sched_build_groups(d->this_sibling_map, cpu_map, | |
8743 | &cpu_to_cpu_group, | |
8744 | d->send_covered, d->tmpmask); | |
8745 | break; | |
8746 | #endif | |
8747 | #ifdef CONFIG_SCHED_MC | |
8748 | case SD_LV_MC: /* set up multi-core groups */ | |
8749 | cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu)); | |
8750 | if (cpu == cpumask_first(d->this_core_map)) | |
8751 | init_sched_build_groups(d->this_core_map, cpu_map, | |
8752 | &cpu_to_core_group, | |
8753 | d->send_covered, d->tmpmask); | |
8754 | break; | |
8755 | #endif | |
8756 | case SD_LV_CPU: /* set up physical groups */ | |
8757 | cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map); | |
8758 | if (!cpumask_empty(d->nodemask)) | |
8759 | init_sched_build_groups(d->nodemask, cpu_map, | |
8760 | &cpu_to_phys_group, | |
8761 | d->send_covered, d->tmpmask); | |
8762 | break; | |
8763 | #ifdef CONFIG_NUMA | |
8764 | case SD_LV_ALLNODES: | |
8765 | init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group, | |
8766 | d->send_covered, d->tmpmask); | |
8767 | break; | |
8768 | #endif | |
8769 | default: | |
8770 | break; | |
8771 | } | |
8772 | } | |
8773 | ||
8774 | /* | |
8775 | * Build sched domains for a given set of cpus and attach the sched domains | |
8776 | * to the individual cpus | |
8777 | */ | |
8778 | static int __build_sched_domains(const struct cpumask *cpu_map, | |
8779 | struct sched_domain_attr *attr) | |
8780 | { | |
8781 | enum s_alloc alloc_state = sa_none; | |
8782 | struct s_data d; | |
8783 | struct sched_domain *sd; | |
8784 | int i; | |
8785 | #ifdef CONFIG_NUMA | |
8786 | d.sd_allnodes = 0; | |
8787 | #endif | |
8788 | ||
8789 | alloc_state = __visit_domain_allocation_hell(&d, cpu_map); | |
8790 | if (alloc_state != sa_rootdomain) | |
8791 | goto error; | |
8792 | alloc_state = sa_sched_groups; | |
8793 | ||
8794 | /* | |
8795 | * Set up domains for cpus specified by the cpu_map. | |
8796 | */ | |
8797 | for_each_cpu(i, cpu_map) { | |
8798 | cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)), | |
8799 | cpu_map); | |
8800 | ||
8801 | sd = __build_numa_sched_domains(&d, cpu_map, attr, i); | |
8802 | sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i); | |
8803 | sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i); | |
8804 | sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i); | |
8805 | } | |
8806 | ||
8807 | for_each_cpu(i, cpu_map) { | |
8808 | build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i); | |
8809 | build_sched_groups(&d, SD_LV_MC, cpu_map, i); | |
8810 | } | |
8811 | ||
8812 | /* Set up physical groups */ | |
8813 | for (i = 0; i < nr_node_ids; i++) | |
8814 | build_sched_groups(&d, SD_LV_CPU, cpu_map, i); | |
8815 | ||
8816 | #ifdef CONFIG_NUMA | |
8817 | /* Set up node groups */ | |
8818 | if (d.sd_allnodes) | |
8819 | build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0); | |
8820 | ||
8821 | for (i = 0; i < nr_node_ids; i++) | |
8822 | if (build_numa_sched_groups(&d, cpu_map, i)) | |
8823 | goto error; | |
8824 | #endif | |
8825 | ||
8826 | /* Calculate CPU power for physical packages and nodes */ | |
8827 | #ifdef CONFIG_SCHED_SMT | |
8828 | for_each_cpu(i, cpu_map) { | |
8829 | sd = &per_cpu(cpu_domains, i).sd; | |
8830 | init_sched_groups_power(i, sd); | |
8831 | } | |
8832 | #endif | |
8833 | #ifdef CONFIG_SCHED_MC | |
8834 | for_each_cpu(i, cpu_map) { | |
8835 | sd = &per_cpu(core_domains, i).sd; | |
8836 | init_sched_groups_power(i, sd); | |
8837 | } | |
8838 | #endif | |
8839 | ||
8840 | for_each_cpu(i, cpu_map) { | |
8841 | sd = &per_cpu(phys_domains, i).sd; | |
8842 | init_sched_groups_power(i, sd); | |
8843 | } | |
8844 | ||
8845 | #ifdef CONFIG_NUMA | |
8846 | for (i = 0; i < nr_node_ids; i++) | |
8847 | init_numa_sched_groups_power(d.sched_group_nodes[i]); | |
8848 | ||
8849 | if (d.sd_allnodes) { | |
8850 | struct sched_group *sg; | |
8851 | ||
8852 | cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg, | |
8853 | d.tmpmask); | |
8854 | init_numa_sched_groups_power(sg); | |
8855 | } | |
8856 | #endif | |
8857 | ||
8858 | /* Attach the domains */ | |
8859 | for_each_cpu(i, cpu_map) { | |
8860 | #ifdef CONFIG_SCHED_SMT | |
8861 | sd = &per_cpu(cpu_domains, i).sd; | |
8862 | #elif defined(CONFIG_SCHED_MC) | |
8863 | sd = &per_cpu(core_domains, i).sd; | |
8864 | #else | |
8865 | sd = &per_cpu(phys_domains, i).sd; | |
8866 | #endif | |
8867 | cpu_attach_domain(sd, d.rd, i); | |
8868 | } | |
8869 | ||
8870 | d.sched_group_nodes = NULL; /* don't free this we still need it */ | |
8871 | __free_domain_allocs(&d, sa_tmpmask, cpu_map); | |
8872 | return 0; | |
8873 | ||
8874 | error: | |
8875 | __free_domain_allocs(&d, alloc_state, cpu_map); | |
8876 | return -ENOMEM; | |
8877 | } | |
8878 | ||
8879 | static int build_sched_domains(const struct cpumask *cpu_map) | |
8880 | { | |
8881 | return __build_sched_domains(cpu_map, NULL); | |
8882 | } | |
8883 | ||
8884 | static struct cpumask *doms_cur; /* current sched domains */ | |
8885 | static int ndoms_cur; /* number of sched domains in 'doms_cur' */ | |
8886 | static struct sched_domain_attr *dattr_cur; | |
8887 | /* attribues of custom domains in 'doms_cur' */ | |
8888 | ||
8889 | /* | |
8890 | * Special case: If a kmalloc of a doms_cur partition (array of | |
8891 | * cpumask) fails, then fallback to a single sched domain, | |
8892 | * as determined by the single cpumask fallback_doms. | |
8893 | */ | |
8894 | static cpumask_var_t fallback_doms; | |
8895 | ||
8896 | /* | |
8897 | * arch_update_cpu_topology lets virtualized architectures update the | |
8898 | * cpu core maps. It is supposed to return 1 if the topology changed | |
8899 | * or 0 if it stayed the same. | |
8900 | */ | |
8901 | int __attribute__((weak)) arch_update_cpu_topology(void) | |
8902 | { | |
8903 | return 0; | |
8904 | } | |
8905 | ||
8906 | /* | |
8907 | * Set up scheduler domains and groups. Callers must hold the hotplug lock. | |
8908 | * For now this just excludes isolated cpus, but could be used to | |
8909 | * exclude other special cases in the future. | |
8910 | */ | |
8911 | static int arch_init_sched_domains(const struct cpumask *cpu_map) | |
8912 | { | |
8913 | int err; | |
8914 | ||
8915 | arch_update_cpu_topology(); | |
8916 | ndoms_cur = 1; | |
8917 | doms_cur = kmalloc(cpumask_size(), GFP_KERNEL); | |
8918 | if (!doms_cur) | |
8919 | doms_cur = fallback_doms; | |
8920 | cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map); | |
8921 | dattr_cur = NULL; | |
8922 | err = build_sched_domains(doms_cur); | |
8923 | register_sched_domain_sysctl(); | |
8924 | ||
8925 | return err; | |
8926 | } | |
8927 | ||
8928 | static void arch_destroy_sched_domains(const struct cpumask *cpu_map, | |
8929 | struct cpumask *tmpmask) | |
8930 | { | |
8931 | free_sched_groups(cpu_map, tmpmask); | |
8932 | } | |
8933 | ||
8934 | /* | |
8935 | * Detach sched domains from a group of cpus specified in cpu_map | |
8936 | * These cpus will now be attached to the NULL domain | |
8937 | */ | |
8938 | static void detach_destroy_domains(const struct cpumask *cpu_map) | |
8939 | { | |
8940 | /* Save because hotplug lock held. */ | |
8941 | static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS); | |
8942 | int i; | |
8943 | ||
8944 | for_each_cpu(i, cpu_map) | |
8945 | cpu_attach_domain(NULL, &def_root_domain, i); | |
8946 | synchronize_sched(); | |
8947 | arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask)); | |
8948 | } | |
8949 | ||
8950 | /* handle null as "default" */ | |
8951 | static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, | |
8952 | struct sched_domain_attr *new, int idx_new) | |
8953 | { | |
8954 | struct sched_domain_attr tmp; | |
8955 | ||
8956 | /* fast path */ | |
8957 | if (!new && !cur) | |
8958 | return 1; | |
8959 | ||
8960 | tmp = SD_ATTR_INIT; | |
8961 | return !memcmp(cur ? (cur + idx_cur) : &tmp, | |
8962 | new ? (new + idx_new) : &tmp, | |
8963 | sizeof(struct sched_domain_attr)); | |
8964 | } | |
8965 | ||
8966 | /* | |
8967 | * Partition sched domains as specified by the 'ndoms_new' | |
8968 | * cpumasks in the array doms_new[] of cpumasks. This compares | |
8969 | * doms_new[] to the current sched domain partitioning, doms_cur[]. | |
8970 | * It destroys each deleted domain and builds each new domain. | |
8971 | * | |
8972 | * 'doms_new' is an array of cpumask's of length 'ndoms_new'. | |
8973 | * The masks don't intersect (don't overlap.) We should setup one | |
8974 | * sched domain for each mask. CPUs not in any of the cpumasks will | |
8975 | * not be load balanced. If the same cpumask appears both in the | |
8976 | * current 'doms_cur' domains and in the new 'doms_new', we can leave | |
8977 | * it as it is. | |
8978 | * | |
8979 | * The passed in 'doms_new' should be kmalloc'd. This routine takes | |
8980 | * ownership of it and will kfree it when done with it. If the caller | |
8981 | * failed the kmalloc call, then it can pass in doms_new == NULL && | |
8982 | * ndoms_new == 1, and partition_sched_domains() will fallback to | |
8983 | * the single partition 'fallback_doms', it also forces the domains | |
8984 | * to be rebuilt. | |
8985 | * | |
8986 | * If doms_new == NULL it will be replaced with cpu_online_mask. | |
8987 | * ndoms_new == 0 is a special case for destroying existing domains, | |
8988 | * and it will not create the default domain. | |
8989 | * | |
8990 | * Call with hotplug lock held | |
8991 | */ | |
8992 | /* FIXME: Change to struct cpumask *doms_new[] */ | |
8993 | void partition_sched_domains(int ndoms_new, struct cpumask *doms_new, | |
8994 | struct sched_domain_attr *dattr_new) | |
8995 | { | |
8996 | int i, j, n; | |
8997 | int new_topology; | |
8998 | ||
8999 | mutex_lock(&sched_domains_mutex); | |
9000 | ||
9001 | /* always unregister in case we don't destroy any domains */ | |
9002 | unregister_sched_domain_sysctl(); | |
9003 | ||
9004 | /* Let architecture update cpu core mappings. */ | |
9005 | new_topology = arch_update_cpu_topology(); | |
9006 | ||
9007 | n = doms_new ? ndoms_new : 0; | |
9008 | ||
9009 | /* Destroy deleted domains */ | |
9010 | for (i = 0; i < ndoms_cur; i++) { | |
9011 | for (j = 0; j < n && !new_topology; j++) { | |
9012 | if (cpumask_equal(&doms_cur[i], &doms_new[j]) | |
9013 | && dattrs_equal(dattr_cur, i, dattr_new, j)) | |
9014 | goto match1; | |
9015 | } | |
9016 | /* no match - a current sched domain not in new doms_new[] */ | |
9017 | detach_destroy_domains(doms_cur + i); | |
9018 | match1: | |
9019 | ; | |
9020 | } | |
9021 | ||
9022 | if (doms_new == NULL) { | |
9023 | ndoms_cur = 0; | |
9024 | doms_new = fallback_doms; | |
9025 | cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map); | |
9026 | WARN_ON_ONCE(dattr_new); | |
9027 | } | |
9028 | ||
9029 | /* Build new domains */ | |
9030 | for (i = 0; i < ndoms_new; i++) { | |
9031 | for (j = 0; j < ndoms_cur && !new_topology; j++) { | |
9032 | if (cpumask_equal(&doms_new[i], &doms_cur[j]) | |
9033 | && dattrs_equal(dattr_new, i, dattr_cur, j)) | |
9034 | goto match2; | |
9035 | } | |
9036 | /* no match - add a new doms_new */ | |
9037 | __build_sched_domains(doms_new + i, | |
9038 | dattr_new ? dattr_new + i : NULL); | |
9039 | match2: | |
9040 | ; | |
9041 | } | |
9042 | ||
9043 | /* Remember the new sched domains */ | |
9044 | if (doms_cur != fallback_doms) | |
9045 | kfree(doms_cur); | |
9046 | kfree(dattr_cur); /* kfree(NULL) is safe */ | |
9047 | doms_cur = doms_new; | |
9048 | dattr_cur = dattr_new; | |
9049 | ndoms_cur = ndoms_new; | |
9050 | ||
9051 | register_sched_domain_sysctl(); | |
9052 | ||
9053 | mutex_unlock(&sched_domains_mutex); | |
9054 | } | |
9055 | ||
9056 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
9057 | static void arch_reinit_sched_domains(void) | |
9058 | { | |
9059 | get_online_cpus(); | |
9060 | ||
9061 | /* Destroy domains first to force the rebuild */ | |
9062 | partition_sched_domains(0, NULL, NULL); | |
9063 | ||
9064 | rebuild_sched_domains(); | |
9065 | put_online_cpus(); | |
9066 | } | |
9067 | ||
9068 | static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt) | |
9069 | { | |
9070 | unsigned int level = 0; | |
9071 | ||
9072 | if (sscanf(buf, "%u", &level) != 1) | |
9073 | return -EINVAL; | |
9074 | ||
9075 | /* | |
9076 | * level is always be positive so don't check for | |
9077 | * level < POWERSAVINGS_BALANCE_NONE which is 0 | |
9078 | * What happens on 0 or 1 byte write, | |
9079 | * need to check for count as well? | |
9080 | */ | |
9081 | ||
9082 | if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS) | |
9083 | return -EINVAL; | |
9084 | ||
9085 | if (smt) | |
9086 | sched_smt_power_savings = level; | |
9087 | else | |
9088 | sched_mc_power_savings = level; | |
9089 | ||
9090 | arch_reinit_sched_domains(); | |
9091 | ||
9092 | return count; | |
9093 | } | |
9094 | ||
9095 | #ifdef CONFIG_SCHED_MC | |
9096 | static ssize_t sched_mc_power_savings_show(struct sysdev_class *class, | |
9097 | char *page) | |
9098 | { | |
9099 | return sprintf(page, "%u\n", sched_mc_power_savings); | |
9100 | } | |
9101 | static ssize_t sched_mc_power_savings_store(struct sysdev_class *class, | |
9102 | const char *buf, size_t count) | |
9103 | { | |
9104 | return sched_power_savings_store(buf, count, 0); | |
9105 | } | |
9106 | static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644, | |
9107 | sched_mc_power_savings_show, | |
9108 | sched_mc_power_savings_store); | |
9109 | #endif | |
9110 | ||
9111 | #ifdef CONFIG_SCHED_SMT | |
9112 | static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev, | |
9113 | char *page) | |
9114 | { | |
9115 | return sprintf(page, "%u\n", sched_smt_power_savings); | |
9116 | } | |
9117 | static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev, | |
9118 | const char *buf, size_t count) | |
9119 | { | |
9120 | return sched_power_savings_store(buf, count, 1); | |
9121 | } | |
9122 | static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644, | |
9123 | sched_smt_power_savings_show, | |
9124 | sched_smt_power_savings_store); | |
9125 | #endif | |
9126 | ||
9127 | int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls) | |
9128 | { | |
9129 | int err = 0; | |
9130 | ||
9131 | #ifdef CONFIG_SCHED_SMT | |
9132 | if (smt_capable()) | |
9133 | err = sysfs_create_file(&cls->kset.kobj, | |
9134 | &attr_sched_smt_power_savings.attr); | |
9135 | #endif | |
9136 | #ifdef CONFIG_SCHED_MC | |
9137 | if (!err && mc_capable()) | |
9138 | err = sysfs_create_file(&cls->kset.kobj, | |
9139 | &attr_sched_mc_power_savings.attr); | |
9140 | #endif | |
9141 | return err; | |
9142 | } | |
9143 | #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ | |
9144 | ||
9145 | #ifndef CONFIG_CPUSETS | |
9146 | /* | |
9147 | * Add online and remove offline CPUs from the scheduler domains. | |
9148 | * When cpusets are enabled they take over this function. | |
9149 | */ | |
9150 | static int update_sched_domains(struct notifier_block *nfb, | |
9151 | unsigned long action, void *hcpu) | |
9152 | { | |
9153 | switch (action) { | |
9154 | case CPU_ONLINE: | |
9155 | case CPU_ONLINE_FROZEN: | |
9156 | case CPU_DEAD: | |
9157 | case CPU_DEAD_FROZEN: | |
9158 | partition_sched_domains(1, NULL, NULL); | |
9159 | return NOTIFY_OK; | |
9160 | ||
9161 | default: | |
9162 | return NOTIFY_DONE; | |
9163 | } | |
9164 | } | |
9165 | #endif | |
9166 | ||
9167 | static int update_runtime(struct notifier_block *nfb, | |
9168 | unsigned long action, void *hcpu) | |
9169 | { | |
9170 | int cpu = (int)(long)hcpu; | |
9171 | ||
9172 | switch (action) { | |
9173 | case CPU_DOWN_PREPARE: | |
9174 | case CPU_DOWN_PREPARE_FROZEN: | |
9175 | disable_runtime(cpu_rq(cpu)); | |
9176 | return NOTIFY_OK; | |
9177 | ||
9178 | case CPU_DOWN_FAILED: | |
9179 | case CPU_DOWN_FAILED_FROZEN: | |
9180 | case CPU_ONLINE: | |
9181 | case CPU_ONLINE_FROZEN: | |
9182 | enable_runtime(cpu_rq(cpu)); | |
9183 | return NOTIFY_OK; | |
9184 | ||
9185 | default: | |
9186 | return NOTIFY_DONE; | |
9187 | } | |
9188 | } | |
9189 | ||
9190 | void __init sched_init_smp(void) | |
9191 | { | |
9192 | cpumask_var_t non_isolated_cpus; | |
9193 | ||
9194 | alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); | |
9195 | alloc_cpumask_var(&fallback_doms, GFP_KERNEL); | |
9196 | ||
9197 | #if defined(CONFIG_NUMA) | |
9198 | sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **), | |
9199 | GFP_KERNEL); | |
9200 | BUG_ON(sched_group_nodes_bycpu == NULL); | |
9201 | #endif | |
9202 | get_online_cpus(); | |
9203 | mutex_lock(&sched_domains_mutex); | |
9204 | arch_init_sched_domains(cpu_online_mask); | |
9205 | cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); | |
9206 | if (cpumask_empty(non_isolated_cpus)) | |
9207 | cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); | |
9208 | mutex_unlock(&sched_domains_mutex); | |
9209 | put_online_cpus(); | |
9210 | ||
9211 | #ifndef CONFIG_CPUSETS | |
9212 | /* XXX: Theoretical race here - CPU may be hotplugged now */ | |
9213 | hotcpu_notifier(update_sched_domains, 0); | |
9214 | #endif | |
9215 | ||
9216 | /* RT runtime code needs to handle some hotplug events */ | |
9217 | hotcpu_notifier(update_runtime, 0); | |
9218 | ||
9219 | init_hrtick(); | |
9220 | ||
9221 | /* Move init over to a non-isolated CPU */ | |
9222 | if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) | |
9223 | BUG(); | |
9224 | sched_init_granularity(); | |
9225 | free_cpumask_var(non_isolated_cpus); | |
9226 | ||
9227 | init_sched_rt_class(); | |
9228 | } | |
9229 | #else | |
9230 | void __init sched_init_smp(void) | |
9231 | { | |
9232 | sched_init_granularity(); | |
9233 | } | |
9234 | #endif /* CONFIG_SMP */ | |
9235 | ||
9236 | const_debug unsigned int sysctl_timer_migration = 1; | |
9237 | ||
9238 | int in_sched_functions(unsigned long addr) | |
9239 | { | |
9240 | return in_lock_functions(addr) || | |
9241 | (addr >= (unsigned long)__sched_text_start | |
9242 | && addr < (unsigned long)__sched_text_end); | |
9243 | } | |
9244 | ||
9245 | static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq) | |
9246 | { | |
9247 | cfs_rq->tasks_timeline = RB_ROOT; | |
9248 | INIT_LIST_HEAD(&cfs_rq->tasks); | |
9249 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
9250 | cfs_rq->rq = rq; | |
9251 | #endif | |
9252 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); | |
9253 | } | |
9254 | ||
9255 | static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq) | |
9256 | { | |
9257 | struct rt_prio_array *array; | |
9258 | int i; | |
9259 | ||
9260 | array = &rt_rq->active; | |
9261 | for (i = 0; i < MAX_RT_PRIO; i++) { | |
9262 | INIT_LIST_HEAD(array->queue + i); | |
9263 | __clear_bit(i, array->bitmap); | |
9264 | } | |
9265 | /* delimiter for bitsearch: */ | |
9266 | __set_bit(MAX_RT_PRIO, array->bitmap); | |
9267 | ||
9268 | #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED | |
9269 | rt_rq->highest_prio.curr = MAX_RT_PRIO; | |
9270 | #ifdef CONFIG_SMP | |
9271 | rt_rq->highest_prio.next = MAX_RT_PRIO; | |
9272 | #endif | |
9273 | #endif | |
9274 | #ifdef CONFIG_SMP | |
9275 | rt_rq->rt_nr_migratory = 0; | |
9276 | rt_rq->overloaded = 0; | |
9277 | plist_head_init(&rt_rq->pushable_tasks, &rq->lock); | |
9278 | #endif | |
9279 | ||
9280 | rt_rq->rt_time = 0; | |
9281 | rt_rq->rt_throttled = 0; | |
9282 | rt_rq->rt_runtime = 0; | |
9283 | spin_lock_init(&rt_rq->rt_runtime_lock); | |
9284 | ||
9285 | #ifdef CONFIG_RT_GROUP_SCHED | |
9286 | rt_rq->rt_nr_boosted = 0; | |
9287 | rt_rq->rq = rq; | |
9288 | #endif | |
9289 | } | |
9290 | ||
9291 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
9292 | static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
9293 | struct sched_entity *se, int cpu, int add, | |
9294 | struct sched_entity *parent) | |
9295 | { | |
9296 | struct rq *rq = cpu_rq(cpu); | |
9297 | tg->cfs_rq[cpu] = cfs_rq; | |
9298 | init_cfs_rq(cfs_rq, rq); | |
9299 | cfs_rq->tg = tg; | |
9300 | if (add) | |
9301 | list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list); | |
9302 | ||
9303 | tg->se[cpu] = se; | |
9304 | /* se could be NULL for init_task_group */ | |
9305 | if (!se) | |
9306 | return; | |
9307 | ||
9308 | if (!parent) | |
9309 | se->cfs_rq = &rq->cfs; | |
9310 | else | |
9311 | se->cfs_rq = parent->my_q; | |
9312 | ||
9313 | se->my_q = cfs_rq; | |
9314 | se->load.weight = tg->shares; | |
9315 | se->load.inv_weight = 0; | |
9316 | se->parent = parent; | |
9317 | } | |
9318 | #endif | |
9319 | ||
9320 | #ifdef CONFIG_RT_GROUP_SCHED | |
9321 | static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq, | |
9322 | struct sched_rt_entity *rt_se, int cpu, int add, | |
9323 | struct sched_rt_entity *parent) | |
9324 | { | |
9325 | struct rq *rq = cpu_rq(cpu); | |
9326 | ||
9327 | tg->rt_rq[cpu] = rt_rq; | |
9328 | init_rt_rq(rt_rq, rq); | |
9329 | rt_rq->tg = tg; | |
9330 | rt_rq->rt_se = rt_se; | |
9331 | rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime; | |
9332 | if (add) | |
9333 | list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list); | |
9334 | ||
9335 | tg->rt_se[cpu] = rt_se; | |
9336 | if (!rt_se) | |
9337 | return; | |
9338 | ||
9339 | if (!parent) | |
9340 | rt_se->rt_rq = &rq->rt; | |
9341 | else | |
9342 | rt_se->rt_rq = parent->my_q; | |
9343 | ||
9344 | rt_se->my_q = rt_rq; | |
9345 | rt_se->parent = parent; | |
9346 | INIT_LIST_HEAD(&rt_se->run_list); | |
9347 | } | |
9348 | #endif | |
9349 | ||
9350 | void __init sched_init(void) | |
9351 | { | |
9352 | int i, j; | |
9353 | unsigned long alloc_size = 0, ptr; | |
9354 | ||
9355 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
9356 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); | |
9357 | #endif | |
9358 | #ifdef CONFIG_RT_GROUP_SCHED | |
9359 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); | |
9360 | #endif | |
9361 | #ifdef CONFIG_USER_SCHED | |
9362 | alloc_size *= 2; | |
9363 | #endif | |
9364 | #ifdef CONFIG_CPUMASK_OFFSTACK | |
9365 | alloc_size += num_possible_cpus() * cpumask_size(); | |
9366 | #endif | |
9367 | /* | |
9368 | * As sched_init() is called before page_alloc is setup, | |
9369 | * we use alloc_bootmem(). | |
9370 | */ | |
9371 | if (alloc_size) { | |
9372 | ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); | |
9373 | ||
9374 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
9375 | init_task_group.se = (struct sched_entity **)ptr; | |
9376 | ptr += nr_cpu_ids * sizeof(void **); | |
9377 | ||
9378 | init_task_group.cfs_rq = (struct cfs_rq **)ptr; | |
9379 | ptr += nr_cpu_ids * sizeof(void **); | |
9380 | ||
9381 | #ifdef CONFIG_USER_SCHED | |
9382 | root_task_group.se = (struct sched_entity **)ptr; | |
9383 | ptr += nr_cpu_ids * sizeof(void **); | |
9384 | ||
9385 | root_task_group.cfs_rq = (struct cfs_rq **)ptr; | |
9386 | ptr += nr_cpu_ids * sizeof(void **); | |
9387 | #endif /* CONFIG_USER_SCHED */ | |
9388 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
9389 | #ifdef CONFIG_RT_GROUP_SCHED | |
9390 | init_task_group.rt_se = (struct sched_rt_entity **)ptr; | |
9391 | ptr += nr_cpu_ids * sizeof(void **); | |
9392 | ||
9393 | init_task_group.rt_rq = (struct rt_rq **)ptr; | |
9394 | ptr += nr_cpu_ids * sizeof(void **); | |
9395 | ||
9396 | #ifdef CONFIG_USER_SCHED | |
9397 | root_task_group.rt_se = (struct sched_rt_entity **)ptr; | |
9398 | ptr += nr_cpu_ids * sizeof(void **); | |
9399 | ||
9400 | root_task_group.rt_rq = (struct rt_rq **)ptr; | |
9401 | ptr += nr_cpu_ids * sizeof(void **); | |
9402 | #endif /* CONFIG_USER_SCHED */ | |
9403 | #endif /* CONFIG_RT_GROUP_SCHED */ | |
9404 | #ifdef CONFIG_CPUMASK_OFFSTACK | |
9405 | for_each_possible_cpu(i) { | |
9406 | per_cpu(load_balance_tmpmask, i) = (void *)ptr; | |
9407 | ptr += cpumask_size(); | |
9408 | } | |
9409 | #endif /* CONFIG_CPUMASK_OFFSTACK */ | |
9410 | } | |
9411 | ||
9412 | #ifdef CONFIG_SMP | |
9413 | init_defrootdomain(); | |
9414 | #endif | |
9415 | ||
9416 | init_rt_bandwidth(&def_rt_bandwidth, | |
9417 | global_rt_period(), global_rt_runtime()); | |
9418 | ||
9419 | #ifdef CONFIG_RT_GROUP_SCHED | |
9420 | init_rt_bandwidth(&init_task_group.rt_bandwidth, | |
9421 | global_rt_period(), global_rt_runtime()); | |
9422 | #ifdef CONFIG_USER_SCHED | |
9423 | init_rt_bandwidth(&root_task_group.rt_bandwidth, | |
9424 | global_rt_period(), RUNTIME_INF); | |
9425 | #endif /* CONFIG_USER_SCHED */ | |
9426 | #endif /* CONFIG_RT_GROUP_SCHED */ | |
9427 | ||
9428 | #ifdef CONFIG_GROUP_SCHED | |
9429 | list_add(&init_task_group.list, &task_groups); | |
9430 | INIT_LIST_HEAD(&init_task_group.children); | |
9431 | ||
9432 | #ifdef CONFIG_USER_SCHED | |
9433 | INIT_LIST_HEAD(&root_task_group.children); | |
9434 | init_task_group.parent = &root_task_group; | |
9435 | list_add(&init_task_group.siblings, &root_task_group.children); | |
9436 | #endif /* CONFIG_USER_SCHED */ | |
9437 | #endif /* CONFIG_GROUP_SCHED */ | |
9438 | ||
9439 | #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP | |
9440 | update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long), | |
9441 | __alignof__(unsigned long)); | |
9442 | #endif | |
9443 | for_each_possible_cpu(i) { | |
9444 | struct rq *rq; | |
9445 | ||
9446 | rq = cpu_rq(i); | |
9447 | spin_lock_init(&rq->lock); | |
9448 | rq->nr_running = 0; | |
9449 | rq->calc_load_active = 0; | |
9450 | rq->calc_load_update = jiffies + LOAD_FREQ; | |
9451 | init_cfs_rq(&rq->cfs, rq); | |
9452 | init_rt_rq(&rq->rt, rq); | |
9453 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
9454 | init_task_group.shares = init_task_group_load; | |
9455 | INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); | |
9456 | #ifdef CONFIG_CGROUP_SCHED | |
9457 | /* | |
9458 | * How much cpu bandwidth does init_task_group get? | |
9459 | * | |
9460 | * In case of task-groups formed thr' the cgroup filesystem, it | |
9461 | * gets 100% of the cpu resources in the system. This overall | |
9462 | * system cpu resource is divided among the tasks of | |
9463 | * init_task_group and its child task-groups in a fair manner, | |
9464 | * based on each entity's (task or task-group's) weight | |
9465 | * (se->load.weight). | |
9466 | * | |
9467 | * In other words, if init_task_group has 10 tasks of weight | |
9468 | * 1024) and two child groups A0 and A1 (of weight 1024 each), | |
9469 | * then A0's share of the cpu resource is: | |
9470 | * | |
9471 | * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% | |
9472 | * | |
9473 | * We achieve this by letting init_task_group's tasks sit | |
9474 | * directly in rq->cfs (i.e init_task_group->se[] = NULL). | |
9475 | */ | |
9476 | init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL); | |
9477 | #elif defined CONFIG_USER_SCHED | |
9478 | root_task_group.shares = NICE_0_LOAD; | |
9479 | init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL); | |
9480 | /* | |
9481 | * In case of task-groups formed thr' the user id of tasks, | |
9482 | * init_task_group represents tasks belonging to root user. | |
9483 | * Hence it forms a sibling of all subsequent groups formed. | |
9484 | * In this case, init_task_group gets only a fraction of overall | |
9485 | * system cpu resource, based on the weight assigned to root | |
9486 | * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished | |
9487 | * by letting tasks of init_task_group sit in a separate cfs_rq | |
9488 | * (init_tg_cfs_rq) and having one entity represent this group of | |
9489 | * tasks in rq->cfs (i.e init_task_group->se[] != NULL). | |
9490 | */ | |
9491 | init_tg_cfs_entry(&init_task_group, | |
9492 | &per_cpu(init_tg_cfs_rq, i), | |
9493 | &per_cpu(init_sched_entity, i), i, 1, | |
9494 | root_task_group.se[i]); | |
9495 | ||
9496 | #endif | |
9497 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
9498 | ||
9499 | rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; | |
9500 | #ifdef CONFIG_RT_GROUP_SCHED | |
9501 | INIT_LIST_HEAD(&rq->leaf_rt_rq_list); | |
9502 | #ifdef CONFIG_CGROUP_SCHED | |
9503 | init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL); | |
9504 | #elif defined CONFIG_USER_SCHED | |
9505 | init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL); | |
9506 | init_tg_rt_entry(&init_task_group, | |
9507 | &per_cpu(init_rt_rq, i), | |
9508 | &per_cpu(init_sched_rt_entity, i), i, 1, | |
9509 | root_task_group.rt_se[i]); | |
9510 | #endif | |
9511 | #endif | |
9512 | ||
9513 | for (j = 0; j < CPU_LOAD_IDX_MAX; j++) | |
9514 | rq->cpu_load[j] = 0; | |
9515 | #ifdef CONFIG_SMP | |
9516 | rq->sd = NULL; | |
9517 | rq->rd = NULL; | |
9518 | rq->post_schedule = 0; | |
9519 | rq->active_balance = 0; | |
9520 | rq->next_balance = jiffies; | |
9521 | rq->push_cpu = 0; | |
9522 | rq->cpu = i; | |
9523 | rq->online = 0; | |
9524 | rq->migration_thread = NULL; | |
9525 | INIT_LIST_HEAD(&rq->migration_queue); | |
9526 | rq_attach_root(rq, &def_root_domain); | |
9527 | #endif | |
9528 | init_rq_hrtick(rq); | |
9529 | atomic_set(&rq->nr_iowait, 0); | |
9530 | } | |
9531 | ||
9532 | set_load_weight(&init_task); | |
9533 | ||
9534 | #ifdef CONFIG_PREEMPT_NOTIFIERS | |
9535 | INIT_HLIST_HEAD(&init_task.preempt_notifiers); | |
9536 | #endif | |
9537 | ||
9538 | #ifdef CONFIG_SMP | |
9539 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | |
9540 | #endif | |
9541 | ||
9542 | #ifdef CONFIG_RT_MUTEXES | |
9543 | plist_head_init(&init_task.pi_waiters, &init_task.pi_lock); | |
9544 | #endif | |
9545 | ||
9546 | /* | |
9547 | * The boot idle thread does lazy MMU switching as well: | |
9548 | */ | |
9549 | atomic_inc(&init_mm.mm_count); | |
9550 | enter_lazy_tlb(&init_mm, current); | |
9551 | ||
9552 | /* | |
9553 | * Make us the idle thread. Technically, schedule() should not be | |
9554 | * called from this thread, however somewhere below it might be, | |
9555 | * but because we are the idle thread, we just pick up running again | |
9556 | * when this runqueue becomes "idle". | |
9557 | */ | |
9558 | init_idle(current, smp_processor_id()); | |
9559 | ||
9560 | calc_load_update = jiffies + LOAD_FREQ; | |
9561 | ||
9562 | /* | |
9563 | * During early bootup we pretend to be a normal task: | |
9564 | */ | |
9565 | current->sched_class = &fair_sched_class; | |
9566 | ||
9567 | /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */ | |
9568 | zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT); | |
9569 | #ifdef CONFIG_SMP | |
9570 | #ifdef CONFIG_NO_HZ | |
9571 | zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT); | |
9572 | alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT); | |
9573 | #endif | |
9574 | zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); | |
9575 | #endif /* SMP */ | |
9576 | ||
9577 | perf_event_init(); | |
9578 | ||
9579 | scheduler_running = 1; | |
9580 | } | |
9581 | ||
9582 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP | |
9583 | static inline int preempt_count_equals(int preempt_offset) | |
9584 | { | |
9585 | int nested = preempt_count() & ~PREEMPT_ACTIVE; | |
9586 | ||
9587 | return (nested == PREEMPT_INATOMIC_BASE + preempt_offset); | |
9588 | } | |
9589 | ||
9590 | void __might_sleep(char *file, int line, int preempt_offset) | |
9591 | { | |
9592 | #ifdef in_atomic | |
9593 | static unsigned long prev_jiffy; /* ratelimiting */ | |
9594 | ||
9595 | if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || | |
9596 | system_state != SYSTEM_RUNNING || oops_in_progress) | |
9597 | return; | |
9598 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) | |
9599 | return; | |
9600 | prev_jiffy = jiffies; | |
9601 | ||
9602 | printk(KERN_ERR | |
9603 | "BUG: sleeping function called from invalid context at %s:%d\n", | |
9604 | file, line); | |
9605 | printk(KERN_ERR | |
9606 | "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", | |
9607 | in_atomic(), irqs_disabled(), | |
9608 | current->pid, current->comm); | |
9609 | ||
9610 | debug_show_held_locks(current); | |
9611 | if (irqs_disabled()) | |
9612 | print_irqtrace_events(current); | |
9613 | dump_stack(); | |
9614 | #endif | |
9615 | } | |
9616 | EXPORT_SYMBOL(__might_sleep); | |
9617 | #endif | |
9618 | ||
9619 | #ifdef CONFIG_MAGIC_SYSRQ | |
9620 | static void normalize_task(struct rq *rq, struct task_struct *p) | |
9621 | { | |
9622 | int on_rq; | |
9623 | ||
9624 | update_rq_clock(rq); | |
9625 | on_rq = p->se.on_rq; | |
9626 | if (on_rq) | |
9627 | deactivate_task(rq, p, 0); | |
9628 | __setscheduler(rq, p, SCHED_NORMAL, 0); | |
9629 | if (on_rq) { | |
9630 | activate_task(rq, p, 0); | |
9631 | resched_task(rq->curr); | |
9632 | } | |
9633 | } | |
9634 | ||
9635 | void normalize_rt_tasks(void) | |
9636 | { | |
9637 | struct task_struct *g, *p; | |
9638 | unsigned long flags; | |
9639 | struct rq *rq; | |
9640 | ||
9641 | read_lock_irqsave(&tasklist_lock, flags); | |
9642 | do_each_thread(g, p) { | |
9643 | /* | |
9644 | * Only normalize user tasks: | |
9645 | */ | |
9646 | if (!p->mm) | |
9647 | continue; | |
9648 | ||
9649 | p->se.exec_start = 0; | |
9650 | #ifdef CONFIG_SCHEDSTATS | |
9651 | p->se.wait_start = 0; | |
9652 | p->se.sleep_start = 0; | |
9653 | p->se.block_start = 0; | |
9654 | #endif | |
9655 | ||
9656 | if (!rt_task(p)) { | |
9657 | /* | |
9658 | * Renice negative nice level userspace | |
9659 | * tasks back to 0: | |
9660 | */ | |
9661 | if (TASK_NICE(p) < 0 && p->mm) | |
9662 | set_user_nice(p, 0); | |
9663 | continue; | |
9664 | } | |
9665 | ||
9666 | spin_lock(&p->pi_lock); | |
9667 | rq = __task_rq_lock(p); | |
9668 | ||
9669 | normalize_task(rq, p); | |
9670 | ||
9671 | __task_rq_unlock(rq); | |
9672 | spin_unlock(&p->pi_lock); | |
9673 | } while_each_thread(g, p); | |
9674 | ||
9675 | read_unlock_irqrestore(&tasklist_lock, flags); | |
9676 | } | |
9677 | ||
9678 | #endif /* CONFIG_MAGIC_SYSRQ */ | |
9679 | ||
9680 | #ifdef CONFIG_IA64 | |
9681 | /* | |
9682 | * These functions are only useful for the IA64 MCA handling. | |
9683 | * | |
9684 | * They can only be called when the whole system has been | |
9685 | * stopped - every CPU needs to be quiescent, and no scheduling | |
9686 | * activity can take place. Using them for anything else would | |
9687 | * be a serious bug, and as a result, they aren't even visible | |
9688 | * under any other configuration. | |
9689 | */ | |
9690 | ||
9691 | /** | |
9692 | * curr_task - return the current task for a given cpu. | |
9693 | * @cpu: the processor in question. | |
9694 | * | |
9695 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
9696 | */ | |
9697 | struct task_struct *curr_task(int cpu) | |
9698 | { | |
9699 | return cpu_curr(cpu); | |
9700 | } | |
9701 | ||
9702 | /** | |
9703 | * set_curr_task - set the current task for a given cpu. | |
9704 | * @cpu: the processor in question. | |
9705 | * @p: the task pointer to set. | |
9706 | * | |
9707 | * Description: This function must only be used when non-maskable interrupts | |
9708 | * are serviced on a separate stack. It allows the architecture to switch the | |
9709 | * notion of the current task on a cpu in a non-blocking manner. This function | |
9710 | * must be called with all CPU's synchronized, and interrupts disabled, the | |
9711 | * and caller must save the original value of the current task (see | |
9712 | * curr_task() above) and restore that value before reenabling interrupts and | |
9713 | * re-starting the system. | |
9714 | * | |
9715 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
9716 | */ | |
9717 | void set_curr_task(int cpu, struct task_struct *p) | |
9718 | { | |
9719 | cpu_curr(cpu) = p; | |
9720 | } | |
9721 | ||
9722 | #endif | |
9723 | ||
9724 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
9725 | static void free_fair_sched_group(struct task_group *tg) | |
9726 | { | |
9727 | int i; | |
9728 | ||
9729 | for_each_possible_cpu(i) { | |
9730 | if (tg->cfs_rq) | |
9731 | kfree(tg->cfs_rq[i]); | |
9732 | if (tg->se) | |
9733 | kfree(tg->se[i]); | |
9734 | } | |
9735 | ||
9736 | kfree(tg->cfs_rq); | |
9737 | kfree(tg->se); | |
9738 | } | |
9739 | ||
9740 | static | |
9741 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
9742 | { | |
9743 | struct cfs_rq *cfs_rq; | |
9744 | struct sched_entity *se; | |
9745 | struct rq *rq; | |
9746 | int i; | |
9747 | ||
9748 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); | |
9749 | if (!tg->cfs_rq) | |
9750 | goto err; | |
9751 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); | |
9752 | if (!tg->se) | |
9753 | goto err; | |
9754 | ||
9755 | tg->shares = NICE_0_LOAD; | |
9756 | ||
9757 | for_each_possible_cpu(i) { | |
9758 | rq = cpu_rq(i); | |
9759 | ||
9760 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
9761 | GFP_KERNEL, cpu_to_node(i)); | |
9762 | if (!cfs_rq) | |
9763 | goto err; | |
9764 | ||
9765 | se = kzalloc_node(sizeof(struct sched_entity), | |
9766 | GFP_KERNEL, cpu_to_node(i)); | |
9767 | if (!se) | |
9768 | goto err; | |
9769 | ||
9770 | init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]); | |
9771 | } | |
9772 | ||
9773 | return 1; | |
9774 | ||
9775 | err: | |
9776 | return 0; | |
9777 | } | |
9778 | ||
9779 | static inline void register_fair_sched_group(struct task_group *tg, int cpu) | |
9780 | { | |
9781 | list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list, | |
9782 | &cpu_rq(cpu)->leaf_cfs_rq_list); | |
9783 | } | |
9784 | ||
9785 | static inline void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
9786 | { | |
9787 | list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list); | |
9788 | } | |
9789 | #else /* !CONFG_FAIR_GROUP_SCHED */ | |
9790 | static inline void free_fair_sched_group(struct task_group *tg) | |
9791 | { | |
9792 | } | |
9793 | ||
9794 | static inline | |
9795 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
9796 | { | |
9797 | return 1; | |
9798 | } | |
9799 | ||
9800 | static inline void register_fair_sched_group(struct task_group *tg, int cpu) | |
9801 | { | |
9802 | } | |
9803 | ||
9804 | static inline void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
9805 | { | |
9806 | } | |
9807 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
9808 | ||
9809 | #ifdef CONFIG_RT_GROUP_SCHED | |
9810 | static void free_rt_sched_group(struct task_group *tg) | |
9811 | { | |
9812 | int i; | |
9813 | ||
9814 | destroy_rt_bandwidth(&tg->rt_bandwidth); | |
9815 | ||
9816 | for_each_possible_cpu(i) { | |
9817 | if (tg->rt_rq) | |
9818 | kfree(tg->rt_rq[i]); | |
9819 | if (tg->rt_se) | |
9820 | kfree(tg->rt_se[i]); | |
9821 | } | |
9822 | ||
9823 | kfree(tg->rt_rq); | |
9824 | kfree(tg->rt_se); | |
9825 | } | |
9826 | ||
9827 | static | |
9828 | int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) | |
9829 | { | |
9830 | struct rt_rq *rt_rq; | |
9831 | struct sched_rt_entity *rt_se; | |
9832 | struct rq *rq; | |
9833 | int i; | |
9834 | ||
9835 | tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL); | |
9836 | if (!tg->rt_rq) | |
9837 | goto err; | |
9838 | tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL); | |
9839 | if (!tg->rt_se) | |
9840 | goto err; | |
9841 | ||
9842 | init_rt_bandwidth(&tg->rt_bandwidth, | |
9843 | ktime_to_ns(def_rt_bandwidth.rt_period), 0); | |
9844 | ||
9845 | for_each_possible_cpu(i) { | |
9846 | rq = cpu_rq(i); | |
9847 | ||
9848 | rt_rq = kzalloc_node(sizeof(struct rt_rq), | |
9849 | GFP_KERNEL, cpu_to_node(i)); | |
9850 | if (!rt_rq) | |
9851 | goto err; | |
9852 | ||
9853 | rt_se = kzalloc_node(sizeof(struct sched_rt_entity), | |
9854 | GFP_KERNEL, cpu_to_node(i)); | |
9855 | if (!rt_se) | |
9856 | goto err; | |
9857 | ||
9858 | init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]); | |
9859 | } | |
9860 | ||
9861 | return 1; | |
9862 | ||
9863 | err: | |
9864 | return 0; | |
9865 | } | |
9866 | ||
9867 | static inline void register_rt_sched_group(struct task_group *tg, int cpu) | |
9868 | { | |
9869 | list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list, | |
9870 | &cpu_rq(cpu)->leaf_rt_rq_list); | |
9871 | } | |
9872 | ||
9873 | static inline void unregister_rt_sched_group(struct task_group *tg, int cpu) | |
9874 | { | |
9875 | list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list); | |
9876 | } | |
9877 | #else /* !CONFIG_RT_GROUP_SCHED */ | |
9878 | static inline void free_rt_sched_group(struct task_group *tg) | |
9879 | { | |
9880 | } | |
9881 | ||
9882 | static inline | |
9883 | int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) | |
9884 | { | |
9885 | return 1; | |
9886 | } | |
9887 | ||
9888 | static inline void register_rt_sched_group(struct task_group *tg, int cpu) | |
9889 | { | |
9890 | } | |
9891 | ||
9892 | static inline void unregister_rt_sched_group(struct task_group *tg, int cpu) | |
9893 | { | |
9894 | } | |
9895 | #endif /* CONFIG_RT_GROUP_SCHED */ | |
9896 | ||
9897 | #ifdef CONFIG_GROUP_SCHED | |
9898 | static void free_sched_group(struct task_group *tg) | |
9899 | { | |
9900 | free_fair_sched_group(tg); | |
9901 | free_rt_sched_group(tg); | |
9902 | kfree(tg); | |
9903 | } | |
9904 | ||
9905 | /* allocate runqueue etc for a new task group */ | |
9906 | struct task_group *sched_create_group(struct task_group *parent) | |
9907 | { | |
9908 | struct task_group *tg; | |
9909 | unsigned long flags; | |
9910 | int i; | |
9911 | ||
9912 | tg = kzalloc(sizeof(*tg), GFP_KERNEL); | |
9913 | if (!tg) | |
9914 | return ERR_PTR(-ENOMEM); | |
9915 | ||
9916 | if (!alloc_fair_sched_group(tg, parent)) | |
9917 | goto err; | |
9918 | ||
9919 | if (!alloc_rt_sched_group(tg, parent)) | |
9920 | goto err; | |
9921 | ||
9922 | spin_lock_irqsave(&task_group_lock, flags); | |
9923 | for_each_possible_cpu(i) { | |
9924 | register_fair_sched_group(tg, i); | |
9925 | register_rt_sched_group(tg, i); | |
9926 | } | |
9927 | list_add_rcu(&tg->list, &task_groups); | |
9928 | ||
9929 | WARN_ON(!parent); /* root should already exist */ | |
9930 | ||
9931 | tg->parent = parent; | |
9932 | INIT_LIST_HEAD(&tg->children); | |
9933 | list_add_rcu(&tg->siblings, &parent->children); | |
9934 | spin_unlock_irqrestore(&task_group_lock, flags); | |
9935 | ||
9936 | return tg; | |
9937 | ||
9938 | err: | |
9939 | free_sched_group(tg); | |
9940 | return ERR_PTR(-ENOMEM); | |
9941 | } | |
9942 | ||
9943 | /* rcu callback to free various structures associated with a task group */ | |
9944 | static void free_sched_group_rcu(struct rcu_head *rhp) | |
9945 | { | |
9946 | /* now it should be safe to free those cfs_rqs */ | |
9947 | free_sched_group(container_of(rhp, struct task_group, rcu)); | |
9948 | } | |
9949 | ||
9950 | /* Destroy runqueue etc associated with a task group */ | |
9951 | void sched_destroy_group(struct task_group *tg) | |
9952 | { | |
9953 | unsigned long flags; | |
9954 | int i; | |
9955 | ||
9956 | spin_lock_irqsave(&task_group_lock, flags); | |
9957 | for_each_possible_cpu(i) { | |
9958 | unregister_fair_sched_group(tg, i); | |
9959 | unregister_rt_sched_group(tg, i); | |
9960 | } | |
9961 | list_del_rcu(&tg->list); | |
9962 | list_del_rcu(&tg->siblings); | |
9963 | spin_unlock_irqrestore(&task_group_lock, flags); | |
9964 | ||
9965 | /* wait for possible concurrent references to cfs_rqs complete */ | |
9966 | call_rcu(&tg->rcu, free_sched_group_rcu); | |
9967 | } | |
9968 | ||
9969 | /* change task's runqueue when it moves between groups. | |
9970 | * The caller of this function should have put the task in its new group | |
9971 | * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to | |
9972 | * reflect its new group. | |
9973 | */ | |
9974 | void sched_move_task(struct task_struct *tsk) | |
9975 | { | |
9976 | int on_rq, running; | |
9977 | unsigned long flags; | |
9978 | struct rq *rq; | |
9979 | ||
9980 | rq = task_rq_lock(tsk, &flags); | |
9981 | ||
9982 | update_rq_clock(rq); | |
9983 | ||
9984 | running = task_current(rq, tsk); | |
9985 | on_rq = tsk->se.on_rq; | |
9986 | ||
9987 | if (on_rq) | |
9988 | dequeue_task(rq, tsk, 0); | |
9989 | if (unlikely(running)) | |
9990 | tsk->sched_class->put_prev_task(rq, tsk); | |
9991 | ||
9992 | set_task_rq(tsk, task_cpu(tsk)); | |
9993 | ||
9994 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
9995 | if (tsk->sched_class->moved_group) | |
9996 | tsk->sched_class->moved_group(tsk); | |
9997 | #endif | |
9998 | ||
9999 | if (unlikely(running)) | |
10000 | tsk->sched_class->set_curr_task(rq); | |
10001 | if (on_rq) | |
10002 | enqueue_task(rq, tsk, 0); | |
10003 | ||
10004 | task_rq_unlock(rq, &flags); | |
10005 | } | |
10006 | #endif /* CONFIG_GROUP_SCHED */ | |
10007 | ||
10008 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
10009 | static void __set_se_shares(struct sched_entity *se, unsigned long shares) | |
10010 | { | |
10011 | struct cfs_rq *cfs_rq = se->cfs_rq; | |
10012 | int on_rq; | |
10013 | ||
10014 | on_rq = se->on_rq; | |
10015 | if (on_rq) | |
10016 | dequeue_entity(cfs_rq, se, 0); | |
10017 | ||
10018 | se->load.weight = shares; | |
10019 | se->load.inv_weight = 0; | |
10020 | ||
10021 | if (on_rq) | |
10022 | enqueue_entity(cfs_rq, se, 0); | |
10023 | } | |
10024 | ||
10025 | static void set_se_shares(struct sched_entity *se, unsigned long shares) | |
10026 | { | |
10027 | struct cfs_rq *cfs_rq = se->cfs_rq; | |
10028 | struct rq *rq = cfs_rq->rq; | |
10029 | unsigned long flags; | |
10030 | ||
10031 | spin_lock_irqsave(&rq->lock, flags); | |
10032 | __set_se_shares(se, shares); | |
10033 | spin_unlock_irqrestore(&rq->lock, flags); | |
10034 | } | |
10035 | ||
10036 | static DEFINE_MUTEX(shares_mutex); | |
10037 | ||
10038 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
10039 | { | |
10040 | int i; | |
10041 | unsigned long flags; | |
10042 | ||
10043 | /* | |
10044 | * We can't change the weight of the root cgroup. | |
10045 | */ | |
10046 | if (!tg->se[0]) | |
10047 | return -EINVAL; | |
10048 | ||
10049 | if (shares < MIN_SHARES) | |
10050 | shares = MIN_SHARES; | |
10051 | else if (shares > MAX_SHARES) | |
10052 | shares = MAX_SHARES; | |
10053 | ||
10054 | mutex_lock(&shares_mutex); | |
10055 | if (tg->shares == shares) | |
10056 | goto done; | |
10057 | ||
10058 | spin_lock_irqsave(&task_group_lock, flags); | |
10059 | for_each_possible_cpu(i) | |
10060 | unregister_fair_sched_group(tg, i); | |
10061 | list_del_rcu(&tg->siblings); | |
10062 | spin_unlock_irqrestore(&task_group_lock, flags); | |
10063 | ||
10064 | /* wait for any ongoing reference to this group to finish */ | |
10065 | synchronize_sched(); | |
10066 | ||
10067 | /* | |
10068 | * Now we are free to modify the group's share on each cpu | |
10069 | * w/o tripping rebalance_share or load_balance_fair. | |
10070 | */ | |
10071 | tg->shares = shares; | |
10072 | for_each_possible_cpu(i) { | |
10073 | /* | |
10074 | * force a rebalance | |
10075 | */ | |
10076 | cfs_rq_set_shares(tg->cfs_rq[i], 0); | |
10077 | set_se_shares(tg->se[i], shares); | |
10078 | } | |
10079 | ||
10080 | /* | |
10081 | * Enable load balance activity on this group, by inserting it back on | |
10082 | * each cpu's rq->leaf_cfs_rq_list. | |
10083 | */ | |
10084 | spin_lock_irqsave(&task_group_lock, flags); | |
10085 | for_each_possible_cpu(i) | |
10086 | register_fair_sched_group(tg, i); | |
10087 | list_add_rcu(&tg->siblings, &tg->parent->children); | |
10088 | spin_unlock_irqrestore(&task_group_lock, flags); | |
10089 | done: | |
10090 | mutex_unlock(&shares_mutex); | |
10091 | return 0; | |
10092 | } | |
10093 | ||
10094 | unsigned long sched_group_shares(struct task_group *tg) | |
10095 | { | |
10096 | return tg->shares; | |
10097 | } | |
10098 | #endif | |
10099 | ||
10100 | #ifdef CONFIG_RT_GROUP_SCHED | |
10101 | /* | |
10102 | * Ensure that the real time constraints are schedulable. | |
10103 | */ | |
10104 | static DEFINE_MUTEX(rt_constraints_mutex); | |
10105 | ||
10106 | static unsigned long to_ratio(u64 period, u64 runtime) | |
10107 | { | |
10108 | if (runtime == RUNTIME_INF) | |
10109 | return 1ULL << 20; | |
10110 | ||
10111 | return div64_u64(runtime << 20, period); | |
10112 | } | |
10113 | ||
10114 | /* Must be called with tasklist_lock held */ | |
10115 | static inline int tg_has_rt_tasks(struct task_group *tg) | |
10116 | { | |
10117 | struct task_struct *g, *p; | |
10118 | ||
10119 | do_each_thread(g, p) { | |
10120 | if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg) | |
10121 | return 1; | |
10122 | } while_each_thread(g, p); | |
10123 | ||
10124 | return 0; | |
10125 | } | |
10126 | ||
10127 | struct rt_schedulable_data { | |
10128 | struct task_group *tg; | |
10129 | u64 rt_period; | |
10130 | u64 rt_runtime; | |
10131 | }; | |
10132 | ||
10133 | static int tg_schedulable(struct task_group *tg, void *data) | |
10134 | { | |
10135 | struct rt_schedulable_data *d = data; | |
10136 | struct task_group *child; | |
10137 | unsigned long total, sum = 0; | |
10138 | u64 period, runtime; | |
10139 | ||
10140 | period = ktime_to_ns(tg->rt_bandwidth.rt_period); | |
10141 | runtime = tg->rt_bandwidth.rt_runtime; | |
10142 | ||
10143 | if (tg == d->tg) { | |
10144 | period = d->rt_period; | |
10145 | runtime = d->rt_runtime; | |
10146 | } | |
10147 | ||
10148 | #ifdef CONFIG_USER_SCHED | |
10149 | if (tg == &root_task_group) { | |
10150 | period = global_rt_period(); | |
10151 | runtime = global_rt_runtime(); | |
10152 | } | |
10153 | #endif | |
10154 | ||
10155 | /* | |
10156 | * Cannot have more runtime than the period. | |
10157 | */ | |
10158 | if (runtime > period && runtime != RUNTIME_INF) | |
10159 | return -EINVAL; | |
10160 | ||
10161 | /* | |
10162 | * Ensure we don't starve existing RT tasks. | |
10163 | */ | |
10164 | if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) | |
10165 | return -EBUSY; | |
10166 | ||
10167 | total = to_ratio(period, runtime); | |
10168 | ||
10169 | /* | |
10170 | * Nobody can have more than the global setting allows. | |
10171 | */ | |
10172 | if (total > to_ratio(global_rt_period(), global_rt_runtime())) | |
10173 | return -EINVAL; | |
10174 | ||
10175 | /* | |
10176 | * The sum of our children's runtime should not exceed our own. | |
10177 | */ | |
10178 | list_for_each_entry_rcu(child, &tg->children, siblings) { | |
10179 | period = ktime_to_ns(child->rt_bandwidth.rt_period); | |
10180 | runtime = child->rt_bandwidth.rt_runtime; | |
10181 | ||
10182 | if (child == d->tg) { | |
10183 | period = d->rt_period; | |
10184 | runtime = d->rt_runtime; | |
10185 | } | |
10186 | ||
10187 | sum += to_ratio(period, runtime); | |
10188 | } | |
10189 | ||
10190 | if (sum > total) | |
10191 | return -EINVAL; | |
10192 | ||
10193 | return 0; | |
10194 | } | |
10195 | ||
10196 | static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) | |
10197 | { | |
10198 | struct rt_schedulable_data data = { | |
10199 | .tg = tg, | |
10200 | .rt_period = period, | |
10201 | .rt_runtime = runtime, | |
10202 | }; | |
10203 | ||
10204 | return walk_tg_tree(tg_schedulable, tg_nop, &data); | |
10205 | } | |
10206 | ||
10207 | static int tg_set_bandwidth(struct task_group *tg, | |
10208 | u64 rt_period, u64 rt_runtime) | |
10209 | { | |
10210 | int i, err = 0; | |
10211 | ||
10212 | mutex_lock(&rt_constraints_mutex); | |
10213 | read_lock(&tasklist_lock); | |
10214 | err = __rt_schedulable(tg, rt_period, rt_runtime); | |
10215 | if (err) | |
10216 | goto unlock; | |
10217 | ||
10218 | spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); | |
10219 | tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); | |
10220 | tg->rt_bandwidth.rt_runtime = rt_runtime; | |
10221 | ||
10222 | for_each_possible_cpu(i) { | |
10223 | struct rt_rq *rt_rq = tg->rt_rq[i]; | |
10224 | ||
10225 | spin_lock(&rt_rq->rt_runtime_lock); | |
10226 | rt_rq->rt_runtime = rt_runtime; | |
10227 | spin_unlock(&rt_rq->rt_runtime_lock); | |
10228 | } | |
10229 | spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); | |
10230 | unlock: | |
10231 | read_unlock(&tasklist_lock); | |
10232 | mutex_unlock(&rt_constraints_mutex); | |
10233 | ||
10234 | return err; | |
10235 | } | |
10236 | ||
10237 | int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) | |
10238 | { | |
10239 | u64 rt_runtime, rt_period; | |
10240 | ||
10241 | rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); | |
10242 | rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; | |
10243 | if (rt_runtime_us < 0) | |
10244 | rt_runtime = RUNTIME_INF; | |
10245 | ||
10246 | return tg_set_bandwidth(tg, rt_period, rt_runtime); | |
10247 | } | |
10248 | ||
10249 | long sched_group_rt_runtime(struct task_group *tg) | |
10250 | { | |
10251 | u64 rt_runtime_us; | |
10252 | ||
10253 | if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) | |
10254 | return -1; | |
10255 | ||
10256 | rt_runtime_us = tg->rt_bandwidth.rt_runtime; | |
10257 | do_div(rt_runtime_us, NSEC_PER_USEC); | |
10258 | return rt_runtime_us; | |
10259 | } | |
10260 | ||
10261 | int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) | |
10262 | { | |
10263 | u64 rt_runtime, rt_period; | |
10264 | ||
10265 | rt_period = (u64)rt_period_us * NSEC_PER_USEC; | |
10266 | rt_runtime = tg->rt_bandwidth.rt_runtime; | |
10267 | ||
10268 | if (rt_period == 0) | |
10269 | return -EINVAL; | |
10270 | ||
10271 | return tg_set_bandwidth(tg, rt_period, rt_runtime); | |
10272 | } | |
10273 | ||
10274 | long sched_group_rt_period(struct task_group *tg) | |
10275 | { | |
10276 | u64 rt_period_us; | |
10277 | ||
10278 | rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); | |
10279 | do_div(rt_period_us, NSEC_PER_USEC); | |
10280 | return rt_period_us; | |
10281 | } | |
10282 | ||
10283 | static int sched_rt_global_constraints(void) | |
10284 | { | |
10285 | u64 runtime, period; | |
10286 | int ret = 0; | |
10287 | ||
10288 | if (sysctl_sched_rt_period <= 0) | |
10289 | return -EINVAL; | |
10290 | ||
10291 | runtime = global_rt_runtime(); | |
10292 | period = global_rt_period(); | |
10293 | ||
10294 | /* | |
10295 | * Sanity check on the sysctl variables. | |
10296 | */ | |
10297 | if (runtime > period && runtime != RUNTIME_INF) | |
10298 | return -EINVAL; | |
10299 | ||
10300 | mutex_lock(&rt_constraints_mutex); | |
10301 | read_lock(&tasklist_lock); | |
10302 | ret = __rt_schedulable(NULL, 0, 0); | |
10303 | read_unlock(&tasklist_lock); | |
10304 | mutex_unlock(&rt_constraints_mutex); | |
10305 | ||
10306 | return ret; | |
10307 | } | |
10308 | ||
10309 | int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) | |
10310 | { | |
10311 | /* Don't accept realtime tasks when there is no way for them to run */ | |
10312 | if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) | |
10313 | return 0; | |
10314 | ||
10315 | return 1; | |
10316 | } | |
10317 | ||
10318 | #else /* !CONFIG_RT_GROUP_SCHED */ | |
10319 | static int sched_rt_global_constraints(void) | |
10320 | { | |
10321 | unsigned long flags; | |
10322 | int i; | |
10323 | ||
10324 | if (sysctl_sched_rt_period <= 0) | |
10325 | return -EINVAL; | |
10326 | ||
10327 | /* | |
10328 | * There's always some RT tasks in the root group | |
10329 | * -- migration, kstopmachine etc.. | |
10330 | */ | |
10331 | if (sysctl_sched_rt_runtime == 0) | |
10332 | return -EBUSY; | |
10333 | ||
10334 | spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); | |
10335 | for_each_possible_cpu(i) { | |
10336 | struct rt_rq *rt_rq = &cpu_rq(i)->rt; | |
10337 | ||
10338 | spin_lock(&rt_rq->rt_runtime_lock); | |
10339 | rt_rq->rt_runtime = global_rt_runtime(); | |
10340 | spin_unlock(&rt_rq->rt_runtime_lock); | |
10341 | } | |
10342 | spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); | |
10343 | ||
10344 | return 0; | |
10345 | } | |
10346 | #endif /* CONFIG_RT_GROUP_SCHED */ | |
10347 | ||
10348 | int sched_rt_handler(struct ctl_table *table, int write, | |
10349 | void __user *buffer, size_t *lenp, | |
10350 | loff_t *ppos) | |
10351 | { | |
10352 | int ret; | |
10353 | int old_period, old_runtime; | |
10354 | static DEFINE_MUTEX(mutex); | |
10355 | ||
10356 | mutex_lock(&mutex); | |
10357 | old_period = sysctl_sched_rt_period; | |
10358 | old_runtime = sysctl_sched_rt_runtime; | |
10359 | ||
10360 | ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
10361 | ||
10362 | if (!ret && write) { | |
10363 | ret = sched_rt_global_constraints(); | |
10364 | if (ret) { | |
10365 | sysctl_sched_rt_period = old_period; | |
10366 | sysctl_sched_rt_runtime = old_runtime; | |
10367 | } else { | |
10368 | def_rt_bandwidth.rt_runtime = global_rt_runtime(); | |
10369 | def_rt_bandwidth.rt_period = | |
10370 | ns_to_ktime(global_rt_period()); | |
10371 | } | |
10372 | } | |
10373 | mutex_unlock(&mutex); | |
10374 | ||
10375 | return ret; | |
10376 | } | |
10377 | ||
10378 | #ifdef CONFIG_CGROUP_SCHED | |
10379 | ||
10380 | /* return corresponding task_group object of a cgroup */ | |
10381 | static inline struct task_group *cgroup_tg(struct cgroup *cgrp) | |
10382 | { | |
10383 | return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id), | |
10384 | struct task_group, css); | |
10385 | } | |
10386 | ||
10387 | static struct cgroup_subsys_state * | |
10388 | cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp) | |
10389 | { | |
10390 | struct task_group *tg, *parent; | |
10391 | ||
10392 | if (!cgrp->parent) { | |
10393 | /* This is early initialization for the top cgroup */ | |
10394 | return &init_task_group.css; | |
10395 | } | |
10396 | ||
10397 | parent = cgroup_tg(cgrp->parent); | |
10398 | tg = sched_create_group(parent); | |
10399 | if (IS_ERR(tg)) | |
10400 | return ERR_PTR(-ENOMEM); | |
10401 | ||
10402 | return &tg->css; | |
10403 | } | |
10404 | ||
10405 | static void | |
10406 | cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp) | |
10407 | { | |
10408 | struct task_group *tg = cgroup_tg(cgrp); | |
10409 | ||
10410 | sched_destroy_group(tg); | |
10411 | } | |
10412 | ||
10413 | static int | |
10414 | cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk) | |
10415 | { | |
10416 | #ifdef CONFIG_RT_GROUP_SCHED | |
10417 | if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk)) | |
10418 | return -EINVAL; | |
10419 | #else | |
10420 | /* We don't support RT-tasks being in separate groups */ | |
10421 | if (tsk->sched_class != &fair_sched_class) | |
10422 | return -EINVAL; | |
10423 | #endif | |
10424 | return 0; | |
10425 | } | |
10426 | ||
10427 | static int | |
10428 | cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp, | |
10429 | struct task_struct *tsk, bool threadgroup) | |
10430 | { | |
10431 | int retval = cpu_cgroup_can_attach_task(cgrp, tsk); | |
10432 | if (retval) | |
10433 | return retval; | |
10434 | if (threadgroup) { | |
10435 | struct task_struct *c; | |
10436 | rcu_read_lock(); | |
10437 | list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) { | |
10438 | retval = cpu_cgroup_can_attach_task(cgrp, c); | |
10439 | if (retval) { | |
10440 | rcu_read_unlock(); | |
10441 | return retval; | |
10442 | } | |
10443 | } | |
10444 | rcu_read_unlock(); | |
10445 | } | |
10446 | return 0; | |
10447 | } | |
10448 | ||
10449 | static void | |
10450 | cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp, | |
10451 | struct cgroup *old_cont, struct task_struct *tsk, | |
10452 | bool threadgroup) | |
10453 | { | |
10454 | sched_move_task(tsk); | |
10455 | if (threadgroup) { | |
10456 | struct task_struct *c; | |
10457 | rcu_read_lock(); | |
10458 | list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) { | |
10459 | sched_move_task(c); | |
10460 | } | |
10461 | rcu_read_unlock(); | |
10462 | } | |
10463 | } | |
10464 | ||
10465 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
10466 | static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype, | |
10467 | u64 shareval) | |
10468 | { | |
10469 | return sched_group_set_shares(cgroup_tg(cgrp), shareval); | |
10470 | } | |
10471 | ||
10472 | static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft) | |
10473 | { | |
10474 | struct task_group *tg = cgroup_tg(cgrp); | |
10475 | ||
10476 | return (u64) tg->shares; | |
10477 | } | |
10478 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
10479 | ||
10480 | #ifdef CONFIG_RT_GROUP_SCHED | |
10481 | static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft, | |
10482 | s64 val) | |
10483 | { | |
10484 | return sched_group_set_rt_runtime(cgroup_tg(cgrp), val); | |
10485 | } | |
10486 | ||
10487 | static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft) | |
10488 | { | |
10489 | return sched_group_rt_runtime(cgroup_tg(cgrp)); | |
10490 | } | |
10491 | ||
10492 | static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype, | |
10493 | u64 rt_period_us) | |
10494 | { | |
10495 | return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us); | |
10496 | } | |
10497 | ||
10498 | static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft) | |
10499 | { | |
10500 | return sched_group_rt_period(cgroup_tg(cgrp)); | |
10501 | } | |
10502 | #endif /* CONFIG_RT_GROUP_SCHED */ | |
10503 | ||
10504 | static struct cftype cpu_files[] = { | |
10505 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
10506 | { | |
10507 | .name = "shares", | |
10508 | .read_u64 = cpu_shares_read_u64, | |
10509 | .write_u64 = cpu_shares_write_u64, | |
10510 | }, | |
10511 | #endif | |
10512 | #ifdef CONFIG_RT_GROUP_SCHED | |
10513 | { | |
10514 | .name = "rt_runtime_us", | |
10515 | .read_s64 = cpu_rt_runtime_read, | |
10516 | .write_s64 = cpu_rt_runtime_write, | |
10517 | }, | |
10518 | { | |
10519 | .name = "rt_period_us", | |
10520 | .read_u64 = cpu_rt_period_read_uint, | |
10521 | .write_u64 = cpu_rt_period_write_uint, | |
10522 | }, | |
10523 | #endif | |
10524 | }; | |
10525 | ||
10526 | static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont) | |
10527 | { | |
10528 | return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files)); | |
10529 | } | |
10530 | ||
10531 | struct cgroup_subsys cpu_cgroup_subsys = { | |
10532 | .name = "cpu", | |
10533 | .create = cpu_cgroup_create, | |
10534 | .destroy = cpu_cgroup_destroy, | |
10535 | .can_attach = cpu_cgroup_can_attach, | |
10536 | .attach = cpu_cgroup_attach, | |
10537 | .populate = cpu_cgroup_populate, | |
10538 | .subsys_id = cpu_cgroup_subsys_id, | |
10539 | .early_init = 1, | |
10540 | }; | |
10541 | ||
10542 | #endif /* CONFIG_CGROUP_SCHED */ | |
10543 | ||
10544 | #ifdef CONFIG_CGROUP_CPUACCT | |
10545 | ||
10546 | /* | |
10547 | * CPU accounting code for task groups. | |
10548 | * | |
10549 | * Based on the work by Paul Menage (menage@google.com) and Balbir Singh | |
10550 | * (balbir@in.ibm.com). | |
10551 | */ | |
10552 | ||
10553 | /* track cpu usage of a group of tasks and its child groups */ | |
10554 | struct cpuacct { | |
10555 | struct cgroup_subsys_state css; | |
10556 | /* cpuusage holds pointer to a u64-type object on every cpu */ | |
10557 | u64 *cpuusage; | |
10558 | struct percpu_counter cpustat[CPUACCT_STAT_NSTATS]; | |
10559 | struct cpuacct *parent; | |
10560 | }; | |
10561 | ||
10562 | struct cgroup_subsys cpuacct_subsys; | |
10563 | ||
10564 | /* return cpu accounting group corresponding to this container */ | |
10565 | static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp) | |
10566 | { | |
10567 | return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id), | |
10568 | struct cpuacct, css); | |
10569 | } | |
10570 | ||
10571 | /* return cpu accounting group to which this task belongs */ | |
10572 | static inline struct cpuacct *task_ca(struct task_struct *tsk) | |
10573 | { | |
10574 | return container_of(task_subsys_state(tsk, cpuacct_subsys_id), | |
10575 | struct cpuacct, css); | |
10576 | } | |
10577 | ||
10578 | /* create a new cpu accounting group */ | |
10579 | static struct cgroup_subsys_state *cpuacct_create( | |
10580 | struct cgroup_subsys *ss, struct cgroup *cgrp) | |
10581 | { | |
10582 | struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL); | |
10583 | int i; | |
10584 | ||
10585 | if (!ca) | |
10586 | goto out; | |
10587 | ||
10588 | ca->cpuusage = alloc_percpu(u64); | |
10589 | if (!ca->cpuusage) | |
10590 | goto out_free_ca; | |
10591 | ||
10592 | for (i = 0; i < CPUACCT_STAT_NSTATS; i++) | |
10593 | if (percpu_counter_init(&ca->cpustat[i], 0)) | |
10594 | goto out_free_counters; | |
10595 | ||
10596 | if (cgrp->parent) | |
10597 | ca->parent = cgroup_ca(cgrp->parent); | |
10598 | ||
10599 | return &ca->css; | |
10600 | ||
10601 | out_free_counters: | |
10602 | while (--i >= 0) | |
10603 | percpu_counter_destroy(&ca->cpustat[i]); | |
10604 | free_percpu(ca->cpuusage); | |
10605 | out_free_ca: | |
10606 | kfree(ca); | |
10607 | out: | |
10608 | return ERR_PTR(-ENOMEM); | |
10609 | } | |
10610 | ||
10611 | /* destroy an existing cpu accounting group */ | |
10612 | static void | |
10613 | cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp) | |
10614 | { | |
10615 | struct cpuacct *ca = cgroup_ca(cgrp); | |
10616 | int i; | |
10617 | ||
10618 | for (i = 0; i < CPUACCT_STAT_NSTATS; i++) | |
10619 | percpu_counter_destroy(&ca->cpustat[i]); | |
10620 | free_percpu(ca->cpuusage); | |
10621 | kfree(ca); | |
10622 | } | |
10623 | ||
10624 | static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu) | |
10625 | { | |
10626 | u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); | |
10627 | u64 data; | |
10628 | ||
10629 | #ifndef CONFIG_64BIT | |
10630 | /* | |
10631 | * Take rq->lock to make 64-bit read safe on 32-bit platforms. | |
10632 | */ | |
10633 | spin_lock_irq(&cpu_rq(cpu)->lock); | |
10634 | data = *cpuusage; | |
10635 | spin_unlock_irq(&cpu_rq(cpu)->lock); | |
10636 | #else | |
10637 | data = *cpuusage; | |
10638 | #endif | |
10639 | ||
10640 | return data; | |
10641 | } | |
10642 | ||
10643 | static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val) | |
10644 | { | |
10645 | u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); | |
10646 | ||
10647 | #ifndef CONFIG_64BIT | |
10648 | /* | |
10649 | * Take rq->lock to make 64-bit write safe on 32-bit platforms. | |
10650 | */ | |
10651 | spin_lock_irq(&cpu_rq(cpu)->lock); | |
10652 | *cpuusage = val; | |
10653 | spin_unlock_irq(&cpu_rq(cpu)->lock); | |
10654 | #else | |
10655 | *cpuusage = val; | |
10656 | #endif | |
10657 | } | |
10658 | ||
10659 | /* return total cpu usage (in nanoseconds) of a group */ | |
10660 | static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft) | |
10661 | { | |
10662 | struct cpuacct *ca = cgroup_ca(cgrp); | |
10663 | u64 totalcpuusage = 0; | |
10664 | int i; | |
10665 | ||
10666 | for_each_present_cpu(i) | |
10667 | totalcpuusage += cpuacct_cpuusage_read(ca, i); | |
10668 | ||
10669 | return totalcpuusage; | |
10670 | } | |
10671 | ||
10672 | static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype, | |
10673 | u64 reset) | |
10674 | { | |
10675 | struct cpuacct *ca = cgroup_ca(cgrp); | |
10676 | int err = 0; | |
10677 | int i; | |
10678 | ||
10679 | if (reset) { | |
10680 | err = -EINVAL; | |
10681 | goto out; | |
10682 | } | |
10683 | ||
10684 | for_each_present_cpu(i) | |
10685 | cpuacct_cpuusage_write(ca, i, 0); | |
10686 | ||
10687 | out: | |
10688 | return err; | |
10689 | } | |
10690 | ||
10691 | static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft, | |
10692 | struct seq_file *m) | |
10693 | { | |
10694 | struct cpuacct *ca = cgroup_ca(cgroup); | |
10695 | u64 percpu; | |
10696 | int i; | |
10697 | ||
10698 | for_each_present_cpu(i) { | |
10699 | percpu = cpuacct_cpuusage_read(ca, i); | |
10700 | seq_printf(m, "%llu ", (unsigned long long) percpu); | |
10701 | } | |
10702 | seq_printf(m, "\n"); | |
10703 | return 0; | |
10704 | } | |
10705 | ||
10706 | static const char *cpuacct_stat_desc[] = { | |
10707 | [CPUACCT_STAT_USER] = "user", | |
10708 | [CPUACCT_STAT_SYSTEM] = "system", | |
10709 | }; | |
10710 | ||
10711 | static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft, | |
10712 | struct cgroup_map_cb *cb) | |
10713 | { | |
10714 | struct cpuacct *ca = cgroup_ca(cgrp); | |
10715 | int i; | |
10716 | ||
10717 | for (i = 0; i < CPUACCT_STAT_NSTATS; i++) { | |
10718 | s64 val = percpu_counter_read(&ca->cpustat[i]); | |
10719 | val = cputime64_to_clock_t(val); | |
10720 | cb->fill(cb, cpuacct_stat_desc[i], val); | |
10721 | } | |
10722 | return 0; | |
10723 | } | |
10724 | ||
10725 | static struct cftype files[] = { | |
10726 | { | |
10727 | .name = "usage", | |
10728 | .read_u64 = cpuusage_read, | |
10729 | .write_u64 = cpuusage_write, | |
10730 | }, | |
10731 | { | |
10732 | .name = "usage_percpu", | |
10733 | .read_seq_string = cpuacct_percpu_seq_read, | |
10734 | }, | |
10735 | { | |
10736 | .name = "stat", | |
10737 | .read_map = cpuacct_stats_show, | |
10738 | }, | |
10739 | }; | |
10740 | ||
10741 | static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp) | |
10742 | { | |
10743 | return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files)); | |
10744 | } | |
10745 | ||
10746 | /* | |
10747 | * charge this task's execution time to its accounting group. | |
10748 | * | |
10749 | * called with rq->lock held. | |
10750 | */ | |
10751 | static void cpuacct_charge(struct task_struct *tsk, u64 cputime) | |
10752 | { | |
10753 | struct cpuacct *ca; | |
10754 | int cpu; | |
10755 | ||
10756 | if (unlikely(!cpuacct_subsys.active)) | |
10757 | return; | |
10758 | ||
10759 | cpu = task_cpu(tsk); | |
10760 | ||
10761 | rcu_read_lock(); | |
10762 | ||
10763 | ca = task_ca(tsk); | |
10764 | ||
10765 | for (; ca; ca = ca->parent) { | |
10766 | u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); | |
10767 | *cpuusage += cputime; | |
10768 | } | |
10769 | ||
10770 | rcu_read_unlock(); | |
10771 | } | |
10772 | ||
10773 | /* | |
10774 | * Charge the system/user time to the task's accounting group. | |
10775 | */ | |
10776 | static void cpuacct_update_stats(struct task_struct *tsk, | |
10777 | enum cpuacct_stat_index idx, cputime_t val) | |
10778 | { | |
10779 | struct cpuacct *ca; | |
10780 | ||
10781 | if (unlikely(!cpuacct_subsys.active)) | |
10782 | return; | |
10783 | ||
10784 | rcu_read_lock(); | |
10785 | ca = task_ca(tsk); | |
10786 | ||
10787 | do { | |
10788 | percpu_counter_add(&ca->cpustat[idx], val); | |
10789 | ca = ca->parent; | |
10790 | } while (ca); | |
10791 | rcu_read_unlock(); | |
10792 | } | |
10793 | ||
10794 | struct cgroup_subsys cpuacct_subsys = { | |
10795 | .name = "cpuacct", | |
10796 | .create = cpuacct_create, | |
10797 | .destroy = cpuacct_destroy, | |
10798 | .populate = cpuacct_populate, | |
10799 | .subsys_id = cpuacct_subsys_id, | |
10800 | }; | |
10801 | #endif /* CONFIG_CGROUP_CPUACCT */ | |
10802 | ||
10803 | #ifndef CONFIG_SMP | |
10804 | ||
10805 | int rcu_expedited_torture_stats(char *page) | |
10806 | { | |
10807 | return 0; | |
10808 | } | |
10809 | EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats); | |
10810 | ||
10811 | void synchronize_sched_expedited(void) | |
10812 | { | |
10813 | } | |
10814 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited); | |
10815 | ||
10816 | #else /* #ifndef CONFIG_SMP */ | |
10817 | ||
10818 | static DEFINE_PER_CPU(struct migration_req, rcu_migration_req); | |
10819 | static DEFINE_MUTEX(rcu_sched_expedited_mutex); | |
10820 | ||
10821 | #define RCU_EXPEDITED_STATE_POST -2 | |
10822 | #define RCU_EXPEDITED_STATE_IDLE -1 | |
10823 | ||
10824 | static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE; | |
10825 | ||
10826 | int rcu_expedited_torture_stats(char *page) | |
10827 | { | |
10828 | int cnt = 0; | |
10829 | int cpu; | |
10830 | ||
10831 | cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state); | |
10832 | for_each_online_cpu(cpu) { | |
10833 | cnt += sprintf(&page[cnt], " %d:%d", | |
10834 | cpu, per_cpu(rcu_migration_req, cpu).dest_cpu); | |
10835 | } | |
10836 | cnt += sprintf(&page[cnt], "\n"); | |
10837 | return cnt; | |
10838 | } | |
10839 | EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats); | |
10840 | ||
10841 | static long synchronize_sched_expedited_count; | |
10842 | ||
10843 | /* | |
10844 | * Wait for an rcu-sched grace period to elapse, but use "big hammer" | |
10845 | * approach to force grace period to end quickly. This consumes | |
10846 | * significant time on all CPUs, and is thus not recommended for | |
10847 | * any sort of common-case code. | |
10848 | * | |
10849 | * Note that it is illegal to call this function while holding any | |
10850 | * lock that is acquired by a CPU-hotplug notifier. Failing to | |
10851 | * observe this restriction will result in deadlock. | |
10852 | */ | |
10853 | void synchronize_sched_expedited(void) | |
10854 | { | |
10855 | int cpu; | |
10856 | unsigned long flags; | |
10857 | bool need_full_sync = 0; | |
10858 | struct rq *rq; | |
10859 | struct migration_req *req; | |
10860 | long snap; | |
10861 | int trycount = 0; | |
10862 | ||
10863 | smp_mb(); /* ensure prior mod happens before capturing snap. */ | |
10864 | snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1; | |
10865 | get_online_cpus(); | |
10866 | while (!mutex_trylock(&rcu_sched_expedited_mutex)) { | |
10867 | put_online_cpus(); | |
10868 | if (trycount++ < 10) | |
10869 | udelay(trycount * num_online_cpus()); | |
10870 | else { | |
10871 | synchronize_sched(); | |
10872 | return; | |
10873 | } | |
10874 | if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) { | |
10875 | smp_mb(); /* ensure test happens before caller kfree */ | |
10876 | return; | |
10877 | } | |
10878 | get_online_cpus(); | |
10879 | } | |
10880 | rcu_expedited_state = RCU_EXPEDITED_STATE_POST; | |
10881 | for_each_online_cpu(cpu) { | |
10882 | rq = cpu_rq(cpu); | |
10883 | req = &per_cpu(rcu_migration_req, cpu); | |
10884 | init_completion(&req->done); | |
10885 | req->task = NULL; | |
10886 | req->dest_cpu = RCU_MIGRATION_NEED_QS; | |
10887 | spin_lock_irqsave(&rq->lock, flags); | |
10888 | list_add(&req->list, &rq->migration_queue); | |
10889 | spin_unlock_irqrestore(&rq->lock, flags); | |
10890 | wake_up_process(rq->migration_thread); | |
10891 | } | |
10892 | for_each_online_cpu(cpu) { | |
10893 | rcu_expedited_state = cpu; | |
10894 | req = &per_cpu(rcu_migration_req, cpu); | |
10895 | rq = cpu_rq(cpu); | |
10896 | wait_for_completion(&req->done); | |
10897 | spin_lock_irqsave(&rq->lock, flags); | |
10898 | if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC)) | |
10899 | need_full_sync = 1; | |
10900 | req->dest_cpu = RCU_MIGRATION_IDLE; | |
10901 | spin_unlock_irqrestore(&rq->lock, flags); | |
10902 | } | |
10903 | rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE; | |
10904 | mutex_unlock(&rcu_sched_expedited_mutex); | |
10905 | put_online_cpus(); | |
10906 | if (need_full_sync) | |
10907 | synchronize_sched(); | |
10908 | } | |
10909 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited); | |
10910 | ||
10911 | #endif /* #else #ifndef CONFIG_SMP */ |