]>
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/proc_fs.h> | |
59 | #include <linux/seq_file.h> | |
60 | #include <linux/stop_machine.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 | #include <linux/slab.h> | |
75 | ||
76 | #include <asm/tlb.h> | |
77 | #include <asm/irq_regs.h> | |
78 | ||
79 | #include "sched_cpupri.h" | |
80 | #include "workqueue_sched.h" | |
81 | ||
82 | #define CREATE_TRACE_POINTS | |
83 | #include <trace/events/sched.h> | |
84 | ||
85 | /* | |
86 | * Convert user-nice values [ -20 ... 0 ... 19 ] | |
87 | * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ], | |
88 | * and back. | |
89 | */ | |
90 | #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20) | |
91 | #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20) | |
92 | #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio) | |
93 | ||
94 | /* | |
95 | * 'User priority' is the nice value converted to something we | |
96 | * can work with better when scaling various scheduler parameters, | |
97 | * it's a [ 0 ... 39 ] range. | |
98 | */ | |
99 | #define USER_PRIO(p) ((p)-MAX_RT_PRIO) | |
100 | #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio) | |
101 | #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO)) | |
102 | ||
103 | /* | |
104 | * Helpers for converting nanosecond timing to jiffy resolution | |
105 | */ | |
106 | #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ)) | |
107 | ||
108 | #define NICE_0_LOAD SCHED_LOAD_SCALE | |
109 | #define NICE_0_SHIFT SCHED_LOAD_SHIFT | |
110 | ||
111 | /* | |
112 | * These are the 'tuning knobs' of the scheduler: | |
113 | * | |
114 | * default timeslice is 100 msecs (used only for SCHED_RR tasks). | |
115 | * Timeslices get refilled after they expire. | |
116 | */ | |
117 | #define DEF_TIMESLICE (100 * HZ / 1000) | |
118 | ||
119 | /* | |
120 | * single value that denotes runtime == period, ie unlimited time. | |
121 | */ | |
122 | #define RUNTIME_INF ((u64)~0ULL) | |
123 | ||
124 | static inline int rt_policy(int policy) | |
125 | { | |
126 | if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR)) | |
127 | return 1; | |
128 | return 0; | |
129 | } | |
130 | ||
131 | static inline int task_has_rt_policy(struct task_struct *p) | |
132 | { | |
133 | return rt_policy(p->policy); | |
134 | } | |
135 | ||
136 | /* | |
137 | * This is the priority-queue data structure of the RT scheduling class: | |
138 | */ | |
139 | struct rt_prio_array { | |
140 | DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */ | |
141 | struct list_head queue[MAX_RT_PRIO]; | |
142 | }; | |
143 | ||
144 | struct rt_bandwidth { | |
145 | /* nests inside the rq lock: */ | |
146 | raw_spinlock_t rt_runtime_lock; | |
147 | ktime_t rt_period; | |
148 | u64 rt_runtime; | |
149 | struct hrtimer rt_period_timer; | |
150 | }; | |
151 | ||
152 | static struct rt_bandwidth def_rt_bandwidth; | |
153 | ||
154 | static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun); | |
155 | ||
156 | static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer) | |
157 | { | |
158 | struct rt_bandwidth *rt_b = | |
159 | container_of(timer, struct rt_bandwidth, rt_period_timer); | |
160 | ktime_t now; | |
161 | int overrun; | |
162 | int idle = 0; | |
163 | ||
164 | for (;;) { | |
165 | now = hrtimer_cb_get_time(timer); | |
166 | overrun = hrtimer_forward(timer, now, rt_b->rt_period); | |
167 | ||
168 | if (!overrun) | |
169 | break; | |
170 | ||
171 | idle = do_sched_rt_period_timer(rt_b, overrun); | |
172 | } | |
173 | ||
174 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
175 | } | |
176 | ||
177 | static | |
178 | void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime) | |
179 | { | |
180 | rt_b->rt_period = ns_to_ktime(period); | |
181 | rt_b->rt_runtime = runtime; | |
182 | ||
183 | raw_spin_lock_init(&rt_b->rt_runtime_lock); | |
184 | ||
185 | hrtimer_init(&rt_b->rt_period_timer, | |
186 | CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
187 | rt_b->rt_period_timer.function = sched_rt_period_timer; | |
188 | } | |
189 | ||
190 | static inline int rt_bandwidth_enabled(void) | |
191 | { | |
192 | return sysctl_sched_rt_runtime >= 0; | |
193 | } | |
194 | ||
195 | static void start_rt_bandwidth(struct rt_bandwidth *rt_b) | |
196 | { | |
197 | ktime_t now; | |
198 | ||
199 | if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF) | |
200 | return; | |
201 | ||
202 | if (hrtimer_active(&rt_b->rt_period_timer)) | |
203 | return; | |
204 | ||
205 | raw_spin_lock(&rt_b->rt_runtime_lock); | |
206 | for (;;) { | |
207 | unsigned long delta; | |
208 | ktime_t soft, hard; | |
209 | ||
210 | if (hrtimer_active(&rt_b->rt_period_timer)) | |
211 | break; | |
212 | ||
213 | now = hrtimer_cb_get_time(&rt_b->rt_period_timer); | |
214 | hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period); | |
215 | ||
216 | soft = hrtimer_get_softexpires(&rt_b->rt_period_timer); | |
217 | hard = hrtimer_get_expires(&rt_b->rt_period_timer); | |
218 | delta = ktime_to_ns(ktime_sub(hard, soft)); | |
219 | __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta, | |
220 | HRTIMER_MODE_ABS_PINNED, 0); | |
221 | } | |
222 | raw_spin_unlock(&rt_b->rt_runtime_lock); | |
223 | } | |
224 | ||
225 | #ifdef CONFIG_RT_GROUP_SCHED | |
226 | static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b) | |
227 | { | |
228 | hrtimer_cancel(&rt_b->rt_period_timer); | |
229 | } | |
230 | #endif | |
231 | ||
232 | /* | |
233 | * sched_domains_mutex serializes calls to arch_init_sched_domains, | |
234 | * detach_destroy_domains and partition_sched_domains. | |
235 | */ | |
236 | static DEFINE_MUTEX(sched_domains_mutex); | |
237 | ||
238 | #ifdef CONFIG_CGROUP_SCHED | |
239 | ||
240 | #include <linux/cgroup.h> | |
241 | ||
242 | struct cfs_rq; | |
243 | ||
244 | static LIST_HEAD(task_groups); | |
245 | ||
246 | /* task group related information */ | |
247 | struct task_group { | |
248 | struct cgroup_subsys_state css; | |
249 | ||
250 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
251 | /* schedulable entities of this group on each cpu */ | |
252 | struct sched_entity **se; | |
253 | /* runqueue "owned" by this group on each cpu */ | |
254 | struct cfs_rq **cfs_rq; | |
255 | unsigned long shares; | |
256 | #endif | |
257 | ||
258 | #ifdef CONFIG_RT_GROUP_SCHED | |
259 | struct sched_rt_entity **rt_se; | |
260 | struct rt_rq **rt_rq; | |
261 | ||
262 | struct rt_bandwidth rt_bandwidth; | |
263 | #endif | |
264 | ||
265 | struct rcu_head rcu; | |
266 | struct list_head list; | |
267 | ||
268 | struct task_group *parent; | |
269 | struct list_head siblings; | |
270 | struct list_head children; | |
271 | }; | |
272 | ||
273 | #define root_task_group init_task_group | |
274 | ||
275 | /* task_group_lock serializes add/remove of task groups and also changes to | |
276 | * a task group's cpu shares. | |
277 | */ | |
278 | static DEFINE_SPINLOCK(task_group_lock); | |
279 | ||
280 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
281 | ||
282 | #ifdef CONFIG_SMP | |
283 | static int root_task_group_empty(void) | |
284 | { | |
285 | return list_empty(&root_task_group.children); | |
286 | } | |
287 | #endif | |
288 | ||
289 | # define INIT_TASK_GROUP_LOAD NICE_0_LOAD | |
290 | ||
291 | /* | |
292 | * A weight of 0 or 1 can cause arithmetics problems. | |
293 | * A weight of a cfs_rq is the sum of weights of which entities | |
294 | * are queued on this cfs_rq, so a weight of a entity should not be | |
295 | * too large, so as the shares value of a task group. | |
296 | * (The default weight is 1024 - so there's no practical | |
297 | * limitation from this.) | |
298 | */ | |
299 | #define MIN_SHARES 2 | |
300 | #define MAX_SHARES (1UL << 18) | |
301 | ||
302 | static int init_task_group_load = INIT_TASK_GROUP_LOAD; | |
303 | #endif | |
304 | ||
305 | /* Default task group. | |
306 | * Every task in system belong to this group at bootup. | |
307 | */ | |
308 | struct task_group init_task_group; | |
309 | ||
310 | #endif /* CONFIG_CGROUP_SCHED */ | |
311 | ||
312 | /* CFS-related fields in a runqueue */ | |
313 | struct cfs_rq { | |
314 | struct load_weight load; | |
315 | unsigned long nr_running; | |
316 | ||
317 | u64 exec_clock; | |
318 | u64 min_vruntime; | |
319 | ||
320 | struct rb_root tasks_timeline; | |
321 | struct rb_node *rb_leftmost; | |
322 | ||
323 | struct list_head tasks; | |
324 | struct list_head *balance_iterator; | |
325 | ||
326 | /* | |
327 | * 'curr' points to currently running entity on this cfs_rq. | |
328 | * It is set to NULL otherwise (i.e when none are currently running). | |
329 | */ | |
330 | struct sched_entity *curr, *next, *last; | |
331 | ||
332 | unsigned int nr_spread_over; | |
333 | ||
334 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
335 | struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */ | |
336 | ||
337 | /* | |
338 | * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in | |
339 | * a hierarchy). Non-leaf lrqs hold other higher schedulable entities | |
340 | * (like users, containers etc.) | |
341 | * | |
342 | * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This | |
343 | * list is used during load balance. | |
344 | */ | |
345 | struct list_head leaf_cfs_rq_list; | |
346 | struct task_group *tg; /* group that "owns" this runqueue */ | |
347 | ||
348 | #ifdef CONFIG_SMP | |
349 | /* | |
350 | * the part of load.weight contributed by tasks | |
351 | */ | |
352 | unsigned long task_weight; | |
353 | ||
354 | /* | |
355 | * h_load = weight * f(tg) | |
356 | * | |
357 | * Where f(tg) is the recursive weight fraction assigned to | |
358 | * this group. | |
359 | */ | |
360 | unsigned long h_load; | |
361 | ||
362 | /* | |
363 | * this cpu's part of tg->shares | |
364 | */ | |
365 | unsigned long shares; | |
366 | ||
367 | /* | |
368 | * load.weight at the time we set shares | |
369 | */ | |
370 | unsigned long rq_weight; | |
371 | #endif | |
372 | #endif | |
373 | }; | |
374 | ||
375 | /* Real-Time classes' related field in a runqueue: */ | |
376 | struct rt_rq { | |
377 | struct rt_prio_array active; | |
378 | unsigned long rt_nr_running; | |
379 | #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED | |
380 | struct { | |
381 | int curr; /* highest queued rt task prio */ | |
382 | #ifdef CONFIG_SMP | |
383 | int next; /* next highest */ | |
384 | #endif | |
385 | } highest_prio; | |
386 | #endif | |
387 | #ifdef CONFIG_SMP | |
388 | unsigned long rt_nr_migratory; | |
389 | unsigned long rt_nr_total; | |
390 | int overloaded; | |
391 | struct plist_head pushable_tasks; | |
392 | #endif | |
393 | int rt_throttled; | |
394 | u64 rt_time; | |
395 | u64 rt_runtime; | |
396 | /* Nests inside the rq lock: */ | |
397 | raw_spinlock_t rt_runtime_lock; | |
398 | ||
399 | #ifdef CONFIG_RT_GROUP_SCHED | |
400 | unsigned long rt_nr_boosted; | |
401 | ||
402 | struct rq *rq; | |
403 | struct list_head leaf_rt_rq_list; | |
404 | struct task_group *tg; | |
405 | #endif | |
406 | }; | |
407 | ||
408 | #ifdef CONFIG_SMP | |
409 | ||
410 | /* | |
411 | * We add the notion of a root-domain which will be used to define per-domain | |
412 | * variables. Each exclusive cpuset essentially defines an island domain by | |
413 | * fully partitioning the member cpus from any other cpuset. Whenever a new | |
414 | * exclusive cpuset is created, we also create and attach a new root-domain | |
415 | * object. | |
416 | * | |
417 | */ | |
418 | struct root_domain { | |
419 | atomic_t refcount; | |
420 | cpumask_var_t span; | |
421 | cpumask_var_t online; | |
422 | ||
423 | /* | |
424 | * The "RT overload" flag: it gets set if a CPU has more than | |
425 | * one runnable RT task. | |
426 | */ | |
427 | cpumask_var_t rto_mask; | |
428 | atomic_t rto_count; | |
429 | struct cpupri cpupri; | |
430 | }; | |
431 | ||
432 | /* | |
433 | * By default the system creates a single root-domain with all cpus as | |
434 | * members (mimicking the global state we have today). | |
435 | */ | |
436 | static struct root_domain def_root_domain; | |
437 | ||
438 | #endif /* CONFIG_SMP */ | |
439 | ||
440 | /* | |
441 | * This is the main, per-CPU runqueue data structure. | |
442 | * | |
443 | * Locking rule: those places that want to lock multiple runqueues | |
444 | * (such as the load balancing or the thread migration code), lock | |
445 | * acquire operations must be ordered by ascending &runqueue. | |
446 | */ | |
447 | struct rq { | |
448 | /* runqueue lock: */ | |
449 | raw_spinlock_t lock; | |
450 | ||
451 | /* | |
452 | * nr_running and cpu_load should be in the same cacheline because | |
453 | * remote CPUs use both these fields when doing load calculation. | |
454 | */ | |
455 | unsigned long nr_running; | |
456 | #define CPU_LOAD_IDX_MAX 5 | |
457 | unsigned long cpu_load[CPU_LOAD_IDX_MAX]; | |
458 | unsigned long last_load_update_tick; | |
459 | #ifdef CONFIG_NO_HZ | |
460 | u64 nohz_stamp; | |
461 | unsigned char nohz_balance_kick; | |
462 | #endif | |
463 | unsigned int skip_clock_update; | |
464 | ||
465 | /* capture load from *all* tasks on this cpu: */ | |
466 | struct load_weight load; | |
467 | unsigned long nr_load_updates; | |
468 | u64 nr_switches; | |
469 | ||
470 | struct cfs_rq cfs; | |
471 | struct rt_rq rt; | |
472 | ||
473 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
474 | /* list of leaf cfs_rq on this cpu: */ | |
475 | struct list_head leaf_cfs_rq_list; | |
476 | #endif | |
477 | #ifdef CONFIG_RT_GROUP_SCHED | |
478 | struct list_head leaf_rt_rq_list; | |
479 | #endif | |
480 | ||
481 | /* | |
482 | * This is part of a global counter where only the total sum | |
483 | * over all CPUs matters. A task can increase this counter on | |
484 | * one CPU and if it got migrated afterwards it may decrease | |
485 | * it on another CPU. Always updated under the runqueue lock: | |
486 | */ | |
487 | unsigned long nr_uninterruptible; | |
488 | ||
489 | struct task_struct *curr, *idle, *stop; | |
490 | unsigned long next_balance; | |
491 | struct mm_struct *prev_mm; | |
492 | ||
493 | u64 clock; | |
494 | u64 clock_task; | |
495 | ||
496 | atomic_t nr_iowait; | |
497 | ||
498 | #ifdef CONFIG_SMP | |
499 | struct root_domain *rd; | |
500 | struct sched_domain *sd; | |
501 | ||
502 | unsigned long cpu_power; | |
503 | ||
504 | unsigned char idle_at_tick; | |
505 | /* For active balancing */ | |
506 | int post_schedule; | |
507 | int active_balance; | |
508 | int push_cpu; | |
509 | struct cpu_stop_work active_balance_work; | |
510 | /* cpu of this runqueue: */ | |
511 | int cpu; | |
512 | int online; | |
513 | ||
514 | unsigned long avg_load_per_task; | |
515 | ||
516 | u64 rt_avg; | |
517 | u64 age_stamp; | |
518 | u64 idle_stamp; | |
519 | u64 avg_idle; | |
520 | #endif | |
521 | ||
522 | #ifdef CONFIG_IRQ_TIME_ACCOUNTING | |
523 | u64 prev_irq_time; | |
524 | #endif | |
525 | ||
526 | /* calc_load related fields */ | |
527 | unsigned long calc_load_update; | |
528 | long calc_load_active; | |
529 | ||
530 | #ifdef CONFIG_SCHED_HRTICK | |
531 | #ifdef CONFIG_SMP | |
532 | int hrtick_csd_pending; | |
533 | struct call_single_data hrtick_csd; | |
534 | #endif | |
535 | struct hrtimer hrtick_timer; | |
536 | #endif | |
537 | ||
538 | #ifdef CONFIG_SCHEDSTATS | |
539 | /* latency stats */ | |
540 | struct sched_info rq_sched_info; | |
541 | unsigned long long rq_cpu_time; | |
542 | /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */ | |
543 | ||
544 | /* sys_sched_yield() stats */ | |
545 | unsigned int yld_count; | |
546 | ||
547 | /* schedule() stats */ | |
548 | unsigned int sched_switch; | |
549 | unsigned int sched_count; | |
550 | unsigned int sched_goidle; | |
551 | ||
552 | /* try_to_wake_up() stats */ | |
553 | unsigned int ttwu_count; | |
554 | unsigned int ttwu_local; | |
555 | ||
556 | /* BKL stats */ | |
557 | unsigned int bkl_count; | |
558 | #endif | |
559 | }; | |
560 | ||
561 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); | |
562 | ||
563 | ||
564 | static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags); | |
565 | ||
566 | static inline int cpu_of(struct rq *rq) | |
567 | { | |
568 | #ifdef CONFIG_SMP | |
569 | return rq->cpu; | |
570 | #else | |
571 | return 0; | |
572 | #endif | |
573 | } | |
574 | ||
575 | #define rcu_dereference_check_sched_domain(p) \ | |
576 | rcu_dereference_check((p), \ | |
577 | rcu_read_lock_sched_held() || \ | |
578 | lockdep_is_held(&sched_domains_mutex)) | |
579 | ||
580 | /* | |
581 | * The domain tree (rq->sd) is protected by RCU's quiescent state transition. | |
582 | * See detach_destroy_domains: synchronize_sched for details. | |
583 | * | |
584 | * The domain tree of any CPU may only be accessed from within | |
585 | * preempt-disabled sections. | |
586 | */ | |
587 | #define for_each_domain(cpu, __sd) \ | |
588 | for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent) | |
589 | ||
590 | #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) | |
591 | #define this_rq() (&__get_cpu_var(runqueues)) | |
592 | #define task_rq(p) cpu_rq(task_cpu(p)) | |
593 | #define cpu_curr(cpu) (cpu_rq(cpu)->curr) | |
594 | #define raw_rq() (&__raw_get_cpu_var(runqueues)) | |
595 | ||
596 | #ifdef CONFIG_CGROUP_SCHED | |
597 | ||
598 | /* | |
599 | * Return the group to which this tasks belongs. | |
600 | * | |
601 | * We use task_subsys_state_check() and extend the RCU verification | |
602 | * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach() | |
603 | * holds that lock for each task it moves into the cgroup. Therefore | |
604 | * by holding that lock, we pin the task to the current cgroup. | |
605 | */ | |
606 | static inline struct task_group *task_group(struct task_struct *p) | |
607 | { | |
608 | struct cgroup_subsys_state *css; | |
609 | ||
610 | css = task_subsys_state_check(p, cpu_cgroup_subsys_id, | |
611 | lockdep_is_held(&task_rq(p)->lock)); | |
612 | return container_of(css, struct task_group, css); | |
613 | } | |
614 | ||
615 | /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */ | |
616 | static inline void set_task_rq(struct task_struct *p, unsigned int cpu) | |
617 | { | |
618 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
619 | p->se.cfs_rq = task_group(p)->cfs_rq[cpu]; | |
620 | p->se.parent = task_group(p)->se[cpu]; | |
621 | #endif | |
622 | ||
623 | #ifdef CONFIG_RT_GROUP_SCHED | |
624 | p->rt.rt_rq = task_group(p)->rt_rq[cpu]; | |
625 | p->rt.parent = task_group(p)->rt_se[cpu]; | |
626 | #endif | |
627 | } | |
628 | ||
629 | #else /* CONFIG_CGROUP_SCHED */ | |
630 | ||
631 | static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { } | |
632 | static inline struct task_group *task_group(struct task_struct *p) | |
633 | { | |
634 | return NULL; | |
635 | } | |
636 | ||
637 | #endif /* CONFIG_CGROUP_SCHED */ | |
638 | ||
639 | static u64 irq_time_cpu(int cpu); | |
640 | static void sched_irq_time_avg_update(struct rq *rq, u64 irq_time); | |
641 | ||
642 | inline void update_rq_clock(struct rq *rq) | |
643 | { | |
644 | if (!rq->skip_clock_update) { | |
645 | int cpu = cpu_of(rq); | |
646 | u64 irq_time; | |
647 | ||
648 | rq->clock = sched_clock_cpu(cpu); | |
649 | irq_time = irq_time_cpu(cpu); | |
650 | if (rq->clock - irq_time > rq->clock_task) | |
651 | rq->clock_task = rq->clock - irq_time; | |
652 | ||
653 | sched_irq_time_avg_update(rq, irq_time); | |
654 | } | |
655 | } | |
656 | ||
657 | /* | |
658 | * Tunables that become constants when CONFIG_SCHED_DEBUG is off: | |
659 | */ | |
660 | #ifdef CONFIG_SCHED_DEBUG | |
661 | # define const_debug __read_mostly | |
662 | #else | |
663 | # define const_debug static const | |
664 | #endif | |
665 | ||
666 | /** | |
667 | * runqueue_is_locked | |
668 | * @cpu: the processor in question. | |
669 | * | |
670 | * Returns true if the current cpu runqueue is locked. | |
671 | * This interface allows printk to be called with the runqueue lock | |
672 | * held and know whether or not it is OK to wake up the klogd. | |
673 | */ | |
674 | int runqueue_is_locked(int cpu) | |
675 | { | |
676 | return raw_spin_is_locked(&cpu_rq(cpu)->lock); | |
677 | } | |
678 | ||
679 | /* | |
680 | * Debugging: various feature bits | |
681 | */ | |
682 | ||
683 | #define SCHED_FEAT(name, enabled) \ | |
684 | __SCHED_FEAT_##name , | |
685 | ||
686 | enum { | |
687 | #include "sched_features.h" | |
688 | }; | |
689 | ||
690 | #undef SCHED_FEAT | |
691 | ||
692 | #define SCHED_FEAT(name, enabled) \ | |
693 | (1UL << __SCHED_FEAT_##name) * enabled | | |
694 | ||
695 | const_debug unsigned int sysctl_sched_features = | |
696 | #include "sched_features.h" | |
697 | 0; | |
698 | ||
699 | #undef SCHED_FEAT | |
700 | ||
701 | #ifdef CONFIG_SCHED_DEBUG | |
702 | #define SCHED_FEAT(name, enabled) \ | |
703 | #name , | |
704 | ||
705 | static __read_mostly char *sched_feat_names[] = { | |
706 | #include "sched_features.h" | |
707 | NULL | |
708 | }; | |
709 | ||
710 | #undef SCHED_FEAT | |
711 | ||
712 | static int sched_feat_show(struct seq_file *m, void *v) | |
713 | { | |
714 | int i; | |
715 | ||
716 | for (i = 0; sched_feat_names[i]; i++) { | |
717 | if (!(sysctl_sched_features & (1UL << i))) | |
718 | seq_puts(m, "NO_"); | |
719 | seq_printf(m, "%s ", sched_feat_names[i]); | |
720 | } | |
721 | seq_puts(m, "\n"); | |
722 | ||
723 | return 0; | |
724 | } | |
725 | ||
726 | static ssize_t | |
727 | sched_feat_write(struct file *filp, const char __user *ubuf, | |
728 | size_t cnt, loff_t *ppos) | |
729 | { | |
730 | char buf[64]; | |
731 | char *cmp; | |
732 | int neg = 0; | |
733 | int i; | |
734 | ||
735 | if (cnt > 63) | |
736 | cnt = 63; | |
737 | ||
738 | if (copy_from_user(&buf, ubuf, cnt)) | |
739 | return -EFAULT; | |
740 | ||
741 | buf[cnt] = 0; | |
742 | cmp = strstrip(buf); | |
743 | ||
744 | if (strncmp(buf, "NO_", 3) == 0) { | |
745 | neg = 1; | |
746 | cmp += 3; | |
747 | } | |
748 | ||
749 | for (i = 0; sched_feat_names[i]; i++) { | |
750 | if (strcmp(cmp, sched_feat_names[i]) == 0) { | |
751 | if (neg) | |
752 | sysctl_sched_features &= ~(1UL << i); | |
753 | else | |
754 | sysctl_sched_features |= (1UL << i); | |
755 | break; | |
756 | } | |
757 | } | |
758 | ||
759 | if (!sched_feat_names[i]) | |
760 | return -EINVAL; | |
761 | ||
762 | *ppos += cnt; | |
763 | ||
764 | return cnt; | |
765 | } | |
766 | ||
767 | static int sched_feat_open(struct inode *inode, struct file *filp) | |
768 | { | |
769 | return single_open(filp, sched_feat_show, NULL); | |
770 | } | |
771 | ||
772 | static const struct file_operations sched_feat_fops = { | |
773 | .open = sched_feat_open, | |
774 | .write = sched_feat_write, | |
775 | .read = seq_read, | |
776 | .llseek = seq_lseek, | |
777 | .release = single_release, | |
778 | }; | |
779 | ||
780 | static __init int sched_init_debug(void) | |
781 | { | |
782 | debugfs_create_file("sched_features", 0644, NULL, NULL, | |
783 | &sched_feat_fops); | |
784 | ||
785 | return 0; | |
786 | } | |
787 | late_initcall(sched_init_debug); | |
788 | ||
789 | #endif | |
790 | ||
791 | #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x)) | |
792 | ||
793 | /* | |
794 | * Number of tasks to iterate in a single balance run. | |
795 | * Limited because this is done with IRQs disabled. | |
796 | */ | |
797 | const_debug unsigned int sysctl_sched_nr_migrate = 32; | |
798 | ||
799 | /* | |
800 | * ratelimit for updating the group shares. | |
801 | * default: 0.25ms | |
802 | */ | |
803 | unsigned int sysctl_sched_shares_ratelimit = 250000; | |
804 | unsigned int normalized_sysctl_sched_shares_ratelimit = 250000; | |
805 | ||
806 | /* | |
807 | * Inject some fuzzyness into changing the per-cpu group shares | |
808 | * this avoids remote rq-locks at the expense of fairness. | |
809 | * default: 4 | |
810 | */ | |
811 | unsigned int sysctl_sched_shares_thresh = 4; | |
812 | ||
813 | /* | |
814 | * period over which we average the RT time consumption, measured | |
815 | * in ms. | |
816 | * | |
817 | * default: 1s | |
818 | */ | |
819 | const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; | |
820 | ||
821 | /* | |
822 | * period over which we measure -rt task cpu usage in us. | |
823 | * default: 1s | |
824 | */ | |
825 | unsigned int sysctl_sched_rt_period = 1000000; | |
826 | ||
827 | static __read_mostly int scheduler_running; | |
828 | ||
829 | /* | |
830 | * part of the period that we allow rt tasks to run in us. | |
831 | * default: 0.95s | |
832 | */ | |
833 | int sysctl_sched_rt_runtime = 950000; | |
834 | ||
835 | static inline u64 global_rt_period(void) | |
836 | { | |
837 | return (u64)sysctl_sched_rt_period * NSEC_PER_USEC; | |
838 | } | |
839 | ||
840 | static inline u64 global_rt_runtime(void) | |
841 | { | |
842 | if (sysctl_sched_rt_runtime < 0) | |
843 | return RUNTIME_INF; | |
844 | ||
845 | return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC; | |
846 | } | |
847 | ||
848 | #ifndef prepare_arch_switch | |
849 | # define prepare_arch_switch(next) do { } while (0) | |
850 | #endif | |
851 | #ifndef finish_arch_switch | |
852 | # define finish_arch_switch(prev) do { } while (0) | |
853 | #endif | |
854 | ||
855 | static inline int task_current(struct rq *rq, struct task_struct *p) | |
856 | { | |
857 | return rq->curr == p; | |
858 | } | |
859 | ||
860 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW | |
861 | static inline int task_running(struct rq *rq, struct task_struct *p) | |
862 | { | |
863 | return task_current(rq, p); | |
864 | } | |
865 | ||
866 | static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) | |
867 | { | |
868 | } | |
869 | ||
870 | static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) | |
871 | { | |
872 | #ifdef CONFIG_DEBUG_SPINLOCK | |
873 | /* this is a valid case when another task releases the spinlock */ | |
874 | rq->lock.owner = current; | |
875 | #endif | |
876 | /* | |
877 | * If we are tracking spinlock dependencies then we have to | |
878 | * fix up the runqueue lock - which gets 'carried over' from | |
879 | * prev into current: | |
880 | */ | |
881 | spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_); | |
882 | ||
883 | raw_spin_unlock_irq(&rq->lock); | |
884 | } | |
885 | ||
886 | #else /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
887 | static inline int task_running(struct rq *rq, struct task_struct *p) | |
888 | { | |
889 | #ifdef CONFIG_SMP | |
890 | return p->oncpu; | |
891 | #else | |
892 | return task_current(rq, p); | |
893 | #endif | |
894 | } | |
895 | ||
896 | static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) | |
897 | { | |
898 | #ifdef CONFIG_SMP | |
899 | /* | |
900 | * We can optimise this out completely for !SMP, because the | |
901 | * SMP rebalancing from interrupt is the only thing that cares | |
902 | * here. | |
903 | */ | |
904 | next->oncpu = 1; | |
905 | #endif | |
906 | #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
907 | raw_spin_unlock_irq(&rq->lock); | |
908 | #else | |
909 | raw_spin_unlock(&rq->lock); | |
910 | #endif | |
911 | } | |
912 | ||
913 | static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) | |
914 | { | |
915 | #ifdef CONFIG_SMP | |
916 | /* | |
917 | * After ->oncpu is cleared, the task can be moved to a different CPU. | |
918 | * We must ensure this doesn't happen until the switch is completely | |
919 | * finished. | |
920 | */ | |
921 | smp_wmb(); | |
922 | prev->oncpu = 0; | |
923 | #endif | |
924 | #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
925 | local_irq_enable(); | |
926 | #endif | |
927 | } | |
928 | #endif /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
929 | ||
930 | /* | |
931 | * Check whether the task is waking, we use this to synchronize ->cpus_allowed | |
932 | * against ttwu(). | |
933 | */ | |
934 | static inline int task_is_waking(struct task_struct *p) | |
935 | { | |
936 | return unlikely(p->state == TASK_WAKING); | |
937 | } | |
938 | ||
939 | /* | |
940 | * __task_rq_lock - lock the runqueue a given task resides on. | |
941 | * Must be called interrupts disabled. | |
942 | */ | |
943 | static inline struct rq *__task_rq_lock(struct task_struct *p) | |
944 | __acquires(rq->lock) | |
945 | { | |
946 | struct rq *rq; | |
947 | ||
948 | for (;;) { | |
949 | rq = task_rq(p); | |
950 | raw_spin_lock(&rq->lock); | |
951 | if (likely(rq == task_rq(p))) | |
952 | return rq; | |
953 | raw_spin_unlock(&rq->lock); | |
954 | } | |
955 | } | |
956 | ||
957 | /* | |
958 | * task_rq_lock - lock the runqueue a given task resides on and disable | |
959 | * interrupts. Note the ordering: we can safely lookup the task_rq without | |
960 | * explicitly disabling preemption. | |
961 | */ | |
962 | static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) | |
963 | __acquires(rq->lock) | |
964 | { | |
965 | struct rq *rq; | |
966 | ||
967 | for (;;) { | |
968 | local_irq_save(*flags); | |
969 | rq = task_rq(p); | |
970 | raw_spin_lock(&rq->lock); | |
971 | if (likely(rq == task_rq(p))) | |
972 | return rq; | |
973 | raw_spin_unlock_irqrestore(&rq->lock, *flags); | |
974 | } | |
975 | } | |
976 | ||
977 | static void __task_rq_unlock(struct rq *rq) | |
978 | __releases(rq->lock) | |
979 | { | |
980 | raw_spin_unlock(&rq->lock); | |
981 | } | |
982 | ||
983 | static inline void task_rq_unlock(struct rq *rq, unsigned long *flags) | |
984 | __releases(rq->lock) | |
985 | { | |
986 | raw_spin_unlock_irqrestore(&rq->lock, *flags); | |
987 | } | |
988 | ||
989 | /* | |
990 | * this_rq_lock - lock this runqueue and disable interrupts. | |
991 | */ | |
992 | static struct rq *this_rq_lock(void) | |
993 | __acquires(rq->lock) | |
994 | { | |
995 | struct rq *rq; | |
996 | ||
997 | local_irq_disable(); | |
998 | rq = this_rq(); | |
999 | raw_spin_lock(&rq->lock); | |
1000 | ||
1001 | return rq; | |
1002 | } | |
1003 | ||
1004 | #ifdef CONFIG_SCHED_HRTICK | |
1005 | /* | |
1006 | * Use HR-timers to deliver accurate preemption points. | |
1007 | * | |
1008 | * Its all a bit involved since we cannot program an hrt while holding the | |
1009 | * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a | |
1010 | * reschedule event. | |
1011 | * | |
1012 | * When we get rescheduled we reprogram the hrtick_timer outside of the | |
1013 | * rq->lock. | |
1014 | */ | |
1015 | ||
1016 | /* | |
1017 | * Use hrtick when: | |
1018 | * - enabled by features | |
1019 | * - hrtimer is actually high res | |
1020 | */ | |
1021 | static inline int hrtick_enabled(struct rq *rq) | |
1022 | { | |
1023 | if (!sched_feat(HRTICK)) | |
1024 | return 0; | |
1025 | if (!cpu_active(cpu_of(rq))) | |
1026 | return 0; | |
1027 | return hrtimer_is_hres_active(&rq->hrtick_timer); | |
1028 | } | |
1029 | ||
1030 | static void hrtick_clear(struct rq *rq) | |
1031 | { | |
1032 | if (hrtimer_active(&rq->hrtick_timer)) | |
1033 | hrtimer_cancel(&rq->hrtick_timer); | |
1034 | } | |
1035 | ||
1036 | /* | |
1037 | * High-resolution timer tick. | |
1038 | * Runs from hardirq context with interrupts disabled. | |
1039 | */ | |
1040 | static enum hrtimer_restart hrtick(struct hrtimer *timer) | |
1041 | { | |
1042 | struct rq *rq = container_of(timer, struct rq, hrtick_timer); | |
1043 | ||
1044 | WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); | |
1045 | ||
1046 | raw_spin_lock(&rq->lock); | |
1047 | update_rq_clock(rq); | |
1048 | rq->curr->sched_class->task_tick(rq, rq->curr, 1); | |
1049 | raw_spin_unlock(&rq->lock); | |
1050 | ||
1051 | return HRTIMER_NORESTART; | |
1052 | } | |
1053 | ||
1054 | #ifdef CONFIG_SMP | |
1055 | /* | |
1056 | * called from hardirq (IPI) context | |
1057 | */ | |
1058 | static void __hrtick_start(void *arg) | |
1059 | { | |
1060 | struct rq *rq = arg; | |
1061 | ||
1062 | raw_spin_lock(&rq->lock); | |
1063 | hrtimer_restart(&rq->hrtick_timer); | |
1064 | rq->hrtick_csd_pending = 0; | |
1065 | raw_spin_unlock(&rq->lock); | |
1066 | } | |
1067 | ||
1068 | /* | |
1069 | * Called to set the hrtick timer state. | |
1070 | * | |
1071 | * called with rq->lock held and irqs disabled | |
1072 | */ | |
1073 | static void hrtick_start(struct rq *rq, u64 delay) | |
1074 | { | |
1075 | struct hrtimer *timer = &rq->hrtick_timer; | |
1076 | ktime_t time = ktime_add_ns(timer->base->get_time(), delay); | |
1077 | ||
1078 | hrtimer_set_expires(timer, time); | |
1079 | ||
1080 | if (rq == this_rq()) { | |
1081 | hrtimer_restart(timer); | |
1082 | } else if (!rq->hrtick_csd_pending) { | |
1083 | __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0); | |
1084 | rq->hrtick_csd_pending = 1; | |
1085 | } | |
1086 | } | |
1087 | ||
1088 | static int | |
1089 | hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) | |
1090 | { | |
1091 | int cpu = (int)(long)hcpu; | |
1092 | ||
1093 | switch (action) { | |
1094 | case CPU_UP_CANCELED: | |
1095 | case CPU_UP_CANCELED_FROZEN: | |
1096 | case CPU_DOWN_PREPARE: | |
1097 | case CPU_DOWN_PREPARE_FROZEN: | |
1098 | case CPU_DEAD: | |
1099 | case CPU_DEAD_FROZEN: | |
1100 | hrtick_clear(cpu_rq(cpu)); | |
1101 | return NOTIFY_OK; | |
1102 | } | |
1103 | ||
1104 | return NOTIFY_DONE; | |
1105 | } | |
1106 | ||
1107 | static __init void init_hrtick(void) | |
1108 | { | |
1109 | hotcpu_notifier(hotplug_hrtick, 0); | |
1110 | } | |
1111 | #else | |
1112 | /* | |
1113 | * Called to set the hrtick timer state. | |
1114 | * | |
1115 | * called with rq->lock held and irqs disabled | |
1116 | */ | |
1117 | static void hrtick_start(struct rq *rq, u64 delay) | |
1118 | { | |
1119 | __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, | |
1120 | HRTIMER_MODE_REL_PINNED, 0); | |
1121 | } | |
1122 | ||
1123 | static inline void init_hrtick(void) | |
1124 | { | |
1125 | } | |
1126 | #endif /* CONFIG_SMP */ | |
1127 | ||
1128 | static void init_rq_hrtick(struct rq *rq) | |
1129 | { | |
1130 | #ifdef CONFIG_SMP | |
1131 | rq->hrtick_csd_pending = 0; | |
1132 | ||
1133 | rq->hrtick_csd.flags = 0; | |
1134 | rq->hrtick_csd.func = __hrtick_start; | |
1135 | rq->hrtick_csd.info = rq; | |
1136 | #endif | |
1137 | ||
1138 | hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
1139 | rq->hrtick_timer.function = hrtick; | |
1140 | } | |
1141 | #else /* CONFIG_SCHED_HRTICK */ | |
1142 | static inline void hrtick_clear(struct rq *rq) | |
1143 | { | |
1144 | } | |
1145 | ||
1146 | static inline void init_rq_hrtick(struct rq *rq) | |
1147 | { | |
1148 | } | |
1149 | ||
1150 | static inline void init_hrtick(void) | |
1151 | { | |
1152 | } | |
1153 | #endif /* CONFIG_SCHED_HRTICK */ | |
1154 | ||
1155 | /* | |
1156 | * resched_task - mark a task 'to be rescheduled now'. | |
1157 | * | |
1158 | * On UP this means the setting of the need_resched flag, on SMP it | |
1159 | * might also involve a cross-CPU call to trigger the scheduler on | |
1160 | * the target CPU. | |
1161 | */ | |
1162 | #ifdef CONFIG_SMP | |
1163 | ||
1164 | #ifndef tsk_is_polling | |
1165 | #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG) | |
1166 | #endif | |
1167 | ||
1168 | static void resched_task(struct task_struct *p) | |
1169 | { | |
1170 | int cpu; | |
1171 | ||
1172 | assert_raw_spin_locked(&task_rq(p)->lock); | |
1173 | ||
1174 | if (test_tsk_need_resched(p)) | |
1175 | return; | |
1176 | ||
1177 | set_tsk_need_resched(p); | |
1178 | ||
1179 | cpu = task_cpu(p); | |
1180 | if (cpu == smp_processor_id()) | |
1181 | return; | |
1182 | ||
1183 | /* NEED_RESCHED must be visible before we test polling */ | |
1184 | smp_mb(); | |
1185 | if (!tsk_is_polling(p)) | |
1186 | smp_send_reschedule(cpu); | |
1187 | } | |
1188 | ||
1189 | static void resched_cpu(int cpu) | |
1190 | { | |
1191 | struct rq *rq = cpu_rq(cpu); | |
1192 | unsigned long flags; | |
1193 | ||
1194 | if (!raw_spin_trylock_irqsave(&rq->lock, flags)) | |
1195 | return; | |
1196 | resched_task(cpu_curr(cpu)); | |
1197 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
1198 | } | |
1199 | ||
1200 | #ifdef CONFIG_NO_HZ | |
1201 | /* | |
1202 | * In the semi idle case, use the nearest busy cpu for migrating timers | |
1203 | * from an idle cpu. This is good for power-savings. | |
1204 | * | |
1205 | * We don't do similar optimization for completely idle system, as | |
1206 | * selecting an idle cpu will add more delays to the timers than intended | |
1207 | * (as that cpu's timer base may not be uptodate wrt jiffies etc). | |
1208 | */ | |
1209 | int get_nohz_timer_target(void) | |
1210 | { | |
1211 | int cpu = smp_processor_id(); | |
1212 | int i; | |
1213 | struct sched_domain *sd; | |
1214 | ||
1215 | for_each_domain(cpu, sd) { | |
1216 | for_each_cpu(i, sched_domain_span(sd)) | |
1217 | if (!idle_cpu(i)) | |
1218 | return i; | |
1219 | } | |
1220 | return cpu; | |
1221 | } | |
1222 | /* | |
1223 | * When add_timer_on() enqueues a timer into the timer wheel of an | |
1224 | * idle CPU then this timer might expire before the next timer event | |
1225 | * which is scheduled to wake up that CPU. In case of a completely | |
1226 | * idle system the next event might even be infinite time into the | |
1227 | * future. wake_up_idle_cpu() ensures that the CPU is woken up and | |
1228 | * leaves the inner idle loop so the newly added timer is taken into | |
1229 | * account when the CPU goes back to idle and evaluates the timer | |
1230 | * wheel for the next timer event. | |
1231 | */ | |
1232 | void wake_up_idle_cpu(int cpu) | |
1233 | { | |
1234 | struct rq *rq = cpu_rq(cpu); | |
1235 | ||
1236 | if (cpu == smp_processor_id()) | |
1237 | return; | |
1238 | ||
1239 | /* | |
1240 | * This is safe, as this function is called with the timer | |
1241 | * wheel base lock of (cpu) held. When the CPU is on the way | |
1242 | * to idle and has not yet set rq->curr to idle then it will | |
1243 | * be serialized on the timer wheel base lock and take the new | |
1244 | * timer into account automatically. | |
1245 | */ | |
1246 | if (rq->curr != rq->idle) | |
1247 | return; | |
1248 | ||
1249 | /* | |
1250 | * We can set TIF_RESCHED on the idle task of the other CPU | |
1251 | * lockless. The worst case is that the other CPU runs the | |
1252 | * idle task through an additional NOOP schedule() | |
1253 | */ | |
1254 | set_tsk_need_resched(rq->idle); | |
1255 | ||
1256 | /* NEED_RESCHED must be visible before we test polling */ | |
1257 | smp_mb(); | |
1258 | if (!tsk_is_polling(rq->idle)) | |
1259 | smp_send_reschedule(cpu); | |
1260 | } | |
1261 | ||
1262 | #endif /* CONFIG_NO_HZ */ | |
1263 | ||
1264 | static u64 sched_avg_period(void) | |
1265 | { | |
1266 | return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2; | |
1267 | } | |
1268 | ||
1269 | static void sched_avg_update(struct rq *rq) | |
1270 | { | |
1271 | s64 period = sched_avg_period(); | |
1272 | ||
1273 | while ((s64)(rq->clock - rq->age_stamp) > period) { | |
1274 | /* | |
1275 | * Inline assembly required to prevent the compiler | |
1276 | * optimising this loop into a divmod call. | |
1277 | * See __iter_div_u64_rem() for another example of this. | |
1278 | */ | |
1279 | asm("" : "+rm" (rq->age_stamp)); | |
1280 | rq->age_stamp += period; | |
1281 | rq->rt_avg /= 2; | |
1282 | } | |
1283 | } | |
1284 | ||
1285 | static void sched_rt_avg_update(struct rq *rq, u64 rt_delta) | |
1286 | { | |
1287 | rq->rt_avg += rt_delta; | |
1288 | sched_avg_update(rq); | |
1289 | } | |
1290 | ||
1291 | #else /* !CONFIG_SMP */ | |
1292 | static void resched_task(struct task_struct *p) | |
1293 | { | |
1294 | assert_raw_spin_locked(&task_rq(p)->lock); | |
1295 | set_tsk_need_resched(p); | |
1296 | } | |
1297 | ||
1298 | static void sched_rt_avg_update(struct rq *rq, u64 rt_delta) | |
1299 | { | |
1300 | } | |
1301 | ||
1302 | static void sched_avg_update(struct rq *rq) | |
1303 | { | |
1304 | } | |
1305 | #endif /* CONFIG_SMP */ | |
1306 | ||
1307 | #if BITS_PER_LONG == 32 | |
1308 | # define WMULT_CONST (~0UL) | |
1309 | #else | |
1310 | # define WMULT_CONST (1UL << 32) | |
1311 | #endif | |
1312 | ||
1313 | #define WMULT_SHIFT 32 | |
1314 | ||
1315 | /* | |
1316 | * Shift right and round: | |
1317 | */ | |
1318 | #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y)) | |
1319 | ||
1320 | /* | |
1321 | * delta *= weight / lw | |
1322 | */ | |
1323 | static unsigned long | |
1324 | calc_delta_mine(unsigned long delta_exec, unsigned long weight, | |
1325 | struct load_weight *lw) | |
1326 | { | |
1327 | u64 tmp; | |
1328 | ||
1329 | if (!lw->inv_weight) { | |
1330 | if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST)) | |
1331 | lw->inv_weight = 1; | |
1332 | else | |
1333 | lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2) | |
1334 | / (lw->weight+1); | |
1335 | } | |
1336 | ||
1337 | tmp = (u64)delta_exec * weight; | |
1338 | /* | |
1339 | * Check whether we'd overflow the 64-bit multiplication: | |
1340 | */ | |
1341 | if (unlikely(tmp > WMULT_CONST)) | |
1342 | tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight, | |
1343 | WMULT_SHIFT/2); | |
1344 | else | |
1345 | tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT); | |
1346 | ||
1347 | return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX); | |
1348 | } | |
1349 | ||
1350 | static inline void update_load_add(struct load_weight *lw, unsigned long inc) | |
1351 | { | |
1352 | lw->weight += inc; | |
1353 | lw->inv_weight = 0; | |
1354 | } | |
1355 | ||
1356 | static inline void update_load_sub(struct load_weight *lw, unsigned long dec) | |
1357 | { | |
1358 | lw->weight -= dec; | |
1359 | lw->inv_weight = 0; | |
1360 | } | |
1361 | ||
1362 | /* | |
1363 | * To aid in avoiding the subversion of "niceness" due to uneven distribution | |
1364 | * of tasks with abnormal "nice" values across CPUs the contribution that | |
1365 | * each task makes to its run queue's load is weighted according to its | |
1366 | * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a | |
1367 | * scaled version of the new time slice allocation that they receive on time | |
1368 | * slice expiry etc. | |
1369 | */ | |
1370 | ||
1371 | #define WEIGHT_IDLEPRIO 3 | |
1372 | #define WMULT_IDLEPRIO 1431655765 | |
1373 | ||
1374 | /* | |
1375 | * Nice levels are multiplicative, with a gentle 10% change for every | |
1376 | * nice level changed. I.e. when a CPU-bound task goes from nice 0 to | |
1377 | * nice 1, it will get ~10% less CPU time than another CPU-bound task | |
1378 | * that remained on nice 0. | |
1379 | * | |
1380 | * The "10% effect" is relative and cumulative: from _any_ nice level, | |
1381 | * if you go up 1 level, it's -10% CPU usage, if you go down 1 level | |
1382 | * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25. | |
1383 | * If a task goes up by ~10% and another task goes down by ~10% then | |
1384 | * the relative distance between them is ~25%.) | |
1385 | */ | |
1386 | static const int prio_to_weight[40] = { | |
1387 | /* -20 */ 88761, 71755, 56483, 46273, 36291, | |
1388 | /* -15 */ 29154, 23254, 18705, 14949, 11916, | |
1389 | /* -10 */ 9548, 7620, 6100, 4904, 3906, | |
1390 | /* -5 */ 3121, 2501, 1991, 1586, 1277, | |
1391 | /* 0 */ 1024, 820, 655, 526, 423, | |
1392 | /* 5 */ 335, 272, 215, 172, 137, | |
1393 | /* 10 */ 110, 87, 70, 56, 45, | |
1394 | /* 15 */ 36, 29, 23, 18, 15, | |
1395 | }; | |
1396 | ||
1397 | /* | |
1398 | * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated. | |
1399 | * | |
1400 | * In cases where the weight does not change often, we can use the | |
1401 | * precalculated inverse to speed up arithmetics by turning divisions | |
1402 | * into multiplications: | |
1403 | */ | |
1404 | static const u32 prio_to_wmult[40] = { | |
1405 | /* -20 */ 48388, 59856, 76040, 92818, 118348, | |
1406 | /* -15 */ 147320, 184698, 229616, 287308, 360437, | |
1407 | /* -10 */ 449829, 563644, 704093, 875809, 1099582, | |
1408 | /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326, | |
1409 | /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587, | |
1410 | /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126, | |
1411 | /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717, | |
1412 | /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153, | |
1413 | }; | |
1414 | ||
1415 | /* Time spent by the tasks of the cpu accounting group executing in ... */ | |
1416 | enum cpuacct_stat_index { | |
1417 | CPUACCT_STAT_USER, /* ... user mode */ | |
1418 | CPUACCT_STAT_SYSTEM, /* ... kernel mode */ | |
1419 | ||
1420 | CPUACCT_STAT_NSTATS, | |
1421 | }; | |
1422 | ||
1423 | #ifdef CONFIG_CGROUP_CPUACCT | |
1424 | static void cpuacct_charge(struct task_struct *tsk, u64 cputime); | |
1425 | static void cpuacct_update_stats(struct task_struct *tsk, | |
1426 | enum cpuacct_stat_index idx, cputime_t val); | |
1427 | #else | |
1428 | static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {} | |
1429 | static inline void cpuacct_update_stats(struct task_struct *tsk, | |
1430 | enum cpuacct_stat_index idx, cputime_t val) {} | |
1431 | #endif | |
1432 | ||
1433 | static inline void inc_cpu_load(struct rq *rq, unsigned long load) | |
1434 | { | |
1435 | update_load_add(&rq->load, load); | |
1436 | } | |
1437 | ||
1438 | static inline void dec_cpu_load(struct rq *rq, unsigned long load) | |
1439 | { | |
1440 | update_load_sub(&rq->load, load); | |
1441 | } | |
1442 | ||
1443 | #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED) | |
1444 | typedef int (*tg_visitor)(struct task_group *, void *); | |
1445 | ||
1446 | /* | |
1447 | * Iterate the full tree, calling @down when first entering a node and @up when | |
1448 | * leaving it for the final time. | |
1449 | */ | |
1450 | static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data) | |
1451 | { | |
1452 | struct task_group *parent, *child; | |
1453 | int ret; | |
1454 | ||
1455 | rcu_read_lock(); | |
1456 | parent = &root_task_group; | |
1457 | down: | |
1458 | ret = (*down)(parent, data); | |
1459 | if (ret) | |
1460 | goto out_unlock; | |
1461 | list_for_each_entry_rcu(child, &parent->children, siblings) { | |
1462 | parent = child; | |
1463 | goto down; | |
1464 | ||
1465 | up: | |
1466 | continue; | |
1467 | } | |
1468 | ret = (*up)(parent, data); | |
1469 | if (ret) | |
1470 | goto out_unlock; | |
1471 | ||
1472 | child = parent; | |
1473 | parent = parent->parent; | |
1474 | if (parent) | |
1475 | goto up; | |
1476 | out_unlock: | |
1477 | rcu_read_unlock(); | |
1478 | ||
1479 | return ret; | |
1480 | } | |
1481 | ||
1482 | static int tg_nop(struct task_group *tg, void *data) | |
1483 | { | |
1484 | return 0; | |
1485 | } | |
1486 | #endif | |
1487 | ||
1488 | #ifdef CONFIG_SMP | |
1489 | /* Used instead of source_load when we know the type == 0 */ | |
1490 | static unsigned long weighted_cpuload(const int cpu) | |
1491 | { | |
1492 | return cpu_rq(cpu)->load.weight; | |
1493 | } | |
1494 | ||
1495 | /* | |
1496 | * Return a low guess at the load of a migration-source cpu weighted | |
1497 | * according to the scheduling class and "nice" value. | |
1498 | * | |
1499 | * We want to under-estimate the load of migration sources, to | |
1500 | * balance conservatively. | |
1501 | */ | |
1502 | static unsigned long source_load(int cpu, int type) | |
1503 | { | |
1504 | struct rq *rq = cpu_rq(cpu); | |
1505 | unsigned long total = weighted_cpuload(cpu); | |
1506 | ||
1507 | if (type == 0 || !sched_feat(LB_BIAS)) | |
1508 | return total; | |
1509 | ||
1510 | return min(rq->cpu_load[type-1], total); | |
1511 | } | |
1512 | ||
1513 | /* | |
1514 | * Return a high guess at the load of a migration-target cpu weighted | |
1515 | * according to the scheduling class and "nice" value. | |
1516 | */ | |
1517 | static unsigned long target_load(int cpu, int type) | |
1518 | { | |
1519 | struct rq *rq = cpu_rq(cpu); | |
1520 | unsigned long total = weighted_cpuload(cpu); | |
1521 | ||
1522 | if (type == 0 || !sched_feat(LB_BIAS)) | |
1523 | return total; | |
1524 | ||
1525 | return max(rq->cpu_load[type-1], total); | |
1526 | } | |
1527 | ||
1528 | static unsigned long power_of(int cpu) | |
1529 | { | |
1530 | return cpu_rq(cpu)->cpu_power; | |
1531 | } | |
1532 | ||
1533 | static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd); | |
1534 | ||
1535 | static unsigned long cpu_avg_load_per_task(int cpu) | |
1536 | { | |
1537 | struct rq *rq = cpu_rq(cpu); | |
1538 | unsigned long nr_running = ACCESS_ONCE(rq->nr_running); | |
1539 | ||
1540 | if (nr_running) | |
1541 | rq->avg_load_per_task = rq->load.weight / nr_running; | |
1542 | else | |
1543 | rq->avg_load_per_task = 0; | |
1544 | ||
1545 | return rq->avg_load_per_task; | |
1546 | } | |
1547 | ||
1548 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1549 | ||
1550 | static __read_mostly unsigned long __percpu *update_shares_data; | |
1551 | ||
1552 | static void __set_se_shares(struct sched_entity *se, unsigned long shares); | |
1553 | ||
1554 | /* | |
1555 | * Calculate and set the cpu's group shares. | |
1556 | */ | |
1557 | static void update_group_shares_cpu(struct task_group *tg, int cpu, | |
1558 | unsigned long sd_shares, | |
1559 | unsigned long sd_rq_weight, | |
1560 | unsigned long *usd_rq_weight) | |
1561 | { | |
1562 | unsigned long shares, rq_weight; | |
1563 | int boost = 0; | |
1564 | ||
1565 | rq_weight = usd_rq_weight[cpu]; | |
1566 | if (!rq_weight) { | |
1567 | boost = 1; | |
1568 | rq_weight = NICE_0_LOAD; | |
1569 | } | |
1570 | ||
1571 | /* | |
1572 | * \Sum_j shares_j * rq_weight_i | |
1573 | * shares_i = ----------------------------- | |
1574 | * \Sum_j rq_weight_j | |
1575 | */ | |
1576 | shares = (sd_shares * rq_weight) / sd_rq_weight; | |
1577 | shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES); | |
1578 | ||
1579 | if (abs(shares - tg->se[cpu]->load.weight) > | |
1580 | sysctl_sched_shares_thresh) { | |
1581 | struct rq *rq = cpu_rq(cpu); | |
1582 | unsigned long flags; | |
1583 | ||
1584 | raw_spin_lock_irqsave(&rq->lock, flags); | |
1585 | tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight; | |
1586 | tg->cfs_rq[cpu]->shares = boost ? 0 : shares; | |
1587 | __set_se_shares(tg->se[cpu], shares); | |
1588 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
1589 | } | |
1590 | } | |
1591 | ||
1592 | /* | |
1593 | * Re-compute the task group their per cpu shares over the given domain. | |
1594 | * This needs to be done in a bottom-up fashion because the rq weight of a | |
1595 | * parent group depends on the shares of its child groups. | |
1596 | */ | |
1597 | static int tg_shares_up(struct task_group *tg, void *data) | |
1598 | { | |
1599 | unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0; | |
1600 | unsigned long *usd_rq_weight; | |
1601 | struct sched_domain *sd = data; | |
1602 | unsigned long flags; | |
1603 | int i; | |
1604 | ||
1605 | if (!tg->se[0]) | |
1606 | return 0; | |
1607 | ||
1608 | local_irq_save(flags); | |
1609 | usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id()); | |
1610 | ||
1611 | for_each_cpu(i, sched_domain_span(sd)) { | |
1612 | weight = tg->cfs_rq[i]->load.weight; | |
1613 | usd_rq_weight[i] = weight; | |
1614 | ||
1615 | rq_weight += weight; | |
1616 | /* | |
1617 | * If there are currently no tasks on the cpu pretend there | |
1618 | * is one of average load so that when a new task gets to | |
1619 | * run here it will not get delayed by group starvation. | |
1620 | */ | |
1621 | if (!weight) | |
1622 | weight = NICE_0_LOAD; | |
1623 | ||
1624 | sum_weight += weight; | |
1625 | shares += tg->cfs_rq[i]->shares; | |
1626 | } | |
1627 | ||
1628 | if (!rq_weight) | |
1629 | rq_weight = sum_weight; | |
1630 | ||
1631 | if ((!shares && rq_weight) || shares > tg->shares) | |
1632 | shares = tg->shares; | |
1633 | ||
1634 | if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE)) | |
1635 | shares = tg->shares; | |
1636 | ||
1637 | for_each_cpu(i, sched_domain_span(sd)) | |
1638 | update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight); | |
1639 | ||
1640 | local_irq_restore(flags); | |
1641 | ||
1642 | return 0; | |
1643 | } | |
1644 | ||
1645 | /* | |
1646 | * Compute the cpu's hierarchical load factor for each task group. | |
1647 | * This needs to be done in a top-down fashion because the load of a child | |
1648 | * group is a fraction of its parents load. | |
1649 | */ | |
1650 | static int tg_load_down(struct task_group *tg, void *data) | |
1651 | { | |
1652 | unsigned long load; | |
1653 | long cpu = (long)data; | |
1654 | ||
1655 | if (!tg->parent) { | |
1656 | load = cpu_rq(cpu)->load.weight; | |
1657 | } else { | |
1658 | load = tg->parent->cfs_rq[cpu]->h_load; | |
1659 | load *= tg->cfs_rq[cpu]->shares; | |
1660 | load /= tg->parent->cfs_rq[cpu]->load.weight + 1; | |
1661 | } | |
1662 | ||
1663 | tg->cfs_rq[cpu]->h_load = load; | |
1664 | ||
1665 | return 0; | |
1666 | } | |
1667 | ||
1668 | static void update_shares(struct sched_domain *sd) | |
1669 | { | |
1670 | s64 elapsed; | |
1671 | u64 now; | |
1672 | ||
1673 | if (root_task_group_empty()) | |
1674 | return; | |
1675 | ||
1676 | now = local_clock(); | |
1677 | elapsed = now - sd->last_update; | |
1678 | ||
1679 | if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) { | |
1680 | sd->last_update = now; | |
1681 | walk_tg_tree(tg_nop, tg_shares_up, sd); | |
1682 | } | |
1683 | } | |
1684 | ||
1685 | static void update_h_load(long cpu) | |
1686 | { | |
1687 | walk_tg_tree(tg_load_down, tg_nop, (void *)cpu); | |
1688 | } | |
1689 | ||
1690 | #else | |
1691 | ||
1692 | static inline void update_shares(struct sched_domain *sd) | |
1693 | { | |
1694 | } | |
1695 | ||
1696 | #endif | |
1697 | ||
1698 | #ifdef CONFIG_PREEMPT | |
1699 | ||
1700 | static void double_rq_lock(struct rq *rq1, struct rq *rq2); | |
1701 | ||
1702 | /* | |
1703 | * fair double_lock_balance: Safely acquires both rq->locks in a fair | |
1704 | * way at the expense of forcing extra atomic operations in all | |
1705 | * invocations. This assures that the double_lock is acquired using the | |
1706 | * same underlying policy as the spinlock_t on this architecture, which | |
1707 | * reduces latency compared to the unfair variant below. However, it | |
1708 | * also adds more overhead and therefore may reduce throughput. | |
1709 | */ | |
1710 | static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest) | |
1711 | __releases(this_rq->lock) | |
1712 | __acquires(busiest->lock) | |
1713 | __acquires(this_rq->lock) | |
1714 | { | |
1715 | raw_spin_unlock(&this_rq->lock); | |
1716 | double_rq_lock(this_rq, busiest); | |
1717 | ||
1718 | return 1; | |
1719 | } | |
1720 | ||
1721 | #else | |
1722 | /* | |
1723 | * Unfair double_lock_balance: Optimizes throughput at the expense of | |
1724 | * latency by eliminating extra atomic operations when the locks are | |
1725 | * already in proper order on entry. This favors lower cpu-ids and will | |
1726 | * grant the double lock to lower cpus over higher ids under contention, | |
1727 | * regardless of entry order into the function. | |
1728 | */ | |
1729 | static int _double_lock_balance(struct rq *this_rq, struct rq *busiest) | |
1730 | __releases(this_rq->lock) | |
1731 | __acquires(busiest->lock) | |
1732 | __acquires(this_rq->lock) | |
1733 | { | |
1734 | int ret = 0; | |
1735 | ||
1736 | if (unlikely(!raw_spin_trylock(&busiest->lock))) { | |
1737 | if (busiest < this_rq) { | |
1738 | raw_spin_unlock(&this_rq->lock); | |
1739 | raw_spin_lock(&busiest->lock); | |
1740 | raw_spin_lock_nested(&this_rq->lock, | |
1741 | SINGLE_DEPTH_NESTING); | |
1742 | ret = 1; | |
1743 | } else | |
1744 | raw_spin_lock_nested(&busiest->lock, | |
1745 | SINGLE_DEPTH_NESTING); | |
1746 | } | |
1747 | return ret; | |
1748 | } | |
1749 | ||
1750 | #endif /* CONFIG_PREEMPT */ | |
1751 | ||
1752 | /* | |
1753 | * double_lock_balance - lock the busiest runqueue, this_rq is locked already. | |
1754 | */ | |
1755 | static int double_lock_balance(struct rq *this_rq, struct rq *busiest) | |
1756 | { | |
1757 | if (unlikely(!irqs_disabled())) { | |
1758 | /* printk() doesn't work good under rq->lock */ | |
1759 | raw_spin_unlock(&this_rq->lock); | |
1760 | BUG_ON(1); | |
1761 | } | |
1762 | ||
1763 | return _double_lock_balance(this_rq, busiest); | |
1764 | } | |
1765 | ||
1766 | static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest) | |
1767 | __releases(busiest->lock) | |
1768 | { | |
1769 | raw_spin_unlock(&busiest->lock); | |
1770 | lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_); | |
1771 | } | |
1772 | ||
1773 | /* | |
1774 | * double_rq_lock - safely lock two runqueues | |
1775 | * | |
1776 | * Note this does not disable interrupts like task_rq_lock, | |
1777 | * you need to do so manually before calling. | |
1778 | */ | |
1779 | static void double_rq_lock(struct rq *rq1, struct rq *rq2) | |
1780 | __acquires(rq1->lock) | |
1781 | __acquires(rq2->lock) | |
1782 | { | |
1783 | BUG_ON(!irqs_disabled()); | |
1784 | if (rq1 == rq2) { | |
1785 | raw_spin_lock(&rq1->lock); | |
1786 | __acquire(rq2->lock); /* Fake it out ;) */ | |
1787 | } else { | |
1788 | if (rq1 < rq2) { | |
1789 | raw_spin_lock(&rq1->lock); | |
1790 | raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING); | |
1791 | } else { | |
1792 | raw_spin_lock(&rq2->lock); | |
1793 | raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING); | |
1794 | } | |
1795 | } | |
1796 | } | |
1797 | ||
1798 | /* | |
1799 | * double_rq_unlock - safely unlock two runqueues | |
1800 | * | |
1801 | * Note this does not restore interrupts like task_rq_unlock, | |
1802 | * you need to do so manually after calling. | |
1803 | */ | |
1804 | static void double_rq_unlock(struct rq *rq1, struct rq *rq2) | |
1805 | __releases(rq1->lock) | |
1806 | __releases(rq2->lock) | |
1807 | { | |
1808 | raw_spin_unlock(&rq1->lock); | |
1809 | if (rq1 != rq2) | |
1810 | raw_spin_unlock(&rq2->lock); | |
1811 | else | |
1812 | __release(rq2->lock); | |
1813 | } | |
1814 | ||
1815 | #endif | |
1816 | ||
1817 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1818 | static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares) | |
1819 | { | |
1820 | #ifdef CONFIG_SMP | |
1821 | cfs_rq->shares = shares; | |
1822 | #endif | |
1823 | } | |
1824 | #endif | |
1825 | ||
1826 | static void calc_load_account_idle(struct rq *this_rq); | |
1827 | static void update_sysctl(void); | |
1828 | static int get_update_sysctl_factor(void); | |
1829 | static void update_cpu_load(struct rq *this_rq); | |
1830 | ||
1831 | static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu) | |
1832 | { | |
1833 | set_task_rq(p, cpu); | |
1834 | #ifdef CONFIG_SMP | |
1835 | /* | |
1836 | * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be | |
1837 | * successfuly executed on another CPU. We must ensure that updates of | |
1838 | * per-task data have been completed by this moment. | |
1839 | */ | |
1840 | smp_wmb(); | |
1841 | task_thread_info(p)->cpu = cpu; | |
1842 | #endif | |
1843 | } | |
1844 | ||
1845 | static const struct sched_class rt_sched_class; | |
1846 | ||
1847 | #define sched_class_highest (&stop_sched_class) | |
1848 | #define for_each_class(class) \ | |
1849 | for (class = sched_class_highest; class; class = class->next) | |
1850 | ||
1851 | #include "sched_stats.h" | |
1852 | ||
1853 | static void inc_nr_running(struct rq *rq) | |
1854 | { | |
1855 | rq->nr_running++; | |
1856 | } | |
1857 | ||
1858 | static void dec_nr_running(struct rq *rq) | |
1859 | { | |
1860 | rq->nr_running--; | |
1861 | } | |
1862 | ||
1863 | static void set_load_weight(struct task_struct *p) | |
1864 | { | |
1865 | /* | |
1866 | * SCHED_IDLE tasks get minimal weight: | |
1867 | */ | |
1868 | if (p->policy == SCHED_IDLE) { | |
1869 | p->se.load.weight = WEIGHT_IDLEPRIO; | |
1870 | p->se.load.inv_weight = WMULT_IDLEPRIO; | |
1871 | return; | |
1872 | } | |
1873 | ||
1874 | p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO]; | |
1875 | p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO]; | |
1876 | } | |
1877 | ||
1878 | static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) | |
1879 | { | |
1880 | update_rq_clock(rq); | |
1881 | sched_info_queued(p); | |
1882 | p->sched_class->enqueue_task(rq, p, flags); | |
1883 | p->se.on_rq = 1; | |
1884 | } | |
1885 | ||
1886 | static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) | |
1887 | { | |
1888 | update_rq_clock(rq); | |
1889 | sched_info_dequeued(p); | |
1890 | p->sched_class->dequeue_task(rq, p, flags); | |
1891 | p->se.on_rq = 0; | |
1892 | } | |
1893 | ||
1894 | /* | |
1895 | * activate_task - move a task to the runqueue. | |
1896 | */ | |
1897 | static void activate_task(struct rq *rq, struct task_struct *p, int flags) | |
1898 | { | |
1899 | if (task_contributes_to_load(p)) | |
1900 | rq->nr_uninterruptible--; | |
1901 | ||
1902 | enqueue_task(rq, p, flags); | |
1903 | inc_nr_running(rq); | |
1904 | } | |
1905 | ||
1906 | /* | |
1907 | * deactivate_task - remove a task from the runqueue. | |
1908 | */ | |
1909 | static void deactivate_task(struct rq *rq, struct task_struct *p, int flags) | |
1910 | { | |
1911 | if (task_contributes_to_load(p)) | |
1912 | rq->nr_uninterruptible++; | |
1913 | ||
1914 | dequeue_task(rq, p, flags); | |
1915 | dec_nr_running(rq); | |
1916 | } | |
1917 | ||
1918 | #ifdef CONFIG_IRQ_TIME_ACCOUNTING | |
1919 | ||
1920 | /* | |
1921 | * There are no locks covering percpu hardirq/softirq time. | |
1922 | * They are only modified in account_system_vtime, on corresponding CPU | |
1923 | * with interrupts disabled. So, writes are safe. | |
1924 | * They are read and saved off onto struct rq in update_rq_clock(). | |
1925 | * This may result in other CPU reading this CPU's irq time and can | |
1926 | * race with irq/account_system_vtime on this CPU. We would either get old | |
1927 | * or new value (or semi updated value on 32 bit) with a side effect of | |
1928 | * accounting a slice of irq time to wrong task when irq is in progress | |
1929 | * while we read rq->clock. That is a worthy compromise in place of having | |
1930 | * locks on each irq in account_system_time. | |
1931 | */ | |
1932 | static DEFINE_PER_CPU(u64, cpu_hardirq_time); | |
1933 | static DEFINE_PER_CPU(u64, cpu_softirq_time); | |
1934 | ||
1935 | static DEFINE_PER_CPU(u64, irq_start_time); | |
1936 | static int sched_clock_irqtime; | |
1937 | ||
1938 | void enable_sched_clock_irqtime(void) | |
1939 | { | |
1940 | sched_clock_irqtime = 1; | |
1941 | } | |
1942 | ||
1943 | void disable_sched_clock_irqtime(void) | |
1944 | { | |
1945 | sched_clock_irqtime = 0; | |
1946 | } | |
1947 | ||
1948 | static u64 irq_time_cpu(int cpu) | |
1949 | { | |
1950 | if (!sched_clock_irqtime) | |
1951 | return 0; | |
1952 | ||
1953 | return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu); | |
1954 | } | |
1955 | ||
1956 | void account_system_vtime(struct task_struct *curr) | |
1957 | { | |
1958 | unsigned long flags; | |
1959 | int cpu; | |
1960 | u64 now, delta; | |
1961 | ||
1962 | if (!sched_clock_irqtime) | |
1963 | return; | |
1964 | ||
1965 | local_irq_save(flags); | |
1966 | ||
1967 | cpu = smp_processor_id(); | |
1968 | now = sched_clock_cpu(cpu); | |
1969 | delta = now - per_cpu(irq_start_time, cpu); | |
1970 | per_cpu(irq_start_time, cpu) = now; | |
1971 | /* | |
1972 | * We do not account for softirq time from ksoftirqd here. | |
1973 | * We want to continue accounting softirq time to ksoftirqd thread | |
1974 | * in that case, so as not to confuse scheduler with a special task | |
1975 | * that do not consume any time, but still wants to run. | |
1976 | */ | |
1977 | if (hardirq_count()) | |
1978 | per_cpu(cpu_hardirq_time, cpu) += delta; | |
1979 | else if (in_serving_softirq() && !(curr->flags & PF_KSOFTIRQD)) | |
1980 | per_cpu(cpu_softirq_time, cpu) += delta; | |
1981 | ||
1982 | local_irq_restore(flags); | |
1983 | } | |
1984 | EXPORT_SYMBOL_GPL(account_system_vtime); | |
1985 | ||
1986 | static void sched_irq_time_avg_update(struct rq *rq, u64 curr_irq_time) | |
1987 | { | |
1988 | if (sched_clock_irqtime && sched_feat(NONIRQ_POWER)) { | |
1989 | u64 delta_irq = curr_irq_time - rq->prev_irq_time; | |
1990 | rq->prev_irq_time = curr_irq_time; | |
1991 | sched_rt_avg_update(rq, delta_irq); | |
1992 | } | |
1993 | } | |
1994 | ||
1995 | #else | |
1996 | ||
1997 | static u64 irq_time_cpu(int cpu) | |
1998 | { | |
1999 | return 0; | |
2000 | } | |
2001 | ||
2002 | static void sched_irq_time_avg_update(struct rq *rq, u64 curr_irq_time) { } | |
2003 | ||
2004 | #endif | |
2005 | ||
2006 | #include "sched_idletask.c" | |
2007 | #include "sched_fair.c" | |
2008 | #include "sched_rt.c" | |
2009 | #include "sched_stoptask.c" | |
2010 | #ifdef CONFIG_SCHED_DEBUG | |
2011 | # include "sched_debug.c" | |
2012 | #endif | |
2013 | ||
2014 | void sched_set_stop_task(int cpu, struct task_struct *stop) | |
2015 | { | |
2016 | struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; | |
2017 | struct task_struct *old_stop = cpu_rq(cpu)->stop; | |
2018 | ||
2019 | if (stop) { | |
2020 | /* | |
2021 | * Make it appear like a SCHED_FIFO task, its something | |
2022 | * userspace knows about and won't get confused about. | |
2023 | * | |
2024 | * Also, it will make PI more or less work without too | |
2025 | * much confusion -- but then, stop work should not | |
2026 | * rely on PI working anyway. | |
2027 | */ | |
2028 | sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); | |
2029 | ||
2030 | stop->sched_class = &stop_sched_class; | |
2031 | } | |
2032 | ||
2033 | cpu_rq(cpu)->stop = stop; | |
2034 | ||
2035 | if (old_stop) { | |
2036 | /* | |
2037 | * Reset it back to a normal scheduling class so that | |
2038 | * it can die in pieces. | |
2039 | */ | |
2040 | old_stop->sched_class = &rt_sched_class; | |
2041 | } | |
2042 | } | |
2043 | ||
2044 | /* | |
2045 | * __normal_prio - return the priority that is based on the static prio | |
2046 | */ | |
2047 | static inline int __normal_prio(struct task_struct *p) | |
2048 | { | |
2049 | return p->static_prio; | |
2050 | } | |
2051 | ||
2052 | /* | |
2053 | * Calculate the expected normal priority: i.e. priority | |
2054 | * without taking RT-inheritance into account. Might be | |
2055 | * boosted by interactivity modifiers. Changes upon fork, | |
2056 | * setprio syscalls, and whenever the interactivity | |
2057 | * estimator recalculates. | |
2058 | */ | |
2059 | static inline int normal_prio(struct task_struct *p) | |
2060 | { | |
2061 | int prio; | |
2062 | ||
2063 | if (task_has_rt_policy(p)) | |
2064 | prio = MAX_RT_PRIO-1 - p->rt_priority; | |
2065 | else | |
2066 | prio = __normal_prio(p); | |
2067 | return prio; | |
2068 | } | |
2069 | ||
2070 | /* | |
2071 | * Calculate the current priority, i.e. the priority | |
2072 | * taken into account by the scheduler. This value might | |
2073 | * be boosted by RT tasks, or might be boosted by | |
2074 | * interactivity modifiers. Will be RT if the task got | |
2075 | * RT-boosted. If not then it returns p->normal_prio. | |
2076 | */ | |
2077 | static int effective_prio(struct task_struct *p) | |
2078 | { | |
2079 | p->normal_prio = normal_prio(p); | |
2080 | /* | |
2081 | * If we are RT tasks or we were boosted to RT priority, | |
2082 | * keep the priority unchanged. Otherwise, update priority | |
2083 | * to the normal priority: | |
2084 | */ | |
2085 | if (!rt_prio(p->prio)) | |
2086 | return p->normal_prio; | |
2087 | return p->prio; | |
2088 | } | |
2089 | ||
2090 | /** | |
2091 | * task_curr - is this task currently executing on a CPU? | |
2092 | * @p: the task in question. | |
2093 | */ | |
2094 | inline int task_curr(const struct task_struct *p) | |
2095 | { | |
2096 | return cpu_curr(task_cpu(p)) == p; | |
2097 | } | |
2098 | ||
2099 | static inline void check_class_changed(struct rq *rq, struct task_struct *p, | |
2100 | const struct sched_class *prev_class, | |
2101 | int oldprio, int running) | |
2102 | { | |
2103 | if (prev_class != p->sched_class) { | |
2104 | if (prev_class->switched_from) | |
2105 | prev_class->switched_from(rq, p, running); | |
2106 | p->sched_class->switched_to(rq, p, running); | |
2107 | } else | |
2108 | p->sched_class->prio_changed(rq, p, oldprio, running); | |
2109 | } | |
2110 | ||
2111 | static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) | |
2112 | { | |
2113 | const struct sched_class *class; | |
2114 | ||
2115 | if (p->sched_class == rq->curr->sched_class) { | |
2116 | rq->curr->sched_class->check_preempt_curr(rq, p, flags); | |
2117 | } else { | |
2118 | for_each_class(class) { | |
2119 | if (class == rq->curr->sched_class) | |
2120 | break; | |
2121 | if (class == p->sched_class) { | |
2122 | resched_task(rq->curr); | |
2123 | break; | |
2124 | } | |
2125 | } | |
2126 | } | |
2127 | ||
2128 | /* | |
2129 | * A queue event has occurred, and we're going to schedule. In | |
2130 | * this case, we can save a useless back to back clock update. | |
2131 | */ | |
2132 | if (test_tsk_need_resched(rq->curr)) | |
2133 | rq->skip_clock_update = 1; | |
2134 | } | |
2135 | ||
2136 | #ifdef CONFIG_SMP | |
2137 | /* | |
2138 | * Is this task likely cache-hot: | |
2139 | */ | |
2140 | static int | |
2141 | task_hot(struct task_struct *p, u64 now, struct sched_domain *sd) | |
2142 | { | |
2143 | s64 delta; | |
2144 | ||
2145 | if (p->sched_class != &fair_sched_class) | |
2146 | return 0; | |
2147 | ||
2148 | if (unlikely(p->policy == SCHED_IDLE)) | |
2149 | return 0; | |
2150 | ||
2151 | /* | |
2152 | * Buddy candidates are cache hot: | |
2153 | */ | |
2154 | if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running && | |
2155 | (&p->se == cfs_rq_of(&p->se)->next || | |
2156 | &p->se == cfs_rq_of(&p->se)->last)) | |
2157 | return 1; | |
2158 | ||
2159 | if (sysctl_sched_migration_cost == -1) | |
2160 | return 1; | |
2161 | if (sysctl_sched_migration_cost == 0) | |
2162 | return 0; | |
2163 | ||
2164 | delta = now - p->se.exec_start; | |
2165 | ||
2166 | return delta < (s64)sysctl_sched_migration_cost; | |
2167 | } | |
2168 | ||
2169 | void set_task_cpu(struct task_struct *p, unsigned int new_cpu) | |
2170 | { | |
2171 | #ifdef CONFIG_SCHED_DEBUG | |
2172 | /* | |
2173 | * We should never call set_task_cpu() on a blocked task, | |
2174 | * ttwu() will sort out the placement. | |
2175 | */ | |
2176 | WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && | |
2177 | !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE)); | |
2178 | #endif | |
2179 | ||
2180 | trace_sched_migrate_task(p, new_cpu); | |
2181 | ||
2182 | if (task_cpu(p) != new_cpu) { | |
2183 | p->se.nr_migrations++; | |
2184 | perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0); | |
2185 | } | |
2186 | ||
2187 | __set_task_cpu(p, new_cpu); | |
2188 | } | |
2189 | ||
2190 | struct migration_arg { | |
2191 | struct task_struct *task; | |
2192 | int dest_cpu; | |
2193 | }; | |
2194 | ||
2195 | static int migration_cpu_stop(void *data); | |
2196 | ||
2197 | /* | |
2198 | * The task's runqueue lock must be held. | |
2199 | * Returns true if you have to wait for migration thread. | |
2200 | */ | |
2201 | static bool migrate_task(struct task_struct *p, int dest_cpu) | |
2202 | { | |
2203 | struct rq *rq = task_rq(p); | |
2204 | ||
2205 | /* | |
2206 | * If the task is not on a runqueue (and not running), then | |
2207 | * the next wake-up will properly place the task. | |
2208 | */ | |
2209 | return p->se.on_rq || task_running(rq, p); | |
2210 | } | |
2211 | ||
2212 | /* | |
2213 | * wait_task_inactive - wait for a thread to unschedule. | |
2214 | * | |
2215 | * If @match_state is nonzero, it's the @p->state value just checked and | |
2216 | * not expected to change. If it changes, i.e. @p might have woken up, | |
2217 | * then return zero. When we succeed in waiting for @p to be off its CPU, | |
2218 | * we return a positive number (its total switch count). If a second call | |
2219 | * a short while later returns the same number, the caller can be sure that | |
2220 | * @p has remained unscheduled the whole time. | |
2221 | * | |
2222 | * The caller must ensure that the task *will* unschedule sometime soon, | |
2223 | * else this function might spin for a *long* time. This function can't | |
2224 | * be called with interrupts off, or it may introduce deadlock with | |
2225 | * smp_call_function() if an IPI is sent by the same process we are | |
2226 | * waiting to become inactive. | |
2227 | */ | |
2228 | unsigned long wait_task_inactive(struct task_struct *p, long match_state) | |
2229 | { | |
2230 | unsigned long flags; | |
2231 | int running, on_rq; | |
2232 | unsigned long ncsw; | |
2233 | struct rq *rq; | |
2234 | ||
2235 | for (;;) { | |
2236 | /* | |
2237 | * We do the initial early heuristics without holding | |
2238 | * any task-queue locks at all. We'll only try to get | |
2239 | * the runqueue lock when things look like they will | |
2240 | * work out! | |
2241 | */ | |
2242 | rq = task_rq(p); | |
2243 | ||
2244 | /* | |
2245 | * If the task is actively running on another CPU | |
2246 | * still, just relax and busy-wait without holding | |
2247 | * any locks. | |
2248 | * | |
2249 | * NOTE! Since we don't hold any locks, it's not | |
2250 | * even sure that "rq" stays as the right runqueue! | |
2251 | * But we don't care, since "task_running()" will | |
2252 | * return false if the runqueue has changed and p | |
2253 | * is actually now running somewhere else! | |
2254 | */ | |
2255 | while (task_running(rq, p)) { | |
2256 | if (match_state && unlikely(p->state != match_state)) | |
2257 | return 0; | |
2258 | cpu_relax(); | |
2259 | } | |
2260 | ||
2261 | /* | |
2262 | * Ok, time to look more closely! We need the rq | |
2263 | * lock now, to be *sure*. If we're wrong, we'll | |
2264 | * just go back and repeat. | |
2265 | */ | |
2266 | rq = task_rq_lock(p, &flags); | |
2267 | trace_sched_wait_task(p); | |
2268 | running = task_running(rq, p); | |
2269 | on_rq = p->se.on_rq; | |
2270 | ncsw = 0; | |
2271 | if (!match_state || p->state == match_state) | |
2272 | ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ | |
2273 | task_rq_unlock(rq, &flags); | |
2274 | ||
2275 | /* | |
2276 | * If it changed from the expected state, bail out now. | |
2277 | */ | |
2278 | if (unlikely(!ncsw)) | |
2279 | break; | |
2280 | ||
2281 | /* | |
2282 | * Was it really running after all now that we | |
2283 | * checked with the proper locks actually held? | |
2284 | * | |
2285 | * Oops. Go back and try again.. | |
2286 | */ | |
2287 | if (unlikely(running)) { | |
2288 | cpu_relax(); | |
2289 | continue; | |
2290 | } | |
2291 | ||
2292 | /* | |
2293 | * It's not enough that it's not actively running, | |
2294 | * it must be off the runqueue _entirely_, and not | |
2295 | * preempted! | |
2296 | * | |
2297 | * So if it was still runnable (but just not actively | |
2298 | * running right now), it's preempted, and we should | |
2299 | * yield - it could be a while. | |
2300 | */ | |
2301 | if (unlikely(on_rq)) { | |
2302 | schedule_timeout_uninterruptible(1); | |
2303 | continue; | |
2304 | } | |
2305 | ||
2306 | /* | |
2307 | * Ahh, all good. It wasn't running, and it wasn't | |
2308 | * runnable, which means that it will never become | |
2309 | * running in the future either. We're all done! | |
2310 | */ | |
2311 | break; | |
2312 | } | |
2313 | ||
2314 | return ncsw; | |
2315 | } | |
2316 | ||
2317 | /*** | |
2318 | * kick_process - kick a running thread to enter/exit the kernel | |
2319 | * @p: the to-be-kicked thread | |
2320 | * | |
2321 | * Cause a process which is running on another CPU to enter | |
2322 | * kernel-mode, without any delay. (to get signals handled.) | |
2323 | * | |
2324 | * NOTE: this function doesnt have to take the runqueue lock, | |
2325 | * because all it wants to ensure is that the remote task enters | |
2326 | * the kernel. If the IPI races and the task has been migrated | |
2327 | * to another CPU then no harm is done and the purpose has been | |
2328 | * achieved as well. | |
2329 | */ | |
2330 | void kick_process(struct task_struct *p) | |
2331 | { | |
2332 | int cpu; | |
2333 | ||
2334 | preempt_disable(); | |
2335 | cpu = task_cpu(p); | |
2336 | if ((cpu != smp_processor_id()) && task_curr(p)) | |
2337 | smp_send_reschedule(cpu); | |
2338 | preempt_enable(); | |
2339 | } | |
2340 | EXPORT_SYMBOL_GPL(kick_process); | |
2341 | #endif /* CONFIG_SMP */ | |
2342 | ||
2343 | /** | |
2344 | * task_oncpu_function_call - call a function on the cpu on which a task runs | |
2345 | * @p: the task to evaluate | |
2346 | * @func: the function to be called | |
2347 | * @info: the function call argument | |
2348 | * | |
2349 | * Calls the function @func when the task is currently running. This might | |
2350 | * be on the current CPU, which just calls the function directly | |
2351 | */ | |
2352 | void task_oncpu_function_call(struct task_struct *p, | |
2353 | void (*func) (void *info), void *info) | |
2354 | { | |
2355 | int cpu; | |
2356 | ||
2357 | preempt_disable(); | |
2358 | cpu = task_cpu(p); | |
2359 | if (task_curr(p)) | |
2360 | smp_call_function_single(cpu, func, info, 1); | |
2361 | preempt_enable(); | |
2362 | } | |
2363 | ||
2364 | #ifdef CONFIG_SMP | |
2365 | /* | |
2366 | * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held. | |
2367 | */ | |
2368 | static int select_fallback_rq(int cpu, struct task_struct *p) | |
2369 | { | |
2370 | int dest_cpu; | |
2371 | const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu)); | |
2372 | ||
2373 | /* Look for allowed, online CPU in same node. */ | |
2374 | for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask) | |
2375 | if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed)) | |
2376 | return dest_cpu; | |
2377 | ||
2378 | /* Any allowed, online CPU? */ | |
2379 | dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask); | |
2380 | if (dest_cpu < nr_cpu_ids) | |
2381 | return dest_cpu; | |
2382 | ||
2383 | /* No more Mr. Nice Guy. */ | |
2384 | if (unlikely(dest_cpu >= nr_cpu_ids)) { | |
2385 | dest_cpu = cpuset_cpus_allowed_fallback(p); | |
2386 | /* | |
2387 | * Don't tell them about moving exiting tasks or | |
2388 | * kernel threads (both mm NULL), since they never | |
2389 | * leave kernel. | |
2390 | */ | |
2391 | if (p->mm && printk_ratelimit()) { | |
2392 | printk(KERN_INFO "process %d (%s) no " | |
2393 | "longer affine to cpu%d\n", | |
2394 | task_pid_nr(p), p->comm, cpu); | |
2395 | } | |
2396 | } | |
2397 | ||
2398 | return dest_cpu; | |
2399 | } | |
2400 | ||
2401 | /* | |
2402 | * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable. | |
2403 | */ | |
2404 | static inline | |
2405 | int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags) | |
2406 | { | |
2407 | int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags); | |
2408 | ||
2409 | /* | |
2410 | * In order not to call set_task_cpu() on a blocking task we need | |
2411 | * to rely on ttwu() to place the task on a valid ->cpus_allowed | |
2412 | * cpu. | |
2413 | * | |
2414 | * Since this is common to all placement strategies, this lives here. | |
2415 | * | |
2416 | * [ this allows ->select_task() to simply return task_cpu(p) and | |
2417 | * not worry about this generic constraint ] | |
2418 | */ | |
2419 | if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) || | |
2420 | !cpu_online(cpu))) | |
2421 | cpu = select_fallback_rq(task_cpu(p), p); | |
2422 | ||
2423 | return cpu; | |
2424 | } | |
2425 | ||
2426 | static void update_avg(u64 *avg, u64 sample) | |
2427 | { | |
2428 | s64 diff = sample - *avg; | |
2429 | *avg += diff >> 3; | |
2430 | } | |
2431 | #endif | |
2432 | ||
2433 | static inline void ttwu_activate(struct task_struct *p, struct rq *rq, | |
2434 | bool is_sync, bool is_migrate, bool is_local, | |
2435 | unsigned long en_flags) | |
2436 | { | |
2437 | schedstat_inc(p, se.statistics.nr_wakeups); | |
2438 | if (is_sync) | |
2439 | schedstat_inc(p, se.statistics.nr_wakeups_sync); | |
2440 | if (is_migrate) | |
2441 | schedstat_inc(p, se.statistics.nr_wakeups_migrate); | |
2442 | if (is_local) | |
2443 | schedstat_inc(p, se.statistics.nr_wakeups_local); | |
2444 | else | |
2445 | schedstat_inc(p, se.statistics.nr_wakeups_remote); | |
2446 | ||
2447 | activate_task(rq, p, en_flags); | |
2448 | } | |
2449 | ||
2450 | static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq, | |
2451 | int wake_flags, bool success) | |
2452 | { | |
2453 | trace_sched_wakeup(p, success); | |
2454 | check_preempt_curr(rq, p, wake_flags); | |
2455 | ||
2456 | p->state = TASK_RUNNING; | |
2457 | #ifdef CONFIG_SMP | |
2458 | if (p->sched_class->task_woken) | |
2459 | p->sched_class->task_woken(rq, p); | |
2460 | ||
2461 | if (unlikely(rq->idle_stamp)) { | |
2462 | u64 delta = rq->clock - rq->idle_stamp; | |
2463 | u64 max = 2*sysctl_sched_migration_cost; | |
2464 | ||
2465 | if (delta > max) | |
2466 | rq->avg_idle = max; | |
2467 | else | |
2468 | update_avg(&rq->avg_idle, delta); | |
2469 | rq->idle_stamp = 0; | |
2470 | } | |
2471 | #endif | |
2472 | /* if a worker is waking up, notify workqueue */ | |
2473 | if ((p->flags & PF_WQ_WORKER) && success) | |
2474 | wq_worker_waking_up(p, cpu_of(rq)); | |
2475 | } | |
2476 | ||
2477 | /** | |
2478 | * try_to_wake_up - wake up a thread | |
2479 | * @p: the thread to be awakened | |
2480 | * @state: the mask of task states that can be woken | |
2481 | * @wake_flags: wake modifier flags (WF_*) | |
2482 | * | |
2483 | * Put it on the run-queue if it's not already there. The "current" | |
2484 | * thread is always on the run-queue (except when the actual | |
2485 | * re-schedule is in progress), and as such you're allowed to do | |
2486 | * the simpler "current->state = TASK_RUNNING" to mark yourself | |
2487 | * runnable without the overhead of this. | |
2488 | * | |
2489 | * Returns %true if @p was woken up, %false if it was already running | |
2490 | * or @state didn't match @p's state. | |
2491 | */ | |
2492 | static int try_to_wake_up(struct task_struct *p, unsigned int state, | |
2493 | int wake_flags) | |
2494 | { | |
2495 | int cpu, orig_cpu, this_cpu, success = 0; | |
2496 | unsigned long flags; | |
2497 | unsigned long en_flags = ENQUEUE_WAKEUP; | |
2498 | struct rq *rq; | |
2499 | ||
2500 | this_cpu = get_cpu(); | |
2501 | ||
2502 | smp_wmb(); | |
2503 | rq = task_rq_lock(p, &flags); | |
2504 | if (!(p->state & state)) | |
2505 | goto out; | |
2506 | ||
2507 | if (p->se.on_rq) | |
2508 | goto out_running; | |
2509 | ||
2510 | cpu = task_cpu(p); | |
2511 | orig_cpu = cpu; | |
2512 | ||
2513 | #ifdef CONFIG_SMP | |
2514 | if (unlikely(task_running(rq, p))) | |
2515 | goto out_activate; | |
2516 | ||
2517 | /* | |
2518 | * In order to handle concurrent wakeups and release the rq->lock | |
2519 | * we put the task in TASK_WAKING state. | |
2520 | * | |
2521 | * First fix up the nr_uninterruptible count: | |
2522 | */ | |
2523 | if (task_contributes_to_load(p)) { | |
2524 | if (likely(cpu_online(orig_cpu))) | |
2525 | rq->nr_uninterruptible--; | |
2526 | else | |
2527 | this_rq()->nr_uninterruptible--; | |
2528 | } | |
2529 | p->state = TASK_WAKING; | |
2530 | ||
2531 | if (p->sched_class->task_waking) { | |
2532 | p->sched_class->task_waking(rq, p); | |
2533 | en_flags |= ENQUEUE_WAKING; | |
2534 | } | |
2535 | ||
2536 | cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags); | |
2537 | if (cpu != orig_cpu) | |
2538 | set_task_cpu(p, cpu); | |
2539 | __task_rq_unlock(rq); | |
2540 | ||
2541 | rq = cpu_rq(cpu); | |
2542 | raw_spin_lock(&rq->lock); | |
2543 | ||
2544 | /* | |
2545 | * We migrated the task without holding either rq->lock, however | |
2546 | * since the task is not on the task list itself, nobody else | |
2547 | * will try and migrate the task, hence the rq should match the | |
2548 | * cpu we just moved it to. | |
2549 | */ | |
2550 | WARN_ON(task_cpu(p) != cpu); | |
2551 | WARN_ON(p->state != TASK_WAKING); | |
2552 | ||
2553 | #ifdef CONFIG_SCHEDSTATS | |
2554 | schedstat_inc(rq, ttwu_count); | |
2555 | if (cpu == this_cpu) | |
2556 | schedstat_inc(rq, ttwu_local); | |
2557 | else { | |
2558 | struct sched_domain *sd; | |
2559 | for_each_domain(this_cpu, sd) { | |
2560 | if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { | |
2561 | schedstat_inc(sd, ttwu_wake_remote); | |
2562 | break; | |
2563 | } | |
2564 | } | |
2565 | } | |
2566 | #endif /* CONFIG_SCHEDSTATS */ | |
2567 | ||
2568 | out_activate: | |
2569 | #endif /* CONFIG_SMP */ | |
2570 | ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu, | |
2571 | cpu == this_cpu, en_flags); | |
2572 | success = 1; | |
2573 | out_running: | |
2574 | ttwu_post_activation(p, rq, wake_flags, success); | |
2575 | out: | |
2576 | task_rq_unlock(rq, &flags); | |
2577 | put_cpu(); | |
2578 | ||
2579 | return success; | |
2580 | } | |
2581 | ||
2582 | /** | |
2583 | * try_to_wake_up_local - try to wake up a local task with rq lock held | |
2584 | * @p: the thread to be awakened | |
2585 | * | |
2586 | * Put @p on the run-queue if it's not alredy there. The caller must | |
2587 | * ensure that this_rq() is locked, @p is bound to this_rq() and not | |
2588 | * the current task. this_rq() stays locked over invocation. | |
2589 | */ | |
2590 | static void try_to_wake_up_local(struct task_struct *p) | |
2591 | { | |
2592 | struct rq *rq = task_rq(p); | |
2593 | bool success = false; | |
2594 | ||
2595 | BUG_ON(rq != this_rq()); | |
2596 | BUG_ON(p == current); | |
2597 | lockdep_assert_held(&rq->lock); | |
2598 | ||
2599 | if (!(p->state & TASK_NORMAL)) | |
2600 | return; | |
2601 | ||
2602 | if (!p->se.on_rq) { | |
2603 | if (likely(!task_running(rq, p))) { | |
2604 | schedstat_inc(rq, ttwu_count); | |
2605 | schedstat_inc(rq, ttwu_local); | |
2606 | } | |
2607 | ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP); | |
2608 | success = true; | |
2609 | } | |
2610 | ttwu_post_activation(p, rq, 0, success); | |
2611 | } | |
2612 | ||
2613 | /** | |
2614 | * wake_up_process - Wake up a specific process | |
2615 | * @p: The process to be woken up. | |
2616 | * | |
2617 | * Attempt to wake up the nominated process and move it to the set of runnable | |
2618 | * processes. Returns 1 if the process was woken up, 0 if it was already | |
2619 | * running. | |
2620 | * | |
2621 | * It may be assumed that this function implies a write memory barrier before | |
2622 | * changing the task state if and only if any tasks are woken up. | |
2623 | */ | |
2624 | int wake_up_process(struct task_struct *p) | |
2625 | { | |
2626 | return try_to_wake_up(p, TASK_ALL, 0); | |
2627 | } | |
2628 | EXPORT_SYMBOL(wake_up_process); | |
2629 | ||
2630 | int wake_up_state(struct task_struct *p, unsigned int state) | |
2631 | { | |
2632 | return try_to_wake_up(p, state, 0); | |
2633 | } | |
2634 | ||
2635 | /* | |
2636 | * Perform scheduler related setup for a newly forked process p. | |
2637 | * p is forked by current. | |
2638 | * | |
2639 | * __sched_fork() is basic setup used by init_idle() too: | |
2640 | */ | |
2641 | static void __sched_fork(struct task_struct *p) | |
2642 | { | |
2643 | p->se.exec_start = 0; | |
2644 | p->se.sum_exec_runtime = 0; | |
2645 | p->se.prev_sum_exec_runtime = 0; | |
2646 | p->se.nr_migrations = 0; | |
2647 | ||
2648 | #ifdef CONFIG_SCHEDSTATS | |
2649 | memset(&p->se.statistics, 0, sizeof(p->se.statistics)); | |
2650 | #endif | |
2651 | ||
2652 | INIT_LIST_HEAD(&p->rt.run_list); | |
2653 | p->se.on_rq = 0; | |
2654 | INIT_LIST_HEAD(&p->se.group_node); | |
2655 | ||
2656 | #ifdef CONFIG_PREEMPT_NOTIFIERS | |
2657 | INIT_HLIST_HEAD(&p->preempt_notifiers); | |
2658 | #endif | |
2659 | } | |
2660 | ||
2661 | /* | |
2662 | * fork()/clone()-time setup: | |
2663 | */ | |
2664 | void sched_fork(struct task_struct *p, int clone_flags) | |
2665 | { | |
2666 | int cpu = get_cpu(); | |
2667 | ||
2668 | __sched_fork(p); | |
2669 | /* | |
2670 | * We mark the process as running here. This guarantees that | |
2671 | * nobody will actually run it, and a signal or other external | |
2672 | * event cannot wake it up and insert it on the runqueue either. | |
2673 | */ | |
2674 | p->state = TASK_RUNNING; | |
2675 | ||
2676 | /* | |
2677 | * Revert to default priority/policy on fork if requested. | |
2678 | */ | |
2679 | if (unlikely(p->sched_reset_on_fork)) { | |
2680 | if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) { | |
2681 | p->policy = SCHED_NORMAL; | |
2682 | p->normal_prio = p->static_prio; | |
2683 | } | |
2684 | ||
2685 | if (PRIO_TO_NICE(p->static_prio) < 0) { | |
2686 | p->static_prio = NICE_TO_PRIO(0); | |
2687 | p->normal_prio = p->static_prio; | |
2688 | set_load_weight(p); | |
2689 | } | |
2690 | ||
2691 | /* | |
2692 | * We don't need the reset flag anymore after the fork. It has | |
2693 | * fulfilled its duty: | |
2694 | */ | |
2695 | p->sched_reset_on_fork = 0; | |
2696 | } | |
2697 | ||
2698 | /* | |
2699 | * Make sure we do not leak PI boosting priority to the child. | |
2700 | */ | |
2701 | p->prio = current->normal_prio; | |
2702 | ||
2703 | if (!rt_prio(p->prio)) | |
2704 | p->sched_class = &fair_sched_class; | |
2705 | ||
2706 | if (p->sched_class->task_fork) | |
2707 | p->sched_class->task_fork(p); | |
2708 | ||
2709 | /* | |
2710 | * The child is not yet in the pid-hash so no cgroup attach races, | |
2711 | * and the cgroup is pinned to this child due to cgroup_fork() | |
2712 | * is ran before sched_fork(). | |
2713 | * | |
2714 | * Silence PROVE_RCU. | |
2715 | */ | |
2716 | rcu_read_lock(); | |
2717 | set_task_cpu(p, cpu); | |
2718 | rcu_read_unlock(); | |
2719 | ||
2720 | #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) | |
2721 | if (likely(sched_info_on())) | |
2722 | memset(&p->sched_info, 0, sizeof(p->sched_info)); | |
2723 | #endif | |
2724 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) | |
2725 | p->oncpu = 0; | |
2726 | #endif | |
2727 | #ifdef CONFIG_PREEMPT | |
2728 | /* Want to start with kernel preemption disabled. */ | |
2729 | task_thread_info(p)->preempt_count = 1; | |
2730 | #endif | |
2731 | plist_node_init(&p->pushable_tasks, MAX_PRIO); | |
2732 | ||
2733 | put_cpu(); | |
2734 | } | |
2735 | ||
2736 | /* | |
2737 | * wake_up_new_task - wake up a newly created task for the first time. | |
2738 | * | |
2739 | * This function will do some initial scheduler statistics housekeeping | |
2740 | * that must be done for every newly created context, then puts the task | |
2741 | * on the runqueue and wakes it. | |
2742 | */ | |
2743 | void wake_up_new_task(struct task_struct *p, unsigned long clone_flags) | |
2744 | { | |
2745 | unsigned long flags; | |
2746 | struct rq *rq; | |
2747 | int cpu __maybe_unused = get_cpu(); | |
2748 | ||
2749 | #ifdef CONFIG_SMP | |
2750 | rq = task_rq_lock(p, &flags); | |
2751 | p->state = TASK_WAKING; | |
2752 | ||
2753 | /* | |
2754 | * Fork balancing, do it here and not earlier because: | |
2755 | * - cpus_allowed can change in the fork path | |
2756 | * - any previously selected cpu might disappear through hotplug | |
2757 | * | |
2758 | * We set TASK_WAKING so that select_task_rq() can drop rq->lock | |
2759 | * without people poking at ->cpus_allowed. | |
2760 | */ | |
2761 | cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0); | |
2762 | set_task_cpu(p, cpu); | |
2763 | ||
2764 | p->state = TASK_RUNNING; | |
2765 | task_rq_unlock(rq, &flags); | |
2766 | #endif | |
2767 | ||
2768 | rq = task_rq_lock(p, &flags); | |
2769 | activate_task(rq, p, 0); | |
2770 | trace_sched_wakeup_new(p, 1); | |
2771 | check_preempt_curr(rq, p, WF_FORK); | |
2772 | #ifdef CONFIG_SMP | |
2773 | if (p->sched_class->task_woken) | |
2774 | p->sched_class->task_woken(rq, p); | |
2775 | #endif | |
2776 | task_rq_unlock(rq, &flags); | |
2777 | put_cpu(); | |
2778 | } | |
2779 | ||
2780 | #ifdef CONFIG_PREEMPT_NOTIFIERS | |
2781 | ||
2782 | /** | |
2783 | * preempt_notifier_register - tell me when current is being preempted & rescheduled | |
2784 | * @notifier: notifier struct to register | |
2785 | */ | |
2786 | void preempt_notifier_register(struct preempt_notifier *notifier) | |
2787 | { | |
2788 | hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); | |
2789 | } | |
2790 | EXPORT_SYMBOL_GPL(preempt_notifier_register); | |
2791 | ||
2792 | /** | |
2793 | * preempt_notifier_unregister - no longer interested in preemption notifications | |
2794 | * @notifier: notifier struct to unregister | |
2795 | * | |
2796 | * This is safe to call from within a preemption notifier. | |
2797 | */ | |
2798 | void preempt_notifier_unregister(struct preempt_notifier *notifier) | |
2799 | { | |
2800 | hlist_del(¬ifier->link); | |
2801 | } | |
2802 | EXPORT_SYMBOL_GPL(preempt_notifier_unregister); | |
2803 | ||
2804 | static void fire_sched_in_preempt_notifiers(struct task_struct *curr) | |
2805 | { | |
2806 | struct preempt_notifier *notifier; | |
2807 | struct hlist_node *node; | |
2808 | ||
2809 | hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) | |
2810 | notifier->ops->sched_in(notifier, raw_smp_processor_id()); | |
2811 | } | |
2812 | ||
2813 | static void | |
2814 | fire_sched_out_preempt_notifiers(struct task_struct *curr, | |
2815 | struct task_struct *next) | |
2816 | { | |
2817 | struct preempt_notifier *notifier; | |
2818 | struct hlist_node *node; | |
2819 | ||
2820 | hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) | |
2821 | notifier->ops->sched_out(notifier, next); | |
2822 | } | |
2823 | ||
2824 | #else /* !CONFIG_PREEMPT_NOTIFIERS */ | |
2825 | ||
2826 | static void fire_sched_in_preempt_notifiers(struct task_struct *curr) | |
2827 | { | |
2828 | } | |
2829 | ||
2830 | static void | |
2831 | fire_sched_out_preempt_notifiers(struct task_struct *curr, | |
2832 | struct task_struct *next) | |
2833 | { | |
2834 | } | |
2835 | ||
2836 | #endif /* CONFIG_PREEMPT_NOTIFIERS */ | |
2837 | ||
2838 | /** | |
2839 | * prepare_task_switch - prepare to switch tasks | |
2840 | * @rq: the runqueue preparing to switch | |
2841 | * @prev: the current task that is being switched out | |
2842 | * @next: the task we are going to switch to. | |
2843 | * | |
2844 | * This is called with the rq lock held and interrupts off. It must | |
2845 | * be paired with a subsequent finish_task_switch after the context | |
2846 | * switch. | |
2847 | * | |
2848 | * prepare_task_switch sets up locking and calls architecture specific | |
2849 | * hooks. | |
2850 | */ | |
2851 | static inline void | |
2852 | prepare_task_switch(struct rq *rq, struct task_struct *prev, | |
2853 | struct task_struct *next) | |
2854 | { | |
2855 | fire_sched_out_preempt_notifiers(prev, next); | |
2856 | prepare_lock_switch(rq, next); | |
2857 | prepare_arch_switch(next); | |
2858 | } | |
2859 | ||
2860 | /** | |
2861 | * finish_task_switch - clean up after a task-switch | |
2862 | * @rq: runqueue associated with task-switch | |
2863 | * @prev: the thread we just switched away from. | |
2864 | * | |
2865 | * finish_task_switch must be called after the context switch, paired | |
2866 | * with a prepare_task_switch call before the context switch. | |
2867 | * finish_task_switch will reconcile locking set up by prepare_task_switch, | |
2868 | * and do any other architecture-specific cleanup actions. | |
2869 | * | |
2870 | * Note that we may have delayed dropping an mm in context_switch(). If | |
2871 | * so, we finish that here outside of the runqueue lock. (Doing it | |
2872 | * with the lock held can cause deadlocks; see schedule() for | |
2873 | * details.) | |
2874 | */ | |
2875 | static void finish_task_switch(struct rq *rq, struct task_struct *prev) | |
2876 | __releases(rq->lock) | |
2877 | { | |
2878 | struct mm_struct *mm = rq->prev_mm; | |
2879 | long prev_state; | |
2880 | ||
2881 | rq->prev_mm = NULL; | |
2882 | ||
2883 | /* | |
2884 | * A task struct has one reference for the use as "current". | |
2885 | * If a task dies, then it sets TASK_DEAD in tsk->state and calls | |
2886 | * schedule one last time. The schedule call will never return, and | |
2887 | * the scheduled task must drop that reference. | |
2888 | * The test for TASK_DEAD must occur while the runqueue locks are | |
2889 | * still held, otherwise prev could be scheduled on another cpu, die | |
2890 | * there before we look at prev->state, and then the reference would | |
2891 | * be dropped twice. | |
2892 | * Manfred Spraul <manfred@colorfullife.com> | |
2893 | */ | |
2894 | prev_state = prev->state; | |
2895 | finish_arch_switch(prev); | |
2896 | #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
2897 | local_irq_disable(); | |
2898 | #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */ | |
2899 | perf_event_task_sched_in(current); | |
2900 | #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
2901 | local_irq_enable(); | |
2902 | #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */ | |
2903 | finish_lock_switch(rq, prev); | |
2904 | ||
2905 | fire_sched_in_preempt_notifiers(current); | |
2906 | if (mm) | |
2907 | mmdrop(mm); | |
2908 | if (unlikely(prev_state == TASK_DEAD)) { | |
2909 | /* | |
2910 | * Remove function-return probe instances associated with this | |
2911 | * task and put them back on the free list. | |
2912 | */ | |
2913 | kprobe_flush_task(prev); | |
2914 | put_task_struct(prev); | |
2915 | } | |
2916 | } | |
2917 | ||
2918 | #ifdef CONFIG_SMP | |
2919 | ||
2920 | /* assumes rq->lock is held */ | |
2921 | static inline void pre_schedule(struct rq *rq, struct task_struct *prev) | |
2922 | { | |
2923 | if (prev->sched_class->pre_schedule) | |
2924 | prev->sched_class->pre_schedule(rq, prev); | |
2925 | } | |
2926 | ||
2927 | /* rq->lock is NOT held, but preemption is disabled */ | |
2928 | static inline void post_schedule(struct rq *rq) | |
2929 | { | |
2930 | if (rq->post_schedule) { | |
2931 | unsigned long flags; | |
2932 | ||
2933 | raw_spin_lock_irqsave(&rq->lock, flags); | |
2934 | if (rq->curr->sched_class->post_schedule) | |
2935 | rq->curr->sched_class->post_schedule(rq); | |
2936 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
2937 | ||
2938 | rq->post_schedule = 0; | |
2939 | } | |
2940 | } | |
2941 | ||
2942 | #else | |
2943 | ||
2944 | static inline void pre_schedule(struct rq *rq, struct task_struct *p) | |
2945 | { | |
2946 | } | |
2947 | ||
2948 | static inline void post_schedule(struct rq *rq) | |
2949 | { | |
2950 | } | |
2951 | ||
2952 | #endif | |
2953 | ||
2954 | /** | |
2955 | * schedule_tail - first thing a freshly forked thread must call. | |
2956 | * @prev: the thread we just switched away from. | |
2957 | */ | |
2958 | asmlinkage void schedule_tail(struct task_struct *prev) | |
2959 | __releases(rq->lock) | |
2960 | { | |
2961 | struct rq *rq = this_rq(); | |
2962 | ||
2963 | finish_task_switch(rq, prev); | |
2964 | ||
2965 | /* | |
2966 | * FIXME: do we need to worry about rq being invalidated by the | |
2967 | * task_switch? | |
2968 | */ | |
2969 | post_schedule(rq); | |
2970 | ||
2971 | #ifdef __ARCH_WANT_UNLOCKED_CTXSW | |
2972 | /* In this case, finish_task_switch does not reenable preemption */ | |
2973 | preempt_enable(); | |
2974 | #endif | |
2975 | if (current->set_child_tid) | |
2976 | put_user(task_pid_vnr(current), current->set_child_tid); | |
2977 | } | |
2978 | ||
2979 | /* | |
2980 | * context_switch - switch to the new MM and the new | |
2981 | * thread's register state. | |
2982 | */ | |
2983 | static inline void | |
2984 | context_switch(struct rq *rq, struct task_struct *prev, | |
2985 | struct task_struct *next) | |
2986 | { | |
2987 | struct mm_struct *mm, *oldmm; | |
2988 | ||
2989 | prepare_task_switch(rq, prev, next); | |
2990 | trace_sched_switch(prev, next); | |
2991 | mm = next->mm; | |
2992 | oldmm = prev->active_mm; | |
2993 | /* | |
2994 | * For paravirt, this is coupled with an exit in switch_to to | |
2995 | * combine the page table reload and the switch backend into | |
2996 | * one hypercall. | |
2997 | */ | |
2998 | arch_start_context_switch(prev); | |
2999 | ||
3000 | if (!mm) { | |
3001 | next->active_mm = oldmm; | |
3002 | atomic_inc(&oldmm->mm_count); | |
3003 | enter_lazy_tlb(oldmm, next); | |
3004 | } else | |
3005 | switch_mm(oldmm, mm, next); | |
3006 | ||
3007 | if (!prev->mm) { | |
3008 | prev->active_mm = NULL; | |
3009 | rq->prev_mm = oldmm; | |
3010 | } | |
3011 | /* | |
3012 | * Since the runqueue lock will be released by the next | |
3013 | * task (which is an invalid locking op but in the case | |
3014 | * of the scheduler it's an obvious special-case), so we | |
3015 | * do an early lockdep release here: | |
3016 | */ | |
3017 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW | |
3018 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); | |
3019 | #endif | |
3020 | ||
3021 | /* Here we just switch the register state and the stack. */ | |
3022 | switch_to(prev, next, prev); | |
3023 | ||
3024 | barrier(); | |
3025 | /* | |
3026 | * this_rq must be evaluated again because prev may have moved | |
3027 | * CPUs since it called schedule(), thus the 'rq' on its stack | |
3028 | * frame will be invalid. | |
3029 | */ | |
3030 | finish_task_switch(this_rq(), prev); | |
3031 | } | |
3032 | ||
3033 | /* | |
3034 | * nr_running, nr_uninterruptible and nr_context_switches: | |
3035 | * | |
3036 | * externally visible scheduler statistics: current number of runnable | |
3037 | * threads, current number of uninterruptible-sleeping threads, total | |
3038 | * number of context switches performed since bootup. | |
3039 | */ | |
3040 | unsigned long nr_running(void) | |
3041 | { | |
3042 | unsigned long i, sum = 0; | |
3043 | ||
3044 | for_each_online_cpu(i) | |
3045 | sum += cpu_rq(i)->nr_running; | |
3046 | ||
3047 | return sum; | |
3048 | } | |
3049 | ||
3050 | unsigned long nr_uninterruptible(void) | |
3051 | { | |
3052 | unsigned long i, sum = 0; | |
3053 | ||
3054 | for_each_possible_cpu(i) | |
3055 | sum += cpu_rq(i)->nr_uninterruptible; | |
3056 | ||
3057 | /* | |
3058 | * Since we read the counters lockless, it might be slightly | |
3059 | * inaccurate. Do not allow it to go below zero though: | |
3060 | */ | |
3061 | if (unlikely((long)sum < 0)) | |
3062 | sum = 0; | |
3063 | ||
3064 | return sum; | |
3065 | } | |
3066 | ||
3067 | unsigned long long nr_context_switches(void) | |
3068 | { | |
3069 | int i; | |
3070 | unsigned long long sum = 0; | |
3071 | ||
3072 | for_each_possible_cpu(i) | |
3073 | sum += cpu_rq(i)->nr_switches; | |
3074 | ||
3075 | return sum; | |
3076 | } | |
3077 | ||
3078 | unsigned long nr_iowait(void) | |
3079 | { | |
3080 | unsigned long i, sum = 0; | |
3081 | ||
3082 | for_each_possible_cpu(i) | |
3083 | sum += atomic_read(&cpu_rq(i)->nr_iowait); | |
3084 | ||
3085 | return sum; | |
3086 | } | |
3087 | ||
3088 | unsigned long nr_iowait_cpu(int cpu) | |
3089 | { | |
3090 | struct rq *this = cpu_rq(cpu); | |
3091 | return atomic_read(&this->nr_iowait); | |
3092 | } | |
3093 | ||
3094 | unsigned long this_cpu_load(void) | |
3095 | { | |
3096 | struct rq *this = this_rq(); | |
3097 | return this->cpu_load[0]; | |
3098 | } | |
3099 | ||
3100 | ||
3101 | /* Variables and functions for calc_load */ | |
3102 | static atomic_long_t calc_load_tasks; | |
3103 | static unsigned long calc_load_update; | |
3104 | unsigned long avenrun[3]; | |
3105 | EXPORT_SYMBOL(avenrun); | |
3106 | ||
3107 | static long calc_load_fold_active(struct rq *this_rq) | |
3108 | { | |
3109 | long nr_active, delta = 0; | |
3110 | ||
3111 | nr_active = this_rq->nr_running; | |
3112 | nr_active += (long) this_rq->nr_uninterruptible; | |
3113 | ||
3114 | if (nr_active != this_rq->calc_load_active) { | |
3115 | delta = nr_active - this_rq->calc_load_active; | |
3116 | this_rq->calc_load_active = nr_active; | |
3117 | } | |
3118 | ||
3119 | return delta; | |
3120 | } | |
3121 | ||
3122 | #ifdef CONFIG_NO_HZ | |
3123 | /* | |
3124 | * For NO_HZ we delay the active fold to the next LOAD_FREQ update. | |
3125 | * | |
3126 | * When making the ILB scale, we should try to pull this in as well. | |
3127 | */ | |
3128 | static atomic_long_t calc_load_tasks_idle; | |
3129 | ||
3130 | static void calc_load_account_idle(struct rq *this_rq) | |
3131 | { | |
3132 | long delta; | |
3133 | ||
3134 | delta = calc_load_fold_active(this_rq); | |
3135 | if (delta) | |
3136 | atomic_long_add(delta, &calc_load_tasks_idle); | |
3137 | } | |
3138 | ||
3139 | static long calc_load_fold_idle(void) | |
3140 | { | |
3141 | long delta = 0; | |
3142 | ||
3143 | /* | |
3144 | * Its got a race, we don't care... | |
3145 | */ | |
3146 | if (atomic_long_read(&calc_load_tasks_idle)) | |
3147 | delta = atomic_long_xchg(&calc_load_tasks_idle, 0); | |
3148 | ||
3149 | return delta; | |
3150 | } | |
3151 | #else | |
3152 | static void calc_load_account_idle(struct rq *this_rq) | |
3153 | { | |
3154 | } | |
3155 | ||
3156 | static inline long calc_load_fold_idle(void) | |
3157 | { | |
3158 | return 0; | |
3159 | } | |
3160 | #endif | |
3161 | ||
3162 | /** | |
3163 | * get_avenrun - get the load average array | |
3164 | * @loads: pointer to dest load array | |
3165 | * @offset: offset to add | |
3166 | * @shift: shift count to shift the result left | |
3167 | * | |
3168 | * These values are estimates at best, so no need for locking. | |
3169 | */ | |
3170 | void get_avenrun(unsigned long *loads, unsigned long offset, int shift) | |
3171 | { | |
3172 | loads[0] = (avenrun[0] + offset) << shift; | |
3173 | loads[1] = (avenrun[1] + offset) << shift; | |
3174 | loads[2] = (avenrun[2] + offset) << shift; | |
3175 | } | |
3176 | ||
3177 | static unsigned long | |
3178 | calc_load(unsigned long load, unsigned long exp, unsigned long active) | |
3179 | { | |
3180 | load *= exp; | |
3181 | load += active * (FIXED_1 - exp); | |
3182 | return load >> FSHIFT; | |
3183 | } | |
3184 | ||
3185 | /* | |
3186 | * calc_load - update the avenrun load estimates 10 ticks after the | |
3187 | * CPUs have updated calc_load_tasks. | |
3188 | */ | |
3189 | void calc_global_load(void) | |
3190 | { | |
3191 | unsigned long upd = calc_load_update + 10; | |
3192 | long active; | |
3193 | ||
3194 | if (time_before(jiffies, upd)) | |
3195 | return; | |
3196 | ||
3197 | active = atomic_long_read(&calc_load_tasks); | |
3198 | active = active > 0 ? active * FIXED_1 : 0; | |
3199 | ||
3200 | avenrun[0] = calc_load(avenrun[0], EXP_1, active); | |
3201 | avenrun[1] = calc_load(avenrun[1], EXP_5, active); | |
3202 | avenrun[2] = calc_load(avenrun[2], EXP_15, active); | |
3203 | ||
3204 | calc_load_update += LOAD_FREQ; | |
3205 | } | |
3206 | ||
3207 | /* | |
3208 | * Called from update_cpu_load() to periodically update this CPU's | |
3209 | * active count. | |
3210 | */ | |
3211 | static void calc_load_account_active(struct rq *this_rq) | |
3212 | { | |
3213 | long delta; | |
3214 | ||
3215 | if (time_before(jiffies, this_rq->calc_load_update)) | |
3216 | return; | |
3217 | ||
3218 | delta = calc_load_fold_active(this_rq); | |
3219 | delta += calc_load_fold_idle(); | |
3220 | if (delta) | |
3221 | atomic_long_add(delta, &calc_load_tasks); | |
3222 | ||
3223 | this_rq->calc_load_update += LOAD_FREQ; | |
3224 | } | |
3225 | ||
3226 | /* | |
3227 | * The exact cpuload at various idx values, calculated at every tick would be | |
3228 | * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load | |
3229 | * | |
3230 | * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called | |
3231 | * on nth tick when cpu may be busy, then we have: | |
3232 | * load = ((2^idx - 1) / 2^idx)^(n-1) * load | |
3233 | * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load | |
3234 | * | |
3235 | * decay_load_missed() below does efficient calculation of | |
3236 | * load = ((2^idx - 1) / 2^idx)^(n-1) * load | |
3237 | * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load | |
3238 | * | |
3239 | * The calculation is approximated on a 128 point scale. | |
3240 | * degrade_zero_ticks is the number of ticks after which load at any | |
3241 | * particular idx is approximated to be zero. | |
3242 | * degrade_factor is a precomputed table, a row for each load idx. | |
3243 | * Each column corresponds to degradation factor for a power of two ticks, | |
3244 | * based on 128 point scale. | |
3245 | * Example: | |
3246 | * row 2, col 3 (=12) says that the degradation at load idx 2 after | |
3247 | * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8). | |
3248 | * | |
3249 | * With this power of 2 load factors, we can degrade the load n times | |
3250 | * by looking at 1 bits in n and doing as many mult/shift instead of | |
3251 | * n mult/shifts needed by the exact degradation. | |
3252 | */ | |
3253 | #define DEGRADE_SHIFT 7 | |
3254 | static const unsigned char | |
3255 | degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128}; | |
3256 | static const unsigned char | |
3257 | degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = { | |
3258 | {0, 0, 0, 0, 0, 0, 0, 0}, | |
3259 | {64, 32, 8, 0, 0, 0, 0, 0}, | |
3260 | {96, 72, 40, 12, 1, 0, 0}, | |
3261 | {112, 98, 75, 43, 15, 1, 0}, | |
3262 | {120, 112, 98, 76, 45, 16, 2} }; | |
3263 | ||
3264 | /* | |
3265 | * Update cpu_load for any missed ticks, due to tickless idle. The backlog | |
3266 | * would be when CPU is idle and so we just decay the old load without | |
3267 | * adding any new load. | |
3268 | */ | |
3269 | static unsigned long | |
3270 | decay_load_missed(unsigned long load, unsigned long missed_updates, int idx) | |
3271 | { | |
3272 | int j = 0; | |
3273 | ||
3274 | if (!missed_updates) | |
3275 | return load; | |
3276 | ||
3277 | if (missed_updates >= degrade_zero_ticks[idx]) | |
3278 | return 0; | |
3279 | ||
3280 | if (idx == 1) | |
3281 | return load >> missed_updates; | |
3282 | ||
3283 | while (missed_updates) { | |
3284 | if (missed_updates % 2) | |
3285 | load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT; | |
3286 | ||
3287 | missed_updates >>= 1; | |
3288 | j++; | |
3289 | } | |
3290 | return load; | |
3291 | } | |
3292 | ||
3293 | /* | |
3294 | * Update rq->cpu_load[] statistics. This function is usually called every | |
3295 | * scheduler tick (TICK_NSEC). With tickless idle this will not be called | |
3296 | * every tick. We fix it up based on jiffies. | |
3297 | */ | |
3298 | static void update_cpu_load(struct rq *this_rq) | |
3299 | { | |
3300 | unsigned long this_load = this_rq->load.weight; | |
3301 | unsigned long curr_jiffies = jiffies; | |
3302 | unsigned long pending_updates; | |
3303 | int i, scale; | |
3304 | ||
3305 | this_rq->nr_load_updates++; | |
3306 | ||
3307 | /* Avoid repeated calls on same jiffy, when moving in and out of idle */ | |
3308 | if (curr_jiffies == this_rq->last_load_update_tick) | |
3309 | return; | |
3310 | ||
3311 | pending_updates = curr_jiffies - this_rq->last_load_update_tick; | |
3312 | this_rq->last_load_update_tick = curr_jiffies; | |
3313 | ||
3314 | /* Update our load: */ | |
3315 | this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */ | |
3316 | for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { | |
3317 | unsigned long old_load, new_load; | |
3318 | ||
3319 | /* scale is effectively 1 << i now, and >> i divides by scale */ | |
3320 | ||
3321 | old_load = this_rq->cpu_load[i]; | |
3322 | old_load = decay_load_missed(old_load, pending_updates - 1, i); | |
3323 | new_load = this_load; | |
3324 | /* | |
3325 | * Round up the averaging division if load is increasing. This | |
3326 | * prevents us from getting stuck on 9 if the load is 10, for | |
3327 | * example. | |
3328 | */ | |
3329 | if (new_load > old_load) | |
3330 | new_load += scale - 1; | |
3331 | ||
3332 | this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i; | |
3333 | } | |
3334 | ||
3335 | sched_avg_update(this_rq); | |
3336 | } | |
3337 | ||
3338 | static void update_cpu_load_active(struct rq *this_rq) | |
3339 | { | |
3340 | update_cpu_load(this_rq); | |
3341 | ||
3342 | calc_load_account_active(this_rq); | |
3343 | } | |
3344 | ||
3345 | #ifdef CONFIG_SMP | |
3346 | ||
3347 | /* | |
3348 | * sched_exec - execve() is a valuable balancing opportunity, because at | |
3349 | * this point the task has the smallest effective memory and cache footprint. | |
3350 | */ | |
3351 | void sched_exec(void) | |
3352 | { | |
3353 | struct task_struct *p = current; | |
3354 | unsigned long flags; | |
3355 | struct rq *rq; | |
3356 | int dest_cpu; | |
3357 | ||
3358 | rq = task_rq_lock(p, &flags); | |
3359 | dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0); | |
3360 | if (dest_cpu == smp_processor_id()) | |
3361 | goto unlock; | |
3362 | ||
3363 | /* | |
3364 | * select_task_rq() can race against ->cpus_allowed | |
3365 | */ | |
3366 | if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) && | |
3367 | likely(cpu_active(dest_cpu)) && migrate_task(p, dest_cpu)) { | |
3368 | struct migration_arg arg = { p, dest_cpu }; | |
3369 | ||
3370 | task_rq_unlock(rq, &flags); | |
3371 | stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); | |
3372 | return; | |
3373 | } | |
3374 | unlock: | |
3375 | task_rq_unlock(rq, &flags); | |
3376 | } | |
3377 | ||
3378 | #endif | |
3379 | ||
3380 | DEFINE_PER_CPU(struct kernel_stat, kstat); | |
3381 | ||
3382 | EXPORT_PER_CPU_SYMBOL(kstat); | |
3383 | ||
3384 | /* | |
3385 | * Return any ns on the sched_clock that have not yet been accounted in | |
3386 | * @p in case that task is currently running. | |
3387 | * | |
3388 | * Called with task_rq_lock() held on @rq. | |
3389 | */ | |
3390 | static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) | |
3391 | { | |
3392 | u64 ns = 0; | |
3393 | ||
3394 | if (task_current(rq, p)) { | |
3395 | update_rq_clock(rq); | |
3396 | ns = rq->clock_task - p->se.exec_start; | |
3397 | if ((s64)ns < 0) | |
3398 | ns = 0; | |
3399 | } | |
3400 | ||
3401 | return ns; | |
3402 | } | |
3403 | ||
3404 | unsigned long long task_delta_exec(struct task_struct *p) | |
3405 | { | |
3406 | unsigned long flags; | |
3407 | struct rq *rq; | |
3408 | u64 ns = 0; | |
3409 | ||
3410 | rq = task_rq_lock(p, &flags); | |
3411 | ns = do_task_delta_exec(p, rq); | |
3412 | task_rq_unlock(rq, &flags); | |
3413 | ||
3414 | return ns; | |
3415 | } | |
3416 | ||
3417 | /* | |
3418 | * Return accounted runtime for the task. | |
3419 | * In case the task is currently running, return the runtime plus current's | |
3420 | * pending runtime that have not been accounted yet. | |
3421 | */ | |
3422 | unsigned long long task_sched_runtime(struct task_struct *p) | |
3423 | { | |
3424 | unsigned long flags; | |
3425 | struct rq *rq; | |
3426 | u64 ns = 0; | |
3427 | ||
3428 | rq = task_rq_lock(p, &flags); | |
3429 | ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); | |
3430 | task_rq_unlock(rq, &flags); | |
3431 | ||
3432 | return ns; | |
3433 | } | |
3434 | ||
3435 | /* | |
3436 | * Return sum_exec_runtime for the thread group. | |
3437 | * In case the task is currently running, return the sum plus current's | |
3438 | * pending runtime that have not been accounted yet. | |
3439 | * | |
3440 | * Note that the thread group might have other running tasks as well, | |
3441 | * so the return value not includes other pending runtime that other | |
3442 | * running tasks might have. | |
3443 | */ | |
3444 | unsigned long long thread_group_sched_runtime(struct task_struct *p) | |
3445 | { | |
3446 | struct task_cputime totals; | |
3447 | unsigned long flags; | |
3448 | struct rq *rq; | |
3449 | u64 ns; | |
3450 | ||
3451 | rq = task_rq_lock(p, &flags); | |
3452 | thread_group_cputime(p, &totals); | |
3453 | ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq); | |
3454 | task_rq_unlock(rq, &flags); | |
3455 | ||
3456 | return ns; | |
3457 | } | |
3458 | ||
3459 | /* | |
3460 | * Account user cpu time to a process. | |
3461 | * @p: the process that the cpu time gets accounted to | |
3462 | * @cputime: the cpu time spent in user space since the last update | |
3463 | * @cputime_scaled: cputime scaled by cpu frequency | |
3464 | */ | |
3465 | void account_user_time(struct task_struct *p, cputime_t cputime, | |
3466 | cputime_t cputime_scaled) | |
3467 | { | |
3468 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
3469 | cputime64_t tmp; | |
3470 | ||
3471 | /* Add user time to process. */ | |
3472 | p->utime = cputime_add(p->utime, cputime); | |
3473 | p->utimescaled = cputime_add(p->utimescaled, cputime_scaled); | |
3474 | account_group_user_time(p, cputime); | |
3475 | ||
3476 | /* Add user time to cpustat. */ | |
3477 | tmp = cputime_to_cputime64(cputime); | |
3478 | if (TASK_NICE(p) > 0) | |
3479 | cpustat->nice = cputime64_add(cpustat->nice, tmp); | |
3480 | else | |
3481 | cpustat->user = cputime64_add(cpustat->user, tmp); | |
3482 | ||
3483 | cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime); | |
3484 | /* Account for user time used */ | |
3485 | acct_update_integrals(p); | |
3486 | } | |
3487 | ||
3488 | /* | |
3489 | * Account guest cpu time to a process. | |
3490 | * @p: the process that the cpu time gets accounted to | |
3491 | * @cputime: the cpu time spent in virtual machine since the last update | |
3492 | * @cputime_scaled: cputime scaled by cpu frequency | |
3493 | */ | |
3494 | static void account_guest_time(struct task_struct *p, cputime_t cputime, | |
3495 | cputime_t cputime_scaled) | |
3496 | { | |
3497 | cputime64_t tmp; | |
3498 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
3499 | ||
3500 | tmp = cputime_to_cputime64(cputime); | |
3501 | ||
3502 | /* Add guest time to process. */ | |
3503 | p->utime = cputime_add(p->utime, cputime); | |
3504 | p->utimescaled = cputime_add(p->utimescaled, cputime_scaled); | |
3505 | account_group_user_time(p, cputime); | |
3506 | p->gtime = cputime_add(p->gtime, cputime); | |
3507 | ||
3508 | /* Add guest time to cpustat. */ | |
3509 | if (TASK_NICE(p) > 0) { | |
3510 | cpustat->nice = cputime64_add(cpustat->nice, tmp); | |
3511 | cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp); | |
3512 | } else { | |
3513 | cpustat->user = cputime64_add(cpustat->user, tmp); | |
3514 | cpustat->guest = cputime64_add(cpustat->guest, tmp); | |
3515 | } | |
3516 | } | |
3517 | ||
3518 | /* | |
3519 | * Account system cpu time to a process. | |
3520 | * @p: the process that the cpu time gets accounted to | |
3521 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
3522 | * @cputime: the cpu time spent in kernel space since the last update | |
3523 | * @cputime_scaled: cputime scaled by cpu frequency | |
3524 | */ | |
3525 | void account_system_time(struct task_struct *p, int hardirq_offset, | |
3526 | cputime_t cputime, cputime_t cputime_scaled) | |
3527 | { | |
3528 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
3529 | cputime64_t tmp; | |
3530 | ||
3531 | if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) { | |
3532 | account_guest_time(p, cputime, cputime_scaled); | |
3533 | return; | |
3534 | } | |
3535 | ||
3536 | /* Add system time to process. */ | |
3537 | p->stime = cputime_add(p->stime, cputime); | |
3538 | p->stimescaled = cputime_add(p->stimescaled, cputime_scaled); | |
3539 | account_group_system_time(p, cputime); | |
3540 | ||
3541 | /* Add system time to cpustat. */ | |
3542 | tmp = cputime_to_cputime64(cputime); | |
3543 | if (hardirq_count() - hardirq_offset) | |
3544 | cpustat->irq = cputime64_add(cpustat->irq, tmp); | |
3545 | else if (in_serving_softirq()) | |
3546 | cpustat->softirq = cputime64_add(cpustat->softirq, tmp); | |
3547 | else | |
3548 | cpustat->system = cputime64_add(cpustat->system, tmp); | |
3549 | ||
3550 | cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime); | |
3551 | ||
3552 | /* Account for system time used */ | |
3553 | acct_update_integrals(p); | |
3554 | } | |
3555 | ||
3556 | /* | |
3557 | * Account for involuntary wait time. | |
3558 | * @steal: the cpu time spent in involuntary wait | |
3559 | */ | |
3560 | void account_steal_time(cputime_t cputime) | |
3561 | { | |
3562 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
3563 | cputime64_t cputime64 = cputime_to_cputime64(cputime); | |
3564 | ||
3565 | cpustat->steal = cputime64_add(cpustat->steal, cputime64); | |
3566 | } | |
3567 | ||
3568 | /* | |
3569 | * Account for idle time. | |
3570 | * @cputime: the cpu time spent in idle wait | |
3571 | */ | |
3572 | void account_idle_time(cputime_t cputime) | |
3573 | { | |
3574 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
3575 | cputime64_t cputime64 = cputime_to_cputime64(cputime); | |
3576 | struct rq *rq = this_rq(); | |
3577 | ||
3578 | if (atomic_read(&rq->nr_iowait) > 0) | |
3579 | cpustat->iowait = cputime64_add(cpustat->iowait, cputime64); | |
3580 | else | |
3581 | cpustat->idle = cputime64_add(cpustat->idle, cputime64); | |
3582 | } | |
3583 | ||
3584 | #ifndef CONFIG_VIRT_CPU_ACCOUNTING | |
3585 | ||
3586 | /* | |
3587 | * Account a single tick of cpu time. | |
3588 | * @p: the process that the cpu time gets accounted to | |
3589 | * @user_tick: indicates if the tick is a user or a system tick | |
3590 | */ | |
3591 | void account_process_tick(struct task_struct *p, int user_tick) | |
3592 | { | |
3593 | cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy); | |
3594 | struct rq *rq = this_rq(); | |
3595 | ||
3596 | if (user_tick) | |
3597 | account_user_time(p, cputime_one_jiffy, one_jiffy_scaled); | |
3598 | else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET)) | |
3599 | account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy, | |
3600 | one_jiffy_scaled); | |
3601 | else | |
3602 | account_idle_time(cputime_one_jiffy); | |
3603 | } | |
3604 | ||
3605 | /* | |
3606 | * Account multiple ticks of steal time. | |
3607 | * @p: the process from which the cpu time has been stolen | |
3608 | * @ticks: number of stolen ticks | |
3609 | */ | |
3610 | void account_steal_ticks(unsigned long ticks) | |
3611 | { | |
3612 | account_steal_time(jiffies_to_cputime(ticks)); | |
3613 | } | |
3614 | ||
3615 | /* | |
3616 | * Account multiple ticks of idle time. | |
3617 | * @ticks: number of stolen ticks | |
3618 | */ | |
3619 | void account_idle_ticks(unsigned long ticks) | |
3620 | { | |
3621 | account_idle_time(jiffies_to_cputime(ticks)); | |
3622 | } | |
3623 | ||
3624 | #endif | |
3625 | ||
3626 | /* | |
3627 | * Use precise platform statistics if available: | |
3628 | */ | |
3629 | #ifdef CONFIG_VIRT_CPU_ACCOUNTING | |
3630 | void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st) | |
3631 | { | |
3632 | *ut = p->utime; | |
3633 | *st = p->stime; | |
3634 | } | |
3635 | ||
3636 | void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st) | |
3637 | { | |
3638 | struct task_cputime cputime; | |
3639 | ||
3640 | thread_group_cputime(p, &cputime); | |
3641 | ||
3642 | *ut = cputime.utime; | |
3643 | *st = cputime.stime; | |
3644 | } | |
3645 | #else | |
3646 | ||
3647 | #ifndef nsecs_to_cputime | |
3648 | # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs) | |
3649 | #endif | |
3650 | ||
3651 | void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st) | |
3652 | { | |
3653 | cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime); | |
3654 | ||
3655 | /* | |
3656 | * Use CFS's precise accounting: | |
3657 | */ | |
3658 | rtime = nsecs_to_cputime(p->se.sum_exec_runtime); | |
3659 | ||
3660 | if (total) { | |
3661 | u64 temp = rtime; | |
3662 | ||
3663 | temp *= utime; | |
3664 | do_div(temp, total); | |
3665 | utime = (cputime_t)temp; | |
3666 | } else | |
3667 | utime = rtime; | |
3668 | ||
3669 | /* | |
3670 | * Compare with previous values, to keep monotonicity: | |
3671 | */ | |
3672 | p->prev_utime = max(p->prev_utime, utime); | |
3673 | p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime)); | |
3674 | ||
3675 | *ut = p->prev_utime; | |
3676 | *st = p->prev_stime; | |
3677 | } | |
3678 | ||
3679 | /* | |
3680 | * Must be called with siglock held. | |
3681 | */ | |
3682 | void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st) | |
3683 | { | |
3684 | struct signal_struct *sig = p->signal; | |
3685 | struct task_cputime cputime; | |
3686 | cputime_t rtime, utime, total; | |
3687 | ||
3688 | thread_group_cputime(p, &cputime); | |
3689 | ||
3690 | total = cputime_add(cputime.utime, cputime.stime); | |
3691 | rtime = nsecs_to_cputime(cputime.sum_exec_runtime); | |
3692 | ||
3693 | if (total) { | |
3694 | u64 temp = rtime; | |
3695 | ||
3696 | temp *= cputime.utime; | |
3697 | do_div(temp, total); | |
3698 | utime = (cputime_t)temp; | |
3699 | } else | |
3700 | utime = rtime; | |
3701 | ||
3702 | sig->prev_utime = max(sig->prev_utime, utime); | |
3703 | sig->prev_stime = max(sig->prev_stime, | |
3704 | cputime_sub(rtime, sig->prev_utime)); | |
3705 | ||
3706 | *ut = sig->prev_utime; | |
3707 | *st = sig->prev_stime; | |
3708 | } | |
3709 | #endif | |
3710 | ||
3711 | /* | |
3712 | * This function gets called by the timer code, with HZ frequency. | |
3713 | * We call it with interrupts disabled. | |
3714 | * | |
3715 | * It also gets called by the fork code, when changing the parent's | |
3716 | * timeslices. | |
3717 | */ | |
3718 | void scheduler_tick(void) | |
3719 | { | |
3720 | int cpu = smp_processor_id(); | |
3721 | struct rq *rq = cpu_rq(cpu); | |
3722 | struct task_struct *curr = rq->curr; | |
3723 | ||
3724 | sched_clock_tick(); | |
3725 | ||
3726 | raw_spin_lock(&rq->lock); | |
3727 | update_rq_clock(rq); | |
3728 | update_cpu_load_active(rq); | |
3729 | curr->sched_class->task_tick(rq, curr, 0); | |
3730 | raw_spin_unlock(&rq->lock); | |
3731 | ||
3732 | perf_event_task_tick(); | |
3733 | ||
3734 | #ifdef CONFIG_SMP | |
3735 | rq->idle_at_tick = idle_cpu(cpu); | |
3736 | trigger_load_balance(rq, cpu); | |
3737 | #endif | |
3738 | } | |
3739 | ||
3740 | notrace unsigned long get_parent_ip(unsigned long addr) | |
3741 | { | |
3742 | if (in_lock_functions(addr)) { | |
3743 | addr = CALLER_ADDR2; | |
3744 | if (in_lock_functions(addr)) | |
3745 | addr = CALLER_ADDR3; | |
3746 | } | |
3747 | return addr; | |
3748 | } | |
3749 | ||
3750 | #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ | |
3751 | defined(CONFIG_PREEMPT_TRACER)) | |
3752 | ||
3753 | void __kprobes add_preempt_count(int val) | |
3754 | { | |
3755 | #ifdef CONFIG_DEBUG_PREEMPT | |
3756 | /* | |
3757 | * Underflow? | |
3758 | */ | |
3759 | if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) | |
3760 | return; | |
3761 | #endif | |
3762 | preempt_count() += val; | |
3763 | #ifdef CONFIG_DEBUG_PREEMPT | |
3764 | /* | |
3765 | * Spinlock count overflowing soon? | |
3766 | */ | |
3767 | DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= | |
3768 | PREEMPT_MASK - 10); | |
3769 | #endif | |
3770 | if (preempt_count() == val) | |
3771 | trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); | |
3772 | } | |
3773 | EXPORT_SYMBOL(add_preempt_count); | |
3774 | ||
3775 | void __kprobes sub_preempt_count(int val) | |
3776 | { | |
3777 | #ifdef CONFIG_DEBUG_PREEMPT | |
3778 | /* | |
3779 | * Underflow? | |
3780 | */ | |
3781 | if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) | |
3782 | return; | |
3783 | /* | |
3784 | * Is the spinlock portion underflowing? | |
3785 | */ | |
3786 | if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && | |
3787 | !(preempt_count() & PREEMPT_MASK))) | |
3788 | return; | |
3789 | #endif | |
3790 | ||
3791 | if (preempt_count() == val) | |
3792 | trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); | |
3793 | preempt_count() -= val; | |
3794 | } | |
3795 | EXPORT_SYMBOL(sub_preempt_count); | |
3796 | ||
3797 | #endif | |
3798 | ||
3799 | /* | |
3800 | * Print scheduling while atomic bug: | |
3801 | */ | |
3802 | static noinline void __schedule_bug(struct task_struct *prev) | |
3803 | { | |
3804 | struct pt_regs *regs = get_irq_regs(); | |
3805 | ||
3806 | printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", | |
3807 | prev->comm, prev->pid, preempt_count()); | |
3808 | ||
3809 | debug_show_held_locks(prev); | |
3810 | print_modules(); | |
3811 | if (irqs_disabled()) | |
3812 | print_irqtrace_events(prev); | |
3813 | ||
3814 | if (regs) | |
3815 | show_regs(regs); | |
3816 | else | |
3817 | dump_stack(); | |
3818 | } | |
3819 | ||
3820 | /* | |
3821 | * Various schedule()-time debugging checks and statistics: | |
3822 | */ | |
3823 | static inline void schedule_debug(struct task_struct *prev) | |
3824 | { | |
3825 | /* | |
3826 | * Test if we are atomic. Since do_exit() needs to call into | |
3827 | * schedule() atomically, we ignore that path for now. | |
3828 | * Otherwise, whine if we are scheduling when we should not be. | |
3829 | */ | |
3830 | if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) | |
3831 | __schedule_bug(prev); | |
3832 | ||
3833 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); | |
3834 | ||
3835 | schedstat_inc(this_rq(), sched_count); | |
3836 | #ifdef CONFIG_SCHEDSTATS | |
3837 | if (unlikely(prev->lock_depth >= 0)) { | |
3838 | schedstat_inc(this_rq(), bkl_count); | |
3839 | schedstat_inc(prev, sched_info.bkl_count); | |
3840 | } | |
3841 | #endif | |
3842 | } | |
3843 | ||
3844 | static void put_prev_task(struct rq *rq, struct task_struct *prev) | |
3845 | { | |
3846 | if (prev->se.on_rq) | |
3847 | update_rq_clock(rq); | |
3848 | rq->skip_clock_update = 0; | |
3849 | prev->sched_class->put_prev_task(rq, prev); | |
3850 | } | |
3851 | ||
3852 | /* | |
3853 | * Pick up the highest-prio task: | |
3854 | */ | |
3855 | static inline struct task_struct * | |
3856 | pick_next_task(struct rq *rq) | |
3857 | { | |
3858 | const struct sched_class *class; | |
3859 | struct task_struct *p; | |
3860 | ||
3861 | /* | |
3862 | * Optimization: we know that if all tasks are in | |
3863 | * the fair class we can call that function directly: | |
3864 | */ | |
3865 | if (likely(rq->nr_running == rq->cfs.nr_running)) { | |
3866 | p = fair_sched_class.pick_next_task(rq); | |
3867 | if (likely(p)) | |
3868 | return p; | |
3869 | } | |
3870 | ||
3871 | for_each_class(class) { | |
3872 | p = class->pick_next_task(rq); | |
3873 | if (p) | |
3874 | return p; | |
3875 | } | |
3876 | ||
3877 | BUG(); /* the idle class will always have a runnable task */ | |
3878 | } | |
3879 | ||
3880 | /* | |
3881 | * schedule() is the main scheduler function. | |
3882 | */ | |
3883 | asmlinkage void __sched schedule(void) | |
3884 | { | |
3885 | struct task_struct *prev, *next; | |
3886 | unsigned long *switch_count; | |
3887 | struct rq *rq; | |
3888 | int cpu; | |
3889 | ||
3890 | need_resched: | |
3891 | preempt_disable(); | |
3892 | cpu = smp_processor_id(); | |
3893 | rq = cpu_rq(cpu); | |
3894 | rcu_note_context_switch(cpu); | |
3895 | prev = rq->curr; | |
3896 | ||
3897 | release_kernel_lock(prev); | |
3898 | need_resched_nonpreemptible: | |
3899 | ||
3900 | schedule_debug(prev); | |
3901 | ||
3902 | if (sched_feat(HRTICK)) | |
3903 | hrtick_clear(rq); | |
3904 | ||
3905 | raw_spin_lock_irq(&rq->lock); | |
3906 | clear_tsk_need_resched(prev); | |
3907 | ||
3908 | switch_count = &prev->nivcsw; | |
3909 | if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { | |
3910 | if (unlikely(signal_pending_state(prev->state, prev))) { | |
3911 | prev->state = TASK_RUNNING; | |
3912 | } else { | |
3913 | /* | |
3914 | * If a worker is going to sleep, notify and | |
3915 | * ask workqueue whether it wants to wake up a | |
3916 | * task to maintain concurrency. If so, wake | |
3917 | * up the task. | |
3918 | */ | |
3919 | if (prev->flags & PF_WQ_WORKER) { | |
3920 | struct task_struct *to_wakeup; | |
3921 | ||
3922 | to_wakeup = wq_worker_sleeping(prev, cpu); | |
3923 | if (to_wakeup) | |
3924 | try_to_wake_up_local(to_wakeup); | |
3925 | } | |
3926 | deactivate_task(rq, prev, DEQUEUE_SLEEP); | |
3927 | } | |
3928 | switch_count = &prev->nvcsw; | |
3929 | } | |
3930 | ||
3931 | pre_schedule(rq, prev); | |
3932 | ||
3933 | if (unlikely(!rq->nr_running)) | |
3934 | idle_balance(cpu, rq); | |
3935 | ||
3936 | put_prev_task(rq, prev); | |
3937 | next = pick_next_task(rq); | |
3938 | ||
3939 | if (likely(prev != next)) { | |
3940 | sched_info_switch(prev, next); | |
3941 | perf_event_task_sched_out(prev, next); | |
3942 | ||
3943 | rq->nr_switches++; | |
3944 | rq->curr = next; | |
3945 | ++*switch_count; | |
3946 | ||
3947 | context_switch(rq, prev, next); /* unlocks the rq */ | |
3948 | /* | |
3949 | * The context switch have flipped the stack from under us | |
3950 | * and restored the local variables which were saved when | |
3951 | * this task called schedule() in the past. prev == current | |
3952 | * is still correct, but it can be moved to another cpu/rq. | |
3953 | */ | |
3954 | cpu = smp_processor_id(); | |
3955 | rq = cpu_rq(cpu); | |
3956 | } else | |
3957 | raw_spin_unlock_irq(&rq->lock); | |
3958 | ||
3959 | post_schedule(rq); | |
3960 | ||
3961 | if (unlikely(reacquire_kernel_lock(prev))) | |
3962 | goto need_resched_nonpreemptible; | |
3963 | ||
3964 | preempt_enable_no_resched(); | |
3965 | if (need_resched()) | |
3966 | goto need_resched; | |
3967 | } | |
3968 | EXPORT_SYMBOL(schedule); | |
3969 | ||
3970 | #ifdef CONFIG_MUTEX_SPIN_ON_OWNER | |
3971 | /* | |
3972 | * Look out! "owner" is an entirely speculative pointer | |
3973 | * access and not reliable. | |
3974 | */ | |
3975 | int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner) | |
3976 | { | |
3977 | unsigned int cpu; | |
3978 | struct rq *rq; | |
3979 | ||
3980 | if (!sched_feat(OWNER_SPIN)) | |
3981 | return 0; | |
3982 | ||
3983 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3984 | /* | |
3985 | * Need to access the cpu field knowing that | |
3986 | * DEBUG_PAGEALLOC could have unmapped it if | |
3987 | * the mutex owner just released it and exited. | |
3988 | */ | |
3989 | if (probe_kernel_address(&owner->cpu, cpu)) | |
3990 | return 0; | |
3991 | #else | |
3992 | cpu = owner->cpu; | |
3993 | #endif | |
3994 | ||
3995 | /* | |
3996 | * Even if the access succeeded (likely case), | |
3997 | * the cpu field may no longer be valid. | |
3998 | */ | |
3999 | if (cpu >= nr_cpumask_bits) | |
4000 | return 0; | |
4001 | ||
4002 | /* | |
4003 | * We need to validate that we can do a | |
4004 | * get_cpu() and that we have the percpu area. | |
4005 | */ | |
4006 | if (!cpu_online(cpu)) | |
4007 | return 0; | |
4008 | ||
4009 | rq = cpu_rq(cpu); | |
4010 | ||
4011 | for (;;) { | |
4012 | /* | |
4013 | * Owner changed, break to re-assess state. | |
4014 | */ | |
4015 | if (lock->owner != owner) { | |
4016 | /* | |
4017 | * If the lock has switched to a different owner, | |
4018 | * we likely have heavy contention. Return 0 to quit | |
4019 | * optimistic spinning and not contend further: | |
4020 | */ | |
4021 | if (lock->owner) | |
4022 | return 0; | |
4023 | break; | |
4024 | } | |
4025 | ||
4026 | /* | |
4027 | * Is that owner really running on that cpu? | |
4028 | */ | |
4029 | if (task_thread_info(rq->curr) != owner || need_resched()) | |
4030 | return 0; | |
4031 | ||
4032 | cpu_relax(); | |
4033 | } | |
4034 | ||
4035 | return 1; | |
4036 | } | |
4037 | #endif | |
4038 | ||
4039 | #ifdef CONFIG_PREEMPT | |
4040 | /* | |
4041 | * this is the entry point to schedule() from in-kernel preemption | |
4042 | * off of preempt_enable. Kernel preemptions off return from interrupt | |
4043 | * occur there and call schedule directly. | |
4044 | */ | |
4045 | asmlinkage void __sched notrace preempt_schedule(void) | |
4046 | { | |
4047 | struct thread_info *ti = current_thread_info(); | |
4048 | ||
4049 | /* | |
4050 | * If there is a non-zero preempt_count or interrupts are disabled, | |
4051 | * we do not want to preempt the current task. Just return.. | |
4052 | */ | |
4053 | if (likely(ti->preempt_count || irqs_disabled())) | |
4054 | return; | |
4055 | ||
4056 | do { | |
4057 | add_preempt_count_notrace(PREEMPT_ACTIVE); | |
4058 | schedule(); | |
4059 | sub_preempt_count_notrace(PREEMPT_ACTIVE); | |
4060 | ||
4061 | /* | |
4062 | * Check again in case we missed a preemption opportunity | |
4063 | * between schedule and now. | |
4064 | */ | |
4065 | barrier(); | |
4066 | } while (need_resched()); | |
4067 | } | |
4068 | EXPORT_SYMBOL(preempt_schedule); | |
4069 | ||
4070 | /* | |
4071 | * this is the entry point to schedule() from kernel preemption | |
4072 | * off of irq context. | |
4073 | * Note, that this is called and return with irqs disabled. This will | |
4074 | * protect us against recursive calling from irq. | |
4075 | */ | |
4076 | asmlinkage void __sched preempt_schedule_irq(void) | |
4077 | { | |
4078 | struct thread_info *ti = current_thread_info(); | |
4079 | ||
4080 | /* Catch callers which need to be fixed */ | |
4081 | BUG_ON(ti->preempt_count || !irqs_disabled()); | |
4082 | ||
4083 | do { | |
4084 | add_preempt_count(PREEMPT_ACTIVE); | |
4085 | local_irq_enable(); | |
4086 | schedule(); | |
4087 | local_irq_disable(); | |
4088 | sub_preempt_count(PREEMPT_ACTIVE); | |
4089 | ||
4090 | /* | |
4091 | * Check again in case we missed a preemption opportunity | |
4092 | * between schedule and now. | |
4093 | */ | |
4094 | barrier(); | |
4095 | } while (need_resched()); | |
4096 | } | |
4097 | ||
4098 | #endif /* CONFIG_PREEMPT */ | |
4099 | ||
4100 | int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, | |
4101 | void *key) | |
4102 | { | |
4103 | return try_to_wake_up(curr->private, mode, wake_flags); | |
4104 | } | |
4105 | EXPORT_SYMBOL(default_wake_function); | |
4106 | ||
4107 | /* | |
4108 | * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just | |
4109 | * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve | |
4110 | * number) then we wake all the non-exclusive tasks and one exclusive task. | |
4111 | * | |
4112 | * There are circumstances in which we can try to wake a task which has already | |
4113 | * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns | |
4114 | * zero in this (rare) case, and we handle it by continuing to scan the queue. | |
4115 | */ | |
4116 | static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, | |
4117 | int nr_exclusive, int wake_flags, void *key) | |
4118 | { | |
4119 | wait_queue_t *curr, *next; | |
4120 | ||
4121 | list_for_each_entry_safe(curr, next, &q->task_list, task_list) { | |
4122 | unsigned flags = curr->flags; | |
4123 | ||
4124 | if (curr->func(curr, mode, wake_flags, key) && | |
4125 | (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) | |
4126 | break; | |
4127 | } | |
4128 | } | |
4129 | ||
4130 | /** | |
4131 | * __wake_up - wake up threads blocked on a waitqueue. | |
4132 | * @q: the waitqueue | |
4133 | * @mode: which threads | |
4134 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
4135 | * @key: is directly passed to the wakeup function | |
4136 | * | |
4137 | * It may be assumed that this function implies a write memory barrier before | |
4138 | * changing the task state if and only if any tasks are woken up. | |
4139 | */ | |
4140 | void __wake_up(wait_queue_head_t *q, unsigned int mode, | |
4141 | int nr_exclusive, void *key) | |
4142 | { | |
4143 | unsigned long flags; | |
4144 | ||
4145 | spin_lock_irqsave(&q->lock, flags); | |
4146 | __wake_up_common(q, mode, nr_exclusive, 0, key); | |
4147 | spin_unlock_irqrestore(&q->lock, flags); | |
4148 | } | |
4149 | EXPORT_SYMBOL(__wake_up); | |
4150 | ||
4151 | /* | |
4152 | * Same as __wake_up but called with the spinlock in wait_queue_head_t held. | |
4153 | */ | |
4154 | void __wake_up_locked(wait_queue_head_t *q, unsigned int mode) | |
4155 | { | |
4156 | __wake_up_common(q, mode, 1, 0, NULL); | |
4157 | } | |
4158 | EXPORT_SYMBOL_GPL(__wake_up_locked); | |
4159 | ||
4160 | void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key) | |
4161 | { | |
4162 | __wake_up_common(q, mode, 1, 0, key); | |
4163 | } | |
4164 | ||
4165 | /** | |
4166 | * __wake_up_sync_key - wake up threads blocked on a waitqueue. | |
4167 | * @q: the waitqueue | |
4168 | * @mode: which threads | |
4169 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
4170 | * @key: opaque value to be passed to wakeup targets | |
4171 | * | |
4172 | * The sync wakeup differs that the waker knows that it will schedule | |
4173 | * away soon, so while the target thread will be woken up, it will not | |
4174 | * be migrated to another CPU - ie. the two threads are 'synchronized' | |
4175 | * with each other. This can prevent needless bouncing between CPUs. | |
4176 | * | |
4177 | * On UP it can prevent extra preemption. | |
4178 | * | |
4179 | * It may be assumed that this function implies a write memory barrier before | |
4180 | * changing the task state if and only if any tasks are woken up. | |
4181 | */ | |
4182 | void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode, | |
4183 | int nr_exclusive, void *key) | |
4184 | { | |
4185 | unsigned long flags; | |
4186 | int wake_flags = WF_SYNC; | |
4187 | ||
4188 | if (unlikely(!q)) | |
4189 | return; | |
4190 | ||
4191 | if (unlikely(!nr_exclusive)) | |
4192 | wake_flags = 0; | |
4193 | ||
4194 | spin_lock_irqsave(&q->lock, flags); | |
4195 | __wake_up_common(q, mode, nr_exclusive, wake_flags, key); | |
4196 | spin_unlock_irqrestore(&q->lock, flags); | |
4197 | } | |
4198 | EXPORT_SYMBOL_GPL(__wake_up_sync_key); | |
4199 | ||
4200 | /* | |
4201 | * __wake_up_sync - see __wake_up_sync_key() | |
4202 | */ | |
4203 | void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) | |
4204 | { | |
4205 | __wake_up_sync_key(q, mode, nr_exclusive, NULL); | |
4206 | } | |
4207 | EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ | |
4208 | ||
4209 | /** | |
4210 | * complete: - signals a single thread waiting on this completion | |
4211 | * @x: holds the state of this particular completion | |
4212 | * | |
4213 | * This will wake up a single thread waiting on this completion. Threads will be | |
4214 | * awakened in the same order in which they were queued. | |
4215 | * | |
4216 | * See also complete_all(), wait_for_completion() and related routines. | |
4217 | * | |
4218 | * It may be assumed that this function implies a write memory barrier before | |
4219 | * changing the task state if and only if any tasks are woken up. | |
4220 | */ | |
4221 | void complete(struct completion *x) | |
4222 | { | |
4223 | unsigned long flags; | |
4224 | ||
4225 | spin_lock_irqsave(&x->wait.lock, flags); | |
4226 | x->done++; | |
4227 | __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); | |
4228 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
4229 | } | |
4230 | EXPORT_SYMBOL(complete); | |
4231 | ||
4232 | /** | |
4233 | * complete_all: - signals all threads waiting on this completion | |
4234 | * @x: holds the state of this particular completion | |
4235 | * | |
4236 | * This will wake up all threads waiting on this particular completion event. | |
4237 | * | |
4238 | * It may be assumed that this function implies a write memory barrier before | |
4239 | * changing the task state if and only if any tasks are woken up. | |
4240 | */ | |
4241 | void complete_all(struct completion *x) | |
4242 | { | |
4243 | unsigned long flags; | |
4244 | ||
4245 | spin_lock_irqsave(&x->wait.lock, flags); | |
4246 | x->done += UINT_MAX/2; | |
4247 | __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); | |
4248 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
4249 | } | |
4250 | EXPORT_SYMBOL(complete_all); | |
4251 | ||
4252 | static inline long __sched | |
4253 | do_wait_for_common(struct completion *x, long timeout, int state) | |
4254 | { | |
4255 | if (!x->done) { | |
4256 | DECLARE_WAITQUEUE(wait, current); | |
4257 | ||
4258 | __add_wait_queue_tail_exclusive(&x->wait, &wait); | |
4259 | do { | |
4260 | if (signal_pending_state(state, current)) { | |
4261 | timeout = -ERESTARTSYS; | |
4262 | break; | |
4263 | } | |
4264 | __set_current_state(state); | |
4265 | spin_unlock_irq(&x->wait.lock); | |
4266 | timeout = schedule_timeout(timeout); | |
4267 | spin_lock_irq(&x->wait.lock); | |
4268 | } while (!x->done && timeout); | |
4269 | __remove_wait_queue(&x->wait, &wait); | |
4270 | if (!x->done) | |
4271 | return timeout; | |
4272 | } | |
4273 | x->done--; | |
4274 | return timeout ?: 1; | |
4275 | } | |
4276 | ||
4277 | static long __sched | |
4278 | wait_for_common(struct completion *x, long timeout, int state) | |
4279 | { | |
4280 | might_sleep(); | |
4281 | ||
4282 | spin_lock_irq(&x->wait.lock); | |
4283 | timeout = do_wait_for_common(x, timeout, state); | |
4284 | spin_unlock_irq(&x->wait.lock); | |
4285 | return timeout; | |
4286 | } | |
4287 | ||
4288 | /** | |
4289 | * wait_for_completion: - waits for completion of a task | |
4290 | * @x: holds the state of this particular completion | |
4291 | * | |
4292 | * This waits to be signaled for completion of a specific task. It is NOT | |
4293 | * interruptible and there is no timeout. | |
4294 | * | |
4295 | * See also similar routines (i.e. wait_for_completion_timeout()) with timeout | |
4296 | * and interrupt capability. Also see complete(). | |
4297 | */ | |
4298 | void __sched wait_for_completion(struct completion *x) | |
4299 | { | |
4300 | wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); | |
4301 | } | |
4302 | EXPORT_SYMBOL(wait_for_completion); | |
4303 | ||
4304 | /** | |
4305 | * wait_for_completion_timeout: - waits for completion of a task (w/timeout) | |
4306 | * @x: holds the state of this particular completion | |
4307 | * @timeout: timeout value in jiffies | |
4308 | * | |
4309 | * This waits for either a completion of a specific task to be signaled or for a | |
4310 | * specified timeout to expire. The timeout is in jiffies. It is not | |
4311 | * interruptible. | |
4312 | */ | |
4313 | unsigned long __sched | |
4314 | wait_for_completion_timeout(struct completion *x, unsigned long timeout) | |
4315 | { | |
4316 | return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); | |
4317 | } | |
4318 | EXPORT_SYMBOL(wait_for_completion_timeout); | |
4319 | ||
4320 | /** | |
4321 | * wait_for_completion_interruptible: - waits for completion of a task (w/intr) | |
4322 | * @x: holds the state of this particular completion | |
4323 | * | |
4324 | * This waits for completion of a specific task to be signaled. It is | |
4325 | * interruptible. | |
4326 | */ | |
4327 | int __sched wait_for_completion_interruptible(struct completion *x) | |
4328 | { | |
4329 | long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); | |
4330 | if (t == -ERESTARTSYS) | |
4331 | return t; | |
4332 | return 0; | |
4333 | } | |
4334 | EXPORT_SYMBOL(wait_for_completion_interruptible); | |
4335 | ||
4336 | /** | |
4337 | * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) | |
4338 | * @x: holds the state of this particular completion | |
4339 | * @timeout: timeout value in jiffies | |
4340 | * | |
4341 | * This waits for either a completion of a specific task to be signaled or for a | |
4342 | * specified timeout to expire. It is interruptible. The timeout is in jiffies. | |
4343 | */ | |
4344 | unsigned long __sched | |
4345 | wait_for_completion_interruptible_timeout(struct completion *x, | |
4346 | unsigned long timeout) | |
4347 | { | |
4348 | return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); | |
4349 | } | |
4350 | EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); | |
4351 | ||
4352 | /** | |
4353 | * wait_for_completion_killable: - waits for completion of a task (killable) | |
4354 | * @x: holds the state of this particular completion | |
4355 | * | |
4356 | * This waits to be signaled for completion of a specific task. It can be | |
4357 | * interrupted by a kill signal. | |
4358 | */ | |
4359 | int __sched wait_for_completion_killable(struct completion *x) | |
4360 | { | |
4361 | long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); | |
4362 | if (t == -ERESTARTSYS) | |
4363 | return t; | |
4364 | return 0; | |
4365 | } | |
4366 | EXPORT_SYMBOL(wait_for_completion_killable); | |
4367 | ||
4368 | /** | |
4369 | * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable)) | |
4370 | * @x: holds the state of this particular completion | |
4371 | * @timeout: timeout value in jiffies | |
4372 | * | |
4373 | * This waits for either a completion of a specific task to be | |
4374 | * signaled or for a specified timeout to expire. It can be | |
4375 | * interrupted by a kill signal. The timeout is in jiffies. | |
4376 | */ | |
4377 | unsigned long __sched | |
4378 | wait_for_completion_killable_timeout(struct completion *x, | |
4379 | unsigned long timeout) | |
4380 | { | |
4381 | return wait_for_common(x, timeout, TASK_KILLABLE); | |
4382 | } | |
4383 | EXPORT_SYMBOL(wait_for_completion_killable_timeout); | |
4384 | ||
4385 | /** | |
4386 | * try_wait_for_completion - try to decrement a completion without blocking | |
4387 | * @x: completion structure | |
4388 | * | |
4389 | * Returns: 0 if a decrement cannot be done without blocking | |
4390 | * 1 if a decrement succeeded. | |
4391 | * | |
4392 | * If a completion is being used as a counting completion, | |
4393 | * attempt to decrement the counter without blocking. This | |
4394 | * enables us to avoid waiting if the resource the completion | |
4395 | * is protecting is not available. | |
4396 | */ | |
4397 | bool try_wait_for_completion(struct completion *x) | |
4398 | { | |
4399 | unsigned long flags; | |
4400 | int ret = 1; | |
4401 | ||
4402 | spin_lock_irqsave(&x->wait.lock, flags); | |
4403 | if (!x->done) | |
4404 | ret = 0; | |
4405 | else | |
4406 | x->done--; | |
4407 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
4408 | return ret; | |
4409 | } | |
4410 | EXPORT_SYMBOL(try_wait_for_completion); | |
4411 | ||
4412 | /** | |
4413 | * completion_done - Test to see if a completion has any waiters | |
4414 | * @x: completion structure | |
4415 | * | |
4416 | * Returns: 0 if there are waiters (wait_for_completion() in progress) | |
4417 | * 1 if there are no waiters. | |
4418 | * | |
4419 | */ | |
4420 | bool completion_done(struct completion *x) | |
4421 | { | |
4422 | unsigned long flags; | |
4423 | int ret = 1; | |
4424 | ||
4425 | spin_lock_irqsave(&x->wait.lock, flags); | |
4426 | if (!x->done) | |
4427 | ret = 0; | |
4428 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
4429 | return ret; | |
4430 | } | |
4431 | EXPORT_SYMBOL(completion_done); | |
4432 | ||
4433 | static long __sched | |
4434 | sleep_on_common(wait_queue_head_t *q, int state, long timeout) | |
4435 | { | |
4436 | unsigned long flags; | |
4437 | wait_queue_t wait; | |
4438 | ||
4439 | init_waitqueue_entry(&wait, current); | |
4440 | ||
4441 | __set_current_state(state); | |
4442 | ||
4443 | spin_lock_irqsave(&q->lock, flags); | |
4444 | __add_wait_queue(q, &wait); | |
4445 | spin_unlock(&q->lock); | |
4446 | timeout = schedule_timeout(timeout); | |
4447 | spin_lock_irq(&q->lock); | |
4448 | __remove_wait_queue(q, &wait); | |
4449 | spin_unlock_irqrestore(&q->lock, flags); | |
4450 | ||
4451 | return timeout; | |
4452 | } | |
4453 | ||
4454 | void __sched interruptible_sleep_on(wait_queue_head_t *q) | |
4455 | { | |
4456 | sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); | |
4457 | } | |
4458 | EXPORT_SYMBOL(interruptible_sleep_on); | |
4459 | ||
4460 | long __sched | |
4461 | interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) | |
4462 | { | |
4463 | return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); | |
4464 | } | |
4465 | EXPORT_SYMBOL(interruptible_sleep_on_timeout); | |
4466 | ||
4467 | void __sched sleep_on(wait_queue_head_t *q) | |
4468 | { | |
4469 | sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); | |
4470 | } | |
4471 | EXPORT_SYMBOL(sleep_on); | |
4472 | ||
4473 | long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) | |
4474 | { | |
4475 | return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); | |
4476 | } | |
4477 | EXPORT_SYMBOL(sleep_on_timeout); | |
4478 | ||
4479 | #ifdef CONFIG_RT_MUTEXES | |
4480 | ||
4481 | /* | |
4482 | * rt_mutex_setprio - set the current priority of a task | |
4483 | * @p: task | |
4484 | * @prio: prio value (kernel-internal form) | |
4485 | * | |
4486 | * This function changes the 'effective' priority of a task. It does | |
4487 | * not touch ->normal_prio like __setscheduler(). | |
4488 | * | |
4489 | * Used by the rt_mutex code to implement priority inheritance logic. | |
4490 | */ | |
4491 | void rt_mutex_setprio(struct task_struct *p, int prio) | |
4492 | { | |
4493 | unsigned long flags; | |
4494 | int oldprio, on_rq, running; | |
4495 | struct rq *rq; | |
4496 | const struct sched_class *prev_class; | |
4497 | ||
4498 | BUG_ON(prio < 0 || prio > MAX_PRIO); | |
4499 | ||
4500 | rq = task_rq_lock(p, &flags); | |
4501 | ||
4502 | trace_sched_pi_setprio(p, prio); | |
4503 | oldprio = p->prio; | |
4504 | prev_class = p->sched_class; | |
4505 | on_rq = p->se.on_rq; | |
4506 | running = task_current(rq, p); | |
4507 | if (on_rq) | |
4508 | dequeue_task(rq, p, 0); | |
4509 | if (running) | |
4510 | p->sched_class->put_prev_task(rq, p); | |
4511 | ||
4512 | if (rt_prio(prio)) | |
4513 | p->sched_class = &rt_sched_class; | |
4514 | else | |
4515 | p->sched_class = &fair_sched_class; | |
4516 | ||
4517 | p->prio = prio; | |
4518 | ||
4519 | if (running) | |
4520 | p->sched_class->set_curr_task(rq); | |
4521 | if (on_rq) { | |
4522 | enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0); | |
4523 | ||
4524 | check_class_changed(rq, p, prev_class, oldprio, running); | |
4525 | } | |
4526 | task_rq_unlock(rq, &flags); | |
4527 | } | |
4528 | ||
4529 | #endif | |
4530 | ||
4531 | void set_user_nice(struct task_struct *p, long nice) | |
4532 | { | |
4533 | int old_prio, delta, on_rq; | |
4534 | unsigned long flags; | |
4535 | struct rq *rq; | |
4536 | ||
4537 | if (TASK_NICE(p) == nice || nice < -20 || nice > 19) | |
4538 | return; | |
4539 | /* | |
4540 | * We have to be careful, if called from sys_setpriority(), | |
4541 | * the task might be in the middle of scheduling on another CPU. | |
4542 | */ | |
4543 | rq = task_rq_lock(p, &flags); | |
4544 | /* | |
4545 | * The RT priorities are set via sched_setscheduler(), but we still | |
4546 | * allow the 'normal' nice value to be set - but as expected | |
4547 | * it wont have any effect on scheduling until the task is | |
4548 | * SCHED_FIFO/SCHED_RR: | |
4549 | */ | |
4550 | if (task_has_rt_policy(p)) { | |
4551 | p->static_prio = NICE_TO_PRIO(nice); | |
4552 | goto out_unlock; | |
4553 | } | |
4554 | on_rq = p->se.on_rq; | |
4555 | if (on_rq) | |
4556 | dequeue_task(rq, p, 0); | |
4557 | ||
4558 | p->static_prio = NICE_TO_PRIO(nice); | |
4559 | set_load_weight(p); | |
4560 | old_prio = p->prio; | |
4561 | p->prio = effective_prio(p); | |
4562 | delta = p->prio - old_prio; | |
4563 | ||
4564 | if (on_rq) { | |
4565 | enqueue_task(rq, p, 0); | |
4566 | /* | |
4567 | * If the task increased its priority or is running and | |
4568 | * lowered its priority, then reschedule its CPU: | |
4569 | */ | |
4570 | if (delta < 0 || (delta > 0 && task_running(rq, p))) | |
4571 | resched_task(rq->curr); | |
4572 | } | |
4573 | out_unlock: | |
4574 | task_rq_unlock(rq, &flags); | |
4575 | } | |
4576 | EXPORT_SYMBOL(set_user_nice); | |
4577 | ||
4578 | /* | |
4579 | * can_nice - check if a task can reduce its nice value | |
4580 | * @p: task | |
4581 | * @nice: nice value | |
4582 | */ | |
4583 | int can_nice(const struct task_struct *p, const int nice) | |
4584 | { | |
4585 | /* convert nice value [19,-20] to rlimit style value [1,40] */ | |
4586 | int nice_rlim = 20 - nice; | |
4587 | ||
4588 | return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || | |
4589 | capable(CAP_SYS_NICE)); | |
4590 | } | |
4591 | ||
4592 | #ifdef __ARCH_WANT_SYS_NICE | |
4593 | ||
4594 | /* | |
4595 | * sys_nice - change the priority of the current process. | |
4596 | * @increment: priority increment | |
4597 | * | |
4598 | * sys_setpriority is a more generic, but much slower function that | |
4599 | * does similar things. | |
4600 | */ | |
4601 | SYSCALL_DEFINE1(nice, int, increment) | |
4602 | { | |
4603 | long nice, retval; | |
4604 | ||
4605 | /* | |
4606 | * Setpriority might change our priority at the same moment. | |
4607 | * We don't have to worry. Conceptually one call occurs first | |
4608 | * and we have a single winner. | |
4609 | */ | |
4610 | if (increment < -40) | |
4611 | increment = -40; | |
4612 | if (increment > 40) | |
4613 | increment = 40; | |
4614 | ||
4615 | nice = TASK_NICE(current) + increment; | |
4616 | if (nice < -20) | |
4617 | nice = -20; | |
4618 | if (nice > 19) | |
4619 | nice = 19; | |
4620 | ||
4621 | if (increment < 0 && !can_nice(current, nice)) | |
4622 | return -EPERM; | |
4623 | ||
4624 | retval = security_task_setnice(current, nice); | |
4625 | if (retval) | |
4626 | return retval; | |
4627 | ||
4628 | set_user_nice(current, nice); | |
4629 | return 0; | |
4630 | } | |
4631 | ||
4632 | #endif | |
4633 | ||
4634 | /** | |
4635 | * task_prio - return the priority value of a given task. | |
4636 | * @p: the task in question. | |
4637 | * | |
4638 | * This is the priority value as seen by users in /proc. | |
4639 | * RT tasks are offset by -200. Normal tasks are centered | |
4640 | * around 0, value goes from -16 to +15. | |
4641 | */ | |
4642 | int task_prio(const struct task_struct *p) | |
4643 | { | |
4644 | return p->prio - MAX_RT_PRIO; | |
4645 | } | |
4646 | ||
4647 | /** | |
4648 | * task_nice - return the nice value of a given task. | |
4649 | * @p: the task in question. | |
4650 | */ | |
4651 | int task_nice(const struct task_struct *p) | |
4652 | { | |
4653 | return TASK_NICE(p); | |
4654 | } | |
4655 | EXPORT_SYMBOL(task_nice); | |
4656 | ||
4657 | /** | |
4658 | * idle_cpu - is a given cpu idle currently? | |
4659 | * @cpu: the processor in question. | |
4660 | */ | |
4661 | int idle_cpu(int cpu) | |
4662 | { | |
4663 | return cpu_curr(cpu) == cpu_rq(cpu)->idle; | |
4664 | } | |
4665 | ||
4666 | /** | |
4667 | * idle_task - return the idle task for a given cpu. | |
4668 | * @cpu: the processor in question. | |
4669 | */ | |
4670 | struct task_struct *idle_task(int cpu) | |
4671 | { | |
4672 | return cpu_rq(cpu)->idle; | |
4673 | } | |
4674 | ||
4675 | /** | |
4676 | * find_process_by_pid - find a process with a matching PID value. | |
4677 | * @pid: the pid in question. | |
4678 | */ | |
4679 | static struct task_struct *find_process_by_pid(pid_t pid) | |
4680 | { | |
4681 | return pid ? find_task_by_vpid(pid) : current; | |
4682 | } | |
4683 | ||
4684 | /* Actually do priority change: must hold rq lock. */ | |
4685 | static void | |
4686 | __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) | |
4687 | { | |
4688 | BUG_ON(p->se.on_rq); | |
4689 | ||
4690 | p->policy = policy; | |
4691 | p->rt_priority = prio; | |
4692 | p->normal_prio = normal_prio(p); | |
4693 | /* we are holding p->pi_lock already */ | |
4694 | p->prio = rt_mutex_getprio(p); | |
4695 | if (rt_prio(p->prio)) | |
4696 | p->sched_class = &rt_sched_class; | |
4697 | else | |
4698 | p->sched_class = &fair_sched_class; | |
4699 | set_load_weight(p); | |
4700 | } | |
4701 | ||
4702 | /* | |
4703 | * check the target process has a UID that matches the current process's | |
4704 | */ | |
4705 | static bool check_same_owner(struct task_struct *p) | |
4706 | { | |
4707 | const struct cred *cred = current_cred(), *pcred; | |
4708 | bool match; | |
4709 | ||
4710 | rcu_read_lock(); | |
4711 | pcred = __task_cred(p); | |
4712 | match = (cred->euid == pcred->euid || | |
4713 | cred->euid == pcred->uid); | |
4714 | rcu_read_unlock(); | |
4715 | return match; | |
4716 | } | |
4717 | ||
4718 | static int __sched_setscheduler(struct task_struct *p, int policy, | |
4719 | struct sched_param *param, bool user) | |
4720 | { | |
4721 | int retval, oldprio, oldpolicy = -1, on_rq, running; | |
4722 | unsigned long flags; | |
4723 | const struct sched_class *prev_class; | |
4724 | struct rq *rq; | |
4725 | int reset_on_fork; | |
4726 | ||
4727 | /* may grab non-irq protected spin_locks */ | |
4728 | BUG_ON(in_interrupt()); | |
4729 | recheck: | |
4730 | /* double check policy once rq lock held */ | |
4731 | if (policy < 0) { | |
4732 | reset_on_fork = p->sched_reset_on_fork; | |
4733 | policy = oldpolicy = p->policy; | |
4734 | } else { | |
4735 | reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); | |
4736 | policy &= ~SCHED_RESET_ON_FORK; | |
4737 | ||
4738 | if (policy != SCHED_FIFO && policy != SCHED_RR && | |
4739 | policy != SCHED_NORMAL && policy != SCHED_BATCH && | |
4740 | policy != SCHED_IDLE) | |
4741 | return -EINVAL; | |
4742 | } | |
4743 | ||
4744 | /* | |
4745 | * Valid priorities for SCHED_FIFO and SCHED_RR are | |
4746 | * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, | |
4747 | * SCHED_BATCH and SCHED_IDLE is 0. | |
4748 | */ | |
4749 | if (param->sched_priority < 0 || | |
4750 | (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || | |
4751 | (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) | |
4752 | return -EINVAL; | |
4753 | if (rt_policy(policy) != (param->sched_priority != 0)) | |
4754 | return -EINVAL; | |
4755 | ||
4756 | /* | |
4757 | * Allow unprivileged RT tasks to decrease priority: | |
4758 | */ | |
4759 | if (user && !capable(CAP_SYS_NICE)) { | |
4760 | if (rt_policy(policy)) { | |
4761 | unsigned long rlim_rtprio = | |
4762 | task_rlimit(p, RLIMIT_RTPRIO); | |
4763 | ||
4764 | /* can't set/change the rt policy */ | |
4765 | if (policy != p->policy && !rlim_rtprio) | |
4766 | return -EPERM; | |
4767 | ||
4768 | /* can't increase priority */ | |
4769 | if (param->sched_priority > p->rt_priority && | |
4770 | param->sched_priority > rlim_rtprio) | |
4771 | return -EPERM; | |
4772 | } | |
4773 | /* | |
4774 | * Like positive nice levels, dont allow tasks to | |
4775 | * move out of SCHED_IDLE either: | |
4776 | */ | |
4777 | if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) | |
4778 | return -EPERM; | |
4779 | ||
4780 | /* can't change other user's priorities */ | |
4781 | if (!check_same_owner(p)) | |
4782 | return -EPERM; | |
4783 | ||
4784 | /* Normal users shall not reset the sched_reset_on_fork flag */ | |
4785 | if (p->sched_reset_on_fork && !reset_on_fork) | |
4786 | return -EPERM; | |
4787 | } | |
4788 | ||
4789 | if (user) { | |
4790 | retval = security_task_setscheduler(p); | |
4791 | if (retval) | |
4792 | return retval; | |
4793 | } | |
4794 | ||
4795 | /* | |
4796 | * make sure no PI-waiters arrive (or leave) while we are | |
4797 | * changing the priority of the task: | |
4798 | */ | |
4799 | raw_spin_lock_irqsave(&p->pi_lock, flags); | |
4800 | /* | |
4801 | * To be able to change p->policy safely, the apropriate | |
4802 | * runqueue lock must be held. | |
4803 | */ | |
4804 | rq = __task_rq_lock(p); | |
4805 | ||
4806 | /* | |
4807 | * Changing the policy of the stop threads its a very bad idea | |
4808 | */ | |
4809 | if (p == rq->stop) { | |
4810 | __task_rq_unlock(rq); | |
4811 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); | |
4812 | return -EINVAL; | |
4813 | } | |
4814 | ||
4815 | #ifdef CONFIG_RT_GROUP_SCHED | |
4816 | if (user) { | |
4817 | /* | |
4818 | * Do not allow realtime tasks into groups that have no runtime | |
4819 | * assigned. | |
4820 | */ | |
4821 | if (rt_bandwidth_enabled() && rt_policy(policy) && | |
4822 | task_group(p)->rt_bandwidth.rt_runtime == 0) { | |
4823 | __task_rq_unlock(rq); | |
4824 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); | |
4825 | return -EPERM; | |
4826 | } | |
4827 | } | |
4828 | #endif | |
4829 | ||
4830 | /* recheck policy now with rq lock held */ | |
4831 | if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { | |
4832 | policy = oldpolicy = -1; | |
4833 | __task_rq_unlock(rq); | |
4834 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); | |
4835 | goto recheck; | |
4836 | } | |
4837 | on_rq = p->se.on_rq; | |
4838 | running = task_current(rq, p); | |
4839 | if (on_rq) | |
4840 | deactivate_task(rq, p, 0); | |
4841 | if (running) | |
4842 | p->sched_class->put_prev_task(rq, p); | |
4843 | ||
4844 | p->sched_reset_on_fork = reset_on_fork; | |
4845 | ||
4846 | oldprio = p->prio; | |
4847 | prev_class = p->sched_class; | |
4848 | __setscheduler(rq, p, policy, param->sched_priority); | |
4849 | ||
4850 | if (running) | |
4851 | p->sched_class->set_curr_task(rq); | |
4852 | if (on_rq) { | |
4853 | activate_task(rq, p, 0); | |
4854 | ||
4855 | check_class_changed(rq, p, prev_class, oldprio, running); | |
4856 | } | |
4857 | __task_rq_unlock(rq); | |
4858 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); | |
4859 | ||
4860 | rt_mutex_adjust_pi(p); | |
4861 | ||
4862 | return 0; | |
4863 | } | |
4864 | ||
4865 | /** | |
4866 | * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. | |
4867 | * @p: the task in question. | |
4868 | * @policy: new policy. | |
4869 | * @param: structure containing the new RT priority. | |
4870 | * | |
4871 | * NOTE that the task may be already dead. | |
4872 | */ | |
4873 | int sched_setscheduler(struct task_struct *p, int policy, | |
4874 | struct sched_param *param) | |
4875 | { | |
4876 | return __sched_setscheduler(p, policy, param, true); | |
4877 | } | |
4878 | EXPORT_SYMBOL_GPL(sched_setscheduler); | |
4879 | ||
4880 | /** | |
4881 | * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. | |
4882 | * @p: the task in question. | |
4883 | * @policy: new policy. | |
4884 | * @param: structure containing the new RT priority. | |
4885 | * | |
4886 | * Just like sched_setscheduler, only don't bother checking if the | |
4887 | * current context has permission. For example, this is needed in | |
4888 | * stop_machine(): we create temporary high priority worker threads, | |
4889 | * but our caller might not have that capability. | |
4890 | */ | |
4891 | int sched_setscheduler_nocheck(struct task_struct *p, int policy, | |
4892 | struct sched_param *param) | |
4893 | { | |
4894 | return __sched_setscheduler(p, policy, param, false); | |
4895 | } | |
4896 | ||
4897 | static int | |
4898 | do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) | |
4899 | { | |
4900 | struct sched_param lparam; | |
4901 | struct task_struct *p; | |
4902 | int retval; | |
4903 | ||
4904 | if (!param || pid < 0) | |
4905 | return -EINVAL; | |
4906 | if (copy_from_user(&lparam, param, sizeof(struct sched_param))) | |
4907 | return -EFAULT; | |
4908 | ||
4909 | rcu_read_lock(); | |
4910 | retval = -ESRCH; | |
4911 | p = find_process_by_pid(pid); | |
4912 | if (p != NULL) | |
4913 | retval = sched_setscheduler(p, policy, &lparam); | |
4914 | rcu_read_unlock(); | |
4915 | ||
4916 | return retval; | |
4917 | } | |
4918 | ||
4919 | /** | |
4920 | * sys_sched_setscheduler - set/change the scheduler policy and RT priority | |
4921 | * @pid: the pid in question. | |
4922 | * @policy: new policy. | |
4923 | * @param: structure containing the new RT priority. | |
4924 | */ | |
4925 | SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, | |
4926 | struct sched_param __user *, param) | |
4927 | { | |
4928 | /* negative values for policy are not valid */ | |
4929 | if (policy < 0) | |
4930 | return -EINVAL; | |
4931 | ||
4932 | return do_sched_setscheduler(pid, policy, param); | |
4933 | } | |
4934 | ||
4935 | /** | |
4936 | * sys_sched_setparam - set/change the RT priority of a thread | |
4937 | * @pid: the pid in question. | |
4938 | * @param: structure containing the new RT priority. | |
4939 | */ | |
4940 | SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) | |
4941 | { | |
4942 | return do_sched_setscheduler(pid, -1, param); | |
4943 | } | |
4944 | ||
4945 | /** | |
4946 | * sys_sched_getscheduler - get the policy (scheduling class) of a thread | |
4947 | * @pid: the pid in question. | |
4948 | */ | |
4949 | SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) | |
4950 | { | |
4951 | struct task_struct *p; | |
4952 | int retval; | |
4953 | ||
4954 | if (pid < 0) | |
4955 | return -EINVAL; | |
4956 | ||
4957 | retval = -ESRCH; | |
4958 | rcu_read_lock(); | |
4959 | p = find_process_by_pid(pid); | |
4960 | if (p) { | |
4961 | retval = security_task_getscheduler(p); | |
4962 | if (!retval) | |
4963 | retval = p->policy | |
4964 | | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); | |
4965 | } | |
4966 | rcu_read_unlock(); | |
4967 | return retval; | |
4968 | } | |
4969 | ||
4970 | /** | |
4971 | * sys_sched_getparam - get the RT priority of a thread | |
4972 | * @pid: the pid in question. | |
4973 | * @param: structure containing the RT priority. | |
4974 | */ | |
4975 | SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) | |
4976 | { | |
4977 | struct sched_param lp; | |
4978 | struct task_struct *p; | |
4979 | int retval; | |
4980 | ||
4981 | if (!param || pid < 0) | |
4982 | return -EINVAL; | |
4983 | ||
4984 | rcu_read_lock(); | |
4985 | p = find_process_by_pid(pid); | |
4986 | retval = -ESRCH; | |
4987 | if (!p) | |
4988 | goto out_unlock; | |
4989 | ||
4990 | retval = security_task_getscheduler(p); | |
4991 | if (retval) | |
4992 | goto out_unlock; | |
4993 | ||
4994 | lp.sched_priority = p->rt_priority; | |
4995 | rcu_read_unlock(); | |
4996 | ||
4997 | /* | |
4998 | * This one might sleep, we cannot do it with a spinlock held ... | |
4999 | */ | |
5000 | retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; | |
5001 | ||
5002 | return retval; | |
5003 | ||
5004 | out_unlock: | |
5005 | rcu_read_unlock(); | |
5006 | return retval; | |
5007 | } | |
5008 | ||
5009 | long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) | |
5010 | { | |
5011 | cpumask_var_t cpus_allowed, new_mask; | |
5012 | struct task_struct *p; | |
5013 | int retval; | |
5014 | ||
5015 | get_online_cpus(); | |
5016 | rcu_read_lock(); | |
5017 | ||
5018 | p = find_process_by_pid(pid); | |
5019 | if (!p) { | |
5020 | rcu_read_unlock(); | |
5021 | put_online_cpus(); | |
5022 | return -ESRCH; | |
5023 | } | |
5024 | ||
5025 | /* Prevent p going away */ | |
5026 | get_task_struct(p); | |
5027 | rcu_read_unlock(); | |
5028 | ||
5029 | if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { | |
5030 | retval = -ENOMEM; | |
5031 | goto out_put_task; | |
5032 | } | |
5033 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { | |
5034 | retval = -ENOMEM; | |
5035 | goto out_free_cpus_allowed; | |
5036 | } | |
5037 | retval = -EPERM; | |
5038 | if (!check_same_owner(p) && !capable(CAP_SYS_NICE)) | |
5039 | goto out_unlock; | |
5040 | ||
5041 | retval = security_task_setscheduler(p); | |
5042 | if (retval) | |
5043 | goto out_unlock; | |
5044 | ||
5045 | cpuset_cpus_allowed(p, cpus_allowed); | |
5046 | cpumask_and(new_mask, in_mask, cpus_allowed); | |
5047 | again: | |
5048 | retval = set_cpus_allowed_ptr(p, new_mask); | |
5049 | ||
5050 | if (!retval) { | |
5051 | cpuset_cpus_allowed(p, cpus_allowed); | |
5052 | if (!cpumask_subset(new_mask, cpus_allowed)) { | |
5053 | /* | |
5054 | * We must have raced with a concurrent cpuset | |
5055 | * update. Just reset the cpus_allowed to the | |
5056 | * cpuset's cpus_allowed | |
5057 | */ | |
5058 | cpumask_copy(new_mask, cpus_allowed); | |
5059 | goto again; | |
5060 | } | |
5061 | } | |
5062 | out_unlock: | |
5063 | free_cpumask_var(new_mask); | |
5064 | out_free_cpus_allowed: | |
5065 | free_cpumask_var(cpus_allowed); | |
5066 | out_put_task: | |
5067 | put_task_struct(p); | |
5068 | put_online_cpus(); | |
5069 | return retval; | |
5070 | } | |
5071 | ||
5072 | static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, | |
5073 | struct cpumask *new_mask) | |
5074 | { | |
5075 | if (len < cpumask_size()) | |
5076 | cpumask_clear(new_mask); | |
5077 | else if (len > cpumask_size()) | |
5078 | len = cpumask_size(); | |
5079 | ||
5080 | return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; | |
5081 | } | |
5082 | ||
5083 | /** | |
5084 | * sys_sched_setaffinity - set the cpu affinity of a process | |
5085 | * @pid: pid of the process | |
5086 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
5087 | * @user_mask_ptr: user-space pointer to the new cpu mask | |
5088 | */ | |
5089 | SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, | |
5090 | unsigned long __user *, user_mask_ptr) | |
5091 | { | |
5092 | cpumask_var_t new_mask; | |
5093 | int retval; | |
5094 | ||
5095 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) | |
5096 | return -ENOMEM; | |
5097 | ||
5098 | retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); | |
5099 | if (retval == 0) | |
5100 | retval = sched_setaffinity(pid, new_mask); | |
5101 | free_cpumask_var(new_mask); | |
5102 | return retval; | |
5103 | } | |
5104 | ||
5105 | long sched_getaffinity(pid_t pid, struct cpumask *mask) | |
5106 | { | |
5107 | struct task_struct *p; | |
5108 | unsigned long flags; | |
5109 | struct rq *rq; | |
5110 | int retval; | |
5111 | ||
5112 | get_online_cpus(); | |
5113 | rcu_read_lock(); | |
5114 | ||
5115 | retval = -ESRCH; | |
5116 | p = find_process_by_pid(pid); | |
5117 | if (!p) | |
5118 | goto out_unlock; | |
5119 | ||
5120 | retval = security_task_getscheduler(p); | |
5121 | if (retval) | |
5122 | goto out_unlock; | |
5123 | ||
5124 | rq = task_rq_lock(p, &flags); | |
5125 | cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); | |
5126 | task_rq_unlock(rq, &flags); | |
5127 | ||
5128 | out_unlock: | |
5129 | rcu_read_unlock(); | |
5130 | put_online_cpus(); | |
5131 | ||
5132 | return retval; | |
5133 | } | |
5134 | ||
5135 | /** | |
5136 | * sys_sched_getaffinity - get the cpu affinity of a process | |
5137 | * @pid: pid of the process | |
5138 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
5139 | * @user_mask_ptr: user-space pointer to hold the current cpu mask | |
5140 | */ | |
5141 | SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, | |
5142 | unsigned long __user *, user_mask_ptr) | |
5143 | { | |
5144 | int ret; | |
5145 | cpumask_var_t mask; | |
5146 | ||
5147 | if ((len * BITS_PER_BYTE) < nr_cpu_ids) | |
5148 | return -EINVAL; | |
5149 | if (len & (sizeof(unsigned long)-1)) | |
5150 | return -EINVAL; | |
5151 | ||
5152 | if (!alloc_cpumask_var(&mask, GFP_KERNEL)) | |
5153 | return -ENOMEM; | |
5154 | ||
5155 | ret = sched_getaffinity(pid, mask); | |
5156 | if (ret == 0) { | |
5157 | size_t retlen = min_t(size_t, len, cpumask_size()); | |
5158 | ||
5159 | if (copy_to_user(user_mask_ptr, mask, retlen)) | |
5160 | ret = -EFAULT; | |
5161 | else | |
5162 | ret = retlen; | |
5163 | } | |
5164 | free_cpumask_var(mask); | |
5165 | ||
5166 | return ret; | |
5167 | } | |
5168 | ||
5169 | /** | |
5170 | * sys_sched_yield - yield the current processor to other threads. | |
5171 | * | |
5172 | * This function yields the current CPU to other tasks. If there are no | |
5173 | * other threads running on this CPU then this function will return. | |
5174 | */ | |
5175 | SYSCALL_DEFINE0(sched_yield) | |
5176 | { | |
5177 | struct rq *rq = this_rq_lock(); | |
5178 | ||
5179 | schedstat_inc(rq, yld_count); | |
5180 | current->sched_class->yield_task(rq); | |
5181 | ||
5182 | /* | |
5183 | * Since we are going to call schedule() anyway, there's | |
5184 | * no need to preempt or enable interrupts: | |
5185 | */ | |
5186 | __release(rq->lock); | |
5187 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); | |
5188 | do_raw_spin_unlock(&rq->lock); | |
5189 | preempt_enable_no_resched(); | |
5190 | ||
5191 | schedule(); | |
5192 | ||
5193 | return 0; | |
5194 | } | |
5195 | ||
5196 | static inline int should_resched(void) | |
5197 | { | |
5198 | return need_resched() && !(preempt_count() & PREEMPT_ACTIVE); | |
5199 | } | |
5200 | ||
5201 | static void __cond_resched(void) | |
5202 | { | |
5203 | add_preempt_count(PREEMPT_ACTIVE); | |
5204 | schedule(); | |
5205 | sub_preempt_count(PREEMPT_ACTIVE); | |
5206 | } | |
5207 | ||
5208 | int __sched _cond_resched(void) | |
5209 | { | |
5210 | if (should_resched()) { | |
5211 | __cond_resched(); | |
5212 | return 1; | |
5213 | } | |
5214 | return 0; | |
5215 | } | |
5216 | EXPORT_SYMBOL(_cond_resched); | |
5217 | ||
5218 | /* | |
5219 | * __cond_resched_lock() - if a reschedule is pending, drop the given lock, | |
5220 | * call schedule, and on return reacquire the lock. | |
5221 | * | |
5222 | * This works OK both with and without CONFIG_PREEMPT. We do strange low-level | |
5223 | * operations here to prevent schedule() from being called twice (once via | |
5224 | * spin_unlock(), once by hand). | |
5225 | */ | |
5226 | int __cond_resched_lock(spinlock_t *lock) | |
5227 | { | |
5228 | int resched = should_resched(); | |
5229 | int ret = 0; | |
5230 | ||
5231 | lockdep_assert_held(lock); | |
5232 | ||
5233 | if (spin_needbreak(lock) || resched) { | |
5234 | spin_unlock(lock); | |
5235 | if (resched) | |
5236 | __cond_resched(); | |
5237 | else | |
5238 | cpu_relax(); | |
5239 | ret = 1; | |
5240 | spin_lock(lock); | |
5241 | } | |
5242 | return ret; | |
5243 | } | |
5244 | EXPORT_SYMBOL(__cond_resched_lock); | |
5245 | ||
5246 | int __sched __cond_resched_softirq(void) | |
5247 | { | |
5248 | BUG_ON(!in_softirq()); | |
5249 | ||
5250 | if (should_resched()) { | |
5251 | local_bh_enable(); | |
5252 | __cond_resched(); | |
5253 | local_bh_disable(); | |
5254 | return 1; | |
5255 | } | |
5256 | return 0; | |
5257 | } | |
5258 | EXPORT_SYMBOL(__cond_resched_softirq); | |
5259 | ||
5260 | /** | |
5261 | * yield - yield the current processor to other threads. | |
5262 | * | |
5263 | * This is a shortcut for kernel-space yielding - it marks the | |
5264 | * thread runnable and calls sys_sched_yield(). | |
5265 | */ | |
5266 | void __sched yield(void) | |
5267 | { | |
5268 | set_current_state(TASK_RUNNING); | |
5269 | sys_sched_yield(); | |
5270 | } | |
5271 | EXPORT_SYMBOL(yield); | |
5272 | ||
5273 | /* | |
5274 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so | |
5275 | * that process accounting knows that this is a task in IO wait state. | |
5276 | */ | |
5277 | void __sched io_schedule(void) | |
5278 | { | |
5279 | struct rq *rq = raw_rq(); | |
5280 | ||
5281 | delayacct_blkio_start(); | |
5282 | atomic_inc(&rq->nr_iowait); | |
5283 | current->in_iowait = 1; | |
5284 | schedule(); | |
5285 | current->in_iowait = 0; | |
5286 | atomic_dec(&rq->nr_iowait); | |
5287 | delayacct_blkio_end(); | |
5288 | } | |
5289 | EXPORT_SYMBOL(io_schedule); | |
5290 | ||
5291 | long __sched io_schedule_timeout(long timeout) | |
5292 | { | |
5293 | struct rq *rq = raw_rq(); | |
5294 | long ret; | |
5295 | ||
5296 | delayacct_blkio_start(); | |
5297 | atomic_inc(&rq->nr_iowait); | |
5298 | current->in_iowait = 1; | |
5299 | ret = schedule_timeout(timeout); | |
5300 | current->in_iowait = 0; | |
5301 | atomic_dec(&rq->nr_iowait); | |
5302 | delayacct_blkio_end(); | |
5303 | return ret; | |
5304 | } | |
5305 | ||
5306 | /** | |
5307 | * sys_sched_get_priority_max - return maximum RT priority. | |
5308 | * @policy: scheduling class. | |
5309 | * | |
5310 | * this syscall returns the maximum rt_priority that can be used | |
5311 | * by a given scheduling class. | |
5312 | */ | |
5313 | SYSCALL_DEFINE1(sched_get_priority_max, int, policy) | |
5314 | { | |
5315 | int ret = -EINVAL; | |
5316 | ||
5317 | switch (policy) { | |
5318 | case SCHED_FIFO: | |
5319 | case SCHED_RR: | |
5320 | ret = MAX_USER_RT_PRIO-1; | |
5321 | break; | |
5322 | case SCHED_NORMAL: | |
5323 | case SCHED_BATCH: | |
5324 | case SCHED_IDLE: | |
5325 | ret = 0; | |
5326 | break; | |
5327 | } | |
5328 | return ret; | |
5329 | } | |
5330 | ||
5331 | /** | |
5332 | * sys_sched_get_priority_min - return minimum RT priority. | |
5333 | * @policy: scheduling class. | |
5334 | * | |
5335 | * this syscall returns the minimum rt_priority that can be used | |
5336 | * by a given scheduling class. | |
5337 | */ | |
5338 | SYSCALL_DEFINE1(sched_get_priority_min, int, policy) | |
5339 | { | |
5340 | int ret = -EINVAL; | |
5341 | ||
5342 | switch (policy) { | |
5343 | case SCHED_FIFO: | |
5344 | case SCHED_RR: | |
5345 | ret = 1; | |
5346 | break; | |
5347 | case SCHED_NORMAL: | |
5348 | case SCHED_BATCH: | |
5349 | case SCHED_IDLE: | |
5350 | ret = 0; | |
5351 | } | |
5352 | return ret; | |
5353 | } | |
5354 | ||
5355 | /** | |
5356 | * sys_sched_rr_get_interval - return the default timeslice of a process. | |
5357 | * @pid: pid of the process. | |
5358 | * @interval: userspace pointer to the timeslice value. | |
5359 | * | |
5360 | * this syscall writes the default timeslice value of a given process | |
5361 | * into the user-space timespec buffer. A value of '0' means infinity. | |
5362 | */ | |
5363 | SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, | |
5364 | struct timespec __user *, interval) | |
5365 | { | |
5366 | struct task_struct *p; | |
5367 | unsigned int time_slice; | |
5368 | unsigned long flags; | |
5369 | struct rq *rq; | |
5370 | int retval; | |
5371 | struct timespec t; | |
5372 | ||
5373 | if (pid < 0) | |
5374 | return -EINVAL; | |
5375 | ||
5376 | retval = -ESRCH; | |
5377 | rcu_read_lock(); | |
5378 | p = find_process_by_pid(pid); | |
5379 | if (!p) | |
5380 | goto out_unlock; | |
5381 | ||
5382 | retval = security_task_getscheduler(p); | |
5383 | if (retval) | |
5384 | goto out_unlock; | |
5385 | ||
5386 | rq = task_rq_lock(p, &flags); | |
5387 | time_slice = p->sched_class->get_rr_interval(rq, p); | |
5388 | task_rq_unlock(rq, &flags); | |
5389 | ||
5390 | rcu_read_unlock(); | |
5391 | jiffies_to_timespec(time_slice, &t); | |
5392 | retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; | |
5393 | return retval; | |
5394 | ||
5395 | out_unlock: | |
5396 | rcu_read_unlock(); | |
5397 | return retval; | |
5398 | } | |
5399 | ||
5400 | static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; | |
5401 | ||
5402 | void sched_show_task(struct task_struct *p) | |
5403 | { | |
5404 | unsigned long free = 0; | |
5405 | unsigned state; | |
5406 | ||
5407 | state = p->state ? __ffs(p->state) + 1 : 0; | |
5408 | printk(KERN_INFO "%-13.13s %c", p->comm, | |
5409 | state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); | |
5410 | #if BITS_PER_LONG == 32 | |
5411 | if (state == TASK_RUNNING) | |
5412 | printk(KERN_CONT " running "); | |
5413 | else | |
5414 | printk(KERN_CONT " %08lx ", thread_saved_pc(p)); | |
5415 | #else | |
5416 | if (state == TASK_RUNNING) | |
5417 | printk(KERN_CONT " running task "); | |
5418 | else | |
5419 | printk(KERN_CONT " %016lx ", thread_saved_pc(p)); | |
5420 | #endif | |
5421 | #ifdef CONFIG_DEBUG_STACK_USAGE | |
5422 | free = stack_not_used(p); | |
5423 | #endif | |
5424 | printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, | |
5425 | task_pid_nr(p), task_pid_nr(p->real_parent), | |
5426 | (unsigned long)task_thread_info(p)->flags); | |
5427 | ||
5428 | show_stack(p, NULL); | |
5429 | } | |
5430 | ||
5431 | void show_state_filter(unsigned long state_filter) | |
5432 | { | |
5433 | struct task_struct *g, *p; | |
5434 | ||
5435 | #if BITS_PER_LONG == 32 | |
5436 | printk(KERN_INFO | |
5437 | " task PC stack pid father\n"); | |
5438 | #else | |
5439 | printk(KERN_INFO | |
5440 | " task PC stack pid father\n"); | |
5441 | #endif | |
5442 | read_lock(&tasklist_lock); | |
5443 | do_each_thread(g, p) { | |
5444 | /* | |
5445 | * reset the NMI-timeout, listing all files on a slow | |
5446 | * console might take alot of time: | |
5447 | */ | |
5448 | touch_nmi_watchdog(); | |
5449 | if (!state_filter || (p->state & state_filter)) | |
5450 | sched_show_task(p); | |
5451 | } while_each_thread(g, p); | |
5452 | ||
5453 | touch_all_softlockup_watchdogs(); | |
5454 | ||
5455 | #ifdef CONFIG_SCHED_DEBUG | |
5456 | sysrq_sched_debug_show(); | |
5457 | #endif | |
5458 | read_unlock(&tasklist_lock); | |
5459 | /* | |
5460 | * Only show locks if all tasks are dumped: | |
5461 | */ | |
5462 | if (!state_filter) | |
5463 | debug_show_all_locks(); | |
5464 | } | |
5465 | ||
5466 | void __cpuinit init_idle_bootup_task(struct task_struct *idle) | |
5467 | { | |
5468 | idle->sched_class = &idle_sched_class; | |
5469 | } | |
5470 | ||
5471 | /** | |
5472 | * init_idle - set up an idle thread for a given CPU | |
5473 | * @idle: task in question | |
5474 | * @cpu: cpu the idle task belongs to | |
5475 | * | |
5476 | * NOTE: this function does not set the idle thread's NEED_RESCHED | |
5477 | * flag, to make booting more robust. | |
5478 | */ | |
5479 | void __cpuinit init_idle(struct task_struct *idle, int cpu) | |
5480 | { | |
5481 | struct rq *rq = cpu_rq(cpu); | |
5482 | unsigned long flags; | |
5483 | ||
5484 | raw_spin_lock_irqsave(&rq->lock, flags); | |
5485 | ||
5486 | __sched_fork(idle); | |
5487 | idle->state = TASK_RUNNING; | |
5488 | idle->se.exec_start = sched_clock(); | |
5489 | ||
5490 | cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu)); | |
5491 | /* | |
5492 | * We're having a chicken and egg problem, even though we are | |
5493 | * holding rq->lock, the cpu isn't yet set to this cpu so the | |
5494 | * lockdep check in task_group() will fail. | |
5495 | * | |
5496 | * Similar case to sched_fork(). / Alternatively we could | |
5497 | * use task_rq_lock() here and obtain the other rq->lock. | |
5498 | * | |
5499 | * Silence PROVE_RCU | |
5500 | */ | |
5501 | rcu_read_lock(); | |
5502 | __set_task_cpu(idle, cpu); | |
5503 | rcu_read_unlock(); | |
5504 | ||
5505 | rq->curr = rq->idle = idle; | |
5506 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) | |
5507 | idle->oncpu = 1; | |
5508 | #endif | |
5509 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
5510 | ||
5511 | /* Set the preempt count _outside_ the spinlocks! */ | |
5512 | #if defined(CONFIG_PREEMPT) | |
5513 | task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0); | |
5514 | #else | |
5515 | task_thread_info(idle)->preempt_count = 0; | |
5516 | #endif | |
5517 | /* | |
5518 | * The idle tasks have their own, simple scheduling class: | |
5519 | */ | |
5520 | idle->sched_class = &idle_sched_class; | |
5521 | ftrace_graph_init_task(idle); | |
5522 | } | |
5523 | ||
5524 | /* | |
5525 | * In a system that switches off the HZ timer nohz_cpu_mask | |
5526 | * indicates which cpus entered this state. This is used | |
5527 | * in the rcu update to wait only for active cpus. For system | |
5528 | * which do not switch off the HZ timer nohz_cpu_mask should | |
5529 | * always be CPU_BITS_NONE. | |
5530 | */ | |
5531 | cpumask_var_t nohz_cpu_mask; | |
5532 | ||
5533 | /* | |
5534 | * Increase the granularity value when there are more CPUs, | |
5535 | * because with more CPUs the 'effective latency' as visible | |
5536 | * to users decreases. But the relationship is not linear, | |
5537 | * so pick a second-best guess by going with the log2 of the | |
5538 | * number of CPUs. | |
5539 | * | |
5540 | * This idea comes from the SD scheduler of Con Kolivas: | |
5541 | */ | |
5542 | static int get_update_sysctl_factor(void) | |
5543 | { | |
5544 | unsigned int cpus = min_t(int, num_online_cpus(), 8); | |
5545 | unsigned int factor; | |
5546 | ||
5547 | switch (sysctl_sched_tunable_scaling) { | |
5548 | case SCHED_TUNABLESCALING_NONE: | |
5549 | factor = 1; | |
5550 | break; | |
5551 | case SCHED_TUNABLESCALING_LINEAR: | |
5552 | factor = cpus; | |
5553 | break; | |
5554 | case SCHED_TUNABLESCALING_LOG: | |
5555 | default: | |
5556 | factor = 1 + ilog2(cpus); | |
5557 | break; | |
5558 | } | |
5559 | ||
5560 | return factor; | |
5561 | } | |
5562 | ||
5563 | static void update_sysctl(void) | |
5564 | { | |
5565 | unsigned int factor = get_update_sysctl_factor(); | |
5566 | ||
5567 | #define SET_SYSCTL(name) \ | |
5568 | (sysctl_##name = (factor) * normalized_sysctl_##name) | |
5569 | SET_SYSCTL(sched_min_granularity); | |
5570 | SET_SYSCTL(sched_latency); | |
5571 | SET_SYSCTL(sched_wakeup_granularity); | |
5572 | SET_SYSCTL(sched_shares_ratelimit); | |
5573 | #undef SET_SYSCTL | |
5574 | } | |
5575 | ||
5576 | static inline void sched_init_granularity(void) | |
5577 | { | |
5578 | update_sysctl(); | |
5579 | } | |
5580 | ||
5581 | #ifdef CONFIG_SMP | |
5582 | /* | |
5583 | * This is how migration works: | |
5584 | * | |
5585 | * 1) we invoke migration_cpu_stop() on the target CPU using | |
5586 | * stop_one_cpu(). | |
5587 | * 2) stopper starts to run (implicitly forcing the migrated thread | |
5588 | * off the CPU) | |
5589 | * 3) it checks whether the migrated task is still in the wrong runqueue. | |
5590 | * 4) if it's in the wrong runqueue then the migration thread removes | |
5591 | * it and puts it into the right queue. | |
5592 | * 5) stopper completes and stop_one_cpu() returns and the migration | |
5593 | * is done. | |
5594 | */ | |
5595 | ||
5596 | /* | |
5597 | * Change a given task's CPU affinity. Migrate the thread to a | |
5598 | * proper CPU and schedule it away if the CPU it's executing on | |
5599 | * is removed from the allowed bitmask. | |
5600 | * | |
5601 | * NOTE: the caller must have a valid reference to the task, the | |
5602 | * task must not exit() & deallocate itself prematurely. The | |
5603 | * call is not atomic; no spinlocks may be held. | |
5604 | */ | |
5605 | int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) | |
5606 | { | |
5607 | unsigned long flags; | |
5608 | struct rq *rq; | |
5609 | unsigned int dest_cpu; | |
5610 | int ret = 0; | |
5611 | ||
5612 | /* | |
5613 | * Serialize against TASK_WAKING so that ttwu() and wunt() can | |
5614 | * drop the rq->lock and still rely on ->cpus_allowed. | |
5615 | */ | |
5616 | again: | |
5617 | while (task_is_waking(p)) | |
5618 | cpu_relax(); | |
5619 | rq = task_rq_lock(p, &flags); | |
5620 | if (task_is_waking(p)) { | |
5621 | task_rq_unlock(rq, &flags); | |
5622 | goto again; | |
5623 | } | |
5624 | ||
5625 | if (!cpumask_intersects(new_mask, cpu_active_mask)) { | |
5626 | ret = -EINVAL; | |
5627 | goto out; | |
5628 | } | |
5629 | ||
5630 | if (unlikely((p->flags & PF_THREAD_BOUND) && p != current && | |
5631 | !cpumask_equal(&p->cpus_allowed, new_mask))) { | |
5632 | ret = -EINVAL; | |
5633 | goto out; | |
5634 | } | |
5635 | ||
5636 | if (p->sched_class->set_cpus_allowed) | |
5637 | p->sched_class->set_cpus_allowed(p, new_mask); | |
5638 | else { | |
5639 | cpumask_copy(&p->cpus_allowed, new_mask); | |
5640 | p->rt.nr_cpus_allowed = cpumask_weight(new_mask); | |
5641 | } | |
5642 | ||
5643 | /* Can the task run on the task's current CPU? If so, we're done */ | |
5644 | if (cpumask_test_cpu(task_cpu(p), new_mask)) | |
5645 | goto out; | |
5646 | ||
5647 | dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); | |
5648 | if (migrate_task(p, dest_cpu)) { | |
5649 | struct migration_arg arg = { p, dest_cpu }; | |
5650 | /* Need help from migration thread: drop lock and wait. */ | |
5651 | task_rq_unlock(rq, &flags); | |
5652 | stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); | |
5653 | tlb_migrate_finish(p->mm); | |
5654 | return 0; | |
5655 | } | |
5656 | out: | |
5657 | task_rq_unlock(rq, &flags); | |
5658 | ||
5659 | return ret; | |
5660 | } | |
5661 | EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); | |
5662 | ||
5663 | /* | |
5664 | * Move (not current) task off this cpu, onto dest cpu. We're doing | |
5665 | * this because either it can't run here any more (set_cpus_allowed() | |
5666 | * away from this CPU, or CPU going down), or because we're | |
5667 | * attempting to rebalance this task on exec (sched_exec). | |
5668 | * | |
5669 | * So we race with normal scheduler movements, but that's OK, as long | |
5670 | * as the task is no longer on this CPU. | |
5671 | * | |
5672 | * Returns non-zero if task was successfully migrated. | |
5673 | */ | |
5674 | static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) | |
5675 | { | |
5676 | struct rq *rq_dest, *rq_src; | |
5677 | int ret = 0; | |
5678 | ||
5679 | if (unlikely(!cpu_active(dest_cpu))) | |
5680 | return ret; | |
5681 | ||
5682 | rq_src = cpu_rq(src_cpu); | |
5683 | rq_dest = cpu_rq(dest_cpu); | |
5684 | ||
5685 | double_rq_lock(rq_src, rq_dest); | |
5686 | /* Already moved. */ | |
5687 | if (task_cpu(p) != src_cpu) | |
5688 | goto done; | |
5689 | /* Affinity changed (again). */ | |
5690 | if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)) | |
5691 | goto fail; | |
5692 | ||
5693 | /* | |
5694 | * If we're not on a rq, the next wake-up will ensure we're | |
5695 | * placed properly. | |
5696 | */ | |
5697 | if (p->se.on_rq) { | |
5698 | deactivate_task(rq_src, p, 0); | |
5699 | set_task_cpu(p, dest_cpu); | |
5700 | activate_task(rq_dest, p, 0); | |
5701 | check_preempt_curr(rq_dest, p, 0); | |
5702 | } | |
5703 | done: | |
5704 | ret = 1; | |
5705 | fail: | |
5706 | double_rq_unlock(rq_src, rq_dest); | |
5707 | return ret; | |
5708 | } | |
5709 | ||
5710 | /* | |
5711 | * migration_cpu_stop - this will be executed by a highprio stopper thread | |
5712 | * and performs thread migration by bumping thread off CPU then | |
5713 | * 'pushing' onto another runqueue. | |
5714 | */ | |
5715 | static int migration_cpu_stop(void *data) | |
5716 | { | |
5717 | struct migration_arg *arg = data; | |
5718 | ||
5719 | /* | |
5720 | * The original target cpu might have gone down and we might | |
5721 | * be on another cpu but it doesn't matter. | |
5722 | */ | |
5723 | local_irq_disable(); | |
5724 | __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); | |
5725 | local_irq_enable(); | |
5726 | return 0; | |
5727 | } | |
5728 | ||
5729 | #ifdef CONFIG_HOTPLUG_CPU | |
5730 | /* | |
5731 | * Figure out where task on dead CPU should go, use force if necessary. | |
5732 | */ | |
5733 | void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p) | |
5734 | { | |
5735 | struct rq *rq = cpu_rq(dead_cpu); | |
5736 | int needs_cpu, uninitialized_var(dest_cpu); | |
5737 | unsigned long flags; | |
5738 | ||
5739 | local_irq_save(flags); | |
5740 | ||
5741 | raw_spin_lock(&rq->lock); | |
5742 | needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING); | |
5743 | if (needs_cpu) | |
5744 | dest_cpu = select_fallback_rq(dead_cpu, p); | |
5745 | raw_spin_unlock(&rq->lock); | |
5746 | /* | |
5747 | * It can only fail if we race with set_cpus_allowed(), | |
5748 | * in the racer should migrate the task anyway. | |
5749 | */ | |
5750 | if (needs_cpu) | |
5751 | __migrate_task(p, dead_cpu, dest_cpu); | |
5752 | local_irq_restore(flags); | |
5753 | } | |
5754 | ||
5755 | /* | |
5756 | * While a dead CPU has no uninterruptible tasks queued at this point, | |
5757 | * it might still have a nonzero ->nr_uninterruptible counter, because | |
5758 | * for performance reasons the counter is not stricly tracking tasks to | |
5759 | * their home CPUs. So we just add the counter to another CPU's counter, | |
5760 | * to keep the global sum constant after CPU-down: | |
5761 | */ | |
5762 | static void migrate_nr_uninterruptible(struct rq *rq_src) | |
5763 | { | |
5764 | struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask)); | |
5765 | unsigned long flags; | |
5766 | ||
5767 | local_irq_save(flags); | |
5768 | double_rq_lock(rq_src, rq_dest); | |
5769 | rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible; | |
5770 | rq_src->nr_uninterruptible = 0; | |
5771 | double_rq_unlock(rq_src, rq_dest); | |
5772 | local_irq_restore(flags); | |
5773 | } | |
5774 | ||
5775 | /* Run through task list and migrate tasks from the dead cpu. */ | |
5776 | static void migrate_live_tasks(int src_cpu) | |
5777 | { | |
5778 | struct task_struct *p, *t; | |
5779 | ||
5780 | read_lock(&tasklist_lock); | |
5781 | ||
5782 | do_each_thread(t, p) { | |
5783 | if (p == current) | |
5784 | continue; | |
5785 | ||
5786 | if (task_cpu(p) == src_cpu) | |
5787 | move_task_off_dead_cpu(src_cpu, p); | |
5788 | } while_each_thread(t, p); | |
5789 | ||
5790 | read_unlock(&tasklist_lock); | |
5791 | } | |
5792 | ||
5793 | /* | |
5794 | * Schedules idle task to be the next runnable task on current CPU. | |
5795 | * It does so by boosting its priority to highest possible. | |
5796 | * Used by CPU offline code. | |
5797 | */ | |
5798 | void sched_idle_next(void) | |
5799 | { | |
5800 | int this_cpu = smp_processor_id(); | |
5801 | struct rq *rq = cpu_rq(this_cpu); | |
5802 | struct task_struct *p = rq->idle; | |
5803 | unsigned long flags; | |
5804 | ||
5805 | /* cpu has to be offline */ | |
5806 | BUG_ON(cpu_online(this_cpu)); | |
5807 | ||
5808 | /* | |
5809 | * Strictly not necessary since rest of the CPUs are stopped by now | |
5810 | * and interrupts disabled on the current cpu. | |
5811 | */ | |
5812 | raw_spin_lock_irqsave(&rq->lock, flags); | |
5813 | ||
5814 | __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1); | |
5815 | ||
5816 | activate_task(rq, p, 0); | |
5817 | ||
5818 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
5819 | } | |
5820 | ||
5821 | /* | |
5822 | * Ensures that the idle task is using init_mm right before its cpu goes | |
5823 | * offline. | |
5824 | */ | |
5825 | void idle_task_exit(void) | |
5826 | { | |
5827 | struct mm_struct *mm = current->active_mm; | |
5828 | ||
5829 | BUG_ON(cpu_online(smp_processor_id())); | |
5830 | ||
5831 | if (mm != &init_mm) | |
5832 | switch_mm(mm, &init_mm, current); | |
5833 | mmdrop(mm); | |
5834 | } | |
5835 | ||
5836 | /* called under rq->lock with disabled interrupts */ | |
5837 | static void migrate_dead(unsigned int dead_cpu, struct task_struct *p) | |
5838 | { | |
5839 | struct rq *rq = cpu_rq(dead_cpu); | |
5840 | ||
5841 | /* Must be exiting, otherwise would be on tasklist. */ | |
5842 | BUG_ON(!p->exit_state); | |
5843 | ||
5844 | /* Cannot have done final schedule yet: would have vanished. */ | |
5845 | BUG_ON(p->state == TASK_DEAD); | |
5846 | ||
5847 | get_task_struct(p); | |
5848 | ||
5849 | /* | |
5850 | * Drop lock around migration; if someone else moves it, | |
5851 | * that's OK. No task can be added to this CPU, so iteration is | |
5852 | * fine. | |
5853 | */ | |
5854 | raw_spin_unlock_irq(&rq->lock); | |
5855 | move_task_off_dead_cpu(dead_cpu, p); | |
5856 | raw_spin_lock_irq(&rq->lock); | |
5857 | ||
5858 | put_task_struct(p); | |
5859 | } | |
5860 | ||
5861 | /* release_task() removes task from tasklist, so we won't find dead tasks. */ | |
5862 | static void migrate_dead_tasks(unsigned int dead_cpu) | |
5863 | { | |
5864 | struct rq *rq = cpu_rq(dead_cpu); | |
5865 | struct task_struct *next; | |
5866 | ||
5867 | for ( ; ; ) { | |
5868 | if (!rq->nr_running) | |
5869 | break; | |
5870 | next = pick_next_task(rq); | |
5871 | if (!next) | |
5872 | break; | |
5873 | next->sched_class->put_prev_task(rq, next); | |
5874 | migrate_dead(dead_cpu, next); | |
5875 | ||
5876 | } | |
5877 | } | |
5878 | ||
5879 | /* | |
5880 | * remove the tasks which were accounted by rq from calc_load_tasks. | |
5881 | */ | |
5882 | static void calc_global_load_remove(struct rq *rq) | |
5883 | { | |
5884 | atomic_long_sub(rq->calc_load_active, &calc_load_tasks); | |
5885 | rq->calc_load_active = 0; | |
5886 | } | |
5887 | #endif /* CONFIG_HOTPLUG_CPU */ | |
5888 | ||
5889 | #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) | |
5890 | ||
5891 | static struct ctl_table sd_ctl_dir[] = { | |
5892 | { | |
5893 | .procname = "sched_domain", | |
5894 | .mode = 0555, | |
5895 | }, | |
5896 | {} | |
5897 | }; | |
5898 | ||
5899 | static struct ctl_table sd_ctl_root[] = { | |
5900 | { | |
5901 | .procname = "kernel", | |
5902 | .mode = 0555, | |
5903 | .child = sd_ctl_dir, | |
5904 | }, | |
5905 | {} | |
5906 | }; | |
5907 | ||
5908 | static struct ctl_table *sd_alloc_ctl_entry(int n) | |
5909 | { | |
5910 | struct ctl_table *entry = | |
5911 | kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); | |
5912 | ||
5913 | return entry; | |
5914 | } | |
5915 | ||
5916 | static void sd_free_ctl_entry(struct ctl_table **tablep) | |
5917 | { | |
5918 | struct ctl_table *entry; | |
5919 | ||
5920 | /* | |
5921 | * In the intermediate directories, both the child directory and | |
5922 | * procname are dynamically allocated and could fail but the mode | |
5923 | * will always be set. In the lowest directory the names are | |
5924 | * static strings and all have proc handlers. | |
5925 | */ | |
5926 | for (entry = *tablep; entry->mode; entry++) { | |
5927 | if (entry->child) | |
5928 | sd_free_ctl_entry(&entry->child); | |
5929 | if (entry->proc_handler == NULL) | |
5930 | kfree(entry->procname); | |
5931 | } | |
5932 | ||
5933 | kfree(*tablep); | |
5934 | *tablep = NULL; | |
5935 | } | |
5936 | ||
5937 | static void | |
5938 | set_table_entry(struct ctl_table *entry, | |
5939 | const char *procname, void *data, int maxlen, | |
5940 | mode_t mode, proc_handler *proc_handler) | |
5941 | { | |
5942 | entry->procname = procname; | |
5943 | entry->data = data; | |
5944 | entry->maxlen = maxlen; | |
5945 | entry->mode = mode; | |
5946 | entry->proc_handler = proc_handler; | |
5947 | } | |
5948 | ||
5949 | static struct ctl_table * | |
5950 | sd_alloc_ctl_domain_table(struct sched_domain *sd) | |
5951 | { | |
5952 | struct ctl_table *table = sd_alloc_ctl_entry(13); | |
5953 | ||
5954 | if (table == NULL) | |
5955 | return NULL; | |
5956 | ||
5957 | set_table_entry(&table[0], "min_interval", &sd->min_interval, | |
5958 | sizeof(long), 0644, proc_doulongvec_minmax); | |
5959 | set_table_entry(&table[1], "max_interval", &sd->max_interval, | |
5960 | sizeof(long), 0644, proc_doulongvec_minmax); | |
5961 | set_table_entry(&table[2], "busy_idx", &sd->busy_idx, | |
5962 | sizeof(int), 0644, proc_dointvec_minmax); | |
5963 | set_table_entry(&table[3], "idle_idx", &sd->idle_idx, | |
5964 | sizeof(int), 0644, proc_dointvec_minmax); | |
5965 | set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, | |
5966 | sizeof(int), 0644, proc_dointvec_minmax); | |
5967 | set_table_entry(&table[5], "wake_idx", &sd->wake_idx, | |
5968 | sizeof(int), 0644, proc_dointvec_minmax); | |
5969 | set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, | |
5970 | sizeof(int), 0644, proc_dointvec_minmax); | |
5971 | set_table_entry(&table[7], "busy_factor", &sd->busy_factor, | |
5972 | sizeof(int), 0644, proc_dointvec_minmax); | |
5973 | set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, | |
5974 | sizeof(int), 0644, proc_dointvec_minmax); | |
5975 | set_table_entry(&table[9], "cache_nice_tries", | |
5976 | &sd->cache_nice_tries, | |
5977 | sizeof(int), 0644, proc_dointvec_minmax); | |
5978 | set_table_entry(&table[10], "flags", &sd->flags, | |
5979 | sizeof(int), 0644, proc_dointvec_minmax); | |
5980 | set_table_entry(&table[11], "name", sd->name, | |
5981 | CORENAME_MAX_SIZE, 0444, proc_dostring); | |
5982 | /* &table[12] is terminator */ | |
5983 | ||
5984 | return table; | |
5985 | } | |
5986 | ||
5987 | static ctl_table *sd_alloc_ctl_cpu_table(int cpu) | |
5988 | { | |
5989 | struct ctl_table *entry, *table; | |
5990 | struct sched_domain *sd; | |
5991 | int domain_num = 0, i; | |
5992 | char buf[32]; | |
5993 | ||
5994 | for_each_domain(cpu, sd) | |
5995 | domain_num++; | |
5996 | entry = table = sd_alloc_ctl_entry(domain_num + 1); | |
5997 | if (table == NULL) | |
5998 | return NULL; | |
5999 | ||
6000 | i = 0; | |
6001 | for_each_domain(cpu, sd) { | |
6002 | snprintf(buf, 32, "domain%d", i); | |
6003 | entry->procname = kstrdup(buf, GFP_KERNEL); | |
6004 | entry->mode = 0555; | |
6005 | entry->child = sd_alloc_ctl_domain_table(sd); | |
6006 | entry++; | |
6007 | i++; | |
6008 | } | |
6009 | return table; | |
6010 | } | |
6011 | ||
6012 | static struct ctl_table_header *sd_sysctl_header; | |
6013 | static void register_sched_domain_sysctl(void) | |
6014 | { | |
6015 | int i, cpu_num = num_possible_cpus(); | |
6016 | struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); | |
6017 | char buf[32]; | |
6018 | ||
6019 | WARN_ON(sd_ctl_dir[0].child); | |
6020 | sd_ctl_dir[0].child = entry; | |
6021 | ||
6022 | if (entry == NULL) | |
6023 | return; | |
6024 | ||
6025 | for_each_possible_cpu(i) { | |
6026 | snprintf(buf, 32, "cpu%d", i); | |
6027 | entry->procname = kstrdup(buf, GFP_KERNEL); | |
6028 | entry->mode = 0555; | |
6029 | entry->child = sd_alloc_ctl_cpu_table(i); | |
6030 | entry++; | |
6031 | } | |
6032 | ||
6033 | WARN_ON(sd_sysctl_header); | |
6034 | sd_sysctl_header = register_sysctl_table(sd_ctl_root); | |
6035 | } | |
6036 | ||
6037 | /* may be called multiple times per register */ | |
6038 | static void unregister_sched_domain_sysctl(void) | |
6039 | { | |
6040 | if (sd_sysctl_header) | |
6041 | unregister_sysctl_table(sd_sysctl_header); | |
6042 | sd_sysctl_header = NULL; | |
6043 | if (sd_ctl_dir[0].child) | |
6044 | sd_free_ctl_entry(&sd_ctl_dir[0].child); | |
6045 | } | |
6046 | #else | |
6047 | static void register_sched_domain_sysctl(void) | |
6048 | { | |
6049 | } | |
6050 | static void unregister_sched_domain_sysctl(void) | |
6051 | { | |
6052 | } | |
6053 | #endif | |
6054 | ||
6055 | static void set_rq_online(struct rq *rq) | |
6056 | { | |
6057 | if (!rq->online) { | |
6058 | const struct sched_class *class; | |
6059 | ||
6060 | cpumask_set_cpu(rq->cpu, rq->rd->online); | |
6061 | rq->online = 1; | |
6062 | ||
6063 | for_each_class(class) { | |
6064 | if (class->rq_online) | |
6065 | class->rq_online(rq); | |
6066 | } | |
6067 | } | |
6068 | } | |
6069 | ||
6070 | static void set_rq_offline(struct rq *rq) | |
6071 | { | |
6072 | if (rq->online) { | |
6073 | const struct sched_class *class; | |
6074 | ||
6075 | for_each_class(class) { | |
6076 | if (class->rq_offline) | |
6077 | class->rq_offline(rq); | |
6078 | } | |
6079 | ||
6080 | cpumask_clear_cpu(rq->cpu, rq->rd->online); | |
6081 | rq->online = 0; | |
6082 | } | |
6083 | } | |
6084 | ||
6085 | /* | |
6086 | * migration_call - callback that gets triggered when a CPU is added. | |
6087 | * Here we can start up the necessary migration thread for the new CPU. | |
6088 | */ | |
6089 | static int __cpuinit | |
6090 | migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) | |
6091 | { | |
6092 | int cpu = (long)hcpu; | |
6093 | unsigned long flags; | |
6094 | struct rq *rq = cpu_rq(cpu); | |
6095 | ||
6096 | switch (action) { | |
6097 | ||
6098 | case CPU_UP_PREPARE: | |
6099 | case CPU_UP_PREPARE_FROZEN: | |
6100 | rq->calc_load_update = calc_load_update; | |
6101 | break; | |
6102 | ||
6103 | case CPU_ONLINE: | |
6104 | case CPU_ONLINE_FROZEN: | |
6105 | /* Update our root-domain */ | |
6106 | raw_spin_lock_irqsave(&rq->lock, flags); | |
6107 | if (rq->rd) { | |
6108 | BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); | |
6109 | ||
6110 | set_rq_online(rq); | |
6111 | } | |
6112 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
6113 | break; | |
6114 | ||
6115 | #ifdef CONFIG_HOTPLUG_CPU | |
6116 | case CPU_DEAD: | |
6117 | case CPU_DEAD_FROZEN: | |
6118 | migrate_live_tasks(cpu); | |
6119 | /* Idle task back to normal (off runqueue, low prio) */ | |
6120 | raw_spin_lock_irq(&rq->lock); | |
6121 | deactivate_task(rq, rq->idle, 0); | |
6122 | __setscheduler(rq, rq->idle, SCHED_NORMAL, 0); | |
6123 | rq->idle->sched_class = &idle_sched_class; | |
6124 | migrate_dead_tasks(cpu); | |
6125 | raw_spin_unlock_irq(&rq->lock); | |
6126 | migrate_nr_uninterruptible(rq); | |
6127 | BUG_ON(rq->nr_running != 0); | |
6128 | calc_global_load_remove(rq); | |
6129 | break; | |
6130 | ||
6131 | case CPU_DYING: | |
6132 | case CPU_DYING_FROZEN: | |
6133 | /* Update our root-domain */ | |
6134 | raw_spin_lock_irqsave(&rq->lock, flags); | |
6135 | if (rq->rd) { | |
6136 | BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); | |
6137 | set_rq_offline(rq); | |
6138 | } | |
6139 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
6140 | break; | |
6141 | #endif | |
6142 | } | |
6143 | return NOTIFY_OK; | |
6144 | } | |
6145 | ||
6146 | /* | |
6147 | * Register at high priority so that task migration (migrate_all_tasks) | |
6148 | * happens before everything else. This has to be lower priority than | |
6149 | * the notifier in the perf_event subsystem, though. | |
6150 | */ | |
6151 | static struct notifier_block __cpuinitdata migration_notifier = { | |
6152 | .notifier_call = migration_call, | |
6153 | .priority = CPU_PRI_MIGRATION, | |
6154 | }; | |
6155 | ||
6156 | static int __cpuinit sched_cpu_active(struct notifier_block *nfb, | |
6157 | unsigned long action, void *hcpu) | |
6158 | { | |
6159 | switch (action & ~CPU_TASKS_FROZEN) { | |
6160 | case CPU_ONLINE: | |
6161 | case CPU_DOWN_FAILED: | |
6162 | set_cpu_active((long)hcpu, true); | |
6163 | return NOTIFY_OK; | |
6164 | default: | |
6165 | return NOTIFY_DONE; | |
6166 | } | |
6167 | } | |
6168 | ||
6169 | static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb, | |
6170 | unsigned long action, void *hcpu) | |
6171 | { | |
6172 | switch (action & ~CPU_TASKS_FROZEN) { | |
6173 | case CPU_DOWN_PREPARE: | |
6174 | set_cpu_active((long)hcpu, false); | |
6175 | return NOTIFY_OK; | |
6176 | default: | |
6177 | return NOTIFY_DONE; | |
6178 | } | |
6179 | } | |
6180 | ||
6181 | static int __init migration_init(void) | |
6182 | { | |
6183 | void *cpu = (void *)(long)smp_processor_id(); | |
6184 | int err; | |
6185 | ||
6186 | /* Initialize migration for the boot CPU */ | |
6187 | err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); | |
6188 | BUG_ON(err == NOTIFY_BAD); | |
6189 | migration_call(&migration_notifier, CPU_ONLINE, cpu); | |
6190 | register_cpu_notifier(&migration_notifier); | |
6191 | ||
6192 | /* Register cpu active notifiers */ | |
6193 | cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); | |
6194 | cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); | |
6195 | ||
6196 | return 0; | |
6197 | } | |
6198 | early_initcall(migration_init); | |
6199 | #endif | |
6200 | ||
6201 | #ifdef CONFIG_SMP | |
6202 | ||
6203 | #ifdef CONFIG_SCHED_DEBUG | |
6204 | ||
6205 | static __read_mostly int sched_domain_debug_enabled; | |
6206 | ||
6207 | static int __init sched_domain_debug_setup(char *str) | |
6208 | { | |
6209 | sched_domain_debug_enabled = 1; | |
6210 | ||
6211 | return 0; | |
6212 | } | |
6213 | early_param("sched_debug", sched_domain_debug_setup); | |
6214 | ||
6215 | static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, | |
6216 | struct cpumask *groupmask) | |
6217 | { | |
6218 | struct sched_group *group = sd->groups; | |
6219 | char str[256]; | |
6220 | ||
6221 | cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); | |
6222 | cpumask_clear(groupmask); | |
6223 | ||
6224 | printk(KERN_DEBUG "%*s domain %d: ", level, "", level); | |
6225 | ||
6226 | if (!(sd->flags & SD_LOAD_BALANCE)) { | |
6227 | printk("does not load-balance\n"); | |
6228 | if (sd->parent) | |
6229 | printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" | |
6230 | " has parent"); | |
6231 | return -1; | |
6232 | } | |
6233 | ||
6234 | printk(KERN_CONT "span %s level %s\n", str, sd->name); | |
6235 | ||
6236 | if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { | |
6237 | printk(KERN_ERR "ERROR: domain->span does not contain " | |
6238 | "CPU%d\n", cpu); | |
6239 | } | |
6240 | if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { | |
6241 | printk(KERN_ERR "ERROR: domain->groups does not contain" | |
6242 | " CPU%d\n", cpu); | |
6243 | } | |
6244 | ||
6245 | printk(KERN_DEBUG "%*s groups:", level + 1, ""); | |
6246 | do { | |
6247 | if (!group) { | |
6248 | printk("\n"); | |
6249 | printk(KERN_ERR "ERROR: group is NULL\n"); | |
6250 | break; | |
6251 | } | |
6252 | ||
6253 | if (!group->cpu_power) { | |
6254 | printk(KERN_CONT "\n"); | |
6255 | printk(KERN_ERR "ERROR: domain->cpu_power not " | |
6256 | "set\n"); | |
6257 | break; | |
6258 | } | |
6259 | ||
6260 | if (!cpumask_weight(sched_group_cpus(group))) { | |
6261 | printk(KERN_CONT "\n"); | |
6262 | printk(KERN_ERR "ERROR: empty group\n"); | |
6263 | break; | |
6264 | } | |
6265 | ||
6266 | if (cpumask_intersects(groupmask, sched_group_cpus(group))) { | |
6267 | printk(KERN_CONT "\n"); | |
6268 | printk(KERN_ERR "ERROR: repeated CPUs\n"); | |
6269 | break; | |
6270 | } | |
6271 | ||
6272 | cpumask_or(groupmask, groupmask, sched_group_cpus(group)); | |
6273 | ||
6274 | cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); | |
6275 | ||
6276 | printk(KERN_CONT " %s", str); | |
6277 | if (group->cpu_power != SCHED_LOAD_SCALE) { | |
6278 | printk(KERN_CONT " (cpu_power = %d)", | |
6279 | group->cpu_power); | |
6280 | } | |
6281 | ||
6282 | group = group->next; | |
6283 | } while (group != sd->groups); | |
6284 | printk(KERN_CONT "\n"); | |
6285 | ||
6286 | if (!cpumask_equal(sched_domain_span(sd), groupmask)) | |
6287 | printk(KERN_ERR "ERROR: groups don't span domain->span\n"); | |
6288 | ||
6289 | if (sd->parent && | |
6290 | !cpumask_subset(groupmask, sched_domain_span(sd->parent))) | |
6291 | printk(KERN_ERR "ERROR: parent span is not a superset " | |
6292 | "of domain->span\n"); | |
6293 | return 0; | |
6294 | } | |
6295 | ||
6296 | static void sched_domain_debug(struct sched_domain *sd, int cpu) | |
6297 | { | |
6298 | cpumask_var_t groupmask; | |
6299 | int level = 0; | |
6300 | ||
6301 | if (!sched_domain_debug_enabled) | |
6302 | return; | |
6303 | ||
6304 | if (!sd) { | |
6305 | printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); | |
6306 | return; | |
6307 | } | |
6308 | ||
6309 | printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); | |
6310 | ||
6311 | if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) { | |
6312 | printk(KERN_DEBUG "Cannot load-balance (out of memory)\n"); | |
6313 | return; | |
6314 | } | |
6315 | ||
6316 | for (;;) { | |
6317 | if (sched_domain_debug_one(sd, cpu, level, groupmask)) | |
6318 | break; | |
6319 | level++; | |
6320 | sd = sd->parent; | |
6321 | if (!sd) | |
6322 | break; | |
6323 | } | |
6324 | free_cpumask_var(groupmask); | |
6325 | } | |
6326 | #else /* !CONFIG_SCHED_DEBUG */ | |
6327 | # define sched_domain_debug(sd, cpu) do { } while (0) | |
6328 | #endif /* CONFIG_SCHED_DEBUG */ | |
6329 | ||
6330 | static int sd_degenerate(struct sched_domain *sd) | |
6331 | { | |
6332 | if (cpumask_weight(sched_domain_span(sd)) == 1) | |
6333 | return 1; | |
6334 | ||
6335 | /* Following flags need at least 2 groups */ | |
6336 | if (sd->flags & (SD_LOAD_BALANCE | | |
6337 | SD_BALANCE_NEWIDLE | | |
6338 | SD_BALANCE_FORK | | |
6339 | SD_BALANCE_EXEC | | |
6340 | SD_SHARE_CPUPOWER | | |
6341 | SD_SHARE_PKG_RESOURCES)) { | |
6342 | if (sd->groups != sd->groups->next) | |
6343 | return 0; | |
6344 | } | |
6345 | ||
6346 | /* Following flags don't use groups */ | |
6347 | if (sd->flags & (SD_WAKE_AFFINE)) | |
6348 | return 0; | |
6349 | ||
6350 | return 1; | |
6351 | } | |
6352 | ||
6353 | static int | |
6354 | sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) | |
6355 | { | |
6356 | unsigned long cflags = sd->flags, pflags = parent->flags; | |
6357 | ||
6358 | if (sd_degenerate(parent)) | |
6359 | return 1; | |
6360 | ||
6361 | if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) | |
6362 | return 0; | |
6363 | ||
6364 | /* Flags needing groups don't count if only 1 group in parent */ | |
6365 | if (parent->groups == parent->groups->next) { | |
6366 | pflags &= ~(SD_LOAD_BALANCE | | |
6367 | SD_BALANCE_NEWIDLE | | |
6368 | SD_BALANCE_FORK | | |
6369 | SD_BALANCE_EXEC | | |
6370 | SD_SHARE_CPUPOWER | | |
6371 | SD_SHARE_PKG_RESOURCES); | |
6372 | if (nr_node_ids == 1) | |
6373 | pflags &= ~SD_SERIALIZE; | |
6374 | } | |
6375 | if (~cflags & pflags) | |
6376 | return 0; | |
6377 | ||
6378 | return 1; | |
6379 | } | |
6380 | ||
6381 | static void free_rootdomain(struct root_domain *rd) | |
6382 | { | |
6383 | synchronize_sched(); | |
6384 | ||
6385 | cpupri_cleanup(&rd->cpupri); | |
6386 | ||
6387 | free_cpumask_var(rd->rto_mask); | |
6388 | free_cpumask_var(rd->online); | |
6389 | free_cpumask_var(rd->span); | |
6390 | kfree(rd); | |
6391 | } | |
6392 | ||
6393 | static void rq_attach_root(struct rq *rq, struct root_domain *rd) | |
6394 | { | |
6395 | struct root_domain *old_rd = NULL; | |
6396 | unsigned long flags; | |
6397 | ||
6398 | raw_spin_lock_irqsave(&rq->lock, flags); | |
6399 | ||
6400 | if (rq->rd) { | |
6401 | old_rd = rq->rd; | |
6402 | ||
6403 | if (cpumask_test_cpu(rq->cpu, old_rd->online)) | |
6404 | set_rq_offline(rq); | |
6405 | ||
6406 | cpumask_clear_cpu(rq->cpu, old_rd->span); | |
6407 | ||
6408 | /* | |
6409 | * If we dont want to free the old_rt yet then | |
6410 | * set old_rd to NULL to skip the freeing later | |
6411 | * in this function: | |
6412 | */ | |
6413 | if (!atomic_dec_and_test(&old_rd->refcount)) | |
6414 | old_rd = NULL; | |
6415 | } | |
6416 | ||
6417 | atomic_inc(&rd->refcount); | |
6418 | rq->rd = rd; | |
6419 | ||
6420 | cpumask_set_cpu(rq->cpu, rd->span); | |
6421 | if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) | |
6422 | set_rq_online(rq); | |
6423 | ||
6424 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
6425 | ||
6426 | if (old_rd) | |
6427 | free_rootdomain(old_rd); | |
6428 | } | |
6429 | ||
6430 | static int init_rootdomain(struct root_domain *rd) | |
6431 | { | |
6432 | memset(rd, 0, sizeof(*rd)); | |
6433 | ||
6434 | if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) | |
6435 | goto out; | |
6436 | if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) | |
6437 | goto free_span; | |
6438 | if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) | |
6439 | goto free_online; | |
6440 | ||
6441 | if (cpupri_init(&rd->cpupri) != 0) | |
6442 | goto free_rto_mask; | |
6443 | return 0; | |
6444 | ||
6445 | free_rto_mask: | |
6446 | free_cpumask_var(rd->rto_mask); | |
6447 | free_online: | |
6448 | free_cpumask_var(rd->online); | |
6449 | free_span: | |
6450 | free_cpumask_var(rd->span); | |
6451 | out: | |
6452 | return -ENOMEM; | |
6453 | } | |
6454 | ||
6455 | static void init_defrootdomain(void) | |
6456 | { | |
6457 | init_rootdomain(&def_root_domain); | |
6458 | ||
6459 | atomic_set(&def_root_domain.refcount, 1); | |
6460 | } | |
6461 | ||
6462 | static struct root_domain *alloc_rootdomain(void) | |
6463 | { | |
6464 | struct root_domain *rd; | |
6465 | ||
6466 | rd = kmalloc(sizeof(*rd), GFP_KERNEL); | |
6467 | if (!rd) | |
6468 | return NULL; | |
6469 | ||
6470 | if (init_rootdomain(rd) != 0) { | |
6471 | kfree(rd); | |
6472 | return NULL; | |
6473 | } | |
6474 | ||
6475 | return rd; | |
6476 | } | |
6477 | ||
6478 | /* | |
6479 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must | |
6480 | * hold the hotplug lock. | |
6481 | */ | |
6482 | static void | |
6483 | cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) | |
6484 | { | |
6485 | struct rq *rq = cpu_rq(cpu); | |
6486 | struct sched_domain *tmp; | |
6487 | ||
6488 | for (tmp = sd; tmp; tmp = tmp->parent) | |
6489 | tmp->span_weight = cpumask_weight(sched_domain_span(tmp)); | |
6490 | ||
6491 | /* Remove the sched domains which do not contribute to scheduling. */ | |
6492 | for (tmp = sd; tmp; ) { | |
6493 | struct sched_domain *parent = tmp->parent; | |
6494 | if (!parent) | |
6495 | break; | |
6496 | ||
6497 | if (sd_parent_degenerate(tmp, parent)) { | |
6498 | tmp->parent = parent->parent; | |
6499 | if (parent->parent) | |
6500 | parent->parent->child = tmp; | |
6501 | } else | |
6502 | tmp = tmp->parent; | |
6503 | } | |
6504 | ||
6505 | if (sd && sd_degenerate(sd)) { | |
6506 | sd = sd->parent; | |
6507 | if (sd) | |
6508 | sd->child = NULL; | |
6509 | } | |
6510 | ||
6511 | sched_domain_debug(sd, cpu); | |
6512 | ||
6513 | rq_attach_root(rq, rd); | |
6514 | rcu_assign_pointer(rq->sd, sd); | |
6515 | } | |
6516 | ||
6517 | /* cpus with isolated domains */ | |
6518 | static cpumask_var_t cpu_isolated_map; | |
6519 | ||
6520 | /* Setup the mask of cpus configured for isolated domains */ | |
6521 | static int __init isolated_cpu_setup(char *str) | |
6522 | { | |
6523 | alloc_bootmem_cpumask_var(&cpu_isolated_map); | |
6524 | cpulist_parse(str, cpu_isolated_map); | |
6525 | return 1; | |
6526 | } | |
6527 | ||
6528 | __setup("isolcpus=", isolated_cpu_setup); | |
6529 | ||
6530 | /* | |
6531 | * init_sched_build_groups takes the cpumask we wish to span, and a pointer | |
6532 | * to a function which identifies what group(along with sched group) a CPU | |
6533 | * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids | |
6534 | * (due to the fact that we keep track of groups covered with a struct cpumask). | |
6535 | * | |
6536 | * init_sched_build_groups will build a circular linked list of the groups | |
6537 | * covered by the given span, and will set each group's ->cpumask correctly, | |
6538 | * and ->cpu_power to 0. | |
6539 | */ | |
6540 | static void | |
6541 | init_sched_build_groups(const struct cpumask *span, | |
6542 | const struct cpumask *cpu_map, | |
6543 | int (*group_fn)(int cpu, const struct cpumask *cpu_map, | |
6544 | struct sched_group **sg, | |
6545 | struct cpumask *tmpmask), | |
6546 | struct cpumask *covered, struct cpumask *tmpmask) | |
6547 | { | |
6548 | struct sched_group *first = NULL, *last = NULL; | |
6549 | int i; | |
6550 | ||
6551 | cpumask_clear(covered); | |
6552 | ||
6553 | for_each_cpu(i, span) { | |
6554 | struct sched_group *sg; | |
6555 | int group = group_fn(i, cpu_map, &sg, tmpmask); | |
6556 | int j; | |
6557 | ||
6558 | if (cpumask_test_cpu(i, covered)) | |
6559 | continue; | |
6560 | ||
6561 | cpumask_clear(sched_group_cpus(sg)); | |
6562 | sg->cpu_power = 0; | |
6563 | ||
6564 | for_each_cpu(j, span) { | |
6565 | if (group_fn(j, cpu_map, NULL, tmpmask) != group) | |
6566 | continue; | |
6567 | ||
6568 | cpumask_set_cpu(j, covered); | |
6569 | cpumask_set_cpu(j, sched_group_cpus(sg)); | |
6570 | } | |
6571 | if (!first) | |
6572 | first = sg; | |
6573 | if (last) | |
6574 | last->next = sg; | |
6575 | last = sg; | |
6576 | } | |
6577 | last->next = first; | |
6578 | } | |
6579 | ||
6580 | #define SD_NODES_PER_DOMAIN 16 | |
6581 | ||
6582 | #ifdef CONFIG_NUMA | |
6583 | ||
6584 | /** | |
6585 | * find_next_best_node - find the next node to include in a sched_domain | |
6586 | * @node: node whose sched_domain we're building | |
6587 | * @used_nodes: nodes already in the sched_domain | |
6588 | * | |
6589 | * Find the next node to include in a given scheduling domain. Simply | |
6590 | * finds the closest node not already in the @used_nodes map. | |
6591 | * | |
6592 | * Should use nodemask_t. | |
6593 | */ | |
6594 | static int find_next_best_node(int node, nodemask_t *used_nodes) | |
6595 | { | |
6596 | int i, n, val, min_val, best_node = 0; | |
6597 | ||
6598 | min_val = INT_MAX; | |
6599 | ||
6600 | for (i = 0; i < nr_node_ids; i++) { | |
6601 | /* Start at @node */ | |
6602 | n = (node + i) % nr_node_ids; | |
6603 | ||
6604 | if (!nr_cpus_node(n)) | |
6605 | continue; | |
6606 | ||
6607 | /* Skip already used nodes */ | |
6608 | if (node_isset(n, *used_nodes)) | |
6609 | continue; | |
6610 | ||
6611 | /* Simple min distance search */ | |
6612 | val = node_distance(node, n); | |
6613 | ||
6614 | if (val < min_val) { | |
6615 | min_val = val; | |
6616 | best_node = n; | |
6617 | } | |
6618 | } | |
6619 | ||
6620 | node_set(best_node, *used_nodes); | |
6621 | return best_node; | |
6622 | } | |
6623 | ||
6624 | /** | |
6625 | * sched_domain_node_span - get a cpumask for a node's sched_domain | |
6626 | * @node: node whose cpumask we're constructing | |
6627 | * @span: resulting cpumask | |
6628 | * | |
6629 | * Given a node, construct a good cpumask for its sched_domain to span. It | |
6630 | * should be one that prevents unnecessary balancing, but also spreads tasks | |
6631 | * out optimally. | |
6632 | */ | |
6633 | static void sched_domain_node_span(int node, struct cpumask *span) | |
6634 | { | |
6635 | nodemask_t used_nodes; | |
6636 | int i; | |
6637 | ||
6638 | cpumask_clear(span); | |
6639 | nodes_clear(used_nodes); | |
6640 | ||
6641 | cpumask_or(span, span, cpumask_of_node(node)); | |
6642 | node_set(node, used_nodes); | |
6643 | ||
6644 | for (i = 1; i < SD_NODES_PER_DOMAIN; i++) { | |
6645 | int next_node = find_next_best_node(node, &used_nodes); | |
6646 | ||
6647 | cpumask_or(span, span, cpumask_of_node(next_node)); | |
6648 | } | |
6649 | } | |
6650 | #endif /* CONFIG_NUMA */ | |
6651 | ||
6652 | int sched_smt_power_savings = 0, sched_mc_power_savings = 0; | |
6653 | ||
6654 | /* | |
6655 | * The cpus mask in sched_group and sched_domain hangs off the end. | |
6656 | * | |
6657 | * ( See the the comments in include/linux/sched.h:struct sched_group | |
6658 | * and struct sched_domain. ) | |
6659 | */ | |
6660 | struct static_sched_group { | |
6661 | struct sched_group sg; | |
6662 | DECLARE_BITMAP(cpus, CONFIG_NR_CPUS); | |
6663 | }; | |
6664 | ||
6665 | struct static_sched_domain { | |
6666 | struct sched_domain sd; | |
6667 | DECLARE_BITMAP(span, CONFIG_NR_CPUS); | |
6668 | }; | |
6669 | ||
6670 | struct s_data { | |
6671 | #ifdef CONFIG_NUMA | |
6672 | int sd_allnodes; | |
6673 | cpumask_var_t domainspan; | |
6674 | cpumask_var_t covered; | |
6675 | cpumask_var_t notcovered; | |
6676 | #endif | |
6677 | cpumask_var_t nodemask; | |
6678 | cpumask_var_t this_sibling_map; | |
6679 | cpumask_var_t this_core_map; | |
6680 | cpumask_var_t this_book_map; | |
6681 | cpumask_var_t send_covered; | |
6682 | cpumask_var_t tmpmask; | |
6683 | struct sched_group **sched_group_nodes; | |
6684 | struct root_domain *rd; | |
6685 | }; | |
6686 | ||
6687 | enum s_alloc { | |
6688 | sa_sched_groups = 0, | |
6689 | sa_rootdomain, | |
6690 | sa_tmpmask, | |
6691 | sa_send_covered, | |
6692 | sa_this_book_map, | |
6693 | sa_this_core_map, | |
6694 | sa_this_sibling_map, | |
6695 | sa_nodemask, | |
6696 | sa_sched_group_nodes, | |
6697 | #ifdef CONFIG_NUMA | |
6698 | sa_notcovered, | |
6699 | sa_covered, | |
6700 | sa_domainspan, | |
6701 | #endif | |
6702 | sa_none, | |
6703 | }; | |
6704 | ||
6705 | /* | |
6706 | * SMT sched-domains: | |
6707 | */ | |
6708 | #ifdef CONFIG_SCHED_SMT | |
6709 | static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains); | |
6710 | static DEFINE_PER_CPU(struct static_sched_group, sched_groups); | |
6711 | ||
6712 | static int | |
6713 | cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map, | |
6714 | struct sched_group **sg, struct cpumask *unused) | |
6715 | { | |
6716 | if (sg) | |
6717 | *sg = &per_cpu(sched_groups, cpu).sg; | |
6718 | return cpu; | |
6719 | } | |
6720 | #endif /* CONFIG_SCHED_SMT */ | |
6721 | ||
6722 | /* | |
6723 | * multi-core sched-domains: | |
6724 | */ | |
6725 | #ifdef CONFIG_SCHED_MC | |
6726 | static DEFINE_PER_CPU(struct static_sched_domain, core_domains); | |
6727 | static DEFINE_PER_CPU(struct static_sched_group, sched_group_core); | |
6728 | ||
6729 | static int | |
6730 | cpu_to_core_group(int cpu, const struct cpumask *cpu_map, | |
6731 | struct sched_group **sg, struct cpumask *mask) | |
6732 | { | |
6733 | int group; | |
6734 | #ifdef CONFIG_SCHED_SMT | |
6735 | cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map); | |
6736 | group = cpumask_first(mask); | |
6737 | #else | |
6738 | group = cpu; | |
6739 | #endif | |
6740 | if (sg) | |
6741 | *sg = &per_cpu(sched_group_core, group).sg; | |
6742 | return group; | |
6743 | } | |
6744 | #endif /* CONFIG_SCHED_MC */ | |
6745 | ||
6746 | /* | |
6747 | * book sched-domains: | |
6748 | */ | |
6749 | #ifdef CONFIG_SCHED_BOOK | |
6750 | static DEFINE_PER_CPU(struct static_sched_domain, book_domains); | |
6751 | static DEFINE_PER_CPU(struct static_sched_group, sched_group_book); | |
6752 | ||
6753 | static int | |
6754 | cpu_to_book_group(int cpu, const struct cpumask *cpu_map, | |
6755 | struct sched_group **sg, struct cpumask *mask) | |
6756 | { | |
6757 | int group = cpu; | |
6758 | #ifdef CONFIG_SCHED_MC | |
6759 | cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map); | |
6760 | group = cpumask_first(mask); | |
6761 | #elif defined(CONFIG_SCHED_SMT) | |
6762 | cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map); | |
6763 | group = cpumask_first(mask); | |
6764 | #endif | |
6765 | if (sg) | |
6766 | *sg = &per_cpu(sched_group_book, group).sg; | |
6767 | return group; | |
6768 | } | |
6769 | #endif /* CONFIG_SCHED_BOOK */ | |
6770 | ||
6771 | static DEFINE_PER_CPU(struct static_sched_domain, phys_domains); | |
6772 | static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys); | |
6773 | ||
6774 | static int | |
6775 | cpu_to_phys_group(int cpu, const struct cpumask *cpu_map, | |
6776 | struct sched_group **sg, struct cpumask *mask) | |
6777 | { | |
6778 | int group; | |
6779 | #ifdef CONFIG_SCHED_BOOK | |
6780 | cpumask_and(mask, cpu_book_mask(cpu), cpu_map); | |
6781 | group = cpumask_first(mask); | |
6782 | #elif defined(CONFIG_SCHED_MC) | |
6783 | cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map); | |
6784 | group = cpumask_first(mask); | |
6785 | #elif defined(CONFIG_SCHED_SMT) | |
6786 | cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map); | |
6787 | group = cpumask_first(mask); | |
6788 | #else | |
6789 | group = cpu; | |
6790 | #endif | |
6791 | if (sg) | |
6792 | *sg = &per_cpu(sched_group_phys, group).sg; | |
6793 | return group; | |
6794 | } | |
6795 | ||
6796 | #ifdef CONFIG_NUMA | |
6797 | /* | |
6798 | * The init_sched_build_groups can't handle what we want to do with node | |
6799 | * groups, so roll our own. Now each node has its own list of groups which | |
6800 | * gets dynamically allocated. | |
6801 | */ | |
6802 | static DEFINE_PER_CPU(struct static_sched_domain, node_domains); | |
6803 | static struct sched_group ***sched_group_nodes_bycpu; | |
6804 | ||
6805 | static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains); | |
6806 | static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes); | |
6807 | ||
6808 | static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map, | |
6809 | struct sched_group **sg, | |
6810 | struct cpumask *nodemask) | |
6811 | { | |
6812 | int group; | |
6813 | ||
6814 | cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map); | |
6815 | group = cpumask_first(nodemask); | |
6816 | ||
6817 | if (sg) | |
6818 | *sg = &per_cpu(sched_group_allnodes, group).sg; | |
6819 | return group; | |
6820 | } | |
6821 | ||
6822 | static void init_numa_sched_groups_power(struct sched_group *group_head) | |
6823 | { | |
6824 | struct sched_group *sg = group_head; | |
6825 | int j; | |
6826 | ||
6827 | if (!sg) | |
6828 | return; | |
6829 | do { | |
6830 | for_each_cpu(j, sched_group_cpus(sg)) { | |
6831 | struct sched_domain *sd; | |
6832 | ||
6833 | sd = &per_cpu(phys_domains, j).sd; | |
6834 | if (j != group_first_cpu(sd->groups)) { | |
6835 | /* | |
6836 | * Only add "power" once for each | |
6837 | * physical package. | |
6838 | */ | |
6839 | continue; | |
6840 | } | |
6841 | ||
6842 | sg->cpu_power += sd->groups->cpu_power; | |
6843 | } | |
6844 | sg = sg->next; | |
6845 | } while (sg != group_head); | |
6846 | } | |
6847 | ||
6848 | static int build_numa_sched_groups(struct s_data *d, | |
6849 | const struct cpumask *cpu_map, int num) | |
6850 | { | |
6851 | struct sched_domain *sd; | |
6852 | struct sched_group *sg, *prev; | |
6853 | int n, j; | |
6854 | ||
6855 | cpumask_clear(d->covered); | |
6856 | cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map); | |
6857 | if (cpumask_empty(d->nodemask)) { | |
6858 | d->sched_group_nodes[num] = NULL; | |
6859 | goto out; | |
6860 | } | |
6861 | ||
6862 | sched_domain_node_span(num, d->domainspan); | |
6863 | cpumask_and(d->domainspan, d->domainspan, cpu_map); | |
6864 | ||
6865 | sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(), | |
6866 | GFP_KERNEL, num); | |
6867 | if (!sg) { | |
6868 | printk(KERN_WARNING "Can not alloc domain group for node %d\n", | |
6869 | num); | |
6870 | return -ENOMEM; | |
6871 | } | |
6872 | d->sched_group_nodes[num] = sg; | |
6873 | ||
6874 | for_each_cpu(j, d->nodemask) { | |
6875 | sd = &per_cpu(node_domains, j).sd; | |
6876 | sd->groups = sg; | |
6877 | } | |
6878 | ||
6879 | sg->cpu_power = 0; | |
6880 | cpumask_copy(sched_group_cpus(sg), d->nodemask); | |
6881 | sg->next = sg; | |
6882 | cpumask_or(d->covered, d->covered, d->nodemask); | |
6883 | ||
6884 | prev = sg; | |
6885 | for (j = 0; j < nr_node_ids; j++) { | |
6886 | n = (num + j) % nr_node_ids; | |
6887 | cpumask_complement(d->notcovered, d->covered); | |
6888 | cpumask_and(d->tmpmask, d->notcovered, cpu_map); | |
6889 | cpumask_and(d->tmpmask, d->tmpmask, d->domainspan); | |
6890 | if (cpumask_empty(d->tmpmask)) | |
6891 | break; | |
6892 | cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n)); | |
6893 | if (cpumask_empty(d->tmpmask)) | |
6894 | continue; | |
6895 | sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(), | |
6896 | GFP_KERNEL, num); | |
6897 | if (!sg) { | |
6898 | printk(KERN_WARNING | |
6899 | "Can not alloc domain group for node %d\n", j); | |
6900 | return -ENOMEM; | |
6901 | } | |
6902 | sg->cpu_power = 0; | |
6903 | cpumask_copy(sched_group_cpus(sg), d->tmpmask); | |
6904 | sg->next = prev->next; | |
6905 | cpumask_or(d->covered, d->covered, d->tmpmask); | |
6906 | prev->next = sg; | |
6907 | prev = sg; | |
6908 | } | |
6909 | out: | |
6910 | return 0; | |
6911 | } | |
6912 | #endif /* CONFIG_NUMA */ | |
6913 | ||
6914 | #ifdef CONFIG_NUMA | |
6915 | /* Free memory allocated for various sched_group structures */ | |
6916 | static void free_sched_groups(const struct cpumask *cpu_map, | |
6917 | struct cpumask *nodemask) | |
6918 | { | |
6919 | int cpu, i; | |
6920 | ||
6921 | for_each_cpu(cpu, cpu_map) { | |
6922 | struct sched_group **sched_group_nodes | |
6923 | = sched_group_nodes_bycpu[cpu]; | |
6924 | ||
6925 | if (!sched_group_nodes) | |
6926 | continue; | |
6927 | ||
6928 | for (i = 0; i < nr_node_ids; i++) { | |
6929 | struct sched_group *oldsg, *sg = sched_group_nodes[i]; | |
6930 | ||
6931 | cpumask_and(nodemask, cpumask_of_node(i), cpu_map); | |
6932 | if (cpumask_empty(nodemask)) | |
6933 | continue; | |
6934 | ||
6935 | if (sg == NULL) | |
6936 | continue; | |
6937 | sg = sg->next; | |
6938 | next_sg: | |
6939 | oldsg = sg; | |
6940 | sg = sg->next; | |
6941 | kfree(oldsg); | |
6942 | if (oldsg != sched_group_nodes[i]) | |
6943 | goto next_sg; | |
6944 | } | |
6945 | kfree(sched_group_nodes); | |
6946 | sched_group_nodes_bycpu[cpu] = NULL; | |
6947 | } | |
6948 | } | |
6949 | #else /* !CONFIG_NUMA */ | |
6950 | static void free_sched_groups(const struct cpumask *cpu_map, | |
6951 | struct cpumask *nodemask) | |
6952 | { | |
6953 | } | |
6954 | #endif /* CONFIG_NUMA */ | |
6955 | ||
6956 | /* | |
6957 | * Initialize sched groups cpu_power. | |
6958 | * | |
6959 | * cpu_power indicates the capacity of sched group, which is used while | |
6960 | * distributing the load between different sched groups in a sched domain. | |
6961 | * Typically cpu_power for all the groups in a sched domain will be same unless | |
6962 | * there are asymmetries in the topology. If there are asymmetries, group | |
6963 | * having more cpu_power will pickup more load compared to the group having | |
6964 | * less cpu_power. | |
6965 | */ | |
6966 | static void init_sched_groups_power(int cpu, struct sched_domain *sd) | |
6967 | { | |
6968 | struct sched_domain *child; | |
6969 | struct sched_group *group; | |
6970 | long power; | |
6971 | int weight; | |
6972 | ||
6973 | WARN_ON(!sd || !sd->groups); | |
6974 | ||
6975 | if (cpu != group_first_cpu(sd->groups)) | |
6976 | return; | |
6977 | ||
6978 | sd->groups->group_weight = cpumask_weight(sched_group_cpus(sd->groups)); | |
6979 | ||
6980 | child = sd->child; | |
6981 | ||
6982 | sd->groups->cpu_power = 0; | |
6983 | ||
6984 | if (!child) { | |
6985 | power = SCHED_LOAD_SCALE; | |
6986 | weight = cpumask_weight(sched_domain_span(sd)); | |
6987 | /* | |
6988 | * SMT siblings share the power of a single core. | |
6989 | * Usually multiple threads get a better yield out of | |
6990 | * that one core than a single thread would have, | |
6991 | * reflect that in sd->smt_gain. | |
6992 | */ | |
6993 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { | |
6994 | power *= sd->smt_gain; | |
6995 | power /= weight; | |
6996 | power >>= SCHED_LOAD_SHIFT; | |
6997 | } | |
6998 | sd->groups->cpu_power += power; | |
6999 | return; | |
7000 | } | |
7001 | ||
7002 | /* | |
7003 | * Add cpu_power of each child group to this groups cpu_power. | |
7004 | */ | |
7005 | group = child->groups; | |
7006 | do { | |
7007 | sd->groups->cpu_power += group->cpu_power; | |
7008 | group = group->next; | |
7009 | } while (group != child->groups); | |
7010 | } | |
7011 | ||
7012 | /* | |
7013 | * Initializers for schedule domains | |
7014 | * Non-inlined to reduce accumulated stack pressure in build_sched_domains() | |
7015 | */ | |
7016 | ||
7017 | #ifdef CONFIG_SCHED_DEBUG | |
7018 | # define SD_INIT_NAME(sd, type) sd->name = #type | |
7019 | #else | |
7020 | # define SD_INIT_NAME(sd, type) do { } while (0) | |
7021 | #endif | |
7022 | ||
7023 | #define SD_INIT(sd, type) sd_init_##type(sd) | |
7024 | ||
7025 | #define SD_INIT_FUNC(type) \ | |
7026 | static noinline void sd_init_##type(struct sched_domain *sd) \ | |
7027 | { \ | |
7028 | memset(sd, 0, sizeof(*sd)); \ | |
7029 | *sd = SD_##type##_INIT; \ | |
7030 | sd->level = SD_LV_##type; \ | |
7031 | SD_INIT_NAME(sd, type); \ | |
7032 | } | |
7033 | ||
7034 | SD_INIT_FUNC(CPU) | |
7035 | #ifdef CONFIG_NUMA | |
7036 | SD_INIT_FUNC(ALLNODES) | |
7037 | SD_INIT_FUNC(NODE) | |
7038 | #endif | |
7039 | #ifdef CONFIG_SCHED_SMT | |
7040 | SD_INIT_FUNC(SIBLING) | |
7041 | #endif | |
7042 | #ifdef CONFIG_SCHED_MC | |
7043 | SD_INIT_FUNC(MC) | |
7044 | #endif | |
7045 | #ifdef CONFIG_SCHED_BOOK | |
7046 | SD_INIT_FUNC(BOOK) | |
7047 | #endif | |
7048 | ||
7049 | static int default_relax_domain_level = -1; | |
7050 | ||
7051 | static int __init setup_relax_domain_level(char *str) | |
7052 | { | |
7053 | unsigned long val; | |
7054 | ||
7055 | val = simple_strtoul(str, NULL, 0); | |
7056 | if (val < SD_LV_MAX) | |
7057 | default_relax_domain_level = val; | |
7058 | ||
7059 | return 1; | |
7060 | } | |
7061 | __setup("relax_domain_level=", setup_relax_domain_level); | |
7062 | ||
7063 | static void set_domain_attribute(struct sched_domain *sd, | |
7064 | struct sched_domain_attr *attr) | |
7065 | { | |
7066 | int request; | |
7067 | ||
7068 | if (!attr || attr->relax_domain_level < 0) { | |
7069 | if (default_relax_domain_level < 0) | |
7070 | return; | |
7071 | else | |
7072 | request = default_relax_domain_level; | |
7073 | } else | |
7074 | request = attr->relax_domain_level; | |
7075 | if (request < sd->level) { | |
7076 | /* turn off idle balance on this domain */ | |
7077 | sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); | |
7078 | } else { | |
7079 | /* turn on idle balance on this domain */ | |
7080 | sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); | |
7081 | } | |
7082 | } | |
7083 | ||
7084 | static void __free_domain_allocs(struct s_data *d, enum s_alloc what, | |
7085 | const struct cpumask *cpu_map) | |
7086 | { | |
7087 | switch (what) { | |
7088 | case sa_sched_groups: | |
7089 | free_sched_groups(cpu_map, d->tmpmask); /* fall through */ | |
7090 | d->sched_group_nodes = NULL; | |
7091 | case sa_rootdomain: | |
7092 | free_rootdomain(d->rd); /* fall through */ | |
7093 | case sa_tmpmask: | |
7094 | free_cpumask_var(d->tmpmask); /* fall through */ | |
7095 | case sa_send_covered: | |
7096 | free_cpumask_var(d->send_covered); /* fall through */ | |
7097 | case sa_this_book_map: | |
7098 | free_cpumask_var(d->this_book_map); /* fall through */ | |
7099 | case sa_this_core_map: | |
7100 | free_cpumask_var(d->this_core_map); /* fall through */ | |
7101 | case sa_this_sibling_map: | |
7102 | free_cpumask_var(d->this_sibling_map); /* fall through */ | |
7103 | case sa_nodemask: | |
7104 | free_cpumask_var(d->nodemask); /* fall through */ | |
7105 | case sa_sched_group_nodes: | |
7106 | #ifdef CONFIG_NUMA | |
7107 | kfree(d->sched_group_nodes); /* fall through */ | |
7108 | case sa_notcovered: | |
7109 | free_cpumask_var(d->notcovered); /* fall through */ | |
7110 | case sa_covered: | |
7111 | free_cpumask_var(d->covered); /* fall through */ | |
7112 | case sa_domainspan: | |
7113 | free_cpumask_var(d->domainspan); /* fall through */ | |
7114 | #endif | |
7115 | case sa_none: | |
7116 | break; | |
7117 | } | |
7118 | } | |
7119 | ||
7120 | static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, | |
7121 | const struct cpumask *cpu_map) | |
7122 | { | |
7123 | #ifdef CONFIG_NUMA | |
7124 | if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL)) | |
7125 | return sa_none; | |
7126 | if (!alloc_cpumask_var(&d->covered, GFP_KERNEL)) | |
7127 | return sa_domainspan; | |
7128 | if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL)) | |
7129 | return sa_covered; | |
7130 | /* Allocate the per-node list of sched groups */ | |
7131 | d->sched_group_nodes = kcalloc(nr_node_ids, | |
7132 | sizeof(struct sched_group *), GFP_KERNEL); | |
7133 | if (!d->sched_group_nodes) { | |
7134 | printk(KERN_WARNING "Can not alloc sched group node list\n"); | |
7135 | return sa_notcovered; | |
7136 | } | |
7137 | sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes; | |
7138 | #endif | |
7139 | if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL)) | |
7140 | return sa_sched_group_nodes; | |
7141 | if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL)) | |
7142 | return sa_nodemask; | |
7143 | if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL)) | |
7144 | return sa_this_sibling_map; | |
7145 | if (!alloc_cpumask_var(&d->this_book_map, GFP_KERNEL)) | |
7146 | return sa_this_core_map; | |
7147 | if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL)) | |
7148 | return sa_this_book_map; | |
7149 | if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL)) | |
7150 | return sa_send_covered; | |
7151 | d->rd = alloc_rootdomain(); | |
7152 | if (!d->rd) { | |
7153 | printk(KERN_WARNING "Cannot alloc root domain\n"); | |
7154 | return sa_tmpmask; | |
7155 | } | |
7156 | return sa_rootdomain; | |
7157 | } | |
7158 | ||
7159 | static struct sched_domain *__build_numa_sched_domains(struct s_data *d, | |
7160 | const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i) | |
7161 | { | |
7162 | struct sched_domain *sd = NULL; | |
7163 | #ifdef CONFIG_NUMA | |
7164 | struct sched_domain *parent; | |
7165 | ||
7166 | d->sd_allnodes = 0; | |
7167 | if (cpumask_weight(cpu_map) > | |
7168 | SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) { | |
7169 | sd = &per_cpu(allnodes_domains, i).sd; | |
7170 | SD_INIT(sd, ALLNODES); | |
7171 | set_domain_attribute(sd, attr); | |
7172 | cpumask_copy(sched_domain_span(sd), cpu_map); | |
7173 | cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask); | |
7174 | d->sd_allnodes = 1; | |
7175 | } | |
7176 | parent = sd; | |
7177 | ||
7178 | sd = &per_cpu(node_domains, i).sd; | |
7179 | SD_INIT(sd, NODE); | |
7180 | set_domain_attribute(sd, attr); | |
7181 | sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd)); | |
7182 | sd->parent = parent; | |
7183 | if (parent) | |
7184 | parent->child = sd; | |
7185 | cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map); | |
7186 | #endif | |
7187 | return sd; | |
7188 | } | |
7189 | ||
7190 | static struct sched_domain *__build_cpu_sched_domain(struct s_data *d, | |
7191 | const struct cpumask *cpu_map, struct sched_domain_attr *attr, | |
7192 | struct sched_domain *parent, int i) | |
7193 | { | |
7194 | struct sched_domain *sd; | |
7195 | sd = &per_cpu(phys_domains, i).sd; | |
7196 | SD_INIT(sd, CPU); | |
7197 | set_domain_attribute(sd, attr); | |
7198 | cpumask_copy(sched_domain_span(sd), d->nodemask); | |
7199 | sd->parent = parent; | |
7200 | if (parent) | |
7201 | parent->child = sd; | |
7202 | cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask); | |
7203 | return sd; | |
7204 | } | |
7205 | ||
7206 | static struct sched_domain *__build_book_sched_domain(struct s_data *d, | |
7207 | const struct cpumask *cpu_map, struct sched_domain_attr *attr, | |
7208 | struct sched_domain *parent, int i) | |
7209 | { | |
7210 | struct sched_domain *sd = parent; | |
7211 | #ifdef CONFIG_SCHED_BOOK | |
7212 | sd = &per_cpu(book_domains, i).sd; | |
7213 | SD_INIT(sd, BOOK); | |
7214 | set_domain_attribute(sd, attr); | |
7215 | cpumask_and(sched_domain_span(sd), cpu_map, cpu_book_mask(i)); | |
7216 | sd->parent = parent; | |
7217 | parent->child = sd; | |
7218 | cpu_to_book_group(i, cpu_map, &sd->groups, d->tmpmask); | |
7219 | #endif | |
7220 | return sd; | |
7221 | } | |
7222 | ||
7223 | static struct sched_domain *__build_mc_sched_domain(struct s_data *d, | |
7224 | const struct cpumask *cpu_map, struct sched_domain_attr *attr, | |
7225 | struct sched_domain *parent, int i) | |
7226 | { | |
7227 | struct sched_domain *sd = parent; | |
7228 | #ifdef CONFIG_SCHED_MC | |
7229 | sd = &per_cpu(core_domains, i).sd; | |
7230 | SD_INIT(sd, MC); | |
7231 | set_domain_attribute(sd, attr); | |
7232 | cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i)); | |
7233 | sd->parent = parent; | |
7234 | parent->child = sd; | |
7235 | cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask); | |
7236 | #endif | |
7237 | return sd; | |
7238 | } | |
7239 | ||
7240 | static struct sched_domain *__build_smt_sched_domain(struct s_data *d, | |
7241 | const struct cpumask *cpu_map, struct sched_domain_attr *attr, | |
7242 | struct sched_domain *parent, int i) | |
7243 | { | |
7244 | struct sched_domain *sd = parent; | |
7245 | #ifdef CONFIG_SCHED_SMT | |
7246 | sd = &per_cpu(cpu_domains, i).sd; | |
7247 | SD_INIT(sd, SIBLING); | |
7248 | set_domain_attribute(sd, attr); | |
7249 | cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i)); | |
7250 | sd->parent = parent; | |
7251 | parent->child = sd; | |
7252 | cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask); | |
7253 | #endif | |
7254 | return sd; | |
7255 | } | |
7256 | ||
7257 | static void build_sched_groups(struct s_data *d, enum sched_domain_level l, | |
7258 | const struct cpumask *cpu_map, int cpu) | |
7259 | { | |
7260 | switch (l) { | |
7261 | #ifdef CONFIG_SCHED_SMT | |
7262 | case SD_LV_SIBLING: /* set up CPU (sibling) groups */ | |
7263 | cpumask_and(d->this_sibling_map, cpu_map, | |
7264 | topology_thread_cpumask(cpu)); | |
7265 | if (cpu == cpumask_first(d->this_sibling_map)) | |
7266 | init_sched_build_groups(d->this_sibling_map, cpu_map, | |
7267 | &cpu_to_cpu_group, | |
7268 | d->send_covered, d->tmpmask); | |
7269 | break; | |
7270 | #endif | |
7271 | #ifdef CONFIG_SCHED_MC | |
7272 | case SD_LV_MC: /* set up multi-core groups */ | |
7273 | cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu)); | |
7274 | if (cpu == cpumask_first(d->this_core_map)) | |
7275 | init_sched_build_groups(d->this_core_map, cpu_map, | |
7276 | &cpu_to_core_group, | |
7277 | d->send_covered, d->tmpmask); | |
7278 | break; | |
7279 | #endif | |
7280 | #ifdef CONFIG_SCHED_BOOK | |
7281 | case SD_LV_BOOK: /* set up book groups */ | |
7282 | cpumask_and(d->this_book_map, cpu_map, cpu_book_mask(cpu)); | |
7283 | if (cpu == cpumask_first(d->this_book_map)) | |
7284 | init_sched_build_groups(d->this_book_map, cpu_map, | |
7285 | &cpu_to_book_group, | |
7286 | d->send_covered, d->tmpmask); | |
7287 | break; | |
7288 | #endif | |
7289 | case SD_LV_CPU: /* set up physical groups */ | |
7290 | cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map); | |
7291 | if (!cpumask_empty(d->nodemask)) | |
7292 | init_sched_build_groups(d->nodemask, cpu_map, | |
7293 | &cpu_to_phys_group, | |
7294 | d->send_covered, d->tmpmask); | |
7295 | break; | |
7296 | #ifdef CONFIG_NUMA | |
7297 | case SD_LV_ALLNODES: | |
7298 | init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group, | |
7299 | d->send_covered, d->tmpmask); | |
7300 | break; | |
7301 | #endif | |
7302 | default: | |
7303 | break; | |
7304 | } | |
7305 | } | |
7306 | ||
7307 | /* | |
7308 | * Build sched domains for a given set of cpus and attach the sched domains | |
7309 | * to the individual cpus | |
7310 | */ | |
7311 | static int __build_sched_domains(const struct cpumask *cpu_map, | |
7312 | struct sched_domain_attr *attr) | |
7313 | { | |
7314 | enum s_alloc alloc_state = sa_none; | |
7315 | struct s_data d; | |
7316 | struct sched_domain *sd; | |
7317 | int i; | |
7318 | #ifdef CONFIG_NUMA | |
7319 | d.sd_allnodes = 0; | |
7320 | #endif | |
7321 | ||
7322 | alloc_state = __visit_domain_allocation_hell(&d, cpu_map); | |
7323 | if (alloc_state != sa_rootdomain) | |
7324 | goto error; | |
7325 | alloc_state = sa_sched_groups; | |
7326 | ||
7327 | /* | |
7328 | * Set up domains for cpus specified by the cpu_map. | |
7329 | */ | |
7330 | for_each_cpu(i, cpu_map) { | |
7331 | cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)), | |
7332 | cpu_map); | |
7333 | ||
7334 | sd = __build_numa_sched_domains(&d, cpu_map, attr, i); | |
7335 | sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i); | |
7336 | sd = __build_book_sched_domain(&d, cpu_map, attr, sd, i); | |
7337 | sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i); | |
7338 | sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i); | |
7339 | } | |
7340 | ||
7341 | for_each_cpu(i, cpu_map) { | |
7342 | build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i); | |
7343 | build_sched_groups(&d, SD_LV_BOOK, cpu_map, i); | |
7344 | build_sched_groups(&d, SD_LV_MC, cpu_map, i); | |
7345 | } | |
7346 | ||
7347 | /* Set up physical groups */ | |
7348 | for (i = 0; i < nr_node_ids; i++) | |
7349 | build_sched_groups(&d, SD_LV_CPU, cpu_map, i); | |
7350 | ||
7351 | #ifdef CONFIG_NUMA | |
7352 | /* Set up node groups */ | |
7353 | if (d.sd_allnodes) | |
7354 | build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0); | |
7355 | ||
7356 | for (i = 0; i < nr_node_ids; i++) | |
7357 | if (build_numa_sched_groups(&d, cpu_map, i)) | |
7358 | goto error; | |
7359 | #endif | |
7360 | ||
7361 | /* Calculate CPU power for physical packages and nodes */ | |
7362 | #ifdef CONFIG_SCHED_SMT | |
7363 | for_each_cpu(i, cpu_map) { | |
7364 | sd = &per_cpu(cpu_domains, i).sd; | |
7365 | init_sched_groups_power(i, sd); | |
7366 | } | |
7367 | #endif | |
7368 | #ifdef CONFIG_SCHED_MC | |
7369 | for_each_cpu(i, cpu_map) { | |
7370 | sd = &per_cpu(core_domains, i).sd; | |
7371 | init_sched_groups_power(i, sd); | |
7372 | } | |
7373 | #endif | |
7374 | #ifdef CONFIG_SCHED_BOOK | |
7375 | for_each_cpu(i, cpu_map) { | |
7376 | sd = &per_cpu(book_domains, i).sd; | |
7377 | init_sched_groups_power(i, sd); | |
7378 | } | |
7379 | #endif | |
7380 | ||
7381 | for_each_cpu(i, cpu_map) { | |
7382 | sd = &per_cpu(phys_domains, i).sd; | |
7383 | init_sched_groups_power(i, sd); | |
7384 | } | |
7385 | ||
7386 | #ifdef CONFIG_NUMA | |
7387 | for (i = 0; i < nr_node_ids; i++) | |
7388 | init_numa_sched_groups_power(d.sched_group_nodes[i]); | |
7389 | ||
7390 | if (d.sd_allnodes) { | |
7391 | struct sched_group *sg; | |
7392 | ||
7393 | cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg, | |
7394 | d.tmpmask); | |
7395 | init_numa_sched_groups_power(sg); | |
7396 | } | |
7397 | #endif | |
7398 | ||
7399 | /* Attach the domains */ | |
7400 | for_each_cpu(i, cpu_map) { | |
7401 | #ifdef CONFIG_SCHED_SMT | |
7402 | sd = &per_cpu(cpu_domains, i).sd; | |
7403 | #elif defined(CONFIG_SCHED_MC) | |
7404 | sd = &per_cpu(core_domains, i).sd; | |
7405 | #elif defined(CONFIG_SCHED_BOOK) | |
7406 | sd = &per_cpu(book_domains, i).sd; | |
7407 | #else | |
7408 | sd = &per_cpu(phys_domains, i).sd; | |
7409 | #endif | |
7410 | cpu_attach_domain(sd, d.rd, i); | |
7411 | } | |
7412 | ||
7413 | d.sched_group_nodes = NULL; /* don't free this we still need it */ | |
7414 | __free_domain_allocs(&d, sa_tmpmask, cpu_map); | |
7415 | return 0; | |
7416 | ||
7417 | error: | |
7418 | __free_domain_allocs(&d, alloc_state, cpu_map); | |
7419 | return -ENOMEM; | |
7420 | } | |
7421 | ||
7422 | static int build_sched_domains(const struct cpumask *cpu_map) | |
7423 | { | |
7424 | return __build_sched_domains(cpu_map, NULL); | |
7425 | } | |
7426 | ||
7427 | static cpumask_var_t *doms_cur; /* current sched domains */ | |
7428 | static int ndoms_cur; /* number of sched domains in 'doms_cur' */ | |
7429 | static struct sched_domain_attr *dattr_cur; | |
7430 | /* attribues of custom domains in 'doms_cur' */ | |
7431 | ||
7432 | /* | |
7433 | * Special case: If a kmalloc of a doms_cur partition (array of | |
7434 | * cpumask) fails, then fallback to a single sched domain, | |
7435 | * as determined by the single cpumask fallback_doms. | |
7436 | */ | |
7437 | static cpumask_var_t fallback_doms; | |
7438 | ||
7439 | /* | |
7440 | * arch_update_cpu_topology lets virtualized architectures update the | |
7441 | * cpu core maps. It is supposed to return 1 if the topology changed | |
7442 | * or 0 if it stayed the same. | |
7443 | */ | |
7444 | int __attribute__((weak)) arch_update_cpu_topology(void) | |
7445 | { | |
7446 | return 0; | |
7447 | } | |
7448 | ||
7449 | cpumask_var_t *alloc_sched_domains(unsigned int ndoms) | |
7450 | { | |
7451 | int i; | |
7452 | cpumask_var_t *doms; | |
7453 | ||
7454 | doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); | |
7455 | if (!doms) | |
7456 | return NULL; | |
7457 | for (i = 0; i < ndoms; i++) { | |
7458 | if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { | |
7459 | free_sched_domains(doms, i); | |
7460 | return NULL; | |
7461 | } | |
7462 | } | |
7463 | return doms; | |
7464 | } | |
7465 | ||
7466 | void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) | |
7467 | { | |
7468 | unsigned int i; | |
7469 | for (i = 0; i < ndoms; i++) | |
7470 | free_cpumask_var(doms[i]); | |
7471 | kfree(doms); | |
7472 | } | |
7473 | ||
7474 | /* | |
7475 | * Set up scheduler domains and groups. Callers must hold the hotplug lock. | |
7476 | * For now this just excludes isolated cpus, but could be used to | |
7477 | * exclude other special cases in the future. | |
7478 | */ | |
7479 | static int arch_init_sched_domains(const struct cpumask *cpu_map) | |
7480 | { | |
7481 | int err; | |
7482 | ||
7483 | arch_update_cpu_topology(); | |
7484 | ndoms_cur = 1; | |
7485 | doms_cur = alloc_sched_domains(ndoms_cur); | |
7486 | if (!doms_cur) | |
7487 | doms_cur = &fallback_doms; | |
7488 | cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); | |
7489 | dattr_cur = NULL; | |
7490 | err = build_sched_domains(doms_cur[0]); | |
7491 | register_sched_domain_sysctl(); | |
7492 | ||
7493 | return err; | |
7494 | } | |
7495 | ||
7496 | static void arch_destroy_sched_domains(const struct cpumask *cpu_map, | |
7497 | struct cpumask *tmpmask) | |
7498 | { | |
7499 | free_sched_groups(cpu_map, tmpmask); | |
7500 | } | |
7501 | ||
7502 | /* | |
7503 | * Detach sched domains from a group of cpus specified in cpu_map | |
7504 | * These cpus will now be attached to the NULL domain | |
7505 | */ | |
7506 | static void detach_destroy_domains(const struct cpumask *cpu_map) | |
7507 | { | |
7508 | /* Save because hotplug lock held. */ | |
7509 | static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS); | |
7510 | int i; | |
7511 | ||
7512 | for_each_cpu(i, cpu_map) | |
7513 | cpu_attach_domain(NULL, &def_root_domain, i); | |
7514 | synchronize_sched(); | |
7515 | arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask)); | |
7516 | } | |
7517 | ||
7518 | /* handle null as "default" */ | |
7519 | static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, | |
7520 | struct sched_domain_attr *new, int idx_new) | |
7521 | { | |
7522 | struct sched_domain_attr tmp; | |
7523 | ||
7524 | /* fast path */ | |
7525 | if (!new && !cur) | |
7526 | return 1; | |
7527 | ||
7528 | tmp = SD_ATTR_INIT; | |
7529 | return !memcmp(cur ? (cur + idx_cur) : &tmp, | |
7530 | new ? (new + idx_new) : &tmp, | |
7531 | sizeof(struct sched_domain_attr)); | |
7532 | } | |
7533 | ||
7534 | /* | |
7535 | * Partition sched domains as specified by the 'ndoms_new' | |
7536 | * cpumasks in the array doms_new[] of cpumasks. This compares | |
7537 | * doms_new[] to the current sched domain partitioning, doms_cur[]. | |
7538 | * It destroys each deleted domain and builds each new domain. | |
7539 | * | |
7540 | * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. | |
7541 | * The masks don't intersect (don't overlap.) We should setup one | |
7542 | * sched domain for each mask. CPUs not in any of the cpumasks will | |
7543 | * not be load balanced. If the same cpumask appears both in the | |
7544 | * current 'doms_cur' domains and in the new 'doms_new', we can leave | |
7545 | * it as it is. | |
7546 | * | |
7547 | * The passed in 'doms_new' should be allocated using | |
7548 | * alloc_sched_domains. This routine takes ownership of it and will | |
7549 | * free_sched_domains it when done with it. If the caller failed the | |
7550 | * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, | |
7551 | * and partition_sched_domains() will fallback to the single partition | |
7552 | * 'fallback_doms', it also forces the domains to be rebuilt. | |
7553 | * | |
7554 | * If doms_new == NULL it will be replaced with cpu_online_mask. | |
7555 | * ndoms_new == 0 is a special case for destroying existing domains, | |
7556 | * and it will not create the default domain. | |
7557 | * | |
7558 | * Call with hotplug lock held | |
7559 | */ | |
7560 | void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], | |
7561 | struct sched_domain_attr *dattr_new) | |
7562 | { | |
7563 | int i, j, n; | |
7564 | int new_topology; | |
7565 | ||
7566 | mutex_lock(&sched_domains_mutex); | |
7567 | ||
7568 | /* always unregister in case we don't destroy any domains */ | |
7569 | unregister_sched_domain_sysctl(); | |
7570 | ||
7571 | /* Let architecture update cpu core mappings. */ | |
7572 | new_topology = arch_update_cpu_topology(); | |
7573 | ||
7574 | n = doms_new ? ndoms_new : 0; | |
7575 | ||
7576 | /* Destroy deleted domains */ | |
7577 | for (i = 0; i < ndoms_cur; i++) { | |
7578 | for (j = 0; j < n && !new_topology; j++) { | |
7579 | if (cpumask_equal(doms_cur[i], doms_new[j]) | |
7580 | && dattrs_equal(dattr_cur, i, dattr_new, j)) | |
7581 | goto match1; | |
7582 | } | |
7583 | /* no match - a current sched domain not in new doms_new[] */ | |
7584 | detach_destroy_domains(doms_cur[i]); | |
7585 | match1: | |
7586 | ; | |
7587 | } | |
7588 | ||
7589 | if (doms_new == NULL) { | |
7590 | ndoms_cur = 0; | |
7591 | doms_new = &fallback_doms; | |
7592 | cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); | |
7593 | WARN_ON_ONCE(dattr_new); | |
7594 | } | |
7595 | ||
7596 | /* Build new domains */ | |
7597 | for (i = 0; i < ndoms_new; i++) { | |
7598 | for (j = 0; j < ndoms_cur && !new_topology; j++) { | |
7599 | if (cpumask_equal(doms_new[i], doms_cur[j]) | |
7600 | && dattrs_equal(dattr_new, i, dattr_cur, j)) | |
7601 | goto match2; | |
7602 | } | |
7603 | /* no match - add a new doms_new */ | |
7604 | __build_sched_domains(doms_new[i], | |
7605 | dattr_new ? dattr_new + i : NULL); | |
7606 | match2: | |
7607 | ; | |
7608 | } | |
7609 | ||
7610 | /* Remember the new sched domains */ | |
7611 | if (doms_cur != &fallback_doms) | |
7612 | free_sched_domains(doms_cur, ndoms_cur); | |
7613 | kfree(dattr_cur); /* kfree(NULL) is safe */ | |
7614 | doms_cur = doms_new; | |
7615 | dattr_cur = dattr_new; | |
7616 | ndoms_cur = ndoms_new; | |
7617 | ||
7618 | register_sched_domain_sysctl(); | |
7619 | ||
7620 | mutex_unlock(&sched_domains_mutex); | |
7621 | } | |
7622 | ||
7623 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
7624 | static void arch_reinit_sched_domains(void) | |
7625 | { | |
7626 | get_online_cpus(); | |
7627 | ||
7628 | /* Destroy domains first to force the rebuild */ | |
7629 | partition_sched_domains(0, NULL, NULL); | |
7630 | ||
7631 | rebuild_sched_domains(); | |
7632 | put_online_cpus(); | |
7633 | } | |
7634 | ||
7635 | static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt) | |
7636 | { | |
7637 | unsigned int level = 0; | |
7638 | ||
7639 | if (sscanf(buf, "%u", &level) != 1) | |
7640 | return -EINVAL; | |
7641 | ||
7642 | /* | |
7643 | * level is always be positive so don't check for | |
7644 | * level < POWERSAVINGS_BALANCE_NONE which is 0 | |
7645 | * What happens on 0 or 1 byte write, | |
7646 | * need to check for count as well? | |
7647 | */ | |
7648 | ||
7649 | if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS) | |
7650 | return -EINVAL; | |
7651 | ||
7652 | if (smt) | |
7653 | sched_smt_power_savings = level; | |
7654 | else | |
7655 | sched_mc_power_savings = level; | |
7656 | ||
7657 | arch_reinit_sched_domains(); | |
7658 | ||
7659 | return count; | |
7660 | } | |
7661 | ||
7662 | #ifdef CONFIG_SCHED_MC | |
7663 | static ssize_t sched_mc_power_savings_show(struct sysdev_class *class, | |
7664 | struct sysdev_class_attribute *attr, | |
7665 | char *page) | |
7666 | { | |
7667 | return sprintf(page, "%u\n", sched_mc_power_savings); | |
7668 | } | |
7669 | static ssize_t sched_mc_power_savings_store(struct sysdev_class *class, | |
7670 | struct sysdev_class_attribute *attr, | |
7671 | const char *buf, size_t count) | |
7672 | { | |
7673 | return sched_power_savings_store(buf, count, 0); | |
7674 | } | |
7675 | static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644, | |
7676 | sched_mc_power_savings_show, | |
7677 | sched_mc_power_savings_store); | |
7678 | #endif | |
7679 | ||
7680 | #ifdef CONFIG_SCHED_SMT | |
7681 | static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev, | |
7682 | struct sysdev_class_attribute *attr, | |
7683 | char *page) | |
7684 | { | |
7685 | return sprintf(page, "%u\n", sched_smt_power_savings); | |
7686 | } | |
7687 | static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev, | |
7688 | struct sysdev_class_attribute *attr, | |
7689 | const char *buf, size_t count) | |
7690 | { | |
7691 | return sched_power_savings_store(buf, count, 1); | |
7692 | } | |
7693 | static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644, | |
7694 | sched_smt_power_savings_show, | |
7695 | sched_smt_power_savings_store); | |
7696 | #endif | |
7697 | ||
7698 | int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls) | |
7699 | { | |
7700 | int err = 0; | |
7701 | ||
7702 | #ifdef CONFIG_SCHED_SMT | |
7703 | if (smt_capable()) | |
7704 | err = sysfs_create_file(&cls->kset.kobj, | |
7705 | &attr_sched_smt_power_savings.attr); | |
7706 | #endif | |
7707 | #ifdef CONFIG_SCHED_MC | |
7708 | if (!err && mc_capable()) | |
7709 | err = sysfs_create_file(&cls->kset.kobj, | |
7710 | &attr_sched_mc_power_savings.attr); | |
7711 | #endif | |
7712 | return err; | |
7713 | } | |
7714 | #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ | |
7715 | ||
7716 | /* | |
7717 | * Update cpusets according to cpu_active mask. If cpusets are | |
7718 | * disabled, cpuset_update_active_cpus() becomes a simple wrapper | |
7719 | * around partition_sched_domains(). | |
7720 | */ | |
7721 | static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, | |
7722 | void *hcpu) | |
7723 | { | |
7724 | switch (action & ~CPU_TASKS_FROZEN) { | |
7725 | case CPU_ONLINE: | |
7726 | case CPU_DOWN_FAILED: | |
7727 | cpuset_update_active_cpus(); | |
7728 | return NOTIFY_OK; | |
7729 | default: | |
7730 | return NOTIFY_DONE; | |
7731 | } | |
7732 | } | |
7733 | ||
7734 | static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, | |
7735 | void *hcpu) | |
7736 | { | |
7737 | switch (action & ~CPU_TASKS_FROZEN) { | |
7738 | case CPU_DOWN_PREPARE: | |
7739 | cpuset_update_active_cpus(); | |
7740 | return NOTIFY_OK; | |
7741 | default: | |
7742 | return NOTIFY_DONE; | |
7743 | } | |
7744 | } | |
7745 | ||
7746 | static int update_runtime(struct notifier_block *nfb, | |
7747 | unsigned long action, void *hcpu) | |
7748 | { | |
7749 | int cpu = (int)(long)hcpu; | |
7750 | ||
7751 | switch (action) { | |
7752 | case CPU_DOWN_PREPARE: | |
7753 | case CPU_DOWN_PREPARE_FROZEN: | |
7754 | disable_runtime(cpu_rq(cpu)); | |
7755 | return NOTIFY_OK; | |
7756 | ||
7757 | case CPU_DOWN_FAILED: | |
7758 | case CPU_DOWN_FAILED_FROZEN: | |
7759 | case CPU_ONLINE: | |
7760 | case CPU_ONLINE_FROZEN: | |
7761 | enable_runtime(cpu_rq(cpu)); | |
7762 | return NOTIFY_OK; | |
7763 | ||
7764 | default: | |
7765 | return NOTIFY_DONE; | |
7766 | } | |
7767 | } | |
7768 | ||
7769 | void __init sched_init_smp(void) | |
7770 | { | |
7771 | cpumask_var_t non_isolated_cpus; | |
7772 | ||
7773 | alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); | |
7774 | alloc_cpumask_var(&fallback_doms, GFP_KERNEL); | |
7775 | ||
7776 | #if defined(CONFIG_NUMA) | |
7777 | sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **), | |
7778 | GFP_KERNEL); | |
7779 | BUG_ON(sched_group_nodes_bycpu == NULL); | |
7780 | #endif | |
7781 | get_online_cpus(); | |
7782 | mutex_lock(&sched_domains_mutex); | |
7783 | arch_init_sched_domains(cpu_active_mask); | |
7784 | cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); | |
7785 | if (cpumask_empty(non_isolated_cpus)) | |
7786 | cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); | |
7787 | mutex_unlock(&sched_domains_mutex); | |
7788 | put_online_cpus(); | |
7789 | ||
7790 | hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); | |
7791 | hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); | |
7792 | ||
7793 | /* RT runtime code needs to handle some hotplug events */ | |
7794 | hotcpu_notifier(update_runtime, 0); | |
7795 | ||
7796 | init_hrtick(); | |
7797 | ||
7798 | /* Move init over to a non-isolated CPU */ | |
7799 | if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) | |
7800 | BUG(); | |
7801 | sched_init_granularity(); | |
7802 | free_cpumask_var(non_isolated_cpus); | |
7803 | ||
7804 | init_sched_rt_class(); | |
7805 | } | |
7806 | #else | |
7807 | void __init sched_init_smp(void) | |
7808 | { | |
7809 | sched_init_granularity(); | |
7810 | } | |
7811 | #endif /* CONFIG_SMP */ | |
7812 | ||
7813 | const_debug unsigned int sysctl_timer_migration = 1; | |
7814 | ||
7815 | int in_sched_functions(unsigned long addr) | |
7816 | { | |
7817 | return in_lock_functions(addr) || | |
7818 | (addr >= (unsigned long)__sched_text_start | |
7819 | && addr < (unsigned long)__sched_text_end); | |
7820 | } | |
7821 | ||
7822 | static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq) | |
7823 | { | |
7824 | cfs_rq->tasks_timeline = RB_ROOT; | |
7825 | INIT_LIST_HEAD(&cfs_rq->tasks); | |
7826 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
7827 | cfs_rq->rq = rq; | |
7828 | #endif | |
7829 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); | |
7830 | } | |
7831 | ||
7832 | static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq) | |
7833 | { | |
7834 | struct rt_prio_array *array; | |
7835 | int i; | |
7836 | ||
7837 | array = &rt_rq->active; | |
7838 | for (i = 0; i < MAX_RT_PRIO; i++) { | |
7839 | INIT_LIST_HEAD(array->queue + i); | |
7840 | __clear_bit(i, array->bitmap); | |
7841 | } | |
7842 | /* delimiter for bitsearch: */ | |
7843 | __set_bit(MAX_RT_PRIO, array->bitmap); | |
7844 | ||
7845 | #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED | |
7846 | rt_rq->highest_prio.curr = MAX_RT_PRIO; | |
7847 | #ifdef CONFIG_SMP | |
7848 | rt_rq->highest_prio.next = MAX_RT_PRIO; | |
7849 | #endif | |
7850 | #endif | |
7851 | #ifdef CONFIG_SMP | |
7852 | rt_rq->rt_nr_migratory = 0; | |
7853 | rt_rq->overloaded = 0; | |
7854 | plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock); | |
7855 | #endif | |
7856 | ||
7857 | rt_rq->rt_time = 0; | |
7858 | rt_rq->rt_throttled = 0; | |
7859 | rt_rq->rt_runtime = 0; | |
7860 | raw_spin_lock_init(&rt_rq->rt_runtime_lock); | |
7861 | ||
7862 | #ifdef CONFIG_RT_GROUP_SCHED | |
7863 | rt_rq->rt_nr_boosted = 0; | |
7864 | rt_rq->rq = rq; | |
7865 | #endif | |
7866 | } | |
7867 | ||
7868 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
7869 | static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
7870 | struct sched_entity *se, int cpu, int add, | |
7871 | struct sched_entity *parent) | |
7872 | { | |
7873 | struct rq *rq = cpu_rq(cpu); | |
7874 | tg->cfs_rq[cpu] = cfs_rq; | |
7875 | init_cfs_rq(cfs_rq, rq); | |
7876 | cfs_rq->tg = tg; | |
7877 | if (add) | |
7878 | list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list); | |
7879 | ||
7880 | tg->se[cpu] = se; | |
7881 | /* se could be NULL for init_task_group */ | |
7882 | if (!se) | |
7883 | return; | |
7884 | ||
7885 | if (!parent) | |
7886 | se->cfs_rq = &rq->cfs; | |
7887 | else | |
7888 | se->cfs_rq = parent->my_q; | |
7889 | ||
7890 | se->my_q = cfs_rq; | |
7891 | se->load.weight = tg->shares; | |
7892 | se->load.inv_weight = 0; | |
7893 | se->parent = parent; | |
7894 | } | |
7895 | #endif | |
7896 | ||
7897 | #ifdef CONFIG_RT_GROUP_SCHED | |
7898 | static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq, | |
7899 | struct sched_rt_entity *rt_se, int cpu, int add, | |
7900 | struct sched_rt_entity *parent) | |
7901 | { | |
7902 | struct rq *rq = cpu_rq(cpu); | |
7903 | ||
7904 | tg->rt_rq[cpu] = rt_rq; | |
7905 | init_rt_rq(rt_rq, rq); | |
7906 | rt_rq->tg = tg; | |
7907 | rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime; | |
7908 | if (add) | |
7909 | list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list); | |
7910 | ||
7911 | tg->rt_se[cpu] = rt_se; | |
7912 | if (!rt_se) | |
7913 | return; | |
7914 | ||
7915 | if (!parent) | |
7916 | rt_se->rt_rq = &rq->rt; | |
7917 | else | |
7918 | rt_se->rt_rq = parent->my_q; | |
7919 | ||
7920 | rt_se->my_q = rt_rq; | |
7921 | rt_se->parent = parent; | |
7922 | INIT_LIST_HEAD(&rt_se->run_list); | |
7923 | } | |
7924 | #endif | |
7925 | ||
7926 | void __init sched_init(void) | |
7927 | { | |
7928 | int i, j; | |
7929 | unsigned long alloc_size = 0, ptr; | |
7930 | ||
7931 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
7932 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); | |
7933 | #endif | |
7934 | #ifdef CONFIG_RT_GROUP_SCHED | |
7935 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); | |
7936 | #endif | |
7937 | #ifdef CONFIG_CPUMASK_OFFSTACK | |
7938 | alloc_size += num_possible_cpus() * cpumask_size(); | |
7939 | #endif | |
7940 | if (alloc_size) { | |
7941 | ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); | |
7942 | ||
7943 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
7944 | init_task_group.se = (struct sched_entity **)ptr; | |
7945 | ptr += nr_cpu_ids * sizeof(void **); | |
7946 | ||
7947 | init_task_group.cfs_rq = (struct cfs_rq **)ptr; | |
7948 | ptr += nr_cpu_ids * sizeof(void **); | |
7949 | ||
7950 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
7951 | #ifdef CONFIG_RT_GROUP_SCHED | |
7952 | init_task_group.rt_se = (struct sched_rt_entity **)ptr; | |
7953 | ptr += nr_cpu_ids * sizeof(void **); | |
7954 | ||
7955 | init_task_group.rt_rq = (struct rt_rq **)ptr; | |
7956 | ptr += nr_cpu_ids * sizeof(void **); | |
7957 | ||
7958 | #endif /* CONFIG_RT_GROUP_SCHED */ | |
7959 | #ifdef CONFIG_CPUMASK_OFFSTACK | |
7960 | for_each_possible_cpu(i) { | |
7961 | per_cpu(load_balance_tmpmask, i) = (void *)ptr; | |
7962 | ptr += cpumask_size(); | |
7963 | } | |
7964 | #endif /* CONFIG_CPUMASK_OFFSTACK */ | |
7965 | } | |
7966 | ||
7967 | #ifdef CONFIG_SMP | |
7968 | init_defrootdomain(); | |
7969 | #endif | |
7970 | ||
7971 | init_rt_bandwidth(&def_rt_bandwidth, | |
7972 | global_rt_period(), global_rt_runtime()); | |
7973 | ||
7974 | #ifdef CONFIG_RT_GROUP_SCHED | |
7975 | init_rt_bandwidth(&init_task_group.rt_bandwidth, | |
7976 | global_rt_period(), global_rt_runtime()); | |
7977 | #endif /* CONFIG_RT_GROUP_SCHED */ | |
7978 | ||
7979 | #ifdef CONFIG_CGROUP_SCHED | |
7980 | list_add(&init_task_group.list, &task_groups); | |
7981 | INIT_LIST_HEAD(&init_task_group.children); | |
7982 | ||
7983 | #endif /* CONFIG_CGROUP_SCHED */ | |
7984 | ||
7985 | #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP | |
7986 | update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long), | |
7987 | __alignof__(unsigned long)); | |
7988 | #endif | |
7989 | for_each_possible_cpu(i) { | |
7990 | struct rq *rq; | |
7991 | ||
7992 | rq = cpu_rq(i); | |
7993 | raw_spin_lock_init(&rq->lock); | |
7994 | rq->nr_running = 0; | |
7995 | rq->calc_load_active = 0; | |
7996 | rq->calc_load_update = jiffies + LOAD_FREQ; | |
7997 | init_cfs_rq(&rq->cfs, rq); | |
7998 | init_rt_rq(&rq->rt, rq); | |
7999 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
8000 | init_task_group.shares = init_task_group_load; | |
8001 | INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); | |
8002 | #ifdef CONFIG_CGROUP_SCHED | |
8003 | /* | |
8004 | * How much cpu bandwidth does init_task_group get? | |
8005 | * | |
8006 | * In case of task-groups formed thr' the cgroup filesystem, it | |
8007 | * gets 100% of the cpu resources in the system. This overall | |
8008 | * system cpu resource is divided among the tasks of | |
8009 | * init_task_group and its child task-groups in a fair manner, | |
8010 | * based on each entity's (task or task-group's) weight | |
8011 | * (se->load.weight). | |
8012 | * | |
8013 | * In other words, if init_task_group has 10 tasks of weight | |
8014 | * 1024) and two child groups A0 and A1 (of weight 1024 each), | |
8015 | * then A0's share of the cpu resource is: | |
8016 | * | |
8017 | * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% | |
8018 | * | |
8019 | * We achieve this by letting init_task_group's tasks sit | |
8020 | * directly in rq->cfs (i.e init_task_group->se[] = NULL). | |
8021 | */ | |
8022 | init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL); | |
8023 | #endif | |
8024 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
8025 | ||
8026 | rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; | |
8027 | #ifdef CONFIG_RT_GROUP_SCHED | |
8028 | INIT_LIST_HEAD(&rq->leaf_rt_rq_list); | |
8029 | #ifdef CONFIG_CGROUP_SCHED | |
8030 | init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL); | |
8031 | #endif | |
8032 | #endif | |
8033 | ||
8034 | for (j = 0; j < CPU_LOAD_IDX_MAX; j++) | |
8035 | rq->cpu_load[j] = 0; | |
8036 | ||
8037 | rq->last_load_update_tick = jiffies; | |
8038 | ||
8039 | #ifdef CONFIG_SMP | |
8040 | rq->sd = NULL; | |
8041 | rq->rd = NULL; | |
8042 | rq->cpu_power = SCHED_LOAD_SCALE; | |
8043 | rq->post_schedule = 0; | |
8044 | rq->active_balance = 0; | |
8045 | rq->next_balance = jiffies; | |
8046 | rq->push_cpu = 0; | |
8047 | rq->cpu = i; | |
8048 | rq->online = 0; | |
8049 | rq->idle_stamp = 0; | |
8050 | rq->avg_idle = 2*sysctl_sched_migration_cost; | |
8051 | rq_attach_root(rq, &def_root_domain); | |
8052 | #ifdef CONFIG_NO_HZ | |
8053 | rq->nohz_balance_kick = 0; | |
8054 | init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i)); | |
8055 | #endif | |
8056 | #endif | |
8057 | init_rq_hrtick(rq); | |
8058 | atomic_set(&rq->nr_iowait, 0); | |
8059 | } | |
8060 | ||
8061 | set_load_weight(&init_task); | |
8062 | ||
8063 | #ifdef CONFIG_PREEMPT_NOTIFIERS | |
8064 | INIT_HLIST_HEAD(&init_task.preempt_notifiers); | |
8065 | #endif | |
8066 | ||
8067 | #ifdef CONFIG_SMP | |
8068 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | |
8069 | #endif | |
8070 | ||
8071 | #ifdef CONFIG_RT_MUTEXES | |
8072 | plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock); | |
8073 | #endif | |
8074 | ||
8075 | /* | |
8076 | * The boot idle thread does lazy MMU switching as well: | |
8077 | */ | |
8078 | atomic_inc(&init_mm.mm_count); | |
8079 | enter_lazy_tlb(&init_mm, current); | |
8080 | ||
8081 | /* | |
8082 | * Make us the idle thread. Technically, schedule() should not be | |
8083 | * called from this thread, however somewhere below it might be, | |
8084 | * but because we are the idle thread, we just pick up running again | |
8085 | * when this runqueue becomes "idle". | |
8086 | */ | |
8087 | init_idle(current, smp_processor_id()); | |
8088 | ||
8089 | calc_load_update = jiffies + LOAD_FREQ; | |
8090 | ||
8091 | /* | |
8092 | * During early bootup we pretend to be a normal task: | |
8093 | */ | |
8094 | current->sched_class = &fair_sched_class; | |
8095 | ||
8096 | /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */ | |
8097 | zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT); | |
8098 | #ifdef CONFIG_SMP | |
8099 | #ifdef CONFIG_NO_HZ | |
8100 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); | |
8101 | alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT); | |
8102 | atomic_set(&nohz.load_balancer, nr_cpu_ids); | |
8103 | atomic_set(&nohz.first_pick_cpu, nr_cpu_ids); | |
8104 | atomic_set(&nohz.second_pick_cpu, nr_cpu_ids); | |
8105 | #endif | |
8106 | /* May be allocated at isolcpus cmdline parse time */ | |
8107 | if (cpu_isolated_map == NULL) | |
8108 | zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); | |
8109 | #endif /* SMP */ | |
8110 | ||
8111 | perf_event_init(); | |
8112 | ||
8113 | scheduler_running = 1; | |
8114 | } | |
8115 | ||
8116 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP | |
8117 | static inline int preempt_count_equals(int preempt_offset) | |
8118 | { | |
8119 | int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); | |
8120 | ||
8121 | return (nested == PREEMPT_INATOMIC_BASE + preempt_offset); | |
8122 | } | |
8123 | ||
8124 | void __might_sleep(const char *file, int line, int preempt_offset) | |
8125 | { | |
8126 | #ifdef in_atomic | |
8127 | static unsigned long prev_jiffy; /* ratelimiting */ | |
8128 | ||
8129 | if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || | |
8130 | system_state != SYSTEM_RUNNING || oops_in_progress) | |
8131 | return; | |
8132 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) | |
8133 | return; | |
8134 | prev_jiffy = jiffies; | |
8135 | ||
8136 | printk(KERN_ERR | |
8137 | "BUG: sleeping function called from invalid context at %s:%d\n", | |
8138 | file, line); | |
8139 | printk(KERN_ERR | |
8140 | "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", | |
8141 | in_atomic(), irqs_disabled(), | |
8142 | current->pid, current->comm); | |
8143 | ||
8144 | debug_show_held_locks(current); | |
8145 | if (irqs_disabled()) | |
8146 | print_irqtrace_events(current); | |
8147 | dump_stack(); | |
8148 | #endif | |
8149 | } | |
8150 | EXPORT_SYMBOL(__might_sleep); | |
8151 | #endif | |
8152 | ||
8153 | #ifdef CONFIG_MAGIC_SYSRQ | |
8154 | static void normalize_task(struct rq *rq, struct task_struct *p) | |
8155 | { | |
8156 | int on_rq; | |
8157 | ||
8158 | on_rq = p->se.on_rq; | |
8159 | if (on_rq) | |
8160 | deactivate_task(rq, p, 0); | |
8161 | __setscheduler(rq, p, SCHED_NORMAL, 0); | |
8162 | if (on_rq) { | |
8163 | activate_task(rq, p, 0); | |
8164 | resched_task(rq->curr); | |
8165 | } | |
8166 | } | |
8167 | ||
8168 | void normalize_rt_tasks(void) | |
8169 | { | |
8170 | struct task_struct *g, *p; | |
8171 | unsigned long flags; | |
8172 | struct rq *rq; | |
8173 | ||
8174 | read_lock_irqsave(&tasklist_lock, flags); | |
8175 | do_each_thread(g, p) { | |
8176 | /* | |
8177 | * Only normalize user tasks: | |
8178 | */ | |
8179 | if (!p->mm) | |
8180 | continue; | |
8181 | ||
8182 | p->se.exec_start = 0; | |
8183 | #ifdef CONFIG_SCHEDSTATS | |
8184 | p->se.statistics.wait_start = 0; | |
8185 | p->se.statistics.sleep_start = 0; | |
8186 | p->se.statistics.block_start = 0; | |
8187 | #endif | |
8188 | ||
8189 | if (!rt_task(p)) { | |
8190 | /* | |
8191 | * Renice negative nice level userspace | |
8192 | * tasks back to 0: | |
8193 | */ | |
8194 | if (TASK_NICE(p) < 0 && p->mm) | |
8195 | set_user_nice(p, 0); | |
8196 | continue; | |
8197 | } | |
8198 | ||
8199 | raw_spin_lock(&p->pi_lock); | |
8200 | rq = __task_rq_lock(p); | |
8201 | ||
8202 | normalize_task(rq, p); | |
8203 | ||
8204 | __task_rq_unlock(rq); | |
8205 | raw_spin_unlock(&p->pi_lock); | |
8206 | } while_each_thread(g, p); | |
8207 | ||
8208 | read_unlock_irqrestore(&tasklist_lock, flags); | |
8209 | } | |
8210 | ||
8211 | #endif /* CONFIG_MAGIC_SYSRQ */ | |
8212 | ||
8213 | #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) | |
8214 | /* | |
8215 | * These functions are only useful for the IA64 MCA handling, or kdb. | |
8216 | * | |
8217 | * They can only be called when the whole system has been | |
8218 | * stopped - every CPU needs to be quiescent, and no scheduling | |
8219 | * activity can take place. Using them for anything else would | |
8220 | * be a serious bug, and as a result, they aren't even visible | |
8221 | * under any other configuration. | |
8222 | */ | |
8223 | ||
8224 | /** | |
8225 | * curr_task - return the current task for a given cpu. | |
8226 | * @cpu: the processor in question. | |
8227 | * | |
8228 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
8229 | */ | |
8230 | struct task_struct *curr_task(int cpu) | |
8231 | { | |
8232 | return cpu_curr(cpu); | |
8233 | } | |
8234 | ||
8235 | #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ | |
8236 | ||
8237 | #ifdef CONFIG_IA64 | |
8238 | /** | |
8239 | * set_curr_task - set the current task for a given cpu. | |
8240 | * @cpu: the processor in question. | |
8241 | * @p: the task pointer to set. | |
8242 | * | |
8243 | * Description: This function must only be used when non-maskable interrupts | |
8244 | * are serviced on a separate stack. It allows the architecture to switch the | |
8245 | * notion of the current task on a cpu in a non-blocking manner. This function | |
8246 | * must be called with all CPU's synchronized, and interrupts disabled, the | |
8247 | * and caller must save the original value of the current task (see | |
8248 | * curr_task() above) and restore that value before reenabling interrupts and | |
8249 | * re-starting the system. | |
8250 | * | |
8251 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
8252 | */ | |
8253 | void set_curr_task(int cpu, struct task_struct *p) | |
8254 | { | |
8255 | cpu_curr(cpu) = p; | |
8256 | } | |
8257 | ||
8258 | #endif | |
8259 | ||
8260 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
8261 | static void free_fair_sched_group(struct task_group *tg) | |
8262 | { | |
8263 | int i; | |
8264 | ||
8265 | for_each_possible_cpu(i) { | |
8266 | if (tg->cfs_rq) | |
8267 | kfree(tg->cfs_rq[i]); | |
8268 | if (tg->se) | |
8269 | kfree(tg->se[i]); | |
8270 | } | |
8271 | ||
8272 | kfree(tg->cfs_rq); | |
8273 | kfree(tg->se); | |
8274 | } | |
8275 | ||
8276 | static | |
8277 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
8278 | { | |
8279 | struct cfs_rq *cfs_rq; | |
8280 | struct sched_entity *se; | |
8281 | struct rq *rq; | |
8282 | int i; | |
8283 | ||
8284 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); | |
8285 | if (!tg->cfs_rq) | |
8286 | goto err; | |
8287 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); | |
8288 | if (!tg->se) | |
8289 | goto err; | |
8290 | ||
8291 | tg->shares = NICE_0_LOAD; | |
8292 | ||
8293 | for_each_possible_cpu(i) { | |
8294 | rq = cpu_rq(i); | |
8295 | ||
8296 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
8297 | GFP_KERNEL, cpu_to_node(i)); | |
8298 | if (!cfs_rq) | |
8299 | goto err; | |
8300 | ||
8301 | se = kzalloc_node(sizeof(struct sched_entity), | |
8302 | GFP_KERNEL, cpu_to_node(i)); | |
8303 | if (!se) | |
8304 | goto err_free_rq; | |
8305 | ||
8306 | init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]); | |
8307 | } | |
8308 | ||
8309 | return 1; | |
8310 | ||
8311 | err_free_rq: | |
8312 | kfree(cfs_rq); | |
8313 | err: | |
8314 | return 0; | |
8315 | } | |
8316 | ||
8317 | static inline void register_fair_sched_group(struct task_group *tg, int cpu) | |
8318 | { | |
8319 | list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list, | |
8320 | &cpu_rq(cpu)->leaf_cfs_rq_list); | |
8321 | } | |
8322 | ||
8323 | static inline void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
8324 | { | |
8325 | list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list); | |
8326 | } | |
8327 | #else /* !CONFG_FAIR_GROUP_SCHED */ | |
8328 | static inline void free_fair_sched_group(struct task_group *tg) | |
8329 | { | |
8330 | } | |
8331 | ||
8332 | static inline | |
8333 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
8334 | { | |
8335 | return 1; | |
8336 | } | |
8337 | ||
8338 | static inline void register_fair_sched_group(struct task_group *tg, int cpu) | |
8339 | { | |
8340 | } | |
8341 | ||
8342 | static inline void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
8343 | { | |
8344 | } | |
8345 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
8346 | ||
8347 | #ifdef CONFIG_RT_GROUP_SCHED | |
8348 | static void free_rt_sched_group(struct task_group *tg) | |
8349 | { | |
8350 | int i; | |
8351 | ||
8352 | destroy_rt_bandwidth(&tg->rt_bandwidth); | |
8353 | ||
8354 | for_each_possible_cpu(i) { | |
8355 | if (tg->rt_rq) | |
8356 | kfree(tg->rt_rq[i]); | |
8357 | if (tg->rt_se) | |
8358 | kfree(tg->rt_se[i]); | |
8359 | } | |
8360 | ||
8361 | kfree(tg->rt_rq); | |
8362 | kfree(tg->rt_se); | |
8363 | } | |
8364 | ||
8365 | static | |
8366 | int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) | |
8367 | { | |
8368 | struct rt_rq *rt_rq; | |
8369 | struct sched_rt_entity *rt_se; | |
8370 | struct rq *rq; | |
8371 | int i; | |
8372 | ||
8373 | tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL); | |
8374 | if (!tg->rt_rq) | |
8375 | goto err; | |
8376 | tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL); | |
8377 | if (!tg->rt_se) | |
8378 | goto err; | |
8379 | ||
8380 | init_rt_bandwidth(&tg->rt_bandwidth, | |
8381 | ktime_to_ns(def_rt_bandwidth.rt_period), 0); | |
8382 | ||
8383 | for_each_possible_cpu(i) { | |
8384 | rq = cpu_rq(i); | |
8385 | ||
8386 | rt_rq = kzalloc_node(sizeof(struct rt_rq), | |
8387 | GFP_KERNEL, cpu_to_node(i)); | |
8388 | if (!rt_rq) | |
8389 | goto err; | |
8390 | ||
8391 | rt_se = kzalloc_node(sizeof(struct sched_rt_entity), | |
8392 | GFP_KERNEL, cpu_to_node(i)); | |
8393 | if (!rt_se) | |
8394 | goto err_free_rq; | |
8395 | ||
8396 | init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]); | |
8397 | } | |
8398 | ||
8399 | return 1; | |
8400 | ||
8401 | err_free_rq: | |
8402 | kfree(rt_rq); | |
8403 | err: | |
8404 | return 0; | |
8405 | } | |
8406 | ||
8407 | static inline void register_rt_sched_group(struct task_group *tg, int cpu) | |
8408 | { | |
8409 | list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list, | |
8410 | &cpu_rq(cpu)->leaf_rt_rq_list); | |
8411 | } | |
8412 | ||
8413 | static inline void unregister_rt_sched_group(struct task_group *tg, int cpu) | |
8414 | { | |
8415 | list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list); | |
8416 | } | |
8417 | #else /* !CONFIG_RT_GROUP_SCHED */ | |
8418 | static inline void free_rt_sched_group(struct task_group *tg) | |
8419 | { | |
8420 | } | |
8421 | ||
8422 | static inline | |
8423 | int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) | |
8424 | { | |
8425 | return 1; | |
8426 | } | |
8427 | ||
8428 | static inline void register_rt_sched_group(struct task_group *tg, int cpu) | |
8429 | { | |
8430 | } | |
8431 | ||
8432 | static inline void unregister_rt_sched_group(struct task_group *tg, int cpu) | |
8433 | { | |
8434 | } | |
8435 | #endif /* CONFIG_RT_GROUP_SCHED */ | |
8436 | ||
8437 | #ifdef CONFIG_CGROUP_SCHED | |
8438 | static void free_sched_group(struct task_group *tg) | |
8439 | { | |
8440 | free_fair_sched_group(tg); | |
8441 | free_rt_sched_group(tg); | |
8442 | kfree(tg); | |
8443 | } | |
8444 | ||
8445 | /* allocate runqueue etc for a new task group */ | |
8446 | struct task_group *sched_create_group(struct task_group *parent) | |
8447 | { | |
8448 | struct task_group *tg; | |
8449 | unsigned long flags; | |
8450 | int i; | |
8451 | ||
8452 | tg = kzalloc(sizeof(*tg), GFP_KERNEL); | |
8453 | if (!tg) | |
8454 | return ERR_PTR(-ENOMEM); | |
8455 | ||
8456 | if (!alloc_fair_sched_group(tg, parent)) | |
8457 | goto err; | |
8458 | ||
8459 | if (!alloc_rt_sched_group(tg, parent)) | |
8460 | goto err; | |
8461 | ||
8462 | spin_lock_irqsave(&task_group_lock, flags); | |
8463 | for_each_possible_cpu(i) { | |
8464 | register_fair_sched_group(tg, i); | |
8465 | register_rt_sched_group(tg, i); | |
8466 | } | |
8467 | list_add_rcu(&tg->list, &task_groups); | |
8468 | ||
8469 | WARN_ON(!parent); /* root should already exist */ | |
8470 | ||
8471 | tg->parent = parent; | |
8472 | INIT_LIST_HEAD(&tg->children); | |
8473 | list_add_rcu(&tg->siblings, &parent->children); | |
8474 | spin_unlock_irqrestore(&task_group_lock, flags); | |
8475 | ||
8476 | return tg; | |
8477 | ||
8478 | err: | |
8479 | free_sched_group(tg); | |
8480 | return ERR_PTR(-ENOMEM); | |
8481 | } | |
8482 | ||
8483 | /* rcu callback to free various structures associated with a task group */ | |
8484 | static void free_sched_group_rcu(struct rcu_head *rhp) | |
8485 | { | |
8486 | /* now it should be safe to free those cfs_rqs */ | |
8487 | free_sched_group(container_of(rhp, struct task_group, rcu)); | |
8488 | } | |
8489 | ||
8490 | /* Destroy runqueue etc associated with a task group */ | |
8491 | void sched_destroy_group(struct task_group *tg) | |
8492 | { | |
8493 | unsigned long flags; | |
8494 | int i; | |
8495 | ||
8496 | spin_lock_irqsave(&task_group_lock, flags); | |
8497 | for_each_possible_cpu(i) { | |
8498 | unregister_fair_sched_group(tg, i); | |
8499 | unregister_rt_sched_group(tg, i); | |
8500 | } | |
8501 | list_del_rcu(&tg->list); | |
8502 | list_del_rcu(&tg->siblings); | |
8503 | spin_unlock_irqrestore(&task_group_lock, flags); | |
8504 | ||
8505 | /* wait for possible concurrent references to cfs_rqs complete */ | |
8506 | call_rcu(&tg->rcu, free_sched_group_rcu); | |
8507 | } | |
8508 | ||
8509 | /* change task's runqueue when it moves between groups. | |
8510 | * The caller of this function should have put the task in its new group | |
8511 | * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to | |
8512 | * reflect its new group. | |
8513 | */ | |
8514 | void sched_move_task(struct task_struct *tsk) | |
8515 | { | |
8516 | int on_rq, running; | |
8517 | unsigned long flags; | |
8518 | struct rq *rq; | |
8519 | ||
8520 | rq = task_rq_lock(tsk, &flags); | |
8521 | ||
8522 | running = task_current(rq, tsk); | |
8523 | on_rq = tsk->se.on_rq; | |
8524 | ||
8525 | if (on_rq) | |
8526 | dequeue_task(rq, tsk, 0); | |
8527 | if (unlikely(running)) | |
8528 | tsk->sched_class->put_prev_task(rq, tsk); | |
8529 | ||
8530 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
8531 | if (tsk->sched_class->task_move_group) | |
8532 | tsk->sched_class->task_move_group(tsk, on_rq); | |
8533 | else | |
8534 | #endif | |
8535 | set_task_rq(tsk, task_cpu(tsk)); | |
8536 | ||
8537 | if (unlikely(running)) | |
8538 | tsk->sched_class->set_curr_task(rq); | |
8539 | if (on_rq) | |
8540 | enqueue_task(rq, tsk, 0); | |
8541 | ||
8542 | task_rq_unlock(rq, &flags); | |
8543 | } | |
8544 | #endif /* CONFIG_CGROUP_SCHED */ | |
8545 | ||
8546 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
8547 | static void __set_se_shares(struct sched_entity *se, unsigned long shares) | |
8548 | { | |
8549 | struct cfs_rq *cfs_rq = se->cfs_rq; | |
8550 | int on_rq; | |
8551 | ||
8552 | on_rq = se->on_rq; | |
8553 | if (on_rq) | |
8554 | dequeue_entity(cfs_rq, se, 0); | |
8555 | ||
8556 | se->load.weight = shares; | |
8557 | se->load.inv_weight = 0; | |
8558 | ||
8559 | if (on_rq) | |
8560 | enqueue_entity(cfs_rq, se, 0); | |
8561 | } | |
8562 | ||
8563 | static void set_se_shares(struct sched_entity *se, unsigned long shares) | |
8564 | { | |
8565 | struct cfs_rq *cfs_rq = se->cfs_rq; | |
8566 | struct rq *rq = cfs_rq->rq; | |
8567 | unsigned long flags; | |
8568 | ||
8569 | raw_spin_lock_irqsave(&rq->lock, flags); | |
8570 | __set_se_shares(se, shares); | |
8571 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
8572 | } | |
8573 | ||
8574 | static DEFINE_MUTEX(shares_mutex); | |
8575 | ||
8576 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
8577 | { | |
8578 | int i; | |
8579 | unsigned long flags; | |
8580 | ||
8581 | /* | |
8582 | * We can't change the weight of the root cgroup. | |
8583 | */ | |
8584 | if (!tg->se[0]) | |
8585 | return -EINVAL; | |
8586 | ||
8587 | if (shares < MIN_SHARES) | |
8588 | shares = MIN_SHARES; | |
8589 | else if (shares > MAX_SHARES) | |
8590 | shares = MAX_SHARES; | |
8591 | ||
8592 | mutex_lock(&shares_mutex); | |
8593 | if (tg->shares == shares) | |
8594 | goto done; | |
8595 | ||
8596 | spin_lock_irqsave(&task_group_lock, flags); | |
8597 | for_each_possible_cpu(i) | |
8598 | unregister_fair_sched_group(tg, i); | |
8599 | list_del_rcu(&tg->siblings); | |
8600 | spin_unlock_irqrestore(&task_group_lock, flags); | |
8601 | ||
8602 | /* wait for any ongoing reference to this group to finish */ | |
8603 | synchronize_sched(); | |
8604 | ||
8605 | /* | |
8606 | * Now we are free to modify the group's share on each cpu | |
8607 | * w/o tripping rebalance_share or load_balance_fair. | |
8608 | */ | |
8609 | tg->shares = shares; | |
8610 | for_each_possible_cpu(i) { | |
8611 | /* | |
8612 | * force a rebalance | |
8613 | */ | |
8614 | cfs_rq_set_shares(tg->cfs_rq[i], 0); | |
8615 | set_se_shares(tg->se[i], shares); | |
8616 | } | |
8617 | ||
8618 | /* | |
8619 | * Enable load balance activity on this group, by inserting it back on | |
8620 | * each cpu's rq->leaf_cfs_rq_list. | |
8621 | */ | |
8622 | spin_lock_irqsave(&task_group_lock, flags); | |
8623 | for_each_possible_cpu(i) | |
8624 | register_fair_sched_group(tg, i); | |
8625 | list_add_rcu(&tg->siblings, &tg->parent->children); | |
8626 | spin_unlock_irqrestore(&task_group_lock, flags); | |
8627 | done: | |
8628 | mutex_unlock(&shares_mutex); | |
8629 | return 0; | |
8630 | } | |
8631 | ||
8632 | unsigned long sched_group_shares(struct task_group *tg) | |
8633 | { | |
8634 | return tg->shares; | |
8635 | } | |
8636 | #endif | |
8637 | ||
8638 | #ifdef CONFIG_RT_GROUP_SCHED | |
8639 | /* | |
8640 | * Ensure that the real time constraints are schedulable. | |
8641 | */ | |
8642 | static DEFINE_MUTEX(rt_constraints_mutex); | |
8643 | ||
8644 | static unsigned long to_ratio(u64 period, u64 runtime) | |
8645 | { | |
8646 | if (runtime == RUNTIME_INF) | |
8647 | return 1ULL << 20; | |
8648 | ||
8649 | return div64_u64(runtime << 20, period); | |
8650 | } | |
8651 | ||
8652 | /* Must be called with tasklist_lock held */ | |
8653 | static inline int tg_has_rt_tasks(struct task_group *tg) | |
8654 | { | |
8655 | struct task_struct *g, *p; | |
8656 | ||
8657 | do_each_thread(g, p) { | |
8658 | if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg) | |
8659 | return 1; | |
8660 | } while_each_thread(g, p); | |
8661 | ||
8662 | return 0; | |
8663 | } | |
8664 | ||
8665 | struct rt_schedulable_data { | |
8666 | struct task_group *tg; | |
8667 | u64 rt_period; | |
8668 | u64 rt_runtime; | |
8669 | }; | |
8670 | ||
8671 | static int tg_schedulable(struct task_group *tg, void *data) | |
8672 | { | |
8673 | struct rt_schedulable_data *d = data; | |
8674 | struct task_group *child; | |
8675 | unsigned long total, sum = 0; | |
8676 | u64 period, runtime; | |
8677 | ||
8678 | period = ktime_to_ns(tg->rt_bandwidth.rt_period); | |
8679 | runtime = tg->rt_bandwidth.rt_runtime; | |
8680 | ||
8681 | if (tg == d->tg) { | |
8682 | period = d->rt_period; | |
8683 | runtime = d->rt_runtime; | |
8684 | } | |
8685 | ||
8686 | /* | |
8687 | * Cannot have more runtime than the period. | |
8688 | */ | |
8689 | if (runtime > period && runtime != RUNTIME_INF) | |
8690 | return -EINVAL; | |
8691 | ||
8692 | /* | |
8693 | * Ensure we don't starve existing RT tasks. | |
8694 | */ | |
8695 | if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) | |
8696 | return -EBUSY; | |
8697 | ||
8698 | total = to_ratio(period, runtime); | |
8699 | ||
8700 | /* | |
8701 | * Nobody can have more than the global setting allows. | |
8702 | */ | |
8703 | if (total > to_ratio(global_rt_period(), global_rt_runtime())) | |
8704 | return -EINVAL; | |
8705 | ||
8706 | /* | |
8707 | * The sum of our children's runtime should not exceed our own. | |
8708 | */ | |
8709 | list_for_each_entry_rcu(child, &tg->children, siblings) { | |
8710 | period = ktime_to_ns(child->rt_bandwidth.rt_period); | |
8711 | runtime = child->rt_bandwidth.rt_runtime; | |
8712 | ||
8713 | if (child == d->tg) { | |
8714 | period = d->rt_period; | |
8715 | runtime = d->rt_runtime; | |
8716 | } | |
8717 | ||
8718 | sum += to_ratio(period, runtime); | |
8719 | } | |
8720 | ||
8721 | if (sum > total) | |
8722 | return -EINVAL; | |
8723 | ||
8724 | return 0; | |
8725 | } | |
8726 | ||
8727 | static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) | |
8728 | { | |
8729 | struct rt_schedulable_data data = { | |
8730 | .tg = tg, | |
8731 | .rt_period = period, | |
8732 | .rt_runtime = runtime, | |
8733 | }; | |
8734 | ||
8735 | return walk_tg_tree(tg_schedulable, tg_nop, &data); | |
8736 | } | |
8737 | ||
8738 | static int tg_set_bandwidth(struct task_group *tg, | |
8739 | u64 rt_period, u64 rt_runtime) | |
8740 | { | |
8741 | int i, err = 0; | |
8742 | ||
8743 | mutex_lock(&rt_constraints_mutex); | |
8744 | read_lock(&tasklist_lock); | |
8745 | err = __rt_schedulable(tg, rt_period, rt_runtime); | |
8746 | if (err) | |
8747 | goto unlock; | |
8748 | ||
8749 | raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); | |
8750 | tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); | |
8751 | tg->rt_bandwidth.rt_runtime = rt_runtime; | |
8752 | ||
8753 | for_each_possible_cpu(i) { | |
8754 | struct rt_rq *rt_rq = tg->rt_rq[i]; | |
8755 | ||
8756 | raw_spin_lock(&rt_rq->rt_runtime_lock); | |
8757 | rt_rq->rt_runtime = rt_runtime; | |
8758 | raw_spin_unlock(&rt_rq->rt_runtime_lock); | |
8759 | } | |
8760 | raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); | |
8761 | unlock: | |
8762 | read_unlock(&tasklist_lock); | |
8763 | mutex_unlock(&rt_constraints_mutex); | |
8764 | ||
8765 | return err; | |
8766 | } | |
8767 | ||
8768 | int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) | |
8769 | { | |
8770 | u64 rt_runtime, rt_period; | |
8771 | ||
8772 | rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); | |
8773 | rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; | |
8774 | if (rt_runtime_us < 0) | |
8775 | rt_runtime = RUNTIME_INF; | |
8776 | ||
8777 | return tg_set_bandwidth(tg, rt_period, rt_runtime); | |
8778 | } | |
8779 | ||
8780 | long sched_group_rt_runtime(struct task_group *tg) | |
8781 | { | |
8782 | u64 rt_runtime_us; | |
8783 | ||
8784 | if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) | |
8785 | return -1; | |
8786 | ||
8787 | rt_runtime_us = tg->rt_bandwidth.rt_runtime; | |
8788 | do_div(rt_runtime_us, NSEC_PER_USEC); | |
8789 | return rt_runtime_us; | |
8790 | } | |
8791 | ||
8792 | int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) | |
8793 | { | |
8794 | u64 rt_runtime, rt_period; | |
8795 | ||
8796 | rt_period = (u64)rt_period_us * NSEC_PER_USEC; | |
8797 | rt_runtime = tg->rt_bandwidth.rt_runtime; | |
8798 | ||
8799 | if (rt_period == 0) | |
8800 | return -EINVAL; | |
8801 | ||
8802 | return tg_set_bandwidth(tg, rt_period, rt_runtime); | |
8803 | } | |
8804 | ||
8805 | long sched_group_rt_period(struct task_group *tg) | |
8806 | { | |
8807 | u64 rt_period_us; | |
8808 | ||
8809 | rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); | |
8810 | do_div(rt_period_us, NSEC_PER_USEC); | |
8811 | return rt_period_us; | |
8812 | } | |
8813 | ||
8814 | static int sched_rt_global_constraints(void) | |
8815 | { | |
8816 | u64 runtime, period; | |
8817 | int ret = 0; | |
8818 | ||
8819 | if (sysctl_sched_rt_period <= 0) | |
8820 | return -EINVAL; | |
8821 | ||
8822 | runtime = global_rt_runtime(); | |
8823 | period = global_rt_period(); | |
8824 | ||
8825 | /* | |
8826 | * Sanity check on the sysctl variables. | |
8827 | */ | |
8828 | if (runtime > period && runtime != RUNTIME_INF) | |
8829 | return -EINVAL; | |
8830 | ||
8831 | mutex_lock(&rt_constraints_mutex); | |
8832 | read_lock(&tasklist_lock); | |
8833 | ret = __rt_schedulable(NULL, 0, 0); | |
8834 | read_unlock(&tasklist_lock); | |
8835 | mutex_unlock(&rt_constraints_mutex); | |
8836 | ||
8837 | return ret; | |
8838 | } | |
8839 | ||
8840 | int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) | |
8841 | { | |
8842 | /* Don't accept realtime tasks when there is no way for them to run */ | |
8843 | if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) | |
8844 | return 0; | |
8845 | ||
8846 | return 1; | |
8847 | } | |
8848 | ||
8849 | #else /* !CONFIG_RT_GROUP_SCHED */ | |
8850 | static int sched_rt_global_constraints(void) | |
8851 | { | |
8852 | unsigned long flags; | |
8853 | int i; | |
8854 | ||
8855 | if (sysctl_sched_rt_period <= 0) | |
8856 | return -EINVAL; | |
8857 | ||
8858 | /* | |
8859 | * There's always some RT tasks in the root group | |
8860 | * -- migration, kstopmachine etc.. | |
8861 | */ | |
8862 | if (sysctl_sched_rt_runtime == 0) | |
8863 | return -EBUSY; | |
8864 | ||
8865 | raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); | |
8866 | for_each_possible_cpu(i) { | |
8867 | struct rt_rq *rt_rq = &cpu_rq(i)->rt; | |
8868 | ||
8869 | raw_spin_lock(&rt_rq->rt_runtime_lock); | |
8870 | rt_rq->rt_runtime = global_rt_runtime(); | |
8871 | raw_spin_unlock(&rt_rq->rt_runtime_lock); | |
8872 | } | |
8873 | raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); | |
8874 | ||
8875 | return 0; | |
8876 | } | |
8877 | #endif /* CONFIG_RT_GROUP_SCHED */ | |
8878 | ||
8879 | int sched_rt_handler(struct ctl_table *table, int write, | |
8880 | void __user *buffer, size_t *lenp, | |
8881 | loff_t *ppos) | |
8882 | { | |
8883 | int ret; | |
8884 | int old_period, old_runtime; | |
8885 | static DEFINE_MUTEX(mutex); | |
8886 | ||
8887 | mutex_lock(&mutex); | |
8888 | old_period = sysctl_sched_rt_period; | |
8889 | old_runtime = sysctl_sched_rt_runtime; | |
8890 | ||
8891 | ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
8892 | ||
8893 | if (!ret && write) { | |
8894 | ret = sched_rt_global_constraints(); | |
8895 | if (ret) { | |
8896 | sysctl_sched_rt_period = old_period; | |
8897 | sysctl_sched_rt_runtime = old_runtime; | |
8898 | } else { | |
8899 | def_rt_bandwidth.rt_runtime = global_rt_runtime(); | |
8900 | def_rt_bandwidth.rt_period = | |
8901 | ns_to_ktime(global_rt_period()); | |
8902 | } | |
8903 | } | |
8904 | mutex_unlock(&mutex); | |
8905 | ||
8906 | return ret; | |
8907 | } | |
8908 | ||
8909 | #ifdef CONFIG_CGROUP_SCHED | |
8910 | ||
8911 | /* return corresponding task_group object of a cgroup */ | |
8912 | static inline struct task_group *cgroup_tg(struct cgroup *cgrp) | |
8913 | { | |
8914 | return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id), | |
8915 | struct task_group, css); | |
8916 | } | |
8917 | ||
8918 | static struct cgroup_subsys_state * | |
8919 | cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp) | |
8920 | { | |
8921 | struct task_group *tg, *parent; | |
8922 | ||
8923 | if (!cgrp->parent) { | |
8924 | /* This is early initialization for the top cgroup */ | |
8925 | return &init_task_group.css; | |
8926 | } | |
8927 | ||
8928 | parent = cgroup_tg(cgrp->parent); | |
8929 | tg = sched_create_group(parent); | |
8930 | if (IS_ERR(tg)) | |
8931 | return ERR_PTR(-ENOMEM); | |
8932 | ||
8933 | return &tg->css; | |
8934 | } | |
8935 | ||
8936 | static void | |
8937 | cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp) | |
8938 | { | |
8939 | struct task_group *tg = cgroup_tg(cgrp); | |
8940 | ||
8941 | sched_destroy_group(tg); | |
8942 | } | |
8943 | ||
8944 | static int | |
8945 | cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk) | |
8946 | { | |
8947 | #ifdef CONFIG_RT_GROUP_SCHED | |
8948 | if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk)) | |
8949 | return -EINVAL; | |
8950 | #else | |
8951 | /* We don't support RT-tasks being in separate groups */ | |
8952 | if (tsk->sched_class != &fair_sched_class) | |
8953 | return -EINVAL; | |
8954 | #endif | |
8955 | return 0; | |
8956 | } | |
8957 | ||
8958 | static int | |
8959 | cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp, | |
8960 | struct task_struct *tsk, bool threadgroup) | |
8961 | { | |
8962 | int retval = cpu_cgroup_can_attach_task(cgrp, tsk); | |
8963 | if (retval) | |
8964 | return retval; | |
8965 | if (threadgroup) { | |
8966 | struct task_struct *c; | |
8967 | rcu_read_lock(); | |
8968 | list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) { | |
8969 | retval = cpu_cgroup_can_attach_task(cgrp, c); | |
8970 | if (retval) { | |
8971 | rcu_read_unlock(); | |
8972 | return retval; | |
8973 | } | |
8974 | } | |
8975 | rcu_read_unlock(); | |
8976 | } | |
8977 | return 0; | |
8978 | } | |
8979 | ||
8980 | static void | |
8981 | cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp, | |
8982 | struct cgroup *old_cont, struct task_struct *tsk, | |
8983 | bool threadgroup) | |
8984 | { | |
8985 | sched_move_task(tsk); | |
8986 | if (threadgroup) { | |
8987 | struct task_struct *c; | |
8988 | rcu_read_lock(); | |
8989 | list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) { | |
8990 | sched_move_task(c); | |
8991 | } | |
8992 | rcu_read_unlock(); | |
8993 | } | |
8994 | } | |
8995 | ||
8996 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
8997 | static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype, | |
8998 | u64 shareval) | |
8999 | { | |
9000 | return sched_group_set_shares(cgroup_tg(cgrp), shareval); | |
9001 | } | |
9002 | ||
9003 | static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft) | |
9004 | { | |
9005 | struct task_group *tg = cgroup_tg(cgrp); | |
9006 | ||
9007 | return (u64) tg->shares; | |
9008 | } | |
9009 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
9010 | ||
9011 | #ifdef CONFIG_RT_GROUP_SCHED | |
9012 | static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft, | |
9013 | s64 val) | |
9014 | { | |
9015 | return sched_group_set_rt_runtime(cgroup_tg(cgrp), val); | |
9016 | } | |
9017 | ||
9018 | static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft) | |
9019 | { | |
9020 | return sched_group_rt_runtime(cgroup_tg(cgrp)); | |
9021 | } | |
9022 | ||
9023 | static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype, | |
9024 | u64 rt_period_us) | |
9025 | { | |
9026 | return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us); | |
9027 | } | |
9028 | ||
9029 | static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft) | |
9030 | { | |
9031 | return sched_group_rt_period(cgroup_tg(cgrp)); | |
9032 | } | |
9033 | #endif /* CONFIG_RT_GROUP_SCHED */ | |
9034 | ||
9035 | static struct cftype cpu_files[] = { | |
9036 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
9037 | { | |
9038 | .name = "shares", | |
9039 | .read_u64 = cpu_shares_read_u64, | |
9040 | .write_u64 = cpu_shares_write_u64, | |
9041 | }, | |
9042 | #endif | |
9043 | #ifdef CONFIG_RT_GROUP_SCHED | |
9044 | { | |
9045 | .name = "rt_runtime_us", | |
9046 | .read_s64 = cpu_rt_runtime_read, | |
9047 | .write_s64 = cpu_rt_runtime_write, | |
9048 | }, | |
9049 | { | |
9050 | .name = "rt_period_us", | |
9051 | .read_u64 = cpu_rt_period_read_uint, | |
9052 | .write_u64 = cpu_rt_period_write_uint, | |
9053 | }, | |
9054 | #endif | |
9055 | }; | |
9056 | ||
9057 | static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont) | |
9058 | { | |
9059 | return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files)); | |
9060 | } | |
9061 | ||
9062 | struct cgroup_subsys cpu_cgroup_subsys = { | |
9063 | .name = "cpu", | |
9064 | .create = cpu_cgroup_create, | |
9065 | .destroy = cpu_cgroup_destroy, | |
9066 | .can_attach = cpu_cgroup_can_attach, | |
9067 | .attach = cpu_cgroup_attach, | |
9068 | .populate = cpu_cgroup_populate, | |
9069 | .subsys_id = cpu_cgroup_subsys_id, | |
9070 | .early_init = 1, | |
9071 | }; | |
9072 | ||
9073 | #endif /* CONFIG_CGROUP_SCHED */ | |
9074 | ||
9075 | #ifdef CONFIG_CGROUP_CPUACCT | |
9076 | ||
9077 | /* | |
9078 | * CPU accounting code for task groups. | |
9079 | * | |
9080 | * Based on the work by Paul Menage (menage@google.com) and Balbir Singh | |
9081 | * (balbir@in.ibm.com). | |
9082 | */ | |
9083 | ||
9084 | /* track cpu usage of a group of tasks and its child groups */ | |
9085 | struct cpuacct { | |
9086 | struct cgroup_subsys_state css; | |
9087 | /* cpuusage holds pointer to a u64-type object on every cpu */ | |
9088 | u64 __percpu *cpuusage; | |
9089 | struct percpu_counter cpustat[CPUACCT_STAT_NSTATS]; | |
9090 | struct cpuacct *parent; | |
9091 | }; | |
9092 | ||
9093 | struct cgroup_subsys cpuacct_subsys; | |
9094 | ||
9095 | /* return cpu accounting group corresponding to this container */ | |
9096 | static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp) | |
9097 | { | |
9098 | return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id), | |
9099 | struct cpuacct, css); | |
9100 | } | |
9101 | ||
9102 | /* return cpu accounting group to which this task belongs */ | |
9103 | static inline struct cpuacct *task_ca(struct task_struct *tsk) | |
9104 | { | |
9105 | return container_of(task_subsys_state(tsk, cpuacct_subsys_id), | |
9106 | struct cpuacct, css); | |
9107 | } | |
9108 | ||
9109 | /* create a new cpu accounting group */ | |
9110 | static struct cgroup_subsys_state *cpuacct_create( | |
9111 | struct cgroup_subsys *ss, struct cgroup *cgrp) | |
9112 | { | |
9113 | struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL); | |
9114 | int i; | |
9115 | ||
9116 | if (!ca) | |
9117 | goto out; | |
9118 | ||
9119 | ca->cpuusage = alloc_percpu(u64); | |
9120 | if (!ca->cpuusage) | |
9121 | goto out_free_ca; | |
9122 | ||
9123 | for (i = 0; i < CPUACCT_STAT_NSTATS; i++) | |
9124 | if (percpu_counter_init(&ca->cpustat[i], 0)) | |
9125 | goto out_free_counters; | |
9126 | ||
9127 | if (cgrp->parent) | |
9128 | ca->parent = cgroup_ca(cgrp->parent); | |
9129 | ||
9130 | return &ca->css; | |
9131 | ||
9132 | out_free_counters: | |
9133 | while (--i >= 0) | |
9134 | percpu_counter_destroy(&ca->cpustat[i]); | |
9135 | free_percpu(ca->cpuusage); | |
9136 | out_free_ca: | |
9137 | kfree(ca); | |
9138 | out: | |
9139 | return ERR_PTR(-ENOMEM); | |
9140 | } | |
9141 | ||
9142 | /* destroy an existing cpu accounting group */ | |
9143 | static void | |
9144 | cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp) | |
9145 | { | |
9146 | struct cpuacct *ca = cgroup_ca(cgrp); | |
9147 | int i; | |
9148 | ||
9149 | for (i = 0; i < CPUACCT_STAT_NSTATS; i++) | |
9150 | percpu_counter_destroy(&ca->cpustat[i]); | |
9151 | free_percpu(ca->cpuusage); | |
9152 | kfree(ca); | |
9153 | } | |
9154 | ||
9155 | static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu) | |
9156 | { | |
9157 | u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); | |
9158 | u64 data; | |
9159 | ||
9160 | #ifndef CONFIG_64BIT | |
9161 | /* | |
9162 | * Take rq->lock to make 64-bit read safe on 32-bit platforms. | |
9163 | */ | |
9164 | raw_spin_lock_irq(&cpu_rq(cpu)->lock); | |
9165 | data = *cpuusage; | |
9166 | raw_spin_unlock_irq(&cpu_rq(cpu)->lock); | |
9167 | #else | |
9168 | data = *cpuusage; | |
9169 | #endif | |
9170 | ||
9171 | return data; | |
9172 | } | |
9173 | ||
9174 | static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val) | |
9175 | { | |
9176 | u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); | |
9177 | ||
9178 | #ifndef CONFIG_64BIT | |
9179 | /* | |
9180 | * Take rq->lock to make 64-bit write safe on 32-bit platforms. | |
9181 | */ | |
9182 | raw_spin_lock_irq(&cpu_rq(cpu)->lock); | |
9183 | *cpuusage = val; | |
9184 | raw_spin_unlock_irq(&cpu_rq(cpu)->lock); | |
9185 | #else | |
9186 | *cpuusage = val; | |
9187 | #endif | |
9188 | } | |
9189 | ||
9190 | /* return total cpu usage (in nanoseconds) of a group */ | |
9191 | static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft) | |
9192 | { | |
9193 | struct cpuacct *ca = cgroup_ca(cgrp); | |
9194 | u64 totalcpuusage = 0; | |
9195 | int i; | |
9196 | ||
9197 | for_each_present_cpu(i) | |
9198 | totalcpuusage += cpuacct_cpuusage_read(ca, i); | |
9199 | ||
9200 | return totalcpuusage; | |
9201 | } | |
9202 | ||
9203 | static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype, | |
9204 | u64 reset) | |
9205 | { | |
9206 | struct cpuacct *ca = cgroup_ca(cgrp); | |
9207 | int err = 0; | |
9208 | int i; | |
9209 | ||
9210 | if (reset) { | |
9211 | err = -EINVAL; | |
9212 | goto out; | |
9213 | } | |
9214 | ||
9215 | for_each_present_cpu(i) | |
9216 | cpuacct_cpuusage_write(ca, i, 0); | |
9217 | ||
9218 | out: | |
9219 | return err; | |
9220 | } | |
9221 | ||
9222 | static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft, | |
9223 | struct seq_file *m) | |
9224 | { | |
9225 | struct cpuacct *ca = cgroup_ca(cgroup); | |
9226 | u64 percpu; | |
9227 | int i; | |
9228 | ||
9229 | for_each_present_cpu(i) { | |
9230 | percpu = cpuacct_cpuusage_read(ca, i); | |
9231 | seq_printf(m, "%llu ", (unsigned long long) percpu); | |
9232 | } | |
9233 | seq_printf(m, "\n"); | |
9234 | return 0; | |
9235 | } | |
9236 | ||
9237 | static const char *cpuacct_stat_desc[] = { | |
9238 | [CPUACCT_STAT_USER] = "user", | |
9239 | [CPUACCT_STAT_SYSTEM] = "system", | |
9240 | }; | |
9241 | ||
9242 | static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft, | |
9243 | struct cgroup_map_cb *cb) | |
9244 | { | |
9245 | struct cpuacct *ca = cgroup_ca(cgrp); | |
9246 | int i; | |
9247 | ||
9248 | for (i = 0; i < CPUACCT_STAT_NSTATS; i++) { | |
9249 | s64 val = percpu_counter_read(&ca->cpustat[i]); | |
9250 | val = cputime64_to_clock_t(val); | |
9251 | cb->fill(cb, cpuacct_stat_desc[i], val); | |
9252 | } | |
9253 | return 0; | |
9254 | } | |
9255 | ||
9256 | static struct cftype files[] = { | |
9257 | { | |
9258 | .name = "usage", | |
9259 | .read_u64 = cpuusage_read, | |
9260 | .write_u64 = cpuusage_write, | |
9261 | }, | |
9262 | { | |
9263 | .name = "usage_percpu", | |
9264 | .read_seq_string = cpuacct_percpu_seq_read, | |
9265 | }, | |
9266 | { | |
9267 | .name = "stat", | |
9268 | .read_map = cpuacct_stats_show, | |
9269 | }, | |
9270 | }; | |
9271 | ||
9272 | static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp) | |
9273 | { | |
9274 | return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files)); | |
9275 | } | |
9276 | ||
9277 | /* | |
9278 | * charge this task's execution time to its accounting group. | |
9279 | * | |
9280 | * called with rq->lock held. | |
9281 | */ | |
9282 | static void cpuacct_charge(struct task_struct *tsk, u64 cputime) | |
9283 | { | |
9284 | struct cpuacct *ca; | |
9285 | int cpu; | |
9286 | ||
9287 | if (unlikely(!cpuacct_subsys.active)) | |
9288 | return; | |
9289 | ||
9290 | cpu = task_cpu(tsk); | |
9291 | ||
9292 | rcu_read_lock(); | |
9293 | ||
9294 | ca = task_ca(tsk); | |
9295 | ||
9296 | for (; ca; ca = ca->parent) { | |
9297 | u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); | |
9298 | *cpuusage += cputime; | |
9299 | } | |
9300 | ||
9301 | rcu_read_unlock(); | |
9302 | } | |
9303 | ||
9304 | /* | |
9305 | * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large | |
9306 | * in cputime_t units. As a result, cpuacct_update_stats calls | |
9307 | * percpu_counter_add with values large enough to always overflow the | |
9308 | * per cpu batch limit causing bad SMP scalability. | |
9309 | * | |
9310 | * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we | |
9311 | * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled | |
9312 | * and enabled. We cap it at INT_MAX which is the largest allowed batch value. | |
9313 | */ | |
9314 | #ifdef CONFIG_SMP | |
9315 | #define CPUACCT_BATCH \ | |
9316 | min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX) | |
9317 | #else | |
9318 | #define CPUACCT_BATCH 0 | |
9319 | #endif | |
9320 | ||
9321 | /* | |
9322 | * Charge the system/user time to the task's accounting group. | |
9323 | */ | |
9324 | static void cpuacct_update_stats(struct task_struct *tsk, | |
9325 | enum cpuacct_stat_index idx, cputime_t val) | |
9326 | { | |
9327 | struct cpuacct *ca; | |
9328 | int batch = CPUACCT_BATCH; | |
9329 | ||
9330 | if (unlikely(!cpuacct_subsys.active)) | |
9331 | return; | |
9332 | ||
9333 | rcu_read_lock(); | |
9334 | ca = task_ca(tsk); | |
9335 | ||
9336 | do { | |
9337 | __percpu_counter_add(&ca->cpustat[idx], val, batch); | |
9338 | ca = ca->parent; | |
9339 | } while (ca); | |
9340 | rcu_read_unlock(); | |
9341 | } | |
9342 | ||
9343 | struct cgroup_subsys cpuacct_subsys = { | |
9344 | .name = "cpuacct", | |
9345 | .create = cpuacct_create, | |
9346 | .destroy = cpuacct_destroy, | |
9347 | .populate = cpuacct_populate, | |
9348 | .subsys_id = cpuacct_subsys_id, | |
9349 | }; | |
9350 | #endif /* CONFIG_CGROUP_CPUACCT */ | |
9351 | ||
9352 | #ifndef CONFIG_SMP | |
9353 | ||
9354 | void synchronize_sched_expedited(void) | |
9355 | { | |
9356 | barrier(); | |
9357 | } | |
9358 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited); | |
9359 | ||
9360 | #else /* #ifndef CONFIG_SMP */ | |
9361 | ||
9362 | static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0); | |
9363 | ||
9364 | static int synchronize_sched_expedited_cpu_stop(void *data) | |
9365 | { | |
9366 | /* | |
9367 | * There must be a full memory barrier on each affected CPU | |
9368 | * between the time that try_stop_cpus() is called and the | |
9369 | * time that it returns. | |
9370 | * | |
9371 | * In the current initial implementation of cpu_stop, the | |
9372 | * above condition is already met when the control reaches | |
9373 | * this point and the following smp_mb() is not strictly | |
9374 | * necessary. Do smp_mb() anyway for documentation and | |
9375 | * robustness against future implementation changes. | |
9376 | */ | |
9377 | smp_mb(); /* See above comment block. */ | |
9378 | return 0; | |
9379 | } | |
9380 | ||
9381 | /* | |
9382 | * Wait for an rcu-sched grace period to elapse, but use "big hammer" | |
9383 | * approach to force grace period to end quickly. This consumes | |
9384 | * significant time on all CPUs, and is thus not recommended for | |
9385 | * any sort of common-case code. | |
9386 | * | |
9387 | * Note that it is illegal to call this function while holding any | |
9388 | * lock that is acquired by a CPU-hotplug notifier. Failing to | |
9389 | * observe this restriction will result in deadlock. | |
9390 | */ | |
9391 | void synchronize_sched_expedited(void) | |
9392 | { | |
9393 | int snap, trycount = 0; | |
9394 | ||
9395 | smp_mb(); /* ensure prior mod happens before capturing snap. */ | |
9396 | snap = atomic_read(&synchronize_sched_expedited_count) + 1; | |
9397 | get_online_cpus(); | |
9398 | while (try_stop_cpus(cpu_online_mask, | |
9399 | synchronize_sched_expedited_cpu_stop, | |
9400 | NULL) == -EAGAIN) { | |
9401 | put_online_cpus(); | |
9402 | if (trycount++ < 10) | |
9403 | udelay(trycount * num_online_cpus()); | |
9404 | else { | |
9405 | synchronize_sched(); | |
9406 | return; | |
9407 | } | |
9408 | if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) { | |
9409 | smp_mb(); /* ensure test happens before caller kfree */ | |
9410 | return; | |
9411 | } | |
9412 | get_online_cpus(); | |
9413 | } | |
9414 | atomic_inc(&synchronize_sched_expedited_count); | |
9415 | smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */ | |
9416 | put_online_cpus(); | |
9417 | } | |
9418 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited); | |
9419 | ||
9420 | #endif /* #else #ifndef CONFIG_SMP */ |