Root/
Source at commit be977234bfb4a6dca8a39e7c52165e4cd536ad71 created 12 years 9 months ago. By Lars-Peter Clausen, jz4740: Fix compile error | |
---|---|
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 <asm/mmu_context.h> |
36 | #include <linux/interrupt.h> |
37 | #include <linux/capability.h> |
38 | #include <linux/completion.h> |
39 | #include <linux/kernel_stat.h> |
40 | #include <linux/debug_locks.h> |
41 | #include <linux/perf_event.h> |
42 | #include <linux/security.h> |
43 | #include <linux/notifier.h> |
44 | #include <linux/profile.h> |
45 | #include <linux/freezer.h> |
46 | #include <linux/vmalloc.h> |
47 | #include <linux/blkdev.h> |
48 | #include <linux/delay.h> |
49 | #include <linux/pid_namespace.h> |
50 | #include <linux/smp.h> |
51 | #include <linux/threads.h> |
52 | #include <linux/timer.h> |
53 | #include <linux/rcupdate.h> |
54 | #include <linux/cpu.h> |
55 | #include <linux/cpuset.h> |
56 | #include <linux/percpu.h> |
57 | #include <linux/proc_fs.h> |
58 | #include <linux/seq_file.h> |
59 | #include <linux/stop_machine.h> |
60 | #include <linux/sysctl.h> |
61 | #include <linux/syscalls.h> |
62 | #include <linux/times.h> |
63 | #include <linux/tsacct_kern.h> |
64 | #include <linux/kprobes.h> |
65 | #include <linux/delayacct.h> |
66 | #include <linux/unistd.h> |
67 | #include <linux/pagemap.h> |
68 | #include <linux/hrtimer.h> |
69 | #include <linux/tick.h> |
70 | #include <linux/debugfs.h> |
71 | #include <linux/ctype.h> |
72 | #include <linux/ftrace.h> |
73 | #include <linux/slab.h> |
74 | |
75 | #include <asm/tlb.h> |
76 | #include <asm/irq_regs.h> |
77 | #include <asm/mutex.h> |
78 | |
79 | #include "sched_cpupri.h" |
80 | #include "workqueue_sched.h" |
81 | #include "sched_autogroup.h" |
82 | |
83 | #define CREATE_TRACE_POINTS |
84 | #include <trace/events/sched.h> |
85 | |
86 | /* |
87 | * Convert user-nice values [ -20 ... 0 ... 19 ] |
88 | * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ], |
89 | * and back. |
90 | */ |
91 | #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20) |
92 | #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20) |
93 | #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio) |
94 | |
95 | /* |
96 | * 'User priority' is the nice value converted to something we |
97 | * can work with better when scaling various scheduler parameters, |
98 | * it's a [ 0 ... 39 ] range. |
99 | */ |
100 | #define USER_PRIO(p) ((p)-MAX_RT_PRIO) |
101 | #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio) |
102 | #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO)) |
103 | |
104 | /* |
105 | * Helpers for converting nanosecond timing to jiffy resolution |
106 | */ |
107 | #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ)) |
108 | |
109 | #define NICE_0_LOAD SCHED_LOAD_SCALE |
110 | #define NICE_0_SHIFT SCHED_LOAD_SHIFT |
111 | |
112 | /* |
113 | * These are the 'tuning knobs' of the scheduler: |
114 | * |
115 | * default timeslice is 100 msecs (used only for SCHED_RR tasks). |
116 | * Timeslices get refilled after they expire. |
117 | */ |
118 | #define DEF_TIMESLICE (100 * HZ / 1000) |
119 | |
120 | /* |
121 | * single value that denotes runtime == period, ie unlimited time. |
122 | */ |
123 | #define RUNTIME_INF ((u64)~0ULL) |
124 | |
125 | static inline int rt_policy(int policy) |
126 | { |
127 | if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR)) |
128 | return 1; |
129 | return 0; |
130 | } |
131 | |
132 | static inline int task_has_rt_policy(struct task_struct *p) |
133 | { |
134 | return rt_policy(p->policy); |
135 | } |
136 | |
137 | /* |
138 | * This is the priority-queue data structure of the RT scheduling class: |
139 | */ |
140 | struct rt_prio_array { |
141 | DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */ |
142 | struct list_head queue[MAX_RT_PRIO]; |
143 | }; |
144 | |
145 | struct rt_bandwidth { |
146 | /* nests inside the rq lock: */ |
147 | raw_spinlock_t rt_runtime_lock; |
148 | ktime_t rt_period; |
149 | u64 rt_runtime; |
150 | struct hrtimer rt_period_timer; |
151 | }; |
152 | |
153 | static struct rt_bandwidth def_rt_bandwidth; |
154 | |
155 | static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun); |
156 | |
157 | static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer) |
158 | { |
159 | struct rt_bandwidth *rt_b = |
160 | container_of(timer, struct rt_bandwidth, rt_period_timer); |
161 | ktime_t now; |
162 | int overrun; |
163 | int idle = 0; |
164 | |
165 | for (;;) { |
166 | now = hrtimer_cb_get_time(timer); |
167 | overrun = hrtimer_forward(timer, now, rt_b->rt_period); |
168 | |
169 | if (!overrun) |
170 | break; |
171 | |
172 | idle = do_sched_rt_period_timer(rt_b, overrun); |
173 | } |
174 | |
175 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; |
176 | } |
177 | |
178 | static |
179 | void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime) |
180 | { |
181 | rt_b->rt_period = ns_to_ktime(period); |
182 | rt_b->rt_runtime = runtime; |
183 | |
184 | raw_spin_lock_init(&rt_b->rt_runtime_lock); |
185 | |
186 | hrtimer_init(&rt_b->rt_period_timer, |
187 | CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
188 | rt_b->rt_period_timer.function = sched_rt_period_timer; |
189 | } |
190 | |
191 | static inline int rt_bandwidth_enabled(void) |
192 | { |
193 | return sysctl_sched_rt_runtime >= 0; |
194 | } |
195 | |
196 | static void start_rt_bandwidth(struct rt_bandwidth *rt_b) |
197 | { |
198 | ktime_t now; |
199 | |
200 | if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF) |
201 | return; |
202 | |
203 | if (hrtimer_active(&rt_b->rt_period_timer)) |
204 | return; |
205 | |
206 | raw_spin_lock(&rt_b->rt_runtime_lock); |
207 | for (;;) { |
208 | unsigned long delta; |
209 | ktime_t soft, hard; |
210 | |
211 | if (hrtimer_active(&rt_b->rt_period_timer)) |
212 | break; |
213 | |
214 | now = hrtimer_cb_get_time(&rt_b->rt_period_timer); |
215 | hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period); |
216 | |
217 | soft = hrtimer_get_softexpires(&rt_b->rt_period_timer); |
218 | hard = hrtimer_get_expires(&rt_b->rt_period_timer); |
219 | delta = ktime_to_ns(ktime_sub(hard, soft)); |
220 | __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta, |
221 | HRTIMER_MODE_ABS_PINNED, 0); |
222 | } |
223 | raw_spin_unlock(&rt_b->rt_runtime_lock); |
224 | } |
225 | |
226 | #ifdef CONFIG_RT_GROUP_SCHED |
227 | static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b) |
228 | { |
229 | hrtimer_cancel(&rt_b->rt_period_timer); |
230 | } |
231 | #endif |
232 | |
233 | /* |
234 | * sched_domains_mutex serializes calls to arch_init_sched_domains, |
235 | * detach_destroy_domains and partition_sched_domains. |
236 | */ |
237 | static DEFINE_MUTEX(sched_domains_mutex); |
238 | |
239 | #ifdef CONFIG_CGROUP_SCHED |
240 | |
241 | #include <linux/cgroup.h> |
242 | |
243 | struct cfs_rq; |
244 | |
245 | static LIST_HEAD(task_groups); |
246 | |
247 | /* task group related information */ |
248 | struct task_group { |
249 | struct cgroup_subsys_state css; |
250 | |
251 | #ifdef CONFIG_FAIR_GROUP_SCHED |
252 | /* schedulable entities of this group on each cpu */ |
253 | struct sched_entity **se; |
254 | /* runqueue "owned" by this group on each cpu */ |
255 | struct cfs_rq **cfs_rq; |
256 | unsigned long shares; |
257 | |
258 | atomic_t load_weight; |
259 | #endif |
260 | |
261 | #ifdef CONFIG_RT_GROUP_SCHED |
262 | struct sched_rt_entity **rt_se; |
263 | struct rt_rq **rt_rq; |
264 | |
265 | struct rt_bandwidth rt_bandwidth; |
266 | #endif |
267 | |
268 | struct rcu_head rcu; |
269 | struct list_head list; |
270 | |
271 | struct task_group *parent; |
272 | struct list_head siblings; |
273 | struct list_head children; |
274 | |
275 | #ifdef CONFIG_SCHED_AUTOGROUP |
276 | struct autogroup *autogroup; |
277 | #endif |
278 | }; |
279 | |
280 | /* task_group_lock serializes the addition/removal of task groups */ |
281 | static DEFINE_SPINLOCK(task_group_lock); |
282 | |
283 | #ifdef CONFIG_FAIR_GROUP_SCHED |
284 | |
285 | # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD |
286 | |
287 | /* |
288 | * A weight of 0 or 1 can cause arithmetics problems. |
289 | * A weight of a cfs_rq is the sum of weights of which entities |
290 | * are queued on this cfs_rq, so a weight of a entity should not be |
291 | * too large, so as the shares value of a task group. |
292 | * (The default weight is 1024 - so there's no practical |
293 | * limitation from this.) |
294 | */ |
295 | #define MIN_SHARES 2 |
296 | #define MAX_SHARES (1UL << 18) |
297 | |
298 | static int root_task_group_load = ROOT_TASK_GROUP_LOAD; |
299 | #endif |
300 | |
301 | /* Default task group. |
302 | * Every task in system belong to this group at bootup. |
303 | */ |
304 | struct task_group root_task_group; |
305 | |
306 | #endif /* CONFIG_CGROUP_SCHED */ |
307 | |
308 | /* CFS-related fields in a runqueue */ |
309 | struct cfs_rq { |
310 | struct load_weight load; |
311 | unsigned long nr_running; |
312 | |
313 | u64 exec_clock; |
314 | u64 min_vruntime; |
315 | |
316 | struct rb_root tasks_timeline; |
317 | struct rb_node *rb_leftmost; |
318 | |
319 | struct list_head tasks; |
320 | struct list_head *balance_iterator; |
321 | |
322 | /* |
323 | * 'curr' points to currently running entity on this cfs_rq. |
324 | * It is set to NULL otherwise (i.e when none are currently running). |
325 | */ |
326 | struct sched_entity *curr, *next, *last, *skip; |
327 | |
328 | unsigned int nr_spread_over; |
329 | |
330 | #ifdef CONFIG_FAIR_GROUP_SCHED |
331 | struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */ |
332 | |
333 | /* |
334 | * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in |
335 | * a hierarchy). Non-leaf lrqs hold other higher schedulable entities |
336 | * (like users, containers etc.) |
337 | * |
338 | * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This |
339 | * list is used during load balance. |
340 | */ |
341 | int on_list; |
342 | struct list_head leaf_cfs_rq_list; |
343 | struct task_group *tg; /* group that "owns" this runqueue */ |
344 | |
345 | #ifdef CONFIG_SMP |
346 | /* |
347 | * the part of load.weight contributed by tasks |
348 | */ |
349 | unsigned long task_weight; |
350 | |
351 | /* |
352 | * h_load = weight * f(tg) |
353 | * |
354 | * Where f(tg) is the recursive weight fraction assigned to |
355 | * this group. |
356 | */ |
357 | unsigned long h_load; |
358 | |
359 | /* |
360 | * Maintaining per-cpu shares distribution for group scheduling |
361 | * |
362 | * load_stamp is the last time we updated the load average |
363 | * load_last is the last time we updated the load average and saw load |
364 | * load_unacc_exec_time is currently unaccounted execution time |
365 | */ |
366 | u64 load_avg; |
367 | u64 load_period; |
368 | u64 load_stamp, load_last, load_unacc_exec_time; |
369 | |
370 | unsigned long load_contribution; |
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 | #endif |
556 | }; |
557 | |
558 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); |
559 | |
560 | |
561 | static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags); |
562 | |
563 | static inline int cpu_of(struct rq *rq) |
564 | { |
565 | #ifdef CONFIG_SMP |
566 | return rq->cpu; |
567 | #else |
568 | return 0; |
569 | #endif |
570 | } |
571 | |
572 | #define rcu_dereference_check_sched_domain(p) \ |
573 | rcu_dereference_check((p), \ |
574 | rcu_read_lock_sched_held() || \ |
575 | lockdep_is_held(&sched_domains_mutex)) |
576 | |
577 | /* |
578 | * The domain tree (rq->sd) is protected by RCU's quiescent state transition. |
579 | * See detach_destroy_domains: synchronize_sched for details. |
580 | * |
581 | * The domain tree of any CPU may only be accessed from within |
582 | * preempt-disabled sections. |
583 | */ |
584 | #define for_each_domain(cpu, __sd) \ |
585 | for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent) |
586 | |
587 | #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) |
588 | #define this_rq() (&__get_cpu_var(runqueues)) |
589 | #define task_rq(p) cpu_rq(task_cpu(p)) |
590 | #define cpu_curr(cpu) (cpu_rq(cpu)->curr) |
591 | #define raw_rq() (&__raw_get_cpu_var(runqueues)) |
592 | |
593 | #ifdef CONFIG_CGROUP_SCHED |
594 | |
595 | /* |
596 | * Return the group to which this tasks belongs. |
597 | * |
598 | * We use task_subsys_state_check() and extend the RCU verification |
599 | * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach() |
600 | * holds that lock for each task it moves into the cgroup. Therefore |
601 | * by holding that lock, we pin the task to the current cgroup. |
602 | */ |
603 | static inline struct task_group *task_group(struct task_struct *p) |
604 | { |
605 | struct task_group *tg; |
606 | struct cgroup_subsys_state *css; |
607 | |
608 | css = task_subsys_state_check(p, cpu_cgroup_subsys_id, |
609 | lockdep_is_held(&task_rq(p)->lock)); |
610 | tg = container_of(css, struct task_group, css); |
611 | |
612 | return autogroup_task_group(p, tg); |
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 void update_rq_clock_task(struct rq *rq, s64 delta); |
640 | |
641 | static void update_rq_clock(struct rq *rq) |
642 | { |
643 | s64 delta; |
644 | |
645 | if (rq->skip_clock_update) |
646 | return; |
647 | |
648 | delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; |
649 | rq->clock += delta; |
650 | update_rq_clock_task(rq, delta); |
651 | } |
652 | |
653 | /* |
654 | * Tunables that become constants when CONFIG_SCHED_DEBUG is off: |
655 | */ |
656 | #ifdef CONFIG_SCHED_DEBUG |
657 | # define const_debug __read_mostly |
658 | #else |
659 | # define const_debug static const |
660 | #endif |
661 | |
662 | /** |
663 | * runqueue_is_locked - Returns true if the current cpu runqueue is locked |
664 | * @cpu: the processor in question. |
665 | * |
666 | * This interface allows printk to be called with the runqueue lock |
667 | * held and know whether or not it is OK to wake up the klogd. |
668 | */ |
669 | int runqueue_is_locked(int cpu) |
670 | { |
671 | return raw_spin_is_locked(&cpu_rq(cpu)->lock); |
672 | } |
673 | |
674 | /* |
675 | * Debugging: various feature bits |
676 | */ |
677 | |
678 | #define SCHED_FEAT(name, enabled) \ |
679 | __SCHED_FEAT_##name , |
680 | |
681 | enum { |
682 | #include "sched_features.h" |
683 | }; |
684 | |
685 | #undef SCHED_FEAT |
686 | |
687 | #define SCHED_FEAT(name, enabled) \ |
688 | (1UL << __SCHED_FEAT_##name) * enabled | |
689 | |
690 | const_debug unsigned int sysctl_sched_features = |
691 | #include "sched_features.h" |
692 | 0; |
693 | |
694 | #undef SCHED_FEAT |
695 | |
696 | #ifdef CONFIG_SCHED_DEBUG |
697 | #define SCHED_FEAT(name, enabled) \ |
698 | #name , |
699 | |
700 | static __read_mostly char *sched_feat_names[] = { |
701 | #include "sched_features.h" |
702 | NULL |
703 | }; |
704 | |
705 | #undef SCHED_FEAT |
706 | |
707 | static int sched_feat_show(struct seq_file *m, void *v) |
708 | { |
709 | int i; |
710 | |
711 | for (i = 0; sched_feat_names[i]; i++) { |
712 | if (!(sysctl_sched_features & (1UL << i))) |
713 | seq_puts(m, "NO_"); |
714 | seq_printf(m, "%s ", sched_feat_names[i]); |
715 | } |
716 | seq_puts(m, "\n"); |
717 | |
718 | return 0; |
719 | } |
720 | |
721 | static ssize_t |
722 | sched_feat_write(struct file *filp, const char __user *ubuf, |
723 | size_t cnt, loff_t *ppos) |
724 | { |
725 | char buf[64]; |
726 | char *cmp; |
727 | int neg = 0; |
728 | int i; |
729 | |
730 | if (cnt > 63) |
731 | cnt = 63; |
732 | |
733 | if (copy_from_user(&buf, ubuf, cnt)) |
734 | return -EFAULT; |
735 | |
736 | buf[cnt] = 0; |
737 | cmp = strstrip(buf); |
738 | |
739 | if (strncmp(cmp, "NO_", 3) == 0) { |
740 | neg = 1; |
741 | cmp += 3; |
742 | } |
743 | |
744 | for (i = 0; sched_feat_names[i]; i++) { |
745 | if (strcmp(cmp, sched_feat_names[i]) == 0) { |
746 | if (neg) |
747 | sysctl_sched_features &= ~(1UL << i); |
748 | else |
749 | sysctl_sched_features |= (1UL << i); |
750 | break; |
751 | } |
752 | } |
753 | |
754 | if (!sched_feat_names[i]) |
755 | return -EINVAL; |
756 | |
757 | *ppos += cnt; |
758 | |
759 | return cnt; |
760 | } |
761 | |
762 | static int sched_feat_open(struct inode *inode, struct file *filp) |
763 | { |
764 | return single_open(filp, sched_feat_show, NULL); |
765 | } |
766 | |
767 | static const struct file_operations sched_feat_fops = { |
768 | .open = sched_feat_open, |
769 | .write = sched_feat_write, |
770 | .read = seq_read, |
771 | .llseek = seq_lseek, |
772 | .release = single_release, |
773 | }; |
774 | |
775 | static __init int sched_init_debug(void) |
776 | { |
777 | debugfs_create_file("sched_features", 0644, NULL, NULL, |
778 | &sched_feat_fops); |
779 | |
780 | return 0; |
781 | } |
782 | late_initcall(sched_init_debug); |
783 | |
784 | #endif |
785 | |
786 | #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x)) |
787 | |
788 | /* |
789 | * Number of tasks to iterate in a single balance run. |
790 | * Limited because this is done with IRQs disabled. |
791 | */ |
792 | const_debug unsigned int sysctl_sched_nr_migrate = 32; |
793 | |
794 | /* |
795 | * period over which we average the RT time consumption, measured |
796 | * in ms. |
797 | * |
798 | * default: 1s |
799 | */ |
800 | const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; |
801 | |
802 | /* |
803 | * period over which we measure -rt task cpu usage in us. |
804 | * default: 1s |
805 | */ |
806 | unsigned int sysctl_sched_rt_period = 1000000; |
807 | |
808 | static __read_mostly int scheduler_running; |
809 | |
810 | /* |
811 | * part of the period that we allow rt tasks to run in us. |
812 | * default: 0.95s |
813 | */ |
814 | int sysctl_sched_rt_runtime = 950000; |
815 | |
816 | static inline u64 global_rt_period(void) |
817 | { |
818 | return (u64)sysctl_sched_rt_period * NSEC_PER_USEC; |
819 | } |
820 | |
821 | static inline u64 global_rt_runtime(void) |
822 | { |
823 | if (sysctl_sched_rt_runtime < 0) |
824 | return RUNTIME_INF; |
825 | |
826 | return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC; |
827 | } |
828 | |
829 | #ifndef prepare_arch_switch |
830 | # define prepare_arch_switch(next) do { } while (0) |
831 | #endif |
832 | #ifndef finish_arch_switch |
833 | # define finish_arch_switch(prev) do { } while (0) |
834 | #endif |
835 | |
836 | static inline int task_current(struct rq *rq, struct task_struct *p) |
837 | { |
838 | return rq->curr == p; |
839 | } |
840 | |
841 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW |
842 | static inline int task_running(struct rq *rq, struct task_struct *p) |
843 | { |
844 | return task_current(rq, p); |
845 | } |
846 | |
847 | static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) |
848 | { |
849 | } |
850 | |
851 | static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) |
852 | { |
853 | #ifdef CONFIG_DEBUG_SPINLOCK |
854 | /* this is a valid case when another task releases the spinlock */ |
855 | rq->lock.owner = current; |
856 | #endif |
857 | /* |
858 | * If we are tracking spinlock dependencies then we have to |
859 | * fix up the runqueue lock - which gets 'carried over' from |
860 | * prev into current: |
861 | */ |
862 | spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_); |
863 | |
864 | raw_spin_unlock_irq(&rq->lock); |
865 | } |
866 | |
867 | #else /* __ARCH_WANT_UNLOCKED_CTXSW */ |
868 | static inline int task_running(struct rq *rq, struct task_struct *p) |
869 | { |
870 | #ifdef CONFIG_SMP |
871 | return p->oncpu; |
872 | #else |
873 | return task_current(rq, p); |
874 | #endif |
875 | } |
876 | |
877 | static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) |
878 | { |
879 | #ifdef CONFIG_SMP |
880 | /* |
881 | * We can optimise this out completely for !SMP, because the |
882 | * SMP rebalancing from interrupt is the only thing that cares |
883 | * here. |
884 | */ |
885 | next->oncpu = 1; |
886 | #endif |
887 | #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW |
888 | raw_spin_unlock_irq(&rq->lock); |
889 | #else |
890 | raw_spin_unlock(&rq->lock); |
891 | #endif |
892 | } |
893 | |
894 | static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) |
895 | { |
896 | #ifdef CONFIG_SMP |
897 | /* |
898 | * After ->oncpu is cleared, the task can be moved to a different CPU. |
899 | * We must ensure this doesn't happen until the switch is completely |
900 | * finished. |
901 | */ |
902 | smp_wmb(); |
903 | prev->oncpu = 0; |
904 | #endif |
905 | #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW |
906 | local_irq_enable(); |
907 | #endif |
908 | } |
909 | #endif /* __ARCH_WANT_UNLOCKED_CTXSW */ |
910 | |
911 | /* |
912 | * Check whether the task is waking, we use this to synchronize ->cpus_allowed |
913 | * against ttwu(). |
914 | */ |
915 | static inline int task_is_waking(struct task_struct *p) |
916 | { |
917 | return unlikely(p->state == TASK_WAKING); |
918 | } |
919 | |
920 | /* |
921 | * __task_rq_lock - lock the runqueue a given task resides on. |
922 | * Must be called interrupts disabled. |
923 | */ |
924 | static inline struct rq *__task_rq_lock(struct task_struct *p) |
925 | __acquires(rq->lock) |
926 | { |
927 | struct rq *rq; |
928 | |
929 | for (;;) { |
930 | rq = task_rq(p); |
931 | raw_spin_lock(&rq->lock); |
932 | if (likely(rq == task_rq(p))) |
933 | return rq; |
934 | raw_spin_unlock(&rq->lock); |
935 | } |
936 | } |
937 | |
938 | /* |
939 | * task_rq_lock - lock the runqueue a given task resides on and disable |
940 | * interrupts. Note the ordering: we can safely lookup the task_rq without |
941 | * explicitly disabling preemption. |
942 | */ |
943 | static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) |
944 | __acquires(rq->lock) |
945 | { |
946 | struct rq *rq; |
947 | |
948 | for (;;) { |
949 | local_irq_save(*flags); |
950 | rq = task_rq(p); |
951 | raw_spin_lock(&rq->lock); |
952 | if (likely(rq == task_rq(p))) |
953 | return rq; |
954 | raw_spin_unlock_irqrestore(&rq->lock, *flags); |
955 | } |
956 | } |
957 | |
958 | static void __task_rq_unlock(struct rq *rq) |
959 | __releases(rq->lock) |
960 | { |
961 | raw_spin_unlock(&rq->lock); |
962 | } |
963 | |
964 | static inline void task_rq_unlock(struct rq *rq, unsigned long *flags) |
965 | __releases(rq->lock) |
966 | { |
967 | raw_spin_unlock_irqrestore(&rq->lock, *flags); |
968 | } |
969 | |
970 | /* |
971 | * this_rq_lock - lock this runqueue and disable interrupts. |
972 | */ |
973 | static struct rq *this_rq_lock(void) |
974 | __acquires(rq->lock) |
975 | { |
976 | struct rq *rq; |
977 | |
978 | local_irq_disable(); |
979 | rq = this_rq(); |
980 | raw_spin_lock(&rq->lock); |
981 | |
982 | return rq; |
983 | } |
984 | |
985 | #ifdef CONFIG_SCHED_HRTICK |
986 | /* |
987 | * Use HR-timers to deliver accurate preemption points. |
988 | * |
989 | * Its all a bit involved since we cannot program an hrt while holding the |
990 | * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a |
991 | * reschedule event. |
992 | * |
993 | * When we get rescheduled we reprogram the hrtick_timer outside of the |
994 | * rq->lock. |
995 | */ |
996 | |
997 | /* |
998 | * Use hrtick when: |
999 | * - enabled by features |
1000 | * - hrtimer is actually high res |
1001 | */ |
1002 | static inline int hrtick_enabled(struct rq *rq) |
1003 | { |
1004 | if (!sched_feat(HRTICK)) |
1005 | return 0; |
1006 | if (!cpu_active(cpu_of(rq))) |
1007 | return 0; |
1008 | return hrtimer_is_hres_active(&rq->hrtick_timer); |
1009 | } |
1010 | |
1011 | static void hrtick_clear(struct rq *rq) |
1012 | { |
1013 | if (hrtimer_active(&rq->hrtick_timer)) |
1014 | hrtimer_cancel(&rq->hrtick_timer); |
1015 | } |
1016 | |
1017 | /* |
1018 | * High-resolution timer tick. |
1019 | * Runs from hardirq context with interrupts disabled. |
1020 | */ |
1021 | static enum hrtimer_restart hrtick(struct hrtimer *timer) |
1022 | { |
1023 | struct rq *rq = container_of(timer, struct rq, hrtick_timer); |
1024 | |
1025 | WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); |
1026 | |
1027 | raw_spin_lock(&rq->lock); |
1028 | update_rq_clock(rq); |
1029 | rq->curr->sched_class->task_tick(rq, rq->curr, 1); |
1030 | raw_spin_unlock(&rq->lock); |
1031 | |
1032 | return HRTIMER_NORESTART; |
1033 | } |
1034 | |
1035 | #ifdef CONFIG_SMP |
1036 | /* |
1037 | * called from hardirq (IPI) context |
1038 | */ |
1039 | static void __hrtick_start(void *arg) |
1040 | { |
1041 | struct rq *rq = arg; |
1042 | |
1043 | raw_spin_lock(&rq->lock); |
1044 | hrtimer_restart(&rq->hrtick_timer); |
1045 | rq->hrtick_csd_pending = 0; |
1046 | raw_spin_unlock(&rq->lock); |
1047 | } |
1048 | |
1049 | /* |
1050 | * Called to set the hrtick timer state. |
1051 | * |
1052 | * called with rq->lock held and irqs disabled |
1053 | */ |
1054 | static void hrtick_start(struct rq *rq, u64 delay) |
1055 | { |
1056 | struct hrtimer *timer = &rq->hrtick_timer; |
1057 | ktime_t time = ktime_add_ns(timer->base->get_time(), delay); |
1058 | |
1059 | hrtimer_set_expires(timer, time); |
1060 | |
1061 | if (rq == this_rq()) { |
1062 | hrtimer_restart(timer); |
1063 | } else if (!rq->hrtick_csd_pending) { |
1064 | __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0); |
1065 | rq->hrtick_csd_pending = 1; |
1066 | } |
1067 | } |
1068 | |
1069 | static int |
1070 | hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) |
1071 | { |
1072 | int cpu = (int)(long)hcpu; |
1073 | |
1074 | switch (action) { |
1075 | case CPU_UP_CANCELED: |
1076 | case CPU_UP_CANCELED_FROZEN: |
1077 | case CPU_DOWN_PREPARE: |
1078 | case CPU_DOWN_PREPARE_FROZEN: |
1079 | case CPU_DEAD: |
1080 | case CPU_DEAD_FROZEN: |
1081 | hrtick_clear(cpu_rq(cpu)); |
1082 | return NOTIFY_OK; |
1083 | } |
1084 | |
1085 | return NOTIFY_DONE; |
1086 | } |
1087 | |
1088 | static __init void init_hrtick(void) |
1089 | { |
1090 | hotcpu_notifier(hotplug_hrtick, 0); |
1091 | } |
1092 | #else |
1093 | /* |
1094 | * Called to set the hrtick timer state. |
1095 | * |
1096 | * called with rq->lock held and irqs disabled |
1097 | */ |
1098 | static void hrtick_start(struct rq *rq, u64 delay) |
1099 | { |
1100 | __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, |
1101 | HRTIMER_MODE_REL_PINNED, 0); |
1102 | } |
1103 | |
1104 | static inline void init_hrtick(void) |
1105 | { |
1106 | } |
1107 | #endif /* CONFIG_SMP */ |
1108 | |
1109 | static void init_rq_hrtick(struct rq *rq) |
1110 | { |
1111 | #ifdef CONFIG_SMP |
1112 | rq->hrtick_csd_pending = 0; |
1113 | |
1114 | rq->hrtick_csd.flags = 0; |
1115 | rq->hrtick_csd.func = __hrtick_start; |
1116 | rq->hrtick_csd.info = rq; |
1117 | #endif |
1118 | |
1119 | hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
1120 | rq->hrtick_timer.function = hrtick; |
1121 | } |
1122 | #else /* CONFIG_SCHED_HRTICK */ |
1123 | static inline void hrtick_clear(struct rq *rq) |
1124 | { |
1125 | } |
1126 | |
1127 | static inline void init_rq_hrtick(struct rq *rq) |
1128 | { |
1129 | } |
1130 | |
1131 | static inline void init_hrtick(void) |
1132 | { |
1133 | } |
1134 | #endif /* CONFIG_SCHED_HRTICK */ |
1135 | |
1136 | /* |
1137 | * resched_task - mark a task 'to be rescheduled now'. |
1138 | * |
1139 | * On UP this means the setting of the need_resched flag, on SMP it |
1140 | * might also involve a cross-CPU call to trigger the scheduler on |
1141 | * the target CPU. |
1142 | */ |
1143 | #ifdef CONFIG_SMP |
1144 | |
1145 | #ifndef tsk_is_polling |
1146 | #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG) |
1147 | #endif |
1148 | |
1149 | static void resched_task(struct task_struct *p) |
1150 | { |
1151 | int cpu; |
1152 | |
1153 | assert_raw_spin_locked(&task_rq(p)->lock); |
1154 | |
1155 | if (test_tsk_need_resched(p)) |
1156 | return; |
1157 | |
1158 | set_tsk_need_resched(p); |
1159 | |
1160 | cpu = task_cpu(p); |
1161 | if (cpu == smp_processor_id()) |
1162 | return; |
1163 | |
1164 | /* NEED_RESCHED must be visible before we test polling */ |
1165 | smp_mb(); |
1166 | if (!tsk_is_polling(p)) |
1167 | smp_send_reschedule(cpu); |
1168 | } |
1169 | |
1170 | static void resched_cpu(int cpu) |
1171 | { |
1172 | struct rq *rq = cpu_rq(cpu); |
1173 | unsigned long flags; |
1174 | |
1175 | if (!raw_spin_trylock_irqsave(&rq->lock, flags)) |
1176 | return; |
1177 | resched_task(cpu_curr(cpu)); |
1178 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
1179 | } |
1180 | |
1181 | #ifdef CONFIG_NO_HZ |
1182 | /* |
1183 | * In the semi idle case, use the nearest busy cpu for migrating timers |
1184 | * from an idle cpu. This is good for power-savings. |
1185 | * |
1186 | * We don't do similar optimization for completely idle system, as |
1187 | * selecting an idle cpu will add more delays to the timers than intended |
1188 | * (as that cpu's timer base may not be uptodate wrt jiffies etc). |
1189 | */ |
1190 | int get_nohz_timer_target(void) |
1191 | { |
1192 | int cpu = smp_processor_id(); |
1193 | int i; |
1194 | struct sched_domain *sd; |
1195 | |
1196 | for_each_domain(cpu, sd) { |
1197 | for_each_cpu(i, sched_domain_span(sd)) |
1198 | if (!idle_cpu(i)) |
1199 | return i; |
1200 | } |
1201 | return cpu; |
1202 | } |
1203 | /* |
1204 | * When add_timer_on() enqueues a timer into the timer wheel of an |
1205 | * idle CPU then this timer might expire before the next timer event |
1206 | * which is scheduled to wake up that CPU. In case of a completely |
1207 | * idle system the next event might even be infinite time into the |
1208 | * future. wake_up_idle_cpu() ensures that the CPU is woken up and |
1209 | * leaves the inner idle loop so the newly added timer is taken into |
1210 | * account when the CPU goes back to idle and evaluates the timer |
1211 | * wheel for the next timer event. |
1212 | */ |
1213 | void wake_up_idle_cpu(int cpu) |
1214 | { |
1215 | struct rq *rq = cpu_rq(cpu); |
1216 | |
1217 | if (cpu == smp_processor_id()) |
1218 | return; |
1219 | |
1220 | /* |
1221 | * This is safe, as this function is called with the timer |
1222 | * wheel base lock of (cpu) held. When the CPU is on the way |
1223 | * to idle and has not yet set rq->curr to idle then it will |
1224 | * be serialized on the timer wheel base lock and take the new |
1225 | * timer into account automatically. |
1226 | */ |
1227 | if (rq->curr != rq->idle) |
1228 | return; |
1229 | |
1230 | /* |
1231 | * We can set TIF_RESCHED on the idle task of the other CPU |
1232 | * lockless. The worst case is that the other CPU runs the |
1233 | * idle task through an additional NOOP schedule() |
1234 | */ |
1235 | set_tsk_need_resched(rq->idle); |
1236 | |
1237 | /* NEED_RESCHED must be visible before we test polling */ |
1238 | smp_mb(); |
1239 | if (!tsk_is_polling(rq->idle)) |
1240 | smp_send_reschedule(cpu); |
1241 | } |
1242 | |
1243 | #endif /* CONFIG_NO_HZ */ |
1244 | |
1245 | static u64 sched_avg_period(void) |
1246 | { |
1247 | return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2; |
1248 | } |
1249 | |
1250 | static void sched_avg_update(struct rq *rq) |
1251 | { |
1252 | s64 period = sched_avg_period(); |
1253 | |
1254 | while ((s64)(rq->clock - rq->age_stamp) > period) { |
1255 | /* |
1256 | * Inline assembly required to prevent the compiler |
1257 | * optimising this loop into a divmod call. |
1258 | * See __iter_div_u64_rem() for another example of this. |
1259 | */ |
1260 | asm("" : "+rm" (rq->age_stamp)); |
1261 | rq->age_stamp += period; |
1262 | rq->rt_avg /= 2; |
1263 | } |
1264 | } |
1265 | |
1266 | static void sched_rt_avg_update(struct rq *rq, u64 rt_delta) |
1267 | { |
1268 | rq->rt_avg += rt_delta; |
1269 | sched_avg_update(rq); |
1270 | } |
1271 | |
1272 | #else /* !CONFIG_SMP */ |
1273 | static void resched_task(struct task_struct *p) |
1274 | { |
1275 | assert_raw_spin_locked(&task_rq(p)->lock); |
1276 | set_tsk_need_resched(p); |
1277 | } |
1278 | |
1279 | static void sched_rt_avg_update(struct rq *rq, u64 rt_delta) |
1280 | { |
1281 | } |
1282 | |
1283 | static void sched_avg_update(struct rq *rq) |
1284 | { |
1285 | } |
1286 | #endif /* CONFIG_SMP */ |
1287 | |
1288 | #if BITS_PER_LONG == 32 |
1289 | # define WMULT_CONST (~0UL) |
1290 | #else |
1291 | # define WMULT_CONST (1UL << 32) |
1292 | #endif |
1293 | |
1294 | #define WMULT_SHIFT 32 |
1295 | |
1296 | /* |
1297 | * Shift right and round: |
1298 | */ |
1299 | #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y)) |
1300 | |
1301 | /* |
1302 | * delta *= weight / lw |
1303 | */ |
1304 | static unsigned long |
1305 | calc_delta_mine(unsigned long delta_exec, unsigned long weight, |
1306 | struct load_weight *lw) |
1307 | { |
1308 | u64 tmp; |
1309 | |
1310 | if (!lw->inv_weight) { |
1311 | if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST)) |
1312 | lw->inv_weight = 1; |
1313 | else |
1314 | lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2) |
1315 | / (lw->weight+1); |
1316 | } |
1317 | |
1318 | tmp = (u64)delta_exec * weight; |
1319 | /* |
1320 | * Check whether we'd overflow the 64-bit multiplication: |
1321 | */ |
1322 | if (unlikely(tmp > WMULT_CONST)) |
1323 | tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight, |
1324 | WMULT_SHIFT/2); |
1325 | else |
1326 | tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT); |
1327 | |
1328 | return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX); |
1329 | } |
1330 | |
1331 | static inline void update_load_add(struct load_weight *lw, unsigned long inc) |
1332 | { |
1333 | lw->weight += inc; |
1334 | lw->inv_weight = 0; |
1335 | } |
1336 | |
1337 | static inline void update_load_sub(struct load_weight *lw, unsigned long dec) |
1338 | { |
1339 | lw->weight -= dec; |
1340 | lw->inv_weight = 0; |
1341 | } |
1342 | |
1343 | static inline void update_load_set(struct load_weight *lw, unsigned long w) |
1344 | { |
1345 | lw->weight = w; |
1346 | lw->inv_weight = 0; |
1347 | } |
1348 | |
1349 | /* |
1350 | * To aid in avoiding the subversion of "niceness" due to uneven distribution |
1351 | * of tasks with abnormal "nice" values across CPUs the contribution that |
1352 | * each task makes to its run queue's load is weighted according to its |
1353 | * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a |
1354 | * scaled version of the new time slice allocation that they receive on time |
1355 | * slice expiry etc. |
1356 | */ |
1357 | |
1358 | #define WEIGHT_IDLEPRIO 3 |
1359 | #define WMULT_IDLEPRIO 1431655765 |
1360 | |
1361 | /* |
1362 | * Nice levels are multiplicative, with a gentle 10% change for every |
1363 | * nice level changed. I.e. when a CPU-bound task goes from nice 0 to |
1364 | * nice 1, it will get ~10% less CPU time than another CPU-bound task |
1365 | * that remained on nice 0. |
1366 | * |
1367 | * The "10% effect" is relative and cumulative: from _any_ nice level, |
1368 | * if you go up 1 level, it's -10% CPU usage, if you go down 1 level |
1369 | * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25. |
1370 | * If a task goes up by ~10% and another task goes down by ~10% then |
1371 | * the relative distance between them is ~25%.) |
1372 | */ |
1373 | static const int prio_to_weight[40] = { |
1374 | /* -20 */ 88761, 71755, 56483, 46273, 36291, |
1375 | /* -15 */ 29154, 23254, 18705, 14949, 11916, |
1376 | /* -10 */ 9548, 7620, 6100, 4904, 3906, |
1377 | /* -5 */ 3121, 2501, 1991, 1586, 1277, |
1378 | /* 0 */ 1024, 820, 655, 526, 423, |
1379 | /* 5 */ 335, 272, 215, 172, 137, |
1380 | /* 10 */ 110, 87, 70, 56, 45, |
1381 | /* 15 */ 36, 29, 23, 18, 15, |
1382 | }; |
1383 | |
1384 | /* |
1385 | * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated. |
1386 | * |
1387 | * In cases where the weight does not change often, we can use the |
1388 | * precalculated inverse to speed up arithmetics by turning divisions |
1389 | * into multiplications: |
1390 | */ |
1391 | static const u32 prio_to_wmult[40] = { |
1392 | /* -20 */ 48388, 59856, 76040, 92818, 118348, |
1393 | /* -15 */ 147320, 184698, 229616, 287308, 360437, |
1394 | /* -10 */ 449829, 563644, 704093, 875809, 1099582, |
1395 | /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326, |
1396 | /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587, |
1397 | /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126, |
1398 | /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717, |
1399 | /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153, |
1400 | }; |
1401 | |
1402 | /* Time spent by the tasks of the cpu accounting group executing in ... */ |
1403 | enum cpuacct_stat_index { |
1404 | CPUACCT_STAT_USER, /* ... user mode */ |
1405 | CPUACCT_STAT_SYSTEM, /* ... kernel mode */ |
1406 | |
1407 | CPUACCT_STAT_NSTATS, |
1408 | }; |
1409 | |
1410 | #ifdef CONFIG_CGROUP_CPUACCT |
1411 | static void cpuacct_charge(struct task_struct *tsk, u64 cputime); |
1412 | static void cpuacct_update_stats(struct task_struct *tsk, |
1413 | enum cpuacct_stat_index idx, cputime_t val); |
1414 | #else |
1415 | static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {} |
1416 | static inline void cpuacct_update_stats(struct task_struct *tsk, |
1417 | enum cpuacct_stat_index idx, cputime_t val) {} |
1418 | #endif |
1419 | |
1420 | static inline void inc_cpu_load(struct rq *rq, unsigned long load) |
1421 | { |
1422 | update_load_add(&rq->load, load); |
1423 | } |
1424 | |
1425 | static inline void dec_cpu_load(struct rq *rq, unsigned long load) |
1426 | { |
1427 | update_load_sub(&rq->load, load); |
1428 | } |
1429 | |
1430 | #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED) |
1431 | typedef int (*tg_visitor)(struct task_group *, void *); |
1432 | |
1433 | /* |
1434 | * Iterate the full tree, calling @down when first entering a node and @up when |
1435 | * leaving it for the final time. |
1436 | */ |
1437 | static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data) |
1438 | { |
1439 | struct task_group *parent, *child; |
1440 | int ret; |
1441 | |
1442 | rcu_read_lock(); |
1443 | parent = &root_task_group; |
1444 | down: |
1445 | ret = (*down)(parent, data); |
1446 | if (ret) |
1447 | goto out_unlock; |
1448 | list_for_each_entry_rcu(child, &parent->children, siblings) { |
1449 | parent = child; |
1450 | goto down; |
1451 | |
1452 | up: |
1453 | continue; |
1454 | } |
1455 | ret = (*up)(parent, data); |
1456 | if (ret) |
1457 | goto out_unlock; |
1458 | |
1459 | child = parent; |
1460 | parent = parent->parent; |
1461 | if (parent) |
1462 | goto up; |
1463 | out_unlock: |
1464 | rcu_read_unlock(); |
1465 | |
1466 | return ret; |
1467 | } |
1468 | |
1469 | static int tg_nop(struct task_group *tg, void *data) |
1470 | { |
1471 | return 0; |
1472 | } |
1473 | #endif |
1474 | |
1475 | #ifdef CONFIG_SMP |
1476 | /* Used instead of source_load when we know the type == 0 */ |
1477 | static unsigned long weighted_cpuload(const int cpu) |
1478 | { |
1479 | return cpu_rq(cpu)->load.weight; |
1480 | } |
1481 | |
1482 | /* |
1483 | * Return a low guess at the load of a migration-source cpu weighted |
1484 | * according to the scheduling class and "nice" value. |
1485 | * |
1486 | * We want to under-estimate the load of migration sources, to |
1487 | * balance conservatively. |
1488 | */ |
1489 | static unsigned long source_load(int cpu, int type) |
1490 | { |
1491 | struct rq *rq = cpu_rq(cpu); |
1492 | unsigned long total = weighted_cpuload(cpu); |
1493 | |
1494 | if (type == 0 || !sched_feat(LB_BIAS)) |
1495 | return total; |
1496 | |
1497 | return min(rq->cpu_load[type-1], total); |
1498 | } |
1499 | |
1500 | /* |
1501 | * Return a high guess at the load of a migration-target cpu weighted |
1502 | * according to the scheduling class and "nice" value. |
1503 | */ |
1504 | static unsigned long target_load(int cpu, int type) |
1505 | { |
1506 | struct rq *rq = cpu_rq(cpu); |
1507 | unsigned long total = weighted_cpuload(cpu); |
1508 | |
1509 | if (type == 0 || !sched_feat(LB_BIAS)) |
1510 | return total; |
1511 | |
1512 | return max(rq->cpu_load[type-1], total); |
1513 | } |
1514 | |
1515 | static unsigned long power_of(int cpu) |
1516 | { |
1517 | return cpu_rq(cpu)->cpu_power; |
1518 | } |
1519 | |
1520 | static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd); |
1521 | |
1522 | static unsigned long cpu_avg_load_per_task(int cpu) |
1523 | { |
1524 | struct rq *rq = cpu_rq(cpu); |
1525 | unsigned long nr_running = ACCESS_ONCE(rq->nr_running); |
1526 | |
1527 | if (nr_running) |
1528 | rq->avg_load_per_task = rq->load.weight / nr_running; |
1529 | else |
1530 | rq->avg_load_per_task = 0; |
1531 | |
1532 | return rq->avg_load_per_task; |
1533 | } |
1534 | |
1535 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1536 | |
1537 | /* |
1538 | * Compute the cpu's hierarchical load factor for each task group. |
1539 | * This needs to be done in a top-down fashion because the load of a child |
1540 | * group is a fraction of its parents load. |
1541 | */ |
1542 | static int tg_load_down(struct task_group *tg, void *data) |
1543 | { |
1544 | unsigned long load; |
1545 | long cpu = (long)data; |
1546 | |
1547 | if (!tg->parent) { |
1548 | load = cpu_rq(cpu)->load.weight; |
1549 | } else { |
1550 | load = tg->parent->cfs_rq[cpu]->h_load; |
1551 | load *= tg->se[cpu]->load.weight; |
1552 | load /= tg->parent->cfs_rq[cpu]->load.weight + 1; |
1553 | } |
1554 | |
1555 | tg->cfs_rq[cpu]->h_load = load; |
1556 | |
1557 | return 0; |
1558 | } |
1559 | |
1560 | static void update_h_load(long cpu) |
1561 | { |
1562 | walk_tg_tree(tg_load_down, tg_nop, (void *)cpu); |
1563 | } |
1564 | |
1565 | #endif |
1566 | |
1567 | #ifdef CONFIG_PREEMPT |
1568 | |
1569 | static void double_rq_lock(struct rq *rq1, struct rq *rq2); |
1570 | |
1571 | /* |
1572 | * fair double_lock_balance: Safely acquires both rq->locks in a fair |
1573 | * way at the expense of forcing extra atomic operations in all |
1574 | * invocations. This assures that the double_lock is acquired using the |
1575 | * same underlying policy as the spinlock_t on this architecture, which |
1576 | * reduces latency compared to the unfair variant below. However, it |
1577 | * also adds more overhead and therefore may reduce throughput. |
1578 | */ |
1579 | static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest) |
1580 | __releases(this_rq->lock) |
1581 | __acquires(busiest->lock) |
1582 | __acquires(this_rq->lock) |
1583 | { |
1584 | raw_spin_unlock(&this_rq->lock); |
1585 | double_rq_lock(this_rq, busiest); |
1586 | |
1587 | return 1; |
1588 | } |
1589 | |
1590 | #else |
1591 | /* |
1592 | * Unfair double_lock_balance: Optimizes throughput at the expense of |
1593 | * latency by eliminating extra atomic operations when the locks are |
1594 | * already in proper order on entry. This favors lower cpu-ids and will |
1595 | * grant the double lock to lower cpus over higher ids under contention, |
1596 | * regardless of entry order into the function. |
1597 | */ |
1598 | static int _double_lock_balance(struct rq *this_rq, struct rq *busiest) |
1599 | __releases(this_rq->lock) |
1600 | __acquires(busiest->lock) |
1601 | __acquires(this_rq->lock) |
1602 | { |
1603 | int ret = 0; |
1604 | |
1605 | if (unlikely(!raw_spin_trylock(&busiest->lock))) { |
1606 | if (busiest < this_rq) { |
1607 | raw_spin_unlock(&this_rq->lock); |
1608 | raw_spin_lock(&busiest->lock); |
1609 | raw_spin_lock_nested(&this_rq->lock, |
1610 | SINGLE_DEPTH_NESTING); |
1611 | ret = 1; |
1612 | } else |
1613 | raw_spin_lock_nested(&busiest->lock, |
1614 | SINGLE_DEPTH_NESTING); |
1615 | } |
1616 | return ret; |
1617 | } |
1618 | |
1619 | #endif /* CONFIG_PREEMPT */ |
1620 | |
1621 | /* |
1622 | * double_lock_balance - lock the busiest runqueue, this_rq is locked already. |
1623 | */ |
1624 | static int double_lock_balance(struct rq *this_rq, struct rq *busiest) |
1625 | { |
1626 | if (unlikely(!irqs_disabled())) { |
1627 | /* printk() doesn't work good under rq->lock */ |
1628 | raw_spin_unlock(&this_rq->lock); |
1629 | BUG_ON(1); |
1630 | } |
1631 | |
1632 | return _double_lock_balance(this_rq, busiest); |
1633 | } |
1634 | |
1635 | static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest) |
1636 | __releases(busiest->lock) |
1637 | { |
1638 | raw_spin_unlock(&busiest->lock); |
1639 | lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_); |
1640 | } |
1641 | |
1642 | /* |
1643 | * double_rq_lock - safely lock two runqueues |
1644 | * |
1645 | * Note this does not disable interrupts like task_rq_lock, |
1646 | * you need to do so manually before calling. |
1647 | */ |
1648 | static void double_rq_lock(struct rq *rq1, struct rq *rq2) |
1649 | __acquires(rq1->lock) |
1650 | __acquires(rq2->lock) |
1651 | { |
1652 | BUG_ON(!irqs_disabled()); |
1653 | if (rq1 == rq2) { |
1654 | raw_spin_lock(&rq1->lock); |
1655 | __acquire(rq2->lock); /* Fake it out ;) */ |
1656 | } else { |
1657 | if (rq1 < rq2) { |
1658 | raw_spin_lock(&rq1->lock); |
1659 | raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING); |
1660 | } else { |
1661 | raw_spin_lock(&rq2->lock); |
1662 | raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING); |
1663 | } |
1664 | } |
1665 | } |
1666 | |
1667 | /* |
1668 | * double_rq_unlock - safely unlock two runqueues |
1669 | * |
1670 | * Note this does not restore interrupts like task_rq_unlock, |
1671 | * you need to do so manually after calling. |
1672 | */ |
1673 | static void double_rq_unlock(struct rq *rq1, struct rq *rq2) |
1674 | __releases(rq1->lock) |
1675 | __releases(rq2->lock) |
1676 | { |
1677 | raw_spin_unlock(&rq1->lock); |
1678 | if (rq1 != rq2) |
1679 | raw_spin_unlock(&rq2->lock); |
1680 | else |
1681 | __release(rq2->lock); |
1682 | } |
1683 | |
1684 | #else /* CONFIG_SMP */ |
1685 | |
1686 | /* |
1687 | * double_rq_lock - safely lock two runqueues |
1688 | * |
1689 | * Note this does not disable interrupts like task_rq_lock, |
1690 | * you need to do so manually before calling. |
1691 | */ |
1692 | static void double_rq_lock(struct rq *rq1, struct rq *rq2) |
1693 | __acquires(rq1->lock) |
1694 | __acquires(rq2->lock) |
1695 | { |
1696 | BUG_ON(!irqs_disabled()); |
1697 | BUG_ON(rq1 != rq2); |
1698 | raw_spin_lock(&rq1->lock); |
1699 | __acquire(rq2->lock); /* Fake it out ;) */ |
1700 | } |
1701 | |
1702 | /* |
1703 | * double_rq_unlock - safely unlock two runqueues |
1704 | * |
1705 | * Note this does not restore interrupts like task_rq_unlock, |
1706 | * you need to do so manually after calling. |
1707 | */ |
1708 | static void double_rq_unlock(struct rq *rq1, struct rq *rq2) |
1709 | __releases(rq1->lock) |
1710 | __releases(rq2->lock) |
1711 | { |
1712 | BUG_ON(rq1 != rq2); |
1713 | raw_spin_unlock(&rq1->lock); |
1714 | __release(rq2->lock); |
1715 | } |
1716 | |
1717 | #endif |
1718 | |
1719 | static void calc_load_account_idle(struct rq *this_rq); |
1720 | static void update_sysctl(void); |
1721 | static int get_update_sysctl_factor(void); |
1722 | static void update_cpu_load(struct rq *this_rq); |
1723 | |
1724 | static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu) |
1725 | { |
1726 | set_task_rq(p, cpu); |
1727 | #ifdef CONFIG_SMP |
1728 | /* |
1729 | * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be |
1730 | * successfuly executed on another CPU. We must ensure that updates of |
1731 | * per-task data have been completed by this moment. |
1732 | */ |
1733 | smp_wmb(); |
1734 | task_thread_info(p)->cpu = cpu; |
1735 | #endif |
1736 | } |
1737 | |
1738 | static const struct sched_class rt_sched_class; |
1739 | |
1740 | #define sched_class_highest (&stop_sched_class) |
1741 | #define for_each_class(class) \ |
1742 | for (class = sched_class_highest; class; class = class->next) |
1743 | |
1744 | #include "sched_stats.h" |
1745 | |
1746 | static void inc_nr_running(struct rq *rq) |
1747 | { |
1748 | rq->nr_running++; |
1749 | } |
1750 | |
1751 | static void dec_nr_running(struct rq *rq) |
1752 | { |
1753 | rq->nr_running--; |
1754 | } |
1755 | |
1756 | static void set_load_weight(struct task_struct *p) |
1757 | { |
1758 | /* |
1759 | * SCHED_IDLE tasks get minimal weight: |
1760 | */ |
1761 | if (p->policy == SCHED_IDLE) { |
1762 | p->se.load.weight = WEIGHT_IDLEPRIO; |
1763 | p->se.load.inv_weight = WMULT_IDLEPRIO; |
1764 | return; |
1765 | } |
1766 | |
1767 | p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO]; |
1768 | p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO]; |
1769 | } |
1770 | |
1771 | static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) |
1772 | { |
1773 | update_rq_clock(rq); |
1774 | sched_info_queued(p); |
1775 | p->sched_class->enqueue_task(rq, p, flags); |
1776 | p->se.on_rq = 1; |
1777 | } |
1778 | |
1779 | static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) |
1780 | { |
1781 | update_rq_clock(rq); |
1782 | sched_info_dequeued(p); |
1783 | p->sched_class->dequeue_task(rq, p, flags); |
1784 | p->se.on_rq = 0; |
1785 | } |
1786 | |
1787 | /* |
1788 | * activate_task - move a task to the runqueue. |
1789 | */ |
1790 | static void activate_task(struct rq *rq, struct task_struct *p, int flags) |
1791 | { |
1792 | if (task_contributes_to_load(p)) |
1793 | rq->nr_uninterruptible--; |
1794 | |
1795 | enqueue_task(rq, p, flags); |
1796 | inc_nr_running(rq); |
1797 | } |
1798 | |
1799 | /* |
1800 | * deactivate_task - remove a task from the runqueue. |
1801 | */ |
1802 | static void deactivate_task(struct rq *rq, struct task_struct *p, int flags) |
1803 | { |
1804 | if (task_contributes_to_load(p)) |
1805 | rq->nr_uninterruptible++; |
1806 | |
1807 | dequeue_task(rq, p, flags); |
1808 | dec_nr_running(rq); |
1809 | } |
1810 | |
1811 | #ifdef CONFIG_IRQ_TIME_ACCOUNTING |
1812 | |
1813 | /* |
1814 | * There are no locks covering percpu hardirq/softirq time. |
1815 | * They are only modified in account_system_vtime, on corresponding CPU |
1816 | * with interrupts disabled. So, writes are safe. |
1817 | * They are read and saved off onto struct rq in update_rq_clock(). |
1818 | * This may result in other CPU reading this CPU's irq time and can |
1819 | * race with irq/account_system_vtime on this CPU. We would either get old |
1820 | * or new value with a side effect of accounting a slice of irq time to wrong |
1821 | * task when irq is in progress while we read rq->clock. That is a worthy |
1822 | * compromise in place of having locks on each irq in account_system_time. |
1823 | */ |
1824 | static DEFINE_PER_CPU(u64, cpu_hardirq_time); |
1825 | static DEFINE_PER_CPU(u64, cpu_softirq_time); |
1826 | |
1827 | static DEFINE_PER_CPU(u64, irq_start_time); |
1828 | static int sched_clock_irqtime; |
1829 | |
1830 | void enable_sched_clock_irqtime(void) |
1831 | { |
1832 | sched_clock_irqtime = 1; |
1833 | } |
1834 | |
1835 | void disable_sched_clock_irqtime(void) |
1836 | { |
1837 | sched_clock_irqtime = 0; |
1838 | } |
1839 | |
1840 | #ifndef CONFIG_64BIT |
1841 | static DEFINE_PER_CPU(seqcount_t, irq_time_seq); |
1842 | |
1843 | static inline void irq_time_write_begin(void) |
1844 | { |
1845 | __this_cpu_inc(irq_time_seq.sequence); |
1846 | smp_wmb(); |
1847 | } |
1848 | |
1849 | static inline void irq_time_write_end(void) |
1850 | { |
1851 | smp_wmb(); |
1852 | __this_cpu_inc(irq_time_seq.sequence); |
1853 | } |
1854 | |
1855 | static inline u64 irq_time_read(int cpu) |
1856 | { |
1857 | u64 irq_time; |
1858 | unsigned seq; |
1859 | |
1860 | do { |
1861 | seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu)); |
1862 | irq_time = per_cpu(cpu_softirq_time, cpu) + |
1863 | per_cpu(cpu_hardirq_time, cpu); |
1864 | } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq)); |
1865 | |
1866 | return irq_time; |
1867 | } |
1868 | #else /* CONFIG_64BIT */ |
1869 | static inline void irq_time_write_begin(void) |
1870 | { |
1871 | } |
1872 | |
1873 | static inline void irq_time_write_end(void) |
1874 | { |
1875 | } |
1876 | |
1877 | static inline u64 irq_time_read(int cpu) |
1878 | { |
1879 | return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu); |
1880 | } |
1881 | #endif /* CONFIG_64BIT */ |
1882 | |
1883 | /* |
1884 | * Called before incrementing preempt_count on {soft,}irq_enter |
1885 | * and before decrementing preempt_count on {soft,}irq_exit. |
1886 | */ |
1887 | void account_system_vtime(struct task_struct *curr) |
1888 | { |
1889 | unsigned long flags; |
1890 | s64 delta; |
1891 | int cpu; |
1892 | |
1893 | if (!sched_clock_irqtime) |
1894 | return; |
1895 | |
1896 | local_irq_save(flags); |
1897 | |
1898 | cpu = smp_processor_id(); |
1899 | delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time); |
1900 | __this_cpu_add(irq_start_time, delta); |
1901 | |
1902 | irq_time_write_begin(); |
1903 | /* |
1904 | * We do not account for softirq time from ksoftirqd here. |
1905 | * We want to continue accounting softirq time to ksoftirqd thread |
1906 | * in that case, so as not to confuse scheduler with a special task |
1907 | * that do not consume any time, but still wants to run. |
1908 | */ |
1909 | if (hardirq_count()) |
1910 | __this_cpu_add(cpu_hardirq_time, delta); |
1911 | else if (in_serving_softirq() && curr != this_cpu_ksoftirqd()) |
1912 | __this_cpu_add(cpu_softirq_time, delta); |
1913 | |
1914 | irq_time_write_end(); |
1915 | local_irq_restore(flags); |
1916 | } |
1917 | EXPORT_SYMBOL_GPL(account_system_vtime); |
1918 | |
1919 | static void update_rq_clock_task(struct rq *rq, s64 delta) |
1920 | { |
1921 | s64 irq_delta; |
1922 | |
1923 | irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; |
1924 | |
1925 | /* |
1926 | * Since irq_time is only updated on {soft,}irq_exit, we might run into |
1927 | * this case when a previous update_rq_clock() happened inside a |
1928 | * {soft,}irq region. |
1929 | * |
1930 | * When this happens, we stop ->clock_task and only update the |
1931 | * prev_irq_time stamp to account for the part that fit, so that a next |
1932 | * update will consume the rest. This ensures ->clock_task is |
1933 | * monotonic. |
1934 | * |
1935 | * It does however cause some slight miss-attribution of {soft,}irq |
1936 | * time, a more accurate solution would be to update the irq_time using |
1937 | * the current rq->clock timestamp, except that would require using |
1938 | * atomic ops. |
1939 | */ |
1940 | if (irq_delta > delta) |
1941 | irq_delta = delta; |
1942 | |
1943 | rq->prev_irq_time += irq_delta; |
1944 | delta -= irq_delta; |
1945 | rq->clock_task += delta; |
1946 | |
1947 | if (irq_delta && sched_feat(NONIRQ_POWER)) |
1948 | sched_rt_avg_update(rq, irq_delta); |
1949 | } |
1950 | |
1951 | static int irqtime_account_hi_update(void) |
1952 | { |
1953 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
1954 | unsigned long flags; |
1955 | u64 latest_ns; |
1956 | int ret = 0; |
1957 | |
1958 | local_irq_save(flags); |
1959 | latest_ns = this_cpu_read(cpu_hardirq_time); |
1960 | if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq)) |
1961 | ret = 1; |
1962 | local_irq_restore(flags); |
1963 | return ret; |
1964 | } |
1965 | |
1966 | static int irqtime_account_si_update(void) |
1967 | { |
1968 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
1969 | unsigned long flags; |
1970 | u64 latest_ns; |
1971 | int ret = 0; |
1972 | |
1973 | local_irq_save(flags); |
1974 | latest_ns = this_cpu_read(cpu_softirq_time); |
1975 | if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq)) |
1976 | ret = 1; |
1977 | local_irq_restore(flags); |
1978 | return ret; |
1979 | } |
1980 | |
1981 | #else /* CONFIG_IRQ_TIME_ACCOUNTING */ |
1982 | |
1983 | #define sched_clock_irqtime (0) |
1984 | |
1985 | static void update_rq_clock_task(struct rq *rq, s64 delta) |
1986 | { |
1987 | rq->clock_task += delta; |
1988 | } |
1989 | |
1990 | #endif /* CONFIG_IRQ_TIME_ACCOUNTING */ |
1991 | |
1992 | #include "sched_idletask.c" |
1993 | #include "sched_fair.c" |
1994 | #include "sched_rt.c" |
1995 | #include "sched_autogroup.c" |
1996 | #include "sched_stoptask.c" |
1997 | #ifdef CONFIG_SCHED_DEBUG |
1998 | # include "sched_debug.c" |
1999 | #endif |
2000 | |
2001 | void sched_set_stop_task(int cpu, struct task_struct *stop) |
2002 | { |
2003 | struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; |
2004 | struct task_struct *old_stop = cpu_rq(cpu)->stop; |
2005 | |
2006 | if (stop) { |
2007 | /* |
2008 | * Make it appear like a SCHED_FIFO task, its something |
2009 | * userspace knows about and won't get confused about. |
2010 | * |
2011 | * Also, it will make PI more or less work without too |
2012 | * much confusion -- but then, stop work should not |
2013 | * rely on PI working anyway. |
2014 | */ |
2015 | sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); |
2016 | |
2017 | stop->sched_class = &stop_sched_class; |
2018 | } |
2019 | |
2020 | cpu_rq(cpu)->stop = stop; |
2021 | |
2022 | if (old_stop) { |
2023 | /* |
2024 | * Reset it back to a normal scheduling class so that |
2025 | * it can die in pieces. |
2026 | */ |
2027 | old_stop->sched_class = &rt_sched_class; |
2028 | } |
2029 | } |
2030 | |
2031 | /* |
2032 | * __normal_prio - return the priority that is based on the static prio |
2033 | */ |
2034 | static inline int __normal_prio(struct task_struct *p) |
2035 | { |
2036 | return p->static_prio; |
2037 | } |
2038 | |
2039 | /* |
2040 | * Calculate the expected normal priority: i.e. priority |
2041 | * without taking RT-inheritance into account. Might be |
2042 | * boosted by interactivity modifiers. Changes upon fork, |
2043 | * setprio syscalls, and whenever the interactivity |
2044 | * estimator recalculates. |
2045 | */ |
2046 | static inline int normal_prio(struct task_struct *p) |
2047 | { |
2048 | int prio; |
2049 | |
2050 | if (task_has_rt_policy(p)) |
2051 | prio = MAX_RT_PRIO-1 - p->rt_priority; |
2052 | else |
2053 | prio = __normal_prio(p); |
2054 | return prio; |
2055 | } |
2056 | |
2057 | /* |
2058 | * Calculate the current priority, i.e. the priority |
2059 | * taken into account by the scheduler. This value might |
2060 | * be boosted by RT tasks, or might be boosted by |
2061 | * interactivity modifiers. Will be RT if the task got |
2062 | * RT-boosted. If not then it returns p->normal_prio. |
2063 | */ |
2064 | static int effective_prio(struct task_struct *p) |
2065 | { |
2066 | p->normal_prio = normal_prio(p); |
2067 | /* |
2068 | * If we are RT tasks or we were boosted to RT priority, |
2069 | * keep the priority unchanged. Otherwise, update priority |
2070 | * to the normal priority: |
2071 | */ |
2072 | if (!rt_prio(p->prio)) |
2073 | return p->normal_prio; |
2074 | return p->prio; |
2075 | } |
2076 | |
2077 | /** |
2078 | * task_curr - is this task currently executing on a CPU? |
2079 | * @p: the task in question. |
2080 | */ |
2081 | inline int task_curr(const struct task_struct *p) |
2082 | { |
2083 | return cpu_curr(task_cpu(p)) == p; |
2084 | } |
2085 | |
2086 | static inline void check_class_changed(struct rq *rq, struct task_struct *p, |
2087 | const struct sched_class *prev_class, |
2088 | int oldprio) |
2089 | { |
2090 | if (prev_class != p->sched_class) { |
2091 | if (prev_class->switched_from) |
2092 | prev_class->switched_from(rq, p); |
2093 | p->sched_class->switched_to(rq, p); |
2094 | } else if (oldprio != p->prio) |
2095 | p->sched_class->prio_changed(rq, p, oldprio); |
2096 | } |
2097 | |
2098 | static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) |
2099 | { |
2100 | const struct sched_class *class; |
2101 | |
2102 | if (p->sched_class == rq->curr->sched_class) { |
2103 | rq->curr->sched_class->check_preempt_curr(rq, p, flags); |
2104 | } else { |
2105 | for_each_class(class) { |
2106 | if (class == rq->curr->sched_class) |
2107 | break; |
2108 | if (class == p->sched_class) { |
2109 | resched_task(rq->curr); |
2110 | break; |
2111 | } |
2112 | } |
2113 | } |
2114 | |
2115 | /* |
2116 | * A queue event has occurred, and we're going to schedule. In |
2117 | * this case, we can save a useless back to back clock update. |
2118 | */ |
2119 | if (rq->curr->se.on_rq && test_tsk_need_resched(rq->curr)) |
2120 | rq->skip_clock_update = 1; |
2121 | } |
2122 | |
2123 | #ifdef CONFIG_SMP |
2124 | /* |
2125 | * Is this task likely cache-hot: |
2126 | */ |
2127 | static int |
2128 | task_hot(struct task_struct *p, u64 now, struct sched_domain *sd) |
2129 | { |
2130 | s64 delta; |
2131 | |
2132 | if (p->sched_class != &fair_sched_class) |
2133 | return 0; |
2134 | |
2135 | if (unlikely(p->policy == SCHED_IDLE)) |
2136 | return 0; |
2137 | |
2138 | /* |
2139 | * Buddy candidates are cache hot: |
2140 | */ |
2141 | if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running && |
2142 | (&p->se == cfs_rq_of(&p->se)->next || |
2143 | &p->se == cfs_rq_of(&p->se)->last)) |
2144 | return 1; |
2145 | |
2146 | if (sysctl_sched_migration_cost == -1) |
2147 | return 1; |
2148 | if (sysctl_sched_migration_cost == 0) |
2149 | return 0; |
2150 | |
2151 | delta = now - p->se.exec_start; |
2152 | |
2153 | return delta < (s64)sysctl_sched_migration_cost; |
2154 | } |
2155 | |
2156 | void set_task_cpu(struct task_struct *p, unsigned int new_cpu) |
2157 | { |
2158 | #ifdef CONFIG_SCHED_DEBUG |
2159 | /* |
2160 | * We should never call set_task_cpu() on a blocked task, |
2161 | * ttwu() will sort out the placement. |
2162 | */ |
2163 | WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && |
2164 | !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE)); |
2165 | #endif |
2166 | |
2167 | trace_sched_migrate_task(p, new_cpu); |
2168 | |
2169 | if (task_cpu(p) != new_cpu) { |
2170 | p->se.nr_migrations++; |
2171 | perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0); |
2172 | } |
2173 | |
2174 | __set_task_cpu(p, new_cpu); |
2175 | } |
2176 | |
2177 | struct migration_arg { |
2178 | struct task_struct *task; |
2179 | int dest_cpu; |
2180 | }; |
2181 | |
2182 | static int migration_cpu_stop(void *data); |
2183 | |
2184 | /* |
2185 | * The task's runqueue lock must be held. |
2186 | * Returns true if you have to wait for migration thread. |
2187 | */ |
2188 | static bool migrate_task(struct task_struct *p, struct rq *rq) |
2189 | { |
2190 | /* |
2191 | * If the task is not on a runqueue (and not running), then |
2192 | * the next wake-up will properly place the task. |
2193 | */ |
2194 | return p->se.on_rq || task_running(rq, p); |
2195 | } |
2196 | |
2197 | /* |
2198 | * wait_task_inactive - wait for a thread to unschedule. |
2199 | * |
2200 | * If @match_state is nonzero, it's the @p->state value just checked and |
2201 | * not expected to change. If it changes, i.e. @p might have woken up, |
2202 | * then return zero. When we succeed in waiting for @p to be off its CPU, |
2203 | * we return a positive number (its total switch count). If a second call |
2204 | * a short while later returns the same number, the caller can be sure that |
2205 | * @p has remained unscheduled the whole time. |
2206 | * |
2207 | * The caller must ensure that the task *will* unschedule sometime soon, |
2208 | * else this function might spin for a *long* time. This function can't |
2209 | * be called with interrupts off, or it may introduce deadlock with |
2210 | * smp_call_function() if an IPI is sent by the same process we are |
2211 | * waiting to become inactive. |
2212 | */ |
2213 | unsigned long wait_task_inactive(struct task_struct *p, long match_state) |
2214 | { |
2215 | unsigned long flags; |
2216 | int running, on_rq; |
2217 | unsigned long ncsw; |
2218 | struct rq *rq; |
2219 | |
2220 | for (;;) { |
2221 | /* |
2222 | * We do the initial early heuristics without holding |
2223 | * any task-queue locks at all. We'll only try to get |
2224 | * the runqueue lock when things look like they will |
2225 | * work out! |
2226 | */ |
2227 | rq = task_rq(p); |
2228 | |
2229 | /* |
2230 | * If the task is actively running on another CPU |
2231 | * still, just relax and busy-wait without holding |
2232 | * any locks. |
2233 | * |
2234 | * NOTE! Since we don't hold any locks, it's not |
2235 | * even sure that "rq" stays as the right runqueue! |
2236 | * But we don't care, since "task_running()" will |
2237 | * return false if the runqueue has changed and p |
2238 | * is actually now running somewhere else! |
2239 | */ |
2240 | while (task_running(rq, p)) { |
2241 | if (match_state && unlikely(p->state != match_state)) |
2242 | return 0; |
2243 | cpu_relax(); |
2244 | } |
2245 | |
2246 | /* |
2247 | * Ok, time to look more closely! We need the rq |
2248 | * lock now, to be *sure*. If we're wrong, we'll |
2249 | * just go back and repeat. |
2250 | */ |
2251 | rq = task_rq_lock(p, &flags); |
2252 | trace_sched_wait_task(p); |
2253 | running = task_running(rq, p); |
2254 | on_rq = p->se.on_rq; |
2255 | ncsw = 0; |
2256 | if (!match_state || p->state == match_state) |
2257 | ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ |
2258 | task_rq_unlock(rq, &flags); |
2259 | |
2260 | /* |
2261 | * If it changed from the expected state, bail out now. |
2262 | */ |
2263 | if (unlikely(!ncsw)) |
2264 | break; |
2265 | |
2266 | /* |
2267 | * Was it really running after all now that we |
2268 | * checked with the proper locks actually held? |
2269 | * |
2270 | * Oops. Go back and try again.. |
2271 | */ |
2272 | if (unlikely(running)) { |
2273 | cpu_relax(); |
2274 | continue; |
2275 | } |
2276 | |
2277 | /* |
2278 | * It's not enough that it's not actively running, |
2279 | * it must be off the runqueue _entirely_, and not |
2280 | * preempted! |
2281 | * |
2282 | * So if it was still runnable (but just not actively |
2283 | * running right now), it's preempted, and we should |
2284 | * yield - it could be a while. |
2285 | */ |
2286 | if (unlikely(on_rq)) { |
2287 | ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); |
2288 | |
2289 | set_current_state(TASK_UNINTERRUPTIBLE); |
2290 | schedule_hrtimeout(&to, HRTIMER_MODE_REL); |
2291 | continue; |
2292 | } |
2293 | |
2294 | /* |
2295 | * Ahh, all good. It wasn't running, and it wasn't |
2296 | * runnable, which means that it will never become |
2297 | * running in the future either. We're all done! |
2298 | */ |
2299 | break; |
2300 | } |
2301 | |
2302 | return ncsw; |
2303 | } |
2304 | |
2305 | /*** |
2306 | * kick_process - kick a running thread to enter/exit the kernel |
2307 | * @p: the to-be-kicked thread |
2308 | * |
2309 | * Cause a process which is running on another CPU to enter |
2310 | * kernel-mode, without any delay. (to get signals handled.) |
2311 | * |
2312 | * NOTE: this function doesn't have to take the runqueue lock, |
2313 | * because all it wants to ensure is that the remote task enters |
2314 | * the kernel. If the IPI races and the task has been migrated |
2315 | * to another CPU then no harm is done and the purpose has been |
2316 | * achieved as well. |
2317 | */ |
2318 | void kick_process(struct task_struct *p) |
2319 | { |
2320 | int cpu; |
2321 | |
2322 | preempt_disable(); |
2323 | cpu = task_cpu(p); |
2324 | if ((cpu != smp_processor_id()) && task_curr(p)) |
2325 | smp_send_reschedule(cpu); |
2326 | preempt_enable(); |
2327 | } |
2328 | EXPORT_SYMBOL_GPL(kick_process); |
2329 | #endif /* CONFIG_SMP */ |
2330 | |
2331 | #ifdef CONFIG_SMP |
2332 | /* |
2333 | * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held. |
2334 | */ |
2335 | static int select_fallback_rq(int cpu, struct task_struct *p) |
2336 | { |
2337 | int dest_cpu; |
2338 | const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu)); |
2339 | |
2340 | /* Look for allowed, online CPU in same node. */ |
2341 | for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask) |
2342 | if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed)) |
2343 | return dest_cpu; |
2344 | |
2345 | /* Any allowed, online CPU? */ |
2346 | dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask); |
2347 | if (dest_cpu < nr_cpu_ids) |
2348 | return dest_cpu; |
2349 | |
2350 | /* No more Mr. Nice Guy. */ |
2351 | dest_cpu = cpuset_cpus_allowed_fallback(p); |
2352 | /* |
2353 | * Don't tell them about moving exiting tasks or |
2354 | * kernel threads (both mm NULL), since they never |
2355 | * leave kernel. |
2356 | */ |
2357 | if (p->mm && printk_ratelimit()) { |
2358 | printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n", |
2359 | task_pid_nr(p), p->comm, cpu); |
2360 | } |
2361 | |
2362 | return dest_cpu; |
2363 | } |
2364 | |
2365 | /* |
2366 | * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable. |
2367 | */ |
2368 | static inline |
2369 | int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags) |
2370 | { |
2371 | int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags); |
2372 | |
2373 | /* |
2374 | * In order not to call set_task_cpu() on a blocking task we need |
2375 | * to rely on ttwu() to place the task on a valid ->cpus_allowed |
2376 | * cpu. |
2377 | * |
2378 | * Since this is common to all placement strategies, this lives here. |
2379 | * |
2380 | * [ this allows ->select_task() to simply return task_cpu(p) and |
2381 | * not worry about this generic constraint ] |
2382 | */ |
2383 | if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) || |
2384 | !cpu_online(cpu))) |
2385 | cpu = select_fallback_rq(task_cpu(p), p); |
2386 | |
2387 | return cpu; |
2388 | } |
2389 | |
2390 | static void update_avg(u64 *avg, u64 sample) |
2391 | { |
2392 | s64 diff = sample - *avg; |
2393 | *avg += diff >> 3; |
2394 | } |
2395 | #endif |
2396 | |
2397 | static inline void ttwu_activate(struct task_struct *p, struct rq *rq, |
2398 | bool is_sync, bool is_migrate, bool is_local, |
2399 | unsigned long en_flags) |
2400 | { |
2401 | schedstat_inc(p, se.statistics.nr_wakeups); |
2402 | if (is_sync) |
2403 | schedstat_inc(p, se.statistics.nr_wakeups_sync); |
2404 | if (is_migrate) |
2405 | schedstat_inc(p, se.statistics.nr_wakeups_migrate); |
2406 | if (is_local) |
2407 | schedstat_inc(p, se.statistics.nr_wakeups_local); |
2408 | else |
2409 | schedstat_inc(p, se.statistics.nr_wakeups_remote); |
2410 | |
2411 | activate_task(rq, p, en_flags); |
2412 | } |
2413 | |
2414 | static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq, |
2415 | int wake_flags, bool success) |
2416 | { |
2417 | trace_sched_wakeup(p, success); |
2418 | check_preempt_curr(rq, p, wake_flags); |
2419 | |
2420 | p->state = TASK_RUNNING; |
2421 | #ifdef CONFIG_SMP |
2422 | if (p->sched_class->task_woken) |
2423 | p->sched_class->task_woken(rq, p); |
2424 | |
2425 | if (unlikely(rq->idle_stamp)) { |
2426 | u64 delta = rq->clock - rq->idle_stamp; |
2427 | u64 max = 2*sysctl_sched_migration_cost; |
2428 | |
2429 | if (delta > max) |
2430 | rq->avg_idle = max; |
2431 | else |
2432 | update_avg(&rq->avg_idle, delta); |
2433 | rq->idle_stamp = 0; |
2434 | } |
2435 | #endif |
2436 | /* if a worker is waking up, notify workqueue */ |
2437 | if ((p->flags & PF_WQ_WORKER) && success) |
2438 | wq_worker_waking_up(p, cpu_of(rq)); |
2439 | } |
2440 | |
2441 | /** |
2442 | * try_to_wake_up - wake up a thread |
2443 | * @p: the thread to be awakened |
2444 | * @state: the mask of task states that can be woken |
2445 | * @wake_flags: wake modifier flags (WF_*) |
2446 | * |
2447 | * Put it on the run-queue if it's not already there. The "current" |
2448 | * thread is always on the run-queue (except when the actual |
2449 | * re-schedule is in progress), and as such you're allowed to do |
2450 | * the simpler "current->state = TASK_RUNNING" to mark yourself |
2451 | * runnable without the overhead of this. |
2452 | * |
2453 | * Returns %true if @p was woken up, %false if it was already running |
2454 | * or @state didn't match @p's state. |
2455 | */ |
2456 | static int try_to_wake_up(struct task_struct *p, unsigned int state, |
2457 | int wake_flags) |
2458 | { |
2459 | int cpu, orig_cpu, this_cpu, success = 0; |
2460 | unsigned long flags; |
2461 | unsigned long en_flags = ENQUEUE_WAKEUP; |
2462 | struct rq *rq; |
2463 | |
2464 | this_cpu = get_cpu(); |
2465 | |
2466 | smp_wmb(); |
2467 | rq = task_rq_lock(p, &flags); |
2468 | if (!(p->state & state)) |
2469 | goto out; |
2470 | |
2471 | if (p->se.on_rq) |
2472 | goto out_running; |
2473 | |
2474 | cpu = task_cpu(p); |
2475 | orig_cpu = cpu; |
2476 | |
2477 | #ifdef CONFIG_SMP |
2478 | if (unlikely(task_running(rq, p))) |
2479 | goto out_activate; |
2480 | |
2481 | /* |
2482 | * In order to handle concurrent wakeups and release the rq->lock |
2483 | * we put the task in TASK_WAKING state. |
2484 | * |
2485 | * First fix up the nr_uninterruptible count: |
2486 | */ |
2487 | if (task_contributes_to_load(p)) { |
2488 | if (likely(cpu_online(orig_cpu))) |
2489 | rq->nr_uninterruptible--; |
2490 | else |
2491 | this_rq()->nr_uninterruptible--; |
2492 | } |
2493 | p->state = TASK_WAKING; |
2494 | |
2495 | if (p->sched_class->task_waking) { |
2496 | p->sched_class->task_waking(rq, p); |
2497 | en_flags |= ENQUEUE_WAKING; |
2498 | } |
2499 | |
2500 | cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags); |
2501 | if (cpu != orig_cpu) |
2502 | set_task_cpu(p, cpu); |
2503 | __task_rq_unlock(rq); |
2504 | |
2505 | rq = cpu_rq(cpu); |
2506 | raw_spin_lock(&rq->lock); |
2507 | |
2508 | /* |
2509 | * We migrated the task without holding either rq->lock, however |
2510 | * since the task is not on the task list itself, nobody else |
2511 | * will try and migrate the task, hence the rq should match the |
2512 | * cpu we just moved it to. |
2513 | */ |
2514 | WARN_ON(task_cpu(p) != cpu); |
2515 | WARN_ON(p->state != TASK_WAKING); |
2516 | |
2517 | #ifdef CONFIG_SCHEDSTATS |
2518 | schedstat_inc(rq, ttwu_count); |
2519 | if (cpu == this_cpu) |
2520 | schedstat_inc(rq, ttwu_local); |
2521 | else { |
2522 | struct sched_domain *sd; |
2523 | for_each_domain(this_cpu, sd) { |
2524 | if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { |
2525 | schedstat_inc(sd, ttwu_wake_remote); |
2526 | break; |
2527 | } |
2528 | } |
2529 | } |
2530 | #endif /* CONFIG_SCHEDSTATS */ |
2531 | |
2532 | out_activate: |
2533 | #endif /* CONFIG_SMP */ |
2534 | ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu, |
2535 | cpu == this_cpu, en_flags); |
2536 | success = 1; |
2537 | out_running: |
2538 | ttwu_post_activation(p, rq, wake_flags, success); |
2539 | out: |
2540 | task_rq_unlock(rq, &flags); |
2541 | put_cpu(); |
2542 | |
2543 | return success; |
2544 | } |
2545 | |
2546 | /** |
2547 | * try_to_wake_up_local - try to wake up a local task with rq lock held |
2548 | * @p: the thread to be awakened |
2549 | * |
2550 | * Put @p on the run-queue if it's not already there. The caller must |
2551 | * ensure that this_rq() is locked, @p is bound to this_rq() and not |
2552 | * the current task. this_rq() stays locked over invocation. |
2553 | */ |
2554 | static void try_to_wake_up_local(struct task_struct *p) |
2555 | { |
2556 | struct rq *rq = task_rq(p); |
2557 | bool success = false; |
2558 | |
2559 | BUG_ON(rq != this_rq()); |
2560 | BUG_ON(p == current); |
2561 | lockdep_assert_held(&rq->lock); |
2562 | |
2563 | if (!(p->state & TASK_NORMAL)) |
2564 | return; |
2565 | |
2566 | if (!p->se.on_rq) { |
2567 | if (likely(!task_running(rq, p))) { |
2568 | schedstat_inc(rq, ttwu_count); |
2569 | schedstat_inc(rq, ttwu_local); |
2570 | } |
2571 | ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP); |
2572 | success = true; |
2573 | } |
2574 | ttwu_post_activation(p, rq, 0, success); |
2575 | } |
2576 | |
2577 | /** |
2578 | * wake_up_process - Wake up a specific process |
2579 | * @p: The process to be woken up. |
2580 | * |
2581 | * Attempt to wake up the nominated process and move it to the set of runnable |
2582 | * processes. Returns 1 if the process was woken up, 0 if it was already |
2583 | * running. |
2584 | * |
2585 | * It may be assumed that this function implies a write memory barrier before |
2586 | * changing the task state if and only if any tasks are woken up. |
2587 | */ |
2588 | int wake_up_process(struct task_struct *p) |
2589 | { |
2590 | return try_to_wake_up(p, TASK_ALL, 0); |
2591 | } |
2592 | EXPORT_SYMBOL(wake_up_process); |
2593 | |
2594 | int wake_up_state(struct task_struct *p, unsigned int state) |
2595 | { |
2596 | return try_to_wake_up(p, state, 0); |
2597 | } |
2598 | |
2599 | /* |
2600 | * Perform scheduler related setup for a newly forked process p. |
2601 | * p is forked by current. |
2602 | * |
2603 | * __sched_fork() is basic setup used by init_idle() too: |
2604 | */ |
2605 | static void __sched_fork(struct task_struct *p) |
2606 | { |
2607 | p->se.exec_start = 0; |
2608 | p->se.sum_exec_runtime = 0; |
2609 | p->se.prev_sum_exec_runtime = 0; |
2610 | p->se.nr_migrations = 0; |
2611 | p->se.vruntime = 0; |
2612 | |
2613 | #ifdef CONFIG_SCHEDSTATS |
2614 | memset(&p->se.statistics, 0, sizeof(p->se.statistics)); |
2615 | #endif |
2616 | |
2617 | INIT_LIST_HEAD(&p->rt.run_list); |
2618 | p->se.on_rq = 0; |
2619 | INIT_LIST_HEAD(&p->se.group_node); |
2620 | |
2621 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
2622 | INIT_HLIST_HEAD(&p->preempt_notifiers); |
2623 | #endif |
2624 | } |
2625 | |
2626 | /* |
2627 | * fork()/clone()-time setup: |
2628 | */ |
2629 | void sched_fork(struct task_struct *p, int clone_flags) |
2630 | { |
2631 | int cpu = get_cpu(); |
2632 | |
2633 | __sched_fork(p); |
2634 | /* |
2635 | * We mark the process as running here. This guarantees that |
2636 | * nobody will actually run it, and a signal or other external |
2637 | * event cannot wake it up and insert it on the runqueue either. |
2638 | */ |
2639 | p->state = TASK_RUNNING; |
2640 | |
2641 | /* |
2642 | * Revert to default priority/policy on fork if requested. |
2643 | */ |
2644 | if (unlikely(p->sched_reset_on_fork)) { |
2645 | if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) { |
2646 | p->policy = SCHED_NORMAL; |
2647 | p->normal_prio = p->static_prio; |
2648 | } |
2649 | |
2650 | if (PRIO_TO_NICE(p->static_prio) < 0) { |
2651 | p->static_prio = NICE_TO_PRIO(0); |
2652 | p->normal_prio = p->static_prio; |
2653 | set_load_weight(p); |
2654 | } |
2655 | |
2656 | /* |
2657 | * We don't need the reset flag anymore after the fork. It has |
2658 | * fulfilled its duty: |
2659 | */ |
2660 | p->sched_reset_on_fork = 0; |
2661 | } |
2662 | |
2663 | /* |
2664 | * Make sure we do not leak PI boosting priority to the child. |
2665 | */ |
2666 | p->prio = current->normal_prio; |
2667 | |
2668 | if (!rt_prio(p->prio)) |
2669 | p->sched_class = &fair_sched_class; |
2670 | |
2671 | if (p->sched_class->task_fork) |
2672 | p->sched_class->task_fork(p); |
2673 | |
2674 | /* |
2675 | * The child is not yet in the pid-hash so no cgroup attach races, |
2676 | * and the cgroup is pinned to this child due to cgroup_fork() |
2677 | * is ran before sched_fork(). |
2678 | * |
2679 | * Silence PROVE_RCU. |
2680 | */ |
2681 | rcu_read_lock(); |
2682 | set_task_cpu(p, cpu); |
2683 | rcu_read_unlock(); |
2684 | |
2685 | #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) |
2686 | if (likely(sched_info_on())) |
2687 | memset(&p->sched_info, 0, sizeof(p->sched_info)); |
2688 | #endif |
2689 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
2690 | p->oncpu = 0; |
2691 | #endif |
2692 | #ifdef CONFIG_PREEMPT |
2693 | /* Want to start with kernel preemption disabled. */ |
2694 | task_thread_info(p)->preempt_count = 1; |
2695 | #endif |
2696 | #ifdef CONFIG_SMP |
2697 | plist_node_init(&p->pushable_tasks, MAX_PRIO); |
2698 | #endif |
2699 | |
2700 | put_cpu(); |
2701 | } |
2702 | |
2703 | /* |
2704 | * wake_up_new_task - wake up a newly created task for the first time. |
2705 | * |
2706 | * This function will do some initial scheduler statistics housekeeping |
2707 | * that must be done for every newly created context, then puts the task |
2708 | * on the runqueue and wakes it. |
2709 | */ |
2710 | void wake_up_new_task(struct task_struct *p, unsigned long clone_flags) |
2711 | { |
2712 | unsigned long flags; |
2713 | struct rq *rq; |
2714 | int cpu __maybe_unused = get_cpu(); |
2715 | |
2716 | #ifdef CONFIG_SMP |
2717 | rq = task_rq_lock(p, &flags); |
2718 | p->state = TASK_WAKING; |
2719 | |
2720 | /* |
2721 | * Fork balancing, do it here and not earlier because: |
2722 | * - cpus_allowed can change in the fork path |
2723 | * - any previously selected cpu might disappear through hotplug |
2724 | * |
2725 | * We set TASK_WAKING so that select_task_rq() can drop rq->lock |
2726 | * without people poking at ->cpus_allowed. |
2727 | */ |
2728 | cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0); |
2729 | set_task_cpu(p, cpu); |
2730 | |
2731 | p->state = TASK_RUNNING; |
2732 | task_rq_unlock(rq, &flags); |
2733 | #endif |
2734 | |
2735 | rq = task_rq_lock(p, &flags); |
2736 | activate_task(rq, p, 0); |
2737 | trace_sched_wakeup_new(p, 1); |
2738 | check_preempt_curr(rq, p, WF_FORK); |
2739 | #ifdef CONFIG_SMP |
2740 | if (p->sched_class->task_woken) |
2741 | p->sched_class->task_woken(rq, p); |
2742 | #endif |
2743 | task_rq_unlock(rq, &flags); |
2744 | put_cpu(); |
2745 | } |
2746 | |
2747 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
2748 | |
2749 | /** |
2750 | * preempt_notifier_register - tell me when current is being preempted & rescheduled |
2751 | * @notifier: notifier struct to register |
2752 | */ |
2753 | void preempt_notifier_register(struct preempt_notifier *notifier) |
2754 | { |
2755 | hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); |
2756 | } |
2757 | EXPORT_SYMBOL_GPL(preempt_notifier_register); |
2758 | |
2759 | /** |
2760 | * preempt_notifier_unregister - no longer interested in preemption notifications |
2761 | * @notifier: notifier struct to unregister |
2762 | * |
2763 | * This is safe to call from within a preemption notifier. |
2764 | */ |
2765 | void preempt_notifier_unregister(struct preempt_notifier *notifier) |
2766 | { |
2767 | hlist_del(¬ifier->link); |
2768 | } |
2769 | EXPORT_SYMBOL_GPL(preempt_notifier_unregister); |
2770 | |
2771 | static void fire_sched_in_preempt_notifiers(struct task_struct *curr) |
2772 | { |
2773 | struct preempt_notifier *notifier; |
2774 | struct hlist_node *node; |
2775 | |
2776 | hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) |
2777 | notifier->ops->sched_in(notifier, raw_smp_processor_id()); |
2778 | } |
2779 | |
2780 | static void |
2781 | fire_sched_out_preempt_notifiers(struct task_struct *curr, |
2782 | struct task_struct *next) |
2783 | { |
2784 | struct preempt_notifier *notifier; |
2785 | struct hlist_node *node; |
2786 | |
2787 | hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) |
2788 | notifier->ops->sched_out(notifier, next); |
2789 | } |
2790 | |
2791 | #else /* !CONFIG_PREEMPT_NOTIFIERS */ |
2792 | |
2793 | static void fire_sched_in_preempt_notifiers(struct task_struct *curr) |
2794 | { |
2795 | } |
2796 | |
2797 | static void |
2798 | fire_sched_out_preempt_notifiers(struct task_struct *curr, |
2799 | struct task_struct *next) |
2800 | { |
2801 | } |
2802 | |
2803 | #endif /* CONFIG_PREEMPT_NOTIFIERS */ |
2804 | |
2805 | /** |
2806 | * prepare_task_switch - prepare to switch tasks |
2807 | * @rq: the runqueue preparing to switch |
2808 | * @prev: the current task that is being switched out |
2809 | * @next: the task we are going to switch to. |
2810 | * |
2811 | * This is called with the rq lock held and interrupts off. It must |
2812 | * be paired with a subsequent finish_task_switch after the context |
2813 | * switch. |
2814 | * |
2815 | * prepare_task_switch sets up locking and calls architecture specific |
2816 | * hooks. |
2817 | */ |
2818 | static inline void |
2819 | prepare_task_switch(struct rq *rq, struct task_struct *prev, |
2820 | struct task_struct *next) |
2821 | { |
2822 | sched_info_switch(prev, next); |
2823 | perf_event_task_sched_out(prev, next); |
2824 | fire_sched_out_preempt_notifiers(prev, next); |
2825 | prepare_lock_switch(rq, next); |
2826 | prepare_arch_switch(next); |
2827 | trace_sched_switch(prev, next); |
2828 | } |
2829 | |
2830 | /** |
2831 | * finish_task_switch - clean up after a task-switch |
2832 | * @rq: runqueue associated with task-switch |
2833 | * @prev: the thread we just switched away from. |
2834 | * |
2835 | * finish_task_switch must be called after the context switch, paired |
2836 | * with a prepare_task_switch call before the context switch. |
2837 | * finish_task_switch will reconcile locking set up by prepare_task_switch, |
2838 | * and do any other architecture-specific cleanup actions. |
2839 | * |
2840 | * Note that we may have delayed dropping an mm in context_switch(). If |
2841 | * so, we finish that here outside of the runqueue lock. (Doing it |
2842 | * with the lock held can cause deadlocks; see schedule() for |
2843 | * details.) |
2844 | */ |
2845 | static void finish_task_switch(struct rq *rq, struct task_struct *prev) |
2846 | __releases(rq->lock) |
2847 | { |
2848 | struct mm_struct *mm = rq->prev_mm; |
2849 | long prev_state; |
2850 | |
2851 | rq->prev_mm = NULL; |
2852 | |
2853 | /* |
2854 | * A task struct has one reference for the use as "current". |
2855 | * If a task dies, then it sets TASK_DEAD in tsk->state and calls |
2856 | * schedule one last time. The schedule call will never return, and |
2857 | * the scheduled task must drop that reference. |
2858 | * The test for TASK_DEAD must occur while the runqueue locks are |
2859 | * still held, otherwise prev could be scheduled on another cpu, die |
2860 | * there before we look at prev->state, and then the reference would |
2861 | * be dropped twice. |
2862 | * Manfred Spraul <manfred@colorfullife.com> |
2863 | */ |
2864 | prev_state = prev->state; |
2865 | finish_arch_switch(prev); |
2866 | #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW |
2867 | local_irq_disable(); |
2868 | #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */ |
2869 | perf_event_task_sched_in(current); |
2870 | #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW |
2871 | local_irq_enable(); |
2872 | #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */ |
2873 | finish_lock_switch(rq, prev); |
2874 | |
2875 | fire_sched_in_preempt_notifiers(current); |
2876 | if (mm) |
2877 | mmdrop(mm); |
2878 | if (unlikely(prev_state == TASK_DEAD)) { |
2879 | /* |
2880 | * Remove function-return probe instances associated with this |
2881 | * task and put them back on the free list. |
2882 | */ |
2883 | kprobe_flush_task(prev); |
2884 | put_task_struct(prev); |
2885 | } |
2886 | } |
2887 | |
2888 | #ifdef CONFIG_SMP |
2889 | |
2890 | /* assumes rq->lock is held */ |
2891 | static inline void pre_schedule(struct rq *rq, struct task_struct *prev) |
2892 | { |
2893 | if (prev->sched_class->pre_schedule) |
2894 | prev->sched_class->pre_schedule(rq, prev); |
2895 | } |
2896 | |
2897 | /* rq->lock is NOT held, but preemption is disabled */ |
2898 | static inline void post_schedule(struct rq *rq) |
2899 | { |
2900 | if (rq->post_schedule) { |
2901 | unsigned long flags; |
2902 | |
2903 | raw_spin_lock_irqsave(&rq->lock, flags); |
2904 | if (rq->curr->sched_class->post_schedule) |
2905 | rq->curr->sched_class->post_schedule(rq); |
2906 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
2907 | |
2908 | rq->post_schedule = 0; |
2909 | } |
2910 | } |
2911 | |
2912 | #else |
2913 | |
2914 | static inline void pre_schedule(struct rq *rq, struct task_struct *p) |
2915 | { |
2916 | } |
2917 | |
2918 | static inline void post_schedule(struct rq *rq) |
2919 | { |
2920 | } |
2921 | |
2922 | #endif |
2923 | |
2924 | /** |
2925 | * schedule_tail - first thing a freshly forked thread must call. |
2926 | * @prev: the thread we just switched away from. |
2927 | */ |
2928 | asmlinkage void schedule_tail(struct task_struct *prev) |
2929 | __releases(rq->lock) |
2930 | { |
2931 | struct rq *rq = this_rq(); |
2932 | |
2933 | finish_task_switch(rq, prev); |
2934 | |
2935 | /* |
2936 | * FIXME: do we need to worry about rq being invalidated by the |
2937 | * task_switch? |
2938 | */ |
2939 | post_schedule(rq); |
2940 | |
2941 | #ifdef __ARCH_WANT_UNLOCKED_CTXSW |
2942 | /* In this case, finish_task_switch does not reenable preemption */ |
2943 | preempt_enable(); |
2944 | #endif |
2945 | if (current->set_child_tid) |
2946 | put_user(task_pid_vnr(current), current->set_child_tid); |
2947 | } |
2948 | |
2949 | /* |
2950 | * context_switch - switch to the new MM and the new |
2951 | * thread's register state. |
2952 | */ |
2953 | static inline void |
2954 | context_switch(struct rq *rq, struct task_struct *prev, |
2955 | struct task_struct *next) |
2956 | { |
2957 | struct mm_struct *mm, *oldmm; |
2958 | |
2959 | prepare_task_switch(rq, prev, next); |
2960 | |
2961 | mm = next->mm; |
2962 | oldmm = prev->active_mm; |
2963 | /* |
2964 | * For paravirt, this is coupled with an exit in switch_to to |
2965 | * combine the page table reload and the switch backend into |
2966 | * one hypercall. |
2967 | */ |
2968 | arch_start_context_switch(prev); |
2969 | |
2970 | if (!mm) { |
2971 | next->active_mm = oldmm; |
2972 | atomic_inc(&oldmm->mm_count); |
2973 | enter_lazy_tlb(oldmm, next); |
2974 | } else |
2975 | switch_mm(oldmm, mm, next); |
2976 | |
2977 | if (!prev->mm) { |
2978 | prev->active_mm = NULL; |
2979 | rq->prev_mm = oldmm; |
2980 | } |
2981 | /* |
2982 | * Since the runqueue lock will be released by the next |
2983 | * task (which is an invalid locking op but in the case |
2984 | * of the scheduler it's an obvious special-case), so we |
2985 | * do an early lockdep release here: |
2986 | */ |
2987 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW |
2988 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); |
2989 | #endif |
2990 | |
2991 | /* Here we just switch the register state and the stack. */ |
2992 | switch_to(prev, next, prev); |
2993 | |
2994 | barrier(); |
2995 | /* |
2996 | * this_rq must be evaluated again because prev may have moved |
2997 | * CPUs since it called schedule(), thus the 'rq' on its stack |
2998 | * frame will be invalid. |
2999 | */ |
3000 | finish_task_switch(this_rq(), prev); |
3001 | } |
3002 | |
3003 | /* |
3004 | * nr_running, nr_uninterruptible and nr_context_switches: |
3005 | * |
3006 | * externally visible scheduler statistics: current number of runnable |
3007 | * threads, current number of uninterruptible-sleeping threads, total |
3008 | * number of context switches performed since bootup. |
3009 | */ |
3010 | unsigned long nr_running(void) |
3011 | { |
3012 | unsigned long i, sum = 0; |
3013 | |
3014 | for_each_online_cpu(i) |
3015 | sum += cpu_rq(i)->nr_running; |
3016 | |
3017 | return sum; |
3018 | } |
3019 | |
3020 | unsigned long nr_uninterruptible(void) |
3021 | { |
3022 | unsigned long i, sum = 0; |
3023 | |
3024 | for_each_possible_cpu(i) |
3025 | sum += cpu_rq(i)->nr_uninterruptible; |
3026 | |
3027 | /* |
3028 | * Since we read the counters lockless, it might be slightly |
3029 | * inaccurate. Do not allow it to go below zero though: |
3030 | */ |
3031 | if (unlikely((long)sum < 0)) |
3032 | sum = 0; |
3033 | |
3034 | return sum; |
3035 | } |
3036 | |
3037 | unsigned long long nr_context_switches(void) |
3038 | { |
3039 | int i; |
3040 | unsigned long long sum = 0; |
3041 | |
3042 | for_each_possible_cpu(i) |
3043 | sum += cpu_rq(i)->nr_switches; |
3044 | |
3045 | return sum; |
3046 | } |
3047 | |
3048 | unsigned long nr_iowait(void) |
3049 | { |
3050 | unsigned long i, sum = 0; |
3051 | |
3052 | for_each_possible_cpu(i) |
3053 | sum += atomic_read(&cpu_rq(i)->nr_iowait); |
3054 | |
3055 | return sum; |
3056 | } |
3057 | |
3058 | unsigned long nr_iowait_cpu(int cpu) |
3059 | { |
3060 | struct rq *this = cpu_rq(cpu); |
3061 | return atomic_read(&this->nr_iowait); |
3062 | } |
3063 | |
3064 | unsigned long this_cpu_load(void) |
3065 | { |
3066 | struct rq *this = this_rq(); |
3067 | return this->cpu_load[0]; |
3068 | } |
3069 | |
3070 | |
3071 | /* Variables and functions for calc_load */ |
3072 | static atomic_long_t calc_load_tasks; |
3073 | static unsigned long calc_load_update; |
3074 | unsigned long avenrun[3]; |
3075 | EXPORT_SYMBOL(avenrun); |
3076 | |
3077 | static long calc_load_fold_active(struct rq *this_rq) |
3078 | { |
3079 | long nr_active, delta = 0; |
3080 | |
3081 | nr_active = this_rq->nr_running; |
3082 | nr_active += (long) this_rq->nr_uninterruptible; |
3083 | |
3084 | if (nr_active != this_rq->calc_load_active) { |
3085 | delta = nr_active - this_rq->calc_load_active; |
3086 | this_rq->calc_load_active = nr_active; |
3087 | } |
3088 | |
3089 | return delta; |
3090 | } |
3091 | |
3092 | static unsigned long |
3093 | calc_load(unsigned long load, unsigned long exp, unsigned long active) |
3094 | { |
3095 | load *= exp; |
3096 | load += active * (FIXED_1 - exp); |
3097 | load += 1UL << (FSHIFT - 1); |
3098 | return load >> FSHIFT; |
3099 | } |
3100 | |
3101 | #ifdef CONFIG_NO_HZ |
3102 | /* |
3103 | * For NO_HZ we delay the active fold to the next LOAD_FREQ update. |
3104 | * |
3105 | * When making the ILB scale, we should try to pull this in as well. |
3106 | */ |
3107 | static atomic_long_t calc_load_tasks_idle; |
3108 | |
3109 | static void calc_load_account_idle(struct rq *this_rq) |
3110 | { |
3111 | long delta; |
3112 | |
3113 | delta = calc_load_fold_active(this_rq); |
3114 | if (delta) |
3115 | atomic_long_add(delta, &calc_load_tasks_idle); |
3116 | } |
3117 | |
3118 | static long calc_load_fold_idle(void) |
3119 | { |
3120 | long delta = 0; |
3121 | |
3122 | /* |
3123 | * Its got a race, we don't care... |
3124 | */ |
3125 | if (atomic_long_read(&calc_load_tasks_idle)) |
3126 | delta = atomic_long_xchg(&calc_load_tasks_idle, 0); |
3127 | |
3128 | return delta; |
3129 | } |
3130 | |
3131 | /** |
3132 | * fixed_power_int - compute: x^n, in O(log n) time |
3133 | * |
3134 | * @x: base of the power |
3135 | * @frac_bits: fractional bits of @x |
3136 | * @n: power to raise @x to. |
3137 | * |
3138 | * By exploiting the relation between the definition of the natural power |
3139 | * function: x^n := x*x*...*x (x multiplied by itself for n times), and |
3140 | * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i, |
3141 | * (where: n_i \elem {0, 1}, the binary vector representing n), |
3142 | * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is |
3143 | * of course trivially computable in O(log_2 n), the length of our binary |
3144 | * vector. |
3145 | */ |
3146 | static unsigned long |
3147 | fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n) |
3148 | { |
3149 | unsigned long result = 1UL << frac_bits; |
3150 | |
3151 | if (n) for (;;) { |
3152 | if (n & 1) { |
3153 | result *= x; |
3154 | result += 1UL << (frac_bits - 1); |
3155 | result >>= frac_bits; |
3156 | } |
3157 | n >>= 1; |
3158 | if (!n) |
3159 | break; |
3160 | x *= x; |
3161 | x += 1UL << (frac_bits - 1); |
3162 | x >>= frac_bits; |
3163 | } |
3164 | |
3165 | return result; |
3166 | } |
3167 | |
3168 | /* |
3169 | * a1 = a0 * e + a * (1 - e) |
3170 | * |
3171 | * a2 = a1 * e + a * (1 - e) |
3172 | * = (a0 * e + a * (1 - e)) * e + a * (1 - e) |
3173 | * = a0 * e^2 + a * (1 - e) * (1 + e) |
3174 | * |
3175 | * a3 = a2 * e + a * (1 - e) |
3176 | * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e) |
3177 | * = a0 * e^3 + a * (1 - e) * (1 + e + e^2) |
3178 | * |
3179 | * ... |
3180 | * |
3181 | * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1] |
3182 | * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e) |
3183 | * = a0 * e^n + a * (1 - e^n) |
3184 | * |
3185 | * [1] application of the geometric series: |
3186 | * |
3187 | * n 1 - x^(n+1) |
3188 | * S_n := \Sum x^i = ------------- |
3189 | * i=0 1 - x |
3190 | */ |
3191 | static unsigned long |
3192 | calc_load_n(unsigned long load, unsigned long exp, |
3193 | unsigned long active, unsigned int n) |
3194 | { |
3195 | |
3196 | return calc_load(load, fixed_power_int(exp, FSHIFT, n), active); |
3197 | } |
3198 | |
3199 | /* |
3200 | * NO_HZ can leave us missing all per-cpu ticks calling |
3201 | * calc_load_account_active(), but since an idle CPU folds its delta into |
3202 | * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold |
3203 | * in the pending idle delta if our idle period crossed a load cycle boundary. |
3204 | * |
3205 | * Once we've updated the global active value, we need to apply the exponential |
3206 | * weights adjusted to the number of cycles missed. |
3207 | */ |
3208 | static void calc_global_nohz(unsigned long ticks) |
3209 | { |
3210 | long delta, active, n; |
3211 | |
3212 | if (time_before(jiffies, calc_load_update)) |
3213 | return; |
3214 | |
3215 | /* |
3216 | * If we crossed a calc_load_update boundary, make sure to fold |
3217 | * any pending idle changes, the respective CPUs might have |
3218 | * missed the tick driven calc_load_account_active() update |
3219 | * due to NO_HZ. |
3220 | */ |
3221 | delta = calc_load_fold_idle(); |
3222 | if (delta) |
3223 | atomic_long_add(delta, &calc_load_tasks); |
3224 | |
3225 | /* |
3226 | * If we were idle for multiple load cycles, apply them. |
3227 | */ |
3228 | if (ticks >= LOAD_FREQ) { |
3229 | n = ticks / LOAD_FREQ; |
3230 | |
3231 | active = atomic_long_read(&calc_load_tasks); |
3232 | active = active > 0 ? active * FIXED_1 : 0; |
3233 | |
3234 | avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n); |
3235 | avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n); |
3236 | avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n); |
3237 | |
3238 | calc_load_update += n * LOAD_FREQ; |
3239 | } |
3240 | |
3241 | /* |
3242 | * Its possible the remainder of the above division also crosses |
3243 | * a LOAD_FREQ period, the regular check in calc_global_load() |
3244 | * which comes after this will take care of that. |
3245 | * |
3246 | * Consider us being 11 ticks before a cycle completion, and us |
3247 | * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will |
3248 | * age us 4 cycles, and the test in calc_global_load() will |
3249 | * pick up the final one. |
3250 | */ |
3251 | } |
3252 | #else |
3253 | static void calc_load_account_idle(struct rq *this_rq) |
3254 | { |
3255 | } |
3256 | |
3257 | static inline long calc_load_fold_idle(void) |
3258 | { |
3259 | return 0; |
3260 | } |
3261 | |
3262 | static void calc_global_nohz(unsigned long ticks) |
3263 | { |
3264 | } |
3265 | #endif |
3266 | |
3267 | /** |
3268 | * get_avenrun - get the load average array |
3269 | * @loads: pointer to dest load array |
3270 | * @offset: offset to add |
3271 | * @shift: shift count to shift the result left |
3272 | * |
3273 | * These values are estimates at best, so no need for locking. |
3274 | */ |
3275 | void get_avenrun(unsigned long *loads, unsigned long offset, int shift) |
3276 | { |
3277 | loads[0] = (avenrun[0] + offset) << shift; |
3278 | loads[1] = (avenrun[1] + offset) << shift; |
3279 | loads[2] = (avenrun[2] + offset) << shift; |
3280 | } |
3281 | |
3282 | /* |
3283 | * calc_load - update the avenrun load estimates 10 ticks after the |
3284 | * CPUs have updated calc_load_tasks. |
3285 | */ |
3286 | void calc_global_load(unsigned long ticks) |
3287 | { |
3288 | long active; |
3289 | |
3290 | calc_global_nohz(ticks); |
3291 | |
3292 | if (time_before(jiffies, calc_load_update + 10)) |
3293 | return; |
3294 | |
3295 | active = atomic_long_read(&calc_load_tasks); |
3296 | active = active > 0 ? active * FIXED_1 : 0; |
3297 | |
3298 | avenrun[0] = calc_load(avenrun[0], EXP_1, active); |
3299 | avenrun[1] = calc_load(avenrun[1], EXP_5, active); |
3300 | avenrun[2] = calc_load(avenrun[2], EXP_15, active); |
3301 | |
3302 | calc_load_update += LOAD_FREQ; |
3303 | } |
3304 | |
3305 | /* |
3306 | * Called from update_cpu_load() to periodically update this CPU's |
3307 | * active count. |
3308 | */ |
3309 | static void calc_load_account_active(struct rq *this_rq) |
3310 | { |
3311 | long delta; |
3312 | |
3313 | if (time_before(jiffies, this_rq->calc_load_update)) |
3314 | return; |
3315 | |
3316 | delta = calc_load_fold_active(this_rq); |
3317 | delta += calc_load_fold_idle(); |
3318 | if (delta) |
3319 | atomic_long_add(delta, &calc_load_tasks); |
3320 | |
3321 | this_rq->calc_load_update += LOAD_FREQ; |
3322 | } |
3323 | |
3324 | /* |
3325 | * The exact cpuload at various idx values, calculated at every tick would be |
3326 | * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load |
3327 | * |
3328 | * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called |
3329 | * on nth tick when cpu may be busy, then we have: |
3330 | * load = ((2^idx - 1) / 2^idx)^(n-1) * load |
3331 | * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load |
3332 | * |
3333 | * decay_load_missed() below does efficient calculation of |
3334 | * load = ((2^idx - 1) / 2^idx)^(n-1) * load |
3335 | * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load |
3336 | * |
3337 | * The calculation is approximated on a 128 point scale. |
3338 | * degrade_zero_ticks is the number of ticks after which load at any |
3339 | * particular idx is approximated to be zero. |
3340 | * degrade_factor is a precomputed table, a row for each load idx. |
3341 | * Each column corresponds to degradation factor for a power of two ticks, |
3342 | * based on 128 point scale. |
3343 | * Example: |
3344 | * row 2, col 3 (=12) says that the degradation at load idx 2 after |
3345 | * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8). |
3346 | * |
3347 | * With this power of 2 load factors, we can degrade the load n times |
3348 | * by looking at 1 bits in n and doing as many mult/shift instead of |
3349 | * n mult/shifts needed by the exact degradation. |
3350 | */ |
3351 | #define DEGRADE_SHIFT 7 |
3352 | static const unsigned char |
3353 | degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128}; |
3354 | static const unsigned char |
3355 | degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = { |
3356 | {0, 0, 0, 0, 0, 0, 0, 0}, |
3357 | {64, 32, 8, 0, 0, 0, 0, 0}, |
3358 | {96, 72, 40, 12, 1, 0, 0}, |
3359 | {112, 98, 75, 43, 15, 1, 0}, |
3360 | {120, 112, 98, 76, 45, 16, 2} }; |
3361 | |
3362 | /* |
3363 | * Update cpu_load for any missed ticks, due to tickless idle. The backlog |
3364 | * would be when CPU is idle and so we just decay the old load without |
3365 | * adding any new load. |
3366 | */ |
3367 | static unsigned long |
3368 | decay_load_missed(unsigned long load, unsigned long missed_updates, int idx) |
3369 | { |
3370 | int j = 0; |
3371 | |
3372 | if (!missed_updates) |
3373 | return load; |
3374 | |
3375 | if (missed_updates >= degrade_zero_ticks[idx]) |
3376 | return 0; |
3377 | |
3378 | if (idx == 1) |
3379 | return load >> missed_updates; |
3380 | |
3381 | while (missed_updates) { |
3382 | if (missed_updates % 2) |
3383 | load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT; |
3384 | |
3385 | missed_updates >>= 1; |
3386 | j++; |
3387 | } |
3388 | return load; |
3389 | } |
3390 | |
3391 | /* |
3392 | * Update rq->cpu_load[] statistics. This function is usually called every |
3393 | * scheduler tick (TICK_NSEC). With tickless idle this will not be called |
3394 | * every tick. We fix it up based on jiffies. |
3395 | */ |
3396 | static void update_cpu_load(struct rq *this_rq) |
3397 | { |
3398 | unsigned long this_load = this_rq->load.weight; |
3399 | unsigned long curr_jiffies = jiffies; |
3400 | unsigned long pending_updates; |
3401 | int i, scale; |
3402 | |
3403 | this_rq->nr_load_updates++; |
3404 | |
3405 | /* Avoid repeated calls on same jiffy, when moving in and out of idle */ |
3406 | if (curr_jiffies == this_rq->last_load_update_tick) |
3407 | return; |
3408 | |
3409 | pending_updates = curr_jiffies - this_rq->last_load_update_tick; |
3410 | this_rq->last_load_update_tick = curr_jiffies; |
3411 | |
3412 | /* Update our load: */ |
3413 | this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */ |
3414 | for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { |
3415 | unsigned long old_load, new_load; |
3416 | |
3417 | /* scale is effectively 1 << i now, and >> i divides by scale */ |
3418 | |
3419 | old_load = this_rq->cpu_load[i]; |
3420 | old_load = decay_load_missed(old_load, pending_updates - 1, i); |
3421 | new_load = this_load; |
3422 | /* |
3423 | * Round up the averaging division if load is increasing. This |
3424 | * prevents us from getting stuck on 9 if the load is 10, for |
3425 | * example. |
3426 | */ |
3427 | if (new_load > old_load) |
3428 | new_load += scale - 1; |
3429 | |
3430 | this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i; |
3431 | } |
3432 | |
3433 | sched_avg_update(this_rq); |
3434 | } |
3435 | |
3436 | static void update_cpu_load_active(struct rq *this_rq) |
3437 | { |
3438 | update_cpu_load(this_rq); |
3439 | |
3440 | calc_load_account_active(this_rq); |
3441 | } |
3442 | |
3443 | #ifdef CONFIG_SMP |
3444 | |
3445 | /* |
3446 | * sched_exec - execve() is a valuable balancing opportunity, because at |
3447 | * this point the task has the smallest effective memory and cache footprint. |
3448 | */ |
3449 | void sched_exec(void) |
3450 | { |
3451 | struct task_struct *p = current; |
3452 | unsigned long flags; |
3453 | struct rq *rq; |
3454 | int dest_cpu; |
3455 | |
3456 | rq = task_rq_lock(p, &flags); |
3457 | dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0); |
3458 | if (dest_cpu == smp_processor_id()) |
3459 | goto unlock; |
3460 | |
3461 | /* |
3462 | * select_task_rq() can race against ->cpus_allowed |
3463 | */ |
3464 | if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) && |
3465 | likely(cpu_active(dest_cpu)) && migrate_task(p, rq)) { |
3466 | struct migration_arg arg = { p, dest_cpu }; |
3467 | |
3468 | task_rq_unlock(rq, &flags); |
3469 | stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); |
3470 | return; |
3471 | } |
3472 | unlock: |
3473 | task_rq_unlock(rq, &flags); |
3474 | } |
3475 | |
3476 | #endif |
3477 | |
3478 | DEFINE_PER_CPU(struct kernel_stat, kstat); |
3479 | |
3480 | EXPORT_PER_CPU_SYMBOL(kstat); |
3481 | |
3482 | /* |
3483 | * Return any ns on the sched_clock that have not yet been accounted in |
3484 | * @p in case that task is currently running. |
3485 | * |
3486 | * Called with task_rq_lock() held on @rq. |
3487 | */ |
3488 | static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) |
3489 | { |
3490 | u64 ns = 0; |
3491 | |
3492 | if (task_current(rq, p)) { |
3493 | update_rq_clock(rq); |
3494 | ns = rq->clock_task - p->se.exec_start; |
3495 | if ((s64)ns < 0) |
3496 | ns = 0; |
3497 | } |
3498 | |
3499 | return ns; |
3500 | } |
3501 | |
3502 | unsigned long long task_delta_exec(struct task_struct *p) |
3503 | { |
3504 | unsigned long flags; |
3505 | struct rq *rq; |
3506 | u64 ns = 0; |
3507 | |
3508 | rq = task_rq_lock(p, &flags); |
3509 | ns = do_task_delta_exec(p, rq); |
3510 | task_rq_unlock(rq, &flags); |
3511 | |
3512 | return ns; |
3513 | } |
3514 | |
3515 | /* |
3516 | * Return accounted runtime for the task. |
3517 | * In case the task is currently running, return the runtime plus current's |
3518 | * pending runtime that have not been accounted yet. |
3519 | */ |
3520 | unsigned long long task_sched_runtime(struct task_struct *p) |
3521 | { |
3522 | unsigned long flags; |
3523 | struct rq *rq; |
3524 | u64 ns = 0; |
3525 | |
3526 | rq = task_rq_lock(p, &flags); |
3527 | ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); |
3528 | task_rq_unlock(rq, &flags); |
3529 | |
3530 | return ns; |
3531 | } |
3532 | |
3533 | /* |
3534 | * Return sum_exec_runtime for the thread group. |
3535 | * In case the task is currently running, return the sum plus current's |
3536 | * pending runtime that have not been accounted yet. |
3537 | * |
3538 | * Note that the thread group might have other running tasks as well, |
3539 | * so the return value not includes other pending runtime that other |
3540 | * running tasks might have. |
3541 | */ |
3542 | unsigned long long thread_group_sched_runtime(struct task_struct *p) |
3543 | { |
3544 | struct task_cputime totals; |
3545 | unsigned long flags; |
3546 | struct rq *rq; |
3547 | u64 ns; |
3548 | |
3549 | rq = task_rq_lock(p, &flags); |
3550 | thread_group_cputime(p, &totals); |
3551 | ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq); |
3552 | task_rq_unlock(rq, &flags); |
3553 | |
3554 | return ns; |
3555 | } |
3556 | |
3557 | /* |
3558 | * Account user cpu time to a process. |
3559 | * @p: the process that the cpu time gets accounted to |
3560 | * @cputime: the cpu time spent in user space since the last update |
3561 | * @cputime_scaled: cputime scaled by cpu frequency |
3562 | */ |
3563 | void account_user_time(struct task_struct *p, cputime_t cputime, |
3564 | cputime_t cputime_scaled) |
3565 | { |
3566 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
3567 | cputime64_t tmp; |
3568 | |
3569 | /* Add user time to process. */ |
3570 | p->utime = cputime_add(p->utime, cputime); |
3571 | p->utimescaled = cputime_add(p->utimescaled, cputime_scaled); |
3572 | account_group_user_time(p, cputime); |
3573 | |
3574 | /* Add user time to cpustat. */ |
3575 | tmp = cputime_to_cputime64(cputime); |
3576 | if (TASK_NICE(p) > 0) |
3577 | cpustat->nice = cputime64_add(cpustat->nice, tmp); |
3578 | else |
3579 | cpustat->user = cputime64_add(cpustat->user, tmp); |
3580 | |
3581 | cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime); |
3582 | /* Account for user time used */ |
3583 | acct_update_integrals(p); |
3584 | } |
3585 | |
3586 | /* |
3587 | * Account guest cpu time to a process. |
3588 | * @p: the process that the cpu time gets accounted to |
3589 | * @cputime: the cpu time spent in virtual machine since the last update |
3590 | * @cputime_scaled: cputime scaled by cpu frequency |
3591 | */ |
3592 | static void account_guest_time(struct task_struct *p, cputime_t cputime, |
3593 | cputime_t cputime_scaled) |
3594 | { |
3595 | cputime64_t tmp; |
3596 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
3597 | |
3598 | tmp = cputime_to_cputime64(cputime); |
3599 | |
3600 | /* Add guest time to process. */ |
3601 | p->utime = cputime_add(p->utime, cputime); |
3602 | p->utimescaled = cputime_add(p->utimescaled, cputime_scaled); |
3603 | account_group_user_time(p, cputime); |
3604 | p->gtime = cputime_add(p->gtime, cputime); |
3605 | |
3606 | /* Add guest time to cpustat. */ |
3607 | if (TASK_NICE(p) > 0) { |
3608 | cpustat->nice = cputime64_add(cpustat->nice, tmp); |
3609 | cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp); |
3610 | } else { |
3611 | cpustat->user = cputime64_add(cpustat->user, tmp); |
3612 | cpustat->guest = cputime64_add(cpustat->guest, tmp); |
3613 | } |
3614 | } |
3615 | |
3616 | /* |
3617 | * Account system cpu time to a process and desired cpustat field |
3618 | * @p: the process that the cpu time gets accounted to |
3619 | * @cputime: the cpu time spent in kernel space since the last update |
3620 | * @cputime_scaled: cputime scaled by cpu frequency |
3621 | * @target_cputime64: pointer to cpustat field that has to be updated |
3622 | */ |
3623 | static inline |
3624 | void __account_system_time(struct task_struct *p, cputime_t cputime, |
3625 | cputime_t cputime_scaled, cputime64_t *target_cputime64) |
3626 | { |
3627 | cputime64_t tmp = cputime_to_cputime64(cputime); |
3628 | |
3629 | /* Add system time to process. */ |
3630 | p->stime = cputime_add(p->stime, cputime); |
3631 | p->stimescaled = cputime_add(p->stimescaled, cputime_scaled); |
3632 | account_group_system_time(p, cputime); |
3633 | |
3634 | /* Add system time to cpustat. */ |
3635 | *target_cputime64 = cputime64_add(*target_cputime64, tmp); |
3636 | cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime); |
3637 | |
3638 | /* Account for system time used */ |
3639 | acct_update_integrals(p); |
3640 | } |
3641 | |
3642 | /* |
3643 | * Account system cpu time to a process. |
3644 | * @p: the process that the cpu time gets accounted to |
3645 | * @hardirq_offset: the offset to subtract from hardirq_count() |
3646 | * @cputime: the cpu time spent in kernel space since the last update |
3647 | * @cputime_scaled: cputime scaled by cpu frequency |
3648 | */ |
3649 | void account_system_time(struct task_struct *p, int hardirq_offset, |
3650 | cputime_t cputime, cputime_t cputime_scaled) |
3651 | { |
3652 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
3653 | cputime64_t *target_cputime64; |
3654 | |
3655 | if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) { |
3656 | account_guest_time(p, cputime, cputime_scaled); |
3657 | return; |
3658 | } |
3659 | |
3660 | if (hardirq_count() - hardirq_offset) |
3661 | target_cputime64 = &cpustat->irq; |
3662 | else if (in_serving_softirq()) |
3663 | target_cputime64 = &cpustat->softirq; |
3664 | else |
3665 | target_cputime64 = &cpustat->system; |
3666 | |
3667 | __account_system_time(p, cputime, cputime_scaled, target_cputime64); |
3668 | } |
3669 | |
3670 | /* |
3671 | * Account for involuntary wait time. |
3672 | * @cputime: the cpu time spent in involuntary wait |
3673 | */ |
3674 | void account_steal_time(cputime_t cputime) |
3675 | { |
3676 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
3677 | cputime64_t cputime64 = cputime_to_cputime64(cputime); |
3678 | |
3679 | cpustat->steal = cputime64_add(cpustat->steal, cputime64); |
3680 | } |
3681 | |
3682 | /* |
3683 | * Account for idle time. |
3684 | * @cputime: the cpu time spent in idle wait |
3685 | */ |
3686 | void account_idle_time(cputime_t cputime) |
3687 | { |
3688 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
3689 | cputime64_t cputime64 = cputime_to_cputime64(cputime); |
3690 | struct rq *rq = this_rq(); |
3691 | |
3692 | if (atomic_read(&rq->nr_iowait) > 0) |
3693 | cpustat->iowait = cputime64_add(cpustat->iowait, cputime64); |
3694 | else |
3695 | cpustat->idle = cputime64_add(cpustat->idle, cputime64); |
3696 | } |
3697 | |
3698 | #ifndef CONFIG_VIRT_CPU_ACCOUNTING |
3699 | |
3700 | #ifdef CONFIG_IRQ_TIME_ACCOUNTING |
3701 | /* |
3702 | * Account a tick to a process and cpustat |
3703 | * @p: the process that the cpu time gets accounted to |
3704 | * @user_tick: is the tick from userspace |
3705 | * @rq: the pointer to rq |
3706 | * |
3707 | * Tick demultiplexing follows the order |
3708 | * - pending hardirq update |
3709 | * - pending softirq update |
3710 | * - user_time |
3711 | * - idle_time |
3712 | * - system time |
3713 | * - check for guest_time |
3714 | * - else account as system_time |
3715 | * |
3716 | * Check for hardirq is done both for system and user time as there is |
3717 | * no timer going off while we are on hardirq and hence we may never get an |
3718 | * opportunity to update it solely in system time. |
3719 | * p->stime and friends are only updated on system time and not on irq |
3720 | * softirq as those do not count in task exec_runtime any more. |
3721 | */ |
3722 | static void irqtime_account_process_tick(struct task_struct *p, int user_tick, |
3723 | struct rq *rq) |
3724 | { |
3725 | cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy); |
3726 | cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy); |
3727 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
3728 | |
3729 | if (irqtime_account_hi_update()) { |
3730 | cpustat->irq = cputime64_add(cpustat->irq, tmp); |
3731 | } else if (irqtime_account_si_update()) { |
3732 | cpustat->softirq = cputime64_add(cpustat->softirq, tmp); |
3733 | } else if (this_cpu_ksoftirqd() == p) { |
3734 | /* |
3735 | * ksoftirqd time do not get accounted in cpu_softirq_time. |
3736 | * So, we have to handle it separately here. |
3737 | * Also, p->stime needs to be updated for ksoftirqd. |
3738 | */ |
3739 | __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled, |
3740 | &cpustat->softirq); |
3741 | } else if (user_tick) { |
3742 | account_user_time(p, cputime_one_jiffy, one_jiffy_scaled); |
3743 | } else if (p == rq->idle) { |
3744 | account_idle_time(cputime_one_jiffy); |
3745 | } else if (p->flags & PF_VCPU) { /* System time or guest time */ |
3746 | account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled); |
3747 | } else { |
3748 | __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled, |
3749 | &cpustat->system); |
3750 | } |
3751 | } |
3752 | |
3753 | static void irqtime_account_idle_ticks(int ticks) |
3754 | { |
3755 | int i; |
3756 | struct rq *rq = this_rq(); |
3757 | |
3758 | for (i = 0; i < ticks; i++) |
3759 | irqtime_account_process_tick(current, 0, rq); |
3760 | } |
3761 | #else /* CONFIG_IRQ_TIME_ACCOUNTING */ |
3762 | static void irqtime_account_idle_ticks(int ticks) {} |
3763 | static void irqtime_account_process_tick(struct task_struct *p, int user_tick, |
3764 | struct rq *rq) {} |
3765 | #endif /* CONFIG_IRQ_TIME_ACCOUNTING */ |
3766 | |
3767 | /* |
3768 | * Account a single tick of cpu time. |
3769 | * @p: the process that the cpu time gets accounted to |
3770 | * @user_tick: indicates if the tick is a user or a system tick |
3771 | */ |
3772 | void account_process_tick(struct task_struct *p, int user_tick) |
3773 | { |
3774 | cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy); |
3775 | struct rq *rq = this_rq(); |
3776 | |
3777 | if (sched_clock_irqtime) { |
3778 | irqtime_account_process_tick(p, user_tick, rq); |
3779 | return; |
3780 | } |
3781 | |
3782 | if (user_tick) |
3783 | account_user_time(p, cputime_one_jiffy, one_jiffy_scaled); |
3784 | else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET)) |
3785 | account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy, |
3786 | one_jiffy_scaled); |
3787 | else |
3788 | account_idle_time(cputime_one_jiffy); |
3789 | } |
3790 | |
3791 | /* |
3792 | * Account multiple ticks of steal time. |
3793 | * @p: the process from which the cpu time has been stolen |
3794 | * @ticks: number of stolen ticks |
3795 | */ |
3796 | void account_steal_ticks(unsigned long ticks) |
3797 | { |
3798 | account_steal_time(jiffies_to_cputime(ticks)); |
3799 | } |
3800 | |
3801 | /* |
3802 | * Account multiple ticks of idle time. |
3803 | * @ticks: number of stolen ticks |
3804 | */ |
3805 | void account_idle_ticks(unsigned long ticks) |
3806 | { |
3807 | |
3808 | if (sched_clock_irqtime) { |
3809 | irqtime_account_idle_ticks(ticks); |
3810 | return; |
3811 | } |
3812 | |
3813 | account_idle_time(jiffies_to_cputime(ticks)); |
3814 | } |
3815 | |
3816 | #endif |
3817 | |
3818 | /* |
3819 | * Use precise platform statistics if available: |
3820 | */ |
3821 | #ifdef CONFIG_VIRT_CPU_ACCOUNTING |
3822 | void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st) |
3823 | { |
3824 | *ut = p->utime; |
3825 | *st = p->stime; |
3826 | } |
3827 | |
3828 | void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st) |
3829 | { |
3830 | struct task_cputime cputime; |
3831 | |
3832 | thread_group_cputime(p, &cputime); |
3833 | |
3834 | *ut = cputime.utime; |
3835 | *st = cputime.stime; |
3836 | } |
3837 | #else |
3838 | |
3839 | #ifndef nsecs_to_cputime |
3840 | # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs) |
3841 | #endif |
3842 | |
3843 | void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st) |
3844 | { |
3845 | cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime); |
3846 | |
3847 | /* |
3848 | * Use CFS's precise accounting: |
3849 | */ |
3850 | rtime = nsecs_to_cputime(p->se.sum_exec_runtime); |
3851 | |
3852 | if (total) { |
3853 | u64 temp = rtime; |
3854 | |
3855 | temp *= utime; |
3856 | do_div(temp, total); |
3857 | utime = (cputime_t)temp; |
3858 | } else |
3859 | utime = rtime; |
3860 | |
3861 | /* |
3862 | * Compare with previous values, to keep monotonicity: |
3863 | */ |
3864 | p->prev_utime = max(p->prev_utime, utime); |
3865 | p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime)); |
3866 | |
3867 | *ut = p->prev_utime; |
3868 | *st = p->prev_stime; |
3869 | } |
3870 | |
3871 | /* |
3872 | * Must be called with siglock held. |
3873 | */ |
3874 | void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st) |
3875 | { |
3876 | struct signal_struct *sig = p->signal; |
3877 | struct task_cputime cputime; |
3878 | cputime_t rtime, utime, total; |
3879 | |
3880 | thread_group_cputime(p, &cputime); |
3881 | |
3882 | total = cputime_add(cputime.utime, cputime.stime); |
3883 | rtime = nsecs_to_cputime(cputime.sum_exec_runtime); |
3884 | |
3885 | if (total) { |
3886 | u64 temp = rtime; |
3887 | |
3888 | temp *= cputime.utime; |
3889 | do_div(temp, total); |
3890 | utime = (cputime_t)temp; |
3891 | } else |
3892 | utime = rtime; |
3893 | |
3894 | sig->prev_utime = max(sig->prev_utime, utime); |
3895 | sig->prev_stime = max(sig->prev_stime, |
3896 | cputime_sub(rtime, sig->prev_utime)); |
3897 | |
3898 | *ut = sig->prev_utime; |
3899 | *st = sig->prev_stime; |
3900 | } |
3901 | #endif |
3902 | |
3903 | /* |
3904 | * This function gets called by the timer code, with HZ frequency. |
3905 | * We call it with interrupts disabled. |
3906 | * |
3907 | * It also gets called by the fork code, when changing the parent's |
3908 | * timeslices. |
3909 | */ |
3910 | void scheduler_tick(void) |
3911 | { |
3912 | int cpu = smp_processor_id(); |
3913 | struct rq *rq = cpu_rq(cpu); |
3914 | struct task_struct *curr = rq->curr; |
3915 | |
3916 | sched_clock_tick(); |
3917 | |
3918 | raw_spin_lock(&rq->lock); |
3919 | update_rq_clock(rq); |
3920 | update_cpu_load_active(rq); |
3921 | curr->sched_class->task_tick(rq, curr, 0); |
3922 | raw_spin_unlock(&rq->lock); |
3923 | |
3924 | perf_event_task_tick(); |
3925 | |
3926 | #ifdef CONFIG_SMP |
3927 | rq->idle_at_tick = idle_cpu(cpu); |
3928 | trigger_load_balance(rq, cpu); |
3929 | #endif |
3930 | } |
3931 | |
3932 | notrace unsigned long get_parent_ip(unsigned long addr) |
3933 | { |
3934 | if (in_lock_functions(addr)) { |
3935 | addr = CALLER_ADDR2; |
3936 | if (in_lock_functions(addr)) |
3937 | addr = CALLER_ADDR3; |
3938 | } |
3939 | return addr; |
3940 | } |
3941 | |
3942 | #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ |
3943 | defined(CONFIG_PREEMPT_TRACER)) |
3944 | |
3945 | void __kprobes add_preempt_count(int val) |
3946 | { |
3947 | #ifdef CONFIG_DEBUG_PREEMPT |
3948 | /* |
3949 | * Underflow? |
3950 | */ |
3951 | if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) |
3952 | return; |
3953 | #endif |
3954 | preempt_count() += val; |
3955 | #ifdef CONFIG_DEBUG_PREEMPT |
3956 | /* |
3957 | * Spinlock count overflowing soon? |
3958 | */ |
3959 | DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= |
3960 | PREEMPT_MASK - 10); |
3961 | #endif |
3962 | if (preempt_count() == val) |
3963 | trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); |
3964 | } |
3965 | EXPORT_SYMBOL(add_preempt_count); |
3966 | |
3967 | void __kprobes sub_preempt_count(int val) |
3968 | { |
3969 | #ifdef CONFIG_DEBUG_PREEMPT |
3970 | /* |
3971 | * Underflow? |
3972 | */ |
3973 | if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) |
3974 | return; |
3975 | /* |
3976 | * Is the spinlock portion underflowing? |
3977 | */ |
3978 | if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && |
3979 | !(preempt_count() & PREEMPT_MASK))) |
3980 | return; |
3981 | #endif |
3982 | |
3983 | if (preempt_count() == val) |
3984 | trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); |
3985 | preempt_count() -= val; |
3986 | } |
3987 | EXPORT_SYMBOL(sub_preempt_count); |
3988 | |
3989 | #endif |
3990 | |
3991 | /* |
3992 | * Print scheduling while atomic bug: |
3993 | */ |
3994 | static noinline void __schedule_bug(struct task_struct *prev) |
3995 | { |
3996 | struct pt_regs *regs = get_irq_regs(); |
3997 | |
3998 | printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", |
3999 | prev->comm, prev->pid, preempt_count()); |
4000 | |
4001 | debug_show_held_locks(prev); |
4002 | print_modules(); |
4003 | if (irqs_disabled()) |
4004 | print_irqtrace_events(prev); |
4005 | |
4006 | if (regs) |
4007 | show_regs(regs); |
4008 | else |
4009 | dump_stack(); |
4010 | } |
4011 | |
4012 | /* |
4013 | * Various schedule()-time debugging checks and statistics: |
4014 | */ |
4015 | static inline void schedule_debug(struct task_struct *prev) |
4016 | { |
4017 | /* |
4018 | * Test if we are atomic. Since do_exit() needs to call into |
4019 | * schedule() atomically, we ignore that path for now. |
4020 | * Otherwise, whine if we are scheduling when we should not be. |
4021 | */ |
4022 | if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) |
4023 | __schedule_bug(prev); |
4024 | |
4025 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); |
4026 | |
4027 | schedstat_inc(this_rq(), sched_count); |
4028 | #ifdef CONFIG_SCHEDSTATS |
4029 | if (unlikely(prev->lock_depth >= 0)) { |
4030 | schedstat_inc(this_rq(), rq_sched_info.bkl_count); |
4031 | schedstat_inc(prev, sched_info.bkl_count); |
4032 | } |
4033 | #endif |
4034 | } |
4035 | |
4036 | static void put_prev_task(struct rq *rq, struct task_struct *prev) |
4037 | { |
4038 | if (prev->se.on_rq) |
4039 | update_rq_clock(rq); |
4040 | prev->sched_class->put_prev_task(rq, prev); |
4041 | } |
4042 | |
4043 | /* |
4044 | * Pick up the highest-prio task: |
4045 | */ |
4046 | static inline struct task_struct * |
4047 | pick_next_task(struct rq *rq) |
4048 | { |
4049 | const struct sched_class *class; |
4050 | struct task_struct *p; |
4051 | |
4052 | /* |
4053 | * Optimization: we know that if all tasks are in |
4054 | * the fair class we can call that function directly: |
4055 | */ |
4056 | if (likely(rq->nr_running == rq->cfs.nr_running)) { |
4057 | p = fair_sched_class.pick_next_task(rq); |
4058 | if (likely(p)) |
4059 | return p; |
4060 | } |
4061 | |
4062 | for_each_class(class) { |
4063 | p = class->pick_next_task(rq); |
4064 | if (p) |
4065 | return p; |
4066 | } |
4067 | |
4068 | BUG(); /* the idle class will always have a runnable task */ |
4069 | } |
4070 | |
4071 | /* |
4072 | * schedule() is the main scheduler function. |
4073 | */ |
4074 | asmlinkage void __sched schedule(void) |
4075 | { |
4076 | struct task_struct *prev, *next; |
4077 | unsigned long *switch_count; |
4078 | struct rq *rq; |
4079 | int cpu; |
4080 | |
4081 | need_resched: |
4082 | preempt_disable(); |
4083 | cpu = smp_processor_id(); |
4084 | rq = cpu_rq(cpu); |
4085 | rcu_note_context_switch(cpu); |
4086 | prev = rq->curr; |
4087 | |
4088 | schedule_debug(prev); |
4089 | |
4090 | if (sched_feat(HRTICK)) |
4091 | hrtick_clear(rq); |
4092 | |
4093 | raw_spin_lock_irq(&rq->lock); |
4094 | |
4095 | switch_count = &prev->nivcsw; |
4096 | if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { |
4097 | if (unlikely(signal_pending_state(prev->state, prev))) { |
4098 | prev->state = TASK_RUNNING; |
4099 | } else { |
4100 | /* |
4101 | * If a worker is going to sleep, notify and |
4102 | * ask workqueue whether it wants to wake up a |
4103 | * task to maintain concurrency. If so, wake |
4104 | * up the task. |
4105 | */ |
4106 | if (prev->flags & PF_WQ_WORKER) { |
4107 | struct task_struct *to_wakeup; |
4108 | |
4109 | to_wakeup = wq_worker_sleeping(prev, cpu); |
4110 | if (to_wakeup) |
4111 | try_to_wake_up_local(to_wakeup); |
4112 | } |
4113 | deactivate_task(rq, prev, DEQUEUE_SLEEP); |
4114 | |
4115 | /* |
4116 | * If we are going to sleep and we have plugged IO queued, make |
4117 | * sure to submit it to avoid deadlocks. |
4118 | */ |
4119 | if (blk_needs_flush_plug(prev)) { |
4120 | raw_spin_unlock(&rq->lock); |
4121 | blk_schedule_flush_plug(prev); |
4122 | raw_spin_lock(&rq->lock); |
4123 | } |
4124 | } |
4125 | switch_count = &prev->nvcsw; |
4126 | } |
4127 | |
4128 | pre_schedule(rq, prev); |
4129 | |
4130 | if (unlikely(!rq->nr_running)) |
4131 | idle_balance(cpu, rq); |
4132 | |
4133 | put_prev_task(rq, prev); |
4134 | next = pick_next_task(rq); |
4135 | clear_tsk_need_resched(prev); |
4136 | rq->skip_clock_update = 0; |
4137 | |
4138 | if (likely(prev != next)) { |
4139 | rq->nr_switches++; |
4140 | rq->curr = next; |
4141 | ++*switch_count; |
4142 | |
4143 | context_switch(rq, prev, next); /* unlocks the rq */ |
4144 | /* |
4145 | * The context switch have flipped the stack from under us |
4146 | * and restored the local variables which were saved when |
4147 | * this task called schedule() in the past. prev == current |
4148 | * is still correct, but it can be moved to another cpu/rq. |
4149 | */ |
4150 | cpu = smp_processor_id(); |
4151 | rq = cpu_rq(cpu); |
4152 | } else |
4153 | raw_spin_unlock_irq(&rq->lock); |
4154 | |
4155 | post_schedule(rq); |
4156 | |
4157 | preempt_enable_no_resched(); |
4158 | if (need_resched()) |
4159 | goto need_resched; |
4160 | } |
4161 | EXPORT_SYMBOL(schedule); |
4162 | |
4163 | #ifdef CONFIG_MUTEX_SPIN_ON_OWNER |
4164 | /* |
4165 | * Look out! "owner" is an entirely speculative pointer |
4166 | * access and not reliable. |
4167 | */ |
4168 | int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner) |
4169 | { |
4170 | unsigned int cpu; |
4171 | struct rq *rq; |
4172 | |
4173 | if (!sched_feat(OWNER_SPIN)) |
4174 | return 0; |
4175 | |
4176 | #ifdef CONFIG_DEBUG_PAGEALLOC |
4177 | /* |
4178 | * Need to access the cpu field knowing that |
4179 | * DEBUG_PAGEALLOC could have unmapped it if |
4180 | * the mutex owner just released it and exited. |
4181 | */ |
4182 | if (probe_kernel_address(&owner->cpu, cpu)) |
4183 | return 0; |
4184 | #else |
4185 | cpu = owner->cpu; |
4186 | #endif |
4187 | |
4188 | /* |
4189 | * Even if the access succeeded (likely case), |
4190 | * the cpu field may no longer be valid. |
4191 | */ |
4192 | if (cpu >= nr_cpumask_bits) |
4193 | return 0; |
4194 | |
4195 | /* |
4196 | * We need to validate that we can do a |
4197 | * get_cpu() and that we have the percpu area. |
4198 | */ |
4199 | if (!cpu_online(cpu)) |
4200 | return 0; |
4201 | |
4202 | rq = cpu_rq(cpu); |
4203 | |
4204 | for (;;) { |
4205 | /* |
4206 | * Owner changed, break to re-assess state. |
4207 | */ |
4208 | if (lock->owner != owner) { |
4209 | /* |
4210 | * If the lock has switched to a different owner, |
4211 | * we likely have heavy contention. Return 0 to quit |
4212 | * optimistic spinning and not contend further: |
4213 | */ |
4214 | if (lock->owner) |
4215 | return 0; |
4216 | break; |
4217 | } |
4218 | |
4219 | /* |
4220 | * Is that owner really running on that cpu? |
4221 | */ |
4222 | if (task_thread_info(rq->curr) != owner || need_resched()) |
4223 | return 0; |
4224 | |
4225 | arch_mutex_cpu_relax(); |
4226 | } |
4227 | |
4228 | return 1; |
4229 | } |
4230 | #endif |
4231 | |
4232 | #ifdef CONFIG_PREEMPT |
4233 | /* |
4234 | * this is the entry point to schedule() from in-kernel preemption |
4235 | * off of preempt_enable. Kernel preemptions off return from interrupt |
4236 | * occur there and call schedule directly. |
4237 | */ |
4238 | asmlinkage void __sched notrace preempt_schedule(void) |
4239 | { |
4240 | struct thread_info *ti = current_thread_info(); |
4241 | |
4242 | /* |
4243 | * If there is a non-zero preempt_count or interrupts are disabled, |
4244 | * we do not want to preempt the current task. Just return.. |
4245 | */ |
4246 | if (likely(ti->preempt_count || irqs_disabled())) |
4247 | return; |
4248 | |
4249 | do { |
4250 | add_preempt_count_notrace(PREEMPT_ACTIVE); |
4251 | schedule(); |
4252 | sub_preempt_count_notrace(PREEMPT_ACTIVE); |
4253 | |
4254 | /* |
4255 | * Check again in case we missed a preemption opportunity |
4256 | * between schedule and now. |
4257 | */ |
4258 | barrier(); |
4259 | } while (need_resched()); |
4260 | } |
4261 | EXPORT_SYMBOL(preempt_schedule); |
4262 | |
4263 | /* |
4264 | * this is the entry point to schedule() from kernel preemption |
4265 | * off of irq context. |
4266 | * Note, that this is called and return with irqs disabled. This will |
4267 | * protect us against recursive calling from irq. |
4268 | */ |
4269 | asmlinkage void __sched preempt_schedule_irq(void) |
4270 | { |
4271 | struct thread_info *ti = current_thread_info(); |
4272 | |
4273 | /* Catch callers which need to be fixed */ |
4274 | BUG_ON(ti->preempt_count || !irqs_disabled()); |
4275 | |
4276 | do { |
4277 | add_preempt_count(PREEMPT_ACTIVE); |
4278 | local_irq_enable(); |
4279 | schedule(); |
4280 | local_irq_disable(); |
4281 | sub_preempt_count(PREEMPT_ACTIVE); |
4282 | |
4283 | /* |
4284 | * Check again in case we missed a preemption opportunity |
4285 | * between schedule and now. |
4286 | */ |
4287 | barrier(); |
4288 | } while (need_resched()); |
4289 | } |
4290 | |
4291 | #endif /* CONFIG_PREEMPT */ |
4292 | |
4293 | int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, |
4294 | void *key) |
4295 | { |
4296 | return try_to_wake_up(curr->private, mode, wake_flags); |
4297 | } |
4298 | EXPORT_SYMBOL(default_wake_function); |
4299 | |
4300 | /* |
4301 | * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just |
4302 | * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve |
4303 | * number) then we wake all the non-exclusive tasks and one exclusive task. |
4304 | * |
4305 | * There are circumstances in which we can try to wake a task which has already |
4306 | * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns |
4307 | * zero in this (rare) case, and we handle it by continuing to scan the queue. |
4308 | */ |
4309 | static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, |
4310 | int nr_exclusive, int wake_flags, void *key) |
4311 | { |
4312 | wait_queue_t *curr, *next; |
4313 | |
4314 | list_for_each_entry_safe(curr, next, &q->task_list, task_list) { |
4315 | unsigned flags = curr->flags; |
4316 | |
4317 | if (curr->func(curr, mode, wake_flags, key) && |
4318 | (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) |
4319 | break; |
4320 | } |
4321 | } |
4322 | |
4323 | /** |
4324 | * __wake_up - wake up threads blocked on a waitqueue. |
4325 | * @q: the waitqueue |
4326 | * @mode: which threads |
4327 | * @nr_exclusive: how many wake-one or wake-many threads to wake up |
4328 | * @key: is directly passed to the wakeup function |
4329 | * |
4330 | * It may be assumed that this function implies a write memory barrier before |
4331 | * changing the task state if and only if any tasks are woken up. |
4332 | */ |
4333 | void __wake_up(wait_queue_head_t *q, unsigned int mode, |
4334 | int nr_exclusive, void *key) |
4335 | { |
4336 | unsigned long flags; |
4337 | |
4338 | spin_lock_irqsave(&q->lock, flags); |
4339 | __wake_up_common(q, mode, nr_exclusive, 0, key); |
4340 | spin_unlock_irqrestore(&q->lock, flags); |
4341 | } |
4342 | EXPORT_SYMBOL(__wake_up); |
4343 | |
4344 | /* |
4345 | * Same as __wake_up but called with the spinlock in wait_queue_head_t held. |
4346 | */ |
4347 | void __wake_up_locked(wait_queue_head_t *q, unsigned int mode) |
4348 | { |
4349 | __wake_up_common(q, mode, 1, 0, NULL); |
4350 | } |
4351 | EXPORT_SYMBOL_GPL(__wake_up_locked); |
4352 | |
4353 | void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key) |
4354 | { |
4355 | __wake_up_common(q, mode, 1, 0, key); |
4356 | } |
4357 | EXPORT_SYMBOL_GPL(__wake_up_locked_key); |
4358 | |
4359 | /** |
4360 | * __wake_up_sync_key - wake up threads blocked on a waitqueue. |
4361 | * @q: the waitqueue |
4362 | * @mode: which threads |
4363 | * @nr_exclusive: how many wake-one or wake-many threads to wake up |
4364 | * @key: opaque value to be passed to wakeup targets |
4365 | * |
4366 | * The sync wakeup differs that the waker knows that it will schedule |
4367 | * away soon, so while the target thread will be woken up, it will not |
4368 | * be migrated to another CPU - ie. the two threads are 'synchronized' |
4369 | * with each other. This can prevent needless bouncing between CPUs. |
4370 | * |
4371 | * On UP it can prevent extra preemption. |
4372 | * |
4373 | * It may be assumed that this function implies a write memory barrier before |
4374 | * changing the task state if and only if any tasks are woken up. |
4375 | */ |
4376 | void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode, |
4377 | int nr_exclusive, void *key) |
4378 | { |
4379 | unsigned long flags; |
4380 | int wake_flags = WF_SYNC; |
4381 | |
4382 | if (unlikely(!q)) |
4383 | return; |
4384 | |
4385 | if (unlikely(!nr_exclusive)) |
4386 | wake_flags = 0; |
4387 | |
4388 | spin_lock_irqsave(&q->lock, flags); |
4389 | __wake_up_common(q, mode, nr_exclusive, wake_flags, key); |
4390 | spin_unlock_irqrestore(&q->lock, flags); |
4391 | } |
4392 | EXPORT_SYMBOL_GPL(__wake_up_sync_key); |
4393 | |
4394 | /* |
4395 | * __wake_up_sync - see __wake_up_sync_key() |
4396 | */ |
4397 | void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) |
4398 | { |
4399 | __wake_up_sync_key(q, mode, nr_exclusive, NULL); |
4400 | } |
4401 | EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ |
4402 | |
4403 | /** |
4404 | * complete: - signals a single thread waiting on this completion |
4405 | * @x: holds the state of this particular completion |
4406 | * |
4407 | * This will wake up a single thread waiting on this completion. Threads will be |
4408 | * awakened in the same order in which they were queued. |
4409 | * |
4410 | * See also complete_all(), wait_for_completion() and related routines. |
4411 | * |
4412 | * It may be assumed that this function implies a write memory barrier before |
4413 | * changing the task state if and only if any tasks are woken up. |
4414 | */ |
4415 | void complete(struct completion *x) |
4416 | { |
4417 | unsigned long flags; |
4418 | |
4419 | spin_lock_irqsave(&x->wait.lock, flags); |
4420 | x->done++; |
4421 | __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); |
4422 | spin_unlock_irqrestore(&x->wait.lock, flags); |
4423 | } |
4424 | EXPORT_SYMBOL(complete); |
4425 | |
4426 | /** |
4427 | * complete_all: - signals all threads waiting on this completion |
4428 | * @x: holds the state of this particular completion |
4429 | * |
4430 | * This will wake up all threads waiting on this particular completion event. |
4431 | * |
4432 | * It may be assumed that this function implies a write memory barrier before |
4433 | * changing the task state if and only if any tasks are woken up. |
4434 | */ |
4435 | void complete_all(struct completion *x) |
4436 | { |
4437 | unsigned long flags; |
4438 | |
4439 | spin_lock_irqsave(&x->wait.lock, flags); |
4440 | x->done += UINT_MAX/2; |
4441 | __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); |
4442 | spin_unlock_irqrestore(&x->wait.lock, flags); |
4443 | } |
4444 | EXPORT_SYMBOL(complete_all); |
4445 | |
4446 | static inline long __sched |
4447 | do_wait_for_common(struct completion *x, long timeout, int state) |
4448 | { |
4449 | if (!x->done) { |
4450 | DECLARE_WAITQUEUE(wait, current); |
4451 | |
4452 | __add_wait_queue_tail_exclusive(&x->wait, &wait); |
4453 | do { |
4454 | if (signal_pending_state(state, current)) { |
4455 | timeout = -ERESTARTSYS; |
4456 | break; |
4457 | } |
4458 | __set_current_state(state); |
4459 | spin_unlock_irq(&x->wait.lock); |
4460 | timeout = schedule_timeout(timeout); |
4461 | spin_lock_irq(&x->wait.lock); |
4462 | } while (!x->done && timeout); |
4463 | __remove_wait_queue(&x->wait, &wait); |
4464 | if (!x->done) |
4465 | return timeout; |
4466 | } |
4467 | x->done--; |
4468 | return timeout ?: 1; |
4469 | } |
4470 | |
4471 | static long __sched |
4472 | wait_for_common(struct completion *x, long timeout, int state) |
4473 | { |
4474 | might_sleep(); |
4475 | |
4476 | spin_lock_irq(&x->wait.lock); |
4477 | timeout = do_wait_for_common(x, timeout, state); |
4478 | spin_unlock_irq(&x->wait.lock); |
4479 | return timeout; |
4480 | } |
4481 | |
4482 | /** |
4483 | * wait_for_completion: - waits for completion of a task |
4484 | * @x: holds the state of this particular completion |
4485 | * |
4486 | * This waits to be signaled for completion of a specific task. It is NOT |
4487 | * interruptible and there is no timeout. |
4488 | * |
4489 | * See also similar routines (i.e. wait_for_completion_timeout()) with timeout |
4490 | * and interrupt capability. Also see complete(). |
4491 | */ |
4492 | void __sched wait_for_completion(struct completion *x) |
4493 | { |
4494 | wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); |
4495 | } |
4496 | EXPORT_SYMBOL(wait_for_completion); |
4497 | |
4498 | /** |
4499 | * wait_for_completion_timeout: - waits for completion of a task (w/timeout) |
4500 | * @x: holds the state of this particular completion |
4501 | * @timeout: timeout value in jiffies |
4502 | * |
4503 | * This waits for either a completion of a specific task to be signaled or for a |
4504 | * specified timeout to expire. The timeout is in jiffies. It is not |
4505 | * interruptible. |
4506 | */ |
4507 | unsigned long __sched |
4508 | wait_for_completion_timeout(struct completion *x, unsigned long timeout) |
4509 | { |
4510 | return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); |
4511 | } |
4512 | EXPORT_SYMBOL(wait_for_completion_timeout); |
4513 | |
4514 | /** |
4515 | * wait_for_completion_interruptible: - waits for completion of a task (w/intr) |
4516 | * @x: holds the state of this particular completion |
4517 | * |
4518 | * This waits for completion of a specific task to be signaled. It is |
4519 | * interruptible. |
4520 | */ |
4521 | int __sched wait_for_completion_interruptible(struct completion *x) |
4522 | { |
4523 | long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); |
4524 | if (t == -ERESTARTSYS) |
4525 | return t; |
4526 | return 0; |
4527 | } |
4528 | EXPORT_SYMBOL(wait_for_completion_interruptible); |
4529 | |
4530 | /** |
4531 | * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) |
4532 | * @x: holds the state of this particular completion |
4533 | * @timeout: timeout value in jiffies |
4534 | * |
4535 | * This waits for either a completion of a specific task to be signaled or for a |
4536 | * specified timeout to expire. It is interruptible. The timeout is in jiffies. |
4537 | */ |
4538 | long __sched |
4539 | wait_for_completion_interruptible_timeout(struct completion *x, |
4540 | unsigned long timeout) |
4541 | { |
4542 | return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); |
4543 | } |
4544 | EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); |
4545 | |
4546 | /** |
4547 | * wait_for_completion_killable: - waits for completion of a task (killable) |
4548 | * @x: holds the state of this particular completion |
4549 | * |
4550 | * This waits to be signaled for completion of a specific task. It can be |
4551 | * interrupted by a kill signal. |
4552 | */ |
4553 | int __sched wait_for_completion_killable(struct completion *x) |
4554 | { |
4555 | long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); |
4556 | if (t == -ERESTARTSYS) |
4557 | return t; |
4558 | return 0; |
4559 | } |
4560 | EXPORT_SYMBOL(wait_for_completion_killable); |
4561 | |
4562 | /** |
4563 | * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable)) |
4564 | * @x: holds the state of this particular completion |
4565 | * @timeout: timeout value in jiffies |
4566 | * |
4567 | * This waits for either a completion of a specific task to be |
4568 | * signaled or for a specified timeout to expire. It can be |
4569 | * interrupted by a kill signal. The timeout is in jiffies. |
4570 | */ |
4571 | long __sched |
4572 | wait_for_completion_killable_timeout(struct completion *x, |
4573 | unsigned long timeout) |
4574 | { |
4575 | return wait_for_common(x, timeout, TASK_KILLABLE); |
4576 | } |
4577 | EXPORT_SYMBOL(wait_for_completion_killable_timeout); |
4578 | |
4579 | /** |
4580 | * try_wait_for_completion - try to decrement a completion without blocking |
4581 | * @x: completion structure |
4582 | * |
4583 | * Returns: 0 if a decrement cannot be done without blocking |
4584 | * 1 if a decrement succeeded. |
4585 | * |
4586 | * If a completion is being used as a counting completion, |
4587 | * attempt to decrement the counter without blocking. This |
4588 | * enables us to avoid waiting if the resource the completion |
4589 | * is protecting is not available. |
4590 | */ |
4591 | bool try_wait_for_completion(struct completion *x) |
4592 | { |
4593 | unsigned long flags; |
4594 | int ret = 1; |
4595 | |
4596 | spin_lock_irqsave(&x->wait.lock, flags); |
4597 | if (!x->done) |
4598 | ret = 0; |
4599 | else |
4600 | x->done--; |
4601 | spin_unlock_irqrestore(&x->wait.lock, flags); |
4602 | return ret; |
4603 | } |
4604 | EXPORT_SYMBOL(try_wait_for_completion); |
4605 | |
4606 | /** |
4607 | * completion_done - Test to see if a completion has any waiters |
4608 | * @x: completion structure |
4609 | * |
4610 | * Returns: 0 if there are waiters (wait_for_completion() in progress) |
4611 | * 1 if there are no waiters. |
4612 | * |
4613 | */ |
4614 | bool completion_done(struct completion *x) |
4615 | { |
4616 | unsigned long flags; |
4617 | int ret = 1; |
4618 | |
4619 | spin_lock_irqsave(&x->wait.lock, flags); |
4620 | if (!x->done) |
4621 | ret = 0; |
4622 | spin_unlock_irqrestore(&x->wait.lock, flags); |
4623 | return ret; |
4624 | } |
4625 | EXPORT_SYMBOL(completion_done); |
4626 | |
4627 | static long __sched |
4628 | sleep_on_common(wait_queue_head_t *q, int state, long timeout) |
4629 | { |
4630 | unsigned long flags; |
4631 | wait_queue_t wait; |
4632 | |
4633 | init_waitqueue_entry(&wait, current); |
4634 | |
4635 | __set_current_state(state); |
4636 | |
4637 | spin_lock_irqsave(&q->lock, flags); |
4638 | __add_wait_queue(q, &wait); |
4639 | spin_unlock(&q->lock); |
4640 | timeout = schedule_timeout(timeout); |
4641 | spin_lock_irq(&q->lock); |
4642 | __remove_wait_queue(q, &wait); |
4643 | spin_unlock_irqrestore(&q->lock, flags); |
4644 | |
4645 | return timeout; |
4646 | } |
4647 | |
4648 | void __sched interruptible_sleep_on(wait_queue_head_t *q) |
4649 | { |
4650 | sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); |
4651 | } |
4652 | EXPORT_SYMBOL(interruptible_sleep_on); |
4653 | |
4654 | long __sched |
4655 | interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) |
4656 | { |
4657 | return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); |
4658 | } |
4659 | EXPORT_SYMBOL(interruptible_sleep_on_timeout); |
4660 | |
4661 | void __sched sleep_on(wait_queue_head_t *q) |
4662 | { |
4663 | sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); |
4664 | } |
4665 | EXPORT_SYMBOL(sleep_on); |
4666 | |
4667 | long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) |
4668 | { |
4669 | return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); |
4670 | } |
4671 | EXPORT_SYMBOL(sleep_on_timeout); |
4672 | |
4673 | #ifdef CONFIG_RT_MUTEXES |
4674 | |
4675 | /* |
4676 | * rt_mutex_setprio - set the current priority of a task |
4677 | * @p: task |
4678 | * @prio: prio value (kernel-internal form) |
4679 | * |
4680 | * This function changes the 'effective' priority of a task. It does |
4681 | * not touch ->normal_prio like __setscheduler(). |
4682 | * |
4683 | * Used by the rt_mutex code to implement priority inheritance logic. |
4684 | */ |
4685 | void rt_mutex_setprio(struct task_struct *p, int prio) |
4686 | { |
4687 | unsigned long flags; |
4688 | int oldprio, on_rq, running; |
4689 | struct rq *rq; |
4690 | const struct sched_class *prev_class; |
4691 | |
4692 | BUG_ON(prio < 0 || prio > MAX_PRIO); |
4693 | |
4694 | rq = task_rq_lock(p, &flags); |
4695 | |
4696 | trace_sched_pi_setprio(p, prio); |
4697 | oldprio = p->prio; |
4698 | prev_class = p->sched_class; |
4699 | on_rq = p->se.on_rq; |
4700 | running = task_current(rq, p); |
4701 | if (on_rq) |
4702 | dequeue_task(rq, p, 0); |
4703 | if (running) |
4704 | p->sched_class->put_prev_task(rq, p); |
4705 | |
4706 | if (rt_prio(prio)) |
4707 | p->sched_class = &rt_sched_class; |
4708 | else |
4709 | p->sched_class = &fair_sched_class; |
4710 | |
4711 | p->prio = prio; |
4712 | |
4713 | if (running) |
4714 | p->sched_class->set_curr_task(rq); |
4715 | if (on_rq) |
4716 | enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0); |
4717 | |
4718 | check_class_changed(rq, p, prev_class, oldprio); |
4719 | task_rq_unlock(rq, &flags); |
4720 | } |
4721 | |
4722 | #endif |
4723 | |
4724 | void set_user_nice(struct task_struct *p, long nice) |
4725 | { |
4726 | int old_prio, delta, on_rq; |
4727 | unsigned long flags; |
4728 | struct rq *rq; |
4729 | |
4730 | if (TASK_NICE(p) == nice || nice < -20 || nice > 19) |
4731 | return; |
4732 | /* |
4733 | * We have to be careful, if called from sys_setpriority(), |
4734 | * the task might be in the middle of scheduling on another CPU. |
4735 | */ |
4736 | rq = task_rq_lock(p, &flags); |
4737 | /* |
4738 | * The RT priorities are set via sched_setscheduler(), but we still |
4739 | * allow the 'normal' nice value to be set - but as expected |
4740 | * it wont have any effect on scheduling until the task is |
4741 | * SCHED_FIFO/SCHED_RR: |
4742 | */ |
4743 | if (task_has_rt_policy(p)) { |
4744 | p->static_prio = NICE_TO_PRIO(nice); |
4745 | goto out_unlock; |
4746 | } |
4747 | on_rq = p->se.on_rq; |
4748 | if (on_rq) |
4749 | dequeue_task(rq, p, 0); |
4750 | |
4751 | p->static_prio = NICE_TO_PRIO(nice); |
4752 | set_load_weight(p); |
4753 | old_prio = p->prio; |
4754 | p->prio = effective_prio(p); |
4755 | delta = p->prio - old_prio; |
4756 | |
4757 | if (on_rq) { |
4758 | enqueue_task(rq, p, 0); |
4759 | /* |
4760 | * If the task increased its priority or is running and |
4761 | * lowered its priority, then reschedule its CPU: |
4762 | */ |
4763 | if (delta < 0 || (delta > 0 && task_running(rq, p))) |
4764 | resched_task(rq->curr); |
4765 | } |
4766 | out_unlock: |
4767 | task_rq_unlock(rq, &flags); |
4768 | } |
4769 | EXPORT_SYMBOL(set_user_nice); |
4770 | |
4771 | /* |
4772 | * can_nice - check if a task can reduce its nice value |
4773 | * @p: task |
4774 | * @nice: nice value |
4775 | */ |
4776 | int can_nice(const struct task_struct *p, const int nice) |
4777 | { |
4778 | /* convert nice value [19,-20] to rlimit style value [1,40] */ |
4779 | int nice_rlim = 20 - nice; |
4780 | |
4781 | return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || |
4782 | capable(CAP_SYS_NICE)); |
4783 | } |
4784 | |
4785 | #ifdef __ARCH_WANT_SYS_NICE |
4786 | |
4787 | /* |
4788 | * sys_nice - change the priority of the current process. |
4789 | * @increment: priority increment |
4790 | * |
4791 | * sys_setpriority is a more generic, but much slower function that |
4792 | * does similar things. |
4793 | */ |
4794 | SYSCALL_DEFINE1(nice, int, increment) |
4795 | { |
4796 | long nice, retval; |
4797 | |
4798 | /* |
4799 | * Setpriority might change our priority at the same moment. |
4800 | * We don't have to worry. Conceptually one call occurs first |
4801 | * and we have a single winner. |
4802 | */ |
4803 | if (increment < -40) |
4804 | increment = -40; |
4805 | if (increment > 40) |
4806 | increment = 40; |
4807 | |
4808 | nice = TASK_NICE(current) + increment; |
4809 | if (nice < -20) |
4810 | nice = -20; |
4811 | if (nice > 19) |
4812 | nice = 19; |
4813 | |
4814 | if (increment < 0 && !can_nice(current, nice)) |
4815 | return -EPERM; |
4816 | |
4817 | retval = security_task_setnice(current, nice); |
4818 | if (retval) |
4819 | return retval; |
4820 | |
4821 | set_user_nice(current, nice); |
4822 | return 0; |
4823 | } |
4824 | |
4825 | #endif |
4826 | |
4827 | /** |
4828 | * task_prio - return the priority value of a given task. |
4829 | * @p: the task in question. |
4830 | * |
4831 | * This is the priority value as seen by users in /proc. |
4832 | * RT tasks are offset by -200. Normal tasks are centered |
4833 | * around 0, value goes from -16 to +15. |
4834 | */ |
4835 | int task_prio(const struct task_struct *p) |
4836 | { |
4837 | return p->prio - MAX_RT_PRIO; |
4838 | } |
4839 | |
4840 | /** |
4841 | * task_nice - return the nice value of a given task. |
4842 | * @p: the task in question. |
4843 | */ |
4844 | int task_nice(const struct task_struct *p) |
4845 | { |
4846 | return TASK_NICE(p); |
4847 | } |
4848 | EXPORT_SYMBOL(task_nice); |
4849 | |
4850 | /** |
4851 | * idle_cpu - is a given cpu idle currently? |
4852 | * @cpu: the processor in question. |
4853 | */ |
4854 | int idle_cpu(int cpu) |
4855 | { |
4856 | return cpu_curr(cpu) == cpu_rq(cpu)->idle; |
4857 | } |
4858 | |
4859 | /** |
4860 | * idle_task - return the idle task for a given cpu. |
4861 | * @cpu: the processor in question. |
4862 | */ |
4863 | struct task_struct *idle_task(int cpu) |
4864 | { |
4865 | return cpu_rq(cpu)->idle; |
4866 | } |
4867 | |
4868 | /** |
4869 | * find_process_by_pid - find a process with a matching PID value. |
4870 | * @pid: the pid in question. |
4871 | */ |
4872 | static struct task_struct *find_process_by_pid(pid_t pid) |
4873 | { |
4874 | return pid ? find_task_by_vpid(pid) : current; |
4875 | } |
4876 | |
4877 | /* Actually do priority change: must hold rq lock. */ |
4878 | static void |
4879 | __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) |
4880 | { |
4881 | BUG_ON(p->se.on_rq); |
4882 | |
4883 | p->policy = policy; |
4884 | p->rt_priority = prio; |
4885 | p->normal_prio = normal_prio(p); |
4886 | /* we are holding p->pi_lock already */ |
4887 | p->prio = rt_mutex_getprio(p); |
4888 | if (rt_prio(p->prio)) |
4889 | p->sched_class = &rt_sched_class; |
4890 | else |
4891 | p->sched_class = &fair_sched_class; |
4892 | set_load_weight(p); |
4893 | } |
4894 | |
4895 | /* |
4896 | * check the target process has a UID that matches the current process's |
4897 | */ |
4898 | static bool check_same_owner(struct task_struct *p) |
4899 | { |
4900 | const struct cred *cred = current_cred(), *pcred; |
4901 | bool match; |
4902 | |
4903 | rcu_read_lock(); |
4904 | pcred = __task_cred(p); |
4905 | if (cred->user->user_ns == pcred->user->user_ns) |
4906 | match = (cred->euid == pcred->euid || |
4907 | cred->euid == pcred->uid); |
4908 | else |
4909 | match = false; |
4910 | rcu_read_unlock(); |
4911 | return match; |
4912 | } |
4913 | |
4914 | static int __sched_setscheduler(struct task_struct *p, int policy, |
4915 | const struct sched_param *param, bool user) |
4916 | { |
4917 | int retval, oldprio, oldpolicy = -1, on_rq, running; |
4918 | unsigned long flags; |
4919 | const struct sched_class *prev_class; |
4920 | struct rq *rq; |
4921 | int reset_on_fork; |
4922 | |
4923 | /* may grab non-irq protected spin_locks */ |
4924 | BUG_ON(in_interrupt()); |
4925 | recheck: |
4926 | /* double check policy once rq lock held */ |
4927 | if (policy < 0) { |
4928 | reset_on_fork = p->sched_reset_on_fork; |
4929 | policy = oldpolicy = p->policy; |
4930 | } else { |
4931 | reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); |
4932 | policy &= ~SCHED_RESET_ON_FORK; |
4933 | |
4934 | if (policy != SCHED_FIFO && policy != SCHED_RR && |
4935 | policy != SCHED_NORMAL && policy != SCHED_BATCH && |
4936 | policy != SCHED_IDLE) |
4937 | return -EINVAL; |
4938 | } |
4939 | |
4940 | /* |
4941 | * Valid priorities for SCHED_FIFO and SCHED_RR are |
4942 | * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, |
4943 | * SCHED_BATCH and SCHED_IDLE is 0. |
4944 | */ |
4945 | if (param->sched_priority < 0 || |
4946 | (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || |
4947 | (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) |
4948 | return -EINVAL; |
4949 | if (rt_policy(policy) != (param->sched_priority != 0)) |
4950 | return -EINVAL; |
4951 | |
4952 | /* |
4953 | * Allow unprivileged RT tasks to decrease priority: |
4954 | */ |
4955 | if (user && !capable(CAP_SYS_NICE)) { |
4956 | if (rt_policy(policy)) { |
4957 | unsigned long rlim_rtprio = |
4958 | task_rlimit(p, RLIMIT_RTPRIO); |
4959 | |
4960 | /* can't set/change the rt policy */ |
4961 | if (policy != p->policy && !rlim_rtprio) |
4962 | return -EPERM; |
4963 | |
4964 | /* can't increase priority */ |
4965 | if (param->sched_priority > p->rt_priority && |
4966 | param->sched_priority > rlim_rtprio) |
4967 | return -EPERM; |
4968 | } |
4969 | |
4970 | /* |
4971 | * Treat SCHED_IDLE as nice 20. Only allow a switch to |
4972 | * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. |
4973 | */ |
4974 | if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { |
4975 | if (!can_nice(p, TASK_NICE(p))) |
4976 | return -EPERM; |
4977 | } |
4978 | |
4979 | /* can't change other user's priorities */ |
4980 | if (!check_same_owner(p)) |
4981 | return -EPERM; |
4982 | |
4983 | /* Normal users shall not reset the sched_reset_on_fork flag */ |
4984 | if (p->sched_reset_on_fork && !reset_on_fork) |
4985 | return -EPERM; |
4986 | } |
4987 | |
4988 | if (user) { |
4989 | retval = security_task_setscheduler(p); |
4990 | if (retval) |
4991 | return retval; |
4992 | } |
4993 | |
4994 | /* |
4995 | * make sure no PI-waiters arrive (or leave) while we are |
4996 | * changing the priority of the task: |
4997 | */ |
4998 | raw_spin_lock_irqsave(&p->pi_lock, flags); |
4999 | /* |
5000 | * To be able to change p->policy safely, the appropriate |
5001 | * runqueue lock must be held. |
5002 | */ |
5003 | rq = __task_rq_lock(p); |
5004 | |
5005 | /* |
5006 | * Changing the policy of the stop threads its a very bad idea |
5007 | */ |
5008 | if (p == rq->stop) { |
5009 | __task_rq_unlock(rq); |
5010 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); |
5011 | return -EINVAL; |
5012 | } |
5013 | |
5014 | /* |
5015 | * If not changing anything there's no need to proceed further: |
5016 | */ |
5017 | if (unlikely(policy == p->policy && (!rt_policy(policy) || |
5018 | param->sched_priority == p->rt_priority))) { |
5019 | |
5020 | __task_rq_unlock(rq); |
5021 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); |
5022 | return 0; |
5023 | } |
5024 | |
5025 | #ifdef CONFIG_RT_GROUP_SCHED |
5026 | if (user) { |
5027 | /* |
5028 | * Do not allow realtime tasks into groups that have no runtime |
5029 | * assigned. |
5030 | */ |
5031 | if (rt_bandwidth_enabled() && rt_policy(policy) && |
5032 | task_group(p)->rt_bandwidth.rt_runtime == 0 && |
5033 | !task_group_is_autogroup(task_group(p))) { |
5034 | __task_rq_unlock(rq); |
5035 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); |
5036 | return -EPERM; |
5037 | } |
5038 | } |
5039 | #endif |
5040 | |
5041 | /* recheck policy now with rq lock held */ |
5042 | if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { |
5043 | policy = oldpolicy = -1; |
5044 | __task_rq_unlock(rq); |
5045 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); |
5046 | goto recheck; |
5047 | } |
5048 | on_rq = p->se.on_rq; |
5049 | running = task_current(rq, p); |
5050 | if (on_rq) |
5051 | deactivate_task(rq, p, 0); |
5052 | if (running) |
5053 | p->sched_class->put_prev_task(rq, p); |
5054 | |
5055 | p->sched_reset_on_fork = reset_on_fork; |
5056 | |
5057 | oldprio = p->prio; |
5058 | prev_class = p->sched_class; |
5059 | __setscheduler(rq, p, policy, param->sched_priority); |
5060 | |
5061 | if (running) |
5062 | p->sched_class->set_curr_task(rq); |
5063 | if (on_rq) |
5064 | activate_task(rq, p, 0); |
5065 | |
5066 | check_class_changed(rq, p, prev_class, oldprio); |
5067 | __task_rq_unlock(rq); |
5068 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); |
5069 | |
5070 | rt_mutex_adjust_pi(p); |
5071 | |
5072 | return 0; |
5073 | } |
5074 | |
5075 | /** |
5076 | * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. |
5077 | * @p: the task in question. |
5078 | * @policy: new policy. |
5079 | * @param: structure containing the new RT priority. |
5080 | * |
5081 | * NOTE that the task may be already dead. |
5082 | */ |
5083 | int sched_setscheduler(struct task_struct *p, int policy, |
5084 | const struct sched_param *param) |
5085 | { |
5086 | return __sched_setscheduler(p, policy, param, true); |
5087 | } |
5088 | EXPORT_SYMBOL_GPL(sched_setscheduler); |
5089 | |
5090 | /** |
5091 | * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. |
5092 | * @p: the task in question. |
5093 | * @policy: new policy. |
5094 | * @param: structure containing the new RT priority. |
5095 | * |
5096 | * Just like sched_setscheduler, only don't bother checking if the |
5097 | * current context has permission. For example, this is needed in |
5098 | * stop_machine(): we create temporary high priority worker threads, |
5099 | * but our caller might not have that capability. |
5100 | */ |
5101 | int sched_setscheduler_nocheck(struct task_struct *p, int policy, |
5102 | const struct sched_param *param) |
5103 | { |
5104 | return __sched_setscheduler(p, policy, param, false); |
5105 | } |
5106 | |
5107 | static int |
5108 | do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) |
5109 | { |
5110 | struct sched_param lparam; |
5111 | struct task_struct *p; |
5112 | int retval; |
5113 | |
5114 | if (!param || pid < 0) |
5115 | return -EINVAL; |
5116 | if (copy_from_user(&lparam, param, sizeof(struct sched_param))) |
5117 | return -EFAULT; |
5118 | |
5119 | rcu_read_lock(); |
5120 | retval = -ESRCH; |
5121 | p = find_process_by_pid(pid); |
5122 | if (p != NULL) |
5123 | retval = sched_setscheduler(p, policy, &lparam); |
5124 | rcu_read_unlock(); |
5125 | |
5126 | return retval; |
5127 | } |
5128 | |
5129 | /** |
5130 | * sys_sched_setscheduler - set/change the scheduler policy and RT priority |
5131 | * @pid: the pid in question. |
5132 | * @policy: new policy. |
5133 | * @param: structure containing the new RT priority. |
5134 | */ |
5135 | SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, |
5136 | struct sched_param __user *, param) |
5137 | { |
5138 | /* negative values for policy are not valid */ |
5139 | if (policy < 0) |
5140 | return -EINVAL; |
5141 | |
5142 | return do_sched_setscheduler(pid, policy, param); |
5143 | } |
5144 | |
5145 | /** |
5146 | * sys_sched_setparam - set/change the RT priority of a thread |
5147 | * @pid: the pid in question. |
5148 | * @param: structure containing the new RT priority. |
5149 | */ |
5150 | SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) |
5151 | { |
5152 | return do_sched_setscheduler(pid, -1, param); |
5153 | } |
5154 | |
5155 | /** |
5156 | * sys_sched_getscheduler - get the policy (scheduling class) of a thread |
5157 | * @pid: the pid in question. |
5158 | */ |
5159 | SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) |
5160 | { |
5161 | struct task_struct *p; |
5162 | int retval; |
5163 | |
5164 | if (pid < 0) |
5165 | return -EINVAL; |
5166 | |
5167 | retval = -ESRCH; |
5168 | rcu_read_lock(); |
5169 | p = find_process_by_pid(pid); |
5170 | if (p) { |
5171 | retval = security_task_getscheduler(p); |
5172 | if (!retval) |
5173 | retval = p->policy |
5174 | | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); |
5175 | } |
5176 | rcu_read_unlock(); |
5177 | return retval; |
5178 | } |
5179 | |
5180 | /** |
5181 | * sys_sched_getparam - get the RT priority of a thread |
5182 | * @pid: the pid in question. |
5183 | * @param: structure containing the RT priority. |
5184 | */ |
5185 | SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) |
5186 | { |
5187 | struct sched_param lp; |
5188 | struct task_struct *p; |
5189 | int retval; |
5190 | |
5191 | if (!param || pid < 0) |
5192 | return -EINVAL; |
5193 | |
5194 | rcu_read_lock(); |
5195 | p = find_process_by_pid(pid); |
5196 | retval = -ESRCH; |
5197 | if (!p) |
5198 | goto out_unlock; |
5199 | |
5200 | retval = security_task_getscheduler(p); |
5201 | if (retval) |
5202 | goto out_unlock; |
5203 | |
5204 | lp.sched_priority = p->rt_priority; |
5205 | rcu_read_unlock(); |
5206 | |
5207 | /* |
5208 | * This one might sleep, we cannot do it with a spinlock held ... |
5209 | */ |
5210 | retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; |
5211 | |
5212 | return retval; |
5213 | |
5214 | out_unlock: |
5215 | rcu_read_unlock(); |
5216 | return retval; |
5217 | } |
5218 | |
5219 | long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) |
5220 | { |
5221 | cpumask_var_t cpus_allowed, new_mask; |
5222 | struct task_struct *p; |
5223 | int retval; |
5224 | |
5225 | get_online_cpus(); |
5226 | rcu_read_lock(); |
5227 | |
5228 | p = find_process_by_pid(pid); |
5229 | if (!p) { |
5230 | rcu_read_unlock(); |
5231 | put_online_cpus(); |
5232 | return -ESRCH; |
5233 | } |
5234 | |
5235 | /* Prevent p going away */ |
5236 | get_task_struct(p); |
5237 | rcu_read_unlock(); |
5238 | |
5239 | if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { |
5240 | retval = -ENOMEM; |
5241 | goto out_put_task; |
5242 | } |
5243 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { |
5244 | retval = -ENOMEM; |
5245 | goto out_free_cpus_allowed; |
5246 | } |
5247 | retval = -EPERM; |
5248 | if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE)) |
5249 | goto out_unlock; |
5250 | |
5251 | retval = security_task_setscheduler(p); |
5252 | if (retval) |
5253 | goto out_unlock; |
5254 | |
5255 | cpuset_cpus_allowed(p, cpus_allowed); |
5256 | cpumask_and(new_mask, in_mask, cpus_allowed); |
5257 | again: |
5258 | retval = set_cpus_allowed_ptr(p, new_mask); |
5259 | |
5260 | if (!retval) { |
5261 | cpuset_cpus_allowed(p, cpus_allowed); |
5262 | if (!cpumask_subset(new_mask, cpus_allowed)) { |
5263 | /* |
5264 | * We must have raced with a concurrent cpuset |
5265 | * update. Just reset the cpus_allowed to the |
5266 | * cpuset's cpus_allowed |
5267 | */ |
5268 | cpumask_copy(new_mask, cpus_allowed); |
5269 | goto again; |
5270 | } |
5271 | } |
5272 | out_unlock: |
5273 | free_cpumask_var(new_mask); |
5274 | out_free_cpus_allowed: |
5275 | free_cpumask_var(cpus_allowed); |
5276 | out_put_task: |
5277 | put_task_struct(p); |
5278 | put_online_cpus(); |
5279 | return retval; |
5280 | } |
5281 | |
5282 | static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, |
5283 | struct cpumask *new_mask) |
5284 | { |
5285 | if (len < cpumask_size()) |
5286 | cpumask_clear(new_mask); |
5287 | else if (len > cpumask_size()) |
5288 | len = cpumask_size(); |
5289 | |
5290 | return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; |
5291 | } |
5292 | |
5293 | /** |
5294 | * sys_sched_setaffinity - set the cpu affinity of a process |
5295 | * @pid: pid of the process |
5296 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr |
5297 | * @user_mask_ptr: user-space pointer to the new cpu mask |
5298 | */ |
5299 | SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, |
5300 | unsigned long __user *, user_mask_ptr) |
5301 | { |
5302 | cpumask_var_t new_mask; |
5303 | int retval; |
5304 | |
5305 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) |
5306 | return -ENOMEM; |
5307 | |
5308 | retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); |
5309 | if (retval == 0) |
5310 | retval = sched_setaffinity(pid, new_mask); |
5311 | free_cpumask_var(new_mask); |
5312 | return retval; |
5313 | } |
5314 | |
5315 | long sched_getaffinity(pid_t pid, struct cpumask *mask) |
5316 | { |
5317 | struct task_struct *p; |
5318 | unsigned long flags; |
5319 | struct rq *rq; |
5320 | int retval; |
5321 | |
5322 | get_online_cpus(); |
5323 | rcu_read_lock(); |
5324 | |
5325 | retval = -ESRCH; |
5326 | p = find_process_by_pid(pid); |
5327 | if (!p) |
5328 | goto out_unlock; |
5329 | |
5330 | retval = security_task_getscheduler(p); |
5331 | if (retval) |
5332 | goto out_unlock; |
5333 | |
5334 | rq = task_rq_lock(p, &flags); |
5335 | cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); |
5336 | task_rq_unlock(rq, &flags); |
5337 | |
5338 | out_unlock: |
5339 | rcu_read_unlock(); |
5340 | put_online_cpus(); |
5341 | |
5342 | return retval; |
5343 | } |
5344 | |
5345 | /** |
5346 | * sys_sched_getaffinity - get the cpu affinity of a process |
5347 | * @pid: pid of the process |
5348 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr |
5349 | * @user_mask_ptr: user-space pointer to hold the current cpu mask |
5350 | */ |
5351 | SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, |
5352 | unsigned long __user *, user_mask_ptr) |
5353 | { |
5354 | int ret; |
5355 | cpumask_var_t mask; |
5356 | |
5357 | if ((len * BITS_PER_BYTE) < nr_cpu_ids) |
5358 | return -EINVAL; |
5359 | if (len & (sizeof(unsigned long)-1)) |
5360 | return -EINVAL; |
5361 | |
5362 | if (!alloc_cpumask_var(&mask, GFP_KERNEL)) |
5363 | return -ENOMEM; |
5364 | |
5365 | ret = sched_getaffinity(pid, mask); |
5366 | if (ret == 0) { |
5367 | size_t retlen = min_t(size_t, len, cpumask_size()); |
5368 | |
5369 | if (copy_to_user(user_mask_ptr, mask, retlen)) |
5370 | ret = -EFAULT; |
5371 | else |
5372 | ret = retlen; |
5373 | } |
5374 | free_cpumask_var(mask); |
5375 | |
5376 | return ret; |
5377 | } |
5378 | |
5379 | /** |
5380 | * sys_sched_yield - yield the current processor to other threads. |
5381 | * |
5382 | * This function yields the current CPU to other tasks. If there are no |
5383 | * other threads running on this CPU then this function will return. |
5384 | */ |
5385 | SYSCALL_DEFINE0(sched_yield) |
5386 | { |
5387 | struct rq *rq = this_rq_lock(); |
5388 | |
5389 | schedstat_inc(rq, yld_count); |
5390 | current->sched_class->yield_task(rq); |
5391 | |
5392 | /* |
5393 | * Since we are going to call schedule() anyway, there's |
5394 | * no need to preempt or enable interrupts: |
5395 | */ |
5396 | __release(rq->lock); |
5397 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); |
5398 | do_raw_spin_unlock(&rq->lock); |
5399 | preempt_enable_no_resched(); |
5400 | |
5401 | schedule(); |
5402 | |
5403 | return 0; |
5404 | } |
5405 | |
5406 | static inline int should_resched(void) |
5407 | { |
5408 | return need_resched() && !(preempt_count() & PREEMPT_ACTIVE); |
5409 | } |
5410 | |
5411 | static void __cond_resched(void) |
5412 | { |
5413 | add_preempt_count(PREEMPT_ACTIVE); |
5414 | schedule(); |
5415 | sub_preempt_count(PREEMPT_ACTIVE); |
5416 | } |
5417 | |
5418 | int __sched _cond_resched(void) |
5419 | { |
5420 | if (should_resched()) { |
5421 | __cond_resched(); |
5422 | return 1; |
5423 | } |
5424 | return 0; |
5425 | } |
5426 | EXPORT_SYMBOL(_cond_resched); |
5427 | |
5428 | /* |
5429 | * __cond_resched_lock() - if a reschedule is pending, drop the given lock, |
5430 | * call schedule, and on return reacquire the lock. |
5431 | * |
5432 | * This works OK both with and without CONFIG_PREEMPT. We do strange low-level |
5433 | * operations here to prevent schedule() from being called twice (once via |
5434 | * spin_unlock(), once by hand). |
5435 | */ |
5436 | int __cond_resched_lock(spinlock_t *lock) |
5437 | { |
5438 | int resched = should_resched(); |
5439 | int ret = 0; |
5440 | |
5441 | lockdep_assert_held(lock); |
5442 | |
5443 | if (spin_needbreak(lock) || resched) { |
5444 | spin_unlock(lock); |
5445 | if (resched) |
5446 | __cond_resched(); |
5447 | else |
5448 | cpu_relax(); |
5449 | ret = 1; |
5450 | spin_lock(lock); |
5451 | } |
5452 | return ret; |
5453 | } |
5454 | EXPORT_SYMBOL(__cond_resched_lock); |
5455 | |
5456 | int __sched __cond_resched_softirq(void) |
5457 | { |
5458 | BUG_ON(!in_softirq()); |
5459 | |
5460 | if (should_resched()) { |
5461 | local_bh_enable(); |
5462 | __cond_resched(); |
5463 | local_bh_disable(); |
5464 | return 1; |
5465 | } |
5466 | return 0; |
5467 | } |
5468 | EXPORT_SYMBOL(__cond_resched_softirq); |
5469 | |
5470 | /** |
5471 | * yield - yield the current processor to other threads. |
5472 | * |
5473 | * This is a shortcut for kernel-space yielding - it marks the |
5474 | * thread runnable and calls sys_sched_yield(). |
5475 | */ |
5476 | void __sched yield(void) |
5477 | { |
5478 | set_current_state(TASK_RUNNING); |
5479 | sys_sched_yield(); |
5480 | } |
5481 | EXPORT_SYMBOL(yield); |
5482 | |
5483 | /** |
5484 | * yield_to - yield the current processor to another thread in |
5485 | * your thread group, or accelerate that thread toward the |
5486 | * processor it's on. |
5487 | * @p: target task |
5488 | * @preempt: whether task preemption is allowed or not |
5489 | * |
5490 | * It's the caller's job to ensure that the target task struct |
5491 | * can't go away on us before we can do any checks. |
5492 | * |
5493 | * Returns true if we indeed boosted the target task. |
5494 | */ |
5495 | bool __sched yield_to(struct task_struct *p, bool preempt) |
5496 | { |
5497 | struct task_struct *curr = current; |
5498 | struct rq *rq, *p_rq; |
5499 | unsigned long flags; |
5500 | bool yielded = 0; |
5501 | |
5502 | local_irq_save(flags); |
5503 | rq = this_rq(); |
5504 | |
5505 | again: |
5506 | p_rq = task_rq(p); |
5507 | double_rq_lock(rq, p_rq); |
5508 | while (task_rq(p) != p_rq) { |
5509 | double_rq_unlock(rq, p_rq); |
5510 | goto again; |
5511 | } |
5512 | |
5513 | if (!curr->sched_class->yield_to_task) |
5514 | goto out; |
5515 | |
5516 | if (curr->sched_class != p->sched_class) |
5517 | goto out; |
5518 | |
5519 | if (task_running(p_rq, p) || p->state) |
5520 | goto out; |
5521 | |
5522 | yielded = curr->sched_class->yield_to_task(rq, p, preempt); |
5523 | if (yielded) { |
5524 | schedstat_inc(rq, yld_count); |
5525 | /* |
5526 | * Make p's CPU reschedule; pick_next_entity takes care of |
5527 | * fairness. |
5528 | */ |
5529 | if (preempt && rq != p_rq) |
5530 | resched_task(p_rq->curr); |
5531 | } |
5532 | |
5533 | out: |
5534 | double_rq_unlock(rq, p_rq); |
5535 | local_irq_restore(flags); |
5536 | |
5537 | if (yielded) |
5538 | schedule(); |
5539 | |
5540 | return yielded; |
5541 | } |
5542 | EXPORT_SYMBOL_GPL(yield_to); |
5543 | |
5544 | /* |
5545 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so |
5546 | * that process accounting knows that this is a task in IO wait state. |
5547 | */ |
5548 | void __sched io_schedule(void) |
5549 | { |
5550 | struct rq *rq = raw_rq(); |
5551 | |
5552 | delayacct_blkio_start(); |
5553 | atomic_inc(&rq->nr_iowait); |
5554 | blk_flush_plug(current); |
5555 | current->in_iowait = 1; |
5556 | schedule(); |
5557 | current->in_iowait = 0; |
5558 | atomic_dec(&rq->nr_iowait); |
5559 | delayacct_blkio_end(); |
5560 | } |
5561 | EXPORT_SYMBOL(io_schedule); |
5562 | |
5563 | long __sched io_schedule_timeout(long timeout) |
5564 | { |
5565 | struct rq *rq = raw_rq(); |
5566 | long ret; |
5567 | |
5568 | delayacct_blkio_start(); |
5569 | atomic_inc(&rq->nr_iowait); |
5570 | blk_flush_plug(current); |
5571 | current->in_iowait = 1; |
5572 | ret = schedule_timeout(timeout); |
5573 | current->in_iowait = 0; |
5574 | atomic_dec(&rq->nr_iowait); |
5575 | delayacct_blkio_end(); |
5576 | return ret; |
5577 | } |
5578 | |
5579 | /** |
5580 | * sys_sched_get_priority_max - return maximum RT priority. |
5581 | * @policy: scheduling class. |
5582 | * |
5583 | * this syscall returns the maximum rt_priority that can be used |
5584 | * by a given scheduling class. |
5585 | */ |
5586 | SYSCALL_DEFINE1(sched_get_priority_max, int, policy) |
5587 | { |
5588 | int ret = -EINVAL; |
5589 | |
5590 | switch (policy) { |
5591 | case SCHED_FIFO: |
5592 | case SCHED_RR: |
5593 | ret = MAX_USER_RT_PRIO-1; |
5594 | break; |
5595 | case SCHED_NORMAL: |
5596 | case SCHED_BATCH: |
5597 | case SCHED_IDLE: |
5598 | ret = 0; |
5599 | break; |
5600 | } |
5601 | return ret; |
5602 | } |
5603 | |
5604 | /** |
5605 | * sys_sched_get_priority_min - return minimum RT priority. |
5606 | * @policy: scheduling class. |
5607 | * |
5608 | * this syscall returns the minimum rt_priority that can be used |
5609 | * by a given scheduling class. |
5610 | */ |
5611 | SYSCALL_DEFINE1(sched_get_priority_min, int, policy) |
5612 | { |
5613 | int ret = -EINVAL; |
5614 | |
5615 | switch (policy) { |
5616 | case SCHED_FIFO: |
5617 | case SCHED_RR: |
5618 | ret = 1; |
5619 | break; |
5620 | case SCHED_NORMAL: |
5621 | case SCHED_BATCH: |
5622 | case SCHED_IDLE: |
5623 | ret = 0; |
5624 | } |
5625 | return ret; |
5626 | } |
5627 | |
5628 | /** |
5629 | * sys_sched_rr_get_interval - return the default timeslice of a process. |
5630 | * @pid: pid of the process. |
5631 | * @interval: userspace pointer to the timeslice value. |
5632 | * |
5633 | * this syscall writes the default timeslice value of a given process |
5634 | * into the user-space timespec buffer. A value of '0' means infinity. |
5635 | */ |
5636 | SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, |
5637 | struct timespec __user *, interval) |
5638 | { |
5639 | struct task_struct *p; |
5640 | unsigned int time_slice; |
5641 | unsigned long flags; |
5642 | struct rq *rq; |
5643 | int retval; |
5644 | struct timespec t; |
5645 | |
5646 | if (pid < 0) |
5647 | return -EINVAL; |
5648 | |
5649 | retval = -ESRCH; |
5650 | rcu_read_lock(); |
5651 | p = find_process_by_pid(pid); |
5652 | if (!p) |
5653 | goto out_unlock; |
5654 | |
5655 | retval = security_task_getscheduler(p); |
5656 | if (retval) |
5657 | goto out_unlock; |
5658 | |
5659 | rq = task_rq_lock(p, &flags); |
5660 | time_slice = p->sched_class->get_rr_interval(rq, p); |
5661 | task_rq_unlock(rq, &flags); |
5662 | |
5663 | rcu_read_unlock(); |
5664 | jiffies_to_timespec(time_slice, &t); |
5665 | retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; |
5666 | return retval; |
5667 | |
5668 | out_unlock: |
5669 | rcu_read_unlock(); |
5670 | return retval; |
5671 | } |
5672 | |
5673 | static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; |
5674 | |
5675 | void sched_show_task(struct task_struct *p) |
5676 | { |
5677 | unsigned long free = 0; |
5678 | unsigned state; |
5679 | |
5680 | state = p->state ? __ffs(p->state) + 1 : 0; |
5681 | printk(KERN_INFO "%-15.15s %c", p->comm, |
5682 | state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); |
5683 | #if BITS_PER_LONG == 32 |
5684 | if (state == TASK_RUNNING) |
5685 | printk(KERN_CONT " running "); |
5686 | else |
5687 | printk(KERN_CONT " %08lx ", thread_saved_pc(p)); |
5688 | #else |
5689 | if (state == TASK_RUNNING) |
5690 | printk(KERN_CONT " running task "); |
5691 | else |
5692 | printk(KERN_CONT " %016lx ", thread_saved_pc(p)); |
5693 | #endif |
5694 | #ifdef CONFIG_DEBUG_STACK_USAGE |
5695 | free = stack_not_used(p); |
5696 | #endif |
5697 | printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, |
5698 | task_pid_nr(p), task_pid_nr(p->real_parent), |
5699 | (unsigned long)task_thread_info(p)->flags); |
5700 | |
5701 | show_stack(p, NULL); |
5702 | } |
5703 | |
5704 | void show_state_filter(unsigned long state_filter) |
5705 | { |
5706 | struct task_struct *g, *p; |
5707 | |
5708 | #if BITS_PER_LONG == 32 |
5709 | printk(KERN_INFO |
5710 | " task PC stack pid father\n"); |
5711 | #else |
5712 | printk(KERN_INFO |
5713 | " task PC stack pid father\n"); |
5714 | #endif |
5715 | read_lock(&tasklist_lock); |
5716 | do_each_thread(g, p) { |
5717 | /* |
5718 | * reset the NMI-timeout, listing all files on a slow |
5719 | * console might take a lot of time: |
5720 | */ |
5721 | touch_nmi_watchdog(); |
5722 | if (!state_filter || (p->state & state_filter)) |
5723 | sched_show_task(p); |
5724 | } while_each_thread(g, p); |
5725 | |
5726 | touch_all_softlockup_watchdogs(); |
5727 | |
5728 | #ifdef CONFIG_SCHED_DEBUG |
5729 | sysrq_sched_debug_show(); |
5730 | #endif |
5731 | read_unlock(&tasklist_lock); |
5732 | /* |
5733 | * Only show locks if all tasks are dumped: |
5734 | */ |
5735 | if (!state_filter) |
5736 | debug_show_all_locks(); |
5737 | } |
5738 | |
5739 | void __cpuinit init_idle_bootup_task(struct task_struct *idle) |
5740 | { |
5741 | idle->sched_class = &idle_sched_class; |
5742 | } |
5743 | |
5744 | /** |
5745 | * init_idle - set up an idle thread for a given CPU |
5746 | * @idle: task in question |
5747 | * @cpu: cpu the idle task belongs to |
5748 | * |
5749 | * NOTE: this function does not set the idle thread's NEED_RESCHED |
5750 | * flag, to make booting more robust. |
5751 | */ |
5752 | void __cpuinit init_idle(struct task_struct *idle, int cpu) |
5753 | { |
5754 | struct rq *rq = cpu_rq(cpu); |
5755 | unsigned long flags; |
5756 | |
5757 | raw_spin_lock_irqsave(&rq->lock, flags); |
5758 | |
5759 | __sched_fork(idle); |
5760 | idle->state = TASK_RUNNING; |
5761 | idle->se.exec_start = sched_clock(); |
5762 | |
5763 | cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu)); |
5764 | /* |
5765 | * We're having a chicken and egg problem, even though we are |
5766 | * holding rq->lock, the cpu isn't yet set to this cpu so the |
5767 | * lockdep check in task_group() will fail. |
5768 | * |
5769 | * Similar case to sched_fork(). / Alternatively we could |
5770 | * use task_rq_lock() here and obtain the other rq->lock. |
5771 | * |
5772 | * Silence PROVE_RCU |
5773 | */ |
5774 | rcu_read_lock(); |
5775 | __set_task_cpu(idle, cpu); |
5776 | rcu_read_unlock(); |
5777 | |
5778 | rq->curr = rq->idle = idle; |
5779 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
5780 | idle->oncpu = 1; |
5781 | #endif |
5782 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
5783 | |
5784 | /* Set the preempt count _outside_ the spinlocks! */ |
5785 | #if defined(CONFIG_PREEMPT) |
5786 | task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0); |
5787 | #else |
5788 | task_thread_info(idle)->preempt_count = 0; |
5789 | #endif |
5790 | /* |
5791 | * The idle tasks have their own, simple scheduling class: |
5792 | */ |
5793 | idle->sched_class = &idle_sched_class; |
5794 | ftrace_graph_init_idle_task(idle, cpu); |
5795 | } |
5796 | |
5797 | /* |
5798 | * In a system that switches off the HZ timer nohz_cpu_mask |
5799 | * indicates which cpus entered this state. This is used |
5800 | * in the rcu update to wait only for active cpus. For system |
5801 | * which do not switch off the HZ timer nohz_cpu_mask should |
5802 | * always be CPU_BITS_NONE. |
5803 | */ |
5804 | cpumask_var_t nohz_cpu_mask; |
5805 | |
5806 | /* |
5807 | * Increase the granularity value when there are more CPUs, |
5808 | * because with more CPUs the 'effective latency' as visible |
5809 | * to users decreases. But the relationship is not linear, |
5810 | * so pick a second-best guess by going with the log2 of the |
5811 | * number of CPUs. |
5812 | * |
5813 | * This idea comes from the SD scheduler of Con Kolivas: |
5814 | */ |
5815 | static int get_update_sysctl_factor(void) |
5816 | { |
5817 | unsigned int cpus = min_t(int, num_online_cpus(), 8); |
5818 | unsigned int factor; |
5819 | |
5820 | switch (sysctl_sched_tunable_scaling) { |
5821 | case SCHED_TUNABLESCALING_NONE: |
5822 | factor = 1; |
5823 | break; |
5824 | case SCHED_TUNABLESCALING_LINEAR: |
5825 | factor = cpus; |
5826 | break; |
5827 | case SCHED_TUNABLESCALING_LOG: |
5828 | default: |
5829 | factor = 1 + ilog2(cpus); |
5830 | break; |
5831 | } |
5832 | |
5833 | return factor; |
5834 | } |
5835 | |
5836 | static void update_sysctl(void) |
5837 | { |
5838 | unsigned int factor = get_update_sysctl_factor(); |
5839 | |
5840 | #define SET_SYSCTL(name) \ |
5841 | (sysctl_##name = (factor) * normalized_sysctl_##name) |
5842 | SET_SYSCTL(sched_min_granularity); |
5843 | SET_SYSCTL(sched_latency); |
5844 | SET_SYSCTL(sched_wakeup_granularity); |
5845 | #undef SET_SYSCTL |
5846 | } |
5847 | |
5848 | static inline void sched_init_granularity(void) |
5849 | { |
5850 | update_sysctl(); |
5851 | } |
5852 | |
5853 | #ifdef CONFIG_SMP |
5854 | /* |
5855 | * This is how migration works: |
5856 | * |
5857 | * 1) we invoke migration_cpu_stop() on the target CPU using |
5858 | * stop_one_cpu(). |
5859 | * 2) stopper starts to run (implicitly forcing the migrated thread |
5860 | * off the CPU) |
5861 | * 3) it checks whether the migrated task is still in the wrong runqueue. |
5862 | * 4) if it's in the wrong runqueue then the migration thread removes |
5863 | * it and puts it into the right queue. |
5864 | * 5) stopper completes and stop_one_cpu() returns and the migration |
5865 | * is done. |
5866 | */ |
5867 | |
5868 | /* |
5869 | * Change a given task's CPU affinity. Migrate the thread to a |
5870 | * proper CPU and schedule it away if the CPU it's executing on |
5871 | * is removed from the allowed bitmask. |
5872 | * |
5873 | * NOTE: the caller must have a valid reference to the task, the |
5874 | * task must not exit() & deallocate itself prematurely. The |
5875 | * call is not atomic; no spinlocks may be held. |
5876 | */ |
5877 | int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) |
5878 | { |
5879 | unsigned long flags; |
5880 | struct rq *rq; |
5881 | unsigned int dest_cpu; |
5882 | int ret = 0; |
5883 | |
5884 | /* |
5885 | * Serialize against TASK_WAKING so that ttwu() and wunt() can |
5886 | * drop the rq->lock and still rely on ->cpus_allowed. |
5887 | */ |
5888 | again: |
5889 | while (task_is_waking(p)) |
5890 | cpu_relax(); |
5891 | rq = task_rq_lock(p, &flags); |
5892 | if (task_is_waking(p)) { |
5893 | task_rq_unlock(rq, &flags); |
5894 | goto again; |
5895 | } |
5896 | |
5897 | if (!cpumask_intersects(new_mask, cpu_active_mask)) { |
5898 | ret = -EINVAL; |
5899 | goto out; |
5900 | } |
5901 | |
5902 | if (unlikely((p->flags & PF_THREAD_BOUND) && p != current && |
5903 | !cpumask_equal(&p->cpus_allowed, new_mask))) { |
5904 | ret = -EINVAL; |
5905 | goto out; |
5906 | } |
5907 | |
5908 | if (p->sched_class->set_cpus_allowed) |
5909 | p->sched_class->set_cpus_allowed(p, new_mask); |
5910 | else { |
5911 | cpumask_copy(&p->cpus_allowed, new_mask); |
5912 | p->rt.nr_cpus_allowed = cpumask_weight(new_mask); |
5913 | } |
5914 | |
5915 | /* Can the task run on the task's current CPU? If so, we're done */ |
5916 | if (cpumask_test_cpu(task_cpu(p), new_mask)) |
5917 | goto out; |
5918 | |
5919 | dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); |
5920 | if (migrate_task(p, rq)) { |
5921 | struct migration_arg arg = { p, dest_cpu }; |
5922 | /* Need help from migration thread: drop lock and wait. */ |
5923 | task_rq_unlock(rq, &flags); |
5924 | stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); |
5925 | tlb_migrate_finish(p->mm); |
5926 | return 0; |
5927 | } |
5928 | out: |
5929 | task_rq_unlock(rq, &flags); |
5930 | |
5931 | return ret; |
5932 | } |
5933 | EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); |
5934 | |
5935 | /* |
5936 | * Move (not current) task off this cpu, onto dest cpu. We're doing |
5937 | * this because either it can't run here any more (set_cpus_allowed() |
5938 | * away from this CPU, or CPU going down), or because we're |
5939 | * attempting to rebalance this task on exec (sched_exec). |
5940 | * |
5941 | * So we race with normal scheduler movements, but that's OK, as long |
5942 | * as the task is no longer on this CPU. |
5943 | * |
5944 | * Returns non-zero if task was successfully migrated. |
5945 | */ |
5946 | static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) |
5947 | { |
5948 | struct rq *rq_dest, *rq_src; |
5949 | int ret = 0; |
5950 | |
5951 | if (unlikely(!cpu_active(dest_cpu))) |
5952 | return ret; |
5953 | |
5954 | rq_src = cpu_rq(src_cpu); |
5955 | rq_dest = cpu_rq(dest_cpu); |
5956 | |
5957 | double_rq_lock(rq_src, rq_dest); |
5958 | /* Already moved. */ |
5959 | if (task_cpu(p) != src_cpu) |
5960 | goto done; |
5961 | /* Affinity changed (again). */ |
5962 | if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)) |
5963 | goto fail; |
5964 | |
5965 | /* |
5966 | * If we're not on a rq, the next wake-up will ensure we're |
5967 | * placed properly. |
5968 | */ |
5969 | if (p->se.on_rq) { |
5970 | deactivate_task(rq_src, p, 0); |
5971 | set_task_cpu(p, dest_cpu); |
5972 | activate_task(rq_dest, p, 0); |
5973 | check_preempt_curr(rq_dest, p, 0); |
5974 | } |
5975 | done: |
5976 | ret = 1; |
5977 | fail: |
5978 | double_rq_unlock(rq_src, rq_dest); |
5979 | return ret; |
5980 | } |
5981 | |
5982 | /* |
5983 | * migration_cpu_stop - this will be executed by a highprio stopper thread |
5984 | * and performs thread migration by bumping thread off CPU then |
5985 | * 'pushing' onto another runqueue. |
5986 | */ |
5987 | static int migration_cpu_stop(void *data) |
5988 | { |
5989 | struct migration_arg *arg = data; |
5990 | |
5991 | /* |
5992 | * The original target cpu might have gone down and we might |
5993 | * be on another cpu but it doesn't matter. |
5994 | */ |
5995 | local_irq_disable(); |
5996 | __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); |
5997 | local_irq_enable(); |
5998 | return 0; |
5999 | } |
6000 | |
6001 | #ifdef CONFIG_HOTPLUG_CPU |
6002 | |
6003 | /* |
6004 | * Ensures that the idle task is using init_mm right before its cpu goes |
6005 | * offline. |
6006 | */ |
6007 | void idle_task_exit(void) |
6008 | { |
6009 | struct mm_struct *mm = current->active_mm; |
6010 | |
6011 | BUG_ON(cpu_online(smp_processor_id())); |
6012 | |
6013 | if (mm != &init_mm) |
6014 | switch_mm(mm, &init_mm, current); |
6015 | mmdrop(mm); |
6016 | } |
6017 | |
6018 | /* |
6019 | * While a dead CPU has no uninterruptible tasks queued at this point, |
6020 | * it might still have a nonzero ->nr_uninterruptible counter, because |
6021 | * for performance reasons the counter is not stricly tracking tasks to |
6022 | * their home CPUs. So we just add the counter to another CPU's counter, |
6023 | * to keep the global sum constant after CPU-down: |
6024 | */ |
6025 | static void migrate_nr_uninterruptible(struct rq *rq_src) |
6026 | { |
6027 | struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask)); |
6028 | |
6029 | rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible; |
6030 | rq_src->nr_uninterruptible = 0; |
6031 | } |
6032 | |
6033 | /* |
6034 | * remove the tasks which were accounted by rq from calc_load_tasks. |
6035 | */ |
6036 | static void calc_global_load_remove(struct rq *rq) |
6037 | { |
6038 | atomic_long_sub(rq->calc_load_active, &calc_load_tasks); |
6039 | rq->calc_load_active = 0; |
6040 | } |
6041 | |
6042 | /* |
6043 | * Migrate all tasks from the rq, sleeping tasks will be migrated by |
6044 | * try_to_wake_up()->select_task_rq(). |
6045 | * |
6046 | * Called with rq->lock held even though we'er in stop_machine() and |
6047 | * there's no concurrency possible, we hold the required locks anyway |
6048 | * because of lock validation efforts. |
6049 | */ |
6050 | static void migrate_tasks(unsigned int dead_cpu) |
6051 | { |
6052 | struct rq *rq = cpu_rq(dead_cpu); |
6053 | struct task_struct *next, *stop = rq->stop; |
6054 | int dest_cpu; |
6055 | |
6056 | /* |
6057 | * Fudge the rq selection such that the below task selection loop |
6058 | * doesn't get stuck on the currently eligible stop task. |
6059 | * |
6060 | * We're currently inside stop_machine() and the rq is either stuck |
6061 | * in the stop_machine_cpu_stop() loop, or we're executing this code, |
6062 | * either way we should never end up calling schedule() until we're |
6063 | * done here. |
6064 | */ |
6065 | rq->stop = NULL; |
6066 | |
6067 | for ( ; ; ) { |
6068 | /* |
6069 | * There's this thread running, bail when that's the only |
6070 | * remaining thread. |
6071 | */ |
6072 | if (rq->nr_running == 1) |
6073 | break; |
6074 | |
6075 | next = pick_next_task(rq); |
6076 | BUG_ON(!next); |
6077 | next->sched_class->put_prev_task(rq, next); |
6078 | |
6079 | /* Find suitable destination for @next, with force if needed. */ |
6080 | dest_cpu = select_fallback_rq(dead_cpu, next); |
6081 | raw_spin_unlock(&rq->lock); |
6082 | |
6083 | __migrate_task(next, dead_cpu, dest_cpu); |
6084 | |
6085 | raw_spin_lock(&rq->lock); |
6086 | } |
6087 | |
6088 | rq->stop = stop; |
6089 | } |
6090 | |
6091 | #endif /* CONFIG_HOTPLUG_CPU */ |
6092 | |
6093 | #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) |
6094 | |
6095 | static struct ctl_table sd_ctl_dir[] = { |
6096 | { |
6097 | .procname = "sched_domain", |
6098 | .mode = 0555, |
6099 | }, |
6100 | {} |
6101 | }; |
6102 | |
6103 | static struct ctl_table sd_ctl_root[] = { |
6104 | { |
6105 | .procname = "kernel", |
6106 | .mode = 0555, |
6107 | .child = sd_ctl_dir, |
6108 | }, |
6109 | {} |
6110 | }; |
6111 | |
6112 | static struct ctl_table *sd_alloc_ctl_entry(int n) |
6113 | { |
6114 | struct ctl_table *entry = |
6115 | kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); |
6116 | |
6117 | return entry; |
6118 | } |
6119 | |
6120 | static void sd_free_ctl_entry(struct ctl_table **tablep) |
6121 | { |
6122 | struct ctl_table *entry; |
6123 | |
6124 | /* |
6125 | * In the intermediate directories, both the child directory and |
6126 | * procname are dynamically allocated and could fail but the mode |
6127 | * will always be set. In the lowest directory the names are |
6128 | * static strings and all have proc handlers. |
6129 | */ |
6130 | for (entry = *tablep; entry->mode; entry++) { |
6131 | if (entry->child) |
6132 | sd_free_ctl_entry(&entry->child); |
6133 | if (entry->proc_handler == NULL) |
6134 | kfree(entry->procname); |
6135 | } |
6136 | |
6137 | kfree(*tablep); |
6138 | *tablep = NULL; |
6139 | } |
6140 | |
6141 | static void |
6142 | set_table_entry(struct ctl_table *entry, |
6143 | const char *procname, void *data, int maxlen, |
6144 | mode_t mode, proc_handler *proc_handler) |
6145 | { |
6146 | entry->procname = procname; |
6147 | entry->data = data; |
6148 | entry->maxlen = maxlen; |
6149 | entry->mode = mode; |
6150 | entry->proc_handler = proc_handler; |
6151 | } |
6152 | |
6153 | static struct ctl_table * |
6154 | sd_alloc_ctl_domain_table(struct sched_domain *sd) |
6155 | { |
6156 | struct ctl_table *table = sd_alloc_ctl_entry(13); |
6157 | |
6158 | if (table == NULL) |
6159 | return NULL; |
6160 | |
6161 | set_table_entry(&table[0], "min_interval", &sd->min_interval, |
6162 | sizeof(long), 0644, proc_doulongvec_minmax); |
6163 | set_table_entry(&table[1], "max_interval", &sd->max_interval, |
6164 | sizeof(long), 0644, proc_doulongvec_minmax); |
6165 | set_table_entry(&table[2], "busy_idx", &sd->busy_idx, |
6166 | sizeof(int), 0644, proc_dointvec_minmax); |
6167 | set_table_entry(&table[3], "idle_idx", &sd->idle_idx, |
6168 | sizeof(int), 0644, proc_dointvec_minmax); |
6169 | set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, |
6170 | sizeof(int), 0644, proc_dointvec_minmax); |
6171 | set_table_entry(&table[5], "wake_idx", &sd->wake_idx, |
6172 | sizeof(int), 0644, proc_dointvec_minmax); |
6173 | set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, |
6174 | sizeof(int), 0644, proc_dointvec_minmax); |
6175 | set_table_entry(&table[7], "busy_factor", &sd->busy_factor, |
6176 | sizeof(int), 0644, proc_dointvec_minmax); |
6177 | set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, |
6178 | sizeof(int), 0644, proc_dointvec_minmax); |
6179 | set_table_entry(&table[9], "cache_nice_tries", |
6180 | &sd->cache_nice_tries, |
6181 | sizeof(int), 0644, proc_dointvec_minmax); |
6182 | set_table_entry(&table[10], "flags", &sd->flags, |
6183 | sizeof(int), 0644, proc_dointvec_minmax); |
6184 | set_table_entry(&table[11], "name", sd->name, |
6185 | CORENAME_MAX_SIZE, 0444, proc_dostring); |
6186 | /* &table[12] is terminator */ |
6187 | |
6188 | return table; |
6189 | } |
6190 | |
6191 | static ctl_table *sd_alloc_ctl_cpu_table(int cpu) |
6192 | { |
6193 | struct ctl_table *entry, *table; |
6194 | struct sched_domain *sd; |
6195 | int domain_num = 0, i; |
6196 | char buf[32]; |
6197 | |
6198 | for_each_domain(cpu, sd) |
6199 | domain_num++; |
6200 | entry = table = sd_alloc_ctl_entry(domain_num + 1); |
6201 | if (table == NULL) |
6202 | return NULL; |
6203 | |
6204 | i = 0; |
6205 | for_each_domain(cpu, sd) { |
6206 | snprintf(buf, 32, "domain%d", i); |
6207 | entry->procname = kstrdup(buf, GFP_KERNEL); |
6208 | entry->mode = 0555; |
6209 | entry->child = sd_alloc_ctl_domain_table(sd); |
6210 | entry++; |
6211 | i++; |
6212 | } |
6213 | return table; |
6214 | } |
6215 | |
6216 | static struct ctl_table_header *sd_sysctl_header; |
6217 | static void register_sched_domain_sysctl(void) |
6218 | { |
6219 | int i, cpu_num = num_possible_cpus(); |
6220 | struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); |
6221 | char buf[32]; |
6222 | |
6223 | WARN_ON(sd_ctl_dir[0].child); |
6224 | sd_ctl_dir[0].child = entry; |
6225 | |
6226 | if (entry == NULL) |
6227 | return; |
6228 | |
6229 | for_each_possible_cpu(i) { |
6230 | snprintf(buf, 32, "cpu%d", i); |
6231 | entry->procname = kstrdup(buf, GFP_KERNEL); |
6232 | entry->mode = 0555; |
6233 | entry->child = sd_alloc_ctl_cpu_table(i); |
6234 | entry++; |
6235 | } |
6236 | |
6237 | WARN_ON(sd_sysctl_header); |
6238 | sd_sysctl_header = register_sysctl_table(sd_ctl_root); |
6239 | } |
6240 | |
6241 | /* may be called multiple times per register */ |
6242 | static void unregister_sched_domain_sysctl(void) |
6243 | { |
6244 | if (sd_sysctl_header) |
6245 | unregister_sysctl_table(sd_sysctl_header); |
6246 | sd_sysctl_header = NULL; |
6247 | if (sd_ctl_dir[0].child) |
6248 | sd_free_ctl_entry(&sd_ctl_dir[0].child); |
6249 | } |
6250 | #else |
6251 | static void register_sched_domain_sysctl(void) |
6252 | { |
6253 | } |
6254 | static void unregister_sched_domain_sysctl(void) |
6255 | { |
6256 | } |
6257 | #endif |
6258 | |
6259 | static void set_rq_online(struct rq *rq) |
6260 | { |
6261 | if (!rq->online) { |
6262 | const struct sched_class *class; |
6263 | |
6264 | cpumask_set_cpu(rq->cpu, rq->rd->online); |
6265 | rq->online = 1; |
6266 | |
6267 | for_each_class(class) { |
6268 | if (class->rq_online) |
6269 | class->rq_online(rq); |
6270 | } |
6271 | } |
6272 | } |
6273 | |
6274 | static void set_rq_offline(struct rq *rq) |
6275 | { |
6276 | if (rq->online) { |
6277 | const struct sched_class *class; |
6278 | |
6279 | for_each_class(class) { |
6280 | if (class->rq_offline) |
6281 | class->rq_offline(rq); |
6282 | } |
6283 | |
6284 | cpumask_clear_cpu(rq->cpu, rq->rd->online); |
6285 | rq->online = 0; |
6286 | } |
6287 | } |
6288 | |
6289 | /* |
6290 | * migration_call - callback that gets triggered when a CPU is added. |
6291 | * Here we can start up the necessary migration thread for the new CPU. |
6292 | */ |
6293 | static int __cpuinit |
6294 | migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) |
6295 | { |
6296 | int cpu = (long)hcpu; |
6297 | unsigned long flags; |
6298 | struct rq *rq = cpu_rq(cpu); |
6299 | |
6300 | switch (action & ~CPU_TASKS_FROZEN) { |
6301 | |
6302 | case CPU_UP_PREPARE: |
6303 | rq->calc_load_update = calc_load_update; |
6304 | break; |
6305 | |
6306 | case CPU_ONLINE: |
6307 | /* Update our root-domain */ |
6308 | raw_spin_lock_irqsave(&rq->lock, flags); |
6309 | if (rq->rd) { |
6310 | BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); |
6311 | |
6312 | set_rq_online(rq); |
6313 | } |
6314 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
6315 | break; |
6316 | |
6317 | #ifdef CONFIG_HOTPLUG_CPU |
6318 | case CPU_DYING: |
6319 | /* Update our root-domain */ |
6320 | raw_spin_lock_irqsave(&rq->lock, flags); |
6321 | if (rq->rd) { |
6322 | BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); |
6323 | set_rq_offline(rq); |
6324 | } |
6325 | migrate_tasks(cpu); |
6326 | BUG_ON(rq->nr_running != 1); /* the migration thread */ |
6327 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
6328 | |
6329 | migrate_nr_uninterruptible(rq); |
6330 | calc_global_load_remove(rq); |
6331 | break; |
6332 | #endif |
6333 | } |
6334 | |
6335 | update_max_interval(); |
6336 | |
6337 | return NOTIFY_OK; |
6338 | } |
6339 | |
6340 | /* |
6341 | * Register at high priority so that task migration (migrate_all_tasks) |
6342 | * happens before everything else. This has to be lower priority than |
6343 | * the notifier in the perf_event subsystem, though. |
6344 | */ |
6345 | static struct notifier_block __cpuinitdata migration_notifier = { |
6346 | .notifier_call = migration_call, |
6347 | .priority = CPU_PRI_MIGRATION, |
6348 | }; |
6349 | |
6350 | static int __cpuinit sched_cpu_active(struct notifier_block *nfb, |
6351 | unsigned long action, void *hcpu) |
6352 | { |
6353 | switch (action & ~CPU_TASKS_FROZEN) { |
6354 | case CPU_ONLINE: |
6355 | case CPU_DOWN_FAILED: |
6356 | set_cpu_active((long)hcpu, true); |
6357 | return NOTIFY_OK; |
6358 | default: |
6359 | return NOTIFY_DONE; |
6360 | } |
6361 | } |
6362 | |
6363 | static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb, |
6364 | unsigned long action, void *hcpu) |
6365 | { |
6366 | switch (action & ~CPU_TASKS_FROZEN) { |
6367 | case CPU_DOWN_PREPARE: |
6368 | set_cpu_active((long)hcpu, false); |
6369 | return NOTIFY_OK; |
6370 | default: |
6371 | return NOTIFY_DONE; |
6372 | } |
6373 | } |
6374 | |
6375 | static int __init migration_init(void) |
6376 | { |
6377 | void *cpu = (void *)(long)smp_processor_id(); |
6378 | int err; |
6379 | |
6380 | /* Initialize migration for the boot CPU */ |
6381 | err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); |
6382 | BUG_ON(err == NOTIFY_BAD); |
6383 | migration_call(&migration_notifier, CPU_ONLINE, cpu); |
6384 | register_cpu_notifier(&migration_notifier); |
6385 | |
6386 | /* Register cpu active notifiers */ |
6387 | cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); |
6388 | cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); |
6389 | |
6390 | return 0; |
6391 | } |
6392 | early_initcall(migration_init); |
6393 | #endif |
6394 | |
6395 | #ifdef CONFIG_SMP |
6396 | |
6397 | #ifdef CONFIG_SCHED_DEBUG |
6398 | |
6399 | static __read_mostly int sched_domain_debug_enabled; |
6400 | |
6401 | static int __init sched_domain_debug_setup(char *str) |
6402 | { |
6403 | sched_domain_debug_enabled = 1; |
6404 | |
6405 | return 0; |
6406 | } |
6407 | early_param("sched_debug", sched_domain_debug_setup); |
6408 | |
6409 | static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, |
6410 | struct cpumask *groupmask) |
6411 | { |
6412 | struct sched_group *group = sd->groups; |
6413 | char str[256]; |
6414 | |
6415 | cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); |
6416 | cpumask_clear(groupmask); |
6417 | |
6418 | printk(KERN_DEBUG "%*s domain %d: ", level, "", level); |
6419 | |
6420 | if (!(sd->flags & SD_LOAD_BALANCE)) { |
6421 | printk("does not load-balance\n"); |
6422 | if (sd->parent) |
6423 | printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" |
6424 | " has parent"); |
6425 | return -1; |
6426 | } |
6427 | |
6428 | printk(KERN_CONT "span %s level %s\n", str, sd->name); |
6429 | |
6430 | if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { |
6431 | printk(KERN_ERR "ERROR: domain->span does not contain " |
6432 | "CPU%d\n", cpu); |
6433 | } |
6434 | if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { |
6435 | printk(KERN_ERR "ERROR: domain->groups does not contain" |
6436 | " CPU%d\n", cpu); |
6437 | } |
6438 | |
6439 | printk(KERN_DEBUG "%*s groups:", level + 1, ""); |
6440 | do { |
6441 | if (!group) { |
6442 | printk("\n"); |
6443 | printk(KERN_ERR "ERROR: group is NULL\n"); |
6444 | break; |
6445 | } |
6446 | |
6447 | if (!group->cpu_power) { |
6448 | printk(KERN_CONT "\n"); |
6449 | printk(KERN_ERR "ERROR: domain->cpu_power not " |
6450 | "set\n"); |
6451 | break; |
6452 | } |
6453 | |
6454 | if (!cpumask_weight(sched_group_cpus(group))) { |
6455 | printk(KERN_CONT "\n"); |
6456 | printk(KERN_ERR "ERROR: empty group\n"); |
6457 | break; |
6458 | } |
6459 | |
6460 | if (cpumask_intersects(groupmask, sched_group_cpus(group))) { |
6461 | printk(KERN_CONT "\n"); |
6462 | printk(KERN_ERR "ERROR: repeated CPUs\n"); |
6463 | break; |
6464 | } |
6465 | |
6466 | cpumask_or(groupmask, groupmask, sched_group_cpus(group)); |
6467 | |
6468 | cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); |
6469 | |
6470 | printk(KERN_CONT " %s", str); |
6471 | if (group->cpu_power != SCHED_LOAD_SCALE) { |
6472 | printk(KERN_CONT " (cpu_power = %d)", |
6473 | group->cpu_power); |
6474 | } |
6475 | |
6476 | group = group->next; |
6477 | } while (group != sd->groups); |
6478 | printk(KERN_CONT "\n"); |
6479 | |
6480 | if (!cpumask_equal(sched_domain_span(sd), groupmask)) |
6481 | printk(KERN_ERR "ERROR: groups don't span domain->span\n"); |
6482 | |
6483 | if (sd->parent && |
6484 | !cpumask_subset(groupmask, sched_domain_span(sd->parent))) |
6485 | printk(KERN_ERR "ERROR: parent span is not a superset " |
6486 | "of domain->span\n"); |
6487 | return 0; |
6488 | } |
6489 | |
6490 | static void sched_domain_debug(struct sched_domain *sd, int cpu) |
6491 | { |
6492 | cpumask_var_t groupmask; |
6493 | int level = 0; |
6494 | |
6495 | if (!sched_domain_debug_enabled) |
6496 | return; |
6497 | |
6498 | if (!sd) { |
6499 | printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); |
6500 | return; |
6501 | } |
6502 | |
6503 | printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); |
6504 | |
6505 | if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) { |
6506 | printk(KERN_DEBUG "Cannot load-balance (out of memory)\n"); |
6507 | return; |
6508 | } |
6509 | |
6510 | for (;;) { |
6511 | if (sched_domain_debug_one(sd, cpu, level, groupmask)) |
6512 | break; |
6513 | level++; |
6514 | sd = sd->parent; |
6515 | if (!sd) |
6516 | break; |
6517 | } |
6518 | free_cpumask_var(groupmask); |
6519 | } |
6520 | #else /* !CONFIG_SCHED_DEBUG */ |
6521 | # define sched_domain_debug(sd, cpu) do { } while (0) |
6522 | #endif /* CONFIG_SCHED_DEBUG */ |
6523 | |
6524 | static int sd_degenerate(struct sched_domain *sd) |
6525 | { |
6526 | if (cpumask_weight(sched_domain_span(sd)) == 1) |
6527 | return 1; |
6528 | |
6529 | /* Following flags need at least 2 groups */ |
6530 | if (sd->flags & (SD_LOAD_BALANCE | |
6531 | SD_BALANCE_NEWIDLE | |
6532 | SD_BALANCE_FORK | |
6533 | SD_BALANCE_EXEC | |
6534 | SD_SHARE_CPUPOWER | |
6535 | SD_SHARE_PKG_RESOURCES)) { |
6536 | if (sd->groups != sd->groups->next) |
6537 | return 0; |
6538 | } |
6539 | |
6540 | /* Following flags don't use groups */ |
6541 | if (sd->flags & (SD_WAKE_AFFINE)) |
6542 | return 0; |
6543 | |
6544 | return 1; |
6545 | } |
6546 | |
6547 | static int |
6548 | sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) |
6549 | { |
6550 | unsigned long cflags = sd->flags, pflags = parent->flags; |
6551 | |
6552 | if (sd_degenerate(parent)) |
6553 | return 1; |
6554 | |
6555 | if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) |
6556 | return 0; |
6557 | |
6558 | /* Flags needing groups don't count if only 1 group in parent */ |
6559 | if (parent->groups == parent->groups->next) { |
6560 | pflags &= ~(SD_LOAD_BALANCE | |
6561 | SD_BALANCE_NEWIDLE | |
6562 | SD_BALANCE_FORK | |
6563 | SD_BALANCE_EXEC | |
6564 | SD_SHARE_CPUPOWER | |
6565 | SD_SHARE_PKG_RESOURCES); |
6566 | if (nr_node_ids == 1) |
6567 | pflags &= ~SD_SERIALIZE; |
6568 | } |
6569 | if (~cflags & pflags) |
6570 | return 0; |
6571 | |
6572 | return 1; |
6573 | } |
6574 | |
6575 | static void free_rootdomain(struct root_domain *rd) |
6576 | { |
6577 | synchronize_sched(); |
6578 | |
6579 | cpupri_cleanup(&rd->cpupri); |
6580 | |
6581 | free_cpumask_var(rd->rto_mask); |
6582 | free_cpumask_var(rd->online); |
6583 | free_cpumask_var(rd->span); |
6584 | kfree(rd); |
6585 | } |
6586 | |
6587 | static void rq_attach_root(struct rq *rq, struct root_domain *rd) |
6588 | { |
6589 | struct root_domain *old_rd = NULL; |
6590 | unsigned long flags; |
6591 | |
6592 | raw_spin_lock_irqsave(&rq->lock, flags); |
6593 | |
6594 | if (rq->rd) { |
6595 | old_rd = rq->rd; |
6596 | |
6597 | if (cpumask_test_cpu(rq->cpu, old_rd->online)) |
6598 | set_rq_offline(rq); |
6599 | |
6600 | cpumask_clear_cpu(rq->cpu, old_rd->span); |
6601 | |
6602 | /* |
6603 | * If we dont want to free the old_rt yet then |
6604 | * set old_rd to NULL to skip the freeing later |
6605 | * in this function: |
6606 | */ |
6607 | if (!atomic_dec_and_test(&old_rd->refcount)) |
6608 | old_rd = NULL; |
6609 | } |
6610 | |
6611 | atomic_inc(&rd->refcount); |
6612 | rq->rd = rd; |
6613 | |
6614 | cpumask_set_cpu(rq->cpu, rd->span); |
6615 | if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) |
6616 | set_rq_online(rq); |
6617 | |
6618 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
6619 | |
6620 | if (old_rd) |
6621 | free_rootdomain(old_rd); |
6622 | } |
6623 | |
6624 | static int init_rootdomain(struct root_domain *rd) |
6625 | { |
6626 | memset(rd, 0, sizeof(*rd)); |
6627 | |
6628 | if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) |
6629 | goto out; |
6630 | if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) |
6631 | goto free_span; |
6632 | if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) |
6633 | goto free_online; |
6634 | |
6635 | if (cpupri_init(&rd->cpupri) != 0) |
6636 | goto free_rto_mask; |
6637 | return 0; |
6638 | |
6639 | free_rto_mask: |
6640 | free_cpumask_var(rd->rto_mask); |
6641 | free_online: |
6642 | free_cpumask_var(rd->online); |
6643 | free_span: |
6644 | free_cpumask_var(rd->span); |
6645 | out: |
6646 | return -ENOMEM; |
6647 | } |
6648 | |
6649 | static void init_defrootdomain(void) |
6650 | { |
6651 | init_rootdomain(&def_root_domain); |
6652 | |
6653 | atomic_set(&def_root_domain.refcount, 1); |
6654 | } |
6655 | |
6656 | static struct root_domain *alloc_rootdomain(void) |
6657 | { |
6658 | struct root_domain *rd; |
6659 | |
6660 | rd = kmalloc(sizeof(*rd), GFP_KERNEL); |
6661 | if (!rd) |
6662 | return NULL; |
6663 | |
6664 | if (init_rootdomain(rd) != 0) { |
6665 | kfree(rd); |
6666 | return NULL; |
6667 | } |
6668 | |
6669 | return rd; |
6670 | } |
6671 | |
6672 | /* |
6673 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must |
6674 | * hold the hotplug lock. |
6675 | */ |
6676 | static void |
6677 | cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) |
6678 | { |
6679 | struct rq *rq = cpu_rq(cpu); |
6680 | struct sched_domain *tmp; |
6681 | |
6682 | for (tmp = sd; tmp; tmp = tmp->parent) |
6683 | tmp->span_weight = cpumask_weight(sched_domain_span(tmp)); |
6684 | |
6685 | /* Remove the sched domains which do not contribute to scheduling. */ |
6686 | for (tmp = sd; tmp; ) { |
6687 | struct sched_domain *parent = tmp->parent; |
6688 | if (!parent) |
6689 | break; |
6690 | |
6691 | if (sd_parent_degenerate(tmp, parent)) { |
6692 | tmp->parent = parent->parent; |
6693 | if (parent->parent) |
6694 | parent->parent->child = tmp; |
6695 | } else |
6696 | tmp = tmp->parent; |
6697 | } |
6698 | |
6699 | if (sd && sd_degenerate(sd)) { |
6700 | sd = sd->parent; |
6701 | if (sd) |
6702 | sd->child = NULL; |
6703 | } |
6704 | |
6705 | sched_domain_debug(sd, cpu); |
6706 | |
6707 | rq_attach_root(rq, rd); |
6708 | rcu_assign_pointer(rq->sd, sd); |
6709 | } |
6710 | |
6711 | /* cpus with isolated domains */ |
6712 | static cpumask_var_t cpu_isolated_map; |
6713 | |
6714 | /* Setup the mask of cpus configured for isolated domains */ |
6715 | static int __init isolated_cpu_setup(char *str) |
6716 | { |
6717 | alloc_bootmem_cpumask_var(&cpu_isolated_map); |
6718 | cpulist_parse(str, cpu_isolated_map); |
6719 | return 1; |
6720 | } |
6721 | |
6722 | __setup("isolcpus=", isolated_cpu_setup); |
6723 | |
6724 | /* |
6725 | * init_sched_build_groups takes the cpumask we wish to span, and a pointer |
6726 | * to a function which identifies what group(along with sched group) a CPU |
6727 | * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids |
6728 | * (due to the fact that we keep track of groups covered with a struct cpumask). |
6729 | * |
6730 | * init_sched_build_groups will build a circular linked list of the groups |
6731 | * covered by the given span, and will set each group's ->cpumask correctly, |
6732 | * and ->cpu_power to 0. |
6733 | */ |
6734 | static void |
6735 | init_sched_build_groups(const struct cpumask *span, |
6736 | const struct cpumask *cpu_map, |
6737 | int (*group_fn)(int cpu, const struct cpumask *cpu_map, |
6738 | struct sched_group **sg, |
6739 | struct cpumask *tmpmask), |
6740 | struct cpumask *covered, struct cpumask *tmpmask) |
6741 | { |
6742 | struct sched_group *first = NULL, *last = NULL; |
6743 | int i; |
6744 | |
6745 | cpumask_clear(covered); |
6746 | |
6747 | for_each_cpu(i, span) { |
6748 | struct sched_group *sg; |
6749 | int group = group_fn(i, cpu_map, &sg, tmpmask); |
6750 | int j; |
6751 | |
6752 | if (cpumask_test_cpu(i, covered)) |
6753 | continue; |
6754 | |
6755 | cpumask_clear(sched_group_cpus(sg)); |
6756 | sg->cpu_power = 0; |
6757 | |
6758 | for_each_cpu(j, span) { |
6759 | if (group_fn(j, cpu_map, NULL, tmpmask) != group) |
6760 | continue; |
6761 | |
6762 | cpumask_set_cpu(j, covered); |
6763 | cpumask_set_cpu(j, sched_group_cpus(sg)); |
6764 | } |
6765 | if (!first) |
6766 | first = sg; |
6767 | if (last) |
6768 | last->next = sg; |
6769 | last = sg; |
6770 | } |
6771 | last->next = first; |
6772 | } |
6773 | |
6774 | #define SD_NODES_PER_DOMAIN 16 |
6775 | |
6776 | #ifdef CONFIG_NUMA |
6777 | |
6778 | /** |
6779 | * find_next_best_node - find the next node to include in a sched_domain |
6780 | * @node: node whose sched_domain we're building |
6781 | * @used_nodes: nodes already in the sched_domain |
6782 | * |
6783 | * Find the next node to include in a given scheduling domain. Simply |
6784 | * finds the closest node not already in the @used_nodes map. |
6785 | * |
6786 | * Should use nodemask_t. |
6787 | */ |
6788 | static int find_next_best_node(int node, nodemask_t *used_nodes) |
6789 | { |
6790 | int i, n, val, min_val, best_node = 0; |
6791 | |
6792 | min_val = INT_MAX; |
6793 | |
6794 | for (i = 0; i < nr_node_ids; i++) { |
6795 | /* Start at @node */ |
6796 | n = (node + i) % nr_node_ids; |
6797 | |
6798 | if (!nr_cpus_node(n)) |
6799 | continue; |
6800 | |
6801 | /* Skip already used nodes */ |
6802 | if (node_isset(n, *used_nodes)) |
6803 | continue; |
6804 | |
6805 | /* Simple min distance search */ |
6806 | val = node_distance(node, n); |
6807 | |
6808 | if (val < min_val) { |
6809 | min_val = val; |
6810 | best_node = n; |
6811 | } |
6812 | } |
6813 | |
6814 | node_set(best_node, *used_nodes); |
6815 | return best_node; |
6816 | } |
6817 | |
6818 | /** |
6819 | * sched_domain_node_span - get a cpumask for a node's sched_domain |
6820 | * @node: node whose cpumask we're constructing |
6821 | * @span: resulting cpumask |
6822 | * |
6823 | * Given a node, construct a good cpumask for its sched_domain to span. It |
6824 | * should be one that prevents unnecessary balancing, but also spreads tasks |
6825 | * out optimally. |
6826 | */ |
6827 | static void sched_domain_node_span(int node, struct cpumask *span) |
6828 | { |
6829 | nodemask_t used_nodes; |
6830 | int i; |
6831 | |
6832 | cpumask_clear(span); |
6833 | nodes_clear(used_nodes); |
6834 | |
6835 | cpumask_or(span, span, cpumask_of_node(node)); |
6836 | node_set(node, used_nodes); |
6837 | |
6838 | for (i = 1; i < SD_NODES_PER_DOMAIN; i++) { |
6839 | int next_node = find_next_best_node(node, &used_nodes); |
6840 | |
6841 | cpumask_or(span, span, cpumask_of_node(next_node)); |
6842 | } |
6843 | } |
6844 | #endif /* CONFIG_NUMA */ |
6845 | |
6846 | int sched_smt_power_savings = 0, sched_mc_power_savings = 0; |
6847 | |
6848 | /* |
6849 | * The cpus mask in sched_group and sched_domain hangs off the end. |
6850 | * |
6851 | * ( See the the comments in include/linux/sched.h:struct sched_group |
6852 | * and struct sched_domain. ) |
6853 | */ |
6854 | struct static_sched_group { |
6855 | struct sched_group sg; |
6856 | DECLARE_BITMAP(cpus, CONFIG_NR_CPUS); |
6857 | }; |
6858 | |
6859 | struct static_sched_domain { |
6860 | struct sched_domain sd; |
6861 | DECLARE_BITMAP(span, CONFIG_NR_CPUS); |
6862 | }; |
6863 | |
6864 | struct s_data { |
6865 | #ifdef CONFIG_NUMA |
6866 | int sd_allnodes; |
6867 | cpumask_var_t domainspan; |
6868 | cpumask_var_t covered; |
6869 | cpumask_var_t notcovered; |
6870 | #endif |
6871 | cpumask_var_t nodemask; |
6872 | cpumask_var_t this_sibling_map; |
6873 | cpumask_var_t this_core_map; |
6874 | cpumask_var_t this_book_map; |
6875 | cpumask_var_t send_covered; |
6876 | cpumask_var_t tmpmask; |
6877 | struct sched_group **sched_group_nodes; |
6878 | struct root_domain *rd; |
6879 | }; |
6880 | |
6881 | enum s_alloc { |
6882 | sa_sched_groups = 0, |
6883 | sa_rootdomain, |
6884 | sa_tmpmask, |
6885 | sa_send_covered, |
6886 | sa_this_book_map, |
6887 | sa_this_core_map, |
6888 | sa_this_sibling_map, |
6889 | sa_nodemask, |
6890 | sa_sched_group_nodes, |
6891 | #ifdef CONFIG_NUMA |
6892 | sa_notcovered, |
6893 | sa_covered, |
6894 | sa_domainspan, |
6895 | #endif |
6896 | sa_none, |
6897 | }; |
6898 | |
6899 | /* |
6900 | * SMT sched-domains: |
6901 | */ |
6902 | #ifdef CONFIG_SCHED_SMT |
6903 | static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains); |
6904 | static DEFINE_PER_CPU(struct static_sched_group, sched_groups); |
6905 | |
6906 | static int |
6907 | cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map, |
6908 | struct sched_group **sg, struct cpumask *unused) |
6909 | { |
6910 | if (sg) |
6911 | *sg = &per_cpu(sched_groups, cpu).sg; |
6912 | return cpu; |
6913 | } |
6914 | #endif /* CONFIG_SCHED_SMT */ |
6915 | |
6916 | /* |
6917 | * multi-core sched-domains: |
6918 | */ |
6919 | #ifdef CONFIG_SCHED_MC |
6920 | static DEFINE_PER_CPU(struct static_sched_domain, core_domains); |
6921 | static DEFINE_PER_CPU(struct static_sched_group, sched_group_core); |
6922 | |
6923 | static int |
6924 | cpu_to_core_group(int cpu, const struct cpumask *cpu_map, |
6925 | struct sched_group **sg, struct cpumask *mask) |
6926 | { |
6927 | int group; |
6928 | #ifdef CONFIG_SCHED_SMT |
6929 | cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map); |
6930 | group = cpumask_first(mask); |
6931 | #else |
6932 | group = cpu; |
6933 | #endif |
6934 | if (sg) |
6935 | *sg = &per_cpu(sched_group_core, group).sg; |
6936 | return group; |
6937 | } |
6938 | #endif /* CONFIG_SCHED_MC */ |
6939 | |
6940 | /* |
6941 | * book sched-domains: |
6942 | */ |
6943 | #ifdef CONFIG_SCHED_BOOK |
6944 | static DEFINE_PER_CPU(struct static_sched_domain, book_domains); |
6945 | static DEFINE_PER_CPU(struct static_sched_group, sched_group_book); |
6946 | |
6947 | static int |
6948 | cpu_to_book_group(int cpu, const struct cpumask *cpu_map, |
6949 | struct sched_group **sg, struct cpumask *mask) |
6950 | { |
6951 | int group = cpu; |
6952 | #ifdef CONFIG_SCHED_MC |
6953 | cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map); |
6954 | group = cpumask_first(mask); |
6955 | #elif defined(CONFIG_SCHED_SMT) |
6956 | cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map); |
6957 | group = cpumask_first(mask); |
6958 | #endif |
6959 | if (sg) |
6960 | *sg = &per_cpu(sched_group_book, group).sg; |
6961 | return group; |
6962 | } |
6963 | #endif /* CONFIG_SCHED_BOOK */ |
6964 | |
6965 | static DEFINE_PER_CPU(struct static_sched_domain, phys_domains); |
6966 | static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys); |
6967 | |
6968 | static int |
6969 | cpu_to_phys_group(int cpu, const struct cpumask *cpu_map, |
6970 | struct sched_group **sg, struct cpumask *mask) |
6971 | { |
6972 | int group; |
6973 | #ifdef CONFIG_SCHED_BOOK |
6974 | cpumask_and(mask, cpu_book_mask(cpu), cpu_map); |
6975 | group = cpumask_first(mask); |
6976 | #elif defined(CONFIG_SCHED_MC) |
6977 | cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map); |
6978 | group = cpumask_first(mask); |
6979 | #elif defined(CONFIG_SCHED_SMT) |
6980 | cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map); |
6981 | group = cpumask_first(mask); |
6982 | #else |
6983 | group = cpu; |
6984 | #endif |
6985 | if (sg) |
6986 | *sg = &per_cpu(sched_group_phys, group).sg; |
6987 | return group; |
6988 | } |
6989 | |
6990 | #ifdef CONFIG_NUMA |
6991 | /* |
6992 | * The init_sched_build_groups can't handle what we want to do with node |
6993 | * groups, so roll our own. Now each node has its own list of groups which |
6994 | * gets dynamically allocated. |
6995 | */ |
6996 | static DEFINE_PER_CPU(struct static_sched_domain, node_domains); |
6997 | static struct sched_group ***sched_group_nodes_bycpu; |
6998 | |
6999 | static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains); |
7000 | static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes); |
7001 | |
7002 | static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map, |
7003 | struct sched_group **sg, |
7004 | struct cpumask *nodemask) |
7005 | { |
7006 | int group; |
7007 | |
7008 | cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map); |
7009 | group = cpumask_first(nodemask); |
7010 | |
7011 | if (sg) |
7012 | *sg = &per_cpu(sched_group_allnodes, group).sg; |
7013 | return group; |
7014 | } |
7015 | |
7016 | static void init_numa_sched_groups_power(struct sched_group *group_head) |
7017 | { |
7018 | struct sched_group *sg = group_head; |
7019 | int j; |
7020 | |
7021 | if (!sg) |
7022 | return; |
7023 | do { |
7024 | for_each_cpu(j, sched_group_cpus(sg)) { |
7025 | struct sched_domain *sd; |
7026 | |
7027 | sd = &per_cpu(phys_domains, j).sd; |
7028 | if (j != group_first_cpu(sd->groups)) { |
7029 | /* |
7030 | * Only add "power" once for each |
7031 | * physical package. |
7032 | */ |
7033 | continue; |
7034 | } |
7035 | |
7036 | sg->cpu_power += sd->groups->cpu_power; |
7037 | } |
7038 | sg = sg->next; |
7039 | } while (sg != group_head); |
7040 | } |
7041 | |
7042 | static int build_numa_sched_groups(struct s_data *d, |
7043 | const struct cpumask *cpu_map, int num) |
7044 | { |
7045 | struct sched_domain *sd; |
7046 | struct sched_group *sg, *prev; |
7047 | int n, j; |
7048 | |
7049 | cpumask_clear(d->covered); |
7050 | cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map); |
7051 | if (cpumask_empty(d->nodemask)) { |
7052 | d->sched_group_nodes[num] = NULL; |
7053 | goto out; |
7054 | } |
7055 | |
7056 | sched_domain_node_span(num, d->domainspan); |
7057 | cpumask_and(d->domainspan, d->domainspan, cpu_map); |
7058 | |
7059 | sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(), |
7060 | GFP_KERNEL, num); |
7061 | if (!sg) { |
7062 | printk(KERN_WARNING "Can not alloc domain group for node %d\n", |
7063 | num); |
7064 | return -ENOMEM; |
7065 | } |
7066 | d->sched_group_nodes[num] = sg; |
7067 | |
7068 | for_each_cpu(j, d->nodemask) { |
7069 | sd = &per_cpu(node_domains, j).sd; |
7070 | sd->groups = sg; |
7071 | } |
7072 | |
7073 | sg->cpu_power = 0; |
7074 | cpumask_copy(sched_group_cpus(sg), d->nodemask); |
7075 | sg->next = sg; |
7076 | cpumask_or(d->covered, d->covered, d->nodemask); |
7077 | |
7078 | prev = sg; |
7079 | for (j = 0; j < nr_node_ids; j++) { |
7080 | n = (num + j) % nr_node_ids; |
7081 | cpumask_complement(d->notcovered, d->covered); |
7082 | cpumask_and(d->tmpmask, d->notcovered, cpu_map); |
7083 | cpumask_and(d->tmpmask, d->tmpmask, d->domainspan); |
7084 | if (cpumask_empty(d->tmpmask)) |
7085 | break; |
7086 | cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n)); |
7087 | if (cpumask_empty(d->tmpmask)) |
7088 | continue; |
7089 | sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(), |
7090 | GFP_KERNEL, num); |
7091 | if (!sg) { |
7092 | printk(KERN_WARNING |
7093 | "Can not alloc domain group for node %d\n", j); |
7094 | return -ENOMEM; |
7095 | } |
7096 | sg->cpu_power = 0; |
7097 | cpumask_copy(sched_group_cpus(sg), d->tmpmask); |
7098 | sg->next = prev->next; |
7099 | cpumask_or(d->covered, d->covered, d->tmpmask); |
7100 | prev->next = sg; |
7101 | prev = sg; |
7102 | } |
7103 | out: |
7104 | return 0; |
7105 | } |
7106 | #endif /* CONFIG_NUMA */ |
7107 | |
7108 | #ifdef CONFIG_NUMA |
7109 | /* Free memory allocated for various sched_group structures */ |
7110 | static void free_sched_groups(const struct cpumask *cpu_map, |
7111 | struct cpumask *nodemask) |
7112 | { |
7113 | int cpu, i; |
7114 | |
7115 | for_each_cpu(cpu, cpu_map) { |
7116 | struct sched_group **sched_group_nodes |
7117 | = sched_group_nodes_bycpu[cpu]; |
7118 | |
7119 | if (!sched_group_nodes) |
7120 | continue; |
7121 | |
7122 | for (i = 0; i < nr_node_ids; i++) { |
7123 | struct sched_group *oldsg, *sg = sched_group_nodes[i]; |
7124 | |
7125 | cpumask_and(nodemask, cpumask_of_node(i), cpu_map); |
7126 | if (cpumask_empty(nodemask)) |
7127 | continue; |
7128 | |
7129 | if (sg == NULL) |
7130 | continue; |
7131 | sg = sg->next; |
7132 | next_sg: |
7133 | oldsg = sg; |
7134 | sg = sg->next; |
7135 | kfree(oldsg); |
7136 | if (oldsg != sched_group_nodes[i]) |
7137 | goto next_sg; |
7138 | } |
7139 | kfree(sched_group_nodes); |
7140 | sched_group_nodes_bycpu[cpu] = NULL; |
7141 | } |
7142 | } |
7143 | #else /* !CONFIG_NUMA */ |
7144 | static void free_sched_groups(const struct cpumask *cpu_map, |
7145 | struct cpumask *nodemask) |
7146 | { |
7147 | } |
7148 | #endif /* CONFIG_NUMA */ |
7149 | |
7150 | /* |
7151 | * Initialize sched groups cpu_power. |
7152 | * |
7153 | * cpu_power indicates the capacity of sched group, which is used while |
7154 | * distributing the load between different sched groups in a sched domain. |
7155 | * Typically cpu_power for all the groups in a sched domain will be same unless |
7156 | * there are asymmetries in the topology. If there are asymmetries, group |
7157 | * having more cpu_power will pickup more load compared to the group having |
7158 | * less cpu_power. |
7159 | */ |
7160 | static void init_sched_groups_power(int cpu, struct sched_domain *sd) |
7161 | { |
7162 | struct sched_domain *child; |
7163 | struct sched_group *group; |
7164 | long power; |
7165 | int weight; |
7166 | |
7167 | WARN_ON(!sd || !sd->groups); |
7168 | |
7169 | if (cpu != group_first_cpu(sd->groups)) |
7170 | return; |
7171 | |
7172 | sd->groups->group_weight = cpumask_weight(sched_group_cpus(sd->groups)); |
7173 | |
7174 | child = sd->child; |
7175 | |
7176 | sd->groups->cpu_power = 0; |
7177 | |
7178 | if (!child) { |
7179 | power = SCHED_LOAD_SCALE; |
7180 | weight = cpumask_weight(sched_domain_span(sd)); |
7181 | /* |
7182 | * SMT siblings share the power of a single core. |
7183 | * Usually multiple threads get a better yield out of |
7184 | * that one core than a single thread would have, |
7185 | * reflect that in sd->smt_gain. |
7186 | */ |
7187 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { |
7188 | power *= sd->smt_gain; |
7189 | power /= weight; |
7190 | power >>= SCHED_LOAD_SHIFT; |
7191 | } |
7192 | sd->groups->cpu_power += power; |
7193 | return; |
7194 | } |
7195 | |
7196 | /* |
7197 | * Add cpu_power of each child group to this groups cpu_power. |
7198 | */ |
7199 | group = child->groups; |
7200 | do { |
7201 | sd->groups->cpu_power += group->cpu_power; |
7202 | group = group->next; |
7203 | } while (group != child->groups); |
7204 | } |
7205 | |
7206 | /* |
7207 | * Initializers for schedule domains |
7208 | * Non-inlined to reduce accumulated stack pressure in build_sched_domains() |
7209 | */ |
7210 | |
7211 | #ifdef CONFIG_SCHED_DEBUG |
7212 | # define SD_INIT_NAME(sd, type) sd->name = #type |
7213 | #else |
7214 | # define SD_INIT_NAME(sd, type) do { } while (0) |
7215 | #endif |
7216 | |
7217 | #define SD_INIT(sd, type) sd_init_##type(sd) |
7218 | |
7219 | #define SD_INIT_FUNC(type) \ |
7220 | static noinline void sd_init_##type(struct sched_domain *sd) \ |
7221 | { \ |
7222 | memset(sd, 0, sizeof(*sd)); \ |
7223 | *sd = SD_##type##_INIT; \ |
7224 | sd->level = SD_LV_##type; \ |
7225 | SD_INIT_NAME(sd, type); \ |
7226 | } |
7227 | |
7228 | SD_INIT_FUNC(CPU) |
7229 | #ifdef CONFIG_NUMA |
7230 | SD_INIT_FUNC(ALLNODES) |
7231 | SD_INIT_FUNC(NODE) |
7232 | #endif |
7233 | #ifdef CONFIG_SCHED_SMT |
7234 | SD_INIT_FUNC(SIBLING) |
7235 | #endif |
7236 | #ifdef CONFIG_SCHED_MC |
7237 | SD_INIT_FUNC(MC) |
7238 | #endif |
7239 | #ifdef CONFIG_SCHED_BOOK |
7240 | SD_INIT_FUNC(BOOK) |
7241 | #endif |
7242 | |
7243 | static int default_relax_domain_level = -1; |
7244 | |
7245 | static int __init setup_relax_domain_level(char *str) |
7246 | { |
7247 | unsigned long val; |
7248 | |
7249 | val = simple_strtoul(str, NULL, 0); |
7250 | if (val < SD_LV_MAX) |
7251 | default_relax_domain_level = val; |
7252 | |
7253 | return 1; |
7254 | } |
7255 | __setup("relax_domain_level=", setup_relax_domain_level); |
7256 | |
7257 | static void set_domain_attribute(struct sched_domain *sd, |
7258 | struct sched_domain_attr *attr) |
7259 | { |
7260 | int request; |
7261 | |
7262 | if (!attr || attr->relax_domain_level < 0) { |
7263 | if (default_relax_domain_level < 0) |
7264 | return; |
7265 | else |
7266 | request = default_relax_domain_level; |
7267 | } else |
7268 | request = attr->relax_domain_level; |
7269 | if (request < sd->level) { |
7270 | /* turn off idle balance on this domain */ |
7271 | sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); |
7272 | } else { |
7273 | /* turn on idle balance on this domain */ |
7274 | sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); |
7275 | } |
7276 | } |
7277 | |
7278 | static void __free_domain_allocs(struct s_data *d, enum s_alloc what, |
7279 | const struct cpumask *cpu_map) |
7280 | { |
7281 | switch (what) { |
7282 | case sa_sched_groups: |
7283 | free_sched_groups(cpu_map, d->tmpmask); /* fall through */ |
7284 | d->sched_group_nodes = NULL; |
7285 | case sa_rootdomain: |
7286 | free_rootdomain(d->rd); /* fall through */ |
7287 | case sa_tmpmask: |
7288 | free_cpumask_var(d->tmpmask); /* fall through */ |
7289 | case sa_send_covered: |
7290 | free_cpumask_var(d->send_covered); /* fall through */ |
7291 | case sa_this_book_map: |
7292 | free_cpumask_var(d->this_book_map); /* fall through */ |
7293 | case sa_this_core_map: |
7294 | free_cpumask_var(d->this_core_map); /* fall through */ |
7295 | case sa_this_sibling_map: |
7296 | free_cpumask_var(d->this_sibling_map); /* fall through */ |
7297 | case sa_nodemask: |
7298 | free_cpumask_var(d->nodemask); /* fall through */ |
7299 | case sa_sched_group_nodes: |
7300 | #ifdef CONFIG_NUMA |
7301 | kfree(d->sched_group_nodes); /* fall through */ |
7302 | case sa_notcovered: |
7303 | free_cpumask_var(d->notcovered); /* fall through */ |
7304 | case sa_covered: |
7305 | free_cpumask_var(d->covered); /* fall through */ |
7306 | case sa_domainspan: |
7307 | free_cpumask_var(d->domainspan); /* fall through */ |
7308 | #endif |
7309 | case sa_none: |
7310 | break; |
7311 | } |
7312 | } |
7313 | |
7314 | static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, |
7315 | const struct cpumask *cpu_map) |
7316 | { |
7317 | #ifdef CONFIG_NUMA |
7318 | if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL)) |
7319 | return sa_none; |
7320 | if (!alloc_cpumask_var(&d->covered, GFP_KERNEL)) |
7321 | return sa_domainspan; |
7322 | if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL)) |
7323 | return sa_covered; |
7324 | /* Allocate the per-node list of sched groups */ |
7325 | d->sched_group_nodes = kcalloc(nr_node_ids, |
7326 | sizeof(struct sched_group *), GFP_KERNEL); |
7327 | if (!d->sched_group_nodes) { |
7328 | printk(KERN_WARNING "Can not alloc sched group node list\n"); |
7329 | return sa_notcovered; |
7330 | } |
7331 | sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes; |
7332 | #endif |
7333 | if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL)) |
7334 | return sa_sched_group_nodes; |
7335 | if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL)) |
7336 | return sa_nodemask; |
7337 | if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL)) |
7338 | return sa_this_sibling_map; |
7339 | if (!alloc_cpumask_var(&d->this_book_map, GFP_KERNEL)) |
7340 | return sa_this_core_map; |
7341 | if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL)) |
7342 | return sa_this_book_map; |
7343 | if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL)) |
7344 | return sa_send_covered; |
7345 | d->rd = alloc_rootdomain(); |
7346 | if (!d->rd) { |
7347 | printk(KERN_WARNING "Cannot alloc root domain\n"); |
7348 | return sa_tmpmask; |
7349 | } |
7350 | return sa_rootdomain; |
7351 | } |
7352 | |
7353 | static struct sched_domain *__build_numa_sched_domains(struct s_data *d, |
7354 | const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i) |
7355 | { |
7356 | struct sched_domain *sd = NULL; |
7357 | #ifdef CONFIG_NUMA |
7358 | struct sched_domain *parent; |
7359 | |
7360 | d->sd_allnodes = 0; |
7361 | if (cpumask_weight(cpu_map) > |
7362 | SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) { |
7363 | sd = &per_cpu(allnodes_domains, i).sd; |
7364 | SD_INIT(sd, ALLNODES); |
7365 | set_domain_attribute(sd, attr); |
7366 | cpumask_copy(sched_domain_span(sd), cpu_map); |
7367 | cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask); |
7368 | d->sd_allnodes = 1; |
7369 | } |
7370 | parent = sd; |
7371 | |
7372 | sd = &per_cpu(node_domains, i).sd; |
7373 | SD_INIT(sd, NODE); |
7374 | set_domain_attribute(sd, attr); |
7375 | sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd)); |
7376 | sd->parent = parent; |
7377 | if (parent) |
7378 | parent->child = sd; |
7379 | cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map); |
7380 | #endif |
7381 | return sd; |
7382 | } |
7383 | |
7384 | static struct sched_domain *__build_cpu_sched_domain(struct s_data *d, |
7385 | const struct cpumask *cpu_map, struct sched_domain_attr *attr, |
7386 | struct sched_domain *parent, int i) |
7387 | { |
7388 | struct sched_domain *sd; |
7389 | sd = &per_cpu(phys_domains, i).sd; |
7390 | SD_INIT(sd, CPU); |
7391 | set_domain_attribute(sd, attr); |
7392 | cpumask_copy(sched_domain_span(sd), d->nodemask); |
7393 | sd->parent = parent; |
7394 | if (parent) |
7395 | parent->child = sd; |
7396 | cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask); |
7397 | return sd; |
7398 | } |
7399 | |
7400 | static struct sched_domain *__build_book_sched_domain(struct s_data *d, |
7401 | const struct cpumask *cpu_map, struct sched_domain_attr *attr, |
7402 | struct sched_domain *parent, int i) |
7403 | { |
7404 | struct sched_domain *sd = parent; |
7405 | #ifdef CONFIG_SCHED_BOOK |
7406 | sd = &per_cpu(book_domains, i).sd; |
7407 | SD_INIT(sd, BOOK); |
7408 | set_domain_attribute(sd, attr); |
7409 | cpumask_and(sched_domain_span(sd), cpu_map, cpu_book_mask(i)); |
7410 | sd->parent = parent; |
7411 | parent->child = sd; |
7412 | cpu_to_book_group(i, cpu_map, &sd->groups, d->tmpmask); |
7413 | #endif |
7414 | return sd; |
7415 | } |
7416 | |
7417 | static struct sched_domain *__build_mc_sched_domain(struct s_data *d, |
7418 | const struct cpumask *cpu_map, struct sched_domain_attr *attr, |
7419 | struct sched_domain *parent, int i) |
7420 | { |
7421 | struct sched_domain *sd = parent; |
7422 | #ifdef CONFIG_SCHED_MC |
7423 | sd = &per_cpu(core_domains, i).sd; |
7424 | SD_INIT(sd, MC); |
7425 | set_domain_attribute(sd, attr); |
7426 | cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i)); |
7427 | sd->parent = parent; |
7428 | parent->child = sd; |
7429 | cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask); |
7430 | #endif |
7431 | return sd; |
7432 | } |
7433 | |
7434 | static struct sched_domain *__build_smt_sched_domain(struct s_data *d, |
7435 | const struct cpumask *cpu_map, struct sched_domain_attr *attr, |
7436 | struct sched_domain *parent, int i) |
7437 | { |
7438 | struct sched_domain *sd = parent; |
7439 | #ifdef CONFIG_SCHED_SMT |
7440 | sd = &per_cpu(cpu_domains, i).sd; |
7441 | SD_INIT(sd, SIBLING); |
7442 | set_domain_attribute(sd, attr); |
7443 | cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i)); |
7444 | sd->parent = parent; |
7445 | parent->child = sd; |
7446 | cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask); |
7447 | #endif |
7448 | return sd; |
7449 | } |
7450 | |
7451 | static void build_sched_groups(struct s_data *d, enum sched_domain_level l, |
7452 | const struct cpumask *cpu_map, int cpu) |
7453 | { |
7454 | switch (l) { |
7455 | #ifdef CONFIG_SCHED_SMT |
7456 | case SD_LV_SIBLING: /* set up CPU (sibling) groups */ |
7457 | cpumask_and(d->this_sibling_map, cpu_map, |
7458 | topology_thread_cpumask(cpu)); |
7459 | if (cpu == cpumask_first(d->this_sibling_map)) |
7460 | init_sched_build_groups(d->this_sibling_map, cpu_map, |
7461 | &cpu_to_cpu_group, |
7462 | d->send_covered, d->tmpmask); |
7463 | break; |
7464 | #endif |
7465 | #ifdef CONFIG_SCHED_MC |
7466 | case SD_LV_MC: /* set up multi-core groups */ |
7467 | cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu)); |
7468 | if (cpu == cpumask_first(d->this_core_map)) |
7469 | init_sched_build_groups(d->this_core_map, cpu_map, |
7470 | &cpu_to_core_group, |
7471 | d->send_covered, d->tmpmask); |
7472 | break; |
7473 | #endif |
7474 | #ifdef CONFIG_SCHED_BOOK |
7475 | case SD_LV_BOOK: /* set up book groups */ |
7476 | cpumask_and(d->this_book_map, cpu_map, cpu_book_mask(cpu)); |
7477 | if (cpu == cpumask_first(d->this_book_map)) |
7478 | init_sched_build_groups(d->this_book_map, cpu_map, |
7479 | &cpu_to_book_group, |
7480 | d->send_covered, d->tmpmask); |
7481 | break; |
7482 | #endif |
7483 | case SD_LV_CPU: /* set up physical groups */ |
7484 | cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map); |
7485 | if (!cpumask_empty(d->nodemask)) |
7486 | init_sched_build_groups(d->nodemask, cpu_map, |
7487 | &cpu_to_phys_group, |
7488 | d->send_covered, d->tmpmask); |
7489 | break; |
7490 | #ifdef CONFIG_NUMA |
7491 | case SD_LV_ALLNODES: |
7492 | init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group, |
7493 | d->send_covered, d->tmpmask); |
7494 | break; |
7495 | #endif |
7496 | default: |
7497 | break; |
7498 | } |
7499 | } |
7500 | |
7501 | /* |
7502 | * Build sched domains for a given set of cpus and attach the sched domains |
7503 | * to the individual cpus |
7504 | */ |
7505 | static int __build_sched_domains(const struct cpumask *cpu_map, |
7506 | struct sched_domain_attr *attr) |
7507 | { |
7508 | enum s_alloc alloc_state = sa_none; |
7509 | struct s_data d; |
7510 | struct sched_domain *sd; |
7511 | int i; |
7512 | #ifdef CONFIG_NUMA |
7513 | d.sd_allnodes = 0; |
7514 | #endif |
7515 | |
7516 | alloc_state = __visit_domain_allocation_hell(&d, cpu_map); |
7517 | if (alloc_state != sa_rootdomain) |
7518 | goto error; |
7519 | alloc_state = sa_sched_groups; |
7520 | |
7521 | /* |
7522 | * Set up domains for cpus specified by the cpu_map. |
7523 | */ |
7524 | for_each_cpu(i, cpu_map) { |
7525 | cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)), |
7526 | cpu_map); |
7527 | |
7528 | sd = __build_numa_sched_domains(&d, cpu_map, attr, i); |
7529 | sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i); |
7530 | sd = __build_book_sched_domain(&d, cpu_map, attr, sd, i); |
7531 | sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i); |
7532 | sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i); |
7533 | } |
7534 | |
7535 | for_each_cpu(i, cpu_map) { |
7536 | build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i); |
7537 | build_sched_groups(&d, SD_LV_BOOK, cpu_map, i); |
7538 | build_sched_groups(&d, SD_LV_MC, cpu_map, i); |
7539 | } |
7540 | |
7541 | /* Set up physical groups */ |
7542 | for (i = 0; i < nr_node_ids; i++) |
7543 | build_sched_groups(&d, SD_LV_CPU, cpu_map, i); |
7544 | |
7545 | #ifdef CONFIG_NUMA |
7546 | /* Set up node groups */ |
7547 | if (d.sd_allnodes) |
7548 | build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0); |
7549 | |
7550 | for (i = 0; i < nr_node_ids; i++) |
7551 | if (build_numa_sched_groups(&d, cpu_map, i)) |
7552 | goto error; |
7553 | #endif |
7554 | |
7555 | /* Calculate CPU power for physical packages and nodes */ |
7556 | #ifdef CONFIG_SCHED_SMT |
7557 | for_each_cpu(i, cpu_map) { |
7558 | sd = &per_cpu(cpu_domains, i).sd; |
7559 | init_sched_groups_power(i, sd); |
7560 | } |
7561 | #endif |
7562 | #ifdef CONFIG_SCHED_MC |
7563 | for_each_cpu(i, cpu_map) { |
7564 | sd = &per_cpu(core_domains, i).sd; |
7565 | init_sched_groups_power(i, sd); |
7566 | } |
7567 | #endif |
7568 | #ifdef CONFIG_SCHED_BOOK |
7569 | for_each_cpu(i, cpu_map) { |
7570 | sd = &per_cpu(book_domains, i).sd; |
7571 | init_sched_groups_power(i, sd); |
7572 | } |
7573 | #endif |
7574 | |
7575 | for_each_cpu(i, cpu_map) { |
7576 | sd = &per_cpu(phys_domains, i).sd; |
7577 | init_sched_groups_power(i, sd); |
7578 | } |
7579 | |
7580 | #ifdef CONFIG_NUMA |
7581 | for (i = 0; i < nr_node_ids; i++) |
7582 | init_numa_sched_groups_power(d.sched_group_nodes[i]); |
7583 | |
7584 | if (d.sd_allnodes) { |
7585 | struct sched_group *sg; |
7586 | |
7587 | cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg, |
7588 | d.tmpmask); |
7589 | init_numa_sched_groups_power(sg); |
7590 | } |
7591 | #endif |
7592 | |
7593 | /* Attach the domains */ |
7594 | for_each_cpu(i, cpu_map) { |
7595 | #ifdef CONFIG_SCHED_SMT |
7596 | sd = &per_cpu(cpu_domains, i).sd; |
7597 | #elif defined(CONFIG_SCHED_MC) |
7598 | sd = &per_cpu(core_domains, i).sd; |
7599 | #elif defined(CONFIG_SCHED_BOOK) |
7600 | sd = &per_cpu(book_domains, i).sd; |
7601 | #else |
7602 | sd = &per_cpu(phys_domains, i).sd; |
7603 | #endif |
7604 | cpu_attach_domain(sd, d.rd, i); |
7605 | } |
7606 | |
7607 | d.sched_group_nodes = NULL; /* don't free this we still need it */ |
7608 | __free_domain_allocs(&d, sa_tmpmask, cpu_map); |
7609 | return 0; |
7610 | |
7611 | error: |
7612 | __free_domain_allocs(&d, alloc_state, cpu_map); |
7613 | return -ENOMEM; |
7614 | } |
7615 | |
7616 | static int build_sched_domains(const struct cpumask *cpu_map) |
7617 | { |
7618 | return __build_sched_domains(cpu_map, NULL); |
7619 | } |
7620 | |
7621 | static cpumask_var_t *doms_cur; /* current sched domains */ |
7622 | static int ndoms_cur; /* number of sched domains in 'doms_cur' */ |
7623 | static struct sched_domain_attr *dattr_cur; |
7624 | /* attribues of custom domains in 'doms_cur' */ |
7625 | |
7626 | /* |
7627 | * Special case: If a kmalloc of a doms_cur partition (array of |
7628 | * cpumask) fails, then fallback to a single sched domain, |
7629 | * as determined by the single cpumask fallback_doms. |
7630 | */ |
7631 | static cpumask_var_t fallback_doms; |
7632 | |
7633 | /* |
7634 | * arch_update_cpu_topology lets virtualized architectures update the |
7635 | * cpu core maps. It is supposed to return 1 if the topology changed |
7636 | * or 0 if it stayed the same. |
7637 | */ |
7638 | int __attribute__((weak)) arch_update_cpu_topology(void) |
7639 | { |
7640 | return 0; |
7641 | } |
7642 | |
7643 | cpumask_var_t *alloc_sched_domains(unsigned int ndoms) |
7644 | { |
7645 | int i; |
7646 | cpumask_var_t *doms; |
7647 | |
7648 | doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); |
7649 | if (!doms) |
7650 | return NULL; |
7651 | for (i = 0; i < ndoms; i++) { |
7652 | if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { |
7653 | free_sched_domains(doms, i); |
7654 | return NULL; |
7655 | } |
7656 | } |
7657 | return doms; |
7658 | } |
7659 | |
7660 | void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) |
7661 | { |
7662 | unsigned int i; |
7663 | for (i = 0; i < ndoms; i++) |
7664 | free_cpumask_var(doms[i]); |
7665 | kfree(doms); |
7666 | } |
7667 | |
7668 | /* |
7669 | * Set up scheduler domains and groups. Callers must hold the hotplug lock. |
7670 | * For now this just excludes isolated cpus, but could be used to |
7671 | * exclude other special cases in the future. |
7672 | */ |
7673 | static int arch_init_sched_domains(const struct cpumask *cpu_map) |
7674 | { |
7675 | int err; |
7676 | |
7677 | arch_update_cpu_topology(); |
7678 | ndoms_cur = 1; |
7679 | doms_cur = alloc_sched_domains(ndoms_cur); |
7680 | if (!doms_cur) |
7681 | doms_cur = &fallback_doms; |
7682 | cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); |
7683 | dattr_cur = NULL; |
7684 | err = build_sched_domains(doms_cur[0]); |
7685 | register_sched_domain_sysctl(); |
7686 | |
7687 | return err; |
7688 | } |
7689 | |
7690 | static void arch_destroy_sched_domains(const struct cpumask *cpu_map, |
7691 | struct cpumask *tmpmask) |
7692 | { |
7693 | free_sched_groups(cpu_map, tmpmask); |
7694 | } |
7695 | |
7696 | /* |
7697 | * Detach sched domains from a group of cpus specified in cpu_map |
7698 | * These cpus will now be attached to the NULL domain |
7699 | */ |
7700 | static void detach_destroy_domains(const struct cpumask *cpu_map) |
7701 | { |
7702 | /* Save because hotplug lock held. */ |
7703 | static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS); |
7704 | int i; |
7705 | |
7706 | for_each_cpu(i, cpu_map) |
7707 | cpu_attach_domain(NULL, &def_root_domain, i); |
7708 | synchronize_sched(); |
7709 | arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask)); |
7710 | } |
7711 | |
7712 | /* handle null as "default" */ |
7713 | static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, |
7714 | struct sched_domain_attr *new, int idx_new) |
7715 | { |
7716 | struct sched_domain_attr tmp; |
7717 | |
7718 | /* fast path */ |
7719 | if (!new && !cur) |
7720 | return 1; |
7721 | |
7722 | tmp = SD_ATTR_INIT; |
7723 | return !memcmp(cur ? (cur + idx_cur) : &tmp, |
7724 | new ? (new + idx_new) : &tmp, |
7725 | sizeof(struct sched_domain_attr)); |
7726 | } |
7727 | |
7728 | /* |
7729 | * Partition sched domains as specified by the 'ndoms_new' |
7730 | * cpumasks in the array doms_new[] of cpumasks. This compares |
7731 | * doms_new[] to the current sched domain partitioning, doms_cur[]. |
7732 | * It destroys each deleted domain and builds each new domain. |
7733 | * |
7734 | * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. |
7735 | * The masks don't intersect (don't overlap.) We should setup one |
7736 | * sched domain for each mask. CPUs not in any of the cpumasks will |
7737 | * not be load balanced. If the same cpumask appears both in the |
7738 | * current 'doms_cur' domains and in the new 'doms_new', we can leave |
7739 | * it as it is. |
7740 | * |
7741 | * The passed in 'doms_new' should be allocated using |
7742 | * alloc_sched_domains. This routine takes ownership of it and will |
7743 | * free_sched_domains it when done with it. If the caller failed the |
7744 | * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, |
7745 | * and partition_sched_domains() will fallback to the single partition |
7746 | * 'fallback_doms', it also forces the domains to be rebuilt. |
7747 | * |
7748 | * If doms_new == NULL it will be replaced with cpu_online_mask. |
7749 | * ndoms_new == 0 is a special case for destroying existing domains, |
7750 | * and it will not create the default domain. |
7751 | * |
7752 | * Call with hotplug lock held |
7753 | */ |
7754 | void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], |
7755 | struct sched_domain_attr *dattr_new) |
7756 | { |
7757 | int i, j, n; |
7758 | int new_topology; |
7759 | |
7760 | mutex_lock(&sched_domains_mutex); |
7761 | |
7762 | /* always unregister in case we don't destroy any domains */ |
7763 | unregister_sched_domain_sysctl(); |
7764 | |
7765 | /* Let architecture update cpu core mappings. */ |
7766 | new_topology = arch_update_cpu_topology(); |
7767 | |
7768 | n = doms_new ? ndoms_new : 0; |
7769 | |
7770 | /* Destroy deleted domains */ |
7771 | for (i = 0; i < ndoms_cur; i++) { |
7772 | for (j = 0; j < n && !new_topology; j++) { |
7773 | if (cpumask_equal(doms_cur[i], doms_new[j]) |
7774 | && dattrs_equal(dattr_cur, i, dattr_new, j)) |
7775 | goto match1; |
7776 | } |
7777 | /* no match - a current sched domain not in new doms_new[] */ |
7778 | detach_destroy_domains(doms_cur[i]); |
7779 | match1: |
7780 | ; |
7781 | } |
7782 | |
7783 | if (doms_new == NULL) { |
7784 | ndoms_cur = 0; |
7785 | doms_new = &fallback_doms; |
7786 | cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); |
7787 | WARN_ON_ONCE(dattr_new); |
7788 | } |
7789 | |
7790 | /* Build new domains */ |
7791 | for (i = 0; i < ndoms_new; i++) { |
7792 | for (j = 0; j < ndoms_cur && !new_topology; j++) { |
7793 | if (cpumask_equal(doms_new[i], doms_cur[j]) |
7794 | && dattrs_equal(dattr_new, i, dattr_cur, j)) |
7795 | goto match2; |
7796 | } |
7797 | /* no match - add a new doms_new */ |
7798 | __build_sched_domains(doms_new[i], |
7799 | dattr_new ? dattr_new + i : NULL); |
7800 | match2: |
7801 | ; |
7802 | } |
7803 | |
7804 | /* Remember the new sched domains */ |
7805 | if (doms_cur != &fallback_doms) |
7806 | free_sched_domains(doms_cur, ndoms_cur); |
7807 | kfree(dattr_cur); /* kfree(NULL) is safe */ |
7808 | doms_cur = doms_new; |
7809 | dattr_cur = dattr_new; |
7810 | ndoms_cur = ndoms_new; |
7811 | |
7812 | register_sched_domain_sysctl(); |
7813 | |
7814 | mutex_unlock(&sched_domains_mutex); |
7815 | } |
7816 | |
7817 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
7818 | static void arch_reinit_sched_domains(void) |
7819 | { |
7820 | get_online_cpus(); |
7821 | |
7822 | /* Destroy domains first to force the rebuild */ |
7823 | partition_sched_domains(0, NULL, NULL); |
7824 | |
7825 | rebuild_sched_domains(); |
7826 | put_online_cpus(); |
7827 | } |
7828 | |
7829 | static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt) |
7830 | { |
7831 | unsigned int level = 0; |
7832 | |
7833 | if (sscanf(buf, "%u", &level) != 1) |
7834 | return -EINVAL; |
7835 | |
7836 | /* |
7837 | * level is always be positive so don't check for |
7838 | * level < POWERSAVINGS_BALANCE_NONE which is 0 |
7839 | * What happens on 0 or 1 byte write, |
7840 | * need to check for count as well? |
7841 | */ |
7842 | |
7843 | if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS) |
7844 | return -EINVAL; |
7845 | |
7846 | if (smt) |
7847 | sched_smt_power_savings = level; |
7848 | else |
7849 | sched_mc_power_savings = level; |
7850 | |
7851 | arch_reinit_sched_domains(); |
7852 | |
7853 | return count; |
7854 | } |
7855 | |
7856 | #ifdef CONFIG_SCHED_MC |
7857 | static ssize_t sched_mc_power_savings_show(struct sysdev_class *class, |
7858 | struct sysdev_class_attribute *attr, |
7859 | char *page) |
7860 | { |
7861 | return sprintf(page, "%u\n", sched_mc_power_savings); |
7862 | } |
7863 | static ssize_t sched_mc_power_savings_store(struct sysdev_class *class, |
7864 | struct sysdev_class_attribute *attr, |
7865 | const char *buf, size_t count) |
7866 | { |
7867 | return sched_power_savings_store(buf, count, 0); |
7868 | } |
7869 | static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644, |
7870 | sched_mc_power_savings_show, |
7871 | sched_mc_power_savings_store); |
7872 | #endif |
7873 | |
7874 | #ifdef CONFIG_SCHED_SMT |
7875 | static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev, |
7876 | struct sysdev_class_attribute *attr, |
7877 | char *page) |
7878 | { |
7879 | return sprintf(page, "%u\n", sched_smt_power_savings); |
7880 | } |
7881 | static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev, |
7882 | struct sysdev_class_attribute *attr, |
7883 | const char *buf, size_t count) |
7884 | { |
7885 | return sched_power_savings_store(buf, count, 1); |
7886 | } |
7887 | static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644, |
7888 | sched_smt_power_savings_show, |
7889 | sched_smt_power_savings_store); |
7890 | #endif |
7891 | |
7892 | int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls) |
7893 | { |
7894 | int err = 0; |
7895 | |
7896 | #ifdef CONFIG_SCHED_SMT |
7897 | if (smt_capable()) |
7898 | err = sysfs_create_file(&cls->kset.kobj, |
7899 | &attr_sched_smt_power_savings.attr); |
7900 | #endif |
7901 | #ifdef CONFIG_SCHED_MC |
7902 | if (!err && mc_capable()) |
7903 | err = sysfs_create_file(&cls->kset.kobj, |
7904 | &attr_sched_mc_power_savings.attr); |
7905 | #endif |
7906 | return err; |
7907 | } |
7908 | #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ |
7909 | |
7910 | /* |
7911 | * Update cpusets according to cpu_active mask. If cpusets are |
7912 | * disabled, cpuset_update_active_cpus() becomes a simple wrapper |
7913 | * around partition_sched_domains(). |
7914 | */ |
7915 | static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, |
7916 | void *hcpu) |
7917 | { |
7918 | switch (action & ~CPU_TASKS_FROZEN) { |
7919 | case CPU_ONLINE: |
7920 | case CPU_DOWN_FAILED: |
7921 | cpuset_update_active_cpus(); |
7922 | return NOTIFY_OK; |
7923 | default: |
7924 | return NOTIFY_DONE; |
7925 | } |
7926 | } |
7927 | |
7928 | static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, |
7929 | void *hcpu) |
7930 | { |
7931 | switch (action & ~CPU_TASKS_FROZEN) { |
7932 | case CPU_DOWN_PREPARE: |
7933 | cpuset_update_active_cpus(); |
7934 | return NOTIFY_OK; |
7935 | default: |
7936 | return NOTIFY_DONE; |
7937 | } |
7938 | } |
7939 | |
7940 | static int update_runtime(struct notifier_block *nfb, |
7941 | unsigned long action, void *hcpu) |
7942 | { |
7943 | int cpu = (int)(long)hcpu; |
7944 | |
7945 | switch (action) { |
7946 | case CPU_DOWN_PREPARE: |
7947 | case CPU_DOWN_PREPARE_FROZEN: |
7948 | disable_runtime(cpu_rq(cpu)); |
7949 | return NOTIFY_OK; |
7950 | |
7951 | case CPU_DOWN_FAILED: |
7952 | case CPU_DOWN_FAILED_FROZEN: |
7953 | case CPU_ONLINE: |
7954 | case CPU_ONLINE_FROZEN: |
7955 | enable_runtime(cpu_rq(cpu)); |
7956 | return NOTIFY_OK; |
7957 | |
7958 | default: |
7959 | return NOTIFY_DONE; |
7960 | } |
7961 | } |
7962 | |
7963 | void __init sched_init_smp(void) |
7964 | { |
7965 | cpumask_var_t non_isolated_cpus; |
7966 | |
7967 | alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); |
7968 | alloc_cpumask_var(&fallback_doms, GFP_KERNEL); |
7969 | |
7970 | #if defined(CONFIG_NUMA) |
7971 | sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **), |
7972 | GFP_KERNEL); |
7973 | BUG_ON(sched_group_nodes_bycpu == NULL); |
7974 | #endif |
7975 | get_online_cpus(); |
7976 | mutex_lock(&sched_domains_mutex); |
7977 | arch_init_sched_domains(cpu_active_mask); |
7978 | cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); |
7979 | if (cpumask_empty(non_isolated_cpus)) |
7980 | cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); |
7981 | mutex_unlock(&sched_domains_mutex); |
7982 | put_online_cpus(); |
7983 | |
7984 | hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); |
7985 | hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); |
7986 | |
7987 | /* RT runtime code needs to handle some hotplug events */ |
7988 | hotcpu_notifier(update_runtime, 0); |
7989 | |
7990 | init_hrtick(); |
7991 | |
7992 | /* Move init over to a non-isolated CPU */ |
7993 | if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) |
7994 | BUG(); |
7995 | sched_init_granularity(); |
7996 | free_cpumask_var(non_isolated_cpus); |
7997 | |
7998 | init_sched_rt_class(); |
7999 | } |
8000 | #else |
8001 | void __init sched_init_smp(void) |
8002 | { |
8003 | sched_init_granularity(); |
8004 | } |
8005 | #endif /* CONFIG_SMP */ |
8006 | |
8007 | const_debug unsigned int sysctl_timer_migration = 1; |
8008 | |
8009 | int in_sched_functions(unsigned long addr) |
8010 | { |
8011 | return in_lock_functions(addr) || |
8012 | (addr >= (unsigned long)__sched_text_start |
8013 | && addr < (unsigned long)__sched_text_end); |
8014 | } |
8015 | |
8016 | static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq) |
8017 | { |
8018 | cfs_rq->tasks_timeline = RB_ROOT; |
8019 | INIT_LIST_HEAD(&cfs_rq->tasks); |
8020 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8021 | cfs_rq->rq = rq; |
8022 | /* allow initial update_cfs_load() to truncate */ |
8023 | #ifdef CONFIG_SMP |
8024 | cfs_rq->load_stamp = 1; |
8025 | #endif |
8026 | #endif |
8027 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
8028 | } |
8029 | |
8030 | static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq) |
8031 | { |
8032 | struct rt_prio_array *array; |
8033 | int i; |
8034 | |
8035 | array = &rt_rq->active; |
8036 | for (i = 0; i < MAX_RT_PRIO; i++) { |
8037 | INIT_LIST_HEAD(array->queue + i); |
8038 | __clear_bit(i, array->bitmap); |
8039 | } |
8040 | /* delimiter for bitsearch: */ |
8041 | __set_bit(MAX_RT_PRIO, array->bitmap); |
8042 | |
8043 | #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED |
8044 | rt_rq->highest_prio.curr = MAX_RT_PRIO; |
8045 | #ifdef CONFIG_SMP |
8046 | rt_rq->highest_prio.next = MAX_RT_PRIO; |
8047 | #endif |
8048 | #endif |
8049 | #ifdef CONFIG_SMP |
8050 | rt_rq->rt_nr_migratory = 0; |
8051 | rt_rq->overloaded = 0; |
8052 | plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock); |
8053 | #endif |
8054 | |
8055 | rt_rq->rt_time = 0; |
8056 | rt_rq->rt_throttled = 0; |
8057 | rt_rq->rt_runtime = 0; |
8058 | raw_spin_lock_init(&rt_rq->rt_runtime_lock); |
8059 | |
8060 | #ifdef CONFIG_RT_GROUP_SCHED |
8061 | rt_rq->rt_nr_boosted = 0; |
8062 | rt_rq->rq = rq; |
8063 | #endif |
8064 | } |
8065 | |
8066 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8067 | static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, |
8068 | struct sched_entity *se, int cpu, |
8069 | struct sched_entity *parent) |
8070 | { |
8071 | struct rq *rq = cpu_rq(cpu); |
8072 | tg->cfs_rq[cpu] = cfs_rq; |
8073 | init_cfs_rq(cfs_rq, rq); |
8074 | cfs_rq->tg = tg; |
8075 | |
8076 | tg->se[cpu] = se; |
8077 | /* se could be NULL for root_task_group */ |
8078 | if (!se) |
8079 | return; |
8080 | |
8081 | if (!parent) |
8082 | se->cfs_rq = &rq->cfs; |
8083 | else |
8084 | se->cfs_rq = parent->my_q; |
8085 | |
8086 | se->my_q = cfs_rq; |
8087 | update_load_set(&se->load, 0); |
8088 | se->parent = parent; |
8089 | } |
8090 | #endif |
8091 | |
8092 | #ifdef CONFIG_RT_GROUP_SCHED |
8093 | static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq, |
8094 | struct sched_rt_entity *rt_se, int cpu, |
8095 | struct sched_rt_entity *parent) |
8096 | { |
8097 | struct rq *rq = cpu_rq(cpu); |
8098 | |
8099 | tg->rt_rq[cpu] = rt_rq; |
8100 | init_rt_rq(rt_rq, rq); |
8101 | rt_rq->tg = tg; |
8102 | rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime; |
8103 | |
8104 | tg->rt_se[cpu] = rt_se; |
8105 | if (!rt_se) |
8106 | return; |
8107 | |
8108 | if (!parent) |
8109 | rt_se->rt_rq = &rq->rt; |
8110 | else |
8111 | rt_se->rt_rq = parent->my_q; |
8112 | |
8113 | rt_se->my_q = rt_rq; |
8114 | rt_se->parent = parent; |
8115 | INIT_LIST_HEAD(&rt_se->run_list); |
8116 | } |
8117 | #endif |
8118 | |
8119 | void __init sched_init(void) |
8120 | { |
8121 | int i, j; |
8122 | unsigned long alloc_size = 0, ptr; |
8123 | |
8124 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8125 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); |
8126 | #endif |
8127 | #ifdef CONFIG_RT_GROUP_SCHED |
8128 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); |
8129 | #endif |
8130 | #ifdef CONFIG_CPUMASK_OFFSTACK |
8131 | alloc_size += num_possible_cpus() * cpumask_size(); |
8132 | #endif |
8133 | if (alloc_size) { |
8134 | ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); |
8135 | |
8136 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8137 | root_task_group.se = (struct sched_entity **)ptr; |
8138 | ptr += nr_cpu_ids * sizeof(void **); |
8139 | |
8140 | root_task_group.cfs_rq = (struct cfs_rq **)ptr; |
8141 | ptr += nr_cpu_ids * sizeof(void **); |
8142 | |
8143 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
8144 | #ifdef CONFIG_RT_GROUP_SCHED |
8145 | root_task_group.rt_se = (struct sched_rt_entity **)ptr; |
8146 | ptr += nr_cpu_ids * sizeof(void **); |
8147 | |
8148 | root_task_group.rt_rq = (struct rt_rq **)ptr; |
8149 | ptr += nr_cpu_ids * sizeof(void **); |
8150 | |
8151 | #endif /* CONFIG_RT_GROUP_SCHED */ |
8152 | #ifdef CONFIG_CPUMASK_OFFSTACK |
8153 | for_each_possible_cpu(i) { |
8154 | per_cpu(load_balance_tmpmask, i) = (void *)ptr; |
8155 | ptr += cpumask_size(); |
8156 | } |
8157 | #endif /* CONFIG_CPUMASK_OFFSTACK */ |
8158 | } |
8159 | |
8160 | #ifdef CONFIG_SMP |
8161 | init_defrootdomain(); |
8162 | #endif |
8163 | |
8164 | init_rt_bandwidth(&def_rt_bandwidth, |
8165 | global_rt_period(), global_rt_runtime()); |
8166 | |
8167 | #ifdef CONFIG_RT_GROUP_SCHED |
8168 | init_rt_bandwidth(&root_task_group.rt_bandwidth, |
8169 | global_rt_period(), global_rt_runtime()); |
8170 | #endif /* CONFIG_RT_GROUP_SCHED */ |
8171 | |
8172 | #ifdef CONFIG_CGROUP_SCHED |
8173 | list_add(&root_task_group.list, &task_groups); |
8174 | INIT_LIST_HEAD(&root_task_group.children); |
8175 | autogroup_init(&init_task); |
8176 | #endif /* CONFIG_CGROUP_SCHED */ |
8177 | |
8178 | for_each_possible_cpu(i) { |
8179 | struct rq *rq; |
8180 | |
8181 | rq = cpu_rq(i); |
8182 | raw_spin_lock_init(&rq->lock); |
8183 | rq->nr_running = 0; |
8184 | rq->calc_load_active = 0; |
8185 | rq->calc_load_update = jiffies + LOAD_FREQ; |
8186 | init_cfs_rq(&rq->cfs, rq); |
8187 | init_rt_rq(&rq->rt, rq); |
8188 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8189 | root_task_group.shares = root_task_group_load; |
8190 | INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); |
8191 | /* |
8192 | * How much cpu bandwidth does root_task_group get? |
8193 | * |
8194 | * In case of task-groups formed thr' the cgroup filesystem, it |
8195 | * gets 100% of the cpu resources in the system. This overall |
8196 | * system cpu resource is divided among the tasks of |
8197 | * root_task_group and its child task-groups in a fair manner, |
8198 | * based on each entity's (task or task-group's) weight |
8199 | * (se->load.weight). |
8200 | * |
8201 | * In other words, if root_task_group has 10 tasks of weight |
8202 | * 1024) and two child groups A0 and A1 (of weight 1024 each), |
8203 | * then A0's share of the cpu resource is: |
8204 | * |
8205 | * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% |
8206 | * |
8207 | * We achieve this by letting root_task_group's tasks sit |
8208 | * directly in rq->cfs (i.e root_task_group->se[] = NULL). |
8209 | */ |
8210 | init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); |
8211 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
8212 | |
8213 | rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; |
8214 | #ifdef CONFIG_RT_GROUP_SCHED |
8215 | INIT_LIST_HEAD(&rq->leaf_rt_rq_list); |
8216 | init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); |
8217 | #endif |
8218 | |
8219 | for (j = 0; j < CPU_LOAD_IDX_MAX; j++) |
8220 | rq->cpu_load[j] = 0; |
8221 | |
8222 | rq->last_load_update_tick = jiffies; |
8223 | |
8224 | #ifdef CONFIG_SMP |
8225 | rq->sd = NULL; |
8226 | rq->rd = NULL; |
8227 | rq->cpu_power = SCHED_LOAD_SCALE; |
8228 | rq->post_schedule = 0; |
8229 | rq->active_balance = 0; |
8230 | rq->next_balance = jiffies; |
8231 | rq->push_cpu = 0; |
8232 | rq->cpu = i; |
8233 | rq->online = 0; |
8234 | rq->idle_stamp = 0; |
8235 | rq->avg_idle = 2*sysctl_sched_migration_cost; |
8236 | rq_attach_root(rq, &def_root_domain); |
8237 | #ifdef CONFIG_NO_HZ |
8238 | rq->nohz_balance_kick = 0; |
8239 | init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i)); |
8240 | #endif |
8241 | #endif |
8242 | init_rq_hrtick(rq); |
8243 | atomic_set(&rq->nr_iowait, 0); |
8244 | } |
8245 | |
8246 | set_load_weight(&init_task); |
8247 | |
8248 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
8249 | INIT_HLIST_HEAD(&init_task.preempt_notifiers); |
8250 | #endif |
8251 | |
8252 | #ifdef CONFIG_SMP |
8253 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); |
8254 | #endif |
8255 | |
8256 | #ifdef CONFIG_RT_MUTEXES |
8257 | plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock); |
8258 | #endif |
8259 | |
8260 | /* |
8261 | * The boot idle thread does lazy MMU switching as well: |
8262 | */ |
8263 | atomic_inc(&init_mm.mm_count); |
8264 | enter_lazy_tlb(&init_mm, current); |
8265 | |
8266 | /* |
8267 | * Make us the idle thread. Technically, schedule() should not be |
8268 | * called from this thread, however somewhere below it might be, |
8269 | * but because we are the idle thread, we just pick up running again |
8270 | * when this runqueue becomes "idle". |
8271 | */ |
8272 | init_idle(current, smp_processor_id()); |
8273 | |
8274 | calc_load_update = jiffies + LOAD_FREQ; |
8275 | |
8276 | /* |
8277 | * During early bootup we pretend to be a normal task: |
8278 | */ |
8279 | current->sched_class = &fair_sched_class; |
8280 | |
8281 | /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */ |
8282 | zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT); |
8283 | #ifdef CONFIG_SMP |
8284 | #ifdef CONFIG_NO_HZ |
8285 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
8286 | alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT); |
8287 | atomic_set(&nohz.load_balancer, nr_cpu_ids); |
8288 | atomic_set(&nohz.first_pick_cpu, nr_cpu_ids); |
8289 | atomic_set(&nohz.second_pick_cpu, nr_cpu_ids); |
8290 | #endif |
8291 | /* May be allocated at isolcpus cmdline parse time */ |
8292 | if (cpu_isolated_map == NULL) |
8293 | zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); |
8294 | #endif /* SMP */ |
8295 | |
8296 | scheduler_running = 1; |
8297 | } |
8298 | |
8299 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP |
8300 | static inline int preempt_count_equals(int preempt_offset) |
8301 | { |
8302 | int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); |
8303 | |
8304 | return (nested == preempt_offset); |
8305 | } |
8306 | |
8307 | void __might_sleep(const char *file, int line, int preempt_offset) |
8308 | { |
8309 | #ifdef in_atomic |
8310 | static unsigned long prev_jiffy; /* ratelimiting */ |
8311 | |
8312 | if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || |
8313 | system_state != SYSTEM_RUNNING || oops_in_progress) |
8314 | return; |
8315 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) |
8316 | return; |
8317 | prev_jiffy = jiffies; |
8318 | |
8319 | printk(KERN_ERR |
8320 | "BUG: sleeping function called from invalid context at %s:%d\n", |
8321 | file, line); |
8322 | printk(KERN_ERR |
8323 | "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", |
8324 | in_atomic(), irqs_disabled(), |
8325 | current->pid, current->comm); |
8326 | |
8327 | debug_show_held_locks(current); |
8328 | if (irqs_disabled()) |
8329 | print_irqtrace_events(current); |
8330 | dump_stack(); |
8331 | #endif |
8332 | } |
8333 | EXPORT_SYMBOL(__might_sleep); |
8334 | #endif |
8335 | |
8336 | #ifdef CONFIG_MAGIC_SYSRQ |
8337 | static void normalize_task(struct rq *rq, struct task_struct *p) |
8338 | { |
8339 | const struct sched_class *prev_class = p->sched_class; |
8340 | int old_prio = p->prio; |
8341 | int on_rq; |
8342 | |
8343 | on_rq = p->se.on_rq; |
8344 | if (on_rq) |
8345 | deactivate_task(rq, p, 0); |
8346 | __setscheduler(rq, p, SCHED_NORMAL, 0); |
8347 | if (on_rq) { |
8348 | activate_task(rq, p, 0); |
8349 | resched_task(rq->curr); |
8350 | } |
8351 | |
8352 | check_class_changed(rq, p, prev_class, old_prio); |
8353 | } |
8354 | |
8355 | void normalize_rt_tasks(void) |
8356 | { |
8357 | struct task_struct *g, *p; |
8358 | unsigned long flags; |
8359 | struct rq *rq; |
8360 | |
8361 | read_lock_irqsave(&tasklist_lock, flags); |
8362 | do_each_thread(g, p) { |
8363 | /* |
8364 | * Only normalize user tasks: |
8365 | */ |
8366 | if (!p->mm) |
8367 | continue; |
8368 | |
8369 | p->se.exec_start = 0; |
8370 | #ifdef CONFIG_SCHEDSTATS |
8371 | p->se.statistics.wait_start = 0; |
8372 | p->se.statistics.sleep_start = 0; |
8373 | p->se.statistics.block_start = 0; |
8374 | #endif |
8375 | |
8376 | if (!rt_task(p)) { |
8377 | /* |
8378 | * Renice negative nice level userspace |
8379 | * tasks back to 0: |
8380 | */ |
8381 | if (TASK_NICE(p) < 0 && p->mm) |
8382 | set_user_nice(p, 0); |
8383 | continue; |
8384 | } |
8385 | |
8386 | raw_spin_lock(&p->pi_lock); |
8387 | rq = __task_rq_lock(p); |
8388 | |
8389 | normalize_task(rq, p); |
8390 | |
8391 | __task_rq_unlock(rq); |
8392 | raw_spin_unlock(&p->pi_lock); |
8393 | } while_each_thread(g, p); |
8394 | |
8395 | read_unlock_irqrestore(&tasklist_lock, flags); |
8396 | } |
8397 | |
8398 | #endif /* CONFIG_MAGIC_SYSRQ */ |
8399 | |
8400 | #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) |
8401 | /* |
8402 | * These functions are only useful for the IA64 MCA handling, or kdb. |
8403 | * |
8404 | * They can only be called when the whole system has been |
8405 | * stopped - every CPU needs to be quiescent, and no scheduling |
8406 | * activity can take place. Using them for anything else would |
8407 | * be a serious bug, and as a result, they aren't even visible |
8408 | * under any other configuration. |
8409 | */ |
8410 | |
8411 | /** |
8412 | * curr_task - return the current task for a given cpu. |
8413 | * @cpu: the processor in question. |
8414 | * |
8415 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! |
8416 | */ |
8417 | struct task_struct *curr_task(int cpu) |
8418 | { |
8419 | return cpu_curr(cpu); |
8420 | } |
8421 | |
8422 | #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ |
8423 | |
8424 | #ifdef CONFIG_IA64 |
8425 | /** |
8426 | * set_curr_task - set the current task for a given cpu. |
8427 | * @cpu: the processor in question. |
8428 | * @p: the task pointer to set. |
8429 | * |
8430 | * Description: This function must only be used when non-maskable interrupts |
8431 | * are serviced on a separate stack. It allows the architecture to switch the |
8432 | * notion of the current task on a cpu in a non-blocking manner. This function |
8433 | * must be called with all CPU's synchronized, and interrupts disabled, the |
8434 | * and caller must save the original value of the current task (see |
8435 | * curr_task() above) and restore that value before reenabling interrupts and |
8436 | * re-starting the system. |
8437 | * |
8438 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! |
8439 | */ |
8440 | void set_curr_task(int cpu, struct task_struct *p) |
8441 | { |
8442 | cpu_curr(cpu) = p; |
8443 | } |
8444 | |
8445 | #endif |
8446 | |
8447 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8448 | static void free_fair_sched_group(struct task_group *tg) |
8449 | { |
8450 | int i; |
8451 | |
8452 | for_each_possible_cpu(i) { |
8453 | if (tg->cfs_rq) |
8454 | kfree(tg->cfs_rq[i]); |
8455 | if (tg->se) |
8456 | kfree(tg->se[i]); |
8457 | } |
8458 | |
8459 | kfree(tg->cfs_rq); |
8460 | kfree(tg->se); |
8461 | } |
8462 | |
8463 | static |
8464 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) |
8465 | { |
8466 | struct cfs_rq *cfs_rq; |
8467 | struct sched_entity *se; |
8468 | int i; |
8469 | |
8470 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); |
8471 | if (!tg->cfs_rq) |
8472 | goto err; |
8473 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); |
8474 | if (!tg->se) |
8475 | goto err; |
8476 | |
8477 | tg->shares = NICE_0_LOAD; |
8478 | |
8479 | for_each_possible_cpu(i) { |
8480 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), |
8481 | GFP_KERNEL, cpu_to_node(i)); |
8482 | if (!cfs_rq) |
8483 | goto err; |
8484 | |
8485 | se = kzalloc_node(sizeof(struct sched_entity), |
8486 | GFP_KERNEL, cpu_to_node(i)); |
8487 | if (!se) |
8488 | goto err_free_rq; |
8489 | |
8490 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); |
8491 | } |
8492 | |
8493 | return 1; |
8494 | |
8495 | err_free_rq: |
8496 | kfree(cfs_rq); |
8497 | err: |
8498 | return 0; |
8499 | } |
8500 | |
8501 | static inline void unregister_fair_sched_group(struct task_group *tg, int cpu) |
8502 | { |
8503 | struct rq *rq = cpu_rq(cpu); |
8504 | unsigned long flags; |
8505 | |
8506 | /* |
8507 | * Only empty task groups can be destroyed; so we can speculatively |
8508 | * check on_list without danger of it being re-added. |
8509 | */ |
8510 | if (!tg->cfs_rq[cpu]->on_list) |
8511 | return; |
8512 | |
8513 | raw_spin_lock_irqsave(&rq->lock, flags); |
8514 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); |
8515 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
8516 | } |
8517 | #else /* !CONFG_FAIR_GROUP_SCHED */ |
8518 | static inline void free_fair_sched_group(struct task_group *tg) |
8519 | { |
8520 | } |
8521 | |
8522 | static inline |
8523 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) |
8524 | { |
8525 | return 1; |
8526 | } |
8527 | |
8528 | static inline void unregister_fair_sched_group(struct task_group *tg, int cpu) |
8529 | { |
8530 | } |
8531 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
8532 | |
8533 | #ifdef CONFIG_RT_GROUP_SCHED |
8534 | static void free_rt_sched_group(struct task_group *tg) |
8535 | { |
8536 | int i; |
8537 | |
8538 | destroy_rt_bandwidth(&tg->rt_bandwidth); |
8539 | |
8540 | for_each_possible_cpu(i) { |
8541 | if (tg->rt_rq) |
8542 | kfree(tg->rt_rq[i]); |
8543 | if (tg->rt_se) |
8544 | kfree(tg->rt_se[i]); |
8545 | } |
8546 | |
8547 | kfree(tg->rt_rq); |
8548 | kfree(tg->rt_se); |
8549 | } |
8550 | |
8551 | static |
8552 | int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) |
8553 | { |
8554 | struct rt_rq *rt_rq; |
8555 | struct sched_rt_entity *rt_se; |
8556 | struct rq *rq; |
8557 | int i; |
8558 | |
8559 | tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL); |
8560 | if (!tg->rt_rq) |
8561 | goto err; |
8562 | tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL); |
8563 | if (!tg->rt_se) |
8564 | goto err; |
8565 | |
8566 | init_rt_bandwidth(&tg->rt_bandwidth, |
8567 | ktime_to_ns(def_rt_bandwidth.rt_period), 0); |
8568 | |
8569 | for_each_possible_cpu(i) { |
8570 | rq = cpu_rq(i); |
8571 | |
8572 | rt_rq = kzalloc_node(sizeof(struct rt_rq), |
8573 | GFP_KERNEL, cpu_to_node(i)); |
8574 | if (!rt_rq) |
8575 | goto err; |
8576 | |
8577 | rt_se = kzalloc_node(sizeof(struct sched_rt_entity), |
8578 | GFP_KERNEL, cpu_to_node(i)); |
8579 | if (!rt_se) |
8580 | goto err_free_rq; |
8581 | |
8582 | init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]); |
8583 | } |
8584 | |
8585 | return 1; |
8586 | |
8587 | err_free_rq: |
8588 | kfree(rt_rq); |
8589 | err: |
8590 | return 0; |
8591 | } |
8592 | #else /* !CONFIG_RT_GROUP_SCHED */ |
8593 | static inline void free_rt_sched_group(struct task_group *tg) |
8594 | { |
8595 | } |
8596 | |
8597 | static inline |
8598 | int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) |
8599 | { |
8600 | return 1; |
8601 | } |
8602 | #endif /* CONFIG_RT_GROUP_SCHED */ |
8603 | |
8604 | #ifdef CONFIG_CGROUP_SCHED |
8605 | static void free_sched_group(struct task_group *tg) |
8606 | { |
8607 | free_fair_sched_group(tg); |
8608 | free_rt_sched_group(tg); |
8609 | autogroup_free(tg); |
8610 | kfree(tg); |
8611 | } |
8612 | |
8613 | /* allocate runqueue etc for a new task group */ |
8614 | struct task_group *sched_create_group(struct task_group *parent) |
8615 | { |
8616 | struct task_group *tg; |
8617 | unsigned long flags; |
8618 | |
8619 | tg = kzalloc(sizeof(*tg), GFP_KERNEL); |
8620 | if (!tg) |
8621 | return ERR_PTR(-ENOMEM); |
8622 | |
8623 | if (!alloc_fair_sched_group(tg, parent)) |
8624 | goto err; |
8625 | |
8626 | if (!alloc_rt_sched_group(tg, parent)) |
8627 | goto err; |
8628 | |
8629 | spin_lock_irqsave(&task_group_lock, flags); |
8630 | list_add_rcu(&tg->list, &task_groups); |
8631 | |
8632 | WARN_ON(!parent); /* root should already exist */ |
8633 | |
8634 | tg->parent = parent; |
8635 | INIT_LIST_HEAD(&tg->children); |
8636 | list_add_rcu(&tg->siblings, &parent->children); |
8637 | spin_unlock_irqrestore(&task_group_lock, flags); |
8638 | |
8639 | return tg; |
8640 | |
8641 | err: |
8642 | free_sched_group(tg); |
8643 | return ERR_PTR(-ENOMEM); |
8644 | } |
8645 | |
8646 | /* rcu callback to free various structures associated with a task group */ |
8647 | static void free_sched_group_rcu(struct rcu_head *rhp) |
8648 | { |
8649 | /* now it should be safe to free those cfs_rqs */ |
8650 | free_sched_group(container_of(rhp, struct task_group, rcu)); |
8651 | } |
8652 | |
8653 | /* Destroy runqueue etc associated with a task group */ |
8654 | void sched_destroy_group(struct task_group *tg) |
8655 | { |
8656 | unsigned long flags; |
8657 | int i; |
8658 | |
8659 | /* end participation in shares distribution */ |
8660 | for_each_possible_cpu(i) |
8661 | unregister_fair_sched_group(tg, i); |
8662 | |
8663 | spin_lock_irqsave(&task_group_lock, flags); |
8664 | list_del_rcu(&tg->list); |
8665 | list_del_rcu(&tg->siblings); |
8666 | spin_unlock_irqrestore(&task_group_lock, flags); |
8667 | |
8668 | /* wait for possible concurrent references to cfs_rqs complete */ |
8669 | call_rcu(&tg->rcu, free_sched_group_rcu); |
8670 | } |
8671 | |
8672 | /* change task's runqueue when it moves between groups. |
8673 | * The caller of this function should have put the task in its new group |
8674 | * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to |
8675 | * reflect its new group. |
8676 | */ |
8677 | void sched_move_task(struct task_struct *tsk) |
8678 | { |
8679 | int on_rq, running; |
8680 | unsigned long flags; |
8681 | struct rq *rq; |
8682 | |
8683 | rq = task_rq_lock(tsk, &flags); |
8684 | |
8685 | running = task_current(rq, tsk); |
8686 | on_rq = tsk->se.on_rq; |
8687 | |
8688 | if (on_rq) |
8689 | dequeue_task(rq, tsk, 0); |
8690 | if (unlikely(running)) |
8691 | tsk->sched_class->put_prev_task(rq, tsk); |
8692 | |
8693 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8694 | if (tsk->sched_class->task_move_group) |
8695 | tsk->sched_class->task_move_group(tsk, on_rq); |
8696 | else |
8697 | #endif |
8698 | set_task_rq(tsk, task_cpu(tsk)); |
8699 | |
8700 | if (unlikely(running)) |
8701 | tsk->sched_class->set_curr_task(rq); |
8702 | if (on_rq) |
8703 | enqueue_task(rq, tsk, 0); |
8704 | |
8705 | task_rq_unlock(rq, &flags); |
8706 | } |
8707 | #endif /* CONFIG_CGROUP_SCHED */ |
8708 | |
8709 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8710 | static DEFINE_MUTEX(shares_mutex); |
8711 | |
8712 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) |
8713 | { |
8714 | int i; |
8715 | unsigned long flags; |
8716 | |
8717 | /* |
8718 | * We can't change the weight of the root cgroup. |
8719 | */ |
8720 | if (!tg->se[0]) |
8721 | return -EINVAL; |
8722 | |
8723 | if (shares < MIN_SHARES) |
8724 | shares = MIN_SHARES; |
8725 | else if (shares > MAX_SHARES) |
8726 | shares = MAX_SHARES; |
8727 | |
8728 | mutex_lock(&shares_mutex); |
8729 | if (tg->shares == shares) |
8730 | goto done; |
8731 | |
8732 | tg->shares = shares; |
8733 | for_each_possible_cpu(i) { |
8734 | struct rq *rq = cpu_rq(i); |
8735 | struct sched_entity *se; |
8736 | |
8737 | se = tg->se[i]; |
8738 | /* Propagate contribution to hierarchy */ |
8739 | raw_spin_lock_irqsave(&rq->lock, flags); |
8740 | for_each_sched_entity(se) |
8741 | update_cfs_shares(group_cfs_rq(se)); |
8742 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
8743 | } |
8744 | |
8745 | done: |
8746 | mutex_unlock(&shares_mutex); |
8747 | return 0; |
8748 | } |
8749 | |
8750 | unsigned long sched_group_shares(struct task_group *tg) |
8751 | { |
8752 | return tg->shares; |
8753 | } |
8754 | #endif |
8755 | |
8756 | #ifdef CONFIG_RT_GROUP_SCHED |
8757 | /* |
8758 | * Ensure that the real time constraints are schedulable. |
8759 | */ |
8760 | static DEFINE_MUTEX(rt_constraints_mutex); |
8761 | |
8762 | static unsigned long to_ratio(u64 period, u64 runtime) |
8763 | { |
8764 | if (runtime == RUNTIME_INF) |
8765 | return 1ULL << 20; |
8766 | |
8767 | return div64_u64(runtime << 20, period); |
8768 | } |
8769 | |
8770 | /* Must be called with tasklist_lock held */ |
8771 | static inline int tg_has_rt_tasks(struct task_group *tg) |
8772 | { |
8773 | struct task_struct *g, *p; |
8774 | |
8775 | do_each_thread(g, p) { |
8776 | if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg) |
8777 | return 1; |
8778 | } while_each_thread(g, p); |
8779 | |
8780 | return 0; |
8781 | } |
8782 | |
8783 | struct rt_schedulable_data { |
8784 | struct task_group *tg; |
8785 | u64 rt_period; |
8786 | u64 rt_runtime; |
8787 | }; |
8788 | |
8789 | static int tg_schedulable(struct task_group *tg, void *data) |
8790 | { |
8791 | struct rt_schedulable_data *d = data; |
8792 | struct task_group *child; |
8793 | unsigned long total, sum = 0; |
8794 | u64 period, runtime; |
8795 | |
8796 | period = ktime_to_ns(tg->rt_bandwidth.rt_period); |
8797 | runtime = tg->rt_bandwidth.rt_runtime; |
8798 | |
8799 | if (tg == d->tg) { |
8800 | period = d->rt_period; |
8801 | runtime = d->rt_runtime; |
8802 | } |
8803 | |
8804 | /* |
8805 | * Cannot have more runtime than the period. |
8806 | */ |
8807 | if (runtime > period && runtime != RUNTIME_INF) |
8808 | return -EINVAL; |
8809 | |
8810 | /* |
8811 | * Ensure we don't starve existing RT tasks. |
8812 | */ |
8813 | if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) |
8814 | return -EBUSY; |
8815 | |
8816 | total = to_ratio(period, runtime); |
8817 | |
8818 | /* |
8819 | * Nobody can have more than the global setting allows. |
8820 | */ |
8821 | if (total > to_ratio(global_rt_period(), global_rt_runtime())) |
8822 | return -EINVAL; |
8823 | |
8824 | /* |
8825 | * The sum of our children's runtime should not exceed our own. |
8826 | */ |
8827 | list_for_each_entry_rcu(child, &tg->children, siblings) { |
8828 | period = ktime_to_ns(child->rt_bandwidth.rt_period); |
8829 | runtime = child->rt_bandwidth.rt_runtime; |
8830 | |
8831 | if (child == d->tg) { |
8832 | period = d->rt_period; |
8833 | runtime = d->rt_runtime; |
8834 | } |
8835 | |
8836 | sum += to_ratio(period, runtime); |
8837 | } |
8838 | |
8839 | if (sum > total) |
8840 | return -EINVAL; |
8841 | |
8842 | return 0; |
8843 | } |
8844 | |
8845 | static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) |
8846 | { |
8847 | struct rt_schedulable_data data = { |
8848 | .tg = tg, |
8849 | .rt_period = period, |
8850 | .rt_runtime = runtime, |
8851 | }; |
8852 | |
8853 | return walk_tg_tree(tg_schedulable, tg_nop, &data); |
8854 | } |
8855 | |
8856 | static int tg_set_bandwidth(struct task_group *tg, |
8857 | u64 rt_period, u64 rt_runtime) |
8858 | { |
8859 | int i, err = 0; |
8860 | |
8861 | mutex_lock(&rt_constraints_mutex); |
8862 | read_lock(&tasklist_lock); |
8863 | err = __rt_schedulable(tg, rt_period, rt_runtime); |
8864 | if (err) |
8865 | goto unlock; |
8866 | |
8867 | raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); |
8868 | tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); |
8869 | tg->rt_bandwidth.rt_runtime = rt_runtime; |
8870 | |
8871 | for_each_possible_cpu(i) { |
8872 | struct rt_rq *rt_rq = tg->rt_rq[i]; |
8873 | |
8874 | raw_spin_lock(&rt_rq->rt_runtime_lock); |
8875 | rt_rq->rt_runtime = rt_runtime; |
8876 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
8877 | } |
8878 | raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); |
8879 | unlock: |
8880 | read_unlock(&tasklist_lock); |
8881 | mutex_unlock(&rt_constraints_mutex); |
8882 | |
8883 | return err; |
8884 | } |
8885 | |
8886 | int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) |
8887 | { |
8888 | u64 rt_runtime, rt_period; |
8889 | |
8890 | rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); |
8891 | rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; |
8892 | if (rt_runtime_us < 0) |
8893 | rt_runtime = RUNTIME_INF; |
8894 | |
8895 | return tg_set_bandwidth(tg, rt_period, rt_runtime); |
8896 | } |
8897 | |
8898 | long sched_group_rt_runtime(struct task_group *tg) |
8899 | { |
8900 | u64 rt_runtime_us; |
8901 | |
8902 | if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) |
8903 | return -1; |
8904 | |
8905 | rt_runtime_us = tg->rt_bandwidth.rt_runtime; |
8906 | do_div(rt_runtime_us, NSEC_PER_USEC); |
8907 | return rt_runtime_us; |
8908 | } |
8909 | |
8910 | int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) |
8911 | { |
8912 | u64 rt_runtime, rt_period; |
8913 | |
8914 | rt_period = (u64)rt_period_us * NSEC_PER_USEC; |
8915 | rt_runtime = tg->rt_bandwidth.rt_runtime; |
8916 | |
8917 | if (rt_period == 0) |
8918 | return -EINVAL; |
8919 | |
8920 | return tg_set_bandwidth(tg, rt_period, rt_runtime); |
8921 | } |
8922 | |
8923 | long sched_group_rt_period(struct task_group *tg) |
8924 | { |
8925 | u64 rt_period_us; |
8926 | |
8927 | rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); |
8928 | do_div(rt_period_us, NSEC_PER_USEC); |
8929 | return rt_period_us; |
8930 | } |
8931 | |
8932 | static int sched_rt_global_constraints(void) |
8933 | { |
8934 | u64 runtime, period; |
8935 | int ret = 0; |
8936 | |
8937 | if (sysctl_sched_rt_period <= 0) |
8938 | return -EINVAL; |
8939 | |
8940 | runtime = global_rt_runtime(); |
8941 | period = global_rt_period(); |
8942 | |
8943 | /* |
8944 | * Sanity check on the sysctl variables. |
8945 | */ |
8946 | if (runtime > period && runtime != RUNTIME_INF) |
8947 | return -EINVAL; |
8948 | |
8949 | mutex_lock(&rt_constraints_mutex); |
8950 | read_lock(&tasklist_lock); |
8951 | ret = __rt_schedulable(NULL, 0, 0); |
8952 | read_unlock(&tasklist_lock); |
8953 | mutex_unlock(&rt_constraints_mutex); |
8954 | |
8955 | return ret; |
8956 | } |
8957 | |
8958 | int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) |
8959 | { |
8960 | /* Don't accept realtime tasks when there is no way for them to run */ |
8961 | if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) |
8962 | return 0; |
8963 | |
8964 | return 1; |
8965 | } |
8966 | |
8967 | #else /* !CONFIG_RT_GROUP_SCHED */ |
8968 | static int sched_rt_global_constraints(void) |
8969 | { |
8970 | unsigned long flags; |
8971 | int i; |
8972 | |
8973 | if (sysctl_sched_rt_period <= 0) |
8974 | return -EINVAL; |
8975 | |
8976 | /* |
8977 | * There's always some RT tasks in the root group |
8978 | * -- migration, kstopmachine etc.. |
8979 | */ |
8980 | if (sysctl_sched_rt_runtime == 0) |
8981 | return -EBUSY; |
8982 | |
8983 | raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); |
8984 | for_each_possible_cpu(i) { |
8985 | struct rt_rq *rt_rq = &cpu_rq(i)->rt; |
8986 | |
8987 | raw_spin_lock(&rt_rq->rt_runtime_lock); |
8988 | rt_rq->rt_runtime = global_rt_runtime(); |
8989 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
8990 | } |
8991 | raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); |
8992 | |
8993 | return 0; |
8994 | } |
8995 | #endif /* CONFIG_RT_GROUP_SCHED */ |
8996 | |
8997 | int sched_rt_handler(struct ctl_table *table, int write, |
8998 | void __user *buffer, size_t *lenp, |
8999 | loff_t *ppos) |
9000 | { |
9001 | int ret; |
9002 | int old_period, old_runtime; |
9003 | static DEFINE_MUTEX(mutex); |
9004 | |
9005 | mutex_lock(&mutex); |
9006 | old_period = sysctl_sched_rt_period; |
9007 | old_runtime = sysctl_sched_rt_runtime; |
9008 | |
9009 | ret = proc_dointvec(table, write, buffer, lenp, ppos); |
9010 | |
9011 | if (!ret && write) { |
9012 | ret = sched_rt_global_constraints(); |
9013 | if (ret) { |
9014 | sysctl_sched_rt_period = old_period; |
9015 | sysctl_sched_rt_runtime = old_runtime; |
9016 | } else { |
9017 | def_rt_bandwidth.rt_runtime = global_rt_runtime(); |
9018 | def_rt_bandwidth.rt_period = |
9019 | ns_to_ktime(global_rt_period()); |
9020 | } |
9021 | } |
9022 | mutex_unlock(&mutex); |
9023 | |
9024 | return ret; |
9025 | } |
9026 | |
9027 | #ifdef CONFIG_CGROUP_SCHED |
9028 | |
9029 | /* return corresponding task_group object of a cgroup */ |
9030 | static inline struct task_group *cgroup_tg(struct cgroup *cgrp) |
9031 | { |
9032 | return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id), |
9033 | struct task_group, css); |
9034 | } |
9035 | |
9036 | static struct cgroup_subsys_state * |
9037 | cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp) |
9038 | { |
9039 | struct task_group *tg, *parent; |
9040 | |
9041 | if (!cgrp->parent) { |
9042 | /* This is early initialization for the top cgroup */ |
9043 | return &root_task_group.css; |
9044 | } |
9045 | |
9046 | parent = cgroup_tg(cgrp->parent); |
9047 | tg = sched_create_group(parent); |
9048 | if (IS_ERR(tg)) |
9049 | return ERR_PTR(-ENOMEM); |
9050 | |
9051 | return &tg->css; |
9052 | } |
9053 | |
9054 | static void |
9055 | cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp) |
9056 | { |
9057 | struct task_group *tg = cgroup_tg(cgrp); |
9058 | |
9059 | sched_destroy_group(tg); |
9060 | } |
9061 | |
9062 | static int |
9063 | cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk) |
9064 | { |
9065 | #ifdef CONFIG_RT_GROUP_SCHED |
9066 | if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk)) |
9067 | return -EINVAL; |
9068 | #else |
9069 | /* We don't support RT-tasks being in separate groups */ |
9070 | if (tsk->sched_class != &fair_sched_class) |
9071 | return -EINVAL; |
9072 | #endif |
9073 | return 0; |
9074 | } |
9075 | |
9076 | static int |
9077 | cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp, |
9078 | struct task_struct *tsk, bool threadgroup) |
9079 | { |
9080 | int retval = cpu_cgroup_can_attach_task(cgrp, tsk); |
9081 | if (retval) |
9082 | return retval; |
9083 | if (threadgroup) { |
9084 | struct task_struct *c; |
9085 | rcu_read_lock(); |
9086 | list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) { |
9087 | retval = cpu_cgroup_can_attach_task(cgrp, c); |
9088 | if (retval) { |
9089 | rcu_read_unlock(); |
9090 | return retval; |
9091 | } |
9092 | } |
9093 | rcu_read_unlock(); |
9094 | } |
9095 | return 0; |
9096 | } |
9097 | |
9098 | static void |
9099 | cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp, |
9100 | struct cgroup *old_cont, struct task_struct *tsk, |
9101 | bool threadgroup) |
9102 | { |
9103 | sched_move_task(tsk); |
9104 | if (threadgroup) { |
9105 | struct task_struct *c; |
9106 | rcu_read_lock(); |
9107 | list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) { |
9108 | sched_move_task(c); |
9109 | } |
9110 | rcu_read_unlock(); |
9111 | } |
9112 | } |
9113 | |
9114 | static void |
9115 | cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp, |
9116 | struct cgroup *old_cgrp, struct task_struct *task) |
9117 | { |
9118 | /* |
9119 | * cgroup_exit() is called in the copy_process() failure path. |
9120 | * Ignore this case since the task hasn't ran yet, this avoids |
9121 | * trying to poke a half freed task state from generic code. |
9122 | */ |
9123 | if (!(task->flags & PF_EXITING)) |
9124 | return; |
9125 | |
9126 | sched_move_task(task); |
9127 | } |
9128 | |
9129 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9130 | static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype, |
9131 | u64 shareval) |
9132 | { |
9133 | return sched_group_set_shares(cgroup_tg(cgrp), shareval); |
9134 | } |
9135 | |
9136 | static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft) |
9137 | { |
9138 | struct task_group *tg = cgroup_tg(cgrp); |
9139 | |
9140 | return (u64) tg->shares; |
9141 | } |
9142 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
9143 | |
9144 | #ifdef CONFIG_RT_GROUP_SCHED |
9145 | static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft, |
9146 | s64 val) |
9147 | { |
9148 | return sched_group_set_rt_runtime(cgroup_tg(cgrp), val); |
9149 | } |
9150 | |
9151 | static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft) |
9152 | { |
9153 | return sched_group_rt_runtime(cgroup_tg(cgrp)); |
9154 | } |
9155 | |
9156 | static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype, |
9157 | u64 rt_period_us) |
9158 | { |
9159 | return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us); |
9160 | } |
9161 | |
9162 | static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft) |
9163 | { |
9164 | return sched_group_rt_period(cgroup_tg(cgrp)); |
9165 | } |
9166 | #endif /* CONFIG_RT_GROUP_SCHED */ |
9167 | |
9168 | static struct cftype cpu_files[] = { |
9169 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9170 | { |
9171 | .name = "shares", |
9172 | .read_u64 = cpu_shares_read_u64, |
9173 | .write_u64 = cpu_shares_write_u64, |
9174 | }, |
9175 | #endif |
9176 | #ifdef CONFIG_RT_GROUP_SCHED |
9177 | { |
9178 | .name = "rt_runtime_us", |
9179 | .read_s64 = cpu_rt_runtime_read, |
9180 | .write_s64 = cpu_rt_runtime_write, |
9181 | }, |
9182 | { |
9183 | .name = "rt_period_us", |
9184 | .read_u64 = cpu_rt_period_read_uint, |
9185 | .write_u64 = cpu_rt_period_write_uint, |
9186 | }, |
9187 | #endif |
9188 | }; |
9189 | |
9190 | static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont) |
9191 | { |
9192 | return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files)); |
9193 | } |
9194 | |
9195 | struct cgroup_subsys cpu_cgroup_subsys = { |
9196 | .name = "cpu", |
9197 | .create = cpu_cgroup_create, |
9198 | .destroy = cpu_cgroup_destroy, |
9199 | .can_attach = cpu_cgroup_can_attach, |
9200 | .attach = cpu_cgroup_attach, |
9201 | .exit = cpu_cgroup_exit, |
9202 | .populate = cpu_cgroup_populate, |
9203 | .subsys_id = cpu_cgroup_subsys_id, |
9204 | .early_init = 1, |
9205 | }; |
9206 | |
9207 | #endif /* CONFIG_CGROUP_SCHED */ |
9208 | |
9209 | #ifdef CONFIG_CGROUP_CPUACCT |
9210 | |
9211 | /* |
9212 | * CPU accounting code for task groups. |
9213 | * |
9214 | * Based on the work by Paul Menage (menage@google.com) and Balbir Singh |
9215 | * (balbir@in.ibm.com). |
9216 | */ |
9217 | |
9218 | /* track cpu usage of a group of tasks and its child groups */ |
9219 | struct cpuacct { |
9220 | struct cgroup_subsys_state css; |
9221 | /* cpuusage holds pointer to a u64-type object on every cpu */ |
9222 | u64 __percpu *cpuusage; |
9223 | struct percpu_counter cpustat[CPUACCT_STAT_NSTATS]; |
9224 | struct cpuacct *parent; |
9225 | }; |
9226 | |
9227 | struct cgroup_subsys cpuacct_subsys; |
9228 | |
9229 | /* return cpu accounting group corresponding to this container */ |
9230 | static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp) |
9231 | { |
9232 | return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id), |
9233 | struct cpuacct, css); |
9234 | } |
9235 | |
9236 | /* return cpu accounting group to which this task belongs */ |
9237 | static inline struct cpuacct *task_ca(struct task_struct *tsk) |
9238 | { |
9239 | return container_of(task_subsys_state(tsk, cpuacct_subsys_id), |
9240 | struct cpuacct, css); |
9241 | } |
9242 | |
9243 | /* create a new cpu accounting group */ |
9244 | static struct cgroup_subsys_state *cpuacct_create( |
9245 | struct cgroup_subsys *ss, struct cgroup *cgrp) |
9246 | { |
9247 | struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL); |
9248 | int i; |
9249 | |
9250 | if (!ca) |
9251 | goto out; |
9252 | |
9253 | ca->cpuusage = alloc_percpu(u64); |
9254 | if (!ca->cpuusage) |
9255 | goto out_free_ca; |
9256 | |
9257 | for (i = 0; i < CPUACCT_STAT_NSTATS; i++) |
9258 | if (percpu_counter_init(&ca->cpustat[i], 0)) |
9259 | goto out_free_counters; |
9260 | |
9261 | if (cgrp->parent) |
9262 | ca->parent = cgroup_ca(cgrp->parent); |
9263 | |
9264 | return &ca->css; |
9265 | |
9266 | out_free_counters: |
9267 | while (--i >= 0) |
9268 | percpu_counter_destroy(&ca->cpustat[i]); |
9269 | free_percpu(ca->cpuusage); |
9270 | out_free_ca: |
9271 | kfree(ca); |
9272 | out: |
9273 | return ERR_PTR(-ENOMEM); |
9274 | } |
9275 | |
9276 | /* destroy an existing cpu accounting group */ |
9277 | static void |
9278 | cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp) |
9279 | { |
9280 | struct cpuacct *ca = cgroup_ca(cgrp); |
9281 | int i; |
9282 | |
9283 | for (i = 0; i < CPUACCT_STAT_NSTATS; i++) |
9284 | percpu_counter_destroy(&ca->cpustat[i]); |
9285 | free_percpu(ca->cpuusage); |
9286 | kfree(ca); |
9287 | } |
9288 | |
9289 | static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu) |
9290 | { |
9291 | u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); |
9292 | u64 data; |
9293 | |
9294 | #ifndef CONFIG_64BIT |
9295 | /* |
9296 | * Take rq->lock to make 64-bit read safe on 32-bit platforms. |
9297 | */ |
9298 | raw_spin_lock_irq(&cpu_rq(cpu)->lock); |
9299 | data = *cpuusage; |
9300 | raw_spin_unlock_irq(&cpu_rq(cpu)->lock); |
9301 | #else |
9302 | data = *cpuusage; |
9303 | #endif |
9304 | |
9305 | return data; |
9306 | } |
9307 | |
9308 | static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val) |
9309 | { |
9310 | u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); |
9311 | |
9312 | #ifndef CONFIG_64BIT |
9313 | /* |
9314 | * Take rq->lock to make 64-bit write safe on 32-bit platforms. |
9315 | */ |
9316 | raw_spin_lock_irq(&cpu_rq(cpu)->lock); |
9317 | *cpuusage = val; |
9318 | raw_spin_unlock_irq(&cpu_rq(cpu)->lock); |
9319 | #else |
9320 | *cpuusage = val; |
9321 | #endif |
9322 | } |
9323 | |
9324 | /* return total cpu usage (in nanoseconds) of a group */ |
9325 | static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft) |
9326 | { |
9327 | struct cpuacct *ca = cgroup_ca(cgrp); |
9328 | u64 totalcpuusage = 0; |
9329 | int i; |
9330 | |
9331 | for_each_present_cpu(i) |
9332 | totalcpuusage += cpuacct_cpuusage_read(ca, i); |
9333 | |
9334 | return totalcpuusage; |
9335 | } |
9336 | |
9337 | static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype, |
9338 | u64 reset) |
9339 | { |
9340 | struct cpuacct *ca = cgroup_ca(cgrp); |
9341 | int err = 0; |
9342 | int i; |
9343 | |
9344 | if (reset) { |
9345 | err = -EINVAL; |
9346 | goto out; |
9347 | } |
9348 | |
9349 | for_each_present_cpu(i) |
9350 | cpuacct_cpuusage_write(ca, i, 0); |
9351 | |
9352 | out: |
9353 | return err; |
9354 | } |
9355 | |
9356 | static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft, |
9357 | struct seq_file *m) |
9358 | { |
9359 | struct cpuacct *ca = cgroup_ca(cgroup); |
9360 | u64 percpu; |
9361 | int i; |
9362 | |
9363 | for_each_present_cpu(i) { |
9364 | percpu = cpuacct_cpuusage_read(ca, i); |
9365 | seq_printf(m, "%llu ", (unsigned long long) percpu); |
9366 | } |
9367 | seq_printf(m, "\n"); |
9368 | return 0; |
9369 | } |
9370 | |
9371 | static const char *cpuacct_stat_desc[] = { |
9372 | [CPUACCT_STAT_USER] = "user", |
9373 | [CPUACCT_STAT_SYSTEM] = "system", |
9374 | }; |
9375 | |
9376 | static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft, |
9377 | struct cgroup_map_cb *cb) |
9378 | { |
9379 | struct cpuacct *ca = cgroup_ca(cgrp); |
9380 | int i; |
9381 | |
9382 | for (i = 0; i < CPUACCT_STAT_NSTATS; i++) { |
9383 | s64 val = percpu_counter_read(&ca->cpustat[i]); |
9384 | val = cputime64_to_clock_t(val); |
9385 | cb->fill(cb, cpuacct_stat_desc[i], val); |
9386 | } |
9387 | return 0; |
9388 | } |
9389 | |
9390 | static struct cftype files[] = { |
9391 | { |
9392 | .name = "usage", |
9393 | .read_u64 = cpuusage_read, |
9394 | .write_u64 = cpuusage_write, |
9395 | }, |
9396 | { |
9397 | .name = "usage_percpu", |
9398 | .read_seq_string = cpuacct_percpu_seq_read, |
9399 | }, |
9400 | { |
9401 | .name = "stat", |
9402 | .read_map = cpuacct_stats_show, |
9403 | }, |
9404 | }; |
9405 | |
9406 | static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp) |
9407 | { |
9408 | return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files)); |
9409 | } |
9410 | |
9411 | /* |
9412 | * charge this task's execution time to its accounting group. |
9413 | * |
9414 | * called with rq->lock held. |
9415 | */ |
9416 | static void cpuacct_charge(struct task_struct *tsk, u64 cputime) |
9417 | { |
9418 | struct cpuacct *ca; |
9419 | int cpu; |
9420 | |
9421 | if (unlikely(!cpuacct_subsys.active)) |
9422 | return; |
9423 | |
9424 | cpu = task_cpu(tsk); |
9425 | |
9426 | rcu_read_lock(); |
9427 | |
9428 | ca = task_ca(tsk); |
9429 | |
9430 | for (; ca; ca = ca->parent) { |
9431 | u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); |
9432 | *cpuusage += cputime; |
9433 | } |
9434 | |
9435 | rcu_read_unlock(); |
9436 | } |
9437 | |
9438 | /* |
9439 | * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large |
9440 | * in cputime_t units. As a result, cpuacct_update_stats calls |
9441 | * percpu_counter_add with values large enough to always overflow the |
9442 | * per cpu batch limit causing bad SMP scalability. |
9443 | * |
9444 | * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we |
9445 | * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled |
9446 | * and enabled. We cap it at INT_MAX which is the largest allowed batch value. |
9447 | */ |
9448 | #ifdef CONFIG_SMP |
9449 | #define CPUACCT_BATCH \ |
9450 | min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX) |
9451 | #else |
9452 | #define CPUACCT_BATCH 0 |
9453 | #endif |
9454 | |
9455 | /* |
9456 | * Charge the system/user time to the task's accounting group. |
9457 | */ |
9458 | static void cpuacct_update_stats(struct task_struct *tsk, |
9459 | enum cpuacct_stat_index idx, cputime_t val) |
9460 | { |
9461 | struct cpuacct *ca; |
9462 | int batch = CPUACCT_BATCH; |
9463 | |
9464 | if (unlikely(!cpuacct_subsys.active)) |
9465 | return; |
9466 | |
9467 | rcu_read_lock(); |
9468 | ca = task_ca(tsk); |
9469 | |
9470 | do { |
9471 | __percpu_counter_add(&ca->cpustat[idx], val, batch); |
9472 | ca = ca->parent; |
9473 | } while (ca); |
9474 | rcu_read_unlock(); |
9475 | } |
9476 | |
9477 | struct cgroup_subsys cpuacct_subsys = { |
9478 | .name = "cpuacct", |
9479 | .create = cpuacct_create, |
9480 | .destroy = cpuacct_destroy, |
9481 | .populate = cpuacct_populate, |
9482 | .subsys_id = cpuacct_subsys_id, |
9483 | }; |
9484 | #endif /* CONFIG_CGROUP_CPUACCT */ |
9485 | |
9486 |
Branches:
ben-wpan
ben-wpan-stefan
javiroman/ks7010
jz-2.6.34
jz-2.6.34-rc5
jz-2.6.34-rc6
jz-2.6.34-rc7
jz-2.6.35
jz-2.6.36
jz-2.6.37
jz-2.6.38
jz-2.6.39
jz-3.0
jz-3.1
jz-3.11
jz-3.12
jz-3.13
jz-3.15
jz-3.16
jz-3.18-dt
jz-3.2
jz-3.3
jz-3.4
jz-3.5
jz-3.6
jz-3.6-rc2-pwm
jz-3.9
jz-3.9-clk
jz-3.9-rc8
jz47xx
jz47xx-2.6.38
master
Tags:
od-2011-09-04
od-2011-09-18
v2.6.34-rc5
v2.6.34-rc6
v2.6.34-rc7
v3.9