]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * kernel/sched.c | |
3 | * | |
4 | * Kernel scheduler and related syscalls | |
5 | * | |
6 | * Copyright (C) 1991-2002 Linus Torvalds | |
7 | * | |
8 | * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and | |
9 | * make semaphores SMP safe | |
10 | * 1998-11-19 Implemented schedule_timeout() and related stuff | |
11 | * by Andrea Arcangeli | |
12 | * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar: | |
13 | * hybrid priority-list and round-robin design with | |
14 | * an array-switch method of distributing timeslices | |
15 | * and per-CPU runqueues. Cleanups and useful suggestions | |
16 | * by Davide Libenzi, preemptible kernel bits by Robert Love. | |
17 | * 2003-09-03 Interactivity tuning by Con Kolivas. | |
18 | * 2004-04-02 Scheduler domains code by Nick Piggin | |
19 | * 2007-04-15 Work begun on replacing all interactivity tuning with a | |
20 | * fair scheduling design by Con Kolivas. | |
21 | * 2007-05-05 Load balancing (smp-nice) and other improvements | |
22 | * by Peter Williams | |
23 | * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith | |
24 | * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri | |
25 | * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins, | |
26 | * Thomas Gleixner, Mike Kravetz | |
27 | */ | |
28 | ||
29 | #include <linux/mm.h> | |
30 | #include <linux/module.h> | |
31 | #include <linux/nmi.h> | |
32 | #include <linux/init.h> | |
33 | #include <linux/uaccess.h> | |
34 | #include <linux/highmem.h> | |
35 | #include <linux/smp_lock.h> | |
36 | #include <asm/mmu_context.h> | |
37 | #include <linux/interrupt.h> | |
38 | #include <linux/capability.h> | |
39 | #include <linux/completion.h> | |
40 | #include <linux/kernel_stat.h> | |
41 | #include <linux/debug_locks.h> | |
42 | #include <linux/security.h> | |
43 | #include <linux/notifier.h> | |
44 | #include <linux/profile.h> | |
45 | #include <linux/freezer.h> | |
46 | #include <linux/vmalloc.h> | |
47 | #include <linux/blkdev.h> | |
48 | #include <linux/delay.h> | |
49 | #include <linux/pid_namespace.h> | |
50 | #include <linux/smp.h> | |
51 | #include <linux/threads.h> | |
52 | #include <linux/timer.h> | |
53 | #include <linux/rcupdate.h> | |
54 | #include <linux/cpu.h> | |
55 | #include <linux/cpuset.h> | |
56 | #include <linux/percpu.h> | |
57 | #include <linux/kthread.h> | |
58 | #include <linux/seq_file.h> | |
59 | #include <linux/sysctl.h> | |
60 | #include <linux/syscalls.h> | |
61 | #include <linux/times.h> | |
62 | #include <linux/tsacct_kern.h> | |
63 | #include <linux/kprobes.h> | |
64 | #include <linux/delayacct.h> | |
65 | #include <linux/reciprocal_div.h> | |
66 | #include <linux/unistd.h> | |
67 | #include <linux/pagemap.h> | |
68 | #include <linux/hrtimer.h> | |
69 | #include <linux/tick.h> | |
70 | #include <linux/bootmem.h> | |
71 | ||
72 | #include <asm/tlb.h> | |
73 | #include <asm/irq_regs.h> | |
74 | ||
75 | /* | |
76 | * Scheduler clock - returns current time in nanosec units. | |
77 | * This is default implementation. | |
78 | * Architectures and sub-architectures can override this. | |
79 | */ | |
80 | unsigned long long __attribute__((weak)) sched_clock(void) | |
81 | { | |
82 | return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ); | |
83 | } | |
84 | ||
85 | /* | |
86 | * Convert user-nice values [ -20 ... 0 ... 19 ] | |
87 | * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ], | |
88 | * and back. | |
89 | */ | |
90 | #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20) | |
91 | #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20) | |
92 | #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio) | |
93 | ||
94 | /* | |
95 | * 'User priority' is the nice value converted to something we | |
96 | * can work with better when scaling various scheduler parameters, | |
97 | * it's a [ 0 ... 39 ] range. | |
98 | */ | |
99 | #define USER_PRIO(p) ((p)-MAX_RT_PRIO) | |
100 | #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio) | |
101 | #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO)) | |
102 | ||
103 | /* | |
104 | * Helpers for converting nanosecond timing to jiffy resolution | |
105 | */ | |
106 | #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ)) | |
107 | ||
108 | #define NICE_0_LOAD SCHED_LOAD_SCALE | |
109 | #define NICE_0_SHIFT SCHED_LOAD_SHIFT | |
110 | ||
111 | /* | |
112 | * These are the 'tuning knobs' of the scheduler: | |
113 | * | |
114 | * default timeslice is 100 msecs (used only for SCHED_RR tasks). | |
115 | * Timeslices get refilled after they expire. | |
116 | */ | |
117 | #define DEF_TIMESLICE (100 * HZ / 1000) | |
118 | ||
119 | /* | |
120 | * single value that denotes runtime == period, ie unlimited time. | |
121 | */ | |
122 | #define RUNTIME_INF ((u64)~0ULL) | |
123 | ||
124 | #ifdef CONFIG_SMP | |
125 | /* | |
126 | * Divide a load by a sched group cpu_power : (load / sg->__cpu_power) | |
127 | * Since cpu_power is a 'constant', we can use a reciprocal divide. | |
128 | */ | |
129 | static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load) | |
130 | { | |
131 | return reciprocal_divide(load, sg->reciprocal_cpu_power); | |
132 | } | |
133 | ||
134 | /* | |
135 | * Each time a sched group cpu_power is changed, | |
136 | * we must compute its reciprocal value | |
137 | */ | |
138 | static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val) | |
139 | { | |
140 | sg->__cpu_power += val; | |
141 | sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power); | |
142 | } | |
143 | #endif | |
144 | ||
145 | static inline int rt_policy(int policy) | |
146 | { | |
147 | if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR)) | |
148 | return 1; | |
149 | return 0; | |
150 | } | |
151 | ||
152 | static inline int task_has_rt_policy(struct task_struct *p) | |
153 | { | |
154 | return rt_policy(p->policy); | |
155 | } | |
156 | ||
157 | /* | |
158 | * This is the priority-queue data structure of the RT scheduling class: | |
159 | */ | |
160 | struct rt_prio_array { | |
161 | DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */ | |
162 | struct list_head queue[MAX_RT_PRIO]; | |
163 | }; | |
164 | ||
165 | struct rt_bandwidth { | |
166 | /* nests inside the rq lock: */ | |
167 | spinlock_t rt_runtime_lock; | |
168 | ktime_t rt_period; | |
169 | u64 rt_runtime; | |
170 | struct hrtimer rt_period_timer; | |
171 | }; | |
172 | ||
173 | static struct rt_bandwidth def_rt_bandwidth; | |
174 | ||
175 | static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun); | |
176 | ||
177 | static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer) | |
178 | { | |
179 | struct rt_bandwidth *rt_b = | |
180 | container_of(timer, struct rt_bandwidth, rt_period_timer); | |
181 | ktime_t now; | |
182 | int overrun; | |
183 | int idle = 0; | |
184 | ||
185 | for (;;) { | |
186 | now = hrtimer_cb_get_time(timer); | |
187 | overrun = hrtimer_forward(timer, now, rt_b->rt_period); | |
188 | ||
189 | if (!overrun) | |
190 | break; | |
191 | ||
192 | idle = do_sched_rt_period_timer(rt_b, overrun); | |
193 | } | |
194 | ||
195 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
196 | } | |
197 | ||
198 | static | |
199 | void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime) | |
200 | { | |
201 | rt_b->rt_period = ns_to_ktime(period); | |
202 | rt_b->rt_runtime = runtime; | |
203 | ||
204 | spin_lock_init(&rt_b->rt_runtime_lock); | |
205 | ||
206 | hrtimer_init(&rt_b->rt_period_timer, | |
207 | CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
208 | rt_b->rt_period_timer.function = sched_rt_period_timer; | |
209 | rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ; | |
210 | } | |
211 | ||
212 | static void start_rt_bandwidth(struct rt_bandwidth *rt_b) | |
213 | { | |
214 | ktime_t now; | |
215 | ||
216 | if (rt_b->rt_runtime == RUNTIME_INF) | |
217 | return; | |
218 | ||
219 | if (hrtimer_active(&rt_b->rt_period_timer)) | |
220 | return; | |
221 | ||
222 | spin_lock(&rt_b->rt_runtime_lock); | |
223 | for (;;) { | |
224 | if (hrtimer_active(&rt_b->rt_period_timer)) | |
225 | break; | |
226 | ||
227 | now = hrtimer_cb_get_time(&rt_b->rt_period_timer); | |
228 | hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period); | |
229 | hrtimer_start(&rt_b->rt_period_timer, | |
230 | rt_b->rt_period_timer.expires, | |
231 | HRTIMER_MODE_ABS); | |
232 | } | |
233 | spin_unlock(&rt_b->rt_runtime_lock); | |
234 | } | |
235 | ||
236 | #ifdef CONFIG_RT_GROUP_SCHED | |
237 | static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b) | |
238 | { | |
239 | hrtimer_cancel(&rt_b->rt_period_timer); | |
240 | } | |
241 | #endif | |
242 | ||
243 | #ifdef CONFIG_GROUP_SCHED | |
244 | ||
245 | #include <linux/cgroup.h> | |
246 | ||
247 | struct cfs_rq; | |
248 | ||
249 | static LIST_HEAD(task_groups); | |
250 | ||
251 | /* task group related information */ | |
252 | struct task_group { | |
253 | #ifdef CONFIG_CGROUP_SCHED | |
254 | struct cgroup_subsys_state css; | |
255 | #endif | |
256 | ||
257 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
258 | /* schedulable entities of this group on each cpu */ | |
259 | struct sched_entity **se; | |
260 | /* runqueue "owned" by this group on each cpu */ | |
261 | struct cfs_rq **cfs_rq; | |
262 | unsigned long shares; | |
263 | #endif | |
264 | ||
265 | #ifdef CONFIG_RT_GROUP_SCHED | |
266 | struct sched_rt_entity **rt_se; | |
267 | struct rt_rq **rt_rq; | |
268 | ||
269 | struct rt_bandwidth rt_bandwidth; | |
270 | #endif | |
271 | ||
272 | struct rcu_head rcu; | |
273 | struct list_head list; | |
274 | ||
275 | struct task_group *parent; | |
276 | struct list_head siblings; | |
277 | struct list_head children; | |
278 | }; | |
279 | ||
280 | #ifdef CONFIG_USER_SCHED | |
281 | ||
282 | /* | |
283 | * Root task group. | |
284 | * Every UID task group (including init_task_group aka UID-0) will | |
285 | * be a child to this group. | |
286 | */ | |
287 | struct task_group root_task_group; | |
288 | ||
289 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
290 | /* Default task group's sched entity on each cpu */ | |
291 | static DEFINE_PER_CPU(struct sched_entity, init_sched_entity); | |
292 | /* Default task group's cfs_rq on each cpu */ | |
293 | static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp; | |
294 | #endif | |
295 | ||
296 | #ifdef CONFIG_RT_GROUP_SCHED | |
297 | static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity); | |
298 | static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp; | |
299 | #endif | |
300 | #else | |
301 | #define root_task_group init_task_group | |
302 | #endif | |
303 | ||
304 | /* task_group_lock serializes add/remove of task groups and also changes to | |
305 | * a task group's cpu shares. | |
306 | */ | |
307 | static DEFINE_SPINLOCK(task_group_lock); | |
308 | ||
309 | /* doms_cur_mutex serializes access to doms_cur[] array */ | |
310 | static DEFINE_MUTEX(doms_cur_mutex); | |
311 | ||
312 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
313 | #ifdef CONFIG_USER_SCHED | |
314 | # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD) | |
315 | #else | |
316 | # define INIT_TASK_GROUP_LOAD NICE_0_LOAD | |
317 | #endif | |
318 | ||
319 | #define MIN_SHARES 2 | |
320 | ||
321 | static int init_task_group_load = INIT_TASK_GROUP_LOAD; | |
322 | #endif | |
323 | ||
324 | /* Default task group. | |
325 | * Every task in system belong to this group at bootup. | |
326 | */ | |
327 | struct task_group init_task_group; | |
328 | ||
329 | /* return group to which a task belongs */ | |
330 | static inline struct task_group *task_group(struct task_struct *p) | |
331 | { | |
332 | struct task_group *tg; | |
333 | ||
334 | #ifdef CONFIG_USER_SCHED | |
335 | tg = p->user->tg; | |
336 | #elif defined(CONFIG_CGROUP_SCHED) | |
337 | tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id), | |
338 | struct task_group, css); | |
339 | #else | |
340 | tg = &init_task_group; | |
341 | #endif | |
342 | return tg; | |
343 | } | |
344 | ||
345 | /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */ | |
346 | static inline void set_task_rq(struct task_struct *p, unsigned int cpu) | |
347 | { | |
348 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
349 | p->se.cfs_rq = task_group(p)->cfs_rq[cpu]; | |
350 | p->se.parent = task_group(p)->se[cpu]; | |
351 | #endif | |
352 | ||
353 | #ifdef CONFIG_RT_GROUP_SCHED | |
354 | p->rt.rt_rq = task_group(p)->rt_rq[cpu]; | |
355 | p->rt.parent = task_group(p)->rt_se[cpu]; | |
356 | #endif | |
357 | } | |
358 | ||
359 | static inline void lock_doms_cur(void) | |
360 | { | |
361 | mutex_lock(&doms_cur_mutex); | |
362 | } | |
363 | ||
364 | static inline void unlock_doms_cur(void) | |
365 | { | |
366 | mutex_unlock(&doms_cur_mutex); | |
367 | } | |
368 | ||
369 | #else | |
370 | ||
371 | static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { } | |
372 | static inline void lock_doms_cur(void) { } | |
373 | static inline void unlock_doms_cur(void) { } | |
374 | ||
375 | #endif /* CONFIG_GROUP_SCHED */ | |
376 | ||
377 | /* CFS-related fields in a runqueue */ | |
378 | struct cfs_rq { | |
379 | struct load_weight load; | |
380 | unsigned long nr_running; | |
381 | ||
382 | u64 exec_clock; | |
383 | u64 min_vruntime; | |
384 | ||
385 | struct rb_root tasks_timeline; | |
386 | struct rb_node *rb_leftmost; | |
387 | ||
388 | struct list_head tasks; | |
389 | struct list_head *balance_iterator; | |
390 | ||
391 | /* | |
392 | * 'curr' points to currently running entity on this cfs_rq. | |
393 | * It is set to NULL otherwise (i.e when none are currently running). | |
394 | */ | |
395 | struct sched_entity *curr, *next; | |
396 | ||
397 | unsigned long nr_spread_over; | |
398 | ||
399 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
400 | struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */ | |
401 | ||
402 | /* | |
403 | * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in | |
404 | * a hierarchy). Non-leaf lrqs hold other higher schedulable entities | |
405 | * (like users, containers etc.) | |
406 | * | |
407 | * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This | |
408 | * list is used during load balance. | |
409 | */ | |
410 | struct list_head leaf_cfs_rq_list; | |
411 | struct task_group *tg; /* group that "owns" this runqueue */ | |
412 | ||
413 | #ifdef CONFIG_SMP | |
414 | unsigned long task_weight; | |
415 | unsigned long shares; | |
416 | /* | |
417 | * We need space to build a sched_domain wide view of the full task | |
418 | * group tree, in order to avoid depending on dynamic memory allocation | |
419 | * during the load balancing we place this in the per cpu task group | |
420 | * hierarchy. This limits the load balancing to one instance per cpu, | |
421 | * but more should not be needed anyway. | |
422 | */ | |
423 | struct aggregate_struct { | |
424 | /* | |
425 | * load = weight(cpus) * f(tg) | |
426 | * | |
427 | * Where f(tg) is the recursive weight fraction assigned to | |
428 | * this group. | |
429 | */ | |
430 | unsigned long load; | |
431 | ||
432 | /* | |
433 | * part of the group weight distributed to this span. | |
434 | */ | |
435 | unsigned long shares; | |
436 | ||
437 | /* | |
438 | * The sum of all runqueue weights within this span. | |
439 | */ | |
440 | unsigned long rq_weight; | |
441 | ||
442 | /* | |
443 | * Weight contributed by tasks; this is the part we can | |
444 | * influence by moving tasks around. | |
445 | */ | |
446 | unsigned long task_weight; | |
447 | } aggregate; | |
448 | #endif | |
449 | #endif | |
450 | }; | |
451 | ||
452 | /* Real-Time classes' related field in a runqueue: */ | |
453 | struct rt_rq { | |
454 | struct rt_prio_array active; | |
455 | unsigned long rt_nr_running; | |
456 | #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED | |
457 | int highest_prio; /* highest queued rt task prio */ | |
458 | #endif | |
459 | #ifdef CONFIG_SMP | |
460 | unsigned long rt_nr_migratory; | |
461 | int overloaded; | |
462 | #endif | |
463 | int rt_throttled; | |
464 | u64 rt_time; | |
465 | u64 rt_runtime; | |
466 | /* Nests inside the rq lock: */ | |
467 | spinlock_t rt_runtime_lock; | |
468 | ||
469 | #ifdef CONFIG_RT_GROUP_SCHED | |
470 | unsigned long rt_nr_boosted; | |
471 | ||
472 | struct rq *rq; | |
473 | struct list_head leaf_rt_rq_list; | |
474 | struct task_group *tg; | |
475 | struct sched_rt_entity *rt_se; | |
476 | #endif | |
477 | }; | |
478 | ||
479 | #ifdef CONFIG_SMP | |
480 | ||
481 | /* | |
482 | * We add the notion of a root-domain which will be used to define per-domain | |
483 | * variables. Each exclusive cpuset essentially defines an island domain by | |
484 | * fully partitioning the member cpus from any other cpuset. Whenever a new | |
485 | * exclusive cpuset is created, we also create and attach a new root-domain | |
486 | * object. | |
487 | * | |
488 | */ | |
489 | struct root_domain { | |
490 | atomic_t refcount; | |
491 | cpumask_t span; | |
492 | cpumask_t online; | |
493 | ||
494 | /* | |
495 | * The "RT overload" flag: it gets set if a CPU has more than | |
496 | * one runnable RT task. | |
497 | */ | |
498 | cpumask_t rto_mask; | |
499 | atomic_t rto_count; | |
500 | }; | |
501 | ||
502 | /* | |
503 | * By default the system creates a single root-domain with all cpus as | |
504 | * members (mimicking the global state we have today). | |
505 | */ | |
506 | static struct root_domain def_root_domain; | |
507 | ||
508 | #endif | |
509 | ||
510 | /* | |
511 | * This is the main, per-CPU runqueue data structure. | |
512 | * | |
513 | * Locking rule: those places that want to lock multiple runqueues | |
514 | * (such as the load balancing or the thread migration code), lock | |
515 | * acquire operations must be ordered by ascending &runqueue. | |
516 | */ | |
517 | struct rq { | |
518 | /* runqueue lock: */ | |
519 | spinlock_t lock; | |
520 | ||
521 | /* | |
522 | * nr_running and cpu_load should be in the same cacheline because | |
523 | * remote CPUs use both these fields when doing load calculation. | |
524 | */ | |
525 | unsigned long nr_running; | |
526 | #define CPU_LOAD_IDX_MAX 5 | |
527 | unsigned long cpu_load[CPU_LOAD_IDX_MAX]; | |
528 | unsigned char idle_at_tick; | |
529 | #ifdef CONFIG_NO_HZ | |
530 | unsigned long last_tick_seen; | |
531 | unsigned char in_nohz_recently; | |
532 | #endif | |
533 | /* capture load from *all* tasks on this cpu: */ | |
534 | struct load_weight load; | |
535 | unsigned long nr_load_updates; | |
536 | u64 nr_switches; | |
537 | ||
538 | struct cfs_rq cfs; | |
539 | struct rt_rq rt; | |
540 | ||
541 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
542 | /* list of leaf cfs_rq on this cpu: */ | |
543 | struct list_head leaf_cfs_rq_list; | |
544 | #endif | |
545 | #ifdef CONFIG_RT_GROUP_SCHED | |
546 | struct list_head leaf_rt_rq_list; | |
547 | #endif | |
548 | ||
549 | /* | |
550 | * This is part of a global counter where only the total sum | |
551 | * over all CPUs matters. A task can increase this counter on | |
552 | * one CPU and if it got migrated afterwards it may decrease | |
553 | * it on another CPU. Always updated under the runqueue lock: | |
554 | */ | |
555 | unsigned long nr_uninterruptible; | |
556 | ||
557 | struct task_struct *curr, *idle; | |
558 | unsigned long next_balance; | |
559 | struct mm_struct *prev_mm; | |
560 | ||
561 | u64 clock, prev_clock_raw; | |
562 | s64 clock_max_delta; | |
563 | ||
564 | unsigned int clock_warps, clock_overflows, clock_underflows; | |
565 | u64 idle_clock; | |
566 | unsigned int clock_deep_idle_events; | |
567 | u64 tick_timestamp; | |
568 | ||
569 | atomic_t nr_iowait; | |
570 | ||
571 | #ifdef CONFIG_SMP | |
572 | struct root_domain *rd; | |
573 | struct sched_domain *sd; | |
574 | ||
575 | /* For active balancing */ | |
576 | int active_balance; | |
577 | int push_cpu; | |
578 | /* cpu of this runqueue: */ | |
579 | int cpu; | |
580 | ||
581 | struct task_struct *migration_thread; | |
582 | struct list_head migration_queue; | |
583 | #endif | |
584 | ||
585 | #ifdef CONFIG_SCHED_HRTICK | |
586 | unsigned long hrtick_flags; | |
587 | ktime_t hrtick_expire; | |
588 | struct hrtimer hrtick_timer; | |
589 | #endif | |
590 | ||
591 | #ifdef CONFIG_SCHEDSTATS | |
592 | /* latency stats */ | |
593 | struct sched_info rq_sched_info; | |
594 | ||
595 | /* sys_sched_yield() stats */ | |
596 | unsigned int yld_exp_empty; | |
597 | unsigned int yld_act_empty; | |
598 | unsigned int yld_both_empty; | |
599 | unsigned int yld_count; | |
600 | ||
601 | /* schedule() stats */ | |
602 | unsigned int sched_switch; | |
603 | unsigned int sched_count; | |
604 | unsigned int sched_goidle; | |
605 | ||
606 | /* try_to_wake_up() stats */ | |
607 | unsigned int ttwu_count; | |
608 | unsigned int ttwu_local; | |
609 | ||
610 | /* BKL stats */ | |
611 | unsigned int bkl_count; | |
612 | #endif | |
613 | struct lock_class_key rq_lock_key; | |
614 | }; | |
615 | ||
616 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); | |
617 | ||
618 | static inline void check_preempt_curr(struct rq *rq, struct task_struct *p) | |
619 | { | |
620 | rq->curr->sched_class->check_preempt_curr(rq, p); | |
621 | } | |
622 | ||
623 | static inline int cpu_of(struct rq *rq) | |
624 | { | |
625 | #ifdef CONFIG_SMP | |
626 | return rq->cpu; | |
627 | #else | |
628 | return 0; | |
629 | #endif | |
630 | } | |
631 | ||
632 | #ifdef CONFIG_NO_HZ | |
633 | static inline bool nohz_on(int cpu) | |
634 | { | |
635 | return tick_get_tick_sched(cpu)->nohz_mode != NOHZ_MODE_INACTIVE; | |
636 | } | |
637 | ||
638 | static inline u64 max_skipped_ticks(struct rq *rq) | |
639 | { | |
640 | return nohz_on(cpu_of(rq)) ? jiffies - rq->last_tick_seen + 2 : 1; | |
641 | } | |
642 | ||
643 | static inline void update_last_tick_seen(struct rq *rq) | |
644 | { | |
645 | rq->last_tick_seen = jiffies; | |
646 | } | |
647 | #else | |
648 | static inline u64 max_skipped_ticks(struct rq *rq) | |
649 | { | |
650 | return 1; | |
651 | } | |
652 | ||
653 | static inline void update_last_tick_seen(struct rq *rq) | |
654 | { | |
655 | } | |
656 | #endif | |
657 | ||
658 | /* | |
659 | * Update the per-runqueue clock, as finegrained as the platform can give | |
660 | * us, but without assuming monotonicity, etc.: | |
661 | */ | |
662 | static void __update_rq_clock(struct rq *rq) | |
663 | { | |
664 | u64 prev_raw = rq->prev_clock_raw; | |
665 | u64 now = sched_clock(); | |
666 | s64 delta = now - prev_raw; | |
667 | u64 clock = rq->clock; | |
668 | ||
669 | #ifdef CONFIG_SCHED_DEBUG | |
670 | WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); | |
671 | #endif | |
672 | /* | |
673 | * Protect against sched_clock() occasionally going backwards: | |
674 | */ | |
675 | if (unlikely(delta < 0)) { | |
676 | clock++; | |
677 | rq->clock_warps++; | |
678 | } else { | |
679 | /* | |
680 | * Catch too large forward jumps too: | |
681 | */ | |
682 | u64 max_jump = max_skipped_ticks(rq) * TICK_NSEC; | |
683 | u64 max_time = rq->tick_timestamp + max_jump; | |
684 | ||
685 | if (unlikely(clock + delta > max_time)) { | |
686 | if (clock < max_time) | |
687 | clock = max_time; | |
688 | else | |
689 | clock++; | |
690 | rq->clock_overflows++; | |
691 | } else { | |
692 | if (unlikely(delta > rq->clock_max_delta)) | |
693 | rq->clock_max_delta = delta; | |
694 | clock += delta; | |
695 | } | |
696 | } | |
697 | ||
698 | rq->prev_clock_raw = now; | |
699 | rq->clock = clock; | |
700 | } | |
701 | ||
702 | static void update_rq_clock(struct rq *rq) | |
703 | { | |
704 | if (likely(smp_processor_id() == cpu_of(rq))) | |
705 | __update_rq_clock(rq); | |
706 | } | |
707 | ||
708 | /* | |
709 | * The domain tree (rq->sd) is protected by RCU's quiescent state transition. | |
710 | * See detach_destroy_domains: synchronize_sched for details. | |
711 | * | |
712 | * The domain tree of any CPU may only be accessed from within | |
713 | * preempt-disabled sections. | |
714 | */ | |
715 | #define for_each_domain(cpu, __sd) \ | |
716 | for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent) | |
717 | ||
718 | #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) | |
719 | #define this_rq() (&__get_cpu_var(runqueues)) | |
720 | #define task_rq(p) cpu_rq(task_cpu(p)) | |
721 | #define cpu_curr(cpu) (cpu_rq(cpu)->curr) | |
722 | ||
723 | /* | |
724 | * Tunables that become constants when CONFIG_SCHED_DEBUG is off: | |
725 | */ | |
726 | #ifdef CONFIG_SCHED_DEBUG | |
727 | # define const_debug __read_mostly | |
728 | #else | |
729 | # define const_debug static const | |
730 | #endif | |
731 | ||
732 | /* | |
733 | * Debugging: various feature bits | |
734 | */ | |
735 | enum { | |
736 | SCHED_FEAT_NEW_FAIR_SLEEPERS = 1, | |
737 | SCHED_FEAT_WAKEUP_PREEMPT = 2, | |
738 | SCHED_FEAT_START_DEBIT = 4, | |
739 | SCHED_FEAT_AFFINE_WAKEUPS = 8, | |
740 | SCHED_FEAT_CACHE_HOT_BUDDY = 16, | |
741 | SCHED_FEAT_SYNC_WAKEUPS = 32, | |
742 | SCHED_FEAT_HRTICK = 64, | |
743 | SCHED_FEAT_DOUBLE_TICK = 128, | |
744 | SCHED_FEAT_NORMALIZED_SLEEPER = 256, | |
745 | }; | |
746 | ||
747 | const_debug unsigned int sysctl_sched_features = | |
748 | SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 | | |
749 | SCHED_FEAT_WAKEUP_PREEMPT * 1 | | |
750 | SCHED_FEAT_START_DEBIT * 1 | | |
751 | SCHED_FEAT_AFFINE_WAKEUPS * 1 | | |
752 | SCHED_FEAT_CACHE_HOT_BUDDY * 1 | | |
753 | SCHED_FEAT_SYNC_WAKEUPS * 1 | | |
754 | SCHED_FEAT_HRTICK * 1 | | |
755 | SCHED_FEAT_DOUBLE_TICK * 0 | | |
756 | SCHED_FEAT_NORMALIZED_SLEEPER * 1; | |
757 | ||
758 | #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x) | |
759 | ||
760 | /* | |
761 | * Number of tasks to iterate in a single balance run. | |
762 | * Limited because this is done with IRQs disabled. | |
763 | */ | |
764 | const_debug unsigned int sysctl_sched_nr_migrate = 32; | |
765 | ||
766 | /* | |
767 | * period over which we measure -rt task cpu usage in us. | |
768 | * default: 1s | |
769 | */ | |
770 | unsigned int sysctl_sched_rt_period = 1000000; | |
771 | ||
772 | static __read_mostly int scheduler_running; | |
773 | ||
774 | /* | |
775 | * part of the period that we allow rt tasks to run in us. | |
776 | * default: 0.95s | |
777 | */ | |
778 | int sysctl_sched_rt_runtime = 950000; | |
779 | ||
780 | static inline u64 global_rt_period(void) | |
781 | { | |
782 | return (u64)sysctl_sched_rt_period * NSEC_PER_USEC; | |
783 | } | |
784 | ||
785 | static inline u64 global_rt_runtime(void) | |
786 | { | |
787 | if (sysctl_sched_rt_period < 0) | |
788 | return RUNTIME_INF; | |
789 | ||
790 | return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC; | |
791 | } | |
792 | ||
793 | static const unsigned long long time_sync_thresh = 100000; | |
794 | ||
795 | static DEFINE_PER_CPU(unsigned long long, time_offset); | |
796 | static DEFINE_PER_CPU(unsigned long long, prev_cpu_time); | |
797 | ||
798 | /* | |
799 | * Global lock which we take every now and then to synchronize | |
800 | * the CPUs time. This method is not warp-safe, but it's good | |
801 | * enough to synchronize slowly diverging time sources and thus | |
802 | * it's good enough for tracing: | |
803 | */ | |
804 | static DEFINE_SPINLOCK(time_sync_lock); | |
805 | static unsigned long long prev_global_time; | |
806 | ||
807 | static unsigned long long __sync_cpu_clock(cycles_t time, int cpu) | |
808 | { | |
809 | unsigned long flags; | |
810 | ||
811 | spin_lock_irqsave(&time_sync_lock, flags); | |
812 | ||
813 | if (time < prev_global_time) { | |
814 | per_cpu(time_offset, cpu) += prev_global_time - time; | |
815 | time = prev_global_time; | |
816 | } else { | |
817 | prev_global_time = time; | |
818 | } | |
819 | ||
820 | spin_unlock_irqrestore(&time_sync_lock, flags); | |
821 | ||
822 | return time; | |
823 | } | |
824 | ||
825 | static unsigned long long __cpu_clock(int cpu) | |
826 | { | |
827 | unsigned long long now; | |
828 | unsigned long flags; | |
829 | struct rq *rq; | |
830 | ||
831 | /* | |
832 | * Only call sched_clock() if the scheduler has already been | |
833 | * initialized (some code might call cpu_clock() very early): | |
834 | */ | |
835 | if (unlikely(!scheduler_running)) | |
836 | return 0; | |
837 | ||
838 | local_irq_save(flags); | |
839 | rq = cpu_rq(cpu); | |
840 | update_rq_clock(rq); | |
841 | now = rq->clock; | |
842 | local_irq_restore(flags); | |
843 | ||
844 | return now; | |
845 | } | |
846 | ||
847 | /* | |
848 | * For kernel-internal use: high-speed (but slightly incorrect) per-cpu | |
849 | * clock constructed from sched_clock(): | |
850 | */ | |
851 | unsigned long long cpu_clock(int cpu) | |
852 | { | |
853 | unsigned long long prev_cpu_time, time, delta_time; | |
854 | ||
855 | prev_cpu_time = per_cpu(prev_cpu_time, cpu); | |
856 | time = __cpu_clock(cpu) + per_cpu(time_offset, cpu); | |
857 | delta_time = time-prev_cpu_time; | |
858 | ||
859 | if (unlikely(delta_time > time_sync_thresh)) | |
860 | time = __sync_cpu_clock(time, cpu); | |
861 | ||
862 | return time; | |
863 | } | |
864 | EXPORT_SYMBOL_GPL(cpu_clock); | |
865 | ||
866 | #ifndef prepare_arch_switch | |
867 | # define prepare_arch_switch(next) do { } while (0) | |
868 | #endif | |
869 | #ifndef finish_arch_switch | |
870 | # define finish_arch_switch(prev) do { } while (0) | |
871 | #endif | |
872 | ||
873 | static inline int task_current(struct rq *rq, struct task_struct *p) | |
874 | { | |
875 | return rq->curr == p; | |
876 | } | |
877 | ||
878 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW | |
879 | static inline int task_running(struct rq *rq, struct task_struct *p) | |
880 | { | |
881 | return task_current(rq, p); | |
882 | } | |
883 | ||
884 | static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) | |
885 | { | |
886 | } | |
887 | ||
888 | static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) | |
889 | { | |
890 | #ifdef CONFIG_DEBUG_SPINLOCK | |
891 | /* this is a valid case when another task releases the spinlock */ | |
892 | rq->lock.owner = current; | |
893 | #endif | |
894 | /* | |
895 | * If we are tracking spinlock dependencies then we have to | |
896 | * fix up the runqueue lock - which gets 'carried over' from | |
897 | * prev into current: | |
898 | */ | |
899 | spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_); | |
900 | ||
901 | spin_unlock_irq(&rq->lock); | |
902 | } | |
903 | ||
904 | #else /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
905 | static inline int task_running(struct rq *rq, struct task_struct *p) | |
906 | { | |
907 | #ifdef CONFIG_SMP | |
908 | return p->oncpu; | |
909 | #else | |
910 | return task_current(rq, p); | |
911 | #endif | |
912 | } | |
913 | ||
914 | static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) | |
915 | { | |
916 | #ifdef CONFIG_SMP | |
917 | /* | |
918 | * We can optimise this out completely for !SMP, because the | |
919 | * SMP rebalancing from interrupt is the only thing that cares | |
920 | * here. | |
921 | */ | |
922 | next->oncpu = 1; | |
923 | #endif | |
924 | #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
925 | spin_unlock_irq(&rq->lock); | |
926 | #else | |
927 | spin_unlock(&rq->lock); | |
928 | #endif | |
929 | } | |
930 | ||
931 | static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) | |
932 | { | |
933 | #ifdef CONFIG_SMP | |
934 | /* | |
935 | * After ->oncpu is cleared, the task can be moved to a different CPU. | |
936 | * We must ensure this doesn't happen until the switch is completely | |
937 | * finished. | |
938 | */ | |
939 | smp_wmb(); | |
940 | prev->oncpu = 0; | |
941 | #endif | |
942 | #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
943 | local_irq_enable(); | |
944 | #endif | |
945 | } | |
946 | #endif /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
947 | ||
948 | /* | |
949 | * __task_rq_lock - lock the runqueue a given task resides on. | |
950 | * Must be called interrupts disabled. | |
951 | */ | |
952 | static inline struct rq *__task_rq_lock(struct task_struct *p) | |
953 | __acquires(rq->lock) | |
954 | { | |
955 | for (;;) { | |
956 | struct rq *rq = task_rq(p); | |
957 | spin_lock(&rq->lock); | |
958 | if (likely(rq == task_rq(p))) | |
959 | return rq; | |
960 | spin_unlock(&rq->lock); | |
961 | } | |
962 | } | |
963 | ||
964 | /* | |
965 | * task_rq_lock - lock the runqueue a given task resides on and disable | |
966 | * interrupts. Note the ordering: we can safely lookup the task_rq without | |
967 | * explicitly disabling preemption. | |
968 | */ | |
969 | static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) | |
970 | __acquires(rq->lock) | |
971 | { | |
972 | struct rq *rq; | |
973 | ||
974 | for (;;) { | |
975 | local_irq_save(*flags); | |
976 | rq = task_rq(p); | |
977 | spin_lock(&rq->lock); | |
978 | if (likely(rq == task_rq(p))) | |
979 | return rq; | |
980 | spin_unlock_irqrestore(&rq->lock, *flags); | |
981 | } | |
982 | } | |
983 | ||
984 | static void __task_rq_unlock(struct rq *rq) | |
985 | __releases(rq->lock) | |
986 | { | |
987 | spin_unlock(&rq->lock); | |
988 | } | |
989 | ||
990 | static inline void task_rq_unlock(struct rq *rq, unsigned long *flags) | |
991 | __releases(rq->lock) | |
992 | { | |
993 | spin_unlock_irqrestore(&rq->lock, *flags); | |
994 | } | |
995 | ||
996 | /* | |
997 | * this_rq_lock - lock this runqueue and disable interrupts. | |
998 | */ | |
999 | static struct rq *this_rq_lock(void) | |
1000 | __acquires(rq->lock) | |
1001 | { | |
1002 | struct rq *rq; | |
1003 | ||
1004 | local_irq_disable(); | |
1005 | rq = this_rq(); | |
1006 | spin_lock(&rq->lock); | |
1007 | ||
1008 | return rq; | |
1009 | } | |
1010 | ||
1011 | /* | |
1012 | * We are going deep-idle (irqs are disabled): | |
1013 | */ | |
1014 | void sched_clock_idle_sleep_event(void) | |
1015 | { | |
1016 | struct rq *rq = cpu_rq(smp_processor_id()); | |
1017 | ||
1018 | spin_lock(&rq->lock); | |
1019 | __update_rq_clock(rq); | |
1020 | spin_unlock(&rq->lock); | |
1021 | rq->clock_deep_idle_events++; | |
1022 | } | |
1023 | EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event); | |
1024 | ||
1025 | /* | |
1026 | * We just idled delta nanoseconds (called with irqs disabled): | |
1027 | */ | |
1028 | void sched_clock_idle_wakeup_event(u64 delta_ns) | |
1029 | { | |
1030 | struct rq *rq = cpu_rq(smp_processor_id()); | |
1031 | u64 now = sched_clock(); | |
1032 | ||
1033 | rq->idle_clock += delta_ns; | |
1034 | /* | |
1035 | * Override the previous timestamp and ignore all | |
1036 | * sched_clock() deltas that occured while we idled, | |
1037 | * and use the PM-provided delta_ns to advance the | |
1038 | * rq clock: | |
1039 | */ | |
1040 | spin_lock(&rq->lock); | |
1041 | rq->prev_clock_raw = now; | |
1042 | rq->clock += delta_ns; | |
1043 | spin_unlock(&rq->lock); | |
1044 | touch_softlockup_watchdog(); | |
1045 | } | |
1046 | EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event); | |
1047 | ||
1048 | static void __resched_task(struct task_struct *p, int tif_bit); | |
1049 | ||
1050 | static inline void resched_task(struct task_struct *p) | |
1051 | { | |
1052 | __resched_task(p, TIF_NEED_RESCHED); | |
1053 | } | |
1054 | ||
1055 | #ifdef CONFIG_SCHED_HRTICK | |
1056 | /* | |
1057 | * Use HR-timers to deliver accurate preemption points. | |
1058 | * | |
1059 | * Its all a bit involved since we cannot program an hrt while holding the | |
1060 | * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a | |
1061 | * reschedule event. | |
1062 | * | |
1063 | * When we get rescheduled we reprogram the hrtick_timer outside of the | |
1064 | * rq->lock. | |
1065 | */ | |
1066 | static inline void resched_hrt(struct task_struct *p) | |
1067 | { | |
1068 | __resched_task(p, TIF_HRTICK_RESCHED); | |
1069 | } | |
1070 | ||
1071 | static inline void resched_rq(struct rq *rq) | |
1072 | { | |
1073 | unsigned long flags; | |
1074 | ||
1075 | spin_lock_irqsave(&rq->lock, flags); | |
1076 | resched_task(rq->curr); | |
1077 | spin_unlock_irqrestore(&rq->lock, flags); | |
1078 | } | |
1079 | ||
1080 | enum { | |
1081 | HRTICK_SET, /* re-programm hrtick_timer */ | |
1082 | HRTICK_RESET, /* not a new slice */ | |
1083 | }; | |
1084 | ||
1085 | /* | |
1086 | * Use hrtick when: | |
1087 | * - enabled by features | |
1088 | * - hrtimer is actually high res | |
1089 | */ | |
1090 | static inline int hrtick_enabled(struct rq *rq) | |
1091 | { | |
1092 | if (!sched_feat(HRTICK)) | |
1093 | return 0; | |
1094 | return hrtimer_is_hres_active(&rq->hrtick_timer); | |
1095 | } | |
1096 | ||
1097 | /* | |
1098 | * Called to set the hrtick timer state. | |
1099 | * | |
1100 | * called with rq->lock held and irqs disabled | |
1101 | */ | |
1102 | static void hrtick_start(struct rq *rq, u64 delay, int reset) | |
1103 | { | |
1104 | assert_spin_locked(&rq->lock); | |
1105 | ||
1106 | /* | |
1107 | * preempt at: now + delay | |
1108 | */ | |
1109 | rq->hrtick_expire = | |
1110 | ktime_add_ns(rq->hrtick_timer.base->get_time(), delay); | |
1111 | /* | |
1112 | * indicate we need to program the timer | |
1113 | */ | |
1114 | __set_bit(HRTICK_SET, &rq->hrtick_flags); | |
1115 | if (reset) | |
1116 | __set_bit(HRTICK_RESET, &rq->hrtick_flags); | |
1117 | ||
1118 | /* | |
1119 | * New slices are called from the schedule path and don't need a | |
1120 | * forced reschedule. | |
1121 | */ | |
1122 | if (reset) | |
1123 | resched_hrt(rq->curr); | |
1124 | } | |
1125 | ||
1126 | static void hrtick_clear(struct rq *rq) | |
1127 | { | |
1128 | if (hrtimer_active(&rq->hrtick_timer)) | |
1129 | hrtimer_cancel(&rq->hrtick_timer); | |
1130 | } | |
1131 | ||
1132 | /* | |
1133 | * Update the timer from the possible pending state. | |
1134 | */ | |
1135 | static void hrtick_set(struct rq *rq) | |
1136 | { | |
1137 | ktime_t time; | |
1138 | int set, reset; | |
1139 | unsigned long flags; | |
1140 | ||
1141 | WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); | |
1142 | ||
1143 | spin_lock_irqsave(&rq->lock, flags); | |
1144 | set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags); | |
1145 | reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags); | |
1146 | time = rq->hrtick_expire; | |
1147 | clear_thread_flag(TIF_HRTICK_RESCHED); | |
1148 | spin_unlock_irqrestore(&rq->lock, flags); | |
1149 | ||
1150 | if (set) { | |
1151 | hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS); | |
1152 | if (reset && !hrtimer_active(&rq->hrtick_timer)) | |
1153 | resched_rq(rq); | |
1154 | } else | |
1155 | hrtick_clear(rq); | |
1156 | } | |
1157 | ||
1158 | /* | |
1159 | * High-resolution timer tick. | |
1160 | * Runs from hardirq context with interrupts disabled. | |
1161 | */ | |
1162 | static enum hrtimer_restart hrtick(struct hrtimer *timer) | |
1163 | { | |
1164 | struct rq *rq = container_of(timer, struct rq, hrtick_timer); | |
1165 | ||
1166 | WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); | |
1167 | ||
1168 | spin_lock(&rq->lock); | |
1169 | __update_rq_clock(rq); | |
1170 | rq->curr->sched_class->task_tick(rq, rq->curr, 1); | |
1171 | spin_unlock(&rq->lock); | |
1172 | ||
1173 | return HRTIMER_NORESTART; | |
1174 | } | |
1175 | ||
1176 | static inline void init_rq_hrtick(struct rq *rq) | |
1177 | { | |
1178 | rq->hrtick_flags = 0; | |
1179 | hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
1180 | rq->hrtick_timer.function = hrtick; | |
1181 | rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ; | |
1182 | } | |
1183 | ||
1184 | void hrtick_resched(void) | |
1185 | { | |
1186 | struct rq *rq; | |
1187 | unsigned long flags; | |
1188 | ||
1189 | if (!test_thread_flag(TIF_HRTICK_RESCHED)) | |
1190 | return; | |
1191 | ||
1192 | local_irq_save(flags); | |
1193 | rq = cpu_rq(smp_processor_id()); | |
1194 | hrtick_set(rq); | |
1195 | local_irq_restore(flags); | |
1196 | } | |
1197 | #else | |
1198 | static inline void hrtick_clear(struct rq *rq) | |
1199 | { | |
1200 | } | |
1201 | ||
1202 | static inline void hrtick_set(struct rq *rq) | |
1203 | { | |
1204 | } | |
1205 | ||
1206 | static inline void init_rq_hrtick(struct rq *rq) | |
1207 | { | |
1208 | } | |
1209 | ||
1210 | void hrtick_resched(void) | |
1211 | { | |
1212 | } | |
1213 | #endif | |
1214 | ||
1215 | /* | |
1216 | * resched_task - mark a task 'to be rescheduled now'. | |
1217 | * | |
1218 | * On UP this means the setting of the need_resched flag, on SMP it | |
1219 | * might also involve a cross-CPU call to trigger the scheduler on | |
1220 | * the target CPU. | |
1221 | */ | |
1222 | #ifdef CONFIG_SMP | |
1223 | ||
1224 | #ifndef tsk_is_polling | |
1225 | #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG) | |
1226 | #endif | |
1227 | ||
1228 | static void __resched_task(struct task_struct *p, int tif_bit) | |
1229 | { | |
1230 | int cpu; | |
1231 | ||
1232 | assert_spin_locked(&task_rq(p)->lock); | |
1233 | ||
1234 | if (unlikely(test_tsk_thread_flag(p, tif_bit))) | |
1235 | return; | |
1236 | ||
1237 | set_tsk_thread_flag(p, tif_bit); | |
1238 | ||
1239 | cpu = task_cpu(p); | |
1240 | if (cpu == smp_processor_id()) | |
1241 | return; | |
1242 | ||
1243 | /* NEED_RESCHED must be visible before we test polling */ | |
1244 | smp_mb(); | |
1245 | if (!tsk_is_polling(p)) | |
1246 | smp_send_reschedule(cpu); | |
1247 | } | |
1248 | ||
1249 | static void resched_cpu(int cpu) | |
1250 | { | |
1251 | struct rq *rq = cpu_rq(cpu); | |
1252 | unsigned long flags; | |
1253 | ||
1254 | if (!spin_trylock_irqsave(&rq->lock, flags)) | |
1255 | return; | |
1256 | resched_task(cpu_curr(cpu)); | |
1257 | spin_unlock_irqrestore(&rq->lock, flags); | |
1258 | } | |
1259 | ||
1260 | #ifdef CONFIG_NO_HZ | |
1261 | /* | |
1262 | * When add_timer_on() enqueues a timer into the timer wheel of an | |
1263 | * idle CPU then this timer might expire before the next timer event | |
1264 | * which is scheduled to wake up that CPU. In case of a completely | |
1265 | * idle system the next event might even be infinite time into the | |
1266 | * future. wake_up_idle_cpu() ensures that the CPU is woken up and | |
1267 | * leaves the inner idle loop so the newly added timer is taken into | |
1268 | * account when the CPU goes back to idle and evaluates the timer | |
1269 | * wheel for the next timer event. | |
1270 | */ | |
1271 | void wake_up_idle_cpu(int cpu) | |
1272 | { | |
1273 | struct rq *rq = cpu_rq(cpu); | |
1274 | ||
1275 | if (cpu == smp_processor_id()) | |
1276 | return; | |
1277 | ||
1278 | /* | |
1279 | * This is safe, as this function is called with the timer | |
1280 | * wheel base lock of (cpu) held. When the CPU is on the way | |
1281 | * to idle and has not yet set rq->curr to idle then it will | |
1282 | * be serialized on the timer wheel base lock and take the new | |
1283 | * timer into account automatically. | |
1284 | */ | |
1285 | if (rq->curr != rq->idle) | |
1286 | return; | |
1287 | ||
1288 | /* | |
1289 | * We can set TIF_RESCHED on the idle task of the other CPU | |
1290 | * lockless. The worst case is that the other CPU runs the | |
1291 | * idle task through an additional NOOP schedule() | |
1292 | */ | |
1293 | set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED); | |
1294 | ||
1295 | /* NEED_RESCHED must be visible before we test polling */ | |
1296 | smp_mb(); | |
1297 | if (!tsk_is_polling(rq->idle)) | |
1298 | smp_send_reschedule(cpu); | |
1299 | } | |
1300 | #endif | |
1301 | ||
1302 | #else | |
1303 | static void __resched_task(struct task_struct *p, int tif_bit) | |
1304 | { | |
1305 | assert_spin_locked(&task_rq(p)->lock); | |
1306 | set_tsk_thread_flag(p, tif_bit); | |
1307 | } | |
1308 | #endif | |
1309 | ||
1310 | #if BITS_PER_LONG == 32 | |
1311 | # define WMULT_CONST (~0UL) | |
1312 | #else | |
1313 | # define WMULT_CONST (1UL << 32) | |
1314 | #endif | |
1315 | ||
1316 | #define WMULT_SHIFT 32 | |
1317 | ||
1318 | /* | |
1319 | * Shift right and round: | |
1320 | */ | |
1321 | #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y)) | |
1322 | ||
1323 | /* | |
1324 | * delta *= weight / lw | |
1325 | */ | |
1326 | static unsigned long | |
1327 | calc_delta_mine(unsigned long delta_exec, unsigned long weight, | |
1328 | struct load_weight *lw) | |
1329 | { | |
1330 | u64 tmp; | |
1331 | ||
1332 | if (unlikely(!lw->inv_weight)) | |
1333 | lw->inv_weight = (WMULT_CONST-lw->weight/2) / (lw->weight+1); | |
1334 | ||
1335 | tmp = (u64)delta_exec * weight; | |
1336 | /* | |
1337 | * Check whether we'd overflow the 64-bit multiplication: | |
1338 | */ | |
1339 | if (unlikely(tmp > WMULT_CONST)) | |
1340 | tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight, | |
1341 | WMULT_SHIFT/2); | |
1342 | else | |
1343 | tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT); | |
1344 | ||
1345 | return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX); | |
1346 | } | |
1347 | ||
1348 | static inline void update_load_add(struct load_weight *lw, unsigned long inc) | |
1349 | { | |
1350 | lw->weight += inc; | |
1351 | lw->inv_weight = 0; | |
1352 | } | |
1353 | ||
1354 | static inline void update_load_sub(struct load_weight *lw, unsigned long dec) | |
1355 | { | |
1356 | lw->weight -= dec; | |
1357 | lw->inv_weight = 0; | |
1358 | } | |
1359 | ||
1360 | /* | |
1361 | * To aid in avoiding the subversion of "niceness" due to uneven distribution | |
1362 | * of tasks with abnormal "nice" values across CPUs the contribution that | |
1363 | * each task makes to its run queue's load is weighted according to its | |
1364 | * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a | |
1365 | * scaled version of the new time slice allocation that they receive on time | |
1366 | * slice expiry etc. | |
1367 | */ | |
1368 | ||
1369 | #define WEIGHT_IDLEPRIO 2 | |
1370 | #define WMULT_IDLEPRIO (1 << 31) | |
1371 | ||
1372 | /* | |
1373 | * Nice levels are multiplicative, with a gentle 10% change for every | |
1374 | * nice level changed. I.e. when a CPU-bound task goes from nice 0 to | |
1375 | * nice 1, it will get ~10% less CPU time than another CPU-bound task | |
1376 | * that remained on nice 0. | |
1377 | * | |
1378 | * The "10% effect" is relative and cumulative: from _any_ nice level, | |
1379 | * if you go up 1 level, it's -10% CPU usage, if you go down 1 level | |
1380 | * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25. | |
1381 | * If a task goes up by ~10% and another task goes down by ~10% then | |
1382 | * the relative distance between them is ~25%.) | |
1383 | */ | |
1384 | static const int prio_to_weight[40] = { | |
1385 | /* -20 */ 88761, 71755, 56483, 46273, 36291, | |
1386 | /* -15 */ 29154, 23254, 18705, 14949, 11916, | |
1387 | /* -10 */ 9548, 7620, 6100, 4904, 3906, | |
1388 | /* -5 */ 3121, 2501, 1991, 1586, 1277, | |
1389 | /* 0 */ 1024, 820, 655, 526, 423, | |
1390 | /* 5 */ 335, 272, 215, 172, 137, | |
1391 | /* 10 */ 110, 87, 70, 56, 45, | |
1392 | /* 15 */ 36, 29, 23, 18, 15, | |
1393 | }; | |
1394 | ||
1395 | /* | |
1396 | * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated. | |
1397 | * | |
1398 | * In cases where the weight does not change often, we can use the | |
1399 | * precalculated inverse to speed up arithmetics by turning divisions | |
1400 | * into multiplications: | |
1401 | */ | |
1402 | static const u32 prio_to_wmult[40] = { | |
1403 | /* -20 */ 48388, 59856, 76040, 92818, 118348, | |
1404 | /* -15 */ 147320, 184698, 229616, 287308, 360437, | |
1405 | /* -10 */ 449829, 563644, 704093, 875809, 1099582, | |
1406 | /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326, | |
1407 | /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587, | |
1408 | /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126, | |
1409 | /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717, | |
1410 | /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153, | |
1411 | }; | |
1412 | ||
1413 | static void activate_task(struct rq *rq, struct task_struct *p, int wakeup); | |
1414 | ||
1415 | /* | |
1416 | * runqueue iterator, to support SMP load-balancing between different | |
1417 | * scheduling classes, without having to expose their internal data | |
1418 | * structures to the load-balancing proper: | |
1419 | */ | |
1420 | struct rq_iterator { | |
1421 | void *arg; | |
1422 | struct task_struct *(*start)(void *); | |
1423 | struct task_struct *(*next)(void *); | |
1424 | }; | |
1425 | ||
1426 | #ifdef CONFIG_SMP | |
1427 | static unsigned long | |
1428 | balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1429 | unsigned long max_load_move, struct sched_domain *sd, | |
1430 | enum cpu_idle_type idle, int *all_pinned, | |
1431 | int *this_best_prio, struct rq_iterator *iterator); | |
1432 | ||
1433 | static int | |
1434 | iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1435 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1436 | struct rq_iterator *iterator); | |
1437 | #endif | |
1438 | ||
1439 | #ifdef CONFIG_CGROUP_CPUACCT | |
1440 | static void cpuacct_charge(struct task_struct *tsk, u64 cputime); | |
1441 | #else | |
1442 | static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {} | |
1443 | #endif | |
1444 | ||
1445 | static inline void inc_cpu_load(struct rq *rq, unsigned long load) | |
1446 | { | |
1447 | update_load_add(&rq->load, load); | |
1448 | } | |
1449 | ||
1450 | static inline void dec_cpu_load(struct rq *rq, unsigned long load) | |
1451 | { | |
1452 | update_load_sub(&rq->load, load); | |
1453 | } | |
1454 | ||
1455 | #ifdef CONFIG_SMP | |
1456 | static unsigned long source_load(int cpu, int type); | |
1457 | static unsigned long target_load(int cpu, int type); | |
1458 | static unsigned long cpu_avg_load_per_task(int cpu); | |
1459 | static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd); | |
1460 | ||
1461 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1462 | ||
1463 | /* | |
1464 | * Group load balancing. | |
1465 | * | |
1466 | * We calculate a few balance domain wide aggregate numbers; load and weight. | |
1467 | * Given the pictures below, and assuming each item has equal weight: | |
1468 | * | |
1469 | * root 1 - thread | |
1470 | * / | \ A - group | |
1471 | * A 1 B | |
1472 | * /|\ / \ | |
1473 | * C 2 D 3 4 | |
1474 | * | | | |
1475 | * 5 6 | |
1476 | * | |
1477 | * load: | |
1478 | * A and B get 1/3-rd of the total load. C and D get 1/3-rd of A's 1/3-rd, | |
1479 | * which equals 1/9-th of the total load. | |
1480 | * | |
1481 | * shares: | |
1482 | * The weight of this group on the selected cpus. | |
1483 | * | |
1484 | * rq_weight: | |
1485 | * Direct sum of all the cpu's their rq weight, e.g. A would get 3 while | |
1486 | * B would get 2. | |
1487 | * | |
1488 | * task_weight: | |
1489 | * Part of the rq_weight contributed by tasks; all groups except B would | |
1490 | * get 1, B gets 2. | |
1491 | */ | |
1492 | ||
1493 | static inline struct aggregate_struct * | |
1494 | aggregate(struct task_group *tg, struct sched_domain *sd) | |
1495 | { | |
1496 | return &tg->cfs_rq[sd->first_cpu]->aggregate; | |
1497 | } | |
1498 | ||
1499 | typedef void (*aggregate_func)(struct task_group *, struct sched_domain *); | |
1500 | ||
1501 | /* | |
1502 | * Iterate the full tree, calling @down when first entering a node and @up when | |
1503 | * leaving it for the final time. | |
1504 | */ | |
1505 | static | |
1506 | void aggregate_walk_tree(aggregate_func down, aggregate_func up, | |
1507 | struct sched_domain *sd) | |
1508 | { | |
1509 | struct task_group *parent, *child; | |
1510 | ||
1511 | rcu_read_lock(); | |
1512 | parent = &root_task_group; | |
1513 | down: | |
1514 | (*down)(parent, sd); | |
1515 | list_for_each_entry_rcu(child, &parent->children, siblings) { | |
1516 | parent = child; | |
1517 | goto down; | |
1518 | ||
1519 | up: | |
1520 | continue; | |
1521 | } | |
1522 | (*up)(parent, sd); | |
1523 | ||
1524 | child = parent; | |
1525 | parent = parent->parent; | |
1526 | if (parent) | |
1527 | goto up; | |
1528 | rcu_read_unlock(); | |
1529 | } | |
1530 | ||
1531 | /* | |
1532 | * Calculate the aggregate runqueue weight. | |
1533 | */ | |
1534 | static | |
1535 | void aggregate_group_weight(struct task_group *tg, struct sched_domain *sd) | |
1536 | { | |
1537 | unsigned long rq_weight = 0; | |
1538 | unsigned long task_weight = 0; | |
1539 | int i; | |
1540 | ||
1541 | for_each_cpu_mask(i, sd->span) { | |
1542 | rq_weight += tg->cfs_rq[i]->load.weight; | |
1543 | task_weight += tg->cfs_rq[i]->task_weight; | |
1544 | } | |
1545 | ||
1546 | aggregate(tg, sd)->rq_weight = rq_weight; | |
1547 | aggregate(tg, sd)->task_weight = task_weight; | |
1548 | } | |
1549 | ||
1550 | /* | |
1551 | * Redistribute tg->shares amongst all tg->cfs_rq[]s. | |
1552 | */ | |
1553 | static void __aggregate_redistribute_shares(struct task_group *tg) | |
1554 | { | |
1555 | int i, max_cpu = smp_processor_id(); | |
1556 | unsigned long rq_weight = 0; | |
1557 | unsigned long shares, max_shares = 0, shares_rem = tg->shares; | |
1558 | ||
1559 | for_each_possible_cpu(i) | |
1560 | rq_weight += tg->cfs_rq[i]->load.weight; | |
1561 | ||
1562 | for_each_possible_cpu(i) { | |
1563 | /* | |
1564 | * divide shares proportional to the rq_weights. | |
1565 | */ | |
1566 | shares = tg->shares * tg->cfs_rq[i]->load.weight; | |
1567 | shares /= rq_weight + 1; | |
1568 | ||
1569 | tg->cfs_rq[i]->shares = shares; | |
1570 | ||
1571 | if (shares > max_shares) { | |
1572 | max_shares = shares; | |
1573 | max_cpu = i; | |
1574 | } | |
1575 | shares_rem -= shares; | |
1576 | } | |
1577 | ||
1578 | /* | |
1579 | * Ensure it all adds up to tg->shares; we can loose a few | |
1580 | * due to rounding down when computing the per-cpu shares. | |
1581 | */ | |
1582 | if (shares_rem) | |
1583 | tg->cfs_rq[max_cpu]->shares += shares_rem; | |
1584 | } | |
1585 | ||
1586 | /* | |
1587 | * Compute the weight of this group on the given cpus. | |
1588 | */ | |
1589 | static | |
1590 | void aggregate_group_shares(struct task_group *tg, struct sched_domain *sd) | |
1591 | { | |
1592 | unsigned long shares = 0; | |
1593 | int i; | |
1594 | ||
1595 | again: | |
1596 | for_each_cpu_mask(i, sd->span) | |
1597 | shares += tg->cfs_rq[i]->shares; | |
1598 | ||
1599 | /* | |
1600 | * When the span doesn't have any shares assigned, but does have | |
1601 | * tasks to run do a machine wide rebalance (should be rare). | |
1602 | */ | |
1603 | if (unlikely(!shares && aggregate(tg, sd)->rq_weight)) { | |
1604 | __aggregate_redistribute_shares(tg); | |
1605 | goto again; | |
1606 | } | |
1607 | ||
1608 | aggregate(tg, sd)->shares = shares; | |
1609 | } | |
1610 | ||
1611 | /* | |
1612 | * Compute the load fraction assigned to this group, relies on the aggregate | |
1613 | * weight and this group's parent's load, i.e. top-down. | |
1614 | */ | |
1615 | static | |
1616 | void aggregate_group_load(struct task_group *tg, struct sched_domain *sd) | |
1617 | { | |
1618 | unsigned long load; | |
1619 | ||
1620 | if (!tg->parent) { | |
1621 | int i; | |
1622 | ||
1623 | load = 0; | |
1624 | for_each_cpu_mask(i, sd->span) | |
1625 | load += cpu_rq(i)->load.weight; | |
1626 | ||
1627 | } else { | |
1628 | load = aggregate(tg->parent, sd)->load; | |
1629 | ||
1630 | /* | |
1631 | * shares is our weight in the parent's rq so | |
1632 | * shares/parent->rq_weight gives our fraction of the load | |
1633 | */ | |
1634 | load *= aggregate(tg, sd)->shares; | |
1635 | load /= aggregate(tg->parent, sd)->rq_weight + 1; | |
1636 | } | |
1637 | ||
1638 | aggregate(tg, sd)->load = load; | |
1639 | } | |
1640 | ||
1641 | static void __set_se_shares(struct sched_entity *se, unsigned long shares); | |
1642 | ||
1643 | /* | |
1644 | * Calculate and set the cpu's group shares. | |
1645 | */ | |
1646 | static void | |
1647 | __update_group_shares_cpu(struct task_group *tg, struct sched_domain *sd, | |
1648 | int tcpu) | |
1649 | { | |
1650 | int boost = 0; | |
1651 | unsigned long shares; | |
1652 | unsigned long rq_weight; | |
1653 | ||
1654 | if (!tg->se[tcpu]) | |
1655 | return; | |
1656 | ||
1657 | rq_weight = tg->cfs_rq[tcpu]->load.weight; | |
1658 | ||
1659 | /* | |
1660 | * If there are currently no tasks on the cpu pretend there is one of | |
1661 | * average load so that when a new task gets to run here it will not | |
1662 | * get delayed by group starvation. | |
1663 | */ | |
1664 | if (!rq_weight) { | |
1665 | boost = 1; | |
1666 | rq_weight = NICE_0_LOAD; | |
1667 | } | |
1668 | ||
1669 | /* | |
1670 | * \Sum shares * rq_weight | |
1671 | * shares = ----------------------- | |
1672 | * \Sum rq_weight | |
1673 | * | |
1674 | */ | |
1675 | shares = aggregate(tg, sd)->shares * rq_weight; | |
1676 | shares /= aggregate(tg, sd)->rq_weight + 1; | |
1677 | ||
1678 | /* | |
1679 | * record the actual number of shares, not the boosted amount. | |
1680 | */ | |
1681 | tg->cfs_rq[tcpu]->shares = boost ? 0 : shares; | |
1682 | ||
1683 | if (shares < MIN_SHARES) | |
1684 | shares = MIN_SHARES; | |
1685 | ||
1686 | __set_se_shares(tg->se[tcpu], shares); | |
1687 | } | |
1688 | ||
1689 | /* | |
1690 | * Re-adjust the weights on the cpu the task came from and on the cpu the | |
1691 | * task went to. | |
1692 | */ | |
1693 | static void | |
1694 | __move_group_shares(struct task_group *tg, struct sched_domain *sd, | |
1695 | int scpu, int dcpu) | |
1696 | { | |
1697 | unsigned long shares; | |
1698 | ||
1699 | shares = tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares; | |
1700 | ||
1701 | __update_group_shares_cpu(tg, sd, scpu); | |
1702 | __update_group_shares_cpu(tg, sd, dcpu); | |
1703 | ||
1704 | /* | |
1705 | * ensure we never loose shares due to rounding errors in the | |
1706 | * above redistribution. | |
1707 | */ | |
1708 | shares -= tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares; | |
1709 | if (shares) | |
1710 | tg->cfs_rq[dcpu]->shares += shares; | |
1711 | } | |
1712 | ||
1713 | /* | |
1714 | * Because changing a group's shares changes the weight of the super-group | |
1715 | * we need to walk up the tree and change all shares until we hit the root. | |
1716 | */ | |
1717 | static void | |
1718 | move_group_shares(struct task_group *tg, struct sched_domain *sd, | |
1719 | int scpu, int dcpu) | |
1720 | { | |
1721 | while (tg) { | |
1722 | __move_group_shares(tg, sd, scpu, dcpu); | |
1723 | tg = tg->parent; | |
1724 | } | |
1725 | } | |
1726 | ||
1727 | static | |
1728 | void aggregate_group_set_shares(struct task_group *tg, struct sched_domain *sd) | |
1729 | { | |
1730 | unsigned long shares = aggregate(tg, sd)->shares; | |
1731 | int i; | |
1732 | ||
1733 | for_each_cpu_mask(i, sd->span) { | |
1734 | struct rq *rq = cpu_rq(i); | |
1735 | unsigned long flags; | |
1736 | ||
1737 | spin_lock_irqsave(&rq->lock, flags); | |
1738 | __update_group_shares_cpu(tg, sd, i); | |
1739 | spin_unlock_irqrestore(&rq->lock, flags); | |
1740 | } | |
1741 | ||
1742 | aggregate_group_shares(tg, sd); | |
1743 | ||
1744 | /* | |
1745 | * ensure we never loose shares due to rounding errors in the | |
1746 | * above redistribution. | |
1747 | */ | |
1748 | shares -= aggregate(tg, sd)->shares; | |
1749 | if (shares) { | |
1750 | tg->cfs_rq[sd->first_cpu]->shares += shares; | |
1751 | aggregate(tg, sd)->shares += shares; | |
1752 | } | |
1753 | } | |
1754 | ||
1755 | /* | |
1756 | * Calculate the accumulative weight and recursive load of each task group | |
1757 | * while walking down the tree. | |
1758 | */ | |
1759 | static | |
1760 | void aggregate_get_down(struct task_group *tg, struct sched_domain *sd) | |
1761 | { | |
1762 | aggregate_group_weight(tg, sd); | |
1763 | aggregate_group_shares(tg, sd); | |
1764 | aggregate_group_load(tg, sd); | |
1765 | } | |
1766 | ||
1767 | /* | |
1768 | * Rebalance the cpu shares while walking back up the tree. | |
1769 | */ | |
1770 | static | |
1771 | void aggregate_get_up(struct task_group *tg, struct sched_domain *sd) | |
1772 | { | |
1773 | aggregate_group_set_shares(tg, sd); | |
1774 | } | |
1775 | ||
1776 | static DEFINE_PER_CPU(spinlock_t, aggregate_lock); | |
1777 | ||
1778 | static void __init init_aggregate(void) | |
1779 | { | |
1780 | int i; | |
1781 | ||
1782 | for_each_possible_cpu(i) | |
1783 | spin_lock_init(&per_cpu(aggregate_lock, i)); | |
1784 | } | |
1785 | ||
1786 | static int get_aggregate(struct sched_domain *sd) | |
1787 | { | |
1788 | if (!spin_trylock(&per_cpu(aggregate_lock, sd->first_cpu))) | |
1789 | return 0; | |
1790 | ||
1791 | aggregate_walk_tree(aggregate_get_down, aggregate_get_up, sd); | |
1792 | return 1; | |
1793 | } | |
1794 | ||
1795 | static void put_aggregate(struct sched_domain *sd) | |
1796 | { | |
1797 | spin_unlock(&per_cpu(aggregate_lock, sd->first_cpu)); | |
1798 | } | |
1799 | ||
1800 | static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares) | |
1801 | { | |
1802 | cfs_rq->shares = shares; | |
1803 | } | |
1804 | ||
1805 | #else | |
1806 | ||
1807 | static inline void init_aggregate(void) | |
1808 | { | |
1809 | } | |
1810 | ||
1811 | static inline int get_aggregate(struct sched_domain *sd) | |
1812 | { | |
1813 | return 0; | |
1814 | } | |
1815 | ||
1816 | static inline void put_aggregate(struct sched_domain *sd) | |
1817 | { | |
1818 | } | |
1819 | #endif | |
1820 | ||
1821 | #else /* CONFIG_SMP */ | |
1822 | ||
1823 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1824 | static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares) | |
1825 | { | |
1826 | } | |
1827 | #endif | |
1828 | ||
1829 | #endif /* CONFIG_SMP */ | |
1830 | ||
1831 | #include "sched_stats.h" | |
1832 | #include "sched_idletask.c" | |
1833 | #include "sched_fair.c" | |
1834 | #include "sched_rt.c" | |
1835 | #ifdef CONFIG_SCHED_DEBUG | |
1836 | # include "sched_debug.c" | |
1837 | #endif | |
1838 | ||
1839 | #define sched_class_highest (&rt_sched_class) | |
1840 | ||
1841 | static void inc_nr_running(struct rq *rq) | |
1842 | { | |
1843 | rq->nr_running++; | |
1844 | } | |
1845 | ||
1846 | static void dec_nr_running(struct rq *rq) | |
1847 | { | |
1848 | rq->nr_running--; | |
1849 | } | |
1850 | ||
1851 | static void set_load_weight(struct task_struct *p) | |
1852 | { | |
1853 | if (task_has_rt_policy(p)) { | |
1854 | p->se.load.weight = prio_to_weight[0] * 2; | |
1855 | p->se.load.inv_weight = prio_to_wmult[0] >> 1; | |
1856 | return; | |
1857 | } | |
1858 | ||
1859 | /* | |
1860 | * SCHED_IDLE tasks get minimal weight: | |
1861 | */ | |
1862 | if (p->policy == SCHED_IDLE) { | |
1863 | p->se.load.weight = WEIGHT_IDLEPRIO; | |
1864 | p->se.load.inv_weight = WMULT_IDLEPRIO; | |
1865 | return; | |
1866 | } | |
1867 | ||
1868 | p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO]; | |
1869 | p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO]; | |
1870 | } | |
1871 | ||
1872 | static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup) | |
1873 | { | |
1874 | sched_info_queued(p); | |
1875 | p->sched_class->enqueue_task(rq, p, wakeup); | |
1876 | p->se.on_rq = 1; | |
1877 | } | |
1878 | ||
1879 | static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep) | |
1880 | { | |
1881 | p->sched_class->dequeue_task(rq, p, sleep); | |
1882 | p->se.on_rq = 0; | |
1883 | } | |
1884 | ||
1885 | /* | |
1886 | * __normal_prio - return the priority that is based on the static prio | |
1887 | */ | |
1888 | static inline int __normal_prio(struct task_struct *p) | |
1889 | { | |
1890 | return p->static_prio; | |
1891 | } | |
1892 | ||
1893 | /* | |
1894 | * Calculate the expected normal priority: i.e. priority | |
1895 | * without taking RT-inheritance into account. Might be | |
1896 | * boosted by interactivity modifiers. Changes upon fork, | |
1897 | * setprio syscalls, and whenever the interactivity | |
1898 | * estimator recalculates. | |
1899 | */ | |
1900 | static inline int normal_prio(struct task_struct *p) | |
1901 | { | |
1902 | int prio; | |
1903 | ||
1904 | if (task_has_rt_policy(p)) | |
1905 | prio = MAX_RT_PRIO-1 - p->rt_priority; | |
1906 | else | |
1907 | prio = __normal_prio(p); | |
1908 | return prio; | |
1909 | } | |
1910 | ||
1911 | /* | |
1912 | * Calculate the current priority, i.e. the priority | |
1913 | * taken into account by the scheduler. This value might | |
1914 | * be boosted by RT tasks, or might be boosted by | |
1915 | * interactivity modifiers. Will be RT if the task got | |
1916 | * RT-boosted. If not then it returns p->normal_prio. | |
1917 | */ | |
1918 | static int effective_prio(struct task_struct *p) | |
1919 | { | |
1920 | p->normal_prio = normal_prio(p); | |
1921 | /* | |
1922 | * If we are RT tasks or we were boosted to RT priority, | |
1923 | * keep the priority unchanged. Otherwise, update priority | |
1924 | * to the normal priority: | |
1925 | */ | |
1926 | if (!rt_prio(p->prio)) | |
1927 | return p->normal_prio; | |
1928 | return p->prio; | |
1929 | } | |
1930 | ||
1931 | /* | |
1932 | * activate_task - move a task to the runqueue. | |
1933 | */ | |
1934 | static void activate_task(struct rq *rq, struct task_struct *p, int wakeup) | |
1935 | { | |
1936 | if (task_contributes_to_load(p)) | |
1937 | rq->nr_uninterruptible--; | |
1938 | ||
1939 | enqueue_task(rq, p, wakeup); | |
1940 | inc_nr_running(rq); | |
1941 | } | |
1942 | ||
1943 | /* | |
1944 | * deactivate_task - remove a task from the runqueue. | |
1945 | */ | |
1946 | static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep) | |
1947 | { | |
1948 | if (task_contributes_to_load(p)) | |
1949 | rq->nr_uninterruptible++; | |
1950 | ||
1951 | dequeue_task(rq, p, sleep); | |
1952 | dec_nr_running(rq); | |
1953 | } | |
1954 | ||
1955 | /** | |
1956 | * task_curr - is this task currently executing on a CPU? | |
1957 | * @p: the task in question. | |
1958 | */ | |
1959 | inline int task_curr(const struct task_struct *p) | |
1960 | { | |
1961 | return cpu_curr(task_cpu(p)) == p; | |
1962 | } | |
1963 | ||
1964 | /* Used instead of source_load when we know the type == 0 */ | |
1965 | unsigned long weighted_cpuload(const int cpu) | |
1966 | { | |
1967 | return cpu_rq(cpu)->load.weight; | |
1968 | } | |
1969 | ||
1970 | static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu) | |
1971 | { | |
1972 | set_task_rq(p, cpu); | |
1973 | #ifdef CONFIG_SMP | |
1974 | /* | |
1975 | * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be | |
1976 | * successfuly executed on another CPU. We must ensure that updates of | |
1977 | * per-task data have been completed by this moment. | |
1978 | */ | |
1979 | smp_wmb(); | |
1980 | task_thread_info(p)->cpu = cpu; | |
1981 | #endif | |
1982 | } | |
1983 | ||
1984 | static inline void check_class_changed(struct rq *rq, struct task_struct *p, | |
1985 | const struct sched_class *prev_class, | |
1986 | int oldprio, int running) | |
1987 | { | |
1988 | if (prev_class != p->sched_class) { | |
1989 | if (prev_class->switched_from) | |
1990 | prev_class->switched_from(rq, p, running); | |
1991 | p->sched_class->switched_to(rq, p, running); | |
1992 | } else | |
1993 | p->sched_class->prio_changed(rq, p, oldprio, running); | |
1994 | } | |
1995 | ||
1996 | #ifdef CONFIG_SMP | |
1997 | ||
1998 | /* | |
1999 | * Is this task likely cache-hot: | |
2000 | */ | |
2001 | static int | |
2002 | task_hot(struct task_struct *p, u64 now, struct sched_domain *sd) | |
2003 | { | |
2004 | s64 delta; | |
2005 | ||
2006 | /* | |
2007 | * Buddy candidates are cache hot: | |
2008 | */ | |
2009 | if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next)) | |
2010 | return 1; | |
2011 | ||
2012 | if (p->sched_class != &fair_sched_class) | |
2013 | return 0; | |
2014 | ||
2015 | if (sysctl_sched_migration_cost == -1) | |
2016 | return 1; | |
2017 | if (sysctl_sched_migration_cost == 0) | |
2018 | return 0; | |
2019 | ||
2020 | delta = now - p->se.exec_start; | |
2021 | ||
2022 | return delta < (s64)sysctl_sched_migration_cost; | |
2023 | } | |
2024 | ||
2025 | ||
2026 | void set_task_cpu(struct task_struct *p, unsigned int new_cpu) | |
2027 | { | |
2028 | int old_cpu = task_cpu(p); | |
2029 | struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu); | |
2030 | struct cfs_rq *old_cfsrq = task_cfs_rq(p), | |
2031 | *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu); | |
2032 | u64 clock_offset; | |
2033 | ||
2034 | clock_offset = old_rq->clock - new_rq->clock; | |
2035 | ||
2036 | #ifdef CONFIG_SCHEDSTATS | |
2037 | if (p->se.wait_start) | |
2038 | p->se.wait_start -= clock_offset; | |
2039 | if (p->se.sleep_start) | |
2040 | p->se.sleep_start -= clock_offset; | |
2041 | if (p->se.block_start) | |
2042 | p->se.block_start -= clock_offset; | |
2043 | if (old_cpu != new_cpu) { | |
2044 | schedstat_inc(p, se.nr_migrations); | |
2045 | if (task_hot(p, old_rq->clock, NULL)) | |
2046 | schedstat_inc(p, se.nr_forced2_migrations); | |
2047 | } | |
2048 | #endif | |
2049 | p->se.vruntime -= old_cfsrq->min_vruntime - | |
2050 | new_cfsrq->min_vruntime; | |
2051 | ||
2052 | __set_task_cpu(p, new_cpu); | |
2053 | } | |
2054 | ||
2055 | struct migration_req { | |
2056 | struct list_head list; | |
2057 | ||
2058 | struct task_struct *task; | |
2059 | int dest_cpu; | |
2060 | ||
2061 | struct completion done; | |
2062 | }; | |
2063 | ||
2064 | /* | |
2065 | * The task's runqueue lock must be held. | |
2066 | * Returns true if you have to wait for migration thread. | |
2067 | */ | |
2068 | static int | |
2069 | migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req) | |
2070 | { | |
2071 | struct rq *rq = task_rq(p); | |
2072 | ||
2073 | /* | |
2074 | * If the task is not on a runqueue (and not running), then | |
2075 | * it is sufficient to simply update the task's cpu field. | |
2076 | */ | |
2077 | if (!p->se.on_rq && !task_running(rq, p)) { | |
2078 | set_task_cpu(p, dest_cpu); | |
2079 | return 0; | |
2080 | } | |
2081 | ||
2082 | init_completion(&req->done); | |
2083 | req->task = p; | |
2084 | req->dest_cpu = dest_cpu; | |
2085 | list_add(&req->list, &rq->migration_queue); | |
2086 | ||
2087 | return 1; | |
2088 | } | |
2089 | ||
2090 | /* | |
2091 | * wait_task_inactive - wait for a thread to unschedule. | |
2092 | * | |
2093 | * The caller must ensure that the task *will* unschedule sometime soon, | |
2094 | * else this function might spin for a *long* time. This function can't | |
2095 | * be called with interrupts off, or it may introduce deadlock with | |
2096 | * smp_call_function() if an IPI is sent by the same process we are | |
2097 | * waiting to become inactive. | |
2098 | */ | |
2099 | void wait_task_inactive(struct task_struct *p) | |
2100 | { | |
2101 | unsigned long flags; | |
2102 | int running, on_rq; | |
2103 | struct rq *rq; | |
2104 | ||
2105 | for (;;) { | |
2106 | /* | |
2107 | * We do the initial early heuristics without holding | |
2108 | * any task-queue locks at all. We'll only try to get | |
2109 | * the runqueue lock when things look like they will | |
2110 | * work out! | |
2111 | */ | |
2112 | rq = task_rq(p); | |
2113 | ||
2114 | /* | |
2115 | * If the task is actively running on another CPU | |
2116 | * still, just relax and busy-wait without holding | |
2117 | * any locks. | |
2118 | * | |
2119 | * NOTE! Since we don't hold any locks, it's not | |
2120 | * even sure that "rq" stays as the right runqueue! | |
2121 | * But we don't care, since "task_running()" will | |
2122 | * return false if the runqueue has changed and p | |
2123 | * is actually now running somewhere else! | |
2124 | */ | |
2125 | while (task_running(rq, p)) | |
2126 | cpu_relax(); | |
2127 | ||
2128 | /* | |
2129 | * Ok, time to look more closely! We need the rq | |
2130 | * lock now, to be *sure*. If we're wrong, we'll | |
2131 | * just go back and repeat. | |
2132 | */ | |
2133 | rq = task_rq_lock(p, &flags); | |
2134 | running = task_running(rq, p); | |
2135 | on_rq = p->se.on_rq; | |
2136 | task_rq_unlock(rq, &flags); | |
2137 | ||
2138 | /* | |
2139 | * Was it really running after all now that we | |
2140 | * checked with the proper locks actually held? | |
2141 | * | |
2142 | * Oops. Go back and try again.. | |
2143 | */ | |
2144 | if (unlikely(running)) { | |
2145 | cpu_relax(); | |
2146 | continue; | |
2147 | } | |
2148 | ||
2149 | /* | |
2150 | * It's not enough that it's not actively running, | |
2151 | * it must be off the runqueue _entirely_, and not | |
2152 | * preempted! | |
2153 | * | |
2154 | * So if it wa still runnable (but just not actively | |
2155 | * running right now), it's preempted, and we should | |
2156 | * yield - it could be a while. | |
2157 | */ | |
2158 | if (unlikely(on_rq)) { | |
2159 | schedule_timeout_uninterruptible(1); | |
2160 | continue; | |
2161 | } | |
2162 | ||
2163 | /* | |
2164 | * Ahh, all good. It wasn't running, and it wasn't | |
2165 | * runnable, which means that it will never become | |
2166 | * running in the future either. We're all done! | |
2167 | */ | |
2168 | break; | |
2169 | } | |
2170 | } | |
2171 | ||
2172 | /*** | |
2173 | * kick_process - kick a running thread to enter/exit the kernel | |
2174 | * @p: the to-be-kicked thread | |
2175 | * | |
2176 | * Cause a process which is running on another CPU to enter | |
2177 | * kernel-mode, without any delay. (to get signals handled.) | |
2178 | * | |
2179 | * NOTE: this function doesnt have to take the runqueue lock, | |
2180 | * because all it wants to ensure is that the remote task enters | |
2181 | * the kernel. If the IPI races and the task has been migrated | |
2182 | * to another CPU then no harm is done and the purpose has been | |
2183 | * achieved as well. | |
2184 | */ | |
2185 | void kick_process(struct task_struct *p) | |
2186 | { | |
2187 | int cpu; | |
2188 | ||
2189 | preempt_disable(); | |
2190 | cpu = task_cpu(p); | |
2191 | if ((cpu != smp_processor_id()) && task_curr(p)) | |
2192 | smp_send_reschedule(cpu); | |
2193 | preempt_enable(); | |
2194 | } | |
2195 | ||
2196 | /* | |
2197 | * Return a low guess at the load of a migration-source cpu weighted | |
2198 | * according to the scheduling class and "nice" value. | |
2199 | * | |
2200 | * We want to under-estimate the load of migration sources, to | |
2201 | * balance conservatively. | |
2202 | */ | |
2203 | static unsigned long source_load(int cpu, int type) | |
2204 | { | |
2205 | struct rq *rq = cpu_rq(cpu); | |
2206 | unsigned long total = weighted_cpuload(cpu); | |
2207 | ||
2208 | if (type == 0) | |
2209 | return total; | |
2210 | ||
2211 | return min(rq->cpu_load[type-1], total); | |
2212 | } | |
2213 | ||
2214 | /* | |
2215 | * Return a high guess at the load of a migration-target cpu weighted | |
2216 | * according to the scheduling class and "nice" value. | |
2217 | */ | |
2218 | static unsigned long target_load(int cpu, int type) | |
2219 | { | |
2220 | struct rq *rq = cpu_rq(cpu); | |
2221 | unsigned long total = weighted_cpuload(cpu); | |
2222 | ||
2223 | if (type == 0) | |
2224 | return total; | |
2225 | ||
2226 | return max(rq->cpu_load[type-1], total); | |
2227 | } | |
2228 | ||
2229 | /* | |
2230 | * Return the average load per task on the cpu's run queue | |
2231 | */ | |
2232 | static unsigned long cpu_avg_load_per_task(int cpu) | |
2233 | { | |
2234 | struct rq *rq = cpu_rq(cpu); | |
2235 | unsigned long total = weighted_cpuload(cpu); | |
2236 | unsigned long n = rq->nr_running; | |
2237 | ||
2238 | return n ? total / n : SCHED_LOAD_SCALE; | |
2239 | } | |
2240 | ||
2241 | /* | |
2242 | * find_idlest_group finds and returns the least busy CPU group within the | |
2243 | * domain. | |
2244 | */ | |
2245 | static struct sched_group * | |
2246 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu) | |
2247 | { | |
2248 | struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups; | |
2249 | unsigned long min_load = ULONG_MAX, this_load = 0; | |
2250 | int load_idx = sd->forkexec_idx; | |
2251 | int imbalance = 100 + (sd->imbalance_pct-100)/2; | |
2252 | ||
2253 | do { | |
2254 | unsigned long load, avg_load; | |
2255 | int local_group; | |
2256 | int i; | |
2257 | ||
2258 | /* Skip over this group if it has no CPUs allowed */ | |
2259 | if (!cpus_intersects(group->cpumask, p->cpus_allowed)) | |
2260 | continue; | |
2261 | ||
2262 | local_group = cpu_isset(this_cpu, group->cpumask); | |
2263 | ||
2264 | /* Tally up the load of all CPUs in the group */ | |
2265 | avg_load = 0; | |
2266 | ||
2267 | for_each_cpu_mask(i, group->cpumask) { | |
2268 | /* Bias balancing toward cpus of our domain */ | |
2269 | if (local_group) | |
2270 | load = source_load(i, load_idx); | |
2271 | else | |
2272 | load = target_load(i, load_idx); | |
2273 | ||
2274 | avg_load += load; | |
2275 | } | |
2276 | ||
2277 | /* Adjust by relative CPU power of the group */ | |
2278 | avg_load = sg_div_cpu_power(group, | |
2279 | avg_load * SCHED_LOAD_SCALE); | |
2280 | ||
2281 | if (local_group) { | |
2282 | this_load = avg_load; | |
2283 | this = group; | |
2284 | } else if (avg_load < min_load) { | |
2285 | min_load = avg_load; | |
2286 | idlest = group; | |
2287 | } | |
2288 | } while (group = group->next, group != sd->groups); | |
2289 | ||
2290 | if (!idlest || 100*this_load < imbalance*min_load) | |
2291 | return NULL; | |
2292 | return idlest; | |
2293 | } | |
2294 | ||
2295 | /* | |
2296 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
2297 | */ | |
2298 | static int | |
2299 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu, | |
2300 | cpumask_t *tmp) | |
2301 | { | |
2302 | unsigned long load, min_load = ULONG_MAX; | |
2303 | int idlest = -1; | |
2304 | int i; | |
2305 | ||
2306 | /* Traverse only the allowed CPUs */ | |
2307 | cpus_and(*tmp, group->cpumask, p->cpus_allowed); | |
2308 | ||
2309 | for_each_cpu_mask(i, *tmp) { | |
2310 | load = weighted_cpuload(i); | |
2311 | ||
2312 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
2313 | min_load = load; | |
2314 | idlest = i; | |
2315 | } | |
2316 | } | |
2317 | ||
2318 | return idlest; | |
2319 | } | |
2320 | ||
2321 | /* | |
2322 | * sched_balance_self: balance the current task (running on cpu) in domains | |
2323 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
2324 | * SD_BALANCE_EXEC. | |
2325 | * | |
2326 | * Balance, ie. select the least loaded group. | |
2327 | * | |
2328 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
2329 | * | |
2330 | * preempt must be disabled. | |
2331 | */ | |
2332 | static int sched_balance_self(int cpu, int flag) | |
2333 | { | |
2334 | struct task_struct *t = current; | |
2335 | struct sched_domain *tmp, *sd = NULL; | |
2336 | ||
2337 | for_each_domain(cpu, tmp) { | |
2338 | /* | |
2339 | * If power savings logic is enabled for a domain, stop there. | |
2340 | */ | |
2341 | if (tmp->flags & SD_POWERSAVINGS_BALANCE) | |
2342 | break; | |
2343 | if (tmp->flags & flag) | |
2344 | sd = tmp; | |
2345 | } | |
2346 | ||
2347 | while (sd) { | |
2348 | cpumask_t span, tmpmask; | |
2349 | struct sched_group *group; | |
2350 | int new_cpu, weight; | |
2351 | ||
2352 | if (!(sd->flags & flag)) { | |
2353 | sd = sd->child; | |
2354 | continue; | |
2355 | } | |
2356 | ||
2357 | span = sd->span; | |
2358 | group = find_idlest_group(sd, t, cpu); | |
2359 | if (!group) { | |
2360 | sd = sd->child; | |
2361 | continue; | |
2362 | } | |
2363 | ||
2364 | new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask); | |
2365 | if (new_cpu == -1 || new_cpu == cpu) { | |
2366 | /* Now try balancing at a lower domain level of cpu */ | |
2367 | sd = sd->child; | |
2368 | continue; | |
2369 | } | |
2370 | ||
2371 | /* Now try balancing at a lower domain level of new_cpu */ | |
2372 | cpu = new_cpu; | |
2373 | sd = NULL; | |
2374 | weight = cpus_weight(span); | |
2375 | for_each_domain(cpu, tmp) { | |
2376 | if (weight <= cpus_weight(tmp->span)) | |
2377 | break; | |
2378 | if (tmp->flags & flag) | |
2379 | sd = tmp; | |
2380 | } | |
2381 | /* while loop will break here if sd == NULL */ | |
2382 | } | |
2383 | ||
2384 | return cpu; | |
2385 | } | |
2386 | ||
2387 | #endif /* CONFIG_SMP */ | |
2388 | ||
2389 | /*** | |
2390 | * try_to_wake_up - wake up a thread | |
2391 | * @p: the to-be-woken-up thread | |
2392 | * @state: the mask of task states that can be woken | |
2393 | * @sync: do a synchronous wakeup? | |
2394 | * | |
2395 | * Put it on the run-queue if it's not already there. The "current" | |
2396 | * thread is always on the run-queue (except when the actual | |
2397 | * re-schedule is in progress), and as such you're allowed to do | |
2398 | * the simpler "current->state = TASK_RUNNING" to mark yourself | |
2399 | * runnable without the overhead of this. | |
2400 | * | |
2401 | * returns failure only if the task is already active. | |
2402 | */ | |
2403 | static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync) | |
2404 | { | |
2405 | int cpu, orig_cpu, this_cpu, success = 0; | |
2406 | unsigned long flags; | |
2407 | long old_state; | |
2408 | struct rq *rq; | |
2409 | ||
2410 | if (!sched_feat(SYNC_WAKEUPS)) | |
2411 | sync = 0; | |
2412 | ||
2413 | smp_wmb(); | |
2414 | rq = task_rq_lock(p, &flags); | |
2415 | old_state = p->state; | |
2416 | if (!(old_state & state)) | |
2417 | goto out; | |
2418 | ||
2419 | if (p->se.on_rq) | |
2420 | goto out_running; | |
2421 | ||
2422 | cpu = task_cpu(p); | |
2423 | orig_cpu = cpu; | |
2424 | this_cpu = smp_processor_id(); | |
2425 | ||
2426 | #ifdef CONFIG_SMP | |
2427 | if (unlikely(task_running(rq, p))) | |
2428 | goto out_activate; | |
2429 | ||
2430 | cpu = p->sched_class->select_task_rq(p, sync); | |
2431 | if (cpu != orig_cpu) { | |
2432 | set_task_cpu(p, cpu); | |
2433 | task_rq_unlock(rq, &flags); | |
2434 | /* might preempt at this point */ | |
2435 | rq = task_rq_lock(p, &flags); | |
2436 | old_state = p->state; | |
2437 | if (!(old_state & state)) | |
2438 | goto out; | |
2439 | if (p->se.on_rq) | |
2440 | goto out_running; | |
2441 | ||
2442 | this_cpu = smp_processor_id(); | |
2443 | cpu = task_cpu(p); | |
2444 | } | |
2445 | ||
2446 | #ifdef CONFIG_SCHEDSTATS | |
2447 | schedstat_inc(rq, ttwu_count); | |
2448 | if (cpu == this_cpu) | |
2449 | schedstat_inc(rq, ttwu_local); | |
2450 | else { | |
2451 | struct sched_domain *sd; | |
2452 | for_each_domain(this_cpu, sd) { | |
2453 | if (cpu_isset(cpu, sd->span)) { | |
2454 | schedstat_inc(sd, ttwu_wake_remote); | |
2455 | break; | |
2456 | } | |
2457 | } | |
2458 | } | |
2459 | #endif | |
2460 | ||
2461 | out_activate: | |
2462 | #endif /* CONFIG_SMP */ | |
2463 | schedstat_inc(p, se.nr_wakeups); | |
2464 | if (sync) | |
2465 | schedstat_inc(p, se.nr_wakeups_sync); | |
2466 | if (orig_cpu != cpu) | |
2467 | schedstat_inc(p, se.nr_wakeups_migrate); | |
2468 | if (cpu == this_cpu) | |
2469 | schedstat_inc(p, se.nr_wakeups_local); | |
2470 | else | |
2471 | schedstat_inc(p, se.nr_wakeups_remote); | |
2472 | update_rq_clock(rq); | |
2473 | activate_task(rq, p, 1); | |
2474 | success = 1; | |
2475 | ||
2476 | out_running: | |
2477 | check_preempt_curr(rq, p); | |
2478 | ||
2479 | p->state = TASK_RUNNING; | |
2480 | #ifdef CONFIG_SMP | |
2481 | if (p->sched_class->task_wake_up) | |
2482 | p->sched_class->task_wake_up(rq, p); | |
2483 | #endif | |
2484 | out: | |
2485 | task_rq_unlock(rq, &flags); | |
2486 | ||
2487 | return success; | |
2488 | } | |
2489 | ||
2490 | int wake_up_process(struct task_struct *p) | |
2491 | { | |
2492 | return try_to_wake_up(p, TASK_ALL, 0); | |
2493 | } | |
2494 | EXPORT_SYMBOL(wake_up_process); | |
2495 | ||
2496 | int wake_up_state(struct task_struct *p, unsigned int state) | |
2497 | { | |
2498 | return try_to_wake_up(p, state, 0); | |
2499 | } | |
2500 | ||
2501 | /* | |
2502 | * Perform scheduler related setup for a newly forked process p. | |
2503 | * p is forked by current. | |
2504 | * | |
2505 | * __sched_fork() is basic setup used by init_idle() too: | |
2506 | */ | |
2507 | static void __sched_fork(struct task_struct *p) | |
2508 | { | |
2509 | p->se.exec_start = 0; | |
2510 | p->se.sum_exec_runtime = 0; | |
2511 | p->se.prev_sum_exec_runtime = 0; | |
2512 | p->se.last_wakeup = 0; | |
2513 | p->se.avg_overlap = 0; | |
2514 | ||
2515 | #ifdef CONFIG_SCHEDSTATS | |
2516 | p->se.wait_start = 0; | |
2517 | p->se.sum_sleep_runtime = 0; | |
2518 | p->se.sleep_start = 0; | |
2519 | p->se.block_start = 0; | |
2520 | p->se.sleep_max = 0; | |
2521 | p->se.block_max = 0; | |
2522 | p->se.exec_max = 0; | |
2523 | p->se.slice_max = 0; | |
2524 | p->se.wait_max = 0; | |
2525 | #endif | |
2526 | ||
2527 | INIT_LIST_HEAD(&p->rt.run_list); | |
2528 | p->se.on_rq = 0; | |
2529 | INIT_LIST_HEAD(&p->se.group_node); | |
2530 | ||
2531 | #ifdef CONFIG_PREEMPT_NOTIFIERS | |
2532 | INIT_HLIST_HEAD(&p->preempt_notifiers); | |
2533 | #endif | |
2534 | ||
2535 | /* | |
2536 | * We mark the process as running here, but have not actually | |
2537 | * inserted it onto the runqueue yet. This guarantees that | |
2538 | * nobody will actually run it, and a signal or other external | |
2539 | * event cannot wake it up and insert it on the runqueue either. | |
2540 | */ | |
2541 | p->state = TASK_RUNNING; | |
2542 | } | |
2543 | ||
2544 | /* | |
2545 | * fork()/clone()-time setup: | |
2546 | */ | |
2547 | void sched_fork(struct task_struct *p, int clone_flags) | |
2548 | { | |
2549 | int cpu = get_cpu(); | |
2550 | ||
2551 | __sched_fork(p); | |
2552 | ||
2553 | #ifdef CONFIG_SMP | |
2554 | cpu = sched_balance_self(cpu, SD_BALANCE_FORK); | |
2555 | #endif | |
2556 | set_task_cpu(p, cpu); | |
2557 | ||
2558 | /* | |
2559 | * Make sure we do not leak PI boosting priority to the child: | |
2560 | */ | |
2561 | p->prio = current->normal_prio; | |
2562 | if (!rt_prio(p->prio)) | |
2563 | p->sched_class = &fair_sched_class; | |
2564 | ||
2565 | #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) | |
2566 | if (likely(sched_info_on())) | |
2567 | memset(&p->sched_info, 0, sizeof(p->sched_info)); | |
2568 | #endif | |
2569 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) | |
2570 | p->oncpu = 0; | |
2571 | #endif | |
2572 | #ifdef CONFIG_PREEMPT | |
2573 | /* Want to start with kernel preemption disabled. */ | |
2574 | task_thread_info(p)->preempt_count = 1; | |
2575 | #endif | |
2576 | put_cpu(); | |
2577 | } | |
2578 | ||
2579 | /* | |
2580 | * wake_up_new_task - wake up a newly created task for the first time. | |
2581 | * | |
2582 | * This function will do some initial scheduler statistics housekeeping | |
2583 | * that must be done for every newly created context, then puts the task | |
2584 | * on the runqueue and wakes it. | |
2585 | */ | |
2586 | void wake_up_new_task(struct task_struct *p, unsigned long clone_flags) | |
2587 | { | |
2588 | unsigned long flags; | |
2589 | struct rq *rq; | |
2590 | ||
2591 | rq = task_rq_lock(p, &flags); | |
2592 | BUG_ON(p->state != TASK_RUNNING); | |
2593 | update_rq_clock(rq); | |
2594 | ||
2595 | p->prio = effective_prio(p); | |
2596 | ||
2597 | if (!p->sched_class->task_new || !current->se.on_rq) { | |
2598 | activate_task(rq, p, 0); | |
2599 | } else { | |
2600 | /* | |
2601 | * Let the scheduling class do new task startup | |
2602 | * management (if any): | |
2603 | */ | |
2604 | p->sched_class->task_new(rq, p); | |
2605 | inc_nr_running(rq); | |
2606 | } | |
2607 | check_preempt_curr(rq, p); | |
2608 | #ifdef CONFIG_SMP | |
2609 | if (p->sched_class->task_wake_up) | |
2610 | p->sched_class->task_wake_up(rq, p); | |
2611 | #endif | |
2612 | task_rq_unlock(rq, &flags); | |
2613 | } | |
2614 | ||
2615 | #ifdef CONFIG_PREEMPT_NOTIFIERS | |
2616 | ||
2617 | /** | |
2618 | * preempt_notifier_register - tell me when current is being being preempted & rescheduled | |
2619 | * @notifier: notifier struct to register | |
2620 | */ | |
2621 | void preempt_notifier_register(struct preempt_notifier *notifier) | |
2622 | { | |
2623 | hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); | |
2624 | } | |
2625 | EXPORT_SYMBOL_GPL(preempt_notifier_register); | |
2626 | ||
2627 | /** | |
2628 | * preempt_notifier_unregister - no longer interested in preemption notifications | |
2629 | * @notifier: notifier struct to unregister | |
2630 | * | |
2631 | * This is safe to call from within a preemption notifier. | |
2632 | */ | |
2633 | void preempt_notifier_unregister(struct preempt_notifier *notifier) | |
2634 | { | |
2635 | hlist_del(¬ifier->link); | |
2636 | } | |
2637 | EXPORT_SYMBOL_GPL(preempt_notifier_unregister); | |
2638 | ||
2639 | static void fire_sched_in_preempt_notifiers(struct task_struct *curr) | |
2640 | { | |
2641 | struct preempt_notifier *notifier; | |
2642 | struct hlist_node *node; | |
2643 | ||
2644 | hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) | |
2645 | notifier->ops->sched_in(notifier, raw_smp_processor_id()); | |
2646 | } | |
2647 | ||
2648 | static void | |
2649 | fire_sched_out_preempt_notifiers(struct task_struct *curr, | |
2650 | struct task_struct *next) | |
2651 | { | |
2652 | struct preempt_notifier *notifier; | |
2653 | struct hlist_node *node; | |
2654 | ||
2655 | hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) | |
2656 | notifier->ops->sched_out(notifier, next); | |
2657 | } | |
2658 | ||
2659 | #else | |
2660 | ||
2661 | static void fire_sched_in_preempt_notifiers(struct task_struct *curr) | |
2662 | { | |
2663 | } | |
2664 | ||
2665 | static void | |
2666 | fire_sched_out_preempt_notifiers(struct task_struct *curr, | |
2667 | struct task_struct *next) | |
2668 | { | |
2669 | } | |
2670 | ||
2671 | #endif | |
2672 | ||
2673 | /** | |
2674 | * prepare_task_switch - prepare to switch tasks | |
2675 | * @rq: the runqueue preparing to switch | |
2676 | * @prev: the current task that is being switched out | |
2677 | * @next: the task we are going to switch to. | |
2678 | * | |
2679 | * This is called with the rq lock held and interrupts off. It must | |
2680 | * be paired with a subsequent finish_task_switch after the context | |
2681 | * switch. | |
2682 | * | |
2683 | * prepare_task_switch sets up locking and calls architecture specific | |
2684 | * hooks. | |
2685 | */ | |
2686 | static inline void | |
2687 | prepare_task_switch(struct rq *rq, struct task_struct *prev, | |
2688 | struct task_struct *next) | |
2689 | { | |
2690 | fire_sched_out_preempt_notifiers(prev, next); | |
2691 | prepare_lock_switch(rq, next); | |
2692 | prepare_arch_switch(next); | |
2693 | } | |
2694 | ||
2695 | /** | |
2696 | * finish_task_switch - clean up after a task-switch | |
2697 | * @rq: runqueue associated with task-switch | |
2698 | * @prev: the thread we just switched away from. | |
2699 | * | |
2700 | * finish_task_switch must be called after the context switch, paired | |
2701 | * with a prepare_task_switch call before the context switch. | |
2702 | * finish_task_switch will reconcile locking set up by prepare_task_switch, | |
2703 | * and do any other architecture-specific cleanup actions. | |
2704 | * | |
2705 | * Note that we may have delayed dropping an mm in context_switch(). If | |
2706 | * so, we finish that here outside of the runqueue lock. (Doing it | |
2707 | * with the lock held can cause deadlocks; see schedule() for | |
2708 | * details.) | |
2709 | */ | |
2710 | static void finish_task_switch(struct rq *rq, struct task_struct *prev) | |
2711 | __releases(rq->lock) | |
2712 | { | |
2713 | struct mm_struct *mm = rq->prev_mm; | |
2714 | long prev_state; | |
2715 | ||
2716 | rq->prev_mm = NULL; | |
2717 | ||
2718 | /* | |
2719 | * A task struct has one reference for the use as "current". | |
2720 | * If a task dies, then it sets TASK_DEAD in tsk->state and calls | |
2721 | * schedule one last time. The schedule call will never return, and | |
2722 | * the scheduled task must drop that reference. | |
2723 | * The test for TASK_DEAD must occur while the runqueue locks are | |
2724 | * still held, otherwise prev could be scheduled on another cpu, die | |
2725 | * there before we look at prev->state, and then the reference would | |
2726 | * be dropped twice. | |
2727 | * Manfred Spraul <manfred@colorfullife.com> | |
2728 | */ | |
2729 | prev_state = prev->state; | |
2730 | finish_arch_switch(prev); | |
2731 | finish_lock_switch(rq, prev); | |
2732 | #ifdef CONFIG_SMP | |
2733 | if (current->sched_class->post_schedule) | |
2734 | current->sched_class->post_schedule(rq); | |
2735 | #endif | |
2736 | ||
2737 | fire_sched_in_preempt_notifiers(current); | |
2738 | if (mm) | |
2739 | mmdrop(mm); | |
2740 | if (unlikely(prev_state == TASK_DEAD)) { | |
2741 | /* | |
2742 | * Remove function-return probe instances associated with this | |
2743 | * task and put them back on the free list. | |
2744 | */ | |
2745 | kprobe_flush_task(prev); | |
2746 | put_task_struct(prev); | |
2747 | } | |
2748 | } | |
2749 | ||
2750 | /** | |
2751 | * schedule_tail - first thing a freshly forked thread must call. | |
2752 | * @prev: the thread we just switched away from. | |
2753 | */ | |
2754 | asmlinkage void schedule_tail(struct task_struct *prev) | |
2755 | __releases(rq->lock) | |
2756 | { | |
2757 | struct rq *rq = this_rq(); | |
2758 | ||
2759 | finish_task_switch(rq, prev); | |
2760 | #ifdef __ARCH_WANT_UNLOCKED_CTXSW | |
2761 | /* In this case, finish_task_switch does not reenable preemption */ | |
2762 | preempt_enable(); | |
2763 | #endif | |
2764 | if (current->set_child_tid) | |
2765 | put_user(task_pid_vnr(current), current->set_child_tid); | |
2766 | } | |
2767 | ||
2768 | /* | |
2769 | * context_switch - switch to the new MM and the new | |
2770 | * thread's register state. | |
2771 | */ | |
2772 | static inline void | |
2773 | context_switch(struct rq *rq, struct task_struct *prev, | |
2774 | struct task_struct *next) | |
2775 | { | |
2776 | struct mm_struct *mm, *oldmm; | |
2777 | ||
2778 | prepare_task_switch(rq, prev, next); | |
2779 | mm = next->mm; | |
2780 | oldmm = prev->active_mm; | |
2781 | /* | |
2782 | * For paravirt, this is coupled with an exit in switch_to to | |
2783 | * combine the page table reload and the switch backend into | |
2784 | * one hypercall. | |
2785 | */ | |
2786 | arch_enter_lazy_cpu_mode(); | |
2787 | ||
2788 | if (unlikely(!mm)) { | |
2789 | next->active_mm = oldmm; | |
2790 | atomic_inc(&oldmm->mm_count); | |
2791 | enter_lazy_tlb(oldmm, next); | |
2792 | } else | |
2793 | switch_mm(oldmm, mm, next); | |
2794 | ||
2795 | if (unlikely(!prev->mm)) { | |
2796 | prev->active_mm = NULL; | |
2797 | rq->prev_mm = oldmm; | |
2798 | } | |
2799 | /* | |
2800 | * Since the runqueue lock will be released by the next | |
2801 | * task (which is an invalid locking op but in the case | |
2802 | * of the scheduler it's an obvious special-case), so we | |
2803 | * do an early lockdep release here: | |
2804 | */ | |
2805 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW | |
2806 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); | |
2807 | #endif | |
2808 | ||
2809 | /* Here we just switch the register state and the stack. */ | |
2810 | switch_to(prev, next, prev); | |
2811 | ||
2812 | barrier(); | |
2813 | /* | |
2814 | * this_rq must be evaluated again because prev may have moved | |
2815 | * CPUs since it called schedule(), thus the 'rq' on its stack | |
2816 | * frame will be invalid. | |
2817 | */ | |
2818 | finish_task_switch(this_rq(), prev); | |
2819 | } | |
2820 | ||
2821 | /* | |
2822 | * nr_running, nr_uninterruptible and nr_context_switches: | |
2823 | * | |
2824 | * externally visible scheduler statistics: current number of runnable | |
2825 | * threads, current number of uninterruptible-sleeping threads, total | |
2826 | * number of context switches performed since bootup. | |
2827 | */ | |
2828 | unsigned long nr_running(void) | |
2829 | { | |
2830 | unsigned long i, sum = 0; | |
2831 | ||
2832 | for_each_online_cpu(i) | |
2833 | sum += cpu_rq(i)->nr_running; | |
2834 | ||
2835 | return sum; | |
2836 | } | |
2837 | ||
2838 | unsigned long nr_uninterruptible(void) | |
2839 | { | |
2840 | unsigned long i, sum = 0; | |
2841 | ||
2842 | for_each_possible_cpu(i) | |
2843 | sum += cpu_rq(i)->nr_uninterruptible; | |
2844 | ||
2845 | /* | |
2846 | * Since we read the counters lockless, it might be slightly | |
2847 | * inaccurate. Do not allow it to go below zero though: | |
2848 | */ | |
2849 | if (unlikely((long)sum < 0)) | |
2850 | sum = 0; | |
2851 | ||
2852 | return sum; | |
2853 | } | |
2854 | ||
2855 | unsigned long long nr_context_switches(void) | |
2856 | { | |
2857 | int i; | |
2858 | unsigned long long sum = 0; | |
2859 | ||
2860 | for_each_possible_cpu(i) | |
2861 | sum += cpu_rq(i)->nr_switches; | |
2862 | ||
2863 | return sum; | |
2864 | } | |
2865 | ||
2866 | unsigned long nr_iowait(void) | |
2867 | { | |
2868 | unsigned long i, sum = 0; | |
2869 | ||
2870 | for_each_possible_cpu(i) | |
2871 | sum += atomic_read(&cpu_rq(i)->nr_iowait); | |
2872 | ||
2873 | return sum; | |
2874 | } | |
2875 | ||
2876 | unsigned long nr_active(void) | |
2877 | { | |
2878 | unsigned long i, running = 0, uninterruptible = 0; | |
2879 | ||
2880 | for_each_online_cpu(i) { | |
2881 | running += cpu_rq(i)->nr_running; | |
2882 | uninterruptible += cpu_rq(i)->nr_uninterruptible; | |
2883 | } | |
2884 | ||
2885 | if (unlikely((long)uninterruptible < 0)) | |
2886 | uninterruptible = 0; | |
2887 | ||
2888 | return running + uninterruptible; | |
2889 | } | |
2890 | ||
2891 | /* | |
2892 | * Update rq->cpu_load[] statistics. This function is usually called every | |
2893 | * scheduler tick (TICK_NSEC). | |
2894 | */ | |
2895 | static void update_cpu_load(struct rq *this_rq) | |
2896 | { | |
2897 | unsigned long this_load = this_rq->load.weight; | |
2898 | int i, scale; | |
2899 | ||
2900 | this_rq->nr_load_updates++; | |
2901 | ||
2902 | /* Update our load: */ | |
2903 | for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { | |
2904 | unsigned long old_load, new_load; | |
2905 | ||
2906 | /* scale is effectively 1 << i now, and >> i divides by scale */ | |
2907 | ||
2908 | old_load = this_rq->cpu_load[i]; | |
2909 | new_load = this_load; | |
2910 | /* | |
2911 | * Round up the averaging division if load is increasing. This | |
2912 | * prevents us from getting stuck on 9 if the load is 10, for | |
2913 | * example. | |
2914 | */ | |
2915 | if (new_load > old_load) | |
2916 | new_load += scale-1; | |
2917 | this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i; | |
2918 | } | |
2919 | } | |
2920 | ||
2921 | #ifdef CONFIG_SMP | |
2922 | ||
2923 | /* | |
2924 | * double_rq_lock - safely lock two runqueues | |
2925 | * | |
2926 | * Note this does not disable interrupts like task_rq_lock, | |
2927 | * you need to do so manually before calling. | |
2928 | */ | |
2929 | static void double_rq_lock(struct rq *rq1, struct rq *rq2) | |
2930 | __acquires(rq1->lock) | |
2931 | __acquires(rq2->lock) | |
2932 | { | |
2933 | BUG_ON(!irqs_disabled()); | |
2934 | if (rq1 == rq2) { | |
2935 | spin_lock(&rq1->lock); | |
2936 | __acquire(rq2->lock); /* Fake it out ;) */ | |
2937 | } else { | |
2938 | if (rq1 < rq2) { | |
2939 | spin_lock(&rq1->lock); | |
2940 | spin_lock(&rq2->lock); | |
2941 | } else { | |
2942 | spin_lock(&rq2->lock); | |
2943 | spin_lock(&rq1->lock); | |
2944 | } | |
2945 | } | |
2946 | update_rq_clock(rq1); | |
2947 | update_rq_clock(rq2); | |
2948 | } | |
2949 | ||
2950 | /* | |
2951 | * double_rq_unlock - safely unlock two runqueues | |
2952 | * | |
2953 | * Note this does not restore interrupts like task_rq_unlock, | |
2954 | * you need to do so manually after calling. | |
2955 | */ | |
2956 | static void double_rq_unlock(struct rq *rq1, struct rq *rq2) | |
2957 | __releases(rq1->lock) | |
2958 | __releases(rq2->lock) | |
2959 | { | |
2960 | spin_unlock(&rq1->lock); | |
2961 | if (rq1 != rq2) | |
2962 | spin_unlock(&rq2->lock); | |
2963 | else | |
2964 | __release(rq2->lock); | |
2965 | } | |
2966 | ||
2967 | /* | |
2968 | * double_lock_balance - lock the busiest runqueue, this_rq is locked already. | |
2969 | */ | |
2970 | static int double_lock_balance(struct rq *this_rq, struct rq *busiest) | |
2971 | __releases(this_rq->lock) | |
2972 | __acquires(busiest->lock) | |
2973 | __acquires(this_rq->lock) | |
2974 | { | |
2975 | int ret = 0; | |
2976 | ||
2977 | if (unlikely(!irqs_disabled())) { | |
2978 | /* printk() doesn't work good under rq->lock */ | |
2979 | spin_unlock(&this_rq->lock); | |
2980 | BUG_ON(1); | |
2981 | } | |
2982 | if (unlikely(!spin_trylock(&busiest->lock))) { | |
2983 | if (busiest < this_rq) { | |
2984 | spin_unlock(&this_rq->lock); | |
2985 | spin_lock(&busiest->lock); | |
2986 | spin_lock(&this_rq->lock); | |
2987 | ret = 1; | |
2988 | } else | |
2989 | spin_lock(&busiest->lock); | |
2990 | } | |
2991 | return ret; | |
2992 | } | |
2993 | ||
2994 | /* | |
2995 | * If dest_cpu is allowed for this process, migrate the task to it. | |
2996 | * This is accomplished by forcing the cpu_allowed mask to only | |
2997 | * allow dest_cpu, which will force the cpu onto dest_cpu. Then | |
2998 | * the cpu_allowed mask is restored. | |
2999 | */ | |
3000 | static void sched_migrate_task(struct task_struct *p, int dest_cpu) | |
3001 | { | |
3002 | struct migration_req req; | |
3003 | unsigned long flags; | |
3004 | struct rq *rq; | |
3005 | ||
3006 | rq = task_rq_lock(p, &flags); | |
3007 | if (!cpu_isset(dest_cpu, p->cpus_allowed) | |
3008 | || unlikely(cpu_is_offline(dest_cpu))) | |
3009 | goto out; | |
3010 | ||
3011 | /* force the process onto the specified CPU */ | |
3012 | if (migrate_task(p, dest_cpu, &req)) { | |
3013 | /* Need to wait for migration thread (might exit: take ref). */ | |
3014 | struct task_struct *mt = rq->migration_thread; | |
3015 | ||
3016 | get_task_struct(mt); | |
3017 | task_rq_unlock(rq, &flags); | |
3018 | wake_up_process(mt); | |
3019 | put_task_struct(mt); | |
3020 | wait_for_completion(&req.done); | |
3021 | ||
3022 | return; | |
3023 | } | |
3024 | out: | |
3025 | task_rq_unlock(rq, &flags); | |
3026 | } | |
3027 | ||
3028 | /* | |
3029 | * sched_exec - execve() is a valuable balancing opportunity, because at | |
3030 | * this point the task has the smallest effective memory and cache footprint. | |
3031 | */ | |
3032 | void sched_exec(void) | |
3033 | { | |
3034 | int new_cpu, this_cpu = get_cpu(); | |
3035 | new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC); | |
3036 | put_cpu(); | |
3037 | if (new_cpu != this_cpu) | |
3038 | sched_migrate_task(current, new_cpu); | |
3039 | } | |
3040 | ||
3041 | /* | |
3042 | * pull_task - move a task from a remote runqueue to the local runqueue. | |
3043 | * Both runqueues must be locked. | |
3044 | */ | |
3045 | static void pull_task(struct rq *src_rq, struct task_struct *p, | |
3046 | struct rq *this_rq, int this_cpu) | |
3047 | { | |
3048 | deactivate_task(src_rq, p, 0); | |
3049 | set_task_cpu(p, this_cpu); | |
3050 | activate_task(this_rq, p, 0); | |
3051 | /* | |
3052 | * Note that idle threads have a prio of MAX_PRIO, for this test | |
3053 | * to be always true for them. | |
3054 | */ | |
3055 | check_preempt_curr(this_rq, p); | |
3056 | } | |
3057 | ||
3058 | /* | |
3059 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
3060 | */ | |
3061 | static | |
3062 | int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu, | |
3063 | struct sched_domain *sd, enum cpu_idle_type idle, | |
3064 | int *all_pinned) | |
3065 | { | |
3066 | /* | |
3067 | * We do not migrate tasks that are: | |
3068 | * 1) running (obviously), or | |
3069 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
3070 | * 3) are cache-hot on their current CPU. | |
3071 | */ | |
3072 | if (!cpu_isset(this_cpu, p->cpus_allowed)) { | |
3073 | schedstat_inc(p, se.nr_failed_migrations_affine); | |
3074 | return 0; | |
3075 | } | |
3076 | *all_pinned = 0; | |
3077 | ||
3078 | if (task_running(rq, p)) { | |
3079 | schedstat_inc(p, se.nr_failed_migrations_running); | |
3080 | return 0; | |
3081 | } | |
3082 | ||
3083 | /* | |
3084 | * Aggressive migration if: | |
3085 | * 1) task is cache cold, or | |
3086 | * 2) too many balance attempts have failed. | |
3087 | */ | |
3088 | ||
3089 | if (!task_hot(p, rq->clock, sd) || | |
3090 | sd->nr_balance_failed > sd->cache_nice_tries) { | |
3091 | #ifdef CONFIG_SCHEDSTATS | |
3092 | if (task_hot(p, rq->clock, sd)) { | |
3093 | schedstat_inc(sd, lb_hot_gained[idle]); | |
3094 | schedstat_inc(p, se.nr_forced_migrations); | |
3095 | } | |
3096 | #endif | |
3097 | return 1; | |
3098 | } | |
3099 | ||
3100 | if (task_hot(p, rq->clock, sd)) { | |
3101 | schedstat_inc(p, se.nr_failed_migrations_hot); | |
3102 | return 0; | |
3103 | } | |
3104 | return 1; | |
3105 | } | |
3106 | ||
3107 | static unsigned long | |
3108 | balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
3109 | unsigned long max_load_move, struct sched_domain *sd, | |
3110 | enum cpu_idle_type idle, int *all_pinned, | |
3111 | int *this_best_prio, struct rq_iterator *iterator) | |
3112 | { | |
3113 | int loops = 0, pulled = 0, pinned = 0, skip_for_load; | |
3114 | struct task_struct *p; | |
3115 | long rem_load_move = max_load_move; | |
3116 | ||
3117 | if (max_load_move == 0) | |
3118 | goto out; | |
3119 | ||
3120 | pinned = 1; | |
3121 | ||
3122 | /* | |
3123 | * Start the load-balancing iterator: | |
3124 | */ | |
3125 | p = iterator->start(iterator->arg); | |
3126 | next: | |
3127 | if (!p || loops++ > sysctl_sched_nr_migrate) | |
3128 | goto out; | |
3129 | /* | |
3130 | * To help distribute high priority tasks across CPUs we don't | |
3131 | * skip a task if it will be the highest priority task (i.e. smallest | |
3132 | * prio value) on its new queue regardless of its load weight | |
3133 | */ | |
3134 | skip_for_load = (p->se.load.weight >> 1) > rem_load_move + | |
3135 | SCHED_LOAD_SCALE_FUZZ; | |
3136 | if ((skip_for_load && p->prio >= *this_best_prio) || | |
3137 | !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) { | |
3138 | p = iterator->next(iterator->arg); | |
3139 | goto next; | |
3140 | } | |
3141 | ||
3142 | pull_task(busiest, p, this_rq, this_cpu); | |
3143 | pulled++; | |
3144 | rem_load_move -= p->se.load.weight; | |
3145 | ||
3146 | /* | |
3147 | * We only want to steal up to the prescribed amount of weighted load. | |
3148 | */ | |
3149 | if (rem_load_move > 0) { | |
3150 | if (p->prio < *this_best_prio) | |
3151 | *this_best_prio = p->prio; | |
3152 | p = iterator->next(iterator->arg); | |
3153 | goto next; | |
3154 | } | |
3155 | out: | |
3156 | /* | |
3157 | * Right now, this is one of only two places pull_task() is called, | |
3158 | * so we can safely collect pull_task() stats here rather than | |
3159 | * inside pull_task(). | |
3160 | */ | |
3161 | schedstat_add(sd, lb_gained[idle], pulled); | |
3162 | ||
3163 | if (all_pinned) | |
3164 | *all_pinned = pinned; | |
3165 | ||
3166 | return max_load_move - rem_load_move; | |
3167 | } | |
3168 | ||
3169 | /* | |
3170 | * move_tasks tries to move up to max_load_move weighted load from busiest to | |
3171 | * this_rq, as part of a balancing operation within domain "sd". | |
3172 | * Returns 1 if successful and 0 otherwise. | |
3173 | * | |
3174 | * Called with both runqueues locked. | |
3175 | */ | |
3176 | static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
3177 | unsigned long max_load_move, | |
3178 | struct sched_domain *sd, enum cpu_idle_type idle, | |
3179 | int *all_pinned) | |
3180 | { | |
3181 | const struct sched_class *class = sched_class_highest; | |
3182 | unsigned long total_load_moved = 0; | |
3183 | int this_best_prio = this_rq->curr->prio; | |
3184 | ||
3185 | do { | |
3186 | total_load_moved += | |
3187 | class->load_balance(this_rq, this_cpu, busiest, | |
3188 | max_load_move - total_load_moved, | |
3189 | sd, idle, all_pinned, &this_best_prio); | |
3190 | class = class->next; | |
3191 | } while (class && max_load_move > total_load_moved); | |
3192 | ||
3193 | return total_load_moved > 0; | |
3194 | } | |
3195 | ||
3196 | static int | |
3197 | iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
3198 | struct sched_domain *sd, enum cpu_idle_type idle, | |
3199 | struct rq_iterator *iterator) | |
3200 | { | |
3201 | struct task_struct *p = iterator->start(iterator->arg); | |
3202 | int pinned = 0; | |
3203 | ||
3204 | while (p) { | |
3205 | if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) { | |
3206 | pull_task(busiest, p, this_rq, this_cpu); | |
3207 | /* | |
3208 | * Right now, this is only the second place pull_task() | |
3209 | * is called, so we can safely collect pull_task() | |
3210 | * stats here rather than inside pull_task(). | |
3211 | */ | |
3212 | schedstat_inc(sd, lb_gained[idle]); | |
3213 | ||
3214 | return 1; | |
3215 | } | |
3216 | p = iterator->next(iterator->arg); | |
3217 | } | |
3218 | ||
3219 | return 0; | |
3220 | } | |
3221 | ||
3222 | /* | |
3223 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
3224 | * part of active balancing operations within "domain". | |
3225 | * Returns 1 if successful and 0 otherwise. | |
3226 | * | |
3227 | * Called with both runqueues locked. | |
3228 | */ | |
3229 | static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
3230 | struct sched_domain *sd, enum cpu_idle_type idle) | |
3231 | { | |
3232 | const struct sched_class *class; | |
3233 | ||
3234 | for (class = sched_class_highest; class; class = class->next) | |
3235 | if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle)) | |
3236 | return 1; | |
3237 | ||
3238 | return 0; | |
3239 | } | |
3240 | ||
3241 | /* | |
3242 | * find_busiest_group finds and returns the busiest CPU group within the | |
3243 | * domain. It calculates and returns the amount of weighted load which | |
3244 | * should be moved to restore balance via the imbalance parameter. | |
3245 | */ | |
3246 | static struct sched_group * | |
3247 | find_busiest_group(struct sched_domain *sd, int this_cpu, | |
3248 | unsigned long *imbalance, enum cpu_idle_type idle, | |
3249 | int *sd_idle, const cpumask_t *cpus, int *balance) | |
3250 | { | |
3251 | struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups; | |
3252 | unsigned long max_load, avg_load, total_load, this_load, total_pwr; | |
3253 | unsigned long max_pull; | |
3254 | unsigned long busiest_load_per_task, busiest_nr_running; | |
3255 | unsigned long this_load_per_task, this_nr_running; | |
3256 | int load_idx, group_imb = 0; | |
3257 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
3258 | int power_savings_balance = 1; | |
3259 | unsigned long leader_nr_running = 0, min_load_per_task = 0; | |
3260 | unsigned long min_nr_running = ULONG_MAX; | |
3261 | struct sched_group *group_min = NULL, *group_leader = NULL; | |
3262 | #endif | |
3263 | ||
3264 | max_load = this_load = total_load = total_pwr = 0; | |
3265 | busiest_load_per_task = busiest_nr_running = 0; | |
3266 | this_load_per_task = this_nr_running = 0; | |
3267 | if (idle == CPU_NOT_IDLE) | |
3268 | load_idx = sd->busy_idx; | |
3269 | else if (idle == CPU_NEWLY_IDLE) | |
3270 | load_idx = sd->newidle_idx; | |
3271 | else | |
3272 | load_idx = sd->idle_idx; | |
3273 | ||
3274 | do { | |
3275 | unsigned long load, group_capacity, max_cpu_load, min_cpu_load; | |
3276 | int local_group; | |
3277 | int i; | |
3278 | int __group_imb = 0; | |
3279 | unsigned int balance_cpu = -1, first_idle_cpu = 0; | |
3280 | unsigned long sum_nr_running, sum_weighted_load; | |
3281 | ||
3282 | local_group = cpu_isset(this_cpu, group->cpumask); | |
3283 | ||
3284 | if (local_group) | |
3285 | balance_cpu = first_cpu(group->cpumask); | |
3286 | ||
3287 | /* Tally up the load of all CPUs in the group */ | |
3288 | sum_weighted_load = sum_nr_running = avg_load = 0; | |
3289 | max_cpu_load = 0; | |
3290 | min_cpu_load = ~0UL; | |
3291 | ||
3292 | for_each_cpu_mask(i, group->cpumask) { | |
3293 | struct rq *rq; | |
3294 | ||
3295 | if (!cpu_isset(i, *cpus)) | |
3296 | continue; | |
3297 | ||
3298 | rq = cpu_rq(i); | |
3299 | ||
3300 | if (*sd_idle && rq->nr_running) | |
3301 | *sd_idle = 0; | |
3302 | ||
3303 | /* Bias balancing toward cpus of our domain */ | |
3304 | if (local_group) { | |
3305 | if (idle_cpu(i) && !first_idle_cpu) { | |
3306 | first_idle_cpu = 1; | |
3307 | balance_cpu = i; | |
3308 | } | |
3309 | ||
3310 | load = target_load(i, load_idx); | |
3311 | } else { | |
3312 | load = source_load(i, load_idx); | |
3313 | if (load > max_cpu_load) | |
3314 | max_cpu_load = load; | |
3315 | if (min_cpu_load > load) | |
3316 | min_cpu_load = load; | |
3317 | } | |
3318 | ||
3319 | avg_load += load; | |
3320 | sum_nr_running += rq->nr_running; | |
3321 | sum_weighted_load += weighted_cpuload(i); | |
3322 | } | |
3323 | ||
3324 | /* | |
3325 | * First idle cpu or the first cpu(busiest) in this sched group | |
3326 | * is eligible for doing load balancing at this and above | |
3327 | * domains. In the newly idle case, we will allow all the cpu's | |
3328 | * to do the newly idle load balance. | |
3329 | */ | |
3330 | if (idle != CPU_NEWLY_IDLE && local_group && | |
3331 | balance_cpu != this_cpu && balance) { | |
3332 | *balance = 0; | |
3333 | goto ret; | |
3334 | } | |
3335 | ||
3336 | total_load += avg_load; | |
3337 | total_pwr += group->__cpu_power; | |
3338 | ||
3339 | /* Adjust by relative CPU power of the group */ | |
3340 | avg_load = sg_div_cpu_power(group, | |
3341 | avg_load * SCHED_LOAD_SCALE); | |
3342 | ||
3343 | if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE) | |
3344 | __group_imb = 1; | |
3345 | ||
3346 | group_capacity = group->__cpu_power / SCHED_LOAD_SCALE; | |
3347 | ||
3348 | if (local_group) { | |
3349 | this_load = avg_load; | |
3350 | this = group; | |
3351 | this_nr_running = sum_nr_running; | |
3352 | this_load_per_task = sum_weighted_load; | |
3353 | } else if (avg_load > max_load && | |
3354 | (sum_nr_running > group_capacity || __group_imb)) { | |
3355 | max_load = avg_load; | |
3356 | busiest = group; | |
3357 | busiest_nr_running = sum_nr_running; | |
3358 | busiest_load_per_task = sum_weighted_load; | |
3359 | group_imb = __group_imb; | |
3360 | } | |
3361 | ||
3362 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
3363 | /* | |
3364 | * Busy processors will not participate in power savings | |
3365 | * balance. | |
3366 | */ | |
3367 | if (idle == CPU_NOT_IDLE || | |
3368 | !(sd->flags & SD_POWERSAVINGS_BALANCE)) | |
3369 | goto group_next; | |
3370 | ||
3371 | /* | |
3372 | * If the local group is idle or completely loaded | |
3373 | * no need to do power savings balance at this domain | |
3374 | */ | |
3375 | if (local_group && (this_nr_running >= group_capacity || | |
3376 | !this_nr_running)) | |
3377 | power_savings_balance = 0; | |
3378 | ||
3379 | /* | |
3380 | * If a group is already running at full capacity or idle, | |
3381 | * don't include that group in power savings calculations | |
3382 | */ | |
3383 | if (!power_savings_balance || sum_nr_running >= group_capacity | |
3384 | || !sum_nr_running) | |
3385 | goto group_next; | |
3386 | ||
3387 | /* | |
3388 | * Calculate the group which has the least non-idle load. | |
3389 | * This is the group from where we need to pick up the load | |
3390 | * for saving power | |
3391 | */ | |
3392 | if ((sum_nr_running < min_nr_running) || | |
3393 | (sum_nr_running == min_nr_running && | |
3394 | first_cpu(group->cpumask) < | |
3395 | first_cpu(group_min->cpumask))) { | |
3396 | group_min = group; | |
3397 | min_nr_running = sum_nr_running; | |
3398 | min_load_per_task = sum_weighted_load / | |
3399 | sum_nr_running; | |
3400 | } | |
3401 | ||
3402 | /* | |
3403 | * Calculate the group which is almost near its | |
3404 | * capacity but still has some space to pick up some load | |
3405 | * from other group and save more power | |
3406 | */ | |
3407 | if (sum_nr_running <= group_capacity - 1) { | |
3408 | if (sum_nr_running > leader_nr_running || | |
3409 | (sum_nr_running == leader_nr_running && | |
3410 | first_cpu(group->cpumask) > | |
3411 | first_cpu(group_leader->cpumask))) { | |
3412 | group_leader = group; | |
3413 | leader_nr_running = sum_nr_running; | |
3414 | } | |
3415 | } | |
3416 | group_next: | |
3417 | #endif | |
3418 | group = group->next; | |
3419 | } while (group != sd->groups); | |
3420 | ||
3421 | if (!busiest || this_load >= max_load || busiest_nr_running == 0) | |
3422 | goto out_balanced; | |
3423 | ||
3424 | avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr; | |
3425 | ||
3426 | if (this_load >= avg_load || | |
3427 | 100*max_load <= sd->imbalance_pct*this_load) | |
3428 | goto out_balanced; | |
3429 | ||
3430 | busiest_load_per_task /= busiest_nr_running; | |
3431 | if (group_imb) | |
3432 | busiest_load_per_task = min(busiest_load_per_task, avg_load); | |
3433 | ||
3434 | /* | |
3435 | * We're trying to get all the cpus to the average_load, so we don't | |
3436 | * want to push ourselves above the average load, nor do we wish to | |
3437 | * reduce the max loaded cpu below the average load, as either of these | |
3438 | * actions would just result in more rebalancing later, and ping-pong | |
3439 | * tasks around. Thus we look for the minimum possible imbalance. | |
3440 | * Negative imbalances (*we* are more loaded than anyone else) will | |
3441 | * be counted as no imbalance for these purposes -- we can't fix that | |
3442 | * by pulling tasks to us. Be careful of negative numbers as they'll | |
3443 | * appear as very large values with unsigned longs. | |
3444 | */ | |
3445 | if (max_load <= busiest_load_per_task) | |
3446 | goto out_balanced; | |
3447 | ||
3448 | /* | |
3449 | * In the presence of smp nice balancing, certain scenarios can have | |
3450 | * max load less than avg load(as we skip the groups at or below | |
3451 | * its cpu_power, while calculating max_load..) | |
3452 | */ | |
3453 | if (max_load < avg_load) { | |
3454 | *imbalance = 0; | |
3455 | goto small_imbalance; | |
3456 | } | |
3457 | ||
3458 | /* Don't want to pull so many tasks that a group would go idle */ | |
3459 | max_pull = min(max_load - avg_load, max_load - busiest_load_per_task); | |
3460 | ||
3461 | /* How much load to actually move to equalise the imbalance */ | |
3462 | *imbalance = min(max_pull * busiest->__cpu_power, | |
3463 | (avg_load - this_load) * this->__cpu_power) | |
3464 | / SCHED_LOAD_SCALE; | |
3465 | ||
3466 | /* | |
3467 | * if *imbalance is less than the average load per runnable task | |
3468 | * there is no gaurantee that any tasks will be moved so we'll have | |
3469 | * a think about bumping its value to force at least one task to be | |
3470 | * moved | |
3471 | */ | |
3472 | if (*imbalance < busiest_load_per_task) { | |
3473 | unsigned long tmp, pwr_now, pwr_move; | |
3474 | unsigned int imbn; | |
3475 | ||
3476 | small_imbalance: | |
3477 | pwr_move = pwr_now = 0; | |
3478 | imbn = 2; | |
3479 | if (this_nr_running) { | |
3480 | this_load_per_task /= this_nr_running; | |
3481 | if (busiest_load_per_task > this_load_per_task) | |
3482 | imbn = 1; | |
3483 | } else | |
3484 | this_load_per_task = SCHED_LOAD_SCALE; | |
3485 | ||
3486 | if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >= | |
3487 | busiest_load_per_task * imbn) { | |
3488 | *imbalance = busiest_load_per_task; | |
3489 | return busiest; | |
3490 | } | |
3491 | ||
3492 | /* | |
3493 | * OK, we don't have enough imbalance to justify moving tasks, | |
3494 | * however we may be able to increase total CPU power used by | |
3495 | * moving them. | |
3496 | */ | |
3497 | ||
3498 | pwr_now += busiest->__cpu_power * | |
3499 | min(busiest_load_per_task, max_load); | |
3500 | pwr_now += this->__cpu_power * | |
3501 | min(this_load_per_task, this_load); | |
3502 | pwr_now /= SCHED_LOAD_SCALE; | |
3503 | ||
3504 | /* Amount of load we'd subtract */ | |
3505 | tmp = sg_div_cpu_power(busiest, | |
3506 | busiest_load_per_task * SCHED_LOAD_SCALE); | |
3507 | if (max_load > tmp) | |
3508 | pwr_move += busiest->__cpu_power * | |
3509 | min(busiest_load_per_task, max_load - tmp); | |
3510 | ||
3511 | /* Amount of load we'd add */ | |
3512 | if (max_load * busiest->__cpu_power < | |
3513 | busiest_load_per_task * SCHED_LOAD_SCALE) | |
3514 | tmp = sg_div_cpu_power(this, | |
3515 | max_load * busiest->__cpu_power); | |
3516 | else | |
3517 | tmp = sg_div_cpu_power(this, | |
3518 | busiest_load_per_task * SCHED_LOAD_SCALE); | |
3519 | pwr_move += this->__cpu_power * | |
3520 | min(this_load_per_task, this_load + tmp); | |
3521 | pwr_move /= SCHED_LOAD_SCALE; | |
3522 | ||
3523 | /* Move if we gain throughput */ | |
3524 | if (pwr_move > pwr_now) | |
3525 | *imbalance = busiest_load_per_task; | |
3526 | } | |
3527 | ||
3528 | return busiest; | |
3529 | ||
3530 | out_balanced: | |
3531 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
3532 | if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE)) | |
3533 | goto ret; | |
3534 | ||
3535 | if (this == group_leader && group_leader != group_min) { | |
3536 | *imbalance = min_load_per_task; | |
3537 | return group_min; | |
3538 | } | |
3539 | #endif | |
3540 | ret: | |
3541 | *imbalance = 0; | |
3542 | return NULL; | |
3543 | } | |
3544 | ||
3545 | /* | |
3546 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
3547 | */ | |
3548 | static struct rq * | |
3549 | find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle, | |
3550 | unsigned long imbalance, const cpumask_t *cpus) | |
3551 | { | |
3552 | struct rq *busiest = NULL, *rq; | |
3553 | unsigned long max_load = 0; | |
3554 | int i; | |
3555 | ||
3556 | for_each_cpu_mask(i, group->cpumask) { | |
3557 | unsigned long wl; | |
3558 | ||
3559 | if (!cpu_isset(i, *cpus)) | |
3560 | continue; | |
3561 | ||
3562 | rq = cpu_rq(i); | |
3563 | wl = weighted_cpuload(i); | |
3564 | ||
3565 | if (rq->nr_running == 1 && wl > imbalance) | |
3566 | continue; | |
3567 | ||
3568 | if (wl > max_load) { | |
3569 | max_load = wl; | |
3570 | busiest = rq; | |
3571 | } | |
3572 | } | |
3573 | ||
3574 | return busiest; | |
3575 | } | |
3576 | ||
3577 | /* | |
3578 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
3579 | * so long as it is large enough. | |
3580 | */ | |
3581 | #define MAX_PINNED_INTERVAL 512 | |
3582 | ||
3583 | /* | |
3584 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
3585 | * tasks if there is an imbalance. | |
3586 | */ | |
3587 | static int load_balance(int this_cpu, struct rq *this_rq, | |
3588 | struct sched_domain *sd, enum cpu_idle_type idle, | |
3589 | int *balance, cpumask_t *cpus) | |
3590 | { | |
3591 | int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0; | |
3592 | struct sched_group *group; | |
3593 | unsigned long imbalance; | |
3594 | struct rq *busiest; | |
3595 | unsigned long flags; | |
3596 | int unlock_aggregate; | |
3597 | ||
3598 | cpus_setall(*cpus); | |
3599 | ||
3600 | unlock_aggregate = get_aggregate(sd); | |
3601 | ||
3602 | /* | |
3603 | * When power savings policy is enabled for the parent domain, idle | |
3604 | * sibling can pick up load irrespective of busy siblings. In this case, | |
3605 | * let the state of idle sibling percolate up as CPU_IDLE, instead of | |
3606 | * portraying it as CPU_NOT_IDLE. | |
3607 | */ | |
3608 | if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER && | |
3609 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
3610 | sd_idle = 1; | |
3611 | ||
3612 | schedstat_inc(sd, lb_count[idle]); | |
3613 | ||
3614 | redo: | |
3615 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle, | |
3616 | cpus, balance); | |
3617 | ||
3618 | if (*balance == 0) | |
3619 | goto out_balanced; | |
3620 | ||
3621 | if (!group) { | |
3622 | schedstat_inc(sd, lb_nobusyg[idle]); | |
3623 | goto out_balanced; | |
3624 | } | |
3625 | ||
3626 | busiest = find_busiest_queue(group, idle, imbalance, cpus); | |
3627 | if (!busiest) { | |
3628 | schedstat_inc(sd, lb_nobusyq[idle]); | |
3629 | goto out_balanced; | |
3630 | } | |
3631 | ||
3632 | BUG_ON(busiest == this_rq); | |
3633 | ||
3634 | schedstat_add(sd, lb_imbalance[idle], imbalance); | |
3635 | ||
3636 | ld_moved = 0; | |
3637 | if (busiest->nr_running > 1) { | |
3638 | /* | |
3639 | * Attempt to move tasks. If find_busiest_group has found | |
3640 | * an imbalance but busiest->nr_running <= 1, the group is | |
3641 | * still unbalanced. ld_moved simply stays zero, so it is | |
3642 | * correctly treated as an imbalance. | |
3643 | */ | |
3644 | local_irq_save(flags); | |
3645 | double_rq_lock(this_rq, busiest); | |
3646 | ld_moved = move_tasks(this_rq, this_cpu, busiest, | |
3647 | imbalance, sd, idle, &all_pinned); | |
3648 | double_rq_unlock(this_rq, busiest); | |
3649 | local_irq_restore(flags); | |
3650 | ||
3651 | /* | |
3652 | * some other cpu did the load balance for us. | |
3653 | */ | |
3654 | if (ld_moved && this_cpu != smp_processor_id()) | |
3655 | resched_cpu(this_cpu); | |
3656 | ||
3657 | /* All tasks on this runqueue were pinned by CPU affinity */ | |
3658 | if (unlikely(all_pinned)) { | |
3659 | cpu_clear(cpu_of(busiest), *cpus); | |
3660 | if (!cpus_empty(*cpus)) | |
3661 | goto redo; | |
3662 | goto out_balanced; | |
3663 | } | |
3664 | } | |
3665 | ||
3666 | if (!ld_moved) { | |
3667 | schedstat_inc(sd, lb_failed[idle]); | |
3668 | sd->nr_balance_failed++; | |
3669 | ||
3670 | if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) { | |
3671 | ||
3672 | spin_lock_irqsave(&busiest->lock, flags); | |
3673 | ||
3674 | /* don't kick the migration_thread, if the curr | |
3675 | * task on busiest cpu can't be moved to this_cpu | |
3676 | */ | |
3677 | if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) { | |
3678 | spin_unlock_irqrestore(&busiest->lock, flags); | |
3679 | all_pinned = 1; | |
3680 | goto out_one_pinned; | |
3681 | } | |
3682 | ||
3683 | if (!busiest->active_balance) { | |
3684 | busiest->active_balance = 1; | |
3685 | busiest->push_cpu = this_cpu; | |
3686 | active_balance = 1; | |
3687 | } | |
3688 | spin_unlock_irqrestore(&busiest->lock, flags); | |
3689 | if (active_balance) | |
3690 | wake_up_process(busiest->migration_thread); | |
3691 | ||
3692 | /* | |
3693 | * We've kicked active balancing, reset the failure | |
3694 | * counter. | |
3695 | */ | |
3696 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
3697 | } | |
3698 | } else | |
3699 | sd->nr_balance_failed = 0; | |
3700 | ||
3701 | if (likely(!active_balance)) { | |
3702 | /* We were unbalanced, so reset the balancing interval */ | |
3703 | sd->balance_interval = sd->min_interval; | |
3704 | } else { | |
3705 | /* | |
3706 | * If we've begun active balancing, start to back off. This | |
3707 | * case may not be covered by the all_pinned logic if there | |
3708 | * is only 1 task on the busy runqueue (because we don't call | |
3709 | * move_tasks). | |
3710 | */ | |
3711 | if (sd->balance_interval < sd->max_interval) | |
3712 | sd->balance_interval *= 2; | |
3713 | } | |
3714 | ||
3715 | if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER && | |
3716 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
3717 | ld_moved = -1; | |
3718 | ||
3719 | goto out; | |
3720 | ||
3721 | out_balanced: | |
3722 | schedstat_inc(sd, lb_balanced[idle]); | |
3723 | ||
3724 | sd->nr_balance_failed = 0; | |
3725 | ||
3726 | out_one_pinned: | |
3727 | /* tune up the balancing interval */ | |
3728 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || | |
3729 | (sd->balance_interval < sd->max_interval)) | |
3730 | sd->balance_interval *= 2; | |
3731 | ||
3732 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && | |
3733 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
3734 | ld_moved = -1; | |
3735 | else | |
3736 | ld_moved = 0; | |
3737 | out: | |
3738 | if (unlock_aggregate) | |
3739 | put_aggregate(sd); | |
3740 | return ld_moved; | |
3741 | } | |
3742 | ||
3743 | /* | |
3744 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
3745 | * tasks if there is an imbalance. | |
3746 | * | |
3747 | * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE). | |
3748 | * this_rq is locked. | |
3749 | */ | |
3750 | static int | |
3751 | load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd, | |
3752 | cpumask_t *cpus) | |
3753 | { | |
3754 | struct sched_group *group; | |
3755 | struct rq *busiest = NULL; | |
3756 | unsigned long imbalance; | |
3757 | int ld_moved = 0; | |
3758 | int sd_idle = 0; | |
3759 | int all_pinned = 0; | |
3760 | ||
3761 | cpus_setall(*cpus); | |
3762 | ||
3763 | /* | |
3764 | * When power savings policy is enabled for the parent domain, idle | |
3765 | * sibling can pick up load irrespective of busy siblings. In this case, | |
3766 | * let the state of idle sibling percolate up as IDLE, instead of | |
3767 | * portraying it as CPU_NOT_IDLE. | |
3768 | */ | |
3769 | if (sd->flags & SD_SHARE_CPUPOWER && | |
3770 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
3771 | sd_idle = 1; | |
3772 | ||
3773 | schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]); | |
3774 | redo: | |
3775 | group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE, | |
3776 | &sd_idle, cpus, NULL); | |
3777 | if (!group) { | |
3778 | schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]); | |
3779 | goto out_balanced; | |
3780 | } | |
3781 | ||
3782 | busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus); | |
3783 | if (!busiest) { | |
3784 | schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]); | |
3785 | goto out_balanced; | |
3786 | } | |
3787 | ||
3788 | BUG_ON(busiest == this_rq); | |
3789 | ||
3790 | schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance); | |
3791 | ||
3792 | ld_moved = 0; | |
3793 | if (busiest->nr_running > 1) { | |
3794 | /* Attempt to move tasks */ | |
3795 | double_lock_balance(this_rq, busiest); | |
3796 | /* this_rq->clock is already updated */ | |
3797 | update_rq_clock(busiest); | |
3798 | ld_moved = move_tasks(this_rq, this_cpu, busiest, | |
3799 | imbalance, sd, CPU_NEWLY_IDLE, | |
3800 | &all_pinned); | |
3801 | spin_unlock(&busiest->lock); | |
3802 | ||
3803 | if (unlikely(all_pinned)) { | |
3804 | cpu_clear(cpu_of(busiest), *cpus); | |
3805 | if (!cpus_empty(*cpus)) | |
3806 | goto redo; | |
3807 | } | |
3808 | } | |
3809 | ||
3810 | if (!ld_moved) { | |
3811 | schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]); | |
3812 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && | |
3813 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
3814 | return -1; | |
3815 | } else | |
3816 | sd->nr_balance_failed = 0; | |
3817 | ||
3818 | return ld_moved; | |
3819 | ||
3820 | out_balanced: | |
3821 | schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]); | |
3822 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && | |
3823 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
3824 | return -1; | |
3825 | sd->nr_balance_failed = 0; | |
3826 | ||
3827 | return 0; | |
3828 | } | |
3829 | ||
3830 | /* | |
3831 | * idle_balance is called by schedule() if this_cpu is about to become | |
3832 | * idle. Attempts to pull tasks from other CPUs. | |
3833 | */ | |
3834 | static void idle_balance(int this_cpu, struct rq *this_rq) | |
3835 | { | |
3836 | struct sched_domain *sd; | |
3837 | int pulled_task = -1; | |
3838 | unsigned long next_balance = jiffies + HZ; | |
3839 | cpumask_t tmpmask; | |
3840 | ||
3841 | for_each_domain(this_cpu, sd) { | |
3842 | unsigned long interval; | |
3843 | ||
3844 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
3845 | continue; | |
3846 | ||
3847 | if (sd->flags & SD_BALANCE_NEWIDLE) | |
3848 | /* If we've pulled tasks over stop searching: */ | |
3849 | pulled_task = load_balance_newidle(this_cpu, this_rq, | |
3850 | sd, &tmpmask); | |
3851 | ||
3852 | interval = msecs_to_jiffies(sd->balance_interval); | |
3853 | if (time_after(next_balance, sd->last_balance + interval)) | |
3854 | next_balance = sd->last_balance + interval; | |
3855 | if (pulled_task) | |
3856 | break; | |
3857 | } | |
3858 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { | |
3859 | /* | |
3860 | * We are going idle. next_balance may be set based on | |
3861 | * a busy processor. So reset next_balance. | |
3862 | */ | |
3863 | this_rq->next_balance = next_balance; | |
3864 | } | |
3865 | } | |
3866 | ||
3867 | /* | |
3868 | * active_load_balance is run by migration threads. It pushes running tasks | |
3869 | * off the busiest CPU onto idle CPUs. It requires at least 1 task to be | |
3870 | * running on each physical CPU where possible, and avoids physical / | |
3871 | * logical imbalances. | |
3872 | * | |
3873 | * Called with busiest_rq locked. | |
3874 | */ | |
3875 | static void active_load_balance(struct rq *busiest_rq, int busiest_cpu) | |
3876 | { | |
3877 | int target_cpu = busiest_rq->push_cpu; | |
3878 | struct sched_domain *sd; | |
3879 | struct rq *target_rq; | |
3880 | ||
3881 | /* Is there any task to move? */ | |
3882 | if (busiest_rq->nr_running <= 1) | |
3883 | return; | |
3884 | ||
3885 | target_rq = cpu_rq(target_cpu); | |
3886 | ||
3887 | /* | |
3888 | * This condition is "impossible", if it occurs | |
3889 | * we need to fix it. Originally reported by | |
3890 | * Bjorn Helgaas on a 128-cpu setup. | |
3891 | */ | |
3892 | BUG_ON(busiest_rq == target_rq); | |
3893 | ||
3894 | /* move a task from busiest_rq to target_rq */ | |
3895 | double_lock_balance(busiest_rq, target_rq); | |
3896 | update_rq_clock(busiest_rq); | |
3897 | update_rq_clock(target_rq); | |
3898 | ||
3899 | /* Search for an sd spanning us and the target CPU. */ | |
3900 | for_each_domain(target_cpu, sd) { | |
3901 | if ((sd->flags & SD_LOAD_BALANCE) && | |
3902 | cpu_isset(busiest_cpu, sd->span)) | |
3903 | break; | |
3904 | } | |
3905 | ||
3906 | if (likely(sd)) { | |
3907 | schedstat_inc(sd, alb_count); | |
3908 | ||
3909 | if (move_one_task(target_rq, target_cpu, busiest_rq, | |
3910 | sd, CPU_IDLE)) | |
3911 | schedstat_inc(sd, alb_pushed); | |
3912 | else | |
3913 | schedstat_inc(sd, alb_failed); | |
3914 | } | |
3915 | spin_unlock(&target_rq->lock); | |
3916 | } | |
3917 | ||
3918 | #ifdef CONFIG_NO_HZ | |
3919 | static struct { | |
3920 | atomic_t load_balancer; | |
3921 | cpumask_t cpu_mask; | |
3922 | } nohz ____cacheline_aligned = { | |
3923 | .load_balancer = ATOMIC_INIT(-1), | |
3924 | .cpu_mask = CPU_MASK_NONE, | |
3925 | }; | |
3926 | ||
3927 | /* | |
3928 | * This routine will try to nominate the ilb (idle load balancing) | |
3929 | * owner among the cpus whose ticks are stopped. ilb owner will do the idle | |
3930 | * load balancing on behalf of all those cpus. If all the cpus in the system | |
3931 | * go into this tickless mode, then there will be no ilb owner (as there is | |
3932 | * no need for one) and all the cpus will sleep till the next wakeup event | |
3933 | * arrives... | |
3934 | * | |
3935 | * For the ilb owner, tick is not stopped. And this tick will be used | |
3936 | * for idle load balancing. ilb owner will still be part of | |
3937 | * nohz.cpu_mask.. | |
3938 | * | |
3939 | * While stopping the tick, this cpu will become the ilb owner if there | |
3940 | * is no other owner. And will be the owner till that cpu becomes busy | |
3941 | * or if all cpus in the system stop their ticks at which point | |
3942 | * there is no need for ilb owner. | |
3943 | * | |
3944 | * When the ilb owner becomes busy, it nominates another owner, during the | |
3945 | * next busy scheduler_tick() | |
3946 | */ | |
3947 | int select_nohz_load_balancer(int stop_tick) | |
3948 | { | |
3949 | int cpu = smp_processor_id(); | |
3950 | ||
3951 | if (stop_tick) { | |
3952 | cpu_set(cpu, nohz.cpu_mask); | |
3953 | cpu_rq(cpu)->in_nohz_recently = 1; | |
3954 | ||
3955 | /* | |
3956 | * If we are going offline and still the leader, give up! | |
3957 | */ | |
3958 | if (cpu_is_offline(cpu) && | |
3959 | atomic_read(&nohz.load_balancer) == cpu) { | |
3960 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu) | |
3961 | BUG(); | |
3962 | return 0; | |
3963 | } | |
3964 | ||
3965 | /* time for ilb owner also to sleep */ | |
3966 | if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) { | |
3967 | if (atomic_read(&nohz.load_balancer) == cpu) | |
3968 | atomic_set(&nohz.load_balancer, -1); | |
3969 | return 0; | |
3970 | } | |
3971 | ||
3972 | if (atomic_read(&nohz.load_balancer) == -1) { | |
3973 | /* make me the ilb owner */ | |
3974 | if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1) | |
3975 | return 1; | |
3976 | } else if (atomic_read(&nohz.load_balancer) == cpu) | |
3977 | return 1; | |
3978 | } else { | |
3979 | if (!cpu_isset(cpu, nohz.cpu_mask)) | |
3980 | return 0; | |
3981 | ||
3982 | cpu_clear(cpu, nohz.cpu_mask); | |
3983 | ||
3984 | if (atomic_read(&nohz.load_balancer) == cpu) | |
3985 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu) | |
3986 | BUG(); | |
3987 | } | |
3988 | return 0; | |
3989 | } | |
3990 | #endif | |
3991 | ||
3992 | static DEFINE_SPINLOCK(balancing); | |
3993 | ||
3994 | /* | |
3995 | * It checks each scheduling domain to see if it is due to be balanced, | |
3996 | * and initiates a balancing operation if so. | |
3997 | * | |
3998 | * Balancing parameters are set up in arch_init_sched_domains. | |
3999 | */ | |
4000 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) | |
4001 | { | |
4002 | int balance = 1; | |
4003 | struct rq *rq = cpu_rq(cpu); | |
4004 | unsigned long interval; | |
4005 | struct sched_domain *sd; | |
4006 | /* Earliest time when we have to do rebalance again */ | |
4007 | unsigned long next_balance = jiffies + 60*HZ; | |
4008 | int update_next_balance = 0; | |
4009 | cpumask_t tmp; | |
4010 | ||
4011 | for_each_domain(cpu, sd) { | |
4012 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
4013 | continue; | |
4014 | ||
4015 | interval = sd->balance_interval; | |
4016 | if (idle != CPU_IDLE) | |
4017 | interval *= sd->busy_factor; | |
4018 | ||
4019 | /* scale ms to jiffies */ | |
4020 | interval = msecs_to_jiffies(interval); | |
4021 | if (unlikely(!interval)) | |
4022 | interval = 1; | |
4023 | if (interval > HZ*NR_CPUS/10) | |
4024 | interval = HZ*NR_CPUS/10; | |
4025 | ||
4026 | ||
4027 | if (sd->flags & SD_SERIALIZE) { | |
4028 | if (!spin_trylock(&balancing)) | |
4029 | goto out; | |
4030 | } | |
4031 | ||
4032 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
4033 | if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) { | |
4034 | /* | |
4035 | * We've pulled tasks over so either we're no | |
4036 | * longer idle, or one of our SMT siblings is | |
4037 | * not idle. | |
4038 | */ | |
4039 | idle = CPU_NOT_IDLE; | |
4040 | } | |
4041 | sd->last_balance = jiffies; | |
4042 | } | |
4043 | if (sd->flags & SD_SERIALIZE) | |
4044 | spin_unlock(&balancing); | |
4045 | out: | |
4046 | if (time_after(next_balance, sd->last_balance + interval)) { | |
4047 | next_balance = sd->last_balance + interval; | |
4048 | update_next_balance = 1; | |
4049 | } | |
4050 | ||
4051 | /* | |
4052 | * Stop the load balance at this level. There is another | |
4053 | * CPU in our sched group which is doing load balancing more | |
4054 | * actively. | |
4055 | */ | |
4056 | if (!balance) | |
4057 | break; | |
4058 | } | |
4059 | ||
4060 | /* | |
4061 | * next_balance will be updated only when there is a need. | |
4062 | * When the cpu is attached to null domain for ex, it will not be | |
4063 | * updated. | |
4064 | */ | |
4065 | if (likely(update_next_balance)) | |
4066 | rq->next_balance = next_balance; | |
4067 | } | |
4068 | ||
4069 | /* | |
4070 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
4071 | * In CONFIG_NO_HZ case, the idle load balance owner will do the | |
4072 | * rebalancing for all the cpus for whom scheduler ticks are stopped. | |
4073 | */ | |
4074 | static void run_rebalance_domains(struct softirq_action *h) | |
4075 | { | |
4076 | int this_cpu = smp_processor_id(); | |
4077 | struct rq *this_rq = cpu_rq(this_cpu); | |
4078 | enum cpu_idle_type idle = this_rq->idle_at_tick ? | |
4079 | CPU_IDLE : CPU_NOT_IDLE; | |
4080 | ||
4081 | rebalance_domains(this_cpu, idle); | |
4082 | ||
4083 | #ifdef CONFIG_NO_HZ | |
4084 | /* | |
4085 | * If this cpu is the owner for idle load balancing, then do the | |
4086 | * balancing on behalf of the other idle cpus whose ticks are | |
4087 | * stopped. | |
4088 | */ | |
4089 | if (this_rq->idle_at_tick && | |
4090 | atomic_read(&nohz.load_balancer) == this_cpu) { | |
4091 | cpumask_t cpus = nohz.cpu_mask; | |
4092 | struct rq *rq; | |
4093 | int balance_cpu; | |
4094 | ||
4095 | cpu_clear(this_cpu, cpus); | |
4096 | for_each_cpu_mask(balance_cpu, cpus) { | |
4097 | /* | |
4098 | * If this cpu gets work to do, stop the load balancing | |
4099 | * work being done for other cpus. Next load | |
4100 | * balancing owner will pick it up. | |
4101 | */ | |
4102 | if (need_resched()) | |
4103 | break; | |
4104 | ||
4105 | rebalance_domains(balance_cpu, CPU_IDLE); | |
4106 | ||
4107 | rq = cpu_rq(balance_cpu); | |
4108 | if (time_after(this_rq->next_balance, rq->next_balance)) | |
4109 | this_rq->next_balance = rq->next_balance; | |
4110 | } | |
4111 | } | |
4112 | #endif | |
4113 | } | |
4114 | ||
4115 | /* | |
4116 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
4117 | * | |
4118 | * In case of CONFIG_NO_HZ, this is the place where we nominate a new | |
4119 | * idle load balancing owner or decide to stop the periodic load balancing, | |
4120 | * if the whole system is idle. | |
4121 | */ | |
4122 | static inline void trigger_load_balance(struct rq *rq, int cpu) | |
4123 | { | |
4124 | #ifdef CONFIG_NO_HZ | |
4125 | /* | |
4126 | * If we were in the nohz mode recently and busy at the current | |
4127 | * scheduler tick, then check if we need to nominate new idle | |
4128 | * load balancer. | |
4129 | */ | |
4130 | if (rq->in_nohz_recently && !rq->idle_at_tick) { | |
4131 | rq->in_nohz_recently = 0; | |
4132 | ||
4133 | if (atomic_read(&nohz.load_balancer) == cpu) { | |
4134 | cpu_clear(cpu, nohz.cpu_mask); | |
4135 | atomic_set(&nohz.load_balancer, -1); | |
4136 | } | |
4137 | ||
4138 | if (atomic_read(&nohz.load_balancer) == -1) { | |
4139 | /* | |
4140 | * simple selection for now: Nominate the | |
4141 | * first cpu in the nohz list to be the next | |
4142 | * ilb owner. | |
4143 | * | |
4144 | * TBD: Traverse the sched domains and nominate | |
4145 | * the nearest cpu in the nohz.cpu_mask. | |
4146 | */ | |
4147 | int ilb = first_cpu(nohz.cpu_mask); | |
4148 | ||
4149 | if (ilb < nr_cpu_ids) | |
4150 | resched_cpu(ilb); | |
4151 | } | |
4152 | } | |
4153 | ||
4154 | /* | |
4155 | * If this cpu is idle and doing idle load balancing for all the | |
4156 | * cpus with ticks stopped, is it time for that to stop? | |
4157 | */ | |
4158 | if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu && | |
4159 | cpus_weight(nohz.cpu_mask) == num_online_cpus()) { | |
4160 | resched_cpu(cpu); | |
4161 | return; | |
4162 | } | |
4163 | ||
4164 | /* | |
4165 | * If this cpu is idle and the idle load balancing is done by | |
4166 | * someone else, then no need raise the SCHED_SOFTIRQ | |
4167 | */ | |
4168 | if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu && | |
4169 | cpu_isset(cpu, nohz.cpu_mask)) | |
4170 | return; | |
4171 | #endif | |
4172 | if (time_after_eq(jiffies, rq->next_balance)) | |
4173 | raise_softirq(SCHED_SOFTIRQ); | |
4174 | } | |
4175 | ||
4176 | #else /* CONFIG_SMP */ | |
4177 | ||
4178 | /* | |
4179 | * on UP we do not need to balance between CPUs: | |
4180 | */ | |
4181 | static inline void idle_balance(int cpu, struct rq *rq) | |
4182 | { | |
4183 | } | |
4184 | ||
4185 | #endif | |
4186 | ||
4187 | DEFINE_PER_CPU(struct kernel_stat, kstat); | |
4188 | ||
4189 | EXPORT_PER_CPU_SYMBOL(kstat); | |
4190 | ||
4191 | /* | |
4192 | * Return p->sum_exec_runtime plus any more ns on the sched_clock | |
4193 | * that have not yet been banked in case the task is currently running. | |
4194 | */ | |
4195 | unsigned long long task_sched_runtime(struct task_struct *p) | |
4196 | { | |
4197 | unsigned long flags; | |
4198 | u64 ns, delta_exec; | |
4199 | struct rq *rq; | |
4200 | ||
4201 | rq = task_rq_lock(p, &flags); | |
4202 | ns = p->se.sum_exec_runtime; | |
4203 | if (task_current(rq, p)) { | |
4204 | update_rq_clock(rq); | |
4205 | delta_exec = rq->clock - p->se.exec_start; | |
4206 | if ((s64)delta_exec > 0) | |
4207 | ns += delta_exec; | |
4208 | } | |
4209 | task_rq_unlock(rq, &flags); | |
4210 | ||
4211 | return ns; | |
4212 | } | |
4213 | ||
4214 | /* | |
4215 | * Account user cpu time to a process. | |
4216 | * @p: the process that the cpu time gets accounted to | |
4217 | * @cputime: the cpu time spent in user space since the last update | |
4218 | */ | |
4219 | void account_user_time(struct task_struct *p, cputime_t cputime) | |
4220 | { | |
4221 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
4222 | cputime64_t tmp; | |
4223 | ||
4224 | p->utime = cputime_add(p->utime, cputime); | |
4225 | ||
4226 | /* Add user time to cpustat. */ | |
4227 | tmp = cputime_to_cputime64(cputime); | |
4228 | if (TASK_NICE(p) > 0) | |
4229 | cpustat->nice = cputime64_add(cpustat->nice, tmp); | |
4230 | else | |
4231 | cpustat->user = cputime64_add(cpustat->user, tmp); | |
4232 | } | |
4233 | ||
4234 | /* | |
4235 | * Account guest cpu time to a process. | |
4236 | * @p: the process that the cpu time gets accounted to | |
4237 | * @cputime: the cpu time spent in virtual machine since the last update | |
4238 | */ | |
4239 | static void account_guest_time(struct task_struct *p, cputime_t cputime) | |
4240 | { | |
4241 | cputime64_t tmp; | |
4242 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
4243 | ||
4244 | tmp = cputime_to_cputime64(cputime); | |
4245 | ||
4246 | p->utime = cputime_add(p->utime, cputime); | |
4247 | p->gtime = cputime_add(p->gtime, cputime); | |
4248 | ||
4249 | cpustat->user = cputime64_add(cpustat->user, tmp); | |
4250 | cpustat->guest = cputime64_add(cpustat->guest, tmp); | |
4251 | } | |
4252 | ||
4253 | /* | |
4254 | * Account scaled user cpu time to a process. | |
4255 | * @p: the process that the cpu time gets accounted to | |
4256 | * @cputime: the cpu time spent in user space since the last update | |
4257 | */ | |
4258 | void account_user_time_scaled(struct task_struct *p, cputime_t cputime) | |
4259 | { | |
4260 | p->utimescaled = cputime_add(p->utimescaled, cputime); | |
4261 | } | |
4262 | ||
4263 | /* | |
4264 | * Account system cpu time to a process. | |
4265 | * @p: the process that the cpu time gets accounted to | |
4266 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
4267 | * @cputime: the cpu time spent in kernel space since the last update | |
4268 | */ | |
4269 | void account_system_time(struct task_struct *p, int hardirq_offset, | |
4270 | cputime_t cputime) | |
4271 | { | |
4272 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
4273 | struct rq *rq = this_rq(); | |
4274 | cputime64_t tmp; | |
4275 | ||
4276 | if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) | |
4277 | return account_guest_time(p, cputime); | |
4278 | ||
4279 | p->stime = cputime_add(p->stime, cputime); | |
4280 | ||
4281 | /* Add system time to cpustat. */ | |
4282 | tmp = cputime_to_cputime64(cputime); | |
4283 | if (hardirq_count() - hardirq_offset) | |
4284 | cpustat->irq = cputime64_add(cpustat->irq, tmp); | |
4285 | else if (softirq_count()) | |
4286 | cpustat->softirq = cputime64_add(cpustat->softirq, tmp); | |
4287 | else if (p != rq->idle) | |
4288 | cpustat->system = cputime64_add(cpustat->system, tmp); | |
4289 | else if (atomic_read(&rq->nr_iowait) > 0) | |
4290 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); | |
4291 | else | |
4292 | cpustat->idle = cputime64_add(cpustat->idle, tmp); | |
4293 | /* Account for system time used */ | |
4294 | acct_update_integrals(p); | |
4295 | } | |
4296 | ||
4297 | /* | |
4298 | * Account scaled system cpu time to a process. | |
4299 | * @p: the process that the cpu time gets accounted to | |
4300 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
4301 | * @cputime: the cpu time spent in kernel space since the last update | |
4302 | */ | |
4303 | void account_system_time_scaled(struct task_struct *p, cputime_t cputime) | |
4304 | { | |
4305 | p->stimescaled = cputime_add(p->stimescaled, cputime); | |
4306 | } | |
4307 | ||
4308 | /* | |
4309 | * Account for involuntary wait time. | |
4310 | * @p: the process from which the cpu time has been stolen | |
4311 | * @steal: the cpu time spent in involuntary wait | |
4312 | */ | |
4313 | void account_steal_time(struct task_struct *p, cputime_t steal) | |
4314 | { | |
4315 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
4316 | cputime64_t tmp = cputime_to_cputime64(steal); | |
4317 | struct rq *rq = this_rq(); | |
4318 | ||
4319 | if (p == rq->idle) { | |
4320 | p->stime = cputime_add(p->stime, steal); | |
4321 | if (atomic_read(&rq->nr_iowait) > 0) | |
4322 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); | |
4323 | else | |
4324 | cpustat->idle = cputime64_add(cpustat->idle, tmp); | |
4325 | } else | |
4326 | cpustat->steal = cputime64_add(cpustat->steal, tmp); | |
4327 | } | |
4328 | ||
4329 | /* | |
4330 | * This function gets called by the timer code, with HZ frequency. | |
4331 | * We call it with interrupts disabled. | |
4332 | * | |
4333 | * It also gets called by the fork code, when changing the parent's | |
4334 | * timeslices. | |
4335 | */ | |
4336 | void scheduler_tick(void) | |
4337 | { | |
4338 | int cpu = smp_processor_id(); | |
4339 | struct rq *rq = cpu_rq(cpu); | |
4340 | struct task_struct *curr = rq->curr; | |
4341 | u64 next_tick = rq->tick_timestamp + TICK_NSEC; | |
4342 | ||
4343 | spin_lock(&rq->lock); | |
4344 | __update_rq_clock(rq); | |
4345 | /* | |
4346 | * Let rq->clock advance by at least TICK_NSEC: | |
4347 | */ | |
4348 | if (unlikely(rq->clock < next_tick)) { | |
4349 | rq->clock = next_tick; | |
4350 | rq->clock_underflows++; | |
4351 | } | |
4352 | rq->tick_timestamp = rq->clock; | |
4353 | update_last_tick_seen(rq); | |
4354 | update_cpu_load(rq); | |
4355 | curr->sched_class->task_tick(rq, curr, 0); | |
4356 | spin_unlock(&rq->lock); | |
4357 | ||
4358 | #ifdef CONFIG_SMP | |
4359 | rq->idle_at_tick = idle_cpu(cpu); | |
4360 | trigger_load_balance(rq, cpu); | |
4361 | #endif | |
4362 | } | |
4363 | ||
4364 | #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT) | |
4365 | ||
4366 | void __kprobes add_preempt_count(int val) | |
4367 | { | |
4368 | /* | |
4369 | * Underflow? | |
4370 | */ | |
4371 | if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) | |
4372 | return; | |
4373 | preempt_count() += val; | |
4374 | /* | |
4375 | * Spinlock count overflowing soon? | |
4376 | */ | |
4377 | DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= | |
4378 | PREEMPT_MASK - 10); | |
4379 | } | |
4380 | EXPORT_SYMBOL(add_preempt_count); | |
4381 | ||
4382 | void __kprobes sub_preempt_count(int val) | |
4383 | { | |
4384 | /* | |
4385 | * Underflow? | |
4386 | */ | |
4387 | if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) | |
4388 | return; | |
4389 | /* | |
4390 | * Is the spinlock portion underflowing? | |
4391 | */ | |
4392 | if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && | |
4393 | !(preempt_count() & PREEMPT_MASK))) | |
4394 | return; | |
4395 | ||
4396 | preempt_count() -= val; | |
4397 | } | |
4398 | EXPORT_SYMBOL(sub_preempt_count); | |
4399 | ||
4400 | #endif | |
4401 | ||
4402 | /* | |
4403 | * Print scheduling while atomic bug: | |
4404 | */ | |
4405 | static noinline void __schedule_bug(struct task_struct *prev) | |
4406 | { | |
4407 | struct pt_regs *regs = get_irq_regs(); | |
4408 | ||
4409 | printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", | |
4410 | prev->comm, prev->pid, preempt_count()); | |
4411 | ||
4412 | debug_show_held_locks(prev); | |
4413 | if (irqs_disabled()) | |
4414 | print_irqtrace_events(prev); | |
4415 | ||
4416 | if (regs) | |
4417 | show_regs(regs); | |
4418 | else | |
4419 | dump_stack(); | |
4420 | } | |
4421 | ||
4422 | /* | |
4423 | * Various schedule()-time debugging checks and statistics: | |
4424 | */ | |
4425 | static inline void schedule_debug(struct task_struct *prev) | |
4426 | { | |
4427 | /* | |
4428 | * Test if we are atomic. Since do_exit() needs to call into | |
4429 | * schedule() atomically, we ignore that path for now. | |
4430 | * Otherwise, whine if we are scheduling when we should not be. | |
4431 | */ | |
4432 | if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state)) | |
4433 | __schedule_bug(prev); | |
4434 | ||
4435 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); | |
4436 | ||
4437 | schedstat_inc(this_rq(), sched_count); | |
4438 | #ifdef CONFIG_SCHEDSTATS | |
4439 | if (unlikely(prev->lock_depth >= 0)) { | |
4440 | schedstat_inc(this_rq(), bkl_count); | |
4441 | schedstat_inc(prev, sched_info.bkl_count); | |
4442 | } | |
4443 | #endif | |
4444 | } | |
4445 | ||
4446 | /* | |
4447 | * Pick up the highest-prio task: | |
4448 | */ | |
4449 | static inline struct task_struct * | |
4450 | pick_next_task(struct rq *rq, struct task_struct *prev) | |
4451 | { | |
4452 | const struct sched_class *class; | |
4453 | struct task_struct *p; | |
4454 | ||
4455 | /* | |
4456 | * Optimization: we know that if all tasks are in | |
4457 | * the fair class we can call that function directly: | |
4458 | */ | |
4459 | if (likely(rq->nr_running == rq->cfs.nr_running)) { | |
4460 | p = fair_sched_class.pick_next_task(rq); | |
4461 | if (likely(p)) | |
4462 | return p; | |
4463 | } | |
4464 | ||
4465 | class = sched_class_highest; | |
4466 | for ( ; ; ) { | |
4467 | p = class->pick_next_task(rq); | |
4468 | if (p) | |
4469 | return p; | |
4470 | /* | |
4471 | * Will never be NULL as the idle class always | |
4472 | * returns a non-NULL p: | |
4473 | */ | |
4474 | class = class->next; | |
4475 | } | |
4476 | } | |
4477 | ||
4478 | /* | |
4479 | * schedule() is the main scheduler function. | |
4480 | */ | |
4481 | asmlinkage void __sched schedule(void) | |
4482 | { | |
4483 | struct task_struct *prev, *next; | |
4484 | unsigned long *switch_count; | |
4485 | struct rq *rq; | |
4486 | int cpu; | |
4487 | ||
4488 | need_resched: | |
4489 | preempt_disable(); | |
4490 | cpu = smp_processor_id(); | |
4491 | rq = cpu_rq(cpu); | |
4492 | rcu_qsctr_inc(cpu); | |
4493 | prev = rq->curr; | |
4494 | switch_count = &prev->nivcsw; | |
4495 | ||
4496 | release_kernel_lock(prev); | |
4497 | need_resched_nonpreemptible: | |
4498 | ||
4499 | schedule_debug(prev); | |
4500 | ||
4501 | hrtick_clear(rq); | |
4502 | ||
4503 | /* | |
4504 | * Do the rq-clock update outside the rq lock: | |
4505 | */ | |
4506 | local_irq_disable(); | |
4507 | __update_rq_clock(rq); | |
4508 | spin_lock(&rq->lock); | |
4509 | clear_tsk_need_resched(prev); | |
4510 | ||
4511 | if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { | |
4512 | if (unlikely((prev->state & TASK_INTERRUPTIBLE) && | |
4513 | signal_pending(prev))) { | |
4514 | prev->state = TASK_RUNNING; | |
4515 | } else { | |
4516 | deactivate_task(rq, prev, 1); | |
4517 | } | |
4518 | switch_count = &prev->nvcsw; | |
4519 | } | |
4520 | ||
4521 | #ifdef CONFIG_SMP | |
4522 | if (prev->sched_class->pre_schedule) | |
4523 | prev->sched_class->pre_schedule(rq, prev); | |
4524 | #endif | |
4525 | ||
4526 | if (unlikely(!rq->nr_running)) | |
4527 | idle_balance(cpu, rq); | |
4528 | ||
4529 | prev->sched_class->put_prev_task(rq, prev); | |
4530 | next = pick_next_task(rq, prev); | |
4531 | ||
4532 | sched_info_switch(prev, next); | |
4533 | ||
4534 | if (likely(prev != next)) { | |
4535 | rq->nr_switches++; | |
4536 | rq->curr = next; | |
4537 | ++*switch_count; | |
4538 | ||
4539 | context_switch(rq, prev, next); /* unlocks the rq */ | |
4540 | /* | |
4541 | * the context switch might have flipped the stack from under | |
4542 | * us, hence refresh the local variables. | |
4543 | */ | |
4544 | cpu = smp_processor_id(); | |
4545 | rq = cpu_rq(cpu); | |
4546 | } else | |
4547 | spin_unlock_irq(&rq->lock); | |
4548 | ||
4549 | hrtick_set(rq); | |
4550 | ||
4551 | if (unlikely(reacquire_kernel_lock(current) < 0)) | |
4552 | goto need_resched_nonpreemptible; | |
4553 | ||
4554 | preempt_enable_no_resched(); | |
4555 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
4556 | goto need_resched; | |
4557 | } | |
4558 | EXPORT_SYMBOL(schedule); | |
4559 | ||
4560 | #ifdef CONFIG_PREEMPT | |
4561 | /* | |
4562 | * this is the entry point to schedule() from in-kernel preemption | |
4563 | * off of preempt_enable. Kernel preemptions off return from interrupt | |
4564 | * occur there and call schedule directly. | |
4565 | */ | |
4566 | asmlinkage void __sched preempt_schedule(void) | |
4567 | { | |
4568 | struct thread_info *ti = current_thread_info(); | |
4569 | struct task_struct *task = current; | |
4570 | int saved_lock_depth; | |
4571 | ||
4572 | /* | |
4573 | * If there is a non-zero preempt_count or interrupts are disabled, | |
4574 | * we do not want to preempt the current task. Just return.. | |
4575 | */ | |
4576 | if (likely(ti->preempt_count || irqs_disabled())) | |
4577 | return; | |
4578 | ||
4579 | do { | |
4580 | add_preempt_count(PREEMPT_ACTIVE); | |
4581 | ||
4582 | /* | |
4583 | * We keep the big kernel semaphore locked, but we | |
4584 | * clear ->lock_depth so that schedule() doesnt | |
4585 | * auto-release the semaphore: | |
4586 | */ | |
4587 | saved_lock_depth = task->lock_depth; | |
4588 | task->lock_depth = -1; | |
4589 | schedule(); | |
4590 | task->lock_depth = saved_lock_depth; | |
4591 | sub_preempt_count(PREEMPT_ACTIVE); | |
4592 | ||
4593 | /* | |
4594 | * Check again in case we missed a preemption opportunity | |
4595 | * between schedule and now. | |
4596 | */ | |
4597 | barrier(); | |
4598 | } while (unlikely(test_thread_flag(TIF_NEED_RESCHED))); | |
4599 | } | |
4600 | EXPORT_SYMBOL(preempt_schedule); | |
4601 | ||
4602 | /* | |
4603 | * this is the entry point to schedule() from kernel preemption | |
4604 | * off of irq context. | |
4605 | * Note, that this is called and return with irqs disabled. This will | |
4606 | * protect us against recursive calling from irq. | |
4607 | */ | |
4608 | asmlinkage void __sched preempt_schedule_irq(void) | |
4609 | { | |
4610 | struct thread_info *ti = current_thread_info(); | |
4611 | struct task_struct *task = current; | |
4612 | int saved_lock_depth; | |
4613 | ||
4614 | /* Catch callers which need to be fixed */ | |
4615 | BUG_ON(ti->preempt_count || !irqs_disabled()); | |
4616 | ||
4617 | do { | |
4618 | add_preempt_count(PREEMPT_ACTIVE); | |
4619 | ||
4620 | /* | |
4621 | * We keep the big kernel semaphore locked, but we | |
4622 | * clear ->lock_depth so that schedule() doesnt | |
4623 | * auto-release the semaphore: | |
4624 | */ | |
4625 | saved_lock_depth = task->lock_depth; | |
4626 | task->lock_depth = -1; | |
4627 | local_irq_enable(); | |
4628 | schedule(); | |
4629 | local_irq_disable(); | |
4630 | task->lock_depth = saved_lock_depth; | |
4631 | sub_preempt_count(PREEMPT_ACTIVE); | |
4632 | ||
4633 | /* | |
4634 | * Check again in case we missed a preemption opportunity | |
4635 | * between schedule and now. | |
4636 | */ | |
4637 | barrier(); | |
4638 | } while (unlikely(test_thread_flag(TIF_NEED_RESCHED))); | |
4639 | } | |
4640 | ||
4641 | #endif /* CONFIG_PREEMPT */ | |
4642 | ||
4643 | int default_wake_function(wait_queue_t *curr, unsigned mode, int sync, | |
4644 | void *key) | |
4645 | { | |
4646 | return try_to_wake_up(curr->private, mode, sync); | |
4647 | } | |
4648 | EXPORT_SYMBOL(default_wake_function); | |
4649 | ||
4650 | /* | |
4651 | * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just | |
4652 | * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve | |
4653 | * number) then we wake all the non-exclusive tasks and one exclusive task. | |
4654 | * | |
4655 | * There are circumstances in which we can try to wake a task which has already | |
4656 | * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns | |
4657 | * zero in this (rare) case, and we handle it by continuing to scan the queue. | |
4658 | */ | |
4659 | static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, | |
4660 | int nr_exclusive, int sync, void *key) | |
4661 | { | |
4662 | wait_queue_t *curr, *next; | |
4663 | ||
4664 | list_for_each_entry_safe(curr, next, &q->task_list, task_list) { | |
4665 | unsigned flags = curr->flags; | |
4666 | ||
4667 | if (curr->func(curr, mode, sync, key) && | |
4668 | (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) | |
4669 | break; | |
4670 | } | |
4671 | } | |
4672 | ||
4673 | /** | |
4674 | * __wake_up - wake up threads blocked on a waitqueue. | |
4675 | * @q: the waitqueue | |
4676 | * @mode: which threads | |
4677 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
4678 | * @key: is directly passed to the wakeup function | |
4679 | */ | |
4680 | void __wake_up(wait_queue_head_t *q, unsigned int mode, | |
4681 | int nr_exclusive, void *key) | |
4682 | { | |
4683 | unsigned long flags; | |
4684 | ||
4685 | spin_lock_irqsave(&q->lock, flags); | |
4686 | __wake_up_common(q, mode, nr_exclusive, 0, key); | |
4687 | spin_unlock_irqrestore(&q->lock, flags); | |
4688 | } | |
4689 | EXPORT_SYMBOL(__wake_up); | |
4690 | ||
4691 | /* | |
4692 | * Same as __wake_up but called with the spinlock in wait_queue_head_t held. | |
4693 | */ | |
4694 | void __wake_up_locked(wait_queue_head_t *q, unsigned int mode) | |
4695 | { | |
4696 | __wake_up_common(q, mode, 1, 0, NULL); | |
4697 | } | |
4698 | ||
4699 | /** | |
4700 | * __wake_up_sync - wake up threads blocked on a waitqueue. | |
4701 | * @q: the waitqueue | |
4702 | * @mode: which threads | |
4703 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
4704 | * | |
4705 | * The sync wakeup differs that the waker knows that it will schedule | |
4706 | * away soon, so while the target thread will be woken up, it will not | |
4707 | * be migrated to another CPU - ie. the two threads are 'synchronized' | |
4708 | * with each other. This can prevent needless bouncing between CPUs. | |
4709 | * | |
4710 | * On UP it can prevent extra preemption. | |
4711 | */ | |
4712 | void | |
4713 | __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) | |
4714 | { | |
4715 | unsigned long flags; | |
4716 | int sync = 1; | |
4717 | ||
4718 | if (unlikely(!q)) | |
4719 | return; | |
4720 | ||
4721 | if (unlikely(!nr_exclusive)) | |
4722 | sync = 0; | |
4723 | ||
4724 | spin_lock_irqsave(&q->lock, flags); | |
4725 | __wake_up_common(q, mode, nr_exclusive, sync, NULL); | |
4726 | spin_unlock_irqrestore(&q->lock, flags); | |
4727 | } | |
4728 | EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ | |
4729 | ||
4730 | void complete(struct completion *x) | |
4731 | { | |
4732 | unsigned long flags; | |
4733 | ||
4734 | spin_lock_irqsave(&x->wait.lock, flags); | |
4735 | x->done++; | |
4736 | __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); | |
4737 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
4738 | } | |
4739 | EXPORT_SYMBOL(complete); | |
4740 | ||
4741 | void complete_all(struct completion *x) | |
4742 | { | |
4743 | unsigned long flags; | |
4744 | ||
4745 | spin_lock_irqsave(&x->wait.lock, flags); | |
4746 | x->done += UINT_MAX/2; | |
4747 | __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); | |
4748 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
4749 | } | |
4750 | EXPORT_SYMBOL(complete_all); | |
4751 | ||
4752 | static inline long __sched | |
4753 | do_wait_for_common(struct completion *x, long timeout, int state) | |
4754 | { | |
4755 | if (!x->done) { | |
4756 | DECLARE_WAITQUEUE(wait, current); | |
4757 | ||
4758 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
4759 | __add_wait_queue_tail(&x->wait, &wait); | |
4760 | do { | |
4761 | if ((state == TASK_INTERRUPTIBLE && | |
4762 | signal_pending(current)) || | |
4763 | (state == TASK_KILLABLE && | |
4764 | fatal_signal_pending(current))) { | |
4765 | __remove_wait_queue(&x->wait, &wait); | |
4766 | return -ERESTARTSYS; | |
4767 | } | |
4768 | __set_current_state(state); | |
4769 | spin_unlock_irq(&x->wait.lock); | |
4770 | timeout = schedule_timeout(timeout); | |
4771 | spin_lock_irq(&x->wait.lock); | |
4772 | if (!timeout) { | |
4773 | __remove_wait_queue(&x->wait, &wait); | |
4774 | return timeout; | |
4775 | } | |
4776 | } while (!x->done); | |
4777 | __remove_wait_queue(&x->wait, &wait); | |
4778 | } | |
4779 | x->done--; | |
4780 | return timeout; | |
4781 | } | |
4782 | ||
4783 | static long __sched | |
4784 | wait_for_common(struct completion *x, long timeout, int state) | |
4785 | { | |
4786 | might_sleep(); | |
4787 | ||
4788 | spin_lock_irq(&x->wait.lock); | |
4789 | timeout = do_wait_for_common(x, timeout, state); | |
4790 | spin_unlock_irq(&x->wait.lock); | |
4791 | return timeout; | |
4792 | } | |
4793 | ||
4794 | void __sched wait_for_completion(struct completion *x) | |
4795 | { | |
4796 | wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); | |
4797 | } | |
4798 | EXPORT_SYMBOL(wait_for_completion); | |
4799 | ||
4800 | unsigned long __sched | |
4801 | wait_for_completion_timeout(struct completion *x, unsigned long timeout) | |
4802 | { | |
4803 | return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); | |
4804 | } | |
4805 | EXPORT_SYMBOL(wait_for_completion_timeout); | |
4806 | ||
4807 | int __sched wait_for_completion_interruptible(struct completion *x) | |
4808 | { | |
4809 | long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); | |
4810 | if (t == -ERESTARTSYS) | |
4811 | return t; | |
4812 | return 0; | |
4813 | } | |
4814 | EXPORT_SYMBOL(wait_for_completion_interruptible); | |
4815 | ||
4816 | unsigned long __sched | |
4817 | wait_for_completion_interruptible_timeout(struct completion *x, | |
4818 | unsigned long timeout) | |
4819 | { | |
4820 | return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); | |
4821 | } | |
4822 | EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); | |
4823 | ||
4824 | int __sched wait_for_completion_killable(struct completion *x) | |
4825 | { | |
4826 | long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); | |
4827 | if (t == -ERESTARTSYS) | |
4828 | return t; | |
4829 | return 0; | |
4830 | } | |
4831 | EXPORT_SYMBOL(wait_for_completion_killable); | |
4832 | ||
4833 | static long __sched | |
4834 | sleep_on_common(wait_queue_head_t *q, int state, long timeout) | |
4835 | { | |
4836 | unsigned long flags; | |
4837 | wait_queue_t wait; | |
4838 | ||
4839 | init_waitqueue_entry(&wait, current); | |
4840 | ||
4841 | __set_current_state(state); | |
4842 | ||
4843 | spin_lock_irqsave(&q->lock, flags); | |
4844 | __add_wait_queue(q, &wait); | |
4845 | spin_unlock(&q->lock); | |
4846 | timeout = schedule_timeout(timeout); | |
4847 | spin_lock_irq(&q->lock); | |
4848 | __remove_wait_queue(q, &wait); | |
4849 | spin_unlock_irqrestore(&q->lock, flags); | |
4850 | ||
4851 | return timeout; | |
4852 | } | |
4853 | ||
4854 | void __sched interruptible_sleep_on(wait_queue_head_t *q) | |
4855 | { | |
4856 | sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); | |
4857 | } | |
4858 | EXPORT_SYMBOL(interruptible_sleep_on); | |
4859 | ||
4860 | long __sched | |
4861 | interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) | |
4862 | { | |
4863 | return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); | |
4864 | } | |
4865 | EXPORT_SYMBOL(interruptible_sleep_on_timeout); | |
4866 | ||
4867 | void __sched sleep_on(wait_queue_head_t *q) | |
4868 | { | |
4869 | sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); | |
4870 | } | |
4871 | EXPORT_SYMBOL(sleep_on); | |
4872 | ||
4873 | long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) | |
4874 | { | |
4875 | return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); | |
4876 | } | |
4877 | EXPORT_SYMBOL(sleep_on_timeout); | |
4878 | ||
4879 | #ifdef CONFIG_RT_MUTEXES | |
4880 | ||
4881 | /* | |
4882 | * rt_mutex_setprio - set the current priority of a task | |
4883 | * @p: task | |
4884 | * @prio: prio value (kernel-internal form) | |
4885 | * | |
4886 | * This function changes the 'effective' priority of a task. It does | |
4887 | * not touch ->normal_prio like __setscheduler(). | |
4888 | * | |
4889 | * Used by the rt_mutex code to implement priority inheritance logic. | |
4890 | */ | |
4891 | void rt_mutex_setprio(struct task_struct *p, int prio) | |
4892 | { | |
4893 | unsigned long flags; | |
4894 | int oldprio, on_rq, running; | |
4895 | struct rq *rq; | |
4896 | const struct sched_class *prev_class = p->sched_class; | |
4897 | ||
4898 | BUG_ON(prio < 0 || prio > MAX_PRIO); | |
4899 | ||
4900 | rq = task_rq_lock(p, &flags); | |
4901 | update_rq_clock(rq); | |
4902 | ||
4903 | oldprio = p->prio; | |
4904 | on_rq = p->se.on_rq; | |
4905 | running = task_current(rq, p); | |
4906 | if (on_rq) | |
4907 | dequeue_task(rq, p, 0); | |
4908 | if (running) | |
4909 | p->sched_class->put_prev_task(rq, p); | |
4910 | ||
4911 | if (rt_prio(prio)) | |
4912 | p->sched_class = &rt_sched_class; | |
4913 | else | |
4914 | p->sched_class = &fair_sched_class; | |
4915 | ||
4916 | p->prio = prio; | |
4917 | ||
4918 | if (running) | |
4919 | p->sched_class->set_curr_task(rq); | |
4920 | if (on_rq) { | |
4921 | enqueue_task(rq, p, 0); | |
4922 | ||
4923 | check_class_changed(rq, p, prev_class, oldprio, running); | |
4924 | } | |
4925 | task_rq_unlock(rq, &flags); | |
4926 | } | |
4927 | ||
4928 | #endif | |
4929 | ||
4930 | void set_user_nice(struct task_struct *p, long nice) | |
4931 | { | |
4932 | int old_prio, delta, on_rq; | |
4933 | unsigned long flags; | |
4934 | struct rq *rq; | |
4935 | ||
4936 | if (TASK_NICE(p) == nice || nice < -20 || nice > 19) | |
4937 | return; | |
4938 | /* | |
4939 | * We have to be careful, if called from sys_setpriority(), | |
4940 | * the task might be in the middle of scheduling on another CPU. | |
4941 | */ | |
4942 | rq = task_rq_lock(p, &flags); | |
4943 | update_rq_clock(rq); | |
4944 | /* | |
4945 | * The RT priorities are set via sched_setscheduler(), but we still | |
4946 | * allow the 'normal' nice value to be set - but as expected | |
4947 | * it wont have any effect on scheduling until the task is | |
4948 | * SCHED_FIFO/SCHED_RR: | |
4949 | */ | |
4950 | if (task_has_rt_policy(p)) { | |
4951 | p->static_prio = NICE_TO_PRIO(nice); | |
4952 | goto out_unlock; | |
4953 | } | |
4954 | on_rq = p->se.on_rq; | |
4955 | if (on_rq) | |
4956 | dequeue_task(rq, p, 0); | |
4957 | ||
4958 | p->static_prio = NICE_TO_PRIO(nice); | |
4959 | set_load_weight(p); | |
4960 | old_prio = p->prio; | |
4961 | p->prio = effective_prio(p); | |
4962 | delta = p->prio - old_prio; | |
4963 | ||
4964 | if (on_rq) { | |
4965 | enqueue_task(rq, p, 0); | |
4966 | /* | |
4967 | * If the task increased its priority or is running and | |
4968 | * lowered its priority, then reschedule its CPU: | |
4969 | */ | |
4970 | if (delta < 0 || (delta > 0 && task_running(rq, p))) | |
4971 | resched_task(rq->curr); | |
4972 | } | |
4973 | out_unlock: | |
4974 | task_rq_unlock(rq, &flags); | |
4975 | } | |
4976 | EXPORT_SYMBOL(set_user_nice); | |
4977 | ||
4978 | /* | |
4979 | * can_nice - check if a task can reduce its nice value | |
4980 | * @p: task | |
4981 | * @nice: nice value | |
4982 | */ | |
4983 | int can_nice(const struct task_struct *p, const int nice) | |
4984 | { | |
4985 | /* convert nice value [19,-20] to rlimit style value [1,40] */ | |
4986 | int nice_rlim = 20 - nice; | |
4987 | ||
4988 | return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur || | |
4989 | capable(CAP_SYS_NICE)); | |
4990 | } | |
4991 | ||
4992 | #ifdef __ARCH_WANT_SYS_NICE | |
4993 | ||
4994 | /* | |
4995 | * sys_nice - change the priority of the current process. | |
4996 | * @increment: priority increment | |
4997 | * | |
4998 | * sys_setpriority is a more generic, but much slower function that | |
4999 | * does similar things. | |
5000 | */ | |
5001 | asmlinkage long sys_nice(int increment) | |
5002 | { | |
5003 | long nice, retval; | |
5004 | ||
5005 | /* | |
5006 | * Setpriority might change our priority at the same moment. | |
5007 | * We don't have to worry. Conceptually one call occurs first | |
5008 | * and we have a single winner. | |
5009 | */ | |
5010 | if (increment < -40) | |
5011 | increment = -40; | |
5012 | if (increment > 40) | |
5013 | increment = 40; | |
5014 | ||
5015 | nice = PRIO_TO_NICE(current->static_prio) + increment; | |
5016 | if (nice < -20) | |
5017 | nice = -20; | |
5018 | if (nice > 19) | |
5019 | nice = 19; | |
5020 | ||
5021 | if (increment < 0 && !can_nice(current, nice)) | |
5022 | return -EPERM; | |
5023 | ||
5024 | retval = security_task_setnice(current, nice); | |
5025 | if (retval) | |
5026 | return retval; | |
5027 | ||
5028 | set_user_nice(current, nice); | |
5029 | return 0; | |
5030 | } | |
5031 | ||
5032 | #endif | |
5033 | ||
5034 | /** | |
5035 | * task_prio - return the priority value of a given task. | |
5036 | * @p: the task in question. | |
5037 | * | |
5038 | * This is the priority value as seen by users in /proc. | |
5039 | * RT tasks are offset by -200. Normal tasks are centered | |
5040 | * around 0, value goes from -16 to +15. | |
5041 | */ | |
5042 | int task_prio(const struct task_struct *p) | |
5043 | { | |
5044 | return p->prio - MAX_RT_PRIO; | |
5045 | } | |
5046 | ||
5047 | /** | |
5048 | * task_nice - return the nice value of a given task. | |
5049 | * @p: the task in question. | |
5050 | */ | |
5051 | int task_nice(const struct task_struct *p) | |
5052 | { | |
5053 | return TASK_NICE(p); | |
5054 | } | |
5055 | EXPORT_SYMBOL(task_nice); | |
5056 | ||
5057 | /** | |
5058 | * idle_cpu - is a given cpu idle currently? | |
5059 | * @cpu: the processor in question. | |
5060 | */ | |
5061 | int idle_cpu(int cpu) | |
5062 | { | |
5063 | return cpu_curr(cpu) == cpu_rq(cpu)->idle; | |
5064 | } | |
5065 | ||
5066 | /** | |
5067 | * idle_task - return the idle task for a given cpu. | |
5068 | * @cpu: the processor in question. | |
5069 | */ | |
5070 | struct task_struct *idle_task(int cpu) | |
5071 | { | |
5072 | return cpu_rq(cpu)->idle; | |
5073 | } | |
5074 | ||
5075 | /** | |
5076 | * find_process_by_pid - find a process with a matching PID value. | |
5077 | * @pid: the pid in question. | |
5078 | */ | |
5079 | static struct task_struct *find_process_by_pid(pid_t pid) | |
5080 | { | |
5081 | return pid ? find_task_by_vpid(pid) : current; | |
5082 | } | |
5083 | ||
5084 | /* Actually do priority change: must hold rq lock. */ | |
5085 | static void | |
5086 | __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) | |
5087 | { | |
5088 | BUG_ON(p->se.on_rq); | |
5089 | ||
5090 | p->policy = policy; | |
5091 | switch (p->policy) { | |
5092 | case SCHED_NORMAL: | |
5093 | case SCHED_BATCH: | |
5094 | case SCHED_IDLE: | |
5095 | p->sched_class = &fair_sched_class; | |
5096 | break; | |
5097 | case SCHED_FIFO: | |
5098 | case SCHED_RR: | |
5099 | p->sched_class = &rt_sched_class; | |
5100 | break; | |
5101 | } | |
5102 | ||
5103 | p->rt_priority = prio; | |
5104 | p->normal_prio = normal_prio(p); | |
5105 | /* we are holding p->pi_lock already */ | |
5106 | p->prio = rt_mutex_getprio(p); | |
5107 | set_load_weight(p); | |
5108 | } | |
5109 | ||
5110 | /** | |
5111 | * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. | |
5112 | * @p: the task in question. | |
5113 | * @policy: new policy. | |
5114 | * @param: structure containing the new RT priority. | |
5115 | * | |
5116 | * NOTE that the task may be already dead. | |
5117 | */ | |
5118 | int sched_setscheduler(struct task_struct *p, int policy, | |
5119 | struct sched_param *param) | |
5120 | { | |
5121 | int retval, oldprio, oldpolicy = -1, on_rq, running; | |
5122 | unsigned long flags; | |
5123 | const struct sched_class *prev_class = p->sched_class; | |
5124 | struct rq *rq; | |
5125 | ||
5126 | /* may grab non-irq protected spin_locks */ | |
5127 | BUG_ON(in_interrupt()); | |
5128 | recheck: | |
5129 | /* double check policy once rq lock held */ | |
5130 | if (policy < 0) | |
5131 | policy = oldpolicy = p->policy; | |
5132 | else if (policy != SCHED_FIFO && policy != SCHED_RR && | |
5133 | policy != SCHED_NORMAL && policy != SCHED_BATCH && | |
5134 | policy != SCHED_IDLE) | |
5135 | return -EINVAL; | |
5136 | /* | |
5137 | * Valid priorities for SCHED_FIFO and SCHED_RR are | |
5138 | * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, | |
5139 | * SCHED_BATCH and SCHED_IDLE is 0. | |
5140 | */ | |
5141 | if (param->sched_priority < 0 || | |
5142 | (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || | |
5143 | (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) | |
5144 | return -EINVAL; | |
5145 | if (rt_policy(policy) != (param->sched_priority != 0)) | |
5146 | return -EINVAL; | |
5147 | ||
5148 | /* | |
5149 | * Allow unprivileged RT tasks to decrease priority: | |
5150 | */ | |
5151 | if (!capable(CAP_SYS_NICE)) { | |
5152 | if (rt_policy(policy)) { | |
5153 | unsigned long rlim_rtprio; | |
5154 | ||
5155 | if (!lock_task_sighand(p, &flags)) | |
5156 | return -ESRCH; | |
5157 | rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur; | |
5158 | unlock_task_sighand(p, &flags); | |
5159 | ||
5160 | /* can't set/change the rt policy */ | |
5161 | if (policy != p->policy && !rlim_rtprio) | |
5162 | return -EPERM; | |
5163 | ||
5164 | /* can't increase priority */ | |
5165 | if (param->sched_priority > p->rt_priority && | |
5166 | param->sched_priority > rlim_rtprio) | |
5167 | return -EPERM; | |
5168 | } | |
5169 | /* | |
5170 | * Like positive nice levels, dont allow tasks to | |
5171 | * move out of SCHED_IDLE either: | |
5172 | */ | |
5173 | if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) | |
5174 | return -EPERM; | |
5175 | ||
5176 | /* can't change other user's priorities */ | |
5177 | if ((current->euid != p->euid) && | |
5178 | (current->euid != p->uid)) | |
5179 | return -EPERM; | |
5180 | } | |
5181 | ||
5182 | #ifdef CONFIG_RT_GROUP_SCHED | |
5183 | /* | |
5184 | * Do not allow realtime tasks into groups that have no runtime | |
5185 | * assigned. | |
5186 | */ | |
5187 | if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0) | |
5188 | return -EPERM; | |
5189 | #endif | |
5190 | ||
5191 | retval = security_task_setscheduler(p, policy, param); | |
5192 | if (retval) | |
5193 | return retval; | |
5194 | /* | |
5195 | * make sure no PI-waiters arrive (or leave) while we are | |
5196 | * changing the priority of the task: | |
5197 | */ | |
5198 | spin_lock_irqsave(&p->pi_lock, flags); | |
5199 | /* | |
5200 | * To be able to change p->policy safely, the apropriate | |
5201 | * runqueue lock must be held. | |
5202 | */ | |
5203 | rq = __task_rq_lock(p); | |
5204 | /* recheck policy now with rq lock held */ | |
5205 | if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { | |
5206 | policy = oldpolicy = -1; | |
5207 | __task_rq_unlock(rq); | |
5208 | spin_unlock_irqrestore(&p->pi_lock, flags); | |
5209 | goto recheck; | |
5210 | } | |
5211 | update_rq_clock(rq); | |
5212 | on_rq = p->se.on_rq; | |
5213 | running = task_current(rq, p); | |
5214 | if (on_rq) | |
5215 | deactivate_task(rq, p, 0); | |
5216 | if (running) | |
5217 | p->sched_class->put_prev_task(rq, p); | |
5218 | ||
5219 | oldprio = p->prio; | |
5220 | __setscheduler(rq, p, policy, param->sched_priority); | |
5221 | ||
5222 | if (running) | |
5223 | p->sched_class->set_curr_task(rq); | |
5224 | if (on_rq) { | |
5225 | activate_task(rq, p, 0); | |
5226 | ||
5227 | check_class_changed(rq, p, prev_class, oldprio, running); | |
5228 | } | |
5229 | __task_rq_unlock(rq); | |
5230 | spin_unlock_irqrestore(&p->pi_lock, flags); | |
5231 | ||
5232 | rt_mutex_adjust_pi(p); | |
5233 | ||
5234 | return 0; | |
5235 | } | |
5236 | EXPORT_SYMBOL_GPL(sched_setscheduler); | |
5237 | ||
5238 | static int | |
5239 | do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) | |
5240 | { | |
5241 | struct sched_param lparam; | |
5242 | struct task_struct *p; | |
5243 | int retval; | |
5244 | ||
5245 | if (!param || pid < 0) | |
5246 | return -EINVAL; | |
5247 | if (copy_from_user(&lparam, param, sizeof(struct sched_param))) | |
5248 | return -EFAULT; | |
5249 | ||
5250 | rcu_read_lock(); | |
5251 | retval = -ESRCH; | |
5252 | p = find_process_by_pid(pid); | |
5253 | if (p != NULL) | |
5254 | retval = sched_setscheduler(p, policy, &lparam); | |
5255 | rcu_read_unlock(); | |
5256 | ||
5257 | return retval; | |
5258 | } | |
5259 | ||
5260 | /** | |
5261 | * sys_sched_setscheduler - set/change the scheduler policy and RT priority | |
5262 | * @pid: the pid in question. | |
5263 | * @policy: new policy. | |
5264 | * @param: structure containing the new RT priority. | |
5265 | */ | |
5266 | asmlinkage long | |
5267 | sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) | |
5268 | { | |
5269 | /* negative values for policy are not valid */ | |
5270 | if (policy < 0) | |
5271 | return -EINVAL; | |
5272 | ||
5273 | return do_sched_setscheduler(pid, policy, param); | |
5274 | } | |
5275 | ||
5276 | /** | |
5277 | * sys_sched_setparam - set/change the RT priority of a thread | |
5278 | * @pid: the pid in question. | |
5279 | * @param: structure containing the new RT priority. | |
5280 | */ | |
5281 | asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param) | |
5282 | { | |
5283 | return do_sched_setscheduler(pid, -1, param); | |
5284 | } | |
5285 | ||
5286 | /** | |
5287 | * sys_sched_getscheduler - get the policy (scheduling class) of a thread | |
5288 | * @pid: the pid in question. | |
5289 | */ | |
5290 | asmlinkage long sys_sched_getscheduler(pid_t pid) | |
5291 | { | |
5292 | struct task_struct *p; | |
5293 | int retval; | |
5294 | ||
5295 | if (pid < 0) | |
5296 | return -EINVAL; | |
5297 | ||
5298 | retval = -ESRCH; | |
5299 | read_lock(&tasklist_lock); | |
5300 | p = find_process_by_pid(pid); | |
5301 | if (p) { | |
5302 | retval = security_task_getscheduler(p); | |
5303 | if (!retval) | |
5304 | retval = p->policy; | |
5305 | } | |
5306 | read_unlock(&tasklist_lock); | |
5307 | return retval; | |
5308 | } | |
5309 | ||
5310 | /** | |
5311 | * sys_sched_getscheduler - get the RT priority of a thread | |
5312 | * @pid: the pid in question. | |
5313 | * @param: structure containing the RT priority. | |
5314 | */ | |
5315 | asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param) | |
5316 | { | |
5317 | struct sched_param lp; | |
5318 | struct task_struct *p; | |
5319 | int retval; | |
5320 | ||
5321 | if (!param || pid < 0) | |
5322 | return -EINVAL; | |
5323 | ||
5324 | read_lock(&tasklist_lock); | |
5325 | p = find_process_by_pid(pid); | |
5326 | retval = -ESRCH; | |
5327 | if (!p) | |
5328 | goto out_unlock; | |
5329 | ||
5330 | retval = security_task_getscheduler(p); | |
5331 | if (retval) | |
5332 | goto out_unlock; | |
5333 | ||
5334 | lp.sched_priority = p->rt_priority; | |
5335 | read_unlock(&tasklist_lock); | |
5336 | ||
5337 | /* | |
5338 | * This one might sleep, we cannot do it with a spinlock held ... | |
5339 | */ | |
5340 | retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; | |
5341 | ||
5342 | return retval; | |
5343 | ||
5344 | out_unlock: | |
5345 | read_unlock(&tasklist_lock); | |
5346 | return retval; | |
5347 | } | |
5348 | ||
5349 | long sched_setaffinity(pid_t pid, const cpumask_t *in_mask) | |
5350 | { | |
5351 | cpumask_t cpus_allowed; | |
5352 | cpumask_t new_mask = *in_mask; | |
5353 | struct task_struct *p; | |
5354 | int retval; | |
5355 | ||
5356 | get_online_cpus(); | |
5357 | read_lock(&tasklist_lock); | |
5358 | ||
5359 | p = find_process_by_pid(pid); | |
5360 | if (!p) { | |
5361 | read_unlock(&tasklist_lock); | |
5362 | put_online_cpus(); | |
5363 | return -ESRCH; | |
5364 | } | |
5365 | ||
5366 | /* | |
5367 | * It is not safe to call set_cpus_allowed with the | |
5368 | * tasklist_lock held. We will bump the task_struct's | |
5369 | * usage count and then drop tasklist_lock. | |
5370 | */ | |
5371 | get_task_struct(p); | |
5372 | read_unlock(&tasklist_lock); | |
5373 | ||
5374 | retval = -EPERM; | |
5375 | if ((current->euid != p->euid) && (current->euid != p->uid) && | |
5376 | !capable(CAP_SYS_NICE)) | |
5377 | goto out_unlock; | |
5378 | ||
5379 | retval = security_task_setscheduler(p, 0, NULL); | |
5380 | if (retval) | |
5381 | goto out_unlock; | |
5382 | ||
5383 | cpuset_cpus_allowed(p, &cpus_allowed); | |
5384 | cpus_and(new_mask, new_mask, cpus_allowed); | |
5385 | again: | |
5386 | retval = set_cpus_allowed_ptr(p, &new_mask); | |
5387 | ||
5388 | if (!retval) { | |
5389 | cpuset_cpus_allowed(p, &cpus_allowed); | |
5390 | if (!cpus_subset(new_mask, cpus_allowed)) { | |
5391 | /* | |
5392 | * We must have raced with a concurrent cpuset | |
5393 | * update. Just reset the cpus_allowed to the | |
5394 | * cpuset's cpus_allowed | |
5395 | */ | |
5396 | new_mask = cpus_allowed; | |
5397 | goto again; | |
5398 | } | |
5399 | } | |
5400 | out_unlock: | |
5401 | put_task_struct(p); | |
5402 | put_online_cpus(); | |
5403 | return retval; | |
5404 | } | |
5405 | ||
5406 | static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, | |
5407 | cpumask_t *new_mask) | |
5408 | { | |
5409 | if (len < sizeof(cpumask_t)) { | |
5410 | memset(new_mask, 0, sizeof(cpumask_t)); | |
5411 | } else if (len > sizeof(cpumask_t)) { | |
5412 | len = sizeof(cpumask_t); | |
5413 | } | |
5414 | return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; | |
5415 | } | |
5416 | ||
5417 | /** | |
5418 | * sys_sched_setaffinity - set the cpu affinity of a process | |
5419 | * @pid: pid of the process | |
5420 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
5421 | * @user_mask_ptr: user-space pointer to the new cpu mask | |
5422 | */ | |
5423 | asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len, | |
5424 | unsigned long __user *user_mask_ptr) | |
5425 | { | |
5426 | cpumask_t new_mask; | |
5427 | int retval; | |
5428 | ||
5429 | retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask); | |
5430 | if (retval) | |
5431 | return retval; | |
5432 | ||
5433 | return sched_setaffinity(pid, &new_mask); | |
5434 | } | |
5435 | ||
5436 | /* | |
5437 | * Represents all cpu's present in the system | |
5438 | * In systems capable of hotplug, this map could dynamically grow | |
5439 | * as new cpu's are detected in the system via any platform specific | |
5440 | * method, such as ACPI for e.g. | |
5441 | */ | |
5442 | ||
5443 | cpumask_t cpu_present_map __read_mostly; | |
5444 | EXPORT_SYMBOL(cpu_present_map); | |
5445 | ||
5446 | #ifndef CONFIG_SMP | |
5447 | cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL; | |
5448 | EXPORT_SYMBOL(cpu_online_map); | |
5449 | ||
5450 | cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL; | |
5451 | EXPORT_SYMBOL(cpu_possible_map); | |
5452 | #endif | |
5453 | ||
5454 | long sched_getaffinity(pid_t pid, cpumask_t *mask) | |
5455 | { | |
5456 | struct task_struct *p; | |
5457 | int retval; | |
5458 | ||
5459 | get_online_cpus(); | |
5460 | read_lock(&tasklist_lock); | |
5461 | ||
5462 | retval = -ESRCH; | |
5463 | p = find_process_by_pid(pid); | |
5464 | if (!p) | |
5465 | goto out_unlock; | |
5466 | ||
5467 | retval = security_task_getscheduler(p); | |
5468 | if (retval) | |
5469 | goto out_unlock; | |
5470 | ||
5471 | cpus_and(*mask, p->cpus_allowed, cpu_online_map); | |
5472 | ||
5473 | out_unlock: | |
5474 | read_unlock(&tasklist_lock); | |
5475 | put_online_cpus(); | |
5476 | ||
5477 | return retval; | |
5478 | } | |
5479 | ||
5480 | /** | |
5481 | * sys_sched_getaffinity - get the cpu affinity of a process | |
5482 | * @pid: pid of the process | |
5483 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
5484 | * @user_mask_ptr: user-space pointer to hold the current cpu mask | |
5485 | */ | |
5486 | asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len, | |
5487 | unsigned long __user *user_mask_ptr) | |
5488 | { | |
5489 | int ret; | |
5490 | cpumask_t mask; | |
5491 | ||
5492 | if (len < sizeof(cpumask_t)) | |
5493 | return -EINVAL; | |
5494 | ||
5495 | ret = sched_getaffinity(pid, &mask); | |
5496 | if (ret < 0) | |
5497 | return ret; | |
5498 | ||
5499 | if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t))) | |
5500 | return -EFAULT; | |
5501 | ||
5502 | return sizeof(cpumask_t); | |
5503 | } | |
5504 | ||
5505 | /** | |
5506 | * sys_sched_yield - yield the current processor to other threads. | |
5507 | * | |
5508 | * This function yields the current CPU to other tasks. If there are no | |
5509 | * other threads running on this CPU then this function will return. | |
5510 | */ | |
5511 | asmlinkage long sys_sched_yield(void) | |
5512 | { | |
5513 | struct rq *rq = this_rq_lock(); | |
5514 | ||
5515 | schedstat_inc(rq, yld_count); | |
5516 | current->sched_class->yield_task(rq); | |
5517 | ||
5518 | /* | |
5519 | * Since we are going to call schedule() anyway, there's | |
5520 | * no need to preempt or enable interrupts: | |
5521 | */ | |
5522 | __release(rq->lock); | |
5523 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); | |
5524 | _raw_spin_unlock(&rq->lock); | |
5525 | preempt_enable_no_resched(); | |
5526 | ||
5527 | schedule(); | |
5528 | ||
5529 | return 0; | |
5530 | } | |
5531 | ||
5532 | static void __cond_resched(void) | |
5533 | { | |
5534 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP | |
5535 | __might_sleep(__FILE__, __LINE__); | |
5536 | #endif | |
5537 | /* | |
5538 | * The BKS might be reacquired before we have dropped | |
5539 | * PREEMPT_ACTIVE, which could trigger a second | |
5540 | * cond_resched() call. | |
5541 | */ | |
5542 | do { | |
5543 | add_preempt_count(PREEMPT_ACTIVE); | |
5544 | schedule(); | |
5545 | sub_preempt_count(PREEMPT_ACTIVE); | |
5546 | } while (need_resched()); | |
5547 | } | |
5548 | ||
5549 | #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY) | |
5550 | int __sched _cond_resched(void) | |
5551 | { | |
5552 | if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) && | |
5553 | system_state == SYSTEM_RUNNING) { | |
5554 | __cond_resched(); | |
5555 | return 1; | |
5556 | } | |
5557 | return 0; | |
5558 | } | |
5559 | EXPORT_SYMBOL(_cond_resched); | |
5560 | #endif | |
5561 | ||
5562 | /* | |
5563 | * cond_resched_lock() - if a reschedule is pending, drop the given lock, | |
5564 | * call schedule, and on return reacquire the lock. | |
5565 | * | |
5566 | * This works OK both with and without CONFIG_PREEMPT. We do strange low-level | |
5567 | * operations here to prevent schedule() from being called twice (once via | |
5568 | * spin_unlock(), once by hand). | |
5569 | */ | |
5570 | int cond_resched_lock(spinlock_t *lock) | |
5571 | { | |
5572 | int resched = need_resched() && system_state == SYSTEM_RUNNING; | |
5573 | int ret = 0; | |
5574 | ||
5575 | if (spin_needbreak(lock) || resched) { | |
5576 | spin_unlock(lock); | |
5577 | if (resched && need_resched()) | |
5578 | __cond_resched(); | |
5579 | else | |
5580 | cpu_relax(); | |
5581 | ret = 1; | |
5582 | spin_lock(lock); | |
5583 | } | |
5584 | return ret; | |
5585 | } | |
5586 | EXPORT_SYMBOL(cond_resched_lock); | |
5587 | ||
5588 | int __sched cond_resched_softirq(void) | |
5589 | { | |
5590 | BUG_ON(!in_softirq()); | |
5591 | ||
5592 | if (need_resched() && system_state == SYSTEM_RUNNING) { | |
5593 | local_bh_enable(); | |
5594 | __cond_resched(); | |
5595 | local_bh_disable(); | |
5596 | return 1; | |
5597 | } | |
5598 | return 0; | |
5599 | } | |
5600 | EXPORT_SYMBOL(cond_resched_softirq); | |
5601 | ||
5602 | /** | |
5603 | * yield - yield the current processor to other threads. | |
5604 | * | |
5605 | * This is a shortcut for kernel-space yielding - it marks the | |
5606 | * thread runnable and calls sys_sched_yield(). | |
5607 | */ | |
5608 | void __sched yield(void) | |
5609 | { | |
5610 | set_current_state(TASK_RUNNING); | |
5611 | sys_sched_yield(); | |
5612 | } | |
5613 | EXPORT_SYMBOL(yield); | |
5614 | ||
5615 | /* | |
5616 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so | |
5617 | * that process accounting knows that this is a task in IO wait state. | |
5618 | * | |
5619 | * But don't do that if it is a deliberate, throttling IO wait (this task | |
5620 | * has set its backing_dev_info: the queue against which it should throttle) | |
5621 | */ | |
5622 | void __sched io_schedule(void) | |
5623 | { | |
5624 | struct rq *rq = &__raw_get_cpu_var(runqueues); | |
5625 | ||
5626 | delayacct_blkio_start(); | |
5627 | atomic_inc(&rq->nr_iowait); | |
5628 | schedule(); | |
5629 | atomic_dec(&rq->nr_iowait); | |
5630 | delayacct_blkio_end(); | |
5631 | } | |
5632 | EXPORT_SYMBOL(io_schedule); | |
5633 | ||
5634 | long __sched io_schedule_timeout(long timeout) | |
5635 | { | |
5636 | struct rq *rq = &__raw_get_cpu_var(runqueues); | |
5637 | long ret; | |
5638 | ||
5639 | delayacct_blkio_start(); | |
5640 | atomic_inc(&rq->nr_iowait); | |
5641 | ret = schedule_timeout(timeout); | |
5642 | atomic_dec(&rq->nr_iowait); | |
5643 | delayacct_blkio_end(); | |
5644 | return ret; | |
5645 | } | |
5646 | ||
5647 | /** | |
5648 | * sys_sched_get_priority_max - return maximum RT priority. | |
5649 | * @policy: scheduling class. | |
5650 | * | |
5651 | * this syscall returns the maximum rt_priority that can be used | |
5652 | * by a given scheduling class. | |
5653 | */ | |
5654 | asmlinkage long sys_sched_get_priority_max(int policy) | |
5655 | { | |
5656 | int ret = -EINVAL; | |
5657 | ||
5658 | switch (policy) { | |
5659 | case SCHED_FIFO: | |
5660 | case SCHED_RR: | |
5661 | ret = MAX_USER_RT_PRIO-1; | |
5662 | break; | |
5663 | case SCHED_NORMAL: | |
5664 | case SCHED_BATCH: | |
5665 | case SCHED_IDLE: | |
5666 | ret = 0; | |
5667 | break; | |
5668 | } | |
5669 | return ret; | |
5670 | } | |
5671 | ||
5672 | /** | |
5673 | * sys_sched_get_priority_min - return minimum RT priority. | |
5674 | * @policy: scheduling class. | |
5675 | * | |
5676 | * this syscall returns the minimum rt_priority that can be used | |
5677 | * by a given scheduling class. | |
5678 | */ | |
5679 | asmlinkage long sys_sched_get_priority_min(int policy) | |
5680 | { | |
5681 | int ret = -EINVAL; | |
5682 | ||
5683 | switch (policy) { | |
5684 | case SCHED_FIFO: | |
5685 | case SCHED_RR: | |
5686 | ret = 1; | |
5687 | break; | |
5688 | case SCHED_NORMAL: | |
5689 | case SCHED_BATCH: | |
5690 | case SCHED_IDLE: | |
5691 | ret = 0; | |
5692 | } | |
5693 | return ret; | |
5694 | } | |
5695 | ||
5696 | /** | |
5697 | * sys_sched_rr_get_interval - return the default timeslice of a process. | |
5698 | * @pid: pid of the process. | |
5699 | * @interval: userspace pointer to the timeslice value. | |
5700 | * | |
5701 | * this syscall writes the default timeslice value of a given process | |
5702 | * into the user-space timespec buffer. A value of '0' means infinity. | |
5703 | */ | |
5704 | asmlinkage | |
5705 | long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval) | |
5706 | { | |
5707 | struct task_struct *p; | |
5708 | unsigned int time_slice; | |
5709 | int retval; | |
5710 | struct timespec t; | |
5711 | ||
5712 | if (pid < 0) | |
5713 | return -EINVAL; | |
5714 | ||
5715 | retval = -ESRCH; | |
5716 | read_lock(&tasklist_lock); | |
5717 | p = find_process_by_pid(pid); | |
5718 | if (!p) | |
5719 | goto out_unlock; | |
5720 | ||
5721 | retval = security_task_getscheduler(p); | |
5722 | if (retval) | |
5723 | goto out_unlock; | |
5724 | ||
5725 | /* | |
5726 | * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER | |
5727 | * tasks that are on an otherwise idle runqueue: | |
5728 | */ | |
5729 | time_slice = 0; | |
5730 | if (p->policy == SCHED_RR) { | |
5731 | time_slice = DEF_TIMESLICE; | |
5732 | } else if (p->policy != SCHED_FIFO) { | |
5733 | struct sched_entity *se = &p->se; | |
5734 | unsigned long flags; | |
5735 | struct rq *rq; | |
5736 | ||
5737 | rq = task_rq_lock(p, &flags); | |
5738 | if (rq->cfs.load.weight) | |
5739 | time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se)); | |
5740 | task_rq_unlock(rq, &flags); | |
5741 | } | |
5742 | read_unlock(&tasklist_lock); | |
5743 | jiffies_to_timespec(time_slice, &t); | |
5744 | retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; | |
5745 | return retval; | |
5746 | ||
5747 | out_unlock: | |
5748 | read_unlock(&tasklist_lock); | |
5749 | return retval; | |
5750 | } | |
5751 | ||
5752 | static const char stat_nam[] = "RSDTtZX"; | |
5753 | ||
5754 | void sched_show_task(struct task_struct *p) | |
5755 | { | |
5756 | unsigned long free = 0; | |
5757 | unsigned state; | |
5758 | ||
5759 | state = p->state ? __ffs(p->state) + 1 : 0; | |
5760 | printk(KERN_INFO "%-13.13s %c", p->comm, | |
5761 | state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); | |
5762 | #if BITS_PER_LONG == 32 | |
5763 | if (state == TASK_RUNNING) | |
5764 | printk(KERN_CONT " running "); | |
5765 | else | |
5766 | printk(KERN_CONT " %08lx ", thread_saved_pc(p)); | |
5767 | #else | |
5768 | if (state == TASK_RUNNING) | |
5769 | printk(KERN_CONT " running task "); | |
5770 | else | |
5771 | printk(KERN_CONT " %016lx ", thread_saved_pc(p)); | |
5772 | #endif | |
5773 | #ifdef CONFIG_DEBUG_STACK_USAGE | |
5774 | { | |
5775 | unsigned long *n = end_of_stack(p); | |
5776 | while (!*n) | |
5777 | n++; | |
5778 | free = (unsigned long)n - (unsigned long)end_of_stack(p); | |
5779 | } | |
5780 | #endif | |
5781 | printk(KERN_CONT "%5lu %5d %6d\n", free, | |
5782 | task_pid_nr(p), task_pid_nr(p->real_parent)); | |
5783 | ||
5784 | show_stack(p, NULL); | |
5785 | } | |
5786 | ||
5787 | void show_state_filter(unsigned long state_filter) | |
5788 | { | |
5789 | struct task_struct *g, *p; | |
5790 | ||
5791 | #if BITS_PER_LONG == 32 | |
5792 | printk(KERN_INFO | |
5793 | " task PC stack pid father\n"); | |
5794 | #else | |
5795 | printk(KERN_INFO | |
5796 | " task PC stack pid father\n"); | |
5797 | #endif | |
5798 | read_lock(&tasklist_lock); | |
5799 | do_each_thread(g, p) { | |
5800 | /* | |
5801 | * reset the NMI-timeout, listing all files on a slow | |
5802 | * console might take alot of time: | |
5803 | */ | |
5804 | touch_nmi_watchdog(); | |
5805 | if (!state_filter || (p->state & state_filter)) | |
5806 | sched_show_task(p); | |
5807 | } while_each_thread(g, p); | |
5808 | ||
5809 | touch_all_softlockup_watchdogs(); | |
5810 | ||
5811 | #ifdef CONFIG_SCHED_DEBUG | |
5812 | sysrq_sched_debug_show(); | |
5813 | #endif | |
5814 | read_unlock(&tasklist_lock); | |
5815 | /* | |
5816 | * Only show locks if all tasks are dumped: | |
5817 | */ | |
5818 | if (state_filter == -1) | |
5819 | debug_show_all_locks(); | |
5820 | } | |
5821 | ||
5822 | void __cpuinit init_idle_bootup_task(struct task_struct *idle) | |
5823 | { | |
5824 | idle->sched_class = &idle_sched_class; | |
5825 | } | |
5826 | ||
5827 | /** | |
5828 | * init_idle - set up an idle thread for a given CPU | |
5829 | * @idle: task in question | |
5830 | * @cpu: cpu the idle task belongs to | |
5831 | * | |
5832 | * NOTE: this function does not set the idle thread's NEED_RESCHED | |
5833 | * flag, to make booting more robust. | |
5834 | */ | |
5835 | void __cpuinit init_idle(struct task_struct *idle, int cpu) | |
5836 | { | |
5837 | struct rq *rq = cpu_rq(cpu); | |
5838 | unsigned long flags; | |
5839 | ||
5840 | __sched_fork(idle); | |
5841 | idle->se.exec_start = sched_clock(); | |
5842 | ||
5843 | idle->prio = idle->normal_prio = MAX_PRIO; | |
5844 | idle->cpus_allowed = cpumask_of_cpu(cpu); | |
5845 | __set_task_cpu(idle, cpu); | |
5846 | ||
5847 | spin_lock_irqsave(&rq->lock, flags); | |
5848 | rq->curr = rq->idle = idle; | |
5849 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) | |
5850 | idle->oncpu = 1; | |
5851 | #endif | |
5852 | spin_unlock_irqrestore(&rq->lock, flags); | |
5853 | ||
5854 | /* Set the preempt count _outside_ the spinlocks! */ | |
5855 | task_thread_info(idle)->preempt_count = 0; | |
5856 | ||
5857 | /* | |
5858 | * The idle tasks have their own, simple scheduling class: | |
5859 | */ | |
5860 | idle->sched_class = &idle_sched_class; | |
5861 | } | |
5862 | ||
5863 | /* | |
5864 | * In a system that switches off the HZ timer nohz_cpu_mask | |
5865 | * indicates which cpus entered this state. This is used | |
5866 | * in the rcu update to wait only for active cpus. For system | |
5867 | * which do not switch off the HZ timer nohz_cpu_mask should | |
5868 | * always be CPU_MASK_NONE. | |
5869 | */ | |
5870 | cpumask_t nohz_cpu_mask = CPU_MASK_NONE; | |
5871 | ||
5872 | /* | |
5873 | * Increase the granularity value when there are more CPUs, | |
5874 | * because with more CPUs the 'effective latency' as visible | |
5875 | * to users decreases. But the relationship is not linear, | |
5876 | * so pick a second-best guess by going with the log2 of the | |
5877 | * number of CPUs. | |
5878 | * | |
5879 | * This idea comes from the SD scheduler of Con Kolivas: | |
5880 | */ | |
5881 | static inline void sched_init_granularity(void) | |
5882 | { | |
5883 | unsigned int factor = 1 + ilog2(num_online_cpus()); | |
5884 | const unsigned long limit = 200000000; | |
5885 | ||
5886 | sysctl_sched_min_granularity *= factor; | |
5887 | if (sysctl_sched_min_granularity > limit) | |
5888 | sysctl_sched_min_granularity = limit; | |
5889 | ||
5890 | sysctl_sched_latency *= factor; | |
5891 | if (sysctl_sched_latency > limit) | |
5892 | sysctl_sched_latency = limit; | |
5893 | ||
5894 | sysctl_sched_wakeup_granularity *= factor; | |
5895 | } | |
5896 | ||
5897 | #ifdef CONFIG_SMP | |
5898 | /* | |
5899 | * This is how migration works: | |
5900 | * | |
5901 | * 1) we queue a struct migration_req structure in the source CPU's | |
5902 | * runqueue and wake up that CPU's migration thread. | |
5903 | * 2) we down() the locked semaphore => thread blocks. | |
5904 | * 3) migration thread wakes up (implicitly it forces the migrated | |
5905 | * thread off the CPU) | |
5906 | * 4) it gets the migration request and checks whether the migrated | |
5907 | * task is still in the wrong runqueue. | |
5908 | * 5) if it's in the wrong runqueue then the migration thread removes | |
5909 | * it and puts it into the right queue. | |
5910 | * 6) migration thread up()s the semaphore. | |
5911 | * 7) we wake up and the migration is done. | |
5912 | */ | |
5913 | ||
5914 | /* | |
5915 | * Change a given task's CPU affinity. Migrate the thread to a | |
5916 | * proper CPU and schedule it away if the CPU it's executing on | |
5917 | * is removed from the allowed bitmask. | |
5918 | * | |
5919 | * NOTE: the caller must have a valid reference to the task, the | |
5920 | * task must not exit() & deallocate itself prematurely. The | |
5921 | * call is not atomic; no spinlocks may be held. | |
5922 | */ | |
5923 | int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask) | |
5924 | { | |
5925 | struct migration_req req; | |
5926 | unsigned long flags; | |
5927 | struct rq *rq; | |
5928 | int ret = 0; | |
5929 | ||
5930 | rq = task_rq_lock(p, &flags); | |
5931 | if (!cpus_intersects(*new_mask, cpu_online_map)) { | |
5932 | ret = -EINVAL; | |
5933 | goto out; | |
5934 | } | |
5935 | ||
5936 | if (p->sched_class->set_cpus_allowed) | |
5937 | p->sched_class->set_cpus_allowed(p, new_mask); | |
5938 | else { | |
5939 | p->cpus_allowed = *new_mask; | |
5940 | p->rt.nr_cpus_allowed = cpus_weight(*new_mask); | |
5941 | } | |
5942 | ||
5943 | /* Can the task run on the task's current CPU? If so, we're done */ | |
5944 | if (cpu_isset(task_cpu(p), *new_mask)) | |
5945 | goto out; | |
5946 | ||
5947 | if (migrate_task(p, any_online_cpu(*new_mask), &req)) { | |
5948 | /* Need help from migration thread: drop lock and wait. */ | |
5949 | task_rq_unlock(rq, &flags); | |
5950 | wake_up_process(rq->migration_thread); | |
5951 | wait_for_completion(&req.done); | |
5952 | tlb_migrate_finish(p->mm); | |
5953 | return 0; | |
5954 | } | |
5955 | out: | |
5956 | task_rq_unlock(rq, &flags); | |
5957 | ||
5958 | return ret; | |
5959 | } | |
5960 | EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); | |
5961 | ||
5962 | /* | |
5963 | * Move (not current) task off this cpu, onto dest cpu. We're doing | |
5964 | * this because either it can't run here any more (set_cpus_allowed() | |
5965 | * away from this CPU, or CPU going down), or because we're | |
5966 | * attempting to rebalance this task on exec (sched_exec). | |
5967 | * | |
5968 | * So we race with normal scheduler movements, but that's OK, as long | |
5969 | * as the task is no longer on this CPU. | |
5970 | * | |
5971 | * Returns non-zero if task was successfully migrated. | |
5972 | */ | |
5973 | static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) | |
5974 | { | |
5975 | struct rq *rq_dest, *rq_src; | |
5976 | int ret = 0, on_rq; | |
5977 | ||
5978 | if (unlikely(cpu_is_offline(dest_cpu))) | |
5979 | return ret; | |
5980 | ||
5981 | rq_src = cpu_rq(src_cpu); | |
5982 | rq_dest = cpu_rq(dest_cpu); | |
5983 | ||
5984 | double_rq_lock(rq_src, rq_dest); | |
5985 | /* Already moved. */ | |
5986 | if (task_cpu(p) != src_cpu) | |
5987 | goto out; | |
5988 | /* Affinity changed (again). */ | |
5989 | if (!cpu_isset(dest_cpu, p->cpus_allowed)) | |
5990 | goto out; | |
5991 | ||
5992 | on_rq = p->se.on_rq; | |
5993 | if (on_rq) | |
5994 | deactivate_task(rq_src, p, 0); | |
5995 | ||
5996 | set_task_cpu(p, dest_cpu); | |
5997 | if (on_rq) { | |
5998 | activate_task(rq_dest, p, 0); | |
5999 | check_preempt_curr(rq_dest, p); | |
6000 | } | |
6001 | ret = 1; | |
6002 | out: | |
6003 | double_rq_unlock(rq_src, rq_dest); | |
6004 | return ret; | |
6005 | } | |
6006 | ||
6007 | /* | |
6008 | * migration_thread - this is a highprio system thread that performs | |
6009 | * thread migration by bumping thread off CPU then 'pushing' onto | |
6010 | * another runqueue. | |
6011 | */ | |
6012 | static int migration_thread(void *data) | |
6013 | { | |
6014 | int cpu = (long)data; | |
6015 | struct rq *rq; | |
6016 | ||
6017 | rq = cpu_rq(cpu); | |
6018 | BUG_ON(rq->migration_thread != current); | |
6019 | ||
6020 | set_current_state(TASK_INTERRUPTIBLE); | |
6021 | while (!kthread_should_stop()) { | |
6022 | struct migration_req *req; | |
6023 | struct list_head *head; | |
6024 | ||
6025 | spin_lock_irq(&rq->lock); | |
6026 | ||
6027 | if (cpu_is_offline(cpu)) { | |
6028 | spin_unlock_irq(&rq->lock); | |
6029 | goto wait_to_die; | |
6030 | } | |
6031 | ||
6032 | if (rq->active_balance) { | |
6033 | active_load_balance(rq, cpu); | |
6034 | rq->active_balance = 0; | |
6035 | } | |
6036 | ||
6037 | head = &rq->migration_queue; | |
6038 | ||
6039 | if (list_empty(head)) { | |
6040 | spin_unlock_irq(&rq->lock); | |
6041 | schedule(); | |
6042 | set_current_state(TASK_INTERRUPTIBLE); | |
6043 | continue; | |
6044 | } | |
6045 | req = list_entry(head->next, struct migration_req, list); | |
6046 | list_del_init(head->next); | |
6047 | ||
6048 | spin_unlock(&rq->lock); | |
6049 | __migrate_task(req->task, cpu, req->dest_cpu); | |
6050 | local_irq_enable(); | |
6051 | ||
6052 | complete(&req->done); | |
6053 | } | |
6054 | __set_current_state(TASK_RUNNING); | |
6055 | return 0; | |
6056 | ||
6057 | wait_to_die: | |
6058 | /* Wait for kthread_stop */ | |
6059 | set_current_state(TASK_INTERRUPTIBLE); | |
6060 | while (!kthread_should_stop()) { | |
6061 | schedule(); | |
6062 | set_current_state(TASK_INTERRUPTIBLE); | |
6063 | } | |
6064 | __set_current_state(TASK_RUNNING); | |
6065 | return 0; | |
6066 | } | |
6067 | ||
6068 | #ifdef CONFIG_HOTPLUG_CPU | |
6069 | ||
6070 | static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu) | |
6071 | { | |
6072 | int ret; | |
6073 | ||
6074 | local_irq_disable(); | |
6075 | ret = __migrate_task(p, src_cpu, dest_cpu); | |
6076 | local_irq_enable(); | |
6077 | return ret; | |
6078 | } | |
6079 | ||
6080 | /* | |
6081 | * Figure out where task on dead CPU should go, use force if necessary. | |
6082 | * NOTE: interrupts should be disabled by the caller | |
6083 | */ | |
6084 | static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p) | |
6085 | { | |
6086 | unsigned long flags; | |
6087 | cpumask_t mask; | |
6088 | struct rq *rq; | |
6089 | int dest_cpu; | |
6090 | ||
6091 | do { | |
6092 | /* On same node? */ | |
6093 | mask = node_to_cpumask(cpu_to_node(dead_cpu)); | |
6094 | cpus_and(mask, mask, p->cpus_allowed); | |
6095 | dest_cpu = any_online_cpu(mask); | |
6096 | ||
6097 | /* On any allowed CPU? */ | |
6098 | if (dest_cpu >= nr_cpu_ids) | |
6099 | dest_cpu = any_online_cpu(p->cpus_allowed); | |
6100 | ||
6101 | /* No more Mr. Nice Guy. */ | |
6102 | if (dest_cpu >= nr_cpu_ids) { | |
6103 | cpumask_t cpus_allowed; | |
6104 | ||
6105 | cpuset_cpus_allowed_locked(p, &cpus_allowed); | |
6106 | /* | |
6107 | * Try to stay on the same cpuset, where the | |
6108 | * current cpuset may be a subset of all cpus. | |
6109 | * The cpuset_cpus_allowed_locked() variant of | |
6110 | * cpuset_cpus_allowed() will not block. It must be | |
6111 | * called within calls to cpuset_lock/cpuset_unlock. | |
6112 | */ | |
6113 | rq = task_rq_lock(p, &flags); | |
6114 | p->cpus_allowed = cpus_allowed; | |
6115 | dest_cpu = any_online_cpu(p->cpus_allowed); | |
6116 | task_rq_unlock(rq, &flags); | |
6117 | ||
6118 | /* | |
6119 | * Don't tell them about moving exiting tasks or | |
6120 | * kernel threads (both mm NULL), since they never | |
6121 | * leave kernel. | |
6122 | */ | |
6123 | if (p->mm && printk_ratelimit()) { | |
6124 | printk(KERN_INFO "process %d (%s) no " | |
6125 | "longer affine to cpu%d\n", | |
6126 | task_pid_nr(p), p->comm, dead_cpu); | |
6127 | } | |
6128 | } | |
6129 | } while (!__migrate_task_irq(p, dead_cpu, dest_cpu)); | |
6130 | } | |
6131 | ||
6132 | /* | |
6133 | * While a dead CPU has no uninterruptible tasks queued at this point, | |
6134 | * it might still have a nonzero ->nr_uninterruptible counter, because | |
6135 | * for performance reasons the counter is not stricly tracking tasks to | |
6136 | * their home CPUs. So we just add the counter to another CPU's counter, | |
6137 | * to keep the global sum constant after CPU-down: | |
6138 | */ | |
6139 | static void migrate_nr_uninterruptible(struct rq *rq_src) | |
6140 | { | |
6141 | struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR)); | |
6142 | unsigned long flags; | |
6143 | ||
6144 | local_irq_save(flags); | |
6145 | double_rq_lock(rq_src, rq_dest); | |
6146 | rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible; | |
6147 | rq_src->nr_uninterruptible = 0; | |
6148 | double_rq_unlock(rq_src, rq_dest); | |
6149 | local_irq_restore(flags); | |
6150 | } | |
6151 | ||
6152 | /* Run through task list and migrate tasks from the dead cpu. */ | |
6153 | static void migrate_live_tasks(int src_cpu) | |
6154 | { | |
6155 | struct task_struct *p, *t; | |
6156 | ||
6157 | read_lock(&tasklist_lock); | |
6158 | ||
6159 | do_each_thread(t, p) { | |
6160 | if (p == current) | |
6161 | continue; | |
6162 | ||
6163 | if (task_cpu(p) == src_cpu) | |
6164 | move_task_off_dead_cpu(src_cpu, p); | |
6165 | } while_each_thread(t, p); | |
6166 | ||
6167 | read_unlock(&tasklist_lock); | |
6168 | } | |
6169 | ||
6170 | /* | |
6171 | * Schedules idle task to be the next runnable task on current CPU. | |
6172 | * It does so by boosting its priority to highest possible. | |
6173 | * Used by CPU offline code. | |
6174 | */ | |
6175 | void sched_idle_next(void) | |
6176 | { | |
6177 | int this_cpu = smp_processor_id(); | |
6178 | struct rq *rq = cpu_rq(this_cpu); | |
6179 | struct task_struct *p = rq->idle; | |
6180 | unsigned long flags; | |
6181 | ||
6182 | /* cpu has to be offline */ | |
6183 | BUG_ON(cpu_online(this_cpu)); | |
6184 | ||
6185 | /* | |
6186 | * Strictly not necessary since rest of the CPUs are stopped by now | |
6187 | * and interrupts disabled on the current cpu. | |
6188 | */ | |
6189 | spin_lock_irqsave(&rq->lock, flags); | |
6190 | ||
6191 | __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1); | |
6192 | ||
6193 | update_rq_clock(rq); | |
6194 | activate_task(rq, p, 0); | |
6195 | ||
6196 | spin_unlock_irqrestore(&rq->lock, flags); | |
6197 | } | |
6198 | ||
6199 | /* | |
6200 | * Ensures that the idle task is using init_mm right before its cpu goes | |
6201 | * offline. | |
6202 | */ | |
6203 | void idle_task_exit(void) | |
6204 | { | |
6205 | struct mm_struct *mm = current->active_mm; | |
6206 | ||
6207 | BUG_ON(cpu_online(smp_processor_id())); | |
6208 | ||
6209 | if (mm != &init_mm) | |
6210 | switch_mm(mm, &init_mm, current); | |
6211 | mmdrop(mm); | |
6212 | } | |
6213 | ||
6214 | /* called under rq->lock with disabled interrupts */ | |
6215 | static void migrate_dead(unsigned int dead_cpu, struct task_struct *p) | |
6216 | { | |
6217 | struct rq *rq = cpu_rq(dead_cpu); | |
6218 | ||
6219 | /* Must be exiting, otherwise would be on tasklist. */ | |
6220 | BUG_ON(!p->exit_state); | |
6221 | ||
6222 | /* Cannot have done final schedule yet: would have vanished. */ | |
6223 | BUG_ON(p->state == TASK_DEAD); | |
6224 | ||
6225 | get_task_struct(p); | |
6226 | ||
6227 | /* | |
6228 | * Drop lock around migration; if someone else moves it, | |
6229 | * that's OK. No task can be added to this CPU, so iteration is | |
6230 | * fine. | |
6231 | */ | |
6232 | spin_unlock_irq(&rq->lock); | |
6233 | move_task_off_dead_cpu(dead_cpu, p); | |
6234 | spin_lock_irq(&rq->lock); | |
6235 | ||
6236 | put_task_struct(p); | |
6237 | } | |
6238 | ||
6239 | /* release_task() removes task from tasklist, so we won't find dead tasks. */ | |
6240 | static void migrate_dead_tasks(unsigned int dead_cpu) | |
6241 | { | |
6242 | struct rq *rq = cpu_rq(dead_cpu); | |
6243 | struct task_struct *next; | |
6244 | ||
6245 | for ( ; ; ) { | |
6246 | if (!rq->nr_running) | |
6247 | break; | |
6248 | update_rq_clock(rq); | |
6249 | next = pick_next_task(rq, rq->curr); | |
6250 | if (!next) | |
6251 | break; | |
6252 | migrate_dead(dead_cpu, next); | |
6253 | ||
6254 | } | |
6255 | } | |
6256 | #endif /* CONFIG_HOTPLUG_CPU */ | |
6257 | ||
6258 | #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) | |
6259 | ||
6260 | static struct ctl_table sd_ctl_dir[] = { | |
6261 | { | |
6262 | .procname = "sched_domain", | |
6263 | .mode = 0555, | |
6264 | }, | |
6265 | {0, }, | |
6266 | }; | |
6267 | ||
6268 | static struct ctl_table sd_ctl_root[] = { | |
6269 | { | |
6270 | .ctl_name = CTL_KERN, | |
6271 | .procname = "kernel", | |
6272 | .mode = 0555, | |
6273 | .child = sd_ctl_dir, | |
6274 | }, | |
6275 | {0, }, | |
6276 | }; | |
6277 | ||
6278 | static struct ctl_table *sd_alloc_ctl_entry(int n) | |
6279 | { | |
6280 | struct ctl_table *entry = | |
6281 | kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); | |
6282 | ||
6283 | return entry; | |
6284 | } | |
6285 | ||
6286 | static void sd_free_ctl_entry(struct ctl_table **tablep) | |
6287 | { | |
6288 | struct ctl_table *entry; | |
6289 | ||
6290 | /* | |
6291 | * In the intermediate directories, both the child directory and | |
6292 | * procname are dynamically allocated and could fail but the mode | |
6293 | * will always be set. In the lowest directory the names are | |
6294 | * static strings and all have proc handlers. | |
6295 | */ | |
6296 | for (entry = *tablep; entry->mode; entry++) { | |
6297 | if (entry->child) | |
6298 | sd_free_ctl_entry(&entry->child); | |
6299 | if (entry->proc_handler == NULL) | |
6300 | kfree(entry->procname); | |
6301 | } | |
6302 | ||
6303 | kfree(*tablep); | |
6304 | *tablep = NULL; | |
6305 | } | |
6306 | ||
6307 | static void | |
6308 | set_table_entry(struct ctl_table *entry, | |
6309 | const char *procname, void *data, int maxlen, | |
6310 | mode_t mode, proc_handler *proc_handler) | |
6311 | { | |
6312 | entry->procname = procname; | |
6313 | entry->data = data; | |
6314 | entry->maxlen = maxlen; | |
6315 | entry->mode = mode; | |
6316 | entry->proc_handler = proc_handler; | |
6317 | } | |
6318 | ||
6319 | static struct ctl_table * | |
6320 | sd_alloc_ctl_domain_table(struct sched_domain *sd) | |
6321 | { | |
6322 | struct ctl_table *table = sd_alloc_ctl_entry(12); | |
6323 | ||
6324 | if (table == NULL) | |
6325 | return NULL; | |
6326 | ||
6327 | set_table_entry(&table[0], "min_interval", &sd->min_interval, | |
6328 | sizeof(long), 0644, proc_doulongvec_minmax); | |
6329 | set_table_entry(&table[1], "max_interval", &sd->max_interval, | |
6330 | sizeof(long), 0644, proc_doulongvec_minmax); | |
6331 | set_table_entry(&table[2], "busy_idx", &sd->busy_idx, | |
6332 | sizeof(int), 0644, proc_dointvec_minmax); | |
6333 | set_table_entry(&table[3], "idle_idx", &sd->idle_idx, | |
6334 | sizeof(int), 0644, proc_dointvec_minmax); | |
6335 | set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, | |
6336 | sizeof(int), 0644, proc_dointvec_minmax); | |
6337 | set_table_entry(&table[5], "wake_idx", &sd->wake_idx, | |
6338 | sizeof(int), 0644, proc_dointvec_minmax); | |
6339 | set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, | |
6340 | sizeof(int), 0644, proc_dointvec_minmax); | |
6341 | set_table_entry(&table[7], "busy_factor", &sd->busy_factor, | |
6342 | sizeof(int), 0644, proc_dointvec_minmax); | |
6343 | set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, | |
6344 | sizeof(int), 0644, proc_dointvec_minmax); | |
6345 | set_table_entry(&table[9], "cache_nice_tries", | |
6346 | &sd->cache_nice_tries, | |
6347 | sizeof(int), 0644, proc_dointvec_minmax); | |
6348 | set_table_entry(&table[10], "flags", &sd->flags, | |
6349 | sizeof(int), 0644, proc_dointvec_minmax); | |
6350 | /* &table[11] is terminator */ | |
6351 | ||
6352 | return table; | |
6353 | } | |
6354 | ||
6355 | static ctl_table *sd_alloc_ctl_cpu_table(int cpu) | |
6356 | { | |
6357 | struct ctl_table *entry, *table; | |
6358 | struct sched_domain *sd; | |
6359 | int domain_num = 0, i; | |
6360 | char buf[32]; | |
6361 | ||
6362 | for_each_domain(cpu, sd) | |
6363 | domain_num++; | |
6364 | entry = table = sd_alloc_ctl_entry(domain_num + 1); | |
6365 | if (table == NULL) | |
6366 | return NULL; | |
6367 | ||
6368 | i = 0; | |
6369 | for_each_domain(cpu, sd) { | |
6370 | snprintf(buf, 32, "domain%d", i); | |
6371 | entry->procname = kstrdup(buf, GFP_KERNEL); | |
6372 | entry->mode = 0555; | |
6373 | entry->child = sd_alloc_ctl_domain_table(sd); | |
6374 | entry++; | |
6375 | i++; | |
6376 | } | |
6377 | return table; | |
6378 | } | |
6379 | ||
6380 | static struct ctl_table_header *sd_sysctl_header; | |
6381 | static void register_sched_domain_sysctl(void) | |
6382 | { | |
6383 | int i, cpu_num = num_online_cpus(); | |
6384 | struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); | |
6385 | char buf[32]; | |
6386 | ||
6387 | WARN_ON(sd_ctl_dir[0].child); | |
6388 | sd_ctl_dir[0].child = entry; | |
6389 | ||
6390 | if (entry == NULL) | |
6391 | return; | |
6392 | ||
6393 | for_each_online_cpu(i) { | |
6394 | snprintf(buf, 32, "cpu%d", i); | |
6395 | entry->procname = kstrdup(buf, GFP_KERNEL); | |
6396 | entry->mode = 0555; | |
6397 | entry->child = sd_alloc_ctl_cpu_table(i); | |
6398 | entry++; | |
6399 | } | |
6400 | ||
6401 | WARN_ON(sd_sysctl_header); | |
6402 | sd_sysctl_header = register_sysctl_table(sd_ctl_root); | |
6403 | } | |
6404 | ||
6405 | /* may be called multiple times per register */ | |
6406 | static void unregister_sched_domain_sysctl(void) | |
6407 | { | |
6408 | if (sd_sysctl_header) | |
6409 | unregister_sysctl_table(sd_sysctl_header); | |
6410 | sd_sysctl_header = NULL; | |
6411 | if (sd_ctl_dir[0].child) | |
6412 | sd_free_ctl_entry(&sd_ctl_dir[0].child); | |
6413 | } | |
6414 | #else | |
6415 | static void register_sched_domain_sysctl(void) | |
6416 | { | |
6417 | } | |
6418 | static void unregister_sched_domain_sysctl(void) | |
6419 | { | |
6420 | } | |
6421 | #endif | |
6422 | ||
6423 | /* | |
6424 | * migration_call - callback that gets triggered when a CPU is added. | |
6425 | * Here we can start up the necessary migration thread for the new CPU. | |
6426 | */ | |
6427 | static int __cpuinit | |
6428 | migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) | |
6429 | { | |
6430 | struct task_struct *p; | |
6431 | int cpu = (long)hcpu; | |
6432 | unsigned long flags; | |
6433 | struct rq *rq; | |
6434 | ||
6435 | switch (action) { | |
6436 | ||
6437 | case CPU_UP_PREPARE: | |
6438 | case CPU_UP_PREPARE_FROZEN: | |
6439 | p = kthread_create(migration_thread, hcpu, "migration/%d", cpu); | |
6440 | if (IS_ERR(p)) | |
6441 | return NOTIFY_BAD; | |
6442 | kthread_bind(p, cpu); | |
6443 | /* Must be high prio: stop_machine expects to yield to it. */ | |
6444 | rq = task_rq_lock(p, &flags); | |
6445 | __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1); | |
6446 | task_rq_unlock(rq, &flags); | |
6447 | cpu_rq(cpu)->migration_thread = p; | |
6448 | break; | |
6449 | ||
6450 | case CPU_ONLINE: | |
6451 | case CPU_ONLINE_FROZEN: | |
6452 | /* Strictly unnecessary, as first user will wake it. */ | |
6453 | wake_up_process(cpu_rq(cpu)->migration_thread); | |
6454 | ||
6455 | /* Update our root-domain */ | |
6456 | rq = cpu_rq(cpu); | |
6457 | spin_lock_irqsave(&rq->lock, flags); | |
6458 | if (rq->rd) { | |
6459 | BUG_ON(!cpu_isset(cpu, rq->rd->span)); | |
6460 | cpu_set(cpu, rq->rd->online); | |
6461 | } | |
6462 | spin_unlock_irqrestore(&rq->lock, flags); | |
6463 | break; | |
6464 | ||
6465 | #ifdef CONFIG_HOTPLUG_CPU | |
6466 | case CPU_UP_CANCELED: | |
6467 | case CPU_UP_CANCELED_FROZEN: | |
6468 | if (!cpu_rq(cpu)->migration_thread) | |
6469 | break; | |
6470 | /* Unbind it from offline cpu so it can run. Fall thru. */ | |
6471 | kthread_bind(cpu_rq(cpu)->migration_thread, | |
6472 | any_online_cpu(cpu_online_map)); | |
6473 | kthread_stop(cpu_rq(cpu)->migration_thread); | |
6474 | cpu_rq(cpu)->migration_thread = NULL; | |
6475 | break; | |
6476 | ||
6477 | case CPU_DEAD: | |
6478 | case CPU_DEAD_FROZEN: | |
6479 | cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */ | |
6480 | migrate_live_tasks(cpu); | |
6481 | rq = cpu_rq(cpu); | |
6482 | kthread_stop(rq->migration_thread); | |
6483 | rq->migration_thread = NULL; | |
6484 | /* Idle task back to normal (off runqueue, low prio) */ | |
6485 | spin_lock_irq(&rq->lock); | |
6486 | update_rq_clock(rq); | |
6487 | deactivate_task(rq, rq->idle, 0); | |
6488 | rq->idle->static_prio = MAX_PRIO; | |
6489 | __setscheduler(rq, rq->idle, SCHED_NORMAL, 0); | |
6490 | rq->idle->sched_class = &idle_sched_class; | |
6491 | migrate_dead_tasks(cpu); | |
6492 | spin_unlock_irq(&rq->lock); | |
6493 | cpuset_unlock(); | |
6494 | migrate_nr_uninterruptible(rq); | |
6495 | BUG_ON(rq->nr_running != 0); | |
6496 | ||
6497 | /* | |
6498 | * No need to migrate the tasks: it was best-effort if | |
6499 | * they didn't take sched_hotcpu_mutex. Just wake up | |
6500 | * the requestors. | |
6501 | */ | |
6502 | spin_lock_irq(&rq->lock); | |
6503 | while (!list_empty(&rq->migration_queue)) { | |
6504 | struct migration_req *req; | |
6505 | ||
6506 | req = list_entry(rq->migration_queue.next, | |
6507 | struct migration_req, list); | |
6508 | list_del_init(&req->list); | |
6509 | complete(&req->done); | |
6510 | } | |
6511 | spin_unlock_irq(&rq->lock); | |
6512 | break; | |
6513 | ||
6514 | case CPU_DYING: | |
6515 | case CPU_DYING_FROZEN: | |
6516 | /* Update our root-domain */ | |
6517 | rq = cpu_rq(cpu); | |
6518 | spin_lock_irqsave(&rq->lock, flags); | |
6519 | if (rq->rd) { | |
6520 | BUG_ON(!cpu_isset(cpu, rq->rd->span)); | |
6521 | cpu_clear(cpu, rq->rd->online); | |
6522 | } | |
6523 | spin_unlock_irqrestore(&rq->lock, flags); | |
6524 | break; | |
6525 | #endif | |
6526 | } | |
6527 | return NOTIFY_OK; | |
6528 | } | |
6529 | ||
6530 | /* Register at highest priority so that task migration (migrate_all_tasks) | |
6531 | * happens before everything else. | |
6532 | */ | |
6533 | static struct notifier_block __cpuinitdata migration_notifier = { | |
6534 | .notifier_call = migration_call, | |
6535 | .priority = 10 | |
6536 | }; | |
6537 | ||
6538 | void __init migration_init(void) | |
6539 | { | |
6540 | void *cpu = (void *)(long)smp_processor_id(); | |
6541 | int err; | |
6542 | ||
6543 | /* Start one for the boot CPU: */ | |
6544 | err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); | |
6545 | BUG_ON(err == NOTIFY_BAD); | |
6546 | migration_call(&migration_notifier, CPU_ONLINE, cpu); | |
6547 | register_cpu_notifier(&migration_notifier); | |
6548 | } | |
6549 | #endif | |
6550 | ||
6551 | #ifdef CONFIG_SMP | |
6552 | ||
6553 | #ifdef CONFIG_SCHED_DEBUG | |
6554 | ||
6555 | static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, | |
6556 | cpumask_t *groupmask) | |
6557 | { | |
6558 | struct sched_group *group = sd->groups; | |
6559 | char str[256]; | |
6560 | ||
6561 | cpulist_scnprintf(str, sizeof(str), sd->span); | |
6562 | cpus_clear(*groupmask); | |
6563 | ||
6564 | printk(KERN_DEBUG "%*s domain %d: ", level, "", level); | |
6565 | ||
6566 | if (!(sd->flags & SD_LOAD_BALANCE)) { | |
6567 | printk("does not load-balance\n"); | |
6568 | if (sd->parent) | |
6569 | printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" | |
6570 | " has parent"); | |
6571 | return -1; | |
6572 | } | |
6573 | ||
6574 | printk(KERN_CONT "span %s\n", str); | |
6575 | ||
6576 | if (!cpu_isset(cpu, sd->span)) { | |
6577 | printk(KERN_ERR "ERROR: domain->span does not contain " | |
6578 | "CPU%d\n", cpu); | |
6579 | } | |
6580 | if (!cpu_isset(cpu, group->cpumask)) { | |
6581 | printk(KERN_ERR "ERROR: domain->groups does not contain" | |
6582 | " CPU%d\n", cpu); | |
6583 | } | |
6584 | ||
6585 | printk(KERN_DEBUG "%*s groups:", level + 1, ""); | |
6586 | do { | |
6587 | if (!group) { | |
6588 | printk("\n"); | |
6589 | printk(KERN_ERR "ERROR: group is NULL\n"); | |
6590 | break; | |
6591 | } | |
6592 | ||
6593 | if (!group->__cpu_power) { | |
6594 | printk(KERN_CONT "\n"); | |
6595 | printk(KERN_ERR "ERROR: domain->cpu_power not " | |
6596 | "set\n"); | |
6597 | break; | |
6598 | } | |
6599 | ||
6600 | if (!cpus_weight(group->cpumask)) { | |
6601 | printk(KERN_CONT "\n"); | |
6602 | printk(KERN_ERR "ERROR: empty group\n"); | |
6603 | break; | |
6604 | } | |
6605 | ||
6606 | if (cpus_intersects(*groupmask, group->cpumask)) { | |
6607 | printk(KERN_CONT "\n"); | |
6608 | printk(KERN_ERR "ERROR: repeated CPUs\n"); | |
6609 | break; | |
6610 | } | |
6611 | ||
6612 | cpus_or(*groupmask, *groupmask, group->cpumask); | |
6613 | ||
6614 | cpulist_scnprintf(str, sizeof(str), group->cpumask); | |
6615 | printk(KERN_CONT " %s", str); | |
6616 | ||
6617 | group = group->next; | |
6618 | } while (group != sd->groups); | |
6619 | printk(KERN_CONT "\n"); | |
6620 | ||
6621 | if (!cpus_equal(sd->span, *groupmask)) | |
6622 | printk(KERN_ERR "ERROR: groups don't span domain->span\n"); | |
6623 | ||
6624 | if (sd->parent && !cpus_subset(*groupmask, sd->parent->span)) | |
6625 | printk(KERN_ERR "ERROR: parent span is not a superset " | |
6626 | "of domain->span\n"); | |
6627 | return 0; | |
6628 | } | |
6629 | ||
6630 | static void sched_domain_debug(struct sched_domain *sd, int cpu) | |
6631 | { | |
6632 | cpumask_t *groupmask; | |
6633 | int level = 0; | |
6634 | ||
6635 | if (!sd) { | |
6636 | printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); | |
6637 | return; | |
6638 | } | |
6639 | ||
6640 | printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); | |
6641 | ||
6642 | groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL); | |
6643 | if (!groupmask) { | |
6644 | printk(KERN_DEBUG "Cannot load-balance (out of memory)\n"); | |
6645 | return; | |
6646 | } | |
6647 | ||
6648 | for (;;) { | |
6649 | if (sched_domain_debug_one(sd, cpu, level, groupmask)) | |
6650 | break; | |
6651 | level++; | |
6652 | sd = sd->parent; | |
6653 | if (!sd) | |
6654 | break; | |
6655 | } | |
6656 | kfree(groupmask); | |
6657 | } | |
6658 | #else | |
6659 | # define sched_domain_debug(sd, cpu) do { } while (0) | |
6660 | #endif | |
6661 | ||
6662 | static int sd_degenerate(struct sched_domain *sd) | |
6663 | { | |
6664 | if (cpus_weight(sd->span) == 1) | |
6665 | return 1; | |
6666 | ||
6667 | /* Following flags need at least 2 groups */ | |
6668 | if (sd->flags & (SD_LOAD_BALANCE | | |
6669 | SD_BALANCE_NEWIDLE | | |
6670 | SD_BALANCE_FORK | | |
6671 | SD_BALANCE_EXEC | | |
6672 | SD_SHARE_CPUPOWER | | |
6673 | SD_SHARE_PKG_RESOURCES)) { | |
6674 | if (sd->groups != sd->groups->next) | |
6675 | return 0; | |
6676 | } | |
6677 | ||
6678 | /* Following flags don't use groups */ | |
6679 | if (sd->flags & (SD_WAKE_IDLE | | |
6680 | SD_WAKE_AFFINE | | |
6681 | SD_WAKE_BALANCE)) | |
6682 | return 0; | |
6683 | ||
6684 | return 1; | |
6685 | } | |
6686 | ||
6687 | static int | |
6688 | sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) | |
6689 | { | |
6690 | unsigned long cflags = sd->flags, pflags = parent->flags; | |
6691 | ||
6692 | if (sd_degenerate(parent)) | |
6693 | return 1; | |
6694 | ||
6695 | if (!cpus_equal(sd->span, parent->span)) | |
6696 | return 0; | |
6697 | ||
6698 | /* Does parent contain flags not in child? */ | |
6699 | /* WAKE_BALANCE is a subset of WAKE_AFFINE */ | |
6700 | if (cflags & SD_WAKE_AFFINE) | |
6701 | pflags &= ~SD_WAKE_BALANCE; | |
6702 | /* Flags needing groups don't count if only 1 group in parent */ | |
6703 | if (parent->groups == parent->groups->next) { | |
6704 | pflags &= ~(SD_LOAD_BALANCE | | |
6705 | SD_BALANCE_NEWIDLE | | |
6706 | SD_BALANCE_FORK | | |
6707 | SD_BALANCE_EXEC | | |
6708 | SD_SHARE_CPUPOWER | | |
6709 | SD_SHARE_PKG_RESOURCES); | |
6710 | } | |
6711 | if (~cflags & pflags) | |
6712 | return 0; | |
6713 | ||
6714 | return 1; | |
6715 | } | |
6716 | ||
6717 | static void rq_attach_root(struct rq *rq, struct root_domain *rd) | |
6718 | { | |
6719 | unsigned long flags; | |
6720 | const struct sched_class *class; | |
6721 | ||
6722 | spin_lock_irqsave(&rq->lock, flags); | |
6723 | ||
6724 | if (rq->rd) { | |
6725 | struct root_domain *old_rd = rq->rd; | |
6726 | ||
6727 | for (class = sched_class_highest; class; class = class->next) { | |
6728 | if (class->leave_domain) | |
6729 | class->leave_domain(rq); | |
6730 | } | |
6731 | ||
6732 | cpu_clear(rq->cpu, old_rd->span); | |
6733 | cpu_clear(rq->cpu, old_rd->online); | |
6734 | ||
6735 | if (atomic_dec_and_test(&old_rd->refcount)) | |
6736 | kfree(old_rd); | |
6737 | } | |
6738 | ||
6739 | atomic_inc(&rd->refcount); | |
6740 | rq->rd = rd; | |
6741 | ||
6742 | cpu_set(rq->cpu, rd->span); | |
6743 | if (cpu_isset(rq->cpu, cpu_online_map)) | |
6744 | cpu_set(rq->cpu, rd->online); | |
6745 | ||
6746 | for (class = sched_class_highest; class; class = class->next) { | |
6747 | if (class->join_domain) | |
6748 | class->join_domain(rq); | |
6749 | } | |
6750 | ||
6751 | spin_unlock_irqrestore(&rq->lock, flags); | |
6752 | } | |
6753 | ||
6754 | static void init_rootdomain(struct root_domain *rd) | |
6755 | { | |
6756 | memset(rd, 0, sizeof(*rd)); | |
6757 | ||
6758 | cpus_clear(rd->span); | |
6759 | cpus_clear(rd->online); | |
6760 | } | |
6761 | ||
6762 | static void init_defrootdomain(void) | |
6763 | { | |
6764 | init_rootdomain(&def_root_domain); | |
6765 | atomic_set(&def_root_domain.refcount, 1); | |
6766 | } | |
6767 | ||
6768 | static struct root_domain *alloc_rootdomain(void) | |
6769 | { | |
6770 | struct root_domain *rd; | |
6771 | ||
6772 | rd = kmalloc(sizeof(*rd), GFP_KERNEL); | |
6773 | if (!rd) | |
6774 | return NULL; | |
6775 | ||
6776 | init_rootdomain(rd); | |
6777 | ||
6778 | return rd; | |
6779 | } | |
6780 | ||
6781 | /* | |
6782 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must | |
6783 | * hold the hotplug lock. | |
6784 | */ | |
6785 | static void | |
6786 | cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) | |
6787 | { | |
6788 | struct rq *rq = cpu_rq(cpu); | |
6789 | struct sched_domain *tmp; | |
6790 | ||
6791 | /* Remove the sched domains which do not contribute to scheduling. */ | |
6792 | for (tmp = sd; tmp; tmp = tmp->parent) { | |
6793 | struct sched_domain *parent = tmp->parent; | |
6794 | if (!parent) | |
6795 | break; | |
6796 | if (sd_parent_degenerate(tmp, parent)) { | |
6797 | tmp->parent = parent->parent; | |
6798 | if (parent->parent) | |
6799 | parent->parent->child = tmp; | |
6800 | } | |
6801 | } | |
6802 | ||
6803 | if (sd && sd_degenerate(sd)) { | |
6804 | sd = sd->parent; | |
6805 | if (sd) | |
6806 | sd->child = NULL; | |
6807 | } | |
6808 | ||
6809 | sched_domain_debug(sd, cpu); | |
6810 | ||
6811 | rq_attach_root(rq, rd); | |
6812 | rcu_assign_pointer(rq->sd, sd); | |
6813 | } | |
6814 | ||
6815 | /* cpus with isolated domains */ | |
6816 | static cpumask_t cpu_isolated_map = CPU_MASK_NONE; | |
6817 | ||
6818 | /* Setup the mask of cpus configured for isolated domains */ | |
6819 | static int __init isolated_cpu_setup(char *str) | |
6820 | { | |
6821 | int ints[NR_CPUS], i; | |
6822 | ||
6823 | str = get_options(str, ARRAY_SIZE(ints), ints); | |
6824 | cpus_clear(cpu_isolated_map); | |
6825 | for (i = 1; i <= ints[0]; i++) | |
6826 | if (ints[i] < NR_CPUS) | |
6827 | cpu_set(ints[i], cpu_isolated_map); | |
6828 | return 1; | |
6829 | } | |
6830 | ||
6831 | __setup("isolcpus=", isolated_cpu_setup); | |
6832 | ||
6833 | /* | |
6834 | * init_sched_build_groups takes the cpumask we wish to span, and a pointer | |
6835 | * to a function which identifies what group(along with sched group) a CPU | |
6836 | * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS | |
6837 | * (due to the fact that we keep track of groups covered with a cpumask_t). | |
6838 | * | |
6839 | * init_sched_build_groups will build a circular linked list of the groups | |
6840 | * covered by the given span, and will set each group's ->cpumask correctly, | |
6841 | * and ->cpu_power to 0. | |
6842 | */ | |
6843 | static void | |
6844 | init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map, | |
6845 | int (*group_fn)(int cpu, const cpumask_t *cpu_map, | |
6846 | struct sched_group **sg, | |
6847 | cpumask_t *tmpmask), | |
6848 | cpumask_t *covered, cpumask_t *tmpmask) | |
6849 | { | |
6850 | struct sched_group *first = NULL, *last = NULL; | |
6851 | int i; | |
6852 | ||
6853 | cpus_clear(*covered); | |
6854 | ||
6855 | for_each_cpu_mask(i, *span) { | |
6856 | struct sched_group *sg; | |
6857 | int group = group_fn(i, cpu_map, &sg, tmpmask); | |
6858 | int j; | |
6859 | ||
6860 | if (cpu_isset(i, *covered)) | |
6861 | continue; | |
6862 | ||
6863 | cpus_clear(sg->cpumask); | |
6864 | sg->__cpu_power = 0; | |
6865 | ||
6866 | for_each_cpu_mask(j, *span) { | |
6867 | if (group_fn(j, cpu_map, NULL, tmpmask) != group) | |
6868 | continue; | |
6869 | ||
6870 | cpu_set(j, *covered); | |
6871 | cpu_set(j, sg->cpumask); | |
6872 | } | |
6873 | if (!first) | |
6874 | first = sg; | |
6875 | if (last) | |
6876 | last->next = sg; | |
6877 | last = sg; | |
6878 | } | |
6879 | last->next = first; | |
6880 | } | |
6881 | ||
6882 | #define SD_NODES_PER_DOMAIN 16 | |
6883 | ||
6884 | #ifdef CONFIG_NUMA | |
6885 | ||
6886 | /** | |
6887 | * find_next_best_node - find the next node to include in a sched_domain | |
6888 | * @node: node whose sched_domain we're building | |
6889 | * @used_nodes: nodes already in the sched_domain | |
6890 | * | |
6891 | * Find the next node to include in a given scheduling domain. Simply | |
6892 | * finds the closest node not already in the @used_nodes map. | |
6893 | * | |
6894 | * Should use nodemask_t. | |
6895 | */ | |
6896 | static int find_next_best_node(int node, nodemask_t *used_nodes) | |
6897 | { | |
6898 | int i, n, val, min_val, best_node = 0; | |
6899 | ||
6900 | min_val = INT_MAX; | |
6901 | ||
6902 | for (i = 0; i < MAX_NUMNODES; i++) { | |
6903 | /* Start at @node */ | |
6904 | n = (node + i) % MAX_NUMNODES; | |
6905 | ||
6906 | if (!nr_cpus_node(n)) | |
6907 | continue; | |
6908 | ||
6909 | /* Skip already used nodes */ | |
6910 | if (node_isset(n, *used_nodes)) | |
6911 | continue; | |
6912 | ||
6913 | /* Simple min distance search */ | |
6914 | val = node_distance(node, n); | |
6915 | ||
6916 | if (val < min_val) { | |
6917 | min_val = val; | |
6918 | best_node = n; | |
6919 | } | |
6920 | } | |
6921 | ||
6922 | node_set(best_node, *used_nodes); | |
6923 | return best_node; | |
6924 | } | |
6925 | ||
6926 | /** | |
6927 | * sched_domain_node_span - get a cpumask for a node's sched_domain | |
6928 | * @node: node whose cpumask we're constructing | |
6929 | * | |
6930 | * Given a node, construct a good cpumask for its sched_domain to span. It | |
6931 | * should be one that prevents unnecessary balancing, but also spreads tasks | |
6932 | * out optimally. | |
6933 | */ | |
6934 | static void sched_domain_node_span(int node, cpumask_t *span) | |
6935 | { | |
6936 | nodemask_t used_nodes; | |
6937 | node_to_cpumask_ptr(nodemask, node); | |
6938 | int i; | |
6939 | ||
6940 | cpus_clear(*span); | |
6941 | nodes_clear(used_nodes); | |
6942 | ||
6943 | cpus_or(*span, *span, *nodemask); | |
6944 | node_set(node, used_nodes); | |
6945 | ||
6946 | for (i = 1; i < SD_NODES_PER_DOMAIN; i++) { | |
6947 | int next_node = find_next_best_node(node, &used_nodes); | |
6948 | ||
6949 | node_to_cpumask_ptr_next(nodemask, next_node); | |
6950 | cpus_or(*span, *span, *nodemask); | |
6951 | } | |
6952 | } | |
6953 | #endif | |
6954 | ||
6955 | int sched_smt_power_savings = 0, sched_mc_power_savings = 0; | |
6956 | ||
6957 | /* | |
6958 | * SMT sched-domains: | |
6959 | */ | |
6960 | #ifdef CONFIG_SCHED_SMT | |
6961 | static DEFINE_PER_CPU(struct sched_domain, cpu_domains); | |
6962 | static DEFINE_PER_CPU(struct sched_group, sched_group_cpus); | |
6963 | ||
6964 | static int | |
6965 | cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg, | |
6966 | cpumask_t *unused) | |
6967 | { | |
6968 | if (sg) | |
6969 | *sg = &per_cpu(sched_group_cpus, cpu); | |
6970 | return cpu; | |
6971 | } | |
6972 | #endif | |
6973 | ||
6974 | /* | |
6975 | * multi-core sched-domains: | |
6976 | */ | |
6977 | #ifdef CONFIG_SCHED_MC | |
6978 | static DEFINE_PER_CPU(struct sched_domain, core_domains); | |
6979 | static DEFINE_PER_CPU(struct sched_group, sched_group_core); | |
6980 | #endif | |
6981 | ||
6982 | #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT) | |
6983 | static int | |
6984 | cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg, | |
6985 | cpumask_t *mask) | |
6986 | { | |
6987 | int group; | |
6988 | ||
6989 | *mask = per_cpu(cpu_sibling_map, cpu); | |
6990 | cpus_and(*mask, *mask, *cpu_map); | |
6991 | group = first_cpu(*mask); | |
6992 | if (sg) | |
6993 | *sg = &per_cpu(sched_group_core, group); | |
6994 | return group; | |
6995 | } | |
6996 | #elif defined(CONFIG_SCHED_MC) | |
6997 | static int | |
6998 | cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg, | |
6999 | cpumask_t *unused) | |
7000 | { | |
7001 | if (sg) | |
7002 | *sg = &per_cpu(sched_group_core, cpu); | |
7003 | return cpu; | |
7004 | } | |
7005 | #endif | |
7006 | ||
7007 | static DEFINE_PER_CPU(struct sched_domain, phys_domains); | |
7008 | static DEFINE_PER_CPU(struct sched_group, sched_group_phys); | |
7009 | ||
7010 | static int | |
7011 | cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg, | |
7012 | cpumask_t *mask) | |
7013 | { | |
7014 | int group; | |
7015 | #ifdef CONFIG_SCHED_MC | |
7016 | *mask = cpu_coregroup_map(cpu); | |
7017 | cpus_and(*mask, *mask, *cpu_map); | |
7018 | group = first_cpu(*mask); | |
7019 | #elif defined(CONFIG_SCHED_SMT) | |
7020 | *mask = per_cpu(cpu_sibling_map, cpu); | |
7021 | cpus_and(*mask, *mask, *cpu_map); | |
7022 | group = first_cpu(*mask); | |
7023 | #else | |
7024 | group = cpu; | |
7025 | #endif | |
7026 | if (sg) | |
7027 | *sg = &per_cpu(sched_group_phys, group); | |
7028 | return group; | |
7029 | } | |
7030 | ||
7031 | #ifdef CONFIG_NUMA | |
7032 | /* | |
7033 | * The init_sched_build_groups can't handle what we want to do with node | |
7034 | * groups, so roll our own. Now each node has its own list of groups which | |
7035 | * gets dynamically allocated. | |
7036 | */ | |
7037 | static DEFINE_PER_CPU(struct sched_domain, node_domains); | |
7038 | static struct sched_group ***sched_group_nodes_bycpu; | |
7039 | ||
7040 | static DEFINE_PER_CPU(struct sched_domain, allnodes_domains); | |
7041 | static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes); | |
7042 | ||
7043 | static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map, | |
7044 | struct sched_group **sg, cpumask_t *nodemask) | |
7045 | { | |
7046 | int group; | |
7047 | ||
7048 | *nodemask = node_to_cpumask(cpu_to_node(cpu)); | |
7049 | cpus_and(*nodemask, *nodemask, *cpu_map); | |
7050 | group = first_cpu(*nodemask); | |
7051 | ||
7052 | if (sg) | |
7053 | *sg = &per_cpu(sched_group_allnodes, group); | |
7054 | return group; | |
7055 | } | |
7056 | ||
7057 | static void init_numa_sched_groups_power(struct sched_group *group_head) | |
7058 | { | |
7059 | struct sched_group *sg = group_head; | |
7060 | int j; | |
7061 | ||
7062 | if (!sg) | |
7063 | return; | |
7064 | do { | |
7065 | for_each_cpu_mask(j, sg->cpumask) { | |
7066 | struct sched_domain *sd; | |
7067 | ||
7068 | sd = &per_cpu(phys_domains, j); | |
7069 | if (j != first_cpu(sd->groups->cpumask)) { | |
7070 | /* | |
7071 | * Only add "power" once for each | |
7072 | * physical package. | |
7073 | */ | |
7074 | continue; | |
7075 | } | |
7076 | ||
7077 | sg_inc_cpu_power(sg, sd->groups->__cpu_power); | |
7078 | } | |
7079 | sg = sg->next; | |
7080 | } while (sg != group_head); | |
7081 | } | |
7082 | #endif | |
7083 | ||
7084 | #ifdef CONFIG_NUMA | |
7085 | /* Free memory allocated for various sched_group structures */ | |
7086 | static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask) | |
7087 | { | |
7088 | int cpu, i; | |
7089 | ||
7090 | for_each_cpu_mask(cpu, *cpu_map) { | |
7091 | struct sched_group **sched_group_nodes | |
7092 | = sched_group_nodes_bycpu[cpu]; | |
7093 | ||
7094 | if (!sched_group_nodes) | |
7095 | continue; | |
7096 | ||
7097 | for (i = 0; i < MAX_NUMNODES; i++) { | |
7098 | struct sched_group *oldsg, *sg = sched_group_nodes[i]; | |
7099 | ||
7100 | *nodemask = node_to_cpumask(i); | |
7101 | cpus_and(*nodemask, *nodemask, *cpu_map); | |
7102 | if (cpus_empty(*nodemask)) | |
7103 | continue; | |
7104 | ||
7105 | if (sg == NULL) | |
7106 | continue; | |
7107 | sg = sg->next; | |
7108 | next_sg: | |
7109 | oldsg = sg; | |
7110 | sg = sg->next; | |
7111 | kfree(oldsg); | |
7112 | if (oldsg != sched_group_nodes[i]) | |
7113 | goto next_sg; | |
7114 | } | |
7115 | kfree(sched_group_nodes); | |
7116 | sched_group_nodes_bycpu[cpu] = NULL; | |
7117 | } | |
7118 | } | |
7119 | #else | |
7120 | static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask) | |
7121 | { | |
7122 | } | |
7123 | #endif | |
7124 | ||
7125 | /* | |
7126 | * Initialize sched groups cpu_power. | |
7127 | * | |
7128 | * cpu_power indicates the capacity of sched group, which is used while | |
7129 | * distributing the load between different sched groups in a sched domain. | |
7130 | * Typically cpu_power for all the groups in a sched domain will be same unless | |
7131 | * there are asymmetries in the topology. If there are asymmetries, group | |
7132 | * having more cpu_power will pickup more load compared to the group having | |
7133 | * less cpu_power. | |
7134 | * | |
7135 | * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents | |
7136 | * the maximum number of tasks a group can handle in the presence of other idle | |
7137 | * or lightly loaded groups in the same sched domain. | |
7138 | */ | |
7139 | static void init_sched_groups_power(int cpu, struct sched_domain *sd) | |
7140 | { | |
7141 | struct sched_domain *child; | |
7142 | struct sched_group *group; | |
7143 | ||
7144 | WARN_ON(!sd || !sd->groups); | |
7145 | ||
7146 | if (cpu != first_cpu(sd->groups->cpumask)) | |
7147 | return; | |
7148 | ||
7149 | child = sd->child; | |
7150 | ||
7151 | sd->groups->__cpu_power = 0; | |
7152 | ||
7153 | /* | |
7154 | * For perf policy, if the groups in child domain share resources | |
7155 | * (for example cores sharing some portions of the cache hierarchy | |
7156 | * or SMT), then set this domain groups cpu_power such that each group | |
7157 | * can handle only one task, when there are other idle groups in the | |
7158 | * same sched domain. | |
7159 | */ | |
7160 | if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) && | |
7161 | (child->flags & | |
7162 | (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) { | |
7163 | sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE); | |
7164 | return; | |
7165 | } | |
7166 | ||
7167 | /* | |
7168 | * add cpu_power of each child group to this groups cpu_power | |
7169 | */ | |
7170 | group = child->groups; | |
7171 | do { | |
7172 | sg_inc_cpu_power(sd->groups, group->__cpu_power); | |
7173 | group = group->next; | |
7174 | } while (group != child->groups); | |
7175 | } | |
7176 | ||
7177 | /* | |
7178 | * Initializers for schedule domains | |
7179 | * Non-inlined to reduce accumulated stack pressure in build_sched_domains() | |
7180 | */ | |
7181 | ||
7182 | #define SD_INIT(sd, type) sd_init_##type(sd) | |
7183 | #define SD_INIT_FUNC(type) \ | |
7184 | static noinline void sd_init_##type(struct sched_domain *sd) \ | |
7185 | { \ | |
7186 | memset(sd, 0, sizeof(*sd)); \ | |
7187 | *sd = SD_##type##_INIT; \ | |
7188 | sd->level = SD_LV_##type; \ | |
7189 | } | |
7190 | ||
7191 | SD_INIT_FUNC(CPU) | |
7192 | #ifdef CONFIG_NUMA | |
7193 | SD_INIT_FUNC(ALLNODES) | |
7194 | SD_INIT_FUNC(NODE) | |
7195 | #endif | |
7196 | #ifdef CONFIG_SCHED_SMT | |
7197 | SD_INIT_FUNC(SIBLING) | |
7198 | #endif | |
7199 | #ifdef CONFIG_SCHED_MC | |
7200 | SD_INIT_FUNC(MC) | |
7201 | #endif | |
7202 | ||
7203 | /* | |
7204 | * To minimize stack usage kmalloc room for cpumasks and share the | |
7205 | * space as the usage in build_sched_domains() dictates. Used only | |
7206 | * if the amount of space is significant. | |
7207 | */ | |
7208 | struct allmasks { | |
7209 | cpumask_t tmpmask; /* make this one first */ | |
7210 | union { | |
7211 | cpumask_t nodemask; | |
7212 | cpumask_t this_sibling_map; | |
7213 | cpumask_t this_core_map; | |
7214 | }; | |
7215 | cpumask_t send_covered; | |
7216 | ||
7217 | #ifdef CONFIG_NUMA | |
7218 | cpumask_t domainspan; | |
7219 | cpumask_t covered; | |
7220 | cpumask_t notcovered; | |
7221 | #endif | |
7222 | }; | |
7223 | ||
7224 | #if NR_CPUS > 128 | |
7225 | #define SCHED_CPUMASK_ALLOC 1 | |
7226 | #define SCHED_CPUMASK_FREE(v) kfree(v) | |
7227 | #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v | |
7228 | #else | |
7229 | #define SCHED_CPUMASK_ALLOC 0 | |
7230 | #define SCHED_CPUMASK_FREE(v) | |
7231 | #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v | |
7232 | #endif | |
7233 | ||
7234 | #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \ | |
7235 | ((unsigned long)(a) + offsetof(struct allmasks, v)) | |
7236 | ||
7237 | static int default_relax_domain_level = -1; | |
7238 | ||
7239 | static int __init setup_relax_domain_level(char *str) | |
7240 | { | |
7241 | default_relax_domain_level = simple_strtoul(str, NULL, 0); | |
7242 | return 1; | |
7243 | } | |
7244 | __setup("relax_domain_level=", setup_relax_domain_level); | |
7245 | ||
7246 | static void set_domain_attribute(struct sched_domain *sd, | |
7247 | struct sched_domain_attr *attr) | |
7248 | { | |
7249 | int request; | |
7250 | ||
7251 | if (!attr || attr->relax_domain_level < 0) { | |
7252 | if (default_relax_domain_level < 0) | |
7253 | return; | |
7254 | else | |
7255 | request = default_relax_domain_level; | |
7256 | } else | |
7257 | request = attr->relax_domain_level; | |
7258 | if (request < sd->level) { | |
7259 | /* turn off idle balance on this domain */ | |
7260 | sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE); | |
7261 | } else { | |
7262 | /* turn on idle balance on this domain */ | |
7263 | sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE); | |
7264 | } | |
7265 | } | |
7266 | ||
7267 | /* | |
7268 | * Build sched domains for a given set of cpus and attach the sched domains | |
7269 | * to the individual cpus | |
7270 | */ | |
7271 | static int __build_sched_domains(const cpumask_t *cpu_map, | |
7272 | struct sched_domain_attr *attr) | |
7273 | { | |
7274 | int i; | |
7275 | struct root_domain *rd; | |
7276 | SCHED_CPUMASK_DECLARE(allmasks); | |
7277 | cpumask_t *tmpmask; | |
7278 | #ifdef CONFIG_NUMA | |
7279 | struct sched_group **sched_group_nodes = NULL; | |
7280 | int sd_allnodes = 0; | |
7281 | ||
7282 | /* | |
7283 | * Allocate the per-node list of sched groups | |
7284 | */ | |
7285 | sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *), | |
7286 | GFP_KERNEL); | |
7287 | if (!sched_group_nodes) { | |
7288 | printk(KERN_WARNING "Can not alloc sched group node list\n"); | |
7289 | return -ENOMEM; | |
7290 | } | |
7291 | #endif | |
7292 | ||
7293 | rd = alloc_rootdomain(); | |
7294 | if (!rd) { | |
7295 | printk(KERN_WARNING "Cannot alloc root domain\n"); | |
7296 | #ifdef CONFIG_NUMA | |
7297 | kfree(sched_group_nodes); | |
7298 | #endif | |
7299 | return -ENOMEM; | |
7300 | } | |
7301 | ||
7302 | #if SCHED_CPUMASK_ALLOC | |
7303 | /* get space for all scratch cpumask variables */ | |
7304 | allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL); | |
7305 | if (!allmasks) { | |
7306 | printk(KERN_WARNING "Cannot alloc cpumask array\n"); | |
7307 | kfree(rd); | |
7308 | #ifdef CONFIG_NUMA | |
7309 | kfree(sched_group_nodes); | |
7310 | #endif | |
7311 | return -ENOMEM; | |
7312 | } | |
7313 | #endif | |
7314 | tmpmask = (cpumask_t *)allmasks; | |
7315 | ||
7316 | ||
7317 | #ifdef CONFIG_NUMA | |
7318 | sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes; | |
7319 | #endif | |
7320 | ||
7321 | /* | |
7322 | * Set up domains for cpus specified by the cpu_map. | |
7323 | */ | |
7324 | for_each_cpu_mask(i, *cpu_map) { | |
7325 | struct sched_domain *sd = NULL, *p; | |
7326 | SCHED_CPUMASK_VAR(nodemask, allmasks); | |
7327 | ||
7328 | *nodemask = node_to_cpumask(cpu_to_node(i)); | |
7329 | cpus_and(*nodemask, *nodemask, *cpu_map); | |
7330 | ||
7331 | #ifdef CONFIG_NUMA | |
7332 | if (cpus_weight(*cpu_map) > | |
7333 | SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) { | |
7334 | sd = &per_cpu(allnodes_domains, i); | |
7335 | SD_INIT(sd, ALLNODES); | |
7336 | set_domain_attribute(sd, attr); | |
7337 | sd->span = *cpu_map; | |
7338 | sd->first_cpu = first_cpu(sd->span); | |
7339 | cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask); | |
7340 | p = sd; | |
7341 | sd_allnodes = 1; | |
7342 | } else | |
7343 | p = NULL; | |
7344 | ||
7345 | sd = &per_cpu(node_domains, i); | |
7346 | SD_INIT(sd, NODE); | |
7347 | set_domain_attribute(sd, attr); | |
7348 | sched_domain_node_span(cpu_to_node(i), &sd->span); | |
7349 | sd->first_cpu = first_cpu(sd->span); | |
7350 | sd->parent = p; | |
7351 | if (p) | |
7352 | p->child = sd; | |
7353 | cpus_and(sd->span, sd->span, *cpu_map); | |
7354 | #endif | |
7355 | ||
7356 | p = sd; | |
7357 | sd = &per_cpu(phys_domains, i); | |
7358 | SD_INIT(sd, CPU); | |
7359 | set_domain_attribute(sd, attr); | |
7360 | sd->span = *nodemask; | |
7361 | sd->first_cpu = first_cpu(sd->span); | |
7362 | sd->parent = p; | |
7363 | if (p) | |
7364 | p->child = sd; | |
7365 | cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask); | |
7366 | ||
7367 | #ifdef CONFIG_SCHED_MC | |
7368 | p = sd; | |
7369 | sd = &per_cpu(core_domains, i); | |
7370 | SD_INIT(sd, MC); | |
7371 | set_domain_attribute(sd, attr); | |
7372 | sd->span = cpu_coregroup_map(i); | |
7373 | sd->first_cpu = first_cpu(sd->span); | |
7374 | cpus_and(sd->span, sd->span, *cpu_map); | |
7375 | sd->parent = p; | |
7376 | p->child = sd; | |
7377 | cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask); | |
7378 | #endif | |
7379 | ||
7380 | #ifdef CONFIG_SCHED_SMT | |
7381 | p = sd; | |
7382 | sd = &per_cpu(cpu_domains, i); | |
7383 | SD_INIT(sd, SIBLING); | |
7384 | set_domain_attribute(sd, attr); | |
7385 | sd->span = per_cpu(cpu_sibling_map, i); | |
7386 | sd->first_cpu = first_cpu(sd->span); | |
7387 | cpus_and(sd->span, sd->span, *cpu_map); | |
7388 | sd->parent = p; | |
7389 | p->child = sd; | |
7390 | cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask); | |
7391 | #endif | |
7392 | } | |
7393 | ||
7394 | #ifdef CONFIG_SCHED_SMT | |
7395 | /* Set up CPU (sibling) groups */ | |
7396 | for_each_cpu_mask(i, *cpu_map) { | |
7397 | SCHED_CPUMASK_VAR(this_sibling_map, allmasks); | |
7398 | SCHED_CPUMASK_VAR(send_covered, allmasks); | |
7399 | ||
7400 | *this_sibling_map = per_cpu(cpu_sibling_map, i); | |
7401 | cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map); | |
7402 | if (i != first_cpu(*this_sibling_map)) | |
7403 | continue; | |
7404 | ||
7405 | init_sched_build_groups(this_sibling_map, cpu_map, | |
7406 | &cpu_to_cpu_group, | |
7407 | send_covered, tmpmask); | |
7408 | } | |
7409 | #endif | |
7410 | ||
7411 | #ifdef CONFIG_SCHED_MC | |
7412 | /* Set up multi-core groups */ | |
7413 | for_each_cpu_mask(i, *cpu_map) { | |
7414 | SCHED_CPUMASK_VAR(this_core_map, allmasks); | |
7415 | SCHED_CPUMASK_VAR(send_covered, allmasks); | |
7416 | ||
7417 | *this_core_map = cpu_coregroup_map(i); | |
7418 | cpus_and(*this_core_map, *this_core_map, *cpu_map); | |
7419 | if (i != first_cpu(*this_core_map)) | |
7420 | continue; | |
7421 | ||
7422 | init_sched_build_groups(this_core_map, cpu_map, | |
7423 | &cpu_to_core_group, | |
7424 | send_covered, tmpmask); | |
7425 | } | |
7426 | #endif | |
7427 | ||
7428 | /* Set up physical groups */ | |
7429 | for (i = 0; i < MAX_NUMNODES; i++) { | |
7430 | SCHED_CPUMASK_VAR(nodemask, allmasks); | |
7431 | SCHED_CPUMASK_VAR(send_covered, allmasks); | |
7432 | ||
7433 | *nodemask = node_to_cpumask(i); | |
7434 | cpus_and(*nodemask, *nodemask, *cpu_map); | |
7435 | if (cpus_empty(*nodemask)) | |
7436 | continue; | |
7437 | ||
7438 | init_sched_build_groups(nodemask, cpu_map, | |
7439 | &cpu_to_phys_group, | |
7440 | send_covered, tmpmask); | |
7441 | } | |
7442 | ||
7443 | #ifdef CONFIG_NUMA | |
7444 | /* Set up node groups */ | |
7445 | if (sd_allnodes) { | |
7446 | SCHED_CPUMASK_VAR(send_covered, allmasks); | |
7447 | ||
7448 | init_sched_build_groups(cpu_map, cpu_map, | |
7449 | &cpu_to_allnodes_group, | |
7450 | send_covered, tmpmask); | |
7451 | } | |
7452 | ||
7453 | for (i = 0; i < MAX_NUMNODES; i++) { | |
7454 | /* Set up node groups */ | |
7455 | struct sched_group *sg, *prev; | |
7456 | SCHED_CPUMASK_VAR(nodemask, allmasks); | |
7457 | SCHED_CPUMASK_VAR(domainspan, allmasks); | |
7458 | SCHED_CPUMASK_VAR(covered, allmasks); | |
7459 | int j; | |
7460 | ||
7461 | *nodemask = node_to_cpumask(i); | |
7462 | cpus_clear(*covered); | |
7463 | ||
7464 | cpus_and(*nodemask, *nodemask, *cpu_map); | |
7465 | if (cpus_empty(*nodemask)) { | |
7466 | sched_group_nodes[i] = NULL; | |
7467 | continue; | |
7468 | } | |
7469 | ||
7470 | sched_domain_node_span(i, domainspan); | |
7471 | cpus_and(*domainspan, *domainspan, *cpu_map); | |
7472 | ||
7473 | sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i); | |
7474 | if (!sg) { | |
7475 | printk(KERN_WARNING "Can not alloc domain group for " | |
7476 | "node %d\n", i); | |
7477 | goto error; | |
7478 | } | |
7479 | sched_group_nodes[i] = sg; | |
7480 | for_each_cpu_mask(j, *nodemask) { | |
7481 | struct sched_domain *sd; | |
7482 | ||
7483 | sd = &per_cpu(node_domains, j); | |
7484 | sd->groups = sg; | |
7485 | } | |
7486 | sg->__cpu_power = 0; | |
7487 | sg->cpumask = *nodemask; | |
7488 | sg->next = sg; | |
7489 | cpus_or(*covered, *covered, *nodemask); | |
7490 | prev = sg; | |
7491 | ||
7492 | for (j = 0; j < MAX_NUMNODES; j++) { | |
7493 | SCHED_CPUMASK_VAR(notcovered, allmasks); | |
7494 | int n = (i + j) % MAX_NUMNODES; | |
7495 | node_to_cpumask_ptr(pnodemask, n); | |
7496 | ||
7497 | cpus_complement(*notcovered, *covered); | |
7498 | cpus_and(*tmpmask, *notcovered, *cpu_map); | |
7499 | cpus_and(*tmpmask, *tmpmask, *domainspan); | |
7500 | if (cpus_empty(*tmpmask)) | |
7501 | break; | |
7502 | ||
7503 | cpus_and(*tmpmask, *tmpmask, *pnodemask); | |
7504 | if (cpus_empty(*tmpmask)) | |
7505 | continue; | |
7506 | ||
7507 | sg = kmalloc_node(sizeof(struct sched_group), | |
7508 | GFP_KERNEL, i); | |
7509 | if (!sg) { | |
7510 | printk(KERN_WARNING | |
7511 | "Can not alloc domain group for node %d\n", j); | |
7512 | goto error; | |
7513 | } | |
7514 | sg->__cpu_power = 0; | |
7515 | sg->cpumask = *tmpmask; | |
7516 | sg->next = prev->next; | |
7517 | cpus_or(*covered, *covered, *tmpmask); | |
7518 | prev->next = sg; | |
7519 | prev = sg; | |
7520 | } | |
7521 | } | |
7522 | #endif | |
7523 | ||
7524 | /* Calculate CPU power for physical packages and nodes */ | |
7525 | #ifdef CONFIG_SCHED_SMT | |
7526 | for_each_cpu_mask(i, *cpu_map) { | |
7527 | struct sched_domain *sd = &per_cpu(cpu_domains, i); | |
7528 | ||
7529 | init_sched_groups_power(i, sd); | |
7530 | } | |
7531 | #endif | |
7532 | #ifdef CONFIG_SCHED_MC | |
7533 | for_each_cpu_mask(i, *cpu_map) { | |
7534 | struct sched_domain *sd = &per_cpu(core_domains, i); | |
7535 | ||
7536 | init_sched_groups_power(i, sd); | |
7537 | } | |
7538 | #endif | |
7539 | ||
7540 | for_each_cpu_mask(i, *cpu_map) { | |
7541 | struct sched_domain *sd = &per_cpu(phys_domains, i); | |
7542 | ||
7543 | init_sched_groups_power(i, sd); | |
7544 | } | |
7545 | ||
7546 | #ifdef CONFIG_NUMA | |
7547 | for (i = 0; i < MAX_NUMNODES; i++) | |
7548 | init_numa_sched_groups_power(sched_group_nodes[i]); | |
7549 | ||
7550 | if (sd_allnodes) { | |
7551 | struct sched_group *sg; | |
7552 | ||
7553 | cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg, | |
7554 | tmpmask); | |
7555 | init_numa_sched_groups_power(sg); | |
7556 | } | |
7557 | #endif | |
7558 | ||
7559 | /* Attach the domains */ | |
7560 | for_each_cpu_mask(i, *cpu_map) { | |
7561 | struct sched_domain *sd; | |
7562 | #ifdef CONFIG_SCHED_SMT | |
7563 | sd = &per_cpu(cpu_domains, i); | |
7564 | #elif defined(CONFIG_SCHED_MC) | |
7565 | sd = &per_cpu(core_domains, i); | |
7566 | #else | |
7567 | sd = &per_cpu(phys_domains, i); | |
7568 | #endif | |
7569 | cpu_attach_domain(sd, rd, i); | |
7570 | } | |
7571 | ||
7572 | SCHED_CPUMASK_FREE((void *)allmasks); | |
7573 | return 0; | |
7574 | ||
7575 | #ifdef CONFIG_NUMA | |
7576 | error: | |
7577 | free_sched_groups(cpu_map, tmpmask); | |
7578 | SCHED_CPUMASK_FREE((void *)allmasks); | |
7579 | return -ENOMEM; | |
7580 | #endif | |
7581 | } | |
7582 | ||
7583 | static int build_sched_domains(const cpumask_t *cpu_map) | |
7584 | { | |
7585 | return __build_sched_domains(cpu_map, NULL); | |
7586 | } | |
7587 | ||
7588 | static cpumask_t *doms_cur; /* current sched domains */ | |
7589 | static int ndoms_cur; /* number of sched domains in 'doms_cur' */ | |
7590 | static struct sched_domain_attr *dattr_cur; /* attribues of custom domains | |
7591 | in 'doms_cur' */ | |
7592 | ||
7593 | /* | |
7594 | * Special case: If a kmalloc of a doms_cur partition (array of | |
7595 | * cpumask_t) fails, then fallback to a single sched domain, | |
7596 | * as determined by the single cpumask_t fallback_doms. | |
7597 | */ | |
7598 | static cpumask_t fallback_doms; | |
7599 | ||
7600 | void __attribute__((weak)) arch_update_cpu_topology(void) | |
7601 | { | |
7602 | } | |
7603 | ||
7604 | /* | |
7605 | * Set up scheduler domains and groups. Callers must hold the hotplug lock. | |
7606 | * For now this just excludes isolated cpus, but could be used to | |
7607 | * exclude other special cases in the future. | |
7608 | */ | |
7609 | static int arch_init_sched_domains(const cpumask_t *cpu_map) | |
7610 | { | |
7611 | int err; | |
7612 | ||
7613 | arch_update_cpu_topology(); | |
7614 | ndoms_cur = 1; | |
7615 | doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL); | |
7616 | if (!doms_cur) | |
7617 | doms_cur = &fallback_doms; | |
7618 | cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map); | |
7619 | dattr_cur = NULL; | |
7620 | err = build_sched_domains(doms_cur); | |
7621 | register_sched_domain_sysctl(); | |
7622 | ||
7623 | return err; | |
7624 | } | |
7625 | ||
7626 | static void arch_destroy_sched_domains(const cpumask_t *cpu_map, | |
7627 | cpumask_t *tmpmask) | |
7628 | { | |
7629 | free_sched_groups(cpu_map, tmpmask); | |
7630 | } | |
7631 | ||
7632 | /* | |
7633 | * Detach sched domains from a group of cpus specified in cpu_map | |
7634 | * These cpus will now be attached to the NULL domain | |
7635 | */ | |
7636 | static void detach_destroy_domains(const cpumask_t *cpu_map) | |
7637 | { | |
7638 | cpumask_t tmpmask; | |
7639 | int i; | |
7640 | ||
7641 | unregister_sched_domain_sysctl(); | |
7642 | ||
7643 | for_each_cpu_mask(i, *cpu_map) | |
7644 | cpu_attach_domain(NULL, &def_root_domain, i); | |
7645 | synchronize_sched(); | |
7646 | arch_destroy_sched_domains(cpu_map, &tmpmask); | |
7647 | } | |
7648 | ||
7649 | /* handle null as "default" */ | |
7650 | static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, | |
7651 | struct sched_domain_attr *new, int idx_new) | |
7652 | { | |
7653 | struct sched_domain_attr tmp; | |
7654 | ||
7655 | /* fast path */ | |
7656 | if (!new && !cur) | |
7657 | return 1; | |
7658 | ||
7659 | tmp = SD_ATTR_INIT; | |
7660 | return !memcmp(cur ? (cur + idx_cur) : &tmp, | |
7661 | new ? (new + idx_new) : &tmp, | |
7662 | sizeof(struct sched_domain_attr)); | |
7663 | } | |
7664 | ||
7665 | /* | |
7666 | * Partition sched domains as specified by the 'ndoms_new' | |
7667 | * cpumasks in the array doms_new[] of cpumasks. This compares | |
7668 | * doms_new[] to the current sched domain partitioning, doms_cur[]. | |
7669 | * It destroys each deleted domain and builds each new domain. | |
7670 | * | |
7671 | * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'. | |
7672 | * The masks don't intersect (don't overlap.) We should setup one | |
7673 | * sched domain for each mask. CPUs not in any of the cpumasks will | |
7674 | * not be load balanced. If the same cpumask appears both in the | |
7675 | * current 'doms_cur' domains and in the new 'doms_new', we can leave | |
7676 | * it as it is. | |
7677 | * | |
7678 | * The passed in 'doms_new' should be kmalloc'd. This routine takes | |
7679 | * ownership of it and will kfree it when done with it. If the caller | |
7680 | * failed the kmalloc call, then it can pass in doms_new == NULL, | |
7681 | * and partition_sched_domains() will fallback to the single partition | |
7682 | * 'fallback_doms'. | |
7683 | * | |
7684 | * Call with hotplug lock held | |
7685 | */ | |
7686 | void partition_sched_domains(int ndoms_new, cpumask_t *doms_new, | |
7687 | struct sched_domain_attr *dattr_new) | |
7688 | { | |
7689 | int i, j; | |
7690 | ||
7691 | lock_doms_cur(); | |
7692 | ||
7693 | /* always unregister in case we don't destroy any domains */ | |
7694 | unregister_sched_domain_sysctl(); | |
7695 | ||
7696 | if (doms_new == NULL) { | |
7697 | ndoms_new = 1; | |
7698 | doms_new = &fallback_doms; | |
7699 | cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map); | |
7700 | dattr_new = NULL; | |
7701 | } | |
7702 | ||
7703 | /* Destroy deleted domains */ | |
7704 | for (i = 0; i < ndoms_cur; i++) { | |
7705 | for (j = 0; j < ndoms_new; j++) { | |
7706 | if (cpus_equal(doms_cur[i], doms_new[j]) | |
7707 | && dattrs_equal(dattr_cur, i, dattr_new, j)) | |
7708 | goto match1; | |
7709 | } | |
7710 | /* no match - a current sched domain not in new doms_new[] */ | |
7711 | detach_destroy_domains(doms_cur + i); | |
7712 | match1: | |
7713 | ; | |
7714 | } | |
7715 | ||
7716 | /* Build new domains */ | |
7717 | for (i = 0; i < ndoms_new; i++) { | |
7718 | for (j = 0; j < ndoms_cur; j++) { | |
7719 | if (cpus_equal(doms_new[i], doms_cur[j]) | |
7720 | && dattrs_equal(dattr_new, i, dattr_cur, j)) | |
7721 | goto match2; | |
7722 | } | |
7723 | /* no match - add a new doms_new */ | |
7724 | __build_sched_domains(doms_new + i, | |
7725 | dattr_new ? dattr_new + i : NULL); | |
7726 | match2: | |
7727 | ; | |
7728 | } | |
7729 | ||
7730 | /* Remember the new sched domains */ | |
7731 | if (doms_cur != &fallback_doms) | |
7732 | kfree(doms_cur); | |
7733 | kfree(dattr_cur); /* kfree(NULL) is safe */ | |
7734 | doms_cur = doms_new; | |
7735 | dattr_cur = dattr_new; | |
7736 | ndoms_cur = ndoms_new; | |
7737 | ||
7738 | register_sched_domain_sysctl(); | |
7739 | ||
7740 | unlock_doms_cur(); | |
7741 | } | |
7742 | ||
7743 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
7744 | int arch_reinit_sched_domains(void) | |
7745 | { | |
7746 | int err; | |
7747 | ||
7748 | get_online_cpus(); | |
7749 | detach_destroy_domains(&cpu_online_map); | |
7750 | err = arch_init_sched_domains(&cpu_online_map); | |
7751 | put_online_cpus(); | |
7752 | ||
7753 | return err; | |
7754 | } | |
7755 | ||
7756 | static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt) | |
7757 | { | |
7758 | int ret; | |
7759 | ||
7760 | if (buf[0] != '0' && buf[0] != '1') | |
7761 | return -EINVAL; | |
7762 | ||
7763 | if (smt) | |
7764 | sched_smt_power_savings = (buf[0] == '1'); | |
7765 | else | |
7766 | sched_mc_power_savings = (buf[0] == '1'); | |
7767 | ||
7768 | ret = arch_reinit_sched_domains(); | |
7769 | ||
7770 | return ret ? ret : count; | |
7771 | } | |
7772 | ||
7773 | #ifdef CONFIG_SCHED_MC | |
7774 | static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page) | |
7775 | { | |
7776 | return sprintf(page, "%u\n", sched_mc_power_savings); | |
7777 | } | |
7778 | static ssize_t sched_mc_power_savings_store(struct sys_device *dev, | |
7779 | const char *buf, size_t count) | |
7780 | { | |
7781 | return sched_power_savings_store(buf, count, 0); | |
7782 | } | |
7783 | static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show, | |
7784 | sched_mc_power_savings_store); | |
7785 | #endif | |
7786 | ||
7787 | #ifdef CONFIG_SCHED_SMT | |
7788 | static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page) | |
7789 | { | |
7790 | return sprintf(page, "%u\n", sched_smt_power_savings); | |
7791 | } | |
7792 | static ssize_t sched_smt_power_savings_store(struct sys_device *dev, | |
7793 | const char *buf, size_t count) | |
7794 | { | |
7795 | return sched_power_savings_store(buf, count, 1); | |
7796 | } | |
7797 | static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show, | |
7798 | sched_smt_power_savings_store); | |
7799 | #endif | |
7800 | ||
7801 | int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls) | |
7802 | { | |
7803 | int err = 0; | |
7804 | ||
7805 | #ifdef CONFIG_SCHED_SMT | |
7806 | if (smt_capable()) | |
7807 | err = sysfs_create_file(&cls->kset.kobj, | |
7808 | &attr_sched_smt_power_savings.attr); | |
7809 | #endif | |
7810 | #ifdef CONFIG_SCHED_MC | |
7811 | if (!err && mc_capable()) | |
7812 | err = sysfs_create_file(&cls->kset.kobj, | |
7813 | &attr_sched_mc_power_savings.attr); | |
7814 | #endif | |
7815 | return err; | |
7816 | } | |
7817 | #endif | |
7818 | ||
7819 | /* | |
7820 | * Force a reinitialization of the sched domains hierarchy. The domains | |
7821 | * and groups cannot be updated in place without racing with the balancing | |
7822 | * code, so we temporarily attach all running cpus to the NULL domain | |
7823 | * which will prevent rebalancing while the sched domains are recalculated. | |
7824 | */ | |
7825 | static int update_sched_domains(struct notifier_block *nfb, | |
7826 | unsigned long action, void *hcpu) | |
7827 | { | |
7828 | switch (action) { | |
7829 | case CPU_UP_PREPARE: | |
7830 | case CPU_UP_PREPARE_FROZEN: | |
7831 | case CPU_DOWN_PREPARE: | |
7832 | case CPU_DOWN_PREPARE_FROZEN: | |
7833 | detach_destroy_domains(&cpu_online_map); | |
7834 | return NOTIFY_OK; | |
7835 | ||
7836 | case CPU_UP_CANCELED: | |
7837 | case CPU_UP_CANCELED_FROZEN: | |
7838 | case CPU_DOWN_FAILED: | |
7839 | case CPU_DOWN_FAILED_FROZEN: | |
7840 | case CPU_ONLINE: | |
7841 | case CPU_ONLINE_FROZEN: | |
7842 | case CPU_DEAD: | |
7843 | case CPU_DEAD_FROZEN: | |
7844 | /* | |
7845 | * Fall through and re-initialise the domains. | |
7846 | */ | |
7847 | break; | |
7848 | default: | |
7849 | return NOTIFY_DONE; | |
7850 | } | |
7851 | ||
7852 | /* The hotplug lock is already held by cpu_up/cpu_down */ | |
7853 | arch_init_sched_domains(&cpu_online_map); | |
7854 | ||
7855 | return NOTIFY_OK; | |
7856 | } | |
7857 | ||
7858 | void __init sched_init_smp(void) | |
7859 | { | |
7860 | cpumask_t non_isolated_cpus; | |
7861 | ||
7862 | #if defined(CONFIG_NUMA) | |
7863 | sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **), | |
7864 | GFP_KERNEL); | |
7865 | BUG_ON(sched_group_nodes_bycpu == NULL); | |
7866 | #endif | |
7867 | get_online_cpus(); | |
7868 | arch_init_sched_domains(&cpu_online_map); | |
7869 | cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map); | |
7870 | if (cpus_empty(non_isolated_cpus)) | |
7871 | cpu_set(smp_processor_id(), non_isolated_cpus); | |
7872 | put_online_cpus(); | |
7873 | /* XXX: Theoretical race here - CPU may be hotplugged now */ | |
7874 | hotcpu_notifier(update_sched_domains, 0); | |
7875 | ||
7876 | /* Move init over to a non-isolated CPU */ | |
7877 | if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0) | |
7878 | BUG(); | |
7879 | sched_init_granularity(); | |
7880 | } | |
7881 | #else | |
7882 | void __init sched_init_smp(void) | |
7883 | { | |
7884 | #if defined(CONFIG_NUMA) | |
7885 | sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **), | |
7886 | GFP_KERNEL); | |
7887 | BUG_ON(sched_group_nodes_bycpu == NULL); | |
7888 | #endif | |
7889 | sched_init_granularity(); | |
7890 | } | |
7891 | #endif /* CONFIG_SMP */ | |
7892 | ||
7893 | int in_sched_functions(unsigned long addr) | |
7894 | { | |
7895 | return in_lock_functions(addr) || | |
7896 | (addr >= (unsigned long)__sched_text_start | |
7897 | && addr < (unsigned long)__sched_text_end); | |
7898 | } | |
7899 | ||
7900 | static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq) | |
7901 | { | |
7902 | cfs_rq->tasks_timeline = RB_ROOT; | |
7903 | INIT_LIST_HEAD(&cfs_rq->tasks); | |
7904 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
7905 | cfs_rq->rq = rq; | |
7906 | #endif | |
7907 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); | |
7908 | } | |
7909 | ||
7910 | static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq) | |
7911 | { | |
7912 | struct rt_prio_array *array; | |
7913 | int i; | |
7914 | ||
7915 | array = &rt_rq->active; | |
7916 | for (i = 0; i < MAX_RT_PRIO; i++) { | |
7917 | INIT_LIST_HEAD(array->queue + i); | |
7918 | __clear_bit(i, array->bitmap); | |
7919 | } | |
7920 | /* delimiter for bitsearch: */ | |
7921 | __set_bit(MAX_RT_PRIO, array->bitmap); | |
7922 | ||
7923 | #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED | |
7924 | rt_rq->highest_prio = MAX_RT_PRIO; | |
7925 | #endif | |
7926 | #ifdef CONFIG_SMP | |
7927 | rt_rq->rt_nr_migratory = 0; | |
7928 | rt_rq->overloaded = 0; | |
7929 | #endif | |
7930 | ||
7931 | rt_rq->rt_time = 0; | |
7932 | rt_rq->rt_throttled = 0; | |
7933 | rt_rq->rt_runtime = 0; | |
7934 | spin_lock_init(&rt_rq->rt_runtime_lock); | |
7935 | ||
7936 | #ifdef CONFIG_RT_GROUP_SCHED | |
7937 | rt_rq->rt_nr_boosted = 0; | |
7938 | rt_rq->rq = rq; | |
7939 | #endif | |
7940 | } | |
7941 | ||
7942 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
7943 | static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
7944 | struct sched_entity *se, int cpu, int add, | |
7945 | struct sched_entity *parent) | |
7946 | { | |
7947 | struct rq *rq = cpu_rq(cpu); | |
7948 | tg->cfs_rq[cpu] = cfs_rq; | |
7949 | init_cfs_rq(cfs_rq, rq); | |
7950 | cfs_rq->tg = tg; | |
7951 | if (add) | |
7952 | list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list); | |
7953 | ||
7954 | tg->se[cpu] = se; | |
7955 | /* se could be NULL for init_task_group */ | |
7956 | if (!se) | |
7957 | return; | |
7958 | ||
7959 | if (!parent) | |
7960 | se->cfs_rq = &rq->cfs; | |
7961 | else | |
7962 | se->cfs_rq = parent->my_q; | |
7963 | ||
7964 | se->my_q = cfs_rq; | |
7965 | se->load.weight = tg->shares; | |
7966 | se->load.inv_weight = div64_64(1ULL<<32, se->load.weight); | |
7967 | se->parent = parent; | |
7968 | } | |
7969 | #endif | |
7970 | ||
7971 | #ifdef CONFIG_RT_GROUP_SCHED | |
7972 | static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq, | |
7973 | struct sched_rt_entity *rt_se, int cpu, int add, | |
7974 | struct sched_rt_entity *parent) | |
7975 | { | |
7976 | struct rq *rq = cpu_rq(cpu); | |
7977 | ||
7978 | tg->rt_rq[cpu] = rt_rq; | |
7979 | init_rt_rq(rt_rq, rq); | |
7980 | rt_rq->tg = tg; | |
7981 | rt_rq->rt_se = rt_se; | |
7982 | rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime; | |
7983 | if (add) | |
7984 | list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list); | |
7985 | ||
7986 | tg->rt_se[cpu] = rt_se; | |
7987 | if (!rt_se) | |
7988 | return; | |
7989 | ||
7990 | if (!parent) | |
7991 | rt_se->rt_rq = &rq->rt; | |
7992 | else | |
7993 | rt_se->rt_rq = parent->my_q; | |
7994 | ||
7995 | rt_se->rt_rq = &rq->rt; | |
7996 | rt_se->my_q = rt_rq; | |
7997 | rt_se->parent = parent; | |
7998 | INIT_LIST_HEAD(&rt_se->run_list); | |
7999 | } | |
8000 | #endif | |
8001 | ||
8002 | void __init sched_init(void) | |
8003 | { | |
8004 | int i, j; | |
8005 | unsigned long alloc_size = 0, ptr; | |
8006 | ||
8007 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
8008 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); | |
8009 | #endif | |
8010 | #ifdef CONFIG_RT_GROUP_SCHED | |
8011 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); | |
8012 | #endif | |
8013 | #ifdef CONFIG_USER_SCHED | |
8014 | alloc_size *= 2; | |
8015 | #endif | |
8016 | /* | |
8017 | * As sched_init() is called before page_alloc is setup, | |
8018 | * we use alloc_bootmem(). | |
8019 | */ | |
8020 | if (alloc_size) { | |
8021 | ptr = (unsigned long)alloc_bootmem_low(alloc_size); | |
8022 | ||
8023 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
8024 | init_task_group.se = (struct sched_entity **)ptr; | |
8025 | ptr += nr_cpu_ids * sizeof(void **); | |
8026 | ||
8027 | init_task_group.cfs_rq = (struct cfs_rq **)ptr; | |
8028 | ptr += nr_cpu_ids * sizeof(void **); | |
8029 | ||
8030 | #ifdef CONFIG_USER_SCHED | |
8031 | root_task_group.se = (struct sched_entity **)ptr; | |
8032 | ptr += nr_cpu_ids * sizeof(void **); | |
8033 | ||
8034 | root_task_group.cfs_rq = (struct cfs_rq **)ptr; | |
8035 | ptr += nr_cpu_ids * sizeof(void **); | |
8036 | #endif | |
8037 | #endif | |
8038 | #ifdef CONFIG_RT_GROUP_SCHED | |
8039 | init_task_group.rt_se = (struct sched_rt_entity **)ptr; | |
8040 | ptr += nr_cpu_ids * sizeof(void **); | |
8041 | ||
8042 | init_task_group.rt_rq = (struct rt_rq **)ptr; | |
8043 | ptr += nr_cpu_ids * sizeof(void **); | |
8044 | ||
8045 | #ifdef CONFIG_USER_SCHED | |
8046 | root_task_group.rt_se = (struct sched_rt_entity **)ptr; | |
8047 | ptr += nr_cpu_ids * sizeof(void **); | |
8048 | ||
8049 | root_task_group.rt_rq = (struct rt_rq **)ptr; | |
8050 | ptr += nr_cpu_ids * sizeof(void **); | |
8051 | #endif | |
8052 | #endif | |
8053 | } | |
8054 | ||
8055 | #ifdef CONFIG_SMP | |
8056 | init_aggregate(); | |
8057 | init_defrootdomain(); | |
8058 | #endif | |
8059 | ||
8060 | init_rt_bandwidth(&def_rt_bandwidth, | |
8061 | global_rt_period(), global_rt_runtime()); | |
8062 | ||
8063 | #ifdef CONFIG_RT_GROUP_SCHED | |
8064 | init_rt_bandwidth(&init_task_group.rt_bandwidth, | |
8065 | global_rt_period(), global_rt_runtime()); | |
8066 | #ifdef CONFIG_USER_SCHED | |
8067 | init_rt_bandwidth(&root_task_group.rt_bandwidth, | |
8068 | global_rt_period(), RUNTIME_INF); | |
8069 | #endif | |
8070 | #endif | |
8071 | ||
8072 | #ifdef CONFIG_GROUP_SCHED | |
8073 | list_add(&init_task_group.list, &task_groups); | |
8074 | INIT_LIST_HEAD(&init_task_group.children); | |
8075 | ||
8076 | #ifdef CONFIG_USER_SCHED | |
8077 | INIT_LIST_HEAD(&root_task_group.children); | |
8078 | init_task_group.parent = &root_task_group; | |
8079 | list_add(&init_task_group.siblings, &root_task_group.children); | |
8080 | #endif | |
8081 | #endif | |
8082 | ||
8083 | for_each_possible_cpu(i) { | |
8084 | struct rq *rq; | |
8085 | ||
8086 | rq = cpu_rq(i); | |
8087 | spin_lock_init(&rq->lock); | |
8088 | lockdep_set_class(&rq->lock, &rq->rq_lock_key); | |
8089 | rq->nr_running = 0; | |
8090 | rq->clock = 1; | |
8091 | update_last_tick_seen(rq); | |
8092 | init_cfs_rq(&rq->cfs, rq); | |
8093 | init_rt_rq(&rq->rt, rq); | |
8094 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
8095 | init_task_group.shares = init_task_group_load; | |
8096 | INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); | |
8097 | #ifdef CONFIG_CGROUP_SCHED | |
8098 | /* | |
8099 | * How much cpu bandwidth does init_task_group get? | |
8100 | * | |
8101 | * In case of task-groups formed thr' the cgroup filesystem, it | |
8102 | * gets 100% of the cpu resources in the system. This overall | |
8103 | * system cpu resource is divided among the tasks of | |
8104 | * init_task_group and its child task-groups in a fair manner, | |
8105 | * based on each entity's (task or task-group's) weight | |
8106 | * (se->load.weight). | |
8107 | * | |
8108 | * In other words, if init_task_group has 10 tasks of weight | |
8109 | * 1024) and two child groups A0 and A1 (of weight 1024 each), | |
8110 | * then A0's share of the cpu resource is: | |
8111 | * | |
8112 | * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% | |
8113 | * | |
8114 | * We achieve this by letting init_task_group's tasks sit | |
8115 | * directly in rq->cfs (i.e init_task_group->se[] = NULL). | |
8116 | */ | |
8117 | init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL); | |
8118 | #elif defined CONFIG_USER_SCHED | |
8119 | root_task_group.shares = NICE_0_LOAD; | |
8120 | init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL); | |
8121 | /* | |
8122 | * In case of task-groups formed thr' the user id of tasks, | |
8123 | * init_task_group represents tasks belonging to root user. | |
8124 | * Hence it forms a sibling of all subsequent groups formed. | |
8125 | * In this case, init_task_group gets only a fraction of overall | |
8126 | * system cpu resource, based on the weight assigned to root | |
8127 | * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished | |
8128 | * by letting tasks of init_task_group sit in a separate cfs_rq | |
8129 | * (init_cfs_rq) and having one entity represent this group of | |
8130 | * tasks in rq->cfs (i.e init_task_group->se[] != NULL). | |
8131 | */ | |
8132 | init_tg_cfs_entry(&init_task_group, | |
8133 | &per_cpu(init_cfs_rq, i), | |
8134 | &per_cpu(init_sched_entity, i), i, 1, | |
8135 | root_task_group.se[i]); | |
8136 | ||
8137 | #endif | |
8138 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
8139 | ||
8140 | rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; | |
8141 | #ifdef CONFIG_RT_GROUP_SCHED | |
8142 | INIT_LIST_HEAD(&rq->leaf_rt_rq_list); | |
8143 | #ifdef CONFIG_CGROUP_SCHED | |
8144 | init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL); | |
8145 | #elif defined CONFIG_USER_SCHED | |
8146 | init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL); | |
8147 | init_tg_rt_entry(&init_task_group, | |
8148 | &per_cpu(init_rt_rq, i), | |
8149 | &per_cpu(init_sched_rt_entity, i), i, 1, | |
8150 | root_task_group.rt_se[i]); | |
8151 | #endif | |
8152 | #endif | |
8153 | ||
8154 | for (j = 0; j < CPU_LOAD_IDX_MAX; j++) | |
8155 | rq->cpu_load[j] = 0; | |
8156 | #ifdef CONFIG_SMP | |
8157 | rq->sd = NULL; | |
8158 | rq->rd = NULL; | |
8159 | rq->active_balance = 0; | |
8160 | rq->next_balance = jiffies; | |
8161 | rq->push_cpu = 0; | |
8162 | rq->cpu = i; | |
8163 | rq->migration_thread = NULL; | |
8164 | INIT_LIST_HEAD(&rq->migration_queue); | |
8165 | rq_attach_root(rq, &def_root_domain); | |
8166 | #endif | |
8167 | init_rq_hrtick(rq); | |
8168 | atomic_set(&rq->nr_iowait, 0); | |
8169 | } | |
8170 | ||
8171 | set_load_weight(&init_task); | |
8172 | ||
8173 | #ifdef CONFIG_PREEMPT_NOTIFIERS | |
8174 | INIT_HLIST_HEAD(&init_task.preempt_notifiers); | |
8175 | #endif | |
8176 | ||
8177 | #ifdef CONFIG_SMP | |
8178 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL); | |
8179 | #endif | |
8180 | ||
8181 | #ifdef CONFIG_RT_MUTEXES | |
8182 | plist_head_init(&init_task.pi_waiters, &init_task.pi_lock); | |
8183 | #endif | |
8184 | ||
8185 | /* | |
8186 | * The boot idle thread does lazy MMU switching as well: | |
8187 | */ | |
8188 | atomic_inc(&init_mm.mm_count); | |
8189 | enter_lazy_tlb(&init_mm, current); | |
8190 | ||
8191 | /* | |
8192 | * Make us the idle thread. Technically, schedule() should not be | |
8193 | * called from this thread, however somewhere below it might be, | |
8194 | * but because we are the idle thread, we just pick up running again | |
8195 | * when this runqueue becomes "idle". | |
8196 | */ | |
8197 | init_idle(current, smp_processor_id()); | |
8198 | /* | |
8199 | * During early bootup we pretend to be a normal task: | |
8200 | */ | |
8201 | current->sched_class = &fair_sched_class; | |
8202 | ||
8203 | scheduler_running = 1; | |
8204 | } | |
8205 | ||
8206 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP | |
8207 | void __might_sleep(char *file, int line) | |
8208 | { | |
8209 | #ifdef in_atomic | |
8210 | static unsigned long prev_jiffy; /* ratelimiting */ | |
8211 | ||
8212 | if ((in_atomic() || irqs_disabled()) && | |
8213 | system_state == SYSTEM_RUNNING && !oops_in_progress) { | |
8214 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) | |
8215 | return; | |
8216 | prev_jiffy = jiffies; | |
8217 | printk(KERN_ERR "BUG: sleeping function called from invalid" | |
8218 | " context at %s:%d\n", file, line); | |
8219 | printk("in_atomic():%d, irqs_disabled():%d\n", | |
8220 | in_atomic(), irqs_disabled()); | |
8221 | debug_show_held_locks(current); | |
8222 | if (irqs_disabled()) | |
8223 | print_irqtrace_events(current); | |
8224 | dump_stack(); | |
8225 | } | |
8226 | #endif | |
8227 | } | |
8228 | EXPORT_SYMBOL(__might_sleep); | |
8229 | #endif | |
8230 | ||
8231 | #ifdef CONFIG_MAGIC_SYSRQ | |
8232 | static void normalize_task(struct rq *rq, struct task_struct *p) | |
8233 | { | |
8234 | int on_rq; | |
8235 | update_rq_clock(rq); | |
8236 | on_rq = p->se.on_rq; | |
8237 | if (on_rq) | |
8238 | deactivate_task(rq, p, 0); | |
8239 | __setscheduler(rq, p, SCHED_NORMAL, 0); | |
8240 | if (on_rq) { | |
8241 | activate_task(rq, p, 0); | |
8242 | resched_task(rq->curr); | |
8243 | } | |
8244 | } | |
8245 | ||
8246 | void normalize_rt_tasks(void) | |
8247 | { | |
8248 | struct task_struct *g, *p; | |
8249 | unsigned long flags; | |
8250 | struct rq *rq; | |
8251 | ||
8252 | read_lock_irqsave(&tasklist_lock, flags); | |
8253 | do_each_thread(g, p) { | |
8254 | /* | |
8255 | * Only normalize user tasks: | |
8256 | */ | |
8257 | if (!p->mm) | |
8258 | continue; | |
8259 | ||
8260 | p->se.exec_start = 0; | |
8261 | #ifdef CONFIG_SCHEDSTATS | |
8262 | p->se.wait_start = 0; | |
8263 | p->se.sleep_start = 0; | |
8264 | p->se.block_start = 0; | |
8265 | #endif | |
8266 | task_rq(p)->clock = 0; | |
8267 | ||
8268 | if (!rt_task(p)) { | |
8269 | /* | |
8270 | * Renice negative nice level userspace | |
8271 | * tasks back to 0: | |
8272 | */ | |
8273 | if (TASK_NICE(p) < 0 && p->mm) | |
8274 | set_user_nice(p, 0); | |
8275 | continue; | |
8276 | } | |
8277 | ||
8278 | spin_lock(&p->pi_lock); | |
8279 | rq = __task_rq_lock(p); | |
8280 | ||
8281 | normalize_task(rq, p); | |
8282 | ||
8283 | __task_rq_unlock(rq); | |
8284 | spin_unlock(&p->pi_lock); | |
8285 | } while_each_thread(g, p); | |
8286 | ||
8287 | read_unlock_irqrestore(&tasklist_lock, flags); | |
8288 | } | |
8289 | ||
8290 | #endif /* CONFIG_MAGIC_SYSRQ */ | |
8291 | ||
8292 | #ifdef CONFIG_IA64 | |
8293 | /* | |
8294 | * These functions are only useful for the IA64 MCA handling. | |
8295 | * | |
8296 | * They can only be called when the whole system has been | |
8297 | * stopped - every CPU needs to be quiescent, and no scheduling | |
8298 | * activity can take place. Using them for anything else would | |
8299 | * be a serious bug, and as a result, they aren't even visible | |
8300 | * under any other configuration. | |
8301 | */ | |
8302 | ||
8303 | /** | |
8304 | * curr_task - return the current task for a given cpu. | |
8305 | * @cpu: the processor in question. | |
8306 | * | |
8307 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
8308 | */ | |
8309 | struct task_struct *curr_task(int cpu) | |
8310 | { | |
8311 | return cpu_curr(cpu); | |
8312 | } | |
8313 | ||
8314 | /** | |
8315 | * set_curr_task - set the current task for a given cpu. | |
8316 | * @cpu: the processor in question. | |
8317 | * @p: the task pointer to set. | |
8318 | * | |
8319 | * Description: This function must only be used when non-maskable interrupts | |
8320 | * are serviced on a separate stack. It allows the architecture to switch the | |
8321 | * notion of the current task on a cpu in a non-blocking manner. This function | |
8322 | * must be called with all CPU's synchronized, and interrupts disabled, the | |
8323 | * and caller must save the original value of the current task (see | |
8324 | * curr_task() above) and restore that value before reenabling interrupts and | |
8325 | * re-starting the system. | |
8326 | * | |
8327 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
8328 | */ | |
8329 | void set_curr_task(int cpu, struct task_struct *p) | |
8330 | { | |
8331 | cpu_curr(cpu) = p; | |
8332 | } | |
8333 | ||
8334 | #endif | |
8335 | ||
8336 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
8337 | static void free_fair_sched_group(struct task_group *tg) | |
8338 | { | |
8339 | int i; | |
8340 | ||
8341 | for_each_possible_cpu(i) { | |
8342 | if (tg->cfs_rq) | |
8343 | kfree(tg->cfs_rq[i]); | |
8344 | if (tg->se) | |
8345 | kfree(tg->se[i]); | |
8346 | } | |
8347 | ||
8348 | kfree(tg->cfs_rq); | |
8349 | kfree(tg->se); | |
8350 | } | |
8351 | ||
8352 | static | |
8353 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
8354 | { | |
8355 | struct cfs_rq *cfs_rq; | |
8356 | struct sched_entity *se, *parent_se; | |
8357 | struct rq *rq; | |
8358 | int i; | |
8359 | ||
8360 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); | |
8361 | if (!tg->cfs_rq) | |
8362 | goto err; | |
8363 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); | |
8364 | if (!tg->se) | |
8365 | goto err; | |
8366 | ||
8367 | tg->shares = NICE_0_LOAD; | |
8368 | ||
8369 | for_each_possible_cpu(i) { | |
8370 | rq = cpu_rq(i); | |
8371 | ||
8372 | cfs_rq = kmalloc_node(sizeof(struct cfs_rq), | |
8373 | GFP_KERNEL|__GFP_ZERO, cpu_to_node(i)); | |
8374 | if (!cfs_rq) | |
8375 | goto err; | |
8376 | ||
8377 | se = kmalloc_node(sizeof(struct sched_entity), | |
8378 | GFP_KERNEL|__GFP_ZERO, cpu_to_node(i)); | |
8379 | if (!se) | |
8380 | goto err; | |
8381 | ||
8382 | parent_se = parent ? parent->se[i] : NULL; | |
8383 | init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se); | |
8384 | } | |
8385 | ||
8386 | return 1; | |
8387 | ||
8388 | err: | |
8389 | return 0; | |
8390 | } | |
8391 | ||
8392 | static inline void register_fair_sched_group(struct task_group *tg, int cpu) | |
8393 | { | |
8394 | list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list, | |
8395 | &cpu_rq(cpu)->leaf_cfs_rq_list); | |
8396 | } | |
8397 | ||
8398 | static inline void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
8399 | { | |
8400 | list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list); | |
8401 | } | |
8402 | #else | |
8403 | static inline void free_fair_sched_group(struct task_group *tg) | |
8404 | { | |
8405 | } | |
8406 | ||
8407 | static inline | |
8408 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
8409 | { | |
8410 | return 1; | |
8411 | } | |
8412 | ||
8413 | static inline void register_fair_sched_group(struct task_group *tg, int cpu) | |
8414 | { | |
8415 | } | |
8416 | ||
8417 | static inline void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
8418 | { | |
8419 | } | |
8420 | #endif | |
8421 | ||
8422 | #ifdef CONFIG_RT_GROUP_SCHED | |
8423 | static void free_rt_sched_group(struct task_group *tg) | |
8424 | { | |
8425 | int i; | |
8426 | ||
8427 | destroy_rt_bandwidth(&tg->rt_bandwidth); | |
8428 | ||
8429 | for_each_possible_cpu(i) { | |
8430 | if (tg->rt_rq) | |
8431 | kfree(tg->rt_rq[i]); | |
8432 | if (tg->rt_se) | |
8433 | kfree(tg->rt_se[i]); | |
8434 | } | |
8435 | ||
8436 | kfree(tg->rt_rq); | |
8437 | kfree(tg->rt_se); | |
8438 | } | |
8439 | ||
8440 | static | |
8441 | int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) | |
8442 | { | |
8443 | struct rt_rq *rt_rq; | |
8444 | struct sched_rt_entity *rt_se, *parent_se; | |
8445 | struct rq *rq; | |
8446 | int i; | |
8447 | ||
8448 | tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL); | |
8449 | if (!tg->rt_rq) | |
8450 | goto err; | |
8451 | tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL); | |
8452 | if (!tg->rt_se) | |
8453 | goto err; | |
8454 | ||
8455 | init_rt_bandwidth(&tg->rt_bandwidth, | |
8456 | ktime_to_ns(def_rt_bandwidth.rt_period), 0); | |
8457 | ||
8458 | for_each_possible_cpu(i) { | |
8459 | rq = cpu_rq(i); | |
8460 | ||
8461 | rt_rq = kmalloc_node(sizeof(struct rt_rq), | |
8462 | GFP_KERNEL|__GFP_ZERO, cpu_to_node(i)); | |
8463 | if (!rt_rq) | |
8464 | goto err; | |
8465 | ||
8466 | rt_se = kmalloc_node(sizeof(struct sched_rt_entity), | |
8467 | GFP_KERNEL|__GFP_ZERO, cpu_to_node(i)); | |
8468 | if (!rt_se) | |
8469 | goto err; | |
8470 | ||
8471 | parent_se = parent ? parent->rt_se[i] : NULL; | |
8472 | init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se); | |
8473 | } | |
8474 | ||
8475 | return 1; | |
8476 | ||
8477 | err: | |
8478 | return 0; | |
8479 | } | |
8480 | ||
8481 | static inline void register_rt_sched_group(struct task_group *tg, int cpu) | |
8482 | { | |
8483 | list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list, | |
8484 | &cpu_rq(cpu)->leaf_rt_rq_list); | |
8485 | } | |
8486 | ||
8487 | static inline void unregister_rt_sched_group(struct task_group *tg, int cpu) | |
8488 | { | |
8489 | list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list); | |
8490 | } | |
8491 | #else | |
8492 | static inline void free_rt_sched_group(struct task_group *tg) | |
8493 | { | |
8494 | } | |
8495 | ||
8496 | static inline | |
8497 | int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) | |
8498 | { | |
8499 | return 1; | |
8500 | } | |
8501 | ||
8502 | static inline void register_rt_sched_group(struct task_group *tg, int cpu) | |
8503 | { | |
8504 | } | |
8505 | ||
8506 | static inline void unregister_rt_sched_group(struct task_group *tg, int cpu) | |
8507 | { | |
8508 | } | |
8509 | #endif | |
8510 | ||
8511 | #ifdef CONFIG_GROUP_SCHED | |
8512 | static void free_sched_group(struct task_group *tg) | |
8513 | { | |
8514 | free_fair_sched_group(tg); | |
8515 | free_rt_sched_group(tg); | |
8516 | kfree(tg); | |
8517 | } | |
8518 | ||
8519 | /* allocate runqueue etc for a new task group */ | |
8520 | struct task_group *sched_create_group(struct task_group *parent) | |
8521 | { | |
8522 | struct task_group *tg; | |
8523 | unsigned long flags; | |
8524 | int i; | |
8525 | ||
8526 | tg = kzalloc(sizeof(*tg), GFP_KERNEL); | |
8527 | if (!tg) | |
8528 | return ERR_PTR(-ENOMEM); | |
8529 | ||
8530 | if (!alloc_fair_sched_group(tg, parent)) | |
8531 | goto err; | |
8532 | ||
8533 | if (!alloc_rt_sched_group(tg, parent)) | |
8534 | goto err; | |
8535 | ||
8536 | spin_lock_irqsave(&task_group_lock, flags); | |
8537 | for_each_possible_cpu(i) { | |
8538 | register_fair_sched_group(tg, i); | |
8539 | register_rt_sched_group(tg, i); | |
8540 | } | |
8541 | list_add_rcu(&tg->list, &task_groups); | |
8542 | ||
8543 | WARN_ON(!parent); /* root should already exist */ | |
8544 | ||
8545 | tg->parent = parent; | |
8546 | list_add_rcu(&tg->siblings, &parent->children); | |
8547 | INIT_LIST_HEAD(&tg->children); | |
8548 | spin_unlock_irqrestore(&task_group_lock, flags); | |
8549 | ||
8550 | return tg; | |
8551 | ||
8552 | err: | |
8553 | free_sched_group(tg); | |
8554 | return ERR_PTR(-ENOMEM); | |
8555 | } | |
8556 | ||
8557 | /* rcu callback to free various structures associated with a task group */ | |
8558 | static void free_sched_group_rcu(struct rcu_head *rhp) | |
8559 | { | |
8560 | /* now it should be safe to free those cfs_rqs */ | |
8561 | free_sched_group(container_of(rhp, struct task_group, rcu)); | |
8562 | } | |
8563 | ||
8564 | /* Destroy runqueue etc associated with a task group */ | |
8565 | void sched_destroy_group(struct task_group *tg) | |
8566 | { | |
8567 | unsigned long flags; | |
8568 | int i; | |
8569 | ||
8570 | spin_lock_irqsave(&task_group_lock, flags); | |
8571 | for_each_possible_cpu(i) { | |
8572 | unregister_fair_sched_group(tg, i); | |
8573 | unregister_rt_sched_group(tg, i); | |
8574 | } | |
8575 | list_del_rcu(&tg->list); | |
8576 | list_del_rcu(&tg->siblings); | |
8577 | spin_unlock_irqrestore(&task_group_lock, flags); | |
8578 | ||
8579 | /* wait for possible concurrent references to cfs_rqs complete */ | |
8580 | call_rcu(&tg->rcu, free_sched_group_rcu); | |
8581 | } | |
8582 | ||
8583 | /* change task's runqueue when it moves between groups. | |
8584 | * The caller of this function should have put the task in its new group | |
8585 | * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to | |
8586 | * reflect its new group. | |
8587 | */ | |
8588 | void sched_move_task(struct task_struct *tsk) | |
8589 | { | |
8590 | int on_rq, running; | |
8591 | unsigned long flags; | |
8592 | struct rq *rq; | |
8593 | ||
8594 | rq = task_rq_lock(tsk, &flags); | |
8595 | ||
8596 | update_rq_clock(rq); | |
8597 | ||
8598 | running = task_current(rq, tsk); | |
8599 | on_rq = tsk->se.on_rq; | |
8600 | ||
8601 | if (on_rq) | |
8602 | dequeue_task(rq, tsk, 0); | |
8603 | if (unlikely(running)) | |
8604 | tsk->sched_class->put_prev_task(rq, tsk); | |
8605 | ||
8606 | set_task_rq(tsk, task_cpu(tsk)); | |
8607 | ||
8608 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
8609 | if (tsk->sched_class->moved_group) | |
8610 | tsk->sched_class->moved_group(tsk); | |
8611 | #endif | |
8612 | ||
8613 | if (unlikely(running)) | |
8614 | tsk->sched_class->set_curr_task(rq); | |
8615 | if (on_rq) | |
8616 | enqueue_task(rq, tsk, 0); | |
8617 | ||
8618 | task_rq_unlock(rq, &flags); | |
8619 | } | |
8620 | #endif | |
8621 | ||
8622 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
8623 | static void __set_se_shares(struct sched_entity *se, unsigned long shares) | |
8624 | { | |
8625 | struct cfs_rq *cfs_rq = se->cfs_rq; | |
8626 | int on_rq; | |
8627 | ||
8628 | on_rq = se->on_rq; | |
8629 | if (on_rq) | |
8630 | dequeue_entity(cfs_rq, se, 0); | |
8631 | ||
8632 | se->load.weight = shares; | |
8633 | se->load.inv_weight = div64_64((1ULL<<32), shares); | |
8634 | ||
8635 | if (on_rq) | |
8636 | enqueue_entity(cfs_rq, se, 0); | |
8637 | } | |
8638 | ||
8639 | static void set_se_shares(struct sched_entity *se, unsigned long shares) | |
8640 | { | |
8641 | struct cfs_rq *cfs_rq = se->cfs_rq; | |
8642 | struct rq *rq = cfs_rq->rq; | |
8643 | unsigned long flags; | |
8644 | ||
8645 | spin_lock_irqsave(&rq->lock, flags); | |
8646 | __set_se_shares(se, shares); | |
8647 | spin_unlock_irqrestore(&rq->lock, flags); | |
8648 | } | |
8649 | ||
8650 | static DEFINE_MUTEX(shares_mutex); | |
8651 | ||
8652 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
8653 | { | |
8654 | int i; | |
8655 | unsigned long flags; | |
8656 | ||
8657 | /* | |
8658 | * We can't change the weight of the root cgroup. | |
8659 | */ | |
8660 | if (!tg->se[0]) | |
8661 | return -EINVAL; | |
8662 | ||
8663 | /* | |
8664 | * A weight of 0 or 1 can cause arithmetics problems. | |
8665 | * (The default weight is 1024 - so there's no practical | |
8666 | * limitation from this.) | |
8667 | */ | |
8668 | if (shares < MIN_SHARES) | |
8669 | shares = MIN_SHARES; | |
8670 | ||
8671 | mutex_lock(&shares_mutex); | |
8672 | if (tg->shares == shares) | |
8673 | goto done; | |
8674 | ||
8675 | spin_lock_irqsave(&task_group_lock, flags); | |
8676 | for_each_possible_cpu(i) | |
8677 | unregister_fair_sched_group(tg, i); | |
8678 | list_del_rcu(&tg->siblings); | |
8679 | spin_unlock_irqrestore(&task_group_lock, flags); | |
8680 | ||
8681 | /* wait for any ongoing reference to this group to finish */ | |
8682 | synchronize_sched(); | |
8683 | ||
8684 | /* | |
8685 | * Now we are free to modify the group's share on each cpu | |
8686 | * w/o tripping rebalance_share or load_balance_fair. | |
8687 | */ | |
8688 | tg->shares = shares; | |
8689 | for_each_possible_cpu(i) { | |
8690 | /* | |
8691 | * force a rebalance | |
8692 | */ | |
8693 | cfs_rq_set_shares(tg->cfs_rq[i], 0); | |
8694 | set_se_shares(tg->se[i], shares/nr_cpu_ids); | |
8695 | } | |
8696 | ||
8697 | /* | |
8698 | * Enable load balance activity on this group, by inserting it back on | |
8699 | * each cpu's rq->leaf_cfs_rq_list. | |
8700 | */ | |
8701 | spin_lock_irqsave(&task_group_lock, flags); | |
8702 | for_each_possible_cpu(i) | |
8703 | register_fair_sched_group(tg, i); | |
8704 | list_add_rcu(&tg->siblings, &tg->parent->children); | |
8705 | spin_unlock_irqrestore(&task_group_lock, flags); | |
8706 | done: | |
8707 | mutex_unlock(&shares_mutex); | |
8708 | return 0; | |
8709 | } | |
8710 | ||
8711 | unsigned long sched_group_shares(struct task_group *tg) | |
8712 | { | |
8713 | return tg->shares; | |
8714 | } | |
8715 | #endif | |
8716 | ||
8717 | #ifdef CONFIG_RT_GROUP_SCHED | |
8718 | /* | |
8719 | * Ensure that the real time constraints are schedulable. | |
8720 | */ | |
8721 | static DEFINE_MUTEX(rt_constraints_mutex); | |
8722 | ||
8723 | static unsigned long to_ratio(u64 period, u64 runtime) | |
8724 | { | |
8725 | if (runtime == RUNTIME_INF) | |
8726 | return 1ULL << 16; | |
8727 | ||
8728 | return div64_64(runtime << 16, period); | |
8729 | } | |
8730 | ||
8731 | #ifdef CONFIG_CGROUP_SCHED | |
8732 | static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) | |
8733 | { | |
8734 | struct task_group *tgi, *parent = tg->parent; | |
8735 | unsigned long total = 0; | |
8736 | ||
8737 | if (!parent) { | |
8738 | if (global_rt_period() < period) | |
8739 | return 0; | |
8740 | ||
8741 | return to_ratio(period, runtime) < | |
8742 | to_ratio(global_rt_period(), global_rt_runtime()); | |
8743 | } | |
8744 | ||
8745 | if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period) | |
8746 | return 0; | |
8747 | ||
8748 | rcu_read_lock(); | |
8749 | list_for_each_entry_rcu(tgi, &parent->children, siblings) { | |
8750 | if (tgi == tg) | |
8751 | continue; | |
8752 | ||
8753 | total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period), | |
8754 | tgi->rt_bandwidth.rt_runtime); | |
8755 | } | |
8756 | rcu_read_unlock(); | |
8757 | ||
8758 | return total + to_ratio(period, runtime) < | |
8759 | to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period), | |
8760 | parent->rt_bandwidth.rt_runtime); | |
8761 | } | |
8762 | #elif defined CONFIG_USER_SCHED | |
8763 | static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) | |
8764 | { | |
8765 | struct task_group *tgi; | |
8766 | unsigned long total = 0; | |
8767 | unsigned long global_ratio = | |
8768 | to_ratio(global_rt_period(), global_rt_runtime()); | |
8769 | ||
8770 | rcu_read_lock(); | |
8771 | list_for_each_entry_rcu(tgi, &task_groups, list) { | |
8772 | if (tgi == tg) | |
8773 | continue; | |
8774 | ||
8775 | total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period), | |
8776 | tgi->rt_bandwidth.rt_runtime); | |
8777 | } | |
8778 | rcu_read_unlock(); | |
8779 | ||
8780 | return total + to_ratio(period, runtime) < global_ratio; | |
8781 | } | |
8782 | #endif | |
8783 | ||
8784 | /* Must be called with tasklist_lock held */ | |
8785 | static inline int tg_has_rt_tasks(struct task_group *tg) | |
8786 | { | |
8787 | struct task_struct *g, *p; | |
8788 | do_each_thread(g, p) { | |
8789 | if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg) | |
8790 | return 1; | |
8791 | } while_each_thread(g, p); | |
8792 | return 0; | |
8793 | } | |
8794 | ||
8795 | static int tg_set_bandwidth(struct task_group *tg, | |
8796 | u64 rt_period, u64 rt_runtime) | |
8797 | { | |
8798 | int i, err = 0; | |
8799 | ||
8800 | mutex_lock(&rt_constraints_mutex); | |
8801 | read_lock(&tasklist_lock); | |
8802 | if (rt_runtime == 0 && tg_has_rt_tasks(tg)) { | |
8803 | err = -EBUSY; | |
8804 | goto unlock; | |
8805 | } | |
8806 | if (!__rt_schedulable(tg, rt_period, rt_runtime)) { | |
8807 | err = -EINVAL; | |
8808 | goto unlock; | |
8809 | } | |
8810 | ||
8811 | spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); | |
8812 | tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); | |
8813 | tg->rt_bandwidth.rt_runtime = rt_runtime; | |
8814 | ||
8815 | for_each_possible_cpu(i) { | |
8816 | struct rt_rq *rt_rq = tg->rt_rq[i]; | |
8817 | ||
8818 | spin_lock(&rt_rq->rt_runtime_lock); | |
8819 | rt_rq->rt_runtime = rt_runtime; | |
8820 | spin_unlock(&rt_rq->rt_runtime_lock); | |
8821 | } | |
8822 | spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); | |
8823 | unlock: | |
8824 | read_unlock(&tasklist_lock); | |
8825 | mutex_unlock(&rt_constraints_mutex); | |
8826 | ||
8827 | return err; | |
8828 | } | |
8829 | ||
8830 | int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) | |
8831 | { | |
8832 | u64 rt_runtime, rt_period; | |
8833 | ||
8834 | rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); | |
8835 | rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; | |
8836 | if (rt_runtime_us < 0) | |
8837 | rt_runtime = RUNTIME_INF; | |
8838 | ||
8839 | return tg_set_bandwidth(tg, rt_period, rt_runtime); | |
8840 | } | |
8841 | ||
8842 | long sched_group_rt_runtime(struct task_group *tg) | |
8843 | { | |
8844 | u64 rt_runtime_us; | |
8845 | ||
8846 | if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) | |
8847 | return -1; | |
8848 | ||
8849 | rt_runtime_us = tg->rt_bandwidth.rt_runtime; | |
8850 | do_div(rt_runtime_us, NSEC_PER_USEC); | |
8851 | return rt_runtime_us; | |
8852 | } | |
8853 | ||
8854 | int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) | |
8855 | { | |
8856 | u64 rt_runtime, rt_period; | |
8857 | ||
8858 | rt_period = (u64)rt_period_us * NSEC_PER_USEC; | |
8859 | rt_runtime = tg->rt_bandwidth.rt_runtime; | |
8860 | ||
8861 | return tg_set_bandwidth(tg, rt_period, rt_runtime); | |
8862 | } | |
8863 | ||
8864 | long sched_group_rt_period(struct task_group *tg) | |
8865 | { | |
8866 | u64 rt_period_us; | |
8867 | ||
8868 | rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); | |
8869 | do_div(rt_period_us, NSEC_PER_USEC); | |
8870 | return rt_period_us; | |
8871 | } | |
8872 | ||
8873 | static int sched_rt_global_constraints(void) | |
8874 | { | |
8875 | int ret = 0; | |
8876 | ||
8877 | mutex_lock(&rt_constraints_mutex); | |
8878 | if (!__rt_schedulable(NULL, 1, 0)) | |
8879 | ret = -EINVAL; | |
8880 | mutex_unlock(&rt_constraints_mutex); | |
8881 | ||
8882 | return ret; | |
8883 | } | |
8884 | #else | |
8885 | static int sched_rt_global_constraints(void) | |
8886 | { | |
8887 | unsigned long flags; | |
8888 | int i; | |
8889 | ||
8890 | spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); | |
8891 | for_each_possible_cpu(i) { | |
8892 | struct rt_rq *rt_rq = &cpu_rq(i)->rt; | |
8893 | ||
8894 | spin_lock(&rt_rq->rt_runtime_lock); | |
8895 | rt_rq->rt_runtime = global_rt_runtime(); | |
8896 | spin_unlock(&rt_rq->rt_runtime_lock); | |
8897 | } | |
8898 | spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); | |
8899 | ||
8900 | return 0; | |
8901 | } | |
8902 | #endif | |
8903 | ||
8904 | int sched_rt_handler(struct ctl_table *table, int write, | |
8905 | struct file *filp, void __user *buffer, size_t *lenp, | |
8906 | loff_t *ppos) | |
8907 | { | |
8908 | int ret; | |
8909 | int old_period, old_runtime; | |
8910 | static DEFINE_MUTEX(mutex); | |
8911 | ||
8912 | mutex_lock(&mutex); | |
8913 | old_period = sysctl_sched_rt_period; | |
8914 | old_runtime = sysctl_sched_rt_runtime; | |
8915 | ||
8916 | ret = proc_dointvec(table, write, filp, buffer, lenp, ppos); | |
8917 | ||
8918 | if (!ret && write) { | |
8919 | ret = sched_rt_global_constraints(); | |
8920 | if (ret) { | |
8921 | sysctl_sched_rt_period = old_period; | |
8922 | sysctl_sched_rt_runtime = old_runtime; | |
8923 | } else { | |
8924 | def_rt_bandwidth.rt_runtime = global_rt_runtime(); | |
8925 | def_rt_bandwidth.rt_period = | |
8926 | ns_to_ktime(global_rt_period()); | |
8927 | } | |
8928 | } | |
8929 | mutex_unlock(&mutex); | |
8930 | ||
8931 | return ret; | |
8932 | } | |
8933 | ||
8934 | #ifdef CONFIG_CGROUP_SCHED | |
8935 | ||
8936 | /* return corresponding task_group object of a cgroup */ | |
8937 | static inline struct task_group *cgroup_tg(struct cgroup *cgrp) | |
8938 | { | |
8939 | return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id), | |
8940 | struct task_group, css); | |
8941 | } | |
8942 | ||
8943 | static struct cgroup_subsys_state * | |
8944 | cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp) | |
8945 | { | |
8946 | struct task_group *tg, *parent; | |
8947 | ||
8948 | if (!cgrp->parent) { | |
8949 | /* This is early initialization for the top cgroup */ | |
8950 | init_task_group.css.cgroup = cgrp; | |
8951 | return &init_task_group.css; | |
8952 | } | |
8953 | ||
8954 | parent = cgroup_tg(cgrp->parent); | |
8955 | tg = sched_create_group(parent); | |
8956 | if (IS_ERR(tg)) | |
8957 | return ERR_PTR(-ENOMEM); | |
8958 | ||
8959 | /* Bind the cgroup to task_group object we just created */ | |
8960 | tg->css.cgroup = cgrp; | |
8961 | ||
8962 | return &tg->css; | |
8963 | } | |
8964 | ||
8965 | static void | |
8966 | cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp) | |
8967 | { | |
8968 | struct task_group *tg = cgroup_tg(cgrp); | |
8969 | ||
8970 | sched_destroy_group(tg); | |
8971 | } | |
8972 | ||
8973 | static int | |
8974 | cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp, | |
8975 | struct task_struct *tsk) | |
8976 | { | |
8977 | #ifdef CONFIG_RT_GROUP_SCHED | |
8978 | /* Don't accept realtime tasks when there is no way for them to run */ | |
8979 | if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0) | |
8980 | return -EINVAL; | |
8981 | #else | |
8982 | /* We don't support RT-tasks being in separate groups */ | |
8983 | if (tsk->sched_class != &fair_sched_class) | |
8984 | return -EINVAL; | |
8985 | #endif | |
8986 | ||
8987 | return 0; | |
8988 | } | |
8989 | ||
8990 | static void | |
8991 | cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp, | |
8992 | struct cgroup *old_cont, struct task_struct *tsk) | |
8993 | { | |
8994 | sched_move_task(tsk); | |
8995 | } | |
8996 | ||
8997 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
8998 | static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype, | |
8999 | u64 shareval) | |
9000 | { | |
9001 | return sched_group_set_shares(cgroup_tg(cgrp), shareval); | |
9002 | } | |
9003 | ||
9004 | static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft) | |
9005 | { | |
9006 | struct task_group *tg = cgroup_tg(cgrp); | |
9007 | ||
9008 | return (u64) tg->shares; | |
9009 | } | |
9010 | #endif | |
9011 | ||
9012 | #ifdef CONFIG_RT_GROUP_SCHED | |
9013 | static ssize_t cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft, | |
9014 | struct file *file, | |
9015 | const char __user *userbuf, | |
9016 | size_t nbytes, loff_t *unused_ppos) | |
9017 | { | |
9018 | char buffer[64]; | |
9019 | int retval = 0; | |
9020 | s64 val; | |
9021 | char *end; | |
9022 | ||
9023 | if (!nbytes) | |
9024 | return -EINVAL; | |
9025 | if (nbytes >= sizeof(buffer)) | |
9026 | return -E2BIG; | |
9027 | if (copy_from_user(buffer, userbuf, nbytes)) | |
9028 | return -EFAULT; | |
9029 | ||
9030 | buffer[nbytes] = 0; /* nul-terminate */ | |
9031 | ||
9032 | /* strip newline if necessary */ | |
9033 | if (nbytes && (buffer[nbytes-1] == '\n')) | |
9034 | buffer[nbytes-1] = 0; | |
9035 | val = simple_strtoll(buffer, &end, 0); | |
9036 | if (*end) | |
9037 | return -EINVAL; | |
9038 | ||
9039 | /* Pass to subsystem */ | |
9040 | retval = sched_group_set_rt_runtime(cgroup_tg(cgrp), val); | |
9041 | if (!retval) | |
9042 | retval = nbytes; | |
9043 | return retval; | |
9044 | } | |
9045 | ||
9046 | static ssize_t cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft, | |
9047 | struct file *file, | |
9048 | char __user *buf, size_t nbytes, | |
9049 | loff_t *ppos) | |
9050 | { | |
9051 | char tmp[64]; | |
9052 | long val = sched_group_rt_runtime(cgroup_tg(cgrp)); | |
9053 | int len = sprintf(tmp, "%ld\n", val); | |
9054 | ||
9055 | return simple_read_from_buffer(buf, nbytes, ppos, tmp, len); | |
9056 | } | |
9057 | ||
9058 | static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype, | |
9059 | u64 rt_period_us) | |
9060 | { | |
9061 | return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us); | |
9062 | } | |
9063 | ||
9064 | static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft) | |
9065 | { | |
9066 | return sched_group_rt_period(cgroup_tg(cgrp)); | |
9067 | } | |
9068 | #endif | |
9069 | ||
9070 | static struct cftype cpu_files[] = { | |
9071 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
9072 | { | |
9073 | .name = "shares", | |
9074 | .read_uint = cpu_shares_read_uint, | |
9075 | .write_uint = cpu_shares_write_uint, | |
9076 | }, | |
9077 | #endif | |
9078 | #ifdef CONFIG_RT_GROUP_SCHED | |
9079 | { | |
9080 | .name = "rt_runtime_us", | |
9081 | .read = cpu_rt_runtime_read, | |
9082 | .write = cpu_rt_runtime_write, | |
9083 | }, | |
9084 | { | |
9085 | .name = "rt_period_us", | |
9086 | .read_uint = cpu_rt_period_read_uint, | |
9087 | .write_uint = cpu_rt_period_write_uint, | |
9088 | }, | |
9089 | #endif | |
9090 | }; | |
9091 | ||
9092 | static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont) | |
9093 | { | |
9094 | return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files)); | |
9095 | } | |
9096 | ||
9097 | struct cgroup_subsys cpu_cgroup_subsys = { | |
9098 | .name = "cpu", | |
9099 | .create = cpu_cgroup_create, | |
9100 | .destroy = cpu_cgroup_destroy, | |
9101 | .can_attach = cpu_cgroup_can_attach, | |
9102 | .attach = cpu_cgroup_attach, | |
9103 | .populate = cpu_cgroup_populate, | |
9104 | .subsys_id = cpu_cgroup_subsys_id, | |
9105 | .early_init = 1, | |
9106 | }; | |
9107 | ||
9108 | #endif /* CONFIG_CGROUP_SCHED */ | |
9109 | ||
9110 | #ifdef CONFIG_CGROUP_CPUACCT | |
9111 | ||
9112 | /* | |
9113 | * CPU accounting code for task groups. | |
9114 | * | |
9115 | * Based on the work by Paul Menage (menage@google.com) and Balbir Singh | |
9116 | * (balbir@in.ibm.com). | |
9117 | */ | |
9118 | ||
9119 | /* track cpu usage of a group of tasks */ | |
9120 | struct cpuacct { | |
9121 | struct cgroup_subsys_state css; | |
9122 | /* cpuusage holds pointer to a u64-type object on every cpu */ | |
9123 | u64 *cpuusage; | |
9124 | }; | |
9125 | ||
9126 | struct cgroup_subsys cpuacct_subsys; | |
9127 | ||
9128 | /* return cpu accounting group corresponding to this container */ | |
9129 | static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp) | |
9130 | { | |
9131 | return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id), | |
9132 | struct cpuacct, css); | |
9133 | } | |
9134 | ||
9135 | /* return cpu accounting group to which this task belongs */ | |
9136 | static inline struct cpuacct *task_ca(struct task_struct *tsk) | |
9137 | { | |
9138 | return container_of(task_subsys_state(tsk, cpuacct_subsys_id), | |
9139 | struct cpuacct, css); | |
9140 | } | |
9141 | ||
9142 | /* create a new cpu accounting group */ | |
9143 | static struct cgroup_subsys_state *cpuacct_create( | |
9144 | struct cgroup_subsys *ss, struct cgroup *cgrp) | |
9145 | { | |
9146 | struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL); | |
9147 | ||
9148 | if (!ca) | |
9149 | return ERR_PTR(-ENOMEM); | |
9150 | ||
9151 | ca->cpuusage = alloc_percpu(u64); | |
9152 | if (!ca->cpuusage) { | |
9153 | kfree(ca); | |
9154 | return ERR_PTR(-ENOMEM); | |
9155 | } | |
9156 | ||
9157 | return &ca->css; | |
9158 | } | |
9159 | ||
9160 | /* destroy an existing cpu accounting group */ | |
9161 | static void | |
9162 | cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp) | |
9163 | { | |
9164 | struct cpuacct *ca = cgroup_ca(cgrp); | |
9165 | ||
9166 | free_percpu(ca->cpuusage); | |
9167 | kfree(ca); | |
9168 | } | |
9169 | ||
9170 | /* return total cpu usage (in nanoseconds) of a group */ | |
9171 | static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft) | |
9172 | { | |
9173 | struct cpuacct *ca = cgroup_ca(cgrp); | |
9174 | u64 totalcpuusage = 0; | |
9175 | int i; | |
9176 | ||
9177 | for_each_possible_cpu(i) { | |
9178 | u64 *cpuusage = percpu_ptr(ca->cpuusage, i); | |
9179 | ||
9180 | /* | |
9181 | * Take rq->lock to make 64-bit addition safe on 32-bit | |
9182 | * platforms. | |
9183 | */ | |
9184 | spin_lock_irq(&cpu_rq(i)->lock); | |
9185 | totalcpuusage += *cpuusage; | |
9186 | spin_unlock_irq(&cpu_rq(i)->lock); | |
9187 | } | |
9188 | ||
9189 | return totalcpuusage; | |
9190 | } | |
9191 | ||
9192 | static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype, | |
9193 | u64 reset) | |
9194 | { | |
9195 | struct cpuacct *ca = cgroup_ca(cgrp); | |
9196 | int err = 0; | |
9197 | int i; | |
9198 | ||
9199 | if (reset) { | |
9200 | err = -EINVAL; | |
9201 | goto out; | |
9202 | } | |
9203 | ||
9204 | for_each_possible_cpu(i) { | |
9205 | u64 *cpuusage = percpu_ptr(ca->cpuusage, i); | |
9206 | ||
9207 | spin_lock_irq(&cpu_rq(i)->lock); | |
9208 | *cpuusage = 0; | |
9209 | spin_unlock_irq(&cpu_rq(i)->lock); | |
9210 | } | |
9211 | out: | |
9212 | return err; | |
9213 | } | |
9214 | ||
9215 | static struct cftype files[] = { | |
9216 | { | |
9217 | .name = "usage", | |
9218 | .read_uint = cpuusage_read, | |
9219 | .write_uint = cpuusage_write, | |
9220 | }, | |
9221 | }; | |
9222 | ||
9223 | static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp) | |
9224 | { | |
9225 | return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files)); | |
9226 | } | |
9227 | ||
9228 | /* | |
9229 | * charge this task's execution time to its accounting group. | |
9230 | * | |
9231 | * called with rq->lock held. | |
9232 | */ | |
9233 | static void cpuacct_charge(struct task_struct *tsk, u64 cputime) | |
9234 | { | |
9235 | struct cpuacct *ca; | |
9236 | ||
9237 | if (!cpuacct_subsys.active) | |
9238 | return; | |
9239 | ||
9240 | ca = task_ca(tsk); | |
9241 | if (ca) { | |
9242 | u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk)); | |
9243 | ||
9244 | *cpuusage += cputime; | |
9245 | } | |
9246 | } | |
9247 | ||
9248 | struct cgroup_subsys cpuacct_subsys = { | |
9249 | .name = "cpuacct", | |
9250 | .create = cpuacct_create, | |
9251 | .destroy = cpuacct_destroy, | |
9252 | .populate = cpuacct_populate, | |
9253 | .subsys_id = cpuacct_subsys_id, | |
9254 | }; | |
9255 | #endif /* CONFIG_CGROUP_CPUACCT */ |