Contents of /alx-src/tags/kernel26-2.6.12-alx-r9/kernel/sched.c
Parent Directory | Revision Log
Revision 630 -
(show annotations)
(download)
Wed Mar 4 11:03:09 2009 UTC (15 years, 6 months ago) by niro
File MIME type: text/plain
File size: 122889 byte(s)
Wed Mar 4 11:03:09 2009 UTC (15 years, 6 months ago) by niro
File MIME type: text/plain
File size: 122889 byte(s)
Tag kernel26-2.6.12-alx-r9
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 | * 2005-08-20 New staircase scheduling policy by Con Kolivas with help |
20 | * from William Lee Irwin III, Zwane Mwaikambo & Peter Williams. |
21 | * Staircase v11.6 |
22 | */ |
23 | |
24 | #include <linux/mm.h> |
25 | #include <linux/module.h> |
26 | #include <linux/nmi.h> |
27 | #include <linux/init.h> |
28 | #include <asm/uaccess.h> |
29 | #include <linux/highmem.h> |
30 | #include <linux/smp_lock.h> |
31 | #include <asm/mmu_context.h> |
32 | #include <linux/interrupt.h> |
33 | #include <linux/completion.h> |
34 | #include <linux/kernel_stat.h> |
35 | #include <linux/security.h> |
36 | #include <linux/notifier.h> |
37 | #include <linux/profile.h> |
38 | #include <linux/suspend.h> |
39 | #include <linux/blkdev.h> |
40 | #include <linux/delay.h> |
41 | #include <linux/smp.h> |
42 | #include <linux/threads.h> |
43 | #include <linux/timer.h> |
44 | #include <linux/rcupdate.h> |
45 | #include <linux/cpu.h> |
46 | #include <linux/cpuset.h> |
47 | #include <linux/percpu.h> |
48 | #include <linux/kthread.h> |
49 | #include <linux/seq_file.h> |
50 | #include <linux/syscalls.h> |
51 | #include <linux/times.h> |
52 | #include <linux/acct.h> |
53 | #include <asm/tlb.h> |
54 | |
55 | #include <asm/unistd.h> |
56 | |
57 | /* |
58 | * Convert user-nice values [ -20 ... 0 ... 19 ] |
59 | * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ], |
60 | * and back. |
61 | */ |
62 | #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20) |
63 | #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20) |
64 | #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio) |
65 | |
66 | /* |
67 | * 'User priority' is the nice value converted to something we |
68 | * can work with better when scaling various scheduler parameters, |
69 | * it's a [ 0 ... 39 ] range. |
70 | */ |
71 | #define USER_PRIO(p) ((p)-MAX_RT_PRIO) |
72 | #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio) |
73 | #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO)) |
74 | |
75 | /* |
76 | * Some helpers for converting nanosecond timing to jiffy resolution |
77 | */ |
78 | #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ)) |
79 | #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ)) |
80 | #define NSJIFFY (1000000000 / HZ) /* One jiffy in ns */ |
81 | |
82 | int sched_compute = 0; |
83 | /* |
84 | *This is the time all tasks within the same priority round robin. |
85 | *compute setting is reserved for dedicated computational scheduling |
86 | *and has ten times larger intervals. Set to a minimum of 5ms. |
87 | */ |
88 | #define _RR_INTERVAL ((5 * HZ / 1001) + 1) |
89 | #define RR_INTERVAL() (_RR_INTERVAL * (1 + 19 * sched_compute)) |
90 | #define DEF_TIMESLICE (RR_INTERVAL() * 19) |
91 | |
92 | #define task_hot(p, now, sd) ((long long) ((now) - (p)->timestamp) \ |
93 | < (long long) (sd)->cache_hot_time) |
94 | |
95 | /* |
96 | * These are the runqueue data structures: |
97 | */ |
98 | |
99 | typedef struct runqueue runqueue_t; |
100 | |
101 | /* |
102 | * This is the main, per-CPU runqueue data structure. |
103 | * |
104 | * Locking rule: those places that want to lock multiple runqueues |
105 | * (such as the load balancing or the thread migration code), lock |
106 | * acquire operations must be ordered by ascending &runqueue. |
107 | */ |
108 | struct runqueue { |
109 | spinlock_t lock; |
110 | |
111 | /* |
112 | * nr_running and cpu_load should be in the same cacheline because |
113 | * remote CPUs use both these fields when doing load calculation. |
114 | */ |
115 | unsigned long nr_running; |
116 | #ifdef CONFIG_SMP |
117 | unsigned long prio_bias; |
118 | unsigned long cpu_load; |
119 | #endif |
120 | unsigned long long nr_switches; |
121 | |
122 | /* |
123 | * This is part of a global counter where only the total sum |
124 | * over all CPUs matters. A task can increase this counter on |
125 | * one CPU and if it got migrated afterwards it may decrease |
126 | * it on another CPU. Always updated under the runqueue lock: |
127 | */ |
128 | unsigned long nr_uninterruptible; |
129 | |
130 | unsigned long long timestamp_last_tick; |
131 | unsigned int cache_ticks, preempted; |
132 | task_t *curr, *idle; |
133 | struct mm_struct *prev_mm; |
134 | unsigned long bitmap[BITS_TO_LONGS(MAX_PRIO + 1)]; |
135 | struct list_head queue[MAX_PRIO]; |
136 | atomic_t nr_iowait; |
137 | |
138 | #ifdef CONFIG_SMP |
139 | struct sched_domain *sd; |
140 | |
141 | /* For active balancing */ |
142 | int active_balance; |
143 | int push_cpu; |
144 | |
145 | task_t *migration_thread; |
146 | struct list_head migration_queue; |
147 | #endif |
148 | |
149 | #ifdef CONFIG_SCHEDSTATS |
150 | /* latency stats */ |
151 | struct sched_info rq_sched_info; |
152 | |
153 | /* sys_sched_yield() stats */ |
154 | unsigned long yld_exp_empty; |
155 | unsigned long yld_act_empty; |
156 | unsigned long yld_both_empty; |
157 | unsigned long yld_cnt; |
158 | |
159 | /* schedule() stats */ |
160 | unsigned long sched_switch; |
161 | unsigned long sched_cnt; |
162 | unsigned long sched_goidle; |
163 | |
164 | /* try_to_wake_up() stats */ |
165 | unsigned long ttwu_cnt; |
166 | unsigned long ttwu_local; |
167 | #endif |
168 | }; |
169 | |
170 | static DEFINE_PER_CPU(struct runqueue, runqueues); |
171 | |
172 | #define for_each_domain(cpu, domain) \ |
173 | for (domain = cpu_rq(cpu)->sd; domain; domain = domain->parent) |
174 | |
175 | #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) |
176 | #define this_rq() (&__get_cpu_var(runqueues)) |
177 | #define task_rq(p) cpu_rq(task_cpu(p)) |
178 | #define cpu_curr(cpu) (cpu_rq(cpu)->curr) |
179 | |
180 | /* |
181 | * Default context-switch locking: |
182 | */ |
183 | #ifndef prepare_arch_switch |
184 | # define prepare_arch_switch(rq, next) do { } while (0) |
185 | # define finish_arch_switch(rq, next) spin_unlock_irq(&(rq)->lock) |
186 | # define task_running(rq, p) ((rq)->curr == (p)) |
187 | #endif |
188 | |
189 | /* |
190 | * task_rq_lock - lock the runqueue a given task resides on and disable |
191 | * interrupts. Note the ordering: we can safely lookup the task_rq without |
192 | * explicitly disabling preemption. |
193 | */ |
194 | static inline runqueue_t *task_rq_lock(task_t *p, unsigned long *flags) |
195 | __acquires(rq->lock) |
196 | { |
197 | struct runqueue *rq; |
198 | |
199 | repeat_lock_task: |
200 | local_irq_save(*flags); |
201 | rq = task_rq(p); |
202 | spin_lock(&rq->lock); |
203 | if (unlikely(rq != task_rq(p))) { |
204 | spin_unlock_irqrestore(&rq->lock, *flags); |
205 | goto repeat_lock_task; |
206 | } |
207 | return rq; |
208 | } |
209 | |
210 | static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags) |
211 | __releases(rq->lock) |
212 | { |
213 | spin_unlock_irqrestore(&rq->lock, *flags); |
214 | } |
215 | |
216 | #ifdef CONFIG_SCHEDSTATS |
217 | /* |
218 | * bump this up when changing the output format or the meaning of an existing |
219 | * format, so that tools can adapt (or abort) |
220 | */ |
221 | #define SCHEDSTAT_VERSION 11 |
222 | |
223 | static int show_schedstat(struct seq_file *seq, void *v) |
224 | { |
225 | int cpu; |
226 | |
227 | seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION); |
228 | seq_printf(seq, "timestamp %lu\n", jiffies); |
229 | for_each_online_cpu(cpu) { |
230 | runqueue_t *rq = cpu_rq(cpu); |
231 | #ifdef CONFIG_SMP |
232 | struct sched_domain *sd; |
233 | int dcnt = 0; |
234 | #endif |
235 | |
236 | /* runqueue-specific stats */ |
237 | seq_printf(seq, |
238 | "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu", |
239 | cpu, rq->yld_both_empty, |
240 | rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt, |
241 | rq->sched_switch, rq->sched_cnt, rq->sched_goidle, |
242 | rq->ttwu_cnt, rq->ttwu_local, |
243 | rq->rq_sched_info.cpu_time, |
244 | rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt); |
245 | |
246 | seq_printf(seq, "\n"); |
247 | |
248 | #ifdef CONFIG_SMP |
249 | /* domain-specific stats */ |
250 | for_each_domain(cpu, sd) { |
251 | enum idle_type itype; |
252 | char mask_str[NR_CPUS]; |
253 | |
254 | cpumask_scnprintf(mask_str, NR_CPUS, sd->span); |
255 | seq_printf(seq, "domain%d %s", dcnt++, mask_str); |
256 | for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES; |
257 | itype++) { |
258 | seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu", |
259 | sd->lb_cnt[itype], |
260 | sd->lb_balanced[itype], |
261 | sd->lb_failed[itype], |
262 | sd->lb_imbalance[itype], |
263 | sd->lb_gained[itype], |
264 | sd->lb_hot_gained[itype], |
265 | sd->lb_nobusyq[itype], |
266 | sd->lb_nobusyg[itype]); |
267 | } |
268 | seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu\n", |
269 | sd->alb_cnt, sd->alb_failed, sd->alb_pushed, |
270 | sd->sbe_pushed, sd->sbe_attempts, |
271 | sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance); |
272 | } |
273 | #endif |
274 | } |
275 | return 0; |
276 | } |
277 | |
278 | static int schedstat_open(struct inode *inode, struct file *file) |
279 | { |
280 | unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32); |
281 | char *buf = kmalloc(size, GFP_KERNEL); |
282 | struct seq_file *m; |
283 | int res; |
284 | |
285 | if (!buf) |
286 | return -ENOMEM; |
287 | res = single_open(file, show_schedstat, NULL); |
288 | if (!res) { |
289 | m = file->private_data; |
290 | m->buf = buf; |
291 | m->size = size; |
292 | } else |
293 | kfree(buf); |
294 | return res; |
295 | } |
296 | |
297 | struct file_operations proc_schedstat_operations = { |
298 | .open = schedstat_open, |
299 | .read = seq_read, |
300 | .llseek = seq_lseek, |
301 | .release = single_release, |
302 | }; |
303 | |
304 | # define schedstat_inc(rq, field) do { (rq)->field++; } while (0) |
305 | # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0) |
306 | #else /* !CONFIG_SCHEDSTATS */ |
307 | # define schedstat_inc(rq, field) do { } while (0) |
308 | # define schedstat_add(rq, field, amt) do { } while (0) |
309 | #endif |
310 | |
311 | /* |
312 | * rq_lock - lock a given runqueue and disable interrupts. |
313 | */ |
314 | static inline runqueue_t *this_rq_lock(void) |
315 | __acquires(rq->lock) |
316 | { |
317 | runqueue_t *rq; |
318 | |
319 | local_irq_disable(); |
320 | rq = this_rq(); |
321 | spin_lock(&rq->lock); |
322 | |
323 | return rq; |
324 | } |
325 | |
326 | #ifdef CONFIG_SCHED_SMT |
327 | static int cpu_and_siblings_are_idle(int cpu) |
328 | { |
329 | int sib; |
330 | for_each_cpu_mask(sib, cpu_sibling_map[cpu]) { |
331 | if (idle_cpu(sib)) |
332 | continue; |
333 | return 0; |
334 | } |
335 | |
336 | return 1; |
337 | } |
338 | #else |
339 | #define cpu_and_siblings_are_idle(A) idle_cpu(A) |
340 | #endif |
341 | |
342 | #ifdef CONFIG_SCHEDSTATS |
343 | /* |
344 | * Called when a process is dequeued from the active array and given |
345 | * the cpu. We should note that with the exception of interactive |
346 | * tasks, the expired queue will become the active queue after the active |
347 | * queue is empty, without explicitly dequeuing and requeuing tasks in the |
348 | * expired queue. (Interactive tasks may be requeued directly to the |
349 | * active queue, thus delaying tasks in the expired queue from running; |
350 | * see scheduler_tick()). |
351 | * |
352 | * This function is only called from sched_info_arrive(), rather than |
353 | * dequeue_task(). Even though a task may be queued and dequeued multiple |
354 | * times as it is shuffled about, we're really interested in knowing how |
355 | * long it was from the *first* time it was queued to the time that it |
356 | * finally hit a cpu. |
357 | */ |
358 | static inline void sched_info_dequeued(task_t *t) |
359 | { |
360 | t->sched_info.last_queued = 0; |
361 | } |
362 | |
363 | /* |
364 | * Called when a task finally hits the cpu. We can now calculate how |
365 | * long it was waiting to run. We also note when it began so that we |
366 | * can keep stats on how long its timeslice is. |
367 | */ |
368 | static inline void sched_info_arrive(task_t *t) |
369 | { |
370 | unsigned long now = jiffies, diff = 0; |
371 | struct runqueue *rq = task_rq(t); |
372 | |
373 | if (t->sched_info.last_queued) |
374 | diff = now - t->sched_info.last_queued; |
375 | sched_info_dequeued(t); |
376 | t->sched_info.run_delay += diff; |
377 | t->sched_info.last_arrival = now; |
378 | t->sched_info.pcnt++; |
379 | |
380 | if (!rq) |
381 | return; |
382 | |
383 | rq->rq_sched_info.run_delay += diff; |
384 | rq->rq_sched_info.pcnt++; |
385 | } |
386 | |
387 | /* |
388 | * Called when a process is queued into either the active or expired |
389 | * array. The time is noted and later used to determine how long we |
390 | * had to wait for us to reach the cpu. Since the expired queue will |
391 | * become the active queue after active queue is empty, without dequeuing |
392 | * and requeuing any tasks, we are interested in queuing to either. It |
393 | * is unusual but not impossible for tasks to be dequeued and immediately |
394 | * requeued in the same or another array: this can happen in sched_yield(), |
395 | * set_user_nice(), and even load_balance() as it moves tasks from runqueue |
396 | * to runqueue. |
397 | * |
398 | * This function is only called from enqueue_task(), but also only updates |
399 | * the timestamp if it is already not set. It's assumed that |
400 | * sched_info_dequeued() will clear that stamp when appropriate. |
401 | */ |
402 | static inline void sched_info_queued(task_t *t) |
403 | { |
404 | if (!t->sched_info.last_queued) |
405 | t->sched_info.last_queued = jiffies; |
406 | } |
407 | |
408 | /* |
409 | * Called when a process ceases being the active-running process, either |
410 | * voluntarily or involuntarily. Now we can calculate how long we ran. |
411 | */ |
412 | static inline void sched_info_depart(task_t *t) |
413 | { |
414 | struct runqueue *rq = task_rq(t); |
415 | unsigned long diff = jiffies - t->sched_info.last_arrival; |
416 | |
417 | t->sched_info.cpu_time += diff; |
418 | |
419 | if (rq) |
420 | rq->rq_sched_info.cpu_time += diff; |
421 | } |
422 | |
423 | /* |
424 | * Called when tasks are switched involuntarily due, typically, to expiring |
425 | * their time slice. (This may also be called when switching to or from |
426 | * the idle task.) We are only called when prev != next. |
427 | */ |
428 | static inline void sched_info_switch(task_t *prev, task_t *next) |
429 | { |
430 | struct runqueue *rq = task_rq(prev); |
431 | |
432 | /* |
433 | * prev now departs the cpu. It's not interesting to record |
434 | * stats about how efficient we were at scheduling the idle |
435 | * process, however. |
436 | */ |
437 | if (prev != rq->idle) |
438 | sched_info_depart(prev); |
439 | |
440 | if (next != rq->idle) |
441 | sched_info_arrive(next); |
442 | } |
443 | #else |
444 | #define sched_info_queued(t) do { } while (0) |
445 | #define sched_info_switch(t, next) do { } while (0) |
446 | #endif /* CONFIG_SCHEDSTATS */ |
447 | |
448 | /* |
449 | * Get nanosecond clock difference without overflowing unsigned long. |
450 | */ |
451 | static inline unsigned long ns_diff(unsigned long long v1, unsigned long long v2) |
452 | { |
453 | unsigned long long vdiff; |
454 | if (likely(v1 > v2)) { |
455 | vdiff = v1 - v2; |
456 | if (vdiff > (1 << 31)) |
457 | vdiff = 1 << 31; |
458 | } else |
459 | /* |
460 | * Rarely the clock appears to go backwards. There should |
461 | * always be a positive difference so return 1. |
462 | */ |
463 | vdiff = 1; |
464 | return (unsigned long)vdiff; |
465 | } |
466 | |
467 | static inline int task_queued(task_t *task) |
468 | { |
469 | return !list_empty(&task->run_list); |
470 | } |
471 | |
472 | /* |
473 | * Adding/removing a task to/from a runqueue: |
474 | */ |
475 | static inline void dequeue_task(struct task_struct *p, runqueue_t *rq) |
476 | { |
477 | list_del_init(&p->run_list); |
478 | if (list_empty(rq->queue + p->prio)) |
479 | __clear_bit(p->prio, rq->bitmap); |
480 | p->ns_debit = 0; |
481 | } |
482 | |
483 | static void enqueue_task(struct task_struct *p, runqueue_t *rq) |
484 | { |
485 | list_add_tail(&p->run_list, rq->queue + p->prio); |
486 | __set_bit(p->prio, rq->bitmap); |
487 | } |
488 | |
489 | /* |
490 | * Put task to the end of the run list without the overhead of dequeue |
491 | * followed by enqueue. |
492 | */ |
493 | static inline void requeue_task(struct task_struct *p, runqueue_t *rq) |
494 | { |
495 | list_move_tail(&p->run_list, rq->queue + p->prio); |
496 | } |
497 | |
498 | static inline void enqueue_task_head(struct task_struct *p, runqueue_t *rq) |
499 | { |
500 | list_add(&p->run_list, rq->queue + p->prio); |
501 | __set_bit(p->prio, rq->bitmap); |
502 | } |
503 | |
504 | #ifdef CONFIG_SMP |
505 | static inline void inc_prio_bias(runqueue_t *rq, int prio) |
506 | { |
507 | rq->prio_bias += MAX_PRIO - prio; |
508 | } |
509 | |
510 | static inline void dec_prio_bias(runqueue_t *rq, int prio) |
511 | { |
512 | rq->prio_bias -= MAX_PRIO - prio; |
513 | } |
514 | #else |
515 | static inline void inc_prio_bias(runqueue_t *rq, int prio) |
516 | { |
517 | } |
518 | |
519 | static inline void dec_prio_bias(runqueue_t *rq, int prio) |
520 | { |
521 | } |
522 | #endif |
523 | |
524 | static inline void inc_nr_running(task_t *p, runqueue_t *rq) |
525 | { |
526 | rq->nr_running++; |
527 | #ifdef CONFIG_SMP |
528 | if (rt_task(p)) { |
529 | if (p != rq->migration_thread) |
530 | inc_prio_bias(rq, p->prio); |
531 | } else |
532 | inc_prio_bias(rq, p->static_prio); |
533 | #endif |
534 | } |
535 | |
536 | static inline void dec_nr_running(task_t *p, runqueue_t *rq) |
537 | { |
538 | rq->nr_running--; |
539 | #ifdef CONFIG_SMP |
540 | if (rt_task(p)) { |
541 | if (p != rq->migration_thread) |
542 | dec_prio_bias(rq, p->prio); |
543 | } else |
544 | dec_prio_bias(rq, p->static_prio); |
545 | #endif |
546 | } |
547 | |
548 | /* |
549 | * __activate_task - move a task to the runqueue. |
550 | */ |
551 | static void __activate_task(task_t *p, runqueue_t *rq) |
552 | { |
553 | enqueue_task(p, rq); |
554 | inc_nr_running(p, rq); |
555 | } |
556 | |
557 | /* |
558 | * __activate_idle_task - move idle task to the _front_ of runqueue. |
559 | */ |
560 | static inline void __activate_idle_task(task_t *p, runqueue_t *rq) |
561 | { |
562 | enqueue_task_head(p, rq); |
563 | inc_nr_running(p, rq); |
564 | } |
565 | |
566 | /* |
567 | * burst - extra intervals an interactive task can run for at best priority |
568 | * instead of descending priorities. |
569 | */ |
570 | static inline unsigned int burst(task_t *p) |
571 | { |
572 | if (likely(!rt_task(p))) { |
573 | unsigned int task_user_prio = TASK_USER_PRIO(p); |
574 | return 39 - task_user_prio; |
575 | } else |
576 | return p->burst; |
577 | } |
578 | |
579 | static void inc_burst(task_t *p) |
580 | { |
581 | unsigned int best_burst; |
582 | best_burst = burst(p); |
583 | if (p->burst < best_burst) |
584 | p->burst++; |
585 | } |
586 | |
587 | static void dec_burst(task_t *p) |
588 | { |
589 | if (p->burst) |
590 | p->burst--; |
591 | } |
592 | |
593 | static inline unsigned int rr_interval(task_t * p) |
594 | { |
595 | unsigned int rr_interval = RR_INTERVAL(); |
596 | int nice = TASK_NICE(p); |
597 | |
598 | if (nice < 0 && !rt_task(p)) |
599 | rr_interval += -(nice); |
600 | return rr_interval; |
601 | } |
602 | |
603 | /* |
604 | * slice - the duration a task runs before getting requeued at its best |
605 | * priority and has its burst decremented. |
606 | */ |
607 | static inline unsigned int slice(task_t *p) |
608 | { |
609 | unsigned int slice, rr; |
610 | slice = rr = rr_interval(p); |
611 | if (likely(!rt_task(p))) |
612 | slice += burst(p) * rr; |
613 | return slice; |
614 | } |
615 | |
616 | /* |
617 | * sched_interactive - sysctl which allows interactive tasks to have bursts |
618 | */ |
619 | int sched_interactive = 1; |
620 | |
621 | /* |
622 | * effective_prio - dynamic priority dependent on burst. |
623 | * The priority normally decreases by one each RR_INTERVAL. |
624 | * As the burst increases the priority stays at the top "stair" or |
625 | * priority for longer. |
626 | */ |
627 | static int effective_prio(task_t *p) |
628 | { |
629 | int prio; |
630 | unsigned int full_slice, used_slice, first_slice; |
631 | unsigned int best_burst, rr; |
632 | if (rt_task(p)) |
633 | return p->prio; |
634 | if (batch_task(p)) { |
635 | if (unlikely(p->flags & (PF_NONSLEEP | PF_FREEZE))) { |
636 | /* |
637 | * If batch is waking up from in kernel activity |
638 | * or being frozen, reschedule at a normal priority |
639 | * to begin with. |
640 | */ |
641 | p->flags |= PF_YIELDED; |
642 | return MAX_PRIO - 2; |
643 | } |
644 | return MAX_PRIO - 1; |
645 | } |
646 | |
647 | best_burst = burst(p); |
648 | full_slice = slice(p); |
649 | rr = rr_interval(p); |
650 | used_slice = full_slice - p->slice; |
651 | if (p->burst > best_burst) |
652 | p->burst = best_burst; |
653 | first_slice = rr; |
654 | if (sched_interactive && !sched_compute && p->mm) |
655 | first_slice *= (p->burst + 1); |
656 | prio = MAX_PRIO - 2 - best_burst; |
657 | |
658 | if (used_slice < first_slice) |
659 | return prio; |
660 | prio += 1 + (used_slice - first_slice) / rr; |
661 | if (prio >= MAX_PRIO - 2) |
662 | prio = MAX_PRIO - 2; |
663 | return prio; |
664 | } |
665 | |
666 | static void continue_slice(task_t *p) |
667 | { |
668 | unsigned long total_run = NS_TO_JIFFIES(p->totalrun); |
669 | |
670 | if (total_run >= p->slice) { |
671 | p->totalrun -= JIFFIES_TO_NS(p->slice); |
672 | dec_burst(p); |
673 | } else { |
674 | unsigned int remainder; |
675 | p->slice -= total_run; |
676 | remainder = p->slice % rr_interval(p); |
677 | if (remainder) |
678 | p->time_slice = remainder; |
679 | } |
680 | } |
681 | |
682 | /* |
683 | * recalc_task_prio - this checks for tasks that run ultra short timeslices |
684 | * or have just forked a thread/process and make them continue their old |
685 | * slice instead of starting a new one at high priority. |
686 | */ |
687 | static inline void recalc_task_prio(task_t *p, unsigned long long now, |
688 | unsigned long rq_running) |
689 | { |
690 | unsigned long sleep_time = ns_diff(now, p->timestamp); |
691 | |
692 | /* |
693 | * Priority is elevated back to best by amount of sleep_time. |
694 | * sleep_time is scaled down by number of tasks currently running. |
695 | */ |
696 | if (rq_running > 1) |
697 | sleep_time /= rq_running; |
698 | |
699 | p->totalrun += p->runtime; |
700 | if (NS_TO_JIFFIES(p->totalrun) >= p->slice && |
701 | NS_TO_JIFFIES(sleep_time) < p->slice) { |
702 | p->flags &= ~PF_NONSLEEP; |
703 | dec_burst(p); |
704 | p->totalrun += sleep_time - JIFFIES_TO_NS(p->slice); |
705 | goto out; |
706 | } |
707 | |
708 | if (p->flags & PF_NONSLEEP) { |
709 | continue_slice(p); |
710 | p->flags &= ~PF_NONSLEEP; |
711 | goto out; |
712 | } |
713 | |
714 | if (sched_compute) { |
715 | continue_slice(p); |
716 | goto out; |
717 | } |
718 | |
719 | if (sleep_time >= p->totalrun) { |
720 | if (!(p->flags & PF_NONSLEEP)) |
721 | inc_burst(p); |
722 | p->totalrun = 0; |
723 | goto out; |
724 | } |
725 | |
726 | p->totalrun -= sleep_time; |
727 | continue_slice(p); |
728 | out: |
729 | return; |
730 | } |
731 | |
732 | /* |
733 | * activate_task - move a task to the runqueue and do priority recalculation |
734 | * |
735 | * Update all the scheduling statistics stuff. (sleep average |
736 | * calculation, priority modifiers, etc.) |
737 | */ |
738 | static void activate_task(task_t *p, runqueue_t *rq, int local) |
739 | { |
740 | unsigned long long now = sched_clock(); |
741 | #ifdef CONFIG_SMP |
742 | if (!local) { |
743 | /* Compensate for drifting sched_clock */ |
744 | runqueue_t *this_rq = this_rq(); |
745 | now = (now - this_rq->timestamp_last_tick) |
746 | + rq->timestamp_last_tick; |
747 | } |
748 | #endif |
749 | p->slice = slice(p); |
750 | p->time_slice = rr_interval(p); |
751 | recalc_task_prio(p, now, rq->nr_running); |
752 | p->flags &= ~PF_NONSLEEP; |
753 | p->prio = effective_prio(p); |
754 | p->timestamp = now; |
755 | __activate_task(p, rq); |
756 | } |
757 | |
758 | /* |
759 | * deactivate_task - remove a task from the runqueue. |
760 | */ |
761 | static inline void deactivate_task(struct task_struct *p, runqueue_t *rq) |
762 | { |
763 | dec_nr_running(p, rq); |
764 | dequeue_task(p, rq); |
765 | } |
766 | |
767 | /* |
768 | * resched_task - mark a task 'to be rescheduled now'. |
769 | * |
770 | * On UP this means the setting of the need_resched flag, on SMP it |
771 | * might also involve a cross-CPU call to trigger the scheduler on |
772 | * the target CPU. |
773 | */ |
774 | #ifdef CONFIG_SMP |
775 | static void resched_task(task_t *p) |
776 | { |
777 | int need_resched, nrpolling; |
778 | |
779 | assert_spin_locked(&task_rq(p)->lock); |
780 | |
781 | /* minimise the chance of sending an interrupt to poll_idle() */ |
782 | nrpolling = test_tsk_thread_flag(p,TIF_POLLING_NRFLAG); |
783 | need_resched = test_and_set_tsk_thread_flag(p,TIF_NEED_RESCHED); |
784 | nrpolling |= test_tsk_thread_flag(p,TIF_POLLING_NRFLAG); |
785 | |
786 | if (!need_resched && !nrpolling && (task_cpu(p) != smp_processor_id())) |
787 | smp_send_reschedule(task_cpu(p)); |
788 | } |
789 | #else |
790 | static inline void resched_task(task_t *p) |
791 | { |
792 | set_tsk_need_resched(p); |
793 | } |
794 | #endif |
795 | |
796 | /** |
797 | * task_curr - is this task currently executing on a CPU? |
798 | * @p: the task in question. |
799 | */ |
800 | inline int task_curr(const task_t *p) |
801 | { |
802 | return cpu_curr(task_cpu(p)) == p; |
803 | } |
804 | |
805 | #ifdef CONFIG_SMP |
806 | enum request_type { |
807 | REQ_MOVE_TASK, |
808 | REQ_SET_DOMAIN, |
809 | }; |
810 | |
811 | typedef struct { |
812 | struct list_head list; |
813 | enum request_type type; |
814 | |
815 | /* For REQ_MOVE_TASK */ |
816 | task_t *task; |
817 | int dest_cpu; |
818 | |
819 | /* For REQ_SET_DOMAIN */ |
820 | struct sched_domain *sd; |
821 | |
822 | struct completion done; |
823 | } migration_req_t; |
824 | |
825 | /* |
826 | * The task's runqueue lock must be held. |
827 | * Returns true if you have to wait for migration thread. |
828 | */ |
829 | static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req) |
830 | { |
831 | runqueue_t *rq = task_rq(p); |
832 | |
833 | /* |
834 | * If the task is not on a runqueue (and not running), then |
835 | * it is sufficient to simply update the task's cpu field. |
836 | */ |
837 | if (!task_queued(p) && !task_running(rq, p)) { |
838 | set_task_cpu(p, dest_cpu); |
839 | return 0; |
840 | } |
841 | |
842 | init_completion(&req->done); |
843 | req->type = REQ_MOVE_TASK; |
844 | req->task = p; |
845 | req->dest_cpu = dest_cpu; |
846 | list_add(&req->list, &rq->migration_queue); |
847 | return 1; |
848 | } |
849 | |
850 | /* |
851 | * wait_task_inactive - wait for a thread to unschedule. |
852 | * |
853 | * The caller must ensure that the task *will* unschedule sometime soon, |
854 | * else this function might spin for a *long* time. This function can't |
855 | * be called with interrupts off, or it may introduce deadlock with |
856 | * smp_call_function() if an IPI is sent by the same process we are |
857 | * waiting to become inactive. |
858 | */ |
859 | void wait_task_inactive(task_t * p) |
860 | { |
861 | unsigned long flags; |
862 | runqueue_t *rq; |
863 | int preempted; |
864 | |
865 | repeat: |
866 | rq = task_rq_lock(p, &flags); |
867 | /* Must be off runqueue entirely, not preempted. */ |
868 | if (unlikely(task_queued(p) || task_running(rq, p))) { |
869 | /* If it's preempted, we yield. It could be a while. */ |
870 | preempted = !task_running(rq, p); |
871 | task_rq_unlock(rq, &flags); |
872 | cpu_relax(); |
873 | if (preempted) |
874 | yield(); |
875 | goto repeat; |
876 | } |
877 | task_rq_unlock(rq, &flags); |
878 | } |
879 | |
880 | /*** |
881 | * kick_process - kick a running thread to enter/exit the kernel |
882 | * @p: the to-be-kicked thread |
883 | * |
884 | * Cause a process which is running on another CPU to enter |
885 | * kernel-mode, without any delay. (to get signals handled.) |
886 | * |
887 | * NOTE: this function doesnt have to take the runqueue lock, |
888 | * because all it wants to ensure is that the remote task enters |
889 | * the kernel. If the IPI races and the task has been migrated |
890 | * to another CPU then no harm is done and the purpose has been |
891 | * achieved as well. |
892 | */ |
893 | void kick_process(task_t *p) |
894 | { |
895 | int cpu; |
896 | |
897 | preempt_disable(); |
898 | cpu = task_cpu(p); |
899 | if ((cpu != smp_processor_id()) && task_curr(p)) |
900 | smp_send_reschedule(cpu); |
901 | preempt_enable(); |
902 | } |
903 | |
904 | /* |
905 | * Return a low guess at the load of a migration-source cpu. |
906 | * |
907 | * We want to under-estimate the load of migration sources, to |
908 | * balance conservatively. |
909 | */ |
910 | static inline unsigned long __source_load(int cpu, enum idle_type idle) |
911 | { |
912 | runqueue_t *rq = cpu_rq(cpu); |
913 | unsigned long running = rq->nr_running; |
914 | unsigned long source_load, cpu_load = rq->cpu_load, |
915 | load_now = running * SCHED_LOAD_SCALE; |
916 | |
917 | source_load = min(cpu_load, load_now); |
918 | |
919 | if (running > 1 || (idle == NOT_IDLE && running)) |
920 | /* |
921 | * If we are busy rebalancing the load is biased by |
922 | * priority to create 'nice' support across cpus. When |
923 | * idle rebalancing we should only bias the source_load if |
924 | * there is more than one task running on that queue to |
925 | * prevent idle rebalance from trying to pull tasks from a |
926 | * queue with only one running task. |
927 | */ |
928 | source_load = source_load * rq->prio_bias / running; |
929 | |
930 | return source_load; |
931 | } |
932 | |
933 | static inline unsigned long source_load(int cpu) |
934 | { |
935 | return __source_load(cpu, NOT_IDLE); |
936 | } |
937 | |
938 | /* |
939 | * Return a high guess at the load of a migration-target cpu |
940 | */ |
941 | static inline unsigned long __target_load(int cpu, enum idle_type idle) |
942 | { |
943 | runqueue_t *rq = cpu_rq(cpu); |
944 | unsigned long running = rq->nr_running; |
945 | unsigned long target_load, cpu_load = rq->cpu_load, |
946 | load_now = running * SCHED_LOAD_SCALE; |
947 | |
948 | target_load = max(cpu_load, load_now); |
949 | |
950 | if (running > 1 || (idle == NOT_IDLE && running)) |
951 | target_load = target_load * rq->prio_bias / running; |
952 | |
953 | return target_load; |
954 | } |
955 | |
956 | static inline unsigned long target_load(int cpu) |
957 | { |
958 | return __target_load(cpu, NOT_IDLE); |
959 | } |
960 | #endif |
961 | |
962 | /* |
963 | * wake_idle() will wake a task on an idle cpu if task->cpu is |
964 | * not idle and an idle cpu is available. The span of cpus to |
965 | * search starts with cpus closest then further out as needed, |
966 | * so we always favor a closer, idle cpu. |
967 | * |
968 | * Returns the CPU we should wake onto. |
969 | */ |
970 | #if defined(ARCH_HAS_SCHED_WAKE_IDLE) |
971 | static inline int wake_idle(int cpu, task_t *p) |
972 | { |
973 | cpumask_t tmp; |
974 | struct sched_domain *sd; |
975 | int i; |
976 | |
977 | if (idle_cpu(cpu)) |
978 | return cpu; |
979 | |
980 | for_each_domain(cpu, sd) { |
981 | if (sd->flags & SD_WAKE_IDLE) { |
982 | cpus_and(tmp, sd->span, cpu_online_map); |
983 | cpus_and(tmp, tmp, p->cpus_allowed); |
984 | for_each_cpu_mask(i, tmp) { |
985 | if (idle_cpu(i)) |
986 | return i; |
987 | } |
988 | } |
989 | else break; |
990 | } |
991 | return cpu; |
992 | } |
993 | #else |
994 | static inline int wake_idle(int cpu, task_t *p) |
995 | { |
996 | return cpu; |
997 | } |
998 | #endif |
999 | |
1000 | /* |
1001 | * cache_delay is the time preemption is delayed in sched_compute mode |
1002 | * and is set to a nominal 10ms. |
1003 | */ |
1004 | static int cache_delay = 10 * HZ / 1000; |
1005 | |
1006 | /* |
1007 | * Check to see if p preempts rq->curr and resched if it does. In compute |
1008 | * mode we do not preempt for at least cache_delay and set rq->preempted. |
1009 | */ |
1010 | static void preempt(task_t *p, runqueue_t *rq) |
1011 | { |
1012 | if (p->prio >= rq->curr->prio) |
1013 | return; |
1014 | if (!sched_compute || rq->cache_ticks >= cache_delay || |
1015 | !p->mm || rt_task(p)) |
1016 | resched_task(rq->curr); |
1017 | rq->preempted = 1; |
1018 | } |
1019 | |
1020 | /*** |
1021 | * try_to_wake_up - wake up a thread |
1022 | * @p: the to-be-woken-up thread |
1023 | * @state: the mask of task states that can be woken |
1024 | * @sync: do a synchronous wakeup? |
1025 | * |
1026 | * Put it on the run-queue if it's not already there. The "current" |
1027 | * thread is always on the run-queue (except when the actual |
1028 | * re-schedule is in progress), and as such you're allowed to do |
1029 | * the simpler "current->state = TASK_RUNNING" to mark yourself |
1030 | * runnable without the overhead of this. |
1031 | * |
1032 | * returns failure only if the task is already active. |
1033 | */ |
1034 | static int try_to_wake_up(task_t * p, unsigned int state, int sync) |
1035 | { |
1036 | int cpu, this_cpu, success = 0; |
1037 | unsigned long flags; |
1038 | long old_state; |
1039 | runqueue_t *rq; |
1040 | #ifdef CONFIG_SMP |
1041 | unsigned long load, this_load; |
1042 | struct sched_domain *sd; |
1043 | int new_cpu; |
1044 | #endif |
1045 | |
1046 | rq = task_rq_lock(p, &flags); |
1047 | old_state = p->state; |
1048 | if (!(old_state & state)) |
1049 | goto out; |
1050 | |
1051 | if (task_queued(p)) |
1052 | goto out_running; |
1053 | |
1054 | cpu = task_cpu(p); |
1055 | this_cpu = smp_processor_id(); |
1056 | |
1057 | #ifdef CONFIG_SMP |
1058 | if (unlikely(task_running(rq, p))) |
1059 | goto out_activate; |
1060 | |
1061 | #ifdef CONFIG_SCHEDSTATS |
1062 | schedstat_inc(rq, ttwu_cnt); |
1063 | if (cpu == this_cpu) { |
1064 | schedstat_inc(rq, ttwu_local); |
1065 | } else { |
1066 | for_each_domain(this_cpu, sd) { |
1067 | if (cpu_isset(cpu, sd->span)) { |
1068 | schedstat_inc(sd, ttwu_wake_remote); |
1069 | break; |
1070 | } |
1071 | } |
1072 | } |
1073 | #endif |
1074 | |
1075 | new_cpu = cpu; |
1076 | if (cpu == this_cpu || unlikely(!cpu_isset(this_cpu, p->cpus_allowed))) |
1077 | goto out_set_cpu; |
1078 | |
1079 | load = source_load(cpu); |
1080 | this_load = target_load(this_cpu); |
1081 | |
1082 | /* |
1083 | * If sync wakeup then subtract the (maximum possible) effect of |
1084 | * the currently running task from the load of the current CPU: |
1085 | */ |
1086 | if (sync) |
1087 | this_load -= SCHED_LOAD_SCALE; |
1088 | |
1089 | /* Don't pull the task off an idle CPU to a busy one */ |
1090 | if (load < SCHED_LOAD_SCALE/2 && this_load > SCHED_LOAD_SCALE/2) |
1091 | goto out_set_cpu; |
1092 | |
1093 | new_cpu = this_cpu; /* Wake to this CPU if we can */ |
1094 | |
1095 | /* |
1096 | * Scan domains for affine wakeup and passive balancing |
1097 | * possibilities. |
1098 | */ |
1099 | for_each_domain(this_cpu, sd) { |
1100 | unsigned int imbalance; |
1101 | /* |
1102 | * Start passive balancing when half the imbalance_pct |
1103 | * limit is reached. |
1104 | */ |
1105 | imbalance = sd->imbalance_pct + (sd->imbalance_pct - 100) / 2; |
1106 | |
1107 | if ((sd->flags & SD_WAKE_AFFINE) && |
1108 | !task_hot(p, rq->timestamp_last_tick, sd)) { |
1109 | /* |
1110 | * This domain has SD_WAKE_AFFINE and p is cache cold |
1111 | * in this domain. |
1112 | */ |
1113 | if (cpu_isset(cpu, sd->span)) { |
1114 | schedstat_inc(sd, ttwu_move_affine); |
1115 | goto out_set_cpu; |
1116 | } |
1117 | } else if ((sd->flags & SD_WAKE_BALANCE) && |
1118 | imbalance*this_load <= 100*load) { |
1119 | /* |
1120 | * This domain has SD_WAKE_BALANCE and there is |
1121 | * an imbalance. |
1122 | */ |
1123 | if (cpu_isset(cpu, sd->span)) { |
1124 | schedstat_inc(sd, ttwu_move_balance); |
1125 | goto out_set_cpu; |
1126 | } |
1127 | } |
1128 | } |
1129 | |
1130 | new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */ |
1131 | out_set_cpu: |
1132 | new_cpu = wake_idle(new_cpu, p); |
1133 | if (new_cpu != cpu) { |
1134 | set_task_cpu(p, new_cpu); |
1135 | task_rq_unlock(rq, &flags); |
1136 | /* might preempt at this point */ |
1137 | rq = task_rq_lock(p, &flags); |
1138 | old_state = p->state; |
1139 | if (!(old_state & state)) |
1140 | goto out; |
1141 | if (task_queued(p)) |
1142 | goto out_running; |
1143 | |
1144 | this_cpu = smp_processor_id(); |
1145 | cpu = task_cpu(p); |
1146 | } |
1147 | |
1148 | out_activate: |
1149 | #endif /* CONFIG_SMP */ |
1150 | if (old_state == TASK_UNINTERRUPTIBLE) |
1151 | rq->nr_uninterruptible--; |
1152 | |
1153 | /* |
1154 | * Tasks that have marked their sleep as noninteractive get |
1155 | * woken up without their sleep counting. |
1156 | */ |
1157 | if (old_state & TASK_NONINTERACTIVE) |
1158 | p->flags |= PF_NONSLEEP; |
1159 | |
1160 | /* |
1161 | * Sync wakeups (i.e. those types of wakeups where the waker |
1162 | * has indicated that it will leave the CPU in short order) |
1163 | * don't trigger a preemption, if the woken up task will run on |
1164 | * this cpu. (in this case the 'I will reschedule' promise of |
1165 | * the waker guarantees that the freshly woken up task is going |
1166 | * to be considered on this CPU.) |
1167 | */ |
1168 | activate_task(p, rq, cpu == this_cpu); |
1169 | if (!sync || cpu != this_cpu) { |
1170 | preempt(p, rq); |
1171 | } |
1172 | success = 1; |
1173 | |
1174 | out_running: |
1175 | p->state = TASK_RUNNING; |
1176 | out: |
1177 | task_rq_unlock(rq, &flags); |
1178 | |
1179 | return success; |
1180 | } |
1181 | |
1182 | int fastcall wake_up_process(task_t * p) |
1183 | { |
1184 | return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED | |
1185 | TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0); |
1186 | } |
1187 | |
1188 | EXPORT_SYMBOL(wake_up_process); |
1189 | |
1190 | int fastcall wake_up_state(task_t *p, unsigned int state) |
1191 | { |
1192 | return try_to_wake_up(p, state, 0); |
1193 | } |
1194 | |
1195 | #ifdef CONFIG_SMP |
1196 | static int find_idlest_cpu(struct task_struct *p, int this_cpu, |
1197 | struct sched_domain *sd); |
1198 | #endif |
1199 | |
1200 | /* |
1201 | * Perform scheduler related setup for a newly forked process p. |
1202 | * p is forked by current. |
1203 | */ |
1204 | void fastcall sched_fork(task_t *p) |
1205 | { |
1206 | /* |
1207 | * We mark the process as running here, but have not actually |
1208 | * inserted it onto the runqueue yet. This guarantees that |
1209 | * nobody will actually run it, and a signal or other external |
1210 | * event cannot wake it up and insert it on the runqueue either. |
1211 | */ |
1212 | p->state = TASK_RUNNING; |
1213 | INIT_LIST_HEAD(&p->run_list); |
1214 | spin_lock_init(&p->switch_lock); |
1215 | #ifdef CONFIG_SCHEDSTATS |
1216 | memset(&p->sched_info, 0, sizeof(p->sched_info)); |
1217 | #endif |
1218 | #ifdef CONFIG_PREEMPT |
1219 | /* |
1220 | * During context-switch we hold precisely one spinlock, which |
1221 | * schedule_tail drops. (in the common case it's this_rq()->lock, |
1222 | * but it also can be p->switch_lock.) So we compensate with a count |
1223 | * of 1. Also, we want to start with kernel preemption disabled. |
1224 | */ |
1225 | p->thread_info->preempt_count = 1; |
1226 | #endif |
1227 | } |
1228 | |
1229 | /* |
1230 | * wake_up_new_task - wake up a newly created task for the first time. |
1231 | * |
1232 | * This function will do some initial scheduler statistics housekeeping |
1233 | * that must be done for every newly created context, then puts the task |
1234 | * on the runqueue and wakes it. |
1235 | */ |
1236 | void fastcall wake_up_new_task(task_t * p, unsigned long clone_flags) |
1237 | { |
1238 | unsigned long flags; |
1239 | int this_cpu, cpu; |
1240 | runqueue_t *rq, *this_rq; |
1241 | |
1242 | rq = task_rq_lock(p, &flags); |
1243 | cpu = task_cpu(p); |
1244 | this_cpu = smp_processor_id(); |
1245 | |
1246 | BUG_ON(p->state != TASK_RUNNING); |
1247 | |
1248 | /* |
1249 | * Forked process gets no burst to prevent fork bombs. |
1250 | */ |
1251 | p->burst = 0; |
1252 | |
1253 | if (likely(cpu == this_cpu)) { |
1254 | current->flags |= PF_NONSLEEP; |
1255 | activate_task(p, rq, 1); |
1256 | if (!(clone_flags & CLONE_VM)) |
1257 | /* |
1258 | * The VM isn't cloned, so we're in a good position to |
1259 | * do child-runs-first in anticipation of an exec. This |
1260 | * usually avoids a lot of COW overhead. |
1261 | */ |
1262 | set_need_resched(); |
1263 | /* |
1264 | * We skip the following code due to cpu == this_cpu |
1265 | * |
1266 | * task_rq_unlock(rq, &flags); |
1267 | * this_rq = task_rq_lock(current, &flags); |
1268 | */ |
1269 | this_rq = rq; |
1270 | } else { |
1271 | this_rq = cpu_rq(this_cpu); |
1272 | |
1273 | /* |
1274 | * Not the local CPU - must adjust timestamp. This should |
1275 | * get optimised away in the !CONFIG_SMP case. |
1276 | */ |
1277 | p->timestamp = (p->timestamp - this_rq->timestamp_last_tick) |
1278 | + rq->timestamp_last_tick; |
1279 | activate_task(p, rq, 0); |
1280 | preempt(p, rq); |
1281 | |
1282 | /* |
1283 | * Parent and child are on different CPUs, now get the |
1284 | * parent runqueue to update the parent's ->flags: |
1285 | */ |
1286 | task_rq_unlock(rq, &flags); |
1287 | this_rq = task_rq_lock(current, &flags); |
1288 | current->flags |= PF_NONSLEEP; |
1289 | } |
1290 | task_rq_unlock(this_rq, &flags); |
1291 | } |
1292 | |
1293 | /** |
1294 | * finish_task_switch - clean up after a task-switch |
1295 | * @prev: the thread we just switched away from. |
1296 | * |
1297 | * We enter this with the runqueue still locked, and finish_arch_switch() |
1298 | * will unlock it along with doing any other architecture-specific cleanup |
1299 | * actions. |
1300 | * |
1301 | * Note that we may have delayed dropping an mm in context_switch(). If |
1302 | * so, we finish that here outside of the runqueue lock. (Doing it |
1303 | * with the lock held can cause deadlocks; see schedule() for |
1304 | * details.) |
1305 | */ |
1306 | static inline void finish_task_switch(task_t *prev) |
1307 | __releases(rq->lock) |
1308 | { |
1309 | runqueue_t *rq = this_rq(); |
1310 | struct mm_struct *mm = rq->prev_mm; |
1311 | unsigned long prev_task_flags; |
1312 | |
1313 | rq->prev_mm = NULL; |
1314 | |
1315 | /* |
1316 | * A task struct has one reference for the use as "current". |
1317 | * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and |
1318 | * calls schedule one last time. The schedule call will never return, |
1319 | * and the scheduled task must drop that reference. |
1320 | * The test for EXIT_ZOMBIE must occur while the runqueue locks are |
1321 | * still held, otherwise prev could be scheduled on another cpu, die |
1322 | * there before we look at prev->state, and then the reference would |
1323 | * be dropped twice. |
1324 | * Manfred Spraul <manfred@colorfullife.com> |
1325 | */ |
1326 | prev_task_flags = prev->flags; |
1327 | finish_arch_switch(rq, prev); |
1328 | if (mm) |
1329 | mmdrop(mm); |
1330 | if (unlikely(prev_task_flags & PF_DEAD)) |
1331 | put_task_struct(prev); |
1332 | } |
1333 | |
1334 | /** |
1335 | * schedule_tail - first thing a freshly forked thread must call. |
1336 | * @prev: the thread we just switched away from. |
1337 | */ |
1338 | asmlinkage void schedule_tail(task_t *prev) |
1339 | __releases(rq->lock) |
1340 | { |
1341 | finish_task_switch(prev); |
1342 | |
1343 | if (current->set_child_tid) |
1344 | put_user(current->pid, current->set_child_tid); |
1345 | } |
1346 | |
1347 | /* |
1348 | * context_switch - switch to the new MM and the new |
1349 | * thread's register state. |
1350 | */ |
1351 | static inline |
1352 | task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next) |
1353 | { |
1354 | struct mm_struct *mm = next->mm; |
1355 | struct mm_struct *oldmm = prev->active_mm; |
1356 | |
1357 | if (unlikely(!mm)) { |
1358 | next->active_mm = oldmm; |
1359 | atomic_inc(&oldmm->mm_count); |
1360 | enter_lazy_tlb(oldmm, next); |
1361 | } else |
1362 | switch_mm(oldmm, mm, next); |
1363 | |
1364 | if (unlikely(!prev->mm)) { |
1365 | prev->active_mm = NULL; |
1366 | WARN_ON(rq->prev_mm); |
1367 | rq->prev_mm = oldmm; |
1368 | } |
1369 | |
1370 | /* Here we just switch the register state and the stack. */ |
1371 | switch_to(prev, next, prev); |
1372 | |
1373 | return prev; |
1374 | } |
1375 | |
1376 | /* |
1377 | * nr_running, nr_uninterruptible and nr_context_switches: |
1378 | * |
1379 | * externally visible scheduler statistics: current number of runnable |
1380 | * threads, current number of uninterruptible-sleeping threads, total |
1381 | * number of context switches performed since bootup. |
1382 | */ |
1383 | unsigned long nr_running(void) |
1384 | { |
1385 | unsigned long i, sum = 0; |
1386 | |
1387 | for_each_online_cpu(i) |
1388 | sum += cpu_rq(i)->nr_running; |
1389 | |
1390 | return sum; |
1391 | } |
1392 | |
1393 | unsigned long nr_uninterruptible(void) |
1394 | { |
1395 | unsigned long i, sum = 0; |
1396 | |
1397 | for_each_cpu(i) |
1398 | sum += cpu_rq(i)->nr_uninterruptible; |
1399 | |
1400 | /* |
1401 | * Since we read the counters lockless, it might be slightly |
1402 | * inaccurate. Do not allow it to go below zero though: |
1403 | */ |
1404 | if (unlikely((long)sum < 0)) |
1405 | sum = 0; |
1406 | |
1407 | return sum; |
1408 | } |
1409 | |
1410 | unsigned long long nr_context_switches(void) |
1411 | { |
1412 | unsigned long long i, sum = 0; |
1413 | |
1414 | for_each_cpu(i) |
1415 | sum += cpu_rq(i)->nr_switches; |
1416 | |
1417 | return sum; |
1418 | } |
1419 | |
1420 | unsigned long nr_iowait(void) |
1421 | { |
1422 | unsigned long i, sum = 0; |
1423 | |
1424 | for_each_cpu(i) |
1425 | sum += atomic_read(&cpu_rq(i)->nr_iowait); |
1426 | |
1427 | return sum; |
1428 | } |
1429 | |
1430 | #ifdef CONFIG_SMP |
1431 | |
1432 | /* |
1433 | * double_rq_lock - safely lock two runqueues |
1434 | * |
1435 | * Note this does not disable interrupts like task_rq_lock, |
1436 | * you need to do so manually before calling. |
1437 | */ |
1438 | static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2) |
1439 | __acquires(rq1->lock) |
1440 | __acquires(rq2->lock) |
1441 | { |
1442 | if (rq1 == rq2) { |
1443 | spin_lock(&rq1->lock); |
1444 | __acquire(rq2->lock); /* Fake it out ;) */ |
1445 | } else { |
1446 | if (rq1 < rq2) { |
1447 | spin_lock(&rq1->lock); |
1448 | spin_lock(&rq2->lock); |
1449 | } else { |
1450 | spin_lock(&rq2->lock); |
1451 | spin_lock(&rq1->lock); |
1452 | } |
1453 | } |
1454 | } |
1455 | |
1456 | /* |
1457 | * double_rq_unlock - safely unlock two runqueues |
1458 | * |
1459 | * Note this does not restore interrupts like task_rq_unlock, |
1460 | * you need to do so manually after calling. |
1461 | */ |
1462 | static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2) |
1463 | __releases(rq1->lock) |
1464 | __releases(rq2->lock) |
1465 | { |
1466 | spin_unlock(&rq1->lock); |
1467 | if (rq1 != rq2) |
1468 | spin_unlock(&rq2->lock); |
1469 | else |
1470 | __release(rq2->lock); |
1471 | } |
1472 | |
1473 | /* |
1474 | * double_lock_balance - lock the busiest runqueue, this_rq is locked already. |
1475 | */ |
1476 | static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest) |
1477 | __releases(this_rq->lock) |
1478 | __acquires(busiest->lock) |
1479 | __acquires(this_rq->lock) |
1480 | { |
1481 | if (unlikely(!spin_trylock(&busiest->lock))) { |
1482 | if (busiest < this_rq) { |
1483 | spin_unlock(&this_rq->lock); |
1484 | spin_lock(&busiest->lock); |
1485 | spin_lock(&this_rq->lock); |
1486 | } else |
1487 | spin_lock(&busiest->lock); |
1488 | } |
1489 | } |
1490 | |
1491 | /* |
1492 | * find_idlest_cpu - find the least busy runqueue. |
1493 | */ |
1494 | static int find_idlest_cpu(struct task_struct *p, int this_cpu, |
1495 | struct sched_domain *sd) |
1496 | { |
1497 | unsigned long load, min_load, this_load; |
1498 | int i, min_cpu; |
1499 | cpumask_t mask; |
1500 | |
1501 | min_cpu = UINT_MAX; |
1502 | min_load = ULONG_MAX; |
1503 | |
1504 | cpus_and(mask, sd->span, p->cpus_allowed); |
1505 | |
1506 | for_each_cpu_mask(i, mask) { |
1507 | load = target_load(i); |
1508 | |
1509 | if (load < min_load) { |
1510 | min_cpu = i; |
1511 | min_load = load; |
1512 | |
1513 | /* break out early on an idle CPU: */ |
1514 | if (!min_load) |
1515 | break; |
1516 | } |
1517 | } |
1518 | |
1519 | /* add +1 to account for the new task */ |
1520 | this_load = source_load(this_cpu) + SCHED_LOAD_SCALE; |
1521 | |
1522 | /* |
1523 | * Would with the addition of the new task to the |
1524 | * current CPU there be an imbalance between this |
1525 | * CPU and the idlest CPU? |
1526 | * |
1527 | * Use half of the balancing threshold - new-context is |
1528 | * a good opportunity to balance. |
1529 | */ |
1530 | if (min_load*(100 + (sd->imbalance_pct-100)/2) < this_load*100) |
1531 | return min_cpu; |
1532 | |
1533 | return this_cpu; |
1534 | } |
1535 | |
1536 | /* |
1537 | * If dest_cpu is allowed for this process, migrate the task to it. |
1538 | * This is accomplished by forcing the cpu_allowed mask to only |
1539 | * allow dest_cpu, which will force the cpu onto dest_cpu. Then |
1540 | * the cpu_allowed mask is restored. |
1541 | */ |
1542 | static inline void sched_migrate_task(task_t *p, int dest_cpu) |
1543 | { |
1544 | migration_req_t req; |
1545 | runqueue_t *rq; |
1546 | unsigned long flags; |
1547 | |
1548 | rq = task_rq_lock(p, &flags); |
1549 | if (!cpu_isset(dest_cpu, p->cpus_allowed) |
1550 | || unlikely(cpu_is_offline(dest_cpu))) |
1551 | goto out; |
1552 | |
1553 | /* force the process onto the specified CPU */ |
1554 | if (migrate_task(p, dest_cpu, &req)) { |
1555 | /* Need to wait for migration thread (might exit: take ref). */ |
1556 | struct task_struct *mt = rq->migration_thread; |
1557 | get_task_struct(mt); |
1558 | task_rq_unlock(rq, &flags); |
1559 | wake_up_process(mt); |
1560 | put_task_struct(mt); |
1561 | wait_for_completion(&req.done); |
1562 | return; |
1563 | } |
1564 | out: |
1565 | task_rq_unlock(rq, &flags); |
1566 | } |
1567 | |
1568 | /* |
1569 | * sched_exec(): find the highest-level, exec-balance-capable |
1570 | * domain and try to migrate the task to the least loaded CPU. |
1571 | * |
1572 | * execve() is a valuable balancing opportunity, because at this point |
1573 | * the task has the smallest effective memory and cache footprint. |
1574 | */ |
1575 | void sched_exec(void) |
1576 | { |
1577 | struct sched_domain *tmp, *sd = NULL; |
1578 | int new_cpu, this_cpu = get_cpu(); |
1579 | |
1580 | /* Prefer the current CPU if there's only this task running */ |
1581 | if (this_rq()->nr_running <= 1) |
1582 | goto out; |
1583 | |
1584 | for_each_domain(this_cpu, tmp) |
1585 | if (tmp->flags & SD_BALANCE_EXEC) |
1586 | sd = tmp; |
1587 | |
1588 | if (sd) { |
1589 | schedstat_inc(sd, sbe_attempts); |
1590 | new_cpu = find_idlest_cpu(current, this_cpu, sd); |
1591 | if (new_cpu != this_cpu) { |
1592 | schedstat_inc(sd, sbe_pushed); |
1593 | put_cpu(); |
1594 | sched_migrate_task(current, new_cpu); |
1595 | return; |
1596 | } |
1597 | } |
1598 | out: |
1599 | put_cpu(); |
1600 | } |
1601 | |
1602 | /* |
1603 | * pull_task - move a task from a remote runqueue to the local runqueue. |
1604 | * Both runqueues must be locked. |
1605 | */ |
1606 | static inline void pull_task(runqueue_t *src_rq, task_t *p, |
1607 | runqueue_t *this_rq, int this_cpu) |
1608 | { |
1609 | dequeue_task(p, src_rq); |
1610 | dec_nr_running(p, src_rq); |
1611 | set_task_cpu(p, this_cpu); |
1612 | inc_nr_running(p, this_rq); |
1613 | enqueue_task(p, this_rq); |
1614 | p->timestamp = (p->timestamp - src_rq->timestamp_last_tick) |
1615 | + this_rq->timestamp_last_tick; |
1616 | /* |
1617 | * Note that idle threads have a prio of MAX_PRIO, for this test |
1618 | * to be always true for them. |
1619 | */ |
1620 | preempt(p, this_rq); |
1621 | } |
1622 | |
1623 | /* |
1624 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? |
1625 | */ |
1626 | static inline int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu, |
1627 | struct sched_domain *sd, enum idle_type idle) |
1628 | { |
1629 | /* |
1630 | * We do not migrate tasks that are: |
1631 | * 1) running (obviously), or |
1632 | * 2) cannot be migrated to this CPU due to cpus_allowed, or |
1633 | * 3) are cache-hot on their current CPU. |
1634 | */ |
1635 | if (task_running(rq, p)) |
1636 | return 0; |
1637 | if (!cpu_isset(this_cpu, p->cpus_allowed)) |
1638 | return 0; |
1639 | |
1640 | /* |
1641 | * Aggressive migration if: |
1642 | * 1) the [whole] cpu is idle, or |
1643 | * 2) too many balance attempts have failed. |
1644 | */ |
1645 | |
1646 | if (cpu_and_siblings_are_idle(this_cpu) || \ |
1647 | sd->nr_balance_failed > sd->cache_nice_tries) |
1648 | return 1; |
1649 | |
1650 | if (task_hot(p, rq->timestamp_last_tick, sd)) |
1651 | return 0; |
1652 | return 1; |
1653 | } |
1654 | |
1655 | /* |
1656 | * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq, |
1657 | * as part of a balancing operation within "domain". Returns the number of |
1658 | * tasks moved. |
1659 | * |
1660 | * Called with both runqueues locked. |
1661 | */ |
1662 | static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest, |
1663 | unsigned long max_nr_move, struct sched_domain *sd, |
1664 | enum idle_type idle) |
1665 | { |
1666 | struct list_head *head, *curr; |
1667 | int idx, pulled = 0; |
1668 | task_t *tmp; |
1669 | |
1670 | if (max_nr_move <= 0 || busiest->nr_running <= 1) |
1671 | goto out; |
1672 | |
1673 | /* Start searching at priority 0: */ |
1674 | idx = 0; |
1675 | skip_bitmap: |
1676 | if (!idx) |
1677 | idx = sched_find_first_bit(busiest->bitmap); |
1678 | else |
1679 | idx = find_next_bit(busiest->bitmap, MAX_PRIO, idx); |
1680 | if (idx >= MAX_PRIO) |
1681 | goto out; |
1682 | |
1683 | head = busiest->queue + idx; |
1684 | curr = head->prev; |
1685 | skip_queue: |
1686 | tmp = list_entry(curr, task_t, run_list); |
1687 | |
1688 | curr = curr->prev; |
1689 | |
1690 | if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle)) { |
1691 | if (curr != head) |
1692 | goto skip_queue; |
1693 | idx++; |
1694 | goto skip_bitmap; |
1695 | } |
1696 | |
1697 | #ifdef CONFIG_SCHEDSTATS |
1698 | if (task_hot(tmp, busiest->timestamp_last_tick, sd)) |
1699 | schedstat_inc(sd, lb_hot_gained[idle]); |
1700 | #endif |
1701 | |
1702 | pull_task(busiest, tmp, this_rq, this_cpu); |
1703 | pulled++; |
1704 | |
1705 | /* We only want to steal up to the prescribed number of tasks. */ |
1706 | if (pulled < max_nr_move) { |
1707 | if (curr != head) |
1708 | goto skip_queue; |
1709 | idx++; |
1710 | goto skip_bitmap; |
1711 | } |
1712 | out: |
1713 | /* |
1714 | * Right now, this is the only place pull_task() is called, |
1715 | * so we can safely collect pull_task() stats here rather than |
1716 | * inside pull_task(). |
1717 | */ |
1718 | schedstat_add(sd, lb_gained[idle], pulled); |
1719 | return pulled; |
1720 | } |
1721 | |
1722 | /* |
1723 | * find_busiest_group finds and returns the busiest CPU group within the |
1724 | * domain. It calculates and returns the number of tasks which should be |
1725 | * moved to restore balance via the imbalance parameter. |
1726 | */ |
1727 | static inline struct sched_group * |
1728 | find_busiest_group(struct sched_domain *sd, int this_cpu, |
1729 | unsigned long *imbalance, enum idle_type idle) |
1730 | { |
1731 | struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups; |
1732 | unsigned long max_load, avg_load, total_load, this_load, total_pwr; |
1733 | |
1734 | max_load = this_load = total_load = total_pwr = 0; |
1735 | |
1736 | do { |
1737 | unsigned long load; |
1738 | int local_group; |
1739 | int i; |
1740 | |
1741 | local_group = cpu_isset(this_cpu, group->cpumask); |
1742 | |
1743 | /* Tally up the load of all CPUs in the group */ |
1744 | avg_load = 0; |
1745 | |
1746 | for_each_cpu_mask(i, group->cpumask) { |
1747 | /* Bias balancing toward cpus of our domain */ |
1748 | if (local_group) |
1749 | load = __target_load(i, idle); |
1750 | else |
1751 | load = __source_load(i, idle); |
1752 | |
1753 | avg_load += load; |
1754 | } |
1755 | |
1756 | total_load += avg_load; |
1757 | total_pwr += group->cpu_power; |
1758 | |
1759 | /* Adjust by relative CPU power of the group */ |
1760 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; |
1761 | |
1762 | if (local_group) { |
1763 | this_load = avg_load; |
1764 | this = group; |
1765 | goto nextgroup; |
1766 | } else if (avg_load > max_load) { |
1767 | max_load = avg_load; |
1768 | busiest = group; |
1769 | } |
1770 | nextgroup: |
1771 | group = group->next; |
1772 | } while (group != sd->groups); |
1773 | |
1774 | if (!busiest || this_load >= max_load) |
1775 | goto out_balanced; |
1776 | |
1777 | avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr; |
1778 | |
1779 | if (this_load >= avg_load || |
1780 | 100*max_load <= sd->imbalance_pct*this_load) |
1781 | goto out_balanced; |
1782 | |
1783 | /* |
1784 | * We're trying to get all the cpus to the average_load, so we don't |
1785 | * want to push ourselves above the average load, nor do we wish to |
1786 | * reduce the max loaded cpu below the average load, as either of these |
1787 | * actions would just result in more rebalancing later, and ping-pong |
1788 | * tasks around. Thus we look for the minimum possible imbalance. |
1789 | * Negative imbalances (*we* are more loaded than anyone else) will |
1790 | * be counted as no imbalance for these purposes -- we can't fix that |
1791 | * by pulling tasks to us. Be careful of negative numbers as they'll |
1792 | * appear as very large values with unsigned longs. |
1793 | */ |
1794 | /* How much load to actually move to equalise the imbalance */ |
1795 | *imbalance = min((max_load - avg_load) * busiest->cpu_power, |
1796 | (avg_load - this_load) * this->cpu_power) |
1797 | / SCHED_LOAD_SCALE; |
1798 | |
1799 | if (*imbalance < SCHED_LOAD_SCALE) { |
1800 | unsigned long pwr_now = 0, pwr_move = 0; |
1801 | unsigned long tmp; |
1802 | |
1803 | if (max_load - this_load >= SCHED_LOAD_SCALE*2) { |
1804 | *imbalance = 1; |
1805 | return busiest; |
1806 | } |
1807 | |
1808 | /* |
1809 | * OK, we don't have enough imbalance to justify moving tasks, |
1810 | * however we may be able to increase total CPU power used by |
1811 | * moving them. |
1812 | */ |
1813 | |
1814 | pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load); |
1815 | pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load); |
1816 | pwr_now /= SCHED_LOAD_SCALE; |
1817 | |
1818 | /* Amount of load we'd subtract */ |
1819 | tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power; |
1820 | if (max_load > tmp) |
1821 | pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE, |
1822 | max_load - tmp); |
1823 | |
1824 | /* Amount of load we'd add */ |
1825 | if (max_load*busiest->cpu_power < |
1826 | SCHED_LOAD_SCALE*SCHED_LOAD_SCALE) |
1827 | tmp = max_load*busiest->cpu_power/this->cpu_power; |
1828 | else |
1829 | tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power; |
1830 | pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp); |
1831 | pwr_move /= SCHED_LOAD_SCALE; |
1832 | |
1833 | /* Move if we gain throughput */ |
1834 | if (pwr_move <= pwr_now) |
1835 | goto out_balanced; |
1836 | |
1837 | *imbalance = 1; |
1838 | return busiest; |
1839 | } |
1840 | |
1841 | /* Get rid of the scaling factor, rounding down as we divide */ |
1842 | *imbalance = *imbalance / SCHED_LOAD_SCALE; |
1843 | |
1844 | return busiest; |
1845 | |
1846 | out_balanced: |
1847 | if (busiest && (idle == NEWLY_IDLE || |
1848 | (idle == SCHED_IDLE && max_load > SCHED_LOAD_SCALE)) ) { |
1849 | *imbalance = 1; |
1850 | return busiest; |
1851 | } |
1852 | |
1853 | *imbalance = 0; |
1854 | return NULL; |
1855 | } |
1856 | |
1857 | /* |
1858 | * find_busiest_queue - find the busiest runqueue among the cpus in group. |
1859 | */ |
1860 | static runqueue_t *find_busiest_queue(struct sched_group *group, |
1861 | enum idle_type idle) |
1862 | { |
1863 | unsigned long load, max_load = 0; |
1864 | runqueue_t *busiest = NULL; |
1865 | int i; |
1866 | |
1867 | for_each_cpu_mask(i, group->cpumask) { |
1868 | load = __source_load(i, idle); |
1869 | |
1870 | if (load > max_load) { |
1871 | max_load = load; |
1872 | busiest = cpu_rq(i); |
1873 | } |
1874 | } |
1875 | |
1876 | return busiest; |
1877 | } |
1878 | |
1879 | /* |
1880 | * Check this_cpu to ensure it is balanced within domain. Attempt to move |
1881 | * tasks if there is an imbalance. |
1882 | * |
1883 | * Called with this_rq unlocked. |
1884 | */ |
1885 | static inline int load_balance(int this_cpu, runqueue_t *this_rq, |
1886 | struct sched_domain *sd, enum idle_type idle) |
1887 | { |
1888 | struct sched_group *group; |
1889 | runqueue_t *busiest; |
1890 | unsigned long imbalance; |
1891 | int nr_moved; |
1892 | |
1893 | spin_lock(&this_rq->lock); |
1894 | schedstat_inc(sd, lb_cnt[idle]); |
1895 | |
1896 | group = find_busiest_group(sd, this_cpu, &imbalance, idle); |
1897 | if (!group) { |
1898 | schedstat_inc(sd, lb_nobusyg[idle]); |
1899 | goto out_balanced; |
1900 | } |
1901 | |
1902 | busiest = find_busiest_queue(group, idle); |
1903 | if (!busiest) { |
1904 | schedstat_inc(sd, lb_nobusyq[idle]); |
1905 | goto out_balanced; |
1906 | } |
1907 | |
1908 | /* |
1909 | * This should be "impossible", but since load |
1910 | * balancing is inherently racy and statistical, |
1911 | * it could happen in theory. |
1912 | */ |
1913 | if (unlikely(busiest == this_rq)) { |
1914 | WARN_ON(1); |
1915 | goto out_balanced; |
1916 | } |
1917 | |
1918 | schedstat_add(sd, lb_imbalance[idle], imbalance); |
1919 | |
1920 | nr_moved = 0; |
1921 | if (busiest->nr_running > 1) { |
1922 | /* |
1923 | * Attempt to move tasks. If find_busiest_group has found |
1924 | * an imbalance but busiest->nr_running <= 1, the group is |
1925 | * still unbalanced. nr_moved simply stays zero, so it is |
1926 | * correctly treated as an imbalance. |
1927 | */ |
1928 | double_lock_balance(this_rq, busiest); |
1929 | nr_moved = move_tasks(this_rq, this_cpu, busiest, |
1930 | imbalance, sd, idle); |
1931 | spin_unlock(&busiest->lock); |
1932 | } |
1933 | spin_unlock(&this_rq->lock); |
1934 | |
1935 | if (!nr_moved) { |
1936 | schedstat_inc(sd, lb_failed[idle]); |
1937 | sd->nr_balance_failed++; |
1938 | |
1939 | if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) { |
1940 | int wake = 0; |
1941 | |
1942 | spin_lock(&busiest->lock); |
1943 | if (!busiest->active_balance) { |
1944 | busiest->active_balance = 1; |
1945 | busiest->push_cpu = this_cpu; |
1946 | wake = 1; |
1947 | } |
1948 | spin_unlock(&busiest->lock); |
1949 | if (wake) |
1950 | wake_up_process(busiest->migration_thread); |
1951 | |
1952 | /* |
1953 | * We've kicked active balancing, reset the failure |
1954 | * counter. |
1955 | */ |
1956 | sd->nr_balance_failed = sd->cache_nice_tries; |
1957 | } |
1958 | |
1959 | /* |
1960 | * We were unbalanced, but unsuccessful in move_tasks(), |
1961 | * so bump the balance_interval to lessen the lock contention. |
1962 | */ |
1963 | if (sd->balance_interval < sd->max_interval) |
1964 | sd->balance_interval++; |
1965 | } else { |
1966 | sd->nr_balance_failed = 0; |
1967 | |
1968 | /* We were unbalanced, so reset the balancing interval */ |
1969 | sd->balance_interval = sd->min_interval; |
1970 | } |
1971 | |
1972 | return nr_moved; |
1973 | |
1974 | out_balanced: |
1975 | spin_unlock(&this_rq->lock); |
1976 | |
1977 | schedstat_inc(sd, lb_balanced[idle]); |
1978 | |
1979 | /* tune up the balancing interval */ |
1980 | if (sd->balance_interval < sd->max_interval) |
1981 | sd->balance_interval *= 2; |
1982 | |
1983 | return 0; |
1984 | } |
1985 | |
1986 | /* |
1987 | * Check this_cpu to ensure it is balanced within domain. Attempt to move |
1988 | * tasks if there is an imbalance. |
1989 | * |
1990 | * Called from schedule when this_rq is about to become idle (NEWLY_IDLE). |
1991 | * this_rq is locked. |
1992 | */ |
1993 | static inline int load_balance_newidle(int this_cpu, runqueue_t *this_rq, |
1994 | struct sched_domain *sd) |
1995 | { |
1996 | struct sched_group *group; |
1997 | runqueue_t *busiest = NULL; |
1998 | unsigned long imbalance; |
1999 | int nr_moved = 0; |
2000 | |
2001 | schedstat_inc(sd, lb_cnt[NEWLY_IDLE]); |
2002 | group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE); |
2003 | if (!group) { |
2004 | schedstat_inc(sd, lb_balanced[NEWLY_IDLE]); |
2005 | schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]); |
2006 | goto out; |
2007 | } |
2008 | |
2009 | busiest = find_busiest_queue(group, NEWLY_IDLE); |
2010 | if (!busiest || busiest == this_rq) { |
2011 | schedstat_inc(sd, lb_balanced[NEWLY_IDLE]); |
2012 | schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]); |
2013 | goto out; |
2014 | } |
2015 | |
2016 | /* Attempt to move tasks */ |
2017 | double_lock_balance(this_rq, busiest); |
2018 | |
2019 | schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance); |
2020 | nr_moved = move_tasks(this_rq, this_cpu, busiest, |
2021 | imbalance, sd, NEWLY_IDLE); |
2022 | if (!nr_moved) |
2023 | schedstat_inc(sd, lb_failed[NEWLY_IDLE]); |
2024 | |
2025 | spin_unlock(&busiest->lock); |
2026 | |
2027 | out: |
2028 | return nr_moved; |
2029 | } |
2030 | |
2031 | /* |
2032 | * idle_balance is called by schedule() if this_cpu is about to become |
2033 | * idle. Attempts to pull tasks from other CPUs. |
2034 | */ |
2035 | static inline void idle_balance(int this_cpu, runqueue_t *this_rq) |
2036 | { |
2037 | struct sched_domain *sd; |
2038 | |
2039 | for_each_domain(this_cpu, sd) { |
2040 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
2041 | if (load_balance_newidle(this_cpu, this_rq, sd)) { |
2042 | /* We've pulled tasks over so stop searching */ |
2043 | break; |
2044 | } |
2045 | } |
2046 | } |
2047 | } |
2048 | |
2049 | /* |
2050 | * active_load_balance is run by migration threads. It pushes running tasks |
2051 | * off the busiest CPU onto idle CPUs. It requires at least 1 task to be |
2052 | * running on each physical CPU where possible, and avoids physical / |
2053 | * logical imbalances. |
2054 | * |
2055 | * Called with busiest_rq locked. |
2056 | */ |
2057 | static inline void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu) |
2058 | { |
2059 | struct sched_domain *sd; |
2060 | struct sched_group *cpu_group; |
2061 | runqueue_t *target_rq; |
2062 | cpumask_t visited_cpus; |
2063 | int cpu; |
2064 | |
2065 | /* |
2066 | * Search for suitable CPUs to push tasks to in successively higher |
2067 | * domains with SD_LOAD_BALANCE set. |
2068 | */ |
2069 | visited_cpus = CPU_MASK_NONE; |
2070 | for_each_domain(busiest_cpu, sd) { |
2071 | if (!(sd->flags & SD_LOAD_BALANCE)) |
2072 | /* no more domains to search */ |
2073 | break; |
2074 | |
2075 | schedstat_inc(sd, alb_cnt); |
2076 | |
2077 | cpu_group = sd->groups; |
2078 | do { |
2079 | for_each_cpu_mask(cpu, cpu_group->cpumask) { |
2080 | if (busiest_rq->nr_running <= 1) |
2081 | /* no more tasks left to move */ |
2082 | return; |
2083 | if (cpu_isset(cpu, visited_cpus)) |
2084 | continue; |
2085 | cpu_set(cpu, visited_cpus); |
2086 | if (!cpu_and_siblings_are_idle(cpu) || cpu == busiest_cpu) |
2087 | continue; |
2088 | |
2089 | target_rq = cpu_rq(cpu); |
2090 | /* |
2091 | * This condition is "impossible", if it occurs |
2092 | * we need to fix it. Originally reported by |
2093 | * Bjorn Helgaas on a 128-cpu setup. |
2094 | */ |
2095 | BUG_ON(busiest_rq == target_rq); |
2096 | |
2097 | /* move a task from busiest_rq to target_rq */ |
2098 | double_lock_balance(busiest_rq, target_rq); |
2099 | if (move_tasks(target_rq, cpu, busiest_rq, |
2100 | 1, sd, SCHED_IDLE)) { |
2101 | schedstat_inc(sd, alb_pushed); |
2102 | } else { |
2103 | schedstat_inc(sd, alb_failed); |
2104 | } |
2105 | spin_unlock(&target_rq->lock); |
2106 | } |
2107 | cpu_group = cpu_group->next; |
2108 | } while (cpu_group != sd->groups); |
2109 | } |
2110 | } |
2111 | |
2112 | /* |
2113 | * rebalance_tick will get called every timer tick, on every CPU. |
2114 | * |
2115 | * It checks each scheduling domain to see if it is due to be balanced, |
2116 | * and initiates a balancing operation if so. |
2117 | * |
2118 | * Balancing parameters are set up in arch_init_sched_domains. |
2119 | */ |
2120 | |
2121 | /* Don't have all balancing operations going off at once */ |
2122 | #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS) |
2123 | |
2124 | static void rebalance_tick(int this_cpu, runqueue_t *this_rq, |
2125 | enum idle_type idle) |
2126 | { |
2127 | unsigned long old_load, this_load; |
2128 | unsigned long j = jiffies + CPU_OFFSET(this_cpu); |
2129 | struct sched_domain *sd; |
2130 | |
2131 | /* Update our load */ |
2132 | old_load = this_rq->cpu_load; |
2133 | this_load = this_rq->nr_running * SCHED_LOAD_SCALE; |
2134 | /* |
2135 | * Round up the averaging division if load is increasing. This |
2136 | * prevents us from getting stuck on 9 if the load is 10, for |
2137 | * example. |
2138 | */ |
2139 | if (this_load > old_load) |
2140 | old_load++; |
2141 | this_rq->cpu_load = (old_load + this_load) / 2; |
2142 | |
2143 | for_each_domain(this_cpu, sd) { |
2144 | unsigned long interval; |
2145 | |
2146 | if (!(sd->flags & SD_LOAD_BALANCE)) |
2147 | continue; |
2148 | |
2149 | interval = sd->balance_interval; |
2150 | |
2151 | /* scale ms to jiffies */ |
2152 | interval = msecs_to_jiffies(interval); |
2153 | if (unlikely(!interval)) |
2154 | interval = 1; |
2155 | |
2156 | if (idle != SCHED_IDLE || j - sd->last_balance >= interval) { |
2157 | if (load_balance(this_cpu, this_rq, sd, idle)) { |
2158 | /* We've pulled tasks over so no longer idle */ |
2159 | idle = NOT_IDLE; |
2160 | } |
2161 | sd->last_balance += interval; |
2162 | } |
2163 | } |
2164 | } |
2165 | #else |
2166 | /* |
2167 | * on UP we do not need to balance between CPUs: |
2168 | */ |
2169 | static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle) |
2170 | { |
2171 | } |
2172 | static inline void idle_balance(int cpu, runqueue_t *rq) |
2173 | { |
2174 | } |
2175 | #endif |
2176 | |
2177 | static inline int wake_priority_sleeper(runqueue_t *rq) |
2178 | { |
2179 | int ret = 0; |
2180 | #ifdef CONFIG_SCHED_SMT |
2181 | spin_lock(&rq->lock); |
2182 | /* |
2183 | * If an SMT sibling task has been put to sleep for priority |
2184 | * reasons reschedule the idle task to see if it can now run. |
2185 | */ |
2186 | if (rq->nr_running) { |
2187 | resched_task(rq->idle); |
2188 | ret = 1; |
2189 | } |
2190 | spin_unlock(&rq->lock); |
2191 | #endif |
2192 | return ret; |
2193 | } |
2194 | |
2195 | DEFINE_PER_CPU(struct kernel_stat, kstat); |
2196 | |
2197 | EXPORT_PER_CPU_SYMBOL(kstat); |
2198 | |
2199 | /* |
2200 | * This is called on clock ticks and on context switches. |
2201 | * Bank in p->sched_time the ns elapsed since the last tick or switch. |
2202 | */ |
2203 | static inline void update_cpu_clock(task_t *p, runqueue_t *rq, |
2204 | unsigned long long now) |
2205 | { |
2206 | unsigned long long last = max(p->timestamp, rq->timestamp_last_tick); |
2207 | p->sched_time += now - last; |
2208 | } |
2209 | |
2210 | /* |
2211 | * Return current->sched_time plus any more ns on the sched_clock |
2212 | * that have not yet been banked. |
2213 | */ |
2214 | unsigned long long current_sched_time(const task_t *tsk) |
2215 | { |
2216 | unsigned long long ns; |
2217 | unsigned long flags; |
2218 | local_irq_save(flags); |
2219 | ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick); |
2220 | ns = tsk->sched_time + (sched_clock() - ns); |
2221 | local_irq_restore(flags); |
2222 | return ns; |
2223 | } |
2224 | |
2225 | /* |
2226 | * Account user cpu time to a process. |
2227 | * @p: the process that the cpu time gets accounted to |
2228 | * @hardirq_offset: the offset to subtract from hardirq_count() |
2229 | * @cputime: the cpu time spent in user space since the last update |
2230 | */ |
2231 | void account_user_time(struct task_struct *p, cputime_t cputime) |
2232 | { |
2233 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
2234 | cputime64_t tmp; |
2235 | |
2236 | p->utime = cputime_add(p->utime, cputime); |
2237 | |
2238 | /* Add user time to cpustat. */ |
2239 | tmp = cputime_to_cputime64(cputime); |
2240 | if (TASK_NICE(p) > 0 || batch_task(p)) |
2241 | cpustat->nice = cputime64_add(cpustat->nice, tmp); |
2242 | else |
2243 | cpustat->user = cputime64_add(cpustat->user, tmp); |
2244 | } |
2245 | |
2246 | /* |
2247 | * Account system cpu time to a process. |
2248 | * @p: the process that the cpu time gets accounted to |
2249 | * @hardirq_offset: the offset to subtract from hardirq_count() |
2250 | * @cputime: the cpu time spent in kernel space since the last update |
2251 | */ |
2252 | void account_system_time(struct task_struct *p, int hardirq_offset, |
2253 | cputime_t cputime) |
2254 | { |
2255 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
2256 | runqueue_t *rq = this_rq(); |
2257 | cputime64_t tmp; |
2258 | |
2259 | p->stime = cputime_add(p->stime, cputime); |
2260 | |
2261 | /* Add system time to cpustat. */ |
2262 | tmp = cputime_to_cputime64(cputime); |
2263 | if (hardirq_count() - hardirq_offset) |
2264 | cpustat->irq = cputime64_add(cpustat->irq, tmp); |
2265 | else if (softirq_count()) |
2266 | cpustat->softirq = cputime64_add(cpustat->softirq, tmp); |
2267 | else if (p != rq->idle) |
2268 | cpustat->system = cputime64_add(cpustat->system, tmp); |
2269 | else if (atomic_read(&rq->nr_iowait) > 0) |
2270 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); |
2271 | else |
2272 | cpustat->idle = cputime64_add(cpustat->idle, tmp); |
2273 | |
2274 | /* Account for system time used */ |
2275 | acct_update_integrals(p); |
2276 | /* Update rss highwater mark */ |
2277 | update_mem_hiwater(p); |
2278 | } |
2279 | |
2280 | /* |
2281 | * Account for involuntary wait time. |
2282 | * @p: the process from which the cpu time has been stolen |
2283 | * @steal: the cpu time spent in involuntary wait |
2284 | */ |
2285 | void account_steal_time(struct task_struct *p, cputime_t steal) |
2286 | { |
2287 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
2288 | cputime64_t tmp = cputime_to_cputime64(steal); |
2289 | runqueue_t *rq = this_rq(); |
2290 | |
2291 | if (p == rq->idle) { |
2292 | p->stime = cputime_add(p->stime, steal); |
2293 | if (atomic_read(&rq->nr_iowait) > 0) |
2294 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); |
2295 | else |
2296 | cpustat->idle = cputime64_add(cpustat->idle, tmp); |
2297 | } else |
2298 | cpustat->steal = cputime64_add(cpustat->steal, tmp); |
2299 | } |
2300 | |
2301 | static void time_slice_expired(task_t *p, runqueue_t *rq) |
2302 | { |
2303 | set_tsk_need_resched(p); |
2304 | dequeue_task(p, rq); |
2305 | p->prio = effective_prio(p); |
2306 | p->time_slice = rr_interval(p); |
2307 | enqueue_task(p, rq); |
2308 | } |
2309 | |
2310 | /* |
2311 | * This function gets called by the timer code, with HZ frequency. |
2312 | * We call it with interrupts disabled. |
2313 | */ |
2314 | void scheduler_tick(void) |
2315 | { |
2316 | int cpu = smp_processor_id(); |
2317 | runqueue_t *rq = this_rq(); |
2318 | task_t *p = current; |
2319 | unsigned long debit, expired_balance = rq->nr_running; |
2320 | unsigned long long now = sched_clock(); |
2321 | |
2322 | update_cpu_clock(p, rq, now); |
2323 | |
2324 | rq->timestamp_last_tick = now; |
2325 | |
2326 | if (p == rq->idle) { |
2327 | if (wake_priority_sleeper(rq)) |
2328 | goto out; |
2329 | rebalance_tick(cpu, rq, SCHED_IDLE); |
2330 | return; |
2331 | } |
2332 | |
2333 | /* Task might have expired already, but not scheduled off yet */ |
2334 | if (unlikely(!task_queued(p))) { |
2335 | set_tsk_need_resched(p); |
2336 | goto out; |
2337 | } |
2338 | /* |
2339 | * SCHED_FIFO tasks never run out of timeslice. |
2340 | */ |
2341 | if (unlikely(p->policy == SCHED_FIFO)) { |
2342 | expired_balance = 0; |
2343 | goto out; |
2344 | } |
2345 | |
2346 | spin_lock(&rq->lock); |
2347 | debit = ns_diff(rq->timestamp_last_tick, p->timestamp); |
2348 | p->ns_debit += debit; |
2349 | if (p->ns_debit < NSJIFFY) |
2350 | goto out_unlock; |
2351 | p->ns_debit %= NSJIFFY; |
2352 | /* |
2353 | * Tasks lose burst each time they use up a full slice(). |
2354 | */ |
2355 | if (!--p->slice) { |
2356 | dec_burst(p); |
2357 | p->slice = slice(p); |
2358 | time_slice_expired(p, rq); |
2359 | p->totalrun = 0; |
2360 | goto out_unlock; |
2361 | } |
2362 | /* |
2363 | * Tasks that run out of time_slice but still have slice left get |
2364 | * requeued with a lower priority && RR_INTERVAL time_slice. |
2365 | */ |
2366 | if (!--p->time_slice) { |
2367 | time_slice_expired(p, rq); |
2368 | goto out_unlock; |
2369 | } |
2370 | rq->cache_ticks++; |
2371 | if (rq->preempted && rq->cache_ticks >= cache_delay) { |
2372 | set_tsk_need_resched(p); |
2373 | goto out_unlock; |
2374 | } |
2375 | expired_balance = 0; |
2376 | out_unlock: |
2377 | spin_unlock(&rq->lock); |
2378 | out: |
2379 | if (expired_balance > 1) |
2380 | rebalance_tick(cpu, rq, NOT_IDLE); |
2381 | } |
2382 | |
2383 | #ifdef CONFIG_SCHED_SMT |
2384 | static inline void wakeup_busy_runqueue(runqueue_t *rq) |
2385 | { |
2386 | /* If an SMT runqueue is sleeping due to priority reasons wake it up */ |
2387 | if (rq->curr == rq->idle && rq->nr_running) |
2388 | resched_task(rq->idle); |
2389 | } |
2390 | |
2391 | static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq) |
2392 | { |
2393 | struct sched_domain *sd = this_rq->sd; |
2394 | cpumask_t sibling_map; |
2395 | int i; |
2396 | |
2397 | if (!(sd->flags & SD_SHARE_CPUPOWER)) |
2398 | return; |
2399 | |
2400 | /* |
2401 | * Unlock the current runqueue because we have to lock in |
2402 | * CPU order to avoid deadlocks. Caller knows that we might |
2403 | * unlock. We keep IRQs disabled. |
2404 | */ |
2405 | spin_unlock(&this_rq->lock); |
2406 | |
2407 | sibling_map = sd->span; |
2408 | |
2409 | for_each_cpu_mask(i, sibling_map) |
2410 | spin_lock(&cpu_rq(i)->lock); |
2411 | /* |
2412 | * We clear this CPU from the mask. This both simplifies the |
2413 | * inner loop and keps this_rq locked when we exit: |
2414 | */ |
2415 | cpu_clear(this_cpu, sibling_map); |
2416 | |
2417 | for_each_cpu_mask(i, sibling_map) { |
2418 | runqueue_t *smt_rq = cpu_rq(i); |
2419 | |
2420 | wakeup_busy_runqueue(smt_rq); |
2421 | } |
2422 | |
2423 | for_each_cpu_mask(i, sibling_map) |
2424 | spin_unlock(&cpu_rq(i)->lock); |
2425 | /* |
2426 | * We exit with this_cpu's rq still held and IRQs |
2427 | * still disabled: |
2428 | */ |
2429 | } |
2430 | |
2431 | static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq) |
2432 | { |
2433 | struct sched_domain *sd = this_rq->sd; |
2434 | cpumask_t sibling_map; |
2435 | int ret = 0, i; |
2436 | task_t *p; |
2437 | |
2438 | if (!(sd->flags & SD_SHARE_CPUPOWER)) |
2439 | return 0; |
2440 | |
2441 | /* |
2442 | * The same locking rules and details apply as for |
2443 | * wake_sleeping_dependent(): |
2444 | */ |
2445 | spin_unlock(&this_rq->lock); |
2446 | sibling_map = sd->span; |
2447 | for_each_cpu_mask(i, sibling_map) |
2448 | spin_lock(&cpu_rq(i)->lock); |
2449 | cpu_clear(this_cpu, sibling_map); |
2450 | |
2451 | /* |
2452 | * Establish next task to be run - it might have gone away because |
2453 | * we released the runqueue lock above: |
2454 | */ |
2455 | if (!this_rq->nr_running) |
2456 | goto out_unlock; |
2457 | |
2458 | p = list_entry(this_rq->queue[sched_find_first_bit(this_rq->bitmap)].next, |
2459 | task_t, run_list); |
2460 | |
2461 | for_each_cpu_mask(i, sibling_map) { |
2462 | runqueue_t *smt_rq = cpu_rq(i); |
2463 | task_t *smt_curr = smt_rq->curr; |
2464 | |
2465 | /* Kernel threads do not participate in dependent sleeping */ |
2466 | if (!p->mm || !smt_curr->mm || rt_task(p)) |
2467 | goto check_smt_task; |
2468 | |
2469 | /* |
2470 | * If a user task with lower static priority than the |
2471 | * running task on the SMT sibling is trying to schedule, |
2472 | * delay it till there is proportionately less timeslice |
2473 | * left of the sibling task to prevent a lower priority |
2474 | * task from using an unfair proportion of the |
2475 | * physical cpu's resources. -ck |
2476 | */ |
2477 | if (rt_task(smt_curr)) { |
2478 | /* |
2479 | * With real time tasks we run non-rt tasks only |
2480 | * per_cpu_gain% of the time. |
2481 | */ |
2482 | if ((jiffies % DEF_TIMESLICE) > |
2483 | (sd->per_cpu_gain * DEF_TIMESLICE / 100)) |
2484 | ret = 1; |
2485 | else if (batch_task(p)) |
2486 | ret = 1; |
2487 | } else { |
2488 | if (((smt_curr->slice * (100 - sd->per_cpu_gain) / |
2489 | 100) > slice(p))) |
2490 | ret = 1; |
2491 | else if (batch_task(p) && !batch_task(smt_curr) && |
2492 | smt_curr->slice * sd->per_cpu_gain > |
2493 | slice(smt_curr)) |
2494 | /* |
2495 | * With batch tasks they run just the last |
2496 | * per_cpu_gain percent of the smt task's slice. |
2497 | */ |
2498 | ret = 1; |
2499 | } |
2500 | |
2501 | check_smt_task: |
2502 | if ((!smt_curr->mm && smt_curr != smt_rq->idle) || |
2503 | rt_task(smt_curr)) |
2504 | continue; |
2505 | if (!p->mm) { |
2506 | wakeup_busy_runqueue(smt_rq); |
2507 | continue; |
2508 | } |
2509 | |
2510 | /* |
2511 | * Reschedule a lower priority task on the SMT sibling, |
2512 | * or wake it up if it has been put to sleep for priority |
2513 | * reasons to see if it should run now. |
2514 | */ |
2515 | if (rt_task(p)) { |
2516 | if ((jiffies % DEF_TIMESLICE) > |
2517 | (sd->per_cpu_gain * DEF_TIMESLICE / 100)) |
2518 | resched_task(smt_curr); |
2519 | else if (batch_task(smt_curr)) |
2520 | resched_task(smt_curr); |
2521 | } else { |
2522 | if ((p->slice * (100 - sd->per_cpu_gain) / 100) > |
2523 | slice(smt_curr)) |
2524 | resched_task(smt_curr); |
2525 | else if (batch_task(smt_curr) && !batch_task(p) && |
2526 | p->slice * sd->per_cpu_gain > slice(p)) |
2527 | resched_task(smt_curr); |
2528 | else |
2529 | wakeup_busy_runqueue(smt_rq); |
2530 | } |
2531 | } |
2532 | out_unlock: |
2533 | for_each_cpu_mask(i, sibling_map) |
2534 | spin_unlock(&cpu_rq(i)->lock); |
2535 | return ret; |
2536 | } |
2537 | #else |
2538 | static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq) |
2539 | { |
2540 | } |
2541 | |
2542 | static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq) |
2543 | { |
2544 | return 0; |
2545 | } |
2546 | #endif |
2547 | |
2548 | #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT) |
2549 | |
2550 | void fastcall add_preempt_count(int val) |
2551 | { |
2552 | /* |
2553 | * Underflow? |
2554 | */ |
2555 | BUG_ON(((int)preempt_count() < 0)); |
2556 | preempt_count() += val; |
2557 | /* |
2558 | * Spinlock count overflowing soon? |
2559 | */ |
2560 | BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10); |
2561 | } |
2562 | EXPORT_SYMBOL(add_preempt_count); |
2563 | |
2564 | void fastcall sub_preempt_count(int val) |
2565 | { |
2566 | /* |
2567 | * Underflow? |
2568 | */ |
2569 | BUG_ON(val > preempt_count()); |
2570 | /* |
2571 | * Is the spinlock portion underflowing? |
2572 | */ |
2573 | BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK)); |
2574 | preempt_count() -= val; |
2575 | } |
2576 | EXPORT_SYMBOL(sub_preempt_count); |
2577 | |
2578 | #endif |
2579 | |
2580 | /* |
2581 | * schedule() is the main scheduler function. |
2582 | */ |
2583 | asmlinkage void __sched schedule(void) |
2584 | { |
2585 | long *switch_count; |
2586 | task_t *prev, *next; |
2587 | runqueue_t *rq; |
2588 | struct list_head *queue; |
2589 | unsigned long long now; |
2590 | unsigned long debit; |
2591 | int cpu, idx; |
2592 | |
2593 | /* |
2594 | * Test if we are atomic. Since do_exit() needs to call into |
2595 | * schedule() atomically, we ignore that path for now. |
2596 | * Otherwise, whine if we are scheduling when we should not be. |
2597 | */ |
2598 | if (likely(!current->exit_state)) { |
2599 | if (unlikely(in_atomic())) { |
2600 | printk(KERN_ERR "scheduling while atomic: " |
2601 | "%s/0x%08x/%d\n", |
2602 | current->comm, preempt_count(), current->pid); |
2603 | dump_stack(); |
2604 | } |
2605 | } |
2606 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); |
2607 | |
2608 | need_resched: |
2609 | preempt_disable(); |
2610 | prev = current; |
2611 | release_kernel_lock(prev); |
2612 | need_resched_nonpreemptible: |
2613 | rq = this_rq(); |
2614 | |
2615 | /* |
2616 | * The idle thread is not allowed to schedule! |
2617 | * Remove this check after it has been exercised a bit. |
2618 | */ |
2619 | if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) { |
2620 | printk(KERN_ERR "bad: scheduling from the idle thread!\n"); |
2621 | dump_stack(); |
2622 | } |
2623 | |
2624 | schedstat_inc(rq, sched_cnt); |
2625 | now = sched_clock(); |
2626 | |
2627 | spin_lock_irq(&rq->lock); |
2628 | prev->runtime = ns_diff(now, prev->timestamp); |
2629 | debit = ns_diff(now, rq->timestamp_last_tick) % NSJIFFY; |
2630 | prev->ns_debit += debit; |
2631 | |
2632 | if (unlikely(prev->flags & PF_DEAD)) |
2633 | prev->state = EXIT_DEAD; |
2634 | |
2635 | switch_count = &prev->nivcsw; |
2636 | if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { |
2637 | switch_count = &prev->nvcsw; |
2638 | if (unlikely((prev->state & TASK_INTERRUPTIBLE) && |
2639 | unlikely(signal_pending(prev)))) |
2640 | prev->state = TASK_RUNNING; |
2641 | else { |
2642 | if (prev->state == TASK_UNINTERRUPTIBLE) { |
2643 | prev->flags |= PF_NONSLEEP; |
2644 | rq->nr_uninterruptible++; |
2645 | } |
2646 | deactivate_task(prev, rq); |
2647 | } |
2648 | } |
2649 | |
2650 | cpu = smp_processor_id(); |
2651 | if (unlikely(!rq->nr_running)) { |
2652 | go_idle: |
2653 | idle_balance(cpu, rq); |
2654 | if (!rq->nr_running) { |
2655 | next = rq->idle; |
2656 | wake_sleeping_dependent(cpu, rq); |
2657 | /* |
2658 | * wake_sleeping_dependent() might have released |
2659 | * the runqueue, so break out if we got new |
2660 | * tasks meanwhile: |
2661 | */ |
2662 | if (!rq->nr_running) |
2663 | goto switch_tasks; |
2664 | } |
2665 | } else { |
2666 | if (dependent_sleeper(cpu, rq)) { |
2667 | next = rq->idle; |
2668 | goto switch_tasks; |
2669 | } |
2670 | /* |
2671 | * dependent_sleeper() releases and reacquires the runqueue |
2672 | * lock, hence go into the idle loop if the rq went |
2673 | * empty meanwhile: |
2674 | */ |
2675 | if (unlikely(!rq->nr_running)) |
2676 | goto go_idle; |
2677 | } |
2678 | |
2679 | idx = sched_find_first_bit(rq->bitmap); |
2680 | queue = rq->queue + idx; |
2681 | next = list_entry(queue->next, task_t, run_list); |
2682 | |
2683 | switch_tasks: |
2684 | if (next == rq->idle) |
2685 | schedstat_inc(rq, sched_goidle); |
2686 | prev->timestamp = now; |
2687 | if (unlikely(next->flags & PF_YIELDED)) { |
2688 | /* |
2689 | * Tasks that have yield()ed get requeued at normal priority |
2690 | */ |
2691 | int newprio = effective_prio(next); |
2692 | next->flags &= ~PF_YIELDED; |
2693 | if (newprio != next->prio) { |
2694 | dequeue_task(next, rq); |
2695 | next->prio = newprio; |
2696 | enqueue_task(next, rq); |
2697 | } |
2698 | } |
2699 | |
2700 | prefetch(next); |
2701 | clear_tsk_need_resched(prev); |
2702 | rcu_qsctr_inc(task_cpu(prev)); |
2703 | |
2704 | update_cpu_clock(prev, rq, now); |
2705 | |
2706 | sched_info_switch(prev, next); |
2707 | if (likely(prev != next)) { |
2708 | rq->preempted = 0; |
2709 | rq->cache_ticks = 0; |
2710 | next->timestamp = now; |
2711 | rq->nr_switches++; |
2712 | rq->curr = next; |
2713 | ++*switch_count; |
2714 | |
2715 | prepare_arch_switch(rq, next); |
2716 | prev = context_switch(rq, prev, next); |
2717 | barrier(); |
2718 | |
2719 | finish_task_switch(prev); |
2720 | } else |
2721 | spin_unlock_irq(&rq->lock); |
2722 | |
2723 | prev = current; |
2724 | if (unlikely(reacquire_kernel_lock(prev) < 0)) |
2725 | goto need_resched_nonpreemptible; |
2726 | preempt_enable_no_resched(); |
2727 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) |
2728 | goto need_resched; |
2729 | } |
2730 | |
2731 | EXPORT_SYMBOL(schedule); |
2732 | |
2733 | #ifdef CONFIG_PREEMPT |
2734 | /* |
2735 | * this is is the entry point to schedule() from in-kernel preemption |
2736 | * off of preempt_enable. Kernel preemptions off return from interrupt |
2737 | * occur there and call schedule directly. |
2738 | */ |
2739 | asmlinkage void __sched preempt_schedule(void) |
2740 | { |
2741 | struct thread_info *ti = current_thread_info(); |
2742 | #ifdef CONFIG_PREEMPT_BKL |
2743 | struct task_struct *task = current; |
2744 | int saved_lock_depth; |
2745 | #endif |
2746 | /* |
2747 | * If there is a non-zero preempt_count or interrupts are disabled, |
2748 | * we do not want to preempt the current task. Just return.. |
2749 | */ |
2750 | if (unlikely(ti->preempt_count || irqs_disabled())) |
2751 | return; |
2752 | |
2753 | need_resched: |
2754 | add_preempt_count(PREEMPT_ACTIVE); |
2755 | /* |
2756 | * We keep the big kernel semaphore locked, but we |
2757 | * clear ->lock_depth so that schedule() doesnt |
2758 | * auto-release the semaphore: |
2759 | */ |
2760 | #ifdef CONFIG_PREEMPT_BKL |
2761 | saved_lock_depth = task->lock_depth; |
2762 | task->lock_depth = -1; |
2763 | #endif |
2764 | schedule(); |
2765 | #ifdef CONFIG_PREEMPT_BKL |
2766 | task->lock_depth = saved_lock_depth; |
2767 | #endif |
2768 | sub_preempt_count(PREEMPT_ACTIVE); |
2769 | |
2770 | /* we could miss a preemption opportunity between schedule and now */ |
2771 | barrier(); |
2772 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) |
2773 | goto need_resched; |
2774 | } |
2775 | |
2776 | EXPORT_SYMBOL(preempt_schedule); |
2777 | |
2778 | /* |
2779 | * this is is the entry point to schedule() from kernel preemption |
2780 | * off of irq context. |
2781 | * Note, that this is called and return with irqs disabled. This will |
2782 | * protect us against recursive calling from irq. |
2783 | */ |
2784 | asmlinkage void __sched preempt_schedule_irq(void) |
2785 | { |
2786 | struct thread_info *ti = current_thread_info(); |
2787 | #ifdef CONFIG_PREEMPT_BKL |
2788 | struct task_struct *task = current; |
2789 | int saved_lock_depth; |
2790 | #endif |
2791 | /* Catch callers which need to be fixed*/ |
2792 | BUG_ON(ti->preempt_count || !irqs_disabled()); |
2793 | |
2794 | need_resched: |
2795 | add_preempt_count(PREEMPT_ACTIVE); |
2796 | /* |
2797 | * We keep the big kernel semaphore locked, but we |
2798 | * clear ->lock_depth so that schedule() doesnt |
2799 | * auto-release the semaphore: |
2800 | */ |
2801 | #ifdef CONFIG_PREEMPT_BKL |
2802 | saved_lock_depth = task->lock_depth; |
2803 | task->lock_depth = -1; |
2804 | #endif |
2805 | local_irq_enable(); |
2806 | schedule(); |
2807 | local_irq_disable(); |
2808 | #ifdef CONFIG_PREEMPT_BKL |
2809 | task->lock_depth = saved_lock_depth; |
2810 | #endif |
2811 | sub_preempt_count(PREEMPT_ACTIVE); |
2812 | |
2813 | /* we could miss a preemption opportunity between schedule and now */ |
2814 | barrier(); |
2815 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) |
2816 | goto need_resched; |
2817 | } |
2818 | |
2819 | #endif /* CONFIG_PREEMPT */ |
2820 | |
2821 | int default_wake_function(wait_queue_t *curr, unsigned mode, int sync, void *key) |
2822 | { |
2823 | task_t *p = curr->task; |
2824 | return try_to_wake_up(p, mode, sync); |
2825 | } |
2826 | |
2827 | EXPORT_SYMBOL(default_wake_function); |
2828 | |
2829 | /* |
2830 | * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just |
2831 | * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve |
2832 | * number) then we wake all the non-exclusive tasks and one exclusive task. |
2833 | * |
2834 | * There are circumstances in which we can try to wake a task which has already |
2835 | * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns |
2836 | * zero in this (rare) case, and we handle it by continuing to scan the queue. |
2837 | */ |
2838 | static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, |
2839 | int nr_exclusive, int sync, void *key) |
2840 | { |
2841 | struct list_head *tmp, *next; |
2842 | |
2843 | list_for_each_safe(tmp, next, &q->task_list) { |
2844 | wait_queue_t *curr; |
2845 | unsigned flags; |
2846 | curr = list_entry(tmp, wait_queue_t, task_list); |
2847 | flags = curr->flags; |
2848 | if (curr->func(curr, mode, sync, key) && |
2849 | (flags & WQ_FLAG_EXCLUSIVE) && |
2850 | !--nr_exclusive) |
2851 | break; |
2852 | } |
2853 | } |
2854 | |
2855 | /** |
2856 | * __wake_up - wake up threads blocked on a waitqueue. |
2857 | * @q: the waitqueue |
2858 | * @mode: which threads |
2859 | * @nr_exclusive: how many wake-one or wake-many threads to wake up |
2860 | * @key: is directly passed to the wakeup function |
2861 | */ |
2862 | void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode, |
2863 | int nr_exclusive, void *key) |
2864 | { |
2865 | unsigned long flags; |
2866 | |
2867 | spin_lock_irqsave(&q->lock, flags); |
2868 | __wake_up_common(q, mode, nr_exclusive, 0, key); |
2869 | spin_unlock_irqrestore(&q->lock, flags); |
2870 | } |
2871 | |
2872 | EXPORT_SYMBOL(__wake_up); |
2873 | |
2874 | /* |
2875 | * Same as __wake_up but called with the spinlock in wait_queue_head_t held. |
2876 | */ |
2877 | void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode) |
2878 | { |
2879 | __wake_up_common(q, mode, 1, 0, NULL); |
2880 | } |
2881 | |
2882 | /** |
2883 | * __wake_up_sync - wake up threads blocked on a waitqueue. |
2884 | * @q: the waitqueue |
2885 | * @mode: which threads |
2886 | * @nr_exclusive: how many wake-one or wake-many threads to wake up |
2887 | * |
2888 | * The sync wakeup differs that the waker knows that it will schedule |
2889 | * away soon, so while the target thread will be woken up, it will not |
2890 | * be migrated to another CPU - ie. the two threads are 'synchronized' |
2891 | * with each other. This can prevent needless bouncing between CPUs. |
2892 | * |
2893 | * On UP it can prevent extra preemption. |
2894 | */ |
2895 | void fastcall __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) |
2896 | { |
2897 | unsigned long flags; |
2898 | int sync = 1; |
2899 | |
2900 | if (unlikely(!q)) |
2901 | return; |
2902 | |
2903 | if (unlikely(!nr_exclusive)) |
2904 | sync = 0; |
2905 | |
2906 | spin_lock_irqsave(&q->lock, flags); |
2907 | __wake_up_common(q, mode, nr_exclusive, sync, NULL); |
2908 | spin_unlock_irqrestore(&q->lock, flags); |
2909 | } |
2910 | EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ |
2911 | |
2912 | void fastcall complete(struct completion *x) |
2913 | { |
2914 | unsigned long flags; |
2915 | |
2916 | spin_lock_irqsave(&x->wait.lock, flags); |
2917 | x->done++; |
2918 | __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, |
2919 | 1, 0, NULL); |
2920 | spin_unlock_irqrestore(&x->wait.lock, flags); |
2921 | } |
2922 | EXPORT_SYMBOL(complete); |
2923 | |
2924 | void fastcall complete_all(struct completion *x) |
2925 | { |
2926 | unsigned long flags; |
2927 | |
2928 | spin_lock_irqsave(&x->wait.lock, flags); |
2929 | x->done += UINT_MAX/2; |
2930 | __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, |
2931 | 0, 0, NULL); |
2932 | spin_unlock_irqrestore(&x->wait.lock, flags); |
2933 | } |
2934 | EXPORT_SYMBOL(complete_all); |
2935 | |
2936 | void fastcall __sched wait_for_completion(struct completion *x) |
2937 | { |
2938 | might_sleep(); |
2939 | spin_lock_irq(&x->wait.lock); |
2940 | if (!x->done) { |
2941 | DECLARE_WAITQUEUE(wait, current); |
2942 | |
2943 | wait.flags |= WQ_FLAG_EXCLUSIVE; |
2944 | __add_wait_queue_tail(&x->wait, &wait); |
2945 | do { |
2946 | __set_current_state(TASK_UNINTERRUPTIBLE); |
2947 | spin_unlock_irq(&x->wait.lock); |
2948 | schedule(); |
2949 | spin_lock_irq(&x->wait.lock); |
2950 | } while (!x->done); |
2951 | __remove_wait_queue(&x->wait, &wait); |
2952 | } |
2953 | x->done--; |
2954 | spin_unlock_irq(&x->wait.lock); |
2955 | } |
2956 | EXPORT_SYMBOL(wait_for_completion); |
2957 | |
2958 | unsigned long fastcall __sched |
2959 | wait_for_completion_timeout(struct completion *x, unsigned long timeout) |
2960 | { |
2961 | might_sleep(); |
2962 | |
2963 | spin_lock_irq(&x->wait.lock); |
2964 | if (!x->done) { |
2965 | DECLARE_WAITQUEUE(wait, current); |
2966 | |
2967 | wait.flags |= WQ_FLAG_EXCLUSIVE; |
2968 | __add_wait_queue_tail(&x->wait, &wait); |
2969 | do { |
2970 | __set_current_state(TASK_UNINTERRUPTIBLE); |
2971 | spin_unlock_irq(&x->wait.lock); |
2972 | timeout = schedule_timeout(timeout); |
2973 | spin_lock_irq(&x->wait.lock); |
2974 | if (!timeout) { |
2975 | __remove_wait_queue(&x->wait, &wait); |
2976 | goto out; |
2977 | } |
2978 | } while (!x->done); |
2979 | __remove_wait_queue(&x->wait, &wait); |
2980 | } |
2981 | x->done--; |
2982 | out: |
2983 | spin_unlock_irq(&x->wait.lock); |
2984 | return timeout; |
2985 | } |
2986 | EXPORT_SYMBOL(wait_for_completion_timeout); |
2987 | |
2988 | int fastcall __sched wait_for_completion_interruptible(struct completion *x) |
2989 | { |
2990 | int ret = 0; |
2991 | |
2992 | might_sleep(); |
2993 | |
2994 | spin_lock_irq(&x->wait.lock); |
2995 | if (!x->done) { |
2996 | DECLARE_WAITQUEUE(wait, current); |
2997 | |
2998 | wait.flags |= WQ_FLAG_EXCLUSIVE; |
2999 | __add_wait_queue_tail(&x->wait, &wait); |
3000 | do { |
3001 | if (signal_pending(current)) { |
3002 | ret = -ERESTARTSYS; |
3003 | __remove_wait_queue(&x->wait, &wait); |
3004 | goto out; |
3005 | } |
3006 | __set_current_state(TASK_INTERRUPTIBLE); |
3007 | spin_unlock_irq(&x->wait.lock); |
3008 | schedule(); |
3009 | spin_lock_irq(&x->wait.lock); |
3010 | } while (!x->done); |
3011 | __remove_wait_queue(&x->wait, &wait); |
3012 | } |
3013 | x->done--; |
3014 | out: |
3015 | spin_unlock_irq(&x->wait.lock); |
3016 | |
3017 | return ret; |
3018 | } |
3019 | EXPORT_SYMBOL(wait_for_completion_interruptible); |
3020 | |
3021 | unsigned long fastcall __sched |
3022 | wait_for_completion_interruptible_timeout(struct completion *x, |
3023 | unsigned long timeout) |
3024 | { |
3025 | might_sleep(); |
3026 | |
3027 | spin_lock_irq(&x->wait.lock); |
3028 | if (!x->done) { |
3029 | DECLARE_WAITQUEUE(wait, current); |
3030 | |
3031 | wait.flags |= WQ_FLAG_EXCLUSIVE; |
3032 | __add_wait_queue_tail(&x->wait, &wait); |
3033 | do { |
3034 | if (signal_pending(current)) { |
3035 | timeout = -ERESTARTSYS; |
3036 | __remove_wait_queue(&x->wait, &wait); |
3037 | goto out; |
3038 | } |
3039 | __set_current_state(TASK_INTERRUPTIBLE); |
3040 | spin_unlock_irq(&x->wait.lock); |
3041 | timeout = schedule_timeout(timeout); |
3042 | spin_lock_irq(&x->wait.lock); |
3043 | if (!timeout) { |
3044 | __remove_wait_queue(&x->wait, &wait); |
3045 | goto out; |
3046 | } |
3047 | } while (!x->done); |
3048 | __remove_wait_queue(&x->wait, &wait); |
3049 | } |
3050 | x->done--; |
3051 | out: |
3052 | spin_unlock_irq(&x->wait.lock); |
3053 | return timeout; |
3054 | } |
3055 | EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); |
3056 | |
3057 | |
3058 | #define SLEEP_ON_VAR \ |
3059 | unsigned long flags; \ |
3060 | wait_queue_t wait; \ |
3061 | init_waitqueue_entry(&wait, current); |
3062 | |
3063 | #define SLEEP_ON_HEAD \ |
3064 | spin_lock_irqsave(&q->lock,flags); \ |
3065 | __add_wait_queue(q, &wait); \ |
3066 | spin_unlock(&q->lock); |
3067 | |
3068 | #define SLEEP_ON_TAIL \ |
3069 | spin_lock_irq(&q->lock); \ |
3070 | __remove_wait_queue(q, &wait); \ |
3071 | spin_unlock_irqrestore(&q->lock, flags); |
3072 | |
3073 | void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q) |
3074 | { |
3075 | SLEEP_ON_VAR |
3076 | |
3077 | current->state = TASK_INTERRUPTIBLE; |
3078 | |
3079 | SLEEP_ON_HEAD |
3080 | schedule(); |
3081 | SLEEP_ON_TAIL |
3082 | } |
3083 | |
3084 | EXPORT_SYMBOL(interruptible_sleep_on); |
3085 | |
3086 | long fastcall __sched interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) |
3087 | { |
3088 | SLEEP_ON_VAR |
3089 | |
3090 | current->state = TASK_INTERRUPTIBLE; |
3091 | |
3092 | SLEEP_ON_HEAD |
3093 | timeout = schedule_timeout(timeout); |
3094 | SLEEP_ON_TAIL |
3095 | |
3096 | return timeout; |
3097 | } |
3098 | |
3099 | EXPORT_SYMBOL(interruptible_sleep_on_timeout); |
3100 | |
3101 | void fastcall __sched sleep_on(wait_queue_head_t *q) |
3102 | { |
3103 | SLEEP_ON_VAR |
3104 | |
3105 | current->state = TASK_UNINTERRUPTIBLE; |
3106 | |
3107 | SLEEP_ON_HEAD |
3108 | schedule(); |
3109 | SLEEP_ON_TAIL |
3110 | } |
3111 | |
3112 | EXPORT_SYMBOL(sleep_on); |
3113 | |
3114 | long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) |
3115 | { |
3116 | SLEEP_ON_VAR |
3117 | |
3118 | current->state = TASK_UNINTERRUPTIBLE; |
3119 | |
3120 | SLEEP_ON_HEAD |
3121 | timeout = schedule_timeout(timeout); |
3122 | SLEEP_ON_TAIL |
3123 | |
3124 | return timeout; |
3125 | } |
3126 | |
3127 | EXPORT_SYMBOL(sleep_on_timeout); |
3128 | |
3129 | void set_user_nice(task_t *p, long nice) |
3130 | { |
3131 | unsigned long flags; |
3132 | runqueue_t *rq; |
3133 | int queued, old_prio, new_prio, delta; |
3134 | |
3135 | if (TASK_NICE(p) == nice || nice < -20 || nice > 19) |
3136 | return; |
3137 | /* |
3138 | * We have to be careful, if called from sys_setpriority(), |
3139 | * the task might be in the middle of scheduling on another CPU. |
3140 | */ |
3141 | rq = task_rq_lock(p, &flags); |
3142 | /* |
3143 | * The RT priorities are set via sched_setscheduler(), but we still |
3144 | * allow the 'normal' nice value to be set - but as expected |
3145 | * it wont have any effect on scheduling until the task is |
3146 | * not SCHED_NORMAL: |
3147 | */ |
3148 | if (rt_task(p)) { |
3149 | p->static_prio = NICE_TO_PRIO(nice); |
3150 | goto out_unlock; |
3151 | } |
3152 | if ((queued = task_queued(p))) { |
3153 | dequeue_task(p, rq); |
3154 | dec_prio_bias(rq, p->static_prio); |
3155 | } |
3156 | |
3157 | old_prio = p->prio; |
3158 | new_prio = NICE_TO_PRIO(nice); |
3159 | delta = new_prio - old_prio; |
3160 | p->static_prio = NICE_TO_PRIO(nice); |
3161 | p->prio += delta; |
3162 | |
3163 | if (queued) { |
3164 | enqueue_task(p, rq); |
3165 | inc_prio_bias(rq, p->static_prio); |
3166 | /* |
3167 | * If the task increased its priority or is running and |
3168 | * lowered its priority, then reschedule its CPU: |
3169 | */ |
3170 | if (delta < 0 || ((delta > 0 || batch_task(p)) && |
3171 | task_running(rq, p))) |
3172 | resched_task(rq->curr); |
3173 | } |
3174 | out_unlock: |
3175 | task_rq_unlock(rq, &flags); |
3176 | } |
3177 | |
3178 | EXPORT_SYMBOL(set_user_nice); |
3179 | |
3180 | /* |
3181 | * can_nice - check if a task can reduce its nice value |
3182 | * @p: task |
3183 | * @nice: nice value |
3184 | */ |
3185 | int can_nice(const task_t *p, const int nice) |
3186 | { |
3187 | /* convert nice value [19,-20] to rlimit style value [0,39] */ |
3188 | int nice_rlim = 19 - nice; |
3189 | return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur || |
3190 | capable(CAP_SYS_NICE)); |
3191 | } |
3192 | |
3193 | #ifdef __ARCH_WANT_SYS_NICE |
3194 | |
3195 | /* |
3196 | * sys_nice - change the priority of the current process. |
3197 | * @increment: priority increment |
3198 | * |
3199 | * sys_setpriority is a more generic, but much slower function that |
3200 | * does similar things. |
3201 | */ |
3202 | asmlinkage long sys_nice(int increment) |
3203 | { |
3204 | int retval; |
3205 | long nice; |
3206 | |
3207 | /* |
3208 | * Setpriority might change our priority at the same moment. |
3209 | * We don't have to worry. Conceptually one call occurs first |
3210 | * and we have a single winner. |
3211 | */ |
3212 | if (increment < -40) |
3213 | increment = -40; |
3214 | if (increment > 40) |
3215 | increment = 40; |
3216 | |
3217 | nice = PRIO_TO_NICE(current->static_prio) + increment; |
3218 | if (nice < -20) |
3219 | nice = -20; |
3220 | if (nice > 19) |
3221 | nice = 19; |
3222 | |
3223 | if (increment < 0 && !can_nice(current, nice)) |
3224 | return -EPERM; |
3225 | |
3226 | retval = security_task_setnice(current, nice); |
3227 | if (retval) |
3228 | return retval; |
3229 | |
3230 | set_user_nice(current, nice); |
3231 | return 0; |
3232 | } |
3233 | |
3234 | #endif |
3235 | |
3236 | /** |
3237 | * task_prio - return the priority value of a given task. |
3238 | * @p: the task in question. |
3239 | * |
3240 | * This is the priority value as seen by users in /proc. |
3241 | * RT tasks are offset by -200. Normal tasks are centered |
3242 | * around 0, value goes from -16 to +15. |
3243 | */ |
3244 | int task_prio(const task_t *p) |
3245 | { |
3246 | return p->prio - MAX_RT_PRIO; |
3247 | } |
3248 | |
3249 | /** |
3250 | * task_nice - return the nice value of a given task. |
3251 | * @p: the task in question. |
3252 | */ |
3253 | int task_nice(const task_t *p) |
3254 | { |
3255 | return TASK_NICE(p); |
3256 | } |
3257 | EXPORT_SYMBOL_GPL(task_nice); |
3258 | |
3259 | /** |
3260 | * idle_cpu - is a given cpu idle currently? |
3261 | * @cpu: the processor in question. |
3262 | */ |
3263 | int idle_cpu(int cpu) |
3264 | { |
3265 | return cpu_curr(cpu) == cpu_rq(cpu)->idle; |
3266 | } |
3267 | |
3268 | EXPORT_SYMBOL_GPL(idle_cpu); |
3269 | |
3270 | /** |
3271 | * idle_task - return the idle task for a given cpu. |
3272 | * @cpu: the processor in question. |
3273 | */ |
3274 | task_t *idle_task(int cpu) |
3275 | { |
3276 | return cpu_rq(cpu)->idle; |
3277 | } |
3278 | |
3279 | /** |
3280 | * find_process_by_pid - find a process with a matching PID value. |
3281 | * @pid: the pid in question. |
3282 | */ |
3283 | static inline task_t *find_process_by_pid(pid_t pid) |
3284 | { |
3285 | return pid ? find_task_by_pid(pid) : current; |
3286 | } |
3287 | |
3288 | /* Actually do priority change: must hold rq lock. */ |
3289 | static void __setscheduler(struct task_struct *p, int policy, int prio) |
3290 | { |
3291 | BUG_ON(task_queued(p)); |
3292 | p->policy = policy; |
3293 | p->rt_priority = prio; |
3294 | if (SCHED_RT(policy)) |
3295 | p->prio = MAX_USER_RT_PRIO-1 - p->rt_priority; |
3296 | else |
3297 | p->prio = p->static_prio; |
3298 | } |
3299 | |
3300 | /** |
3301 | * sched_setscheduler - change the scheduling policy and/or RT priority of |
3302 | * a thread. |
3303 | * @p: the task in question. |
3304 | * @policy: new policy. |
3305 | * @param: structure containing the new RT priority. |
3306 | */ |
3307 | int sched_setscheduler(struct task_struct *p, int policy, struct sched_param *param) |
3308 | { |
3309 | int retval; |
3310 | int queued, oldprio, oldpolicy = -1; |
3311 | unsigned long flags; |
3312 | runqueue_t *rq; |
3313 | |
3314 | recheck: |
3315 | /* double check policy once rq lock held */ |
3316 | if (policy < 0) |
3317 | policy = oldpolicy = p->policy; |
3318 | else if (!SCHED_RANGE(policy)) |
3319 | return -EINVAL; |
3320 | /* |
3321 | * Valid priorities for SCHED_FIFO and SCHED_RR are |
3322 | * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL is 0. |
3323 | */ |
3324 | if (param->sched_priority < 0 || |
3325 | param->sched_priority > MAX_USER_RT_PRIO-1) |
3326 | return -EINVAL; |
3327 | if ((!SCHED_RT(policy)) != (param->sched_priority == 0)) |
3328 | return -EINVAL; |
3329 | |
3330 | if (SCHED_RT(policy) && |
3331 | param->sched_priority > p->signal->rlim[RLIMIT_RTPRIO].rlim_cur && |
3332 | !capable(CAP_SYS_NICE)) |
3333 | return -EPERM; |
3334 | if ((current->euid != p->euid) && (current->euid != p->uid) && |
3335 | !capable(CAP_SYS_NICE)) |
3336 | return -EPERM; |
3337 | |
3338 | if (!(p->mm) && policy == SCHED_BATCH) |
3339 | /* |
3340 | * Don't allow kernel threads to be SCHED_BATCH. |
3341 | */ |
3342 | return -EINVAL; |
3343 | |
3344 | retval = security_task_setscheduler(p, policy, param); |
3345 | if (retval) |
3346 | return retval; |
3347 | /* |
3348 | * To be able to change p->policy safely, the apropriate |
3349 | * runqueue lock must be held. |
3350 | */ |
3351 | rq = task_rq_lock(p, &flags); |
3352 | /* recheck policy now with rq lock held */ |
3353 | if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { |
3354 | policy = oldpolicy = -1; |
3355 | task_rq_unlock(rq, &flags); |
3356 | goto recheck; |
3357 | } |
3358 | if ((queued = task_queued(p))) |
3359 | deactivate_task(p, rq); |
3360 | oldprio = p->prio; |
3361 | __setscheduler(p, policy, param->sched_priority); |
3362 | if (queued) { |
3363 | __activate_task(p, rq); |
3364 | /* |
3365 | * Reschedule if we are currently running on this runqueue and |
3366 | * our priority decreased, or if we are not currently running on |
3367 | * this runqueue and our priority is higher than the current's |
3368 | */ |
3369 | if (task_running(rq, p)) { |
3370 | if (p->prio > oldprio) |
3371 | resched_task(rq->curr); |
3372 | } else |
3373 | preempt(p, rq); |
3374 | } |
3375 | task_rq_unlock(rq, &flags); |
3376 | return 0; |
3377 | } |
3378 | EXPORT_SYMBOL_GPL(sched_setscheduler); |
3379 | |
3380 | static int do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) |
3381 | { |
3382 | int retval; |
3383 | struct sched_param lparam; |
3384 | struct task_struct *p; |
3385 | |
3386 | if (!param || pid < 0) |
3387 | return -EINVAL; |
3388 | if (copy_from_user(&lparam, param, sizeof(struct sched_param))) |
3389 | return -EFAULT; |
3390 | read_lock_irq(&tasklist_lock); |
3391 | p = find_process_by_pid(pid); |
3392 | if (!p) { |
3393 | read_unlock_irq(&tasklist_lock); |
3394 | return -ESRCH; |
3395 | } |
3396 | retval = sched_setscheduler(p, policy, &lparam); |
3397 | read_unlock_irq(&tasklist_lock); |
3398 | return retval; |
3399 | } |
3400 | |
3401 | /** |
3402 | * sys_sched_setscheduler - set/change the scheduler policy and RT priority |
3403 | * @pid: the pid in question. |
3404 | * @policy: new policy. |
3405 | * @param: structure containing the new RT priority. |
3406 | */ |
3407 | asmlinkage long sys_sched_setscheduler(pid_t pid, int policy, |
3408 | struct sched_param __user *param) |
3409 | { |
3410 | return do_sched_setscheduler(pid, policy, param); |
3411 | } |
3412 | |
3413 | /** |
3414 | * sys_sched_setparam - set/change the RT priority of a thread |
3415 | * @pid: the pid in question. |
3416 | * @param: structure containing the new RT priority. |
3417 | */ |
3418 | asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param) |
3419 | { |
3420 | return do_sched_setscheduler(pid, -1, param); |
3421 | } |
3422 | |
3423 | /** |
3424 | * sys_sched_getscheduler - get the policy (scheduling class) of a thread |
3425 | * @pid: the pid in question. |
3426 | */ |
3427 | asmlinkage long sys_sched_getscheduler(pid_t pid) |
3428 | { |
3429 | int retval = -EINVAL; |
3430 | task_t *p; |
3431 | |
3432 | if (pid < 0) |
3433 | goto out_nounlock; |
3434 | |
3435 | retval = -ESRCH; |
3436 | read_lock(&tasklist_lock); |
3437 | p = find_process_by_pid(pid); |
3438 | if (p) { |
3439 | retval = security_task_getscheduler(p); |
3440 | if (!retval) |
3441 | retval = p->policy; |
3442 | } |
3443 | read_unlock(&tasklist_lock); |
3444 | |
3445 | out_nounlock: |
3446 | return retval; |
3447 | } |
3448 | |
3449 | /** |
3450 | * sys_sched_getscheduler - get the RT priority of a thread |
3451 | * @pid: the pid in question. |
3452 | * @param: structure containing the RT priority. |
3453 | */ |
3454 | asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param) |
3455 | { |
3456 | struct sched_param lp; |
3457 | int retval = -EINVAL; |
3458 | task_t *p; |
3459 | |
3460 | if (!param || pid < 0) |
3461 | goto out_nounlock; |
3462 | |
3463 | read_lock(&tasklist_lock); |
3464 | p = find_process_by_pid(pid); |
3465 | retval = -ESRCH; |
3466 | if (!p) |
3467 | goto out_unlock; |
3468 | |
3469 | retval = security_task_getscheduler(p); |
3470 | if (retval) |
3471 | goto out_unlock; |
3472 | |
3473 | lp.sched_priority = p->rt_priority; |
3474 | read_unlock(&tasklist_lock); |
3475 | |
3476 | /* |
3477 | * This one might sleep, we cannot do it with a spinlock held ... |
3478 | */ |
3479 | retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; |
3480 | |
3481 | out_nounlock: |
3482 | return retval; |
3483 | |
3484 | out_unlock: |
3485 | read_unlock(&tasklist_lock); |
3486 | return retval; |
3487 | } |
3488 | |
3489 | long sched_setaffinity(pid_t pid, cpumask_t new_mask) |
3490 | { |
3491 | task_t *p; |
3492 | int retval; |
3493 | cpumask_t cpus_allowed; |
3494 | |
3495 | lock_cpu_hotplug(); |
3496 | read_lock(&tasklist_lock); |
3497 | |
3498 | p = find_process_by_pid(pid); |
3499 | if (!p) { |
3500 | read_unlock(&tasklist_lock); |
3501 | unlock_cpu_hotplug(); |
3502 | return -ESRCH; |
3503 | } |
3504 | |
3505 | /* |
3506 | * It is not safe to call set_cpus_allowed with the |
3507 | * tasklist_lock held. We will bump the task_struct's |
3508 | * usage count and then drop tasklist_lock. |
3509 | */ |
3510 | get_task_struct(p); |
3511 | read_unlock(&tasklist_lock); |
3512 | |
3513 | retval = -EPERM; |
3514 | if ((current->euid != p->euid) && (current->euid != p->uid) && |
3515 | !capable(CAP_SYS_NICE)) |
3516 | goto out_unlock; |
3517 | |
3518 | cpus_allowed = cpuset_cpus_allowed(p); |
3519 | cpus_and(new_mask, new_mask, cpus_allowed); |
3520 | retval = set_cpus_allowed(p, new_mask); |
3521 | |
3522 | out_unlock: |
3523 | put_task_struct(p); |
3524 | unlock_cpu_hotplug(); |
3525 | return retval; |
3526 | } |
3527 | |
3528 | static inline int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, |
3529 | cpumask_t *new_mask) |
3530 | { |
3531 | if (len < sizeof(cpumask_t)) { |
3532 | memset(new_mask, 0, sizeof(cpumask_t)); |
3533 | } else if (len > sizeof(cpumask_t)) { |
3534 | len = sizeof(cpumask_t); |
3535 | } |
3536 | return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; |
3537 | } |
3538 | |
3539 | /** |
3540 | * sys_sched_setaffinity - set the cpu affinity of a process |
3541 | * @pid: pid of the process |
3542 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr |
3543 | * @user_mask_ptr: user-space pointer to the new cpu mask |
3544 | */ |
3545 | asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len, |
3546 | unsigned long __user *user_mask_ptr) |
3547 | { |
3548 | cpumask_t new_mask; |
3549 | int retval; |
3550 | |
3551 | retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask); |
3552 | if (retval) |
3553 | return retval; |
3554 | |
3555 | return sched_setaffinity(pid, new_mask); |
3556 | } |
3557 | |
3558 | /* |
3559 | * Represents all cpu's present in the system |
3560 | * In systems capable of hotplug, this map could dynamically grow |
3561 | * as new cpu's are detected in the system via any platform specific |
3562 | * method, such as ACPI for e.g. |
3563 | */ |
3564 | |
3565 | cpumask_t cpu_present_map; |
3566 | EXPORT_SYMBOL(cpu_present_map); |
3567 | |
3568 | #ifndef CONFIG_SMP |
3569 | cpumask_t cpu_online_map = CPU_MASK_ALL; |
3570 | cpumask_t cpu_possible_map = CPU_MASK_ALL; |
3571 | #endif |
3572 | |
3573 | long sched_getaffinity(pid_t pid, cpumask_t *mask) |
3574 | { |
3575 | int retval; |
3576 | task_t *p; |
3577 | |
3578 | lock_cpu_hotplug(); |
3579 | read_lock(&tasklist_lock); |
3580 | |
3581 | retval = -ESRCH; |
3582 | p = find_process_by_pid(pid); |
3583 | if (!p) |
3584 | goto out_unlock; |
3585 | |
3586 | retval = 0; |
3587 | cpus_and(*mask, p->cpus_allowed, cpu_possible_map); |
3588 | |
3589 | out_unlock: |
3590 | read_unlock(&tasklist_lock); |
3591 | unlock_cpu_hotplug(); |
3592 | if (retval) |
3593 | return retval; |
3594 | |
3595 | return 0; |
3596 | } |
3597 | |
3598 | /** |
3599 | * sys_sched_getaffinity - get the cpu affinity of a process |
3600 | * @pid: pid of the process |
3601 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr |
3602 | * @user_mask_ptr: user-space pointer to hold the current cpu mask |
3603 | */ |
3604 | asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len, |
3605 | unsigned long __user *user_mask_ptr) |
3606 | { |
3607 | int ret; |
3608 | cpumask_t mask; |
3609 | |
3610 | if (len < sizeof(cpumask_t)) |
3611 | return -EINVAL; |
3612 | |
3613 | ret = sched_getaffinity(pid, &mask); |
3614 | if (ret < 0) |
3615 | return ret; |
3616 | |
3617 | if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t))) |
3618 | return -EFAULT; |
3619 | |
3620 | return sizeof(cpumask_t); |
3621 | } |
3622 | |
3623 | /** |
3624 | * sys_sched_yield - yield the current processor to other threads. |
3625 | * This function yields the current CPU by dropping the priority of current |
3626 | * to the lowest priority and setting the PF_YIELDED flag. |
3627 | */ |
3628 | asmlinkage long sys_sched_yield(void) |
3629 | { |
3630 | int newprio; |
3631 | runqueue_t *rq = this_rq_lock(); |
3632 | |
3633 | newprio = current->prio; |
3634 | schedstat_inc(rq, yld_cnt); |
3635 | current->slice = slice(current); |
3636 | current->time_slice = rr_interval(current); |
3637 | if (likely(!rt_task(current) && !batch_task(current))) { |
3638 | current->flags |= PF_YIELDED; |
3639 | newprio = MAX_PRIO - 2; |
3640 | } |
3641 | |
3642 | if (newprio != current->prio) { |
3643 | dequeue_task(current, rq); |
3644 | current->prio = newprio; |
3645 | enqueue_task(current, rq); |
3646 | } else |
3647 | requeue_task(current, rq); |
3648 | |
3649 | /* |
3650 | * Since we are going to call schedule() anyway, there's |
3651 | * no need to preempt or enable interrupts: |
3652 | */ |
3653 | __release(rq->lock); |
3654 | _raw_spin_unlock(&rq->lock); |
3655 | preempt_enable_no_resched(); |
3656 | |
3657 | schedule(); |
3658 | |
3659 | return 0; |
3660 | } |
3661 | |
3662 | static inline void __cond_resched(void) |
3663 | { |
3664 | do { |
3665 | add_preempt_count(PREEMPT_ACTIVE); |
3666 | schedule(); |
3667 | sub_preempt_count(PREEMPT_ACTIVE); |
3668 | } while (need_resched()); |
3669 | } |
3670 | |
3671 | int __sched cond_resched(void) |
3672 | { |
3673 | if (need_resched()) { |
3674 | __cond_resched(); |
3675 | return 1; |
3676 | } |
3677 | return 0; |
3678 | } |
3679 | |
3680 | EXPORT_SYMBOL(cond_resched); |
3681 | |
3682 | /* |
3683 | * cond_resched_lock() - if a reschedule is pending, drop the given lock, |
3684 | * call schedule, and on return reacquire the lock. |
3685 | * |
3686 | * This works OK both with and without CONFIG_PREEMPT. We do strange low-level |
3687 | * operations here to prevent schedule() from being called twice (once via |
3688 | * spin_unlock(), once by hand). |
3689 | */ |
3690 | int cond_resched_lock(spinlock_t * lock) |
3691 | { |
3692 | int ret = 0; |
3693 | |
3694 | if (need_lockbreak(lock)) { |
3695 | spin_unlock(lock); |
3696 | cpu_relax(); |
3697 | ret = 1; |
3698 | spin_lock(lock); |
3699 | } |
3700 | if (need_resched()) { |
3701 | _raw_spin_unlock(lock); |
3702 | preempt_enable_no_resched(); |
3703 | __cond_resched(); |
3704 | ret = 1; |
3705 | spin_lock(lock); |
3706 | } |
3707 | return ret; |
3708 | } |
3709 | |
3710 | EXPORT_SYMBOL(cond_resched_lock); |
3711 | |
3712 | int __sched cond_resched_softirq(void) |
3713 | { |
3714 | BUG_ON(!in_softirq()); |
3715 | |
3716 | if (need_resched()) { |
3717 | __local_bh_enable(); |
3718 | __cond_resched(); |
3719 | local_bh_disable(); |
3720 | return 1; |
3721 | } |
3722 | return 0; |
3723 | } |
3724 | |
3725 | EXPORT_SYMBOL(cond_resched_softirq); |
3726 | |
3727 | |
3728 | /** |
3729 | * yield - yield the current processor to other threads. |
3730 | * |
3731 | * this is a shortcut for kernel-space yielding - it marks the |
3732 | * thread runnable and calls sys_sched_yield(). |
3733 | */ |
3734 | void __sched yield(void) |
3735 | { |
3736 | set_current_state(TASK_RUNNING); |
3737 | sys_sched_yield(); |
3738 | } |
3739 | |
3740 | EXPORT_SYMBOL(yield); |
3741 | |
3742 | /* |
3743 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so |
3744 | * that process accounting knows that this is a task in IO wait state. |
3745 | * |
3746 | * But don't do that if it is a deliberate, throttling IO wait (this task |
3747 | * has set its backing_dev_info: the queue against which it should throttle) |
3748 | */ |
3749 | void __sched io_schedule(void) |
3750 | { |
3751 | struct runqueue *rq = &per_cpu(runqueues, _smp_processor_id()); |
3752 | |
3753 | atomic_inc(&rq->nr_iowait); |
3754 | schedule(); |
3755 | atomic_dec(&rq->nr_iowait); |
3756 | } |
3757 | |
3758 | EXPORT_SYMBOL(io_schedule); |
3759 | |
3760 | long __sched io_schedule_timeout(long timeout) |
3761 | { |
3762 | struct runqueue *rq = &per_cpu(runqueues, _smp_processor_id()); |
3763 | long ret; |
3764 | |
3765 | atomic_inc(&rq->nr_iowait); |
3766 | ret = schedule_timeout(timeout); |
3767 | atomic_dec(&rq->nr_iowait); |
3768 | return ret; |
3769 | } |
3770 | |
3771 | /** |
3772 | * sys_sched_get_priority_max - return maximum RT priority. |
3773 | * @policy: scheduling class. |
3774 | * |
3775 | * this syscall returns the maximum rt_priority that can be used |
3776 | * by a given scheduling class. |
3777 | */ |
3778 | asmlinkage long sys_sched_get_priority_max(int policy) |
3779 | { |
3780 | int ret = -EINVAL; |
3781 | |
3782 | switch (policy) { |
3783 | case SCHED_FIFO: |
3784 | case SCHED_RR: |
3785 | ret = MAX_USER_RT_PRIO-1; |
3786 | break; |
3787 | case SCHED_NORMAL: |
3788 | case SCHED_BATCH: |
3789 | ret = 0; |
3790 | break; |
3791 | } |
3792 | return ret; |
3793 | } |
3794 | |
3795 | /** |
3796 | * sys_sched_get_priority_min - return minimum RT priority. |
3797 | * @policy: scheduling class. |
3798 | * |
3799 | * this syscall returns the minimum rt_priority that can be used |
3800 | * by a given scheduling class. |
3801 | */ |
3802 | asmlinkage long sys_sched_get_priority_min(int policy) |
3803 | { |
3804 | int ret = -EINVAL; |
3805 | |
3806 | switch (policy) { |
3807 | case SCHED_FIFO: |
3808 | case SCHED_RR: |
3809 | ret = 1; |
3810 | break; |
3811 | case SCHED_NORMAL: |
3812 | case SCHED_BATCH: |
3813 | ret = 0; |
3814 | } |
3815 | return ret; |
3816 | } |
3817 | |
3818 | /** |
3819 | * sys_sched_rr_get_interval - return the default timeslice of a process. |
3820 | * @pid: pid of the process. |
3821 | * @interval: userspace pointer to the timeslice value. |
3822 | * |
3823 | * this syscall writes the default timeslice value of a given process |
3824 | * into the user-space timespec buffer. A value of '0' means infinity. |
3825 | */ |
3826 | asmlinkage |
3827 | long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval) |
3828 | { |
3829 | int retval = -EINVAL; |
3830 | struct timespec t; |
3831 | task_t *p; |
3832 | |
3833 | if (pid < 0) |
3834 | goto out_nounlock; |
3835 | |
3836 | retval = -ESRCH; |
3837 | read_lock(&tasklist_lock); |
3838 | p = find_process_by_pid(pid); |
3839 | if (!p) |
3840 | goto out_unlock; |
3841 | |
3842 | retval = security_task_getscheduler(p); |
3843 | if (retval) |
3844 | goto out_unlock; |
3845 | |
3846 | jiffies_to_timespec(p->policy & SCHED_FIFO ? |
3847 | 0 : slice(p), &t); |
3848 | read_unlock(&tasklist_lock); |
3849 | retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; |
3850 | out_nounlock: |
3851 | return retval; |
3852 | out_unlock: |
3853 | read_unlock(&tasklist_lock); |
3854 | return retval; |
3855 | } |
3856 | |
3857 | static inline struct task_struct *eldest_child(struct task_struct *p) |
3858 | { |
3859 | if (list_empty(&p->children)) return NULL; |
3860 | return list_entry(p->children.next,struct task_struct,sibling); |
3861 | } |
3862 | |
3863 | static inline struct task_struct *older_sibling(struct task_struct *p) |
3864 | { |
3865 | if (p->sibling.prev==&p->parent->children) return NULL; |
3866 | return list_entry(p->sibling.prev,struct task_struct,sibling); |
3867 | } |
3868 | |
3869 | static inline struct task_struct *younger_sibling(struct task_struct *p) |
3870 | { |
3871 | if (p->sibling.next==&p->parent->children) return NULL; |
3872 | return list_entry(p->sibling.next,struct task_struct,sibling); |
3873 | } |
3874 | |
3875 | static inline void show_task(task_t * p) |
3876 | { |
3877 | task_t *relative; |
3878 | unsigned state; |
3879 | unsigned long free = 0; |
3880 | static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" }; |
3881 | |
3882 | printk("%-13.13s ", p->comm); |
3883 | state = p->state ? __ffs(p->state) + 1 : 0; |
3884 | if (state < ARRAY_SIZE(stat_nam)) |
3885 | printk(stat_nam[state]); |
3886 | else |
3887 | printk("?"); |
3888 | #if (BITS_PER_LONG == 32) |
3889 | if (state == TASK_RUNNING) |
3890 | printk(" running "); |
3891 | else |
3892 | printk(" %08lX ", thread_saved_pc(p)); |
3893 | #else |
3894 | if (state == TASK_RUNNING) |
3895 | printk(" running task "); |
3896 | else |
3897 | printk(" %016lx ", thread_saved_pc(p)); |
3898 | #endif |
3899 | #ifdef CONFIG_DEBUG_STACK_USAGE |
3900 | { |
3901 | unsigned long * n = (unsigned long *) (p->thread_info+1); |
3902 | while (!*n) |
3903 | n++; |
3904 | free = (unsigned long) n - (unsigned long)(p->thread_info+1); |
3905 | } |
3906 | #endif |
3907 | printk("%5lu %5d %6d ", free, p->pid, p->parent->pid); |
3908 | if ((relative = eldest_child(p))) |
3909 | printk("%5d ", relative->pid); |
3910 | else |
3911 | printk(" "); |
3912 | if ((relative = younger_sibling(p))) |
3913 | printk("%7d", relative->pid); |
3914 | else |
3915 | printk(" "); |
3916 | if ((relative = older_sibling(p))) |
3917 | printk(" %5d", relative->pid); |
3918 | else |
3919 | printk(" "); |
3920 | if (!p->mm) |
3921 | printk(" (L-TLB)\n"); |
3922 | else |
3923 | printk(" (NOTLB)\n"); |
3924 | |
3925 | if (state != TASK_RUNNING) |
3926 | show_stack(p, NULL); |
3927 | } |
3928 | |
3929 | void show_state(void) |
3930 | { |
3931 | task_t *g, *p; |
3932 | |
3933 | #if (BITS_PER_LONG == 32) |
3934 | printk("\n" |
3935 | " sibling\n"); |
3936 | printk(" task PC pid father child younger older\n"); |
3937 | #else |
3938 | printk("\n" |
3939 | " sibling\n"); |
3940 | printk(" task PC pid father child younger older\n"); |
3941 | #endif |
3942 | read_lock(&tasklist_lock); |
3943 | do_each_thread(g, p) { |
3944 | /* |
3945 | * reset the NMI-timeout, listing all files on a slow |
3946 | * console might take alot of time: |
3947 | */ |
3948 | touch_nmi_watchdog(); |
3949 | show_task(p); |
3950 | } while_each_thread(g, p); |
3951 | |
3952 | read_unlock(&tasklist_lock); |
3953 | } |
3954 | |
3955 | void __devinit init_idle(task_t *idle, int cpu) |
3956 | { |
3957 | runqueue_t *rq = cpu_rq(cpu); |
3958 | unsigned long flags; |
3959 | |
3960 | idle->prio = MAX_PRIO; |
3961 | idle->state = TASK_RUNNING; |
3962 | idle->cpus_allowed = cpumask_of_cpu(cpu); |
3963 | set_task_cpu(idle, cpu); |
3964 | |
3965 | spin_lock_irqsave(&rq->lock, flags); |
3966 | rq->curr = rq->idle = idle; |
3967 | set_tsk_need_resched(idle); |
3968 | spin_unlock_irqrestore(&rq->lock, flags); |
3969 | |
3970 | /* Set the preempt count _outside_ the spinlocks! */ |
3971 | #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL) |
3972 | idle->thread_info->preempt_count = (idle->lock_depth >= 0); |
3973 | #else |
3974 | idle->thread_info->preempt_count = 0; |
3975 | #endif |
3976 | } |
3977 | |
3978 | /* |
3979 | * In a system that switches off the HZ timer nohz_cpu_mask |
3980 | * indicates which cpus entered this state. This is used |
3981 | * in the rcu update to wait only for active cpus. For system |
3982 | * which do not switch off the HZ timer nohz_cpu_mask should |
3983 | * always be CPU_MASK_NONE. |
3984 | */ |
3985 | cpumask_t nohz_cpu_mask = CPU_MASK_NONE; |
3986 | |
3987 | #ifdef CONFIG_SMP |
3988 | /* |
3989 | * This is how migration works: |
3990 | * |
3991 | * 1) we queue a migration_req_t structure in the source CPU's |
3992 | * runqueue and wake up that CPU's migration thread. |
3993 | * 2) we down() the locked semaphore => thread blocks. |
3994 | * 3) migration thread wakes up (implicitly it forces the migrated |
3995 | * thread off the CPU) |
3996 | * 4) it gets the migration request and checks whether the migrated |
3997 | * task is still in the wrong runqueue. |
3998 | * 5) if it's in the wrong runqueue then the migration thread removes |
3999 | * it and puts it into the right queue. |
4000 | * 6) migration thread up()s the semaphore. |
4001 | * 7) we wake up and the migration is done. |
4002 | */ |
4003 | |
4004 | /* |
4005 | * Change a given task's CPU affinity. Migrate the thread to a |
4006 | * proper CPU and schedule it away if the CPU it's executing on |
4007 | * is removed from the allowed bitmask. |
4008 | * |
4009 | * NOTE: the caller must have a valid reference to the task, the |
4010 | * task must not exit() & deallocate itself prematurely. The |
4011 | * call is not atomic; no spinlocks may be held. |
4012 | */ |
4013 | int set_cpus_allowed(task_t *p, cpumask_t new_mask) |
4014 | { |
4015 | unsigned long flags; |
4016 | int ret = 0; |
4017 | migration_req_t req; |
4018 | runqueue_t *rq; |
4019 | |
4020 | rq = task_rq_lock(p, &flags); |
4021 | if (!cpus_intersects(new_mask, cpu_online_map)) { |
4022 | ret = -EINVAL; |
4023 | goto out; |
4024 | } |
4025 | |
4026 | p->cpus_allowed = new_mask; |
4027 | /* Can the task run on the task's current CPU? If so, we're done */ |
4028 | if (cpu_isset(task_cpu(p), new_mask)) |
4029 | goto out; |
4030 | |
4031 | if (migrate_task(p, any_online_cpu(new_mask), &req)) { |
4032 | /* Need help from migration thread: drop lock and wait. */ |
4033 | task_rq_unlock(rq, &flags); |
4034 | wake_up_process(rq->migration_thread); |
4035 | wait_for_completion(&req.done); |
4036 | tlb_migrate_finish(p->mm); |
4037 | return 0; |
4038 | } |
4039 | out: |
4040 | task_rq_unlock(rq, &flags); |
4041 | return ret; |
4042 | } |
4043 | |
4044 | EXPORT_SYMBOL_GPL(set_cpus_allowed); |
4045 | |
4046 | /* |
4047 | * Move (not current) task off this cpu, onto dest cpu. We're doing |
4048 | * this because either it can't run here any more (set_cpus_allowed() |
4049 | * away from this CPU, or CPU going down), or because we're |
4050 | * attempting to rebalance this task on exec (sched_exec). |
4051 | * |
4052 | * So we race with normal scheduler movements, but that's OK, as long |
4053 | * as the task is no longer on this CPU. |
4054 | */ |
4055 | static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) |
4056 | { |
4057 | runqueue_t *rq_dest, *rq_src; |
4058 | |
4059 | if (unlikely(cpu_is_offline(dest_cpu))) |
4060 | return; |
4061 | |
4062 | rq_src = cpu_rq(src_cpu); |
4063 | rq_dest = cpu_rq(dest_cpu); |
4064 | |
4065 | double_rq_lock(rq_src, rq_dest); |
4066 | /* Already moved. */ |
4067 | if (task_cpu(p) != src_cpu) |
4068 | goto out; |
4069 | /* Affinity changed (again). */ |
4070 | if (!cpu_isset(dest_cpu, p->cpus_allowed)) |
4071 | goto out; |
4072 | |
4073 | set_task_cpu(p, dest_cpu); |
4074 | if (task_queued(p)) { |
4075 | /* |
4076 | * Sync timestamp with rq_dest's before activating. |
4077 | * The same thing could be achieved by doing this step |
4078 | * afterwards, and pretending it was a local activate. |
4079 | * This way is cleaner and logically correct. |
4080 | */ |
4081 | p->timestamp = p->timestamp - rq_src->timestamp_last_tick |
4082 | + rq_dest->timestamp_last_tick; |
4083 | deactivate_task(p, rq_src); |
4084 | activate_task(p, rq_dest, 0); |
4085 | preempt(p, rq_dest); |
4086 | } |
4087 | |
4088 | out: |
4089 | double_rq_unlock(rq_src, rq_dest); |
4090 | } |
4091 | |
4092 | /* |
4093 | * migration_thread - this is a highprio system thread that performs |
4094 | * thread migration by bumping thread off CPU then 'pushing' onto |
4095 | * another runqueue. |
4096 | */ |
4097 | static int migration_thread(void * data) |
4098 | { |
4099 | runqueue_t *rq; |
4100 | int cpu = (long)data; |
4101 | |
4102 | rq = cpu_rq(cpu); |
4103 | BUG_ON(rq->migration_thread != current); |
4104 | |
4105 | set_current_state(TASK_INTERRUPTIBLE); |
4106 | while (!kthread_should_stop()) { |
4107 | struct list_head *head; |
4108 | migration_req_t *req; |
4109 | |
4110 | if (current->flags & PF_FREEZE) |
4111 | refrigerator(PF_FREEZE); |
4112 | |
4113 | spin_lock_irq(&rq->lock); |
4114 | |
4115 | if (cpu_is_offline(cpu)) { |
4116 | spin_unlock_irq(&rq->lock); |
4117 | goto wait_to_die; |
4118 | } |
4119 | |
4120 | if (rq->active_balance) { |
4121 | active_load_balance(rq, cpu); |
4122 | rq->active_balance = 0; |
4123 | } |
4124 | |
4125 | head = &rq->migration_queue; |
4126 | |
4127 | if (list_empty(head)) { |
4128 | spin_unlock_irq(&rq->lock); |
4129 | schedule(); |
4130 | set_current_state(TASK_INTERRUPTIBLE); |
4131 | continue; |
4132 | } |
4133 | req = list_entry(head->next, migration_req_t, list); |
4134 | list_del_init(head->next); |
4135 | |
4136 | if (req->type == REQ_MOVE_TASK) { |
4137 | spin_unlock(&rq->lock); |
4138 | __migrate_task(req->task, cpu, req->dest_cpu); |
4139 | local_irq_enable(); |
4140 | } else if (req->type == REQ_SET_DOMAIN) { |
4141 | rq->sd = req->sd; |
4142 | spin_unlock_irq(&rq->lock); |
4143 | } else { |
4144 | spin_unlock_irq(&rq->lock); |
4145 | WARN_ON(1); |
4146 | } |
4147 | |
4148 | complete(&req->done); |
4149 | } |
4150 | __set_current_state(TASK_RUNNING); |
4151 | return 0; |
4152 | |
4153 | wait_to_die: |
4154 | /* Wait for kthread_stop */ |
4155 | set_current_state(TASK_INTERRUPTIBLE); |
4156 | while (!kthread_should_stop()) { |
4157 | schedule(); |
4158 | set_current_state(TASK_INTERRUPTIBLE); |
4159 | } |
4160 | __set_current_state(TASK_RUNNING); |
4161 | return 0; |
4162 | } |
4163 | |
4164 | #ifdef CONFIG_HOTPLUG_CPU |
4165 | /* Figure out where task on dead CPU should go, use force if neccessary. */ |
4166 | static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk) |
4167 | { |
4168 | int dest_cpu; |
4169 | cpumask_t mask; |
4170 | |
4171 | /* On same node? */ |
4172 | mask = node_to_cpumask(cpu_to_node(dead_cpu)); |
4173 | cpus_and(mask, mask, tsk->cpus_allowed); |
4174 | dest_cpu = any_online_cpu(mask); |
4175 | |
4176 | /* On any allowed CPU? */ |
4177 | if (dest_cpu == NR_CPUS) |
4178 | dest_cpu = any_online_cpu(tsk->cpus_allowed); |
4179 | |
4180 | /* No more Mr. Nice Guy. */ |
4181 | if (dest_cpu == NR_CPUS) { |
4182 | cpus_setall(tsk->cpus_allowed); |
4183 | dest_cpu = any_online_cpu(tsk->cpus_allowed); |
4184 | |
4185 | /* |
4186 | * Don't tell them about moving exiting tasks or |
4187 | * kernel threads (both mm NULL), since they never |
4188 | * leave kernel. |
4189 | */ |
4190 | if (tsk->mm && printk_ratelimit()) |
4191 | printk(KERN_INFO "process %d (%s) no " |
4192 | "longer affine to cpu%d\n", |
4193 | tsk->pid, tsk->comm, dead_cpu); |
4194 | } |
4195 | __migrate_task(tsk, dead_cpu, dest_cpu); |
4196 | } |
4197 | |
4198 | /* |
4199 | * While a dead CPU has no uninterruptible tasks queued at this point, |
4200 | * it might still have a nonzero ->nr_uninterruptible counter, because |
4201 | * for performance reasons the counter is not stricly tracking tasks to |
4202 | * their home CPUs. So we just add the counter to another CPU's counter, |
4203 | * to keep the global sum constant after CPU-down: |
4204 | */ |
4205 | static void migrate_nr_uninterruptible(runqueue_t *rq_src) |
4206 | { |
4207 | runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL)); |
4208 | unsigned long flags; |
4209 | |
4210 | local_irq_save(flags); |
4211 | double_rq_lock(rq_src, rq_dest); |
4212 | rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible; |
4213 | rq_src->nr_uninterruptible = 0; |
4214 | double_rq_unlock(rq_src, rq_dest); |
4215 | local_irq_restore(flags); |
4216 | } |
4217 | |
4218 | /* Run through task list and migrate tasks from the dead cpu. */ |
4219 | static void migrate_live_tasks(int src_cpu) |
4220 | { |
4221 | struct task_struct *tsk, *t; |
4222 | |
4223 | write_lock_irq(&tasklist_lock); |
4224 | |
4225 | do_each_thread(t, tsk) { |
4226 | if (tsk == current) |
4227 | continue; |
4228 | |
4229 | if (task_cpu(tsk) == src_cpu) |
4230 | move_task_off_dead_cpu(src_cpu, tsk); |
4231 | } while_each_thread(t, tsk); |
4232 | |
4233 | write_unlock_irq(&tasklist_lock); |
4234 | } |
4235 | |
4236 | /* Schedules idle task to be the next runnable task on current CPU. |
4237 | * It does so by boosting its priority to highest possible and adding it to |
4238 | * the _front_ of runqueue. Used by CPU offline code. |
4239 | */ |
4240 | void sched_idle_next(void) |
4241 | { |
4242 | int cpu = smp_processor_id(); |
4243 | runqueue_t *rq = this_rq(); |
4244 | struct task_struct *p = rq->idle; |
4245 | unsigned long flags; |
4246 | |
4247 | /* cpu has to be offline */ |
4248 | BUG_ON(cpu_online(cpu)); |
4249 | |
4250 | /* Strictly not necessary since rest of the CPUs are stopped by now |
4251 | * and interrupts disabled on current cpu. |
4252 | */ |
4253 | spin_lock_irqsave(&rq->lock, flags); |
4254 | |
4255 | __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1); |
4256 | /* Add idle task to _front_ of it's priority queue */ |
4257 | __activate_idle_task(p, rq); |
4258 | |
4259 | spin_unlock_irqrestore(&rq->lock, flags); |
4260 | } |
4261 | |
4262 | /* Ensures that the idle task is using init_mm right before its cpu goes |
4263 | * offline. |
4264 | */ |
4265 | void idle_task_exit(void) |
4266 | { |
4267 | struct mm_struct *mm = current->active_mm; |
4268 | |
4269 | BUG_ON(cpu_online(smp_processor_id())); |
4270 | |
4271 | if (mm != &init_mm) |
4272 | switch_mm(mm, &init_mm, current); |
4273 | mmdrop(mm); |
4274 | } |
4275 | |
4276 | static void migrate_dead(unsigned int dead_cpu, task_t *tsk) |
4277 | { |
4278 | struct runqueue *rq = cpu_rq(dead_cpu); |
4279 | |
4280 | /* Must be exiting, otherwise would be on tasklist. */ |
4281 | BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD); |
4282 | |
4283 | /* Cannot have done final schedule yet: would have vanished. */ |
4284 | BUG_ON(tsk->flags & PF_DEAD); |
4285 | |
4286 | get_task_struct(tsk); |
4287 | |
4288 | /* |
4289 | * Drop lock around migration; if someone else moves it, |
4290 | * that's OK. No task can be added to this CPU, so iteration is |
4291 | * fine. |
4292 | */ |
4293 | spin_unlock_irq(&rq->lock); |
4294 | move_task_off_dead_cpu(dead_cpu, tsk); |
4295 | spin_lock_irq(&rq->lock); |
4296 | |
4297 | put_task_struct(tsk); |
4298 | } |
4299 | |
4300 | /* release_task() removes task from tasklist, so we won't find dead tasks. */ |
4301 | static void migrate_dead_tasks(unsigned int dead_cpu) |
4302 | { |
4303 | unsigned arr, i; |
4304 | struct runqueue *rq = cpu_rq(dead_cpu); |
4305 | |
4306 | for (arr = 0; arr < 2; arr++) { |
4307 | for (i = 0; i < MAX_PRIO; i++) { |
4308 | struct list_head *list = &rq->queue[i]; |
4309 | while (!list_empty(list)) |
4310 | migrate_dead(dead_cpu, |
4311 | list_entry(list->next, task_t, |
4312 | run_list)); |
4313 | } |
4314 | } |
4315 | } |
4316 | #endif /* CONFIG_HOTPLUG_CPU */ |
4317 | |
4318 | /* |
4319 | * migration_call - callback that gets triggered when a CPU is added. |
4320 | * Here we can start up the necessary migration thread for the new CPU. |
4321 | */ |
4322 | static int migration_call(struct notifier_block *nfb, unsigned long action, |
4323 | void *hcpu) |
4324 | { |
4325 | int cpu = (long)hcpu; |
4326 | struct task_struct *p; |
4327 | struct runqueue *rq; |
4328 | unsigned long flags; |
4329 | |
4330 | switch (action) { |
4331 | case CPU_UP_PREPARE: |
4332 | p = kthread_create(migration_thread, hcpu, "migration/%d",cpu); |
4333 | if (IS_ERR(p)) |
4334 | return NOTIFY_BAD; |
4335 | p->flags |= PF_NOFREEZE; |
4336 | kthread_bind(p, cpu); |
4337 | /* Must be high prio: stop_machine expects to yield to it. */ |
4338 | rq = task_rq_lock(p, &flags); |
4339 | __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1); |
4340 | task_rq_unlock(rq, &flags); |
4341 | cpu_rq(cpu)->migration_thread = p; |
4342 | break; |
4343 | case CPU_ONLINE: |
4344 | /* Strictly unneccessary, as first user will wake it. */ |
4345 | wake_up_process(cpu_rq(cpu)->migration_thread); |
4346 | break; |
4347 | #ifdef CONFIG_HOTPLUG_CPU |
4348 | case CPU_UP_CANCELED: |
4349 | /* Unbind it from offline cpu so it can run. Fall thru. */ |
4350 | kthread_bind(cpu_rq(cpu)->migration_thread,smp_processor_id()); |
4351 | kthread_stop(cpu_rq(cpu)->migration_thread); |
4352 | cpu_rq(cpu)->migration_thread = NULL; |
4353 | break; |
4354 | case CPU_DEAD: |
4355 | migrate_live_tasks(cpu); |
4356 | rq = cpu_rq(cpu); |
4357 | kthread_stop(rq->migration_thread); |
4358 | rq->migration_thread = NULL; |
4359 | /* Idle task back to normal (off runqueue, low prio) */ |
4360 | rq = task_rq_lock(rq->idle, &flags); |
4361 | deactivate_task(rq->idle, rq); |
4362 | rq->idle->static_prio = MAX_PRIO; |
4363 | __setscheduler(rq->idle, SCHED_NORMAL, 0); |
4364 | migrate_dead_tasks(cpu); |
4365 | task_rq_unlock(rq, &flags); |
4366 | migrate_nr_uninterruptible(rq); |
4367 | BUG_ON(rq->nr_running != 0); |
4368 | |
4369 | /* No need to migrate the tasks: it was best-effort if |
4370 | * they didn't do lock_cpu_hotplug(). Just wake up |
4371 | * the requestors. */ |
4372 | spin_lock_irq(&rq->lock); |
4373 | while (!list_empty(&rq->migration_queue)) { |
4374 | migration_req_t *req; |
4375 | req = list_entry(rq->migration_queue.next, |
4376 | migration_req_t, list); |
4377 | BUG_ON(req->type != REQ_MOVE_TASK); |
4378 | list_del_init(&req->list); |
4379 | complete(&req->done); |
4380 | } |
4381 | spin_unlock_irq(&rq->lock); |
4382 | break; |
4383 | #endif |
4384 | } |
4385 | return NOTIFY_OK; |
4386 | } |
4387 | |
4388 | /* Register at highest priority so that task migration (migrate_all_tasks) |
4389 | * happens before everything else. |
4390 | */ |
4391 | static struct notifier_block __devinitdata migration_notifier = { |
4392 | .notifier_call = migration_call, |
4393 | .priority = 10 |
4394 | }; |
4395 | |
4396 | int __init migration_init(void) |
4397 | { |
4398 | void *cpu = (void *)(long)smp_processor_id(); |
4399 | /* Start one for boot CPU. */ |
4400 | migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); |
4401 | migration_call(&migration_notifier, CPU_ONLINE, cpu); |
4402 | register_cpu_notifier(&migration_notifier); |
4403 | return 0; |
4404 | } |
4405 | #endif |
4406 | |
4407 | #ifdef CONFIG_SMP |
4408 | #define SCHED_DOMAIN_DEBUG |
4409 | #ifdef SCHED_DOMAIN_DEBUG |
4410 | static void sched_domain_debug(struct sched_domain *sd, int cpu) |
4411 | { |
4412 | int level = 0; |
4413 | |
4414 | printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); |
4415 | |
4416 | do { |
4417 | int i; |
4418 | char str[NR_CPUS]; |
4419 | struct sched_group *group = sd->groups; |
4420 | cpumask_t groupmask; |
4421 | |
4422 | cpumask_scnprintf(str, NR_CPUS, sd->span); |
4423 | cpus_clear(groupmask); |
4424 | |
4425 | printk(KERN_DEBUG); |
4426 | for (i = 0; i < level + 1; i++) |
4427 | printk(" "); |
4428 | printk("domain %d: ", level); |
4429 | |
4430 | if (!(sd->flags & SD_LOAD_BALANCE)) { |
4431 | printk("does not load-balance\n"); |
4432 | if (sd->parent) |
4433 | printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent"); |
4434 | break; |
4435 | } |
4436 | |
4437 | printk("span %s\n", str); |
4438 | |
4439 | if (!cpu_isset(cpu, sd->span)) |
4440 | printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu); |
4441 | if (!cpu_isset(cpu, group->cpumask)) |
4442 | printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu); |
4443 | |
4444 | printk(KERN_DEBUG); |
4445 | for (i = 0; i < level + 2; i++) |
4446 | printk(" "); |
4447 | printk("groups:"); |
4448 | do { |
4449 | if (!group) { |
4450 | printk("\n"); |
4451 | printk(KERN_ERR "ERROR: group is NULL\n"); |
4452 | break; |
4453 | } |
4454 | |
4455 | if (!group->cpu_power) { |
4456 | printk("\n"); |
4457 | printk(KERN_ERR "ERROR: domain->cpu_power not set\n"); |
4458 | } |
4459 | |
4460 | if (!cpus_weight(group->cpumask)) { |
4461 | printk("\n"); |
4462 | printk(KERN_ERR "ERROR: empty group\n"); |
4463 | } |
4464 | |
4465 | if (cpus_intersects(groupmask, group->cpumask)) { |
4466 | printk("\n"); |
4467 | printk(KERN_ERR "ERROR: repeated CPUs\n"); |
4468 | } |
4469 | |
4470 | cpus_or(groupmask, groupmask, group->cpumask); |
4471 | |
4472 | cpumask_scnprintf(str, NR_CPUS, group->cpumask); |
4473 | printk(" %s", str); |
4474 | |
4475 | group = group->next; |
4476 | } while (group != sd->groups); |
4477 | printk("\n"); |
4478 | |
4479 | if (!cpus_equal(sd->span, groupmask)) |
4480 | printk(KERN_ERR "ERROR: groups don't span domain->span\n"); |
4481 | |
4482 | level++; |
4483 | sd = sd->parent; |
4484 | |
4485 | if (sd) { |
4486 | if (!cpus_subset(groupmask, sd->span)) |
4487 | printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n"); |
4488 | } |
4489 | |
4490 | } while (sd); |
4491 | } |
4492 | #else |
4493 | #define sched_domain_debug(sd, cpu) {} |
4494 | #endif |
4495 | |
4496 | /* |
4497 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must |
4498 | * hold the hotplug lock. |
4499 | */ |
4500 | void __devinit cpu_attach_domain(struct sched_domain *sd, int cpu) |
4501 | { |
4502 | migration_req_t req; |
4503 | unsigned long flags; |
4504 | runqueue_t *rq = cpu_rq(cpu); |
4505 | int local = 1; |
4506 | |
4507 | sched_domain_debug(sd, cpu); |
4508 | |
4509 | spin_lock_irqsave(&rq->lock, flags); |
4510 | |
4511 | if (cpu == smp_processor_id() || !cpu_online(cpu)) { |
4512 | rq->sd = sd; |
4513 | } else { |
4514 | init_completion(&req.done); |
4515 | req.type = REQ_SET_DOMAIN; |
4516 | req.sd = sd; |
4517 | list_add(&req.list, &rq->migration_queue); |
4518 | local = 0; |
4519 | } |
4520 | |
4521 | spin_unlock_irqrestore(&rq->lock, flags); |
4522 | |
4523 | if (!local) { |
4524 | wake_up_process(rq->migration_thread); |
4525 | wait_for_completion(&req.done); |
4526 | } |
4527 | } |
4528 | |
4529 | /* cpus with isolated domains */ |
4530 | cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE; |
4531 | |
4532 | /* Setup the mask of cpus configured for isolated domains */ |
4533 | static int __init isolated_cpu_setup(char *str) |
4534 | { |
4535 | int ints[NR_CPUS], i; |
4536 | |
4537 | str = get_options(str, ARRAY_SIZE(ints), ints); |
4538 | cpus_clear(cpu_isolated_map); |
4539 | for (i = 1; i <= ints[0]; i++) |
4540 | if (ints[i] < NR_CPUS) |
4541 | cpu_set(ints[i], cpu_isolated_map); |
4542 | return 1; |
4543 | } |
4544 | |
4545 | __setup ("isolcpus=", isolated_cpu_setup); |
4546 | |
4547 | /* |
4548 | * init_sched_build_groups takes an array of groups, the cpumask we wish |
4549 | * to span, and a pointer to a function which identifies what group a CPU |
4550 | * belongs to. The return value of group_fn must be a valid index into the |
4551 | * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we |
4552 | * keep track of groups covered with a cpumask_t). |
4553 | * |
4554 | * init_sched_build_groups will build a circular linked list of the groups |
4555 | * covered by the given span, and will set each group's ->cpumask correctly, |
4556 | * and ->cpu_power to 0. |
4557 | */ |
4558 | void __devinit init_sched_build_groups(struct sched_group groups[], |
4559 | cpumask_t span, int (*group_fn)(int cpu)) |
4560 | { |
4561 | struct sched_group *first = NULL, *last = NULL; |
4562 | cpumask_t covered = CPU_MASK_NONE; |
4563 | int i; |
4564 | |
4565 | for_each_cpu_mask(i, span) { |
4566 | int group = group_fn(i); |
4567 | struct sched_group *sg = &groups[group]; |
4568 | int j; |
4569 | |
4570 | if (cpu_isset(i, covered)) |
4571 | continue; |
4572 | |
4573 | sg->cpumask = CPU_MASK_NONE; |
4574 | sg->cpu_power = 0; |
4575 | |
4576 | for_each_cpu_mask(j, span) { |
4577 | if (group_fn(j) != group) |
4578 | continue; |
4579 | |
4580 | cpu_set(j, covered); |
4581 | cpu_set(j, sg->cpumask); |
4582 | } |
4583 | if (!first) |
4584 | first = sg; |
4585 | if (last) |
4586 | last->next = sg; |
4587 | last = sg; |
4588 | } |
4589 | last->next = first; |
4590 | } |
4591 | |
4592 | |
4593 | #ifdef ARCH_HAS_SCHED_DOMAIN |
4594 | extern void __devinit arch_init_sched_domains(void); |
4595 | extern void __devinit arch_destroy_sched_domains(void); |
4596 | #else |
4597 | #ifdef CONFIG_SCHED_SMT |
4598 | static DEFINE_PER_CPU(struct sched_domain, cpu_domains); |
4599 | static struct sched_group sched_group_cpus[NR_CPUS]; |
4600 | static int __devinit cpu_to_cpu_group(int cpu) |
4601 | { |
4602 | return cpu; |
4603 | } |
4604 | #endif |
4605 | |
4606 | static DEFINE_PER_CPU(struct sched_domain, phys_domains); |
4607 | static struct sched_group sched_group_phys[NR_CPUS]; |
4608 | static int __devinit cpu_to_phys_group(int cpu) |
4609 | { |
4610 | #ifdef CONFIG_SCHED_SMT |
4611 | return first_cpu(cpu_sibling_map[cpu]); |
4612 | #else |
4613 | return cpu; |
4614 | #endif |
4615 | } |
4616 | |
4617 | #ifdef CONFIG_NUMA |
4618 | |
4619 | static DEFINE_PER_CPU(struct sched_domain, node_domains); |
4620 | static struct sched_group sched_group_nodes[MAX_NUMNODES]; |
4621 | static int __devinit cpu_to_node_group(int cpu) |
4622 | { |
4623 | return cpu_to_node(cpu); |
4624 | } |
4625 | #endif |
4626 | |
4627 | #if defined(CONFIG_SCHED_SMT) && defined(CONFIG_NUMA) |
4628 | /* |
4629 | * The domains setup code relies on siblings not spanning |
4630 | * multiple nodes. Make sure the architecture has a proper |
4631 | * siblings map: |
4632 | */ |
4633 | static void check_sibling_maps(void) |
4634 | { |
4635 | int i, j; |
4636 | |
4637 | for_each_online_cpu(i) { |
4638 | for_each_cpu_mask(j, cpu_sibling_map[i]) { |
4639 | if (cpu_to_node(i) != cpu_to_node(j)) { |
4640 | printk(KERN_INFO "warning: CPU %d siblings map " |
4641 | "to different node - isolating " |
4642 | "them.\n", i); |
4643 | cpu_sibling_map[i] = cpumask_of_cpu(i); |
4644 | break; |
4645 | } |
4646 | } |
4647 | } |
4648 | } |
4649 | #endif |
4650 | |
4651 | /* |
4652 | * Set up scheduler domains and groups. Callers must hold the hotplug lock. |
4653 | */ |
4654 | static void __devinit arch_init_sched_domains(void) |
4655 | { |
4656 | int i; |
4657 | cpumask_t cpu_default_map; |
4658 | |
4659 | #if defined(CONFIG_SCHED_SMT) && defined(CONFIG_NUMA) |
4660 | check_sibling_maps(); |
4661 | #endif |
4662 | /* |
4663 | * Setup mask for cpus without special case scheduling requirements. |
4664 | * For now this just excludes isolated cpus, but could be used to |
4665 | * exclude other special cases in the future. |
4666 | */ |
4667 | cpus_complement(cpu_default_map, cpu_isolated_map); |
4668 | cpus_and(cpu_default_map, cpu_default_map, cpu_online_map); |
4669 | |
4670 | /* |
4671 | * Set up domains. Isolated domains just stay on the dummy domain. |
4672 | */ |
4673 | for_each_cpu_mask(i, cpu_default_map) { |
4674 | int group; |
4675 | struct sched_domain *sd = NULL, *p; |
4676 | cpumask_t nodemask = node_to_cpumask(cpu_to_node(i)); |
4677 | |
4678 | cpus_and(nodemask, nodemask, cpu_default_map); |
4679 | |
4680 | #ifdef CONFIG_NUMA |
4681 | sd = &per_cpu(node_domains, i); |
4682 | group = cpu_to_node_group(i); |
4683 | *sd = SD_NODE_INIT; |
4684 | sd->span = cpu_default_map; |
4685 | sd->groups = &sched_group_nodes[group]; |
4686 | #endif |
4687 | |
4688 | p = sd; |
4689 | sd = &per_cpu(phys_domains, i); |
4690 | group = cpu_to_phys_group(i); |
4691 | *sd = SD_CPU_INIT; |
4692 | sd->span = nodemask; |
4693 | sd->parent = p; |
4694 | sd->groups = &sched_group_phys[group]; |
4695 | |
4696 | #ifdef CONFIG_SCHED_SMT |
4697 | p = sd; |
4698 | sd = &per_cpu(cpu_domains, i); |
4699 | group = cpu_to_cpu_group(i); |
4700 | *sd = SD_SIBLING_INIT; |
4701 | sd->span = cpu_sibling_map[i]; |
4702 | cpus_and(sd->span, sd->span, cpu_default_map); |
4703 | sd->parent = p; |
4704 | sd->groups = &sched_group_cpus[group]; |
4705 | #endif |
4706 | } |
4707 | |
4708 | #ifdef CONFIG_SCHED_SMT |
4709 | /* Set up CPU (sibling) groups */ |
4710 | for_each_online_cpu(i) { |
4711 | cpumask_t this_sibling_map = cpu_sibling_map[i]; |
4712 | cpus_and(this_sibling_map, this_sibling_map, cpu_default_map); |
4713 | if (i != first_cpu(this_sibling_map)) |
4714 | continue; |
4715 | |
4716 | init_sched_build_groups(sched_group_cpus, this_sibling_map, |
4717 | &cpu_to_cpu_group); |
4718 | } |
4719 | #endif |
4720 | |
4721 | /* Set up physical groups */ |
4722 | for (i = 0; i < MAX_NUMNODES; i++) { |
4723 | cpumask_t nodemask = node_to_cpumask(i); |
4724 | |
4725 | cpus_and(nodemask, nodemask, cpu_default_map); |
4726 | if (cpus_empty(nodemask)) |
4727 | continue; |
4728 | |
4729 | init_sched_build_groups(sched_group_phys, nodemask, |
4730 | &cpu_to_phys_group); |
4731 | } |
4732 | |
4733 | #ifdef CONFIG_NUMA |
4734 | /* Set up node groups */ |
4735 | init_sched_build_groups(sched_group_nodes, cpu_default_map, |
4736 | &cpu_to_node_group); |
4737 | #endif |
4738 | |
4739 | /* Calculate CPU power for physical packages and nodes */ |
4740 | for_each_cpu_mask(i, cpu_default_map) { |
4741 | int power; |
4742 | struct sched_domain *sd; |
4743 | #ifdef CONFIG_SCHED_SMT |
4744 | sd = &per_cpu(cpu_domains, i); |
4745 | power = SCHED_LOAD_SCALE; |
4746 | sd->groups->cpu_power = power; |
4747 | #endif |
4748 | |
4749 | sd = &per_cpu(phys_domains, i); |
4750 | power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE * |
4751 | (cpus_weight(sd->groups->cpumask)-1) / 10; |
4752 | sd->groups->cpu_power = power; |
4753 | |
4754 | #ifdef CONFIG_NUMA |
4755 | if (i == first_cpu(sd->groups->cpumask)) { |
4756 | /* Only add "power" once for each physical package. */ |
4757 | sd = &per_cpu(node_domains, i); |
4758 | sd->groups->cpu_power += power; |
4759 | } |
4760 | #endif |
4761 | } |
4762 | |
4763 | /* Attach the domains */ |
4764 | for_each_online_cpu(i) { |
4765 | struct sched_domain *sd; |
4766 | #ifdef CONFIG_SCHED_SMT |
4767 | sd = &per_cpu(cpu_domains, i); |
4768 | #else |
4769 | sd = &per_cpu(phys_domains, i); |
4770 | #endif |
4771 | cpu_attach_domain(sd, i); |
4772 | } |
4773 | } |
4774 | |
4775 | #ifdef CONFIG_HOTPLUG_CPU |
4776 | static void __devinit arch_destroy_sched_domains(void) |
4777 | { |
4778 | /* Do nothing: everything is statically allocated. */ |
4779 | } |
4780 | #endif |
4781 | |
4782 | #endif /* ARCH_HAS_SCHED_DOMAIN */ |
4783 | |
4784 | /* |
4785 | * Initial dummy domain for early boot and for hotplug cpu. Being static, |
4786 | * it is initialized to zero, so all balancing flags are cleared which is |
4787 | * what we want. |
4788 | */ |
4789 | static struct sched_domain sched_domain_dummy; |
4790 | |
4791 | #ifdef CONFIG_HOTPLUG_CPU |
4792 | /* |
4793 | * Force a reinitialization of the sched domains hierarchy. The domains |
4794 | * and groups cannot be updated in place without racing with the balancing |
4795 | * code, so we temporarily attach all running cpus to a "dummy" domain |
4796 | * which will prevent rebalancing while the sched domains are recalculated. |
4797 | */ |
4798 | static int update_sched_domains(struct notifier_block *nfb, |
4799 | unsigned long action, void *hcpu) |
4800 | { |
4801 | int i; |
4802 | |
4803 | switch (action) { |
4804 | case CPU_UP_PREPARE: |
4805 | case CPU_DOWN_PREPARE: |
4806 | for_each_online_cpu(i) |
4807 | cpu_attach_domain(&sched_domain_dummy, i); |
4808 | arch_destroy_sched_domains(); |
4809 | return NOTIFY_OK; |
4810 | |
4811 | case CPU_UP_CANCELED: |
4812 | case CPU_DOWN_FAILED: |
4813 | case CPU_ONLINE: |
4814 | case CPU_DEAD: |
4815 | /* |
4816 | * Fall through and re-initialise the domains. |
4817 | */ |
4818 | break; |
4819 | default: |
4820 | return NOTIFY_DONE; |
4821 | } |
4822 | |
4823 | /* The hotplug lock is already held by cpu_up/cpu_down */ |
4824 | arch_init_sched_domains(); |
4825 | |
4826 | return NOTIFY_OK; |
4827 | } |
4828 | #endif |
4829 | |
4830 | void __init sched_init_smp(void) |
4831 | { |
4832 | lock_cpu_hotplug(); |
4833 | arch_init_sched_domains(); |
4834 | unlock_cpu_hotplug(); |
4835 | /* XXX: Theoretical race here - CPU may be hotplugged now */ |
4836 | hotcpu_notifier(update_sched_domains, 0); |
4837 | } |
4838 | #else |
4839 | void __init sched_init_smp(void) |
4840 | { |
4841 | } |
4842 | #endif /* CONFIG_SMP */ |
4843 | |
4844 | int in_sched_functions(unsigned long addr) |
4845 | { |
4846 | /* Linker adds these: start and end of __sched functions */ |
4847 | extern char __sched_text_start[], __sched_text_end[]; |
4848 | return in_lock_functions(addr) || |
4849 | (addr >= (unsigned long)__sched_text_start |
4850 | && addr < (unsigned long)__sched_text_end); |
4851 | } |
4852 | |
4853 | void __init sched_init(void) |
4854 | { |
4855 | runqueue_t *rq; |
4856 | int i, j; |
4857 | |
4858 | for (i = 0; i < NR_CPUS; i++) { |
4859 | |
4860 | rq = cpu_rq(i); |
4861 | spin_lock_init(&rq->lock); |
4862 | rq->cache_ticks = 0; |
4863 | rq->preempted = 0; |
4864 | |
4865 | #ifdef CONFIG_SMP |
4866 | rq->sd = &sched_domain_dummy; |
4867 | rq->cpu_load = 0; |
4868 | rq->active_balance = 0; |
4869 | rq->push_cpu = 0; |
4870 | rq->migration_thread = NULL; |
4871 | INIT_LIST_HEAD(&rq->migration_queue); |
4872 | #endif |
4873 | atomic_set(&rq->nr_iowait, 0); |
4874 | for (j = 0; j < MAX_PRIO; j++) |
4875 | INIT_LIST_HEAD(&rq->queue[j]); |
4876 | memset(rq->bitmap, 0, BITS_TO_LONGS(MAX_PRIO)*sizeof(long)); |
4877 | /* |
4878 | * delimiter for bitsearch |
4879 | */ |
4880 | __set_bit(MAX_PRIO, rq->bitmap); |
4881 | } |
4882 | |
4883 | /* |
4884 | * The boot idle thread does lazy MMU switching as well: |
4885 | */ |
4886 | atomic_inc(&init_mm.mm_count); |
4887 | enter_lazy_tlb(&init_mm, current); |
4888 | |
4889 | /* |
4890 | * Make us the idle thread. Technically, schedule() should not be |
4891 | * called from this thread, however somewhere below it might be, |
4892 | * but because we are the idle thread, we just pick up running again |
4893 | * when this runqueue becomes "idle". |
4894 | */ |
4895 | init_idle(current, smp_processor_id()); |
4896 | } |
4897 | |
4898 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP |
4899 | void __might_sleep(char *file, int line) |
4900 | { |
4901 | #if defined(in_atomic) |
4902 | static unsigned long prev_jiffy; /* ratelimiting */ |
4903 | |
4904 | if ((in_atomic() || irqs_disabled()) && |
4905 | system_state == SYSTEM_RUNNING && !oops_in_progress) { |
4906 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) |
4907 | return; |
4908 | prev_jiffy = jiffies; |
4909 | printk(KERN_ERR "Debug: sleeping function called from invalid" |
4910 | " context at %s:%d\n", file, line); |
4911 | printk("in_atomic():%d, irqs_disabled():%d\n", |
4912 | in_atomic(), irqs_disabled()); |
4913 | dump_stack(); |
4914 | } |
4915 | #endif |
4916 | } |
4917 | EXPORT_SYMBOL(__might_sleep); |
4918 | #endif |
4919 | |
4920 | #ifdef CONFIG_MAGIC_SYSRQ |
4921 | void normalize_rt_tasks(void) |
4922 | { |
4923 | struct task_struct *p; |
4924 | unsigned long flags; |
4925 | runqueue_t *rq; |
4926 | int queued; |
4927 | |
4928 | read_lock_irq(&tasklist_lock); |
4929 | for_each_process (p) { |
4930 | if (!rt_task(p)) |
4931 | continue; |
4932 | |
4933 | rq = task_rq_lock(p, &flags); |
4934 | |
4935 | if ((queued = task_queued(p))) |
4936 | deactivate_task(p, task_rq(p)); |
4937 | __setscheduler(p, SCHED_NORMAL, 0); |
4938 | if (queued) { |
4939 | __activate_task(p, task_rq(p)); |
4940 | resched_task(rq->curr); |
4941 | } |
4942 | |
4943 | task_rq_unlock(rq, &flags); |
4944 | } |
4945 | read_unlock_irq(&tasklist_lock); |
4946 | } |
4947 | |
4948 | #endif /* CONFIG_MAGIC_SYSRQ */ |