]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
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 | */ | |
20 | ||
21 | #include <linux/mm.h> | |
22 | #include <linux/module.h> | |
23 | #include <linux/nmi.h> | |
24 | #include <linux/init.h> | |
25 | #include <asm/uaccess.h> | |
26 | #include <linux/highmem.h> | |
27 | #include <linux/smp_lock.h> | |
28 | #include <asm/mmu_context.h> | |
29 | #include <linux/interrupt.h> | |
30 | #include <linux/completion.h> | |
31 | #include <linux/kernel_stat.h> | |
32 | #include <linux/security.h> | |
33 | #include <linux/notifier.h> | |
34 | #include <linux/profile.h> | |
35 | #include <linux/suspend.h> | |
36 | #include <linux/blkdev.h> | |
37 | #include <linux/delay.h> | |
38 | #include <linux/smp.h> | |
39 | #include <linux/threads.h> | |
40 | #include <linux/timer.h> | |
41 | #include <linux/rcupdate.h> | |
42 | #include <linux/cpu.h> | |
43 | #include <linux/cpuset.h> | |
44 | #include <linux/percpu.h> | |
45 | #include <linux/kthread.h> | |
46 | #include <linux/seq_file.h> | |
47 | #include <linux/syscalls.h> | |
48 | #include <linux/times.h> | |
49 | #include <linux/acct.h> | |
50 | #include <asm/tlb.h> | |
51 | ||
52 | #include <asm/unistd.h> | |
53 | ||
54 | /* | |
55 | * Convert user-nice values [ -20 ... 0 ... 19 ] | |
56 | * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ], | |
57 | * and back. | |
58 | */ | |
59 | #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20) | |
60 | #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20) | |
61 | #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio) | |
62 | ||
63 | /* | |
64 | * 'User priority' is the nice value converted to something we | |
65 | * can work with better when scaling various scheduler parameters, | |
66 | * it's a [ 0 ... 39 ] range. | |
67 | */ | |
68 | #define USER_PRIO(p) ((p)-MAX_RT_PRIO) | |
69 | #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio) | |
70 | #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO)) | |
71 | ||
72 | /* | |
73 | * Some helpers for converting nanosecond timing to jiffy resolution | |
74 | */ | |
75 | #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ)) | |
76 | #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ)) | |
77 | ||
78 | /* | |
79 | * These are the 'tuning knobs' of the scheduler: | |
80 | * | |
81 | * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger), | |
82 | * default timeslice is 100 msecs, maximum timeslice is 800 msecs. | |
83 | * Timeslices get refilled after they expire. | |
84 | */ | |
85 | #define MIN_TIMESLICE max(5 * HZ / 1000, 1) | |
86 | #define DEF_TIMESLICE (100 * HZ / 1000) | |
87 | #define ON_RUNQUEUE_WEIGHT 30 | |
88 | #define CHILD_PENALTY 95 | |
89 | #define PARENT_PENALTY 100 | |
90 | #define EXIT_WEIGHT 3 | |
91 | #define PRIO_BONUS_RATIO 25 | |
92 | #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100) | |
93 | #define INTERACTIVE_DELTA 2 | |
94 | #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS) | |
95 | #define STARVATION_LIMIT (MAX_SLEEP_AVG) | |
96 | #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG)) | |
97 | ||
98 | /* | |
99 | * If a task is 'interactive' then we reinsert it in the active | |
100 | * array after it has expired its current timeslice. (it will not | |
101 | * continue to run immediately, it will still roundrobin with | |
102 | * other interactive tasks.) | |
103 | * | |
104 | * This part scales the interactivity limit depending on niceness. | |
105 | * | |
106 | * We scale it linearly, offset by the INTERACTIVE_DELTA delta. | |
107 | * Here are a few examples of different nice levels: | |
108 | * | |
109 | * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0] | |
110 | * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0] | |
111 | * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0] | |
112 | * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0] | |
113 | * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0] | |
114 | * | |
115 | * (the X axis represents the possible -5 ... 0 ... +5 dynamic | |
116 | * priority range a task can explore, a value of '1' means the | |
117 | * task is rated interactive.) | |
118 | * | |
119 | * Ie. nice +19 tasks can never get 'interactive' enough to be | |
120 | * reinserted into the active array. And only heavily CPU-hog nice -20 | |
121 | * tasks will be expired. Default nice 0 tasks are somewhere between, | |
122 | * it takes some effort for them to get interactive, but it's not | |
123 | * too hard. | |
124 | */ | |
125 | ||
126 | #define CURRENT_BONUS(p) \ | |
127 | (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \ | |
128 | MAX_SLEEP_AVG) | |
129 | ||
130 | #define GRANULARITY (10 * HZ / 1000 ? : 1) | |
131 | ||
132 | #ifdef CONFIG_SMP | |
133 | #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \ | |
134 | (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \ | |
135 | num_online_cpus()) | |
136 | #else | |
137 | #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \ | |
138 | (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1))) | |
139 | #endif | |
140 | ||
141 | #define SCALE(v1,v1_max,v2_max) \ | |
142 | (v1) * (v2_max) / (v1_max) | |
143 | ||
144 | #define DELTA(p) \ | |
145 | (SCALE(TASK_NICE(p), 40, MAX_BONUS) + INTERACTIVE_DELTA) | |
146 | ||
147 | #define TASK_INTERACTIVE(p) \ | |
148 | ((p)->prio <= (p)->static_prio - DELTA(p)) | |
149 | ||
150 | #define INTERACTIVE_SLEEP(p) \ | |
151 | (JIFFIES_TO_NS(MAX_SLEEP_AVG * \ | |
152 | (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1)) | |
153 | ||
154 | #define TASK_PREEMPTS_CURR(p, rq) \ | |
155 | ((p)->prio < (rq)->curr->prio) | |
156 | ||
157 | /* | |
158 | * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ] | |
159 | * to time slice values: [800ms ... 100ms ... 5ms] | |
160 | * | |
161 | * The higher a thread's priority, the bigger timeslices | |
162 | * it gets during one round of execution. But even the lowest | |
163 | * priority thread gets MIN_TIMESLICE worth of execution time. | |
164 | */ | |
165 | ||
166 | #define SCALE_PRIO(x, prio) \ | |
167 | max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO/2), MIN_TIMESLICE) | |
168 | ||
48c08d3f | 169 | static unsigned int task_timeslice(task_t *p) |
1da177e4 LT |
170 | { |
171 | if (p->static_prio < NICE_TO_PRIO(0)) | |
172 | return SCALE_PRIO(DEF_TIMESLICE*4, p->static_prio); | |
173 | else | |
174 | return SCALE_PRIO(DEF_TIMESLICE, p->static_prio); | |
175 | } | |
176 | #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran) \ | |
177 | < (long long) (sd)->cache_hot_time) | |
178 | ||
179 | /* | |
180 | * These are the runqueue data structures: | |
181 | */ | |
182 | ||
183 | #define BITMAP_SIZE ((((MAX_PRIO+1+7)/8)+sizeof(long)-1)/sizeof(long)) | |
184 | ||
185 | typedef struct runqueue runqueue_t; | |
186 | ||
187 | struct prio_array { | |
188 | unsigned int nr_active; | |
189 | unsigned long bitmap[BITMAP_SIZE]; | |
190 | struct list_head queue[MAX_PRIO]; | |
191 | }; | |
192 | ||
193 | /* | |
194 | * This is the main, per-CPU runqueue data structure. | |
195 | * | |
196 | * Locking rule: those places that want to lock multiple runqueues | |
197 | * (such as the load balancing or the thread migration code), lock | |
198 | * acquire operations must be ordered by ascending &runqueue. | |
199 | */ | |
200 | struct runqueue { | |
201 | spinlock_t lock; | |
202 | ||
203 | /* | |
204 | * nr_running and cpu_load should be in the same cacheline because | |
205 | * remote CPUs use both these fields when doing load calculation. | |
206 | */ | |
207 | unsigned long nr_running; | |
208 | #ifdef CONFIG_SMP | |
7897986b | 209 | unsigned long cpu_load[3]; |
1da177e4 LT |
210 | #endif |
211 | unsigned long long nr_switches; | |
212 | ||
213 | /* | |
214 | * This is part of a global counter where only the total sum | |
215 | * over all CPUs matters. A task can increase this counter on | |
216 | * one CPU and if it got migrated afterwards it may decrease | |
217 | * it on another CPU. Always updated under the runqueue lock: | |
218 | */ | |
219 | unsigned long nr_uninterruptible; | |
220 | ||
221 | unsigned long expired_timestamp; | |
222 | unsigned long long timestamp_last_tick; | |
223 | task_t *curr, *idle; | |
224 | struct mm_struct *prev_mm; | |
225 | prio_array_t *active, *expired, arrays[2]; | |
226 | int best_expired_prio; | |
227 | atomic_t nr_iowait; | |
228 | ||
229 | #ifdef CONFIG_SMP | |
230 | struct sched_domain *sd; | |
231 | ||
232 | /* For active balancing */ | |
233 | int active_balance; | |
234 | int push_cpu; | |
235 | ||
236 | task_t *migration_thread; | |
237 | struct list_head migration_queue; | |
238 | #endif | |
239 | ||
240 | #ifdef CONFIG_SCHEDSTATS | |
241 | /* latency stats */ | |
242 | struct sched_info rq_sched_info; | |
243 | ||
244 | /* sys_sched_yield() stats */ | |
245 | unsigned long yld_exp_empty; | |
246 | unsigned long yld_act_empty; | |
247 | unsigned long yld_both_empty; | |
248 | unsigned long yld_cnt; | |
249 | ||
250 | /* schedule() stats */ | |
251 | unsigned long sched_switch; | |
252 | unsigned long sched_cnt; | |
253 | unsigned long sched_goidle; | |
254 | ||
255 | /* try_to_wake_up() stats */ | |
256 | unsigned long ttwu_cnt; | |
257 | unsigned long ttwu_local; | |
258 | #endif | |
259 | }; | |
260 | ||
261 | static DEFINE_PER_CPU(struct runqueue, runqueues); | |
262 | ||
674311d5 NP |
263 | /* |
264 | * The domain tree (rq->sd) is protected by RCU's quiescent state transition. | |
1a20ff27 | 265 | * See detach_destroy_domains: synchronize_sched for details. |
674311d5 NP |
266 | * |
267 | * The domain tree of any CPU may only be accessed from within | |
268 | * preempt-disabled sections. | |
269 | */ | |
1da177e4 | 270 | #define for_each_domain(cpu, domain) \ |
674311d5 | 271 | for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent) |
1da177e4 LT |
272 | |
273 | #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) | |
274 | #define this_rq() (&__get_cpu_var(runqueues)) | |
275 | #define task_rq(p) cpu_rq(task_cpu(p)) | |
276 | #define cpu_curr(cpu) (cpu_rq(cpu)->curr) | |
277 | ||
1da177e4 | 278 | #ifndef prepare_arch_switch |
4866cde0 NP |
279 | # define prepare_arch_switch(next) do { } while (0) |
280 | #endif | |
281 | #ifndef finish_arch_switch | |
282 | # define finish_arch_switch(prev) do { } while (0) | |
283 | #endif | |
284 | ||
285 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW | |
286 | static inline int task_running(runqueue_t *rq, task_t *p) | |
287 | { | |
288 | return rq->curr == p; | |
289 | } | |
290 | ||
291 | static inline void prepare_lock_switch(runqueue_t *rq, task_t *next) | |
292 | { | |
293 | } | |
294 | ||
295 | static inline void finish_lock_switch(runqueue_t *rq, task_t *prev) | |
296 | { | |
da04c035 IM |
297 | #ifdef CONFIG_DEBUG_SPINLOCK |
298 | /* this is a valid case when another task releases the spinlock */ | |
299 | rq->lock.owner = current; | |
300 | #endif | |
4866cde0 NP |
301 | spin_unlock_irq(&rq->lock); |
302 | } | |
303 | ||
304 | #else /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
305 | static inline int task_running(runqueue_t *rq, task_t *p) | |
306 | { | |
307 | #ifdef CONFIG_SMP | |
308 | return p->oncpu; | |
309 | #else | |
310 | return rq->curr == p; | |
311 | #endif | |
312 | } | |
313 | ||
314 | static inline void prepare_lock_switch(runqueue_t *rq, task_t *next) | |
315 | { | |
316 | #ifdef CONFIG_SMP | |
317 | /* | |
318 | * We can optimise this out completely for !SMP, because the | |
319 | * SMP rebalancing from interrupt is the only thing that cares | |
320 | * here. | |
321 | */ | |
322 | next->oncpu = 1; | |
323 | #endif | |
324 | #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
325 | spin_unlock_irq(&rq->lock); | |
326 | #else | |
327 | spin_unlock(&rq->lock); | |
328 | #endif | |
329 | } | |
330 | ||
331 | static inline void finish_lock_switch(runqueue_t *rq, task_t *prev) | |
332 | { | |
333 | #ifdef CONFIG_SMP | |
334 | /* | |
335 | * After ->oncpu is cleared, the task can be moved to a different CPU. | |
336 | * We must ensure this doesn't happen until the switch is completely | |
337 | * finished. | |
338 | */ | |
339 | smp_wmb(); | |
340 | prev->oncpu = 0; | |
341 | #endif | |
342 | #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
343 | local_irq_enable(); | |
1da177e4 | 344 | #endif |
4866cde0 NP |
345 | } |
346 | #endif /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
1da177e4 LT |
347 | |
348 | /* | |
349 | * task_rq_lock - lock the runqueue a given task resides on and disable | |
350 | * interrupts. Note the ordering: we can safely lookup the task_rq without | |
351 | * explicitly disabling preemption. | |
352 | */ | |
353 | static inline runqueue_t *task_rq_lock(task_t *p, unsigned long *flags) | |
354 | __acquires(rq->lock) | |
355 | { | |
356 | struct runqueue *rq; | |
357 | ||
358 | repeat_lock_task: | |
359 | local_irq_save(*flags); | |
360 | rq = task_rq(p); | |
361 | spin_lock(&rq->lock); | |
362 | if (unlikely(rq != task_rq(p))) { | |
363 | spin_unlock_irqrestore(&rq->lock, *flags); | |
364 | goto repeat_lock_task; | |
365 | } | |
366 | return rq; | |
367 | } | |
368 | ||
369 | static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags) | |
370 | __releases(rq->lock) | |
371 | { | |
372 | spin_unlock_irqrestore(&rq->lock, *flags); | |
373 | } | |
374 | ||
375 | #ifdef CONFIG_SCHEDSTATS | |
376 | /* | |
377 | * bump this up when changing the output format or the meaning of an existing | |
378 | * format, so that tools can adapt (or abort) | |
379 | */ | |
68767a0a | 380 | #define SCHEDSTAT_VERSION 12 |
1da177e4 LT |
381 | |
382 | static int show_schedstat(struct seq_file *seq, void *v) | |
383 | { | |
384 | int cpu; | |
385 | ||
386 | seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION); | |
387 | seq_printf(seq, "timestamp %lu\n", jiffies); | |
388 | for_each_online_cpu(cpu) { | |
389 | runqueue_t *rq = cpu_rq(cpu); | |
390 | #ifdef CONFIG_SMP | |
391 | struct sched_domain *sd; | |
392 | int dcnt = 0; | |
393 | #endif | |
394 | ||
395 | /* runqueue-specific stats */ | |
396 | seq_printf(seq, | |
397 | "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu", | |
398 | cpu, rq->yld_both_empty, | |
399 | rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt, | |
400 | rq->sched_switch, rq->sched_cnt, rq->sched_goidle, | |
401 | rq->ttwu_cnt, rq->ttwu_local, | |
402 | rq->rq_sched_info.cpu_time, | |
403 | rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt); | |
404 | ||
405 | seq_printf(seq, "\n"); | |
406 | ||
407 | #ifdef CONFIG_SMP | |
408 | /* domain-specific stats */ | |
674311d5 | 409 | preempt_disable(); |
1da177e4 LT |
410 | for_each_domain(cpu, sd) { |
411 | enum idle_type itype; | |
412 | char mask_str[NR_CPUS]; | |
413 | ||
414 | cpumask_scnprintf(mask_str, NR_CPUS, sd->span); | |
415 | seq_printf(seq, "domain%d %s", dcnt++, mask_str); | |
416 | for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES; | |
417 | itype++) { | |
418 | seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu", | |
419 | sd->lb_cnt[itype], | |
420 | sd->lb_balanced[itype], | |
421 | sd->lb_failed[itype], | |
422 | sd->lb_imbalance[itype], | |
423 | sd->lb_gained[itype], | |
424 | sd->lb_hot_gained[itype], | |
425 | sd->lb_nobusyq[itype], | |
426 | sd->lb_nobusyg[itype]); | |
427 | } | |
68767a0a | 428 | seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n", |
1da177e4 | 429 | sd->alb_cnt, sd->alb_failed, sd->alb_pushed, |
68767a0a NP |
430 | sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed, |
431 | sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed, | |
1da177e4 LT |
432 | sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance); |
433 | } | |
674311d5 | 434 | preempt_enable(); |
1da177e4 LT |
435 | #endif |
436 | } | |
437 | return 0; | |
438 | } | |
439 | ||
440 | static int schedstat_open(struct inode *inode, struct file *file) | |
441 | { | |
442 | unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32); | |
443 | char *buf = kmalloc(size, GFP_KERNEL); | |
444 | struct seq_file *m; | |
445 | int res; | |
446 | ||
447 | if (!buf) | |
448 | return -ENOMEM; | |
449 | res = single_open(file, show_schedstat, NULL); | |
450 | if (!res) { | |
451 | m = file->private_data; | |
452 | m->buf = buf; | |
453 | m->size = size; | |
454 | } else | |
455 | kfree(buf); | |
456 | return res; | |
457 | } | |
458 | ||
459 | struct file_operations proc_schedstat_operations = { | |
460 | .open = schedstat_open, | |
461 | .read = seq_read, | |
462 | .llseek = seq_lseek, | |
463 | .release = single_release, | |
464 | }; | |
465 | ||
466 | # define schedstat_inc(rq, field) do { (rq)->field++; } while (0) | |
467 | # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0) | |
468 | #else /* !CONFIG_SCHEDSTATS */ | |
469 | # define schedstat_inc(rq, field) do { } while (0) | |
470 | # define schedstat_add(rq, field, amt) do { } while (0) | |
471 | #endif | |
472 | ||
473 | /* | |
474 | * rq_lock - lock a given runqueue and disable interrupts. | |
475 | */ | |
476 | static inline runqueue_t *this_rq_lock(void) | |
477 | __acquires(rq->lock) | |
478 | { | |
479 | runqueue_t *rq; | |
480 | ||
481 | local_irq_disable(); | |
482 | rq = this_rq(); | |
483 | spin_lock(&rq->lock); | |
484 | ||
485 | return rq; | |
486 | } | |
487 | ||
1da177e4 LT |
488 | #ifdef CONFIG_SCHEDSTATS |
489 | /* | |
490 | * Called when a process is dequeued from the active array and given | |
491 | * the cpu. We should note that with the exception of interactive | |
492 | * tasks, the expired queue will become the active queue after the active | |
493 | * queue is empty, without explicitly dequeuing and requeuing tasks in the | |
494 | * expired queue. (Interactive tasks may be requeued directly to the | |
495 | * active queue, thus delaying tasks in the expired queue from running; | |
496 | * see scheduler_tick()). | |
497 | * | |
498 | * This function is only called from sched_info_arrive(), rather than | |
499 | * dequeue_task(). Even though a task may be queued and dequeued multiple | |
500 | * times as it is shuffled about, we're really interested in knowing how | |
501 | * long it was from the *first* time it was queued to the time that it | |
502 | * finally hit a cpu. | |
503 | */ | |
504 | static inline void sched_info_dequeued(task_t *t) | |
505 | { | |
506 | t->sched_info.last_queued = 0; | |
507 | } | |
508 | ||
509 | /* | |
510 | * Called when a task finally hits the cpu. We can now calculate how | |
511 | * long it was waiting to run. We also note when it began so that we | |
512 | * can keep stats on how long its timeslice is. | |
513 | */ | |
514 | static inline void sched_info_arrive(task_t *t) | |
515 | { | |
516 | unsigned long now = jiffies, diff = 0; | |
517 | struct runqueue *rq = task_rq(t); | |
518 | ||
519 | if (t->sched_info.last_queued) | |
520 | diff = now - t->sched_info.last_queued; | |
521 | sched_info_dequeued(t); | |
522 | t->sched_info.run_delay += diff; | |
523 | t->sched_info.last_arrival = now; | |
524 | t->sched_info.pcnt++; | |
525 | ||
526 | if (!rq) | |
527 | return; | |
528 | ||
529 | rq->rq_sched_info.run_delay += diff; | |
530 | rq->rq_sched_info.pcnt++; | |
531 | } | |
532 | ||
533 | /* | |
534 | * Called when a process is queued into either the active or expired | |
535 | * array. The time is noted and later used to determine how long we | |
536 | * had to wait for us to reach the cpu. Since the expired queue will | |
537 | * become the active queue after active queue is empty, without dequeuing | |
538 | * and requeuing any tasks, we are interested in queuing to either. It | |
539 | * is unusual but not impossible for tasks to be dequeued and immediately | |
540 | * requeued in the same or another array: this can happen in sched_yield(), | |
541 | * set_user_nice(), and even load_balance() as it moves tasks from runqueue | |
542 | * to runqueue. | |
543 | * | |
544 | * This function is only called from enqueue_task(), but also only updates | |
545 | * the timestamp if it is already not set. It's assumed that | |
546 | * sched_info_dequeued() will clear that stamp when appropriate. | |
547 | */ | |
548 | static inline void sched_info_queued(task_t *t) | |
549 | { | |
550 | if (!t->sched_info.last_queued) | |
551 | t->sched_info.last_queued = jiffies; | |
552 | } | |
553 | ||
554 | /* | |
555 | * Called when a process ceases being the active-running process, either | |
556 | * voluntarily or involuntarily. Now we can calculate how long we ran. | |
557 | */ | |
558 | static inline void sched_info_depart(task_t *t) | |
559 | { | |
560 | struct runqueue *rq = task_rq(t); | |
561 | unsigned long diff = jiffies - t->sched_info.last_arrival; | |
562 | ||
563 | t->sched_info.cpu_time += diff; | |
564 | ||
565 | if (rq) | |
566 | rq->rq_sched_info.cpu_time += diff; | |
567 | } | |
568 | ||
569 | /* | |
570 | * Called when tasks are switched involuntarily due, typically, to expiring | |
571 | * their time slice. (This may also be called when switching to or from | |
572 | * the idle task.) We are only called when prev != next. | |
573 | */ | |
574 | static inline void sched_info_switch(task_t *prev, task_t *next) | |
575 | { | |
576 | struct runqueue *rq = task_rq(prev); | |
577 | ||
578 | /* | |
579 | * prev now departs the cpu. It's not interesting to record | |
580 | * stats about how efficient we were at scheduling the idle | |
581 | * process, however. | |
582 | */ | |
583 | if (prev != rq->idle) | |
584 | sched_info_depart(prev); | |
585 | ||
586 | if (next != rq->idle) | |
587 | sched_info_arrive(next); | |
588 | } | |
589 | #else | |
590 | #define sched_info_queued(t) do { } while (0) | |
591 | #define sched_info_switch(t, next) do { } while (0) | |
592 | #endif /* CONFIG_SCHEDSTATS */ | |
593 | ||
594 | /* | |
595 | * Adding/removing a task to/from a priority array: | |
596 | */ | |
597 | static void dequeue_task(struct task_struct *p, prio_array_t *array) | |
598 | { | |
599 | array->nr_active--; | |
600 | list_del(&p->run_list); | |
601 | if (list_empty(array->queue + p->prio)) | |
602 | __clear_bit(p->prio, array->bitmap); | |
603 | } | |
604 | ||
605 | static void enqueue_task(struct task_struct *p, prio_array_t *array) | |
606 | { | |
607 | sched_info_queued(p); | |
608 | list_add_tail(&p->run_list, array->queue + p->prio); | |
609 | __set_bit(p->prio, array->bitmap); | |
610 | array->nr_active++; | |
611 | p->array = array; | |
612 | } | |
613 | ||
614 | /* | |
615 | * Put task to the end of the run list without the overhead of dequeue | |
616 | * followed by enqueue. | |
617 | */ | |
618 | static void requeue_task(struct task_struct *p, prio_array_t *array) | |
619 | { | |
620 | list_move_tail(&p->run_list, array->queue + p->prio); | |
621 | } | |
622 | ||
623 | static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array) | |
624 | { | |
625 | list_add(&p->run_list, array->queue + p->prio); | |
626 | __set_bit(p->prio, array->bitmap); | |
627 | array->nr_active++; | |
628 | p->array = array; | |
629 | } | |
630 | ||
631 | /* | |
632 | * effective_prio - return the priority that is based on the static | |
633 | * priority but is modified by bonuses/penalties. | |
634 | * | |
635 | * We scale the actual sleep average [0 .... MAX_SLEEP_AVG] | |
636 | * into the -5 ... 0 ... +5 bonus/penalty range. | |
637 | * | |
638 | * We use 25% of the full 0...39 priority range so that: | |
639 | * | |
640 | * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs. | |
641 | * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks. | |
642 | * | |
643 | * Both properties are important to certain workloads. | |
644 | */ | |
645 | static int effective_prio(task_t *p) | |
646 | { | |
647 | int bonus, prio; | |
648 | ||
649 | if (rt_task(p)) | |
650 | return p->prio; | |
651 | ||
652 | bonus = CURRENT_BONUS(p) - MAX_BONUS / 2; | |
653 | ||
654 | prio = p->static_prio - bonus; | |
655 | if (prio < MAX_RT_PRIO) | |
656 | prio = MAX_RT_PRIO; | |
657 | if (prio > MAX_PRIO-1) | |
658 | prio = MAX_PRIO-1; | |
659 | return prio; | |
660 | } | |
661 | ||
662 | /* | |
663 | * __activate_task - move a task to the runqueue. | |
664 | */ | |
665 | static inline void __activate_task(task_t *p, runqueue_t *rq) | |
666 | { | |
667 | enqueue_task(p, rq->active); | |
668 | rq->nr_running++; | |
669 | } | |
670 | ||
671 | /* | |
672 | * __activate_idle_task - move idle task to the _front_ of runqueue. | |
673 | */ | |
674 | static inline void __activate_idle_task(task_t *p, runqueue_t *rq) | |
675 | { | |
676 | enqueue_task_head(p, rq->active); | |
677 | rq->nr_running++; | |
678 | } | |
679 | ||
a3464a10 | 680 | static int recalc_task_prio(task_t *p, unsigned long long now) |
1da177e4 LT |
681 | { |
682 | /* Caller must always ensure 'now >= p->timestamp' */ | |
683 | unsigned long long __sleep_time = now - p->timestamp; | |
684 | unsigned long sleep_time; | |
685 | ||
686 | if (__sleep_time > NS_MAX_SLEEP_AVG) | |
687 | sleep_time = NS_MAX_SLEEP_AVG; | |
688 | else | |
689 | sleep_time = (unsigned long)__sleep_time; | |
690 | ||
691 | if (likely(sleep_time > 0)) { | |
692 | /* | |
693 | * User tasks that sleep a long time are categorised as | |
694 | * idle and will get just interactive status to stay active & | |
695 | * prevent them suddenly becoming cpu hogs and starving | |
696 | * other processes. | |
697 | */ | |
698 | if (p->mm && p->activated != -1 && | |
699 | sleep_time > INTERACTIVE_SLEEP(p)) { | |
700 | p->sleep_avg = JIFFIES_TO_NS(MAX_SLEEP_AVG - | |
701 | DEF_TIMESLICE); | |
702 | } else { | |
703 | /* | |
704 | * The lower the sleep avg a task has the more | |
705 | * rapidly it will rise with sleep time. | |
706 | */ | |
707 | sleep_time *= (MAX_BONUS - CURRENT_BONUS(p)) ? : 1; | |
708 | ||
709 | /* | |
710 | * Tasks waking from uninterruptible sleep are | |
711 | * limited in their sleep_avg rise as they | |
712 | * are likely to be waiting on I/O | |
713 | */ | |
714 | if (p->activated == -1 && p->mm) { | |
715 | if (p->sleep_avg >= INTERACTIVE_SLEEP(p)) | |
716 | sleep_time = 0; | |
717 | else if (p->sleep_avg + sleep_time >= | |
718 | INTERACTIVE_SLEEP(p)) { | |
719 | p->sleep_avg = INTERACTIVE_SLEEP(p); | |
720 | sleep_time = 0; | |
721 | } | |
722 | } | |
723 | ||
724 | /* | |
725 | * This code gives a bonus to interactive tasks. | |
726 | * | |
727 | * The boost works by updating the 'average sleep time' | |
728 | * value here, based on ->timestamp. The more time a | |
729 | * task spends sleeping, the higher the average gets - | |
730 | * and the higher the priority boost gets as well. | |
731 | */ | |
732 | p->sleep_avg += sleep_time; | |
733 | ||
734 | if (p->sleep_avg > NS_MAX_SLEEP_AVG) | |
735 | p->sleep_avg = NS_MAX_SLEEP_AVG; | |
736 | } | |
737 | } | |
738 | ||
a3464a10 | 739 | return effective_prio(p); |
1da177e4 LT |
740 | } |
741 | ||
742 | /* | |
743 | * activate_task - move a task to the runqueue and do priority recalculation | |
744 | * | |
745 | * Update all the scheduling statistics stuff. (sleep average | |
746 | * calculation, priority modifiers, etc.) | |
747 | */ | |
748 | static void activate_task(task_t *p, runqueue_t *rq, int local) | |
749 | { | |
750 | unsigned long long now; | |
751 | ||
752 | now = sched_clock(); | |
753 | #ifdef CONFIG_SMP | |
754 | if (!local) { | |
755 | /* Compensate for drifting sched_clock */ | |
756 | runqueue_t *this_rq = this_rq(); | |
757 | now = (now - this_rq->timestamp_last_tick) | |
758 | + rq->timestamp_last_tick; | |
759 | } | |
760 | #endif | |
761 | ||
a3464a10 | 762 | p->prio = recalc_task_prio(p, now); |
1da177e4 LT |
763 | |
764 | /* | |
765 | * This checks to make sure it's not an uninterruptible task | |
766 | * that is now waking up. | |
767 | */ | |
768 | if (!p->activated) { | |
769 | /* | |
770 | * Tasks which were woken up by interrupts (ie. hw events) | |
771 | * are most likely of interactive nature. So we give them | |
772 | * the credit of extending their sleep time to the period | |
773 | * of time they spend on the runqueue, waiting for execution | |
774 | * on a CPU, first time around: | |
775 | */ | |
776 | if (in_interrupt()) | |
777 | p->activated = 2; | |
778 | else { | |
779 | /* | |
780 | * Normal first-time wakeups get a credit too for | |
781 | * on-runqueue time, but it will be weighted down: | |
782 | */ | |
783 | p->activated = 1; | |
784 | } | |
785 | } | |
786 | p->timestamp = now; | |
787 | ||
788 | __activate_task(p, rq); | |
789 | } | |
790 | ||
791 | /* | |
792 | * deactivate_task - remove a task from the runqueue. | |
793 | */ | |
794 | static void deactivate_task(struct task_struct *p, runqueue_t *rq) | |
795 | { | |
796 | rq->nr_running--; | |
797 | dequeue_task(p, p->array); | |
798 | p->array = NULL; | |
799 | } | |
800 | ||
801 | /* | |
802 | * resched_task - mark a task 'to be rescheduled now'. | |
803 | * | |
804 | * On UP this means the setting of the need_resched flag, on SMP it | |
805 | * might also involve a cross-CPU call to trigger the scheduler on | |
806 | * the target CPU. | |
807 | */ | |
808 | #ifdef CONFIG_SMP | |
809 | static void resched_task(task_t *p) | |
810 | { | |
811 | int need_resched, nrpolling; | |
812 | ||
813 | assert_spin_locked(&task_rq(p)->lock); | |
814 | ||
815 | /* minimise the chance of sending an interrupt to poll_idle() */ | |
816 | nrpolling = test_tsk_thread_flag(p,TIF_POLLING_NRFLAG); | |
817 | need_resched = test_and_set_tsk_thread_flag(p,TIF_NEED_RESCHED); | |
818 | nrpolling |= test_tsk_thread_flag(p,TIF_POLLING_NRFLAG); | |
819 | ||
820 | if (!need_resched && !nrpolling && (task_cpu(p) != smp_processor_id())) | |
821 | smp_send_reschedule(task_cpu(p)); | |
822 | } | |
823 | #else | |
824 | static inline void resched_task(task_t *p) | |
825 | { | |
826 | set_tsk_need_resched(p); | |
827 | } | |
828 | #endif | |
829 | ||
830 | /** | |
831 | * task_curr - is this task currently executing on a CPU? | |
832 | * @p: the task in question. | |
833 | */ | |
834 | inline int task_curr(const task_t *p) | |
835 | { | |
836 | return cpu_curr(task_cpu(p)) == p; | |
837 | } | |
838 | ||
839 | #ifdef CONFIG_SMP | |
1da177e4 LT |
840 | typedef struct { |
841 | struct list_head list; | |
1da177e4 | 842 | |
1da177e4 LT |
843 | task_t *task; |
844 | int dest_cpu; | |
845 | ||
1da177e4 LT |
846 | struct completion done; |
847 | } migration_req_t; | |
848 | ||
849 | /* | |
850 | * The task's runqueue lock must be held. | |
851 | * Returns true if you have to wait for migration thread. | |
852 | */ | |
853 | static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req) | |
854 | { | |
855 | runqueue_t *rq = task_rq(p); | |
856 | ||
857 | /* | |
858 | * If the task is not on a runqueue (and not running), then | |
859 | * it is sufficient to simply update the task's cpu field. | |
860 | */ | |
861 | if (!p->array && !task_running(rq, p)) { | |
862 | set_task_cpu(p, dest_cpu); | |
863 | return 0; | |
864 | } | |
865 | ||
866 | init_completion(&req->done); | |
1da177e4 LT |
867 | req->task = p; |
868 | req->dest_cpu = dest_cpu; | |
869 | list_add(&req->list, &rq->migration_queue); | |
870 | return 1; | |
871 | } | |
872 | ||
873 | /* | |
874 | * wait_task_inactive - wait for a thread to unschedule. | |
875 | * | |
876 | * The caller must ensure that the task *will* unschedule sometime soon, | |
877 | * else this function might spin for a *long* time. This function can't | |
878 | * be called with interrupts off, or it may introduce deadlock with | |
879 | * smp_call_function() if an IPI is sent by the same process we are | |
880 | * waiting to become inactive. | |
881 | */ | |
95cdf3b7 | 882 | void wait_task_inactive(task_t *p) |
1da177e4 LT |
883 | { |
884 | unsigned long flags; | |
885 | runqueue_t *rq; | |
886 | int preempted; | |
887 | ||
888 | repeat: | |
889 | rq = task_rq_lock(p, &flags); | |
890 | /* Must be off runqueue entirely, not preempted. */ | |
891 | if (unlikely(p->array || task_running(rq, p))) { | |
892 | /* If it's preempted, we yield. It could be a while. */ | |
893 | preempted = !task_running(rq, p); | |
894 | task_rq_unlock(rq, &flags); | |
895 | cpu_relax(); | |
896 | if (preempted) | |
897 | yield(); | |
898 | goto repeat; | |
899 | } | |
900 | task_rq_unlock(rq, &flags); | |
901 | } | |
902 | ||
903 | /*** | |
904 | * kick_process - kick a running thread to enter/exit the kernel | |
905 | * @p: the to-be-kicked thread | |
906 | * | |
907 | * Cause a process which is running on another CPU to enter | |
908 | * kernel-mode, without any delay. (to get signals handled.) | |
909 | * | |
910 | * NOTE: this function doesnt have to take the runqueue lock, | |
911 | * because all it wants to ensure is that the remote task enters | |
912 | * the kernel. If the IPI races and the task has been migrated | |
913 | * to another CPU then no harm is done and the purpose has been | |
914 | * achieved as well. | |
915 | */ | |
916 | void kick_process(task_t *p) | |
917 | { | |
918 | int cpu; | |
919 | ||
920 | preempt_disable(); | |
921 | cpu = task_cpu(p); | |
922 | if ((cpu != smp_processor_id()) && task_curr(p)) | |
923 | smp_send_reschedule(cpu); | |
924 | preempt_enable(); | |
925 | } | |
926 | ||
927 | /* | |
928 | * Return a low guess at the load of a migration-source cpu. | |
929 | * | |
930 | * We want to under-estimate the load of migration sources, to | |
931 | * balance conservatively. | |
932 | */ | |
7897986b | 933 | static inline unsigned long source_load(int cpu, int type) |
1da177e4 LT |
934 | { |
935 | runqueue_t *rq = cpu_rq(cpu); | |
936 | unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE; | |
7897986b NP |
937 | if (type == 0) |
938 | return load_now; | |
1da177e4 | 939 | |
7897986b | 940 | return min(rq->cpu_load[type-1], load_now); |
1da177e4 LT |
941 | } |
942 | ||
943 | /* | |
944 | * Return a high guess at the load of a migration-target cpu | |
945 | */ | |
7897986b | 946 | static inline unsigned long target_load(int cpu, int type) |
1da177e4 LT |
947 | { |
948 | runqueue_t *rq = cpu_rq(cpu); | |
949 | unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE; | |
7897986b NP |
950 | if (type == 0) |
951 | return load_now; | |
1da177e4 | 952 | |
7897986b | 953 | return max(rq->cpu_load[type-1], load_now); |
1da177e4 LT |
954 | } |
955 | ||
147cbb4b NP |
956 | /* |
957 | * find_idlest_group finds and returns the least busy CPU group within the | |
958 | * domain. | |
959 | */ | |
960 | static struct sched_group * | |
961 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu) | |
962 | { | |
963 | struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups; | |
964 | unsigned long min_load = ULONG_MAX, this_load = 0; | |
965 | int load_idx = sd->forkexec_idx; | |
966 | int imbalance = 100 + (sd->imbalance_pct-100)/2; | |
967 | ||
968 | do { | |
969 | unsigned long load, avg_load; | |
970 | int local_group; | |
971 | int i; | |
972 | ||
da5a5522 BD |
973 | /* Skip over this group if it has no CPUs allowed */ |
974 | if (!cpus_intersects(group->cpumask, p->cpus_allowed)) | |
975 | goto nextgroup; | |
976 | ||
147cbb4b | 977 | local_group = cpu_isset(this_cpu, group->cpumask); |
147cbb4b NP |
978 | |
979 | /* Tally up the load of all CPUs in the group */ | |
980 | avg_load = 0; | |
981 | ||
982 | for_each_cpu_mask(i, group->cpumask) { | |
983 | /* Bias balancing toward cpus of our domain */ | |
984 | if (local_group) | |
985 | load = source_load(i, load_idx); | |
986 | else | |
987 | load = target_load(i, load_idx); | |
988 | ||
989 | avg_load += load; | |
990 | } | |
991 | ||
992 | /* Adjust by relative CPU power of the group */ | |
993 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
994 | ||
995 | if (local_group) { | |
996 | this_load = avg_load; | |
997 | this = group; | |
998 | } else if (avg_load < min_load) { | |
999 | min_load = avg_load; | |
1000 | idlest = group; | |
1001 | } | |
da5a5522 | 1002 | nextgroup: |
147cbb4b NP |
1003 | group = group->next; |
1004 | } while (group != sd->groups); | |
1005 | ||
1006 | if (!idlest || 100*this_load < imbalance*min_load) | |
1007 | return NULL; | |
1008 | return idlest; | |
1009 | } | |
1010 | ||
1011 | /* | |
1012 | * find_idlest_queue - find the idlest runqueue among the cpus in group. | |
1013 | */ | |
95cdf3b7 IM |
1014 | static int |
1015 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
147cbb4b | 1016 | { |
da5a5522 | 1017 | cpumask_t tmp; |
147cbb4b NP |
1018 | unsigned long load, min_load = ULONG_MAX; |
1019 | int idlest = -1; | |
1020 | int i; | |
1021 | ||
da5a5522 BD |
1022 | /* Traverse only the allowed CPUs */ |
1023 | cpus_and(tmp, group->cpumask, p->cpus_allowed); | |
1024 | ||
1025 | for_each_cpu_mask(i, tmp) { | |
147cbb4b NP |
1026 | load = source_load(i, 0); |
1027 | ||
1028 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
1029 | min_load = load; | |
1030 | idlest = i; | |
1031 | } | |
1032 | } | |
1033 | ||
1034 | return idlest; | |
1035 | } | |
1036 | ||
476d139c NP |
1037 | /* |
1038 | * sched_balance_self: balance the current task (running on cpu) in domains | |
1039 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
1040 | * SD_BALANCE_EXEC. | |
1041 | * | |
1042 | * Balance, ie. select the least loaded group. | |
1043 | * | |
1044 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
1045 | * | |
1046 | * preempt must be disabled. | |
1047 | */ | |
1048 | static int sched_balance_self(int cpu, int flag) | |
1049 | { | |
1050 | struct task_struct *t = current; | |
1051 | struct sched_domain *tmp, *sd = NULL; | |
147cbb4b | 1052 | |
476d139c NP |
1053 | for_each_domain(cpu, tmp) |
1054 | if (tmp->flags & flag) | |
1055 | sd = tmp; | |
1056 | ||
1057 | while (sd) { | |
1058 | cpumask_t span; | |
1059 | struct sched_group *group; | |
1060 | int new_cpu; | |
1061 | int weight; | |
1062 | ||
1063 | span = sd->span; | |
1064 | group = find_idlest_group(sd, t, cpu); | |
1065 | if (!group) | |
1066 | goto nextlevel; | |
1067 | ||
da5a5522 | 1068 | new_cpu = find_idlest_cpu(group, t, cpu); |
476d139c NP |
1069 | if (new_cpu == -1 || new_cpu == cpu) |
1070 | goto nextlevel; | |
1071 | ||
1072 | /* Now try balancing at a lower domain level */ | |
1073 | cpu = new_cpu; | |
1074 | nextlevel: | |
1075 | sd = NULL; | |
1076 | weight = cpus_weight(span); | |
1077 | for_each_domain(cpu, tmp) { | |
1078 | if (weight <= cpus_weight(tmp->span)) | |
1079 | break; | |
1080 | if (tmp->flags & flag) | |
1081 | sd = tmp; | |
1082 | } | |
1083 | /* while loop will break here if sd == NULL */ | |
1084 | } | |
1085 | ||
1086 | return cpu; | |
1087 | } | |
1088 | ||
1089 | #endif /* CONFIG_SMP */ | |
1da177e4 LT |
1090 | |
1091 | /* | |
1092 | * wake_idle() will wake a task on an idle cpu if task->cpu is | |
1093 | * not idle and an idle cpu is available. The span of cpus to | |
1094 | * search starts with cpus closest then further out as needed, | |
1095 | * so we always favor a closer, idle cpu. | |
1096 | * | |
1097 | * Returns the CPU we should wake onto. | |
1098 | */ | |
1099 | #if defined(ARCH_HAS_SCHED_WAKE_IDLE) | |
1100 | static int wake_idle(int cpu, task_t *p) | |
1101 | { | |
1102 | cpumask_t tmp; | |
1103 | struct sched_domain *sd; | |
1104 | int i; | |
1105 | ||
1106 | if (idle_cpu(cpu)) | |
1107 | return cpu; | |
1108 | ||
1109 | for_each_domain(cpu, sd) { | |
1110 | if (sd->flags & SD_WAKE_IDLE) { | |
e0f364f4 | 1111 | cpus_and(tmp, sd->span, p->cpus_allowed); |
1da177e4 LT |
1112 | for_each_cpu_mask(i, tmp) { |
1113 | if (idle_cpu(i)) | |
1114 | return i; | |
1115 | } | |
1116 | } | |
e0f364f4 NP |
1117 | else |
1118 | break; | |
1da177e4 LT |
1119 | } |
1120 | return cpu; | |
1121 | } | |
1122 | #else | |
1123 | static inline int wake_idle(int cpu, task_t *p) | |
1124 | { | |
1125 | return cpu; | |
1126 | } | |
1127 | #endif | |
1128 | ||
1129 | /*** | |
1130 | * try_to_wake_up - wake up a thread | |
1131 | * @p: the to-be-woken-up thread | |
1132 | * @state: the mask of task states that can be woken | |
1133 | * @sync: do a synchronous wakeup? | |
1134 | * | |
1135 | * Put it on the run-queue if it's not already there. The "current" | |
1136 | * thread is always on the run-queue (except when the actual | |
1137 | * re-schedule is in progress), and as such you're allowed to do | |
1138 | * the simpler "current->state = TASK_RUNNING" to mark yourself | |
1139 | * runnable without the overhead of this. | |
1140 | * | |
1141 | * returns failure only if the task is already active. | |
1142 | */ | |
95cdf3b7 | 1143 | static int try_to_wake_up(task_t *p, unsigned int state, int sync) |
1da177e4 LT |
1144 | { |
1145 | int cpu, this_cpu, success = 0; | |
1146 | unsigned long flags; | |
1147 | long old_state; | |
1148 | runqueue_t *rq; | |
1149 | #ifdef CONFIG_SMP | |
1150 | unsigned long load, this_load; | |
7897986b | 1151 | struct sched_domain *sd, *this_sd = NULL; |
1da177e4 LT |
1152 | int new_cpu; |
1153 | #endif | |
1154 | ||
1155 | rq = task_rq_lock(p, &flags); | |
1156 | old_state = p->state; | |
1157 | if (!(old_state & state)) | |
1158 | goto out; | |
1159 | ||
1160 | if (p->array) | |
1161 | goto out_running; | |
1162 | ||
1163 | cpu = task_cpu(p); | |
1164 | this_cpu = smp_processor_id(); | |
1165 | ||
1166 | #ifdef CONFIG_SMP | |
1167 | if (unlikely(task_running(rq, p))) | |
1168 | goto out_activate; | |
1169 | ||
7897986b NP |
1170 | new_cpu = cpu; |
1171 | ||
1da177e4 LT |
1172 | schedstat_inc(rq, ttwu_cnt); |
1173 | if (cpu == this_cpu) { | |
1174 | schedstat_inc(rq, ttwu_local); | |
7897986b NP |
1175 | goto out_set_cpu; |
1176 | } | |
1177 | ||
1178 | for_each_domain(this_cpu, sd) { | |
1179 | if (cpu_isset(cpu, sd->span)) { | |
1180 | schedstat_inc(sd, ttwu_wake_remote); | |
1181 | this_sd = sd; | |
1182 | break; | |
1da177e4 LT |
1183 | } |
1184 | } | |
1da177e4 | 1185 | |
7897986b | 1186 | if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed))) |
1da177e4 LT |
1187 | goto out_set_cpu; |
1188 | ||
1da177e4 | 1189 | /* |
7897986b | 1190 | * Check for affine wakeup and passive balancing possibilities. |
1da177e4 | 1191 | */ |
7897986b NP |
1192 | if (this_sd) { |
1193 | int idx = this_sd->wake_idx; | |
1194 | unsigned int imbalance; | |
1da177e4 | 1195 | |
a3f21bce NP |
1196 | imbalance = 100 + (this_sd->imbalance_pct - 100) / 2; |
1197 | ||
7897986b NP |
1198 | load = source_load(cpu, idx); |
1199 | this_load = target_load(this_cpu, idx); | |
1da177e4 | 1200 | |
7897986b NP |
1201 | new_cpu = this_cpu; /* Wake to this CPU if we can */ |
1202 | ||
a3f21bce NP |
1203 | if (this_sd->flags & SD_WAKE_AFFINE) { |
1204 | unsigned long tl = this_load; | |
1da177e4 | 1205 | /* |
a3f21bce NP |
1206 | * If sync wakeup then subtract the (maximum possible) |
1207 | * effect of the currently running task from the load | |
1208 | * of the current CPU: | |
1da177e4 | 1209 | */ |
a3f21bce NP |
1210 | if (sync) |
1211 | tl -= SCHED_LOAD_SCALE; | |
1212 | ||
1213 | if ((tl <= load && | |
1214 | tl + target_load(cpu, idx) <= SCHED_LOAD_SCALE) || | |
1215 | 100*(tl + SCHED_LOAD_SCALE) <= imbalance*load) { | |
1216 | /* | |
1217 | * This domain has SD_WAKE_AFFINE and | |
1218 | * p is cache cold in this domain, and | |
1219 | * there is no bad imbalance. | |
1220 | */ | |
1221 | schedstat_inc(this_sd, ttwu_move_affine); | |
1222 | goto out_set_cpu; | |
1223 | } | |
1224 | } | |
1225 | ||
1226 | /* | |
1227 | * Start passive balancing when half the imbalance_pct | |
1228 | * limit is reached. | |
1229 | */ | |
1230 | if (this_sd->flags & SD_WAKE_BALANCE) { | |
1231 | if (imbalance*this_load <= 100*load) { | |
1232 | schedstat_inc(this_sd, ttwu_move_balance); | |
1233 | goto out_set_cpu; | |
1234 | } | |
1da177e4 LT |
1235 | } |
1236 | } | |
1237 | ||
1238 | new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */ | |
1239 | out_set_cpu: | |
1240 | new_cpu = wake_idle(new_cpu, p); | |
1241 | if (new_cpu != cpu) { | |
1242 | set_task_cpu(p, new_cpu); | |
1243 | task_rq_unlock(rq, &flags); | |
1244 | /* might preempt at this point */ | |
1245 | rq = task_rq_lock(p, &flags); | |
1246 | old_state = p->state; | |
1247 | if (!(old_state & state)) | |
1248 | goto out; | |
1249 | if (p->array) | |
1250 | goto out_running; | |
1251 | ||
1252 | this_cpu = smp_processor_id(); | |
1253 | cpu = task_cpu(p); | |
1254 | } | |
1255 | ||
1256 | out_activate: | |
1257 | #endif /* CONFIG_SMP */ | |
1258 | if (old_state == TASK_UNINTERRUPTIBLE) { | |
1259 | rq->nr_uninterruptible--; | |
1260 | /* | |
1261 | * Tasks on involuntary sleep don't earn | |
1262 | * sleep_avg beyond just interactive state. | |
1263 | */ | |
1264 | p->activated = -1; | |
1265 | } | |
1266 | ||
d79fc0fc IM |
1267 | /* |
1268 | * Tasks that have marked their sleep as noninteractive get | |
1269 | * woken up without updating their sleep average. (i.e. their | |
1270 | * sleep is handled in a priority-neutral manner, no priority | |
1271 | * boost and no penalty.) | |
1272 | */ | |
1273 | if (old_state & TASK_NONINTERACTIVE) | |
1274 | __activate_task(p, rq); | |
1275 | else | |
1276 | activate_task(p, rq, cpu == this_cpu); | |
1da177e4 LT |
1277 | /* |
1278 | * Sync wakeups (i.e. those types of wakeups where the waker | |
1279 | * has indicated that it will leave the CPU in short order) | |
1280 | * don't trigger a preemption, if the woken up task will run on | |
1281 | * this cpu. (in this case the 'I will reschedule' promise of | |
1282 | * the waker guarantees that the freshly woken up task is going | |
1283 | * to be considered on this CPU.) | |
1284 | */ | |
1da177e4 LT |
1285 | if (!sync || cpu != this_cpu) { |
1286 | if (TASK_PREEMPTS_CURR(p, rq)) | |
1287 | resched_task(rq->curr); | |
1288 | } | |
1289 | success = 1; | |
1290 | ||
1291 | out_running: | |
1292 | p->state = TASK_RUNNING; | |
1293 | out: | |
1294 | task_rq_unlock(rq, &flags); | |
1295 | ||
1296 | return success; | |
1297 | } | |
1298 | ||
95cdf3b7 | 1299 | int fastcall wake_up_process(task_t *p) |
1da177e4 LT |
1300 | { |
1301 | return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED | | |
1302 | TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0); | |
1303 | } | |
1304 | ||
1305 | EXPORT_SYMBOL(wake_up_process); | |
1306 | ||
1307 | int fastcall wake_up_state(task_t *p, unsigned int state) | |
1308 | { | |
1309 | return try_to_wake_up(p, state, 0); | |
1310 | } | |
1311 | ||
1da177e4 LT |
1312 | /* |
1313 | * Perform scheduler related setup for a newly forked process p. | |
1314 | * p is forked by current. | |
1315 | */ | |
476d139c | 1316 | void fastcall sched_fork(task_t *p, int clone_flags) |
1da177e4 | 1317 | { |
476d139c NP |
1318 | int cpu = get_cpu(); |
1319 | ||
1320 | #ifdef CONFIG_SMP | |
1321 | cpu = sched_balance_self(cpu, SD_BALANCE_FORK); | |
1322 | #endif | |
1323 | set_task_cpu(p, cpu); | |
1324 | ||
1da177e4 LT |
1325 | /* |
1326 | * We mark the process as running here, but have not actually | |
1327 | * inserted it onto the runqueue yet. This guarantees that | |
1328 | * nobody will actually run it, and a signal or other external | |
1329 | * event cannot wake it up and insert it on the runqueue either. | |
1330 | */ | |
1331 | p->state = TASK_RUNNING; | |
1332 | INIT_LIST_HEAD(&p->run_list); | |
1333 | p->array = NULL; | |
1da177e4 LT |
1334 | #ifdef CONFIG_SCHEDSTATS |
1335 | memset(&p->sched_info, 0, sizeof(p->sched_info)); | |
1336 | #endif | |
4866cde0 NP |
1337 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
1338 | p->oncpu = 0; | |
1339 | #endif | |
1da177e4 | 1340 | #ifdef CONFIG_PREEMPT |
4866cde0 | 1341 | /* Want to start with kernel preemption disabled. */ |
1da177e4 LT |
1342 | p->thread_info->preempt_count = 1; |
1343 | #endif | |
1344 | /* | |
1345 | * Share the timeslice between parent and child, thus the | |
1346 | * total amount of pending timeslices in the system doesn't change, | |
1347 | * resulting in more scheduling fairness. | |
1348 | */ | |
1349 | local_irq_disable(); | |
1350 | p->time_slice = (current->time_slice + 1) >> 1; | |
1351 | /* | |
1352 | * The remainder of the first timeslice might be recovered by | |
1353 | * the parent if the child exits early enough. | |
1354 | */ | |
1355 | p->first_time_slice = 1; | |
1356 | current->time_slice >>= 1; | |
1357 | p->timestamp = sched_clock(); | |
1358 | if (unlikely(!current->time_slice)) { | |
1359 | /* | |
1360 | * This case is rare, it happens when the parent has only | |
1361 | * a single jiffy left from its timeslice. Taking the | |
1362 | * runqueue lock is not a problem. | |
1363 | */ | |
1364 | current->time_slice = 1; | |
1da177e4 | 1365 | scheduler_tick(); |
476d139c NP |
1366 | } |
1367 | local_irq_enable(); | |
1368 | put_cpu(); | |
1da177e4 LT |
1369 | } |
1370 | ||
1371 | /* | |
1372 | * wake_up_new_task - wake up a newly created task for the first time. | |
1373 | * | |
1374 | * This function will do some initial scheduler statistics housekeeping | |
1375 | * that must be done for every newly created context, then puts the task | |
1376 | * on the runqueue and wakes it. | |
1377 | */ | |
95cdf3b7 | 1378 | void fastcall wake_up_new_task(task_t *p, unsigned long clone_flags) |
1da177e4 LT |
1379 | { |
1380 | unsigned long flags; | |
1381 | int this_cpu, cpu; | |
1382 | runqueue_t *rq, *this_rq; | |
1383 | ||
1384 | rq = task_rq_lock(p, &flags); | |
147cbb4b | 1385 | BUG_ON(p->state != TASK_RUNNING); |
1da177e4 | 1386 | this_cpu = smp_processor_id(); |
147cbb4b | 1387 | cpu = task_cpu(p); |
1da177e4 | 1388 | |
1da177e4 LT |
1389 | /* |
1390 | * We decrease the sleep average of forking parents | |
1391 | * and children as well, to keep max-interactive tasks | |
1392 | * from forking tasks that are max-interactive. The parent | |
1393 | * (current) is done further down, under its lock. | |
1394 | */ | |
1395 | p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) * | |
1396 | CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS); | |
1397 | ||
1398 | p->prio = effective_prio(p); | |
1399 | ||
1400 | if (likely(cpu == this_cpu)) { | |
1401 | if (!(clone_flags & CLONE_VM)) { | |
1402 | /* | |
1403 | * The VM isn't cloned, so we're in a good position to | |
1404 | * do child-runs-first in anticipation of an exec. This | |
1405 | * usually avoids a lot of COW overhead. | |
1406 | */ | |
1407 | if (unlikely(!current->array)) | |
1408 | __activate_task(p, rq); | |
1409 | else { | |
1410 | p->prio = current->prio; | |
1411 | list_add_tail(&p->run_list, ¤t->run_list); | |
1412 | p->array = current->array; | |
1413 | p->array->nr_active++; | |
1414 | rq->nr_running++; | |
1415 | } | |
1416 | set_need_resched(); | |
1417 | } else | |
1418 | /* Run child last */ | |
1419 | __activate_task(p, rq); | |
1420 | /* | |
1421 | * We skip the following code due to cpu == this_cpu | |
1422 | * | |
1423 | * task_rq_unlock(rq, &flags); | |
1424 | * this_rq = task_rq_lock(current, &flags); | |
1425 | */ | |
1426 | this_rq = rq; | |
1427 | } else { | |
1428 | this_rq = cpu_rq(this_cpu); | |
1429 | ||
1430 | /* | |
1431 | * Not the local CPU - must adjust timestamp. This should | |
1432 | * get optimised away in the !CONFIG_SMP case. | |
1433 | */ | |
1434 | p->timestamp = (p->timestamp - this_rq->timestamp_last_tick) | |
1435 | + rq->timestamp_last_tick; | |
1436 | __activate_task(p, rq); | |
1437 | if (TASK_PREEMPTS_CURR(p, rq)) | |
1438 | resched_task(rq->curr); | |
1439 | ||
1440 | /* | |
1441 | * Parent and child are on different CPUs, now get the | |
1442 | * parent runqueue to update the parent's ->sleep_avg: | |
1443 | */ | |
1444 | task_rq_unlock(rq, &flags); | |
1445 | this_rq = task_rq_lock(current, &flags); | |
1446 | } | |
1447 | current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) * | |
1448 | PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS); | |
1449 | task_rq_unlock(this_rq, &flags); | |
1450 | } | |
1451 | ||
1452 | /* | |
1453 | * Potentially available exiting-child timeslices are | |
1454 | * retrieved here - this way the parent does not get | |
1455 | * penalized for creating too many threads. | |
1456 | * | |
1457 | * (this cannot be used to 'generate' timeslices | |
1458 | * artificially, because any timeslice recovered here | |
1459 | * was given away by the parent in the first place.) | |
1460 | */ | |
95cdf3b7 | 1461 | void fastcall sched_exit(task_t *p) |
1da177e4 LT |
1462 | { |
1463 | unsigned long flags; | |
1464 | runqueue_t *rq; | |
1465 | ||
1466 | /* | |
1467 | * If the child was a (relative-) CPU hog then decrease | |
1468 | * the sleep_avg of the parent as well. | |
1469 | */ | |
1470 | rq = task_rq_lock(p->parent, &flags); | |
1471 | if (p->first_time_slice) { | |
1472 | p->parent->time_slice += p->time_slice; | |
1473 | if (unlikely(p->parent->time_slice > task_timeslice(p))) | |
1474 | p->parent->time_slice = task_timeslice(p); | |
1475 | } | |
1476 | if (p->sleep_avg < p->parent->sleep_avg) | |
1477 | p->parent->sleep_avg = p->parent->sleep_avg / | |
1478 | (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg / | |
1479 | (EXIT_WEIGHT + 1); | |
1480 | task_rq_unlock(rq, &flags); | |
1481 | } | |
1482 | ||
4866cde0 NP |
1483 | /** |
1484 | * prepare_task_switch - prepare to switch tasks | |
1485 | * @rq: the runqueue preparing to switch | |
1486 | * @next: the task we are going to switch to. | |
1487 | * | |
1488 | * This is called with the rq lock held and interrupts off. It must | |
1489 | * be paired with a subsequent finish_task_switch after the context | |
1490 | * switch. | |
1491 | * | |
1492 | * prepare_task_switch sets up locking and calls architecture specific | |
1493 | * hooks. | |
1494 | */ | |
1495 | static inline void prepare_task_switch(runqueue_t *rq, task_t *next) | |
1496 | { | |
1497 | prepare_lock_switch(rq, next); | |
1498 | prepare_arch_switch(next); | |
1499 | } | |
1500 | ||
1da177e4 LT |
1501 | /** |
1502 | * finish_task_switch - clean up after a task-switch | |
344babaa | 1503 | * @rq: runqueue associated with task-switch |
1da177e4 LT |
1504 | * @prev: the thread we just switched away from. |
1505 | * | |
4866cde0 NP |
1506 | * finish_task_switch must be called after the context switch, paired |
1507 | * with a prepare_task_switch call before the context switch. | |
1508 | * finish_task_switch will reconcile locking set up by prepare_task_switch, | |
1509 | * and do any other architecture-specific cleanup actions. | |
1da177e4 LT |
1510 | * |
1511 | * Note that we may have delayed dropping an mm in context_switch(). If | |
1512 | * so, we finish that here outside of the runqueue lock. (Doing it | |
1513 | * with the lock held can cause deadlocks; see schedule() for | |
1514 | * details.) | |
1515 | */ | |
4866cde0 | 1516 | static inline void finish_task_switch(runqueue_t *rq, task_t *prev) |
1da177e4 LT |
1517 | __releases(rq->lock) |
1518 | { | |
1da177e4 LT |
1519 | struct mm_struct *mm = rq->prev_mm; |
1520 | unsigned long prev_task_flags; | |
1521 | ||
1522 | rq->prev_mm = NULL; | |
1523 | ||
1524 | /* | |
1525 | * A task struct has one reference for the use as "current". | |
1526 | * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and | |
1527 | * calls schedule one last time. The schedule call will never return, | |
1528 | * and the scheduled task must drop that reference. | |
1529 | * The test for EXIT_ZOMBIE must occur while the runqueue locks are | |
1530 | * still held, otherwise prev could be scheduled on another cpu, die | |
1531 | * there before we look at prev->state, and then the reference would | |
1532 | * be dropped twice. | |
1533 | * Manfred Spraul <manfred@colorfullife.com> | |
1534 | */ | |
1535 | prev_task_flags = prev->flags; | |
4866cde0 NP |
1536 | finish_arch_switch(prev); |
1537 | finish_lock_switch(rq, prev); | |
1da177e4 LT |
1538 | if (mm) |
1539 | mmdrop(mm); | |
1540 | if (unlikely(prev_task_flags & PF_DEAD)) | |
1541 | put_task_struct(prev); | |
1542 | } | |
1543 | ||
1544 | /** | |
1545 | * schedule_tail - first thing a freshly forked thread must call. | |
1546 | * @prev: the thread we just switched away from. | |
1547 | */ | |
1548 | asmlinkage void schedule_tail(task_t *prev) | |
1549 | __releases(rq->lock) | |
1550 | { | |
4866cde0 NP |
1551 | runqueue_t *rq = this_rq(); |
1552 | finish_task_switch(rq, prev); | |
1553 | #ifdef __ARCH_WANT_UNLOCKED_CTXSW | |
1554 | /* In this case, finish_task_switch does not reenable preemption */ | |
1555 | preempt_enable(); | |
1556 | #endif | |
1da177e4 LT |
1557 | if (current->set_child_tid) |
1558 | put_user(current->pid, current->set_child_tid); | |
1559 | } | |
1560 | ||
1561 | /* | |
1562 | * context_switch - switch to the new MM and the new | |
1563 | * thread's register state. | |
1564 | */ | |
1565 | static inline | |
1566 | task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next) | |
1567 | { | |
1568 | struct mm_struct *mm = next->mm; | |
1569 | struct mm_struct *oldmm = prev->active_mm; | |
1570 | ||
1571 | if (unlikely(!mm)) { | |
1572 | next->active_mm = oldmm; | |
1573 | atomic_inc(&oldmm->mm_count); | |
1574 | enter_lazy_tlb(oldmm, next); | |
1575 | } else | |
1576 | switch_mm(oldmm, mm, next); | |
1577 | ||
1578 | if (unlikely(!prev->mm)) { | |
1579 | prev->active_mm = NULL; | |
1580 | WARN_ON(rq->prev_mm); | |
1581 | rq->prev_mm = oldmm; | |
1582 | } | |
1583 | ||
1584 | /* Here we just switch the register state and the stack. */ | |
1585 | switch_to(prev, next, prev); | |
1586 | ||
1587 | return prev; | |
1588 | } | |
1589 | ||
1590 | /* | |
1591 | * nr_running, nr_uninterruptible and nr_context_switches: | |
1592 | * | |
1593 | * externally visible scheduler statistics: current number of runnable | |
1594 | * threads, current number of uninterruptible-sleeping threads, total | |
1595 | * number of context switches performed since bootup. | |
1596 | */ | |
1597 | unsigned long nr_running(void) | |
1598 | { | |
1599 | unsigned long i, sum = 0; | |
1600 | ||
1601 | for_each_online_cpu(i) | |
1602 | sum += cpu_rq(i)->nr_running; | |
1603 | ||
1604 | return sum; | |
1605 | } | |
1606 | ||
1607 | unsigned long nr_uninterruptible(void) | |
1608 | { | |
1609 | unsigned long i, sum = 0; | |
1610 | ||
1611 | for_each_cpu(i) | |
1612 | sum += cpu_rq(i)->nr_uninterruptible; | |
1613 | ||
1614 | /* | |
1615 | * Since we read the counters lockless, it might be slightly | |
1616 | * inaccurate. Do not allow it to go below zero though: | |
1617 | */ | |
1618 | if (unlikely((long)sum < 0)) | |
1619 | sum = 0; | |
1620 | ||
1621 | return sum; | |
1622 | } | |
1623 | ||
1624 | unsigned long long nr_context_switches(void) | |
1625 | { | |
1626 | unsigned long long i, sum = 0; | |
1627 | ||
1628 | for_each_cpu(i) | |
1629 | sum += cpu_rq(i)->nr_switches; | |
1630 | ||
1631 | return sum; | |
1632 | } | |
1633 | ||
1634 | unsigned long nr_iowait(void) | |
1635 | { | |
1636 | unsigned long i, sum = 0; | |
1637 | ||
1638 | for_each_cpu(i) | |
1639 | sum += atomic_read(&cpu_rq(i)->nr_iowait); | |
1640 | ||
1641 | return sum; | |
1642 | } | |
1643 | ||
1644 | #ifdef CONFIG_SMP | |
1645 | ||
1646 | /* | |
1647 | * double_rq_lock - safely lock two runqueues | |
1648 | * | |
1649 | * Note this does not disable interrupts like task_rq_lock, | |
1650 | * you need to do so manually before calling. | |
1651 | */ | |
1652 | static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2) | |
1653 | __acquires(rq1->lock) | |
1654 | __acquires(rq2->lock) | |
1655 | { | |
1656 | if (rq1 == rq2) { | |
1657 | spin_lock(&rq1->lock); | |
1658 | __acquire(rq2->lock); /* Fake it out ;) */ | |
1659 | } else { | |
1660 | if (rq1 < rq2) { | |
1661 | spin_lock(&rq1->lock); | |
1662 | spin_lock(&rq2->lock); | |
1663 | } else { | |
1664 | spin_lock(&rq2->lock); | |
1665 | spin_lock(&rq1->lock); | |
1666 | } | |
1667 | } | |
1668 | } | |
1669 | ||
1670 | /* | |
1671 | * double_rq_unlock - safely unlock two runqueues | |
1672 | * | |
1673 | * Note this does not restore interrupts like task_rq_unlock, | |
1674 | * you need to do so manually after calling. | |
1675 | */ | |
1676 | static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2) | |
1677 | __releases(rq1->lock) | |
1678 | __releases(rq2->lock) | |
1679 | { | |
1680 | spin_unlock(&rq1->lock); | |
1681 | if (rq1 != rq2) | |
1682 | spin_unlock(&rq2->lock); | |
1683 | else | |
1684 | __release(rq2->lock); | |
1685 | } | |
1686 | ||
1687 | /* | |
1688 | * double_lock_balance - lock the busiest runqueue, this_rq is locked already. | |
1689 | */ | |
1690 | static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest) | |
1691 | __releases(this_rq->lock) | |
1692 | __acquires(busiest->lock) | |
1693 | __acquires(this_rq->lock) | |
1694 | { | |
1695 | if (unlikely(!spin_trylock(&busiest->lock))) { | |
1696 | if (busiest < this_rq) { | |
1697 | spin_unlock(&this_rq->lock); | |
1698 | spin_lock(&busiest->lock); | |
1699 | spin_lock(&this_rq->lock); | |
1700 | } else | |
1701 | spin_lock(&busiest->lock); | |
1702 | } | |
1703 | } | |
1704 | ||
1da177e4 LT |
1705 | /* |
1706 | * If dest_cpu is allowed for this process, migrate the task to it. | |
1707 | * This is accomplished by forcing the cpu_allowed mask to only | |
1708 | * allow dest_cpu, which will force the cpu onto dest_cpu. Then | |
1709 | * the cpu_allowed mask is restored. | |
1710 | */ | |
1711 | static void sched_migrate_task(task_t *p, int dest_cpu) | |
1712 | { | |
1713 | migration_req_t req; | |
1714 | runqueue_t *rq; | |
1715 | unsigned long flags; | |
1716 | ||
1717 | rq = task_rq_lock(p, &flags); | |
1718 | if (!cpu_isset(dest_cpu, p->cpus_allowed) | |
1719 | || unlikely(cpu_is_offline(dest_cpu))) | |
1720 | goto out; | |
1721 | ||
1722 | /* force the process onto the specified CPU */ | |
1723 | if (migrate_task(p, dest_cpu, &req)) { | |
1724 | /* Need to wait for migration thread (might exit: take ref). */ | |
1725 | struct task_struct *mt = rq->migration_thread; | |
1726 | get_task_struct(mt); | |
1727 | task_rq_unlock(rq, &flags); | |
1728 | wake_up_process(mt); | |
1729 | put_task_struct(mt); | |
1730 | wait_for_completion(&req.done); | |
1731 | return; | |
1732 | } | |
1733 | out: | |
1734 | task_rq_unlock(rq, &flags); | |
1735 | } | |
1736 | ||
1737 | /* | |
476d139c NP |
1738 | * sched_exec - execve() is a valuable balancing opportunity, because at |
1739 | * this point the task has the smallest effective memory and cache footprint. | |
1da177e4 LT |
1740 | */ |
1741 | void sched_exec(void) | |
1742 | { | |
1da177e4 | 1743 | int new_cpu, this_cpu = get_cpu(); |
476d139c | 1744 | new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC); |
1da177e4 | 1745 | put_cpu(); |
476d139c NP |
1746 | if (new_cpu != this_cpu) |
1747 | sched_migrate_task(current, new_cpu); | |
1da177e4 LT |
1748 | } |
1749 | ||
1750 | /* | |
1751 | * pull_task - move a task from a remote runqueue to the local runqueue. | |
1752 | * Both runqueues must be locked. | |
1753 | */ | |
1754 | static inline | |
1755 | void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p, | |
1756 | runqueue_t *this_rq, prio_array_t *this_array, int this_cpu) | |
1757 | { | |
1758 | dequeue_task(p, src_array); | |
1759 | src_rq->nr_running--; | |
1760 | set_task_cpu(p, this_cpu); | |
1761 | this_rq->nr_running++; | |
1762 | enqueue_task(p, this_array); | |
1763 | p->timestamp = (p->timestamp - src_rq->timestamp_last_tick) | |
1764 | + this_rq->timestamp_last_tick; | |
1765 | /* | |
1766 | * Note that idle threads have a prio of MAX_PRIO, for this test | |
1767 | * to be always true for them. | |
1768 | */ | |
1769 | if (TASK_PREEMPTS_CURR(p, this_rq)) | |
1770 | resched_task(this_rq->curr); | |
1771 | } | |
1772 | ||
1773 | /* | |
1774 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
1775 | */ | |
1776 | static inline | |
1777 | int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu, | |
95cdf3b7 IM |
1778 | struct sched_domain *sd, enum idle_type idle, |
1779 | int *all_pinned) | |
1da177e4 LT |
1780 | { |
1781 | /* | |
1782 | * We do not migrate tasks that are: | |
1783 | * 1) running (obviously), or | |
1784 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
1785 | * 3) are cache-hot on their current CPU. | |
1786 | */ | |
1da177e4 LT |
1787 | if (!cpu_isset(this_cpu, p->cpus_allowed)) |
1788 | return 0; | |
81026794 NP |
1789 | *all_pinned = 0; |
1790 | ||
1791 | if (task_running(rq, p)) | |
1792 | return 0; | |
1da177e4 LT |
1793 | |
1794 | /* | |
1795 | * Aggressive migration if: | |
cafb20c1 | 1796 | * 1) task is cache cold, or |
1da177e4 LT |
1797 | * 2) too many balance attempts have failed. |
1798 | */ | |
1799 | ||
cafb20c1 | 1800 | if (sd->nr_balance_failed > sd->cache_nice_tries) |
1da177e4 LT |
1801 | return 1; |
1802 | ||
1803 | if (task_hot(p, rq->timestamp_last_tick, sd)) | |
81026794 | 1804 | return 0; |
1da177e4 LT |
1805 | return 1; |
1806 | } | |
1807 | ||
1808 | /* | |
1809 | * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq, | |
1810 | * as part of a balancing operation within "domain". Returns the number of | |
1811 | * tasks moved. | |
1812 | * | |
1813 | * Called with both runqueues locked. | |
1814 | */ | |
1815 | static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest, | |
1816 | unsigned long max_nr_move, struct sched_domain *sd, | |
81026794 | 1817 | enum idle_type idle, int *all_pinned) |
1da177e4 LT |
1818 | { |
1819 | prio_array_t *array, *dst_array; | |
1820 | struct list_head *head, *curr; | |
81026794 | 1821 | int idx, pulled = 0, pinned = 0; |
1da177e4 LT |
1822 | task_t *tmp; |
1823 | ||
81026794 | 1824 | if (max_nr_move == 0) |
1da177e4 LT |
1825 | goto out; |
1826 | ||
81026794 NP |
1827 | pinned = 1; |
1828 | ||
1da177e4 LT |
1829 | /* |
1830 | * We first consider expired tasks. Those will likely not be | |
1831 | * executed in the near future, and they are most likely to | |
1832 | * be cache-cold, thus switching CPUs has the least effect | |
1833 | * on them. | |
1834 | */ | |
1835 | if (busiest->expired->nr_active) { | |
1836 | array = busiest->expired; | |
1837 | dst_array = this_rq->expired; | |
1838 | } else { | |
1839 | array = busiest->active; | |
1840 | dst_array = this_rq->active; | |
1841 | } | |
1842 | ||
1843 | new_array: | |
1844 | /* Start searching at priority 0: */ | |
1845 | idx = 0; | |
1846 | skip_bitmap: | |
1847 | if (!idx) | |
1848 | idx = sched_find_first_bit(array->bitmap); | |
1849 | else | |
1850 | idx = find_next_bit(array->bitmap, MAX_PRIO, idx); | |
1851 | if (idx >= MAX_PRIO) { | |
1852 | if (array == busiest->expired && busiest->active->nr_active) { | |
1853 | array = busiest->active; | |
1854 | dst_array = this_rq->active; | |
1855 | goto new_array; | |
1856 | } | |
1857 | goto out; | |
1858 | } | |
1859 | ||
1860 | head = array->queue + idx; | |
1861 | curr = head->prev; | |
1862 | skip_queue: | |
1863 | tmp = list_entry(curr, task_t, run_list); | |
1864 | ||
1865 | curr = curr->prev; | |
1866 | ||
81026794 | 1867 | if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) { |
1da177e4 LT |
1868 | if (curr != head) |
1869 | goto skip_queue; | |
1870 | idx++; | |
1871 | goto skip_bitmap; | |
1872 | } | |
1873 | ||
1874 | #ifdef CONFIG_SCHEDSTATS | |
1875 | if (task_hot(tmp, busiest->timestamp_last_tick, sd)) | |
1876 | schedstat_inc(sd, lb_hot_gained[idle]); | |
1877 | #endif | |
1878 | ||
1879 | pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu); | |
1880 | pulled++; | |
1881 | ||
1882 | /* We only want to steal up to the prescribed number of tasks. */ | |
1883 | if (pulled < max_nr_move) { | |
1884 | if (curr != head) | |
1885 | goto skip_queue; | |
1886 | idx++; | |
1887 | goto skip_bitmap; | |
1888 | } | |
1889 | out: | |
1890 | /* | |
1891 | * Right now, this is the only place pull_task() is called, | |
1892 | * so we can safely collect pull_task() stats here rather than | |
1893 | * inside pull_task(). | |
1894 | */ | |
1895 | schedstat_add(sd, lb_gained[idle], pulled); | |
81026794 NP |
1896 | |
1897 | if (all_pinned) | |
1898 | *all_pinned = pinned; | |
1da177e4 LT |
1899 | return pulled; |
1900 | } | |
1901 | ||
1902 | /* | |
1903 | * find_busiest_group finds and returns the busiest CPU group within the | |
1904 | * domain. It calculates and returns the number of tasks which should be | |
1905 | * moved to restore balance via the imbalance parameter. | |
1906 | */ | |
1907 | static struct sched_group * | |
1908 | find_busiest_group(struct sched_domain *sd, int this_cpu, | |
5969fe06 | 1909 | unsigned long *imbalance, enum idle_type idle, int *sd_idle) |
1da177e4 LT |
1910 | { |
1911 | struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups; | |
1912 | unsigned long max_load, avg_load, total_load, this_load, total_pwr; | |
0c117f1b | 1913 | unsigned long max_pull; |
7897986b | 1914 | int load_idx; |
1da177e4 LT |
1915 | |
1916 | max_load = this_load = total_load = total_pwr = 0; | |
7897986b NP |
1917 | if (idle == NOT_IDLE) |
1918 | load_idx = sd->busy_idx; | |
1919 | else if (idle == NEWLY_IDLE) | |
1920 | load_idx = sd->newidle_idx; | |
1921 | else | |
1922 | load_idx = sd->idle_idx; | |
1da177e4 LT |
1923 | |
1924 | do { | |
1925 | unsigned long load; | |
1926 | int local_group; | |
1927 | int i; | |
1928 | ||
1929 | local_group = cpu_isset(this_cpu, group->cpumask); | |
1930 | ||
1931 | /* Tally up the load of all CPUs in the group */ | |
1932 | avg_load = 0; | |
1933 | ||
1934 | for_each_cpu_mask(i, group->cpumask) { | |
5969fe06 NP |
1935 | if (*sd_idle && !idle_cpu(i)) |
1936 | *sd_idle = 0; | |
1937 | ||
1da177e4 LT |
1938 | /* Bias balancing toward cpus of our domain */ |
1939 | if (local_group) | |
7897986b | 1940 | load = target_load(i, load_idx); |
1da177e4 | 1941 | else |
7897986b | 1942 | load = source_load(i, load_idx); |
1da177e4 LT |
1943 | |
1944 | avg_load += load; | |
1945 | } | |
1946 | ||
1947 | total_load += avg_load; | |
1948 | total_pwr += group->cpu_power; | |
1949 | ||
1950 | /* Adjust by relative CPU power of the group */ | |
1951 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
1952 | ||
1953 | if (local_group) { | |
1954 | this_load = avg_load; | |
1955 | this = group; | |
1da177e4 LT |
1956 | } else if (avg_load > max_load) { |
1957 | max_load = avg_load; | |
1958 | busiest = group; | |
1959 | } | |
1da177e4 LT |
1960 | group = group->next; |
1961 | } while (group != sd->groups); | |
1962 | ||
0c117f1b | 1963 | if (!busiest || this_load >= max_load || max_load <= SCHED_LOAD_SCALE) |
1da177e4 LT |
1964 | goto out_balanced; |
1965 | ||
1966 | avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr; | |
1967 | ||
1968 | if (this_load >= avg_load || | |
1969 | 100*max_load <= sd->imbalance_pct*this_load) | |
1970 | goto out_balanced; | |
1971 | ||
1972 | /* | |
1973 | * We're trying to get all the cpus to the average_load, so we don't | |
1974 | * want to push ourselves above the average load, nor do we wish to | |
1975 | * reduce the max loaded cpu below the average load, as either of these | |
1976 | * actions would just result in more rebalancing later, and ping-pong | |
1977 | * tasks around. Thus we look for the minimum possible imbalance. | |
1978 | * Negative imbalances (*we* are more loaded than anyone else) will | |
1979 | * be counted as no imbalance for these purposes -- we can't fix that | |
1980 | * by pulling tasks to us. Be careful of negative numbers as they'll | |
1981 | * appear as very large values with unsigned longs. | |
1982 | */ | |
0c117f1b SS |
1983 | |
1984 | /* Don't want to pull so many tasks that a group would go idle */ | |
1985 | max_pull = min(max_load - avg_load, max_load - SCHED_LOAD_SCALE); | |
1986 | ||
1da177e4 | 1987 | /* How much load to actually move to equalise the imbalance */ |
0c117f1b | 1988 | *imbalance = min(max_pull * busiest->cpu_power, |
1da177e4 LT |
1989 | (avg_load - this_load) * this->cpu_power) |
1990 | / SCHED_LOAD_SCALE; | |
1991 | ||
1992 | if (*imbalance < SCHED_LOAD_SCALE) { | |
1993 | unsigned long pwr_now = 0, pwr_move = 0; | |
1994 | unsigned long tmp; | |
1995 | ||
1996 | if (max_load - this_load >= SCHED_LOAD_SCALE*2) { | |
1997 | *imbalance = 1; | |
1998 | return busiest; | |
1999 | } | |
2000 | ||
2001 | /* | |
2002 | * OK, we don't have enough imbalance to justify moving tasks, | |
2003 | * however we may be able to increase total CPU power used by | |
2004 | * moving them. | |
2005 | */ | |
2006 | ||
2007 | pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load); | |
2008 | pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load); | |
2009 | pwr_now /= SCHED_LOAD_SCALE; | |
2010 | ||
2011 | /* Amount of load we'd subtract */ | |
2012 | tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power; | |
2013 | if (max_load > tmp) | |
2014 | pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE, | |
2015 | max_load - tmp); | |
2016 | ||
2017 | /* Amount of load we'd add */ | |
2018 | if (max_load*busiest->cpu_power < | |
2019 | SCHED_LOAD_SCALE*SCHED_LOAD_SCALE) | |
2020 | tmp = max_load*busiest->cpu_power/this->cpu_power; | |
2021 | else | |
2022 | tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power; | |
2023 | pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp); | |
2024 | pwr_move /= SCHED_LOAD_SCALE; | |
2025 | ||
2026 | /* Move if we gain throughput */ | |
2027 | if (pwr_move <= pwr_now) | |
2028 | goto out_balanced; | |
2029 | ||
2030 | *imbalance = 1; | |
2031 | return busiest; | |
2032 | } | |
2033 | ||
2034 | /* Get rid of the scaling factor, rounding down as we divide */ | |
2035 | *imbalance = *imbalance / SCHED_LOAD_SCALE; | |
1da177e4 LT |
2036 | return busiest; |
2037 | ||
2038 | out_balanced: | |
1da177e4 LT |
2039 | |
2040 | *imbalance = 0; | |
2041 | return NULL; | |
2042 | } | |
2043 | ||
2044 | /* | |
2045 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
2046 | */ | |
2047 | static runqueue_t *find_busiest_queue(struct sched_group *group) | |
2048 | { | |
2049 | unsigned long load, max_load = 0; | |
2050 | runqueue_t *busiest = NULL; | |
2051 | int i; | |
2052 | ||
2053 | for_each_cpu_mask(i, group->cpumask) { | |
7897986b | 2054 | load = source_load(i, 0); |
1da177e4 LT |
2055 | |
2056 | if (load > max_load) { | |
2057 | max_load = load; | |
2058 | busiest = cpu_rq(i); | |
2059 | } | |
2060 | } | |
2061 | ||
2062 | return busiest; | |
2063 | } | |
2064 | ||
77391d71 NP |
2065 | /* |
2066 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
2067 | * so long as it is large enough. | |
2068 | */ | |
2069 | #define MAX_PINNED_INTERVAL 512 | |
2070 | ||
1da177e4 LT |
2071 | /* |
2072 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
2073 | * tasks if there is an imbalance. | |
2074 | * | |
2075 | * Called with this_rq unlocked. | |
2076 | */ | |
2077 | static int load_balance(int this_cpu, runqueue_t *this_rq, | |
2078 | struct sched_domain *sd, enum idle_type idle) | |
2079 | { | |
2080 | struct sched_group *group; | |
2081 | runqueue_t *busiest; | |
2082 | unsigned long imbalance; | |
77391d71 | 2083 | int nr_moved, all_pinned = 0; |
81026794 | 2084 | int active_balance = 0; |
5969fe06 NP |
2085 | int sd_idle = 0; |
2086 | ||
2087 | if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER) | |
2088 | sd_idle = 1; | |
1da177e4 | 2089 | |
1da177e4 LT |
2090 | schedstat_inc(sd, lb_cnt[idle]); |
2091 | ||
5969fe06 | 2092 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle); |
1da177e4 LT |
2093 | if (!group) { |
2094 | schedstat_inc(sd, lb_nobusyg[idle]); | |
2095 | goto out_balanced; | |
2096 | } | |
2097 | ||
2098 | busiest = find_busiest_queue(group); | |
2099 | if (!busiest) { | |
2100 | schedstat_inc(sd, lb_nobusyq[idle]); | |
2101 | goto out_balanced; | |
2102 | } | |
2103 | ||
db935dbd | 2104 | BUG_ON(busiest == this_rq); |
1da177e4 LT |
2105 | |
2106 | schedstat_add(sd, lb_imbalance[idle], imbalance); | |
2107 | ||
2108 | nr_moved = 0; | |
2109 | if (busiest->nr_running > 1) { | |
2110 | /* | |
2111 | * Attempt to move tasks. If find_busiest_group has found | |
2112 | * an imbalance but busiest->nr_running <= 1, the group is | |
2113 | * still unbalanced. nr_moved simply stays zero, so it is | |
2114 | * correctly treated as an imbalance. | |
2115 | */ | |
e17224bf | 2116 | double_rq_lock(this_rq, busiest); |
1da177e4 | 2117 | nr_moved = move_tasks(this_rq, this_cpu, busiest, |
d6d5cfaf | 2118 | imbalance, sd, idle, &all_pinned); |
e17224bf | 2119 | double_rq_unlock(this_rq, busiest); |
81026794 NP |
2120 | |
2121 | /* All tasks on this runqueue were pinned by CPU affinity */ | |
2122 | if (unlikely(all_pinned)) | |
2123 | goto out_balanced; | |
1da177e4 | 2124 | } |
81026794 | 2125 | |
1da177e4 LT |
2126 | if (!nr_moved) { |
2127 | schedstat_inc(sd, lb_failed[idle]); | |
2128 | sd->nr_balance_failed++; | |
2129 | ||
2130 | if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) { | |
1da177e4 LT |
2131 | |
2132 | spin_lock(&busiest->lock); | |
fa3b6ddc SS |
2133 | |
2134 | /* don't kick the migration_thread, if the curr | |
2135 | * task on busiest cpu can't be moved to this_cpu | |
2136 | */ | |
2137 | if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) { | |
2138 | spin_unlock(&busiest->lock); | |
2139 | all_pinned = 1; | |
2140 | goto out_one_pinned; | |
2141 | } | |
2142 | ||
1da177e4 LT |
2143 | if (!busiest->active_balance) { |
2144 | busiest->active_balance = 1; | |
2145 | busiest->push_cpu = this_cpu; | |
81026794 | 2146 | active_balance = 1; |
1da177e4 LT |
2147 | } |
2148 | spin_unlock(&busiest->lock); | |
81026794 | 2149 | if (active_balance) |
1da177e4 LT |
2150 | wake_up_process(busiest->migration_thread); |
2151 | ||
2152 | /* | |
2153 | * We've kicked active balancing, reset the failure | |
2154 | * counter. | |
2155 | */ | |
39507451 | 2156 | sd->nr_balance_failed = sd->cache_nice_tries+1; |
1da177e4 | 2157 | } |
81026794 | 2158 | } else |
1da177e4 LT |
2159 | sd->nr_balance_failed = 0; |
2160 | ||
81026794 | 2161 | if (likely(!active_balance)) { |
1da177e4 LT |
2162 | /* We were unbalanced, so reset the balancing interval */ |
2163 | sd->balance_interval = sd->min_interval; | |
81026794 NP |
2164 | } else { |
2165 | /* | |
2166 | * If we've begun active balancing, start to back off. This | |
2167 | * case may not be covered by the all_pinned logic if there | |
2168 | * is only 1 task on the busy runqueue (because we don't call | |
2169 | * move_tasks). | |
2170 | */ | |
2171 | if (sd->balance_interval < sd->max_interval) | |
2172 | sd->balance_interval *= 2; | |
1da177e4 LT |
2173 | } |
2174 | ||
5969fe06 NP |
2175 | if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
2176 | return -1; | |
1da177e4 LT |
2177 | return nr_moved; |
2178 | ||
2179 | out_balanced: | |
1da177e4 LT |
2180 | schedstat_inc(sd, lb_balanced[idle]); |
2181 | ||
16cfb1c0 | 2182 | sd->nr_balance_failed = 0; |
fa3b6ddc SS |
2183 | |
2184 | out_one_pinned: | |
1da177e4 | 2185 | /* tune up the balancing interval */ |
77391d71 NP |
2186 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || |
2187 | (sd->balance_interval < sd->max_interval)) | |
1da177e4 LT |
2188 | sd->balance_interval *= 2; |
2189 | ||
5969fe06 NP |
2190 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
2191 | return -1; | |
1da177e4 LT |
2192 | return 0; |
2193 | } | |
2194 | ||
2195 | /* | |
2196 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
2197 | * tasks if there is an imbalance. | |
2198 | * | |
2199 | * Called from schedule when this_rq is about to become idle (NEWLY_IDLE). | |
2200 | * this_rq is locked. | |
2201 | */ | |
2202 | static int load_balance_newidle(int this_cpu, runqueue_t *this_rq, | |
2203 | struct sched_domain *sd) | |
2204 | { | |
2205 | struct sched_group *group; | |
2206 | runqueue_t *busiest = NULL; | |
2207 | unsigned long imbalance; | |
2208 | int nr_moved = 0; | |
5969fe06 NP |
2209 | int sd_idle = 0; |
2210 | ||
2211 | if (sd->flags & SD_SHARE_CPUPOWER) | |
2212 | sd_idle = 1; | |
1da177e4 LT |
2213 | |
2214 | schedstat_inc(sd, lb_cnt[NEWLY_IDLE]); | |
5969fe06 | 2215 | group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE, &sd_idle); |
1da177e4 | 2216 | if (!group) { |
1da177e4 | 2217 | schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]); |
16cfb1c0 | 2218 | goto out_balanced; |
1da177e4 LT |
2219 | } |
2220 | ||
2221 | busiest = find_busiest_queue(group); | |
db935dbd | 2222 | if (!busiest) { |
1da177e4 | 2223 | schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]); |
16cfb1c0 | 2224 | goto out_balanced; |
1da177e4 LT |
2225 | } |
2226 | ||
db935dbd NP |
2227 | BUG_ON(busiest == this_rq); |
2228 | ||
1da177e4 | 2229 | schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance); |
d6d5cfaf NP |
2230 | |
2231 | nr_moved = 0; | |
2232 | if (busiest->nr_running > 1) { | |
2233 | /* Attempt to move tasks */ | |
2234 | double_lock_balance(this_rq, busiest); | |
2235 | nr_moved = move_tasks(this_rq, this_cpu, busiest, | |
81026794 | 2236 | imbalance, sd, NEWLY_IDLE, NULL); |
d6d5cfaf NP |
2237 | spin_unlock(&busiest->lock); |
2238 | } | |
2239 | ||
5969fe06 | 2240 | if (!nr_moved) { |
1da177e4 | 2241 | schedstat_inc(sd, lb_failed[NEWLY_IDLE]); |
5969fe06 NP |
2242 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
2243 | return -1; | |
2244 | } else | |
16cfb1c0 | 2245 | sd->nr_balance_failed = 0; |
1da177e4 | 2246 | |
1da177e4 | 2247 | return nr_moved; |
16cfb1c0 NP |
2248 | |
2249 | out_balanced: | |
2250 | schedstat_inc(sd, lb_balanced[NEWLY_IDLE]); | |
5969fe06 NP |
2251 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
2252 | return -1; | |
16cfb1c0 NP |
2253 | sd->nr_balance_failed = 0; |
2254 | return 0; | |
1da177e4 LT |
2255 | } |
2256 | ||
2257 | /* | |
2258 | * idle_balance is called by schedule() if this_cpu is about to become | |
2259 | * idle. Attempts to pull tasks from other CPUs. | |
2260 | */ | |
2261 | static inline void idle_balance(int this_cpu, runqueue_t *this_rq) | |
2262 | { | |
2263 | struct sched_domain *sd; | |
2264 | ||
2265 | for_each_domain(this_cpu, sd) { | |
2266 | if (sd->flags & SD_BALANCE_NEWIDLE) { | |
2267 | if (load_balance_newidle(this_cpu, this_rq, sd)) { | |
2268 | /* We've pulled tasks over so stop searching */ | |
2269 | break; | |
2270 | } | |
2271 | } | |
2272 | } | |
2273 | } | |
2274 | ||
2275 | /* | |
2276 | * active_load_balance is run by migration threads. It pushes running tasks | |
2277 | * off the busiest CPU onto idle CPUs. It requires at least 1 task to be | |
2278 | * running on each physical CPU where possible, and avoids physical / | |
2279 | * logical imbalances. | |
2280 | * | |
2281 | * Called with busiest_rq locked. | |
2282 | */ | |
2283 | static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu) | |
2284 | { | |
2285 | struct sched_domain *sd; | |
1da177e4 | 2286 | runqueue_t *target_rq; |
39507451 NP |
2287 | int target_cpu = busiest_rq->push_cpu; |
2288 | ||
2289 | if (busiest_rq->nr_running <= 1) | |
2290 | /* no task to move */ | |
2291 | return; | |
2292 | ||
2293 | target_rq = cpu_rq(target_cpu); | |
1da177e4 LT |
2294 | |
2295 | /* | |
39507451 NP |
2296 | * This condition is "impossible", if it occurs |
2297 | * we need to fix it. Originally reported by | |
2298 | * Bjorn Helgaas on a 128-cpu setup. | |
1da177e4 | 2299 | */ |
39507451 | 2300 | BUG_ON(busiest_rq == target_rq); |
1da177e4 | 2301 | |
39507451 NP |
2302 | /* move a task from busiest_rq to target_rq */ |
2303 | double_lock_balance(busiest_rq, target_rq); | |
2304 | ||
2305 | /* Search for an sd spanning us and the target CPU. */ | |
2306 | for_each_domain(target_cpu, sd) | |
2307 | if ((sd->flags & SD_LOAD_BALANCE) && | |
2308 | cpu_isset(busiest_cpu, sd->span)) | |
2309 | break; | |
2310 | ||
2311 | if (unlikely(sd == NULL)) | |
2312 | goto out; | |
2313 | ||
2314 | schedstat_inc(sd, alb_cnt); | |
2315 | ||
2316 | if (move_tasks(target_rq, target_cpu, busiest_rq, 1, sd, SCHED_IDLE, NULL)) | |
2317 | schedstat_inc(sd, alb_pushed); | |
2318 | else | |
2319 | schedstat_inc(sd, alb_failed); | |
2320 | out: | |
2321 | spin_unlock(&target_rq->lock); | |
1da177e4 LT |
2322 | } |
2323 | ||
2324 | /* | |
2325 | * rebalance_tick will get called every timer tick, on every CPU. | |
2326 | * | |
2327 | * It checks each scheduling domain to see if it is due to be balanced, | |
2328 | * and initiates a balancing operation if so. | |
2329 | * | |
2330 | * Balancing parameters are set up in arch_init_sched_domains. | |
2331 | */ | |
2332 | ||
2333 | /* Don't have all balancing operations going off at once */ | |
2334 | #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS) | |
2335 | ||
2336 | static void rebalance_tick(int this_cpu, runqueue_t *this_rq, | |
2337 | enum idle_type idle) | |
2338 | { | |
2339 | unsigned long old_load, this_load; | |
2340 | unsigned long j = jiffies + CPU_OFFSET(this_cpu); | |
2341 | struct sched_domain *sd; | |
7897986b | 2342 | int i; |
1da177e4 | 2343 | |
1da177e4 | 2344 | this_load = this_rq->nr_running * SCHED_LOAD_SCALE; |
7897986b NP |
2345 | /* Update our load */ |
2346 | for (i = 0; i < 3; i++) { | |
2347 | unsigned long new_load = this_load; | |
2348 | int scale = 1 << i; | |
2349 | old_load = this_rq->cpu_load[i]; | |
2350 | /* | |
2351 | * Round up the averaging division if load is increasing. This | |
2352 | * prevents us from getting stuck on 9 if the load is 10, for | |
2353 | * example. | |
2354 | */ | |
2355 | if (new_load > old_load) | |
2356 | new_load += scale-1; | |
2357 | this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale; | |
2358 | } | |
1da177e4 LT |
2359 | |
2360 | for_each_domain(this_cpu, sd) { | |
2361 | unsigned long interval; | |
2362 | ||
2363 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
2364 | continue; | |
2365 | ||
2366 | interval = sd->balance_interval; | |
2367 | if (idle != SCHED_IDLE) | |
2368 | interval *= sd->busy_factor; | |
2369 | ||
2370 | /* scale ms to jiffies */ | |
2371 | interval = msecs_to_jiffies(interval); | |
2372 | if (unlikely(!interval)) | |
2373 | interval = 1; | |
2374 | ||
2375 | if (j - sd->last_balance >= interval) { | |
2376 | if (load_balance(this_cpu, this_rq, sd, idle)) { | |
fa3b6ddc SS |
2377 | /* |
2378 | * We've pulled tasks over so either we're no | |
5969fe06 NP |
2379 | * longer idle, or one of our SMT siblings is |
2380 | * not idle. | |
2381 | */ | |
1da177e4 LT |
2382 | idle = NOT_IDLE; |
2383 | } | |
2384 | sd->last_balance += interval; | |
2385 | } | |
2386 | } | |
2387 | } | |
2388 | #else | |
2389 | /* | |
2390 | * on UP we do not need to balance between CPUs: | |
2391 | */ | |
2392 | static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle) | |
2393 | { | |
2394 | } | |
2395 | static inline void idle_balance(int cpu, runqueue_t *rq) | |
2396 | { | |
2397 | } | |
2398 | #endif | |
2399 | ||
2400 | static inline int wake_priority_sleeper(runqueue_t *rq) | |
2401 | { | |
2402 | int ret = 0; | |
2403 | #ifdef CONFIG_SCHED_SMT | |
2404 | spin_lock(&rq->lock); | |
2405 | /* | |
2406 | * If an SMT sibling task has been put to sleep for priority | |
2407 | * reasons reschedule the idle task to see if it can now run. | |
2408 | */ | |
2409 | if (rq->nr_running) { | |
2410 | resched_task(rq->idle); | |
2411 | ret = 1; | |
2412 | } | |
2413 | spin_unlock(&rq->lock); | |
2414 | #endif | |
2415 | return ret; | |
2416 | } | |
2417 | ||
2418 | DEFINE_PER_CPU(struct kernel_stat, kstat); | |
2419 | ||
2420 | EXPORT_PER_CPU_SYMBOL(kstat); | |
2421 | ||
2422 | /* | |
2423 | * This is called on clock ticks and on context switches. | |
2424 | * Bank in p->sched_time the ns elapsed since the last tick or switch. | |
2425 | */ | |
2426 | static inline void update_cpu_clock(task_t *p, runqueue_t *rq, | |
2427 | unsigned long long now) | |
2428 | { | |
2429 | unsigned long long last = max(p->timestamp, rq->timestamp_last_tick); | |
2430 | p->sched_time += now - last; | |
2431 | } | |
2432 | ||
2433 | /* | |
2434 | * Return current->sched_time plus any more ns on the sched_clock | |
2435 | * that have not yet been banked. | |
2436 | */ | |
2437 | unsigned long long current_sched_time(const task_t *tsk) | |
2438 | { | |
2439 | unsigned long long ns; | |
2440 | unsigned long flags; | |
2441 | local_irq_save(flags); | |
2442 | ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick); | |
2443 | ns = tsk->sched_time + (sched_clock() - ns); | |
2444 | local_irq_restore(flags); | |
2445 | return ns; | |
2446 | } | |
2447 | ||
2448 | /* | |
2449 | * We place interactive tasks back into the active array, if possible. | |
2450 | * | |
2451 | * To guarantee that this does not starve expired tasks we ignore the | |
2452 | * interactivity of a task if the first expired task had to wait more | |
2453 | * than a 'reasonable' amount of time. This deadline timeout is | |
2454 | * load-dependent, as the frequency of array switched decreases with | |
2455 | * increasing number of running tasks. We also ignore the interactivity | |
2456 | * if a better static_prio task has expired: | |
2457 | */ | |
2458 | #define EXPIRED_STARVING(rq) \ | |
2459 | ((STARVATION_LIMIT && ((rq)->expired_timestamp && \ | |
2460 | (jiffies - (rq)->expired_timestamp >= \ | |
2461 | STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \ | |
2462 | ((rq)->curr->static_prio > (rq)->best_expired_prio)) | |
2463 | ||
2464 | /* | |
2465 | * Account user cpu time to a process. | |
2466 | * @p: the process that the cpu time gets accounted to | |
2467 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
2468 | * @cputime: the cpu time spent in user space since the last update | |
2469 | */ | |
2470 | void account_user_time(struct task_struct *p, cputime_t cputime) | |
2471 | { | |
2472 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
2473 | cputime64_t tmp; | |
2474 | ||
2475 | p->utime = cputime_add(p->utime, cputime); | |
2476 | ||
2477 | /* Add user time to cpustat. */ | |
2478 | tmp = cputime_to_cputime64(cputime); | |
2479 | if (TASK_NICE(p) > 0) | |
2480 | cpustat->nice = cputime64_add(cpustat->nice, tmp); | |
2481 | else | |
2482 | cpustat->user = cputime64_add(cpustat->user, tmp); | |
2483 | } | |
2484 | ||
2485 | /* | |
2486 | * Account system cpu time to a process. | |
2487 | * @p: the process that the cpu time gets accounted to | |
2488 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
2489 | * @cputime: the cpu time spent in kernel space since the last update | |
2490 | */ | |
2491 | void account_system_time(struct task_struct *p, int hardirq_offset, | |
2492 | cputime_t cputime) | |
2493 | { | |
2494 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
2495 | runqueue_t *rq = this_rq(); | |
2496 | cputime64_t tmp; | |
2497 | ||
2498 | p->stime = cputime_add(p->stime, cputime); | |
2499 | ||
2500 | /* Add system time to cpustat. */ | |
2501 | tmp = cputime_to_cputime64(cputime); | |
2502 | if (hardirq_count() - hardirq_offset) | |
2503 | cpustat->irq = cputime64_add(cpustat->irq, tmp); | |
2504 | else if (softirq_count()) | |
2505 | cpustat->softirq = cputime64_add(cpustat->softirq, tmp); | |
2506 | else if (p != rq->idle) | |
2507 | cpustat->system = cputime64_add(cpustat->system, tmp); | |
2508 | else if (atomic_read(&rq->nr_iowait) > 0) | |
2509 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); | |
2510 | else | |
2511 | cpustat->idle = cputime64_add(cpustat->idle, tmp); | |
2512 | /* Account for system time used */ | |
2513 | acct_update_integrals(p); | |
2514 | /* Update rss highwater mark */ | |
2515 | update_mem_hiwater(p); | |
2516 | } | |
2517 | ||
2518 | /* | |
2519 | * Account for involuntary wait time. | |
2520 | * @p: the process from which the cpu time has been stolen | |
2521 | * @steal: the cpu time spent in involuntary wait | |
2522 | */ | |
2523 | void account_steal_time(struct task_struct *p, cputime_t steal) | |
2524 | { | |
2525 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
2526 | cputime64_t tmp = cputime_to_cputime64(steal); | |
2527 | runqueue_t *rq = this_rq(); | |
2528 | ||
2529 | if (p == rq->idle) { | |
2530 | p->stime = cputime_add(p->stime, steal); | |
2531 | if (atomic_read(&rq->nr_iowait) > 0) | |
2532 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); | |
2533 | else | |
2534 | cpustat->idle = cputime64_add(cpustat->idle, tmp); | |
2535 | } else | |
2536 | cpustat->steal = cputime64_add(cpustat->steal, tmp); | |
2537 | } | |
2538 | ||
2539 | /* | |
2540 | * This function gets called by the timer code, with HZ frequency. | |
2541 | * We call it with interrupts disabled. | |
2542 | * | |
2543 | * It also gets called by the fork code, when changing the parent's | |
2544 | * timeslices. | |
2545 | */ | |
2546 | void scheduler_tick(void) | |
2547 | { | |
2548 | int cpu = smp_processor_id(); | |
2549 | runqueue_t *rq = this_rq(); | |
2550 | task_t *p = current; | |
2551 | unsigned long long now = sched_clock(); | |
2552 | ||
2553 | update_cpu_clock(p, rq, now); | |
2554 | ||
2555 | rq->timestamp_last_tick = now; | |
2556 | ||
2557 | if (p == rq->idle) { | |
2558 | if (wake_priority_sleeper(rq)) | |
2559 | goto out; | |
2560 | rebalance_tick(cpu, rq, SCHED_IDLE); | |
2561 | return; | |
2562 | } | |
2563 | ||
2564 | /* Task might have expired already, but not scheduled off yet */ | |
2565 | if (p->array != rq->active) { | |
2566 | set_tsk_need_resched(p); | |
2567 | goto out; | |
2568 | } | |
2569 | spin_lock(&rq->lock); | |
2570 | /* | |
2571 | * The task was running during this tick - update the | |
2572 | * time slice counter. Note: we do not update a thread's | |
2573 | * priority until it either goes to sleep or uses up its | |
2574 | * timeslice. This makes it possible for interactive tasks | |
2575 | * to use up their timeslices at their highest priority levels. | |
2576 | */ | |
2577 | if (rt_task(p)) { | |
2578 | /* | |
2579 | * RR tasks need a special form of timeslice management. | |
2580 | * FIFO tasks have no timeslices. | |
2581 | */ | |
2582 | if ((p->policy == SCHED_RR) && !--p->time_slice) { | |
2583 | p->time_slice = task_timeslice(p); | |
2584 | p->first_time_slice = 0; | |
2585 | set_tsk_need_resched(p); | |
2586 | ||
2587 | /* put it at the end of the queue: */ | |
2588 | requeue_task(p, rq->active); | |
2589 | } | |
2590 | goto out_unlock; | |
2591 | } | |
2592 | if (!--p->time_slice) { | |
2593 | dequeue_task(p, rq->active); | |
2594 | set_tsk_need_resched(p); | |
2595 | p->prio = effective_prio(p); | |
2596 | p->time_slice = task_timeslice(p); | |
2597 | p->first_time_slice = 0; | |
2598 | ||
2599 | if (!rq->expired_timestamp) | |
2600 | rq->expired_timestamp = jiffies; | |
2601 | if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) { | |
2602 | enqueue_task(p, rq->expired); | |
2603 | if (p->static_prio < rq->best_expired_prio) | |
2604 | rq->best_expired_prio = p->static_prio; | |
2605 | } else | |
2606 | enqueue_task(p, rq->active); | |
2607 | } else { | |
2608 | /* | |
2609 | * Prevent a too long timeslice allowing a task to monopolize | |
2610 | * the CPU. We do this by splitting up the timeslice into | |
2611 | * smaller pieces. | |
2612 | * | |
2613 | * Note: this does not mean the task's timeslices expire or | |
2614 | * get lost in any way, they just might be preempted by | |
2615 | * another task of equal priority. (one with higher | |
2616 | * priority would have preempted this task already.) We | |
2617 | * requeue this task to the end of the list on this priority | |
2618 | * level, which is in essence a round-robin of tasks with | |
2619 | * equal priority. | |
2620 | * | |
2621 | * This only applies to tasks in the interactive | |
2622 | * delta range with at least TIMESLICE_GRANULARITY to requeue. | |
2623 | */ | |
2624 | if (TASK_INTERACTIVE(p) && !((task_timeslice(p) - | |
2625 | p->time_slice) % TIMESLICE_GRANULARITY(p)) && | |
2626 | (p->time_slice >= TIMESLICE_GRANULARITY(p)) && | |
2627 | (p->array == rq->active)) { | |
2628 | ||
2629 | requeue_task(p, rq->active); | |
2630 | set_tsk_need_resched(p); | |
2631 | } | |
2632 | } | |
2633 | out_unlock: | |
2634 | spin_unlock(&rq->lock); | |
2635 | out: | |
2636 | rebalance_tick(cpu, rq, NOT_IDLE); | |
2637 | } | |
2638 | ||
2639 | #ifdef CONFIG_SCHED_SMT | |
fc38ed75 CK |
2640 | static inline void wakeup_busy_runqueue(runqueue_t *rq) |
2641 | { | |
2642 | /* If an SMT runqueue is sleeping due to priority reasons wake it up */ | |
2643 | if (rq->curr == rq->idle && rq->nr_running) | |
2644 | resched_task(rq->idle); | |
2645 | } | |
2646 | ||
1da177e4 LT |
2647 | static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq) |
2648 | { | |
41c7ce9a | 2649 | struct sched_domain *tmp, *sd = NULL; |
1da177e4 LT |
2650 | cpumask_t sibling_map; |
2651 | int i; | |
2652 | ||
41c7ce9a NP |
2653 | for_each_domain(this_cpu, tmp) |
2654 | if (tmp->flags & SD_SHARE_CPUPOWER) | |
2655 | sd = tmp; | |
2656 | ||
2657 | if (!sd) | |
1da177e4 LT |
2658 | return; |
2659 | ||
2660 | /* | |
2661 | * Unlock the current runqueue because we have to lock in | |
2662 | * CPU order to avoid deadlocks. Caller knows that we might | |
2663 | * unlock. We keep IRQs disabled. | |
2664 | */ | |
2665 | spin_unlock(&this_rq->lock); | |
2666 | ||
2667 | sibling_map = sd->span; | |
2668 | ||
2669 | for_each_cpu_mask(i, sibling_map) | |
2670 | spin_lock(&cpu_rq(i)->lock); | |
2671 | /* | |
2672 | * We clear this CPU from the mask. This both simplifies the | |
2673 | * inner loop and keps this_rq locked when we exit: | |
2674 | */ | |
2675 | cpu_clear(this_cpu, sibling_map); | |
2676 | ||
2677 | for_each_cpu_mask(i, sibling_map) { | |
2678 | runqueue_t *smt_rq = cpu_rq(i); | |
2679 | ||
fc38ed75 | 2680 | wakeup_busy_runqueue(smt_rq); |
1da177e4 LT |
2681 | } |
2682 | ||
2683 | for_each_cpu_mask(i, sibling_map) | |
2684 | spin_unlock(&cpu_rq(i)->lock); | |
2685 | /* | |
2686 | * We exit with this_cpu's rq still held and IRQs | |
2687 | * still disabled: | |
2688 | */ | |
2689 | } | |
2690 | ||
67f9a619 IM |
2691 | /* |
2692 | * number of 'lost' timeslices this task wont be able to fully | |
2693 | * utilize, if another task runs on a sibling. This models the | |
2694 | * slowdown effect of other tasks running on siblings: | |
2695 | */ | |
2696 | static inline unsigned long smt_slice(task_t *p, struct sched_domain *sd) | |
2697 | { | |
2698 | return p->time_slice * (100 - sd->per_cpu_gain) / 100; | |
2699 | } | |
2700 | ||
1da177e4 LT |
2701 | static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq) |
2702 | { | |
41c7ce9a | 2703 | struct sched_domain *tmp, *sd = NULL; |
1da177e4 LT |
2704 | cpumask_t sibling_map; |
2705 | prio_array_t *array; | |
2706 | int ret = 0, i; | |
2707 | task_t *p; | |
2708 | ||
41c7ce9a NP |
2709 | for_each_domain(this_cpu, tmp) |
2710 | if (tmp->flags & SD_SHARE_CPUPOWER) | |
2711 | sd = tmp; | |
2712 | ||
2713 | if (!sd) | |
1da177e4 LT |
2714 | return 0; |
2715 | ||
2716 | /* | |
2717 | * The same locking rules and details apply as for | |
2718 | * wake_sleeping_dependent(): | |
2719 | */ | |
2720 | spin_unlock(&this_rq->lock); | |
2721 | sibling_map = sd->span; | |
2722 | for_each_cpu_mask(i, sibling_map) | |
2723 | spin_lock(&cpu_rq(i)->lock); | |
2724 | cpu_clear(this_cpu, sibling_map); | |
2725 | ||
2726 | /* | |
2727 | * Establish next task to be run - it might have gone away because | |
2728 | * we released the runqueue lock above: | |
2729 | */ | |
2730 | if (!this_rq->nr_running) | |
2731 | goto out_unlock; | |
2732 | array = this_rq->active; | |
2733 | if (!array->nr_active) | |
2734 | array = this_rq->expired; | |
2735 | BUG_ON(!array->nr_active); | |
2736 | ||
2737 | p = list_entry(array->queue[sched_find_first_bit(array->bitmap)].next, | |
2738 | task_t, run_list); | |
2739 | ||
2740 | for_each_cpu_mask(i, sibling_map) { | |
2741 | runqueue_t *smt_rq = cpu_rq(i); | |
2742 | task_t *smt_curr = smt_rq->curr; | |
2743 | ||
fc38ed75 CK |
2744 | /* Kernel threads do not participate in dependent sleeping */ |
2745 | if (!p->mm || !smt_curr->mm || rt_task(p)) | |
2746 | goto check_smt_task; | |
2747 | ||
1da177e4 LT |
2748 | /* |
2749 | * If a user task with lower static priority than the | |
2750 | * running task on the SMT sibling is trying to schedule, | |
2751 | * delay it till there is proportionately less timeslice | |
2752 | * left of the sibling task to prevent a lower priority | |
2753 | * task from using an unfair proportion of the | |
2754 | * physical cpu's resources. -ck | |
2755 | */ | |
fc38ed75 CK |
2756 | if (rt_task(smt_curr)) { |
2757 | /* | |
2758 | * With real time tasks we run non-rt tasks only | |
2759 | * per_cpu_gain% of the time. | |
2760 | */ | |
2761 | if ((jiffies % DEF_TIMESLICE) > | |
2762 | (sd->per_cpu_gain * DEF_TIMESLICE / 100)) | |
2763 | ret = 1; | |
2764 | } else | |
67f9a619 IM |
2765 | if (smt_curr->static_prio < p->static_prio && |
2766 | !TASK_PREEMPTS_CURR(p, smt_rq) && | |
2767 | smt_slice(smt_curr, sd) > task_timeslice(p)) | |
fc38ed75 CK |
2768 | ret = 1; |
2769 | ||
2770 | check_smt_task: | |
2771 | if ((!smt_curr->mm && smt_curr != smt_rq->idle) || | |
2772 | rt_task(smt_curr)) | |
2773 | continue; | |
2774 | if (!p->mm) { | |
2775 | wakeup_busy_runqueue(smt_rq); | |
2776 | continue; | |
2777 | } | |
1da177e4 LT |
2778 | |
2779 | /* | |
fc38ed75 CK |
2780 | * Reschedule a lower priority task on the SMT sibling for |
2781 | * it to be put to sleep, or wake it up if it has been put to | |
2782 | * sleep for priority reasons to see if it should run now. | |
1da177e4 | 2783 | */ |
fc38ed75 CK |
2784 | if (rt_task(p)) { |
2785 | if ((jiffies % DEF_TIMESLICE) > | |
2786 | (sd->per_cpu_gain * DEF_TIMESLICE / 100)) | |
2787 | resched_task(smt_curr); | |
2788 | } else { | |
67f9a619 IM |
2789 | if (TASK_PREEMPTS_CURR(p, smt_rq) && |
2790 | smt_slice(p, sd) > task_timeslice(smt_curr)) | |
fc38ed75 CK |
2791 | resched_task(smt_curr); |
2792 | else | |
2793 | wakeup_busy_runqueue(smt_rq); | |
2794 | } | |
1da177e4 LT |
2795 | } |
2796 | out_unlock: | |
2797 | for_each_cpu_mask(i, sibling_map) | |
2798 | spin_unlock(&cpu_rq(i)->lock); | |
2799 | return ret; | |
2800 | } | |
2801 | #else | |
2802 | static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq) | |
2803 | { | |
2804 | } | |
2805 | ||
2806 | static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq) | |
2807 | { | |
2808 | return 0; | |
2809 | } | |
2810 | #endif | |
2811 | ||
2812 | #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT) | |
2813 | ||
2814 | void fastcall add_preempt_count(int val) | |
2815 | { | |
2816 | /* | |
2817 | * Underflow? | |
2818 | */ | |
be5b4fbd | 2819 | BUG_ON((preempt_count() < 0)); |
1da177e4 LT |
2820 | preempt_count() += val; |
2821 | /* | |
2822 | * Spinlock count overflowing soon? | |
2823 | */ | |
2824 | BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10); | |
2825 | } | |
2826 | EXPORT_SYMBOL(add_preempt_count); | |
2827 | ||
2828 | void fastcall sub_preempt_count(int val) | |
2829 | { | |
2830 | /* | |
2831 | * Underflow? | |
2832 | */ | |
2833 | BUG_ON(val > preempt_count()); | |
2834 | /* | |
2835 | * Is the spinlock portion underflowing? | |
2836 | */ | |
2837 | BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK)); | |
2838 | preempt_count() -= val; | |
2839 | } | |
2840 | EXPORT_SYMBOL(sub_preempt_count); | |
2841 | ||
2842 | #endif | |
2843 | ||
2844 | /* | |
2845 | * schedule() is the main scheduler function. | |
2846 | */ | |
2847 | asmlinkage void __sched schedule(void) | |
2848 | { | |
2849 | long *switch_count; | |
2850 | task_t *prev, *next; | |
2851 | runqueue_t *rq; | |
2852 | prio_array_t *array; | |
2853 | struct list_head *queue; | |
2854 | unsigned long long now; | |
2855 | unsigned long run_time; | |
a3464a10 | 2856 | int cpu, idx, new_prio; |
1da177e4 LT |
2857 | |
2858 | /* | |
2859 | * Test if we are atomic. Since do_exit() needs to call into | |
2860 | * schedule() atomically, we ignore that path for now. | |
2861 | * Otherwise, whine if we are scheduling when we should not be. | |
2862 | */ | |
2863 | if (likely(!current->exit_state)) { | |
2864 | if (unlikely(in_atomic())) { | |
2865 | printk(KERN_ERR "scheduling while atomic: " | |
2866 | "%s/0x%08x/%d\n", | |
2867 | current->comm, preempt_count(), current->pid); | |
2868 | dump_stack(); | |
2869 | } | |
2870 | } | |
2871 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); | |
2872 | ||
2873 | need_resched: | |
2874 | preempt_disable(); | |
2875 | prev = current; | |
2876 | release_kernel_lock(prev); | |
2877 | need_resched_nonpreemptible: | |
2878 | rq = this_rq(); | |
2879 | ||
2880 | /* | |
2881 | * The idle thread is not allowed to schedule! | |
2882 | * Remove this check after it has been exercised a bit. | |
2883 | */ | |
2884 | if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) { | |
2885 | printk(KERN_ERR "bad: scheduling from the idle thread!\n"); | |
2886 | dump_stack(); | |
2887 | } | |
2888 | ||
2889 | schedstat_inc(rq, sched_cnt); | |
2890 | now = sched_clock(); | |
238628ed | 2891 | if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) { |
1da177e4 | 2892 | run_time = now - prev->timestamp; |
238628ed | 2893 | if (unlikely((long long)(now - prev->timestamp) < 0)) |
1da177e4 LT |
2894 | run_time = 0; |
2895 | } else | |
2896 | run_time = NS_MAX_SLEEP_AVG; | |
2897 | ||
2898 | /* | |
2899 | * Tasks charged proportionately less run_time at high sleep_avg to | |
2900 | * delay them losing their interactive status | |
2901 | */ | |
2902 | run_time /= (CURRENT_BONUS(prev) ? : 1); | |
2903 | ||
2904 | spin_lock_irq(&rq->lock); | |
2905 | ||
2906 | if (unlikely(prev->flags & PF_DEAD)) | |
2907 | prev->state = EXIT_DEAD; | |
2908 | ||
2909 | switch_count = &prev->nivcsw; | |
2910 | if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { | |
2911 | switch_count = &prev->nvcsw; | |
2912 | if (unlikely((prev->state & TASK_INTERRUPTIBLE) && | |
2913 | unlikely(signal_pending(prev)))) | |
2914 | prev->state = TASK_RUNNING; | |
2915 | else { | |
2916 | if (prev->state == TASK_UNINTERRUPTIBLE) | |
2917 | rq->nr_uninterruptible++; | |
2918 | deactivate_task(prev, rq); | |
2919 | } | |
2920 | } | |
2921 | ||
2922 | cpu = smp_processor_id(); | |
2923 | if (unlikely(!rq->nr_running)) { | |
2924 | go_idle: | |
2925 | idle_balance(cpu, rq); | |
2926 | if (!rq->nr_running) { | |
2927 | next = rq->idle; | |
2928 | rq->expired_timestamp = 0; | |
2929 | wake_sleeping_dependent(cpu, rq); | |
2930 | /* | |
2931 | * wake_sleeping_dependent() might have released | |
2932 | * the runqueue, so break out if we got new | |
2933 | * tasks meanwhile: | |
2934 | */ | |
2935 | if (!rq->nr_running) | |
2936 | goto switch_tasks; | |
2937 | } | |
2938 | } else { | |
2939 | if (dependent_sleeper(cpu, rq)) { | |
2940 | next = rq->idle; | |
2941 | goto switch_tasks; | |
2942 | } | |
2943 | /* | |
2944 | * dependent_sleeper() releases and reacquires the runqueue | |
2945 | * lock, hence go into the idle loop if the rq went | |
2946 | * empty meanwhile: | |
2947 | */ | |
2948 | if (unlikely(!rq->nr_running)) | |
2949 | goto go_idle; | |
2950 | } | |
2951 | ||
2952 | array = rq->active; | |
2953 | if (unlikely(!array->nr_active)) { | |
2954 | /* | |
2955 | * Switch the active and expired arrays. | |
2956 | */ | |
2957 | schedstat_inc(rq, sched_switch); | |
2958 | rq->active = rq->expired; | |
2959 | rq->expired = array; | |
2960 | array = rq->active; | |
2961 | rq->expired_timestamp = 0; | |
2962 | rq->best_expired_prio = MAX_PRIO; | |
2963 | } | |
2964 | ||
2965 | idx = sched_find_first_bit(array->bitmap); | |
2966 | queue = array->queue + idx; | |
2967 | next = list_entry(queue->next, task_t, run_list); | |
2968 | ||
2969 | if (!rt_task(next) && next->activated > 0) { | |
2970 | unsigned long long delta = now - next->timestamp; | |
238628ed | 2971 | if (unlikely((long long)(now - next->timestamp) < 0)) |
1da177e4 LT |
2972 | delta = 0; |
2973 | ||
2974 | if (next->activated == 1) | |
2975 | delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128; | |
2976 | ||
2977 | array = next->array; | |
a3464a10 CS |
2978 | new_prio = recalc_task_prio(next, next->timestamp + delta); |
2979 | ||
2980 | if (unlikely(next->prio != new_prio)) { | |
2981 | dequeue_task(next, array); | |
2982 | next->prio = new_prio; | |
2983 | enqueue_task(next, array); | |
2984 | } else | |
2985 | requeue_task(next, array); | |
1da177e4 LT |
2986 | } |
2987 | next->activated = 0; | |
2988 | switch_tasks: | |
2989 | if (next == rq->idle) | |
2990 | schedstat_inc(rq, sched_goidle); | |
2991 | prefetch(next); | |
383f2835 | 2992 | prefetch_stack(next); |
1da177e4 LT |
2993 | clear_tsk_need_resched(prev); |
2994 | rcu_qsctr_inc(task_cpu(prev)); | |
2995 | ||
2996 | update_cpu_clock(prev, rq, now); | |
2997 | ||
2998 | prev->sleep_avg -= run_time; | |
2999 | if ((long)prev->sleep_avg <= 0) | |
3000 | prev->sleep_avg = 0; | |
3001 | prev->timestamp = prev->last_ran = now; | |
3002 | ||
3003 | sched_info_switch(prev, next); | |
3004 | if (likely(prev != next)) { | |
3005 | next->timestamp = now; | |
3006 | rq->nr_switches++; | |
3007 | rq->curr = next; | |
3008 | ++*switch_count; | |
3009 | ||
4866cde0 | 3010 | prepare_task_switch(rq, next); |
1da177e4 LT |
3011 | prev = context_switch(rq, prev, next); |
3012 | barrier(); | |
4866cde0 NP |
3013 | /* |
3014 | * this_rq must be evaluated again because prev may have moved | |
3015 | * CPUs since it called schedule(), thus the 'rq' on its stack | |
3016 | * frame will be invalid. | |
3017 | */ | |
3018 | finish_task_switch(this_rq(), prev); | |
1da177e4 LT |
3019 | } else |
3020 | spin_unlock_irq(&rq->lock); | |
3021 | ||
3022 | prev = current; | |
3023 | if (unlikely(reacquire_kernel_lock(prev) < 0)) | |
3024 | goto need_resched_nonpreemptible; | |
3025 | preempt_enable_no_resched(); | |
3026 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
3027 | goto need_resched; | |
3028 | } | |
3029 | ||
3030 | EXPORT_SYMBOL(schedule); | |
3031 | ||
3032 | #ifdef CONFIG_PREEMPT | |
3033 | /* | |
3034 | * this is is the entry point to schedule() from in-kernel preemption | |
3035 | * off of preempt_enable. Kernel preemptions off return from interrupt | |
3036 | * occur there and call schedule directly. | |
3037 | */ | |
3038 | asmlinkage void __sched preempt_schedule(void) | |
3039 | { | |
3040 | struct thread_info *ti = current_thread_info(); | |
3041 | #ifdef CONFIG_PREEMPT_BKL | |
3042 | struct task_struct *task = current; | |
3043 | int saved_lock_depth; | |
3044 | #endif | |
3045 | /* | |
3046 | * If there is a non-zero preempt_count or interrupts are disabled, | |
3047 | * we do not want to preempt the current task. Just return.. | |
3048 | */ | |
3049 | if (unlikely(ti->preempt_count || irqs_disabled())) | |
3050 | return; | |
3051 | ||
3052 | need_resched: | |
3053 | add_preempt_count(PREEMPT_ACTIVE); | |
3054 | /* | |
3055 | * We keep the big kernel semaphore locked, but we | |
3056 | * clear ->lock_depth so that schedule() doesnt | |
3057 | * auto-release the semaphore: | |
3058 | */ | |
3059 | #ifdef CONFIG_PREEMPT_BKL | |
3060 | saved_lock_depth = task->lock_depth; | |
3061 | task->lock_depth = -1; | |
3062 | #endif | |
3063 | schedule(); | |
3064 | #ifdef CONFIG_PREEMPT_BKL | |
3065 | task->lock_depth = saved_lock_depth; | |
3066 | #endif | |
3067 | sub_preempt_count(PREEMPT_ACTIVE); | |
3068 | ||
3069 | /* we could miss a preemption opportunity between schedule and now */ | |
3070 | barrier(); | |
3071 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
3072 | goto need_resched; | |
3073 | } | |
3074 | ||
3075 | EXPORT_SYMBOL(preempt_schedule); | |
3076 | ||
3077 | /* | |
3078 | * this is is the entry point to schedule() from kernel preemption | |
3079 | * off of irq context. | |
3080 | * Note, that this is called and return with irqs disabled. This will | |
3081 | * protect us against recursive calling from irq. | |
3082 | */ | |
3083 | asmlinkage void __sched preempt_schedule_irq(void) | |
3084 | { | |
3085 | struct thread_info *ti = current_thread_info(); | |
3086 | #ifdef CONFIG_PREEMPT_BKL | |
3087 | struct task_struct *task = current; | |
3088 | int saved_lock_depth; | |
3089 | #endif | |
3090 | /* Catch callers which need to be fixed*/ | |
3091 | BUG_ON(ti->preempt_count || !irqs_disabled()); | |
3092 | ||
3093 | need_resched: | |
3094 | add_preempt_count(PREEMPT_ACTIVE); | |
3095 | /* | |
3096 | * We keep the big kernel semaphore locked, but we | |
3097 | * clear ->lock_depth so that schedule() doesnt | |
3098 | * auto-release the semaphore: | |
3099 | */ | |
3100 | #ifdef CONFIG_PREEMPT_BKL | |
3101 | saved_lock_depth = task->lock_depth; | |
3102 | task->lock_depth = -1; | |
3103 | #endif | |
3104 | local_irq_enable(); | |
3105 | schedule(); | |
3106 | local_irq_disable(); | |
3107 | #ifdef CONFIG_PREEMPT_BKL | |
3108 | task->lock_depth = saved_lock_depth; | |
3109 | #endif | |
3110 | sub_preempt_count(PREEMPT_ACTIVE); | |
3111 | ||
3112 | /* we could miss a preemption opportunity between schedule and now */ | |
3113 | barrier(); | |
3114 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
3115 | goto need_resched; | |
3116 | } | |
3117 | ||
3118 | #endif /* CONFIG_PREEMPT */ | |
3119 | ||
95cdf3b7 IM |
3120 | int default_wake_function(wait_queue_t *curr, unsigned mode, int sync, |
3121 | void *key) | |
1da177e4 | 3122 | { |
c43dc2fd | 3123 | task_t *p = curr->private; |
1da177e4 LT |
3124 | return try_to_wake_up(p, mode, sync); |
3125 | } | |
3126 | ||
3127 | EXPORT_SYMBOL(default_wake_function); | |
3128 | ||
3129 | /* | |
3130 | * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just | |
3131 | * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve | |
3132 | * number) then we wake all the non-exclusive tasks and one exclusive task. | |
3133 | * | |
3134 | * There are circumstances in which we can try to wake a task which has already | |
3135 | * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns | |
3136 | * zero in this (rare) case, and we handle it by continuing to scan the queue. | |
3137 | */ | |
3138 | static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, | |
3139 | int nr_exclusive, int sync, void *key) | |
3140 | { | |
3141 | struct list_head *tmp, *next; | |
3142 | ||
3143 | list_for_each_safe(tmp, next, &q->task_list) { | |
3144 | wait_queue_t *curr; | |
3145 | unsigned flags; | |
3146 | curr = list_entry(tmp, wait_queue_t, task_list); | |
3147 | flags = curr->flags; | |
3148 | if (curr->func(curr, mode, sync, key) && | |
3149 | (flags & WQ_FLAG_EXCLUSIVE) && | |
3150 | !--nr_exclusive) | |
3151 | break; | |
3152 | } | |
3153 | } | |
3154 | ||
3155 | /** | |
3156 | * __wake_up - wake up threads blocked on a waitqueue. | |
3157 | * @q: the waitqueue | |
3158 | * @mode: which threads | |
3159 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
67be2dd1 | 3160 | * @key: is directly passed to the wakeup function |
1da177e4 LT |
3161 | */ |
3162 | void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode, | |
95cdf3b7 | 3163 | int nr_exclusive, void *key) |
1da177e4 LT |
3164 | { |
3165 | unsigned long flags; | |
3166 | ||
3167 | spin_lock_irqsave(&q->lock, flags); | |
3168 | __wake_up_common(q, mode, nr_exclusive, 0, key); | |
3169 | spin_unlock_irqrestore(&q->lock, flags); | |
3170 | } | |
3171 | ||
3172 | EXPORT_SYMBOL(__wake_up); | |
3173 | ||
3174 | /* | |
3175 | * Same as __wake_up but called with the spinlock in wait_queue_head_t held. | |
3176 | */ | |
3177 | void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode) | |
3178 | { | |
3179 | __wake_up_common(q, mode, 1, 0, NULL); | |
3180 | } | |
3181 | ||
3182 | /** | |
67be2dd1 | 3183 | * __wake_up_sync - wake up threads blocked on a waitqueue. |
1da177e4 LT |
3184 | * @q: the waitqueue |
3185 | * @mode: which threads | |
3186 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
3187 | * | |
3188 | * The sync wakeup differs that the waker knows that it will schedule | |
3189 | * away soon, so while the target thread will be woken up, it will not | |
3190 | * be migrated to another CPU - ie. the two threads are 'synchronized' | |
3191 | * with each other. This can prevent needless bouncing between CPUs. | |
3192 | * | |
3193 | * On UP it can prevent extra preemption. | |
3194 | */ | |
95cdf3b7 IM |
3195 | void fastcall |
3196 | __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) | |
1da177e4 LT |
3197 | { |
3198 | unsigned long flags; | |
3199 | int sync = 1; | |
3200 | ||
3201 | if (unlikely(!q)) | |
3202 | return; | |
3203 | ||
3204 | if (unlikely(!nr_exclusive)) | |
3205 | sync = 0; | |
3206 | ||
3207 | spin_lock_irqsave(&q->lock, flags); | |
3208 | __wake_up_common(q, mode, nr_exclusive, sync, NULL); | |
3209 | spin_unlock_irqrestore(&q->lock, flags); | |
3210 | } | |
3211 | EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ | |
3212 | ||
3213 | void fastcall complete(struct completion *x) | |
3214 | { | |
3215 | unsigned long flags; | |
3216 | ||
3217 | spin_lock_irqsave(&x->wait.lock, flags); | |
3218 | x->done++; | |
3219 | __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, | |
3220 | 1, 0, NULL); | |
3221 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
3222 | } | |
3223 | EXPORT_SYMBOL(complete); | |
3224 | ||
3225 | void fastcall complete_all(struct completion *x) | |
3226 | { | |
3227 | unsigned long flags; | |
3228 | ||
3229 | spin_lock_irqsave(&x->wait.lock, flags); | |
3230 | x->done += UINT_MAX/2; | |
3231 | __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, | |
3232 | 0, 0, NULL); | |
3233 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
3234 | } | |
3235 | EXPORT_SYMBOL(complete_all); | |
3236 | ||
3237 | void fastcall __sched wait_for_completion(struct completion *x) | |
3238 | { | |
3239 | might_sleep(); | |
3240 | spin_lock_irq(&x->wait.lock); | |
3241 | if (!x->done) { | |
3242 | DECLARE_WAITQUEUE(wait, current); | |
3243 | ||
3244 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3245 | __add_wait_queue_tail(&x->wait, &wait); | |
3246 | do { | |
3247 | __set_current_state(TASK_UNINTERRUPTIBLE); | |
3248 | spin_unlock_irq(&x->wait.lock); | |
3249 | schedule(); | |
3250 | spin_lock_irq(&x->wait.lock); | |
3251 | } while (!x->done); | |
3252 | __remove_wait_queue(&x->wait, &wait); | |
3253 | } | |
3254 | x->done--; | |
3255 | spin_unlock_irq(&x->wait.lock); | |
3256 | } | |
3257 | EXPORT_SYMBOL(wait_for_completion); | |
3258 | ||
3259 | unsigned long fastcall __sched | |
3260 | wait_for_completion_timeout(struct completion *x, unsigned long timeout) | |
3261 | { | |
3262 | might_sleep(); | |
3263 | ||
3264 | spin_lock_irq(&x->wait.lock); | |
3265 | if (!x->done) { | |
3266 | DECLARE_WAITQUEUE(wait, current); | |
3267 | ||
3268 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3269 | __add_wait_queue_tail(&x->wait, &wait); | |
3270 | do { | |
3271 | __set_current_state(TASK_UNINTERRUPTIBLE); | |
3272 | spin_unlock_irq(&x->wait.lock); | |
3273 | timeout = schedule_timeout(timeout); | |
3274 | spin_lock_irq(&x->wait.lock); | |
3275 | if (!timeout) { | |
3276 | __remove_wait_queue(&x->wait, &wait); | |
3277 | goto out; | |
3278 | } | |
3279 | } while (!x->done); | |
3280 | __remove_wait_queue(&x->wait, &wait); | |
3281 | } | |
3282 | x->done--; | |
3283 | out: | |
3284 | spin_unlock_irq(&x->wait.lock); | |
3285 | return timeout; | |
3286 | } | |
3287 | EXPORT_SYMBOL(wait_for_completion_timeout); | |
3288 | ||
3289 | int fastcall __sched wait_for_completion_interruptible(struct completion *x) | |
3290 | { | |
3291 | int ret = 0; | |
3292 | ||
3293 | might_sleep(); | |
3294 | ||
3295 | spin_lock_irq(&x->wait.lock); | |
3296 | if (!x->done) { | |
3297 | DECLARE_WAITQUEUE(wait, current); | |
3298 | ||
3299 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3300 | __add_wait_queue_tail(&x->wait, &wait); | |
3301 | do { | |
3302 | if (signal_pending(current)) { | |
3303 | ret = -ERESTARTSYS; | |
3304 | __remove_wait_queue(&x->wait, &wait); | |
3305 | goto out; | |
3306 | } | |
3307 | __set_current_state(TASK_INTERRUPTIBLE); | |
3308 | spin_unlock_irq(&x->wait.lock); | |
3309 | schedule(); | |
3310 | spin_lock_irq(&x->wait.lock); | |
3311 | } while (!x->done); | |
3312 | __remove_wait_queue(&x->wait, &wait); | |
3313 | } | |
3314 | x->done--; | |
3315 | out: | |
3316 | spin_unlock_irq(&x->wait.lock); | |
3317 | ||
3318 | return ret; | |
3319 | } | |
3320 | EXPORT_SYMBOL(wait_for_completion_interruptible); | |
3321 | ||
3322 | unsigned long fastcall __sched | |
3323 | wait_for_completion_interruptible_timeout(struct completion *x, | |
3324 | unsigned long timeout) | |
3325 | { | |
3326 | might_sleep(); | |
3327 | ||
3328 | spin_lock_irq(&x->wait.lock); | |
3329 | if (!x->done) { | |
3330 | DECLARE_WAITQUEUE(wait, current); | |
3331 | ||
3332 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3333 | __add_wait_queue_tail(&x->wait, &wait); | |
3334 | do { | |
3335 | if (signal_pending(current)) { | |
3336 | timeout = -ERESTARTSYS; | |
3337 | __remove_wait_queue(&x->wait, &wait); | |
3338 | goto out; | |
3339 | } | |
3340 | __set_current_state(TASK_INTERRUPTIBLE); | |
3341 | spin_unlock_irq(&x->wait.lock); | |
3342 | timeout = schedule_timeout(timeout); | |
3343 | spin_lock_irq(&x->wait.lock); | |
3344 | if (!timeout) { | |
3345 | __remove_wait_queue(&x->wait, &wait); | |
3346 | goto out; | |
3347 | } | |
3348 | } while (!x->done); | |
3349 | __remove_wait_queue(&x->wait, &wait); | |
3350 | } | |
3351 | x->done--; | |
3352 | out: | |
3353 | spin_unlock_irq(&x->wait.lock); | |
3354 | return timeout; | |
3355 | } | |
3356 | EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); | |
3357 | ||
3358 | ||
3359 | #define SLEEP_ON_VAR \ | |
3360 | unsigned long flags; \ | |
3361 | wait_queue_t wait; \ | |
3362 | init_waitqueue_entry(&wait, current); | |
3363 | ||
3364 | #define SLEEP_ON_HEAD \ | |
3365 | spin_lock_irqsave(&q->lock,flags); \ | |
3366 | __add_wait_queue(q, &wait); \ | |
3367 | spin_unlock(&q->lock); | |
3368 | ||
3369 | #define SLEEP_ON_TAIL \ | |
3370 | spin_lock_irq(&q->lock); \ | |
3371 | __remove_wait_queue(q, &wait); \ | |
3372 | spin_unlock_irqrestore(&q->lock, flags); | |
3373 | ||
3374 | void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q) | |
3375 | { | |
3376 | SLEEP_ON_VAR | |
3377 | ||
3378 | current->state = TASK_INTERRUPTIBLE; | |
3379 | ||
3380 | SLEEP_ON_HEAD | |
3381 | schedule(); | |
3382 | SLEEP_ON_TAIL | |
3383 | } | |
3384 | ||
3385 | EXPORT_SYMBOL(interruptible_sleep_on); | |
3386 | ||
95cdf3b7 IM |
3387 | long fastcall __sched |
3388 | interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) | |
1da177e4 LT |
3389 | { |
3390 | SLEEP_ON_VAR | |
3391 | ||
3392 | current->state = TASK_INTERRUPTIBLE; | |
3393 | ||
3394 | SLEEP_ON_HEAD | |
3395 | timeout = schedule_timeout(timeout); | |
3396 | SLEEP_ON_TAIL | |
3397 | ||
3398 | return timeout; | |
3399 | } | |
3400 | ||
3401 | EXPORT_SYMBOL(interruptible_sleep_on_timeout); | |
3402 | ||
3403 | void fastcall __sched sleep_on(wait_queue_head_t *q) | |
3404 | { | |
3405 | SLEEP_ON_VAR | |
3406 | ||
3407 | current->state = TASK_UNINTERRUPTIBLE; | |
3408 | ||
3409 | SLEEP_ON_HEAD | |
3410 | schedule(); | |
3411 | SLEEP_ON_TAIL | |
3412 | } | |
3413 | ||
3414 | EXPORT_SYMBOL(sleep_on); | |
3415 | ||
3416 | long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) | |
3417 | { | |
3418 | SLEEP_ON_VAR | |
3419 | ||
3420 | current->state = TASK_UNINTERRUPTIBLE; | |
3421 | ||
3422 | SLEEP_ON_HEAD | |
3423 | timeout = schedule_timeout(timeout); | |
3424 | SLEEP_ON_TAIL | |
3425 | ||
3426 | return timeout; | |
3427 | } | |
3428 | ||
3429 | EXPORT_SYMBOL(sleep_on_timeout); | |
3430 | ||
3431 | void set_user_nice(task_t *p, long nice) | |
3432 | { | |
3433 | unsigned long flags; | |
3434 | prio_array_t *array; | |
3435 | runqueue_t *rq; | |
3436 | int old_prio, new_prio, delta; | |
3437 | ||
3438 | if (TASK_NICE(p) == nice || nice < -20 || nice > 19) | |
3439 | return; | |
3440 | /* | |
3441 | * We have to be careful, if called from sys_setpriority(), | |
3442 | * the task might be in the middle of scheduling on another CPU. | |
3443 | */ | |
3444 | rq = task_rq_lock(p, &flags); | |
3445 | /* | |
3446 | * The RT priorities are set via sched_setscheduler(), but we still | |
3447 | * allow the 'normal' nice value to be set - but as expected | |
3448 | * it wont have any effect on scheduling until the task is | |
3449 | * not SCHED_NORMAL: | |
3450 | */ | |
3451 | if (rt_task(p)) { | |
3452 | p->static_prio = NICE_TO_PRIO(nice); | |
3453 | goto out_unlock; | |
3454 | } | |
3455 | array = p->array; | |
3456 | if (array) | |
3457 | dequeue_task(p, array); | |
3458 | ||
3459 | old_prio = p->prio; | |
3460 | new_prio = NICE_TO_PRIO(nice); | |
3461 | delta = new_prio - old_prio; | |
3462 | p->static_prio = NICE_TO_PRIO(nice); | |
3463 | p->prio += delta; | |
3464 | ||
3465 | if (array) { | |
3466 | enqueue_task(p, array); | |
3467 | /* | |
3468 | * If the task increased its priority or is running and | |
3469 | * lowered its priority, then reschedule its CPU: | |
3470 | */ | |
3471 | if (delta < 0 || (delta > 0 && task_running(rq, p))) | |
3472 | resched_task(rq->curr); | |
3473 | } | |
3474 | out_unlock: | |
3475 | task_rq_unlock(rq, &flags); | |
3476 | } | |
3477 | ||
3478 | EXPORT_SYMBOL(set_user_nice); | |
3479 | ||
e43379f1 MM |
3480 | /* |
3481 | * can_nice - check if a task can reduce its nice value | |
3482 | * @p: task | |
3483 | * @nice: nice value | |
3484 | */ | |
3485 | int can_nice(const task_t *p, const int nice) | |
3486 | { | |
024f4747 MM |
3487 | /* convert nice value [19,-20] to rlimit style value [1,40] */ |
3488 | int nice_rlim = 20 - nice; | |
e43379f1 MM |
3489 | return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur || |
3490 | capable(CAP_SYS_NICE)); | |
3491 | } | |
3492 | ||
1da177e4 LT |
3493 | #ifdef __ARCH_WANT_SYS_NICE |
3494 | ||
3495 | /* | |
3496 | * sys_nice - change the priority of the current process. | |
3497 | * @increment: priority increment | |
3498 | * | |
3499 | * sys_setpriority is a more generic, but much slower function that | |
3500 | * does similar things. | |
3501 | */ | |
3502 | asmlinkage long sys_nice(int increment) | |
3503 | { | |
3504 | int retval; | |
3505 | long nice; | |
3506 | ||
3507 | /* | |
3508 | * Setpriority might change our priority at the same moment. | |
3509 | * We don't have to worry. Conceptually one call occurs first | |
3510 | * and we have a single winner. | |
3511 | */ | |
e43379f1 MM |
3512 | if (increment < -40) |
3513 | increment = -40; | |
1da177e4 LT |
3514 | if (increment > 40) |
3515 | increment = 40; | |
3516 | ||
3517 | nice = PRIO_TO_NICE(current->static_prio) + increment; | |
3518 | if (nice < -20) | |
3519 | nice = -20; | |
3520 | if (nice > 19) | |
3521 | nice = 19; | |
3522 | ||
e43379f1 MM |
3523 | if (increment < 0 && !can_nice(current, nice)) |
3524 | return -EPERM; | |
3525 | ||
1da177e4 LT |
3526 | retval = security_task_setnice(current, nice); |
3527 | if (retval) | |
3528 | return retval; | |
3529 | ||
3530 | set_user_nice(current, nice); | |
3531 | return 0; | |
3532 | } | |
3533 | ||
3534 | #endif | |
3535 | ||
3536 | /** | |
3537 | * task_prio - return the priority value of a given task. | |
3538 | * @p: the task in question. | |
3539 | * | |
3540 | * This is the priority value as seen by users in /proc. | |
3541 | * RT tasks are offset by -200. Normal tasks are centered | |
3542 | * around 0, value goes from -16 to +15. | |
3543 | */ | |
3544 | int task_prio(const task_t *p) | |
3545 | { | |
3546 | return p->prio - MAX_RT_PRIO; | |
3547 | } | |
3548 | ||
3549 | /** | |
3550 | * task_nice - return the nice value of a given task. | |
3551 | * @p: the task in question. | |
3552 | */ | |
3553 | int task_nice(const task_t *p) | |
3554 | { | |
3555 | return TASK_NICE(p); | |
3556 | } | |
1da177e4 | 3557 | EXPORT_SYMBOL_GPL(task_nice); |
1da177e4 LT |
3558 | |
3559 | /** | |
3560 | * idle_cpu - is a given cpu idle currently? | |
3561 | * @cpu: the processor in question. | |
3562 | */ | |
3563 | int idle_cpu(int cpu) | |
3564 | { | |
3565 | return cpu_curr(cpu) == cpu_rq(cpu)->idle; | |
3566 | } | |
3567 | ||
3568 | EXPORT_SYMBOL_GPL(idle_cpu); | |
3569 | ||
3570 | /** | |
3571 | * idle_task - return the idle task for a given cpu. | |
3572 | * @cpu: the processor in question. | |
3573 | */ | |
3574 | task_t *idle_task(int cpu) | |
3575 | { | |
3576 | return cpu_rq(cpu)->idle; | |
3577 | } | |
3578 | ||
3579 | /** | |
3580 | * find_process_by_pid - find a process with a matching PID value. | |
3581 | * @pid: the pid in question. | |
3582 | */ | |
3583 | static inline task_t *find_process_by_pid(pid_t pid) | |
3584 | { | |
3585 | return pid ? find_task_by_pid(pid) : current; | |
3586 | } | |
3587 | ||
3588 | /* Actually do priority change: must hold rq lock. */ | |
3589 | static void __setscheduler(struct task_struct *p, int policy, int prio) | |
3590 | { | |
3591 | BUG_ON(p->array); | |
3592 | p->policy = policy; | |
3593 | p->rt_priority = prio; | |
3594 | if (policy != SCHED_NORMAL) | |
d46523ea | 3595 | p->prio = MAX_RT_PRIO-1 - p->rt_priority; |
1da177e4 LT |
3596 | else |
3597 | p->prio = p->static_prio; | |
3598 | } | |
3599 | ||
3600 | /** | |
3601 | * sched_setscheduler - change the scheduling policy and/or RT priority of | |
3602 | * a thread. | |
3603 | * @p: the task in question. | |
3604 | * @policy: new policy. | |
3605 | * @param: structure containing the new RT priority. | |
3606 | */ | |
95cdf3b7 IM |
3607 | int sched_setscheduler(struct task_struct *p, int policy, |
3608 | struct sched_param *param) | |
1da177e4 LT |
3609 | { |
3610 | int retval; | |
3611 | int oldprio, oldpolicy = -1; | |
3612 | prio_array_t *array; | |
3613 | unsigned long flags; | |
3614 | runqueue_t *rq; | |
3615 | ||
3616 | recheck: | |
3617 | /* double check policy once rq lock held */ | |
3618 | if (policy < 0) | |
3619 | policy = oldpolicy = p->policy; | |
3620 | else if (policy != SCHED_FIFO && policy != SCHED_RR && | |
3621 | policy != SCHED_NORMAL) | |
3622 | return -EINVAL; | |
3623 | /* | |
3624 | * Valid priorities for SCHED_FIFO and SCHED_RR are | |
3625 | * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL is 0. | |
3626 | */ | |
3627 | if (param->sched_priority < 0 || | |
95cdf3b7 | 3628 | (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || |
d46523ea | 3629 | (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) |
1da177e4 LT |
3630 | return -EINVAL; |
3631 | if ((policy == SCHED_NORMAL) != (param->sched_priority == 0)) | |
3632 | return -EINVAL; | |
3633 | ||
37e4ab3f OC |
3634 | /* |
3635 | * Allow unprivileged RT tasks to decrease priority: | |
3636 | */ | |
3637 | if (!capable(CAP_SYS_NICE)) { | |
3638 | /* can't change policy */ | |
18586e72 AS |
3639 | if (policy != p->policy && |
3640 | !p->signal->rlim[RLIMIT_RTPRIO].rlim_cur) | |
37e4ab3f OC |
3641 | return -EPERM; |
3642 | /* can't increase priority */ | |
3643 | if (policy != SCHED_NORMAL && | |
3644 | param->sched_priority > p->rt_priority && | |
3645 | param->sched_priority > | |
3646 | p->signal->rlim[RLIMIT_RTPRIO].rlim_cur) | |
3647 | return -EPERM; | |
3648 | /* can't change other user's priorities */ | |
3649 | if ((current->euid != p->euid) && | |
3650 | (current->euid != p->uid)) | |
3651 | return -EPERM; | |
3652 | } | |
1da177e4 LT |
3653 | |
3654 | retval = security_task_setscheduler(p, policy, param); | |
3655 | if (retval) | |
3656 | return retval; | |
3657 | /* | |
3658 | * To be able to change p->policy safely, the apropriate | |
3659 | * runqueue lock must be held. | |
3660 | */ | |
3661 | rq = task_rq_lock(p, &flags); | |
3662 | /* recheck policy now with rq lock held */ | |
3663 | if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { | |
3664 | policy = oldpolicy = -1; | |
3665 | task_rq_unlock(rq, &flags); | |
3666 | goto recheck; | |
3667 | } | |
3668 | array = p->array; | |
3669 | if (array) | |
3670 | deactivate_task(p, rq); | |
3671 | oldprio = p->prio; | |
3672 | __setscheduler(p, policy, param->sched_priority); | |
3673 | if (array) { | |
3674 | __activate_task(p, rq); | |
3675 | /* | |
3676 | * Reschedule if we are currently running on this runqueue and | |
3677 | * our priority decreased, or if we are not currently running on | |
3678 | * this runqueue and our priority is higher than the current's | |
3679 | */ | |
3680 | if (task_running(rq, p)) { | |
3681 | if (p->prio > oldprio) | |
3682 | resched_task(rq->curr); | |
3683 | } else if (TASK_PREEMPTS_CURR(p, rq)) | |
3684 | resched_task(rq->curr); | |
3685 | } | |
3686 | task_rq_unlock(rq, &flags); | |
3687 | return 0; | |
3688 | } | |
3689 | EXPORT_SYMBOL_GPL(sched_setscheduler); | |
3690 | ||
95cdf3b7 IM |
3691 | static int |
3692 | do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) | |
1da177e4 LT |
3693 | { |
3694 | int retval; | |
3695 | struct sched_param lparam; | |
3696 | struct task_struct *p; | |
3697 | ||
3698 | if (!param || pid < 0) | |
3699 | return -EINVAL; | |
3700 | if (copy_from_user(&lparam, param, sizeof(struct sched_param))) | |
3701 | return -EFAULT; | |
3702 | read_lock_irq(&tasklist_lock); | |
3703 | p = find_process_by_pid(pid); | |
3704 | if (!p) { | |
3705 | read_unlock_irq(&tasklist_lock); | |
3706 | return -ESRCH; | |
3707 | } | |
3708 | retval = sched_setscheduler(p, policy, &lparam); | |
3709 | read_unlock_irq(&tasklist_lock); | |
3710 | return retval; | |
3711 | } | |
3712 | ||
3713 | /** | |
3714 | * sys_sched_setscheduler - set/change the scheduler policy and RT priority | |
3715 | * @pid: the pid in question. | |
3716 | * @policy: new policy. | |
3717 | * @param: structure containing the new RT priority. | |
3718 | */ | |
3719 | asmlinkage long sys_sched_setscheduler(pid_t pid, int policy, | |
3720 | struct sched_param __user *param) | |
3721 | { | |
3722 | return do_sched_setscheduler(pid, policy, param); | |
3723 | } | |
3724 | ||
3725 | /** | |
3726 | * sys_sched_setparam - set/change the RT priority of a thread | |
3727 | * @pid: the pid in question. | |
3728 | * @param: structure containing the new RT priority. | |
3729 | */ | |
3730 | asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param) | |
3731 | { | |
3732 | return do_sched_setscheduler(pid, -1, param); | |
3733 | } | |
3734 | ||
3735 | /** | |
3736 | * sys_sched_getscheduler - get the policy (scheduling class) of a thread | |
3737 | * @pid: the pid in question. | |
3738 | */ | |
3739 | asmlinkage long sys_sched_getscheduler(pid_t pid) | |
3740 | { | |
3741 | int retval = -EINVAL; | |
3742 | task_t *p; | |
3743 | ||
3744 | if (pid < 0) | |
3745 | goto out_nounlock; | |
3746 | ||
3747 | retval = -ESRCH; | |
3748 | read_lock(&tasklist_lock); | |
3749 | p = find_process_by_pid(pid); | |
3750 | if (p) { | |
3751 | retval = security_task_getscheduler(p); | |
3752 | if (!retval) | |
3753 | retval = p->policy; | |
3754 | } | |
3755 | read_unlock(&tasklist_lock); | |
3756 | ||
3757 | out_nounlock: | |
3758 | return retval; | |
3759 | } | |
3760 | ||
3761 | /** | |
3762 | * sys_sched_getscheduler - get the RT priority of a thread | |
3763 | * @pid: the pid in question. | |
3764 | * @param: structure containing the RT priority. | |
3765 | */ | |
3766 | asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param) | |
3767 | { | |
3768 | struct sched_param lp; | |
3769 | int retval = -EINVAL; | |
3770 | task_t *p; | |
3771 | ||
3772 | if (!param || pid < 0) | |
3773 | goto out_nounlock; | |
3774 | ||
3775 | read_lock(&tasklist_lock); | |
3776 | p = find_process_by_pid(pid); | |
3777 | retval = -ESRCH; | |
3778 | if (!p) | |
3779 | goto out_unlock; | |
3780 | ||
3781 | retval = security_task_getscheduler(p); | |
3782 | if (retval) | |
3783 | goto out_unlock; | |
3784 | ||
3785 | lp.sched_priority = p->rt_priority; | |
3786 | read_unlock(&tasklist_lock); | |
3787 | ||
3788 | /* | |
3789 | * This one might sleep, we cannot do it with a spinlock held ... | |
3790 | */ | |
3791 | retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; | |
3792 | ||
3793 | out_nounlock: | |
3794 | return retval; | |
3795 | ||
3796 | out_unlock: | |
3797 | read_unlock(&tasklist_lock); | |
3798 | return retval; | |
3799 | } | |
3800 | ||
3801 | long sched_setaffinity(pid_t pid, cpumask_t new_mask) | |
3802 | { | |
3803 | task_t *p; | |
3804 | int retval; | |
3805 | cpumask_t cpus_allowed; | |
3806 | ||
3807 | lock_cpu_hotplug(); | |
3808 | read_lock(&tasklist_lock); | |
3809 | ||
3810 | p = find_process_by_pid(pid); | |
3811 | if (!p) { | |
3812 | read_unlock(&tasklist_lock); | |
3813 | unlock_cpu_hotplug(); | |
3814 | return -ESRCH; | |
3815 | } | |
3816 | ||
3817 | /* | |
3818 | * It is not safe to call set_cpus_allowed with the | |
3819 | * tasklist_lock held. We will bump the task_struct's | |
3820 | * usage count and then drop tasklist_lock. | |
3821 | */ | |
3822 | get_task_struct(p); | |
3823 | read_unlock(&tasklist_lock); | |
3824 | ||
3825 | retval = -EPERM; | |
3826 | if ((current->euid != p->euid) && (current->euid != p->uid) && | |
3827 | !capable(CAP_SYS_NICE)) | |
3828 | goto out_unlock; | |
3829 | ||
3830 | cpus_allowed = cpuset_cpus_allowed(p); | |
3831 | cpus_and(new_mask, new_mask, cpus_allowed); | |
3832 | retval = set_cpus_allowed(p, new_mask); | |
3833 | ||
3834 | out_unlock: | |
3835 | put_task_struct(p); | |
3836 | unlock_cpu_hotplug(); | |
3837 | return retval; | |
3838 | } | |
3839 | ||
3840 | static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, | |
3841 | cpumask_t *new_mask) | |
3842 | { | |
3843 | if (len < sizeof(cpumask_t)) { | |
3844 | memset(new_mask, 0, sizeof(cpumask_t)); | |
3845 | } else if (len > sizeof(cpumask_t)) { | |
3846 | len = sizeof(cpumask_t); | |
3847 | } | |
3848 | return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; | |
3849 | } | |
3850 | ||
3851 | /** | |
3852 | * sys_sched_setaffinity - set the cpu affinity of a process | |
3853 | * @pid: pid of the process | |
3854 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
3855 | * @user_mask_ptr: user-space pointer to the new cpu mask | |
3856 | */ | |
3857 | asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len, | |
3858 | unsigned long __user *user_mask_ptr) | |
3859 | { | |
3860 | cpumask_t new_mask; | |
3861 | int retval; | |
3862 | ||
3863 | retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask); | |
3864 | if (retval) | |
3865 | return retval; | |
3866 | ||
3867 | return sched_setaffinity(pid, new_mask); | |
3868 | } | |
3869 | ||
3870 | /* | |
3871 | * Represents all cpu's present in the system | |
3872 | * In systems capable of hotplug, this map could dynamically grow | |
3873 | * as new cpu's are detected in the system via any platform specific | |
3874 | * method, such as ACPI for e.g. | |
3875 | */ | |
3876 | ||
3877 | cpumask_t cpu_present_map; | |
3878 | EXPORT_SYMBOL(cpu_present_map); | |
3879 | ||
3880 | #ifndef CONFIG_SMP | |
3881 | cpumask_t cpu_online_map = CPU_MASK_ALL; | |
3882 | cpumask_t cpu_possible_map = CPU_MASK_ALL; | |
3883 | #endif | |
3884 | ||
3885 | long sched_getaffinity(pid_t pid, cpumask_t *mask) | |
3886 | { | |
3887 | int retval; | |
3888 | task_t *p; | |
3889 | ||
3890 | lock_cpu_hotplug(); | |
3891 | read_lock(&tasklist_lock); | |
3892 | ||
3893 | retval = -ESRCH; | |
3894 | p = find_process_by_pid(pid); | |
3895 | if (!p) | |
3896 | goto out_unlock; | |
3897 | ||
3898 | retval = 0; | |
3899 | cpus_and(*mask, p->cpus_allowed, cpu_possible_map); | |
3900 | ||
3901 | out_unlock: | |
3902 | read_unlock(&tasklist_lock); | |
3903 | unlock_cpu_hotplug(); | |
3904 | if (retval) | |
3905 | return retval; | |
3906 | ||
3907 | return 0; | |
3908 | } | |
3909 | ||
3910 | /** | |
3911 | * sys_sched_getaffinity - get the cpu affinity of a process | |
3912 | * @pid: pid of the process | |
3913 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
3914 | * @user_mask_ptr: user-space pointer to hold the current cpu mask | |
3915 | */ | |
3916 | asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len, | |
3917 | unsigned long __user *user_mask_ptr) | |
3918 | { | |
3919 | int ret; | |
3920 | cpumask_t mask; | |
3921 | ||
3922 | if (len < sizeof(cpumask_t)) | |
3923 | return -EINVAL; | |
3924 | ||
3925 | ret = sched_getaffinity(pid, &mask); | |
3926 | if (ret < 0) | |
3927 | return ret; | |
3928 | ||
3929 | if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t))) | |
3930 | return -EFAULT; | |
3931 | ||
3932 | return sizeof(cpumask_t); | |
3933 | } | |
3934 | ||
3935 | /** | |
3936 | * sys_sched_yield - yield the current processor to other threads. | |
3937 | * | |
3938 | * this function yields the current CPU by moving the calling thread | |
3939 | * to the expired array. If there are no other threads running on this | |
3940 | * CPU then this function will return. | |
3941 | */ | |
3942 | asmlinkage long sys_sched_yield(void) | |
3943 | { | |
3944 | runqueue_t *rq = this_rq_lock(); | |
3945 | prio_array_t *array = current->array; | |
3946 | prio_array_t *target = rq->expired; | |
3947 | ||
3948 | schedstat_inc(rq, yld_cnt); | |
3949 | /* | |
3950 | * We implement yielding by moving the task into the expired | |
3951 | * queue. | |
3952 | * | |
3953 | * (special rule: RT tasks will just roundrobin in the active | |
3954 | * array.) | |
3955 | */ | |
3956 | if (rt_task(current)) | |
3957 | target = rq->active; | |
3958 | ||
5927ad78 | 3959 | if (array->nr_active == 1) { |
1da177e4 LT |
3960 | schedstat_inc(rq, yld_act_empty); |
3961 | if (!rq->expired->nr_active) | |
3962 | schedstat_inc(rq, yld_both_empty); | |
3963 | } else if (!rq->expired->nr_active) | |
3964 | schedstat_inc(rq, yld_exp_empty); | |
3965 | ||
3966 | if (array != target) { | |
3967 | dequeue_task(current, array); | |
3968 | enqueue_task(current, target); | |
3969 | } else | |
3970 | /* | |
3971 | * requeue_task is cheaper so perform that if possible. | |
3972 | */ | |
3973 | requeue_task(current, array); | |
3974 | ||
3975 | /* | |
3976 | * Since we are going to call schedule() anyway, there's | |
3977 | * no need to preempt or enable interrupts: | |
3978 | */ | |
3979 | __release(rq->lock); | |
3980 | _raw_spin_unlock(&rq->lock); | |
3981 | preempt_enable_no_resched(); | |
3982 | ||
3983 | schedule(); | |
3984 | ||
3985 | return 0; | |
3986 | } | |
3987 | ||
3988 | static inline void __cond_resched(void) | |
3989 | { | |
5bbcfd90 IM |
3990 | /* |
3991 | * The BKS might be reacquired before we have dropped | |
3992 | * PREEMPT_ACTIVE, which could trigger a second | |
3993 | * cond_resched() call. | |
3994 | */ | |
3995 | if (unlikely(preempt_count())) | |
3996 | return; | |
1da177e4 LT |
3997 | do { |
3998 | add_preempt_count(PREEMPT_ACTIVE); | |
3999 | schedule(); | |
4000 | sub_preempt_count(PREEMPT_ACTIVE); | |
4001 | } while (need_resched()); | |
4002 | } | |
4003 | ||
4004 | int __sched cond_resched(void) | |
4005 | { | |
4006 | if (need_resched()) { | |
4007 | __cond_resched(); | |
4008 | return 1; | |
4009 | } | |
4010 | return 0; | |
4011 | } | |
4012 | ||
4013 | EXPORT_SYMBOL(cond_resched); | |
4014 | ||
4015 | /* | |
4016 | * cond_resched_lock() - if a reschedule is pending, drop the given lock, | |
4017 | * call schedule, and on return reacquire the lock. | |
4018 | * | |
4019 | * This works OK both with and without CONFIG_PREEMPT. We do strange low-level | |
4020 | * operations here to prevent schedule() from being called twice (once via | |
4021 | * spin_unlock(), once by hand). | |
4022 | */ | |
95cdf3b7 | 4023 | int cond_resched_lock(spinlock_t *lock) |
1da177e4 | 4024 | { |
6df3cecb JK |
4025 | int ret = 0; |
4026 | ||
1da177e4 LT |
4027 | if (need_lockbreak(lock)) { |
4028 | spin_unlock(lock); | |
4029 | cpu_relax(); | |
6df3cecb | 4030 | ret = 1; |
1da177e4 LT |
4031 | spin_lock(lock); |
4032 | } | |
4033 | if (need_resched()) { | |
4034 | _raw_spin_unlock(lock); | |
4035 | preempt_enable_no_resched(); | |
4036 | __cond_resched(); | |
6df3cecb | 4037 | ret = 1; |
1da177e4 | 4038 | spin_lock(lock); |
1da177e4 | 4039 | } |
6df3cecb | 4040 | return ret; |
1da177e4 LT |
4041 | } |
4042 | ||
4043 | EXPORT_SYMBOL(cond_resched_lock); | |
4044 | ||
4045 | int __sched cond_resched_softirq(void) | |
4046 | { | |
4047 | BUG_ON(!in_softirq()); | |
4048 | ||
4049 | if (need_resched()) { | |
4050 | __local_bh_enable(); | |
4051 | __cond_resched(); | |
4052 | local_bh_disable(); | |
4053 | return 1; | |
4054 | } | |
4055 | return 0; | |
4056 | } | |
4057 | ||
4058 | EXPORT_SYMBOL(cond_resched_softirq); | |
4059 | ||
4060 | ||
4061 | /** | |
4062 | * yield - yield the current processor to other threads. | |
4063 | * | |
4064 | * this is a shortcut for kernel-space yielding - it marks the | |
4065 | * thread runnable and calls sys_sched_yield(). | |
4066 | */ | |
4067 | void __sched yield(void) | |
4068 | { | |
4069 | set_current_state(TASK_RUNNING); | |
4070 | sys_sched_yield(); | |
4071 | } | |
4072 | ||
4073 | EXPORT_SYMBOL(yield); | |
4074 | ||
4075 | /* | |
4076 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so | |
4077 | * that process accounting knows that this is a task in IO wait state. | |
4078 | * | |
4079 | * But don't do that if it is a deliberate, throttling IO wait (this task | |
4080 | * has set its backing_dev_info: the queue against which it should throttle) | |
4081 | */ | |
4082 | void __sched io_schedule(void) | |
4083 | { | |
39c715b7 | 4084 | struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id()); |
1da177e4 LT |
4085 | |
4086 | atomic_inc(&rq->nr_iowait); | |
4087 | schedule(); | |
4088 | atomic_dec(&rq->nr_iowait); | |
4089 | } | |
4090 | ||
4091 | EXPORT_SYMBOL(io_schedule); | |
4092 | ||
4093 | long __sched io_schedule_timeout(long timeout) | |
4094 | { | |
39c715b7 | 4095 | struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id()); |
1da177e4 LT |
4096 | long ret; |
4097 | ||
4098 | atomic_inc(&rq->nr_iowait); | |
4099 | ret = schedule_timeout(timeout); | |
4100 | atomic_dec(&rq->nr_iowait); | |
4101 | return ret; | |
4102 | } | |
4103 | ||
4104 | /** | |
4105 | * sys_sched_get_priority_max - return maximum RT priority. | |
4106 | * @policy: scheduling class. | |
4107 | * | |
4108 | * this syscall returns the maximum rt_priority that can be used | |
4109 | * by a given scheduling class. | |
4110 | */ | |
4111 | asmlinkage long sys_sched_get_priority_max(int policy) | |
4112 | { | |
4113 | int ret = -EINVAL; | |
4114 | ||
4115 | switch (policy) { | |
4116 | case SCHED_FIFO: | |
4117 | case SCHED_RR: | |
4118 | ret = MAX_USER_RT_PRIO-1; | |
4119 | break; | |
4120 | case SCHED_NORMAL: | |
4121 | ret = 0; | |
4122 | break; | |
4123 | } | |
4124 | return ret; | |
4125 | } | |
4126 | ||
4127 | /** | |
4128 | * sys_sched_get_priority_min - return minimum RT priority. | |
4129 | * @policy: scheduling class. | |
4130 | * | |
4131 | * this syscall returns the minimum rt_priority that can be used | |
4132 | * by a given scheduling class. | |
4133 | */ | |
4134 | asmlinkage long sys_sched_get_priority_min(int policy) | |
4135 | { | |
4136 | int ret = -EINVAL; | |
4137 | ||
4138 | switch (policy) { | |
4139 | case SCHED_FIFO: | |
4140 | case SCHED_RR: | |
4141 | ret = 1; | |
4142 | break; | |
4143 | case SCHED_NORMAL: | |
4144 | ret = 0; | |
4145 | } | |
4146 | return ret; | |
4147 | } | |
4148 | ||
4149 | /** | |
4150 | * sys_sched_rr_get_interval - return the default timeslice of a process. | |
4151 | * @pid: pid of the process. | |
4152 | * @interval: userspace pointer to the timeslice value. | |
4153 | * | |
4154 | * this syscall writes the default timeslice value of a given process | |
4155 | * into the user-space timespec buffer. A value of '0' means infinity. | |
4156 | */ | |
4157 | asmlinkage | |
4158 | long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval) | |
4159 | { | |
4160 | int retval = -EINVAL; | |
4161 | struct timespec t; | |
4162 | task_t *p; | |
4163 | ||
4164 | if (pid < 0) | |
4165 | goto out_nounlock; | |
4166 | ||
4167 | retval = -ESRCH; | |
4168 | read_lock(&tasklist_lock); | |
4169 | p = find_process_by_pid(pid); | |
4170 | if (!p) | |
4171 | goto out_unlock; | |
4172 | ||
4173 | retval = security_task_getscheduler(p); | |
4174 | if (retval) | |
4175 | goto out_unlock; | |
4176 | ||
4177 | jiffies_to_timespec(p->policy & SCHED_FIFO ? | |
4178 | 0 : task_timeslice(p), &t); | |
4179 | read_unlock(&tasklist_lock); | |
4180 | retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; | |
4181 | out_nounlock: | |
4182 | return retval; | |
4183 | out_unlock: | |
4184 | read_unlock(&tasklist_lock); | |
4185 | return retval; | |
4186 | } | |
4187 | ||
4188 | static inline struct task_struct *eldest_child(struct task_struct *p) | |
4189 | { | |
4190 | if (list_empty(&p->children)) return NULL; | |
4191 | return list_entry(p->children.next,struct task_struct,sibling); | |
4192 | } | |
4193 | ||
4194 | static inline struct task_struct *older_sibling(struct task_struct *p) | |
4195 | { | |
4196 | if (p->sibling.prev==&p->parent->children) return NULL; | |
4197 | return list_entry(p->sibling.prev,struct task_struct,sibling); | |
4198 | } | |
4199 | ||
4200 | static inline struct task_struct *younger_sibling(struct task_struct *p) | |
4201 | { | |
4202 | if (p->sibling.next==&p->parent->children) return NULL; | |
4203 | return list_entry(p->sibling.next,struct task_struct,sibling); | |
4204 | } | |
4205 | ||
95cdf3b7 | 4206 | static void show_task(task_t *p) |
1da177e4 LT |
4207 | { |
4208 | task_t *relative; | |
4209 | unsigned state; | |
4210 | unsigned long free = 0; | |
4211 | static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" }; | |
4212 | ||
4213 | printk("%-13.13s ", p->comm); | |
4214 | state = p->state ? __ffs(p->state) + 1 : 0; | |
4215 | if (state < ARRAY_SIZE(stat_nam)) | |
4216 | printk(stat_nam[state]); | |
4217 | else | |
4218 | printk("?"); | |
4219 | #if (BITS_PER_LONG == 32) | |
4220 | if (state == TASK_RUNNING) | |
4221 | printk(" running "); | |
4222 | else | |
4223 | printk(" %08lX ", thread_saved_pc(p)); | |
4224 | #else | |
4225 | if (state == TASK_RUNNING) | |
4226 | printk(" running task "); | |
4227 | else | |
4228 | printk(" %016lx ", thread_saved_pc(p)); | |
4229 | #endif | |
4230 | #ifdef CONFIG_DEBUG_STACK_USAGE | |
4231 | { | |
95cdf3b7 | 4232 | unsigned long *n = (unsigned long *) (p->thread_info+1); |
1da177e4 LT |
4233 | while (!*n) |
4234 | n++; | |
4235 | free = (unsigned long) n - (unsigned long)(p->thread_info+1); | |
4236 | } | |
4237 | #endif | |
4238 | printk("%5lu %5d %6d ", free, p->pid, p->parent->pid); | |
4239 | if ((relative = eldest_child(p))) | |
4240 | printk("%5d ", relative->pid); | |
4241 | else | |
4242 | printk(" "); | |
4243 | if ((relative = younger_sibling(p))) | |
4244 | printk("%7d", relative->pid); | |
4245 | else | |
4246 | printk(" "); | |
4247 | if ((relative = older_sibling(p))) | |
4248 | printk(" %5d", relative->pid); | |
4249 | else | |
4250 | printk(" "); | |
4251 | if (!p->mm) | |
4252 | printk(" (L-TLB)\n"); | |
4253 | else | |
4254 | printk(" (NOTLB)\n"); | |
4255 | ||
4256 | if (state != TASK_RUNNING) | |
4257 | show_stack(p, NULL); | |
4258 | } | |
4259 | ||
4260 | void show_state(void) | |
4261 | { | |
4262 | task_t *g, *p; | |
4263 | ||
4264 | #if (BITS_PER_LONG == 32) | |
4265 | printk("\n" | |
4266 | " sibling\n"); | |
4267 | printk(" task PC pid father child younger older\n"); | |
4268 | #else | |
4269 | printk("\n" | |
4270 | " sibling\n"); | |
4271 | printk(" task PC pid father child younger older\n"); | |
4272 | #endif | |
4273 | read_lock(&tasklist_lock); | |
4274 | do_each_thread(g, p) { | |
4275 | /* | |
4276 | * reset the NMI-timeout, listing all files on a slow | |
4277 | * console might take alot of time: | |
4278 | */ | |
4279 | touch_nmi_watchdog(); | |
4280 | show_task(p); | |
4281 | } while_each_thread(g, p); | |
4282 | ||
4283 | read_unlock(&tasklist_lock); | |
4284 | } | |
4285 | ||
f340c0d1 IM |
4286 | /** |
4287 | * init_idle - set up an idle thread for a given CPU | |
4288 | * @idle: task in question | |
4289 | * @cpu: cpu the idle task belongs to | |
4290 | * | |
4291 | * NOTE: this function does not set the idle thread's NEED_RESCHED | |
4292 | * flag, to make booting more robust. | |
4293 | */ | |
1da177e4 LT |
4294 | void __devinit init_idle(task_t *idle, int cpu) |
4295 | { | |
4296 | runqueue_t *rq = cpu_rq(cpu); | |
4297 | unsigned long flags; | |
4298 | ||
4299 | idle->sleep_avg = 0; | |
4300 | idle->array = NULL; | |
4301 | idle->prio = MAX_PRIO; | |
4302 | idle->state = TASK_RUNNING; | |
4303 | idle->cpus_allowed = cpumask_of_cpu(cpu); | |
4304 | set_task_cpu(idle, cpu); | |
4305 | ||
4306 | spin_lock_irqsave(&rq->lock, flags); | |
4307 | rq->curr = rq->idle = idle; | |
4866cde0 NP |
4308 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
4309 | idle->oncpu = 1; | |
4310 | #endif | |
1da177e4 LT |
4311 | spin_unlock_irqrestore(&rq->lock, flags); |
4312 | ||
4313 | /* Set the preempt count _outside_ the spinlocks! */ | |
4314 | #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL) | |
4315 | idle->thread_info->preempt_count = (idle->lock_depth >= 0); | |
4316 | #else | |
4317 | idle->thread_info->preempt_count = 0; | |
4318 | #endif | |
4319 | } | |
4320 | ||
4321 | /* | |
4322 | * In a system that switches off the HZ timer nohz_cpu_mask | |
4323 | * indicates which cpus entered this state. This is used | |
4324 | * in the rcu update to wait only for active cpus. For system | |
4325 | * which do not switch off the HZ timer nohz_cpu_mask should | |
4326 | * always be CPU_MASK_NONE. | |
4327 | */ | |
4328 | cpumask_t nohz_cpu_mask = CPU_MASK_NONE; | |
4329 | ||
4330 | #ifdef CONFIG_SMP | |
4331 | /* | |
4332 | * This is how migration works: | |
4333 | * | |
4334 | * 1) we queue a migration_req_t structure in the source CPU's | |
4335 | * runqueue and wake up that CPU's migration thread. | |
4336 | * 2) we down() the locked semaphore => thread blocks. | |
4337 | * 3) migration thread wakes up (implicitly it forces the migrated | |
4338 | * thread off the CPU) | |
4339 | * 4) it gets the migration request and checks whether the migrated | |
4340 | * task is still in the wrong runqueue. | |
4341 | * 5) if it's in the wrong runqueue then the migration thread removes | |
4342 | * it and puts it into the right queue. | |
4343 | * 6) migration thread up()s the semaphore. | |
4344 | * 7) we wake up and the migration is done. | |
4345 | */ | |
4346 | ||
4347 | /* | |
4348 | * Change a given task's CPU affinity. Migrate the thread to a | |
4349 | * proper CPU and schedule it away if the CPU it's executing on | |
4350 | * is removed from the allowed bitmask. | |
4351 | * | |
4352 | * NOTE: the caller must have a valid reference to the task, the | |
4353 | * task must not exit() & deallocate itself prematurely. The | |
4354 | * call is not atomic; no spinlocks may be held. | |
4355 | */ | |
4356 | int set_cpus_allowed(task_t *p, cpumask_t new_mask) | |
4357 | { | |
4358 | unsigned long flags; | |
4359 | int ret = 0; | |
4360 | migration_req_t req; | |
4361 | runqueue_t *rq; | |
4362 | ||
4363 | rq = task_rq_lock(p, &flags); | |
4364 | if (!cpus_intersects(new_mask, cpu_online_map)) { | |
4365 | ret = -EINVAL; | |
4366 | goto out; | |
4367 | } | |
4368 | ||
4369 | p->cpus_allowed = new_mask; | |
4370 | /* Can the task run on the task's current CPU? If so, we're done */ | |
4371 | if (cpu_isset(task_cpu(p), new_mask)) | |
4372 | goto out; | |
4373 | ||
4374 | if (migrate_task(p, any_online_cpu(new_mask), &req)) { | |
4375 | /* Need help from migration thread: drop lock and wait. */ | |
4376 | task_rq_unlock(rq, &flags); | |
4377 | wake_up_process(rq->migration_thread); | |
4378 | wait_for_completion(&req.done); | |
4379 | tlb_migrate_finish(p->mm); | |
4380 | return 0; | |
4381 | } | |
4382 | out: | |
4383 | task_rq_unlock(rq, &flags); | |
4384 | return ret; | |
4385 | } | |
4386 | ||
4387 | EXPORT_SYMBOL_GPL(set_cpus_allowed); | |
4388 | ||
4389 | /* | |
4390 | * Move (not current) task off this cpu, onto dest cpu. We're doing | |
4391 | * this because either it can't run here any more (set_cpus_allowed() | |
4392 | * away from this CPU, or CPU going down), or because we're | |
4393 | * attempting to rebalance this task on exec (sched_exec). | |
4394 | * | |
4395 | * So we race with normal scheduler movements, but that's OK, as long | |
4396 | * as the task is no longer on this CPU. | |
4397 | */ | |
4398 | static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) | |
4399 | { | |
4400 | runqueue_t *rq_dest, *rq_src; | |
4401 | ||
4402 | if (unlikely(cpu_is_offline(dest_cpu))) | |
4403 | return; | |
4404 | ||
4405 | rq_src = cpu_rq(src_cpu); | |
4406 | rq_dest = cpu_rq(dest_cpu); | |
4407 | ||
4408 | double_rq_lock(rq_src, rq_dest); | |
4409 | /* Already moved. */ | |
4410 | if (task_cpu(p) != src_cpu) | |
4411 | goto out; | |
4412 | /* Affinity changed (again). */ | |
4413 | if (!cpu_isset(dest_cpu, p->cpus_allowed)) | |
4414 | goto out; | |
4415 | ||
4416 | set_task_cpu(p, dest_cpu); | |
4417 | if (p->array) { | |
4418 | /* | |
4419 | * Sync timestamp with rq_dest's before activating. | |
4420 | * The same thing could be achieved by doing this step | |
4421 | * afterwards, and pretending it was a local activate. | |
4422 | * This way is cleaner and logically correct. | |
4423 | */ | |
4424 | p->timestamp = p->timestamp - rq_src->timestamp_last_tick | |
4425 | + rq_dest->timestamp_last_tick; | |
4426 | deactivate_task(p, rq_src); | |
4427 | activate_task(p, rq_dest, 0); | |
4428 | if (TASK_PREEMPTS_CURR(p, rq_dest)) | |
4429 | resched_task(rq_dest->curr); | |
4430 | } | |
4431 | ||
4432 | out: | |
4433 | double_rq_unlock(rq_src, rq_dest); | |
4434 | } | |
4435 | ||
4436 | /* | |
4437 | * migration_thread - this is a highprio system thread that performs | |
4438 | * thread migration by bumping thread off CPU then 'pushing' onto | |
4439 | * another runqueue. | |
4440 | */ | |
95cdf3b7 | 4441 | static int migration_thread(void *data) |
1da177e4 LT |
4442 | { |
4443 | runqueue_t *rq; | |
4444 | int cpu = (long)data; | |
4445 | ||
4446 | rq = cpu_rq(cpu); | |
4447 | BUG_ON(rq->migration_thread != current); | |
4448 | ||
4449 | set_current_state(TASK_INTERRUPTIBLE); | |
4450 | while (!kthread_should_stop()) { | |
4451 | struct list_head *head; | |
4452 | migration_req_t *req; | |
4453 | ||
3e1d1d28 | 4454 | try_to_freeze(); |
1da177e4 LT |
4455 | |
4456 | spin_lock_irq(&rq->lock); | |
4457 | ||
4458 | if (cpu_is_offline(cpu)) { | |
4459 | spin_unlock_irq(&rq->lock); | |
4460 | goto wait_to_die; | |
4461 | } | |
4462 | ||
4463 | if (rq->active_balance) { | |
4464 | active_load_balance(rq, cpu); | |
4465 | rq->active_balance = 0; | |
4466 | } | |
4467 | ||
4468 | head = &rq->migration_queue; | |
4469 | ||
4470 | if (list_empty(head)) { | |
4471 | spin_unlock_irq(&rq->lock); | |
4472 | schedule(); | |
4473 | set_current_state(TASK_INTERRUPTIBLE); | |
4474 | continue; | |
4475 | } | |
4476 | req = list_entry(head->next, migration_req_t, list); | |
4477 | list_del_init(head->next); | |
4478 | ||
674311d5 NP |
4479 | spin_unlock(&rq->lock); |
4480 | __migrate_task(req->task, cpu, req->dest_cpu); | |
4481 | local_irq_enable(); | |
1da177e4 LT |
4482 | |
4483 | complete(&req->done); | |
4484 | } | |
4485 | __set_current_state(TASK_RUNNING); | |
4486 | return 0; | |
4487 | ||
4488 | wait_to_die: | |
4489 | /* Wait for kthread_stop */ | |
4490 | set_current_state(TASK_INTERRUPTIBLE); | |
4491 | while (!kthread_should_stop()) { | |
4492 | schedule(); | |
4493 | set_current_state(TASK_INTERRUPTIBLE); | |
4494 | } | |
4495 | __set_current_state(TASK_RUNNING); | |
4496 | return 0; | |
4497 | } | |
4498 | ||
4499 | #ifdef CONFIG_HOTPLUG_CPU | |
4500 | /* Figure out where task on dead CPU should go, use force if neccessary. */ | |
4501 | static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk) | |
4502 | { | |
4503 | int dest_cpu; | |
4504 | cpumask_t mask; | |
4505 | ||
4506 | /* On same node? */ | |
4507 | mask = node_to_cpumask(cpu_to_node(dead_cpu)); | |
4508 | cpus_and(mask, mask, tsk->cpus_allowed); | |
4509 | dest_cpu = any_online_cpu(mask); | |
4510 | ||
4511 | /* On any allowed CPU? */ | |
4512 | if (dest_cpu == NR_CPUS) | |
4513 | dest_cpu = any_online_cpu(tsk->cpus_allowed); | |
4514 | ||
4515 | /* No more Mr. Nice Guy. */ | |
4516 | if (dest_cpu == NR_CPUS) { | |
b39c4fab | 4517 | cpus_setall(tsk->cpus_allowed); |
1da177e4 LT |
4518 | dest_cpu = any_online_cpu(tsk->cpus_allowed); |
4519 | ||
4520 | /* | |
4521 | * Don't tell them about moving exiting tasks or | |
4522 | * kernel threads (both mm NULL), since they never | |
4523 | * leave kernel. | |
4524 | */ | |
4525 | if (tsk->mm && printk_ratelimit()) | |
4526 | printk(KERN_INFO "process %d (%s) no " | |
4527 | "longer affine to cpu%d\n", | |
4528 | tsk->pid, tsk->comm, dead_cpu); | |
4529 | } | |
4530 | __migrate_task(tsk, dead_cpu, dest_cpu); | |
4531 | } | |
4532 | ||
4533 | /* | |
4534 | * While a dead CPU has no uninterruptible tasks queued at this point, | |
4535 | * it might still have a nonzero ->nr_uninterruptible counter, because | |
4536 | * for performance reasons the counter is not stricly tracking tasks to | |
4537 | * their home CPUs. So we just add the counter to another CPU's counter, | |
4538 | * to keep the global sum constant after CPU-down: | |
4539 | */ | |
4540 | static void migrate_nr_uninterruptible(runqueue_t *rq_src) | |
4541 | { | |
4542 | runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL)); | |
4543 | unsigned long flags; | |
4544 | ||
4545 | local_irq_save(flags); | |
4546 | double_rq_lock(rq_src, rq_dest); | |
4547 | rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible; | |
4548 | rq_src->nr_uninterruptible = 0; | |
4549 | double_rq_unlock(rq_src, rq_dest); | |
4550 | local_irq_restore(flags); | |
4551 | } | |
4552 | ||
4553 | /* Run through task list and migrate tasks from the dead cpu. */ | |
4554 | static void migrate_live_tasks(int src_cpu) | |
4555 | { | |
4556 | struct task_struct *tsk, *t; | |
4557 | ||
4558 | write_lock_irq(&tasklist_lock); | |
4559 | ||
4560 | do_each_thread(t, tsk) { | |
4561 | if (tsk == current) | |
4562 | continue; | |
4563 | ||
4564 | if (task_cpu(tsk) == src_cpu) | |
4565 | move_task_off_dead_cpu(src_cpu, tsk); | |
4566 | } while_each_thread(t, tsk); | |
4567 | ||
4568 | write_unlock_irq(&tasklist_lock); | |
4569 | } | |
4570 | ||
4571 | /* Schedules idle task to be the next runnable task on current CPU. | |
4572 | * It does so by boosting its priority to highest possible and adding it to | |
4573 | * the _front_ of runqueue. Used by CPU offline code. | |
4574 | */ | |
4575 | void sched_idle_next(void) | |
4576 | { | |
4577 | int cpu = smp_processor_id(); | |
4578 | runqueue_t *rq = this_rq(); | |
4579 | struct task_struct *p = rq->idle; | |
4580 | unsigned long flags; | |
4581 | ||
4582 | /* cpu has to be offline */ | |
4583 | BUG_ON(cpu_online(cpu)); | |
4584 | ||
4585 | /* Strictly not necessary since rest of the CPUs are stopped by now | |
4586 | * and interrupts disabled on current cpu. | |
4587 | */ | |
4588 | spin_lock_irqsave(&rq->lock, flags); | |
4589 | ||
4590 | __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1); | |
4591 | /* Add idle task to _front_ of it's priority queue */ | |
4592 | __activate_idle_task(p, rq); | |
4593 | ||
4594 | spin_unlock_irqrestore(&rq->lock, flags); | |
4595 | } | |
4596 | ||
4597 | /* Ensures that the idle task is using init_mm right before its cpu goes | |
4598 | * offline. | |
4599 | */ | |
4600 | void idle_task_exit(void) | |
4601 | { | |
4602 | struct mm_struct *mm = current->active_mm; | |
4603 | ||
4604 | BUG_ON(cpu_online(smp_processor_id())); | |
4605 | ||
4606 | if (mm != &init_mm) | |
4607 | switch_mm(mm, &init_mm, current); | |
4608 | mmdrop(mm); | |
4609 | } | |
4610 | ||
4611 | static void migrate_dead(unsigned int dead_cpu, task_t *tsk) | |
4612 | { | |
4613 | struct runqueue *rq = cpu_rq(dead_cpu); | |
4614 | ||
4615 | /* Must be exiting, otherwise would be on tasklist. */ | |
4616 | BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD); | |
4617 | ||
4618 | /* Cannot have done final schedule yet: would have vanished. */ | |
4619 | BUG_ON(tsk->flags & PF_DEAD); | |
4620 | ||
4621 | get_task_struct(tsk); | |
4622 | ||
4623 | /* | |
4624 | * Drop lock around migration; if someone else moves it, | |
4625 | * that's OK. No task can be added to this CPU, so iteration is | |
4626 | * fine. | |
4627 | */ | |
4628 | spin_unlock_irq(&rq->lock); | |
4629 | move_task_off_dead_cpu(dead_cpu, tsk); | |
4630 | spin_lock_irq(&rq->lock); | |
4631 | ||
4632 | put_task_struct(tsk); | |
4633 | } | |
4634 | ||
4635 | /* release_task() removes task from tasklist, so we won't find dead tasks. */ | |
4636 | static void migrate_dead_tasks(unsigned int dead_cpu) | |
4637 | { | |
4638 | unsigned arr, i; | |
4639 | struct runqueue *rq = cpu_rq(dead_cpu); | |
4640 | ||
4641 | for (arr = 0; arr < 2; arr++) { | |
4642 | for (i = 0; i < MAX_PRIO; i++) { | |
4643 | struct list_head *list = &rq->arrays[arr].queue[i]; | |
4644 | while (!list_empty(list)) | |
4645 | migrate_dead(dead_cpu, | |
4646 | list_entry(list->next, task_t, | |
4647 | run_list)); | |
4648 | } | |
4649 | } | |
4650 | } | |
4651 | #endif /* CONFIG_HOTPLUG_CPU */ | |
4652 | ||
4653 | /* | |
4654 | * migration_call - callback that gets triggered when a CPU is added. | |
4655 | * Here we can start up the necessary migration thread for the new CPU. | |
4656 | */ | |
4657 | static int migration_call(struct notifier_block *nfb, unsigned long action, | |
4658 | void *hcpu) | |
4659 | { | |
4660 | int cpu = (long)hcpu; | |
4661 | struct task_struct *p; | |
4662 | struct runqueue *rq; | |
4663 | unsigned long flags; | |
4664 | ||
4665 | switch (action) { | |
4666 | case CPU_UP_PREPARE: | |
4667 | p = kthread_create(migration_thread, hcpu, "migration/%d",cpu); | |
4668 | if (IS_ERR(p)) | |
4669 | return NOTIFY_BAD; | |
4670 | p->flags |= PF_NOFREEZE; | |
4671 | kthread_bind(p, cpu); | |
4672 | /* Must be high prio: stop_machine expects to yield to it. */ | |
4673 | rq = task_rq_lock(p, &flags); | |
4674 | __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1); | |
4675 | task_rq_unlock(rq, &flags); | |
4676 | cpu_rq(cpu)->migration_thread = p; | |
4677 | break; | |
4678 | case CPU_ONLINE: | |
4679 | /* Strictly unneccessary, as first user will wake it. */ | |
4680 | wake_up_process(cpu_rq(cpu)->migration_thread); | |
4681 | break; | |
4682 | #ifdef CONFIG_HOTPLUG_CPU | |
4683 | case CPU_UP_CANCELED: | |
4684 | /* Unbind it from offline cpu so it can run. Fall thru. */ | |
4685 | kthread_bind(cpu_rq(cpu)->migration_thread,smp_processor_id()); | |
4686 | kthread_stop(cpu_rq(cpu)->migration_thread); | |
4687 | cpu_rq(cpu)->migration_thread = NULL; | |
4688 | break; | |
4689 | case CPU_DEAD: | |
4690 | migrate_live_tasks(cpu); | |
4691 | rq = cpu_rq(cpu); | |
4692 | kthread_stop(rq->migration_thread); | |
4693 | rq->migration_thread = NULL; | |
4694 | /* Idle task back to normal (off runqueue, low prio) */ | |
4695 | rq = task_rq_lock(rq->idle, &flags); | |
4696 | deactivate_task(rq->idle, rq); | |
4697 | rq->idle->static_prio = MAX_PRIO; | |
4698 | __setscheduler(rq->idle, SCHED_NORMAL, 0); | |
4699 | migrate_dead_tasks(cpu); | |
4700 | task_rq_unlock(rq, &flags); | |
4701 | migrate_nr_uninterruptible(rq); | |
4702 | BUG_ON(rq->nr_running != 0); | |
4703 | ||
4704 | /* No need to migrate the tasks: it was best-effort if | |
4705 | * they didn't do lock_cpu_hotplug(). Just wake up | |
4706 | * the requestors. */ | |
4707 | spin_lock_irq(&rq->lock); | |
4708 | while (!list_empty(&rq->migration_queue)) { | |
4709 | migration_req_t *req; | |
4710 | req = list_entry(rq->migration_queue.next, | |
4711 | migration_req_t, list); | |
1da177e4 LT |
4712 | list_del_init(&req->list); |
4713 | complete(&req->done); | |
4714 | } | |
4715 | spin_unlock_irq(&rq->lock); | |
4716 | break; | |
4717 | #endif | |
4718 | } | |
4719 | return NOTIFY_OK; | |
4720 | } | |
4721 | ||
4722 | /* Register at highest priority so that task migration (migrate_all_tasks) | |
4723 | * happens before everything else. | |
4724 | */ | |
4725 | static struct notifier_block __devinitdata migration_notifier = { | |
4726 | .notifier_call = migration_call, | |
4727 | .priority = 10 | |
4728 | }; | |
4729 | ||
4730 | int __init migration_init(void) | |
4731 | { | |
4732 | void *cpu = (void *)(long)smp_processor_id(); | |
4733 | /* Start one for boot CPU. */ | |
4734 | migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); | |
4735 | migration_call(&migration_notifier, CPU_ONLINE, cpu); | |
4736 | register_cpu_notifier(&migration_notifier); | |
4737 | return 0; | |
4738 | } | |
4739 | #endif | |
4740 | ||
4741 | #ifdef CONFIG_SMP | |
1a20ff27 | 4742 | #undef SCHED_DOMAIN_DEBUG |
1da177e4 LT |
4743 | #ifdef SCHED_DOMAIN_DEBUG |
4744 | static void sched_domain_debug(struct sched_domain *sd, int cpu) | |
4745 | { | |
4746 | int level = 0; | |
4747 | ||
41c7ce9a NP |
4748 | if (!sd) { |
4749 | printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); | |
4750 | return; | |
4751 | } | |
4752 | ||
1da177e4 LT |
4753 | printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); |
4754 | ||
4755 | do { | |
4756 | int i; | |
4757 | char str[NR_CPUS]; | |
4758 | struct sched_group *group = sd->groups; | |
4759 | cpumask_t groupmask; | |
4760 | ||
4761 | cpumask_scnprintf(str, NR_CPUS, sd->span); | |
4762 | cpus_clear(groupmask); | |
4763 | ||
4764 | printk(KERN_DEBUG); | |
4765 | for (i = 0; i < level + 1; i++) | |
4766 | printk(" "); | |
4767 | printk("domain %d: ", level); | |
4768 | ||
4769 | if (!(sd->flags & SD_LOAD_BALANCE)) { | |
4770 | printk("does not load-balance\n"); | |
4771 | if (sd->parent) | |
4772 | printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent"); | |
4773 | break; | |
4774 | } | |
4775 | ||
4776 | printk("span %s\n", str); | |
4777 | ||
4778 | if (!cpu_isset(cpu, sd->span)) | |
4779 | printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu); | |
4780 | if (!cpu_isset(cpu, group->cpumask)) | |
4781 | printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu); | |
4782 | ||
4783 | printk(KERN_DEBUG); | |
4784 | for (i = 0; i < level + 2; i++) | |
4785 | printk(" "); | |
4786 | printk("groups:"); | |
4787 | do { | |
4788 | if (!group) { | |
4789 | printk("\n"); | |
4790 | printk(KERN_ERR "ERROR: group is NULL\n"); | |
4791 | break; | |
4792 | } | |
4793 | ||
4794 | if (!group->cpu_power) { | |
4795 | printk("\n"); | |
4796 | printk(KERN_ERR "ERROR: domain->cpu_power not set\n"); | |
4797 | } | |
4798 | ||
4799 | if (!cpus_weight(group->cpumask)) { | |
4800 | printk("\n"); | |
4801 | printk(KERN_ERR "ERROR: empty group\n"); | |
4802 | } | |
4803 | ||
4804 | if (cpus_intersects(groupmask, group->cpumask)) { | |
4805 | printk("\n"); | |
4806 | printk(KERN_ERR "ERROR: repeated CPUs\n"); | |
4807 | } | |
4808 | ||
4809 | cpus_or(groupmask, groupmask, group->cpumask); | |
4810 | ||
4811 | cpumask_scnprintf(str, NR_CPUS, group->cpumask); | |
4812 | printk(" %s", str); | |
4813 | ||
4814 | group = group->next; | |
4815 | } while (group != sd->groups); | |
4816 | printk("\n"); | |
4817 | ||
4818 | if (!cpus_equal(sd->span, groupmask)) | |
4819 | printk(KERN_ERR "ERROR: groups don't span domain->span\n"); | |
4820 | ||
4821 | level++; | |
4822 | sd = sd->parent; | |
4823 | ||
4824 | if (sd) { | |
4825 | if (!cpus_subset(groupmask, sd->span)) | |
4826 | printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n"); | |
4827 | } | |
4828 | ||
4829 | } while (sd); | |
4830 | } | |
4831 | #else | |
4832 | #define sched_domain_debug(sd, cpu) {} | |
4833 | #endif | |
4834 | ||
1a20ff27 | 4835 | static int sd_degenerate(struct sched_domain *sd) |
245af2c7 SS |
4836 | { |
4837 | if (cpus_weight(sd->span) == 1) | |
4838 | return 1; | |
4839 | ||
4840 | /* Following flags need at least 2 groups */ | |
4841 | if (sd->flags & (SD_LOAD_BALANCE | | |
4842 | SD_BALANCE_NEWIDLE | | |
4843 | SD_BALANCE_FORK | | |
4844 | SD_BALANCE_EXEC)) { | |
4845 | if (sd->groups != sd->groups->next) | |
4846 | return 0; | |
4847 | } | |
4848 | ||
4849 | /* Following flags don't use groups */ | |
4850 | if (sd->flags & (SD_WAKE_IDLE | | |
4851 | SD_WAKE_AFFINE | | |
4852 | SD_WAKE_BALANCE)) | |
4853 | return 0; | |
4854 | ||
4855 | return 1; | |
4856 | } | |
4857 | ||
1a20ff27 | 4858 | static int sd_parent_degenerate(struct sched_domain *sd, |
245af2c7 SS |
4859 | struct sched_domain *parent) |
4860 | { | |
4861 | unsigned long cflags = sd->flags, pflags = parent->flags; | |
4862 | ||
4863 | if (sd_degenerate(parent)) | |
4864 | return 1; | |
4865 | ||
4866 | if (!cpus_equal(sd->span, parent->span)) | |
4867 | return 0; | |
4868 | ||
4869 | /* Does parent contain flags not in child? */ | |
4870 | /* WAKE_BALANCE is a subset of WAKE_AFFINE */ | |
4871 | if (cflags & SD_WAKE_AFFINE) | |
4872 | pflags &= ~SD_WAKE_BALANCE; | |
4873 | /* Flags needing groups don't count if only 1 group in parent */ | |
4874 | if (parent->groups == parent->groups->next) { | |
4875 | pflags &= ~(SD_LOAD_BALANCE | | |
4876 | SD_BALANCE_NEWIDLE | | |
4877 | SD_BALANCE_FORK | | |
4878 | SD_BALANCE_EXEC); | |
4879 | } | |
4880 | if (~cflags & pflags) | |
4881 | return 0; | |
4882 | ||
4883 | return 1; | |
4884 | } | |
4885 | ||
1da177e4 LT |
4886 | /* |
4887 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must | |
4888 | * hold the hotplug lock. | |
4889 | */ | |
9c1cfda2 | 4890 | static void cpu_attach_domain(struct sched_domain *sd, int cpu) |
1da177e4 | 4891 | { |
1da177e4 | 4892 | runqueue_t *rq = cpu_rq(cpu); |
245af2c7 SS |
4893 | struct sched_domain *tmp; |
4894 | ||
4895 | /* Remove the sched domains which do not contribute to scheduling. */ | |
4896 | for (tmp = sd; tmp; tmp = tmp->parent) { | |
4897 | struct sched_domain *parent = tmp->parent; | |
4898 | if (!parent) | |
4899 | break; | |
4900 | if (sd_parent_degenerate(tmp, parent)) | |
4901 | tmp->parent = parent->parent; | |
4902 | } | |
4903 | ||
4904 | if (sd && sd_degenerate(sd)) | |
4905 | sd = sd->parent; | |
1da177e4 LT |
4906 | |
4907 | sched_domain_debug(sd, cpu); | |
4908 | ||
674311d5 | 4909 | rcu_assign_pointer(rq->sd, sd); |
1da177e4 LT |
4910 | } |
4911 | ||
4912 | /* cpus with isolated domains */ | |
9c1cfda2 | 4913 | static cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE; |
1da177e4 LT |
4914 | |
4915 | /* Setup the mask of cpus configured for isolated domains */ | |
4916 | static int __init isolated_cpu_setup(char *str) | |
4917 | { | |
4918 | int ints[NR_CPUS], i; | |
4919 | ||
4920 | str = get_options(str, ARRAY_SIZE(ints), ints); | |
4921 | cpus_clear(cpu_isolated_map); | |
4922 | for (i = 1; i <= ints[0]; i++) | |
4923 | if (ints[i] < NR_CPUS) | |
4924 | cpu_set(ints[i], cpu_isolated_map); | |
4925 | return 1; | |
4926 | } | |
4927 | ||
4928 | __setup ("isolcpus=", isolated_cpu_setup); | |
4929 | ||
4930 | /* | |
4931 | * init_sched_build_groups takes an array of groups, the cpumask we wish | |
4932 | * to span, and a pointer to a function which identifies what group a CPU | |
4933 | * belongs to. The return value of group_fn must be a valid index into the | |
4934 | * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we | |
4935 | * keep track of groups covered with a cpumask_t). | |
4936 | * | |
4937 | * init_sched_build_groups will build a circular linked list of the groups | |
4938 | * covered by the given span, and will set each group's ->cpumask correctly, | |
4939 | * and ->cpu_power to 0. | |
4940 | */ | |
9c1cfda2 JH |
4941 | static void init_sched_build_groups(struct sched_group groups[], cpumask_t span, |
4942 | int (*group_fn)(int cpu)) | |
1da177e4 LT |
4943 | { |
4944 | struct sched_group *first = NULL, *last = NULL; | |
4945 | cpumask_t covered = CPU_MASK_NONE; | |
4946 | int i; | |
4947 | ||
4948 | for_each_cpu_mask(i, span) { | |
4949 | int group = group_fn(i); | |
4950 | struct sched_group *sg = &groups[group]; | |
4951 | int j; | |
4952 | ||
4953 | if (cpu_isset(i, covered)) | |
4954 | continue; | |
4955 | ||
4956 | sg->cpumask = CPU_MASK_NONE; | |
4957 | sg->cpu_power = 0; | |
4958 | ||
4959 | for_each_cpu_mask(j, span) { | |
4960 | if (group_fn(j) != group) | |
4961 | continue; | |
4962 | ||
4963 | cpu_set(j, covered); | |
4964 | cpu_set(j, sg->cpumask); | |
4965 | } | |
4966 | if (!first) | |
4967 | first = sg; | |
4968 | if (last) | |
4969 | last->next = sg; | |
4970 | last = sg; | |
4971 | } | |
4972 | last->next = first; | |
4973 | } | |
4974 | ||
9c1cfda2 | 4975 | #define SD_NODES_PER_DOMAIN 16 |
1da177e4 | 4976 | |
9c1cfda2 JH |
4977 | #ifdef CONFIG_NUMA |
4978 | /** | |
4979 | * find_next_best_node - find the next node to include in a sched_domain | |
4980 | * @node: node whose sched_domain we're building | |
4981 | * @used_nodes: nodes already in the sched_domain | |
4982 | * | |
4983 | * Find the next node to include in a given scheduling domain. Simply | |
4984 | * finds the closest node not already in the @used_nodes map. | |
4985 | * | |
4986 | * Should use nodemask_t. | |
4987 | */ | |
4988 | static int find_next_best_node(int node, unsigned long *used_nodes) | |
4989 | { | |
4990 | int i, n, val, min_val, best_node = 0; | |
4991 | ||
4992 | min_val = INT_MAX; | |
4993 | ||
4994 | for (i = 0; i < MAX_NUMNODES; i++) { | |
4995 | /* Start at @node */ | |
4996 | n = (node + i) % MAX_NUMNODES; | |
4997 | ||
4998 | if (!nr_cpus_node(n)) | |
4999 | continue; | |
5000 | ||
5001 | /* Skip already used nodes */ | |
5002 | if (test_bit(n, used_nodes)) | |
5003 | continue; | |
5004 | ||
5005 | /* Simple min distance search */ | |
5006 | val = node_distance(node, n); | |
5007 | ||
5008 | if (val < min_val) { | |
5009 | min_val = val; | |
5010 | best_node = n; | |
5011 | } | |
5012 | } | |
5013 | ||
5014 | set_bit(best_node, used_nodes); | |
5015 | return best_node; | |
5016 | } | |
5017 | ||
5018 | /** | |
5019 | * sched_domain_node_span - get a cpumask for a node's sched_domain | |
5020 | * @node: node whose cpumask we're constructing | |
5021 | * @size: number of nodes to include in this span | |
5022 | * | |
5023 | * Given a node, construct a good cpumask for its sched_domain to span. It | |
5024 | * should be one that prevents unnecessary balancing, but also spreads tasks | |
5025 | * out optimally. | |
5026 | */ | |
5027 | static cpumask_t sched_domain_node_span(int node) | |
5028 | { | |
5029 | int i; | |
5030 | cpumask_t span, nodemask; | |
5031 | DECLARE_BITMAP(used_nodes, MAX_NUMNODES); | |
5032 | ||
5033 | cpus_clear(span); | |
5034 | bitmap_zero(used_nodes, MAX_NUMNODES); | |
5035 | ||
5036 | nodemask = node_to_cpumask(node); | |
5037 | cpus_or(span, span, nodemask); | |
5038 | set_bit(node, used_nodes); | |
5039 | ||
5040 | for (i = 1; i < SD_NODES_PER_DOMAIN; i++) { | |
5041 | int next_node = find_next_best_node(node, used_nodes); | |
5042 | nodemask = node_to_cpumask(next_node); | |
5043 | cpus_or(span, span, nodemask); | |
5044 | } | |
5045 | ||
5046 | return span; | |
5047 | } | |
5048 | #endif | |
5049 | ||
5050 | /* | |
5051 | * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we | |
5052 | * can switch it on easily if needed. | |
5053 | */ | |
1da177e4 LT |
5054 | #ifdef CONFIG_SCHED_SMT |
5055 | static DEFINE_PER_CPU(struct sched_domain, cpu_domains); | |
5056 | static struct sched_group sched_group_cpus[NR_CPUS]; | |
1a20ff27 | 5057 | static int cpu_to_cpu_group(int cpu) |
1da177e4 LT |
5058 | { |
5059 | return cpu; | |
5060 | } | |
5061 | #endif | |
5062 | ||
5063 | static DEFINE_PER_CPU(struct sched_domain, phys_domains); | |
5064 | static struct sched_group sched_group_phys[NR_CPUS]; | |
1a20ff27 | 5065 | static int cpu_to_phys_group(int cpu) |
1da177e4 LT |
5066 | { |
5067 | #ifdef CONFIG_SCHED_SMT | |
5068 | return first_cpu(cpu_sibling_map[cpu]); | |
5069 | #else | |
5070 | return cpu; | |
5071 | #endif | |
5072 | } | |
5073 | ||
5074 | #ifdef CONFIG_NUMA | |
1da177e4 | 5075 | /* |
9c1cfda2 JH |
5076 | * The init_sched_build_groups can't handle what we want to do with node |
5077 | * groups, so roll our own. Now each node has its own list of groups which | |
5078 | * gets dynamically allocated. | |
1da177e4 | 5079 | */ |
9c1cfda2 | 5080 | static DEFINE_PER_CPU(struct sched_domain, node_domains); |
d1b55138 | 5081 | static struct sched_group **sched_group_nodes_bycpu[NR_CPUS]; |
1da177e4 | 5082 | |
9c1cfda2 | 5083 | static DEFINE_PER_CPU(struct sched_domain, allnodes_domains); |
d1b55138 | 5084 | static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS]; |
9c1cfda2 JH |
5085 | |
5086 | static int cpu_to_allnodes_group(int cpu) | |
5087 | { | |
5088 | return cpu_to_node(cpu); | |
1da177e4 LT |
5089 | } |
5090 | #endif | |
5091 | ||
5092 | /* | |
1a20ff27 DG |
5093 | * Build sched domains for a given set of cpus and attach the sched domains |
5094 | * to the individual cpus | |
1da177e4 | 5095 | */ |
9c1cfda2 | 5096 | void build_sched_domains(const cpumask_t *cpu_map) |
1da177e4 LT |
5097 | { |
5098 | int i; | |
d1b55138 JH |
5099 | #ifdef CONFIG_NUMA |
5100 | struct sched_group **sched_group_nodes = NULL; | |
5101 | struct sched_group *sched_group_allnodes = NULL; | |
5102 | ||
5103 | /* | |
5104 | * Allocate the per-node list of sched groups | |
5105 | */ | |
5106 | sched_group_nodes = kmalloc(sizeof(struct sched_group*)*MAX_NUMNODES, | |
5107 | GFP_ATOMIC); | |
5108 | if (!sched_group_nodes) { | |
5109 | printk(KERN_WARNING "Can not alloc sched group node list\n"); | |
5110 | return; | |
5111 | } | |
5112 | sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes; | |
5113 | #endif | |
1da177e4 LT |
5114 | |
5115 | /* | |
1a20ff27 | 5116 | * Set up domains for cpus specified by the cpu_map. |
1da177e4 | 5117 | */ |
1a20ff27 | 5118 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 LT |
5119 | int group; |
5120 | struct sched_domain *sd = NULL, *p; | |
5121 | cpumask_t nodemask = node_to_cpumask(cpu_to_node(i)); | |
5122 | ||
1a20ff27 | 5123 | cpus_and(nodemask, nodemask, *cpu_map); |
1da177e4 LT |
5124 | |
5125 | #ifdef CONFIG_NUMA | |
d1b55138 | 5126 | if (cpus_weight(*cpu_map) |
9c1cfda2 | 5127 | > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) { |
d1b55138 JH |
5128 | if (!sched_group_allnodes) { |
5129 | sched_group_allnodes | |
5130 | = kmalloc(sizeof(struct sched_group) | |
5131 | * MAX_NUMNODES, | |
5132 | GFP_KERNEL); | |
5133 | if (!sched_group_allnodes) { | |
5134 | printk(KERN_WARNING | |
5135 | "Can not alloc allnodes sched group\n"); | |
5136 | break; | |
5137 | } | |
5138 | sched_group_allnodes_bycpu[i] | |
5139 | = sched_group_allnodes; | |
5140 | } | |
9c1cfda2 JH |
5141 | sd = &per_cpu(allnodes_domains, i); |
5142 | *sd = SD_ALLNODES_INIT; | |
5143 | sd->span = *cpu_map; | |
5144 | group = cpu_to_allnodes_group(i); | |
5145 | sd->groups = &sched_group_allnodes[group]; | |
5146 | p = sd; | |
5147 | } else | |
5148 | p = NULL; | |
5149 | ||
1da177e4 | 5150 | sd = &per_cpu(node_domains, i); |
1da177e4 | 5151 | *sd = SD_NODE_INIT; |
9c1cfda2 JH |
5152 | sd->span = sched_domain_node_span(cpu_to_node(i)); |
5153 | sd->parent = p; | |
5154 | cpus_and(sd->span, sd->span, *cpu_map); | |
1da177e4 LT |
5155 | #endif |
5156 | ||
5157 | p = sd; | |
5158 | sd = &per_cpu(phys_domains, i); | |
5159 | group = cpu_to_phys_group(i); | |
5160 | *sd = SD_CPU_INIT; | |
5161 | sd->span = nodemask; | |
5162 | sd->parent = p; | |
5163 | sd->groups = &sched_group_phys[group]; | |
5164 | ||
5165 | #ifdef CONFIG_SCHED_SMT | |
5166 | p = sd; | |
5167 | sd = &per_cpu(cpu_domains, i); | |
5168 | group = cpu_to_cpu_group(i); | |
5169 | *sd = SD_SIBLING_INIT; | |
5170 | sd->span = cpu_sibling_map[i]; | |
1a20ff27 | 5171 | cpus_and(sd->span, sd->span, *cpu_map); |
1da177e4 LT |
5172 | sd->parent = p; |
5173 | sd->groups = &sched_group_cpus[group]; | |
5174 | #endif | |
5175 | } | |
5176 | ||
5177 | #ifdef CONFIG_SCHED_SMT | |
5178 | /* Set up CPU (sibling) groups */ | |
9c1cfda2 | 5179 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 | 5180 | cpumask_t this_sibling_map = cpu_sibling_map[i]; |
1a20ff27 | 5181 | cpus_and(this_sibling_map, this_sibling_map, *cpu_map); |
1da177e4 LT |
5182 | if (i != first_cpu(this_sibling_map)) |
5183 | continue; | |
5184 | ||
5185 | init_sched_build_groups(sched_group_cpus, this_sibling_map, | |
5186 | &cpu_to_cpu_group); | |
5187 | } | |
5188 | #endif | |
5189 | ||
5190 | /* Set up physical groups */ | |
5191 | for (i = 0; i < MAX_NUMNODES; i++) { | |
5192 | cpumask_t nodemask = node_to_cpumask(i); | |
5193 | ||
1a20ff27 | 5194 | cpus_and(nodemask, nodemask, *cpu_map); |
1da177e4 LT |
5195 | if (cpus_empty(nodemask)) |
5196 | continue; | |
5197 | ||
5198 | init_sched_build_groups(sched_group_phys, nodemask, | |
5199 | &cpu_to_phys_group); | |
5200 | } | |
5201 | ||
5202 | #ifdef CONFIG_NUMA | |
5203 | /* Set up node groups */ | |
d1b55138 JH |
5204 | if (sched_group_allnodes) |
5205 | init_sched_build_groups(sched_group_allnodes, *cpu_map, | |
5206 | &cpu_to_allnodes_group); | |
9c1cfda2 JH |
5207 | |
5208 | for (i = 0; i < MAX_NUMNODES; i++) { | |
5209 | /* Set up node groups */ | |
5210 | struct sched_group *sg, *prev; | |
5211 | cpumask_t nodemask = node_to_cpumask(i); | |
5212 | cpumask_t domainspan; | |
5213 | cpumask_t covered = CPU_MASK_NONE; | |
5214 | int j; | |
5215 | ||
5216 | cpus_and(nodemask, nodemask, *cpu_map); | |
d1b55138 JH |
5217 | if (cpus_empty(nodemask)) { |
5218 | sched_group_nodes[i] = NULL; | |
9c1cfda2 | 5219 | continue; |
d1b55138 | 5220 | } |
9c1cfda2 JH |
5221 | |
5222 | domainspan = sched_domain_node_span(i); | |
5223 | cpus_and(domainspan, domainspan, *cpu_map); | |
5224 | ||
5225 | sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL); | |
5226 | sched_group_nodes[i] = sg; | |
5227 | for_each_cpu_mask(j, nodemask) { | |
5228 | struct sched_domain *sd; | |
5229 | sd = &per_cpu(node_domains, j); | |
5230 | sd->groups = sg; | |
5231 | if (sd->groups == NULL) { | |
5232 | /* Turn off balancing if we have no groups */ | |
5233 | sd->flags = 0; | |
5234 | } | |
5235 | } | |
5236 | if (!sg) { | |
5237 | printk(KERN_WARNING | |
5238 | "Can not alloc domain group for node %d\n", i); | |
5239 | continue; | |
5240 | } | |
5241 | sg->cpu_power = 0; | |
5242 | sg->cpumask = nodemask; | |
5243 | cpus_or(covered, covered, nodemask); | |
5244 | prev = sg; | |
5245 | ||
5246 | for (j = 0; j < MAX_NUMNODES; j++) { | |
5247 | cpumask_t tmp, notcovered; | |
5248 | int n = (i + j) % MAX_NUMNODES; | |
5249 | ||
5250 | cpus_complement(notcovered, covered); | |
5251 | cpus_and(tmp, notcovered, *cpu_map); | |
5252 | cpus_and(tmp, tmp, domainspan); | |
5253 | if (cpus_empty(tmp)) | |
5254 | break; | |
5255 | ||
5256 | nodemask = node_to_cpumask(n); | |
5257 | cpus_and(tmp, tmp, nodemask); | |
5258 | if (cpus_empty(tmp)) | |
5259 | continue; | |
5260 | ||
5261 | sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL); | |
5262 | if (!sg) { | |
5263 | printk(KERN_WARNING | |
5264 | "Can not alloc domain group for node %d\n", j); | |
5265 | break; | |
5266 | } | |
5267 | sg->cpu_power = 0; | |
5268 | sg->cpumask = tmp; | |
5269 | cpus_or(covered, covered, tmp); | |
5270 | prev->next = sg; | |
5271 | prev = sg; | |
5272 | } | |
5273 | prev->next = sched_group_nodes[i]; | |
5274 | } | |
1da177e4 LT |
5275 | #endif |
5276 | ||
5277 | /* Calculate CPU power for physical packages and nodes */ | |
1a20ff27 | 5278 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 LT |
5279 | int power; |
5280 | struct sched_domain *sd; | |
5281 | #ifdef CONFIG_SCHED_SMT | |
5282 | sd = &per_cpu(cpu_domains, i); | |
5283 | power = SCHED_LOAD_SCALE; | |
5284 | sd->groups->cpu_power = power; | |
5285 | #endif | |
5286 | ||
5287 | sd = &per_cpu(phys_domains, i); | |
5288 | power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE * | |
5289 | (cpus_weight(sd->groups->cpumask)-1) / 10; | |
5290 | sd->groups->cpu_power = power; | |
5291 | ||
5292 | #ifdef CONFIG_NUMA | |
9c1cfda2 JH |
5293 | sd = &per_cpu(allnodes_domains, i); |
5294 | if (sd->groups) { | |
5295 | power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE * | |
5296 | (cpus_weight(sd->groups->cpumask)-1) / 10; | |
5297 | sd->groups->cpu_power = power; | |
1da177e4 LT |
5298 | } |
5299 | #endif | |
5300 | } | |
5301 | ||
9c1cfda2 JH |
5302 | #ifdef CONFIG_NUMA |
5303 | for (i = 0; i < MAX_NUMNODES; i++) { | |
5304 | struct sched_group *sg = sched_group_nodes[i]; | |
5305 | int j; | |
5306 | ||
5307 | if (sg == NULL) | |
5308 | continue; | |
5309 | next_sg: | |
5310 | for_each_cpu_mask(j, sg->cpumask) { | |
5311 | struct sched_domain *sd; | |
5312 | int power; | |
5313 | ||
5314 | sd = &per_cpu(phys_domains, j); | |
5315 | if (j != first_cpu(sd->groups->cpumask)) { | |
5316 | /* | |
5317 | * Only add "power" once for each | |
5318 | * physical package. | |
5319 | */ | |
5320 | continue; | |
5321 | } | |
5322 | power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE * | |
5323 | (cpus_weight(sd->groups->cpumask)-1) / 10; | |
5324 | ||
5325 | sg->cpu_power += power; | |
5326 | } | |
5327 | sg = sg->next; | |
5328 | if (sg != sched_group_nodes[i]) | |
5329 | goto next_sg; | |
5330 | } | |
5331 | #endif | |
5332 | ||
1da177e4 | 5333 | /* Attach the domains */ |
1a20ff27 | 5334 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 LT |
5335 | struct sched_domain *sd; |
5336 | #ifdef CONFIG_SCHED_SMT | |
5337 | sd = &per_cpu(cpu_domains, i); | |
5338 | #else | |
5339 | sd = &per_cpu(phys_domains, i); | |
5340 | #endif | |
5341 | cpu_attach_domain(sd, i); | |
5342 | } | |
5343 | } | |
1a20ff27 DG |
5344 | /* |
5345 | * Set up scheduler domains and groups. Callers must hold the hotplug lock. | |
5346 | */ | |
9c1cfda2 | 5347 | static void arch_init_sched_domains(const cpumask_t *cpu_map) |
1a20ff27 DG |
5348 | { |
5349 | cpumask_t cpu_default_map; | |
1da177e4 | 5350 | |
1a20ff27 DG |
5351 | /* |
5352 | * Setup mask for cpus without special case scheduling requirements. | |
5353 | * For now this just excludes isolated cpus, but could be used to | |
5354 | * exclude other special cases in the future. | |
5355 | */ | |
5356 | cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map); | |
5357 | ||
5358 | build_sched_domains(&cpu_default_map); | |
5359 | } | |
5360 | ||
5361 | static void arch_destroy_sched_domains(const cpumask_t *cpu_map) | |
1da177e4 | 5362 | { |
9c1cfda2 JH |
5363 | #ifdef CONFIG_NUMA |
5364 | int i; | |
d1b55138 | 5365 | int cpu; |
1da177e4 | 5366 | |
d1b55138 JH |
5367 | for_each_cpu_mask(cpu, *cpu_map) { |
5368 | struct sched_group *sched_group_allnodes | |
5369 | = sched_group_allnodes_bycpu[cpu]; | |
5370 | struct sched_group **sched_group_nodes | |
5371 | = sched_group_nodes_bycpu[cpu]; | |
9c1cfda2 | 5372 | |
d1b55138 JH |
5373 | if (sched_group_allnodes) { |
5374 | kfree(sched_group_allnodes); | |
5375 | sched_group_allnodes_bycpu[cpu] = NULL; | |
5376 | } | |
5377 | ||
5378 | if (!sched_group_nodes) | |
9c1cfda2 | 5379 | continue; |
d1b55138 JH |
5380 | |
5381 | for (i = 0; i < MAX_NUMNODES; i++) { | |
5382 | cpumask_t nodemask = node_to_cpumask(i); | |
5383 | struct sched_group *oldsg, *sg = sched_group_nodes[i]; | |
5384 | ||
5385 | cpus_and(nodemask, nodemask, *cpu_map); | |
5386 | if (cpus_empty(nodemask)) | |
5387 | continue; | |
5388 | ||
5389 | if (sg == NULL) | |
5390 | continue; | |
5391 | sg = sg->next; | |
9c1cfda2 | 5392 | next_sg: |
d1b55138 JH |
5393 | oldsg = sg; |
5394 | sg = sg->next; | |
5395 | kfree(oldsg); | |
5396 | if (oldsg != sched_group_nodes[i]) | |
5397 | goto next_sg; | |
5398 | } | |
5399 | kfree(sched_group_nodes); | |
5400 | sched_group_nodes_bycpu[cpu] = NULL; | |
9c1cfda2 JH |
5401 | } |
5402 | #endif | |
5403 | } | |
1da177e4 | 5404 | |
1a20ff27 DG |
5405 | /* |
5406 | * Detach sched domains from a group of cpus specified in cpu_map | |
5407 | * These cpus will now be attached to the NULL domain | |
5408 | */ | |
5409 | static inline void detach_destroy_domains(const cpumask_t *cpu_map) | |
5410 | { | |
5411 | int i; | |
5412 | ||
5413 | for_each_cpu_mask(i, *cpu_map) | |
5414 | cpu_attach_domain(NULL, i); | |
5415 | synchronize_sched(); | |
5416 | arch_destroy_sched_domains(cpu_map); | |
5417 | } | |
5418 | ||
5419 | /* | |
5420 | * Partition sched domains as specified by the cpumasks below. | |
5421 | * This attaches all cpus from the cpumasks to the NULL domain, | |
5422 | * waits for a RCU quiescent period, recalculates sched | |
5423 | * domain information and then attaches them back to the | |
5424 | * correct sched domains | |
5425 | * Call with hotplug lock held | |
5426 | */ | |
5427 | void partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2) | |
5428 | { | |
5429 | cpumask_t change_map; | |
5430 | ||
5431 | cpus_and(*partition1, *partition1, cpu_online_map); | |
5432 | cpus_and(*partition2, *partition2, cpu_online_map); | |
5433 | cpus_or(change_map, *partition1, *partition2); | |
5434 | ||
5435 | /* Detach sched domains from all of the affected cpus */ | |
5436 | detach_destroy_domains(&change_map); | |
5437 | if (!cpus_empty(*partition1)) | |
5438 | build_sched_domains(partition1); | |
5439 | if (!cpus_empty(*partition2)) | |
5440 | build_sched_domains(partition2); | |
5441 | } | |
5442 | ||
1da177e4 LT |
5443 | #ifdef CONFIG_HOTPLUG_CPU |
5444 | /* | |
5445 | * Force a reinitialization of the sched domains hierarchy. The domains | |
5446 | * and groups cannot be updated in place without racing with the balancing | |
41c7ce9a | 5447 | * code, so we temporarily attach all running cpus to the NULL domain |
1da177e4 LT |
5448 | * which will prevent rebalancing while the sched domains are recalculated. |
5449 | */ | |
5450 | static int update_sched_domains(struct notifier_block *nfb, | |
5451 | unsigned long action, void *hcpu) | |
5452 | { | |
1da177e4 LT |
5453 | switch (action) { |
5454 | case CPU_UP_PREPARE: | |
5455 | case CPU_DOWN_PREPARE: | |
1a20ff27 | 5456 | detach_destroy_domains(&cpu_online_map); |
1da177e4 LT |
5457 | return NOTIFY_OK; |
5458 | ||
5459 | case CPU_UP_CANCELED: | |
5460 | case CPU_DOWN_FAILED: | |
5461 | case CPU_ONLINE: | |
5462 | case CPU_DEAD: | |
5463 | /* | |
5464 | * Fall through and re-initialise the domains. | |
5465 | */ | |
5466 | break; | |
5467 | default: | |
5468 | return NOTIFY_DONE; | |
5469 | } | |
5470 | ||
5471 | /* The hotplug lock is already held by cpu_up/cpu_down */ | |
1a20ff27 | 5472 | arch_init_sched_domains(&cpu_online_map); |
1da177e4 LT |
5473 | |
5474 | return NOTIFY_OK; | |
5475 | } | |
5476 | #endif | |
5477 | ||
5478 | void __init sched_init_smp(void) | |
5479 | { | |
5480 | lock_cpu_hotplug(); | |
1a20ff27 | 5481 | arch_init_sched_domains(&cpu_online_map); |
1da177e4 LT |
5482 | unlock_cpu_hotplug(); |
5483 | /* XXX: Theoretical race here - CPU may be hotplugged now */ | |
5484 | hotcpu_notifier(update_sched_domains, 0); | |
5485 | } | |
5486 | #else | |
5487 | void __init sched_init_smp(void) | |
5488 | { | |
5489 | } | |
5490 | #endif /* CONFIG_SMP */ | |
5491 | ||
5492 | int in_sched_functions(unsigned long addr) | |
5493 | { | |
5494 | /* Linker adds these: start and end of __sched functions */ | |
5495 | extern char __sched_text_start[], __sched_text_end[]; | |
5496 | return in_lock_functions(addr) || | |
5497 | (addr >= (unsigned long)__sched_text_start | |
5498 | && addr < (unsigned long)__sched_text_end); | |
5499 | } | |
5500 | ||
5501 | void __init sched_init(void) | |
5502 | { | |
5503 | runqueue_t *rq; | |
5504 | int i, j, k; | |
5505 | ||
5506 | for (i = 0; i < NR_CPUS; i++) { | |
5507 | prio_array_t *array; | |
5508 | ||
5509 | rq = cpu_rq(i); | |
5510 | spin_lock_init(&rq->lock); | |
7897986b | 5511 | rq->nr_running = 0; |
1da177e4 LT |
5512 | rq->active = rq->arrays; |
5513 | rq->expired = rq->arrays + 1; | |
5514 | rq->best_expired_prio = MAX_PRIO; | |
5515 | ||
5516 | #ifdef CONFIG_SMP | |
41c7ce9a | 5517 | rq->sd = NULL; |
7897986b NP |
5518 | for (j = 1; j < 3; j++) |
5519 | rq->cpu_load[j] = 0; | |
1da177e4 LT |
5520 | rq->active_balance = 0; |
5521 | rq->push_cpu = 0; | |
5522 | rq->migration_thread = NULL; | |
5523 | INIT_LIST_HEAD(&rq->migration_queue); | |
5524 | #endif | |
5525 | atomic_set(&rq->nr_iowait, 0); | |
5526 | ||
5527 | for (j = 0; j < 2; j++) { | |
5528 | array = rq->arrays + j; | |
5529 | for (k = 0; k < MAX_PRIO; k++) { | |
5530 | INIT_LIST_HEAD(array->queue + k); | |
5531 | __clear_bit(k, array->bitmap); | |
5532 | } | |
5533 | // delimiter for bitsearch | |
5534 | __set_bit(MAX_PRIO, array->bitmap); | |
5535 | } | |
5536 | } | |
5537 | ||
5538 | /* | |
5539 | * The boot idle thread does lazy MMU switching as well: | |
5540 | */ | |
5541 | atomic_inc(&init_mm.mm_count); | |
5542 | enter_lazy_tlb(&init_mm, current); | |
5543 | ||
5544 | /* | |
5545 | * Make us the idle thread. Technically, schedule() should not be | |
5546 | * called from this thread, however somewhere below it might be, | |
5547 | * but because we are the idle thread, we just pick up running again | |
5548 | * when this runqueue becomes "idle". | |
5549 | */ | |
5550 | init_idle(current, smp_processor_id()); | |
5551 | } | |
5552 | ||
5553 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP | |
5554 | void __might_sleep(char *file, int line) | |
5555 | { | |
5556 | #if defined(in_atomic) | |
5557 | static unsigned long prev_jiffy; /* ratelimiting */ | |
5558 | ||
5559 | if ((in_atomic() || irqs_disabled()) && | |
5560 | system_state == SYSTEM_RUNNING && !oops_in_progress) { | |
5561 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) | |
5562 | return; | |
5563 | prev_jiffy = jiffies; | |
5564 | printk(KERN_ERR "Debug: sleeping function called from invalid" | |
5565 | " context at %s:%d\n", file, line); | |
5566 | printk("in_atomic():%d, irqs_disabled():%d\n", | |
5567 | in_atomic(), irqs_disabled()); | |
5568 | dump_stack(); | |
5569 | } | |
5570 | #endif | |
5571 | } | |
5572 | EXPORT_SYMBOL(__might_sleep); | |
5573 | #endif | |
5574 | ||
5575 | #ifdef CONFIG_MAGIC_SYSRQ | |
5576 | void normalize_rt_tasks(void) | |
5577 | { | |
5578 | struct task_struct *p; | |
5579 | prio_array_t *array; | |
5580 | unsigned long flags; | |
5581 | runqueue_t *rq; | |
5582 | ||
5583 | read_lock_irq(&tasklist_lock); | |
5584 | for_each_process (p) { | |
5585 | if (!rt_task(p)) | |
5586 | continue; | |
5587 | ||
5588 | rq = task_rq_lock(p, &flags); | |
5589 | ||
5590 | array = p->array; | |
5591 | if (array) | |
5592 | deactivate_task(p, task_rq(p)); | |
5593 | __setscheduler(p, SCHED_NORMAL, 0); | |
5594 | if (array) { | |
5595 | __activate_task(p, task_rq(p)); | |
5596 | resched_task(rq->curr); | |
5597 | } | |
5598 | ||
5599 | task_rq_unlock(rq, &flags); | |
5600 | } | |
5601 | read_unlock_irq(&tasklist_lock); | |
5602 | } | |
5603 | ||
5604 | #endif /* CONFIG_MAGIC_SYSRQ */ | |
1df5c10a LT |
5605 | |
5606 | #ifdef CONFIG_IA64 | |
5607 | /* | |
5608 | * These functions are only useful for the IA64 MCA handling. | |
5609 | * | |
5610 | * They can only be called when the whole system has been | |
5611 | * stopped - every CPU needs to be quiescent, and no scheduling | |
5612 | * activity can take place. Using them for anything else would | |
5613 | * be a serious bug, and as a result, they aren't even visible | |
5614 | * under any other configuration. | |
5615 | */ | |
5616 | ||
5617 | /** | |
5618 | * curr_task - return the current task for a given cpu. | |
5619 | * @cpu: the processor in question. | |
5620 | * | |
5621 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
5622 | */ | |
5623 | task_t *curr_task(int cpu) | |
5624 | { | |
5625 | return cpu_curr(cpu); | |
5626 | } | |
5627 | ||
5628 | /** | |
5629 | * set_curr_task - set the current task for a given cpu. | |
5630 | * @cpu: the processor in question. | |
5631 | * @p: the task pointer to set. | |
5632 | * | |
5633 | * Description: This function must only be used when non-maskable interrupts | |
5634 | * are serviced on a separate stack. It allows the architecture to switch the | |
5635 | * notion of the current task on a cpu in a non-blocking manner. This function | |
5636 | * must be called with all CPU's synchronized, and interrupts disabled, the | |
5637 | * and caller must save the original value of the current task (see | |
5638 | * curr_task() above) and restore that value before reenabling interrupts and | |
5639 | * re-starting the system. | |
5640 | * | |
5641 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
5642 | */ | |
5643 | void set_curr_task(int cpu, task_t *p) | |
5644 | { | |
5645 | cpu_curr(cpu) = p; | |
5646 | } | |
5647 | ||
5648 | #endif |