]>
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 | |
b910472d | 209 | unsigned long prio_bias; |
7897986b | 210 | unsigned long cpu_load[3]; |
1da177e4 LT |
211 | #endif |
212 | unsigned long long nr_switches; | |
213 | ||
214 | /* | |
215 | * This is part of a global counter where only the total sum | |
216 | * over all CPUs matters. A task can increase this counter on | |
217 | * one CPU and if it got migrated afterwards it may decrease | |
218 | * it on another CPU. Always updated under the runqueue lock: | |
219 | */ | |
220 | unsigned long nr_uninterruptible; | |
221 | ||
222 | unsigned long expired_timestamp; | |
223 | unsigned long long timestamp_last_tick; | |
224 | task_t *curr, *idle; | |
225 | struct mm_struct *prev_mm; | |
226 | prio_array_t *active, *expired, arrays[2]; | |
227 | int best_expired_prio; | |
228 | atomic_t nr_iowait; | |
229 | ||
230 | #ifdef CONFIG_SMP | |
231 | struct sched_domain *sd; | |
232 | ||
233 | /* For active balancing */ | |
234 | int active_balance; | |
235 | int push_cpu; | |
236 | ||
237 | task_t *migration_thread; | |
238 | struct list_head migration_queue; | |
239 | #endif | |
240 | ||
241 | #ifdef CONFIG_SCHEDSTATS | |
242 | /* latency stats */ | |
243 | struct sched_info rq_sched_info; | |
244 | ||
245 | /* sys_sched_yield() stats */ | |
246 | unsigned long yld_exp_empty; | |
247 | unsigned long yld_act_empty; | |
248 | unsigned long yld_both_empty; | |
249 | unsigned long yld_cnt; | |
250 | ||
251 | /* schedule() stats */ | |
252 | unsigned long sched_switch; | |
253 | unsigned long sched_cnt; | |
254 | unsigned long sched_goidle; | |
255 | ||
256 | /* try_to_wake_up() stats */ | |
257 | unsigned long ttwu_cnt; | |
258 | unsigned long ttwu_local; | |
259 | #endif | |
260 | }; | |
261 | ||
262 | static DEFINE_PER_CPU(struct runqueue, runqueues); | |
263 | ||
674311d5 NP |
264 | /* |
265 | * The domain tree (rq->sd) is protected by RCU's quiescent state transition. | |
1a20ff27 | 266 | * See detach_destroy_domains: synchronize_sched for details. |
674311d5 NP |
267 | * |
268 | * The domain tree of any CPU may only be accessed from within | |
269 | * preempt-disabled sections. | |
270 | */ | |
1da177e4 | 271 | #define for_each_domain(cpu, domain) \ |
674311d5 | 272 | for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent) |
1da177e4 LT |
273 | |
274 | #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) | |
275 | #define this_rq() (&__get_cpu_var(runqueues)) | |
276 | #define task_rq(p) cpu_rq(task_cpu(p)) | |
277 | #define cpu_curr(cpu) (cpu_rq(cpu)->curr) | |
278 | ||
1da177e4 | 279 | #ifndef prepare_arch_switch |
4866cde0 NP |
280 | # define prepare_arch_switch(next) do { } while (0) |
281 | #endif | |
282 | #ifndef finish_arch_switch | |
283 | # define finish_arch_switch(prev) do { } while (0) | |
284 | #endif | |
285 | ||
286 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW | |
287 | static inline int task_running(runqueue_t *rq, task_t *p) | |
288 | { | |
289 | return rq->curr == p; | |
290 | } | |
291 | ||
292 | static inline void prepare_lock_switch(runqueue_t *rq, task_t *next) | |
293 | { | |
294 | } | |
295 | ||
296 | static inline void finish_lock_switch(runqueue_t *rq, task_t *prev) | |
297 | { | |
da04c035 IM |
298 | #ifdef CONFIG_DEBUG_SPINLOCK |
299 | /* this is a valid case when another task releases the spinlock */ | |
300 | rq->lock.owner = current; | |
301 | #endif | |
4866cde0 NP |
302 | spin_unlock_irq(&rq->lock); |
303 | } | |
304 | ||
305 | #else /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
306 | static inline int task_running(runqueue_t *rq, task_t *p) | |
307 | { | |
308 | #ifdef CONFIG_SMP | |
309 | return p->oncpu; | |
310 | #else | |
311 | return rq->curr == p; | |
312 | #endif | |
313 | } | |
314 | ||
315 | static inline void prepare_lock_switch(runqueue_t *rq, task_t *next) | |
316 | { | |
317 | #ifdef CONFIG_SMP | |
318 | /* | |
319 | * We can optimise this out completely for !SMP, because the | |
320 | * SMP rebalancing from interrupt is the only thing that cares | |
321 | * here. | |
322 | */ | |
323 | next->oncpu = 1; | |
324 | #endif | |
325 | #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
326 | spin_unlock_irq(&rq->lock); | |
327 | #else | |
328 | spin_unlock(&rq->lock); | |
329 | #endif | |
330 | } | |
331 | ||
332 | static inline void finish_lock_switch(runqueue_t *rq, task_t *prev) | |
333 | { | |
334 | #ifdef CONFIG_SMP | |
335 | /* | |
336 | * After ->oncpu is cleared, the task can be moved to a different CPU. | |
337 | * We must ensure this doesn't happen until the switch is completely | |
338 | * finished. | |
339 | */ | |
340 | smp_wmb(); | |
341 | prev->oncpu = 0; | |
342 | #endif | |
343 | #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
344 | local_irq_enable(); | |
1da177e4 | 345 | #endif |
4866cde0 NP |
346 | } |
347 | #endif /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
1da177e4 LT |
348 | |
349 | /* | |
350 | * task_rq_lock - lock the runqueue a given task resides on and disable | |
351 | * interrupts. Note the ordering: we can safely lookup the task_rq without | |
352 | * explicitly disabling preemption. | |
353 | */ | |
354 | static inline runqueue_t *task_rq_lock(task_t *p, unsigned long *flags) | |
355 | __acquires(rq->lock) | |
356 | { | |
357 | struct runqueue *rq; | |
358 | ||
359 | repeat_lock_task: | |
360 | local_irq_save(*flags); | |
361 | rq = task_rq(p); | |
362 | spin_lock(&rq->lock); | |
363 | if (unlikely(rq != task_rq(p))) { | |
364 | spin_unlock_irqrestore(&rq->lock, *flags); | |
365 | goto repeat_lock_task; | |
366 | } | |
367 | return rq; | |
368 | } | |
369 | ||
370 | static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags) | |
371 | __releases(rq->lock) | |
372 | { | |
373 | spin_unlock_irqrestore(&rq->lock, *flags); | |
374 | } | |
375 | ||
376 | #ifdef CONFIG_SCHEDSTATS | |
377 | /* | |
378 | * bump this up when changing the output format or the meaning of an existing | |
379 | * format, so that tools can adapt (or abort) | |
380 | */ | |
68767a0a | 381 | #define SCHEDSTAT_VERSION 12 |
1da177e4 LT |
382 | |
383 | static int show_schedstat(struct seq_file *seq, void *v) | |
384 | { | |
385 | int cpu; | |
386 | ||
387 | seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION); | |
388 | seq_printf(seq, "timestamp %lu\n", jiffies); | |
389 | for_each_online_cpu(cpu) { | |
390 | runqueue_t *rq = cpu_rq(cpu); | |
391 | #ifdef CONFIG_SMP | |
392 | struct sched_domain *sd; | |
393 | int dcnt = 0; | |
394 | #endif | |
395 | ||
396 | /* runqueue-specific stats */ | |
397 | seq_printf(seq, | |
398 | "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu", | |
399 | cpu, rq->yld_both_empty, | |
400 | rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt, | |
401 | rq->sched_switch, rq->sched_cnt, rq->sched_goidle, | |
402 | rq->ttwu_cnt, rq->ttwu_local, | |
403 | rq->rq_sched_info.cpu_time, | |
404 | rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt); | |
405 | ||
406 | seq_printf(seq, "\n"); | |
407 | ||
408 | #ifdef CONFIG_SMP | |
409 | /* domain-specific stats */ | |
674311d5 | 410 | preempt_disable(); |
1da177e4 LT |
411 | for_each_domain(cpu, sd) { |
412 | enum idle_type itype; | |
413 | char mask_str[NR_CPUS]; | |
414 | ||
415 | cpumask_scnprintf(mask_str, NR_CPUS, sd->span); | |
416 | seq_printf(seq, "domain%d %s", dcnt++, mask_str); | |
417 | for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES; | |
418 | itype++) { | |
419 | seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu", | |
420 | sd->lb_cnt[itype], | |
421 | sd->lb_balanced[itype], | |
422 | sd->lb_failed[itype], | |
423 | sd->lb_imbalance[itype], | |
424 | sd->lb_gained[itype], | |
425 | sd->lb_hot_gained[itype], | |
426 | sd->lb_nobusyq[itype], | |
427 | sd->lb_nobusyg[itype]); | |
428 | } | |
68767a0a | 429 | seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n", |
1da177e4 | 430 | sd->alb_cnt, sd->alb_failed, sd->alb_pushed, |
68767a0a NP |
431 | sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed, |
432 | sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed, | |
1da177e4 LT |
433 | sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance); |
434 | } | |
674311d5 | 435 | preempt_enable(); |
1da177e4 LT |
436 | #endif |
437 | } | |
438 | return 0; | |
439 | } | |
440 | ||
441 | static int schedstat_open(struct inode *inode, struct file *file) | |
442 | { | |
443 | unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32); | |
444 | char *buf = kmalloc(size, GFP_KERNEL); | |
445 | struct seq_file *m; | |
446 | int res; | |
447 | ||
448 | if (!buf) | |
449 | return -ENOMEM; | |
450 | res = single_open(file, show_schedstat, NULL); | |
451 | if (!res) { | |
452 | m = file->private_data; | |
453 | m->buf = buf; | |
454 | m->size = size; | |
455 | } else | |
456 | kfree(buf); | |
457 | return res; | |
458 | } | |
459 | ||
460 | struct file_operations proc_schedstat_operations = { | |
461 | .open = schedstat_open, | |
462 | .read = seq_read, | |
463 | .llseek = seq_lseek, | |
464 | .release = single_release, | |
465 | }; | |
466 | ||
467 | # define schedstat_inc(rq, field) do { (rq)->field++; } while (0) | |
468 | # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0) | |
469 | #else /* !CONFIG_SCHEDSTATS */ | |
470 | # define schedstat_inc(rq, field) do { } while (0) | |
471 | # define schedstat_add(rq, field, amt) do { } while (0) | |
472 | #endif | |
473 | ||
474 | /* | |
475 | * rq_lock - lock a given runqueue and disable interrupts. | |
476 | */ | |
477 | static inline runqueue_t *this_rq_lock(void) | |
478 | __acquires(rq->lock) | |
479 | { | |
480 | runqueue_t *rq; | |
481 | ||
482 | local_irq_disable(); | |
483 | rq = this_rq(); | |
484 | spin_lock(&rq->lock); | |
485 | ||
486 | return rq; | |
487 | } | |
488 | ||
1da177e4 LT |
489 | #ifdef CONFIG_SCHEDSTATS |
490 | /* | |
491 | * Called when a process is dequeued from the active array and given | |
492 | * the cpu. We should note that with the exception of interactive | |
493 | * tasks, the expired queue will become the active queue after the active | |
494 | * queue is empty, without explicitly dequeuing and requeuing tasks in the | |
495 | * expired queue. (Interactive tasks may be requeued directly to the | |
496 | * active queue, thus delaying tasks in the expired queue from running; | |
497 | * see scheduler_tick()). | |
498 | * | |
499 | * This function is only called from sched_info_arrive(), rather than | |
500 | * dequeue_task(). Even though a task may be queued and dequeued multiple | |
501 | * times as it is shuffled about, we're really interested in knowing how | |
502 | * long it was from the *first* time it was queued to the time that it | |
503 | * finally hit a cpu. | |
504 | */ | |
505 | static inline void sched_info_dequeued(task_t *t) | |
506 | { | |
507 | t->sched_info.last_queued = 0; | |
508 | } | |
509 | ||
510 | /* | |
511 | * Called when a task finally hits the cpu. We can now calculate how | |
512 | * long it was waiting to run. We also note when it began so that we | |
513 | * can keep stats on how long its timeslice is. | |
514 | */ | |
515 | static inline void sched_info_arrive(task_t *t) | |
516 | { | |
517 | unsigned long now = jiffies, diff = 0; | |
518 | struct runqueue *rq = task_rq(t); | |
519 | ||
520 | if (t->sched_info.last_queued) | |
521 | diff = now - t->sched_info.last_queued; | |
522 | sched_info_dequeued(t); | |
523 | t->sched_info.run_delay += diff; | |
524 | t->sched_info.last_arrival = now; | |
525 | t->sched_info.pcnt++; | |
526 | ||
527 | if (!rq) | |
528 | return; | |
529 | ||
530 | rq->rq_sched_info.run_delay += diff; | |
531 | rq->rq_sched_info.pcnt++; | |
532 | } | |
533 | ||
534 | /* | |
535 | * Called when a process is queued into either the active or expired | |
536 | * array. The time is noted and later used to determine how long we | |
537 | * had to wait for us to reach the cpu. Since the expired queue will | |
538 | * become the active queue after active queue is empty, without dequeuing | |
539 | * and requeuing any tasks, we are interested in queuing to either. It | |
540 | * is unusual but not impossible for tasks to be dequeued and immediately | |
541 | * requeued in the same or another array: this can happen in sched_yield(), | |
542 | * set_user_nice(), and even load_balance() as it moves tasks from runqueue | |
543 | * to runqueue. | |
544 | * | |
545 | * This function is only called from enqueue_task(), but also only updates | |
546 | * the timestamp if it is already not set. It's assumed that | |
547 | * sched_info_dequeued() will clear that stamp when appropriate. | |
548 | */ | |
549 | static inline void sched_info_queued(task_t *t) | |
550 | { | |
551 | if (!t->sched_info.last_queued) | |
552 | t->sched_info.last_queued = jiffies; | |
553 | } | |
554 | ||
555 | /* | |
556 | * Called when a process ceases being the active-running process, either | |
557 | * voluntarily or involuntarily. Now we can calculate how long we ran. | |
558 | */ | |
559 | static inline void sched_info_depart(task_t *t) | |
560 | { | |
561 | struct runqueue *rq = task_rq(t); | |
562 | unsigned long diff = jiffies - t->sched_info.last_arrival; | |
563 | ||
564 | t->sched_info.cpu_time += diff; | |
565 | ||
566 | if (rq) | |
567 | rq->rq_sched_info.cpu_time += diff; | |
568 | } | |
569 | ||
570 | /* | |
571 | * Called when tasks are switched involuntarily due, typically, to expiring | |
572 | * their time slice. (This may also be called when switching to or from | |
573 | * the idle task.) We are only called when prev != next. | |
574 | */ | |
575 | static inline void sched_info_switch(task_t *prev, task_t *next) | |
576 | { | |
577 | struct runqueue *rq = task_rq(prev); | |
578 | ||
579 | /* | |
580 | * prev now departs the cpu. It's not interesting to record | |
581 | * stats about how efficient we were at scheduling the idle | |
582 | * process, however. | |
583 | */ | |
584 | if (prev != rq->idle) | |
585 | sched_info_depart(prev); | |
586 | ||
587 | if (next != rq->idle) | |
588 | sched_info_arrive(next); | |
589 | } | |
590 | #else | |
591 | #define sched_info_queued(t) do { } while (0) | |
592 | #define sched_info_switch(t, next) do { } while (0) | |
593 | #endif /* CONFIG_SCHEDSTATS */ | |
594 | ||
595 | /* | |
596 | * Adding/removing a task to/from a priority array: | |
597 | */ | |
598 | static void dequeue_task(struct task_struct *p, prio_array_t *array) | |
599 | { | |
600 | array->nr_active--; | |
601 | list_del(&p->run_list); | |
602 | if (list_empty(array->queue + p->prio)) | |
603 | __clear_bit(p->prio, array->bitmap); | |
604 | } | |
605 | ||
606 | static void enqueue_task(struct task_struct *p, prio_array_t *array) | |
607 | { | |
608 | sched_info_queued(p); | |
609 | list_add_tail(&p->run_list, array->queue + p->prio); | |
610 | __set_bit(p->prio, array->bitmap); | |
611 | array->nr_active++; | |
612 | p->array = array; | |
613 | } | |
614 | ||
615 | /* | |
616 | * Put task to the end of the run list without the overhead of dequeue | |
617 | * followed by enqueue. | |
618 | */ | |
619 | static void requeue_task(struct task_struct *p, prio_array_t *array) | |
620 | { | |
621 | list_move_tail(&p->run_list, array->queue + p->prio); | |
622 | } | |
623 | ||
624 | static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array) | |
625 | { | |
626 | list_add(&p->run_list, array->queue + p->prio); | |
627 | __set_bit(p->prio, array->bitmap); | |
628 | array->nr_active++; | |
629 | p->array = array; | |
630 | } | |
631 | ||
632 | /* | |
633 | * effective_prio - return the priority that is based on the static | |
634 | * priority but is modified by bonuses/penalties. | |
635 | * | |
636 | * We scale the actual sleep average [0 .... MAX_SLEEP_AVG] | |
637 | * into the -5 ... 0 ... +5 bonus/penalty range. | |
638 | * | |
639 | * We use 25% of the full 0...39 priority range so that: | |
640 | * | |
641 | * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs. | |
642 | * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks. | |
643 | * | |
644 | * Both properties are important to certain workloads. | |
645 | */ | |
646 | static int effective_prio(task_t *p) | |
647 | { | |
648 | int bonus, prio; | |
649 | ||
650 | if (rt_task(p)) | |
651 | return p->prio; | |
652 | ||
653 | bonus = CURRENT_BONUS(p) - MAX_BONUS / 2; | |
654 | ||
655 | prio = p->static_prio - bonus; | |
656 | if (prio < MAX_RT_PRIO) | |
657 | prio = MAX_RT_PRIO; | |
658 | if (prio > MAX_PRIO-1) | |
659 | prio = MAX_PRIO-1; | |
660 | return prio; | |
661 | } | |
662 | ||
b910472d | 663 | #ifdef CONFIG_SMP |
dad1c65c | 664 | static inline void inc_prio_bias(runqueue_t *rq, int prio) |
b910472d | 665 | { |
dad1c65c | 666 | rq->prio_bias += MAX_PRIO - prio; |
b910472d CK |
667 | } |
668 | ||
dad1c65c | 669 | static inline void dec_prio_bias(runqueue_t *rq, int prio) |
b910472d | 670 | { |
dad1c65c | 671 | rq->prio_bias -= MAX_PRIO - prio; |
b910472d | 672 | } |
ede3d0fb CK |
673 | |
674 | static inline void inc_nr_running(task_t *p, runqueue_t *rq) | |
675 | { | |
676 | rq->nr_running++; | |
677 | if (rt_task(p)) { | |
678 | if (p != rq->migration_thread) | |
679 | /* | |
680 | * The migration thread does the actual balancing. Do | |
681 | * not bias by its priority as the ultra high priority | |
682 | * will skew balancing adversely. | |
683 | */ | |
684 | inc_prio_bias(rq, p->prio); | |
685 | } else | |
686 | inc_prio_bias(rq, p->static_prio); | |
687 | } | |
688 | ||
689 | static inline void dec_nr_running(task_t *p, runqueue_t *rq) | |
690 | { | |
691 | rq->nr_running--; | |
692 | if (rt_task(p)) { | |
693 | if (p != rq->migration_thread) | |
694 | dec_prio_bias(rq, p->prio); | |
695 | } else | |
696 | dec_prio_bias(rq, p->static_prio); | |
697 | } | |
b910472d | 698 | #else |
dad1c65c | 699 | static inline void inc_prio_bias(runqueue_t *rq, int prio) |
b910472d CK |
700 | { |
701 | } | |
702 | ||
dad1c65c | 703 | static inline void dec_prio_bias(runqueue_t *rq, int prio) |
b910472d CK |
704 | { |
705 | } | |
b910472d CK |
706 | |
707 | static inline void inc_nr_running(task_t *p, runqueue_t *rq) | |
708 | { | |
709 | rq->nr_running++; | |
b910472d CK |
710 | } |
711 | ||
712 | static inline void dec_nr_running(task_t *p, runqueue_t *rq) | |
713 | { | |
714 | rq->nr_running--; | |
b910472d | 715 | } |
ede3d0fb | 716 | #endif |
b910472d | 717 | |
1da177e4 LT |
718 | /* |
719 | * __activate_task - move a task to the runqueue. | |
720 | */ | |
721 | static inline void __activate_task(task_t *p, runqueue_t *rq) | |
722 | { | |
723 | enqueue_task(p, rq->active); | |
b910472d | 724 | inc_nr_running(p, rq); |
1da177e4 LT |
725 | } |
726 | ||
727 | /* | |
728 | * __activate_idle_task - move idle task to the _front_ of runqueue. | |
729 | */ | |
730 | static inline void __activate_idle_task(task_t *p, runqueue_t *rq) | |
731 | { | |
732 | enqueue_task_head(p, rq->active); | |
b910472d | 733 | inc_nr_running(p, rq); |
1da177e4 LT |
734 | } |
735 | ||
a3464a10 | 736 | static int recalc_task_prio(task_t *p, unsigned long long now) |
1da177e4 LT |
737 | { |
738 | /* Caller must always ensure 'now >= p->timestamp' */ | |
739 | unsigned long long __sleep_time = now - p->timestamp; | |
740 | unsigned long sleep_time; | |
741 | ||
742 | if (__sleep_time > NS_MAX_SLEEP_AVG) | |
743 | sleep_time = NS_MAX_SLEEP_AVG; | |
744 | else | |
745 | sleep_time = (unsigned long)__sleep_time; | |
746 | ||
747 | if (likely(sleep_time > 0)) { | |
748 | /* | |
749 | * User tasks that sleep a long time are categorised as | |
750 | * idle and will get just interactive status to stay active & | |
751 | * prevent them suddenly becoming cpu hogs and starving | |
752 | * other processes. | |
753 | */ | |
754 | if (p->mm && p->activated != -1 && | |
755 | sleep_time > INTERACTIVE_SLEEP(p)) { | |
756 | p->sleep_avg = JIFFIES_TO_NS(MAX_SLEEP_AVG - | |
757 | DEF_TIMESLICE); | |
758 | } else { | |
759 | /* | |
760 | * The lower the sleep avg a task has the more | |
761 | * rapidly it will rise with sleep time. | |
762 | */ | |
763 | sleep_time *= (MAX_BONUS - CURRENT_BONUS(p)) ? : 1; | |
764 | ||
765 | /* | |
766 | * Tasks waking from uninterruptible sleep are | |
767 | * limited in their sleep_avg rise as they | |
768 | * are likely to be waiting on I/O | |
769 | */ | |
770 | if (p->activated == -1 && p->mm) { | |
771 | if (p->sleep_avg >= INTERACTIVE_SLEEP(p)) | |
772 | sleep_time = 0; | |
773 | else if (p->sleep_avg + sleep_time >= | |
774 | INTERACTIVE_SLEEP(p)) { | |
775 | p->sleep_avg = INTERACTIVE_SLEEP(p); | |
776 | sleep_time = 0; | |
777 | } | |
778 | } | |
779 | ||
780 | /* | |
781 | * This code gives a bonus to interactive tasks. | |
782 | * | |
783 | * The boost works by updating the 'average sleep time' | |
784 | * value here, based on ->timestamp. The more time a | |
785 | * task spends sleeping, the higher the average gets - | |
786 | * and the higher the priority boost gets as well. | |
787 | */ | |
788 | p->sleep_avg += sleep_time; | |
789 | ||
790 | if (p->sleep_avg > NS_MAX_SLEEP_AVG) | |
791 | p->sleep_avg = NS_MAX_SLEEP_AVG; | |
792 | } | |
793 | } | |
794 | ||
a3464a10 | 795 | return effective_prio(p); |
1da177e4 LT |
796 | } |
797 | ||
798 | /* | |
799 | * activate_task - move a task to the runqueue and do priority recalculation | |
800 | * | |
801 | * Update all the scheduling statistics stuff. (sleep average | |
802 | * calculation, priority modifiers, etc.) | |
803 | */ | |
804 | static void activate_task(task_t *p, runqueue_t *rq, int local) | |
805 | { | |
806 | unsigned long long now; | |
807 | ||
808 | now = sched_clock(); | |
809 | #ifdef CONFIG_SMP | |
810 | if (!local) { | |
811 | /* Compensate for drifting sched_clock */ | |
812 | runqueue_t *this_rq = this_rq(); | |
813 | now = (now - this_rq->timestamp_last_tick) | |
814 | + rq->timestamp_last_tick; | |
815 | } | |
816 | #endif | |
817 | ||
a3464a10 | 818 | p->prio = recalc_task_prio(p, now); |
1da177e4 LT |
819 | |
820 | /* | |
821 | * This checks to make sure it's not an uninterruptible task | |
822 | * that is now waking up. | |
823 | */ | |
824 | if (!p->activated) { | |
825 | /* | |
826 | * Tasks which were woken up by interrupts (ie. hw events) | |
827 | * are most likely of interactive nature. So we give them | |
828 | * the credit of extending their sleep time to the period | |
829 | * of time they spend on the runqueue, waiting for execution | |
830 | * on a CPU, first time around: | |
831 | */ | |
832 | if (in_interrupt()) | |
833 | p->activated = 2; | |
834 | else { | |
835 | /* | |
836 | * Normal first-time wakeups get a credit too for | |
837 | * on-runqueue time, but it will be weighted down: | |
838 | */ | |
839 | p->activated = 1; | |
840 | } | |
841 | } | |
842 | p->timestamp = now; | |
843 | ||
844 | __activate_task(p, rq); | |
845 | } | |
846 | ||
847 | /* | |
848 | * deactivate_task - remove a task from the runqueue. | |
849 | */ | |
850 | static void deactivate_task(struct task_struct *p, runqueue_t *rq) | |
851 | { | |
b910472d | 852 | dec_nr_running(p, rq); |
1da177e4 LT |
853 | dequeue_task(p, p->array); |
854 | p->array = NULL; | |
855 | } | |
856 | ||
857 | /* | |
858 | * resched_task - mark a task 'to be rescheduled now'. | |
859 | * | |
860 | * On UP this means the setting of the need_resched flag, on SMP it | |
861 | * might also involve a cross-CPU call to trigger the scheduler on | |
862 | * the target CPU. | |
863 | */ | |
864 | #ifdef CONFIG_SMP | |
865 | static void resched_task(task_t *p) | |
866 | { | |
64c7c8f8 | 867 | int cpu; |
1da177e4 LT |
868 | |
869 | assert_spin_locked(&task_rq(p)->lock); | |
870 | ||
64c7c8f8 NP |
871 | if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED))) |
872 | return; | |
873 | ||
874 | set_tsk_thread_flag(p, TIF_NEED_RESCHED); | |
1da177e4 | 875 | |
64c7c8f8 NP |
876 | cpu = task_cpu(p); |
877 | if (cpu == smp_processor_id()) | |
878 | return; | |
879 | ||
880 | /* NEED_RESCHED must be visible before we test POLLING_NRFLAG */ | |
881 | smp_mb(); | |
882 | if (!test_tsk_thread_flag(p, TIF_POLLING_NRFLAG)) | |
883 | smp_send_reschedule(cpu); | |
1da177e4 LT |
884 | } |
885 | #else | |
886 | static inline void resched_task(task_t *p) | |
887 | { | |
64c7c8f8 | 888 | assert_spin_locked(&task_rq(p)->lock); |
1da177e4 LT |
889 | set_tsk_need_resched(p); |
890 | } | |
891 | #endif | |
892 | ||
893 | /** | |
894 | * task_curr - is this task currently executing on a CPU? | |
895 | * @p: the task in question. | |
896 | */ | |
897 | inline int task_curr(const task_t *p) | |
898 | { | |
899 | return cpu_curr(task_cpu(p)) == p; | |
900 | } | |
901 | ||
902 | #ifdef CONFIG_SMP | |
1da177e4 LT |
903 | typedef struct { |
904 | struct list_head list; | |
1da177e4 | 905 | |
1da177e4 LT |
906 | task_t *task; |
907 | int dest_cpu; | |
908 | ||
1da177e4 LT |
909 | struct completion done; |
910 | } migration_req_t; | |
911 | ||
912 | /* | |
913 | * The task's runqueue lock must be held. | |
914 | * Returns true if you have to wait for migration thread. | |
915 | */ | |
916 | static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req) | |
917 | { | |
918 | runqueue_t *rq = task_rq(p); | |
919 | ||
920 | /* | |
921 | * If the task is not on a runqueue (and not running), then | |
922 | * it is sufficient to simply update the task's cpu field. | |
923 | */ | |
924 | if (!p->array && !task_running(rq, p)) { | |
925 | set_task_cpu(p, dest_cpu); | |
926 | return 0; | |
927 | } | |
928 | ||
929 | init_completion(&req->done); | |
1da177e4 LT |
930 | req->task = p; |
931 | req->dest_cpu = dest_cpu; | |
932 | list_add(&req->list, &rq->migration_queue); | |
933 | return 1; | |
934 | } | |
935 | ||
936 | /* | |
937 | * wait_task_inactive - wait for a thread to unschedule. | |
938 | * | |
939 | * The caller must ensure that the task *will* unschedule sometime soon, | |
940 | * else this function might spin for a *long* time. This function can't | |
941 | * be called with interrupts off, or it may introduce deadlock with | |
942 | * smp_call_function() if an IPI is sent by the same process we are | |
943 | * waiting to become inactive. | |
944 | */ | |
95cdf3b7 | 945 | void wait_task_inactive(task_t *p) |
1da177e4 LT |
946 | { |
947 | unsigned long flags; | |
948 | runqueue_t *rq; | |
949 | int preempted; | |
950 | ||
951 | repeat: | |
952 | rq = task_rq_lock(p, &flags); | |
953 | /* Must be off runqueue entirely, not preempted. */ | |
954 | if (unlikely(p->array || task_running(rq, p))) { | |
955 | /* If it's preempted, we yield. It could be a while. */ | |
956 | preempted = !task_running(rq, p); | |
957 | task_rq_unlock(rq, &flags); | |
958 | cpu_relax(); | |
959 | if (preempted) | |
960 | yield(); | |
961 | goto repeat; | |
962 | } | |
963 | task_rq_unlock(rq, &flags); | |
964 | } | |
965 | ||
966 | /*** | |
967 | * kick_process - kick a running thread to enter/exit the kernel | |
968 | * @p: the to-be-kicked thread | |
969 | * | |
970 | * Cause a process which is running on another CPU to enter | |
971 | * kernel-mode, without any delay. (to get signals handled.) | |
972 | * | |
973 | * NOTE: this function doesnt have to take the runqueue lock, | |
974 | * because all it wants to ensure is that the remote task enters | |
975 | * the kernel. If the IPI races and the task has been migrated | |
976 | * to another CPU then no harm is done and the purpose has been | |
977 | * achieved as well. | |
978 | */ | |
979 | void kick_process(task_t *p) | |
980 | { | |
981 | int cpu; | |
982 | ||
983 | preempt_disable(); | |
984 | cpu = task_cpu(p); | |
985 | if ((cpu != smp_processor_id()) && task_curr(p)) | |
986 | smp_send_reschedule(cpu); | |
987 | preempt_enable(); | |
988 | } | |
989 | ||
990 | /* | |
991 | * Return a low guess at the load of a migration-source cpu. | |
992 | * | |
993 | * We want to under-estimate the load of migration sources, to | |
994 | * balance conservatively. | |
995 | */ | |
b910472d | 996 | static inline unsigned long __source_load(int cpu, int type, enum idle_type idle) |
1da177e4 LT |
997 | { |
998 | runqueue_t *rq = cpu_rq(cpu); | |
6dd4a85b | 999 | unsigned long running = rq->nr_running; |
3b0bd9bc | 1000 | unsigned long source_load, cpu_load = rq->cpu_load[type-1], |
6dd4a85b | 1001 | load_now = running * SCHED_LOAD_SCALE; |
b910472d | 1002 | |
3b0bd9bc CK |
1003 | if (type == 0) |
1004 | source_load = load_now; | |
1005 | else | |
1006 | source_load = min(cpu_load, load_now); | |
1007 | ||
6dd4a85b | 1008 | if (running > 1 || (idle == NOT_IDLE && running)) |
b910472d | 1009 | /* |
3b0bd9bc CK |
1010 | * If we are busy rebalancing the load is biased by |
1011 | * priority to create 'nice' support across cpus. When | |
1012 | * idle rebalancing we should only bias the source_load if | |
1013 | * there is more than one task running on that queue to | |
1014 | * prevent idle rebalance from trying to pull tasks from a | |
1015 | * queue with only one running task. | |
b910472d | 1016 | */ |
6dd4a85b | 1017 | source_load = source_load * rq->prio_bias / running; |
b910472d | 1018 | |
3b0bd9bc | 1019 | return source_load; |
b910472d CK |
1020 | } |
1021 | ||
1022 | static inline unsigned long source_load(int cpu, int type) | |
1023 | { | |
1024 | return __source_load(cpu, type, NOT_IDLE); | |
1da177e4 LT |
1025 | } |
1026 | ||
1027 | /* | |
1028 | * Return a high guess at the load of a migration-target cpu | |
1029 | */ | |
b910472d | 1030 | static inline unsigned long __target_load(int cpu, int type, enum idle_type idle) |
1da177e4 LT |
1031 | { |
1032 | runqueue_t *rq = cpu_rq(cpu); | |
6dd4a85b | 1033 | unsigned long running = rq->nr_running; |
3b0bd9bc | 1034 | unsigned long target_load, cpu_load = rq->cpu_load[type-1], |
6dd4a85b | 1035 | load_now = running * SCHED_LOAD_SCALE; |
b910472d | 1036 | |
7897986b | 1037 | if (type == 0) |
3b0bd9bc CK |
1038 | target_load = load_now; |
1039 | else | |
1040 | target_load = max(cpu_load, load_now); | |
1da177e4 | 1041 | |
6dd4a85b CK |
1042 | if (running > 1 || (idle == NOT_IDLE && running)) |
1043 | target_load = target_load * rq->prio_bias / running; | |
3b0bd9bc CK |
1044 | |
1045 | return target_load; | |
b910472d CK |
1046 | } |
1047 | ||
1048 | static inline unsigned long target_load(int cpu, int type) | |
1049 | { | |
1050 | return __target_load(cpu, type, NOT_IDLE); | |
1da177e4 LT |
1051 | } |
1052 | ||
147cbb4b NP |
1053 | /* |
1054 | * find_idlest_group finds and returns the least busy CPU group within the | |
1055 | * domain. | |
1056 | */ | |
1057 | static struct sched_group * | |
1058 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu) | |
1059 | { | |
1060 | struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups; | |
1061 | unsigned long min_load = ULONG_MAX, this_load = 0; | |
1062 | int load_idx = sd->forkexec_idx; | |
1063 | int imbalance = 100 + (sd->imbalance_pct-100)/2; | |
1064 | ||
1065 | do { | |
1066 | unsigned long load, avg_load; | |
1067 | int local_group; | |
1068 | int i; | |
1069 | ||
da5a5522 BD |
1070 | /* Skip over this group if it has no CPUs allowed */ |
1071 | if (!cpus_intersects(group->cpumask, p->cpus_allowed)) | |
1072 | goto nextgroup; | |
1073 | ||
147cbb4b | 1074 | local_group = cpu_isset(this_cpu, group->cpumask); |
147cbb4b NP |
1075 | |
1076 | /* Tally up the load of all CPUs in the group */ | |
1077 | avg_load = 0; | |
1078 | ||
1079 | for_each_cpu_mask(i, group->cpumask) { | |
1080 | /* Bias balancing toward cpus of our domain */ | |
1081 | if (local_group) | |
1082 | load = source_load(i, load_idx); | |
1083 | else | |
1084 | load = target_load(i, load_idx); | |
1085 | ||
1086 | avg_load += load; | |
1087 | } | |
1088 | ||
1089 | /* Adjust by relative CPU power of the group */ | |
1090 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
1091 | ||
1092 | if (local_group) { | |
1093 | this_load = avg_load; | |
1094 | this = group; | |
1095 | } else if (avg_load < min_load) { | |
1096 | min_load = avg_load; | |
1097 | idlest = group; | |
1098 | } | |
da5a5522 | 1099 | nextgroup: |
147cbb4b NP |
1100 | group = group->next; |
1101 | } while (group != sd->groups); | |
1102 | ||
1103 | if (!idlest || 100*this_load < imbalance*min_load) | |
1104 | return NULL; | |
1105 | return idlest; | |
1106 | } | |
1107 | ||
1108 | /* | |
1109 | * find_idlest_queue - find the idlest runqueue among the cpus in group. | |
1110 | */ | |
95cdf3b7 IM |
1111 | static int |
1112 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
147cbb4b | 1113 | { |
da5a5522 | 1114 | cpumask_t tmp; |
147cbb4b NP |
1115 | unsigned long load, min_load = ULONG_MAX; |
1116 | int idlest = -1; | |
1117 | int i; | |
1118 | ||
da5a5522 BD |
1119 | /* Traverse only the allowed CPUs */ |
1120 | cpus_and(tmp, group->cpumask, p->cpus_allowed); | |
1121 | ||
1122 | for_each_cpu_mask(i, tmp) { | |
147cbb4b NP |
1123 | load = source_load(i, 0); |
1124 | ||
1125 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
1126 | min_load = load; | |
1127 | idlest = i; | |
1128 | } | |
1129 | } | |
1130 | ||
1131 | return idlest; | |
1132 | } | |
1133 | ||
476d139c NP |
1134 | /* |
1135 | * sched_balance_self: balance the current task (running on cpu) in domains | |
1136 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
1137 | * SD_BALANCE_EXEC. | |
1138 | * | |
1139 | * Balance, ie. select the least loaded group. | |
1140 | * | |
1141 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
1142 | * | |
1143 | * preempt must be disabled. | |
1144 | */ | |
1145 | static int sched_balance_self(int cpu, int flag) | |
1146 | { | |
1147 | struct task_struct *t = current; | |
1148 | struct sched_domain *tmp, *sd = NULL; | |
147cbb4b | 1149 | |
476d139c NP |
1150 | for_each_domain(cpu, tmp) |
1151 | if (tmp->flags & flag) | |
1152 | sd = tmp; | |
1153 | ||
1154 | while (sd) { | |
1155 | cpumask_t span; | |
1156 | struct sched_group *group; | |
1157 | int new_cpu; | |
1158 | int weight; | |
1159 | ||
1160 | span = sd->span; | |
1161 | group = find_idlest_group(sd, t, cpu); | |
1162 | if (!group) | |
1163 | goto nextlevel; | |
1164 | ||
da5a5522 | 1165 | new_cpu = find_idlest_cpu(group, t, cpu); |
476d139c NP |
1166 | if (new_cpu == -1 || new_cpu == cpu) |
1167 | goto nextlevel; | |
1168 | ||
1169 | /* Now try balancing at a lower domain level */ | |
1170 | cpu = new_cpu; | |
1171 | nextlevel: | |
1172 | sd = NULL; | |
1173 | weight = cpus_weight(span); | |
1174 | for_each_domain(cpu, tmp) { | |
1175 | if (weight <= cpus_weight(tmp->span)) | |
1176 | break; | |
1177 | if (tmp->flags & flag) | |
1178 | sd = tmp; | |
1179 | } | |
1180 | /* while loop will break here if sd == NULL */ | |
1181 | } | |
1182 | ||
1183 | return cpu; | |
1184 | } | |
1185 | ||
1186 | #endif /* CONFIG_SMP */ | |
1da177e4 LT |
1187 | |
1188 | /* | |
1189 | * wake_idle() will wake a task on an idle cpu if task->cpu is | |
1190 | * not idle and an idle cpu is available. The span of cpus to | |
1191 | * search starts with cpus closest then further out as needed, | |
1192 | * so we always favor a closer, idle cpu. | |
1193 | * | |
1194 | * Returns the CPU we should wake onto. | |
1195 | */ | |
1196 | #if defined(ARCH_HAS_SCHED_WAKE_IDLE) | |
1197 | static int wake_idle(int cpu, task_t *p) | |
1198 | { | |
1199 | cpumask_t tmp; | |
1200 | struct sched_domain *sd; | |
1201 | int i; | |
1202 | ||
1203 | if (idle_cpu(cpu)) | |
1204 | return cpu; | |
1205 | ||
1206 | for_each_domain(cpu, sd) { | |
1207 | if (sd->flags & SD_WAKE_IDLE) { | |
e0f364f4 | 1208 | cpus_and(tmp, sd->span, p->cpus_allowed); |
1da177e4 LT |
1209 | for_each_cpu_mask(i, tmp) { |
1210 | if (idle_cpu(i)) | |
1211 | return i; | |
1212 | } | |
1213 | } | |
e0f364f4 NP |
1214 | else |
1215 | break; | |
1da177e4 LT |
1216 | } |
1217 | return cpu; | |
1218 | } | |
1219 | #else | |
1220 | static inline int wake_idle(int cpu, task_t *p) | |
1221 | { | |
1222 | return cpu; | |
1223 | } | |
1224 | #endif | |
1225 | ||
1226 | /*** | |
1227 | * try_to_wake_up - wake up a thread | |
1228 | * @p: the to-be-woken-up thread | |
1229 | * @state: the mask of task states that can be woken | |
1230 | * @sync: do a synchronous wakeup? | |
1231 | * | |
1232 | * Put it on the run-queue if it's not already there. The "current" | |
1233 | * thread is always on the run-queue (except when the actual | |
1234 | * re-schedule is in progress), and as such you're allowed to do | |
1235 | * the simpler "current->state = TASK_RUNNING" to mark yourself | |
1236 | * runnable without the overhead of this. | |
1237 | * | |
1238 | * returns failure only if the task is already active. | |
1239 | */ | |
95cdf3b7 | 1240 | static int try_to_wake_up(task_t *p, unsigned int state, int sync) |
1da177e4 LT |
1241 | { |
1242 | int cpu, this_cpu, success = 0; | |
1243 | unsigned long flags; | |
1244 | long old_state; | |
1245 | runqueue_t *rq; | |
1246 | #ifdef CONFIG_SMP | |
1247 | unsigned long load, this_load; | |
7897986b | 1248 | struct sched_domain *sd, *this_sd = NULL; |
1da177e4 LT |
1249 | int new_cpu; |
1250 | #endif | |
1251 | ||
1252 | rq = task_rq_lock(p, &flags); | |
1253 | old_state = p->state; | |
1254 | if (!(old_state & state)) | |
1255 | goto out; | |
1256 | ||
1257 | if (p->array) | |
1258 | goto out_running; | |
1259 | ||
1260 | cpu = task_cpu(p); | |
1261 | this_cpu = smp_processor_id(); | |
1262 | ||
1263 | #ifdef CONFIG_SMP | |
1264 | if (unlikely(task_running(rq, p))) | |
1265 | goto out_activate; | |
1266 | ||
7897986b NP |
1267 | new_cpu = cpu; |
1268 | ||
1da177e4 LT |
1269 | schedstat_inc(rq, ttwu_cnt); |
1270 | if (cpu == this_cpu) { | |
1271 | schedstat_inc(rq, ttwu_local); | |
7897986b NP |
1272 | goto out_set_cpu; |
1273 | } | |
1274 | ||
1275 | for_each_domain(this_cpu, sd) { | |
1276 | if (cpu_isset(cpu, sd->span)) { | |
1277 | schedstat_inc(sd, ttwu_wake_remote); | |
1278 | this_sd = sd; | |
1279 | break; | |
1da177e4 LT |
1280 | } |
1281 | } | |
1da177e4 | 1282 | |
7897986b | 1283 | if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed))) |
1da177e4 LT |
1284 | goto out_set_cpu; |
1285 | ||
1da177e4 | 1286 | /* |
7897986b | 1287 | * Check for affine wakeup and passive balancing possibilities. |
1da177e4 | 1288 | */ |
7897986b NP |
1289 | if (this_sd) { |
1290 | int idx = this_sd->wake_idx; | |
1291 | unsigned int imbalance; | |
1da177e4 | 1292 | |
a3f21bce NP |
1293 | imbalance = 100 + (this_sd->imbalance_pct - 100) / 2; |
1294 | ||
7897986b NP |
1295 | load = source_load(cpu, idx); |
1296 | this_load = target_load(this_cpu, idx); | |
1da177e4 | 1297 | |
7897986b NP |
1298 | new_cpu = this_cpu; /* Wake to this CPU if we can */ |
1299 | ||
a3f21bce NP |
1300 | if (this_sd->flags & SD_WAKE_AFFINE) { |
1301 | unsigned long tl = this_load; | |
1da177e4 | 1302 | /* |
a3f21bce NP |
1303 | * If sync wakeup then subtract the (maximum possible) |
1304 | * effect of the currently running task from the load | |
1305 | * of the current CPU: | |
1da177e4 | 1306 | */ |
a3f21bce NP |
1307 | if (sync) |
1308 | tl -= SCHED_LOAD_SCALE; | |
1309 | ||
1310 | if ((tl <= load && | |
1311 | tl + target_load(cpu, idx) <= SCHED_LOAD_SCALE) || | |
1312 | 100*(tl + SCHED_LOAD_SCALE) <= imbalance*load) { | |
1313 | /* | |
1314 | * This domain has SD_WAKE_AFFINE and | |
1315 | * p is cache cold in this domain, and | |
1316 | * there is no bad imbalance. | |
1317 | */ | |
1318 | schedstat_inc(this_sd, ttwu_move_affine); | |
1319 | goto out_set_cpu; | |
1320 | } | |
1321 | } | |
1322 | ||
1323 | /* | |
1324 | * Start passive balancing when half the imbalance_pct | |
1325 | * limit is reached. | |
1326 | */ | |
1327 | if (this_sd->flags & SD_WAKE_BALANCE) { | |
1328 | if (imbalance*this_load <= 100*load) { | |
1329 | schedstat_inc(this_sd, ttwu_move_balance); | |
1330 | goto out_set_cpu; | |
1331 | } | |
1da177e4 LT |
1332 | } |
1333 | } | |
1334 | ||
1335 | new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */ | |
1336 | out_set_cpu: | |
1337 | new_cpu = wake_idle(new_cpu, p); | |
1338 | if (new_cpu != cpu) { | |
1339 | set_task_cpu(p, new_cpu); | |
1340 | task_rq_unlock(rq, &flags); | |
1341 | /* might preempt at this point */ | |
1342 | rq = task_rq_lock(p, &flags); | |
1343 | old_state = p->state; | |
1344 | if (!(old_state & state)) | |
1345 | goto out; | |
1346 | if (p->array) | |
1347 | goto out_running; | |
1348 | ||
1349 | this_cpu = smp_processor_id(); | |
1350 | cpu = task_cpu(p); | |
1351 | } | |
1352 | ||
1353 | out_activate: | |
1354 | #endif /* CONFIG_SMP */ | |
1355 | if (old_state == TASK_UNINTERRUPTIBLE) { | |
1356 | rq->nr_uninterruptible--; | |
1357 | /* | |
1358 | * Tasks on involuntary sleep don't earn | |
1359 | * sleep_avg beyond just interactive state. | |
1360 | */ | |
1361 | p->activated = -1; | |
1362 | } | |
1363 | ||
d79fc0fc IM |
1364 | /* |
1365 | * Tasks that have marked their sleep as noninteractive get | |
1366 | * woken up without updating their sleep average. (i.e. their | |
1367 | * sleep is handled in a priority-neutral manner, no priority | |
1368 | * boost and no penalty.) | |
1369 | */ | |
1370 | if (old_state & TASK_NONINTERACTIVE) | |
1371 | __activate_task(p, rq); | |
1372 | else | |
1373 | activate_task(p, rq, cpu == this_cpu); | |
1da177e4 LT |
1374 | /* |
1375 | * Sync wakeups (i.e. those types of wakeups where the waker | |
1376 | * has indicated that it will leave the CPU in short order) | |
1377 | * don't trigger a preemption, if the woken up task will run on | |
1378 | * this cpu. (in this case the 'I will reschedule' promise of | |
1379 | * the waker guarantees that the freshly woken up task is going | |
1380 | * to be considered on this CPU.) | |
1381 | */ | |
1da177e4 LT |
1382 | if (!sync || cpu != this_cpu) { |
1383 | if (TASK_PREEMPTS_CURR(p, rq)) | |
1384 | resched_task(rq->curr); | |
1385 | } | |
1386 | success = 1; | |
1387 | ||
1388 | out_running: | |
1389 | p->state = TASK_RUNNING; | |
1390 | out: | |
1391 | task_rq_unlock(rq, &flags); | |
1392 | ||
1393 | return success; | |
1394 | } | |
1395 | ||
95cdf3b7 | 1396 | int fastcall wake_up_process(task_t *p) |
1da177e4 LT |
1397 | { |
1398 | return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED | | |
1399 | TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0); | |
1400 | } | |
1401 | ||
1402 | EXPORT_SYMBOL(wake_up_process); | |
1403 | ||
1404 | int fastcall wake_up_state(task_t *p, unsigned int state) | |
1405 | { | |
1406 | return try_to_wake_up(p, state, 0); | |
1407 | } | |
1408 | ||
1da177e4 LT |
1409 | /* |
1410 | * Perform scheduler related setup for a newly forked process p. | |
1411 | * p is forked by current. | |
1412 | */ | |
476d139c | 1413 | void fastcall sched_fork(task_t *p, int clone_flags) |
1da177e4 | 1414 | { |
476d139c NP |
1415 | int cpu = get_cpu(); |
1416 | ||
1417 | #ifdef CONFIG_SMP | |
1418 | cpu = sched_balance_self(cpu, SD_BALANCE_FORK); | |
1419 | #endif | |
1420 | set_task_cpu(p, cpu); | |
1421 | ||
1da177e4 LT |
1422 | /* |
1423 | * We mark the process as running here, but have not actually | |
1424 | * inserted it onto the runqueue yet. This guarantees that | |
1425 | * nobody will actually run it, and a signal or other external | |
1426 | * event cannot wake it up and insert it on the runqueue either. | |
1427 | */ | |
1428 | p->state = TASK_RUNNING; | |
1429 | INIT_LIST_HEAD(&p->run_list); | |
1430 | p->array = NULL; | |
1da177e4 LT |
1431 | #ifdef CONFIG_SCHEDSTATS |
1432 | memset(&p->sched_info, 0, sizeof(p->sched_info)); | |
1433 | #endif | |
4866cde0 NP |
1434 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
1435 | p->oncpu = 0; | |
1436 | #endif | |
1da177e4 | 1437 | #ifdef CONFIG_PREEMPT |
4866cde0 | 1438 | /* Want to start with kernel preemption disabled. */ |
1da177e4 LT |
1439 | p->thread_info->preempt_count = 1; |
1440 | #endif | |
1441 | /* | |
1442 | * Share the timeslice between parent and child, thus the | |
1443 | * total amount of pending timeslices in the system doesn't change, | |
1444 | * resulting in more scheduling fairness. | |
1445 | */ | |
1446 | local_irq_disable(); | |
1447 | p->time_slice = (current->time_slice + 1) >> 1; | |
1448 | /* | |
1449 | * The remainder of the first timeslice might be recovered by | |
1450 | * the parent if the child exits early enough. | |
1451 | */ | |
1452 | p->first_time_slice = 1; | |
1453 | current->time_slice >>= 1; | |
1454 | p->timestamp = sched_clock(); | |
1455 | if (unlikely(!current->time_slice)) { | |
1456 | /* | |
1457 | * This case is rare, it happens when the parent has only | |
1458 | * a single jiffy left from its timeslice. Taking the | |
1459 | * runqueue lock is not a problem. | |
1460 | */ | |
1461 | current->time_slice = 1; | |
1da177e4 | 1462 | scheduler_tick(); |
476d139c NP |
1463 | } |
1464 | local_irq_enable(); | |
1465 | put_cpu(); | |
1da177e4 LT |
1466 | } |
1467 | ||
1468 | /* | |
1469 | * wake_up_new_task - wake up a newly created task for the first time. | |
1470 | * | |
1471 | * This function will do some initial scheduler statistics housekeeping | |
1472 | * that must be done for every newly created context, then puts the task | |
1473 | * on the runqueue and wakes it. | |
1474 | */ | |
95cdf3b7 | 1475 | void fastcall wake_up_new_task(task_t *p, unsigned long clone_flags) |
1da177e4 LT |
1476 | { |
1477 | unsigned long flags; | |
1478 | int this_cpu, cpu; | |
1479 | runqueue_t *rq, *this_rq; | |
1480 | ||
1481 | rq = task_rq_lock(p, &flags); | |
147cbb4b | 1482 | BUG_ON(p->state != TASK_RUNNING); |
1da177e4 | 1483 | this_cpu = smp_processor_id(); |
147cbb4b | 1484 | cpu = task_cpu(p); |
1da177e4 | 1485 | |
1da177e4 LT |
1486 | /* |
1487 | * We decrease the sleep average of forking parents | |
1488 | * and children as well, to keep max-interactive tasks | |
1489 | * from forking tasks that are max-interactive. The parent | |
1490 | * (current) is done further down, under its lock. | |
1491 | */ | |
1492 | p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) * | |
1493 | CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS); | |
1494 | ||
1495 | p->prio = effective_prio(p); | |
1496 | ||
1497 | if (likely(cpu == this_cpu)) { | |
1498 | if (!(clone_flags & CLONE_VM)) { | |
1499 | /* | |
1500 | * The VM isn't cloned, so we're in a good position to | |
1501 | * do child-runs-first in anticipation of an exec. This | |
1502 | * usually avoids a lot of COW overhead. | |
1503 | */ | |
1504 | if (unlikely(!current->array)) | |
1505 | __activate_task(p, rq); | |
1506 | else { | |
1507 | p->prio = current->prio; | |
1508 | list_add_tail(&p->run_list, ¤t->run_list); | |
1509 | p->array = current->array; | |
1510 | p->array->nr_active++; | |
b910472d | 1511 | inc_nr_running(p, rq); |
1da177e4 LT |
1512 | } |
1513 | set_need_resched(); | |
1514 | } else | |
1515 | /* Run child last */ | |
1516 | __activate_task(p, rq); | |
1517 | /* | |
1518 | * We skip the following code due to cpu == this_cpu | |
1519 | * | |
1520 | * task_rq_unlock(rq, &flags); | |
1521 | * this_rq = task_rq_lock(current, &flags); | |
1522 | */ | |
1523 | this_rq = rq; | |
1524 | } else { | |
1525 | this_rq = cpu_rq(this_cpu); | |
1526 | ||
1527 | /* | |
1528 | * Not the local CPU - must adjust timestamp. This should | |
1529 | * get optimised away in the !CONFIG_SMP case. | |
1530 | */ | |
1531 | p->timestamp = (p->timestamp - this_rq->timestamp_last_tick) | |
1532 | + rq->timestamp_last_tick; | |
1533 | __activate_task(p, rq); | |
1534 | if (TASK_PREEMPTS_CURR(p, rq)) | |
1535 | resched_task(rq->curr); | |
1536 | ||
1537 | /* | |
1538 | * Parent and child are on different CPUs, now get the | |
1539 | * parent runqueue to update the parent's ->sleep_avg: | |
1540 | */ | |
1541 | task_rq_unlock(rq, &flags); | |
1542 | this_rq = task_rq_lock(current, &flags); | |
1543 | } | |
1544 | current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) * | |
1545 | PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS); | |
1546 | task_rq_unlock(this_rq, &flags); | |
1547 | } | |
1548 | ||
1549 | /* | |
1550 | * Potentially available exiting-child timeslices are | |
1551 | * retrieved here - this way the parent does not get | |
1552 | * penalized for creating too many threads. | |
1553 | * | |
1554 | * (this cannot be used to 'generate' timeslices | |
1555 | * artificially, because any timeslice recovered here | |
1556 | * was given away by the parent in the first place.) | |
1557 | */ | |
95cdf3b7 | 1558 | void fastcall sched_exit(task_t *p) |
1da177e4 LT |
1559 | { |
1560 | unsigned long flags; | |
1561 | runqueue_t *rq; | |
1562 | ||
1563 | /* | |
1564 | * If the child was a (relative-) CPU hog then decrease | |
1565 | * the sleep_avg of the parent as well. | |
1566 | */ | |
1567 | rq = task_rq_lock(p->parent, &flags); | |
889dfafe | 1568 | if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) { |
1da177e4 LT |
1569 | p->parent->time_slice += p->time_slice; |
1570 | if (unlikely(p->parent->time_slice > task_timeslice(p))) | |
1571 | p->parent->time_slice = task_timeslice(p); | |
1572 | } | |
1573 | if (p->sleep_avg < p->parent->sleep_avg) | |
1574 | p->parent->sleep_avg = p->parent->sleep_avg / | |
1575 | (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg / | |
1576 | (EXIT_WEIGHT + 1); | |
1577 | task_rq_unlock(rq, &flags); | |
1578 | } | |
1579 | ||
4866cde0 NP |
1580 | /** |
1581 | * prepare_task_switch - prepare to switch tasks | |
1582 | * @rq: the runqueue preparing to switch | |
1583 | * @next: the task we are going to switch to. | |
1584 | * | |
1585 | * This is called with the rq lock held and interrupts off. It must | |
1586 | * be paired with a subsequent finish_task_switch after the context | |
1587 | * switch. | |
1588 | * | |
1589 | * prepare_task_switch sets up locking and calls architecture specific | |
1590 | * hooks. | |
1591 | */ | |
1592 | static inline void prepare_task_switch(runqueue_t *rq, task_t *next) | |
1593 | { | |
1594 | prepare_lock_switch(rq, next); | |
1595 | prepare_arch_switch(next); | |
1596 | } | |
1597 | ||
1da177e4 LT |
1598 | /** |
1599 | * finish_task_switch - clean up after a task-switch | |
344babaa | 1600 | * @rq: runqueue associated with task-switch |
1da177e4 LT |
1601 | * @prev: the thread we just switched away from. |
1602 | * | |
4866cde0 NP |
1603 | * finish_task_switch must be called after the context switch, paired |
1604 | * with a prepare_task_switch call before the context switch. | |
1605 | * finish_task_switch will reconcile locking set up by prepare_task_switch, | |
1606 | * and do any other architecture-specific cleanup actions. | |
1da177e4 LT |
1607 | * |
1608 | * Note that we may have delayed dropping an mm in context_switch(). If | |
1609 | * so, we finish that here outside of the runqueue lock. (Doing it | |
1610 | * with the lock held can cause deadlocks; see schedule() for | |
1611 | * details.) | |
1612 | */ | |
4866cde0 | 1613 | static inline void finish_task_switch(runqueue_t *rq, task_t *prev) |
1da177e4 LT |
1614 | __releases(rq->lock) |
1615 | { | |
1da177e4 LT |
1616 | struct mm_struct *mm = rq->prev_mm; |
1617 | unsigned long prev_task_flags; | |
1618 | ||
1619 | rq->prev_mm = NULL; | |
1620 | ||
1621 | /* | |
1622 | * A task struct has one reference for the use as "current". | |
1623 | * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and | |
1624 | * calls schedule one last time. The schedule call will never return, | |
1625 | * and the scheduled task must drop that reference. | |
1626 | * The test for EXIT_ZOMBIE must occur while the runqueue locks are | |
1627 | * still held, otherwise prev could be scheduled on another cpu, die | |
1628 | * there before we look at prev->state, and then the reference would | |
1629 | * be dropped twice. | |
1630 | * Manfred Spraul <manfred@colorfullife.com> | |
1631 | */ | |
1632 | prev_task_flags = prev->flags; | |
4866cde0 NP |
1633 | finish_arch_switch(prev); |
1634 | finish_lock_switch(rq, prev); | |
1da177e4 LT |
1635 | if (mm) |
1636 | mmdrop(mm); | |
1637 | if (unlikely(prev_task_flags & PF_DEAD)) | |
1638 | put_task_struct(prev); | |
1639 | } | |
1640 | ||
1641 | /** | |
1642 | * schedule_tail - first thing a freshly forked thread must call. | |
1643 | * @prev: the thread we just switched away from. | |
1644 | */ | |
1645 | asmlinkage void schedule_tail(task_t *prev) | |
1646 | __releases(rq->lock) | |
1647 | { | |
4866cde0 NP |
1648 | runqueue_t *rq = this_rq(); |
1649 | finish_task_switch(rq, prev); | |
1650 | #ifdef __ARCH_WANT_UNLOCKED_CTXSW | |
1651 | /* In this case, finish_task_switch does not reenable preemption */ | |
1652 | preempt_enable(); | |
1653 | #endif | |
1da177e4 LT |
1654 | if (current->set_child_tid) |
1655 | put_user(current->pid, current->set_child_tid); | |
1656 | } | |
1657 | ||
1658 | /* | |
1659 | * context_switch - switch to the new MM and the new | |
1660 | * thread's register state. | |
1661 | */ | |
1662 | static inline | |
1663 | task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next) | |
1664 | { | |
1665 | struct mm_struct *mm = next->mm; | |
1666 | struct mm_struct *oldmm = prev->active_mm; | |
1667 | ||
1668 | if (unlikely(!mm)) { | |
1669 | next->active_mm = oldmm; | |
1670 | atomic_inc(&oldmm->mm_count); | |
1671 | enter_lazy_tlb(oldmm, next); | |
1672 | } else | |
1673 | switch_mm(oldmm, mm, next); | |
1674 | ||
1675 | if (unlikely(!prev->mm)) { | |
1676 | prev->active_mm = NULL; | |
1677 | WARN_ON(rq->prev_mm); | |
1678 | rq->prev_mm = oldmm; | |
1679 | } | |
1680 | ||
1681 | /* Here we just switch the register state and the stack. */ | |
1682 | switch_to(prev, next, prev); | |
1683 | ||
1684 | return prev; | |
1685 | } | |
1686 | ||
1687 | /* | |
1688 | * nr_running, nr_uninterruptible and nr_context_switches: | |
1689 | * | |
1690 | * externally visible scheduler statistics: current number of runnable | |
1691 | * threads, current number of uninterruptible-sleeping threads, total | |
1692 | * number of context switches performed since bootup. | |
1693 | */ | |
1694 | unsigned long nr_running(void) | |
1695 | { | |
1696 | unsigned long i, sum = 0; | |
1697 | ||
1698 | for_each_online_cpu(i) | |
1699 | sum += cpu_rq(i)->nr_running; | |
1700 | ||
1701 | return sum; | |
1702 | } | |
1703 | ||
1704 | unsigned long nr_uninterruptible(void) | |
1705 | { | |
1706 | unsigned long i, sum = 0; | |
1707 | ||
1708 | for_each_cpu(i) | |
1709 | sum += cpu_rq(i)->nr_uninterruptible; | |
1710 | ||
1711 | /* | |
1712 | * Since we read the counters lockless, it might be slightly | |
1713 | * inaccurate. Do not allow it to go below zero though: | |
1714 | */ | |
1715 | if (unlikely((long)sum < 0)) | |
1716 | sum = 0; | |
1717 | ||
1718 | return sum; | |
1719 | } | |
1720 | ||
1721 | unsigned long long nr_context_switches(void) | |
1722 | { | |
1723 | unsigned long long i, sum = 0; | |
1724 | ||
1725 | for_each_cpu(i) | |
1726 | sum += cpu_rq(i)->nr_switches; | |
1727 | ||
1728 | return sum; | |
1729 | } | |
1730 | ||
1731 | unsigned long nr_iowait(void) | |
1732 | { | |
1733 | unsigned long i, sum = 0; | |
1734 | ||
1735 | for_each_cpu(i) | |
1736 | sum += atomic_read(&cpu_rq(i)->nr_iowait); | |
1737 | ||
1738 | return sum; | |
1739 | } | |
1740 | ||
1741 | #ifdef CONFIG_SMP | |
1742 | ||
1743 | /* | |
1744 | * double_rq_lock - safely lock two runqueues | |
1745 | * | |
1746 | * Note this does not disable interrupts like task_rq_lock, | |
1747 | * you need to do so manually before calling. | |
1748 | */ | |
1749 | static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2) | |
1750 | __acquires(rq1->lock) | |
1751 | __acquires(rq2->lock) | |
1752 | { | |
1753 | if (rq1 == rq2) { | |
1754 | spin_lock(&rq1->lock); | |
1755 | __acquire(rq2->lock); /* Fake it out ;) */ | |
1756 | } else { | |
1757 | if (rq1 < rq2) { | |
1758 | spin_lock(&rq1->lock); | |
1759 | spin_lock(&rq2->lock); | |
1760 | } else { | |
1761 | spin_lock(&rq2->lock); | |
1762 | spin_lock(&rq1->lock); | |
1763 | } | |
1764 | } | |
1765 | } | |
1766 | ||
1767 | /* | |
1768 | * double_rq_unlock - safely unlock two runqueues | |
1769 | * | |
1770 | * Note this does not restore interrupts like task_rq_unlock, | |
1771 | * you need to do so manually after calling. | |
1772 | */ | |
1773 | static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2) | |
1774 | __releases(rq1->lock) | |
1775 | __releases(rq2->lock) | |
1776 | { | |
1777 | spin_unlock(&rq1->lock); | |
1778 | if (rq1 != rq2) | |
1779 | spin_unlock(&rq2->lock); | |
1780 | else | |
1781 | __release(rq2->lock); | |
1782 | } | |
1783 | ||
1784 | /* | |
1785 | * double_lock_balance - lock the busiest runqueue, this_rq is locked already. | |
1786 | */ | |
1787 | static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest) | |
1788 | __releases(this_rq->lock) | |
1789 | __acquires(busiest->lock) | |
1790 | __acquires(this_rq->lock) | |
1791 | { | |
1792 | if (unlikely(!spin_trylock(&busiest->lock))) { | |
1793 | if (busiest < this_rq) { | |
1794 | spin_unlock(&this_rq->lock); | |
1795 | spin_lock(&busiest->lock); | |
1796 | spin_lock(&this_rq->lock); | |
1797 | } else | |
1798 | spin_lock(&busiest->lock); | |
1799 | } | |
1800 | } | |
1801 | ||
1da177e4 LT |
1802 | /* |
1803 | * If dest_cpu is allowed for this process, migrate the task to it. | |
1804 | * This is accomplished by forcing the cpu_allowed mask to only | |
1805 | * allow dest_cpu, which will force the cpu onto dest_cpu. Then | |
1806 | * the cpu_allowed mask is restored. | |
1807 | */ | |
1808 | static void sched_migrate_task(task_t *p, int dest_cpu) | |
1809 | { | |
1810 | migration_req_t req; | |
1811 | runqueue_t *rq; | |
1812 | unsigned long flags; | |
1813 | ||
1814 | rq = task_rq_lock(p, &flags); | |
1815 | if (!cpu_isset(dest_cpu, p->cpus_allowed) | |
1816 | || unlikely(cpu_is_offline(dest_cpu))) | |
1817 | goto out; | |
1818 | ||
1819 | /* force the process onto the specified CPU */ | |
1820 | if (migrate_task(p, dest_cpu, &req)) { | |
1821 | /* Need to wait for migration thread (might exit: take ref). */ | |
1822 | struct task_struct *mt = rq->migration_thread; | |
1823 | get_task_struct(mt); | |
1824 | task_rq_unlock(rq, &flags); | |
1825 | wake_up_process(mt); | |
1826 | put_task_struct(mt); | |
1827 | wait_for_completion(&req.done); | |
1828 | return; | |
1829 | } | |
1830 | out: | |
1831 | task_rq_unlock(rq, &flags); | |
1832 | } | |
1833 | ||
1834 | /* | |
476d139c NP |
1835 | * sched_exec - execve() is a valuable balancing opportunity, because at |
1836 | * this point the task has the smallest effective memory and cache footprint. | |
1da177e4 LT |
1837 | */ |
1838 | void sched_exec(void) | |
1839 | { | |
1da177e4 | 1840 | int new_cpu, this_cpu = get_cpu(); |
476d139c | 1841 | new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC); |
1da177e4 | 1842 | put_cpu(); |
476d139c NP |
1843 | if (new_cpu != this_cpu) |
1844 | sched_migrate_task(current, new_cpu); | |
1da177e4 LT |
1845 | } |
1846 | ||
1847 | /* | |
1848 | * pull_task - move a task from a remote runqueue to the local runqueue. | |
1849 | * Both runqueues must be locked. | |
1850 | */ | |
1851 | static inline | |
1852 | void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p, | |
1853 | runqueue_t *this_rq, prio_array_t *this_array, int this_cpu) | |
1854 | { | |
1855 | dequeue_task(p, src_array); | |
b910472d | 1856 | dec_nr_running(p, src_rq); |
1da177e4 | 1857 | set_task_cpu(p, this_cpu); |
b910472d | 1858 | inc_nr_running(p, this_rq); |
1da177e4 LT |
1859 | enqueue_task(p, this_array); |
1860 | p->timestamp = (p->timestamp - src_rq->timestamp_last_tick) | |
1861 | + this_rq->timestamp_last_tick; | |
1862 | /* | |
1863 | * Note that idle threads have a prio of MAX_PRIO, for this test | |
1864 | * to be always true for them. | |
1865 | */ | |
1866 | if (TASK_PREEMPTS_CURR(p, this_rq)) | |
1867 | resched_task(this_rq->curr); | |
1868 | } | |
1869 | ||
1870 | /* | |
1871 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
1872 | */ | |
1873 | static inline | |
1874 | int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu, | |
95cdf3b7 IM |
1875 | struct sched_domain *sd, enum idle_type idle, |
1876 | int *all_pinned) | |
1da177e4 LT |
1877 | { |
1878 | /* | |
1879 | * We do not migrate tasks that are: | |
1880 | * 1) running (obviously), or | |
1881 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
1882 | * 3) are cache-hot on their current CPU. | |
1883 | */ | |
1da177e4 LT |
1884 | if (!cpu_isset(this_cpu, p->cpus_allowed)) |
1885 | return 0; | |
81026794 NP |
1886 | *all_pinned = 0; |
1887 | ||
1888 | if (task_running(rq, p)) | |
1889 | return 0; | |
1da177e4 LT |
1890 | |
1891 | /* | |
1892 | * Aggressive migration if: | |
cafb20c1 | 1893 | * 1) task is cache cold, or |
1da177e4 LT |
1894 | * 2) too many balance attempts have failed. |
1895 | */ | |
1896 | ||
cafb20c1 | 1897 | if (sd->nr_balance_failed > sd->cache_nice_tries) |
1da177e4 LT |
1898 | return 1; |
1899 | ||
1900 | if (task_hot(p, rq->timestamp_last_tick, sd)) | |
81026794 | 1901 | return 0; |
1da177e4 LT |
1902 | return 1; |
1903 | } | |
1904 | ||
1905 | /* | |
1906 | * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq, | |
1907 | * as part of a balancing operation within "domain". Returns the number of | |
1908 | * tasks moved. | |
1909 | * | |
1910 | * Called with both runqueues locked. | |
1911 | */ | |
1912 | static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest, | |
1913 | unsigned long max_nr_move, struct sched_domain *sd, | |
81026794 | 1914 | enum idle_type idle, int *all_pinned) |
1da177e4 LT |
1915 | { |
1916 | prio_array_t *array, *dst_array; | |
1917 | struct list_head *head, *curr; | |
81026794 | 1918 | int idx, pulled = 0, pinned = 0; |
1da177e4 LT |
1919 | task_t *tmp; |
1920 | ||
81026794 | 1921 | if (max_nr_move == 0) |
1da177e4 LT |
1922 | goto out; |
1923 | ||
81026794 NP |
1924 | pinned = 1; |
1925 | ||
1da177e4 LT |
1926 | /* |
1927 | * We first consider expired tasks. Those will likely not be | |
1928 | * executed in the near future, and they are most likely to | |
1929 | * be cache-cold, thus switching CPUs has the least effect | |
1930 | * on them. | |
1931 | */ | |
1932 | if (busiest->expired->nr_active) { | |
1933 | array = busiest->expired; | |
1934 | dst_array = this_rq->expired; | |
1935 | } else { | |
1936 | array = busiest->active; | |
1937 | dst_array = this_rq->active; | |
1938 | } | |
1939 | ||
1940 | new_array: | |
1941 | /* Start searching at priority 0: */ | |
1942 | idx = 0; | |
1943 | skip_bitmap: | |
1944 | if (!idx) | |
1945 | idx = sched_find_first_bit(array->bitmap); | |
1946 | else | |
1947 | idx = find_next_bit(array->bitmap, MAX_PRIO, idx); | |
1948 | if (idx >= MAX_PRIO) { | |
1949 | if (array == busiest->expired && busiest->active->nr_active) { | |
1950 | array = busiest->active; | |
1951 | dst_array = this_rq->active; | |
1952 | goto new_array; | |
1953 | } | |
1954 | goto out; | |
1955 | } | |
1956 | ||
1957 | head = array->queue + idx; | |
1958 | curr = head->prev; | |
1959 | skip_queue: | |
1960 | tmp = list_entry(curr, task_t, run_list); | |
1961 | ||
1962 | curr = curr->prev; | |
1963 | ||
81026794 | 1964 | if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) { |
1da177e4 LT |
1965 | if (curr != head) |
1966 | goto skip_queue; | |
1967 | idx++; | |
1968 | goto skip_bitmap; | |
1969 | } | |
1970 | ||
1971 | #ifdef CONFIG_SCHEDSTATS | |
1972 | if (task_hot(tmp, busiest->timestamp_last_tick, sd)) | |
1973 | schedstat_inc(sd, lb_hot_gained[idle]); | |
1974 | #endif | |
1975 | ||
1976 | pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu); | |
1977 | pulled++; | |
1978 | ||
1979 | /* We only want to steal up to the prescribed number of tasks. */ | |
1980 | if (pulled < max_nr_move) { | |
1981 | if (curr != head) | |
1982 | goto skip_queue; | |
1983 | idx++; | |
1984 | goto skip_bitmap; | |
1985 | } | |
1986 | out: | |
1987 | /* | |
1988 | * Right now, this is the only place pull_task() is called, | |
1989 | * so we can safely collect pull_task() stats here rather than | |
1990 | * inside pull_task(). | |
1991 | */ | |
1992 | schedstat_add(sd, lb_gained[idle], pulled); | |
81026794 NP |
1993 | |
1994 | if (all_pinned) | |
1995 | *all_pinned = pinned; | |
1da177e4 LT |
1996 | return pulled; |
1997 | } | |
1998 | ||
1999 | /* | |
2000 | * find_busiest_group finds and returns the busiest CPU group within the | |
2001 | * domain. It calculates and returns the number of tasks which should be | |
2002 | * moved to restore balance via the imbalance parameter. | |
2003 | */ | |
2004 | static struct sched_group * | |
2005 | find_busiest_group(struct sched_domain *sd, int this_cpu, | |
5969fe06 | 2006 | unsigned long *imbalance, enum idle_type idle, int *sd_idle) |
1da177e4 LT |
2007 | { |
2008 | struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups; | |
2009 | unsigned long max_load, avg_load, total_load, this_load, total_pwr; | |
0c117f1b | 2010 | unsigned long max_pull; |
7897986b | 2011 | int load_idx; |
1da177e4 LT |
2012 | |
2013 | max_load = this_load = total_load = total_pwr = 0; | |
7897986b NP |
2014 | if (idle == NOT_IDLE) |
2015 | load_idx = sd->busy_idx; | |
2016 | else if (idle == NEWLY_IDLE) | |
2017 | load_idx = sd->newidle_idx; | |
2018 | else | |
2019 | load_idx = sd->idle_idx; | |
1da177e4 LT |
2020 | |
2021 | do { | |
2022 | unsigned long load; | |
2023 | int local_group; | |
2024 | int i; | |
2025 | ||
2026 | local_group = cpu_isset(this_cpu, group->cpumask); | |
2027 | ||
2028 | /* Tally up the load of all CPUs in the group */ | |
2029 | avg_load = 0; | |
2030 | ||
2031 | for_each_cpu_mask(i, group->cpumask) { | |
5969fe06 NP |
2032 | if (*sd_idle && !idle_cpu(i)) |
2033 | *sd_idle = 0; | |
2034 | ||
1da177e4 LT |
2035 | /* Bias balancing toward cpus of our domain */ |
2036 | if (local_group) | |
b910472d | 2037 | load = __target_load(i, load_idx, idle); |
1da177e4 | 2038 | else |
b910472d | 2039 | load = __source_load(i, load_idx, idle); |
1da177e4 LT |
2040 | |
2041 | avg_load += load; | |
2042 | } | |
2043 | ||
2044 | total_load += avg_load; | |
2045 | total_pwr += group->cpu_power; | |
2046 | ||
2047 | /* Adjust by relative CPU power of the group */ | |
2048 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
2049 | ||
2050 | if (local_group) { | |
2051 | this_load = avg_load; | |
2052 | this = group; | |
1da177e4 LT |
2053 | } else if (avg_load > max_load) { |
2054 | max_load = avg_load; | |
2055 | busiest = group; | |
2056 | } | |
1da177e4 LT |
2057 | group = group->next; |
2058 | } while (group != sd->groups); | |
2059 | ||
0c117f1b | 2060 | if (!busiest || this_load >= max_load || max_load <= SCHED_LOAD_SCALE) |
1da177e4 LT |
2061 | goto out_balanced; |
2062 | ||
2063 | avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr; | |
2064 | ||
2065 | if (this_load >= avg_load || | |
2066 | 100*max_load <= sd->imbalance_pct*this_load) | |
2067 | goto out_balanced; | |
2068 | ||
2069 | /* | |
2070 | * We're trying to get all the cpus to the average_load, so we don't | |
2071 | * want to push ourselves above the average load, nor do we wish to | |
2072 | * reduce the max loaded cpu below the average load, as either of these | |
2073 | * actions would just result in more rebalancing later, and ping-pong | |
2074 | * tasks around. Thus we look for the minimum possible imbalance. | |
2075 | * Negative imbalances (*we* are more loaded than anyone else) will | |
2076 | * be counted as no imbalance for these purposes -- we can't fix that | |
2077 | * by pulling tasks to us. Be careful of negative numbers as they'll | |
2078 | * appear as very large values with unsigned longs. | |
2079 | */ | |
0c117f1b SS |
2080 | |
2081 | /* Don't want to pull so many tasks that a group would go idle */ | |
2082 | max_pull = min(max_load - avg_load, max_load - SCHED_LOAD_SCALE); | |
2083 | ||
1da177e4 | 2084 | /* How much load to actually move to equalise the imbalance */ |
0c117f1b | 2085 | *imbalance = min(max_pull * busiest->cpu_power, |
1da177e4 LT |
2086 | (avg_load - this_load) * this->cpu_power) |
2087 | / SCHED_LOAD_SCALE; | |
2088 | ||
2089 | if (*imbalance < SCHED_LOAD_SCALE) { | |
2090 | unsigned long pwr_now = 0, pwr_move = 0; | |
2091 | unsigned long tmp; | |
2092 | ||
2093 | if (max_load - this_load >= SCHED_LOAD_SCALE*2) { | |
2094 | *imbalance = 1; | |
2095 | return busiest; | |
2096 | } | |
2097 | ||
2098 | /* | |
2099 | * OK, we don't have enough imbalance to justify moving tasks, | |
2100 | * however we may be able to increase total CPU power used by | |
2101 | * moving them. | |
2102 | */ | |
2103 | ||
2104 | pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load); | |
2105 | pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load); | |
2106 | pwr_now /= SCHED_LOAD_SCALE; | |
2107 | ||
2108 | /* Amount of load we'd subtract */ | |
2109 | tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power; | |
2110 | if (max_load > tmp) | |
2111 | pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE, | |
2112 | max_load - tmp); | |
2113 | ||
2114 | /* Amount of load we'd add */ | |
2115 | if (max_load*busiest->cpu_power < | |
2116 | SCHED_LOAD_SCALE*SCHED_LOAD_SCALE) | |
2117 | tmp = max_load*busiest->cpu_power/this->cpu_power; | |
2118 | else | |
2119 | tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power; | |
2120 | pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp); | |
2121 | pwr_move /= SCHED_LOAD_SCALE; | |
2122 | ||
2123 | /* Move if we gain throughput */ | |
2124 | if (pwr_move <= pwr_now) | |
2125 | goto out_balanced; | |
2126 | ||
2127 | *imbalance = 1; | |
2128 | return busiest; | |
2129 | } | |
2130 | ||
2131 | /* Get rid of the scaling factor, rounding down as we divide */ | |
2132 | *imbalance = *imbalance / SCHED_LOAD_SCALE; | |
1da177e4 LT |
2133 | return busiest; |
2134 | ||
2135 | out_balanced: | |
1da177e4 LT |
2136 | |
2137 | *imbalance = 0; | |
2138 | return NULL; | |
2139 | } | |
2140 | ||
2141 | /* | |
2142 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
2143 | */ | |
b910472d CK |
2144 | static runqueue_t *find_busiest_queue(struct sched_group *group, |
2145 | enum idle_type idle) | |
1da177e4 LT |
2146 | { |
2147 | unsigned long load, max_load = 0; | |
2148 | runqueue_t *busiest = NULL; | |
2149 | int i; | |
2150 | ||
2151 | for_each_cpu_mask(i, group->cpumask) { | |
b910472d | 2152 | load = __source_load(i, 0, idle); |
1da177e4 LT |
2153 | |
2154 | if (load > max_load) { | |
2155 | max_load = load; | |
2156 | busiest = cpu_rq(i); | |
2157 | } | |
2158 | } | |
2159 | ||
2160 | return busiest; | |
2161 | } | |
2162 | ||
77391d71 NP |
2163 | /* |
2164 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
2165 | * so long as it is large enough. | |
2166 | */ | |
2167 | #define MAX_PINNED_INTERVAL 512 | |
2168 | ||
1da177e4 LT |
2169 | /* |
2170 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
2171 | * tasks if there is an imbalance. | |
2172 | * | |
2173 | * Called with this_rq unlocked. | |
2174 | */ | |
2175 | static int load_balance(int this_cpu, runqueue_t *this_rq, | |
2176 | struct sched_domain *sd, enum idle_type idle) | |
2177 | { | |
2178 | struct sched_group *group; | |
2179 | runqueue_t *busiest; | |
2180 | unsigned long imbalance; | |
77391d71 | 2181 | int nr_moved, all_pinned = 0; |
81026794 | 2182 | int active_balance = 0; |
5969fe06 NP |
2183 | int sd_idle = 0; |
2184 | ||
2185 | if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER) | |
2186 | sd_idle = 1; | |
1da177e4 | 2187 | |
1da177e4 LT |
2188 | schedstat_inc(sd, lb_cnt[idle]); |
2189 | ||
5969fe06 | 2190 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle); |
1da177e4 LT |
2191 | if (!group) { |
2192 | schedstat_inc(sd, lb_nobusyg[idle]); | |
2193 | goto out_balanced; | |
2194 | } | |
2195 | ||
b910472d | 2196 | busiest = find_busiest_queue(group, idle); |
1da177e4 LT |
2197 | if (!busiest) { |
2198 | schedstat_inc(sd, lb_nobusyq[idle]); | |
2199 | goto out_balanced; | |
2200 | } | |
2201 | ||
db935dbd | 2202 | BUG_ON(busiest == this_rq); |
1da177e4 LT |
2203 | |
2204 | schedstat_add(sd, lb_imbalance[idle], imbalance); | |
2205 | ||
2206 | nr_moved = 0; | |
2207 | if (busiest->nr_running > 1) { | |
2208 | /* | |
2209 | * Attempt to move tasks. If find_busiest_group has found | |
2210 | * an imbalance but busiest->nr_running <= 1, the group is | |
2211 | * still unbalanced. nr_moved simply stays zero, so it is | |
2212 | * correctly treated as an imbalance. | |
2213 | */ | |
e17224bf | 2214 | double_rq_lock(this_rq, busiest); |
1da177e4 | 2215 | nr_moved = move_tasks(this_rq, this_cpu, busiest, |
d6d5cfaf | 2216 | imbalance, sd, idle, &all_pinned); |
e17224bf | 2217 | double_rq_unlock(this_rq, busiest); |
81026794 NP |
2218 | |
2219 | /* All tasks on this runqueue were pinned by CPU affinity */ | |
2220 | if (unlikely(all_pinned)) | |
2221 | goto out_balanced; | |
1da177e4 | 2222 | } |
81026794 | 2223 | |
1da177e4 LT |
2224 | if (!nr_moved) { |
2225 | schedstat_inc(sd, lb_failed[idle]); | |
2226 | sd->nr_balance_failed++; | |
2227 | ||
2228 | if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) { | |
1da177e4 LT |
2229 | |
2230 | spin_lock(&busiest->lock); | |
fa3b6ddc SS |
2231 | |
2232 | /* don't kick the migration_thread, if the curr | |
2233 | * task on busiest cpu can't be moved to this_cpu | |
2234 | */ | |
2235 | if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) { | |
2236 | spin_unlock(&busiest->lock); | |
2237 | all_pinned = 1; | |
2238 | goto out_one_pinned; | |
2239 | } | |
2240 | ||
1da177e4 LT |
2241 | if (!busiest->active_balance) { |
2242 | busiest->active_balance = 1; | |
2243 | busiest->push_cpu = this_cpu; | |
81026794 | 2244 | active_balance = 1; |
1da177e4 LT |
2245 | } |
2246 | spin_unlock(&busiest->lock); | |
81026794 | 2247 | if (active_balance) |
1da177e4 LT |
2248 | wake_up_process(busiest->migration_thread); |
2249 | ||
2250 | /* | |
2251 | * We've kicked active balancing, reset the failure | |
2252 | * counter. | |
2253 | */ | |
39507451 | 2254 | sd->nr_balance_failed = sd->cache_nice_tries+1; |
1da177e4 | 2255 | } |
81026794 | 2256 | } else |
1da177e4 LT |
2257 | sd->nr_balance_failed = 0; |
2258 | ||
81026794 | 2259 | if (likely(!active_balance)) { |
1da177e4 LT |
2260 | /* We were unbalanced, so reset the balancing interval */ |
2261 | sd->balance_interval = sd->min_interval; | |
81026794 NP |
2262 | } else { |
2263 | /* | |
2264 | * If we've begun active balancing, start to back off. This | |
2265 | * case may not be covered by the all_pinned logic if there | |
2266 | * is only 1 task on the busy runqueue (because we don't call | |
2267 | * move_tasks). | |
2268 | */ | |
2269 | if (sd->balance_interval < sd->max_interval) | |
2270 | sd->balance_interval *= 2; | |
1da177e4 LT |
2271 | } |
2272 | ||
5969fe06 NP |
2273 | if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
2274 | return -1; | |
1da177e4 LT |
2275 | return nr_moved; |
2276 | ||
2277 | out_balanced: | |
1da177e4 LT |
2278 | schedstat_inc(sd, lb_balanced[idle]); |
2279 | ||
16cfb1c0 | 2280 | sd->nr_balance_failed = 0; |
fa3b6ddc SS |
2281 | |
2282 | out_one_pinned: | |
1da177e4 | 2283 | /* tune up the balancing interval */ |
77391d71 NP |
2284 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || |
2285 | (sd->balance_interval < sd->max_interval)) | |
1da177e4 LT |
2286 | sd->balance_interval *= 2; |
2287 | ||
5969fe06 NP |
2288 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
2289 | return -1; | |
1da177e4 LT |
2290 | return 0; |
2291 | } | |
2292 | ||
2293 | /* | |
2294 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
2295 | * tasks if there is an imbalance. | |
2296 | * | |
2297 | * Called from schedule when this_rq is about to become idle (NEWLY_IDLE). | |
2298 | * this_rq is locked. | |
2299 | */ | |
2300 | static int load_balance_newidle(int this_cpu, runqueue_t *this_rq, | |
2301 | struct sched_domain *sd) | |
2302 | { | |
2303 | struct sched_group *group; | |
2304 | runqueue_t *busiest = NULL; | |
2305 | unsigned long imbalance; | |
2306 | int nr_moved = 0; | |
5969fe06 NP |
2307 | int sd_idle = 0; |
2308 | ||
2309 | if (sd->flags & SD_SHARE_CPUPOWER) | |
2310 | sd_idle = 1; | |
1da177e4 LT |
2311 | |
2312 | schedstat_inc(sd, lb_cnt[NEWLY_IDLE]); | |
5969fe06 | 2313 | group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE, &sd_idle); |
1da177e4 | 2314 | if (!group) { |
1da177e4 | 2315 | schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]); |
16cfb1c0 | 2316 | goto out_balanced; |
1da177e4 LT |
2317 | } |
2318 | ||
b910472d | 2319 | busiest = find_busiest_queue(group, NEWLY_IDLE); |
db935dbd | 2320 | if (!busiest) { |
1da177e4 | 2321 | schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]); |
16cfb1c0 | 2322 | goto out_balanced; |
1da177e4 LT |
2323 | } |
2324 | ||
db935dbd NP |
2325 | BUG_ON(busiest == this_rq); |
2326 | ||
1da177e4 | 2327 | schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance); |
d6d5cfaf NP |
2328 | |
2329 | nr_moved = 0; | |
2330 | if (busiest->nr_running > 1) { | |
2331 | /* Attempt to move tasks */ | |
2332 | double_lock_balance(this_rq, busiest); | |
2333 | nr_moved = move_tasks(this_rq, this_cpu, busiest, | |
81026794 | 2334 | imbalance, sd, NEWLY_IDLE, NULL); |
d6d5cfaf NP |
2335 | spin_unlock(&busiest->lock); |
2336 | } | |
2337 | ||
5969fe06 | 2338 | if (!nr_moved) { |
1da177e4 | 2339 | schedstat_inc(sd, lb_failed[NEWLY_IDLE]); |
5969fe06 NP |
2340 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
2341 | return -1; | |
2342 | } else | |
16cfb1c0 | 2343 | sd->nr_balance_failed = 0; |
1da177e4 | 2344 | |
1da177e4 | 2345 | return nr_moved; |
16cfb1c0 NP |
2346 | |
2347 | out_balanced: | |
2348 | schedstat_inc(sd, lb_balanced[NEWLY_IDLE]); | |
5969fe06 NP |
2349 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
2350 | return -1; | |
16cfb1c0 NP |
2351 | sd->nr_balance_failed = 0; |
2352 | return 0; | |
1da177e4 LT |
2353 | } |
2354 | ||
2355 | /* | |
2356 | * idle_balance is called by schedule() if this_cpu is about to become | |
2357 | * idle. Attempts to pull tasks from other CPUs. | |
2358 | */ | |
2359 | static inline void idle_balance(int this_cpu, runqueue_t *this_rq) | |
2360 | { | |
2361 | struct sched_domain *sd; | |
2362 | ||
2363 | for_each_domain(this_cpu, sd) { | |
2364 | if (sd->flags & SD_BALANCE_NEWIDLE) { | |
2365 | if (load_balance_newidle(this_cpu, this_rq, sd)) { | |
2366 | /* We've pulled tasks over so stop searching */ | |
2367 | break; | |
2368 | } | |
2369 | } | |
2370 | } | |
2371 | } | |
2372 | ||
2373 | /* | |
2374 | * active_load_balance is run by migration threads. It pushes running tasks | |
2375 | * off the busiest CPU onto idle CPUs. It requires at least 1 task to be | |
2376 | * running on each physical CPU where possible, and avoids physical / | |
2377 | * logical imbalances. | |
2378 | * | |
2379 | * Called with busiest_rq locked. | |
2380 | */ | |
2381 | static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu) | |
2382 | { | |
2383 | struct sched_domain *sd; | |
1da177e4 | 2384 | runqueue_t *target_rq; |
39507451 NP |
2385 | int target_cpu = busiest_rq->push_cpu; |
2386 | ||
2387 | if (busiest_rq->nr_running <= 1) | |
2388 | /* no task to move */ | |
2389 | return; | |
2390 | ||
2391 | target_rq = cpu_rq(target_cpu); | |
1da177e4 LT |
2392 | |
2393 | /* | |
39507451 NP |
2394 | * This condition is "impossible", if it occurs |
2395 | * we need to fix it. Originally reported by | |
2396 | * Bjorn Helgaas on a 128-cpu setup. | |
1da177e4 | 2397 | */ |
39507451 | 2398 | BUG_ON(busiest_rq == target_rq); |
1da177e4 | 2399 | |
39507451 NP |
2400 | /* move a task from busiest_rq to target_rq */ |
2401 | double_lock_balance(busiest_rq, target_rq); | |
2402 | ||
2403 | /* Search for an sd spanning us and the target CPU. */ | |
2404 | for_each_domain(target_cpu, sd) | |
2405 | if ((sd->flags & SD_LOAD_BALANCE) && | |
2406 | cpu_isset(busiest_cpu, sd->span)) | |
2407 | break; | |
2408 | ||
2409 | if (unlikely(sd == NULL)) | |
2410 | goto out; | |
2411 | ||
2412 | schedstat_inc(sd, alb_cnt); | |
2413 | ||
2414 | if (move_tasks(target_rq, target_cpu, busiest_rq, 1, sd, SCHED_IDLE, NULL)) | |
2415 | schedstat_inc(sd, alb_pushed); | |
2416 | else | |
2417 | schedstat_inc(sd, alb_failed); | |
2418 | out: | |
2419 | spin_unlock(&target_rq->lock); | |
1da177e4 LT |
2420 | } |
2421 | ||
2422 | /* | |
2423 | * rebalance_tick will get called every timer tick, on every CPU. | |
2424 | * | |
2425 | * It checks each scheduling domain to see if it is due to be balanced, | |
2426 | * and initiates a balancing operation if so. | |
2427 | * | |
2428 | * Balancing parameters are set up in arch_init_sched_domains. | |
2429 | */ | |
2430 | ||
2431 | /* Don't have all balancing operations going off at once */ | |
2432 | #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS) | |
2433 | ||
2434 | static void rebalance_tick(int this_cpu, runqueue_t *this_rq, | |
2435 | enum idle_type idle) | |
2436 | { | |
2437 | unsigned long old_load, this_load; | |
2438 | unsigned long j = jiffies + CPU_OFFSET(this_cpu); | |
2439 | struct sched_domain *sd; | |
7897986b | 2440 | int i; |
1da177e4 | 2441 | |
1da177e4 | 2442 | this_load = this_rq->nr_running * SCHED_LOAD_SCALE; |
7897986b NP |
2443 | /* Update our load */ |
2444 | for (i = 0; i < 3; i++) { | |
2445 | unsigned long new_load = this_load; | |
2446 | int scale = 1 << i; | |
2447 | old_load = this_rq->cpu_load[i]; | |
2448 | /* | |
2449 | * Round up the averaging division if load is increasing. This | |
2450 | * prevents us from getting stuck on 9 if the load is 10, for | |
2451 | * example. | |
2452 | */ | |
2453 | if (new_load > old_load) | |
2454 | new_load += scale-1; | |
2455 | this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale; | |
2456 | } | |
1da177e4 LT |
2457 | |
2458 | for_each_domain(this_cpu, sd) { | |
2459 | unsigned long interval; | |
2460 | ||
2461 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
2462 | continue; | |
2463 | ||
2464 | interval = sd->balance_interval; | |
2465 | if (idle != SCHED_IDLE) | |
2466 | interval *= sd->busy_factor; | |
2467 | ||
2468 | /* scale ms to jiffies */ | |
2469 | interval = msecs_to_jiffies(interval); | |
2470 | if (unlikely(!interval)) | |
2471 | interval = 1; | |
2472 | ||
2473 | if (j - sd->last_balance >= interval) { | |
2474 | if (load_balance(this_cpu, this_rq, sd, idle)) { | |
fa3b6ddc SS |
2475 | /* |
2476 | * We've pulled tasks over so either we're no | |
5969fe06 NP |
2477 | * longer idle, or one of our SMT siblings is |
2478 | * not idle. | |
2479 | */ | |
1da177e4 LT |
2480 | idle = NOT_IDLE; |
2481 | } | |
2482 | sd->last_balance += interval; | |
2483 | } | |
2484 | } | |
2485 | } | |
2486 | #else | |
2487 | /* | |
2488 | * on UP we do not need to balance between CPUs: | |
2489 | */ | |
2490 | static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle) | |
2491 | { | |
2492 | } | |
2493 | static inline void idle_balance(int cpu, runqueue_t *rq) | |
2494 | { | |
2495 | } | |
2496 | #endif | |
2497 | ||
2498 | static inline int wake_priority_sleeper(runqueue_t *rq) | |
2499 | { | |
2500 | int ret = 0; | |
2501 | #ifdef CONFIG_SCHED_SMT | |
2502 | spin_lock(&rq->lock); | |
2503 | /* | |
2504 | * If an SMT sibling task has been put to sleep for priority | |
2505 | * reasons reschedule the idle task to see if it can now run. | |
2506 | */ | |
2507 | if (rq->nr_running) { | |
2508 | resched_task(rq->idle); | |
2509 | ret = 1; | |
2510 | } | |
2511 | spin_unlock(&rq->lock); | |
2512 | #endif | |
2513 | return ret; | |
2514 | } | |
2515 | ||
2516 | DEFINE_PER_CPU(struct kernel_stat, kstat); | |
2517 | ||
2518 | EXPORT_PER_CPU_SYMBOL(kstat); | |
2519 | ||
2520 | /* | |
2521 | * This is called on clock ticks and on context switches. | |
2522 | * Bank in p->sched_time the ns elapsed since the last tick or switch. | |
2523 | */ | |
2524 | static inline void update_cpu_clock(task_t *p, runqueue_t *rq, | |
2525 | unsigned long long now) | |
2526 | { | |
2527 | unsigned long long last = max(p->timestamp, rq->timestamp_last_tick); | |
2528 | p->sched_time += now - last; | |
2529 | } | |
2530 | ||
2531 | /* | |
2532 | * Return current->sched_time plus any more ns on the sched_clock | |
2533 | * that have not yet been banked. | |
2534 | */ | |
2535 | unsigned long long current_sched_time(const task_t *tsk) | |
2536 | { | |
2537 | unsigned long long ns; | |
2538 | unsigned long flags; | |
2539 | local_irq_save(flags); | |
2540 | ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick); | |
2541 | ns = tsk->sched_time + (sched_clock() - ns); | |
2542 | local_irq_restore(flags); | |
2543 | return ns; | |
2544 | } | |
2545 | ||
2546 | /* | |
2547 | * We place interactive tasks back into the active array, if possible. | |
2548 | * | |
2549 | * To guarantee that this does not starve expired tasks we ignore the | |
2550 | * interactivity of a task if the first expired task had to wait more | |
2551 | * than a 'reasonable' amount of time. This deadline timeout is | |
2552 | * load-dependent, as the frequency of array switched decreases with | |
2553 | * increasing number of running tasks. We also ignore the interactivity | |
2554 | * if a better static_prio task has expired: | |
2555 | */ | |
2556 | #define EXPIRED_STARVING(rq) \ | |
2557 | ((STARVATION_LIMIT && ((rq)->expired_timestamp && \ | |
2558 | (jiffies - (rq)->expired_timestamp >= \ | |
2559 | STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \ | |
2560 | ((rq)->curr->static_prio > (rq)->best_expired_prio)) | |
2561 | ||
2562 | /* | |
2563 | * Account user cpu time to a process. | |
2564 | * @p: the process that the cpu time gets accounted to | |
2565 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
2566 | * @cputime: the cpu time spent in user space since the last update | |
2567 | */ | |
2568 | void account_user_time(struct task_struct *p, cputime_t cputime) | |
2569 | { | |
2570 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
2571 | cputime64_t tmp; | |
2572 | ||
2573 | p->utime = cputime_add(p->utime, cputime); | |
2574 | ||
2575 | /* Add user time to cpustat. */ | |
2576 | tmp = cputime_to_cputime64(cputime); | |
2577 | if (TASK_NICE(p) > 0) | |
2578 | cpustat->nice = cputime64_add(cpustat->nice, tmp); | |
2579 | else | |
2580 | cpustat->user = cputime64_add(cpustat->user, tmp); | |
2581 | } | |
2582 | ||
2583 | /* | |
2584 | * Account system cpu time to a process. | |
2585 | * @p: the process that the cpu time gets accounted to | |
2586 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
2587 | * @cputime: the cpu time spent in kernel space since the last update | |
2588 | */ | |
2589 | void account_system_time(struct task_struct *p, int hardirq_offset, | |
2590 | cputime_t cputime) | |
2591 | { | |
2592 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
2593 | runqueue_t *rq = this_rq(); | |
2594 | cputime64_t tmp; | |
2595 | ||
2596 | p->stime = cputime_add(p->stime, cputime); | |
2597 | ||
2598 | /* Add system time to cpustat. */ | |
2599 | tmp = cputime_to_cputime64(cputime); | |
2600 | if (hardirq_count() - hardirq_offset) | |
2601 | cpustat->irq = cputime64_add(cpustat->irq, tmp); | |
2602 | else if (softirq_count()) | |
2603 | cpustat->softirq = cputime64_add(cpustat->softirq, tmp); | |
2604 | else if (p != rq->idle) | |
2605 | cpustat->system = cputime64_add(cpustat->system, tmp); | |
2606 | else if (atomic_read(&rq->nr_iowait) > 0) | |
2607 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); | |
2608 | else | |
2609 | cpustat->idle = cputime64_add(cpustat->idle, tmp); | |
2610 | /* Account for system time used */ | |
2611 | acct_update_integrals(p); | |
1da177e4 LT |
2612 | } |
2613 | ||
2614 | /* | |
2615 | * Account for involuntary wait time. | |
2616 | * @p: the process from which the cpu time has been stolen | |
2617 | * @steal: the cpu time spent in involuntary wait | |
2618 | */ | |
2619 | void account_steal_time(struct task_struct *p, cputime_t steal) | |
2620 | { | |
2621 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
2622 | cputime64_t tmp = cputime_to_cputime64(steal); | |
2623 | runqueue_t *rq = this_rq(); | |
2624 | ||
2625 | if (p == rq->idle) { | |
2626 | p->stime = cputime_add(p->stime, steal); | |
2627 | if (atomic_read(&rq->nr_iowait) > 0) | |
2628 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); | |
2629 | else | |
2630 | cpustat->idle = cputime64_add(cpustat->idle, tmp); | |
2631 | } else | |
2632 | cpustat->steal = cputime64_add(cpustat->steal, tmp); | |
2633 | } | |
2634 | ||
2635 | /* | |
2636 | * This function gets called by the timer code, with HZ frequency. | |
2637 | * We call it with interrupts disabled. | |
2638 | * | |
2639 | * It also gets called by the fork code, when changing the parent's | |
2640 | * timeslices. | |
2641 | */ | |
2642 | void scheduler_tick(void) | |
2643 | { | |
2644 | int cpu = smp_processor_id(); | |
2645 | runqueue_t *rq = this_rq(); | |
2646 | task_t *p = current; | |
2647 | unsigned long long now = sched_clock(); | |
2648 | ||
2649 | update_cpu_clock(p, rq, now); | |
2650 | ||
2651 | rq->timestamp_last_tick = now; | |
2652 | ||
2653 | if (p == rq->idle) { | |
2654 | if (wake_priority_sleeper(rq)) | |
2655 | goto out; | |
2656 | rebalance_tick(cpu, rq, SCHED_IDLE); | |
2657 | return; | |
2658 | } | |
2659 | ||
2660 | /* Task might have expired already, but not scheduled off yet */ | |
2661 | if (p->array != rq->active) { | |
2662 | set_tsk_need_resched(p); | |
2663 | goto out; | |
2664 | } | |
2665 | spin_lock(&rq->lock); | |
2666 | /* | |
2667 | * The task was running during this tick - update the | |
2668 | * time slice counter. Note: we do not update a thread's | |
2669 | * priority until it either goes to sleep or uses up its | |
2670 | * timeslice. This makes it possible for interactive tasks | |
2671 | * to use up their timeslices at their highest priority levels. | |
2672 | */ | |
2673 | if (rt_task(p)) { | |
2674 | /* | |
2675 | * RR tasks need a special form of timeslice management. | |
2676 | * FIFO tasks have no timeslices. | |
2677 | */ | |
2678 | if ((p->policy == SCHED_RR) && !--p->time_slice) { | |
2679 | p->time_slice = task_timeslice(p); | |
2680 | p->first_time_slice = 0; | |
2681 | set_tsk_need_resched(p); | |
2682 | ||
2683 | /* put it at the end of the queue: */ | |
2684 | requeue_task(p, rq->active); | |
2685 | } | |
2686 | goto out_unlock; | |
2687 | } | |
2688 | if (!--p->time_slice) { | |
2689 | dequeue_task(p, rq->active); | |
2690 | set_tsk_need_resched(p); | |
2691 | p->prio = effective_prio(p); | |
2692 | p->time_slice = task_timeslice(p); | |
2693 | p->first_time_slice = 0; | |
2694 | ||
2695 | if (!rq->expired_timestamp) | |
2696 | rq->expired_timestamp = jiffies; | |
2697 | if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) { | |
2698 | enqueue_task(p, rq->expired); | |
2699 | if (p->static_prio < rq->best_expired_prio) | |
2700 | rq->best_expired_prio = p->static_prio; | |
2701 | } else | |
2702 | enqueue_task(p, rq->active); | |
2703 | } else { | |
2704 | /* | |
2705 | * Prevent a too long timeslice allowing a task to monopolize | |
2706 | * the CPU. We do this by splitting up the timeslice into | |
2707 | * smaller pieces. | |
2708 | * | |
2709 | * Note: this does not mean the task's timeslices expire or | |
2710 | * get lost in any way, they just might be preempted by | |
2711 | * another task of equal priority. (one with higher | |
2712 | * priority would have preempted this task already.) We | |
2713 | * requeue this task to the end of the list on this priority | |
2714 | * level, which is in essence a round-robin of tasks with | |
2715 | * equal priority. | |
2716 | * | |
2717 | * This only applies to tasks in the interactive | |
2718 | * delta range with at least TIMESLICE_GRANULARITY to requeue. | |
2719 | */ | |
2720 | if (TASK_INTERACTIVE(p) && !((task_timeslice(p) - | |
2721 | p->time_slice) % TIMESLICE_GRANULARITY(p)) && | |
2722 | (p->time_slice >= TIMESLICE_GRANULARITY(p)) && | |
2723 | (p->array == rq->active)) { | |
2724 | ||
2725 | requeue_task(p, rq->active); | |
2726 | set_tsk_need_resched(p); | |
2727 | } | |
2728 | } | |
2729 | out_unlock: | |
2730 | spin_unlock(&rq->lock); | |
2731 | out: | |
2732 | rebalance_tick(cpu, rq, NOT_IDLE); | |
2733 | } | |
2734 | ||
2735 | #ifdef CONFIG_SCHED_SMT | |
fc38ed75 CK |
2736 | static inline void wakeup_busy_runqueue(runqueue_t *rq) |
2737 | { | |
2738 | /* If an SMT runqueue is sleeping due to priority reasons wake it up */ | |
2739 | if (rq->curr == rq->idle && rq->nr_running) | |
2740 | resched_task(rq->idle); | |
2741 | } | |
2742 | ||
1da177e4 LT |
2743 | static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq) |
2744 | { | |
41c7ce9a | 2745 | struct sched_domain *tmp, *sd = NULL; |
1da177e4 LT |
2746 | cpumask_t sibling_map; |
2747 | int i; | |
2748 | ||
41c7ce9a NP |
2749 | for_each_domain(this_cpu, tmp) |
2750 | if (tmp->flags & SD_SHARE_CPUPOWER) | |
2751 | sd = tmp; | |
2752 | ||
2753 | if (!sd) | |
1da177e4 LT |
2754 | return; |
2755 | ||
2756 | /* | |
2757 | * Unlock the current runqueue because we have to lock in | |
2758 | * CPU order to avoid deadlocks. Caller knows that we might | |
2759 | * unlock. We keep IRQs disabled. | |
2760 | */ | |
2761 | spin_unlock(&this_rq->lock); | |
2762 | ||
2763 | sibling_map = sd->span; | |
2764 | ||
2765 | for_each_cpu_mask(i, sibling_map) | |
2766 | spin_lock(&cpu_rq(i)->lock); | |
2767 | /* | |
2768 | * We clear this CPU from the mask. This both simplifies the | |
2769 | * inner loop and keps this_rq locked when we exit: | |
2770 | */ | |
2771 | cpu_clear(this_cpu, sibling_map); | |
2772 | ||
2773 | for_each_cpu_mask(i, sibling_map) { | |
2774 | runqueue_t *smt_rq = cpu_rq(i); | |
2775 | ||
fc38ed75 | 2776 | wakeup_busy_runqueue(smt_rq); |
1da177e4 LT |
2777 | } |
2778 | ||
2779 | for_each_cpu_mask(i, sibling_map) | |
2780 | spin_unlock(&cpu_rq(i)->lock); | |
2781 | /* | |
2782 | * We exit with this_cpu's rq still held and IRQs | |
2783 | * still disabled: | |
2784 | */ | |
2785 | } | |
2786 | ||
67f9a619 IM |
2787 | /* |
2788 | * number of 'lost' timeslices this task wont be able to fully | |
2789 | * utilize, if another task runs on a sibling. This models the | |
2790 | * slowdown effect of other tasks running on siblings: | |
2791 | */ | |
2792 | static inline unsigned long smt_slice(task_t *p, struct sched_domain *sd) | |
2793 | { | |
2794 | return p->time_slice * (100 - sd->per_cpu_gain) / 100; | |
2795 | } | |
2796 | ||
1da177e4 LT |
2797 | static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq) |
2798 | { | |
41c7ce9a | 2799 | struct sched_domain *tmp, *sd = NULL; |
1da177e4 LT |
2800 | cpumask_t sibling_map; |
2801 | prio_array_t *array; | |
2802 | int ret = 0, i; | |
2803 | task_t *p; | |
2804 | ||
41c7ce9a NP |
2805 | for_each_domain(this_cpu, tmp) |
2806 | if (tmp->flags & SD_SHARE_CPUPOWER) | |
2807 | sd = tmp; | |
2808 | ||
2809 | if (!sd) | |
1da177e4 LT |
2810 | return 0; |
2811 | ||
2812 | /* | |
2813 | * The same locking rules and details apply as for | |
2814 | * wake_sleeping_dependent(): | |
2815 | */ | |
2816 | spin_unlock(&this_rq->lock); | |
2817 | sibling_map = sd->span; | |
2818 | for_each_cpu_mask(i, sibling_map) | |
2819 | spin_lock(&cpu_rq(i)->lock); | |
2820 | cpu_clear(this_cpu, sibling_map); | |
2821 | ||
2822 | /* | |
2823 | * Establish next task to be run - it might have gone away because | |
2824 | * we released the runqueue lock above: | |
2825 | */ | |
2826 | if (!this_rq->nr_running) | |
2827 | goto out_unlock; | |
2828 | array = this_rq->active; | |
2829 | if (!array->nr_active) | |
2830 | array = this_rq->expired; | |
2831 | BUG_ON(!array->nr_active); | |
2832 | ||
2833 | p = list_entry(array->queue[sched_find_first_bit(array->bitmap)].next, | |
2834 | task_t, run_list); | |
2835 | ||
2836 | for_each_cpu_mask(i, sibling_map) { | |
2837 | runqueue_t *smt_rq = cpu_rq(i); | |
2838 | task_t *smt_curr = smt_rq->curr; | |
2839 | ||
fc38ed75 CK |
2840 | /* Kernel threads do not participate in dependent sleeping */ |
2841 | if (!p->mm || !smt_curr->mm || rt_task(p)) | |
2842 | goto check_smt_task; | |
2843 | ||
1da177e4 LT |
2844 | /* |
2845 | * If a user task with lower static priority than the | |
2846 | * running task on the SMT sibling is trying to schedule, | |
2847 | * delay it till there is proportionately less timeslice | |
2848 | * left of the sibling task to prevent a lower priority | |
2849 | * task from using an unfair proportion of the | |
2850 | * physical cpu's resources. -ck | |
2851 | */ | |
fc38ed75 CK |
2852 | if (rt_task(smt_curr)) { |
2853 | /* | |
2854 | * With real time tasks we run non-rt tasks only | |
2855 | * per_cpu_gain% of the time. | |
2856 | */ | |
2857 | if ((jiffies % DEF_TIMESLICE) > | |
2858 | (sd->per_cpu_gain * DEF_TIMESLICE / 100)) | |
2859 | ret = 1; | |
2860 | } else | |
67f9a619 IM |
2861 | if (smt_curr->static_prio < p->static_prio && |
2862 | !TASK_PREEMPTS_CURR(p, smt_rq) && | |
2863 | smt_slice(smt_curr, sd) > task_timeslice(p)) | |
fc38ed75 CK |
2864 | ret = 1; |
2865 | ||
2866 | check_smt_task: | |
2867 | if ((!smt_curr->mm && smt_curr != smt_rq->idle) || | |
2868 | rt_task(smt_curr)) | |
2869 | continue; | |
2870 | if (!p->mm) { | |
2871 | wakeup_busy_runqueue(smt_rq); | |
2872 | continue; | |
2873 | } | |
1da177e4 LT |
2874 | |
2875 | /* | |
fc38ed75 CK |
2876 | * Reschedule a lower priority task on the SMT sibling for |
2877 | * it to be put to sleep, or wake it up if it has been put to | |
2878 | * sleep for priority reasons to see if it should run now. | |
1da177e4 | 2879 | */ |
fc38ed75 CK |
2880 | if (rt_task(p)) { |
2881 | if ((jiffies % DEF_TIMESLICE) > | |
2882 | (sd->per_cpu_gain * DEF_TIMESLICE / 100)) | |
2883 | resched_task(smt_curr); | |
2884 | } else { | |
67f9a619 IM |
2885 | if (TASK_PREEMPTS_CURR(p, smt_rq) && |
2886 | smt_slice(p, sd) > task_timeslice(smt_curr)) | |
fc38ed75 CK |
2887 | resched_task(smt_curr); |
2888 | else | |
2889 | wakeup_busy_runqueue(smt_rq); | |
2890 | } | |
1da177e4 LT |
2891 | } |
2892 | out_unlock: | |
2893 | for_each_cpu_mask(i, sibling_map) | |
2894 | spin_unlock(&cpu_rq(i)->lock); | |
2895 | return ret; | |
2896 | } | |
2897 | #else | |
2898 | static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq) | |
2899 | { | |
2900 | } | |
2901 | ||
2902 | static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq) | |
2903 | { | |
2904 | return 0; | |
2905 | } | |
2906 | #endif | |
2907 | ||
2908 | #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT) | |
2909 | ||
2910 | void fastcall add_preempt_count(int val) | |
2911 | { | |
2912 | /* | |
2913 | * Underflow? | |
2914 | */ | |
be5b4fbd | 2915 | BUG_ON((preempt_count() < 0)); |
1da177e4 LT |
2916 | preempt_count() += val; |
2917 | /* | |
2918 | * Spinlock count overflowing soon? | |
2919 | */ | |
2920 | BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10); | |
2921 | } | |
2922 | EXPORT_SYMBOL(add_preempt_count); | |
2923 | ||
2924 | void fastcall sub_preempt_count(int val) | |
2925 | { | |
2926 | /* | |
2927 | * Underflow? | |
2928 | */ | |
2929 | BUG_ON(val > preempt_count()); | |
2930 | /* | |
2931 | * Is the spinlock portion underflowing? | |
2932 | */ | |
2933 | BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK)); | |
2934 | preempt_count() -= val; | |
2935 | } | |
2936 | EXPORT_SYMBOL(sub_preempt_count); | |
2937 | ||
2938 | #endif | |
2939 | ||
2940 | /* | |
2941 | * schedule() is the main scheduler function. | |
2942 | */ | |
2943 | asmlinkage void __sched schedule(void) | |
2944 | { | |
2945 | long *switch_count; | |
2946 | task_t *prev, *next; | |
2947 | runqueue_t *rq; | |
2948 | prio_array_t *array; | |
2949 | struct list_head *queue; | |
2950 | unsigned long long now; | |
2951 | unsigned long run_time; | |
a3464a10 | 2952 | int cpu, idx, new_prio; |
1da177e4 LT |
2953 | |
2954 | /* | |
2955 | * Test if we are atomic. Since do_exit() needs to call into | |
2956 | * schedule() atomically, we ignore that path for now. | |
2957 | * Otherwise, whine if we are scheduling when we should not be. | |
2958 | */ | |
2959 | if (likely(!current->exit_state)) { | |
2960 | if (unlikely(in_atomic())) { | |
2961 | printk(KERN_ERR "scheduling while atomic: " | |
2962 | "%s/0x%08x/%d\n", | |
2963 | current->comm, preempt_count(), current->pid); | |
2964 | dump_stack(); | |
2965 | } | |
2966 | } | |
2967 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); | |
2968 | ||
2969 | need_resched: | |
2970 | preempt_disable(); | |
2971 | prev = current; | |
2972 | release_kernel_lock(prev); | |
2973 | need_resched_nonpreemptible: | |
2974 | rq = this_rq(); | |
2975 | ||
2976 | /* | |
2977 | * The idle thread is not allowed to schedule! | |
2978 | * Remove this check after it has been exercised a bit. | |
2979 | */ | |
2980 | if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) { | |
2981 | printk(KERN_ERR "bad: scheduling from the idle thread!\n"); | |
2982 | dump_stack(); | |
2983 | } | |
2984 | ||
2985 | schedstat_inc(rq, sched_cnt); | |
2986 | now = sched_clock(); | |
238628ed | 2987 | if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) { |
1da177e4 | 2988 | run_time = now - prev->timestamp; |
238628ed | 2989 | if (unlikely((long long)(now - prev->timestamp) < 0)) |
1da177e4 LT |
2990 | run_time = 0; |
2991 | } else | |
2992 | run_time = NS_MAX_SLEEP_AVG; | |
2993 | ||
2994 | /* | |
2995 | * Tasks charged proportionately less run_time at high sleep_avg to | |
2996 | * delay them losing their interactive status | |
2997 | */ | |
2998 | run_time /= (CURRENT_BONUS(prev) ? : 1); | |
2999 | ||
3000 | spin_lock_irq(&rq->lock); | |
3001 | ||
3002 | if (unlikely(prev->flags & PF_DEAD)) | |
3003 | prev->state = EXIT_DEAD; | |
3004 | ||
3005 | switch_count = &prev->nivcsw; | |
3006 | if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { | |
3007 | switch_count = &prev->nvcsw; | |
3008 | if (unlikely((prev->state & TASK_INTERRUPTIBLE) && | |
3009 | unlikely(signal_pending(prev)))) | |
3010 | prev->state = TASK_RUNNING; | |
3011 | else { | |
3012 | if (prev->state == TASK_UNINTERRUPTIBLE) | |
3013 | rq->nr_uninterruptible++; | |
3014 | deactivate_task(prev, rq); | |
3015 | } | |
3016 | } | |
3017 | ||
3018 | cpu = smp_processor_id(); | |
3019 | if (unlikely(!rq->nr_running)) { | |
3020 | go_idle: | |
3021 | idle_balance(cpu, rq); | |
3022 | if (!rq->nr_running) { | |
3023 | next = rq->idle; | |
3024 | rq->expired_timestamp = 0; | |
3025 | wake_sleeping_dependent(cpu, rq); | |
3026 | /* | |
3027 | * wake_sleeping_dependent() might have released | |
3028 | * the runqueue, so break out if we got new | |
3029 | * tasks meanwhile: | |
3030 | */ | |
3031 | if (!rq->nr_running) | |
3032 | goto switch_tasks; | |
3033 | } | |
3034 | } else { | |
3035 | if (dependent_sleeper(cpu, rq)) { | |
3036 | next = rq->idle; | |
3037 | goto switch_tasks; | |
3038 | } | |
3039 | /* | |
3040 | * dependent_sleeper() releases and reacquires the runqueue | |
3041 | * lock, hence go into the idle loop if the rq went | |
3042 | * empty meanwhile: | |
3043 | */ | |
3044 | if (unlikely(!rq->nr_running)) | |
3045 | goto go_idle; | |
3046 | } | |
3047 | ||
3048 | array = rq->active; | |
3049 | if (unlikely(!array->nr_active)) { | |
3050 | /* | |
3051 | * Switch the active and expired arrays. | |
3052 | */ | |
3053 | schedstat_inc(rq, sched_switch); | |
3054 | rq->active = rq->expired; | |
3055 | rq->expired = array; | |
3056 | array = rq->active; | |
3057 | rq->expired_timestamp = 0; | |
3058 | rq->best_expired_prio = MAX_PRIO; | |
3059 | } | |
3060 | ||
3061 | idx = sched_find_first_bit(array->bitmap); | |
3062 | queue = array->queue + idx; | |
3063 | next = list_entry(queue->next, task_t, run_list); | |
3064 | ||
3065 | if (!rt_task(next) && next->activated > 0) { | |
3066 | unsigned long long delta = now - next->timestamp; | |
238628ed | 3067 | if (unlikely((long long)(now - next->timestamp) < 0)) |
1da177e4 LT |
3068 | delta = 0; |
3069 | ||
3070 | if (next->activated == 1) | |
3071 | delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128; | |
3072 | ||
3073 | array = next->array; | |
a3464a10 CS |
3074 | new_prio = recalc_task_prio(next, next->timestamp + delta); |
3075 | ||
3076 | if (unlikely(next->prio != new_prio)) { | |
3077 | dequeue_task(next, array); | |
3078 | next->prio = new_prio; | |
3079 | enqueue_task(next, array); | |
3080 | } else | |
3081 | requeue_task(next, array); | |
1da177e4 LT |
3082 | } |
3083 | next->activated = 0; | |
3084 | switch_tasks: | |
3085 | if (next == rq->idle) | |
3086 | schedstat_inc(rq, sched_goidle); | |
3087 | prefetch(next); | |
383f2835 | 3088 | prefetch_stack(next); |
1da177e4 LT |
3089 | clear_tsk_need_resched(prev); |
3090 | rcu_qsctr_inc(task_cpu(prev)); | |
3091 | ||
3092 | update_cpu_clock(prev, rq, now); | |
3093 | ||
3094 | prev->sleep_avg -= run_time; | |
3095 | if ((long)prev->sleep_avg <= 0) | |
3096 | prev->sleep_avg = 0; | |
3097 | prev->timestamp = prev->last_ran = now; | |
3098 | ||
3099 | sched_info_switch(prev, next); | |
3100 | if (likely(prev != next)) { | |
3101 | next->timestamp = now; | |
3102 | rq->nr_switches++; | |
3103 | rq->curr = next; | |
3104 | ++*switch_count; | |
3105 | ||
4866cde0 | 3106 | prepare_task_switch(rq, next); |
1da177e4 LT |
3107 | prev = context_switch(rq, prev, next); |
3108 | barrier(); | |
4866cde0 NP |
3109 | /* |
3110 | * this_rq must be evaluated again because prev may have moved | |
3111 | * CPUs since it called schedule(), thus the 'rq' on its stack | |
3112 | * frame will be invalid. | |
3113 | */ | |
3114 | finish_task_switch(this_rq(), prev); | |
1da177e4 LT |
3115 | } else |
3116 | spin_unlock_irq(&rq->lock); | |
3117 | ||
3118 | prev = current; | |
3119 | if (unlikely(reacquire_kernel_lock(prev) < 0)) | |
3120 | goto need_resched_nonpreemptible; | |
3121 | preempt_enable_no_resched(); | |
3122 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
3123 | goto need_resched; | |
3124 | } | |
3125 | ||
3126 | EXPORT_SYMBOL(schedule); | |
3127 | ||
3128 | #ifdef CONFIG_PREEMPT | |
3129 | /* | |
3130 | * this is is the entry point to schedule() from in-kernel preemption | |
3131 | * off of preempt_enable. Kernel preemptions off return from interrupt | |
3132 | * occur there and call schedule directly. | |
3133 | */ | |
3134 | asmlinkage void __sched preempt_schedule(void) | |
3135 | { | |
3136 | struct thread_info *ti = current_thread_info(); | |
3137 | #ifdef CONFIG_PREEMPT_BKL | |
3138 | struct task_struct *task = current; | |
3139 | int saved_lock_depth; | |
3140 | #endif | |
3141 | /* | |
3142 | * If there is a non-zero preempt_count or interrupts are disabled, | |
3143 | * we do not want to preempt the current task. Just return.. | |
3144 | */ | |
3145 | if (unlikely(ti->preempt_count || irqs_disabled())) | |
3146 | return; | |
3147 | ||
3148 | need_resched: | |
3149 | add_preempt_count(PREEMPT_ACTIVE); | |
3150 | /* | |
3151 | * We keep the big kernel semaphore locked, but we | |
3152 | * clear ->lock_depth so that schedule() doesnt | |
3153 | * auto-release the semaphore: | |
3154 | */ | |
3155 | #ifdef CONFIG_PREEMPT_BKL | |
3156 | saved_lock_depth = task->lock_depth; | |
3157 | task->lock_depth = -1; | |
3158 | #endif | |
3159 | schedule(); | |
3160 | #ifdef CONFIG_PREEMPT_BKL | |
3161 | task->lock_depth = saved_lock_depth; | |
3162 | #endif | |
3163 | sub_preempt_count(PREEMPT_ACTIVE); | |
3164 | ||
3165 | /* we could miss a preemption opportunity between schedule and now */ | |
3166 | barrier(); | |
3167 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
3168 | goto need_resched; | |
3169 | } | |
3170 | ||
3171 | EXPORT_SYMBOL(preempt_schedule); | |
3172 | ||
3173 | /* | |
3174 | * this is is the entry point to schedule() from kernel preemption | |
3175 | * off of irq context. | |
3176 | * Note, that this is called and return with irqs disabled. This will | |
3177 | * protect us against recursive calling from irq. | |
3178 | */ | |
3179 | asmlinkage void __sched preempt_schedule_irq(void) | |
3180 | { | |
3181 | struct thread_info *ti = current_thread_info(); | |
3182 | #ifdef CONFIG_PREEMPT_BKL | |
3183 | struct task_struct *task = current; | |
3184 | int saved_lock_depth; | |
3185 | #endif | |
3186 | /* Catch callers which need to be fixed*/ | |
3187 | BUG_ON(ti->preempt_count || !irqs_disabled()); | |
3188 | ||
3189 | need_resched: | |
3190 | add_preempt_count(PREEMPT_ACTIVE); | |
3191 | /* | |
3192 | * We keep the big kernel semaphore locked, but we | |
3193 | * clear ->lock_depth so that schedule() doesnt | |
3194 | * auto-release the semaphore: | |
3195 | */ | |
3196 | #ifdef CONFIG_PREEMPT_BKL | |
3197 | saved_lock_depth = task->lock_depth; | |
3198 | task->lock_depth = -1; | |
3199 | #endif | |
3200 | local_irq_enable(); | |
3201 | schedule(); | |
3202 | local_irq_disable(); | |
3203 | #ifdef CONFIG_PREEMPT_BKL | |
3204 | task->lock_depth = saved_lock_depth; | |
3205 | #endif | |
3206 | sub_preempt_count(PREEMPT_ACTIVE); | |
3207 | ||
3208 | /* we could miss a preemption opportunity between schedule and now */ | |
3209 | barrier(); | |
3210 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
3211 | goto need_resched; | |
3212 | } | |
3213 | ||
3214 | #endif /* CONFIG_PREEMPT */ | |
3215 | ||
95cdf3b7 IM |
3216 | int default_wake_function(wait_queue_t *curr, unsigned mode, int sync, |
3217 | void *key) | |
1da177e4 | 3218 | { |
c43dc2fd | 3219 | task_t *p = curr->private; |
1da177e4 LT |
3220 | return try_to_wake_up(p, mode, sync); |
3221 | } | |
3222 | ||
3223 | EXPORT_SYMBOL(default_wake_function); | |
3224 | ||
3225 | /* | |
3226 | * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just | |
3227 | * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve | |
3228 | * number) then we wake all the non-exclusive tasks and one exclusive task. | |
3229 | * | |
3230 | * There are circumstances in which we can try to wake a task which has already | |
3231 | * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns | |
3232 | * zero in this (rare) case, and we handle it by continuing to scan the queue. | |
3233 | */ | |
3234 | static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, | |
3235 | int nr_exclusive, int sync, void *key) | |
3236 | { | |
3237 | struct list_head *tmp, *next; | |
3238 | ||
3239 | list_for_each_safe(tmp, next, &q->task_list) { | |
3240 | wait_queue_t *curr; | |
3241 | unsigned flags; | |
3242 | curr = list_entry(tmp, wait_queue_t, task_list); | |
3243 | flags = curr->flags; | |
3244 | if (curr->func(curr, mode, sync, key) && | |
3245 | (flags & WQ_FLAG_EXCLUSIVE) && | |
3246 | !--nr_exclusive) | |
3247 | break; | |
3248 | } | |
3249 | } | |
3250 | ||
3251 | /** | |
3252 | * __wake_up - wake up threads blocked on a waitqueue. | |
3253 | * @q: the waitqueue | |
3254 | * @mode: which threads | |
3255 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
67be2dd1 | 3256 | * @key: is directly passed to the wakeup function |
1da177e4 LT |
3257 | */ |
3258 | void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode, | |
95cdf3b7 | 3259 | int nr_exclusive, void *key) |
1da177e4 LT |
3260 | { |
3261 | unsigned long flags; | |
3262 | ||
3263 | spin_lock_irqsave(&q->lock, flags); | |
3264 | __wake_up_common(q, mode, nr_exclusive, 0, key); | |
3265 | spin_unlock_irqrestore(&q->lock, flags); | |
3266 | } | |
3267 | ||
3268 | EXPORT_SYMBOL(__wake_up); | |
3269 | ||
3270 | /* | |
3271 | * Same as __wake_up but called with the spinlock in wait_queue_head_t held. | |
3272 | */ | |
3273 | void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode) | |
3274 | { | |
3275 | __wake_up_common(q, mode, 1, 0, NULL); | |
3276 | } | |
3277 | ||
3278 | /** | |
67be2dd1 | 3279 | * __wake_up_sync - wake up threads blocked on a waitqueue. |
1da177e4 LT |
3280 | * @q: the waitqueue |
3281 | * @mode: which threads | |
3282 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
3283 | * | |
3284 | * The sync wakeup differs that the waker knows that it will schedule | |
3285 | * away soon, so while the target thread will be woken up, it will not | |
3286 | * be migrated to another CPU - ie. the two threads are 'synchronized' | |
3287 | * with each other. This can prevent needless bouncing between CPUs. | |
3288 | * | |
3289 | * On UP it can prevent extra preemption. | |
3290 | */ | |
95cdf3b7 IM |
3291 | void fastcall |
3292 | __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) | |
1da177e4 LT |
3293 | { |
3294 | unsigned long flags; | |
3295 | int sync = 1; | |
3296 | ||
3297 | if (unlikely(!q)) | |
3298 | return; | |
3299 | ||
3300 | if (unlikely(!nr_exclusive)) | |
3301 | sync = 0; | |
3302 | ||
3303 | spin_lock_irqsave(&q->lock, flags); | |
3304 | __wake_up_common(q, mode, nr_exclusive, sync, NULL); | |
3305 | spin_unlock_irqrestore(&q->lock, flags); | |
3306 | } | |
3307 | EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ | |
3308 | ||
3309 | void fastcall complete(struct completion *x) | |
3310 | { | |
3311 | unsigned long flags; | |
3312 | ||
3313 | spin_lock_irqsave(&x->wait.lock, flags); | |
3314 | x->done++; | |
3315 | __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, | |
3316 | 1, 0, NULL); | |
3317 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
3318 | } | |
3319 | EXPORT_SYMBOL(complete); | |
3320 | ||
3321 | void fastcall complete_all(struct completion *x) | |
3322 | { | |
3323 | unsigned long flags; | |
3324 | ||
3325 | spin_lock_irqsave(&x->wait.lock, flags); | |
3326 | x->done += UINT_MAX/2; | |
3327 | __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, | |
3328 | 0, 0, NULL); | |
3329 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
3330 | } | |
3331 | EXPORT_SYMBOL(complete_all); | |
3332 | ||
3333 | void fastcall __sched wait_for_completion(struct completion *x) | |
3334 | { | |
3335 | might_sleep(); | |
3336 | spin_lock_irq(&x->wait.lock); | |
3337 | if (!x->done) { | |
3338 | DECLARE_WAITQUEUE(wait, current); | |
3339 | ||
3340 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3341 | __add_wait_queue_tail(&x->wait, &wait); | |
3342 | do { | |
3343 | __set_current_state(TASK_UNINTERRUPTIBLE); | |
3344 | spin_unlock_irq(&x->wait.lock); | |
3345 | schedule(); | |
3346 | spin_lock_irq(&x->wait.lock); | |
3347 | } while (!x->done); | |
3348 | __remove_wait_queue(&x->wait, &wait); | |
3349 | } | |
3350 | x->done--; | |
3351 | spin_unlock_irq(&x->wait.lock); | |
3352 | } | |
3353 | EXPORT_SYMBOL(wait_for_completion); | |
3354 | ||
3355 | unsigned long fastcall __sched | |
3356 | wait_for_completion_timeout(struct completion *x, unsigned long timeout) | |
3357 | { | |
3358 | might_sleep(); | |
3359 | ||
3360 | spin_lock_irq(&x->wait.lock); | |
3361 | if (!x->done) { | |
3362 | DECLARE_WAITQUEUE(wait, current); | |
3363 | ||
3364 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3365 | __add_wait_queue_tail(&x->wait, &wait); | |
3366 | do { | |
3367 | __set_current_state(TASK_UNINTERRUPTIBLE); | |
3368 | spin_unlock_irq(&x->wait.lock); | |
3369 | timeout = schedule_timeout(timeout); | |
3370 | spin_lock_irq(&x->wait.lock); | |
3371 | if (!timeout) { | |
3372 | __remove_wait_queue(&x->wait, &wait); | |
3373 | goto out; | |
3374 | } | |
3375 | } while (!x->done); | |
3376 | __remove_wait_queue(&x->wait, &wait); | |
3377 | } | |
3378 | x->done--; | |
3379 | out: | |
3380 | spin_unlock_irq(&x->wait.lock); | |
3381 | return timeout; | |
3382 | } | |
3383 | EXPORT_SYMBOL(wait_for_completion_timeout); | |
3384 | ||
3385 | int fastcall __sched wait_for_completion_interruptible(struct completion *x) | |
3386 | { | |
3387 | int ret = 0; | |
3388 | ||
3389 | might_sleep(); | |
3390 | ||
3391 | spin_lock_irq(&x->wait.lock); | |
3392 | if (!x->done) { | |
3393 | DECLARE_WAITQUEUE(wait, current); | |
3394 | ||
3395 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3396 | __add_wait_queue_tail(&x->wait, &wait); | |
3397 | do { | |
3398 | if (signal_pending(current)) { | |
3399 | ret = -ERESTARTSYS; | |
3400 | __remove_wait_queue(&x->wait, &wait); | |
3401 | goto out; | |
3402 | } | |
3403 | __set_current_state(TASK_INTERRUPTIBLE); | |
3404 | spin_unlock_irq(&x->wait.lock); | |
3405 | schedule(); | |
3406 | spin_lock_irq(&x->wait.lock); | |
3407 | } while (!x->done); | |
3408 | __remove_wait_queue(&x->wait, &wait); | |
3409 | } | |
3410 | x->done--; | |
3411 | out: | |
3412 | spin_unlock_irq(&x->wait.lock); | |
3413 | ||
3414 | return ret; | |
3415 | } | |
3416 | EXPORT_SYMBOL(wait_for_completion_interruptible); | |
3417 | ||
3418 | unsigned long fastcall __sched | |
3419 | wait_for_completion_interruptible_timeout(struct completion *x, | |
3420 | unsigned long timeout) | |
3421 | { | |
3422 | might_sleep(); | |
3423 | ||
3424 | spin_lock_irq(&x->wait.lock); | |
3425 | if (!x->done) { | |
3426 | DECLARE_WAITQUEUE(wait, current); | |
3427 | ||
3428 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3429 | __add_wait_queue_tail(&x->wait, &wait); | |
3430 | do { | |
3431 | if (signal_pending(current)) { | |
3432 | timeout = -ERESTARTSYS; | |
3433 | __remove_wait_queue(&x->wait, &wait); | |
3434 | goto out; | |
3435 | } | |
3436 | __set_current_state(TASK_INTERRUPTIBLE); | |
3437 | spin_unlock_irq(&x->wait.lock); | |
3438 | timeout = schedule_timeout(timeout); | |
3439 | spin_lock_irq(&x->wait.lock); | |
3440 | if (!timeout) { | |
3441 | __remove_wait_queue(&x->wait, &wait); | |
3442 | goto out; | |
3443 | } | |
3444 | } while (!x->done); | |
3445 | __remove_wait_queue(&x->wait, &wait); | |
3446 | } | |
3447 | x->done--; | |
3448 | out: | |
3449 | spin_unlock_irq(&x->wait.lock); | |
3450 | return timeout; | |
3451 | } | |
3452 | EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); | |
3453 | ||
3454 | ||
3455 | #define SLEEP_ON_VAR \ | |
3456 | unsigned long flags; \ | |
3457 | wait_queue_t wait; \ | |
3458 | init_waitqueue_entry(&wait, current); | |
3459 | ||
3460 | #define SLEEP_ON_HEAD \ | |
3461 | spin_lock_irqsave(&q->lock,flags); \ | |
3462 | __add_wait_queue(q, &wait); \ | |
3463 | spin_unlock(&q->lock); | |
3464 | ||
3465 | #define SLEEP_ON_TAIL \ | |
3466 | spin_lock_irq(&q->lock); \ | |
3467 | __remove_wait_queue(q, &wait); \ | |
3468 | spin_unlock_irqrestore(&q->lock, flags); | |
3469 | ||
3470 | void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q) | |
3471 | { | |
3472 | SLEEP_ON_VAR | |
3473 | ||
3474 | current->state = TASK_INTERRUPTIBLE; | |
3475 | ||
3476 | SLEEP_ON_HEAD | |
3477 | schedule(); | |
3478 | SLEEP_ON_TAIL | |
3479 | } | |
3480 | ||
3481 | EXPORT_SYMBOL(interruptible_sleep_on); | |
3482 | ||
95cdf3b7 IM |
3483 | long fastcall __sched |
3484 | interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) | |
1da177e4 LT |
3485 | { |
3486 | SLEEP_ON_VAR | |
3487 | ||
3488 | current->state = TASK_INTERRUPTIBLE; | |
3489 | ||
3490 | SLEEP_ON_HEAD | |
3491 | timeout = schedule_timeout(timeout); | |
3492 | SLEEP_ON_TAIL | |
3493 | ||
3494 | return timeout; | |
3495 | } | |
3496 | ||
3497 | EXPORT_SYMBOL(interruptible_sleep_on_timeout); | |
3498 | ||
3499 | void fastcall __sched sleep_on(wait_queue_head_t *q) | |
3500 | { | |
3501 | SLEEP_ON_VAR | |
3502 | ||
3503 | current->state = TASK_UNINTERRUPTIBLE; | |
3504 | ||
3505 | SLEEP_ON_HEAD | |
3506 | schedule(); | |
3507 | SLEEP_ON_TAIL | |
3508 | } | |
3509 | ||
3510 | EXPORT_SYMBOL(sleep_on); | |
3511 | ||
3512 | long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) | |
3513 | { | |
3514 | SLEEP_ON_VAR | |
3515 | ||
3516 | current->state = TASK_UNINTERRUPTIBLE; | |
3517 | ||
3518 | SLEEP_ON_HEAD | |
3519 | timeout = schedule_timeout(timeout); | |
3520 | SLEEP_ON_TAIL | |
3521 | ||
3522 | return timeout; | |
3523 | } | |
3524 | ||
3525 | EXPORT_SYMBOL(sleep_on_timeout); | |
3526 | ||
3527 | void set_user_nice(task_t *p, long nice) | |
3528 | { | |
3529 | unsigned long flags; | |
3530 | prio_array_t *array; | |
3531 | runqueue_t *rq; | |
3532 | int old_prio, new_prio, delta; | |
3533 | ||
3534 | if (TASK_NICE(p) == nice || nice < -20 || nice > 19) | |
3535 | return; | |
3536 | /* | |
3537 | * We have to be careful, if called from sys_setpriority(), | |
3538 | * the task might be in the middle of scheduling on another CPU. | |
3539 | */ | |
3540 | rq = task_rq_lock(p, &flags); | |
3541 | /* | |
3542 | * The RT priorities are set via sched_setscheduler(), but we still | |
3543 | * allow the 'normal' nice value to be set - but as expected | |
3544 | * it wont have any effect on scheduling until the task is | |
3545 | * not SCHED_NORMAL: | |
3546 | */ | |
3547 | if (rt_task(p)) { | |
3548 | p->static_prio = NICE_TO_PRIO(nice); | |
3549 | goto out_unlock; | |
3550 | } | |
3551 | array = p->array; | |
738a2ccb | 3552 | if (array) { |
1da177e4 | 3553 | dequeue_task(p, array); |
738a2ccb CK |
3554 | dec_prio_bias(rq, p->static_prio); |
3555 | } | |
1da177e4 LT |
3556 | |
3557 | old_prio = p->prio; | |
3558 | new_prio = NICE_TO_PRIO(nice); | |
3559 | delta = new_prio - old_prio; | |
3560 | p->static_prio = NICE_TO_PRIO(nice); | |
3561 | p->prio += delta; | |
3562 | ||
3563 | if (array) { | |
3564 | enqueue_task(p, array); | |
738a2ccb | 3565 | inc_prio_bias(rq, p->static_prio); |
1da177e4 LT |
3566 | /* |
3567 | * If the task increased its priority or is running and | |
3568 | * lowered its priority, then reschedule its CPU: | |
3569 | */ | |
3570 | if (delta < 0 || (delta > 0 && task_running(rq, p))) | |
3571 | resched_task(rq->curr); | |
3572 | } | |
3573 | out_unlock: | |
3574 | task_rq_unlock(rq, &flags); | |
3575 | } | |
3576 | ||
3577 | EXPORT_SYMBOL(set_user_nice); | |
3578 | ||
e43379f1 MM |
3579 | /* |
3580 | * can_nice - check if a task can reduce its nice value | |
3581 | * @p: task | |
3582 | * @nice: nice value | |
3583 | */ | |
3584 | int can_nice(const task_t *p, const int nice) | |
3585 | { | |
024f4747 MM |
3586 | /* convert nice value [19,-20] to rlimit style value [1,40] */ |
3587 | int nice_rlim = 20 - nice; | |
e43379f1 MM |
3588 | return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur || |
3589 | capable(CAP_SYS_NICE)); | |
3590 | } | |
3591 | ||
1da177e4 LT |
3592 | #ifdef __ARCH_WANT_SYS_NICE |
3593 | ||
3594 | /* | |
3595 | * sys_nice - change the priority of the current process. | |
3596 | * @increment: priority increment | |
3597 | * | |
3598 | * sys_setpriority is a more generic, but much slower function that | |
3599 | * does similar things. | |
3600 | */ | |
3601 | asmlinkage long sys_nice(int increment) | |
3602 | { | |
3603 | int retval; | |
3604 | long nice; | |
3605 | ||
3606 | /* | |
3607 | * Setpriority might change our priority at the same moment. | |
3608 | * We don't have to worry. Conceptually one call occurs first | |
3609 | * and we have a single winner. | |
3610 | */ | |
e43379f1 MM |
3611 | if (increment < -40) |
3612 | increment = -40; | |
1da177e4 LT |
3613 | if (increment > 40) |
3614 | increment = 40; | |
3615 | ||
3616 | nice = PRIO_TO_NICE(current->static_prio) + increment; | |
3617 | if (nice < -20) | |
3618 | nice = -20; | |
3619 | if (nice > 19) | |
3620 | nice = 19; | |
3621 | ||
e43379f1 MM |
3622 | if (increment < 0 && !can_nice(current, nice)) |
3623 | return -EPERM; | |
3624 | ||
1da177e4 LT |
3625 | retval = security_task_setnice(current, nice); |
3626 | if (retval) | |
3627 | return retval; | |
3628 | ||
3629 | set_user_nice(current, nice); | |
3630 | return 0; | |
3631 | } | |
3632 | ||
3633 | #endif | |
3634 | ||
3635 | /** | |
3636 | * task_prio - return the priority value of a given task. | |
3637 | * @p: the task in question. | |
3638 | * | |
3639 | * This is the priority value as seen by users in /proc. | |
3640 | * RT tasks are offset by -200. Normal tasks are centered | |
3641 | * around 0, value goes from -16 to +15. | |
3642 | */ | |
3643 | int task_prio(const task_t *p) | |
3644 | { | |
3645 | return p->prio - MAX_RT_PRIO; | |
3646 | } | |
3647 | ||
3648 | /** | |
3649 | * task_nice - return the nice value of a given task. | |
3650 | * @p: the task in question. | |
3651 | */ | |
3652 | int task_nice(const task_t *p) | |
3653 | { | |
3654 | return TASK_NICE(p); | |
3655 | } | |
1da177e4 | 3656 | EXPORT_SYMBOL_GPL(task_nice); |
1da177e4 LT |
3657 | |
3658 | /** | |
3659 | * idle_cpu - is a given cpu idle currently? | |
3660 | * @cpu: the processor in question. | |
3661 | */ | |
3662 | int idle_cpu(int cpu) | |
3663 | { | |
3664 | return cpu_curr(cpu) == cpu_rq(cpu)->idle; | |
3665 | } | |
3666 | ||
1da177e4 LT |
3667 | /** |
3668 | * idle_task - return the idle task for a given cpu. | |
3669 | * @cpu: the processor in question. | |
3670 | */ | |
3671 | task_t *idle_task(int cpu) | |
3672 | { | |
3673 | return cpu_rq(cpu)->idle; | |
3674 | } | |
3675 | ||
3676 | /** | |
3677 | * find_process_by_pid - find a process with a matching PID value. | |
3678 | * @pid: the pid in question. | |
3679 | */ | |
3680 | static inline task_t *find_process_by_pid(pid_t pid) | |
3681 | { | |
3682 | return pid ? find_task_by_pid(pid) : current; | |
3683 | } | |
3684 | ||
3685 | /* Actually do priority change: must hold rq lock. */ | |
3686 | static void __setscheduler(struct task_struct *p, int policy, int prio) | |
3687 | { | |
3688 | BUG_ON(p->array); | |
3689 | p->policy = policy; | |
3690 | p->rt_priority = prio; | |
3691 | if (policy != SCHED_NORMAL) | |
d46523ea | 3692 | p->prio = MAX_RT_PRIO-1 - p->rt_priority; |
1da177e4 LT |
3693 | else |
3694 | p->prio = p->static_prio; | |
3695 | } | |
3696 | ||
3697 | /** | |
3698 | * sched_setscheduler - change the scheduling policy and/or RT priority of | |
3699 | * a thread. | |
3700 | * @p: the task in question. | |
3701 | * @policy: new policy. | |
3702 | * @param: structure containing the new RT priority. | |
3703 | */ | |
95cdf3b7 IM |
3704 | int sched_setscheduler(struct task_struct *p, int policy, |
3705 | struct sched_param *param) | |
1da177e4 LT |
3706 | { |
3707 | int retval; | |
3708 | int oldprio, oldpolicy = -1; | |
3709 | prio_array_t *array; | |
3710 | unsigned long flags; | |
3711 | runqueue_t *rq; | |
3712 | ||
3713 | recheck: | |
3714 | /* double check policy once rq lock held */ | |
3715 | if (policy < 0) | |
3716 | policy = oldpolicy = p->policy; | |
3717 | else if (policy != SCHED_FIFO && policy != SCHED_RR && | |
3718 | policy != SCHED_NORMAL) | |
3719 | return -EINVAL; | |
3720 | /* | |
3721 | * Valid priorities for SCHED_FIFO and SCHED_RR are | |
3722 | * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL is 0. | |
3723 | */ | |
3724 | if (param->sched_priority < 0 || | |
95cdf3b7 | 3725 | (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || |
d46523ea | 3726 | (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) |
1da177e4 LT |
3727 | return -EINVAL; |
3728 | if ((policy == SCHED_NORMAL) != (param->sched_priority == 0)) | |
3729 | return -EINVAL; | |
3730 | ||
37e4ab3f OC |
3731 | /* |
3732 | * Allow unprivileged RT tasks to decrease priority: | |
3733 | */ | |
3734 | if (!capable(CAP_SYS_NICE)) { | |
3735 | /* can't change policy */ | |
18586e72 AS |
3736 | if (policy != p->policy && |
3737 | !p->signal->rlim[RLIMIT_RTPRIO].rlim_cur) | |
37e4ab3f OC |
3738 | return -EPERM; |
3739 | /* can't increase priority */ | |
3740 | if (policy != SCHED_NORMAL && | |
3741 | param->sched_priority > p->rt_priority && | |
3742 | param->sched_priority > | |
3743 | p->signal->rlim[RLIMIT_RTPRIO].rlim_cur) | |
3744 | return -EPERM; | |
3745 | /* can't change other user's priorities */ | |
3746 | if ((current->euid != p->euid) && | |
3747 | (current->euid != p->uid)) | |
3748 | return -EPERM; | |
3749 | } | |
1da177e4 LT |
3750 | |
3751 | retval = security_task_setscheduler(p, policy, param); | |
3752 | if (retval) | |
3753 | return retval; | |
3754 | /* | |
3755 | * To be able to change p->policy safely, the apropriate | |
3756 | * runqueue lock must be held. | |
3757 | */ | |
3758 | rq = task_rq_lock(p, &flags); | |
3759 | /* recheck policy now with rq lock held */ | |
3760 | if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { | |
3761 | policy = oldpolicy = -1; | |
3762 | task_rq_unlock(rq, &flags); | |
3763 | goto recheck; | |
3764 | } | |
3765 | array = p->array; | |
3766 | if (array) | |
3767 | deactivate_task(p, rq); | |
3768 | oldprio = p->prio; | |
3769 | __setscheduler(p, policy, param->sched_priority); | |
3770 | if (array) { | |
3771 | __activate_task(p, rq); | |
3772 | /* | |
3773 | * Reschedule if we are currently running on this runqueue and | |
3774 | * our priority decreased, or if we are not currently running on | |
3775 | * this runqueue and our priority is higher than the current's | |
3776 | */ | |
3777 | if (task_running(rq, p)) { | |
3778 | if (p->prio > oldprio) | |
3779 | resched_task(rq->curr); | |
3780 | } else if (TASK_PREEMPTS_CURR(p, rq)) | |
3781 | resched_task(rq->curr); | |
3782 | } | |
3783 | task_rq_unlock(rq, &flags); | |
3784 | return 0; | |
3785 | } | |
3786 | EXPORT_SYMBOL_GPL(sched_setscheduler); | |
3787 | ||
95cdf3b7 IM |
3788 | static int |
3789 | do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) | |
1da177e4 LT |
3790 | { |
3791 | int retval; | |
3792 | struct sched_param lparam; | |
3793 | struct task_struct *p; | |
3794 | ||
3795 | if (!param || pid < 0) | |
3796 | return -EINVAL; | |
3797 | if (copy_from_user(&lparam, param, sizeof(struct sched_param))) | |
3798 | return -EFAULT; | |
3799 | read_lock_irq(&tasklist_lock); | |
3800 | p = find_process_by_pid(pid); | |
3801 | if (!p) { | |
3802 | read_unlock_irq(&tasklist_lock); | |
3803 | return -ESRCH; | |
3804 | } | |
3805 | retval = sched_setscheduler(p, policy, &lparam); | |
3806 | read_unlock_irq(&tasklist_lock); | |
3807 | return retval; | |
3808 | } | |
3809 | ||
3810 | /** | |
3811 | * sys_sched_setscheduler - set/change the scheduler policy and RT priority | |
3812 | * @pid: the pid in question. | |
3813 | * @policy: new policy. | |
3814 | * @param: structure containing the new RT priority. | |
3815 | */ | |
3816 | asmlinkage long sys_sched_setscheduler(pid_t pid, int policy, | |
3817 | struct sched_param __user *param) | |
3818 | { | |
3819 | return do_sched_setscheduler(pid, policy, param); | |
3820 | } | |
3821 | ||
3822 | /** | |
3823 | * sys_sched_setparam - set/change the RT priority of a thread | |
3824 | * @pid: the pid in question. | |
3825 | * @param: structure containing the new RT priority. | |
3826 | */ | |
3827 | asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param) | |
3828 | { | |
3829 | return do_sched_setscheduler(pid, -1, param); | |
3830 | } | |
3831 | ||
3832 | /** | |
3833 | * sys_sched_getscheduler - get the policy (scheduling class) of a thread | |
3834 | * @pid: the pid in question. | |
3835 | */ | |
3836 | asmlinkage long sys_sched_getscheduler(pid_t pid) | |
3837 | { | |
3838 | int retval = -EINVAL; | |
3839 | task_t *p; | |
3840 | ||
3841 | if (pid < 0) | |
3842 | goto out_nounlock; | |
3843 | ||
3844 | retval = -ESRCH; | |
3845 | read_lock(&tasklist_lock); | |
3846 | p = find_process_by_pid(pid); | |
3847 | if (p) { | |
3848 | retval = security_task_getscheduler(p); | |
3849 | if (!retval) | |
3850 | retval = p->policy; | |
3851 | } | |
3852 | read_unlock(&tasklist_lock); | |
3853 | ||
3854 | out_nounlock: | |
3855 | return retval; | |
3856 | } | |
3857 | ||
3858 | /** | |
3859 | * sys_sched_getscheduler - get the RT priority of a thread | |
3860 | * @pid: the pid in question. | |
3861 | * @param: structure containing the RT priority. | |
3862 | */ | |
3863 | asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param) | |
3864 | { | |
3865 | struct sched_param lp; | |
3866 | int retval = -EINVAL; | |
3867 | task_t *p; | |
3868 | ||
3869 | if (!param || pid < 0) | |
3870 | goto out_nounlock; | |
3871 | ||
3872 | read_lock(&tasklist_lock); | |
3873 | p = find_process_by_pid(pid); | |
3874 | retval = -ESRCH; | |
3875 | if (!p) | |
3876 | goto out_unlock; | |
3877 | ||
3878 | retval = security_task_getscheduler(p); | |
3879 | if (retval) | |
3880 | goto out_unlock; | |
3881 | ||
3882 | lp.sched_priority = p->rt_priority; | |
3883 | read_unlock(&tasklist_lock); | |
3884 | ||
3885 | /* | |
3886 | * This one might sleep, we cannot do it with a spinlock held ... | |
3887 | */ | |
3888 | retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; | |
3889 | ||
3890 | out_nounlock: | |
3891 | return retval; | |
3892 | ||
3893 | out_unlock: | |
3894 | read_unlock(&tasklist_lock); | |
3895 | return retval; | |
3896 | } | |
3897 | ||
3898 | long sched_setaffinity(pid_t pid, cpumask_t new_mask) | |
3899 | { | |
3900 | task_t *p; | |
3901 | int retval; | |
3902 | cpumask_t cpus_allowed; | |
3903 | ||
3904 | lock_cpu_hotplug(); | |
3905 | read_lock(&tasklist_lock); | |
3906 | ||
3907 | p = find_process_by_pid(pid); | |
3908 | if (!p) { | |
3909 | read_unlock(&tasklist_lock); | |
3910 | unlock_cpu_hotplug(); | |
3911 | return -ESRCH; | |
3912 | } | |
3913 | ||
3914 | /* | |
3915 | * It is not safe to call set_cpus_allowed with the | |
3916 | * tasklist_lock held. We will bump the task_struct's | |
3917 | * usage count and then drop tasklist_lock. | |
3918 | */ | |
3919 | get_task_struct(p); | |
3920 | read_unlock(&tasklist_lock); | |
3921 | ||
3922 | retval = -EPERM; | |
3923 | if ((current->euid != p->euid) && (current->euid != p->uid) && | |
3924 | !capable(CAP_SYS_NICE)) | |
3925 | goto out_unlock; | |
3926 | ||
3927 | cpus_allowed = cpuset_cpus_allowed(p); | |
3928 | cpus_and(new_mask, new_mask, cpus_allowed); | |
3929 | retval = set_cpus_allowed(p, new_mask); | |
3930 | ||
3931 | out_unlock: | |
3932 | put_task_struct(p); | |
3933 | unlock_cpu_hotplug(); | |
3934 | return retval; | |
3935 | } | |
3936 | ||
3937 | static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, | |
3938 | cpumask_t *new_mask) | |
3939 | { | |
3940 | if (len < sizeof(cpumask_t)) { | |
3941 | memset(new_mask, 0, sizeof(cpumask_t)); | |
3942 | } else if (len > sizeof(cpumask_t)) { | |
3943 | len = sizeof(cpumask_t); | |
3944 | } | |
3945 | return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; | |
3946 | } | |
3947 | ||
3948 | /** | |
3949 | * sys_sched_setaffinity - set the cpu affinity of a process | |
3950 | * @pid: pid of the process | |
3951 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
3952 | * @user_mask_ptr: user-space pointer to the new cpu mask | |
3953 | */ | |
3954 | asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len, | |
3955 | unsigned long __user *user_mask_ptr) | |
3956 | { | |
3957 | cpumask_t new_mask; | |
3958 | int retval; | |
3959 | ||
3960 | retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask); | |
3961 | if (retval) | |
3962 | return retval; | |
3963 | ||
3964 | return sched_setaffinity(pid, new_mask); | |
3965 | } | |
3966 | ||
3967 | /* | |
3968 | * Represents all cpu's present in the system | |
3969 | * In systems capable of hotplug, this map could dynamically grow | |
3970 | * as new cpu's are detected in the system via any platform specific | |
3971 | * method, such as ACPI for e.g. | |
3972 | */ | |
3973 | ||
3974 | cpumask_t cpu_present_map; | |
3975 | EXPORT_SYMBOL(cpu_present_map); | |
3976 | ||
3977 | #ifndef CONFIG_SMP | |
3978 | cpumask_t cpu_online_map = CPU_MASK_ALL; | |
3979 | cpumask_t cpu_possible_map = CPU_MASK_ALL; | |
3980 | #endif | |
3981 | ||
3982 | long sched_getaffinity(pid_t pid, cpumask_t *mask) | |
3983 | { | |
3984 | int retval; | |
3985 | task_t *p; | |
3986 | ||
3987 | lock_cpu_hotplug(); | |
3988 | read_lock(&tasklist_lock); | |
3989 | ||
3990 | retval = -ESRCH; | |
3991 | p = find_process_by_pid(pid); | |
3992 | if (!p) | |
3993 | goto out_unlock; | |
3994 | ||
3995 | retval = 0; | |
3996 | cpus_and(*mask, p->cpus_allowed, cpu_possible_map); | |
3997 | ||
3998 | out_unlock: | |
3999 | read_unlock(&tasklist_lock); | |
4000 | unlock_cpu_hotplug(); | |
4001 | if (retval) | |
4002 | return retval; | |
4003 | ||
4004 | return 0; | |
4005 | } | |
4006 | ||
4007 | /** | |
4008 | * sys_sched_getaffinity - get the cpu affinity of a process | |
4009 | * @pid: pid of the process | |
4010 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
4011 | * @user_mask_ptr: user-space pointer to hold the current cpu mask | |
4012 | */ | |
4013 | asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len, | |
4014 | unsigned long __user *user_mask_ptr) | |
4015 | { | |
4016 | int ret; | |
4017 | cpumask_t mask; | |
4018 | ||
4019 | if (len < sizeof(cpumask_t)) | |
4020 | return -EINVAL; | |
4021 | ||
4022 | ret = sched_getaffinity(pid, &mask); | |
4023 | if (ret < 0) | |
4024 | return ret; | |
4025 | ||
4026 | if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t))) | |
4027 | return -EFAULT; | |
4028 | ||
4029 | return sizeof(cpumask_t); | |
4030 | } | |
4031 | ||
4032 | /** | |
4033 | * sys_sched_yield - yield the current processor to other threads. | |
4034 | * | |
4035 | * this function yields the current CPU by moving the calling thread | |
4036 | * to the expired array. If there are no other threads running on this | |
4037 | * CPU then this function will return. | |
4038 | */ | |
4039 | asmlinkage long sys_sched_yield(void) | |
4040 | { | |
4041 | runqueue_t *rq = this_rq_lock(); | |
4042 | prio_array_t *array = current->array; | |
4043 | prio_array_t *target = rq->expired; | |
4044 | ||
4045 | schedstat_inc(rq, yld_cnt); | |
4046 | /* | |
4047 | * We implement yielding by moving the task into the expired | |
4048 | * queue. | |
4049 | * | |
4050 | * (special rule: RT tasks will just roundrobin in the active | |
4051 | * array.) | |
4052 | */ | |
4053 | if (rt_task(current)) | |
4054 | target = rq->active; | |
4055 | ||
5927ad78 | 4056 | if (array->nr_active == 1) { |
1da177e4 LT |
4057 | schedstat_inc(rq, yld_act_empty); |
4058 | if (!rq->expired->nr_active) | |
4059 | schedstat_inc(rq, yld_both_empty); | |
4060 | } else if (!rq->expired->nr_active) | |
4061 | schedstat_inc(rq, yld_exp_empty); | |
4062 | ||
4063 | if (array != target) { | |
4064 | dequeue_task(current, array); | |
4065 | enqueue_task(current, target); | |
4066 | } else | |
4067 | /* | |
4068 | * requeue_task is cheaper so perform that if possible. | |
4069 | */ | |
4070 | requeue_task(current, array); | |
4071 | ||
4072 | /* | |
4073 | * Since we are going to call schedule() anyway, there's | |
4074 | * no need to preempt or enable interrupts: | |
4075 | */ | |
4076 | __release(rq->lock); | |
4077 | _raw_spin_unlock(&rq->lock); | |
4078 | preempt_enable_no_resched(); | |
4079 | ||
4080 | schedule(); | |
4081 | ||
4082 | return 0; | |
4083 | } | |
4084 | ||
4085 | static inline void __cond_resched(void) | |
4086 | { | |
5bbcfd90 IM |
4087 | /* |
4088 | * The BKS might be reacquired before we have dropped | |
4089 | * PREEMPT_ACTIVE, which could trigger a second | |
4090 | * cond_resched() call. | |
4091 | */ | |
4092 | if (unlikely(preempt_count())) | |
4093 | return; | |
1da177e4 LT |
4094 | do { |
4095 | add_preempt_count(PREEMPT_ACTIVE); | |
4096 | schedule(); | |
4097 | sub_preempt_count(PREEMPT_ACTIVE); | |
4098 | } while (need_resched()); | |
4099 | } | |
4100 | ||
4101 | int __sched cond_resched(void) | |
4102 | { | |
4103 | if (need_resched()) { | |
4104 | __cond_resched(); | |
4105 | return 1; | |
4106 | } | |
4107 | return 0; | |
4108 | } | |
4109 | ||
4110 | EXPORT_SYMBOL(cond_resched); | |
4111 | ||
4112 | /* | |
4113 | * cond_resched_lock() - if a reschedule is pending, drop the given lock, | |
4114 | * call schedule, and on return reacquire the lock. | |
4115 | * | |
4116 | * This works OK both with and without CONFIG_PREEMPT. We do strange low-level | |
4117 | * operations here to prevent schedule() from being called twice (once via | |
4118 | * spin_unlock(), once by hand). | |
4119 | */ | |
95cdf3b7 | 4120 | int cond_resched_lock(spinlock_t *lock) |
1da177e4 | 4121 | { |
6df3cecb JK |
4122 | int ret = 0; |
4123 | ||
1da177e4 LT |
4124 | if (need_lockbreak(lock)) { |
4125 | spin_unlock(lock); | |
4126 | cpu_relax(); | |
6df3cecb | 4127 | ret = 1; |
1da177e4 LT |
4128 | spin_lock(lock); |
4129 | } | |
4130 | if (need_resched()) { | |
4131 | _raw_spin_unlock(lock); | |
4132 | preempt_enable_no_resched(); | |
4133 | __cond_resched(); | |
6df3cecb | 4134 | ret = 1; |
1da177e4 | 4135 | spin_lock(lock); |
1da177e4 | 4136 | } |
6df3cecb | 4137 | return ret; |
1da177e4 LT |
4138 | } |
4139 | ||
4140 | EXPORT_SYMBOL(cond_resched_lock); | |
4141 | ||
4142 | int __sched cond_resched_softirq(void) | |
4143 | { | |
4144 | BUG_ON(!in_softirq()); | |
4145 | ||
4146 | if (need_resched()) { | |
4147 | __local_bh_enable(); | |
4148 | __cond_resched(); | |
4149 | local_bh_disable(); | |
4150 | return 1; | |
4151 | } | |
4152 | return 0; | |
4153 | } | |
4154 | ||
4155 | EXPORT_SYMBOL(cond_resched_softirq); | |
4156 | ||
4157 | ||
4158 | /** | |
4159 | * yield - yield the current processor to other threads. | |
4160 | * | |
4161 | * this is a shortcut for kernel-space yielding - it marks the | |
4162 | * thread runnable and calls sys_sched_yield(). | |
4163 | */ | |
4164 | void __sched yield(void) | |
4165 | { | |
4166 | set_current_state(TASK_RUNNING); | |
4167 | sys_sched_yield(); | |
4168 | } | |
4169 | ||
4170 | EXPORT_SYMBOL(yield); | |
4171 | ||
4172 | /* | |
4173 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so | |
4174 | * that process accounting knows that this is a task in IO wait state. | |
4175 | * | |
4176 | * But don't do that if it is a deliberate, throttling IO wait (this task | |
4177 | * has set its backing_dev_info: the queue against which it should throttle) | |
4178 | */ | |
4179 | void __sched io_schedule(void) | |
4180 | { | |
39c715b7 | 4181 | struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id()); |
1da177e4 LT |
4182 | |
4183 | atomic_inc(&rq->nr_iowait); | |
4184 | schedule(); | |
4185 | atomic_dec(&rq->nr_iowait); | |
4186 | } | |
4187 | ||
4188 | EXPORT_SYMBOL(io_schedule); | |
4189 | ||
4190 | long __sched io_schedule_timeout(long timeout) | |
4191 | { | |
39c715b7 | 4192 | struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id()); |
1da177e4 LT |
4193 | long ret; |
4194 | ||
4195 | atomic_inc(&rq->nr_iowait); | |
4196 | ret = schedule_timeout(timeout); | |
4197 | atomic_dec(&rq->nr_iowait); | |
4198 | return ret; | |
4199 | } | |
4200 | ||
4201 | /** | |
4202 | * sys_sched_get_priority_max - return maximum RT priority. | |
4203 | * @policy: scheduling class. | |
4204 | * | |
4205 | * this syscall returns the maximum rt_priority that can be used | |
4206 | * by a given scheduling class. | |
4207 | */ | |
4208 | asmlinkage long sys_sched_get_priority_max(int policy) | |
4209 | { | |
4210 | int ret = -EINVAL; | |
4211 | ||
4212 | switch (policy) { | |
4213 | case SCHED_FIFO: | |
4214 | case SCHED_RR: | |
4215 | ret = MAX_USER_RT_PRIO-1; | |
4216 | break; | |
4217 | case SCHED_NORMAL: | |
4218 | ret = 0; | |
4219 | break; | |
4220 | } | |
4221 | return ret; | |
4222 | } | |
4223 | ||
4224 | /** | |
4225 | * sys_sched_get_priority_min - return minimum RT priority. | |
4226 | * @policy: scheduling class. | |
4227 | * | |
4228 | * this syscall returns the minimum rt_priority that can be used | |
4229 | * by a given scheduling class. | |
4230 | */ | |
4231 | asmlinkage long sys_sched_get_priority_min(int policy) | |
4232 | { | |
4233 | int ret = -EINVAL; | |
4234 | ||
4235 | switch (policy) { | |
4236 | case SCHED_FIFO: | |
4237 | case SCHED_RR: | |
4238 | ret = 1; | |
4239 | break; | |
4240 | case SCHED_NORMAL: | |
4241 | ret = 0; | |
4242 | } | |
4243 | return ret; | |
4244 | } | |
4245 | ||
4246 | /** | |
4247 | * sys_sched_rr_get_interval - return the default timeslice of a process. | |
4248 | * @pid: pid of the process. | |
4249 | * @interval: userspace pointer to the timeslice value. | |
4250 | * | |
4251 | * this syscall writes the default timeslice value of a given process | |
4252 | * into the user-space timespec buffer. A value of '0' means infinity. | |
4253 | */ | |
4254 | asmlinkage | |
4255 | long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval) | |
4256 | { | |
4257 | int retval = -EINVAL; | |
4258 | struct timespec t; | |
4259 | task_t *p; | |
4260 | ||
4261 | if (pid < 0) | |
4262 | goto out_nounlock; | |
4263 | ||
4264 | retval = -ESRCH; | |
4265 | read_lock(&tasklist_lock); | |
4266 | p = find_process_by_pid(pid); | |
4267 | if (!p) | |
4268 | goto out_unlock; | |
4269 | ||
4270 | retval = security_task_getscheduler(p); | |
4271 | if (retval) | |
4272 | goto out_unlock; | |
4273 | ||
4274 | jiffies_to_timespec(p->policy & SCHED_FIFO ? | |
4275 | 0 : task_timeslice(p), &t); | |
4276 | read_unlock(&tasklist_lock); | |
4277 | retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; | |
4278 | out_nounlock: | |
4279 | return retval; | |
4280 | out_unlock: | |
4281 | read_unlock(&tasklist_lock); | |
4282 | return retval; | |
4283 | } | |
4284 | ||
4285 | static inline struct task_struct *eldest_child(struct task_struct *p) | |
4286 | { | |
4287 | if (list_empty(&p->children)) return NULL; | |
4288 | return list_entry(p->children.next,struct task_struct,sibling); | |
4289 | } | |
4290 | ||
4291 | static inline struct task_struct *older_sibling(struct task_struct *p) | |
4292 | { | |
4293 | if (p->sibling.prev==&p->parent->children) return NULL; | |
4294 | return list_entry(p->sibling.prev,struct task_struct,sibling); | |
4295 | } | |
4296 | ||
4297 | static inline struct task_struct *younger_sibling(struct task_struct *p) | |
4298 | { | |
4299 | if (p->sibling.next==&p->parent->children) return NULL; | |
4300 | return list_entry(p->sibling.next,struct task_struct,sibling); | |
4301 | } | |
4302 | ||
95cdf3b7 | 4303 | static void show_task(task_t *p) |
1da177e4 LT |
4304 | { |
4305 | task_t *relative; | |
4306 | unsigned state; | |
4307 | unsigned long free = 0; | |
4308 | static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" }; | |
4309 | ||
4310 | printk("%-13.13s ", p->comm); | |
4311 | state = p->state ? __ffs(p->state) + 1 : 0; | |
4312 | if (state < ARRAY_SIZE(stat_nam)) | |
4313 | printk(stat_nam[state]); | |
4314 | else | |
4315 | printk("?"); | |
4316 | #if (BITS_PER_LONG == 32) | |
4317 | if (state == TASK_RUNNING) | |
4318 | printk(" running "); | |
4319 | else | |
4320 | printk(" %08lX ", thread_saved_pc(p)); | |
4321 | #else | |
4322 | if (state == TASK_RUNNING) | |
4323 | printk(" running task "); | |
4324 | else | |
4325 | printk(" %016lx ", thread_saved_pc(p)); | |
4326 | #endif | |
4327 | #ifdef CONFIG_DEBUG_STACK_USAGE | |
4328 | { | |
95cdf3b7 | 4329 | unsigned long *n = (unsigned long *) (p->thread_info+1); |
1da177e4 LT |
4330 | while (!*n) |
4331 | n++; | |
4332 | free = (unsigned long) n - (unsigned long)(p->thread_info+1); | |
4333 | } | |
4334 | #endif | |
4335 | printk("%5lu %5d %6d ", free, p->pid, p->parent->pid); | |
4336 | if ((relative = eldest_child(p))) | |
4337 | printk("%5d ", relative->pid); | |
4338 | else | |
4339 | printk(" "); | |
4340 | if ((relative = younger_sibling(p))) | |
4341 | printk("%7d", relative->pid); | |
4342 | else | |
4343 | printk(" "); | |
4344 | if ((relative = older_sibling(p))) | |
4345 | printk(" %5d", relative->pid); | |
4346 | else | |
4347 | printk(" "); | |
4348 | if (!p->mm) | |
4349 | printk(" (L-TLB)\n"); | |
4350 | else | |
4351 | printk(" (NOTLB)\n"); | |
4352 | ||
4353 | if (state != TASK_RUNNING) | |
4354 | show_stack(p, NULL); | |
4355 | } | |
4356 | ||
4357 | void show_state(void) | |
4358 | { | |
4359 | task_t *g, *p; | |
4360 | ||
4361 | #if (BITS_PER_LONG == 32) | |
4362 | printk("\n" | |
4363 | " sibling\n"); | |
4364 | printk(" task PC pid father child younger older\n"); | |
4365 | #else | |
4366 | printk("\n" | |
4367 | " sibling\n"); | |
4368 | printk(" task PC pid father child younger older\n"); | |
4369 | #endif | |
4370 | read_lock(&tasklist_lock); | |
4371 | do_each_thread(g, p) { | |
4372 | /* | |
4373 | * reset the NMI-timeout, listing all files on a slow | |
4374 | * console might take alot of time: | |
4375 | */ | |
4376 | touch_nmi_watchdog(); | |
4377 | show_task(p); | |
4378 | } while_each_thread(g, p); | |
4379 | ||
4380 | read_unlock(&tasklist_lock); | |
4381 | } | |
4382 | ||
f340c0d1 IM |
4383 | /** |
4384 | * init_idle - set up an idle thread for a given CPU | |
4385 | * @idle: task in question | |
4386 | * @cpu: cpu the idle task belongs to | |
4387 | * | |
4388 | * NOTE: this function does not set the idle thread's NEED_RESCHED | |
4389 | * flag, to make booting more robust. | |
4390 | */ | |
1da177e4 LT |
4391 | void __devinit init_idle(task_t *idle, int cpu) |
4392 | { | |
4393 | runqueue_t *rq = cpu_rq(cpu); | |
4394 | unsigned long flags; | |
4395 | ||
4396 | idle->sleep_avg = 0; | |
4397 | idle->array = NULL; | |
4398 | idle->prio = MAX_PRIO; | |
4399 | idle->state = TASK_RUNNING; | |
4400 | idle->cpus_allowed = cpumask_of_cpu(cpu); | |
4401 | set_task_cpu(idle, cpu); | |
4402 | ||
4403 | spin_lock_irqsave(&rq->lock, flags); | |
4404 | rq->curr = rq->idle = idle; | |
4866cde0 NP |
4405 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
4406 | idle->oncpu = 1; | |
4407 | #endif | |
1da177e4 LT |
4408 | spin_unlock_irqrestore(&rq->lock, flags); |
4409 | ||
4410 | /* Set the preempt count _outside_ the spinlocks! */ | |
4411 | #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL) | |
4412 | idle->thread_info->preempt_count = (idle->lock_depth >= 0); | |
4413 | #else | |
4414 | idle->thread_info->preempt_count = 0; | |
4415 | #endif | |
4416 | } | |
4417 | ||
4418 | /* | |
4419 | * In a system that switches off the HZ timer nohz_cpu_mask | |
4420 | * indicates which cpus entered this state. This is used | |
4421 | * in the rcu update to wait only for active cpus. For system | |
4422 | * which do not switch off the HZ timer nohz_cpu_mask should | |
4423 | * always be CPU_MASK_NONE. | |
4424 | */ | |
4425 | cpumask_t nohz_cpu_mask = CPU_MASK_NONE; | |
4426 | ||
4427 | #ifdef CONFIG_SMP | |
4428 | /* | |
4429 | * This is how migration works: | |
4430 | * | |
4431 | * 1) we queue a migration_req_t structure in the source CPU's | |
4432 | * runqueue and wake up that CPU's migration thread. | |
4433 | * 2) we down() the locked semaphore => thread blocks. | |
4434 | * 3) migration thread wakes up (implicitly it forces the migrated | |
4435 | * thread off the CPU) | |
4436 | * 4) it gets the migration request and checks whether the migrated | |
4437 | * task is still in the wrong runqueue. | |
4438 | * 5) if it's in the wrong runqueue then the migration thread removes | |
4439 | * it and puts it into the right queue. | |
4440 | * 6) migration thread up()s the semaphore. | |
4441 | * 7) we wake up and the migration is done. | |
4442 | */ | |
4443 | ||
4444 | /* | |
4445 | * Change a given task's CPU affinity. Migrate the thread to a | |
4446 | * proper CPU and schedule it away if the CPU it's executing on | |
4447 | * is removed from the allowed bitmask. | |
4448 | * | |
4449 | * NOTE: the caller must have a valid reference to the task, the | |
4450 | * task must not exit() & deallocate itself prematurely. The | |
4451 | * call is not atomic; no spinlocks may be held. | |
4452 | */ | |
4453 | int set_cpus_allowed(task_t *p, cpumask_t new_mask) | |
4454 | { | |
4455 | unsigned long flags; | |
4456 | int ret = 0; | |
4457 | migration_req_t req; | |
4458 | runqueue_t *rq; | |
4459 | ||
4460 | rq = task_rq_lock(p, &flags); | |
4461 | if (!cpus_intersects(new_mask, cpu_online_map)) { | |
4462 | ret = -EINVAL; | |
4463 | goto out; | |
4464 | } | |
4465 | ||
4466 | p->cpus_allowed = new_mask; | |
4467 | /* Can the task run on the task's current CPU? If so, we're done */ | |
4468 | if (cpu_isset(task_cpu(p), new_mask)) | |
4469 | goto out; | |
4470 | ||
4471 | if (migrate_task(p, any_online_cpu(new_mask), &req)) { | |
4472 | /* Need help from migration thread: drop lock and wait. */ | |
4473 | task_rq_unlock(rq, &flags); | |
4474 | wake_up_process(rq->migration_thread); | |
4475 | wait_for_completion(&req.done); | |
4476 | tlb_migrate_finish(p->mm); | |
4477 | return 0; | |
4478 | } | |
4479 | out: | |
4480 | task_rq_unlock(rq, &flags); | |
4481 | return ret; | |
4482 | } | |
4483 | ||
4484 | EXPORT_SYMBOL_GPL(set_cpus_allowed); | |
4485 | ||
4486 | /* | |
4487 | * Move (not current) task off this cpu, onto dest cpu. We're doing | |
4488 | * this because either it can't run here any more (set_cpus_allowed() | |
4489 | * away from this CPU, or CPU going down), or because we're | |
4490 | * attempting to rebalance this task on exec (sched_exec). | |
4491 | * | |
4492 | * So we race with normal scheduler movements, but that's OK, as long | |
4493 | * as the task is no longer on this CPU. | |
4494 | */ | |
4495 | static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) | |
4496 | { | |
4497 | runqueue_t *rq_dest, *rq_src; | |
4498 | ||
4499 | if (unlikely(cpu_is_offline(dest_cpu))) | |
4500 | return; | |
4501 | ||
4502 | rq_src = cpu_rq(src_cpu); | |
4503 | rq_dest = cpu_rq(dest_cpu); | |
4504 | ||
4505 | double_rq_lock(rq_src, rq_dest); | |
4506 | /* Already moved. */ | |
4507 | if (task_cpu(p) != src_cpu) | |
4508 | goto out; | |
4509 | /* Affinity changed (again). */ | |
4510 | if (!cpu_isset(dest_cpu, p->cpus_allowed)) | |
4511 | goto out; | |
4512 | ||
4513 | set_task_cpu(p, dest_cpu); | |
4514 | if (p->array) { | |
4515 | /* | |
4516 | * Sync timestamp with rq_dest's before activating. | |
4517 | * The same thing could be achieved by doing this step | |
4518 | * afterwards, and pretending it was a local activate. | |
4519 | * This way is cleaner and logically correct. | |
4520 | */ | |
4521 | p->timestamp = p->timestamp - rq_src->timestamp_last_tick | |
4522 | + rq_dest->timestamp_last_tick; | |
4523 | deactivate_task(p, rq_src); | |
4524 | activate_task(p, rq_dest, 0); | |
4525 | if (TASK_PREEMPTS_CURR(p, rq_dest)) | |
4526 | resched_task(rq_dest->curr); | |
4527 | } | |
4528 | ||
4529 | out: | |
4530 | double_rq_unlock(rq_src, rq_dest); | |
4531 | } | |
4532 | ||
4533 | /* | |
4534 | * migration_thread - this is a highprio system thread that performs | |
4535 | * thread migration by bumping thread off CPU then 'pushing' onto | |
4536 | * another runqueue. | |
4537 | */ | |
95cdf3b7 | 4538 | static int migration_thread(void *data) |
1da177e4 LT |
4539 | { |
4540 | runqueue_t *rq; | |
4541 | int cpu = (long)data; | |
4542 | ||
4543 | rq = cpu_rq(cpu); | |
4544 | BUG_ON(rq->migration_thread != current); | |
4545 | ||
4546 | set_current_state(TASK_INTERRUPTIBLE); | |
4547 | while (!kthread_should_stop()) { | |
4548 | struct list_head *head; | |
4549 | migration_req_t *req; | |
4550 | ||
3e1d1d28 | 4551 | try_to_freeze(); |
1da177e4 LT |
4552 | |
4553 | spin_lock_irq(&rq->lock); | |
4554 | ||
4555 | if (cpu_is_offline(cpu)) { | |
4556 | spin_unlock_irq(&rq->lock); | |
4557 | goto wait_to_die; | |
4558 | } | |
4559 | ||
4560 | if (rq->active_balance) { | |
4561 | active_load_balance(rq, cpu); | |
4562 | rq->active_balance = 0; | |
4563 | } | |
4564 | ||
4565 | head = &rq->migration_queue; | |
4566 | ||
4567 | if (list_empty(head)) { | |
4568 | spin_unlock_irq(&rq->lock); | |
4569 | schedule(); | |
4570 | set_current_state(TASK_INTERRUPTIBLE); | |
4571 | continue; | |
4572 | } | |
4573 | req = list_entry(head->next, migration_req_t, list); | |
4574 | list_del_init(head->next); | |
4575 | ||
674311d5 NP |
4576 | spin_unlock(&rq->lock); |
4577 | __migrate_task(req->task, cpu, req->dest_cpu); | |
4578 | local_irq_enable(); | |
1da177e4 LT |
4579 | |
4580 | complete(&req->done); | |
4581 | } | |
4582 | __set_current_state(TASK_RUNNING); | |
4583 | return 0; | |
4584 | ||
4585 | wait_to_die: | |
4586 | /* Wait for kthread_stop */ | |
4587 | set_current_state(TASK_INTERRUPTIBLE); | |
4588 | while (!kthread_should_stop()) { | |
4589 | schedule(); | |
4590 | set_current_state(TASK_INTERRUPTIBLE); | |
4591 | } | |
4592 | __set_current_state(TASK_RUNNING); | |
4593 | return 0; | |
4594 | } | |
4595 | ||
4596 | #ifdef CONFIG_HOTPLUG_CPU | |
4597 | /* Figure out where task on dead CPU should go, use force if neccessary. */ | |
4598 | static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk) | |
4599 | { | |
4600 | int dest_cpu; | |
4601 | cpumask_t mask; | |
4602 | ||
4603 | /* On same node? */ | |
4604 | mask = node_to_cpumask(cpu_to_node(dead_cpu)); | |
4605 | cpus_and(mask, mask, tsk->cpus_allowed); | |
4606 | dest_cpu = any_online_cpu(mask); | |
4607 | ||
4608 | /* On any allowed CPU? */ | |
4609 | if (dest_cpu == NR_CPUS) | |
4610 | dest_cpu = any_online_cpu(tsk->cpus_allowed); | |
4611 | ||
4612 | /* No more Mr. Nice Guy. */ | |
4613 | if (dest_cpu == NR_CPUS) { | |
b39c4fab | 4614 | cpus_setall(tsk->cpus_allowed); |
1da177e4 LT |
4615 | dest_cpu = any_online_cpu(tsk->cpus_allowed); |
4616 | ||
4617 | /* | |
4618 | * Don't tell them about moving exiting tasks or | |
4619 | * kernel threads (both mm NULL), since they never | |
4620 | * leave kernel. | |
4621 | */ | |
4622 | if (tsk->mm && printk_ratelimit()) | |
4623 | printk(KERN_INFO "process %d (%s) no " | |
4624 | "longer affine to cpu%d\n", | |
4625 | tsk->pid, tsk->comm, dead_cpu); | |
4626 | } | |
4627 | __migrate_task(tsk, dead_cpu, dest_cpu); | |
4628 | } | |
4629 | ||
4630 | /* | |
4631 | * While a dead CPU has no uninterruptible tasks queued at this point, | |
4632 | * it might still have a nonzero ->nr_uninterruptible counter, because | |
4633 | * for performance reasons the counter is not stricly tracking tasks to | |
4634 | * their home CPUs. So we just add the counter to another CPU's counter, | |
4635 | * to keep the global sum constant after CPU-down: | |
4636 | */ | |
4637 | static void migrate_nr_uninterruptible(runqueue_t *rq_src) | |
4638 | { | |
4639 | runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL)); | |
4640 | unsigned long flags; | |
4641 | ||
4642 | local_irq_save(flags); | |
4643 | double_rq_lock(rq_src, rq_dest); | |
4644 | rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible; | |
4645 | rq_src->nr_uninterruptible = 0; | |
4646 | double_rq_unlock(rq_src, rq_dest); | |
4647 | local_irq_restore(flags); | |
4648 | } | |
4649 | ||
4650 | /* Run through task list and migrate tasks from the dead cpu. */ | |
4651 | static void migrate_live_tasks(int src_cpu) | |
4652 | { | |
4653 | struct task_struct *tsk, *t; | |
4654 | ||
4655 | write_lock_irq(&tasklist_lock); | |
4656 | ||
4657 | do_each_thread(t, tsk) { | |
4658 | if (tsk == current) | |
4659 | continue; | |
4660 | ||
4661 | if (task_cpu(tsk) == src_cpu) | |
4662 | move_task_off_dead_cpu(src_cpu, tsk); | |
4663 | } while_each_thread(t, tsk); | |
4664 | ||
4665 | write_unlock_irq(&tasklist_lock); | |
4666 | } | |
4667 | ||
4668 | /* Schedules idle task to be the next runnable task on current CPU. | |
4669 | * It does so by boosting its priority to highest possible and adding it to | |
4670 | * the _front_ of runqueue. Used by CPU offline code. | |
4671 | */ | |
4672 | void sched_idle_next(void) | |
4673 | { | |
4674 | int cpu = smp_processor_id(); | |
4675 | runqueue_t *rq = this_rq(); | |
4676 | struct task_struct *p = rq->idle; | |
4677 | unsigned long flags; | |
4678 | ||
4679 | /* cpu has to be offline */ | |
4680 | BUG_ON(cpu_online(cpu)); | |
4681 | ||
4682 | /* Strictly not necessary since rest of the CPUs are stopped by now | |
4683 | * and interrupts disabled on current cpu. | |
4684 | */ | |
4685 | spin_lock_irqsave(&rq->lock, flags); | |
4686 | ||
4687 | __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1); | |
4688 | /* Add idle task to _front_ of it's priority queue */ | |
4689 | __activate_idle_task(p, rq); | |
4690 | ||
4691 | spin_unlock_irqrestore(&rq->lock, flags); | |
4692 | } | |
4693 | ||
4694 | /* Ensures that the idle task is using init_mm right before its cpu goes | |
4695 | * offline. | |
4696 | */ | |
4697 | void idle_task_exit(void) | |
4698 | { | |
4699 | struct mm_struct *mm = current->active_mm; | |
4700 | ||
4701 | BUG_ON(cpu_online(smp_processor_id())); | |
4702 | ||
4703 | if (mm != &init_mm) | |
4704 | switch_mm(mm, &init_mm, current); | |
4705 | mmdrop(mm); | |
4706 | } | |
4707 | ||
4708 | static void migrate_dead(unsigned int dead_cpu, task_t *tsk) | |
4709 | { | |
4710 | struct runqueue *rq = cpu_rq(dead_cpu); | |
4711 | ||
4712 | /* Must be exiting, otherwise would be on tasklist. */ | |
4713 | BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD); | |
4714 | ||
4715 | /* Cannot have done final schedule yet: would have vanished. */ | |
4716 | BUG_ON(tsk->flags & PF_DEAD); | |
4717 | ||
4718 | get_task_struct(tsk); | |
4719 | ||
4720 | /* | |
4721 | * Drop lock around migration; if someone else moves it, | |
4722 | * that's OK. No task can be added to this CPU, so iteration is | |
4723 | * fine. | |
4724 | */ | |
4725 | spin_unlock_irq(&rq->lock); | |
4726 | move_task_off_dead_cpu(dead_cpu, tsk); | |
4727 | spin_lock_irq(&rq->lock); | |
4728 | ||
4729 | put_task_struct(tsk); | |
4730 | } | |
4731 | ||
4732 | /* release_task() removes task from tasklist, so we won't find dead tasks. */ | |
4733 | static void migrate_dead_tasks(unsigned int dead_cpu) | |
4734 | { | |
4735 | unsigned arr, i; | |
4736 | struct runqueue *rq = cpu_rq(dead_cpu); | |
4737 | ||
4738 | for (arr = 0; arr < 2; arr++) { | |
4739 | for (i = 0; i < MAX_PRIO; i++) { | |
4740 | struct list_head *list = &rq->arrays[arr].queue[i]; | |
4741 | while (!list_empty(list)) | |
4742 | migrate_dead(dead_cpu, | |
4743 | list_entry(list->next, task_t, | |
4744 | run_list)); | |
4745 | } | |
4746 | } | |
4747 | } | |
4748 | #endif /* CONFIG_HOTPLUG_CPU */ | |
4749 | ||
4750 | /* | |
4751 | * migration_call - callback that gets triggered when a CPU is added. | |
4752 | * Here we can start up the necessary migration thread for the new CPU. | |
4753 | */ | |
4754 | static int migration_call(struct notifier_block *nfb, unsigned long action, | |
4755 | void *hcpu) | |
4756 | { | |
4757 | int cpu = (long)hcpu; | |
4758 | struct task_struct *p; | |
4759 | struct runqueue *rq; | |
4760 | unsigned long flags; | |
4761 | ||
4762 | switch (action) { | |
4763 | case CPU_UP_PREPARE: | |
4764 | p = kthread_create(migration_thread, hcpu, "migration/%d",cpu); | |
4765 | if (IS_ERR(p)) | |
4766 | return NOTIFY_BAD; | |
4767 | p->flags |= PF_NOFREEZE; | |
4768 | kthread_bind(p, cpu); | |
4769 | /* Must be high prio: stop_machine expects to yield to it. */ | |
4770 | rq = task_rq_lock(p, &flags); | |
4771 | __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1); | |
4772 | task_rq_unlock(rq, &flags); | |
4773 | cpu_rq(cpu)->migration_thread = p; | |
4774 | break; | |
4775 | case CPU_ONLINE: | |
4776 | /* Strictly unneccessary, as first user will wake it. */ | |
4777 | wake_up_process(cpu_rq(cpu)->migration_thread); | |
4778 | break; | |
4779 | #ifdef CONFIG_HOTPLUG_CPU | |
4780 | case CPU_UP_CANCELED: | |
4781 | /* Unbind it from offline cpu so it can run. Fall thru. */ | |
a4c4af7c HC |
4782 | kthread_bind(cpu_rq(cpu)->migration_thread, |
4783 | any_online_cpu(cpu_online_map)); | |
1da177e4 LT |
4784 | kthread_stop(cpu_rq(cpu)->migration_thread); |
4785 | cpu_rq(cpu)->migration_thread = NULL; | |
4786 | break; | |
4787 | case CPU_DEAD: | |
4788 | migrate_live_tasks(cpu); | |
4789 | rq = cpu_rq(cpu); | |
4790 | kthread_stop(rq->migration_thread); | |
4791 | rq->migration_thread = NULL; | |
4792 | /* Idle task back to normal (off runqueue, low prio) */ | |
4793 | rq = task_rq_lock(rq->idle, &flags); | |
4794 | deactivate_task(rq->idle, rq); | |
4795 | rq->idle->static_prio = MAX_PRIO; | |
4796 | __setscheduler(rq->idle, SCHED_NORMAL, 0); | |
4797 | migrate_dead_tasks(cpu); | |
4798 | task_rq_unlock(rq, &flags); | |
4799 | migrate_nr_uninterruptible(rq); | |
4800 | BUG_ON(rq->nr_running != 0); | |
4801 | ||
4802 | /* No need to migrate the tasks: it was best-effort if | |
4803 | * they didn't do lock_cpu_hotplug(). Just wake up | |
4804 | * the requestors. */ | |
4805 | spin_lock_irq(&rq->lock); | |
4806 | while (!list_empty(&rq->migration_queue)) { | |
4807 | migration_req_t *req; | |
4808 | req = list_entry(rq->migration_queue.next, | |
4809 | migration_req_t, list); | |
1da177e4 LT |
4810 | list_del_init(&req->list); |
4811 | complete(&req->done); | |
4812 | } | |
4813 | spin_unlock_irq(&rq->lock); | |
4814 | break; | |
4815 | #endif | |
4816 | } | |
4817 | return NOTIFY_OK; | |
4818 | } | |
4819 | ||
4820 | /* Register at highest priority so that task migration (migrate_all_tasks) | |
4821 | * happens before everything else. | |
4822 | */ | |
4823 | static struct notifier_block __devinitdata migration_notifier = { | |
4824 | .notifier_call = migration_call, | |
4825 | .priority = 10 | |
4826 | }; | |
4827 | ||
4828 | int __init migration_init(void) | |
4829 | { | |
4830 | void *cpu = (void *)(long)smp_processor_id(); | |
4831 | /* Start one for boot CPU. */ | |
4832 | migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); | |
4833 | migration_call(&migration_notifier, CPU_ONLINE, cpu); | |
4834 | register_cpu_notifier(&migration_notifier); | |
4835 | return 0; | |
4836 | } | |
4837 | #endif | |
4838 | ||
4839 | #ifdef CONFIG_SMP | |
1a20ff27 | 4840 | #undef SCHED_DOMAIN_DEBUG |
1da177e4 LT |
4841 | #ifdef SCHED_DOMAIN_DEBUG |
4842 | static void sched_domain_debug(struct sched_domain *sd, int cpu) | |
4843 | { | |
4844 | int level = 0; | |
4845 | ||
41c7ce9a NP |
4846 | if (!sd) { |
4847 | printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); | |
4848 | return; | |
4849 | } | |
4850 | ||
1da177e4 LT |
4851 | printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); |
4852 | ||
4853 | do { | |
4854 | int i; | |
4855 | char str[NR_CPUS]; | |
4856 | struct sched_group *group = sd->groups; | |
4857 | cpumask_t groupmask; | |
4858 | ||
4859 | cpumask_scnprintf(str, NR_CPUS, sd->span); | |
4860 | cpus_clear(groupmask); | |
4861 | ||
4862 | printk(KERN_DEBUG); | |
4863 | for (i = 0; i < level + 1; i++) | |
4864 | printk(" "); | |
4865 | printk("domain %d: ", level); | |
4866 | ||
4867 | if (!(sd->flags & SD_LOAD_BALANCE)) { | |
4868 | printk("does not load-balance\n"); | |
4869 | if (sd->parent) | |
4870 | printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent"); | |
4871 | break; | |
4872 | } | |
4873 | ||
4874 | printk("span %s\n", str); | |
4875 | ||
4876 | if (!cpu_isset(cpu, sd->span)) | |
4877 | printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu); | |
4878 | if (!cpu_isset(cpu, group->cpumask)) | |
4879 | printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu); | |
4880 | ||
4881 | printk(KERN_DEBUG); | |
4882 | for (i = 0; i < level + 2; i++) | |
4883 | printk(" "); | |
4884 | printk("groups:"); | |
4885 | do { | |
4886 | if (!group) { | |
4887 | printk("\n"); | |
4888 | printk(KERN_ERR "ERROR: group is NULL\n"); | |
4889 | break; | |
4890 | } | |
4891 | ||
4892 | if (!group->cpu_power) { | |
4893 | printk("\n"); | |
4894 | printk(KERN_ERR "ERROR: domain->cpu_power not set\n"); | |
4895 | } | |
4896 | ||
4897 | if (!cpus_weight(group->cpumask)) { | |
4898 | printk("\n"); | |
4899 | printk(KERN_ERR "ERROR: empty group\n"); | |
4900 | } | |
4901 | ||
4902 | if (cpus_intersects(groupmask, group->cpumask)) { | |
4903 | printk("\n"); | |
4904 | printk(KERN_ERR "ERROR: repeated CPUs\n"); | |
4905 | } | |
4906 | ||
4907 | cpus_or(groupmask, groupmask, group->cpumask); | |
4908 | ||
4909 | cpumask_scnprintf(str, NR_CPUS, group->cpumask); | |
4910 | printk(" %s", str); | |
4911 | ||
4912 | group = group->next; | |
4913 | } while (group != sd->groups); | |
4914 | printk("\n"); | |
4915 | ||
4916 | if (!cpus_equal(sd->span, groupmask)) | |
4917 | printk(KERN_ERR "ERROR: groups don't span domain->span\n"); | |
4918 | ||
4919 | level++; | |
4920 | sd = sd->parent; | |
4921 | ||
4922 | if (sd) { | |
4923 | if (!cpus_subset(groupmask, sd->span)) | |
4924 | printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n"); | |
4925 | } | |
4926 | ||
4927 | } while (sd); | |
4928 | } | |
4929 | #else | |
4930 | #define sched_domain_debug(sd, cpu) {} | |
4931 | #endif | |
4932 | ||
1a20ff27 | 4933 | static int sd_degenerate(struct sched_domain *sd) |
245af2c7 SS |
4934 | { |
4935 | if (cpus_weight(sd->span) == 1) | |
4936 | return 1; | |
4937 | ||
4938 | /* Following flags need at least 2 groups */ | |
4939 | if (sd->flags & (SD_LOAD_BALANCE | | |
4940 | SD_BALANCE_NEWIDLE | | |
4941 | SD_BALANCE_FORK | | |
4942 | SD_BALANCE_EXEC)) { | |
4943 | if (sd->groups != sd->groups->next) | |
4944 | return 0; | |
4945 | } | |
4946 | ||
4947 | /* Following flags don't use groups */ | |
4948 | if (sd->flags & (SD_WAKE_IDLE | | |
4949 | SD_WAKE_AFFINE | | |
4950 | SD_WAKE_BALANCE)) | |
4951 | return 0; | |
4952 | ||
4953 | return 1; | |
4954 | } | |
4955 | ||
1a20ff27 | 4956 | static int sd_parent_degenerate(struct sched_domain *sd, |
245af2c7 SS |
4957 | struct sched_domain *parent) |
4958 | { | |
4959 | unsigned long cflags = sd->flags, pflags = parent->flags; | |
4960 | ||
4961 | if (sd_degenerate(parent)) | |
4962 | return 1; | |
4963 | ||
4964 | if (!cpus_equal(sd->span, parent->span)) | |
4965 | return 0; | |
4966 | ||
4967 | /* Does parent contain flags not in child? */ | |
4968 | /* WAKE_BALANCE is a subset of WAKE_AFFINE */ | |
4969 | if (cflags & SD_WAKE_AFFINE) | |
4970 | pflags &= ~SD_WAKE_BALANCE; | |
4971 | /* Flags needing groups don't count if only 1 group in parent */ | |
4972 | if (parent->groups == parent->groups->next) { | |
4973 | pflags &= ~(SD_LOAD_BALANCE | | |
4974 | SD_BALANCE_NEWIDLE | | |
4975 | SD_BALANCE_FORK | | |
4976 | SD_BALANCE_EXEC); | |
4977 | } | |
4978 | if (~cflags & pflags) | |
4979 | return 0; | |
4980 | ||
4981 | return 1; | |
4982 | } | |
4983 | ||
1da177e4 LT |
4984 | /* |
4985 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must | |
4986 | * hold the hotplug lock. | |
4987 | */ | |
9c1cfda2 | 4988 | static void cpu_attach_domain(struct sched_domain *sd, int cpu) |
1da177e4 | 4989 | { |
1da177e4 | 4990 | runqueue_t *rq = cpu_rq(cpu); |
245af2c7 SS |
4991 | struct sched_domain *tmp; |
4992 | ||
4993 | /* Remove the sched domains which do not contribute to scheduling. */ | |
4994 | for (tmp = sd; tmp; tmp = tmp->parent) { | |
4995 | struct sched_domain *parent = tmp->parent; | |
4996 | if (!parent) | |
4997 | break; | |
4998 | if (sd_parent_degenerate(tmp, parent)) | |
4999 | tmp->parent = parent->parent; | |
5000 | } | |
5001 | ||
5002 | if (sd && sd_degenerate(sd)) | |
5003 | sd = sd->parent; | |
1da177e4 LT |
5004 | |
5005 | sched_domain_debug(sd, cpu); | |
5006 | ||
674311d5 | 5007 | rcu_assign_pointer(rq->sd, sd); |
1da177e4 LT |
5008 | } |
5009 | ||
5010 | /* cpus with isolated domains */ | |
9c1cfda2 | 5011 | static cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE; |
1da177e4 LT |
5012 | |
5013 | /* Setup the mask of cpus configured for isolated domains */ | |
5014 | static int __init isolated_cpu_setup(char *str) | |
5015 | { | |
5016 | int ints[NR_CPUS], i; | |
5017 | ||
5018 | str = get_options(str, ARRAY_SIZE(ints), ints); | |
5019 | cpus_clear(cpu_isolated_map); | |
5020 | for (i = 1; i <= ints[0]; i++) | |
5021 | if (ints[i] < NR_CPUS) | |
5022 | cpu_set(ints[i], cpu_isolated_map); | |
5023 | return 1; | |
5024 | } | |
5025 | ||
5026 | __setup ("isolcpus=", isolated_cpu_setup); | |
5027 | ||
5028 | /* | |
5029 | * init_sched_build_groups takes an array of groups, the cpumask we wish | |
5030 | * to span, and a pointer to a function which identifies what group a CPU | |
5031 | * belongs to. The return value of group_fn must be a valid index into the | |
5032 | * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we | |
5033 | * keep track of groups covered with a cpumask_t). | |
5034 | * | |
5035 | * init_sched_build_groups will build a circular linked list of the groups | |
5036 | * covered by the given span, and will set each group's ->cpumask correctly, | |
5037 | * and ->cpu_power to 0. | |
5038 | */ | |
9c1cfda2 JH |
5039 | static void init_sched_build_groups(struct sched_group groups[], cpumask_t span, |
5040 | int (*group_fn)(int cpu)) | |
1da177e4 LT |
5041 | { |
5042 | struct sched_group *first = NULL, *last = NULL; | |
5043 | cpumask_t covered = CPU_MASK_NONE; | |
5044 | int i; | |
5045 | ||
5046 | for_each_cpu_mask(i, span) { | |
5047 | int group = group_fn(i); | |
5048 | struct sched_group *sg = &groups[group]; | |
5049 | int j; | |
5050 | ||
5051 | if (cpu_isset(i, covered)) | |
5052 | continue; | |
5053 | ||
5054 | sg->cpumask = CPU_MASK_NONE; | |
5055 | sg->cpu_power = 0; | |
5056 | ||
5057 | for_each_cpu_mask(j, span) { | |
5058 | if (group_fn(j) != group) | |
5059 | continue; | |
5060 | ||
5061 | cpu_set(j, covered); | |
5062 | cpu_set(j, sg->cpumask); | |
5063 | } | |
5064 | if (!first) | |
5065 | first = sg; | |
5066 | if (last) | |
5067 | last->next = sg; | |
5068 | last = sg; | |
5069 | } | |
5070 | last->next = first; | |
5071 | } | |
5072 | ||
9c1cfda2 | 5073 | #define SD_NODES_PER_DOMAIN 16 |
1da177e4 | 5074 | |
9c1cfda2 JH |
5075 | #ifdef CONFIG_NUMA |
5076 | /** | |
5077 | * find_next_best_node - find the next node to include in a sched_domain | |
5078 | * @node: node whose sched_domain we're building | |
5079 | * @used_nodes: nodes already in the sched_domain | |
5080 | * | |
5081 | * Find the next node to include in a given scheduling domain. Simply | |
5082 | * finds the closest node not already in the @used_nodes map. | |
5083 | * | |
5084 | * Should use nodemask_t. | |
5085 | */ | |
5086 | static int find_next_best_node(int node, unsigned long *used_nodes) | |
5087 | { | |
5088 | int i, n, val, min_val, best_node = 0; | |
5089 | ||
5090 | min_val = INT_MAX; | |
5091 | ||
5092 | for (i = 0; i < MAX_NUMNODES; i++) { | |
5093 | /* Start at @node */ | |
5094 | n = (node + i) % MAX_NUMNODES; | |
5095 | ||
5096 | if (!nr_cpus_node(n)) | |
5097 | continue; | |
5098 | ||
5099 | /* Skip already used nodes */ | |
5100 | if (test_bit(n, used_nodes)) | |
5101 | continue; | |
5102 | ||
5103 | /* Simple min distance search */ | |
5104 | val = node_distance(node, n); | |
5105 | ||
5106 | if (val < min_val) { | |
5107 | min_val = val; | |
5108 | best_node = n; | |
5109 | } | |
5110 | } | |
5111 | ||
5112 | set_bit(best_node, used_nodes); | |
5113 | return best_node; | |
5114 | } | |
5115 | ||
5116 | /** | |
5117 | * sched_domain_node_span - get a cpumask for a node's sched_domain | |
5118 | * @node: node whose cpumask we're constructing | |
5119 | * @size: number of nodes to include in this span | |
5120 | * | |
5121 | * Given a node, construct a good cpumask for its sched_domain to span. It | |
5122 | * should be one that prevents unnecessary balancing, but also spreads tasks | |
5123 | * out optimally. | |
5124 | */ | |
5125 | static cpumask_t sched_domain_node_span(int node) | |
5126 | { | |
5127 | int i; | |
5128 | cpumask_t span, nodemask; | |
5129 | DECLARE_BITMAP(used_nodes, MAX_NUMNODES); | |
5130 | ||
5131 | cpus_clear(span); | |
5132 | bitmap_zero(used_nodes, MAX_NUMNODES); | |
5133 | ||
5134 | nodemask = node_to_cpumask(node); | |
5135 | cpus_or(span, span, nodemask); | |
5136 | set_bit(node, used_nodes); | |
5137 | ||
5138 | for (i = 1; i < SD_NODES_PER_DOMAIN; i++) { | |
5139 | int next_node = find_next_best_node(node, used_nodes); | |
5140 | nodemask = node_to_cpumask(next_node); | |
5141 | cpus_or(span, span, nodemask); | |
5142 | } | |
5143 | ||
5144 | return span; | |
5145 | } | |
5146 | #endif | |
5147 | ||
5148 | /* | |
5149 | * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we | |
5150 | * can switch it on easily if needed. | |
5151 | */ | |
1da177e4 LT |
5152 | #ifdef CONFIG_SCHED_SMT |
5153 | static DEFINE_PER_CPU(struct sched_domain, cpu_domains); | |
5154 | static struct sched_group sched_group_cpus[NR_CPUS]; | |
1a20ff27 | 5155 | static int cpu_to_cpu_group(int cpu) |
1da177e4 LT |
5156 | { |
5157 | return cpu; | |
5158 | } | |
5159 | #endif | |
5160 | ||
5161 | static DEFINE_PER_CPU(struct sched_domain, phys_domains); | |
5162 | static struct sched_group sched_group_phys[NR_CPUS]; | |
1a20ff27 | 5163 | static int cpu_to_phys_group(int cpu) |
1da177e4 LT |
5164 | { |
5165 | #ifdef CONFIG_SCHED_SMT | |
5166 | return first_cpu(cpu_sibling_map[cpu]); | |
5167 | #else | |
5168 | return cpu; | |
5169 | #endif | |
5170 | } | |
5171 | ||
5172 | #ifdef CONFIG_NUMA | |
1da177e4 | 5173 | /* |
9c1cfda2 JH |
5174 | * The init_sched_build_groups can't handle what we want to do with node |
5175 | * groups, so roll our own. Now each node has its own list of groups which | |
5176 | * gets dynamically allocated. | |
1da177e4 | 5177 | */ |
9c1cfda2 | 5178 | static DEFINE_PER_CPU(struct sched_domain, node_domains); |
d1b55138 | 5179 | static struct sched_group **sched_group_nodes_bycpu[NR_CPUS]; |
1da177e4 | 5180 | |
9c1cfda2 | 5181 | static DEFINE_PER_CPU(struct sched_domain, allnodes_domains); |
d1b55138 | 5182 | static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS]; |
9c1cfda2 JH |
5183 | |
5184 | static int cpu_to_allnodes_group(int cpu) | |
5185 | { | |
5186 | return cpu_to_node(cpu); | |
1da177e4 LT |
5187 | } |
5188 | #endif | |
5189 | ||
5190 | /* | |
1a20ff27 DG |
5191 | * Build sched domains for a given set of cpus and attach the sched domains |
5192 | * to the individual cpus | |
1da177e4 | 5193 | */ |
9c1cfda2 | 5194 | void build_sched_domains(const cpumask_t *cpu_map) |
1da177e4 LT |
5195 | { |
5196 | int i; | |
d1b55138 JH |
5197 | #ifdef CONFIG_NUMA |
5198 | struct sched_group **sched_group_nodes = NULL; | |
5199 | struct sched_group *sched_group_allnodes = NULL; | |
5200 | ||
5201 | /* | |
5202 | * Allocate the per-node list of sched groups | |
5203 | */ | |
5204 | sched_group_nodes = kmalloc(sizeof(struct sched_group*)*MAX_NUMNODES, | |
5205 | GFP_ATOMIC); | |
5206 | if (!sched_group_nodes) { | |
5207 | printk(KERN_WARNING "Can not alloc sched group node list\n"); | |
5208 | return; | |
5209 | } | |
5210 | sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes; | |
5211 | #endif | |
1da177e4 LT |
5212 | |
5213 | /* | |
1a20ff27 | 5214 | * Set up domains for cpus specified by the cpu_map. |
1da177e4 | 5215 | */ |
1a20ff27 | 5216 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 LT |
5217 | int group; |
5218 | struct sched_domain *sd = NULL, *p; | |
5219 | cpumask_t nodemask = node_to_cpumask(cpu_to_node(i)); | |
5220 | ||
1a20ff27 | 5221 | cpus_and(nodemask, nodemask, *cpu_map); |
1da177e4 LT |
5222 | |
5223 | #ifdef CONFIG_NUMA | |
d1b55138 | 5224 | if (cpus_weight(*cpu_map) |
9c1cfda2 | 5225 | > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) { |
d1b55138 JH |
5226 | if (!sched_group_allnodes) { |
5227 | sched_group_allnodes | |
5228 | = kmalloc(sizeof(struct sched_group) | |
5229 | * MAX_NUMNODES, | |
5230 | GFP_KERNEL); | |
5231 | if (!sched_group_allnodes) { | |
5232 | printk(KERN_WARNING | |
5233 | "Can not alloc allnodes sched group\n"); | |
5234 | break; | |
5235 | } | |
5236 | sched_group_allnodes_bycpu[i] | |
5237 | = sched_group_allnodes; | |
5238 | } | |
9c1cfda2 JH |
5239 | sd = &per_cpu(allnodes_domains, i); |
5240 | *sd = SD_ALLNODES_INIT; | |
5241 | sd->span = *cpu_map; | |
5242 | group = cpu_to_allnodes_group(i); | |
5243 | sd->groups = &sched_group_allnodes[group]; | |
5244 | p = sd; | |
5245 | } else | |
5246 | p = NULL; | |
5247 | ||
1da177e4 | 5248 | sd = &per_cpu(node_domains, i); |
1da177e4 | 5249 | *sd = SD_NODE_INIT; |
9c1cfda2 JH |
5250 | sd->span = sched_domain_node_span(cpu_to_node(i)); |
5251 | sd->parent = p; | |
5252 | cpus_and(sd->span, sd->span, *cpu_map); | |
1da177e4 LT |
5253 | #endif |
5254 | ||
5255 | p = sd; | |
5256 | sd = &per_cpu(phys_domains, i); | |
5257 | group = cpu_to_phys_group(i); | |
5258 | *sd = SD_CPU_INIT; | |
5259 | sd->span = nodemask; | |
5260 | sd->parent = p; | |
5261 | sd->groups = &sched_group_phys[group]; | |
5262 | ||
5263 | #ifdef CONFIG_SCHED_SMT | |
5264 | p = sd; | |
5265 | sd = &per_cpu(cpu_domains, i); | |
5266 | group = cpu_to_cpu_group(i); | |
5267 | *sd = SD_SIBLING_INIT; | |
5268 | sd->span = cpu_sibling_map[i]; | |
1a20ff27 | 5269 | cpus_and(sd->span, sd->span, *cpu_map); |
1da177e4 LT |
5270 | sd->parent = p; |
5271 | sd->groups = &sched_group_cpus[group]; | |
5272 | #endif | |
5273 | } | |
5274 | ||
5275 | #ifdef CONFIG_SCHED_SMT | |
5276 | /* Set up CPU (sibling) groups */ | |
9c1cfda2 | 5277 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 | 5278 | cpumask_t this_sibling_map = cpu_sibling_map[i]; |
1a20ff27 | 5279 | cpus_and(this_sibling_map, this_sibling_map, *cpu_map); |
1da177e4 LT |
5280 | if (i != first_cpu(this_sibling_map)) |
5281 | continue; | |
5282 | ||
5283 | init_sched_build_groups(sched_group_cpus, this_sibling_map, | |
5284 | &cpu_to_cpu_group); | |
5285 | } | |
5286 | #endif | |
5287 | ||
5288 | /* Set up physical groups */ | |
5289 | for (i = 0; i < MAX_NUMNODES; i++) { | |
5290 | cpumask_t nodemask = node_to_cpumask(i); | |
5291 | ||
1a20ff27 | 5292 | cpus_and(nodemask, nodemask, *cpu_map); |
1da177e4 LT |
5293 | if (cpus_empty(nodemask)) |
5294 | continue; | |
5295 | ||
5296 | init_sched_build_groups(sched_group_phys, nodemask, | |
5297 | &cpu_to_phys_group); | |
5298 | } | |
5299 | ||
5300 | #ifdef CONFIG_NUMA | |
5301 | /* Set up node groups */ | |
d1b55138 JH |
5302 | if (sched_group_allnodes) |
5303 | init_sched_build_groups(sched_group_allnodes, *cpu_map, | |
5304 | &cpu_to_allnodes_group); | |
9c1cfda2 JH |
5305 | |
5306 | for (i = 0; i < MAX_NUMNODES; i++) { | |
5307 | /* Set up node groups */ | |
5308 | struct sched_group *sg, *prev; | |
5309 | cpumask_t nodemask = node_to_cpumask(i); | |
5310 | cpumask_t domainspan; | |
5311 | cpumask_t covered = CPU_MASK_NONE; | |
5312 | int j; | |
5313 | ||
5314 | cpus_and(nodemask, nodemask, *cpu_map); | |
d1b55138 JH |
5315 | if (cpus_empty(nodemask)) { |
5316 | sched_group_nodes[i] = NULL; | |
9c1cfda2 | 5317 | continue; |
d1b55138 | 5318 | } |
9c1cfda2 JH |
5319 | |
5320 | domainspan = sched_domain_node_span(i); | |
5321 | cpus_and(domainspan, domainspan, *cpu_map); | |
5322 | ||
5323 | sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL); | |
5324 | sched_group_nodes[i] = sg; | |
5325 | for_each_cpu_mask(j, nodemask) { | |
5326 | struct sched_domain *sd; | |
5327 | sd = &per_cpu(node_domains, j); | |
5328 | sd->groups = sg; | |
5329 | if (sd->groups == NULL) { | |
5330 | /* Turn off balancing if we have no groups */ | |
5331 | sd->flags = 0; | |
5332 | } | |
5333 | } | |
5334 | if (!sg) { | |
5335 | printk(KERN_WARNING | |
5336 | "Can not alloc domain group for node %d\n", i); | |
5337 | continue; | |
5338 | } | |
5339 | sg->cpu_power = 0; | |
5340 | sg->cpumask = nodemask; | |
5341 | cpus_or(covered, covered, nodemask); | |
5342 | prev = sg; | |
5343 | ||
5344 | for (j = 0; j < MAX_NUMNODES; j++) { | |
5345 | cpumask_t tmp, notcovered; | |
5346 | int n = (i + j) % MAX_NUMNODES; | |
5347 | ||
5348 | cpus_complement(notcovered, covered); | |
5349 | cpus_and(tmp, notcovered, *cpu_map); | |
5350 | cpus_and(tmp, tmp, domainspan); | |
5351 | if (cpus_empty(tmp)) | |
5352 | break; | |
5353 | ||
5354 | nodemask = node_to_cpumask(n); | |
5355 | cpus_and(tmp, tmp, nodemask); | |
5356 | if (cpus_empty(tmp)) | |
5357 | continue; | |
5358 | ||
5359 | sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL); | |
5360 | if (!sg) { | |
5361 | printk(KERN_WARNING | |
5362 | "Can not alloc domain group for node %d\n", j); | |
5363 | break; | |
5364 | } | |
5365 | sg->cpu_power = 0; | |
5366 | sg->cpumask = tmp; | |
5367 | cpus_or(covered, covered, tmp); | |
5368 | prev->next = sg; | |
5369 | prev = sg; | |
5370 | } | |
5371 | prev->next = sched_group_nodes[i]; | |
5372 | } | |
1da177e4 LT |
5373 | #endif |
5374 | ||
5375 | /* Calculate CPU power for physical packages and nodes */ | |
1a20ff27 | 5376 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 LT |
5377 | int power; |
5378 | struct sched_domain *sd; | |
5379 | #ifdef CONFIG_SCHED_SMT | |
5380 | sd = &per_cpu(cpu_domains, i); | |
5381 | power = SCHED_LOAD_SCALE; | |
5382 | sd->groups->cpu_power = power; | |
5383 | #endif | |
5384 | ||
5385 | sd = &per_cpu(phys_domains, i); | |
5386 | power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE * | |
5387 | (cpus_weight(sd->groups->cpumask)-1) / 10; | |
5388 | sd->groups->cpu_power = power; | |
5389 | ||
5390 | #ifdef CONFIG_NUMA | |
9c1cfda2 JH |
5391 | sd = &per_cpu(allnodes_domains, i); |
5392 | if (sd->groups) { | |
5393 | power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE * | |
5394 | (cpus_weight(sd->groups->cpumask)-1) / 10; | |
5395 | sd->groups->cpu_power = power; | |
1da177e4 LT |
5396 | } |
5397 | #endif | |
5398 | } | |
5399 | ||
9c1cfda2 JH |
5400 | #ifdef CONFIG_NUMA |
5401 | for (i = 0; i < MAX_NUMNODES; i++) { | |
5402 | struct sched_group *sg = sched_group_nodes[i]; | |
5403 | int j; | |
5404 | ||
5405 | if (sg == NULL) | |
5406 | continue; | |
5407 | next_sg: | |
5408 | for_each_cpu_mask(j, sg->cpumask) { | |
5409 | struct sched_domain *sd; | |
5410 | int power; | |
5411 | ||
5412 | sd = &per_cpu(phys_domains, j); | |
5413 | if (j != first_cpu(sd->groups->cpumask)) { | |
5414 | /* | |
5415 | * Only add "power" once for each | |
5416 | * physical package. | |
5417 | */ | |
5418 | continue; | |
5419 | } | |
5420 | power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE * | |
5421 | (cpus_weight(sd->groups->cpumask)-1) / 10; | |
5422 | ||
5423 | sg->cpu_power += power; | |
5424 | } | |
5425 | sg = sg->next; | |
5426 | if (sg != sched_group_nodes[i]) | |
5427 | goto next_sg; | |
5428 | } | |
5429 | #endif | |
5430 | ||
1da177e4 | 5431 | /* Attach the domains */ |
1a20ff27 | 5432 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 LT |
5433 | struct sched_domain *sd; |
5434 | #ifdef CONFIG_SCHED_SMT | |
5435 | sd = &per_cpu(cpu_domains, i); | |
5436 | #else | |
5437 | sd = &per_cpu(phys_domains, i); | |
5438 | #endif | |
5439 | cpu_attach_domain(sd, i); | |
5440 | } | |
5441 | } | |
1a20ff27 DG |
5442 | /* |
5443 | * Set up scheduler domains and groups. Callers must hold the hotplug lock. | |
5444 | */ | |
9c1cfda2 | 5445 | static void arch_init_sched_domains(const cpumask_t *cpu_map) |
1a20ff27 DG |
5446 | { |
5447 | cpumask_t cpu_default_map; | |
1da177e4 | 5448 | |
1a20ff27 DG |
5449 | /* |
5450 | * Setup mask for cpus without special case scheduling requirements. | |
5451 | * For now this just excludes isolated cpus, but could be used to | |
5452 | * exclude other special cases in the future. | |
5453 | */ | |
5454 | cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map); | |
5455 | ||
5456 | build_sched_domains(&cpu_default_map); | |
5457 | } | |
5458 | ||
5459 | static void arch_destroy_sched_domains(const cpumask_t *cpu_map) | |
1da177e4 | 5460 | { |
9c1cfda2 JH |
5461 | #ifdef CONFIG_NUMA |
5462 | int i; | |
d1b55138 | 5463 | int cpu; |
1da177e4 | 5464 | |
d1b55138 JH |
5465 | for_each_cpu_mask(cpu, *cpu_map) { |
5466 | struct sched_group *sched_group_allnodes | |
5467 | = sched_group_allnodes_bycpu[cpu]; | |
5468 | struct sched_group **sched_group_nodes | |
5469 | = sched_group_nodes_bycpu[cpu]; | |
9c1cfda2 | 5470 | |
d1b55138 JH |
5471 | if (sched_group_allnodes) { |
5472 | kfree(sched_group_allnodes); | |
5473 | sched_group_allnodes_bycpu[cpu] = NULL; | |
5474 | } | |
5475 | ||
5476 | if (!sched_group_nodes) | |
9c1cfda2 | 5477 | continue; |
d1b55138 JH |
5478 | |
5479 | for (i = 0; i < MAX_NUMNODES; i++) { | |
5480 | cpumask_t nodemask = node_to_cpumask(i); | |
5481 | struct sched_group *oldsg, *sg = sched_group_nodes[i]; | |
5482 | ||
5483 | cpus_and(nodemask, nodemask, *cpu_map); | |
5484 | if (cpus_empty(nodemask)) | |
5485 | continue; | |
5486 | ||
5487 | if (sg == NULL) | |
5488 | continue; | |
5489 | sg = sg->next; | |
9c1cfda2 | 5490 | next_sg: |
d1b55138 JH |
5491 | oldsg = sg; |
5492 | sg = sg->next; | |
5493 | kfree(oldsg); | |
5494 | if (oldsg != sched_group_nodes[i]) | |
5495 | goto next_sg; | |
5496 | } | |
5497 | kfree(sched_group_nodes); | |
5498 | sched_group_nodes_bycpu[cpu] = NULL; | |
9c1cfda2 JH |
5499 | } |
5500 | #endif | |
5501 | } | |
1da177e4 | 5502 | |
1a20ff27 DG |
5503 | /* |
5504 | * Detach sched domains from a group of cpus specified in cpu_map | |
5505 | * These cpus will now be attached to the NULL domain | |
5506 | */ | |
5507 | static inline void detach_destroy_domains(const cpumask_t *cpu_map) | |
5508 | { | |
5509 | int i; | |
5510 | ||
5511 | for_each_cpu_mask(i, *cpu_map) | |
5512 | cpu_attach_domain(NULL, i); | |
5513 | synchronize_sched(); | |
5514 | arch_destroy_sched_domains(cpu_map); | |
5515 | } | |
5516 | ||
5517 | /* | |
5518 | * Partition sched domains as specified by the cpumasks below. | |
5519 | * This attaches all cpus from the cpumasks to the NULL domain, | |
5520 | * waits for a RCU quiescent period, recalculates sched | |
5521 | * domain information and then attaches them back to the | |
5522 | * correct sched domains | |
5523 | * Call with hotplug lock held | |
5524 | */ | |
5525 | void partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2) | |
5526 | { | |
5527 | cpumask_t change_map; | |
5528 | ||
5529 | cpus_and(*partition1, *partition1, cpu_online_map); | |
5530 | cpus_and(*partition2, *partition2, cpu_online_map); | |
5531 | cpus_or(change_map, *partition1, *partition2); | |
5532 | ||
5533 | /* Detach sched domains from all of the affected cpus */ | |
5534 | detach_destroy_domains(&change_map); | |
5535 | if (!cpus_empty(*partition1)) | |
5536 | build_sched_domains(partition1); | |
5537 | if (!cpus_empty(*partition2)) | |
5538 | build_sched_domains(partition2); | |
5539 | } | |
5540 | ||
1da177e4 LT |
5541 | #ifdef CONFIG_HOTPLUG_CPU |
5542 | /* | |
5543 | * Force a reinitialization of the sched domains hierarchy. The domains | |
5544 | * and groups cannot be updated in place without racing with the balancing | |
41c7ce9a | 5545 | * code, so we temporarily attach all running cpus to the NULL domain |
1da177e4 LT |
5546 | * which will prevent rebalancing while the sched domains are recalculated. |
5547 | */ | |
5548 | static int update_sched_domains(struct notifier_block *nfb, | |
5549 | unsigned long action, void *hcpu) | |
5550 | { | |
1da177e4 LT |
5551 | switch (action) { |
5552 | case CPU_UP_PREPARE: | |
5553 | case CPU_DOWN_PREPARE: | |
1a20ff27 | 5554 | detach_destroy_domains(&cpu_online_map); |
1da177e4 LT |
5555 | return NOTIFY_OK; |
5556 | ||
5557 | case CPU_UP_CANCELED: | |
5558 | case CPU_DOWN_FAILED: | |
5559 | case CPU_ONLINE: | |
5560 | case CPU_DEAD: | |
5561 | /* | |
5562 | * Fall through and re-initialise the domains. | |
5563 | */ | |
5564 | break; | |
5565 | default: | |
5566 | return NOTIFY_DONE; | |
5567 | } | |
5568 | ||
5569 | /* The hotplug lock is already held by cpu_up/cpu_down */ | |
1a20ff27 | 5570 | arch_init_sched_domains(&cpu_online_map); |
1da177e4 LT |
5571 | |
5572 | return NOTIFY_OK; | |
5573 | } | |
5574 | #endif | |
5575 | ||
5576 | void __init sched_init_smp(void) | |
5577 | { | |
5578 | lock_cpu_hotplug(); | |
1a20ff27 | 5579 | arch_init_sched_domains(&cpu_online_map); |
1da177e4 LT |
5580 | unlock_cpu_hotplug(); |
5581 | /* XXX: Theoretical race here - CPU may be hotplugged now */ | |
5582 | hotcpu_notifier(update_sched_domains, 0); | |
5583 | } | |
5584 | #else | |
5585 | void __init sched_init_smp(void) | |
5586 | { | |
5587 | } | |
5588 | #endif /* CONFIG_SMP */ | |
5589 | ||
5590 | int in_sched_functions(unsigned long addr) | |
5591 | { | |
5592 | /* Linker adds these: start and end of __sched functions */ | |
5593 | extern char __sched_text_start[], __sched_text_end[]; | |
5594 | return in_lock_functions(addr) || | |
5595 | (addr >= (unsigned long)__sched_text_start | |
5596 | && addr < (unsigned long)__sched_text_end); | |
5597 | } | |
5598 | ||
5599 | void __init sched_init(void) | |
5600 | { | |
5601 | runqueue_t *rq; | |
5602 | int i, j, k; | |
5603 | ||
5604 | for (i = 0; i < NR_CPUS; i++) { | |
5605 | prio_array_t *array; | |
5606 | ||
5607 | rq = cpu_rq(i); | |
5608 | spin_lock_init(&rq->lock); | |
7897986b | 5609 | rq->nr_running = 0; |
1da177e4 LT |
5610 | rq->active = rq->arrays; |
5611 | rq->expired = rq->arrays + 1; | |
5612 | rq->best_expired_prio = MAX_PRIO; | |
5613 | ||
5614 | #ifdef CONFIG_SMP | |
41c7ce9a | 5615 | rq->sd = NULL; |
7897986b NP |
5616 | for (j = 1; j < 3; j++) |
5617 | rq->cpu_load[j] = 0; | |
1da177e4 LT |
5618 | rq->active_balance = 0; |
5619 | rq->push_cpu = 0; | |
5620 | rq->migration_thread = NULL; | |
5621 | INIT_LIST_HEAD(&rq->migration_queue); | |
5622 | #endif | |
5623 | atomic_set(&rq->nr_iowait, 0); | |
5624 | ||
5625 | for (j = 0; j < 2; j++) { | |
5626 | array = rq->arrays + j; | |
5627 | for (k = 0; k < MAX_PRIO; k++) { | |
5628 | INIT_LIST_HEAD(array->queue + k); | |
5629 | __clear_bit(k, array->bitmap); | |
5630 | } | |
5631 | // delimiter for bitsearch | |
5632 | __set_bit(MAX_PRIO, array->bitmap); | |
5633 | } | |
5634 | } | |
5635 | ||
5636 | /* | |
5637 | * The boot idle thread does lazy MMU switching as well: | |
5638 | */ | |
5639 | atomic_inc(&init_mm.mm_count); | |
5640 | enter_lazy_tlb(&init_mm, current); | |
5641 | ||
5642 | /* | |
5643 | * Make us the idle thread. Technically, schedule() should not be | |
5644 | * called from this thread, however somewhere below it might be, | |
5645 | * but because we are the idle thread, we just pick up running again | |
5646 | * when this runqueue becomes "idle". | |
5647 | */ | |
5648 | init_idle(current, smp_processor_id()); | |
5649 | } | |
5650 | ||
5651 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP | |
5652 | void __might_sleep(char *file, int line) | |
5653 | { | |
5654 | #if defined(in_atomic) | |
5655 | static unsigned long prev_jiffy; /* ratelimiting */ | |
5656 | ||
5657 | if ((in_atomic() || irqs_disabled()) && | |
5658 | system_state == SYSTEM_RUNNING && !oops_in_progress) { | |
5659 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) | |
5660 | return; | |
5661 | prev_jiffy = jiffies; | |
5662 | printk(KERN_ERR "Debug: sleeping function called from invalid" | |
5663 | " context at %s:%d\n", file, line); | |
5664 | printk("in_atomic():%d, irqs_disabled():%d\n", | |
5665 | in_atomic(), irqs_disabled()); | |
5666 | dump_stack(); | |
5667 | } | |
5668 | #endif | |
5669 | } | |
5670 | EXPORT_SYMBOL(__might_sleep); | |
5671 | #endif | |
5672 | ||
5673 | #ifdef CONFIG_MAGIC_SYSRQ | |
5674 | void normalize_rt_tasks(void) | |
5675 | { | |
5676 | struct task_struct *p; | |
5677 | prio_array_t *array; | |
5678 | unsigned long flags; | |
5679 | runqueue_t *rq; | |
5680 | ||
5681 | read_lock_irq(&tasklist_lock); | |
5682 | for_each_process (p) { | |
5683 | if (!rt_task(p)) | |
5684 | continue; | |
5685 | ||
5686 | rq = task_rq_lock(p, &flags); | |
5687 | ||
5688 | array = p->array; | |
5689 | if (array) | |
5690 | deactivate_task(p, task_rq(p)); | |
5691 | __setscheduler(p, SCHED_NORMAL, 0); | |
5692 | if (array) { | |
5693 | __activate_task(p, task_rq(p)); | |
5694 | resched_task(rq->curr); | |
5695 | } | |
5696 | ||
5697 | task_rq_unlock(rq, &flags); | |
5698 | } | |
5699 | read_unlock_irq(&tasklist_lock); | |
5700 | } | |
5701 | ||
5702 | #endif /* CONFIG_MAGIC_SYSRQ */ | |
1df5c10a LT |
5703 | |
5704 | #ifdef CONFIG_IA64 | |
5705 | /* | |
5706 | * These functions are only useful for the IA64 MCA handling. | |
5707 | * | |
5708 | * They can only be called when the whole system has been | |
5709 | * stopped - every CPU needs to be quiescent, and no scheduling | |
5710 | * activity can take place. Using them for anything else would | |
5711 | * be a serious bug, and as a result, they aren't even visible | |
5712 | * under any other configuration. | |
5713 | */ | |
5714 | ||
5715 | /** | |
5716 | * curr_task - return the current task for a given cpu. | |
5717 | * @cpu: the processor in question. | |
5718 | * | |
5719 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
5720 | */ | |
5721 | task_t *curr_task(int cpu) | |
5722 | { | |
5723 | return cpu_curr(cpu); | |
5724 | } | |
5725 | ||
5726 | /** | |
5727 | * set_curr_task - set the current task for a given cpu. | |
5728 | * @cpu: the processor in question. | |
5729 | * @p: the task pointer to set. | |
5730 | * | |
5731 | * Description: This function must only be used when non-maskable interrupts | |
5732 | * are serviced on a separate stack. It allows the architecture to switch the | |
5733 | * notion of the current task on a cpu in a non-blocking manner. This function | |
5734 | * must be called with all CPU's synchronized, and interrupts disabled, the | |
5735 | * and caller must save the original value of the current task (see | |
5736 | * curr_task() above) and restore that value before reenabling interrupts and | |
5737 | * re-starting the system. | |
5738 | * | |
5739 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
5740 | */ | |
5741 | void set_curr_task(int cpu, task_t *p) | |
5742 | { | |
5743 | cpu_curr(cpu) = p; | |
5744 | } | |
5745 | ||
5746 | #endif |