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> | |
c59ede7b | 30 | #include <linux/capability.h> |
1da177e4 LT |
31 | #include <linux/completion.h> |
32 | #include <linux/kernel_stat.h> | |
33 | #include <linux/security.h> | |
34 | #include <linux/notifier.h> | |
35 | #include <linux/profile.h> | |
36 | #include <linux/suspend.h> | |
198e2f18 | 37 | #include <linux/vmalloc.h> |
1da177e4 LT |
38 | #include <linux/blkdev.h> |
39 | #include <linux/delay.h> | |
40 | #include <linux/smp.h> | |
41 | #include <linux/threads.h> | |
42 | #include <linux/timer.h> | |
43 | #include <linux/rcupdate.h> | |
44 | #include <linux/cpu.h> | |
45 | #include <linux/cpuset.h> | |
46 | #include <linux/percpu.h> | |
47 | #include <linux/kthread.h> | |
48 | #include <linux/seq_file.h> | |
49 | #include <linux/syscalls.h> | |
50 | #include <linux/times.h> | |
51 | #include <linux/acct.h> | |
c6fd91f0 | 52 | #include <linux/kprobes.h> |
1da177e4 LT |
53 | #include <asm/tlb.h> |
54 | ||
55 | #include <asm/unistd.h> | |
56 | ||
57 | /* | |
58 | * Convert user-nice values [ -20 ... 0 ... 19 ] | |
59 | * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ], | |
60 | * and back. | |
61 | */ | |
62 | #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20) | |
63 | #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20) | |
64 | #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio) | |
65 | ||
66 | /* | |
67 | * 'User priority' is the nice value converted to something we | |
68 | * can work with better when scaling various scheduler parameters, | |
69 | * it's a [ 0 ... 39 ] range. | |
70 | */ | |
71 | #define USER_PRIO(p) ((p)-MAX_RT_PRIO) | |
72 | #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio) | |
73 | #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO)) | |
74 | ||
75 | /* | |
76 | * Some helpers for converting nanosecond timing to jiffy resolution | |
77 | */ | |
78 | #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ)) | |
79 | #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ)) | |
80 | ||
81 | /* | |
82 | * These are the 'tuning knobs' of the scheduler: | |
83 | * | |
84 | * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger), | |
85 | * default timeslice is 100 msecs, maximum timeslice is 800 msecs. | |
86 | * Timeslices get refilled after they expire. | |
87 | */ | |
88 | #define MIN_TIMESLICE max(5 * HZ / 1000, 1) | |
89 | #define DEF_TIMESLICE (100 * HZ / 1000) | |
90 | #define ON_RUNQUEUE_WEIGHT 30 | |
91 | #define CHILD_PENALTY 95 | |
92 | #define PARENT_PENALTY 100 | |
93 | #define EXIT_WEIGHT 3 | |
94 | #define PRIO_BONUS_RATIO 25 | |
95 | #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100) | |
96 | #define INTERACTIVE_DELTA 2 | |
97 | #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS) | |
98 | #define STARVATION_LIMIT (MAX_SLEEP_AVG) | |
99 | #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG)) | |
100 | ||
101 | /* | |
102 | * If a task is 'interactive' then we reinsert it in the active | |
103 | * array after it has expired its current timeslice. (it will not | |
104 | * continue to run immediately, it will still roundrobin with | |
105 | * other interactive tasks.) | |
106 | * | |
107 | * This part scales the interactivity limit depending on niceness. | |
108 | * | |
109 | * We scale it linearly, offset by the INTERACTIVE_DELTA delta. | |
110 | * Here are a few examples of different nice levels: | |
111 | * | |
112 | * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0] | |
113 | * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0] | |
114 | * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0] | |
115 | * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0] | |
116 | * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0] | |
117 | * | |
118 | * (the X axis represents the possible -5 ... 0 ... +5 dynamic | |
119 | * priority range a task can explore, a value of '1' means the | |
120 | * task is rated interactive.) | |
121 | * | |
122 | * Ie. nice +19 tasks can never get 'interactive' enough to be | |
123 | * reinserted into the active array. And only heavily CPU-hog nice -20 | |
124 | * tasks will be expired. Default nice 0 tasks are somewhere between, | |
125 | * it takes some effort for them to get interactive, but it's not | |
126 | * too hard. | |
127 | */ | |
128 | ||
129 | #define CURRENT_BONUS(p) \ | |
130 | (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \ | |
131 | MAX_SLEEP_AVG) | |
132 | ||
133 | #define GRANULARITY (10 * HZ / 1000 ? : 1) | |
134 | ||
135 | #ifdef CONFIG_SMP | |
136 | #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \ | |
137 | (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \ | |
138 | num_online_cpus()) | |
139 | #else | |
140 | #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \ | |
141 | (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1))) | |
142 | #endif | |
143 | ||
144 | #define SCALE(v1,v1_max,v2_max) \ | |
145 | (v1) * (v2_max) / (v1_max) | |
146 | ||
147 | #define DELTA(p) \ | |
013d3868 MA |
148 | (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \ |
149 | INTERACTIVE_DELTA) | |
1da177e4 LT |
150 | |
151 | #define TASK_INTERACTIVE(p) \ | |
152 | ((p)->prio <= (p)->static_prio - DELTA(p)) | |
153 | ||
154 | #define INTERACTIVE_SLEEP(p) \ | |
155 | (JIFFIES_TO_NS(MAX_SLEEP_AVG * \ | |
156 | (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1)) | |
157 | ||
158 | #define TASK_PREEMPTS_CURR(p, rq) \ | |
159 | ((p)->prio < (rq)->curr->prio) | |
160 | ||
161 | /* | |
162 | * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ] | |
163 | * to time slice values: [800ms ... 100ms ... 5ms] | |
164 | * | |
165 | * The higher a thread's priority, the bigger timeslices | |
166 | * it gets during one round of execution. But even the lowest | |
167 | * priority thread gets MIN_TIMESLICE worth of execution time. | |
168 | */ | |
169 | ||
170 | #define SCALE_PRIO(x, prio) \ | |
171 | max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO/2), MIN_TIMESLICE) | |
172 | ||
48c08d3f | 173 | static unsigned int task_timeslice(task_t *p) |
1da177e4 LT |
174 | { |
175 | if (p->static_prio < NICE_TO_PRIO(0)) | |
176 | return SCALE_PRIO(DEF_TIMESLICE*4, p->static_prio); | |
177 | else | |
178 | return SCALE_PRIO(DEF_TIMESLICE, p->static_prio); | |
179 | } | |
180 | #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran) \ | |
181 | < (long long) (sd)->cache_hot_time) | |
182 | ||
183 | /* | |
184 | * These are the runqueue data structures: | |
185 | */ | |
186 | ||
1da177e4 LT |
187 | typedef struct runqueue runqueue_t; |
188 | ||
189 | struct prio_array { | |
190 | unsigned int nr_active; | |
d444886e | 191 | DECLARE_BITMAP(bitmap, MAX_PRIO+1); /* include 1 bit for delimiter */ |
1da177e4 LT |
192 | struct list_head queue[MAX_PRIO]; |
193 | }; | |
194 | ||
195 | /* | |
196 | * This is the main, per-CPU runqueue data structure. | |
197 | * | |
198 | * Locking rule: those places that want to lock multiple runqueues | |
199 | * (such as the load balancing or the thread migration code), lock | |
200 | * acquire operations must be ordered by ascending &runqueue. | |
201 | */ | |
202 | struct runqueue { | |
203 | spinlock_t lock; | |
204 | ||
205 | /* | |
206 | * nr_running and cpu_load should be in the same cacheline because | |
207 | * remote CPUs use both these fields when doing load calculation. | |
208 | */ | |
209 | unsigned long nr_running; | |
210 | #ifdef CONFIG_SMP | |
7897986b | 211 | unsigned long cpu_load[3]; |
1da177e4 LT |
212 | #endif |
213 | unsigned long long nr_switches; | |
214 | ||
215 | /* | |
216 | * This is part of a global counter where only the total sum | |
217 | * over all CPUs matters. A task can increase this counter on | |
218 | * one CPU and if it got migrated afterwards it may decrease | |
219 | * it on another CPU. Always updated under the runqueue lock: | |
220 | */ | |
221 | unsigned long nr_uninterruptible; | |
222 | ||
223 | unsigned long expired_timestamp; | |
224 | unsigned long long timestamp_last_tick; | |
225 | task_t *curr, *idle; | |
226 | struct mm_struct *prev_mm; | |
227 | prio_array_t *active, *expired, arrays[2]; | |
228 | int best_expired_prio; | |
229 | atomic_t nr_iowait; | |
230 | ||
231 | #ifdef CONFIG_SMP | |
232 | struct sched_domain *sd; | |
233 | ||
234 | /* For active balancing */ | |
235 | int active_balance; | |
236 | int push_cpu; | |
237 | ||
238 | task_t *migration_thread; | |
239 | struct list_head migration_queue; | |
240 | #endif | |
241 | ||
242 | #ifdef CONFIG_SCHEDSTATS | |
243 | /* latency stats */ | |
244 | struct sched_info rq_sched_info; | |
245 | ||
246 | /* sys_sched_yield() stats */ | |
247 | unsigned long yld_exp_empty; | |
248 | unsigned long yld_act_empty; | |
249 | unsigned long yld_both_empty; | |
250 | unsigned long yld_cnt; | |
251 | ||
252 | /* schedule() stats */ | |
253 | unsigned long sched_switch; | |
254 | unsigned long sched_cnt; | |
255 | unsigned long sched_goidle; | |
256 | ||
257 | /* try_to_wake_up() stats */ | |
258 | unsigned long ttwu_cnt; | |
259 | unsigned long ttwu_local; | |
260 | #endif | |
261 | }; | |
262 | ||
263 | static DEFINE_PER_CPU(struct runqueue, runqueues); | |
264 | ||
674311d5 NP |
265 | /* |
266 | * The domain tree (rq->sd) is protected by RCU's quiescent state transition. | |
1a20ff27 | 267 | * See detach_destroy_domains: synchronize_sched for details. |
674311d5 NP |
268 | * |
269 | * The domain tree of any CPU may only be accessed from within | |
270 | * preempt-disabled sections. | |
271 | */ | |
1da177e4 | 272 | #define for_each_domain(cpu, domain) \ |
674311d5 | 273 | for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent) |
1da177e4 LT |
274 | |
275 | #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) | |
276 | #define this_rq() (&__get_cpu_var(runqueues)) | |
277 | #define task_rq(p) cpu_rq(task_cpu(p)) | |
278 | #define cpu_curr(cpu) (cpu_rq(cpu)->curr) | |
279 | ||
1da177e4 | 280 | #ifndef prepare_arch_switch |
4866cde0 NP |
281 | # define prepare_arch_switch(next) do { } while (0) |
282 | #endif | |
283 | #ifndef finish_arch_switch | |
284 | # define finish_arch_switch(prev) do { } while (0) | |
285 | #endif | |
286 | ||
287 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW | |
288 | static inline int task_running(runqueue_t *rq, task_t *p) | |
289 | { | |
290 | return rq->curr == p; | |
291 | } | |
292 | ||
293 | static inline void prepare_lock_switch(runqueue_t *rq, task_t *next) | |
294 | { | |
295 | } | |
296 | ||
297 | static inline void finish_lock_switch(runqueue_t *rq, task_t *prev) | |
298 | { | |
da04c035 IM |
299 | #ifdef CONFIG_DEBUG_SPINLOCK |
300 | /* this is a valid case when another task releases the spinlock */ | |
301 | rq->lock.owner = current; | |
302 | #endif | |
4866cde0 NP |
303 | spin_unlock_irq(&rq->lock); |
304 | } | |
305 | ||
306 | #else /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
307 | static inline int task_running(runqueue_t *rq, task_t *p) | |
308 | { | |
309 | #ifdef CONFIG_SMP | |
310 | return p->oncpu; | |
311 | #else | |
312 | return rq->curr == p; | |
313 | #endif | |
314 | } | |
315 | ||
316 | static inline void prepare_lock_switch(runqueue_t *rq, task_t *next) | |
317 | { | |
318 | #ifdef CONFIG_SMP | |
319 | /* | |
320 | * We can optimise this out completely for !SMP, because the | |
321 | * SMP rebalancing from interrupt is the only thing that cares | |
322 | * here. | |
323 | */ | |
324 | next->oncpu = 1; | |
325 | #endif | |
326 | #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
327 | spin_unlock_irq(&rq->lock); | |
328 | #else | |
329 | spin_unlock(&rq->lock); | |
330 | #endif | |
331 | } | |
332 | ||
333 | static inline void finish_lock_switch(runqueue_t *rq, task_t *prev) | |
334 | { | |
335 | #ifdef CONFIG_SMP | |
336 | /* | |
337 | * After ->oncpu is cleared, the task can be moved to a different CPU. | |
338 | * We must ensure this doesn't happen until the switch is completely | |
339 | * finished. | |
340 | */ | |
341 | smp_wmb(); | |
342 | prev->oncpu = 0; | |
343 | #endif | |
344 | #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
345 | local_irq_enable(); | |
1da177e4 | 346 | #endif |
4866cde0 NP |
347 | } |
348 | #endif /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
1da177e4 LT |
349 | |
350 | /* | |
351 | * task_rq_lock - lock the runqueue a given task resides on and disable | |
352 | * interrupts. Note the ordering: we can safely lookup the task_rq without | |
353 | * explicitly disabling preemption. | |
354 | */ | |
355 | static inline runqueue_t *task_rq_lock(task_t *p, unsigned long *flags) | |
356 | __acquires(rq->lock) | |
357 | { | |
358 | struct runqueue *rq; | |
359 | ||
360 | repeat_lock_task: | |
361 | local_irq_save(*flags); | |
362 | rq = task_rq(p); | |
363 | spin_lock(&rq->lock); | |
364 | if (unlikely(rq != task_rq(p))) { | |
365 | spin_unlock_irqrestore(&rq->lock, *flags); | |
366 | goto repeat_lock_task; | |
367 | } | |
368 | return rq; | |
369 | } | |
370 | ||
371 | static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags) | |
372 | __releases(rq->lock) | |
373 | { | |
374 | spin_unlock_irqrestore(&rq->lock, *flags); | |
375 | } | |
376 | ||
377 | #ifdef CONFIG_SCHEDSTATS | |
378 | /* | |
379 | * bump this up when changing the output format or the meaning of an existing | |
380 | * format, so that tools can adapt (or abort) | |
381 | */ | |
68767a0a | 382 | #define SCHEDSTAT_VERSION 12 |
1da177e4 LT |
383 | |
384 | static int show_schedstat(struct seq_file *seq, void *v) | |
385 | { | |
386 | int cpu; | |
387 | ||
388 | seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION); | |
389 | seq_printf(seq, "timestamp %lu\n", jiffies); | |
390 | for_each_online_cpu(cpu) { | |
391 | runqueue_t *rq = cpu_rq(cpu); | |
392 | #ifdef CONFIG_SMP | |
393 | struct sched_domain *sd; | |
394 | int dcnt = 0; | |
395 | #endif | |
396 | ||
397 | /* runqueue-specific stats */ | |
398 | seq_printf(seq, | |
399 | "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu", | |
400 | cpu, rq->yld_both_empty, | |
401 | rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt, | |
402 | rq->sched_switch, rq->sched_cnt, rq->sched_goidle, | |
403 | rq->ttwu_cnt, rq->ttwu_local, | |
404 | rq->rq_sched_info.cpu_time, | |
405 | rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt); | |
406 | ||
407 | seq_printf(seq, "\n"); | |
408 | ||
409 | #ifdef CONFIG_SMP | |
410 | /* domain-specific stats */ | |
674311d5 | 411 | preempt_disable(); |
1da177e4 LT |
412 | for_each_domain(cpu, sd) { |
413 | enum idle_type itype; | |
414 | char mask_str[NR_CPUS]; | |
415 | ||
416 | cpumask_scnprintf(mask_str, NR_CPUS, sd->span); | |
417 | seq_printf(seq, "domain%d %s", dcnt++, mask_str); | |
418 | for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES; | |
419 | itype++) { | |
420 | seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu", | |
421 | sd->lb_cnt[itype], | |
422 | sd->lb_balanced[itype], | |
423 | sd->lb_failed[itype], | |
424 | sd->lb_imbalance[itype], | |
425 | sd->lb_gained[itype], | |
426 | sd->lb_hot_gained[itype], | |
427 | sd->lb_nobusyq[itype], | |
428 | sd->lb_nobusyg[itype]); | |
429 | } | |
68767a0a | 430 | seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n", |
1da177e4 | 431 | sd->alb_cnt, sd->alb_failed, sd->alb_pushed, |
68767a0a NP |
432 | sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed, |
433 | sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed, | |
1da177e4 LT |
434 | sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance); |
435 | } | |
674311d5 | 436 | preempt_enable(); |
1da177e4 LT |
437 | #endif |
438 | } | |
439 | return 0; | |
440 | } | |
441 | ||
442 | static int schedstat_open(struct inode *inode, struct file *file) | |
443 | { | |
444 | unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32); | |
445 | char *buf = kmalloc(size, GFP_KERNEL); | |
446 | struct seq_file *m; | |
447 | int res; | |
448 | ||
449 | if (!buf) | |
450 | return -ENOMEM; | |
451 | res = single_open(file, show_schedstat, NULL); | |
452 | if (!res) { | |
453 | m = file->private_data; | |
454 | m->buf = buf; | |
455 | m->size = size; | |
456 | } else | |
457 | kfree(buf); | |
458 | return res; | |
459 | } | |
460 | ||
461 | struct file_operations proc_schedstat_operations = { | |
462 | .open = schedstat_open, | |
463 | .read = seq_read, | |
464 | .llseek = seq_lseek, | |
465 | .release = single_release, | |
466 | }; | |
467 | ||
468 | # define schedstat_inc(rq, field) do { (rq)->field++; } while (0) | |
469 | # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0) | |
470 | #else /* !CONFIG_SCHEDSTATS */ | |
471 | # define schedstat_inc(rq, field) do { } while (0) | |
472 | # define schedstat_add(rq, field, amt) do { } while (0) | |
473 | #endif | |
474 | ||
475 | /* | |
476 | * rq_lock - lock a given runqueue and disable interrupts. | |
477 | */ | |
478 | static inline runqueue_t *this_rq_lock(void) | |
479 | __acquires(rq->lock) | |
480 | { | |
481 | runqueue_t *rq; | |
482 | ||
483 | local_irq_disable(); | |
484 | rq = this_rq(); | |
485 | spin_lock(&rq->lock); | |
486 | ||
487 | return rq; | |
488 | } | |
489 | ||
1da177e4 LT |
490 | #ifdef CONFIG_SCHEDSTATS |
491 | /* | |
492 | * Called when a process is dequeued from the active array and given | |
493 | * the cpu. We should note that with the exception of interactive | |
494 | * tasks, the expired queue will become the active queue after the active | |
495 | * queue is empty, without explicitly dequeuing and requeuing tasks in the | |
496 | * expired queue. (Interactive tasks may be requeued directly to the | |
497 | * active queue, thus delaying tasks in the expired queue from running; | |
498 | * see scheduler_tick()). | |
499 | * | |
500 | * This function is only called from sched_info_arrive(), rather than | |
501 | * dequeue_task(). Even though a task may be queued and dequeued multiple | |
502 | * times as it is shuffled about, we're really interested in knowing how | |
503 | * long it was from the *first* time it was queued to the time that it | |
504 | * finally hit a cpu. | |
505 | */ | |
506 | static inline void sched_info_dequeued(task_t *t) | |
507 | { | |
508 | t->sched_info.last_queued = 0; | |
509 | } | |
510 | ||
511 | /* | |
512 | * Called when a task finally hits the cpu. We can now calculate how | |
513 | * long it was waiting to run. We also note when it began so that we | |
514 | * can keep stats on how long its timeslice is. | |
515 | */ | |
858119e1 | 516 | static void sched_info_arrive(task_t *t) |
1da177e4 LT |
517 | { |
518 | unsigned long now = jiffies, diff = 0; | |
519 | struct runqueue *rq = task_rq(t); | |
520 | ||
521 | if (t->sched_info.last_queued) | |
522 | diff = now - t->sched_info.last_queued; | |
523 | sched_info_dequeued(t); | |
524 | t->sched_info.run_delay += diff; | |
525 | t->sched_info.last_arrival = now; | |
526 | t->sched_info.pcnt++; | |
527 | ||
528 | if (!rq) | |
529 | return; | |
530 | ||
531 | rq->rq_sched_info.run_delay += diff; | |
532 | rq->rq_sched_info.pcnt++; | |
533 | } | |
534 | ||
535 | /* | |
536 | * Called when a process is queued into either the active or expired | |
537 | * array. The time is noted and later used to determine how long we | |
538 | * had to wait for us to reach the cpu. Since the expired queue will | |
539 | * become the active queue after active queue is empty, without dequeuing | |
540 | * and requeuing any tasks, we are interested in queuing to either. It | |
541 | * is unusual but not impossible for tasks to be dequeued and immediately | |
542 | * requeued in the same or another array: this can happen in sched_yield(), | |
543 | * set_user_nice(), and even load_balance() as it moves tasks from runqueue | |
544 | * to runqueue. | |
545 | * | |
546 | * This function is only called from enqueue_task(), but also only updates | |
547 | * the timestamp if it is already not set. It's assumed that | |
548 | * sched_info_dequeued() will clear that stamp when appropriate. | |
549 | */ | |
550 | static inline void sched_info_queued(task_t *t) | |
551 | { | |
552 | if (!t->sched_info.last_queued) | |
553 | t->sched_info.last_queued = jiffies; | |
554 | } | |
555 | ||
556 | /* | |
557 | * Called when a process ceases being the active-running process, either | |
558 | * voluntarily or involuntarily. Now we can calculate how long we ran. | |
559 | */ | |
560 | static inline void sched_info_depart(task_t *t) | |
561 | { | |
562 | struct runqueue *rq = task_rq(t); | |
563 | unsigned long diff = jiffies - t->sched_info.last_arrival; | |
564 | ||
565 | t->sched_info.cpu_time += diff; | |
566 | ||
567 | if (rq) | |
568 | rq->rq_sched_info.cpu_time += diff; | |
569 | } | |
570 | ||
571 | /* | |
572 | * Called when tasks are switched involuntarily due, typically, to expiring | |
573 | * their time slice. (This may also be called when switching to or from | |
574 | * the idle task.) We are only called when prev != next. | |
575 | */ | |
576 | static inline void sched_info_switch(task_t *prev, task_t *next) | |
577 | { | |
578 | struct runqueue *rq = task_rq(prev); | |
579 | ||
580 | /* | |
581 | * prev now departs the cpu. It's not interesting to record | |
582 | * stats about how efficient we were at scheduling the idle | |
583 | * process, however. | |
584 | */ | |
585 | if (prev != rq->idle) | |
586 | sched_info_depart(prev); | |
587 | ||
588 | if (next != rq->idle) | |
589 | sched_info_arrive(next); | |
590 | } | |
591 | #else | |
592 | #define sched_info_queued(t) do { } while (0) | |
593 | #define sched_info_switch(t, next) do { } while (0) | |
594 | #endif /* CONFIG_SCHEDSTATS */ | |
595 | ||
596 | /* | |
597 | * Adding/removing a task to/from a priority array: | |
598 | */ | |
599 | static void dequeue_task(struct task_struct *p, prio_array_t *array) | |
600 | { | |
601 | array->nr_active--; | |
602 | list_del(&p->run_list); | |
603 | if (list_empty(array->queue + p->prio)) | |
604 | __clear_bit(p->prio, array->bitmap); | |
605 | } | |
606 | ||
607 | static void enqueue_task(struct task_struct *p, prio_array_t *array) | |
608 | { | |
609 | sched_info_queued(p); | |
610 | list_add_tail(&p->run_list, array->queue + p->prio); | |
611 | __set_bit(p->prio, array->bitmap); | |
612 | array->nr_active++; | |
613 | p->array = array; | |
614 | } | |
615 | ||
616 | /* | |
617 | * Put task to the end of the run list without the overhead of dequeue | |
618 | * followed by enqueue. | |
619 | */ | |
620 | static void requeue_task(struct task_struct *p, prio_array_t *array) | |
621 | { | |
622 | list_move_tail(&p->run_list, array->queue + p->prio); | |
623 | } | |
624 | ||
625 | static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array) | |
626 | { | |
627 | list_add(&p->run_list, array->queue + p->prio); | |
628 | __set_bit(p->prio, array->bitmap); | |
629 | array->nr_active++; | |
630 | p->array = array; | |
631 | } | |
632 | ||
633 | /* | |
634 | * effective_prio - return the priority that is based on the static | |
635 | * priority but is modified by bonuses/penalties. | |
636 | * | |
637 | * We scale the actual sleep average [0 .... MAX_SLEEP_AVG] | |
638 | * into the -5 ... 0 ... +5 bonus/penalty range. | |
639 | * | |
640 | * We use 25% of the full 0...39 priority range so that: | |
641 | * | |
642 | * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs. | |
643 | * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks. | |
644 | * | |
645 | * Both properties are important to certain workloads. | |
646 | */ | |
647 | static int effective_prio(task_t *p) | |
648 | { | |
649 | int bonus, prio; | |
650 | ||
651 | if (rt_task(p)) | |
652 | return p->prio; | |
653 | ||
654 | bonus = CURRENT_BONUS(p) - MAX_BONUS / 2; | |
655 | ||
656 | prio = p->static_prio - bonus; | |
657 | if (prio < MAX_RT_PRIO) | |
658 | prio = MAX_RT_PRIO; | |
659 | if (prio > MAX_PRIO-1) | |
660 | prio = MAX_PRIO-1; | |
661 | return prio; | |
662 | } | |
663 | ||
664 | /* | |
665 | * __activate_task - move a task to the runqueue. | |
666 | */ | |
d425b274 | 667 | static void __activate_task(task_t *p, runqueue_t *rq) |
1da177e4 | 668 | { |
d425b274 CK |
669 | prio_array_t *target = rq->active; |
670 | ||
f1adad78 | 671 | if (batch_task(p)) |
d425b274 CK |
672 | target = rq->expired; |
673 | enqueue_task(p, target); | |
a2000572 | 674 | rq->nr_running++; |
1da177e4 LT |
675 | } |
676 | ||
677 | /* | |
678 | * __activate_idle_task - move idle task to the _front_ of runqueue. | |
679 | */ | |
680 | static inline void __activate_idle_task(task_t *p, runqueue_t *rq) | |
681 | { | |
682 | enqueue_task_head(p, rq->active); | |
a2000572 | 683 | rq->nr_running++; |
1da177e4 LT |
684 | } |
685 | ||
a3464a10 | 686 | static int recalc_task_prio(task_t *p, unsigned long long now) |
1da177e4 LT |
687 | { |
688 | /* Caller must always ensure 'now >= p->timestamp' */ | |
689 | unsigned long long __sleep_time = now - p->timestamp; | |
690 | unsigned long sleep_time; | |
691 | ||
d425b274 | 692 | if (batch_task(p)) |
b0a9499c IM |
693 | sleep_time = 0; |
694 | else { | |
695 | if (__sleep_time > NS_MAX_SLEEP_AVG) | |
696 | sleep_time = NS_MAX_SLEEP_AVG; | |
697 | else | |
698 | sleep_time = (unsigned long)__sleep_time; | |
699 | } | |
1da177e4 LT |
700 | |
701 | if (likely(sleep_time > 0)) { | |
702 | /* | |
703 | * User tasks that sleep a long time are categorised as | |
e72ff0bb CK |
704 | * idle. They will only have their sleep_avg increased to a |
705 | * level that makes them just interactive priority to stay | |
706 | * active yet prevent them suddenly becoming cpu hogs and | |
707 | * starving other processes. | |
1da177e4 | 708 | */ |
5138930e | 709 | if (p->mm && sleep_time > INTERACTIVE_SLEEP(p)) { |
e72ff0bb CK |
710 | unsigned long ceiling; |
711 | ||
712 | ceiling = JIFFIES_TO_NS(MAX_SLEEP_AVG - | |
713 | DEF_TIMESLICE); | |
714 | if (p->sleep_avg < ceiling) | |
715 | p->sleep_avg = ceiling; | |
1da177e4 | 716 | } else { |
1da177e4 LT |
717 | /* |
718 | * Tasks waking from uninterruptible sleep are | |
719 | * limited in their sleep_avg rise as they | |
720 | * are likely to be waiting on I/O | |
721 | */ | |
3dee386e | 722 | if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) { |
1da177e4 LT |
723 | if (p->sleep_avg >= INTERACTIVE_SLEEP(p)) |
724 | sleep_time = 0; | |
725 | else if (p->sleep_avg + sleep_time >= | |
726 | INTERACTIVE_SLEEP(p)) { | |
727 | p->sleep_avg = INTERACTIVE_SLEEP(p); | |
728 | sleep_time = 0; | |
729 | } | |
730 | } | |
731 | ||
732 | /* | |
733 | * This code gives a bonus to interactive tasks. | |
734 | * | |
735 | * The boost works by updating the 'average sleep time' | |
736 | * value here, based on ->timestamp. The more time a | |
737 | * task spends sleeping, the higher the average gets - | |
738 | * and the higher the priority boost gets as well. | |
739 | */ | |
740 | p->sleep_avg += sleep_time; | |
741 | ||
742 | if (p->sleep_avg > NS_MAX_SLEEP_AVG) | |
743 | p->sleep_avg = NS_MAX_SLEEP_AVG; | |
744 | } | |
745 | } | |
746 | ||
a3464a10 | 747 | return effective_prio(p); |
1da177e4 LT |
748 | } |
749 | ||
750 | /* | |
751 | * activate_task - move a task to the runqueue and do priority recalculation | |
752 | * | |
753 | * Update all the scheduling statistics stuff. (sleep average | |
754 | * calculation, priority modifiers, etc.) | |
755 | */ | |
756 | static void activate_task(task_t *p, runqueue_t *rq, int local) | |
757 | { | |
758 | unsigned long long now; | |
759 | ||
760 | now = sched_clock(); | |
761 | #ifdef CONFIG_SMP | |
762 | if (!local) { | |
763 | /* Compensate for drifting sched_clock */ | |
764 | runqueue_t *this_rq = this_rq(); | |
765 | now = (now - this_rq->timestamp_last_tick) | |
766 | + rq->timestamp_last_tick; | |
767 | } | |
768 | #endif | |
769 | ||
a47ab937 CK |
770 | if (!rt_task(p)) |
771 | p->prio = recalc_task_prio(p, now); | |
1da177e4 LT |
772 | |
773 | /* | |
774 | * This checks to make sure it's not an uninterruptible task | |
775 | * that is now waking up. | |
776 | */ | |
3dee386e | 777 | if (p->sleep_type == SLEEP_NORMAL) { |
1da177e4 LT |
778 | /* |
779 | * Tasks which were woken up by interrupts (ie. hw events) | |
780 | * are most likely of interactive nature. So we give them | |
781 | * the credit of extending their sleep time to the period | |
782 | * of time they spend on the runqueue, waiting for execution | |
783 | * on a CPU, first time around: | |
784 | */ | |
785 | if (in_interrupt()) | |
3dee386e | 786 | p->sleep_type = SLEEP_INTERRUPTED; |
1da177e4 LT |
787 | else { |
788 | /* | |
789 | * Normal first-time wakeups get a credit too for | |
790 | * on-runqueue time, but it will be weighted down: | |
791 | */ | |
3dee386e | 792 | p->sleep_type = SLEEP_INTERACTIVE; |
1da177e4 LT |
793 | } |
794 | } | |
795 | p->timestamp = now; | |
796 | ||
797 | __activate_task(p, rq); | |
798 | } | |
799 | ||
800 | /* | |
801 | * deactivate_task - remove a task from the runqueue. | |
802 | */ | |
803 | static void deactivate_task(struct task_struct *p, runqueue_t *rq) | |
804 | { | |
a2000572 | 805 | rq->nr_running--; |
1da177e4 LT |
806 | dequeue_task(p, p->array); |
807 | p->array = NULL; | |
808 | } | |
809 | ||
810 | /* | |
811 | * resched_task - mark a task 'to be rescheduled now'. | |
812 | * | |
813 | * On UP this means the setting of the need_resched flag, on SMP it | |
814 | * might also involve a cross-CPU call to trigger the scheduler on | |
815 | * the target CPU. | |
816 | */ | |
817 | #ifdef CONFIG_SMP | |
495ab9c0 AK |
818 | |
819 | #ifndef tsk_is_polling | |
820 | #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG) | |
821 | #endif | |
822 | ||
1da177e4 LT |
823 | static void resched_task(task_t *p) |
824 | { | |
64c7c8f8 | 825 | int cpu; |
1da177e4 LT |
826 | |
827 | assert_spin_locked(&task_rq(p)->lock); | |
828 | ||
64c7c8f8 NP |
829 | if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED))) |
830 | return; | |
831 | ||
832 | set_tsk_thread_flag(p, TIF_NEED_RESCHED); | |
1da177e4 | 833 | |
64c7c8f8 NP |
834 | cpu = task_cpu(p); |
835 | if (cpu == smp_processor_id()) | |
836 | return; | |
837 | ||
495ab9c0 | 838 | /* NEED_RESCHED must be visible before we test polling */ |
64c7c8f8 | 839 | smp_mb(); |
495ab9c0 | 840 | if (!tsk_is_polling(p)) |
64c7c8f8 | 841 | smp_send_reschedule(cpu); |
1da177e4 LT |
842 | } |
843 | #else | |
844 | static inline void resched_task(task_t *p) | |
845 | { | |
64c7c8f8 | 846 | assert_spin_locked(&task_rq(p)->lock); |
1da177e4 LT |
847 | set_tsk_need_resched(p); |
848 | } | |
849 | #endif | |
850 | ||
851 | /** | |
852 | * task_curr - is this task currently executing on a CPU? | |
853 | * @p: the task in question. | |
854 | */ | |
855 | inline int task_curr(const task_t *p) | |
856 | { | |
857 | return cpu_curr(task_cpu(p)) == p; | |
858 | } | |
859 | ||
860 | #ifdef CONFIG_SMP | |
1da177e4 LT |
861 | typedef struct { |
862 | struct list_head list; | |
1da177e4 | 863 | |
1da177e4 LT |
864 | task_t *task; |
865 | int dest_cpu; | |
866 | ||
1da177e4 LT |
867 | struct completion done; |
868 | } migration_req_t; | |
869 | ||
870 | /* | |
871 | * The task's runqueue lock must be held. | |
872 | * Returns true if you have to wait for migration thread. | |
873 | */ | |
874 | static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req) | |
875 | { | |
876 | runqueue_t *rq = task_rq(p); | |
877 | ||
878 | /* | |
879 | * If the task is not on a runqueue (and not running), then | |
880 | * it is sufficient to simply update the task's cpu field. | |
881 | */ | |
882 | if (!p->array && !task_running(rq, p)) { | |
883 | set_task_cpu(p, dest_cpu); | |
884 | return 0; | |
885 | } | |
886 | ||
887 | init_completion(&req->done); | |
1da177e4 LT |
888 | req->task = p; |
889 | req->dest_cpu = dest_cpu; | |
890 | list_add(&req->list, &rq->migration_queue); | |
891 | return 1; | |
892 | } | |
893 | ||
894 | /* | |
895 | * wait_task_inactive - wait for a thread to unschedule. | |
896 | * | |
897 | * The caller must ensure that the task *will* unschedule sometime soon, | |
898 | * else this function might spin for a *long* time. This function can't | |
899 | * be called with interrupts off, or it may introduce deadlock with | |
900 | * smp_call_function() if an IPI is sent by the same process we are | |
901 | * waiting to become inactive. | |
902 | */ | |
95cdf3b7 | 903 | void wait_task_inactive(task_t *p) |
1da177e4 LT |
904 | { |
905 | unsigned long flags; | |
906 | runqueue_t *rq; | |
907 | int preempted; | |
908 | ||
909 | repeat: | |
910 | rq = task_rq_lock(p, &flags); | |
911 | /* Must be off runqueue entirely, not preempted. */ | |
912 | if (unlikely(p->array || task_running(rq, p))) { | |
913 | /* If it's preempted, we yield. It could be a while. */ | |
914 | preempted = !task_running(rq, p); | |
915 | task_rq_unlock(rq, &flags); | |
916 | cpu_relax(); | |
917 | if (preempted) | |
918 | yield(); | |
919 | goto repeat; | |
920 | } | |
921 | task_rq_unlock(rq, &flags); | |
922 | } | |
923 | ||
924 | /*** | |
925 | * kick_process - kick a running thread to enter/exit the kernel | |
926 | * @p: the to-be-kicked thread | |
927 | * | |
928 | * Cause a process which is running on another CPU to enter | |
929 | * kernel-mode, without any delay. (to get signals handled.) | |
930 | * | |
931 | * NOTE: this function doesnt have to take the runqueue lock, | |
932 | * because all it wants to ensure is that the remote task enters | |
933 | * the kernel. If the IPI races and the task has been migrated | |
934 | * to another CPU then no harm is done and the purpose has been | |
935 | * achieved as well. | |
936 | */ | |
937 | void kick_process(task_t *p) | |
938 | { | |
939 | int cpu; | |
940 | ||
941 | preempt_disable(); | |
942 | cpu = task_cpu(p); | |
943 | if ((cpu != smp_processor_id()) && task_curr(p)) | |
944 | smp_send_reschedule(cpu); | |
945 | preempt_enable(); | |
946 | } | |
947 | ||
948 | /* | |
949 | * Return a low guess at the load of a migration-source cpu. | |
950 | * | |
951 | * We want to under-estimate the load of migration sources, to | |
952 | * balance conservatively. | |
953 | */ | |
a2000572 | 954 | static inline unsigned long source_load(int cpu, int type) |
1da177e4 LT |
955 | { |
956 | runqueue_t *rq = cpu_rq(cpu); | |
a2000572 | 957 | unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE; |
3b0bd9bc | 958 | if (type == 0) |
a2000572 | 959 | return load_now; |
b910472d | 960 | |
a2000572 | 961 | return min(rq->cpu_load[type-1], load_now); |
1da177e4 LT |
962 | } |
963 | ||
964 | /* | |
965 | * Return a high guess at the load of a migration-target cpu | |
966 | */ | |
a2000572 | 967 | static inline unsigned long target_load(int cpu, int type) |
1da177e4 LT |
968 | { |
969 | runqueue_t *rq = cpu_rq(cpu); | |
a2000572 | 970 | unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE; |
7897986b | 971 | if (type == 0) |
a2000572 | 972 | return load_now; |
3b0bd9bc | 973 | |
a2000572 | 974 | return max(rq->cpu_load[type-1], load_now); |
1da177e4 LT |
975 | } |
976 | ||
147cbb4b NP |
977 | /* |
978 | * find_idlest_group finds and returns the least busy CPU group within the | |
979 | * domain. | |
980 | */ | |
981 | static struct sched_group * | |
982 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu) | |
983 | { | |
984 | struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups; | |
985 | unsigned long min_load = ULONG_MAX, this_load = 0; | |
986 | int load_idx = sd->forkexec_idx; | |
987 | int imbalance = 100 + (sd->imbalance_pct-100)/2; | |
988 | ||
989 | do { | |
990 | unsigned long load, avg_load; | |
991 | int local_group; | |
992 | int i; | |
993 | ||
da5a5522 BD |
994 | /* Skip over this group if it has no CPUs allowed */ |
995 | if (!cpus_intersects(group->cpumask, p->cpus_allowed)) | |
996 | goto nextgroup; | |
997 | ||
147cbb4b | 998 | local_group = cpu_isset(this_cpu, group->cpumask); |
147cbb4b NP |
999 | |
1000 | /* Tally up the load of all CPUs in the group */ | |
1001 | avg_load = 0; | |
1002 | ||
1003 | for_each_cpu_mask(i, group->cpumask) { | |
1004 | /* Bias balancing toward cpus of our domain */ | |
1005 | if (local_group) | |
1006 | load = source_load(i, load_idx); | |
1007 | else | |
1008 | load = target_load(i, load_idx); | |
1009 | ||
1010 | avg_load += load; | |
1011 | } | |
1012 | ||
1013 | /* Adjust by relative CPU power of the group */ | |
1014 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
1015 | ||
1016 | if (local_group) { | |
1017 | this_load = avg_load; | |
1018 | this = group; | |
1019 | } else if (avg_load < min_load) { | |
1020 | min_load = avg_load; | |
1021 | idlest = group; | |
1022 | } | |
da5a5522 | 1023 | nextgroup: |
147cbb4b NP |
1024 | group = group->next; |
1025 | } while (group != sd->groups); | |
1026 | ||
1027 | if (!idlest || 100*this_load < imbalance*min_load) | |
1028 | return NULL; | |
1029 | return idlest; | |
1030 | } | |
1031 | ||
1032 | /* | |
1033 | * find_idlest_queue - find the idlest runqueue among the cpus in group. | |
1034 | */ | |
95cdf3b7 IM |
1035 | static int |
1036 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
147cbb4b | 1037 | { |
da5a5522 | 1038 | cpumask_t tmp; |
147cbb4b NP |
1039 | unsigned long load, min_load = ULONG_MAX; |
1040 | int idlest = -1; | |
1041 | int i; | |
1042 | ||
da5a5522 BD |
1043 | /* Traverse only the allowed CPUs */ |
1044 | cpus_and(tmp, group->cpumask, p->cpus_allowed); | |
1045 | ||
1046 | for_each_cpu_mask(i, tmp) { | |
147cbb4b NP |
1047 | load = source_load(i, 0); |
1048 | ||
1049 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
1050 | min_load = load; | |
1051 | idlest = i; | |
1052 | } | |
1053 | } | |
1054 | ||
1055 | return idlest; | |
1056 | } | |
1057 | ||
476d139c NP |
1058 | /* |
1059 | * sched_balance_self: balance the current task (running on cpu) in domains | |
1060 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
1061 | * SD_BALANCE_EXEC. | |
1062 | * | |
1063 | * Balance, ie. select the least loaded group. | |
1064 | * | |
1065 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
1066 | * | |
1067 | * preempt must be disabled. | |
1068 | */ | |
1069 | static int sched_balance_self(int cpu, int flag) | |
1070 | { | |
1071 | struct task_struct *t = current; | |
1072 | struct sched_domain *tmp, *sd = NULL; | |
147cbb4b | 1073 | |
c96d145e | 1074 | for_each_domain(cpu, tmp) { |
476d139c NP |
1075 | if (tmp->flags & flag) |
1076 | sd = tmp; | |
c96d145e | 1077 | } |
476d139c NP |
1078 | |
1079 | while (sd) { | |
1080 | cpumask_t span; | |
1081 | struct sched_group *group; | |
1082 | int new_cpu; | |
1083 | int weight; | |
1084 | ||
1085 | span = sd->span; | |
1086 | group = find_idlest_group(sd, t, cpu); | |
1087 | if (!group) | |
1088 | goto nextlevel; | |
1089 | ||
da5a5522 | 1090 | new_cpu = find_idlest_cpu(group, t, cpu); |
476d139c NP |
1091 | if (new_cpu == -1 || new_cpu == cpu) |
1092 | goto nextlevel; | |
1093 | ||
1094 | /* Now try balancing at a lower domain level */ | |
1095 | cpu = new_cpu; | |
1096 | nextlevel: | |
1097 | sd = NULL; | |
1098 | weight = cpus_weight(span); | |
1099 | for_each_domain(cpu, tmp) { | |
1100 | if (weight <= cpus_weight(tmp->span)) | |
1101 | break; | |
1102 | if (tmp->flags & flag) | |
1103 | sd = tmp; | |
1104 | } | |
1105 | /* while loop will break here if sd == NULL */ | |
1106 | } | |
1107 | ||
1108 | return cpu; | |
1109 | } | |
1110 | ||
1111 | #endif /* CONFIG_SMP */ | |
1da177e4 LT |
1112 | |
1113 | /* | |
1114 | * wake_idle() will wake a task on an idle cpu if task->cpu is | |
1115 | * not idle and an idle cpu is available. The span of cpus to | |
1116 | * search starts with cpus closest then further out as needed, | |
1117 | * so we always favor a closer, idle cpu. | |
1118 | * | |
1119 | * Returns the CPU we should wake onto. | |
1120 | */ | |
1121 | #if defined(ARCH_HAS_SCHED_WAKE_IDLE) | |
1122 | static int wake_idle(int cpu, task_t *p) | |
1123 | { | |
1124 | cpumask_t tmp; | |
1125 | struct sched_domain *sd; | |
1126 | int i; | |
1127 | ||
1128 | if (idle_cpu(cpu)) | |
1129 | return cpu; | |
1130 | ||
1131 | for_each_domain(cpu, sd) { | |
1132 | if (sd->flags & SD_WAKE_IDLE) { | |
e0f364f4 | 1133 | cpus_and(tmp, sd->span, p->cpus_allowed); |
1da177e4 LT |
1134 | for_each_cpu_mask(i, tmp) { |
1135 | if (idle_cpu(i)) | |
1136 | return i; | |
1137 | } | |
1138 | } | |
e0f364f4 NP |
1139 | else |
1140 | break; | |
1da177e4 LT |
1141 | } |
1142 | return cpu; | |
1143 | } | |
1144 | #else | |
1145 | static inline int wake_idle(int cpu, task_t *p) | |
1146 | { | |
1147 | return cpu; | |
1148 | } | |
1149 | #endif | |
1150 | ||
1151 | /*** | |
1152 | * try_to_wake_up - wake up a thread | |
1153 | * @p: the to-be-woken-up thread | |
1154 | * @state: the mask of task states that can be woken | |
1155 | * @sync: do a synchronous wakeup? | |
1156 | * | |
1157 | * Put it on the run-queue if it's not already there. The "current" | |
1158 | * thread is always on the run-queue (except when the actual | |
1159 | * re-schedule is in progress), and as such you're allowed to do | |
1160 | * the simpler "current->state = TASK_RUNNING" to mark yourself | |
1161 | * runnable without the overhead of this. | |
1162 | * | |
1163 | * returns failure only if the task is already active. | |
1164 | */ | |
95cdf3b7 | 1165 | static int try_to_wake_up(task_t *p, unsigned int state, int sync) |
1da177e4 LT |
1166 | { |
1167 | int cpu, this_cpu, success = 0; | |
1168 | unsigned long flags; | |
1169 | long old_state; | |
1170 | runqueue_t *rq; | |
1171 | #ifdef CONFIG_SMP | |
1172 | unsigned long load, this_load; | |
7897986b | 1173 | struct sched_domain *sd, *this_sd = NULL; |
1da177e4 LT |
1174 | int new_cpu; |
1175 | #endif | |
1176 | ||
1177 | rq = task_rq_lock(p, &flags); | |
1178 | old_state = p->state; | |
1179 | if (!(old_state & state)) | |
1180 | goto out; | |
1181 | ||
1182 | if (p->array) | |
1183 | goto out_running; | |
1184 | ||
1185 | cpu = task_cpu(p); | |
1186 | this_cpu = smp_processor_id(); | |
1187 | ||
1188 | #ifdef CONFIG_SMP | |
1189 | if (unlikely(task_running(rq, p))) | |
1190 | goto out_activate; | |
1191 | ||
7897986b NP |
1192 | new_cpu = cpu; |
1193 | ||
1da177e4 LT |
1194 | schedstat_inc(rq, ttwu_cnt); |
1195 | if (cpu == this_cpu) { | |
1196 | schedstat_inc(rq, ttwu_local); | |
7897986b NP |
1197 | goto out_set_cpu; |
1198 | } | |
1199 | ||
1200 | for_each_domain(this_cpu, sd) { | |
1201 | if (cpu_isset(cpu, sd->span)) { | |
1202 | schedstat_inc(sd, ttwu_wake_remote); | |
1203 | this_sd = sd; | |
1204 | break; | |
1da177e4 LT |
1205 | } |
1206 | } | |
1da177e4 | 1207 | |
7897986b | 1208 | if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed))) |
1da177e4 LT |
1209 | goto out_set_cpu; |
1210 | ||
1da177e4 | 1211 | /* |
7897986b | 1212 | * Check for affine wakeup and passive balancing possibilities. |
1da177e4 | 1213 | */ |
7897986b NP |
1214 | if (this_sd) { |
1215 | int idx = this_sd->wake_idx; | |
1216 | unsigned int imbalance; | |
1da177e4 | 1217 | |
a3f21bce NP |
1218 | imbalance = 100 + (this_sd->imbalance_pct - 100) / 2; |
1219 | ||
7897986b NP |
1220 | load = source_load(cpu, idx); |
1221 | this_load = target_load(this_cpu, idx); | |
1da177e4 | 1222 | |
7897986b NP |
1223 | new_cpu = this_cpu; /* Wake to this CPU if we can */ |
1224 | ||
a3f21bce NP |
1225 | if (this_sd->flags & SD_WAKE_AFFINE) { |
1226 | unsigned long tl = this_load; | |
1da177e4 | 1227 | /* |
a3f21bce NP |
1228 | * If sync wakeup then subtract the (maximum possible) |
1229 | * effect of the currently running task from the load | |
1230 | * of the current CPU: | |
1da177e4 | 1231 | */ |
a3f21bce NP |
1232 | if (sync) |
1233 | tl -= SCHED_LOAD_SCALE; | |
1234 | ||
1235 | if ((tl <= load && | |
1236 | tl + target_load(cpu, idx) <= SCHED_LOAD_SCALE) || | |
1237 | 100*(tl + SCHED_LOAD_SCALE) <= imbalance*load) { | |
1238 | /* | |
1239 | * This domain has SD_WAKE_AFFINE and | |
1240 | * p is cache cold in this domain, and | |
1241 | * there is no bad imbalance. | |
1242 | */ | |
1243 | schedstat_inc(this_sd, ttwu_move_affine); | |
1244 | goto out_set_cpu; | |
1245 | } | |
1246 | } | |
1247 | ||
1248 | /* | |
1249 | * Start passive balancing when half the imbalance_pct | |
1250 | * limit is reached. | |
1251 | */ | |
1252 | if (this_sd->flags & SD_WAKE_BALANCE) { | |
1253 | if (imbalance*this_load <= 100*load) { | |
1254 | schedstat_inc(this_sd, ttwu_move_balance); | |
1255 | goto out_set_cpu; | |
1256 | } | |
1da177e4 LT |
1257 | } |
1258 | } | |
1259 | ||
1260 | new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */ | |
1261 | out_set_cpu: | |
1262 | new_cpu = wake_idle(new_cpu, p); | |
1263 | if (new_cpu != cpu) { | |
1264 | set_task_cpu(p, new_cpu); | |
1265 | task_rq_unlock(rq, &flags); | |
1266 | /* might preempt at this point */ | |
1267 | rq = task_rq_lock(p, &flags); | |
1268 | old_state = p->state; | |
1269 | if (!(old_state & state)) | |
1270 | goto out; | |
1271 | if (p->array) | |
1272 | goto out_running; | |
1273 | ||
1274 | this_cpu = smp_processor_id(); | |
1275 | cpu = task_cpu(p); | |
1276 | } | |
1277 | ||
1278 | out_activate: | |
1279 | #endif /* CONFIG_SMP */ | |
1280 | if (old_state == TASK_UNINTERRUPTIBLE) { | |
1281 | rq->nr_uninterruptible--; | |
1282 | /* | |
1283 | * Tasks on involuntary sleep don't earn | |
1284 | * sleep_avg beyond just interactive state. | |
1285 | */ | |
3dee386e | 1286 | p->sleep_type = SLEEP_NONINTERACTIVE; |
e7c38cb4 | 1287 | } else |
1da177e4 | 1288 | |
d79fc0fc IM |
1289 | /* |
1290 | * Tasks that have marked their sleep as noninteractive get | |
e7c38cb4 CK |
1291 | * woken up with their sleep average not weighted in an |
1292 | * interactive way. | |
d79fc0fc | 1293 | */ |
e7c38cb4 CK |
1294 | if (old_state & TASK_NONINTERACTIVE) |
1295 | p->sleep_type = SLEEP_NONINTERACTIVE; | |
1296 | ||
1297 | ||
1298 | activate_task(p, rq, cpu == this_cpu); | |
1da177e4 LT |
1299 | /* |
1300 | * Sync wakeups (i.e. those types of wakeups where the waker | |
1301 | * has indicated that it will leave the CPU in short order) | |
1302 | * don't trigger a preemption, if the woken up task will run on | |
1303 | * this cpu. (in this case the 'I will reschedule' promise of | |
1304 | * the waker guarantees that the freshly woken up task is going | |
1305 | * to be considered on this CPU.) | |
1306 | */ | |
1da177e4 LT |
1307 | if (!sync || cpu != this_cpu) { |
1308 | if (TASK_PREEMPTS_CURR(p, rq)) | |
1309 | resched_task(rq->curr); | |
1310 | } | |
1311 | success = 1; | |
1312 | ||
1313 | out_running: | |
1314 | p->state = TASK_RUNNING; | |
1315 | out: | |
1316 | task_rq_unlock(rq, &flags); | |
1317 | ||
1318 | return success; | |
1319 | } | |
1320 | ||
95cdf3b7 | 1321 | int fastcall wake_up_process(task_t *p) |
1da177e4 LT |
1322 | { |
1323 | return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED | | |
1324 | TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0); | |
1325 | } | |
1326 | ||
1327 | EXPORT_SYMBOL(wake_up_process); | |
1328 | ||
1329 | int fastcall wake_up_state(task_t *p, unsigned int state) | |
1330 | { | |
1331 | return try_to_wake_up(p, state, 0); | |
1332 | } | |
1333 | ||
1da177e4 LT |
1334 | /* |
1335 | * Perform scheduler related setup for a newly forked process p. | |
1336 | * p is forked by current. | |
1337 | */ | |
476d139c | 1338 | void fastcall sched_fork(task_t *p, int clone_flags) |
1da177e4 | 1339 | { |
476d139c NP |
1340 | int cpu = get_cpu(); |
1341 | ||
1342 | #ifdef CONFIG_SMP | |
1343 | cpu = sched_balance_self(cpu, SD_BALANCE_FORK); | |
1344 | #endif | |
1345 | set_task_cpu(p, cpu); | |
1346 | ||
1da177e4 LT |
1347 | /* |
1348 | * We mark the process as running here, but have not actually | |
1349 | * inserted it onto the runqueue yet. This guarantees that | |
1350 | * nobody will actually run it, and a signal or other external | |
1351 | * event cannot wake it up and insert it on the runqueue either. | |
1352 | */ | |
1353 | p->state = TASK_RUNNING; | |
1354 | INIT_LIST_HEAD(&p->run_list); | |
1355 | p->array = NULL; | |
1da177e4 LT |
1356 | #ifdef CONFIG_SCHEDSTATS |
1357 | memset(&p->sched_info, 0, sizeof(p->sched_info)); | |
1358 | #endif | |
d6077cb8 | 1359 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
4866cde0 NP |
1360 | p->oncpu = 0; |
1361 | #endif | |
1da177e4 | 1362 | #ifdef CONFIG_PREEMPT |
4866cde0 | 1363 | /* Want to start with kernel preemption disabled. */ |
a1261f54 | 1364 | task_thread_info(p)->preempt_count = 1; |
1da177e4 LT |
1365 | #endif |
1366 | /* | |
1367 | * Share the timeslice between parent and child, thus the | |
1368 | * total amount of pending timeslices in the system doesn't change, | |
1369 | * resulting in more scheduling fairness. | |
1370 | */ | |
1371 | local_irq_disable(); | |
1372 | p->time_slice = (current->time_slice + 1) >> 1; | |
1373 | /* | |
1374 | * The remainder of the first timeslice might be recovered by | |
1375 | * the parent if the child exits early enough. | |
1376 | */ | |
1377 | p->first_time_slice = 1; | |
1378 | current->time_slice >>= 1; | |
1379 | p->timestamp = sched_clock(); | |
1380 | if (unlikely(!current->time_slice)) { | |
1381 | /* | |
1382 | * This case is rare, it happens when the parent has only | |
1383 | * a single jiffy left from its timeslice. Taking the | |
1384 | * runqueue lock is not a problem. | |
1385 | */ | |
1386 | current->time_slice = 1; | |
1da177e4 | 1387 | scheduler_tick(); |
476d139c NP |
1388 | } |
1389 | local_irq_enable(); | |
1390 | put_cpu(); | |
1da177e4 LT |
1391 | } |
1392 | ||
1393 | /* | |
1394 | * wake_up_new_task - wake up a newly created task for the first time. | |
1395 | * | |
1396 | * This function will do some initial scheduler statistics housekeeping | |
1397 | * that must be done for every newly created context, then puts the task | |
1398 | * on the runqueue and wakes it. | |
1399 | */ | |
95cdf3b7 | 1400 | void fastcall wake_up_new_task(task_t *p, unsigned long clone_flags) |
1da177e4 LT |
1401 | { |
1402 | unsigned long flags; | |
1403 | int this_cpu, cpu; | |
1404 | runqueue_t *rq, *this_rq; | |
1405 | ||
1406 | rq = task_rq_lock(p, &flags); | |
147cbb4b | 1407 | BUG_ON(p->state != TASK_RUNNING); |
1da177e4 | 1408 | this_cpu = smp_processor_id(); |
147cbb4b | 1409 | cpu = task_cpu(p); |
1da177e4 | 1410 | |
1da177e4 LT |
1411 | /* |
1412 | * We decrease the sleep average of forking parents | |
1413 | * and children as well, to keep max-interactive tasks | |
1414 | * from forking tasks that are max-interactive. The parent | |
1415 | * (current) is done further down, under its lock. | |
1416 | */ | |
1417 | p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) * | |
1418 | CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS); | |
1419 | ||
1420 | p->prio = effective_prio(p); | |
1421 | ||
1422 | if (likely(cpu == this_cpu)) { | |
1423 | if (!(clone_flags & CLONE_VM)) { | |
1424 | /* | |
1425 | * The VM isn't cloned, so we're in a good position to | |
1426 | * do child-runs-first in anticipation of an exec. This | |
1427 | * usually avoids a lot of COW overhead. | |
1428 | */ | |
1429 | if (unlikely(!current->array)) | |
1430 | __activate_task(p, rq); | |
1431 | else { | |
1432 | p->prio = current->prio; | |
1433 | list_add_tail(&p->run_list, ¤t->run_list); | |
1434 | p->array = current->array; | |
1435 | p->array->nr_active++; | |
a2000572 | 1436 | rq->nr_running++; |
1da177e4 LT |
1437 | } |
1438 | set_need_resched(); | |
1439 | } else | |
1440 | /* Run child last */ | |
1441 | __activate_task(p, rq); | |
1442 | /* | |
1443 | * We skip the following code due to cpu == this_cpu | |
1444 | * | |
1445 | * task_rq_unlock(rq, &flags); | |
1446 | * this_rq = task_rq_lock(current, &flags); | |
1447 | */ | |
1448 | this_rq = rq; | |
1449 | } else { | |
1450 | this_rq = cpu_rq(this_cpu); | |
1451 | ||
1452 | /* | |
1453 | * Not the local CPU - must adjust timestamp. This should | |
1454 | * get optimised away in the !CONFIG_SMP case. | |
1455 | */ | |
1456 | p->timestamp = (p->timestamp - this_rq->timestamp_last_tick) | |
1457 | + rq->timestamp_last_tick; | |
1458 | __activate_task(p, rq); | |
1459 | if (TASK_PREEMPTS_CURR(p, rq)) | |
1460 | resched_task(rq->curr); | |
1461 | ||
1462 | /* | |
1463 | * Parent and child are on different CPUs, now get the | |
1464 | * parent runqueue to update the parent's ->sleep_avg: | |
1465 | */ | |
1466 | task_rq_unlock(rq, &flags); | |
1467 | this_rq = task_rq_lock(current, &flags); | |
1468 | } | |
1469 | current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) * | |
1470 | PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS); | |
1471 | task_rq_unlock(this_rq, &flags); | |
1472 | } | |
1473 | ||
1474 | /* | |
1475 | * Potentially available exiting-child timeslices are | |
1476 | * retrieved here - this way the parent does not get | |
1477 | * penalized for creating too many threads. | |
1478 | * | |
1479 | * (this cannot be used to 'generate' timeslices | |
1480 | * artificially, because any timeslice recovered here | |
1481 | * was given away by the parent in the first place.) | |
1482 | */ | |
95cdf3b7 | 1483 | void fastcall sched_exit(task_t *p) |
1da177e4 LT |
1484 | { |
1485 | unsigned long flags; | |
1486 | runqueue_t *rq; | |
1487 | ||
1488 | /* | |
1489 | * If the child was a (relative-) CPU hog then decrease | |
1490 | * the sleep_avg of the parent as well. | |
1491 | */ | |
1492 | rq = task_rq_lock(p->parent, &flags); | |
889dfafe | 1493 | if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) { |
1da177e4 LT |
1494 | p->parent->time_slice += p->time_slice; |
1495 | if (unlikely(p->parent->time_slice > task_timeslice(p))) | |
1496 | p->parent->time_slice = task_timeslice(p); | |
1497 | } | |
1498 | if (p->sleep_avg < p->parent->sleep_avg) | |
1499 | p->parent->sleep_avg = p->parent->sleep_avg / | |
1500 | (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg / | |
1501 | (EXIT_WEIGHT + 1); | |
1502 | task_rq_unlock(rq, &flags); | |
1503 | } | |
1504 | ||
4866cde0 NP |
1505 | /** |
1506 | * prepare_task_switch - prepare to switch tasks | |
1507 | * @rq: the runqueue preparing to switch | |
1508 | * @next: the task we are going to switch to. | |
1509 | * | |
1510 | * This is called with the rq lock held and interrupts off. It must | |
1511 | * be paired with a subsequent finish_task_switch after the context | |
1512 | * switch. | |
1513 | * | |
1514 | * prepare_task_switch sets up locking and calls architecture specific | |
1515 | * hooks. | |
1516 | */ | |
1517 | static inline void prepare_task_switch(runqueue_t *rq, task_t *next) | |
1518 | { | |
1519 | prepare_lock_switch(rq, next); | |
1520 | prepare_arch_switch(next); | |
1521 | } | |
1522 | ||
1da177e4 LT |
1523 | /** |
1524 | * finish_task_switch - clean up after a task-switch | |
344babaa | 1525 | * @rq: runqueue associated with task-switch |
1da177e4 LT |
1526 | * @prev: the thread we just switched away from. |
1527 | * | |
4866cde0 NP |
1528 | * finish_task_switch must be called after the context switch, paired |
1529 | * with a prepare_task_switch call before the context switch. | |
1530 | * finish_task_switch will reconcile locking set up by prepare_task_switch, | |
1531 | * and do any other architecture-specific cleanup actions. | |
1da177e4 LT |
1532 | * |
1533 | * Note that we may have delayed dropping an mm in context_switch(). If | |
1534 | * so, we finish that here outside of the runqueue lock. (Doing it | |
1535 | * with the lock held can cause deadlocks; see schedule() for | |
1536 | * details.) | |
1537 | */ | |
4866cde0 | 1538 | static inline void finish_task_switch(runqueue_t *rq, task_t *prev) |
1da177e4 LT |
1539 | __releases(rq->lock) |
1540 | { | |
1da177e4 LT |
1541 | struct mm_struct *mm = rq->prev_mm; |
1542 | unsigned long prev_task_flags; | |
1543 | ||
1544 | rq->prev_mm = NULL; | |
1545 | ||
1546 | /* | |
1547 | * A task struct has one reference for the use as "current". | |
1548 | * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and | |
1549 | * calls schedule one last time. The schedule call will never return, | |
1550 | * and the scheduled task must drop that reference. | |
1551 | * The test for EXIT_ZOMBIE must occur while the runqueue locks are | |
1552 | * still held, otherwise prev could be scheduled on another cpu, die | |
1553 | * there before we look at prev->state, and then the reference would | |
1554 | * be dropped twice. | |
1555 | * Manfred Spraul <manfred@colorfullife.com> | |
1556 | */ | |
1557 | prev_task_flags = prev->flags; | |
4866cde0 NP |
1558 | finish_arch_switch(prev); |
1559 | finish_lock_switch(rq, prev); | |
1da177e4 LT |
1560 | if (mm) |
1561 | mmdrop(mm); | |
c6fd91f0 | 1562 | if (unlikely(prev_task_flags & PF_DEAD)) { |
1563 | /* | |
1564 | * Remove function-return probe instances associated with this | |
1565 | * task and put them back on the free list. | |
1566 | */ | |
1567 | kprobe_flush_task(prev); | |
1da177e4 | 1568 | put_task_struct(prev); |
c6fd91f0 | 1569 | } |
1da177e4 LT |
1570 | } |
1571 | ||
1572 | /** | |
1573 | * schedule_tail - first thing a freshly forked thread must call. | |
1574 | * @prev: the thread we just switched away from. | |
1575 | */ | |
1576 | asmlinkage void schedule_tail(task_t *prev) | |
1577 | __releases(rq->lock) | |
1578 | { | |
4866cde0 NP |
1579 | runqueue_t *rq = this_rq(); |
1580 | finish_task_switch(rq, prev); | |
1581 | #ifdef __ARCH_WANT_UNLOCKED_CTXSW | |
1582 | /* In this case, finish_task_switch does not reenable preemption */ | |
1583 | preempt_enable(); | |
1584 | #endif | |
1da177e4 LT |
1585 | if (current->set_child_tid) |
1586 | put_user(current->pid, current->set_child_tid); | |
1587 | } | |
1588 | ||
1589 | /* | |
1590 | * context_switch - switch to the new MM and the new | |
1591 | * thread's register state. | |
1592 | */ | |
1593 | static inline | |
1594 | task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next) | |
1595 | { | |
1596 | struct mm_struct *mm = next->mm; | |
1597 | struct mm_struct *oldmm = prev->active_mm; | |
1598 | ||
1599 | if (unlikely(!mm)) { | |
1600 | next->active_mm = oldmm; | |
1601 | atomic_inc(&oldmm->mm_count); | |
1602 | enter_lazy_tlb(oldmm, next); | |
1603 | } else | |
1604 | switch_mm(oldmm, mm, next); | |
1605 | ||
1606 | if (unlikely(!prev->mm)) { | |
1607 | prev->active_mm = NULL; | |
1608 | WARN_ON(rq->prev_mm); | |
1609 | rq->prev_mm = oldmm; | |
1610 | } | |
1611 | ||
1612 | /* Here we just switch the register state and the stack. */ | |
1613 | switch_to(prev, next, prev); | |
1614 | ||
1615 | return prev; | |
1616 | } | |
1617 | ||
1618 | /* | |
1619 | * nr_running, nr_uninterruptible and nr_context_switches: | |
1620 | * | |
1621 | * externally visible scheduler statistics: current number of runnable | |
1622 | * threads, current number of uninterruptible-sleeping threads, total | |
1623 | * number of context switches performed since bootup. | |
1624 | */ | |
1625 | unsigned long nr_running(void) | |
1626 | { | |
1627 | unsigned long i, sum = 0; | |
1628 | ||
1629 | for_each_online_cpu(i) | |
1630 | sum += cpu_rq(i)->nr_running; | |
1631 | ||
1632 | return sum; | |
1633 | } | |
1634 | ||
1635 | unsigned long nr_uninterruptible(void) | |
1636 | { | |
1637 | unsigned long i, sum = 0; | |
1638 | ||
0a945022 | 1639 | for_each_possible_cpu(i) |
1da177e4 LT |
1640 | sum += cpu_rq(i)->nr_uninterruptible; |
1641 | ||
1642 | /* | |
1643 | * Since we read the counters lockless, it might be slightly | |
1644 | * inaccurate. Do not allow it to go below zero though: | |
1645 | */ | |
1646 | if (unlikely((long)sum < 0)) | |
1647 | sum = 0; | |
1648 | ||
1649 | return sum; | |
1650 | } | |
1651 | ||
1652 | unsigned long long nr_context_switches(void) | |
1653 | { | |
1654 | unsigned long long i, sum = 0; | |
1655 | ||
0a945022 | 1656 | for_each_possible_cpu(i) |
1da177e4 LT |
1657 | sum += cpu_rq(i)->nr_switches; |
1658 | ||
1659 | return sum; | |
1660 | } | |
1661 | ||
1662 | unsigned long nr_iowait(void) | |
1663 | { | |
1664 | unsigned long i, sum = 0; | |
1665 | ||
0a945022 | 1666 | for_each_possible_cpu(i) |
1da177e4 LT |
1667 | sum += atomic_read(&cpu_rq(i)->nr_iowait); |
1668 | ||
1669 | return sum; | |
1670 | } | |
1671 | ||
db1b1fef JS |
1672 | unsigned long nr_active(void) |
1673 | { | |
1674 | unsigned long i, running = 0, uninterruptible = 0; | |
1675 | ||
1676 | for_each_online_cpu(i) { | |
1677 | running += cpu_rq(i)->nr_running; | |
1678 | uninterruptible += cpu_rq(i)->nr_uninterruptible; | |
1679 | } | |
1680 | ||
1681 | if (unlikely((long)uninterruptible < 0)) | |
1682 | uninterruptible = 0; | |
1683 | ||
1684 | return running + uninterruptible; | |
1685 | } | |
1686 | ||
1da177e4 LT |
1687 | #ifdef CONFIG_SMP |
1688 | ||
1689 | /* | |
1690 | * double_rq_lock - safely lock two runqueues | |
1691 | * | |
1692 | * Note this does not disable interrupts like task_rq_lock, | |
1693 | * you need to do so manually before calling. | |
1694 | */ | |
1695 | static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2) | |
1696 | __acquires(rq1->lock) | |
1697 | __acquires(rq2->lock) | |
1698 | { | |
1699 | if (rq1 == rq2) { | |
1700 | spin_lock(&rq1->lock); | |
1701 | __acquire(rq2->lock); /* Fake it out ;) */ | |
1702 | } else { | |
c96d145e | 1703 | if (rq1 < rq2) { |
1da177e4 LT |
1704 | spin_lock(&rq1->lock); |
1705 | spin_lock(&rq2->lock); | |
1706 | } else { | |
1707 | spin_lock(&rq2->lock); | |
1708 | spin_lock(&rq1->lock); | |
1709 | } | |
1710 | } | |
1711 | } | |
1712 | ||
1713 | /* | |
1714 | * double_rq_unlock - safely unlock two runqueues | |
1715 | * | |
1716 | * Note this does not restore interrupts like task_rq_unlock, | |
1717 | * you need to do so manually after calling. | |
1718 | */ | |
1719 | static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2) | |
1720 | __releases(rq1->lock) | |
1721 | __releases(rq2->lock) | |
1722 | { | |
1723 | spin_unlock(&rq1->lock); | |
1724 | if (rq1 != rq2) | |
1725 | spin_unlock(&rq2->lock); | |
1726 | else | |
1727 | __release(rq2->lock); | |
1728 | } | |
1729 | ||
1730 | /* | |
1731 | * double_lock_balance - lock the busiest runqueue, this_rq is locked already. | |
1732 | */ | |
1733 | static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest) | |
1734 | __releases(this_rq->lock) | |
1735 | __acquires(busiest->lock) | |
1736 | __acquires(this_rq->lock) | |
1737 | { | |
1738 | if (unlikely(!spin_trylock(&busiest->lock))) { | |
c96d145e | 1739 | if (busiest < this_rq) { |
1da177e4 LT |
1740 | spin_unlock(&this_rq->lock); |
1741 | spin_lock(&busiest->lock); | |
1742 | spin_lock(&this_rq->lock); | |
1743 | } else | |
1744 | spin_lock(&busiest->lock); | |
1745 | } | |
1746 | } | |
1747 | ||
1da177e4 LT |
1748 | /* |
1749 | * If dest_cpu is allowed for this process, migrate the task to it. | |
1750 | * This is accomplished by forcing the cpu_allowed mask to only | |
1751 | * allow dest_cpu, which will force the cpu onto dest_cpu. Then | |
1752 | * the cpu_allowed mask is restored. | |
1753 | */ | |
1754 | static void sched_migrate_task(task_t *p, int dest_cpu) | |
1755 | { | |
1756 | migration_req_t req; | |
1757 | runqueue_t *rq; | |
1758 | unsigned long flags; | |
1759 | ||
1760 | rq = task_rq_lock(p, &flags); | |
1761 | if (!cpu_isset(dest_cpu, p->cpus_allowed) | |
1762 | || unlikely(cpu_is_offline(dest_cpu))) | |
1763 | goto out; | |
1764 | ||
1765 | /* force the process onto the specified CPU */ | |
1766 | if (migrate_task(p, dest_cpu, &req)) { | |
1767 | /* Need to wait for migration thread (might exit: take ref). */ | |
1768 | struct task_struct *mt = rq->migration_thread; | |
1769 | get_task_struct(mt); | |
1770 | task_rq_unlock(rq, &flags); | |
1771 | wake_up_process(mt); | |
1772 | put_task_struct(mt); | |
1773 | wait_for_completion(&req.done); | |
1774 | return; | |
1775 | } | |
1776 | out: | |
1777 | task_rq_unlock(rq, &flags); | |
1778 | } | |
1779 | ||
1780 | /* | |
476d139c NP |
1781 | * sched_exec - execve() is a valuable balancing opportunity, because at |
1782 | * this point the task has the smallest effective memory and cache footprint. | |
1da177e4 LT |
1783 | */ |
1784 | void sched_exec(void) | |
1785 | { | |
1da177e4 | 1786 | int new_cpu, this_cpu = get_cpu(); |
476d139c | 1787 | new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC); |
1da177e4 | 1788 | put_cpu(); |
476d139c NP |
1789 | if (new_cpu != this_cpu) |
1790 | sched_migrate_task(current, new_cpu); | |
1da177e4 LT |
1791 | } |
1792 | ||
1793 | /* | |
1794 | * pull_task - move a task from a remote runqueue to the local runqueue. | |
1795 | * Both runqueues must be locked. | |
1796 | */ | |
858119e1 | 1797 | static |
1da177e4 LT |
1798 | void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p, |
1799 | runqueue_t *this_rq, prio_array_t *this_array, int this_cpu) | |
1800 | { | |
1801 | dequeue_task(p, src_array); | |
a2000572 | 1802 | src_rq->nr_running--; |
1da177e4 | 1803 | set_task_cpu(p, this_cpu); |
a2000572 | 1804 | this_rq->nr_running++; |
1da177e4 LT |
1805 | enqueue_task(p, this_array); |
1806 | p->timestamp = (p->timestamp - src_rq->timestamp_last_tick) | |
1807 | + this_rq->timestamp_last_tick; | |
1808 | /* | |
1809 | * Note that idle threads have a prio of MAX_PRIO, for this test | |
1810 | * to be always true for them. | |
1811 | */ | |
1812 | if (TASK_PREEMPTS_CURR(p, this_rq)) | |
1813 | resched_task(this_rq->curr); | |
1814 | } | |
1815 | ||
1816 | /* | |
1817 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
1818 | */ | |
858119e1 | 1819 | static |
1da177e4 | 1820 | int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu, |
95cdf3b7 IM |
1821 | struct sched_domain *sd, enum idle_type idle, |
1822 | int *all_pinned) | |
1da177e4 LT |
1823 | { |
1824 | /* | |
1825 | * We do not migrate tasks that are: | |
1826 | * 1) running (obviously), or | |
1827 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
1828 | * 3) are cache-hot on their current CPU. | |
1829 | */ | |
1da177e4 LT |
1830 | if (!cpu_isset(this_cpu, p->cpus_allowed)) |
1831 | return 0; | |
81026794 NP |
1832 | *all_pinned = 0; |
1833 | ||
1834 | if (task_running(rq, p)) | |
1835 | return 0; | |
1da177e4 LT |
1836 | |
1837 | /* | |
1838 | * Aggressive migration if: | |
cafb20c1 | 1839 | * 1) task is cache cold, or |
1da177e4 LT |
1840 | * 2) too many balance attempts have failed. |
1841 | */ | |
1842 | ||
cafb20c1 | 1843 | if (sd->nr_balance_failed > sd->cache_nice_tries) |
1da177e4 LT |
1844 | return 1; |
1845 | ||
1846 | if (task_hot(p, rq->timestamp_last_tick, sd)) | |
81026794 | 1847 | return 0; |
1da177e4 LT |
1848 | return 1; |
1849 | } | |
1850 | ||
1851 | /* | |
1852 | * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq, | |
1853 | * as part of a balancing operation within "domain". Returns the number of | |
1854 | * tasks moved. | |
1855 | * | |
1856 | * Called with both runqueues locked. | |
1857 | */ | |
1858 | static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest, | |
1859 | unsigned long max_nr_move, struct sched_domain *sd, | |
81026794 | 1860 | enum idle_type idle, int *all_pinned) |
1da177e4 LT |
1861 | { |
1862 | prio_array_t *array, *dst_array; | |
1863 | struct list_head *head, *curr; | |
81026794 | 1864 | int idx, pulled = 0, pinned = 0; |
1da177e4 LT |
1865 | task_t *tmp; |
1866 | ||
81026794 | 1867 | if (max_nr_move == 0) |
1da177e4 LT |
1868 | goto out; |
1869 | ||
81026794 NP |
1870 | pinned = 1; |
1871 | ||
1da177e4 LT |
1872 | /* |
1873 | * We first consider expired tasks. Those will likely not be | |
1874 | * executed in the near future, and they are most likely to | |
1875 | * be cache-cold, thus switching CPUs has the least effect | |
1876 | * on them. | |
1877 | */ | |
1878 | if (busiest->expired->nr_active) { | |
1879 | array = busiest->expired; | |
1880 | dst_array = this_rq->expired; | |
1881 | } else { | |
1882 | array = busiest->active; | |
1883 | dst_array = this_rq->active; | |
1884 | } | |
1885 | ||
1886 | new_array: | |
1887 | /* Start searching at priority 0: */ | |
1888 | idx = 0; | |
1889 | skip_bitmap: | |
1890 | if (!idx) | |
1891 | idx = sched_find_first_bit(array->bitmap); | |
1892 | else | |
1893 | idx = find_next_bit(array->bitmap, MAX_PRIO, idx); | |
1894 | if (idx >= MAX_PRIO) { | |
1895 | if (array == busiest->expired && busiest->active->nr_active) { | |
1896 | array = busiest->active; | |
1897 | dst_array = this_rq->active; | |
1898 | goto new_array; | |
1899 | } | |
1900 | goto out; | |
1901 | } | |
1902 | ||
1903 | head = array->queue + idx; | |
1904 | curr = head->prev; | |
1905 | skip_queue: | |
1906 | tmp = list_entry(curr, task_t, run_list); | |
1907 | ||
1908 | curr = curr->prev; | |
1909 | ||
81026794 | 1910 | if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) { |
1da177e4 LT |
1911 | if (curr != head) |
1912 | goto skip_queue; | |
1913 | idx++; | |
1914 | goto skip_bitmap; | |
1915 | } | |
1916 | ||
1917 | #ifdef CONFIG_SCHEDSTATS | |
1918 | if (task_hot(tmp, busiest->timestamp_last_tick, sd)) | |
1919 | schedstat_inc(sd, lb_hot_gained[idle]); | |
1920 | #endif | |
1921 | ||
1922 | pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu); | |
1923 | pulled++; | |
1924 | ||
1925 | /* We only want to steal up to the prescribed number of tasks. */ | |
1926 | if (pulled < max_nr_move) { | |
1927 | if (curr != head) | |
1928 | goto skip_queue; | |
1929 | idx++; | |
1930 | goto skip_bitmap; | |
1931 | } | |
1932 | out: | |
1933 | /* | |
1934 | * Right now, this is the only place pull_task() is called, | |
1935 | * so we can safely collect pull_task() stats here rather than | |
1936 | * inside pull_task(). | |
1937 | */ | |
1938 | schedstat_add(sd, lb_gained[idle], pulled); | |
81026794 NP |
1939 | |
1940 | if (all_pinned) | |
1941 | *all_pinned = pinned; | |
1da177e4 LT |
1942 | return pulled; |
1943 | } | |
1944 | ||
1945 | /* | |
1946 | * find_busiest_group finds and returns the busiest CPU group within the | |
1947 | * domain. It calculates and returns the number of tasks which should be | |
1948 | * moved to restore balance via the imbalance parameter. | |
1949 | */ | |
1950 | static struct sched_group * | |
1951 | find_busiest_group(struct sched_domain *sd, int this_cpu, | |
5969fe06 | 1952 | unsigned long *imbalance, enum idle_type idle, int *sd_idle) |
1da177e4 LT |
1953 | { |
1954 | struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups; | |
1955 | unsigned long max_load, avg_load, total_load, this_load, total_pwr; | |
0c117f1b | 1956 | unsigned long max_pull; |
7897986b | 1957 | int load_idx; |
1da177e4 LT |
1958 | |
1959 | max_load = this_load = total_load = total_pwr = 0; | |
7897986b NP |
1960 | if (idle == NOT_IDLE) |
1961 | load_idx = sd->busy_idx; | |
1962 | else if (idle == NEWLY_IDLE) | |
1963 | load_idx = sd->newidle_idx; | |
1964 | else | |
1965 | load_idx = sd->idle_idx; | |
1da177e4 LT |
1966 | |
1967 | do { | |
1968 | unsigned long load; | |
1969 | int local_group; | |
1970 | int i; | |
1971 | ||
1972 | local_group = cpu_isset(this_cpu, group->cpumask); | |
1973 | ||
1974 | /* Tally up the load of all CPUs in the group */ | |
1975 | avg_load = 0; | |
1976 | ||
1977 | for_each_cpu_mask(i, group->cpumask) { | |
5969fe06 NP |
1978 | if (*sd_idle && !idle_cpu(i)) |
1979 | *sd_idle = 0; | |
1980 | ||
1da177e4 LT |
1981 | /* Bias balancing toward cpus of our domain */ |
1982 | if (local_group) | |
a2000572 | 1983 | load = target_load(i, load_idx); |
1da177e4 | 1984 | else |
a2000572 | 1985 | load = source_load(i, load_idx); |
1da177e4 LT |
1986 | |
1987 | avg_load += load; | |
1988 | } | |
1989 | ||
1990 | total_load += avg_load; | |
1991 | total_pwr += group->cpu_power; | |
1992 | ||
1993 | /* Adjust by relative CPU power of the group */ | |
1994 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
1995 | ||
1996 | if (local_group) { | |
1997 | this_load = avg_load; | |
1998 | this = group; | |
1da177e4 LT |
1999 | } else if (avg_load > max_load) { |
2000 | max_load = avg_load; | |
2001 | busiest = group; | |
2002 | } | |
1da177e4 LT |
2003 | group = group->next; |
2004 | } while (group != sd->groups); | |
2005 | ||
0c117f1b | 2006 | if (!busiest || this_load >= max_load || max_load <= SCHED_LOAD_SCALE) |
1da177e4 LT |
2007 | goto out_balanced; |
2008 | ||
2009 | avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr; | |
2010 | ||
2011 | if (this_load >= avg_load || | |
2012 | 100*max_load <= sd->imbalance_pct*this_load) | |
2013 | goto out_balanced; | |
2014 | ||
2015 | /* | |
2016 | * We're trying to get all the cpus to the average_load, so we don't | |
2017 | * want to push ourselves above the average load, nor do we wish to | |
2018 | * reduce the max loaded cpu below the average load, as either of these | |
2019 | * actions would just result in more rebalancing later, and ping-pong | |
2020 | * tasks around. Thus we look for the minimum possible imbalance. | |
2021 | * Negative imbalances (*we* are more loaded than anyone else) will | |
2022 | * be counted as no imbalance for these purposes -- we can't fix that | |
2023 | * by pulling tasks to us. Be careful of negative numbers as they'll | |
2024 | * appear as very large values with unsigned longs. | |
2025 | */ | |
0c117f1b SS |
2026 | |
2027 | /* Don't want to pull so many tasks that a group would go idle */ | |
2028 | max_pull = min(max_load - avg_load, max_load - SCHED_LOAD_SCALE); | |
2029 | ||
1da177e4 | 2030 | /* How much load to actually move to equalise the imbalance */ |
0c117f1b | 2031 | *imbalance = min(max_pull * busiest->cpu_power, |
1da177e4 LT |
2032 | (avg_load - this_load) * this->cpu_power) |
2033 | / SCHED_LOAD_SCALE; | |
2034 | ||
2035 | if (*imbalance < SCHED_LOAD_SCALE) { | |
2036 | unsigned long pwr_now = 0, pwr_move = 0; | |
2037 | unsigned long tmp; | |
2038 | ||
2039 | if (max_load - this_load >= SCHED_LOAD_SCALE*2) { | |
2040 | *imbalance = 1; | |
2041 | return busiest; | |
2042 | } | |
2043 | ||
2044 | /* | |
2045 | * OK, we don't have enough imbalance to justify moving tasks, | |
2046 | * however we may be able to increase total CPU power used by | |
2047 | * moving them. | |
2048 | */ | |
2049 | ||
2050 | pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load); | |
2051 | pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load); | |
2052 | pwr_now /= SCHED_LOAD_SCALE; | |
2053 | ||
2054 | /* Amount of load we'd subtract */ | |
2055 | tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power; | |
2056 | if (max_load > tmp) | |
2057 | pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE, | |
2058 | max_load - tmp); | |
2059 | ||
2060 | /* Amount of load we'd add */ | |
2061 | if (max_load*busiest->cpu_power < | |
2062 | SCHED_LOAD_SCALE*SCHED_LOAD_SCALE) | |
2063 | tmp = max_load*busiest->cpu_power/this->cpu_power; | |
2064 | else | |
2065 | tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power; | |
2066 | pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp); | |
2067 | pwr_move /= SCHED_LOAD_SCALE; | |
2068 | ||
2069 | /* Move if we gain throughput */ | |
2070 | if (pwr_move <= pwr_now) | |
2071 | goto out_balanced; | |
2072 | ||
2073 | *imbalance = 1; | |
2074 | return busiest; | |
2075 | } | |
2076 | ||
2077 | /* Get rid of the scaling factor, rounding down as we divide */ | |
2078 | *imbalance = *imbalance / SCHED_LOAD_SCALE; | |
1da177e4 LT |
2079 | return busiest; |
2080 | ||
2081 | out_balanced: | |
1da177e4 LT |
2082 | |
2083 | *imbalance = 0; | |
2084 | return NULL; | |
2085 | } | |
2086 | ||
2087 | /* | |
2088 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
2089 | */ | |
b910472d CK |
2090 | static runqueue_t *find_busiest_queue(struct sched_group *group, |
2091 | enum idle_type idle) | |
1da177e4 LT |
2092 | { |
2093 | unsigned long load, max_load = 0; | |
2094 | runqueue_t *busiest = NULL; | |
2095 | int i; | |
2096 | ||
2097 | for_each_cpu_mask(i, group->cpumask) { | |
a2000572 | 2098 | load = source_load(i, 0); |
1da177e4 LT |
2099 | |
2100 | if (load > max_load) { | |
2101 | max_load = load; | |
2102 | busiest = cpu_rq(i); | |
2103 | } | |
2104 | } | |
2105 | ||
2106 | return busiest; | |
2107 | } | |
2108 | ||
77391d71 NP |
2109 | /* |
2110 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
2111 | * so long as it is large enough. | |
2112 | */ | |
2113 | #define MAX_PINNED_INTERVAL 512 | |
2114 | ||
1da177e4 LT |
2115 | /* |
2116 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
2117 | * tasks if there is an imbalance. | |
2118 | * | |
2119 | * Called with this_rq unlocked. | |
2120 | */ | |
2121 | static int load_balance(int this_cpu, runqueue_t *this_rq, | |
2122 | struct sched_domain *sd, enum idle_type idle) | |
2123 | { | |
2124 | struct sched_group *group; | |
2125 | runqueue_t *busiest; | |
2126 | unsigned long imbalance; | |
77391d71 | 2127 | int nr_moved, all_pinned = 0; |
81026794 | 2128 | int active_balance = 0; |
5969fe06 NP |
2129 | int sd_idle = 0; |
2130 | ||
2131 | if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER) | |
2132 | sd_idle = 1; | |
1da177e4 | 2133 | |
1da177e4 LT |
2134 | schedstat_inc(sd, lb_cnt[idle]); |
2135 | ||
5969fe06 | 2136 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle); |
1da177e4 LT |
2137 | if (!group) { |
2138 | schedstat_inc(sd, lb_nobusyg[idle]); | |
2139 | goto out_balanced; | |
2140 | } | |
2141 | ||
b910472d | 2142 | busiest = find_busiest_queue(group, idle); |
1da177e4 LT |
2143 | if (!busiest) { |
2144 | schedstat_inc(sd, lb_nobusyq[idle]); | |
2145 | goto out_balanced; | |
2146 | } | |
2147 | ||
db935dbd | 2148 | BUG_ON(busiest == this_rq); |
1da177e4 LT |
2149 | |
2150 | schedstat_add(sd, lb_imbalance[idle], imbalance); | |
2151 | ||
2152 | nr_moved = 0; | |
2153 | if (busiest->nr_running > 1) { | |
2154 | /* | |
2155 | * Attempt to move tasks. If find_busiest_group has found | |
2156 | * an imbalance but busiest->nr_running <= 1, the group is | |
2157 | * still unbalanced. nr_moved simply stays zero, so it is | |
2158 | * correctly treated as an imbalance. | |
2159 | */ | |
e17224bf | 2160 | double_rq_lock(this_rq, busiest); |
1da177e4 | 2161 | nr_moved = move_tasks(this_rq, this_cpu, busiest, |
d6d5cfaf | 2162 | imbalance, sd, idle, &all_pinned); |
e17224bf | 2163 | double_rq_unlock(this_rq, busiest); |
81026794 NP |
2164 | |
2165 | /* All tasks on this runqueue were pinned by CPU affinity */ | |
2166 | if (unlikely(all_pinned)) | |
2167 | goto out_balanced; | |
1da177e4 | 2168 | } |
81026794 | 2169 | |
1da177e4 LT |
2170 | if (!nr_moved) { |
2171 | schedstat_inc(sd, lb_failed[idle]); | |
2172 | sd->nr_balance_failed++; | |
2173 | ||
2174 | if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) { | |
1da177e4 LT |
2175 | |
2176 | spin_lock(&busiest->lock); | |
fa3b6ddc SS |
2177 | |
2178 | /* don't kick the migration_thread, if the curr | |
2179 | * task on busiest cpu can't be moved to this_cpu | |
2180 | */ | |
2181 | if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) { | |
2182 | spin_unlock(&busiest->lock); | |
2183 | all_pinned = 1; | |
2184 | goto out_one_pinned; | |
2185 | } | |
2186 | ||
1da177e4 LT |
2187 | if (!busiest->active_balance) { |
2188 | busiest->active_balance = 1; | |
2189 | busiest->push_cpu = this_cpu; | |
81026794 | 2190 | active_balance = 1; |
1da177e4 LT |
2191 | } |
2192 | spin_unlock(&busiest->lock); | |
81026794 | 2193 | if (active_balance) |
1da177e4 LT |
2194 | wake_up_process(busiest->migration_thread); |
2195 | ||
2196 | /* | |
2197 | * We've kicked active balancing, reset the failure | |
2198 | * counter. | |
2199 | */ | |
39507451 | 2200 | sd->nr_balance_failed = sd->cache_nice_tries+1; |
1da177e4 | 2201 | } |
81026794 | 2202 | } else |
1da177e4 LT |
2203 | sd->nr_balance_failed = 0; |
2204 | ||
81026794 | 2205 | if (likely(!active_balance)) { |
1da177e4 LT |
2206 | /* We were unbalanced, so reset the balancing interval */ |
2207 | sd->balance_interval = sd->min_interval; | |
81026794 NP |
2208 | } else { |
2209 | /* | |
2210 | * If we've begun active balancing, start to back off. This | |
2211 | * case may not be covered by the all_pinned logic if there | |
2212 | * is only 1 task on the busy runqueue (because we don't call | |
2213 | * move_tasks). | |
2214 | */ | |
2215 | if (sd->balance_interval < sd->max_interval) | |
2216 | sd->balance_interval *= 2; | |
1da177e4 LT |
2217 | } |
2218 | ||
5969fe06 NP |
2219 | if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
2220 | return -1; | |
1da177e4 LT |
2221 | return nr_moved; |
2222 | ||
2223 | out_balanced: | |
1da177e4 LT |
2224 | schedstat_inc(sd, lb_balanced[idle]); |
2225 | ||
16cfb1c0 | 2226 | sd->nr_balance_failed = 0; |
fa3b6ddc SS |
2227 | |
2228 | out_one_pinned: | |
1da177e4 | 2229 | /* tune up the balancing interval */ |
77391d71 NP |
2230 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || |
2231 | (sd->balance_interval < sd->max_interval)) | |
1da177e4 LT |
2232 | sd->balance_interval *= 2; |
2233 | ||
5969fe06 NP |
2234 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
2235 | return -1; | |
1da177e4 LT |
2236 | return 0; |
2237 | } | |
2238 | ||
2239 | /* | |
2240 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
2241 | * tasks if there is an imbalance. | |
2242 | * | |
2243 | * Called from schedule when this_rq is about to become idle (NEWLY_IDLE). | |
2244 | * this_rq is locked. | |
2245 | */ | |
2246 | static int load_balance_newidle(int this_cpu, runqueue_t *this_rq, | |
2247 | struct sched_domain *sd) | |
2248 | { | |
2249 | struct sched_group *group; | |
2250 | runqueue_t *busiest = NULL; | |
2251 | unsigned long imbalance; | |
2252 | int nr_moved = 0; | |
5969fe06 NP |
2253 | int sd_idle = 0; |
2254 | ||
2255 | if (sd->flags & SD_SHARE_CPUPOWER) | |
2256 | sd_idle = 1; | |
1da177e4 LT |
2257 | |
2258 | schedstat_inc(sd, lb_cnt[NEWLY_IDLE]); | |
5969fe06 | 2259 | group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE, &sd_idle); |
1da177e4 | 2260 | if (!group) { |
1da177e4 | 2261 | schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]); |
16cfb1c0 | 2262 | goto out_balanced; |
1da177e4 LT |
2263 | } |
2264 | ||
b910472d | 2265 | busiest = find_busiest_queue(group, NEWLY_IDLE); |
db935dbd | 2266 | if (!busiest) { |
1da177e4 | 2267 | schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]); |
16cfb1c0 | 2268 | goto out_balanced; |
1da177e4 LT |
2269 | } |
2270 | ||
db935dbd NP |
2271 | BUG_ON(busiest == this_rq); |
2272 | ||
1da177e4 | 2273 | schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance); |
d6d5cfaf NP |
2274 | |
2275 | nr_moved = 0; | |
2276 | if (busiest->nr_running > 1) { | |
2277 | /* Attempt to move tasks */ | |
2278 | double_lock_balance(this_rq, busiest); | |
2279 | nr_moved = move_tasks(this_rq, this_cpu, busiest, | |
81026794 | 2280 | imbalance, sd, NEWLY_IDLE, NULL); |
d6d5cfaf NP |
2281 | spin_unlock(&busiest->lock); |
2282 | } | |
2283 | ||
5969fe06 | 2284 | if (!nr_moved) { |
1da177e4 | 2285 | schedstat_inc(sd, lb_failed[NEWLY_IDLE]); |
5969fe06 NP |
2286 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
2287 | return -1; | |
2288 | } else | |
16cfb1c0 | 2289 | sd->nr_balance_failed = 0; |
1da177e4 | 2290 | |
1da177e4 | 2291 | return nr_moved; |
16cfb1c0 NP |
2292 | |
2293 | out_balanced: | |
2294 | schedstat_inc(sd, lb_balanced[NEWLY_IDLE]); | |
5969fe06 NP |
2295 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
2296 | return -1; | |
16cfb1c0 NP |
2297 | sd->nr_balance_failed = 0; |
2298 | return 0; | |
1da177e4 LT |
2299 | } |
2300 | ||
2301 | /* | |
2302 | * idle_balance is called by schedule() if this_cpu is about to become | |
2303 | * idle. Attempts to pull tasks from other CPUs. | |
2304 | */ | |
858119e1 | 2305 | static void idle_balance(int this_cpu, runqueue_t *this_rq) |
1da177e4 LT |
2306 | { |
2307 | struct sched_domain *sd; | |
2308 | ||
2309 | for_each_domain(this_cpu, sd) { | |
2310 | if (sd->flags & SD_BALANCE_NEWIDLE) { | |
2311 | if (load_balance_newidle(this_cpu, this_rq, sd)) { | |
2312 | /* We've pulled tasks over so stop searching */ | |
2313 | break; | |
2314 | } | |
2315 | } | |
2316 | } | |
2317 | } | |
2318 | ||
2319 | /* | |
2320 | * active_load_balance is run by migration threads. It pushes running tasks | |
2321 | * off the busiest CPU onto idle CPUs. It requires at least 1 task to be | |
2322 | * running on each physical CPU where possible, and avoids physical / | |
2323 | * logical imbalances. | |
2324 | * | |
2325 | * Called with busiest_rq locked. | |
2326 | */ | |
2327 | static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu) | |
2328 | { | |
2329 | struct sched_domain *sd; | |
1da177e4 | 2330 | runqueue_t *target_rq; |
39507451 NP |
2331 | int target_cpu = busiest_rq->push_cpu; |
2332 | ||
2333 | if (busiest_rq->nr_running <= 1) | |
2334 | /* no task to move */ | |
2335 | return; | |
2336 | ||
2337 | target_rq = cpu_rq(target_cpu); | |
1da177e4 LT |
2338 | |
2339 | /* | |
39507451 NP |
2340 | * This condition is "impossible", if it occurs |
2341 | * we need to fix it. Originally reported by | |
2342 | * Bjorn Helgaas on a 128-cpu setup. | |
1da177e4 | 2343 | */ |
39507451 | 2344 | BUG_ON(busiest_rq == target_rq); |
1da177e4 | 2345 | |
39507451 NP |
2346 | /* move a task from busiest_rq to target_rq */ |
2347 | double_lock_balance(busiest_rq, target_rq); | |
2348 | ||
2349 | /* Search for an sd spanning us and the target CPU. */ | |
c96d145e | 2350 | for_each_domain(target_cpu, sd) { |
39507451 NP |
2351 | if ((sd->flags & SD_LOAD_BALANCE) && |
2352 | cpu_isset(busiest_cpu, sd->span)) | |
2353 | break; | |
c96d145e | 2354 | } |
39507451 NP |
2355 | |
2356 | if (unlikely(sd == NULL)) | |
2357 | goto out; | |
2358 | ||
2359 | schedstat_inc(sd, alb_cnt); | |
2360 | ||
2361 | if (move_tasks(target_rq, target_cpu, busiest_rq, 1, sd, SCHED_IDLE, NULL)) | |
2362 | schedstat_inc(sd, alb_pushed); | |
2363 | else | |
2364 | schedstat_inc(sd, alb_failed); | |
2365 | out: | |
2366 | spin_unlock(&target_rq->lock); | |
1da177e4 LT |
2367 | } |
2368 | ||
2369 | /* | |
2370 | * rebalance_tick will get called every timer tick, on every CPU. | |
2371 | * | |
2372 | * It checks each scheduling domain to see if it is due to be balanced, | |
2373 | * and initiates a balancing operation if so. | |
2374 | * | |
2375 | * Balancing parameters are set up in arch_init_sched_domains. | |
2376 | */ | |
2377 | ||
2378 | /* Don't have all balancing operations going off at once */ | |
2379 | #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS) | |
2380 | ||
2381 | static void rebalance_tick(int this_cpu, runqueue_t *this_rq, | |
2382 | enum idle_type idle) | |
2383 | { | |
2384 | unsigned long old_load, this_load; | |
2385 | unsigned long j = jiffies + CPU_OFFSET(this_cpu); | |
2386 | struct sched_domain *sd; | |
7897986b | 2387 | int i; |
1da177e4 | 2388 | |
1da177e4 | 2389 | this_load = this_rq->nr_running * SCHED_LOAD_SCALE; |
7897986b NP |
2390 | /* Update our load */ |
2391 | for (i = 0; i < 3; i++) { | |
2392 | unsigned long new_load = this_load; | |
2393 | int scale = 1 << i; | |
2394 | old_load = this_rq->cpu_load[i]; | |
2395 | /* | |
2396 | * Round up the averaging division if load is increasing. This | |
2397 | * prevents us from getting stuck on 9 if the load is 10, for | |
2398 | * example. | |
2399 | */ | |
2400 | if (new_load > old_load) | |
2401 | new_load += scale-1; | |
2402 | this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale; | |
2403 | } | |
1da177e4 LT |
2404 | |
2405 | for_each_domain(this_cpu, sd) { | |
2406 | unsigned long interval; | |
2407 | ||
2408 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
2409 | continue; | |
2410 | ||
2411 | interval = sd->balance_interval; | |
2412 | if (idle != SCHED_IDLE) | |
2413 | interval *= sd->busy_factor; | |
2414 | ||
2415 | /* scale ms to jiffies */ | |
2416 | interval = msecs_to_jiffies(interval); | |
2417 | if (unlikely(!interval)) | |
2418 | interval = 1; | |
2419 | ||
2420 | if (j - sd->last_balance >= interval) { | |
2421 | if (load_balance(this_cpu, this_rq, sd, idle)) { | |
fa3b6ddc SS |
2422 | /* |
2423 | * We've pulled tasks over so either we're no | |
5969fe06 NP |
2424 | * longer idle, or one of our SMT siblings is |
2425 | * not idle. | |
2426 | */ | |
1da177e4 LT |
2427 | idle = NOT_IDLE; |
2428 | } | |
2429 | sd->last_balance += interval; | |
2430 | } | |
2431 | } | |
2432 | } | |
2433 | #else | |
2434 | /* | |
2435 | * on UP we do not need to balance between CPUs: | |
2436 | */ | |
2437 | static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle) | |
2438 | { | |
2439 | } | |
2440 | static inline void idle_balance(int cpu, runqueue_t *rq) | |
2441 | { | |
2442 | } | |
2443 | #endif | |
2444 | ||
2445 | static inline int wake_priority_sleeper(runqueue_t *rq) | |
2446 | { | |
2447 | int ret = 0; | |
2448 | #ifdef CONFIG_SCHED_SMT | |
2449 | spin_lock(&rq->lock); | |
2450 | /* | |
2451 | * If an SMT sibling task has been put to sleep for priority | |
2452 | * reasons reschedule the idle task to see if it can now run. | |
2453 | */ | |
2454 | if (rq->nr_running) { | |
2455 | resched_task(rq->idle); | |
2456 | ret = 1; | |
2457 | } | |
2458 | spin_unlock(&rq->lock); | |
2459 | #endif | |
2460 | return ret; | |
2461 | } | |
2462 | ||
2463 | DEFINE_PER_CPU(struct kernel_stat, kstat); | |
2464 | ||
2465 | EXPORT_PER_CPU_SYMBOL(kstat); | |
2466 | ||
2467 | /* | |
2468 | * This is called on clock ticks and on context switches. | |
2469 | * Bank in p->sched_time the ns elapsed since the last tick or switch. | |
2470 | */ | |
2471 | static inline void update_cpu_clock(task_t *p, runqueue_t *rq, | |
2472 | unsigned long long now) | |
2473 | { | |
2474 | unsigned long long last = max(p->timestamp, rq->timestamp_last_tick); | |
2475 | p->sched_time += now - last; | |
2476 | } | |
2477 | ||
2478 | /* | |
2479 | * Return current->sched_time plus any more ns on the sched_clock | |
2480 | * that have not yet been banked. | |
2481 | */ | |
2482 | unsigned long long current_sched_time(const task_t *tsk) | |
2483 | { | |
2484 | unsigned long long ns; | |
2485 | unsigned long flags; | |
2486 | local_irq_save(flags); | |
2487 | ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick); | |
2488 | ns = tsk->sched_time + (sched_clock() - ns); | |
2489 | local_irq_restore(flags); | |
2490 | return ns; | |
2491 | } | |
2492 | ||
f1adad78 LT |
2493 | /* |
2494 | * We place interactive tasks back into the active array, if possible. | |
2495 | * | |
2496 | * To guarantee that this does not starve expired tasks we ignore the | |
2497 | * interactivity of a task if the first expired task had to wait more | |
2498 | * than a 'reasonable' amount of time. This deadline timeout is | |
2499 | * load-dependent, as the frequency of array switched decreases with | |
2500 | * increasing number of running tasks. We also ignore the interactivity | |
2501 | * if a better static_prio task has expired: | |
2502 | */ | |
2503 | #define EXPIRED_STARVING(rq) \ | |
2504 | ((STARVATION_LIMIT && ((rq)->expired_timestamp && \ | |
2505 | (jiffies - (rq)->expired_timestamp >= \ | |
2506 | STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \ | |
2507 | ((rq)->curr->static_prio > (rq)->best_expired_prio)) | |
2508 | ||
1da177e4 LT |
2509 | /* |
2510 | * Account user cpu time to a process. | |
2511 | * @p: the process that the cpu time gets accounted to | |
2512 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
2513 | * @cputime: the cpu time spent in user space since the last update | |
2514 | */ | |
2515 | void account_user_time(struct task_struct *p, cputime_t cputime) | |
2516 | { | |
2517 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
2518 | cputime64_t tmp; | |
2519 | ||
2520 | p->utime = cputime_add(p->utime, cputime); | |
2521 | ||
2522 | /* Add user time to cpustat. */ | |
2523 | tmp = cputime_to_cputime64(cputime); | |
2524 | if (TASK_NICE(p) > 0) | |
2525 | cpustat->nice = cputime64_add(cpustat->nice, tmp); | |
2526 | else | |
2527 | cpustat->user = cputime64_add(cpustat->user, tmp); | |
2528 | } | |
2529 | ||
2530 | /* | |
2531 | * Account system cpu time to a process. | |
2532 | * @p: the process that the cpu time gets accounted to | |
2533 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
2534 | * @cputime: the cpu time spent in kernel space since the last update | |
2535 | */ | |
2536 | void account_system_time(struct task_struct *p, int hardirq_offset, | |
2537 | cputime_t cputime) | |
2538 | { | |
2539 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
2540 | runqueue_t *rq = this_rq(); | |
2541 | cputime64_t tmp; | |
2542 | ||
2543 | p->stime = cputime_add(p->stime, cputime); | |
2544 | ||
2545 | /* Add system time to cpustat. */ | |
2546 | tmp = cputime_to_cputime64(cputime); | |
2547 | if (hardirq_count() - hardirq_offset) | |
2548 | cpustat->irq = cputime64_add(cpustat->irq, tmp); | |
2549 | else if (softirq_count()) | |
2550 | cpustat->softirq = cputime64_add(cpustat->softirq, tmp); | |
2551 | else if (p != rq->idle) | |
2552 | cpustat->system = cputime64_add(cpustat->system, tmp); | |
2553 | else if (atomic_read(&rq->nr_iowait) > 0) | |
2554 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); | |
2555 | else | |
2556 | cpustat->idle = cputime64_add(cpustat->idle, tmp); | |
2557 | /* Account for system time used */ | |
2558 | acct_update_integrals(p); | |
1da177e4 LT |
2559 | } |
2560 | ||
2561 | /* | |
2562 | * Account for involuntary wait time. | |
2563 | * @p: the process from which the cpu time has been stolen | |
2564 | * @steal: the cpu time spent in involuntary wait | |
2565 | */ | |
2566 | void account_steal_time(struct task_struct *p, cputime_t steal) | |
2567 | { | |
2568 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
2569 | cputime64_t tmp = cputime_to_cputime64(steal); | |
2570 | runqueue_t *rq = this_rq(); | |
2571 | ||
2572 | if (p == rq->idle) { | |
2573 | p->stime = cputime_add(p->stime, steal); | |
2574 | if (atomic_read(&rq->nr_iowait) > 0) | |
2575 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); | |
2576 | else | |
2577 | cpustat->idle = cputime64_add(cpustat->idle, tmp); | |
2578 | } else | |
2579 | cpustat->steal = cputime64_add(cpustat->steal, tmp); | |
2580 | } | |
2581 | ||
2582 | /* | |
2583 | * This function gets called by the timer code, with HZ frequency. | |
2584 | * We call it with interrupts disabled. | |
2585 | * | |
2586 | * It also gets called by the fork code, when changing the parent's | |
2587 | * timeslices. | |
2588 | */ | |
2589 | void scheduler_tick(void) | |
2590 | { | |
2591 | int cpu = smp_processor_id(); | |
2592 | runqueue_t *rq = this_rq(); | |
2593 | task_t *p = current; | |
2594 | unsigned long long now = sched_clock(); | |
2595 | ||
2596 | update_cpu_clock(p, rq, now); | |
2597 | ||
2598 | rq->timestamp_last_tick = now; | |
2599 | ||
2600 | if (p == rq->idle) { | |
2601 | if (wake_priority_sleeper(rq)) | |
2602 | goto out; | |
2603 | rebalance_tick(cpu, rq, SCHED_IDLE); | |
2604 | return; | |
2605 | } | |
2606 | ||
2607 | /* Task might have expired already, but not scheduled off yet */ | |
2608 | if (p->array != rq->active) { | |
2609 | set_tsk_need_resched(p); | |
2610 | goto out; | |
2611 | } | |
2612 | spin_lock(&rq->lock); | |
2613 | /* | |
2614 | * The task was running during this tick - update the | |
2615 | * time slice counter. Note: we do not update a thread's | |
2616 | * priority until it either goes to sleep or uses up its | |
2617 | * timeslice. This makes it possible for interactive tasks | |
2618 | * to use up their timeslices at their highest priority levels. | |
2619 | */ | |
2620 | if (rt_task(p)) { | |
2621 | /* | |
2622 | * RR tasks need a special form of timeslice management. | |
2623 | * FIFO tasks have no timeslices. | |
2624 | */ | |
2625 | if ((p->policy == SCHED_RR) && !--p->time_slice) { | |
2626 | p->time_slice = task_timeslice(p); | |
2627 | p->first_time_slice = 0; | |
2628 | set_tsk_need_resched(p); | |
2629 | ||
2630 | /* put it at the end of the queue: */ | |
2631 | requeue_task(p, rq->active); | |
2632 | } | |
2633 | goto out_unlock; | |
2634 | } | |
2635 | if (!--p->time_slice) { | |
2636 | dequeue_task(p, rq->active); | |
2637 | set_tsk_need_resched(p); | |
2638 | p->prio = effective_prio(p); | |
2639 | p->time_slice = task_timeslice(p); | |
2640 | p->first_time_slice = 0; | |
2641 | ||
2642 | if (!rq->expired_timestamp) | |
2643 | rq->expired_timestamp = jiffies; | |
f1adad78 | 2644 | if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) { |
1da177e4 LT |
2645 | enqueue_task(p, rq->expired); |
2646 | if (p->static_prio < rq->best_expired_prio) | |
2647 | rq->best_expired_prio = p->static_prio; | |
2648 | } else | |
2649 | enqueue_task(p, rq->active); | |
2650 | } else { | |
2651 | /* | |
2652 | * Prevent a too long timeslice allowing a task to monopolize | |
2653 | * the CPU. We do this by splitting up the timeslice into | |
2654 | * smaller pieces. | |
2655 | * | |
2656 | * Note: this does not mean the task's timeslices expire or | |
2657 | * get lost in any way, they just might be preempted by | |
2658 | * another task of equal priority. (one with higher | |
2659 | * priority would have preempted this task already.) We | |
2660 | * requeue this task to the end of the list on this priority | |
2661 | * level, which is in essence a round-robin of tasks with | |
2662 | * equal priority. | |
2663 | * | |
2664 | * This only applies to tasks in the interactive | |
2665 | * delta range with at least TIMESLICE_GRANULARITY to requeue. | |
2666 | */ | |
2667 | if (TASK_INTERACTIVE(p) && !((task_timeslice(p) - | |
2668 | p->time_slice) % TIMESLICE_GRANULARITY(p)) && | |
2669 | (p->time_slice >= TIMESLICE_GRANULARITY(p)) && | |
2670 | (p->array == rq->active)) { | |
2671 | ||
2672 | requeue_task(p, rq->active); | |
2673 | set_tsk_need_resched(p); | |
2674 | } | |
2675 | } | |
2676 | out_unlock: | |
2677 | spin_unlock(&rq->lock); | |
2678 | out: | |
2679 | rebalance_tick(cpu, rq, NOT_IDLE); | |
2680 | } | |
2681 | ||
2682 | #ifdef CONFIG_SCHED_SMT | |
fc38ed75 CK |
2683 | static inline void wakeup_busy_runqueue(runqueue_t *rq) |
2684 | { | |
2685 | /* If an SMT runqueue is sleeping due to priority reasons wake it up */ | |
2686 | if (rq->curr == rq->idle && rq->nr_running) | |
2687 | resched_task(rq->idle); | |
2688 | } | |
2689 | ||
c96d145e CK |
2690 | /* |
2691 | * Called with interrupt disabled and this_rq's runqueue locked. | |
2692 | */ | |
2693 | static void wake_sleeping_dependent(int this_cpu) | |
1da177e4 | 2694 | { |
41c7ce9a | 2695 | struct sched_domain *tmp, *sd = NULL; |
1da177e4 LT |
2696 | int i; |
2697 | ||
c96d145e CK |
2698 | for_each_domain(this_cpu, tmp) { |
2699 | if (tmp->flags & SD_SHARE_CPUPOWER) { | |
41c7ce9a | 2700 | sd = tmp; |
c96d145e CK |
2701 | break; |
2702 | } | |
2703 | } | |
41c7ce9a NP |
2704 | |
2705 | if (!sd) | |
1da177e4 LT |
2706 | return; |
2707 | ||
c96d145e | 2708 | for_each_cpu_mask(i, sd->span) { |
1da177e4 LT |
2709 | runqueue_t *smt_rq = cpu_rq(i); |
2710 | ||
c96d145e CK |
2711 | if (i == this_cpu) |
2712 | continue; | |
2713 | if (unlikely(!spin_trylock(&smt_rq->lock))) | |
2714 | continue; | |
2715 | ||
fc38ed75 | 2716 | wakeup_busy_runqueue(smt_rq); |
c96d145e | 2717 | spin_unlock(&smt_rq->lock); |
1da177e4 | 2718 | } |
1da177e4 LT |
2719 | } |
2720 | ||
67f9a619 IM |
2721 | /* |
2722 | * number of 'lost' timeslices this task wont be able to fully | |
2723 | * utilize, if another task runs on a sibling. This models the | |
2724 | * slowdown effect of other tasks running on siblings: | |
2725 | */ | |
2726 | static inline unsigned long smt_slice(task_t *p, struct sched_domain *sd) | |
2727 | { | |
2728 | return p->time_slice * (100 - sd->per_cpu_gain) / 100; | |
2729 | } | |
2730 | ||
c96d145e CK |
2731 | /* |
2732 | * To minimise lock contention and not have to drop this_rq's runlock we only | |
2733 | * trylock the sibling runqueues and bypass those runqueues if we fail to | |
2734 | * acquire their lock. As we only trylock the normal locking order does not | |
2735 | * need to be obeyed. | |
2736 | */ | |
2737 | static int dependent_sleeper(int this_cpu, runqueue_t *this_rq, task_t *p) | |
1da177e4 | 2738 | { |
41c7ce9a | 2739 | struct sched_domain *tmp, *sd = NULL; |
1da177e4 | 2740 | int ret = 0, i; |
1da177e4 | 2741 | |
c96d145e CK |
2742 | /* kernel/rt threads do not participate in dependent sleeping */ |
2743 | if (!p->mm || rt_task(p)) | |
2744 | return 0; | |
2745 | ||
2746 | for_each_domain(this_cpu, tmp) { | |
2747 | if (tmp->flags & SD_SHARE_CPUPOWER) { | |
41c7ce9a | 2748 | sd = tmp; |
c96d145e CK |
2749 | break; |
2750 | } | |
2751 | } | |
41c7ce9a NP |
2752 | |
2753 | if (!sd) | |
1da177e4 LT |
2754 | return 0; |
2755 | ||
c96d145e CK |
2756 | for_each_cpu_mask(i, sd->span) { |
2757 | runqueue_t *smt_rq; | |
2758 | task_t *smt_curr; | |
1da177e4 | 2759 | |
c96d145e CK |
2760 | if (i == this_cpu) |
2761 | continue; | |
1da177e4 | 2762 | |
c96d145e CK |
2763 | smt_rq = cpu_rq(i); |
2764 | if (unlikely(!spin_trylock(&smt_rq->lock))) | |
2765 | continue; | |
1da177e4 | 2766 | |
c96d145e | 2767 | smt_curr = smt_rq->curr; |
1da177e4 | 2768 | |
c96d145e CK |
2769 | if (!smt_curr->mm) |
2770 | goto unlock; | |
fc38ed75 | 2771 | |
1da177e4 LT |
2772 | /* |
2773 | * If a user task with lower static priority than the | |
2774 | * running task on the SMT sibling is trying to schedule, | |
2775 | * delay it till there is proportionately less timeslice | |
2776 | * left of the sibling task to prevent a lower priority | |
2777 | * task from using an unfair proportion of the | |
2778 | * physical cpu's resources. -ck | |
2779 | */ | |
fc38ed75 CK |
2780 | if (rt_task(smt_curr)) { |
2781 | /* | |
2782 | * With real time tasks we run non-rt tasks only | |
2783 | * per_cpu_gain% of the time. | |
2784 | */ | |
2785 | if ((jiffies % DEF_TIMESLICE) > | |
2786 | (sd->per_cpu_gain * DEF_TIMESLICE / 100)) | |
2787 | ret = 1; | |
c96d145e | 2788 | } else { |
67f9a619 IM |
2789 | if (smt_curr->static_prio < p->static_prio && |
2790 | !TASK_PREEMPTS_CURR(p, smt_rq) && | |
2791 | smt_slice(smt_curr, sd) > task_timeslice(p)) | |
fc38ed75 | 2792 | ret = 1; |
fc38ed75 | 2793 | } |
c96d145e CK |
2794 | unlock: |
2795 | spin_unlock(&smt_rq->lock); | |
1da177e4 | 2796 | } |
1da177e4 LT |
2797 | return ret; |
2798 | } | |
2799 | #else | |
c96d145e | 2800 | static inline void wake_sleeping_dependent(int this_cpu) |
1da177e4 LT |
2801 | { |
2802 | } | |
2803 | ||
c96d145e CK |
2804 | static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq, |
2805 | task_t *p) | |
1da177e4 LT |
2806 | { |
2807 | return 0; | |
2808 | } | |
2809 | #endif | |
2810 | ||
2811 | #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT) | |
2812 | ||
2813 | void fastcall add_preempt_count(int val) | |
2814 | { | |
2815 | /* | |
2816 | * Underflow? | |
2817 | */ | |
be5b4fbd | 2818 | BUG_ON((preempt_count() < 0)); |
1da177e4 LT |
2819 | preempt_count() += val; |
2820 | /* | |
2821 | * Spinlock count overflowing soon? | |
2822 | */ | |
2823 | BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10); | |
2824 | } | |
2825 | EXPORT_SYMBOL(add_preempt_count); | |
2826 | ||
2827 | void fastcall sub_preempt_count(int val) | |
2828 | { | |
2829 | /* | |
2830 | * Underflow? | |
2831 | */ | |
2832 | BUG_ON(val > preempt_count()); | |
2833 | /* | |
2834 | * Is the spinlock portion underflowing? | |
2835 | */ | |
2836 | BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK)); | |
2837 | preempt_count() -= val; | |
2838 | } | |
2839 | EXPORT_SYMBOL(sub_preempt_count); | |
2840 | ||
2841 | #endif | |
2842 | ||
3dee386e CK |
2843 | static inline int interactive_sleep(enum sleep_type sleep_type) |
2844 | { | |
2845 | return (sleep_type == SLEEP_INTERACTIVE || | |
2846 | sleep_type == SLEEP_INTERRUPTED); | |
2847 | } | |
2848 | ||
1da177e4 LT |
2849 | /* |
2850 | * schedule() is the main scheduler function. | |
2851 | */ | |
2852 | asmlinkage void __sched schedule(void) | |
2853 | { | |
2854 | long *switch_count; | |
2855 | task_t *prev, *next; | |
2856 | runqueue_t *rq; | |
2857 | prio_array_t *array; | |
2858 | struct list_head *queue; | |
2859 | unsigned long long now; | |
2860 | unsigned long run_time; | |
a3464a10 | 2861 | int cpu, idx, new_prio; |
1da177e4 LT |
2862 | |
2863 | /* | |
2864 | * Test if we are atomic. Since do_exit() needs to call into | |
2865 | * schedule() atomically, we ignore that path for now. | |
2866 | * Otherwise, whine if we are scheduling when we should not be. | |
2867 | */ | |
77e4bfbc AM |
2868 | if (unlikely(in_atomic() && !current->exit_state)) { |
2869 | printk(KERN_ERR "BUG: scheduling while atomic: " | |
2870 | "%s/0x%08x/%d\n", | |
2871 | current->comm, preempt_count(), current->pid); | |
2872 | dump_stack(); | |
1da177e4 LT |
2873 | } |
2874 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); | |
2875 | ||
2876 | need_resched: | |
2877 | preempt_disable(); | |
2878 | prev = current; | |
2879 | release_kernel_lock(prev); | |
2880 | need_resched_nonpreemptible: | |
2881 | rq = this_rq(); | |
2882 | ||
2883 | /* | |
2884 | * The idle thread is not allowed to schedule! | |
2885 | * Remove this check after it has been exercised a bit. | |
2886 | */ | |
2887 | if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) { | |
2888 | printk(KERN_ERR "bad: scheduling from the idle thread!\n"); | |
2889 | dump_stack(); | |
2890 | } | |
2891 | ||
2892 | schedstat_inc(rq, sched_cnt); | |
2893 | now = sched_clock(); | |
238628ed | 2894 | if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) { |
1da177e4 | 2895 | run_time = now - prev->timestamp; |
238628ed | 2896 | if (unlikely((long long)(now - prev->timestamp) < 0)) |
1da177e4 LT |
2897 | run_time = 0; |
2898 | } else | |
2899 | run_time = NS_MAX_SLEEP_AVG; | |
2900 | ||
2901 | /* | |
2902 | * Tasks charged proportionately less run_time at high sleep_avg to | |
2903 | * delay them losing their interactive status | |
2904 | */ | |
2905 | run_time /= (CURRENT_BONUS(prev) ? : 1); | |
2906 | ||
2907 | spin_lock_irq(&rq->lock); | |
2908 | ||
2909 | if (unlikely(prev->flags & PF_DEAD)) | |
2910 | prev->state = EXIT_DEAD; | |
2911 | ||
2912 | switch_count = &prev->nivcsw; | |
2913 | if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { | |
2914 | switch_count = &prev->nvcsw; | |
2915 | if (unlikely((prev->state & TASK_INTERRUPTIBLE) && | |
2916 | unlikely(signal_pending(prev)))) | |
2917 | prev->state = TASK_RUNNING; | |
2918 | else { | |
2919 | if (prev->state == TASK_UNINTERRUPTIBLE) | |
2920 | rq->nr_uninterruptible++; | |
2921 | deactivate_task(prev, rq); | |
2922 | } | |
2923 | } | |
2924 | ||
2925 | cpu = smp_processor_id(); | |
2926 | if (unlikely(!rq->nr_running)) { | |
1da177e4 LT |
2927 | idle_balance(cpu, rq); |
2928 | if (!rq->nr_running) { | |
2929 | next = rq->idle; | |
2930 | rq->expired_timestamp = 0; | |
c96d145e | 2931 | wake_sleeping_dependent(cpu); |
1da177e4 LT |
2932 | goto switch_tasks; |
2933 | } | |
1da177e4 LT |
2934 | } |
2935 | ||
2936 | array = rq->active; | |
2937 | if (unlikely(!array->nr_active)) { | |
2938 | /* | |
2939 | * Switch the active and expired arrays. | |
2940 | */ | |
2941 | schedstat_inc(rq, sched_switch); | |
2942 | rq->active = rq->expired; | |
2943 | rq->expired = array; | |
2944 | array = rq->active; | |
2945 | rq->expired_timestamp = 0; | |
2946 | rq->best_expired_prio = MAX_PRIO; | |
2947 | } | |
2948 | ||
2949 | idx = sched_find_first_bit(array->bitmap); | |
2950 | queue = array->queue + idx; | |
2951 | next = list_entry(queue->next, task_t, run_list); | |
2952 | ||
3dee386e | 2953 | if (!rt_task(next) && interactive_sleep(next->sleep_type)) { |
1da177e4 | 2954 | unsigned long long delta = now - next->timestamp; |
238628ed | 2955 | if (unlikely((long long)(now - next->timestamp) < 0)) |
1da177e4 LT |
2956 | delta = 0; |
2957 | ||
3dee386e | 2958 | if (next->sleep_type == SLEEP_INTERACTIVE) |
1da177e4 LT |
2959 | delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128; |
2960 | ||
2961 | array = next->array; | |
a3464a10 CS |
2962 | new_prio = recalc_task_prio(next, next->timestamp + delta); |
2963 | ||
2964 | if (unlikely(next->prio != new_prio)) { | |
2965 | dequeue_task(next, array); | |
2966 | next->prio = new_prio; | |
2967 | enqueue_task(next, array); | |
7c4bb1f9 | 2968 | } |
1da177e4 | 2969 | } |
3dee386e | 2970 | next->sleep_type = SLEEP_NORMAL; |
c96d145e CK |
2971 | if (dependent_sleeper(cpu, rq, next)) |
2972 | next = rq->idle; | |
1da177e4 LT |
2973 | switch_tasks: |
2974 | if (next == rq->idle) | |
2975 | schedstat_inc(rq, sched_goidle); | |
2976 | prefetch(next); | |
383f2835 | 2977 | prefetch_stack(next); |
1da177e4 LT |
2978 | clear_tsk_need_resched(prev); |
2979 | rcu_qsctr_inc(task_cpu(prev)); | |
2980 | ||
2981 | update_cpu_clock(prev, rq, now); | |
2982 | ||
2983 | prev->sleep_avg -= run_time; | |
2984 | if ((long)prev->sleep_avg <= 0) | |
2985 | prev->sleep_avg = 0; | |
2986 | prev->timestamp = prev->last_ran = now; | |
2987 | ||
2988 | sched_info_switch(prev, next); | |
2989 | if (likely(prev != next)) { | |
2990 | next->timestamp = now; | |
2991 | rq->nr_switches++; | |
2992 | rq->curr = next; | |
2993 | ++*switch_count; | |
2994 | ||
4866cde0 | 2995 | prepare_task_switch(rq, next); |
1da177e4 LT |
2996 | prev = context_switch(rq, prev, next); |
2997 | barrier(); | |
4866cde0 NP |
2998 | /* |
2999 | * this_rq must be evaluated again because prev may have moved | |
3000 | * CPUs since it called schedule(), thus the 'rq' on its stack | |
3001 | * frame will be invalid. | |
3002 | */ | |
3003 | finish_task_switch(this_rq(), prev); | |
1da177e4 LT |
3004 | } else |
3005 | spin_unlock_irq(&rq->lock); | |
3006 | ||
3007 | prev = current; | |
3008 | if (unlikely(reacquire_kernel_lock(prev) < 0)) | |
3009 | goto need_resched_nonpreemptible; | |
3010 | preempt_enable_no_resched(); | |
3011 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
3012 | goto need_resched; | |
3013 | } | |
3014 | ||
3015 | EXPORT_SYMBOL(schedule); | |
3016 | ||
3017 | #ifdef CONFIG_PREEMPT | |
3018 | /* | |
3019 | * this is is the entry point to schedule() from in-kernel preemption | |
3020 | * off of preempt_enable. Kernel preemptions off return from interrupt | |
3021 | * occur there and call schedule directly. | |
3022 | */ | |
3023 | asmlinkage void __sched preempt_schedule(void) | |
3024 | { | |
3025 | struct thread_info *ti = current_thread_info(); | |
3026 | #ifdef CONFIG_PREEMPT_BKL | |
3027 | struct task_struct *task = current; | |
3028 | int saved_lock_depth; | |
3029 | #endif | |
3030 | /* | |
3031 | * If there is a non-zero preempt_count or interrupts are disabled, | |
3032 | * we do not want to preempt the current task. Just return.. | |
3033 | */ | |
3034 | if (unlikely(ti->preempt_count || irqs_disabled())) | |
3035 | return; | |
3036 | ||
3037 | need_resched: | |
3038 | add_preempt_count(PREEMPT_ACTIVE); | |
3039 | /* | |
3040 | * We keep the big kernel semaphore locked, but we | |
3041 | * clear ->lock_depth so that schedule() doesnt | |
3042 | * auto-release the semaphore: | |
3043 | */ | |
3044 | #ifdef CONFIG_PREEMPT_BKL | |
3045 | saved_lock_depth = task->lock_depth; | |
3046 | task->lock_depth = -1; | |
3047 | #endif | |
3048 | schedule(); | |
3049 | #ifdef CONFIG_PREEMPT_BKL | |
3050 | task->lock_depth = saved_lock_depth; | |
3051 | #endif | |
3052 | sub_preempt_count(PREEMPT_ACTIVE); | |
3053 | ||
3054 | /* we could miss a preemption opportunity between schedule and now */ | |
3055 | barrier(); | |
3056 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
3057 | goto need_resched; | |
3058 | } | |
3059 | ||
3060 | EXPORT_SYMBOL(preempt_schedule); | |
3061 | ||
3062 | /* | |
3063 | * this is is the entry point to schedule() from kernel preemption | |
3064 | * off of irq context. | |
3065 | * Note, that this is called and return with irqs disabled. This will | |
3066 | * protect us against recursive calling from irq. | |
3067 | */ | |
3068 | asmlinkage void __sched preempt_schedule_irq(void) | |
3069 | { | |
3070 | struct thread_info *ti = current_thread_info(); | |
3071 | #ifdef CONFIG_PREEMPT_BKL | |
3072 | struct task_struct *task = current; | |
3073 | int saved_lock_depth; | |
3074 | #endif | |
3075 | /* Catch callers which need to be fixed*/ | |
3076 | BUG_ON(ti->preempt_count || !irqs_disabled()); | |
3077 | ||
3078 | need_resched: | |
3079 | add_preempt_count(PREEMPT_ACTIVE); | |
3080 | /* | |
3081 | * We keep the big kernel semaphore locked, but we | |
3082 | * clear ->lock_depth so that schedule() doesnt | |
3083 | * auto-release the semaphore: | |
3084 | */ | |
3085 | #ifdef CONFIG_PREEMPT_BKL | |
3086 | saved_lock_depth = task->lock_depth; | |
3087 | task->lock_depth = -1; | |
3088 | #endif | |
3089 | local_irq_enable(); | |
3090 | schedule(); | |
3091 | local_irq_disable(); | |
3092 | #ifdef CONFIG_PREEMPT_BKL | |
3093 | task->lock_depth = saved_lock_depth; | |
3094 | #endif | |
3095 | sub_preempt_count(PREEMPT_ACTIVE); | |
3096 | ||
3097 | /* we could miss a preemption opportunity between schedule and now */ | |
3098 | barrier(); | |
3099 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
3100 | goto need_resched; | |
3101 | } | |
3102 | ||
3103 | #endif /* CONFIG_PREEMPT */ | |
3104 | ||
95cdf3b7 IM |
3105 | int default_wake_function(wait_queue_t *curr, unsigned mode, int sync, |
3106 | void *key) | |
1da177e4 | 3107 | { |
c43dc2fd | 3108 | task_t *p = curr->private; |
1da177e4 LT |
3109 | return try_to_wake_up(p, mode, sync); |
3110 | } | |
3111 | ||
3112 | EXPORT_SYMBOL(default_wake_function); | |
3113 | ||
3114 | /* | |
3115 | * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just | |
3116 | * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve | |
3117 | * number) then we wake all the non-exclusive tasks and one exclusive task. | |
3118 | * | |
3119 | * There are circumstances in which we can try to wake a task which has already | |
3120 | * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns | |
3121 | * zero in this (rare) case, and we handle it by continuing to scan the queue. | |
3122 | */ | |
3123 | static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, | |
3124 | int nr_exclusive, int sync, void *key) | |
3125 | { | |
3126 | struct list_head *tmp, *next; | |
3127 | ||
3128 | list_for_each_safe(tmp, next, &q->task_list) { | |
3129 | wait_queue_t *curr; | |
3130 | unsigned flags; | |
3131 | curr = list_entry(tmp, wait_queue_t, task_list); | |
3132 | flags = curr->flags; | |
3133 | if (curr->func(curr, mode, sync, key) && | |
3134 | (flags & WQ_FLAG_EXCLUSIVE) && | |
3135 | !--nr_exclusive) | |
3136 | break; | |
3137 | } | |
3138 | } | |
3139 | ||
3140 | /** | |
3141 | * __wake_up - wake up threads blocked on a waitqueue. | |
3142 | * @q: the waitqueue | |
3143 | * @mode: which threads | |
3144 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
67be2dd1 | 3145 | * @key: is directly passed to the wakeup function |
1da177e4 LT |
3146 | */ |
3147 | void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode, | |
95cdf3b7 | 3148 | int nr_exclusive, void *key) |
1da177e4 LT |
3149 | { |
3150 | unsigned long flags; | |
3151 | ||
3152 | spin_lock_irqsave(&q->lock, flags); | |
3153 | __wake_up_common(q, mode, nr_exclusive, 0, key); | |
3154 | spin_unlock_irqrestore(&q->lock, flags); | |
3155 | } | |
3156 | ||
3157 | EXPORT_SYMBOL(__wake_up); | |
3158 | ||
3159 | /* | |
3160 | * Same as __wake_up but called with the spinlock in wait_queue_head_t held. | |
3161 | */ | |
3162 | void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode) | |
3163 | { | |
3164 | __wake_up_common(q, mode, 1, 0, NULL); | |
3165 | } | |
3166 | ||
3167 | /** | |
67be2dd1 | 3168 | * __wake_up_sync - wake up threads blocked on a waitqueue. |
1da177e4 LT |
3169 | * @q: the waitqueue |
3170 | * @mode: which threads | |
3171 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
3172 | * | |
3173 | * The sync wakeup differs that the waker knows that it will schedule | |
3174 | * away soon, so while the target thread will be woken up, it will not | |
3175 | * be migrated to another CPU - ie. the two threads are 'synchronized' | |
3176 | * with each other. This can prevent needless bouncing between CPUs. | |
3177 | * | |
3178 | * On UP it can prevent extra preemption. | |
3179 | */ | |
95cdf3b7 IM |
3180 | void fastcall |
3181 | __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) | |
1da177e4 LT |
3182 | { |
3183 | unsigned long flags; | |
3184 | int sync = 1; | |
3185 | ||
3186 | if (unlikely(!q)) | |
3187 | return; | |
3188 | ||
3189 | if (unlikely(!nr_exclusive)) | |
3190 | sync = 0; | |
3191 | ||
3192 | spin_lock_irqsave(&q->lock, flags); | |
3193 | __wake_up_common(q, mode, nr_exclusive, sync, NULL); | |
3194 | spin_unlock_irqrestore(&q->lock, flags); | |
3195 | } | |
3196 | EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ | |
3197 | ||
3198 | void fastcall complete(struct completion *x) | |
3199 | { | |
3200 | unsigned long flags; | |
3201 | ||
3202 | spin_lock_irqsave(&x->wait.lock, flags); | |
3203 | x->done++; | |
3204 | __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, | |
3205 | 1, 0, NULL); | |
3206 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
3207 | } | |
3208 | EXPORT_SYMBOL(complete); | |
3209 | ||
3210 | void fastcall complete_all(struct completion *x) | |
3211 | { | |
3212 | unsigned long flags; | |
3213 | ||
3214 | spin_lock_irqsave(&x->wait.lock, flags); | |
3215 | x->done += UINT_MAX/2; | |
3216 | __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, | |
3217 | 0, 0, NULL); | |
3218 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
3219 | } | |
3220 | EXPORT_SYMBOL(complete_all); | |
3221 | ||
3222 | void fastcall __sched wait_for_completion(struct completion *x) | |
3223 | { | |
3224 | might_sleep(); | |
3225 | spin_lock_irq(&x->wait.lock); | |
3226 | if (!x->done) { | |
3227 | DECLARE_WAITQUEUE(wait, current); | |
3228 | ||
3229 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3230 | __add_wait_queue_tail(&x->wait, &wait); | |
3231 | do { | |
3232 | __set_current_state(TASK_UNINTERRUPTIBLE); | |
3233 | spin_unlock_irq(&x->wait.lock); | |
3234 | schedule(); | |
3235 | spin_lock_irq(&x->wait.lock); | |
3236 | } while (!x->done); | |
3237 | __remove_wait_queue(&x->wait, &wait); | |
3238 | } | |
3239 | x->done--; | |
3240 | spin_unlock_irq(&x->wait.lock); | |
3241 | } | |
3242 | EXPORT_SYMBOL(wait_for_completion); | |
3243 | ||
3244 | unsigned long fastcall __sched | |
3245 | wait_for_completion_timeout(struct completion *x, unsigned long timeout) | |
3246 | { | |
3247 | might_sleep(); | |
3248 | ||
3249 | spin_lock_irq(&x->wait.lock); | |
3250 | if (!x->done) { | |
3251 | DECLARE_WAITQUEUE(wait, current); | |
3252 | ||
3253 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3254 | __add_wait_queue_tail(&x->wait, &wait); | |
3255 | do { | |
3256 | __set_current_state(TASK_UNINTERRUPTIBLE); | |
3257 | spin_unlock_irq(&x->wait.lock); | |
3258 | timeout = schedule_timeout(timeout); | |
3259 | spin_lock_irq(&x->wait.lock); | |
3260 | if (!timeout) { | |
3261 | __remove_wait_queue(&x->wait, &wait); | |
3262 | goto out; | |
3263 | } | |
3264 | } while (!x->done); | |
3265 | __remove_wait_queue(&x->wait, &wait); | |
3266 | } | |
3267 | x->done--; | |
3268 | out: | |
3269 | spin_unlock_irq(&x->wait.lock); | |
3270 | return timeout; | |
3271 | } | |
3272 | EXPORT_SYMBOL(wait_for_completion_timeout); | |
3273 | ||
3274 | int fastcall __sched wait_for_completion_interruptible(struct completion *x) | |
3275 | { | |
3276 | int ret = 0; | |
3277 | ||
3278 | might_sleep(); | |
3279 | ||
3280 | spin_lock_irq(&x->wait.lock); | |
3281 | if (!x->done) { | |
3282 | DECLARE_WAITQUEUE(wait, current); | |
3283 | ||
3284 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3285 | __add_wait_queue_tail(&x->wait, &wait); | |
3286 | do { | |
3287 | if (signal_pending(current)) { | |
3288 | ret = -ERESTARTSYS; | |
3289 | __remove_wait_queue(&x->wait, &wait); | |
3290 | goto out; | |
3291 | } | |
3292 | __set_current_state(TASK_INTERRUPTIBLE); | |
3293 | spin_unlock_irq(&x->wait.lock); | |
3294 | schedule(); | |
3295 | spin_lock_irq(&x->wait.lock); | |
3296 | } while (!x->done); | |
3297 | __remove_wait_queue(&x->wait, &wait); | |
3298 | } | |
3299 | x->done--; | |
3300 | out: | |
3301 | spin_unlock_irq(&x->wait.lock); | |
3302 | ||
3303 | return ret; | |
3304 | } | |
3305 | EXPORT_SYMBOL(wait_for_completion_interruptible); | |
3306 | ||
3307 | unsigned long fastcall __sched | |
3308 | wait_for_completion_interruptible_timeout(struct completion *x, | |
3309 | unsigned long timeout) | |
3310 | { | |
3311 | might_sleep(); | |
3312 | ||
3313 | spin_lock_irq(&x->wait.lock); | |
3314 | if (!x->done) { | |
3315 | DECLARE_WAITQUEUE(wait, current); | |
3316 | ||
3317 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3318 | __add_wait_queue_tail(&x->wait, &wait); | |
3319 | do { | |
3320 | if (signal_pending(current)) { | |
3321 | timeout = -ERESTARTSYS; | |
3322 | __remove_wait_queue(&x->wait, &wait); | |
3323 | goto out; | |
3324 | } | |
3325 | __set_current_state(TASK_INTERRUPTIBLE); | |
3326 | spin_unlock_irq(&x->wait.lock); | |
3327 | timeout = schedule_timeout(timeout); | |
3328 | spin_lock_irq(&x->wait.lock); | |
3329 | if (!timeout) { | |
3330 | __remove_wait_queue(&x->wait, &wait); | |
3331 | goto out; | |
3332 | } | |
3333 | } while (!x->done); | |
3334 | __remove_wait_queue(&x->wait, &wait); | |
3335 | } | |
3336 | x->done--; | |
3337 | out: | |
3338 | spin_unlock_irq(&x->wait.lock); | |
3339 | return timeout; | |
3340 | } | |
3341 | EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); | |
3342 | ||
3343 | ||
3344 | #define SLEEP_ON_VAR \ | |
3345 | unsigned long flags; \ | |
3346 | wait_queue_t wait; \ | |
3347 | init_waitqueue_entry(&wait, current); | |
3348 | ||
3349 | #define SLEEP_ON_HEAD \ | |
3350 | spin_lock_irqsave(&q->lock,flags); \ | |
3351 | __add_wait_queue(q, &wait); \ | |
3352 | spin_unlock(&q->lock); | |
3353 | ||
3354 | #define SLEEP_ON_TAIL \ | |
3355 | spin_lock_irq(&q->lock); \ | |
3356 | __remove_wait_queue(q, &wait); \ | |
3357 | spin_unlock_irqrestore(&q->lock, flags); | |
3358 | ||
3359 | void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q) | |
3360 | { | |
3361 | SLEEP_ON_VAR | |
3362 | ||
3363 | current->state = TASK_INTERRUPTIBLE; | |
3364 | ||
3365 | SLEEP_ON_HEAD | |
3366 | schedule(); | |
3367 | SLEEP_ON_TAIL | |
3368 | } | |
3369 | ||
3370 | EXPORT_SYMBOL(interruptible_sleep_on); | |
3371 | ||
95cdf3b7 IM |
3372 | long fastcall __sched |
3373 | interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) | |
1da177e4 LT |
3374 | { |
3375 | SLEEP_ON_VAR | |
3376 | ||
3377 | current->state = TASK_INTERRUPTIBLE; | |
3378 | ||
3379 | SLEEP_ON_HEAD | |
3380 | timeout = schedule_timeout(timeout); | |
3381 | SLEEP_ON_TAIL | |
3382 | ||
3383 | return timeout; | |
3384 | } | |
3385 | ||
3386 | EXPORT_SYMBOL(interruptible_sleep_on_timeout); | |
3387 | ||
3388 | void fastcall __sched sleep_on(wait_queue_head_t *q) | |
3389 | { | |
3390 | SLEEP_ON_VAR | |
3391 | ||
3392 | current->state = TASK_UNINTERRUPTIBLE; | |
3393 | ||
3394 | SLEEP_ON_HEAD | |
3395 | schedule(); | |
3396 | SLEEP_ON_TAIL | |
3397 | } | |
3398 | ||
3399 | EXPORT_SYMBOL(sleep_on); | |
3400 | ||
3401 | long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) | |
3402 | { | |
3403 | SLEEP_ON_VAR | |
3404 | ||
3405 | current->state = TASK_UNINTERRUPTIBLE; | |
3406 | ||
3407 | SLEEP_ON_HEAD | |
3408 | timeout = schedule_timeout(timeout); | |
3409 | SLEEP_ON_TAIL | |
3410 | ||
3411 | return timeout; | |
3412 | } | |
3413 | ||
3414 | EXPORT_SYMBOL(sleep_on_timeout); | |
3415 | ||
3416 | void set_user_nice(task_t *p, long nice) | |
3417 | { | |
3418 | unsigned long flags; | |
3419 | prio_array_t *array; | |
3420 | runqueue_t *rq; | |
3421 | int old_prio, new_prio, delta; | |
3422 | ||
3423 | if (TASK_NICE(p) == nice || nice < -20 || nice > 19) | |
3424 | return; | |
3425 | /* | |
3426 | * We have to be careful, if called from sys_setpriority(), | |
3427 | * the task might be in the middle of scheduling on another CPU. | |
3428 | */ | |
3429 | rq = task_rq_lock(p, &flags); | |
3430 | /* | |
3431 | * The RT priorities are set via sched_setscheduler(), but we still | |
3432 | * allow the 'normal' nice value to be set - but as expected | |
3433 | * it wont have any effect on scheduling until the task is | |
b0a9499c | 3434 | * not SCHED_NORMAL/SCHED_BATCH: |
1da177e4 LT |
3435 | */ |
3436 | if (rt_task(p)) { | |
3437 | p->static_prio = NICE_TO_PRIO(nice); | |
3438 | goto out_unlock; | |
3439 | } | |
3440 | array = p->array; | |
a2000572 | 3441 | if (array) |
1da177e4 LT |
3442 | dequeue_task(p, array); |
3443 | ||
3444 | old_prio = p->prio; | |
3445 | new_prio = NICE_TO_PRIO(nice); | |
3446 | delta = new_prio - old_prio; | |
3447 | p->static_prio = NICE_TO_PRIO(nice); | |
3448 | p->prio += delta; | |
3449 | ||
3450 | if (array) { | |
3451 | enqueue_task(p, array); | |
3452 | /* | |
3453 | * If the task increased its priority or is running and | |
3454 | * lowered its priority, then reschedule its CPU: | |
3455 | */ | |
3456 | if (delta < 0 || (delta > 0 && task_running(rq, p))) | |
3457 | resched_task(rq->curr); | |
3458 | } | |
3459 | out_unlock: | |
3460 | task_rq_unlock(rq, &flags); | |
3461 | } | |
3462 | ||
3463 | EXPORT_SYMBOL(set_user_nice); | |
3464 | ||
e43379f1 MM |
3465 | /* |
3466 | * can_nice - check if a task can reduce its nice value | |
3467 | * @p: task | |
3468 | * @nice: nice value | |
3469 | */ | |
3470 | int can_nice(const task_t *p, const int nice) | |
3471 | { | |
024f4747 MM |
3472 | /* convert nice value [19,-20] to rlimit style value [1,40] */ |
3473 | int nice_rlim = 20 - nice; | |
e43379f1 MM |
3474 | return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur || |
3475 | capable(CAP_SYS_NICE)); | |
3476 | } | |
3477 | ||
1da177e4 LT |
3478 | #ifdef __ARCH_WANT_SYS_NICE |
3479 | ||
3480 | /* | |
3481 | * sys_nice - change the priority of the current process. | |
3482 | * @increment: priority increment | |
3483 | * | |
3484 | * sys_setpriority is a more generic, but much slower function that | |
3485 | * does similar things. | |
3486 | */ | |
3487 | asmlinkage long sys_nice(int increment) | |
3488 | { | |
3489 | int retval; | |
3490 | long nice; | |
3491 | ||
3492 | /* | |
3493 | * Setpriority might change our priority at the same moment. | |
3494 | * We don't have to worry. Conceptually one call occurs first | |
3495 | * and we have a single winner. | |
3496 | */ | |
e43379f1 MM |
3497 | if (increment < -40) |
3498 | increment = -40; | |
1da177e4 LT |
3499 | if (increment > 40) |
3500 | increment = 40; | |
3501 | ||
3502 | nice = PRIO_TO_NICE(current->static_prio) + increment; | |
3503 | if (nice < -20) | |
3504 | nice = -20; | |
3505 | if (nice > 19) | |
3506 | nice = 19; | |
3507 | ||
e43379f1 MM |
3508 | if (increment < 0 && !can_nice(current, nice)) |
3509 | return -EPERM; | |
3510 | ||
1da177e4 LT |
3511 | retval = security_task_setnice(current, nice); |
3512 | if (retval) | |
3513 | return retval; | |
3514 | ||
3515 | set_user_nice(current, nice); | |
3516 | return 0; | |
3517 | } | |
3518 | ||
3519 | #endif | |
3520 | ||
3521 | /** | |
3522 | * task_prio - return the priority value of a given task. | |
3523 | * @p: the task in question. | |
3524 | * | |
3525 | * This is the priority value as seen by users in /proc. | |
3526 | * RT tasks are offset by -200. Normal tasks are centered | |
3527 | * around 0, value goes from -16 to +15. | |
3528 | */ | |
3529 | int task_prio(const task_t *p) | |
3530 | { | |
3531 | return p->prio - MAX_RT_PRIO; | |
3532 | } | |
3533 | ||
3534 | /** | |
3535 | * task_nice - return the nice value of a given task. | |
3536 | * @p: the task in question. | |
3537 | */ | |
3538 | int task_nice(const task_t *p) | |
3539 | { | |
3540 | return TASK_NICE(p); | |
3541 | } | |
1da177e4 | 3542 | EXPORT_SYMBOL_GPL(task_nice); |
1da177e4 LT |
3543 | |
3544 | /** | |
3545 | * idle_cpu - is a given cpu idle currently? | |
3546 | * @cpu: the processor in question. | |
3547 | */ | |
3548 | int idle_cpu(int cpu) | |
3549 | { | |
3550 | return cpu_curr(cpu) == cpu_rq(cpu)->idle; | |
3551 | } | |
3552 | ||
1da177e4 LT |
3553 | /** |
3554 | * idle_task - return the idle task for a given cpu. | |
3555 | * @cpu: the processor in question. | |
3556 | */ | |
3557 | task_t *idle_task(int cpu) | |
3558 | { | |
3559 | return cpu_rq(cpu)->idle; | |
3560 | } | |
3561 | ||
3562 | /** | |
3563 | * find_process_by_pid - find a process with a matching PID value. | |
3564 | * @pid: the pid in question. | |
3565 | */ | |
3566 | static inline task_t *find_process_by_pid(pid_t pid) | |
3567 | { | |
3568 | return pid ? find_task_by_pid(pid) : current; | |
3569 | } | |
3570 | ||
3571 | /* Actually do priority change: must hold rq lock. */ | |
3572 | static void __setscheduler(struct task_struct *p, int policy, int prio) | |
3573 | { | |
3574 | BUG_ON(p->array); | |
3575 | p->policy = policy; | |
3576 | p->rt_priority = prio; | |
b0a9499c | 3577 | if (policy != SCHED_NORMAL && policy != SCHED_BATCH) { |
d46523ea | 3578 | p->prio = MAX_RT_PRIO-1 - p->rt_priority; |
b0a9499c | 3579 | } else { |
1da177e4 | 3580 | p->prio = p->static_prio; |
b0a9499c IM |
3581 | /* |
3582 | * SCHED_BATCH tasks are treated as perpetual CPU hogs: | |
3583 | */ | |
3584 | if (policy == SCHED_BATCH) | |
3585 | p->sleep_avg = 0; | |
3586 | } | |
1da177e4 LT |
3587 | } |
3588 | ||
3589 | /** | |
3590 | * sched_setscheduler - change the scheduling policy and/or RT priority of | |
3591 | * a thread. | |
3592 | * @p: the task in question. | |
3593 | * @policy: new policy. | |
3594 | * @param: structure containing the new RT priority. | |
3595 | */ | |
95cdf3b7 IM |
3596 | int sched_setscheduler(struct task_struct *p, int policy, |
3597 | struct sched_param *param) | |
1da177e4 LT |
3598 | { |
3599 | int retval; | |
3600 | int oldprio, oldpolicy = -1; | |
3601 | prio_array_t *array; | |
3602 | unsigned long flags; | |
3603 | runqueue_t *rq; | |
3604 | ||
3605 | recheck: | |
3606 | /* double check policy once rq lock held */ | |
3607 | if (policy < 0) | |
3608 | policy = oldpolicy = p->policy; | |
3609 | else if (policy != SCHED_FIFO && policy != SCHED_RR && | |
b0a9499c IM |
3610 | policy != SCHED_NORMAL && policy != SCHED_BATCH) |
3611 | return -EINVAL; | |
1da177e4 LT |
3612 | /* |
3613 | * Valid priorities for SCHED_FIFO and SCHED_RR are | |
b0a9499c IM |
3614 | * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and |
3615 | * SCHED_BATCH is 0. | |
1da177e4 LT |
3616 | */ |
3617 | if (param->sched_priority < 0 || | |
95cdf3b7 | 3618 | (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || |
d46523ea | 3619 | (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) |
1da177e4 | 3620 | return -EINVAL; |
b0a9499c IM |
3621 | if ((policy == SCHED_NORMAL || policy == SCHED_BATCH) |
3622 | != (param->sched_priority == 0)) | |
1da177e4 LT |
3623 | return -EINVAL; |
3624 | ||
37e4ab3f OC |
3625 | /* |
3626 | * Allow unprivileged RT tasks to decrease priority: | |
3627 | */ | |
3628 | if (!capable(CAP_SYS_NICE)) { | |
b0a9499c IM |
3629 | /* |
3630 | * can't change policy, except between SCHED_NORMAL | |
3631 | * and SCHED_BATCH: | |
3632 | */ | |
3633 | if (((policy != SCHED_NORMAL && p->policy != SCHED_BATCH) && | |
3634 | (policy != SCHED_BATCH && p->policy != SCHED_NORMAL)) && | |
3635 | !p->signal->rlim[RLIMIT_RTPRIO].rlim_cur) | |
37e4ab3f OC |
3636 | return -EPERM; |
3637 | /* can't increase priority */ | |
b0a9499c | 3638 | if ((policy != SCHED_NORMAL && policy != SCHED_BATCH) && |
37e4ab3f OC |
3639 | param->sched_priority > p->rt_priority && |
3640 | param->sched_priority > | |
3641 | p->signal->rlim[RLIMIT_RTPRIO].rlim_cur) | |
3642 | return -EPERM; | |
3643 | /* can't change other user's priorities */ | |
3644 | if ((current->euid != p->euid) && | |
3645 | (current->euid != p->uid)) | |
3646 | return -EPERM; | |
3647 | } | |
1da177e4 LT |
3648 | |
3649 | retval = security_task_setscheduler(p, policy, param); | |
3650 | if (retval) | |
3651 | return retval; | |
3652 | /* | |
3653 | * To be able to change p->policy safely, the apropriate | |
3654 | * runqueue lock must be held. | |
3655 | */ | |
3656 | rq = task_rq_lock(p, &flags); | |
3657 | /* recheck policy now with rq lock held */ | |
3658 | if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { | |
3659 | policy = oldpolicy = -1; | |
3660 | task_rq_unlock(rq, &flags); | |
3661 | goto recheck; | |
3662 | } | |
3663 | array = p->array; | |
3664 | if (array) | |
3665 | deactivate_task(p, rq); | |
3666 | oldprio = p->prio; | |
3667 | __setscheduler(p, policy, param->sched_priority); | |
3668 | if (array) { | |
3669 | __activate_task(p, rq); | |
3670 | /* | |
3671 | * Reschedule if we are currently running on this runqueue and | |
3672 | * our priority decreased, or if we are not currently running on | |
3673 | * this runqueue and our priority is higher than the current's | |
3674 | */ | |
3675 | if (task_running(rq, p)) { | |
3676 | if (p->prio > oldprio) | |
3677 | resched_task(rq->curr); | |
3678 | } else if (TASK_PREEMPTS_CURR(p, rq)) | |
3679 | resched_task(rq->curr); | |
3680 | } | |
3681 | task_rq_unlock(rq, &flags); | |
3682 | return 0; | |
3683 | } | |
3684 | EXPORT_SYMBOL_GPL(sched_setscheduler); | |
3685 | ||
95cdf3b7 IM |
3686 | static int |
3687 | do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) | |
1da177e4 LT |
3688 | { |
3689 | int retval; | |
3690 | struct sched_param lparam; | |
3691 | struct task_struct *p; | |
3692 | ||
3693 | if (!param || pid < 0) | |
3694 | return -EINVAL; | |
3695 | if (copy_from_user(&lparam, param, sizeof(struct sched_param))) | |
3696 | return -EFAULT; | |
3697 | read_lock_irq(&tasklist_lock); | |
3698 | p = find_process_by_pid(pid); | |
3699 | if (!p) { | |
3700 | read_unlock_irq(&tasklist_lock); | |
3701 | return -ESRCH; | |
3702 | } | |
3703 | retval = sched_setscheduler(p, policy, &lparam); | |
3704 | read_unlock_irq(&tasklist_lock); | |
3705 | return retval; | |
3706 | } | |
3707 | ||
3708 | /** | |
3709 | * sys_sched_setscheduler - set/change the scheduler policy and RT priority | |
3710 | * @pid: the pid in question. | |
3711 | * @policy: new policy. | |
3712 | * @param: structure containing the new RT priority. | |
3713 | */ | |
3714 | asmlinkage long sys_sched_setscheduler(pid_t pid, int policy, | |
3715 | struct sched_param __user *param) | |
3716 | { | |
c21761f1 JB |
3717 | /* negative values for policy are not valid */ |
3718 | if (policy < 0) | |
3719 | return -EINVAL; | |
3720 | ||
1da177e4 LT |
3721 | return do_sched_setscheduler(pid, policy, param); |
3722 | } | |
3723 | ||
3724 | /** | |
3725 | * sys_sched_setparam - set/change the RT priority of a thread | |
3726 | * @pid: the pid in question. | |
3727 | * @param: structure containing the new RT priority. | |
3728 | */ | |
3729 | asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param) | |
3730 | { | |
3731 | return do_sched_setscheduler(pid, -1, param); | |
3732 | } | |
3733 | ||
3734 | /** | |
3735 | * sys_sched_getscheduler - get the policy (scheduling class) of a thread | |
3736 | * @pid: the pid in question. | |
3737 | */ | |
3738 | asmlinkage long sys_sched_getscheduler(pid_t pid) | |
3739 | { | |
3740 | int retval = -EINVAL; | |
3741 | task_t *p; | |
3742 | ||
3743 | if (pid < 0) | |
3744 | goto out_nounlock; | |
3745 | ||
3746 | retval = -ESRCH; | |
3747 | read_lock(&tasklist_lock); | |
3748 | p = find_process_by_pid(pid); | |
3749 | if (p) { | |
3750 | retval = security_task_getscheduler(p); | |
3751 | if (!retval) | |
3752 | retval = p->policy; | |
3753 | } | |
3754 | read_unlock(&tasklist_lock); | |
3755 | ||
3756 | out_nounlock: | |
3757 | return retval; | |
3758 | } | |
3759 | ||
3760 | /** | |
3761 | * sys_sched_getscheduler - get the RT priority of a thread | |
3762 | * @pid: the pid in question. | |
3763 | * @param: structure containing the RT priority. | |
3764 | */ | |
3765 | asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param) | |
3766 | { | |
3767 | struct sched_param lp; | |
3768 | int retval = -EINVAL; | |
3769 | task_t *p; | |
3770 | ||
3771 | if (!param || pid < 0) | |
3772 | goto out_nounlock; | |
3773 | ||
3774 | read_lock(&tasklist_lock); | |
3775 | p = find_process_by_pid(pid); | |
3776 | retval = -ESRCH; | |
3777 | if (!p) | |
3778 | goto out_unlock; | |
3779 | ||
3780 | retval = security_task_getscheduler(p); | |
3781 | if (retval) | |
3782 | goto out_unlock; | |
3783 | ||
3784 | lp.sched_priority = p->rt_priority; | |
3785 | read_unlock(&tasklist_lock); | |
3786 | ||
3787 | /* | |
3788 | * This one might sleep, we cannot do it with a spinlock held ... | |
3789 | */ | |
3790 | retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; | |
3791 | ||
3792 | out_nounlock: | |
3793 | return retval; | |
3794 | ||
3795 | out_unlock: | |
3796 | read_unlock(&tasklist_lock); | |
3797 | return retval; | |
3798 | } | |
3799 | ||
3800 | long sched_setaffinity(pid_t pid, cpumask_t new_mask) | |
3801 | { | |
3802 | task_t *p; | |
3803 | int retval; | |
3804 | cpumask_t cpus_allowed; | |
3805 | ||
3806 | lock_cpu_hotplug(); | |
3807 | read_lock(&tasklist_lock); | |
3808 | ||
3809 | p = find_process_by_pid(pid); | |
3810 | if (!p) { | |
3811 | read_unlock(&tasklist_lock); | |
3812 | unlock_cpu_hotplug(); | |
3813 | return -ESRCH; | |
3814 | } | |
3815 | ||
3816 | /* | |
3817 | * It is not safe to call set_cpus_allowed with the | |
3818 | * tasklist_lock held. We will bump the task_struct's | |
3819 | * usage count and then drop tasklist_lock. | |
3820 | */ | |
3821 | get_task_struct(p); | |
3822 | read_unlock(&tasklist_lock); | |
3823 | ||
3824 | retval = -EPERM; | |
3825 | if ((current->euid != p->euid) && (current->euid != p->uid) && | |
3826 | !capable(CAP_SYS_NICE)) | |
3827 | goto out_unlock; | |
3828 | ||
e7834f8f DQ |
3829 | retval = security_task_setscheduler(p, 0, NULL); |
3830 | if (retval) | |
3831 | goto out_unlock; | |
3832 | ||
1da177e4 LT |
3833 | cpus_allowed = cpuset_cpus_allowed(p); |
3834 | cpus_and(new_mask, new_mask, cpus_allowed); | |
3835 | retval = set_cpus_allowed(p, new_mask); | |
3836 | ||
3837 | out_unlock: | |
3838 | put_task_struct(p); | |
3839 | unlock_cpu_hotplug(); | |
3840 | return retval; | |
3841 | } | |
3842 | ||
3843 | static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, | |
3844 | cpumask_t *new_mask) | |
3845 | { | |
3846 | if (len < sizeof(cpumask_t)) { | |
3847 | memset(new_mask, 0, sizeof(cpumask_t)); | |
3848 | } else if (len > sizeof(cpumask_t)) { | |
3849 | len = sizeof(cpumask_t); | |
3850 | } | |
3851 | return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; | |
3852 | } | |
3853 | ||
3854 | /** | |
3855 | * sys_sched_setaffinity - set the cpu affinity of a process | |
3856 | * @pid: pid of the process | |
3857 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
3858 | * @user_mask_ptr: user-space pointer to the new cpu mask | |
3859 | */ | |
3860 | asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len, | |
3861 | unsigned long __user *user_mask_ptr) | |
3862 | { | |
3863 | cpumask_t new_mask; | |
3864 | int retval; | |
3865 | ||
3866 | retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask); | |
3867 | if (retval) | |
3868 | return retval; | |
3869 | ||
3870 | return sched_setaffinity(pid, new_mask); | |
3871 | } | |
3872 | ||
3873 | /* | |
3874 | * Represents all cpu's present in the system | |
3875 | * In systems capable of hotplug, this map could dynamically grow | |
3876 | * as new cpu's are detected in the system via any platform specific | |
3877 | * method, such as ACPI for e.g. | |
3878 | */ | |
3879 | ||
4cef0c61 | 3880 | cpumask_t cpu_present_map __read_mostly; |
1da177e4 LT |
3881 | EXPORT_SYMBOL(cpu_present_map); |
3882 | ||
3883 | #ifndef CONFIG_SMP | |
4cef0c61 AK |
3884 | cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL; |
3885 | cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL; | |
1da177e4 LT |
3886 | #endif |
3887 | ||
3888 | long sched_getaffinity(pid_t pid, cpumask_t *mask) | |
3889 | { | |
3890 | int retval; | |
3891 | task_t *p; | |
3892 | ||
3893 | lock_cpu_hotplug(); | |
3894 | read_lock(&tasklist_lock); | |
3895 | ||
3896 | retval = -ESRCH; | |
3897 | p = find_process_by_pid(pid); | |
3898 | if (!p) | |
3899 | goto out_unlock; | |
3900 | ||
e7834f8f DQ |
3901 | retval = security_task_getscheduler(p); |
3902 | if (retval) | |
3903 | goto out_unlock; | |
3904 | ||
2f7016d9 | 3905 | cpus_and(*mask, p->cpus_allowed, cpu_online_map); |
1da177e4 LT |
3906 | |
3907 | out_unlock: | |
3908 | read_unlock(&tasklist_lock); | |
3909 | unlock_cpu_hotplug(); | |
3910 | if (retval) | |
3911 | return retval; | |
3912 | ||
3913 | return 0; | |
3914 | } | |
3915 | ||
3916 | /** | |
3917 | * sys_sched_getaffinity - get the cpu affinity of a process | |
3918 | * @pid: pid of the process | |
3919 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
3920 | * @user_mask_ptr: user-space pointer to hold the current cpu mask | |
3921 | */ | |
3922 | asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len, | |
3923 | unsigned long __user *user_mask_ptr) | |
3924 | { | |
3925 | int ret; | |
3926 | cpumask_t mask; | |
3927 | ||
3928 | if (len < sizeof(cpumask_t)) | |
3929 | return -EINVAL; | |
3930 | ||
3931 | ret = sched_getaffinity(pid, &mask); | |
3932 | if (ret < 0) | |
3933 | return ret; | |
3934 | ||
3935 | if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t))) | |
3936 | return -EFAULT; | |
3937 | ||
3938 | return sizeof(cpumask_t); | |
3939 | } | |
3940 | ||
3941 | /** | |
3942 | * sys_sched_yield - yield the current processor to other threads. | |
3943 | * | |
3944 | * this function yields the current CPU by moving the calling thread | |
3945 | * to the expired array. If there are no other threads running on this | |
3946 | * CPU then this function will return. | |
3947 | */ | |
3948 | asmlinkage long sys_sched_yield(void) | |
3949 | { | |
3950 | runqueue_t *rq = this_rq_lock(); | |
3951 | prio_array_t *array = current->array; | |
3952 | prio_array_t *target = rq->expired; | |
3953 | ||
3954 | schedstat_inc(rq, yld_cnt); | |
3955 | /* | |
3956 | * We implement yielding by moving the task into the expired | |
3957 | * queue. | |
3958 | * | |
3959 | * (special rule: RT tasks will just roundrobin in the active | |
3960 | * array.) | |
3961 | */ | |
3962 | if (rt_task(current)) | |
3963 | target = rq->active; | |
3964 | ||
5927ad78 | 3965 | if (array->nr_active == 1) { |
1da177e4 LT |
3966 | schedstat_inc(rq, yld_act_empty); |
3967 | if (!rq->expired->nr_active) | |
3968 | schedstat_inc(rq, yld_both_empty); | |
3969 | } else if (!rq->expired->nr_active) | |
3970 | schedstat_inc(rq, yld_exp_empty); | |
3971 | ||
3972 | if (array != target) { | |
3973 | dequeue_task(current, array); | |
3974 | enqueue_task(current, target); | |
3975 | } else | |
3976 | /* | |
3977 | * requeue_task is cheaper so perform that if possible. | |
3978 | */ | |
3979 | requeue_task(current, array); | |
3980 | ||
3981 | /* | |
3982 | * Since we are going to call schedule() anyway, there's | |
3983 | * no need to preempt or enable interrupts: | |
3984 | */ | |
3985 | __release(rq->lock); | |
3986 | _raw_spin_unlock(&rq->lock); | |
3987 | preempt_enable_no_resched(); | |
3988 | ||
3989 | schedule(); | |
3990 | ||
3991 | return 0; | |
3992 | } | |
3993 | ||
3994 | static inline void __cond_resched(void) | |
3995 | { | |
8e0a43d8 IM |
3996 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP |
3997 | __might_sleep(__FILE__, __LINE__); | |
3998 | #endif | |
5bbcfd90 IM |
3999 | /* |
4000 | * The BKS might be reacquired before we have dropped | |
4001 | * PREEMPT_ACTIVE, which could trigger a second | |
4002 | * cond_resched() call. | |
4003 | */ | |
4004 | if (unlikely(preempt_count())) | |
4005 | return; | |
8ba7b0a1 LT |
4006 | if (unlikely(system_state != SYSTEM_RUNNING)) |
4007 | return; | |
1da177e4 LT |
4008 | do { |
4009 | add_preempt_count(PREEMPT_ACTIVE); | |
4010 | schedule(); | |
4011 | sub_preempt_count(PREEMPT_ACTIVE); | |
4012 | } while (need_resched()); | |
4013 | } | |
4014 | ||
4015 | int __sched cond_resched(void) | |
4016 | { | |
4017 | if (need_resched()) { | |
4018 | __cond_resched(); | |
4019 | return 1; | |
4020 | } | |
4021 | return 0; | |
4022 | } | |
4023 | ||
4024 | EXPORT_SYMBOL(cond_resched); | |
4025 | ||
4026 | /* | |
4027 | * cond_resched_lock() - if a reschedule is pending, drop the given lock, | |
4028 | * call schedule, and on return reacquire the lock. | |
4029 | * | |
4030 | * This works OK both with and without CONFIG_PREEMPT. We do strange low-level | |
4031 | * operations here to prevent schedule() from being called twice (once via | |
4032 | * spin_unlock(), once by hand). | |
4033 | */ | |
95cdf3b7 | 4034 | int cond_resched_lock(spinlock_t *lock) |
1da177e4 | 4035 | { |
6df3cecb JK |
4036 | int ret = 0; |
4037 | ||
1da177e4 LT |
4038 | if (need_lockbreak(lock)) { |
4039 | spin_unlock(lock); | |
4040 | cpu_relax(); | |
6df3cecb | 4041 | ret = 1; |
1da177e4 LT |
4042 | spin_lock(lock); |
4043 | } | |
4044 | if (need_resched()) { | |
4045 | _raw_spin_unlock(lock); | |
4046 | preempt_enable_no_resched(); | |
4047 | __cond_resched(); | |
6df3cecb | 4048 | ret = 1; |
1da177e4 | 4049 | spin_lock(lock); |
1da177e4 | 4050 | } |
6df3cecb | 4051 | return ret; |
1da177e4 LT |
4052 | } |
4053 | ||
4054 | EXPORT_SYMBOL(cond_resched_lock); | |
4055 | ||
4056 | int __sched cond_resched_softirq(void) | |
4057 | { | |
4058 | BUG_ON(!in_softirq()); | |
4059 | ||
4060 | if (need_resched()) { | |
4061 | __local_bh_enable(); | |
4062 | __cond_resched(); | |
4063 | local_bh_disable(); | |
4064 | return 1; | |
4065 | } | |
4066 | return 0; | |
4067 | } | |
4068 | ||
4069 | EXPORT_SYMBOL(cond_resched_softirq); | |
4070 | ||
4071 | ||
4072 | /** | |
4073 | * yield - yield the current processor to other threads. | |
4074 | * | |
4075 | * this is a shortcut for kernel-space yielding - it marks the | |
4076 | * thread runnable and calls sys_sched_yield(). | |
4077 | */ | |
4078 | void __sched yield(void) | |
4079 | { | |
4080 | set_current_state(TASK_RUNNING); | |
4081 | sys_sched_yield(); | |
4082 | } | |
4083 | ||
4084 | EXPORT_SYMBOL(yield); | |
4085 | ||
4086 | /* | |
4087 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so | |
4088 | * that process accounting knows that this is a task in IO wait state. | |
4089 | * | |
4090 | * But don't do that if it is a deliberate, throttling IO wait (this task | |
4091 | * has set its backing_dev_info: the queue against which it should throttle) | |
4092 | */ | |
4093 | void __sched io_schedule(void) | |
4094 | { | |
bfe5d834 | 4095 | struct runqueue *rq = &__raw_get_cpu_var(runqueues); |
1da177e4 LT |
4096 | |
4097 | atomic_inc(&rq->nr_iowait); | |
4098 | schedule(); | |
4099 | atomic_dec(&rq->nr_iowait); | |
4100 | } | |
4101 | ||
4102 | EXPORT_SYMBOL(io_schedule); | |
4103 | ||
4104 | long __sched io_schedule_timeout(long timeout) | |
4105 | { | |
bfe5d834 | 4106 | struct runqueue *rq = &__raw_get_cpu_var(runqueues); |
1da177e4 LT |
4107 | long ret; |
4108 | ||
4109 | atomic_inc(&rq->nr_iowait); | |
4110 | ret = schedule_timeout(timeout); | |
4111 | atomic_dec(&rq->nr_iowait); | |
4112 | return ret; | |
4113 | } | |
4114 | ||
4115 | /** | |
4116 | * sys_sched_get_priority_max - return maximum RT priority. | |
4117 | * @policy: scheduling class. | |
4118 | * | |
4119 | * this syscall returns the maximum rt_priority that can be used | |
4120 | * by a given scheduling class. | |
4121 | */ | |
4122 | asmlinkage long sys_sched_get_priority_max(int policy) | |
4123 | { | |
4124 | int ret = -EINVAL; | |
4125 | ||
4126 | switch (policy) { | |
4127 | case SCHED_FIFO: | |
4128 | case SCHED_RR: | |
4129 | ret = MAX_USER_RT_PRIO-1; | |
4130 | break; | |
4131 | case SCHED_NORMAL: | |
b0a9499c | 4132 | case SCHED_BATCH: |
1da177e4 LT |
4133 | ret = 0; |
4134 | break; | |
4135 | } | |
4136 | return ret; | |
4137 | } | |
4138 | ||
4139 | /** | |
4140 | * sys_sched_get_priority_min - return minimum RT priority. | |
4141 | * @policy: scheduling class. | |
4142 | * | |
4143 | * this syscall returns the minimum rt_priority that can be used | |
4144 | * by a given scheduling class. | |
4145 | */ | |
4146 | asmlinkage long sys_sched_get_priority_min(int policy) | |
4147 | { | |
4148 | int ret = -EINVAL; | |
4149 | ||
4150 | switch (policy) { | |
4151 | case SCHED_FIFO: | |
4152 | case SCHED_RR: | |
4153 | ret = 1; | |
4154 | break; | |
4155 | case SCHED_NORMAL: | |
b0a9499c | 4156 | case SCHED_BATCH: |
1da177e4 LT |
4157 | ret = 0; |
4158 | } | |
4159 | return ret; | |
4160 | } | |
4161 | ||
4162 | /** | |
4163 | * sys_sched_rr_get_interval - return the default timeslice of a process. | |
4164 | * @pid: pid of the process. | |
4165 | * @interval: userspace pointer to the timeslice value. | |
4166 | * | |
4167 | * this syscall writes the default timeslice value of a given process | |
4168 | * into the user-space timespec buffer. A value of '0' means infinity. | |
4169 | */ | |
4170 | asmlinkage | |
4171 | long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval) | |
4172 | { | |
4173 | int retval = -EINVAL; | |
4174 | struct timespec t; | |
4175 | task_t *p; | |
4176 | ||
4177 | if (pid < 0) | |
4178 | goto out_nounlock; | |
4179 | ||
4180 | retval = -ESRCH; | |
4181 | read_lock(&tasklist_lock); | |
4182 | p = find_process_by_pid(pid); | |
4183 | if (!p) | |
4184 | goto out_unlock; | |
4185 | ||
4186 | retval = security_task_getscheduler(p); | |
4187 | if (retval) | |
4188 | goto out_unlock; | |
4189 | ||
b78709cf | 4190 | jiffies_to_timespec(p->policy == SCHED_FIFO ? |
1da177e4 LT |
4191 | 0 : task_timeslice(p), &t); |
4192 | read_unlock(&tasklist_lock); | |
4193 | retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; | |
4194 | out_nounlock: | |
4195 | return retval; | |
4196 | out_unlock: | |
4197 | read_unlock(&tasklist_lock); | |
4198 | return retval; | |
4199 | } | |
4200 | ||
4201 | static inline struct task_struct *eldest_child(struct task_struct *p) | |
4202 | { | |
4203 | if (list_empty(&p->children)) return NULL; | |
4204 | return list_entry(p->children.next,struct task_struct,sibling); | |
4205 | } | |
4206 | ||
4207 | static inline struct task_struct *older_sibling(struct task_struct *p) | |
4208 | { | |
4209 | if (p->sibling.prev==&p->parent->children) return NULL; | |
4210 | return list_entry(p->sibling.prev,struct task_struct,sibling); | |
4211 | } | |
4212 | ||
4213 | static inline struct task_struct *younger_sibling(struct task_struct *p) | |
4214 | { | |
4215 | if (p->sibling.next==&p->parent->children) return NULL; | |
4216 | return list_entry(p->sibling.next,struct task_struct,sibling); | |
4217 | } | |
4218 | ||
95cdf3b7 | 4219 | static void show_task(task_t *p) |
1da177e4 LT |
4220 | { |
4221 | task_t *relative; | |
4222 | unsigned state; | |
4223 | unsigned long free = 0; | |
4224 | static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" }; | |
4225 | ||
4226 | printk("%-13.13s ", p->comm); | |
4227 | state = p->state ? __ffs(p->state) + 1 : 0; | |
4228 | if (state < ARRAY_SIZE(stat_nam)) | |
4229 | printk(stat_nam[state]); | |
4230 | else | |
4231 | printk("?"); | |
4232 | #if (BITS_PER_LONG == 32) | |
4233 | if (state == TASK_RUNNING) | |
4234 | printk(" running "); | |
4235 | else | |
4236 | printk(" %08lX ", thread_saved_pc(p)); | |
4237 | #else | |
4238 | if (state == TASK_RUNNING) | |
4239 | printk(" running task "); | |
4240 | else | |
4241 | printk(" %016lx ", thread_saved_pc(p)); | |
4242 | #endif | |
4243 | #ifdef CONFIG_DEBUG_STACK_USAGE | |
4244 | { | |
10ebffde | 4245 | unsigned long *n = end_of_stack(p); |
1da177e4 LT |
4246 | while (!*n) |
4247 | n++; | |
10ebffde | 4248 | free = (unsigned long)n - (unsigned long)end_of_stack(p); |
1da177e4 LT |
4249 | } |
4250 | #endif | |
4251 | printk("%5lu %5d %6d ", free, p->pid, p->parent->pid); | |
4252 | if ((relative = eldest_child(p))) | |
4253 | printk("%5d ", relative->pid); | |
4254 | else | |
4255 | printk(" "); | |
4256 | if ((relative = younger_sibling(p))) | |
4257 | printk("%7d", relative->pid); | |
4258 | else | |
4259 | printk(" "); | |
4260 | if ((relative = older_sibling(p))) | |
4261 | printk(" %5d", relative->pid); | |
4262 | else | |
4263 | printk(" "); | |
4264 | if (!p->mm) | |
4265 | printk(" (L-TLB)\n"); | |
4266 | else | |
4267 | printk(" (NOTLB)\n"); | |
4268 | ||
4269 | if (state != TASK_RUNNING) | |
4270 | show_stack(p, NULL); | |
4271 | } | |
4272 | ||
4273 | void show_state(void) | |
4274 | { | |
4275 | task_t *g, *p; | |
4276 | ||
4277 | #if (BITS_PER_LONG == 32) | |
4278 | printk("\n" | |
4279 | " sibling\n"); | |
4280 | printk(" task PC pid father child younger older\n"); | |
4281 | #else | |
4282 | printk("\n" | |
4283 | " sibling\n"); | |
4284 | printk(" task PC pid father child younger older\n"); | |
4285 | #endif | |
4286 | read_lock(&tasklist_lock); | |
4287 | do_each_thread(g, p) { | |
4288 | /* | |
4289 | * reset the NMI-timeout, listing all files on a slow | |
4290 | * console might take alot of time: | |
4291 | */ | |
4292 | touch_nmi_watchdog(); | |
4293 | show_task(p); | |
4294 | } while_each_thread(g, p); | |
4295 | ||
4296 | read_unlock(&tasklist_lock); | |
de5097c2 | 4297 | mutex_debug_show_all_locks(); |
1da177e4 LT |
4298 | } |
4299 | ||
f340c0d1 IM |
4300 | /** |
4301 | * init_idle - set up an idle thread for a given CPU | |
4302 | * @idle: task in question | |
4303 | * @cpu: cpu the idle task belongs to | |
4304 | * | |
4305 | * NOTE: this function does not set the idle thread's NEED_RESCHED | |
4306 | * flag, to make booting more robust. | |
4307 | */ | |
1da177e4 LT |
4308 | void __devinit init_idle(task_t *idle, int cpu) |
4309 | { | |
4310 | runqueue_t *rq = cpu_rq(cpu); | |
4311 | unsigned long flags; | |
4312 | ||
81c29a85 | 4313 | idle->timestamp = sched_clock(); |
1da177e4 LT |
4314 | idle->sleep_avg = 0; |
4315 | idle->array = NULL; | |
4316 | idle->prio = MAX_PRIO; | |
4317 | idle->state = TASK_RUNNING; | |
4318 | idle->cpus_allowed = cpumask_of_cpu(cpu); | |
4319 | set_task_cpu(idle, cpu); | |
4320 | ||
4321 | spin_lock_irqsave(&rq->lock, flags); | |
4322 | rq->curr = rq->idle = idle; | |
4866cde0 NP |
4323 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
4324 | idle->oncpu = 1; | |
4325 | #endif | |
1da177e4 LT |
4326 | spin_unlock_irqrestore(&rq->lock, flags); |
4327 | ||
4328 | /* Set the preempt count _outside_ the spinlocks! */ | |
4329 | #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL) | |
a1261f54 | 4330 | task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0); |
1da177e4 | 4331 | #else |
a1261f54 | 4332 | task_thread_info(idle)->preempt_count = 0; |
1da177e4 LT |
4333 | #endif |
4334 | } | |
4335 | ||
4336 | /* | |
4337 | * In a system that switches off the HZ timer nohz_cpu_mask | |
4338 | * indicates which cpus entered this state. This is used | |
4339 | * in the rcu update to wait only for active cpus. For system | |
4340 | * which do not switch off the HZ timer nohz_cpu_mask should | |
4341 | * always be CPU_MASK_NONE. | |
4342 | */ | |
4343 | cpumask_t nohz_cpu_mask = CPU_MASK_NONE; | |
4344 | ||
4345 | #ifdef CONFIG_SMP | |
4346 | /* | |
4347 | * This is how migration works: | |
4348 | * | |
4349 | * 1) we queue a migration_req_t structure in the source CPU's | |
4350 | * runqueue and wake up that CPU's migration thread. | |
4351 | * 2) we down() the locked semaphore => thread blocks. | |
4352 | * 3) migration thread wakes up (implicitly it forces the migrated | |
4353 | * thread off the CPU) | |
4354 | * 4) it gets the migration request and checks whether the migrated | |
4355 | * task is still in the wrong runqueue. | |
4356 | * 5) if it's in the wrong runqueue then the migration thread removes | |
4357 | * it and puts it into the right queue. | |
4358 | * 6) migration thread up()s the semaphore. | |
4359 | * 7) we wake up and the migration is done. | |
4360 | */ | |
4361 | ||
4362 | /* | |
4363 | * Change a given task's CPU affinity. Migrate the thread to a | |
4364 | * proper CPU and schedule it away if the CPU it's executing on | |
4365 | * is removed from the allowed bitmask. | |
4366 | * | |
4367 | * NOTE: the caller must have a valid reference to the task, the | |
4368 | * task must not exit() & deallocate itself prematurely. The | |
4369 | * call is not atomic; no spinlocks may be held. | |
4370 | */ | |
4371 | int set_cpus_allowed(task_t *p, cpumask_t new_mask) | |
4372 | { | |
4373 | unsigned long flags; | |
4374 | int ret = 0; | |
4375 | migration_req_t req; | |
4376 | runqueue_t *rq; | |
4377 | ||
4378 | rq = task_rq_lock(p, &flags); | |
4379 | if (!cpus_intersects(new_mask, cpu_online_map)) { | |
4380 | ret = -EINVAL; | |
4381 | goto out; | |
4382 | } | |
4383 | ||
4384 | p->cpus_allowed = new_mask; | |
4385 | /* Can the task run on the task's current CPU? If so, we're done */ | |
4386 | if (cpu_isset(task_cpu(p), new_mask)) | |
4387 | goto out; | |
4388 | ||
4389 | if (migrate_task(p, any_online_cpu(new_mask), &req)) { | |
4390 | /* Need help from migration thread: drop lock and wait. */ | |
4391 | task_rq_unlock(rq, &flags); | |
4392 | wake_up_process(rq->migration_thread); | |
4393 | wait_for_completion(&req.done); | |
4394 | tlb_migrate_finish(p->mm); | |
4395 | return 0; | |
4396 | } | |
4397 | out: | |
4398 | task_rq_unlock(rq, &flags); | |
4399 | return ret; | |
4400 | } | |
4401 | ||
4402 | EXPORT_SYMBOL_GPL(set_cpus_allowed); | |
4403 | ||
4404 | /* | |
4405 | * Move (not current) task off this cpu, onto dest cpu. We're doing | |
4406 | * this because either it can't run here any more (set_cpus_allowed() | |
4407 | * away from this CPU, or CPU going down), or because we're | |
4408 | * attempting to rebalance this task on exec (sched_exec). | |
4409 | * | |
4410 | * So we race with normal scheduler movements, but that's OK, as long | |
4411 | * as the task is no longer on this CPU. | |
4412 | */ | |
4413 | static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) | |
4414 | { | |
4415 | runqueue_t *rq_dest, *rq_src; | |
4416 | ||
4417 | if (unlikely(cpu_is_offline(dest_cpu))) | |
4418 | return; | |
4419 | ||
4420 | rq_src = cpu_rq(src_cpu); | |
4421 | rq_dest = cpu_rq(dest_cpu); | |
4422 | ||
4423 | double_rq_lock(rq_src, rq_dest); | |
4424 | /* Already moved. */ | |
4425 | if (task_cpu(p) != src_cpu) | |
4426 | goto out; | |
4427 | /* Affinity changed (again). */ | |
4428 | if (!cpu_isset(dest_cpu, p->cpus_allowed)) | |
4429 | goto out; | |
4430 | ||
4431 | set_task_cpu(p, dest_cpu); | |
4432 | if (p->array) { | |
4433 | /* | |
4434 | * Sync timestamp with rq_dest's before activating. | |
4435 | * The same thing could be achieved by doing this step | |
4436 | * afterwards, and pretending it was a local activate. | |
4437 | * This way is cleaner and logically correct. | |
4438 | */ | |
4439 | p->timestamp = p->timestamp - rq_src->timestamp_last_tick | |
4440 | + rq_dest->timestamp_last_tick; | |
4441 | deactivate_task(p, rq_src); | |
4442 | activate_task(p, rq_dest, 0); | |
4443 | if (TASK_PREEMPTS_CURR(p, rq_dest)) | |
4444 | resched_task(rq_dest->curr); | |
4445 | } | |
4446 | ||
4447 | out: | |
4448 | double_rq_unlock(rq_src, rq_dest); | |
4449 | } | |
4450 | ||
4451 | /* | |
4452 | * migration_thread - this is a highprio system thread that performs | |
4453 | * thread migration by bumping thread off CPU then 'pushing' onto | |
4454 | * another runqueue. | |
4455 | */ | |
95cdf3b7 | 4456 | static int migration_thread(void *data) |
1da177e4 LT |
4457 | { |
4458 | runqueue_t *rq; | |
4459 | int cpu = (long)data; | |
4460 | ||
4461 | rq = cpu_rq(cpu); | |
4462 | BUG_ON(rq->migration_thread != current); | |
4463 | ||
4464 | set_current_state(TASK_INTERRUPTIBLE); | |
4465 | while (!kthread_should_stop()) { | |
4466 | struct list_head *head; | |
4467 | migration_req_t *req; | |
4468 | ||
3e1d1d28 | 4469 | try_to_freeze(); |
1da177e4 LT |
4470 | |
4471 | spin_lock_irq(&rq->lock); | |
4472 | ||
4473 | if (cpu_is_offline(cpu)) { | |
4474 | spin_unlock_irq(&rq->lock); | |
4475 | goto wait_to_die; | |
4476 | } | |
4477 | ||
4478 | if (rq->active_balance) { | |
4479 | active_load_balance(rq, cpu); | |
4480 | rq->active_balance = 0; | |
4481 | } | |
4482 | ||
4483 | head = &rq->migration_queue; | |
4484 | ||
4485 | if (list_empty(head)) { | |
4486 | spin_unlock_irq(&rq->lock); | |
4487 | schedule(); | |
4488 | set_current_state(TASK_INTERRUPTIBLE); | |
4489 | continue; | |
4490 | } | |
4491 | req = list_entry(head->next, migration_req_t, list); | |
4492 | list_del_init(head->next); | |
4493 | ||
674311d5 NP |
4494 | spin_unlock(&rq->lock); |
4495 | __migrate_task(req->task, cpu, req->dest_cpu); | |
4496 | local_irq_enable(); | |
1da177e4 LT |
4497 | |
4498 | complete(&req->done); | |
4499 | } | |
4500 | __set_current_state(TASK_RUNNING); | |
4501 | return 0; | |
4502 | ||
4503 | wait_to_die: | |
4504 | /* Wait for kthread_stop */ | |
4505 | set_current_state(TASK_INTERRUPTIBLE); | |
4506 | while (!kthread_should_stop()) { | |
4507 | schedule(); | |
4508 | set_current_state(TASK_INTERRUPTIBLE); | |
4509 | } | |
4510 | __set_current_state(TASK_RUNNING); | |
4511 | return 0; | |
4512 | } | |
4513 | ||
4514 | #ifdef CONFIG_HOTPLUG_CPU | |
4515 | /* Figure out where task on dead CPU should go, use force if neccessary. */ | |
4516 | static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk) | |
4517 | { | |
4518 | int dest_cpu; | |
4519 | cpumask_t mask; | |
4520 | ||
4521 | /* On same node? */ | |
4522 | mask = node_to_cpumask(cpu_to_node(dead_cpu)); | |
4523 | cpus_and(mask, mask, tsk->cpus_allowed); | |
4524 | dest_cpu = any_online_cpu(mask); | |
4525 | ||
4526 | /* On any allowed CPU? */ | |
4527 | if (dest_cpu == NR_CPUS) | |
4528 | dest_cpu = any_online_cpu(tsk->cpus_allowed); | |
4529 | ||
4530 | /* No more Mr. Nice Guy. */ | |
4531 | if (dest_cpu == NR_CPUS) { | |
b39c4fab | 4532 | cpus_setall(tsk->cpus_allowed); |
1da177e4 LT |
4533 | dest_cpu = any_online_cpu(tsk->cpus_allowed); |
4534 | ||
4535 | /* | |
4536 | * Don't tell them about moving exiting tasks or | |
4537 | * kernel threads (both mm NULL), since they never | |
4538 | * leave kernel. | |
4539 | */ | |
4540 | if (tsk->mm && printk_ratelimit()) | |
4541 | printk(KERN_INFO "process %d (%s) no " | |
4542 | "longer affine to cpu%d\n", | |
4543 | tsk->pid, tsk->comm, dead_cpu); | |
4544 | } | |
4545 | __migrate_task(tsk, dead_cpu, dest_cpu); | |
4546 | } | |
4547 | ||
4548 | /* | |
4549 | * While a dead CPU has no uninterruptible tasks queued at this point, | |
4550 | * it might still have a nonzero ->nr_uninterruptible counter, because | |
4551 | * for performance reasons the counter is not stricly tracking tasks to | |
4552 | * their home CPUs. So we just add the counter to another CPU's counter, | |
4553 | * to keep the global sum constant after CPU-down: | |
4554 | */ | |
4555 | static void migrate_nr_uninterruptible(runqueue_t *rq_src) | |
4556 | { | |
4557 | runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL)); | |
4558 | unsigned long flags; | |
4559 | ||
4560 | local_irq_save(flags); | |
4561 | double_rq_lock(rq_src, rq_dest); | |
4562 | rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible; | |
4563 | rq_src->nr_uninterruptible = 0; | |
4564 | double_rq_unlock(rq_src, rq_dest); | |
4565 | local_irq_restore(flags); | |
4566 | } | |
4567 | ||
4568 | /* Run through task list and migrate tasks from the dead cpu. */ | |
4569 | static void migrate_live_tasks(int src_cpu) | |
4570 | { | |
4571 | struct task_struct *tsk, *t; | |
4572 | ||
4573 | write_lock_irq(&tasklist_lock); | |
4574 | ||
4575 | do_each_thread(t, tsk) { | |
4576 | if (tsk == current) | |
4577 | continue; | |
4578 | ||
4579 | if (task_cpu(tsk) == src_cpu) | |
4580 | move_task_off_dead_cpu(src_cpu, tsk); | |
4581 | } while_each_thread(t, tsk); | |
4582 | ||
4583 | write_unlock_irq(&tasklist_lock); | |
4584 | } | |
4585 | ||
4586 | /* Schedules idle task to be the next runnable task on current CPU. | |
4587 | * It does so by boosting its priority to highest possible and adding it to | |
4588 | * the _front_ of runqueue. Used by CPU offline code. | |
4589 | */ | |
4590 | void sched_idle_next(void) | |
4591 | { | |
4592 | int cpu = smp_processor_id(); | |
4593 | runqueue_t *rq = this_rq(); | |
4594 | struct task_struct *p = rq->idle; | |
4595 | unsigned long flags; | |
4596 | ||
4597 | /* cpu has to be offline */ | |
4598 | BUG_ON(cpu_online(cpu)); | |
4599 | ||
4600 | /* Strictly not necessary since rest of the CPUs are stopped by now | |
4601 | * and interrupts disabled on current cpu. | |
4602 | */ | |
4603 | spin_lock_irqsave(&rq->lock, flags); | |
4604 | ||
4605 | __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1); | |
4606 | /* Add idle task to _front_ of it's priority queue */ | |
4607 | __activate_idle_task(p, rq); | |
4608 | ||
4609 | spin_unlock_irqrestore(&rq->lock, flags); | |
4610 | } | |
4611 | ||
4612 | /* Ensures that the idle task is using init_mm right before its cpu goes | |
4613 | * offline. | |
4614 | */ | |
4615 | void idle_task_exit(void) | |
4616 | { | |
4617 | struct mm_struct *mm = current->active_mm; | |
4618 | ||
4619 | BUG_ON(cpu_online(smp_processor_id())); | |
4620 | ||
4621 | if (mm != &init_mm) | |
4622 | switch_mm(mm, &init_mm, current); | |
4623 | mmdrop(mm); | |
4624 | } | |
4625 | ||
4626 | static void migrate_dead(unsigned int dead_cpu, task_t *tsk) | |
4627 | { | |
4628 | struct runqueue *rq = cpu_rq(dead_cpu); | |
4629 | ||
4630 | /* Must be exiting, otherwise would be on tasklist. */ | |
4631 | BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD); | |
4632 | ||
4633 | /* Cannot have done final schedule yet: would have vanished. */ | |
4634 | BUG_ON(tsk->flags & PF_DEAD); | |
4635 | ||
4636 | get_task_struct(tsk); | |
4637 | ||
4638 | /* | |
4639 | * Drop lock around migration; if someone else moves it, | |
4640 | * that's OK. No task can be added to this CPU, so iteration is | |
4641 | * fine. | |
4642 | */ | |
4643 | spin_unlock_irq(&rq->lock); | |
4644 | move_task_off_dead_cpu(dead_cpu, tsk); | |
4645 | spin_lock_irq(&rq->lock); | |
4646 | ||
4647 | put_task_struct(tsk); | |
4648 | } | |
4649 | ||
4650 | /* release_task() removes task from tasklist, so we won't find dead tasks. */ | |
4651 | static void migrate_dead_tasks(unsigned int dead_cpu) | |
4652 | { | |
4653 | unsigned arr, i; | |
4654 | struct runqueue *rq = cpu_rq(dead_cpu); | |
4655 | ||
4656 | for (arr = 0; arr < 2; arr++) { | |
4657 | for (i = 0; i < MAX_PRIO; i++) { | |
4658 | struct list_head *list = &rq->arrays[arr].queue[i]; | |
4659 | while (!list_empty(list)) | |
4660 | migrate_dead(dead_cpu, | |
4661 | list_entry(list->next, task_t, | |
4662 | run_list)); | |
4663 | } | |
4664 | } | |
4665 | } | |
4666 | #endif /* CONFIG_HOTPLUG_CPU */ | |
4667 | ||
4668 | /* | |
4669 | * migration_call - callback that gets triggered when a CPU is added. | |
4670 | * Here we can start up the necessary migration thread for the new CPU. | |
4671 | */ | |
26c2143b CS |
4672 | static int __cpuinit migration_call(struct notifier_block *nfb, |
4673 | unsigned long action, | |
4674 | void *hcpu) | |
1da177e4 LT |
4675 | { |
4676 | int cpu = (long)hcpu; | |
4677 | struct task_struct *p; | |
4678 | struct runqueue *rq; | |
4679 | unsigned long flags; | |
4680 | ||
4681 | switch (action) { | |
4682 | case CPU_UP_PREPARE: | |
4683 | p = kthread_create(migration_thread, hcpu, "migration/%d",cpu); | |
4684 | if (IS_ERR(p)) | |
4685 | return NOTIFY_BAD; | |
4686 | p->flags |= PF_NOFREEZE; | |
4687 | kthread_bind(p, cpu); | |
4688 | /* Must be high prio: stop_machine expects to yield to it. */ | |
4689 | rq = task_rq_lock(p, &flags); | |
4690 | __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1); | |
4691 | task_rq_unlock(rq, &flags); | |
4692 | cpu_rq(cpu)->migration_thread = p; | |
4693 | break; | |
4694 | case CPU_ONLINE: | |
4695 | /* Strictly unneccessary, as first user will wake it. */ | |
4696 | wake_up_process(cpu_rq(cpu)->migration_thread); | |
4697 | break; | |
4698 | #ifdef CONFIG_HOTPLUG_CPU | |
4699 | case CPU_UP_CANCELED: | |
fc75cdfa HC |
4700 | if (!cpu_rq(cpu)->migration_thread) |
4701 | break; | |
1da177e4 | 4702 | /* Unbind it from offline cpu so it can run. Fall thru. */ |
a4c4af7c HC |
4703 | kthread_bind(cpu_rq(cpu)->migration_thread, |
4704 | any_online_cpu(cpu_online_map)); | |
1da177e4 LT |
4705 | kthread_stop(cpu_rq(cpu)->migration_thread); |
4706 | cpu_rq(cpu)->migration_thread = NULL; | |
4707 | break; | |
4708 | case CPU_DEAD: | |
4709 | migrate_live_tasks(cpu); | |
4710 | rq = cpu_rq(cpu); | |
4711 | kthread_stop(rq->migration_thread); | |
4712 | rq->migration_thread = NULL; | |
4713 | /* Idle task back to normal (off runqueue, low prio) */ | |
4714 | rq = task_rq_lock(rq->idle, &flags); | |
4715 | deactivate_task(rq->idle, rq); | |
4716 | rq->idle->static_prio = MAX_PRIO; | |
4717 | __setscheduler(rq->idle, SCHED_NORMAL, 0); | |
4718 | migrate_dead_tasks(cpu); | |
4719 | task_rq_unlock(rq, &flags); | |
4720 | migrate_nr_uninterruptible(rq); | |
4721 | BUG_ON(rq->nr_running != 0); | |
4722 | ||
4723 | /* No need to migrate the tasks: it was best-effort if | |
4724 | * they didn't do lock_cpu_hotplug(). Just wake up | |
4725 | * the requestors. */ | |
4726 | spin_lock_irq(&rq->lock); | |
4727 | while (!list_empty(&rq->migration_queue)) { | |
4728 | migration_req_t *req; | |
4729 | req = list_entry(rq->migration_queue.next, | |
4730 | migration_req_t, list); | |
1da177e4 LT |
4731 | list_del_init(&req->list); |
4732 | complete(&req->done); | |
4733 | } | |
4734 | spin_unlock_irq(&rq->lock); | |
4735 | break; | |
4736 | #endif | |
4737 | } | |
4738 | return NOTIFY_OK; | |
4739 | } | |
4740 | ||
4741 | /* Register at highest priority so that task migration (migrate_all_tasks) | |
4742 | * happens before everything else. | |
4743 | */ | |
26c2143b | 4744 | static struct notifier_block __cpuinitdata migration_notifier = { |
1da177e4 LT |
4745 | .notifier_call = migration_call, |
4746 | .priority = 10 | |
4747 | }; | |
4748 | ||
4749 | int __init migration_init(void) | |
4750 | { | |
4751 | void *cpu = (void *)(long)smp_processor_id(); | |
4752 | /* Start one for boot CPU. */ | |
4753 | migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); | |
4754 | migration_call(&migration_notifier, CPU_ONLINE, cpu); | |
4755 | register_cpu_notifier(&migration_notifier); | |
4756 | return 0; | |
4757 | } | |
4758 | #endif | |
4759 | ||
4760 | #ifdef CONFIG_SMP | |
1a20ff27 | 4761 | #undef SCHED_DOMAIN_DEBUG |
1da177e4 LT |
4762 | #ifdef SCHED_DOMAIN_DEBUG |
4763 | static void sched_domain_debug(struct sched_domain *sd, int cpu) | |
4764 | { | |
4765 | int level = 0; | |
4766 | ||
41c7ce9a NP |
4767 | if (!sd) { |
4768 | printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); | |
4769 | return; | |
4770 | } | |
4771 | ||
1da177e4 LT |
4772 | printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); |
4773 | ||
4774 | do { | |
4775 | int i; | |
4776 | char str[NR_CPUS]; | |
4777 | struct sched_group *group = sd->groups; | |
4778 | cpumask_t groupmask; | |
4779 | ||
4780 | cpumask_scnprintf(str, NR_CPUS, sd->span); | |
4781 | cpus_clear(groupmask); | |
4782 | ||
4783 | printk(KERN_DEBUG); | |
4784 | for (i = 0; i < level + 1; i++) | |
4785 | printk(" "); | |
4786 | printk("domain %d: ", level); | |
4787 | ||
4788 | if (!(sd->flags & SD_LOAD_BALANCE)) { | |
4789 | printk("does not load-balance\n"); | |
4790 | if (sd->parent) | |
4791 | printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent"); | |
4792 | break; | |
4793 | } | |
4794 | ||
4795 | printk("span %s\n", str); | |
4796 | ||
4797 | if (!cpu_isset(cpu, sd->span)) | |
4798 | printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu); | |
4799 | if (!cpu_isset(cpu, group->cpumask)) | |
4800 | printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu); | |
4801 | ||
4802 | printk(KERN_DEBUG); | |
4803 | for (i = 0; i < level + 2; i++) | |
4804 | printk(" "); | |
4805 | printk("groups:"); | |
4806 | do { | |
4807 | if (!group) { | |
4808 | printk("\n"); | |
4809 | printk(KERN_ERR "ERROR: group is NULL\n"); | |
4810 | break; | |
4811 | } | |
4812 | ||
4813 | if (!group->cpu_power) { | |
4814 | printk("\n"); | |
4815 | printk(KERN_ERR "ERROR: domain->cpu_power not set\n"); | |
4816 | } | |
4817 | ||
4818 | if (!cpus_weight(group->cpumask)) { | |
4819 | printk("\n"); | |
4820 | printk(KERN_ERR "ERROR: empty group\n"); | |
4821 | } | |
4822 | ||
4823 | if (cpus_intersects(groupmask, group->cpumask)) { | |
4824 | printk("\n"); | |
4825 | printk(KERN_ERR "ERROR: repeated CPUs\n"); | |
4826 | } | |
4827 | ||
4828 | cpus_or(groupmask, groupmask, group->cpumask); | |
4829 | ||
4830 | cpumask_scnprintf(str, NR_CPUS, group->cpumask); | |
4831 | printk(" %s", str); | |
4832 | ||
4833 | group = group->next; | |
4834 | } while (group != sd->groups); | |
4835 | printk("\n"); | |
4836 | ||
4837 | if (!cpus_equal(sd->span, groupmask)) | |
4838 | printk(KERN_ERR "ERROR: groups don't span domain->span\n"); | |
4839 | ||
4840 | level++; | |
4841 | sd = sd->parent; | |
4842 | ||
4843 | if (sd) { | |
4844 | if (!cpus_subset(groupmask, sd->span)) | |
4845 | printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n"); | |
4846 | } | |
4847 | ||
4848 | } while (sd); | |
4849 | } | |
4850 | #else | |
4851 | #define sched_domain_debug(sd, cpu) {} | |
4852 | #endif | |
4853 | ||
1a20ff27 | 4854 | static int sd_degenerate(struct sched_domain *sd) |
245af2c7 SS |
4855 | { |
4856 | if (cpus_weight(sd->span) == 1) | |
4857 | return 1; | |
4858 | ||
4859 | /* Following flags need at least 2 groups */ | |
4860 | if (sd->flags & (SD_LOAD_BALANCE | | |
4861 | SD_BALANCE_NEWIDLE | | |
4862 | SD_BALANCE_FORK | | |
4863 | SD_BALANCE_EXEC)) { | |
4864 | if (sd->groups != sd->groups->next) | |
4865 | return 0; | |
4866 | } | |
4867 | ||
4868 | /* Following flags don't use groups */ | |
4869 | if (sd->flags & (SD_WAKE_IDLE | | |
4870 | SD_WAKE_AFFINE | | |
4871 | SD_WAKE_BALANCE)) | |
4872 | return 0; | |
4873 | ||
4874 | return 1; | |
4875 | } | |
4876 | ||
1a20ff27 | 4877 | static int sd_parent_degenerate(struct sched_domain *sd, |
245af2c7 SS |
4878 | struct sched_domain *parent) |
4879 | { | |
4880 | unsigned long cflags = sd->flags, pflags = parent->flags; | |
4881 | ||
4882 | if (sd_degenerate(parent)) | |
4883 | return 1; | |
4884 | ||
4885 | if (!cpus_equal(sd->span, parent->span)) | |
4886 | return 0; | |
4887 | ||
4888 | /* Does parent contain flags not in child? */ | |
4889 | /* WAKE_BALANCE is a subset of WAKE_AFFINE */ | |
4890 | if (cflags & SD_WAKE_AFFINE) | |
4891 | pflags &= ~SD_WAKE_BALANCE; | |
4892 | /* Flags needing groups don't count if only 1 group in parent */ | |
4893 | if (parent->groups == parent->groups->next) { | |
4894 | pflags &= ~(SD_LOAD_BALANCE | | |
4895 | SD_BALANCE_NEWIDLE | | |
4896 | SD_BALANCE_FORK | | |
4897 | SD_BALANCE_EXEC); | |
4898 | } | |
4899 | if (~cflags & pflags) | |
4900 | return 0; | |
4901 | ||
4902 | return 1; | |
4903 | } | |
4904 | ||
1da177e4 LT |
4905 | /* |
4906 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must | |
4907 | * hold the hotplug lock. | |
4908 | */ | |
9c1cfda2 | 4909 | static void cpu_attach_domain(struct sched_domain *sd, int cpu) |
1da177e4 | 4910 | { |
1da177e4 | 4911 | runqueue_t *rq = cpu_rq(cpu); |
245af2c7 SS |
4912 | struct sched_domain *tmp; |
4913 | ||
4914 | /* Remove the sched domains which do not contribute to scheduling. */ | |
4915 | for (tmp = sd; tmp; tmp = tmp->parent) { | |
4916 | struct sched_domain *parent = tmp->parent; | |
4917 | if (!parent) | |
4918 | break; | |
4919 | if (sd_parent_degenerate(tmp, parent)) | |
4920 | tmp->parent = parent->parent; | |
4921 | } | |
4922 | ||
4923 | if (sd && sd_degenerate(sd)) | |
4924 | sd = sd->parent; | |
1da177e4 LT |
4925 | |
4926 | sched_domain_debug(sd, cpu); | |
4927 | ||
674311d5 | 4928 | rcu_assign_pointer(rq->sd, sd); |
1da177e4 LT |
4929 | } |
4930 | ||
4931 | /* cpus with isolated domains */ | |
9c1cfda2 | 4932 | static cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE; |
1da177e4 LT |
4933 | |
4934 | /* Setup the mask of cpus configured for isolated domains */ | |
4935 | static int __init isolated_cpu_setup(char *str) | |
4936 | { | |
4937 | int ints[NR_CPUS], i; | |
4938 | ||
4939 | str = get_options(str, ARRAY_SIZE(ints), ints); | |
4940 | cpus_clear(cpu_isolated_map); | |
4941 | for (i = 1; i <= ints[0]; i++) | |
4942 | if (ints[i] < NR_CPUS) | |
4943 | cpu_set(ints[i], cpu_isolated_map); | |
4944 | return 1; | |
4945 | } | |
4946 | ||
4947 | __setup ("isolcpus=", isolated_cpu_setup); | |
4948 | ||
4949 | /* | |
4950 | * init_sched_build_groups takes an array of groups, the cpumask we wish | |
4951 | * to span, and a pointer to a function which identifies what group a CPU | |
4952 | * belongs to. The return value of group_fn must be a valid index into the | |
4953 | * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we | |
4954 | * keep track of groups covered with a cpumask_t). | |
4955 | * | |
4956 | * init_sched_build_groups will build a circular linked list of the groups | |
4957 | * covered by the given span, and will set each group's ->cpumask correctly, | |
4958 | * and ->cpu_power to 0. | |
4959 | */ | |
9c1cfda2 JH |
4960 | static void init_sched_build_groups(struct sched_group groups[], cpumask_t span, |
4961 | int (*group_fn)(int cpu)) | |
1da177e4 LT |
4962 | { |
4963 | struct sched_group *first = NULL, *last = NULL; | |
4964 | cpumask_t covered = CPU_MASK_NONE; | |
4965 | int i; | |
4966 | ||
4967 | for_each_cpu_mask(i, span) { | |
4968 | int group = group_fn(i); | |
4969 | struct sched_group *sg = &groups[group]; | |
4970 | int j; | |
4971 | ||
4972 | if (cpu_isset(i, covered)) | |
4973 | continue; | |
4974 | ||
4975 | sg->cpumask = CPU_MASK_NONE; | |
4976 | sg->cpu_power = 0; | |
4977 | ||
4978 | for_each_cpu_mask(j, span) { | |
4979 | if (group_fn(j) != group) | |
4980 | continue; | |
4981 | ||
4982 | cpu_set(j, covered); | |
4983 | cpu_set(j, sg->cpumask); | |
4984 | } | |
4985 | if (!first) | |
4986 | first = sg; | |
4987 | if (last) | |
4988 | last->next = sg; | |
4989 | last = sg; | |
4990 | } | |
4991 | last->next = first; | |
4992 | } | |
4993 | ||
9c1cfda2 | 4994 | #define SD_NODES_PER_DOMAIN 16 |
1da177e4 | 4995 | |
198e2f18 | 4996 | /* |
4997 | * Self-tuning task migration cost measurement between source and target CPUs. | |
4998 | * | |
4999 | * This is done by measuring the cost of manipulating buffers of varying | |
5000 | * sizes. For a given buffer-size here are the steps that are taken: | |
5001 | * | |
5002 | * 1) the source CPU reads+dirties a shared buffer | |
5003 | * 2) the target CPU reads+dirties the same shared buffer | |
5004 | * | |
5005 | * We measure how long they take, in the following 4 scenarios: | |
5006 | * | |
5007 | * - source: CPU1, target: CPU2 | cost1 | |
5008 | * - source: CPU2, target: CPU1 | cost2 | |
5009 | * - source: CPU1, target: CPU1 | cost3 | |
5010 | * - source: CPU2, target: CPU2 | cost4 | |
5011 | * | |
5012 | * We then calculate the cost3+cost4-cost1-cost2 difference - this is | |
5013 | * the cost of migration. | |
5014 | * | |
5015 | * We then start off from a small buffer-size and iterate up to larger | |
5016 | * buffer sizes, in 5% steps - measuring each buffer-size separately, and | |
5017 | * doing a maximum search for the cost. (The maximum cost for a migration | |
5018 | * normally occurs when the working set size is around the effective cache | |
5019 | * size.) | |
5020 | */ | |
5021 | #define SEARCH_SCOPE 2 | |
5022 | #define MIN_CACHE_SIZE (64*1024U) | |
5023 | #define DEFAULT_CACHE_SIZE (5*1024*1024U) | |
70b4d63e | 5024 | #define ITERATIONS 1 |
198e2f18 | 5025 | #define SIZE_THRESH 130 |
5026 | #define COST_THRESH 130 | |
5027 | ||
5028 | /* | |
5029 | * The migration cost is a function of 'domain distance'. Domain | |
5030 | * distance is the number of steps a CPU has to iterate down its | |
5031 | * domain tree to share a domain with the other CPU. The farther | |
5032 | * two CPUs are from each other, the larger the distance gets. | |
5033 | * | |
5034 | * Note that we use the distance only to cache measurement results, | |
5035 | * the distance value is not used numerically otherwise. When two | |
5036 | * CPUs have the same distance it is assumed that the migration | |
5037 | * cost is the same. (this is a simplification but quite practical) | |
5038 | */ | |
5039 | #define MAX_DOMAIN_DISTANCE 32 | |
5040 | ||
5041 | static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] = | |
4bbf39c2 IM |
5042 | { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] = |
5043 | /* | |
5044 | * Architectures may override the migration cost and thus avoid | |
5045 | * boot-time calibration. Unit is nanoseconds. Mostly useful for | |
5046 | * virtualized hardware: | |
5047 | */ | |
5048 | #ifdef CONFIG_DEFAULT_MIGRATION_COST | |
5049 | CONFIG_DEFAULT_MIGRATION_COST | |
5050 | #else | |
5051 | -1LL | |
5052 | #endif | |
5053 | }; | |
198e2f18 | 5054 | |
5055 | /* | |
5056 | * Allow override of migration cost - in units of microseconds. | |
5057 | * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost | |
5058 | * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs: | |
5059 | */ | |
5060 | static int __init migration_cost_setup(char *str) | |
5061 | { | |
5062 | int ints[MAX_DOMAIN_DISTANCE+1], i; | |
5063 | ||
5064 | str = get_options(str, ARRAY_SIZE(ints), ints); | |
5065 | ||
5066 | printk("#ints: %d\n", ints[0]); | |
5067 | for (i = 1; i <= ints[0]; i++) { | |
5068 | migration_cost[i-1] = (unsigned long long)ints[i]*1000; | |
5069 | printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]); | |
5070 | } | |
5071 | return 1; | |
5072 | } | |
5073 | ||
5074 | __setup ("migration_cost=", migration_cost_setup); | |
5075 | ||
5076 | /* | |
5077 | * Global multiplier (divisor) for migration-cutoff values, | |
5078 | * in percentiles. E.g. use a value of 150 to get 1.5 times | |
5079 | * longer cache-hot cutoff times. | |
5080 | * | |
5081 | * (We scale it from 100 to 128 to long long handling easier.) | |
5082 | */ | |
5083 | ||
5084 | #define MIGRATION_FACTOR_SCALE 128 | |
5085 | ||
5086 | static unsigned int migration_factor = MIGRATION_FACTOR_SCALE; | |
5087 | ||
5088 | static int __init setup_migration_factor(char *str) | |
5089 | { | |
5090 | get_option(&str, &migration_factor); | |
5091 | migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100; | |
5092 | return 1; | |
5093 | } | |
5094 | ||
5095 | __setup("migration_factor=", setup_migration_factor); | |
5096 | ||
5097 | /* | |
5098 | * Estimated distance of two CPUs, measured via the number of domains | |
5099 | * we have to pass for the two CPUs to be in the same span: | |
5100 | */ | |
5101 | static unsigned long domain_distance(int cpu1, int cpu2) | |
5102 | { | |
5103 | unsigned long distance = 0; | |
5104 | struct sched_domain *sd; | |
5105 | ||
5106 | for_each_domain(cpu1, sd) { | |
5107 | WARN_ON(!cpu_isset(cpu1, sd->span)); | |
5108 | if (cpu_isset(cpu2, sd->span)) | |
5109 | return distance; | |
5110 | distance++; | |
5111 | } | |
5112 | if (distance >= MAX_DOMAIN_DISTANCE) { | |
5113 | WARN_ON(1); | |
5114 | distance = MAX_DOMAIN_DISTANCE-1; | |
5115 | } | |
5116 | ||
5117 | return distance; | |
5118 | } | |
5119 | ||
5120 | static unsigned int migration_debug; | |
5121 | ||
5122 | static int __init setup_migration_debug(char *str) | |
5123 | { | |
5124 | get_option(&str, &migration_debug); | |
5125 | return 1; | |
5126 | } | |
5127 | ||
5128 | __setup("migration_debug=", setup_migration_debug); | |
5129 | ||
5130 | /* | |
5131 | * Maximum cache-size that the scheduler should try to measure. | |
5132 | * Architectures with larger caches should tune this up during | |
5133 | * bootup. Gets used in the domain-setup code (i.e. during SMP | |
5134 | * bootup). | |
5135 | */ | |
5136 | unsigned int max_cache_size; | |
5137 | ||
5138 | static int __init setup_max_cache_size(char *str) | |
5139 | { | |
5140 | get_option(&str, &max_cache_size); | |
5141 | return 1; | |
5142 | } | |
5143 | ||
5144 | __setup("max_cache_size=", setup_max_cache_size); | |
5145 | ||
5146 | /* | |
5147 | * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This | |
5148 | * is the operation that is timed, so we try to generate unpredictable | |
5149 | * cachemisses that still end up filling the L2 cache: | |
5150 | */ | |
5151 | static void touch_cache(void *__cache, unsigned long __size) | |
5152 | { | |
5153 | unsigned long size = __size/sizeof(long), chunk1 = size/3, | |
5154 | chunk2 = 2*size/3; | |
5155 | unsigned long *cache = __cache; | |
5156 | int i; | |
5157 | ||
5158 | for (i = 0; i < size/6; i += 8) { | |
5159 | switch (i % 6) { | |
5160 | case 0: cache[i]++; | |
5161 | case 1: cache[size-1-i]++; | |
5162 | case 2: cache[chunk1-i]++; | |
5163 | case 3: cache[chunk1+i]++; | |
5164 | case 4: cache[chunk2-i]++; | |
5165 | case 5: cache[chunk2+i]++; | |
5166 | } | |
5167 | } | |
5168 | } | |
5169 | ||
5170 | /* | |
5171 | * Measure the cache-cost of one task migration. Returns in units of nsec. | |
5172 | */ | |
5173 | static unsigned long long measure_one(void *cache, unsigned long size, | |
5174 | int source, int target) | |
5175 | { | |
5176 | cpumask_t mask, saved_mask; | |
5177 | unsigned long long t0, t1, t2, t3, cost; | |
5178 | ||
5179 | saved_mask = current->cpus_allowed; | |
5180 | ||
5181 | /* | |
5182 | * Flush source caches to RAM and invalidate them: | |
5183 | */ | |
5184 | sched_cacheflush(); | |
5185 | ||
5186 | /* | |
5187 | * Migrate to the source CPU: | |
5188 | */ | |
5189 | mask = cpumask_of_cpu(source); | |
5190 | set_cpus_allowed(current, mask); | |
5191 | WARN_ON(smp_processor_id() != source); | |
5192 | ||
5193 | /* | |
5194 | * Dirty the working set: | |
5195 | */ | |
5196 | t0 = sched_clock(); | |
5197 | touch_cache(cache, size); | |
5198 | t1 = sched_clock(); | |
5199 | ||
5200 | /* | |
5201 | * Migrate to the target CPU, dirty the L2 cache and access | |
5202 | * the shared buffer. (which represents the working set | |
5203 | * of a migrated task.) | |
5204 | */ | |
5205 | mask = cpumask_of_cpu(target); | |
5206 | set_cpus_allowed(current, mask); | |
5207 | WARN_ON(smp_processor_id() != target); | |
5208 | ||
5209 | t2 = sched_clock(); | |
5210 | touch_cache(cache, size); | |
5211 | t3 = sched_clock(); | |
5212 | ||
5213 | cost = t1-t0 + t3-t2; | |
5214 | ||
5215 | if (migration_debug >= 2) | |
5216 | printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n", | |
5217 | source, target, t1-t0, t1-t0, t3-t2, cost); | |
5218 | /* | |
5219 | * Flush target caches to RAM and invalidate them: | |
5220 | */ | |
5221 | sched_cacheflush(); | |
5222 | ||
5223 | set_cpus_allowed(current, saved_mask); | |
5224 | ||
5225 | return cost; | |
5226 | } | |
5227 | ||
5228 | /* | |
5229 | * Measure a series of task migrations and return the average | |
5230 | * result. Since this code runs early during bootup the system | |
5231 | * is 'undisturbed' and the average latency makes sense. | |
5232 | * | |
5233 | * The algorithm in essence auto-detects the relevant cache-size, | |
5234 | * so it will properly detect different cachesizes for different | |
5235 | * cache-hierarchies, depending on how the CPUs are connected. | |
5236 | * | |
5237 | * Architectures can prime the upper limit of the search range via | |
5238 | * max_cache_size, otherwise the search range defaults to 20MB...64K. | |
5239 | */ | |
5240 | static unsigned long long | |
5241 | measure_cost(int cpu1, int cpu2, void *cache, unsigned int size) | |
5242 | { | |
5243 | unsigned long long cost1, cost2; | |
5244 | int i; | |
5245 | ||
5246 | /* | |
5247 | * Measure the migration cost of 'size' bytes, over an | |
5248 | * average of 10 runs: | |
5249 | * | |
5250 | * (We perturb the cache size by a small (0..4k) | |
5251 | * value to compensate size/alignment related artifacts. | |
5252 | * We also subtract the cost of the operation done on | |
5253 | * the same CPU.) | |
5254 | */ | |
5255 | cost1 = 0; | |
5256 | ||
5257 | /* | |
5258 | * dry run, to make sure we start off cache-cold on cpu1, | |
5259 | * and to get any vmalloc pagefaults in advance: | |
5260 | */ | |
5261 | measure_one(cache, size, cpu1, cpu2); | |
5262 | for (i = 0; i < ITERATIONS; i++) | |
5263 | cost1 += measure_one(cache, size - i*1024, cpu1, cpu2); | |
5264 | ||
5265 | measure_one(cache, size, cpu2, cpu1); | |
5266 | for (i = 0; i < ITERATIONS; i++) | |
5267 | cost1 += measure_one(cache, size - i*1024, cpu2, cpu1); | |
5268 | ||
5269 | /* | |
5270 | * (We measure the non-migrating [cached] cost on both | |
5271 | * cpu1 and cpu2, to handle CPUs with different speeds) | |
5272 | */ | |
5273 | cost2 = 0; | |
5274 | ||
5275 | measure_one(cache, size, cpu1, cpu1); | |
5276 | for (i = 0; i < ITERATIONS; i++) | |
5277 | cost2 += measure_one(cache, size - i*1024, cpu1, cpu1); | |
5278 | ||
5279 | measure_one(cache, size, cpu2, cpu2); | |
5280 | for (i = 0; i < ITERATIONS; i++) | |
5281 | cost2 += measure_one(cache, size - i*1024, cpu2, cpu2); | |
5282 | ||
5283 | /* | |
5284 | * Get the per-iteration migration cost: | |
5285 | */ | |
5286 | do_div(cost1, 2*ITERATIONS); | |
5287 | do_div(cost2, 2*ITERATIONS); | |
5288 | ||
5289 | return cost1 - cost2; | |
5290 | } | |
5291 | ||
5292 | static unsigned long long measure_migration_cost(int cpu1, int cpu2) | |
5293 | { | |
5294 | unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0; | |
5295 | unsigned int max_size, size, size_found = 0; | |
5296 | long long cost = 0, prev_cost; | |
5297 | void *cache; | |
5298 | ||
5299 | /* | |
5300 | * Search from max_cache_size*5 down to 64K - the real relevant | |
5301 | * cachesize has to lie somewhere inbetween. | |
5302 | */ | |
5303 | if (max_cache_size) { | |
5304 | max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE); | |
5305 | size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE); | |
5306 | } else { | |
5307 | /* | |
5308 | * Since we have no estimation about the relevant | |
5309 | * search range | |
5310 | */ | |
5311 | max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE; | |
5312 | size = MIN_CACHE_SIZE; | |
5313 | } | |
5314 | ||
5315 | if (!cpu_online(cpu1) || !cpu_online(cpu2)) { | |
5316 | printk("cpu %d and %d not both online!\n", cpu1, cpu2); | |
5317 | return 0; | |
5318 | } | |
5319 | ||
5320 | /* | |
5321 | * Allocate the working set: | |
5322 | */ | |
5323 | cache = vmalloc(max_size); | |
5324 | if (!cache) { | |
5325 | printk("could not vmalloc %d bytes for cache!\n", 2*max_size); | |
5326 | return 1000000; // return 1 msec on very small boxen | |
5327 | } | |
5328 | ||
5329 | while (size <= max_size) { | |
5330 | prev_cost = cost; | |
5331 | cost = measure_cost(cpu1, cpu2, cache, size); | |
5332 | ||
5333 | /* | |
5334 | * Update the max: | |
5335 | */ | |
5336 | if (cost > 0) { | |
5337 | if (max_cost < cost) { | |
5338 | max_cost = cost; | |
5339 | size_found = size; | |
5340 | } | |
5341 | } | |
5342 | /* | |
5343 | * Calculate average fluctuation, we use this to prevent | |
5344 | * noise from triggering an early break out of the loop: | |
5345 | */ | |
5346 | fluct = abs(cost - prev_cost); | |
5347 | avg_fluct = (avg_fluct + fluct)/2; | |
5348 | ||
5349 | if (migration_debug) | |
5350 | printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n", | |
5351 | cpu1, cpu2, size, | |
5352 | (long)cost / 1000000, | |
5353 | ((long)cost / 100000) % 10, | |
5354 | (long)max_cost / 1000000, | |
5355 | ((long)max_cost / 100000) % 10, | |
5356 | domain_distance(cpu1, cpu2), | |
5357 | cost, avg_fluct); | |
5358 | ||
5359 | /* | |
5360 | * If we iterated at least 20% past the previous maximum, | |
5361 | * and the cost has dropped by more than 20% already, | |
5362 | * (taking fluctuations into account) then we assume to | |
5363 | * have found the maximum and break out of the loop early: | |
5364 | */ | |
5365 | if (size_found && (size*100 > size_found*SIZE_THRESH)) | |
5366 | if (cost+avg_fluct <= 0 || | |
5367 | max_cost*100 > (cost+avg_fluct)*COST_THRESH) { | |
5368 | ||
5369 | if (migration_debug) | |
5370 | printk("-> found max.\n"); | |
5371 | break; | |
5372 | } | |
5373 | /* | |
70b4d63e | 5374 | * Increase the cachesize in 10% steps: |
198e2f18 | 5375 | */ |
70b4d63e | 5376 | size = size * 10 / 9; |
198e2f18 | 5377 | } |
5378 | ||
5379 | if (migration_debug) | |
5380 | printk("[%d][%d] working set size found: %d, cost: %Ld\n", | |
5381 | cpu1, cpu2, size_found, max_cost); | |
5382 | ||
5383 | vfree(cache); | |
5384 | ||
5385 | /* | |
5386 | * A task is considered 'cache cold' if at least 2 times | |
5387 | * the worst-case cost of migration has passed. | |
5388 | * | |
5389 | * (this limit is only listened to if the load-balancing | |
5390 | * situation is 'nice' - if there is a large imbalance we | |
5391 | * ignore it for the sake of CPU utilization and | |
5392 | * processing fairness.) | |
5393 | */ | |
5394 | return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE; | |
5395 | } | |
5396 | ||
5397 | static void calibrate_migration_costs(const cpumask_t *cpu_map) | |
5398 | { | |
5399 | int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id(); | |
5400 | unsigned long j0, j1, distance, max_distance = 0; | |
5401 | struct sched_domain *sd; | |
5402 | ||
5403 | j0 = jiffies; | |
5404 | ||
5405 | /* | |
5406 | * First pass - calculate the cacheflush times: | |
5407 | */ | |
5408 | for_each_cpu_mask(cpu1, *cpu_map) { | |
5409 | for_each_cpu_mask(cpu2, *cpu_map) { | |
5410 | if (cpu1 == cpu2) | |
5411 | continue; | |
5412 | distance = domain_distance(cpu1, cpu2); | |
5413 | max_distance = max(max_distance, distance); | |
5414 | /* | |
5415 | * No result cached yet? | |
5416 | */ | |
5417 | if (migration_cost[distance] == -1LL) | |
5418 | migration_cost[distance] = | |
5419 | measure_migration_cost(cpu1, cpu2); | |
5420 | } | |
5421 | } | |
5422 | /* | |
5423 | * Second pass - update the sched domain hierarchy with | |
5424 | * the new cache-hot-time estimations: | |
5425 | */ | |
5426 | for_each_cpu_mask(cpu, *cpu_map) { | |
5427 | distance = 0; | |
5428 | for_each_domain(cpu, sd) { | |
5429 | sd->cache_hot_time = migration_cost[distance]; | |
5430 | distance++; | |
5431 | } | |
5432 | } | |
5433 | /* | |
5434 | * Print the matrix: | |
5435 | */ | |
5436 | if (migration_debug) | |
5437 | printk("migration: max_cache_size: %d, cpu: %d MHz:\n", | |
5438 | max_cache_size, | |
5439 | #ifdef CONFIG_X86 | |
5440 | cpu_khz/1000 | |
5441 | #else | |
5442 | -1 | |
5443 | #endif | |
5444 | ); | |
bd576c95 CE |
5445 | if (system_state == SYSTEM_BOOTING) { |
5446 | printk("migration_cost="); | |
5447 | for (distance = 0; distance <= max_distance; distance++) { | |
5448 | if (distance) | |
5449 | printk(","); | |
5450 | printk("%ld", (long)migration_cost[distance] / 1000); | |
5451 | } | |
5452 | printk("\n"); | |
198e2f18 | 5453 | } |
198e2f18 | 5454 | j1 = jiffies; |
5455 | if (migration_debug) | |
5456 | printk("migration: %ld seconds\n", (j1-j0)/HZ); | |
5457 | ||
5458 | /* | |
5459 | * Move back to the original CPU. NUMA-Q gets confused | |
5460 | * if we migrate to another quad during bootup. | |
5461 | */ | |
5462 | if (raw_smp_processor_id() != orig_cpu) { | |
5463 | cpumask_t mask = cpumask_of_cpu(orig_cpu), | |
5464 | saved_mask = current->cpus_allowed; | |
5465 | ||
5466 | set_cpus_allowed(current, mask); | |
5467 | set_cpus_allowed(current, saved_mask); | |
5468 | } | |
5469 | } | |
5470 | ||
9c1cfda2 | 5471 | #ifdef CONFIG_NUMA |
198e2f18 | 5472 | |
9c1cfda2 JH |
5473 | /** |
5474 | * find_next_best_node - find the next node to include in a sched_domain | |
5475 | * @node: node whose sched_domain we're building | |
5476 | * @used_nodes: nodes already in the sched_domain | |
5477 | * | |
5478 | * Find the next node to include in a given scheduling domain. Simply | |
5479 | * finds the closest node not already in the @used_nodes map. | |
5480 | * | |
5481 | * Should use nodemask_t. | |
5482 | */ | |
5483 | static int find_next_best_node(int node, unsigned long *used_nodes) | |
5484 | { | |
5485 | int i, n, val, min_val, best_node = 0; | |
5486 | ||
5487 | min_val = INT_MAX; | |
5488 | ||
5489 | for (i = 0; i < MAX_NUMNODES; i++) { | |
5490 | /* Start at @node */ | |
5491 | n = (node + i) % MAX_NUMNODES; | |
5492 | ||
5493 | if (!nr_cpus_node(n)) | |
5494 | continue; | |
5495 | ||
5496 | /* Skip already used nodes */ | |
5497 | if (test_bit(n, used_nodes)) | |
5498 | continue; | |
5499 | ||
5500 | /* Simple min distance search */ | |
5501 | val = node_distance(node, n); | |
5502 | ||
5503 | if (val < min_val) { | |
5504 | min_val = val; | |
5505 | best_node = n; | |
5506 | } | |
5507 | } | |
5508 | ||
5509 | set_bit(best_node, used_nodes); | |
5510 | return best_node; | |
5511 | } | |
5512 | ||
5513 | /** | |
5514 | * sched_domain_node_span - get a cpumask for a node's sched_domain | |
5515 | * @node: node whose cpumask we're constructing | |
5516 | * @size: number of nodes to include in this span | |
5517 | * | |
5518 | * Given a node, construct a good cpumask for its sched_domain to span. It | |
5519 | * should be one that prevents unnecessary balancing, but also spreads tasks | |
5520 | * out optimally. | |
5521 | */ | |
5522 | static cpumask_t sched_domain_node_span(int node) | |
5523 | { | |
5524 | int i; | |
5525 | cpumask_t span, nodemask; | |
5526 | DECLARE_BITMAP(used_nodes, MAX_NUMNODES); | |
5527 | ||
5528 | cpus_clear(span); | |
5529 | bitmap_zero(used_nodes, MAX_NUMNODES); | |
5530 | ||
5531 | nodemask = node_to_cpumask(node); | |
5532 | cpus_or(span, span, nodemask); | |
5533 | set_bit(node, used_nodes); | |
5534 | ||
5535 | for (i = 1; i < SD_NODES_PER_DOMAIN; i++) { | |
5536 | int next_node = find_next_best_node(node, used_nodes); | |
5537 | nodemask = node_to_cpumask(next_node); | |
5538 | cpus_or(span, span, nodemask); | |
5539 | } | |
5540 | ||
5541 | return span; | |
5542 | } | |
5543 | #endif | |
5544 | ||
5545 | /* | |
5546 | * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we | |
5547 | * can switch it on easily if needed. | |
5548 | */ | |
1da177e4 LT |
5549 | #ifdef CONFIG_SCHED_SMT |
5550 | static DEFINE_PER_CPU(struct sched_domain, cpu_domains); | |
5551 | static struct sched_group sched_group_cpus[NR_CPUS]; | |
1a20ff27 | 5552 | static int cpu_to_cpu_group(int cpu) |
1da177e4 LT |
5553 | { |
5554 | return cpu; | |
5555 | } | |
5556 | #endif | |
5557 | ||
1e9f28fa SS |
5558 | #ifdef CONFIG_SCHED_MC |
5559 | static DEFINE_PER_CPU(struct sched_domain, core_domains); | |
5560 | static struct sched_group sched_group_core[NR_CPUS]; | |
5561 | #endif | |
5562 | ||
5563 | #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT) | |
5564 | static int cpu_to_core_group(int cpu) | |
5565 | { | |
5566 | return first_cpu(cpu_sibling_map[cpu]); | |
5567 | } | |
5568 | #elif defined(CONFIG_SCHED_MC) | |
5569 | static int cpu_to_core_group(int cpu) | |
5570 | { | |
5571 | return cpu; | |
5572 | } | |
5573 | #endif | |
5574 | ||
1da177e4 LT |
5575 | static DEFINE_PER_CPU(struct sched_domain, phys_domains); |
5576 | static struct sched_group sched_group_phys[NR_CPUS]; | |
1a20ff27 | 5577 | static int cpu_to_phys_group(int cpu) |
1da177e4 | 5578 | { |
1e9f28fa SS |
5579 | #if defined(CONFIG_SCHED_MC) |
5580 | cpumask_t mask = cpu_coregroup_map(cpu); | |
5581 | return first_cpu(mask); | |
5582 | #elif defined(CONFIG_SCHED_SMT) | |
1da177e4 LT |
5583 | return first_cpu(cpu_sibling_map[cpu]); |
5584 | #else | |
5585 | return cpu; | |
5586 | #endif | |
5587 | } | |
5588 | ||
5589 | #ifdef CONFIG_NUMA | |
1da177e4 | 5590 | /* |
9c1cfda2 JH |
5591 | * The init_sched_build_groups can't handle what we want to do with node |
5592 | * groups, so roll our own. Now each node has its own list of groups which | |
5593 | * gets dynamically allocated. | |
1da177e4 | 5594 | */ |
9c1cfda2 | 5595 | static DEFINE_PER_CPU(struct sched_domain, node_domains); |
d1b55138 | 5596 | static struct sched_group **sched_group_nodes_bycpu[NR_CPUS]; |
1da177e4 | 5597 | |
9c1cfda2 | 5598 | static DEFINE_PER_CPU(struct sched_domain, allnodes_domains); |
d1b55138 | 5599 | static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS]; |
9c1cfda2 JH |
5600 | |
5601 | static int cpu_to_allnodes_group(int cpu) | |
5602 | { | |
5603 | return cpu_to_node(cpu); | |
1da177e4 | 5604 | } |
08069033 SS |
5605 | static void init_numa_sched_groups_power(struct sched_group *group_head) |
5606 | { | |
5607 | struct sched_group *sg = group_head; | |
5608 | int j; | |
5609 | ||
5610 | if (!sg) | |
5611 | return; | |
5612 | next_sg: | |
5613 | for_each_cpu_mask(j, sg->cpumask) { | |
5614 | struct sched_domain *sd; | |
5615 | ||
5616 | sd = &per_cpu(phys_domains, j); | |
5617 | if (j != first_cpu(sd->groups->cpumask)) { | |
5618 | /* | |
5619 | * Only add "power" once for each | |
5620 | * physical package. | |
5621 | */ | |
5622 | continue; | |
5623 | } | |
5624 | ||
5625 | sg->cpu_power += sd->groups->cpu_power; | |
5626 | } | |
5627 | sg = sg->next; | |
5628 | if (sg != group_head) | |
5629 | goto next_sg; | |
5630 | } | |
1da177e4 LT |
5631 | #endif |
5632 | ||
5633 | /* | |
1a20ff27 DG |
5634 | * Build sched domains for a given set of cpus and attach the sched domains |
5635 | * to the individual cpus | |
1da177e4 | 5636 | */ |
9c1cfda2 | 5637 | void build_sched_domains(const cpumask_t *cpu_map) |
1da177e4 LT |
5638 | { |
5639 | int i; | |
d1b55138 JH |
5640 | #ifdef CONFIG_NUMA |
5641 | struct sched_group **sched_group_nodes = NULL; | |
5642 | struct sched_group *sched_group_allnodes = NULL; | |
5643 | ||
5644 | /* | |
5645 | * Allocate the per-node list of sched groups | |
5646 | */ | |
5647 | sched_group_nodes = kmalloc(sizeof(struct sched_group*)*MAX_NUMNODES, | |
5648 | GFP_ATOMIC); | |
5649 | if (!sched_group_nodes) { | |
5650 | printk(KERN_WARNING "Can not alloc sched group node list\n"); | |
5651 | return; | |
5652 | } | |
5653 | sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes; | |
5654 | #endif | |
1da177e4 LT |
5655 | |
5656 | /* | |
1a20ff27 | 5657 | * Set up domains for cpus specified by the cpu_map. |
1da177e4 | 5658 | */ |
1a20ff27 | 5659 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 LT |
5660 | int group; |
5661 | struct sched_domain *sd = NULL, *p; | |
5662 | cpumask_t nodemask = node_to_cpumask(cpu_to_node(i)); | |
5663 | ||
1a20ff27 | 5664 | cpus_and(nodemask, nodemask, *cpu_map); |
1da177e4 LT |
5665 | |
5666 | #ifdef CONFIG_NUMA | |
d1b55138 | 5667 | if (cpus_weight(*cpu_map) |
9c1cfda2 | 5668 | > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) { |
d1b55138 JH |
5669 | if (!sched_group_allnodes) { |
5670 | sched_group_allnodes | |
5671 | = kmalloc(sizeof(struct sched_group) | |
5672 | * MAX_NUMNODES, | |
5673 | GFP_KERNEL); | |
5674 | if (!sched_group_allnodes) { | |
5675 | printk(KERN_WARNING | |
5676 | "Can not alloc allnodes sched group\n"); | |
5677 | break; | |
5678 | } | |
5679 | sched_group_allnodes_bycpu[i] | |
5680 | = sched_group_allnodes; | |
5681 | } | |
9c1cfda2 JH |
5682 | sd = &per_cpu(allnodes_domains, i); |
5683 | *sd = SD_ALLNODES_INIT; | |
5684 | sd->span = *cpu_map; | |
5685 | group = cpu_to_allnodes_group(i); | |
5686 | sd->groups = &sched_group_allnodes[group]; | |
5687 | p = sd; | |
5688 | } else | |
5689 | p = NULL; | |
5690 | ||
1da177e4 | 5691 | sd = &per_cpu(node_domains, i); |
1da177e4 | 5692 | *sd = SD_NODE_INIT; |
9c1cfda2 JH |
5693 | sd->span = sched_domain_node_span(cpu_to_node(i)); |
5694 | sd->parent = p; | |
5695 | cpus_and(sd->span, sd->span, *cpu_map); | |
1da177e4 LT |
5696 | #endif |
5697 | ||
5698 | p = sd; | |
5699 | sd = &per_cpu(phys_domains, i); | |
5700 | group = cpu_to_phys_group(i); | |
5701 | *sd = SD_CPU_INIT; | |
5702 | sd->span = nodemask; | |
5703 | sd->parent = p; | |
5704 | sd->groups = &sched_group_phys[group]; | |
5705 | ||
1e9f28fa SS |
5706 | #ifdef CONFIG_SCHED_MC |
5707 | p = sd; | |
5708 | sd = &per_cpu(core_domains, i); | |
5709 | group = cpu_to_core_group(i); | |
5710 | *sd = SD_MC_INIT; | |
5711 | sd->span = cpu_coregroup_map(i); | |
5712 | cpus_and(sd->span, sd->span, *cpu_map); | |
5713 | sd->parent = p; | |
5714 | sd->groups = &sched_group_core[group]; | |
5715 | #endif | |
5716 | ||
1da177e4 LT |
5717 | #ifdef CONFIG_SCHED_SMT |
5718 | p = sd; | |
5719 | sd = &per_cpu(cpu_domains, i); | |
5720 | group = cpu_to_cpu_group(i); | |
5721 | *sd = SD_SIBLING_INIT; | |
5722 | sd->span = cpu_sibling_map[i]; | |
1a20ff27 | 5723 | cpus_and(sd->span, sd->span, *cpu_map); |
1da177e4 LT |
5724 | sd->parent = p; |
5725 | sd->groups = &sched_group_cpus[group]; | |
5726 | #endif | |
5727 | } | |
5728 | ||
5729 | #ifdef CONFIG_SCHED_SMT | |
5730 | /* Set up CPU (sibling) groups */ | |
9c1cfda2 | 5731 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 | 5732 | cpumask_t this_sibling_map = cpu_sibling_map[i]; |
1a20ff27 | 5733 | cpus_and(this_sibling_map, this_sibling_map, *cpu_map); |
1da177e4 LT |
5734 | if (i != first_cpu(this_sibling_map)) |
5735 | continue; | |
5736 | ||
5737 | init_sched_build_groups(sched_group_cpus, this_sibling_map, | |
5738 | &cpu_to_cpu_group); | |
5739 | } | |
5740 | #endif | |
5741 | ||
1e9f28fa SS |
5742 | #ifdef CONFIG_SCHED_MC |
5743 | /* Set up multi-core groups */ | |
5744 | for_each_cpu_mask(i, *cpu_map) { | |
5745 | cpumask_t this_core_map = cpu_coregroup_map(i); | |
5746 | cpus_and(this_core_map, this_core_map, *cpu_map); | |
5747 | if (i != first_cpu(this_core_map)) | |
5748 | continue; | |
5749 | init_sched_build_groups(sched_group_core, this_core_map, | |
5750 | &cpu_to_core_group); | |
5751 | } | |
5752 | #endif | |
5753 | ||
5754 | ||
1da177e4 LT |
5755 | /* Set up physical groups */ |
5756 | for (i = 0; i < MAX_NUMNODES; i++) { | |
5757 | cpumask_t nodemask = node_to_cpumask(i); | |
5758 | ||
1a20ff27 | 5759 | cpus_and(nodemask, nodemask, *cpu_map); |
1da177e4 LT |
5760 | if (cpus_empty(nodemask)) |
5761 | continue; | |
5762 | ||
5763 | init_sched_build_groups(sched_group_phys, nodemask, | |
5764 | &cpu_to_phys_group); | |
5765 | } | |
5766 | ||
5767 | #ifdef CONFIG_NUMA | |
5768 | /* Set up node groups */ | |
d1b55138 JH |
5769 | if (sched_group_allnodes) |
5770 | init_sched_build_groups(sched_group_allnodes, *cpu_map, | |
5771 | &cpu_to_allnodes_group); | |
9c1cfda2 JH |
5772 | |
5773 | for (i = 0; i < MAX_NUMNODES; i++) { | |
5774 | /* Set up node groups */ | |
5775 | struct sched_group *sg, *prev; | |
5776 | cpumask_t nodemask = node_to_cpumask(i); | |
5777 | cpumask_t domainspan; | |
5778 | cpumask_t covered = CPU_MASK_NONE; | |
5779 | int j; | |
5780 | ||
5781 | cpus_and(nodemask, nodemask, *cpu_map); | |
d1b55138 JH |
5782 | if (cpus_empty(nodemask)) { |
5783 | sched_group_nodes[i] = NULL; | |
9c1cfda2 | 5784 | continue; |
d1b55138 | 5785 | } |
9c1cfda2 JH |
5786 | |
5787 | domainspan = sched_domain_node_span(i); | |
5788 | cpus_and(domainspan, domainspan, *cpu_map); | |
5789 | ||
5790 | sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL); | |
5791 | sched_group_nodes[i] = sg; | |
5792 | for_each_cpu_mask(j, nodemask) { | |
5793 | struct sched_domain *sd; | |
5794 | sd = &per_cpu(node_domains, j); | |
5795 | sd->groups = sg; | |
5796 | if (sd->groups == NULL) { | |
5797 | /* Turn off balancing if we have no groups */ | |
5798 | sd->flags = 0; | |
5799 | } | |
5800 | } | |
5801 | if (!sg) { | |
5802 | printk(KERN_WARNING | |
5803 | "Can not alloc domain group for node %d\n", i); | |
5804 | continue; | |
5805 | } | |
5806 | sg->cpu_power = 0; | |
5807 | sg->cpumask = nodemask; | |
5808 | cpus_or(covered, covered, nodemask); | |
5809 | prev = sg; | |
5810 | ||
5811 | for (j = 0; j < MAX_NUMNODES; j++) { | |
5812 | cpumask_t tmp, notcovered; | |
5813 | int n = (i + j) % MAX_NUMNODES; | |
5814 | ||
5815 | cpus_complement(notcovered, covered); | |
5816 | cpus_and(tmp, notcovered, *cpu_map); | |
5817 | cpus_and(tmp, tmp, domainspan); | |
5818 | if (cpus_empty(tmp)) | |
5819 | break; | |
5820 | ||
5821 | nodemask = node_to_cpumask(n); | |
5822 | cpus_and(tmp, tmp, nodemask); | |
5823 | if (cpus_empty(tmp)) | |
5824 | continue; | |
5825 | ||
5826 | sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL); | |
5827 | if (!sg) { | |
5828 | printk(KERN_WARNING | |
5829 | "Can not alloc domain group for node %d\n", j); | |
5830 | break; | |
5831 | } | |
5832 | sg->cpu_power = 0; | |
5833 | sg->cpumask = tmp; | |
5834 | cpus_or(covered, covered, tmp); | |
5835 | prev->next = sg; | |
5836 | prev = sg; | |
5837 | } | |
5838 | prev->next = sched_group_nodes[i]; | |
5839 | } | |
1da177e4 LT |
5840 | #endif |
5841 | ||
5842 | /* Calculate CPU power for physical packages and nodes */ | |
1a20ff27 | 5843 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 LT |
5844 | int power; |
5845 | struct sched_domain *sd; | |
5846 | #ifdef CONFIG_SCHED_SMT | |
5847 | sd = &per_cpu(cpu_domains, i); | |
5848 | power = SCHED_LOAD_SCALE; | |
5849 | sd->groups->cpu_power = power; | |
5850 | #endif | |
1e9f28fa SS |
5851 | #ifdef CONFIG_SCHED_MC |
5852 | sd = &per_cpu(core_domains, i); | |
5853 | power = SCHED_LOAD_SCALE + (cpus_weight(sd->groups->cpumask)-1) | |
5854 | * SCHED_LOAD_SCALE / 10; | |
5855 | sd->groups->cpu_power = power; | |
5856 | ||
5857 | sd = &per_cpu(phys_domains, i); | |
1da177e4 | 5858 | |
1e9f28fa SS |
5859 | /* |
5860 | * This has to be < 2 * SCHED_LOAD_SCALE | |
5861 | * Lets keep it SCHED_LOAD_SCALE, so that | |
5862 | * while calculating NUMA group's cpu_power | |
5863 | * we can simply do | |
5864 | * numa_group->cpu_power += phys_group->cpu_power; | |
5865 | * | |
5866 | * See "only add power once for each physical pkg" | |
5867 | * comment below | |
5868 | */ | |
5869 | sd->groups->cpu_power = SCHED_LOAD_SCALE; | |
5870 | #else | |
1da177e4 LT |
5871 | sd = &per_cpu(phys_domains, i); |
5872 | power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE * | |
5873 | (cpus_weight(sd->groups->cpumask)-1) / 10; | |
5874 | sd->groups->cpu_power = power; | |
1e9f28fa | 5875 | #endif |
1da177e4 LT |
5876 | } |
5877 | ||
9c1cfda2 | 5878 | #ifdef CONFIG_NUMA |
08069033 SS |
5879 | for (i = 0; i < MAX_NUMNODES; i++) |
5880 | init_numa_sched_groups_power(sched_group_nodes[i]); | |
9c1cfda2 | 5881 | |
08069033 | 5882 | init_numa_sched_groups_power(sched_group_allnodes); |
9c1cfda2 JH |
5883 | #endif |
5884 | ||
1da177e4 | 5885 | /* Attach the domains */ |
1a20ff27 | 5886 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 LT |
5887 | struct sched_domain *sd; |
5888 | #ifdef CONFIG_SCHED_SMT | |
5889 | sd = &per_cpu(cpu_domains, i); | |
1e9f28fa SS |
5890 | #elif defined(CONFIG_SCHED_MC) |
5891 | sd = &per_cpu(core_domains, i); | |
1da177e4 LT |
5892 | #else |
5893 | sd = &per_cpu(phys_domains, i); | |
5894 | #endif | |
5895 | cpu_attach_domain(sd, i); | |
5896 | } | |
198e2f18 | 5897 | /* |
5898 | * Tune cache-hot values: | |
5899 | */ | |
5900 | calibrate_migration_costs(cpu_map); | |
1da177e4 | 5901 | } |
1a20ff27 DG |
5902 | /* |
5903 | * Set up scheduler domains and groups. Callers must hold the hotplug lock. | |
5904 | */ | |
9c1cfda2 | 5905 | static void arch_init_sched_domains(const cpumask_t *cpu_map) |
1a20ff27 DG |
5906 | { |
5907 | cpumask_t cpu_default_map; | |
1da177e4 | 5908 | |
1a20ff27 DG |
5909 | /* |
5910 | * Setup mask for cpus without special case scheduling requirements. | |
5911 | * For now this just excludes isolated cpus, but could be used to | |
5912 | * exclude other special cases in the future. | |
5913 | */ | |
5914 | cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map); | |
5915 | ||
5916 | build_sched_domains(&cpu_default_map); | |
5917 | } | |
5918 | ||
5919 | static void arch_destroy_sched_domains(const cpumask_t *cpu_map) | |
1da177e4 | 5920 | { |
9c1cfda2 JH |
5921 | #ifdef CONFIG_NUMA |
5922 | int i; | |
d1b55138 | 5923 | int cpu; |
1da177e4 | 5924 | |
d1b55138 JH |
5925 | for_each_cpu_mask(cpu, *cpu_map) { |
5926 | struct sched_group *sched_group_allnodes | |
5927 | = sched_group_allnodes_bycpu[cpu]; | |
5928 | struct sched_group **sched_group_nodes | |
5929 | = sched_group_nodes_bycpu[cpu]; | |
9c1cfda2 | 5930 | |
d1b55138 JH |
5931 | if (sched_group_allnodes) { |
5932 | kfree(sched_group_allnodes); | |
5933 | sched_group_allnodes_bycpu[cpu] = NULL; | |
5934 | } | |
5935 | ||
5936 | if (!sched_group_nodes) | |
9c1cfda2 | 5937 | continue; |
d1b55138 JH |
5938 | |
5939 | for (i = 0; i < MAX_NUMNODES; i++) { | |
5940 | cpumask_t nodemask = node_to_cpumask(i); | |
5941 | struct sched_group *oldsg, *sg = sched_group_nodes[i]; | |
5942 | ||
5943 | cpus_and(nodemask, nodemask, *cpu_map); | |
5944 | if (cpus_empty(nodemask)) | |
5945 | continue; | |
5946 | ||
5947 | if (sg == NULL) | |
5948 | continue; | |
5949 | sg = sg->next; | |
9c1cfda2 | 5950 | next_sg: |
d1b55138 JH |
5951 | oldsg = sg; |
5952 | sg = sg->next; | |
5953 | kfree(oldsg); | |
5954 | if (oldsg != sched_group_nodes[i]) | |
5955 | goto next_sg; | |
5956 | } | |
5957 | kfree(sched_group_nodes); | |
5958 | sched_group_nodes_bycpu[cpu] = NULL; | |
9c1cfda2 JH |
5959 | } |
5960 | #endif | |
5961 | } | |
1da177e4 | 5962 | |
1a20ff27 DG |
5963 | /* |
5964 | * Detach sched domains from a group of cpus specified in cpu_map | |
5965 | * These cpus will now be attached to the NULL domain | |
5966 | */ | |
858119e1 | 5967 | static void detach_destroy_domains(const cpumask_t *cpu_map) |
1a20ff27 DG |
5968 | { |
5969 | int i; | |
5970 | ||
5971 | for_each_cpu_mask(i, *cpu_map) | |
5972 | cpu_attach_domain(NULL, i); | |
5973 | synchronize_sched(); | |
5974 | arch_destroy_sched_domains(cpu_map); | |
5975 | } | |
5976 | ||
5977 | /* | |
5978 | * Partition sched domains as specified by the cpumasks below. | |
5979 | * This attaches all cpus from the cpumasks to the NULL domain, | |
5980 | * waits for a RCU quiescent period, recalculates sched | |
5981 | * domain information and then attaches them back to the | |
5982 | * correct sched domains | |
5983 | * Call with hotplug lock held | |
5984 | */ | |
5985 | void partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2) | |
5986 | { | |
5987 | cpumask_t change_map; | |
5988 | ||
5989 | cpus_and(*partition1, *partition1, cpu_online_map); | |
5990 | cpus_and(*partition2, *partition2, cpu_online_map); | |
5991 | cpus_or(change_map, *partition1, *partition2); | |
5992 | ||
5993 | /* Detach sched domains from all of the affected cpus */ | |
5994 | detach_destroy_domains(&change_map); | |
5995 | if (!cpus_empty(*partition1)) | |
5996 | build_sched_domains(partition1); | |
5997 | if (!cpus_empty(*partition2)) | |
5998 | build_sched_domains(partition2); | |
5999 | } | |
6000 | ||
1da177e4 LT |
6001 | #ifdef CONFIG_HOTPLUG_CPU |
6002 | /* | |
6003 | * Force a reinitialization of the sched domains hierarchy. The domains | |
6004 | * and groups cannot be updated in place without racing with the balancing | |
41c7ce9a | 6005 | * code, so we temporarily attach all running cpus to the NULL domain |
1da177e4 LT |
6006 | * which will prevent rebalancing while the sched domains are recalculated. |
6007 | */ | |
6008 | static int update_sched_domains(struct notifier_block *nfb, | |
6009 | unsigned long action, void *hcpu) | |
6010 | { | |
1da177e4 LT |
6011 | switch (action) { |
6012 | case CPU_UP_PREPARE: | |
6013 | case CPU_DOWN_PREPARE: | |
1a20ff27 | 6014 | detach_destroy_domains(&cpu_online_map); |
1da177e4 LT |
6015 | return NOTIFY_OK; |
6016 | ||
6017 | case CPU_UP_CANCELED: | |
6018 | case CPU_DOWN_FAILED: | |
6019 | case CPU_ONLINE: | |
6020 | case CPU_DEAD: | |
6021 | /* | |
6022 | * Fall through and re-initialise the domains. | |
6023 | */ | |
6024 | break; | |
6025 | default: | |
6026 | return NOTIFY_DONE; | |
6027 | } | |
6028 | ||
6029 | /* The hotplug lock is already held by cpu_up/cpu_down */ | |
1a20ff27 | 6030 | arch_init_sched_domains(&cpu_online_map); |
1da177e4 LT |
6031 | |
6032 | return NOTIFY_OK; | |
6033 | } | |
6034 | #endif | |
6035 | ||
6036 | void __init sched_init_smp(void) | |
6037 | { | |
6038 | lock_cpu_hotplug(); | |
1a20ff27 | 6039 | arch_init_sched_domains(&cpu_online_map); |
1da177e4 LT |
6040 | unlock_cpu_hotplug(); |
6041 | /* XXX: Theoretical race here - CPU may be hotplugged now */ | |
6042 | hotcpu_notifier(update_sched_domains, 0); | |
6043 | } | |
6044 | #else | |
6045 | void __init sched_init_smp(void) | |
6046 | { | |
6047 | } | |
6048 | #endif /* CONFIG_SMP */ | |
6049 | ||
6050 | int in_sched_functions(unsigned long addr) | |
6051 | { | |
6052 | /* Linker adds these: start and end of __sched functions */ | |
6053 | extern char __sched_text_start[], __sched_text_end[]; | |
6054 | return in_lock_functions(addr) || | |
6055 | (addr >= (unsigned long)__sched_text_start | |
6056 | && addr < (unsigned long)__sched_text_end); | |
6057 | } | |
6058 | ||
6059 | void __init sched_init(void) | |
6060 | { | |
6061 | runqueue_t *rq; | |
6062 | int i, j, k; | |
6063 | ||
0a945022 | 6064 | for_each_possible_cpu(i) { |
1da177e4 LT |
6065 | prio_array_t *array; |
6066 | ||
6067 | rq = cpu_rq(i); | |
6068 | spin_lock_init(&rq->lock); | |
7897986b | 6069 | rq->nr_running = 0; |
1da177e4 LT |
6070 | rq->active = rq->arrays; |
6071 | rq->expired = rq->arrays + 1; | |
6072 | rq->best_expired_prio = MAX_PRIO; | |
6073 | ||
6074 | #ifdef CONFIG_SMP | |
41c7ce9a | 6075 | rq->sd = NULL; |
7897986b NP |
6076 | for (j = 1; j < 3; j++) |
6077 | rq->cpu_load[j] = 0; | |
1da177e4 LT |
6078 | rq->active_balance = 0; |
6079 | rq->push_cpu = 0; | |
6080 | rq->migration_thread = NULL; | |
6081 | INIT_LIST_HEAD(&rq->migration_queue); | |
6082 | #endif | |
6083 | atomic_set(&rq->nr_iowait, 0); | |
6084 | ||
6085 | for (j = 0; j < 2; j++) { | |
6086 | array = rq->arrays + j; | |
6087 | for (k = 0; k < MAX_PRIO; k++) { | |
6088 | INIT_LIST_HEAD(array->queue + k); | |
6089 | __clear_bit(k, array->bitmap); | |
6090 | } | |
6091 | // delimiter for bitsearch | |
6092 | __set_bit(MAX_PRIO, array->bitmap); | |
6093 | } | |
6094 | } | |
6095 | ||
6096 | /* | |
6097 | * The boot idle thread does lazy MMU switching as well: | |
6098 | */ | |
6099 | atomic_inc(&init_mm.mm_count); | |
6100 | enter_lazy_tlb(&init_mm, current); | |
6101 | ||
6102 | /* | |
6103 | * Make us the idle thread. Technically, schedule() should not be | |
6104 | * called from this thread, however somewhere below it might be, | |
6105 | * but because we are the idle thread, we just pick up running again | |
6106 | * when this runqueue becomes "idle". | |
6107 | */ | |
6108 | init_idle(current, smp_processor_id()); | |
6109 | } | |
6110 | ||
6111 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP | |
6112 | void __might_sleep(char *file, int line) | |
6113 | { | |
6114 | #if defined(in_atomic) | |
6115 | static unsigned long prev_jiffy; /* ratelimiting */ | |
6116 | ||
6117 | if ((in_atomic() || irqs_disabled()) && | |
6118 | system_state == SYSTEM_RUNNING && !oops_in_progress) { | |
6119 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) | |
6120 | return; | |
6121 | prev_jiffy = jiffies; | |
91368d73 | 6122 | printk(KERN_ERR "BUG: sleeping function called from invalid" |
1da177e4 LT |
6123 | " context at %s:%d\n", file, line); |
6124 | printk("in_atomic():%d, irqs_disabled():%d\n", | |
6125 | in_atomic(), irqs_disabled()); | |
6126 | dump_stack(); | |
6127 | } | |
6128 | #endif | |
6129 | } | |
6130 | EXPORT_SYMBOL(__might_sleep); | |
6131 | #endif | |
6132 | ||
6133 | #ifdef CONFIG_MAGIC_SYSRQ | |
6134 | void normalize_rt_tasks(void) | |
6135 | { | |
6136 | struct task_struct *p; | |
6137 | prio_array_t *array; | |
6138 | unsigned long flags; | |
6139 | runqueue_t *rq; | |
6140 | ||
6141 | read_lock_irq(&tasklist_lock); | |
c96d145e | 6142 | for_each_process(p) { |
1da177e4 LT |
6143 | if (!rt_task(p)) |
6144 | continue; | |
6145 | ||
6146 | rq = task_rq_lock(p, &flags); | |
6147 | ||
6148 | array = p->array; | |
6149 | if (array) | |
6150 | deactivate_task(p, task_rq(p)); | |
6151 | __setscheduler(p, SCHED_NORMAL, 0); | |
6152 | if (array) { | |
6153 | __activate_task(p, task_rq(p)); | |
6154 | resched_task(rq->curr); | |
6155 | } | |
6156 | ||
6157 | task_rq_unlock(rq, &flags); | |
6158 | } | |
6159 | read_unlock_irq(&tasklist_lock); | |
6160 | } | |
6161 | ||
6162 | #endif /* CONFIG_MAGIC_SYSRQ */ | |
1df5c10a LT |
6163 | |
6164 | #ifdef CONFIG_IA64 | |
6165 | /* | |
6166 | * These functions are only useful for the IA64 MCA handling. | |
6167 | * | |
6168 | * They can only be called when the whole system has been | |
6169 | * stopped - every CPU needs to be quiescent, and no scheduling | |
6170 | * activity can take place. Using them for anything else would | |
6171 | * be a serious bug, and as a result, they aren't even visible | |
6172 | * under any other configuration. | |
6173 | */ | |
6174 | ||
6175 | /** | |
6176 | * curr_task - return the current task for a given cpu. | |
6177 | * @cpu: the processor in question. | |
6178 | * | |
6179 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
6180 | */ | |
6181 | task_t *curr_task(int cpu) | |
6182 | { | |
6183 | return cpu_curr(cpu); | |
6184 | } | |
6185 | ||
6186 | /** | |
6187 | * set_curr_task - set the current task for a given cpu. | |
6188 | * @cpu: the processor in question. | |
6189 | * @p: the task pointer to set. | |
6190 | * | |
6191 | * Description: This function must only be used when non-maskable interrupts | |
6192 | * are serviced on a separate stack. It allows the architecture to switch the | |
6193 | * notion of the current task on a cpu in a non-blocking manner. This function | |
6194 | * must be called with all CPU's synchronized, and interrupts disabled, the | |
6195 | * and caller must save the original value of the current task (see | |
6196 | * curr_task() above) and restore that value before reenabling interrupts and | |
6197 | * re-starting the system. | |
6198 | * | |
6199 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
6200 | */ | |
6201 | void set_curr_task(int cpu, task_t *p) | |
6202 | { | |
6203 | cpu_curr(cpu) = p; | |
6204 | } | |
6205 | ||
6206 | #endif |