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 | ||
187 | #define BITMAP_SIZE ((((MAX_PRIO+1+7)/8)+sizeof(long)-1)/sizeof(long)) | |
188 | ||
189 | typedef struct runqueue runqueue_t; | |
190 | ||
191 | struct prio_array { | |
192 | unsigned int nr_active; | |
193 | unsigned long bitmap[BITMAP_SIZE]; | |
194 | struct list_head queue[MAX_PRIO]; | |
195 | }; | |
196 | ||
197 | /* | |
198 | * This is the main, per-CPU runqueue data structure. | |
199 | * | |
200 | * Locking rule: those places that want to lock multiple runqueues | |
201 | * (such as the load balancing or the thread migration code), lock | |
202 | * acquire operations must be ordered by ascending &runqueue. | |
203 | */ | |
204 | struct runqueue { | |
205 | spinlock_t lock; | |
206 | ||
207 | /* | |
208 | * nr_running and cpu_load should be in the same cacheline because | |
209 | * remote CPUs use both these fields when doing load calculation. | |
210 | */ | |
211 | unsigned long nr_running; | |
212 | #ifdef CONFIG_SMP | |
7897986b | 213 | unsigned long cpu_load[3]; |
1da177e4 LT |
214 | #endif |
215 | unsigned long long nr_switches; | |
216 | ||
217 | /* | |
218 | * This is part of a global counter where only the total sum | |
219 | * over all CPUs matters. A task can increase this counter on | |
220 | * one CPU and if it got migrated afterwards it may decrease | |
221 | * it on another CPU. Always updated under the runqueue lock: | |
222 | */ | |
223 | unsigned long nr_uninterruptible; | |
224 | ||
225 | unsigned long expired_timestamp; | |
226 | unsigned long long timestamp_last_tick; | |
227 | task_t *curr, *idle; | |
228 | struct mm_struct *prev_mm; | |
229 | prio_array_t *active, *expired, arrays[2]; | |
230 | int best_expired_prio; | |
231 | atomic_t nr_iowait; | |
232 | ||
233 | #ifdef CONFIG_SMP | |
234 | struct sched_domain *sd; | |
235 | ||
236 | /* For active balancing */ | |
237 | int active_balance; | |
238 | int push_cpu; | |
239 | ||
240 | task_t *migration_thread; | |
241 | struct list_head migration_queue; | |
e9028b0f | 242 | int cpu; |
1da177e4 LT |
243 | #endif |
244 | ||
245 | #ifdef CONFIG_SCHEDSTATS | |
246 | /* latency stats */ | |
247 | struct sched_info rq_sched_info; | |
248 | ||
249 | /* sys_sched_yield() stats */ | |
250 | unsigned long yld_exp_empty; | |
251 | unsigned long yld_act_empty; | |
252 | unsigned long yld_both_empty; | |
253 | unsigned long yld_cnt; | |
254 | ||
255 | /* schedule() stats */ | |
256 | unsigned long sched_switch; | |
257 | unsigned long sched_cnt; | |
258 | unsigned long sched_goidle; | |
259 | ||
260 | /* try_to_wake_up() stats */ | |
261 | unsigned long ttwu_cnt; | |
262 | unsigned long ttwu_local; | |
263 | #endif | |
264 | }; | |
265 | ||
266 | static DEFINE_PER_CPU(struct runqueue, runqueues); | |
267 | ||
674311d5 NP |
268 | /* |
269 | * The domain tree (rq->sd) is protected by RCU's quiescent state transition. | |
1a20ff27 | 270 | * See detach_destroy_domains: synchronize_sched for details. |
674311d5 NP |
271 | * |
272 | * The domain tree of any CPU may only be accessed from within | |
273 | * preempt-disabled sections. | |
274 | */ | |
1da177e4 | 275 | #define for_each_domain(cpu, domain) \ |
674311d5 | 276 | for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent) |
1da177e4 LT |
277 | |
278 | #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) | |
279 | #define this_rq() (&__get_cpu_var(runqueues)) | |
280 | #define task_rq(p) cpu_rq(task_cpu(p)) | |
281 | #define cpu_curr(cpu) (cpu_rq(cpu)->curr) | |
282 | ||
1da177e4 | 283 | #ifndef prepare_arch_switch |
4866cde0 NP |
284 | # define prepare_arch_switch(next) do { } while (0) |
285 | #endif | |
286 | #ifndef finish_arch_switch | |
287 | # define finish_arch_switch(prev) do { } while (0) | |
288 | #endif | |
289 | ||
290 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW | |
291 | static inline int task_running(runqueue_t *rq, task_t *p) | |
292 | { | |
293 | return rq->curr == p; | |
294 | } | |
295 | ||
296 | static inline void prepare_lock_switch(runqueue_t *rq, task_t *next) | |
297 | { | |
298 | } | |
299 | ||
300 | static inline void finish_lock_switch(runqueue_t *rq, task_t *prev) | |
301 | { | |
da04c035 IM |
302 | #ifdef CONFIG_DEBUG_SPINLOCK |
303 | /* this is a valid case when another task releases the spinlock */ | |
304 | rq->lock.owner = current; | |
305 | #endif | |
4866cde0 NP |
306 | spin_unlock_irq(&rq->lock); |
307 | } | |
308 | ||
309 | #else /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
310 | static inline int task_running(runqueue_t *rq, task_t *p) | |
311 | { | |
312 | #ifdef CONFIG_SMP | |
313 | return p->oncpu; | |
314 | #else | |
315 | return rq->curr == p; | |
316 | #endif | |
317 | } | |
318 | ||
319 | static inline void prepare_lock_switch(runqueue_t *rq, task_t *next) | |
320 | { | |
321 | #ifdef CONFIG_SMP | |
322 | /* | |
323 | * We can optimise this out completely for !SMP, because the | |
324 | * SMP rebalancing from interrupt is the only thing that cares | |
325 | * here. | |
326 | */ | |
327 | next->oncpu = 1; | |
328 | #endif | |
329 | #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
330 | spin_unlock_irq(&rq->lock); | |
331 | #else | |
332 | spin_unlock(&rq->lock); | |
333 | #endif | |
334 | } | |
335 | ||
336 | static inline void finish_lock_switch(runqueue_t *rq, task_t *prev) | |
337 | { | |
338 | #ifdef CONFIG_SMP | |
339 | /* | |
340 | * After ->oncpu is cleared, the task can be moved to a different CPU. | |
341 | * We must ensure this doesn't happen until the switch is completely | |
342 | * finished. | |
343 | */ | |
344 | smp_wmb(); | |
345 | prev->oncpu = 0; | |
346 | #endif | |
347 | #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
348 | local_irq_enable(); | |
1da177e4 | 349 | #endif |
4866cde0 NP |
350 | } |
351 | #endif /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
1da177e4 LT |
352 | |
353 | /* | |
354 | * task_rq_lock - lock the runqueue a given task resides on and disable | |
355 | * interrupts. Note the ordering: we can safely lookup the task_rq without | |
356 | * explicitly disabling preemption. | |
357 | */ | |
358 | static inline runqueue_t *task_rq_lock(task_t *p, unsigned long *flags) | |
359 | __acquires(rq->lock) | |
360 | { | |
361 | struct runqueue *rq; | |
362 | ||
363 | repeat_lock_task: | |
364 | local_irq_save(*flags); | |
365 | rq = task_rq(p); | |
366 | spin_lock(&rq->lock); | |
367 | if (unlikely(rq != task_rq(p))) { | |
368 | spin_unlock_irqrestore(&rq->lock, *flags); | |
369 | goto repeat_lock_task; | |
370 | } | |
371 | return rq; | |
372 | } | |
373 | ||
374 | static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags) | |
375 | __releases(rq->lock) | |
376 | { | |
377 | spin_unlock_irqrestore(&rq->lock, *flags); | |
378 | } | |
379 | ||
380 | #ifdef CONFIG_SCHEDSTATS | |
381 | /* | |
382 | * bump this up when changing the output format or the meaning of an existing | |
383 | * format, so that tools can adapt (or abort) | |
384 | */ | |
68767a0a | 385 | #define SCHEDSTAT_VERSION 12 |
1da177e4 LT |
386 | |
387 | static int show_schedstat(struct seq_file *seq, void *v) | |
388 | { | |
389 | int cpu; | |
390 | ||
391 | seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION); | |
392 | seq_printf(seq, "timestamp %lu\n", jiffies); | |
393 | for_each_online_cpu(cpu) { | |
394 | runqueue_t *rq = cpu_rq(cpu); | |
395 | #ifdef CONFIG_SMP | |
396 | struct sched_domain *sd; | |
397 | int dcnt = 0; | |
398 | #endif | |
399 | ||
400 | /* runqueue-specific stats */ | |
401 | seq_printf(seq, | |
402 | "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu", | |
403 | cpu, rq->yld_both_empty, | |
404 | rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt, | |
405 | rq->sched_switch, rq->sched_cnt, rq->sched_goidle, | |
406 | rq->ttwu_cnt, rq->ttwu_local, | |
407 | rq->rq_sched_info.cpu_time, | |
408 | rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt); | |
409 | ||
410 | seq_printf(seq, "\n"); | |
411 | ||
412 | #ifdef CONFIG_SMP | |
413 | /* domain-specific stats */ | |
674311d5 | 414 | preempt_disable(); |
1da177e4 LT |
415 | for_each_domain(cpu, sd) { |
416 | enum idle_type itype; | |
417 | char mask_str[NR_CPUS]; | |
418 | ||
419 | cpumask_scnprintf(mask_str, NR_CPUS, sd->span); | |
420 | seq_printf(seq, "domain%d %s", dcnt++, mask_str); | |
421 | for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES; | |
422 | itype++) { | |
423 | seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu", | |
424 | sd->lb_cnt[itype], | |
425 | sd->lb_balanced[itype], | |
426 | sd->lb_failed[itype], | |
427 | sd->lb_imbalance[itype], | |
428 | sd->lb_gained[itype], | |
429 | sd->lb_hot_gained[itype], | |
430 | sd->lb_nobusyq[itype], | |
431 | sd->lb_nobusyg[itype]); | |
432 | } | |
68767a0a | 433 | seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n", |
1da177e4 | 434 | sd->alb_cnt, sd->alb_failed, sd->alb_pushed, |
68767a0a NP |
435 | sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed, |
436 | sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed, | |
1da177e4 LT |
437 | sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance); |
438 | } | |
674311d5 | 439 | preempt_enable(); |
1da177e4 LT |
440 | #endif |
441 | } | |
442 | return 0; | |
443 | } | |
444 | ||
445 | static int schedstat_open(struct inode *inode, struct file *file) | |
446 | { | |
447 | unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32); | |
448 | char *buf = kmalloc(size, GFP_KERNEL); | |
449 | struct seq_file *m; | |
450 | int res; | |
451 | ||
452 | if (!buf) | |
453 | return -ENOMEM; | |
454 | res = single_open(file, show_schedstat, NULL); | |
455 | if (!res) { | |
456 | m = file->private_data; | |
457 | m->buf = buf; | |
458 | m->size = size; | |
459 | } else | |
460 | kfree(buf); | |
461 | return res; | |
462 | } | |
463 | ||
464 | struct file_operations proc_schedstat_operations = { | |
465 | .open = schedstat_open, | |
466 | .read = seq_read, | |
467 | .llseek = seq_lseek, | |
468 | .release = single_release, | |
469 | }; | |
470 | ||
471 | # define schedstat_inc(rq, field) do { (rq)->field++; } while (0) | |
472 | # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0) | |
473 | #else /* !CONFIG_SCHEDSTATS */ | |
474 | # define schedstat_inc(rq, field) do { } while (0) | |
475 | # define schedstat_add(rq, field, amt) do { } while (0) | |
476 | #endif | |
477 | ||
478 | /* | |
479 | * rq_lock - lock a given runqueue and disable interrupts. | |
480 | */ | |
481 | static inline runqueue_t *this_rq_lock(void) | |
482 | __acquires(rq->lock) | |
483 | { | |
484 | runqueue_t *rq; | |
485 | ||
486 | local_irq_disable(); | |
487 | rq = this_rq(); | |
488 | spin_lock(&rq->lock); | |
489 | ||
490 | return rq; | |
491 | } | |
492 | ||
1da177e4 LT |
493 | #ifdef CONFIG_SCHEDSTATS |
494 | /* | |
495 | * Called when a process is dequeued from the active array and given | |
496 | * the cpu. We should note that with the exception of interactive | |
497 | * tasks, the expired queue will become the active queue after the active | |
498 | * queue is empty, without explicitly dequeuing and requeuing tasks in the | |
499 | * expired queue. (Interactive tasks may be requeued directly to the | |
500 | * active queue, thus delaying tasks in the expired queue from running; | |
501 | * see scheduler_tick()). | |
502 | * | |
503 | * This function is only called from sched_info_arrive(), rather than | |
504 | * dequeue_task(). Even though a task may be queued and dequeued multiple | |
505 | * times as it is shuffled about, we're really interested in knowing how | |
506 | * long it was from the *first* time it was queued to the time that it | |
507 | * finally hit a cpu. | |
508 | */ | |
509 | static inline void sched_info_dequeued(task_t *t) | |
510 | { | |
511 | t->sched_info.last_queued = 0; | |
512 | } | |
513 | ||
514 | /* | |
515 | * Called when a task finally hits the cpu. We can now calculate how | |
516 | * long it was waiting to run. We also note when it began so that we | |
517 | * can keep stats on how long its timeslice is. | |
518 | */ | |
858119e1 | 519 | static void sched_info_arrive(task_t *t) |
1da177e4 LT |
520 | { |
521 | unsigned long now = jiffies, diff = 0; | |
522 | struct runqueue *rq = task_rq(t); | |
523 | ||
524 | if (t->sched_info.last_queued) | |
525 | diff = now - t->sched_info.last_queued; | |
526 | sched_info_dequeued(t); | |
527 | t->sched_info.run_delay += diff; | |
528 | t->sched_info.last_arrival = now; | |
529 | t->sched_info.pcnt++; | |
530 | ||
531 | if (!rq) | |
532 | return; | |
533 | ||
534 | rq->rq_sched_info.run_delay += diff; | |
535 | rq->rq_sched_info.pcnt++; | |
536 | } | |
537 | ||
538 | /* | |
539 | * Called when a process is queued into either the active or expired | |
540 | * array. The time is noted and later used to determine how long we | |
541 | * had to wait for us to reach the cpu. Since the expired queue will | |
542 | * become the active queue after active queue is empty, without dequeuing | |
543 | * and requeuing any tasks, we are interested in queuing to either. It | |
544 | * is unusual but not impossible for tasks to be dequeued and immediately | |
545 | * requeued in the same or another array: this can happen in sched_yield(), | |
546 | * set_user_nice(), and even load_balance() as it moves tasks from runqueue | |
547 | * to runqueue. | |
548 | * | |
549 | * This function is only called from enqueue_task(), but also only updates | |
550 | * the timestamp if it is already not set. It's assumed that | |
551 | * sched_info_dequeued() will clear that stamp when appropriate. | |
552 | */ | |
553 | static inline void sched_info_queued(task_t *t) | |
554 | { | |
555 | if (!t->sched_info.last_queued) | |
556 | t->sched_info.last_queued = jiffies; | |
557 | } | |
558 | ||
559 | /* | |
560 | * Called when a process ceases being the active-running process, either | |
561 | * voluntarily or involuntarily. Now we can calculate how long we ran. | |
562 | */ | |
563 | static inline void sched_info_depart(task_t *t) | |
564 | { | |
565 | struct runqueue *rq = task_rq(t); | |
566 | unsigned long diff = jiffies - t->sched_info.last_arrival; | |
567 | ||
568 | t->sched_info.cpu_time += diff; | |
569 | ||
570 | if (rq) | |
571 | rq->rq_sched_info.cpu_time += diff; | |
572 | } | |
573 | ||
574 | /* | |
575 | * Called when tasks are switched involuntarily due, typically, to expiring | |
576 | * their time slice. (This may also be called when switching to or from | |
577 | * the idle task.) We are only called when prev != next. | |
578 | */ | |
579 | static inline void sched_info_switch(task_t *prev, task_t *next) | |
580 | { | |
581 | struct runqueue *rq = task_rq(prev); | |
582 | ||
583 | /* | |
584 | * prev now departs the cpu. It's not interesting to record | |
585 | * stats about how efficient we were at scheduling the idle | |
586 | * process, however. | |
587 | */ | |
588 | if (prev != rq->idle) | |
589 | sched_info_depart(prev); | |
590 | ||
591 | if (next != rq->idle) | |
592 | sched_info_arrive(next); | |
593 | } | |
594 | #else | |
595 | #define sched_info_queued(t) do { } while (0) | |
596 | #define sched_info_switch(t, next) do { } while (0) | |
597 | #endif /* CONFIG_SCHEDSTATS */ | |
598 | ||
599 | /* | |
600 | * Adding/removing a task to/from a priority array: | |
601 | */ | |
602 | static void dequeue_task(struct task_struct *p, prio_array_t *array) | |
603 | { | |
604 | array->nr_active--; | |
605 | list_del(&p->run_list); | |
606 | if (list_empty(array->queue + p->prio)) | |
607 | __clear_bit(p->prio, array->bitmap); | |
608 | } | |
609 | ||
610 | static void enqueue_task(struct task_struct *p, prio_array_t *array) | |
611 | { | |
612 | sched_info_queued(p); | |
613 | list_add_tail(&p->run_list, array->queue + p->prio); | |
614 | __set_bit(p->prio, array->bitmap); | |
615 | array->nr_active++; | |
616 | p->array = array; | |
617 | } | |
618 | ||
619 | /* | |
620 | * Put task to the end of the run list without the overhead of dequeue | |
621 | * followed by enqueue. | |
622 | */ | |
623 | static void requeue_task(struct task_struct *p, prio_array_t *array) | |
624 | { | |
625 | list_move_tail(&p->run_list, array->queue + p->prio); | |
626 | } | |
627 | ||
628 | static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array) | |
629 | { | |
630 | list_add(&p->run_list, array->queue + p->prio); | |
631 | __set_bit(p->prio, array->bitmap); | |
632 | array->nr_active++; | |
633 | p->array = array; | |
634 | } | |
635 | ||
636 | /* | |
637 | * effective_prio - return the priority that is based on the static | |
638 | * priority but is modified by bonuses/penalties. | |
639 | * | |
640 | * We scale the actual sleep average [0 .... MAX_SLEEP_AVG] | |
641 | * into the -5 ... 0 ... +5 bonus/penalty range. | |
642 | * | |
643 | * We use 25% of the full 0...39 priority range so that: | |
644 | * | |
645 | * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs. | |
646 | * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks. | |
647 | * | |
648 | * Both properties are important to certain workloads. | |
649 | */ | |
650 | static int effective_prio(task_t *p) | |
651 | { | |
652 | int bonus, prio; | |
653 | ||
654 | if (rt_task(p)) | |
655 | return p->prio; | |
656 | ||
657 | bonus = CURRENT_BONUS(p) - MAX_BONUS / 2; | |
658 | ||
659 | prio = p->static_prio - bonus; | |
660 | if (prio < MAX_RT_PRIO) | |
661 | prio = MAX_RT_PRIO; | |
662 | if (prio > MAX_PRIO-1) | |
663 | prio = MAX_PRIO-1; | |
664 | return prio; | |
665 | } | |
666 | ||
667 | /* | |
668 | * __activate_task - move a task to the runqueue. | |
669 | */ | |
670 | static inline void __activate_task(task_t *p, runqueue_t *rq) | |
671 | { | |
672 | enqueue_task(p, rq->active); | |
a2000572 | 673 | rq->nr_running++; |
1da177e4 LT |
674 | } |
675 | ||
676 | /* | |
677 | * __activate_idle_task - move idle task to the _front_ of runqueue. | |
678 | */ | |
679 | static inline void __activate_idle_task(task_t *p, runqueue_t *rq) | |
680 | { | |
681 | enqueue_task_head(p, rq->active); | |
a2000572 | 682 | rq->nr_running++; |
1da177e4 LT |
683 | } |
684 | ||
a3464a10 | 685 | static int recalc_task_prio(task_t *p, unsigned long long now) |
1da177e4 LT |
686 | { |
687 | /* Caller must always ensure 'now >= p->timestamp' */ | |
688 | unsigned long long __sleep_time = now - p->timestamp; | |
689 | unsigned long sleep_time; | |
690 | ||
b0a9499c IM |
691 | if (unlikely(p->policy == SCHED_BATCH)) |
692 | sleep_time = 0; | |
693 | else { | |
694 | if (__sleep_time > NS_MAX_SLEEP_AVG) | |
695 | sleep_time = NS_MAX_SLEEP_AVG; | |
696 | else | |
697 | sleep_time = (unsigned long)__sleep_time; | |
698 | } | |
1da177e4 LT |
699 | |
700 | if (likely(sleep_time > 0)) { | |
701 | /* | |
702 | * User tasks that sleep a long time are categorised as | |
703 | * idle and will get just interactive status to stay active & | |
704 | * prevent them suddenly becoming cpu hogs and starving | |
705 | * other processes. | |
706 | */ | |
3dee386e | 707 | if (p->mm && p->sleep_type != SLEEP_NONINTERACTIVE && |
1da177e4 LT |
708 | sleep_time > INTERACTIVE_SLEEP(p)) { |
709 | p->sleep_avg = JIFFIES_TO_NS(MAX_SLEEP_AVG - | |
710 | DEF_TIMESLICE); | |
711 | } else { | |
1da177e4 LT |
712 | /* |
713 | * Tasks waking from uninterruptible sleep are | |
714 | * limited in their sleep_avg rise as they | |
715 | * are likely to be waiting on I/O | |
716 | */ | |
3dee386e | 717 | if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) { |
1da177e4 LT |
718 | if (p->sleep_avg >= INTERACTIVE_SLEEP(p)) |
719 | sleep_time = 0; | |
720 | else if (p->sleep_avg + sleep_time >= | |
721 | INTERACTIVE_SLEEP(p)) { | |
722 | p->sleep_avg = INTERACTIVE_SLEEP(p); | |
723 | sleep_time = 0; | |
724 | } | |
725 | } | |
726 | ||
727 | /* | |
728 | * This code gives a bonus to interactive tasks. | |
729 | * | |
730 | * The boost works by updating the 'average sleep time' | |
731 | * value here, based on ->timestamp. The more time a | |
732 | * task spends sleeping, the higher the average gets - | |
733 | * and the higher the priority boost gets as well. | |
734 | */ | |
735 | p->sleep_avg += sleep_time; | |
736 | ||
737 | if (p->sleep_avg > NS_MAX_SLEEP_AVG) | |
738 | p->sleep_avg = NS_MAX_SLEEP_AVG; | |
739 | } | |
740 | } | |
741 | ||
a3464a10 | 742 | return effective_prio(p); |
1da177e4 LT |
743 | } |
744 | ||
745 | /* | |
746 | * activate_task - move a task to the runqueue and do priority recalculation | |
747 | * | |
748 | * Update all the scheduling statistics stuff. (sleep average | |
749 | * calculation, priority modifiers, etc.) | |
750 | */ | |
751 | static void activate_task(task_t *p, runqueue_t *rq, int local) | |
752 | { | |
753 | unsigned long long now; | |
754 | ||
755 | now = sched_clock(); | |
756 | #ifdef CONFIG_SMP | |
757 | if (!local) { | |
758 | /* Compensate for drifting sched_clock */ | |
759 | runqueue_t *this_rq = this_rq(); | |
760 | now = (now - this_rq->timestamp_last_tick) | |
761 | + rq->timestamp_last_tick; | |
762 | } | |
763 | #endif | |
764 | ||
a47ab937 KC |
765 | if (!rt_task(p)) |
766 | p->prio = recalc_task_prio(p, now); | |
1da177e4 LT |
767 | |
768 | /* | |
769 | * This checks to make sure it's not an uninterruptible task | |
770 | * that is now waking up. | |
771 | */ | |
3dee386e | 772 | if (p->sleep_type == SLEEP_NORMAL) { |
1da177e4 LT |
773 | /* |
774 | * Tasks which were woken up by interrupts (ie. hw events) | |
775 | * are most likely of interactive nature. So we give them | |
776 | * the credit of extending their sleep time to the period | |
777 | * of time they spend on the runqueue, waiting for execution | |
778 | * on a CPU, first time around: | |
779 | */ | |
780 | if (in_interrupt()) | |
3dee386e | 781 | p->sleep_type = SLEEP_INTERRUPTED; |
1da177e4 LT |
782 | else { |
783 | /* | |
784 | * Normal first-time wakeups get a credit too for | |
785 | * on-runqueue time, but it will be weighted down: | |
786 | */ | |
3dee386e | 787 | p->sleep_type = SLEEP_INTERACTIVE; |
1da177e4 LT |
788 | } |
789 | } | |
790 | p->timestamp = now; | |
791 | ||
792 | __activate_task(p, rq); | |
793 | } | |
794 | ||
795 | /* | |
796 | * deactivate_task - remove a task from the runqueue. | |
797 | */ | |
798 | static void deactivate_task(struct task_struct *p, runqueue_t *rq) | |
799 | { | |
a2000572 | 800 | rq->nr_running--; |
1da177e4 LT |
801 | dequeue_task(p, p->array); |
802 | p->array = NULL; | |
803 | } | |
804 | ||
805 | /* | |
806 | * resched_task - mark a task 'to be rescheduled now'. | |
807 | * | |
808 | * On UP this means the setting of the need_resched flag, on SMP it | |
809 | * might also involve a cross-CPU call to trigger the scheduler on | |
810 | * the target CPU. | |
811 | */ | |
812 | #ifdef CONFIG_SMP | |
813 | static void resched_task(task_t *p) | |
814 | { | |
64c7c8f8 | 815 | int cpu; |
1da177e4 LT |
816 | |
817 | assert_spin_locked(&task_rq(p)->lock); | |
818 | ||
64c7c8f8 NP |
819 | if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED))) |
820 | return; | |
821 | ||
822 | set_tsk_thread_flag(p, TIF_NEED_RESCHED); | |
1da177e4 | 823 | |
64c7c8f8 NP |
824 | cpu = task_cpu(p); |
825 | if (cpu == smp_processor_id()) | |
826 | return; | |
827 | ||
828 | /* NEED_RESCHED must be visible before we test POLLING_NRFLAG */ | |
829 | smp_mb(); | |
830 | if (!test_tsk_thread_flag(p, TIF_POLLING_NRFLAG)) | |
831 | smp_send_reschedule(cpu); | |
1da177e4 LT |
832 | } |
833 | #else | |
834 | static inline void resched_task(task_t *p) | |
835 | { | |
64c7c8f8 | 836 | assert_spin_locked(&task_rq(p)->lock); |
1da177e4 LT |
837 | set_tsk_need_resched(p); |
838 | } | |
839 | #endif | |
840 | ||
841 | /** | |
842 | * task_curr - is this task currently executing on a CPU? | |
843 | * @p: the task in question. | |
844 | */ | |
845 | inline int task_curr(const task_t *p) | |
846 | { | |
847 | return cpu_curr(task_cpu(p)) == p; | |
848 | } | |
849 | ||
850 | #ifdef CONFIG_SMP | |
1da177e4 LT |
851 | typedef struct { |
852 | struct list_head list; | |
1da177e4 | 853 | |
1da177e4 LT |
854 | task_t *task; |
855 | int dest_cpu; | |
856 | ||
1da177e4 LT |
857 | struct completion done; |
858 | } migration_req_t; | |
859 | ||
860 | /* | |
861 | * The task's runqueue lock must be held. | |
862 | * Returns true if you have to wait for migration thread. | |
863 | */ | |
864 | static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req) | |
865 | { | |
866 | runqueue_t *rq = task_rq(p); | |
867 | ||
868 | /* | |
869 | * If the task is not on a runqueue (and not running), then | |
870 | * it is sufficient to simply update the task's cpu field. | |
871 | */ | |
872 | if (!p->array && !task_running(rq, p)) { | |
873 | set_task_cpu(p, dest_cpu); | |
874 | return 0; | |
875 | } | |
876 | ||
877 | init_completion(&req->done); | |
1da177e4 LT |
878 | req->task = p; |
879 | req->dest_cpu = dest_cpu; | |
880 | list_add(&req->list, &rq->migration_queue); | |
881 | return 1; | |
882 | } | |
883 | ||
884 | /* | |
885 | * wait_task_inactive - wait for a thread to unschedule. | |
886 | * | |
887 | * The caller must ensure that the task *will* unschedule sometime soon, | |
888 | * else this function might spin for a *long* time. This function can't | |
889 | * be called with interrupts off, or it may introduce deadlock with | |
890 | * smp_call_function() if an IPI is sent by the same process we are | |
891 | * waiting to become inactive. | |
892 | */ | |
95cdf3b7 | 893 | void wait_task_inactive(task_t *p) |
1da177e4 LT |
894 | { |
895 | unsigned long flags; | |
896 | runqueue_t *rq; | |
897 | int preempted; | |
898 | ||
899 | repeat: | |
900 | rq = task_rq_lock(p, &flags); | |
901 | /* Must be off runqueue entirely, not preempted. */ | |
902 | if (unlikely(p->array || task_running(rq, p))) { | |
903 | /* If it's preempted, we yield. It could be a while. */ | |
904 | preempted = !task_running(rq, p); | |
905 | task_rq_unlock(rq, &flags); | |
906 | cpu_relax(); | |
907 | if (preempted) | |
908 | yield(); | |
909 | goto repeat; | |
910 | } | |
911 | task_rq_unlock(rq, &flags); | |
912 | } | |
913 | ||
914 | /*** | |
915 | * kick_process - kick a running thread to enter/exit the kernel | |
916 | * @p: the to-be-kicked thread | |
917 | * | |
918 | * Cause a process which is running on another CPU to enter | |
919 | * kernel-mode, without any delay. (to get signals handled.) | |
920 | * | |
921 | * NOTE: this function doesnt have to take the runqueue lock, | |
922 | * because all it wants to ensure is that the remote task enters | |
923 | * the kernel. If the IPI races and the task has been migrated | |
924 | * to another CPU then no harm is done and the purpose has been | |
925 | * achieved as well. | |
926 | */ | |
927 | void kick_process(task_t *p) | |
928 | { | |
929 | int cpu; | |
930 | ||
931 | preempt_disable(); | |
932 | cpu = task_cpu(p); | |
933 | if ((cpu != smp_processor_id()) && task_curr(p)) | |
934 | smp_send_reschedule(cpu); | |
935 | preempt_enable(); | |
936 | } | |
937 | ||
938 | /* | |
939 | * Return a low guess at the load of a migration-source cpu. | |
940 | * | |
941 | * We want to under-estimate the load of migration sources, to | |
942 | * balance conservatively. | |
943 | */ | |
a2000572 | 944 | static inline unsigned long source_load(int cpu, int type) |
1da177e4 LT |
945 | { |
946 | runqueue_t *rq = cpu_rq(cpu); | |
a2000572 | 947 | unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE; |
3b0bd9bc | 948 | if (type == 0) |
a2000572 | 949 | return load_now; |
b910472d | 950 | |
a2000572 | 951 | return min(rq->cpu_load[type-1], load_now); |
1da177e4 LT |
952 | } |
953 | ||
954 | /* | |
955 | * Return a high guess at the load of a migration-target cpu | |
956 | */ | |
a2000572 | 957 | static inline unsigned long target_load(int cpu, int type) |
1da177e4 LT |
958 | { |
959 | runqueue_t *rq = cpu_rq(cpu); | |
a2000572 | 960 | unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE; |
7897986b | 961 | if (type == 0) |
a2000572 | 962 | return load_now; |
3b0bd9bc | 963 | |
a2000572 | 964 | return max(rq->cpu_load[type-1], load_now); |
1da177e4 LT |
965 | } |
966 | ||
147cbb4b NP |
967 | /* |
968 | * find_idlest_group finds and returns the least busy CPU group within the | |
969 | * domain. | |
970 | */ | |
971 | static struct sched_group * | |
972 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu) | |
973 | { | |
974 | struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups; | |
975 | unsigned long min_load = ULONG_MAX, this_load = 0; | |
976 | int load_idx = sd->forkexec_idx; | |
977 | int imbalance = 100 + (sd->imbalance_pct-100)/2; | |
978 | ||
979 | do { | |
980 | unsigned long load, avg_load; | |
981 | int local_group; | |
982 | int i; | |
983 | ||
da5a5522 BD |
984 | /* Skip over this group if it has no CPUs allowed */ |
985 | if (!cpus_intersects(group->cpumask, p->cpus_allowed)) | |
986 | goto nextgroup; | |
987 | ||
147cbb4b | 988 | local_group = cpu_isset(this_cpu, group->cpumask); |
147cbb4b NP |
989 | |
990 | /* Tally up the load of all CPUs in the group */ | |
991 | avg_load = 0; | |
992 | ||
993 | for_each_cpu_mask(i, group->cpumask) { | |
994 | /* Bias balancing toward cpus of our domain */ | |
995 | if (local_group) | |
996 | load = source_load(i, load_idx); | |
997 | else | |
998 | load = target_load(i, load_idx); | |
999 | ||
1000 | avg_load += load; | |
1001 | } | |
1002 | ||
1003 | /* Adjust by relative CPU power of the group */ | |
1004 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
1005 | ||
1006 | if (local_group) { | |
1007 | this_load = avg_load; | |
1008 | this = group; | |
1009 | } else if (avg_load < min_load) { | |
1010 | min_load = avg_load; | |
1011 | idlest = group; | |
1012 | } | |
da5a5522 | 1013 | nextgroup: |
147cbb4b NP |
1014 | group = group->next; |
1015 | } while (group != sd->groups); | |
1016 | ||
1017 | if (!idlest || 100*this_load < imbalance*min_load) | |
1018 | return NULL; | |
1019 | return idlest; | |
1020 | } | |
1021 | ||
1022 | /* | |
1023 | * find_idlest_queue - find the idlest runqueue among the cpus in group. | |
1024 | */ | |
95cdf3b7 IM |
1025 | static int |
1026 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
147cbb4b | 1027 | { |
da5a5522 | 1028 | cpumask_t tmp; |
147cbb4b NP |
1029 | unsigned long load, min_load = ULONG_MAX; |
1030 | int idlest = -1; | |
1031 | int i; | |
1032 | ||
da5a5522 BD |
1033 | /* Traverse only the allowed CPUs */ |
1034 | cpus_and(tmp, group->cpumask, p->cpus_allowed); | |
1035 | ||
1036 | for_each_cpu_mask(i, tmp) { | |
147cbb4b NP |
1037 | load = source_load(i, 0); |
1038 | ||
1039 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
1040 | min_load = load; | |
1041 | idlest = i; | |
1042 | } | |
1043 | } | |
1044 | ||
1045 | return idlest; | |
1046 | } | |
1047 | ||
476d139c NP |
1048 | /* |
1049 | * sched_balance_self: balance the current task (running on cpu) in domains | |
1050 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
1051 | * SD_BALANCE_EXEC. | |
1052 | * | |
1053 | * Balance, ie. select the least loaded group. | |
1054 | * | |
1055 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
1056 | * | |
1057 | * preempt must be disabled. | |
1058 | */ | |
1059 | static int sched_balance_self(int cpu, int flag) | |
1060 | { | |
1061 | struct task_struct *t = current; | |
1062 | struct sched_domain *tmp, *sd = NULL; | |
147cbb4b | 1063 | |
476d139c NP |
1064 | for_each_domain(cpu, tmp) |
1065 | if (tmp->flags & flag) | |
1066 | sd = tmp; | |
1067 | ||
1068 | while (sd) { | |
1069 | cpumask_t span; | |
1070 | struct sched_group *group; | |
1071 | int new_cpu; | |
1072 | int weight; | |
1073 | ||
1074 | span = sd->span; | |
1075 | group = find_idlest_group(sd, t, cpu); | |
1076 | if (!group) | |
1077 | goto nextlevel; | |
1078 | ||
da5a5522 | 1079 | new_cpu = find_idlest_cpu(group, t, cpu); |
476d139c NP |
1080 | if (new_cpu == -1 || new_cpu == cpu) |
1081 | goto nextlevel; | |
1082 | ||
1083 | /* Now try balancing at a lower domain level */ | |
1084 | cpu = new_cpu; | |
1085 | nextlevel: | |
1086 | sd = NULL; | |
1087 | weight = cpus_weight(span); | |
1088 | for_each_domain(cpu, tmp) { | |
1089 | if (weight <= cpus_weight(tmp->span)) | |
1090 | break; | |
1091 | if (tmp->flags & flag) | |
1092 | sd = tmp; | |
1093 | } | |
1094 | /* while loop will break here if sd == NULL */ | |
1095 | } | |
1096 | ||
1097 | return cpu; | |
1098 | } | |
1099 | ||
1100 | #endif /* CONFIG_SMP */ | |
1da177e4 LT |
1101 | |
1102 | /* | |
1103 | * wake_idle() will wake a task on an idle cpu if task->cpu is | |
1104 | * not idle and an idle cpu is available. The span of cpus to | |
1105 | * search starts with cpus closest then further out as needed, | |
1106 | * so we always favor a closer, idle cpu. | |
1107 | * | |
1108 | * Returns the CPU we should wake onto. | |
1109 | */ | |
1110 | #if defined(ARCH_HAS_SCHED_WAKE_IDLE) | |
1111 | static int wake_idle(int cpu, task_t *p) | |
1112 | { | |
1113 | cpumask_t tmp; | |
1114 | struct sched_domain *sd; | |
1115 | int i; | |
1116 | ||
1117 | if (idle_cpu(cpu)) | |
1118 | return cpu; | |
1119 | ||
1120 | for_each_domain(cpu, sd) { | |
1121 | if (sd->flags & SD_WAKE_IDLE) { | |
e0f364f4 | 1122 | cpus_and(tmp, sd->span, p->cpus_allowed); |
1da177e4 LT |
1123 | for_each_cpu_mask(i, tmp) { |
1124 | if (idle_cpu(i)) | |
1125 | return i; | |
1126 | } | |
1127 | } | |
e0f364f4 NP |
1128 | else |
1129 | break; | |
1da177e4 LT |
1130 | } |
1131 | return cpu; | |
1132 | } | |
1133 | #else | |
1134 | static inline int wake_idle(int cpu, task_t *p) | |
1135 | { | |
1136 | return cpu; | |
1137 | } | |
1138 | #endif | |
1139 | ||
1140 | /*** | |
1141 | * try_to_wake_up - wake up a thread | |
1142 | * @p: the to-be-woken-up thread | |
1143 | * @state: the mask of task states that can be woken | |
1144 | * @sync: do a synchronous wakeup? | |
1145 | * | |
1146 | * Put it on the run-queue if it's not already there. The "current" | |
1147 | * thread is always on the run-queue (except when the actual | |
1148 | * re-schedule is in progress), and as such you're allowed to do | |
1149 | * the simpler "current->state = TASK_RUNNING" to mark yourself | |
1150 | * runnable without the overhead of this. | |
1151 | * | |
1152 | * returns failure only if the task is already active. | |
1153 | */ | |
95cdf3b7 | 1154 | static int try_to_wake_up(task_t *p, unsigned int state, int sync) |
1da177e4 LT |
1155 | { |
1156 | int cpu, this_cpu, success = 0; | |
1157 | unsigned long flags; | |
1158 | long old_state; | |
1159 | runqueue_t *rq; | |
1160 | #ifdef CONFIG_SMP | |
1161 | unsigned long load, this_load; | |
7897986b | 1162 | struct sched_domain *sd, *this_sd = NULL; |
1da177e4 LT |
1163 | int new_cpu; |
1164 | #endif | |
1165 | ||
1166 | rq = task_rq_lock(p, &flags); | |
1167 | old_state = p->state; | |
1168 | if (!(old_state & state)) | |
1169 | goto out; | |
1170 | ||
1171 | if (p->array) | |
1172 | goto out_running; | |
1173 | ||
1174 | cpu = task_cpu(p); | |
1175 | this_cpu = smp_processor_id(); | |
1176 | ||
1177 | #ifdef CONFIG_SMP | |
1178 | if (unlikely(task_running(rq, p))) | |
1179 | goto out_activate; | |
1180 | ||
7897986b NP |
1181 | new_cpu = cpu; |
1182 | ||
1da177e4 LT |
1183 | schedstat_inc(rq, ttwu_cnt); |
1184 | if (cpu == this_cpu) { | |
1185 | schedstat_inc(rq, ttwu_local); | |
7897986b NP |
1186 | goto out_set_cpu; |
1187 | } | |
1188 | ||
1189 | for_each_domain(this_cpu, sd) { | |
1190 | if (cpu_isset(cpu, sd->span)) { | |
1191 | schedstat_inc(sd, ttwu_wake_remote); | |
1192 | this_sd = sd; | |
1193 | break; | |
1da177e4 LT |
1194 | } |
1195 | } | |
1da177e4 | 1196 | |
7897986b | 1197 | if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed))) |
1da177e4 LT |
1198 | goto out_set_cpu; |
1199 | ||
1da177e4 | 1200 | /* |
7897986b | 1201 | * Check for affine wakeup and passive balancing possibilities. |
1da177e4 | 1202 | */ |
7897986b NP |
1203 | if (this_sd) { |
1204 | int idx = this_sd->wake_idx; | |
1205 | unsigned int imbalance; | |
1da177e4 | 1206 | |
a3f21bce NP |
1207 | imbalance = 100 + (this_sd->imbalance_pct - 100) / 2; |
1208 | ||
7897986b NP |
1209 | load = source_load(cpu, idx); |
1210 | this_load = target_load(this_cpu, idx); | |
1da177e4 | 1211 | |
7897986b NP |
1212 | new_cpu = this_cpu; /* Wake to this CPU if we can */ |
1213 | ||
a3f21bce NP |
1214 | if (this_sd->flags & SD_WAKE_AFFINE) { |
1215 | unsigned long tl = this_load; | |
1da177e4 | 1216 | /* |
a3f21bce NP |
1217 | * If sync wakeup then subtract the (maximum possible) |
1218 | * effect of the currently running task from the load | |
1219 | * of the current CPU: | |
1da177e4 | 1220 | */ |
a3f21bce NP |
1221 | if (sync) |
1222 | tl -= SCHED_LOAD_SCALE; | |
1223 | ||
1224 | if ((tl <= load && | |
1225 | tl + target_load(cpu, idx) <= SCHED_LOAD_SCALE) || | |
1226 | 100*(tl + SCHED_LOAD_SCALE) <= imbalance*load) { | |
1227 | /* | |
1228 | * This domain has SD_WAKE_AFFINE and | |
1229 | * p is cache cold in this domain, and | |
1230 | * there is no bad imbalance. | |
1231 | */ | |
1232 | schedstat_inc(this_sd, ttwu_move_affine); | |
1233 | goto out_set_cpu; | |
1234 | } | |
1235 | } | |
1236 | ||
1237 | /* | |
1238 | * Start passive balancing when half the imbalance_pct | |
1239 | * limit is reached. | |
1240 | */ | |
1241 | if (this_sd->flags & SD_WAKE_BALANCE) { | |
1242 | if (imbalance*this_load <= 100*load) { | |
1243 | schedstat_inc(this_sd, ttwu_move_balance); | |
1244 | goto out_set_cpu; | |
1245 | } | |
1da177e4 LT |
1246 | } |
1247 | } | |
1248 | ||
1249 | new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */ | |
1250 | out_set_cpu: | |
1251 | new_cpu = wake_idle(new_cpu, p); | |
1252 | if (new_cpu != cpu) { | |
1253 | set_task_cpu(p, new_cpu); | |
1254 | task_rq_unlock(rq, &flags); | |
1255 | /* might preempt at this point */ | |
1256 | rq = task_rq_lock(p, &flags); | |
1257 | old_state = p->state; | |
1258 | if (!(old_state & state)) | |
1259 | goto out; | |
1260 | if (p->array) | |
1261 | goto out_running; | |
1262 | ||
1263 | this_cpu = smp_processor_id(); | |
1264 | cpu = task_cpu(p); | |
1265 | } | |
1266 | ||
1267 | out_activate: | |
1268 | #endif /* CONFIG_SMP */ | |
1269 | if (old_state == TASK_UNINTERRUPTIBLE) { | |
1270 | rq->nr_uninterruptible--; | |
1271 | /* | |
1272 | * Tasks on involuntary sleep don't earn | |
1273 | * sleep_avg beyond just interactive state. | |
1274 | */ | |
3dee386e | 1275 | p->sleep_type = SLEEP_NONINTERACTIVE; |
1da177e4 LT |
1276 | } |
1277 | ||
d79fc0fc IM |
1278 | /* |
1279 | * Tasks that have marked their sleep as noninteractive get | |
1280 | * woken up without updating their sleep average. (i.e. their | |
1281 | * sleep is handled in a priority-neutral manner, no priority | |
1282 | * boost and no penalty.) | |
1283 | */ | |
1284 | if (old_state & TASK_NONINTERACTIVE) | |
1285 | __activate_task(p, rq); | |
1286 | else | |
1287 | activate_task(p, rq, cpu == this_cpu); | |
1da177e4 LT |
1288 | /* |
1289 | * Sync wakeups (i.e. those types of wakeups where the waker | |
1290 | * has indicated that it will leave the CPU in short order) | |
1291 | * don't trigger a preemption, if the woken up task will run on | |
1292 | * this cpu. (in this case the 'I will reschedule' promise of | |
1293 | * the waker guarantees that the freshly woken up task is going | |
1294 | * to be considered on this CPU.) | |
1295 | */ | |
1da177e4 LT |
1296 | if (!sync || cpu != this_cpu) { |
1297 | if (TASK_PREEMPTS_CURR(p, rq)) | |
1298 | resched_task(rq->curr); | |
1299 | } | |
1300 | success = 1; | |
1301 | ||
1302 | out_running: | |
1303 | p->state = TASK_RUNNING; | |
1304 | out: | |
1305 | task_rq_unlock(rq, &flags); | |
1306 | ||
1307 | return success; | |
1308 | } | |
1309 | ||
95cdf3b7 | 1310 | int fastcall wake_up_process(task_t *p) |
1da177e4 LT |
1311 | { |
1312 | return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED | | |
1313 | TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0); | |
1314 | } | |
1315 | ||
1316 | EXPORT_SYMBOL(wake_up_process); | |
1317 | ||
1318 | int fastcall wake_up_state(task_t *p, unsigned int state) | |
1319 | { | |
1320 | return try_to_wake_up(p, state, 0); | |
1321 | } | |
1322 | ||
1da177e4 LT |
1323 | /* |
1324 | * Perform scheduler related setup for a newly forked process p. | |
1325 | * p is forked by current. | |
1326 | */ | |
476d139c | 1327 | void fastcall sched_fork(task_t *p, int clone_flags) |
1da177e4 | 1328 | { |
476d139c NP |
1329 | int cpu = get_cpu(); |
1330 | ||
1331 | #ifdef CONFIG_SMP | |
1332 | cpu = sched_balance_self(cpu, SD_BALANCE_FORK); | |
1333 | #endif | |
1334 | set_task_cpu(p, cpu); | |
1335 | ||
1da177e4 LT |
1336 | /* |
1337 | * We mark the process as running here, but have not actually | |
1338 | * inserted it onto the runqueue yet. This guarantees that | |
1339 | * nobody will actually run it, and a signal or other external | |
1340 | * event cannot wake it up and insert it on the runqueue either. | |
1341 | */ | |
1342 | p->state = TASK_RUNNING; | |
1343 | INIT_LIST_HEAD(&p->run_list); | |
1344 | p->array = NULL; | |
1da177e4 LT |
1345 | #ifdef CONFIG_SCHEDSTATS |
1346 | memset(&p->sched_info, 0, sizeof(p->sched_info)); | |
1347 | #endif | |
d6077cb8 | 1348 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
4866cde0 NP |
1349 | p->oncpu = 0; |
1350 | #endif | |
1da177e4 | 1351 | #ifdef CONFIG_PREEMPT |
4866cde0 | 1352 | /* Want to start with kernel preemption disabled. */ |
a1261f54 | 1353 | task_thread_info(p)->preempt_count = 1; |
1da177e4 LT |
1354 | #endif |
1355 | /* | |
1356 | * Share the timeslice between parent and child, thus the | |
1357 | * total amount of pending timeslices in the system doesn't change, | |
1358 | * resulting in more scheduling fairness. | |
1359 | */ | |
1360 | local_irq_disable(); | |
1361 | p->time_slice = (current->time_slice + 1) >> 1; | |
1362 | /* | |
1363 | * The remainder of the first timeslice might be recovered by | |
1364 | * the parent if the child exits early enough. | |
1365 | */ | |
1366 | p->first_time_slice = 1; | |
1367 | current->time_slice >>= 1; | |
1368 | p->timestamp = sched_clock(); | |
1369 | if (unlikely(!current->time_slice)) { | |
1370 | /* | |
1371 | * This case is rare, it happens when the parent has only | |
1372 | * a single jiffy left from its timeslice. Taking the | |
1373 | * runqueue lock is not a problem. | |
1374 | */ | |
1375 | current->time_slice = 1; | |
1da177e4 | 1376 | scheduler_tick(); |
476d139c NP |
1377 | } |
1378 | local_irq_enable(); | |
1379 | put_cpu(); | |
1da177e4 LT |
1380 | } |
1381 | ||
1382 | /* | |
1383 | * wake_up_new_task - wake up a newly created task for the first time. | |
1384 | * | |
1385 | * This function will do some initial scheduler statistics housekeeping | |
1386 | * that must be done for every newly created context, then puts the task | |
1387 | * on the runqueue and wakes it. | |
1388 | */ | |
95cdf3b7 | 1389 | void fastcall wake_up_new_task(task_t *p, unsigned long clone_flags) |
1da177e4 LT |
1390 | { |
1391 | unsigned long flags; | |
1392 | int this_cpu, cpu; | |
1393 | runqueue_t *rq, *this_rq; | |
1394 | ||
1395 | rq = task_rq_lock(p, &flags); | |
147cbb4b | 1396 | BUG_ON(p->state != TASK_RUNNING); |
1da177e4 | 1397 | this_cpu = smp_processor_id(); |
147cbb4b | 1398 | cpu = task_cpu(p); |
1da177e4 | 1399 | |
1da177e4 LT |
1400 | /* |
1401 | * We decrease the sleep average of forking parents | |
1402 | * and children as well, to keep max-interactive tasks | |
1403 | * from forking tasks that are max-interactive. The parent | |
1404 | * (current) is done further down, under its lock. | |
1405 | */ | |
1406 | p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) * | |
1407 | CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS); | |
1408 | ||
1409 | p->prio = effective_prio(p); | |
1410 | ||
1411 | if (likely(cpu == this_cpu)) { | |
1412 | if (!(clone_flags & CLONE_VM)) { | |
1413 | /* | |
1414 | * The VM isn't cloned, so we're in a good position to | |
1415 | * do child-runs-first in anticipation of an exec. This | |
1416 | * usually avoids a lot of COW overhead. | |
1417 | */ | |
1418 | if (unlikely(!current->array)) | |
1419 | __activate_task(p, rq); | |
1420 | else { | |
1421 | p->prio = current->prio; | |
1422 | list_add_tail(&p->run_list, ¤t->run_list); | |
1423 | p->array = current->array; | |
1424 | p->array->nr_active++; | |
a2000572 | 1425 | rq->nr_running++; |
1da177e4 LT |
1426 | } |
1427 | set_need_resched(); | |
1428 | } else | |
1429 | /* Run child last */ | |
1430 | __activate_task(p, rq); | |
1431 | /* | |
1432 | * We skip the following code due to cpu == this_cpu | |
1433 | * | |
1434 | * task_rq_unlock(rq, &flags); | |
1435 | * this_rq = task_rq_lock(current, &flags); | |
1436 | */ | |
1437 | this_rq = rq; | |
1438 | } else { | |
1439 | this_rq = cpu_rq(this_cpu); | |
1440 | ||
1441 | /* | |
1442 | * Not the local CPU - must adjust timestamp. This should | |
1443 | * get optimised away in the !CONFIG_SMP case. | |
1444 | */ | |
1445 | p->timestamp = (p->timestamp - this_rq->timestamp_last_tick) | |
1446 | + rq->timestamp_last_tick; | |
1447 | __activate_task(p, rq); | |
1448 | if (TASK_PREEMPTS_CURR(p, rq)) | |
1449 | resched_task(rq->curr); | |
1450 | ||
1451 | /* | |
1452 | * Parent and child are on different CPUs, now get the | |
1453 | * parent runqueue to update the parent's ->sleep_avg: | |
1454 | */ | |
1455 | task_rq_unlock(rq, &flags); | |
1456 | this_rq = task_rq_lock(current, &flags); | |
1457 | } | |
1458 | current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) * | |
1459 | PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS); | |
1460 | task_rq_unlock(this_rq, &flags); | |
1461 | } | |
1462 | ||
1463 | /* | |
1464 | * Potentially available exiting-child timeslices are | |
1465 | * retrieved here - this way the parent does not get | |
1466 | * penalized for creating too many threads. | |
1467 | * | |
1468 | * (this cannot be used to 'generate' timeslices | |
1469 | * artificially, because any timeslice recovered here | |
1470 | * was given away by the parent in the first place.) | |
1471 | */ | |
95cdf3b7 | 1472 | void fastcall sched_exit(task_t *p) |
1da177e4 LT |
1473 | { |
1474 | unsigned long flags; | |
1475 | runqueue_t *rq; | |
1476 | ||
1477 | /* | |
1478 | * If the child was a (relative-) CPU hog then decrease | |
1479 | * the sleep_avg of the parent as well. | |
1480 | */ | |
1481 | rq = task_rq_lock(p->parent, &flags); | |
889dfafe | 1482 | if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) { |
1da177e4 LT |
1483 | p->parent->time_slice += p->time_slice; |
1484 | if (unlikely(p->parent->time_slice > task_timeslice(p))) | |
1485 | p->parent->time_slice = task_timeslice(p); | |
1486 | } | |
1487 | if (p->sleep_avg < p->parent->sleep_avg) | |
1488 | p->parent->sleep_avg = p->parent->sleep_avg / | |
1489 | (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg / | |
1490 | (EXIT_WEIGHT + 1); | |
1491 | task_rq_unlock(rq, &flags); | |
1492 | } | |
1493 | ||
4866cde0 NP |
1494 | /** |
1495 | * prepare_task_switch - prepare to switch tasks | |
1496 | * @rq: the runqueue preparing to switch | |
1497 | * @next: the task we are going to switch to. | |
1498 | * | |
1499 | * This is called with the rq lock held and interrupts off. It must | |
1500 | * be paired with a subsequent finish_task_switch after the context | |
1501 | * switch. | |
1502 | * | |
1503 | * prepare_task_switch sets up locking and calls architecture specific | |
1504 | * hooks. | |
1505 | */ | |
1506 | static inline void prepare_task_switch(runqueue_t *rq, task_t *next) | |
1507 | { | |
1508 | prepare_lock_switch(rq, next); | |
1509 | prepare_arch_switch(next); | |
1510 | } | |
1511 | ||
1da177e4 LT |
1512 | /** |
1513 | * finish_task_switch - clean up after a task-switch | |
344babaa | 1514 | * @rq: runqueue associated with task-switch |
1da177e4 LT |
1515 | * @prev: the thread we just switched away from. |
1516 | * | |
4866cde0 NP |
1517 | * finish_task_switch must be called after the context switch, paired |
1518 | * with a prepare_task_switch call before the context switch. | |
1519 | * finish_task_switch will reconcile locking set up by prepare_task_switch, | |
1520 | * and do any other architecture-specific cleanup actions. | |
1da177e4 LT |
1521 | * |
1522 | * Note that we may have delayed dropping an mm in context_switch(). If | |
1523 | * so, we finish that here outside of the runqueue lock. (Doing it | |
1524 | * with the lock held can cause deadlocks; see schedule() for | |
1525 | * details.) | |
1526 | */ | |
4866cde0 | 1527 | static inline void finish_task_switch(runqueue_t *rq, task_t *prev) |
1da177e4 LT |
1528 | __releases(rq->lock) |
1529 | { | |
1da177e4 LT |
1530 | struct mm_struct *mm = rq->prev_mm; |
1531 | unsigned long prev_task_flags; | |
1532 | ||
1533 | rq->prev_mm = NULL; | |
1534 | ||
1535 | /* | |
1536 | * A task struct has one reference for the use as "current". | |
1537 | * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and | |
1538 | * calls schedule one last time. The schedule call will never return, | |
1539 | * and the scheduled task must drop that reference. | |
1540 | * The test for EXIT_ZOMBIE must occur while the runqueue locks are | |
1541 | * still held, otherwise prev could be scheduled on another cpu, die | |
1542 | * there before we look at prev->state, and then the reference would | |
1543 | * be dropped twice. | |
1544 | * Manfred Spraul <manfred@colorfullife.com> | |
1545 | */ | |
1546 | prev_task_flags = prev->flags; | |
4866cde0 NP |
1547 | finish_arch_switch(prev); |
1548 | finish_lock_switch(rq, prev); | |
1da177e4 LT |
1549 | if (mm) |
1550 | mmdrop(mm); | |
c6fd91f0 | 1551 | if (unlikely(prev_task_flags & PF_DEAD)) { |
1552 | /* | |
1553 | * Remove function-return probe instances associated with this | |
1554 | * task and put them back on the free list. | |
1555 | */ | |
1556 | kprobe_flush_task(prev); | |
1da177e4 | 1557 | put_task_struct(prev); |
c6fd91f0 | 1558 | } |
1da177e4 LT |
1559 | } |
1560 | ||
1561 | /** | |
1562 | * schedule_tail - first thing a freshly forked thread must call. | |
1563 | * @prev: the thread we just switched away from. | |
1564 | */ | |
1565 | asmlinkage void schedule_tail(task_t *prev) | |
1566 | __releases(rq->lock) | |
1567 | { | |
4866cde0 NP |
1568 | runqueue_t *rq = this_rq(); |
1569 | finish_task_switch(rq, prev); | |
1570 | #ifdef __ARCH_WANT_UNLOCKED_CTXSW | |
1571 | /* In this case, finish_task_switch does not reenable preemption */ | |
1572 | preempt_enable(); | |
1573 | #endif | |
1da177e4 LT |
1574 | if (current->set_child_tid) |
1575 | put_user(current->pid, current->set_child_tid); | |
1576 | } | |
1577 | ||
1578 | /* | |
1579 | * context_switch - switch to the new MM and the new | |
1580 | * thread's register state. | |
1581 | */ | |
1582 | static inline | |
1583 | task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next) | |
1584 | { | |
1585 | struct mm_struct *mm = next->mm; | |
1586 | struct mm_struct *oldmm = prev->active_mm; | |
1587 | ||
1588 | if (unlikely(!mm)) { | |
1589 | next->active_mm = oldmm; | |
1590 | atomic_inc(&oldmm->mm_count); | |
1591 | enter_lazy_tlb(oldmm, next); | |
1592 | } else | |
1593 | switch_mm(oldmm, mm, next); | |
1594 | ||
1595 | if (unlikely(!prev->mm)) { | |
1596 | prev->active_mm = NULL; | |
1597 | WARN_ON(rq->prev_mm); | |
1598 | rq->prev_mm = oldmm; | |
1599 | } | |
1600 | ||
1601 | /* Here we just switch the register state and the stack. */ | |
1602 | switch_to(prev, next, prev); | |
1603 | ||
1604 | return prev; | |
1605 | } | |
1606 | ||
1607 | /* | |
1608 | * nr_running, nr_uninterruptible and nr_context_switches: | |
1609 | * | |
1610 | * externally visible scheduler statistics: current number of runnable | |
1611 | * threads, current number of uninterruptible-sleeping threads, total | |
1612 | * number of context switches performed since bootup. | |
1613 | */ | |
1614 | unsigned long nr_running(void) | |
1615 | { | |
1616 | unsigned long i, sum = 0; | |
1617 | ||
1618 | for_each_online_cpu(i) | |
1619 | sum += cpu_rq(i)->nr_running; | |
1620 | ||
1621 | return sum; | |
1622 | } | |
1623 | ||
1624 | unsigned long nr_uninterruptible(void) | |
1625 | { | |
1626 | unsigned long i, sum = 0; | |
1627 | ||
0a945022 | 1628 | for_each_possible_cpu(i) |
1da177e4 LT |
1629 | sum += cpu_rq(i)->nr_uninterruptible; |
1630 | ||
1631 | /* | |
1632 | * Since we read the counters lockless, it might be slightly | |
1633 | * inaccurate. Do not allow it to go below zero though: | |
1634 | */ | |
1635 | if (unlikely((long)sum < 0)) | |
1636 | sum = 0; | |
1637 | ||
1638 | return sum; | |
1639 | } | |
1640 | ||
1641 | unsigned long long nr_context_switches(void) | |
1642 | { | |
1643 | unsigned long long i, sum = 0; | |
1644 | ||
0a945022 | 1645 | for_each_possible_cpu(i) |
1da177e4 LT |
1646 | sum += cpu_rq(i)->nr_switches; |
1647 | ||
1648 | return sum; | |
1649 | } | |
1650 | ||
1651 | unsigned long nr_iowait(void) | |
1652 | { | |
1653 | unsigned long i, sum = 0; | |
1654 | ||
0a945022 | 1655 | for_each_possible_cpu(i) |
1da177e4 LT |
1656 | sum += atomic_read(&cpu_rq(i)->nr_iowait); |
1657 | ||
1658 | return sum; | |
1659 | } | |
1660 | ||
db1b1fef JS |
1661 | unsigned long nr_active(void) |
1662 | { | |
1663 | unsigned long i, running = 0, uninterruptible = 0; | |
1664 | ||
1665 | for_each_online_cpu(i) { | |
1666 | running += cpu_rq(i)->nr_running; | |
1667 | uninterruptible += cpu_rq(i)->nr_uninterruptible; | |
1668 | } | |
1669 | ||
1670 | if (unlikely((long)uninterruptible < 0)) | |
1671 | uninterruptible = 0; | |
1672 | ||
1673 | return running + uninterruptible; | |
1674 | } | |
1675 | ||
1da177e4 LT |
1676 | #ifdef CONFIG_SMP |
1677 | ||
1678 | /* | |
1679 | * double_rq_lock - safely lock two runqueues | |
1680 | * | |
e9028b0f AB |
1681 | * We must take them in cpu order to match code in |
1682 | * dependent_sleeper and wake_dependent_sleeper. | |
1683 | * | |
1da177e4 LT |
1684 | * Note this does not disable interrupts like task_rq_lock, |
1685 | * you need to do so manually before calling. | |
1686 | */ | |
1687 | static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2) | |
1688 | __acquires(rq1->lock) | |
1689 | __acquires(rq2->lock) | |
1690 | { | |
1691 | if (rq1 == rq2) { | |
1692 | spin_lock(&rq1->lock); | |
1693 | __acquire(rq2->lock); /* Fake it out ;) */ | |
1694 | } else { | |
e9028b0f | 1695 | if (rq1->cpu < rq2->cpu) { |
1da177e4 LT |
1696 | spin_lock(&rq1->lock); |
1697 | spin_lock(&rq2->lock); | |
1698 | } else { | |
1699 | spin_lock(&rq2->lock); | |
1700 | spin_lock(&rq1->lock); | |
1701 | } | |
1702 | } | |
1703 | } | |
1704 | ||
1705 | /* | |
1706 | * double_rq_unlock - safely unlock two runqueues | |
1707 | * | |
1708 | * Note this does not restore interrupts like task_rq_unlock, | |
1709 | * you need to do so manually after calling. | |
1710 | */ | |
1711 | static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2) | |
1712 | __releases(rq1->lock) | |
1713 | __releases(rq2->lock) | |
1714 | { | |
1715 | spin_unlock(&rq1->lock); | |
1716 | if (rq1 != rq2) | |
1717 | spin_unlock(&rq2->lock); | |
1718 | else | |
1719 | __release(rq2->lock); | |
1720 | } | |
1721 | ||
1722 | /* | |
1723 | * double_lock_balance - lock the busiest runqueue, this_rq is locked already. | |
1724 | */ | |
1725 | static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest) | |
1726 | __releases(this_rq->lock) | |
1727 | __acquires(busiest->lock) | |
1728 | __acquires(this_rq->lock) | |
1729 | { | |
1730 | if (unlikely(!spin_trylock(&busiest->lock))) { | |
e9028b0f | 1731 | if (busiest->cpu < this_rq->cpu) { |
1da177e4 LT |
1732 | spin_unlock(&this_rq->lock); |
1733 | spin_lock(&busiest->lock); | |
1734 | spin_lock(&this_rq->lock); | |
1735 | } else | |
1736 | spin_lock(&busiest->lock); | |
1737 | } | |
1738 | } | |
1739 | ||
1da177e4 LT |
1740 | /* |
1741 | * If dest_cpu is allowed for this process, migrate the task to it. | |
1742 | * This is accomplished by forcing the cpu_allowed mask to only | |
1743 | * allow dest_cpu, which will force the cpu onto dest_cpu. Then | |
1744 | * the cpu_allowed mask is restored. | |
1745 | */ | |
1746 | static void sched_migrate_task(task_t *p, int dest_cpu) | |
1747 | { | |
1748 | migration_req_t req; | |
1749 | runqueue_t *rq; | |
1750 | unsigned long flags; | |
1751 | ||
1752 | rq = task_rq_lock(p, &flags); | |
1753 | if (!cpu_isset(dest_cpu, p->cpus_allowed) | |
1754 | || unlikely(cpu_is_offline(dest_cpu))) | |
1755 | goto out; | |
1756 | ||
1757 | /* force the process onto the specified CPU */ | |
1758 | if (migrate_task(p, dest_cpu, &req)) { | |
1759 | /* Need to wait for migration thread (might exit: take ref). */ | |
1760 | struct task_struct *mt = rq->migration_thread; | |
1761 | get_task_struct(mt); | |
1762 | task_rq_unlock(rq, &flags); | |
1763 | wake_up_process(mt); | |
1764 | put_task_struct(mt); | |
1765 | wait_for_completion(&req.done); | |
1766 | return; | |
1767 | } | |
1768 | out: | |
1769 | task_rq_unlock(rq, &flags); | |
1770 | } | |
1771 | ||
1772 | /* | |
476d139c NP |
1773 | * sched_exec - execve() is a valuable balancing opportunity, because at |
1774 | * this point the task has the smallest effective memory and cache footprint. | |
1da177e4 LT |
1775 | */ |
1776 | void sched_exec(void) | |
1777 | { | |
1da177e4 | 1778 | int new_cpu, this_cpu = get_cpu(); |
476d139c | 1779 | new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC); |
1da177e4 | 1780 | put_cpu(); |
476d139c NP |
1781 | if (new_cpu != this_cpu) |
1782 | sched_migrate_task(current, new_cpu); | |
1da177e4 LT |
1783 | } |
1784 | ||
1785 | /* | |
1786 | * pull_task - move a task from a remote runqueue to the local runqueue. | |
1787 | * Both runqueues must be locked. | |
1788 | */ | |
858119e1 | 1789 | static |
1da177e4 LT |
1790 | void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p, |
1791 | runqueue_t *this_rq, prio_array_t *this_array, int this_cpu) | |
1792 | { | |
1793 | dequeue_task(p, src_array); | |
a2000572 | 1794 | src_rq->nr_running--; |
1da177e4 | 1795 | set_task_cpu(p, this_cpu); |
a2000572 | 1796 | this_rq->nr_running++; |
1da177e4 LT |
1797 | enqueue_task(p, this_array); |
1798 | p->timestamp = (p->timestamp - src_rq->timestamp_last_tick) | |
1799 | + this_rq->timestamp_last_tick; | |
1800 | /* | |
1801 | * Note that idle threads have a prio of MAX_PRIO, for this test | |
1802 | * to be always true for them. | |
1803 | */ | |
1804 | if (TASK_PREEMPTS_CURR(p, this_rq)) | |
1805 | resched_task(this_rq->curr); | |
1806 | } | |
1807 | ||
1808 | /* | |
1809 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
1810 | */ | |
858119e1 | 1811 | static |
1da177e4 | 1812 | int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu, |
95cdf3b7 IM |
1813 | struct sched_domain *sd, enum idle_type idle, |
1814 | int *all_pinned) | |
1da177e4 LT |
1815 | { |
1816 | /* | |
1817 | * We do not migrate tasks that are: | |
1818 | * 1) running (obviously), or | |
1819 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
1820 | * 3) are cache-hot on their current CPU. | |
1821 | */ | |
1da177e4 LT |
1822 | if (!cpu_isset(this_cpu, p->cpus_allowed)) |
1823 | return 0; | |
81026794 NP |
1824 | *all_pinned = 0; |
1825 | ||
1826 | if (task_running(rq, p)) | |
1827 | return 0; | |
1da177e4 LT |
1828 | |
1829 | /* | |
1830 | * Aggressive migration if: | |
cafb20c1 | 1831 | * 1) task is cache cold, or |
1da177e4 LT |
1832 | * 2) too many balance attempts have failed. |
1833 | */ | |
1834 | ||
cafb20c1 | 1835 | if (sd->nr_balance_failed > sd->cache_nice_tries) |
1da177e4 LT |
1836 | return 1; |
1837 | ||
1838 | if (task_hot(p, rq->timestamp_last_tick, sd)) | |
81026794 | 1839 | return 0; |
1da177e4 LT |
1840 | return 1; |
1841 | } | |
1842 | ||
1843 | /* | |
1844 | * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq, | |
1845 | * as part of a balancing operation within "domain". Returns the number of | |
1846 | * tasks moved. | |
1847 | * | |
1848 | * Called with both runqueues locked. | |
1849 | */ | |
1850 | static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest, | |
1851 | unsigned long max_nr_move, struct sched_domain *sd, | |
81026794 | 1852 | enum idle_type idle, int *all_pinned) |
1da177e4 LT |
1853 | { |
1854 | prio_array_t *array, *dst_array; | |
1855 | struct list_head *head, *curr; | |
81026794 | 1856 | int idx, pulled = 0, pinned = 0; |
1da177e4 LT |
1857 | task_t *tmp; |
1858 | ||
81026794 | 1859 | if (max_nr_move == 0) |
1da177e4 LT |
1860 | goto out; |
1861 | ||
81026794 NP |
1862 | pinned = 1; |
1863 | ||
1da177e4 LT |
1864 | /* |
1865 | * We first consider expired tasks. Those will likely not be | |
1866 | * executed in the near future, and they are most likely to | |
1867 | * be cache-cold, thus switching CPUs has the least effect | |
1868 | * on them. | |
1869 | */ | |
1870 | if (busiest->expired->nr_active) { | |
1871 | array = busiest->expired; | |
1872 | dst_array = this_rq->expired; | |
1873 | } else { | |
1874 | array = busiest->active; | |
1875 | dst_array = this_rq->active; | |
1876 | } | |
1877 | ||
1878 | new_array: | |
1879 | /* Start searching at priority 0: */ | |
1880 | idx = 0; | |
1881 | skip_bitmap: | |
1882 | if (!idx) | |
1883 | idx = sched_find_first_bit(array->bitmap); | |
1884 | else | |
1885 | idx = find_next_bit(array->bitmap, MAX_PRIO, idx); | |
1886 | if (idx >= MAX_PRIO) { | |
1887 | if (array == busiest->expired && busiest->active->nr_active) { | |
1888 | array = busiest->active; | |
1889 | dst_array = this_rq->active; | |
1890 | goto new_array; | |
1891 | } | |
1892 | goto out; | |
1893 | } | |
1894 | ||
1895 | head = array->queue + idx; | |
1896 | curr = head->prev; | |
1897 | skip_queue: | |
1898 | tmp = list_entry(curr, task_t, run_list); | |
1899 | ||
1900 | curr = curr->prev; | |
1901 | ||
81026794 | 1902 | if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) { |
1da177e4 LT |
1903 | if (curr != head) |
1904 | goto skip_queue; | |
1905 | idx++; | |
1906 | goto skip_bitmap; | |
1907 | } | |
1908 | ||
1909 | #ifdef CONFIG_SCHEDSTATS | |
1910 | if (task_hot(tmp, busiest->timestamp_last_tick, sd)) | |
1911 | schedstat_inc(sd, lb_hot_gained[idle]); | |
1912 | #endif | |
1913 | ||
1914 | pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu); | |
1915 | pulled++; | |
1916 | ||
1917 | /* We only want to steal up to the prescribed number of tasks. */ | |
1918 | if (pulled < max_nr_move) { | |
1919 | if (curr != head) | |
1920 | goto skip_queue; | |
1921 | idx++; | |
1922 | goto skip_bitmap; | |
1923 | } | |
1924 | out: | |
1925 | /* | |
1926 | * Right now, this is the only place pull_task() is called, | |
1927 | * so we can safely collect pull_task() stats here rather than | |
1928 | * inside pull_task(). | |
1929 | */ | |
1930 | schedstat_add(sd, lb_gained[idle], pulled); | |
81026794 NP |
1931 | |
1932 | if (all_pinned) | |
1933 | *all_pinned = pinned; | |
1da177e4 LT |
1934 | return pulled; |
1935 | } | |
1936 | ||
1937 | /* | |
1938 | * find_busiest_group finds and returns the busiest CPU group within the | |
1939 | * domain. It calculates and returns the number of tasks which should be | |
1940 | * moved to restore balance via the imbalance parameter. | |
1941 | */ | |
1942 | static struct sched_group * | |
1943 | find_busiest_group(struct sched_domain *sd, int this_cpu, | |
5969fe06 | 1944 | unsigned long *imbalance, enum idle_type idle, int *sd_idle) |
1da177e4 LT |
1945 | { |
1946 | struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups; | |
1947 | unsigned long max_load, avg_load, total_load, this_load, total_pwr; | |
0c117f1b | 1948 | unsigned long max_pull; |
7897986b | 1949 | int load_idx; |
1da177e4 LT |
1950 | |
1951 | max_load = this_load = total_load = total_pwr = 0; | |
7897986b NP |
1952 | if (idle == NOT_IDLE) |
1953 | load_idx = sd->busy_idx; | |
1954 | else if (idle == NEWLY_IDLE) | |
1955 | load_idx = sd->newidle_idx; | |
1956 | else | |
1957 | load_idx = sd->idle_idx; | |
1da177e4 LT |
1958 | |
1959 | do { | |
1960 | unsigned long load; | |
1961 | int local_group; | |
1962 | int i; | |
1963 | ||
1964 | local_group = cpu_isset(this_cpu, group->cpumask); | |
1965 | ||
1966 | /* Tally up the load of all CPUs in the group */ | |
1967 | avg_load = 0; | |
1968 | ||
1969 | for_each_cpu_mask(i, group->cpumask) { | |
5969fe06 NP |
1970 | if (*sd_idle && !idle_cpu(i)) |
1971 | *sd_idle = 0; | |
1972 | ||
1da177e4 LT |
1973 | /* Bias balancing toward cpus of our domain */ |
1974 | if (local_group) | |
a2000572 | 1975 | load = target_load(i, load_idx); |
1da177e4 | 1976 | else |
a2000572 | 1977 | load = source_load(i, load_idx); |
1da177e4 LT |
1978 | |
1979 | avg_load += load; | |
1980 | } | |
1981 | ||
1982 | total_load += avg_load; | |
1983 | total_pwr += group->cpu_power; | |
1984 | ||
1985 | /* Adjust by relative CPU power of the group */ | |
1986 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
1987 | ||
1988 | if (local_group) { | |
1989 | this_load = avg_load; | |
1990 | this = group; | |
1da177e4 LT |
1991 | } else if (avg_load > max_load) { |
1992 | max_load = avg_load; | |
1993 | busiest = group; | |
1994 | } | |
1da177e4 LT |
1995 | group = group->next; |
1996 | } while (group != sd->groups); | |
1997 | ||
0c117f1b | 1998 | if (!busiest || this_load >= max_load || max_load <= SCHED_LOAD_SCALE) |
1da177e4 LT |
1999 | goto out_balanced; |
2000 | ||
2001 | avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr; | |
2002 | ||
2003 | if (this_load >= avg_load || | |
2004 | 100*max_load <= sd->imbalance_pct*this_load) | |
2005 | goto out_balanced; | |
2006 | ||
2007 | /* | |
2008 | * We're trying to get all the cpus to the average_load, so we don't | |
2009 | * want to push ourselves above the average load, nor do we wish to | |
2010 | * reduce the max loaded cpu below the average load, as either of these | |
2011 | * actions would just result in more rebalancing later, and ping-pong | |
2012 | * tasks around. Thus we look for the minimum possible imbalance. | |
2013 | * Negative imbalances (*we* are more loaded than anyone else) will | |
2014 | * be counted as no imbalance for these purposes -- we can't fix that | |
2015 | * by pulling tasks to us. Be careful of negative numbers as they'll | |
2016 | * appear as very large values with unsigned longs. | |
2017 | */ | |
0c117f1b SS |
2018 | |
2019 | /* Don't want to pull so many tasks that a group would go idle */ | |
2020 | max_pull = min(max_load - avg_load, max_load - SCHED_LOAD_SCALE); | |
2021 | ||
1da177e4 | 2022 | /* How much load to actually move to equalise the imbalance */ |
0c117f1b | 2023 | *imbalance = min(max_pull * busiest->cpu_power, |
1da177e4 LT |
2024 | (avg_load - this_load) * this->cpu_power) |
2025 | / SCHED_LOAD_SCALE; | |
2026 | ||
2027 | if (*imbalance < SCHED_LOAD_SCALE) { | |
2028 | unsigned long pwr_now = 0, pwr_move = 0; | |
2029 | unsigned long tmp; | |
2030 | ||
2031 | if (max_load - this_load >= SCHED_LOAD_SCALE*2) { | |
2032 | *imbalance = 1; | |
2033 | return busiest; | |
2034 | } | |
2035 | ||
2036 | /* | |
2037 | * OK, we don't have enough imbalance to justify moving tasks, | |
2038 | * however we may be able to increase total CPU power used by | |
2039 | * moving them. | |
2040 | */ | |
2041 | ||
2042 | pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load); | |
2043 | pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load); | |
2044 | pwr_now /= SCHED_LOAD_SCALE; | |
2045 | ||
2046 | /* Amount of load we'd subtract */ | |
2047 | tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power; | |
2048 | if (max_load > tmp) | |
2049 | pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE, | |
2050 | max_load - tmp); | |
2051 | ||
2052 | /* Amount of load we'd add */ | |
2053 | if (max_load*busiest->cpu_power < | |
2054 | SCHED_LOAD_SCALE*SCHED_LOAD_SCALE) | |
2055 | tmp = max_load*busiest->cpu_power/this->cpu_power; | |
2056 | else | |
2057 | tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power; | |
2058 | pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp); | |
2059 | pwr_move /= SCHED_LOAD_SCALE; | |
2060 | ||
2061 | /* Move if we gain throughput */ | |
2062 | if (pwr_move <= pwr_now) | |
2063 | goto out_balanced; | |
2064 | ||
2065 | *imbalance = 1; | |
2066 | return busiest; | |
2067 | } | |
2068 | ||
2069 | /* Get rid of the scaling factor, rounding down as we divide */ | |
2070 | *imbalance = *imbalance / SCHED_LOAD_SCALE; | |
1da177e4 LT |
2071 | return busiest; |
2072 | ||
2073 | out_balanced: | |
1da177e4 LT |
2074 | |
2075 | *imbalance = 0; | |
2076 | return NULL; | |
2077 | } | |
2078 | ||
2079 | /* | |
2080 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
2081 | */ | |
b910472d CK |
2082 | static runqueue_t *find_busiest_queue(struct sched_group *group, |
2083 | enum idle_type idle) | |
1da177e4 LT |
2084 | { |
2085 | unsigned long load, max_load = 0; | |
2086 | runqueue_t *busiest = NULL; | |
2087 | int i; | |
2088 | ||
2089 | for_each_cpu_mask(i, group->cpumask) { | |
a2000572 | 2090 | load = source_load(i, 0); |
1da177e4 LT |
2091 | |
2092 | if (load > max_load) { | |
2093 | max_load = load; | |
2094 | busiest = cpu_rq(i); | |
2095 | } | |
2096 | } | |
2097 | ||
2098 | return busiest; | |
2099 | } | |
2100 | ||
77391d71 NP |
2101 | /* |
2102 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
2103 | * so long as it is large enough. | |
2104 | */ | |
2105 | #define MAX_PINNED_INTERVAL 512 | |
2106 | ||
1da177e4 LT |
2107 | /* |
2108 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
2109 | * tasks if there is an imbalance. | |
2110 | * | |
2111 | * Called with this_rq unlocked. | |
2112 | */ | |
2113 | static int load_balance(int this_cpu, runqueue_t *this_rq, | |
2114 | struct sched_domain *sd, enum idle_type idle) | |
2115 | { | |
2116 | struct sched_group *group; | |
2117 | runqueue_t *busiest; | |
2118 | unsigned long imbalance; | |
77391d71 | 2119 | int nr_moved, all_pinned = 0; |
81026794 | 2120 | int active_balance = 0; |
5969fe06 NP |
2121 | int sd_idle = 0; |
2122 | ||
2123 | if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER) | |
2124 | sd_idle = 1; | |
1da177e4 | 2125 | |
1da177e4 LT |
2126 | schedstat_inc(sd, lb_cnt[idle]); |
2127 | ||
5969fe06 | 2128 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle); |
1da177e4 LT |
2129 | if (!group) { |
2130 | schedstat_inc(sd, lb_nobusyg[idle]); | |
2131 | goto out_balanced; | |
2132 | } | |
2133 | ||
b910472d | 2134 | busiest = find_busiest_queue(group, idle); |
1da177e4 LT |
2135 | if (!busiest) { |
2136 | schedstat_inc(sd, lb_nobusyq[idle]); | |
2137 | goto out_balanced; | |
2138 | } | |
2139 | ||
db935dbd | 2140 | BUG_ON(busiest == this_rq); |
1da177e4 LT |
2141 | |
2142 | schedstat_add(sd, lb_imbalance[idle], imbalance); | |
2143 | ||
2144 | nr_moved = 0; | |
2145 | if (busiest->nr_running > 1) { | |
2146 | /* | |
2147 | * Attempt to move tasks. If find_busiest_group has found | |
2148 | * an imbalance but busiest->nr_running <= 1, the group is | |
2149 | * still unbalanced. nr_moved simply stays zero, so it is | |
2150 | * correctly treated as an imbalance. | |
2151 | */ | |
e17224bf | 2152 | double_rq_lock(this_rq, busiest); |
1da177e4 | 2153 | nr_moved = move_tasks(this_rq, this_cpu, busiest, |
d6d5cfaf | 2154 | imbalance, sd, idle, &all_pinned); |
e17224bf | 2155 | double_rq_unlock(this_rq, busiest); |
81026794 NP |
2156 | |
2157 | /* All tasks on this runqueue were pinned by CPU affinity */ | |
2158 | if (unlikely(all_pinned)) | |
2159 | goto out_balanced; | |
1da177e4 | 2160 | } |
81026794 | 2161 | |
1da177e4 LT |
2162 | if (!nr_moved) { |
2163 | schedstat_inc(sd, lb_failed[idle]); | |
2164 | sd->nr_balance_failed++; | |
2165 | ||
2166 | if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) { | |
1da177e4 LT |
2167 | |
2168 | spin_lock(&busiest->lock); | |
fa3b6ddc SS |
2169 | |
2170 | /* don't kick the migration_thread, if the curr | |
2171 | * task on busiest cpu can't be moved to this_cpu | |
2172 | */ | |
2173 | if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) { | |
2174 | spin_unlock(&busiest->lock); | |
2175 | all_pinned = 1; | |
2176 | goto out_one_pinned; | |
2177 | } | |
2178 | ||
1da177e4 LT |
2179 | if (!busiest->active_balance) { |
2180 | busiest->active_balance = 1; | |
2181 | busiest->push_cpu = this_cpu; | |
81026794 | 2182 | active_balance = 1; |
1da177e4 LT |
2183 | } |
2184 | spin_unlock(&busiest->lock); | |
81026794 | 2185 | if (active_balance) |
1da177e4 LT |
2186 | wake_up_process(busiest->migration_thread); |
2187 | ||
2188 | /* | |
2189 | * We've kicked active balancing, reset the failure | |
2190 | * counter. | |
2191 | */ | |
39507451 | 2192 | sd->nr_balance_failed = sd->cache_nice_tries+1; |
1da177e4 | 2193 | } |
81026794 | 2194 | } else |
1da177e4 LT |
2195 | sd->nr_balance_failed = 0; |
2196 | ||
81026794 | 2197 | if (likely(!active_balance)) { |
1da177e4 LT |
2198 | /* We were unbalanced, so reset the balancing interval */ |
2199 | sd->balance_interval = sd->min_interval; | |
81026794 NP |
2200 | } else { |
2201 | /* | |
2202 | * If we've begun active balancing, start to back off. This | |
2203 | * case may not be covered by the all_pinned logic if there | |
2204 | * is only 1 task on the busy runqueue (because we don't call | |
2205 | * move_tasks). | |
2206 | */ | |
2207 | if (sd->balance_interval < sd->max_interval) | |
2208 | sd->balance_interval *= 2; | |
1da177e4 LT |
2209 | } |
2210 | ||
5969fe06 NP |
2211 | if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
2212 | return -1; | |
1da177e4 LT |
2213 | return nr_moved; |
2214 | ||
2215 | out_balanced: | |
1da177e4 LT |
2216 | schedstat_inc(sd, lb_balanced[idle]); |
2217 | ||
16cfb1c0 | 2218 | sd->nr_balance_failed = 0; |
fa3b6ddc SS |
2219 | |
2220 | out_one_pinned: | |
1da177e4 | 2221 | /* tune up the balancing interval */ |
77391d71 NP |
2222 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || |
2223 | (sd->balance_interval < sd->max_interval)) | |
1da177e4 LT |
2224 | sd->balance_interval *= 2; |
2225 | ||
5969fe06 NP |
2226 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
2227 | return -1; | |
1da177e4 LT |
2228 | return 0; |
2229 | } | |
2230 | ||
2231 | /* | |
2232 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
2233 | * tasks if there is an imbalance. | |
2234 | * | |
2235 | * Called from schedule when this_rq is about to become idle (NEWLY_IDLE). | |
2236 | * this_rq is locked. | |
2237 | */ | |
2238 | static int load_balance_newidle(int this_cpu, runqueue_t *this_rq, | |
2239 | struct sched_domain *sd) | |
2240 | { | |
2241 | struct sched_group *group; | |
2242 | runqueue_t *busiest = NULL; | |
2243 | unsigned long imbalance; | |
2244 | int nr_moved = 0; | |
5969fe06 NP |
2245 | int sd_idle = 0; |
2246 | ||
2247 | if (sd->flags & SD_SHARE_CPUPOWER) | |
2248 | sd_idle = 1; | |
1da177e4 LT |
2249 | |
2250 | schedstat_inc(sd, lb_cnt[NEWLY_IDLE]); | |
5969fe06 | 2251 | group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE, &sd_idle); |
1da177e4 | 2252 | if (!group) { |
1da177e4 | 2253 | schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]); |
16cfb1c0 | 2254 | goto out_balanced; |
1da177e4 LT |
2255 | } |
2256 | ||
b910472d | 2257 | busiest = find_busiest_queue(group, NEWLY_IDLE); |
db935dbd | 2258 | if (!busiest) { |
1da177e4 | 2259 | schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]); |
16cfb1c0 | 2260 | goto out_balanced; |
1da177e4 LT |
2261 | } |
2262 | ||
db935dbd NP |
2263 | BUG_ON(busiest == this_rq); |
2264 | ||
1da177e4 | 2265 | schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance); |
d6d5cfaf NP |
2266 | |
2267 | nr_moved = 0; | |
2268 | if (busiest->nr_running > 1) { | |
2269 | /* Attempt to move tasks */ | |
2270 | double_lock_balance(this_rq, busiest); | |
2271 | nr_moved = move_tasks(this_rq, this_cpu, busiest, | |
81026794 | 2272 | imbalance, sd, NEWLY_IDLE, NULL); |
d6d5cfaf NP |
2273 | spin_unlock(&busiest->lock); |
2274 | } | |
2275 | ||
5969fe06 | 2276 | if (!nr_moved) { |
1da177e4 | 2277 | schedstat_inc(sd, lb_failed[NEWLY_IDLE]); |
5969fe06 NP |
2278 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
2279 | return -1; | |
2280 | } else | |
16cfb1c0 | 2281 | sd->nr_balance_failed = 0; |
1da177e4 | 2282 | |
1da177e4 | 2283 | return nr_moved; |
16cfb1c0 NP |
2284 | |
2285 | out_balanced: | |
2286 | schedstat_inc(sd, lb_balanced[NEWLY_IDLE]); | |
5969fe06 NP |
2287 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
2288 | return -1; | |
16cfb1c0 NP |
2289 | sd->nr_balance_failed = 0; |
2290 | return 0; | |
1da177e4 LT |
2291 | } |
2292 | ||
2293 | /* | |
2294 | * idle_balance is called by schedule() if this_cpu is about to become | |
2295 | * idle. Attempts to pull tasks from other CPUs. | |
2296 | */ | |
858119e1 | 2297 | static void idle_balance(int this_cpu, runqueue_t *this_rq) |
1da177e4 LT |
2298 | { |
2299 | struct sched_domain *sd; | |
2300 | ||
2301 | for_each_domain(this_cpu, sd) { | |
2302 | if (sd->flags & SD_BALANCE_NEWIDLE) { | |
2303 | if (load_balance_newidle(this_cpu, this_rq, sd)) { | |
2304 | /* We've pulled tasks over so stop searching */ | |
2305 | break; | |
2306 | } | |
2307 | } | |
2308 | } | |
2309 | } | |
2310 | ||
2311 | /* | |
2312 | * active_load_balance is run by migration threads. It pushes running tasks | |
2313 | * off the busiest CPU onto idle CPUs. It requires at least 1 task to be | |
2314 | * running on each physical CPU where possible, and avoids physical / | |
2315 | * logical imbalances. | |
2316 | * | |
2317 | * Called with busiest_rq locked. | |
2318 | */ | |
2319 | static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu) | |
2320 | { | |
2321 | struct sched_domain *sd; | |
1da177e4 | 2322 | runqueue_t *target_rq; |
39507451 NP |
2323 | int target_cpu = busiest_rq->push_cpu; |
2324 | ||
2325 | if (busiest_rq->nr_running <= 1) | |
2326 | /* no task to move */ | |
2327 | return; | |
2328 | ||
2329 | target_rq = cpu_rq(target_cpu); | |
1da177e4 LT |
2330 | |
2331 | /* | |
39507451 NP |
2332 | * This condition is "impossible", if it occurs |
2333 | * we need to fix it. Originally reported by | |
2334 | * Bjorn Helgaas on a 128-cpu setup. | |
1da177e4 | 2335 | */ |
39507451 | 2336 | BUG_ON(busiest_rq == target_rq); |
1da177e4 | 2337 | |
39507451 NP |
2338 | /* move a task from busiest_rq to target_rq */ |
2339 | double_lock_balance(busiest_rq, target_rq); | |
2340 | ||
2341 | /* Search for an sd spanning us and the target CPU. */ | |
2342 | for_each_domain(target_cpu, sd) | |
2343 | if ((sd->flags & SD_LOAD_BALANCE) && | |
2344 | cpu_isset(busiest_cpu, sd->span)) | |
2345 | break; | |
2346 | ||
2347 | if (unlikely(sd == NULL)) | |
2348 | goto out; | |
2349 | ||
2350 | schedstat_inc(sd, alb_cnt); | |
2351 | ||
2352 | if (move_tasks(target_rq, target_cpu, busiest_rq, 1, sd, SCHED_IDLE, NULL)) | |
2353 | schedstat_inc(sd, alb_pushed); | |
2354 | else | |
2355 | schedstat_inc(sd, alb_failed); | |
2356 | out: | |
2357 | spin_unlock(&target_rq->lock); | |
1da177e4 LT |
2358 | } |
2359 | ||
2360 | /* | |
2361 | * rebalance_tick will get called every timer tick, on every CPU. | |
2362 | * | |
2363 | * It checks each scheduling domain to see if it is due to be balanced, | |
2364 | * and initiates a balancing operation if so. | |
2365 | * | |
2366 | * Balancing parameters are set up in arch_init_sched_domains. | |
2367 | */ | |
2368 | ||
2369 | /* Don't have all balancing operations going off at once */ | |
2370 | #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS) | |
2371 | ||
2372 | static void rebalance_tick(int this_cpu, runqueue_t *this_rq, | |
2373 | enum idle_type idle) | |
2374 | { | |
2375 | unsigned long old_load, this_load; | |
2376 | unsigned long j = jiffies + CPU_OFFSET(this_cpu); | |
2377 | struct sched_domain *sd; | |
7897986b | 2378 | int i; |
1da177e4 | 2379 | |
1da177e4 | 2380 | this_load = this_rq->nr_running * SCHED_LOAD_SCALE; |
7897986b NP |
2381 | /* Update our load */ |
2382 | for (i = 0; i < 3; i++) { | |
2383 | unsigned long new_load = this_load; | |
2384 | int scale = 1 << i; | |
2385 | old_load = this_rq->cpu_load[i]; | |
2386 | /* | |
2387 | * Round up the averaging division if load is increasing. This | |
2388 | * prevents us from getting stuck on 9 if the load is 10, for | |
2389 | * example. | |
2390 | */ | |
2391 | if (new_load > old_load) | |
2392 | new_load += scale-1; | |
2393 | this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale; | |
2394 | } | |
1da177e4 LT |
2395 | |
2396 | for_each_domain(this_cpu, sd) { | |
2397 | unsigned long interval; | |
2398 | ||
2399 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
2400 | continue; | |
2401 | ||
2402 | interval = sd->balance_interval; | |
2403 | if (idle != SCHED_IDLE) | |
2404 | interval *= sd->busy_factor; | |
2405 | ||
2406 | /* scale ms to jiffies */ | |
2407 | interval = msecs_to_jiffies(interval); | |
2408 | if (unlikely(!interval)) | |
2409 | interval = 1; | |
2410 | ||
2411 | if (j - sd->last_balance >= interval) { | |
2412 | if (load_balance(this_cpu, this_rq, sd, idle)) { | |
fa3b6ddc SS |
2413 | /* |
2414 | * We've pulled tasks over so either we're no | |
5969fe06 NP |
2415 | * longer idle, or one of our SMT siblings is |
2416 | * not idle. | |
2417 | */ | |
1da177e4 LT |
2418 | idle = NOT_IDLE; |
2419 | } | |
2420 | sd->last_balance += interval; | |
2421 | } | |
2422 | } | |
2423 | } | |
2424 | #else | |
2425 | /* | |
2426 | * on UP we do not need to balance between CPUs: | |
2427 | */ | |
2428 | static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle) | |
2429 | { | |
2430 | } | |
2431 | static inline void idle_balance(int cpu, runqueue_t *rq) | |
2432 | { | |
2433 | } | |
2434 | #endif | |
2435 | ||
2436 | static inline int wake_priority_sleeper(runqueue_t *rq) | |
2437 | { | |
2438 | int ret = 0; | |
2439 | #ifdef CONFIG_SCHED_SMT | |
2440 | spin_lock(&rq->lock); | |
2441 | /* | |
2442 | * If an SMT sibling task has been put to sleep for priority | |
2443 | * reasons reschedule the idle task to see if it can now run. | |
2444 | */ | |
2445 | if (rq->nr_running) { | |
2446 | resched_task(rq->idle); | |
2447 | ret = 1; | |
2448 | } | |
2449 | spin_unlock(&rq->lock); | |
2450 | #endif | |
2451 | return ret; | |
2452 | } | |
2453 | ||
2454 | DEFINE_PER_CPU(struct kernel_stat, kstat); | |
2455 | ||
2456 | EXPORT_PER_CPU_SYMBOL(kstat); | |
2457 | ||
2458 | /* | |
2459 | * This is called on clock ticks and on context switches. | |
2460 | * Bank in p->sched_time the ns elapsed since the last tick or switch. | |
2461 | */ | |
2462 | static inline void update_cpu_clock(task_t *p, runqueue_t *rq, | |
2463 | unsigned long long now) | |
2464 | { | |
2465 | unsigned long long last = max(p->timestamp, rq->timestamp_last_tick); | |
2466 | p->sched_time += now - last; | |
2467 | } | |
2468 | ||
2469 | /* | |
2470 | * Return current->sched_time plus any more ns on the sched_clock | |
2471 | * that have not yet been banked. | |
2472 | */ | |
2473 | unsigned long long current_sched_time(const task_t *tsk) | |
2474 | { | |
2475 | unsigned long long ns; | |
2476 | unsigned long flags; | |
2477 | local_irq_save(flags); | |
2478 | ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick); | |
2479 | ns = tsk->sched_time + (sched_clock() - ns); | |
2480 | local_irq_restore(flags); | |
2481 | return ns; | |
2482 | } | |
2483 | ||
2484 | /* | |
2485 | * We place interactive tasks back into the active array, if possible. | |
2486 | * | |
2487 | * To guarantee that this does not starve expired tasks we ignore the | |
2488 | * interactivity of a task if the first expired task had to wait more | |
2489 | * than a 'reasonable' amount of time. This deadline timeout is | |
2490 | * load-dependent, as the frequency of array switched decreases with | |
2491 | * increasing number of running tasks. We also ignore the interactivity | |
2492 | * if a better static_prio task has expired: | |
2493 | */ | |
2494 | #define EXPIRED_STARVING(rq) \ | |
2495 | ((STARVATION_LIMIT && ((rq)->expired_timestamp && \ | |
2496 | (jiffies - (rq)->expired_timestamp >= \ | |
2497 | STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \ | |
2498 | ((rq)->curr->static_prio > (rq)->best_expired_prio)) | |
2499 | ||
2500 | /* | |
2501 | * Account user cpu time to a process. | |
2502 | * @p: the process that the cpu time gets accounted to | |
2503 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
2504 | * @cputime: the cpu time spent in user space since the last update | |
2505 | */ | |
2506 | void account_user_time(struct task_struct *p, cputime_t cputime) | |
2507 | { | |
2508 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
2509 | cputime64_t tmp; | |
2510 | ||
2511 | p->utime = cputime_add(p->utime, cputime); | |
2512 | ||
2513 | /* Add user time to cpustat. */ | |
2514 | tmp = cputime_to_cputime64(cputime); | |
2515 | if (TASK_NICE(p) > 0) | |
2516 | cpustat->nice = cputime64_add(cpustat->nice, tmp); | |
2517 | else | |
2518 | cpustat->user = cputime64_add(cpustat->user, tmp); | |
2519 | } | |
2520 | ||
2521 | /* | |
2522 | * Account system cpu time to a process. | |
2523 | * @p: the process that the cpu time gets accounted to | |
2524 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
2525 | * @cputime: the cpu time spent in kernel space since the last update | |
2526 | */ | |
2527 | void account_system_time(struct task_struct *p, int hardirq_offset, | |
2528 | cputime_t cputime) | |
2529 | { | |
2530 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
2531 | runqueue_t *rq = this_rq(); | |
2532 | cputime64_t tmp; | |
2533 | ||
2534 | p->stime = cputime_add(p->stime, cputime); | |
2535 | ||
2536 | /* Add system time to cpustat. */ | |
2537 | tmp = cputime_to_cputime64(cputime); | |
2538 | if (hardirq_count() - hardirq_offset) | |
2539 | cpustat->irq = cputime64_add(cpustat->irq, tmp); | |
2540 | else if (softirq_count()) | |
2541 | cpustat->softirq = cputime64_add(cpustat->softirq, tmp); | |
2542 | else if (p != rq->idle) | |
2543 | cpustat->system = cputime64_add(cpustat->system, tmp); | |
2544 | else if (atomic_read(&rq->nr_iowait) > 0) | |
2545 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); | |
2546 | else | |
2547 | cpustat->idle = cputime64_add(cpustat->idle, tmp); | |
2548 | /* Account for system time used */ | |
2549 | acct_update_integrals(p); | |
1da177e4 LT |
2550 | } |
2551 | ||
2552 | /* | |
2553 | * Account for involuntary wait time. | |
2554 | * @p: the process from which the cpu time has been stolen | |
2555 | * @steal: the cpu time spent in involuntary wait | |
2556 | */ | |
2557 | void account_steal_time(struct task_struct *p, cputime_t steal) | |
2558 | { | |
2559 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
2560 | cputime64_t tmp = cputime_to_cputime64(steal); | |
2561 | runqueue_t *rq = this_rq(); | |
2562 | ||
2563 | if (p == rq->idle) { | |
2564 | p->stime = cputime_add(p->stime, steal); | |
2565 | if (atomic_read(&rq->nr_iowait) > 0) | |
2566 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); | |
2567 | else | |
2568 | cpustat->idle = cputime64_add(cpustat->idle, tmp); | |
2569 | } else | |
2570 | cpustat->steal = cputime64_add(cpustat->steal, tmp); | |
2571 | } | |
2572 | ||
2573 | /* | |
2574 | * This function gets called by the timer code, with HZ frequency. | |
2575 | * We call it with interrupts disabled. | |
2576 | * | |
2577 | * It also gets called by the fork code, when changing the parent's | |
2578 | * timeslices. | |
2579 | */ | |
2580 | void scheduler_tick(void) | |
2581 | { | |
2582 | int cpu = smp_processor_id(); | |
2583 | runqueue_t *rq = this_rq(); | |
2584 | task_t *p = current; | |
2585 | unsigned long long now = sched_clock(); | |
2586 | ||
2587 | update_cpu_clock(p, rq, now); | |
2588 | ||
2589 | rq->timestamp_last_tick = now; | |
2590 | ||
2591 | if (p == rq->idle) { | |
2592 | if (wake_priority_sleeper(rq)) | |
2593 | goto out; | |
2594 | rebalance_tick(cpu, rq, SCHED_IDLE); | |
2595 | return; | |
2596 | } | |
2597 | ||
2598 | /* Task might have expired already, but not scheduled off yet */ | |
2599 | if (p->array != rq->active) { | |
2600 | set_tsk_need_resched(p); | |
2601 | goto out; | |
2602 | } | |
2603 | spin_lock(&rq->lock); | |
2604 | /* | |
2605 | * The task was running during this tick - update the | |
2606 | * time slice counter. Note: we do not update a thread's | |
2607 | * priority until it either goes to sleep or uses up its | |
2608 | * timeslice. This makes it possible for interactive tasks | |
2609 | * to use up their timeslices at their highest priority levels. | |
2610 | */ | |
2611 | if (rt_task(p)) { | |
2612 | /* | |
2613 | * RR tasks need a special form of timeslice management. | |
2614 | * FIFO tasks have no timeslices. | |
2615 | */ | |
2616 | if ((p->policy == SCHED_RR) && !--p->time_slice) { | |
2617 | p->time_slice = task_timeslice(p); | |
2618 | p->first_time_slice = 0; | |
2619 | set_tsk_need_resched(p); | |
2620 | ||
2621 | /* put it at the end of the queue: */ | |
2622 | requeue_task(p, rq->active); | |
2623 | } | |
2624 | goto out_unlock; | |
2625 | } | |
2626 | if (!--p->time_slice) { | |
2627 | dequeue_task(p, rq->active); | |
2628 | set_tsk_need_resched(p); | |
2629 | p->prio = effective_prio(p); | |
2630 | p->time_slice = task_timeslice(p); | |
2631 | p->first_time_slice = 0; | |
2632 | ||
2633 | if (!rq->expired_timestamp) | |
2634 | rq->expired_timestamp = jiffies; | |
2635 | if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) { | |
2636 | enqueue_task(p, rq->expired); | |
2637 | if (p->static_prio < rq->best_expired_prio) | |
2638 | rq->best_expired_prio = p->static_prio; | |
2639 | } else | |
2640 | enqueue_task(p, rq->active); | |
2641 | } else { | |
2642 | /* | |
2643 | * Prevent a too long timeslice allowing a task to monopolize | |
2644 | * the CPU. We do this by splitting up the timeslice into | |
2645 | * smaller pieces. | |
2646 | * | |
2647 | * Note: this does not mean the task's timeslices expire or | |
2648 | * get lost in any way, they just might be preempted by | |
2649 | * another task of equal priority. (one with higher | |
2650 | * priority would have preempted this task already.) We | |
2651 | * requeue this task to the end of the list on this priority | |
2652 | * level, which is in essence a round-robin of tasks with | |
2653 | * equal priority. | |
2654 | * | |
2655 | * This only applies to tasks in the interactive | |
2656 | * delta range with at least TIMESLICE_GRANULARITY to requeue. | |
2657 | */ | |
2658 | if (TASK_INTERACTIVE(p) && !((task_timeslice(p) - | |
2659 | p->time_slice) % TIMESLICE_GRANULARITY(p)) && | |
2660 | (p->time_slice >= TIMESLICE_GRANULARITY(p)) && | |
2661 | (p->array == rq->active)) { | |
2662 | ||
2663 | requeue_task(p, rq->active); | |
2664 | set_tsk_need_resched(p); | |
2665 | } | |
2666 | } | |
2667 | out_unlock: | |
2668 | spin_unlock(&rq->lock); | |
2669 | out: | |
2670 | rebalance_tick(cpu, rq, NOT_IDLE); | |
2671 | } | |
2672 | ||
2673 | #ifdef CONFIG_SCHED_SMT | |
fc38ed75 CK |
2674 | static inline void wakeup_busy_runqueue(runqueue_t *rq) |
2675 | { | |
2676 | /* If an SMT runqueue is sleeping due to priority reasons wake it up */ | |
2677 | if (rq->curr == rq->idle && rq->nr_running) | |
2678 | resched_task(rq->idle); | |
2679 | } | |
2680 | ||
858119e1 | 2681 | static void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq) |
1da177e4 | 2682 | { |
41c7ce9a | 2683 | struct sched_domain *tmp, *sd = NULL; |
1da177e4 LT |
2684 | cpumask_t sibling_map; |
2685 | int i; | |
2686 | ||
41c7ce9a NP |
2687 | for_each_domain(this_cpu, tmp) |
2688 | if (tmp->flags & SD_SHARE_CPUPOWER) | |
2689 | sd = tmp; | |
2690 | ||
2691 | if (!sd) | |
1da177e4 LT |
2692 | return; |
2693 | ||
2694 | /* | |
2695 | * Unlock the current runqueue because we have to lock in | |
2696 | * CPU order to avoid deadlocks. Caller knows that we might | |
2697 | * unlock. We keep IRQs disabled. | |
2698 | */ | |
2699 | spin_unlock(&this_rq->lock); | |
2700 | ||
2701 | sibling_map = sd->span; | |
2702 | ||
2703 | for_each_cpu_mask(i, sibling_map) | |
2704 | spin_lock(&cpu_rq(i)->lock); | |
2705 | /* | |
2706 | * We clear this CPU from the mask. This both simplifies the | |
2707 | * inner loop and keps this_rq locked when we exit: | |
2708 | */ | |
2709 | cpu_clear(this_cpu, sibling_map); | |
2710 | ||
2711 | for_each_cpu_mask(i, sibling_map) { | |
2712 | runqueue_t *smt_rq = cpu_rq(i); | |
2713 | ||
fc38ed75 | 2714 | wakeup_busy_runqueue(smt_rq); |
1da177e4 LT |
2715 | } |
2716 | ||
2717 | for_each_cpu_mask(i, sibling_map) | |
2718 | spin_unlock(&cpu_rq(i)->lock); | |
2719 | /* | |
2720 | * We exit with this_cpu's rq still held and IRQs | |
2721 | * still disabled: | |
2722 | */ | |
2723 | } | |
2724 | ||
67f9a619 IM |
2725 | /* |
2726 | * number of 'lost' timeslices this task wont be able to fully | |
2727 | * utilize, if another task runs on a sibling. This models the | |
2728 | * slowdown effect of other tasks running on siblings: | |
2729 | */ | |
2730 | static inline unsigned long smt_slice(task_t *p, struct sched_domain *sd) | |
2731 | { | |
2732 | return p->time_slice * (100 - sd->per_cpu_gain) / 100; | |
2733 | } | |
2734 | ||
858119e1 | 2735 | static int dependent_sleeper(int this_cpu, runqueue_t *this_rq) |
1da177e4 | 2736 | { |
41c7ce9a | 2737 | struct sched_domain *tmp, *sd = NULL; |
1da177e4 LT |
2738 | cpumask_t sibling_map; |
2739 | prio_array_t *array; | |
2740 | int ret = 0, i; | |
2741 | task_t *p; | |
2742 | ||
41c7ce9a NP |
2743 | for_each_domain(this_cpu, tmp) |
2744 | if (tmp->flags & SD_SHARE_CPUPOWER) | |
2745 | sd = tmp; | |
2746 | ||
2747 | if (!sd) | |
1da177e4 LT |
2748 | return 0; |
2749 | ||
2750 | /* | |
2751 | * The same locking rules and details apply as for | |
2752 | * wake_sleeping_dependent(): | |
2753 | */ | |
2754 | spin_unlock(&this_rq->lock); | |
2755 | sibling_map = sd->span; | |
2756 | for_each_cpu_mask(i, sibling_map) | |
2757 | spin_lock(&cpu_rq(i)->lock); | |
2758 | cpu_clear(this_cpu, sibling_map); | |
2759 | ||
2760 | /* | |
2761 | * Establish next task to be run - it might have gone away because | |
2762 | * we released the runqueue lock above: | |
2763 | */ | |
2764 | if (!this_rq->nr_running) | |
2765 | goto out_unlock; | |
2766 | array = this_rq->active; | |
2767 | if (!array->nr_active) | |
2768 | array = this_rq->expired; | |
2769 | BUG_ON(!array->nr_active); | |
2770 | ||
2771 | p = list_entry(array->queue[sched_find_first_bit(array->bitmap)].next, | |
2772 | task_t, run_list); | |
2773 | ||
2774 | for_each_cpu_mask(i, sibling_map) { | |
2775 | runqueue_t *smt_rq = cpu_rq(i); | |
2776 | task_t *smt_curr = smt_rq->curr; | |
2777 | ||
fc38ed75 CK |
2778 | /* Kernel threads do not participate in dependent sleeping */ |
2779 | if (!p->mm || !smt_curr->mm || rt_task(p)) | |
2780 | goto check_smt_task; | |
2781 | ||
1da177e4 LT |
2782 | /* |
2783 | * If a user task with lower static priority than the | |
2784 | * running task on the SMT sibling is trying to schedule, | |
2785 | * delay it till there is proportionately less timeslice | |
2786 | * left of the sibling task to prevent a lower priority | |
2787 | * task from using an unfair proportion of the | |
2788 | * physical cpu's resources. -ck | |
2789 | */ | |
fc38ed75 CK |
2790 | if (rt_task(smt_curr)) { |
2791 | /* | |
2792 | * With real time tasks we run non-rt tasks only | |
2793 | * per_cpu_gain% of the time. | |
2794 | */ | |
2795 | if ((jiffies % DEF_TIMESLICE) > | |
2796 | (sd->per_cpu_gain * DEF_TIMESLICE / 100)) | |
2797 | ret = 1; | |
2798 | } else | |
67f9a619 IM |
2799 | if (smt_curr->static_prio < p->static_prio && |
2800 | !TASK_PREEMPTS_CURR(p, smt_rq) && | |
2801 | smt_slice(smt_curr, sd) > task_timeslice(p)) | |
fc38ed75 CK |
2802 | ret = 1; |
2803 | ||
2804 | check_smt_task: | |
2805 | if ((!smt_curr->mm && smt_curr != smt_rq->idle) || | |
2806 | rt_task(smt_curr)) | |
2807 | continue; | |
2808 | if (!p->mm) { | |
2809 | wakeup_busy_runqueue(smt_rq); | |
2810 | continue; | |
2811 | } | |
1da177e4 LT |
2812 | |
2813 | /* | |
fc38ed75 CK |
2814 | * Reschedule a lower priority task on the SMT sibling for |
2815 | * it to be put to sleep, or wake it up if it has been put to | |
2816 | * sleep for priority reasons to see if it should run now. | |
1da177e4 | 2817 | */ |
fc38ed75 CK |
2818 | if (rt_task(p)) { |
2819 | if ((jiffies % DEF_TIMESLICE) > | |
2820 | (sd->per_cpu_gain * DEF_TIMESLICE / 100)) | |
2821 | resched_task(smt_curr); | |
2822 | } else { | |
67f9a619 IM |
2823 | if (TASK_PREEMPTS_CURR(p, smt_rq) && |
2824 | smt_slice(p, sd) > task_timeslice(smt_curr)) | |
fc38ed75 CK |
2825 | resched_task(smt_curr); |
2826 | else | |
2827 | wakeup_busy_runqueue(smt_rq); | |
2828 | } | |
1da177e4 LT |
2829 | } |
2830 | out_unlock: | |
2831 | for_each_cpu_mask(i, sibling_map) | |
2832 | spin_unlock(&cpu_rq(i)->lock); | |
2833 | return ret; | |
2834 | } | |
2835 | #else | |
2836 | static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq) | |
2837 | { | |
2838 | } | |
2839 | ||
2840 | static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq) | |
2841 | { | |
2842 | return 0; | |
2843 | } | |
2844 | #endif | |
2845 | ||
2846 | #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT) | |
2847 | ||
2848 | void fastcall add_preempt_count(int val) | |
2849 | { | |
2850 | /* | |
2851 | * Underflow? | |
2852 | */ | |
be5b4fbd | 2853 | BUG_ON((preempt_count() < 0)); |
1da177e4 LT |
2854 | preempt_count() += val; |
2855 | /* | |
2856 | * Spinlock count overflowing soon? | |
2857 | */ | |
2858 | BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10); | |
2859 | } | |
2860 | EXPORT_SYMBOL(add_preempt_count); | |
2861 | ||
2862 | void fastcall sub_preempt_count(int val) | |
2863 | { | |
2864 | /* | |
2865 | * Underflow? | |
2866 | */ | |
2867 | BUG_ON(val > preempt_count()); | |
2868 | /* | |
2869 | * Is the spinlock portion underflowing? | |
2870 | */ | |
2871 | BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK)); | |
2872 | preempt_count() -= val; | |
2873 | } | |
2874 | EXPORT_SYMBOL(sub_preempt_count); | |
2875 | ||
2876 | #endif | |
2877 | ||
3dee386e CK |
2878 | static inline int interactive_sleep(enum sleep_type sleep_type) |
2879 | { | |
2880 | return (sleep_type == SLEEP_INTERACTIVE || | |
2881 | sleep_type == SLEEP_INTERRUPTED); | |
2882 | } | |
2883 | ||
1da177e4 LT |
2884 | /* |
2885 | * schedule() is the main scheduler function. | |
2886 | */ | |
2887 | asmlinkage void __sched schedule(void) | |
2888 | { | |
2889 | long *switch_count; | |
2890 | task_t *prev, *next; | |
2891 | runqueue_t *rq; | |
2892 | prio_array_t *array; | |
2893 | struct list_head *queue; | |
2894 | unsigned long long now; | |
2895 | unsigned long run_time; | |
a3464a10 | 2896 | int cpu, idx, new_prio; |
1da177e4 LT |
2897 | |
2898 | /* | |
2899 | * Test if we are atomic. Since do_exit() needs to call into | |
2900 | * schedule() atomically, we ignore that path for now. | |
2901 | * Otherwise, whine if we are scheduling when we should not be. | |
2902 | */ | |
77e4bfbc AM |
2903 | if (unlikely(in_atomic() && !current->exit_state)) { |
2904 | printk(KERN_ERR "BUG: scheduling while atomic: " | |
2905 | "%s/0x%08x/%d\n", | |
2906 | current->comm, preempt_count(), current->pid); | |
2907 | dump_stack(); | |
1da177e4 LT |
2908 | } |
2909 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); | |
2910 | ||
2911 | need_resched: | |
2912 | preempt_disable(); | |
2913 | prev = current; | |
2914 | release_kernel_lock(prev); | |
2915 | need_resched_nonpreemptible: | |
2916 | rq = this_rq(); | |
2917 | ||
2918 | /* | |
2919 | * The idle thread is not allowed to schedule! | |
2920 | * Remove this check after it has been exercised a bit. | |
2921 | */ | |
2922 | if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) { | |
2923 | printk(KERN_ERR "bad: scheduling from the idle thread!\n"); | |
2924 | dump_stack(); | |
2925 | } | |
2926 | ||
2927 | schedstat_inc(rq, sched_cnt); | |
2928 | now = sched_clock(); | |
238628ed | 2929 | if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) { |
1da177e4 | 2930 | run_time = now - prev->timestamp; |
238628ed | 2931 | if (unlikely((long long)(now - prev->timestamp) < 0)) |
1da177e4 LT |
2932 | run_time = 0; |
2933 | } else | |
2934 | run_time = NS_MAX_SLEEP_AVG; | |
2935 | ||
2936 | /* | |
2937 | * Tasks charged proportionately less run_time at high sleep_avg to | |
2938 | * delay them losing their interactive status | |
2939 | */ | |
2940 | run_time /= (CURRENT_BONUS(prev) ? : 1); | |
2941 | ||
2942 | spin_lock_irq(&rq->lock); | |
2943 | ||
2944 | if (unlikely(prev->flags & PF_DEAD)) | |
2945 | prev->state = EXIT_DEAD; | |
2946 | ||
2947 | switch_count = &prev->nivcsw; | |
2948 | if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { | |
2949 | switch_count = &prev->nvcsw; | |
2950 | if (unlikely((prev->state & TASK_INTERRUPTIBLE) && | |
2951 | unlikely(signal_pending(prev)))) | |
2952 | prev->state = TASK_RUNNING; | |
2953 | else { | |
2954 | if (prev->state == TASK_UNINTERRUPTIBLE) | |
2955 | rq->nr_uninterruptible++; | |
2956 | deactivate_task(prev, rq); | |
2957 | } | |
2958 | } | |
2959 | ||
2960 | cpu = smp_processor_id(); | |
2961 | if (unlikely(!rq->nr_running)) { | |
2962 | go_idle: | |
2963 | idle_balance(cpu, rq); | |
2964 | if (!rq->nr_running) { | |
2965 | next = rq->idle; | |
2966 | rq->expired_timestamp = 0; | |
2967 | wake_sleeping_dependent(cpu, rq); | |
2968 | /* | |
2969 | * wake_sleeping_dependent() might have released | |
2970 | * the runqueue, so break out if we got new | |
2971 | * tasks meanwhile: | |
2972 | */ | |
2973 | if (!rq->nr_running) | |
2974 | goto switch_tasks; | |
2975 | } | |
2976 | } else { | |
2977 | if (dependent_sleeper(cpu, rq)) { | |
2978 | next = rq->idle; | |
2979 | goto switch_tasks; | |
2980 | } | |
2981 | /* | |
2982 | * dependent_sleeper() releases and reacquires the runqueue | |
2983 | * lock, hence go into the idle loop if the rq went | |
2984 | * empty meanwhile: | |
2985 | */ | |
2986 | if (unlikely(!rq->nr_running)) | |
2987 | goto go_idle; | |
2988 | } | |
2989 | ||
2990 | array = rq->active; | |
2991 | if (unlikely(!array->nr_active)) { | |
2992 | /* | |
2993 | * Switch the active and expired arrays. | |
2994 | */ | |
2995 | schedstat_inc(rq, sched_switch); | |
2996 | rq->active = rq->expired; | |
2997 | rq->expired = array; | |
2998 | array = rq->active; | |
2999 | rq->expired_timestamp = 0; | |
3000 | rq->best_expired_prio = MAX_PRIO; | |
3001 | } | |
3002 | ||
3003 | idx = sched_find_first_bit(array->bitmap); | |
3004 | queue = array->queue + idx; | |
3005 | next = list_entry(queue->next, task_t, run_list); | |
3006 | ||
3dee386e | 3007 | if (!rt_task(next) && interactive_sleep(next->sleep_type)) { |
1da177e4 | 3008 | unsigned long long delta = now - next->timestamp; |
238628ed | 3009 | if (unlikely((long long)(now - next->timestamp) < 0)) |
1da177e4 LT |
3010 | delta = 0; |
3011 | ||
3dee386e | 3012 | if (next->sleep_type == SLEEP_INTERACTIVE) |
1da177e4 LT |
3013 | delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128; |
3014 | ||
3015 | array = next->array; | |
a3464a10 CS |
3016 | new_prio = recalc_task_prio(next, next->timestamp + delta); |
3017 | ||
3018 | if (unlikely(next->prio != new_prio)) { | |
3019 | dequeue_task(next, array); | |
3020 | next->prio = new_prio; | |
3021 | enqueue_task(next, array); | |
3022 | } else | |
3023 | requeue_task(next, array); | |
1da177e4 | 3024 | } |
3dee386e | 3025 | next->sleep_type = SLEEP_NORMAL; |
1da177e4 LT |
3026 | switch_tasks: |
3027 | if (next == rq->idle) | |
3028 | schedstat_inc(rq, sched_goidle); | |
3029 | prefetch(next); | |
383f2835 | 3030 | prefetch_stack(next); |
1da177e4 LT |
3031 | clear_tsk_need_resched(prev); |
3032 | rcu_qsctr_inc(task_cpu(prev)); | |
3033 | ||
3034 | update_cpu_clock(prev, rq, now); | |
3035 | ||
3036 | prev->sleep_avg -= run_time; | |
3037 | if ((long)prev->sleep_avg <= 0) | |
3038 | prev->sleep_avg = 0; | |
3039 | prev->timestamp = prev->last_ran = now; | |
3040 | ||
3041 | sched_info_switch(prev, next); | |
3042 | if (likely(prev != next)) { | |
3043 | next->timestamp = now; | |
3044 | rq->nr_switches++; | |
3045 | rq->curr = next; | |
3046 | ++*switch_count; | |
3047 | ||
4866cde0 | 3048 | prepare_task_switch(rq, next); |
1da177e4 LT |
3049 | prev = context_switch(rq, prev, next); |
3050 | barrier(); | |
4866cde0 NP |
3051 | /* |
3052 | * this_rq must be evaluated again because prev may have moved | |
3053 | * CPUs since it called schedule(), thus the 'rq' on its stack | |
3054 | * frame will be invalid. | |
3055 | */ | |
3056 | finish_task_switch(this_rq(), prev); | |
1da177e4 LT |
3057 | } else |
3058 | spin_unlock_irq(&rq->lock); | |
3059 | ||
3060 | prev = current; | |
3061 | if (unlikely(reacquire_kernel_lock(prev) < 0)) | |
3062 | goto need_resched_nonpreemptible; | |
3063 | preempt_enable_no_resched(); | |
3064 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
3065 | goto need_resched; | |
3066 | } | |
3067 | ||
3068 | EXPORT_SYMBOL(schedule); | |
3069 | ||
3070 | #ifdef CONFIG_PREEMPT | |
3071 | /* | |
3072 | * this is is the entry point to schedule() from in-kernel preemption | |
3073 | * off of preempt_enable. Kernel preemptions off return from interrupt | |
3074 | * occur there and call schedule directly. | |
3075 | */ | |
3076 | asmlinkage void __sched preempt_schedule(void) | |
3077 | { | |
3078 | struct thread_info *ti = current_thread_info(); | |
3079 | #ifdef CONFIG_PREEMPT_BKL | |
3080 | struct task_struct *task = current; | |
3081 | int saved_lock_depth; | |
3082 | #endif | |
3083 | /* | |
3084 | * If there is a non-zero preempt_count or interrupts are disabled, | |
3085 | * we do not want to preempt the current task. Just return.. | |
3086 | */ | |
3087 | if (unlikely(ti->preempt_count || irqs_disabled())) | |
3088 | return; | |
3089 | ||
3090 | need_resched: | |
3091 | add_preempt_count(PREEMPT_ACTIVE); | |
3092 | /* | |
3093 | * We keep the big kernel semaphore locked, but we | |
3094 | * clear ->lock_depth so that schedule() doesnt | |
3095 | * auto-release the semaphore: | |
3096 | */ | |
3097 | #ifdef CONFIG_PREEMPT_BKL | |
3098 | saved_lock_depth = task->lock_depth; | |
3099 | task->lock_depth = -1; | |
3100 | #endif | |
3101 | schedule(); | |
3102 | #ifdef CONFIG_PREEMPT_BKL | |
3103 | task->lock_depth = saved_lock_depth; | |
3104 | #endif | |
3105 | sub_preempt_count(PREEMPT_ACTIVE); | |
3106 | ||
3107 | /* we could miss a preemption opportunity between schedule and now */ | |
3108 | barrier(); | |
3109 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
3110 | goto need_resched; | |
3111 | } | |
3112 | ||
3113 | EXPORT_SYMBOL(preempt_schedule); | |
3114 | ||
3115 | /* | |
3116 | * this is is the entry point to schedule() from kernel preemption | |
3117 | * off of irq context. | |
3118 | * Note, that this is called and return with irqs disabled. This will | |
3119 | * protect us against recursive calling from irq. | |
3120 | */ | |
3121 | asmlinkage void __sched preempt_schedule_irq(void) | |
3122 | { | |
3123 | struct thread_info *ti = current_thread_info(); | |
3124 | #ifdef CONFIG_PREEMPT_BKL | |
3125 | struct task_struct *task = current; | |
3126 | int saved_lock_depth; | |
3127 | #endif | |
3128 | /* Catch callers which need to be fixed*/ | |
3129 | BUG_ON(ti->preempt_count || !irqs_disabled()); | |
3130 | ||
3131 | need_resched: | |
3132 | add_preempt_count(PREEMPT_ACTIVE); | |
3133 | /* | |
3134 | * We keep the big kernel semaphore locked, but we | |
3135 | * clear ->lock_depth so that schedule() doesnt | |
3136 | * auto-release the semaphore: | |
3137 | */ | |
3138 | #ifdef CONFIG_PREEMPT_BKL | |
3139 | saved_lock_depth = task->lock_depth; | |
3140 | task->lock_depth = -1; | |
3141 | #endif | |
3142 | local_irq_enable(); | |
3143 | schedule(); | |
3144 | local_irq_disable(); | |
3145 | #ifdef CONFIG_PREEMPT_BKL | |
3146 | task->lock_depth = saved_lock_depth; | |
3147 | #endif | |
3148 | sub_preempt_count(PREEMPT_ACTIVE); | |
3149 | ||
3150 | /* we could miss a preemption opportunity between schedule and now */ | |
3151 | barrier(); | |
3152 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
3153 | goto need_resched; | |
3154 | } | |
3155 | ||
3156 | #endif /* CONFIG_PREEMPT */ | |
3157 | ||
95cdf3b7 IM |
3158 | int default_wake_function(wait_queue_t *curr, unsigned mode, int sync, |
3159 | void *key) | |
1da177e4 | 3160 | { |
c43dc2fd | 3161 | task_t *p = curr->private; |
1da177e4 LT |
3162 | return try_to_wake_up(p, mode, sync); |
3163 | } | |
3164 | ||
3165 | EXPORT_SYMBOL(default_wake_function); | |
3166 | ||
3167 | /* | |
3168 | * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just | |
3169 | * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve | |
3170 | * number) then we wake all the non-exclusive tasks and one exclusive task. | |
3171 | * | |
3172 | * There are circumstances in which we can try to wake a task which has already | |
3173 | * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns | |
3174 | * zero in this (rare) case, and we handle it by continuing to scan the queue. | |
3175 | */ | |
3176 | static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, | |
3177 | int nr_exclusive, int sync, void *key) | |
3178 | { | |
3179 | struct list_head *tmp, *next; | |
3180 | ||
3181 | list_for_each_safe(tmp, next, &q->task_list) { | |
3182 | wait_queue_t *curr; | |
3183 | unsigned flags; | |
3184 | curr = list_entry(tmp, wait_queue_t, task_list); | |
3185 | flags = curr->flags; | |
3186 | if (curr->func(curr, mode, sync, key) && | |
3187 | (flags & WQ_FLAG_EXCLUSIVE) && | |
3188 | !--nr_exclusive) | |
3189 | break; | |
3190 | } | |
3191 | } | |
3192 | ||
3193 | /** | |
3194 | * __wake_up - wake up threads blocked on a waitqueue. | |
3195 | * @q: the waitqueue | |
3196 | * @mode: which threads | |
3197 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
67be2dd1 | 3198 | * @key: is directly passed to the wakeup function |
1da177e4 LT |
3199 | */ |
3200 | void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode, | |
95cdf3b7 | 3201 | int nr_exclusive, void *key) |
1da177e4 LT |
3202 | { |
3203 | unsigned long flags; | |
3204 | ||
3205 | spin_lock_irqsave(&q->lock, flags); | |
3206 | __wake_up_common(q, mode, nr_exclusive, 0, key); | |
3207 | spin_unlock_irqrestore(&q->lock, flags); | |
3208 | } | |
3209 | ||
3210 | EXPORT_SYMBOL(__wake_up); | |
3211 | ||
3212 | /* | |
3213 | * Same as __wake_up but called with the spinlock in wait_queue_head_t held. | |
3214 | */ | |
3215 | void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode) | |
3216 | { | |
3217 | __wake_up_common(q, mode, 1, 0, NULL); | |
3218 | } | |
3219 | ||
3220 | /** | |
67be2dd1 | 3221 | * __wake_up_sync - wake up threads blocked on a waitqueue. |
1da177e4 LT |
3222 | * @q: the waitqueue |
3223 | * @mode: which threads | |
3224 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
3225 | * | |
3226 | * The sync wakeup differs that the waker knows that it will schedule | |
3227 | * away soon, so while the target thread will be woken up, it will not | |
3228 | * be migrated to another CPU - ie. the two threads are 'synchronized' | |
3229 | * with each other. This can prevent needless bouncing between CPUs. | |
3230 | * | |
3231 | * On UP it can prevent extra preemption. | |
3232 | */ | |
95cdf3b7 IM |
3233 | void fastcall |
3234 | __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) | |
1da177e4 LT |
3235 | { |
3236 | unsigned long flags; | |
3237 | int sync = 1; | |
3238 | ||
3239 | if (unlikely(!q)) | |
3240 | return; | |
3241 | ||
3242 | if (unlikely(!nr_exclusive)) | |
3243 | sync = 0; | |
3244 | ||
3245 | spin_lock_irqsave(&q->lock, flags); | |
3246 | __wake_up_common(q, mode, nr_exclusive, sync, NULL); | |
3247 | spin_unlock_irqrestore(&q->lock, flags); | |
3248 | } | |
3249 | EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ | |
3250 | ||
3251 | void fastcall complete(struct completion *x) | |
3252 | { | |
3253 | unsigned long flags; | |
3254 | ||
3255 | spin_lock_irqsave(&x->wait.lock, flags); | |
3256 | x->done++; | |
3257 | __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, | |
3258 | 1, 0, NULL); | |
3259 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
3260 | } | |
3261 | EXPORT_SYMBOL(complete); | |
3262 | ||
3263 | void fastcall complete_all(struct completion *x) | |
3264 | { | |
3265 | unsigned long flags; | |
3266 | ||
3267 | spin_lock_irqsave(&x->wait.lock, flags); | |
3268 | x->done += UINT_MAX/2; | |
3269 | __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, | |
3270 | 0, 0, NULL); | |
3271 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
3272 | } | |
3273 | EXPORT_SYMBOL(complete_all); | |
3274 | ||
3275 | void fastcall __sched wait_for_completion(struct completion *x) | |
3276 | { | |
3277 | might_sleep(); | |
3278 | spin_lock_irq(&x->wait.lock); | |
3279 | if (!x->done) { | |
3280 | DECLARE_WAITQUEUE(wait, current); | |
3281 | ||
3282 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3283 | __add_wait_queue_tail(&x->wait, &wait); | |
3284 | do { | |
3285 | __set_current_state(TASK_UNINTERRUPTIBLE); | |
3286 | spin_unlock_irq(&x->wait.lock); | |
3287 | schedule(); | |
3288 | spin_lock_irq(&x->wait.lock); | |
3289 | } while (!x->done); | |
3290 | __remove_wait_queue(&x->wait, &wait); | |
3291 | } | |
3292 | x->done--; | |
3293 | spin_unlock_irq(&x->wait.lock); | |
3294 | } | |
3295 | EXPORT_SYMBOL(wait_for_completion); | |
3296 | ||
3297 | unsigned long fastcall __sched | |
3298 | wait_for_completion_timeout(struct completion *x, unsigned long timeout) | |
3299 | { | |
3300 | might_sleep(); | |
3301 | ||
3302 | spin_lock_irq(&x->wait.lock); | |
3303 | if (!x->done) { | |
3304 | DECLARE_WAITQUEUE(wait, current); | |
3305 | ||
3306 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3307 | __add_wait_queue_tail(&x->wait, &wait); | |
3308 | do { | |
3309 | __set_current_state(TASK_UNINTERRUPTIBLE); | |
3310 | spin_unlock_irq(&x->wait.lock); | |
3311 | timeout = schedule_timeout(timeout); | |
3312 | spin_lock_irq(&x->wait.lock); | |
3313 | if (!timeout) { | |
3314 | __remove_wait_queue(&x->wait, &wait); | |
3315 | goto out; | |
3316 | } | |
3317 | } while (!x->done); | |
3318 | __remove_wait_queue(&x->wait, &wait); | |
3319 | } | |
3320 | x->done--; | |
3321 | out: | |
3322 | spin_unlock_irq(&x->wait.lock); | |
3323 | return timeout; | |
3324 | } | |
3325 | EXPORT_SYMBOL(wait_for_completion_timeout); | |
3326 | ||
3327 | int fastcall __sched wait_for_completion_interruptible(struct completion *x) | |
3328 | { | |
3329 | int ret = 0; | |
3330 | ||
3331 | might_sleep(); | |
3332 | ||
3333 | spin_lock_irq(&x->wait.lock); | |
3334 | if (!x->done) { | |
3335 | DECLARE_WAITQUEUE(wait, current); | |
3336 | ||
3337 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3338 | __add_wait_queue_tail(&x->wait, &wait); | |
3339 | do { | |
3340 | if (signal_pending(current)) { | |
3341 | ret = -ERESTARTSYS; | |
3342 | __remove_wait_queue(&x->wait, &wait); | |
3343 | goto out; | |
3344 | } | |
3345 | __set_current_state(TASK_INTERRUPTIBLE); | |
3346 | spin_unlock_irq(&x->wait.lock); | |
3347 | schedule(); | |
3348 | spin_lock_irq(&x->wait.lock); | |
3349 | } while (!x->done); | |
3350 | __remove_wait_queue(&x->wait, &wait); | |
3351 | } | |
3352 | x->done--; | |
3353 | out: | |
3354 | spin_unlock_irq(&x->wait.lock); | |
3355 | ||
3356 | return ret; | |
3357 | } | |
3358 | EXPORT_SYMBOL(wait_for_completion_interruptible); | |
3359 | ||
3360 | unsigned long fastcall __sched | |
3361 | wait_for_completion_interruptible_timeout(struct completion *x, | |
3362 | unsigned long timeout) | |
3363 | { | |
3364 | might_sleep(); | |
3365 | ||
3366 | spin_lock_irq(&x->wait.lock); | |
3367 | if (!x->done) { | |
3368 | DECLARE_WAITQUEUE(wait, current); | |
3369 | ||
3370 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3371 | __add_wait_queue_tail(&x->wait, &wait); | |
3372 | do { | |
3373 | if (signal_pending(current)) { | |
3374 | timeout = -ERESTARTSYS; | |
3375 | __remove_wait_queue(&x->wait, &wait); | |
3376 | goto out; | |
3377 | } | |
3378 | __set_current_state(TASK_INTERRUPTIBLE); | |
3379 | spin_unlock_irq(&x->wait.lock); | |
3380 | timeout = schedule_timeout(timeout); | |
3381 | spin_lock_irq(&x->wait.lock); | |
3382 | if (!timeout) { | |
3383 | __remove_wait_queue(&x->wait, &wait); | |
3384 | goto out; | |
3385 | } | |
3386 | } while (!x->done); | |
3387 | __remove_wait_queue(&x->wait, &wait); | |
3388 | } | |
3389 | x->done--; | |
3390 | out: | |
3391 | spin_unlock_irq(&x->wait.lock); | |
3392 | return timeout; | |
3393 | } | |
3394 | EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); | |
3395 | ||
3396 | ||
3397 | #define SLEEP_ON_VAR \ | |
3398 | unsigned long flags; \ | |
3399 | wait_queue_t wait; \ | |
3400 | init_waitqueue_entry(&wait, current); | |
3401 | ||
3402 | #define SLEEP_ON_HEAD \ | |
3403 | spin_lock_irqsave(&q->lock,flags); \ | |
3404 | __add_wait_queue(q, &wait); \ | |
3405 | spin_unlock(&q->lock); | |
3406 | ||
3407 | #define SLEEP_ON_TAIL \ | |
3408 | spin_lock_irq(&q->lock); \ | |
3409 | __remove_wait_queue(q, &wait); \ | |
3410 | spin_unlock_irqrestore(&q->lock, flags); | |
3411 | ||
3412 | void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q) | |
3413 | { | |
3414 | SLEEP_ON_VAR | |
3415 | ||
3416 | current->state = TASK_INTERRUPTIBLE; | |
3417 | ||
3418 | SLEEP_ON_HEAD | |
3419 | schedule(); | |
3420 | SLEEP_ON_TAIL | |
3421 | } | |
3422 | ||
3423 | EXPORT_SYMBOL(interruptible_sleep_on); | |
3424 | ||
95cdf3b7 IM |
3425 | long fastcall __sched |
3426 | interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) | |
1da177e4 LT |
3427 | { |
3428 | SLEEP_ON_VAR | |
3429 | ||
3430 | current->state = TASK_INTERRUPTIBLE; | |
3431 | ||
3432 | SLEEP_ON_HEAD | |
3433 | timeout = schedule_timeout(timeout); | |
3434 | SLEEP_ON_TAIL | |
3435 | ||
3436 | return timeout; | |
3437 | } | |
3438 | ||
3439 | EXPORT_SYMBOL(interruptible_sleep_on_timeout); | |
3440 | ||
3441 | void fastcall __sched sleep_on(wait_queue_head_t *q) | |
3442 | { | |
3443 | SLEEP_ON_VAR | |
3444 | ||
3445 | current->state = TASK_UNINTERRUPTIBLE; | |
3446 | ||
3447 | SLEEP_ON_HEAD | |
3448 | schedule(); | |
3449 | SLEEP_ON_TAIL | |
3450 | } | |
3451 | ||
3452 | EXPORT_SYMBOL(sleep_on); | |
3453 | ||
3454 | long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) | |
3455 | { | |
3456 | SLEEP_ON_VAR | |
3457 | ||
3458 | current->state = TASK_UNINTERRUPTIBLE; | |
3459 | ||
3460 | SLEEP_ON_HEAD | |
3461 | timeout = schedule_timeout(timeout); | |
3462 | SLEEP_ON_TAIL | |
3463 | ||
3464 | return timeout; | |
3465 | } | |
3466 | ||
3467 | EXPORT_SYMBOL(sleep_on_timeout); | |
3468 | ||
3469 | void set_user_nice(task_t *p, long nice) | |
3470 | { | |
3471 | unsigned long flags; | |
3472 | prio_array_t *array; | |
3473 | runqueue_t *rq; | |
3474 | int old_prio, new_prio, delta; | |
3475 | ||
3476 | if (TASK_NICE(p) == nice || nice < -20 || nice > 19) | |
3477 | return; | |
3478 | /* | |
3479 | * We have to be careful, if called from sys_setpriority(), | |
3480 | * the task might be in the middle of scheduling on another CPU. | |
3481 | */ | |
3482 | rq = task_rq_lock(p, &flags); | |
3483 | /* | |
3484 | * The RT priorities are set via sched_setscheduler(), but we still | |
3485 | * allow the 'normal' nice value to be set - but as expected | |
3486 | * it wont have any effect on scheduling until the task is | |
b0a9499c | 3487 | * not SCHED_NORMAL/SCHED_BATCH: |
1da177e4 LT |
3488 | */ |
3489 | if (rt_task(p)) { | |
3490 | p->static_prio = NICE_TO_PRIO(nice); | |
3491 | goto out_unlock; | |
3492 | } | |
3493 | array = p->array; | |
a2000572 | 3494 | if (array) |
1da177e4 LT |
3495 | dequeue_task(p, array); |
3496 | ||
3497 | old_prio = p->prio; | |
3498 | new_prio = NICE_TO_PRIO(nice); | |
3499 | delta = new_prio - old_prio; | |
3500 | p->static_prio = NICE_TO_PRIO(nice); | |
3501 | p->prio += delta; | |
3502 | ||
3503 | if (array) { | |
3504 | enqueue_task(p, array); | |
3505 | /* | |
3506 | * If the task increased its priority or is running and | |
3507 | * lowered its priority, then reschedule its CPU: | |
3508 | */ | |
3509 | if (delta < 0 || (delta > 0 && task_running(rq, p))) | |
3510 | resched_task(rq->curr); | |
3511 | } | |
3512 | out_unlock: | |
3513 | task_rq_unlock(rq, &flags); | |
3514 | } | |
3515 | ||
3516 | EXPORT_SYMBOL(set_user_nice); | |
3517 | ||
e43379f1 MM |
3518 | /* |
3519 | * can_nice - check if a task can reduce its nice value | |
3520 | * @p: task | |
3521 | * @nice: nice value | |
3522 | */ | |
3523 | int can_nice(const task_t *p, const int nice) | |
3524 | { | |
024f4747 MM |
3525 | /* convert nice value [19,-20] to rlimit style value [1,40] */ |
3526 | int nice_rlim = 20 - nice; | |
e43379f1 MM |
3527 | return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur || |
3528 | capable(CAP_SYS_NICE)); | |
3529 | } | |
3530 | ||
1da177e4 LT |
3531 | #ifdef __ARCH_WANT_SYS_NICE |
3532 | ||
3533 | /* | |
3534 | * sys_nice - change the priority of the current process. | |
3535 | * @increment: priority increment | |
3536 | * | |
3537 | * sys_setpriority is a more generic, but much slower function that | |
3538 | * does similar things. | |
3539 | */ | |
3540 | asmlinkage long sys_nice(int increment) | |
3541 | { | |
3542 | int retval; | |
3543 | long nice; | |
3544 | ||
3545 | /* | |
3546 | * Setpriority might change our priority at the same moment. | |
3547 | * We don't have to worry. Conceptually one call occurs first | |
3548 | * and we have a single winner. | |
3549 | */ | |
e43379f1 MM |
3550 | if (increment < -40) |
3551 | increment = -40; | |
1da177e4 LT |
3552 | if (increment > 40) |
3553 | increment = 40; | |
3554 | ||
3555 | nice = PRIO_TO_NICE(current->static_prio) + increment; | |
3556 | if (nice < -20) | |
3557 | nice = -20; | |
3558 | if (nice > 19) | |
3559 | nice = 19; | |
3560 | ||
e43379f1 MM |
3561 | if (increment < 0 && !can_nice(current, nice)) |
3562 | return -EPERM; | |
3563 | ||
1da177e4 LT |
3564 | retval = security_task_setnice(current, nice); |
3565 | if (retval) | |
3566 | return retval; | |
3567 | ||
3568 | set_user_nice(current, nice); | |
3569 | return 0; | |
3570 | } | |
3571 | ||
3572 | #endif | |
3573 | ||
3574 | /** | |
3575 | * task_prio - return the priority value of a given task. | |
3576 | * @p: the task in question. | |
3577 | * | |
3578 | * This is the priority value as seen by users in /proc. | |
3579 | * RT tasks are offset by -200. Normal tasks are centered | |
3580 | * around 0, value goes from -16 to +15. | |
3581 | */ | |
3582 | int task_prio(const task_t *p) | |
3583 | { | |
3584 | return p->prio - MAX_RT_PRIO; | |
3585 | } | |
3586 | ||
3587 | /** | |
3588 | * task_nice - return the nice value of a given task. | |
3589 | * @p: the task in question. | |
3590 | */ | |
3591 | int task_nice(const task_t *p) | |
3592 | { | |
3593 | return TASK_NICE(p); | |
3594 | } | |
1da177e4 | 3595 | EXPORT_SYMBOL_GPL(task_nice); |
1da177e4 LT |
3596 | |
3597 | /** | |
3598 | * idle_cpu - is a given cpu idle currently? | |
3599 | * @cpu: the processor in question. | |
3600 | */ | |
3601 | int idle_cpu(int cpu) | |
3602 | { | |
3603 | return cpu_curr(cpu) == cpu_rq(cpu)->idle; | |
3604 | } | |
3605 | ||
1da177e4 LT |
3606 | /** |
3607 | * idle_task - return the idle task for a given cpu. | |
3608 | * @cpu: the processor in question. | |
3609 | */ | |
3610 | task_t *idle_task(int cpu) | |
3611 | { | |
3612 | return cpu_rq(cpu)->idle; | |
3613 | } | |
3614 | ||
3615 | /** | |
3616 | * find_process_by_pid - find a process with a matching PID value. | |
3617 | * @pid: the pid in question. | |
3618 | */ | |
3619 | static inline task_t *find_process_by_pid(pid_t pid) | |
3620 | { | |
3621 | return pid ? find_task_by_pid(pid) : current; | |
3622 | } | |
3623 | ||
3624 | /* Actually do priority change: must hold rq lock. */ | |
3625 | static void __setscheduler(struct task_struct *p, int policy, int prio) | |
3626 | { | |
3627 | BUG_ON(p->array); | |
3628 | p->policy = policy; | |
3629 | p->rt_priority = prio; | |
b0a9499c | 3630 | if (policy != SCHED_NORMAL && policy != SCHED_BATCH) { |
d46523ea | 3631 | p->prio = MAX_RT_PRIO-1 - p->rt_priority; |
b0a9499c | 3632 | } else { |
1da177e4 | 3633 | p->prio = p->static_prio; |
b0a9499c IM |
3634 | /* |
3635 | * SCHED_BATCH tasks are treated as perpetual CPU hogs: | |
3636 | */ | |
3637 | if (policy == SCHED_BATCH) | |
3638 | p->sleep_avg = 0; | |
3639 | } | |
1da177e4 LT |
3640 | } |
3641 | ||
3642 | /** | |
3643 | * sched_setscheduler - change the scheduling policy and/or RT priority of | |
3644 | * a thread. | |
3645 | * @p: the task in question. | |
3646 | * @policy: new policy. | |
3647 | * @param: structure containing the new RT priority. | |
3648 | */ | |
95cdf3b7 IM |
3649 | int sched_setscheduler(struct task_struct *p, int policy, |
3650 | struct sched_param *param) | |
1da177e4 LT |
3651 | { |
3652 | int retval; | |
3653 | int oldprio, oldpolicy = -1; | |
3654 | prio_array_t *array; | |
3655 | unsigned long flags; | |
3656 | runqueue_t *rq; | |
3657 | ||
3658 | recheck: | |
3659 | /* double check policy once rq lock held */ | |
3660 | if (policy < 0) | |
3661 | policy = oldpolicy = p->policy; | |
3662 | else if (policy != SCHED_FIFO && policy != SCHED_RR && | |
b0a9499c IM |
3663 | policy != SCHED_NORMAL && policy != SCHED_BATCH) |
3664 | return -EINVAL; | |
1da177e4 LT |
3665 | /* |
3666 | * Valid priorities for SCHED_FIFO and SCHED_RR are | |
b0a9499c IM |
3667 | * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and |
3668 | * SCHED_BATCH is 0. | |
1da177e4 LT |
3669 | */ |
3670 | if (param->sched_priority < 0 || | |
95cdf3b7 | 3671 | (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || |
d46523ea | 3672 | (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) |
1da177e4 | 3673 | return -EINVAL; |
b0a9499c IM |
3674 | if ((policy == SCHED_NORMAL || policy == SCHED_BATCH) |
3675 | != (param->sched_priority == 0)) | |
1da177e4 LT |
3676 | return -EINVAL; |
3677 | ||
37e4ab3f OC |
3678 | /* |
3679 | * Allow unprivileged RT tasks to decrease priority: | |
3680 | */ | |
3681 | if (!capable(CAP_SYS_NICE)) { | |
b0a9499c IM |
3682 | /* |
3683 | * can't change policy, except between SCHED_NORMAL | |
3684 | * and SCHED_BATCH: | |
3685 | */ | |
3686 | if (((policy != SCHED_NORMAL && p->policy != SCHED_BATCH) && | |
3687 | (policy != SCHED_BATCH && p->policy != SCHED_NORMAL)) && | |
3688 | !p->signal->rlim[RLIMIT_RTPRIO].rlim_cur) | |
37e4ab3f OC |
3689 | return -EPERM; |
3690 | /* can't increase priority */ | |
b0a9499c | 3691 | if ((policy != SCHED_NORMAL && policy != SCHED_BATCH) && |
37e4ab3f OC |
3692 | param->sched_priority > p->rt_priority && |
3693 | param->sched_priority > | |
3694 | p->signal->rlim[RLIMIT_RTPRIO].rlim_cur) | |
3695 | return -EPERM; | |
3696 | /* can't change other user's priorities */ | |
3697 | if ((current->euid != p->euid) && | |
3698 | (current->euid != p->uid)) | |
3699 | return -EPERM; | |
3700 | } | |
1da177e4 LT |
3701 | |
3702 | retval = security_task_setscheduler(p, policy, param); | |
3703 | if (retval) | |
3704 | return retval; | |
3705 | /* | |
3706 | * To be able to change p->policy safely, the apropriate | |
3707 | * runqueue lock must be held. | |
3708 | */ | |
3709 | rq = task_rq_lock(p, &flags); | |
3710 | /* recheck policy now with rq lock held */ | |
3711 | if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { | |
3712 | policy = oldpolicy = -1; | |
3713 | task_rq_unlock(rq, &flags); | |
3714 | goto recheck; | |
3715 | } | |
3716 | array = p->array; | |
3717 | if (array) | |
3718 | deactivate_task(p, rq); | |
3719 | oldprio = p->prio; | |
3720 | __setscheduler(p, policy, param->sched_priority); | |
3721 | if (array) { | |
3722 | __activate_task(p, rq); | |
3723 | /* | |
3724 | * Reschedule if we are currently running on this runqueue and | |
3725 | * our priority decreased, or if we are not currently running on | |
3726 | * this runqueue and our priority is higher than the current's | |
3727 | */ | |
3728 | if (task_running(rq, p)) { | |
3729 | if (p->prio > oldprio) | |
3730 | resched_task(rq->curr); | |
3731 | } else if (TASK_PREEMPTS_CURR(p, rq)) | |
3732 | resched_task(rq->curr); | |
3733 | } | |
3734 | task_rq_unlock(rq, &flags); | |
3735 | return 0; | |
3736 | } | |
3737 | EXPORT_SYMBOL_GPL(sched_setscheduler); | |
3738 | ||
95cdf3b7 IM |
3739 | static int |
3740 | do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) | |
1da177e4 LT |
3741 | { |
3742 | int retval; | |
3743 | struct sched_param lparam; | |
3744 | struct task_struct *p; | |
3745 | ||
3746 | if (!param || pid < 0) | |
3747 | return -EINVAL; | |
3748 | if (copy_from_user(&lparam, param, sizeof(struct sched_param))) | |
3749 | return -EFAULT; | |
3750 | read_lock_irq(&tasklist_lock); | |
3751 | p = find_process_by_pid(pid); | |
3752 | if (!p) { | |
3753 | read_unlock_irq(&tasklist_lock); | |
3754 | return -ESRCH; | |
3755 | } | |
3756 | retval = sched_setscheduler(p, policy, &lparam); | |
3757 | read_unlock_irq(&tasklist_lock); | |
3758 | return retval; | |
3759 | } | |
3760 | ||
3761 | /** | |
3762 | * sys_sched_setscheduler - set/change the scheduler policy and RT priority | |
3763 | * @pid: the pid in question. | |
3764 | * @policy: new policy. | |
3765 | * @param: structure containing the new RT priority. | |
3766 | */ | |
3767 | asmlinkage long sys_sched_setscheduler(pid_t pid, int policy, | |
3768 | struct sched_param __user *param) | |
3769 | { | |
c21761f1 JB |
3770 | /* negative values for policy are not valid */ |
3771 | if (policy < 0) | |
3772 | return -EINVAL; | |
3773 | ||
1da177e4 LT |
3774 | return do_sched_setscheduler(pid, policy, param); |
3775 | } | |
3776 | ||
3777 | /** | |
3778 | * sys_sched_setparam - set/change the RT priority of a thread | |
3779 | * @pid: the pid in question. | |
3780 | * @param: structure containing the new RT priority. | |
3781 | */ | |
3782 | asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param) | |
3783 | { | |
3784 | return do_sched_setscheduler(pid, -1, param); | |
3785 | } | |
3786 | ||
3787 | /** | |
3788 | * sys_sched_getscheduler - get the policy (scheduling class) of a thread | |
3789 | * @pid: the pid in question. | |
3790 | */ | |
3791 | asmlinkage long sys_sched_getscheduler(pid_t pid) | |
3792 | { | |
3793 | int retval = -EINVAL; | |
3794 | task_t *p; | |
3795 | ||
3796 | if (pid < 0) | |
3797 | goto out_nounlock; | |
3798 | ||
3799 | retval = -ESRCH; | |
3800 | read_lock(&tasklist_lock); | |
3801 | p = find_process_by_pid(pid); | |
3802 | if (p) { | |
3803 | retval = security_task_getscheduler(p); | |
3804 | if (!retval) | |
3805 | retval = p->policy; | |
3806 | } | |
3807 | read_unlock(&tasklist_lock); | |
3808 | ||
3809 | out_nounlock: | |
3810 | return retval; | |
3811 | } | |
3812 | ||
3813 | /** | |
3814 | * sys_sched_getscheduler - get the RT priority of a thread | |
3815 | * @pid: the pid in question. | |
3816 | * @param: structure containing the RT priority. | |
3817 | */ | |
3818 | asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param) | |
3819 | { | |
3820 | struct sched_param lp; | |
3821 | int retval = -EINVAL; | |
3822 | task_t *p; | |
3823 | ||
3824 | if (!param || pid < 0) | |
3825 | goto out_nounlock; | |
3826 | ||
3827 | read_lock(&tasklist_lock); | |
3828 | p = find_process_by_pid(pid); | |
3829 | retval = -ESRCH; | |
3830 | if (!p) | |
3831 | goto out_unlock; | |
3832 | ||
3833 | retval = security_task_getscheduler(p); | |
3834 | if (retval) | |
3835 | goto out_unlock; | |
3836 | ||
3837 | lp.sched_priority = p->rt_priority; | |
3838 | read_unlock(&tasklist_lock); | |
3839 | ||
3840 | /* | |
3841 | * This one might sleep, we cannot do it with a spinlock held ... | |
3842 | */ | |
3843 | retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; | |
3844 | ||
3845 | out_nounlock: | |
3846 | return retval; | |
3847 | ||
3848 | out_unlock: | |
3849 | read_unlock(&tasklist_lock); | |
3850 | return retval; | |
3851 | } | |
3852 | ||
3853 | long sched_setaffinity(pid_t pid, cpumask_t new_mask) | |
3854 | { | |
3855 | task_t *p; | |
3856 | int retval; | |
3857 | cpumask_t cpus_allowed; | |
3858 | ||
3859 | lock_cpu_hotplug(); | |
3860 | read_lock(&tasklist_lock); | |
3861 | ||
3862 | p = find_process_by_pid(pid); | |
3863 | if (!p) { | |
3864 | read_unlock(&tasklist_lock); | |
3865 | unlock_cpu_hotplug(); | |
3866 | return -ESRCH; | |
3867 | } | |
3868 | ||
3869 | /* | |
3870 | * It is not safe to call set_cpus_allowed with the | |
3871 | * tasklist_lock held. We will bump the task_struct's | |
3872 | * usage count and then drop tasklist_lock. | |
3873 | */ | |
3874 | get_task_struct(p); | |
3875 | read_unlock(&tasklist_lock); | |
3876 | ||
3877 | retval = -EPERM; | |
3878 | if ((current->euid != p->euid) && (current->euid != p->uid) && | |
3879 | !capable(CAP_SYS_NICE)) | |
3880 | goto out_unlock; | |
3881 | ||
3882 | cpus_allowed = cpuset_cpus_allowed(p); | |
3883 | cpus_and(new_mask, new_mask, cpus_allowed); | |
3884 | retval = set_cpus_allowed(p, new_mask); | |
3885 | ||
3886 | out_unlock: | |
3887 | put_task_struct(p); | |
3888 | unlock_cpu_hotplug(); | |
3889 | return retval; | |
3890 | } | |
3891 | ||
3892 | static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, | |
3893 | cpumask_t *new_mask) | |
3894 | { | |
3895 | if (len < sizeof(cpumask_t)) { | |
3896 | memset(new_mask, 0, sizeof(cpumask_t)); | |
3897 | } else if (len > sizeof(cpumask_t)) { | |
3898 | len = sizeof(cpumask_t); | |
3899 | } | |
3900 | return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; | |
3901 | } | |
3902 | ||
3903 | /** | |
3904 | * sys_sched_setaffinity - set the cpu affinity of a process | |
3905 | * @pid: pid of the process | |
3906 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
3907 | * @user_mask_ptr: user-space pointer to the new cpu mask | |
3908 | */ | |
3909 | asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len, | |
3910 | unsigned long __user *user_mask_ptr) | |
3911 | { | |
3912 | cpumask_t new_mask; | |
3913 | int retval; | |
3914 | ||
3915 | retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask); | |
3916 | if (retval) | |
3917 | return retval; | |
3918 | ||
3919 | return sched_setaffinity(pid, new_mask); | |
3920 | } | |
3921 | ||
3922 | /* | |
3923 | * Represents all cpu's present in the system | |
3924 | * In systems capable of hotplug, this map could dynamically grow | |
3925 | * as new cpu's are detected in the system via any platform specific | |
3926 | * method, such as ACPI for e.g. | |
3927 | */ | |
3928 | ||
4cef0c61 | 3929 | cpumask_t cpu_present_map __read_mostly; |
1da177e4 LT |
3930 | EXPORT_SYMBOL(cpu_present_map); |
3931 | ||
3932 | #ifndef CONFIG_SMP | |
4cef0c61 AK |
3933 | cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL; |
3934 | cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL; | |
1da177e4 LT |
3935 | #endif |
3936 | ||
3937 | long sched_getaffinity(pid_t pid, cpumask_t *mask) | |
3938 | { | |
3939 | int retval; | |
3940 | task_t *p; | |
3941 | ||
3942 | lock_cpu_hotplug(); | |
3943 | read_lock(&tasklist_lock); | |
3944 | ||
3945 | retval = -ESRCH; | |
3946 | p = find_process_by_pid(pid); | |
3947 | if (!p) | |
3948 | goto out_unlock; | |
3949 | ||
3950 | retval = 0; | |
2f7016d9 | 3951 | cpus_and(*mask, p->cpus_allowed, cpu_online_map); |
1da177e4 LT |
3952 | |
3953 | out_unlock: | |
3954 | read_unlock(&tasklist_lock); | |
3955 | unlock_cpu_hotplug(); | |
3956 | if (retval) | |
3957 | return retval; | |
3958 | ||
3959 | return 0; | |
3960 | } | |
3961 | ||
3962 | /** | |
3963 | * sys_sched_getaffinity - get the cpu affinity of a process | |
3964 | * @pid: pid of the process | |
3965 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
3966 | * @user_mask_ptr: user-space pointer to hold the current cpu mask | |
3967 | */ | |
3968 | asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len, | |
3969 | unsigned long __user *user_mask_ptr) | |
3970 | { | |
3971 | int ret; | |
3972 | cpumask_t mask; | |
3973 | ||
3974 | if (len < sizeof(cpumask_t)) | |
3975 | return -EINVAL; | |
3976 | ||
3977 | ret = sched_getaffinity(pid, &mask); | |
3978 | if (ret < 0) | |
3979 | return ret; | |
3980 | ||
3981 | if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t))) | |
3982 | return -EFAULT; | |
3983 | ||
3984 | return sizeof(cpumask_t); | |
3985 | } | |
3986 | ||
3987 | /** | |
3988 | * sys_sched_yield - yield the current processor to other threads. | |
3989 | * | |
3990 | * this function yields the current CPU by moving the calling thread | |
3991 | * to the expired array. If there are no other threads running on this | |
3992 | * CPU then this function will return. | |
3993 | */ | |
3994 | asmlinkage long sys_sched_yield(void) | |
3995 | { | |
3996 | runqueue_t *rq = this_rq_lock(); | |
3997 | prio_array_t *array = current->array; | |
3998 | prio_array_t *target = rq->expired; | |
3999 | ||
4000 | schedstat_inc(rq, yld_cnt); | |
4001 | /* | |
4002 | * We implement yielding by moving the task into the expired | |
4003 | * queue. | |
4004 | * | |
4005 | * (special rule: RT tasks will just roundrobin in the active | |
4006 | * array.) | |
4007 | */ | |
4008 | if (rt_task(current)) | |
4009 | target = rq->active; | |
4010 | ||
5927ad78 | 4011 | if (array->nr_active == 1) { |
1da177e4 LT |
4012 | schedstat_inc(rq, yld_act_empty); |
4013 | if (!rq->expired->nr_active) | |
4014 | schedstat_inc(rq, yld_both_empty); | |
4015 | } else if (!rq->expired->nr_active) | |
4016 | schedstat_inc(rq, yld_exp_empty); | |
4017 | ||
4018 | if (array != target) { | |
4019 | dequeue_task(current, array); | |
4020 | enqueue_task(current, target); | |
4021 | } else | |
4022 | /* | |
4023 | * requeue_task is cheaper so perform that if possible. | |
4024 | */ | |
4025 | requeue_task(current, array); | |
4026 | ||
4027 | /* | |
4028 | * Since we are going to call schedule() anyway, there's | |
4029 | * no need to preempt or enable interrupts: | |
4030 | */ | |
4031 | __release(rq->lock); | |
4032 | _raw_spin_unlock(&rq->lock); | |
4033 | preempt_enable_no_resched(); | |
4034 | ||
4035 | schedule(); | |
4036 | ||
4037 | return 0; | |
4038 | } | |
4039 | ||
4040 | static inline void __cond_resched(void) | |
4041 | { | |
5bbcfd90 IM |
4042 | /* |
4043 | * The BKS might be reacquired before we have dropped | |
4044 | * PREEMPT_ACTIVE, which could trigger a second | |
4045 | * cond_resched() call. | |
4046 | */ | |
4047 | if (unlikely(preempt_count())) | |
4048 | return; | |
8ba7b0a1 LT |
4049 | if (unlikely(system_state != SYSTEM_RUNNING)) |
4050 | return; | |
1da177e4 LT |
4051 | do { |
4052 | add_preempt_count(PREEMPT_ACTIVE); | |
4053 | schedule(); | |
4054 | sub_preempt_count(PREEMPT_ACTIVE); | |
4055 | } while (need_resched()); | |
4056 | } | |
4057 | ||
4058 | int __sched cond_resched(void) | |
4059 | { | |
4060 | if (need_resched()) { | |
4061 | __cond_resched(); | |
4062 | return 1; | |
4063 | } | |
4064 | return 0; | |
4065 | } | |
4066 | ||
4067 | EXPORT_SYMBOL(cond_resched); | |
4068 | ||
4069 | /* | |
4070 | * cond_resched_lock() - if a reschedule is pending, drop the given lock, | |
4071 | * call schedule, and on return reacquire the lock. | |
4072 | * | |
4073 | * This works OK both with and without CONFIG_PREEMPT. We do strange low-level | |
4074 | * operations here to prevent schedule() from being called twice (once via | |
4075 | * spin_unlock(), once by hand). | |
4076 | */ | |
95cdf3b7 | 4077 | int cond_resched_lock(spinlock_t *lock) |
1da177e4 | 4078 | { |
6df3cecb JK |
4079 | int ret = 0; |
4080 | ||
1da177e4 LT |
4081 | if (need_lockbreak(lock)) { |
4082 | spin_unlock(lock); | |
4083 | cpu_relax(); | |
6df3cecb | 4084 | ret = 1; |
1da177e4 LT |
4085 | spin_lock(lock); |
4086 | } | |
4087 | if (need_resched()) { | |
4088 | _raw_spin_unlock(lock); | |
4089 | preempt_enable_no_resched(); | |
4090 | __cond_resched(); | |
6df3cecb | 4091 | ret = 1; |
1da177e4 | 4092 | spin_lock(lock); |
1da177e4 | 4093 | } |
6df3cecb | 4094 | return ret; |
1da177e4 LT |
4095 | } |
4096 | ||
4097 | EXPORT_SYMBOL(cond_resched_lock); | |
4098 | ||
4099 | int __sched cond_resched_softirq(void) | |
4100 | { | |
4101 | BUG_ON(!in_softirq()); | |
4102 | ||
4103 | if (need_resched()) { | |
4104 | __local_bh_enable(); | |
4105 | __cond_resched(); | |
4106 | local_bh_disable(); | |
4107 | return 1; | |
4108 | } | |
4109 | return 0; | |
4110 | } | |
4111 | ||
4112 | EXPORT_SYMBOL(cond_resched_softirq); | |
4113 | ||
4114 | ||
4115 | /** | |
4116 | * yield - yield the current processor to other threads. | |
4117 | * | |
4118 | * this is a shortcut for kernel-space yielding - it marks the | |
4119 | * thread runnable and calls sys_sched_yield(). | |
4120 | */ | |
4121 | void __sched yield(void) | |
4122 | { | |
4123 | set_current_state(TASK_RUNNING); | |
4124 | sys_sched_yield(); | |
4125 | } | |
4126 | ||
4127 | EXPORT_SYMBOL(yield); | |
4128 | ||
4129 | /* | |
4130 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so | |
4131 | * that process accounting knows that this is a task in IO wait state. | |
4132 | * | |
4133 | * But don't do that if it is a deliberate, throttling IO wait (this task | |
4134 | * has set its backing_dev_info: the queue against which it should throttle) | |
4135 | */ | |
4136 | void __sched io_schedule(void) | |
4137 | { | |
39c715b7 | 4138 | struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id()); |
1da177e4 LT |
4139 | |
4140 | atomic_inc(&rq->nr_iowait); | |
4141 | schedule(); | |
4142 | atomic_dec(&rq->nr_iowait); | |
4143 | } | |
4144 | ||
4145 | EXPORT_SYMBOL(io_schedule); | |
4146 | ||
4147 | long __sched io_schedule_timeout(long timeout) | |
4148 | { | |
39c715b7 | 4149 | struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id()); |
1da177e4 LT |
4150 | long ret; |
4151 | ||
4152 | atomic_inc(&rq->nr_iowait); | |
4153 | ret = schedule_timeout(timeout); | |
4154 | atomic_dec(&rq->nr_iowait); | |
4155 | return ret; | |
4156 | } | |
4157 | ||
4158 | /** | |
4159 | * sys_sched_get_priority_max - return maximum RT priority. | |
4160 | * @policy: scheduling class. | |
4161 | * | |
4162 | * this syscall returns the maximum rt_priority that can be used | |
4163 | * by a given scheduling class. | |
4164 | */ | |
4165 | asmlinkage long sys_sched_get_priority_max(int policy) | |
4166 | { | |
4167 | int ret = -EINVAL; | |
4168 | ||
4169 | switch (policy) { | |
4170 | case SCHED_FIFO: | |
4171 | case SCHED_RR: | |
4172 | ret = MAX_USER_RT_PRIO-1; | |
4173 | break; | |
4174 | case SCHED_NORMAL: | |
b0a9499c | 4175 | case SCHED_BATCH: |
1da177e4 LT |
4176 | ret = 0; |
4177 | break; | |
4178 | } | |
4179 | return ret; | |
4180 | } | |
4181 | ||
4182 | /** | |
4183 | * sys_sched_get_priority_min - return minimum RT priority. | |
4184 | * @policy: scheduling class. | |
4185 | * | |
4186 | * this syscall returns the minimum rt_priority that can be used | |
4187 | * by a given scheduling class. | |
4188 | */ | |
4189 | asmlinkage long sys_sched_get_priority_min(int policy) | |
4190 | { | |
4191 | int ret = -EINVAL; | |
4192 | ||
4193 | switch (policy) { | |
4194 | case SCHED_FIFO: | |
4195 | case SCHED_RR: | |
4196 | ret = 1; | |
4197 | break; | |
4198 | case SCHED_NORMAL: | |
b0a9499c | 4199 | case SCHED_BATCH: |
1da177e4 LT |
4200 | ret = 0; |
4201 | } | |
4202 | return ret; | |
4203 | } | |
4204 | ||
4205 | /** | |
4206 | * sys_sched_rr_get_interval - return the default timeslice of a process. | |
4207 | * @pid: pid of the process. | |
4208 | * @interval: userspace pointer to the timeslice value. | |
4209 | * | |
4210 | * this syscall writes the default timeslice value of a given process | |
4211 | * into the user-space timespec buffer. A value of '0' means infinity. | |
4212 | */ | |
4213 | asmlinkage | |
4214 | long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval) | |
4215 | { | |
4216 | int retval = -EINVAL; | |
4217 | struct timespec t; | |
4218 | task_t *p; | |
4219 | ||
4220 | if (pid < 0) | |
4221 | goto out_nounlock; | |
4222 | ||
4223 | retval = -ESRCH; | |
4224 | read_lock(&tasklist_lock); | |
4225 | p = find_process_by_pid(pid); | |
4226 | if (!p) | |
4227 | goto out_unlock; | |
4228 | ||
4229 | retval = security_task_getscheduler(p); | |
4230 | if (retval) | |
4231 | goto out_unlock; | |
4232 | ||
4233 | jiffies_to_timespec(p->policy & SCHED_FIFO ? | |
4234 | 0 : task_timeslice(p), &t); | |
4235 | read_unlock(&tasklist_lock); | |
4236 | retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; | |
4237 | out_nounlock: | |
4238 | return retval; | |
4239 | out_unlock: | |
4240 | read_unlock(&tasklist_lock); | |
4241 | return retval; | |
4242 | } | |
4243 | ||
4244 | static inline struct task_struct *eldest_child(struct task_struct *p) | |
4245 | { | |
4246 | if (list_empty(&p->children)) return NULL; | |
4247 | return list_entry(p->children.next,struct task_struct,sibling); | |
4248 | } | |
4249 | ||
4250 | static inline struct task_struct *older_sibling(struct task_struct *p) | |
4251 | { | |
4252 | if (p->sibling.prev==&p->parent->children) return NULL; | |
4253 | return list_entry(p->sibling.prev,struct task_struct,sibling); | |
4254 | } | |
4255 | ||
4256 | static inline struct task_struct *younger_sibling(struct task_struct *p) | |
4257 | { | |
4258 | if (p->sibling.next==&p->parent->children) return NULL; | |
4259 | return list_entry(p->sibling.next,struct task_struct,sibling); | |
4260 | } | |
4261 | ||
95cdf3b7 | 4262 | static void show_task(task_t *p) |
1da177e4 LT |
4263 | { |
4264 | task_t *relative; | |
4265 | unsigned state; | |
4266 | unsigned long free = 0; | |
4267 | static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" }; | |
4268 | ||
4269 | printk("%-13.13s ", p->comm); | |
4270 | state = p->state ? __ffs(p->state) + 1 : 0; | |
4271 | if (state < ARRAY_SIZE(stat_nam)) | |
4272 | printk(stat_nam[state]); | |
4273 | else | |
4274 | printk("?"); | |
4275 | #if (BITS_PER_LONG == 32) | |
4276 | if (state == TASK_RUNNING) | |
4277 | printk(" running "); | |
4278 | else | |
4279 | printk(" %08lX ", thread_saved_pc(p)); | |
4280 | #else | |
4281 | if (state == TASK_RUNNING) | |
4282 | printk(" running task "); | |
4283 | else | |
4284 | printk(" %016lx ", thread_saved_pc(p)); | |
4285 | #endif | |
4286 | #ifdef CONFIG_DEBUG_STACK_USAGE | |
4287 | { | |
10ebffde | 4288 | unsigned long *n = end_of_stack(p); |
1da177e4 LT |
4289 | while (!*n) |
4290 | n++; | |
10ebffde | 4291 | free = (unsigned long)n - (unsigned long)end_of_stack(p); |
1da177e4 LT |
4292 | } |
4293 | #endif | |
4294 | printk("%5lu %5d %6d ", free, p->pid, p->parent->pid); | |
4295 | if ((relative = eldest_child(p))) | |
4296 | printk("%5d ", relative->pid); | |
4297 | else | |
4298 | printk(" "); | |
4299 | if ((relative = younger_sibling(p))) | |
4300 | printk("%7d", relative->pid); | |
4301 | else | |
4302 | printk(" "); | |
4303 | if ((relative = older_sibling(p))) | |
4304 | printk(" %5d", relative->pid); | |
4305 | else | |
4306 | printk(" "); | |
4307 | if (!p->mm) | |
4308 | printk(" (L-TLB)\n"); | |
4309 | else | |
4310 | printk(" (NOTLB)\n"); | |
4311 | ||
4312 | if (state != TASK_RUNNING) | |
4313 | show_stack(p, NULL); | |
4314 | } | |
4315 | ||
4316 | void show_state(void) | |
4317 | { | |
4318 | task_t *g, *p; | |
4319 | ||
4320 | #if (BITS_PER_LONG == 32) | |
4321 | printk("\n" | |
4322 | " sibling\n"); | |
4323 | printk(" task PC pid father child younger older\n"); | |
4324 | #else | |
4325 | printk("\n" | |
4326 | " sibling\n"); | |
4327 | printk(" task PC pid father child younger older\n"); | |
4328 | #endif | |
4329 | read_lock(&tasklist_lock); | |
4330 | do_each_thread(g, p) { | |
4331 | /* | |
4332 | * reset the NMI-timeout, listing all files on a slow | |
4333 | * console might take alot of time: | |
4334 | */ | |
4335 | touch_nmi_watchdog(); | |
4336 | show_task(p); | |
4337 | } while_each_thread(g, p); | |
4338 | ||
4339 | read_unlock(&tasklist_lock); | |
de5097c2 | 4340 | mutex_debug_show_all_locks(); |
1da177e4 LT |
4341 | } |
4342 | ||
f340c0d1 IM |
4343 | /** |
4344 | * init_idle - set up an idle thread for a given CPU | |
4345 | * @idle: task in question | |
4346 | * @cpu: cpu the idle task belongs to | |
4347 | * | |
4348 | * NOTE: this function does not set the idle thread's NEED_RESCHED | |
4349 | * flag, to make booting more robust. | |
4350 | */ | |
1da177e4 LT |
4351 | void __devinit init_idle(task_t *idle, int cpu) |
4352 | { | |
4353 | runqueue_t *rq = cpu_rq(cpu); | |
4354 | unsigned long flags; | |
4355 | ||
81c29a85 | 4356 | idle->timestamp = sched_clock(); |
1da177e4 LT |
4357 | idle->sleep_avg = 0; |
4358 | idle->array = NULL; | |
4359 | idle->prio = MAX_PRIO; | |
4360 | idle->state = TASK_RUNNING; | |
4361 | idle->cpus_allowed = cpumask_of_cpu(cpu); | |
4362 | set_task_cpu(idle, cpu); | |
4363 | ||
4364 | spin_lock_irqsave(&rq->lock, flags); | |
4365 | rq->curr = rq->idle = idle; | |
4866cde0 NP |
4366 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
4367 | idle->oncpu = 1; | |
4368 | #endif | |
1da177e4 LT |
4369 | spin_unlock_irqrestore(&rq->lock, flags); |
4370 | ||
4371 | /* Set the preempt count _outside_ the spinlocks! */ | |
4372 | #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL) | |
a1261f54 | 4373 | task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0); |
1da177e4 | 4374 | #else |
a1261f54 | 4375 | task_thread_info(idle)->preempt_count = 0; |
1da177e4 LT |
4376 | #endif |
4377 | } | |
4378 | ||
4379 | /* | |
4380 | * In a system that switches off the HZ timer nohz_cpu_mask | |
4381 | * indicates which cpus entered this state. This is used | |
4382 | * in the rcu update to wait only for active cpus. For system | |
4383 | * which do not switch off the HZ timer nohz_cpu_mask should | |
4384 | * always be CPU_MASK_NONE. | |
4385 | */ | |
4386 | cpumask_t nohz_cpu_mask = CPU_MASK_NONE; | |
4387 | ||
4388 | #ifdef CONFIG_SMP | |
4389 | /* | |
4390 | * This is how migration works: | |
4391 | * | |
4392 | * 1) we queue a migration_req_t structure in the source CPU's | |
4393 | * runqueue and wake up that CPU's migration thread. | |
4394 | * 2) we down() the locked semaphore => thread blocks. | |
4395 | * 3) migration thread wakes up (implicitly it forces the migrated | |
4396 | * thread off the CPU) | |
4397 | * 4) it gets the migration request and checks whether the migrated | |
4398 | * task is still in the wrong runqueue. | |
4399 | * 5) if it's in the wrong runqueue then the migration thread removes | |
4400 | * it and puts it into the right queue. | |
4401 | * 6) migration thread up()s the semaphore. | |
4402 | * 7) we wake up and the migration is done. | |
4403 | */ | |
4404 | ||
4405 | /* | |
4406 | * Change a given task's CPU affinity. Migrate the thread to a | |
4407 | * proper CPU and schedule it away if the CPU it's executing on | |
4408 | * is removed from the allowed bitmask. | |
4409 | * | |
4410 | * NOTE: the caller must have a valid reference to the task, the | |
4411 | * task must not exit() & deallocate itself prematurely. The | |
4412 | * call is not atomic; no spinlocks may be held. | |
4413 | */ | |
4414 | int set_cpus_allowed(task_t *p, cpumask_t new_mask) | |
4415 | { | |
4416 | unsigned long flags; | |
4417 | int ret = 0; | |
4418 | migration_req_t req; | |
4419 | runqueue_t *rq; | |
4420 | ||
4421 | rq = task_rq_lock(p, &flags); | |
4422 | if (!cpus_intersects(new_mask, cpu_online_map)) { | |
4423 | ret = -EINVAL; | |
4424 | goto out; | |
4425 | } | |
4426 | ||
4427 | p->cpus_allowed = new_mask; | |
4428 | /* Can the task run on the task's current CPU? If so, we're done */ | |
4429 | if (cpu_isset(task_cpu(p), new_mask)) | |
4430 | goto out; | |
4431 | ||
4432 | if (migrate_task(p, any_online_cpu(new_mask), &req)) { | |
4433 | /* Need help from migration thread: drop lock and wait. */ | |
4434 | task_rq_unlock(rq, &flags); | |
4435 | wake_up_process(rq->migration_thread); | |
4436 | wait_for_completion(&req.done); | |
4437 | tlb_migrate_finish(p->mm); | |
4438 | return 0; | |
4439 | } | |
4440 | out: | |
4441 | task_rq_unlock(rq, &flags); | |
4442 | return ret; | |
4443 | } | |
4444 | ||
4445 | EXPORT_SYMBOL_GPL(set_cpus_allowed); | |
4446 | ||
4447 | /* | |
4448 | * Move (not current) task off this cpu, onto dest cpu. We're doing | |
4449 | * this because either it can't run here any more (set_cpus_allowed() | |
4450 | * away from this CPU, or CPU going down), or because we're | |
4451 | * attempting to rebalance this task on exec (sched_exec). | |
4452 | * | |
4453 | * So we race with normal scheduler movements, but that's OK, as long | |
4454 | * as the task is no longer on this CPU. | |
4455 | */ | |
4456 | static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) | |
4457 | { | |
4458 | runqueue_t *rq_dest, *rq_src; | |
4459 | ||
4460 | if (unlikely(cpu_is_offline(dest_cpu))) | |
4461 | return; | |
4462 | ||
4463 | rq_src = cpu_rq(src_cpu); | |
4464 | rq_dest = cpu_rq(dest_cpu); | |
4465 | ||
4466 | double_rq_lock(rq_src, rq_dest); | |
4467 | /* Already moved. */ | |
4468 | if (task_cpu(p) != src_cpu) | |
4469 | goto out; | |
4470 | /* Affinity changed (again). */ | |
4471 | if (!cpu_isset(dest_cpu, p->cpus_allowed)) | |
4472 | goto out; | |
4473 | ||
4474 | set_task_cpu(p, dest_cpu); | |
4475 | if (p->array) { | |
4476 | /* | |
4477 | * Sync timestamp with rq_dest's before activating. | |
4478 | * The same thing could be achieved by doing this step | |
4479 | * afterwards, and pretending it was a local activate. | |
4480 | * This way is cleaner and logically correct. | |
4481 | */ | |
4482 | p->timestamp = p->timestamp - rq_src->timestamp_last_tick | |
4483 | + rq_dest->timestamp_last_tick; | |
4484 | deactivate_task(p, rq_src); | |
4485 | activate_task(p, rq_dest, 0); | |
4486 | if (TASK_PREEMPTS_CURR(p, rq_dest)) | |
4487 | resched_task(rq_dest->curr); | |
4488 | } | |
4489 | ||
4490 | out: | |
4491 | double_rq_unlock(rq_src, rq_dest); | |
4492 | } | |
4493 | ||
4494 | /* | |
4495 | * migration_thread - this is a highprio system thread that performs | |
4496 | * thread migration by bumping thread off CPU then 'pushing' onto | |
4497 | * another runqueue. | |
4498 | */ | |
95cdf3b7 | 4499 | static int migration_thread(void *data) |
1da177e4 LT |
4500 | { |
4501 | runqueue_t *rq; | |
4502 | int cpu = (long)data; | |
4503 | ||
4504 | rq = cpu_rq(cpu); | |
4505 | BUG_ON(rq->migration_thread != current); | |
4506 | ||
4507 | set_current_state(TASK_INTERRUPTIBLE); | |
4508 | while (!kthread_should_stop()) { | |
4509 | struct list_head *head; | |
4510 | migration_req_t *req; | |
4511 | ||
3e1d1d28 | 4512 | try_to_freeze(); |
1da177e4 LT |
4513 | |
4514 | spin_lock_irq(&rq->lock); | |
4515 | ||
4516 | if (cpu_is_offline(cpu)) { | |
4517 | spin_unlock_irq(&rq->lock); | |
4518 | goto wait_to_die; | |
4519 | } | |
4520 | ||
4521 | if (rq->active_balance) { | |
4522 | active_load_balance(rq, cpu); | |
4523 | rq->active_balance = 0; | |
4524 | } | |
4525 | ||
4526 | head = &rq->migration_queue; | |
4527 | ||
4528 | if (list_empty(head)) { | |
4529 | spin_unlock_irq(&rq->lock); | |
4530 | schedule(); | |
4531 | set_current_state(TASK_INTERRUPTIBLE); | |
4532 | continue; | |
4533 | } | |
4534 | req = list_entry(head->next, migration_req_t, list); | |
4535 | list_del_init(head->next); | |
4536 | ||
674311d5 NP |
4537 | spin_unlock(&rq->lock); |
4538 | __migrate_task(req->task, cpu, req->dest_cpu); | |
4539 | local_irq_enable(); | |
1da177e4 LT |
4540 | |
4541 | complete(&req->done); | |
4542 | } | |
4543 | __set_current_state(TASK_RUNNING); | |
4544 | return 0; | |
4545 | ||
4546 | wait_to_die: | |
4547 | /* Wait for kthread_stop */ | |
4548 | set_current_state(TASK_INTERRUPTIBLE); | |
4549 | while (!kthread_should_stop()) { | |
4550 | schedule(); | |
4551 | set_current_state(TASK_INTERRUPTIBLE); | |
4552 | } | |
4553 | __set_current_state(TASK_RUNNING); | |
4554 | return 0; | |
4555 | } | |
4556 | ||
4557 | #ifdef CONFIG_HOTPLUG_CPU | |
4558 | /* Figure out where task on dead CPU should go, use force if neccessary. */ | |
4559 | static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk) | |
4560 | { | |
4561 | int dest_cpu; | |
4562 | cpumask_t mask; | |
4563 | ||
4564 | /* On same node? */ | |
4565 | mask = node_to_cpumask(cpu_to_node(dead_cpu)); | |
4566 | cpus_and(mask, mask, tsk->cpus_allowed); | |
4567 | dest_cpu = any_online_cpu(mask); | |
4568 | ||
4569 | /* On any allowed CPU? */ | |
4570 | if (dest_cpu == NR_CPUS) | |
4571 | dest_cpu = any_online_cpu(tsk->cpus_allowed); | |
4572 | ||
4573 | /* No more Mr. Nice Guy. */ | |
4574 | if (dest_cpu == NR_CPUS) { | |
b39c4fab | 4575 | cpus_setall(tsk->cpus_allowed); |
1da177e4 LT |
4576 | dest_cpu = any_online_cpu(tsk->cpus_allowed); |
4577 | ||
4578 | /* | |
4579 | * Don't tell them about moving exiting tasks or | |
4580 | * kernel threads (both mm NULL), since they never | |
4581 | * leave kernel. | |
4582 | */ | |
4583 | if (tsk->mm && printk_ratelimit()) | |
4584 | printk(KERN_INFO "process %d (%s) no " | |
4585 | "longer affine to cpu%d\n", | |
4586 | tsk->pid, tsk->comm, dead_cpu); | |
4587 | } | |
4588 | __migrate_task(tsk, dead_cpu, dest_cpu); | |
4589 | } | |
4590 | ||
4591 | /* | |
4592 | * While a dead CPU has no uninterruptible tasks queued at this point, | |
4593 | * it might still have a nonzero ->nr_uninterruptible counter, because | |
4594 | * for performance reasons the counter is not stricly tracking tasks to | |
4595 | * their home CPUs. So we just add the counter to another CPU's counter, | |
4596 | * to keep the global sum constant after CPU-down: | |
4597 | */ | |
4598 | static void migrate_nr_uninterruptible(runqueue_t *rq_src) | |
4599 | { | |
4600 | runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL)); | |
4601 | unsigned long flags; | |
4602 | ||
4603 | local_irq_save(flags); | |
4604 | double_rq_lock(rq_src, rq_dest); | |
4605 | rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible; | |
4606 | rq_src->nr_uninterruptible = 0; | |
4607 | double_rq_unlock(rq_src, rq_dest); | |
4608 | local_irq_restore(flags); | |
4609 | } | |
4610 | ||
4611 | /* Run through task list and migrate tasks from the dead cpu. */ | |
4612 | static void migrate_live_tasks(int src_cpu) | |
4613 | { | |
4614 | struct task_struct *tsk, *t; | |
4615 | ||
4616 | write_lock_irq(&tasklist_lock); | |
4617 | ||
4618 | do_each_thread(t, tsk) { | |
4619 | if (tsk == current) | |
4620 | continue; | |
4621 | ||
4622 | if (task_cpu(tsk) == src_cpu) | |
4623 | move_task_off_dead_cpu(src_cpu, tsk); | |
4624 | } while_each_thread(t, tsk); | |
4625 | ||
4626 | write_unlock_irq(&tasklist_lock); | |
4627 | } | |
4628 | ||
4629 | /* Schedules idle task to be the next runnable task on current CPU. | |
4630 | * It does so by boosting its priority to highest possible and adding it to | |
4631 | * the _front_ of runqueue. Used by CPU offline code. | |
4632 | */ | |
4633 | void sched_idle_next(void) | |
4634 | { | |
4635 | int cpu = smp_processor_id(); | |
4636 | runqueue_t *rq = this_rq(); | |
4637 | struct task_struct *p = rq->idle; | |
4638 | unsigned long flags; | |
4639 | ||
4640 | /* cpu has to be offline */ | |
4641 | BUG_ON(cpu_online(cpu)); | |
4642 | ||
4643 | /* Strictly not necessary since rest of the CPUs are stopped by now | |
4644 | * and interrupts disabled on current cpu. | |
4645 | */ | |
4646 | spin_lock_irqsave(&rq->lock, flags); | |
4647 | ||
4648 | __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1); | |
4649 | /* Add idle task to _front_ of it's priority queue */ | |
4650 | __activate_idle_task(p, rq); | |
4651 | ||
4652 | spin_unlock_irqrestore(&rq->lock, flags); | |
4653 | } | |
4654 | ||
4655 | /* Ensures that the idle task is using init_mm right before its cpu goes | |
4656 | * offline. | |
4657 | */ | |
4658 | void idle_task_exit(void) | |
4659 | { | |
4660 | struct mm_struct *mm = current->active_mm; | |
4661 | ||
4662 | BUG_ON(cpu_online(smp_processor_id())); | |
4663 | ||
4664 | if (mm != &init_mm) | |
4665 | switch_mm(mm, &init_mm, current); | |
4666 | mmdrop(mm); | |
4667 | } | |
4668 | ||
4669 | static void migrate_dead(unsigned int dead_cpu, task_t *tsk) | |
4670 | { | |
4671 | struct runqueue *rq = cpu_rq(dead_cpu); | |
4672 | ||
4673 | /* Must be exiting, otherwise would be on tasklist. */ | |
4674 | BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD); | |
4675 | ||
4676 | /* Cannot have done final schedule yet: would have vanished. */ | |
4677 | BUG_ON(tsk->flags & PF_DEAD); | |
4678 | ||
4679 | get_task_struct(tsk); | |
4680 | ||
4681 | /* | |
4682 | * Drop lock around migration; if someone else moves it, | |
4683 | * that's OK. No task can be added to this CPU, so iteration is | |
4684 | * fine. | |
4685 | */ | |
4686 | spin_unlock_irq(&rq->lock); | |
4687 | move_task_off_dead_cpu(dead_cpu, tsk); | |
4688 | spin_lock_irq(&rq->lock); | |
4689 | ||
4690 | put_task_struct(tsk); | |
4691 | } | |
4692 | ||
4693 | /* release_task() removes task from tasklist, so we won't find dead tasks. */ | |
4694 | static void migrate_dead_tasks(unsigned int dead_cpu) | |
4695 | { | |
4696 | unsigned arr, i; | |
4697 | struct runqueue *rq = cpu_rq(dead_cpu); | |
4698 | ||
4699 | for (arr = 0; arr < 2; arr++) { | |
4700 | for (i = 0; i < MAX_PRIO; i++) { | |
4701 | struct list_head *list = &rq->arrays[arr].queue[i]; | |
4702 | while (!list_empty(list)) | |
4703 | migrate_dead(dead_cpu, | |
4704 | list_entry(list->next, task_t, | |
4705 | run_list)); | |
4706 | } | |
4707 | } | |
4708 | } | |
4709 | #endif /* CONFIG_HOTPLUG_CPU */ | |
4710 | ||
4711 | /* | |
4712 | * migration_call - callback that gets triggered when a CPU is added. | |
4713 | * Here we can start up the necessary migration thread for the new CPU. | |
4714 | */ | |
4715 | static int migration_call(struct notifier_block *nfb, unsigned long action, | |
4716 | void *hcpu) | |
4717 | { | |
4718 | int cpu = (long)hcpu; | |
4719 | struct task_struct *p; | |
4720 | struct runqueue *rq; | |
4721 | unsigned long flags; | |
4722 | ||
4723 | switch (action) { | |
4724 | case CPU_UP_PREPARE: | |
4725 | p = kthread_create(migration_thread, hcpu, "migration/%d",cpu); | |
4726 | if (IS_ERR(p)) | |
4727 | return NOTIFY_BAD; | |
4728 | p->flags |= PF_NOFREEZE; | |
4729 | kthread_bind(p, cpu); | |
4730 | /* Must be high prio: stop_machine expects to yield to it. */ | |
4731 | rq = task_rq_lock(p, &flags); | |
4732 | __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1); | |
4733 | task_rq_unlock(rq, &flags); | |
4734 | cpu_rq(cpu)->migration_thread = p; | |
4735 | break; | |
4736 | case CPU_ONLINE: | |
4737 | /* Strictly unneccessary, as first user will wake it. */ | |
4738 | wake_up_process(cpu_rq(cpu)->migration_thread); | |
4739 | break; | |
4740 | #ifdef CONFIG_HOTPLUG_CPU | |
4741 | case CPU_UP_CANCELED: | |
4742 | /* Unbind it from offline cpu so it can run. Fall thru. */ | |
a4c4af7c HC |
4743 | kthread_bind(cpu_rq(cpu)->migration_thread, |
4744 | any_online_cpu(cpu_online_map)); | |
1da177e4 LT |
4745 | kthread_stop(cpu_rq(cpu)->migration_thread); |
4746 | cpu_rq(cpu)->migration_thread = NULL; | |
4747 | break; | |
4748 | case CPU_DEAD: | |
4749 | migrate_live_tasks(cpu); | |
4750 | rq = cpu_rq(cpu); | |
4751 | kthread_stop(rq->migration_thread); | |
4752 | rq->migration_thread = NULL; | |
4753 | /* Idle task back to normal (off runqueue, low prio) */ | |
4754 | rq = task_rq_lock(rq->idle, &flags); | |
4755 | deactivate_task(rq->idle, rq); | |
4756 | rq->idle->static_prio = MAX_PRIO; | |
4757 | __setscheduler(rq->idle, SCHED_NORMAL, 0); | |
4758 | migrate_dead_tasks(cpu); | |
4759 | task_rq_unlock(rq, &flags); | |
4760 | migrate_nr_uninterruptible(rq); | |
4761 | BUG_ON(rq->nr_running != 0); | |
4762 | ||
4763 | /* No need to migrate the tasks: it was best-effort if | |
4764 | * they didn't do lock_cpu_hotplug(). Just wake up | |
4765 | * the requestors. */ | |
4766 | spin_lock_irq(&rq->lock); | |
4767 | while (!list_empty(&rq->migration_queue)) { | |
4768 | migration_req_t *req; | |
4769 | req = list_entry(rq->migration_queue.next, | |
4770 | migration_req_t, list); | |
1da177e4 LT |
4771 | list_del_init(&req->list); |
4772 | complete(&req->done); | |
4773 | } | |
4774 | spin_unlock_irq(&rq->lock); | |
4775 | break; | |
4776 | #endif | |
4777 | } | |
4778 | return NOTIFY_OK; | |
4779 | } | |
4780 | ||
4781 | /* Register at highest priority so that task migration (migrate_all_tasks) | |
4782 | * happens before everything else. | |
4783 | */ | |
4784 | static struct notifier_block __devinitdata migration_notifier = { | |
4785 | .notifier_call = migration_call, | |
4786 | .priority = 10 | |
4787 | }; | |
4788 | ||
4789 | int __init migration_init(void) | |
4790 | { | |
4791 | void *cpu = (void *)(long)smp_processor_id(); | |
4792 | /* Start one for boot CPU. */ | |
4793 | migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); | |
4794 | migration_call(&migration_notifier, CPU_ONLINE, cpu); | |
4795 | register_cpu_notifier(&migration_notifier); | |
4796 | return 0; | |
4797 | } | |
4798 | #endif | |
4799 | ||
4800 | #ifdef CONFIG_SMP | |
1a20ff27 | 4801 | #undef SCHED_DOMAIN_DEBUG |
1da177e4 LT |
4802 | #ifdef SCHED_DOMAIN_DEBUG |
4803 | static void sched_domain_debug(struct sched_domain *sd, int cpu) | |
4804 | { | |
4805 | int level = 0; | |
4806 | ||
41c7ce9a NP |
4807 | if (!sd) { |
4808 | printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); | |
4809 | return; | |
4810 | } | |
4811 | ||
1da177e4 LT |
4812 | printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); |
4813 | ||
4814 | do { | |
4815 | int i; | |
4816 | char str[NR_CPUS]; | |
4817 | struct sched_group *group = sd->groups; | |
4818 | cpumask_t groupmask; | |
4819 | ||
4820 | cpumask_scnprintf(str, NR_CPUS, sd->span); | |
4821 | cpus_clear(groupmask); | |
4822 | ||
4823 | printk(KERN_DEBUG); | |
4824 | for (i = 0; i < level + 1; i++) | |
4825 | printk(" "); | |
4826 | printk("domain %d: ", level); | |
4827 | ||
4828 | if (!(sd->flags & SD_LOAD_BALANCE)) { | |
4829 | printk("does not load-balance\n"); | |
4830 | if (sd->parent) | |
4831 | printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent"); | |
4832 | break; | |
4833 | } | |
4834 | ||
4835 | printk("span %s\n", str); | |
4836 | ||
4837 | if (!cpu_isset(cpu, sd->span)) | |
4838 | printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu); | |
4839 | if (!cpu_isset(cpu, group->cpumask)) | |
4840 | printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu); | |
4841 | ||
4842 | printk(KERN_DEBUG); | |
4843 | for (i = 0; i < level + 2; i++) | |
4844 | printk(" "); | |
4845 | printk("groups:"); | |
4846 | do { | |
4847 | if (!group) { | |
4848 | printk("\n"); | |
4849 | printk(KERN_ERR "ERROR: group is NULL\n"); | |
4850 | break; | |
4851 | } | |
4852 | ||
4853 | if (!group->cpu_power) { | |
4854 | printk("\n"); | |
4855 | printk(KERN_ERR "ERROR: domain->cpu_power not set\n"); | |
4856 | } | |
4857 | ||
4858 | if (!cpus_weight(group->cpumask)) { | |
4859 | printk("\n"); | |
4860 | printk(KERN_ERR "ERROR: empty group\n"); | |
4861 | } | |
4862 | ||
4863 | if (cpus_intersects(groupmask, group->cpumask)) { | |
4864 | printk("\n"); | |
4865 | printk(KERN_ERR "ERROR: repeated CPUs\n"); | |
4866 | } | |
4867 | ||
4868 | cpus_or(groupmask, groupmask, group->cpumask); | |
4869 | ||
4870 | cpumask_scnprintf(str, NR_CPUS, group->cpumask); | |
4871 | printk(" %s", str); | |
4872 | ||
4873 | group = group->next; | |
4874 | } while (group != sd->groups); | |
4875 | printk("\n"); | |
4876 | ||
4877 | if (!cpus_equal(sd->span, groupmask)) | |
4878 | printk(KERN_ERR "ERROR: groups don't span domain->span\n"); | |
4879 | ||
4880 | level++; | |
4881 | sd = sd->parent; | |
4882 | ||
4883 | if (sd) { | |
4884 | if (!cpus_subset(groupmask, sd->span)) | |
4885 | printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n"); | |
4886 | } | |
4887 | ||
4888 | } while (sd); | |
4889 | } | |
4890 | #else | |
4891 | #define sched_domain_debug(sd, cpu) {} | |
4892 | #endif | |
4893 | ||
1a20ff27 | 4894 | static int sd_degenerate(struct sched_domain *sd) |
245af2c7 SS |
4895 | { |
4896 | if (cpus_weight(sd->span) == 1) | |
4897 | return 1; | |
4898 | ||
4899 | /* Following flags need at least 2 groups */ | |
4900 | if (sd->flags & (SD_LOAD_BALANCE | | |
4901 | SD_BALANCE_NEWIDLE | | |
4902 | SD_BALANCE_FORK | | |
4903 | SD_BALANCE_EXEC)) { | |
4904 | if (sd->groups != sd->groups->next) | |
4905 | return 0; | |
4906 | } | |
4907 | ||
4908 | /* Following flags don't use groups */ | |
4909 | if (sd->flags & (SD_WAKE_IDLE | | |
4910 | SD_WAKE_AFFINE | | |
4911 | SD_WAKE_BALANCE)) | |
4912 | return 0; | |
4913 | ||
4914 | return 1; | |
4915 | } | |
4916 | ||
1a20ff27 | 4917 | static int sd_parent_degenerate(struct sched_domain *sd, |
245af2c7 SS |
4918 | struct sched_domain *parent) |
4919 | { | |
4920 | unsigned long cflags = sd->flags, pflags = parent->flags; | |
4921 | ||
4922 | if (sd_degenerate(parent)) | |
4923 | return 1; | |
4924 | ||
4925 | if (!cpus_equal(sd->span, parent->span)) | |
4926 | return 0; | |
4927 | ||
4928 | /* Does parent contain flags not in child? */ | |
4929 | /* WAKE_BALANCE is a subset of WAKE_AFFINE */ | |
4930 | if (cflags & SD_WAKE_AFFINE) | |
4931 | pflags &= ~SD_WAKE_BALANCE; | |
4932 | /* Flags needing groups don't count if only 1 group in parent */ | |
4933 | if (parent->groups == parent->groups->next) { | |
4934 | pflags &= ~(SD_LOAD_BALANCE | | |
4935 | SD_BALANCE_NEWIDLE | | |
4936 | SD_BALANCE_FORK | | |
4937 | SD_BALANCE_EXEC); | |
4938 | } | |
4939 | if (~cflags & pflags) | |
4940 | return 0; | |
4941 | ||
4942 | return 1; | |
4943 | } | |
4944 | ||
1da177e4 LT |
4945 | /* |
4946 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must | |
4947 | * hold the hotplug lock. | |
4948 | */ | |
9c1cfda2 | 4949 | static void cpu_attach_domain(struct sched_domain *sd, int cpu) |
1da177e4 | 4950 | { |
1da177e4 | 4951 | runqueue_t *rq = cpu_rq(cpu); |
245af2c7 SS |
4952 | struct sched_domain *tmp; |
4953 | ||
4954 | /* Remove the sched domains which do not contribute to scheduling. */ | |
4955 | for (tmp = sd; tmp; tmp = tmp->parent) { | |
4956 | struct sched_domain *parent = tmp->parent; | |
4957 | if (!parent) | |
4958 | break; | |
4959 | if (sd_parent_degenerate(tmp, parent)) | |
4960 | tmp->parent = parent->parent; | |
4961 | } | |
4962 | ||
4963 | if (sd && sd_degenerate(sd)) | |
4964 | sd = sd->parent; | |
1da177e4 LT |
4965 | |
4966 | sched_domain_debug(sd, cpu); | |
4967 | ||
674311d5 | 4968 | rcu_assign_pointer(rq->sd, sd); |
1da177e4 LT |
4969 | } |
4970 | ||
4971 | /* cpus with isolated domains */ | |
9c1cfda2 | 4972 | static cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE; |
1da177e4 LT |
4973 | |
4974 | /* Setup the mask of cpus configured for isolated domains */ | |
4975 | static int __init isolated_cpu_setup(char *str) | |
4976 | { | |
4977 | int ints[NR_CPUS], i; | |
4978 | ||
4979 | str = get_options(str, ARRAY_SIZE(ints), ints); | |
4980 | cpus_clear(cpu_isolated_map); | |
4981 | for (i = 1; i <= ints[0]; i++) | |
4982 | if (ints[i] < NR_CPUS) | |
4983 | cpu_set(ints[i], cpu_isolated_map); | |
4984 | return 1; | |
4985 | } | |
4986 | ||
4987 | __setup ("isolcpus=", isolated_cpu_setup); | |
4988 | ||
4989 | /* | |
4990 | * init_sched_build_groups takes an array of groups, the cpumask we wish | |
4991 | * to span, and a pointer to a function which identifies what group a CPU | |
4992 | * belongs to. The return value of group_fn must be a valid index into the | |
4993 | * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we | |
4994 | * keep track of groups covered with a cpumask_t). | |
4995 | * | |
4996 | * init_sched_build_groups will build a circular linked list of the groups | |
4997 | * covered by the given span, and will set each group's ->cpumask correctly, | |
4998 | * and ->cpu_power to 0. | |
4999 | */ | |
9c1cfda2 JH |
5000 | static void init_sched_build_groups(struct sched_group groups[], cpumask_t span, |
5001 | int (*group_fn)(int cpu)) | |
1da177e4 LT |
5002 | { |
5003 | struct sched_group *first = NULL, *last = NULL; | |
5004 | cpumask_t covered = CPU_MASK_NONE; | |
5005 | int i; | |
5006 | ||
5007 | for_each_cpu_mask(i, span) { | |
5008 | int group = group_fn(i); | |
5009 | struct sched_group *sg = &groups[group]; | |
5010 | int j; | |
5011 | ||
5012 | if (cpu_isset(i, covered)) | |
5013 | continue; | |
5014 | ||
5015 | sg->cpumask = CPU_MASK_NONE; | |
5016 | sg->cpu_power = 0; | |
5017 | ||
5018 | for_each_cpu_mask(j, span) { | |
5019 | if (group_fn(j) != group) | |
5020 | continue; | |
5021 | ||
5022 | cpu_set(j, covered); | |
5023 | cpu_set(j, sg->cpumask); | |
5024 | } | |
5025 | if (!first) | |
5026 | first = sg; | |
5027 | if (last) | |
5028 | last->next = sg; | |
5029 | last = sg; | |
5030 | } | |
5031 | last->next = first; | |
5032 | } | |
5033 | ||
9c1cfda2 | 5034 | #define SD_NODES_PER_DOMAIN 16 |
1da177e4 | 5035 | |
198e2f18 AM |
5036 | /* |
5037 | * Self-tuning task migration cost measurement between source and target CPUs. | |
5038 | * | |
5039 | * This is done by measuring the cost of manipulating buffers of varying | |
5040 | * sizes. For a given buffer-size here are the steps that are taken: | |
5041 | * | |
5042 | * 1) the source CPU reads+dirties a shared buffer | |
5043 | * 2) the target CPU reads+dirties the same shared buffer | |
5044 | * | |
5045 | * We measure how long they take, in the following 4 scenarios: | |
5046 | * | |
5047 | * - source: CPU1, target: CPU2 | cost1 | |
5048 | * - source: CPU2, target: CPU1 | cost2 | |
5049 | * - source: CPU1, target: CPU1 | cost3 | |
5050 | * - source: CPU2, target: CPU2 | cost4 | |
5051 | * | |
5052 | * We then calculate the cost3+cost4-cost1-cost2 difference - this is | |
5053 | * the cost of migration. | |
5054 | * | |
5055 | * We then start off from a small buffer-size and iterate up to larger | |
5056 | * buffer sizes, in 5% steps - measuring each buffer-size separately, and | |
5057 | * doing a maximum search for the cost. (The maximum cost for a migration | |
5058 | * normally occurs when the working set size is around the effective cache | |
5059 | * size.) | |
5060 | */ | |
5061 | #define SEARCH_SCOPE 2 | |
5062 | #define MIN_CACHE_SIZE (64*1024U) | |
5063 | #define DEFAULT_CACHE_SIZE (5*1024*1024U) | |
70b4d63e | 5064 | #define ITERATIONS 1 |
198e2f18 AM |
5065 | #define SIZE_THRESH 130 |
5066 | #define COST_THRESH 130 | |
5067 | ||
5068 | /* | |
5069 | * The migration cost is a function of 'domain distance'. Domain | |
5070 | * distance is the number of steps a CPU has to iterate down its | |
5071 | * domain tree to share a domain with the other CPU. The farther | |
5072 | * two CPUs are from each other, the larger the distance gets. | |
5073 | * | |
5074 | * Note that we use the distance only to cache measurement results, | |
5075 | * the distance value is not used numerically otherwise. When two | |
5076 | * CPUs have the same distance it is assumed that the migration | |
5077 | * cost is the same. (this is a simplification but quite practical) | |
5078 | */ | |
5079 | #define MAX_DOMAIN_DISTANCE 32 | |
5080 | ||
5081 | static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] = | |
4bbf39c2 IM |
5082 | { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] = |
5083 | /* | |
5084 | * Architectures may override the migration cost and thus avoid | |
5085 | * boot-time calibration. Unit is nanoseconds. Mostly useful for | |
5086 | * virtualized hardware: | |
5087 | */ | |
5088 | #ifdef CONFIG_DEFAULT_MIGRATION_COST | |
5089 | CONFIG_DEFAULT_MIGRATION_COST | |
5090 | #else | |
5091 | -1LL | |
5092 | #endif | |
5093 | }; | |
198e2f18 AM |
5094 | |
5095 | /* | |
5096 | * Allow override of migration cost - in units of microseconds. | |
5097 | * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost | |
5098 | * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs: | |
5099 | */ | |
5100 | static int __init migration_cost_setup(char *str) | |
5101 | { | |
5102 | int ints[MAX_DOMAIN_DISTANCE+1], i; | |
5103 | ||
5104 | str = get_options(str, ARRAY_SIZE(ints), ints); | |
5105 | ||
5106 | printk("#ints: %d\n", ints[0]); | |
5107 | for (i = 1; i <= ints[0]; i++) { | |
5108 | migration_cost[i-1] = (unsigned long long)ints[i]*1000; | |
5109 | printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]); | |
5110 | } | |
5111 | return 1; | |
5112 | } | |
5113 | ||
5114 | __setup ("migration_cost=", migration_cost_setup); | |
5115 | ||
5116 | /* | |
5117 | * Global multiplier (divisor) for migration-cutoff values, | |
5118 | * in percentiles. E.g. use a value of 150 to get 1.5 times | |
5119 | * longer cache-hot cutoff times. | |
5120 | * | |
5121 | * (We scale it from 100 to 128 to long long handling easier.) | |
5122 | */ | |
5123 | ||
5124 | #define MIGRATION_FACTOR_SCALE 128 | |
5125 | ||
5126 | static unsigned int migration_factor = MIGRATION_FACTOR_SCALE; | |
5127 | ||
5128 | static int __init setup_migration_factor(char *str) | |
5129 | { | |
5130 | get_option(&str, &migration_factor); | |
5131 | migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100; | |
5132 | return 1; | |
5133 | } | |
5134 | ||
5135 | __setup("migration_factor=", setup_migration_factor); | |
5136 | ||
5137 | /* | |
5138 | * Estimated distance of two CPUs, measured via the number of domains | |
5139 | * we have to pass for the two CPUs to be in the same span: | |
5140 | */ | |
5141 | static unsigned long domain_distance(int cpu1, int cpu2) | |
5142 | { | |
5143 | unsigned long distance = 0; | |
5144 | struct sched_domain *sd; | |
5145 | ||
5146 | for_each_domain(cpu1, sd) { | |
5147 | WARN_ON(!cpu_isset(cpu1, sd->span)); | |
5148 | if (cpu_isset(cpu2, sd->span)) | |
5149 | return distance; | |
5150 | distance++; | |
5151 | } | |
5152 | if (distance >= MAX_DOMAIN_DISTANCE) { | |
5153 | WARN_ON(1); | |
5154 | distance = MAX_DOMAIN_DISTANCE-1; | |
5155 | } | |
5156 | ||
5157 | return distance; | |
5158 | } | |
5159 | ||
5160 | static unsigned int migration_debug; | |
5161 | ||
5162 | static int __init setup_migration_debug(char *str) | |
5163 | { | |
5164 | get_option(&str, &migration_debug); | |
5165 | return 1; | |
5166 | } | |
5167 | ||
5168 | __setup("migration_debug=", setup_migration_debug); | |
5169 | ||
5170 | /* | |
5171 | * Maximum cache-size that the scheduler should try to measure. | |
5172 | * Architectures with larger caches should tune this up during | |
5173 | * bootup. Gets used in the domain-setup code (i.e. during SMP | |
5174 | * bootup). | |
5175 | */ | |
5176 | unsigned int max_cache_size; | |
5177 | ||
5178 | static int __init setup_max_cache_size(char *str) | |
5179 | { | |
5180 | get_option(&str, &max_cache_size); | |
5181 | return 1; | |
5182 | } | |
5183 | ||
5184 | __setup("max_cache_size=", setup_max_cache_size); | |
5185 | ||
5186 | /* | |
5187 | * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This | |
5188 | * is the operation that is timed, so we try to generate unpredictable | |
5189 | * cachemisses that still end up filling the L2 cache: | |
5190 | */ | |
5191 | static void touch_cache(void *__cache, unsigned long __size) | |
5192 | { | |
5193 | unsigned long size = __size/sizeof(long), chunk1 = size/3, | |
5194 | chunk2 = 2*size/3; | |
5195 | unsigned long *cache = __cache; | |
5196 | int i; | |
5197 | ||
5198 | for (i = 0; i < size/6; i += 8) { | |
5199 | switch (i % 6) { | |
5200 | case 0: cache[i]++; | |
5201 | case 1: cache[size-1-i]++; | |
5202 | case 2: cache[chunk1-i]++; | |
5203 | case 3: cache[chunk1+i]++; | |
5204 | case 4: cache[chunk2-i]++; | |
5205 | case 5: cache[chunk2+i]++; | |
5206 | } | |
5207 | } | |
5208 | } | |
5209 | ||
5210 | /* | |
5211 | * Measure the cache-cost of one task migration. Returns in units of nsec. | |
5212 | */ | |
5213 | static unsigned long long measure_one(void *cache, unsigned long size, | |
5214 | int source, int target) | |
5215 | { | |
5216 | cpumask_t mask, saved_mask; | |
5217 | unsigned long long t0, t1, t2, t3, cost; | |
5218 | ||
5219 | saved_mask = current->cpus_allowed; | |
5220 | ||
5221 | /* | |
5222 | * Flush source caches to RAM and invalidate them: | |
5223 | */ | |
5224 | sched_cacheflush(); | |
5225 | ||
5226 | /* | |
5227 | * Migrate to the source CPU: | |
5228 | */ | |
5229 | mask = cpumask_of_cpu(source); | |
5230 | set_cpus_allowed(current, mask); | |
5231 | WARN_ON(smp_processor_id() != source); | |
5232 | ||
5233 | /* | |
5234 | * Dirty the working set: | |
5235 | */ | |
5236 | t0 = sched_clock(); | |
5237 | touch_cache(cache, size); | |
5238 | t1 = sched_clock(); | |
5239 | ||
5240 | /* | |
5241 | * Migrate to the target CPU, dirty the L2 cache and access | |
5242 | * the shared buffer. (which represents the working set | |
5243 | * of a migrated task.) | |
5244 | */ | |
5245 | mask = cpumask_of_cpu(target); | |
5246 | set_cpus_allowed(current, mask); | |
5247 | WARN_ON(smp_processor_id() != target); | |
5248 | ||
5249 | t2 = sched_clock(); | |
5250 | touch_cache(cache, size); | |
5251 | t3 = sched_clock(); | |
5252 | ||
5253 | cost = t1-t0 + t3-t2; | |
5254 | ||
5255 | if (migration_debug >= 2) | |
5256 | printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n", | |
5257 | source, target, t1-t0, t1-t0, t3-t2, cost); | |
5258 | /* | |
5259 | * Flush target caches to RAM and invalidate them: | |
5260 | */ | |
5261 | sched_cacheflush(); | |
5262 | ||
5263 | set_cpus_allowed(current, saved_mask); | |
5264 | ||
5265 | return cost; | |
5266 | } | |
5267 | ||
5268 | /* | |
5269 | * Measure a series of task migrations and return the average | |
5270 | * result. Since this code runs early during bootup the system | |
5271 | * is 'undisturbed' and the average latency makes sense. | |
5272 | * | |
5273 | * The algorithm in essence auto-detects the relevant cache-size, | |
5274 | * so it will properly detect different cachesizes for different | |
5275 | * cache-hierarchies, depending on how the CPUs are connected. | |
5276 | * | |
5277 | * Architectures can prime the upper limit of the search range via | |
5278 | * max_cache_size, otherwise the search range defaults to 20MB...64K. | |
5279 | */ | |
5280 | static unsigned long long | |
5281 | measure_cost(int cpu1, int cpu2, void *cache, unsigned int size) | |
5282 | { | |
5283 | unsigned long long cost1, cost2; | |
5284 | int i; | |
5285 | ||
5286 | /* | |
5287 | * Measure the migration cost of 'size' bytes, over an | |
5288 | * average of 10 runs: | |
5289 | * | |
5290 | * (We perturb the cache size by a small (0..4k) | |
5291 | * value to compensate size/alignment related artifacts. | |
5292 | * We also subtract the cost of the operation done on | |
5293 | * the same CPU.) | |
5294 | */ | |
5295 | cost1 = 0; | |
5296 | ||
5297 | /* | |
5298 | * dry run, to make sure we start off cache-cold on cpu1, | |
5299 | * and to get any vmalloc pagefaults in advance: | |
5300 | */ | |
5301 | measure_one(cache, size, cpu1, cpu2); | |
5302 | for (i = 0; i < ITERATIONS; i++) | |
5303 | cost1 += measure_one(cache, size - i*1024, cpu1, cpu2); | |
5304 | ||
5305 | measure_one(cache, size, cpu2, cpu1); | |
5306 | for (i = 0; i < ITERATIONS; i++) | |
5307 | cost1 += measure_one(cache, size - i*1024, cpu2, cpu1); | |
5308 | ||
5309 | /* | |
5310 | * (We measure the non-migrating [cached] cost on both | |
5311 | * cpu1 and cpu2, to handle CPUs with different speeds) | |
5312 | */ | |
5313 | cost2 = 0; | |
5314 | ||
5315 | measure_one(cache, size, cpu1, cpu1); | |
5316 | for (i = 0; i < ITERATIONS; i++) | |
5317 | cost2 += measure_one(cache, size - i*1024, cpu1, cpu1); | |
5318 | ||
5319 | measure_one(cache, size, cpu2, cpu2); | |
5320 | for (i = 0; i < ITERATIONS; i++) | |
5321 | cost2 += measure_one(cache, size - i*1024, cpu2, cpu2); | |
5322 | ||
5323 | /* | |
5324 | * Get the per-iteration migration cost: | |
5325 | */ | |
5326 | do_div(cost1, 2*ITERATIONS); | |
5327 | do_div(cost2, 2*ITERATIONS); | |
5328 | ||
5329 | return cost1 - cost2; | |
5330 | } | |
5331 | ||
5332 | static unsigned long long measure_migration_cost(int cpu1, int cpu2) | |
5333 | { | |
5334 | unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0; | |
5335 | unsigned int max_size, size, size_found = 0; | |
5336 | long long cost = 0, prev_cost; | |
5337 | void *cache; | |
5338 | ||
5339 | /* | |
5340 | * Search from max_cache_size*5 down to 64K - the real relevant | |
5341 | * cachesize has to lie somewhere inbetween. | |
5342 | */ | |
5343 | if (max_cache_size) { | |
5344 | max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE); | |
5345 | size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE); | |
5346 | } else { | |
5347 | /* | |
5348 | * Since we have no estimation about the relevant | |
5349 | * search range | |
5350 | */ | |
5351 | max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE; | |
5352 | size = MIN_CACHE_SIZE; | |
5353 | } | |
5354 | ||
5355 | if (!cpu_online(cpu1) || !cpu_online(cpu2)) { | |
5356 | printk("cpu %d and %d not both online!\n", cpu1, cpu2); | |
5357 | return 0; | |
5358 | } | |
5359 | ||
5360 | /* | |
5361 | * Allocate the working set: | |
5362 | */ | |
5363 | cache = vmalloc(max_size); | |
5364 | if (!cache) { | |
5365 | printk("could not vmalloc %d bytes for cache!\n", 2*max_size); | |
5366 | return 1000000; // return 1 msec on very small boxen | |
5367 | } | |
5368 | ||
5369 | while (size <= max_size) { | |
5370 | prev_cost = cost; | |
5371 | cost = measure_cost(cpu1, cpu2, cache, size); | |
5372 | ||
5373 | /* | |
5374 | * Update the max: | |
5375 | */ | |
5376 | if (cost > 0) { | |
5377 | if (max_cost < cost) { | |
5378 | max_cost = cost; | |
5379 | size_found = size; | |
5380 | } | |
5381 | } | |
5382 | /* | |
5383 | * Calculate average fluctuation, we use this to prevent | |
5384 | * noise from triggering an early break out of the loop: | |
5385 | */ | |
5386 | fluct = abs(cost - prev_cost); | |
5387 | avg_fluct = (avg_fluct + fluct)/2; | |
5388 | ||
5389 | if (migration_debug) | |
5390 | printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n", | |
5391 | cpu1, cpu2, size, | |
5392 | (long)cost / 1000000, | |
5393 | ((long)cost / 100000) % 10, | |
5394 | (long)max_cost / 1000000, | |
5395 | ((long)max_cost / 100000) % 10, | |
5396 | domain_distance(cpu1, cpu2), | |
5397 | cost, avg_fluct); | |
5398 | ||
5399 | /* | |
5400 | * If we iterated at least 20% past the previous maximum, | |
5401 | * and the cost has dropped by more than 20% already, | |
5402 | * (taking fluctuations into account) then we assume to | |
5403 | * have found the maximum and break out of the loop early: | |
5404 | */ | |
5405 | if (size_found && (size*100 > size_found*SIZE_THRESH)) | |
5406 | if (cost+avg_fluct <= 0 || | |
5407 | max_cost*100 > (cost+avg_fluct)*COST_THRESH) { | |
5408 | ||
5409 | if (migration_debug) | |
5410 | printk("-> found max.\n"); | |
5411 | break; | |
5412 | } | |
5413 | /* | |
70b4d63e | 5414 | * Increase the cachesize in 10% steps: |
198e2f18 | 5415 | */ |
70b4d63e | 5416 | size = size * 10 / 9; |
198e2f18 AM |
5417 | } |
5418 | ||
5419 | if (migration_debug) | |
5420 | printk("[%d][%d] working set size found: %d, cost: %Ld\n", | |
5421 | cpu1, cpu2, size_found, max_cost); | |
5422 | ||
5423 | vfree(cache); | |
5424 | ||
5425 | /* | |
5426 | * A task is considered 'cache cold' if at least 2 times | |
5427 | * the worst-case cost of migration has passed. | |
5428 | * | |
5429 | * (this limit is only listened to if the load-balancing | |
5430 | * situation is 'nice' - if there is a large imbalance we | |
5431 | * ignore it for the sake of CPU utilization and | |
5432 | * processing fairness.) | |
5433 | */ | |
5434 | return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE; | |
5435 | } | |
5436 | ||
5437 | static void calibrate_migration_costs(const cpumask_t *cpu_map) | |
5438 | { | |
5439 | int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id(); | |
5440 | unsigned long j0, j1, distance, max_distance = 0; | |
5441 | struct sched_domain *sd; | |
5442 | ||
5443 | j0 = jiffies; | |
5444 | ||
5445 | /* | |
5446 | * First pass - calculate the cacheflush times: | |
5447 | */ | |
5448 | for_each_cpu_mask(cpu1, *cpu_map) { | |
5449 | for_each_cpu_mask(cpu2, *cpu_map) { | |
5450 | if (cpu1 == cpu2) | |
5451 | continue; | |
5452 | distance = domain_distance(cpu1, cpu2); | |
5453 | max_distance = max(max_distance, distance); | |
5454 | /* | |
5455 | * No result cached yet? | |
5456 | */ | |
5457 | if (migration_cost[distance] == -1LL) | |
5458 | migration_cost[distance] = | |
5459 | measure_migration_cost(cpu1, cpu2); | |
5460 | } | |
5461 | } | |
5462 | /* | |
5463 | * Second pass - update the sched domain hierarchy with | |
5464 | * the new cache-hot-time estimations: | |
5465 | */ | |
5466 | for_each_cpu_mask(cpu, *cpu_map) { | |
5467 | distance = 0; | |
5468 | for_each_domain(cpu, sd) { | |
5469 | sd->cache_hot_time = migration_cost[distance]; | |
5470 | distance++; | |
5471 | } | |
5472 | } | |
5473 | /* | |
5474 | * Print the matrix: | |
5475 | */ | |
5476 | if (migration_debug) | |
5477 | printk("migration: max_cache_size: %d, cpu: %d MHz:\n", | |
5478 | max_cache_size, | |
5479 | #ifdef CONFIG_X86 | |
5480 | cpu_khz/1000 | |
5481 | #else | |
5482 | -1 | |
5483 | #endif | |
5484 | ); | |
bd576c95 CE |
5485 | if (system_state == SYSTEM_BOOTING) { |
5486 | printk("migration_cost="); | |
5487 | for (distance = 0; distance <= max_distance; distance++) { | |
5488 | if (distance) | |
5489 | printk(","); | |
5490 | printk("%ld", (long)migration_cost[distance] / 1000); | |
5491 | } | |
5492 | printk("\n"); | |
198e2f18 | 5493 | } |
198e2f18 AM |
5494 | j1 = jiffies; |
5495 | if (migration_debug) | |
5496 | printk("migration: %ld seconds\n", (j1-j0)/HZ); | |
5497 | ||
5498 | /* | |
5499 | * Move back to the original CPU. NUMA-Q gets confused | |
5500 | * if we migrate to another quad during bootup. | |
5501 | */ | |
5502 | if (raw_smp_processor_id() != orig_cpu) { | |
5503 | cpumask_t mask = cpumask_of_cpu(orig_cpu), | |
5504 | saved_mask = current->cpus_allowed; | |
5505 | ||
5506 | set_cpus_allowed(current, mask); | |
5507 | set_cpus_allowed(current, saved_mask); | |
5508 | } | |
5509 | } | |
5510 | ||
9c1cfda2 | 5511 | #ifdef CONFIG_NUMA |
198e2f18 | 5512 | |
9c1cfda2 JH |
5513 | /** |
5514 | * find_next_best_node - find the next node to include in a sched_domain | |
5515 | * @node: node whose sched_domain we're building | |
5516 | * @used_nodes: nodes already in the sched_domain | |
5517 | * | |
5518 | * Find the next node to include in a given scheduling domain. Simply | |
5519 | * finds the closest node not already in the @used_nodes map. | |
5520 | * | |
5521 | * Should use nodemask_t. | |
5522 | */ | |
5523 | static int find_next_best_node(int node, unsigned long *used_nodes) | |
5524 | { | |
5525 | int i, n, val, min_val, best_node = 0; | |
5526 | ||
5527 | min_val = INT_MAX; | |
5528 | ||
5529 | for (i = 0; i < MAX_NUMNODES; i++) { | |
5530 | /* Start at @node */ | |
5531 | n = (node + i) % MAX_NUMNODES; | |
5532 | ||
5533 | if (!nr_cpus_node(n)) | |
5534 | continue; | |
5535 | ||
5536 | /* Skip already used nodes */ | |
5537 | if (test_bit(n, used_nodes)) | |
5538 | continue; | |
5539 | ||
5540 | /* Simple min distance search */ | |
5541 | val = node_distance(node, n); | |
5542 | ||
5543 | if (val < min_val) { | |
5544 | min_val = val; | |
5545 | best_node = n; | |
5546 | } | |
5547 | } | |
5548 | ||
5549 | set_bit(best_node, used_nodes); | |
5550 | return best_node; | |
5551 | } | |
5552 | ||
5553 | /** | |
5554 | * sched_domain_node_span - get a cpumask for a node's sched_domain | |
5555 | * @node: node whose cpumask we're constructing | |
5556 | * @size: number of nodes to include in this span | |
5557 | * | |
5558 | * Given a node, construct a good cpumask for its sched_domain to span. It | |
5559 | * should be one that prevents unnecessary balancing, but also spreads tasks | |
5560 | * out optimally. | |
5561 | */ | |
5562 | static cpumask_t sched_domain_node_span(int node) | |
5563 | { | |
5564 | int i; | |
5565 | cpumask_t span, nodemask; | |
5566 | DECLARE_BITMAP(used_nodes, MAX_NUMNODES); | |
5567 | ||
5568 | cpus_clear(span); | |
5569 | bitmap_zero(used_nodes, MAX_NUMNODES); | |
5570 | ||
5571 | nodemask = node_to_cpumask(node); | |
5572 | cpus_or(span, span, nodemask); | |
5573 | set_bit(node, used_nodes); | |
5574 | ||
5575 | for (i = 1; i < SD_NODES_PER_DOMAIN; i++) { | |
5576 | int next_node = find_next_best_node(node, used_nodes); | |
5577 | nodemask = node_to_cpumask(next_node); | |
5578 | cpus_or(span, span, nodemask); | |
5579 | } | |
5580 | ||
5581 | return span; | |
5582 | } | |
5583 | #endif | |
5584 | ||
5585 | /* | |
5586 | * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we | |
5587 | * can switch it on easily if needed. | |
5588 | */ | |
1da177e4 LT |
5589 | #ifdef CONFIG_SCHED_SMT |
5590 | static DEFINE_PER_CPU(struct sched_domain, cpu_domains); | |
5591 | static struct sched_group sched_group_cpus[NR_CPUS]; | |
1a20ff27 | 5592 | static int cpu_to_cpu_group(int cpu) |
1da177e4 LT |
5593 | { |
5594 | return cpu; | |
5595 | } | |
5596 | #endif | |
5597 | ||
1e9f28fa SS |
5598 | #ifdef CONFIG_SCHED_MC |
5599 | static DEFINE_PER_CPU(struct sched_domain, core_domains); | |
5600 | static struct sched_group sched_group_core[NR_CPUS]; | |
5601 | #endif | |
5602 | ||
5603 | #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT) | |
5604 | static int cpu_to_core_group(int cpu) | |
5605 | { | |
5606 | return first_cpu(cpu_sibling_map[cpu]); | |
5607 | } | |
5608 | #elif defined(CONFIG_SCHED_MC) | |
5609 | static int cpu_to_core_group(int cpu) | |
5610 | { | |
5611 | return cpu; | |
5612 | } | |
5613 | #endif | |
5614 | ||
1da177e4 LT |
5615 | static DEFINE_PER_CPU(struct sched_domain, phys_domains); |
5616 | static struct sched_group sched_group_phys[NR_CPUS]; | |
1a20ff27 | 5617 | static int cpu_to_phys_group(int cpu) |
1da177e4 | 5618 | { |
1e9f28fa SS |
5619 | #if defined(CONFIG_SCHED_MC) |
5620 | cpumask_t mask = cpu_coregroup_map(cpu); | |
5621 | return first_cpu(mask); | |
5622 | #elif defined(CONFIG_SCHED_SMT) | |
1da177e4 LT |
5623 | return first_cpu(cpu_sibling_map[cpu]); |
5624 | #else | |
5625 | return cpu; | |
5626 | #endif | |
5627 | } | |
5628 | ||
5629 | #ifdef CONFIG_NUMA | |
1da177e4 | 5630 | /* |
9c1cfda2 JH |
5631 | * The init_sched_build_groups can't handle what we want to do with node |
5632 | * groups, so roll our own. Now each node has its own list of groups which | |
5633 | * gets dynamically allocated. | |
1da177e4 | 5634 | */ |
9c1cfda2 | 5635 | static DEFINE_PER_CPU(struct sched_domain, node_domains); |
d1b55138 | 5636 | static struct sched_group **sched_group_nodes_bycpu[NR_CPUS]; |
1da177e4 | 5637 | |
9c1cfda2 | 5638 | static DEFINE_PER_CPU(struct sched_domain, allnodes_domains); |
d1b55138 | 5639 | static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS]; |
9c1cfda2 JH |
5640 | |
5641 | static int cpu_to_allnodes_group(int cpu) | |
5642 | { | |
5643 | return cpu_to_node(cpu); | |
1da177e4 | 5644 | } |
08069033 SS |
5645 | static void init_numa_sched_groups_power(struct sched_group *group_head) |
5646 | { | |
5647 | struct sched_group *sg = group_head; | |
5648 | int j; | |
5649 | ||
5650 | if (!sg) | |
5651 | return; | |
5652 | next_sg: | |
5653 | for_each_cpu_mask(j, sg->cpumask) { | |
5654 | struct sched_domain *sd; | |
5655 | ||
5656 | sd = &per_cpu(phys_domains, j); | |
5657 | if (j != first_cpu(sd->groups->cpumask)) { | |
5658 | /* | |
5659 | * Only add "power" once for each | |
5660 | * physical package. | |
5661 | */ | |
5662 | continue; | |
5663 | } | |
5664 | ||
5665 | sg->cpu_power += sd->groups->cpu_power; | |
5666 | } | |
5667 | sg = sg->next; | |
5668 | if (sg != group_head) | |
5669 | goto next_sg; | |
5670 | } | |
1da177e4 LT |
5671 | #endif |
5672 | ||
5673 | /* | |
1a20ff27 DG |
5674 | * Build sched domains for a given set of cpus and attach the sched domains |
5675 | * to the individual cpus | |
1da177e4 | 5676 | */ |
9c1cfda2 | 5677 | void build_sched_domains(const cpumask_t *cpu_map) |
1da177e4 LT |
5678 | { |
5679 | int i; | |
d1b55138 JH |
5680 | #ifdef CONFIG_NUMA |
5681 | struct sched_group **sched_group_nodes = NULL; | |
5682 | struct sched_group *sched_group_allnodes = NULL; | |
5683 | ||
5684 | /* | |
5685 | * Allocate the per-node list of sched groups | |
5686 | */ | |
5687 | sched_group_nodes = kmalloc(sizeof(struct sched_group*)*MAX_NUMNODES, | |
5688 | GFP_ATOMIC); | |
5689 | if (!sched_group_nodes) { | |
5690 | printk(KERN_WARNING "Can not alloc sched group node list\n"); | |
5691 | return; | |
5692 | } | |
5693 | sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes; | |
5694 | #endif | |
1da177e4 LT |
5695 | |
5696 | /* | |
1a20ff27 | 5697 | * Set up domains for cpus specified by the cpu_map. |
1da177e4 | 5698 | */ |
1a20ff27 | 5699 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 LT |
5700 | int group; |
5701 | struct sched_domain *sd = NULL, *p; | |
5702 | cpumask_t nodemask = node_to_cpumask(cpu_to_node(i)); | |
5703 | ||
1a20ff27 | 5704 | cpus_and(nodemask, nodemask, *cpu_map); |
1da177e4 LT |
5705 | |
5706 | #ifdef CONFIG_NUMA | |
d1b55138 | 5707 | if (cpus_weight(*cpu_map) |
9c1cfda2 | 5708 | > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) { |
d1b55138 JH |
5709 | if (!sched_group_allnodes) { |
5710 | sched_group_allnodes | |
5711 | = kmalloc(sizeof(struct sched_group) | |
5712 | * MAX_NUMNODES, | |
5713 | GFP_KERNEL); | |
5714 | if (!sched_group_allnodes) { | |
5715 | printk(KERN_WARNING | |
5716 | "Can not alloc allnodes sched group\n"); | |
5717 | break; | |
5718 | } | |
5719 | sched_group_allnodes_bycpu[i] | |
5720 | = sched_group_allnodes; | |
5721 | } | |
9c1cfda2 JH |
5722 | sd = &per_cpu(allnodes_domains, i); |
5723 | *sd = SD_ALLNODES_INIT; | |
5724 | sd->span = *cpu_map; | |
5725 | group = cpu_to_allnodes_group(i); | |
5726 | sd->groups = &sched_group_allnodes[group]; | |
5727 | p = sd; | |
5728 | } else | |
5729 | p = NULL; | |
5730 | ||
1da177e4 | 5731 | sd = &per_cpu(node_domains, i); |
1da177e4 | 5732 | *sd = SD_NODE_INIT; |
9c1cfda2 JH |
5733 | sd->span = sched_domain_node_span(cpu_to_node(i)); |
5734 | sd->parent = p; | |
5735 | cpus_and(sd->span, sd->span, *cpu_map); | |
1da177e4 LT |
5736 | #endif |
5737 | ||
5738 | p = sd; | |
5739 | sd = &per_cpu(phys_domains, i); | |
5740 | group = cpu_to_phys_group(i); | |
5741 | *sd = SD_CPU_INIT; | |
5742 | sd->span = nodemask; | |
5743 | sd->parent = p; | |
5744 | sd->groups = &sched_group_phys[group]; | |
5745 | ||
1e9f28fa SS |
5746 | #ifdef CONFIG_SCHED_MC |
5747 | p = sd; | |
5748 | sd = &per_cpu(core_domains, i); | |
5749 | group = cpu_to_core_group(i); | |
5750 | *sd = SD_MC_INIT; | |
5751 | sd->span = cpu_coregroup_map(i); | |
5752 | cpus_and(sd->span, sd->span, *cpu_map); | |
5753 | sd->parent = p; | |
5754 | sd->groups = &sched_group_core[group]; | |
5755 | #endif | |
5756 | ||
1da177e4 LT |
5757 | #ifdef CONFIG_SCHED_SMT |
5758 | p = sd; | |
5759 | sd = &per_cpu(cpu_domains, i); | |
5760 | group = cpu_to_cpu_group(i); | |
5761 | *sd = SD_SIBLING_INIT; | |
5762 | sd->span = cpu_sibling_map[i]; | |
1a20ff27 | 5763 | cpus_and(sd->span, sd->span, *cpu_map); |
1da177e4 LT |
5764 | sd->parent = p; |
5765 | sd->groups = &sched_group_cpus[group]; | |
5766 | #endif | |
5767 | } | |
5768 | ||
5769 | #ifdef CONFIG_SCHED_SMT | |
5770 | /* Set up CPU (sibling) groups */ | |
9c1cfda2 | 5771 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 | 5772 | cpumask_t this_sibling_map = cpu_sibling_map[i]; |
1a20ff27 | 5773 | cpus_and(this_sibling_map, this_sibling_map, *cpu_map); |
1da177e4 LT |
5774 | if (i != first_cpu(this_sibling_map)) |
5775 | continue; | |
5776 | ||
5777 | init_sched_build_groups(sched_group_cpus, this_sibling_map, | |
5778 | &cpu_to_cpu_group); | |
5779 | } | |
5780 | #endif | |
5781 | ||
1e9f28fa SS |
5782 | #ifdef CONFIG_SCHED_MC |
5783 | /* Set up multi-core groups */ | |
5784 | for_each_cpu_mask(i, *cpu_map) { | |
5785 | cpumask_t this_core_map = cpu_coregroup_map(i); | |
5786 | cpus_and(this_core_map, this_core_map, *cpu_map); | |
5787 | if (i != first_cpu(this_core_map)) | |
5788 | continue; | |
5789 | init_sched_build_groups(sched_group_core, this_core_map, | |
5790 | &cpu_to_core_group); | |
5791 | } | |
5792 | #endif | |
5793 | ||
5794 | ||
1da177e4 LT |
5795 | /* Set up physical groups */ |
5796 | for (i = 0; i < MAX_NUMNODES; i++) { | |
5797 | cpumask_t nodemask = node_to_cpumask(i); | |
5798 | ||
1a20ff27 | 5799 | cpus_and(nodemask, nodemask, *cpu_map); |
1da177e4 LT |
5800 | if (cpus_empty(nodemask)) |
5801 | continue; | |
5802 | ||
5803 | init_sched_build_groups(sched_group_phys, nodemask, | |
5804 | &cpu_to_phys_group); | |
5805 | } | |
5806 | ||
5807 | #ifdef CONFIG_NUMA | |
5808 | /* Set up node groups */ | |
d1b55138 JH |
5809 | if (sched_group_allnodes) |
5810 | init_sched_build_groups(sched_group_allnodes, *cpu_map, | |
5811 | &cpu_to_allnodes_group); | |
9c1cfda2 JH |
5812 | |
5813 | for (i = 0; i < MAX_NUMNODES; i++) { | |
5814 | /* Set up node groups */ | |
5815 | struct sched_group *sg, *prev; | |
5816 | cpumask_t nodemask = node_to_cpumask(i); | |
5817 | cpumask_t domainspan; | |
5818 | cpumask_t covered = CPU_MASK_NONE; | |
5819 | int j; | |
5820 | ||
5821 | cpus_and(nodemask, nodemask, *cpu_map); | |
d1b55138 JH |
5822 | if (cpus_empty(nodemask)) { |
5823 | sched_group_nodes[i] = NULL; | |
9c1cfda2 | 5824 | continue; |
d1b55138 | 5825 | } |
9c1cfda2 JH |
5826 | |
5827 | domainspan = sched_domain_node_span(i); | |
5828 | cpus_and(domainspan, domainspan, *cpu_map); | |
5829 | ||
5830 | sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL); | |
5831 | sched_group_nodes[i] = sg; | |
5832 | for_each_cpu_mask(j, nodemask) { | |
5833 | struct sched_domain *sd; | |
5834 | sd = &per_cpu(node_domains, j); | |
5835 | sd->groups = sg; | |
5836 | if (sd->groups == NULL) { | |
5837 | /* Turn off balancing if we have no groups */ | |
5838 | sd->flags = 0; | |
5839 | } | |
5840 | } | |
5841 | if (!sg) { | |
5842 | printk(KERN_WARNING | |
5843 | "Can not alloc domain group for node %d\n", i); | |
5844 | continue; | |
5845 | } | |
5846 | sg->cpu_power = 0; | |
5847 | sg->cpumask = nodemask; | |
5848 | cpus_or(covered, covered, nodemask); | |
5849 | prev = sg; | |
5850 | ||
5851 | for (j = 0; j < MAX_NUMNODES; j++) { | |
5852 | cpumask_t tmp, notcovered; | |
5853 | int n = (i + j) % MAX_NUMNODES; | |
5854 | ||
5855 | cpus_complement(notcovered, covered); | |
5856 | cpus_and(tmp, notcovered, *cpu_map); | |
5857 | cpus_and(tmp, tmp, domainspan); | |
5858 | if (cpus_empty(tmp)) | |
5859 | break; | |
5860 | ||
5861 | nodemask = node_to_cpumask(n); | |
5862 | cpus_and(tmp, tmp, nodemask); | |
5863 | if (cpus_empty(tmp)) | |
5864 | continue; | |
5865 | ||
5866 | sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL); | |
5867 | if (!sg) { | |
5868 | printk(KERN_WARNING | |
5869 | "Can not alloc domain group for node %d\n", j); | |
5870 | break; | |
5871 | } | |
5872 | sg->cpu_power = 0; | |
5873 | sg->cpumask = tmp; | |
5874 | cpus_or(covered, covered, tmp); | |
5875 | prev->next = sg; | |
5876 | prev = sg; | |
5877 | } | |
5878 | prev->next = sched_group_nodes[i]; | |
5879 | } | |
1da177e4 LT |
5880 | #endif |
5881 | ||
5882 | /* Calculate CPU power for physical packages and nodes */ | |
1a20ff27 | 5883 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 LT |
5884 | int power; |
5885 | struct sched_domain *sd; | |
5886 | #ifdef CONFIG_SCHED_SMT | |
5887 | sd = &per_cpu(cpu_domains, i); | |
5888 | power = SCHED_LOAD_SCALE; | |
5889 | sd->groups->cpu_power = power; | |
5890 | #endif | |
1e9f28fa SS |
5891 | #ifdef CONFIG_SCHED_MC |
5892 | sd = &per_cpu(core_domains, i); | |
5893 | power = SCHED_LOAD_SCALE + (cpus_weight(sd->groups->cpumask)-1) | |
5894 | * SCHED_LOAD_SCALE / 10; | |
5895 | sd->groups->cpu_power = power; | |
5896 | ||
5897 | sd = &per_cpu(phys_domains, i); | |
1da177e4 | 5898 | |
1e9f28fa SS |
5899 | /* |
5900 | * This has to be < 2 * SCHED_LOAD_SCALE | |
5901 | * Lets keep it SCHED_LOAD_SCALE, so that | |
5902 | * while calculating NUMA group's cpu_power | |
5903 | * we can simply do | |
5904 | * numa_group->cpu_power += phys_group->cpu_power; | |
5905 | * | |
5906 | * See "only add power once for each physical pkg" | |
5907 | * comment below | |
5908 | */ | |
5909 | sd->groups->cpu_power = SCHED_LOAD_SCALE; | |
5910 | #else | |
1da177e4 LT |
5911 | sd = &per_cpu(phys_domains, i); |
5912 | power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE * | |
5913 | (cpus_weight(sd->groups->cpumask)-1) / 10; | |
5914 | sd->groups->cpu_power = power; | |
1e9f28fa | 5915 | #endif |
1da177e4 LT |
5916 | } |
5917 | ||
9c1cfda2 | 5918 | #ifdef CONFIG_NUMA |
08069033 SS |
5919 | for (i = 0; i < MAX_NUMNODES; i++) |
5920 | init_numa_sched_groups_power(sched_group_nodes[i]); | |
9c1cfda2 | 5921 | |
08069033 | 5922 | init_numa_sched_groups_power(sched_group_allnodes); |
9c1cfda2 JH |
5923 | #endif |
5924 | ||
1da177e4 | 5925 | /* Attach the domains */ |
1a20ff27 | 5926 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 LT |
5927 | struct sched_domain *sd; |
5928 | #ifdef CONFIG_SCHED_SMT | |
5929 | sd = &per_cpu(cpu_domains, i); | |
1e9f28fa SS |
5930 | #elif defined(CONFIG_SCHED_MC) |
5931 | sd = &per_cpu(core_domains, i); | |
1da177e4 LT |
5932 | #else |
5933 | sd = &per_cpu(phys_domains, i); | |
5934 | #endif | |
5935 | cpu_attach_domain(sd, i); | |
5936 | } | |
198e2f18 AM |
5937 | /* |
5938 | * Tune cache-hot values: | |
5939 | */ | |
5940 | calibrate_migration_costs(cpu_map); | |
1da177e4 | 5941 | } |
1a20ff27 DG |
5942 | /* |
5943 | * Set up scheduler domains and groups. Callers must hold the hotplug lock. | |
5944 | */ | |
9c1cfda2 | 5945 | static void arch_init_sched_domains(const cpumask_t *cpu_map) |
1a20ff27 DG |
5946 | { |
5947 | cpumask_t cpu_default_map; | |
1da177e4 | 5948 | |
1a20ff27 DG |
5949 | /* |
5950 | * Setup mask for cpus without special case scheduling requirements. | |
5951 | * For now this just excludes isolated cpus, but could be used to | |
5952 | * exclude other special cases in the future. | |
5953 | */ | |
5954 | cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map); | |
5955 | ||
5956 | build_sched_domains(&cpu_default_map); | |
5957 | } | |
5958 | ||
5959 | static void arch_destroy_sched_domains(const cpumask_t *cpu_map) | |
1da177e4 | 5960 | { |
9c1cfda2 JH |
5961 | #ifdef CONFIG_NUMA |
5962 | int i; | |
d1b55138 | 5963 | int cpu; |
1da177e4 | 5964 | |
d1b55138 JH |
5965 | for_each_cpu_mask(cpu, *cpu_map) { |
5966 | struct sched_group *sched_group_allnodes | |
5967 | = sched_group_allnodes_bycpu[cpu]; | |
5968 | struct sched_group **sched_group_nodes | |
5969 | = sched_group_nodes_bycpu[cpu]; | |
9c1cfda2 | 5970 | |
d1b55138 JH |
5971 | if (sched_group_allnodes) { |
5972 | kfree(sched_group_allnodes); | |
5973 | sched_group_allnodes_bycpu[cpu] = NULL; | |
5974 | } | |
5975 | ||
5976 | if (!sched_group_nodes) | |
9c1cfda2 | 5977 | continue; |
d1b55138 JH |
5978 | |
5979 | for (i = 0; i < MAX_NUMNODES; i++) { | |
5980 | cpumask_t nodemask = node_to_cpumask(i); | |
5981 | struct sched_group *oldsg, *sg = sched_group_nodes[i]; | |
5982 | ||
5983 | cpus_and(nodemask, nodemask, *cpu_map); | |
5984 | if (cpus_empty(nodemask)) | |
5985 | continue; | |
5986 | ||
5987 | if (sg == NULL) | |
5988 | continue; | |
5989 | sg = sg->next; | |
9c1cfda2 | 5990 | next_sg: |
d1b55138 JH |
5991 | oldsg = sg; |
5992 | sg = sg->next; | |
5993 | kfree(oldsg); | |
5994 | if (oldsg != sched_group_nodes[i]) | |
5995 | goto next_sg; | |
5996 | } | |
5997 | kfree(sched_group_nodes); | |
5998 | sched_group_nodes_bycpu[cpu] = NULL; | |
9c1cfda2 JH |
5999 | } |
6000 | #endif | |
6001 | } | |
1da177e4 | 6002 | |
1a20ff27 DG |
6003 | /* |
6004 | * Detach sched domains from a group of cpus specified in cpu_map | |
6005 | * These cpus will now be attached to the NULL domain | |
6006 | */ | |
858119e1 | 6007 | static void detach_destroy_domains(const cpumask_t *cpu_map) |
1a20ff27 DG |
6008 | { |
6009 | int i; | |
6010 | ||
6011 | for_each_cpu_mask(i, *cpu_map) | |
6012 | cpu_attach_domain(NULL, i); | |
6013 | synchronize_sched(); | |
6014 | arch_destroy_sched_domains(cpu_map); | |
6015 | } | |
6016 | ||
6017 | /* | |
6018 | * Partition sched domains as specified by the cpumasks below. | |
6019 | * This attaches all cpus from the cpumasks to the NULL domain, | |
6020 | * waits for a RCU quiescent period, recalculates sched | |
6021 | * domain information and then attaches them back to the | |
6022 | * correct sched domains | |
6023 | * Call with hotplug lock held | |
6024 | */ | |
6025 | void partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2) | |
6026 | { | |
6027 | cpumask_t change_map; | |
6028 | ||
6029 | cpus_and(*partition1, *partition1, cpu_online_map); | |
6030 | cpus_and(*partition2, *partition2, cpu_online_map); | |
6031 | cpus_or(change_map, *partition1, *partition2); | |
6032 | ||
6033 | /* Detach sched domains from all of the affected cpus */ | |
6034 | detach_destroy_domains(&change_map); | |
6035 | if (!cpus_empty(*partition1)) | |
6036 | build_sched_domains(partition1); | |
6037 | if (!cpus_empty(*partition2)) | |
6038 | build_sched_domains(partition2); | |
6039 | } | |
6040 | ||
1da177e4 LT |
6041 | #ifdef CONFIG_HOTPLUG_CPU |
6042 | /* | |
6043 | * Force a reinitialization of the sched domains hierarchy. The domains | |
6044 | * and groups cannot be updated in place without racing with the balancing | |
41c7ce9a | 6045 | * code, so we temporarily attach all running cpus to the NULL domain |
1da177e4 LT |
6046 | * which will prevent rebalancing while the sched domains are recalculated. |
6047 | */ | |
6048 | static int update_sched_domains(struct notifier_block *nfb, | |
6049 | unsigned long action, void *hcpu) | |
6050 | { | |
1da177e4 LT |
6051 | switch (action) { |
6052 | case CPU_UP_PREPARE: | |
6053 | case CPU_DOWN_PREPARE: | |
1a20ff27 | 6054 | detach_destroy_domains(&cpu_online_map); |
1da177e4 LT |
6055 | return NOTIFY_OK; |
6056 | ||
6057 | case CPU_UP_CANCELED: | |
6058 | case CPU_DOWN_FAILED: | |
6059 | case CPU_ONLINE: | |
6060 | case CPU_DEAD: | |
6061 | /* | |
6062 | * Fall through and re-initialise the domains. | |
6063 | */ | |
6064 | break; | |
6065 | default: | |
6066 | return NOTIFY_DONE; | |
6067 | } | |
6068 | ||
6069 | /* The hotplug lock is already held by cpu_up/cpu_down */ | |
1a20ff27 | 6070 | arch_init_sched_domains(&cpu_online_map); |
1da177e4 LT |
6071 | |
6072 | return NOTIFY_OK; | |
6073 | } | |
6074 | #endif | |
6075 | ||
6076 | void __init sched_init_smp(void) | |
6077 | { | |
6078 | lock_cpu_hotplug(); | |
1a20ff27 | 6079 | arch_init_sched_domains(&cpu_online_map); |
1da177e4 LT |
6080 | unlock_cpu_hotplug(); |
6081 | /* XXX: Theoretical race here - CPU may be hotplugged now */ | |
6082 | hotcpu_notifier(update_sched_domains, 0); | |
6083 | } | |
6084 | #else | |
6085 | void __init sched_init_smp(void) | |
6086 | { | |
6087 | } | |
6088 | #endif /* CONFIG_SMP */ | |
6089 | ||
6090 | int in_sched_functions(unsigned long addr) | |
6091 | { | |
6092 | /* Linker adds these: start and end of __sched functions */ | |
6093 | extern char __sched_text_start[], __sched_text_end[]; | |
6094 | return in_lock_functions(addr) || | |
6095 | (addr >= (unsigned long)__sched_text_start | |
6096 | && addr < (unsigned long)__sched_text_end); | |
6097 | } | |
6098 | ||
6099 | void __init sched_init(void) | |
6100 | { | |
6101 | runqueue_t *rq; | |
6102 | int i, j, k; | |
6103 | ||
0a945022 | 6104 | for_each_possible_cpu(i) { |
1da177e4 LT |
6105 | prio_array_t *array; |
6106 | ||
6107 | rq = cpu_rq(i); | |
6108 | spin_lock_init(&rq->lock); | |
7897986b | 6109 | rq->nr_running = 0; |
1da177e4 LT |
6110 | rq->active = rq->arrays; |
6111 | rq->expired = rq->arrays + 1; | |
6112 | rq->best_expired_prio = MAX_PRIO; | |
6113 | ||
6114 | #ifdef CONFIG_SMP | |
41c7ce9a | 6115 | rq->sd = NULL; |
7897986b NP |
6116 | for (j = 1; j < 3; j++) |
6117 | rq->cpu_load[j] = 0; | |
1da177e4 LT |
6118 | rq->active_balance = 0; |
6119 | rq->push_cpu = 0; | |
6120 | rq->migration_thread = NULL; | |
6121 | INIT_LIST_HEAD(&rq->migration_queue); | |
e9028b0f | 6122 | rq->cpu = i; |
1da177e4 LT |
6123 | #endif |
6124 | atomic_set(&rq->nr_iowait, 0); | |
6125 | ||
6126 | for (j = 0; j < 2; j++) { | |
6127 | array = rq->arrays + j; | |
6128 | for (k = 0; k < MAX_PRIO; k++) { | |
6129 | INIT_LIST_HEAD(array->queue + k); | |
6130 | __clear_bit(k, array->bitmap); | |
6131 | } | |
6132 | // delimiter for bitsearch | |
6133 | __set_bit(MAX_PRIO, array->bitmap); | |
6134 | } | |
6135 | } | |
6136 | ||
6137 | /* | |
6138 | * The boot idle thread does lazy MMU switching as well: | |
6139 | */ | |
6140 | atomic_inc(&init_mm.mm_count); | |
6141 | enter_lazy_tlb(&init_mm, current); | |
6142 | ||
6143 | /* | |
6144 | * Make us the idle thread. Technically, schedule() should not be | |
6145 | * called from this thread, however somewhere below it might be, | |
6146 | * but because we are the idle thread, we just pick up running again | |
6147 | * when this runqueue becomes "idle". | |
6148 | */ | |
6149 | init_idle(current, smp_processor_id()); | |
6150 | } | |
6151 | ||
6152 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP | |
6153 | void __might_sleep(char *file, int line) | |
6154 | { | |
6155 | #if defined(in_atomic) | |
6156 | static unsigned long prev_jiffy; /* ratelimiting */ | |
6157 | ||
6158 | if ((in_atomic() || irqs_disabled()) && | |
6159 | system_state == SYSTEM_RUNNING && !oops_in_progress) { | |
6160 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) | |
6161 | return; | |
6162 | prev_jiffy = jiffies; | |
91368d73 | 6163 | printk(KERN_ERR "BUG: sleeping function called from invalid" |
1da177e4 LT |
6164 | " context at %s:%d\n", file, line); |
6165 | printk("in_atomic():%d, irqs_disabled():%d\n", | |
6166 | in_atomic(), irqs_disabled()); | |
6167 | dump_stack(); | |
6168 | } | |
6169 | #endif | |
6170 | } | |
6171 | EXPORT_SYMBOL(__might_sleep); | |
6172 | #endif | |
6173 | ||
6174 | #ifdef CONFIG_MAGIC_SYSRQ | |
6175 | void normalize_rt_tasks(void) | |
6176 | { | |
6177 | struct task_struct *p; | |
6178 | prio_array_t *array; | |
6179 | unsigned long flags; | |
6180 | runqueue_t *rq; | |
6181 | ||
6182 | read_lock_irq(&tasklist_lock); | |
6183 | for_each_process (p) { | |
6184 | if (!rt_task(p)) | |
6185 | continue; | |
6186 | ||
6187 | rq = task_rq_lock(p, &flags); | |
6188 | ||
6189 | array = p->array; | |
6190 | if (array) | |
6191 | deactivate_task(p, task_rq(p)); | |
6192 | __setscheduler(p, SCHED_NORMAL, 0); | |
6193 | if (array) { | |
6194 | __activate_task(p, task_rq(p)); | |
6195 | resched_task(rq->curr); | |
6196 | } | |
6197 | ||
6198 | task_rq_unlock(rq, &flags); | |
6199 | } | |
6200 | read_unlock_irq(&tasklist_lock); | |
6201 | } | |
6202 | ||
6203 | #endif /* CONFIG_MAGIC_SYSRQ */ | |
1df5c10a LT |
6204 | |
6205 | #ifdef CONFIG_IA64 | |
6206 | /* | |
6207 | * These functions are only useful for the IA64 MCA handling. | |
6208 | * | |
6209 | * They can only be called when the whole system has been | |
6210 | * stopped - every CPU needs to be quiescent, and no scheduling | |
6211 | * activity can take place. Using them for anything else would | |
6212 | * be a serious bug, and as a result, they aren't even visible | |
6213 | * under any other configuration. | |
6214 | */ | |
6215 | ||
6216 | /** | |
6217 | * curr_task - return the current task for a given cpu. | |
6218 | * @cpu: the processor in question. | |
6219 | * | |
6220 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
6221 | */ | |
6222 | task_t *curr_task(int cpu) | |
6223 | { | |
6224 | return cpu_curr(cpu); | |
6225 | } | |
6226 | ||
6227 | /** | |
6228 | * set_curr_task - set the current task for a given cpu. | |
6229 | * @cpu: the processor in question. | |
6230 | * @p: the task pointer to set. | |
6231 | * | |
6232 | * Description: This function must only be used when non-maskable interrupts | |
6233 | * are serviced on a separate stack. It allows the architecture to switch the | |
6234 | * notion of the current task on a cpu in a non-blocking manner. This function | |
6235 | * must be called with all CPU's synchronized, and interrupts disabled, the | |
6236 | * and caller must save the original value of the current task (see | |
6237 | * curr_task() above) and restore that value before reenabling interrupts and | |
6238 | * re-starting the system. | |
6239 | * | |
6240 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
6241 | */ | |
6242 | void set_curr_task(int cpu, task_t *p) | |
6243 | { | |
6244 | cpu_curr(cpu) = p; | |
6245 | } | |
6246 | ||
6247 | #endif |