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) \ | |
2dd73a4f | 171 | max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE) |
1da177e4 | 172 | |
2dd73a4f | 173 | static unsigned int static_prio_timeslice(int static_prio) |
1da177e4 | 174 | { |
2dd73a4f PW |
175 | if (static_prio < NICE_TO_PRIO(0)) |
176 | return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio); | |
1da177e4 | 177 | else |
2dd73a4f | 178 | return SCALE_PRIO(DEF_TIMESLICE, static_prio); |
1da177e4 | 179 | } |
2dd73a4f PW |
180 | |
181 | static inline unsigned int task_timeslice(task_t *p) | |
182 | { | |
183 | return static_prio_timeslice(p->static_prio); | |
184 | } | |
185 | ||
1da177e4 LT |
186 | #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran) \ |
187 | < (long long) (sd)->cache_hot_time) | |
188 | ||
189 | /* | |
190 | * These are the runqueue data structures: | |
191 | */ | |
192 | ||
1da177e4 LT |
193 | typedef struct runqueue runqueue_t; |
194 | ||
195 | struct prio_array { | |
196 | unsigned int nr_active; | |
d444886e | 197 | DECLARE_BITMAP(bitmap, MAX_PRIO+1); /* include 1 bit for delimiter */ |
1da177e4 LT |
198 | struct list_head queue[MAX_PRIO]; |
199 | }; | |
200 | ||
201 | /* | |
202 | * This is the main, per-CPU runqueue data structure. | |
203 | * | |
204 | * Locking rule: those places that want to lock multiple runqueues | |
205 | * (such as the load balancing or the thread migration code), lock | |
206 | * acquire operations must be ordered by ascending &runqueue. | |
207 | */ | |
208 | struct runqueue { | |
209 | spinlock_t lock; | |
210 | ||
211 | /* | |
212 | * nr_running and cpu_load should be in the same cacheline because | |
213 | * remote CPUs use both these fields when doing load calculation. | |
214 | */ | |
215 | unsigned long nr_running; | |
2dd73a4f | 216 | unsigned long raw_weighted_load; |
1da177e4 | 217 | #ifdef CONFIG_SMP |
7897986b | 218 | unsigned long cpu_load[3]; |
1da177e4 LT |
219 | #endif |
220 | unsigned long long nr_switches; | |
221 | ||
222 | /* | |
223 | * This is part of a global counter where only the total sum | |
224 | * over all CPUs matters. A task can increase this counter on | |
225 | * one CPU and if it got migrated afterwards it may decrease | |
226 | * it on another CPU. Always updated under the runqueue lock: | |
227 | */ | |
228 | unsigned long nr_uninterruptible; | |
229 | ||
230 | unsigned long expired_timestamp; | |
231 | unsigned long long timestamp_last_tick; | |
232 | task_t *curr, *idle; | |
233 | struct mm_struct *prev_mm; | |
234 | prio_array_t *active, *expired, arrays[2]; | |
235 | int best_expired_prio; | |
236 | atomic_t nr_iowait; | |
237 | ||
238 | #ifdef CONFIG_SMP | |
239 | struct sched_domain *sd; | |
240 | ||
241 | /* For active balancing */ | |
242 | int active_balance; | |
243 | int push_cpu; | |
244 | ||
245 | task_t *migration_thread; | |
246 | struct list_head migration_queue; | |
247 | #endif | |
248 | ||
249 | #ifdef CONFIG_SCHEDSTATS | |
250 | /* latency stats */ | |
251 | struct sched_info rq_sched_info; | |
252 | ||
253 | /* sys_sched_yield() stats */ | |
254 | unsigned long yld_exp_empty; | |
255 | unsigned long yld_act_empty; | |
256 | unsigned long yld_both_empty; | |
257 | unsigned long yld_cnt; | |
258 | ||
259 | /* schedule() stats */ | |
260 | unsigned long sched_switch; | |
261 | unsigned long sched_cnt; | |
262 | unsigned long sched_goidle; | |
263 | ||
264 | /* try_to_wake_up() stats */ | |
265 | unsigned long ttwu_cnt; | |
266 | unsigned long ttwu_local; | |
267 | #endif | |
268 | }; | |
269 | ||
270 | static DEFINE_PER_CPU(struct runqueue, runqueues); | |
271 | ||
674311d5 NP |
272 | /* |
273 | * The domain tree (rq->sd) is protected by RCU's quiescent state transition. | |
1a20ff27 | 274 | * See detach_destroy_domains: synchronize_sched for details. |
674311d5 NP |
275 | * |
276 | * The domain tree of any CPU may only be accessed from within | |
277 | * preempt-disabled sections. | |
278 | */ | |
1da177e4 | 279 | #define for_each_domain(cpu, domain) \ |
674311d5 | 280 | for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent) |
1da177e4 LT |
281 | |
282 | #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) | |
283 | #define this_rq() (&__get_cpu_var(runqueues)) | |
284 | #define task_rq(p) cpu_rq(task_cpu(p)) | |
285 | #define cpu_curr(cpu) (cpu_rq(cpu)->curr) | |
286 | ||
1da177e4 | 287 | #ifndef prepare_arch_switch |
4866cde0 NP |
288 | # define prepare_arch_switch(next) do { } while (0) |
289 | #endif | |
290 | #ifndef finish_arch_switch | |
291 | # define finish_arch_switch(prev) do { } while (0) | |
292 | #endif | |
293 | ||
294 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW | |
295 | static inline int task_running(runqueue_t *rq, task_t *p) | |
296 | { | |
297 | return rq->curr == p; | |
298 | } | |
299 | ||
300 | static inline void prepare_lock_switch(runqueue_t *rq, task_t *next) | |
301 | { | |
302 | } | |
303 | ||
304 | static inline void finish_lock_switch(runqueue_t *rq, task_t *prev) | |
305 | { | |
da04c035 IM |
306 | #ifdef CONFIG_DEBUG_SPINLOCK |
307 | /* this is a valid case when another task releases the spinlock */ | |
308 | rq->lock.owner = current; | |
309 | #endif | |
4866cde0 NP |
310 | spin_unlock_irq(&rq->lock); |
311 | } | |
312 | ||
313 | #else /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
314 | static inline int task_running(runqueue_t *rq, task_t *p) | |
315 | { | |
316 | #ifdef CONFIG_SMP | |
317 | return p->oncpu; | |
318 | #else | |
319 | return rq->curr == p; | |
320 | #endif | |
321 | } | |
322 | ||
323 | static inline void prepare_lock_switch(runqueue_t *rq, task_t *next) | |
324 | { | |
325 | #ifdef CONFIG_SMP | |
326 | /* | |
327 | * We can optimise this out completely for !SMP, because the | |
328 | * SMP rebalancing from interrupt is the only thing that cares | |
329 | * here. | |
330 | */ | |
331 | next->oncpu = 1; | |
332 | #endif | |
333 | #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
334 | spin_unlock_irq(&rq->lock); | |
335 | #else | |
336 | spin_unlock(&rq->lock); | |
337 | #endif | |
338 | } | |
339 | ||
340 | static inline void finish_lock_switch(runqueue_t *rq, task_t *prev) | |
341 | { | |
342 | #ifdef CONFIG_SMP | |
343 | /* | |
344 | * After ->oncpu is cleared, the task can be moved to a different CPU. | |
345 | * We must ensure this doesn't happen until the switch is completely | |
346 | * finished. | |
347 | */ | |
348 | smp_wmb(); | |
349 | prev->oncpu = 0; | |
350 | #endif | |
351 | #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
352 | local_irq_enable(); | |
1da177e4 | 353 | #endif |
4866cde0 NP |
354 | } |
355 | #endif /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
1da177e4 | 356 | |
b29739f9 IM |
357 | /* |
358 | * __task_rq_lock - lock the runqueue a given task resides on. | |
359 | * Must be called interrupts disabled. | |
360 | */ | |
361 | static inline runqueue_t *__task_rq_lock(task_t *p) | |
362 | __acquires(rq->lock) | |
363 | { | |
364 | struct runqueue *rq; | |
365 | ||
366 | repeat_lock_task: | |
367 | rq = task_rq(p); | |
368 | spin_lock(&rq->lock); | |
369 | if (unlikely(rq != task_rq(p))) { | |
370 | spin_unlock(&rq->lock); | |
371 | goto repeat_lock_task; | |
372 | } | |
373 | return rq; | |
374 | } | |
375 | ||
1da177e4 LT |
376 | /* |
377 | * task_rq_lock - lock the runqueue a given task resides on and disable | |
378 | * interrupts. Note the ordering: we can safely lookup the task_rq without | |
379 | * explicitly disabling preemption. | |
380 | */ | |
9fea80e4 | 381 | static runqueue_t *task_rq_lock(task_t *p, unsigned long *flags) |
1da177e4 LT |
382 | __acquires(rq->lock) |
383 | { | |
384 | struct runqueue *rq; | |
385 | ||
386 | repeat_lock_task: | |
387 | local_irq_save(*flags); | |
388 | rq = task_rq(p); | |
389 | spin_lock(&rq->lock); | |
390 | if (unlikely(rq != task_rq(p))) { | |
391 | spin_unlock_irqrestore(&rq->lock, *flags); | |
392 | goto repeat_lock_task; | |
393 | } | |
394 | return rq; | |
395 | } | |
396 | ||
b29739f9 IM |
397 | static inline void __task_rq_unlock(runqueue_t *rq) |
398 | __releases(rq->lock) | |
399 | { | |
400 | spin_unlock(&rq->lock); | |
401 | } | |
402 | ||
1da177e4 LT |
403 | static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags) |
404 | __releases(rq->lock) | |
405 | { | |
406 | spin_unlock_irqrestore(&rq->lock, *flags); | |
407 | } | |
408 | ||
409 | #ifdef CONFIG_SCHEDSTATS | |
410 | /* | |
411 | * bump this up when changing the output format or the meaning of an existing | |
412 | * format, so that tools can adapt (or abort) | |
413 | */ | |
68767a0a | 414 | #define SCHEDSTAT_VERSION 12 |
1da177e4 LT |
415 | |
416 | static int show_schedstat(struct seq_file *seq, void *v) | |
417 | { | |
418 | int cpu; | |
419 | ||
420 | seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION); | |
421 | seq_printf(seq, "timestamp %lu\n", jiffies); | |
422 | for_each_online_cpu(cpu) { | |
423 | runqueue_t *rq = cpu_rq(cpu); | |
424 | #ifdef CONFIG_SMP | |
425 | struct sched_domain *sd; | |
426 | int dcnt = 0; | |
427 | #endif | |
428 | ||
429 | /* runqueue-specific stats */ | |
430 | seq_printf(seq, | |
431 | "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu", | |
432 | cpu, rq->yld_both_empty, | |
433 | rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt, | |
434 | rq->sched_switch, rq->sched_cnt, rq->sched_goidle, | |
435 | rq->ttwu_cnt, rq->ttwu_local, | |
436 | rq->rq_sched_info.cpu_time, | |
437 | rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt); | |
438 | ||
439 | seq_printf(seq, "\n"); | |
440 | ||
441 | #ifdef CONFIG_SMP | |
442 | /* domain-specific stats */ | |
674311d5 | 443 | preempt_disable(); |
1da177e4 LT |
444 | for_each_domain(cpu, sd) { |
445 | enum idle_type itype; | |
446 | char mask_str[NR_CPUS]; | |
447 | ||
448 | cpumask_scnprintf(mask_str, NR_CPUS, sd->span); | |
449 | seq_printf(seq, "domain%d %s", dcnt++, mask_str); | |
450 | for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES; | |
451 | itype++) { | |
452 | seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu", | |
453 | sd->lb_cnt[itype], | |
454 | sd->lb_balanced[itype], | |
455 | sd->lb_failed[itype], | |
456 | sd->lb_imbalance[itype], | |
457 | sd->lb_gained[itype], | |
458 | sd->lb_hot_gained[itype], | |
459 | sd->lb_nobusyq[itype], | |
460 | sd->lb_nobusyg[itype]); | |
461 | } | |
68767a0a | 462 | seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n", |
1da177e4 | 463 | sd->alb_cnt, sd->alb_failed, sd->alb_pushed, |
68767a0a NP |
464 | sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed, |
465 | sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed, | |
1da177e4 LT |
466 | sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance); |
467 | } | |
674311d5 | 468 | preempt_enable(); |
1da177e4 LT |
469 | #endif |
470 | } | |
471 | return 0; | |
472 | } | |
473 | ||
474 | static int schedstat_open(struct inode *inode, struct file *file) | |
475 | { | |
476 | unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32); | |
477 | char *buf = kmalloc(size, GFP_KERNEL); | |
478 | struct seq_file *m; | |
479 | int res; | |
480 | ||
481 | if (!buf) | |
482 | return -ENOMEM; | |
483 | res = single_open(file, show_schedstat, NULL); | |
484 | if (!res) { | |
485 | m = file->private_data; | |
486 | m->buf = buf; | |
487 | m->size = size; | |
488 | } else | |
489 | kfree(buf); | |
490 | return res; | |
491 | } | |
492 | ||
493 | struct file_operations proc_schedstat_operations = { | |
494 | .open = schedstat_open, | |
495 | .read = seq_read, | |
496 | .llseek = seq_lseek, | |
497 | .release = single_release, | |
498 | }; | |
499 | ||
500 | # define schedstat_inc(rq, field) do { (rq)->field++; } while (0) | |
501 | # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0) | |
502 | #else /* !CONFIG_SCHEDSTATS */ | |
503 | # define schedstat_inc(rq, field) do { } while (0) | |
504 | # define schedstat_add(rq, field, amt) do { } while (0) | |
505 | #endif | |
506 | ||
507 | /* | |
508 | * rq_lock - lock a given runqueue and disable interrupts. | |
509 | */ | |
510 | static inline runqueue_t *this_rq_lock(void) | |
511 | __acquires(rq->lock) | |
512 | { | |
513 | runqueue_t *rq; | |
514 | ||
515 | local_irq_disable(); | |
516 | rq = this_rq(); | |
517 | spin_lock(&rq->lock); | |
518 | ||
519 | return rq; | |
520 | } | |
521 | ||
1da177e4 LT |
522 | #ifdef CONFIG_SCHEDSTATS |
523 | /* | |
524 | * Called when a process is dequeued from the active array and given | |
525 | * the cpu. We should note that with the exception of interactive | |
526 | * tasks, the expired queue will become the active queue after the active | |
527 | * queue is empty, without explicitly dequeuing and requeuing tasks in the | |
528 | * expired queue. (Interactive tasks may be requeued directly to the | |
529 | * active queue, thus delaying tasks in the expired queue from running; | |
530 | * see scheduler_tick()). | |
531 | * | |
532 | * This function is only called from sched_info_arrive(), rather than | |
533 | * dequeue_task(). Even though a task may be queued and dequeued multiple | |
534 | * times as it is shuffled about, we're really interested in knowing how | |
535 | * long it was from the *first* time it was queued to the time that it | |
536 | * finally hit a cpu. | |
537 | */ | |
538 | static inline void sched_info_dequeued(task_t *t) | |
539 | { | |
540 | t->sched_info.last_queued = 0; | |
541 | } | |
542 | ||
543 | /* | |
544 | * Called when a task finally hits the cpu. We can now calculate how | |
545 | * long it was waiting to run. We also note when it began so that we | |
546 | * can keep stats on how long its timeslice is. | |
547 | */ | |
858119e1 | 548 | static void sched_info_arrive(task_t *t) |
1da177e4 LT |
549 | { |
550 | unsigned long now = jiffies, diff = 0; | |
551 | struct runqueue *rq = task_rq(t); | |
552 | ||
553 | if (t->sched_info.last_queued) | |
554 | diff = now - t->sched_info.last_queued; | |
555 | sched_info_dequeued(t); | |
556 | t->sched_info.run_delay += diff; | |
557 | t->sched_info.last_arrival = now; | |
558 | t->sched_info.pcnt++; | |
559 | ||
560 | if (!rq) | |
561 | return; | |
562 | ||
563 | rq->rq_sched_info.run_delay += diff; | |
564 | rq->rq_sched_info.pcnt++; | |
565 | } | |
566 | ||
567 | /* | |
568 | * Called when a process is queued into either the active or expired | |
569 | * array. The time is noted and later used to determine how long we | |
570 | * had to wait for us to reach the cpu. Since the expired queue will | |
571 | * become the active queue after active queue is empty, without dequeuing | |
572 | * and requeuing any tasks, we are interested in queuing to either. It | |
573 | * is unusual but not impossible for tasks to be dequeued and immediately | |
574 | * requeued in the same or another array: this can happen in sched_yield(), | |
575 | * set_user_nice(), and even load_balance() as it moves tasks from runqueue | |
576 | * to runqueue. | |
577 | * | |
578 | * This function is only called from enqueue_task(), but also only updates | |
579 | * the timestamp if it is already not set. It's assumed that | |
580 | * sched_info_dequeued() will clear that stamp when appropriate. | |
581 | */ | |
582 | static inline void sched_info_queued(task_t *t) | |
583 | { | |
584 | if (!t->sched_info.last_queued) | |
585 | t->sched_info.last_queued = jiffies; | |
586 | } | |
587 | ||
588 | /* | |
589 | * Called when a process ceases being the active-running process, either | |
590 | * voluntarily or involuntarily. Now we can calculate how long we ran. | |
591 | */ | |
592 | static inline void sched_info_depart(task_t *t) | |
593 | { | |
594 | struct runqueue *rq = task_rq(t); | |
595 | unsigned long diff = jiffies - t->sched_info.last_arrival; | |
596 | ||
597 | t->sched_info.cpu_time += diff; | |
598 | ||
599 | if (rq) | |
600 | rq->rq_sched_info.cpu_time += diff; | |
601 | } | |
602 | ||
603 | /* | |
604 | * Called when tasks are switched involuntarily due, typically, to expiring | |
605 | * their time slice. (This may also be called when switching to or from | |
606 | * the idle task.) We are only called when prev != next. | |
607 | */ | |
608 | static inline void sched_info_switch(task_t *prev, task_t *next) | |
609 | { | |
610 | struct runqueue *rq = task_rq(prev); | |
611 | ||
612 | /* | |
613 | * prev now departs the cpu. It's not interesting to record | |
614 | * stats about how efficient we were at scheduling the idle | |
615 | * process, however. | |
616 | */ | |
617 | if (prev != rq->idle) | |
618 | sched_info_depart(prev); | |
619 | ||
620 | if (next != rq->idle) | |
621 | sched_info_arrive(next); | |
622 | } | |
623 | #else | |
624 | #define sched_info_queued(t) do { } while (0) | |
625 | #define sched_info_switch(t, next) do { } while (0) | |
626 | #endif /* CONFIG_SCHEDSTATS */ | |
627 | ||
628 | /* | |
629 | * Adding/removing a task to/from a priority array: | |
630 | */ | |
631 | static void dequeue_task(struct task_struct *p, prio_array_t *array) | |
632 | { | |
633 | array->nr_active--; | |
634 | list_del(&p->run_list); | |
635 | if (list_empty(array->queue + p->prio)) | |
636 | __clear_bit(p->prio, array->bitmap); | |
637 | } | |
638 | ||
639 | static void enqueue_task(struct task_struct *p, prio_array_t *array) | |
640 | { | |
641 | sched_info_queued(p); | |
642 | list_add_tail(&p->run_list, array->queue + p->prio); | |
643 | __set_bit(p->prio, array->bitmap); | |
644 | array->nr_active++; | |
645 | p->array = array; | |
646 | } | |
647 | ||
648 | /* | |
649 | * Put task to the end of the run list without the overhead of dequeue | |
650 | * followed by enqueue. | |
651 | */ | |
652 | static void requeue_task(struct task_struct *p, prio_array_t *array) | |
653 | { | |
654 | list_move_tail(&p->run_list, array->queue + p->prio); | |
655 | } | |
656 | ||
657 | static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array) | |
658 | { | |
659 | list_add(&p->run_list, array->queue + p->prio); | |
660 | __set_bit(p->prio, array->bitmap); | |
661 | array->nr_active++; | |
662 | p->array = array; | |
663 | } | |
664 | ||
665 | /* | |
b29739f9 | 666 | * __normal_prio - return the priority that is based on the static |
1da177e4 LT |
667 | * priority but is modified by bonuses/penalties. |
668 | * | |
669 | * We scale the actual sleep average [0 .... MAX_SLEEP_AVG] | |
670 | * into the -5 ... 0 ... +5 bonus/penalty range. | |
671 | * | |
672 | * We use 25% of the full 0...39 priority range so that: | |
673 | * | |
674 | * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs. | |
675 | * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks. | |
676 | * | |
677 | * Both properties are important to certain workloads. | |
678 | */ | |
b29739f9 IM |
679 | |
680 | static inline int __normal_prio(task_t *p) | |
1da177e4 LT |
681 | { |
682 | int bonus, prio; | |
683 | ||
1da177e4 LT |
684 | bonus = CURRENT_BONUS(p) - MAX_BONUS / 2; |
685 | ||
686 | prio = p->static_prio - bonus; | |
687 | if (prio < MAX_RT_PRIO) | |
688 | prio = MAX_RT_PRIO; | |
689 | if (prio > MAX_PRIO-1) | |
690 | prio = MAX_PRIO-1; | |
691 | return prio; | |
692 | } | |
693 | ||
2dd73a4f PW |
694 | /* |
695 | * To aid in avoiding the subversion of "niceness" due to uneven distribution | |
696 | * of tasks with abnormal "nice" values across CPUs the contribution that | |
697 | * each task makes to its run queue's load is weighted according to its | |
698 | * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a | |
699 | * scaled version of the new time slice allocation that they receive on time | |
700 | * slice expiry etc. | |
701 | */ | |
702 | ||
703 | /* | |
704 | * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE | |
705 | * If static_prio_timeslice() is ever changed to break this assumption then | |
706 | * this code will need modification | |
707 | */ | |
708 | #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE | |
709 | #define LOAD_WEIGHT(lp) \ | |
710 | (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO) | |
711 | #define PRIO_TO_LOAD_WEIGHT(prio) \ | |
712 | LOAD_WEIGHT(static_prio_timeslice(prio)) | |
713 | #define RTPRIO_TO_LOAD_WEIGHT(rp) \ | |
714 | (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp)) | |
715 | ||
716 | static void set_load_weight(task_t *p) | |
717 | { | |
b29739f9 | 718 | if (has_rt_policy(p)) { |
2dd73a4f PW |
719 | #ifdef CONFIG_SMP |
720 | if (p == task_rq(p)->migration_thread) | |
721 | /* | |
722 | * The migration thread does the actual balancing. | |
723 | * Giving its load any weight will skew balancing | |
724 | * adversely. | |
725 | */ | |
726 | p->load_weight = 0; | |
727 | else | |
728 | #endif | |
729 | p->load_weight = RTPRIO_TO_LOAD_WEIGHT(p->rt_priority); | |
730 | } else | |
731 | p->load_weight = PRIO_TO_LOAD_WEIGHT(p->static_prio); | |
732 | } | |
733 | ||
734 | static inline void inc_raw_weighted_load(runqueue_t *rq, const task_t *p) | |
735 | { | |
736 | rq->raw_weighted_load += p->load_weight; | |
737 | } | |
738 | ||
739 | static inline void dec_raw_weighted_load(runqueue_t *rq, const task_t *p) | |
740 | { | |
741 | rq->raw_weighted_load -= p->load_weight; | |
742 | } | |
743 | ||
744 | static inline void inc_nr_running(task_t *p, runqueue_t *rq) | |
745 | { | |
746 | rq->nr_running++; | |
747 | inc_raw_weighted_load(rq, p); | |
748 | } | |
749 | ||
750 | static inline void dec_nr_running(task_t *p, runqueue_t *rq) | |
751 | { | |
752 | rq->nr_running--; | |
753 | dec_raw_weighted_load(rq, p); | |
754 | } | |
755 | ||
b29739f9 IM |
756 | /* |
757 | * Calculate the expected normal priority: i.e. priority | |
758 | * without taking RT-inheritance into account. Might be | |
759 | * boosted by interactivity modifiers. Changes upon fork, | |
760 | * setprio syscalls, and whenever the interactivity | |
761 | * estimator recalculates. | |
762 | */ | |
763 | static inline int normal_prio(task_t *p) | |
764 | { | |
765 | int prio; | |
766 | ||
767 | if (has_rt_policy(p)) | |
768 | prio = MAX_RT_PRIO-1 - p->rt_priority; | |
769 | else | |
770 | prio = __normal_prio(p); | |
771 | return prio; | |
772 | } | |
773 | ||
774 | /* | |
775 | * Calculate the current priority, i.e. the priority | |
776 | * taken into account by the scheduler. This value might | |
777 | * be boosted by RT tasks, or might be boosted by | |
778 | * interactivity modifiers. Will be RT if the task got | |
779 | * RT-boosted. If not then it returns p->normal_prio. | |
780 | */ | |
781 | static int effective_prio(task_t *p) | |
782 | { | |
783 | p->normal_prio = normal_prio(p); | |
784 | /* | |
785 | * If we are RT tasks or we were boosted to RT priority, | |
786 | * keep the priority unchanged. Otherwise, update priority | |
787 | * to the normal priority: | |
788 | */ | |
789 | if (!rt_prio(p->prio)) | |
790 | return p->normal_prio; | |
791 | return p->prio; | |
792 | } | |
793 | ||
1da177e4 LT |
794 | /* |
795 | * __activate_task - move a task to the runqueue. | |
796 | */ | |
d425b274 | 797 | static void __activate_task(task_t *p, runqueue_t *rq) |
1da177e4 | 798 | { |
d425b274 CK |
799 | prio_array_t *target = rq->active; |
800 | ||
f1adad78 | 801 | if (batch_task(p)) |
d425b274 CK |
802 | target = rq->expired; |
803 | enqueue_task(p, target); | |
2dd73a4f | 804 | inc_nr_running(p, rq); |
1da177e4 LT |
805 | } |
806 | ||
807 | /* | |
808 | * __activate_idle_task - move idle task to the _front_ of runqueue. | |
809 | */ | |
810 | static inline void __activate_idle_task(task_t *p, runqueue_t *rq) | |
811 | { | |
812 | enqueue_task_head(p, rq->active); | |
2dd73a4f | 813 | inc_nr_running(p, rq); |
1da177e4 LT |
814 | } |
815 | ||
b29739f9 IM |
816 | /* |
817 | * Recalculate p->normal_prio and p->prio after having slept, | |
818 | * updating the sleep-average too: | |
819 | */ | |
a3464a10 | 820 | static int recalc_task_prio(task_t *p, unsigned long long now) |
1da177e4 LT |
821 | { |
822 | /* Caller must always ensure 'now >= p->timestamp' */ | |
72d2854d | 823 | unsigned long sleep_time = now - p->timestamp; |
1da177e4 | 824 | |
d425b274 | 825 | if (batch_task(p)) |
b0a9499c | 826 | sleep_time = 0; |
1da177e4 LT |
827 | |
828 | if (likely(sleep_time > 0)) { | |
829 | /* | |
72d2854d CK |
830 | * This ceiling is set to the lowest priority that would allow |
831 | * a task to be reinserted into the active array on timeslice | |
832 | * completion. | |
1da177e4 | 833 | */ |
72d2854d | 834 | unsigned long ceiling = INTERACTIVE_SLEEP(p); |
e72ff0bb | 835 | |
72d2854d CK |
836 | if (p->mm && sleep_time > ceiling && p->sleep_avg < ceiling) { |
837 | /* | |
838 | * Prevents user tasks from achieving best priority | |
839 | * with one single large enough sleep. | |
840 | */ | |
841 | p->sleep_avg = ceiling; | |
842 | /* | |
843 | * Using INTERACTIVE_SLEEP() as a ceiling places a | |
844 | * nice(0) task 1ms sleep away from promotion, and | |
845 | * gives it 700ms to round-robin with no chance of | |
846 | * being demoted. This is more than generous, so | |
847 | * mark this sleep as non-interactive to prevent the | |
848 | * on-runqueue bonus logic from intervening should | |
849 | * this task not receive cpu immediately. | |
850 | */ | |
851 | p->sleep_type = SLEEP_NONINTERACTIVE; | |
1da177e4 | 852 | } else { |
1da177e4 LT |
853 | /* |
854 | * Tasks waking from uninterruptible sleep are | |
855 | * limited in their sleep_avg rise as they | |
856 | * are likely to be waiting on I/O | |
857 | */ | |
3dee386e | 858 | if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) { |
72d2854d | 859 | if (p->sleep_avg >= ceiling) |
1da177e4 LT |
860 | sleep_time = 0; |
861 | else if (p->sleep_avg + sleep_time >= | |
72d2854d CK |
862 | ceiling) { |
863 | p->sleep_avg = ceiling; | |
864 | sleep_time = 0; | |
1da177e4 LT |
865 | } |
866 | } | |
867 | ||
868 | /* | |
869 | * This code gives a bonus to interactive tasks. | |
870 | * | |
871 | * The boost works by updating the 'average sleep time' | |
872 | * value here, based on ->timestamp. The more time a | |
873 | * task spends sleeping, the higher the average gets - | |
874 | * and the higher the priority boost gets as well. | |
875 | */ | |
876 | p->sleep_avg += sleep_time; | |
877 | ||
1da177e4 | 878 | } |
72d2854d CK |
879 | if (p->sleep_avg > NS_MAX_SLEEP_AVG) |
880 | p->sleep_avg = NS_MAX_SLEEP_AVG; | |
1da177e4 LT |
881 | } |
882 | ||
a3464a10 | 883 | return effective_prio(p); |
1da177e4 LT |
884 | } |
885 | ||
886 | /* | |
887 | * activate_task - move a task to the runqueue and do priority recalculation | |
888 | * | |
889 | * Update all the scheduling statistics stuff. (sleep average | |
890 | * calculation, priority modifiers, etc.) | |
891 | */ | |
892 | static void activate_task(task_t *p, runqueue_t *rq, int local) | |
893 | { | |
894 | unsigned long long now; | |
895 | ||
896 | now = sched_clock(); | |
897 | #ifdef CONFIG_SMP | |
898 | if (!local) { | |
899 | /* Compensate for drifting sched_clock */ | |
900 | runqueue_t *this_rq = this_rq(); | |
901 | now = (now - this_rq->timestamp_last_tick) | |
902 | + rq->timestamp_last_tick; | |
903 | } | |
904 | #endif | |
905 | ||
a47ab937 CK |
906 | if (!rt_task(p)) |
907 | p->prio = recalc_task_prio(p, now); | |
1da177e4 LT |
908 | |
909 | /* | |
910 | * This checks to make sure it's not an uninterruptible task | |
911 | * that is now waking up. | |
912 | */ | |
3dee386e | 913 | if (p->sleep_type == SLEEP_NORMAL) { |
1da177e4 LT |
914 | /* |
915 | * Tasks which were woken up by interrupts (ie. hw events) | |
916 | * are most likely of interactive nature. So we give them | |
917 | * the credit of extending their sleep time to the period | |
918 | * of time they spend on the runqueue, waiting for execution | |
919 | * on a CPU, first time around: | |
920 | */ | |
921 | if (in_interrupt()) | |
3dee386e | 922 | p->sleep_type = SLEEP_INTERRUPTED; |
1da177e4 LT |
923 | else { |
924 | /* | |
925 | * Normal first-time wakeups get a credit too for | |
926 | * on-runqueue time, but it will be weighted down: | |
927 | */ | |
3dee386e | 928 | p->sleep_type = SLEEP_INTERACTIVE; |
1da177e4 LT |
929 | } |
930 | } | |
931 | p->timestamp = now; | |
932 | ||
933 | __activate_task(p, rq); | |
934 | } | |
935 | ||
936 | /* | |
937 | * deactivate_task - remove a task from the runqueue. | |
938 | */ | |
939 | static void deactivate_task(struct task_struct *p, runqueue_t *rq) | |
940 | { | |
2dd73a4f | 941 | dec_nr_running(p, rq); |
1da177e4 LT |
942 | dequeue_task(p, p->array); |
943 | p->array = NULL; | |
944 | } | |
945 | ||
946 | /* | |
947 | * resched_task - mark a task 'to be rescheduled now'. | |
948 | * | |
949 | * On UP this means the setting of the need_resched flag, on SMP it | |
950 | * might also involve a cross-CPU call to trigger the scheduler on | |
951 | * the target CPU. | |
952 | */ | |
953 | #ifdef CONFIG_SMP | |
495ab9c0 AK |
954 | |
955 | #ifndef tsk_is_polling | |
956 | #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG) | |
957 | #endif | |
958 | ||
1da177e4 LT |
959 | static void resched_task(task_t *p) |
960 | { | |
64c7c8f8 | 961 | int cpu; |
1da177e4 LT |
962 | |
963 | assert_spin_locked(&task_rq(p)->lock); | |
964 | ||
64c7c8f8 NP |
965 | if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED))) |
966 | return; | |
967 | ||
968 | set_tsk_thread_flag(p, TIF_NEED_RESCHED); | |
1da177e4 | 969 | |
64c7c8f8 NP |
970 | cpu = task_cpu(p); |
971 | if (cpu == smp_processor_id()) | |
972 | return; | |
973 | ||
495ab9c0 | 974 | /* NEED_RESCHED must be visible before we test polling */ |
64c7c8f8 | 975 | smp_mb(); |
495ab9c0 | 976 | if (!tsk_is_polling(p)) |
64c7c8f8 | 977 | smp_send_reschedule(cpu); |
1da177e4 LT |
978 | } |
979 | #else | |
980 | static inline void resched_task(task_t *p) | |
981 | { | |
64c7c8f8 | 982 | assert_spin_locked(&task_rq(p)->lock); |
1da177e4 LT |
983 | set_tsk_need_resched(p); |
984 | } | |
985 | #endif | |
986 | ||
987 | /** | |
988 | * task_curr - is this task currently executing on a CPU? | |
989 | * @p: the task in question. | |
990 | */ | |
991 | inline int task_curr(const task_t *p) | |
992 | { | |
993 | return cpu_curr(task_cpu(p)) == p; | |
994 | } | |
995 | ||
2dd73a4f PW |
996 | /* Used instead of source_load when we know the type == 0 */ |
997 | unsigned long weighted_cpuload(const int cpu) | |
998 | { | |
999 | return cpu_rq(cpu)->raw_weighted_load; | |
1000 | } | |
1001 | ||
1da177e4 | 1002 | #ifdef CONFIG_SMP |
1da177e4 LT |
1003 | typedef struct { |
1004 | struct list_head list; | |
1da177e4 | 1005 | |
1da177e4 LT |
1006 | task_t *task; |
1007 | int dest_cpu; | |
1008 | ||
1da177e4 LT |
1009 | struct completion done; |
1010 | } migration_req_t; | |
1011 | ||
1012 | /* | |
1013 | * The task's runqueue lock must be held. | |
1014 | * Returns true if you have to wait for migration thread. | |
1015 | */ | |
1016 | static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req) | |
1017 | { | |
1018 | runqueue_t *rq = task_rq(p); | |
1019 | ||
1020 | /* | |
1021 | * If the task is not on a runqueue (and not running), then | |
1022 | * it is sufficient to simply update the task's cpu field. | |
1023 | */ | |
1024 | if (!p->array && !task_running(rq, p)) { | |
1025 | set_task_cpu(p, dest_cpu); | |
1026 | return 0; | |
1027 | } | |
1028 | ||
1029 | init_completion(&req->done); | |
1da177e4 LT |
1030 | req->task = p; |
1031 | req->dest_cpu = dest_cpu; | |
1032 | list_add(&req->list, &rq->migration_queue); | |
1033 | return 1; | |
1034 | } | |
1035 | ||
1036 | /* | |
1037 | * wait_task_inactive - wait for a thread to unschedule. | |
1038 | * | |
1039 | * The caller must ensure that the task *will* unschedule sometime soon, | |
1040 | * else this function might spin for a *long* time. This function can't | |
1041 | * be called with interrupts off, or it may introduce deadlock with | |
1042 | * smp_call_function() if an IPI is sent by the same process we are | |
1043 | * waiting to become inactive. | |
1044 | */ | |
95cdf3b7 | 1045 | void wait_task_inactive(task_t *p) |
1da177e4 LT |
1046 | { |
1047 | unsigned long flags; | |
1048 | runqueue_t *rq; | |
1049 | int preempted; | |
1050 | ||
1051 | repeat: | |
1052 | rq = task_rq_lock(p, &flags); | |
1053 | /* Must be off runqueue entirely, not preempted. */ | |
1054 | if (unlikely(p->array || task_running(rq, p))) { | |
1055 | /* If it's preempted, we yield. It could be a while. */ | |
1056 | preempted = !task_running(rq, p); | |
1057 | task_rq_unlock(rq, &flags); | |
1058 | cpu_relax(); | |
1059 | if (preempted) | |
1060 | yield(); | |
1061 | goto repeat; | |
1062 | } | |
1063 | task_rq_unlock(rq, &flags); | |
1064 | } | |
1065 | ||
1066 | /*** | |
1067 | * kick_process - kick a running thread to enter/exit the kernel | |
1068 | * @p: the to-be-kicked thread | |
1069 | * | |
1070 | * Cause a process which is running on another CPU to enter | |
1071 | * kernel-mode, without any delay. (to get signals handled.) | |
1072 | * | |
1073 | * NOTE: this function doesnt have to take the runqueue lock, | |
1074 | * because all it wants to ensure is that the remote task enters | |
1075 | * the kernel. If the IPI races and the task has been migrated | |
1076 | * to another CPU then no harm is done and the purpose has been | |
1077 | * achieved as well. | |
1078 | */ | |
1079 | void kick_process(task_t *p) | |
1080 | { | |
1081 | int cpu; | |
1082 | ||
1083 | preempt_disable(); | |
1084 | cpu = task_cpu(p); | |
1085 | if ((cpu != smp_processor_id()) && task_curr(p)) | |
1086 | smp_send_reschedule(cpu); | |
1087 | preempt_enable(); | |
1088 | } | |
1089 | ||
1090 | /* | |
2dd73a4f PW |
1091 | * Return a low guess at the load of a migration-source cpu weighted |
1092 | * according to the scheduling class and "nice" value. | |
1da177e4 LT |
1093 | * |
1094 | * We want to under-estimate the load of migration sources, to | |
1095 | * balance conservatively. | |
1096 | */ | |
a2000572 | 1097 | static inline unsigned long source_load(int cpu, int type) |
1da177e4 LT |
1098 | { |
1099 | runqueue_t *rq = cpu_rq(cpu); | |
2dd73a4f | 1100 | |
3b0bd9bc | 1101 | if (type == 0) |
2dd73a4f | 1102 | return rq->raw_weighted_load; |
b910472d | 1103 | |
2dd73a4f | 1104 | return min(rq->cpu_load[type-1], rq->raw_weighted_load); |
1da177e4 LT |
1105 | } |
1106 | ||
1107 | /* | |
2dd73a4f PW |
1108 | * Return a high guess at the load of a migration-target cpu weighted |
1109 | * according to the scheduling class and "nice" value. | |
1da177e4 | 1110 | */ |
a2000572 | 1111 | static inline unsigned long target_load(int cpu, int type) |
1da177e4 LT |
1112 | { |
1113 | runqueue_t *rq = cpu_rq(cpu); | |
2dd73a4f | 1114 | |
7897986b | 1115 | if (type == 0) |
2dd73a4f | 1116 | return rq->raw_weighted_load; |
3b0bd9bc | 1117 | |
2dd73a4f PW |
1118 | return max(rq->cpu_load[type-1], rq->raw_weighted_load); |
1119 | } | |
1120 | ||
1121 | /* | |
1122 | * Return the average load per task on the cpu's run queue | |
1123 | */ | |
1124 | static inline unsigned long cpu_avg_load_per_task(int cpu) | |
1125 | { | |
1126 | runqueue_t *rq = cpu_rq(cpu); | |
1127 | unsigned long n = rq->nr_running; | |
1128 | ||
1129 | return n ? rq->raw_weighted_load / n : SCHED_LOAD_SCALE; | |
1da177e4 LT |
1130 | } |
1131 | ||
147cbb4b NP |
1132 | /* |
1133 | * find_idlest_group finds and returns the least busy CPU group within the | |
1134 | * domain. | |
1135 | */ | |
1136 | static struct sched_group * | |
1137 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu) | |
1138 | { | |
1139 | struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups; | |
1140 | unsigned long min_load = ULONG_MAX, this_load = 0; | |
1141 | int load_idx = sd->forkexec_idx; | |
1142 | int imbalance = 100 + (sd->imbalance_pct-100)/2; | |
1143 | ||
1144 | do { | |
1145 | unsigned long load, avg_load; | |
1146 | int local_group; | |
1147 | int i; | |
1148 | ||
da5a5522 BD |
1149 | /* Skip over this group if it has no CPUs allowed */ |
1150 | if (!cpus_intersects(group->cpumask, p->cpus_allowed)) | |
1151 | goto nextgroup; | |
1152 | ||
147cbb4b | 1153 | local_group = cpu_isset(this_cpu, group->cpumask); |
147cbb4b NP |
1154 | |
1155 | /* Tally up the load of all CPUs in the group */ | |
1156 | avg_load = 0; | |
1157 | ||
1158 | for_each_cpu_mask(i, group->cpumask) { | |
1159 | /* Bias balancing toward cpus of our domain */ | |
1160 | if (local_group) | |
1161 | load = source_load(i, load_idx); | |
1162 | else | |
1163 | load = target_load(i, load_idx); | |
1164 | ||
1165 | avg_load += load; | |
1166 | } | |
1167 | ||
1168 | /* Adjust by relative CPU power of the group */ | |
1169 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
1170 | ||
1171 | if (local_group) { | |
1172 | this_load = avg_load; | |
1173 | this = group; | |
1174 | } else if (avg_load < min_load) { | |
1175 | min_load = avg_load; | |
1176 | idlest = group; | |
1177 | } | |
da5a5522 | 1178 | nextgroup: |
147cbb4b NP |
1179 | group = group->next; |
1180 | } while (group != sd->groups); | |
1181 | ||
1182 | if (!idlest || 100*this_load < imbalance*min_load) | |
1183 | return NULL; | |
1184 | return idlest; | |
1185 | } | |
1186 | ||
1187 | /* | |
1188 | * find_idlest_queue - find the idlest runqueue among the cpus in group. | |
1189 | */ | |
95cdf3b7 IM |
1190 | static int |
1191 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
147cbb4b | 1192 | { |
da5a5522 | 1193 | cpumask_t tmp; |
147cbb4b NP |
1194 | unsigned long load, min_load = ULONG_MAX; |
1195 | int idlest = -1; | |
1196 | int i; | |
1197 | ||
da5a5522 BD |
1198 | /* Traverse only the allowed CPUs */ |
1199 | cpus_and(tmp, group->cpumask, p->cpus_allowed); | |
1200 | ||
1201 | for_each_cpu_mask(i, tmp) { | |
2dd73a4f | 1202 | load = weighted_cpuload(i); |
147cbb4b NP |
1203 | |
1204 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
1205 | min_load = load; | |
1206 | idlest = i; | |
1207 | } | |
1208 | } | |
1209 | ||
1210 | return idlest; | |
1211 | } | |
1212 | ||
476d139c NP |
1213 | /* |
1214 | * sched_balance_self: balance the current task (running on cpu) in domains | |
1215 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
1216 | * SD_BALANCE_EXEC. | |
1217 | * | |
1218 | * Balance, ie. select the least loaded group. | |
1219 | * | |
1220 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
1221 | * | |
1222 | * preempt must be disabled. | |
1223 | */ | |
1224 | static int sched_balance_self(int cpu, int flag) | |
1225 | { | |
1226 | struct task_struct *t = current; | |
1227 | struct sched_domain *tmp, *sd = NULL; | |
147cbb4b | 1228 | |
c96d145e | 1229 | for_each_domain(cpu, tmp) { |
5c45bf27 SS |
1230 | /* |
1231 | * If power savings logic is enabled for a domain, stop there. | |
1232 | */ | |
1233 | if (tmp->flags & SD_POWERSAVINGS_BALANCE) | |
1234 | break; | |
476d139c NP |
1235 | if (tmp->flags & flag) |
1236 | sd = tmp; | |
c96d145e | 1237 | } |
476d139c NP |
1238 | |
1239 | while (sd) { | |
1240 | cpumask_t span; | |
1241 | struct sched_group *group; | |
1242 | int new_cpu; | |
1243 | int weight; | |
1244 | ||
1245 | span = sd->span; | |
1246 | group = find_idlest_group(sd, t, cpu); | |
1247 | if (!group) | |
1248 | goto nextlevel; | |
1249 | ||
da5a5522 | 1250 | new_cpu = find_idlest_cpu(group, t, cpu); |
476d139c NP |
1251 | if (new_cpu == -1 || new_cpu == cpu) |
1252 | goto nextlevel; | |
1253 | ||
1254 | /* Now try balancing at a lower domain level */ | |
1255 | cpu = new_cpu; | |
1256 | nextlevel: | |
1257 | sd = NULL; | |
1258 | weight = cpus_weight(span); | |
1259 | for_each_domain(cpu, tmp) { | |
1260 | if (weight <= cpus_weight(tmp->span)) | |
1261 | break; | |
1262 | if (tmp->flags & flag) | |
1263 | sd = tmp; | |
1264 | } | |
1265 | /* while loop will break here if sd == NULL */ | |
1266 | } | |
1267 | ||
1268 | return cpu; | |
1269 | } | |
1270 | ||
1271 | #endif /* CONFIG_SMP */ | |
1da177e4 LT |
1272 | |
1273 | /* | |
1274 | * wake_idle() will wake a task on an idle cpu if task->cpu is | |
1275 | * not idle and an idle cpu is available. The span of cpus to | |
1276 | * search starts with cpus closest then further out as needed, | |
1277 | * so we always favor a closer, idle cpu. | |
1278 | * | |
1279 | * Returns the CPU we should wake onto. | |
1280 | */ | |
1281 | #if defined(ARCH_HAS_SCHED_WAKE_IDLE) | |
1282 | static int wake_idle(int cpu, task_t *p) | |
1283 | { | |
1284 | cpumask_t tmp; | |
1285 | struct sched_domain *sd; | |
1286 | int i; | |
1287 | ||
1288 | if (idle_cpu(cpu)) | |
1289 | return cpu; | |
1290 | ||
1291 | for_each_domain(cpu, sd) { | |
1292 | if (sd->flags & SD_WAKE_IDLE) { | |
e0f364f4 | 1293 | cpus_and(tmp, sd->span, p->cpus_allowed); |
1da177e4 LT |
1294 | for_each_cpu_mask(i, tmp) { |
1295 | if (idle_cpu(i)) | |
1296 | return i; | |
1297 | } | |
1298 | } | |
e0f364f4 NP |
1299 | else |
1300 | break; | |
1da177e4 LT |
1301 | } |
1302 | return cpu; | |
1303 | } | |
1304 | #else | |
1305 | static inline int wake_idle(int cpu, task_t *p) | |
1306 | { | |
1307 | return cpu; | |
1308 | } | |
1309 | #endif | |
1310 | ||
1311 | /*** | |
1312 | * try_to_wake_up - wake up a thread | |
1313 | * @p: the to-be-woken-up thread | |
1314 | * @state: the mask of task states that can be woken | |
1315 | * @sync: do a synchronous wakeup? | |
1316 | * | |
1317 | * Put it on the run-queue if it's not already there. The "current" | |
1318 | * thread is always on the run-queue (except when the actual | |
1319 | * re-schedule is in progress), and as such you're allowed to do | |
1320 | * the simpler "current->state = TASK_RUNNING" to mark yourself | |
1321 | * runnable without the overhead of this. | |
1322 | * | |
1323 | * returns failure only if the task is already active. | |
1324 | */ | |
95cdf3b7 | 1325 | static int try_to_wake_up(task_t *p, unsigned int state, int sync) |
1da177e4 LT |
1326 | { |
1327 | int cpu, this_cpu, success = 0; | |
1328 | unsigned long flags; | |
1329 | long old_state; | |
1330 | runqueue_t *rq; | |
1331 | #ifdef CONFIG_SMP | |
1332 | unsigned long load, this_load; | |
7897986b | 1333 | struct sched_domain *sd, *this_sd = NULL; |
1da177e4 LT |
1334 | int new_cpu; |
1335 | #endif | |
1336 | ||
1337 | rq = task_rq_lock(p, &flags); | |
1338 | old_state = p->state; | |
1339 | if (!(old_state & state)) | |
1340 | goto out; | |
1341 | ||
1342 | if (p->array) | |
1343 | goto out_running; | |
1344 | ||
1345 | cpu = task_cpu(p); | |
1346 | this_cpu = smp_processor_id(); | |
1347 | ||
1348 | #ifdef CONFIG_SMP | |
1349 | if (unlikely(task_running(rq, p))) | |
1350 | goto out_activate; | |
1351 | ||
7897986b NP |
1352 | new_cpu = cpu; |
1353 | ||
1da177e4 LT |
1354 | schedstat_inc(rq, ttwu_cnt); |
1355 | if (cpu == this_cpu) { | |
1356 | schedstat_inc(rq, ttwu_local); | |
7897986b NP |
1357 | goto out_set_cpu; |
1358 | } | |
1359 | ||
1360 | for_each_domain(this_cpu, sd) { | |
1361 | if (cpu_isset(cpu, sd->span)) { | |
1362 | schedstat_inc(sd, ttwu_wake_remote); | |
1363 | this_sd = sd; | |
1364 | break; | |
1da177e4 LT |
1365 | } |
1366 | } | |
1da177e4 | 1367 | |
7897986b | 1368 | if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed))) |
1da177e4 LT |
1369 | goto out_set_cpu; |
1370 | ||
1da177e4 | 1371 | /* |
7897986b | 1372 | * Check for affine wakeup and passive balancing possibilities. |
1da177e4 | 1373 | */ |
7897986b NP |
1374 | if (this_sd) { |
1375 | int idx = this_sd->wake_idx; | |
1376 | unsigned int imbalance; | |
1da177e4 | 1377 | |
a3f21bce NP |
1378 | imbalance = 100 + (this_sd->imbalance_pct - 100) / 2; |
1379 | ||
7897986b NP |
1380 | load = source_load(cpu, idx); |
1381 | this_load = target_load(this_cpu, idx); | |
1da177e4 | 1382 | |
7897986b NP |
1383 | new_cpu = this_cpu; /* Wake to this CPU if we can */ |
1384 | ||
a3f21bce NP |
1385 | if (this_sd->flags & SD_WAKE_AFFINE) { |
1386 | unsigned long tl = this_load; | |
2dd73a4f PW |
1387 | unsigned long tl_per_task = cpu_avg_load_per_task(this_cpu); |
1388 | ||
1da177e4 | 1389 | /* |
a3f21bce NP |
1390 | * If sync wakeup then subtract the (maximum possible) |
1391 | * effect of the currently running task from the load | |
1392 | * of the current CPU: | |
1da177e4 | 1393 | */ |
a3f21bce | 1394 | if (sync) |
2dd73a4f | 1395 | tl -= current->load_weight; |
a3f21bce NP |
1396 | |
1397 | if ((tl <= load && | |
2dd73a4f PW |
1398 | tl + target_load(cpu, idx) <= tl_per_task) || |
1399 | 100*(tl + p->load_weight) <= imbalance*load) { | |
a3f21bce NP |
1400 | /* |
1401 | * This domain has SD_WAKE_AFFINE and | |
1402 | * p is cache cold in this domain, and | |
1403 | * there is no bad imbalance. | |
1404 | */ | |
1405 | schedstat_inc(this_sd, ttwu_move_affine); | |
1406 | goto out_set_cpu; | |
1407 | } | |
1408 | } | |
1409 | ||
1410 | /* | |
1411 | * Start passive balancing when half the imbalance_pct | |
1412 | * limit is reached. | |
1413 | */ | |
1414 | if (this_sd->flags & SD_WAKE_BALANCE) { | |
1415 | if (imbalance*this_load <= 100*load) { | |
1416 | schedstat_inc(this_sd, ttwu_move_balance); | |
1417 | goto out_set_cpu; | |
1418 | } | |
1da177e4 LT |
1419 | } |
1420 | } | |
1421 | ||
1422 | new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */ | |
1423 | out_set_cpu: | |
1424 | new_cpu = wake_idle(new_cpu, p); | |
1425 | if (new_cpu != cpu) { | |
1426 | set_task_cpu(p, new_cpu); | |
1427 | task_rq_unlock(rq, &flags); | |
1428 | /* might preempt at this point */ | |
1429 | rq = task_rq_lock(p, &flags); | |
1430 | old_state = p->state; | |
1431 | if (!(old_state & state)) | |
1432 | goto out; | |
1433 | if (p->array) | |
1434 | goto out_running; | |
1435 | ||
1436 | this_cpu = smp_processor_id(); | |
1437 | cpu = task_cpu(p); | |
1438 | } | |
1439 | ||
1440 | out_activate: | |
1441 | #endif /* CONFIG_SMP */ | |
1442 | if (old_state == TASK_UNINTERRUPTIBLE) { | |
1443 | rq->nr_uninterruptible--; | |
1444 | /* | |
1445 | * Tasks on involuntary sleep don't earn | |
1446 | * sleep_avg beyond just interactive state. | |
1447 | */ | |
3dee386e | 1448 | p->sleep_type = SLEEP_NONINTERACTIVE; |
e7c38cb4 | 1449 | } else |
1da177e4 | 1450 | |
d79fc0fc IM |
1451 | /* |
1452 | * Tasks that have marked their sleep as noninteractive get | |
e7c38cb4 CK |
1453 | * woken up with their sleep average not weighted in an |
1454 | * interactive way. | |
d79fc0fc | 1455 | */ |
e7c38cb4 CK |
1456 | if (old_state & TASK_NONINTERACTIVE) |
1457 | p->sleep_type = SLEEP_NONINTERACTIVE; | |
1458 | ||
1459 | ||
1460 | activate_task(p, rq, cpu == this_cpu); | |
1da177e4 LT |
1461 | /* |
1462 | * Sync wakeups (i.e. those types of wakeups where the waker | |
1463 | * has indicated that it will leave the CPU in short order) | |
1464 | * don't trigger a preemption, if the woken up task will run on | |
1465 | * this cpu. (in this case the 'I will reschedule' promise of | |
1466 | * the waker guarantees that the freshly woken up task is going | |
1467 | * to be considered on this CPU.) | |
1468 | */ | |
1da177e4 LT |
1469 | if (!sync || cpu != this_cpu) { |
1470 | if (TASK_PREEMPTS_CURR(p, rq)) | |
1471 | resched_task(rq->curr); | |
1472 | } | |
1473 | success = 1; | |
1474 | ||
1475 | out_running: | |
1476 | p->state = TASK_RUNNING; | |
1477 | out: | |
1478 | task_rq_unlock(rq, &flags); | |
1479 | ||
1480 | return success; | |
1481 | } | |
1482 | ||
95cdf3b7 | 1483 | int fastcall wake_up_process(task_t *p) |
1da177e4 LT |
1484 | { |
1485 | return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED | | |
1486 | TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0); | |
1487 | } | |
1488 | ||
1489 | EXPORT_SYMBOL(wake_up_process); | |
1490 | ||
1491 | int fastcall wake_up_state(task_t *p, unsigned int state) | |
1492 | { | |
1493 | return try_to_wake_up(p, state, 0); | |
1494 | } | |
1495 | ||
1da177e4 LT |
1496 | /* |
1497 | * Perform scheduler related setup for a newly forked process p. | |
1498 | * p is forked by current. | |
1499 | */ | |
476d139c | 1500 | void fastcall sched_fork(task_t *p, int clone_flags) |
1da177e4 | 1501 | { |
476d139c NP |
1502 | int cpu = get_cpu(); |
1503 | ||
1504 | #ifdef CONFIG_SMP | |
1505 | cpu = sched_balance_self(cpu, SD_BALANCE_FORK); | |
1506 | #endif | |
1507 | set_task_cpu(p, cpu); | |
1508 | ||
1da177e4 LT |
1509 | /* |
1510 | * We mark the process as running here, but have not actually | |
1511 | * inserted it onto the runqueue yet. This guarantees that | |
1512 | * nobody will actually run it, and a signal or other external | |
1513 | * event cannot wake it up and insert it on the runqueue either. | |
1514 | */ | |
1515 | p->state = TASK_RUNNING; | |
b29739f9 IM |
1516 | |
1517 | /* | |
1518 | * Make sure we do not leak PI boosting priority to the child: | |
1519 | */ | |
1520 | p->prio = current->normal_prio; | |
1521 | ||
1da177e4 LT |
1522 | INIT_LIST_HEAD(&p->run_list); |
1523 | p->array = NULL; | |
1da177e4 LT |
1524 | #ifdef CONFIG_SCHEDSTATS |
1525 | memset(&p->sched_info, 0, sizeof(p->sched_info)); | |
1526 | #endif | |
d6077cb8 | 1527 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
4866cde0 NP |
1528 | p->oncpu = 0; |
1529 | #endif | |
1da177e4 | 1530 | #ifdef CONFIG_PREEMPT |
4866cde0 | 1531 | /* Want to start with kernel preemption disabled. */ |
a1261f54 | 1532 | task_thread_info(p)->preempt_count = 1; |
1da177e4 LT |
1533 | #endif |
1534 | /* | |
1535 | * Share the timeslice between parent and child, thus the | |
1536 | * total amount of pending timeslices in the system doesn't change, | |
1537 | * resulting in more scheduling fairness. | |
1538 | */ | |
1539 | local_irq_disable(); | |
1540 | p->time_slice = (current->time_slice + 1) >> 1; | |
1541 | /* | |
1542 | * The remainder of the first timeslice might be recovered by | |
1543 | * the parent if the child exits early enough. | |
1544 | */ | |
1545 | p->first_time_slice = 1; | |
1546 | current->time_slice >>= 1; | |
1547 | p->timestamp = sched_clock(); | |
1548 | if (unlikely(!current->time_slice)) { | |
1549 | /* | |
1550 | * This case is rare, it happens when the parent has only | |
1551 | * a single jiffy left from its timeslice. Taking the | |
1552 | * runqueue lock is not a problem. | |
1553 | */ | |
1554 | current->time_slice = 1; | |
1da177e4 | 1555 | scheduler_tick(); |
476d139c NP |
1556 | } |
1557 | local_irq_enable(); | |
1558 | put_cpu(); | |
1da177e4 LT |
1559 | } |
1560 | ||
1561 | /* | |
1562 | * wake_up_new_task - wake up a newly created task for the first time. | |
1563 | * | |
1564 | * This function will do some initial scheduler statistics housekeeping | |
1565 | * that must be done for every newly created context, then puts the task | |
1566 | * on the runqueue and wakes it. | |
1567 | */ | |
95cdf3b7 | 1568 | void fastcall wake_up_new_task(task_t *p, unsigned long clone_flags) |
1da177e4 LT |
1569 | { |
1570 | unsigned long flags; | |
1571 | int this_cpu, cpu; | |
1572 | runqueue_t *rq, *this_rq; | |
1573 | ||
1574 | rq = task_rq_lock(p, &flags); | |
147cbb4b | 1575 | BUG_ON(p->state != TASK_RUNNING); |
1da177e4 | 1576 | this_cpu = smp_processor_id(); |
147cbb4b | 1577 | cpu = task_cpu(p); |
1da177e4 | 1578 | |
1da177e4 LT |
1579 | /* |
1580 | * We decrease the sleep average of forking parents | |
1581 | * and children as well, to keep max-interactive tasks | |
1582 | * from forking tasks that are max-interactive. The parent | |
1583 | * (current) is done further down, under its lock. | |
1584 | */ | |
1585 | p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) * | |
1586 | CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS); | |
1587 | ||
1588 | p->prio = effective_prio(p); | |
1589 | ||
1590 | if (likely(cpu == this_cpu)) { | |
1591 | if (!(clone_flags & CLONE_VM)) { | |
1592 | /* | |
1593 | * The VM isn't cloned, so we're in a good position to | |
1594 | * do child-runs-first in anticipation of an exec. This | |
1595 | * usually avoids a lot of COW overhead. | |
1596 | */ | |
1597 | if (unlikely(!current->array)) | |
1598 | __activate_task(p, rq); | |
1599 | else { | |
1600 | p->prio = current->prio; | |
b29739f9 | 1601 | p->normal_prio = current->normal_prio; |
1da177e4 LT |
1602 | list_add_tail(&p->run_list, ¤t->run_list); |
1603 | p->array = current->array; | |
1604 | p->array->nr_active++; | |
2dd73a4f | 1605 | inc_nr_running(p, rq); |
1da177e4 LT |
1606 | } |
1607 | set_need_resched(); | |
1608 | } else | |
1609 | /* Run child last */ | |
1610 | __activate_task(p, rq); | |
1611 | /* | |
1612 | * We skip the following code due to cpu == this_cpu | |
1613 | * | |
1614 | * task_rq_unlock(rq, &flags); | |
1615 | * this_rq = task_rq_lock(current, &flags); | |
1616 | */ | |
1617 | this_rq = rq; | |
1618 | } else { | |
1619 | this_rq = cpu_rq(this_cpu); | |
1620 | ||
1621 | /* | |
1622 | * Not the local CPU - must adjust timestamp. This should | |
1623 | * get optimised away in the !CONFIG_SMP case. | |
1624 | */ | |
1625 | p->timestamp = (p->timestamp - this_rq->timestamp_last_tick) | |
1626 | + rq->timestamp_last_tick; | |
1627 | __activate_task(p, rq); | |
1628 | if (TASK_PREEMPTS_CURR(p, rq)) | |
1629 | resched_task(rq->curr); | |
1630 | ||
1631 | /* | |
1632 | * Parent and child are on different CPUs, now get the | |
1633 | * parent runqueue to update the parent's ->sleep_avg: | |
1634 | */ | |
1635 | task_rq_unlock(rq, &flags); | |
1636 | this_rq = task_rq_lock(current, &flags); | |
1637 | } | |
1638 | current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) * | |
1639 | PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS); | |
1640 | task_rq_unlock(this_rq, &flags); | |
1641 | } | |
1642 | ||
1643 | /* | |
1644 | * Potentially available exiting-child timeslices are | |
1645 | * retrieved here - this way the parent does not get | |
1646 | * penalized for creating too many threads. | |
1647 | * | |
1648 | * (this cannot be used to 'generate' timeslices | |
1649 | * artificially, because any timeslice recovered here | |
1650 | * was given away by the parent in the first place.) | |
1651 | */ | |
95cdf3b7 | 1652 | void fastcall sched_exit(task_t *p) |
1da177e4 LT |
1653 | { |
1654 | unsigned long flags; | |
1655 | runqueue_t *rq; | |
1656 | ||
1657 | /* | |
1658 | * If the child was a (relative-) CPU hog then decrease | |
1659 | * the sleep_avg of the parent as well. | |
1660 | */ | |
1661 | rq = task_rq_lock(p->parent, &flags); | |
889dfafe | 1662 | if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) { |
1da177e4 LT |
1663 | p->parent->time_slice += p->time_slice; |
1664 | if (unlikely(p->parent->time_slice > task_timeslice(p))) | |
1665 | p->parent->time_slice = task_timeslice(p); | |
1666 | } | |
1667 | if (p->sleep_avg < p->parent->sleep_avg) | |
1668 | p->parent->sleep_avg = p->parent->sleep_avg / | |
1669 | (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg / | |
1670 | (EXIT_WEIGHT + 1); | |
1671 | task_rq_unlock(rq, &flags); | |
1672 | } | |
1673 | ||
4866cde0 NP |
1674 | /** |
1675 | * prepare_task_switch - prepare to switch tasks | |
1676 | * @rq: the runqueue preparing to switch | |
1677 | * @next: the task we are going to switch to. | |
1678 | * | |
1679 | * This is called with the rq lock held and interrupts off. It must | |
1680 | * be paired with a subsequent finish_task_switch after the context | |
1681 | * switch. | |
1682 | * | |
1683 | * prepare_task_switch sets up locking and calls architecture specific | |
1684 | * hooks. | |
1685 | */ | |
1686 | static inline void prepare_task_switch(runqueue_t *rq, task_t *next) | |
1687 | { | |
1688 | prepare_lock_switch(rq, next); | |
1689 | prepare_arch_switch(next); | |
1690 | } | |
1691 | ||
1da177e4 LT |
1692 | /** |
1693 | * finish_task_switch - clean up after a task-switch | |
344babaa | 1694 | * @rq: runqueue associated with task-switch |
1da177e4 LT |
1695 | * @prev: the thread we just switched away from. |
1696 | * | |
4866cde0 NP |
1697 | * finish_task_switch must be called after the context switch, paired |
1698 | * with a prepare_task_switch call before the context switch. | |
1699 | * finish_task_switch will reconcile locking set up by prepare_task_switch, | |
1700 | * and do any other architecture-specific cleanup actions. | |
1da177e4 LT |
1701 | * |
1702 | * Note that we may have delayed dropping an mm in context_switch(). If | |
1703 | * so, we finish that here outside of the runqueue lock. (Doing it | |
1704 | * with the lock held can cause deadlocks; see schedule() for | |
1705 | * details.) | |
1706 | */ | |
4866cde0 | 1707 | static inline void finish_task_switch(runqueue_t *rq, task_t *prev) |
1da177e4 LT |
1708 | __releases(rq->lock) |
1709 | { | |
1da177e4 LT |
1710 | struct mm_struct *mm = rq->prev_mm; |
1711 | unsigned long prev_task_flags; | |
1712 | ||
1713 | rq->prev_mm = NULL; | |
1714 | ||
1715 | /* | |
1716 | * A task struct has one reference for the use as "current". | |
1717 | * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and | |
1718 | * calls schedule one last time. The schedule call will never return, | |
1719 | * and the scheduled task must drop that reference. | |
1720 | * The test for EXIT_ZOMBIE must occur while the runqueue locks are | |
1721 | * still held, otherwise prev could be scheduled on another cpu, die | |
1722 | * there before we look at prev->state, and then the reference would | |
1723 | * be dropped twice. | |
1724 | * Manfred Spraul <manfred@colorfullife.com> | |
1725 | */ | |
1726 | prev_task_flags = prev->flags; | |
4866cde0 NP |
1727 | finish_arch_switch(prev); |
1728 | finish_lock_switch(rq, prev); | |
1da177e4 LT |
1729 | if (mm) |
1730 | mmdrop(mm); | |
c6fd91f0 | 1731 | if (unlikely(prev_task_flags & PF_DEAD)) { |
1732 | /* | |
1733 | * Remove function-return probe instances associated with this | |
1734 | * task and put them back on the free list. | |
1735 | */ | |
1736 | kprobe_flush_task(prev); | |
1da177e4 | 1737 | put_task_struct(prev); |
c6fd91f0 | 1738 | } |
1da177e4 LT |
1739 | } |
1740 | ||
1741 | /** | |
1742 | * schedule_tail - first thing a freshly forked thread must call. | |
1743 | * @prev: the thread we just switched away from. | |
1744 | */ | |
1745 | asmlinkage void schedule_tail(task_t *prev) | |
1746 | __releases(rq->lock) | |
1747 | { | |
4866cde0 NP |
1748 | runqueue_t *rq = this_rq(); |
1749 | finish_task_switch(rq, prev); | |
1750 | #ifdef __ARCH_WANT_UNLOCKED_CTXSW | |
1751 | /* In this case, finish_task_switch does not reenable preemption */ | |
1752 | preempt_enable(); | |
1753 | #endif | |
1da177e4 LT |
1754 | if (current->set_child_tid) |
1755 | put_user(current->pid, current->set_child_tid); | |
1756 | } | |
1757 | ||
1758 | /* | |
1759 | * context_switch - switch to the new MM and the new | |
1760 | * thread's register state. | |
1761 | */ | |
1762 | static inline | |
1763 | task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next) | |
1764 | { | |
1765 | struct mm_struct *mm = next->mm; | |
1766 | struct mm_struct *oldmm = prev->active_mm; | |
1767 | ||
1768 | if (unlikely(!mm)) { | |
1769 | next->active_mm = oldmm; | |
1770 | atomic_inc(&oldmm->mm_count); | |
1771 | enter_lazy_tlb(oldmm, next); | |
1772 | } else | |
1773 | switch_mm(oldmm, mm, next); | |
1774 | ||
1775 | if (unlikely(!prev->mm)) { | |
1776 | prev->active_mm = NULL; | |
1777 | WARN_ON(rq->prev_mm); | |
1778 | rq->prev_mm = oldmm; | |
1779 | } | |
1780 | ||
1781 | /* Here we just switch the register state and the stack. */ | |
1782 | switch_to(prev, next, prev); | |
1783 | ||
1784 | return prev; | |
1785 | } | |
1786 | ||
1787 | /* | |
1788 | * nr_running, nr_uninterruptible and nr_context_switches: | |
1789 | * | |
1790 | * externally visible scheduler statistics: current number of runnable | |
1791 | * threads, current number of uninterruptible-sleeping threads, total | |
1792 | * number of context switches performed since bootup. | |
1793 | */ | |
1794 | unsigned long nr_running(void) | |
1795 | { | |
1796 | unsigned long i, sum = 0; | |
1797 | ||
1798 | for_each_online_cpu(i) | |
1799 | sum += cpu_rq(i)->nr_running; | |
1800 | ||
1801 | return sum; | |
1802 | } | |
1803 | ||
1804 | unsigned long nr_uninterruptible(void) | |
1805 | { | |
1806 | unsigned long i, sum = 0; | |
1807 | ||
0a945022 | 1808 | for_each_possible_cpu(i) |
1da177e4 LT |
1809 | sum += cpu_rq(i)->nr_uninterruptible; |
1810 | ||
1811 | /* | |
1812 | * Since we read the counters lockless, it might be slightly | |
1813 | * inaccurate. Do not allow it to go below zero though: | |
1814 | */ | |
1815 | if (unlikely((long)sum < 0)) | |
1816 | sum = 0; | |
1817 | ||
1818 | return sum; | |
1819 | } | |
1820 | ||
1821 | unsigned long long nr_context_switches(void) | |
1822 | { | |
cc94abfc SR |
1823 | int i; |
1824 | unsigned long long sum = 0; | |
1da177e4 | 1825 | |
0a945022 | 1826 | for_each_possible_cpu(i) |
1da177e4 LT |
1827 | sum += cpu_rq(i)->nr_switches; |
1828 | ||
1829 | return sum; | |
1830 | } | |
1831 | ||
1832 | unsigned long nr_iowait(void) | |
1833 | { | |
1834 | unsigned long i, sum = 0; | |
1835 | ||
0a945022 | 1836 | for_each_possible_cpu(i) |
1da177e4 LT |
1837 | sum += atomic_read(&cpu_rq(i)->nr_iowait); |
1838 | ||
1839 | return sum; | |
1840 | } | |
1841 | ||
db1b1fef JS |
1842 | unsigned long nr_active(void) |
1843 | { | |
1844 | unsigned long i, running = 0, uninterruptible = 0; | |
1845 | ||
1846 | for_each_online_cpu(i) { | |
1847 | running += cpu_rq(i)->nr_running; | |
1848 | uninterruptible += cpu_rq(i)->nr_uninterruptible; | |
1849 | } | |
1850 | ||
1851 | if (unlikely((long)uninterruptible < 0)) | |
1852 | uninterruptible = 0; | |
1853 | ||
1854 | return running + uninterruptible; | |
1855 | } | |
1856 | ||
1da177e4 LT |
1857 | #ifdef CONFIG_SMP |
1858 | ||
1859 | /* | |
1860 | * double_rq_lock - safely lock two runqueues | |
1861 | * | |
1862 | * Note this does not disable interrupts like task_rq_lock, | |
1863 | * you need to do so manually before calling. | |
1864 | */ | |
1865 | static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2) | |
1866 | __acquires(rq1->lock) | |
1867 | __acquires(rq2->lock) | |
1868 | { | |
1869 | if (rq1 == rq2) { | |
1870 | spin_lock(&rq1->lock); | |
1871 | __acquire(rq2->lock); /* Fake it out ;) */ | |
1872 | } else { | |
c96d145e | 1873 | if (rq1 < rq2) { |
1da177e4 LT |
1874 | spin_lock(&rq1->lock); |
1875 | spin_lock(&rq2->lock); | |
1876 | } else { | |
1877 | spin_lock(&rq2->lock); | |
1878 | spin_lock(&rq1->lock); | |
1879 | } | |
1880 | } | |
1881 | } | |
1882 | ||
1883 | /* | |
1884 | * double_rq_unlock - safely unlock two runqueues | |
1885 | * | |
1886 | * Note this does not restore interrupts like task_rq_unlock, | |
1887 | * you need to do so manually after calling. | |
1888 | */ | |
1889 | static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2) | |
1890 | __releases(rq1->lock) | |
1891 | __releases(rq2->lock) | |
1892 | { | |
1893 | spin_unlock(&rq1->lock); | |
1894 | if (rq1 != rq2) | |
1895 | spin_unlock(&rq2->lock); | |
1896 | else | |
1897 | __release(rq2->lock); | |
1898 | } | |
1899 | ||
1900 | /* | |
1901 | * double_lock_balance - lock the busiest runqueue, this_rq is locked already. | |
1902 | */ | |
1903 | static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest) | |
1904 | __releases(this_rq->lock) | |
1905 | __acquires(busiest->lock) | |
1906 | __acquires(this_rq->lock) | |
1907 | { | |
1908 | if (unlikely(!spin_trylock(&busiest->lock))) { | |
c96d145e | 1909 | if (busiest < this_rq) { |
1da177e4 LT |
1910 | spin_unlock(&this_rq->lock); |
1911 | spin_lock(&busiest->lock); | |
1912 | spin_lock(&this_rq->lock); | |
1913 | } else | |
1914 | spin_lock(&busiest->lock); | |
1915 | } | |
1916 | } | |
1917 | ||
1da177e4 LT |
1918 | /* |
1919 | * If dest_cpu is allowed for this process, migrate the task to it. | |
1920 | * This is accomplished by forcing the cpu_allowed mask to only | |
1921 | * allow dest_cpu, which will force the cpu onto dest_cpu. Then | |
1922 | * the cpu_allowed mask is restored. | |
1923 | */ | |
1924 | static void sched_migrate_task(task_t *p, int dest_cpu) | |
1925 | { | |
1926 | migration_req_t req; | |
1927 | runqueue_t *rq; | |
1928 | unsigned long flags; | |
1929 | ||
1930 | rq = task_rq_lock(p, &flags); | |
1931 | if (!cpu_isset(dest_cpu, p->cpus_allowed) | |
1932 | || unlikely(cpu_is_offline(dest_cpu))) | |
1933 | goto out; | |
1934 | ||
1935 | /* force the process onto the specified CPU */ | |
1936 | if (migrate_task(p, dest_cpu, &req)) { | |
1937 | /* Need to wait for migration thread (might exit: take ref). */ | |
1938 | struct task_struct *mt = rq->migration_thread; | |
1939 | get_task_struct(mt); | |
1940 | task_rq_unlock(rq, &flags); | |
1941 | wake_up_process(mt); | |
1942 | put_task_struct(mt); | |
1943 | wait_for_completion(&req.done); | |
1944 | return; | |
1945 | } | |
1946 | out: | |
1947 | task_rq_unlock(rq, &flags); | |
1948 | } | |
1949 | ||
1950 | /* | |
476d139c NP |
1951 | * sched_exec - execve() is a valuable balancing opportunity, because at |
1952 | * this point the task has the smallest effective memory and cache footprint. | |
1da177e4 LT |
1953 | */ |
1954 | void sched_exec(void) | |
1955 | { | |
1da177e4 | 1956 | int new_cpu, this_cpu = get_cpu(); |
476d139c | 1957 | new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC); |
1da177e4 | 1958 | put_cpu(); |
476d139c NP |
1959 | if (new_cpu != this_cpu) |
1960 | sched_migrate_task(current, new_cpu); | |
1da177e4 LT |
1961 | } |
1962 | ||
1963 | /* | |
1964 | * pull_task - move a task from a remote runqueue to the local runqueue. | |
1965 | * Both runqueues must be locked. | |
1966 | */ | |
858119e1 | 1967 | static |
1da177e4 LT |
1968 | void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p, |
1969 | runqueue_t *this_rq, prio_array_t *this_array, int this_cpu) | |
1970 | { | |
1971 | dequeue_task(p, src_array); | |
2dd73a4f | 1972 | dec_nr_running(p, src_rq); |
1da177e4 | 1973 | set_task_cpu(p, this_cpu); |
2dd73a4f | 1974 | inc_nr_running(p, this_rq); |
1da177e4 LT |
1975 | enqueue_task(p, this_array); |
1976 | p->timestamp = (p->timestamp - src_rq->timestamp_last_tick) | |
1977 | + this_rq->timestamp_last_tick; | |
1978 | /* | |
1979 | * Note that idle threads have a prio of MAX_PRIO, for this test | |
1980 | * to be always true for them. | |
1981 | */ | |
1982 | if (TASK_PREEMPTS_CURR(p, this_rq)) | |
1983 | resched_task(this_rq->curr); | |
1984 | } | |
1985 | ||
1986 | /* | |
1987 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
1988 | */ | |
858119e1 | 1989 | static |
1da177e4 | 1990 | int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu, |
95cdf3b7 IM |
1991 | struct sched_domain *sd, enum idle_type idle, |
1992 | int *all_pinned) | |
1da177e4 LT |
1993 | { |
1994 | /* | |
1995 | * We do not migrate tasks that are: | |
1996 | * 1) running (obviously), or | |
1997 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
1998 | * 3) are cache-hot on their current CPU. | |
1999 | */ | |
1da177e4 LT |
2000 | if (!cpu_isset(this_cpu, p->cpus_allowed)) |
2001 | return 0; | |
81026794 NP |
2002 | *all_pinned = 0; |
2003 | ||
2004 | if (task_running(rq, p)) | |
2005 | return 0; | |
1da177e4 LT |
2006 | |
2007 | /* | |
2008 | * Aggressive migration if: | |
cafb20c1 | 2009 | * 1) task is cache cold, or |
1da177e4 LT |
2010 | * 2) too many balance attempts have failed. |
2011 | */ | |
2012 | ||
cafb20c1 | 2013 | if (sd->nr_balance_failed > sd->cache_nice_tries) |
1da177e4 LT |
2014 | return 1; |
2015 | ||
2016 | if (task_hot(p, rq->timestamp_last_tick, sd)) | |
81026794 | 2017 | return 0; |
1da177e4 LT |
2018 | return 1; |
2019 | } | |
2020 | ||
615052dc | 2021 | #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio) |
1da177e4 | 2022 | /* |
2dd73a4f PW |
2023 | * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted |
2024 | * load from busiest to this_rq, as part of a balancing operation within | |
2025 | * "domain". Returns the number of tasks moved. | |
1da177e4 LT |
2026 | * |
2027 | * Called with both runqueues locked. | |
2028 | */ | |
2029 | static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest, | |
2dd73a4f PW |
2030 | unsigned long max_nr_move, unsigned long max_load_move, |
2031 | struct sched_domain *sd, enum idle_type idle, | |
2032 | int *all_pinned) | |
1da177e4 LT |
2033 | { |
2034 | prio_array_t *array, *dst_array; | |
2035 | struct list_head *head, *curr; | |
615052dc PW |
2036 | int idx, pulled = 0, pinned = 0, this_best_prio, busiest_best_prio; |
2037 | int busiest_best_prio_seen; | |
2038 | int skip_for_load; /* skip the task based on weighted load issues */ | |
2dd73a4f | 2039 | long rem_load_move; |
1da177e4 LT |
2040 | task_t *tmp; |
2041 | ||
2dd73a4f | 2042 | if (max_nr_move == 0 || max_load_move == 0) |
1da177e4 LT |
2043 | goto out; |
2044 | ||
2dd73a4f | 2045 | rem_load_move = max_load_move; |
81026794 | 2046 | pinned = 1; |
615052dc PW |
2047 | this_best_prio = rq_best_prio(this_rq); |
2048 | busiest_best_prio = rq_best_prio(busiest); | |
2049 | /* | |
2050 | * Enable handling of the case where there is more than one task | |
2051 | * with the best priority. If the current running task is one | |
2052 | * of those with prio==busiest_best_prio we know it won't be moved | |
2053 | * and therefore it's safe to override the skip (based on load) of | |
2054 | * any task we find with that prio. | |
2055 | */ | |
2056 | busiest_best_prio_seen = busiest_best_prio == busiest->curr->prio; | |
81026794 | 2057 | |
1da177e4 LT |
2058 | /* |
2059 | * We first consider expired tasks. Those will likely not be | |
2060 | * executed in the near future, and they are most likely to | |
2061 | * be cache-cold, thus switching CPUs has the least effect | |
2062 | * on them. | |
2063 | */ | |
2064 | if (busiest->expired->nr_active) { | |
2065 | array = busiest->expired; | |
2066 | dst_array = this_rq->expired; | |
2067 | } else { | |
2068 | array = busiest->active; | |
2069 | dst_array = this_rq->active; | |
2070 | } | |
2071 | ||
2072 | new_array: | |
2073 | /* Start searching at priority 0: */ | |
2074 | idx = 0; | |
2075 | skip_bitmap: | |
2076 | if (!idx) | |
2077 | idx = sched_find_first_bit(array->bitmap); | |
2078 | else | |
2079 | idx = find_next_bit(array->bitmap, MAX_PRIO, idx); | |
2080 | if (idx >= MAX_PRIO) { | |
2081 | if (array == busiest->expired && busiest->active->nr_active) { | |
2082 | array = busiest->active; | |
2083 | dst_array = this_rq->active; | |
2084 | goto new_array; | |
2085 | } | |
2086 | goto out; | |
2087 | } | |
2088 | ||
2089 | head = array->queue + idx; | |
2090 | curr = head->prev; | |
2091 | skip_queue: | |
2092 | tmp = list_entry(curr, task_t, run_list); | |
2093 | ||
2094 | curr = curr->prev; | |
2095 | ||
50ddd969 PW |
2096 | /* |
2097 | * To help distribute high priority tasks accross CPUs we don't | |
2098 | * skip a task if it will be the highest priority task (i.e. smallest | |
2099 | * prio value) on its new queue regardless of its load weight | |
2100 | */ | |
615052dc PW |
2101 | skip_for_load = tmp->load_weight > rem_load_move; |
2102 | if (skip_for_load && idx < this_best_prio) | |
2103 | skip_for_load = !busiest_best_prio_seen && idx == busiest_best_prio; | |
2104 | if (skip_for_load || | |
2dd73a4f | 2105 | !can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) { |
615052dc | 2106 | busiest_best_prio_seen |= idx == busiest_best_prio; |
1da177e4 LT |
2107 | if (curr != head) |
2108 | goto skip_queue; | |
2109 | idx++; | |
2110 | goto skip_bitmap; | |
2111 | } | |
2112 | ||
2113 | #ifdef CONFIG_SCHEDSTATS | |
2114 | if (task_hot(tmp, busiest->timestamp_last_tick, sd)) | |
2115 | schedstat_inc(sd, lb_hot_gained[idle]); | |
2116 | #endif | |
2117 | ||
2118 | pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu); | |
2119 | pulled++; | |
2dd73a4f | 2120 | rem_load_move -= tmp->load_weight; |
1da177e4 | 2121 | |
2dd73a4f PW |
2122 | /* |
2123 | * We only want to steal up to the prescribed number of tasks | |
2124 | * and the prescribed amount of weighted load. | |
2125 | */ | |
2126 | if (pulled < max_nr_move && rem_load_move > 0) { | |
615052dc PW |
2127 | if (idx < this_best_prio) |
2128 | this_best_prio = idx; | |
1da177e4 LT |
2129 | if (curr != head) |
2130 | goto skip_queue; | |
2131 | idx++; | |
2132 | goto skip_bitmap; | |
2133 | } | |
2134 | out: | |
2135 | /* | |
2136 | * Right now, this is the only place pull_task() is called, | |
2137 | * so we can safely collect pull_task() stats here rather than | |
2138 | * inside pull_task(). | |
2139 | */ | |
2140 | schedstat_add(sd, lb_gained[idle], pulled); | |
81026794 NP |
2141 | |
2142 | if (all_pinned) | |
2143 | *all_pinned = pinned; | |
1da177e4 LT |
2144 | return pulled; |
2145 | } | |
2146 | ||
2147 | /* | |
2148 | * find_busiest_group finds and returns the busiest CPU group within the | |
2dd73a4f | 2149 | * domain. It calculates and returns the amount of weighted load which should be |
1da177e4 LT |
2150 | * moved to restore balance via the imbalance parameter. |
2151 | */ | |
2152 | static struct sched_group * | |
2153 | find_busiest_group(struct sched_domain *sd, int this_cpu, | |
5969fe06 | 2154 | unsigned long *imbalance, enum idle_type idle, int *sd_idle) |
1da177e4 LT |
2155 | { |
2156 | struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups; | |
2157 | unsigned long max_load, avg_load, total_load, this_load, total_pwr; | |
0c117f1b | 2158 | unsigned long max_pull; |
2dd73a4f PW |
2159 | unsigned long busiest_load_per_task, busiest_nr_running; |
2160 | unsigned long this_load_per_task, this_nr_running; | |
7897986b | 2161 | int load_idx; |
5c45bf27 SS |
2162 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
2163 | int power_savings_balance = 1; | |
2164 | unsigned long leader_nr_running = 0, min_load_per_task = 0; | |
2165 | unsigned long min_nr_running = ULONG_MAX; | |
2166 | struct sched_group *group_min = NULL, *group_leader = NULL; | |
2167 | #endif | |
1da177e4 LT |
2168 | |
2169 | max_load = this_load = total_load = total_pwr = 0; | |
2dd73a4f PW |
2170 | busiest_load_per_task = busiest_nr_running = 0; |
2171 | this_load_per_task = this_nr_running = 0; | |
7897986b NP |
2172 | if (idle == NOT_IDLE) |
2173 | load_idx = sd->busy_idx; | |
2174 | else if (idle == NEWLY_IDLE) | |
2175 | load_idx = sd->newidle_idx; | |
2176 | else | |
2177 | load_idx = sd->idle_idx; | |
1da177e4 LT |
2178 | |
2179 | do { | |
5c45bf27 | 2180 | unsigned long load, group_capacity; |
1da177e4 LT |
2181 | int local_group; |
2182 | int i; | |
2dd73a4f | 2183 | unsigned long sum_nr_running, sum_weighted_load; |
1da177e4 LT |
2184 | |
2185 | local_group = cpu_isset(this_cpu, group->cpumask); | |
2186 | ||
2187 | /* Tally up the load of all CPUs in the group */ | |
2dd73a4f | 2188 | sum_weighted_load = sum_nr_running = avg_load = 0; |
1da177e4 LT |
2189 | |
2190 | for_each_cpu_mask(i, group->cpumask) { | |
2dd73a4f PW |
2191 | runqueue_t *rq = cpu_rq(i); |
2192 | ||
5969fe06 NP |
2193 | if (*sd_idle && !idle_cpu(i)) |
2194 | *sd_idle = 0; | |
2195 | ||
1da177e4 LT |
2196 | /* Bias balancing toward cpus of our domain */ |
2197 | if (local_group) | |
a2000572 | 2198 | load = target_load(i, load_idx); |
1da177e4 | 2199 | else |
a2000572 | 2200 | load = source_load(i, load_idx); |
1da177e4 LT |
2201 | |
2202 | avg_load += load; | |
2dd73a4f PW |
2203 | sum_nr_running += rq->nr_running; |
2204 | sum_weighted_load += rq->raw_weighted_load; | |
1da177e4 LT |
2205 | } |
2206 | ||
2207 | total_load += avg_load; | |
2208 | total_pwr += group->cpu_power; | |
2209 | ||
2210 | /* Adjust by relative CPU power of the group */ | |
2211 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
2212 | ||
5c45bf27 SS |
2213 | group_capacity = group->cpu_power / SCHED_LOAD_SCALE; |
2214 | ||
1da177e4 LT |
2215 | if (local_group) { |
2216 | this_load = avg_load; | |
2217 | this = group; | |
2dd73a4f PW |
2218 | this_nr_running = sum_nr_running; |
2219 | this_load_per_task = sum_weighted_load; | |
2220 | } else if (avg_load > max_load && | |
5c45bf27 | 2221 | sum_nr_running > group_capacity) { |
1da177e4 LT |
2222 | max_load = avg_load; |
2223 | busiest = group; | |
2dd73a4f PW |
2224 | busiest_nr_running = sum_nr_running; |
2225 | busiest_load_per_task = sum_weighted_load; | |
1da177e4 | 2226 | } |
5c45bf27 SS |
2227 | |
2228 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
2229 | /* | |
2230 | * Busy processors will not participate in power savings | |
2231 | * balance. | |
2232 | */ | |
2233 | if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE)) | |
2234 | goto group_next; | |
2235 | ||
2236 | /* | |
2237 | * If the local group is idle or completely loaded | |
2238 | * no need to do power savings balance at this domain | |
2239 | */ | |
2240 | if (local_group && (this_nr_running >= group_capacity || | |
2241 | !this_nr_running)) | |
2242 | power_savings_balance = 0; | |
2243 | ||
2244 | /* | |
2245 | * If a group is already running at full capacity or idle, | |
2246 | * don't include that group in power savings calculations | |
2247 | */ | |
2248 | if (!power_savings_balance || sum_nr_running >= group_capacity | |
2249 | || !sum_nr_running) | |
2250 | goto group_next; | |
2251 | ||
2252 | /* | |
2253 | * Calculate the group which has the least non-idle load. | |
2254 | * This is the group from where we need to pick up the load | |
2255 | * for saving power | |
2256 | */ | |
2257 | if ((sum_nr_running < min_nr_running) || | |
2258 | (sum_nr_running == min_nr_running && | |
2259 | first_cpu(group->cpumask) < | |
2260 | first_cpu(group_min->cpumask))) { | |
2261 | group_min = group; | |
2262 | min_nr_running = sum_nr_running; | |
2263 | min_load_per_task = sum_weighted_load / | |
2264 | sum_nr_running; | |
2265 | } | |
2266 | ||
2267 | /* | |
2268 | * Calculate the group which is almost near its | |
2269 | * capacity but still has some space to pick up some load | |
2270 | * from other group and save more power | |
2271 | */ | |
2272 | if (sum_nr_running <= group_capacity - 1) | |
2273 | if (sum_nr_running > leader_nr_running || | |
2274 | (sum_nr_running == leader_nr_running && | |
2275 | first_cpu(group->cpumask) > | |
2276 | first_cpu(group_leader->cpumask))) { | |
2277 | group_leader = group; | |
2278 | leader_nr_running = sum_nr_running; | |
2279 | } | |
2280 | ||
2281 | group_next: | |
2282 | #endif | |
1da177e4 LT |
2283 | group = group->next; |
2284 | } while (group != sd->groups); | |
2285 | ||
2dd73a4f | 2286 | if (!busiest || this_load >= max_load || busiest_nr_running == 0) |
1da177e4 LT |
2287 | goto out_balanced; |
2288 | ||
2289 | avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr; | |
2290 | ||
2291 | if (this_load >= avg_load || | |
2292 | 100*max_load <= sd->imbalance_pct*this_load) | |
2293 | goto out_balanced; | |
2294 | ||
2dd73a4f | 2295 | busiest_load_per_task /= busiest_nr_running; |
1da177e4 LT |
2296 | /* |
2297 | * We're trying to get all the cpus to the average_load, so we don't | |
2298 | * want to push ourselves above the average load, nor do we wish to | |
2299 | * reduce the max loaded cpu below the average load, as either of these | |
2300 | * actions would just result in more rebalancing later, and ping-pong | |
2301 | * tasks around. Thus we look for the minimum possible imbalance. | |
2302 | * Negative imbalances (*we* are more loaded than anyone else) will | |
2303 | * be counted as no imbalance for these purposes -- we can't fix that | |
2304 | * by pulling tasks to us. Be careful of negative numbers as they'll | |
2305 | * appear as very large values with unsigned longs. | |
2306 | */ | |
2dd73a4f PW |
2307 | if (max_load <= busiest_load_per_task) |
2308 | goto out_balanced; | |
2309 | ||
2310 | /* | |
2311 | * In the presence of smp nice balancing, certain scenarios can have | |
2312 | * max load less than avg load(as we skip the groups at or below | |
2313 | * its cpu_power, while calculating max_load..) | |
2314 | */ | |
2315 | if (max_load < avg_load) { | |
2316 | *imbalance = 0; | |
2317 | goto small_imbalance; | |
2318 | } | |
0c117f1b SS |
2319 | |
2320 | /* Don't want to pull so many tasks that a group would go idle */ | |
2dd73a4f | 2321 | max_pull = min(max_load - avg_load, max_load - busiest_load_per_task); |
0c117f1b | 2322 | |
1da177e4 | 2323 | /* How much load to actually move to equalise the imbalance */ |
0c117f1b | 2324 | *imbalance = min(max_pull * busiest->cpu_power, |
1da177e4 LT |
2325 | (avg_load - this_load) * this->cpu_power) |
2326 | / SCHED_LOAD_SCALE; | |
2327 | ||
2dd73a4f PW |
2328 | /* |
2329 | * if *imbalance is less than the average load per runnable task | |
2330 | * there is no gaurantee that any tasks will be moved so we'll have | |
2331 | * a think about bumping its value to force at least one task to be | |
2332 | * moved | |
2333 | */ | |
2334 | if (*imbalance < busiest_load_per_task) { | |
2335 | unsigned long pwr_now, pwr_move; | |
1da177e4 | 2336 | unsigned long tmp; |
2dd73a4f PW |
2337 | unsigned int imbn; |
2338 | ||
2339 | small_imbalance: | |
2340 | pwr_move = pwr_now = 0; | |
2341 | imbn = 2; | |
2342 | if (this_nr_running) { | |
2343 | this_load_per_task /= this_nr_running; | |
2344 | if (busiest_load_per_task > this_load_per_task) | |
2345 | imbn = 1; | |
2346 | } else | |
2347 | this_load_per_task = SCHED_LOAD_SCALE; | |
1da177e4 | 2348 | |
2dd73a4f PW |
2349 | if (max_load - this_load >= busiest_load_per_task * imbn) { |
2350 | *imbalance = busiest_load_per_task; | |
1da177e4 LT |
2351 | return busiest; |
2352 | } | |
2353 | ||
2354 | /* | |
2355 | * OK, we don't have enough imbalance to justify moving tasks, | |
2356 | * however we may be able to increase total CPU power used by | |
2357 | * moving them. | |
2358 | */ | |
2359 | ||
2dd73a4f PW |
2360 | pwr_now += busiest->cpu_power * |
2361 | min(busiest_load_per_task, max_load); | |
2362 | pwr_now += this->cpu_power * | |
2363 | min(this_load_per_task, this_load); | |
1da177e4 LT |
2364 | pwr_now /= SCHED_LOAD_SCALE; |
2365 | ||
2366 | /* Amount of load we'd subtract */ | |
2dd73a4f | 2367 | tmp = busiest_load_per_task*SCHED_LOAD_SCALE/busiest->cpu_power; |
1da177e4 | 2368 | if (max_load > tmp) |
2dd73a4f PW |
2369 | pwr_move += busiest->cpu_power * |
2370 | min(busiest_load_per_task, max_load - tmp); | |
1da177e4 LT |
2371 | |
2372 | /* Amount of load we'd add */ | |
2373 | if (max_load*busiest->cpu_power < | |
2dd73a4f | 2374 | busiest_load_per_task*SCHED_LOAD_SCALE) |
1da177e4 LT |
2375 | tmp = max_load*busiest->cpu_power/this->cpu_power; |
2376 | else | |
2dd73a4f PW |
2377 | tmp = busiest_load_per_task*SCHED_LOAD_SCALE/this->cpu_power; |
2378 | pwr_move += this->cpu_power*min(this_load_per_task, this_load + tmp); | |
1da177e4 LT |
2379 | pwr_move /= SCHED_LOAD_SCALE; |
2380 | ||
2381 | /* Move if we gain throughput */ | |
2382 | if (pwr_move <= pwr_now) | |
2383 | goto out_balanced; | |
2384 | ||
2dd73a4f | 2385 | *imbalance = busiest_load_per_task; |
1da177e4 LT |
2386 | } |
2387 | ||
1da177e4 LT |
2388 | return busiest; |
2389 | ||
2390 | out_balanced: | |
5c45bf27 SS |
2391 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
2392 | if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE)) | |
2393 | goto ret; | |
1da177e4 | 2394 | |
5c45bf27 SS |
2395 | if (this == group_leader && group_leader != group_min) { |
2396 | *imbalance = min_load_per_task; | |
2397 | return group_min; | |
2398 | } | |
2399 | ret: | |
2400 | #endif | |
1da177e4 LT |
2401 | *imbalance = 0; |
2402 | return NULL; | |
2403 | } | |
2404 | ||
2405 | /* | |
2406 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
2407 | */ | |
b910472d | 2408 | static runqueue_t *find_busiest_queue(struct sched_group *group, |
2dd73a4f | 2409 | enum idle_type idle, unsigned long imbalance) |
1da177e4 | 2410 | { |
2dd73a4f PW |
2411 | unsigned long max_load = 0; |
2412 | runqueue_t *busiest = NULL, *rqi; | |
1da177e4 LT |
2413 | int i; |
2414 | ||
2415 | for_each_cpu_mask(i, group->cpumask) { | |
2dd73a4f PW |
2416 | rqi = cpu_rq(i); |
2417 | ||
2418 | if (rqi->nr_running == 1 && rqi->raw_weighted_load > imbalance) | |
2419 | continue; | |
1da177e4 | 2420 | |
2dd73a4f PW |
2421 | if (rqi->raw_weighted_load > max_load) { |
2422 | max_load = rqi->raw_weighted_load; | |
2423 | busiest = rqi; | |
1da177e4 LT |
2424 | } |
2425 | } | |
2426 | ||
2427 | return busiest; | |
2428 | } | |
2429 | ||
77391d71 NP |
2430 | /* |
2431 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
2432 | * so long as it is large enough. | |
2433 | */ | |
2434 | #define MAX_PINNED_INTERVAL 512 | |
2435 | ||
2dd73a4f | 2436 | #define minus_1_or_zero(n) ((n) > 0 ? (n) - 1 : 0) |
1da177e4 LT |
2437 | /* |
2438 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
2439 | * tasks if there is an imbalance. | |
2440 | * | |
2441 | * Called with this_rq unlocked. | |
2442 | */ | |
2443 | static int load_balance(int this_cpu, runqueue_t *this_rq, | |
2444 | struct sched_domain *sd, enum idle_type idle) | |
2445 | { | |
2446 | struct sched_group *group; | |
2447 | runqueue_t *busiest; | |
2448 | unsigned long imbalance; | |
77391d71 | 2449 | int nr_moved, all_pinned = 0; |
81026794 | 2450 | int active_balance = 0; |
5969fe06 NP |
2451 | int sd_idle = 0; |
2452 | ||
5c45bf27 SS |
2453 | if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER && |
2454 | !sched_smt_power_savings) | |
5969fe06 | 2455 | sd_idle = 1; |
1da177e4 | 2456 | |
1da177e4 LT |
2457 | schedstat_inc(sd, lb_cnt[idle]); |
2458 | ||
5969fe06 | 2459 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle); |
1da177e4 LT |
2460 | if (!group) { |
2461 | schedstat_inc(sd, lb_nobusyg[idle]); | |
2462 | goto out_balanced; | |
2463 | } | |
2464 | ||
2dd73a4f | 2465 | busiest = find_busiest_queue(group, idle, imbalance); |
1da177e4 LT |
2466 | if (!busiest) { |
2467 | schedstat_inc(sd, lb_nobusyq[idle]); | |
2468 | goto out_balanced; | |
2469 | } | |
2470 | ||
db935dbd | 2471 | BUG_ON(busiest == this_rq); |
1da177e4 LT |
2472 | |
2473 | schedstat_add(sd, lb_imbalance[idle], imbalance); | |
2474 | ||
2475 | nr_moved = 0; | |
2476 | if (busiest->nr_running > 1) { | |
2477 | /* | |
2478 | * Attempt to move tasks. If find_busiest_group has found | |
2479 | * an imbalance but busiest->nr_running <= 1, the group is | |
2480 | * still unbalanced. nr_moved simply stays zero, so it is | |
2481 | * correctly treated as an imbalance. | |
2482 | */ | |
e17224bf | 2483 | double_rq_lock(this_rq, busiest); |
1da177e4 | 2484 | nr_moved = move_tasks(this_rq, this_cpu, busiest, |
2dd73a4f | 2485 | minus_1_or_zero(busiest->nr_running), |
d6d5cfaf | 2486 | imbalance, sd, idle, &all_pinned); |
e17224bf | 2487 | double_rq_unlock(this_rq, busiest); |
81026794 NP |
2488 | |
2489 | /* All tasks on this runqueue were pinned by CPU affinity */ | |
2490 | if (unlikely(all_pinned)) | |
2491 | goto out_balanced; | |
1da177e4 | 2492 | } |
81026794 | 2493 | |
1da177e4 LT |
2494 | if (!nr_moved) { |
2495 | schedstat_inc(sd, lb_failed[idle]); | |
2496 | sd->nr_balance_failed++; | |
2497 | ||
2498 | if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) { | |
1da177e4 LT |
2499 | |
2500 | spin_lock(&busiest->lock); | |
fa3b6ddc SS |
2501 | |
2502 | /* don't kick the migration_thread, if the curr | |
2503 | * task on busiest cpu can't be moved to this_cpu | |
2504 | */ | |
2505 | if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) { | |
2506 | spin_unlock(&busiest->lock); | |
2507 | all_pinned = 1; | |
2508 | goto out_one_pinned; | |
2509 | } | |
2510 | ||
1da177e4 LT |
2511 | if (!busiest->active_balance) { |
2512 | busiest->active_balance = 1; | |
2513 | busiest->push_cpu = this_cpu; | |
81026794 | 2514 | active_balance = 1; |
1da177e4 LT |
2515 | } |
2516 | spin_unlock(&busiest->lock); | |
81026794 | 2517 | if (active_balance) |
1da177e4 LT |
2518 | wake_up_process(busiest->migration_thread); |
2519 | ||
2520 | /* | |
2521 | * We've kicked active balancing, reset the failure | |
2522 | * counter. | |
2523 | */ | |
39507451 | 2524 | sd->nr_balance_failed = sd->cache_nice_tries+1; |
1da177e4 | 2525 | } |
81026794 | 2526 | } else |
1da177e4 LT |
2527 | sd->nr_balance_failed = 0; |
2528 | ||
81026794 | 2529 | if (likely(!active_balance)) { |
1da177e4 LT |
2530 | /* We were unbalanced, so reset the balancing interval */ |
2531 | sd->balance_interval = sd->min_interval; | |
81026794 NP |
2532 | } else { |
2533 | /* | |
2534 | * If we've begun active balancing, start to back off. This | |
2535 | * case may not be covered by the all_pinned logic if there | |
2536 | * is only 1 task on the busy runqueue (because we don't call | |
2537 | * move_tasks). | |
2538 | */ | |
2539 | if (sd->balance_interval < sd->max_interval) | |
2540 | sd->balance_interval *= 2; | |
1da177e4 LT |
2541 | } |
2542 | ||
5c45bf27 SS |
2543 | if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
2544 | !sched_smt_power_savings) | |
5969fe06 | 2545 | return -1; |
1da177e4 LT |
2546 | return nr_moved; |
2547 | ||
2548 | out_balanced: | |
1da177e4 LT |
2549 | schedstat_inc(sd, lb_balanced[idle]); |
2550 | ||
16cfb1c0 | 2551 | sd->nr_balance_failed = 0; |
fa3b6ddc SS |
2552 | |
2553 | out_one_pinned: | |
1da177e4 | 2554 | /* tune up the balancing interval */ |
77391d71 NP |
2555 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || |
2556 | (sd->balance_interval < sd->max_interval)) | |
1da177e4 LT |
2557 | sd->balance_interval *= 2; |
2558 | ||
5c45bf27 | 2559 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && !sched_smt_power_savings) |
5969fe06 | 2560 | return -1; |
1da177e4 LT |
2561 | return 0; |
2562 | } | |
2563 | ||
2564 | /* | |
2565 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
2566 | * tasks if there is an imbalance. | |
2567 | * | |
2568 | * Called from schedule when this_rq is about to become idle (NEWLY_IDLE). | |
2569 | * this_rq is locked. | |
2570 | */ | |
2571 | static int load_balance_newidle(int this_cpu, runqueue_t *this_rq, | |
2572 | struct sched_domain *sd) | |
2573 | { | |
2574 | struct sched_group *group; | |
2575 | runqueue_t *busiest = NULL; | |
2576 | unsigned long imbalance; | |
2577 | int nr_moved = 0; | |
5969fe06 NP |
2578 | int sd_idle = 0; |
2579 | ||
5c45bf27 | 2580 | if (sd->flags & SD_SHARE_CPUPOWER && !sched_smt_power_savings) |
5969fe06 | 2581 | sd_idle = 1; |
1da177e4 LT |
2582 | |
2583 | schedstat_inc(sd, lb_cnt[NEWLY_IDLE]); | |
5969fe06 | 2584 | group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE, &sd_idle); |
1da177e4 | 2585 | if (!group) { |
1da177e4 | 2586 | schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]); |
16cfb1c0 | 2587 | goto out_balanced; |
1da177e4 LT |
2588 | } |
2589 | ||
2dd73a4f | 2590 | busiest = find_busiest_queue(group, NEWLY_IDLE, imbalance); |
db935dbd | 2591 | if (!busiest) { |
1da177e4 | 2592 | schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]); |
16cfb1c0 | 2593 | goto out_balanced; |
1da177e4 LT |
2594 | } |
2595 | ||
db935dbd NP |
2596 | BUG_ON(busiest == this_rq); |
2597 | ||
1da177e4 | 2598 | schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance); |
d6d5cfaf NP |
2599 | |
2600 | nr_moved = 0; | |
2601 | if (busiest->nr_running > 1) { | |
2602 | /* Attempt to move tasks */ | |
2603 | double_lock_balance(this_rq, busiest); | |
2604 | nr_moved = move_tasks(this_rq, this_cpu, busiest, | |
2dd73a4f | 2605 | minus_1_or_zero(busiest->nr_running), |
81026794 | 2606 | imbalance, sd, NEWLY_IDLE, NULL); |
d6d5cfaf NP |
2607 | spin_unlock(&busiest->lock); |
2608 | } | |
2609 | ||
5969fe06 | 2610 | if (!nr_moved) { |
1da177e4 | 2611 | schedstat_inc(sd, lb_failed[NEWLY_IDLE]); |
5969fe06 NP |
2612 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
2613 | return -1; | |
2614 | } else | |
16cfb1c0 | 2615 | sd->nr_balance_failed = 0; |
1da177e4 | 2616 | |
1da177e4 | 2617 | return nr_moved; |
16cfb1c0 NP |
2618 | |
2619 | out_balanced: | |
2620 | schedstat_inc(sd, lb_balanced[NEWLY_IDLE]); | |
5c45bf27 | 2621 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && !sched_smt_power_savings) |
5969fe06 | 2622 | return -1; |
16cfb1c0 NP |
2623 | sd->nr_balance_failed = 0; |
2624 | return 0; | |
1da177e4 LT |
2625 | } |
2626 | ||
2627 | /* | |
2628 | * idle_balance is called by schedule() if this_cpu is about to become | |
2629 | * idle. Attempts to pull tasks from other CPUs. | |
2630 | */ | |
858119e1 | 2631 | static void idle_balance(int this_cpu, runqueue_t *this_rq) |
1da177e4 LT |
2632 | { |
2633 | struct sched_domain *sd; | |
2634 | ||
2635 | for_each_domain(this_cpu, sd) { | |
2636 | if (sd->flags & SD_BALANCE_NEWIDLE) { | |
2637 | if (load_balance_newidle(this_cpu, this_rq, sd)) { | |
2638 | /* We've pulled tasks over so stop searching */ | |
2639 | break; | |
2640 | } | |
2641 | } | |
2642 | } | |
2643 | } | |
2644 | ||
2645 | /* | |
2646 | * active_load_balance is run by migration threads. It pushes running tasks | |
2647 | * off the busiest CPU onto idle CPUs. It requires at least 1 task to be | |
2648 | * running on each physical CPU where possible, and avoids physical / | |
2649 | * logical imbalances. | |
2650 | * | |
2651 | * Called with busiest_rq locked. | |
2652 | */ | |
2653 | static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu) | |
2654 | { | |
2655 | struct sched_domain *sd; | |
1da177e4 | 2656 | runqueue_t *target_rq; |
39507451 NP |
2657 | int target_cpu = busiest_rq->push_cpu; |
2658 | ||
2659 | if (busiest_rq->nr_running <= 1) | |
2660 | /* no task to move */ | |
2661 | return; | |
2662 | ||
2663 | target_rq = cpu_rq(target_cpu); | |
1da177e4 LT |
2664 | |
2665 | /* | |
39507451 NP |
2666 | * This condition is "impossible", if it occurs |
2667 | * we need to fix it. Originally reported by | |
2668 | * Bjorn Helgaas on a 128-cpu setup. | |
1da177e4 | 2669 | */ |
39507451 | 2670 | BUG_ON(busiest_rq == target_rq); |
1da177e4 | 2671 | |
39507451 NP |
2672 | /* move a task from busiest_rq to target_rq */ |
2673 | double_lock_balance(busiest_rq, target_rq); | |
2674 | ||
2675 | /* Search for an sd spanning us and the target CPU. */ | |
c96d145e | 2676 | for_each_domain(target_cpu, sd) { |
39507451 NP |
2677 | if ((sd->flags & SD_LOAD_BALANCE) && |
2678 | cpu_isset(busiest_cpu, sd->span)) | |
2679 | break; | |
c96d145e | 2680 | } |
39507451 NP |
2681 | |
2682 | if (unlikely(sd == NULL)) | |
2683 | goto out; | |
2684 | ||
2685 | schedstat_inc(sd, alb_cnt); | |
2686 | ||
2dd73a4f PW |
2687 | if (move_tasks(target_rq, target_cpu, busiest_rq, 1, |
2688 | RTPRIO_TO_LOAD_WEIGHT(100), sd, SCHED_IDLE, NULL)) | |
39507451 NP |
2689 | schedstat_inc(sd, alb_pushed); |
2690 | else | |
2691 | schedstat_inc(sd, alb_failed); | |
2692 | out: | |
2693 | spin_unlock(&target_rq->lock); | |
1da177e4 LT |
2694 | } |
2695 | ||
2696 | /* | |
2697 | * rebalance_tick will get called every timer tick, on every CPU. | |
2698 | * | |
2699 | * It checks each scheduling domain to see if it is due to be balanced, | |
2700 | * and initiates a balancing operation if so. | |
2701 | * | |
2702 | * Balancing parameters are set up in arch_init_sched_domains. | |
2703 | */ | |
2704 | ||
2705 | /* Don't have all balancing operations going off at once */ | |
2706 | #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS) | |
2707 | ||
2708 | static void rebalance_tick(int this_cpu, runqueue_t *this_rq, | |
2709 | enum idle_type idle) | |
2710 | { | |
2711 | unsigned long old_load, this_load; | |
2712 | unsigned long j = jiffies + CPU_OFFSET(this_cpu); | |
2713 | struct sched_domain *sd; | |
7897986b | 2714 | int i; |
1da177e4 | 2715 | |
2dd73a4f | 2716 | this_load = this_rq->raw_weighted_load; |
7897986b NP |
2717 | /* Update our load */ |
2718 | for (i = 0; i < 3; i++) { | |
2719 | unsigned long new_load = this_load; | |
2720 | int scale = 1 << i; | |
2721 | old_load = this_rq->cpu_load[i]; | |
2722 | /* | |
2723 | * Round up the averaging division if load is increasing. This | |
2724 | * prevents us from getting stuck on 9 if the load is 10, for | |
2725 | * example. | |
2726 | */ | |
2727 | if (new_load > old_load) | |
2728 | new_load += scale-1; | |
2729 | this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale; | |
2730 | } | |
1da177e4 LT |
2731 | |
2732 | for_each_domain(this_cpu, sd) { | |
2733 | unsigned long interval; | |
2734 | ||
2735 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
2736 | continue; | |
2737 | ||
2738 | interval = sd->balance_interval; | |
2739 | if (idle != SCHED_IDLE) | |
2740 | interval *= sd->busy_factor; | |
2741 | ||
2742 | /* scale ms to jiffies */ | |
2743 | interval = msecs_to_jiffies(interval); | |
2744 | if (unlikely(!interval)) | |
2745 | interval = 1; | |
2746 | ||
2747 | if (j - sd->last_balance >= interval) { | |
2748 | if (load_balance(this_cpu, this_rq, sd, idle)) { | |
fa3b6ddc SS |
2749 | /* |
2750 | * We've pulled tasks over so either we're no | |
5969fe06 NP |
2751 | * longer idle, or one of our SMT siblings is |
2752 | * not idle. | |
2753 | */ | |
1da177e4 LT |
2754 | idle = NOT_IDLE; |
2755 | } | |
2756 | sd->last_balance += interval; | |
2757 | } | |
2758 | } | |
2759 | } | |
2760 | #else | |
2761 | /* | |
2762 | * on UP we do not need to balance between CPUs: | |
2763 | */ | |
2764 | static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle) | |
2765 | { | |
2766 | } | |
2767 | static inline void idle_balance(int cpu, runqueue_t *rq) | |
2768 | { | |
2769 | } | |
2770 | #endif | |
2771 | ||
2772 | static inline int wake_priority_sleeper(runqueue_t *rq) | |
2773 | { | |
2774 | int ret = 0; | |
2775 | #ifdef CONFIG_SCHED_SMT | |
2776 | spin_lock(&rq->lock); | |
2777 | /* | |
2778 | * If an SMT sibling task has been put to sleep for priority | |
2779 | * reasons reschedule the idle task to see if it can now run. | |
2780 | */ | |
2781 | if (rq->nr_running) { | |
2782 | resched_task(rq->idle); | |
2783 | ret = 1; | |
2784 | } | |
2785 | spin_unlock(&rq->lock); | |
2786 | #endif | |
2787 | return ret; | |
2788 | } | |
2789 | ||
2790 | DEFINE_PER_CPU(struct kernel_stat, kstat); | |
2791 | ||
2792 | EXPORT_PER_CPU_SYMBOL(kstat); | |
2793 | ||
2794 | /* | |
2795 | * This is called on clock ticks and on context switches. | |
2796 | * Bank in p->sched_time the ns elapsed since the last tick or switch. | |
2797 | */ | |
2798 | static inline void update_cpu_clock(task_t *p, runqueue_t *rq, | |
2799 | unsigned long long now) | |
2800 | { | |
2801 | unsigned long long last = max(p->timestamp, rq->timestamp_last_tick); | |
2802 | p->sched_time += now - last; | |
2803 | } | |
2804 | ||
2805 | /* | |
2806 | * Return current->sched_time plus any more ns on the sched_clock | |
2807 | * that have not yet been banked. | |
2808 | */ | |
2809 | unsigned long long current_sched_time(const task_t *tsk) | |
2810 | { | |
2811 | unsigned long long ns; | |
2812 | unsigned long flags; | |
2813 | local_irq_save(flags); | |
2814 | ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick); | |
2815 | ns = tsk->sched_time + (sched_clock() - ns); | |
2816 | local_irq_restore(flags); | |
2817 | return ns; | |
2818 | } | |
2819 | ||
f1adad78 LT |
2820 | /* |
2821 | * We place interactive tasks back into the active array, if possible. | |
2822 | * | |
2823 | * To guarantee that this does not starve expired tasks we ignore the | |
2824 | * interactivity of a task if the first expired task had to wait more | |
2825 | * than a 'reasonable' amount of time. This deadline timeout is | |
2826 | * load-dependent, as the frequency of array switched decreases with | |
2827 | * increasing number of running tasks. We also ignore the interactivity | |
2828 | * if a better static_prio task has expired: | |
2829 | */ | |
2830 | #define EXPIRED_STARVING(rq) \ | |
2831 | ((STARVATION_LIMIT && ((rq)->expired_timestamp && \ | |
2832 | (jiffies - (rq)->expired_timestamp >= \ | |
2833 | STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \ | |
2834 | ((rq)->curr->static_prio > (rq)->best_expired_prio)) | |
2835 | ||
1da177e4 LT |
2836 | /* |
2837 | * Account user cpu time to a process. | |
2838 | * @p: the process that the cpu time gets accounted to | |
2839 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
2840 | * @cputime: the cpu time spent in user space since the last update | |
2841 | */ | |
2842 | void account_user_time(struct task_struct *p, cputime_t cputime) | |
2843 | { | |
2844 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
2845 | cputime64_t tmp; | |
2846 | ||
2847 | p->utime = cputime_add(p->utime, cputime); | |
2848 | ||
2849 | /* Add user time to cpustat. */ | |
2850 | tmp = cputime_to_cputime64(cputime); | |
2851 | if (TASK_NICE(p) > 0) | |
2852 | cpustat->nice = cputime64_add(cpustat->nice, tmp); | |
2853 | else | |
2854 | cpustat->user = cputime64_add(cpustat->user, tmp); | |
2855 | } | |
2856 | ||
2857 | /* | |
2858 | * Account system cpu time to a process. | |
2859 | * @p: the process that the cpu time gets accounted to | |
2860 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
2861 | * @cputime: the cpu time spent in kernel space since the last update | |
2862 | */ | |
2863 | void account_system_time(struct task_struct *p, int hardirq_offset, | |
2864 | cputime_t cputime) | |
2865 | { | |
2866 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
2867 | runqueue_t *rq = this_rq(); | |
2868 | cputime64_t tmp; | |
2869 | ||
2870 | p->stime = cputime_add(p->stime, cputime); | |
2871 | ||
2872 | /* Add system time to cpustat. */ | |
2873 | tmp = cputime_to_cputime64(cputime); | |
2874 | if (hardirq_count() - hardirq_offset) | |
2875 | cpustat->irq = cputime64_add(cpustat->irq, tmp); | |
2876 | else if (softirq_count()) | |
2877 | cpustat->softirq = cputime64_add(cpustat->softirq, tmp); | |
2878 | else if (p != rq->idle) | |
2879 | cpustat->system = cputime64_add(cpustat->system, tmp); | |
2880 | else if (atomic_read(&rq->nr_iowait) > 0) | |
2881 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); | |
2882 | else | |
2883 | cpustat->idle = cputime64_add(cpustat->idle, tmp); | |
2884 | /* Account for system time used */ | |
2885 | acct_update_integrals(p); | |
1da177e4 LT |
2886 | } |
2887 | ||
2888 | /* | |
2889 | * Account for involuntary wait time. | |
2890 | * @p: the process from which the cpu time has been stolen | |
2891 | * @steal: the cpu time spent in involuntary wait | |
2892 | */ | |
2893 | void account_steal_time(struct task_struct *p, cputime_t steal) | |
2894 | { | |
2895 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
2896 | cputime64_t tmp = cputime_to_cputime64(steal); | |
2897 | runqueue_t *rq = this_rq(); | |
2898 | ||
2899 | if (p == rq->idle) { | |
2900 | p->stime = cputime_add(p->stime, steal); | |
2901 | if (atomic_read(&rq->nr_iowait) > 0) | |
2902 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); | |
2903 | else | |
2904 | cpustat->idle = cputime64_add(cpustat->idle, tmp); | |
2905 | } else | |
2906 | cpustat->steal = cputime64_add(cpustat->steal, tmp); | |
2907 | } | |
2908 | ||
2909 | /* | |
2910 | * This function gets called by the timer code, with HZ frequency. | |
2911 | * We call it with interrupts disabled. | |
2912 | * | |
2913 | * It also gets called by the fork code, when changing the parent's | |
2914 | * timeslices. | |
2915 | */ | |
2916 | void scheduler_tick(void) | |
2917 | { | |
2918 | int cpu = smp_processor_id(); | |
2919 | runqueue_t *rq = this_rq(); | |
2920 | task_t *p = current; | |
2921 | unsigned long long now = sched_clock(); | |
2922 | ||
2923 | update_cpu_clock(p, rq, now); | |
2924 | ||
2925 | rq->timestamp_last_tick = now; | |
2926 | ||
2927 | if (p == rq->idle) { | |
2928 | if (wake_priority_sleeper(rq)) | |
2929 | goto out; | |
2930 | rebalance_tick(cpu, rq, SCHED_IDLE); | |
2931 | return; | |
2932 | } | |
2933 | ||
2934 | /* Task might have expired already, but not scheduled off yet */ | |
2935 | if (p->array != rq->active) { | |
2936 | set_tsk_need_resched(p); | |
2937 | goto out; | |
2938 | } | |
2939 | spin_lock(&rq->lock); | |
2940 | /* | |
2941 | * The task was running during this tick - update the | |
2942 | * time slice counter. Note: we do not update a thread's | |
2943 | * priority until it either goes to sleep or uses up its | |
2944 | * timeslice. This makes it possible for interactive tasks | |
2945 | * to use up their timeslices at their highest priority levels. | |
2946 | */ | |
2947 | if (rt_task(p)) { | |
2948 | /* | |
2949 | * RR tasks need a special form of timeslice management. | |
2950 | * FIFO tasks have no timeslices. | |
2951 | */ | |
2952 | if ((p->policy == SCHED_RR) && !--p->time_slice) { | |
2953 | p->time_slice = task_timeslice(p); | |
2954 | p->first_time_slice = 0; | |
2955 | set_tsk_need_resched(p); | |
2956 | ||
2957 | /* put it at the end of the queue: */ | |
2958 | requeue_task(p, rq->active); | |
2959 | } | |
2960 | goto out_unlock; | |
2961 | } | |
2962 | if (!--p->time_slice) { | |
2963 | dequeue_task(p, rq->active); | |
2964 | set_tsk_need_resched(p); | |
2965 | p->prio = effective_prio(p); | |
2966 | p->time_slice = task_timeslice(p); | |
2967 | p->first_time_slice = 0; | |
2968 | ||
2969 | if (!rq->expired_timestamp) | |
2970 | rq->expired_timestamp = jiffies; | |
f1adad78 | 2971 | if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) { |
1da177e4 LT |
2972 | enqueue_task(p, rq->expired); |
2973 | if (p->static_prio < rq->best_expired_prio) | |
2974 | rq->best_expired_prio = p->static_prio; | |
2975 | } else | |
2976 | enqueue_task(p, rq->active); | |
2977 | } else { | |
2978 | /* | |
2979 | * Prevent a too long timeslice allowing a task to monopolize | |
2980 | * the CPU. We do this by splitting up the timeslice into | |
2981 | * smaller pieces. | |
2982 | * | |
2983 | * Note: this does not mean the task's timeslices expire or | |
2984 | * get lost in any way, they just might be preempted by | |
2985 | * another task of equal priority. (one with higher | |
2986 | * priority would have preempted this task already.) We | |
2987 | * requeue this task to the end of the list on this priority | |
2988 | * level, which is in essence a round-robin of tasks with | |
2989 | * equal priority. | |
2990 | * | |
2991 | * This only applies to tasks in the interactive | |
2992 | * delta range with at least TIMESLICE_GRANULARITY to requeue. | |
2993 | */ | |
2994 | if (TASK_INTERACTIVE(p) && !((task_timeslice(p) - | |
2995 | p->time_slice) % TIMESLICE_GRANULARITY(p)) && | |
2996 | (p->time_slice >= TIMESLICE_GRANULARITY(p)) && | |
2997 | (p->array == rq->active)) { | |
2998 | ||
2999 | requeue_task(p, rq->active); | |
3000 | set_tsk_need_resched(p); | |
3001 | } | |
3002 | } | |
3003 | out_unlock: | |
3004 | spin_unlock(&rq->lock); | |
3005 | out: | |
3006 | rebalance_tick(cpu, rq, NOT_IDLE); | |
3007 | } | |
3008 | ||
3009 | #ifdef CONFIG_SCHED_SMT | |
fc38ed75 CK |
3010 | static inline void wakeup_busy_runqueue(runqueue_t *rq) |
3011 | { | |
3012 | /* If an SMT runqueue is sleeping due to priority reasons wake it up */ | |
3013 | if (rq->curr == rq->idle && rq->nr_running) | |
3014 | resched_task(rq->idle); | |
3015 | } | |
3016 | ||
c96d145e CK |
3017 | /* |
3018 | * Called with interrupt disabled and this_rq's runqueue locked. | |
3019 | */ | |
3020 | static void wake_sleeping_dependent(int this_cpu) | |
1da177e4 | 3021 | { |
41c7ce9a | 3022 | struct sched_domain *tmp, *sd = NULL; |
1da177e4 LT |
3023 | int i; |
3024 | ||
c96d145e CK |
3025 | for_each_domain(this_cpu, tmp) { |
3026 | if (tmp->flags & SD_SHARE_CPUPOWER) { | |
41c7ce9a | 3027 | sd = tmp; |
c96d145e CK |
3028 | break; |
3029 | } | |
3030 | } | |
41c7ce9a NP |
3031 | |
3032 | if (!sd) | |
1da177e4 LT |
3033 | return; |
3034 | ||
c96d145e | 3035 | for_each_cpu_mask(i, sd->span) { |
1da177e4 LT |
3036 | runqueue_t *smt_rq = cpu_rq(i); |
3037 | ||
c96d145e CK |
3038 | if (i == this_cpu) |
3039 | continue; | |
3040 | if (unlikely(!spin_trylock(&smt_rq->lock))) | |
3041 | continue; | |
3042 | ||
fc38ed75 | 3043 | wakeup_busy_runqueue(smt_rq); |
c96d145e | 3044 | spin_unlock(&smt_rq->lock); |
1da177e4 | 3045 | } |
1da177e4 LT |
3046 | } |
3047 | ||
67f9a619 IM |
3048 | /* |
3049 | * number of 'lost' timeslices this task wont be able to fully | |
3050 | * utilize, if another task runs on a sibling. This models the | |
3051 | * slowdown effect of other tasks running on siblings: | |
3052 | */ | |
3053 | static inline unsigned long smt_slice(task_t *p, struct sched_domain *sd) | |
3054 | { | |
3055 | return p->time_slice * (100 - sd->per_cpu_gain) / 100; | |
3056 | } | |
3057 | ||
c96d145e CK |
3058 | /* |
3059 | * To minimise lock contention and not have to drop this_rq's runlock we only | |
3060 | * trylock the sibling runqueues and bypass those runqueues if we fail to | |
3061 | * acquire their lock. As we only trylock the normal locking order does not | |
3062 | * need to be obeyed. | |
3063 | */ | |
3064 | static int dependent_sleeper(int this_cpu, runqueue_t *this_rq, task_t *p) | |
1da177e4 | 3065 | { |
41c7ce9a | 3066 | struct sched_domain *tmp, *sd = NULL; |
1da177e4 | 3067 | int ret = 0, i; |
1da177e4 | 3068 | |
c96d145e CK |
3069 | /* kernel/rt threads do not participate in dependent sleeping */ |
3070 | if (!p->mm || rt_task(p)) | |
3071 | return 0; | |
3072 | ||
3073 | for_each_domain(this_cpu, tmp) { | |
3074 | if (tmp->flags & SD_SHARE_CPUPOWER) { | |
41c7ce9a | 3075 | sd = tmp; |
c96d145e CK |
3076 | break; |
3077 | } | |
3078 | } | |
41c7ce9a NP |
3079 | |
3080 | if (!sd) | |
1da177e4 LT |
3081 | return 0; |
3082 | ||
c96d145e CK |
3083 | for_each_cpu_mask(i, sd->span) { |
3084 | runqueue_t *smt_rq; | |
3085 | task_t *smt_curr; | |
1da177e4 | 3086 | |
c96d145e CK |
3087 | if (i == this_cpu) |
3088 | continue; | |
1da177e4 | 3089 | |
c96d145e CK |
3090 | smt_rq = cpu_rq(i); |
3091 | if (unlikely(!spin_trylock(&smt_rq->lock))) | |
3092 | continue; | |
1da177e4 | 3093 | |
c96d145e | 3094 | smt_curr = smt_rq->curr; |
1da177e4 | 3095 | |
c96d145e CK |
3096 | if (!smt_curr->mm) |
3097 | goto unlock; | |
fc38ed75 | 3098 | |
1da177e4 LT |
3099 | /* |
3100 | * If a user task with lower static priority than the | |
3101 | * running task on the SMT sibling is trying to schedule, | |
3102 | * delay it till there is proportionately less timeslice | |
3103 | * left of the sibling task to prevent a lower priority | |
3104 | * task from using an unfair proportion of the | |
3105 | * physical cpu's resources. -ck | |
3106 | */ | |
fc38ed75 CK |
3107 | if (rt_task(smt_curr)) { |
3108 | /* | |
3109 | * With real time tasks we run non-rt tasks only | |
3110 | * per_cpu_gain% of the time. | |
3111 | */ | |
3112 | if ((jiffies % DEF_TIMESLICE) > | |
3113 | (sd->per_cpu_gain * DEF_TIMESLICE / 100)) | |
3114 | ret = 1; | |
c96d145e | 3115 | } else { |
67f9a619 IM |
3116 | if (smt_curr->static_prio < p->static_prio && |
3117 | !TASK_PREEMPTS_CURR(p, smt_rq) && | |
3118 | smt_slice(smt_curr, sd) > task_timeslice(p)) | |
fc38ed75 | 3119 | ret = 1; |
fc38ed75 | 3120 | } |
c96d145e CK |
3121 | unlock: |
3122 | spin_unlock(&smt_rq->lock); | |
1da177e4 | 3123 | } |
1da177e4 LT |
3124 | return ret; |
3125 | } | |
3126 | #else | |
c96d145e | 3127 | static inline void wake_sleeping_dependent(int this_cpu) |
1da177e4 LT |
3128 | { |
3129 | } | |
3130 | ||
c96d145e CK |
3131 | static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq, |
3132 | task_t *p) | |
1da177e4 LT |
3133 | { |
3134 | return 0; | |
3135 | } | |
3136 | #endif | |
3137 | ||
3138 | #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT) | |
3139 | ||
3140 | void fastcall add_preempt_count(int val) | |
3141 | { | |
3142 | /* | |
3143 | * Underflow? | |
3144 | */ | |
be5b4fbd | 3145 | BUG_ON((preempt_count() < 0)); |
1da177e4 LT |
3146 | preempt_count() += val; |
3147 | /* | |
3148 | * Spinlock count overflowing soon? | |
3149 | */ | |
3150 | BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10); | |
3151 | } | |
3152 | EXPORT_SYMBOL(add_preempt_count); | |
3153 | ||
3154 | void fastcall sub_preempt_count(int val) | |
3155 | { | |
3156 | /* | |
3157 | * Underflow? | |
3158 | */ | |
3159 | BUG_ON(val > preempt_count()); | |
3160 | /* | |
3161 | * Is the spinlock portion underflowing? | |
3162 | */ | |
3163 | BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK)); | |
3164 | preempt_count() -= val; | |
3165 | } | |
3166 | EXPORT_SYMBOL(sub_preempt_count); | |
3167 | ||
3168 | #endif | |
3169 | ||
3dee386e CK |
3170 | static inline int interactive_sleep(enum sleep_type sleep_type) |
3171 | { | |
3172 | return (sleep_type == SLEEP_INTERACTIVE || | |
3173 | sleep_type == SLEEP_INTERRUPTED); | |
3174 | } | |
3175 | ||
1da177e4 LT |
3176 | /* |
3177 | * schedule() is the main scheduler function. | |
3178 | */ | |
3179 | asmlinkage void __sched schedule(void) | |
3180 | { | |
3181 | long *switch_count; | |
3182 | task_t *prev, *next; | |
3183 | runqueue_t *rq; | |
3184 | prio_array_t *array; | |
3185 | struct list_head *queue; | |
3186 | unsigned long long now; | |
3187 | unsigned long run_time; | |
a3464a10 | 3188 | int cpu, idx, new_prio; |
1da177e4 LT |
3189 | |
3190 | /* | |
3191 | * Test if we are atomic. Since do_exit() needs to call into | |
3192 | * schedule() atomically, we ignore that path for now. | |
3193 | * Otherwise, whine if we are scheduling when we should not be. | |
3194 | */ | |
77e4bfbc AM |
3195 | if (unlikely(in_atomic() && !current->exit_state)) { |
3196 | printk(KERN_ERR "BUG: scheduling while atomic: " | |
3197 | "%s/0x%08x/%d\n", | |
3198 | current->comm, preempt_count(), current->pid); | |
3199 | dump_stack(); | |
1da177e4 LT |
3200 | } |
3201 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); | |
3202 | ||
3203 | need_resched: | |
3204 | preempt_disable(); | |
3205 | prev = current; | |
3206 | release_kernel_lock(prev); | |
3207 | need_resched_nonpreemptible: | |
3208 | rq = this_rq(); | |
3209 | ||
3210 | /* | |
3211 | * The idle thread is not allowed to schedule! | |
3212 | * Remove this check after it has been exercised a bit. | |
3213 | */ | |
3214 | if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) { | |
3215 | printk(KERN_ERR "bad: scheduling from the idle thread!\n"); | |
3216 | dump_stack(); | |
3217 | } | |
3218 | ||
3219 | schedstat_inc(rq, sched_cnt); | |
3220 | now = sched_clock(); | |
238628ed | 3221 | if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) { |
1da177e4 | 3222 | run_time = now - prev->timestamp; |
238628ed | 3223 | if (unlikely((long long)(now - prev->timestamp) < 0)) |
1da177e4 LT |
3224 | run_time = 0; |
3225 | } else | |
3226 | run_time = NS_MAX_SLEEP_AVG; | |
3227 | ||
3228 | /* | |
3229 | * Tasks charged proportionately less run_time at high sleep_avg to | |
3230 | * delay them losing their interactive status | |
3231 | */ | |
3232 | run_time /= (CURRENT_BONUS(prev) ? : 1); | |
3233 | ||
3234 | spin_lock_irq(&rq->lock); | |
3235 | ||
3236 | if (unlikely(prev->flags & PF_DEAD)) | |
3237 | prev->state = EXIT_DEAD; | |
3238 | ||
3239 | switch_count = &prev->nivcsw; | |
3240 | if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { | |
3241 | switch_count = &prev->nvcsw; | |
3242 | if (unlikely((prev->state & TASK_INTERRUPTIBLE) && | |
3243 | unlikely(signal_pending(prev)))) | |
3244 | prev->state = TASK_RUNNING; | |
3245 | else { | |
3246 | if (prev->state == TASK_UNINTERRUPTIBLE) | |
3247 | rq->nr_uninterruptible++; | |
3248 | deactivate_task(prev, rq); | |
3249 | } | |
3250 | } | |
3251 | ||
3252 | cpu = smp_processor_id(); | |
3253 | if (unlikely(!rq->nr_running)) { | |
1da177e4 LT |
3254 | idle_balance(cpu, rq); |
3255 | if (!rq->nr_running) { | |
3256 | next = rq->idle; | |
3257 | rq->expired_timestamp = 0; | |
c96d145e | 3258 | wake_sleeping_dependent(cpu); |
1da177e4 LT |
3259 | goto switch_tasks; |
3260 | } | |
1da177e4 LT |
3261 | } |
3262 | ||
3263 | array = rq->active; | |
3264 | if (unlikely(!array->nr_active)) { | |
3265 | /* | |
3266 | * Switch the active and expired arrays. | |
3267 | */ | |
3268 | schedstat_inc(rq, sched_switch); | |
3269 | rq->active = rq->expired; | |
3270 | rq->expired = array; | |
3271 | array = rq->active; | |
3272 | rq->expired_timestamp = 0; | |
3273 | rq->best_expired_prio = MAX_PRIO; | |
3274 | } | |
3275 | ||
3276 | idx = sched_find_first_bit(array->bitmap); | |
3277 | queue = array->queue + idx; | |
3278 | next = list_entry(queue->next, task_t, run_list); | |
3279 | ||
3dee386e | 3280 | if (!rt_task(next) && interactive_sleep(next->sleep_type)) { |
1da177e4 | 3281 | unsigned long long delta = now - next->timestamp; |
238628ed | 3282 | if (unlikely((long long)(now - next->timestamp) < 0)) |
1da177e4 LT |
3283 | delta = 0; |
3284 | ||
3dee386e | 3285 | if (next->sleep_type == SLEEP_INTERACTIVE) |
1da177e4 LT |
3286 | delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128; |
3287 | ||
3288 | array = next->array; | |
a3464a10 CS |
3289 | new_prio = recalc_task_prio(next, next->timestamp + delta); |
3290 | ||
3291 | if (unlikely(next->prio != new_prio)) { | |
3292 | dequeue_task(next, array); | |
3293 | next->prio = new_prio; | |
3294 | enqueue_task(next, array); | |
7c4bb1f9 | 3295 | } |
1da177e4 | 3296 | } |
3dee386e | 3297 | next->sleep_type = SLEEP_NORMAL; |
c96d145e CK |
3298 | if (dependent_sleeper(cpu, rq, next)) |
3299 | next = rq->idle; | |
1da177e4 LT |
3300 | switch_tasks: |
3301 | if (next == rq->idle) | |
3302 | schedstat_inc(rq, sched_goidle); | |
3303 | prefetch(next); | |
383f2835 | 3304 | prefetch_stack(next); |
1da177e4 LT |
3305 | clear_tsk_need_resched(prev); |
3306 | rcu_qsctr_inc(task_cpu(prev)); | |
3307 | ||
3308 | update_cpu_clock(prev, rq, now); | |
3309 | ||
3310 | prev->sleep_avg -= run_time; | |
3311 | if ((long)prev->sleep_avg <= 0) | |
3312 | prev->sleep_avg = 0; | |
3313 | prev->timestamp = prev->last_ran = now; | |
3314 | ||
3315 | sched_info_switch(prev, next); | |
3316 | if (likely(prev != next)) { | |
3317 | next->timestamp = now; | |
3318 | rq->nr_switches++; | |
3319 | rq->curr = next; | |
3320 | ++*switch_count; | |
3321 | ||
4866cde0 | 3322 | prepare_task_switch(rq, next); |
1da177e4 LT |
3323 | prev = context_switch(rq, prev, next); |
3324 | barrier(); | |
4866cde0 NP |
3325 | /* |
3326 | * this_rq must be evaluated again because prev may have moved | |
3327 | * CPUs since it called schedule(), thus the 'rq' on its stack | |
3328 | * frame will be invalid. | |
3329 | */ | |
3330 | finish_task_switch(this_rq(), prev); | |
1da177e4 LT |
3331 | } else |
3332 | spin_unlock_irq(&rq->lock); | |
3333 | ||
3334 | prev = current; | |
3335 | if (unlikely(reacquire_kernel_lock(prev) < 0)) | |
3336 | goto need_resched_nonpreemptible; | |
3337 | preempt_enable_no_resched(); | |
3338 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
3339 | goto need_resched; | |
3340 | } | |
3341 | ||
3342 | EXPORT_SYMBOL(schedule); | |
3343 | ||
3344 | #ifdef CONFIG_PREEMPT | |
3345 | /* | |
3346 | * this is is the entry point to schedule() from in-kernel preemption | |
3347 | * off of preempt_enable. Kernel preemptions off return from interrupt | |
3348 | * occur there and call schedule directly. | |
3349 | */ | |
3350 | asmlinkage void __sched preempt_schedule(void) | |
3351 | { | |
3352 | struct thread_info *ti = current_thread_info(); | |
3353 | #ifdef CONFIG_PREEMPT_BKL | |
3354 | struct task_struct *task = current; | |
3355 | int saved_lock_depth; | |
3356 | #endif | |
3357 | /* | |
3358 | * If there is a non-zero preempt_count or interrupts are disabled, | |
3359 | * we do not want to preempt the current task. Just return.. | |
3360 | */ | |
3361 | if (unlikely(ti->preempt_count || irqs_disabled())) | |
3362 | return; | |
3363 | ||
3364 | need_resched: | |
3365 | add_preempt_count(PREEMPT_ACTIVE); | |
3366 | /* | |
3367 | * We keep the big kernel semaphore locked, but we | |
3368 | * clear ->lock_depth so that schedule() doesnt | |
3369 | * auto-release the semaphore: | |
3370 | */ | |
3371 | #ifdef CONFIG_PREEMPT_BKL | |
3372 | saved_lock_depth = task->lock_depth; | |
3373 | task->lock_depth = -1; | |
3374 | #endif | |
3375 | schedule(); | |
3376 | #ifdef CONFIG_PREEMPT_BKL | |
3377 | task->lock_depth = saved_lock_depth; | |
3378 | #endif | |
3379 | sub_preempt_count(PREEMPT_ACTIVE); | |
3380 | ||
3381 | /* we could miss a preemption opportunity between schedule and now */ | |
3382 | barrier(); | |
3383 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
3384 | goto need_resched; | |
3385 | } | |
3386 | ||
3387 | EXPORT_SYMBOL(preempt_schedule); | |
3388 | ||
3389 | /* | |
3390 | * this is is the entry point to schedule() from kernel preemption | |
3391 | * off of irq context. | |
3392 | * Note, that this is called and return with irqs disabled. This will | |
3393 | * protect us against recursive calling from irq. | |
3394 | */ | |
3395 | asmlinkage void __sched preempt_schedule_irq(void) | |
3396 | { | |
3397 | struct thread_info *ti = current_thread_info(); | |
3398 | #ifdef CONFIG_PREEMPT_BKL | |
3399 | struct task_struct *task = current; | |
3400 | int saved_lock_depth; | |
3401 | #endif | |
3402 | /* Catch callers which need to be fixed*/ | |
3403 | BUG_ON(ti->preempt_count || !irqs_disabled()); | |
3404 | ||
3405 | need_resched: | |
3406 | add_preempt_count(PREEMPT_ACTIVE); | |
3407 | /* | |
3408 | * We keep the big kernel semaphore locked, but we | |
3409 | * clear ->lock_depth so that schedule() doesnt | |
3410 | * auto-release the semaphore: | |
3411 | */ | |
3412 | #ifdef CONFIG_PREEMPT_BKL | |
3413 | saved_lock_depth = task->lock_depth; | |
3414 | task->lock_depth = -1; | |
3415 | #endif | |
3416 | local_irq_enable(); | |
3417 | schedule(); | |
3418 | local_irq_disable(); | |
3419 | #ifdef CONFIG_PREEMPT_BKL | |
3420 | task->lock_depth = saved_lock_depth; | |
3421 | #endif | |
3422 | sub_preempt_count(PREEMPT_ACTIVE); | |
3423 | ||
3424 | /* we could miss a preemption opportunity between schedule and now */ | |
3425 | barrier(); | |
3426 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
3427 | goto need_resched; | |
3428 | } | |
3429 | ||
3430 | #endif /* CONFIG_PREEMPT */ | |
3431 | ||
95cdf3b7 IM |
3432 | int default_wake_function(wait_queue_t *curr, unsigned mode, int sync, |
3433 | void *key) | |
1da177e4 | 3434 | { |
c43dc2fd | 3435 | task_t *p = curr->private; |
1da177e4 LT |
3436 | return try_to_wake_up(p, mode, sync); |
3437 | } | |
3438 | ||
3439 | EXPORT_SYMBOL(default_wake_function); | |
3440 | ||
3441 | /* | |
3442 | * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just | |
3443 | * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve | |
3444 | * number) then we wake all the non-exclusive tasks and one exclusive task. | |
3445 | * | |
3446 | * There are circumstances in which we can try to wake a task which has already | |
3447 | * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns | |
3448 | * zero in this (rare) case, and we handle it by continuing to scan the queue. | |
3449 | */ | |
3450 | static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, | |
3451 | int nr_exclusive, int sync, void *key) | |
3452 | { | |
3453 | struct list_head *tmp, *next; | |
3454 | ||
3455 | list_for_each_safe(tmp, next, &q->task_list) { | |
3456 | wait_queue_t *curr; | |
3457 | unsigned flags; | |
3458 | curr = list_entry(tmp, wait_queue_t, task_list); | |
3459 | flags = curr->flags; | |
3460 | if (curr->func(curr, mode, sync, key) && | |
3461 | (flags & WQ_FLAG_EXCLUSIVE) && | |
3462 | !--nr_exclusive) | |
3463 | break; | |
3464 | } | |
3465 | } | |
3466 | ||
3467 | /** | |
3468 | * __wake_up - wake up threads blocked on a waitqueue. | |
3469 | * @q: the waitqueue | |
3470 | * @mode: which threads | |
3471 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
67be2dd1 | 3472 | * @key: is directly passed to the wakeup function |
1da177e4 LT |
3473 | */ |
3474 | void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode, | |
95cdf3b7 | 3475 | int nr_exclusive, void *key) |
1da177e4 LT |
3476 | { |
3477 | unsigned long flags; | |
3478 | ||
3479 | spin_lock_irqsave(&q->lock, flags); | |
3480 | __wake_up_common(q, mode, nr_exclusive, 0, key); | |
3481 | spin_unlock_irqrestore(&q->lock, flags); | |
3482 | } | |
3483 | ||
3484 | EXPORT_SYMBOL(__wake_up); | |
3485 | ||
3486 | /* | |
3487 | * Same as __wake_up but called with the spinlock in wait_queue_head_t held. | |
3488 | */ | |
3489 | void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode) | |
3490 | { | |
3491 | __wake_up_common(q, mode, 1, 0, NULL); | |
3492 | } | |
3493 | ||
3494 | /** | |
67be2dd1 | 3495 | * __wake_up_sync - wake up threads blocked on a waitqueue. |
1da177e4 LT |
3496 | * @q: the waitqueue |
3497 | * @mode: which threads | |
3498 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
3499 | * | |
3500 | * The sync wakeup differs that the waker knows that it will schedule | |
3501 | * away soon, so while the target thread will be woken up, it will not | |
3502 | * be migrated to another CPU - ie. the two threads are 'synchronized' | |
3503 | * with each other. This can prevent needless bouncing between CPUs. | |
3504 | * | |
3505 | * On UP it can prevent extra preemption. | |
3506 | */ | |
95cdf3b7 IM |
3507 | void fastcall |
3508 | __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) | |
1da177e4 LT |
3509 | { |
3510 | unsigned long flags; | |
3511 | int sync = 1; | |
3512 | ||
3513 | if (unlikely(!q)) | |
3514 | return; | |
3515 | ||
3516 | if (unlikely(!nr_exclusive)) | |
3517 | sync = 0; | |
3518 | ||
3519 | spin_lock_irqsave(&q->lock, flags); | |
3520 | __wake_up_common(q, mode, nr_exclusive, sync, NULL); | |
3521 | spin_unlock_irqrestore(&q->lock, flags); | |
3522 | } | |
3523 | EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ | |
3524 | ||
3525 | void fastcall complete(struct completion *x) | |
3526 | { | |
3527 | unsigned long flags; | |
3528 | ||
3529 | spin_lock_irqsave(&x->wait.lock, flags); | |
3530 | x->done++; | |
3531 | __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, | |
3532 | 1, 0, NULL); | |
3533 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
3534 | } | |
3535 | EXPORT_SYMBOL(complete); | |
3536 | ||
3537 | void fastcall complete_all(struct completion *x) | |
3538 | { | |
3539 | unsigned long flags; | |
3540 | ||
3541 | spin_lock_irqsave(&x->wait.lock, flags); | |
3542 | x->done += UINT_MAX/2; | |
3543 | __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, | |
3544 | 0, 0, NULL); | |
3545 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
3546 | } | |
3547 | EXPORT_SYMBOL(complete_all); | |
3548 | ||
3549 | void fastcall __sched wait_for_completion(struct completion *x) | |
3550 | { | |
3551 | might_sleep(); | |
3552 | spin_lock_irq(&x->wait.lock); | |
3553 | if (!x->done) { | |
3554 | DECLARE_WAITQUEUE(wait, current); | |
3555 | ||
3556 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3557 | __add_wait_queue_tail(&x->wait, &wait); | |
3558 | do { | |
3559 | __set_current_state(TASK_UNINTERRUPTIBLE); | |
3560 | spin_unlock_irq(&x->wait.lock); | |
3561 | schedule(); | |
3562 | spin_lock_irq(&x->wait.lock); | |
3563 | } while (!x->done); | |
3564 | __remove_wait_queue(&x->wait, &wait); | |
3565 | } | |
3566 | x->done--; | |
3567 | spin_unlock_irq(&x->wait.lock); | |
3568 | } | |
3569 | EXPORT_SYMBOL(wait_for_completion); | |
3570 | ||
3571 | unsigned long fastcall __sched | |
3572 | wait_for_completion_timeout(struct completion *x, unsigned long timeout) | |
3573 | { | |
3574 | might_sleep(); | |
3575 | ||
3576 | spin_lock_irq(&x->wait.lock); | |
3577 | if (!x->done) { | |
3578 | DECLARE_WAITQUEUE(wait, current); | |
3579 | ||
3580 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3581 | __add_wait_queue_tail(&x->wait, &wait); | |
3582 | do { | |
3583 | __set_current_state(TASK_UNINTERRUPTIBLE); | |
3584 | spin_unlock_irq(&x->wait.lock); | |
3585 | timeout = schedule_timeout(timeout); | |
3586 | spin_lock_irq(&x->wait.lock); | |
3587 | if (!timeout) { | |
3588 | __remove_wait_queue(&x->wait, &wait); | |
3589 | goto out; | |
3590 | } | |
3591 | } while (!x->done); | |
3592 | __remove_wait_queue(&x->wait, &wait); | |
3593 | } | |
3594 | x->done--; | |
3595 | out: | |
3596 | spin_unlock_irq(&x->wait.lock); | |
3597 | return timeout; | |
3598 | } | |
3599 | EXPORT_SYMBOL(wait_for_completion_timeout); | |
3600 | ||
3601 | int fastcall __sched wait_for_completion_interruptible(struct completion *x) | |
3602 | { | |
3603 | int ret = 0; | |
3604 | ||
3605 | might_sleep(); | |
3606 | ||
3607 | spin_lock_irq(&x->wait.lock); | |
3608 | if (!x->done) { | |
3609 | DECLARE_WAITQUEUE(wait, current); | |
3610 | ||
3611 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3612 | __add_wait_queue_tail(&x->wait, &wait); | |
3613 | do { | |
3614 | if (signal_pending(current)) { | |
3615 | ret = -ERESTARTSYS; | |
3616 | __remove_wait_queue(&x->wait, &wait); | |
3617 | goto out; | |
3618 | } | |
3619 | __set_current_state(TASK_INTERRUPTIBLE); | |
3620 | spin_unlock_irq(&x->wait.lock); | |
3621 | schedule(); | |
3622 | spin_lock_irq(&x->wait.lock); | |
3623 | } while (!x->done); | |
3624 | __remove_wait_queue(&x->wait, &wait); | |
3625 | } | |
3626 | x->done--; | |
3627 | out: | |
3628 | spin_unlock_irq(&x->wait.lock); | |
3629 | ||
3630 | return ret; | |
3631 | } | |
3632 | EXPORT_SYMBOL(wait_for_completion_interruptible); | |
3633 | ||
3634 | unsigned long fastcall __sched | |
3635 | wait_for_completion_interruptible_timeout(struct completion *x, | |
3636 | unsigned long timeout) | |
3637 | { | |
3638 | might_sleep(); | |
3639 | ||
3640 | spin_lock_irq(&x->wait.lock); | |
3641 | if (!x->done) { | |
3642 | DECLARE_WAITQUEUE(wait, current); | |
3643 | ||
3644 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3645 | __add_wait_queue_tail(&x->wait, &wait); | |
3646 | do { | |
3647 | if (signal_pending(current)) { | |
3648 | timeout = -ERESTARTSYS; | |
3649 | __remove_wait_queue(&x->wait, &wait); | |
3650 | goto out; | |
3651 | } | |
3652 | __set_current_state(TASK_INTERRUPTIBLE); | |
3653 | spin_unlock_irq(&x->wait.lock); | |
3654 | timeout = schedule_timeout(timeout); | |
3655 | spin_lock_irq(&x->wait.lock); | |
3656 | if (!timeout) { | |
3657 | __remove_wait_queue(&x->wait, &wait); | |
3658 | goto out; | |
3659 | } | |
3660 | } while (!x->done); | |
3661 | __remove_wait_queue(&x->wait, &wait); | |
3662 | } | |
3663 | x->done--; | |
3664 | out: | |
3665 | spin_unlock_irq(&x->wait.lock); | |
3666 | return timeout; | |
3667 | } | |
3668 | EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); | |
3669 | ||
3670 | ||
3671 | #define SLEEP_ON_VAR \ | |
3672 | unsigned long flags; \ | |
3673 | wait_queue_t wait; \ | |
3674 | init_waitqueue_entry(&wait, current); | |
3675 | ||
3676 | #define SLEEP_ON_HEAD \ | |
3677 | spin_lock_irqsave(&q->lock,flags); \ | |
3678 | __add_wait_queue(q, &wait); \ | |
3679 | spin_unlock(&q->lock); | |
3680 | ||
3681 | #define SLEEP_ON_TAIL \ | |
3682 | spin_lock_irq(&q->lock); \ | |
3683 | __remove_wait_queue(q, &wait); \ | |
3684 | spin_unlock_irqrestore(&q->lock, flags); | |
3685 | ||
3686 | void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q) | |
3687 | { | |
3688 | SLEEP_ON_VAR | |
3689 | ||
3690 | current->state = TASK_INTERRUPTIBLE; | |
3691 | ||
3692 | SLEEP_ON_HEAD | |
3693 | schedule(); | |
3694 | SLEEP_ON_TAIL | |
3695 | } | |
3696 | ||
3697 | EXPORT_SYMBOL(interruptible_sleep_on); | |
3698 | ||
95cdf3b7 IM |
3699 | long fastcall __sched |
3700 | interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) | |
1da177e4 LT |
3701 | { |
3702 | SLEEP_ON_VAR | |
3703 | ||
3704 | current->state = TASK_INTERRUPTIBLE; | |
3705 | ||
3706 | SLEEP_ON_HEAD | |
3707 | timeout = schedule_timeout(timeout); | |
3708 | SLEEP_ON_TAIL | |
3709 | ||
3710 | return timeout; | |
3711 | } | |
3712 | ||
3713 | EXPORT_SYMBOL(interruptible_sleep_on_timeout); | |
3714 | ||
3715 | void fastcall __sched sleep_on(wait_queue_head_t *q) | |
3716 | { | |
3717 | SLEEP_ON_VAR | |
3718 | ||
3719 | current->state = TASK_UNINTERRUPTIBLE; | |
3720 | ||
3721 | SLEEP_ON_HEAD | |
3722 | schedule(); | |
3723 | SLEEP_ON_TAIL | |
3724 | } | |
3725 | ||
3726 | EXPORT_SYMBOL(sleep_on); | |
3727 | ||
3728 | long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) | |
3729 | { | |
3730 | SLEEP_ON_VAR | |
3731 | ||
3732 | current->state = TASK_UNINTERRUPTIBLE; | |
3733 | ||
3734 | SLEEP_ON_HEAD | |
3735 | timeout = schedule_timeout(timeout); | |
3736 | SLEEP_ON_TAIL | |
3737 | ||
3738 | return timeout; | |
3739 | } | |
3740 | ||
3741 | EXPORT_SYMBOL(sleep_on_timeout); | |
3742 | ||
b29739f9 IM |
3743 | #ifdef CONFIG_RT_MUTEXES |
3744 | ||
3745 | /* | |
3746 | * rt_mutex_setprio - set the current priority of a task | |
3747 | * @p: task | |
3748 | * @prio: prio value (kernel-internal form) | |
3749 | * | |
3750 | * This function changes the 'effective' priority of a task. It does | |
3751 | * not touch ->normal_prio like __setscheduler(). | |
3752 | * | |
3753 | * Used by the rt_mutex code to implement priority inheritance logic. | |
3754 | */ | |
3755 | void rt_mutex_setprio(task_t *p, int prio) | |
3756 | { | |
3757 | unsigned long flags; | |
3758 | prio_array_t *array; | |
3759 | runqueue_t *rq; | |
3760 | int oldprio; | |
3761 | ||
3762 | BUG_ON(prio < 0 || prio > MAX_PRIO); | |
3763 | ||
3764 | rq = task_rq_lock(p, &flags); | |
3765 | ||
3766 | oldprio = p->prio; | |
3767 | array = p->array; | |
3768 | if (array) | |
3769 | dequeue_task(p, array); | |
3770 | p->prio = prio; | |
3771 | ||
3772 | if (array) { | |
3773 | /* | |
3774 | * If changing to an RT priority then queue it | |
3775 | * in the active array! | |
3776 | */ | |
3777 | if (rt_task(p)) | |
3778 | array = rq->active; | |
3779 | enqueue_task(p, array); | |
3780 | /* | |
3781 | * Reschedule if we are currently running on this runqueue and | |
3782 | * our priority decreased, or if we are not currently running on | |
3783 | * this runqueue and our priority is higher than the current's | |
3784 | */ | |
3785 | if (task_running(rq, p)) { | |
3786 | if (p->prio > oldprio) | |
3787 | resched_task(rq->curr); | |
3788 | } else if (TASK_PREEMPTS_CURR(p, rq)) | |
3789 | resched_task(rq->curr); | |
3790 | } | |
3791 | task_rq_unlock(rq, &flags); | |
3792 | } | |
3793 | ||
3794 | #endif | |
3795 | ||
1da177e4 LT |
3796 | void set_user_nice(task_t *p, long nice) |
3797 | { | |
3798 | unsigned long flags; | |
3799 | prio_array_t *array; | |
3800 | runqueue_t *rq; | |
b29739f9 | 3801 | int old_prio, delta; |
1da177e4 LT |
3802 | |
3803 | if (TASK_NICE(p) == nice || nice < -20 || nice > 19) | |
3804 | return; | |
3805 | /* | |
3806 | * We have to be careful, if called from sys_setpriority(), | |
3807 | * the task might be in the middle of scheduling on another CPU. | |
3808 | */ | |
3809 | rq = task_rq_lock(p, &flags); | |
3810 | /* | |
3811 | * The RT priorities are set via sched_setscheduler(), but we still | |
3812 | * allow the 'normal' nice value to be set - but as expected | |
3813 | * it wont have any effect on scheduling until the task is | |
b0a9499c | 3814 | * not SCHED_NORMAL/SCHED_BATCH: |
1da177e4 | 3815 | */ |
b29739f9 | 3816 | if (has_rt_policy(p)) { |
1da177e4 LT |
3817 | p->static_prio = NICE_TO_PRIO(nice); |
3818 | goto out_unlock; | |
3819 | } | |
3820 | array = p->array; | |
2dd73a4f | 3821 | if (array) { |
1da177e4 | 3822 | dequeue_task(p, array); |
2dd73a4f PW |
3823 | dec_raw_weighted_load(rq, p); |
3824 | } | |
1da177e4 | 3825 | |
1da177e4 | 3826 | p->static_prio = NICE_TO_PRIO(nice); |
2dd73a4f | 3827 | set_load_weight(p); |
b29739f9 IM |
3828 | old_prio = p->prio; |
3829 | p->prio = effective_prio(p); | |
3830 | delta = p->prio - old_prio; | |
1da177e4 LT |
3831 | |
3832 | if (array) { | |
3833 | enqueue_task(p, array); | |
2dd73a4f | 3834 | inc_raw_weighted_load(rq, p); |
1da177e4 LT |
3835 | /* |
3836 | * If the task increased its priority or is running and | |
3837 | * lowered its priority, then reschedule its CPU: | |
3838 | */ | |
3839 | if (delta < 0 || (delta > 0 && task_running(rq, p))) | |
3840 | resched_task(rq->curr); | |
3841 | } | |
3842 | out_unlock: | |
3843 | task_rq_unlock(rq, &flags); | |
3844 | } | |
1da177e4 LT |
3845 | EXPORT_SYMBOL(set_user_nice); |
3846 | ||
e43379f1 MM |
3847 | /* |
3848 | * can_nice - check if a task can reduce its nice value | |
3849 | * @p: task | |
3850 | * @nice: nice value | |
3851 | */ | |
3852 | int can_nice(const task_t *p, const int nice) | |
3853 | { | |
024f4747 MM |
3854 | /* convert nice value [19,-20] to rlimit style value [1,40] */ |
3855 | int nice_rlim = 20 - nice; | |
e43379f1 MM |
3856 | return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur || |
3857 | capable(CAP_SYS_NICE)); | |
3858 | } | |
3859 | ||
1da177e4 LT |
3860 | #ifdef __ARCH_WANT_SYS_NICE |
3861 | ||
3862 | /* | |
3863 | * sys_nice - change the priority of the current process. | |
3864 | * @increment: priority increment | |
3865 | * | |
3866 | * sys_setpriority is a more generic, but much slower function that | |
3867 | * does similar things. | |
3868 | */ | |
3869 | asmlinkage long sys_nice(int increment) | |
3870 | { | |
3871 | int retval; | |
3872 | long nice; | |
3873 | ||
3874 | /* | |
3875 | * Setpriority might change our priority at the same moment. | |
3876 | * We don't have to worry. Conceptually one call occurs first | |
3877 | * and we have a single winner. | |
3878 | */ | |
e43379f1 MM |
3879 | if (increment < -40) |
3880 | increment = -40; | |
1da177e4 LT |
3881 | if (increment > 40) |
3882 | increment = 40; | |
3883 | ||
3884 | nice = PRIO_TO_NICE(current->static_prio) + increment; | |
3885 | if (nice < -20) | |
3886 | nice = -20; | |
3887 | if (nice > 19) | |
3888 | nice = 19; | |
3889 | ||
e43379f1 MM |
3890 | if (increment < 0 && !can_nice(current, nice)) |
3891 | return -EPERM; | |
3892 | ||
1da177e4 LT |
3893 | retval = security_task_setnice(current, nice); |
3894 | if (retval) | |
3895 | return retval; | |
3896 | ||
3897 | set_user_nice(current, nice); | |
3898 | return 0; | |
3899 | } | |
3900 | ||
3901 | #endif | |
3902 | ||
3903 | /** | |
3904 | * task_prio - return the priority value of a given task. | |
3905 | * @p: the task in question. | |
3906 | * | |
3907 | * This is the priority value as seen by users in /proc. | |
3908 | * RT tasks are offset by -200. Normal tasks are centered | |
3909 | * around 0, value goes from -16 to +15. | |
3910 | */ | |
3911 | int task_prio(const task_t *p) | |
3912 | { | |
3913 | return p->prio - MAX_RT_PRIO; | |
3914 | } | |
3915 | ||
3916 | /** | |
3917 | * task_nice - return the nice value of a given task. | |
3918 | * @p: the task in question. | |
3919 | */ | |
3920 | int task_nice(const task_t *p) | |
3921 | { | |
3922 | return TASK_NICE(p); | |
3923 | } | |
1da177e4 | 3924 | EXPORT_SYMBOL_GPL(task_nice); |
1da177e4 LT |
3925 | |
3926 | /** | |
3927 | * idle_cpu - is a given cpu idle currently? | |
3928 | * @cpu: the processor in question. | |
3929 | */ | |
3930 | int idle_cpu(int cpu) | |
3931 | { | |
3932 | return cpu_curr(cpu) == cpu_rq(cpu)->idle; | |
3933 | } | |
3934 | ||
1da177e4 LT |
3935 | /** |
3936 | * idle_task - return the idle task for a given cpu. | |
3937 | * @cpu: the processor in question. | |
3938 | */ | |
3939 | task_t *idle_task(int cpu) | |
3940 | { | |
3941 | return cpu_rq(cpu)->idle; | |
3942 | } | |
3943 | ||
3944 | /** | |
3945 | * find_process_by_pid - find a process with a matching PID value. | |
3946 | * @pid: the pid in question. | |
3947 | */ | |
3948 | static inline task_t *find_process_by_pid(pid_t pid) | |
3949 | { | |
3950 | return pid ? find_task_by_pid(pid) : current; | |
3951 | } | |
3952 | ||
3953 | /* Actually do priority change: must hold rq lock. */ | |
3954 | static void __setscheduler(struct task_struct *p, int policy, int prio) | |
3955 | { | |
3956 | BUG_ON(p->array); | |
3957 | p->policy = policy; | |
3958 | p->rt_priority = prio; | |
b29739f9 IM |
3959 | p->normal_prio = normal_prio(p); |
3960 | /* we are holding p->pi_lock already */ | |
3961 | p->prio = rt_mutex_getprio(p); | |
3962 | /* | |
3963 | * SCHED_BATCH tasks are treated as perpetual CPU hogs: | |
3964 | */ | |
3965 | if (policy == SCHED_BATCH) | |
3966 | p->sleep_avg = 0; | |
2dd73a4f | 3967 | set_load_weight(p); |
1da177e4 LT |
3968 | } |
3969 | ||
3970 | /** | |
3971 | * sched_setscheduler - change the scheduling policy and/or RT priority of | |
3972 | * a thread. | |
3973 | * @p: the task in question. | |
3974 | * @policy: new policy. | |
3975 | * @param: structure containing the new RT priority. | |
3976 | */ | |
95cdf3b7 IM |
3977 | int sched_setscheduler(struct task_struct *p, int policy, |
3978 | struct sched_param *param) | |
1da177e4 LT |
3979 | { |
3980 | int retval; | |
3981 | int oldprio, oldpolicy = -1; | |
3982 | prio_array_t *array; | |
3983 | unsigned long flags; | |
3984 | runqueue_t *rq; | |
3985 | ||
66e5393a SR |
3986 | /* may grab non-irq protected spin_locks */ |
3987 | BUG_ON(in_interrupt()); | |
1da177e4 LT |
3988 | recheck: |
3989 | /* double check policy once rq lock held */ | |
3990 | if (policy < 0) | |
3991 | policy = oldpolicy = p->policy; | |
3992 | else if (policy != SCHED_FIFO && policy != SCHED_RR && | |
b0a9499c IM |
3993 | policy != SCHED_NORMAL && policy != SCHED_BATCH) |
3994 | return -EINVAL; | |
1da177e4 LT |
3995 | /* |
3996 | * Valid priorities for SCHED_FIFO and SCHED_RR are | |
b0a9499c IM |
3997 | * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and |
3998 | * SCHED_BATCH is 0. | |
1da177e4 LT |
3999 | */ |
4000 | if (param->sched_priority < 0 || | |
95cdf3b7 | 4001 | (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || |
d46523ea | 4002 | (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) |
1da177e4 | 4003 | return -EINVAL; |
b0a9499c IM |
4004 | if ((policy == SCHED_NORMAL || policy == SCHED_BATCH) |
4005 | != (param->sched_priority == 0)) | |
1da177e4 LT |
4006 | return -EINVAL; |
4007 | ||
37e4ab3f OC |
4008 | /* |
4009 | * Allow unprivileged RT tasks to decrease priority: | |
4010 | */ | |
4011 | if (!capable(CAP_SYS_NICE)) { | |
b0a9499c IM |
4012 | /* |
4013 | * can't change policy, except between SCHED_NORMAL | |
4014 | * and SCHED_BATCH: | |
4015 | */ | |
4016 | if (((policy != SCHED_NORMAL && p->policy != SCHED_BATCH) && | |
4017 | (policy != SCHED_BATCH && p->policy != SCHED_NORMAL)) && | |
4018 | !p->signal->rlim[RLIMIT_RTPRIO].rlim_cur) | |
37e4ab3f OC |
4019 | return -EPERM; |
4020 | /* can't increase priority */ | |
b0a9499c | 4021 | if ((policy != SCHED_NORMAL && policy != SCHED_BATCH) && |
37e4ab3f OC |
4022 | param->sched_priority > p->rt_priority && |
4023 | param->sched_priority > | |
4024 | p->signal->rlim[RLIMIT_RTPRIO].rlim_cur) | |
4025 | return -EPERM; | |
4026 | /* can't change other user's priorities */ | |
4027 | if ((current->euid != p->euid) && | |
4028 | (current->euid != p->uid)) | |
4029 | return -EPERM; | |
4030 | } | |
1da177e4 LT |
4031 | |
4032 | retval = security_task_setscheduler(p, policy, param); | |
4033 | if (retval) | |
4034 | return retval; | |
b29739f9 IM |
4035 | /* |
4036 | * make sure no PI-waiters arrive (or leave) while we are | |
4037 | * changing the priority of the task: | |
4038 | */ | |
4039 | spin_lock_irqsave(&p->pi_lock, flags); | |
1da177e4 LT |
4040 | /* |
4041 | * To be able to change p->policy safely, the apropriate | |
4042 | * runqueue lock must be held. | |
4043 | */ | |
b29739f9 | 4044 | rq = __task_rq_lock(p); |
1da177e4 LT |
4045 | /* recheck policy now with rq lock held */ |
4046 | if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { | |
4047 | policy = oldpolicy = -1; | |
b29739f9 IM |
4048 | __task_rq_unlock(rq); |
4049 | spin_unlock_irqrestore(&p->pi_lock, flags); | |
1da177e4 LT |
4050 | goto recheck; |
4051 | } | |
4052 | array = p->array; | |
4053 | if (array) | |
4054 | deactivate_task(p, rq); | |
4055 | oldprio = p->prio; | |
4056 | __setscheduler(p, policy, param->sched_priority); | |
4057 | if (array) { | |
4058 | __activate_task(p, rq); | |
4059 | /* | |
4060 | * Reschedule if we are currently running on this runqueue and | |
4061 | * our priority decreased, or if we are not currently running on | |
4062 | * this runqueue and our priority is higher than the current's | |
4063 | */ | |
4064 | if (task_running(rq, p)) { | |
4065 | if (p->prio > oldprio) | |
4066 | resched_task(rq->curr); | |
4067 | } else if (TASK_PREEMPTS_CURR(p, rq)) | |
4068 | resched_task(rq->curr); | |
4069 | } | |
b29739f9 IM |
4070 | __task_rq_unlock(rq); |
4071 | spin_unlock_irqrestore(&p->pi_lock, flags); | |
4072 | ||
95e02ca9 TG |
4073 | rt_mutex_adjust_pi(p); |
4074 | ||
1da177e4 LT |
4075 | return 0; |
4076 | } | |
4077 | EXPORT_SYMBOL_GPL(sched_setscheduler); | |
4078 | ||
95cdf3b7 IM |
4079 | static int |
4080 | do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) | |
1da177e4 LT |
4081 | { |
4082 | int retval; | |
4083 | struct sched_param lparam; | |
4084 | struct task_struct *p; | |
4085 | ||
4086 | if (!param || pid < 0) | |
4087 | return -EINVAL; | |
4088 | if (copy_from_user(&lparam, param, sizeof(struct sched_param))) | |
4089 | return -EFAULT; | |
4090 | read_lock_irq(&tasklist_lock); | |
4091 | p = find_process_by_pid(pid); | |
4092 | if (!p) { | |
4093 | read_unlock_irq(&tasklist_lock); | |
4094 | return -ESRCH; | |
4095 | } | |
e74c69f4 | 4096 | get_task_struct(p); |
1da177e4 | 4097 | read_unlock_irq(&tasklist_lock); |
e74c69f4 TG |
4098 | retval = sched_setscheduler(p, policy, &lparam); |
4099 | put_task_struct(p); | |
1da177e4 LT |
4100 | return retval; |
4101 | } | |
4102 | ||
4103 | /** | |
4104 | * sys_sched_setscheduler - set/change the scheduler policy and RT priority | |
4105 | * @pid: the pid in question. | |
4106 | * @policy: new policy. | |
4107 | * @param: structure containing the new RT priority. | |
4108 | */ | |
4109 | asmlinkage long sys_sched_setscheduler(pid_t pid, int policy, | |
4110 | struct sched_param __user *param) | |
4111 | { | |
c21761f1 JB |
4112 | /* negative values for policy are not valid */ |
4113 | if (policy < 0) | |
4114 | return -EINVAL; | |
4115 | ||
1da177e4 LT |
4116 | return do_sched_setscheduler(pid, policy, param); |
4117 | } | |
4118 | ||
4119 | /** | |
4120 | * sys_sched_setparam - set/change the RT priority of a thread | |
4121 | * @pid: the pid in question. | |
4122 | * @param: structure containing the new RT priority. | |
4123 | */ | |
4124 | asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param) | |
4125 | { | |
4126 | return do_sched_setscheduler(pid, -1, param); | |
4127 | } | |
4128 | ||
4129 | /** | |
4130 | * sys_sched_getscheduler - get the policy (scheduling class) of a thread | |
4131 | * @pid: the pid in question. | |
4132 | */ | |
4133 | asmlinkage long sys_sched_getscheduler(pid_t pid) | |
4134 | { | |
4135 | int retval = -EINVAL; | |
4136 | task_t *p; | |
4137 | ||
4138 | if (pid < 0) | |
4139 | goto out_nounlock; | |
4140 | ||
4141 | retval = -ESRCH; | |
4142 | read_lock(&tasklist_lock); | |
4143 | p = find_process_by_pid(pid); | |
4144 | if (p) { | |
4145 | retval = security_task_getscheduler(p); | |
4146 | if (!retval) | |
4147 | retval = p->policy; | |
4148 | } | |
4149 | read_unlock(&tasklist_lock); | |
4150 | ||
4151 | out_nounlock: | |
4152 | return retval; | |
4153 | } | |
4154 | ||
4155 | /** | |
4156 | * sys_sched_getscheduler - get the RT priority of a thread | |
4157 | * @pid: the pid in question. | |
4158 | * @param: structure containing the RT priority. | |
4159 | */ | |
4160 | asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param) | |
4161 | { | |
4162 | struct sched_param lp; | |
4163 | int retval = -EINVAL; | |
4164 | task_t *p; | |
4165 | ||
4166 | if (!param || pid < 0) | |
4167 | goto out_nounlock; | |
4168 | ||
4169 | read_lock(&tasklist_lock); | |
4170 | p = find_process_by_pid(pid); | |
4171 | retval = -ESRCH; | |
4172 | if (!p) | |
4173 | goto out_unlock; | |
4174 | ||
4175 | retval = security_task_getscheduler(p); | |
4176 | if (retval) | |
4177 | goto out_unlock; | |
4178 | ||
4179 | lp.sched_priority = p->rt_priority; | |
4180 | read_unlock(&tasklist_lock); | |
4181 | ||
4182 | /* | |
4183 | * This one might sleep, we cannot do it with a spinlock held ... | |
4184 | */ | |
4185 | retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; | |
4186 | ||
4187 | out_nounlock: | |
4188 | return retval; | |
4189 | ||
4190 | out_unlock: | |
4191 | read_unlock(&tasklist_lock); | |
4192 | return retval; | |
4193 | } | |
4194 | ||
4195 | long sched_setaffinity(pid_t pid, cpumask_t new_mask) | |
4196 | { | |
4197 | task_t *p; | |
4198 | int retval; | |
4199 | cpumask_t cpus_allowed; | |
4200 | ||
4201 | lock_cpu_hotplug(); | |
4202 | read_lock(&tasklist_lock); | |
4203 | ||
4204 | p = find_process_by_pid(pid); | |
4205 | if (!p) { | |
4206 | read_unlock(&tasklist_lock); | |
4207 | unlock_cpu_hotplug(); | |
4208 | return -ESRCH; | |
4209 | } | |
4210 | ||
4211 | /* | |
4212 | * It is not safe to call set_cpus_allowed with the | |
4213 | * tasklist_lock held. We will bump the task_struct's | |
4214 | * usage count and then drop tasklist_lock. | |
4215 | */ | |
4216 | get_task_struct(p); | |
4217 | read_unlock(&tasklist_lock); | |
4218 | ||
4219 | retval = -EPERM; | |
4220 | if ((current->euid != p->euid) && (current->euid != p->uid) && | |
4221 | !capable(CAP_SYS_NICE)) | |
4222 | goto out_unlock; | |
4223 | ||
e7834f8f DQ |
4224 | retval = security_task_setscheduler(p, 0, NULL); |
4225 | if (retval) | |
4226 | goto out_unlock; | |
4227 | ||
1da177e4 LT |
4228 | cpus_allowed = cpuset_cpus_allowed(p); |
4229 | cpus_and(new_mask, new_mask, cpus_allowed); | |
4230 | retval = set_cpus_allowed(p, new_mask); | |
4231 | ||
4232 | out_unlock: | |
4233 | put_task_struct(p); | |
4234 | unlock_cpu_hotplug(); | |
4235 | return retval; | |
4236 | } | |
4237 | ||
4238 | static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, | |
4239 | cpumask_t *new_mask) | |
4240 | { | |
4241 | if (len < sizeof(cpumask_t)) { | |
4242 | memset(new_mask, 0, sizeof(cpumask_t)); | |
4243 | } else if (len > sizeof(cpumask_t)) { | |
4244 | len = sizeof(cpumask_t); | |
4245 | } | |
4246 | return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; | |
4247 | } | |
4248 | ||
4249 | /** | |
4250 | * sys_sched_setaffinity - set the cpu affinity of a process | |
4251 | * @pid: pid of the process | |
4252 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
4253 | * @user_mask_ptr: user-space pointer to the new cpu mask | |
4254 | */ | |
4255 | asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len, | |
4256 | unsigned long __user *user_mask_ptr) | |
4257 | { | |
4258 | cpumask_t new_mask; | |
4259 | int retval; | |
4260 | ||
4261 | retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask); | |
4262 | if (retval) | |
4263 | return retval; | |
4264 | ||
4265 | return sched_setaffinity(pid, new_mask); | |
4266 | } | |
4267 | ||
4268 | /* | |
4269 | * Represents all cpu's present in the system | |
4270 | * In systems capable of hotplug, this map could dynamically grow | |
4271 | * as new cpu's are detected in the system via any platform specific | |
4272 | * method, such as ACPI for e.g. | |
4273 | */ | |
4274 | ||
4cef0c61 | 4275 | cpumask_t cpu_present_map __read_mostly; |
1da177e4 LT |
4276 | EXPORT_SYMBOL(cpu_present_map); |
4277 | ||
4278 | #ifndef CONFIG_SMP | |
4cef0c61 AK |
4279 | cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL; |
4280 | cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL; | |
1da177e4 LT |
4281 | #endif |
4282 | ||
4283 | long sched_getaffinity(pid_t pid, cpumask_t *mask) | |
4284 | { | |
4285 | int retval; | |
4286 | task_t *p; | |
4287 | ||
4288 | lock_cpu_hotplug(); | |
4289 | read_lock(&tasklist_lock); | |
4290 | ||
4291 | retval = -ESRCH; | |
4292 | p = find_process_by_pid(pid); | |
4293 | if (!p) | |
4294 | goto out_unlock; | |
4295 | ||
e7834f8f DQ |
4296 | retval = security_task_getscheduler(p); |
4297 | if (retval) | |
4298 | goto out_unlock; | |
4299 | ||
2f7016d9 | 4300 | cpus_and(*mask, p->cpus_allowed, cpu_online_map); |
1da177e4 LT |
4301 | |
4302 | out_unlock: | |
4303 | read_unlock(&tasklist_lock); | |
4304 | unlock_cpu_hotplug(); | |
4305 | if (retval) | |
4306 | return retval; | |
4307 | ||
4308 | return 0; | |
4309 | } | |
4310 | ||
4311 | /** | |
4312 | * sys_sched_getaffinity - get the cpu affinity of a process | |
4313 | * @pid: pid of the process | |
4314 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
4315 | * @user_mask_ptr: user-space pointer to hold the current cpu mask | |
4316 | */ | |
4317 | asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len, | |
4318 | unsigned long __user *user_mask_ptr) | |
4319 | { | |
4320 | int ret; | |
4321 | cpumask_t mask; | |
4322 | ||
4323 | if (len < sizeof(cpumask_t)) | |
4324 | return -EINVAL; | |
4325 | ||
4326 | ret = sched_getaffinity(pid, &mask); | |
4327 | if (ret < 0) | |
4328 | return ret; | |
4329 | ||
4330 | if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t))) | |
4331 | return -EFAULT; | |
4332 | ||
4333 | return sizeof(cpumask_t); | |
4334 | } | |
4335 | ||
4336 | /** | |
4337 | * sys_sched_yield - yield the current processor to other threads. | |
4338 | * | |
4339 | * this function yields the current CPU by moving the calling thread | |
4340 | * to the expired array. If there are no other threads running on this | |
4341 | * CPU then this function will return. | |
4342 | */ | |
4343 | asmlinkage long sys_sched_yield(void) | |
4344 | { | |
4345 | runqueue_t *rq = this_rq_lock(); | |
4346 | prio_array_t *array = current->array; | |
4347 | prio_array_t *target = rq->expired; | |
4348 | ||
4349 | schedstat_inc(rq, yld_cnt); | |
4350 | /* | |
4351 | * We implement yielding by moving the task into the expired | |
4352 | * queue. | |
4353 | * | |
4354 | * (special rule: RT tasks will just roundrobin in the active | |
4355 | * array.) | |
4356 | */ | |
4357 | if (rt_task(current)) | |
4358 | target = rq->active; | |
4359 | ||
5927ad78 | 4360 | if (array->nr_active == 1) { |
1da177e4 LT |
4361 | schedstat_inc(rq, yld_act_empty); |
4362 | if (!rq->expired->nr_active) | |
4363 | schedstat_inc(rq, yld_both_empty); | |
4364 | } else if (!rq->expired->nr_active) | |
4365 | schedstat_inc(rq, yld_exp_empty); | |
4366 | ||
4367 | if (array != target) { | |
4368 | dequeue_task(current, array); | |
4369 | enqueue_task(current, target); | |
4370 | } else | |
4371 | /* | |
4372 | * requeue_task is cheaper so perform that if possible. | |
4373 | */ | |
4374 | requeue_task(current, array); | |
4375 | ||
4376 | /* | |
4377 | * Since we are going to call schedule() anyway, there's | |
4378 | * no need to preempt or enable interrupts: | |
4379 | */ | |
4380 | __release(rq->lock); | |
4381 | _raw_spin_unlock(&rq->lock); | |
4382 | preempt_enable_no_resched(); | |
4383 | ||
4384 | schedule(); | |
4385 | ||
4386 | return 0; | |
4387 | } | |
4388 | ||
4389 | static inline void __cond_resched(void) | |
4390 | { | |
8e0a43d8 IM |
4391 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP |
4392 | __might_sleep(__FILE__, __LINE__); | |
4393 | #endif | |
5bbcfd90 IM |
4394 | /* |
4395 | * The BKS might be reacquired before we have dropped | |
4396 | * PREEMPT_ACTIVE, which could trigger a second | |
4397 | * cond_resched() call. | |
4398 | */ | |
4399 | if (unlikely(preempt_count())) | |
4400 | return; | |
8ba7b0a1 LT |
4401 | if (unlikely(system_state != SYSTEM_RUNNING)) |
4402 | return; | |
1da177e4 LT |
4403 | do { |
4404 | add_preempt_count(PREEMPT_ACTIVE); | |
4405 | schedule(); | |
4406 | sub_preempt_count(PREEMPT_ACTIVE); | |
4407 | } while (need_resched()); | |
4408 | } | |
4409 | ||
4410 | int __sched cond_resched(void) | |
4411 | { | |
4412 | if (need_resched()) { | |
4413 | __cond_resched(); | |
4414 | return 1; | |
4415 | } | |
4416 | return 0; | |
4417 | } | |
4418 | ||
4419 | EXPORT_SYMBOL(cond_resched); | |
4420 | ||
4421 | /* | |
4422 | * cond_resched_lock() - if a reschedule is pending, drop the given lock, | |
4423 | * call schedule, and on return reacquire the lock. | |
4424 | * | |
4425 | * This works OK both with and without CONFIG_PREEMPT. We do strange low-level | |
4426 | * operations here to prevent schedule() from being called twice (once via | |
4427 | * spin_unlock(), once by hand). | |
4428 | */ | |
95cdf3b7 | 4429 | int cond_resched_lock(spinlock_t *lock) |
1da177e4 | 4430 | { |
6df3cecb JK |
4431 | int ret = 0; |
4432 | ||
1da177e4 LT |
4433 | if (need_lockbreak(lock)) { |
4434 | spin_unlock(lock); | |
4435 | cpu_relax(); | |
6df3cecb | 4436 | ret = 1; |
1da177e4 LT |
4437 | spin_lock(lock); |
4438 | } | |
4439 | if (need_resched()) { | |
4440 | _raw_spin_unlock(lock); | |
4441 | preempt_enable_no_resched(); | |
4442 | __cond_resched(); | |
6df3cecb | 4443 | ret = 1; |
1da177e4 | 4444 | spin_lock(lock); |
1da177e4 | 4445 | } |
6df3cecb | 4446 | return ret; |
1da177e4 LT |
4447 | } |
4448 | ||
4449 | EXPORT_SYMBOL(cond_resched_lock); | |
4450 | ||
4451 | int __sched cond_resched_softirq(void) | |
4452 | { | |
4453 | BUG_ON(!in_softirq()); | |
4454 | ||
4455 | if (need_resched()) { | |
4456 | __local_bh_enable(); | |
4457 | __cond_resched(); | |
4458 | local_bh_disable(); | |
4459 | return 1; | |
4460 | } | |
4461 | return 0; | |
4462 | } | |
4463 | ||
4464 | EXPORT_SYMBOL(cond_resched_softirq); | |
4465 | ||
4466 | ||
4467 | /** | |
4468 | * yield - yield the current processor to other threads. | |
4469 | * | |
4470 | * this is a shortcut for kernel-space yielding - it marks the | |
4471 | * thread runnable and calls sys_sched_yield(). | |
4472 | */ | |
4473 | void __sched yield(void) | |
4474 | { | |
4475 | set_current_state(TASK_RUNNING); | |
4476 | sys_sched_yield(); | |
4477 | } | |
4478 | ||
4479 | EXPORT_SYMBOL(yield); | |
4480 | ||
4481 | /* | |
4482 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so | |
4483 | * that process accounting knows that this is a task in IO wait state. | |
4484 | * | |
4485 | * But don't do that if it is a deliberate, throttling IO wait (this task | |
4486 | * has set its backing_dev_info: the queue against which it should throttle) | |
4487 | */ | |
4488 | void __sched io_schedule(void) | |
4489 | { | |
bfe5d834 | 4490 | struct runqueue *rq = &__raw_get_cpu_var(runqueues); |
1da177e4 LT |
4491 | |
4492 | atomic_inc(&rq->nr_iowait); | |
4493 | schedule(); | |
4494 | atomic_dec(&rq->nr_iowait); | |
4495 | } | |
4496 | ||
4497 | EXPORT_SYMBOL(io_schedule); | |
4498 | ||
4499 | long __sched io_schedule_timeout(long timeout) | |
4500 | { | |
bfe5d834 | 4501 | struct runqueue *rq = &__raw_get_cpu_var(runqueues); |
1da177e4 LT |
4502 | long ret; |
4503 | ||
4504 | atomic_inc(&rq->nr_iowait); | |
4505 | ret = schedule_timeout(timeout); | |
4506 | atomic_dec(&rq->nr_iowait); | |
4507 | return ret; | |
4508 | } | |
4509 | ||
4510 | /** | |
4511 | * sys_sched_get_priority_max - return maximum RT priority. | |
4512 | * @policy: scheduling class. | |
4513 | * | |
4514 | * this syscall returns the maximum rt_priority that can be used | |
4515 | * by a given scheduling class. | |
4516 | */ | |
4517 | asmlinkage long sys_sched_get_priority_max(int policy) | |
4518 | { | |
4519 | int ret = -EINVAL; | |
4520 | ||
4521 | switch (policy) { | |
4522 | case SCHED_FIFO: | |
4523 | case SCHED_RR: | |
4524 | ret = MAX_USER_RT_PRIO-1; | |
4525 | break; | |
4526 | case SCHED_NORMAL: | |
b0a9499c | 4527 | case SCHED_BATCH: |
1da177e4 LT |
4528 | ret = 0; |
4529 | break; | |
4530 | } | |
4531 | return ret; | |
4532 | } | |
4533 | ||
4534 | /** | |
4535 | * sys_sched_get_priority_min - return minimum RT priority. | |
4536 | * @policy: scheduling class. | |
4537 | * | |
4538 | * this syscall returns the minimum rt_priority that can be used | |
4539 | * by a given scheduling class. | |
4540 | */ | |
4541 | asmlinkage long sys_sched_get_priority_min(int policy) | |
4542 | { | |
4543 | int ret = -EINVAL; | |
4544 | ||
4545 | switch (policy) { | |
4546 | case SCHED_FIFO: | |
4547 | case SCHED_RR: | |
4548 | ret = 1; | |
4549 | break; | |
4550 | case SCHED_NORMAL: | |
b0a9499c | 4551 | case SCHED_BATCH: |
1da177e4 LT |
4552 | ret = 0; |
4553 | } | |
4554 | return ret; | |
4555 | } | |
4556 | ||
4557 | /** | |
4558 | * sys_sched_rr_get_interval - return the default timeslice of a process. | |
4559 | * @pid: pid of the process. | |
4560 | * @interval: userspace pointer to the timeslice value. | |
4561 | * | |
4562 | * this syscall writes the default timeslice value of a given process | |
4563 | * into the user-space timespec buffer. A value of '0' means infinity. | |
4564 | */ | |
4565 | asmlinkage | |
4566 | long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval) | |
4567 | { | |
4568 | int retval = -EINVAL; | |
4569 | struct timespec t; | |
4570 | task_t *p; | |
4571 | ||
4572 | if (pid < 0) | |
4573 | goto out_nounlock; | |
4574 | ||
4575 | retval = -ESRCH; | |
4576 | read_lock(&tasklist_lock); | |
4577 | p = find_process_by_pid(pid); | |
4578 | if (!p) | |
4579 | goto out_unlock; | |
4580 | ||
4581 | retval = security_task_getscheduler(p); | |
4582 | if (retval) | |
4583 | goto out_unlock; | |
4584 | ||
b78709cf | 4585 | jiffies_to_timespec(p->policy == SCHED_FIFO ? |
1da177e4 LT |
4586 | 0 : task_timeslice(p), &t); |
4587 | read_unlock(&tasklist_lock); | |
4588 | retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; | |
4589 | out_nounlock: | |
4590 | return retval; | |
4591 | out_unlock: | |
4592 | read_unlock(&tasklist_lock); | |
4593 | return retval; | |
4594 | } | |
4595 | ||
4596 | static inline struct task_struct *eldest_child(struct task_struct *p) | |
4597 | { | |
4598 | if (list_empty(&p->children)) return NULL; | |
4599 | return list_entry(p->children.next,struct task_struct,sibling); | |
4600 | } | |
4601 | ||
4602 | static inline struct task_struct *older_sibling(struct task_struct *p) | |
4603 | { | |
4604 | if (p->sibling.prev==&p->parent->children) return NULL; | |
4605 | return list_entry(p->sibling.prev,struct task_struct,sibling); | |
4606 | } | |
4607 | ||
4608 | static inline struct task_struct *younger_sibling(struct task_struct *p) | |
4609 | { | |
4610 | if (p->sibling.next==&p->parent->children) return NULL; | |
4611 | return list_entry(p->sibling.next,struct task_struct,sibling); | |
4612 | } | |
4613 | ||
95cdf3b7 | 4614 | static void show_task(task_t *p) |
1da177e4 LT |
4615 | { |
4616 | task_t *relative; | |
4617 | unsigned state; | |
4618 | unsigned long free = 0; | |
4619 | static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" }; | |
4620 | ||
4621 | printk("%-13.13s ", p->comm); | |
4622 | state = p->state ? __ffs(p->state) + 1 : 0; | |
4623 | if (state < ARRAY_SIZE(stat_nam)) | |
4624 | printk(stat_nam[state]); | |
4625 | else | |
4626 | printk("?"); | |
4627 | #if (BITS_PER_LONG == 32) | |
4628 | if (state == TASK_RUNNING) | |
4629 | printk(" running "); | |
4630 | else | |
4631 | printk(" %08lX ", thread_saved_pc(p)); | |
4632 | #else | |
4633 | if (state == TASK_RUNNING) | |
4634 | printk(" running task "); | |
4635 | else | |
4636 | printk(" %016lx ", thread_saved_pc(p)); | |
4637 | #endif | |
4638 | #ifdef CONFIG_DEBUG_STACK_USAGE | |
4639 | { | |
10ebffde | 4640 | unsigned long *n = end_of_stack(p); |
1da177e4 LT |
4641 | while (!*n) |
4642 | n++; | |
10ebffde | 4643 | free = (unsigned long)n - (unsigned long)end_of_stack(p); |
1da177e4 LT |
4644 | } |
4645 | #endif | |
4646 | printk("%5lu %5d %6d ", free, p->pid, p->parent->pid); | |
4647 | if ((relative = eldest_child(p))) | |
4648 | printk("%5d ", relative->pid); | |
4649 | else | |
4650 | printk(" "); | |
4651 | if ((relative = younger_sibling(p))) | |
4652 | printk("%7d", relative->pid); | |
4653 | else | |
4654 | printk(" "); | |
4655 | if ((relative = older_sibling(p))) | |
4656 | printk(" %5d", relative->pid); | |
4657 | else | |
4658 | printk(" "); | |
4659 | if (!p->mm) | |
4660 | printk(" (L-TLB)\n"); | |
4661 | else | |
4662 | printk(" (NOTLB)\n"); | |
4663 | ||
4664 | if (state != TASK_RUNNING) | |
4665 | show_stack(p, NULL); | |
4666 | } | |
4667 | ||
4668 | void show_state(void) | |
4669 | { | |
4670 | task_t *g, *p; | |
4671 | ||
4672 | #if (BITS_PER_LONG == 32) | |
4673 | printk("\n" | |
4674 | " sibling\n"); | |
4675 | printk(" task PC pid father child younger older\n"); | |
4676 | #else | |
4677 | printk("\n" | |
4678 | " sibling\n"); | |
4679 | printk(" task PC pid father child younger older\n"); | |
4680 | #endif | |
4681 | read_lock(&tasklist_lock); | |
4682 | do_each_thread(g, p) { | |
4683 | /* | |
4684 | * reset the NMI-timeout, listing all files on a slow | |
4685 | * console might take alot of time: | |
4686 | */ | |
4687 | touch_nmi_watchdog(); | |
4688 | show_task(p); | |
4689 | } while_each_thread(g, p); | |
4690 | ||
4691 | read_unlock(&tasklist_lock); | |
de5097c2 | 4692 | mutex_debug_show_all_locks(); |
1da177e4 LT |
4693 | } |
4694 | ||
f340c0d1 IM |
4695 | /** |
4696 | * init_idle - set up an idle thread for a given CPU | |
4697 | * @idle: task in question | |
4698 | * @cpu: cpu the idle task belongs to | |
4699 | * | |
4700 | * NOTE: this function does not set the idle thread's NEED_RESCHED | |
4701 | * flag, to make booting more robust. | |
4702 | */ | |
1da177e4 LT |
4703 | void __devinit init_idle(task_t *idle, int cpu) |
4704 | { | |
4705 | runqueue_t *rq = cpu_rq(cpu); | |
4706 | unsigned long flags; | |
4707 | ||
81c29a85 | 4708 | idle->timestamp = sched_clock(); |
1da177e4 LT |
4709 | idle->sleep_avg = 0; |
4710 | idle->array = NULL; | |
b29739f9 | 4711 | idle->prio = idle->normal_prio = MAX_PRIO; |
1da177e4 LT |
4712 | idle->state = TASK_RUNNING; |
4713 | idle->cpus_allowed = cpumask_of_cpu(cpu); | |
4714 | set_task_cpu(idle, cpu); | |
4715 | ||
4716 | spin_lock_irqsave(&rq->lock, flags); | |
4717 | rq->curr = rq->idle = idle; | |
4866cde0 NP |
4718 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
4719 | idle->oncpu = 1; | |
4720 | #endif | |
1da177e4 LT |
4721 | spin_unlock_irqrestore(&rq->lock, flags); |
4722 | ||
4723 | /* Set the preempt count _outside_ the spinlocks! */ | |
4724 | #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL) | |
a1261f54 | 4725 | task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0); |
1da177e4 | 4726 | #else |
a1261f54 | 4727 | task_thread_info(idle)->preempt_count = 0; |
1da177e4 LT |
4728 | #endif |
4729 | } | |
4730 | ||
4731 | /* | |
4732 | * In a system that switches off the HZ timer nohz_cpu_mask | |
4733 | * indicates which cpus entered this state. This is used | |
4734 | * in the rcu update to wait only for active cpus. For system | |
4735 | * which do not switch off the HZ timer nohz_cpu_mask should | |
4736 | * always be CPU_MASK_NONE. | |
4737 | */ | |
4738 | cpumask_t nohz_cpu_mask = CPU_MASK_NONE; | |
4739 | ||
4740 | #ifdef CONFIG_SMP | |
4741 | /* | |
4742 | * This is how migration works: | |
4743 | * | |
4744 | * 1) we queue a migration_req_t structure in the source CPU's | |
4745 | * runqueue and wake up that CPU's migration thread. | |
4746 | * 2) we down() the locked semaphore => thread blocks. | |
4747 | * 3) migration thread wakes up (implicitly it forces the migrated | |
4748 | * thread off the CPU) | |
4749 | * 4) it gets the migration request and checks whether the migrated | |
4750 | * task is still in the wrong runqueue. | |
4751 | * 5) if it's in the wrong runqueue then the migration thread removes | |
4752 | * it and puts it into the right queue. | |
4753 | * 6) migration thread up()s the semaphore. | |
4754 | * 7) we wake up and the migration is done. | |
4755 | */ | |
4756 | ||
4757 | /* | |
4758 | * Change a given task's CPU affinity. Migrate the thread to a | |
4759 | * proper CPU and schedule it away if the CPU it's executing on | |
4760 | * is removed from the allowed bitmask. | |
4761 | * | |
4762 | * NOTE: the caller must have a valid reference to the task, the | |
4763 | * task must not exit() & deallocate itself prematurely. The | |
4764 | * call is not atomic; no spinlocks may be held. | |
4765 | */ | |
4766 | int set_cpus_allowed(task_t *p, cpumask_t new_mask) | |
4767 | { | |
4768 | unsigned long flags; | |
4769 | int ret = 0; | |
4770 | migration_req_t req; | |
4771 | runqueue_t *rq; | |
4772 | ||
4773 | rq = task_rq_lock(p, &flags); | |
4774 | if (!cpus_intersects(new_mask, cpu_online_map)) { | |
4775 | ret = -EINVAL; | |
4776 | goto out; | |
4777 | } | |
4778 | ||
4779 | p->cpus_allowed = new_mask; | |
4780 | /* Can the task run on the task's current CPU? If so, we're done */ | |
4781 | if (cpu_isset(task_cpu(p), new_mask)) | |
4782 | goto out; | |
4783 | ||
4784 | if (migrate_task(p, any_online_cpu(new_mask), &req)) { | |
4785 | /* Need help from migration thread: drop lock and wait. */ | |
4786 | task_rq_unlock(rq, &flags); | |
4787 | wake_up_process(rq->migration_thread); | |
4788 | wait_for_completion(&req.done); | |
4789 | tlb_migrate_finish(p->mm); | |
4790 | return 0; | |
4791 | } | |
4792 | out: | |
4793 | task_rq_unlock(rq, &flags); | |
4794 | return ret; | |
4795 | } | |
4796 | ||
4797 | EXPORT_SYMBOL_GPL(set_cpus_allowed); | |
4798 | ||
4799 | /* | |
4800 | * Move (not current) task off this cpu, onto dest cpu. We're doing | |
4801 | * this because either it can't run here any more (set_cpus_allowed() | |
4802 | * away from this CPU, or CPU going down), or because we're | |
4803 | * attempting to rebalance this task on exec (sched_exec). | |
4804 | * | |
4805 | * So we race with normal scheduler movements, but that's OK, as long | |
4806 | * as the task is no longer on this CPU. | |
efc30814 KK |
4807 | * |
4808 | * Returns non-zero if task was successfully migrated. | |
1da177e4 | 4809 | */ |
efc30814 | 4810 | static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) |
1da177e4 LT |
4811 | { |
4812 | runqueue_t *rq_dest, *rq_src; | |
efc30814 | 4813 | int ret = 0; |
1da177e4 LT |
4814 | |
4815 | if (unlikely(cpu_is_offline(dest_cpu))) | |
efc30814 | 4816 | return ret; |
1da177e4 LT |
4817 | |
4818 | rq_src = cpu_rq(src_cpu); | |
4819 | rq_dest = cpu_rq(dest_cpu); | |
4820 | ||
4821 | double_rq_lock(rq_src, rq_dest); | |
4822 | /* Already moved. */ | |
4823 | if (task_cpu(p) != src_cpu) | |
4824 | goto out; | |
4825 | /* Affinity changed (again). */ | |
4826 | if (!cpu_isset(dest_cpu, p->cpus_allowed)) | |
4827 | goto out; | |
4828 | ||
4829 | set_task_cpu(p, dest_cpu); | |
4830 | if (p->array) { | |
4831 | /* | |
4832 | * Sync timestamp with rq_dest's before activating. | |
4833 | * The same thing could be achieved by doing this step | |
4834 | * afterwards, and pretending it was a local activate. | |
4835 | * This way is cleaner and logically correct. | |
4836 | */ | |
4837 | p->timestamp = p->timestamp - rq_src->timestamp_last_tick | |
4838 | + rq_dest->timestamp_last_tick; | |
4839 | deactivate_task(p, rq_src); | |
4840 | activate_task(p, rq_dest, 0); | |
4841 | if (TASK_PREEMPTS_CURR(p, rq_dest)) | |
4842 | resched_task(rq_dest->curr); | |
4843 | } | |
efc30814 | 4844 | ret = 1; |
1da177e4 LT |
4845 | out: |
4846 | double_rq_unlock(rq_src, rq_dest); | |
efc30814 | 4847 | return ret; |
1da177e4 LT |
4848 | } |
4849 | ||
4850 | /* | |
4851 | * migration_thread - this is a highprio system thread that performs | |
4852 | * thread migration by bumping thread off CPU then 'pushing' onto | |
4853 | * another runqueue. | |
4854 | */ | |
95cdf3b7 | 4855 | static int migration_thread(void *data) |
1da177e4 LT |
4856 | { |
4857 | runqueue_t *rq; | |
4858 | int cpu = (long)data; | |
4859 | ||
4860 | rq = cpu_rq(cpu); | |
4861 | BUG_ON(rq->migration_thread != current); | |
4862 | ||
4863 | set_current_state(TASK_INTERRUPTIBLE); | |
4864 | while (!kthread_should_stop()) { | |
4865 | struct list_head *head; | |
4866 | migration_req_t *req; | |
4867 | ||
3e1d1d28 | 4868 | try_to_freeze(); |
1da177e4 LT |
4869 | |
4870 | spin_lock_irq(&rq->lock); | |
4871 | ||
4872 | if (cpu_is_offline(cpu)) { | |
4873 | spin_unlock_irq(&rq->lock); | |
4874 | goto wait_to_die; | |
4875 | } | |
4876 | ||
4877 | if (rq->active_balance) { | |
4878 | active_load_balance(rq, cpu); | |
4879 | rq->active_balance = 0; | |
4880 | } | |
4881 | ||
4882 | head = &rq->migration_queue; | |
4883 | ||
4884 | if (list_empty(head)) { | |
4885 | spin_unlock_irq(&rq->lock); | |
4886 | schedule(); | |
4887 | set_current_state(TASK_INTERRUPTIBLE); | |
4888 | continue; | |
4889 | } | |
4890 | req = list_entry(head->next, migration_req_t, list); | |
4891 | list_del_init(head->next); | |
4892 | ||
674311d5 NP |
4893 | spin_unlock(&rq->lock); |
4894 | __migrate_task(req->task, cpu, req->dest_cpu); | |
4895 | local_irq_enable(); | |
1da177e4 LT |
4896 | |
4897 | complete(&req->done); | |
4898 | } | |
4899 | __set_current_state(TASK_RUNNING); | |
4900 | return 0; | |
4901 | ||
4902 | wait_to_die: | |
4903 | /* Wait for kthread_stop */ | |
4904 | set_current_state(TASK_INTERRUPTIBLE); | |
4905 | while (!kthread_should_stop()) { | |
4906 | schedule(); | |
4907 | set_current_state(TASK_INTERRUPTIBLE); | |
4908 | } | |
4909 | __set_current_state(TASK_RUNNING); | |
4910 | return 0; | |
4911 | } | |
4912 | ||
4913 | #ifdef CONFIG_HOTPLUG_CPU | |
4914 | /* Figure out where task on dead CPU should go, use force if neccessary. */ | |
4915 | static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk) | |
4916 | { | |
efc30814 KK |
4917 | runqueue_t *rq; |
4918 | unsigned long flags; | |
1da177e4 LT |
4919 | int dest_cpu; |
4920 | cpumask_t mask; | |
4921 | ||
efc30814 | 4922 | restart: |
1da177e4 LT |
4923 | /* On same node? */ |
4924 | mask = node_to_cpumask(cpu_to_node(dead_cpu)); | |
4925 | cpus_and(mask, mask, tsk->cpus_allowed); | |
4926 | dest_cpu = any_online_cpu(mask); | |
4927 | ||
4928 | /* On any allowed CPU? */ | |
4929 | if (dest_cpu == NR_CPUS) | |
4930 | dest_cpu = any_online_cpu(tsk->cpus_allowed); | |
4931 | ||
4932 | /* No more Mr. Nice Guy. */ | |
4933 | if (dest_cpu == NR_CPUS) { | |
efc30814 | 4934 | rq = task_rq_lock(tsk, &flags); |
b39c4fab | 4935 | cpus_setall(tsk->cpus_allowed); |
1da177e4 | 4936 | dest_cpu = any_online_cpu(tsk->cpus_allowed); |
efc30814 | 4937 | task_rq_unlock(rq, &flags); |
1da177e4 LT |
4938 | |
4939 | /* | |
4940 | * Don't tell them about moving exiting tasks or | |
4941 | * kernel threads (both mm NULL), since they never | |
4942 | * leave kernel. | |
4943 | */ | |
4944 | if (tsk->mm && printk_ratelimit()) | |
4945 | printk(KERN_INFO "process %d (%s) no " | |
4946 | "longer affine to cpu%d\n", | |
4947 | tsk->pid, tsk->comm, dead_cpu); | |
4948 | } | |
efc30814 KK |
4949 | if (!__migrate_task(tsk, dead_cpu, dest_cpu)) |
4950 | goto restart; | |
1da177e4 LT |
4951 | } |
4952 | ||
4953 | /* | |
4954 | * While a dead CPU has no uninterruptible tasks queued at this point, | |
4955 | * it might still have a nonzero ->nr_uninterruptible counter, because | |
4956 | * for performance reasons the counter is not stricly tracking tasks to | |
4957 | * their home CPUs. So we just add the counter to another CPU's counter, | |
4958 | * to keep the global sum constant after CPU-down: | |
4959 | */ | |
4960 | static void migrate_nr_uninterruptible(runqueue_t *rq_src) | |
4961 | { | |
4962 | runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL)); | |
4963 | unsigned long flags; | |
4964 | ||
4965 | local_irq_save(flags); | |
4966 | double_rq_lock(rq_src, rq_dest); | |
4967 | rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible; | |
4968 | rq_src->nr_uninterruptible = 0; | |
4969 | double_rq_unlock(rq_src, rq_dest); | |
4970 | local_irq_restore(flags); | |
4971 | } | |
4972 | ||
4973 | /* Run through task list and migrate tasks from the dead cpu. */ | |
4974 | static void migrate_live_tasks(int src_cpu) | |
4975 | { | |
4976 | struct task_struct *tsk, *t; | |
4977 | ||
4978 | write_lock_irq(&tasklist_lock); | |
4979 | ||
4980 | do_each_thread(t, tsk) { | |
4981 | if (tsk == current) | |
4982 | continue; | |
4983 | ||
4984 | if (task_cpu(tsk) == src_cpu) | |
4985 | move_task_off_dead_cpu(src_cpu, tsk); | |
4986 | } while_each_thread(t, tsk); | |
4987 | ||
4988 | write_unlock_irq(&tasklist_lock); | |
4989 | } | |
4990 | ||
4991 | /* Schedules idle task to be the next runnable task on current CPU. | |
4992 | * It does so by boosting its priority to highest possible and adding it to | |
4993 | * the _front_ of runqueue. Used by CPU offline code. | |
4994 | */ | |
4995 | void sched_idle_next(void) | |
4996 | { | |
4997 | int cpu = smp_processor_id(); | |
4998 | runqueue_t *rq = this_rq(); | |
4999 | struct task_struct *p = rq->idle; | |
5000 | unsigned long flags; | |
5001 | ||
5002 | /* cpu has to be offline */ | |
5003 | BUG_ON(cpu_online(cpu)); | |
5004 | ||
5005 | /* Strictly not necessary since rest of the CPUs are stopped by now | |
5006 | * and interrupts disabled on current cpu. | |
5007 | */ | |
5008 | spin_lock_irqsave(&rq->lock, flags); | |
5009 | ||
5010 | __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1); | |
5011 | /* Add idle task to _front_ of it's priority queue */ | |
5012 | __activate_idle_task(p, rq); | |
5013 | ||
5014 | spin_unlock_irqrestore(&rq->lock, flags); | |
5015 | } | |
5016 | ||
5017 | /* Ensures that the idle task is using init_mm right before its cpu goes | |
5018 | * offline. | |
5019 | */ | |
5020 | void idle_task_exit(void) | |
5021 | { | |
5022 | struct mm_struct *mm = current->active_mm; | |
5023 | ||
5024 | BUG_ON(cpu_online(smp_processor_id())); | |
5025 | ||
5026 | if (mm != &init_mm) | |
5027 | switch_mm(mm, &init_mm, current); | |
5028 | mmdrop(mm); | |
5029 | } | |
5030 | ||
5031 | static void migrate_dead(unsigned int dead_cpu, task_t *tsk) | |
5032 | { | |
5033 | struct runqueue *rq = cpu_rq(dead_cpu); | |
5034 | ||
5035 | /* Must be exiting, otherwise would be on tasklist. */ | |
5036 | BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD); | |
5037 | ||
5038 | /* Cannot have done final schedule yet: would have vanished. */ | |
5039 | BUG_ON(tsk->flags & PF_DEAD); | |
5040 | ||
5041 | get_task_struct(tsk); | |
5042 | ||
5043 | /* | |
5044 | * Drop lock around migration; if someone else moves it, | |
5045 | * that's OK. No task can be added to this CPU, so iteration is | |
5046 | * fine. | |
5047 | */ | |
5048 | spin_unlock_irq(&rq->lock); | |
5049 | move_task_off_dead_cpu(dead_cpu, tsk); | |
5050 | spin_lock_irq(&rq->lock); | |
5051 | ||
5052 | put_task_struct(tsk); | |
5053 | } | |
5054 | ||
5055 | /* release_task() removes task from tasklist, so we won't find dead tasks. */ | |
5056 | static void migrate_dead_tasks(unsigned int dead_cpu) | |
5057 | { | |
5058 | unsigned arr, i; | |
5059 | struct runqueue *rq = cpu_rq(dead_cpu); | |
5060 | ||
5061 | for (arr = 0; arr < 2; arr++) { | |
5062 | for (i = 0; i < MAX_PRIO; i++) { | |
5063 | struct list_head *list = &rq->arrays[arr].queue[i]; | |
5064 | while (!list_empty(list)) | |
5065 | migrate_dead(dead_cpu, | |
5066 | list_entry(list->next, task_t, | |
5067 | run_list)); | |
5068 | } | |
5069 | } | |
5070 | } | |
5071 | #endif /* CONFIG_HOTPLUG_CPU */ | |
5072 | ||
5073 | /* | |
5074 | * migration_call - callback that gets triggered when a CPU is added. | |
5075 | * Here we can start up the necessary migration thread for the new CPU. | |
5076 | */ | |
26c2143b CS |
5077 | static int __cpuinit migration_call(struct notifier_block *nfb, |
5078 | unsigned long action, | |
5079 | void *hcpu) | |
1da177e4 LT |
5080 | { |
5081 | int cpu = (long)hcpu; | |
5082 | struct task_struct *p; | |
5083 | struct runqueue *rq; | |
5084 | unsigned long flags; | |
5085 | ||
5086 | switch (action) { | |
5087 | case CPU_UP_PREPARE: | |
5088 | p = kthread_create(migration_thread, hcpu, "migration/%d",cpu); | |
5089 | if (IS_ERR(p)) | |
5090 | return NOTIFY_BAD; | |
5091 | p->flags |= PF_NOFREEZE; | |
5092 | kthread_bind(p, cpu); | |
5093 | /* Must be high prio: stop_machine expects to yield to it. */ | |
5094 | rq = task_rq_lock(p, &flags); | |
5095 | __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1); | |
5096 | task_rq_unlock(rq, &flags); | |
5097 | cpu_rq(cpu)->migration_thread = p; | |
5098 | break; | |
5099 | case CPU_ONLINE: | |
5100 | /* Strictly unneccessary, as first user will wake it. */ | |
5101 | wake_up_process(cpu_rq(cpu)->migration_thread); | |
5102 | break; | |
5103 | #ifdef CONFIG_HOTPLUG_CPU | |
5104 | case CPU_UP_CANCELED: | |
fc75cdfa HC |
5105 | if (!cpu_rq(cpu)->migration_thread) |
5106 | break; | |
1da177e4 | 5107 | /* Unbind it from offline cpu so it can run. Fall thru. */ |
a4c4af7c HC |
5108 | kthread_bind(cpu_rq(cpu)->migration_thread, |
5109 | any_online_cpu(cpu_online_map)); | |
1da177e4 LT |
5110 | kthread_stop(cpu_rq(cpu)->migration_thread); |
5111 | cpu_rq(cpu)->migration_thread = NULL; | |
5112 | break; | |
5113 | case CPU_DEAD: | |
5114 | migrate_live_tasks(cpu); | |
5115 | rq = cpu_rq(cpu); | |
5116 | kthread_stop(rq->migration_thread); | |
5117 | rq->migration_thread = NULL; | |
5118 | /* Idle task back to normal (off runqueue, low prio) */ | |
5119 | rq = task_rq_lock(rq->idle, &flags); | |
5120 | deactivate_task(rq->idle, rq); | |
5121 | rq->idle->static_prio = MAX_PRIO; | |
5122 | __setscheduler(rq->idle, SCHED_NORMAL, 0); | |
5123 | migrate_dead_tasks(cpu); | |
5124 | task_rq_unlock(rq, &flags); | |
5125 | migrate_nr_uninterruptible(rq); | |
5126 | BUG_ON(rq->nr_running != 0); | |
5127 | ||
5128 | /* No need to migrate the tasks: it was best-effort if | |
5129 | * they didn't do lock_cpu_hotplug(). Just wake up | |
5130 | * the requestors. */ | |
5131 | spin_lock_irq(&rq->lock); | |
5132 | while (!list_empty(&rq->migration_queue)) { | |
5133 | migration_req_t *req; | |
5134 | req = list_entry(rq->migration_queue.next, | |
5135 | migration_req_t, list); | |
1da177e4 LT |
5136 | list_del_init(&req->list); |
5137 | complete(&req->done); | |
5138 | } | |
5139 | spin_unlock_irq(&rq->lock); | |
5140 | break; | |
5141 | #endif | |
5142 | } | |
5143 | return NOTIFY_OK; | |
5144 | } | |
5145 | ||
5146 | /* Register at highest priority so that task migration (migrate_all_tasks) | |
5147 | * happens before everything else. | |
5148 | */ | |
26c2143b | 5149 | static struct notifier_block __cpuinitdata migration_notifier = { |
1da177e4 LT |
5150 | .notifier_call = migration_call, |
5151 | .priority = 10 | |
5152 | }; | |
5153 | ||
5154 | int __init migration_init(void) | |
5155 | { | |
5156 | void *cpu = (void *)(long)smp_processor_id(); | |
5157 | /* Start one for boot CPU. */ | |
5158 | migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); | |
5159 | migration_call(&migration_notifier, CPU_ONLINE, cpu); | |
5160 | register_cpu_notifier(&migration_notifier); | |
5161 | return 0; | |
5162 | } | |
5163 | #endif | |
5164 | ||
5165 | #ifdef CONFIG_SMP | |
1a20ff27 | 5166 | #undef SCHED_DOMAIN_DEBUG |
1da177e4 LT |
5167 | #ifdef SCHED_DOMAIN_DEBUG |
5168 | static void sched_domain_debug(struct sched_domain *sd, int cpu) | |
5169 | { | |
5170 | int level = 0; | |
5171 | ||
41c7ce9a NP |
5172 | if (!sd) { |
5173 | printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); | |
5174 | return; | |
5175 | } | |
5176 | ||
1da177e4 LT |
5177 | printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); |
5178 | ||
5179 | do { | |
5180 | int i; | |
5181 | char str[NR_CPUS]; | |
5182 | struct sched_group *group = sd->groups; | |
5183 | cpumask_t groupmask; | |
5184 | ||
5185 | cpumask_scnprintf(str, NR_CPUS, sd->span); | |
5186 | cpus_clear(groupmask); | |
5187 | ||
5188 | printk(KERN_DEBUG); | |
5189 | for (i = 0; i < level + 1; i++) | |
5190 | printk(" "); | |
5191 | printk("domain %d: ", level); | |
5192 | ||
5193 | if (!(sd->flags & SD_LOAD_BALANCE)) { | |
5194 | printk("does not load-balance\n"); | |
5195 | if (sd->parent) | |
5196 | printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent"); | |
5197 | break; | |
5198 | } | |
5199 | ||
5200 | printk("span %s\n", str); | |
5201 | ||
5202 | if (!cpu_isset(cpu, sd->span)) | |
5203 | printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu); | |
5204 | if (!cpu_isset(cpu, group->cpumask)) | |
5205 | printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu); | |
5206 | ||
5207 | printk(KERN_DEBUG); | |
5208 | for (i = 0; i < level + 2; i++) | |
5209 | printk(" "); | |
5210 | printk("groups:"); | |
5211 | do { | |
5212 | if (!group) { | |
5213 | printk("\n"); | |
5214 | printk(KERN_ERR "ERROR: group is NULL\n"); | |
5215 | break; | |
5216 | } | |
5217 | ||
5218 | if (!group->cpu_power) { | |
5219 | printk("\n"); | |
5220 | printk(KERN_ERR "ERROR: domain->cpu_power not set\n"); | |
5221 | } | |
5222 | ||
5223 | if (!cpus_weight(group->cpumask)) { | |
5224 | printk("\n"); | |
5225 | printk(KERN_ERR "ERROR: empty group\n"); | |
5226 | } | |
5227 | ||
5228 | if (cpus_intersects(groupmask, group->cpumask)) { | |
5229 | printk("\n"); | |
5230 | printk(KERN_ERR "ERROR: repeated CPUs\n"); | |
5231 | } | |
5232 | ||
5233 | cpus_or(groupmask, groupmask, group->cpumask); | |
5234 | ||
5235 | cpumask_scnprintf(str, NR_CPUS, group->cpumask); | |
5236 | printk(" %s", str); | |
5237 | ||
5238 | group = group->next; | |
5239 | } while (group != sd->groups); | |
5240 | printk("\n"); | |
5241 | ||
5242 | if (!cpus_equal(sd->span, groupmask)) | |
5243 | printk(KERN_ERR "ERROR: groups don't span domain->span\n"); | |
5244 | ||
5245 | level++; | |
5246 | sd = sd->parent; | |
5247 | ||
5248 | if (sd) { | |
5249 | if (!cpus_subset(groupmask, sd->span)) | |
5250 | printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n"); | |
5251 | } | |
5252 | ||
5253 | } while (sd); | |
5254 | } | |
5255 | #else | |
5256 | #define sched_domain_debug(sd, cpu) {} | |
5257 | #endif | |
5258 | ||
1a20ff27 | 5259 | static int sd_degenerate(struct sched_domain *sd) |
245af2c7 SS |
5260 | { |
5261 | if (cpus_weight(sd->span) == 1) | |
5262 | return 1; | |
5263 | ||
5264 | /* Following flags need at least 2 groups */ | |
5265 | if (sd->flags & (SD_LOAD_BALANCE | | |
5266 | SD_BALANCE_NEWIDLE | | |
5267 | SD_BALANCE_FORK | | |
5268 | SD_BALANCE_EXEC)) { | |
5269 | if (sd->groups != sd->groups->next) | |
5270 | return 0; | |
5271 | } | |
5272 | ||
5273 | /* Following flags don't use groups */ | |
5274 | if (sd->flags & (SD_WAKE_IDLE | | |
5275 | SD_WAKE_AFFINE | | |
5276 | SD_WAKE_BALANCE)) | |
5277 | return 0; | |
5278 | ||
5279 | return 1; | |
5280 | } | |
5281 | ||
1a20ff27 | 5282 | static int sd_parent_degenerate(struct sched_domain *sd, |
245af2c7 SS |
5283 | struct sched_domain *parent) |
5284 | { | |
5285 | unsigned long cflags = sd->flags, pflags = parent->flags; | |
5286 | ||
5287 | if (sd_degenerate(parent)) | |
5288 | return 1; | |
5289 | ||
5290 | if (!cpus_equal(sd->span, parent->span)) | |
5291 | return 0; | |
5292 | ||
5293 | /* Does parent contain flags not in child? */ | |
5294 | /* WAKE_BALANCE is a subset of WAKE_AFFINE */ | |
5295 | if (cflags & SD_WAKE_AFFINE) | |
5296 | pflags &= ~SD_WAKE_BALANCE; | |
5297 | /* Flags needing groups don't count if only 1 group in parent */ | |
5298 | if (parent->groups == parent->groups->next) { | |
5299 | pflags &= ~(SD_LOAD_BALANCE | | |
5300 | SD_BALANCE_NEWIDLE | | |
5301 | SD_BALANCE_FORK | | |
5302 | SD_BALANCE_EXEC); | |
5303 | } | |
5304 | if (~cflags & pflags) | |
5305 | return 0; | |
5306 | ||
5307 | return 1; | |
5308 | } | |
5309 | ||
1da177e4 LT |
5310 | /* |
5311 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must | |
5312 | * hold the hotplug lock. | |
5313 | */ | |
9c1cfda2 | 5314 | static void cpu_attach_domain(struct sched_domain *sd, int cpu) |
1da177e4 | 5315 | { |
1da177e4 | 5316 | runqueue_t *rq = cpu_rq(cpu); |
245af2c7 SS |
5317 | struct sched_domain *tmp; |
5318 | ||
5319 | /* Remove the sched domains which do not contribute to scheduling. */ | |
5320 | for (tmp = sd; tmp; tmp = tmp->parent) { | |
5321 | struct sched_domain *parent = tmp->parent; | |
5322 | if (!parent) | |
5323 | break; | |
5324 | if (sd_parent_degenerate(tmp, parent)) | |
5325 | tmp->parent = parent->parent; | |
5326 | } | |
5327 | ||
5328 | if (sd && sd_degenerate(sd)) | |
5329 | sd = sd->parent; | |
1da177e4 LT |
5330 | |
5331 | sched_domain_debug(sd, cpu); | |
5332 | ||
674311d5 | 5333 | rcu_assign_pointer(rq->sd, sd); |
1da177e4 LT |
5334 | } |
5335 | ||
5336 | /* cpus with isolated domains */ | |
9c1cfda2 | 5337 | static cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE; |
1da177e4 LT |
5338 | |
5339 | /* Setup the mask of cpus configured for isolated domains */ | |
5340 | static int __init isolated_cpu_setup(char *str) | |
5341 | { | |
5342 | int ints[NR_CPUS], i; | |
5343 | ||
5344 | str = get_options(str, ARRAY_SIZE(ints), ints); | |
5345 | cpus_clear(cpu_isolated_map); | |
5346 | for (i = 1; i <= ints[0]; i++) | |
5347 | if (ints[i] < NR_CPUS) | |
5348 | cpu_set(ints[i], cpu_isolated_map); | |
5349 | return 1; | |
5350 | } | |
5351 | ||
5352 | __setup ("isolcpus=", isolated_cpu_setup); | |
5353 | ||
5354 | /* | |
5355 | * init_sched_build_groups takes an array of groups, the cpumask we wish | |
5356 | * to span, and a pointer to a function which identifies what group a CPU | |
5357 | * belongs to. The return value of group_fn must be a valid index into the | |
5358 | * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we | |
5359 | * keep track of groups covered with a cpumask_t). | |
5360 | * | |
5361 | * init_sched_build_groups will build a circular linked list of the groups | |
5362 | * covered by the given span, and will set each group's ->cpumask correctly, | |
5363 | * and ->cpu_power to 0. | |
5364 | */ | |
9c1cfda2 JH |
5365 | static void init_sched_build_groups(struct sched_group groups[], cpumask_t span, |
5366 | int (*group_fn)(int cpu)) | |
1da177e4 LT |
5367 | { |
5368 | struct sched_group *first = NULL, *last = NULL; | |
5369 | cpumask_t covered = CPU_MASK_NONE; | |
5370 | int i; | |
5371 | ||
5372 | for_each_cpu_mask(i, span) { | |
5373 | int group = group_fn(i); | |
5374 | struct sched_group *sg = &groups[group]; | |
5375 | int j; | |
5376 | ||
5377 | if (cpu_isset(i, covered)) | |
5378 | continue; | |
5379 | ||
5380 | sg->cpumask = CPU_MASK_NONE; | |
5381 | sg->cpu_power = 0; | |
5382 | ||
5383 | for_each_cpu_mask(j, span) { | |
5384 | if (group_fn(j) != group) | |
5385 | continue; | |
5386 | ||
5387 | cpu_set(j, covered); | |
5388 | cpu_set(j, sg->cpumask); | |
5389 | } | |
5390 | if (!first) | |
5391 | first = sg; | |
5392 | if (last) | |
5393 | last->next = sg; | |
5394 | last = sg; | |
5395 | } | |
5396 | last->next = first; | |
5397 | } | |
5398 | ||
9c1cfda2 | 5399 | #define SD_NODES_PER_DOMAIN 16 |
1da177e4 | 5400 | |
198e2f18 | 5401 | /* |
5402 | * Self-tuning task migration cost measurement between source and target CPUs. | |
5403 | * | |
5404 | * This is done by measuring the cost of manipulating buffers of varying | |
5405 | * sizes. For a given buffer-size here are the steps that are taken: | |
5406 | * | |
5407 | * 1) the source CPU reads+dirties a shared buffer | |
5408 | * 2) the target CPU reads+dirties the same shared buffer | |
5409 | * | |
5410 | * We measure how long they take, in the following 4 scenarios: | |
5411 | * | |
5412 | * - source: CPU1, target: CPU2 | cost1 | |
5413 | * - source: CPU2, target: CPU1 | cost2 | |
5414 | * - source: CPU1, target: CPU1 | cost3 | |
5415 | * - source: CPU2, target: CPU2 | cost4 | |
5416 | * | |
5417 | * We then calculate the cost3+cost4-cost1-cost2 difference - this is | |
5418 | * the cost of migration. | |
5419 | * | |
5420 | * We then start off from a small buffer-size and iterate up to larger | |
5421 | * buffer sizes, in 5% steps - measuring each buffer-size separately, and | |
5422 | * doing a maximum search for the cost. (The maximum cost for a migration | |
5423 | * normally occurs when the working set size is around the effective cache | |
5424 | * size.) | |
5425 | */ | |
5426 | #define SEARCH_SCOPE 2 | |
5427 | #define MIN_CACHE_SIZE (64*1024U) | |
5428 | #define DEFAULT_CACHE_SIZE (5*1024*1024U) | |
70b4d63e | 5429 | #define ITERATIONS 1 |
198e2f18 | 5430 | #define SIZE_THRESH 130 |
5431 | #define COST_THRESH 130 | |
5432 | ||
5433 | /* | |
5434 | * The migration cost is a function of 'domain distance'. Domain | |
5435 | * distance is the number of steps a CPU has to iterate down its | |
5436 | * domain tree to share a domain with the other CPU. The farther | |
5437 | * two CPUs are from each other, the larger the distance gets. | |
5438 | * | |
5439 | * Note that we use the distance only to cache measurement results, | |
5440 | * the distance value is not used numerically otherwise. When two | |
5441 | * CPUs have the same distance it is assumed that the migration | |
5442 | * cost is the same. (this is a simplification but quite practical) | |
5443 | */ | |
5444 | #define MAX_DOMAIN_DISTANCE 32 | |
5445 | ||
5446 | static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] = | |
4bbf39c2 IM |
5447 | { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] = |
5448 | /* | |
5449 | * Architectures may override the migration cost and thus avoid | |
5450 | * boot-time calibration. Unit is nanoseconds. Mostly useful for | |
5451 | * virtualized hardware: | |
5452 | */ | |
5453 | #ifdef CONFIG_DEFAULT_MIGRATION_COST | |
5454 | CONFIG_DEFAULT_MIGRATION_COST | |
5455 | #else | |
5456 | -1LL | |
5457 | #endif | |
5458 | }; | |
198e2f18 | 5459 | |
5460 | /* | |
5461 | * Allow override of migration cost - in units of microseconds. | |
5462 | * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost | |
5463 | * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs: | |
5464 | */ | |
5465 | static int __init migration_cost_setup(char *str) | |
5466 | { | |
5467 | int ints[MAX_DOMAIN_DISTANCE+1], i; | |
5468 | ||
5469 | str = get_options(str, ARRAY_SIZE(ints), ints); | |
5470 | ||
5471 | printk("#ints: %d\n", ints[0]); | |
5472 | for (i = 1; i <= ints[0]; i++) { | |
5473 | migration_cost[i-1] = (unsigned long long)ints[i]*1000; | |
5474 | printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]); | |
5475 | } | |
5476 | return 1; | |
5477 | } | |
5478 | ||
5479 | __setup ("migration_cost=", migration_cost_setup); | |
5480 | ||
5481 | /* | |
5482 | * Global multiplier (divisor) for migration-cutoff values, | |
5483 | * in percentiles. E.g. use a value of 150 to get 1.5 times | |
5484 | * longer cache-hot cutoff times. | |
5485 | * | |
5486 | * (We scale it from 100 to 128 to long long handling easier.) | |
5487 | */ | |
5488 | ||
5489 | #define MIGRATION_FACTOR_SCALE 128 | |
5490 | ||
5491 | static unsigned int migration_factor = MIGRATION_FACTOR_SCALE; | |
5492 | ||
5493 | static int __init setup_migration_factor(char *str) | |
5494 | { | |
5495 | get_option(&str, &migration_factor); | |
5496 | migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100; | |
5497 | return 1; | |
5498 | } | |
5499 | ||
5500 | __setup("migration_factor=", setup_migration_factor); | |
5501 | ||
5502 | /* | |
5503 | * Estimated distance of two CPUs, measured via the number of domains | |
5504 | * we have to pass for the two CPUs to be in the same span: | |
5505 | */ | |
5506 | static unsigned long domain_distance(int cpu1, int cpu2) | |
5507 | { | |
5508 | unsigned long distance = 0; | |
5509 | struct sched_domain *sd; | |
5510 | ||
5511 | for_each_domain(cpu1, sd) { | |
5512 | WARN_ON(!cpu_isset(cpu1, sd->span)); | |
5513 | if (cpu_isset(cpu2, sd->span)) | |
5514 | return distance; | |
5515 | distance++; | |
5516 | } | |
5517 | if (distance >= MAX_DOMAIN_DISTANCE) { | |
5518 | WARN_ON(1); | |
5519 | distance = MAX_DOMAIN_DISTANCE-1; | |
5520 | } | |
5521 | ||
5522 | return distance; | |
5523 | } | |
5524 | ||
5525 | static unsigned int migration_debug; | |
5526 | ||
5527 | static int __init setup_migration_debug(char *str) | |
5528 | { | |
5529 | get_option(&str, &migration_debug); | |
5530 | return 1; | |
5531 | } | |
5532 | ||
5533 | __setup("migration_debug=", setup_migration_debug); | |
5534 | ||
5535 | /* | |
5536 | * Maximum cache-size that the scheduler should try to measure. | |
5537 | * Architectures with larger caches should tune this up during | |
5538 | * bootup. Gets used in the domain-setup code (i.e. during SMP | |
5539 | * bootup). | |
5540 | */ | |
5541 | unsigned int max_cache_size; | |
5542 | ||
5543 | static int __init setup_max_cache_size(char *str) | |
5544 | { | |
5545 | get_option(&str, &max_cache_size); | |
5546 | return 1; | |
5547 | } | |
5548 | ||
5549 | __setup("max_cache_size=", setup_max_cache_size); | |
5550 | ||
5551 | /* | |
5552 | * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This | |
5553 | * is the operation that is timed, so we try to generate unpredictable | |
5554 | * cachemisses that still end up filling the L2 cache: | |
5555 | */ | |
5556 | static void touch_cache(void *__cache, unsigned long __size) | |
5557 | { | |
5558 | unsigned long size = __size/sizeof(long), chunk1 = size/3, | |
5559 | chunk2 = 2*size/3; | |
5560 | unsigned long *cache = __cache; | |
5561 | int i; | |
5562 | ||
5563 | for (i = 0; i < size/6; i += 8) { | |
5564 | switch (i % 6) { | |
5565 | case 0: cache[i]++; | |
5566 | case 1: cache[size-1-i]++; | |
5567 | case 2: cache[chunk1-i]++; | |
5568 | case 3: cache[chunk1+i]++; | |
5569 | case 4: cache[chunk2-i]++; | |
5570 | case 5: cache[chunk2+i]++; | |
5571 | } | |
5572 | } | |
5573 | } | |
5574 | ||
5575 | /* | |
5576 | * Measure the cache-cost of one task migration. Returns in units of nsec. | |
5577 | */ | |
5578 | static unsigned long long measure_one(void *cache, unsigned long size, | |
5579 | int source, int target) | |
5580 | { | |
5581 | cpumask_t mask, saved_mask; | |
5582 | unsigned long long t0, t1, t2, t3, cost; | |
5583 | ||
5584 | saved_mask = current->cpus_allowed; | |
5585 | ||
5586 | /* | |
5587 | * Flush source caches to RAM and invalidate them: | |
5588 | */ | |
5589 | sched_cacheflush(); | |
5590 | ||
5591 | /* | |
5592 | * Migrate to the source CPU: | |
5593 | */ | |
5594 | mask = cpumask_of_cpu(source); | |
5595 | set_cpus_allowed(current, mask); | |
5596 | WARN_ON(smp_processor_id() != source); | |
5597 | ||
5598 | /* | |
5599 | * Dirty the working set: | |
5600 | */ | |
5601 | t0 = sched_clock(); | |
5602 | touch_cache(cache, size); | |
5603 | t1 = sched_clock(); | |
5604 | ||
5605 | /* | |
5606 | * Migrate to the target CPU, dirty the L2 cache and access | |
5607 | * the shared buffer. (which represents the working set | |
5608 | * of a migrated task.) | |
5609 | */ | |
5610 | mask = cpumask_of_cpu(target); | |
5611 | set_cpus_allowed(current, mask); | |
5612 | WARN_ON(smp_processor_id() != target); | |
5613 | ||
5614 | t2 = sched_clock(); | |
5615 | touch_cache(cache, size); | |
5616 | t3 = sched_clock(); | |
5617 | ||
5618 | cost = t1-t0 + t3-t2; | |
5619 | ||
5620 | if (migration_debug >= 2) | |
5621 | printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n", | |
5622 | source, target, t1-t0, t1-t0, t3-t2, cost); | |
5623 | /* | |
5624 | * Flush target caches to RAM and invalidate them: | |
5625 | */ | |
5626 | sched_cacheflush(); | |
5627 | ||
5628 | set_cpus_allowed(current, saved_mask); | |
5629 | ||
5630 | return cost; | |
5631 | } | |
5632 | ||
5633 | /* | |
5634 | * Measure a series of task migrations and return the average | |
5635 | * result. Since this code runs early during bootup the system | |
5636 | * is 'undisturbed' and the average latency makes sense. | |
5637 | * | |
5638 | * The algorithm in essence auto-detects the relevant cache-size, | |
5639 | * so it will properly detect different cachesizes for different | |
5640 | * cache-hierarchies, depending on how the CPUs are connected. | |
5641 | * | |
5642 | * Architectures can prime the upper limit of the search range via | |
5643 | * max_cache_size, otherwise the search range defaults to 20MB...64K. | |
5644 | */ | |
5645 | static unsigned long long | |
5646 | measure_cost(int cpu1, int cpu2, void *cache, unsigned int size) | |
5647 | { | |
5648 | unsigned long long cost1, cost2; | |
5649 | int i; | |
5650 | ||
5651 | /* | |
5652 | * Measure the migration cost of 'size' bytes, over an | |
5653 | * average of 10 runs: | |
5654 | * | |
5655 | * (We perturb the cache size by a small (0..4k) | |
5656 | * value to compensate size/alignment related artifacts. | |
5657 | * We also subtract the cost of the operation done on | |
5658 | * the same CPU.) | |
5659 | */ | |
5660 | cost1 = 0; | |
5661 | ||
5662 | /* | |
5663 | * dry run, to make sure we start off cache-cold on cpu1, | |
5664 | * and to get any vmalloc pagefaults in advance: | |
5665 | */ | |
5666 | measure_one(cache, size, cpu1, cpu2); | |
5667 | for (i = 0; i < ITERATIONS; i++) | |
5668 | cost1 += measure_one(cache, size - i*1024, cpu1, cpu2); | |
5669 | ||
5670 | measure_one(cache, size, cpu2, cpu1); | |
5671 | for (i = 0; i < ITERATIONS; i++) | |
5672 | cost1 += measure_one(cache, size - i*1024, cpu2, cpu1); | |
5673 | ||
5674 | /* | |
5675 | * (We measure the non-migrating [cached] cost on both | |
5676 | * cpu1 and cpu2, to handle CPUs with different speeds) | |
5677 | */ | |
5678 | cost2 = 0; | |
5679 | ||
5680 | measure_one(cache, size, cpu1, cpu1); | |
5681 | for (i = 0; i < ITERATIONS; i++) | |
5682 | cost2 += measure_one(cache, size - i*1024, cpu1, cpu1); | |
5683 | ||
5684 | measure_one(cache, size, cpu2, cpu2); | |
5685 | for (i = 0; i < ITERATIONS; i++) | |
5686 | cost2 += measure_one(cache, size - i*1024, cpu2, cpu2); | |
5687 | ||
5688 | /* | |
5689 | * Get the per-iteration migration cost: | |
5690 | */ | |
5691 | do_div(cost1, 2*ITERATIONS); | |
5692 | do_div(cost2, 2*ITERATIONS); | |
5693 | ||
5694 | return cost1 - cost2; | |
5695 | } | |
5696 | ||
5697 | static unsigned long long measure_migration_cost(int cpu1, int cpu2) | |
5698 | { | |
5699 | unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0; | |
5700 | unsigned int max_size, size, size_found = 0; | |
5701 | long long cost = 0, prev_cost; | |
5702 | void *cache; | |
5703 | ||
5704 | /* | |
5705 | * Search from max_cache_size*5 down to 64K - the real relevant | |
5706 | * cachesize has to lie somewhere inbetween. | |
5707 | */ | |
5708 | if (max_cache_size) { | |
5709 | max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE); | |
5710 | size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE); | |
5711 | } else { | |
5712 | /* | |
5713 | * Since we have no estimation about the relevant | |
5714 | * search range | |
5715 | */ | |
5716 | max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE; | |
5717 | size = MIN_CACHE_SIZE; | |
5718 | } | |
5719 | ||
5720 | if (!cpu_online(cpu1) || !cpu_online(cpu2)) { | |
5721 | printk("cpu %d and %d not both online!\n", cpu1, cpu2); | |
5722 | return 0; | |
5723 | } | |
5724 | ||
5725 | /* | |
5726 | * Allocate the working set: | |
5727 | */ | |
5728 | cache = vmalloc(max_size); | |
5729 | if (!cache) { | |
5730 | printk("could not vmalloc %d bytes for cache!\n", 2*max_size); | |
5731 | return 1000000; // return 1 msec on very small boxen | |
5732 | } | |
5733 | ||
5734 | while (size <= max_size) { | |
5735 | prev_cost = cost; | |
5736 | cost = measure_cost(cpu1, cpu2, cache, size); | |
5737 | ||
5738 | /* | |
5739 | * Update the max: | |
5740 | */ | |
5741 | if (cost > 0) { | |
5742 | if (max_cost < cost) { | |
5743 | max_cost = cost; | |
5744 | size_found = size; | |
5745 | } | |
5746 | } | |
5747 | /* | |
5748 | * Calculate average fluctuation, we use this to prevent | |
5749 | * noise from triggering an early break out of the loop: | |
5750 | */ | |
5751 | fluct = abs(cost - prev_cost); | |
5752 | avg_fluct = (avg_fluct + fluct)/2; | |
5753 | ||
5754 | if (migration_debug) | |
5755 | printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n", | |
5756 | cpu1, cpu2, size, | |
5757 | (long)cost / 1000000, | |
5758 | ((long)cost / 100000) % 10, | |
5759 | (long)max_cost / 1000000, | |
5760 | ((long)max_cost / 100000) % 10, | |
5761 | domain_distance(cpu1, cpu2), | |
5762 | cost, avg_fluct); | |
5763 | ||
5764 | /* | |
5765 | * If we iterated at least 20% past the previous maximum, | |
5766 | * and the cost has dropped by more than 20% already, | |
5767 | * (taking fluctuations into account) then we assume to | |
5768 | * have found the maximum and break out of the loop early: | |
5769 | */ | |
5770 | if (size_found && (size*100 > size_found*SIZE_THRESH)) | |
5771 | if (cost+avg_fluct <= 0 || | |
5772 | max_cost*100 > (cost+avg_fluct)*COST_THRESH) { | |
5773 | ||
5774 | if (migration_debug) | |
5775 | printk("-> found max.\n"); | |
5776 | break; | |
5777 | } | |
5778 | /* | |
70b4d63e | 5779 | * Increase the cachesize in 10% steps: |
198e2f18 | 5780 | */ |
70b4d63e | 5781 | size = size * 10 / 9; |
198e2f18 | 5782 | } |
5783 | ||
5784 | if (migration_debug) | |
5785 | printk("[%d][%d] working set size found: %d, cost: %Ld\n", | |
5786 | cpu1, cpu2, size_found, max_cost); | |
5787 | ||
5788 | vfree(cache); | |
5789 | ||
5790 | /* | |
5791 | * A task is considered 'cache cold' if at least 2 times | |
5792 | * the worst-case cost of migration has passed. | |
5793 | * | |
5794 | * (this limit is only listened to if the load-balancing | |
5795 | * situation is 'nice' - if there is a large imbalance we | |
5796 | * ignore it for the sake of CPU utilization and | |
5797 | * processing fairness.) | |
5798 | */ | |
5799 | return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE; | |
5800 | } | |
5801 | ||
5802 | static void calibrate_migration_costs(const cpumask_t *cpu_map) | |
5803 | { | |
5804 | int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id(); | |
5805 | unsigned long j0, j1, distance, max_distance = 0; | |
5806 | struct sched_domain *sd; | |
5807 | ||
5808 | j0 = jiffies; | |
5809 | ||
5810 | /* | |
5811 | * First pass - calculate the cacheflush times: | |
5812 | */ | |
5813 | for_each_cpu_mask(cpu1, *cpu_map) { | |
5814 | for_each_cpu_mask(cpu2, *cpu_map) { | |
5815 | if (cpu1 == cpu2) | |
5816 | continue; | |
5817 | distance = domain_distance(cpu1, cpu2); | |
5818 | max_distance = max(max_distance, distance); | |
5819 | /* | |
5820 | * No result cached yet? | |
5821 | */ | |
5822 | if (migration_cost[distance] == -1LL) | |
5823 | migration_cost[distance] = | |
5824 | measure_migration_cost(cpu1, cpu2); | |
5825 | } | |
5826 | } | |
5827 | /* | |
5828 | * Second pass - update the sched domain hierarchy with | |
5829 | * the new cache-hot-time estimations: | |
5830 | */ | |
5831 | for_each_cpu_mask(cpu, *cpu_map) { | |
5832 | distance = 0; | |
5833 | for_each_domain(cpu, sd) { | |
5834 | sd->cache_hot_time = migration_cost[distance]; | |
5835 | distance++; | |
5836 | } | |
5837 | } | |
5838 | /* | |
5839 | * Print the matrix: | |
5840 | */ | |
5841 | if (migration_debug) | |
5842 | printk("migration: max_cache_size: %d, cpu: %d MHz:\n", | |
5843 | max_cache_size, | |
5844 | #ifdef CONFIG_X86 | |
5845 | cpu_khz/1000 | |
5846 | #else | |
5847 | -1 | |
5848 | #endif | |
5849 | ); | |
bd576c95 CE |
5850 | if (system_state == SYSTEM_BOOTING) { |
5851 | printk("migration_cost="); | |
5852 | for (distance = 0; distance <= max_distance; distance++) { | |
5853 | if (distance) | |
5854 | printk(","); | |
5855 | printk("%ld", (long)migration_cost[distance] / 1000); | |
5856 | } | |
5857 | printk("\n"); | |
198e2f18 | 5858 | } |
198e2f18 | 5859 | j1 = jiffies; |
5860 | if (migration_debug) | |
5861 | printk("migration: %ld seconds\n", (j1-j0)/HZ); | |
5862 | ||
5863 | /* | |
5864 | * Move back to the original CPU. NUMA-Q gets confused | |
5865 | * if we migrate to another quad during bootup. | |
5866 | */ | |
5867 | if (raw_smp_processor_id() != orig_cpu) { | |
5868 | cpumask_t mask = cpumask_of_cpu(orig_cpu), | |
5869 | saved_mask = current->cpus_allowed; | |
5870 | ||
5871 | set_cpus_allowed(current, mask); | |
5872 | set_cpus_allowed(current, saved_mask); | |
5873 | } | |
5874 | } | |
5875 | ||
9c1cfda2 | 5876 | #ifdef CONFIG_NUMA |
198e2f18 | 5877 | |
9c1cfda2 JH |
5878 | /** |
5879 | * find_next_best_node - find the next node to include in a sched_domain | |
5880 | * @node: node whose sched_domain we're building | |
5881 | * @used_nodes: nodes already in the sched_domain | |
5882 | * | |
5883 | * Find the next node to include in a given scheduling domain. Simply | |
5884 | * finds the closest node not already in the @used_nodes map. | |
5885 | * | |
5886 | * Should use nodemask_t. | |
5887 | */ | |
5888 | static int find_next_best_node(int node, unsigned long *used_nodes) | |
5889 | { | |
5890 | int i, n, val, min_val, best_node = 0; | |
5891 | ||
5892 | min_val = INT_MAX; | |
5893 | ||
5894 | for (i = 0; i < MAX_NUMNODES; i++) { | |
5895 | /* Start at @node */ | |
5896 | n = (node + i) % MAX_NUMNODES; | |
5897 | ||
5898 | if (!nr_cpus_node(n)) | |
5899 | continue; | |
5900 | ||
5901 | /* Skip already used nodes */ | |
5902 | if (test_bit(n, used_nodes)) | |
5903 | continue; | |
5904 | ||
5905 | /* Simple min distance search */ | |
5906 | val = node_distance(node, n); | |
5907 | ||
5908 | if (val < min_val) { | |
5909 | min_val = val; | |
5910 | best_node = n; | |
5911 | } | |
5912 | } | |
5913 | ||
5914 | set_bit(best_node, used_nodes); | |
5915 | return best_node; | |
5916 | } | |
5917 | ||
5918 | /** | |
5919 | * sched_domain_node_span - get a cpumask for a node's sched_domain | |
5920 | * @node: node whose cpumask we're constructing | |
5921 | * @size: number of nodes to include in this span | |
5922 | * | |
5923 | * Given a node, construct a good cpumask for its sched_domain to span. It | |
5924 | * should be one that prevents unnecessary balancing, but also spreads tasks | |
5925 | * out optimally. | |
5926 | */ | |
5927 | static cpumask_t sched_domain_node_span(int node) | |
5928 | { | |
5929 | int i; | |
5930 | cpumask_t span, nodemask; | |
5931 | DECLARE_BITMAP(used_nodes, MAX_NUMNODES); | |
5932 | ||
5933 | cpus_clear(span); | |
5934 | bitmap_zero(used_nodes, MAX_NUMNODES); | |
5935 | ||
5936 | nodemask = node_to_cpumask(node); | |
5937 | cpus_or(span, span, nodemask); | |
5938 | set_bit(node, used_nodes); | |
5939 | ||
5940 | for (i = 1; i < SD_NODES_PER_DOMAIN; i++) { | |
5941 | int next_node = find_next_best_node(node, used_nodes); | |
5942 | nodemask = node_to_cpumask(next_node); | |
5943 | cpus_or(span, span, nodemask); | |
5944 | } | |
5945 | ||
5946 | return span; | |
5947 | } | |
5948 | #endif | |
5949 | ||
5c45bf27 | 5950 | int sched_smt_power_savings = 0, sched_mc_power_savings = 0; |
9c1cfda2 JH |
5951 | /* |
5952 | * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we | |
5953 | * can switch it on easily if needed. | |
5954 | */ | |
1da177e4 LT |
5955 | #ifdef CONFIG_SCHED_SMT |
5956 | static DEFINE_PER_CPU(struct sched_domain, cpu_domains); | |
5957 | static struct sched_group sched_group_cpus[NR_CPUS]; | |
1a20ff27 | 5958 | static int cpu_to_cpu_group(int cpu) |
1da177e4 LT |
5959 | { |
5960 | return cpu; | |
5961 | } | |
5962 | #endif | |
5963 | ||
1e9f28fa SS |
5964 | #ifdef CONFIG_SCHED_MC |
5965 | static DEFINE_PER_CPU(struct sched_domain, core_domains); | |
36938169 | 5966 | static struct sched_group *sched_group_core_bycpu[NR_CPUS]; |
1e9f28fa SS |
5967 | #endif |
5968 | ||
5969 | #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT) | |
5970 | static int cpu_to_core_group(int cpu) | |
5971 | { | |
5972 | return first_cpu(cpu_sibling_map[cpu]); | |
5973 | } | |
5974 | #elif defined(CONFIG_SCHED_MC) | |
5975 | static int cpu_to_core_group(int cpu) | |
5976 | { | |
5977 | return cpu; | |
5978 | } | |
5979 | #endif | |
5980 | ||
1da177e4 | 5981 | static DEFINE_PER_CPU(struct sched_domain, phys_domains); |
36938169 | 5982 | static struct sched_group *sched_group_phys_bycpu[NR_CPUS]; |
1a20ff27 | 5983 | static int cpu_to_phys_group(int cpu) |
1da177e4 | 5984 | { |
1e9f28fa SS |
5985 | #if defined(CONFIG_SCHED_MC) |
5986 | cpumask_t mask = cpu_coregroup_map(cpu); | |
5987 | return first_cpu(mask); | |
5988 | #elif defined(CONFIG_SCHED_SMT) | |
1da177e4 LT |
5989 | return first_cpu(cpu_sibling_map[cpu]); |
5990 | #else | |
5991 | return cpu; | |
5992 | #endif | |
5993 | } | |
5994 | ||
5995 | #ifdef CONFIG_NUMA | |
1da177e4 | 5996 | /* |
9c1cfda2 JH |
5997 | * The init_sched_build_groups can't handle what we want to do with node |
5998 | * groups, so roll our own. Now each node has its own list of groups which | |
5999 | * gets dynamically allocated. | |
1da177e4 | 6000 | */ |
9c1cfda2 | 6001 | static DEFINE_PER_CPU(struct sched_domain, node_domains); |
d1b55138 | 6002 | static struct sched_group **sched_group_nodes_bycpu[NR_CPUS]; |
1da177e4 | 6003 | |
9c1cfda2 | 6004 | static DEFINE_PER_CPU(struct sched_domain, allnodes_domains); |
d1b55138 | 6005 | static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS]; |
9c1cfda2 JH |
6006 | |
6007 | static int cpu_to_allnodes_group(int cpu) | |
6008 | { | |
6009 | return cpu_to_node(cpu); | |
1da177e4 | 6010 | } |
08069033 SS |
6011 | static void init_numa_sched_groups_power(struct sched_group *group_head) |
6012 | { | |
6013 | struct sched_group *sg = group_head; | |
6014 | int j; | |
6015 | ||
6016 | if (!sg) | |
6017 | return; | |
6018 | next_sg: | |
6019 | for_each_cpu_mask(j, sg->cpumask) { | |
6020 | struct sched_domain *sd; | |
6021 | ||
6022 | sd = &per_cpu(phys_domains, j); | |
6023 | if (j != first_cpu(sd->groups->cpumask)) { | |
6024 | /* | |
6025 | * Only add "power" once for each | |
6026 | * physical package. | |
6027 | */ | |
6028 | continue; | |
6029 | } | |
6030 | ||
6031 | sg->cpu_power += sd->groups->cpu_power; | |
6032 | } | |
6033 | sg = sg->next; | |
6034 | if (sg != group_head) | |
6035 | goto next_sg; | |
6036 | } | |
1da177e4 LT |
6037 | #endif |
6038 | ||
51888ca2 SV |
6039 | /* Free memory allocated for various sched_group structures */ |
6040 | static void free_sched_groups(const cpumask_t *cpu_map) | |
6041 | { | |
36938169 | 6042 | int cpu; |
51888ca2 SV |
6043 | #ifdef CONFIG_NUMA |
6044 | int i; | |
51888ca2 SV |
6045 | |
6046 | for_each_cpu_mask(cpu, *cpu_map) { | |
6047 | struct sched_group *sched_group_allnodes | |
6048 | = sched_group_allnodes_bycpu[cpu]; | |
6049 | struct sched_group **sched_group_nodes | |
6050 | = sched_group_nodes_bycpu[cpu]; | |
6051 | ||
6052 | if (sched_group_allnodes) { | |
6053 | kfree(sched_group_allnodes); | |
6054 | sched_group_allnodes_bycpu[cpu] = NULL; | |
6055 | } | |
6056 | ||
6057 | if (!sched_group_nodes) | |
6058 | continue; | |
6059 | ||
6060 | for (i = 0; i < MAX_NUMNODES; i++) { | |
6061 | cpumask_t nodemask = node_to_cpumask(i); | |
6062 | struct sched_group *oldsg, *sg = sched_group_nodes[i]; | |
6063 | ||
6064 | cpus_and(nodemask, nodemask, *cpu_map); | |
6065 | if (cpus_empty(nodemask)) | |
6066 | continue; | |
6067 | ||
6068 | if (sg == NULL) | |
6069 | continue; | |
6070 | sg = sg->next; | |
6071 | next_sg: | |
6072 | oldsg = sg; | |
6073 | sg = sg->next; | |
6074 | kfree(oldsg); | |
6075 | if (oldsg != sched_group_nodes[i]) | |
6076 | goto next_sg; | |
6077 | } | |
6078 | kfree(sched_group_nodes); | |
6079 | sched_group_nodes_bycpu[cpu] = NULL; | |
6080 | } | |
6081 | #endif | |
36938169 SV |
6082 | for_each_cpu_mask(cpu, *cpu_map) { |
6083 | if (sched_group_phys_bycpu[cpu]) { | |
6084 | kfree(sched_group_phys_bycpu[cpu]); | |
6085 | sched_group_phys_bycpu[cpu] = NULL; | |
6086 | } | |
6087 | #ifdef CONFIG_SCHED_MC | |
6088 | if (sched_group_core_bycpu[cpu]) { | |
6089 | kfree(sched_group_core_bycpu[cpu]); | |
6090 | sched_group_core_bycpu[cpu] = NULL; | |
6091 | } | |
6092 | #endif | |
6093 | } | |
51888ca2 SV |
6094 | } |
6095 | ||
1da177e4 | 6096 | /* |
1a20ff27 DG |
6097 | * Build sched domains for a given set of cpus and attach the sched domains |
6098 | * to the individual cpus | |
1da177e4 | 6099 | */ |
51888ca2 | 6100 | static int build_sched_domains(const cpumask_t *cpu_map) |
1da177e4 LT |
6101 | { |
6102 | int i; | |
36938169 SV |
6103 | struct sched_group *sched_group_phys = NULL; |
6104 | #ifdef CONFIG_SCHED_MC | |
6105 | struct sched_group *sched_group_core = NULL; | |
6106 | #endif | |
d1b55138 JH |
6107 | #ifdef CONFIG_NUMA |
6108 | struct sched_group **sched_group_nodes = NULL; | |
6109 | struct sched_group *sched_group_allnodes = NULL; | |
6110 | ||
6111 | /* | |
6112 | * Allocate the per-node list of sched groups | |
6113 | */ | |
51888ca2 | 6114 | sched_group_nodes = kzalloc(sizeof(struct sched_group*)*MAX_NUMNODES, |
d3a5aa98 | 6115 | GFP_KERNEL); |
d1b55138 JH |
6116 | if (!sched_group_nodes) { |
6117 | printk(KERN_WARNING "Can not alloc sched group node list\n"); | |
51888ca2 | 6118 | return -ENOMEM; |
d1b55138 JH |
6119 | } |
6120 | sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes; | |
6121 | #endif | |
1da177e4 LT |
6122 | |
6123 | /* | |
1a20ff27 | 6124 | * Set up domains for cpus specified by the cpu_map. |
1da177e4 | 6125 | */ |
1a20ff27 | 6126 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 LT |
6127 | int group; |
6128 | struct sched_domain *sd = NULL, *p; | |
6129 | cpumask_t nodemask = node_to_cpumask(cpu_to_node(i)); | |
6130 | ||
1a20ff27 | 6131 | cpus_and(nodemask, nodemask, *cpu_map); |
1da177e4 LT |
6132 | |
6133 | #ifdef CONFIG_NUMA | |
d1b55138 | 6134 | if (cpus_weight(*cpu_map) |
9c1cfda2 | 6135 | > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) { |
d1b55138 JH |
6136 | if (!sched_group_allnodes) { |
6137 | sched_group_allnodes | |
6138 | = kmalloc(sizeof(struct sched_group) | |
6139 | * MAX_NUMNODES, | |
6140 | GFP_KERNEL); | |
6141 | if (!sched_group_allnodes) { | |
6142 | printk(KERN_WARNING | |
6143 | "Can not alloc allnodes sched group\n"); | |
51888ca2 | 6144 | goto error; |
d1b55138 JH |
6145 | } |
6146 | sched_group_allnodes_bycpu[i] | |
6147 | = sched_group_allnodes; | |
6148 | } | |
9c1cfda2 JH |
6149 | sd = &per_cpu(allnodes_domains, i); |
6150 | *sd = SD_ALLNODES_INIT; | |
6151 | sd->span = *cpu_map; | |
6152 | group = cpu_to_allnodes_group(i); | |
6153 | sd->groups = &sched_group_allnodes[group]; | |
6154 | p = sd; | |
6155 | } else | |
6156 | p = NULL; | |
6157 | ||
1da177e4 | 6158 | sd = &per_cpu(node_domains, i); |
1da177e4 | 6159 | *sd = SD_NODE_INIT; |
9c1cfda2 JH |
6160 | sd->span = sched_domain_node_span(cpu_to_node(i)); |
6161 | sd->parent = p; | |
6162 | cpus_and(sd->span, sd->span, *cpu_map); | |
1da177e4 LT |
6163 | #endif |
6164 | ||
36938169 SV |
6165 | if (!sched_group_phys) { |
6166 | sched_group_phys | |
6167 | = kmalloc(sizeof(struct sched_group) * NR_CPUS, | |
6168 | GFP_KERNEL); | |
6169 | if (!sched_group_phys) { | |
6170 | printk (KERN_WARNING "Can not alloc phys sched" | |
6171 | "group\n"); | |
6172 | goto error; | |
6173 | } | |
6174 | sched_group_phys_bycpu[i] = sched_group_phys; | |
6175 | } | |
6176 | ||
1da177e4 LT |
6177 | p = sd; |
6178 | sd = &per_cpu(phys_domains, i); | |
6179 | group = cpu_to_phys_group(i); | |
6180 | *sd = SD_CPU_INIT; | |
6181 | sd->span = nodemask; | |
6182 | sd->parent = p; | |
6183 | sd->groups = &sched_group_phys[group]; | |
6184 | ||
1e9f28fa | 6185 | #ifdef CONFIG_SCHED_MC |
36938169 SV |
6186 | if (!sched_group_core) { |
6187 | sched_group_core | |
6188 | = kmalloc(sizeof(struct sched_group) * NR_CPUS, | |
6189 | GFP_KERNEL); | |
6190 | if (!sched_group_core) { | |
6191 | printk (KERN_WARNING "Can not alloc core sched" | |
6192 | "group\n"); | |
6193 | goto error; | |
6194 | } | |
6195 | sched_group_core_bycpu[i] = sched_group_core; | |
6196 | } | |
6197 | ||
1e9f28fa SS |
6198 | p = sd; |
6199 | sd = &per_cpu(core_domains, i); | |
6200 | group = cpu_to_core_group(i); | |
6201 | *sd = SD_MC_INIT; | |
6202 | sd->span = cpu_coregroup_map(i); | |
6203 | cpus_and(sd->span, sd->span, *cpu_map); | |
6204 | sd->parent = p; | |
6205 | sd->groups = &sched_group_core[group]; | |
6206 | #endif | |
6207 | ||
1da177e4 LT |
6208 | #ifdef CONFIG_SCHED_SMT |
6209 | p = sd; | |
6210 | sd = &per_cpu(cpu_domains, i); | |
6211 | group = cpu_to_cpu_group(i); | |
6212 | *sd = SD_SIBLING_INIT; | |
6213 | sd->span = cpu_sibling_map[i]; | |
1a20ff27 | 6214 | cpus_and(sd->span, sd->span, *cpu_map); |
1da177e4 LT |
6215 | sd->parent = p; |
6216 | sd->groups = &sched_group_cpus[group]; | |
6217 | #endif | |
6218 | } | |
6219 | ||
6220 | #ifdef CONFIG_SCHED_SMT | |
6221 | /* Set up CPU (sibling) groups */ | |
9c1cfda2 | 6222 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 | 6223 | cpumask_t this_sibling_map = cpu_sibling_map[i]; |
1a20ff27 | 6224 | cpus_and(this_sibling_map, this_sibling_map, *cpu_map); |
1da177e4 LT |
6225 | if (i != first_cpu(this_sibling_map)) |
6226 | continue; | |
6227 | ||
6228 | init_sched_build_groups(sched_group_cpus, this_sibling_map, | |
6229 | &cpu_to_cpu_group); | |
6230 | } | |
6231 | #endif | |
6232 | ||
1e9f28fa SS |
6233 | #ifdef CONFIG_SCHED_MC |
6234 | /* Set up multi-core groups */ | |
6235 | for_each_cpu_mask(i, *cpu_map) { | |
6236 | cpumask_t this_core_map = cpu_coregroup_map(i); | |
6237 | cpus_and(this_core_map, this_core_map, *cpu_map); | |
6238 | if (i != first_cpu(this_core_map)) | |
6239 | continue; | |
6240 | init_sched_build_groups(sched_group_core, this_core_map, | |
6241 | &cpu_to_core_group); | |
6242 | } | |
6243 | #endif | |
6244 | ||
6245 | ||
1da177e4 LT |
6246 | /* Set up physical groups */ |
6247 | for (i = 0; i < MAX_NUMNODES; i++) { | |
6248 | cpumask_t nodemask = node_to_cpumask(i); | |
6249 | ||
1a20ff27 | 6250 | cpus_and(nodemask, nodemask, *cpu_map); |
1da177e4 LT |
6251 | if (cpus_empty(nodemask)) |
6252 | continue; | |
6253 | ||
6254 | init_sched_build_groups(sched_group_phys, nodemask, | |
6255 | &cpu_to_phys_group); | |
6256 | } | |
6257 | ||
6258 | #ifdef CONFIG_NUMA | |
6259 | /* Set up node groups */ | |
d1b55138 JH |
6260 | if (sched_group_allnodes) |
6261 | init_sched_build_groups(sched_group_allnodes, *cpu_map, | |
6262 | &cpu_to_allnodes_group); | |
9c1cfda2 JH |
6263 | |
6264 | for (i = 0; i < MAX_NUMNODES; i++) { | |
6265 | /* Set up node groups */ | |
6266 | struct sched_group *sg, *prev; | |
6267 | cpumask_t nodemask = node_to_cpumask(i); | |
6268 | cpumask_t domainspan; | |
6269 | cpumask_t covered = CPU_MASK_NONE; | |
6270 | int j; | |
6271 | ||
6272 | cpus_and(nodemask, nodemask, *cpu_map); | |
d1b55138 JH |
6273 | if (cpus_empty(nodemask)) { |
6274 | sched_group_nodes[i] = NULL; | |
9c1cfda2 | 6275 | continue; |
d1b55138 | 6276 | } |
9c1cfda2 JH |
6277 | |
6278 | domainspan = sched_domain_node_span(i); | |
6279 | cpus_and(domainspan, domainspan, *cpu_map); | |
6280 | ||
15f0b676 | 6281 | sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i); |
51888ca2 SV |
6282 | if (!sg) { |
6283 | printk(KERN_WARNING "Can not alloc domain group for " | |
6284 | "node %d\n", i); | |
6285 | goto error; | |
6286 | } | |
9c1cfda2 JH |
6287 | sched_group_nodes[i] = sg; |
6288 | for_each_cpu_mask(j, nodemask) { | |
6289 | struct sched_domain *sd; | |
6290 | sd = &per_cpu(node_domains, j); | |
6291 | sd->groups = sg; | |
9c1cfda2 JH |
6292 | } |
6293 | sg->cpu_power = 0; | |
6294 | sg->cpumask = nodemask; | |
51888ca2 | 6295 | sg->next = sg; |
9c1cfda2 JH |
6296 | cpus_or(covered, covered, nodemask); |
6297 | prev = sg; | |
6298 | ||
6299 | for (j = 0; j < MAX_NUMNODES; j++) { | |
6300 | cpumask_t tmp, notcovered; | |
6301 | int n = (i + j) % MAX_NUMNODES; | |
6302 | ||
6303 | cpus_complement(notcovered, covered); | |
6304 | cpus_and(tmp, notcovered, *cpu_map); | |
6305 | cpus_and(tmp, tmp, domainspan); | |
6306 | if (cpus_empty(tmp)) | |
6307 | break; | |
6308 | ||
6309 | nodemask = node_to_cpumask(n); | |
6310 | cpus_and(tmp, tmp, nodemask); | |
6311 | if (cpus_empty(tmp)) | |
6312 | continue; | |
6313 | ||
15f0b676 SV |
6314 | sg = kmalloc_node(sizeof(struct sched_group), |
6315 | GFP_KERNEL, i); | |
9c1cfda2 JH |
6316 | if (!sg) { |
6317 | printk(KERN_WARNING | |
6318 | "Can not alloc domain group for node %d\n", j); | |
51888ca2 | 6319 | goto error; |
9c1cfda2 JH |
6320 | } |
6321 | sg->cpu_power = 0; | |
6322 | sg->cpumask = tmp; | |
51888ca2 | 6323 | sg->next = prev->next; |
9c1cfda2 JH |
6324 | cpus_or(covered, covered, tmp); |
6325 | prev->next = sg; | |
6326 | prev = sg; | |
6327 | } | |
9c1cfda2 | 6328 | } |
1da177e4 LT |
6329 | #endif |
6330 | ||
6331 | /* Calculate CPU power for physical packages and nodes */ | |
5c45bf27 | 6332 | #ifdef CONFIG_SCHED_SMT |
1a20ff27 | 6333 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 | 6334 | struct sched_domain *sd; |
1da177e4 | 6335 | sd = &per_cpu(cpu_domains, i); |
5c45bf27 SS |
6336 | sd->groups->cpu_power = SCHED_LOAD_SCALE; |
6337 | } | |
1da177e4 | 6338 | #endif |
1e9f28fa | 6339 | #ifdef CONFIG_SCHED_MC |
5c45bf27 SS |
6340 | for_each_cpu_mask(i, *cpu_map) { |
6341 | int power; | |
6342 | struct sched_domain *sd; | |
1e9f28fa | 6343 | sd = &per_cpu(core_domains, i); |
5c45bf27 SS |
6344 | if (sched_smt_power_savings) |
6345 | power = SCHED_LOAD_SCALE * cpus_weight(sd->groups->cpumask); | |
6346 | else | |
6347 | power = SCHED_LOAD_SCALE + (cpus_weight(sd->groups->cpumask)-1) | |
1e9f28fa SS |
6348 | * SCHED_LOAD_SCALE / 10; |
6349 | sd->groups->cpu_power = power; | |
5c45bf27 SS |
6350 | } |
6351 | #endif | |
1e9f28fa | 6352 | |
5c45bf27 SS |
6353 | for_each_cpu_mask(i, *cpu_map) { |
6354 | struct sched_domain *sd; | |
6355 | #ifdef CONFIG_SCHED_MC | |
1e9f28fa | 6356 | sd = &per_cpu(phys_domains, i); |
5c45bf27 SS |
6357 | if (i != first_cpu(sd->groups->cpumask)) |
6358 | continue; | |
1da177e4 | 6359 | |
5c45bf27 SS |
6360 | sd->groups->cpu_power = 0; |
6361 | if (sched_mc_power_savings || sched_smt_power_savings) { | |
6362 | int j; | |
6363 | ||
6364 | for_each_cpu_mask(j, sd->groups->cpumask) { | |
6365 | struct sched_domain *sd1; | |
6366 | sd1 = &per_cpu(core_domains, j); | |
6367 | /* | |
6368 | * for each core we will add once | |
6369 | * to the group in physical domain | |
6370 | */ | |
6371 | if (j != first_cpu(sd1->groups->cpumask)) | |
6372 | continue; | |
6373 | ||
6374 | if (sched_smt_power_savings) | |
6375 | sd->groups->cpu_power += sd1->groups->cpu_power; | |
6376 | else | |
6377 | sd->groups->cpu_power += SCHED_LOAD_SCALE; | |
6378 | } | |
6379 | } else | |
6380 | /* | |
6381 | * This has to be < 2 * SCHED_LOAD_SCALE | |
6382 | * Lets keep it SCHED_LOAD_SCALE, so that | |
6383 | * while calculating NUMA group's cpu_power | |
6384 | * we can simply do | |
6385 | * numa_group->cpu_power += phys_group->cpu_power; | |
6386 | * | |
6387 | * See "only add power once for each physical pkg" | |
6388 | * comment below | |
6389 | */ | |
6390 | sd->groups->cpu_power = SCHED_LOAD_SCALE; | |
1e9f28fa | 6391 | #else |
5c45bf27 | 6392 | int power; |
1da177e4 | 6393 | sd = &per_cpu(phys_domains, i); |
5c45bf27 SS |
6394 | if (sched_smt_power_savings) |
6395 | power = SCHED_LOAD_SCALE * cpus_weight(sd->groups->cpumask); | |
6396 | else | |
6397 | power = SCHED_LOAD_SCALE; | |
1da177e4 | 6398 | sd->groups->cpu_power = power; |
1e9f28fa | 6399 | #endif |
1da177e4 LT |
6400 | } |
6401 | ||
9c1cfda2 | 6402 | #ifdef CONFIG_NUMA |
08069033 SS |
6403 | for (i = 0; i < MAX_NUMNODES; i++) |
6404 | init_numa_sched_groups_power(sched_group_nodes[i]); | |
9c1cfda2 | 6405 | |
08069033 | 6406 | init_numa_sched_groups_power(sched_group_allnodes); |
9c1cfda2 JH |
6407 | #endif |
6408 | ||
1da177e4 | 6409 | /* Attach the domains */ |
1a20ff27 | 6410 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 LT |
6411 | struct sched_domain *sd; |
6412 | #ifdef CONFIG_SCHED_SMT | |
6413 | sd = &per_cpu(cpu_domains, i); | |
1e9f28fa SS |
6414 | #elif defined(CONFIG_SCHED_MC) |
6415 | sd = &per_cpu(core_domains, i); | |
1da177e4 LT |
6416 | #else |
6417 | sd = &per_cpu(phys_domains, i); | |
6418 | #endif | |
6419 | cpu_attach_domain(sd, i); | |
6420 | } | |
198e2f18 | 6421 | /* |
6422 | * Tune cache-hot values: | |
6423 | */ | |
6424 | calibrate_migration_costs(cpu_map); | |
51888ca2 SV |
6425 | |
6426 | return 0; | |
6427 | ||
51888ca2 SV |
6428 | error: |
6429 | free_sched_groups(cpu_map); | |
6430 | return -ENOMEM; | |
1da177e4 | 6431 | } |
1a20ff27 DG |
6432 | /* |
6433 | * Set up scheduler domains and groups. Callers must hold the hotplug lock. | |
6434 | */ | |
51888ca2 | 6435 | static int arch_init_sched_domains(const cpumask_t *cpu_map) |
1a20ff27 DG |
6436 | { |
6437 | cpumask_t cpu_default_map; | |
51888ca2 | 6438 | int err; |
1da177e4 | 6439 | |
1a20ff27 DG |
6440 | /* |
6441 | * Setup mask for cpus without special case scheduling requirements. | |
6442 | * For now this just excludes isolated cpus, but could be used to | |
6443 | * exclude other special cases in the future. | |
6444 | */ | |
6445 | cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map); | |
6446 | ||
51888ca2 SV |
6447 | err = build_sched_domains(&cpu_default_map); |
6448 | ||
6449 | return err; | |
1a20ff27 DG |
6450 | } |
6451 | ||
6452 | static void arch_destroy_sched_domains(const cpumask_t *cpu_map) | |
1da177e4 | 6453 | { |
51888ca2 | 6454 | free_sched_groups(cpu_map); |
9c1cfda2 | 6455 | } |
1da177e4 | 6456 | |
1a20ff27 DG |
6457 | /* |
6458 | * Detach sched domains from a group of cpus specified in cpu_map | |
6459 | * These cpus will now be attached to the NULL domain | |
6460 | */ | |
858119e1 | 6461 | static void detach_destroy_domains(const cpumask_t *cpu_map) |
1a20ff27 DG |
6462 | { |
6463 | int i; | |
6464 | ||
6465 | for_each_cpu_mask(i, *cpu_map) | |
6466 | cpu_attach_domain(NULL, i); | |
6467 | synchronize_sched(); | |
6468 | arch_destroy_sched_domains(cpu_map); | |
6469 | } | |
6470 | ||
6471 | /* | |
6472 | * Partition sched domains as specified by the cpumasks below. | |
6473 | * This attaches all cpus from the cpumasks to the NULL domain, | |
6474 | * waits for a RCU quiescent period, recalculates sched | |
6475 | * domain information and then attaches them back to the | |
6476 | * correct sched domains | |
6477 | * Call with hotplug lock held | |
6478 | */ | |
51888ca2 | 6479 | int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2) |
1a20ff27 DG |
6480 | { |
6481 | cpumask_t change_map; | |
51888ca2 | 6482 | int err = 0; |
1a20ff27 DG |
6483 | |
6484 | cpus_and(*partition1, *partition1, cpu_online_map); | |
6485 | cpus_and(*partition2, *partition2, cpu_online_map); | |
6486 | cpus_or(change_map, *partition1, *partition2); | |
6487 | ||
6488 | /* Detach sched domains from all of the affected cpus */ | |
6489 | detach_destroy_domains(&change_map); | |
6490 | if (!cpus_empty(*partition1)) | |
51888ca2 SV |
6491 | err = build_sched_domains(partition1); |
6492 | if (!err && !cpus_empty(*partition2)) | |
6493 | err = build_sched_domains(partition2); | |
6494 | ||
6495 | return err; | |
1a20ff27 DG |
6496 | } |
6497 | ||
5c45bf27 SS |
6498 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
6499 | int arch_reinit_sched_domains(void) | |
6500 | { | |
6501 | int err; | |
6502 | ||
6503 | lock_cpu_hotplug(); | |
6504 | detach_destroy_domains(&cpu_online_map); | |
6505 | err = arch_init_sched_domains(&cpu_online_map); | |
6506 | unlock_cpu_hotplug(); | |
6507 | ||
6508 | return err; | |
6509 | } | |
6510 | ||
6511 | static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt) | |
6512 | { | |
6513 | int ret; | |
6514 | ||
6515 | if (buf[0] != '0' && buf[0] != '1') | |
6516 | return -EINVAL; | |
6517 | ||
6518 | if (smt) | |
6519 | sched_smt_power_savings = (buf[0] == '1'); | |
6520 | else | |
6521 | sched_mc_power_savings = (buf[0] == '1'); | |
6522 | ||
6523 | ret = arch_reinit_sched_domains(); | |
6524 | ||
6525 | return ret ? ret : count; | |
6526 | } | |
6527 | ||
6528 | int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls) | |
6529 | { | |
6530 | int err = 0; | |
6531 | #ifdef CONFIG_SCHED_SMT | |
6532 | if (smt_capable()) | |
6533 | err = sysfs_create_file(&cls->kset.kobj, | |
6534 | &attr_sched_smt_power_savings.attr); | |
6535 | #endif | |
6536 | #ifdef CONFIG_SCHED_MC | |
6537 | if (!err && mc_capable()) | |
6538 | err = sysfs_create_file(&cls->kset.kobj, | |
6539 | &attr_sched_mc_power_savings.attr); | |
6540 | #endif | |
6541 | return err; | |
6542 | } | |
6543 | #endif | |
6544 | ||
6545 | #ifdef CONFIG_SCHED_MC | |
6546 | static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page) | |
6547 | { | |
6548 | return sprintf(page, "%u\n", sched_mc_power_savings); | |
6549 | } | |
6550 | static ssize_t sched_mc_power_savings_store(struct sys_device *dev, const char *buf, size_t count) | |
6551 | { | |
6552 | return sched_power_savings_store(buf, count, 0); | |
6553 | } | |
6554 | SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show, | |
6555 | sched_mc_power_savings_store); | |
6556 | #endif | |
6557 | ||
6558 | #ifdef CONFIG_SCHED_SMT | |
6559 | static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page) | |
6560 | { | |
6561 | return sprintf(page, "%u\n", sched_smt_power_savings); | |
6562 | } | |
6563 | static ssize_t sched_smt_power_savings_store(struct sys_device *dev, const char *buf, size_t count) | |
6564 | { | |
6565 | return sched_power_savings_store(buf, count, 1); | |
6566 | } | |
6567 | SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show, | |
6568 | sched_smt_power_savings_store); | |
6569 | #endif | |
6570 | ||
6571 | ||
1da177e4 LT |
6572 | #ifdef CONFIG_HOTPLUG_CPU |
6573 | /* | |
6574 | * Force a reinitialization of the sched domains hierarchy. The domains | |
6575 | * and groups cannot be updated in place without racing with the balancing | |
41c7ce9a | 6576 | * code, so we temporarily attach all running cpus to the NULL domain |
1da177e4 LT |
6577 | * which will prevent rebalancing while the sched domains are recalculated. |
6578 | */ | |
6579 | static int update_sched_domains(struct notifier_block *nfb, | |
6580 | unsigned long action, void *hcpu) | |
6581 | { | |
1da177e4 LT |
6582 | switch (action) { |
6583 | case CPU_UP_PREPARE: | |
6584 | case CPU_DOWN_PREPARE: | |
1a20ff27 | 6585 | detach_destroy_domains(&cpu_online_map); |
1da177e4 LT |
6586 | return NOTIFY_OK; |
6587 | ||
6588 | case CPU_UP_CANCELED: | |
6589 | case CPU_DOWN_FAILED: | |
6590 | case CPU_ONLINE: | |
6591 | case CPU_DEAD: | |
6592 | /* | |
6593 | * Fall through and re-initialise the domains. | |
6594 | */ | |
6595 | break; | |
6596 | default: | |
6597 | return NOTIFY_DONE; | |
6598 | } | |
6599 | ||
6600 | /* The hotplug lock is already held by cpu_up/cpu_down */ | |
1a20ff27 | 6601 | arch_init_sched_domains(&cpu_online_map); |
1da177e4 LT |
6602 | |
6603 | return NOTIFY_OK; | |
6604 | } | |
6605 | #endif | |
6606 | ||
6607 | void __init sched_init_smp(void) | |
6608 | { | |
6609 | lock_cpu_hotplug(); | |
1a20ff27 | 6610 | arch_init_sched_domains(&cpu_online_map); |
1da177e4 LT |
6611 | unlock_cpu_hotplug(); |
6612 | /* XXX: Theoretical race here - CPU may be hotplugged now */ | |
6613 | hotcpu_notifier(update_sched_domains, 0); | |
6614 | } | |
6615 | #else | |
6616 | void __init sched_init_smp(void) | |
6617 | { | |
6618 | } | |
6619 | #endif /* CONFIG_SMP */ | |
6620 | ||
6621 | int in_sched_functions(unsigned long addr) | |
6622 | { | |
6623 | /* Linker adds these: start and end of __sched functions */ | |
6624 | extern char __sched_text_start[], __sched_text_end[]; | |
6625 | return in_lock_functions(addr) || | |
6626 | (addr >= (unsigned long)__sched_text_start | |
6627 | && addr < (unsigned long)__sched_text_end); | |
6628 | } | |
6629 | ||
6630 | void __init sched_init(void) | |
6631 | { | |
6632 | runqueue_t *rq; | |
6633 | int i, j, k; | |
6634 | ||
0a945022 | 6635 | for_each_possible_cpu(i) { |
1da177e4 LT |
6636 | prio_array_t *array; |
6637 | ||
6638 | rq = cpu_rq(i); | |
6639 | spin_lock_init(&rq->lock); | |
7897986b | 6640 | rq->nr_running = 0; |
1da177e4 LT |
6641 | rq->active = rq->arrays; |
6642 | rq->expired = rq->arrays + 1; | |
6643 | rq->best_expired_prio = MAX_PRIO; | |
6644 | ||
6645 | #ifdef CONFIG_SMP | |
41c7ce9a | 6646 | rq->sd = NULL; |
7897986b NP |
6647 | for (j = 1; j < 3; j++) |
6648 | rq->cpu_load[j] = 0; | |
1da177e4 LT |
6649 | rq->active_balance = 0; |
6650 | rq->push_cpu = 0; | |
6651 | rq->migration_thread = NULL; | |
6652 | INIT_LIST_HEAD(&rq->migration_queue); | |
6653 | #endif | |
6654 | atomic_set(&rq->nr_iowait, 0); | |
6655 | ||
6656 | for (j = 0; j < 2; j++) { | |
6657 | array = rq->arrays + j; | |
6658 | for (k = 0; k < MAX_PRIO; k++) { | |
6659 | INIT_LIST_HEAD(array->queue + k); | |
6660 | __clear_bit(k, array->bitmap); | |
6661 | } | |
6662 | // delimiter for bitsearch | |
6663 | __set_bit(MAX_PRIO, array->bitmap); | |
6664 | } | |
6665 | } | |
6666 | ||
2dd73a4f | 6667 | set_load_weight(&init_task); |
1da177e4 LT |
6668 | /* |
6669 | * The boot idle thread does lazy MMU switching as well: | |
6670 | */ | |
6671 | atomic_inc(&init_mm.mm_count); | |
6672 | enter_lazy_tlb(&init_mm, current); | |
6673 | ||
6674 | /* | |
6675 | * Make us the idle thread. Technically, schedule() should not be | |
6676 | * called from this thread, however somewhere below it might be, | |
6677 | * but because we are the idle thread, we just pick up running again | |
6678 | * when this runqueue becomes "idle". | |
6679 | */ | |
6680 | init_idle(current, smp_processor_id()); | |
6681 | } | |
6682 | ||
6683 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP | |
6684 | void __might_sleep(char *file, int line) | |
6685 | { | |
6686 | #if defined(in_atomic) | |
6687 | static unsigned long prev_jiffy; /* ratelimiting */ | |
6688 | ||
6689 | if ((in_atomic() || irqs_disabled()) && | |
6690 | system_state == SYSTEM_RUNNING && !oops_in_progress) { | |
6691 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) | |
6692 | return; | |
6693 | prev_jiffy = jiffies; | |
91368d73 | 6694 | printk(KERN_ERR "BUG: sleeping function called from invalid" |
1da177e4 LT |
6695 | " context at %s:%d\n", file, line); |
6696 | printk("in_atomic():%d, irqs_disabled():%d\n", | |
6697 | in_atomic(), irqs_disabled()); | |
6698 | dump_stack(); | |
6699 | } | |
6700 | #endif | |
6701 | } | |
6702 | EXPORT_SYMBOL(__might_sleep); | |
6703 | #endif | |
6704 | ||
6705 | #ifdef CONFIG_MAGIC_SYSRQ | |
6706 | void normalize_rt_tasks(void) | |
6707 | { | |
6708 | struct task_struct *p; | |
6709 | prio_array_t *array; | |
6710 | unsigned long flags; | |
6711 | runqueue_t *rq; | |
6712 | ||
6713 | read_lock_irq(&tasklist_lock); | |
c96d145e | 6714 | for_each_process(p) { |
1da177e4 LT |
6715 | if (!rt_task(p)) |
6716 | continue; | |
6717 | ||
b29739f9 IM |
6718 | spin_lock_irqsave(&p->pi_lock, flags); |
6719 | rq = __task_rq_lock(p); | |
1da177e4 LT |
6720 | |
6721 | array = p->array; | |
6722 | if (array) | |
6723 | deactivate_task(p, task_rq(p)); | |
6724 | __setscheduler(p, SCHED_NORMAL, 0); | |
6725 | if (array) { | |
6726 | __activate_task(p, task_rq(p)); | |
6727 | resched_task(rq->curr); | |
6728 | } | |
6729 | ||
b29739f9 IM |
6730 | __task_rq_unlock(rq); |
6731 | spin_unlock_irqrestore(&p->pi_lock, flags); | |
1da177e4 LT |
6732 | } |
6733 | read_unlock_irq(&tasklist_lock); | |
6734 | } | |
6735 | ||
6736 | #endif /* CONFIG_MAGIC_SYSRQ */ | |
1df5c10a LT |
6737 | |
6738 | #ifdef CONFIG_IA64 | |
6739 | /* | |
6740 | * These functions are only useful for the IA64 MCA handling. | |
6741 | * | |
6742 | * They can only be called when the whole system has been | |
6743 | * stopped - every CPU needs to be quiescent, and no scheduling | |
6744 | * activity can take place. Using them for anything else would | |
6745 | * be a serious bug, and as a result, they aren't even visible | |
6746 | * under any other configuration. | |
6747 | */ | |
6748 | ||
6749 | /** | |
6750 | * curr_task - return the current task for a given cpu. | |
6751 | * @cpu: the processor in question. | |
6752 | * | |
6753 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
6754 | */ | |
6755 | task_t *curr_task(int cpu) | |
6756 | { | |
6757 | return cpu_curr(cpu); | |
6758 | } | |
6759 | ||
6760 | /** | |
6761 | * set_curr_task - set the current task for a given cpu. | |
6762 | * @cpu: the processor in question. | |
6763 | * @p: the task pointer to set. | |
6764 | * | |
6765 | * Description: This function must only be used when non-maskable interrupts | |
6766 | * are serviced on a separate stack. It allows the architecture to switch the | |
6767 | * notion of the current task on a cpu in a non-blocking manner. This function | |
6768 | * must be called with all CPU's synchronized, and interrupts disabled, the | |
6769 | * and caller must save the original value of the current task (see | |
6770 | * curr_task() above) and restore that value before reenabling interrupts and | |
6771 | * re-starting the system. | |
6772 | * | |
6773 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
6774 | */ | |
6775 | void set_curr_task(int cpu, task_t *p) | |
6776 | { | |
6777 | cpu_curr(cpu) = p; | |
6778 | } | |
6779 | ||
6780 | #endif |