]>
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> | |
9a11b49a | 33 | #include <linux/debug_locks.h> |
1da177e4 LT |
34 | #include <linux/security.h> |
35 | #include <linux/notifier.h> | |
36 | #include <linux/profile.h> | |
37 | #include <linux/suspend.h> | |
198e2f18 | 38 | #include <linux/vmalloc.h> |
1da177e4 LT |
39 | #include <linux/blkdev.h> |
40 | #include <linux/delay.h> | |
41 | #include <linux/smp.h> | |
42 | #include <linux/threads.h> | |
43 | #include <linux/timer.h> | |
44 | #include <linux/rcupdate.h> | |
45 | #include <linux/cpu.h> | |
46 | #include <linux/cpuset.h> | |
47 | #include <linux/percpu.h> | |
48 | #include <linux/kthread.h> | |
49 | #include <linux/seq_file.h> | |
50 | #include <linux/syscalls.h> | |
51 | #include <linux/times.h> | |
52 | #include <linux/acct.h> | |
c6fd91f0 | 53 | #include <linux/kprobes.h> |
1da177e4 LT |
54 | #include <asm/tlb.h> |
55 | ||
56 | #include <asm/unistd.h> | |
57 | ||
58 | /* | |
59 | * Convert user-nice values [ -20 ... 0 ... 19 ] | |
60 | * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ], | |
61 | * and back. | |
62 | */ | |
63 | #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20) | |
64 | #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20) | |
65 | #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio) | |
66 | ||
67 | /* | |
68 | * 'User priority' is the nice value converted to something we | |
69 | * can work with better when scaling various scheduler parameters, | |
70 | * it's a [ 0 ... 39 ] range. | |
71 | */ | |
72 | #define USER_PRIO(p) ((p)-MAX_RT_PRIO) | |
73 | #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio) | |
74 | #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO)) | |
75 | ||
76 | /* | |
77 | * Some helpers for converting nanosecond timing to jiffy resolution | |
78 | */ | |
79 | #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ)) | |
80 | #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ)) | |
81 | ||
82 | /* | |
83 | * These are the 'tuning knobs' of the scheduler: | |
84 | * | |
85 | * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger), | |
86 | * default timeslice is 100 msecs, maximum timeslice is 800 msecs. | |
87 | * Timeslices get refilled after they expire. | |
88 | */ | |
89 | #define MIN_TIMESLICE max(5 * HZ / 1000, 1) | |
90 | #define DEF_TIMESLICE (100 * HZ / 1000) | |
91 | #define ON_RUNQUEUE_WEIGHT 30 | |
92 | #define CHILD_PENALTY 95 | |
93 | #define PARENT_PENALTY 100 | |
94 | #define EXIT_WEIGHT 3 | |
95 | #define PRIO_BONUS_RATIO 25 | |
96 | #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100) | |
97 | #define INTERACTIVE_DELTA 2 | |
98 | #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS) | |
99 | #define STARVATION_LIMIT (MAX_SLEEP_AVG) | |
100 | #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG)) | |
101 | ||
102 | /* | |
103 | * If a task is 'interactive' then we reinsert it in the active | |
104 | * array after it has expired its current timeslice. (it will not | |
105 | * continue to run immediately, it will still roundrobin with | |
106 | * other interactive tasks.) | |
107 | * | |
108 | * This part scales the interactivity limit depending on niceness. | |
109 | * | |
110 | * We scale it linearly, offset by the INTERACTIVE_DELTA delta. | |
111 | * Here are a few examples of different nice levels: | |
112 | * | |
113 | * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0] | |
114 | * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0] | |
115 | * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0] | |
116 | * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0] | |
117 | * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0] | |
118 | * | |
119 | * (the X axis represents the possible -5 ... 0 ... +5 dynamic | |
120 | * priority range a task can explore, a value of '1' means the | |
121 | * task is rated interactive.) | |
122 | * | |
123 | * Ie. nice +19 tasks can never get 'interactive' enough to be | |
124 | * reinserted into the active array. And only heavily CPU-hog nice -20 | |
125 | * tasks will be expired. Default nice 0 tasks are somewhere between, | |
126 | * it takes some effort for them to get interactive, but it's not | |
127 | * too hard. | |
128 | */ | |
129 | ||
130 | #define CURRENT_BONUS(p) \ | |
131 | (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \ | |
132 | MAX_SLEEP_AVG) | |
133 | ||
134 | #define GRANULARITY (10 * HZ / 1000 ? : 1) | |
135 | ||
136 | #ifdef CONFIG_SMP | |
137 | #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \ | |
138 | (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \ | |
139 | num_online_cpus()) | |
140 | #else | |
141 | #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \ | |
142 | (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1))) | |
143 | #endif | |
144 | ||
145 | #define SCALE(v1,v1_max,v2_max) \ | |
146 | (v1) * (v2_max) / (v1_max) | |
147 | ||
148 | #define DELTA(p) \ | |
013d3868 MA |
149 | (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \ |
150 | INTERACTIVE_DELTA) | |
1da177e4 LT |
151 | |
152 | #define TASK_INTERACTIVE(p) \ | |
153 | ((p)->prio <= (p)->static_prio - DELTA(p)) | |
154 | ||
155 | #define INTERACTIVE_SLEEP(p) \ | |
156 | (JIFFIES_TO_NS(MAX_SLEEP_AVG * \ | |
157 | (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1)) | |
158 | ||
159 | #define TASK_PREEMPTS_CURR(p, rq) \ | |
160 | ((p)->prio < (rq)->curr->prio) | |
161 | ||
162 | /* | |
163 | * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ] | |
164 | * to time slice values: [800ms ... 100ms ... 5ms] | |
165 | * | |
166 | * The higher a thread's priority, the bigger timeslices | |
167 | * it gets during one round of execution. But even the lowest | |
168 | * priority thread gets MIN_TIMESLICE worth of execution time. | |
169 | */ | |
170 | ||
171 | #define SCALE_PRIO(x, prio) \ | |
2dd73a4f | 172 | max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE) |
1da177e4 | 173 | |
2dd73a4f | 174 | static unsigned int static_prio_timeslice(int static_prio) |
1da177e4 | 175 | { |
2dd73a4f PW |
176 | if (static_prio < NICE_TO_PRIO(0)) |
177 | return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio); | |
1da177e4 | 178 | else |
2dd73a4f | 179 | return SCALE_PRIO(DEF_TIMESLICE, static_prio); |
1da177e4 | 180 | } |
2dd73a4f PW |
181 | |
182 | static inline unsigned int task_timeslice(task_t *p) | |
183 | { | |
184 | return static_prio_timeslice(p->static_prio); | |
185 | } | |
186 | ||
1da177e4 LT |
187 | #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran) \ |
188 | < (long long) (sd)->cache_hot_time) | |
189 | ||
190 | /* | |
191 | * These are the runqueue data structures: | |
192 | */ | |
193 | ||
1da177e4 LT |
194 | typedef struct runqueue runqueue_t; |
195 | ||
196 | struct prio_array { | |
197 | unsigned int nr_active; | |
d444886e | 198 | DECLARE_BITMAP(bitmap, MAX_PRIO+1); /* include 1 bit for delimiter */ |
1da177e4 LT |
199 | struct list_head queue[MAX_PRIO]; |
200 | }; | |
201 | ||
202 | /* | |
203 | * This is the main, per-CPU runqueue data structure. | |
204 | * | |
205 | * Locking rule: those places that want to lock multiple runqueues | |
206 | * (such as the load balancing or the thread migration code), lock | |
207 | * acquire operations must be ordered by ascending &runqueue. | |
208 | */ | |
209 | struct runqueue { | |
210 | spinlock_t lock; | |
211 | ||
212 | /* | |
213 | * nr_running and cpu_load should be in the same cacheline because | |
214 | * remote CPUs use both these fields when doing load calculation. | |
215 | */ | |
216 | unsigned long nr_running; | |
2dd73a4f | 217 | unsigned long raw_weighted_load; |
1da177e4 | 218 | #ifdef CONFIG_SMP |
7897986b | 219 | unsigned long cpu_load[3]; |
1da177e4 LT |
220 | #endif |
221 | unsigned long long nr_switches; | |
222 | ||
223 | /* | |
224 | * This is part of a global counter where only the total sum | |
225 | * over all CPUs matters. A task can increase this counter on | |
226 | * one CPU and if it got migrated afterwards it may decrease | |
227 | * it on another CPU. Always updated under the runqueue lock: | |
228 | */ | |
229 | unsigned long nr_uninterruptible; | |
230 | ||
231 | unsigned long expired_timestamp; | |
232 | unsigned long long timestamp_last_tick; | |
233 | task_t *curr, *idle; | |
234 | struct mm_struct *prev_mm; | |
235 | prio_array_t *active, *expired, arrays[2]; | |
236 | int best_expired_prio; | |
237 | atomic_t nr_iowait; | |
238 | ||
239 | #ifdef CONFIG_SMP | |
240 | struct sched_domain *sd; | |
241 | ||
242 | /* For active balancing */ | |
243 | int active_balance; | |
244 | int push_cpu; | |
245 | ||
246 | task_t *migration_thread; | |
247 | struct list_head migration_queue; | |
248 | #endif | |
249 | ||
250 | #ifdef CONFIG_SCHEDSTATS | |
251 | /* latency stats */ | |
252 | struct sched_info rq_sched_info; | |
253 | ||
254 | /* sys_sched_yield() stats */ | |
255 | unsigned long yld_exp_empty; | |
256 | unsigned long yld_act_empty; | |
257 | unsigned long yld_both_empty; | |
258 | unsigned long yld_cnt; | |
259 | ||
260 | /* schedule() stats */ | |
261 | unsigned long sched_switch; | |
262 | unsigned long sched_cnt; | |
263 | unsigned long sched_goidle; | |
264 | ||
265 | /* try_to_wake_up() stats */ | |
266 | unsigned long ttwu_cnt; | |
267 | unsigned long ttwu_local; | |
268 | #endif | |
269 | }; | |
270 | ||
271 | static DEFINE_PER_CPU(struct runqueue, runqueues); | |
272 | ||
674311d5 NP |
273 | /* |
274 | * The domain tree (rq->sd) is protected by RCU's quiescent state transition. | |
1a20ff27 | 275 | * See detach_destroy_domains: synchronize_sched for details. |
674311d5 NP |
276 | * |
277 | * The domain tree of any CPU may only be accessed from within | |
278 | * preempt-disabled sections. | |
279 | */ | |
1da177e4 | 280 | #define for_each_domain(cpu, domain) \ |
674311d5 | 281 | for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent) |
1da177e4 LT |
282 | |
283 | #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) | |
284 | #define this_rq() (&__get_cpu_var(runqueues)) | |
285 | #define task_rq(p) cpu_rq(task_cpu(p)) | |
286 | #define cpu_curr(cpu) (cpu_rq(cpu)->curr) | |
287 | ||
1da177e4 | 288 | #ifndef prepare_arch_switch |
4866cde0 NP |
289 | # define prepare_arch_switch(next) do { } while (0) |
290 | #endif | |
291 | #ifndef finish_arch_switch | |
292 | # define finish_arch_switch(prev) do { } while (0) | |
293 | #endif | |
294 | ||
295 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW | |
296 | static inline int task_running(runqueue_t *rq, task_t *p) | |
297 | { | |
298 | return rq->curr == p; | |
299 | } | |
300 | ||
301 | static inline void prepare_lock_switch(runqueue_t *rq, task_t *next) | |
302 | { | |
303 | } | |
304 | ||
305 | static inline void finish_lock_switch(runqueue_t *rq, task_t *prev) | |
306 | { | |
da04c035 IM |
307 | #ifdef CONFIG_DEBUG_SPINLOCK |
308 | /* this is a valid case when another task releases the spinlock */ | |
309 | rq->lock.owner = current; | |
310 | #endif | |
4866cde0 NP |
311 | spin_unlock_irq(&rq->lock); |
312 | } | |
313 | ||
314 | #else /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
315 | static inline int task_running(runqueue_t *rq, task_t *p) | |
316 | { | |
317 | #ifdef CONFIG_SMP | |
318 | return p->oncpu; | |
319 | #else | |
320 | return rq->curr == p; | |
321 | #endif | |
322 | } | |
323 | ||
324 | static inline void prepare_lock_switch(runqueue_t *rq, task_t *next) | |
325 | { | |
326 | #ifdef CONFIG_SMP | |
327 | /* | |
328 | * We can optimise this out completely for !SMP, because the | |
329 | * SMP rebalancing from interrupt is the only thing that cares | |
330 | * here. | |
331 | */ | |
332 | next->oncpu = 1; | |
333 | #endif | |
334 | #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
335 | spin_unlock_irq(&rq->lock); | |
336 | #else | |
337 | spin_unlock(&rq->lock); | |
338 | #endif | |
339 | } | |
340 | ||
341 | static inline void finish_lock_switch(runqueue_t *rq, task_t *prev) | |
342 | { | |
343 | #ifdef CONFIG_SMP | |
344 | /* | |
345 | * After ->oncpu is cleared, the task can be moved to a different CPU. | |
346 | * We must ensure this doesn't happen until the switch is completely | |
347 | * finished. | |
348 | */ | |
349 | smp_wmb(); | |
350 | prev->oncpu = 0; | |
351 | #endif | |
352 | #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
353 | local_irq_enable(); | |
1da177e4 | 354 | #endif |
4866cde0 NP |
355 | } |
356 | #endif /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
1da177e4 | 357 | |
b29739f9 IM |
358 | /* |
359 | * __task_rq_lock - lock the runqueue a given task resides on. | |
360 | * Must be called interrupts disabled. | |
361 | */ | |
362 | static inline runqueue_t *__task_rq_lock(task_t *p) | |
363 | __acquires(rq->lock) | |
364 | { | |
365 | struct runqueue *rq; | |
366 | ||
367 | repeat_lock_task: | |
368 | rq = task_rq(p); | |
369 | spin_lock(&rq->lock); | |
370 | if (unlikely(rq != task_rq(p))) { | |
371 | spin_unlock(&rq->lock); | |
372 | goto repeat_lock_task; | |
373 | } | |
374 | return rq; | |
375 | } | |
376 | ||
1da177e4 LT |
377 | /* |
378 | * task_rq_lock - lock the runqueue a given task resides on and disable | |
379 | * interrupts. Note the ordering: we can safely lookup the task_rq without | |
380 | * explicitly disabling preemption. | |
381 | */ | |
9fea80e4 | 382 | static runqueue_t *task_rq_lock(task_t *p, unsigned long *flags) |
1da177e4 LT |
383 | __acquires(rq->lock) |
384 | { | |
385 | struct runqueue *rq; | |
386 | ||
387 | repeat_lock_task: | |
388 | local_irq_save(*flags); | |
389 | rq = task_rq(p); | |
390 | spin_lock(&rq->lock); | |
391 | if (unlikely(rq != task_rq(p))) { | |
392 | spin_unlock_irqrestore(&rq->lock, *flags); | |
393 | goto repeat_lock_task; | |
394 | } | |
395 | return rq; | |
396 | } | |
397 | ||
b29739f9 IM |
398 | static inline void __task_rq_unlock(runqueue_t *rq) |
399 | __releases(rq->lock) | |
400 | { | |
401 | spin_unlock(&rq->lock); | |
402 | } | |
403 | ||
1da177e4 LT |
404 | static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags) |
405 | __releases(rq->lock) | |
406 | { | |
407 | spin_unlock_irqrestore(&rq->lock, *flags); | |
408 | } | |
409 | ||
410 | #ifdef CONFIG_SCHEDSTATS | |
411 | /* | |
412 | * bump this up when changing the output format or the meaning of an existing | |
413 | * format, so that tools can adapt (or abort) | |
414 | */ | |
68767a0a | 415 | #define SCHEDSTAT_VERSION 12 |
1da177e4 LT |
416 | |
417 | static int show_schedstat(struct seq_file *seq, void *v) | |
418 | { | |
419 | int cpu; | |
420 | ||
421 | seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION); | |
422 | seq_printf(seq, "timestamp %lu\n", jiffies); | |
423 | for_each_online_cpu(cpu) { | |
424 | runqueue_t *rq = cpu_rq(cpu); | |
425 | #ifdef CONFIG_SMP | |
426 | struct sched_domain *sd; | |
427 | int dcnt = 0; | |
428 | #endif | |
429 | ||
430 | /* runqueue-specific stats */ | |
431 | seq_printf(seq, | |
432 | "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu", | |
433 | cpu, rq->yld_both_empty, | |
434 | rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt, | |
435 | rq->sched_switch, rq->sched_cnt, rq->sched_goidle, | |
436 | rq->ttwu_cnt, rq->ttwu_local, | |
437 | rq->rq_sched_info.cpu_time, | |
438 | rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt); | |
439 | ||
440 | seq_printf(seq, "\n"); | |
441 | ||
442 | #ifdef CONFIG_SMP | |
443 | /* domain-specific stats */ | |
674311d5 | 444 | preempt_disable(); |
1da177e4 LT |
445 | for_each_domain(cpu, sd) { |
446 | enum idle_type itype; | |
447 | char mask_str[NR_CPUS]; | |
448 | ||
449 | cpumask_scnprintf(mask_str, NR_CPUS, sd->span); | |
450 | seq_printf(seq, "domain%d %s", dcnt++, mask_str); | |
451 | for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES; | |
452 | itype++) { | |
453 | seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu", | |
454 | sd->lb_cnt[itype], | |
455 | sd->lb_balanced[itype], | |
456 | sd->lb_failed[itype], | |
457 | sd->lb_imbalance[itype], | |
458 | sd->lb_gained[itype], | |
459 | sd->lb_hot_gained[itype], | |
460 | sd->lb_nobusyq[itype], | |
461 | sd->lb_nobusyg[itype]); | |
462 | } | |
68767a0a | 463 | seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n", |
1da177e4 | 464 | sd->alb_cnt, sd->alb_failed, sd->alb_pushed, |
68767a0a NP |
465 | sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed, |
466 | sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed, | |
1da177e4 LT |
467 | sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance); |
468 | } | |
674311d5 | 469 | preempt_enable(); |
1da177e4 LT |
470 | #endif |
471 | } | |
472 | return 0; | |
473 | } | |
474 | ||
475 | static int schedstat_open(struct inode *inode, struct file *file) | |
476 | { | |
477 | unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32); | |
478 | char *buf = kmalloc(size, GFP_KERNEL); | |
479 | struct seq_file *m; | |
480 | int res; | |
481 | ||
482 | if (!buf) | |
483 | return -ENOMEM; | |
484 | res = single_open(file, show_schedstat, NULL); | |
485 | if (!res) { | |
486 | m = file->private_data; | |
487 | m->buf = buf; | |
488 | m->size = size; | |
489 | } else | |
490 | kfree(buf); | |
491 | return res; | |
492 | } | |
493 | ||
494 | struct file_operations proc_schedstat_operations = { | |
495 | .open = schedstat_open, | |
496 | .read = seq_read, | |
497 | .llseek = seq_lseek, | |
498 | .release = single_release, | |
499 | }; | |
500 | ||
501 | # define schedstat_inc(rq, field) do { (rq)->field++; } while (0) | |
502 | # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0) | |
503 | #else /* !CONFIG_SCHEDSTATS */ | |
504 | # define schedstat_inc(rq, field) do { } while (0) | |
505 | # define schedstat_add(rq, field, amt) do { } while (0) | |
506 | #endif | |
507 | ||
508 | /* | |
509 | * rq_lock - lock a given runqueue and disable interrupts. | |
510 | */ | |
511 | static inline runqueue_t *this_rq_lock(void) | |
512 | __acquires(rq->lock) | |
513 | { | |
514 | runqueue_t *rq; | |
515 | ||
516 | local_irq_disable(); | |
517 | rq = this_rq(); | |
518 | spin_lock(&rq->lock); | |
519 | ||
520 | return rq; | |
521 | } | |
522 | ||
1da177e4 LT |
523 | #ifdef CONFIG_SCHEDSTATS |
524 | /* | |
525 | * Called when a process is dequeued from the active array and given | |
526 | * the cpu. We should note that with the exception of interactive | |
527 | * tasks, the expired queue will become the active queue after the active | |
528 | * queue is empty, without explicitly dequeuing and requeuing tasks in the | |
529 | * expired queue. (Interactive tasks may be requeued directly to the | |
530 | * active queue, thus delaying tasks in the expired queue from running; | |
531 | * see scheduler_tick()). | |
532 | * | |
533 | * This function is only called from sched_info_arrive(), rather than | |
534 | * dequeue_task(). Even though a task may be queued and dequeued multiple | |
535 | * times as it is shuffled about, we're really interested in knowing how | |
536 | * long it was from the *first* time it was queued to the time that it | |
537 | * finally hit a cpu. | |
538 | */ | |
539 | static inline void sched_info_dequeued(task_t *t) | |
540 | { | |
541 | t->sched_info.last_queued = 0; | |
542 | } | |
543 | ||
544 | /* | |
545 | * Called when a task finally hits the cpu. We can now calculate how | |
546 | * long it was waiting to run. We also note when it began so that we | |
547 | * can keep stats on how long its timeslice is. | |
548 | */ | |
858119e1 | 549 | static void sched_info_arrive(task_t *t) |
1da177e4 LT |
550 | { |
551 | unsigned long now = jiffies, diff = 0; | |
552 | struct runqueue *rq = task_rq(t); | |
553 | ||
554 | if (t->sched_info.last_queued) | |
555 | diff = now - t->sched_info.last_queued; | |
556 | sched_info_dequeued(t); | |
557 | t->sched_info.run_delay += diff; | |
558 | t->sched_info.last_arrival = now; | |
559 | t->sched_info.pcnt++; | |
560 | ||
561 | if (!rq) | |
562 | return; | |
563 | ||
564 | rq->rq_sched_info.run_delay += diff; | |
565 | rq->rq_sched_info.pcnt++; | |
566 | } | |
567 | ||
568 | /* | |
569 | * Called when a process is queued into either the active or expired | |
570 | * array. The time is noted and later used to determine how long we | |
571 | * had to wait for us to reach the cpu. Since the expired queue will | |
572 | * become the active queue after active queue is empty, without dequeuing | |
573 | * and requeuing any tasks, we are interested in queuing to either. It | |
574 | * is unusual but not impossible for tasks to be dequeued and immediately | |
575 | * requeued in the same or another array: this can happen in sched_yield(), | |
576 | * set_user_nice(), and even load_balance() as it moves tasks from runqueue | |
577 | * to runqueue. | |
578 | * | |
579 | * This function is only called from enqueue_task(), but also only updates | |
580 | * the timestamp if it is already not set. It's assumed that | |
581 | * sched_info_dequeued() will clear that stamp when appropriate. | |
582 | */ | |
583 | static inline void sched_info_queued(task_t *t) | |
584 | { | |
585 | if (!t->sched_info.last_queued) | |
586 | t->sched_info.last_queued = jiffies; | |
587 | } | |
588 | ||
589 | /* | |
590 | * Called when a process ceases being the active-running process, either | |
591 | * voluntarily or involuntarily. Now we can calculate how long we ran. | |
592 | */ | |
593 | static inline void sched_info_depart(task_t *t) | |
594 | { | |
595 | struct runqueue *rq = task_rq(t); | |
596 | unsigned long diff = jiffies - t->sched_info.last_arrival; | |
597 | ||
598 | t->sched_info.cpu_time += diff; | |
599 | ||
600 | if (rq) | |
601 | rq->rq_sched_info.cpu_time += diff; | |
602 | } | |
603 | ||
604 | /* | |
605 | * Called when tasks are switched involuntarily due, typically, to expiring | |
606 | * their time slice. (This may also be called when switching to or from | |
607 | * the idle task.) We are only called when prev != next. | |
608 | */ | |
609 | static inline void sched_info_switch(task_t *prev, task_t *next) | |
610 | { | |
611 | struct runqueue *rq = task_rq(prev); | |
612 | ||
613 | /* | |
614 | * prev now departs the cpu. It's not interesting to record | |
615 | * stats about how efficient we were at scheduling the idle | |
616 | * process, however. | |
617 | */ | |
618 | if (prev != rq->idle) | |
619 | sched_info_depart(prev); | |
620 | ||
621 | if (next != rq->idle) | |
622 | sched_info_arrive(next); | |
623 | } | |
624 | #else | |
625 | #define sched_info_queued(t) do { } while (0) | |
626 | #define sched_info_switch(t, next) do { } while (0) | |
627 | #endif /* CONFIG_SCHEDSTATS */ | |
628 | ||
629 | /* | |
630 | * Adding/removing a task to/from a priority array: | |
631 | */ | |
632 | static void dequeue_task(struct task_struct *p, prio_array_t *array) | |
633 | { | |
634 | array->nr_active--; | |
635 | list_del(&p->run_list); | |
636 | if (list_empty(array->queue + p->prio)) | |
637 | __clear_bit(p->prio, array->bitmap); | |
638 | } | |
639 | ||
640 | static void enqueue_task(struct task_struct *p, prio_array_t *array) | |
641 | { | |
642 | sched_info_queued(p); | |
643 | list_add_tail(&p->run_list, array->queue + p->prio); | |
644 | __set_bit(p->prio, array->bitmap); | |
645 | array->nr_active++; | |
646 | p->array = array; | |
647 | } | |
648 | ||
649 | /* | |
650 | * Put task to the end of the run list without the overhead of dequeue | |
651 | * followed by enqueue. | |
652 | */ | |
653 | static void requeue_task(struct task_struct *p, prio_array_t *array) | |
654 | { | |
655 | list_move_tail(&p->run_list, array->queue + p->prio); | |
656 | } | |
657 | ||
658 | static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array) | |
659 | { | |
660 | list_add(&p->run_list, array->queue + p->prio); | |
661 | __set_bit(p->prio, array->bitmap); | |
662 | array->nr_active++; | |
663 | p->array = array; | |
664 | } | |
665 | ||
666 | /* | |
b29739f9 | 667 | * __normal_prio - return the priority that is based on the static |
1da177e4 LT |
668 | * priority but is modified by bonuses/penalties. |
669 | * | |
670 | * We scale the actual sleep average [0 .... MAX_SLEEP_AVG] | |
671 | * into the -5 ... 0 ... +5 bonus/penalty range. | |
672 | * | |
673 | * We use 25% of the full 0...39 priority range so that: | |
674 | * | |
675 | * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs. | |
676 | * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks. | |
677 | * | |
678 | * Both properties are important to certain workloads. | |
679 | */ | |
b29739f9 IM |
680 | |
681 | static inline int __normal_prio(task_t *p) | |
1da177e4 LT |
682 | { |
683 | int bonus, prio; | |
684 | ||
1da177e4 LT |
685 | bonus = CURRENT_BONUS(p) - MAX_BONUS / 2; |
686 | ||
687 | prio = p->static_prio - bonus; | |
688 | if (prio < MAX_RT_PRIO) | |
689 | prio = MAX_RT_PRIO; | |
690 | if (prio > MAX_PRIO-1) | |
691 | prio = MAX_PRIO-1; | |
692 | return prio; | |
693 | } | |
694 | ||
2dd73a4f PW |
695 | /* |
696 | * To aid in avoiding the subversion of "niceness" due to uneven distribution | |
697 | * of tasks with abnormal "nice" values across CPUs the contribution that | |
698 | * each task makes to its run queue's load is weighted according to its | |
699 | * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a | |
700 | * scaled version of the new time slice allocation that they receive on time | |
701 | * slice expiry etc. | |
702 | */ | |
703 | ||
704 | /* | |
705 | * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE | |
706 | * If static_prio_timeslice() is ever changed to break this assumption then | |
707 | * this code will need modification | |
708 | */ | |
709 | #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE | |
710 | #define LOAD_WEIGHT(lp) \ | |
711 | (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO) | |
712 | #define PRIO_TO_LOAD_WEIGHT(prio) \ | |
713 | LOAD_WEIGHT(static_prio_timeslice(prio)) | |
714 | #define RTPRIO_TO_LOAD_WEIGHT(rp) \ | |
715 | (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp)) | |
716 | ||
717 | static void set_load_weight(task_t *p) | |
718 | { | |
b29739f9 | 719 | if (has_rt_policy(p)) { |
2dd73a4f PW |
720 | #ifdef CONFIG_SMP |
721 | if (p == task_rq(p)->migration_thread) | |
722 | /* | |
723 | * The migration thread does the actual balancing. | |
724 | * Giving its load any weight will skew balancing | |
725 | * adversely. | |
726 | */ | |
727 | p->load_weight = 0; | |
728 | else | |
729 | #endif | |
730 | p->load_weight = RTPRIO_TO_LOAD_WEIGHT(p->rt_priority); | |
731 | } else | |
732 | p->load_weight = PRIO_TO_LOAD_WEIGHT(p->static_prio); | |
733 | } | |
734 | ||
735 | static inline void inc_raw_weighted_load(runqueue_t *rq, const task_t *p) | |
736 | { | |
737 | rq->raw_weighted_load += p->load_weight; | |
738 | } | |
739 | ||
740 | static inline void dec_raw_weighted_load(runqueue_t *rq, const task_t *p) | |
741 | { | |
742 | rq->raw_weighted_load -= p->load_weight; | |
743 | } | |
744 | ||
745 | static inline void inc_nr_running(task_t *p, runqueue_t *rq) | |
746 | { | |
747 | rq->nr_running++; | |
748 | inc_raw_weighted_load(rq, p); | |
749 | } | |
750 | ||
751 | static inline void dec_nr_running(task_t *p, runqueue_t *rq) | |
752 | { | |
753 | rq->nr_running--; | |
754 | dec_raw_weighted_load(rq, p); | |
755 | } | |
756 | ||
b29739f9 IM |
757 | /* |
758 | * Calculate the expected normal priority: i.e. priority | |
759 | * without taking RT-inheritance into account. Might be | |
760 | * boosted by interactivity modifiers. Changes upon fork, | |
761 | * setprio syscalls, and whenever the interactivity | |
762 | * estimator recalculates. | |
763 | */ | |
764 | static inline int normal_prio(task_t *p) | |
765 | { | |
766 | int prio; | |
767 | ||
768 | if (has_rt_policy(p)) | |
769 | prio = MAX_RT_PRIO-1 - p->rt_priority; | |
770 | else | |
771 | prio = __normal_prio(p); | |
772 | return prio; | |
773 | } | |
774 | ||
775 | /* | |
776 | * Calculate the current priority, i.e. the priority | |
777 | * taken into account by the scheduler. This value might | |
778 | * be boosted by RT tasks, or might be boosted by | |
779 | * interactivity modifiers. Will be RT if the task got | |
780 | * RT-boosted. If not then it returns p->normal_prio. | |
781 | */ | |
782 | static int effective_prio(task_t *p) | |
783 | { | |
784 | p->normal_prio = normal_prio(p); | |
785 | /* | |
786 | * If we are RT tasks or we were boosted to RT priority, | |
787 | * keep the priority unchanged. Otherwise, update priority | |
788 | * to the normal priority: | |
789 | */ | |
790 | if (!rt_prio(p->prio)) | |
791 | return p->normal_prio; | |
792 | return p->prio; | |
793 | } | |
794 | ||
1da177e4 LT |
795 | /* |
796 | * __activate_task - move a task to the runqueue. | |
797 | */ | |
d425b274 | 798 | static void __activate_task(task_t *p, runqueue_t *rq) |
1da177e4 | 799 | { |
d425b274 CK |
800 | prio_array_t *target = rq->active; |
801 | ||
f1adad78 | 802 | if (batch_task(p)) |
d425b274 CK |
803 | target = rq->expired; |
804 | enqueue_task(p, target); | |
2dd73a4f | 805 | inc_nr_running(p, rq); |
1da177e4 LT |
806 | } |
807 | ||
808 | /* | |
809 | * __activate_idle_task - move idle task to the _front_ of runqueue. | |
810 | */ | |
811 | static inline void __activate_idle_task(task_t *p, runqueue_t *rq) | |
812 | { | |
813 | enqueue_task_head(p, rq->active); | |
2dd73a4f | 814 | inc_nr_running(p, rq); |
1da177e4 LT |
815 | } |
816 | ||
b29739f9 IM |
817 | /* |
818 | * Recalculate p->normal_prio and p->prio after having slept, | |
819 | * updating the sleep-average too: | |
820 | */ | |
a3464a10 | 821 | static int recalc_task_prio(task_t *p, unsigned long long now) |
1da177e4 LT |
822 | { |
823 | /* Caller must always ensure 'now >= p->timestamp' */ | |
72d2854d | 824 | unsigned long sleep_time = now - p->timestamp; |
1da177e4 | 825 | |
d425b274 | 826 | if (batch_task(p)) |
b0a9499c | 827 | sleep_time = 0; |
1da177e4 LT |
828 | |
829 | if (likely(sleep_time > 0)) { | |
830 | /* | |
72d2854d CK |
831 | * This ceiling is set to the lowest priority that would allow |
832 | * a task to be reinserted into the active array on timeslice | |
833 | * completion. | |
1da177e4 | 834 | */ |
72d2854d | 835 | unsigned long ceiling = INTERACTIVE_SLEEP(p); |
e72ff0bb | 836 | |
72d2854d CK |
837 | if (p->mm && sleep_time > ceiling && p->sleep_avg < ceiling) { |
838 | /* | |
839 | * Prevents user tasks from achieving best priority | |
840 | * with one single large enough sleep. | |
841 | */ | |
842 | p->sleep_avg = ceiling; | |
843 | /* | |
844 | * Using INTERACTIVE_SLEEP() as a ceiling places a | |
845 | * nice(0) task 1ms sleep away from promotion, and | |
846 | * gives it 700ms to round-robin with no chance of | |
847 | * being demoted. This is more than generous, so | |
848 | * mark this sleep as non-interactive to prevent the | |
849 | * on-runqueue bonus logic from intervening should | |
850 | * this task not receive cpu immediately. | |
851 | */ | |
852 | p->sleep_type = SLEEP_NONINTERACTIVE; | |
1da177e4 | 853 | } else { |
1da177e4 LT |
854 | /* |
855 | * Tasks waking from uninterruptible sleep are | |
856 | * limited in their sleep_avg rise as they | |
857 | * are likely to be waiting on I/O | |
858 | */ | |
3dee386e | 859 | if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) { |
72d2854d | 860 | if (p->sleep_avg >= ceiling) |
1da177e4 LT |
861 | sleep_time = 0; |
862 | else if (p->sleep_avg + sleep_time >= | |
72d2854d CK |
863 | ceiling) { |
864 | p->sleep_avg = ceiling; | |
865 | sleep_time = 0; | |
1da177e4 LT |
866 | } |
867 | } | |
868 | ||
869 | /* | |
870 | * This code gives a bonus to interactive tasks. | |
871 | * | |
872 | * The boost works by updating the 'average sleep time' | |
873 | * value here, based on ->timestamp. The more time a | |
874 | * task spends sleeping, the higher the average gets - | |
875 | * and the higher the priority boost gets as well. | |
876 | */ | |
877 | p->sleep_avg += sleep_time; | |
878 | ||
1da177e4 | 879 | } |
72d2854d CK |
880 | if (p->sleep_avg > NS_MAX_SLEEP_AVG) |
881 | p->sleep_avg = NS_MAX_SLEEP_AVG; | |
1da177e4 LT |
882 | } |
883 | ||
a3464a10 | 884 | return effective_prio(p); |
1da177e4 LT |
885 | } |
886 | ||
887 | /* | |
888 | * activate_task - move a task to the runqueue and do priority recalculation | |
889 | * | |
890 | * Update all the scheduling statistics stuff. (sleep average | |
891 | * calculation, priority modifiers, etc.) | |
892 | */ | |
893 | static void activate_task(task_t *p, runqueue_t *rq, int local) | |
894 | { | |
895 | unsigned long long now; | |
896 | ||
897 | now = sched_clock(); | |
898 | #ifdef CONFIG_SMP | |
899 | if (!local) { | |
900 | /* Compensate for drifting sched_clock */ | |
901 | runqueue_t *this_rq = this_rq(); | |
902 | now = (now - this_rq->timestamp_last_tick) | |
903 | + rq->timestamp_last_tick; | |
904 | } | |
905 | #endif | |
906 | ||
a47ab937 KC |
907 | if (!rt_task(p)) |
908 | p->prio = recalc_task_prio(p, now); | |
1da177e4 LT |
909 | |
910 | /* | |
911 | * This checks to make sure it's not an uninterruptible task | |
912 | * that is now waking up. | |
913 | */ | |
3dee386e | 914 | if (p->sleep_type == SLEEP_NORMAL) { |
1da177e4 LT |
915 | /* |
916 | * Tasks which were woken up by interrupts (ie. hw events) | |
917 | * are most likely of interactive nature. So we give them | |
918 | * the credit of extending their sleep time to the period | |
919 | * of time they spend on the runqueue, waiting for execution | |
920 | * on a CPU, first time around: | |
921 | */ | |
922 | if (in_interrupt()) | |
3dee386e | 923 | p->sleep_type = SLEEP_INTERRUPTED; |
1da177e4 LT |
924 | else { |
925 | /* | |
926 | * Normal first-time wakeups get a credit too for | |
927 | * on-runqueue time, but it will be weighted down: | |
928 | */ | |
3dee386e | 929 | p->sleep_type = SLEEP_INTERACTIVE; |
1da177e4 LT |
930 | } |
931 | } | |
932 | p->timestamp = now; | |
933 | ||
934 | __activate_task(p, rq); | |
935 | } | |
936 | ||
937 | /* | |
938 | * deactivate_task - remove a task from the runqueue. | |
939 | */ | |
940 | static void deactivate_task(struct task_struct *p, runqueue_t *rq) | |
941 | { | |
2dd73a4f | 942 | dec_nr_running(p, rq); |
1da177e4 LT |
943 | dequeue_task(p, p->array); |
944 | p->array = NULL; | |
945 | } | |
946 | ||
947 | /* | |
948 | * resched_task - mark a task 'to be rescheduled now'. | |
949 | * | |
950 | * On UP this means the setting of the need_resched flag, on SMP it | |
951 | * might also involve a cross-CPU call to trigger the scheduler on | |
952 | * the target CPU. | |
953 | */ | |
954 | #ifdef CONFIG_SMP | |
495ab9c0 AK |
955 | |
956 | #ifndef tsk_is_polling | |
957 | #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG) | |
958 | #endif | |
959 | ||
1da177e4 LT |
960 | static void resched_task(task_t *p) |
961 | { | |
64c7c8f8 | 962 | int cpu; |
1da177e4 LT |
963 | |
964 | assert_spin_locked(&task_rq(p)->lock); | |
965 | ||
64c7c8f8 NP |
966 | if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED))) |
967 | return; | |
968 | ||
969 | set_tsk_thread_flag(p, TIF_NEED_RESCHED); | |
1da177e4 | 970 | |
64c7c8f8 NP |
971 | cpu = task_cpu(p); |
972 | if (cpu == smp_processor_id()) | |
973 | return; | |
974 | ||
495ab9c0 | 975 | /* NEED_RESCHED must be visible before we test polling */ |
64c7c8f8 | 976 | smp_mb(); |
495ab9c0 | 977 | if (!tsk_is_polling(p)) |
64c7c8f8 | 978 | smp_send_reschedule(cpu); |
1da177e4 LT |
979 | } |
980 | #else | |
981 | static inline void resched_task(task_t *p) | |
982 | { | |
64c7c8f8 | 983 | assert_spin_locked(&task_rq(p)->lock); |
1da177e4 LT |
984 | set_tsk_need_resched(p); |
985 | } | |
986 | #endif | |
987 | ||
988 | /** | |
989 | * task_curr - is this task currently executing on a CPU? | |
990 | * @p: the task in question. | |
991 | */ | |
992 | inline int task_curr(const task_t *p) | |
993 | { | |
994 | return cpu_curr(task_cpu(p)) == p; | |
995 | } | |
996 | ||
2dd73a4f PW |
997 | /* Used instead of source_load when we know the type == 0 */ |
998 | unsigned long weighted_cpuload(const int cpu) | |
999 | { | |
1000 | return cpu_rq(cpu)->raw_weighted_load; | |
1001 | } | |
1002 | ||
1da177e4 | 1003 | #ifdef CONFIG_SMP |
1da177e4 LT |
1004 | typedef struct { |
1005 | struct list_head list; | |
1da177e4 | 1006 | |
1da177e4 LT |
1007 | task_t *task; |
1008 | int dest_cpu; | |
1009 | ||
1da177e4 LT |
1010 | struct completion done; |
1011 | } migration_req_t; | |
1012 | ||
1013 | /* | |
1014 | * The task's runqueue lock must be held. | |
1015 | * Returns true if you have to wait for migration thread. | |
1016 | */ | |
1017 | static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req) | |
1018 | { | |
1019 | runqueue_t *rq = task_rq(p); | |
1020 | ||
1021 | /* | |
1022 | * If the task is not on a runqueue (and not running), then | |
1023 | * it is sufficient to simply update the task's cpu field. | |
1024 | */ | |
1025 | if (!p->array && !task_running(rq, p)) { | |
1026 | set_task_cpu(p, dest_cpu); | |
1027 | return 0; | |
1028 | } | |
1029 | ||
1030 | init_completion(&req->done); | |
1da177e4 LT |
1031 | req->task = p; |
1032 | req->dest_cpu = dest_cpu; | |
1033 | list_add(&req->list, &rq->migration_queue); | |
1034 | return 1; | |
1035 | } | |
1036 | ||
1037 | /* | |
1038 | * wait_task_inactive - wait for a thread to unschedule. | |
1039 | * | |
1040 | * The caller must ensure that the task *will* unschedule sometime soon, | |
1041 | * else this function might spin for a *long* time. This function can't | |
1042 | * be called with interrupts off, or it may introduce deadlock with | |
1043 | * smp_call_function() if an IPI is sent by the same process we are | |
1044 | * waiting to become inactive. | |
1045 | */ | |
95cdf3b7 | 1046 | void wait_task_inactive(task_t *p) |
1da177e4 LT |
1047 | { |
1048 | unsigned long flags; | |
1049 | runqueue_t *rq; | |
1050 | int preempted; | |
1051 | ||
1052 | repeat: | |
1053 | rq = task_rq_lock(p, &flags); | |
1054 | /* Must be off runqueue entirely, not preempted. */ | |
1055 | if (unlikely(p->array || task_running(rq, p))) { | |
1056 | /* If it's preempted, we yield. It could be a while. */ | |
1057 | preempted = !task_running(rq, p); | |
1058 | task_rq_unlock(rq, &flags); | |
1059 | cpu_relax(); | |
1060 | if (preempted) | |
1061 | yield(); | |
1062 | goto repeat; | |
1063 | } | |
1064 | task_rq_unlock(rq, &flags); | |
1065 | } | |
1066 | ||
1067 | /*** | |
1068 | * kick_process - kick a running thread to enter/exit the kernel | |
1069 | * @p: the to-be-kicked thread | |
1070 | * | |
1071 | * Cause a process which is running on another CPU to enter | |
1072 | * kernel-mode, without any delay. (to get signals handled.) | |
1073 | * | |
1074 | * NOTE: this function doesnt have to take the runqueue lock, | |
1075 | * because all it wants to ensure is that the remote task enters | |
1076 | * the kernel. If the IPI races and the task has been migrated | |
1077 | * to another CPU then no harm is done and the purpose has been | |
1078 | * achieved as well. | |
1079 | */ | |
1080 | void kick_process(task_t *p) | |
1081 | { | |
1082 | int cpu; | |
1083 | ||
1084 | preempt_disable(); | |
1085 | cpu = task_cpu(p); | |
1086 | if ((cpu != smp_processor_id()) && task_curr(p)) | |
1087 | smp_send_reschedule(cpu); | |
1088 | preempt_enable(); | |
1089 | } | |
1090 | ||
1091 | /* | |
2dd73a4f PW |
1092 | * Return a low guess at the load of a migration-source cpu weighted |
1093 | * according to the scheduling class and "nice" value. | |
1da177e4 LT |
1094 | * |
1095 | * We want to under-estimate the load of migration sources, to | |
1096 | * balance conservatively. | |
1097 | */ | |
a2000572 | 1098 | static inline unsigned long source_load(int cpu, int type) |
1da177e4 LT |
1099 | { |
1100 | runqueue_t *rq = cpu_rq(cpu); | |
2dd73a4f | 1101 | |
3b0bd9bc | 1102 | if (type == 0) |
2dd73a4f | 1103 | return rq->raw_weighted_load; |
b910472d | 1104 | |
2dd73a4f | 1105 | return min(rq->cpu_load[type-1], rq->raw_weighted_load); |
1da177e4 LT |
1106 | } |
1107 | ||
1108 | /* | |
2dd73a4f PW |
1109 | * Return a high guess at the load of a migration-target cpu weighted |
1110 | * according to the scheduling class and "nice" value. | |
1da177e4 | 1111 | */ |
a2000572 | 1112 | static inline unsigned long target_load(int cpu, int type) |
1da177e4 LT |
1113 | { |
1114 | runqueue_t *rq = cpu_rq(cpu); | |
2dd73a4f | 1115 | |
7897986b | 1116 | if (type == 0) |
2dd73a4f | 1117 | return rq->raw_weighted_load; |
3b0bd9bc | 1118 | |
2dd73a4f PW |
1119 | return max(rq->cpu_load[type-1], rq->raw_weighted_load); |
1120 | } | |
1121 | ||
1122 | /* | |
1123 | * Return the average load per task on the cpu's run queue | |
1124 | */ | |
1125 | static inline unsigned long cpu_avg_load_per_task(int cpu) | |
1126 | { | |
1127 | runqueue_t *rq = cpu_rq(cpu); | |
1128 | unsigned long n = rq->nr_running; | |
1129 | ||
1130 | return n ? rq->raw_weighted_load / n : SCHED_LOAD_SCALE; | |
1da177e4 LT |
1131 | } |
1132 | ||
147cbb4b NP |
1133 | /* |
1134 | * find_idlest_group finds and returns the least busy CPU group within the | |
1135 | * domain. | |
1136 | */ | |
1137 | static struct sched_group * | |
1138 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu) | |
1139 | { | |
1140 | struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups; | |
1141 | unsigned long min_load = ULONG_MAX, this_load = 0; | |
1142 | int load_idx = sd->forkexec_idx; | |
1143 | int imbalance = 100 + (sd->imbalance_pct-100)/2; | |
1144 | ||
1145 | do { | |
1146 | unsigned long load, avg_load; | |
1147 | int local_group; | |
1148 | int i; | |
1149 | ||
da5a5522 BD |
1150 | /* Skip over this group if it has no CPUs allowed */ |
1151 | if (!cpus_intersects(group->cpumask, p->cpus_allowed)) | |
1152 | goto nextgroup; | |
1153 | ||
147cbb4b | 1154 | local_group = cpu_isset(this_cpu, group->cpumask); |
147cbb4b NP |
1155 | |
1156 | /* Tally up the load of all CPUs in the group */ | |
1157 | avg_load = 0; | |
1158 | ||
1159 | for_each_cpu_mask(i, group->cpumask) { | |
1160 | /* Bias balancing toward cpus of our domain */ | |
1161 | if (local_group) | |
1162 | load = source_load(i, load_idx); | |
1163 | else | |
1164 | load = target_load(i, load_idx); | |
1165 | ||
1166 | avg_load += load; | |
1167 | } | |
1168 | ||
1169 | /* Adjust by relative CPU power of the group */ | |
1170 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
1171 | ||
1172 | if (local_group) { | |
1173 | this_load = avg_load; | |
1174 | this = group; | |
1175 | } else if (avg_load < min_load) { | |
1176 | min_load = avg_load; | |
1177 | idlest = group; | |
1178 | } | |
da5a5522 | 1179 | nextgroup: |
147cbb4b NP |
1180 | group = group->next; |
1181 | } while (group != sd->groups); | |
1182 | ||
1183 | if (!idlest || 100*this_load < imbalance*min_load) | |
1184 | return NULL; | |
1185 | return idlest; | |
1186 | } | |
1187 | ||
1188 | /* | |
1189 | * find_idlest_queue - find the idlest runqueue among the cpus in group. | |
1190 | */ | |
95cdf3b7 IM |
1191 | static int |
1192 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
147cbb4b | 1193 | { |
da5a5522 | 1194 | cpumask_t tmp; |
147cbb4b NP |
1195 | unsigned long load, min_load = ULONG_MAX; |
1196 | int idlest = -1; | |
1197 | int i; | |
1198 | ||
da5a5522 BD |
1199 | /* Traverse only the allowed CPUs */ |
1200 | cpus_and(tmp, group->cpumask, p->cpus_allowed); | |
1201 | ||
1202 | for_each_cpu_mask(i, tmp) { | |
2dd73a4f | 1203 | load = weighted_cpuload(i); |
147cbb4b NP |
1204 | |
1205 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
1206 | min_load = load; | |
1207 | idlest = i; | |
1208 | } | |
1209 | } | |
1210 | ||
1211 | return idlest; | |
1212 | } | |
1213 | ||
476d139c NP |
1214 | /* |
1215 | * sched_balance_self: balance the current task (running on cpu) in domains | |
1216 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
1217 | * SD_BALANCE_EXEC. | |
1218 | * | |
1219 | * Balance, ie. select the least loaded group. | |
1220 | * | |
1221 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
1222 | * | |
1223 | * preempt must be disabled. | |
1224 | */ | |
1225 | static int sched_balance_self(int cpu, int flag) | |
1226 | { | |
1227 | struct task_struct *t = current; | |
1228 | struct sched_domain *tmp, *sd = NULL; | |
147cbb4b | 1229 | |
c96d145e | 1230 | for_each_domain(cpu, tmp) { |
5c45bf27 SS |
1231 | /* |
1232 | * If power savings logic is enabled for a domain, stop there. | |
1233 | */ | |
1234 | if (tmp->flags & SD_POWERSAVINGS_BALANCE) | |
1235 | break; | |
476d139c NP |
1236 | if (tmp->flags & flag) |
1237 | sd = tmp; | |
c96d145e | 1238 | } |
476d139c NP |
1239 | |
1240 | while (sd) { | |
1241 | cpumask_t span; | |
1242 | struct sched_group *group; | |
1243 | int new_cpu; | |
1244 | int weight; | |
1245 | ||
1246 | span = sd->span; | |
1247 | group = find_idlest_group(sd, t, cpu); | |
1248 | if (!group) | |
1249 | goto nextlevel; | |
1250 | ||
da5a5522 | 1251 | new_cpu = find_idlest_cpu(group, t, cpu); |
476d139c NP |
1252 | if (new_cpu == -1 || new_cpu == cpu) |
1253 | goto nextlevel; | |
1254 | ||
1255 | /* Now try balancing at a lower domain level */ | |
1256 | cpu = new_cpu; | |
1257 | nextlevel: | |
1258 | sd = NULL; | |
1259 | weight = cpus_weight(span); | |
1260 | for_each_domain(cpu, tmp) { | |
1261 | if (weight <= cpus_weight(tmp->span)) | |
1262 | break; | |
1263 | if (tmp->flags & flag) | |
1264 | sd = tmp; | |
1265 | } | |
1266 | /* while loop will break here if sd == NULL */ | |
1267 | } | |
1268 | ||
1269 | return cpu; | |
1270 | } | |
1271 | ||
1272 | #endif /* CONFIG_SMP */ | |
1da177e4 LT |
1273 | |
1274 | /* | |
1275 | * wake_idle() will wake a task on an idle cpu if task->cpu is | |
1276 | * not idle and an idle cpu is available. The span of cpus to | |
1277 | * search starts with cpus closest then further out as needed, | |
1278 | * so we always favor a closer, idle cpu. | |
1279 | * | |
1280 | * Returns the CPU we should wake onto. | |
1281 | */ | |
1282 | #if defined(ARCH_HAS_SCHED_WAKE_IDLE) | |
1283 | static int wake_idle(int cpu, task_t *p) | |
1284 | { | |
1285 | cpumask_t tmp; | |
1286 | struct sched_domain *sd; | |
1287 | int i; | |
1288 | ||
1289 | if (idle_cpu(cpu)) | |
1290 | return cpu; | |
1291 | ||
1292 | for_each_domain(cpu, sd) { | |
1293 | if (sd->flags & SD_WAKE_IDLE) { | |
e0f364f4 | 1294 | cpus_and(tmp, sd->span, p->cpus_allowed); |
1da177e4 LT |
1295 | for_each_cpu_mask(i, tmp) { |
1296 | if (idle_cpu(i)) | |
1297 | return i; | |
1298 | } | |
1299 | } | |
e0f364f4 NP |
1300 | else |
1301 | break; | |
1da177e4 LT |
1302 | } |
1303 | return cpu; | |
1304 | } | |
1305 | #else | |
1306 | static inline int wake_idle(int cpu, task_t *p) | |
1307 | { | |
1308 | return cpu; | |
1309 | } | |
1310 | #endif | |
1311 | ||
1312 | /*** | |
1313 | * try_to_wake_up - wake up a thread | |
1314 | * @p: the to-be-woken-up thread | |
1315 | * @state: the mask of task states that can be woken | |
1316 | * @sync: do a synchronous wakeup? | |
1317 | * | |
1318 | * Put it on the run-queue if it's not already there. The "current" | |
1319 | * thread is always on the run-queue (except when the actual | |
1320 | * re-schedule is in progress), and as such you're allowed to do | |
1321 | * the simpler "current->state = TASK_RUNNING" to mark yourself | |
1322 | * runnable without the overhead of this. | |
1323 | * | |
1324 | * returns failure only if the task is already active. | |
1325 | */ | |
95cdf3b7 | 1326 | static int try_to_wake_up(task_t *p, unsigned int state, int sync) |
1da177e4 LT |
1327 | { |
1328 | int cpu, this_cpu, success = 0; | |
1329 | unsigned long flags; | |
1330 | long old_state; | |
1331 | runqueue_t *rq; | |
1332 | #ifdef CONFIG_SMP | |
1333 | unsigned long load, this_load; | |
7897986b | 1334 | struct sched_domain *sd, *this_sd = NULL; |
1da177e4 LT |
1335 | int new_cpu; |
1336 | #endif | |
1337 | ||
1338 | rq = task_rq_lock(p, &flags); | |
1339 | old_state = p->state; | |
1340 | if (!(old_state & state)) | |
1341 | goto out; | |
1342 | ||
1343 | if (p->array) | |
1344 | goto out_running; | |
1345 | ||
1346 | cpu = task_cpu(p); | |
1347 | this_cpu = smp_processor_id(); | |
1348 | ||
1349 | #ifdef CONFIG_SMP | |
1350 | if (unlikely(task_running(rq, p))) | |
1351 | goto out_activate; | |
1352 | ||
7897986b NP |
1353 | new_cpu = cpu; |
1354 | ||
1da177e4 LT |
1355 | schedstat_inc(rq, ttwu_cnt); |
1356 | if (cpu == this_cpu) { | |
1357 | schedstat_inc(rq, ttwu_local); | |
7897986b NP |
1358 | goto out_set_cpu; |
1359 | } | |
1360 | ||
1361 | for_each_domain(this_cpu, sd) { | |
1362 | if (cpu_isset(cpu, sd->span)) { | |
1363 | schedstat_inc(sd, ttwu_wake_remote); | |
1364 | this_sd = sd; | |
1365 | break; | |
1da177e4 LT |
1366 | } |
1367 | } | |
1da177e4 | 1368 | |
7897986b | 1369 | if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed))) |
1da177e4 LT |
1370 | goto out_set_cpu; |
1371 | ||
1da177e4 | 1372 | /* |
7897986b | 1373 | * Check for affine wakeup and passive balancing possibilities. |
1da177e4 | 1374 | */ |
7897986b NP |
1375 | if (this_sd) { |
1376 | int idx = this_sd->wake_idx; | |
1377 | unsigned int imbalance; | |
1da177e4 | 1378 | |
a3f21bce NP |
1379 | imbalance = 100 + (this_sd->imbalance_pct - 100) / 2; |
1380 | ||
7897986b NP |
1381 | load = source_load(cpu, idx); |
1382 | this_load = target_load(this_cpu, idx); | |
1da177e4 | 1383 | |
7897986b NP |
1384 | new_cpu = this_cpu; /* Wake to this CPU if we can */ |
1385 | ||
a3f21bce NP |
1386 | if (this_sd->flags & SD_WAKE_AFFINE) { |
1387 | unsigned long tl = this_load; | |
2dd73a4f PW |
1388 | unsigned long tl_per_task = cpu_avg_load_per_task(this_cpu); |
1389 | ||
1da177e4 | 1390 | /* |
a3f21bce NP |
1391 | * If sync wakeup then subtract the (maximum possible) |
1392 | * effect of the currently running task from the load | |
1393 | * of the current CPU: | |
1da177e4 | 1394 | */ |
a3f21bce | 1395 | if (sync) |
2dd73a4f | 1396 | tl -= current->load_weight; |
a3f21bce NP |
1397 | |
1398 | if ((tl <= load && | |
2dd73a4f PW |
1399 | tl + target_load(cpu, idx) <= tl_per_task) || |
1400 | 100*(tl + p->load_weight) <= imbalance*load) { | |
a3f21bce NP |
1401 | /* |
1402 | * This domain has SD_WAKE_AFFINE and | |
1403 | * p is cache cold in this domain, and | |
1404 | * there is no bad imbalance. | |
1405 | */ | |
1406 | schedstat_inc(this_sd, ttwu_move_affine); | |
1407 | goto out_set_cpu; | |
1408 | } | |
1409 | } | |
1410 | ||
1411 | /* | |
1412 | * Start passive balancing when half the imbalance_pct | |
1413 | * limit is reached. | |
1414 | */ | |
1415 | if (this_sd->flags & SD_WAKE_BALANCE) { | |
1416 | if (imbalance*this_load <= 100*load) { | |
1417 | schedstat_inc(this_sd, ttwu_move_balance); | |
1418 | goto out_set_cpu; | |
1419 | } | |
1da177e4 LT |
1420 | } |
1421 | } | |
1422 | ||
1423 | new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */ | |
1424 | out_set_cpu: | |
1425 | new_cpu = wake_idle(new_cpu, p); | |
1426 | if (new_cpu != cpu) { | |
1427 | set_task_cpu(p, new_cpu); | |
1428 | task_rq_unlock(rq, &flags); | |
1429 | /* might preempt at this point */ | |
1430 | rq = task_rq_lock(p, &flags); | |
1431 | old_state = p->state; | |
1432 | if (!(old_state & state)) | |
1433 | goto out; | |
1434 | if (p->array) | |
1435 | goto out_running; | |
1436 | ||
1437 | this_cpu = smp_processor_id(); | |
1438 | cpu = task_cpu(p); | |
1439 | } | |
1440 | ||
1441 | out_activate: | |
1442 | #endif /* CONFIG_SMP */ | |
1443 | if (old_state == TASK_UNINTERRUPTIBLE) { | |
1444 | rq->nr_uninterruptible--; | |
1445 | /* | |
1446 | * Tasks on involuntary sleep don't earn | |
1447 | * sleep_avg beyond just interactive state. | |
1448 | */ | |
3dee386e | 1449 | p->sleep_type = SLEEP_NONINTERACTIVE; |
e7c38cb4 | 1450 | } else |
1da177e4 | 1451 | |
d79fc0fc IM |
1452 | /* |
1453 | * Tasks that have marked their sleep as noninteractive get | |
e7c38cb4 CK |
1454 | * woken up with their sleep average not weighted in an |
1455 | * interactive way. | |
d79fc0fc | 1456 | */ |
e7c38cb4 CK |
1457 | if (old_state & TASK_NONINTERACTIVE) |
1458 | p->sleep_type = SLEEP_NONINTERACTIVE; | |
1459 | ||
1460 | ||
1461 | activate_task(p, rq, cpu == this_cpu); | |
1da177e4 LT |
1462 | /* |
1463 | * Sync wakeups (i.e. those types of wakeups where the waker | |
1464 | * has indicated that it will leave the CPU in short order) | |
1465 | * don't trigger a preemption, if the woken up task will run on | |
1466 | * this cpu. (in this case the 'I will reschedule' promise of | |
1467 | * the waker guarantees that the freshly woken up task is going | |
1468 | * to be considered on this CPU.) | |
1469 | */ | |
1da177e4 LT |
1470 | if (!sync || cpu != this_cpu) { |
1471 | if (TASK_PREEMPTS_CURR(p, rq)) | |
1472 | resched_task(rq->curr); | |
1473 | } | |
1474 | success = 1; | |
1475 | ||
1476 | out_running: | |
1477 | p->state = TASK_RUNNING; | |
1478 | out: | |
1479 | task_rq_unlock(rq, &flags); | |
1480 | ||
1481 | return success; | |
1482 | } | |
1483 | ||
95cdf3b7 | 1484 | int fastcall wake_up_process(task_t *p) |
1da177e4 LT |
1485 | { |
1486 | return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED | | |
1487 | TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0); | |
1488 | } | |
1489 | ||
1490 | EXPORT_SYMBOL(wake_up_process); | |
1491 | ||
1492 | int fastcall wake_up_state(task_t *p, unsigned int state) | |
1493 | { | |
1494 | return try_to_wake_up(p, state, 0); | |
1495 | } | |
1496 | ||
1da177e4 LT |
1497 | /* |
1498 | * Perform scheduler related setup for a newly forked process p. | |
1499 | * p is forked by current. | |
1500 | */ | |
476d139c | 1501 | void fastcall sched_fork(task_t *p, int clone_flags) |
1da177e4 | 1502 | { |
476d139c NP |
1503 | int cpu = get_cpu(); |
1504 | ||
1505 | #ifdef CONFIG_SMP | |
1506 | cpu = sched_balance_self(cpu, SD_BALANCE_FORK); | |
1507 | #endif | |
1508 | set_task_cpu(p, cpu); | |
1509 | ||
1da177e4 LT |
1510 | /* |
1511 | * We mark the process as running here, but have not actually | |
1512 | * inserted it onto the runqueue yet. This guarantees that | |
1513 | * nobody will actually run it, and a signal or other external | |
1514 | * event cannot wake it up and insert it on the runqueue either. | |
1515 | */ | |
1516 | p->state = TASK_RUNNING; | |
b29739f9 IM |
1517 | |
1518 | /* | |
1519 | * Make sure we do not leak PI boosting priority to the child: | |
1520 | */ | |
1521 | p->prio = current->normal_prio; | |
1522 | ||
1da177e4 LT |
1523 | INIT_LIST_HEAD(&p->run_list); |
1524 | p->array = NULL; | |
1da177e4 LT |
1525 | #ifdef CONFIG_SCHEDSTATS |
1526 | memset(&p->sched_info, 0, sizeof(p->sched_info)); | |
1527 | #endif | |
d6077cb8 | 1528 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
4866cde0 NP |
1529 | p->oncpu = 0; |
1530 | #endif | |
1da177e4 | 1531 | #ifdef CONFIG_PREEMPT |
4866cde0 | 1532 | /* Want to start with kernel preemption disabled. */ |
a1261f54 | 1533 | task_thread_info(p)->preempt_count = 1; |
1da177e4 LT |
1534 | #endif |
1535 | /* | |
1536 | * Share the timeslice between parent and child, thus the | |
1537 | * total amount of pending timeslices in the system doesn't change, | |
1538 | * resulting in more scheduling fairness. | |
1539 | */ | |
1540 | local_irq_disable(); | |
1541 | p->time_slice = (current->time_slice + 1) >> 1; | |
1542 | /* | |
1543 | * The remainder of the first timeslice might be recovered by | |
1544 | * the parent if the child exits early enough. | |
1545 | */ | |
1546 | p->first_time_slice = 1; | |
1547 | current->time_slice >>= 1; | |
1548 | p->timestamp = sched_clock(); | |
1549 | if (unlikely(!current->time_slice)) { | |
1550 | /* | |
1551 | * This case is rare, it happens when the parent has only | |
1552 | * a single jiffy left from its timeslice. Taking the | |
1553 | * runqueue lock is not a problem. | |
1554 | */ | |
1555 | current->time_slice = 1; | |
1da177e4 | 1556 | scheduler_tick(); |
476d139c NP |
1557 | } |
1558 | local_irq_enable(); | |
1559 | put_cpu(); | |
1da177e4 LT |
1560 | } |
1561 | ||
1562 | /* | |
1563 | * wake_up_new_task - wake up a newly created task for the first time. | |
1564 | * | |
1565 | * This function will do some initial scheduler statistics housekeeping | |
1566 | * that must be done for every newly created context, then puts the task | |
1567 | * on the runqueue and wakes it. | |
1568 | */ | |
95cdf3b7 | 1569 | void fastcall wake_up_new_task(task_t *p, unsigned long clone_flags) |
1da177e4 LT |
1570 | { |
1571 | unsigned long flags; | |
1572 | int this_cpu, cpu; | |
1573 | runqueue_t *rq, *this_rq; | |
1574 | ||
1575 | rq = task_rq_lock(p, &flags); | |
147cbb4b | 1576 | BUG_ON(p->state != TASK_RUNNING); |
1da177e4 | 1577 | this_cpu = smp_processor_id(); |
147cbb4b | 1578 | cpu = task_cpu(p); |
1da177e4 | 1579 | |
1da177e4 LT |
1580 | /* |
1581 | * We decrease the sleep average of forking parents | |
1582 | * and children as well, to keep max-interactive tasks | |
1583 | * from forking tasks that are max-interactive. The parent | |
1584 | * (current) is done further down, under its lock. | |
1585 | */ | |
1586 | p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) * | |
1587 | CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS); | |
1588 | ||
1589 | p->prio = effective_prio(p); | |
1590 | ||
1591 | if (likely(cpu == this_cpu)) { | |
1592 | if (!(clone_flags & CLONE_VM)) { | |
1593 | /* | |
1594 | * The VM isn't cloned, so we're in a good position to | |
1595 | * do child-runs-first in anticipation of an exec. This | |
1596 | * usually avoids a lot of COW overhead. | |
1597 | */ | |
1598 | if (unlikely(!current->array)) | |
1599 | __activate_task(p, rq); | |
1600 | else { | |
1601 | p->prio = current->prio; | |
b29739f9 | 1602 | p->normal_prio = current->normal_prio; |
1da177e4 LT |
1603 | list_add_tail(&p->run_list, ¤t->run_list); |
1604 | p->array = current->array; | |
1605 | p->array->nr_active++; | |
2dd73a4f | 1606 | inc_nr_running(p, rq); |
1da177e4 LT |
1607 | } |
1608 | set_need_resched(); | |
1609 | } else | |
1610 | /* Run child last */ | |
1611 | __activate_task(p, rq); | |
1612 | /* | |
1613 | * We skip the following code due to cpu == this_cpu | |
1614 | * | |
1615 | * task_rq_unlock(rq, &flags); | |
1616 | * this_rq = task_rq_lock(current, &flags); | |
1617 | */ | |
1618 | this_rq = rq; | |
1619 | } else { | |
1620 | this_rq = cpu_rq(this_cpu); | |
1621 | ||
1622 | /* | |
1623 | * Not the local CPU - must adjust timestamp. This should | |
1624 | * get optimised away in the !CONFIG_SMP case. | |
1625 | */ | |
1626 | p->timestamp = (p->timestamp - this_rq->timestamp_last_tick) | |
1627 | + rq->timestamp_last_tick; | |
1628 | __activate_task(p, rq); | |
1629 | if (TASK_PREEMPTS_CURR(p, rq)) | |
1630 | resched_task(rq->curr); | |
1631 | ||
1632 | /* | |
1633 | * Parent and child are on different CPUs, now get the | |
1634 | * parent runqueue to update the parent's ->sleep_avg: | |
1635 | */ | |
1636 | task_rq_unlock(rq, &flags); | |
1637 | this_rq = task_rq_lock(current, &flags); | |
1638 | } | |
1639 | current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) * | |
1640 | PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS); | |
1641 | task_rq_unlock(this_rq, &flags); | |
1642 | } | |
1643 | ||
1644 | /* | |
1645 | * Potentially available exiting-child timeslices are | |
1646 | * retrieved here - this way the parent does not get | |
1647 | * penalized for creating too many threads. | |
1648 | * | |
1649 | * (this cannot be used to 'generate' timeslices | |
1650 | * artificially, because any timeslice recovered here | |
1651 | * was given away by the parent in the first place.) | |
1652 | */ | |
95cdf3b7 | 1653 | void fastcall sched_exit(task_t *p) |
1da177e4 LT |
1654 | { |
1655 | unsigned long flags; | |
1656 | runqueue_t *rq; | |
1657 | ||
1658 | /* | |
1659 | * If the child was a (relative-) CPU hog then decrease | |
1660 | * the sleep_avg of the parent as well. | |
1661 | */ | |
1662 | rq = task_rq_lock(p->parent, &flags); | |
889dfafe | 1663 | if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) { |
1da177e4 LT |
1664 | p->parent->time_slice += p->time_slice; |
1665 | if (unlikely(p->parent->time_slice > task_timeslice(p))) | |
1666 | p->parent->time_slice = task_timeslice(p); | |
1667 | } | |
1668 | if (p->sleep_avg < p->parent->sleep_avg) | |
1669 | p->parent->sleep_avg = p->parent->sleep_avg / | |
1670 | (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg / | |
1671 | (EXIT_WEIGHT + 1); | |
1672 | task_rq_unlock(rq, &flags); | |
1673 | } | |
1674 | ||
4866cde0 NP |
1675 | /** |
1676 | * prepare_task_switch - prepare to switch tasks | |
1677 | * @rq: the runqueue preparing to switch | |
1678 | * @next: the task we are going to switch to. | |
1679 | * | |
1680 | * This is called with the rq lock held and interrupts off. It must | |
1681 | * be paired with a subsequent finish_task_switch after the context | |
1682 | * switch. | |
1683 | * | |
1684 | * prepare_task_switch sets up locking and calls architecture specific | |
1685 | * hooks. | |
1686 | */ | |
1687 | static inline void prepare_task_switch(runqueue_t *rq, task_t *next) | |
1688 | { | |
1689 | prepare_lock_switch(rq, next); | |
1690 | prepare_arch_switch(next); | |
1691 | } | |
1692 | ||
1da177e4 LT |
1693 | /** |
1694 | * finish_task_switch - clean up after a task-switch | |
344babaa | 1695 | * @rq: runqueue associated with task-switch |
1da177e4 LT |
1696 | * @prev: the thread we just switched away from. |
1697 | * | |
4866cde0 NP |
1698 | * finish_task_switch must be called after the context switch, paired |
1699 | * with a prepare_task_switch call before the context switch. | |
1700 | * finish_task_switch will reconcile locking set up by prepare_task_switch, | |
1701 | * and do any other architecture-specific cleanup actions. | |
1da177e4 LT |
1702 | * |
1703 | * Note that we may have delayed dropping an mm in context_switch(). If | |
1704 | * so, we finish that here outside of the runqueue lock. (Doing it | |
1705 | * with the lock held can cause deadlocks; see schedule() for | |
1706 | * details.) | |
1707 | */ | |
4866cde0 | 1708 | static inline void finish_task_switch(runqueue_t *rq, task_t *prev) |
1da177e4 LT |
1709 | __releases(rq->lock) |
1710 | { | |
1da177e4 LT |
1711 | struct mm_struct *mm = rq->prev_mm; |
1712 | unsigned long prev_task_flags; | |
1713 | ||
1714 | rq->prev_mm = NULL; | |
1715 | ||
1716 | /* | |
1717 | * A task struct has one reference for the use as "current". | |
1718 | * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and | |
1719 | * calls schedule one last time. The schedule call will never return, | |
1720 | * and the scheduled task must drop that reference. | |
1721 | * The test for EXIT_ZOMBIE must occur while the runqueue locks are | |
1722 | * still held, otherwise prev could be scheduled on another cpu, die | |
1723 | * there before we look at prev->state, and then the reference would | |
1724 | * be dropped twice. | |
1725 | * Manfred Spraul <manfred@colorfullife.com> | |
1726 | */ | |
1727 | prev_task_flags = prev->flags; | |
4866cde0 NP |
1728 | finish_arch_switch(prev); |
1729 | finish_lock_switch(rq, prev); | |
1da177e4 LT |
1730 | if (mm) |
1731 | mmdrop(mm); | |
c6fd91f0 | 1732 | if (unlikely(prev_task_flags & PF_DEAD)) { |
1733 | /* | |
1734 | * Remove function-return probe instances associated with this | |
1735 | * task and put them back on the free list. | |
1736 | */ | |
1737 | kprobe_flush_task(prev); | |
1da177e4 | 1738 | put_task_struct(prev); |
c6fd91f0 | 1739 | } |
1da177e4 LT |
1740 | } |
1741 | ||
1742 | /** | |
1743 | * schedule_tail - first thing a freshly forked thread must call. | |
1744 | * @prev: the thread we just switched away from. | |
1745 | */ | |
1746 | asmlinkage void schedule_tail(task_t *prev) | |
1747 | __releases(rq->lock) | |
1748 | { | |
4866cde0 NP |
1749 | runqueue_t *rq = this_rq(); |
1750 | finish_task_switch(rq, prev); | |
1751 | #ifdef __ARCH_WANT_UNLOCKED_CTXSW | |
1752 | /* In this case, finish_task_switch does not reenable preemption */ | |
1753 | preempt_enable(); | |
1754 | #endif | |
1da177e4 LT |
1755 | if (current->set_child_tid) |
1756 | put_user(current->pid, current->set_child_tid); | |
1757 | } | |
1758 | ||
1759 | /* | |
1760 | * context_switch - switch to the new MM and the new | |
1761 | * thread's register state. | |
1762 | */ | |
1763 | static inline | |
1764 | task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next) | |
1765 | { | |
1766 | struct mm_struct *mm = next->mm; | |
1767 | struct mm_struct *oldmm = prev->active_mm; | |
1768 | ||
1769 | if (unlikely(!mm)) { | |
1770 | next->active_mm = oldmm; | |
1771 | atomic_inc(&oldmm->mm_count); | |
1772 | enter_lazy_tlb(oldmm, next); | |
1773 | } else | |
1774 | switch_mm(oldmm, mm, next); | |
1775 | ||
1776 | if (unlikely(!prev->mm)) { | |
1777 | prev->active_mm = NULL; | |
1778 | WARN_ON(rq->prev_mm); | |
1779 | rq->prev_mm = oldmm; | |
1780 | } | |
1781 | ||
1782 | /* Here we just switch the register state and the stack. */ | |
1783 | switch_to(prev, next, prev); | |
1784 | ||
1785 | return prev; | |
1786 | } | |
1787 | ||
1788 | /* | |
1789 | * nr_running, nr_uninterruptible and nr_context_switches: | |
1790 | * | |
1791 | * externally visible scheduler statistics: current number of runnable | |
1792 | * threads, current number of uninterruptible-sleeping threads, total | |
1793 | * number of context switches performed since bootup. | |
1794 | */ | |
1795 | unsigned long nr_running(void) | |
1796 | { | |
1797 | unsigned long i, sum = 0; | |
1798 | ||
1799 | for_each_online_cpu(i) | |
1800 | sum += cpu_rq(i)->nr_running; | |
1801 | ||
1802 | return sum; | |
1803 | } | |
1804 | ||
1805 | unsigned long nr_uninterruptible(void) | |
1806 | { | |
1807 | unsigned long i, sum = 0; | |
1808 | ||
0a945022 | 1809 | for_each_possible_cpu(i) |
1da177e4 LT |
1810 | sum += cpu_rq(i)->nr_uninterruptible; |
1811 | ||
1812 | /* | |
1813 | * Since we read the counters lockless, it might be slightly | |
1814 | * inaccurate. Do not allow it to go below zero though: | |
1815 | */ | |
1816 | if (unlikely((long)sum < 0)) | |
1817 | sum = 0; | |
1818 | ||
1819 | return sum; | |
1820 | } | |
1821 | ||
1822 | unsigned long long nr_context_switches(void) | |
1823 | { | |
cc94abfc SR |
1824 | int i; |
1825 | unsigned long long sum = 0; | |
1da177e4 | 1826 | |
0a945022 | 1827 | for_each_possible_cpu(i) |
1da177e4 LT |
1828 | sum += cpu_rq(i)->nr_switches; |
1829 | ||
1830 | return sum; | |
1831 | } | |
1832 | ||
1833 | unsigned long nr_iowait(void) | |
1834 | { | |
1835 | unsigned long i, sum = 0; | |
1836 | ||
0a945022 | 1837 | for_each_possible_cpu(i) |
1da177e4 LT |
1838 | sum += atomic_read(&cpu_rq(i)->nr_iowait); |
1839 | ||
1840 | return sum; | |
1841 | } | |
1842 | ||
db1b1fef JS |
1843 | unsigned long nr_active(void) |
1844 | { | |
1845 | unsigned long i, running = 0, uninterruptible = 0; | |
1846 | ||
1847 | for_each_online_cpu(i) { | |
1848 | running += cpu_rq(i)->nr_running; | |
1849 | uninterruptible += cpu_rq(i)->nr_uninterruptible; | |
1850 | } | |
1851 | ||
1852 | if (unlikely((long)uninterruptible < 0)) | |
1853 | uninterruptible = 0; | |
1854 | ||
1855 | return running + uninterruptible; | |
1856 | } | |
1857 | ||
1da177e4 LT |
1858 | #ifdef CONFIG_SMP |
1859 | ||
1860 | /* | |
1861 | * double_rq_lock - safely lock two runqueues | |
1862 | * | |
1863 | * Note this does not disable interrupts like task_rq_lock, | |
1864 | * you need to do so manually before calling. | |
1865 | */ | |
1866 | static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2) | |
1867 | __acquires(rq1->lock) | |
1868 | __acquires(rq2->lock) | |
1869 | { | |
1870 | if (rq1 == rq2) { | |
1871 | spin_lock(&rq1->lock); | |
1872 | __acquire(rq2->lock); /* Fake it out ;) */ | |
1873 | } else { | |
c96d145e | 1874 | if (rq1 < rq2) { |
1da177e4 LT |
1875 | spin_lock(&rq1->lock); |
1876 | spin_lock(&rq2->lock); | |
1877 | } else { | |
1878 | spin_lock(&rq2->lock); | |
1879 | spin_lock(&rq1->lock); | |
1880 | } | |
1881 | } | |
1882 | } | |
1883 | ||
1884 | /* | |
1885 | * double_rq_unlock - safely unlock two runqueues | |
1886 | * | |
1887 | * Note this does not restore interrupts like task_rq_unlock, | |
1888 | * you need to do so manually after calling. | |
1889 | */ | |
1890 | static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2) | |
1891 | __releases(rq1->lock) | |
1892 | __releases(rq2->lock) | |
1893 | { | |
1894 | spin_unlock(&rq1->lock); | |
1895 | if (rq1 != rq2) | |
1896 | spin_unlock(&rq2->lock); | |
1897 | else | |
1898 | __release(rq2->lock); | |
1899 | } | |
1900 | ||
1901 | /* | |
1902 | * double_lock_balance - lock the busiest runqueue, this_rq is locked already. | |
1903 | */ | |
1904 | static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest) | |
1905 | __releases(this_rq->lock) | |
1906 | __acquires(busiest->lock) | |
1907 | __acquires(this_rq->lock) | |
1908 | { | |
1909 | if (unlikely(!spin_trylock(&busiest->lock))) { | |
c96d145e | 1910 | if (busiest < this_rq) { |
1da177e4 LT |
1911 | spin_unlock(&this_rq->lock); |
1912 | spin_lock(&busiest->lock); | |
1913 | spin_lock(&this_rq->lock); | |
1914 | } else | |
1915 | spin_lock(&busiest->lock); | |
1916 | } | |
1917 | } | |
1918 | ||
1da177e4 LT |
1919 | /* |
1920 | * If dest_cpu is allowed for this process, migrate the task to it. | |
1921 | * This is accomplished by forcing the cpu_allowed mask to only | |
1922 | * allow dest_cpu, which will force the cpu onto dest_cpu. Then | |
1923 | * the cpu_allowed mask is restored. | |
1924 | */ | |
1925 | static void sched_migrate_task(task_t *p, int dest_cpu) | |
1926 | { | |
1927 | migration_req_t req; | |
1928 | runqueue_t *rq; | |
1929 | unsigned long flags; | |
1930 | ||
1931 | rq = task_rq_lock(p, &flags); | |
1932 | if (!cpu_isset(dest_cpu, p->cpus_allowed) | |
1933 | || unlikely(cpu_is_offline(dest_cpu))) | |
1934 | goto out; | |
1935 | ||
1936 | /* force the process onto the specified CPU */ | |
1937 | if (migrate_task(p, dest_cpu, &req)) { | |
1938 | /* Need to wait for migration thread (might exit: take ref). */ | |
1939 | struct task_struct *mt = rq->migration_thread; | |
1940 | get_task_struct(mt); | |
1941 | task_rq_unlock(rq, &flags); | |
1942 | wake_up_process(mt); | |
1943 | put_task_struct(mt); | |
1944 | wait_for_completion(&req.done); | |
1945 | return; | |
1946 | } | |
1947 | out: | |
1948 | task_rq_unlock(rq, &flags); | |
1949 | } | |
1950 | ||
1951 | /* | |
476d139c NP |
1952 | * sched_exec - execve() is a valuable balancing opportunity, because at |
1953 | * this point the task has the smallest effective memory and cache footprint. | |
1da177e4 LT |
1954 | */ |
1955 | void sched_exec(void) | |
1956 | { | |
1da177e4 | 1957 | int new_cpu, this_cpu = get_cpu(); |
476d139c | 1958 | new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC); |
1da177e4 | 1959 | put_cpu(); |
476d139c NP |
1960 | if (new_cpu != this_cpu) |
1961 | sched_migrate_task(current, new_cpu); | |
1da177e4 LT |
1962 | } |
1963 | ||
1964 | /* | |
1965 | * pull_task - move a task from a remote runqueue to the local runqueue. | |
1966 | * Both runqueues must be locked. | |
1967 | */ | |
858119e1 | 1968 | static |
1da177e4 LT |
1969 | void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p, |
1970 | runqueue_t *this_rq, prio_array_t *this_array, int this_cpu) | |
1971 | { | |
1972 | dequeue_task(p, src_array); | |
2dd73a4f | 1973 | dec_nr_running(p, src_rq); |
1da177e4 | 1974 | set_task_cpu(p, this_cpu); |
2dd73a4f | 1975 | inc_nr_running(p, this_rq); |
1da177e4 LT |
1976 | enqueue_task(p, this_array); |
1977 | p->timestamp = (p->timestamp - src_rq->timestamp_last_tick) | |
1978 | + this_rq->timestamp_last_tick; | |
1979 | /* | |
1980 | * Note that idle threads have a prio of MAX_PRIO, for this test | |
1981 | * to be always true for them. | |
1982 | */ | |
1983 | if (TASK_PREEMPTS_CURR(p, this_rq)) | |
1984 | resched_task(this_rq->curr); | |
1985 | } | |
1986 | ||
1987 | /* | |
1988 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
1989 | */ | |
858119e1 | 1990 | static |
1da177e4 | 1991 | int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu, |
95cdf3b7 IM |
1992 | struct sched_domain *sd, enum idle_type idle, |
1993 | int *all_pinned) | |
1da177e4 LT |
1994 | { |
1995 | /* | |
1996 | * We do not migrate tasks that are: | |
1997 | * 1) running (obviously), or | |
1998 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
1999 | * 3) are cache-hot on their current CPU. | |
2000 | */ | |
1da177e4 LT |
2001 | if (!cpu_isset(this_cpu, p->cpus_allowed)) |
2002 | return 0; | |
81026794 NP |
2003 | *all_pinned = 0; |
2004 | ||
2005 | if (task_running(rq, p)) | |
2006 | return 0; | |
1da177e4 LT |
2007 | |
2008 | /* | |
2009 | * Aggressive migration if: | |
cafb20c1 | 2010 | * 1) task is cache cold, or |
1da177e4 LT |
2011 | * 2) too many balance attempts have failed. |
2012 | */ | |
2013 | ||
cafb20c1 | 2014 | if (sd->nr_balance_failed > sd->cache_nice_tries) |
1da177e4 LT |
2015 | return 1; |
2016 | ||
2017 | if (task_hot(p, rq->timestamp_last_tick, sd)) | |
81026794 | 2018 | return 0; |
1da177e4 LT |
2019 | return 1; |
2020 | } | |
2021 | ||
615052dc | 2022 | #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio) |
1da177e4 | 2023 | /* |
2dd73a4f PW |
2024 | * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted |
2025 | * load from busiest to this_rq, as part of a balancing operation within | |
2026 | * "domain". Returns the number of tasks moved. | |
1da177e4 LT |
2027 | * |
2028 | * Called with both runqueues locked. | |
2029 | */ | |
2030 | static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest, | |
2dd73a4f PW |
2031 | unsigned long max_nr_move, unsigned long max_load_move, |
2032 | struct sched_domain *sd, enum idle_type idle, | |
2033 | int *all_pinned) | |
1da177e4 LT |
2034 | { |
2035 | prio_array_t *array, *dst_array; | |
2036 | struct list_head *head, *curr; | |
615052dc PW |
2037 | int idx, pulled = 0, pinned = 0, this_best_prio, busiest_best_prio; |
2038 | int busiest_best_prio_seen; | |
2039 | int skip_for_load; /* skip the task based on weighted load issues */ | |
2dd73a4f | 2040 | long rem_load_move; |
1da177e4 LT |
2041 | task_t *tmp; |
2042 | ||
2dd73a4f | 2043 | if (max_nr_move == 0 || max_load_move == 0) |
1da177e4 LT |
2044 | goto out; |
2045 | ||
2dd73a4f | 2046 | rem_load_move = max_load_move; |
81026794 | 2047 | pinned = 1; |
615052dc PW |
2048 | this_best_prio = rq_best_prio(this_rq); |
2049 | busiest_best_prio = rq_best_prio(busiest); | |
2050 | /* | |
2051 | * Enable handling of the case where there is more than one task | |
2052 | * with the best priority. If the current running task is one | |
2053 | * of those with prio==busiest_best_prio we know it won't be moved | |
2054 | * and therefore it's safe to override the skip (based on load) of | |
2055 | * any task we find with that prio. | |
2056 | */ | |
2057 | busiest_best_prio_seen = busiest_best_prio == busiest->curr->prio; | |
81026794 | 2058 | |
1da177e4 LT |
2059 | /* |
2060 | * We first consider expired tasks. Those will likely not be | |
2061 | * executed in the near future, and they are most likely to | |
2062 | * be cache-cold, thus switching CPUs has the least effect | |
2063 | * on them. | |
2064 | */ | |
2065 | if (busiest->expired->nr_active) { | |
2066 | array = busiest->expired; | |
2067 | dst_array = this_rq->expired; | |
2068 | } else { | |
2069 | array = busiest->active; | |
2070 | dst_array = this_rq->active; | |
2071 | } | |
2072 | ||
2073 | new_array: | |
2074 | /* Start searching at priority 0: */ | |
2075 | idx = 0; | |
2076 | skip_bitmap: | |
2077 | if (!idx) | |
2078 | idx = sched_find_first_bit(array->bitmap); | |
2079 | else | |
2080 | idx = find_next_bit(array->bitmap, MAX_PRIO, idx); | |
2081 | if (idx >= MAX_PRIO) { | |
2082 | if (array == busiest->expired && busiest->active->nr_active) { | |
2083 | array = busiest->active; | |
2084 | dst_array = this_rq->active; | |
2085 | goto new_array; | |
2086 | } | |
2087 | goto out; | |
2088 | } | |
2089 | ||
2090 | head = array->queue + idx; | |
2091 | curr = head->prev; | |
2092 | skip_queue: | |
2093 | tmp = list_entry(curr, task_t, run_list); | |
2094 | ||
2095 | curr = curr->prev; | |
2096 | ||
50ddd969 PW |
2097 | /* |
2098 | * To help distribute high priority tasks accross CPUs we don't | |
2099 | * skip a task if it will be the highest priority task (i.e. smallest | |
2100 | * prio value) on its new queue regardless of its load weight | |
2101 | */ | |
615052dc PW |
2102 | skip_for_load = tmp->load_weight > rem_load_move; |
2103 | if (skip_for_load && idx < this_best_prio) | |
2104 | skip_for_load = !busiest_best_prio_seen && idx == busiest_best_prio; | |
2105 | if (skip_for_load || | |
2dd73a4f | 2106 | !can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) { |
615052dc | 2107 | busiest_best_prio_seen |= idx == busiest_best_prio; |
1da177e4 LT |
2108 | if (curr != head) |
2109 | goto skip_queue; | |
2110 | idx++; | |
2111 | goto skip_bitmap; | |
2112 | } | |
2113 | ||
2114 | #ifdef CONFIG_SCHEDSTATS | |
2115 | if (task_hot(tmp, busiest->timestamp_last_tick, sd)) | |
2116 | schedstat_inc(sd, lb_hot_gained[idle]); | |
2117 | #endif | |
2118 | ||
2119 | pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu); | |
2120 | pulled++; | |
2dd73a4f | 2121 | rem_load_move -= tmp->load_weight; |
1da177e4 | 2122 | |
2dd73a4f PW |
2123 | /* |
2124 | * We only want to steal up to the prescribed number of tasks | |
2125 | * and the prescribed amount of weighted load. | |
2126 | */ | |
2127 | if (pulled < max_nr_move && rem_load_move > 0) { | |
615052dc PW |
2128 | if (idx < this_best_prio) |
2129 | this_best_prio = idx; | |
1da177e4 LT |
2130 | if (curr != head) |
2131 | goto skip_queue; | |
2132 | idx++; | |
2133 | goto skip_bitmap; | |
2134 | } | |
2135 | out: | |
2136 | /* | |
2137 | * Right now, this is the only place pull_task() is called, | |
2138 | * so we can safely collect pull_task() stats here rather than | |
2139 | * inside pull_task(). | |
2140 | */ | |
2141 | schedstat_add(sd, lb_gained[idle], pulled); | |
81026794 NP |
2142 | |
2143 | if (all_pinned) | |
2144 | *all_pinned = pinned; | |
1da177e4 LT |
2145 | return pulled; |
2146 | } | |
2147 | ||
2148 | /* | |
2149 | * find_busiest_group finds and returns the busiest CPU group within the | |
2dd73a4f | 2150 | * domain. It calculates and returns the amount of weighted load which should be |
1da177e4 LT |
2151 | * moved to restore balance via the imbalance parameter. |
2152 | */ | |
2153 | static struct sched_group * | |
2154 | find_busiest_group(struct sched_domain *sd, int this_cpu, | |
5969fe06 | 2155 | unsigned long *imbalance, enum idle_type idle, int *sd_idle) |
1da177e4 LT |
2156 | { |
2157 | struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups; | |
2158 | unsigned long max_load, avg_load, total_load, this_load, total_pwr; | |
0c117f1b | 2159 | unsigned long max_pull; |
2dd73a4f PW |
2160 | unsigned long busiest_load_per_task, busiest_nr_running; |
2161 | unsigned long this_load_per_task, this_nr_running; | |
7897986b | 2162 | int load_idx; |
5c45bf27 SS |
2163 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
2164 | int power_savings_balance = 1; | |
2165 | unsigned long leader_nr_running = 0, min_load_per_task = 0; | |
2166 | unsigned long min_nr_running = ULONG_MAX; | |
2167 | struct sched_group *group_min = NULL, *group_leader = NULL; | |
2168 | #endif | |
1da177e4 LT |
2169 | |
2170 | max_load = this_load = total_load = total_pwr = 0; | |
2dd73a4f PW |
2171 | busiest_load_per_task = busiest_nr_running = 0; |
2172 | this_load_per_task = this_nr_running = 0; | |
7897986b NP |
2173 | if (idle == NOT_IDLE) |
2174 | load_idx = sd->busy_idx; | |
2175 | else if (idle == NEWLY_IDLE) | |
2176 | load_idx = sd->newidle_idx; | |
2177 | else | |
2178 | load_idx = sd->idle_idx; | |
1da177e4 LT |
2179 | |
2180 | do { | |
5c45bf27 | 2181 | unsigned long load, group_capacity; |
1da177e4 LT |
2182 | int local_group; |
2183 | int i; | |
2dd73a4f | 2184 | unsigned long sum_nr_running, sum_weighted_load; |
1da177e4 LT |
2185 | |
2186 | local_group = cpu_isset(this_cpu, group->cpumask); | |
2187 | ||
2188 | /* Tally up the load of all CPUs in the group */ | |
2dd73a4f | 2189 | sum_weighted_load = sum_nr_running = avg_load = 0; |
1da177e4 LT |
2190 | |
2191 | for_each_cpu_mask(i, group->cpumask) { | |
2dd73a4f PW |
2192 | runqueue_t *rq = cpu_rq(i); |
2193 | ||
5969fe06 NP |
2194 | if (*sd_idle && !idle_cpu(i)) |
2195 | *sd_idle = 0; | |
2196 | ||
1da177e4 LT |
2197 | /* Bias balancing toward cpus of our domain */ |
2198 | if (local_group) | |
a2000572 | 2199 | load = target_load(i, load_idx); |
1da177e4 | 2200 | else |
a2000572 | 2201 | load = source_load(i, load_idx); |
1da177e4 LT |
2202 | |
2203 | avg_load += load; | |
2dd73a4f PW |
2204 | sum_nr_running += rq->nr_running; |
2205 | sum_weighted_load += rq->raw_weighted_load; | |
1da177e4 LT |
2206 | } |
2207 | ||
2208 | total_load += avg_load; | |
2209 | total_pwr += group->cpu_power; | |
2210 | ||
2211 | /* Adjust by relative CPU power of the group */ | |
2212 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
2213 | ||
5c45bf27 SS |
2214 | group_capacity = group->cpu_power / SCHED_LOAD_SCALE; |
2215 | ||
1da177e4 LT |
2216 | if (local_group) { |
2217 | this_load = avg_load; | |
2218 | this = group; | |
2dd73a4f PW |
2219 | this_nr_running = sum_nr_running; |
2220 | this_load_per_task = sum_weighted_load; | |
2221 | } else if (avg_load > max_load && | |
5c45bf27 | 2222 | sum_nr_running > group_capacity) { |
1da177e4 LT |
2223 | max_load = avg_load; |
2224 | busiest = group; | |
2dd73a4f PW |
2225 | busiest_nr_running = sum_nr_running; |
2226 | busiest_load_per_task = sum_weighted_load; | |
1da177e4 | 2227 | } |
5c45bf27 SS |
2228 | |
2229 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
2230 | /* | |
2231 | * Busy processors will not participate in power savings | |
2232 | * balance. | |
2233 | */ | |
2234 | if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE)) | |
2235 | goto group_next; | |
2236 | ||
2237 | /* | |
2238 | * If the local group is idle or completely loaded | |
2239 | * no need to do power savings balance at this domain | |
2240 | */ | |
2241 | if (local_group && (this_nr_running >= group_capacity || | |
2242 | !this_nr_running)) | |
2243 | power_savings_balance = 0; | |
2244 | ||
2245 | /* | |
2246 | * If a group is already running at full capacity or idle, | |
2247 | * don't include that group in power savings calculations | |
2248 | */ | |
2249 | if (!power_savings_balance || sum_nr_running >= group_capacity | |
2250 | || !sum_nr_running) | |
2251 | goto group_next; | |
2252 | ||
2253 | /* | |
2254 | * Calculate the group which has the least non-idle load. | |
2255 | * This is the group from where we need to pick up the load | |
2256 | * for saving power | |
2257 | */ | |
2258 | if ((sum_nr_running < min_nr_running) || | |
2259 | (sum_nr_running == min_nr_running && | |
2260 | first_cpu(group->cpumask) < | |
2261 | first_cpu(group_min->cpumask))) { | |
2262 | group_min = group; | |
2263 | min_nr_running = sum_nr_running; | |
2264 | min_load_per_task = sum_weighted_load / | |
2265 | sum_nr_running; | |
2266 | } | |
2267 | ||
2268 | /* | |
2269 | * Calculate the group which is almost near its | |
2270 | * capacity but still has some space to pick up some load | |
2271 | * from other group and save more power | |
2272 | */ | |
2273 | if (sum_nr_running <= group_capacity - 1) | |
2274 | if (sum_nr_running > leader_nr_running || | |
2275 | (sum_nr_running == leader_nr_running && | |
2276 | first_cpu(group->cpumask) > | |
2277 | first_cpu(group_leader->cpumask))) { | |
2278 | group_leader = group; | |
2279 | leader_nr_running = sum_nr_running; | |
2280 | } | |
2281 | ||
2282 | group_next: | |
2283 | #endif | |
1da177e4 LT |
2284 | group = group->next; |
2285 | } while (group != sd->groups); | |
2286 | ||
2dd73a4f | 2287 | if (!busiest || this_load >= max_load || busiest_nr_running == 0) |
1da177e4 LT |
2288 | goto out_balanced; |
2289 | ||
2290 | avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr; | |
2291 | ||
2292 | if (this_load >= avg_load || | |
2293 | 100*max_load <= sd->imbalance_pct*this_load) | |
2294 | goto out_balanced; | |
2295 | ||
2dd73a4f | 2296 | busiest_load_per_task /= busiest_nr_running; |
1da177e4 LT |
2297 | /* |
2298 | * We're trying to get all the cpus to the average_load, so we don't | |
2299 | * want to push ourselves above the average load, nor do we wish to | |
2300 | * reduce the max loaded cpu below the average load, as either of these | |
2301 | * actions would just result in more rebalancing later, and ping-pong | |
2302 | * tasks around. Thus we look for the minimum possible imbalance. | |
2303 | * Negative imbalances (*we* are more loaded than anyone else) will | |
2304 | * be counted as no imbalance for these purposes -- we can't fix that | |
2305 | * by pulling tasks to us. Be careful of negative numbers as they'll | |
2306 | * appear as very large values with unsigned longs. | |
2307 | */ | |
2dd73a4f PW |
2308 | if (max_load <= busiest_load_per_task) |
2309 | goto out_balanced; | |
2310 | ||
2311 | /* | |
2312 | * In the presence of smp nice balancing, certain scenarios can have | |
2313 | * max load less than avg load(as we skip the groups at or below | |
2314 | * its cpu_power, while calculating max_load..) | |
2315 | */ | |
2316 | if (max_load < avg_load) { | |
2317 | *imbalance = 0; | |
2318 | goto small_imbalance; | |
2319 | } | |
0c117f1b SS |
2320 | |
2321 | /* Don't want to pull so many tasks that a group would go idle */ | |
2dd73a4f | 2322 | max_pull = min(max_load - avg_load, max_load - busiest_load_per_task); |
0c117f1b | 2323 | |
1da177e4 | 2324 | /* How much load to actually move to equalise the imbalance */ |
0c117f1b | 2325 | *imbalance = min(max_pull * busiest->cpu_power, |
1da177e4 LT |
2326 | (avg_load - this_load) * this->cpu_power) |
2327 | / SCHED_LOAD_SCALE; | |
2328 | ||
2dd73a4f PW |
2329 | /* |
2330 | * if *imbalance is less than the average load per runnable task | |
2331 | * there is no gaurantee that any tasks will be moved so we'll have | |
2332 | * a think about bumping its value to force at least one task to be | |
2333 | * moved | |
2334 | */ | |
2335 | if (*imbalance < busiest_load_per_task) { | |
2336 | unsigned long pwr_now, pwr_move; | |
1da177e4 | 2337 | unsigned long tmp; |
2dd73a4f PW |
2338 | unsigned int imbn; |
2339 | ||
2340 | small_imbalance: | |
2341 | pwr_move = pwr_now = 0; | |
2342 | imbn = 2; | |
2343 | if (this_nr_running) { | |
2344 | this_load_per_task /= this_nr_running; | |
2345 | if (busiest_load_per_task > this_load_per_task) | |
2346 | imbn = 1; | |
2347 | } else | |
2348 | this_load_per_task = SCHED_LOAD_SCALE; | |
1da177e4 | 2349 | |
2dd73a4f PW |
2350 | if (max_load - this_load >= busiest_load_per_task * imbn) { |
2351 | *imbalance = busiest_load_per_task; | |
1da177e4 LT |
2352 | return busiest; |
2353 | } | |
2354 | ||
2355 | /* | |
2356 | * OK, we don't have enough imbalance to justify moving tasks, | |
2357 | * however we may be able to increase total CPU power used by | |
2358 | * moving them. | |
2359 | */ | |
2360 | ||
2dd73a4f PW |
2361 | pwr_now += busiest->cpu_power * |
2362 | min(busiest_load_per_task, max_load); | |
2363 | pwr_now += this->cpu_power * | |
2364 | min(this_load_per_task, this_load); | |
1da177e4 LT |
2365 | pwr_now /= SCHED_LOAD_SCALE; |
2366 | ||
2367 | /* Amount of load we'd subtract */ | |
2dd73a4f | 2368 | tmp = busiest_load_per_task*SCHED_LOAD_SCALE/busiest->cpu_power; |
1da177e4 | 2369 | if (max_load > tmp) |
2dd73a4f PW |
2370 | pwr_move += busiest->cpu_power * |
2371 | min(busiest_load_per_task, max_load - tmp); | |
1da177e4 LT |
2372 | |
2373 | /* Amount of load we'd add */ | |
2374 | if (max_load*busiest->cpu_power < | |
2dd73a4f | 2375 | busiest_load_per_task*SCHED_LOAD_SCALE) |
1da177e4 LT |
2376 | tmp = max_load*busiest->cpu_power/this->cpu_power; |
2377 | else | |
2dd73a4f PW |
2378 | tmp = busiest_load_per_task*SCHED_LOAD_SCALE/this->cpu_power; |
2379 | pwr_move += this->cpu_power*min(this_load_per_task, this_load + tmp); | |
1da177e4 LT |
2380 | pwr_move /= SCHED_LOAD_SCALE; |
2381 | ||
2382 | /* Move if we gain throughput */ | |
2383 | if (pwr_move <= pwr_now) | |
2384 | goto out_balanced; | |
2385 | ||
2dd73a4f | 2386 | *imbalance = busiest_load_per_task; |
1da177e4 LT |
2387 | } |
2388 | ||
1da177e4 LT |
2389 | return busiest; |
2390 | ||
2391 | out_balanced: | |
5c45bf27 SS |
2392 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
2393 | if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE)) | |
2394 | goto ret; | |
1da177e4 | 2395 | |
5c45bf27 SS |
2396 | if (this == group_leader && group_leader != group_min) { |
2397 | *imbalance = min_load_per_task; | |
2398 | return group_min; | |
2399 | } | |
2400 | ret: | |
2401 | #endif | |
1da177e4 LT |
2402 | *imbalance = 0; |
2403 | return NULL; | |
2404 | } | |
2405 | ||
2406 | /* | |
2407 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
2408 | */ | |
b910472d | 2409 | static runqueue_t *find_busiest_queue(struct sched_group *group, |
2dd73a4f | 2410 | enum idle_type idle, unsigned long imbalance) |
1da177e4 | 2411 | { |
2dd73a4f PW |
2412 | unsigned long max_load = 0; |
2413 | runqueue_t *busiest = NULL, *rqi; | |
1da177e4 LT |
2414 | int i; |
2415 | ||
2416 | for_each_cpu_mask(i, group->cpumask) { | |
2dd73a4f PW |
2417 | rqi = cpu_rq(i); |
2418 | ||
2419 | if (rqi->nr_running == 1 && rqi->raw_weighted_load > imbalance) | |
2420 | continue; | |
1da177e4 | 2421 | |
2dd73a4f PW |
2422 | if (rqi->raw_weighted_load > max_load) { |
2423 | max_load = rqi->raw_weighted_load; | |
2424 | busiest = rqi; | |
1da177e4 LT |
2425 | } |
2426 | } | |
2427 | ||
2428 | return busiest; | |
2429 | } | |
2430 | ||
77391d71 NP |
2431 | /* |
2432 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
2433 | * so long as it is large enough. | |
2434 | */ | |
2435 | #define MAX_PINNED_INTERVAL 512 | |
2436 | ||
2dd73a4f | 2437 | #define minus_1_or_zero(n) ((n) > 0 ? (n) - 1 : 0) |
1da177e4 LT |
2438 | /* |
2439 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
2440 | * tasks if there is an imbalance. | |
2441 | * | |
2442 | * Called with this_rq unlocked. | |
2443 | */ | |
2444 | static int load_balance(int this_cpu, runqueue_t *this_rq, | |
2445 | struct sched_domain *sd, enum idle_type idle) | |
2446 | { | |
2447 | struct sched_group *group; | |
2448 | runqueue_t *busiest; | |
2449 | unsigned long imbalance; | |
77391d71 | 2450 | int nr_moved, all_pinned = 0; |
81026794 | 2451 | int active_balance = 0; |
5969fe06 NP |
2452 | int sd_idle = 0; |
2453 | ||
5c45bf27 SS |
2454 | if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER && |
2455 | !sched_smt_power_savings) | |
5969fe06 | 2456 | sd_idle = 1; |
1da177e4 | 2457 | |
1da177e4 LT |
2458 | schedstat_inc(sd, lb_cnt[idle]); |
2459 | ||
5969fe06 | 2460 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle); |
1da177e4 LT |
2461 | if (!group) { |
2462 | schedstat_inc(sd, lb_nobusyg[idle]); | |
2463 | goto out_balanced; | |
2464 | } | |
2465 | ||
2dd73a4f | 2466 | busiest = find_busiest_queue(group, idle, imbalance); |
1da177e4 LT |
2467 | if (!busiest) { |
2468 | schedstat_inc(sd, lb_nobusyq[idle]); | |
2469 | goto out_balanced; | |
2470 | } | |
2471 | ||
db935dbd | 2472 | BUG_ON(busiest == this_rq); |
1da177e4 LT |
2473 | |
2474 | schedstat_add(sd, lb_imbalance[idle], imbalance); | |
2475 | ||
2476 | nr_moved = 0; | |
2477 | if (busiest->nr_running > 1) { | |
2478 | /* | |
2479 | * Attempt to move tasks. If find_busiest_group has found | |
2480 | * an imbalance but busiest->nr_running <= 1, the group is | |
2481 | * still unbalanced. nr_moved simply stays zero, so it is | |
2482 | * correctly treated as an imbalance. | |
2483 | */ | |
e17224bf | 2484 | double_rq_lock(this_rq, busiest); |
1da177e4 | 2485 | nr_moved = move_tasks(this_rq, this_cpu, busiest, |
2dd73a4f | 2486 | minus_1_or_zero(busiest->nr_running), |
d6d5cfaf | 2487 | imbalance, sd, idle, &all_pinned); |
e17224bf | 2488 | double_rq_unlock(this_rq, busiest); |
81026794 NP |
2489 | |
2490 | /* All tasks on this runqueue were pinned by CPU affinity */ | |
2491 | if (unlikely(all_pinned)) | |
2492 | goto out_balanced; | |
1da177e4 | 2493 | } |
81026794 | 2494 | |
1da177e4 LT |
2495 | if (!nr_moved) { |
2496 | schedstat_inc(sd, lb_failed[idle]); | |
2497 | sd->nr_balance_failed++; | |
2498 | ||
2499 | if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) { | |
1da177e4 LT |
2500 | |
2501 | spin_lock(&busiest->lock); | |
fa3b6ddc SS |
2502 | |
2503 | /* don't kick the migration_thread, if the curr | |
2504 | * task on busiest cpu can't be moved to this_cpu | |
2505 | */ | |
2506 | if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) { | |
2507 | spin_unlock(&busiest->lock); | |
2508 | all_pinned = 1; | |
2509 | goto out_one_pinned; | |
2510 | } | |
2511 | ||
1da177e4 LT |
2512 | if (!busiest->active_balance) { |
2513 | busiest->active_balance = 1; | |
2514 | busiest->push_cpu = this_cpu; | |
81026794 | 2515 | active_balance = 1; |
1da177e4 LT |
2516 | } |
2517 | spin_unlock(&busiest->lock); | |
81026794 | 2518 | if (active_balance) |
1da177e4 LT |
2519 | wake_up_process(busiest->migration_thread); |
2520 | ||
2521 | /* | |
2522 | * We've kicked active balancing, reset the failure | |
2523 | * counter. | |
2524 | */ | |
39507451 | 2525 | sd->nr_balance_failed = sd->cache_nice_tries+1; |
1da177e4 | 2526 | } |
81026794 | 2527 | } else |
1da177e4 LT |
2528 | sd->nr_balance_failed = 0; |
2529 | ||
81026794 | 2530 | if (likely(!active_balance)) { |
1da177e4 LT |
2531 | /* We were unbalanced, so reset the balancing interval */ |
2532 | sd->balance_interval = sd->min_interval; | |
81026794 NP |
2533 | } else { |
2534 | /* | |
2535 | * If we've begun active balancing, start to back off. This | |
2536 | * case may not be covered by the all_pinned logic if there | |
2537 | * is only 1 task on the busy runqueue (because we don't call | |
2538 | * move_tasks). | |
2539 | */ | |
2540 | if (sd->balance_interval < sd->max_interval) | |
2541 | sd->balance_interval *= 2; | |
1da177e4 LT |
2542 | } |
2543 | ||
5c45bf27 SS |
2544 | if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
2545 | !sched_smt_power_savings) | |
5969fe06 | 2546 | return -1; |
1da177e4 LT |
2547 | return nr_moved; |
2548 | ||
2549 | out_balanced: | |
1da177e4 LT |
2550 | schedstat_inc(sd, lb_balanced[idle]); |
2551 | ||
16cfb1c0 | 2552 | sd->nr_balance_failed = 0; |
fa3b6ddc SS |
2553 | |
2554 | out_one_pinned: | |
1da177e4 | 2555 | /* tune up the balancing interval */ |
77391d71 NP |
2556 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || |
2557 | (sd->balance_interval < sd->max_interval)) | |
1da177e4 LT |
2558 | sd->balance_interval *= 2; |
2559 | ||
5c45bf27 | 2560 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && !sched_smt_power_savings) |
5969fe06 | 2561 | return -1; |
1da177e4 LT |
2562 | return 0; |
2563 | } | |
2564 | ||
2565 | /* | |
2566 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
2567 | * tasks if there is an imbalance. | |
2568 | * | |
2569 | * Called from schedule when this_rq is about to become idle (NEWLY_IDLE). | |
2570 | * this_rq is locked. | |
2571 | */ | |
2572 | static int load_balance_newidle(int this_cpu, runqueue_t *this_rq, | |
2573 | struct sched_domain *sd) | |
2574 | { | |
2575 | struct sched_group *group; | |
2576 | runqueue_t *busiest = NULL; | |
2577 | unsigned long imbalance; | |
2578 | int nr_moved = 0; | |
5969fe06 NP |
2579 | int sd_idle = 0; |
2580 | ||
5c45bf27 | 2581 | if (sd->flags & SD_SHARE_CPUPOWER && !sched_smt_power_savings) |
5969fe06 | 2582 | sd_idle = 1; |
1da177e4 LT |
2583 | |
2584 | schedstat_inc(sd, lb_cnt[NEWLY_IDLE]); | |
5969fe06 | 2585 | group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE, &sd_idle); |
1da177e4 | 2586 | if (!group) { |
1da177e4 | 2587 | schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]); |
16cfb1c0 | 2588 | goto out_balanced; |
1da177e4 LT |
2589 | } |
2590 | ||
2dd73a4f | 2591 | busiest = find_busiest_queue(group, NEWLY_IDLE, imbalance); |
db935dbd | 2592 | if (!busiest) { |
1da177e4 | 2593 | schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]); |
16cfb1c0 | 2594 | goto out_balanced; |
1da177e4 LT |
2595 | } |
2596 | ||
db935dbd NP |
2597 | BUG_ON(busiest == this_rq); |
2598 | ||
1da177e4 | 2599 | schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance); |
d6d5cfaf NP |
2600 | |
2601 | nr_moved = 0; | |
2602 | if (busiest->nr_running > 1) { | |
2603 | /* Attempt to move tasks */ | |
2604 | double_lock_balance(this_rq, busiest); | |
2605 | nr_moved = move_tasks(this_rq, this_cpu, busiest, | |
2dd73a4f | 2606 | minus_1_or_zero(busiest->nr_running), |
81026794 | 2607 | imbalance, sd, NEWLY_IDLE, NULL); |
d6d5cfaf NP |
2608 | spin_unlock(&busiest->lock); |
2609 | } | |
2610 | ||
5969fe06 | 2611 | if (!nr_moved) { |
1da177e4 | 2612 | schedstat_inc(sd, lb_failed[NEWLY_IDLE]); |
5969fe06 NP |
2613 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
2614 | return -1; | |
2615 | } else | |
16cfb1c0 | 2616 | sd->nr_balance_failed = 0; |
1da177e4 | 2617 | |
1da177e4 | 2618 | return nr_moved; |
16cfb1c0 NP |
2619 | |
2620 | out_balanced: | |
2621 | schedstat_inc(sd, lb_balanced[NEWLY_IDLE]); | |
5c45bf27 | 2622 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && !sched_smt_power_savings) |
5969fe06 | 2623 | return -1; |
16cfb1c0 NP |
2624 | sd->nr_balance_failed = 0; |
2625 | return 0; | |
1da177e4 LT |
2626 | } |
2627 | ||
2628 | /* | |
2629 | * idle_balance is called by schedule() if this_cpu is about to become | |
2630 | * idle. Attempts to pull tasks from other CPUs. | |
2631 | */ | |
858119e1 | 2632 | static void idle_balance(int this_cpu, runqueue_t *this_rq) |
1da177e4 LT |
2633 | { |
2634 | struct sched_domain *sd; | |
2635 | ||
2636 | for_each_domain(this_cpu, sd) { | |
2637 | if (sd->flags & SD_BALANCE_NEWIDLE) { | |
2638 | if (load_balance_newidle(this_cpu, this_rq, sd)) { | |
2639 | /* We've pulled tasks over so stop searching */ | |
2640 | break; | |
2641 | } | |
2642 | } | |
2643 | } | |
2644 | } | |
2645 | ||
2646 | /* | |
2647 | * active_load_balance is run by migration threads. It pushes running tasks | |
2648 | * off the busiest CPU onto idle CPUs. It requires at least 1 task to be | |
2649 | * running on each physical CPU where possible, and avoids physical / | |
2650 | * logical imbalances. | |
2651 | * | |
2652 | * Called with busiest_rq locked. | |
2653 | */ | |
2654 | static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu) | |
2655 | { | |
2656 | struct sched_domain *sd; | |
1da177e4 | 2657 | runqueue_t *target_rq; |
39507451 NP |
2658 | int target_cpu = busiest_rq->push_cpu; |
2659 | ||
2660 | if (busiest_rq->nr_running <= 1) | |
2661 | /* no task to move */ | |
2662 | return; | |
2663 | ||
2664 | target_rq = cpu_rq(target_cpu); | |
1da177e4 LT |
2665 | |
2666 | /* | |
39507451 NP |
2667 | * This condition is "impossible", if it occurs |
2668 | * we need to fix it. Originally reported by | |
2669 | * Bjorn Helgaas on a 128-cpu setup. | |
1da177e4 | 2670 | */ |
39507451 | 2671 | BUG_ON(busiest_rq == target_rq); |
1da177e4 | 2672 | |
39507451 NP |
2673 | /* move a task from busiest_rq to target_rq */ |
2674 | double_lock_balance(busiest_rq, target_rq); | |
2675 | ||
2676 | /* Search for an sd spanning us and the target CPU. */ | |
c96d145e | 2677 | for_each_domain(target_cpu, sd) { |
39507451 NP |
2678 | if ((sd->flags & SD_LOAD_BALANCE) && |
2679 | cpu_isset(busiest_cpu, sd->span)) | |
2680 | break; | |
c96d145e | 2681 | } |
39507451 NP |
2682 | |
2683 | if (unlikely(sd == NULL)) | |
2684 | goto out; | |
2685 | ||
2686 | schedstat_inc(sd, alb_cnt); | |
2687 | ||
2dd73a4f PW |
2688 | if (move_tasks(target_rq, target_cpu, busiest_rq, 1, |
2689 | RTPRIO_TO_LOAD_WEIGHT(100), sd, SCHED_IDLE, NULL)) | |
39507451 NP |
2690 | schedstat_inc(sd, alb_pushed); |
2691 | else | |
2692 | schedstat_inc(sd, alb_failed); | |
2693 | out: | |
2694 | spin_unlock(&target_rq->lock); | |
1da177e4 LT |
2695 | } |
2696 | ||
2697 | /* | |
2698 | * rebalance_tick will get called every timer tick, on every CPU. | |
2699 | * | |
2700 | * It checks each scheduling domain to see if it is due to be balanced, | |
2701 | * and initiates a balancing operation if so. | |
2702 | * | |
2703 | * Balancing parameters are set up in arch_init_sched_domains. | |
2704 | */ | |
2705 | ||
2706 | /* Don't have all balancing operations going off at once */ | |
2707 | #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS) | |
2708 | ||
2709 | static void rebalance_tick(int this_cpu, runqueue_t *this_rq, | |
2710 | enum idle_type idle) | |
2711 | { | |
2712 | unsigned long old_load, this_load; | |
2713 | unsigned long j = jiffies + CPU_OFFSET(this_cpu); | |
2714 | struct sched_domain *sd; | |
7897986b | 2715 | int i; |
1da177e4 | 2716 | |
2dd73a4f | 2717 | this_load = this_rq->raw_weighted_load; |
7897986b NP |
2718 | /* Update our load */ |
2719 | for (i = 0; i < 3; i++) { | |
2720 | unsigned long new_load = this_load; | |
2721 | int scale = 1 << i; | |
2722 | old_load = this_rq->cpu_load[i]; | |
2723 | /* | |
2724 | * Round up the averaging division if load is increasing. This | |
2725 | * prevents us from getting stuck on 9 if the load is 10, for | |
2726 | * example. | |
2727 | */ | |
2728 | if (new_load > old_load) | |
2729 | new_load += scale-1; | |
2730 | this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale; | |
2731 | } | |
1da177e4 LT |
2732 | |
2733 | for_each_domain(this_cpu, sd) { | |
2734 | unsigned long interval; | |
2735 | ||
2736 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
2737 | continue; | |
2738 | ||
2739 | interval = sd->balance_interval; | |
2740 | if (idle != SCHED_IDLE) | |
2741 | interval *= sd->busy_factor; | |
2742 | ||
2743 | /* scale ms to jiffies */ | |
2744 | interval = msecs_to_jiffies(interval); | |
2745 | if (unlikely(!interval)) | |
2746 | interval = 1; | |
2747 | ||
2748 | if (j - sd->last_balance >= interval) { | |
2749 | if (load_balance(this_cpu, this_rq, sd, idle)) { | |
fa3b6ddc SS |
2750 | /* |
2751 | * We've pulled tasks over so either we're no | |
5969fe06 NP |
2752 | * longer idle, or one of our SMT siblings is |
2753 | * not idle. | |
2754 | */ | |
1da177e4 LT |
2755 | idle = NOT_IDLE; |
2756 | } | |
2757 | sd->last_balance += interval; | |
2758 | } | |
2759 | } | |
2760 | } | |
2761 | #else | |
2762 | /* | |
2763 | * on UP we do not need to balance between CPUs: | |
2764 | */ | |
2765 | static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle) | |
2766 | { | |
2767 | } | |
2768 | static inline void idle_balance(int cpu, runqueue_t *rq) | |
2769 | { | |
2770 | } | |
2771 | #endif | |
2772 | ||
2773 | static inline int wake_priority_sleeper(runqueue_t *rq) | |
2774 | { | |
2775 | int ret = 0; | |
2776 | #ifdef CONFIG_SCHED_SMT | |
2777 | spin_lock(&rq->lock); | |
2778 | /* | |
2779 | * If an SMT sibling task has been put to sleep for priority | |
2780 | * reasons reschedule the idle task to see if it can now run. | |
2781 | */ | |
2782 | if (rq->nr_running) { | |
2783 | resched_task(rq->idle); | |
2784 | ret = 1; | |
2785 | } | |
2786 | spin_unlock(&rq->lock); | |
2787 | #endif | |
2788 | return ret; | |
2789 | } | |
2790 | ||
2791 | DEFINE_PER_CPU(struct kernel_stat, kstat); | |
2792 | ||
2793 | EXPORT_PER_CPU_SYMBOL(kstat); | |
2794 | ||
2795 | /* | |
2796 | * This is called on clock ticks and on context switches. | |
2797 | * Bank in p->sched_time the ns elapsed since the last tick or switch. | |
2798 | */ | |
2799 | static inline void update_cpu_clock(task_t *p, runqueue_t *rq, | |
2800 | unsigned long long now) | |
2801 | { | |
2802 | unsigned long long last = max(p->timestamp, rq->timestamp_last_tick); | |
2803 | p->sched_time += now - last; | |
2804 | } | |
2805 | ||
2806 | /* | |
2807 | * Return current->sched_time plus any more ns on the sched_clock | |
2808 | * that have not yet been banked. | |
2809 | */ | |
2810 | unsigned long long current_sched_time(const task_t *tsk) | |
2811 | { | |
2812 | unsigned long long ns; | |
2813 | unsigned long flags; | |
2814 | local_irq_save(flags); | |
2815 | ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick); | |
2816 | ns = tsk->sched_time + (sched_clock() - ns); | |
2817 | local_irq_restore(flags); | |
2818 | return ns; | |
2819 | } | |
2820 | ||
f1adad78 LT |
2821 | /* |
2822 | * We place interactive tasks back into the active array, if possible. | |
2823 | * | |
2824 | * To guarantee that this does not starve expired tasks we ignore the | |
2825 | * interactivity of a task if the first expired task had to wait more | |
2826 | * than a 'reasonable' amount of time. This deadline timeout is | |
2827 | * load-dependent, as the frequency of array switched decreases with | |
2828 | * increasing number of running tasks. We also ignore the interactivity | |
2829 | * if a better static_prio task has expired: | |
2830 | */ | |
2831 | #define EXPIRED_STARVING(rq) \ | |
2832 | ((STARVATION_LIMIT && ((rq)->expired_timestamp && \ | |
2833 | (jiffies - (rq)->expired_timestamp >= \ | |
2834 | STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \ | |
2835 | ((rq)->curr->static_prio > (rq)->best_expired_prio)) | |
2836 | ||
1da177e4 LT |
2837 | /* |
2838 | * Account user cpu time to a process. | |
2839 | * @p: the process that the cpu time gets accounted to | |
2840 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
2841 | * @cputime: the cpu time spent in user space since the last update | |
2842 | */ | |
2843 | void account_user_time(struct task_struct *p, cputime_t cputime) | |
2844 | { | |
2845 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
2846 | cputime64_t tmp; | |
2847 | ||
2848 | p->utime = cputime_add(p->utime, cputime); | |
2849 | ||
2850 | /* Add user time to cpustat. */ | |
2851 | tmp = cputime_to_cputime64(cputime); | |
2852 | if (TASK_NICE(p) > 0) | |
2853 | cpustat->nice = cputime64_add(cpustat->nice, tmp); | |
2854 | else | |
2855 | cpustat->user = cputime64_add(cpustat->user, tmp); | |
2856 | } | |
2857 | ||
2858 | /* | |
2859 | * Account system cpu time to a process. | |
2860 | * @p: the process that the cpu time gets accounted to | |
2861 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
2862 | * @cputime: the cpu time spent in kernel space since the last update | |
2863 | */ | |
2864 | void account_system_time(struct task_struct *p, int hardirq_offset, | |
2865 | cputime_t cputime) | |
2866 | { | |
2867 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
2868 | runqueue_t *rq = this_rq(); | |
2869 | cputime64_t tmp; | |
2870 | ||
2871 | p->stime = cputime_add(p->stime, cputime); | |
2872 | ||
2873 | /* Add system time to cpustat. */ | |
2874 | tmp = cputime_to_cputime64(cputime); | |
2875 | if (hardirq_count() - hardirq_offset) | |
2876 | cpustat->irq = cputime64_add(cpustat->irq, tmp); | |
2877 | else if (softirq_count()) | |
2878 | cpustat->softirq = cputime64_add(cpustat->softirq, tmp); | |
2879 | else if (p != rq->idle) | |
2880 | cpustat->system = cputime64_add(cpustat->system, tmp); | |
2881 | else if (atomic_read(&rq->nr_iowait) > 0) | |
2882 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); | |
2883 | else | |
2884 | cpustat->idle = cputime64_add(cpustat->idle, tmp); | |
2885 | /* Account for system time used */ | |
2886 | acct_update_integrals(p); | |
1da177e4 LT |
2887 | } |
2888 | ||
2889 | /* | |
2890 | * Account for involuntary wait time. | |
2891 | * @p: the process from which the cpu time has been stolen | |
2892 | * @steal: the cpu time spent in involuntary wait | |
2893 | */ | |
2894 | void account_steal_time(struct task_struct *p, cputime_t steal) | |
2895 | { | |
2896 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
2897 | cputime64_t tmp = cputime_to_cputime64(steal); | |
2898 | runqueue_t *rq = this_rq(); | |
2899 | ||
2900 | if (p == rq->idle) { | |
2901 | p->stime = cputime_add(p->stime, steal); | |
2902 | if (atomic_read(&rq->nr_iowait) > 0) | |
2903 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); | |
2904 | else | |
2905 | cpustat->idle = cputime64_add(cpustat->idle, tmp); | |
2906 | } else | |
2907 | cpustat->steal = cputime64_add(cpustat->steal, tmp); | |
2908 | } | |
2909 | ||
2910 | /* | |
2911 | * This function gets called by the timer code, with HZ frequency. | |
2912 | * We call it with interrupts disabled. | |
2913 | * | |
2914 | * It also gets called by the fork code, when changing the parent's | |
2915 | * timeslices. | |
2916 | */ | |
2917 | void scheduler_tick(void) | |
2918 | { | |
2919 | int cpu = smp_processor_id(); | |
2920 | runqueue_t *rq = this_rq(); | |
2921 | task_t *p = current; | |
2922 | unsigned long long now = sched_clock(); | |
2923 | ||
2924 | update_cpu_clock(p, rq, now); | |
2925 | ||
2926 | rq->timestamp_last_tick = now; | |
2927 | ||
2928 | if (p == rq->idle) { | |
2929 | if (wake_priority_sleeper(rq)) | |
2930 | goto out; | |
2931 | rebalance_tick(cpu, rq, SCHED_IDLE); | |
2932 | return; | |
2933 | } | |
2934 | ||
2935 | /* Task might have expired already, but not scheduled off yet */ | |
2936 | if (p->array != rq->active) { | |
2937 | set_tsk_need_resched(p); | |
2938 | goto out; | |
2939 | } | |
2940 | spin_lock(&rq->lock); | |
2941 | /* | |
2942 | * The task was running during this tick - update the | |
2943 | * time slice counter. Note: we do not update a thread's | |
2944 | * priority until it either goes to sleep or uses up its | |
2945 | * timeslice. This makes it possible for interactive tasks | |
2946 | * to use up their timeslices at their highest priority levels. | |
2947 | */ | |
2948 | if (rt_task(p)) { | |
2949 | /* | |
2950 | * RR tasks need a special form of timeslice management. | |
2951 | * FIFO tasks have no timeslices. | |
2952 | */ | |
2953 | if ((p->policy == SCHED_RR) && !--p->time_slice) { | |
2954 | p->time_slice = task_timeslice(p); | |
2955 | p->first_time_slice = 0; | |
2956 | set_tsk_need_resched(p); | |
2957 | ||
2958 | /* put it at the end of the queue: */ | |
2959 | requeue_task(p, rq->active); | |
2960 | } | |
2961 | goto out_unlock; | |
2962 | } | |
2963 | if (!--p->time_slice) { | |
2964 | dequeue_task(p, rq->active); | |
2965 | set_tsk_need_resched(p); | |
2966 | p->prio = effective_prio(p); | |
2967 | p->time_slice = task_timeslice(p); | |
2968 | p->first_time_slice = 0; | |
2969 | ||
2970 | if (!rq->expired_timestamp) | |
2971 | rq->expired_timestamp = jiffies; | |
f1adad78 | 2972 | if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) { |
1da177e4 LT |
2973 | enqueue_task(p, rq->expired); |
2974 | if (p->static_prio < rq->best_expired_prio) | |
2975 | rq->best_expired_prio = p->static_prio; | |
2976 | } else | |
2977 | enqueue_task(p, rq->active); | |
2978 | } else { | |
2979 | /* | |
2980 | * Prevent a too long timeslice allowing a task to monopolize | |
2981 | * the CPU. We do this by splitting up the timeslice into | |
2982 | * smaller pieces. | |
2983 | * | |
2984 | * Note: this does not mean the task's timeslices expire or | |
2985 | * get lost in any way, they just might be preempted by | |
2986 | * another task of equal priority. (one with higher | |
2987 | * priority would have preempted this task already.) We | |
2988 | * requeue this task to the end of the list on this priority | |
2989 | * level, which is in essence a round-robin of tasks with | |
2990 | * equal priority. | |
2991 | * | |
2992 | * This only applies to tasks in the interactive | |
2993 | * delta range with at least TIMESLICE_GRANULARITY to requeue. | |
2994 | */ | |
2995 | if (TASK_INTERACTIVE(p) && !((task_timeslice(p) - | |
2996 | p->time_slice) % TIMESLICE_GRANULARITY(p)) && | |
2997 | (p->time_slice >= TIMESLICE_GRANULARITY(p)) && | |
2998 | (p->array == rq->active)) { | |
2999 | ||
3000 | requeue_task(p, rq->active); | |
3001 | set_tsk_need_resched(p); | |
3002 | } | |
3003 | } | |
3004 | out_unlock: | |
3005 | spin_unlock(&rq->lock); | |
3006 | out: | |
3007 | rebalance_tick(cpu, rq, NOT_IDLE); | |
3008 | } | |
3009 | ||
3010 | #ifdef CONFIG_SCHED_SMT | |
fc38ed75 CK |
3011 | static inline void wakeup_busy_runqueue(runqueue_t *rq) |
3012 | { | |
3013 | /* If an SMT runqueue is sleeping due to priority reasons wake it up */ | |
3014 | if (rq->curr == rq->idle && rq->nr_running) | |
3015 | resched_task(rq->idle); | |
3016 | } | |
3017 | ||
c96d145e KC |
3018 | /* |
3019 | * Called with interrupt disabled and this_rq's runqueue locked. | |
3020 | */ | |
3021 | static void wake_sleeping_dependent(int this_cpu) | |
1da177e4 | 3022 | { |
41c7ce9a | 3023 | struct sched_domain *tmp, *sd = NULL; |
1da177e4 LT |
3024 | int i; |
3025 | ||
c96d145e KC |
3026 | for_each_domain(this_cpu, tmp) { |
3027 | if (tmp->flags & SD_SHARE_CPUPOWER) { | |
41c7ce9a | 3028 | sd = tmp; |
c96d145e KC |
3029 | break; |
3030 | } | |
3031 | } | |
41c7ce9a NP |
3032 | |
3033 | if (!sd) | |
1da177e4 LT |
3034 | return; |
3035 | ||
c96d145e | 3036 | for_each_cpu_mask(i, sd->span) { |
1da177e4 LT |
3037 | runqueue_t *smt_rq = cpu_rq(i); |
3038 | ||
c96d145e KC |
3039 | if (i == this_cpu) |
3040 | continue; | |
3041 | if (unlikely(!spin_trylock(&smt_rq->lock))) | |
3042 | continue; | |
3043 | ||
fc38ed75 | 3044 | wakeup_busy_runqueue(smt_rq); |
c96d145e | 3045 | spin_unlock(&smt_rq->lock); |
1da177e4 | 3046 | } |
1da177e4 LT |
3047 | } |
3048 | ||
67f9a619 IM |
3049 | /* |
3050 | * number of 'lost' timeslices this task wont be able to fully | |
3051 | * utilize, if another task runs on a sibling. This models the | |
3052 | * slowdown effect of other tasks running on siblings: | |
3053 | */ | |
3054 | static inline unsigned long smt_slice(task_t *p, struct sched_domain *sd) | |
3055 | { | |
3056 | return p->time_slice * (100 - sd->per_cpu_gain) / 100; | |
3057 | } | |
3058 | ||
c96d145e KC |
3059 | /* |
3060 | * To minimise lock contention and not have to drop this_rq's runlock we only | |
3061 | * trylock the sibling runqueues and bypass those runqueues if we fail to | |
3062 | * acquire their lock. As we only trylock the normal locking order does not | |
3063 | * need to be obeyed. | |
3064 | */ | |
3065 | static int dependent_sleeper(int this_cpu, runqueue_t *this_rq, task_t *p) | |
1da177e4 | 3066 | { |
41c7ce9a | 3067 | struct sched_domain *tmp, *sd = NULL; |
1da177e4 | 3068 | int ret = 0, i; |
1da177e4 | 3069 | |
c96d145e KC |
3070 | /* kernel/rt threads do not participate in dependent sleeping */ |
3071 | if (!p->mm || rt_task(p)) | |
3072 | return 0; | |
3073 | ||
3074 | for_each_domain(this_cpu, tmp) { | |
3075 | if (tmp->flags & SD_SHARE_CPUPOWER) { | |
41c7ce9a | 3076 | sd = tmp; |
c96d145e KC |
3077 | break; |
3078 | } | |
3079 | } | |
41c7ce9a NP |
3080 | |
3081 | if (!sd) | |
1da177e4 LT |
3082 | return 0; |
3083 | ||
c96d145e KC |
3084 | for_each_cpu_mask(i, sd->span) { |
3085 | runqueue_t *smt_rq; | |
3086 | task_t *smt_curr; | |
1da177e4 | 3087 | |
c96d145e KC |
3088 | if (i == this_cpu) |
3089 | continue; | |
1da177e4 | 3090 | |
c96d145e KC |
3091 | smt_rq = cpu_rq(i); |
3092 | if (unlikely(!spin_trylock(&smt_rq->lock))) | |
3093 | continue; | |
1da177e4 | 3094 | |
c96d145e | 3095 | smt_curr = smt_rq->curr; |
1da177e4 | 3096 | |
c96d145e KC |
3097 | if (!smt_curr->mm) |
3098 | goto unlock; | |
fc38ed75 | 3099 | |
1da177e4 LT |
3100 | /* |
3101 | * If a user task with lower static priority than the | |
3102 | * running task on the SMT sibling is trying to schedule, | |
3103 | * delay it till there is proportionately less timeslice | |
3104 | * left of the sibling task to prevent a lower priority | |
3105 | * task from using an unfair proportion of the | |
3106 | * physical cpu's resources. -ck | |
3107 | */ | |
fc38ed75 CK |
3108 | if (rt_task(smt_curr)) { |
3109 | /* | |
3110 | * With real time tasks we run non-rt tasks only | |
3111 | * per_cpu_gain% of the time. | |
3112 | */ | |
3113 | if ((jiffies % DEF_TIMESLICE) > | |
3114 | (sd->per_cpu_gain * DEF_TIMESLICE / 100)) | |
3115 | ret = 1; | |
c96d145e | 3116 | } else { |
67f9a619 IM |
3117 | if (smt_curr->static_prio < p->static_prio && |
3118 | !TASK_PREEMPTS_CURR(p, smt_rq) && | |
3119 | smt_slice(smt_curr, sd) > task_timeslice(p)) | |
fc38ed75 | 3120 | ret = 1; |
fc38ed75 | 3121 | } |
c96d145e KC |
3122 | unlock: |
3123 | spin_unlock(&smt_rq->lock); | |
1da177e4 | 3124 | } |
1da177e4 LT |
3125 | return ret; |
3126 | } | |
3127 | #else | |
c96d145e | 3128 | static inline void wake_sleeping_dependent(int this_cpu) |
1da177e4 LT |
3129 | { |
3130 | } | |
3131 | ||
c96d145e KC |
3132 | static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq, |
3133 | task_t *p) | |
1da177e4 LT |
3134 | { |
3135 | return 0; | |
3136 | } | |
3137 | #endif | |
3138 | ||
3139 | #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT) | |
3140 | ||
3141 | void fastcall add_preempt_count(int val) | |
3142 | { | |
3143 | /* | |
3144 | * Underflow? | |
3145 | */ | |
9a11b49a IM |
3146 | if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) |
3147 | return; | |
1da177e4 LT |
3148 | preempt_count() += val; |
3149 | /* | |
3150 | * Spinlock count overflowing soon? | |
3151 | */ | |
9a11b49a | 3152 | DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10); |
1da177e4 LT |
3153 | } |
3154 | EXPORT_SYMBOL(add_preempt_count); | |
3155 | ||
3156 | void fastcall sub_preempt_count(int val) | |
3157 | { | |
3158 | /* | |
3159 | * Underflow? | |
3160 | */ | |
9a11b49a IM |
3161 | if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) |
3162 | return; | |
1da177e4 LT |
3163 | /* |
3164 | * Is the spinlock portion underflowing? | |
3165 | */ | |
9a11b49a IM |
3166 | if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && |
3167 | !(preempt_count() & PREEMPT_MASK))) | |
3168 | return; | |
3169 | ||
1da177e4 LT |
3170 | preempt_count() -= val; |
3171 | } | |
3172 | EXPORT_SYMBOL(sub_preempt_count); | |
3173 | ||
3174 | #endif | |
3175 | ||
3dee386e CK |
3176 | static inline int interactive_sleep(enum sleep_type sleep_type) |
3177 | { | |
3178 | return (sleep_type == SLEEP_INTERACTIVE || | |
3179 | sleep_type == SLEEP_INTERRUPTED); | |
3180 | } | |
3181 | ||
1da177e4 LT |
3182 | /* |
3183 | * schedule() is the main scheduler function. | |
3184 | */ | |
3185 | asmlinkage void __sched schedule(void) | |
3186 | { | |
3187 | long *switch_count; | |
3188 | task_t *prev, *next; | |
3189 | runqueue_t *rq; | |
3190 | prio_array_t *array; | |
3191 | struct list_head *queue; | |
3192 | unsigned long long now; | |
3193 | unsigned long run_time; | |
a3464a10 | 3194 | int cpu, idx, new_prio; |
1da177e4 LT |
3195 | |
3196 | /* | |
3197 | * Test if we are atomic. Since do_exit() needs to call into | |
3198 | * schedule() atomically, we ignore that path for now. | |
3199 | * Otherwise, whine if we are scheduling when we should not be. | |
3200 | */ | |
77e4bfbc AM |
3201 | if (unlikely(in_atomic() && !current->exit_state)) { |
3202 | printk(KERN_ERR "BUG: scheduling while atomic: " | |
3203 | "%s/0x%08x/%d\n", | |
3204 | current->comm, preempt_count(), current->pid); | |
3205 | dump_stack(); | |
1da177e4 LT |
3206 | } |
3207 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); | |
3208 | ||
3209 | need_resched: | |
3210 | preempt_disable(); | |
3211 | prev = current; | |
3212 | release_kernel_lock(prev); | |
3213 | need_resched_nonpreemptible: | |
3214 | rq = this_rq(); | |
3215 | ||
3216 | /* | |
3217 | * The idle thread is not allowed to schedule! | |
3218 | * Remove this check after it has been exercised a bit. | |
3219 | */ | |
3220 | if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) { | |
3221 | printk(KERN_ERR "bad: scheduling from the idle thread!\n"); | |
3222 | dump_stack(); | |
3223 | } | |
3224 | ||
3225 | schedstat_inc(rq, sched_cnt); | |
3226 | now = sched_clock(); | |
238628ed | 3227 | if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) { |
1da177e4 | 3228 | run_time = now - prev->timestamp; |
238628ed | 3229 | if (unlikely((long long)(now - prev->timestamp) < 0)) |
1da177e4 LT |
3230 | run_time = 0; |
3231 | } else | |
3232 | run_time = NS_MAX_SLEEP_AVG; | |
3233 | ||
3234 | /* | |
3235 | * Tasks charged proportionately less run_time at high sleep_avg to | |
3236 | * delay them losing their interactive status | |
3237 | */ | |
3238 | run_time /= (CURRENT_BONUS(prev) ? : 1); | |
3239 | ||
3240 | spin_lock_irq(&rq->lock); | |
3241 | ||
3242 | if (unlikely(prev->flags & PF_DEAD)) | |
3243 | prev->state = EXIT_DEAD; | |
3244 | ||
3245 | switch_count = &prev->nivcsw; | |
3246 | if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { | |
3247 | switch_count = &prev->nvcsw; | |
3248 | if (unlikely((prev->state & TASK_INTERRUPTIBLE) && | |
3249 | unlikely(signal_pending(prev)))) | |
3250 | prev->state = TASK_RUNNING; | |
3251 | else { | |
3252 | if (prev->state == TASK_UNINTERRUPTIBLE) | |
3253 | rq->nr_uninterruptible++; | |
3254 | deactivate_task(prev, rq); | |
3255 | } | |
3256 | } | |
3257 | ||
3258 | cpu = smp_processor_id(); | |
3259 | if (unlikely(!rq->nr_running)) { | |
1da177e4 LT |
3260 | idle_balance(cpu, rq); |
3261 | if (!rq->nr_running) { | |
3262 | next = rq->idle; | |
3263 | rq->expired_timestamp = 0; | |
c96d145e | 3264 | wake_sleeping_dependent(cpu); |
1da177e4 LT |
3265 | goto switch_tasks; |
3266 | } | |
1da177e4 LT |
3267 | } |
3268 | ||
3269 | array = rq->active; | |
3270 | if (unlikely(!array->nr_active)) { | |
3271 | /* | |
3272 | * Switch the active and expired arrays. | |
3273 | */ | |
3274 | schedstat_inc(rq, sched_switch); | |
3275 | rq->active = rq->expired; | |
3276 | rq->expired = array; | |
3277 | array = rq->active; | |
3278 | rq->expired_timestamp = 0; | |
3279 | rq->best_expired_prio = MAX_PRIO; | |
3280 | } | |
3281 | ||
3282 | idx = sched_find_first_bit(array->bitmap); | |
3283 | queue = array->queue + idx; | |
3284 | next = list_entry(queue->next, task_t, run_list); | |
3285 | ||
3dee386e | 3286 | if (!rt_task(next) && interactive_sleep(next->sleep_type)) { |
1da177e4 | 3287 | unsigned long long delta = now - next->timestamp; |
238628ed | 3288 | if (unlikely((long long)(now - next->timestamp) < 0)) |
1da177e4 LT |
3289 | delta = 0; |
3290 | ||
3dee386e | 3291 | if (next->sleep_type == SLEEP_INTERACTIVE) |
1da177e4 LT |
3292 | delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128; |
3293 | ||
3294 | array = next->array; | |
a3464a10 CS |
3295 | new_prio = recalc_task_prio(next, next->timestamp + delta); |
3296 | ||
3297 | if (unlikely(next->prio != new_prio)) { | |
3298 | dequeue_task(next, array); | |
3299 | next->prio = new_prio; | |
3300 | enqueue_task(next, array); | |
7c4bb1f9 | 3301 | } |
1da177e4 | 3302 | } |
3dee386e | 3303 | next->sleep_type = SLEEP_NORMAL; |
c96d145e KC |
3304 | if (dependent_sleeper(cpu, rq, next)) |
3305 | next = rq->idle; | |
1da177e4 LT |
3306 | switch_tasks: |
3307 | if (next == rq->idle) | |
3308 | schedstat_inc(rq, sched_goidle); | |
3309 | prefetch(next); | |
383f2835 | 3310 | prefetch_stack(next); |
1da177e4 LT |
3311 | clear_tsk_need_resched(prev); |
3312 | rcu_qsctr_inc(task_cpu(prev)); | |
3313 | ||
3314 | update_cpu_clock(prev, rq, now); | |
3315 | ||
3316 | prev->sleep_avg -= run_time; | |
3317 | if ((long)prev->sleep_avg <= 0) | |
3318 | prev->sleep_avg = 0; | |
3319 | prev->timestamp = prev->last_ran = now; | |
3320 | ||
3321 | sched_info_switch(prev, next); | |
3322 | if (likely(prev != next)) { | |
3323 | next->timestamp = now; | |
3324 | rq->nr_switches++; | |
3325 | rq->curr = next; | |
3326 | ++*switch_count; | |
3327 | ||
4866cde0 | 3328 | prepare_task_switch(rq, next); |
1da177e4 LT |
3329 | prev = context_switch(rq, prev, next); |
3330 | barrier(); | |
4866cde0 NP |
3331 | /* |
3332 | * this_rq must be evaluated again because prev may have moved | |
3333 | * CPUs since it called schedule(), thus the 'rq' on its stack | |
3334 | * frame will be invalid. | |
3335 | */ | |
3336 | finish_task_switch(this_rq(), prev); | |
1da177e4 LT |
3337 | } else |
3338 | spin_unlock_irq(&rq->lock); | |
3339 | ||
3340 | prev = current; | |
3341 | if (unlikely(reacquire_kernel_lock(prev) < 0)) | |
3342 | goto need_resched_nonpreemptible; | |
3343 | preempt_enable_no_resched(); | |
3344 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
3345 | goto need_resched; | |
3346 | } | |
3347 | ||
3348 | EXPORT_SYMBOL(schedule); | |
3349 | ||
3350 | #ifdef CONFIG_PREEMPT | |
3351 | /* | |
3352 | * this is is the entry point to schedule() from in-kernel preemption | |
3353 | * off of preempt_enable. Kernel preemptions off return from interrupt | |
3354 | * occur there and call schedule directly. | |
3355 | */ | |
3356 | asmlinkage void __sched preempt_schedule(void) | |
3357 | { | |
3358 | struct thread_info *ti = current_thread_info(); | |
3359 | #ifdef CONFIG_PREEMPT_BKL | |
3360 | struct task_struct *task = current; | |
3361 | int saved_lock_depth; | |
3362 | #endif | |
3363 | /* | |
3364 | * If there is a non-zero preempt_count or interrupts are disabled, | |
3365 | * we do not want to preempt the current task. Just return.. | |
3366 | */ | |
3367 | if (unlikely(ti->preempt_count || irqs_disabled())) | |
3368 | return; | |
3369 | ||
3370 | need_resched: | |
3371 | add_preempt_count(PREEMPT_ACTIVE); | |
3372 | /* | |
3373 | * We keep the big kernel semaphore locked, but we | |
3374 | * clear ->lock_depth so that schedule() doesnt | |
3375 | * auto-release the semaphore: | |
3376 | */ | |
3377 | #ifdef CONFIG_PREEMPT_BKL | |
3378 | saved_lock_depth = task->lock_depth; | |
3379 | task->lock_depth = -1; | |
3380 | #endif | |
3381 | schedule(); | |
3382 | #ifdef CONFIG_PREEMPT_BKL | |
3383 | task->lock_depth = saved_lock_depth; | |
3384 | #endif | |
3385 | sub_preempt_count(PREEMPT_ACTIVE); | |
3386 | ||
3387 | /* we could miss a preemption opportunity between schedule and now */ | |
3388 | barrier(); | |
3389 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
3390 | goto need_resched; | |
3391 | } | |
3392 | ||
3393 | EXPORT_SYMBOL(preempt_schedule); | |
3394 | ||
3395 | /* | |
3396 | * this is is the entry point to schedule() from kernel preemption | |
3397 | * off of irq context. | |
3398 | * Note, that this is called and return with irqs disabled. This will | |
3399 | * protect us against recursive calling from irq. | |
3400 | */ | |
3401 | asmlinkage void __sched preempt_schedule_irq(void) | |
3402 | { | |
3403 | struct thread_info *ti = current_thread_info(); | |
3404 | #ifdef CONFIG_PREEMPT_BKL | |
3405 | struct task_struct *task = current; | |
3406 | int saved_lock_depth; | |
3407 | #endif | |
3408 | /* Catch callers which need to be fixed*/ | |
3409 | BUG_ON(ti->preempt_count || !irqs_disabled()); | |
3410 | ||
3411 | need_resched: | |
3412 | add_preempt_count(PREEMPT_ACTIVE); | |
3413 | /* | |
3414 | * We keep the big kernel semaphore locked, but we | |
3415 | * clear ->lock_depth so that schedule() doesnt | |
3416 | * auto-release the semaphore: | |
3417 | */ | |
3418 | #ifdef CONFIG_PREEMPT_BKL | |
3419 | saved_lock_depth = task->lock_depth; | |
3420 | task->lock_depth = -1; | |
3421 | #endif | |
3422 | local_irq_enable(); | |
3423 | schedule(); | |
3424 | local_irq_disable(); | |
3425 | #ifdef CONFIG_PREEMPT_BKL | |
3426 | task->lock_depth = saved_lock_depth; | |
3427 | #endif | |
3428 | sub_preempt_count(PREEMPT_ACTIVE); | |
3429 | ||
3430 | /* we could miss a preemption opportunity between schedule and now */ | |
3431 | barrier(); | |
3432 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
3433 | goto need_resched; | |
3434 | } | |
3435 | ||
3436 | #endif /* CONFIG_PREEMPT */ | |
3437 | ||
95cdf3b7 IM |
3438 | int default_wake_function(wait_queue_t *curr, unsigned mode, int sync, |
3439 | void *key) | |
1da177e4 | 3440 | { |
c43dc2fd | 3441 | task_t *p = curr->private; |
1da177e4 LT |
3442 | return try_to_wake_up(p, mode, sync); |
3443 | } | |
3444 | ||
3445 | EXPORT_SYMBOL(default_wake_function); | |
3446 | ||
3447 | /* | |
3448 | * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just | |
3449 | * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve | |
3450 | * number) then we wake all the non-exclusive tasks and one exclusive task. | |
3451 | * | |
3452 | * There are circumstances in which we can try to wake a task which has already | |
3453 | * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns | |
3454 | * zero in this (rare) case, and we handle it by continuing to scan the queue. | |
3455 | */ | |
3456 | static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, | |
3457 | int nr_exclusive, int sync, void *key) | |
3458 | { | |
3459 | struct list_head *tmp, *next; | |
3460 | ||
3461 | list_for_each_safe(tmp, next, &q->task_list) { | |
3462 | wait_queue_t *curr; | |
3463 | unsigned flags; | |
3464 | curr = list_entry(tmp, wait_queue_t, task_list); | |
3465 | flags = curr->flags; | |
3466 | if (curr->func(curr, mode, sync, key) && | |
3467 | (flags & WQ_FLAG_EXCLUSIVE) && | |
3468 | !--nr_exclusive) | |
3469 | break; | |
3470 | } | |
3471 | } | |
3472 | ||
3473 | /** | |
3474 | * __wake_up - wake up threads blocked on a waitqueue. | |
3475 | * @q: the waitqueue | |
3476 | * @mode: which threads | |
3477 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
67be2dd1 | 3478 | * @key: is directly passed to the wakeup function |
1da177e4 LT |
3479 | */ |
3480 | void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode, | |
95cdf3b7 | 3481 | int nr_exclusive, void *key) |
1da177e4 LT |
3482 | { |
3483 | unsigned long flags; | |
3484 | ||
3485 | spin_lock_irqsave(&q->lock, flags); | |
3486 | __wake_up_common(q, mode, nr_exclusive, 0, key); | |
3487 | spin_unlock_irqrestore(&q->lock, flags); | |
3488 | } | |
3489 | ||
3490 | EXPORT_SYMBOL(__wake_up); | |
3491 | ||
3492 | /* | |
3493 | * Same as __wake_up but called with the spinlock in wait_queue_head_t held. | |
3494 | */ | |
3495 | void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode) | |
3496 | { | |
3497 | __wake_up_common(q, mode, 1, 0, NULL); | |
3498 | } | |
3499 | ||
3500 | /** | |
67be2dd1 | 3501 | * __wake_up_sync - wake up threads blocked on a waitqueue. |
1da177e4 LT |
3502 | * @q: the waitqueue |
3503 | * @mode: which threads | |
3504 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
3505 | * | |
3506 | * The sync wakeup differs that the waker knows that it will schedule | |
3507 | * away soon, so while the target thread will be woken up, it will not | |
3508 | * be migrated to another CPU - ie. the two threads are 'synchronized' | |
3509 | * with each other. This can prevent needless bouncing between CPUs. | |
3510 | * | |
3511 | * On UP it can prevent extra preemption. | |
3512 | */ | |
95cdf3b7 IM |
3513 | void fastcall |
3514 | __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) | |
1da177e4 LT |
3515 | { |
3516 | unsigned long flags; | |
3517 | int sync = 1; | |
3518 | ||
3519 | if (unlikely(!q)) | |
3520 | return; | |
3521 | ||
3522 | if (unlikely(!nr_exclusive)) | |
3523 | sync = 0; | |
3524 | ||
3525 | spin_lock_irqsave(&q->lock, flags); | |
3526 | __wake_up_common(q, mode, nr_exclusive, sync, NULL); | |
3527 | spin_unlock_irqrestore(&q->lock, flags); | |
3528 | } | |
3529 | EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ | |
3530 | ||
3531 | void fastcall complete(struct completion *x) | |
3532 | { | |
3533 | unsigned long flags; | |
3534 | ||
3535 | spin_lock_irqsave(&x->wait.lock, flags); | |
3536 | x->done++; | |
3537 | __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, | |
3538 | 1, 0, NULL); | |
3539 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
3540 | } | |
3541 | EXPORT_SYMBOL(complete); | |
3542 | ||
3543 | void fastcall complete_all(struct completion *x) | |
3544 | { | |
3545 | unsigned long flags; | |
3546 | ||
3547 | spin_lock_irqsave(&x->wait.lock, flags); | |
3548 | x->done += UINT_MAX/2; | |
3549 | __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, | |
3550 | 0, 0, NULL); | |
3551 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
3552 | } | |
3553 | EXPORT_SYMBOL(complete_all); | |
3554 | ||
3555 | void fastcall __sched wait_for_completion(struct completion *x) | |
3556 | { | |
3557 | might_sleep(); | |
3558 | spin_lock_irq(&x->wait.lock); | |
3559 | if (!x->done) { | |
3560 | DECLARE_WAITQUEUE(wait, current); | |
3561 | ||
3562 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3563 | __add_wait_queue_tail(&x->wait, &wait); | |
3564 | do { | |
3565 | __set_current_state(TASK_UNINTERRUPTIBLE); | |
3566 | spin_unlock_irq(&x->wait.lock); | |
3567 | schedule(); | |
3568 | spin_lock_irq(&x->wait.lock); | |
3569 | } while (!x->done); | |
3570 | __remove_wait_queue(&x->wait, &wait); | |
3571 | } | |
3572 | x->done--; | |
3573 | spin_unlock_irq(&x->wait.lock); | |
3574 | } | |
3575 | EXPORT_SYMBOL(wait_for_completion); | |
3576 | ||
3577 | unsigned long fastcall __sched | |
3578 | wait_for_completion_timeout(struct completion *x, unsigned long timeout) | |
3579 | { | |
3580 | might_sleep(); | |
3581 | ||
3582 | spin_lock_irq(&x->wait.lock); | |
3583 | if (!x->done) { | |
3584 | DECLARE_WAITQUEUE(wait, current); | |
3585 | ||
3586 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3587 | __add_wait_queue_tail(&x->wait, &wait); | |
3588 | do { | |
3589 | __set_current_state(TASK_UNINTERRUPTIBLE); | |
3590 | spin_unlock_irq(&x->wait.lock); | |
3591 | timeout = schedule_timeout(timeout); | |
3592 | spin_lock_irq(&x->wait.lock); | |
3593 | if (!timeout) { | |
3594 | __remove_wait_queue(&x->wait, &wait); | |
3595 | goto out; | |
3596 | } | |
3597 | } while (!x->done); | |
3598 | __remove_wait_queue(&x->wait, &wait); | |
3599 | } | |
3600 | x->done--; | |
3601 | out: | |
3602 | spin_unlock_irq(&x->wait.lock); | |
3603 | return timeout; | |
3604 | } | |
3605 | EXPORT_SYMBOL(wait_for_completion_timeout); | |
3606 | ||
3607 | int fastcall __sched wait_for_completion_interruptible(struct completion *x) | |
3608 | { | |
3609 | int ret = 0; | |
3610 | ||
3611 | might_sleep(); | |
3612 | ||
3613 | spin_lock_irq(&x->wait.lock); | |
3614 | if (!x->done) { | |
3615 | DECLARE_WAITQUEUE(wait, current); | |
3616 | ||
3617 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3618 | __add_wait_queue_tail(&x->wait, &wait); | |
3619 | do { | |
3620 | if (signal_pending(current)) { | |
3621 | ret = -ERESTARTSYS; | |
3622 | __remove_wait_queue(&x->wait, &wait); | |
3623 | goto out; | |
3624 | } | |
3625 | __set_current_state(TASK_INTERRUPTIBLE); | |
3626 | spin_unlock_irq(&x->wait.lock); | |
3627 | schedule(); | |
3628 | spin_lock_irq(&x->wait.lock); | |
3629 | } while (!x->done); | |
3630 | __remove_wait_queue(&x->wait, &wait); | |
3631 | } | |
3632 | x->done--; | |
3633 | out: | |
3634 | spin_unlock_irq(&x->wait.lock); | |
3635 | ||
3636 | return ret; | |
3637 | } | |
3638 | EXPORT_SYMBOL(wait_for_completion_interruptible); | |
3639 | ||
3640 | unsigned long fastcall __sched | |
3641 | wait_for_completion_interruptible_timeout(struct completion *x, | |
3642 | unsigned long timeout) | |
3643 | { | |
3644 | might_sleep(); | |
3645 | ||
3646 | spin_lock_irq(&x->wait.lock); | |
3647 | if (!x->done) { | |
3648 | DECLARE_WAITQUEUE(wait, current); | |
3649 | ||
3650 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3651 | __add_wait_queue_tail(&x->wait, &wait); | |
3652 | do { | |
3653 | if (signal_pending(current)) { | |
3654 | timeout = -ERESTARTSYS; | |
3655 | __remove_wait_queue(&x->wait, &wait); | |
3656 | goto out; | |
3657 | } | |
3658 | __set_current_state(TASK_INTERRUPTIBLE); | |
3659 | spin_unlock_irq(&x->wait.lock); | |
3660 | timeout = schedule_timeout(timeout); | |
3661 | spin_lock_irq(&x->wait.lock); | |
3662 | if (!timeout) { | |
3663 | __remove_wait_queue(&x->wait, &wait); | |
3664 | goto out; | |
3665 | } | |
3666 | } while (!x->done); | |
3667 | __remove_wait_queue(&x->wait, &wait); | |
3668 | } | |
3669 | x->done--; | |
3670 | out: | |
3671 | spin_unlock_irq(&x->wait.lock); | |
3672 | return timeout; | |
3673 | } | |
3674 | EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); | |
3675 | ||
3676 | ||
3677 | #define SLEEP_ON_VAR \ | |
3678 | unsigned long flags; \ | |
3679 | wait_queue_t wait; \ | |
3680 | init_waitqueue_entry(&wait, current); | |
3681 | ||
3682 | #define SLEEP_ON_HEAD \ | |
3683 | spin_lock_irqsave(&q->lock,flags); \ | |
3684 | __add_wait_queue(q, &wait); \ | |
3685 | spin_unlock(&q->lock); | |
3686 | ||
3687 | #define SLEEP_ON_TAIL \ | |
3688 | spin_lock_irq(&q->lock); \ | |
3689 | __remove_wait_queue(q, &wait); \ | |
3690 | spin_unlock_irqrestore(&q->lock, flags); | |
3691 | ||
3692 | void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q) | |
3693 | { | |
3694 | SLEEP_ON_VAR | |
3695 | ||
3696 | current->state = TASK_INTERRUPTIBLE; | |
3697 | ||
3698 | SLEEP_ON_HEAD | |
3699 | schedule(); | |
3700 | SLEEP_ON_TAIL | |
3701 | } | |
3702 | ||
3703 | EXPORT_SYMBOL(interruptible_sleep_on); | |
3704 | ||
95cdf3b7 IM |
3705 | long fastcall __sched |
3706 | interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) | |
1da177e4 LT |
3707 | { |
3708 | SLEEP_ON_VAR | |
3709 | ||
3710 | current->state = TASK_INTERRUPTIBLE; | |
3711 | ||
3712 | SLEEP_ON_HEAD | |
3713 | timeout = schedule_timeout(timeout); | |
3714 | SLEEP_ON_TAIL | |
3715 | ||
3716 | return timeout; | |
3717 | } | |
3718 | ||
3719 | EXPORT_SYMBOL(interruptible_sleep_on_timeout); | |
3720 | ||
3721 | void fastcall __sched sleep_on(wait_queue_head_t *q) | |
3722 | { | |
3723 | SLEEP_ON_VAR | |
3724 | ||
3725 | current->state = TASK_UNINTERRUPTIBLE; | |
3726 | ||
3727 | SLEEP_ON_HEAD | |
3728 | schedule(); | |
3729 | SLEEP_ON_TAIL | |
3730 | } | |
3731 | ||
3732 | EXPORT_SYMBOL(sleep_on); | |
3733 | ||
3734 | long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) | |
3735 | { | |
3736 | SLEEP_ON_VAR | |
3737 | ||
3738 | current->state = TASK_UNINTERRUPTIBLE; | |
3739 | ||
3740 | SLEEP_ON_HEAD | |
3741 | timeout = schedule_timeout(timeout); | |
3742 | SLEEP_ON_TAIL | |
3743 | ||
3744 | return timeout; | |
3745 | } | |
3746 | ||
3747 | EXPORT_SYMBOL(sleep_on_timeout); | |
3748 | ||
b29739f9 IM |
3749 | #ifdef CONFIG_RT_MUTEXES |
3750 | ||
3751 | /* | |
3752 | * rt_mutex_setprio - set the current priority of a task | |
3753 | * @p: task | |
3754 | * @prio: prio value (kernel-internal form) | |
3755 | * | |
3756 | * This function changes the 'effective' priority of a task. It does | |
3757 | * not touch ->normal_prio like __setscheduler(). | |
3758 | * | |
3759 | * Used by the rt_mutex code to implement priority inheritance logic. | |
3760 | */ | |
3761 | void rt_mutex_setprio(task_t *p, int prio) | |
3762 | { | |
3763 | unsigned long flags; | |
3764 | prio_array_t *array; | |
3765 | runqueue_t *rq; | |
3766 | int oldprio; | |
3767 | ||
3768 | BUG_ON(prio < 0 || prio > MAX_PRIO); | |
3769 | ||
3770 | rq = task_rq_lock(p, &flags); | |
3771 | ||
3772 | oldprio = p->prio; | |
3773 | array = p->array; | |
3774 | if (array) | |
3775 | dequeue_task(p, array); | |
3776 | p->prio = prio; | |
3777 | ||
3778 | if (array) { | |
3779 | /* | |
3780 | * If changing to an RT priority then queue it | |
3781 | * in the active array! | |
3782 | */ | |
3783 | if (rt_task(p)) | |
3784 | array = rq->active; | |
3785 | enqueue_task(p, array); | |
3786 | /* | |
3787 | * Reschedule if we are currently running on this runqueue and | |
3788 | * our priority decreased, or if we are not currently running on | |
3789 | * this runqueue and our priority is higher than the current's | |
3790 | */ | |
3791 | if (task_running(rq, p)) { | |
3792 | if (p->prio > oldprio) | |
3793 | resched_task(rq->curr); | |
3794 | } else if (TASK_PREEMPTS_CURR(p, rq)) | |
3795 | resched_task(rq->curr); | |
3796 | } | |
3797 | task_rq_unlock(rq, &flags); | |
3798 | } | |
3799 | ||
3800 | #endif | |
3801 | ||
1da177e4 LT |
3802 | void set_user_nice(task_t *p, long nice) |
3803 | { | |
3804 | unsigned long flags; | |
3805 | prio_array_t *array; | |
3806 | runqueue_t *rq; | |
b29739f9 | 3807 | int old_prio, delta; |
1da177e4 LT |
3808 | |
3809 | if (TASK_NICE(p) == nice || nice < -20 || nice > 19) | |
3810 | return; | |
3811 | /* | |
3812 | * We have to be careful, if called from sys_setpriority(), | |
3813 | * the task might be in the middle of scheduling on another CPU. | |
3814 | */ | |
3815 | rq = task_rq_lock(p, &flags); | |
3816 | /* | |
3817 | * The RT priorities are set via sched_setscheduler(), but we still | |
3818 | * allow the 'normal' nice value to be set - but as expected | |
3819 | * it wont have any effect on scheduling until the task is | |
b0a9499c | 3820 | * not SCHED_NORMAL/SCHED_BATCH: |
1da177e4 | 3821 | */ |
b29739f9 | 3822 | if (has_rt_policy(p)) { |
1da177e4 LT |
3823 | p->static_prio = NICE_TO_PRIO(nice); |
3824 | goto out_unlock; | |
3825 | } | |
3826 | array = p->array; | |
2dd73a4f | 3827 | if (array) { |
1da177e4 | 3828 | dequeue_task(p, array); |
2dd73a4f PW |
3829 | dec_raw_weighted_load(rq, p); |
3830 | } | |
1da177e4 | 3831 | |
1da177e4 | 3832 | p->static_prio = NICE_TO_PRIO(nice); |
2dd73a4f | 3833 | set_load_weight(p); |
b29739f9 IM |
3834 | old_prio = p->prio; |
3835 | p->prio = effective_prio(p); | |
3836 | delta = p->prio - old_prio; | |
1da177e4 LT |
3837 | |
3838 | if (array) { | |
3839 | enqueue_task(p, array); | |
2dd73a4f | 3840 | inc_raw_weighted_load(rq, p); |
1da177e4 LT |
3841 | /* |
3842 | * If the task increased its priority or is running and | |
3843 | * lowered its priority, then reschedule its CPU: | |
3844 | */ | |
3845 | if (delta < 0 || (delta > 0 && task_running(rq, p))) | |
3846 | resched_task(rq->curr); | |
3847 | } | |
3848 | out_unlock: | |
3849 | task_rq_unlock(rq, &flags); | |
3850 | } | |
1da177e4 LT |
3851 | EXPORT_SYMBOL(set_user_nice); |
3852 | ||
e43379f1 MM |
3853 | /* |
3854 | * can_nice - check if a task can reduce its nice value | |
3855 | * @p: task | |
3856 | * @nice: nice value | |
3857 | */ | |
3858 | int can_nice(const task_t *p, const int nice) | |
3859 | { | |
024f4747 MM |
3860 | /* convert nice value [19,-20] to rlimit style value [1,40] */ |
3861 | int nice_rlim = 20 - nice; | |
e43379f1 MM |
3862 | return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur || |
3863 | capable(CAP_SYS_NICE)); | |
3864 | } | |
3865 | ||
1da177e4 LT |
3866 | #ifdef __ARCH_WANT_SYS_NICE |
3867 | ||
3868 | /* | |
3869 | * sys_nice - change the priority of the current process. | |
3870 | * @increment: priority increment | |
3871 | * | |
3872 | * sys_setpriority is a more generic, but much slower function that | |
3873 | * does similar things. | |
3874 | */ | |
3875 | asmlinkage long sys_nice(int increment) | |
3876 | { | |
3877 | int retval; | |
3878 | long nice; | |
3879 | ||
3880 | /* | |
3881 | * Setpriority might change our priority at the same moment. | |
3882 | * We don't have to worry. Conceptually one call occurs first | |
3883 | * and we have a single winner. | |
3884 | */ | |
e43379f1 MM |
3885 | if (increment < -40) |
3886 | increment = -40; | |
1da177e4 LT |
3887 | if (increment > 40) |
3888 | increment = 40; | |
3889 | ||
3890 | nice = PRIO_TO_NICE(current->static_prio) + increment; | |
3891 | if (nice < -20) | |
3892 | nice = -20; | |
3893 | if (nice > 19) | |
3894 | nice = 19; | |
3895 | ||
e43379f1 MM |
3896 | if (increment < 0 && !can_nice(current, nice)) |
3897 | return -EPERM; | |
3898 | ||
1da177e4 LT |
3899 | retval = security_task_setnice(current, nice); |
3900 | if (retval) | |
3901 | return retval; | |
3902 | ||
3903 | set_user_nice(current, nice); | |
3904 | return 0; | |
3905 | } | |
3906 | ||
3907 | #endif | |
3908 | ||
3909 | /** | |
3910 | * task_prio - return the priority value of a given task. | |
3911 | * @p: the task in question. | |
3912 | * | |
3913 | * This is the priority value as seen by users in /proc. | |
3914 | * RT tasks are offset by -200. Normal tasks are centered | |
3915 | * around 0, value goes from -16 to +15. | |
3916 | */ | |
3917 | int task_prio(const task_t *p) | |
3918 | { | |
3919 | return p->prio - MAX_RT_PRIO; | |
3920 | } | |
3921 | ||
3922 | /** | |
3923 | * task_nice - return the nice value of a given task. | |
3924 | * @p: the task in question. | |
3925 | */ | |
3926 | int task_nice(const task_t *p) | |
3927 | { | |
3928 | return TASK_NICE(p); | |
3929 | } | |
1da177e4 | 3930 | EXPORT_SYMBOL_GPL(task_nice); |
1da177e4 LT |
3931 | |
3932 | /** | |
3933 | * idle_cpu - is a given cpu idle currently? | |
3934 | * @cpu: the processor in question. | |
3935 | */ | |
3936 | int idle_cpu(int cpu) | |
3937 | { | |
3938 | return cpu_curr(cpu) == cpu_rq(cpu)->idle; | |
3939 | } | |
3940 | ||
1da177e4 LT |
3941 | /** |
3942 | * idle_task - return the idle task for a given cpu. | |
3943 | * @cpu: the processor in question. | |
3944 | */ | |
3945 | task_t *idle_task(int cpu) | |
3946 | { | |
3947 | return cpu_rq(cpu)->idle; | |
3948 | } | |
3949 | ||
3950 | /** | |
3951 | * find_process_by_pid - find a process with a matching PID value. | |
3952 | * @pid: the pid in question. | |
3953 | */ | |
3954 | static inline task_t *find_process_by_pid(pid_t pid) | |
3955 | { | |
3956 | return pid ? find_task_by_pid(pid) : current; | |
3957 | } | |
3958 | ||
3959 | /* Actually do priority change: must hold rq lock. */ | |
3960 | static void __setscheduler(struct task_struct *p, int policy, int prio) | |
3961 | { | |
3962 | BUG_ON(p->array); | |
3963 | p->policy = policy; | |
3964 | p->rt_priority = prio; | |
b29739f9 IM |
3965 | p->normal_prio = normal_prio(p); |
3966 | /* we are holding p->pi_lock already */ | |
3967 | p->prio = rt_mutex_getprio(p); | |
3968 | /* | |
3969 | * SCHED_BATCH tasks are treated as perpetual CPU hogs: | |
3970 | */ | |
3971 | if (policy == SCHED_BATCH) | |
3972 | p->sleep_avg = 0; | |
2dd73a4f | 3973 | set_load_weight(p); |
1da177e4 LT |
3974 | } |
3975 | ||
3976 | /** | |
3977 | * sched_setscheduler - change the scheduling policy and/or RT priority of | |
3978 | * a thread. | |
3979 | * @p: the task in question. | |
3980 | * @policy: new policy. | |
3981 | * @param: structure containing the new RT priority. | |
3982 | */ | |
95cdf3b7 IM |
3983 | int sched_setscheduler(struct task_struct *p, int policy, |
3984 | struct sched_param *param) | |
1da177e4 LT |
3985 | { |
3986 | int retval; | |
3987 | int oldprio, oldpolicy = -1; | |
3988 | prio_array_t *array; | |
3989 | unsigned long flags; | |
3990 | runqueue_t *rq; | |
3991 | ||
66e5393a SR |
3992 | /* may grab non-irq protected spin_locks */ |
3993 | BUG_ON(in_interrupt()); | |
1da177e4 LT |
3994 | recheck: |
3995 | /* double check policy once rq lock held */ | |
3996 | if (policy < 0) | |
3997 | policy = oldpolicy = p->policy; | |
3998 | else if (policy != SCHED_FIFO && policy != SCHED_RR && | |
b0a9499c IM |
3999 | policy != SCHED_NORMAL && policy != SCHED_BATCH) |
4000 | return -EINVAL; | |
1da177e4 LT |
4001 | /* |
4002 | * Valid priorities for SCHED_FIFO and SCHED_RR are | |
b0a9499c IM |
4003 | * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and |
4004 | * SCHED_BATCH is 0. | |
1da177e4 LT |
4005 | */ |
4006 | if (param->sched_priority < 0 || | |
95cdf3b7 | 4007 | (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || |
d46523ea | 4008 | (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) |
1da177e4 | 4009 | return -EINVAL; |
b0a9499c IM |
4010 | if ((policy == SCHED_NORMAL || policy == SCHED_BATCH) |
4011 | != (param->sched_priority == 0)) | |
1da177e4 LT |
4012 | return -EINVAL; |
4013 | ||
37e4ab3f OC |
4014 | /* |
4015 | * Allow unprivileged RT tasks to decrease priority: | |
4016 | */ | |
4017 | if (!capable(CAP_SYS_NICE)) { | |
b0a9499c IM |
4018 | /* |
4019 | * can't change policy, except between SCHED_NORMAL | |
4020 | * and SCHED_BATCH: | |
4021 | */ | |
4022 | if (((policy != SCHED_NORMAL && p->policy != SCHED_BATCH) && | |
4023 | (policy != SCHED_BATCH && p->policy != SCHED_NORMAL)) && | |
4024 | !p->signal->rlim[RLIMIT_RTPRIO].rlim_cur) | |
37e4ab3f OC |
4025 | return -EPERM; |
4026 | /* can't increase priority */ | |
b0a9499c | 4027 | if ((policy != SCHED_NORMAL && policy != SCHED_BATCH) && |
37e4ab3f OC |
4028 | param->sched_priority > p->rt_priority && |
4029 | param->sched_priority > | |
4030 | p->signal->rlim[RLIMIT_RTPRIO].rlim_cur) | |
4031 | return -EPERM; | |
4032 | /* can't change other user's priorities */ | |
4033 | if ((current->euid != p->euid) && | |
4034 | (current->euid != p->uid)) | |
4035 | return -EPERM; | |
4036 | } | |
1da177e4 LT |
4037 | |
4038 | retval = security_task_setscheduler(p, policy, param); | |
4039 | if (retval) | |
4040 | return retval; | |
b29739f9 IM |
4041 | /* |
4042 | * make sure no PI-waiters arrive (or leave) while we are | |
4043 | * changing the priority of the task: | |
4044 | */ | |
4045 | spin_lock_irqsave(&p->pi_lock, flags); | |
1da177e4 LT |
4046 | /* |
4047 | * To be able to change p->policy safely, the apropriate | |
4048 | * runqueue lock must be held. | |
4049 | */ | |
b29739f9 | 4050 | rq = __task_rq_lock(p); |
1da177e4 LT |
4051 | /* recheck policy now with rq lock held */ |
4052 | if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { | |
4053 | policy = oldpolicy = -1; | |
b29739f9 IM |
4054 | __task_rq_unlock(rq); |
4055 | spin_unlock_irqrestore(&p->pi_lock, flags); | |
1da177e4 LT |
4056 | goto recheck; |
4057 | } | |
4058 | array = p->array; | |
4059 | if (array) | |
4060 | deactivate_task(p, rq); | |
4061 | oldprio = p->prio; | |
4062 | __setscheduler(p, policy, param->sched_priority); | |
4063 | if (array) { | |
4064 | __activate_task(p, rq); | |
4065 | /* | |
4066 | * Reschedule if we are currently running on this runqueue and | |
4067 | * our priority decreased, or if we are not currently running on | |
4068 | * this runqueue and our priority is higher than the current's | |
4069 | */ | |
4070 | if (task_running(rq, p)) { | |
4071 | if (p->prio > oldprio) | |
4072 | resched_task(rq->curr); | |
4073 | } else if (TASK_PREEMPTS_CURR(p, rq)) | |
4074 | resched_task(rq->curr); | |
4075 | } | |
b29739f9 IM |
4076 | __task_rq_unlock(rq); |
4077 | spin_unlock_irqrestore(&p->pi_lock, flags); | |
4078 | ||
95e02ca9 TG |
4079 | rt_mutex_adjust_pi(p); |
4080 | ||
1da177e4 LT |
4081 | return 0; |
4082 | } | |
4083 | EXPORT_SYMBOL_GPL(sched_setscheduler); | |
4084 | ||
95cdf3b7 IM |
4085 | static int |
4086 | do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) | |
1da177e4 LT |
4087 | { |
4088 | int retval; | |
4089 | struct sched_param lparam; | |
4090 | struct task_struct *p; | |
4091 | ||
4092 | if (!param || pid < 0) | |
4093 | return -EINVAL; | |
4094 | if (copy_from_user(&lparam, param, sizeof(struct sched_param))) | |
4095 | return -EFAULT; | |
4096 | read_lock_irq(&tasklist_lock); | |
4097 | p = find_process_by_pid(pid); | |
4098 | if (!p) { | |
4099 | read_unlock_irq(&tasklist_lock); | |
4100 | return -ESRCH; | |
4101 | } | |
e74c69f4 | 4102 | get_task_struct(p); |
1da177e4 | 4103 | read_unlock_irq(&tasklist_lock); |
e74c69f4 TG |
4104 | retval = sched_setscheduler(p, policy, &lparam); |
4105 | put_task_struct(p); | |
1da177e4 LT |
4106 | return retval; |
4107 | } | |
4108 | ||
4109 | /** | |
4110 | * sys_sched_setscheduler - set/change the scheduler policy and RT priority | |
4111 | * @pid: the pid in question. | |
4112 | * @policy: new policy. | |
4113 | * @param: structure containing the new RT priority. | |
4114 | */ | |
4115 | asmlinkage long sys_sched_setscheduler(pid_t pid, int policy, | |
4116 | struct sched_param __user *param) | |
4117 | { | |
c21761f1 JB |
4118 | /* negative values for policy are not valid */ |
4119 | if (policy < 0) | |
4120 | return -EINVAL; | |
4121 | ||
1da177e4 LT |
4122 | return do_sched_setscheduler(pid, policy, param); |
4123 | } | |
4124 | ||
4125 | /** | |
4126 | * sys_sched_setparam - set/change the RT priority of a thread | |
4127 | * @pid: the pid in question. | |
4128 | * @param: structure containing the new RT priority. | |
4129 | */ | |
4130 | asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param) | |
4131 | { | |
4132 | return do_sched_setscheduler(pid, -1, param); | |
4133 | } | |
4134 | ||
4135 | /** | |
4136 | * sys_sched_getscheduler - get the policy (scheduling class) of a thread | |
4137 | * @pid: the pid in question. | |
4138 | */ | |
4139 | asmlinkage long sys_sched_getscheduler(pid_t pid) | |
4140 | { | |
4141 | int retval = -EINVAL; | |
4142 | task_t *p; | |
4143 | ||
4144 | if (pid < 0) | |
4145 | goto out_nounlock; | |
4146 | ||
4147 | retval = -ESRCH; | |
4148 | read_lock(&tasklist_lock); | |
4149 | p = find_process_by_pid(pid); | |
4150 | if (p) { | |
4151 | retval = security_task_getscheduler(p); | |
4152 | if (!retval) | |
4153 | retval = p->policy; | |
4154 | } | |
4155 | read_unlock(&tasklist_lock); | |
4156 | ||
4157 | out_nounlock: | |
4158 | return retval; | |
4159 | } | |
4160 | ||
4161 | /** | |
4162 | * sys_sched_getscheduler - get the RT priority of a thread | |
4163 | * @pid: the pid in question. | |
4164 | * @param: structure containing the RT priority. | |
4165 | */ | |
4166 | asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param) | |
4167 | { | |
4168 | struct sched_param lp; | |
4169 | int retval = -EINVAL; | |
4170 | task_t *p; | |
4171 | ||
4172 | if (!param || pid < 0) | |
4173 | goto out_nounlock; | |
4174 | ||
4175 | read_lock(&tasklist_lock); | |
4176 | p = find_process_by_pid(pid); | |
4177 | retval = -ESRCH; | |
4178 | if (!p) | |
4179 | goto out_unlock; | |
4180 | ||
4181 | retval = security_task_getscheduler(p); | |
4182 | if (retval) | |
4183 | goto out_unlock; | |
4184 | ||
4185 | lp.sched_priority = p->rt_priority; | |
4186 | read_unlock(&tasklist_lock); | |
4187 | ||
4188 | /* | |
4189 | * This one might sleep, we cannot do it with a spinlock held ... | |
4190 | */ | |
4191 | retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; | |
4192 | ||
4193 | out_nounlock: | |
4194 | return retval; | |
4195 | ||
4196 | out_unlock: | |
4197 | read_unlock(&tasklist_lock); | |
4198 | return retval; | |
4199 | } | |
4200 | ||
4201 | long sched_setaffinity(pid_t pid, cpumask_t new_mask) | |
4202 | { | |
4203 | task_t *p; | |
4204 | int retval; | |
4205 | cpumask_t cpus_allowed; | |
4206 | ||
4207 | lock_cpu_hotplug(); | |
4208 | read_lock(&tasklist_lock); | |
4209 | ||
4210 | p = find_process_by_pid(pid); | |
4211 | if (!p) { | |
4212 | read_unlock(&tasklist_lock); | |
4213 | unlock_cpu_hotplug(); | |
4214 | return -ESRCH; | |
4215 | } | |
4216 | ||
4217 | /* | |
4218 | * It is not safe to call set_cpus_allowed with the | |
4219 | * tasklist_lock held. We will bump the task_struct's | |
4220 | * usage count and then drop tasklist_lock. | |
4221 | */ | |
4222 | get_task_struct(p); | |
4223 | read_unlock(&tasklist_lock); | |
4224 | ||
4225 | retval = -EPERM; | |
4226 | if ((current->euid != p->euid) && (current->euid != p->uid) && | |
4227 | !capable(CAP_SYS_NICE)) | |
4228 | goto out_unlock; | |
4229 | ||
e7834f8f DQ |
4230 | retval = security_task_setscheduler(p, 0, NULL); |
4231 | if (retval) | |
4232 | goto out_unlock; | |
4233 | ||
1da177e4 LT |
4234 | cpus_allowed = cpuset_cpus_allowed(p); |
4235 | cpus_and(new_mask, new_mask, cpus_allowed); | |
4236 | retval = set_cpus_allowed(p, new_mask); | |
4237 | ||
4238 | out_unlock: | |
4239 | put_task_struct(p); | |
4240 | unlock_cpu_hotplug(); | |
4241 | return retval; | |
4242 | } | |
4243 | ||
4244 | static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, | |
4245 | cpumask_t *new_mask) | |
4246 | { | |
4247 | if (len < sizeof(cpumask_t)) { | |
4248 | memset(new_mask, 0, sizeof(cpumask_t)); | |
4249 | } else if (len > sizeof(cpumask_t)) { | |
4250 | len = sizeof(cpumask_t); | |
4251 | } | |
4252 | return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; | |
4253 | } | |
4254 | ||
4255 | /** | |
4256 | * sys_sched_setaffinity - set the cpu affinity of a process | |
4257 | * @pid: pid of the process | |
4258 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
4259 | * @user_mask_ptr: user-space pointer to the new cpu mask | |
4260 | */ | |
4261 | asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len, | |
4262 | unsigned long __user *user_mask_ptr) | |
4263 | { | |
4264 | cpumask_t new_mask; | |
4265 | int retval; | |
4266 | ||
4267 | retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask); | |
4268 | if (retval) | |
4269 | return retval; | |
4270 | ||
4271 | return sched_setaffinity(pid, new_mask); | |
4272 | } | |
4273 | ||
4274 | /* | |
4275 | * Represents all cpu's present in the system | |
4276 | * In systems capable of hotplug, this map could dynamically grow | |
4277 | * as new cpu's are detected in the system via any platform specific | |
4278 | * method, such as ACPI for e.g. | |
4279 | */ | |
4280 | ||
4cef0c61 | 4281 | cpumask_t cpu_present_map __read_mostly; |
1da177e4 LT |
4282 | EXPORT_SYMBOL(cpu_present_map); |
4283 | ||
4284 | #ifndef CONFIG_SMP | |
4cef0c61 AK |
4285 | cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL; |
4286 | cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL; | |
1da177e4 LT |
4287 | #endif |
4288 | ||
4289 | long sched_getaffinity(pid_t pid, cpumask_t *mask) | |
4290 | { | |
4291 | int retval; | |
4292 | task_t *p; | |
4293 | ||
4294 | lock_cpu_hotplug(); | |
4295 | read_lock(&tasklist_lock); | |
4296 | ||
4297 | retval = -ESRCH; | |
4298 | p = find_process_by_pid(pid); | |
4299 | if (!p) | |
4300 | goto out_unlock; | |
4301 | ||
e7834f8f DQ |
4302 | retval = security_task_getscheduler(p); |
4303 | if (retval) | |
4304 | goto out_unlock; | |
4305 | ||
2f7016d9 | 4306 | cpus_and(*mask, p->cpus_allowed, cpu_online_map); |
1da177e4 LT |
4307 | |
4308 | out_unlock: | |
4309 | read_unlock(&tasklist_lock); | |
4310 | unlock_cpu_hotplug(); | |
4311 | if (retval) | |
4312 | return retval; | |
4313 | ||
4314 | return 0; | |
4315 | } | |
4316 | ||
4317 | /** | |
4318 | * sys_sched_getaffinity - get the cpu affinity of a process | |
4319 | * @pid: pid of the process | |
4320 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
4321 | * @user_mask_ptr: user-space pointer to hold the current cpu mask | |
4322 | */ | |
4323 | asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len, | |
4324 | unsigned long __user *user_mask_ptr) | |
4325 | { | |
4326 | int ret; | |
4327 | cpumask_t mask; | |
4328 | ||
4329 | if (len < sizeof(cpumask_t)) | |
4330 | return -EINVAL; | |
4331 | ||
4332 | ret = sched_getaffinity(pid, &mask); | |
4333 | if (ret < 0) | |
4334 | return ret; | |
4335 | ||
4336 | if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t))) | |
4337 | return -EFAULT; | |
4338 | ||
4339 | return sizeof(cpumask_t); | |
4340 | } | |
4341 | ||
4342 | /** | |
4343 | * sys_sched_yield - yield the current processor to other threads. | |
4344 | * | |
4345 | * this function yields the current CPU by moving the calling thread | |
4346 | * to the expired array. If there are no other threads running on this | |
4347 | * CPU then this function will return. | |
4348 | */ | |
4349 | asmlinkage long sys_sched_yield(void) | |
4350 | { | |
4351 | runqueue_t *rq = this_rq_lock(); | |
4352 | prio_array_t *array = current->array; | |
4353 | prio_array_t *target = rq->expired; | |
4354 | ||
4355 | schedstat_inc(rq, yld_cnt); | |
4356 | /* | |
4357 | * We implement yielding by moving the task into the expired | |
4358 | * queue. | |
4359 | * | |
4360 | * (special rule: RT tasks will just roundrobin in the active | |
4361 | * array.) | |
4362 | */ | |
4363 | if (rt_task(current)) | |
4364 | target = rq->active; | |
4365 | ||
5927ad78 | 4366 | if (array->nr_active == 1) { |
1da177e4 LT |
4367 | schedstat_inc(rq, yld_act_empty); |
4368 | if (!rq->expired->nr_active) | |
4369 | schedstat_inc(rq, yld_both_empty); | |
4370 | } else if (!rq->expired->nr_active) | |
4371 | schedstat_inc(rq, yld_exp_empty); | |
4372 | ||
4373 | if (array != target) { | |
4374 | dequeue_task(current, array); | |
4375 | enqueue_task(current, target); | |
4376 | } else | |
4377 | /* | |
4378 | * requeue_task is cheaper so perform that if possible. | |
4379 | */ | |
4380 | requeue_task(current, array); | |
4381 | ||
4382 | /* | |
4383 | * Since we are going to call schedule() anyway, there's | |
4384 | * no need to preempt or enable interrupts: | |
4385 | */ | |
4386 | __release(rq->lock); | |
4387 | _raw_spin_unlock(&rq->lock); | |
4388 | preempt_enable_no_resched(); | |
4389 | ||
4390 | schedule(); | |
4391 | ||
4392 | return 0; | |
4393 | } | |
4394 | ||
e7b38404 AM |
4395 | static inline int __resched_legal(void) |
4396 | { | |
4397 | if (unlikely(preempt_count())) | |
4398 | return 0; | |
4399 | if (unlikely(system_state != SYSTEM_RUNNING)) | |
4400 | return 0; | |
4401 | return 1; | |
4402 | } | |
4403 | ||
4404 | static void __cond_resched(void) | |
1da177e4 | 4405 | { |
8e0a43d8 IM |
4406 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP |
4407 | __might_sleep(__FILE__, __LINE__); | |
4408 | #endif | |
5bbcfd90 IM |
4409 | /* |
4410 | * The BKS might be reacquired before we have dropped | |
4411 | * PREEMPT_ACTIVE, which could trigger a second | |
4412 | * cond_resched() call. | |
4413 | */ | |
1da177e4 LT |
4414 | do { |
4415 | add_preempt_count(PREEMPT_ACTIVE); | |
4416 | schedule(); | |
4417 | sub_preempt_count(PREEMPT_ACTIVE); | |
4418 | } while (need_resched()); | |
4419 | } | |
4420 | ||
4421 | int __sched cond_resched(void) | |
4422 | { | |
e7b38404 | 4423 | if (need_resched() && __resched_legal()) { |
1da177e4 LT |
4424 | __cond_resched(); |
4425 | return 1; | |
4426 | } | |
4427 | return 0; | |
4428 | } | |
1da177e4 LT |
4429 | EXPORT_SYMBOL(cond_resched); |
4430 | ||
4431 | /* | |
4432 | * cond_resched_lock() - if a reschedule is pending, drop the given lock, | |
4433 | * call schedule, and on return reacquire the lock. | |
4434 | * | |
4435 | * This works OK both with and without CONFIG_PREEMPT. We do strange low-level | |
4436 | * operations here to prevent schedule() from being called twice (once via | |
4437 | * spin_unlock(), once by hand). | |
4438 | */ | |
95cdf3b7 | 4439 | int cond_resched_lock(spinlock_t *lock) |
1da177e4 | 4440 | { |
6df3cecb JK |
4441 | int ret = 0; |
4442 | ||
1da177e4 LT |
4443 | if (need_lockbreak(lock)) { |
4444 | spin_unlock(lock); | |
4445 | cpu_relax(); | |
6df3cecb | 4446 | ret = 1; |
1da177e4 LT |
4447 | spin_lock(lock); |
4448 | } | |
e7b38404 | 4449 | if (need_resched() && __resched_legal()) { |
1da177e4 LT |
4450 | _raw_spin_unlock(lock); |
4451 | preempt_enable_no_resched(); | |
4452 | __cond_resched(); | |
6df3cecb | 4453 | ret = 1; |
1da177e4 | 4454 | spin_lock(lock); |
1da177e4 | 4455 | } |
6df3cecb | 4456 | return ret; |
1da177e4 | 4457 | } |
1da177e4 LT |
4458 | EXPORT_SYMBOL(cond_resched_lock); |
4459 | ||
4460 | int __sched cond_resched_softirq(void) | |
4461 | { | |
4462 | BUG_ON(!in_softirq()); | |
4463 | ||
e7b38404 | 4464 | if (need_resched() && __resched_legal()) { |
de30a2b3 IM |
4465 | raw_local_irq_disable(); |
4466 | _local_bh_enable(); | |
4467 | raw_local_irq_enable(); | |
1da177e4 LT |
4468 | __cond_resched(); |
4469 | local_bh_disable(); | |
4470 | return 1; | |
4471 | } | |
4472 | return 0; | |
4473 | } | |
1da177e4 LT |
4474 | EXPORT_SYMBOL(cond_resched_softirq); |
4475 | ||
1da177e4 LT |
4476 | /** |
4477 | * yield - yield the current processor to other threads. | |
4478 | * | |
4479 | * this is a shortcut for kernel-space yielding - it marks the | |
4480 | * thread runnable and calls sys_sched_yield(). | |
4481 | */ | |
4482 | void __sched yield(void) | |
4483 | { | |
4484 | set_current_state(TASK_RUNNING); | |
4485 | sys_sched_yield(); | |
4486 | } | |
4487 | ||
4488 | EXPORT_SYMBOL(yield); | |
4489 | ||
4490 | /* | |
4491 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so | |
4492 | * that process accounting knows that this is a task in IO wait state. | |
4493 | * | |
4494 | * But don't do that if it is a deliberate, throttling IO wait (this task | |
4495 | * has set its backing_dev_info: the queue against which it should throttle) | |
4496 | */ | |
4497 | void __sched io_schedule(void) | |
4498 | { | |
bfe5d834 | 4499 | struct runqueue *rq = &__raw_get_cpu_var(runqueues); |
1da177e4 LT |
4500 | |
4501 | atomic_inc(&rq->nr_iowait); | |
4502 | schedule(); | |
4503 | atomic_dec(&rq->nr_iowait); | |
4504 | } | |
4505 | ||
4506 | EXPORT_SYMBOL(io_schedule); | |
4507 | ||
4508 | long __sched io_schedule_timeout(long timeout) | |
4509 | { | |
bfe5d834 | 4510 | struct runqueue *rq = &__raw_get_cpu_var(runqueues); |
1da177e4 LT |
4511 | long ret; |
4512 | ||
4513 | atomic_inc(&rq->nr_iowait); | |
4514 | ret = schedule_timeout(timeout); | |
4515 | atomic_dec(&rq->nr_iowait); | |
4516 | return ret; | |
4517 | } | |
4518 | ||
4519 | /** | |
4520 | * sys_sched_get_priority_max - return maximum RT priority. | |
4521 | * @policy: scheduling class. | |
4522 | * | |
4523 | * this syscall returns the maximum rt_priority that can be used | |
4524 | * by a given scheduling class. | |
4525 | */ | |
4526 | asmlinkage long sys_sched_get_priority_max(int policy) | |
4527 | { | |
4528 | int ret = -EINVAL; | |
4529 | ||
4530 | switch (policy) { | |
4531 | case SCHED_FIFO: | |
4532 | case SCHED_RR: | |
4533 | ret = MAX_USER_RT_PRIO-1; | |
4534 | break; | |
4535 | case SCHED_NORMAL: | |
b0a9499c | 4536 | case SCHED_BATCH: |
1da177e4 LT |
4537 | ret = 0; |
4538 | break; | |
4539 | } | |
4540 | return ret; | |
4541 | } | |
4542 | ||
4543 | /** | |
4544 | * sys_sched_get_priority_min - return minimum RT priority. | |
4545 | * @policy: scheduling class. | |
4546 | * | |
4547 | * this syscall returns the minimum rt_priority that can be used | |
4548 | * by a given scheduling class. | |
4549 | */ | |
4550 | asmlinkage long sys_sched_get_priority_min(int policy) | |
4551 | { | |
4552 | int ret = -EINVAL; | |
4553 | ||
4554 | switch (policy) { | |
4555 | case SCHED_FIFO: | |
4556 | case SCHED_RR: | |
4557 | ret = 1; | |
4558 | break; | |
4559 | case SCHED_NORMAL: | |
b0a9499c | 4560 | case SCHED_BATCH: |
1da177e4 LT |
4561 | ret = 0; |
4562 | } | |
4563 | return ret; | |
4564 | } | |
4565 | ||
4566 | /** | |
4567 | * sys_sched_rr_get_interval - return the default timeslice of a process. | |
4568 | * @pid: pid of the process. | |
4569 | * @interval: userspace pointer to the timeslice value. | |
4570 | * | |
4571 | * this syscall writes the default timeslice value of a given process | |
4572 | * into the user-space timespec buffer. A value of '0' means infinity. | |
4573 | */ | |
4574 | asmlinkage | |
4575 | long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval) | |
4576 | { | |
4577 | int retval = -EINVAL; | |
4578 | struct timespec t; | |
4579 | task_t *p; | |
4580 | ||
4581 | if (pid < 0) | |
4582 | goto out_nounlock; | |
4583 | ||
4584 | retval = -ESRCH; | |
4585 | read_lock(&tasklist_lock); | |
4586 | p = find_process_by_pid(pid); | |
4587 | if (!p) | |
4588 | goto out_unlock; | |
4589 | ||
4590 | retval = security_task_getscheduler(p); | |
4591 | if (retval) | |
4592 | goto out_unlock; | |
4593 | ||
b78709cf | 4594 | jiffies_to_timespec(p->policy == SCHED_FIFO ? |
1da177e4 LT |
4595 | 0 : task_timeslice(p), &t); |
4596 | read_unlock(&tasklist_lock); | |
4597 | retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; | |
4598 | out_nounlock: | |
4599 | return retval; | |
4600 | out_unlock: | |
4601 | read_unlock(&tasklist_lock); | |
4602 | return retval; | |
4603 | } | |
4604 | ||
4605 | static inline struct task_struct *eldest_child(struct task_struct *p) | |
4606 | { | |
4607 | if (list_empty(&p->children)) return NULL; | |
4608 | return list_entry(p->children.next,struct task_struct,sibling); | |
4609 | } | |
4610 | ||
4611 | static inline struct task_struct *older_sibling(struct task_struct *p) | |
4612 | { | |
4613 | if (p->sibling.prev==&p->parent->children) return NULL; | |
4614 | return list_entry(p->sibling.prev,struct task_struct,sibling); | |
4615 | } | |
4616 | ||
4617 | static inline struct task_struct *younger_sibling(struct task_struct *p) | |
4618 | { | |
4619 | if (p->sibling.next==&p->parent->children) return NULL; | |
4620 | return list_entry(p->sibling.next,struct task_struct,sibling); | |
4621 | } | |
4622 | ||
95cdf3b7 | 4623 | static void show_task(task_t *p) |
1da177e4 LT |
4624 | { |
4625 | task_t *relative; | |
4626 | unsigned state; | |
4627 | unsigned long free = 0; | |
4628 | static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" }; | |
4629 | ||
4630 | printk("%-13.13s ", p->comm); | |
4631 | state = p->state ? __ffs(p->state) + 1 : 0; | |
4632 | if (state < ARRAY_SIZE(stat_nam)) | |
4633 | printk(stat_nam[state]); | |
4634 | else | |
4635 | printk("?"); | |
4636 | #if (BITS_PER_LONG == 32) | |
4637 | if (state == TASK_RUNNING) | |
4638 | printk(" running "); | |
4639 | else | |
4640 | printk(" %08lX ", thread_saved_pc(p)); | |
4641 | #else | |
4642 | if (state == TASK_RUNNING) | |
4643 | printk(" running task "); | |
4644 | else | |
4645 | printk(" %016lx ", thread_saved_pc(p)); | |
4646 | #endif | |
4647 | #ifdef CONFIG_DEBUG_STACK_USAGE | |
4648 | { | |
10ebffde | 4649 | unsigned long *n = end_of_stack(p); |
1da177e4 LT |
4650 | while (!*n) |
4651 | n++; | |
10ebffde | 4652 | free = (unsigned long)n - (unsigned long)end_of_stack(p); |
1da177e4 LT |
4653 | } |
4654 | #endif | |
4655 | printk("%5lu %5d %6d ", free, p->pid, p->parent->pid); | |
4656 | if ((relative = eldest_child(p))) | |
4657 | printk("%5d ", relative->pid); | |
4658 | else | |
4659 | printk(" "); | |
4660 | if ((relative = younger_sibling(p))) | |
4661 | printk("%7d", relative->pid); | |
4662 | else | |
4663 | printk(" "); | |
4664 | if ((relative = older_sibling(p))) | |
4665 | printk(" %5d", relative->pid); | |
4666 | else | |
4667 | printk(" "); | |
4668 | if (!p->mm) | |
4669 | printk(" (L-TLB)\n"); | |
4670 | else | |
4671 | printk(" (NOTLB)\n"); | |
4672 | ||
4673 | if (state != TASK_RUNNING) | |
4674 | show_stack(p, NULL); | |
4675 | } | |
4676 | ||
4677 | void show_state(void) | |
4678 | { | |
4679 | task_t *g, *p; | |
4680 | ||
4681 | #if (BITS_PER_LONG == 32) | |
4682 | printk("\n" | |
4683 | " sibling\n"); | |
4684 | printk(" task PC pid father child younger older\n"); | |
4685 | #else | |
4686 | printk("\n" | |
4687 | " sibling\n"); | |
4688 | printk(" task PC pid father child younger older\n"); | |
4689 | #endif | |
4690 | read_lock(&tasklist_lock); | |
4691 | do_each_thread(g, p) { | |
4692 | /* | |
4693 | * reset the NMI-timeout, listing all files on a slow | |
4694 | * console might take alot of time: | |
4695 | */ | |
4696 | touch_nmi_watchdog(); | |
4697 | show_task(p); | |
4698 | } while_each_thread(g, p); | |
4699 | ||
4700 | read_unlock(&tasklist_lock); | |
9a11b49a | 4701 | debug_show_all_locks(); |
1da177e4 LT |
4702 | } |
4703 | ||
f340c0d1 IM |
4704 | /** |
4705 | * init_idle - set up an idle thread for a given CPU | |
4706 | * @idle: task in question | |
4707 | * @cpu: cpu the idle task belongs to | |
4708 | * | |
4709 | * NOTE: this function does not set the idle thread's NEED_RESCHED | |
4710 | * flag, to make booting more robust. | |
4711 | */ | |
1da177e4 LT |
4712 | void __devinit init_idle(task_t *idle, int cpu) |
4713 | { | |
4714 | runqueue_t *rq = cpu_rq(cpu); | |
4715 | unsigned long flags; | |
4716 | ||
81c29a85 | 4717 | idle->timestamp = sched_clock(); |
1da177e4 LT |
4718 | idle->sleep_avg = 0; |
4719 | idle->array = NULL; | |
b29739f9 | 4720 | idle->prio = idle->normal_prio = MAX_PRIO; |
1da177e4 LT |
4721 | idle->state = TASK_RUNNING; |
4722 | idle->cpus_allowed = cpumask_of_cpu(cpu); | |
4723 | set_task_cpu(idle, cpu); | |
4724 | ||
4725 | spin_lock_irqsave(&rq->lock, flags); | |
4726 | rq->curr = rq->idle = idle; | |
4866cde0 NP |
4727 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
4728 | idle->oncpu = 1; | |
4729 | #endif | |
1da177e4 LT |
4730 | spin_unlock_irqrestore(&rq->lock, flags); |
4731 | ||
4732 | /* Set the preempt count _outside_ the spinlocks! */ | |
4733 | #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL) | |
a1261f54 | 4734 | task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0); |
1da177e4 | 4735 | #else |
a1261f54 | 4736 | task_thread_info(idle)->preempt_count = 0; |
1da177e4 LT |
4737 | #endif |
4738 | } | |
4739 | ||
4740 | /* | |
4741 | * In a system that switches off the HZ timer nohz_cpu_mask | |
4742 | * indicates which cpus entered this state. This is used | |
4743 | * in the rcu update to wait only for active cpus. For system | |
4744 | * which do not switch off the HZ timer nohz_cpu_mask should | |
4745 | * always be CPU_MASK_NONE. | |
4746 | */ | |
4747 | cpumask_t nohz_cpu_mask = CPU_MASK_NONE; | |
4748 | ||
4749 | #ifdef CONFIG_SMP | |
4750 | /* | |
4751 | * This is how migration works: | |
4752 | * | |
4753 | * 1) we queue a migration_req_t structure in the source CPU's | |
4754 | * runqueue and wake up that CPU's migration thread. | |
4755 | * 2) we down() the locked semaphore => thread blocks. | |
4756 | * 3) migration thread wakes up (implicitly it forces the migrated | |
4757 | * thread off the CPU) | |
4758 | * 4) it gets the migration request and checks whether the migrated | |
4759 | * task is still in the wrong runqueue. | |
4760 | * 5) if it's in the wrong runqueue then the migration thread removes | |
4761 | * it and puts it into the right queue. | |
4762 | * 6) migration thread up()s the semaphore. | |
4763 | * 7) we wake up and the migration is done. | |
4764 | */ | |
4765 | ||
4766 | /* | |
4767 | * Change a given task's CPU affinity. Migrate the thread to a | |
4768 | * proper CPU and schedule it away if the CPU it's executing on | |
4769 | * is removed from the allowed bitmask. | |
4770 | * | |
4771 | * NOTE: the caller must have a valid reference to the task, the | |
4772 | * task must not exit() & deallocate itself prematurely. The | |
4773 | * call is not atomic; no spinlocks may be held. | |
4774 | */ | |
4775 | int set_cpus_allowed(task_t *p, cpumask_t new_mask) | |
4776 | { | |
4777 | unsigned long flags; | |
4778 | int ret = 0; | |
4779 | migration_req_t req; | |
4780 | runqueue_t *rq; | |
4781 | ||
4782 | rq = task_rq_lock(p, &flags); | |
4783 | if (!cpus_intersects(new_mask, cpu_online_map)) { | |
4784 | ret = -EINVAL; | |
4785 | goto out; | |
4786 | } | |
4787 | ||
4788 | p->cpus_allowed = new_mask; | |
4789 | /* Can the task run on the task's current CPU? If so, we're done */ | |
4790 | if (cpu_isset(task_cpu(p), new_mask)) | |
4791 | goto out; | |
4792 | ||
4793 | if (migrate_task(p, any_online_cpu(new_mask), &req)) { | |
4794 | /* Need help from migration thread: drop lock and wait. */ | |
4795 | task_rq_unlock(rq, &flags); | |
4796 | wake_up_process(rq->migration_thread); | |
4797 | wait_for_completion(&req.done); | |
4798 | tlb_migrate_finish(p->mm); | |
4799 | return 0; | |
4800 | } | |
4801 | out: | |
4802 | task_rq_unlock(rq, &flags); | |
4803 | return ret; | |
4804 | } | |
4805 | ||
4806 | EXPORT_SYMBOL_GPL(set_cpus_allowed); | |
4807 | ||
4808 | /* | |
4809 | * Move (not current) task off this cpu, onto dest cpu. We're doing | |
4810 | * this because either it can't run here any more (set_cpus_allowed() | |
4811 | * away from this CPU, or CPU going down), or because we're | |
4812 | * attempting to rebalance this task on exec (sched_exec). | |
4813 | * | |
4814 | * So we race with normal scheduler movements, but that's OK, as long | |
4815 | * as the task is no longer on this CPU. | |
efc30814 KK |
4816 | * |
4817 | * Returns non-zero if task was successfully migrated. | |
1da177e4 | 4818 | */ |
efc30814 | 4819 | static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) |
1da177e4 LT |
4820 | { |
4821 | runqueue_t *rq_dest, *rq_src; | |
efc30814 | 4822 | int ret = 0; |
1da177e4 LT |
4823 | |
4824 | if (unlikely(cpu_is_offline(dest_cpu))) | |
efc30814 | 4825 | return ret; |
1da177e4 LT |
4826 | |
4827 | rq_src = cpu_rq(src_cpu); | |
4828 | rq_dest = cpu_rq(dest_cpu); | |
4829 | ||
4830 | double_rq_lock(rq_src, rq_dest); | |
4831 | /* Already moved. */ | |
4832 | if (task_cpu(p) != src_cpu) | |
4833 | goto out; | |
4834 | /* Affinity changed (again). */ | |
4835 | if (!cpu_isset(dest_cpu, p->cpus_allowed)) | |
4836 | goto out; | |
4837 | ||
4838 | set_task_cpu(p, dest_cpu); | |
4839 | if (p->array) { | |
4840 | /* | |
4841 | * Sync timestamp with rq_dest's before activating. | |
4842 | * The same thing could be achieved by doing this step | |
4843 | * afterwards, and pretending it was a local activate. | |
4844 | * This way is cleaner and logically correct. | |
4845 | */ | |
4846 | p->timestamp = p->timestamp - rq_src->timestamp_last_tick | |
4847 | + rq_dest->timestamp_last_tick; | |
4848 | deactivate_task(p, rq_src); | |
4849 | activate_task(p, rq_dest, 0); | |
4850 | if (TASK_PREEMPTS_CURR(p, rq_dest)) | |
4851 | resched_task(rq_dest->curr); | |
4852 | } | |
efc30814 | 4853 | ret = 1; |
1da177e4 LT |
4854 | out: |
4855 | double_rq_unlock(rq_src, rq_dest); | |
efc30814 | 4856 | return ret; |
1da177e4 LT |
4857 | } |
4858 | ||
4859 | /* | |
4860 | * migration_thread - this is a highprio system thread that performs | |
4861 | * thread migration by bumping thread off CPU then 'pushing' onto | |
4862 | * another runqueue. | |
4863 | */ | |
95cdf3b7 | 4864 | static int migration_thread(void *data) |
1da177e4 LT |
4865 | { |
4866 | runqueue_t *rq; | |
4867 | int cpu = (long)data; | |
4868 | ||
4869 | rq = cpu_rq(cpu); | |
4870 | BUG_ON(rq->migration_thread != current); | |
4871 | ||
4872 | set_current_state(TASK_INTERRUPTIBLE); | |
4873 | while (!kthread_should_stop()) { | |
4874 | struct list_head *head; | |
4875 | migration_req_t *req; | |
4876 | ||
3e1d1d28 | 4877 | try_to_freeze(); |
1da177e4 LT |
4878 | |
4879 | spin_lock_irq(&rq->lock); | |
4880 | ||
4881 | if (cpu_is_offline(cpu)) { | |
4882 | spin_unlock_irq(&rq->lock); | |
4883 | goto wait_to_die; | |
4884 | } | |
4885 | ||
4886 | if (rq->active_balance) { | |
4887 | active_load_balance(rq, cpu); | |
4888 | rq->active_balance = 0; | |
4889 | } | |
4890 | ||
4891 | head = &rq->migration_queue; | |
4892 | ||
4893 | if (list_empty(head)) { | |
4894 | spin_unlock_irq(&rq->lock); | |
4895 | schedule(); | |
4896 | set_current_state(TASK_INTERRUPTIBLE); | |
4897 | continue; | |
4898 | } | |
4899 | req = list_entry(head->next, migration_req_t, list); | |
4900 | list_del_init(head->next); | |
4901 | ||
674311d5 NP |
4902 | spin_unlock(&rq->lock); |
4903 | __migrate_task(req->task, cpu, req->dest_cpu); | |
4904 | local_irq_enable(); | |
1da177e4 LT |
4905 | |
4906 | complete(&req->done); | |
4907 | } | |
4908 | __set_current_state(TASK_RUNNING); | |
4909 | return 0; | |
4910 | ||
4911 | wait_to_die: | |
4912 | /* Wait for kthread_stop */ | |
4913 | set_current_state(TASK_INTERRUPTIBLE); | |
4914 | while (!kthread_should_stop()) { | |
4915 | schedule(); | |
4916 | set_current_state(TASK_INTERRUPTIBLE); | |
4917 | } | |
4918 | __set_current_state(TASK_RUNNING); | |
4919 | return 0; | |
4920 | } | |
4921 | ||
4922 | #ifdef CONFIG_HOTPLUG_CPU | |
4923 | /* Figure out where task on dead CPU should go, use force if neccessary. */ | |
4924 | static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk) | |
4925 | { | |
efc30814 KK |
4926 | runqueue_t *rq; |
4927 | unsigned long flags; | |
1da177e4 LT |
4928 | int dest_cpu; |
4929 | cpumask_t mask; | |
4930 | ||
efc30814 | 4931 | restart: |
1da177e4 LT |
4932 | /* On same node? */ |
4933 | mask = node_to_cpumask(cpu_to_node(dead_cpu)); | |
4934 | cpus_and(mask, mask, tsk->cpus_allowed); | |
4935 | dest_cpu = any_online_cpu(mask); | |
4936 | ||
4937 | /* On any allowed CPU? */ | |
4938 | if (dest_cpu == NR_CPUS) | |
4939 | dest_cpu = any_online_cpu(tsk->cpus_allowed); | |
4940 | ||
4941 | /* No more Mr. Nice Guy. */ | |
4942 | if (dest_cpu == NR_CPUS) { | |
efc30814 | 4943 | rq = task_rq_lock(tsk, &flags); |
b39c4fab | 4944 | cpus_setall(tsk->cpus_allowed); |
1da177e4 | 4945 | dest_cpu = any_online_cpu(tsk->cpus_allowed); |
efc30814 | 4946 | task_rq_unlock(rq, &flags); |
1da177e4 LT |
4947 | |
4948 | /* | |
4949 | * Don't tell them about moving exiting tasks or | |
4950 | * kernel threads (both mm NULL), since they never | |
4951 | * leave kernel. | |
4952 | */ | |
4953 | if (tsk->mm && printk_ratelimit()) | |
4954 | printk(KERN_INFO "process %d (%s) no " | |
4955 | "longer affine to cpu%d\n", | |
4956 | tsk->pid, tsk->comm, dead_cpu); | |
4957 | } | |
efc30814 KK |
4958 | if (!__migrate_task(tsk, dead_cpu, dest_cpu)) |
4959 | goto restart; | |
1da177e4 LT |
4960 | } |
4961 | ||
4962 | /* | |
4963 | * While a dead CPU has no uninterruptible tasks queued at this point, | |
4964 | * it might still have a nonzero ->nr_uninterruptible counter, because | |
4965 | * for performance reasons the counter is not stricly tracking tasks to | |
4966 | * their home CPUs. So we just add the counter to another CPU's counter, | |
4967 | * to keep the global sum constant after CPU-down: | |
4968 | */ | |
4969 | static void migrate_nr_uninterruptible(runqueue_t *rq_src) | |
4970 | { | |
4971 | runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL)); | |
4972 | unsigned long flags; | |
4973 | ||
4974 | local_irq_save(flags); | |
4975 | double_rq_lock(rq_src, rq_dest); | |
4976 | rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible; | |
4977 | rq_src->nr_uninterruptible = 0; | |
4978 | double_rq_unlock(rq_src, rq_dest); | |
4979 | local_irq_restore(flags); | |
4980 | } | |
4981 | ||
4982 | /* Run through task list and migrate tasks from the dead cpu. */ | |
4983 | static void migrate_live_tasks(int src_cpu) | |
4984 | { | |
4985 | struct task_struct *tsk, *t; | |
4986 | ||
4987 | write_lock_irq(&tasklist_lock); | |
4988 | ||
4989 | do_each_thread(t, tsk) { | |
4990 | if (tsk == current) | |
4991 | continue; | |
4992 | ||
4993 | if (task_cpu(tsk) == src_cpu) | |
4994 | move_task_off_dead_cpu(src_cpu, tsk); | |
4995 | } while_each_thread(t, tsk); | |
4996 | ||
4997 | write_unlock_irq(&tasklist_lock); | |
4998 | } | |
4999 | ||
5000 | /* Schedules idle task to be the next runnable task on current CPU. | |
5001 | * It does so by boosting its priority to highest possible and adding it to | |
5002 | * the _front_ of runqueue. Used by CPU offline code. | |
5003 | */ | |
5004 | void sched_idle_next(void) | |
5005 | { | |
5006 | int cpu = smp_processor_id(); | |
5007 | runqueue_t *rq = this_rq(); | |
5008 | struct task_struct *p = rq->idle; | |
5009 | unsigned long flags; | |
5010 | ||
5011 | /* cpu has to be offline */ | |
5012 | BUG_ON(cpu_online(cpu)); | |
5013 | ||
5014 | /* Strictly not necessary since rest of the CPUs are stopped by now | |
5015 | * and interrupts disabled on current cpu. | |
5016 | */ | |
5017 | spin_lock_irqsave(&rq->lock, flags); | |
5018 | ||
5019 | __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1); | |
5020 | /* Add idle task to _front_ of it's priority queue */ | |
5021 | __activate_idle_task(p, rq); | |
5022 | ||
5023 | spin_unlock_irqrestore(&rq->lock, flags); | |
5024 | } | |
5025 | ||
5026 | /* Ensures that the idle task is using init_mm right before its cpu goes | |
5027 | * offline. | |
5028 | */ | |
5029 | void idle_task_exit(void) | |
5030 | { | |
5031 | struct mm_struct *mm = current->active_mm; | |
5032 | ||
5033 | BUG_ON(cpu_online(smp_processor_id())); | |
5034 | ||
5035 | if (mm != &init_mm) | |
5036 | switch_mm(mm, &init_mm, current); | |
5037 | mmdrop(mm); | |
5038 | } | |
5039 | ||
5040 | static void migrate_dead(unsigned int dead_cpu, task_t *tsk) | |
5041 | { | |
5042 | struct runqueue *rq = cpu_rq(dead_cpu); | |
5043 | ||
5044 | /* Must be exiting, otherwise would be on tasklist. */ | |
5045 | BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD); | |
5046 | ||
5047 | /* Cannot have done final schedule yet: would have vanished. */ | |
5048 | BUG_ON(tsk->flags & PF_DEAD); | |
5049 | ||
5050 | get_task_struct(tsk); | |
5051 | ||
5052 | /* | |
5053 | * Drop lock around migration; if someone else moves it, | |
5054 | * that's OK. No task can be added to this CPU, so iteration is | |
5055 | * fine. | |
5056 | */ | |
5057 | spin_unlock_irq(&rq->lock); | |
5058 | move_task_off_dead_cpu(dead_cpu, tsk); | |
5059 | spin_lock_irq(&rq->lock); | |
5060 | ||
5061 | put_task_struct(tsk); | |
5062 | } | |
5063 | ||
5064 | /* release_task() removes task from tasklist, so we won't find dead tasks. */ | |
5065 | static void migrate_dead_tasks(unsigned int dead_cpu) | |
5066 | { | |
5067 | unsigned arr, i; | |
5068 | struct runqueue *rq = cpu_rq(dead_cpu); | |
5069 | ||
5070 | for (arr = 0; arr < 2; arr++) { | |
5071 | for (i = 0; i < MAX_PRIO; i++) { | |
5072 | struct list_head *list = &rq->arrays[arr].queue[i]; | |
5073 | while (!list_empty(list)) | |
5074 | migrate_dead(dead_cpu, | |
5075 | list_entry(list->next, task_t, | |
5076 | run_list)); | |
5077 | } | |
5078 | } | |
5079 | } | |
5080 | #endif /* CONFIG_HOTPLUG_CPU */ | |
5081 | ||
5082 | /* | |
5083 | * migration_call - callback that gets triggered when a CPU is added. | |
5084 | * Here we can start up the necessary migration thread for the new CPU. | |
5085 | */ | |
26c2143b CS |
5086 | static int __cpuinit migration_call(struct notifier_block *nfb, |
5087 | unsigned long action, | |
5088 | void *hcpu) | |
1da177e4 LT |
5089 | { |
5090 | int cpu = (long)hcpu; | |
5091 | struct task_struct *p; | |
5092 | struct runqueue *rq; | |
5093 | unsigned long flags; | |
5094 | ||
5095 | switch (action) { | |
5096 | case CPU_UP_PREPARE: | |
5097 | p = kthread_create(migration_thread, hcpu, "migration/%d",cpu); | |
5098 | if (IS_ERR(p)) | |
5099 | return NOTIFY_BAD; | |
5100 | p->flags |= PF_NOFREEZE; | |
5101 | kthread_bind(p, cpu); | |
5102 | /* Must be high prio: stop_machine expects to yield to it. */ | |
5103 | rq = task_rq_lock(p, &flags); | |
5104 | __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1); | |
5105 | task_rq_unlock(rq, &flags); | |
5106 | cpu_rq(cpu)->migration_thread = p; | |
5107 | break; | |
5108 | case CPU_ONLINE: | |
5109 | /* Strictly unneccessary, as first user will wake it. */ | |
5110 | wake_up_process(cpu_rq(cpu)->migration_thread); | |
5111 | break; | |
5112 | #ifdef CONFIG_HOTPLUG_CPU | |
5113 | case CPU_UP_CANCELED: | |
fc75cdfa HC |
5114 | if (!cpu_rq(cpu)->migration_thread) |
5115 | break; | |
1da177e4 | 5116 | /* Unbind it from offline cpu so it can run. Fall thru. */ |
a4c4af7c HC |
5117 | kthread_bind(cpu_rq(cpu)->migration_thread, |
5118 | any_online_cpu(cpu_online_map)); | |
1da177e4 LT |
5119 | kthread_stop(cpu_rq(cpu)->migration_thread); |
5120 | cpu_rq(cpu)->migration_thread = NULL; | |
5121 | break; | |
5122 | case CPU_DEAD: | |
5123 | migrate_live_tasks(cpu); | |
5124 | rq = cpu_rq(cpu); | |
5125 | kthread_stop(rq->migration_thread); | |
5126 | rq->migration_thread = NULL; | |
5127 | /* Idle task back to normal (off runqueue, low prio) */ | |
5128 | rq = task_rq_lock(rq->idle, &flags); | |
5129 | deactivate_task(rq->idle, rq); | |
5130 | rq->idle->static_prio = MAX_PRIO; | |
5131 | __setscheduler(rq->idle, SCHED_NORMAL, 0); | |
5132 | migrate_dead_tasks(cpu); | |
5133 | task_rq_unlock(rq, &flags); | |
5134 | migrate_nr_uninterruptible(rq); | |
5135 | BUG_ON(rq->nr_running != 0); | |
5136 | ||
5137 | /* No need to migrate the tasks: it was best-effort if | |
5138 | * they didn't do lock_cpu_hotplug(). Just wake up | |
5139 | * the requestors. */ | |
5140 | spin_lock_irq(&rq->lock); | |
5141 | while (!list_empty(&rq->migration_queue)) { | |
5142 | migration_req_t *req; | |
5143 | req = list_entry(rq->migration_queue.next, | |
5144 | migration_req_t, list); | |
1da177e4 LT |
5145 | list_del_init(&req->list); |
5146 | complete(&req->done); | |
5147 | } | |
5148 | spin_unlock_irq(&rq->lock); | |
5149 | break; | |
5150 | #endif | |
5151 | } | |
5152 | return NOTIFY_OK; | |
5153 | } | |
5154 | ||
5155 | /* Register at highest priority so that task migration (migrate_all_tasks) | |
5156 | * happens before everything else. | |
5157 | */ | |
26c2143b | 5158 | static struct notifier_block __cpuinitdata migration_notifier = { |
1da177e4 LT |
5159 | .notifier_call = migration_call, |
5160 | .priority = 10 | |
5161 | }; | |
5162 | ||
5163 | int __init migration_init(void) | |
5164 | { | |
5165 | void *cpu = (void *)(long)smp_processor_id(); | |
5166 | /* Start one for boot CPU. */ | |
5167 | migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); | |
5168 | migration_call(&migration_notifier, CPU_ONLINE, cpu); | |
5169 | register_cpu_notifier(&migration_notifier); | |
5170 | return 0; | |
5171 | } | |
5172 | #endif | |
5173 | ||
5174 | #ifdef CONFIG_SMP | |
1a20ff27 | 5175 | #undef SCHED_DOMAIN_DEBUG |
1da177e4 LT |
5176 | #ifdef SCHED_DOMAIN_DEBUG |
5177 | static void sched_domain_debug(struct sched_domain *sd, int cpu) | |
5178 | { | |
5179 | int level = 0; | |
5180 | ||
41c7ce9a NP |
5181 | if (!sd) { |
5182 | printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); | |
5183 | return; | |
5184 | } | |
5185 | ||
1da177e4 LT |
5186 | printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); |
5187 | ||
5188 | do { | |
5189 | int i; | |
5190 | char str[NR_CPUS]; | |
5191 | struct sched_group *group = sd->groups; | |
5192 | cpumask_t groupmask; | |
5193 | ||
5194 | cpumask_scnprintf(str, NR_CPUS, sd->span); | |
5195 | cpus_clear(groupmask); | |
5196 | ||
5197 | printk(KERN_DEBUG); | |
5198 | for (i = 0; i < level + 1; i++) | |
5199 | printk(" "); | |
5200 | printk("domain %d: ", level); | |
5201 | ||
5202 | if (!(sd->flags & SD_LOAD_BALANCE)) { | |
5203 | printk("does not load-balance\n"); | |
5204 | if (sd->parent) | |
5205 | printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent"); | |
5206 | break; | |
5207 | } | |
5208 | ||
5209 | printk("span %s\n", str); | |
5210 | ||
5211 | if (!cpu_isset(cpu, sd->span)) | |
5212 | printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu); | |
5213 | if (!cpu_isset(cpu, group->cpumask)) | |
5214 | printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu); | |
5215 | ||
5216 | printk(KERN_DEBUG); | |
5217 | for (i = 0; i < level + 2; i++) | |
5218 | printk(" "); | |
5219 | printk("groups:"); | |
5220 | do { | |
5221 | if (!group) { | |
5222 | printk("\n"); | |
5223 | printk(KERN_ERR "ERROR: group is NULL\n"); | |
5224 | break; | |
5225 | } | |
5226 | ||
5227 | if (!group->cpu_power) { | |
5228 | printk("\n"); | |
5229 | printk(KERN_ERR "ERROR: domain->cpu_power not set\n"); | |
5230 | } | |
5231 | ||
5232 | if (!cpus_weight(group->cpumask)) { | |
5233 | printk("\n"); | |
5234 | printk(KERN_ERR "ERROR: empty group\n"); | |
5235 | } | |
5236 | ||
5237 | if (cpus_intersects(groupmask, group->cpumask)) { | |
5238 | printk("\n"); | |
5239 | printk(KERN_ERR "ERROR: repeated CPUs\n"); | |
5240 | } | |
5241 | ||
5242 | cpus_or(groupmask, groupmask, group->cpumask); | |
5243 | ||
5244 | cpumask_scnprintf(str, NR_CPUS, group->cpumask); | |
5245 | printk(" %s", str); | |
5246 | ||
5247 | group = group->next; | |
5248 | } while (group != sd->groups); | |
5249 | printk("\n"); | |
5250 | ||
5251 | if (!cpus_equal(sd->span, groupmask)) | |
5252 | printk(KERN_ERR "ERROR: groups don't span domain->span\n"); | |
5253 | ||
5254 | level++; | |
5255 | sd = sd->parent; | |
5256 | ||
5257 | if (sd) { | |
5258 | if (!cpus_subset(groupmask, sd->span)) | |
5259 | printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n"); | |
5260 | } | |
5261 | ||
5262 | } while (sd); | |
5263 | } | |
5264 | #else | |
5265 | #define sched_domain_debug(sd, cpu) {} | |
5266 | #endif | |
5267 | ||
1a20ff27 | 5268 | static int sd_degenerate(struct sched_domain *sd) |
245af2c7 SS |
5269 | { |
5270 | if (cpus_weight(sd->span) == 1) | |
5271 | return 1; | |
5272 | ||
5273 | /* Following flags need at least 2 groups */ | |
5274 | if (sd->flags & (SD_LOAD_BALANCE | | |
5275 | SD_BALANCE_NEWIDLE | | |
5276 | SD_BALANCE_FORK | | |
5277 | SD_BALANCE_EXEC)) { | |
5278 | if (sd->groups != sd->groups->next) | |
5279 | return 0; | |
5280 | } | |
5281 | ||
5282 | /* Following flags don't use groups */ | |
5283 | if (sd->flags & (SD_WAKE_IDLE | | |
5284 | SD_WAKE_AFFINE | | |
5285 | SD_WAKE_BALANCE)) | |
5286 | return 0; | |
5287 | ||
5288 | return 1; | |
5289 | } | |
5290 | ||
1a20ff27 | 5291 | static int sd_parent_degenerate(struct sched_domain *sd, |
245af2c7 SS |
5292 | struct sched_domain *parent) |
5293 | { | |
5294 | unsigned long cflags = sd->flags, pflags = parent->flags; | |
5295 | ||
5296 | if (sd_degenerate(parent)) | |
5297 | return 1; | |
5298 | ||
5299 | if (!cpus_equal(sd->span, parent->span)) | |
5300 | return 0; | |
5301 | ||
5302 | /* Does parent contain flags not in child? */ | |
5303 | /* WAKE_BALANCE is a subset of WAKE_AFFINE */ | |
5304 | if (cflags & SD_WAKE_AFFINE) | |
5305 | pflags &= ~SD_WAKE_BALANCE; | |
5306 | /* Flags needing groups don't count if only 1 group in parent */ | |
5307 | if (parent->groups == parent->groups->next) { | |
5308 | pflags &= ~(SD_LOAD_BALANCE | | |
5309 | SD_BALANCE_NEWIDLE | | |
5310 | SD_BALANCE_FORK | | |
5311 | SD_BALANCE_EXEC); | |
5312 | } | |
5313 | if (~cflags & pflags) | |
5314 | return 0; | |
5315 | ||
5316 | return 1; | |
5317 | } | |
5318 | ||
1da177e4 LT |
5319 | /* |
5320 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must | |
5321 | * hold the hotplug lock. | |
5322 | */ | |
9c1cfda2 | 5323 | static void cpu_attach_domain(struct sched_domain *sd, int cpu) |
1da177e4 | 5324 | { |
1da177e4 | 5325 | runqueue_t *rq = cpu_rq(cpu); |
245af2c7 SS |
5326 | struct sched_domain *tmp; |
5327 | ||
5328 | /* Remove the sched domains which do not contribute to scheduling. */ | |
5329 | for (tmp = sd; tmp; tmp = tmp->parent) { | |
5330 | struct sched_domain *parent = tmp->parent; | |
5331 | if (!parent) | |
5332 | break; | |
5333 | if (sd_parent_degenerate(tmp, parent)) | |
5334 | tmp->parent = parent->parent; | |
5335 | } | |
5336 | ||
5337 | if (sd && sd_degenerate(sd)) | |
5338 | sd = sd->parent; | |
1da177e4 LT |
5339 | |
5340 | sched_domain_debug(sd, cpu); | |
5341 | ||
674311d5 | 5342 | rcu_assign_pointer(rq->sd, sd); |
1da177e4 LT |
5343 | } |
5344 | ||
5345 | /* cpus with isolated domains */ | |
9c1cfda2 | 5346 | static cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE; |
1da177e4 LT |
5347 | |
5348 | /* Setup the mask of cpus configured for isolated domains */ | |
5349 | static int __init isolated_cpu_setup(char *str) | |
5350 | { | |
5351 | int ints[NR_CPUS], i; | |
5352 | ||
5353 | str = get_options(str, ARRAY_SIZE(ints), ints); | |
5354 | cpus_clear(cpu_isolated_map); | |
5355 | for (i = 1; i <= ints[0]; i++) | |
5356 | if (ints[i] < NR_CPUS) | |
5357 | cpu_set(ints[i], cpu_isolated_map); | |
5358 | return 1; | |
5359 | } | |
5360 | ||
5361 | __setup ("isolcpus=", isolated_cpu_setup); | |
5362 | ||
5363 | /* | |
5364 | * init_sched_build_groups takes an array of groups, the cpumask we wish | |
5365 | * to span, and a pointer to a function which identifies what group a CPU | |
5366 | * belongs to. The return value of group_fn must be a valid index into the | |
5367 | * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we | |
5368 | * keep track of groups covered with a cpumask_t). | |
5369 | * | |
5370 | * init_sched_build_groups will build a circular linked list of the groups | |
5371 | * covered by the given span, and will set each group's ->cpumask correctly, | |
5372 | * and ->cpu_power to 0. | |
5373 | */ | |
9c1cfda2 JH |
5374 | static void init_sched_build_groups(struct sched_group groups[], cpumask_t span, |
5375 | int (*group_fn)(int cpu)) | |
1da177e4 LT |
5376 | { |
5377 | struct sched_group *first = NULL, *last = NULL; | |
5378 | cpumask_t covered = CPU_MASK_NONE; | |
5379 | int i; | |
5380 | ||
5381 | for_each_cpu_mask(i, span) { | |
5382 | int group = group_fn(i); | |
5383 | struct sched_group *sg = &groups[group]; | |
5384 | int j; | |
5385 | ||
5386 | if (cpu_isset(i, covered)) | |
5387 | continue; | |
5388 | ||
5389 | sg->cpumask = CPU_MASK_NONE; | |
5390 | sg->cpu_power = 0; | |
5391 | ||
5392 | for_each_cpu_mask(j, span) { | |
5393 | if (group_fn(j) != group) | |
5394 | continue; | |
5395 | ||
5396 | cpu_set(j, covered); | |
5397 | cpu_set(j, sg->cpumask); | |
5398 | } | |
5399 | if (!first) | |
5400 | first = sg; | |
5401 | if (last) | |
5402 | last->next = sg; | |
5403 | last = sg; | |
5404 | } | |
5405 | last->next = first; | |
5406 | } | |
5407 | ||
9c1cfda2 | 5408 | #define SD_NODES_PER_DOMAIN 16 |
1da177e4 | 5409 | |
198e2f18 | 5410 | /* |
5411 | * Self-tuning task migration cost measurement between source and target CPUs. | |
5412 | * | |
5413 | * This is done by measuring the cost of manipulating buffers of varying | |
5414 | * sizes. For a given buffer-size here are the steps that are taken: | |
5415 | * | |
5416 | * 1) the source CPU reads+dirties a shared buffer | |
5417 | * 2) the target CPU reads+dirties the same shared buffer | |
5418 | * | |
5419 | * We measure how long they take, in the following 4 scenarios: | |
5420 | * | |
5421 | * - source: CPU1, target: CPU2 | cost1 | |
5422 | * - source: CPU2, target: CPU1 | cost2 | |
5423 | * - source: CPU1, target: CPU1 | cost3 | |
5424 | * - source: CPU2, target: CPU2 | cost4 | |
5425 | * | |
5426 | * We then calculate the cost3+cost4-cost1-cost2 difference - this is | |
5427 | * the cost of migration. | |
5428 | * | |
5429 | * We then start off from a small buffer-size and iterate up to larger | |
5430 | * buffer sizes, in 5% steps - measuring each buffer-size separately, and | |
5431 | * doing a maximum search for the cost. (The maximum cost for a migration | |
5432 | * normally occurs when the working set size is around the effective cache | |
5433 | * size.) | |
5434 | */ | |
5435 | #define SEARCH_SCOPE 2 | |
5436 | #define MIN_CACHE_SIZE (64*1024U) | |
5437 | #define DEFAULT_CACHE_SIZE (5*1024*1024U) | |
70b4d63e | 5438 | #define ITERATIONS 1 |
198e2f18 | 5439 | #define SIZE_THRESH 130 |
5440 | #define COST_THRESH 130 | |
5441 | ||
5442 | /* | |
5443 | * The migration cost is a function of 'domain distance'. Domain | |
5444 | * distance is the number of steps a CPU has to iterate down its | |
5445 | * domain tree to share a domain with the other CPU. The farther | |
5446 | * two CPUs are from each other, the larger the distance gets. | |
5447 | * | |
5448 | * Note that we use the distance only to cache measurement results, | |
5449 | * the distance value is not used numerically otherwise. When two | |
5450 | * CPUs have the same distance it is assumed that the migration | |
5451 | * cost is the same. (this is a simplification but quite practical) | |
5452 | */ | |
5453 | #define MAX_DOMAIN_DISTANCE 32 | |
5454 | ||
5455 | static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] = | |
4bbf39c2 IM |
5456 | { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] = |
5457 | /* | |
5458 | * Architectures may override the migration cost and thus avoid | |
5459 | * boot-time calibration. Unit is nanoseconds. Mostly useful for | |
5460 | * virtualized hardware: | |
5461 | */ | |
5462 | #ifdef CONFIG_DEFAULT_MIGRATION_COST | |
5463 | CONFIG_DEFAULT_MIGRATION_COST | |
5464 | #else | |
5465 | -1LL | |
5466 | #endif | |
5467 | }; | |
198e2f18 | 5468 | |
5469 | /* | |
5470 | * Allow override of migration cost - in units of microseconds. | |
5471 | * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost | |
5472 | * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs: | |
5473 | */ | |
5474 | static int __init migration_cost_setup(char *str) | |
5475 | { | |
5476 | int ints[MAX_DOMAIN_DISTANCE+1], i; | |
5477 | ||
5478 | str = get_options(str, ARRAY_SIZE(ints), ints); | |
5479 | ||
5480 | printk("#ints: %d\n", ints[0]); | |
5481 | for (i = 1; i <= ints[0]; i++) { | |
5482 | migration_cost[i-1] = (unsigned long long)ints[i]*1000; | |
5483 | printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]); | |
5484 | } | |
5485 | return 1; | |
5486 | } | |
5487 | ||
5488 | __setup ("migration_cost=", migration_cost_setup); | |
5489 | ||
5490 | /* | |
5491 | * Global multiplier (divisor) for migration-cutoff values, | |
5492 | * in percentiles. E.g. use a value of 150 to get 1.5 times | |
5493 | * longer cache-hot cutoff times. | |
5494 | * | |
5495 | * (We scale it from 100 to 128 to long long handling easier.) | |
5496 | */ | |
5497 | ||
5498 | #define MIGRATION_FACTOR_SCALE 128 | |
5499 | ||
5500 | static unsigned int migration_factor = MIGRATION_FACTOR_SCALE; | |
5501 | ||
5502 | static int __init setup_migration_factor(char *str) | |
5503 | { | |
5504 | get_option(&str, &migration_factor); | |
5505 | migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100; | |
5506 | return 1; | |
5507 | } | |
5508 | ||
5509 | __setup("migration_factor=", setup_migration_factor); | |
5510 | ||
5511 | /* | |
5512 | * Estimated distance of two CPUs, measured via the number of domains | |
5513 | * we have to pass for the two CPUs to be in the same span: | |
5514 | */ | |
5515 | static unsigned long domain_distance(int cpu1, int cpu2) | |
5516 | { | |
5517 | unsigned long distance = 0; | |
5518 | struct sched_domain *sd; | |
5519 | ||
5520 | for_each_domain(cpu1, sd) { | |
5521 | WARN_ON(!cpu_isset(cpu1, sd->span)); | |
5522 | if (cpu_isset(cpu2, sd->span)) | |
5523 | return distance; | |
5524 | distance++; | |
5525 | } | |
5526 | if (distance >= MAX_DOMAIN_DISTANCE) { | |
5527 | WARN_ON(1); | |
5528 | distance = MAX_DOMAIN_DISTANCE-1; | |
5529 | } | |
5530 | ||
5531 | return distance; | |
5532 | } | |
5533 | ||
5534 | static unsigned int migration_debug; | |
5535 | ||
5536 | static int __init setup_migration_debug(char *str) | |
5537 | { | |
5538 | get_option(&str, &migration_debug); | |
5539 | return 1; | |
5540 | } | |
5541 | ||
5542 | __setup("migration_debug=", setup_migration_debug); | |
5543 | ||
5544 | /* | |
5545 | * Maximum cache-size that the scheduler should try to measure. | |
5546 | * Architectures with larger caches should tune this up during | |
5547 | * bootup. Gets used in the domain-setup code (i.e. during SMP | |
5548 | * bootup). | |
5549 | */ | |
5550 | unsigned int max_cache_size; | |
5551 | ||
5552 | static int __init setup_max_cache_size(char *str) | |
5553 | { | |
5554 | get_option(&str, &max_cache_size); | |
5555 | return 1; | |
5556 | } | |
5557 | ||
5558 | __setup("max_cache_size=", setup_max_cache_size); | |
5559 | ||
5560 | /* | |
5561 | * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This | |
5562 | * is the operation that is timed, so we try to generate unpredictable | |
5563 | * cachemisses that still end up filling the L2 cache: | |
5564 | */ | |
5565 | static void touch_cache(void *__cache, unsigned long __size) | |
5566 | { | |
5567 | unsigned long size = __size/sizeof(long), chunk1 = size/3, | |
5568 | chunk2 = 2*size/3; | |
5569 | unsigned long *cache = __cache; | |
5570 | int i; | |
5571 | ||
5572 | for (i = 0; i < size/6; i += 8) { | |
5573 | switch (i % 6) { | |
5574 | case 0: cache[i]++; | |
5575 | case 1: cache[size-1-i]++; | |
5576 | case 2: cache[chunk1-i]++; | |
5577 | case 3: cache[chunk1+i]++; | |
5578 | case 4: cache[chunk2-i]++; | |
5579 | case 5: cache[chunk2+i]++; | |
5580 | } | |
5581 | } | |
5582 | } | |
5583 | ||
5584 | /* | |
5585 | * Measure the cache-cost of one task migration. Returns in units of nsec. | |
5586 | */ | |
5587 | static unsigned long long measure_one(void *cache, unsigned long size, | |
5588 | int source, int target) | |
5589 | { | |
5590 | cpumask_t mask, saved_mask; | |
5591 | unsigned long long t0, t1, t2, t3, cost; | |
5592 | ||
5593 | saved_mask = current->cpus_allowed; | |
5594 | ||
5595 | /* | |
5596 | * Flush source caches to RAM and invalidate them: | |
5597 | */ | |
5598 | sched_cacheflush(); | |
5599 | ||
5600 | /* | |
5601 | * Migrate to the source CPU: | |
5602 | */ | |
5603 | mask = cpumask_of_cpu(source); | |
5604 | set_cpus_allowed(current, mask); | |
5605 | WARN_ON(smp_processor_id() != source); | |
5606 | ||
5607 | /* | |
5608 | * Dirty the working set: | |
5609 | */ | |
5610 | t0 = sched_clock(); | |
5611 | touch_cache(cache, size); | |
5612 | t1 = sched_clock(); | |
5613 | ||
5614 | /* | |
5615 | * Migrate to the target CPU, dirty the L2 cache and access | |
5616 | * the shared buffer. (which represents the working set | |
5617 | * of a migrated task.) | |
5618 | */ | |
5619 | mask = cpumask_of_cpu(target); | |
5620 | set_cpus_allowed(current, mask); | |
5621 | WARN_ON(smp_processor_id() != target); | |
5622 | ||
5623 | t2 = sched_clock(); | |
5624 | touch_cache(cache, size); | |
5625 | t3 = sched_clock(); | |
5626 | ||
5627 | cost = t1-t0 + t3-t2; | |
5628 | ||
5629 | if (migration_debug >= 2) | |
5630 | printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n", | |
5631 | source, target, t1-t0, t1-t0, t3-t2, cost); | |
5632 | /* | |
5633 | * Flush target caches to RAM and invalidate them: | |
5634 | */ | |
5635 | sched_cacheflush(); | |
5636 | ||
5637 | set_cpus_allowed(current, saved_mask); | |
5638 | ||
5639 | return cost; | |
5640 | } | |
5641 | ||
5642 | /* | |
5643 | * Measure a series of task migrations and return the average | |
5644 | * result. Since this code runs early during bootup the system | |
5645 | * is 'undisturbed' and the average latency makes sense. | |
5646 | * | |
5647 | * The algorithm in essence auto-detects the relevant cache-size, | |
5648 | * so it will properly detect different cachesizes for different | |
5649 | * cache-hierarchies, depending on how the CPUs are connected. | |
5650 | * | |
5651 | * Architectures can prime the upper limit of the search range via | |
5652 | * max_cache_size, otherwise the search range defaults to 20MB...64K. | |
5653 | */ | |
5654 | static unsigned long long | |
5655 | measure_cost(int cpu1, int cpu2, void *cache, unsigned int size) | |
5656 | { | |
5657 | unsigned long long cost1, cost2; | |
5658 | int i; | |
5659 | ||
5660 | /* | |
5661 | * Measure the migration cost of 'size' bytes, over an | |
5662 | * average of 10 runs: | |
5663 | * | |
5664 | * (We perturb the cache size by a small (0..4k) | |
5665 | * value to compensate size/alignment related artifacts. | |
5666 | * We also subtract the cost of the operation done on | |
5667 | * the same CPU.) | |
5668 | */ | |
5669 | cost1 = 0; | |
5670 | ||
5671 | /* | |
5672 | * dry run, to make sure we start off cache-cold on cpu1, | |
5673 | * and to get any vmalloc pagefaults in advance: | |
5674 | */ | |
5675 | measure_one(cache, size, cpu1, cpu2); | |
5676 | for (i = 0; i < ITERATIONS; i++) | |
5677 | cost1 += measure_one(cache, size - i*1024, cpu1, cpu2); | |
5678 | ||
5679 | measure_one(cache, size, cpu2, cpu1); | |
5680 | for (i = 0; i < ITERATIONS; i++) | |
5681 | cost1 += measure_one(cache, size - i*1024, cpu2, cpu1); | |
5682 | ||
5683 | /* | |
5684 | * (We measure the non-migrating [cached] cost on both | |
5685 | * cpu1 and cpu2, to handle CPUs with different speeds) | |
5686 | */ | |
5687 | cost2 = 0; | |
5688 | ||
5689 | measure_one(cache, size, cpu1, cpu1); | |
5690 | for (i = 0; i < ITERATIONS; i++) | |
5691 | cost2 += measure_one(cache, size - i*1024, cpu1, cpu1); | |
5692 | ||
5693 | measure_one(cache, size, cpu2, cpu2); | |
5694 | for (i = 0; i < ITERATIONS; i++) | |
5695 | cost2 += measure_one(cache, size - i*1024, cpu2, cpu2); | |
5696 | ||
5697 | /* | |
5698 | * Get the per-iteration migration cost: | |
5699 | */ | |
5700 | do_div(cost1, 2*ITERATIONS); | |
5701 | do_div(cost2, 2*ITERATIONS); | |
5702 | ||
5703 | return cost1 - cost2; | |
5704 | } | |
5705 | ||
5706 | static unsigned long long measure_migration_cost(int cpu1, int cpu2) | |
5707 | { | |
5708 | unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0; | |
5709 | unsigned int max_size, size, size_found = 0; | |
5710 | long long cost = 0, prev_cost; | |
5711 | void *cache; | |
5712 | ||
5713 | /* | |
5714 | * Search from max_cache_size*5 down to 64K - the real relevant | |
5715 | * cachesize has to lie somewhere inbetween. | |
5716 | */ | |
5717 | if (max_cache_size) { | |
5718 | max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE); | |
5719 | size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE); | |
5720 | } else { | |
5721 | /* | |
5722 | * Since we have no estimation about the relevant | |
5723 | * search range | |
5724 | */ | |
5725 | max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE; | |
5726 | size = MIN_CACHE_SIZE; | |
5727 | } | |
5728 | ||
5729 | if (!cpu_online(cpu1) || !cpu_online(cpu2)) { | |
5730 | printk("cpu %d and %d not both online!\n", cpu1, cpu2); | |
5731 | return 0; | |
5732 | } | |
5733 | ||
5734 | /* | |
5735 | * Allocate the working set: | |
5736 | */ | |
5737 | cache = vmalloc(max_size); | |
5738 | if (!cache) { | |
5739 | printk("could not vmalloc %d bytes for cache!\n", 2*max_size); | |
5740 | return 1000000; // return 1 msec on very small boxen | |
5741 | } | |
5742 | ||
5743 | while (size <= max_size) { | |
5744 | prev_cost = cost; | |
5745 | cost = measure_cost(cpu1, cpu2, cache, size); | |
5746 | ||
5747 | /* | |
5748 | * Update the max: | |
5749 | */ | |
5750 | if (cost > 0) { | |
5751 | if (max_cost < cost) { | |
5752 | max_cost = cost; | |
5753 | size_found = size; | |
5754 | } | |
5755 | } | |
5756 | /* | |
5757 | * Calculate average fluctuation, we use this to prevent | |
5758 | * noise from triggering an early break out of the loop: | |
5759 | */ | |
5760 | fluct = abs(cost - prev_cost); | |
5761 | avg_fluct = (avg_fluct + fluct)/2; | |
5762 | ||
5763 | if (migration_debug) | |
5764 | printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n", | |
5765 | cpu1, cpu2, size, | |
5766 | (long)cost / 1000000, | |
5767 | ((long)cost / 100000) % 10, | |
5768 | (long)max_cost / 1000000, | |
5769 | ((long)max_cost / 100000) % 10, | |
5770 | domain_distance(cpu1, cpu2), | |
5771 | cost, avg_fluct); | |
5772 | ||
5773 | /* | |
5774 | * If we iterated at least 20% past the previous maximum, | |
5775 | * and the cost has dropped by more than 20% already, | |
5776 | * (taking fluctuations into account) then we assume to | |
5777 | * have found the maximum and break out of the loop early: | |
5778 | */ | |
5779 | if (size_found && (size*100 > size_found*SIZE_THRESH)) | |
5780 | if (cost+avg_fluct <= 0 || | |
5781 | max_cost*100 > (cost+avg_fluct)*COST_THRESH) { | |
5782 | ||
5783 | if (migration_debug) | |
5784 | printk("-> found max.\n"); | |
5785 | break; | |
5786 | } | |
5787 | /* | |
70b4d63e | 5788 | * Increase the cachesize in 10% steps: |
198e2f18 | 5789 | */ |
70b4d63e | 5790 | size = size * 10 / 9; |
198e2f18 | 5791 | } |
5792 | ||
5793 | if (migration_debug) | |
5794 | printk("[%d][%d] working set size found: %d, cost: %Ld\n", | |
5795 | cpu1, cpu2, size_found, max_cost); | |
5796 | ||
5797 | vfree(cache); | |
5798 | ||
5799 | /* | |
5800 | * A task is considered 'cache cold' if at least 2 times | |
5801 | * the worst-case cost of migration has passed. | |
5802 | * | |
5803 | * (this limit is only listened to if the load-balancing | |
5804 | * situation is 'nice' - if there is a large imbalance we | |
5805 | * ignore it for the sake of CPU utilization and | |
5806 | * processing fairness.) | |
5807 | */ | |
5808 | return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE; | |
5809 | } | |
5810 | ||
5811 | static void calibrate_migration_costs(const cpumask_t *cpu_map) | |
5812 | { | |
5813 | int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id(); | |
5814 | unsigned long j0, j1, distance, max_distance = 0; | |
5815 | struct sched_domain *sd; | |
5816 | ||
5817 | j0 = jiffies; | |
5818 | ||
5819 | /* | |
5820 | * First pass - calculate the cacheflush times: | |
5821 | */ | |
5822 | for_each_cpu_mask(cpu1, *cpu_map) { | |
5823 | for_each_cpu_mask(cpu2, *cpu_map) { | |
5824 | if (cpu1 == cpu2) | |
5825 | continue; | |
5826 | distance = domain_distance(cpu1, cpu2); | |
5827 | max_distance = max(max_distance, distance); | |
5828 | /* | |
5829 | * No result cached yet? | |
5830 | */ | |
5831 | if (migration_cost[distance] == -1LL) | |
5832 | migration_cost[distance] = | |
5833 | measure_migration_cost(cpu1, cpu2); | |
5834 | } | |
5835 | } | |
5836 | /* | |
5837 | * Second pass - update the sched domain hierarchy with | |
5838 | * the new cache-hot-time estimations: | |
5839 | */ | |
5840 | for_each_cpu_mask(cpu, *cpu_map) { | |
5841 | distance = 0; | |
5842 | for_each_domain(cpu, sd) { | |
5843 | sd->cache_hot_time = migration_cost[distance]; | |
5844 | distance++; | |
5845 | } | |
5846 | } | |
5847 | /* | |
5848 | * Print the matrix: | |
5849 | */ | |
5850 | if (migration_debug) | |
5851 | printk("migration: max_cache_size: %d, cpu: %d MHz:\n", | |
5852 | max_cache_size, | |
5853 | #ifdef CONFIG_X86 | |
5854 | cpu_khz/1000 | |
5855 | #else | |
5856 | -1 | |
5857 | #endif | |
5858 | ); | |
bd576c95 CE |
5859 | if (system_state == SYSTEM_BOOTING) { |
5860 | printk("migration_cost="); | |
5861 | for (distance = 0; distance <= max_distance; distance++) { | |
5862 | if (distance) | |
5863 | printk(","); | |
5864 | printk("%ld", (long)migration_cost[distance] / 1000); | |
5865 | } | |
5866 | printk("\n"); | |
198e2f18 | 5867 | } |
198e2f18 | 5868 | j1 = jiffies; |
5869 | if (migration_debug) | |
5870 | printk("migration: %ld seconds\n", (j1-j0)/HZ); | |
5871 | ||
5872 | /* | |
5873 | * Move back to the original CPU. NUMA-Q gets confused | |
5874 | * if we migrate to another quad during bootup. | |
5875 | */ | |
5876 | if (raw_smp_processor_id() != orig_cpu) { | |
5877 | cpumask_t mask = cpumask_of_cpu(orig_cpu), | |
5878 | saved_mask = current->cpus_allowed; | |
5879 | ||
5880 | set_cpus_allowed(current, mask); | |
5881 | set_cpus_allowed(current, saved_mask); | |
5882 | } | |
5883 | } | |
5884 | ||
9c1cfda2 | 5885 | #ifdef CONFIG_NUMA |
198e2f18 | 5886 | |
9c1cfda2 JH |
5887 | /** |
5888 | * find_next_best_node - find the next node to include in a sched_domain | |
5889 | * @node: node whose sched_domain we're building | |
5890 | * @used_nodes: nodes already in the sched_domain | |
5891 | * | |
5892 | * Find the next node to include in a given scheduling domain. Simply | |
5893 | * finds the closest node not already in the @used_nodes map. | |
5894 | * | |
5895 | * Should use nodemask_t. | |
5896 | */ | |
5897 | static int find_next_best_node(int node, unsigned long *used_nodes) | |
5898 | { | |
5899 | int i, n, val, min_val, best_node = 0; | |
5900 | ||
5901 | min_val = INT_MAX; | |
5902 | ||
5903 | for (i = 0; i < MAX_NUMNODES; i++) { | |
5904 | /* Start at @node */ | |
5905 | n = (node + i) % MAX_NUMNODES; | |
5906 | ||
5907 | if (!nr_cpus_node(n)) | |
5908 | continue; | |
5909 | ||
5910 | /* Skip already used nodes */ | |
5911 | if (test_bit(n, used_nodes)) | |
5912 | continue; | |
5913 | ||
5914 | /* Simple min distance search */ | |
5915 | val = node_distance(node, n); | |
5916 | ||
5917 | if (val < min_val) { | |
5918 | min_val = val; | |
5919 | best_node = n; | |
5920 | } | |
5921 | } | |
5922 | ||
5923 | set_bit(best_node, used_nodes); | |
5924 | return best_node; | |
5925 | } | |
5926 | ||
5927 | /** | |
5928 | * sched_domain_node_span - get a cpumask for a node's sched_domain | |
5929 | * @node: node whose cpumask we're constructing | |
5930 | * @size: number of nodes to include in this span | |
5931 | * | |
5932 | * Given a node, construct a good cpumask for its sched_domain to span. It | |
5933 | * should be one that prevents unnecessary balancing, but also spreads tasks | |
5934 | * out optimally. | |
5935 | */ | |
5936 | static cpumask_t sched_domain_node_span(int node) | |
5937 | { | |
5938 | int i; | |
5939 | cpumask_t span, nodemask; | |
5940 | DECLARE_BITMAP(used_nodes, MAX_NUMNODES); | |
5941 | ||
5942 | cpus_clear(span); | |
5943 | bitmap_zero(used_nodes, MAX_NUMNODES); | |
5944 | ||
5945 | nodemask = node_to_cpumask(node); | |
5946 | cpus_or(span, span, nodemask); | |
5947 | set_bit(node, used_nodes); | |
5948 | ||
5949 | for (i = 1; i < SD_NODES_PER_DOMAIN; i++) { | |
5950 | int next_node = find_next_best_node(node, used_nodes); | |
5951 | nodemask = node_to_cpumask(next_node); | |
5952 | cpus_or(span, span, nodemask); | |
5953 | } | |
5954 | ||
5955 | return span; | |
5956 | } | |
5957 | #endif | |
5958 | ||
5c45bf27 | 5959 | int sched_smt_power_savings = 0, sched_mc_power_savings = 0; |
9c1cfda2 JH |
5960 | /* |
5961 | * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we | |
5962 | * can switch it on easily if needed. | |
5963 | */ | |
1da177e4 LT |
5964 | #ifdef CONFIG_SCHED_SMT |
5965 | static DEFINE_PER_CPU(struct sched_domain, cpu_domains); | |
5966 | static struct sched_group sched_group_cpus[NR_CPUS]; | |
1a20ff27 | 5967 | static int cpu_to_cpu_group(int cpu) |
1da177e4 LT |
5968 | { |
5969 | return cpu; | |
5970 | } | |
5971 | #endif | |
5972 | ||
1e9f28fa SS |
5973 | #ifdef CONFIG_SCHED_MC |
5974 | static DEFINE_PER_CPU(struct sched_domain, core_domains); | |
36938169 | 5975 | static struct sched_group *sched_group_core_bycpu[NR_CPUS]; |
1e9f28fa SS |
5976 | #endif |
5977 | ||
5978 | #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT) | |
5979 | static int cpu_to_core_group(int cpu) | |
5980 | { | |
5981 | return first_cpu(cpu_sibling_map[cpu]); | |
5982 | } | |
5983 | #elif defined(CONFIG_SCHED_MC) | |
5984 | static int cpu_to_core_group(int cpu) | |
5985 | { | |
5986 | return cpu; | |
5987 | } | |
5988 | #endif | |
5989 | ||
1da177e4 | 5990 | static DEFINE_PER_CPU(struct sched_domain, phys_domains); |
36938169 | 5991 | static struct sched_group *sched_group_phys_bycpu[NR_CPUS]; |
1a20ff27 | 5992 | static int cpu_to_phys_group(int cpu) |
1da177e4 | 5993 | { |
1e9f28fa SS |
5994 | #if defined(CONFIG_SCHED_MC) |
5995 | cpumask_t mask = cpu_coregroup_map(cpu); | |
5996 | return first_cpu(mask); | |
5997 | #elif defined(CONFIG_SCHED_SMT) | |
1da177e4 LT |
5998 | return first_cpu(cpu_sibling_map[cpu]); |
5999 | #else | |
6000 | return cpu; | |
6001 | #endif | |
6002 | } | |
6003 | ||
6004 | #ifdef CONFIG_NUMA | |
1da177e4 | 6005 | /* |
9c1cfda2 JH |
6006 | * The init_sched_build_groups can't handle what we want to do with node |
6007 | * groups, so roll our own. Now each node has its own list of groups which | |
6008 | * gets dynamically allocated. | |
1da177e4 | 6009 | */ |
9c1cfda2 | 6010 | static DEFINE_PER_CPU(struct sched_domain, node_domains); |
d1b55138 | 6011 | static struct sched_group **sched_group_nodes_bycpu[NR_CPUS]; |
1da177e4 | 6012 | |
9c1cfda2 | 6013 | static DEFINE_PER_CPU(struct sched_domain, allnodes_domains); |
d1b55138 | 6014 | static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS]; |
9c1cfda2 JH |
6015 | |
6016 | static int cpu_to_allnodes_group(int cpu) | |
6017 | { | |
6018 | return cpu_to_node(cpu); | |
1da177e4 | 6019 | } |
08069033 SS |
6020 | static void init_numa_sched_groups_power(struct sched_group *group_head) |
6021 | { | |
6022 | struct sched_group *sg = group_head; | |
6023 | int j; | |
6024 | ||
6025 | if (!sg) | |
6026 | return; | |
6027 | next_sg: | |
6028 | for_each_cpu_mask(j, sg->cpumask) { | |
6029 | struct sched_domain *sd; | |
6030 | ||
6031 | sd = &per_cpu(phys_domains, j); | |
6032 | if (j != first_cpu(sd->groups->cpumask)) { | |
6033 | /* | |
6034 | * Only add "power" once for each | |
6035 | * physical package. | |
6036 | */ | |
6037 | continue; | |
6038 | } | |
6039 | ||
6040 | sg->cpu_power += sd->groups->cpu_power; | |
6041 | } | |
6042 | sg = sg->next; | |
6043 | if (sg != group_head) | |
6044 | goto next_sg; | |
6045 | } | |
1da177e4 LT |
6046 | #endif |
6047 | ||
51888ca2 SV |
6048 | /* Free memory allocated for various sched_group structures */ |
6049 | static void free_sched_groups(const cpumask_t *cpu_map) | |
6050 | { | |
36938169 | 6051 | int cpu; |
51888ca2 SV |
6052 | #ifdef CONFIG_NUMA |
6053 | int i; | |
51888ca2 SV |
6054 | |
6055 | for_each_cpu_mask(cpu, *cpu_map) { | |
6056 | struct sched_group *sched_group_allnodes | |
6057 | = sched_group_allnodes_bycpu[cpu]; | |
6058 | struct sched_group **sched_group_nodes | |
6059 | = sched_group_nodes_bycpu[cpu]; | |
6060 | ||
6061 | if (sched_group_allnodes) { | |
6062 | kfree(sched_group_allnodes); | |
6063 | sched_group_allnodes_bycpu[cpu] = NULL; | |
6064 | } | |
6065 | ||
6066 | if (!sched_group_nodes) | |
6067 | continue; | |
6068 | ||
6069 | for (i = 0; i < MAX_NUMNODES; i++) { | |
6070 | cpumask_t nodemask = node_to_cpumask(i); | |
6071 | struct sched_group *oldsg, *sg = sched_group_nodes[i]; | |
6072 | ||
6073 | cpus_and(nodemask, nodemask, *cpu_map); | |
6074 | if (cpus_empty(nodemask)) | |
6075 | continue; | |
6076 | ||
6077 | if (sg == NULL) | |
6078 | continue; | |
6079 | sg = sg->next; | |
6080 | next_sg: | |
6081 | oldsg = sg; | |
6082 | sg = sg->next; | |
6083 | kfree(oldsg); | |
6084 | if (oldsg != sched_group_nodes[i]) | |
6085 | goto next_sg; | |
6086 | } | |
6087 | kfree(sched_group_nodes); | |
6088 | sched_group_nodes_bycpu[cpu] = NULL; | |
6089 | } | |
6090 | #endif | |
36938169 SV |
6091 | for_each_cpu_mask(cpu, *cpu_map) { |
6092 | if (sched_group_phys_bycpu[cpu]) { | |
6093 | kfree(sched_group_phys_bycpu[cpu]); | |
6094 | sched_group_phys_bycpu[cpu] = NULL; | |
6095 | } | |
6096 | #ifdef CONFIG_SCHED_MC | |
6097 | if (sched_group_core_bycpu[cpu]) { | |
6098 | kfree(sched_group_core_bycpu[cpu]); | |
6099 | sched_group_core_bycpu[cpu] = NULL; | |
6100 | } | |
6101 | #endif | |
6102 | } | |
51888ca2 SV |
6103 | } |
6104 | ||
1da177e4 | 6105 | /* |
1a20ff27 DG |
6106 | * Build sched domains for a given set of cpus and attach the sched domains |
6107 | * to the individual cpus | |
1da177e4 | 6108 | */ |
51888ca2 | 6109 | static int build_sched_domains(const cpumask_t *cpu_map) |
1da177e4 LT |
6110 | { |
6111 | int i; | |
36938169 SV |
6112 | struct sched_group *sched_group_phys = NULL; |
6113 | #ifdef CONFIG_SCHED_MC | |
6114 | struct sched_group *sched_group_core = NULL; | |
6115 | #endif | |
d1b55138 JH |
6116 | #ifdef CONFIG_NUMA |
6117 | struct sched_group **sched_group_nodes = NULL; | |
6118 | struct sched_group *sched_group_allnodes = NULL; | |
6119 | ||
6120 | /* | |
6121 | * Allocate the per-node list of sched groups | |
6122 | */ | |
51888ca2 | 6123 | sched_group_nodes = kzalloc(sizeof(struct sched_group*)*MAX_NUMNODES, |
d3a5aa98 | 6124 | GFP_KERNEL); |
d1b55138 JH |
6125 | if (!sched_group_nodes) { |
6126 | printk(KERN_WARNING "Can not alloc sched group node list\n"); | |
51888ca2 | 6127 | return -ENOMEM; |
d1b55138 JH |
6128 | } |
6129 | sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes; | |
6130 | #endif | |
1da177e4 LT |
6131 | |
6132 | /* | |
1a20ff27 | 6133 | * Set up domains for cpus specified by the cpu_map. |
1da177e4 | 6134 | */ |
1a20ff27 | 6135 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 LT |
6136 | int group; |
6137 | struct sched_domain *sd = NULL, *p; | |
6138 | cpumask_t nodemask = node_to_cpumask(cpu_to_node(i)); | |
6139 | ||
1a20ff27 | 6140 | cpus_and(nodemask, nodemask, *cpu_map); |
1da177e4 LT |
6141 | |
6142 | #ifdef CONFIG_NUMA | |
d1b55138 | 6143 | if (cpus_weight(*cpu_map) |
9c1cfda2 | 6144 | > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) { |
d1b55138 JH |
6145 | if (!sched_group_allnodes) { |
6146 | sched_group_allnodes | |
6147 | = kmalloc(sizeof(struct sched_group) | |
6148 | * MAX_NUMNODES, | |
6149 | GFP_KERNEL); | |
6150 | if (!sched_group_allnodes) { | |
6151 | printk(KERN_WARNING | |
6152 | "Can not alloc allnodes sched group\n"); | |
51888ca2 | 6153 | goto error; |
d1b55138 JH |
6154 | } |
6155 | sched_group_allnodes_bycpu[i] | |
6156 | = sched_group_allnodes; | |
6157 | } | |
9c1cfda2 JH |
6158 | sd = &per_cpu(allnodes_domains, i); |
6159 | *sd = SD_ALLNODES_INIT; | |
6160 | sd->span = *cpu_map; | |
6161 | group = cpu_to_allnodes_group(i); | |
6162 | sd->groups = &sched_group_allnodes[group]; | |
6163 | p = sd; | |
6164 | } else | |
6165 | p = NULL; | |
6166 | ||
1da177e4 | 6167 | sd = &per_cpu(node_domains, i); |
1da177e4 | 6168 | *sd = SD_NODE_INIT; |
9c1cfda2 JH |
6169 | sd->span = sched_domain_node_span(cpu_to_node(i)); |
6170 | sd->parent = p; | |
6171 | cpus_and(sd->span, sd->span, *cpu_map); | |
1da177e4 LT |
6172 | #endif |
6173 | ||
36938169 SV |
6174 | if (!sched_group_phys) { |
6175 | sched_group_phys | |
6176 | = kmalloc(sizeof(struct sched_group) * NR_CPUS, | |
6177 | GFP_KERNEL); | |
6178 | if (!sched_group_phys) { | |
6179 | printk (KERN_WARNING "Can not alloc phys sched" | |
6180 | "group\n"); | |
6181 | goto error; | |
6182 | } | |
6183 | sched_group_phys_bycpu[i] = sched_group_phys; | |
6184 | } | |
6185 | ||
1da177e4 LT |
6186 | p = sd; |
6187 | sd = &per_cpu(phys_domains, i); | |
6188 | group = cpu_to_phys_group(i); | |
6189 | *sd = SD_CPU_INIT; | |
6190 | sd->span = nodemask; | |
6191 | sd->parent = p; | |
6192 | sd->groups = &sched_group_phys[group]; | |
6193 | ||
1e9f28fa | 6194 | #ifdef CONFIG_SCHED_MC |
36938169 SV |
6195 | if (!sched_group_core) { |
6196 | sched_group_core | |
6197 | = kmalloc(sizeof(struct sched_group) * NR_CPUS, | |
6198 | GFP_KERNEL); | |
6199 | if (!sched_group_core) { | |
6200 | printk (KERN_WARNING "Can not alloc core sched" | |
6201 | "group\n"); | |
6202 | goto error; | |
6203 | } | |
6204 | sched_group_core_bycpu[i] = sched_group_core; | |
6205 | } | |
6206 | ||
1e9f28fa SS |
6207 | p = sd; |
6208 | sd = &per_cpu(core_domains, i); | |
6209 | group = cpu_to_core_group(i); | |
6210 | *sd = SD_MC_INIT; | |
6211 | sd->span = cpu_coregroup_map(i); | |
6212 | cpus_and(sd->span, sd->span, *cpu_map); | |
6213 | sd->parent = p; | |
6214 | sd->groups = &sched_group_core[group]; | |
6215 | #endif | |
6216 | ||
1da177e4 LT |
6217 | #ifdef CONFIG_SCHED_SMT |
6218 | p = sd; | |
6219 | sd = &per_cpu(cpu_domains, i); | |
6220 | group = cpu_to_cpu_group(i); | |
6221 | *sd = SD_SIBLING_INIT; | |
6222 | sd->span = cpu_sibling_map[i]; | |
1a20ff27 | 6223 | cpus_and(sd->span, sd->span, *cpu_map); |
1da177e4 LT |
6224 | sd->parent = p; |
6225 | sd->groups = &sched_group_cpus[group]; | |
6226 | #endif | |
6227 | } | |
6228 | ||
6229 | #ifdef CONFIG_SCHED_SMT | |
6230 | /* Set up CPU (sibling) groups */ | |
9c1cfda2 | 6231 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 | 6232 | cpumask_t this_sibling_map = cpu_sibling_map[i]; |
1a20ff27 | 6233 | cpus_and(this_sibling_map, this_sibling_map, *cpu_map); |
1da177e4 LT |
6234 | if (i != first_cpu(this_sibling_map)) |
6235 | continue; | |
6236 | ||
6237 | init_sched_build_groups(sched_group_cpus, this_sibling_map, | |
6238 | &cpu_to_cpu_group); | |
6239 | } | |
6240 | #endif | |
6241 | ||
1e9f28fa SS |
6242 | #ifdef CONFIG_SCHED_MC |
6243 | /* Set up multi-core groups */ | |
6244 | for_each_cpu_mask(i, *cpu_map) { | |
6245 | cpumask_t this_core_map = cpu_coregroup_map(i); | |
6246 | cpus_and(this_core_map, this_core_map, *cpu_map); | |
6247 | if (i != first_cpu(this_core_map)) | |
6248 | continue; | |
6249 | init_sched_build_groups(sched_group_core, this_core_map, | |
6250 | &cpu_to_core_group); | |
6251 | } | |
6252 | #endif | |
6253 | ||
6254 | ||
1da177e4 LT |
6255 | /* Set up physical groups */ |
6256 | for (i = 0; i < MAX_NUMNODES; i++) { | |
6257 | cpumask_t nodemask = node_to_cpumask(i); | |
6258 | ||
1a20ff27 | 6259 | cpus_and(nodemask, nodemask, *cpu_map); |
1da177e4 LT |
6260 | if (cpus_empty(nodemask)) |
6261 | continue; | |
6262 | ||
6263 | init_sched_build_groups(sched_group_phys, nodemask, | |
6264 | &cpu_to_phys_group); | |
6265 | } | |
6266 | ||
6267 | #ifdef CONFIG_NUMA | |
6268 | /* Set up node groups */ | |
d1b55138 JH |
6269 | if (sched_group_allnodes) |
6270 | init_sched_build_groups(sched_group_allnodes, *cpu_map, | |
6271 | &cpu_to_allnodes_group); | |
9c1cfda2 JH |
6272 | |
6273 | for (i = 0; i < MAX_NUMNODES; i++) { | |
6274 | /* Set up node groups */ | |
6275 | struct sched_group *sg, *prev; | |
6276 | cpumask_t nodemask = node_to_cpumask(i); | |
6277 | cpumask_t domainspan; | |
6278 | cpumask_t covered = CPU_MASK_NONE; | |
6279 | int j; | |
6280 | ||
6281 | cpus_and(nodemask, nodemask, *cpu_map); | |
d1b55138 JH |
6282 | if (cpus_empty(nodemask)) { |
6283 | sched_group_nodes[i] = NULL; | |
9c1cfda2 | 6284 | continue; |
d1b55138 | 6285 | } |
9c1cfda2 JH |
6286 | |
6287 | domainspan = sched_domain_node_span(i); | |
6288 | cpus_and(domainspan, domainspan, *cpu_map); | |
6289 | ||
15f0b676 | 6290 | sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i); |
51888ca2 SV |
6291 | if (!sg) { |
6292 | printk(KERN_WARNING "Can not alloc domain group for " | |
6293 | "node %d\n", i); | |
6294 | goto error; | |
6295 | } | |
9c1cfda2 JH |
6296 | sched_group_nodes[i] = sg; |
6297 | for_each_cpu_mask(j, nodemask) { | |
6298 | struct sched_domain *sd; | |
6299 | sd = &per_cpu(node_domains, j); | |
6300 | sd->groups = sg; | |
9c1cfda2 JH |
6301 | } |
6302 | sg->cpu_power = 0; | |
6303 | sg->cpumask = nodemask; | |
51888ca2 | 6304 | sg->next = sg; |
9c1cfda2 JH |
6305 | cpus_or(covered, covered, nodemask); |
6306 | prev = sg; | |
6307 | ||
6308 | for (j = 0; j < MAX_NUMNODES; j++) { | |
6309 | cpumask_t tmp, notcovered; | |
6310 | int n = (i + j) % MAX_NUMNODES; | |
6311 | ||
6312 | cpus_complement(notcovered, covered); | |
6313 | cpus_and(tmp, notcovered, *cpu_map); | |
6314 | cpus_and(tmp, tmp, domainspan); | |
6315 | if (cpus_empty(tmp)) | |
6316 | break; | |
6317 | ||
6318 | nodemask = node_to_cpumask(n); | |
6319 | cpus_and(tmp, tmp, nodemask); | |
6320 | if (cpus_empty(tmp)) | |
6321 | continue; | |
6322 | ||
15f0b676 SV |
6323 | sg = kmalloc_node(sizeof(struct sched_group), |
6324 | GFP_KERNEL, i); | |
9c1cfda2 JH |
6325 | if (!sg) { |
6326 | printk(KERN_WARNING | |
6327 | "Can not alloc domain group for node %d\n", j); | |
51888ca2 | 6328 | goto error; |
9c1cfda2 JH |
6329 | } |
6330 | sg->cpu_power = 0; | |
6331 | sg->cpumask = tmp; | |
51888ca2 | 6332 | sg->next = prev->next; |
9c1cfda2 JH |
6333 | cpus_or(covered, covered, tmp); |
6334 | prev->next = sg; | |
6335 | prev = sg; | |
6336 | } | |
9c1cfda2 | 6337 | } |
1da177e4 LT |
6338 | #endif |
6339 | ||
6340 | /* Calculate CPU power for physical packages and nodes */ | |
5c45bf27 | 6341 | #ifdef CONFIG_SCHED_SMT |
1a20ff27 | 6342 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 | 6343 | struct sched_domain *sd; |
1da177e4 | 6344 | sd = &per_cpu(cpu_domains, i); |
5c45bf27 SS |
6345 | sd->groups->cpu_power = SCHED_LOAD_SCALE; |
6346 | } | |
1da177e4 | 6347 | #endif |
1e9f28fa | 6348 | #ifdef CONFIG_SCHED_MC |
5c45bf27 SS |
6349 | for_each_cpu_mask(i, *cpu_map) { |
6350 | int power; | |
6351 | struct sched_domain *sd; | |
1e9f28fa | 6352 | sd = &per_cpu(core_domains, i); |
5c45bf27 SS |
6353 | if (sched_smt_power_savings) |
6354 | power = SCHED_LOAD_SCALE * cpus_weight(sd->groups->cpumask); | |
6355 | else | |
6356 | power = SCHED_LOAD_SCALE + (cpus_weight(sd->groups->cpumask)-1) | |
1e9f28fa SS |
6357 | * SCHED_LOAD_SCALE / 10; |
6358 | sd->groups->cpu_power = power; | |
5c45bf27 SS |
6359 | } |
6360 | #endif | |
1e9f28fa | 6361 | |
5c45bf27 SS |
6362 | for_each_cpu_mask(i, *cpu_map) { |
6363 | struct sched_domain *sd; | |
6364 | #ifdef CONFIG_SCHED_MC | |
1e9f28fa | 6365 | sd = &per_cpu(phys_domains, i); |
5c45bf27 SS |
6366 | if (i != first_cpu(sd->groups->cpumask)) |
6367 | continue; | |
1da177e4 | 6368 | |
5c45bf27 SS |
6369 | sd->groups->cpu_power = 0; |
6370 | if (sched_mc_power_savings || sched_smt_power_savings) { | |
6371 | int j; | |
6372 | ||
6373 | for_each_cpu_mask(j, sd->groups->cpumask) { | |
6374 | struct sched_domain *sd1; | |
6375 | sd1 = &per_cpu(core_domains, j); | |
6376 | /* | |
6377 | * for each core we will add once | |
6378 | * to the group in physical domain | |
6379 | */ | |
6380 | if (j != first_cpu(sd1->groups->cpumask)) | |
6381 | continue; | |
6382 | ||
6383 | if (sched_smt_power_savings) | |
6384 | sd->groups->cpu_power += sd1->groups->cpu_power; | |
6385 | else | |
6386 | sd->groups->cpu_power += SCHED_LOAD_SCALE; | |
6387 | } | |
6388 | } else | |
6389 | /* | |
6390 | * This has to be < 2 * SCHED_LOAD_SCALE | |
6391 | * Lets keep it SCHED_LOAD_SCALE, so that | |
6392 | * while calculating NUMA group's cpu_power | |
6393 | * we can simply do | |
6394 | * numa_group->cpu_power += phys_group->cpu_power; | |
6395 | * | |
6396 | * See "only add power once for each physical pkg" | |
6397 | * comment below | |
6398 | */ | |
6399 | sd->groups->cpu_power = SCHED_LOAD_SCALE; | |
1e9f28fa | 6400 | #else |
5c45bf27 | 6401 | int power; |
1da177e4 | 6402 | sd = &per_cpu(phys_domains, i); |
5c45bf27 SS |
6403 | if (sched_smt_power_savings) |
6404 | power = SCHED_LOAD_SCALE * cpus_weight(sd->groups->cpumask); | |
6405 | else | |
6406 | power = SCHED_LOAD_SCALE; | |
1da177e4 | 6407 | sd->groups->cpu_power = power; |
1e9f28fa | 6408 | #endif |
1da177e4 LT |
6409 | } |
6410 | ||
9c1cfda2 | 6411 | #ifdef CONFIG_NUMA |
08069033 SS |
6412 | for (i = 0; i < MAX_NUMNODES; i++) |
6413 | init_numa_sched_groups_power(sched_group_nodes[i]); | |
9c1cfda2 | 6414 | |
08069033 | 6415 | init_numa_sched_groups_power(sched_group_allnodes); |
9c1cfda2 JH |
6416 | #endif |
6417 | ||
1da177e4 | 6418 | /* Attach the domains */ |
1a20ff27 | 6419 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 LT |
6420 | struct sched_domain *sd; |
6421 | #ifdef CONFIG_SCHED_SMT | |
6422 | sd = &per_cpu(cpu_domains, i); | |
1e9f28fa SS |
6423 | #elif defined(CONFIG_SCHED_MC) |
6424 | sd = &per_cpu(core_domains, i); | |
1da177e4 LT |
6425 | #else |
6426 | sd = &per_cpu(phys_domains, i); | |
6427 | #endif | |
6428 | cpu_attach_domain(sd, i); | |
6429 | } | |
198e2f18 | 6430 | /* |
6431 | * Tune cache-hot values: | |
6432 | */ | |
6433 | calibrate_migration_costs(cpu_map); | |
51888ca2 SV |
6434 | |
6435 | return 0; | |
6436 | ||
51888ca2 SV |
6437 | error: |
6438 | free_sched_groups(cpu_map); | |
6439 | return -ENOMEM; | |
1da177e4 | 6440 | } |
1a20ff27 DG |
6441 | /* |
6442 | * Set up scheduler domains and groups. Callers must hold the hotplug lock. | |
6443 | */ | |
51888ca2 | 6444 | static int arch_init_sched_domains(const cpumask_t *cpu_map) |
1a20ff27 DG |
6445 | { |
6446 | cpumask_t cpu_default_map; | |
51888ca2 | 6447 | int err; |
1da177e4 | 6448 | |
1a20ff27 DG |
6449 | /* |
6450 | * Setup mask for cpus without special case scheduling requirements. | |
6451 | * For now this just excludes isolated cpus, but could be used to | |
6452 | * exclude other special cases in the future. | |
6453 | */ | |
6454 | cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map); | |
6455 | ||
51888ca2 SV |
6456 | err = build_sched_domains(&cpu_default_map); |
6457 | ||
6458 | return err; | |
1a20ff27 DG |
6459 | } |
6460 | ||
6461 | static void arch_destroy_sched_domains(const cpumask_t *cpu_map) | |
1da177e4 | 6462 | { |
51888ca2 | 6463 | free_sched_groups(cpu_map); |
9c1cfda2 | 6464 | } |
1da177e4 | 6465 | |
1a20ff27 DG |
6466 | /* |
6467 | * Detach sched domains from a group of cpus specified in cpu_map | |
6468 | * These cpus will now be attached to the NULL domain | |
6469 | */ | |
858119e1 | 6470 | static void detach_destroy_domains(const cpumask_t *cpu_map) |
1a20ff27 DG |
6471 | { |
6472 | int i; | |
6473 | ||
6474 | for_each_cpu_mask(i, *cpu_map) | |
6475 | cpu_attach_domain(NULL, i); | |
6476 | synchronize_sched(); | |
6477 | arch_destroy_sched_domains(cpu_map); | |
6478 | } | |
6479 | ||
6480 | /* | |
6481 | * Partition sched domains as specified by the cpumasks below. | |
6482 | * This attaches all cpus from the cpumasks to the NULL domain, | |
6483 | * waits for a RCU quiescent period, recalculates sched | |
6484 | * domain information and then attaches them back to the | |
6485 | * correct sched domains | |
6486 | * Call with hotplug lock held | |
6487 | */ | |
51888ca2 | 6488 | int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2) |
1a20ff27 DG |
6489 | { |
6490 | cpumask_t change_map; | |
51888ca2 | 6491 | int err = 0; |
1a20ff27 DG |
6492 | |
6493 | cpus_and(*partition1, *partition1, cpu_online_map); | |
6494 | cpus_and(*partition2, *partition2, cpu_online_map); | |
6495 | cpus_or(change_map, *partition1, *partition2); | |
6496 | ||
6497 | /* Detach sched domains from all of the affected cpus */ | |
6498 | detach_destroy_domains(&change_map); | |
6499 | if (!cpus_empty(*partition1)) | |
51888ca2 SV |
6500 | err = build_sched_domains(partition1); |
6501 | if (!err && !cpus_empty(*partition2)) | |
6502 | err = build_sched_domains(partition2); | |
6503 | ||
6504 | return err; | |
1a20ff27 DG |
6505 | } |
6506 | ||
5c45bf27 SS |
6507 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
6508 | int arch_reinit_sched_domains(void) | |
6509 | { | |
6510 | int err; | |
6511 | ||
6512 | lock_cpu_hotplug(); | |
6513 | detach_destroy_domains(&cpu_online_map); | |
6514 | err = arch_init_sched_domains(&cpu_online_map); | |
6515 | unlock_cpu_hotplug(); | |
6516 | ||
6517 | return err; | |
6518 | } | |
6519 | ||
6520 | static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt) | |
6521 | { | |
6522 | int ret; | |
6523 | ||
6524 | if (buf[0] != '0' && buf[0] != '1') | |
6525 | return -EINVAL; | |
6526 | ||
6527 | if (smt) | |
6528 | sched_smt_power_savings = (buf[0] == '1'); | |
6529 | else | |
6530 | sched_mc_power_savings = (buf[0] == '1'); | |
6531 | ||
6532 | ret = arch_reinit_sched_domains(); | |
6533 | ||
6534 | return ret ? ret : count; | |
6535 | } | |
6536 | ||
6537 | int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls) | |
6538 | { | |
6539 | int err = 0; | |
6540 | #ifdef CONFIG_SCHED_SMT | |
6541 | if (smt_capable()) | |
6542 | err = sysfs_create_file(&cls->kset.kobj, | |
6543 | &attr_sched_smt_power_savings.attr); | |
6544 | #endif | |
6545 | #ifdef CONFIG_SCHED_MC | |
6546 | if (!err && mc_capable()) | |
6547 | err = sysfs_create_file(&cls->kset.kobj, | |
6548 | &attr_sched_mc_power_savings.attr); | |
6549 | #endif | |
6550 | return err; | |
6551 | } | |
6552 | #endif | |
6553 | ||
6554 | #ifdef CONFIG_SCHED_MC | |
6555 | static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page) | |
6556 | { | |
6557 | return sprintf(page, "%u\n", sched_mc_power_savings); | |
6558 | } | |
6559 | static ssize_t sched_mc_power_savings_store(struct sys_device *dev, const char *buf, size_t count) | |
6560 | { | |
6561 | return sched_power_savings_store(buf, count, 0); | |
6562 | } | |
6563 | SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show, | |
6564 | sched_mc_power_savings_store); | |
6565 | #endif | |
6566 | ||
6567 | #ifdef CONFIG_SCHED_SMT | |
6568 | static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page) | |
6569 | { | |
6570 | return sprintf(page, "%u\n", sched_smt_power_savings); | |
6571 | } | |
6572 | static ssize_t sched_smt_power_savings_store(struct sys_device *dev, const char *buf, size_t count) | |
6573 | { | |
6574 | return sched_power_savings_store(buf, count, 1); | |
6575 | } | |
6576 | SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show, | |
6577 | sched_smt_power_savings_store); | |
6578 | #endif | |
6579 | ||
6580 | ||
1da177e4 LT |
6581 | #ifdef CONFIG_HOTPLUG_CPU |
6582 | /* | |
6583 | * Force a reinitialization of the sched domains hierarchy. The domains | |
6584 | * and groups cannot be updated in place without racing with the balancing | |
41c7ce9a | 6585 | * code, so we temporarily attach all running cpus to the NULL domain |
1da177e4 LT |
6586 | * which will prevent rebalancing while the sched domains are recalculated. |
6587 | */ | |
6588 | static int update_sched_domains(struct notifier_block *nfb, | |
6589 | unsigned long action, void *hcpu) | |
6590 | { | |
1da177e4 LT |
6591 | switch (action) { |
6592 | case CPU_UP_PREPARE: | |
6593 | case CPU_DOWN_PREPARE: | |
1a20ff27 | 6594 | detach_destroy_domains(&cpu_online_map); |
1da177e4 LT |
6595 | return NOTIFY_OK; |
6596 | ||
6597 | case CPU_UP_CANCELED: | |
6598 | case CPU_DOWN_FAILED: | |
6599 | case CPU_ONLINE: | |
6600 | case CPU_DEAD: | |
6601 | /* | |
6602 | * Fall through and re-initialise the domains. | |
6603 | */ | |
6604 | break; | |
6605 | default: | |
6606 | return NOTIFY_DONE; | |
6607 | } | |
6608 | ||
6609 | /* The hotplug lock is already held by cpu_up/cpu_down */ | |
1a20ff27 | 6610 | arch_init_sched_domains(&cpu_online_map); |
1da177e4 LT |
6611 | |
6612 | return NOTIFY_OK; | |
6613 | } | |
6614 | #endif | |
6615 | ||
6616 | void __init sched_init_smp(void) | |
6617 | { | |
6618 | lock_cpu_hotplug(); | |
1a20ff27 | 6619 | arch_init_sched_domains(&cpu_online_map); |
1da177e4 LT |
6620 | unlock_cpu_hotplug(); |
6621 | /* XXX: Theoretical race here - CPU may be hotplugged now */ | |
6622 | hotcpu_notifier(update_sched_domains, 0); | |
6623 | } | |
6624 | #else | |
6625 | void __init sched_init_smp(void) | |
6626 | { | |
6627 | } | |
6628 | #endif /* CONFIG_SMP */ | |
6629 | ||
6630 | int in_sched_functions(unsigned long addr) | |
6631 | { | |
6632 | /* Linker adds these: start and end of __sched functions */ | |
6633 | extern char __sched_text_start[], __sched_text_end[]; | |
6634 | return in_lock_functions(addr) || | |
6635 | (addr >= (unsigned long)__sched_text_start | |
6636 | && addr < (unsigned long)__sched_text_end); | |
6637 | } | |
6638 | ||
6639 | void __init sched_init(void) | |
6640 | { | |
6641 | runqueue_t *rq; | |
6642 | int i, j, k; | |
6643 | ||
0a945022 | 6644 | for_each_possible_cpu(i) { |
1da177e4 LT |
6645 | prio_array_t *array; |
6646 | ||
6647 | rq = cpu_rq(i); | |
6648 | spin_lock_init(&rq->lock); | |
7897986b | 6649 | rq->nr_running = 0; |
1da177e4 LT |
6650 | rq->active = rq->arrays; |
6651 | rq->expired = rq->arrays + 1; | |
6652 | rq->best_expired_prio = MAX_PRIO; | |
6653 | ||
6654 | #ifdef CONFIG_SMP | |
41c7ce9a | 6655 | rq->sd = NULL; |
7897986b NP |
6656 | for (j = 1; j < 3; j++) |
6657 | rq->cpu_load[j] = 0; | |
1da177e4 LT |
6658 | rq->active_balance = 0; |
6659 | rq->push_cpu = 0; | |
6660 | rq->migration_thread = NULL; | |
6661 | INIT_LIST_HEAD(&rq->migration_queue); | |
6662 | #endif | |
6663 | atomic_set(&rq->nr_iowait, 0); | |
6664 | ||
6665 | for (j = 0; j < 2; j++) { | |
6666 | array = rq->arrays + j; | |
6667 | for (k = 0; k < MAX_PRIO; k++) { | |
6668 | INIT_LIST_HEAD(array->queue + k); | |
6669 | __clear_bit(k, array->bitmap); | |
6670 | } | |
6671 | // delimiter for bitsearch | |
6672 | __set_bit(MAX_PRIO, array->bitmap); | |
6673 | } | |
6674 | } | |
6675 | ||
2dd73a4f | 6676 | set_load_weight(&init_task); |
1da177e4 LT |
6677 | /* |
6678 | * The boot idle thread does lazy MMU switching as well: | |
6679 | */ | |
6680 | atomic_inc(&init_mm.mm_count); | |
6681 | enter_lazy_tlb(&init_mm, current); | |
6682 | ||
6683 | /* | |
6684 | * Make us the idle thread. Technically, schedule() should not be | |
6685 | * called from this thread, however somewhere below it might be, | |
6686 | * but because we are the idle thread, we just pick up running again | |
6687 | * when this runqueue becomes "idle". | |
6688 | */ | |
6689 | init_idle(current, smp_processor_id()); | |
6690 | } | |
6691 | ||
6692 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP | |
6693 | void __might_sleep(char *file, int line) | |
6694 | { | |
6695 | #if defined(in_atomic) | |
6696 | static unsigned long prev_jiffy; /* ratelimiting */ | |
6697 | ||
6698 | if ((in_atomic() || irqs_disabled()) && | |
6699 | system_state == SYSTEM_RUNNING && !oops_in_progress) { | |
6700 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) | |
6701 | return; | |
6702 | prev_jiffy = jiffies; | |
91368d73 | 6703 | printk(KERN_ERR "BUG: sleeping function called from invalid" |
1da177e4 LT |
6704 | " context at %s:%d\n", file, line); |
6705 | printk("in_atomic():%d, irqs_disabled():%d\n", | |
6706 | in_atomic(), irqs_disabled()); | |
6707 | dump_stack(); | |
6708 | } | |
6709 | #endif | |
6710 | } | |
6711 | EXPORT_SYMBOL(__might_sleep); | |
6712 | #endif | |
6713 | ||
6714 | #ifdef CONFIG_MAGIC_SYSRQ | |
6715 | void normalize_rt_tasks(void) | |
6716 | { | |
6717 | struct task_struct *p; | |
6718 | prio_array_t *array; | |
6719 | unsigned long flags; | |
6720 | runqueue_t *rq; | |
6721 | ||
6722 | read_lock_irq(&tasklist_lock); | |
c96d145e | 6723 | for_each_process(p) { |
1da177e4 LT |
6724 | if (!rt_task(p)) |
6725 | continue; | |
6726 | ||
b29739f9 IM |
6727 | spin_lock_irqsave(&p->pi_lock, flags); |
6728 | rq = __task_rq_lock(p); | |
1da177e4 LT |
6729 | |
6730 | array = p->array; | |
6731 | if (array) | |
6732 | deactivate_task(p, task_rq(p)); | |
6733 | __setscheduler(p, SCHED_NORMAL, 0); | |
6734 | if (array) { | |
6735 | __activate_task(p, task_rq(p)); | |
6736 | resched_task(rq->curr); | |
6737 | } | |
6738 | ||
b29739f9 IM |
6739 | __task_rq_unlock(rq); |
6740 | spin_unlock_irqrestore(&p->pi_lock, flags); | |
1da177e4 LT |
6741 | } |
6742 | read_unlock_irq(&tasklist_lock); | |
6743 | } | |
6744 | ||
6745 | #endif /* CONFIG_MAGIC_SYSRQ */ | |
1df5c10a LT |
6746 | |
6747 | #ifdef CONFIG_IA64 | |
6748 | /* | |
6749 | * These functions are only useful for the IA64 MCA handling. | |
6750 | * | |
6751 | * They can only be called when the whole system has been | |
6752 | * stopped - every CPU needs to be quiescent, and no scheduling | |
6753 | * activity can take place. Using them for anything else would | |
6754 | * be a serious bug, and as a result, they aren't even visible | |
6755 | * under any other configuration. | |
6756 | */ | |
6757 | ||
6758 | /** | |
6759 | * curr_task - return the current task for a given cpu. | |
6760 | * @cpu: the processor in question. | |
6761 | * | |
6762 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
6763 | */ | |
6764 | task_t *curr_task(int cpu) | |
6765 | { | |
6766 | return cpu_curr(cpu); | |
6767 | } | |
6768 | ||
6769 | /** | |
6770 | * set_curr_task - set the current task for a given cpu. | |
6771 | * @cpu: the processor in question. | |
6772 | * @p: the task pointer to set. | |
6773 | * | |
6774 | * Description: This function must only be used when non-maskable interrupts | |
6775 | * are serviced on a separate stack. It allows the architecture to switch the | |
6776 | * notion of the current task on a cpu in a non-blocking manner. This function | |
6777 | * must be called with all CPU's synchronized, and interrupts disabled, the | |
6778 | * and caller must save the original value of the current task (see | |
6779 | * curr_task() above) and restore that value before reenabling interrupts and | |
6780 | * re-starting the system. | |
6781 | * | |
6782 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
6783 | */ | |
6784 | void set_curr_task(int cpu, task_t *p) | |
6785 | { | |
6786 | cpu_curr(cpu) = p; | |
6787 | } | |
6788 | ||
6789 | #endif |