]>
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> | |
7dfb7103 | 37 | #include <linux/freezer.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> | |
8f0ab514 | 52 | #include <linux/tsacct_kern.h> |
c6fd91f0 | 53 | #include <linux/kprobes.h> |
0ff92245 | 54 | #include <linux/delayacct.h> |
1da177e4 LT |
55 | #include <asm/tlb.h> |
56 | ||
57 | #include <asm/unistd.h> | |
58 | ||
b035b6de AD |
59 | /* |
60 | * Scheduler clock - returns current time in nanosec units. | |
61 | * This is default implementation. | |
62 | * Architectures and sub-architectures can override this. | |
63 | */ | |
64 | unsigned long long __attribute__((weak)) sched_clock(void) | |
65 | { | |
66 | return (unsigned long long)jiffies * (1000000000 / HZ); | |
67 | } | |
68 | ||
1da177e4 LT |
69 | /* |
70 | * Convert user-nice values [ -20 ... 0 ... 19 ] | |
71 | * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ], | |
72 | * and back. | |
73 | */ | |
74 | #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20) | |
75 | #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20) | |
76 | #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio) | |
77 | ||
78 | /* | |
79 | * 'User priority' is the nice value converted to something we | |
80 | * can work with better when scaling various scheduler parameters, | |
81 | * it's a [ 0 ... 39 ] range. | |
82 | */ | |
83 | #define USER_PRIO(p) ((p)-MAX_RT_PRIO) | |
84 | #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio) | |
85 | #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO)) | |
86 | ||
87 | /* | |
88 | * Some helpers for converting nanosecond timing to jiffy resolution | |
89 | */ | |
90 | #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ)) | |
91 | #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ)) | |
92 | ||
93 | /* | |
94 | * These are the 'tuning knobs' of the scheduler: | |
95 | * | |
96 | * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger), | |
97 | * default timeslice is 100 msecs, maximum timeslice is 800 msecs. | |
98 | * Timeslices get refilled after they expire. | |
99 | */ | |
100 | #define MIN_TIMESLICE max(5 * HZ / 1000, 1) | |
101 | #define DEF_TIMESLICE (100 * HZ / 1000) | |
102 | #define ON_RUNQUEUE_WEIGHT 30 | |
103 | #define CHILD_PENALTY 95 | |
104 | #define PARENT_PENALTY 100 | |
105 | #define EXIT_WEIGHT 3 | |
106 | #define PRIO_BONUS_RATIO 25 | |
107 | #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100) | |
108 | #define INTERACTIVE_DELTA 2 | |
109 | #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS) | |
110 | #define STARVATION_LIMIT (MAX_SLEEP_AVG) | |
111 | #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG)) | |
112 | ||
113 | /* | |
114 | * If a task is 'interactive' then we reinsert it in the active | |
115 | * array after it has expired its current timeslice. (it will not | |
116 | * continue to run immediately, it will still roundrobin with | |
117 | * other interactive tasks.) | |
118 | * | |
119 | * This part scales the interactivity limit depending on niceness. | |
120 | * | |
121 | * We scale it linearly, offset by the INTERACTIVE_DELTA delta. | |
122 | * Here are a few examples of different nice levels: | |
123 | * | |
124 | * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0] | |
125 | * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0] | |
126 | * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0] | |
127 | * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0] | |
128 | * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0] | |
129 | * | |
130 | * (the X axis represents the possible -5 ... 0 ... +5 dynamic | |
131 | * priority range a task can explore, a value of '1' means the | |
132 | * task is rated interactive.) | |
133 | * | |
134 | * Ie. nice +19 tasks can never get 'interactive' enough to be | |
135 | * reinserted into the active array. And only heavily CPU-hog nice -20 | |
136 | * tasks will be expired. Default nice 0 tasks are somewhere between, | |
137 | * it takes some effort for them to get interactive, but it's not | |
138 | * too hard. | |
139 | */ | |
140 | ||
141 | #define CURRENT_BONUS(p) \ | |
142 | (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \ | |
143 | MAX_SLEEP_AVG) | |
144 | ||
145 | #define GRANULARITY (10 * HZ / 1000 ? : 1) | |
146 | ||
147 | #ifdef CONFIG_SMP | |
148 | #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \ | |
149 | (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \ | |
150 | num_online_cpus()) | |
151 | #else | |
152 | #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \ | |
153 | (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1))) | |
154 | #endif | |
155 | ||
156 | #define SCALE(v1,v1_max,v2_max) \ | |
157 | (v1) * (v2_max) / (v1_max) | |
158 | ||
159 | #define DELTA(p) \ | |
013d3868 MA |
160 | (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \ |
161 | INTERACTIVE_DELTA) | |
1da177e4 LT |
162 | |
163 | #define TASK_INTERACTIVE(p) \ | |
164 | ((p)->prio <= (p)->static_prio - DELTA(p)) | |
165 | ||
166 | #define INTERACTIVE_SLEEP(p) \ | |
167 | (JIFFIES_TO_NS(MAX_SLEEP_AVG * \ | |
168 | (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1)) | |
169 | ||
170 | #define TASK_PREEMPTS_CURR(p, rq) \ | |
171 | ((p)->prio < (rq)->curr->prio) | |
172 | ||
1da177e4 | 173 | #define SCALE_PRIO(x, prio) \ |
2dd73a4f | 174 | max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE) |
1da177e4 | 175 | |
2dd73a4f | 176 | static unsigned int static_prio_timeslice(int static_prio) |
1da177e4 | 177 | { |
2dd73a4f PW |
178 | if (static_prio < NICE_TO_PRIO(0)) |
179 | return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio); | |
1da177e4 | 180 | else |
2dd73a4f | 181 | return SCALE_PRIO(DEF_TIMESLICE, static_prio); |
1da177e4 | 182 | } |
2dd73a4f | 183 | |
91fcdd4e BP |
184 | /* |
185 | * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ] | |
186 | * to time slice values: [800ms ... 100ms ... 5ms] | |
187 | * | |
188 | * The higher a thread's priority, the bigger timeslices | |
189 | * it gets during one round of execution. But even the lowest | |
190 | * priority thread gets MIN_TIMESLICE worth of execution time. | |
191 | */ | |
192 | ||
36c8b586 | 193 | static inline unsigned int task_timeslice(struct task_struct *p) |
2dd73a4f PW |
194 | { |
195 | return static_prio_timeslice(p->static_prio); | |
196 | } | |
197 | ||
1da177e4 LT |
198 | /* |
199 | * These are the runqueue data structures: | |
200 | */ | |
201 | ||
1da177e4 LT |
202 | struct prio_array { |
203 | unsigned int nr_active; | |
d444886e | 204 | DECLARE_BITMAP(bitmap, MAX_PRIO+1); /* include 1 bit for delimiter */ |
1da177e4 LT |
205 | struct list_head queue[MAX_PRIO]; |
206 | }; | |
207 | ||
208 | /* | |
209 | * This is the main, per-CPU runqueue data structure. | |
210 | * | |
211 | * Locking rule: those places that want to lock multiple runqueues | |
212 | * (such as the load balancing or the thread migration code), lock | |
213 | * acquire operations must be ordered by ascending &runqueue. | |
214 | */ | |
70b97a7f | 215 | struct rq { |
1da177e4 LT |
216 | spinlock_t lock; |
217 | ||
218 | /* | |
219 | * nr_running and cpu_load should be in the same cacheline because | |
220 | * remote CPUs use both these fields when doing load calculation. | |
221 | */ | |
222 | unsigned long nr_running; | |
2dd73a4f | 223 | unsigned long raw_weighted_load; |
1da177e4 | 224 | #ifdef CONFIG_SMP |
7897986b | 225 | unsigned long cpu_load[3]; |
1da177e4 LT |
226 | #endif |
227 | unsigned long long nr_switches; | |
228 | ||
229 | /* | |
230 | * This is part of a global counter where only the total sum | |
231 | * over all CPUs matters. A task can increase this counter on | |
232 | * one CPU and if it got migrated afterwards it may decrease | |
233 | * it on another CPU. Always updated under the runqueue lock: | |
234 | */ | |
235 | unsigned long nr_uninterruptible; | |
236 | ||
237 | unsigned long expired_timestamp; | |
b18ec803 MG |
238 | /* Cached timestamp set by update_cpu_clock() */ |
239 | unsigned long long most_recent_timestamp; | |
36c8b586 | 240 | struct task_struct *curr, *idle; |
c9819f45 | 241 | unsigned long next_balance; |
1da177e4 | 242 | struct mm_struct *prev_mm; |
70b97a7f | 243 | struct prio_array *active, *expired, arrays[2]; |
1da177e4 LT |
244 | int best_expired_prio; |
245 | atomic_t nr_iowait; | |
246 | ||
247 | #ifdef CONFIG_SMP | |
248 | struct sched_domain *sd; | |
249 | ||
250 | /* For active balancing */ | |
251 | int active_balance; | |
252 | int push_cpu; | |
0a2966b4 | 253 | int cpu; /* cpu of this runqueue */ |
1da177e4 | 254 | |
36c8b586 | 255 | struct task_struct *migration_thread; |
1da177e4 LT |
256 | struct list_head migration_queue; |
257 | #endif | |
258 | ||
259 | #ifdef CONFIG_SCHEDSTATS | |
260 | /* latency stats */ | |
261 | struct sched_info rq_sched_info; | |
262 | ||
263 | /* sys_sched_yield() stats */ | |
264 | unsigned long yld_exp_empty; | |
265 | unsigned long yld_act_empty; | |
266 | unsigned long yld_both_empty; | |
267 | unsigned long yld_cnt; | |
268 | ||
269 | /* schedule() stats */ | |
270 | unsigned long sched_switch; | |
271 | unsigned long sched_cnt; | |
272 | unsigned long sched_goidle; | |
273 | ||
274 | /* try_to_wake_up() stats */ | |
275 | unsigned long ttwu_cnt; | |
276 | unsigned long ttwu_local; | |
277 | #endif | |
fcb99371 | 278 | struct lock_class_key rq_lock_key; |
1da177e4 LT |
279 | }; |
280 | ||
70b97a7f | 281 | static DEFINE_PER_CPU(struct rq, runqueues); |
1da177e4 | 282 | |
0a2966b4 CL |
283 | static inline int cpu_of(struct rq *rq) |
284 | { | |
285 | #ifdef CONFIG_SMP | |
286 | return rq->cpu; | |
287 | #else | |
288 | return 0; | |
289 | #endif | |
290 | } | |
291 | ||
674311d5 NP |
292 | /* |
293 | * The domain tree (rq->sd) is protected by RCU's quiescent state transition. | |
1a20ff27 | 294 | * See detach_destroy_domains: synchronize_sched for details. |
674311d5 NP |
295 | * |
296 | * The domain tree of any CPU may only be accessed from within | |
297 | * preempt-disabled sections. | |
298 | */ | |
48f24c4d IM |
299 | #define for_each_domain(cpu, __sd) \ |
300 | for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent) | |
1da177e4 LT |
301 | |
302 | #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) | |
303 | #define this_rq() (&__get_cpu_var(runqueues)) | |
304 | #define task_rq(p) cpu_rq(task_cpu(p)) | |
305 | #define cpu_curr(cpu) (cpu_rq(cpu)->curr) | |
306 | ||
1da177e4 | 307 | #ifndef prepare_arch_switch |
4866cde0 NP |
308 | # define prepare_arch_switch(next) do { } while (0) |
309 | #endif | |
310 | #ifndef finish_arch_switch | |
311 | # define finish_arch_switch(prev) do { } while (0) | |
312 | #endif | |
313 | ||
314 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW | |
70b97a7f | 315 | static inline int task_running(struct rq *rq, struct task_struct *p) |
4866cde0 NP |
316 | { |
317 | return rq->curr == p; | |
318 | } | |
319 | ||
70b97a7f | 320 | static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) |
4866cde0 NP |
321 | { |
322 | } | |
323 | ||
70b97a7f | 324 | static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) |
4866cde0 | 325 | { |
da04c035 IM |
326 | #ifdef CONFIG_DEBUG_SPINLOCK |
327 | /* this is a valid case when another task releases the spinlock */ | |
328 | rq->lock.owner = current; | |
329 | #endif | |
8a25d5de IM |
330 | /* |
331 | * If we are tracking spinlock dependencies then we have to | |
332 | * fix up the runqueue lock - which gets 'carried over' from | |
333 | * prev into current: | |
334 | */ | |
335 | spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_); | |
336 | ||
4866cde0 NP |
337 | spin_unlock_irq(&rq->lock); |
338 | } | |
339 | ||
340 | #else /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
70b97a7f | 341 | static inline int task_running(struct rq *rq, struct task_struct *p) |
4866cde0 NP |
342 | { |
343 | #ifdef CONFIG_SMP | |
344 | return p->oncpu; | |
345 | #else | |
346 | return rq->curr == p; | |
347 | #endif | |
348 | } | |
349 | ||
70b97a7f | 350 | static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) |
4866cde0 NP |
351 | { |
352 | #ifdef CONFIG_SMP | |
353 | /* | |
354 | * We can optimise this out completely for !SMP, because the | |
355 | * SMP rebalancing from interrupt is the only thing that cares | |
356 | * here. | |
357 | */ | |
358 | next->oncpu = 1; | |
359 | #endif | |
360 | #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
361 | spin_unlock_irq(&rq->lock); | |
362 | #else | |
363 | spin_unlock(&rq->lock); | |
364 | #endif | |
365 | } | |
366 | ||
70b97a7f | 367 | static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) |
4866cde0 NP |
368 | { |
369 | #ifdef CONFIG_SMP | |
370 | /* | |
371 | * After ->oncpu is cleared, the task can be moved to a different CPU. | |
372 | * We must ensure this doesn't happen until the switch is completely | |
373 | * finished. | |
374 | */ | |
375 | smp_wmb(); | |
376 | prev->oncpu = 0; | |
377 | #endif | |
378 | #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
379 | local_irq_enable(); | |
1da177e4 | 380 | #endif |
4866cde0 NP |
381 | } |
382 | #endif /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
1da177e4 | 383 | |
b29739f9 IM |
384 | /* |
385 | * __task_rq_lock - lock the runqueue a given task resides on. | |
386 | * Must be called interrupts disabled. | |
387 | */ | |
70b97a7f | 388 | static inline struct rq *__task_rq_lock(struct task_struct *p) |
b29739f9 IM |
389 | __acquires(rq->lock) |
390 | { | |
70b97a7f | 391 | struct rq *rq; |
b29739f9 IM |
392 | |
393 | repeat_lock_task: | |
394 | rq = task_rq(p); | |
395 | spin_lock(&rq->lock); | |
396 | if (unlikely(rq != task_rq(p))) { | |
397 | spin_unlock(&rq->lock); | |
398 | goto repeat_lock_task; | |
399 | } | |
400 | return rq; | |
401 | } | |
402 | ||
1da177e4 LT |
403 | /* |
404 | * task_rq_lock - lock the runqueue a given task resides on and disable | |
405 | * interrupts. Note the ordering: we can safely lookup the task_rq without | |
406 | * explicitly disabling preemption. | |
407 | */ | |
70b97a7f | 408 | static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) |
1da177e4 LT |
409 | __acquires(rq->lock) |
410 | { | |
70b97a7f | 411 | struct rq *rq; |
1da177e4 LT |
412 | |
413 | repeat_lock_task: | |
414 | local_irq_save(*flags); | |
415 | rq = task_rq(p); | |
416 | spin_lock(&rq->lock); | |
417 | if (unlikely(rq != task_rq(p))) { | |
418 | spin_unlock_irqrestore(&rq->lock, *flags); | |
419 | goto repeat_lock_task; | |
420 | } | |
421 | return rq; | |
422 | } | |
423 | ||
70b97a7f | 424 | static inline void __task_rq_unlock(struct rq *rq) |
b29739f9 IM |
425 | __releases(rq->lock) |
426 | { | |
427 | spin_unlock(&rq->lock); | |
428 | } | |
429 | ||
70b97a7f | 430 | static inline void task_rq_unlock(struct rq *rq, unsigned long *flags) |
1da177e4 LT |
431 | __releases(rq->lock) |
432 | { | |
433 | spin_unlock_irqrestore(&rq->lock, *flags); | |
434 | } | |
435 | ||
436 | #ifdef CONFIG_SCHEDSTATS | |
437 | /* | |
438 | * bump this up when changing the output format or the meaning of an existing | |
439 | * format, so that tools can adapt (or abort) | |
440 | */ | |
06066714 | 441 | #define SCHEDSTAT_VERSION 14 |
1da177e4 LT |
442 | |
443 | static int show_schedstat(struct seq_file *seq, void *v) | |
444 | { | |
445 | int cpu; | |
446 | ||
447 | seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION); | |
448 | seq_printf(seq, "timestamp %lu\n", jiffies); | |
449 | for_each_online_cpu(cpu) { | |
70b97a7f | 450 | struct rq *rq = cpu_rq(cpu); |
1da177e4 LT |
451 | #ifdef CONFIG_SMP |
452 | struct sched_domain *sd; | |
453 | int dcnt = 0; | |
454 | #endif | |
455 | ||
456 | /* runqueue-specific stats */ | |
457 | seq_printf(seq, | |
458 | "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu", | |
459 | cpu, rq->yld_both_empty, | |
460 | rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt, | |
461 | rq->sched_switch, rq->sched_cnt, rq->sched_goidle, | |
462 | rq->ttwu_cnt, rq->ttwu_local, | |
463 | rq->rq_sched_info.cpu_time, | |
464 | rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt); | |
465 | ||
466 | seq_printf(seq, "\n"); | |
467 | ||
468 | #ifdef CONFIG_SMP | |
469 | /* domain-specific stats */ | |
674311d5 | 470 | preempt_disable(); |
1da177e4 LT |
471 | for_each_domain(cpu, sd) { |
472 | enum idle_type itype; | |
473 | char mask_str[NR_CPUS]; | |
474 | ||
475 | cpumask_scnprintf(mask_str, NR_CPUS, sd->span); | |
476 | seq_printf(seq, "domain%d %s", dcnt++, mask_str); | |
477 | for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES; | |
478 | itype++) { | |
33859f7f MOS |
479 | seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu " |
480 | "%lu", | |
1da177e4 LT |
481 | sd->lb_cnt[itype], |
482 | sd->lb_balanced[itype], | |
483 | sd->lb_failed[itype], | |
484 | sd->lb_imbalance[itype], | |
485 | sd->lb_gained[itype], | |
486 | sd->lb_hot_gained[itype], | |
487 | sd->lb_nobusyq[itype], | |
06066714 | 488 | sd->lb_nobusyg[itype]); |
1da177e4 | 489 | } |
33859f7f MOS |
490 | seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu" |
491 | " %lu %lu %lu\n", | |
1da177e4 | 492 | sd->alb_cnt, sd->alb_failed, sd->alb_pushed, |
68767a0a NP |
493 | sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed, |
494 | sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed, | |
33859f7f MOS |
495 | sd->ttwu_wake_remote, sd->ttwu_move_affine, |
496 | sd->ttwu_move_balance); | |
1da177e4 | 497 | } |
674311d5 | 498 | preempt_enable(); |
1da177e4 LT |
499 | #endif |
500 | } | |
501 | return 0; | |
502 | } | |
503 | ||
504 | static int schedstat_open(struct inode *inode, struct file *file) | |
505 | { | |
506 | unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32); | |
507 | char *buf = kmalloc(size, GFP_KERNEL); | |
508 | struct seq_file *m; | |
509 | int res; | |
510 | ||
511 | if (!buf) | |
512 | return -ENOMEM; | |
513 | res = single_open(file, show_schedstat, NULL); | |
514 | if (!res) { | |
515 | m = file->private_data; | |
516 | m->buf = buf; | |
517 | m->size = size; | |
518 | } else | |
519 | kfree(buf); | |
520 | return res; | |
521 | } | |
522 | ||
15ad7cdc | 523 | const struct file_operations proc_schedstat_operations = { |
1da177e4 LT |
524 | .open = schedstat_open, |
525 | .read = seq_read, | |
526 | .llseek = seq_lseek, | |
527 | .release = single_release, | |
528 | }; | |
529 | ||
52f17b6c CS |
530 | /* |
531 | * Expects runqueue lock to be held for atomicity of update | |
532 | */ | |
533 | static inline void | |
534 | rq_sched_info_arrive(struct rq *rq, unsigned long delta_jiffies) | |
535 | { | |
536 | if (rq) { | |
537 | rq->rq_sched_info.run_delay += delta_jiffies; | |
538 | rq->rq_sched_info.pcnt++; | |
539 | } | |
540 | } | |
541 | ||
542 | /* | |
543 | * Expects runqueue lock to be held for atomicity of update | |
544 | */ | |
545 | static inline void | |
546 | rq_sched_info_depart(struct rq *rq, unsigned long delta_jiffies) | |
547 | { | |
548 | if (rq) | |
549 | rq->rq_sched_info.cpu_time += delta_jiffies; | |
550 | } | |
1da177e4 LT |
551 | # define schedstat_inc(rq, field) do { (rq)->field++; } while (0) |
552 | # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0) | |
553 | #else /* !CONFIG_SCHEDSTATS */ | |
52f17b6c CS |
554 | static inline void |
555 | rq_sched_info_arrive(struct rq *rq, unsigned long delta_jiffies) | |
556 | {} | |
557 | static inline void | |
558 | rq_sched_info_depart(struct rq *rq, unsigned long delta_jiffies) | |
559 | {} | |
1da177e4 LT |
560 | # define schedstat_inc(rq, field) do { } while (0) |
561 | # define schedstat_add(rq, field, amt) do { } while (0) | |
562 | #endif | |
563 | ||
564 | /* | |
cc2a73b5 | 565 | * this_rq_lock - lock this runqueue and disable interrupts. |
1da177e4 | 566 | */ |
70b97a7f | 567 | static inline struct rq *this_rq_lock(void) |
1da177e4 LT |
568 | __acquires(rq->lock) |
569 | { | |
70b97a7f | 570 | struct rq *rq; |
1da177e4 LT |
571 | |
572 | local_irq_disable(); | |
573 | rq = this_rq(); | |
574 | spin_lock(&rq->lock); | |
575 | ||
576 | return rq; | |
577 | } | |
578 | ||
52f17b6c | 579 | #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) |
1da177e4 LT |
580 | /* |
581 | * Called when a process is dequeued from the active array and given | |
582 | * the cpu. We should note that with the exception of interactive | |
583 | * tasks, the expired queue will become the active queue after the active | |
584 | * queue is empty, without explicitly dequeuing and requeuing tasks in the | |
585 | * expired queue. (Interactive tasks may be requeued directly to the | |
586 | * active queue, thus delaying tasks in the expired queue from running; | |
587 | * see scheduler_tick()). | |
588 | * | |
589 | * This function is only called from sched_info_arrive(), rather than | |
590 | * dequeue_task(). Even though a task may be queued and dequeued multiple | |
591 | * times as it is shuffled about, we're really interested in knowing how | |
592 | * long it was from the *first* time it was queued to the time that it | |
593 | * finally hit a cpu. | |
594 | */ | |
36c8b586 | 595 | static inline void sched_info_dequeued(struct task_struct *t) |
1da177e4 LT |
596 | { |
597 | t->sched_info.last_queued = 0; | |
598 | } | |
599 | ||
600 | /* | |
601 | * Called when a task finally hits the cpu. We can now calculate how | |
602 | * long it was waiting to run. We also note when it began so that we | |
603 | * can keep stats on how long its timeslice is. | |
604 | */ | |
36c8b586 | 605 | static void sched_info_arrive(struct task_struct *t) |
1da177e4 | 606 | { |
52f17b6c | 607 | unsigned long now = jiffies, delta_jiffies = 0; |
1da177e4 LT |
608 | |
609 | if (t->sched_info.last_queued) | |
52f17b6c | 610 | delta_jiffies = now - t->sched_info.last_queued; |
1da177e4 | 611 | sched_info_dequeued(t); |
52f17b6c | 612 | t->sched_info.run_delay += delta_jiffies; |
1da177e4 LT |
613 | t->sched_info.last_arrival = now; |
614 | t->sched_info.pcnt++; | |
615 | ||
52f17b6c | 616 | rq_sched_info_arrive(task_rq(t), delta_jiffies); |
1da177e4 LT |
617 | } |
618 | ||
619 | /* | |
620 | * Called when a process is queued into either the active or expired | |
621 | * array. The time is noted and later used to determine how long we | |
622 | * had to wait for us to reach the cpu. Since the expired queue will | |
623 | * become the active queue after active queue is empty, without dequeuing | |
624 | * and requeuing any tasks, we are interested in queuing to either. It | |
625 | * is unusual but not impossible for tasks to be dequeued and immediately | |
626 | * requeued in the same or another array: this can happen in sched_yield(), | |
627 | * set_user_nice(), and even load_balance() as it moves tasks from runqueue | |
628 | * to runqueue. | |
629 | * | |
630 | * This function is only called from enqueue_task(), but also only updates | |
631 | * the timestamp if it is already not set. It's assumed that | |
632 | * sched_info_dequeued() will clear that stamp when appropriate. | |
633 | */ | |
36c8b586 | 634 | static inline void sched_info_queued(struct task_struct *t) |
1da177e4 | 635 | { |
52f17b6c CS |
636 | if (unlikely(sched_info_on())) |
637 | if (!t->sched_info.last_queued) | |
638 | t->sched_info.last_queued = jiffies; | |
1da177e4 LT |
639 | } |
640 | ||
641 | /* | |
642 | * Called when a process ceases being the active-running process, either | |
643 | * voluntarily or involuntarily. Now we can calculate how long we ran. | |
644 | */ | |
36c8b586 | 645 | static inline void sched_info_depart(struct task_struct *t) |
1da177e4 | 646 | { |
52f17b6c | 647 | unsigned long delta_jiffies = jiffies - t->sched_info.last_arrival; |
1da177e4 | 648 | |
52f17b6c CS |
649 | t->sched_info.cpu_time += delta_jiffies; |
650 | rq_sched_info_depart(task_rq(t), delta_jiffies); | |
1da177e4 LT |
651 | } |
652 | ||
653 | /* | |
654 | * Called when tasks are switched involuntarily due, typically, to expiring | |
655 | * their time slice. (This may also be called when switching to or from | |
656 | * the idle task.) We are only called when prev != next. | |
657 | */ | |
36c8b586 | 658 | static inline void |
52f17b6c | 659 | __sched_info_switch(struct task_struct *prev, struct task_struct *next) |
1da177e4 | 660 | { |
70b97a7f | 661 | struct rq *rq = task_rq(prev); |
1da177e4 LT |
662 | |
663 | /* | |
664 | * prev now departs the cpu. It's not interesting to record | |
665 | * stats about how efficient we were at scheduling the idle | |
666 | * process, however. | |
667 | */ | |
668 | if (prev != rq->idle) | |
669 | sched_info_depart(prev); | |
670 | ||
671 | if (next != rq->idle) | |
672 | sched_info_arrive(next); | |
673 | } | |
52f17b6c CS |
674 | static inline void |
675 | sched_info_switch(struct task_struct *prev, struct task_struct *next) | |
676 | { | |
677 | if (unlikely(sched_info_on())) | |
678 | __sched_info_switch(prev, next); | |
679 | } | |
1da177e4 LT |
680 | #else |
681 | #define sched_info_queued(t) do { } while (0) | |
682 | #define sched_info_switch(t, next) do { } while (0) | |
52f17b6c | 683 | #endif /* CONFIG_SCHEDSTATS || CONFIG_TASK_DELAY_ACCT */ |
1da177e4 LT |
684 | |
685 | /* | |
686 | * Adding/removing a task to/from a priority array: | |
687 | */ | |
70b97a7f | 688 | static void dequeue_task(struct task_struct *p, struct prio_array *array) |
1da177e4 LT |
689 | { |
690 | array->nr_active--; | |
691 | list_del(&p->run_list); | |
692 | if (list_empty(array->queue + p->prio)) | |
693 | __clear_bit(p->prio, array->bitmap); | |
694 | } | |
695 | ||
70b97a7f | 696 | static void enqueue_task(struct task_struct *p, struct prio_array *array) |
1da177e4 LT |
697 | { |
698 | sched_info_queued(p); | |
699 | list_add_tail(&p->run_list, array->queue + p->prio); | |
700 | __set_bit(p->prio, array->bitmap); | |
701 | array->nr_active++; | |
702 | p->array = array; | |
703 | } | |
704 | ||
705 | /* | |
706 | * Put task to the end of the run list without the overhead of dequeue | |
707 | * followed by enqueue. | |
708 | */ | |
70b97a7f | 709 | static void requeue_task(struct task_struct *p, struct prio_array *array) |
1da177e4 LT |
710 | { |
711 | list_move_tail(&p->run_list, array->queue + p->prio); | |
712 | } | |
713 | ||
70b97a7f IM |
714 | static inline void |
715 | enqueue_task_head(struct task_struct *p, struct prio_array *array) | |
1da177e4 LT |
716 | { |
717 | list_add(&p->run_list, array->queue + p->prio); | |
718 | __set_bit(p->prio, array->bitmap); | |
719 | array->nr_active++; | |
720 | p->array = array; | |
721 | } | |
722 | ||
723 | /* | |
b29739f9 | 724 | * __normal_prio - return the priority that is based on the static |
1da177e4 LT |
725 | * priority but is modified by bonuses/penalties. |
726 | * | |
727 | * We scale the actual sleep average [0 .... MAX_SLEEP_AVG] | |
728 | * into the -5 ... 0 ... +5 bonus/penalty range. | |
729 | * | |
730 | * We use 25% of the full 0...39 priority range so that: | |
731 | * | |
732 | * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs. | |
733 | * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks. | |
734 | * | |
735 | * Both properties are important to certain workloads. | |
736 | */ | |
b29739f9 | 737 | |
36c8b586 | 738 | static inline int __normal_prio(struct task_struct *p) |
1da177e4 LT |
739 | { |
740 | int bonus, prio; | |
741 | ||
1da177e4 LT |
742 | bonus = CURRENT_BONUS(p) - MAX_BONUS / 2; |
743 | ||
744 | prio = p->static_prio - bonus; | |
745 | if (prio < MAX_RT_PRIO) | |
746 | prio = MAX_RT_PRIO; | |
747 | if (prio > MAX_PRIO-1) | |
748 | prio = MAX_PRIO-1; | |
749 | return prio; | |
750 | } | |
751 | ||
2dd73a4f PW |
752 | /* |
753 | * To aid in avoiding the subversion of "niceness" due to uneven distribution | |
754 | * of tasks with abnormal "nice" values across CPUs the contribution that | |
755 | * each task makes to its run queue's load is weighted according to its | |
756 | * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a | |
757 | * scaled version of the new time slice allocation that they receive on time | |
758 | * slice expiry etc. | |
759 | */ | |
760 | ||
761 | /* | |
762 | * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE | |
763 | * If static_prio_timeslice() is ever changed to break this assumption then | |
764 | * this code will need modification | |
765 | */ | |
766 | #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE | |
767 | #define LOAD_WEIGHT(lp) \ | |
768 | (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO) | |
769 | #define PRIO_TO_LOAD_WEIGHT(prio) \ | |
770 | LOAD_WEIGHT(static_prio_timeslice(prio)) | |
771 | #define RTPRIO_TO_LOAD_WEIGHT(rp) \ | |
772 | (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp)) | |
773 | ||
36c8b586 | 774 | static void set_load_weight(struct task_struct *p) |
2dd73a4f | 775 | { |
b29739f9 | 776 | if (has_rt_policy(p)) { |
2dd73a4f PW |
777 | #ifdef CONFIG_SMP |
778 | if (p == task_rq(p)->migration_thread) | |
779 | /* | |
780 | * The migration thread does the actual balancing. | |
781 | * Giving its load any weight will skew balancing | |
782 | * adversely. | |
783 | */ | |
784 | p->load_weight = 0; | |
785 | else | |
786 | #endif | |
787 | p->load_weight = RTPRIO_TO_LOAD_WEIGHT(p->rt_priority); | |
788 | } else | |
789 | p->load_weight = PRIO_TO_LOAD_WEIGHT(p->static_prio); | |
790 | } | |
791 | ||
36c8b586 | 792 | static inline void |
70b97a7f | 793 | inc_raw_weighted_load(struct rq *rq, const struct task_struct *p) |
2dd73a4f PW |
794 | { |
795 | rq->raw_weighted_load += p->load_weight; | |
796 | } | |
797 | ||
36c8b586 | 798 | static inline void |
70b97a7f | 799 | dec_raw_weighted_load(struct rq *rq, const struct task_struct *p) |
2dd73a4f PW |
800 | { |
801 | rq->raw_weighted_load -= p->load_weight; | |
802 | } | |
803 | ||
70b97a7f | 804 | static inline void inc_nr_running(struct task_struct *p, struct rq *rq) |
2dd73a4f PW |
805 | { |
806 | rq->nr_running++; | |
807 | inc_raw_weighted_load(rq, p); | |
808 | } | |
809 | ||
70b97a7f | 810 | static inline void dec_nr_running(struct task_struct *p, struct rq *rq) |
2dd73a4f PW |
811 | { |
812 | rq->nr_running--; | |
813 | dec_raw_weighted_load(rq, p); | |
814 | } | |
815 | ||
b29739f9 IM |
816 | /* |
817 | * Calculate the expected normal priority: i.e. priority | |
818 | * without taking RT-inheritance into account. Might be | |
819 | * boosted by interactivity modifiers. Changes upon fork, | |
820 | * setprio syscalls, and whenever the interactivity | |
821 | * estimator recalculates. | |
822 | */ | |
36c8b586 | 823 | static inline int normal_prio(struct task_struct *p) |
b29739f9 IM |
824 | { |
825 | int prio; | |
826 | ||
827 | if (has_rt_policy(p)) | |
828 | prio = MAX_RT_PRIO-1 - p->rt_priority; | |
829 | else | |
830 | prio = __normal_prio(p); | |
831 | return prio; | |
832 | } | |
833 | ||
834 | /* | |
835 | * Calculate the current priority, i.e. the priority | |
836 | * taken into account by the scheduler. This value might | |
837 | * be boosted by RT tasks, or might be boosted by | |
838 | * interactivity modifiers. Will be RT if the task got | |
839 | * RT-boosted. If not then it returns p->normal_prio. | |
840 | */ | |
36c8b586 | 841 | static int effective_prio(struct task_struct *p) |
b29739f9 IM |
842 | { |
843 | p->normal_prio = normal_prio(p); | |
844 | /* | |
845 | * If we are RT tasks or we were boosted to RT priority, | |
846 | * keep the priority unchanged. Otherwise, update priority | |
847 | * to the normal priority: | |
848 | */ | |
849 | if (!rt_prio(p->prio)) | |
850 | return p->normal_prio; | |
851 | return p->prio; | |
852 | } | |
853 | ||
1da177e4 LT |
854 | /* |
855 | * __activate_task - move a task to the runqueue. | |
856 | */ | |
70b97a7f | 857 | static void __activate_task(struct task_struct *p, struct rq *rq) |
1da177e4 | 858 | { |
70b97a7f | 859 | struct prio_array *target = rq->active; |
d425b274 | 860 | |
f1adad78 | 861 | if (batch_task(p)) |
d425b274 CK |
862 | target = rq->expired; |
863 | enqueue_task(p, target); | |
2dd73a4f | 864 | inc_nr_running(p, rq); |
1da177e4 LT |
865 | } |
866 | ||
867 | /* | |
868 | * __activate_idle_task - move idle task to the _front_ of runqueue. | |
869 | */ | |
70b97a7f | 870 | static inline void __activate_idle_task(struct task_struct *p, struct rq *rq) |
1da177e4 LT |
871 | { |
872 | enqueue_task_head(p, rq->active); | |
2dd73a4f | 873 | inc_nr_running(p, rq); |
1da177e4 LT |
874 | } |
875 | ||
b29739f9 IM |
876 | /* |
877 | * Recalculate p->normal_prio and p->prio after having slept, | |
878 | * updating the sleep-average too: | |
879 | */ | |
36c8b586 | 880 | static int recalc_task_prio(struct task_struct *p, unsigned long long now) |
1da177e4 LT |
881 | { |
882 | /* Caller must always ensure 'now >= p->timestamp' */ | |
72d2854d | 883 | unsigned long sleep_time = now - p->timestamp; |
1da177e4 | 884 | |
d425b274 | 885 | if (batch_task(p)) |
b0a9499c | 886 | sleep_time = 0; |
1da177e4 LT |
887 | |
888 | if (likely(sleep_time > 0)) { | |
889 | /* | |
72d2854d CK |
890 | * This ceiling is set to the lowest priority that would allow |
891 | * a task to be reinserted into the active array on timeslice | |
892 | * completion. | |
1da177e4 | 893 | */ |
72d2854d | 894 | unsigned long ceiling = INTERACTIVE_SLEEP(p); |
e72ff0bb | 895 | |
72d2854d CK |
896 | if (p->mm && sleep_time > ceiling && p->sleep_avg < ceiling) { |
897 | /* | |
898 | * Prevents user tasks from achieving best priority | |
899 | * with one single large enough sleep. | |
900 | */ | |
901 | p->sleep_avg = ceiling; | |
902 | /* | |
903 | * Using INTERACTIVE_SLEEP() as a ceiling places a | |
904 | * nice(0) task 1ms sleep away from promotion, and | |
905 | * gives it 700ms to round-robin with no chance of | |
906 | * being demoted. This is more than generous, so | |
907 | * mark this sleep as non-interactive to prevent the | |
908 | * on-runqueue bonus logic from intervening should | |
909 | * this task not receive cpu immediately. | |
910 | */ | |
911 | p->sleep_type = SLEEP_NONINTERACTIVE; | |
1da177e4 | 912 | } else { |
1da177e4 LT |
913 | /* |
914 | * Tasks waking from uninterruptible sleep are | |
915 | * limited in their sleep_avg rise as they | |
916 | * are likely to be waiting on I/O | |
917 | */ | |
3dee386e | 918 | if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) { |
72d2854d | 919 | if (p->sleep_avg >= ceiling) |
1da177e4 LT |
920 | sleep_time = 0; |
921 | else if (p->sleep_avg + sleep_time >= | |
72d2854d CK |
922 | ceiling) { |
923 | p->sleep_avg = ceiling; | |
924 | sleep_time = 0; | |
1da177e4 LT |
925 | } |
926 | } | |
927 | ||
928 | /* | |
929 | * This code gives a bonus to interactive tasks. | |
930 | * | |
931 | * The boost works by updating the 'average sleep time' | |
932 | * value here, based on ->timestamp. The more time a | |
933 | * task spends sleeping, the higher the average gets - | |
934 | * and the higher the priority boost gets as well. | |
935 | */ | |
936 | p->sleep_avg += sleep_time; | |
937 | ||
1da177e4 | 938 | } |
72d2854d CK |
939 | if (p->sleep_avg > NS_MAX_SLEEP_AVG) |
940 | p->sleep_avg = NS_MAX_SLEEP_AVG; | |
1da177e4 LT |
941 | } |
942 | ||
a3464a10 | 943 | return effective_prio(p); |
1da177e4 LT |
944 | } |
945 | ||
946 | /* | |
947 | * activate_task - move a task to the runqueue and do priority recalculation | |
948 | * | |
949 | * Update all the scheduling statistics stuff. (sleep average | |
950 | * calculation, priority modifiers, etc.) | |
951 | */ | |
70b97a7f | 952 | static void activate_task(struct task_struct *p, struct rq *rq, int local) |
1da177e4 LT |
953 | { |
954 | unsigned long long now; | |
955 | ||
62ab616d KC |
956 | if (rt_task(p)) |
957 | goto out; | |
958 | ||
1da177e4 LT |
959 | now = sched_clock(); |
960 | #ifdef CONFIG_SMP | |
961 | if (!local) { | |
962 | /* Compensate for drifting sched_clock */ | |
70b97a7f | 963 | struct rq *this_rq = this_rq(); |
b18ec803 MG |
964 | now = (now - this_rq->most_recent_timestamp) |
965 | + rq->most_recent_timestamp; | |
1da177e4 LT |
966 | } |
967 | #endif | |
968 | ||
ece8a684 IM |
969 | /* |
970 | * Sleep time is in units of nanosecs, so shift by 20 to get a | |
971 | * milliseconds-range estimation of the amount of time that the task | |
972 | * spent sleeping: | |
973 | */ | |
974 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
975 | if (p->state == TASK_UNINTERRUPTIBLE) | |
976 | profile_hits(SLEEP_PROFILING, (void *)get_wchan(p), | |
977 | (now - p->timestamp) >> 20); | |
978 | } | |
979 | ||
62ab616d | 980 | p->prio = recalc_task_prio(p, now); |
1da177e4 LT |
981 | |
982 | /* | |
983 | * This checks to make sure it's not an uninterruptible task | |
984 | * that is now waking up. | |
985 | */ | |
3dee386e | 986 | if (p->sleep_type == SLEEP_NORMAL) { |
1da177e4 LT |
987 | /* |
988 | * Tasks which were woken up by interrupts (ie. hw events) | |
989 | * are most likely of interactive nature. So we give them | |
990 | * the credit of extending their sleep time to the period | |
991 | * of time they spend on the runqueue, waiting for execution | |
992 | * on a CPU, first time around: | |
993 | */ | |
994 | if (in_interrupt()) | |
3dee386e | 995 | p->sleep_type = SLEEP_INTERRUPTED; |
1da177e4 LT |
996 | else { |
997 | /* | |
998 | * Normal first-time wakeups get a credit too for | |
999 | * on-runqueue time, but it will be weighted down: | |
1000 | */ | |
3dee386e | 1001 | p->sleep_type = SLEEP_INTERACTIVE; |
1da177e4 LT |
1002 | } |
1003 | } | |
1004 | p->timestamp = now; | |
62ab616d | 1005 | out: |
1da177e4 LT |
1006 | __activate_task(p, rq); |
1007 | } | |
1008 | ||
1009 | /* | |
1010 | * deactivate_task - remove a task from the runqueue. | |
1011 | */ | |
70b97a7f | 1012 | static void deactivate_task(struct task_struct *p, struct rq *rq) |
1da177e4 | 1013 | { |
2dd73a4f | 1014 | dec_nr_running(p, rq); |
1da177e4 LT |
1015 | dequeue_task(p, p->array); |
1016 | p->array = NULL; | |
1017 | } | |
1018 | ||
1019 | /* | |
1020 | * resched_task - mark a task 'to be rescheduled now'. | |
1021 | * | |
1022 | * On UP this means the setting of the need_resched flag, on SMP it | |
1023 | * might also involve a cross-CPU call to trigger the scheduler on | |
1024 | * the target CPU. | |
1025 | */ | |
1026 | #ifdef CONFIG_SMP | |
495ab9c0 AK |
1027 | |
1028 | #ifndef tsk_is_polling | |
1029 | #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG) | |
1030 | #endif | |
1031 | ||
36c8b586 | 1032 | static void resched_task(struct task_struct *p) |
1da177e4 | 1033 | { |
64c7c8f8 | 1034 | int cpu; |
1da177e4 LT |
1035 | |
1036 | assert_spin_locked(&task_rq(p)->lock); | |
1037 | ||
64c7c8f8 NP |
1038 | if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED))) |
1039 | return; | |
1040 | ||
1041 | set_tsk_thread_flag(p, TIF_NEED_RESCHED); | |
1da177e4 | 1042 | |
64c7c8f8 NP |
1043 | cpu = task_cpu(p); |
1044 | if (cpu == smp_processor_id()) | |
1045 | return; | |
1046 | ||
495ab9c0 | 1047 | /* NEED_RESCHED must be visible before we test polling */ |
64c7c8f8 | 1048 | smp_mb(); |
495ab9c0 | 1049 | if (!tsk_is_polling(p)) |
64c7c8f8 | 1050 | smp_send_reschedule(cpu); |
1da177e4 LT |
1051 | } |
1052 | #else | |
36c8b586 | 1053 | static inline void resched_task(struct task_struct *p) |
1da177e4 | 1054 | { |
64c7c8f8 | 1055 | assert_spin_locked(&task_rq(p)->lock); |
1da177e4 LT |
1056 | set_tsk_need_resched(p); |
1057 | } | |
1058 | #endif | |
1059 | ||
1060 | /** | |
1061 | * task_curr - is this task currently executing on a CPU? | |
1062 | * @p: the task in question. | |
1063 | */ | |
36c8b586 | 1064 | inline int task_curr(const struct task_struct *p) |
1da177e4 LT |
1065 | { |
1066 | return cpu_curr(task_cpu(p)) == p; | |
1067 | } | |
1068 | ||
2dd73a4f PW |
1069 | /* Used instead of source_load when we know the type == 0 */ |
1070 | unsigned long weighted_cpuload(const int cpu) | |
1071 | { | |
1072 | return cpu_rq(cpu)->raw_weighted_load; | |
1073 | } | |
1074 | ||
1da177e4 | 1075 | #ifdef CONFIG_SMP |
70b97a7f | 1076 | struct migration_req { |
1da177e4 | 1077 | struct list_head list; |
1da177e4 | 1078 | |
36c8b586 | 1079 | struct task_struct *task; |
1da177e4 LT |
1080 | int dest_cpu; |
1081 | ||
1da177e4 | 1082 | struct completion done; |
70b97a7f | 1083 | }; |
1da177e4 LT |
1084 | |
1085 | /* | |
1086 | * The task's runqueue lock must be held. | |
1087 | * Returns true if you have to wait for migration thread. | |
1088 | */ | |
36c8b586 | 1089 | static int |
70b97a7f | 1090 | migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req) |
1da177e4 | 1091 | { |
70b97a7f | 1092 | struct rq *rq = task_rq(p); |
1da177e4 LT |
1093 | |
1094 | /* | |
1095 | * If the task is not on a runqueue (and not running), then | |
1096 | * it is sufficient to simply update the task's cpu field. | |
1097 | */ | |
1098 | if (!p->array && !task_running(rq, p)) { | |
1099 | set_task_cpu(p, dest_cpu); | |
1100 | return 0; | |
1101 | } | |
1102 | ||
1103 | init_completion(&req->done); | |
1da177e4 LT |
1104 | req->task = p; |
1105 | req->dest_cpu = dest_cpu; | |
1106 | list_add(&req->list, &rq->migration_queue); | |
48f24c4d | 1107 | |
1da177e4 LT |
1108 | return 1; |
1109 | } | |
1110 | ||
1111 | /* | |
1112 | * wait_task_inactive - wait for a thread to unschedule. | |
1113 | * | |
1114 | * The caller must ensure that the task *will* unschedule sometime soon, | |
1115 | * else this function might spin for a *long* time. This function can't | |
1116 | * be called with interrupts off, or it may introduce deadlock with | |
1117 | * smp_call_function() if an IPI is sent by the same process we are | |
1118 | * waiting to become inactive. | |
1119 | */ | |
36c8b586 | 1120 | void wait_task_inactive(struct task_struct *p) |
1da177e4 LT |
1121 | { |
1122 | unsigned long flags; | |
70b97a7f | 1123 | struct rq *rq; |
1da177e4 LT |
1124 | int preempted; |
1125 | ||
1126 | repeat: | |
1127 | rq = task_rq_lock(p, &flags); | |
1128 | /* Must be off runqueue entirely, not preempted. */ | |
1129 | if (unlikely(p->array || task_running(rq, p))) { | |
1130 | /* If it's preempted, we yield. It could be a while. */ | |
1131 | preempted = !task_running(rq, p); | |
1132 | task_rq_unlock(rq, &flags); | |
1133 | cpu_relax(); | |
1134 | if (preempted) | |
1135 | yield(); | |
1136 | goto repeat; | |
1137 | } | |
1138 | task_rq_unlock(rq, &flags); | |
1139 | } | |
1140 | ||
1141 | /*** | |
1142 | * kick_process - kick a running thread to enter/exit the kernel | |
1143 | * @p: the to-be-kicked thread | |
1144 | * | |
1145 | * Cause a process which is running on another CPU to enter | |
1146 | * kernel-mode, without any delay. (to get signals handled.) | |
1147 | * | |
1148 | * NOTE: this function doesnt have to take the runqueue lock, | |
1149 | * because all it wants to ensure is that the remote task enters | |
1150 | * the kernel. If the IPI races and the task has been migrated | |
1151 | * to another CPU then no harm is done and the purpose has been | |
1152 | * achieved as well. | |
1153 | */ | |
36c8b586 | 1154 | void kick_process(struct task_struct *p) |
1da177e4 LT |
1155 | { |
1156 | int cpu; | |
1157 | ||
1158 | preempt_disable(); | |
1159 | cpu = task_cpu(p); | |
1160 | if ((cpu != smp_processor_id()) && task_curr(p)) | |
1161 | smp_send_reschedule(cpu); | |
1162 | preempt_enable(); | |
1163 | } | |
1164 | ||
1165 | /* | |
2dd73a4f PW |
1166 | * Return a low guess at the load of a migration-source cpu weighted |
1167 | * according to the scheduling class and "nice" value. | |
1da177e4 LT |
1168 | * |
1169 | * We want to under-estimate the load of migration sources, to | |
1170 | * balance conservatively. | |
1171 | */ | |
a2000572 | 1172 | static inline unsigned long source_load(int cpu, int type) |
1da177e4 | 1173 | { |
70b97a7f | 1174 | struct rq *rq = cpu_rq(cpu); |
2dd73a4f | 1175 | |
3b0bd9bc | 1176 | if (type == 0) |
2dd73a4f | 1177 | return rq->raw_weighted_load; |
b910472d | 1178 | |
2dd73a4f | 1179 | return min(rq->cpu_load[type-1], rq->raw_weighted_load); |
1da177e4 LT |
1180 | } |
1181 | ||
1182 | /* | |
2dd73a4f PW |
1183 | * Return a high guess at the load of a migration-target cpu weighted |
1184 | * according to the scheduling class and "nice" value. | |
1da177e4 | 1185 | */ |
a2000572 | 1186 | static inline unsigned long target_load(int cpu, int type) |
1da177e4 | 1187 | { |
70b97a7f | 1188 | struct rq *rq = cpu_rq(cpu); |
2dd73a4f | 1189 | |
7897986b | 1190 | if (type == 0) |
2dd73a4f | 1191 | return rq->raw_weighted_load; |
3b0bd9bc | 1192 | |
2dd73a4f PW |
1193 | return max(rq->cpu_load[type-1], rq->raw_weighted_load); |
1194 | } | |
1195 | ||
1196 | /* | |
1197 | * Return the average load per task on the cpu's run queue | |
1198 | */ | |
1199 | static inline unsigned long cpu_avg_load_per_task(int cpu) | |
1200 | { | |
70b97a7f | 1201 | struct rq *rq = cpu_rq(cpu); |
2dd73a4f PW |
1202 | unsigned long n = rq->nr_running; |
1203 | ||
48f24c4d | 1204 | return n ? rq->raw_weighted_load / n : SCHED_LOAD_SCALE; |
1da177e4 LT |
1205 | } |
1206 | ||
147cbb4b NP |
1207 | /* |
1208 | * find_idlest_group finds and returns the least busy CPU group within the | |
1209 | * domain. | |
1210 | */ | |
1211 | static struct sched_group * | |
1212 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu) | |
1213 | { | |
1214 | struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups; | |
1215 | unsigned long min_load = ULONG_MAX, this_load = 0; | |
1216 | int load_idx = sd->forkexec_idx; | |
1217 | int imbalance = 100 + (sd->imbalance_pct-100)/2; | |
1218 | ||
1219 | do { | |
1220 | unsigned long load, avg_load; | |
1221 | int local_group; | |
1222 | int i; | |
1223 | ||
da5a5522 BD |
1224 | /* Skip over this group if it has no CPUs allowed */ |
1225 | if (!cpus_intersects(group->cpumask, p->cpus_allowed)) | |
1226 | goto nextgroup; | |
1227 | ||
147cbb4b | 1228 | local_group = cpu_isset(this_cpu, group->cpumask); |
147cbb4b NP |
1229 | |
1230 | /* Tally up the load of all CPUs in the group */ | |
1231 | avg_load = 0; | |
1232 | ||
1233 | for_each_cpu_mask(i, group->cpumask) { | |
1234 | /* Bias balancing toward cpus of our domain */ | |
1235 | if (local_group) | |
1236 | load = source_load(i, load_idx); | |
1237 | else | |
1238 | load = target_load(i, load_idx); | |
1239 | ||
1240 | avg_load += load; | |
1241 | } | |
1242 | ||
1243 | /* Adjust by relative CPU power of the group */ | |
1244 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
1245 | ||
1246 | if (local_group) { | |
1247 | this_load = avg_load; | |
1248 | this = group; | |
1249 | } else if (avg_load < min_load) { | |
1250 | min_load = avg_load; | |
1251 | idlest = group; | |
1252 | } | |
da5a5522 | 1253 | nextgroup: |
147cbb4b NP |
1254 | group = group->next; |
1255 | } while (group != sd->groups); | |
1256 | ||
1257 | if (!idlest || 100*this_load < imbalance*min_load) | |
1258 | return NULL; | |
1259 | return idlest; | |
1260 | } | |
1261 | ||
1262 | /* | |
0feaece9 | 1263 | * find_idlest_cpu - find the idlest cpu among the cpus in group. |
147cbb4b | 1264 | */ |
95cdf3b7 IM |
1265 | static int |
1266 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
147cbb4b | 1267 | { |
da5a5522 | 1268 | cpumask_t tmp; |
147cbb4b NP |
1269 | unsigned long load, min_load = ULONG_MAX; |
1270 | int idlest = -1; | |
1271 | int i; | |
1272 | ||
da5a5522 BD |
1273 | /* Traverse only the allowed CPUs */ |
1274 | cpus_and(tmp, group->cpumask, p->cpus_allowed); | |
1275 | ||
1276 | for_each_cpu_mask(i, tmp) { | |
2dd73a4f | 1277 | load = weighted_cpuload(i); |
147cbb4b NP |
1278 | |
1279 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
1280 | min_load = load; | |
1281 | idlest = i; | |
1282 | } | |
1283 | } | |
1284 | ||
1285 | return idlest; | |
1286 | } | |
1287 | ||
476d139c NP |
1288 | /* |
1289 | * sched_balance_self: balance the current task (running on cpu) in domains | |
1290 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
1291 | * SD_BALANCE_EXEC. | |
1292 | * | |
1293 | * Balance, ie. select the least loaded group. | |
1294 | * | |
1295 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
1296 | * | |
1297 | * preempt must be disabled. | |
1298 | */ | |
1299 | static int sched_balance_self(int cpu, int flag) | |
1300 | { | |
1301 | struct task_struct *t = current; | |
1302 | struct sched_domain *tmp, *sd = NULL; | |
147cbb4b | 1303 | |
c96d145e | 1304 | for_each_domain(cpu, tmp) { |
5c45bf27 SS |
1305 | /* |
1306 | * If power savings logic is enabled for a domain, stop there. | |
1307 | */ | |
1308 | if (tmp->flags & SD_POWERSAVINGS_BALANCE) | |
1309 | break; | |
476d139c NP |
1310 | if (tmp->flags & flag) |
1311 | sd = tmp; | |
c96d145e | 1312 | } |
476d139c NP |
1313 | |
1314 | while (sd) { | |
1315 | cpumask_t span; | |
1316 | struct sched_group *group; | |
1a848870 SS |
1317 | int new_cpu, weight; |
1318 | ||
1319 | if (!(sd->flags & flag)) { | |
1320 | sd = sd->child; | |
1321 | continue; | |
1322 | } | |
476d139c NP |
1323 | |
1324 | span = sd->span; | |
1325 | group = find_idlest_group(sd, t, cpu); | |
1a848870 SS |
1326 | if (!group) { |
1327 | sd = sd->child; | |
1328 | continue; | |
1329 | } | |
476d139c | 1330 | |
da5a5522 | 1331 | new_cpu = find_idlest_cpu(group, t, cpu); |
1a848870 SS |
1332 | if (new_cpu == -1 || new_cpu == cpu) { |
1333 | /* Now try balancing at a lower domain level of cpu */ | |
1334 | sd = sd->child; | |
1335 | continue; | |
1336 | } | |
476d139c | 1337 | |
1a848870 | 1338 | /* Now try balancing at a lower domain level of new_cpu */ |
476d139c | 1339 | cpu = new_cpu; |
476d139c NP |
1340 | sd = NULL; |
1341 | weight = cpus_weight(span); | |
1342 | for_each_domain(cpu, tmp) { | |
1343 | if (weight <= cpus_weight(tmp->span)) | |
1344 | break; | |
1345 | if (tmp->flags & flag) | |
1346 | sd = tmp; | |
1347 | } | |
1348 | /* while loop will break here if sd == NULL */ | |
1349 | } | |
1350 | ||
1351 | return cpu; | |
1352 | } | |
1353 | ||
1354 | #endif /* CONFIG_SMP */ | |
1da177e4 LT |
1355 | |
1356 | /* | |
1357 | * wake_idle() will wake a task on an idle cpu if task->cpu is | |
1358 | * not idle and an idle cpu is available. The span of cpus to | |
1359 | * search starts with cpus closest then further out as needed, | |
1360 | * so we always favor a closer, idle cpu. | |
1361 | * | |
1362 | * Returns the CPU we should wake onto. | |
1363 | */ | |
1364 | #if defined(ARCH_HAS_SCHED_WAKE_IDLE) | |
36c8b586 | 1365 | static int wake_idle(int cpu, struct task_struct *p) |
1da177e4 LT |
1366 | { |
1367 | cpumask_t tmp; | |
1368 | struct sched_domain *sd; | |
1369 | int i; | |
1370 | ||
1371 | if (idle_cpu(cpu)) | |
1372 | return cpu; | |
1373 | ||
1374 | for_each_domain(cpu, sd) { | |
1375 | if (sd->flags & SD_WAKE_IDLE) { | |
e0f364f4 | 1376 | cpus_and(tmp, sd->span, p->cpus_allowed); |
1da177e4 LT |
1377 | for_each_cpu_mask(i, tmp) { |
1378 | if (idle_cpu(i)) | |
1379 | return i; | |
1380 | } | |
1381 | } | |
e0f364f4 NP |
1382 | else |
1383 | break; | |
1da177e4 LT |
1384 | } |
1385 | return cpu; | |
1386 | } | |
1387 | #else | |
36c8b586 | 1388 | static inline int wake_idle(int cpu, struct task_struct *p) |
1da177e4 LT |
1389 | { |
1390 | return cpu; | |
1391 | } | |
1392 | #endif | |
1393 | ||
1394 | /*** | |
1395 | * try_to_wake_up - wake up a thread | |
1396 | * @p: the to-be-woken-up thread | |
1397 | * @state: the mask of task states that can be woken | |
1398 | * @sync: do a synchronous wakeup? | |
1399 | * | |
1400 | * Put it on the run-queue if it's not already there. The "current" | |
1401 | * thread is always on the run-queue (except when the actual | |
1402 | * re-schedule is in progress), and as such you're allowed to do | |
1403 | * the simpler "current->state = TASK_RUNNING" to mark yourself | |
1404 | * runnable without the overhead of this. | |
1405 | * | |
1406 | * returns failure only if the task is already active. | |
1407 | */ | |
36c8b586 | 1408 | static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync) |
1da177e4 LT |
1409 | { |
1410 | int cpu, this_cpu, success = 0; | |
1411 | unsigned long flags; | |
1412 | long old_state; | |
70b97a7f | 1413 | struct rq *rq; |
1da177e4 | 1414 | #ifdef CONFIG_SMP |
7897986b | 1415 | struct sched_domain *sd, *this_sd = NULL; |
70b97a7f | 1416 | unsigned long load, this_load; |
1da177e4 LT |
1417 | int new_cpu; |
1418 | #endif | |
1419 | ||
1420 | rq = task_rq_lock(p, &flags); | |
1421 | old_state = p->state; | |
1422 | if (!(old_state & state)) | |
1423 | goto out; | |
1424 | ||
1425 | if (p->array) | |
1426 | goto out_running; | |
1427 | ||
1428 | cpu = task_cpu(p); | |
1429 | this_cpu = smp_processor_id(); | |
1430 | ||
1431 | #ifdef CONFIG_SMP | |
1432 | if (unlikely(task_running(rq, p))) | |
1433 | goto out_activate; | |
1434 | ||
7897986b NP |
1435 | new_cpu = cpu; |
1436 | ||
1da177e4 LT |
1437 | schedstat_inc(rq, ttwu_cnt); |
1438 | if (cpu == this_cpu) { | |
1439 | schedstat_inc(rq, ttwu_local); | |
7897986b NP |
1440 | goto out_set_cpu; |
1441 | } | |
1442 | ||
1443 | for_each_domain(this_cpu, sd) { | |
1444 | if (cpu_isset(cpu, sd->span)) { | |
1445 | schedstat_inc(sd, ttwu_wake_remote); | |
1446 | this_sd = sd; | |
1447 | break; | |
1da177e4 LT |
1448 | } |
1449 | } | |
1da177e4 | 1450 | |
7897986b | 1451 | if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed))) |
1da177e4 LT |
1452 | goto out_set_cpu; |
1453 | ||
1da177e4 | 1454 | /* |
7897986b | 1455 | * Check for affine wakeup and passive balancing possibilities. |
1da177e4 | 1456 | */ |
7897986b NP |
1457 | if (this_sd) { |
1458 | int idx = this_sd->wake_idx; | |
1459 | unsigned int imbalance; | |
1da177e4 | 1460 | |
a3f21bce NP |
1461 | imbalance = 100 + (this_sd->imbalance_pct - 100) / 2; |
1462 | ||
7897986b NP |
1463 | load = source_load(cpu, idx); |
1464 | this_load = target_load(this_cpu, idx); | |
1da177e4 | 1465 | |
7897986b NP |
1466 | new_cpu = this_cpu; /* Wake to this CPU if we can */ |
1467 | ||
a3f21bce NP |
1468 | if (this_sd->flags & SD_WAKE_AFFINE) { |
1469 | unsigned long tl = this_load; | |
33859f7f MOS |
1470 | unsigned long tl_per_task; |
1471 | ||
1472 | tl_per_task = cpu_avg_load_per_task(this_cpu); | |
2dd73a4f | 1473 | |
1da177e4 | 1474 | /* |
a3f21bce NP |
1475 | * If sync wakeup then subtract the (maximum possible) |
1476 | * effect of the currently running task from the load | |
1477 | * of the current CPU: | |
1da177e4 | 1478 | */ |
a3f21bce | 1479 | if (sync) |
2dd73a4f | 1480 | tl -= current->load_weight; |
a3f21bce NP |
1481 | |
1482 | if ((tl <= load && | |
2dd73a4f PW |
1483 | tl + target_load(cpu, idx) <= tl_per_task) || |
1484 | 100*(tl + p->load_weight) <= imbalance*load) { | |
a3f21bce NP |
1485 | /* |
1486 | * This domain has SD_WAKE_AFFINE and | |
1487 | * p is cache cold in this domain, and | |
1488 | * there is no bad imbalance. | |
1489 | */ | |
1490 | schedstat_inc(this_sd, ttwu_move_affine); | |
1491 | goto out_set_cpu; | |
1492 | } | |
1493 | } | |
1494 | ||
1495 | /* | |
1496 | * Start passive balancing when half the imbalance_pct | |
1497 | * limit is reached. | |
1498 | */ | |
1499 | if (this_sd->flags & SD_WAKE_BALANCE) { | |
1500 | if (imbalance*this_load <= 100*load) { | |
1501 | schedstat_inc(this_sd, ttwu_move_balance); | |
1502 | goto out_set_cpu; | |
1503 | } | |
1da177e4 LT |
1504 | } |
1505 | } | |
1506 | ||
1507 | new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */ | |
1508 | out_set_cpu: | |
1509 | new_cpu = wake_idle(new_cpu, p); | |
1510 | if (new_cpu != cpu) { | |
1511 | set_task_cpu(p, new_cpu); | |
1512 | task_rq_unlock(rq, &flags); | |
1513 | /* might preempt at this point */ | |
1514 | rq = task_rq_lock(p, &flags); | |
1515 | old_state = p->state; | |
1516 | if (!(old_state & state)) | |
1517 | goto out; | |
1518 | if (p->array) | |
1519 | goto out_running; | |
1520 | ||
1521 | this_cpu = smp_processor_id(); | |
1522 | cpu = task_cpu(p); | |
1523 | } | |
1524 | ||
1525 | out_activate: | |
1526 | #endif /* CONFIG_SMP */ | |
1527 | if (old_state == TASK_UNINTERRUPTIBLE) { | |
1528 | rq->nr_uninterruptible--; | |
1529 | /* | |
1530 | * Tasks on involuntary sleep don't earn | |
1531 | * sleep_avg beyond just interactive state. | |
1532 | */ | |
3dee386e | 1533 | p->sleep_type = SLEEP_NONINTERACTIVE; |
e7c38cb4 | 1534 | } else |
1da177e4 | 1535 | |
d79fc0fc IM |
1536 | /* |
1537 | * Tasks that have marked their sleep as noninteractive get | |
e7c38cb4 CK |
1538 | * woken up with their sleep average not weighted in an |
1539 | * interactive way. | |
d79fc0fc | 1540 | */ |
e7c38cb4 CK |
1541 | if (old_state & TASK_NONINTERACTIVE) |
1542 | p->sleep_type = SLEEP_NONINTERACTIVE; | |
1543 | ||
1544 | ||
1545 | activate_task(p, rq, cpu == this_cpu); | |
1da177e4 LT |
1546 | /* |
1547 | * Sync wakeups (i.e. those types of wakeups where the waker | |
1548 | * has indicated that it will leave the CPU in short order) | |
1549 | * don't trigger a preemption, if the woken up task will run on | |
1550 | * this cpu. (in this case the 'I will reschedule' promise of | |
1551 | * the waker guarantees that the freshly woken up task is going | |
1552 | * to be considered on this CPU.) | |
1553 | */ | |
1da177e4 LT |
1554 | if (!sync || cpu != this_cpu) { |
1555 | if (TASK_PREEMPTS_CURR(p, rq)) | |
1556 | resched_task(rq->curr); | |
1557 | } | |
1558 | success = 1; | |
1559 | ||
1560 | out_running: | |
1561 | p->state = TASK_RUNNING; | |
1562 | out: | |
1563 | task_rq_unlock(rq, &flags); | |
1564 | ||
1565 | return success; | |
1566 | } | |
1567 | ||
36c8b586 | 1568 | int fastcall wake_up_process(struct task_struct *p) |
1da177e4 LT |
1569 | { |
1570 | return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED | | |
1571 | TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0); | |
1572 | } | |
1da177e4 LT |
1573 | EXPORT_SYMBOL(wake_up_process); |
1574 | ||
36c8b586 | 1575 | int fastcall wake_up_state(struct task_struct *p, unsigned int state) |
1da177e4 LT |
1576 | { |
1577 | return try_to_wake_up(p, state, 0); | |
1578 | } | |
1579 | ||
bc947631 | 1580 | static void task_running_tick(struct rq *rq, struct task_struct *p); |
1da177e4 LT |
1581 | /* |
1582 | * Perform scheduler related setup for a newly forked process p. | |
1583 | * p is forked by current. | |
1584 | */ | |
36c8b586 | 1585 | void fastcall sched_fork(struct task_struct *p, int clone_flags) |
1da177e4 | 1586 | { |
476d139c NP |
1587 | int cpu = get_cpu(); |
1588 | ||
1589 | #ifdef CONFIG_SMP | |
1590 | cpu = sched_balance_self(cpu, SD_BALANCE_FORK); | |
1591 | #endif | |
1592 | set_task_cpu(p, cpu); | |
1593 | ||
1da177e4 LT |
1594 | /* |
1595 | * We mark the process as running here, but have not actually | |
1596 | * inserted it onto the runqueue yet. This guarantees that | |
1597 | * nobody will actually run it, and a signal or other external | |
1598 | * event cannot wake it up and insert it on the runqueue either. | |
1599 | */ | |
1600 | p->state = TASK_RUNNING; | |
b29739f9 IM |
1601 | |
1602 | /* | |
1603 | * Make sure we do not leak PI boosting priority to the child: | |
1604 | */ | |
1605 | p->prio = current->normal_prio; | |
1606 | ||
1da177e4 LT |
1607 | INIT_LIST_HEAD(&p->run_list); |
1608 | p->array = NULL; | |
52f17b6c CS |
1609 | #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) |
1610 | if (unlikely(sched_info_on())) | |
1611 | memset(&p->sched_info, 0, sizeof(p->sched_info)); | |
1da177e4 | 1612 | #endif |
d6077cb8 | 1613 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
4866cde0 NP |
1614 | p->oncpu = 0; |
1615 | #endif | |
1da177e4 | 1616 | #ifdef CONFIG_PREEMPT |
4866cde0 | 1617 | /* Want to start with kernel preemption disabled. */ |
a1261f54 | 1618 | task_thread_info(p)->preempt_count = 1; |
1da177e4 LT |
1619 | #endif |
1620 | /* | |
1621 | * Share the timeslice between parent and child, thus the | |
1622 | * total amount of pending timeslices in the system doesn't change, | |
1623 | * resulting in more scheduling fairness. | |
1624 | */ | |
1625 | local_irq_disable(); | |
1626 | p->time_slice = (current->time_slice + 1) >> 1; | |
1627 | /* | |
1628 | * The remainder of the first timeslice might be recovered by | |
1629 | * the parent if the child exits early enough. | |
1630 | */ | |
1631 | p->first_time_slice = 1; | |
1632 | current->time_slice >>= 1; | |
1633 | p->timestamp = sched_clock(); | |
1634 | if (unlikely(!current->time_slice)) { | |
1635 | /* | |
1636 | * This case is rare, it happens when the parent has only | |
1637 | * a single jiffy left from its timeslice. Taking the | |
1638 | * runqueue lock is not a problem. | |
1639 | */ | |
1640 | current->time_slice = 1; | |
bc947631 | 1641 | task_running_tick(cpu_rq(cpu), current); |
476d139c NP |
1642 | } |
1643 | local_irq_enable(); | |
1644 | put_cpu(); | |
1da177e4 LT |
1645 | } |
1646 | ||
1647 | /* | |
1648 | * wake_up_new_task - wake up a newly created task for the first time. | |
1649 | * | |
1650 | * This function will do some initial scheduler statistics housekeeping | |
1651 | * that must be done for every newly created context, then puts the task | |
1652 | * on the runqueue and wakes it. | |
1653 | */ | |
36c8b586 | 1654 | void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags) |
1da177e4 | 1655 | { |
70b97a7f | 1656 | struct rq *rq, *this_rq; |
1da177e4 LT |
1657 | unsigned long flags; |
1658 | int this_cpu, cpu; | |
1da177e4 LT |
1659 | |
1660 | rq = task_rq_lock(p, &flags); | |
147cbb4b | 1661 | BUG_ON(p->state != TASK_RUNNING); |
1da177e4 | 1662 | this_cpu = smp_processor_id(); |
147cbb4b | 1663 | cpu = task_cpu(p); |
1da177e4 | 1664 | |
1da177e4 LT |
1665 | /* |
1666 | * We decrease the sleep average of forking parents | |
1667 | * and children as well, to keep max-interactive tasks | |
1668 | * from forking tasks that are max-interactive. The parent | |
1669 | * (current) is done further down, under its lock. | |
1670 | */ | |
1671 | p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) * | |
1672 | CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS); | |
1673 | ||
1674 | p->prio = effective_prio(p); | |
1675 | ||
1676 | if (likely(cpu == this_cpu)) { | |
1677 | if (!(clone_flags & CLONE_VM)) { | |
1678 | /* | |
1679 | * The VM isn't cloned, so we're in a good position to | |
1680 | * do child-runs-first in anticipation of an exec. This | |
1681 | * usually avoids a lot of COW overhead. | |
1682 | */ | |
1683 | if (unlikely(!current->array)) | |
1684 | __activate_task(p, rq); | |
1685 | else { | |
1686 | p->prio = current->prio; | |
b29739f9 | 1687 | p->normal_prio = current->normal_prio; |
1da177e4 LT |
1688 | list_add_tail(&p->run_list, ¤t->run_list); |
1689 | p->array = current->array; | |
1690 | p->array->nr_active++; | |
2dd73a4f | 1691 | inc_nr_running(p, rq); |
1da177e4 LT |
1692 | } |
1693 | set_need_resched(); | |
1694 | } else | |
1695 | /* Run child last */ | |
1696 | __activate_task(p, rq); | |
1697 | /* | |
1698 | * We skip the following code due to cpu == this_cpu | |
1699 | * | |
1700 | * task_rq_unlock(rq, &flags); | |
1701 | * this_rq = task_rq_lock(current, &flags); | |
1702 | */ | |
1703 | this_rq = rq; | |
1704 | } else { | |
1705 | this_rq = cpu_rq(this_cpu); | |
1706 | ||
1707 | /* | |
1708 | * Not the local CPU - must adjust timestamp. This should | |
1709 | * get optimised away in the !CONFIG_SMP case. | |
1710 | */ | |
b18ec803 MG |
1711 | p->timestamp = (p->timestamp - this_rq->most_recent_timestamp) |
1712 | + rq->most_recent_timestamp; | |
1da177e4 LT |
1713 | __activate_task(p, rq); |
1714 | if (TASK_PREEMPTS_CURR(p, rq)) | |
1715 | resched_task(rq->curr); | |
1716 | ||
1717 | /* | |
1718 | * Parent and child are on different CPUs, now get the | |
1719 | * parent runqueue to update the parent's ->sleep_avg: | |
1720 | */ | |
1721 | task_rq_unlock(rq, &flags); | |
1722 | this_rq = task_rq_lock(current, &flags); | |
1723 | } | |
1724 | current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) * | |
1725 | PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS); | |
1726 | task_rq_unlock(this_rq, &flags); | |
1727 | } | |
1728 | ||
1729 | /* | |
1730 | * Potentially available exiting-child timeslices are | |
1731 | * retrieved here - this way the parent does not get | |
1732 | * penalized for creating too many threads. | |
1733 | * | |
1734 | * (this cannot be used to 'generate' timeslices | |
1735 | * artificially, because any timeslice recovered here | |
1736 | * was given away by the parent in the first place.) | |
1737 | */ | |
36c8b586 | 1738 | void fastcall sched_exit(struct task_struct *p) |
1da177e4 LT |
1739 | { |
1740 | unsigned long flags; | |
70b97a7f | 1741 | struct rq *rq; |
1da177e4 LT |
1742 | |
1743 | /* | |
1744 | * If the child was a (relative-) CPU hog then decrease | |
1745 | * the sleep_avg of the parent as well. | |
1746 | */ | |
1747 | rq = task_rq_lock(p->parent, &flags); | |
889dfafe | 1748 | if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) { |
1da177e4 LT |
1749 | p->parent->time_slice += p->time_slice; |
1750 | if (unlikely(p->parent->time_slice > task_timeslice(p))) | |
1751 | p->parent->time_slice = task_timeslice(p); | |
1752 | } | |
1753 | if (p->sleep_avg < p->parent->sleep_avg) | |
1754 | p->parent->sleep_avg = p->parent->sleep_avg / | |
1755 | (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg / | |
1756 | (EXIT_WEIGHT + 1); | |
1757 | task_rq_unlock(rq, &flags); | |
1758 | } | |
1759 | ||
4866cde0 NP |
1760 | /** |
1761 | * prepare_task_switch - prepare to switch tasks | |
1762 | * @rq: the runqueue preparing to switch | |
1763 | * @next: the task we are going to switch to. | |
1764 | * | |
1765 | * This is called with the rq lock held and interrupts off. It must | |
1766 | * be paired with a subsequent finish_task_switch after the context | |
1767 | * switch. | |
1768 | * | |
1769 | * prepare_task_switch sets up locking and calls architecture specific | |
1770 | * hooks. | |
1771 | */ | |
70b97a7f | 1772 | static inline void prepare_task_switch(struct rq *rq, struct task_struct *next) |
4866cde0 NP |
1773 | { |
1774 | prepare_lock_switch(rq, next); | |
1775 | prepare_arch_switch(next); | |
1776 | } | |
1777 | ||
1da177e4 LT |
1778 | /** |
1779 | * finish_task_switch - clean up after a task-switch | |
344babaa | 1780 | * @rq: runqueue associated with task-switch |
1da177e4 LT |
1781 | * @prev: the thread we just switched away from. |
1782 | * | |
4866cde0 NP |
1783 | * finish_task_switch must be called after the context switch, paired |
1784 | * with a prepare_task_switch call before the context switch. | |
1785 | * finish_task_switch will reconcile locking set up by prepare_task_switch, | |
1786 | * and do any other architecture-specific cleanup actions. | |
1da177e4 LT |
1787 | * |
1788 | * Note that we may have delayed dropping an mm in context_switch(). If | |
1789 | * so, we finish that here outside of the runqueue lock. (Doing it | |
1790 | * with the lock held can cause deadlocks; see schedule() for | |
1791 | * details.) | |
1792 | */ | |
70b97a7f | 1793 | static inline void finish_task_switch(struct rq *rq, struct task_struct *prev) |
1da177e4 LT |
1794 | __releases(rq->lock) |
1795 | { | |
1da177e4 | 1796 | struct mm_struct *mm = rq->prev_mm; |
55a101f8 | 1797 | long prev_state; |
1da177e4 LT |
1798 | |
1799 | rq->prev_mm = NULL; | |
1800 | ||
1801 | /* | |
1802 | * A task struct has one reference for the use as "current". | |
c394cc9f | 1803 | * If a task dies, then it sets TASK_DEAD in tsk->state and calls |
55a101f8 ON |
1804 | * schedule one last time. The schedule call will never return, and |
1805 | * the scheduled task must drop that reference. | |
c394cc9f | 1806 | * The test for TASK_DEAD must occur while the runqueue locks are |
1da177e4 LT |
1807 | * still held, otherwise prev could be scheduled on another cpu, die |
1808 | * there before we look at prev->state, and then the reference would | |
1809 | * be dropped twice. | |
1810 | * Manfred Spraul <manfred@colorfullife.com> | |
1811 | */ | |
55a101f8 | 1812 | prev_state = prev->state; |
4866cde0 NP |
1813 | finish_arch_switch(prev); |
1814 | finish_lock_switch(rq, prev); | |
1da177e4 LT |
1815 | if (mm) |
1816 | mmdrop(mm); | |
c394cc9f | 1817 | if (unlikely(prev_state == TASK_DEAD)) { |
c6fd91f0 | 1818 | /* |
1819 | * Remove function-return probe instances associated with this | |
1820 | * task and put them back on the free list. | |
1821 | */ | |
1822 | kprobe_flush_task(prev); | |
1da177e4 | 1823 | put_task_struct(prev); |
c6fd91f0 | 1824 | } |
1da177e4 LT |
1825 | } |
1826 | ||
1827 | /** | |
1828 | * schedule_tail - first thing a freshly forked thread must call. | |
1829 | * @prev: the thread we just switched away from. | |
1830 | */ | |
36c8b586 | 1831 | asmlinkage void schedule_tail(struct task_struct *prev) |
1da177e4 LT |
1832 | __releases(rq->lock) |
1833 | { | |
70b97a7f IM |
1834 | struct rq *rq = this_rq(); |
1835 | ||
4866cde0 NP |
1836 | finish_task_switch(rq, prev); |
1837 | #ifdef __ARCH_WANT_UNLOCKED_CTXSW | |
1838 | /* In this case, finish_task_switch does not reenable preemption */ | |
1839 | preempt_enable(); | |
1840 | #endif | |
1da177e4 LT |
1841 | if (current->set_child_tid) |
1842 | put_user(current->pid, current->set_child_tid); | |
1843 | } | |
1844 | ||
1845 | /* | |
1846 | * context_switch - switch to the new MM and the new | |
1847 | * thread's register state. | |
1848 | */ | |
36c8b586 | 1849 | static inline struct task_struct * |
70b97a7f | 1850 | context_switch(struct rq *rq, struct task_struct *prev, |
36c8b586 | 1851 | struct task_struct *next) |
1da177e4 LT |
1852 | { |
1853 | struct mm_struct *mm = next->mm; | |
1854 | struct mm_struct *oldmm = prev->active_mm; | |
1855 | ||
beed33a8 | 1856 | if (!mm) { |
1da177e4 LT |
1857 | next->active_mm = oldmm; |
1858 | atomic_inc(&oldmm->mm_count); | |
1859 | enter_lazy_tlb(oldmm, next); | |
1860 | } else | |
1861 | switch_mm(oldmm, mm, next); | |
1862 | ||
beed33a8 | 1863 | if (!prev->mm) { |
1da177e4 LT |
1864 | prev->active_mm = NULL; |
1865 | WARN_ON(rq->prev_mm); | |
1866 | rq->prev_mm = oldmm; | |
1867 | } | |
3a5f5e48 IM |
1868 | /* |
1869 | * Since the runqueue lock will be released by the next | |
1870 | * task (which is an invalid locking op but in the case | |
1871 | * of the scheduler it's an obvious special-case), so we | |
1872 | * do an early lockdep release here: | |
1873 | */ | |
1874 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW | |
8a25d5de | 1875 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); |
3a5f5e48 | 1876 | #endif |
1da177e4 LT |
1877 | |
1878 | /* Here we just switch the register state and the stack. */ | |
1879 | switch_to(prev, next, prev); | |
1880 | ||
1881 | return prev; | |
1882 | } | |
1883 | ||
1884 | /* | |
1885 | * nr_running, nr_uninterruptible and nr_context_switches: | |
1886 | * | |
1887 | * externally visible scheduler statistics: current number of runnable | |
1888 | * threads, current number of uninterruptible-sleeping threads, total | |
1889 | * number of context switches performed since bootup. | |
1890 | */ | |
1891 | unsigned long nr_running(void) | |
1892 | { | |
1893 | unsigned long i, sum = 0; | |
1894 | ||
1895 | for_each_online_cpu(i) | |
1896 | sum += cpu_rq(i)->nr_running; | |
1897 | ||
1898 | return sum; | |
1899 | } | |
1900 | ||
1901 | unsigned long nr_uninterruptible(void) | |
1902 | { | |
1903 | unsigned long i, sum = 0; | |
1904 | ||
0a945022 | 1905 | for_each_possible_cpu(i) |
1da177e4 LT |
1906 | sum += cpu_rq(i)->nr_uninterruptible; |
1907 | ||
1908 | /* | |
1909 | * Since we read the counters lockless, it might be slightly | |
1910 | * inaccurate. Do not allow it to go below zero though: | |
1911 | */ | |
1912 | if (unlikely((long)sum < 0)) | |
1913 | sum = 0; | |
1914 | ||
1915 | return sum; | |
1916 | } | |
1917 | ||
1918 | unsigned long long nr_context_switches(void) | |
1919 | { | |
cc94abfc SR |
1920 | int i; |
1921 | unsigned long long sum = 0; | |
1da177e4 | 1922 | |
0a945022 | 1923 | for_each_possible_cpu(i) |
1da177e4 LT |
1924 | sum += cpu_rq(i)->nr_switches; |
1925 | ||
1926 | return sum; | |
1927 | } | |
1928 | ||
1929 | unsigned long nr_iowait(void) | |
1930 | { | |
1931 | unsigned long i, sum = 0; | |
1932 | ||
0a945022 | 1933 | for_each_possible_cpu(i) |
1da177e4 LT |
1934 | sum += atomic_read(&cpu_rq(i)->nr_iowait); |
1935 | ||
1936 | return sum; | |
1937 | } | |
1938 | ||
db1b1fef JS |
1939 | unsigned long nr_active(void) |
1940 | { | |
1941 | unsigned long i, running = 0, uninterruptible = 0; | |
1942 | ||
1943 | for_each_online_cpu(i) { | |
1944 | running += cpu_rq(i)->nr_running; | |
1945 | uninterruptible += cpu_rq(i)->nr_uninterruptible; | |
1946 | } | |
1947 | ||
1948 | if (unlikely((long)uninterruptible < 0)) | |
1949 | uninterruptible = 0; | |
1950 | ||
1951 | return running + uninterruptible; | |
1952 | } | |
1953 | ||
1da177e4 LT |
1954 | #ifdef CONFIG_SMP |
1955 | ||
48f24c4d IM |
1956 | /* |
1957 | * Is this task likely cache-hot: | |
1958 | */ | |
1959 | static inline int | |
1960 | task_hot(struct task_struct *p, unsigned long long now, struct sched_domain *sd) | |
1961 | { | |
1962 | return (long long)(now - p->last_ran) < (long long)sd->cache_hot_time; | |
1963 | } | |
1964 | ||
1da177e4 LT |
1965 | /* |
1966 | * double_rq_lock - safely lock two runqueues | |
1967 | * | |
1968 | * Note this does not disable interrupts like task_rq_lock, | |
1969 | * you need to do so manually before calling. | |
1970 | */ | |
70b97a7f | 1971 | static void double_rq_lock(struct rq *rq1, struct rq *rq2) |
1da177e4 LT |
1972 | __acquires(rq1->lock) |
1973 | __acquires(rq2->lock) | |
1974 | { | |
054b9108 | 1975 | BUG_ON(!irqs_disabled()); |
1da177e4 LT |
1976 | if (rq1 == rq2) { |
1977 | spin_lock(&rq1->lock); | |
1978 | __acquire(rq2->lock); /* Fake it out ;) */ | |
1979 | } else { | |
c96d145e | 1980 | if (rq1 < rq2) { |
1da177e4 LT |
1981 | spin_lock(&rq1->lock); |
1982 | spin_lock(&rq2->lock); | |
1983 | } else { | |
1984 | spin_lock(&rq2->lock); | |
1985 | spin_lock(&rq1->lock); | |
1986 | } | |
1987 | } | |
1988 | } | |
1989 | ||
1990 | /* | |
1991 | * double_rq_unlock - safely unlock two runqueues | |
1992 | * | |
1993 | * Note this does not restore interrupts like task_rq_unlock, | |
1994 | * you need to do so manually after calling. | |
1995 | */ | |
70b97a7f | 1996 | static void double_rq_unlock(struct rq *rq1, struct rq *rq2) |
1da177e4 LT |
1997 | __releases(rq1->lock) |
1998 | __releases(rq2->lock) | |
1999 | { | |
2000 | spin_unlock(&rq1->lock); | |
2001 | if (rq1 != rq2) | |
2002 | spin_unlock(&rq2->lock); | |
2003 | else | |
2004 | __release(rq2->lock); | |
2005 | } | |
2006 | ||
2007 | /* | |
2008 | * double_lock_balance - lock the busiest runqueue, this_rq is locked already. | |
2009 | */ | |
70b97a7f | 2010 | static void double_lock_balance(struct rq *this_rq, struct rq *busiest) |
1da177e4 LT |
2011 | __releases(this_rq->lock) |
2012 | __acquires(busiest->lock) | |
2013 | __acquires(this_rq->lock) | |
2014 | { | |
054b9108 KK |
2015 | if (unlikely(!irqs_disabled())) { |
2016 | /* printk() doesn't work good under rq->lock */ | |
2017 | spin_unlock(&this_rq->lock); | |
2018 | BUG_ON(1); | |
2019 | } | |
1da177e4 | 2020 | if (unlikely(!spin_trylock(&busiest->lock))) { |
c96d145e | 2021 | if (busiest < this_rq) { |
1da177e4 LT |
2022 | spin_unlock(&this_rq->lock); |
2023 | spin_lock(&busiest->lock); | |
2024 | spin_lock(&this_rq->lock); | |
2025 | } else | |
2026 | spin_lock(&busiest->lock); | |
2027 | } | |
2028 | } | |
2029 | ||
1da177e4 LT |
2030 | /* |
2031 | * If dest_cpu is allowed for this process, migrate the task to it. | |
2032 | * This is accomplished by forcing the cpu_allowed mask to only | |
2033 | * allow dest_cpu, which will force the cpu onto dest_cpu. Then | |
2034 | * the cpu_allowed mask is restored. | |
2035 | */ | |
36c8b586 | 2036 | static void sched_migrate_task(struct task_struct *p, int dest_cpu) |
1da177e4 | 2037 | { |
70b97a7f | 2038 | struct migration_req req; |
1da177e4 | 2039 | unsigned long flags; |
70b97a7f | 2040 | struct rq *rq; |
1da177e4 LT |
2041 | |
2042 | rq = task_rq_lock(p, &flags); | |
2043 | if (!cpu_isset(dest_cpu, p->cpus_allowed) | |
2044 | || unlikely(cpu_is_offline(dest_cpu))) | |
2045 | goto out; | |
2046 | ||
2047 | /* force the process onto the specified CPU */ | |
2048 | if (migrate_task(p, dest_cpu, &req)) { | |
2049 | /* Need to wait for migration thread (might exit: take ref). */ | |
2050 | struct task_struct *mt = rq->migration_thread; | |
36c8b586 | 2051 | |
1da177e4 LT |
2052 | get_task_struct(mt); |
2053 | task_rq_unlock(rq, &flags); | |
2054 | wake_up_process(mt); | |
2055 | put_task_struct(mt); | |
2056 | wait_for_completion(&req.done); | |
36c8b586 | 2057 | |
1da177e4 LT |
2058 | return; |
2059 | } | |
2060 | out: | |
2061 | task_rq_unlock(rq, &flags); | |
2062 | } | |
2063 | ||
2064 | /* | |
476d139c NP |
2065 | * sched_exec - execve() is a valuable balancing opportunity, because at |
2066 | * this point the task has the smallest effective memory and cache footprint. | |
1da177e4 LT |
2067 | */ |
2068 | void sched_exec(void) | |
2069 | { | |
1da177e4 | 2070 | int new_cpu, this_cpu = get_cpu(); |
476d139c | 2071 | new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC); |
1da177e4 | 2072 | put_cpu(); |
476d139c NP |
2073 | if (new_cpu != this_cpu) |
2074 | sched_migrate_task(current, new_cpu); | |
1da177e4 LT |
2075 | } |
2076 | ||
2077 | /* | |
2078 | * pull_task - move a task from a remote runqueue to the local runqueue. | |
2079 | * Both runqueues must be locked. | |
2080 | */ | |
70b97a7f IM |
2081 | static void pull_task(struct rq *src_rq, struct prio_array *src_array, |
2082 | struct task_struct *p, struct rq *this_rq, | |
2083 | struct prio_array *this_array, int this_cpu) | |
1da177e4 LT |
2084 | { |
2085 | dequeue_task(p, src_array); | |
2dd73a4f | 2086 | dec_nr_running(p, src_rq); |
1da177e4 | 2087 | set_task_cpu(p, this_cpu); |
2dd73a4f | 2088 | inc_nr_running(p, this_rq); |
1da177e4 | 2089 | enqueue_task(p, this_array); |
b18ec803 MG |
2090 | p->timestamp = (p->timestamp - src_rq->most_recent_timestamp) |
2091 | + this_rq->most_recent_timestamp; | |
1da177e4 LT |
2092 | /* |
2093 | * Note that idle threads have a prio of MAX_PRIO, for this test | |
2094 | * to be always true for them. | |
2095 | */ | |
2096 | if (TASK_PREEMPTS_CURR(p, this_rq)) | |
2097 | resched_task(this_rq->curr); | |
2098 | } | |
2099 | ||
2100 | /* | |
2101 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
2102 | */ | |
858119e1 | 2103 | static |
70b97a7f | 2104 | int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu, |
95cdf3b7 IM |
2105 | struct sched_domain *sd, enum idle_type idle, |
2106 | int *all_pinned) | |
1da177e4 LT |
2107 | { |
2108 | /* | |
2109 | * We do not migrate tasks that are: | |
2110 | * 1) running (obviously), or | |
2111 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
2112 | * 3) are cache-hot on their current CPU. | |
2113 | */ | |
1da177e4 LT |
2114 | if (!cpu_isset(this_cpu, p->cpus_allowed)) |
2115 | return 0; | |
81026794 NP |
2116 | *all_pinned = 0; |
2117 | ||
2118 | if (task_running(rq, p)) | |
2119 | return 0; | |
1da177e4 LT |
2120 | |
2121 | /* | |
2122 | * Aggressive migration if: | |
cafb20c1 | 2123 | * 1) task is cache cold, or |
1da177e4 LT |
2124 | * 2) too many balance attempts have failed. |
2125 | */ | |
2126 | ||
b18ec803 MG |
2127 | if (sd->nr_balance_failed > sd->cache_nice_tries) { |
2128 | #ifdef CONFIG_SCHEDSTATS | |
2129 | if (task_hot(p, rq->most_recent_timestamp, sd)) | |
2130 | schedstat_inc(sd, lb_hot_gained[idle]); | |
2131 | #endif | |
1da177e4 | 2132 | return 1; |
b18ec803 | 2133 | } |
1da177e4 | 2134 | |
b18ec803 | 2135 | if (task_hot(p, rq->most_recent_timestamp, sd)) |
81026794 | 2136 | return 0; |
1da177e4 LT |
2137 | return 1; |
2138 | } | |
2139 | ||
615052dc | 2140 | #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio) |
48f24c4d | 2141 | |
1da177e4 | 2142 | /* |
2dd73a4f PW |
2143 | * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted |
2144 | * load from busiest to this_rq, as part of a balancing operation within | |
2145 | * "domain". Returns the number of tasks moved. | |
1da177e4 LT |
2146 | * |
2147 | * Called with both runqueues locked. | |
2148 | */ | |
70b97a7f | 2149 | static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, |
2dd73a4f PW |
2150 | unsigned long max_nr_move, unsigned long max_load_move, |
2151 | struct sched_domain *sd, enum idle_type idle, | |
2152 | int *all_pinned) | |
1da177e4 | 2153 | { |
48f24c4d IM |
2154 | int idx, pulled = 0, pinned = 0, this_best_prio, best_prio, |
2155 | best_prio_seen, skip_for_load; | |
70b97a7f | 2156 | struct prio_array *array, *dst_array; |
1da177e4 | 2157 | struct list_head *head, *curr; |
36c8b586 | 2158 | struct task_struct *tmp; |
2dd73a4f | 2159 | long rem_load_move; |
1da177e4 | 2160 | |
2dd73a4f | 2161 | if (max_nr_move == 0 || max_load_move == 0) |
1da177e4 LT |
2162 | goto out; |
2163 | ||
2dd73a4f | 2164 | rem_load_move = max_load_move; |
81026794 | 2165 | pinned = 1; |
615052dc | 2166 | this_best_prio = rq_best_prio(this_rq); |
48f24c4d | 2167 | best_prio = rq_best_prio(busiest); |
615052dc PW |
2168 | /* |
2169 | * Enable handling of the case where there is more than one task | |
2170 | * with the best priority. If the current running task is one | |
48f24c4d | 2171 | * of those with prio==best_prio we know it won't be moved |
615052dc PW |
2172 | * and therefore it's safe to override the skip (based on load) of |
2173 | * any task we find with that prio. | |
2174 | */ | |
48f24c4d | 2175 | best_prio_seen = best_prio == busiest->curr->prio; |
81026794 | 2176 | |
1da177e4 LT |
2177 | /* |
2178 | * We first consider expired tasks. Those will likely not be | |
2179 | * executed in the near future, and they are most likely to | |
2180 | * be cache-cold, thus switching CPUs has the least effect | |
2181 | * on them. | |
2182 | */ | |
2183 | if (busiest->expired->nr_active) { | |
2184 | array = busiest->expired; | |
2185 | dst_array = this_rq->expired; | |
2186 | } else { | |
2187 | array = busiest->active; | |
2188 | dst_array = this_rq->active; | |
2189 | } | |
2190 | ||
2191 | new_array: | |
2192 | /* Start searching at priority 0: */ | |
2193 | idx = 0; | |
2194 | skip_bitmap: | |
2195 | if (!idx) | |
2196 | idx = sched_find_first_bit(array->bitmap); | |
2197 | else | |
2198 | idx = find_next_bit(array->bitmap, MAX_PRIO, idx); | |
2199 | if (idx >= MAX_PRIO) { | |
2200 | if (array == busiest->expired && busiest->active->nr_active) { | |
2201 | array = busiest->active; | |
2202 | dst_array = this_rq->active; | |
2203 | goto new_array; | |
2204 | } | |
2205 | goto out; | |
2206 | } | |
2207 | ||
2208 | head = array->queue + idx; | |
2209 | curr = head->prev; | |
2210 | skip_queue: | |
36c8b586 | 2211 | tmp = list_entry(curr, struct task_struct, run_list); |
1da177e4 LT |
2212 | |
2213 | curr = curr->prev; | |
2214 | ||
50ddd969 PW |
2215 | /* |
2216 | * To help distribute high priority tasks accross CPUs we don't | |
2217 | * skip a task if it will be the highest priority task (i.e. smallest | |
2218 | * prio value) on its new queue regardless of its load weight | |
2219 | */ | |
615052dc PW |
2220 | skip_for_load = tmp->load_weight > rem_load_move; |
2221 | if (skip_for_load && idx < this_best_prio) | |
48f24c4d | 2222 | skip_for_load = !best_prio_seen && idx == best_prio; |
615052dc | 2223 | if (skip_for_load || |
2dd73a4f | 2224 | !can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) { |
48f24c4d IM |
2225 | |
2226 | best_prio_seen |= idx == best_prio; | |
1da177e4 LT |
2227 | if (curr != head) |
2228 | goto skip_queue; | |
2229 | idx++; | |
2230 | goto skip_bitmap; | |
2231 | } | |
2232 | ||
1da177e4 LT |
2233 | pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu); |
2234 | pulled++; | |
2dd73a4f | 2235 | rem_load_move -= tmp->load_weight; |
1da177e4 | 2236 | |
2dd73a4f PW |
2237 | /* |
2238 | * We only want to steal up to the prescribed number of tasks | |
2239 | * and the prescribed amount of weighted load. | |
2240 | */ | |
2241 | if (pulled < max_nr_move && rem_load_move > 0) { | |
615052dc PW |
2242 | if (idx < this_best_prio) |
2243 | this_best_prio = idx; | |
1da177e4 LT |
2244 | if (curr != head) |
2245 | goto skip_queue; | |
2246 | idx++; | |
2247 | goto skip_bitmap; | |
2248 | } | |
2249 | out: | |
2250 | /* | |
2251 | * Right now, this is the only place pull_task() is called, | |
2252 | * so we can safely collect pull_task() stats here rather than | |
2253 | * inside pull_task(). | |
2254 | */ | |
2255 | schedstat_add(sd, lb_gained[idle], pulled); | |
81026794 NP |
2256 | |
2257 | if (all_pinned) | |
2258 | *all_pinned = pinned; | |
1da177e4 LT |
2259 | return pulled; |
2260 | } | |
2261 | ||
2262 | /* | |
2263 | * find_busiest_group finds and returns the busiest CPU group within the | |
48f24c4d IM |
2264 | * domain. It calculates and returns the amount of weighted load which |
2265 | * should be moved to restore balance via the imbalance parameter. | |
1da177e4 LT |
2266 | */ |
2267 | static struct sched_group * | |
2268 | find_busiest_group(struct sched_domain *sd, int this_cpu, | |
0a2966b4 | 2269 | unsigned long *imbalance, enum idle_type idle, int *sd_idle, |
783609c6 | 2270 | cpumask_t *cpus, int *balance) |
1da177e4 LT |
2271 | { |
2272 | struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups; | |
2273 | unsigned long max_load, avg_load, total_load, this_load, total_pwr; | |
0c117f1b | 2274 | unsigned long max_pull; |
2dd73a4f PW |
2275 | unsigned long busiest_load_per_task, busiest_nr_running; |
2276 | unsigned long this_load_per_task, this_nr_running; | |
7897986b | 2277 | int load_idx; |
5c45bf27 SS |
2278 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
2279 | int power_savings_balance = 1; | |
2280 | unsigned long leader_nr_running = 0, min_load_per_task = 0; | |
2281 | unsigned long min_nr_running = ULONG_MAX; | |
2282 | struct sched_group *group_min = NULL, *group_leader = NULL; | |
2283 | #endif | |
1da177e4 LT |
2284 | |
2285 | max_load = this_load = total_load = total_pwr = 0; | |
2dd73a4f PW |
2286 | busiest_load_per_task = busiest_nr_running = 0; |
2287 | this_load_per_task = this_nr_running = 0; | |
7897986b NP |
2288 | if (idle == NOT_IDLE) |
2289 | load_idx = sd->busy_idx; | |
2290 | else if (idle == NEWLY_IDLE) | |
2291 | load_idx = sd->newidle_idx; | |
2292 | else | |
2293 | load_idx = sd->idle_idx; | |
1da177e4 LT |
2294 | |
2295 | do { | |
5c45bf27 | 2296 | unsigned long load, group_capacity; |
1da177e4 LT |
2297 | int local_group; |
2298 | int i; | |
783609c6 | 2299 | unsigned int balance_cpu = -1, first_idle_cpu = 0; |
2dd73a4f | 2300 | unsigned long sum_nr_running, sum_weighted_load; |
1da177e4 LT |
2301 | |
2302 | local_group = cpu_isset(this_cpu, group->cpumask); | |
2303 | ||
783609c6 SS |
2304 | if (local_group) |
2305 | balance_cpu = first_cpu(group->cpumask); | |
2306 | ||
1da177e4 | 2307 | /* Tally up the load of all CPUs in the group */ |
2dd73a4f | 2308 | sum_weighted_load = sum_nr_running = avg_load = 0; |
1da177e4 LT |
2309 | |
2310 | for_each_cpu_mask(i, group->cpumask) { | |
0a2966b4 CL |
2311 | struct rq *rq; |
2312 | ||
2313 | if (!cpu_isset(i, *cpus)) | |
2314 | continue; | |
2315 | ||
2316 | rq = cpu_rq(i); | |
2dd73a4f | 2317 | |
5969fe06 NP |
2318 | if (*sd_idle && !idle_cpu(i)) |
2319 | *sd_idle = 0; | |
2320 | ||
1da177e4 | 2321 | /* Bias balancing toward cpus of our domain */ |
783609c6 SS |
2322 | if (local_group) { |
2323 | if (idle_cpu(i) && !first_idle_cpu) { | |
2324 | first_idle_cpu = 1; | |
2325 | balance_cpu = i; | |
2326 | } | |
2327 | ||
a2000572 | 2328 | load = target_load(i, load_idx); |
783609c6 | 2329 | } else |
a2000572 | 2330 | load = source_load(i, load_idx); |
1da177e4 LT |
2331 | |
2332 | avg_load += load; | |
2dd73a4f PW |
2333 | sum_nr_running += rq->nr_running; |
2334 | sum_weighted_load += rq->raw_weighted_load; | |
1da177e4 LT |
2335 | } |
2336 | ||
783609c6 SS |
2337 | /* |
2338 | * First idle cpu or the first cpu(busiest) in this sched group | |
2339 | * is eligible for doing load balancing at this and above | |
2340 | * domains. | |
2341 | */ | |
2342 | if (local_group && balance_cpu != this_cpu && balance) { | |
2343 | *balance = 0; | |
2344 | goto ret; | |
2345 | } | |
2346 | ||
1da177e4 LT |
2347 | total_load += avg_load; |
2348 | total_pwr += group->cpu_power; | |
2349 | ||
2350 | /* Adjust by relative CPU power of the group */ | |
2351 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
2352 | ||
5c45bf27 SS |
2353 | group_capacity = group->cpu_power / SCHED_LOAD_SCALE; |
2354 | ||
1da177e4 LT |
2355 | if (local_group) { |
2356 | this_load = avg_load; | |
2357 | this = group; | |
2dd73a4f PW |
2358 | this_nr_running = sum_nr_running; |
2359 | this_load_per_task = sum_weighted_load; | |
2360 | } else if (avg_load > max_load && | |
5c45bf27 | 2361 | sum_nr_running > group_capacity) { |
1da177e4 LT |
2362 | max_load = avg_load; |
2363 | busiest = group; | |
2dd73a4f PW |
2364 | busiest_nr_running = sum_nr_running; |
2365 | busiest_load_per_task = sum_weighted_load; | |
1da177e4 | 2366 | } |
5c45bf27 SS |
2367 | |
2368 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
2369 | /* | |
2370 | * Busy processors will not participate in power savings | |
2371 | * balance. | |
2372 | */ | |
2373 | if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE)) | |
2374 | goto group_next; | |
2375 | ||
2376 | /* | |
2377 | * If the local group is idle or completely loaded | |
2378 | * no need to do power savings balance at this domain | |
2379 | */ | |
2380 | if (local_group && (this_nr_running >= group_capacity || | |
2381 | !this_nr_running)) | |
2382 | power_savings_balance = 0; | |
2383 | ||
2384 | /* | |
2385 | * If a group is already running at full capacity or idle, | |
2386 | * don't include that group in power savings calculations | |
2387 | */ | |
2388 | if (!power_savings_balance || sum_nr_running >= group_capacity | |
2389 | || !sum_nr_running) | |
2390 | goto group_next; | |
2391 | ||
2392 | /* | |
2393 | * Calculate the group which has the least non-idle load. | |
2394 | * This is the group from where we need to pick up the load | |
2395 | * for saving power | |
2396 | */ | |
2397 | if ((sum_nr_running < min_nr_running) || | |
2398 | (sum_nr_running == min_nr_running && | |
2399 | first_cpu(group->cpumask) < | |
2400 | first_cpu(group_min->cpumask))) { | |
2401 | group_min = group; | |
2402 | min_nr_running = sum_nr_running; | |
2403 | min_load_per_task = sum_weighted_load / | |
2404 | sum_nr_running; | |
2405 | } | |
2406 | ||
2407 | /* | |
2408 | * Calculate the group which is almost near its | |
2409 | * capacity but still has some space to pick up some load | |
2410 | * from other group and save more power | |
2411 | */ | |
48f24c4d | 2412 | if (sum_nr_running <= group_capacity - 1) { |
5c45bf27 SS |
2413 | if (sum_nr_running > leader_nr_running || |
2414 | (sum_nr_running == leader_nr_running && | |
2415 | first_cpu(group->cpumask) > | |
2416 | first_cpu(group_leader->cpumask))) { | |
2417 | group_leader = group; | |
2418 | leader_nr_running = sum_nr_running; | |
2419 | } | |
48f24c4d | 2420 | } |
5c45bf27 SS |
2421 | group_next: |
2422 | #endif | |
1da177e4 LT |
2423 | group = group->next; |
2424 | } while (group != sd->groups); | |
2425 | ||
2dd73a4f | 2426 | if (!busiest || this_load >= max_load || busiest_nr_running == 0) |
1da177e4 LT |
2427 | goto out_balanced; |
2428 | ||
2429 | avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr; | |
2430 | ||
2431 | if (this_load >= avg_load || | |
2432 | 100*max_load <= sd->imbalance_pct*this_load) | |
2433 | goto out_balanced; | |
2434 | ||
2dd73a4f | 2435 | busiest_load_per_task /= busiest_nr_running; |
1da177e4 LT |
2436 | /* |
2437 | * We're trying to get all the cpus to the average_load, so we don't | |
2438 | * want to push ourselves above the average load, nor do we wish to | |
2439 | * reduce the max loaded cpu below the average load, as either of these | |
2440 | * actions would just result in more rebalancing later, and ping-pong | |
2441 | * tasks around. Thus we look for the minimum possible imbalance. | |
2442 | * Negative imbalances (*we* are more loaded than anyone else) will | |
2443 | * be counted as no imbalance for these purposes -- we can't fix that | |
2444 | * by pulling tasks to us. Be careful of negative numbers as they'll | |
2445 | * appear as very large values with unsigned longs. | |
2446 | */ | |
2dd73a4f PW |
2447 | if (max_load <= busiest_load_per_task) |
2448 | goto out_balanced; | |
2449 | ||
2450 | /* | |
2451 | * In the presence of smp nice balancing, certain scenarios can have | |
2452 | * max load less than avg load(as we skip the groups at or below | |
2453 | * its cpu_power, while calculating max_load..) | |
2454 | */ | |
2455 | if (max_load < avg_load) { | |
2456 | *imbalance = 0; | |
2457 | goto small_imbalance; | |
2458 | } | |
0c117f1b SS |
2459 | |
2460 | /* Don't want to pull so many tasks that a group would go idle */ | |
2dd73a4f | 2461 | max_pull = min(max_load - avg_load, max_load - busiest_load_per_task); |
0c117f1b | 2462 | |
1da177e4 | 2463 | /* How much load to actually move to equalise the imbalance */ |
0c117f1b | 2464 | *imbalance = min(max_pull * busiest->cpu_power, |
1da177e4 LT |
2465 | (avg_load - this_load) * this->cpu_power) |
2466 | / SCHED_LOAD_SCALE; | |
2467 | ||
2dd73a4f PW |
2468 | /* |
2469 | * if *imbalance is less than the average load per runnable task | |
2470 | * there is no gaurantee that any tasks will be moved so we'll have | |
2471 | * a think about bumping its value to force at least one task to be | |
2472 | * moved | |
2473 | */ | |
2474 | if (*imbalance < busiest_load_per_task) { | |
48f24c4d | 2475 | unsigned long tmp, pwr_now, pwr_move; |
2dd73a4f PW |
2476 | unsigned int imbn; |
2477 | ||
2478 | small_imbalance: | |
2479 | pwr_move = pwr_now = 0; | |
2480 | imbn = 2; | |
2481 | if (this_nr_running) { | |
2482 | this_load_per_task /= this_nr_running; | |
2483 | if (busiest_load_per_task > this_load_per_task) | |
2484 | imbn = 1; | |
2485 | } else | |
2486 | this_load_per_task = SCHED_LOAD_SCALE; | |
1da177e4 | 2487 | |
2dd73a4f PW |
2488 | if (max_load - this_load >= busiest_load_per_task * imbn) { |
2489 | *imbalance = busiest_load_per_task; | |
1da177e4 LT |
2490 | return busiest; |
2491 | } | |
2492 | ||
2493 | /* | |
2494 | * OK, we don't have enough imbalance to justify moving tasks, | |
2495 | * however we may be able to increase total CPU power used by | |
2496 | * moving them. | |
2497 | */ | |
2498 | ||
2dd73a4f PW |
2499 | pwr_now += busiest->cpu_power * |
2500 | min(busiest_load_per_task, max_load); | |
2501 | pwr_now += this->cpu_power * | |
2502 | min(this_load_per_task, this_load); | |
1da177e4 LT |
2503 | pwr_now /= SCHED_LOAD_SCALE; |
2504 | ||
2505 | /* Amount of load we'd subtract */ | |
33859f7f MOS |
2506 | tmp = busiest_load_per_task * SCHED_LOAD_SCALE / |
2507 | busiest->cpu_power; | |
1da177e4 | 2508 | if (max_load > tmp) |
2dd73a4f PW |
2509 | pwr_move += busiest->cpu_power * |
2510 | min(busiest_load_per_task, max_load - tmp); | |
1da177e4 LT |
2511 | |
2512 | /* Amount of load we'd add */ | |
33859f7f MOS |
2513 | if (max_load * busiest->cpu_power < |
2514 | busiest_load_per_task * SCHED_LOAD_SCALE) | |
2515 | tmp = max_load * busiest->cpu_power / this->cpu_power; | |
1da177e4 | 2516 | else |
33859f7f MOS |
2517 | tmp = busiest_load_per_task * SCHED_LOAD_SCALE / |
2518 | this->cpu_power; | |
2519 | pwr_move += this->cpu_power * | |
2520 | min(this_load_per_task, this_load + tmp); | |
1da177e4 LT |
2521 | pwr_move /= SCHED_LOAD_SCALE; |
2522 | ||
2523 | /* Move if we gain throughput */ | |
2524 | if (pwr_move <= pwr_now) | |
2525 | goto out_balanced; | |
2526 | ||
2dd73a4f | 2527 | *imbalance = busiest_load_per_task; |
1da177e4 LT |
2528 | } |
2529 | ||
1da177e4 LT |
2530 | return busiest; |
2531 | ||
2532 | out_balanced: | |
5c45bf27 SS |
2533 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
2534 | if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE)) | |
2535 | goto ret; | |
1da177e4 | 2536 | |
5c45bf27 SS |
2537 | if (this == group_leader && group_leader != group_min) { |
2538 | *imbalance = min_load_per_task; | |
2539 | return group_min; | |
2540 | } | |
5c45bf27 | 2541 | #endif |
783609c6 | 2542 | ret: |
1da177e4 LT |
2543 | *imbalance = 0; |
2544 | return NULL; | |
2545 | } | |
2546 | ||
2547 | /* | |
2548 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
2549 | */ | |
70b97a7f | 2550 | static struct rq * |
48f24c4d | 2551 | find_busiest_queue(struct sched_group *group, enum idle_type idle, |
0a2966b4 | 2552 | unsigned long imbalance, cpumask_t *cpus) |
1da177e4 | 2553 | { |
70b97a7f | 2554 | struct rq *busiest = NULL, *rq; |
2dd73a4f | 2555 | unsigned long max_load = 0; |
1da177e4 LT |
2556 | int i; |
2557 | ||
2558 | for_each_cpu_mask(i, group->cpumask) { | |
0a2966b4 CL |
2559 | |
2560 | if (!cpu_isset(i, *cpus)) | |
2561 | continue; | |
2562 | ||
48f24c4d | 2563 | rq = cpu_rq(i); |
2dd73a4f | 2564 | |
48f24c4d | 2565 | if (rq->nr_running == 1 && rq->raw_weighted_load > imbalance) |
2dd73a4f | 2566 | continue; |
1da177e4 | 2567 | |
48f24c4d IM |
2568 | if (rq->raw_weighted_load > max_load) { |
2569 | max_load = rq->raw_weighted_load; | |
2570 | busiest = rq; | |
1da177e4 LT |
2571 | } |
2572 | } | |
2573 | ||
2574 | return busiest; | |
2575 | } | |
2576 | ||
77391d71 NP |
2577 | /* |
2578 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
2579 | * so long as it is large enough. | |
2580 | */ | |
2581 | #define MAX_PINNED_INTERVAL 512 | |
2582 | ||
48f24c4d IM |
2583 | static inline unsigned long minus_1_or_zero(unsigned long n) |
2584 | { | |
2585 | return n > 0 ? n - 1 : 0; | |
2586 | } | |
2587 | ||
1da177e4 LT |
2588 | /* |
2589 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
2590 | * tasks if there is an imbalance. | |
1da177e4 | 2591 | */ |
70b97a7f | 2592 | static int load_balance(int this_cpu, struct rq *this_rq, |
783609c6 SS |
2593 | struct sched_domain *sd, enum idle_type idle, |
2594 | int *balance) | |
1da177e4 | 2595 | { |
48f24c4d | 2596 | int nr_moved, all_pinned = 0, active_balance = 0, sd_idle = 0; |
1da177e4 | 2597 | struct sched_group *group; |
1da177e4 | 2598 | unsigned long imbalance; |
70b97a7f | 2599 | struct rq *busiest; |
0a2966b4 | 2600 | cpumask_t cpus = CPU_MASK_ALL; |
fe2eea3f | 2601 | unsigned long flags; |
5969fe06 | 2602 | |
89c4710e SS |
2603 | /* |
2604 | * When power savings policy is enabled for the parent domain, idle | |
2605 | * sibling can pick up load irrespective of busy siblings. In this case, | |
2606 | * let the state of idle sibling percolate up as IDLE, instead of | |
2607 | * portraying it as NOT_IDLE. | |
2608 | */ | |
5c45bf27 | 2609 | if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER && |
89c4710e | 2610 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
5969fe06 | 2611 | sd_idle = 1; |
1da177e4 | 2612 | |
1da177e4 LT |
2613 | schedstat_inc(sd, lb_cnt[idle]); |
2614 | ||
0a2966b4 CL |
2615 | redo: |
2616 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle, | |
783609c6 SS |
2617 | &cpus, balance); |
2618 | ||
06066714 | 2619 | if (*balance == 0) |
783609c6 | 2620 | goto out_balanced; |
783609c6 | 2621 | |
1da177e4 LT |
2622 | if (!group) { |
2623 | schedstat_inc(sd, lb_nobusyg[idle]); | |
2624 | goto out_balanced; | |
2625 | } | |
2626 | ||
0a2966b4 | 2627 | busiest = find_busiest_queue(group, idle, imbalance, &cpus); |
1da177e4 LT |
2628 | if (!busiest) { |
2629 | schedstat_inc(sd, lb_nobusyq[idle]); | |
2630 | goto out_balanced; | |
2631 | } | |
2632 | ||
db935dbd | 2633 | BUG_ON(busiest == this_rq); |
1da177e4 LT |
2634 | |
2635 | schedstat_add(sd, lb_imbalance[idle], imbalance); | |
2636 | ||
2637 | nr_moved = 0; | |
2638 | if (busiest->nr_running > 1) { | |
2639 | /* | |
2640 | * Attempt to move tasks. If find_busiest_group has found | |
2641 | * an imbalance but busiest->nr_running <= 1, the group is | |
2642 | * still unbalanced. nr_moved simply stays zero, so it is | |
2643 | * correctly treated as an imbalance. | |
2644 | */ | |
fe2eea3f | 2645 | local_irq_save(flags); |
e17224bf | 2646 | double_rq_lock(this_rq, busiest); |
1da177e4 | 2647 | nr_moved = move_tasks(this_rq, this_cpu, busiest, |
48f24c4d IM |
2648 | minus_1_or_zero(busiest->nr_running), |
2649 | imbalance, sd, idle, &all_pinned); | |
e17224bf | 2650 | double_rq_unlock(this_rq, busiest); |
fe2eea3f | 2651 | local_irq_restore(flags); |
81026794 NP |
2652 | |
2653 | /* All tasks on this runqueue were pinned by CPU affinity */ | |
0a2966b4 CL |
2654 | if (unlikely(all_pinned)) { |
2655 | cpu_clear(cpu_of(busiest), cpus); | |
2656 | if (!cpus_empty(cpus)) | |
2657 | goto redo; | |
81026794 | 2658 | goto out_balanced; |
0a2966b4 | 2659 | } |
1da177e4 | 2660 | } |
81026794 | 2661 | |
1da177e4 LT |
2662 | if (!nr_moved) { |
2663 | schedstat_inc(sd, lb_failed[idle]); | |
2664 | sd->nr_balance_failed++; | |
2665 | ||
2666 | if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) { | |
1da177e4 | 2667 | |
fe2eea3f | 2668 | spin_lock_irqsave(&busiest->lock, flags); |
fa3b6ddc SS |
2669 | |
2670 | /* don't kick the migration_thread, if the curr | |
2671 | * task on busiest cpu can't be moved to this_cpu | |
2672 | */ | |
2673 | if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) { | |
fe2eea3f | 2674 | spin_unlock_irqrestore(&busiest->lock, flags); |
fa3b6ddc SS |
2675 | all_pinned = 1; |
2676 | goto out_one_pinned; | |
2677 | } | |
2678 | ||
1da177e4 LT |
2679 | if (!busiest->active_balance) { |
2680 | busiest->active_balance = 1; | |
2681 | busiest->push_cpu = this_cpu; | |
81026794 | 2682 | active_balance = 1; |
1da177e4 | 2683 | } |
fe2eea3f | 2684 | spin_unlock_irqrestore(&busiest->lock, flags); |
81026794 | 2685 | if (active_balance) |
1da177e4 LT |
2686 | wake_up_process(busiest->migration_thread); |
2687 | ||
2688 | /* | |
2689 | * We've kicked active balancing, reset the failure | |
2690 | * counter. | |
2691 | */ | |
39507451 | 2692 | sd->nr_balance_failed = sd->cache_nice_tries+1; |
1da177e4 | 2693 | } |
81026794 | 2694 | } else |
1da177e4 LT |
2695 | sd->nr_balance_failed = 0; |
2696 | ||
81026794 | 2697 | if (likely(!active_balance)) { |
1da177e4 LT |
2698 | /* We were unbalanced, so reset the balancing interval */ |
2699 | sd->balance_interval = sd->min_interval; | |
81026794 NP |
2700 | } else { |
2701 | /* | |
2702 | * If we've begun active balancing, start to back off. This | |
2703 | * case may not be covered by the all_pinned logic if there | |
2704 | * is only 1 task on the busy runqueue (because we don't call | |
2705 | * move_tasks). | |
2706 | */ | |
2707 | if (sd->balance_interval < sd->max_interval) | |
2708 | sd->balance_interval *= 2; | |
1da177e4 LT |
2709 | } |
2710 | ||
5c45bf27 | 2711 | if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
89c4710e | 2712 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
5969fe06 | 2713 | return -1; |
1da177e4 LT |
2714 | return nr_moved; |
2715 | ||
2716 | out_balanced: | |
1da177e4 LT |
2717 | schedstat_inc(sd, lb_balanced[idle]); |
2718 | ||
16cfb1c0 | 2719 | sd->nr_balance_failed = 0; |
fa3b6ddc SS |
2720 | |
2721 | out_one_pinned: | |
1da177e4 | 2722 | /* tune up the balancing interval */ |
77391d71 NP |
2723 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || |
2724 | (sd->balance_interval < sd->max_interval)) | |
1da177e4 LT |
2725 | sd->balance_interval *= 2; |
2726 | ||
48f24c4d | 2727 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
89c4710e | 2728 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
5969fe06 | 2729 | return -1; |
1da177e4 LT |
2730 | return 0; |
2731 | } | |
2732 | ||
2733 | /* | |
2734 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
2735 | * tasks if there is an imbalance. | |
2736 | * | |
2737 | * Called from schedule when this_rq is about to become idle (NEWLY_IDLE). | |
2738 | * this_rq is locked. | |
2739 | */ | |
48f24c4d | 2740 | static int |
70b97a7f | 2741 | load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd) |
1da177e4 LT |
2742 | { |
2743 | struct sched_group *group; | |
70b97a7f | 2744 | struct rq *busiest = NULL; |
1da177e4 LT |
2745 | unsigned long imbalance; |
2746 | int nr_moved = 0; | |
5969fe06 | 2747 | int sd_idle = 0; |
0a2966b4 | 2748 | cpumask_t cpus = CPU_MASK_ALL; |
5969fe06 | 2749 | |
89c4710e SS |
2750 | /* |
2751 | * When power savings policy is enabled for the parent domain, idle | |
2752 | * sibling can pick up load irrespective of busy siblings. In this case, | |
2753 | * let the state of idle sibling percolate up as IDLE, instead of | |
2754 | * portraying it as NOT_IDLE. | |
2755 | */ | |
2756 | if (sd->flags & SD_SHARE_CPUPOWER && | |
2757 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
5969fe06 | 2758 | sd_idle = 1; |
1da177e4 LT |
2759 | |
2760 | schedstat_inc(sd, lb_cnt[NEWLY_IDLE]); | |
0a2966b4 CL |
2761 | redo: |
2762 | group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE, | |
783609c6 | 2763 | &sd_idle, &cpus, NULL); |
1da177e4 | 2764 | if (!group) { |
1da177e4 | 2765 | schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]); |
16cfb1c0 | 2766 | goto out_balanced; |
1da177e4 LT |
2767 | } |
2768 | ||
0a2966b4 CL |
2769 | busiest = find_busiest_queue(group, NEWLY_IDLE, imbalance, |
2770 | &cpus); | |
db935dbd | 2771 | if (!busiest) { |
1da177e4 | 2772 | schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]); |
16cfb1c0 | 2773 | goto out_balanced; |
1da177e4 LT |
2774 | } |
2775 | ||
db935dbd NP |
2776 | BUG_ON(busiest == this_rq); |
2777 | ||
1da177e4 | 2778 | schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance); |
d6d5cfaf NP |
2779 | |
2780 | nr_moved = 0; | |
2781 | if (busiest->nr_running > 1) { | |
2782 | /* Attempt to move tasks */ | |
2783 | double_lock_balance(this_rq, busiest); | |
2784 | nr_moved = move_tasks(this_rq, this_cpu, busiest, | |
2dd73a4f | 2785 | minus_1_or_zero(busiest->nr_running), |
81026794 | 2786 | imbalance, sd, NEWLY_IDLE, NULL); |
d6d5cfaf | 2787 | spin_unlock(&busiest->lock); |
0a2966b4 CL |
2788 | |
2789 | if (!nr_moved) { | |
2790 | cpu_clear(cpu_of(busiest), cpus); | |
2791 | if (!cpus_empty(cpus)) | |
2792 | goto redo; | |
2793 | } | |
d6d5cfaf NP |
2794 | } |
2795 | ||
5969fe06 | 2796 | if (!nr_moved) { |
1da177e4 | 2797 | schedstat_inc(sd, lb_failed[NEWLY_IDLE]); |
89c4710e SS |
2798 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
2799 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
5969fe06 NP |
2800 | return -1; |
2801 | } else | |
16cfb1c0 | 2802 | sd->nr_balance_failed = 0; |
1da177e4 | 2803 | |
1da177e4 | 2804 | return nr_moved; |
16cfb1c0 NP |
2805 | |
2806 | out_balanced: | |
2807 | schedstat_inc(sd, lb_balanced[NEWLY_IDLE]); | |
48f24c4d | 2808 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
89c4710e | 2809 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
5969fe06 | 2810 | return -1; |
16cfb1c0 | 2811 | sd->nr_balance_failed = 0; |
48f24c4d | 2812 | |
16cfb1c0 | 2813 | return 0; |
1da177e4 LT |
2814 | } |
2815 | ||
2816 | /* | |
2817 | * idle_balance is called by schedule() if this_cpu is about to become | |
2818 | * idle. Attempts to pull tasks from other CPUs. | |
2819 | */ | |
70b97a7f | 2820 | static void idle_balance(int this_cpu, struct rq *this_rq) |
1da177e4 LT |
2821 | { |
2822 | struct sched_domain *sd; | |
1bd77f2d CL |
2823 | int pulled_task = 0; |
2824 | unsigned long next_balance = jiffies + 60 * HZ; | |
1da177e4 LT |
2825 | |
2826 | for_each_domain(this_cpu, sd) { | |
2827 | if (sd->flags & SD_BALANCE_NEWIDLE) { | |
48f24c4d | 2828 | /* If we've pulled tasks over stop searching: */ |
1bd77f2d CL |
2829 | pulled_task = load_balance_newidle(this_cpu, |
2830 | this_rq, sd); | |
2831 | if (time_after(next_balance, | |
2832 | sd->last_balance + sd->balance_interval)) | |
2833 | next_balance = sd->last_balance | |
2834 | + sd->balance_interval; | |
2835 | if (pulled_task) | |
1da177e4 | 2836 | break; |
1da177e4 LT |
2837 | } |
2838 | } | |
1bd77f2d CL |
2839 | if (!pulled_task) |
2840 | /* | |
2841 | * We are going idle. next_balance may be set based on | |
2842 | * a busy processor. So reset next_balance. | |
2843 | */ | |
2844 | this_rq->next_balance = next_balance; | |
1da177e4 LT |
2845 | } |
2846 | ||
2847 | /* | |
2848 | * active_load_balance is run by migration threads. It pushes running tasks | |
2849 | * off the busiest CPU onto idle CPUs. It requires at least 1 task to be | |
2850 | * running on each physical CPU where possible, and avoids physical / | |
2851 | * logical imbalances. | |
2852 | * | |
2853 | * Called with busiest_rq locked. | |
2854 | */ | |
70b97a7f | 2855 | static void active_load_balance(struct rq *busiest_rq, int busiest_cpu) |
1da177e4 | 2856 | { |
39507451 | 2857 | int target_cpu = busiest_rq->push_cpu; |
70b97a7f IM |
2858 | struct sched_domain *sd; |
2859 | struct rq *target_rq; | |
39507451 | 2860 | |
48f24c4d | 2861 | /* Is there any task to move? */ |
39507451 | 2862 | if (busiest_rq->nr_running <= 1) |
39507451 NP |
2863 | return; |
2864 | ||
2865 | target_rq = cpu_rq(target_cpu); | |
1da177e4 LT |
2866 | |
2867 | /* | |
39507451 NP |
2868 | * This condition is "impossible", if it occurs |
2869 | * we need to fix it. Originally reported by | |
2870 | * Bjorn Helgaas on a 128-cpu setup. | |
1da177e4 | 2871 | */ |
39507451 | 2872 | BUG_ON(busiest_rq == target_rq); |
1da177e4 | 2873 | |
39507451 NP |
2874 | /* move a task from busiest_rq to target_rq */ |
2875 | double_lock_balance(busiest_rq, target_rq); | |
2876 | ||
2877 | /* Search for an sd spanning us and the target CPU. */ | |
c96d145e | 2878 | for_each_domain(target_cpu, sd) { |
39507451 | 2879 | if ((sd->flags & SD_LOAD_BALANCE) && |
48f24c4d | 2880 | cpu_isset(busiest_cpu, sd->span)) |
39507451 | 2881 | break; |
c96d145e | 2882 | } |
39507451 | 2883 | |
48f24c4d IM |
2884 | if (likely(sd)) { |
2885 | schedstat_inc(sd, alb_cnt); | |
39507451 | 2886 | |
48f24c4d IM |
2887 | if (move_tasks(target_rq, target_cpu, busiest_rq, 1, |
2888 | RTPRIO_TO_LOAD_WEIGHT(100), sd, SCHED_IDLE, | |
2889 | NULL)) | |
2890 | schedstat_inc(sd, alb_pushed); | |
2891 | else | |
2892 | schedstat_inc(sd, alb_failed); | |
2893 | } | |
39507451 | 2894 | spin_unlock(&target_rq->lock); |
1da177e4 LT |
2895 | } |
2896 | ||
7835b98b | 2897 | static void update_load(struct rq *this_rq) |
1da177e4 | 2898 | { |
7835b98b | 2899 | unsigned long this_load; |
48f24c4d | 2900 | int i, scale; |
1da177e4 | 2901 | |
2dd73a4f | 2902 | this_load = this_rq->raw_weighted_load; |
48f24c4d IM |
2903 | |
2904 | /* Update our load: */ | |
2905 | for (i = 0, scale = 1; i < 3; i++, scale <<= 1) { | |
2906 | unsigned long old_load, new_load; | |
2907 | ||
7897986b | 2908 | old_load = this_rq->cpu_load[i]; |
48f24c4d | 2909 | new_load = this_load; |
7897986b NP |
2910 | /* |
2911 | * Round up the averaging division if load is increasing. This | |
2912 | * prevents us from getting stuck on 9 if the load is 10, for | |
2913 | * example. | |
2914 | */ | |
2915 | if (new_load > old_load) | |
2916 | new_load += scale-1; | |
2917 | this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale; | |
2918 | } | |
7835b98b CL |
2919 | } |
2920 | ||
2921 | /* | |
c9819f45 | 2922 | * run_rebalance_domains is triggered when needed from the scheduler tick. |
7835b98b CL |
2923 | * |
2924 | * It checks each scheduling domain to see if it is due to be balanced, | |
2925 | * and initiates a balancing operation if so. | |
2926 | * | |
2927 | * Balancing parameters are set up in arch_init_sched_domains. | |
2928 | */ | |
08c183f3 | 2929 | static DEFINE_SPINLOCK(balancing); |
7835b98b | 2930 | |
c9819f45 | 2931 | static void run_rebalance_domains(struct softirq_action *h) |
7835b98b | 2932 | { |
783609c6 | 2933 | int this_cpu = smp_processor_id(), balance = 1; |
c9819f45 | 2934 | struct rq *this_rq = cpu_rq(this_cpu); |
7835b98b CL |
2935 | unsigned long interval; |
2936 | struct sched_domain *sd; | |
e418e1c2 CL |
2937 | /* |
2938 | * We are idle if there are no processes running. This | |
2939 | * is valid even if we are the idle process (SMT). | |
2940 | */ | |
2941 | enum idle_type idle = !this_rq->nr_running ? | |
2942 | SCHED_IDLE : NOT_IDLE; | |
c9819f45 CL |
2943 | /* Earliest time when we have to call run_rebalance_domains again */ |
2944 | unsigned long next_balance = jiffies + 60*HZ; | |
1da177e4 LT |
2945 | |
2946 | for_each_domain(this_cpu, sd) { | |
1da177e4 LT |
2947 | if (!(sd->flags & SD_LOAD_BALANCE)) |
2948 | continue; | |
2949 | ||
2950 | interval = sd->balance_interval; | |
2951 | if (idle != SCHED_IDLE) | |
2952 | interval *= sd->busy_factor; | |
2953 | ||
2954 | /* scale ms to jiffies */ | |
2955 | interval = msecs_to_jiffies(interval); | |
2956 | if (unlikely(!interval)) | |
2957 | interval = 1; | |
2958 | ||
08c183f3 CL |
2959 | if (sd->flags & SD_SERIALIZE) { |
2960 | if (!spin_trylock(&balancing)) | |
2961 | goto out; | |
2962 | } | |
2963 | ||
c9819f45 | 2964 | if (time_after_eq(jiffies, sd->last_balance + interval)) { |
783609c6 | 2965 | if (load_balance(this_cpu, this_rq, sd, idle, &balance)) { |
fa3b6ddc SS |
2966 | /* |
2967 | * We've pulled tasks over so either we're no | |
5969fe06 NP |
2968 | * longer idle, or one of our SMT siblings is |
2969 | * not idle. | |
2970 | */ | |
1da177e4 LT |
2971 | idle = NOT_IDLE; |
2972 | } | |
1bd77f2d | 2973 | sd->last_balance = jiffies; |
1da177e4 | 2974 | } |
08c183f3 CL |
2975 | if (sd->flags & SD_SERIALIZE) |
2976 | spin_unlock(&balancing); | |
2977 | out: | |
c9819f45 CL |
2978 | if (time_after(next_balance, sd->last_balance + interval)) |
2979 | next_balance = sd->last_balance + interval; | |
783609c6 SS |
2980 | |
2981 | /* | |
2982 | * Stop the load balance at this level. There is another | |
2983 | * CPU in our sched group which is doing load balancing more | |
2984 | * actively. | |
2985 | */ | |
2986 | if (!balance) | |
2987 | break; | |
1da177e4 | 2988 | } |
c9819f45 | 2989 | this_rq->next_balance = next_balance; |
1da177e4 LT |
2990 | } |
2991 | #else | |
2992 | /* | |
2993 | * on UP we do not need to balance between CPUs: | |
2994 | */ | |
70b97a7f | 2995 | static inline void idle_balance(int cpu, struct rq *rq) |
1da177e4 LT |
2996 | { |
2997 | } | |
2998 | #endif | |
2999 | ||
e418e1c2 | 3000 | static inline void wake_priority_sleeper(struct rq *rq) |
1da177e4 | 3001 | { |
1da177e4 | 3002 | #ifdef CONFIG_SCHED_SMT |
571f6d2f | 3003 | if (!rq->nr_running) |
e418e1c2 | 3004 | return; |
571f6d2f | 3005 | |
1da177e4 LT |
3006 | spin_lock(&rq->lock); |
3007 | /* | |
3008 | * If an SMT sibling task has been put to sleep for priority | |
3009 | * reasons reschedule the idle task to see if it can now run. | |
3010 | */ | |
e418e1c2 | 3011 | if (rq->nr_running) |
1da177e4 | 3012 | resched_task(rq->idle); |
1da177e4 LT |
3013 | spin_unlock(&rq->lock); |
3014 | #endif | |
1da177e4 LT |
3015 | } |
3016 | ||
3017 | DEFINE_PER_CPU(struct kernel_stat, kstat); | |
3018 | ||
3019 | EXPORT_PER_CPU_SYMBOL(kstat); | |
3020 | ||
3021 | /* | |
3022 | * This is called on clock ticks and on context switches. | |
3023 | * Bank in p->sched_time the ns elapsed since the last tick or switch. | |
3024 | */ | |
48f24c4d | 3025 | static inline void |
70b97a7f | 3026 | update_cpu_clock(struct task_struct *p, struct rq *rq, unsigned long long now) |
1da177e4 | 3027 | { |
b18ec803 MG |
3028 | p->sched_time += now - p->last_ran; |
3029 | p->last_ran = rq->most_recent_timestamp = now; | |
1da177e4 LT |
3030 | } |
3031 | ||
3032 | /* | |
3033 | * Return current->sched_time plus any more ns on the sched_clock | |
3034 | * that have not yet been banked. | |
3035 | */ | |
36c8b586 | 3036 | unsigned long long current_sched_time(const struct task_struct *p) |
1da177e4 LT |
3037 | { |
3038 | unsigned long long ns; | |
3039 | unsigned long flags; | |
48f24c4d | 3040 | |
1da177e4 | 3041 | local_irq_save(flags); |
b18ec803 | 3042 | ns = p->sched_time + sched_clock() - p->last_ran; |
1da177e4 | 3043 | local_irq_restore(flags); |
48f24c4d | 3044 | |
1da177e4 LT |
3045 | return ns; |
3046 | } | |
3047 | ||
f1adad78 LT |
3048 | /* |
3049 | * We place interactive tasks back into the active array, if possible. | |
3050 | * | |
3051 | * To guarantee that this does not starve expired tasks we ignore the | |
3052 | * interactivity of a task if the first expired task had to wait more | |
3053 | * than a 'reasonable' amount of time. This deadline timeout is | |
3054 | * load-dependent, as the frequency of array switched decreases with | |
3055 | * increasing number of running tasks. We also ignore the interactivity | |
3056 | * if a better static_prio task has expired: | |
3057 | */ | |
70b97a7f | 3058 | static inline int expired_starving(struct rq *rq) |
48f24c4d IM |
3059 | { |
3060 | if (rq->curr->static_prio > rq->best_expired_prio) | |
3061 | return 1; | |
3062 | if (!STARVATION_LIMIT || !rq->expired_timestamp) | |
3063 | return 0; | |
3064 | if (jiffies - rq->expired_timestamp > STARVATION_LIMIT * rq->nr_running) | |
3065 | return 1; | |
3066 | return 0; | |
3067 | } | |
f1adad78 | 3068 | |
1da177e4 LT |
3069 | /* |
3070 | * Account user cpu time to a process. | |
3071 | * @p: the process that the cpu time gets accounted to | |
3072 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
3073 | * @cputime: the cpu time spent in user space since the last update | |
3074 | */ | |
3075 | void account_user_time(struct task_struct *p, cputime_t cputime) | |
3076 | { | |
3077 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
3078 | cputime64_t tmp; | |
3079 | ||
3080 | p->utime = cputime_add(p->utime, cputime); | |
3081 | ||
3082 | /* Add user time to cpustat. */ | |
3083 | tmp = cputime_to_cputime64(cputime); | |
3084 | if (TASK_NICE(p) > 0) | |
3085 | cpustat->nice = cputime64_add(cpustat->nice, tmp); | |
3086 | else | |
3087 | cpustat->user = cputime64_add(cpustat->user, tmp); | |
3088 | } | |
3089 | ||
3090 | /* | |
3091 | * Account system cpu time to a process. | |
3092 | * @p: the process that the cpu time gets accounted to | |
3093 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
3094 | * @cputime: the cpu time spent in kernel space since the last update | |
3095 | */ | |
3096 | void account_system_time(struct task_struct *p, int hardirq_offset, | |
3097 | cputime_t cputime) | |
3098 | { | |
3099 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
70b97a7f | 3100 | struct rq *rq = this_rq(); |
1da177e4 LT |
3101 | cputime64_t tmp; |
3102 | ||
3103 | p->stime = cputime_add(p->stime, cputime); | |
3104 | ||
3105 | /* Add system time to cpustat. */ | |
3106 | tmp = cputime_to_cputime64(cputime); | |
3107 | if (hardirq_count() - hardirq_offset) | |
3108 | cpustat->irq = cputime64_add(cpustat->irq, tmp); | |
3109 | else if (softirq_count()) | |
3110 | cpustat->softirq = cputime64_add(cpustat->softirq, tmp); | |
3111 | else if (p != rq->idle) | |
3112 | cpustat->system = cputime64_add(cpustat->system, tmp); | |
3113 | else if (atomic_read(&rq->nr_iowait) > 0) | |
3114 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); | |
3115 | else | |
3116 | cpustat->idle = cputime64_add(cpustat->idle, tmp); | |
3117 | /* Account for system time used */ | |
3118 | acct_update_integrals(p); | |
1da177e4 LT |
3119 | } |
3120 | ||
3121 | /* | |
3122 | * Account for involuntary wait time. | |
3123 | * @p: the process from which the cpu time has been stolen | |
3124 | * @steal: the cpu time spent in involuntary wait | |
3125 | */ | |
3126 | void account_steal_time(struct task_struct *p, cputime_t steal) | |
3127 | { | |
3128 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
3129 | cputime64_t tmp = cputime_to_cputime64(steal); | |
70b97a7f | 3130 | struct rq *rq = this_rq(); |
1da177e4 LT |
3131 | |
3132 | if (p == rq->idle) { | |
3133 | p->stime = cputime_add(p->stime, steal); | |
3134 | if (atomic_read(&rq->nr_iowait) > 0) | |
3135 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); | |
3136 | else | |
3137 | cpustat->idle = cputime64_add(cpustat->idle, tmp); | |
3138 | } else | |
3139 | cpustat->steal = cputime64_add(cpustat->steal, tmp); | |
3140 | } | |
3141 | ||
7835b98b | 3142 | static void task_running_tick(struct rq *rq, struct task_struct *p) |
1da177e4 | 3143 | { |
1da177e4 | 3144 | if (p->array != rq->active) { |
7835b98b | 3145 | /* Task has expired but was not scheduled yet */ |
1da177e4 | 3146 | set_tsk_need_resched(p); |
7835b98b | 3147 | return; |
1da177e4 LT |
3148 | } |
3149 | spin_lock(&rq->lock); | |
3150 | /* | |
3151 | * The task was running during this tick - update the | |
3152 | * time slice counter. Note: we do not update a thread's | |
3153 | * priority until it either goes to sleep or uses up its | |
3154 | * timeslice. This makes it possible for interactive tasks | |
3155 | * to use up their timeslices at their highest priority levels. | |
3156 | */ | |
3157 | if (rt_task(p)) { | |
3158 | /* | |
3159 | * RR tasks need a special form of timeslice management. | |
3160 | * FIFO tasks have no timeslices. | |
3161 | */ | |
3162 | if ((p->policy == SCHED_RR) && !--p->time_slice) { | |
3163 | p->time_slice = task_timeslice(p); | |
3164 | p->first_time_slice = 0; | |
3165 | set_tsk_need_resched(p); | |
3166 | ||
3167 | /* put it at the end of the queue: */ | |
3168 | requeue_task(p, rq->active); | |
3169 | } | |
3170 | goto out_unlock; | |
3171 | } | |
3172 | if (!--p->time_slice) { | |
3173 | dequeue_task(p, rq->active); | |
3174 | set_tsk_need_resched(p); | |
3175 | p->prio = effective_prio(p); | |
3176 | p->time_slice = task_timeslice(p); | |
3177 | p->first_time_slice = 0; | |
3178 | ||
3179 | if (!rq->expired_timestamp) | |
3180 | rq->expired_timestamp = jiffies; | |
48f24c4d | 3181 | if (!TASK_INTERACTIVE(p) || expired_starving(rq)) { |
1da177e4 LT |
3182 | enqueue_task(p, rq->expired); |
3183 | if (p->static_prio < rq->best_expired_prio) | |
3184 | rq->best_expired_prio = p->static_prio; | |
3185 | } else | |
3186 | enqueue_task(p, rq->active); | |
3187 | } else { | |
3188 | /* | |
3189 | * Prevent a too long timeslice allowing a task to monopolize | |
3190 | * the CPU. We do this by splitting up the timeslice into | |
3191 | * smaller pieces. | |
3192 | * | |
3193 | * Note: this does not mean the task's timeslices expire or | |
3194 | * get lost in any way, they just might be preempted by | |
3195 | * another task of equal priority. (one with higher | |
3196 | * priority would have preempted this task already.) We | |
3197 | * requeue this task to the end of the list on this priority | |
3198 | * level, which is in essence a round-robin of tasks with | |
3199 | * equal priority. | |
3200 | * | |
3201 | * This only applies to tasks in the interactive | |
3202 | * delta range with at least TIMESLICE_GRANULARITY to requeue. | |
3203 | */ | |
3204 | if (TASK_INTERACTIVE(p) && !((task_timeslice(p) - | |
3205 | p->time_slice) % TIMESLICE_GRANULARITY(p)) && | |
3206 | (p->time_slice >= TIMESLICE_GRANULARITY(p)) && | |
3207 | (p->array == rq->active)) { | |
3208 | ||
3209 | requeue_task(p, rq->active); | |
3210 | set_tsk_need_resched(p); | |
3211 | } | |
3212 | } | |
3213 | out_unlock: | |
3214 | spin_unlock(&rq->lock); | |
7835b98b CL |
3215 | } |
3216 | ||
3217 | /* | |
3218 | * This function gets called by the timer code, with HZ frequency. | |
3219 | * We call it with interrupts disabled. | |
3220 | * | |
3221 | * It also gets called by the fork code, when changing the parent's | |
3222 | * timeslices. | |
3223 | */ | |
3224 | void scheduler_tick(void) | |
3225 | { | |
3226 | unsigned long long now = sched_clock(); | |
3227 | struct task_struct *p = current; | |
3228 | int cpu = smp_processor_id(); | |
3229 | struct rq *rq = cpu_rq(cpu); | |
7835b98b CL |
3230 | |
3231 | update_cpu_clock(p, rq, now); | |
3232 | ||
e418e1c2 | 3233 | if (p == rq->idle) |
7835b98b | 3234 | /* Task on the idle queue */ |
e418e1c2 CL |
3235 | wake_priority_sleeper(rq); |
3236 | else | |
7835b98b | 3237 | task_running_tick(rq, p); |
e418e1c2 | 3238 | #ifdef CONFIG_SMP |
7835b98b | 3239 | update_load(rq); |
c9819f45 CL |
3240 | if (time_after_eq(jiffies, rq->next_balance)) |
3241 | raise_softirq(SCHED_SOFTIRQ); | |
e418e1c2 | 3242 | #endif |
1da177e4 LT |
3243 | } |
3244 | ||
3245 | #ifdef CONFIG_SCHED_SMT | |
70b97a7f | 3246 | static inline void wakeup_busy_runqueue(struct rq *rq) |
fc38ed75 CK |
3247 | { |
3248 | /* If an SMT runqueue is sleeping due to priority reasons wake it up */ | |
3249 | if (rq->curr == rq->idle && rq->nr_running) | |
3250 | resched_task(rq->idle); | |
3251 | } | |
3252 | ||
c96d145e KC |
3253 | /* |
3254 | * Called with interrupt disabled and this_rq's runqueue locked. | |
3255 | */ | |
3256 | static void wake_sleeping_dependent(int this_cpu) | |
1da177e4 | 3257 | { |
41c7ce9a | 3258 | struct sched_domain *tmp, *sd = NULL; |
1da177e4 LT |
3259 | int i; |
3260 | ||
c96d145e KC |
3261 | for_each_domain(this_cpu, tmp) { |
3262 | if (tmp->flags & SD_SHARE_CPUPOWER) { | |
41c7ce9a | 3263 | sd = tmp; |
c96d145e KC |
3264 | break; |
3265 | } | |
3266 | } | |
41c7ce9a NP |
3267 | |
3268 | if (!sd) | |
1da177e4 LT |
3269 | return; |
3270 | ||
c96d145e | 3271 | for_each_cpu_mask(i, sd->span) { |
70b97a7f | 3272 | struct rq *smt_rq = cpu_rq(i); |
1da177e4 | 3273 | |
c96d145e KC |
3274 | if (i == this_cpu) |
3275 | continue; | |
3276 | if (unlikely(!spin_trylock(&smt_rq->lock))) | |
3277 | continue; | |
3278 | ||
fc38ed75 | 3279 | wakeup_busy_runqueue(smt_rq); |
c96d145e | 3280 | spin_unlock(&smt_rq->lock); |
1da177e4 | 3281 | } |
1da177e4 LT |
3282 | } |
3283 | ||
67f9a619 IM |
3284 | /* |
3285 | * number of 'lost' timeslices this task wont be able to fully | |
3286 | * utilize, if another task runs on a sibling. This models the | |
3287 | * slowdown effect of other tasks running on siblings: | |
3288 | */ | |
36c8b586 IM |
3289 | static inline unsigned long |
3290 | smt_slice(struct task_struct *p, struct sched_domain *sd) | |
67f9a619 IM |
3291 | { |
3292 | return p->time_slice * (100 - sd->per_cpu_gain) / 100; | |
3293 | } | |
3294 | ||
c96d145e KC |
3295 | /* |
3296 | * To minimise lock contention and not have to drop this_rq's runlock we only | |
3297 | * trylock the sibling runqueues and bypass those runqueues if we fail to | |
3298 | * acquire their lock. As we only trylock the normal locking order does not | |
3299 | * need to be obeyed. | |
3300 | */ | |
36c8b586 | 3301 | static int |
70b97a7f | 3302 | dependent_sleeper(int this_cpu, struct rq *this_rq, struct task_struct *p) |
1da177e4 | 3303 | { |
41c7ce9a | 3304 | struct sched_domain *tmp, *sd = NULL; |
1da177e4 | 3305 | int ret = 0, i; |
1da177e4 | 3306 | |
c96d145e KC |
3307 | /* kernel/rt threads do not participate in dependent sleeping */ |
3308 | if (!p->mm || rt_task(p)) | |
3309 | return 0; | |
3310 | ||
3311 | for_each_domain(this_cpu, tmp) { | |
3312 | if (tmp->flags & SD_SHARE_CPUPOWER) { | |
41c7ce9a | 3313 | sd = tmp; |
c96d145e KC |
3314 | break; |
3315 | } | |
3316 | } | |
41c7ce9a NP |
3317 | |
3318 | if (!sd) | |
1da177e4 LT |
3319 | return 0; |
3320 | ||
c96d145e | 3321 | for_each_cpu_mask(i, sd->span) { |
36c8b586 | 3322 | struct task_struct *smt_curr; |
70b97a7f | 3323 | struct rq *smt_rq; |
1da177e4 | 3324 | |
c96d145e KC |
3325 | if (i == this_cpu) |
3326 | continue; | |
1da177e4 | 3327 | |
c96d145e KC |
3328 | smt_rq = cpu_rq(i); |
3329 | if (unlikely(!spin_trylock(&smt_rq->lock))) | |
3330 | continue; | |
1da177e4 | 3331 | |
c96d145e | 3332 | smt_curr = smt_rq->curr; |
1da177e4 | 3333 | |
c96d145e KC |
3334 | if (!smt_curr->mm) |
3335 | goto unlock; | |
fc38ed75 | 3336 | |
1da177e4 LT |
3337 | /* |
3338 | * If a user task with lower static priority than the | |
3339 | * running task on the SMT sibling is trying to schedule, | |
3340 | * delay it till there is proportionately less timeslice | |
3341 | * left of the sibling task to prevent a lower priority | |
3342 | * task from using an unfair proportion of the | |
3343 | * physical cpu's resources. -ck | |
3344 | */ | |
fc38ed75 CK |
3345 | if (rt_task(smt_curr)) { |
3346 | /* | |
3347 | * With real time tasks we run non-rt tasks only | |
3348 | * per_cpu_gain% of the time. | |
3349 | */ | |
3350 | if ((jiffies % DEF_TIMESLICE) > | |
3351 | (sd->per_cpu_gain * DEF_TIMESLICE / 100)) | |
3352 | ret = 1; | |
c96d145e | 3353 | } else { |
67f9a619 IM |
3354 | if (smt_curr->static_prio < p->static_prio && |
3355 | !TASK_PREEMPTS_CURR(p, smt_rq) && | |
3356 | smt_slice(smt_curr, sd) > task_timeslice(p)) | |
fc38ed75 | 3357 | ret = 1; |
fc38ed75 | 3358 | } |
c96d145e KC |
3359 | unlock: |
3360 | spin_unlock(&smt_rq->lock); | |
1da177e4 | 3361 | } |
1da177e4 LT |
3362 | return ret; |
3363 | } | |
3364 | #else | |
c96d145e | 3365 | static inline void wake_sleeping_dependent(int this_cpu) |
1da177e4 LT |
3366 | { |
3367 | } | |
48f24c4d | 3368 | static inline int |
70b97a7f | 3369 | dependent_sleeper(int this_cpu, struct rq *this_rq, struct task_struct *p) |
1da177e4 LT |
3370 | { |
3371 | return 0; | |
3372 | } | |
3373 | #endif | |
3374 | ||
3375 | #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT) | |
3376 | ||
3377 | void fastcall add_preempt_count(int val) | |
3378 | { | |
3379 | /* | |
3380 | * Underflow? | |
3381 | */ | |
9a11b49a IM |
3382 | if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) |
3383 | return; | |
1da177e4 LT |
3384 | preempt_count() += val; |
3385 | /* | |
3386 | * Spinlock count overflowing soon? | |
3387 | */ | |
33859f7f MOS |
3388 | DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= |
3389 | PREEMPT_MASK - 10); | |
1da177e4 LT |
3390 | } |
3391 | EXPORT_SYMBOL(add_preempt_count); | |
3392 | ||
3393 | void fastcall sub_preempt_count(int val) | |
3394 | { | |
3395 | /* | |
3396 | * Underflow? | |
3397 | */ | |
9a11b49a IM |
3398 | if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) |
3399 | return; | |
1da177e4 LT |
3400 | /* |
3401 | * Is the spinlock portion underflowing? | |
3402 | */ | |
9a11b49a IM |
3403 | if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && |
3404 | !(preempt_count() & PREEMPT_MASK))) | |
3405 | return; | |
3406 | ||
1da177e4 LT |
3407 | preempt_count() -= val; |
3408 | } | |
3409 | EXPORT_SYMBOL(sub_preempt_count); | |
3410 | ||
3411 | #endif | |
3412 | ||
3dee386e CK |
3413 | static inline int interactive_sleep(enum sleep_type sleep_type) |
3414 | { | |
3415 | return (sleep_type == SLEEP_INTERACTIVE || | |
3416 | sleep_type == SLEEP_INTERRUPTED); | |
3417 | } | |
3418 | ||
1da177e4 LT |
3419 | /* |
3420 | * schedule() is the main scheduler function. | |
3421 | */ | |
3422 | asmlinkage void __sched schedule(void) | |
3423 | { | |
36c8b586 | 3424 | struct task_struct *prev, *next; |
70b97a7f | 3425 | struct prio_array *array; |
1da177e4 LT |
3426 | struct list_head *queue; |
3427 | unsigned long long now; | |
3428 | unsigned long run_time; | |
a3464a10 | 3429 | int cpu, idx, new_prio; |
48f24c4d | 3430 | long *switch_count; |
70b97a7f | 3431 | struct rq *rq; |
1da177e4 LT |
3432 | |
3433 | /* | |
3434 | * Test if we are atomic. Since do_exit() needs to call into | |
3435 | * schedule() atomically, we ignore that path for now. | |
3436 | * Otherwise, whine if we are scheduling when we should not be. | |
3437 | */ | |
77e4bfbc AM |
3438 | if (unlikely(in_atomic() && !current->exit_state)) { |
3439 | printk(KERN_ERR "BUG: scheduling while atomic: " | |
3440 | "%s/0x%08x/%d\n", | |
3441 | current->comm, preempt_count(), current->pid); | |
a4c410f0 | 3442 | debug_show_held_locks(current); |
3117df04 IM |
3443 | if (irqs_disabled()) |
3444 | print_irqtrace_events(current); | |
77e4bfbc | 3445 | dump_stack(); |
1da177e4 LT |
3446 | } |
3447 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); | |
3448 | ||
3449 | need_resched: | |
3450 | preempt_disable(); | |
3451 | prev = current; | |
3452 | release_kernel_lock(prev); | |
3453 | need_resched_nonpreemptible: | |
3454 | rq = this_rq(); | |
3455 | ||
3456 | /* | |
3457 | * The idle thread is not allowed to schedule! | |
3458 | * Remove this check after it has been exercised a bit. | |
3459 | */ | |
3460 | if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) { | |
3461 | printk(KERN_ERR "bad: scheduling from the idle thread!\n"); | |
3462 | dump_stack(); | |
3463 | } | |
3464 | ||
3465 | schedstat_inc(rq, sched_cnt); | |
3466 | now = sched_clock(); | |
238628ed | 3467 | if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) { |
1da177e4 | 3468 | run_time = now - prev->timestamp; |
238628ed | 3469 | if (unlikely((long long)(now - prev->timestamp) < 0)) |
1da177e4 LT |
3470 | run_time = 0; |
3471 | } else | |
3472 | run_time = NS_MAX_SLEEP_AVG; | |
3473 | ||
3474 | /* | |
3475 | * Tasks charged proportionately less run_time at high sleep_avg to | |
3476 | * delay them losing their interactive status | |
3477 | */ | |
3478 | run_time /= (CURRENT_BONUS(prev) ? : 1); | |
3479 | ||
3480 | spin_lock_irq(&rq->lock); | |
3481 | ||
1da177e4 LT |
3482 | switch_count = &prev->nivcsw; |
3483 | if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { | |
3484 | switch_count = &prev->nvcsw; | |
3485 | if (unlikely((prev->state & TASK_INTERRUPTIBLE) && | |
3486 | unlikely(signal_pending(prev)))) | |
3487 | prev->state = TASK_RUNNING; | |
3488 | else { | |
3489 | if (prev->state == TASK_UNINTERRUPTIBLE) | |
3490 | rq->nr_uninterruptible++; | |
3491 | deactivate_task(prev, rq); | |
3492 | } | |
3493 | } | |
3494 | ||
3495 | cpu = smp_processor_id(); | |
3496 | if (unlikely(!rq->nr_running)) { | |
1da177e4 LT |
3497 | idle_balance(cpu, rq); |
3498 | if (!rq->nr_running) { | |
3499 | next = rq->idle; | |
3500 | rq->expired_timestamp = 0; | |
c96d145e | 3501 | wake_sleeping_dependent(cpu); |
1da177e4 LT |
3502 | goto switch_tasks; |
3503 | } | |
1da177e4 LT |
3504 | } |
3505 | ||
3506 | array = rq->active; | |
3507 | if (unlikely(!array->nr_active)) { | |
3508 | /* | |
3509 | * Switch the active and expired arrays. | |
3510 | */ | |
3511 | schedstat_inc(rq, sched_switch); | |
3512 | rq->active = rq->expired; | |
3513 | rq->expired = array; | |
3514 | array = rq->active; | |
3515 | rq->expired_timestamp = 0; | |
3516 | rq->best_expired_prio = MAX_PRIO; | |
3517 | } | |
3518 | ||
3519 | idx = sched_find_first_bit(array->bitmap); | |
3520 | queue = array->queue + idx; | |
36c8b586 | 3521 | next = list_entry(queue->next, struct task_struct, run_list); |
1da177e4 | 3522 | |
3dee386e | 3523 | if (!rt_task(next) && interactive_sleep(next->sleep_type)) { |
1da177e4 | 3524 | unsigned long long delta = now - next->timestamp; |
238628ed | 3525 | if (unlikely((long long)(now - next->timestamp) < 0)) |
1da177e4 LT |
3526 | delta = 0; |
3527 | ||
3dee386e | 3528 | if (next->sleep_type == SLEEP_INTERACTIVE) |
1da177e4 LT |
3529 | delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128; |
3530 | ||
3531 | array = next->array; | |
a3464a10 CS |
3532 | new_prio = recalc_task_prio(next, next->timestamp + delta); |
3533 | ||
3534 | if (unlikely(next->prio != new_prio)) { | |
3535 | dequeue_task(next, array); | |
3536 | next->prio = new_prio; | |
3537 | enqueue_task(next, array); | |
7c4bb1f9 | 3538 | } |
1da177e4 | 3539 | } |
3dee386e | 3540 | next->sleep_type = SLEEP_NORMAL; |
c96d145e KC |
3541 | if (dependent_sleeper(cpu, rq, next)) |
3542 | next = rq->idle; | |
1da177e4 LT |
3543 | switch_tasks: |
3544 | if (next == rq->idle) | |
3545 | schedstat_inc(rq, sched_goidle); | |
3546 | prefetch(next); | |
383f2835 | 3547 | prefetch_stack(next); |
1da177e4 LT |
3548 | clear_tsk_need_resched(prev); |
3549 | rcu_qsctr_inc(task_cpu(prev)); | |
3550 | ||
3551 | update_cpu_clock(prev, rq, now); | |
3552 | ||
3553 | prev->sleep_avg -= run_time; | |
3554 | if ((long)prev->sleep_avg <= 0) | |
3555 | prev->sleep_avg = 0; | |
3556 | prev->timestamp = prev->last_ran = now; | |
3557 | ||
3558 | sched_info_switch(prev, next); | |
3559 | if (likely(prev != next)) { | |
3560 | next->timestamp = now; | |
3561 | rq->nr_switches++; | |
3562 | rq->curr = next; | |
3563 | ++*switch_count; | |
3564 | ||
4866cde0 | 3565 | prepare_task_switch(rq, next); |
1da177e4 LT |
3566 | prev = context_switch(rq, prev, next); |
3567 | barrier(); | |
4866cde0 NP |
3568 | /* |
3569 | * this_rq must be evaluated again because prev may have moved | |
3570 | * CPUs since it called schedule(), thus the 'rq' on its stack | |
3571 | * frame will be invalid. | |
3572 | */ | |
3573 | finish_task_switch(this_rq(), prev); | |
1da177e4 LT |
3574 | } else |
3575 | spin_unlock_irq(&rq->lock); | |
3576 | ||
3577 | prev = current; | |
3578 | if (unlikely(reacquire_kernel_lock(prev) < 0)) | |
3579 | goto need_resched_nonpreemptible; | |
3580 | preempt_enable_no_resched(); | |
3581 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
3582 | goto need_resched; | |
3583 | } | |
1da177e4 LT |
3584 | EXPORT_SYMBOL(schedule); |
3585 | ||
3586 | #ifdef CONFIG_PREEMPT | |
3587 | /* | |
2ed6e34f | 3588 | * this is the entry point to schedule() from in-kernel preemption |
1da177e4 LT |
3589 | * off of preempt_enable. Kernel preemptions off return from interrupt |
3590 | * occur there and call schedule directly. | |
3591 | */ | |
3592 | asmlinkage void __sched preempt_schedule(void) | |
3593 | { | |
3594 | struct thread_info *ti = current_thread_info(); | |
3595 | #ifdef CONFIG_PREEMPT_BKL | |
3596 | struct task_struct *task = current; | |
3597 | int saved_lock_depth; | |
3598 | #endif | |
3599 | /* | |
3600 | * If there is a non-zero preempt_count or interrupts are disabled, | |
3601 | * we do not want to preempt the current task. Just return.. | |
3602 | */ | |
beed33a8 | 3603 | if (likely(ti->preempt_count || irqs_disabled())) |
1da177e4 LT |
3604 | return; |
3605 | ||
3606 | need_resched: | |
3607 | add_preempt_count(PREEMPT_ACTIVE); | |
3608 | /* | |
3609 | * We keep the big kernel semaphore locked, but we | |
3610 | * clear ->lock_depth so that schedule() doesnt | |
3611 | * auto-release the semaphore: | |
3612 | */ | |
3613 | #ifdef CONFIG_PREEMPT_BKL | |
3614 | saved_lock_depth = task->lock_depth; | |
3615 | task->lock_depth = -1; | |
3616 | #endif | |
3617 | schedule(); | |
3618 | #ifdef CONFIG_PREEMPT_BKL | |
3619 | task->lock_depth = saved_lock_depth; | |
3620 | #endif | |
3621 | sub_preempt_count(PREEMPT_ACTIVE); | |
3622 | ||
3623 | /* we could miss a preemption opportunity between schedule and now */ | |
3624 | barrier(); | |
3625 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
3626 | goto need_resched; | |
3627 | } | |
1da177e4 LT |
3628 | EXPORT_SYMBOL(preempt_schedule); |
3629 | ||
3630 | /* | |
2ed6e34f | 3631 | * this is the entry point to schedule() from kernel preemption |
1da177e4 LT |
3632 | * off of irq context. |
3633 | * Note, that this is called and return with irqs disabled. This will | |
3634 | * protect us against recursive calling from irq. | |
3635 | */ | |
3636 | asmlinkage void __sched preempt_schedule_irq(void) | |
3637 | { | |
3638 | struct thread_info *ti = current_thread_info(); | |
3639 | #ifdef CONFIG_PREEMPT_BKL | |
3640 | struct task_struct *task = current; | |
3641 | int saved_lock_depth; | |
3642 | #endif | |
2ed6e34f | 3643 | /* Catch callers which need to be fixed */ |
1da177e4 LT |
3644 | BUG_ON(ti->preempt_count || !irqs_disabled()); |
3645 | ||
3646 | need_resched: | |
3647 | add_preempt_count(PREEMPT_ACTIVE); | |
3648 | /* | |
3649 | * We keep the big kernel semaphore locked, but we | |
3650 | * clear ->lock_depth so that schedule() doesnt | |
3651 | * auto-release the semaphore: | |
3652 | */ | |
3653 | #ifdef CONFIG_PREEMPT_BKL | |
3654 | saved_lock_depth = task->lock_depth; | |
3655 | task->lock_depth = -1; | |
3656 | #endif | |
3657 | local_irq_enable(); | |
3658 | schedule(); | |
3659 | local_irq_disable(); | |
3660 | #ifdef CONFIG_PREEMPT_BKL | |
3661 | task->lock_depth = saved_lock_depth; | |
3662 | #endif | |
3663 | sub_preempt_count(PREEMPT_ACTIVE); | |
3664 | ||
3665 | /* we could miss a preemption opportunity between schedule and now */ | |
3666 | barrier(); | |
3667 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
3668 | goto need_resched; | |
3669 | } | |
3670 | ||
3671 | #endif /* CONFIG_PREEMPT */ | |
3672 | ||
95cdf3b7 IM |
3673 | int default_wake_function(wait_queue_t *curr, unsigned mode, int sync, |
3674 | void *key) | |
1da177e4 | 3675 | { |
48f24c4d | 3676 | return try_to_wake_up(curr->private, mode, sync); |
1da177e4 | 3677 | } |
1da177e4 LT |
3678 | EXPORT_SYMBOL(default_wake_function); |
3679 | ||
3680 | /* | |
3681 | * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just | |
3682 | * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve | |
3683 | * number) then we wake all the non-exclusive tasks and one exclusive task. | |
3684 | * | |
3685 | * There are circumstances in which we can try to wake a task which has already | |
3686 | * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns | |
3687 | * zero in this (rare) case, and we handle it by continuing to scan the queue. | |
3688 | */ | |
3689 | static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, | |
3690 | int nr_exclusive, int sync, void *key) | |
3691 | { | |
3692 | struct list_head *tmp, *next; | |
3693 | ||
3694 | list_for_each_safe(tmp, next, &q->task_list) { | |
48f24c4d IM |
3695 | wait_queue_t *curr = list_entry(tmp, wait_queue_t, task_list); |
3696 | unsigned flags = curr->flags; | |
3697 | ||
1da177e4 | 3698 | if (curr->func(curr, mode, sync, key) && |
48f24c4d | 3699 | (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) |
1da177e4 LT |
3700 | break; |
3701 | } | |
3702 | } | |
3703 | ||
3704 | /** | |
3705 | * __wake_up - wake up threads blocked on a waitqueue. | |
3706 | * @q: the waitqueue | |
3707 | * @mode: which threads | |
3708 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
67be2dd1 | 3709 | * @key: is directly passed to the wakeup function |
1da177e4 LT |
3710 | */ |
3711 | void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode, | |
95cdf3b7 | 3712 | int nr_exclusive, void *key) |
1da177e4 LT |
3713 | { |
3714 | unsigned long flags; | |
3715 | ||
3716 | spin_lock_irqsave(&q->lock, flags); | |
3717 | __wake_up_common(q, mode, nr_exclusive, 0, key); | |
3718 | spin_unlock_irqrestore(&q->lock, flags); | |
3719 | } | |
1da177e4 LT |
3720 | EXPORT_SYMBOL(__wake_up); |
3721 | ||
3722 | /* | |
3723 | * Same as __wake_up but called with the spinlock in wait_queue_head_t held. | |
3724 | */ | |
3725 | void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode) | |
3726 | { | |
3727 | __wake_up_common(q, mode, 1, 0, NULL); | |
3728 | } | |
3729 | ||
3730 | /** | |
67be2dd1 | 3731 | * __wake_up_sync - wake up threads blocked on a waitqueue. |
1da177e4 LT |
3732 | * @q: the waitqueue |
3733 | * @mode: which threads | |
3734 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
3735 | * | |
3736 | * The sync wakeup differs that the waker knows that it will schedule | |
3737 | * away soon, so while the target thread will be woken up, it will not | |
3738 | * be migrated to another CPU - ie. the two threads are 'synchronized' | |
3739 | * with each other. This can prevent needless bouncing between CPUs. | |
3740 | * | |
3741 | * On UP it can prevent extra preemption. | |
3742 | */ | |
95cdf3b7 IM |
3743 | void fastcall |
3744 | __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) | |
1da177e4 LT |
3745 | { |
3746 | unsigned long flags; | |
3747 | int sync = 1; | |
3748 | ||
3749 | if (unlikely(!q)) | |
3750 | return; | |
3751 | ||
3752 | if (unlikely(!nr_exclusive)) | |
3753 | sync = 0; | |
3754 | ||
3755 | spin_lock_irqsave(&q->lock, flags); | |
3756 | __wake_up_common(q, mode, nr_exclusive, sync, NULL); | |
3757 | spin_unlock_irqrestore(&q->lock, flags); | |
3758 | } | |
3759 | EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ | |
3760 | ||
3761 | void fastcall complete(struct completion *x) | |
3762 | { | |
3763 | unsigned long flags; | |
3764 | ||
3765 | spin_lock_irqsave(&x->wait.lock, flags); | |
3766 | x->done++; | |
3767 | __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, | |
3768 | 1, 0, NULL); | |
3769 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
3770 | } | |
3771 | EXPORT_SYMBOL(complete); | |
3772 | ||
3773 | void fastcall complete_all(struct completion *x) | |
3774 | { | |
3775 | unsigned long flags; | |
3776 | ||
3777 | spin_lock_irqsave(&x->wait.lock, flags); | |
3778 | x->done += UINT_MAX/2; | |
3779 | __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, | |
3780 | 0, 0, NULL); | |
3781 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
3782 | } | |
3783 | EXPORT_SYMBOL(complete_all); | |
3784 | ||
3785 | void fastcall __sched wait_for_completion(struct completion *x) | |
3786 | { | |
3787 | might_sleep(); | |
48f24c4d | 3788 | |
1da177e4 LT |
3789 | spin_lock_irq(&x->wait.lock); |
3790 | if (!x->done) { | |
3791 | DECLARE_WAITQUEUE(wait, current); | |
3792 | ||
3793 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3794 | __add_wait_queue_tail(&x->wait, &wait); | |
3795 | do { | |
3796 | __set_current_state(TASK_UNINTERRUPTIBLE); | |
3797 | spin_unlock_irq(&x->wait.lock); | |
3798 | schedule(); | |
3799 | spin_lock_irq(&x->wait.lock); | |
3800 | } while (!x->done); | |
3801 | __remove_wait_queue(&x->wait, &wait); | |
3802 | } | |
3803 | x->done--; | |
3804 | spin_unlock_irq(&x->wait.lock); | |
3805 | } | |
3806 | EXPORT_SYMBOL(wait_for_completion); | |
3807 | ||
3808 | unsigned long fastcall __sched | |
3809 | wait_for_completion_timeout(struct completion *x, unsigned long timeout) | |
3810 | { | |
3811 | might_sleep(); | |
3812 | ||
3813 | spin_lock_irq(&x->wait.lock); | |
3814 | if (!x->done) { | |
3815 | DECLARE_WAITQUEUE(wait, current); | |
3816 | ||
3817 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3818 | __add_wait_queue_tail(&x->wait, &wait); | |
3819 | do { | |
3820 | __set_current_state(TASK_UNINTERRUPTIBLE); | |
3821 | spin_unlock_irq(&x->wait.lock); | |
3822 | timeout = schedule_timeout(timeout); | |
3823 | spin_lock_irq(&x->wait.lock); | |
3824 | if (!timeout) { | |
3825 | __remove_wait_queue(&x->wait, &wait); | |
3826 | goto out; | |
3827 | } | |
3828 | } while (!x->done); | |
3829 | __remove_wait_queue(&x->wait, &wait); | |
3830 | } | |
3831 | x->done--; | |
3832 | out: | |
3833 | spin_unlock_irq(&x->wait.lock); | |
3834 | return timeout; | |
3835 | } | |
3836 | EXPORT_SYMBOL(wait_for_completion_timeout); | |
3837 | ||
3838 | int fastcall __sched wait_for_completion_interruptible(struct completion *x) | |
3839 | { | |
3840 | int ret = 0; | |
3841 | ||
3842 | might_sleep(); | |
3843 | ||
3844 | spin_lock_irq(&x->wait.lock); | |
3845 | if (!x->done) { | |
3846 | DECLARE_WAITQUEUE(wait, current); | |
3847 | ||
3848 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3849 | __add_wait_queue_tail(&x->wait, &wait); | |
3850 | do { | |
3851 | if (signal_pending(current)) { | |
3852 | ret = -ERESTARTSYS; | |
3853 | __remove_wait_queue(&x->wait, &wait); | |
3854 | goto out; | |
3855 | } | |
3856 | __set_current_state(TASK_INTERRUPTIBLE); | |
3857 | spin_unlock_irq(&x->wait.lock); | |
3858 | schedule(); | |
3859 | spin_lock_irq(&x->wait.lock); | |
3860 | } while (!x->done); | |
3861 | __remove_wait_queue(&x->wait, &wait); | |
3862 | } | |
3863 | x->done--; | |
3864 | out: | |
3865 | spin_unlock_irq(&x->wait.lock); | |
3866 | ||
3867 | return ret; | |
3868 | } | |
3869 | EXPORT_SYMBOL(wait_for_completion_interruptible); | |
3870 | ||
3871 | unsigned long fastcall __sched | |
3872 | wait_for_completion_interruptible_timeout(struct completion *x, | |
3873 | unsigned long timeout) | |
3874 | { | |
3875 | might_sleep(); | |
3876 | ||
3877 | spin_lock_irq(&x->wait.lock); | |
3878 | if (!x->done) { | |
3879 | DECLARE_WAITQUEUE(wait, current); | |
3880 | ||
3881 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3882 | __add_wait_queue_tail(&x->wait, &wait); | |
3883 | do { | |
3884 | if (signal_pending(current)) { | |
3885 | timeout = -ERESTARTSYS; | |
3886 | __remove_wait_queue(&x->wait, &wait); | |
3887 | goto out; | |
3888 | } | |
3889 | __set_current_state(TASK_INTERRUPTIBLE); | |
3890 | spin_unlock_irq(&x->wait.lock); | |
3891 | timeout = schedule_timeout(timeout); | |
3892 | spin_lock_irq(&x->wait.lock); | |
3893 | if (!timeout) { | |
3894 | __remove_wait_queue(&x->wait, &wait); | |
3895 | goto out; | |
3896 | } | |
3897 | } while (!x->done); | |
3898 | __remove_wait_queue(&x->wait, &wait); | |
3899 | } | |
3900 | x->done--; | |
3901 | out: | |
3902 | spin_unlock_irq(&x->wait.lock); | |
3903 | return timeout; | |
3904 | } | |
3905 | EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); | |
3906 | ||
3907 | ||
3908 | #define SLEEP_ON_VAR \ | |
3909 | unsigned long flags; \ | |
3910 | wait_queue_t wait; \ | |
3911 | init_waitqueue_entry(&wait, current); | |
3912 | ||
3913 | #define SLEEP_ON_HEAD \ | |
3914 | spin_lock_irqsave(&q->lock,flags); \ | |
3915 | __add_wait_queue(q, &wait); \ | |
3916 | spin_unlock(&q->lock); | |
3917 | ||
3918 | #define SLEEP_ON_TAIL \ | |
3919 | spin_lock_irq(&q->lock); \ | |
3920 | __remove_wait_queue(q, &wait); \ | |
3921 | spin_unlock_irqrestore(&q->lock, flags); | |
3922 | ||
3923 | void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q) | |
3924 | { | |
3925 | SLEEP_ON_VAR | |
3926 | ||
3927 | current->state = TASK_INTERRUPTIBLE; | |
3928 | ||
3929 | SLEEP_ON_HEAD | |
3930 | schedule(); | |
3931 | SLEEP_ON_TAIL | |
3932 | } | |
1da177e4 LT |
3933 | EXPORT_SYMBOL(interruptible_sleep_on); |
3934 | ||
95cdf3b7 IM |
3935 | long fastcall __sched |
3936 | interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) | |
1da177e4 LT |
3937 | { |
3938 | SLEEP_ON_VAR | |
3939 | ||
3940 | current->state = TASK_INTERRUPTIBLE; | |
3941 | ||
3942 | SLEEP_ON_HEAD | |
3943 | timeout = schedule_timeout(timeout); | |
3944 | SLEEP_ON_TAIL | |
3945 | ||
3946 | return timeout; | |
3947 | } | |
1da177e4 LT |
3948 | EXPORT_SYMBOL(interruptible_sleep_on_timeout); |
3949 | ||
3950 | void fastcall __sched sleep_on(wait_queue_head_t *q) | |
3951 | { | |
3952 | SLEEP_ON_VAR | |
3953 | ||
3954 | current->state = TASK_UNINTERRUPTIBLE; | |
3955 | ||
3956 | SLEEP_ON_HEAD | |
3957 | schedule(); | |
3958 | SLEEP_ON_TAIL | |
3959 | } | |
1da177e4 LT |
3960 | EXPORT_SYMBOL(sleep_on); |
3961 | ||
3962 | long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) | |
3963 | { | |
3964 | SLEEP_ON_VAR | |
3965 | ||
3966 | current->state = TASK_UNINTERRUPTIBLE; | |
3967 | ||
3968 | SLEEP_ON_HEAD | |
3969 | timeout = schedule_timeout(timeout); | |
3970 | SLEEP_ON_TAIL | |
3971 | ||
3972 | return timeout; | |
3973 | } | |
3974 | ||
3975 | EXPORT_SYMBOL(sleep_on_timeout); | |
3976 | ||
b29739f9 IM |
3977 | #ifdef CONFIG_RT_MUTEXES |
3978 | ||
3979 | /* | |
3980 | * rt_mutex_setprio - set the current priority of a task | |
3981 | * @p: task | |
3982 | * @prio: prio value (kernel-internal form) | |
3983 | * | |
3984 | * This function changes the 'effective' priority of a task. It does | |
3985 | * not touch ->normal_prio like __setscheduler(). | |
3986 | * | |
3987 | * Used by the rt_mutex code to implement priority inheritance logic. | |
3988 | */ | |
36c8b586 | 3989 | void rt_mutex_setprio(struct task_struct *p, int prio) |
b29739f9 | 3990 | { |
70b97a7f | 3991 | struct prio_array *array; |
b29739f9 | 3992 | unsigned long flags; |
70b97a7f | 3993 | struct rq *rq; |
b29739f9 IM |
3994 | int oldprio; |
3995 | ||
3996 | BUG_ON(prio < 0 || prio > MAX_PRIO); | |
3997 | ||
3998 | rq = task_rq_lock(p, &flags); | |
3999 | ||
4000 | oldprio = p->prio; | |
4001 | array = p->array; | |
4002 | if (array) | |
4003 | dequeue_task(p, array); | |
4004 | p->prio = prio; | |
4005 | ||
4006 | if (array) { | |
4007 | /* | |
4008 | * If changing to an RT priority then queue it | |
4009 | * in the active array! | |
4010 | */ | |
4011 | if (rt_task(p)) | |
4012 | array = rq->active; | |
4013 | enqueue_task(p, array); | |
4014 | /* | |
4015 | * Reschedule if we are currently running on this runqueue and | |
4016 | * our priority decreased, or if we are not currently running on | |
4017 | * this runqueue and our priority is higher than the current's | |
4018 | */ | |
4019 | if (task_running(rq, p)) { | |
4020 | if (p->prio > oldprio) | |
4021 | resched_task(rq->curr); | |
4022 | } else if (TASK_PREEMPTS_CURR(p, rq)) | |
4023 | resched_task(rq->curr); | |
4024 | } | |
4025 | task_rq_unlock(rq, &flags); | |
4026 | } | |
4027 | ||
4028 | #endif | |
4029 | ||
36c8b586 | 4030 | void set_user_nice(struct task_struct *p, long nice) |
1da177e4 | 4031 | { |
70b97a7f | 4032 | struct prio_array *array; |
48f24c4d | 4033 | int old_prio, delta; |
1da177e4 | 4034 | unsigned long flags; |
70b97a7f | 4035 | struct rq *rq; |
1da177e4 LT |
4036 | |
4037 | if (TASK_NICE(p) == nice || nice < -20 || nice > 19) | |
4038 | return; | |
4039 | /* | |
4040 | * We have to be careful, if called from sys_setpriority(), | |
4041 | * the task might be in the middle of scheduling on another CPU. | |
4042 | */ | |
4043 | rq = task_rq_lock(p, &flags); | |
4044 | /* | |
4045 | * The RT priorities are set via sched_setscheduler(), but we still | |
4046 | * allow the 'normal' nice value to be set - but as expected | |
4047 | * it wont have any effect on scheduling until the task is | |
b0a9499c | 4048 | * not SCHED_NORMAL/SCHED_BATCH: |
1da177e4 | 4049 | */ |
b29739f9 | 4050 | if (has_rt_policy(p)) { |
1da177e4 LT |
4051 | p->static_prio = NICE_TO_PRIO(nice); |
4052 | goto out_unlock; | |
4053 | } | |
4054 | array = p->array; | |
2dd73a4f | 4055 | if (array) { |
1da177e4 | 4056 | dequeue_task(p, array); |
2dd73a4f PW |
4057 | dec_raw_weighted_load(rq, p); |
4058 | } | |
1da177e4 | 4059 | |
1da177e4 | 4060 | p->static_prio = NICE_TO_PRIO(nice); |
2dd73a4f | 4061 | set_load_weight(p); |
b29739f9 IM |
4062 | old_prio = p->prio; |
4063 | p->prio = effective_prio(p); | |
4064 | delta = p->prio - old_prio; | |
1da177e4 LT |
4065 | |
4066 | if (array) { | |
4067 | enqueue_task(p, array); | |
2dd73a4f | 4068 | inc_raw_weighted_load(rq, p); |
1da177e4 LT |
4069 | /* |
4070 | * If the task increased its priority or is running and | |
4071 | * lowered its priority, then reschedule its CPU: | |
4072 | */ | |
4073 | if (delta < 0 || (delta > 0 && task_running(rq, p))) | |
4074 | resched_task(rq->curr); | |
4075 | } | |
4076 | out_unlock: | |
4077 | task_rq_unlock(rq, &flags); | |
4078 | } | |
1da177e4 LT |
4079 | EXPORT_SYMBOL(set_user_nice); |
4080 | ||
e43379f1 MM |
4081 | /* |
4082 | * can_nice - check if a task can reduce its nice value | |
4083 | * @p: task | |
4084 | * @nice: nice value | |
4085 | */ | |
36c8b586 | 4086 | int can_nice(const struct task_struct *p, const int nice) |
e43379f1 | 4087 | { |
024f4747 MM |
4088 | /* convert nice value [19,-20] to rlimit style value [1,40] */ |
4089 | int nice_rlim = 20 - nice; | |
48f24c4d | 4090 | |
e43379f1 MM |
4091 | return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur || |
4092 | capable(CAP_SYS_NICE)); | |
4093 | } | |
4094 | ||
1da177e4 LT |
4095 | #ifdef __ARCH_WANT_SYS_NICE |
4096 | ||
4097 | /* | |
4098 | * sys_nice - change the priority of the current process. | |
4099 | * @increment: priority increment | |
4100 | * | |
4101 | * sys_setpriority is a more generic, but much slower function that | |
4102 | * does similar things. | |
4103 | */ | |
4104 | asmlinkage long sys_nice(int increment) | |
4105 | { | |
48f24c4d | 4106 | long nice, retval; |
1da177e4 LT |
4107 | |
4108 | /* | |
4109 | * Setpriority might change our priority at the same moment. | |
4110 | * We don't have to worry. Conceptually one call occurs first | |
4111 | * and we have a single winner. | |
4112 | */ | |
e43379f1 MM |
4113 | if (increment < -40) |
4114 | increment = -40; | |
1da177e4 LT |
4115 | if (increment > 40) |
4116 | increment = 40; | |
4117 | ||
4118 | nice = PRIO_TO_NICE(current->static_prio) + increment; | |
4119 | if (nice < -20) | |
4120 | nice = -20; | |
4121 | if (nice > 19) | |
4122 | nice = 19; | |
4123 | ||
e43379f1 MM |
4124 | if (increment < 0 && !can_nice(current, nice)) |
4125 | return -EPERM; | |
4126 | ||
1da177e4 LT |
4127 | retval = security_task_setnice(current, nice); |
4128 | if (retval) | |
4129 | return retval; | |
4130 | ||
4131 | set_user_nice(current, nice); | |
4132 | return 0; | |
4133 | } | |
4134 | ||
4135 | #endif | |
4136 | ||
4137 | /** | |
4138 | * task_prio - return the priority value of a given task. | |
4139 | * @p: the task in question. | |
4140 | * | |
4141 | * This is the priority value as seen by users in /proc. | |
4142 | * RT tasks are offset by -200. Normal tasks are centered | |
4143 | * around 0, value goes from -16 to +15. | |
4144 | */ | |
36c8b586 | 4145 | int task_prio(const struct task_struct *p) |
1da177e4 LT |
4146 | { |
4147 | return p->prio - MAX_RT_PRIO; | |
4148 | } | |
4149 | ||
4150 | /** | |
4151 | * task_nice - return the nice value of a given task. | |
4152 | * @p: the task in question. | |
4153 | */ | |
36c8b586 | 4154 | int task_nice(const struct task_struct *p) |
1da177e4 LT |
4155 | { |
4156 | return TASK_NICE(p); | |
4157 | } | |
1da177e4 | 4158 | EXPORT_SYMBOL_GPL(task_nice); |
1da177e4 LT |
4159 | |
4160 | /** | |
4161 | * idle_cpu - is a given cpu idle currently? | |
4162 | * @cpu: the processor in question. | |
4163 | */ | |
4164 | int idle_cpu(int cpu) | |
4165 | { | |
4166 | return cpu_curr(cpu) == cpu_rq(cpu)->idle; | |
4167 | } | |
4168 | ||
1da177e4 LT |
4169 | /** |
4170 | * idle_task - return the idle task for a given cpu. | |
4171 | * @cpu: the processor in question. | |
4172 | */ | |
36c8b586 | 4173 | struct task_struct *idle_task(int cpu) |
1da177e4 LT |
4174 | { |
4175 | return cpu_rq(cpu)->idle; | |
4176 | } | |
4177 | ||
4178 | /** | |
4179 | * find_process_by_pid - find a process with a matching PID value. | |
4180 | * @pid: the pid in question. | |
4181 | */ | |
36c8b586 | 4182 | static inline struct task_struct *find_process_by_pid(pid_t pid) |
1da177e4 LT |
4183 | { |
4184 | return pid ? find_task_by_pid(pid) : current; | |
4185 | } | |
4186 | ||
4187 | /* Actually do priority change: must hold rq lock. */ | |
4188 | static void __setscheduler(struct task_struct *p, int policy, int prio) | |
4189 | { | |
4190 | BUG_ON(p->array); | |
48f24c4d | 4191 | |
1da177e4 LT |
4192 | p->policy = policy; |
4193 | p->rt_priority = prio; | |
b29739f9 IM |
4194 | p->normal_prio = normal_prio(p); |
4195 | /* we are holding p->pi_lock already */ | |
4196 | p->prio = rt_mutex_getprio(p); | |
4197 | /* | |
4198 | * SCHED_BATCH tasks are treated as perpetual CPU hogs: | |
4199 | */ | |
4200 | if (policy == SCHED_BATCH) | |
4201 | p->sleep_avg = 0; | |
2dd73a4f | 4202 | set_load_weight(p); |
1da177e4 LT |
4203 | } |
4204 | ||
4205 | /** | |
4206 | * sched_setscheduler - change the scheduling policy and/or RT priority of | |
4207 | * a thread. | |
4208 | * @p: the task in question. | |
4209 | * @policy: new policy. | |
4210 | * @param: structure containing the new RT priority. | |
5fe1d75f ON |
4211 | * |
4212 | * NOTE: the task may be already dead | |
1da177e4 | 4213 | */ |
95cdf3b7 IM |
4214 | int sched_setscheduler(struct task_struct *p, int policy, |
4215 | struct sched_param *param) | |
1da177e4 | 4216 | { |
48f24c4d | 4217 | int retval, oldprio, oldpolicy = -1; |
70b97a7f | 4218 | struct prio_array *array; |
1da177e4 | 4219 | unsigned long flags; |
70b97a7f | 4220 | struct rq *rq; |
1da177e4 | 4221 | |
66e5393a SR |
4222 | /* may grab non-irq protected spin_locks */ |
4223 | BUG_ON(in_interrupt()); | |
1da177e4 LT |
4224 | recheck: |
4225 | /* double check policy once rq lock held */ | |
4226 | if (policy < 0) | |
4227 | policy = oldpolicy = p->policy; | |
4228 | else if (policy != SCHED_FIFO && policy != SCHED_RR && | |
b0a9499c IM |
4229 | policy != SCHED_NORMAL && policy != SCHED_BATCH) |
4230 | return -EINVAL; | |
1da177e4 LT |
4231 | /* |
4232 | * Valid priorities for SCHED_FIFO and SCHED_RR are | |
b0a9499c IM |
4233 | * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and |
4234 | * SCHED_BATCH is 0. | |
1da177e4 LT |
4235 | */ |
4236 | if (param->sched_priority < 0 || | |
95cdf3b7 | 4237 | (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || |
d46523ea | 4238 | (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) |
1da177e4 | 4239 | return -EINVAL; |
57a6f51c | 4240 | if (is_rt_policy(policy) != (param->sched_priority != 0)) |
1da177e4 LT |
4241 | return -EINVAL; |
4242 | ||
37e4ab3f OC |
4243 | /* |
4244 | * Allow unprivileged RT tasks to decrease priority: | |
4245 | */ | |
4246 | if (!capable(CAP_SYS_NICE)) { | |
8dc3e909 ON |
4247 | if (is_rt_policy(policy)) { |
4248 | unsigned long rlim_rtprio; | |
4249 | unsigned long flags; | |
4250 | ||
4251 | if (!lock_task_sighand(p, &flags)) | |
4252 | return -ESRCH; | |
4253 | rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur; | |
4254 | unlock_task_sighand(p, &flags); | |
4255 | ||
4256 | /* can't set/change the rt policy */ | |
4257 | if (policy != p->policy && !rlim_rtprio) | |
4258 | return -EPERM; | |
4259 | ||
4260 | /* can't increase priority */ | |
4261 | if (param->sched_priority > p->rt_priority && | |
4262 | param->sched_priority > rlim_rtprio) | |
4263 | return -EPERM; | |
4264 | } | |
5fe1d75f | 4265 | |
37e4ab3f OC |
4266 | /* can't change other user's priorities */ |
4267 | if ((current->euid != p->euid) && | |
4268 | (current->euid != p->uid)) | |
4269 | return -EPERM; | |
4270 | } | |
1da177e4 LT |
4271 | |
4272 | retval = security_task_setscheduler(p, policy, param); | |
4273 | if (retval) | |
4274 | return retval; | |
b29739f9 IM |
4275 | /* |
4276 | * make sure no PI-waiters arrive (or leave) while we are | |
4277 | * changing the priority of the task: | |
4278 | */ | |
4279 | spin_lock_irqsave(&p->pi_lock, flags); | |
1da177e4 LT |
4280 | /* |
4281 | * To be able to change p->policy safely, the apropriate | |
4282 | * runqueue lock must be held. | |
4283 | */ | |
b29739f9 | 4284 | rq = __task_rq_lock(p); |
1da177e4 LT |
4285 | /* recheck policy now with rq lock held */ |
4286 | if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { | |
4287 | policy = oldpolicy = -1; | |
b29739f9 IM |
4288 | __task_rq_unlock(rq); |
4289 | spin_unlock_irqrestore(&p->pi_lock, flags); | |
1da177e4 LT |
4290 | goto recheck; |
4291 | } | |
4292 | array = p->array; | |
4293 | if (array) | |
4294 | deactivate_task(p, rq); | |
4295 | oldprio = p->prio; | |
4296 | __setscheduler(p, policy, param->sched_priority); | |
4297 | if (array) { | |
4298 | __activate_task(p, rq); | |
4299 | /* | |
4300 | * Reschedule if we are currently running on this runqueue and | |
4301 | * our priority decreased, or if we are not currently running on | |
4302 | * this runqueue and our priority is higher than the current's | |
4303 | */ | |
4304 | if (task_running(rq, p)) { | |
4305 | if (p->prio > oldprio) | |
4306 | resched_task(rq->curr); | |
4307 | } else if (TASK_PREEMPTS_CURR(p, rq)) | |
4308 | resched_task(rq->curr); | |
4309 | } | |
b29739f9 IM |
4310 | __task_rq_unlock(rq); |
4311 | spin_unlock_irqrestore(&p->pi_lock, flags); | |
4312 | ||
95e02ca9 TG |
4313 | rt_mutex_adjust_pi(p); |
4314 | ||
1da177e4 LT |
4315 | return 0; |
4316 | } | |
4317 | EXPORT_SYMBOL_GPL(sched_setscheduler); | |
4318 | ||
95cdf3b7 IM |
4319 | static int |
4320 | do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) | |
1da177e4 | 4321 | { |
1da177e4 LT |
4322 | struct sched_param lparam; |
4323 | struct task_struct *p; | |
36c8b586 | 4324 | int retval; |
1da177e4 LT |
4325 | |
4326 | if (!param || pid < 0) | |
4327 | return -EINVAL; | |
4328 | if (copy_from_user(&lparam, param, sizeof(struct sched_param))) | |
4329 | return -EFAULT; | |
5fe1d75f ON |
4330 | |
4331 | rcu_read_lock(); | |
4332 | retval = -ESRCH; | |
1da177e4 | 4333 | p = find_process_by_pid(pid); |
5fe1d75f ON |
4334 | if (p != NULL) |
4335 | retval = sched_setscheduler(p, policy, &lparam); | |
4336 | rcu_read_unlock(); | |
36c8b586 | 4337 | |
1da177e4 LT |
4338 | return retval; |
4339 | } | |
4340 | ||
4341 | /** | |
4342 | * sys_sched_setscheduler - set/change the scheduler policy and RT priority | |
4343 | * @pid: the pid in question. | |
4344 | * @policy: new policy. | |
4345 | * @param: structure containing the new RT priority. | |
4346 | */ | |
4347 | asmlinkage long sys_sched_setscheduler(pid_t pid, int policy, | |
4348 | struct sched_param __user *param) | |
4349 | { | |
c21761f1 JB |
4350 | /* negative values for policy are not valid */ |
4351 | if (policy < 0) | |
4352 | return -EINVAL; | |
4353 | ||
1da177e4 LT |
4354 | return do_sched_setscheduler(pid, policy, param); |
4355 | } | |
4356 | ||
4357 | /** | |
4358 | * sys_sched_setparam - set/change the RT priority of a thread | |
4359 | * @pid: the pid in question. | |
4360 | * @param: structure containing the new RT priority. | |
4361 | */ | |
4362 | asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param) | |
4363 | { | |
4364 | return do_sched_setscheduler(pid, -1, param); | |
4365 | } | |
4366 | ||
4367 | /** | |
4368 | * sys_sched_getscheduler - get the policy (scheduling class) of a thread | |
4369 | * @pid: the pid in question. | |
4370 | */ | |
4371 | asmlinkage long sys_sched_getscheduler(pid_t pid) | |
4372 | { | |
36c8b586 | 4373 | struct task_struct *p; |
1da177e4 | 4374 | int retval = -EINVAL; |
1da177e4 LT |
4375 | |
4376 | if (pid < 0) | |
4377 | goto out_nounlock; | |
4378 | ||
4379 | retval = -ESRCH; | |
4380 | read_lock(&tasklist_lock); | |
4381 | p = find_process_by_pid(pid); | |
4382 | if (p) { | |
4383 | retval = security_task_getscheduler(p); | |
4384 | if (!retval) | |
4385 | retval = p->policy; | |
4386 | } | |
4387 | read_unlock(&tasklist_lock); | |
4388 | ||
4389 | out_nounlock: | |
4390 | return retval; | |
4391 | } | |
4392 | ||
4393 | /** | |
4394 | * sys_sched_getscheduler - get the RT priority of a thread | |
4395 | * @pid: the pid in question. | |
4396 | * @param: structure containing the RT priority. | |
4397 | */ | |
4398 | asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param) | |
4399 | { | |
4400 | struct sched_param lp; | |
36c8b586 | 4401 | struct task_struct *p; |
1da177e4 | 4402 | int retval = -EINVAL; |
1da177e4 LT |
4403 | |
4404 | if (!param || pid < 0) | |
4405 | goto out_nounlock; | |
4406 | ||
4407 | read_lock(&tasklist_lock); | |
4408 | p = find_process_by_pid(pid); | |
4409 | retval = -ESRCH; | |
4410 | if (!p) | |
4411 | goto out_unlock; | |
4412 | ||
4413 | retval = security_task_getscheduler(p); | |
4414 | if (retval) | |
4415 | goto out_unlock; | |
4416 | ||
4417 | lp.sched_priority = p->rt_priority; | |
4418 | read_unlock(&tasklist_lock); | |
4419 | ||
4420 | /* | |
4421 | * This one might sleep, we cannot do it with a spinlock held ... | |
4422 | */ | |
4423 | retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; | |
4424 | ||
4425 | out_nounlock: | |
4426 | return retval; | |
4427 | ||
4428 | out_unlock: | |
4429 | read_unlock(&tasklist_lock); | |
4430 | return retval; | |
4431 | } | |
4432 | ||
4433 | long sched_setaffinity(pid_t pid, cpumask_t new_mask) | |
4434 | { | |
1da177e4 | 4435 | cpumask_t cpus_allowed; |
36c8b586 IM |
4436 | struct task_struct *p; |
4437 | int retval; | |
1da177e4 LT |
4438 | |
4439 | lock_cpu_hotplug(); | |
4440 | read_lock(&tasklist_lock); | |
4441 | ||
4442 | p = find_process_by_pid(pid); | |
4443 | if (!p) { | |
4444 | read_unlock(&tasklist_lock); | |
4445 | unlock_cpu_hotplug(); | |
4446 | return -ESRCH; | |
4447 | } | |
4448 | ||
4449 | /* | |
4450 | * It is not safe to call set_cpus_allowed with the | |
4451 | * tasklist_lock held. We will bump the task_struct's | |
4452 | * usage count and then drop tasklist_lock. | |
4453 | */ | |
4454 | get_task_struct(p); | |
4455 | read_unlock(&tasklist_lock); | |
4456 | ||
4457 | retval = -EPERM; | |
4458 | if ((current->euid != p->euid) && (current->euid != p->uid) && | |
4459 | !capable(CAP_SYS_NICE)) | |
4460 | goto out_unlock; | |
4461 | ||
e7834f8f DQ |
4462 | retval = security_task_setscheduler(p, 0, NULL); |
4463 | if (retval) | |
4464 | goto out_unlock; | |
4465 | ||
1da177e4 LT |
4466 | cpus_allowed = cpuset_cpus_allowed(p); |
4467 | cpus_and(new_mask, new_mask, cpus_allowed); | |
4468 | retval = set_cpus_allowed(p, new_mask); | |
4469 | ||
4470 | out_unlock: | |
4471 | put_task_struct(p); | |
4472 | unlock_cpu_hotplug(); | |
4473 | return retval; | |
4474 | } | |
4475 | ||
4476 | static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, | |
4477 | cpumask_t *new_mask) | |
4478 | { | |
4479 | if (len < sizeof(cpumask_t)) { | |
4480 | memset(new_mask, 0, sizeof(cpumask_t)); | |
4481 | } else if (len > sizeof(cpumask_t)) { | |
4482 | len = sizeof(cpumask_t); | |
4483 | } | |
4484 | return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; | |
4485 | } | |
4486 | ||
4487 | /** | |
4488 | * sys_sched_setaffinity - set the cpu affinity of a process | |
4489 | * @pid: pid of the process | |
4490 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
4491 | * @user_mask_ptr: user-space pointer to the new cpu mask | |
4492 | */ | |
4493 | asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len, | |
4494 | unsigned long __user *user_mask_ptr) | |
4495 | { | |
4496 | cpumask_t new_mask; | |
4497 | int retval; | |
4498 | ||
4499 | retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask); | |
4500 | if (retval) | |
4501 | return retval; | |
4502 | ||
4503 | return sched_setaffinity(pid, new_mask); | |
4504 | } | |
4505 | ||
4506 | /* | |
4507 | * Represents all cpu's present in the system | |
4508 | * In systems capable of hotplug, this map could dynamically grow | |
4509 | * as new cpu's are detected in the system via any platform specific | |
4510 | * method, such as ACPI for e.g. | |
4511 | */ | |
4512 | ||
4cef0c61 | 4513 | cpumask_t cpu_present_map __read_mostly; |
1da177e4 LT |
4514 | EXPORT_SYMBOL(cpu_present_map); |
4515 | ||
4516 | #ifndef CONFIG_SMP | |
4cef0c61 | 4517 | cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL; |
e16b38f7 GB |
4518 | EXPORT_SYMBOL(cpu_online_map); |
4519 | ||
4cef0c61 | 4520 | cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL; |
e16b38f7 | 4521 | EXPORT_SYMBOL(cpu_possible_map); |
1da177e4 LT |
4522 | #endif |
4523 | ||
4524 | long sched_getaffinity(pid_t pid, cpumask_t *mask) | |
4525 | { | |
36c8b586 | 4526 | struct task_struct *p; |
1da177e4 | 4527 | int retval; |
1da177e4 LT |
4528 | |
4529 | lock_cpu_hotplug(); | |
4530 | read_lock(&tasklist_lock); | |
4531 | ||
4532 | retval = -ESRCH; | |
4533 | p = find_process_by_pid(pid); | |
4534 | if (!p) | |
4535 | goto out_unlock; | |
4536 | ||
e7834f8f DQ |
4537 | retval = security_task_getscheduler(p); |
4538 | if (retval) | |
4539 | goto out_unlock; | |
4540 | ||
2f7016d9 | 4541 | cpus_and(*mask, p->cpus_allowed, cpu_online_map); |
1da177e4 LT |
4542 | |
4543 | out_unlock: | |
4544 | read_unlock(&tasklist_lock); | |
4545 | unlock_cpu_hotplug(); | |
4546 | if (retval) | |
4547 | return retval; | |
4548 | ||
4549 | return 0; | |
4550 | } | |
4551 | ||
4552 | /** | |
4553 | * sys_sched_getaffinity - get the cpu affinity of a process | |
4554 | * @pid: pid of the process | |
4555 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
4556 | * @user_mask_ptr: user-space pointer to hold the current cpu mask | |
4557 | */ | |
4558 | asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len, | |
4559 | unsigned long __user *user_mask_ptr) | |
4560 | { | |
4561 | int ret; | |
4562 | cpumask_t mask; | |
4563 | ||
4564 | if (len < sizeof(cpumask_t)) | |
4565 | return -EINVAL; | |
4566 | ||
4567 | ret = sched_getaffinity(pid, &mask); | |
4568 | if (ret < 0) | |
4569 | return ret; | |
4570 | ||
4571 | if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t))) | |
4572 | return -EFAULT; | |
4573 | ||
4574 | return sizeof(cpumask_t); | |
4575 | } | |
4576 | ||
4577 | /** | |
4578 | * sys_sched_yield - yield the current processor to other threads. | |
4579 | * | |
4580 | * this function yields the current CPU by moving the calling thread | |
4581 | * to the expired array. If there are no other threads running on this | |
4582 | * CPU then this function will return. | |
4583 | */ | |
4584 | asmlinkage long sys_sched_yield(void) | |
4585 | { | |
70b97a7f IM |
4586 | struct rq *rq = this_rq_lock(); |
4587 | struct prio_array *array = current->array, *target = rq->expired; | |
1da177e4 LT |
4588 | |
4589 | schedstat_inc(rq, yld_cnt); | |
4590 | /* | |
4591 | * We implement yielding by moving the task into the expired | |
4592 | * queue. | |
4593 | * | |
4594 | * (special rule: RT tasks will just roundrobin in the active | |
4595 | * array.) | |
4596 | */ | |
4597 | if (rt_task(current)) | |
4598 | target = rq->active; | |
4599 | ||
5927ad78 | 4600 | if (array->nr_active == 1) { |
1da177e4 LT |
4601 | schedstat_inc(rq, yld_act_empty); |
4602 | if (!rq->expired->nr_active) | |
4603 | schedstat_inc(rq, yld_both_empty); | |
4604 | } else if (!rq->expired->nr_active) | |
4605 | schedstat_inc(rq, yld_exp_empty); | |
4606 | ||
4607 | if (array != target) { | |
4608 | dequeue_task(current, array); | |
4609 | enqueue_task(current, target); | |
4610 | } else | |
4611 | /* | |
4612 | * requeue_task is cheaper so perform that if possible. | |
4613 | */ | |
4614 | requeue_task(current, array); | |
4615 | ||
4616 | /* | |
4617 | * Since we are going to call schedule() anyway, there's | |
4618 | * no need to preempt or enable interrupts: | |
4619 | */ | |
4620 | __release(rq->lock); | |
8a25d5de | 4621 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); |
1da177e4 LT |
4622 | _raw_spin_unlock(&rq->lock); |
4623 | preempt_enable_no_resched(); | |
4624 | ||
4625 | schedule(); | |
4626 | ||
4627 | return 0; | |
4628 | } | |
4629 | ||
e7b38404 | 4630 | static void __cond_resched(void) |
1da177e4 | 4631 | { |
8e0a43d8 IM |
4632 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP |
4633 | __might_sleep(__FILE__, __LINE__); | |
4634 | #endif | |
5bbcfd90 IM |
4635 | /* |
4636 | * The BKS might be reacquired before we have dropped | |
4637 | * PREEMPT_ACTIVE, which could trigger a second | |
4638 | * cond_resched() call. | |
4639 | */ | |
1da177e4 LT |
4640 | do { |
4641 | add_preempt_count(PREEMPT_ACTIVE); | |
4642 | schedule(); | |
4643 | sub_preempt_count(PREEMPT_ACTIVE); | |
4644 | } while (need_resched()); | |
4645 | } | |
4646 | ||
4647 | int __sched cond_resched(void) | |
4648 | { | |
9414232f IM |
4649 | if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) && |
4650 | system_state == SYSTEM_RUNNING) { | |
1da177e4 LT |
4651 | __cond_resched(); |
4652 | return 1; | |
4653 | } | |
4654 | return 0; | |
4655 | } | |
1da177e4 LT |
4656 | EXPORT_SYMBOL(cond_resched); |
4657 | ||
4658 | /* | |
4659 | * cond_resched_lock() - if a reschedule is pending, drop the given lock, | |
4660 | * call schedule, and on return reacquire the lock. | |
4661 | * | |
4662 | * This works OK both with and without CONFIG_PREEMPT. We do strange low-level | |
4663 | * operations here to prevent schedule() from being called twice (once via | |
4664 | * spin_unlock(), once by hand). | |
4665 | */ | |
95cdf3b7 | 4666 | int cond_resched_lock(spinlock_t *lock) |
1da177e4 | 4667 | { |
6df3cecb JK |
4668 | int ret = 0; |
4669 | ||
1da177e4 LT |
4670 | if (need_lockbreak(lock)) { |
4671 | spin_unlock(lock); | |
4672 | cpu_relax(); | |
6df3cecb | 4673 | ret = 1; |
1da177e4 LT |
4674 | spin_lock(lock); |
4675 | } | |
9414232f | 4676 | if (need_resched() && system_state == SYSTEM_RUNNING) { |
8a25d5de | 4677 | spin_release(&lock->dep_map, 1, _THIS_IP_); |
1da177e4 LT |
4678 | _raw_spin_unlock(lock); |
4679 | preempt_enable_no_resched(); | |
4680 | __cond_resched(); | |
6df3cecb | 4681 | ret = 1; |
1da177e4 | 4682 | spin_lock(lock); |
1da177e4 | 4683 | } |
6df3cecb | 4684 | return ret; |
1da177e4 | 4685 | } |
1da177e4 LT |
4686 | EXPORT_SYMBOL(cond_resched_lock); |
4687 | ||
4688 | int __sched cond_resched_softirq(void) | |
4689 | { | |
4690 | BUG_ON(!in_softirq()); | |
4691 | ||
9414232f | 4692 | if (need_resched() && system_state == SYSTEM_RUNNING) { |
de30a2b3 IM |
4693 | raw_local_irq_disable(); |
4694 | _local_bh_enable(); | |
4695 | raw_local_irq_enable(); | |
1da177e4 LT |
4696 | __cond_resched(); |
4697 | local_bh_disable(); | |
4698 | return 1; | |
4699 | } | |
4700 | return 0; | |
4701 | } | |
1da177e4 LT |
4702 | EXPORT_SYMBOL(cond_resched_softirq); |
4703 | ||
1da177e4 LT |
4704 | /** |
4705 | * yield - yield the current processor to other threads. | |
4706 | * | |
4707 | * this is a shortcut for kernel-space yielding - it marks the | |
4708 | * thread runnable and calls sys_sched_yield(). | |
4709 | */ | |
4710 | void __sched yield(void) | |
4711 | { | |
4712 | set_current_state(TASK_RUNNING); | |
4713 | sys_sched_yield(); | |
4714 | } | |
1da177e4 LT |
4715 | EXPORT_SYMBOL(yield); |
4716 | ||
4717 | /* | |
4718 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so | |
4719 | * that process accounting knows that this is a task in IO wait state. | |
4720 | * | |
4721 | * But don't do that if it is a deliberate, throttling IO wait (this task | |
4722 | * has set its backing_dev_info: the queue against which it should throttle) | |
4723 | */ | |
4724 | void __sched io_schedule(void) | |
4725 | { | |
70b97a7f | 4726 | struct rq *rq = &__raw_get_cpu_var(runqueues); |
1da177e4 | 4727 | |
0ff92245 | 4728 | delayacct_blkio_start(); |
1da177e4 LT |
4729 | atomic_inc(&rq->nr_iowait); |
4730 | schedule(); | |
4731 | atomic_dec(&rq->nr_iowait); | |
0ff92245 | 4732 | delayacct_blkio_end(); |
1da177e4 | 4733 | } |
1da177e4 LT |
4734 | EXPORT_SYMBOL(io_schedule); |
4735 | ||
4736 | long __sched io_schedule_timeout(long timeout) | |
4737 | { | |
70b97a7f | 4738 | struct rq *rq = &__raw_get_cpu_var(runqueues); |
1da177e4 LT |
4739 | long ret; |
4740 | ||
0ff92245 | 4741 | delayacct_blkio_start(); |
1da177e4 LT |
4742 | atomic_inc(&rq->nr_iowait); |
4743 | ret = schedule_timeout(timeout); | |
4744 | atomic_dec(&rq->nr_iowait); | |
0ff92245 | 4745 | delayacct_blkio_end(); |
1da177e4 LT |
4746 | return ret; |
4747 | } | |
4748 | ||
4749 | /** | |
4750 | * sys_sched_get_priority_max - return maximum RT priority. | |
4751 | * @policy: scheduling class. | |
4752 | * | |
4753 | * this syscall returns the maximum rt_priority that can be used | |
4754 | * by a given scheduling class. | |
4755 | */ | |
4756 | asmlinkage long sys_sched_get_priority_max(int policy) | |
4757 | { | |
4758 | int ret = -EINVAL; | |
4759 | ||
4760 | switch (policy) { | |
4761 | case SCHED_FIFO: | |
4762 | case SCHED_RR: | |
4763 | ret = MAX_USER_RT_PRIO-1; | |
4764 | break; | |
4765 | case SCHED_NORMAL: | |
b0a9499c | 4766 | case SCHED_BATCH: |
1da177e4 LT |
4767 | ret = 0; |
4768 | break; | |
4769 | } | |
4770 | return ret; | |
4771 | } | |
4772 | ||
4773 | /** | |
4774 | * sys_sched_get_priority_min - return minimum RT priority. | |
4775 | * @policy: scheduling class. | |
4776 | * | |
4777 | * this syscall returns the minimum rt_priority that can be used | |
4778 | * by a given scheduling class. | |
4779 | */ | |
4780 | asmlinkage long sys_sched_get_priority_min(int policy) | |
4781 | { | |
4782 | int ret = -EINVAL; | |
4783 | ||
4784 | switch (policy) { | |
4785 | case SCHED_FIFO: | |
4786 | case SCHED_RR: | |
4787 | ret = 1; | |
4788 | break; | |
4789 | case SCHED_NORMAL: | |
b0a9499c | 4790 | case SCHED_BATCH: |
1da177e4 LT |
4791 | ret = 0; |
4792 | } | |
4793 | return ret; | |
4794 | } | |
4795 | ||
4796 | /** | |
4797 | * sys_sched_rr_get_interval - return the default timeslice of a process. | |
4798 | * @pid: pid of the process. | |
4799 | * @interval: userspace pointer to the timeslice value. | |
4800 | * | |
4801 | * this syscall writes the default timeslice value of a given process | |
4802 | * into the user-space timespec buffer. A value of '0' means infinity. | |
4803 | */ | |
4804 | asmlinkage | |
4805 | long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval) | |
4806 | { | |
36c8b586 | 4807 | struct task_struct *p; |
1da177e4 LT |
4808 | int retval = -EINVAL; |
4809 | struct timespec t; | |
1da177e4 LT |
4810 | |
4811 | if (pid < 0) | |
4812 | goto out_nounlock; | |
4813 | ||
4814 | retval = -ESRCH; | |
4815 | read_lock(&tasklist_lock); | |
4816 | p = find_process_by_pid(pid); | |
4817 | if (!p) | |
4818 | goto out_unlock; | |
4819 | ||
4820 | retval = security_task_getscheduler(p); | |
4821 | if (retval) | |
4822 | goto out_unlock; | |
4823 | ||
b78709cf | 4824 | jiffies_to_timespec(p->policy == SCHED_FIFO ? |
1da177e4 LT |
4825 | 0 : task_timeslice(p), &t); |
4826 | read_unlock(&tasklist_lock); | |
4827 | retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; | |
4828 | out_nounlock: | |
4829 | return retval; | |
4830 | out_unlock: | |
4831 | read_unlock(&tasklist_lock); | |
4832 | return retval; | |
4833 | } | |
4834 | ||
4835 | static inline struct task_struct *eldest_child(struct task_struct *p) | |
4836 | { | |
48f24c4d IM |
4837 | if (list_empty(&p->children)) |
4838 | return NULL; | |
1da177e4 LT |
4839 | return list_entry(p->children.next,struct task_struct,sibling); |
4840 | } | |
4841 | ||
4842 | static inline struct task_struct *older_sibling(struct task_struct *p) | |
4843 | { | |
48f24c4d IM |
4844 | if (p->sibling.prev==&p->parent->children) |
4845 | return NULL; | |
1da177e4 LT |
4846 | return list_entry(p->sibling.prev,struct task_struct,sibling); |
4847 | } | |
4848 | ||
4849 | static inline struct task_struct *younger_sibling(struct task_struct *p) | |
4850 | { | |
48f24c4d IM |
4851 | if (p->sibling.next==&p->parent->children) |
4852 | return NULL; | |
1da177e4 LT |
4853 | return list_entry(p->sibling.next,struct task_struct,sibling); |
4854 | } | |
4855 | ||
2ed6e34f | 4856 | static const char stat_nam[] = "RSDTtZX"; |
36c8b586 IM |
4857 | |
4858 | static void show_task(struct task_struct *p) | |
1da177e4 | 4859 | { |
36c8b586 | 4860 | struct task_struct *relative; |
1da177e4 | 4861 | unsigned long free = 0; |
36c8b586 | 4862 | unsigned state; |
1da177e4 | 4863 | |
1da177e4 | 4864 | state = p->state ? __ffs(p->state) + 1 : 0; |
2ed6e34f AM |
4865 | printk("%-13.13s %c", p->comm, |
4866 | state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); | |
1da177e4 LT |
4867 | #if (BITS_PER_LONG == 32) |
4868 | if (state == TASK_RUNNING) | |
4869 | printk(" running "); | |
4870 | else | |
4871 | printk(" %08lX ", thread_saved_pc(p)); | |
4872 | #else | |
4873 | if (state == TASK_RUNNING) | |
4874 | printk(" running task "); | |
4875 | else | |
4876 | printk(" %016lx ", thread_saved_pc(p)); | |
4877 | #endif | |
4878 | #ifdef CONFIG_DEBUG_STACK_USAGE | |
4879 | { | |
10ebffde | 4880 | unsigned long *n = end_of_stack(p); |
1da177e4 LT |
4881 | while (!*n) |
4882 | n++; | |
10ebffde | 4883 | free = (unsigned long)n - (unsigned long)end_of_stack(p); |
1da177e4 LT |
4884 | } |
4885 | #endif | |
4886 | printk("%5lu %5d %6d ", free, p->pid, p->parent->pid); | |
4887 | if ((relative = eldest_child(p))) | |
4888 | printk("%5d ", relative->pid); | |
4889 | else | |
4890 | printk(" "); | |
4891 | if ((relative = younger_sibling(p))) | |
4892 | printk("%7d", relative->pid); | |
4893 | else | |
4894 | printk(" "); | |
4895 | if ((relative = older_sibling(p))) | |
4896 | printk(" %5d", relative->pid); | |
4897 | else | |
4898 | printk(" "); | |
4899 | if (!p->mm) | |
4900 | printk(" (L-TLB)\n"); | |
4901 | else | |
4902 | printk(" (NOTLB)\n"); | |
4903 | ||
4904 | if (state != TASK_RUNNING) | |
4905 | show_stack(p, NULL); | |
4906 | } | |
4907 | ||
e59e2ae2 | 4908 | void show_state_filter(unsigned long state_filter) |
1da177e4 | 4909 | { |
36c8b586 | 4910 | struct task_struct *g, *p; |
1da177e4 LT |
4911 | |
4912 | #if (BITS_PER_LONG == 32) | |
4913 | printk("\n" | |
301827ac CC |
4914 | " free sibling\n"); |
4915 | printk(" task PC stack pid father child younger older\n"); | |
1da177e4 LT |
4916 | #else |
4917 | printk("\n" | |
301827ac CC |
4918 | " free sibling\n"); |
4919 | printk(" task PC stack pid father child younger older\n"); | |
1da177e4 LT |
4920 | #endif |
4921 | read_lock(&tasklist_lock); | |
4922 | do_each_thread(g, p) { | |
4923 | /* | |
4924 | * reset the NMI-timeout, listing all files on a slow | |
4925 | * console might take alot of time: | |
4926 | */ | |
4927 | touch_nmi_watchdog(); | |
e59e2ae2 IM |
4928 | if (p->state & state_filter) |
4929 | show_task(p); | |
1da177e4 LT |
4930 | } while_each_thread(g, p); |
4931 | ||
4932 | read_unlock(&tasklist_lock); | |
e59e2ae2 IM |
4933 | /* |
4934 | * Only show locks if all tasks are dumped: | |
4935 | */ | |
4936 | if (state_filter == -1) | |
4937 | debug_show_all_locks(); | |
1da177e4 LT |
4938 | } |
4939 | ||
f340c0d1 IM |
4940 | /** |
4941 | * init_idle - set up an idle thread for a given CPU | |
4942 | * @idle: task in question | |
4943 | * @cpu: cpu the idle task belongs to | |
4944 | * | |
4945 | * NOTE: this function does not set the idle thread's NEED_RESCHED | |
4946 | * flag, to make booting more robust. | |
4947 | */ | |
5c1e1767 | 4948 | void __cpuinit init_idle(struct task_struct *idle, int cpu) |
1da177e4 | 4949 | { |
70b97a7f | 4950 | struct rq *rq = cpu_rq(cpu); |
1da177e4 LT |
4951 | unsigned long flags; |
4952 | ||
81c29a85 | 4953 | idle->timestamp = sched_clock(); |
1da177e4 LT |
4954 | idle->sleep_avg = 0; |
4955 | idle->array = NULL; | |
b29739f9 | 4956 | idle->prio = idle->normal_prio = MAX_PRIO; |
1da177e4 LT |
4957 | idle->state = TASK_RUNNING; |
4958 | idle->cpus_allowed = cpumask_of_cpu(cpu); | |
4959 | set_task_cpu(idle, cpu); | |
4960 | ||
4961 | spin_lock_irqsave(&rq->lock, flags); | |
4962 | rq->curr = rq->idle = idle; | |
4866cde0 NP |
4963 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
4964 | idle->oncpu = 1; | |
4965 | #endif | |
1da177e4 LT |
4966 | spin_unlock_irqrestore(&rq->lock, flags); |
4967 | ||
4968 | /* Set the preempt count _outside_ the spinlocks! */ | |
4969 | #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL) | |
a1261f54 | 4970 | task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0); |
1da177e4 | 4971 | #else |
a1261f54 | 4972 | task_thread_info(idle)->preempt_count = 0; |
1da177e4 LT |
4973 | #endif |
4974 | } | |
4975 | ||
4976 | /* | |
4977 | * In a system that switches off the HZ timer nohz_cpu_mask | |
4978 | * indicates which cpus entered this state. This is used | |
4979 | * in the rcu update to wait only for active cpus. For system | |
4980 | * which do not switch off the HZ timer nohz_cpu_mask should | |
4981 | * always be CPU_MASK_NONE. | |
4982 | */ | |
4983 | cpumask_t nohz_cpu_mask = CPU_MASK_NONE; | |
4984 | ||
4985 | #ifdef CONFIG_SMP | |
4986 | /* | |
4987 | * This is how migration works: | |
4988 | * | |
70b97a7f | 4989 | * 1) we queue a struct migration_req structure in the source CPU's |
1da177e4 LT |
4990 | * runqueue and wake up that CPU's migration thread. |
4991 | * 2) we down() the locked semaphore => thread blocks. | |
4992 | * 3) migration thread wakes up (implicitly it forces the migrated | |
4993 | * thread off the CPU) | |
4994 | * 4) it gets the migration request and checks whether the migrated | |
4995 | * task is still in the wrong runqueue. | |
4996 | * 5) if it's in the wrong runqueue then the migration thread removes | |
4997 | * it and puts it into the right queue. | |
4998 | * 6) migration thread up()s the semaphore. | |
4999 | * 7) we wake up and the migration is done. | |
5000 | */ | |
5001 | ||
5002 | /* | |
5003 | * Change a given task's CPU affinity. Migrate the thread to a | |
5004 | * proper CPU and schedule it away if the CPU it's executing on | |
5005 | * is removed from the allowed bitmask. | |
5006 | * | |
5007 | * NOTE: the caller must have a valid reference to the task, the | |
5008 | * task must not exit() & deallocate itself prematurely. The | |
5009 | * call is not atomic; no spinlocks may be held. | |
5010 | */ | |
36c8b586 | 5011 | int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask) |
1da177e4 | 5012 | { |
70b97a7f | 5013 | struct migration_req req; |
1da177e4 | 5014 | unsigned long flags; |
70b97a7f | 5015 | struct rq *rq; |
48f24c4d | 5016 | int ret = 0; |
1da177e4 LT |
5017 | |
5018 | rq = task_rq_lock(p, &flags); | |
5019 | if (!cpus_intersects(new_mask, cpu_online_map)) { | |
5020 | ret = -EINVAL; | |
5021 | goto out; | |
5022 | } | |
5023 | ||
5024 | p->cpus_allowed = new_mask; | |
5025 | /* Can the task run on the task's current CPU? If so, we're done */ | |
5026 | if (cpu_isset(task_cpu(p), new_mask)) | |
5027 | goto out; | |
5028 | ||
5029 | if (migrate_task(p, any_online_cpu(new_mask), &req)) { | |
5030 | /* Need help from migration thread: drop lock and wait. */ | |
5031 | task_rq_unlock(rq, &flags); | |
5032 | wake_up_process(rq->migration_thread); | |
5033 | wait_for_completion(&req.done); | |
5034 | tlb_migrate_finish(p->mm); | |
5035 | return 0; | |
5036 | } | |
5037 | out: | |
5038 | task_rq_unlock(rq, &flags); | |
48f24c4d | 5039 | |
1da177e4 LT |
5040 | return ret; |
5041 | } | |
1da177e4 LT |
5042 | EXPORT_SYMBOL_GPL(set_cpus_allowed); |
5043 | ||
5044 | /* | |
5045 | * Move (not current) task off this cpu, onto dest cpu. We're doing | |
5046 | * this because either it can't run here any more (set_cpus_allowed() | |
5047 | * away from this CPU, or CPU going down), or because we're | |
5048 | * attempting to rebalance this task on exec (sched_exec). | |
5049 | * | |
5050 | * So we race with normal scheduler movements, but that's OK, as long | |
5051 | * as the task is no longer on this CPU. | |
efc30814 KK |
5052 | * |
5053 | * Returns non-zero if task was successfully migrated. | |
1da177e4 | 5054 | */ |
efc30814 | 5055 | static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) |
1da177e4 | 5056 | { |
70b97a7f | 5057 | struct rq *rq_dest, *rq_src; |
efc30814 | 5058 | int ret = 0; |
1da177e4 LT |
5059 | |
5060 | if (unlikely(cpu_is_offline(dest_cpu))) | |
efc30814 | 5061 | return ret; |
1da177e4 LT |
5062 | |
5063 | rq_src = cpu_rq(src_cpu); | |
5064 | rq_dest = cpu_rq(dest_cpu); | |
5065 | ||
5066 | double_rq_lock(rq_src, rq_dest); | |
5067 | /* Already moved. */ | |
5068 | if (task_cpu(p) != src_cpu) | |
5069 | goto out; | |
5070 | /* Affinity changed (again). */ | |
5071 | if (!cpu_isset(dest_cpu, p->cpus_allowed)) | |
5072 | goto out; | |
5073 | ||
5074 | set_task_cpu(p, dest_cpu); | |
5075 | if (p->array) { | |
5076 | /* | |
5077 | * Sync timestamp with rq_dest's before activating. | |
5078 | * The same thing could be achieved by doing this step | |
5079 | * afterwards, and pretending it was a local activate. | |
5080 | * This way is cleaner and logically correct. | |
5081 | */ | |
b18ec803 MG |
5082 | p->timestamp = p->timestamp - rq_src->most_recent_timestamp |
5083 | + rq_dest->most_recent_timestamp; | |
1da177e4 | 5084 | deactivate_task(p, rq_src); |
0a565f79 | 5085 | __activate_task(p, rq_dest); |
1da177e4 LT |
5086 | if (TASK_PREEMPTS_CURR(p, rq_dest)) |
5087 | resched_task(rq_dest->curr); | |
5088 | } | |
efc30814 | 5089 | ret = 1; |
1da177e4 LT |
5090 | out: |
5091 | double_rq_unlock(rq_src, rq_dest); | |
efc30814 | 5092 | return ret; |
1da177e4 LT |
5093 | } |
5094 | ||
5095 | /* | |
5096 | * migration_thread - this is a highprio system thread that performs | |
5097 | * thread migration by bumping thread off CPU then 'pushing' onto | |
5098 | * another runqueue. | |
5099 | */ | |
95cdf3b7 | 5100 | static int migration_thread(void *data) |
1da177e4 | 5101 | { |
1da177e4 | 5102 | int cpu = (long)data; |
70b97a7f | 5103 | struct rq *rq; |
1da177e4 LT |
5104 | |
5105 | rq = cpu_rq(cpu); | |
5106 | BUG_ON(rq->migration_thread != current); | |
5107 | ||
5108 | set_current_state(TASK_INTERRUPTIBLE); | |
5109 | while (!kthread_should_stop()) { | |
70b97a7f | 5110 | struct migration_req *req; |
1da177e4 | 5111 | struct list_head *head; |
1da177e4 | 5112 | |
3e1d1d28 | 5113 | try_to_freeze(); |
1da177e4 LT |
5114 | |
5115 | spin_lock_irq(&rq->lock); | |
5116 | ||
5117 | if (cpu_is_offline(cpu)) { | |
5118 | spin_unlock_irq(&rq->lock); | |
5119 | goto wait_to_die; | |
5120 | } | |
5121 | ||
5122 | if (rq->active_balance) { | |
5123 | active_load_balance(rq, cpu); | |
5124 | rq->active_balance = 0; | |
5125 | } | |
5126 | ||
5127 | head = &rq->migration_queue; | |
5128 | ||
5129 | if (list_empty(head)) { | |
5130 | spin_unlock_irq(&rq->lock); | |
5131 | schedule(); | |
5132 | set_current_state(TASK_INTERRUPTIBLE); | |
5133 | continue; | |
5134 | } | |
70b97a7f | 5135 | req = list_entry(head->next, struct migration_req, list); |
1da177e4 LT |
5136 | list_del_init(head->next); |
5137 | ||
674311d5 NP |
5138 | spin_unlock(&rq->lock); |
5139 | __migrate_task(req->task, cpu, req->dest_cpu); | |
5140 | local_irq_enable(); | |
1da177e4 LT |
5141 | |
5142 | complete(&req->done); | |
5143 | } | |
5144 | __set_current_state(TASK_RUNNING); | |
5145 | return 0; | |
5146 | ||
5147 | wait_to_die: | |
5148 | /* Wait for kthread_stop */ | |
5149 | set_current_state(TASK_INTERRUPTIBLE); | |
5150 | while (!kthread_should_stop()) { | |
5151 | schedule(); | |
5152 | set_current_state(TASK_INTERRUPTIBLE); | |
5153 | } | |
5154 | __set_current_state(TASK_RUNNING); | |
5155 | return 0; | |
5156 | } | |
5157 | ||
5158 | #ifdef CONFIG_HOTPLUG_CPU | |
054b9108 KK |
5159 | /* |
5160 | * Figure out where task on dead CPU should go, use force if neccessary. | |
5161 | * NOTE: interrupts should be disabled by the caller | |
5162 | */ | |
48f24c4d | 5163 | static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p) |
1da177e4 | 5164 | { |
efc30814 | 5165 | unsigned long flags; |
1da177e4 | 5166 | cpumask_t mask; |
70b97a7f IM |
5167 | struct rq *rq; |
5168 | int dest_cpu; | |
1da177e4 | 5169 | |
efc30814 | 5170 | restart: |
1da177e4 LT |
5171 | /* On same node? */ |
5172 | mask = node_to_cpumask(cpu_to_node(dead_cpu)); | |
48f24c4d | 5173 | cpus_and(mask, mask, p->cpus_allowed); |
1da177e4 LT |
5174 | dest_cpu = any_online_cpu(mask); |
5175 | ||
5176 | /* On any allowed CPU? */ | |
5177 | if (dest_cpu == NR_CPUS) | |
48f24c4d | 5178 | dest_cpu = any_online_cpu(p->cpus_allowed); |
1da177e4 LT |
5179 | |
5180 | /* No more Mr. Nice Guy. */ | |
5181 | if (dest_cpu == NR_CPUS) { | |
48f24c4d IM |
5182 | rq = task_rq_lock(p, &flags); |
5183 | cpus_setall(p->cpus_allowed); | |
5184 | dest_cpu = any_online_cpu(p->cpus_allowed); | |
efc30814 | 5185 | task_rq_unlock(rq, &flags); |
1da177e4 LT |
5186 | |
5187 | /* | |
5188 | * Don't tell them about moving exiting tasks or | |
5189 | * kernel threads (both mm NULL), since they never | |
5190 | * leave kernel. | |
5191 | */ | |
48f24c4d | 5192 | if (p->mm && printk_ratelimit()) |
1da177e4 LT |
5193 | printk(KERN_INFO "process %d (%s) no " |
5194 | "longer affine to cpu%d\n", | |
48f24c4d | 5195 | p->pid, p->comm, dead_cpu); |
1da177e4 | 5196 | } |
48f24c4d | 5197 | if (!__migrate_task(p, dead_cpu, dest_cpu)) |
efc30814 | 5198 | goto restart; |
1da177e4 LT |
5199 | } |
5200 | ||
5201 | /* | |
5202 | * While a dead CPU has no uninterruptible tasks queued at this point, | |
5203 | * it might still have a nonzero ->nr_uninterruptible counter, because | |
5204 | * for performance reasons the counter is not stricly tracking tasks to | |
5205 | * their home CPUs. So we just add the counter to another CPU's counter, | |
5206 | * to keep the global sum constant after CPU-down: | |
5207 | */ | |
70b97a7f | 5208 | static void migrate_nr_uninterruptible(struct rq *rq_src) |
1da177e4 | 5209 | { |
70b97a7f | 5210 | struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL)); |
1da177e4 LT |
5211 | unsigned long flags; |
5212 | ||
5213 | local_irq_save(flags); | |
5214 | double_rq_lock(rq_src, rq_dest); | |
5215 | rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible; | |
5216 | rq_src->nr_uninterruptible = 0; | |
5217 | double_rq_unlock(rq_src, rq_dest); | |
5218 | local_irq_restore(flags); | |
5219 | } | |
5220 | ||
5221 | /* Run through task list and migrate tasks from the dead cpu. */ | |
5222 | static void migrate_live_tasks(int src_cpu) | |
5223 | { | |
48f24c4d | 5224 | struct task_struct *p, *t; |
1da177e4 LT |
5225 | |
5226 | write_lock_irq(&tasklist_lock); | |
5227 | ||
48f24c4d IM |
5228 | do_each_thread(t, p) { |
5229 | if (p == current) | |
1da177e4 LT |
5230 | continue; |
5231 | ||
48f24c4d IM |
5232 | if (task_cpu(p) == src_cpu) |
5233 | move_task_off_dead_cpu(src_cpu, p); | |
5234 | } while_each_thread(t, p); | |
1da177e4 LT |
5235 | |
5236 | write_unlock_irq(&tasklist_lock); | |
5237 | } | |
5238 | ||
5239 | /* Schedules idle task to be the next runnable task on current CPU. | |
5240 | * It does so by boosting its priority to highest possible and adding it to | |
48f24c4d | 5241 | * the _front_ of the runqueue. Used by CPU offline code. |
1da177e4 LT |
5242 | */ |
5243 | void sched_idle_next(void) | |
5244 | { | |
48f24c4d | 5245 | int this_cpu = smp_processor_id(); |
70b97a7f | 5246 | struct rq *rq = cpu_rq(this_cpu); |
1da177e4 LT |
5247 | struct task_struct *p = rq->idle; |
5248 | unsigned long flags; | |
5249 | ||
5250 | /* cpu has to be offline */ | |
48f24c4d | 5251 | BUG_ON(cpu_online(this_cpu)); |
1da177e4 | 5252 | |
48f24c4d IM |
5253 | /* |
5254 | * Strictly not necessary since rest of the CPUs are stopped by now | |
5255 | * and interrupts disabled on the current cpu. | |
1da177e4 LT |
5256 | */ |
5257 | spin_lock_irqsave(&rq->lock, flags); | |
5258 | ||
5259 | __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1); | |
48f24c4d IM |
5260 | |
5261 | /* Add idle task to the _front_ of its priority queue: */ | |
1da177e4 LT |
5262 | __activate_idle_task(p, rq); |
5263 | ||
5264 | spin_unlock_irqrestore(&rq->lock, flags); | |
5265 | } | |
5266 | ||
48f24c4d IM |
5267 | /* |
5268 | * Ensures that the idle task is using init_mm right before its cpu goes | |
1da177e4 LT |
5269 | * offline. |
5270 | */ | |
5271 | void idle_task_exit(void) | |
5272 | { | |
5273 | struct mm_struct *mm = current->active_mm; | |
5274 | ||
5275 | BUG_ON(cpu_online(smp_processor_id())); | |
5276 | ||
5277 | if (mm != &init_mm) | |
5278 | switch_mm(mm, &init_mm, current); | |
5279 | mmdrop(mm); | |
5280 | } | |
5281 | ||
054b9108 | 5282 | /* called under rq->lock with disabled interrupts */ |
36c8b586 | 5283 | static void migrate_dead(unsigned int dead_cpu, struct task_struct *p) |
1da177e4 | 5284 | { |
70b97a7f | 5285 | struct rq *rq = cpu_rq(dead_cpu); |
1da177e4 LT |
5286 | |
5287 | /* Must be exiting, otherwise would be on tasklist. */ | |
48f24c4d | 5288 | BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD); |
1da177e4 LT |
5289 | |
5290 | /* Cannot have done final schedule yet: would have vanished. */ | |
c394cc9f | 5291 | BUG_ON(p->state == TASK_DEAD); |
1da177e4 | 5292 | |
48f24c4d | 5293 | get_task_struct(p); |
1da177e4 LT |
5294 | |
5295 | /* | |
5296 | * Drop lock around migration; if someone else moves it, | |
5297 | * that's OK. No task can be added to this CPU, so iteration is | |
5298 | * fine. | |
054b9108 | 5299 | * NOTE: interrupts should be left disabled --dev@ |
1da177e4 | 5300 | */ |
054b9108 | 5301 | spin_unlock(&rq->lock); |
48f24c4d | 5302 | move_task_off_dead_cpu(dead_cpu, p); |
054b9108 | 5303 | spin_lock(&rq->lock); |
1da177e4 | 5304 | |
48f24c4d | 5305 | put_task_struct(p); |
1da177e4 LT |
5306 | } |
5307 | ||
5308 | /* release_task() removes task from tasklist, so we won't find dead tasks. */ | |
5309 | static void migrate_dead_tasks(unsigned int dead_cpu) | |
5310 | { | |
70b97a7f | 5311 | struct rq *rq = cpu_rq(dead_cpu); |
48f24c4d | 5312 | unsigned int arr, i; |
1da177e4 LT |
5313 | |
5314 | for (arr = 0; arr < 2; arr++) { | |
5315 | for (i = 0; i < MAX_PRIO; i++) { | |
5316 | struct list_head *list = &rq->arrays[arr].queue[i]; | |
48f24c4d | 5317 | |
1da177e4 | 5318 | while (!list_empty(list)) |
36c8b586 IM |
5319 | migrate_dead(dead_cpu, list_entry(list->next, |
5320 | struct task_struct, run_list)); | |
1da177e4 LT |
5321 | } |
5322 | } | |
5323 | } | |
5324 | #endif /* CONFIG_HOTPLUG_CPU */ | |
5325 | ||
5326 | /* | |
5327 | * migration_call - callback that gets triggered when a CPU is added. | |
5328 | * Here we can start up the necessary migration thread for the new CPU. | |
5329 | */ | |
48f24c4d IM |
5330 | static int __cpuinit |
5331 | migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) | |
1da177e4 | 5332 | { |
1da177e4 | 5333 | struct task_struct *p; |
48f24c4d | 5334 | int cpu = (long)hcpu; |
1da177e4 | 5335 | unsigned long flags; |
70b97a7f | 5336 | struct rq *rq; |
1da177e4 LT |
5337 | |
5338 | switch (action) { | |
5339 | case CPU_UP_PREPARE: | |
5340 | p = kthread_create(migration_thread, hcpu, "migration/%d",cpu); | |
5341 | if (IS_ERR(p)) | |
5342 | return NOTIFY_BAD; | |
5343 | p->flags |= PF_NOFREEZE; | |
5344 | kthread_bind(p, cpu); | |
5345 | /* Must be high prio: stop_machine expects to yield to it. */ | |
5346 | rq = task_rq_lock(p, &flags); | |
5347 | __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1); | |
5348 | task_rq_unlock(rq, &flags); | |
5349 | cpu_rq(cpu)->migration_thread = p; | |
5350 | break; | |
48f24c4d | 5351 | |
1da177e4 LT |
5352 | case CPU_ONLINE: |
5353 | /* Strictly unneccessary, as first user will wake it. */ | |
5354 | wake_up_process(cpu_rq(cpu)->migration_thread); | |
5355 | break; | |
48f24c4d | 5356 | |
1da177e4 LT |
5357 | #ifdef CONFIG_HOTPLUG_CPU |
5358 | case CPU_UP_CANCELED: | |
fc75cdfa HC |
5359 | if (!cpu_rq(cpu)->migration_thread) |
5360 | break; | |
1da177e4 | 5361 | /* Unbind it from offline cpu so it can run. Fall thru. */ |
a4c4af7c HC |
5362 | kthread_bind(cpu_rq(cpu)->migration_thread, |
5363 | any_online_cpu(cpu_online_map)); | |
1da177e4 LT |
5364 | kthread_stop(cpu_rq(cpu)->migration_thread); |
5365 | cpu_rq(cpu)->migration_thread = NULL; | |
5366 | break; | |
48f24c4d | 5367 | |
1da177e4 LT |
5368 | case CPU_DEAD: |
5369 | migrate_live_tasks(cpu); | |
5370 | rq = cpu_rq(cpu); | |
5371 | kthread_stop(rq->migration_thread); | |
5372 | rq->migration_thread = NULL; | |
5373 | /* Idle task back to normal (off runqueue, low prio) */ | |
5374 | rq = task_rq_lock(rq->idle, &flags); | |
5375 | deactivate_task(rq->idle, rq); | |
5376 | rq->idle->static_prio = MAX_PRIO; | |
5377 | __setscheduler(rq->idle, SCHED_NORMAL, 0); | |
5378 | migrate_dead_tasks(cpu); | |
5379 | task_rq_unlock(rq, &flags); | |
5380 | migrate_nr_uninterruptible(rq); | |
5381 | BUG_ON(rq->nr_running != 0); | |
5382 | ||
5383 | /* No need to migrate the tasks: it was best-effort if | |
5384 | * they didn't do lock_cpu_hotplug(). Just wake up | |
5385 | * the requestors. */ | |
5386 | spin_lock_irq(&rq->lock); | |
5387 | while (!list_empty(&rq->migration_queue)) { | |
70b97a7f IM |
5388 | struct migration_req *req; |
5389 | ||
1da177e4 | 5390 | req = list_entry(rq->migration_queue.next, |
70b97a7f | 5391 | struct migration_req, list); |
1da177e4 LT |
5392 | list_del_init(&req->list); |
5393 | complete(&req->done); | |
5394 | } | |
5395 | spin_unlock_irq(&rq->lock); | |
5396 | break; | |
5397 | #endif | |
5398 | } | |
5399 | return NOTIFY_OK; | |
5400 | } | |
5401 | ||
5402 | /* Register at highest priority so that task migration (migrate_all_tasks) | |
5403 | * happens before everything else. | |
5404 | */ | |
26c2143b | 5405 | static struct notifier_block __cpuinitdata migration_notifier = { |
1da177e4 LT |
5406 | .notifier_call = migration_call, |
5407 | .priority = 10 | |
5408 | }; | |
5409 | ||
5410 | int __init migration_init(void) | |
5411 | { | |
5412 | void *cpu = (void *)(long)smp_processor_id(); | |
07dccf33 | 5413 | int err; |
48f24c4d IM |
5414 | |
5415 | /* Start one for the boot CPU: */ | |
07dccf33 AM |
5416 | err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); |
5417 | BUG_ON(err == NOTIFY_BAD); | |
1da177e4 LT |
5418 | migration_call(&migration_notifier, CPU_ONLINE, cpu); |
5419 | register_cpu_notifier(&migration_notifier); | |
48f24c4d | 5420 | |
1da177e4 LT |
5421 | return 0; |
5422 | } | |
5423 | #endif | |
5424 | ||
5425 | #ifdef CONFIG_SMP | |
1a20ff27 | 5426 | #undef SCHED_DOMAIN_DEBUG |
1da177e4 LT |
5427 | #ifdef SCHED_DOMAIN_DEBUG |
5428 | static void sched_domain_debug(struct sched_domain *sd, int cpu) | |
5429 | { | |
5430 | int level = 0; | |
5431 | ||
41c7ce9a NP |
5432 | if (!sd) { |
5433 | printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); | |
5434 | return; | |
5435 | } | |
5436 | ||
1da177e4 LT |
5437 | printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); |
5438 | ||
5439 | do { | |
5440 | int i; | |
5441 | char str[NR_CPUS]; | |
5442 | struct sched_group *group = sd->groups; | |
5443 | cpumask_t groupmask; | |
5444 | ||
5445 | cpumask_scnprintf(str, NR_CPUS, sd->span); | |
5446 | cpus_clear(groupmask); | |
5447 | ||
5448 | printk(KERN_DEBUG); | |
5449 | for (i = 0; i < level + 1; i++) | |
5450 | printk(" "); | |
5451 | printk("domain %d: ", level); | |
5452 | ||
5453 | if (!(sd->flags & SD_LOAD_BALANCE)) { | |
5454 | printk("does not load-balance\n"); | |
5455 | if (sd->parent) | |
33859f7f MOS |
5456 | printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" |
5457 | " has parent"); | |
1da177e4 LT |
5458 | break; |
5459 | } | |
5460 | ||
5461 | printk("span %s\n", str); | |
5462 | ||
5463 | if (!cpu_isset(cpu, sd->span)) | |
33859f7f MOS |
5464 | printk(KERN_ERR "ERROR: domain->span does not contain " |
5465 | "CPU%d\n", cpu); | |
1da177e4 | 5466 | if (!cpu_isset(cpu, group->cpumask)) |
33859f7f MOS |
5467 | printk(KERN_ERR "ERROR: domain->groups does not contain" |
5468 | " CPU%d\n", cpu); | |
1da177e4 LT |
5469 | |
5470 | printk(KERN_DEBUG); | |
5471 | for (i = 0; i < level + 2; i++) | |
5472 | printk(" "); | |
5473 | printk("groups:"); | |
5474 | do { | |
5475 | if (!group) { | |
5476 | printk("\n"); | |
5477 | printk(KERN_ERR "ERROR: group is NULL\n"); | |
5478 | break; | |
5479 | } | |
5480 | ||
5481 | if (!group->cpu_power) { | |
5482 | printk("\n"); | |
33859f7f MOS |
5483 | printk(KERN_ERR "ERROR: domain->cpu_power not " |
5484 | "set\n"); | |
1da177e4 LT |
5485 | } |
5486 | ||
5487 | if (!cpus_weight(group->cpumask)) { | |
5488 | printk("\n"); | |
5489 | printk(KERN_ERR "ERROR: empty group\n"); | |
5490 | } | |
5491 | ||
5492 | if (cpus_intersects(groupmask, group->cpumask)) { | |
5493 | printk("\n"); | |
5494 | printk(KERN_ERR "ERROR: repeated CPUs\n"); | |
5495 | } | |
5496 | ||
5497 | cpus_or(groupmask, groupmask, group->cpumask); | |
5498 | ||
5499 | cpumask_scnprintf(str, NR_CPUS, group->cpumask); | |
5500 | printk(" %s", str); | |
5501 | ||
5502 | group = group->next; | |
5503 | } while (group != sd->groups); | |
5504 | printk("\n"); | |
5505 | ||
5506 | if (!cpus_equal(sd->span, groupmask)) | |
33859f7f MOS |
5507 | printk(KERN_ERR "ERROR: groups don't span " |
5508 | "domain->span\n"); | |
1da177e4 LT |
5509 | |
5510 | level++; | |
5511 | sd = sd->parent; | |
33859f7f MOS |
5512 | if (!sd) |
5513 | continue; | |
1da177e4 | 5514 | |
33859f7f MOS |
5515 | if (!cpus_subset(groupmask, sd->span)) |
5516 | printk(KERN_ERR "ERROR: parent span is not a superset " | |
5517 | "of domain->span\n"); | |
1da177e4 LT |
5518 | |
5519 | } while (sd); | |
5520 | } | |
5521 | #else | |
48f24c4d | 5522 | # define sched_domain_debug(sd, cpu) do { } while (0) |
1da177e4 LT |
5523 | #endif |
5524 | ||
1a20ff27 | 5525 | static int sd_degenerate(struct sched_domain *sd) |
245af2c7 SS |
5526 | { |
5527 | if (cpus_weight(sd->span) == 1) | |
5528 | return 1; | |
5529 | ||
5530 | /* Following flags need at least 2 groups */ | |
5531 | if (sd->flags & (SD_LOAD_BALANCE | | |
5532 | SD_BALANCE_NEWIDLE | | |
5533 | SD_BALANCE_FORK | | |
89c4710e SS |
5534 | SD_BALANCE_EXEC | |
5535 | SD_SHARE_CPUPOWER | | |
5536 | SD_SHARE_PKG_RESOURCES)) { | |
245af2c7 SS |
5537 | if (sd->groups != sd->groups->next) |
5538 | return 0; | |
5539 | } | |
5540 | ||
5541 | /* Following flags don't use groups */ | |
5542 | if (sd->flags & (SD_WAKE_IDLE | | |
5543 | SD_WAKE_AFFINE | | |
5544 | SD_WAKE_BALANCE)) | |
5545 | return 0; | |
5546 | ||
5547 | return 1; | |
5548 | } | |
5549 | ||
48f24c4d IM |
5550 | static int |
5551 | sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) | |
245af2c7 SS |
5552 | { |
5553 | unsigned long cflags = sd->flags, pflags = parent->flags; | |
5554 | ||
5555 | if (sd_degenerate(parent)) | |
5556 | return 1; | |
5557 | ||
5558 | if (!cpus_equal(sd->span, parent->span)) | |
5559 | return 0; | |
5560 | ||
5561 | /* Does parent contain flags not in child? */ | |
5562 | /* WAKE_BALANCE is a subset of WAKE_AFFINE */ | |
5563 | if (cflags & SD_WAKE_AFFINE) | |
5564 | pflags &= ~SD_WAKE_BALANCE; | |
5565 | /* Flags needing groups don't count if only 1 group in parent */ | |
5566 | if (parent->groups == parent->groups->next) { | |
5567 | pflags &= ~(SD_LOAD_BALANCE | | |
5568 | SD_BALANCE_NEWIDLE | | |
5569 | SD_BALANCE_FORK | | |
89c4710e SS |
5570 | SD_BALANCE_EXEC | |
5571 | SD_SHARE_CPUPOWER | | |
5572 | SD_SHARE_PKG_RESOURCES); | |
245af2c7 SS |
5573 | } |
5574 | if (~cflags & pflags) | |
5575 | return 0; | |
5576 | ||
5577 | return 1; | |
5578 | } | |
5579 | ||
1da177e4 LT |
5580 | /* |
5581 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must | |
5582 | * hold the hotplug lock. | |
5583 | */ | |
9c1cfda2 | 5584 | static void cpu_attach_domain(struct sched_domain *sd, int cpu) |
1da177e4 | 5585 | { |
70b97a7f | 5586 | struct rq *rq = cpu_rq(cpu); |
245af2c7 SS |
5587 | struct sched_domain *tmp; |
5588 | ||
5589 | /* Remove the sched domains which do not contribute to scheduling. */ | |
5590 | for (tmp = sd; tmp; tmp = tmp->parent) { | |
5591 | struct sched_domain *parent = tmp->parent; | |
5592 | if (!parent) | |
5593 | break; | |
1a848870 | 5594 | if (sd_parent_degenerate(tmp, parent)) { |
245af2c7 | 5595 | tmp->parent = parent->parent; |
1a848870 SS |
5596 | if (parent->parent) |
5597 | parent->parent->child = tmp; | |
5598 | } | |
245af2c7 SS |
5599 | } |
5600 | ||
1a848870 | 5601 | if (sd && sd_degenerate(sd)) { |
245af2c7 | 5602 | sd = sd->parent; |
1a848870 SS |
5603 | if (sd) |
5604 | sd->child = NULL; | |
5605 | } | |
1da177e4 LT |
5606 | |
5607 | sched_domain_debug(sd, cpu); | |
5608 | ||
674311d5 | 5609 | rcu_assign_pointer(rq->sd, sd); |
1da177e4 LT |
5610 | } |
5611 | ||
5612 | /* cpus with isolated domains */ | |
67af63a6 | 5613 | static cpumask_t cpu_isolated_map = CPU_MASK_NONE; |
1da177e4 LT |
5614 | |
5615 | /* Setup the mask of cpus configured for isolated domains */ | |
5616 | static int __init isolated_cpu_setup(char *str) | |
5617 | { | |
5618 | int ints[NR_CPUS], i; | |
5619 | ||
5620 | str = get_options(str, ARRAY_SIZE(ints), ints); | |
5621 | cpus_clear(cpu_isolated_map); | |
5622 | for (i = 1; i <= ints[0]; i++) | |
5623 | if (ints[i] < NR_CPUS) | |
5624 | cpu_set(ints[i], cpu_isolated_map); | |
5625 | return 1; | |
5626 | } | |
5627 | ||
5628 | __setup ("isolcpus=", isolated_cpu_setup); | |
5629 | ||
5630 | /* | |
6711cab4 SS |
5631 | * init_sched_build_groups takes the cpumask we wish to span, and a pointer |
5632 | * to a function which identifies what group(along with sched group) a CPU | |
5633 | * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS | |
5634 | * (due to the fact that we keep track of groups covered with a cpumask_t). | |
1da177e4 LT |
5635 | * |
5636 | * init_sched_build_groups will build a circular linked list of the groups | |
5637 | * covered by the given span, and will set each group's ->cpumask correctly, | |
5638 | * and ->cpu_power to 0. | |
5639 | */ | |
a616058b | 5640 | static void |
6711cab4 SS |
5641 | init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map, |
5642 | int (*group_fn)(int cpu, const cpumask_t *cpu_map, | |
5643 | struct sched_group **sg)) | |
1da177e4 LT |
5644 | { |
5645 | struct sched_group *first = NULL, *last = NULL; | |
5646 | cpumask_t covered = CPU_MASK_NONE; | |
5647 | int i; | |
5648 | ||
5649 | for_each_cpu_mask(i, span) { | |
6711cab4 SS |
5650 | struct sched_group *sg; |
5651 | int group = group_fn(i, cpu_map, &sg); | |
1da177e4 LT |
5652 | int j; |
5653 | ||
5654 | if (cpu_isset(i, covered)) | |
5655 | continue; | |
5656 | ||
5657 | sg->cpumask = CPU_MASK_NONE; | |
5658 | sg->cpu_power = 0; | |
5659 | ||
5660 | for_each_cpu_mask(j, span) { | |
6711cab4 | 5661 | if (group_fn(j, cpu_map, NULL) != group) |
1da177e4 LT |
5662 | continue; |
5663 | ||
5664 | cpu_set(j, covered); | |
5665 | cpu_set(j, sg->cpumask); | |
5666 | } | |
5667 | if (!first) | |
5668 | first = sg; | |
5669 | if (last) | |
5670 | last->next = sg; | |
5671 | last = sg; | |
5672 | } | |
5673 | last->next = first; | |
5674 | } | |
5675 | ||
9c1cfda2 | 5676 | #define SD_NODES_PER_DOMAIN 16 |
1da177e4 | 5677 | |
198e2f18 | 5678 | /* |
5679 | * Self-tuning task migration cost measurement between source and target CPUs. | |
5680 | * | |
5681 | * This is done by measuring the cost of manipulating buffers of varying | |
5682 | * sizes. For a given buffer-size here are the steps that are taken: | |
5683 | * | |
5684 | * 1) the source CPU reads+dirties a shared buffer | |
5685 | * 2) the target CPU reads+dirties the same shared buffer | |
5686 | * | |
5687 | * We measure how long they take, in the following 4 scenarios: | |
5688 | * | |
5689 | * - source: CPU1, target: CPU2 | cost1 | |
5690 | * - source: CPU2, target: CPU1 | cost2 | |
5691 | * - source: CPU1, target: CPU1 | cost3 | |
5692 | * - source: CPU2, target: CPU2 | cost4 | |
5693 | * | |
5694 | * We then calculate the cost3+cost4-cost1-cost2 difference - this is | |
5695 | * the cost of migration. | |
5696 | * | |
5697 | * We then start off from a small buffer-size and iterate up to larger | |
5698 | * buffer sizes, in 5% steps - measuring each buffer-size separately, and | |
5699 | * doing a maximum search for the cost. (The maximum cost for a migration | |
5700 | * normally occurs when the working set size is around the effective cache | |
5701 | * size.) | |
5702 | */ | |
5703 | #define SEARCH_SCOPE 2 | |
5704 | #define MIN_CACHE_SIZE (64*1024U) | |
5705 | #define DEFAULT_CACHE_SIZE (5*1024*1024U) | |
70b4d63e | 5706 | #define ITERATIONS 1 |
198e2f18 | 5707 | #define SIZE_THRESH 130 |
5708 | #define COST_THRESH 130 | |
5709 | ||
5710 | /* | |
5711 | * The migration cost is a function of 'domain distance'. Domain | |
5712 | * distance is the number of steps a CPU has to iterate down its | |
5713 | * domain tree to share a domain with the other CPU. The farther | |
5714 | * two CPUs are from each other, the larger the distance gets. | |
5715 | * | |
5716 | * Note that we use the distance only to cache measurement results, | |
5717 | * the distance value is not used numerically otherwise. When two | |
5718 | * CPUs have the same distance it is assumed that the migration | |
5719 | * cost is the same. (this is a simplification but quite practical) | |
5720 | */ | |
5721 | #define MAX_DOMAIN_DISTANCE 32 | |
5722 | ||
5723 | static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] = | |
4bbf39c2 IM |
5724 | { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] = |
5725 | /* | |
5726 | * Architectures may override the migration cost and thus avoid | |
5727 | * boot-time calibration. Unit is nanoseconds. Mostly useful for | |
5728 | * virtualized hardware: | |
5729 | */ | |
5730 | #ifdef CONFIG_DEFAULT_MIGRATION_COST | |
5731 | CONFIG_DEFAULT_MIGRATION_COST | |
5732 | #else | |
5733 | -1LL | |
5734 | #endif | |
5735 | }; | |
198e2f18 | 5736 | |
5737 | /* | |
5738 | * Allow override of migration cost - in units of microseconds. | |
5739 | * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost | |
5740 | * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs: | |
5741 | */ | |
5742 | static int __init migration_cost_setup(char *str) | |
5743 | { | |
5744 | int ints[MAX_DOMAIN_DISTANCE+1], i; | |
5745 | ||
5746 | str = get_options(str, ARRAY_SIZE(ints), ints); | |
5747 | ||
5748 | printk("#ints: %d\n", ints[0]); | |
5749 | for (i = 1; i <= ints[0]; i++) { | |
5750 | migration_cost[i-1] = (unsigned long long)ints[i]*1000; | |
5751 | printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]); | |
5752 | } | |
5753 | return 1; | |
5754 | } | |
5755 | ||
5756 | __setup ("migration_cost=", migration_cost_setup); | |
5757 | ||
5758 | /* | |
5759 | * Global multiplier (divisor) for migration-cutoff values, | |
5760 | * in percentiles. E.g. use a value of 150 to get 1.5 times | |
5761 | * longer cache-hot cutoff times. | |
5762 | * | |
5763 | * (We scale it from 100 to 128 to long long handling easier.) | |
5764 | */ | |
5765 | ||
5766 | #define MIGRATION_FACTOR_SCALE 128 | |
5767 | ||
5768 | static unsigned int migration_factor = MIGRATION_FACTOR_SCALE; | |
5769 | ||
5770 | static int __init setup_migration_factor(char *str) | |
5771 | { | |
5772 | get_option(&str, &migration_factor); | |
5773 | migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100; | |
5774 | return 1; | |
5775 | } | |
5776 | ||
5777 | __setup("migration_factor=", setup_migration_factor); | |
5778 | ||
5779 | /* | |
5780 | * Estimated distance of two CPUs, measured via the number of domains | |
5781 | * we have to pass for the two CPUs to be in the same span: | |
5782 | */ | |
5783 | static unsigned long domain_distance(int cpu1, int cpu2) | |
5784 | { | |
5785 | unsigned long distance = 0; | |
5786 | struct sched_domain *sd; | |
5787 | ||
5788 | for_each_domain(cpu1, sd) { | |
5789 | WARN_ON(!cpu_isset(cpu1, sd->span)); | |
5790 | if (cpu_isset(cpu2, sd->span)) | |
5791 | return distance; | |
5792 | distance++; | |
5793 | } | |
5794 | if (distance >= MAX_DOMAIN_DISTANCE) { | |
5795 | WARN_ON(1); | |
5796 | distance = MAX_DOMAIN_DISTANCE-1; | |
5797 | } | |
5798 | ||
5799 | return distance; | |
5800 | } | |
5801 | ||
5802 | static unsigned int migration_debug; | |
5803 | ||
5804 | static int __init setup_migration_debug(char *str) | |
5805 | { | |
5806 | get_option(&str, &migration_debug); | |
5807 | return 1; | |
5808 | } | |
5809 | ||
5810 | __setup("migration_debug=", setup_migration_debug); | |
5811 | ||
5812 | /* | |
5813 | * Maximum cache-size that the scheduler should try to measure. | |
5814 | * Architectures with larger caches should tune this up during | |
5815 | * bootup. Gets used in the domain-setup code (i.e. during SMP | |
5816 | * bootup). | |
5817 | */ | |
5818 | unsigned int max_cache_size; | |
5819 | ||
5820 | static int __init setup_max_cache_size(char *str) | |
5821 | { | |
5822 | get_option(&str, &max_cache_size); | |
5823 | return 1; | |
5824 | } | |
5825 | ||
5826 | __setup("max_cache_size=", setup_max_cache_size); | |
5827 | ||
5828 | /* | |
5829 | * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This | |
5830 | * is the operation that is timed, so we try to generate unpredictable | |
5831 | * cachemisses that still end up filling the L2 cache: | |
5832 | */ | |
5833 | static void touch_cache(void *__cache, unsigned long __size) | |
5834 | { | |
33859f7f MOS |
5835 | unsigned long size = __size / sizeof(long); |
5836 | unsigned long chunk1 = size / 3; | |
5837 | unsigned long chunk2 = 2 * size / 3; | |
198e2f18 | 5838 | unsigned long *cache = __cache; |
5839 | int i; | |
5840 | ||
5841 | for (i = 0; i < size/6; i += 8) { | |
5842 | switch (i % 6) { | |
5843 | case 0: cache[i]++; | |
5844 | case 1: cache[size-1-i]++; | |
5845 | case 2: cache[chunk1-i]++; | |
5846 | case 3: cache[chunk1+i]++; | |
5847 | case 4: cache[chunk2-i]++; | |
5848 | case 5: cache[chunk2+i]++; | |
5849 | } | |
5850 | } | |
5851 | } | |
5852 | ||
5853 | /* | |
5854 | * Measure the cache-cost of one task migration. Returns in units of nsec. | |
5855 | */ | |
48f24c4d IM |
5856 | static unsigned long long |
5857 | measure_one(void *cache, unsigned long size, int source, int target) | |
198e2f18 | 5858 | { |
5859 | cpumask_t mask, saved_mask; | |
5860 | unsigned long long t0, t1, t2, t3, cost; | |
5861 | ||
5862 | saved_mask = current->cpus_allowed; | |
5863 | ||
5864 | /* | |
5865 | * Flush source caches to RAM and invalidate them: | |
5866 | */ | |
5867 | sched_cacheflush(); | |
5868 | ||
5869 | /* | |
5870 | * Migrate to the source CPU: | |
5871 | */ | |
5872 | mask = cpumask_of_cpu(source); | |
5873 | set_cpus_allowed(current, mask); | |
5874 | WARN_ON(smp_processor_id() != source); | |
5875 | ||
5876 | /* | |
5877 | * Dirty the working set: | |
5878 | */ | |
5879 | t0 = sched_clock(); | |
5880 | touch_cache(cache, size); | |
5881 | t1 = sched_clock(); | |
5882 | ||
5883 | /* | |
5884 | * Migrate to the target CPU, dirty the L2 cache and access | |
5885 | * the shared buffer. (which represents the working set | |
5886 | * of a migrated task.) | |
5887 | */ | |
5888 | mask = cpumask_of_cpu(target); | |
5889 | set_cpus_allowed(current, mask); | |
5890 | WARN_ON(smp_processor_id() != target); | |
5891 | ||
5892 | t2 = sched_clock(); | |
5893 | touch_cache(cache, size); | |
5894 | t3 = sched_clock(); | |
5895 | ||
5896 | cost = t1-t0 + t3-t2; | |
5897 | ||
5898 | if (migration_debug >= 2) | |
5899 | printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n", | |
5900 | source, target, t1-t0, t1-t0, t3-t2, cost); | |
5901 | /* | |
5902 | * Flush target caches to RAM and invalidate them: | |
5903 | */ | |
5904 | sched_cacheflush(); | |
5905 | ||
5906 | set_cpus_allowed(current, saved_mask); | |
5907 | ||
5908 | return cost; | |
5909 | } | |
5910 | ||
5911 | /* | |
5912 | * Measure a series of task migrations and return the average | |
5913 | * result. Since this code runs early during bootup the system | |
5914 | * is 'undisturbed' and the average latency makes sense. | |
5915 | * | |
5916 | * The algorithm in essence auto-detects the relevant cache-size, | |
5917 | * so it will properly detect different cachesizes for different | |
5918 | * cache-hierarchies, depending on how the CPUs are connected. | |
5919 | * | |
5920 | * Architectures can prime the upper limit of the search range via | |
5921 | * max_cache_size, otherwise the search range defaults to 20MB...64K. | |
5922 | */ | |
5923 | static unsigned long long | |
5924 | measure_cost(int cpu1, int cpu2, void *cache, unsigned int size) | |
5925 | { | |
5926 | unsigned long long cost1, cost2; | |
5927 | int i; | |
5928 | ||
5929 | /* | |
5930 | * Measure the migration cost of 'size' bytes, over an | |
5931 | * average of 10 runs: | |
5932 | * | |
5933 | * (We perturb the cache size by a small (0..4k) | |
5934 | * value to compensate size/alignment related artifacts. | |
5935 | * We also subtract the cost of the operation done on | |
5936 | * the same CPU.) | |
5937 | */ | |
5938 | cost1 = 0; | |
5939 | ||
5940 | /* | |
5941 | * dry run, to make sure we start off cache-cold on cpu1, | |
5942 | * and to get any vmalloc pagefaults in advance: | |
5943 | */ | |
5944 | measure_one(cache, size, cpu1, cpu2); | |
5945 | for (i = 0; i < ITERATIONS; i++) | |
33859f7f | 5946 | cost1 += measure_one(cache, size - i * 1024, cpu1, cpu2); |
198e2f18 | 5947 | |
5948 | measure_one(cache, size, cpu2, cpu1); | |
5949 | for (i = 0; i < ITERATIONS; i++) | |
33859f7f | 5950 | cost1 += measure_one(cache, size - i * 1024, cpu2, cpu1); |
198e2f18 | 5951 | |
5952 | /* | |
5953 | * (We measure the non-migrating [cached] cost on both | |
5954 | * cpu1 and cpu2, to handle CPUs with different speeds) | |
5955 | */ | |
5956 | cost2 = 0; | |
5957 | ||
5958 | measure_one(cache, size, cpu1, cpu1); | |
5959 | for (i = 0; i < ITERATIONS; i++) | |
33859f7f | 5960 | cost2 += measure_one(cache, size - i * 1024, cpu1, cpu1); |
198e2f18 | 5961 | |
5962 | measure_one(cache, size, cpu2, cpu2); | |
5963 | for (i = 0; i < ITERATIONS; i++) | |
33859f7f | 5964 | cost2 += measure_one(cache, size - i * 1024, cpu2, cpu2); |
198e2f18 | 5965 | |
5966 | /* | |
5967 | * Get the per-iteration migration cost: | |
5968 | */ | |
33859f7f MOS |
5969 | do_div(cost1, 2 * ITERATIONS); |
5970 | do_div(cost2, 2 * ITERATIONS); | |
198e2f18 | 5971 | |
5972 | return cost1 - cost2; | |
5973 | } | |
5974 | ||
5975 | static unsigned long long measure_migration_cost(int cpu1, int cpu2) | |
5976 | { | |
5977 | unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0; | |
5978 | unsigned int max_size, size, size_found = 0; | |
5979 | long long cost = 0, prev_cost; | |
5980 | void *cache; | |
5981 | ||
5982 | /* | |
5983 | * Search from max_cache_size*5 down to 64K - the real relevant | |
5984 | * cachesize has to lie somewhere inbetween. | |
5985 | */ | |
5986 | if (max_cache_size) { | |
5987 | max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE); | |
5988 | size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE); | |
5989 | } else { | |
5990 | /* | |
5991 | * Since we have no estimation about the relevant | |
5992 | * search range | |
5993 | */ | |
5994 | max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE; | |
5995 | size = MIN_CACHE_SIZE; | |
5996 | } | |
5997 | ||
5998 | if (!cpu_online(cpu1) || !cpu_online(cpu2)) { | |
5999 | printk("cpu %d and %d not both online!\n", cpu1, cpu2); | |
6000 | return 0; | |
6001 | } | |
6002 | ||
6003 | /* | |
6004 | * Allocate the working set: | |
6005 | */ | |
6006 | cache = vmalloc(max_size); | |
6007 | if (!cache) { | |
33859f7f | 6008 | printk("could not vmalloc %d bytes for cache!\n", 2 * max_size); |
2ed6e34f | 6009 | return 1000000; /* return 1 msec on very small boxen */ |
198e2f18 | 6010 | } |
6011 | ||
6012 | while (size <= max_size) { | |
6013 | prev_cost = cost; | |
6014 | cost = measure_cost(cpu1, cpu2, cache, size); | |
6015 | ||
6016 | /* | |
6017 | * Update the max: | |
6018 | */ | |
6019 | if (cost > 0) { | |
6020 | if (max_cost < cost) { | |
6021 | max_cost = cost; | |
6022 | size_found = size; | |
6023 | } | |
6024 | } | |
6025 | /* | |
6026 | * Calculate average fluctuation, we use this to prevent | |
6027 | * noise from triggering an early break out of the loop: | |
6028 | */ | |
6029 | fluct = abs(cost - prev_cost); | |
6030 | avg_fluct = (avg_fluct + fluct)/2; | |
6031 | ||
6032 | if (migration_debug) | |
33859f7f MOS |
6033 | printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): " |
6034 | "(%8Ld %8Ld)\n", | |
198e2f18 | 6035 | cpu1, cpu2, size, |
6036 | (long)cost / 1000000, | |
6037 | ((long)cost / 100000) % 10, | |
6038 | (long)max_cost / 1000000, | |
6039 | ((long)max_cost / 100000) % 10, | |
6040 | domain_distance(cpu1, cpu2), | |
6041 | cost, avg_fluct); | |
6042 | ||
6043 | /* | |
6044 | * If we iterated at least 20% past the previous maximum, | |
6045 | * and the cost has dropped by more than 20% already, | |
6046 | * (taking fluctuations into account) then we assume to | |
6047 | * have found the maximum and break out of the loop early: | |
6048 | */ | |
6049 | if (size_found && (size*100 > size_found*SIZE_THRESH)) | |
6050 | if (cost+avg_fluct <= 0 || | |
6051 | max_cost*100 > (cost+avg_fluct)*COST_THRESH) { | |
6052 | ||
6053 | if (migration_debug) | |
6054 | printk("-> found max.\n"); | |
6055 | break; | |
6056 | } | |
6057 | /* | |
70b4d63e | 6058 | * Increase the cachesize in 10% steps: |
198e2f18 | 6059 | */ |
70b4d63e | 6060 | size = size * 10 / 9; |
198e2f18 | 6061 | } |
6062 | ||
6063 | if (migration_debug) | |
6064 | printk("[%d][%d] working set size found: %d, cost: %Ld\n", | |
6065 | cpu1, cpu2, size_found, max_cost); | |
6066 | ||
6067 | vfree(cache); | |
6068 | ||
6069 | /* | |
6070 | * A task is considered 'cache cold' if at least 2 times | |
6071 | * the worst-case cost of migration has passed. | |
6072 | * | |
6073 | * (this limit is only listened to if the load-balancing | |
6074 | * situation is 'nice' - if there is a large imbalance we | |
6075 | * ignore it for the sake of CPU utilization and | |
6076 | * processing fairness.) | |
6077 | */ | |
6078 | return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE; | |
6079 | } | |
6080 | ||
6081 | static void calibrate_migration_costs(const cpumask_t *cpu_map) | |
6082 | { | |
6083 | int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id(); | |
6084 | unsigned long j0, j1, distance, max_distance = 0; | |
6085 | struct sched_domain *sd; | |
6086 | ||
6087 | j0 = jiffies; | |
6088 | ||
6089 | /* | |
6090 | * First pass - calculate the cacheflush times: | |
6091 | */ | |
6092 | for_each_cpu_mask(cpu1, *cpu_map) { | |
6093 | for_each_cpu_mask(cpu2, *cpu_map) { | |
6094 | if (cpu1 == cpu2) | |
6095 | continue; | |
6096 | distance = domain_distance(cpu1, cpu2); | |
6097 | max_distance = max(max_distance, distance); | |
6098 | /* | |
6099 | * No result cached yet? | |
6100 | */ | |
6101 | if (migration_cost[distance] == -1LL) | |
6102 | migration_cost[distance] = | |
6103 | measure_migration_cost(cpu1, cpu2); | |
6104 | } | |
6105 | } | |
6106 | /* | |
6107 | * Second pass - update the sched domain hierarchy with | |
6108 | * the new cache-hot-time estimations: | |
6109 | */ | |
6110 | for_each_cpu_mask(cpu, *cpu_map) { | |
6111 | distance = 0; | |
6112 | for_each_domain(cpu, sd) { | |
6113 | sd->cache_hot_time = migration_cost[distance]; | |
6114 | distance++; | |
6115 | } | |
6116 | } | |
6117 | /* | |
6118 | * Print the matrix: | |
6119 | */ | |
6120 | if (migration_debug) | |
6121 | printk("migration: max_cache_size: %d, cpu: %d MHz:\n", | |
6122 | max_cache_size, | |
6123 | #ifdef CONFIG_X86 | |
6124 | cpu_khz/1000 | |
6125 | #else | |
6126 | -1 | |
6127 | #endif | |
6128 | ); | |
33859f7f MOS |
6129 | if (system_state == SYSTEM_BOOTING && num_online_cpus() > 1) { |
6130 | printk("migration_cost="); | |
6131 | for (distance = 0; distance <= max_distance; distance++) { | |
6132 | if (distance) | |
6133 | printk(","); | |
6134 | printk("%ld", (long)migration_cost[distance] / 1000); | |
bd576c95 | 6135 | } |
33859f7f | 6136 | printk("\n"); |
198e2f18 | 6137 | } |
198e2f18 | 6138 | j1 = jiffies; |
6139 | if (migration_debug) | |
33859f7f | 6140 | printk("migration: %ld seconds\n", (j1-j0) / HZ); |
198e2f18 | 6141 | |
6142 | /* | |
6143 | * Move back to the original CPU. NUMA-Q gets confused | |
6144 | * if we migrate to another quad during bootup. | |
6145 | */ | |
6146 | if (raw_smp_processor_id() != orig_cpu) { | |
6147 | cpumask_t mask = cpumask_of_cpu(orig_cpu), | |
6148 | saved_mask = current->cpus_allowed; | |
6149 | ||
6150 | set_cpus_allowed(current, mask); | |
6151 | set_cpus_allowed(current, saved_mask); | |
6152 | } | |
6153 | } | |
6154 | ||
9c1cfda2 | 6155 | #ifdef CONFIG_NUMA |
198e2f18 | 6156 | |
9c1cfda2 JH |
6157 | /** |
6158 | * find_next_best_node - find the next node to include in a sched_domain | |
6159 | * @node: node whose sched_domain we're building | |
6160 | * @used_nodes: nodes already in the sched_domain | |
6161 | * | |
6162 | * Find the next node to include in a given scheduling domain. Simply | |
6163 | * finds the closest node not already in the @used_nodes map. | |
6164 | * | |
6165 | * Should use nodemask_t. | |
6166 | */ | |
6167 | static int find_next_best_node(int node, unsigned long *used_nodes) | |
6168 | { | |
6169 | int i, n, val, min_val, best_node = 0; | |
6170 | ||
6171 | min_val = INT_MAX; | |
6172 | ||
6173 | for (i = 0; i < MAX_NUMNODES; i++) { | |
6174 | /* Start at @node */ | |
6175 | n = (node + i) % MAX_NUMNODES; | |
6176 | ||
6177 | if (!nr_cpus_node(n)) | |
6178 | continue; | |
6179 | ||
6180 | /* Skip already used nodes */ | |
6181 | if (test_bit(n, used_nodes)) | |
6182 | continue; | |
6183 | ||
6184 | /* Simple min distance search */ | |
6185 | val = node_distance(node, n); | |
6186 | ||
6187 | if (val < min_val) { | |
6188 | min_val = val; | |
6189 | best_node = n; | |
6190 | } | |
6191 | } | |
6192 | ||
6193 | set_bit(best_node, used_nodes); | |
6194 | return best_node; | |
6195 | } | |
6196 | ||
6197 | /** | |
6198 | * sched_domain_node_span - get a cpumask for a node's sched_domain | |
6199 | * @node: node whose cpumask we're constructing | |
6200 | * @size: number of nodes to include in this span | |
6201 | * | |
6202 | * Given a node, construct a good cpumask for its sched_domain to span. It | |
6203 | * should be one that prevents unnecessary balancing, but also spreads tasks | |
6204 | * out optimally. | |
6205 | */ | |
6206 | static cpumask_t sched_domain_node_span(int node) | |
6207 | { | |
9c1cfda2 | 6208 | DECLARE_BITMAP(used_nodes, MAX_NUMNODES); |
48f24c4d IM |
6209 | cpumask_t span, nodemask; |
6210 | int i; | |
9c1cfda2 JH |
6211 | |
6212 | cpus_clear(span); | |
6213 | bitmap_zero(used_nodes, MAX_NUMNODES); | |
6214 | ||
6215 | nodemask = node_to_cpumask(node); | |
6216 | cpus_or(span, span, nodemask); | |
6217 | set_bit(node, used_nodes); | |
6218 | ||
6219 | for (i = 1; i < SD_NODES_PER_DOMAIN; i++) { | |
6220 | int next_node = find_next_best_node(node, used_nodes); | |
48f24c4d | 6221 | |
9c1cfda2 JH |
6222 | nodemask = node_to_cpumask(next_node); |
6223 | cpus_or(span, span, nodemask); | |
6224 | } | |
6225 | ||
6226 | return span; | |
6227 | } | |
6228 | #endif | |
6229 | ||
5c45bf27 | 6230 | int sched_smt_power_savings = 0, sched_mc_power_savings = 0; |
48f24c4d | 6231 | |
9c1cfda2 | 6232 | /* |
48f24c4d | 6233 | * SMT sched-domains: |
9c1cfda2 | 6234 | */ |
1da177e4 LT |
6235 | #ifdef CONFIG_SCHED_SMT |
6236 | static DEFINE_PER_CPU(struct sched_domain, cpu_domains); | |
6711cab4 | 6237 | static DEFINE_PER_CPU(struct sched_group, sched_group_cpus); |
48f24c4d | 6238 | |
6711cab4 SS |
6239 | static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, |
6240 | struct sched_group **sg) | |
1da177e4 | 6241 | { |
6711cab4 SS |
6242 | if (sg) |
6243 | *sg = &per_cpu(sched_group_cpus, cpu); | |
1da177e4 LT |
6244 | return cpu; |
6245 | } | |
6246 | #endif | |
6247 | ||
48f24c4d IM |
6248 | /* |
6249 | * multi-core sched-domains: | |
6250 | */ | |
1e9f28fa SS |
6251 | #ifdef CONFIG_SCHED_MC |
6252 | static DEFINE_PER_CPU(struct sched_domain, core_domains); | |
6711cab4 | 6253 | static DEFINE_PER_CPU(struct sched_group, sched_group_core); |
1e9f28fa SS |
6254 | #endif |
6255 | ||
6256 | #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT) | |
6711cab4 SS |
6257 | static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map, |
6258 | struct sched_group **sg) | |
1e9f28fa | 6259 | { |
6711cab4 | 6260 | int group; |
a616058b SS |
6261 | cpumask_t mask = cpu_sibling_map[cpu]; |
6262 | cpus_and(mask, mask, *cpu_map); | |
6711cab4 SS |
6263 | group = first_cpu(mask); |
6264 | if (sg) | |
6265 | *sg = &per_cpu(sched_group_core, group); | |
6266 | return group; | |
1e9f28fa SS |
6267 | } |
6268 | #elif defined(CONFIG_SCHED_MC) | |
6711cab4 SS |
6269 | static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map, |
6270 | struct sched_group **sg) | |
1e9f28fa | 6271 | { |
6711cab4 SS |
6272 | if (sg) |
6273 | *sg = &per_cpu(sched_group_core, cpu); | |
1e9f28fa SS |
6274 | return cpu; |
6275 | } | |
6276 | #endif | |
6277 | ||
1da177e4 | 6278 | static DEFINE_PER_CPU(struct sched_domain, phys_domains); |
6711cab4 | 6279 | static DEFINE_PER_CPU(struct sched_group, sched_group_phys); |
48f24c4d | 6280 | |
6711cab4 SS |
6281 | static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, |
6282 | struct sched_group **sg) | |
1da177e4 | 6283 | { |
6711cab4 | 6284 | int group; |
48f24c4d | 6285 | #ifdef CONFIG_SCHED_MC |
1e9f28fa | 6286 | cpumask_t mask = cpu_coregroup_map(cpu); |
a616058b | 6287 | cpus_and(mask, mask, *cpu_map); |
6711cab4 | 6288 | group = first_cpu(mask); |
1e9f28fa | 6289 | #elif defined(CONFIG_SCHED_SMT) |
a616058b SS |
6290 | cpumask_t mask = cpu_sibling_map[cpu]; |
6291 | cpus_and(mask, mask, *cpu_map); | |
6711cab4 | 6292 | group = first_cpu(mask); |
1da177e4 | 6293 | #else |
6711cab4 | 6294 | group = cpu; |
1da177e4 | 6295 | #endif |
6711cab4 SS |
6296 | if (sg) |
6297 | *sg = &per_cpu(sched_group_phys, group); | |
6298 | return group; | |
1da177e4 LT |
6299 | } |
6300 | ||
6301 | #ifdef CONFIG_NUMA | |
1da177e4 | 6302 | /* |
9c1cfda2 JH |
6303 | * The init_sched_build_groups can't handle what we want to do with node |
6304 | * groups, so roll our own. Now each node has its own list of groups which | |
6305 | * gets dynamically allocated. | |
1da177e4 | 6306 | */ |
9c1cfda2 | 6307 | static DEFINE_PER_CPU(struct sched_domain, node_domains); |
d1b55138 | 6308 | static struct sched_group **sched_group_nodes_bycpu[NR_CPUS]; |
1da177e4 | 6309 | |
9c1cfda2 | 6310 | static DEFINE_PER_CPU(struct sched_domain, allnodes_domains); |
6711cab4 | 6311 | static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes); |
9c1cfda2 | 6312 | |
6711cab4 SS |
6313 | static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map, |
6314 | struct sched_group **sg) | |
9c1cfda2 | 6315 | { |
6711cab4 SS |
6316 | cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu)); |
6317 | int group; | |
6318 | ||
6319 | cpus_and(nodemask, nodemask, *cpu_map); | |
6320 | group = first_cpu(nodemask); | |
6321 | ||
6322 | if (sg) | |
6323 | *sg = &per_cpu(sched_group_allnodes, group); | |
6324 | return group; | |
1da177e4 | 6325 | } |
6711cab4 | 6326 | |
08069033 SS |
6327 | static void init_numa_sched_groups_power(struct sched_group *group_head) |
6328 | { | |
6329 | struct sched_group *sg = group_head; | |
6330 | int j; | |
6331 | ||
6332 | if (!sg) | |
6333 | return; | |
6334 | next_sg: | |
6335 | for_each_cpu_mask(j, sg->cpumask) { | |
6336 | struct sched_domain *sd; | |
6337 | ||
6338 | sd = &per_cpu(phys_domains, j); | |
6339 | if (j != first_cpu(sd->groups->cpumask)) { | |
6340 | /* | |
6341 | * Only add "power" once for each | |
6342 | * physical package. | |
6343 | */ | |
6344 | continue; | |
6345 | } | |
6346 | ||
6347 | sg->cpu_power += sd->groups->cpu_power; | |
6348 | } | |
6349 | sg = sg->next; | |
6350 | if (sg != group_head) | |
6351 | goto next_sg; | |
6352 | } | |
1da177e4 LT |
6353 | #endif |
6354 | ||
a616058b | 6355 | #ifdef CONFIG_NUMA |
51888ca2 SV |
6356 | /* Free memory allocated for various sched_group structures */ |
6357 | static void free_sched_groups(const cpumask_t *cpu_map) | |
6358 | { | |
a616058b | 6359 | int cpu, i; |
51888ca2 SV |
6360 | |
6361 | for_each_cpu_mask(cpu, *cpu_map) { | |
51888ca2 SV |
6362 | struct sched_group **sched_group_nodes |
6363 | = sched_group_nodes_bycpu[cpu]; | |
6364 | ||
51888ca2 SV |
6365 | if (!sched_group_nodes) |
6366 | continue; | |
6367 | ||
6368 | for (i = 0; i < MAX_NUMNODES; i++) { | |
6369 | cpumask_t nodemask = node_to_cpumask(i); | |
6370 | struct sched_group *oldsg, *sg = sched_group_nodes[i]; | |
6371 | ||
6372 | cpus_and(nodemask, nodemask, *cpu_map); | |
6373 | if (cpus_empty(nodemask)) | |
6374 | continue; | |
6375 | ||
6376 | if (sg == NULL) | |
6377 | continue; | |
6378 | sg = sg->next; | |
6379 | next_sg: | |
6380 | oldsg = sg; | |
6381 | sg = sg->next; | |
6382 | kfree(oldsg); | |
6383 | if (oldsg != sched_group_nodes[i]) | |
6384 | goto next_sg; | |
6385 | } | |
6386 | kfree(sched_group_nodes); | |
6387 | sched_group_nodes_bycpu[cpu] = NULL; | |
6388 | } | |
51888ca2 | 6389 | } |
a616058b SS |
6390 | #else |
6391 | static void free_sched_groups(const cpumask_t *cpu_map) | |
6392 | { | |
6393 | } | |
6394 | #endif | |
51888ca2 | 6395 | |
89c4710e SS |
6396 | /* |
6397 | * Initialize sched groups cpu_power. | |
6398 | * | |
6399 | * cpu_power indicates the capacity of sched group, which is used while | |
6400 | * distributing the load between different sched groups in a sched domain. | |
6401 | * Typically cpu_power for all the groups in a sched domain will be same unless | |
6402 | * there are asymmetries in the topology. If there are asymmetries, group | |
6403 | * having more cpu_power will pickup more load compared to the group having | |
6404 | * less cpu_power. | |
6405 | * | |
6406 | * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents | |
6407 | * the maximum number of tasks a group can handle in the presence of other idle | |
6408 | * or lightly loaded groups in the same sched domain. | |
6409 | */ | |
6410 | static void init_sched_groups_power(int cpu, struct sched_domain *sd) | |
6411 | { | |
6412 | struct sched_domain *child; | |
6413 | struct sched_group *group; | |
6414 | ||
6415 | WARN_ON(!sd || !sd->groups); | |
6416 | ||
6417 | if (cpu != first_cpu(sd->groups->cpumask)) | |
6418 | return; | |
6419 | ||
6420 | child = sd->child; | |
6421 | ||
6422 | /* | |
6423 | * For perf policy, if the groups in child domain share resources | |
6424 | * (for example cores sharing some portions of the cache hierarchy | |
6425 | * or SMT), then set this domain groups cpu_power such that each group | |
6426 | * can handle only one task, when there are other idle groups in the | |
6427 | * same sched domain. | |
6428 | */ | |
6429 | if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) && | |
6430 | (child->flags & | |
6431 | (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) { | |
6432 | sd->groups->cpu_power = SCHED_LOAD_SCALE; | |
6433 | return; | |
6434 | } | |
6435 | ||
6436 | sd->groups->cpu_power = 0; | |
6437 | ||
6438 | /* | |
6439 | * add cpu_power of each child group to this groups cpu_power | |
6440 | */ | |
6441 | group = child->groups; | |
6442 | do { | |
6443 | sd->groups->cpu_power += group->cpu_power; | |
6444 | group = group->next; | |
6445 | } while (group != child->groups); | |
6446 | } | |
6447 | ||
1da177e4 | 6448 | /* |
1a20ff27 DG |
6449 | * Build sched domains for a given set of cpus and attach the sched domains |
6450 | * to the individual cpus | |
1da177e4 | 6451 | */ |
51888ca2 | 6452 | static int build_sched_domains(const cpumask_t *cpu_map) |
1da177e4 LT |
6453 | { |
6454 | int i; | |
89c4710e | 6455 | struct sched_domain *sd; |
d1b55138 JH |
6456 | #ifdef CONFIG_NUMA |
6457 | struct sched_group **sched_group_nodes = NULL; | |
6711cab4 | 6458 | int sd_allnodes = 0; |
d1b55138 JH |
6459 | |
6460 | /* | |
6461 | * Allocate the per-node list of sched groups | |
6462 | */ | |
51888ca2 | 6463 | sched_group_nodes = kzalloc(sizeof(struct sched_group*)*MAX_NUMNODES, |
d3a5aa98 | 6464 | GFP_KERNEL); |
d1b55138 JH |
6465 | if (!sched_group_nodes) { |
6466 | printk(KERN_WARNING "Can not alloc sched group node list\n"); | |
51888ca2 | 6467 | return -ENOMEM; |
d1b55138 JH |
6468 | } |
6469 | sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes; | |
6470 | #endif | |
1da177e4 LT |
6471 | |
6472 | /* | |
1a20ff27 | 6473 | * Set up domains for cpus specified by the cpu_map. |
1da177e4 | 6474 | */ |
1a20ff27 | 6475 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 LT |
6476 | struct sched_domain *sd = NULL, *p; |
6477 | cpumask_t nodemask = node_to_cpumask(cpu_to_node(i)); | |
6478 | ||
1a20ff27 | 6479 | cpus_and(nodemask, nodemask, *cpu_map); |
1da177e4 LT |
6480 | |
6481 | #ifdef CONFIG_NUMA | |
d1b55138 | 6482 | if (cpus_weight(*cpu_map) |
9c1cfda2 JH |
6483 | > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) { |
6484 | sd = &per_cpu(allnodes_domains, i); | |
6485 | *sd = SD_ALLNODES_INIT; | |
6486 | sd->span = *cpu_map; | |
6711cab4 | 6487 | cpu_to_allnodes_group(i, cpu_map, &sd->groups); |
9c1cfda2 | 6488 | p = sd; |
6711cab4 | 6489 | sd_allnodes = 1; |
9c1cfda2 JH |
6490 | } else |
6491 | p = NULL; | |
6492 | ||
1da177e4 | 6493 | sd = &per_cpu(node_domains, i); |
1da177e4 | 6494 | *sd = SD_NODE_INIT; |
9c1cfda2 JH |
6495 | sd->span = sched_domain_node_span(cpu_to_node(i)); |
6496 | sd->parent = p; | |
1a848870 SS |
6497 | if (p) |
6498 | p->child = sd; | |
9c1cfda2 | 6499 | cpus_and(sd->span, sd->span, *cpu_map); |
1da177e4 LT |
6500 | #endif |
6501 | ||
6502 | p = sd; | |
6503 | sd = &per_cpu(phys_domains, i); | |
1da177e4 LT |
6504 | *sd = SD_CPU_INIT; |
6505 | sd->span = nodemask; | |
6506 | sd->parent = p; | |
1a848870 SS |
6507 | if (p) |
6508 | p->child = sd; | |
6711cab4 | 6509 | cpu_to_phys_group(i, cpu_map, &sd->groups); |
1da177e4 | 6510 | |
1e9f28fa SS |
6511 | #ifdef CONFIG_SCHED_MC |
6512 | p = sd; | |
6513 | sd = &per_cpu(core_domains, i); | |
1e9f28fa SS |
6514 | *sd = SD_MC_INIT; |
6515 | sd->span = cpu_coregroup_map(i); | |
6516 | cpus_and(sd->span, sd->span, *cpu_map); | |
6517 | sd->parent = p; | |
1a848870 | 6518 | p->child = sd; |
6711cab4 | 6519 | cpu_to_core_group(i, cpu_map, &sd->groups); |
1e9f28fa SS |
6520 | #endif |
6521 | ||
1da177e4 LT |
6522 | #ifdef CONFIG_SCHED_SMT |
6523 | p = sd; | |
6524 | sd = &per_cpu(cpu_domains, i); | |
1da177e4 LT |
6525 | *sd = SD_SIBLING_INIT; |
6526 | sd->span = cpu_sibling_map[i]; | |
1a20ff27 | 6527 | cpus_and(sd->span, sd->span, *cpu_map); |
1da177e4 | 6528 | sd->parent = p; |
1a848870 | 6529 | p->child = sd; |
6711cab4 | 6530 | cpu_to_cpu_group(i, cpu_map, &sd->groups); |
1da177e4 LT |
6531 | #endif |
6532 | } | |
6533 | ||
6534 | #ifdef CONFIG_SCHED_SMT | |
6535 | /* Set up CPU (sibling) groups */ | |
9c1cfda2 | 6536 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 | 6537 | cpumask_t this_sibling_map = cpu_sibling_map[i]; |
1a20ff27 | 6538 | cpus_and(this_sibling_map, this_sibling_map, *cpu_map); |
1da177e4 LT |
6539 | if (i != first_cpu(this_sibling_map)) |
6540 | continue; | |
6541 | ||
6711cab4 | 6542 | init_sched_build_groups(this_sibling_map, cpu_map, &cpu_to_cpu_group); |
1da177e4 LT |
6543 | } |
6544 | #endif | |
6545 | ||
1e9f28fa SS |
6546 | #ifdef CONFIG_SCHED_MC |
6547 | /* Set up multi-core groups */ | |
6548 | for_each_cpu_mask(i, *cpu_map) { | |
6549 | cpumask_t this_core_map = cpu_coregroup_map(i); | |
6550 | cpus_and(this_core_map, this_core_map, *cpu_map); | |
6551 | if (i != first_cpu(this_core_map)) | |
6552 | continue; | |
6711cab4 | 6553 | init_sched_build_groups(this_core_map, cpu_map, &cpu_to_core_group); |
1e9f28fa SS |
6554 | } |
6555 | #endif | |
6556 | ||
6557 | ||
1da177e4 LT |
6558 | /* Set up physical groups */ |
6559 | for (i = 0; i < MAX_NUMNODES; i++) { | |
6560 | cpumask_t nodemask = node_to_cpumask(i); | |
6561 | ||
1a20ff27 | 6562 | cpus_and(nodemask, nodemask, *cpu_map); |
1da177e4 LT |
6563 | if (cpus_empty(nodemask)) |
6564 | continue; | |
6565 | ||
6711cab4 | 6566 | init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group); |
1da177e4 LT |
6567 | } |
6568 | ||
6569 | #ifdef CONFIG_NUMA | |
6570 | /* Set up node groups */ | |
6711cab4 SS |
6571 | if (sd_allnodes) |
6572 | init_sched_build_groups(*cpu_map, cpu_map, &cpu_to_allnodes_group); | |
9c1cfda2 JH |
6573 | |
6574 | for (i = 0; i < MAX_NUMNODES; i++) { | |
6575 | /* Set up node groups */ | |
6576 | struct sched_group *sg, *prev; | |
6577 | cpumask_t nodemask = node_to_cpumask(i); | |
6578 | cpumask_t domainspan; | |
6579 | cpumask_t covered = CPU_MASK_NONE; | |
6580 | int j; | |
6581 | ||
6582 | cpus_and(nodemask, nodemask, *cpu_map); | |
d1b55138 JH |
6583 | if (cpus_empty(nodemask)) { |
6584 | sched_group_nodes[i] = NULL; | |
9c1cfda2 | 6585 | continue; |
d1b55138 | 6586 | } |
9c1cfda2 JH |
6587 | |
6588 | domainspan = sched_domain_node_span(i); | |
6589 | cpus_and(domainspan, domainspan, *cpu_map); | |
6590 | ||
15f0b676 | 6591 | sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i); |
51888ca2 SV |
6592 | if (!sg) { |
6593 | printk(KERN_WARNING "Can not alloc domain group for " | |
6594 | "node %d\n", i); | |
6595 | goto error; | |
6596 | } | |
9c1cfda2 JH |
6597 | sched_group_nodes[i] = sg; |
6598 | for_each_cpu_mask(j, nodemask) { | |
6599 | struct sched_domain *sd; | |
6600 | sd = &per_cpu(node_domains, j); | |
6601 | sd->groups = sg; | |
9c1cfda2 JH |
6602 | } |
6603 | sg->cpu_power = 0; | |
6604 | sg->cpumask = nodemask; | |
51888ca2 | 6605 | sg->next = sg; |
9c1cfda2 JH |
6606 | cpus_or(covered, covered, nodemask); |
6607 | prev = sg; | |
6608 | ||
6609 | for (j = 0; j < MAX_NUMNODES; j++) { | |
6610 | cpumask_t tmp, notcovered; | |
6611 | int n = (i + j) % MAX_NUMNODES; | |
6612 | ||
6613 | cpus_complement(notcovered, covered); | |
6614 | cpus_and(tmp, notcovered, *cpu_map); | |
6615 | cpus_and(tmp, tmp, domainspan); | |
6616 | if (cpus_empty(tmp)) | |
6617 | break; | |
6618 | ||
6619 | nodemask = node_to_cpumask(n); | |
6620 | cpus_and(tmp, tmp, nodemask); | |
6621 | if (cpus_empty(tmp)) | |
6622 | continue; | |
6623 | ||
15f0b676 SV |
6624 | sg = kmalloc_node(sizeof(struct sched_group), |
6625 | GFP_KERNEL, i); | |
9c1cfda2 JH |
6626 | if (!sg) { |
6627 | printk(KERN_WARNING | |
6628 | "Can not alloc domain group for node %d\n", j); | |
51888ca2 | 6629 | goto error; |
9c1cfda2 JH |
6630 | } |
6631 | sg->cpu_power = 0; | |
6632 | sg->cpumask = tmp; | |
51888ca2 | 6633 | sg->next = prev->next; |
9c1cfda2 JH |
6634 | cpus_or(covered, covered, tmp); |
6635 | prev->next = sg; | |
6636 | prev = sg; | |
6637 | } | |
9c1cfda2 | 6638 | } |
1da177e4 LT |
6639 | #endif |
6640 | ||
6641 | /* Calculate CPU power for physical packages and nodes */ | |
5c45bf27 | 6642 | #ifdef CONFIG_SCHED_SMT |
1a20ff27 | 6643 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 | 6644 | sd = &per_cpu(cpu_domains, i); |
89c4710e | 6645 | init_sched_groups_power(i, sd); |
5c45bf27 | 6646 | } |
1da177e4 | 6647 | #endif |
1e9f28fa | 6648 | #ifdef CONFIG_SCHED_MC |
5c45bf27 | 6649 | for_each_cpu_mask(i, *cpu_map) { |
1e9f28fa | 6650 | sd = &per_cpu(core_domains, i); |
89c4710e | 6651 | init_sched_groups_power(i, sd); |
5c45bf27 SS |
6652 | } |
6653 | #endif | |
1e9f28fa | 6654 | |
5c45bf27 | 6655 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 | 6656 | sd = &per_cpu(phys_domains, i); |
89c4710e | 6657 | init_sched_groups_power(i, sd); |
1da177e4 LT |
6658 | } |
6659 | ||
9c1cfda2 | 6660 | #ifdef CONFIG_NUMA |
08069033 SS |
6661 | for (i = 0; i < MAX_NUMNODES; i++) |
6662 | init_numa_sched_groups_power(sched_group_nodes[i]); | |
9c1cfda2 | 6663 | |
6711cab4 SS |
6664 | if (sd_allnodes) { |
6665 | struct sched_group *sg; | |
f712c0c7 | 6666 | |
6711cab4 | 6667 | cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg); |
f712c0c7 SS |
6668 | init_numa_sched_groups_power(sg); |
6669 | } | |
9c1cfda2 JH |
6670 | #endif |
6671 | ||
1da177e4 | 6672 | /* Attach the domains */ |
1a20ff27 | 6673 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 LT |
6674 | struct sched_domain *sd; |
6675 | #ifdef CONFIG_SCHED_SMT | |
6676 | sd = &per_cpu(cpu_domains, i); | |
1e9f28fa SS |
6677 | #elif defined(CONFIG_SCHED_MC) |
6678 | sd = &per_cpu(core_domains, i); | |
1da177e4 LT |
6679 | #else |
6680 | sd = &per_cpu(phys_domains, i); | |
6681 | #endif | |
6682 | cpu_attach_domain(sd, i); | |
6683 | } | |
198e2f18 | 6684 | /* |
6685 | * Tune cache-hot values: | |
6686 | */ | |
6687 | calibrate_migration_costs(cpu_map); | |
51888ca2 SV |
6688 | |
6689 | return 0; | |
6690 | ||
a616058b | 6691 | #ifdef CONFIG_NUMA |
51888ca2 SV |
6692 | error: |
6693 | free_sched_groups(cpu_map); | |
6694 | return -ENOMEM; | |
a616058b | 6695 | #endif |
1da177e4 | 6696 | } |
1a20ff27 DG |
6697 | /* |
6698 | * Set up scheduler domains and groups. Callers must hold the hotplug lock. | |
6699 | */ | |
51888ca2 | 6700 | static int arch_init_sched_domains(const cpumask_t *cpu_map) |
1a20ff27 DG |
6701 | { |
6702 | cpumask_t cpu_default_map; | |
51888ca2 | 6703 | int err; |
1da177e4 | 6704 | |
1a20ff27 DG |
6705 | /* |
6706 | * Setup mask for cpus without special case scheduling requirements. | |
6707 | * For now this just excludes isolated cpus, but could be used to | |
6708 | * exclude other special cases in the future. | |
6709 | */ | |
6710 | cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map); | |
6711 | ||
51888ca2 SV |
6712 | err = build_sched_domains(&cpu_default_map); |
6713 | ||
6714 | return err; | |
1a20ff27 DG |
6715 | } |
6716 | ||
6717 | static void arch_destroy_sched_domains(const cpumask_t *cpu_map) | |
1da177e4 | 6718 | { |
51888ca2 | 6719 | free_sched_groups(cpu_map); |
9c1cfda2 | 6720 | } |
1da177e4 | 6721 | |
1a20ff27 DG |
6722 | /* |
6723 | * Detach sched domains from a group of cpus specified in cpu_map | |
6724 | * These cpus will now be attached to the NULL domain | |
6725 | */ | |
858119e1 | 6726 | static void detach_destroy_domains(const cpumask_t *cpu_map) |
1a20ff27 DG |
6727 | { |
6728 | int i; | |
6729 | ||
6730 | for_each_cpu_mask(i, *cpu_map) | |
6731 | cpu_attach_domain(NULL, i); | |
6732 | synchronize_sched(); | |
6733 | arch_destroy_sched_domains(cpu_map); | |
6734 | } | |
6735 | ||
6736 | /* | |
6737 | * Partition sched domains as specified by the cpumasks below. | |
6738 | * This attaches all cpus from the cpumasks to the NULL domain, | |
6739 | * waits for a RCU quiescent period, recalculates sched | |
6740 | * domain information and then attaches them back to the | |
6741 | * correct sched domains | |
6742 | * Call with hotplug lock held | |
6743 | */ | |
51888ca2 | 6744 | int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2) |
1a20ff27 DG |
6745 | { |
6746 | cpumask_t change_map; | |
51888ca2 | 6747 | int err = 0; |
1a20ff27 DG |
6748 | |
6749 | cpus_and(*partition1, *partition1, cpu_online_map); | |
6750 | cpus_and(*partition2, *partition2, cpu_online_map); | |
6751 | cpus_or(change_map, *partition1, *partition2); | |
6752 | ||
6753 | /* Detach sched domains from all of the affected cpus */ | |
6754 | detach_destroy_domains(&change_map); | |
6755 | if (!cpus_empty(*partition1)) | |
51888ca2 SV |
6756 | err = build_sched_domains(partition1); |
6757 | if (!err && !cpus_empty(*partition2)) | |
6758 | err = build_sched_domains(partition2); | |
6759 | ||
6760 | return err; | |
1a20ff27 DG |
6761 | } |
6762 | ||
5c45bf27 SS |
6763 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
6764 | int arch_reinit_sched_domains(void) | |
6765 | { | |
6766 | int err; | |
6767 | ||
6768 | lock_cpu_hotplug(); | |
6769 | detach_destroy_domains(&cpu_online_map); | |
6770 | err = arch_init_sched_domains(&cpu_online_map); | |
6771 | unlock_cpu_hotplug(); | |
6772 | ||
6773 | return err; | |
6774 | } | |
6775 | ||
6776 | static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt) | |
6777 | { | |
6778 | int ret; | |
6779 | ||
6780 | if (buf[0] != '0' && buf[0] != '1') | |
6781 | return -EINVAL; | |
6782 | ||
6783 | if (smt) | |
6784 | sched_smt_power_savings = (buf[0] == '1'); | |
6785 | else | |
6786 | sched_mc_power_savings = (buf[0] == '1'); | |
6787 | ||
6788 | ret = arch_reinit_sched_domains(); | |
6789 | ||
6790 | return ret ? ret : count; | |
6791 | } | |
6792 | ||
6793 | int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls) | |
6794 | { | |
6795 | int err = 0; | |
48f24c4d | 6796 | |
5c45bf27 SS |
6797 | #ifdef CONFIG_SCHED_SMT |
6798 | if (smt_capable()) | |
6799 | err = sysfs_create_file(&cls->kset.kobj, | |
6800 | &attr_sched_smt_power_savings.attr); | |
6801 | #endif | |
6802 | #ifdef CONFIG_SCHED_MC | |
6803 | if (!err && mc_capable()) | |
6804 | err = sysfs_create_file(&cls->kset.kobj, | |
6805 | &attr_sched_mc_power_savings.attr); | |
6806 | #endif | |
6807 | return err; | |
6808 | } | |
6809 | #endif | |
6810 | ||
6811 | #ifdef CONFIG_SCHED_MC | |
6812 | static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page) | |
6813 | { | |
6814 | return sprintf(page, "%u\n", sched_mc_power_savings); | |
6815 | } | |
48f24c4d IM |
6816 | static ssize_t sched_mc_power_savings_store(struct sys_device *dev, |
6817 | const char *buf, size_t count) | |
5c45bf27 SS |
6818 | { |
6819 | return sched_power_savings_store(buf, count, 0); | |
6820 | } | |
6821 | SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show, | |
6822 | sched_mc_power_savings_store); | |
6823 | #endif | |
6824 | ||
6825 | #ifdef CONFIG_SCHED_SMT | |
6826 | static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page) | |
6827 | { | |
6828 | return sprintf(page, "%u\n", sched_smt_power_savings); | |
6829 | } | |
48f24c4d IM |
6830 | static ssize_t sched_smt_power_savings_store(struct sys_device *dev, |
6831 | const char *buf, size_t count) | |
5c45bf27 SS |
6832 | { |
6833 | return sched_power_savings_store(buf, count, 1); | |
6834 | } | |
6835 | SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show, | |
6836 | sched_smt_power_savings_store); | |
6837 | #endif | |
6838 | ||
1da177e4 LT |
6839 | /* |
6840 | * Force a reinitialization of the sched domains hierarchy. The domains | |
6841 | * and groups cannot be updated in place without racing with the balancing | |
41c7ce9a | 6842 | * code, so we temporarily attach all running cpus to the NULL domain |
1da177e4 LT |
6843 | * which will prevent rebalancing while the sched domains are recalculated. |
6844 | */ | |
6845 | static int update_sched_domains(struct notifier_block *nfb, | |
6846 | unsigned long action, void *hcpu) | |
6847 | { | |
1da177e4 LT |
6848 | switch (action) { |
6849 | case CPU_UP_PREPARE: | |
6850 | case CPU_DOWN_PREPARE: | |
1a20ff27 | 6851 | detach_destroy_domains(&cpu_online_map); |
1da177e4 LT |
6852 | return NOTIFY_OK; |
6853 | ||
6854 | case CPU_UP_CANCELED: | |
6855 | case CPU_DOWN_FAILED: | |
6856 | case CPU_ONLINE: | |
6857 | case CPU_DEAD: | |
6858 | /* | |
6859 | * Fall through and re-initialise the domains. | |
6860 | */ | |
6861 | break; | |
6862 | default: | |
6863 | return NOTIFY_DONE; | |
6864 | } | |
6865 | ||
6866 | /* The hotplug lock is already held by cpu_up/cpu_down */ | |
1a20ff27 | 6867 | arch_init_sched_domains(&cpu_online_map); |
1da177e4 LT |
6868 | |
6869 | return NOTIFY_OK; | |
6870 | } | |
1da177e4 LT |
6871 | |
6872 | void __init sched_init_smp(void) | |
6873 | { | |
5c1e1767 NP |
6874 | cpumask_t non_isolated_cpus; |
6875 | ||
1da177e4 | 6876 | lock_cpu_hotplug(); |
1a20ff27 | 6877 | arch_init_sched_domains(&cpu_online_map); |
e5e5673f | 6878 | cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map); |
5c1e1767 NP |
6879 | if (cpus_empty(non_isolated_cpus)) |
6880 | cpu_set(smp_processor_id(), non_isolated_cpus); | |
1da177e4 LT |
6881 | unlock_cpu_hotplug(); |
6882 | /* XXX: Theoretical race here - CPU may be hotplugged now */ | |
6883 | hotcpu_notifier(update_sched_domains, 0); | |
5c1e1767 NP |
6884 | |
6885 | /* Move init over to a non-isolated CPU */ | |
6886 | if (set_cpus_allowed(current, non_isolated_cpus) < 0) | |
6887 | BUG(); | |
1da177e4 LT |
6888 | } |
6889 | #else | |
6890 | void __init sched_init_smp(void) | |
6891 | { | |
6892 | } | |
6893 | #endif /* CONFIG_SMP */ | |
6894 | ||
6895 | int in_sched_functions(unsigned long addr) | |
6896 | { | |
6897 | /* Linker adds these: start and end of __sched functions */ | |
6898 | extern char __sched_text_start[], __sched_text_end[]; | |
48f24c4d | 6899 | |
1da177e4 LT |
6900 | return in_lock_functions(addr) || |
6901 | (addr >= (unsigned long)__sched_text_start | |
6902 | && addr < (unsigned long)__sched_text_end); | |
6903 | } | |
6904 | ||
6905 | void __init sched_init(void) | |
6906 | { | |
1da177e4 LT |
6907 | int i, j, k; |
6908 | ||
0a945022 | 6909 | for_each_possible_cpu(i) { |
70b97a7f IM |
6910 | struct prio_array *array; |
6911 | struct rq *rq; | |
1da177e4 LT |
6912 | |
6913 | rq = cpu_rq(i); | |
6914 | spin_lock_init(&rq->lock); | |
fcb99371 | 6915 | lockdep_set_class(&rq->lock, &rq->rq_lock_key); |
7897986b | 6916 | rq->nr_running = 0; |
1da177e4 LT |
6917 | rq->active = rq->arrays; |
6918 | rq->expired = rq->arrays + 1; | |
6919 | rq->best_expired_prio = MAX_PRIO; | |
6920 | ||
6921 | #ifdef CONFIG_SMP | |
41c7ce9a | 6922 | rq->sd = NULL; |
7897986b NP |
6923 | for (j = 1; j < 3; j++) |
6924 | rq->cpu_load[j] = 0; | |
1da177e4 LT |
6925 | rq->active_balance = 0; |
6926 | rq->push_cpu = 0; | |
0a2966b4 | 6927 | rq->cpu = i; |
1da177e4 LT |
6928 | rq->migration_thread = NULL; |
6929 | INIT_LIST_HEAD(&rq->migration_queue); | |
6930 | #endif | |
6931 | atomic_set(&rq->nr_iowait, 0); | |
6932 | ||
6933 | for (j = 0; j < 2; j++) { | |
6934 | array = rq->arrays + j; | |
6935 | for (k = 0; k < MAX_PRIO; k++) { | |
6936 | INIT_LIST_HEAD(array->queue + k); | |
6937 | __clear_bit(k, array->bitmap); | |
6938 | } | |
6939 | // delimiter for bitsearch | |
6940 | __set_bit(MAX_PRIO, array->bitmap); | |
6941 | } | |
6942 | } | |
6943 | ||
2dd73a4f | 6944 | set_load_weight(&init_task); |
b50f60ce | 6945 | |
c9819f45 CL |
6946 | #ifdef CONFIG_SMP |
6947 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL); | |
6948 | #endif | |
6949 | ||
b50f60ce HC |
6950 | #ifdef CONFIG_RT_MUTEXES |
6951 | plist_head_init(&init_task.pi_waiters, &init_task.pi_lock); | |
6952 | #endif | |
6953 | ||
1da177e4 LT |
6954 | /* |
6955 | * The boot idle thread does lazy MMU switching as well: | |
6956 | */ | |
6957 | atomic_inc(&init_mm.mm_count); | |
6958 | enter_lazy_tlb(&init_mm, current); | |
6959 | ||
6960 | /* | |
6961 | * Make us the idle thread. Technically, schedule() should not be | |
6962 | * called from this thread, however somewhere below it might be, | |
6963 | * but because we are the idle thread, we just pick up running again | |
6964 | * when this runqueue becomes "idle". | |
6965 | */ | |
6966 | init_idle(current, smp_processor_id()); | |
6967 | } | |
6968 | ||
6969 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP | |
6970 | void __might_sleep(char *file, int line) | |
6971 | { | |
48f24c4d | 6972 | #ifdef in_atomic |
1da177e4 LT |
6973 | static unsigned long prev_jiffy; /* ratelimiting */ |
6974 | ||
6975 | if ((in_atomic() || irqs_disabled()) && | |
6976 | system_state == SYSTEM_RUNNING && !oops_in_progress) { | |
6977 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) | |
6978 | return; | |
6979 | prev_jiffy = jiffies; | |
91368d73 | 6980 | printk(KERN_ERR "BUG: sleeping function called from invalid" |
1da177e4 LT |
6981 | " context at %s:%d\n", file, line); |
6982 | printk("in_atomic():%d, irqs_disabled():%d\n", | |
6983 | in_atomic(), irqs_disabled()); | |
a4c410f0 | 6984 | debug_show_held_locks(current); |
3117df04 IM |
6985 | if (irqs_disabled()) |
6986 | print_irqtrace_events(current); | |
1da177e4 LT |
6987 | dump_stack(); |
6988 | } | |
6989 | #endif | |
6990 | } | |
6991 | EXPORT_SYMBOL(__might_sleep); | |
6992 | #endif | |
6993 | ||
6994 | #ifdef CONFIG_MAGIC_SYSRQ | |
6995 | void normalize_rt_tasks(void) | |
6996 | { | |
70b97a7f | 6997 | struct prio_array *array; |
1da177e4 | 6998 | struct task_struct *p; |
1da177e4 | 6999 | unsigned long flags; |
70b97a7f | 7000 | struct rq *rq; |
1da177e4 LT |
7001 | |
7002 | read_lock_irq(&tasklist_lock); | |
c96d145e | 7003 | for_each_process(p) { |
1da177e4 LT |
7004 | if (!rt_task(p)) |
7005 | continue; | |
7006 | ||
b29739f9 IM |
7007 | spin_lock_irqsave(&p->pi_lock, flags); |
7008 | rq = __task_rq_lock(p); | |
1da177e4 LT |
7009 | |
7010 | array = p->array; | |
7011 | if (array) | |
7012 | deactivate_task(p, task_rq(p)); | |
7013 | __setscheduler(p, SCHED_NORMAL, 0); | |
7014 | if (array) { | |
7015 | __activate_task(p, task_rq(p)); | |
7016 | resched_task(rq->curr); | |
7017 | } | |
7018 | ||
b29739f9 IM |
7019 | __task_rq_unlock(rq); |
7020 | spin_unlock_irqrestore(&p->pi_lock, flags); | |
1da177e4 LT |
7021 | } |
7022 | read_unlock_irq(&tasklist_lock); | |
7023 | } | |
7024 | ||
7025 | #endif /* CONFIG_MAGIC_SYSRQ */ | |
1df5c10a LT |
7026 | |
7027 | #ifdef CONFIG_IA64 | |
7028 | /* | |
7029 | * These functions are only useful for the IA64 MCA handling. | |
7030 | * | |
7031 | * They can only be called when the whole system has been | |
7032 | * stopped - every CPU needs to be quiescent, and no scheduling | |
7033 | * activity can take place. Using them for anything else would | |
7034 | * be a serious bug, and as a result, they aren't even visible | |
7035 | * under any other configuration. | |
7036 | */ | |
7037 | ||
7038 | /** | |
7039 | * curr_task - return the current task for a given cpu. | |
7040 | * @cpu: the processor in question. | |
7041 | * | |
7042 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
7043 | */ | |
36c8b586 | 7044 | struct task_struct *curr_task(int cpu) |
1df5c10a LT |
7045 | { |
7046 | return cpu_curr(cpu); | |
7047 | } | |
7048 | ||
7049 | /** | |
7050 | * set_curr_task - set the current task for a given cpu. | |
7051 | * @cpu: the processor in question. | |
7052 | * @p: the task pointer to set. | |
7053 | * | |
7054 | * Description: This function must only be used when non-maskable interrupts | |
7055 | * are serviced on a separate stack. It allows the architecture to switch the | |
7056 | * notion of the current task on a cpu in a non-blocking manner. This function | |
7057 | * must be called with all CPU's synchronized, and interrupts disabled, the | |
7058 | * and caller must save the original value of the current task (see | |
7059 | * curr_task() above) and restore that value before reenabling interrupts and | |
7060 | * re-starting the system. | |
7061 | * | |
7062 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
7063 | */ | |
36c8b586 | 7064 | void set_curr_task(int cpu, struct task_struct *p) |
1df5c10a LT |
7065 | { |
7066 | cpu_curr(cpu) = p; | |
7067 | } | |
7068 | ||
7069 | #endif |