]>
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> |
5517d86b | 55 | #include <linux/reciprocal_div.h> |
1da177e4 | 56 | |
5517d86b | 57 | #include <asm/tlb.h> |
1da177e4 LT |
58 | #include <asm/unistd.h> |
59 | ||
b035b6de AD |
60 | /* |
61 | * Scheduler clock - returns current time in nanosec units. | |
62 | * This is default implementation. | |
63 | * Architectures and sub-architectures can override this. | |
64 | */ | |
65 | unsigned long long __attribute__((weak)) sched_clock(void) | |
66 | { | |
67 | return (unsigned long long)jiffies * (1000000000 / HZ); | |
68 | } | |
69 | ||
1da177e4 LT |
70 | /* |
71 | * Convert user-nice values [ -20 ... 0 ... 19 ] | |
72 | * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ], | |
73 | * and back. | |
74 | */ | |
75 | #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20) | |
76 | #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20) | |
77 | #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio) | |
78 | ||
79 | /* | |
80 | * 'User priority' is the nice value converted to something we | |
81 | * can work with better when scaling various scheduler parameters, | |
82 | * it's a [ 0 ... 39 ] range. | |
83 | */ | |
84 | #define USER_PRIO(p) ((p)-MAX_RT_PRIO) | |
85 | #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio) | |
86 | #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO)) | |
87 | ||
88 | /* | |
89 | * Some helpers for converting nanosecond timing to jiffy resolution | |
90 | */ | |
91 | #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ)) | |
92 | #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ)) | |
93 | ||
94 | /* | |
95 | * These are the 'tuning knobs' of the scheduler: | |
96 | * | |
97 | * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger), | |
98 | * default timeslice is 100 msecs, maximum timeslice is 800 msecs. | |
99 | * Timeslices get refilled after they expire. | |
100 | */ | |
101 | #define MIN_TIMESLICE max(5 * HZ / 1000, 1) | |
102 | #define DEF_TIMESLICE (100 * HZ / 1000) | |
103 | #define ON_RUNQUEUE_WEIGHT 30 | |
104 | #define CHILD_PENALTY 95 | |
105 | #define PARENT_PENALTY 100 | |
106 | #define EXIT_WEIGHT 3 | |
107 | #define PRIO_BONUS_RATIO 25 | |
108 | #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100) | |
109 | #define INTERACTIVE_DELTA 2 | |
110 | #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS) | |
111 | #define STARVATION_LIMIT (MAX_SLEEP_AVG) | |
112 | #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG)) | |
113 | ||
114 | /* | |
115 | * If a task is 'interactive' then we reinsert it in the active | |
116 | * array after it has expired its current timeslice. (it will not | |
117 | * continue to run immediately, it will still roundrobin with | |
118 | * other interactive tasks.) | |
119 | * | |
120 | * This part scales the interactivity limit depending on niceness. | |
121 | * | |
122 | * We scale it linearly, offset by the INTERACTIVE_DELTA delta. | |
123 | * Here are a few examples of different nice levels: | |
124 | * | |
125 | * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0] | |
126 | * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0] | |
127 | * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0] | |
128 | * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0] | |
129 | * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0] | |
130 | * | |
131 | * (the X axis represents the possible -5 ... 0 ... +5 dynamic | |
132 | * priority range a task can explore, a value of '1' means the | |
133 | * task is rated interactive.) | |
134 | * | |
135 | * Ie. nice +19 tasks can never get 'interactive' enough to be | |
136 | * reinserted into the active array. And only heavily CPU-hog nice -20 | |
137 | * tasks will be expired. Default nice 0 tasks are somewhere between, | |
138 | * it takes some effort for them to get interactive, but it's not | |
139 | * too hard. | |
140 | */ | |
141 | ||
142 | #define CURRENT_BONUS(p) \ | |
143 | (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \ | |
144 | MAX_SLEEP_AVG) | |
145 | ||
146 | #define GRANULARITY (10 * HZ / 1000 ? : 1) | |
147 | ||
148 | #ifdef CONFIG_SMP | |
149 | #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \ | |
150 | (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \ | |
151 | num_online_cpus()) | |
152 | #else | |
153 | #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \ | |
154 | (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1))) | |
155 | #endif | |
156 | ||
157 | #define SCALE(v1,v1_max,v2_max) \ | |
158 | (v1) * (v2_max) / (v1_max) | |
159 | ||
160 | #define DELTA(p) \ | |
013d3868 MA |
161 | (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \ |
162 | INTERACTIVE_DELTA) | |
1da177e4 LT |
163 | |
164 | #define TASK_INTERACTIVE(p) \ | |
165 | ((p)->prio <= (p)->static_prio - DELTA(p)) | |
166 | ||
167 | #define INTERACTIVE_SLEEP(p) \ | |
168 | (JIFFIES_TO_NS(MAX_SLEEP_AVG * \ | |
169 | (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1)) | |
170 | ||
171 | #define TASK_PREEMPTS_CURR(p, rq) \ | |
172 | ((p)->prio < (rq)->curr->prio) | |
173 | ||
1da177e4 | 174 | #define SCALE_PRIO(x, prio) \ |
2dd73a4f | 175 | max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE) |
1da177e4 | 176 | |
2dd73a4f | 177 | static unsigned int static_prio_timeslice(int static_prio) |
1da177e4 | 178 | { |
2dd73a4f PW |
179 | if (static_prio < NICE_TO_PRIO(0)) |
180 | return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio); | |
1da177e4 | 181 | else |
2dd73a4f | 182 | return SCALE_PRIO(DEF_TIMESLICE, static_prio); |
1da177e4 | 183 | } |
2dd73a4f | 184 | |
5517d86b ED |
185 | #ifdef CONFIG_SMP |
186 | /* | |
187 | * Divide a load by a sched group cpu_power : (load / sg->__cpu_power) | |
188 | * Since cpu_power is a 'constant', we can use a reciprocal divide. | |
189 | */ | |
190 | static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load) | |
191 | { | |
192 | return reciprocal_divide(load, sg->reciprocal_cpu_power); | |
193 | } | |
194 | ||
195 | /* | |
196 | * Each time a sched group cpu_power is changed, | |
197 | * we must compute its reciprocal value | |
198 | */ | |
199 | static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val) | |
200 | { | |
201 | sg->__cpu_power += val; | |
202 | sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power); | |
203 | } | |
204 | #endif | |
205 | ||
91fcdd4e BP |
206 | /* |
207 | * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ] | |
208 | * to time slice values: [800ms ... 100ms ... 5ms] | |
209 | * | |
210 | * The higher a thread's priority, the bigger timeslices | |
211 | * it gets during one round of execution. But even the lowest | |
212 | * priority thread gets MIN_TIMESLICE worth of execution time. | |
213 | */ | |
214 | ||
36c8b586 | 215 | static inline unsigned int task_timeslice(struct task_struct *p) |
2dd73a4f PW |
216 | { |
217 | return static_prio_timeslice(p->static_prio); | |
218 | } | |
219 | ||
1da177e4 LT |
220 | /* |
221 | * These are the runqueue data structures: | |
222 | */ | |
223 | ||
1da177e4 LT |
224 | struct prio_array { |
225 | unsigned int nr_active; | |
d444886e | 226 | DECLARE_BITMAP(bitmap, MAX_PRIO+1); /* include 1 bit for delimiter */ |
1da177e4 LT |
227 | struct list_head queue[MAX_PRIO]; |
228 | }; | |
229 | ||
230 | /* | |
231 | * This is the main, per-CPU runqueue data structure. | |
232 | * | |
233 | * Locking rule: those places that want to lock multiple runqueues | |
234 | * (such as the load balancing or the thread migration code), lock | |
235 | * acquire operations must be ordered by ascending &runqueue. | |
236 | */ | |
70b97a7f | 237 | struct rq { |
1da177e4 LT |
238 | spinlock_t lock; |
239 | ||
240 | /* | |
241 | * nr_running and cpu_load should be in the same cacheline because | |
242 | * remote CPUs use both these fields when doing load calculation. | |
243 | */ | |
244 | unsigned long nr_running; | |
2dd73a4f | 245 | unsigned long raw_weighted_load; |
1da177e4 | 246 | #ifdef CONFIG_SMP |
7897986b | 247 | unsigned long cpu_load[3]; |
bdecea3a | 248 | unsigned char idle_at_tick; |
46cb4b7c SS |
249 | #ifdef CONFIG_NO_HZ |
250 | unsigned char in_nohz_recently; | |
251 | #endif | |
1da177e4 LT |
252 | #endif |
253 | unsigned long long nr_switches; | |
254 | ||
255 | /* | |
256 | * This is part of a global counter where only the total sum | |
257 | * over all CPUs matters. A task can increase this counter on | |
258 | * one CPU and if it got migrated afterwards it may decrease | |
259 | * it on another CPU. Always updated under the runqueue lock: | |
260 | */ | |
261 | unsigned long nr_uninterruptible; | |
262 | ||
263 | unsigned long expired_timestamp; | |
b18ec803 MG |
264 | /* Cached timestamp set by update_cpu_clock() */ |
265 | unsigned long long most_recent_timestamp; | |
36c8b586 | 266 | struct task_struct *curr, *idle; |
c9819f45 | 267 | unsigned long next_balance; |
1da177e4 | 268 | struct mm_struct *prev_mm; |
70b97a7f | 269 | struct prio_array *active, *expired, arrays[2]; |
1da177e4 LT |
270 | int best_expired_prio; |
271 | atomic_t nr_iowait; | |
272 | ||
273 | #ifdef CONFIG_SMP | |
274 | struct sched_domain *sd; | |
275 | ||
276 | /* For active balancing */ | |
277 | int active_balance; | |
278 | int push_cpu; | |
0a2966b4 | 279 | int cpu; /* cpu of this runqueue */ |
1da177e4 | 280 | |
36c8b586 | 281 | struct task_struct *migration_thread; |
1da177e4 LT |
282 | struct list_head migration_queue; |
283 | #endif | |
284 | ||
285 | #ifdef CONFIG_SCHEDSTATS | |
286 | /* latency stats */ | |
287 | struct sched_info rq_sched_info; | |
288 | ||
289 | /* sys_sched_yield() stats */ | |
290 | unsigned long yld_exp_empty; | |
291 | unsigned long yld_act_empty; | |
292 | unsigned long yld_both_empty; | |
293 | unsigned long yld_cnt; | |
294 | ||
295 | /* schedule() stats */ | |
296 | unsigned long sched_switch; | |
297 | unsigned long sched_cnt; | |
298 | unsigned long sched_goidle; | |
299 | ||
300 | /* try_to_wake_up() stats */ | |
301 | unsigned long ttwu_cnt; | |
302 | unsigned long ttwu_local; | |
303 | #endif | |
fcb99371 | 304 | struct lock_class_key rq_lock_key; |
1da177e4 LT |
305 | }; |
306 | ||
70b97a7f | 307 | static DEFINE_PER_CPU(struct rq, runqueues); |
1da177e4 | 308 | |
0a2966b4 CL |
309 | static inline int cpu_of(struct rq *rq) |
310 | { | |
311 | #ifdef CONFIG_SMP | |
312 | return rq->cpu; | |
313 | #else | |
314 | return 0; | |
315 | #endif | |
316 | } | |
317 | ||
674311d5 NP |
318 | /* |
319 | * The domain tree (rq->sd) is protected by RCU's quiescent state transition. | |
1a20ff27 | 320 | * See detach_destroy_domains: synchronize_sched for details. |
674311d5 NP |
321 | * |
322 | * The domain tree of any CPU may only be accessed from within | |
323 | * preempt-disabled sections. | |
324 | */ | |
48f24c4d IM |
325 | #define for_each_domain(cpu, __sd) \ |
326 | for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent) | |
1da177e4 LT |
327 | |
328 | #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) | |
329 | #define this_rq() (&__get_cpu_var(runqueues)) | |
330 | #define task_rq(p) cpu_rq(task_cpu(p)) | |
331 | #define cpu_curr(cpu) (cpu_rq(cpu)->curr) | |
332 | ||
1da177e4 | 333 | #ifndef prepare_arch_switch |
4866cde0 NP |
334 | # define prepare_arch_switch(next) do { } while (0) |
335 | #endif | |
336 | #ifndef finish_arch_switch | |
337 | # define finish_arch_switch(prev) do { } while (0) | |
338 | #endif | |
339 | ||
340 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW | |
70b97a7f | 341 | static inline int task_running(struct rq *rq, struct task_struct *p) |
4866cde0 NP |
342 | { |
343 | return rq->curr == p; | |
344 | } | |
345 | ||
70b97a7f | 346 | static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) |
4866cde0 NP |
347 | { |
348 | } | |
349 | ||
70b97a7f | 350 | static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) |
4866cde0 | 351 | { |
da04c035 IM |
352 | #ifdef CONFIG_DEBUG_SPINLOCK |
353 | /* this is a valid case when another task releases the spinlock */ | |
354 | rq->lock.owner = current; | |
355 | #endif | |
8a25d5de IM |
356 | /* |
357 | * If we are tracking spinlock dependencies then we have to | |
358 | * fix up the runqueue lock - which gets 'carried over' from | |
359 | * prev into current: | |
360 | */ | |
361 | spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_); | |
362 | ||
4866cde0 NP |
363 | spin_unlock_irq(&rq->lock); |
364 | } | |
365 | ||
366 | #else /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
70b97a7f | 367 | static inline int task_running(struct rq *rq, struct task_struct *p) |
4866cde0 NP |
368 | { |
369 | #ifdef CONFIG_SMP | |
370 | return p->oncpu; | |
371 | #else | |
372 | return rq->curr == p; | |
373 | #endif | |
374 | } | |
375 | ||
70b97a7f | 376 | static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) |
4866cde0 NP |
377 | { |
378 | #ifdef CONFIG_SMP | |
379 | /* | |
380 | * We can optimise this out completely for !SMP, because the | |
381 | * SMP rebalancing from interrupt is the only thing that cares | |
382 | * here. | |
383 | */ | |
384 | next->oncpu = 1; | |
385 | #endif | |
386 | #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
387 | spin_unlock_irq(&rq->lock); | |
388 | #else | |
389 | spin_unlock(&rq->lock); | |
390 | #endif | |
391 | } | |
392 | ||
70b97a7f | 393 | static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) |
4866cde0 NP |
394 | { |
395 | #ifdef CONFIG_SMP | |
396 | /* | |
397 | * After ->oncpu is cleared, the task can be moved to a different CPU. | |
398 | * We must ensure this doesn't happen until the switch is completely | |
399 | * finished. | |
400 | */ | |
401 | smp_wmb(); | |
402 | prev->oncpu = 0; | |
403 | #endif | |
404 | #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
405 | local_irq_enable(); | |
1da177e4 | 406 | #endif |
4866cde0 NP |
407 | } |
408 | #endif /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
1da177e4 | 409 | |
b29739f9 IM |
410 | /* |
411 | * __task_rq_lock - lock the runqueue a given task resides on. | |
412 | * Must be called interrupts disabled. | |
413 | */ | |
70b97a7f | 414 | static inline struct rq *__task_rq_lock(struct task_struct *p) |
b29739f9 IM |
415 | __acquires(rq->lock) |
416 | { | |
70b97a7f | 417 | struct rq *rq; |
b29739f9 IM |
418 | |
419 | repeat_lock_task: | |
420 | rq = task_rq(p); | |
421 | spin_lock(&rq->lock); | |
422 | if (unlikely(rq != task_rq(p))) { | |
423 | spin_unlock(&rq->lock); | |
424 | goto repeat_lock_task; | |
425 | } | |
426 | return rq; | |
427 | } | |
428 | ||
1da177e4 LT |
429 | /* |
430 | * task_rq_lock - lock the runqueue a given task resides on and disable | |
431 | * interrupts. Note the ordering: we can safely lookup the task_rq without | |
432 | * explicitly disabling preemption. | |
433 | */ | |
70b97a7f | 434 | static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) |
1da177e4 LT |
435 | __acquires(rq->lock) |
436 | { | |
70b97a7f | 437 | struct rq *rq; |
1da177e4 LT |
438 | |
439 | repeat_lock_task: | |
440 | local_irq_save(*flags); | |
441 | rq = task_rq(p); | |
442 | spin_lock(&rq->lock); | |
443 | if (unlikely(rq != task_rq(p))) { | |
444 | spin_unlock_irqrestore(&rq->lock, *flags); | |
445 | goto repeat_lock_task; | |
446 | } | |
447 | return rq; | |
448 | } | |
449 | ||
70b97a7f | 450 | static inline void __task_rq_unlock(struct rq *rq) |
b29739f9 IM |
451 | __releases(rq->lock) |
452 | { | |
453 | spin_unlock(&rq->lock); | |
454 | } | |
455 | ||
70b97a7f | 456 | static inline void task_rq_unlock(struct rq *rq, unsigned long *flags) |
1da177e4 LT |
457 | __releases(rq->lock) |
458 | { | |
459 | spin_unlock_irqrestore(&rq->lock, *flags); | |
460 | } | |
461 | ||
462 | #ifdef CONFIG_SCHEDSTATS | |
463 | /* | |
464 | * bump this up when changing the output format or the meaning of an existing | |
465 | * format, so that tools can adapt (or abort) | |
466 | */ | |
06066714 | 467 | #define SCHEDSTAT_VERSION 14 |
1da177e4 LT |
468 | |
469 | static int show_schedstat(struct seq_file *seq, void *v) | |
470 | { | |
471 | int cpu; | |
472 | ||
473 | seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION); | |
474 | seq_printf(seq, "timestamp %lu\n", jiffies); | |
475 | for_each_online_cpu(cpu) { | |
70b97a7f | 476 | struct rq *rq = cpu_rq(cpu); |
1da177e4 LT |
477 | #ifdef CONFIG_SMP |
478 | struct sched_domain *sd; | |
479 | int dcnt = 0; | |
480 | #endif | |
481 | ||
482 | /* runqueue-specific stats */ | |
483 | seq_printf(seq, | |
484 | "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu", | |
485 | cpu, rq->yld_both_empty, | |
486 | rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt, | |
487 | rq->sched_switch, rq->sched_cnt, rq->sched_goidle, | |
488 | rq->ttwu_cnt, rq->ttwu_local, | |
489 | rq->rq_sched_info.cpu_time, | |
490 | rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt); | |
491 | ||
492 | seq_printf(seq, "\n"); | |
493 | ||
494 | #ifdef CONFIG_SMP | |
495 | /* domain-specific stats */ | |
674311d5 | 496 | preempt_disable(); |
1da177e4 LT |
497 | for_each_domain(cpu, sd) { |
498 | enum idle_type itype; | |
499 | char mask_str[NR_CPUS]; | |
500 | ||
501 | cpumask_scnprintf(mask_str, NR_CPUS, sd->span); | |
502 | seq_printf(seq, "domain%d %s", dcnt++, mask_str); | |
503 | for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES; | |
504 | itype++) { | |
33859f7f MOS |
505 | seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu " |
506 | "%lu", | |
1da177e4 LT |
507 | sd->lb_cnt[itype], |
508 | sd->lb_balanced[itype], | |
509 | sd->lb_failed[itype], | |
510 | sd->lb_imbalance[itype], | |
511 | sd->lb_gained[itype], | |
512 | sd->lb_hot_gained[itype], | |
513 | sd->lb_nobusyq[itype], | |
06066714 | 514 | sd->lb_nobusyg[itype]); |
1da177e4 | 515 | } |
33859f7f MOS |
516 | seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu" |
517 | " %lu %lu %lu\n", | |
1da177e4 | 518 | sd->alb_cnt, sd->alb_failed, sd->alb_pushed, |
68767a0a NP |
519 | sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed, |
520 | sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed, | |
33859f7f MOS |
521 | sd->ttwu_wake_remote, sd->ttwu_move_affine, |
522 | sd->ttwu_move_balance); | |
1da177e4 | 523 | } |
674311d5 | 524 | preempt_enable(); |
1da177e4 LT |
525 | #endif |
526 | } | |
527 | return 0; | |
528 | } | |
529 | ||
530 | static int schedstat_open(struct inode *inode, struct file *file) | |
531 | { | |
532 | unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32); | |
533 | char *buf = kmalloc(size, GFP_KERNEL); | |
534 | struct seq_file *m; | |
535 | int res; | |
536 | ||
537 | if (!buf) | |
538 | return -ENOMEM; | |
539 | res = single_open(file, show_schedstat, NULL); | |
540 | if (!res) { | |
541 | m = file->private_data; | |
542 | m->buf = buf; | |
543 | m->size = size; | |
544 | } else | |
545 | kfree(buf); | |
546 | return res; | |
547 | } | |
548 | ||
15ad7cdc | 549 | const struct file_operations proc_schedstat_operations = { |
1da177e4 LT |
550 | .open = schedstat_open, |
551 | .read = seq_read, | |
552 | .llseek = seq_lseek, | |
553 | .release = single_release, | |
554 | }; | |
555 | ||
52f17b6c CS |
556 | /* |
557 | * Expects runqueue lock to be held for atomicity of update | |
558 | */ | |
559 | static inline void | |
560 | rq_sched_info_arrive(struct rq *rq, unsigned long delta_jiffies) | |
561 | { | |
562 | if (rq) { | |
563 | rq->rq_sched_info.run_delay += delta_jiffies; | |
564 | rq->rq_sched_info.pcnt++; | |
565 | } | |
566 | } | |
567 | ||
568 | /* | |
569 | * Expects runqueue lock to be held for atomicity of update | |
570 | */ | |
571 | static inline void | |
572 | rq_sched_info_depart(struct rq *rq, unsigned long delta_jiffies) | |
573 | { | |
574 | if (rq) | |
575 | rq->rq_sched_info.cpu_time += delta_jiffies; | |
576 | } | |
1da177e4 LT |
577 | # define schedstat_inc(rq, field) do { (rq)->field++; } while (0) |
578 | # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0) | |
579 | #else /* !CONFIG_SCHEDSTATS */ | |
52f17b6c CS |
580 | static inline void |
581 | rq_sched_info_arrive(struct rq *rq, unsigned long delta_jiffies) | |
582 | {} | |
583 | static inline void | |
584 | rq_sched_info_depart(struct rq *rq, unsigned long delta_jiffies) | |
585 | {} | |
1da177e4 LT |
586 | # define schedstat_inc(rq, field) do { } while (0) |
587 | # define schedstat_add(rq, field, amt) do { } while (0) | |
588 | #endif | |
589 | ||
590 | /* | |
cc2a73b5 | 591 | * this_rq_lock - lock this runqueue and disable interrupts. |
1da177e4 | 592 | */ |
70b97a7f | 593 | static inline struct rq *this_rq_lock(void) |
1da177e4 LT |
594 | __acquires(rq->lock) |
595 | { | |
70b97a7f | 596 | struct rq *rq; |
1da177e4 LT |
597 | |
598 | local_irq_disable(); | |
599 | rq = this_rq(); | |
600 | spin_lock(&rq->lock); | |
601 | ||
602 | return rq; | |
603 | } | |
604 | ||
52f17b6c | 605 | #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) |
1da177e4 LT |
606 | /* |
607 | * Called when a process is dequeued from the active array and given | |
608 | * the cpu. We should note that with the exception of interactive | |
609 | * tasks, the expired queue will become the active queue after the active | |
610 | * queue is empty, without explicitly dequeuing and requeuing tasks in the | |
611 | * expired queue. (Interactive tasks may be requeued directly to the | |
612 | * active queue, thus delaying tasks in the expired queue from running; | |
613 | * see scheduler_tick()). | |
614 | * | |
615 | * This function is only called from sched_info_arrive(), rather than | |
616 | * dequeue_task(). Even though a task may be queued and dequeued multiple | |
617 | * times as it is shuffled about, we're really interested in knowing how | |
618 | * long it was from the *first* time it was queued to the time that it | |
619 | * finally hit a cpu. | |
620 | */ | |
36c8b586 | 621 | static inline void sched_info_dequeued(struct task_struct *t) |
1da177e4 LT |
622 | { |
623 | t->sched_info.last_queued = 0; | |
624 | } | |
625 | ||
626 | /* | |
627 | * Called when a task finally hits the cpu. We can now calculate how | |
628 | * long it was waiting to run. We also note when it began so that we | |
629 | * can keep stats on how long its timeslice is. | |
630 | */ | |
36c8b586 | 631 | static void sched_info_arrive(struct task_struct *t) |
1da177e4 | 632 | { |
52f17b6c | 633 | unsigned long now = jiffies, delta_jiffies = 0; |
1da177e4 LT |
634 | |
635 | if (t->sched_info.last_queued) | |
52f17b6c | 636 | delta_jiffies = now - t->sched_info.last_queued; |
1da177e4 | 637 | sched_info_dequeued(t); |
52f17b6c | 638 | t->sched_info.run_delay += delta_jiffies; |
1da177e4 LT |
639 | t->sched_info.last_arrival = now; |
640 | t->sched_info.pcnt++; | |
641 | ||
52f17b6c | 642 | rq_sched_info_arrive(task_rq(t), delta_jiffies); |
1da177e4 LT |
643 | } |
644 | ||
645 | /* | |
646 | * Called when a process is queued into either the active or expired | |
647 | * array. The time is noted and later used to determine how long we | |
648 | * had to wait for us to reach the cpu. Since the expired queue will | |
649 | * become the active queue after active queue is empty, without dequeuing | |
650 | * and requeuing any tasks, we are interested in queuing to either. It | |
651 | * is unusual but not impossible for tasks to be dequeued and immediately | |
652 | * requeued in the same or another array: this can happen in sched_yield(), | |
653 | * set_user_nice(), and even load_balance() as it moves tasks from runqueue | |
654 | * to runqueue. | |
655 | * | |
656 | * This function is only called from enqueue_task(), but also only updates | |
657 | * the timestamp if it is already not set. It's assumed that | |
658 | * sched_info_dequeued() will clear that stamp when appropriate. | |
659 | */ | |
36c8b586 | 660 | static inline void sched_info_queued(struct task_struct *t) |
1da177e4 | 661 | { |
52f17b6c CS |
662 | if (unlikely(sched_info_on())) |
663 | if (!t->sched_info.last_queued) | |
664 | t->sched_info.last_queued = jiffies; | |
1da177e4 LT |
665 | } |
666 | ||
667 | /* | |
668 | * Called when a process ceases being the active-running process, either | |
669 | * voluntarily or involuntarily. Now we can calculate how long we ran. | |
670 | */ | |
36c8b586 | 671 | static inline void sched_info_depart(struct task_struct *t) |
1da177e4 | 672 | { |
52f17b6c | 673 | unsigned long delta_jiffies = jiffies - t->sched_info.last_arrival; |
1da177e4 | 674 | |
52f17b6c CS |
675 | t->sched_info.cpu_time += delta_jiffies; |
676 | rq_sched_info_depart(task_rq(t), delta_jiffies); | |
1da177e4 LT |
677 | } |
678 | ||
679 | /* | |
680 | * Called when tasks are switched involuntarily due, typically, to expiring | |
681 | * their time slice. (This may also be called when switching to or from | |
682 | * the idle task.) We are only called when prev != next. | |
683 | */ | |
36c8b586 | 684 | static inline void |
52f17b6c | 685 | __sched_info_switch(struct task_struct *prev, struct task_struct *next) |
1da177e4 | 686 | { |
70b97a7f | 687 | struct rq *rq = task_rq(prev); |
1da177e4 LT |
688 | |
689 | /* | |
690 | * prev now departs the cpu. It's not interesting to record | |
691 | * stats about how efficient we were at scheduling the idle | |
692 | * process, however. | |
693 | */ | |
694 | if (prev != rq->idle) | |
695 | sched_info_depart(prev); | |
696 | ||
697 | if (next != rq->idle) | |
698 | sched_info_arrive(next); | |
699 | } | |
52f17b6c CS |
700 | static inline void |
701 | sched_info_switch(struct task_struct *prev, struct task_struct *next) | |
702 | { | |
703 | if (unlikely(sched_info_on())) | |
704 | __sched_info_switch(prev, next); | |
705 | } | |
1da177e4 LT |
706 | #else |
707 | #define sched_info_queued(t) do { } while (0) | |
708 | #define sched_info_switch(t, next) do { } while (0) | |
52f17b6c | 709 | #endif /* CONFIG_SCHEDSTATS || CONFIG_TASK_DELAY_ACCT */ |
1da177e4 LT |
710 | |
711 | /* | |
712 | * Adding/removing a task to/from a priority array: | |
713 | */ | |
70b97a7f | 714 | static void dequeue_task(struct task_struct *p, struct prio_array *array) |
1da177e4 LT |
715 | { |
716 | array->nr_active--; | |
717 | list_del(&p->run_list); | |
718 | if (list_empty(array->queue + p->prio)) | |
719 | __clear_bit(p->prio, array->bitmap); | |
720 | } | |
721 | ||
70b97a7f | 722 | static void enqueue_task(struct task_struct *p, struct prio_array *array) |
1da177e4 LT |
723 | { |
724 | sched_info_queued(p); | |
725 | list_add_tail(&p->run_list, array->queue + p->prio); | |
726 | __set_bit(p->prio, array->bitmap); | |
727 | array->nr_active++; | |
728 | p->array = array; | |
729 | } | |
730 | ||
731 | /* | |
732 | * Put task to the end of the run list without the overhead of dequeue | |
733 | * followed by enqueue. | |
734 | */ | |
70b97a7f | 735 | static void requeue_task(struct task_struct *p, struct prio_array *array) |
1da177e4 LT |
736 | { |
737 | list_move_tail(&p->run_list, array->queue + p->prio); | |
738 | } | |
739 | ||
70b97a7f IM |
740 | static inline void |
741 | enqueue_task_head(struct task_struct *p, struct prio_array *array) | |
1da177e4 LT |
742 | { |
743 | list_add(&p->run_list, array->queue + p->prio); | |
744 | __set_bit(p->prio, array->bitmap); | |
745 | array->nr_active++; | |
746 | p->array = array; | |
747 | } | |
748 | ||
749 | /* | |
b29739f9 | 750 | * __normal_prio - return the priority that is based on the static |
1da177e4 LT |
751 | * priority but is modified by bonuses/penalties. |
752 | * | |
753 | * We scale the actual sleep average [0 .... MAX_SLEEP_AVG] | |
754 | * into the -5 ... 0 ... +5 bonus/penalty range. | |
755 | * | |
756 | * We use 25% of the full 0...39 priority range so that: | |
757 | * | |
758 | * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs. | |
759 | * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks. | |
760 | * | |
761 | * Both properties are important to certain workloads. | |
762 | */ | |
b29739f9 | 763 | |
36c8b586 | 764 | static inline int __normal_prio(struct task_struct *p) |
1da177e4 LT |
765 | { |
766 | int bonus, prio; | |
767 | ||
1da177e4 LT |
768 | bonus = CURRENT_BONUS(p) - MAX_BONUS / 2; |
769 | ||
770 | prio = p->static_prio - bonus; | |
771 | if (prio < MAX_RT_PRIO) | |
772 | prio = MAX_RT_PRIO; | |
773 | if (prio > MAX_PRIO-1) | |
774 | prio = MAX_PRIO-1; | |
775 | return prio; | |
776 | } | |
777 | ||
2dd73a4f PW |
778 | /* |
779 | * To aid in avoiding the subversion of "niceness" due to uneven distribution | |
780 | * of tasks with abnormal "nice" values across CPUs the contribution that | |
781 | * each task makes to its run queue's load is weighted according to its | |
782 | * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a | |
783 | * scaled version of the new time slice allocation that they receive on time | |
784 | * slice expiry etc. | |
785 | */ | |
786 | ||
787 | /* | |
788 | * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE | |
789 | * If static_prio_timeslice() is ever changed to break this assumption then | |
790 | * this code will need modification | |
791 | */ | |
792 | #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE | |
793 | #define LOAD_WEIGHT(lp) \ | |
794 | (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO) | |
795 | #define PRIO_TO_LOAD_WEIGHT(prio) \ | |
796 | LOAD_WEIGHT(static_prio_timeslice(prio)) | |
797 | #define RTPRIO_TO_LOAD_WEIGHT(rp) \ | |
798 | (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp)) | |
799 | ||
36c8b586 | 800 | static void set_load_weight(struct task_struct *p) |
2dd73a4f | 801 | { |
b29739f9 | 802 | if (has_rt_policy(p)) { |
2dd73a4f PW |
803 | #ifdef CONFIG_SMP |
804 | if (p == task_rq(p)->migration_thread) | |
805 | /* | |
806 | * The migration thread does the actual balancing. | |
807 | * Giving its load any weight will skew balancing | |
808 | * adversely. | |
809 | */ | |
810 | p->load_weight = 0; | |
811 | else | |
812 | #endif | |
813 | p->load_weight = RTPRIO_TO_LOAD_WEIGHT(p->rt_priority); | |
814 | } else | |
815 | p->load_weight = PRIO_TO_LOAD_WEIGHT(p->static_prio); | |
816 | } | |
817 | ||
36c8b586 | 818 | static inline void |
70b97a7f | 819 | inc_raw_weighted_load(struct rq *rq, const struct task_struct *p) |
2dd73a4f PW |
820 | { |
821 | rq->raw_weighted_load += p->load_weight; | |
822 | } | |
823 | ||
36c8b586 | 824 | static inline void |
70b97a7f | 825 | dec_raw_weighted_load(struct rq *rq, const struct task_struct *p) |
2dd73a4f PW |
826 | { |
827 | rq->raw_weighted_load -= p->load_weight; | |
828 | } | |
829 | ||
70b97a7f | 830 | static inline void inc_nr_running(struct task_struct *p, struct rq *rq) |
2dd73a4f PW |
831 | { |
832 | rq->nr_running++; | |
833 | inc_raw_weighted_load(rq, p); | |
834 | } | |
835 | ||
70b97a7f | 836 | static inline void dec_nr_running(struct task_struct *p, struct rq *rq) |
2dd73a4f PW |
837 | { |
838 | rq->nr_running--; | |
839 | dec_raw_weighted_load(rq, p); | |
840 | } | |
841 | ||
b29739f9 IM |
842 | /* |
843 | * Calculate the expected normal priority: i.e. priority | |
844 | * without taking RT-inheritance into account. Might be | |
845 | * boosted by interactivity modifiers. Changes upon fork, | |
846 | * setprio syscalls, and whenever the interactivity | |
847 | * estimator recalculates. | |
848 | */ | |
36c8b586 | 849 | static inline int normal_prio(struct task_struct *p) |
b29739f9 IM |
850 | { |
851 | int prio; | |
852 | ||
853 | if (has_rt_policy(p)) | |
854 | prio = MAX_RT_PRIO-1 - p->rt_priority; | |
855 | else | |
856 | prio = __normal_prio(p); | |
857 | return prio; | |
858 | } | |
859 | ||
860 | /* | |
861 | * Calculate the current priority, i.e. the priority | |
862 | * taken into account by the scheduler. This value might | |
863 | * be boosted by RT tasks, or might be boosted by | |
864 | * interactivity modifiers. Will be RT if the task got | |
865 | * RT-boosted. If not then it returns p->normal_prio. | |
866 | */ | |
36c8b586 | 867 | static int effective_prio(struct task_struct *p) |
b29739f9 IM |
868 | { |
869 | p->normal_prio = normal_prio(p); | |
870 | /* | |
871 | * If we are RT tasks or we were boosted to RT priority, | |
872 | * keep the priority unchanged. Otherwise, update priority | |
873 | * to the normal priority: | |
874 | */ | |
875 | if (!rt_prio(p->prio)) | |
876 | return p->normal_prio; | |
877 | return p->prio; | |
878 | } | |
879 | ||
1da177e4 LT |
880 | /* |
881 | * __activate_task - move a task to the runqueue. | |
882 | */ | |
70b97a7f | 883 | static void __activate_task(struct task_struct *p, struct rq *rq) |
1da177e4 | 884 | { |
70b97a7f | 885 | struct prio_array *target = rq->active; |
d425b274 | 886 | |
f1adad78 | 887 | if (batch_task(p)) |
d425b274 CK |
888 | target = rq->expired; |
889 | enqueue_task(p, target); | |
2dd73a4f | 890 | inc_nr_running(p, rq); |
1da177e4 LT |
891 | } |
892 | ||
893 | /* | |
894 | * __activate_idle_task - move idle task to the _front_ of runqueue. | |
895 | */ | |
70b97a7f | 896 | static inline void __activate_idle_task(struct task_struct *p, struct rq *rq) |
1da177e4 LT |
897 | { |
898 | enqueue_task_head(p, rq->active); | |
2dd73a4f | 899 | inc_nr_running(p, rq); |
1da177e4 LT |
900 | } |
901 | ||
b29739f9 IM |
902 | /* |
903 | * Recalculate p->normal_prio and p->prio after having slept, | |
904 | * updating the sleep-average too: | |
905 | */ | |
36c8b586 | 906 | static int recalc_task_prio(struct task_struct *p, unsigned long long now) |
1da177e4 LT |
907 | { |
908 | /* Caller must always ensure 'now >= p->timestamp' */ | |
72d2854d | 909 | unsigned long sleep_time = now - p->timestamp; |
1da177e4 | 910 | |
d425b274 | 911 | if (batch_task(p)) |
b0a9499c | 912 | sleep_time = 0; |
1da177e4 LT |
913 | |
914 | if (likely(sleep_time > 0)) { | |
915 | /* | |
72d2854d CK |
916 | * This ceiling is set to the lowest priority that would allow |
917 | * a task to be reinserted into the active array on timeslice | |
918 | * completion. | |
1da177e4 | 919 | */ |
72d2854d | 920 | unsigned long ceiling = INTERACTIVE_SLEEP(p); |
e72ff0bb | 921 | |
72d2854d CK |
922 | if (p->mm && sleep_time > ceiling && p->sleep_avg < ceiling) { |
923 | /* | |
924 | * Prevents user tasks from achieving best priority | |
925 | * with one single large enough sleep. | |
926 | */ | |
927 | p->sleep_avg = ceiling; | |
928 | /* | |
929 | * Using INTERACTIVE_SLEEP() as a ceiling places a | |
930 | * nice(0) task 1ms sleep away from promotion, and | |
931 | * gives it 700ms to round-robin with no chance of | |
932 | * being demoted. This is more than generous, so | |
933 | * mark this sleep as non-interactive to prevent the | |
934 | * on-runqueue bonus logic from intervening should | |
935 | * this task not receive cpu immediately. | |
936 | */ | |
937 | p->sleep_type = SLEEP_NONINTERACTIVE; | |
1da177e4 | 938 | } else { |
1da177e4 LT |
939 | /* |
940 | * Tasks waking from uninterruptible sleep are | |
941 | * limited in their sleep_avg rise as they | |
942 | * are likely to be waiting on I/O | |
943 | */ | |
3dee386e | 944 | if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) { |
72d2854d | 945 | if (p->sleep_avg >= ceiling) |
1da177e4 LT |
946 | sleep_time = 0; |
947 | else if (p->sleep_avg + sleep_time >= | |
72d2854d CK |
948 | ceiling) { |
949 | p->sleep_avg = ceiling; | |
950 | sleep_time = 0; | |
1da177e4 LT |
951 | } |
952 | } | |
953 | ||
954 | /* | |
955 | * This code gives a bonus to interactive tasks. | |
956 | * | |
957 | * The boost works by updating the 'average sleep time' | |
958 | * value here, based on ->timestamp. The more time a | |
959 | * task spends sleeping, the higher the average gets - | |
960 | * and the higher the priority boost gets as well. | |
961 | */ | |
962 | p->sleep_avg += sleep_time; | |
963 | ||
1da177e4 | 964 | } |
72d2854d CK |
965 | if (p->sleep_avg > NS_MAX_SLEEP_AVG) |
966 | p->sleep_avg = NS_MAX_SLEEP_AVG; | |
1da177e4 LT |
967 | } |
968 | ||
a3464a10 | 969 | return effective_prio(p); |
1da177e4 LT |
970 | } |
971 | ||
972 | /* | |
973 | * activate_task - move a task to the runqueue and do priority recalculation | |
974 | * | |
975 | * Update all the scheduling statistics stuff. (sleep average | |
976 | * calculation, priority modifiers, etc.) | |
977 | */ | |
70b97a7f | 978 | static void activate_task(struct task_struct *p, struct rq *rq, int local) |
1da177e4 LT |
979 | { |
980 | unsigned long long now; | |
981 | ||
62ab616d KC |
982 | if (rt_task(p)) |
983 | goto out; | |
984 | ||
1da177e4 LT |
985 | now = sched_clock(); |
986 | #ifdef CONFIG_SMP | |
987 | if (!local) { | |
988 | /* Compensate for drifting sched_clock */ | |
70b97a7f | 989 | struct rq *this_rq = this_rq(); |
b18ec803 MG |
990 | now = (now - this_rq->most_recent_timestamp) |
991 | + rq->most_recent_timestamp; | |
1da177e4 LT |
992 | } |
993 | #endif | |
994 | ||
ece8a684 IM |
995 | /* |
996 | * Sleep time is in units of nanosecs, so shift by 20 to get a | |
997 | * milliseconds-range estimation of the amount of time that the task | |
998 | * spent sleeping: | |
999 | */ | |
1000 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
1001 | if (p->state == TASK_UNINTERRUPTIBLE) | |
1002 | profile_hits(SLEEP_PROFILING, (void *)get_wchan(p), | |
1003 | (now - p->timestamp) >> 20); | |
1004 | } | |
1005 | ||
62ab616d | 1006 | p->prio = recalc_task_prio(p, now); |
1da177e4 LT |
1007 | |
1008 | /* | |
1009 | * This checks to make sure it's not an uninterruptible task | |
1010 | * that is now waking up. | |
1011 | */ | |
3dee386e | 1012 | if (p->sleep_type == SLEEP_NORMAL) { |
1da177e4 LT |
1013 | /* |
1014 | * Tasks which were woken up by interrupts (ie. hw events) | |
1015 | * are most likely of interactive nature. So we give them | |
1016 | * the credit of extending their sleep time to the period | |
1017 | * of time they spend on the runqueue, waiting for execution | |
1018 | * on a CPU, first time around: | |
1019 | */ | |
1020 | if (in_interrupt()) | |
3dee386e | 1021 | p->sleep_type = SLEEP_INTERRUPTED; |
1da177e4 LT |
1022 | else { |
1023 | /* | |
1024 | * Normal first-time wakeups get a credit too for | |
1025 | * on-runqueue time, but it will be weighted down: | |
1026 | */ | |
3dee386e | 1027 | p->sleep_type = SLEEP_INTERACTIVE; |
1da177e4 LT |
1028 | } |
1029 | } | |
1030 | p->timestamp = now; | |
62ab616d | 1031 | out: |
1da177e4 LT |
1032 | __activate_task(p, rq); |
1033 | } | |
1034 | ||
1035 | /* | |
1036 | * deactivate_task - remove a task from the runqueue. | |
1037 | */ | |
70b97a7f | 1038 | static void deactivate_task(struct task_struct *p, struct rq *rq) |
1da177e4 | 1039 | { |
2dd73a4f | 1040 | dec_nr_running(p, rq); |
1da177e4 LT |
1041 | dequeue_task(p, p->array); |
1042 | p->array = NULL; | |
1043 | } | |
1044 | ||
1045 | /* | |
1046 | * resched_task - mark a task 'to be rescheduled now'. | |
1047 | * | |
1048 | * On UP this means the setting of the need_resched flag, on SMP it | |
1049 | * might also involve a cross-CPU call to trigger the scheduler on | |
1050 | * the target CPU. | |
1051 | */ | |
1052 | #ifdef CONFIG_SMP | |
495ab9c0 AK |
1053 | |
1054 | #ifndef tsk_is_polling | |
1055 | #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG) | |
1056 | #endif | |
1057 | ||
36c8b586 | 1058 | static void resched_task(struct task_struct *p) |
1da177e4 | 1059 | { |
64c7c8f8 | 1060 | int cpu; |
1da177e4 LT |
1061 | |
1062 | assert_spin_locked(&task_rq(p)->lock); | |
1063 | ||
64c7c8f8 NP |
1064 | if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED))) |
1065 | return; | |
1066 | ||
1067 | set_tsk_thread_flag(p, TIF_NEED_RESCHED); | |
1da177e4 | 1068 | |
64c7c8f8 NP |
1069 | cpu = task_cpu(p); |
1070 | if (cpu == smp_processor_id()) | |
1071 | return; | |
1072 | ||
495ab9c0 | 1073 | /* NEED_RESCHED must be visible before we test polling */ |
64c7c8f8 | 1074 | smp_mb(); |
495ab9c0 | 1075 | if (!tsk_is_polling(p)) |
64c7c8f8 | 1076 | smp_send_reschedule(cpu); |
1da177e4 | 1077 | } |
46cb4b7c SS |
1078 | |
1079 | static void resched_cpu(int cpu) | |
1080 | { | |
1081 | struct rq *rq = cpu_rq(cpu); | |
1082 | unsigned long flags; | |
1083 | ||
1084 | if (!spin_trylock_irqsave(&rq->lock, flags)) | |
1085 | return; | |
1086 | resched_task(cpu_curr(cpu)); | |
1087 | spin_unlock_irqrestore(&rq->lock, flags); | |
1088 | } | |
1da177e4 | 1089 | #else |
36c8b586 | 1090 | static inline void resched_task(struct task_struct *p) |
1da177e4 | 1091 | { |
64c7c8f8 | 1092 | assert_spin_locked(&task_rq(p)->lock); |
1da177e4 LT |
1093 | set_tsk_need_resched(p); |
1094 | } | |
1095 | #endif | |
1096 | ||
1097 | /** | |
1098 | * task_curr - is this task currently executing on a CPU? | |
1099 | * @p: the task in question. | |
1100 | */ | |
36c8b586 | 1101 | inline int task_curr(const struct task_struct *p) |
1da177e4 LT |
1102 | { |
1103 | return cpu_curr(task_cpu(p)) == p; | |
1104 | } | |
1105 | ||
2dd73a4f PW |
1106 | /* Used instead of source_load when we know the type == 0 */ |
1107 | unsigned long weighted_cpuload(const int cpu) | |
1108 | { | |
1109 | return cpu_rq(cpu)->raw_weighted_load; | |
1110 | } | |
1111 | ||
1da177e4 | 1112 | #ifdef CONFIG_SMP |
70b97a7f | 1113 | struct migration_req { |
1da177e4 | 1114 | struct list_head list; |
1da177e4 | 1115 | |
36c8b586 | 1116 | struct task_struct *task; |
1da177e4 LT |
1117 | int dest_cpu; |
1118 | ||
1da177e4 | 1119 | struct completion done; |
70b97a7f | 1120 | }; |
1da177e4 LT |
1121 | |
1122 | /* | |
1123 | * The task's runqueue lock must be held. | |
1124 | * Returns true if you have to wait for migration thread. | |
1125 | */ | |
36c8b586 | 1126 | static int |
70b97a7f | 1127 | migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req) |
1da177e4 | 1128 | { |
70b97a7f | 1129 | struct rq *rq = task_rq(p); |
1da177e4 LT |
1130 | |
1131 | /* | |
1132 | * If the task is not on a runqueue (and not running), then | |
1133 | * it is sufficient to simply update the task's cpu field. | |
1134 | */ | |
1135 | if (!p->array && !task_running(rq, p)) { | |
1136 | set_task_cpu(p, dest_cpu); | |
1137 | return 0; | |
1138 | } | |
1139 | ||
1140 | init_completion(&req->done); | |
1da177e4 LT |
1141 | req->task = p; |
1142 | req->dest_cpu = dest_cpu; | |
1143 | list_add(&req->list, &rq->migration_queue); | |
48f24c4d | 1144 | |
1da177e4 LT |
1145 | return 1; |
1146 | } | |
1147 | ||
1148 | /* | |
1149 | * wait_task_inactive - wait for a thread to unschedule. | |
1150 | * | |
1151 | * The caller must ensure that the task *will* unschedule sometime soon, | |
1152 | * else this function might spin for a *long* time. This function can't | |
1153 | * be called with interrupts off, or it may introduce deadlock with | |
1154 | * smp_call_function() if an IPI is sent by the same process we are | |
1155 | * waiting to become inactive. | |
1156 | */ | |
36c8b586 | 1157 | void wait_task_inactive(struct task_struct *p) |
1da177e4 LT |
1158 | { |
1159 | unsigned long flags; | |
70b97a7f | 1160 | struct rq *rq; |
1da177e4 LT |
1161 | int preempted; |
1162 | ||
1163 | repeat: | |
1164 | rq = task_rq_lock(p, &flags); | |
1165 | /* Must be off runqueue entirely, not preempted. */ | |
1166 | if (unlikely(p->array || task_running(rq, p))) { | |
1167 | /* If it's preempted, we yield. It could be a while. */ | |
1168 | preempted = !task_running(rq, p); | |
1169 | task_rq_unlock(rq, &flags); | |
1170 | cpu_relax(); | |
1171 | if (preempted) | |
1172 | yield(); | |
1173 | goto repeat; | |
1174 | } | |
1175 | task_rq_unlock(rq, &flags); | |
1176 | } | |
1177 | ||
1178 | /*** | |
1179 | * kick_process - kick a running thread to enter/exit the kernel | |
1180 | * @p: the to-be-kicked thread | |
1181 | * | |
1182 | * Cause a process which is running on another CPU to enter | |
1183 | * kernel-mode, without any delay. (to get signals handled.) | |
1184 | * | |
1185 | * NOTE: this function doesnt have to take the runqueue lock, | |
1186 | * because all it wants to ensure is that the remote task enters | |
1187 | * the kernel. If the IPI races and the task has been migrated | |
1188 | * to another CPU then no harm is done and the purpose has been | |
1189 | * achieved as well. | |
1190 | */ | |
36c8b586 | 1191 | void kick_process(struct task_struct *p) |
1da177e4 LT |
1192 | { |
1193 | int cpu; | |
1194 | ||
1195 | preempt_disable(); | |
1196 | cpu = task_cpu(p); | |
1197 | if ((cpu != smp_processor_id()) && task_curr(p)) | |
1198 | smp_send_reschedule(cpu); | |
1199 | preempt_enable(); | |
1200 | } | |
1201 | ||
1202 | /* | |
2dd73a4f PW |
1203 | * Return a low guess at the load of a migration-source cpu weighted |
1204 | * according to the scheduling class and "nice" value. | |
1da177e4 LT |
1205 | * |
1206 | * We want to under-estimate the load of migration sources, to | |
1207 | * balance conservatively. | |
1208 | */ | |
a2000572 | 1209 | static inline unsigned long source_load(int cpu, int type) |
1da177e4 | 1210 | { |
70b97a7f | 1211 | struct rq *rq = cpu_rq(cpu); |
2dd73a4f | 1212 | |
3b0bd9bc | 1213 | if (type == 0) |
2dd73a4f | 1214 | return rq->raw_weighted_load; |
b910472d | 1215 | |
2dd73a4f | 1216 | return min(rq->cpu_load[type-1], rq->raw_weighted_load); |
1da177e4 LT |
1217 | } |
1218 | ||
1219 | /* | |
2dd73a4f PW |
1220 | * Return a high guess at the load of a migration-target cpu weighted |
1221 | * according to the scheduling class and "nice" value. | |
1da177e4 | 1222 | */ |
a2000572 | 1223 | static inline unsigned long target_load(int cpu, int type) |
1da177e4 | 1224 | { |
70b97a7f | 1225 | struct rq *rq = cpu_rq(cpu); |
2dd73a4f | 1226 | |
7897986b | 1227 | if (type == 0) |
2dd73a4f | 1228 | return rq->raw_weighted_load; |
3b0bd9bc | 1229 | |
2dd73a4f PW |
1230 | return max(rq->cpu_load[type-1], rq->raw_weighted_load); |
1231 | } | |
1232 | ||
1233 | /* | |
1234 | * Return the average load per task on the cpu's run queue | |
1235 | */ | |
1236 | static inline unsigned long cpu_avg_load_per_task(int cpu) | |
1237 | { | |
70b97a7f | 1238 | struct rq *rq = cpu_rq(cpu); |
2dd73a4f PW |
1239 | unsigned long n = rq->nr_running; |
1240 | ||
48f24c4d | 1241 | return n ? rq->raw_weighted_load / n : SCHED_LOAD_SCALE; |
1da177e4 LT |
1242 | } |
1243 | ||
147cbb4b NP |
1244 | /* |
1245 | * find_idlest_group finds and returns the least busy CPU group within the | |
1246 | * domain. | |
1247 | */ | |
1248 | static struct sched_group * | |
1249 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu) | |
1250 | { | |
1251 | struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups; | |
1252 | unsigned long min_load = ULONG_MAX, this_load = 0; | |
1253 | int load_idx = sd->forkexec_idx; | |
1254 | int imbalance = 100 + (sd->imbalance_pct-100)/2; | |
1255 | ||
1256 | do { | |
1257 | unsigned long load, avg_load; | |
1258 | int local_group; | |
1259 | int i; | |
1260 | ||
da5a5522 BD |
1261 | /* Skip over this group if it has no CPUs allowed */ |
1262 | if (!cpus_intersects(group->cpumask, p->cpus_allowed)) | |
1263 | goto nextgroup; | |
1264 | ||
147cbb4b | 1265 | local_group = cpu_isset(this_cpu, group->cpumask); |
147cbb4b NP |
1266 | |
1267 | /* Tally up the load of all CPUs in the group */ | |
1268 | avg_load = 0; | |
1269 | ||
1270 | for_each_cpu_mask(i, group->cpumask) { | |
1271 | /* Bias balancing toward cpus of our domain */ | |
1272 | if (local_group) | |
1273 | load = source_load(i, load_idx); | |
1274 | else | |
1275 | load = target_load(i, load_idx); | |
1276 | ||
1277 | avg_load += load; | |
1278 | } | |
1279 | ||
1280 | /* Adjust by relative CPU power of the group */ | |
5517d86b ED |
1281 | avg_load = sg_div_cpu_power(group, |
1282 | avg_load * SCHED_LOAD_SCALE); | |
147cbb4b NP |
1283 | |
1284 | if (local_group) { | |
1285 | this_load = avg_load; | |
1286 | this = group; | |
1287 | } else if (avg_load < min_load) { | |
1288 | min_load = avg_load; | |
1289 | idlest = group; | |
1290 | } | |
da5a5522 | 1291 | nextgroup: |
147cbb4b NP |
1292 | group = group->next; |
1293 | } while (group != sd->groups); | |
1294 | ||
1295 | if (!idlest || 100*this_load < imbalance*min_load) | |
1296 | return NULL; | |
1297 | return idlest; | |
1298 | } | |
1299 | ||
1300 | /* | |
0feaece9 | 1301 | * find_idlest_cpu - find the idlest cpu among the cpus in group. |
147cbb4b | 1302 | */ |
95cdf3b7 IM |
1303 | static int |
1304 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
147cbb4b | 1305 | { |
da5a5522 | 1306 | cpumask_t tmp; |
147cbb4b NP |
1307 | unsigned long load, min_load = ULONG_MAX; |
1308 | int idlest = -1; | |
1309 | int i; | |
1310 | ||
da5a5522 BD |
1311 | /* Traverse only the allowed CPUs */ |
1312 | cpus_and(tmp, group->cpumask, p->cpus_allowed); | |
1313 | ||
1314 | for_each_cpu_mask(i, tmp) { | |
2dd73a4f | 1315 | load = weighted_cpuload(i); |
147cbb4b NP |
1316 | |
1317 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
1318 | min_load = load; | |
1319 | idlest = i; | |
1320 | } | |
1321 | } | |
1322 | ||
1323 | return idlest; | |
1324 | } | |
1325 | ||
476d139c NP |
1326 | /* |
1327 | * sched_balance_self: balance the current task (running on cpu) in domains | |
1328 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
1329 | * SD_BALANCE_EXEC. | |
1330 | * | |
1331 | * Balance, ie. select the least loaded group. | |
1332 | * | |
1333 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
1334 | * | |
1335 | * preempt must be disabled. | |
1336 | */ | |
1337 | static int sched_balance_self(int cpu, int flag) | |
1338 | { | |
1339 | struct task_struct *t = current; | |
1340 | struct sched_domain *tmp, *sd = NULL; | |
147cbb4b | 1341 | |
c96d145e | 1342 | for_each_domain(cpu, tmp) { |
5c45bf27 SS |
1343 | /* |
1344 | * If power savings logic is enabled for a domain, stop there. | |
1345 | */ | |
1346 | if (tmp->flags & SD_POWERSAVINGS_BALANCE) | |
1347 | break; | |
476d139c NP |
1348 | if (tmp->flags & flag) |
1349 | sd = tmp; | |
c96d145e | 1350 | } |
476d139c NP |
1351 | |
1352 | while (sd) { | |
1353 | cpumask_t span; | |
1354 | struct sched_group *group; | |
1a848870 SS |
1355 | int new_cpu, weight; |
1356 | ||
1357 | if (!(sd->flags & flag)) { | |
1358 | sd = sd->child; | |
1359 | continue; | |
1360 | } | |
476d139c NP |
1361 | |
1362 | span = sd->span; | |
1363 | group = find_idlest_group(sd, t, cpu); | |
1a848870 SS |
1364 | if (!group) { |
1365 | sd = sd->child; | |
1366 | continue; | |
1367 | } | |
476d139c | 1368 | |
da5a5522 | 1369 | new_cpu = find_idlest_cpu(group, t, cpu); |
1a848870 SS |
1370 | if (new_cpu == -1 || new_cpu == cpu) { |
1371 | /* Now try balancing at a lower domain level of cpu */ | |
1372 | sd = sd->child; | |
1373 | continue; | |
1374 | } | |
476d139c | 1375 | |
1a848870 | 1376 | /* Now try balancing at a lower domain level of new_cpu */ |
476d139c | 1377 | cpu = new_cpu; |
476d139c NP |
1378 | sd = NULL; |
1379 | weight = cpus_weight(span); | |
1380 | for_each_domain(cpu, tmp) { | |
1381 | if (weight <= cpus_weight(tmp->span)) | |
1382 | break; | |
1383 | if (tmp->flags & flag) | |
1384 | sd = tmp; | |
1385 | } | |
1386 | /* while loop will break here if sd == NULL */ | |
1387 | } | |
1388 | ||
1389 | return cpu; | |
1390 | } | |
1391 | ||
1392 | #endif /* CONFIG_SMP */ | |
1da177e4 LT |
1393 | |
1394 | /* | |
1395 | * wake_idle() will wake a task on an idle cpu if task->cpu is | |
1396 | * not idle and an idle cpu is available. The span of cpus to | |
1397 | * search starts with cpus closest then further out as needed, | |
1398 | * so we always favor a closer, idle cpu. | |
1399 | * | |
1400 | * Returns the CPU we should wake onto. | |
1401 | */ | |
1402 | #if defined(ARCH_HAS_SCHED_WAKE_IDLE) | |
36c8b586 | 1403 | static int wake_idle(int cpu, struct task_struct *p) |
1da177e4 LT |
1404 | { |
1405 | cpumask_t tmp; | |
1406 | struct sched_domain *sd; | |
1407 | int i; | |
1408 | ||
1409 | if (idle_cpu(cpu)) | |
1410 | return cpu; | |
1411 | ||
1412 | for_each_domain(cpu, sd) { | |
1413 | if (sd->flags & SD_WAKE_IDLE) { | |
e0f364f4 | 1414 | cpus_and(tmp, sd->span, p->cpus_allowed); |
1da177e4 LT |
1415 | for_each_cpu_mask(i, tmp) { |
1416 | if (idle_cpu(i)) | |
1417 | return i; | |
1418 | } | |
1419 | } | |
e0f364f4 NP |
1420 | else |
1421 | break; | |
1da177e4 LT |
1422 | } |
1423 | return cpu; | |
1424 | } | |
1425 | #else | |
36c8b586 | 1426 | static inline int wake_idle(int cpu, struct task_struct *p) |
1da177e4 LT |
1427 | { |
1428 | return cpu; | |
1429 | } | |
1430 | #endif | |
1431 | ||
1432 | /*** | |
1433 | * try_to_wake_up - wake up a thread | |
1434 | * @p: the to-be-woken-up thread | |
1435 | * @state: the mask of task states that can be woken | |
1436 | * @sync: do a synchronous wakeup? | |
1437 | * | |
1438 | * Put it on the run-queue if it's not already there. The "current" | |
1439 | * thread is always on the run-queue (except when the actual | |
1440 | * re-schedule is in progress), and as such you're allowed to do | |
1441 | * the simpler "current->state = TASK_RUNNING" to mark yourself | |
1442 | * runnable without the overhead of this. | |
1443 | * | |
1444 | * returns failure only if the task is already active. | |
1445 | */ | |
36c8b586 | 1446 | static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync) |
1da177e4 LT |
1447 | { |
1448 | int cpu, this_cpu, success = 0; | |
1449 | unsigned long flags; | |
1450 | long old_state; | |
70b97a7f | 1451 | struct rq *rq; |
1da177e4 | 1452 | #ifdef CONFIG_SMP |
7897986b | 1453 | struct sched_domain *sd, *this_sd = NULL; |
70b97a7f | 1454 | unsigned long load, this_load; |
1da177e4 LT |
1455 | int new_cpu; |
1456 | #endif | |
1457 | ||
1458 | rq = task_rq_lock(p, &flags); | |
1459 | old_state = p->state; | |
1460 | if (!(old_state & state)) | |
1461 | goto out; | |
1462 | ||
1463 | if (p->array) | |
1464 | goto out_running; | |
1465 | ||
1466 | cpu = task_cpu(p); | |
1467 | this_cpu = smp_processor_id(); | |
1468 | ||
1469 | #ifdef CONFIG_SMP | |
1470 | if (unlikely(task_running(rq, p))) | |
1471 | goto out_activate; | |
1472 | ||
7897986b NP |
1473 | new_cpu = cpu; |
1474 | ||
1da177e4 LT |
1475 | schedstat_inc(rq, ttwu_cnt); |
1476 | if (cpu == this_cpu) { | |
1477 | schedstat_inc(rq, ttwu_local); | |
7897986b NP |
1478 | goto out_set_cpu; |
1479 | } | |
1480 | ||
1481 | for_each_domain(this_cpu, sd) { | |
1482 | if (cpu_isset(cpu, sd->span)) { | |
1483 | schedstat_inc(sd, ttwu_wake_remote); | |
1484 | this_sd = sd; | |
1485 | break; | |
1da177e4 LT |
1486 | } |
1487 | } | |
1da177e4 | 1488 | |
7897986b | 1489 | if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed))) |
1da177e4 LT |
1490 | goto out_set_cpu; |
1491 | ||
1da177e4 | 1492 | /* |
7897986b | 1493 | * Check for affine wakeup and passive balancing possibilities. |
1da177e4 | 1494 | */ |
7897986b NP |
1495 | if (this_sd) { |
1496 | int idx = this_sd->wake_idx; | |
1497 | unsigned int imbalance; | |
1da177e4 | 1498 | |
a3f21bce NP |
1499 | imbalance = 100 + (this_sd->imbalance_pct - 100) / 2; |
1500 | ||
7897986b NP |
1501 | load = source_load(cpu, idx); |
1502 | this_load = target_load(this_cpu, idx); | |
1da177e4 | 1503 | |
7897986b NP |
1504 | new_cpu = this_cpu; /* Wake to this CPU if we can */ |
1505 | ||
a3f21bce NP |
1506 | if (this_sd->flags & SD_WAKE_AFFINE) { |
1507 | unsigned long tl = this_load; | |
33859f7f MOS |
1508 | unsigned long tl_per_task; |
1509 | ||
1510 | tl_per_task = cpu_avg_load_per_task(this_cpu); | |
2dd73a4f | 1511 | |
1da177e4 | 1512 | /* |
a3f21bce NP |
1513 | * If sync wakeup then subtract the (maximum possible) |
1514 | * effect of the currently running task from the load | |
1515 | * of the current CPU: | |
1da177e4 | 1516 | */ |
a3f21bce | 1517 | if (sync) |
2dd73a4f | 1518 | tl -= current->load_weight; |
a3f21bce NP |
1519 | |
1520 | if ((tl <= load && | |
2dd73a4f PW |
1521 | tl + target_load(cpu, idx) <= tl_per_task) || |
1522 | 100*(tl + p->load_weight) <= imbalance*load) { | |
a3f21bce NP |
1523 | /* |
1524 | * This domain has SD_WAKE_AFFINE and | |
1525 | * p is cache cold in this domain, and | |
1526 | * there is no bad imbalance. | |
1527 | */ | |
1528 | schedstat_inc(this_sd, ttwu_move_affine); | |
1529 | goto out_set_cpu; | |
1530 | } | |
1531 | } | |
1532 | ||
1533 | /* | |
1534 | * Start passive balancing when half the imbalance_pct | |
1535 | * limit is reached. | |
1536 | */ | |
1537 | if (this_sd->flags & SD_WAKE_BALANCE) { | |
1538 | if (imbalance*this_load <= 100*load) { | |
1539 | schedstat_inc(this_sd, ttwu_move_balance); | |
1540 | goto out_set_cpu; | |
1541 | } | |
1da177e4 LT |
1542 | } |
1543 | } | |
1544 | ||
1545 | new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */ | |
1546 | out_set_cpu: | |
1547 | new_cpu = wake_idle(new_cpu, p); | |
1548 | if (new_cpu != cpu) { | |
1549 | set_task_cpu(p, new_cpu); | |
1550 | task_rq_unlock(rq, &flags); | |
1551 | /* might preempt at this point */ | |
1552 | rq = task_rq_lock(p, &flags); | |
1553 | old_state = p->state; | |
1554 | if (!(old_state & state)) | |
1555 | goto out; | |
1556 | if (p->array) | |
1557 | goto out_running; | |
1558 | ||
1559 | this_cpu = smp_processor_id(); | |
1560 | cpu = task_cpu(p); | |
1561 | } | |
1562 | ||
1563 | out_activate: | |
1564 | #endif /* CONFIG_SMP */ | |
1565 | if (old_state == TASK_UNINTERRUPTIBLE) { | |
1566 | rq->nr_uninterruptible--; | |
1567 | /* | |
1568 | * Tasks on involuntary sleep don't earn | |
1569 | * sleep_avg beyond just interactive state. | |
1570 | */ | |
3dee386e | 1571 | p->sleep_type = SLEEP_NONINTERACTIVE; |
e7c38cb4 | 1572 | } else |
1da177e4 | 1573 | |
d79fc0fc IM |
1574 | /* |
1575 | * Tasks that have marked their sleep as noninteractive get | |
e7c38cb4 CK |
1576 | * woken up with their sleep average not weighted in an |
1577 | * interactive way. | |
d79fc0fc | 1578 | */ |
e7c38cb4 CK |
1579 | if (old_state & TASK_NONINTERACTIVE) |
1580 | p->sleep_type = SLEEP_NONINTERACTIVE; | |
1581 | ||
1582 | ||
1583 | activate_task(p, rq, cpu == this_cpu); | |
1da177e4 LT |
1584 | /* |
1585 | * Sync wakeups (i.e. those types of wakeups where the waker | |
1586 | * has indicated that it will leave the CPU in short order) | |
1587 | * don't trigger a preemption, if the woken up task will run on | |
1588 | * this cpu. (in this case the 'I will reschedule' promise of | |
1589 | * the waker guarantees that the freshly woken up task is going | |
1590 | * to be considered on this CPU.) | |
1591 | */ | |
1da177e4 LT |
1592 | if (!sync || cpu != this_cpu) { |
1593 | if (TASK_PREEMPTS_CURR(p, rq)) | |
1594 | resched_task(rq->curr); | |
1595 | } | |
1596 | success = 1; | |
1597 | ||
1598 | out_running: | |
1599 | p->state = TASK_RUNNING; | |
1600 | out: | |
1601 | task_rq_unlock(rq, &flags); | |
1602 | ||
1603 | return success; | |
1604 | } | |
1605 | ||
36c8b586 | 1606 | int fastcall wake_up_process(struct task_struct *p) |
1da177e4 LT |
1607 | { |
1608 | return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED | | |
1609 | TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0); | |
1610 | } | |
1da177e4 LT |
1611 | EXPORT_SYMBOL(wake_up_process); |
1612 | ||
36c8b586 | 1613 | int fastcall wake_up_state(struct task_struct *p, unsigned int state) |
1da177e4 LT |
1614 | { |
1615 | return try_to_wake_up(p, state, 0); | |
1616 | } | |
1617 | ||
bc947631 | 1618 | static void task_running_tick(struct rq *rq, struct task_struct *p); |
1da177e4 LT |
1619 | /* |
1620 | * Perform scheduler related setup for a newly forked process p. | |
1621 | * p is forked by current. | |
1622 | */ | |
36c8b586 | 1623 | void fastcall sched_fork(struct task_struct *p, int clone_flags) |
1da177e4 | 1624 | { |
476d139c NP |
1625 | int cpu = get_cpu(); |
1626 | ||
1627 | #ifdef CONFIG_SMP | |
1628 | cpu = sched_balance_self(cpu, SD_BALANCE_FORK); | |
1629 | #endif | |
1630 | set_task_cpu(p, cpu); | |
1631 | ||
1da177e4 LT |
1632 | /* |
1633 | * We mark the process as running here, but have not actually | |
1634 | * inserted it onto the runqueue yet. This guarantees that | |
1635 | * nobody will actually run it, and a signal or other external | |
1636 | * event cannot wake it up and insert it on the runqueue either. | |
1637 | */ | |
1638 | p->state = TASK_RUNNING; | |
b29739f9 IM |
1639 | |
1640 | /* | |
1641 | * Make sure we do not leak PI boosting priority to the child: | |
1642 | */ | |
1643 | p->prio = current->normal_prio; | |
1644 | ||
1da177e4 LT |
1645 | INIT_LIST_HEAD(&p->run_list); |
1646 | p->array = NULL; | |
52f17b6c CS |
1647 | #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) |
1648 | if (unlikely(sched_info_on())) | |
1649 | memset(&p->sched_info, 0, sizeof(p->sched_info)); | |
1da177e4 | 1650 | #endif |
d6077cb8 | 1651 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
4866cde0 NP |
1652 | p->oncpu = 0; |
1653 | #endif | |
1da177e4 | 1654 | #ifdef CONFIG_PREEMPT |
4866cde0 | 1655 | /* Want to start with kernel preemption disabled. */ |
a1261f54 | 1656 | task_thread_info(p)->preempt_count = 1; |
1da177e4 LT |
1657 | #endif |
1658 | /* | |
1659 | * Share the timeslice between parent and child, thus the | |
1660 | * total amount of pending timeslices in the system doesn't change, | |
1661 | * resulting in more scheduling fairness. | |
1662 | */ | |
1663 | local_irq_disable(); | |
1664 | p->time_slice = (current->time_slice + 1) >> 1; | |
1665 | /* | |
1666 | * The remainder of the first timeslice might be recovered by | |
1667 | * the parent if the child exits early enough. | |
1668 | */ | |
1669 | p->first_time_slice = 1; | |
1670 | current->time_slice >>= 1; | |
1671 | p->timestamp = sched_clock(); | |
1672 | if (unlikely(!current->time_slice)) { | |
1673 | /* | |
1674 | * This case is rare, it happens when the parent has only | |
1675 | * a single jiffy left from its timeslice. Taking the | |
1676 | * runqueue lock is not a problem. | |
1677 | */ | |
1678 | current->time_slice = 1; | |
bc947631 | 1679 | task_running_tick(cpu_rq(cpu), current); |
476d139c NP |
1680 | } |
1681 | local_irq_enable(); | |
1682 | put_cpu(); | |
1da177e4 LT |
1683 | } |
1684 | ||
1685 | /* | |
1686 | * wake_up_new_task - wake up a newly created task for the first time. | |
1687 | * | |
1688 | * This function will do some initial scheduler statistics housekeeping | |
1689 | * that must be done for every newly created context, then puts the task | |
1690 | * on the runqueue and wakes it. | |
1691 | */ | |
36c8b586 | 1692 | void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags) |
1da177e4 | 1693 | { |
70b97a7f | 1694 | struct rq *rq, *this_rq; |
1da177e4 LT |
1695 | unsigned long flags; |
1696 | int this_cpu, cpu; | |
1da177e4 LT |
1697 | |
1698 | rq = task_rq_lock(p, &flags); | |
147cbb4b | 1699 | BUG_ON(p->state != TASK_RUNNING); |
1da177e4 | 1700 | this_cpu = smp_processor_id(); |
147cbb4b | 1701 | cpu = task_cpu(p); |
1da177e4 | 1702 | |
1da177e4 LT |
1703 | /* |
1704 | * We decrease the sleep average of forking parents | |
1705 | * and children as well, to keep max-interactive tasks | |
1706 | * from forking tasks that are max-interactive. The parent | |
1707 | * (current) is done further down, under its lock. | |
1708 | */ | |
1709 | p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) * | |
1710 | CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS); | |
1711 | ||
1712 | p->prio = effective_prio(p); | |
1713 | ||
1714 | if (likely(cpu == this_cpu)) { | |
1715 | if (!(clone_flags & CLONE_VM)) { | |
1716 | /* | |
1717 | * The VM isn't cloned, so we're in a good position to | |
1718 | * do child-runs-first in anticipation of an exec. This | |
1719 | * usually avoids a lot of COW overhead. | |
1720 | */ | |
1721 | if (unlikely(!current->array)) | |
1722 | __activate_task(p, rq); | |
1723 | else { | |
1724 | p->prio = current->prio; | |
b29739f9 | 1725 | p->normal_prio = current->normal_prio; |
1da177e4 LT |
1726 | list_add_tail(&p->run_list, ¤t->run_list); |
1727 | p->array = current->array; | |
1728 | p->array->nr_active++; | |
2dd73a4f | 1729 | inc_nr_running(p, rq); |
1da177e4 LT |
1730 | } |
1731 | set_need_resched(); | |
1732 | } else | |
1733 | /* Run child last */ | |
1734 | __activate_task(p, rq); | |
1735 | /* | |
1736 | * We skip the following code due to cpu == this_cpu | |
1737 | * | |
1738 | * task_rq_unlock(rq, &flags); | |
1739 | * this_rq = task_rq_lock(current, &flags); | |
1740 | */ | |
1741 | this_rq = rq; | |
1742 | } else { | |
1743 | this_rq = cpu_rq(this_cpu); | |
1744 | ||
1745 | /* | |
1746 | * Not the local CPU - must adjust timestamp. This should | |
1747 | * get optimised away in the !CONFIG_SMP case. | |
1748 | */ | |
b18ec803 MG |
1749 | p->timestamp = (p->timestamp - this_rq->most_recent_timestamp) |
1750 | + rq->most_recent_timestamp; | |
1da177e4 LT |
1751 | __activate_task(p, rq); |
1752 | if (TASK_PREEMPTS_CURR(p, rq)) | |
1753 | resched_task(rq->curr); | |
1754 | ||
1755 | /* | |
1756 | * Parent and child are on different CPUs, now get the | |
1757 | * parent runqueue to update the parent's ->sleep_avg: | |
1758 | */ | |
1759 | task_rq_unlock(rq, &flags); | |
1760 | this_rq = task_rq_lock(current, &flags); | |
1761 | } | |
1762 | current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) * | |
1763 | PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS); | |
1764 | task_rq_unlock(this_rq, &flags); | |
1765 | } | |
1766 | ||
1767 | /* | |
1768 | * Potentially available exiting-child timeslices are | |
1769 | * retrieved here - this way the parent does not get | |
1770 | * penalized for creating too many threads. | |
1771 | * | |
1772 | * (this cannot be used to 'generate' timeslices | |
1773 | * artificially, because any timeslice recovered here | |
1774 | * was given away by the parent in the first place.) | |
1775 | */ | |
36c8b586 | 1776 | void fastcall sched_exit(struct task_struct *p) |
1da177e4 LT |
1777 | { |
1778 | unsigned long flags; | |
70b97a7f | 1779 | struct rq *rq; |
1da177e4 LT |
1780 | |
1781 | /* | |
1782 | * If the child was a (relative-) CPU hog then decrease | |
1783 | * the sleep_avg of the parent as well. | |
1784 | */ | |
1785 | rq = task_rq_lock(p->parent, &flags); | |
889dfafe | 1786 | if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) { |
1da177e4 LT |
1787 | p->parent->time_slice += p->time_slice; |
1788 | if (unlikely(p->parent->time_slice > task_timeslice(p))) | |
1789 | p->parent->time_slice = task_timeslice(p); | |
1790 | } | |
1791 | if (p->sleep_avg < p->parent->sleep_avg) | |
1792 | p->parent->sleep_avg = p->parent->sleep_avg / | |
1793 | (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg / | |
1794 | (EXIT_WEIGHT + 1); | |
1795 | task_rq_unlock(rq, &flags); | |
1796 | } | |
1797 | ||
4866cde0 NP |
1798 | /** |
1799 | * prepare_task_switch - prepare to switch tasks | |
1800 | * @rq: the runqueue preparing to switch | |
1801 | * @next: the task we are going to switch to. | |
1802 | * | |
1803 | * This is called with the rq lock held and interrupts off. It must | |
1804 | * be paired with a subsequent finish_task_switch after the context | |
1805 | * switch. | |
1806 | * | |
1807 | * prepare_task_switch sets up locking and calls architecture specific | |
1808 | * hooks. | |
1809 | */ | |
70b97a7f | 1810 | static inline void prepare_task_switch(struct rq *rq, struct task_struct *next) |
4866cde0 NP |
1811 | { |
1812 | prepare_lock_switch(rq, next); | |
1813 | prepare_arch_switch(next); | |
1814 | } | |
1815 | ||
1da177e4 LT |
1816 | /** |
1817 | * finish_task_switch - clean up after a task-switch | |
344babaa | 1818 | * @rq: runqueue associated with task-switch |
1da177e4 LT |
1819 | * @prev: the thread we just switched away from. |
1820 | * | |
4866cde0 NP |
1821 | * finish_task_switch must be called after the context switch, paired |
1822 | * with a prepare_task_switch call before the context switch. | |
1823 | * finish_task_switch will reconcile locking set up by prepare_task_switch, | |
1824 | * and do any other architecture-specific cleanup actions. | |
1da177e4 LT |
1825 | * |
1826 | * Note that we may have delayed dropping an mm in context_switch(). If | |
1827 | * so, we finish that here outside of the runqueue lock. (Doing it | |
1828 | * with the lock held can cause deadlocks; see schedule() for | |
1829 | * details.) | |
1830 | */ | |
70b97a7f | 1831 | static inline void finish_task_switch(struct rq *rq, struct task_struct *prev) |
1da177e4 LT |
1832 | __releases(rq->lock) |
1833 | { | |
1da177e4 | 1834 | struct mm_struct *mm = rq->prev_mm; |
55a101f8 | 1835 | long prev_state; |
1da177e4 LT |
1836 | |
1837 | rq->prev_mm = NULL; | |
1838 | ||
1839 | /* | |
1840 | * A task struct has one reference for the use as "current". | |
c394cc9f | 1841 | * If a task dies, then it sets TASK_DEAD in tsk->state and calls |
55a101f8 ON |
1842 | * schedule one last time. The schedule call will never return, and |
1843 | * the scheduled task must drop that reference. | |
c394cc9f | 1844 | * The test for TASK_DEAD must occur while the runqueue locks are |
1da177e4 LT |
1845 | * still held, otherwise prev could be scheduled on another cpu, die |
1846 | * there before we look at prev->state, and then the reference would | |
1847 | * be dropped twice. | |
1848 | * Manfred Spraul <manfred@colorfullife.com> | |
1849 | */ | |
55a101f8 | 1850 | prev_state = prev->state; |
4866cde0 NP |
1851 | finish_arch_switch(prev); |
1852 | finish_lock_switch(rq, prev); | |
1da177e4 LT |
1853 | if (mm) |
1854 | mmdrop(mm); | |
c394cc9f | 1855 | if (unlikely(prev_state == TASK_DEAD)) { |
c6fd91f0 | 1856 | /* |
1857 | * Remove function-return probe instances associated with this | |
1858 | * task and put them back on the free list. | |
1859 | */ | |
1860 | kprobe_flush_task(prev); | |
1da177e4 | 1861 | put_task_struct(prev); |
c6fd91f0 | 1862 | } |
1da177e4 LT |
1863 | } |
1864 | ||
1865 | /** | |
1866 | * schedule_tail - first thing a freshly forked thread must call. | |
1867 | * @prev: the thread we just switched away from. | |
1868 | */ | |
36c8b586 | 1869 | asmlinkage void schedule_tail(struct task_struct *prev) |
1da177e4 LT |
1870 | __releases(rq->lock) |
1871 | { | |
70b97a7f IM |
1872 | struct rq *rq = this_rq(); |
1873 | ||
4866cde0 NP |
1874 | finish_task_switch(rq, prev); |
1875 | #ifdef __ARCH_WANT_UNLOCKED_CTXSW | |
1876 | /* In this case, finish_task_switch does not reenable preemption */ | |
1877 | preempt_enable(); | |
1878 | #endif | |
1da177e4 LT |
1879 | if (current->set_child_tid) |
1880 | put_user(current->pid, current->set_child_tid); | |
1881 | } | |
1882 | ||
1883 | /* | |
1884 | * context_switch - switch to the new MM and the new | |
1885 | * thread's register state. | |
1886 | */ | |
36c8b586 | 1887 | static inline struct task_struct * |
70b97a7f | 1888 | context_switch(struct rq *rq, struct task_struct *prev, |
36c8b586 | 1889 | struct task_struct *next) |
1da177e4 LT |
1890 | { |
1891 | struct mm_struct *mm = next->mm; | |
1892 | struct mm_struct *oldmm = prev->active_mm; | |
1893 | ||
9226d125 ZA |
1894 | /* |
1895 | * For paravirt, this is coupled with an exit in switch_to to | |
1896 | * combine the page table reload and the switch backend into | |
1897 | * one hypercall. | |
1898 | */ | |
1899 | arch_enter_lazy_cpu_mode(); | |
1900 | ||
beed33a8 | 1901 | if (!mm) { |
1da177e4 LT |
1902 | next->active_mm = oldmm; |
1903 | atomic_inc(&oldmm->mm_count); | |
1904 | enter_lazy_tlb(oldmm, next); | |
1905 | } else | |
1906 | switch_mm(oldmm, mm, next); | |
1907 | ||
beed33a8 | 1908 | if (!prev->mm) { |
1da177e4 LT |
1909 | prev->active_mm = NULL; |
1910 | WARN_ON(rq->prev_mm); | |
1911 | rq->prev_mm = oldmm; | |
1912 | } | |
3a5f5e48 IM |
1913 | /* |
1914 | * Since the runqueue lock will be released by the next | |
1915 | * task (which is an invalid locking op but in the case | |
1916 | * of the scheduler it's an obvious special-case), so we | |
1917 | * do an early lockdep release here: | |
1918 | */ | |
1919 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW | |
8a25d5de | 1920 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); |
3a5f5e48 | 1921 | #endif |
1da177e4 LT |
1922 | |
1923 | /* Here we just switch the register state and the stack. */ | |
1924 | switch_to(prev, next, prev); | |
1925 | ||
1926 | return prev; | |
1927 | } | |
1928 | ||
1929 | /* | |
1930 | * nr_running, nr_uninterruptible and nr_context_switches: | |
1931 | * | |
1932 | * externally visible scheduler statistics: current number of runnable | |
1933 | * threads, current number of uninterruptible-sleeping threads, total | |
1934 | * number of context switches performed since bootup. | |
1935 | */ | |
1936 | unsigned long nr_running(void) | |
1937 | { | |
1938 | unsigned long i, sum = 0; | |
1939 | ||
1940 | for_each_online_cpu(i) | |
1941 | sum += cpu_rq(i)->nr_running; | |
1942 | ||
1943 | return sum; | |
1944 | } | |
1945 | ||
1946 | unsigned long nr_uninterruptible(void) | |
1947 | { | |
1948 | unsigned long i, sum = 0; | |
1949 | ||
0a945022 | 1950 | for_each_possible_cpu(i) |
1da177e4 LT |
1951 | sum += cpu_rq(i)->nr_uninterruptible; |
1952 | ||
1953 | /* | |
1954 | * Since we read the counters lockless, it might be slightly | |
1955 | * inaccurate. Do not allow it to go below zero though: | |
1956 | */ | |
1957 | if (unlikely((long)sum < 0)) | |
1958 | sum = 0; | |
1959 | ||
1960 | return sum; | |
1961 | } | |
1962 | ||
1963 | unsigned long long nr_context_switches(void) | |
1964 | { | |
cc94abfc SR |
1965 | int i; |
1966 | unsigned long long sum = 0; | |
1da177e4 | 1967 | |
0a945022 | 1968 | for_each_possible_cpu(i) |
1da177e4 LT |
1969 | sum += cpu_rq(i)->nr_switches; |
1970 | ||
1971 | return sum; | |
1972 | } | |
1973 | ||
1974 | unsigned long nr_iowait(void) | |
1975 | { | |
1976 | unsigned long i, sum = 0; | |
1977 | ||
0a945022 | 1978 | for_each_possible_cpu(i) |
1da177e4 LT |
1979 | sum += atomic_read(&cpu_rq(i)->nr_iowait); |
1980 | ||
1981 | return sum; | |
1982 | } | |
1983 | ||
db1b1fef JS |
1984 | unsigned long nr_active(void) |
1985 | { | |
1986 | unsigned long i, running = 0, uninterruptible = 0; | |
1987 | ||
1988 | for_each_online_cpu(i) { | |
1989 | running += cpu_rq(i)->nr_running; | |
1990 | uninterruptible += cpu_rq(i)->nr_uninterruptible; | |
1991 | } | |
1992 | ||
1993 | if (unlikely((long)uninterruptible < 0)) | |
1994 | uninterruptible = 0; | |
1995 | ||
1996 | return running + uninterruptible; | |
1997 | } | |
1998 | ||
1da177e4 LT |
1999 | #ifdef CONFIG_SMP |
2000 | ||
48f24c4d IM |
2001 | /* |
2002 | * Is this task likely cache-hot: | |
2003 | */ | |
2004 | static inline int | |
2005 | task_hot(struct task_struct *p, unsigned long long now, struct sched_domain *sd) | |
2006 | { | |
2007 | return (long long)(now - p->last_ran) < (long long)sd->cache_hot_time; | |
2008 | } | |
2009 | ||
1da177e4 LT |
2010 | /* |
2011 | * double_rq_lock - safely lock two runqueues | |
2012 | * | |
2013 | * Note this does not disable interrupts like task_rq_lock, | |
2014 | * you need to do so manually before calling. | |
2015 | */ | |
70b97a7f | 2016 | static void double_rq_lock(struct rq *rq1, struct rq *rq2) |
1da177e4 LT |
2017 | __acquires(rq1->lock) |
2018 | __acquires(rq2->lock) | |
2019 | { | |
054b9108 | 2020 | BUG_ON(!irqs_disabled()); |
1da177e4 LT |
2021 | if (rq1 == rq2) { |
2022 | spin_lock(&rq1->lock); | |
2023 | __acquire(rq2->lock); /* Fake it out ;) */ | |
2024 | } else { | |
c96d145e | 2025 | if (rq1 < rq2) { |
1da177e4 LT |
2026 | spin_lock(&rq1->lock); |
2027 | spin_lock(&rq2->lock); | |
2028 | } else { | |
2029 | spin_lock(&rq2->lock); | |
2030 | spin_lock(&rq1->lock); | |
2031 | } | |
2032 | } | |
2033 | } | |
2034 | ||
2035 | /* | |
2036 | * double_rq_unlock - safely unlock two runqueues | |
2037 | * | |
2038 | * Note this does not restore interrupts like task_rq_unlock, | |
2039 | * you need to do so manually after calling. | |
2040 | */ | |
70b97a7f | 2041 | static void double_rq_unlock(struct rq *rq1, struct rq *rq2) |
1da177e4 LT |
2042 | __releases(rq1->lock) |
2043 | __releases(rq2->lock) | |
2044 | { | |
2045 | spin_unlock(&rq1->lock); | |
2046 | if (rq1 != rq2) | |
2047 | spin_unlock(&rq2->lock); | |
2048 | else | |
2049 | __release(rq2->lock); | |
2050 | } | |
2051 | ||
2052 | /* | |
2053 | * double_lock_balance - lock the busiest runqueue, this_rq is locked already. | |
2054 | */ | |
70b97a7f | 2055 | static void double_lock_balance(struct rq *this_rq, struct rq *busiest) |
1da177e4 LT |
2056 | __releases(this_rq->lock) |
2057 | __acquires(busiest->lock) | |
2058 | __acquires(this_rq->lock) | |
2059 | { | |
054b9108 KK |
2060 | if (unlikely(!irqs_disabled())) { |
2061 | /* printk() doesn't work good under rq->lock */ | |
2062 | spin_unlock(&this_rq->lock); | |
2063 | BUG_ON(1); | |
2064 | } | |
1da177e4 | 2065 | if (unlikely(!spin_trylock(&busiest->lock))) { |
c96d145e | 2066 | if (busiest < this_rq) { |
1da177e4 LT |
2067 | spin_unlock(&this_rq->lock); |
2068 | spin_lock(&busiest->lock); | |
2069 | spin_lock(&this_rq->lock); | |
2070 | } else | |
2071 | spin_lock(&busiest->lock); | |
2072 | } | |
2073 | } | |
2074 | ||
1da177e4 LT |
2075 | /* |
2076 | * If dest_cpu is allowed for this process, migrate the task to it. | |
2077 | * This is accomplished by forcing the cpu_allowed mask to only | |
2078 | * allow dest_cpu, which will force the cpu onto dest_cpu. Then | |
2079 | * the cpu_allowed mask is restored. | |
2080 | */ | |
36c8b586 | 2081 | static void sched_migrate_task(struct task_struct *p, int dest_cpu) |
1da177e4 | 2082 | { |
70b97a7f | 2083 | struct migration_req req; |
1da177e4 | 2084 | unsigned long flags; |
70b97a7f | 2085 | struct rq *rq; |
1da177e4 LT |
2086 | |
2087 | rq = task_rq_lock(p, &flags); | |
2088 | if (!cpu_isset(dest_cpu, p->cpus_allowed) | |
2089 | || unlikely(cpu_is_offline(dest_cpu))) | |
2090 | goto out; | |
2091 | ||
2092 | /* force the process onto the specified CPU */ | |
2093 | if (migrate_task(p, dest_cpu, &req)) { | |
2094 | /* Need to wait for migration thread (might exit: take ref). */ | |
2095 | struct task_struct *mt = rq->migration_thread; | |
36c8b586 | 2096 | |
1da177e4 LT |
2097 | get_task_struct(mt); |
2098 | task_rq_unlock(rq, &flags); | |
2099 | wake_up_process(mt); | |
2100 | put_task_struct(mt); | |
2101 | wait_for_completion(&req.done); | |
36c8b586 | 2102 | |
1da177e4 LT |
2103 | return; |
2104 | } | |
2105 | out: | |
2106 | task_rq_unlock(rq, &flags); | |
2107 | } | |
2108 | ||
2109 | /* | |
476d139c NP |
2110 | * sched_exec - execve() is a valuable balancing opportunity, because at |
2111 | * this point the task has the smallest effective memory and cache footprint. | |
1da177e4 LT |
2112 | */ |
2113 | void sched_exec(void) | |
2114 | { | |
1da177e4 | 2115 | int new_cpu, this_cpu = get_cpu(); |
476d139c | 2116 | new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC); |
1da177e4 | 2117 | put_cpu(); |
476d139c NP |
2118 | if (new_cpu != this_cpu) |
2119 | sched_migrate_task(current, new_cpu); | |
1da177e4 LT |
2120 | } |
2121 | ||
2122 | /* | |
2123 | * pull_task - move a task from a remote runqueue to the local runqueue. | |
2124 | * Both runqueues must be locked. | |
2125 | */ | |
70b97a7f IM |
2126 | static void pull_task(struct rq *src_rq, struct prio_array *src_array, |
2127 | struct task_struct *p, struct rq *this_rq, | |
2128 | struct prio_array *this_array, int this_cpu) | |
1da177e4 LT |
2129 | { |
2130 | dequeue_task(p, src_array); | |
2dd73a4f | 2131 | dec_nr_running(p, src_rq); |
1da177e4 | 2132 | set_task_cpu(p, this_cpu); |
2dd73a4f | 2133 | inc_nr_running(p, this_rq); |
1da177e4 | 2134 | enqueue_task(p, this_array); |
b18ec803 MG |
2135 | p->timestamp = (p->timestamp - src_rq->most_recent_timestamp) |
2136 | + this_rq->most_recent_timestamp; | |
1da177e4 LT |
2137 | /* |
2138 | * Note that idle threads have a prio of MAX_PRIO, for this test | |
2139 | * to be always true for them. | |
2140 | */ | |
2141 | if (TASK_PREEMPTS_CURR(p, this_rq)) | |
2142 | resched_task(this_rq->curr); | |
2143 | } | |
2144 | ||
2145 | /* | |
2146 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
2147 | */ | |
858119e1 | 2148 | static |
70b97a7f | 2149 | int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu, |
95cdf3b7 IM |
2150 | struct sched_domain *sd, enum idle_type idle, |
2151 | int *all_pinned) | |
1da177e4 LT |
2152 | { |
2153 | /* | |
2154 | * We do not migrate tasks that are: | |
2155 | * 1) running (obviously), or | |
2156 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
2157 | * 3) are cache-hot on their current CPU. | |
2158 | */ | |
1da177e4 LT |
2159 | if (!cpu_isset(this_cpu, p->cpus_allowed)) |
2160 | return 0; | |
81026794 NP |
2161 | *all_pinned = 0; |
2162 | ||
2163 | if (task_running(rq, p)) | |
2164 | return 0; | |
1da177e4 LT |
2165 | |
2166 | /* | |
2167 | * Aggressive migration if: | |
cafb20c1 | 2168 | * 1) task is cache cold, or |
1da177e4 LT |
2169 | * 2) too many balance attempts have failed. |
2170 | */ | |
2171 | ||
b18ec803 MG |
2172 | if (sd->nr_balance_failed > sd->cache_nice_tries) { |
2173 | #ifdef CONFIG_SCHEDSTATS | |
2174 | if (task_hot(p, rq->most_recent_timestamp, sd)) | |
2175 | schedstat_inc(sd, lb_hot_gained[idle]); | |
2176 | #endif | |
1da177e4 | 2177 | return 1; |
b18ec803 | 2178 | } |
1da177e4 | 2179 | |
b18ec803 | 2180 | if (task_hot(p, rq->most_recent_timestamp, sd)) |
81026794 | 2181 | return 0; |
1da177e4 LT |
2182 | return 1; |
2183 | } | |
2184 | ||
615052dc | 2185 | #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio) |
48f24c4d | 2186 | |
1da177e4 | 2187 | /* |
2dd73a4f PW |
2188 | * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted |
2189 | * load from busiest to this_rq, as part of a balancing operation within | |
2190 | * "domain". Returns the number of tasks moved. | |
1da177e4 LT |
2191 | * |
2192 | * Called with both runqueues locked. | |
2193 | */ | |
70b97a7f | 2194 | static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, |
2dd73a4f PW |
2195 | unsigned long max_nr_move, unsigned long max_load_move, |
2196 | struct sched_domain *sd, enum idle_type idle, | |
2197 | int *all_pinned) | |
1da177e4 | 2198 | { |
48f24c4d IM |
2199 | int idx, pulled = 0, pinned = 0, this_best_prio, best_prio, |
2200 | best_prio_seen, skip_for_load; | |
70b97a7f | 2201 | struct prio_array *array, *dst_array; |
1da177e4 | 2202 | struct list_head *head, *curr; |
36c8b586 | 2203 | struct task_struct *tmp; |
2dd73a4f | 2204 | long rem_load_move; |
1da177e4 | 2205 | |
2dd73a4f | 2206 | if (max_nr_move == 0 || max_load_move == 0) |
1da177e4 LT |
2207 | goto out; |
2208 | ||
2dd73a4f | 2209 | rem_load_move = max_load_move; |
81026794 | 2210 | pinned = 1; |
615052dc | 2211 | this_best_prio = rq_best_prio(this_rq); |
48f24c4d | 2212 | best_prio = rq_best_prio(busiest); |
615052dc PW |
2213 | /* |
2214 | * Enable handling of the case where there is more than one task | |
2215 | * with the best priority. If the current running task is one | |
48f24c4d | 2216 | * of those with prio==best_prio we know it won't be moved |
615052dc PW |
2217 | * and therefore it's safe to override the skip (based on load) of |
2218 | * any task we find with that prio. | |
2219 | */ | |
48f24c4d | 2220 | best_prio_seen = best_prio == busiest->curr->prio; |
81026794 | 2221 | |
1da177e4 LT |
2222 | /* |
2223 | * We first consider expired tasks. Those will likely not be | |
2224 | * executed in the near future, and they are most likely to | |
2225 | * be cache-cold, thus switching CPUs has the least effect | |
2226 | * on them. | |
2227 | */ | |
2228 | if (busiest->expired->nr_active) { | |
2229 | array = busiest->expired; | |
2230 | dst_array = this_rq->expired; | |
2231 | } else { | |
2232 | array = busiest->active; | |
2233 | dst_array = this_rq->active; | |
2234 | } | |
2235 | ||
2236 | new_array: | |
2237 | /* Start searching at priority 0: */ | |
2238 | idx = 0; | |
2239 | skip_bitmap: | |
2240 | if (!idx) | |
2241 | idx = sched_find_first_bit(array->bitmap); | |
2242 | else | |
2243 | idx = find_next_bit(array->bitmap, MAX_PRIO, idx); | |
2244 | if (idx >= MAX_PRIO) { | |
2245 | if (array == busiest->expired && busiest->active->nr_active) { | |
2246 | array = busiest->active; | |
2247 | dst_array = this_rq->active; | |
2248 | goto new_array; | |
2249 | } | |
2250 | goto out; | |
2251 | } | |
2252 | ||
2253 | head = array->queue + idx; | |
2254 | curr = head->prev; | |
2255 | skip_queue: | |
36c8b586 | 2256 | tmp = list_entry(curr, struct task_struct, run_list); |
1da177e4 LT |
2257 | |
2258 | curr = curr->prev; | |
2259 | ||
50ddd969 PW |
2260 | /* |
2261 | * To help distribute high priority tasks accross CPUs we don't | |
2262 | * skip a task if it will be the highest priority task (i.e. smallest | |
2263 | * prio value) on its new queue regardless of its load weight | |
2264 | */ | |
615052dc PW |
2265 | skip_for_load = tmp->load_weight > rem_load_move; |
2266 | if (skip_for_load && idx < this_best_prio) | |
48f24c4d | 2267 | skip_for_load = !best_prio_seen && idx == best_prio; |
615052dc | 2268 | if (skip_for_load || |
2dd73a4f | 2269 | !can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) { |
48f24c4d IM |
2270 | |
2271 | best_prio_seen |= idx == best_prio; | |
1da177e4 LT |
2272 | if (curr != head) |
2273 | goto skip_queue; | |
2274 | idx++; | |
2275 | goto skip_bitmap; | |
2276 | } | |
2277 | ||
1da177e4 LT |
2278 | pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu); |
2279 | pulled++; | |
2dd73a4f | 2280 | rem_load_move -= tmp->load_weight; |
1da177e4 | 2281 | |
2dd73a4f PW |
2282 | /* |
2283 | * We only want to steal up to the prescribed number of tasks | |
2284 | * and the prescribed amount of weighted load. | |
2285 | */ | |
2286 | if (pulled < max_nr_move && rem_load_move > 0) { | |
615052dc PW |
2287 | if (idx < this_best_prio) |
2288 | this_best_prio = idx; | |
1da177e4 LT |
2289 | if (curr != head) |
2290 | goto skip_queue; | |
2291 | idx++; | |
2292 | goto skip_bitmap; | |
2293 | } | |
2294 | out: | |
2295 | /* | |
2296 | * Right now, this is the only place pull_task() is called, | |
2297 | * so we can safely collect pull_task() stats here rather than | |
2298 | * inside pull_task(). | |
2299 | */ | |
2300 | schedstat_add(sd, lb_gained[idle], pulled); | |
81026794 NP |
2301 | |
2302 | if (all_pinned) | |
2303 | *all_pinned = pinned; | |
1da177e4 LT |
2304 | return pulled; |
2305 | } | |
2306 | ||
2307 | /* | |
2308 | * find_busiest_group finds and returns the busiest CPU group within the | |
48f24c4d IM |
2309 | * domain. It calculates and returns the amount of weighted load which |
2310 | * should be moved to restore balance via the imbalance parameter. | |
1da177e4 LT |
2311 | */ |
2312 | static struct sched_group * | |
2313 | find_busiest_group(struct sched_domain *sd, int this_cpu, | |
0a2966b4 | 2314 | unsigned long *imbalance, enum idle_type idle, int *sd_idle, |
783609c6 | 2315 | cpumask_t *cpus, int *balance) |
1da177e4 LT |
2316 | { |
2317 | struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups; | |
2318 | unsigned long max_load, avg_load, total_load, this_load, total_pwr; | |
0c117f1b | 2319 | unsigned long max_pull; |
2dd73a4f PW |
2320 | unsigned long busiest_load_per_task, busiest_nr_running; |
2321 | unsigned long this_load_per_task, this_nr_running; | |
7897986b | 2322 | int load_idx; |
5c45bf27 SS |
2323 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
2324 | int power_savings_balance = 1; | |
2325 | unsigned long leader_nr_running = 0, min_load_per_task = 0; | |
2326 | unsigned long min_nr_running = ULONG_MAX; | |
2327 | struct sched_group *group_min = NULL, *group_leader = NULL; | |
2328 | #endif | |
1da177e4 LT |
2329 | |
2330 | max_load = this_load = total_load = total_pwr = 0; | |
2dd73a4f PW |
2331 | busiest_load_per_task = busiest_nr_running = 0; |
2332 | this_load_per_task = this_nr_running = 0; | |
7897986b NP |
2333 | if (idle == NOT_IDLE) |
2334 | load_idx = sd->busy_idx; | |
2335 | else if (idle == NEWLY_IDLE) | |
2336 | load_idx = sd->newidle_idx; | |
2337 | else | |
2338 | load_idx = sd->idle_idx; | |
1da177e4 LT |
2339 | |
2340 | do { | |
5c45bf27 | 2341 | unsigned long load, group_capacity; |
1da177e4 LT |
2342 | int local_group; |
2343 | int i; | |
783609c6 | 2344 | unsigned int balance_cpu = -1, first_idle_cpu = 0; |
2dd73a4f | 2345 | unsigned long sum_nr_running, sum_weighted_load; |
1da177e4 LT |
2346 | |
2347 | local_group = cpu_isset(this_cpu, group->cpumask); | |
2348 | ||
783609c6 SS |
2349 | if (local_group) |
2350 | balance_cpu = first_cpu(group->cpumask); | |
2351 | ||
1da177e4 | 2352 | /* Tally up the load of all CPUs in the group */ |
2dd73a4f | 2353 | sum_weighted_load = sum_nr_running = avg_load = 0; |
1da177e4 LT |
2354 | |
2355 | for_each_cpu_mask(i, group->cpumask) { | |
0a2966b4 CL |
2356 | struct rq *rq; |
2357 | ||
2358 | if (!cpu_isset(i, *cpus)) | |
2359 | continue; | |
2360 | ||
2361 | rq = cpu_rq(i); | |
2dd73a4f | 2362 | |
5969fe06 NP |
2363 | if (*sd_idle && !idle_cpu(i)) |
2364 | *sd_idle = 0; | |
2365 | ||
1da177e4 | 2366 | /* Bias balancing toward cpus of our domain */ |
783609c6 SS |
2367 | if (local_group) { |
2368 | if (idle_cpu(i) && !first_idle_cpu) { | |
2369 | first_idle_cpu = 1; | |
2370 | balance_cpu = i; | |
2371 | } | |
2372 | ||
a2000572 | 2373 | load = target_load(i, load_idx); |
783609c6 | 2374 | } else |
a2000572 | 2375 | load = source_load(i, load_idx); |
1da177e4 LT |
2376 | |
2377 | avg_load += load; | |
2dd73a4f PW |
2378 | sum_nr_running += rq->nr_running; |
2379 | sum_weighted_load += rq->raw_weighted_load; | |
1da177e4 LT |
2380 | } |
2381 | ||
783609c6 SS |
2382 | /* |
2383 | * First idle cpu or the first cpu(busiest) in this sched group | |
2384 | * is eligible for doing load balancing at this and above | |
2385 | * domains. | |
2386 | */ | |
2387 | if (local_group && balance_cpu != this_cpu && balance) { | |
2388 | *balance = 0; | |
2389 | goto ret; | |
2390 | } | |
2391 | ||
1da177e4 | 2392 | total_load += avg_load; |
5517d86b | 2393 | total_pwr += group->__cpu_power; |
1da177e4 LT |
2394 | |
2395 | /* Adjust by relative CPU power of the group */ | |
5517d86b ED |
2396 | avg_load = sg_div_cpu_power(group, |
2397 | avg_load * SCHED_LOAD_SCALE); | |
1da177e4 | 2398 | |
5517d86b | 2399 | group_capacity = group->__cpu_power / SCHED_LOAD_SCALE; |
5c45bf27 | 2400 | |
1da177e4 LT |
2401 | if (local_group) { |
2402 | this_load = avg_load; | |
2403 | this = group; | |
2dd73a4f PW |
2404 | this_nr_running = sum_nr_running; |
2405 | this_load_per_task = sum_weighted_load; | |
2406 | } else if (avg_load > max_load && | |
5c45bf27 | 2407 | sum_nr_running > group_capacity) { |
1da177e4 LT |
2408 | max_load = avg_load; |
2409 | busiest = group; | |
2dd73a4f PW |
2410 | busiest_nr_running = sum_nr_running; |
2411 | busiest_load_per_task = sum_weighted_load; | |
1da177e4 | 2412 | } |
5c45bf27 SS |
2413 | |
2414 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
2415 | /* | |
2416 | * Busy processors will not participate in power savings | |
2417 | * balance. | |
2418 | */ | |
2419 | if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE)) | |
2420 | goto group_next; | |
2421 | ||
2422 | /* | |
2423 | * If the local group is idle or completely loaded | |
2424 | * no need to do power savings balance at this domain | |
2425 | */ | |
2426 | if (local_group && (this_nr_running >= group_capacity || | |
2427 | !this_nr_running)) | |
2428 | power_savings_balance = 0; | |
2429 | ||
2430 | /* | |
2431 | * If a group is already running at full capacity or idle, | |
2432 | * don't include that group in power savings calculations | |
2433 | */ | |
2434 | if (!power_savings_balance || sum_nr_running >= group_capacity | |
2435 | || !sum_nr_running) | |
2436 | goto group_next; | |
2437 | ||
2438 | /* | |
2439 | * Calculate the group which has the least non-idle load. | |
2440 | * This is the group from where we need to pick up the load | |
2441 | * for saving power | |
2442 | */ | |
2443 | if ((sum_nr_running < min_nr_running) || | |
2444 | (sum_nr_running == min_nr_running && | |
2445 | first_cpu(group->cpumask) < | |
2446 | first_cpu(group_min->cpumask))) { | |
2447 | group_min = group; | |
2448 | min_nr_running = sum_nr_running; | |
2449 | min_load_per_task = sum_weighted_load / | |
2450 | sum_nr_running; | |
2451 | } | |
2452 | ||
2453 | /* | |
2454 | * Calculate the group which is almost near its | |
2455 | * capacity but still has some space to pick up some load | |
2456 | * from other group and save more power | |
2457 | */ | |
48f24c4d | 2458 | if (sum_nr_running <= group_capacity - 1) { |
5c45bf27 SS |
2459 | if (sum_nr_running > leader_nr_running || |
2460 | (sum_nr_running == leader_nr_running && | |
2461 | first_cpu(group->cpumask) > | |
2462 | first_cpu(group_leader->cpumask))) { | |
2463 | group_leader = group; | |
2464 | leader_nr_running = sum_nr_running; | |
2465 | } | |
48f24c4d | 2466 | } |
5c45bf27 SS |
2467 | group_next: |
2468 | #endif | |
1da177e4 LT |
2469 | group = group->next; |
2470 | } while (group != sd->groups); | |
2471 | ||
2dd73a4f | 2472 | if (!busiest || this_load >= max_load || busiest_nr_running == 0) |
1da177e4 LT |
2473 | goto out_balanced; |
2474 | ||
2475 | avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr; | |
2476 | ||
2477 | if (this_load >= avg_load || | |
2478 | 100*max_load <= sd->imbalance_pct*this_load) | |
2479 | goto out_balanced; | |
2480 | ||
2dd73a4f | 2481 | busiest_load_per_task /= busiest_nr_running; |
1da177e4 LT |
2482 | /* |
2483 | * We're trying to get all the cpus to the average_load, so we don't | |
2484 | * want to push ourselves above the average load, nor do we wish to | |
2485 | * reduce the max loaded cpu below the average load, as either of these | |
2486 | * actions would just result in more rebalancing later, and ping-pong | |
2487 | * tasks around. Thus we look for the minimum possible imbalance. | |
2488 | * Negative imbalances (*we* are more loaded than anyone else) will | |
2489 | * be counted as no imbalance for these purposes -- we can't fix that | |
2490 | * by pulling tasks to us. Be careful of negative numbers as they'll | |
2491 | * appear as very large values with unsigned longs. | |
2492 | */ | |
2dd73a4f PW |
2493 | if (max_load <= busiest_load_per_task) |
2494 | goto out_balanced; | |
2495 | ||
2496 | /* | |
2497 | * In the presence of smp nice balancing, certain scenarios can have | |
2498 | * max load less than avg load(as we skip the groups at or below | |
2499 | * its cpu_power, while calculating max_load..) | |
2500 | */ | |
2501 | if (max_load < avg_load) { | |
2502 | *imbalance = 0; | |
2503 | goto small_imbalance; | |
2504 | } | |
0c117f1b SS |
2505 | |
2506 | /* Don't want to pull so many tasks that a group would go idle */ | |
2dd73a4f | 2507 | max_pull = min(max_load - avg_load, max_load - busiest_load_per_task); |
0c117f1b | 2508 | |
1da177e4 | 2509 | /* How much load to actually move to equalise the imbalance */ |
5517d86b ED |
2510 | *imbalance = min(max_pull * busiest->__cpu_power, |
2511 | (avg_load - this_load) * this->__cpu_power) | |
1da177e4 LT |
2512 | / SCHED_LOAD_SCALE; |
2513 | ||
2dd73a4f PW |
2514 | /* |
2515 | * if *imbalance is less than the average load per runnable task | |
2516 | * there is no gaurantee that any tasks will be moved so we'll have | |
2517 | * a think about bumping its value to force at least one task to be | |
2518 | * moved | |
2519 | */ | |
2520 | if (*imbalance < busiest_load_per_task) { | |
48f24c4d | 2521 | unsigned long tmp, pwr_now, pwr_move; |
2dd73a4f PW |
2522 | unsigned int imbn; |
2523 | ||
2524 | small_imbalance: | |
2525 | pwr_move = pwr_now = 0; | |
2526 | imbn = 2; | |
2527 | if (this_nr_running) { | |
2528 | this_load_per_task /= this_nr_running; | |
2529 | if (busiest_load_per_task > this_load_per_task) | |
2530 | imbn = 1; | |
2531 | } else | |
2532 | this_load_per_task = SCHED_LOAD_SCALE; | |
1da177e4 | 2533 | |
2dd73a4f PW |
2534 | if (max_load - this_load >= busiest_load_per_task * imbn) { |
2535 | *imbalance = busiest_load_per_task; | |
1da177e4 LT |
2536 | return busiest; |
2537 | } | |
2538 | ||
2539 | /* | |
2540 | * OK, we don't have enough imbalance to justify moving tasks, | |
2541 | * however we may be able to increase total CPU power used by | |
2542 | * moving them. | |
2543 | */ | |
2544 | ||
5517d86b ED |
2545 | pwr_now += busiest->__cpu_power * |
2546 | min(busiest_load_per_task, max_load); | |
2547 | pwr_now += this->__cpu_power * | |
2548 | min(this_load_per_task, this_load); | |
1da177e4 LT |
2549 | pwr_now /= SCHED_LOAD_SCALE; |
2550 | ||
2551 | /* Amount of load we'd subtract */ | |
5517d86b ED |
2552 | tmp = sg_div_cpu_power(busiest, |
2553 | busiest_load_per_task * SCHED_LOAD_SCALE); | |
1da177e4 | 2554 | if (max_load > tmp) |
5517d86b | 2555 | pwr_move += busiest->__cpu_power * |
2dd73a4f | 2556 | min(busiest_load_per_task, max_load - tmp); |
1da177e4 LT |
2557 | |
2558 | /* Amount of load we'd add */ | |
5517d86b | 2559 | if (max_load * busiest->__cpu_power < |
33859f7f | 2560 | busiest_load_per_task * SCHED_LOAD_SCALE) |
5517d86b ED |
2561 | tmp = sg_div_cpu_power(this, |
2562 | max_load * busiest->__cpu_power); | |
1da177e4 | 2563 | else |
5517d86b ED |
2564 | tmp = sg_div_cpu_power(this, |
2565 | busiest_load_per_task * SCHED_LOAD_SCALE); | |
2566 | pwr_move += this->__cpu_power * | |
2567 | min(this_load_per_task, this_load + tmp); | |
1da177e4 LT |
2568 | pwr_move /= SCHED_LOAD_SCALE; |
2569 | ||
2570 | /* Move if we gain throughput */ | |
2571 | if (pwr_move <= pwr_now) | |
2572 | goto out_balanced; | |
2573 | ||
2dd73a4f | 2574 | *imbalance = busiest_load_per_task; |
1da177e4 LT |
2575 | } |
2576 | ||
1da177e4 LT |
2577 | return busiest; |
2578 | ||
2579 | out_balanced: | |
5c45bf27 SS |
2580 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
2581 | if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE)) | |
2582 | goto ret; | |
1da177e4 | 2583 | |
5c45bf27 SS |
2584 | if (this == group_leader && group_leader != group_min) { |
2585 | *imbalance = min_load_per_task; | |
2586 | return group_min; | |
2587 | } | |
5c45bf27 | 2588 | #endif |
783609c6 | 2589 | ret: |
1da177e4 LT |
2590 | *imbalance = 0; |
2591 | return NULL; | |
2592 | } | |
2593 | ||
2594 | /* | |
2595 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
2596 | */ | |
70b97a7f | 2597 | static struct rq * |
48f24c4d | 2598 | find_busiest_queue(struct sched_group *group, enum idle_type idle, |
0a2966b4 | 2599 | unsigned long imbalance, cpumask_t *cpus) |
1da177e4 | 2600 | { |
70b97a7f | 2601 | struct rq *busiest = NULL, *rq; |
2dd73a4f | 2602 | unsigned long max_load = 0; |
1da177e4 LT |
2603 | int i; |
2604 | ||
2605 | for_each_cpu_mask(i, group->cpumask) { | |
0a2966b4 CL |
2606 | |
2607 | if (!cpu_isset(i, *cpus)) | |
2608 | continue; | |
2609 | ||
48f24c4d | 2610 | rq = cpu_rq(i); |
2dd73a4f | 2611 | |
48f24c4d | 2612 | if (rq->nr_running == 1 && rq->raw_weighted_load > imbalance) |
2dd73a4f | 2613 | continue; |
1da177e4 | 2614 | |
48f24c4d IM |
2615 | if (rq->raw_weighted_load > max_load) { |
2616 | max_load = rq->raw_weighted_load; | |
2617 | busiest = rq; | |
1da177e4 LT |
2618 | } |
2619 | } | |
2620 | ||
2621 | return busiest; | |
2622 | } | |
2623 | ||
77391d71 NP |
2624 | /* |
2625 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
2626 | * so long as it is large enough. | |
2627 | */ | |
2628 | #define MAX_PINNED_INTERVAL 512 | |
2629 | ||
48f24c4d IM |
2630 | static inline unsigned long minus_1_or_zero(unsigned long n) |
2631 | { | |
2632 | return n > 0 ? n - 1 : 0; | |
2633 | } | |
2634 | ||
1da177e4 LT |
2635 | /* |
2636 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
2637 | * tasks if there is an imbalance. | |
1da177e4 | 2638 | */ |
70b97a7f | 2639 | static int load_balance(int this_cpu, struct rq *this_rq, |
783609c6 SS |
2640 | struct sched_domain *sd, enum idle_type idle, |
2641 | int *balance) | |
1da177e4 | 2642 | { |
48f24c4d | 2643 | int nr_moved, all_pinned = 0, active_balance = 0, sd_idle = 0; |
1da177e4 | 2644 | struct sched_group *group; |
1da177e4 | 2645 | unsigned long imbalance; |
70b97a7f | 2646 | struct rq *busiest; |
0a2966b4 | 2647 | cpumask_t cpus = CPU_MASK_ALL; |
fe2eea3f | 2648 | unsigned long flags; |
5969fe06 | 2649 | |
89c4710e SS |
2650 | /* |
2651 | * When power savings policy is enabled for the parent domain, idle | |
2652 | * sibling can pick up load irrespective of busy siblings. In this case, | |
2653 | * let the state of idle sibling percolate up as IDLE, instead of | |
2654 | * portraying it as NOT_IDLE. | |
2655 | */ | |
5c45bf27 | 2656 | if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER && |
89c4710e | 2657 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
5969fe06 | 2658 | sd_idle = 1; |
1da177e4 | 2659 | |
1da177e4 LT |
2660 | schedstat_inc(sd, lb_cnt[idle]); |
2661 | ||
0a2966b4 CL |
2662 | redo: |
2663 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle, | |
783609c6 SS |
2664 | &cpus, balance); |
2665 | ||
06066714 | 2666 | if (*balance == 0) |
783609c6 | 2667 | goto out_balanced; |
783609c6 | 2668 | |
1da177e4 LT |
2669 | if (!group) { |
2670 | schedstat_inc(sd, lb_nobusyg[idle]); | |
2671 | goto out_balanced; | |
2672 | } | |
2673 | ||
0a2966b4 | 2674 | busiest = find_busiest_queue(group, idle, imbalance, &cpus); |
1da177e4 LT |
2675 | if (!busiest) { |
2676 | schedstat_inc(sd, lb_nobusyq[idle]); | |
2677 | goto out_balanced; | |
2678 | } | |
2679 | ||
db935dbd | 2680 | BUG_ON(busiest == this_rq); |
1da177e4 LT |
2681 | |
2682 | schedstat_add(sd, lb_imbalance[idle], imbalance); | |
2683 | ||
2684 | nr_moved = 0; | |
2685 | if (busiest->nr_running > 1) { | |
2686 | /* | |
2687 | * Attempt to move tasks. If find_busiest_group has found | |
2688 | * an imbalance but busiest->nr_running <= 1, the group is | |
2689 | * still unbalanced. nr_moved simply stays zero, so it is | |
2690 | * correctly treated as an imbalance. | |
2691 | */ | |
fe2eea3f | 2692 | local_irq_save(flags); |
e17224bf | 2693 | double_rq_lock(this_rq, busiest); |
1da177e4 | 2694 | nr_moved = move_tasks(this_rq, this_cpu, busiest, |
48f24c4d IM |
2695 | minus_1_or_zero(busiest->nr_running), |
2696 | imbalance, sd, idle, &all_pinned); | |
e17224bf | 2697 | double_rq_unlock(this_rq, busiest); |
fe2eea3f | 2698 | local_irq_restore(flags); |
81026794 | 2699 | |
46cb4b7c SS |
2700 | /* |
2701 | * some other cpu did the load balance for us. | |
2702 | */ | |
2703 | if (nr_moved && this_cpu != smp_processor_id()) | |
2704 | resched_cpu(this_cpu); | |
2705 | ||
81026794 | 2706 | /* All tasks on this runqueue were pinned by CPU affinity */ |
0a2966b4 CL |
2707 | if (unlikely(all_pinned)) { |
2708 | cpu_clear(cpu_of(busiest), cpus); | |
2709 | if (!cpus_empty(cpus)) | |
2710 | goto redo; | |
81026794 | 2711 | goto out_balanced; |
0a2966b4 | 2712 | } |
1da177e4 | 2713 | } |
81026794 | 2714 | |
1da177e4 LT |
2715 | if (!nr_moved) { |
2716 | schedstat_inc(sd, lb_failed[idle]); | |
2717 | sd->nr_balance_failed++; | |
2718 | ||
2719 | if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) { | |
1da177e4 | 2720 | |
fe2eea3f | 2721 | spin_lock_irqsave(&busiest->lock, flags); |
fa3b6ddc SS |
2722 | |
2723 | /* don't kick the migration_thread, if the curr | |
2724 | * task on busiest cpu can't be moved to this_cpu | |
2725 | */ | |
2726 | if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) { | |
fe2eea3f | 2727 | spin_unlock_irqrestore(&busiest->lock, flags); |
fa3b6ddc SS |
2728 | all_pinned = 1; |
2729 | goto out_one_pinned; | |
2730 | } | |
2731 | ||
1da177e4 LT |
2732 | if (!busiest->active_balance) { |
2733 | busiest->active_balance = 1; | |
2734 | busiest->push_cpu = this_cpu; | |
81026794 | 2735 | active_balance = 1; |
1da177e4 | 2736 | } |
fe2eea3f | 2737 | spin_unlock_irqrestore(&busiest->lock, flags); |
81026794 | 2738 | if (active_balance) |
1da177e4 LT |
2739 | wake_up_process(busiest->migration_thread); |
2740 | ||
2741 | /* | |
2742 | * We've kicked active balancing, reset the failure | |
2743 | * counter. | |
2744 | */ | |
39507451 | 2745 | sd->nr_balance_failed = sd->cache_nice_tries+1; |
1da177e4 | 2746 | } |
81026794 | 2747 | } else |
1da177e4 LT |
2748 | sd->nr_balance_failed = 0; |
2749 | ||
81026794 | 2750 | if (likely(!active_balance)) { |
1da177e4 LT |
2751 | /* We were unbalanced, so reset the balancing interval */ |
2752 | sd->balance_interval = sd->min_interval; | |
81026794 NP |
2753 | } else { |
2754 | /* | |
2755 | * If we've begun active balancing, start to back off. This | |
2756 | * case may not be covered by the all_pinned logic if there | |
2757 | * is only 1 task on the busy runqueue (because we don't call | |
2758 | * move_tasks). | |
2759 | */ | |
2760 | if (sd->balance_interval < sd->max_interval) | |
2761 | sd->balance_interval *= 2; | |
1da177e4 LT |
2762 | } |
2763 | ||
5c45bf27 | 2764 | if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
89c4710e | 2765 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
5969fe06 | 2766 | return -1; |
1da177e4 LT |
2767 | return nr_moved; |
2768 | ||
2769 | out_balanced: | |
1da177e4 LT |
2770 | schedstat_inc(sd, lb_balanced[idle]); |
2771 | ||
16cfb1c0 | 2772 | sd->nr_balance_failed = 0; |
fa3b6ddc SS |
2773 | |
2774 | out_one_pinned: | |
1da177e4 | 2775 | /* tune up the balancing interval */ |
77391d71 NP |
2776 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || |
2777 | (sd->balance_interval < sd->max_interval)) | |
1da177e4 LT |
2778 | sd->balance_interval *= 2; |
2779 | ||
48f24c4d | 2780 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
89c4710e | 2781 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
5969fe06 | 2782 | return -1; |
1da177e4 LT |
2783 | return 0; |
2784 | } | |
2785 | ||
2786 | /* | |
2787 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
2788 | * tasks if there is an imbalance. | |
2789 | * | |
2790 | * Called from schedule when this_rq is about to become idle (NEWLY_IDLE). | |
2791 | * this_rq is locked. | |
2792 | */ | |
48f24c4d | 2793 | static int |
70b97a7f | 2794 | load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd) |
1da177e4 LT |
2795 | { |
2796 | struct sched_group *group; | |
70b97a7f | 2797 | struct rq *busiest = NULL; |
1da177e4 LT |
2798 | unsigned long imbalance; |
2799 | int nr_moved = 0; | |
5969fe06 | 2800 | int sd_idle = 0; |
0a2966b4 | 2801 | cpumask_t cpus = CPU_MASK_ALL; |
5969fe06 | 2802 | |
89c4710e SS |
2803 | /* |
2804 | * When power savings policy is enabled for the parent domain, idle | |
2805 | * sibling can pick up load irrespective of busy siblings. In this case, | |
2806 | * let the state of idle sibling percolate up as IDLE, instead of | |
2807 | * portraying it as NOT_IDLE. | |
2808 | */ | |
2809 | if (sd->flags & SD_SHARE_CPUPOWER && | |
2810 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
5969fe06 | 2811 | sd_idle = 1; |
1da177e4 LT |
2812 | |
2813 | schedstat_inc(sd, lb_cnt[NEWLY_IDLE]); | |
0a2966b4 CL |
2814 | redo: |
2815 | group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE, | |
783609c6 | 2816 | &sd_idle, &cpus, NULL); |
1da177e4 | 2817 | if (!group) { |
1da177e4 | 2818 | schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]); |
16cfb1c0 | 2819 | goto out_balanced; |
1da177e4 LT |
2820 | } |
2821 | ||
0a2966b4 CL |
2822 | busiest = find_busiest_queue(group, NEWLY_IDLE, imbalance, |
2823 | &cpus); | |
db935dbd | 2824 | if (!busiest) { |
1da177e4 | 2825 | schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]); |
16cfb1c0 | 2826 | goto out_balanced; |
1da177e4 LT |
2827 | } |
2828 | ||
db935dbd NP |
2829 | BUG_ON(busiest == this_rq); |
2830 | ||
1da177e4 | 2831 | schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance); |
d6d5cfaf NP |
2832 | |
2833 | nr_moved = 0; | |
2834 | if (busiest->nr_running > 1) { | |
2835 | /* Attempt to move tasks */ | |
2836 | double_lock_balance(this_rq, busiest); | |
2837 | nr_moved = move_tasks(this_rq, this_cpu, busiest, | |
2dd73a4f | 2838 | minus_1_or_zero(busiest->nr_running), |
81026794 | 2839 | imbalance, sd, NEWLY_IDLE, NULL); |
d6d5cfaf | 2840 | spin_unlock(&busiest->lock); |
0a2966b4 CL |
2841 | |
2842 | if (!nr_moved) { | |
2843 | cpu_clear(cpu_of(busiest), cpus); | |
2844 | if (!cpus_empty(cpus)) | |
2845 | goto redo; | |
2846 | } | |
d6d5cfaf NP |
2847 | } |
2848 | ||
5969fe06 | 2849 | if (!nr_moved) { |
1da177e4 | 2850 | schedstat_inc(sd, lb_failed[NEWLY_IDLE]); |
89c4710e SS |
2851 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
2852 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
5969fe06 NP |
2853 | return -1; |
2854 | } else | |
16cfb1c0 | 2855 | sd->nr_balance_failed = 0; |
1da177e4 | 2856 | |
1da177e4 | 2857 | return nr_moved; |
16cfb1c0 NP |
2858 | |
2859 | out_balanced: | |
2860 | schedstat_inc(sd, lb_balanced[NEWLY_IDLE]); | |
48f24c4d | 2861 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
89c4710e | 2862 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
5969fe06 | 2863 | return -1; |
16cfb1c0 | 2864 | sd->nr_balance_failed = 0; |
48f24c4d | 2865 | |
16cfb1c0 | 2866 | return 0; |
1da177e4 LT |
2867 | } |
2868 | ||
2869 | /* | |
2870 | * idle_balance is called by schedule() if this_cpu is about to become | |
2871 | * idle. Attempts to pull tasks from other CPUs. | |
2872 | */ | |
70b97a7f | 2873 | static void idle_balance(int this_cpu, struct rq *this_rq) |
1da177e4 LT |
2874 | { |
2875 | struct sched_domain *sd; | |
1bd77f2d CL |
2876 | int pulled_task = 0; |
2877 | unsigned long next_balance = jiffies + 60 * HZ; | |
1da177e4 LT |
2878 | |
2879 | for_each_domain(this_cpu, sd) { | |
2880 | if (sd->flags & SD_BALANCE_NEWIDLE) { | |
48f24c4d | 2881 | /* If we've pulled tasks over stop searching: */ |
1bd77f2d CL |
2882 | pulled_task = load_balance_newidle(this_cpu, |
2883 | this_rq, sd); | |
2884 | if (time_after(next_balance, | |
2885 | sd->last_balance + sd->balance_interval)) | |
2886 | next_balance = sd->last_balance | |
2887 | + sd->balance_interval; | |
2888 | if (pulled_task) | |
1da177e4 | 2889 | break; |
1da177e4 LT |
2890 | } |
2891 | } | |
1bd77f2d CL |
2892 | if (!pulled_task) |
2893 | /* | |
2894 | * We are going idle. next_balance may be set based on | |
2895 | * a busy processor. So reset next_balance. | |
2896 | */ | |
2897 | this_rq->next_balance = next_balance; | |
1da177e4 LT |
2898 | } |
2899 | ||
2900 | /* | |
2901 | * active_load_balance is run by migration threads. It pushes running tasks | |
2902 | * off the busiest CPU onto idle CPUs. It requires at least 1 task to be | |
2903 | * running on each physical CPU where possible, and avoids physical / | |
2904 | * logical imbalances. | |
2905 | * | |
2906 | * Called with busiest_rq locked. | |
2907 | */ | |
70b97a7f | 2908 | static void active_load_balance(struct rq *busiest_rq, int busiest_cpu) |
1da177e4 | 2909 | { |
39507451 | 2910 | int target_cpu = busiest_rq->push_cpu; |
70b97a7f IM |
2911 | struct sched_domain *sd; |
2912 | struct rq *target_rq; | |
39507451 | 2913 | |
48f24c4d | 2914 | /* Is there any task to move? */ |
39507451 | 2915 | if (busiest_rq->nr_running <= 1) |
39507451 NP |
2916 | return; |
2917 | ||
2918 | target_rq = cpu_rq(target_cpu); | |
1da177e4 LT |
2919 | |
2920 | /* | |
39507451 NP |
2921 | * This condition is "impossible", if it occurs |
2922 | * we need to fix it. Originally reported by | |
2923 | * Bjorn Helgaas on a 128-cpu setup. | |
1da177e4 | 2924 | */ |
39507451 | 2925 | BUG_ON(busiest_rq == target_rq); |
1da177e4 | 2926 | |
39507451 NP |
2927 | /* move a task from busiest_rq to target_rq */ |
2928 | double_lock_balance(busiest_rq, target_rq); | |
2929 | ||
2930 | /* Search for an sd spanning us and the target CPU. */ | |
c96d145e | 2931 | for_each_domain(target_cpu, sd) { |
39507451 | 2932 | if ((sd->flags & SD_LOAD_BALANCE) && |
48f24c4d | 2933 | cpu_isset(busiest_cpu, sd->span)) |
39507451 | 2934 | break; |
c96d145e | 2935 | } |
39507451 | 2936 | |
48f24c4d IM |
2937 | if (likely(sd)) { |
2938 | schedstat_inc(sd, alb_cnt); | |
39507451 | 2939 | |
48f24c4d IM |
2940 | if (move_tasks(target_rq, target_cpu, busiest_rq, 1, |
2941 | RTPRIO_TO_LOAD_WEIGHT(100), sd, SCHED_IDLE, | |
2942 | NULL)) | |
2943 | schedstat_inc(sd, alb_pushed); | |
2944 | else | |
2945 | schedstat_inc(sd, alb_failed); | |
2946 | } | |
39507451 | 2947 | spin_unlock(&target_rq->lock); |
1da177e4 LT |
2948 | } |
2949 | ||
7835b98b | 2950 | static void update_load(struct rq *this_rq) |
1da177e4 | 2951 | { |
7835b98b | 2952 | unsigned long this_load; |
ff91691b | 2953 | unsigned int i, scale; |
1da177e4 | 2954 | |
2dd73a4f | 2955 | this_load = this_rq->raw_weighted_load; |
48f24c4d IM |
2956 | |
2957 | /* Update our load: */ | |
ff91691b | 2958 | for (i = 0, scale = 1; i < 3; i++, scale += scale) { |
48f24c4d IM |
2959 | unsigned long old_load, new_load; |
2960 | ||
ff91691b NP |
2961 | /* scale is effectively 1 << i now, and >> i divides by scale */ |
2962 | ||
7897986b | 2963 | old_load = this_rq->cpu_load[i]; |
48f24c4d | 2964 | new_load = this_load; |
7897986b NP |
2965 | /* |
2966 | * Round up the averaging division if load is increasing. This | |
2967 | * prevents us from getting stuck on 9 if the load is 10, for | |
2968 | * example. | |
2969 | */ | |
2970 | if (new_load > old_load) | |
2971 | new_load += scale-1; | |
ff91691b | 2972 | this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i; |
7897986b | 2973 | } |
7835b98b CL |
2974 | } |
2975 | ||
46cb4b7c SS |
2976 | #ifdef CONFIG_NO_HZ |
2977 | static struct { | |
2978 | atomic_t load_balancer; | |
2979 | cpumask_t cpu_mask; | |
2980 | } nohz ____cacheline_aligned = { | |
2981 | .load_balancer = ATOMIC_INIT(-1), | |
2982 | .cpu_mask = CPU_MASK_NONE, | |
2983 | }; | |
2984 | ||
7835b98b | 2985 | /* |
46cb4b7c SS |
2986 | * This routine will try to nominate the ilb (idle load balancing) |
2987 | * owner among the cpus whose ticks are stopped. ilb owner will do the idle | |
2988 | * load balancing on behalf of all those cpus. If all the cpus in the system | |
2989 | * go into this tickless mode, then there will be no ilb owner (as there is | |
2990 | * no need for one) and all the cpus will sleep till the next wakeup event | |
2991 | * arrives... | |
2992 | * | |
2993 | * For the ilb owner, tick is not stopped. And this tick will be used | |
2994 | * for idle load balancing. ilb owner will still be part of | |
2995 | * nohz.cpu_mask.. | |
7835b98b | 2996 | * |
46cb4b7c SS |
2997 | * While stopping the tick, this cpu will become the ilb owner if there |
2998 | * is no other owner. And will be the owner till that cpu becomes busy | |
2999 | * or if all cpus in the system stop their ticks at which point | |
3000 | * there is no need for ilb owner. | |
3001 | * | |
3002 | * When the ilb owner becomes busy, it nominates another owner, during the | |
3003 | * next busy scheduler_tick() | |
3004 | */ | |
3005 | int select_nohz_load_balancer(int stop_tick) | |
3006 | { | |
3007 | int cpu = smp_processor_id(); | |
3008 | ||
3009 | if (stop_tick) { | |
3010 | cpu_set(cpu, nohz.cpu_mask); | |
3011 | cpu_rq(cpu)->in_nohz_recently = 1; | |
3012 | ||
3013 | /* | |
3014 | * If we are going offline and still the leader, give up! | |
3015 | */ | |
3016 | if (cpu_is_offline(cpu) && | |
3017 | atomic_read(&nohz.load_balancer) == cpu) { | |
3018 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu) | |
3019 | BUG(); | |
3020 | return 0; | |
3021 | } | |
3022 | ||
3023 | /* time for ilb owner also to sleep */ | |
3024 | if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) { | |
3025 | if (atomic_read(&nohz.load_balancer) == cpu) | |
3026 | atomic_set(&nohz.load_balancer, -1); | |
3027 | return 0; | |
3028 | } | |
3029 | ||
3030 | if (atomic_read(&nohz.load_balancer) == -1) { | |
3031 | /* make me the ilb owner */ | |
3032 | if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1) | |
3033 | return 1; | |
3034 | } else if (atomic_read(&nohz.load_balancer) == cpu) | |
3035 | return 1; | |
3036 | } else { | |
3037 | if (!cpu_isset(cpu, nohz.cpu_mask)) | |
3038 | return 0; | |
3039 | ||
3040 | cpu_clear(cpu, nohz.cpu_mask); | |
3041 | ||
3042 | if (atomic_read(&nohz.load_balancer) == cpu) | |
3043 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu) | |
3044 | BUG(); | |
3045 | } | |
3046 | return 0; | |
3047 | } | |
3048 | #endif | |
3049 | ||
3050 | static DEFINE_SPINLOCK(balancing); | |
3051 | ||
3052 | /* | |
7835b98b CL |
3053 | * It checks each scheduling domain to see if it is due to be balanced, |
3054 | * and initiates a balancing operation if so. | |
3055 | * | |
3056 | * Balancing parameters are set up in arch_init_sched_domains. | |
3057 | */ | |
46cb4b7c | 3058 | static inline void rebalance_domains(int cpu, enum idle_type idle) |
7835b98b | 3059 | { |
46cb4b7c SS |
3060 | int balance = 1; |
3061 | struct rq *rq = cpu_rq(cpu); | |
7835b98b CL |
3062 | unsigned long interval; |
3063 | struct sched_domain *sd; | |
46cb4b7c | 3064 | /* Earliest time when we have to do rebalance again */ |
c9819f45 | 3065 | unsigned long next_balance = jiffies + 60*HZ; |
1da177e4 | 3066 | |
46cb4b7c | 3067 | for_each_domain(cpu, sd) { |
1da177e4 LT |
3068 | if (!(sd->flags & SD_LOAD_BALANCE)) |
3069 | continue; | |
3070 | ||
3071 | interval = sd->balance_interval; | |
3072 | if (idle != SCHED_IDLE) | |
3073 | interval *= sd->busy_factor; | |
3074 | ||
3075 | /* scale ms to jiffies */ | |
3076 | interval = msecs_to_jiffies(interval); | |
3077 | if (unlikely(!interval)) | |
3078 | interval = 1; | |
3079 | ||
08c183f3 CL |
3080 | if (sd->flags & SD_SERIALIZE) { |
3081 | if (!spin_trylock(&balancing)) | |
3082 | goto out; | |
3083 | } | |
3084 | ||
c9819f45 | 3085 | if (time_after_eq(jiffies, sd->last_balance + interval)) { |
46cb4b7c | 3086 | if (load_balance(cpu, rq, sd, idle, &balance)) { |
fa3b6ddc SS |
3087 | /* |
3088 | * We've pulled tasks over so either we're no | |
5969fe06 NP |
3089 | * longer idle, or one of our SMT siblings is |
3090 | * not idle. | |
3091 | */ | |
1da177e4 LT |
3092 | idle = NOT_IDLE; |
3093 | } | |
1bd77f2d | 3094 | sd->last_balance = jiffies; |
1da177e4 | 3095 | } |
08c183f3 CL |
3096 | if (sd->flags & SD_SERIALIZE) |
3097 | spin_unlock(&balancing); | |
3098 | out: | |
c9819f45 CL |
3099 | if (time_after(next_balance, sd->last_balance + interval)) |
3100 | next_balance = sd->last_balance + interval; | |
783609c6 SS |
3101 | |
3102 | /* | |
3103 | * Stop the load balance at this level. There is another | |
3104 | * CPU in our sched group which is doing load balancing more | |
3105 | * actively. | |
3106 | */ | |
3107 | if (!balance) | |
3108 | break; | |
1da177e4 | 3109 | } |
46cb4b7c SS |
3110 | rq->next_balance = next_balance; |
3111 | } | |
3112 | ||
3113 | /* | |
3114 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
3115 | * In CONFIG_NO_HZ case, the idle load balance owner will do the | |
3116 | * rebalancing for all the cpus for whom scheduler ticks are stopped. | |
3117 | */ | |
3118 | static void run_rebalance_domains(struct softirq_action *h) | |
3119 | { | |
3120 | int local_cpu = smp_processor_id(); | |
3121 | struct rq *local_rq = cpu_rq(local_cpu); | |
3122 | enum idle_type idle = local_rq->idle_at_tick ? SCHED_IDLE : NOT_IDLE; | |
3123 | ||
3124 | rebalance_domains(local_cpu, idle); | |
3125 | ||
3126 | #ifdef CONFIG_NO_HZ | |
3127 | /* | |
3128 | * If this cpu is the owner for idle load balancing, then do the | |
3129 | * balancing on behalf of the other idle cpus whose ticks are | |
3130 | * stopped. | |
3131 | */ | |
3132 | if (local_rq->idle_at_tick && | |
3133 | atomic_read(&nohz.load_balancer) == local_cpu) { | |
3134 | cpumask_t cpus = nohz.cpu_mask; | |
3135 | struct rq *rq; | |
3136 | int balance_cpu; | |
3137 | ||
3138 | cpu_clear(local_cpu, cpus); | |
3139 | for_each_cpu_mask(balance_cpu, cpus) { | |
3140 | /* | |
3141 | * If this cpu gets work to do, stop the load balancing | |
3142 | * work being done for other cpus. Next load | |
3143 | * balancing owner will pick it up. | |
3144 | */ | |
3145 | if (need_resched()) | |
3146 | break; | |
3147 | ||
3148 | rebalance_domains(balance_cpu, SCHED_IDLE); | |
3149 | ||
3150 | rq = cpu_rq(balance_cpu); | |
3151 | if (time_after(local_rq->next_balance, rq->next_balance)) | |
3152 | local_rq->next_balance = rq->next_balance; | |
3153 | } | |
3154 | } | |
3155 | #endif | |
3156 | } | |
3157 | ||
3158 | /* | |
3159 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
3160 | * | |
3161 | * In case of CONFIG_NO_HZ, this is the place where we nominate a new | |
3162 | * idle load balancing owner or decide to stop the periodic load balancing, | |
3163 | * if the whole system is idle. | |
3164 | */ | |
3165 | static inline void trigger_load_balance(int cpu) | |
3166 | { | |
3167 | struct rq *rq = cpu_rq(cpu); | |
3168 | #ifdef CONFIG_NO_HZ | |
3169 | /* | |
3170 | * If we were in the nohz mode recently and busy at the current | |
3171 | * scheduler tick, then check if we need to nominate new idle | |
3172 | * load balancer. | |
3173 | */ | |
3174 | if (rq->in_nohz_recently && !rq->idle_at_tick) { | |
3175 | rq->in_nohz_recently = 0; | |
3176 | ||
3177 | if (atomic_read(&nohz.load_balancer) == cpu) { | |
3178 | cpu_clear(cpu, nohz.cpu_mask); | |
3179 | atomic_set(&nohz.load_balancer, -1); | |
3180 | } | |
3181 | ||
3182 | if (atomic_read(&nohz.load_balancer) == -1) { | |
3183 | /* | |
3184 | * simple selection for now: Nominate the | |
3185 | * first cpu in the nohz list to be the next | |
3186 | * ilb owner. | |
3187 | * | |
3188 | * TBD: Traverse the sched domains and nominate | |
3189 | * the nearest cpu in the nohz.cpu_mask. | |
3190 | */ | |
3191 | int ilb = first_cpu(nohz.cpu_mask); | |
3192 | ||
3193 | if (ilb != NR_CPUS) | |
3194 | resched_cpu(ilb); | |
3195 | } | |
3196 | } | |
3197 | ||
3198 | /* | |
3199 | * If this cpu is idle and doing idle load balancing for all the | |
3200 | * cpus with ticks stopped, is it time for that to stop? | |
3201 | */ | |
3202 | if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu && | |
3203 | cpus_weight(nohz.cpu_mask) == num_online_cpus()) { | |
3204 | resched_cpu(cpu); | |
3205 | return; | |
3206 | } | |
3207 | ||
3208 | /* | |
3209 | * If this cpu is idle and the idle load balancing is done by | |
3210 | * someone else, then no need raise the SCHED_SOFTIRQ | |
3211 | */ | |
3212 | if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu && | |
3213 | cpu_isset(cpu, nohz.cpu_mask)) | |
3214 | return; | |
3215 | #endif | |
3216 | if (time_after_eq(jiffies, rq->next_balance)) | |
3217 | raise_softirq(SCHED_SOFTIRQ); | |
1da177e4 LT |
3218 | } |
3219 | #else | |
3220 | /* | |
3221 | * on UP we do not need to balance between CPUs: | |
3222 | */ | |
70b97a7f | 3223 | static inline void idle_balance(int cpu, struct rq *rq) |
1da177e4 LT |
3224 | { |
3225 | } | |
3226 | #endif | |
3227 | ||
1da177e4 LT |
3228 | DEFINE_PER_CPU(struct kernel_stat, kstat); |
3229 | ||
3230 | EXPORT_PER_CPU_SYMBOL(kstat); | |
3231 | ||
3232 | /* | |
3233 | * This is called on clock ticks and on context switches. | |
3234 | * Bank in p->sched_time the ns elapsed since the last tick or switch. | |
3235 | */ | |
48f24c4d | 3236 | static inline void |
70b97a7f | 3237 | update_cpu_clock(struct task_struct *p, struct rq *rq, unsigned long long now) |
1da177e4 | 3238 | { |
b18ec803 MG |
3239 | p->sched_time += now - p->last_ran; |
3240 | p->last_ran = rq->most_recent_timestamp = now; | |
1da177e4 LT |
3241 | } |
3242 | ||
3243 | /* | |
3244 | * Return current->sched_time plus any more ns on the sched_clock | |
3245 | * that have not yet been banked. | |
3246 | */ | |
36c8b586 | 3247 | unsigned long long current_sched_time(const struct task_struct *p) |
1da177e4 LT |
3248 | { |
3249 | unsigned long long ns; | |
3250 | unsigned long flags; | |
48f24c4d | 3251 | |
1da177e4 | 3252 | local_irq_save(flags); |
b18ec803 | 3253 | ns = p->sched_time + sched_clock() - p->last_ran; |
1da177e4 | 3254 | local_irq_restore(flags); |
48f24c4d | 3255 | |
1da177e4 LT |
3256 | return ns; |
3257 | } | |
3258 | ||
f1adad78 LT |
3259 | /* |
3260 | * We place interactive tasks back into the active array, if possible. | |
3261 | * | |
3262 | * To guarantee that this does not starve expired tasks we ignore the | |
3263 | * interactivity of a task if the first expired task had to wait more | |
3264 | * than a 'reasonable' amount of time. This deadline timeout is | |
3265 | * load-dependent, as the frequency of array switched decreases with | |
3266 | * increasing number of running tasks. We also ignore the interactivity | |
3267 | * if a better static_prio task has expired: | |
3268 | */ | |
70b97a7f | 3269 | static inline int expired_starving(struct rq *rq) |
48f24c4d IM |
3270 | { |
3271 | if (rq->curr->static_prio > rq->best_expired_prio) | |
3272 | return 1; | |
3273 | if (!STARVATION_LIMIT || !rq->expired_timestamp) | |
3274 | return 0; | |
3275 | if (jiffies - rq->expired_timestamp > STARVATION_LIMIT * rq->nr_running) | |
3276 | return 1; | |
3277 | return 0; | |
3278 | } | |
f1adad78 | 3279 | |
1da177e4 LT |
3280 | /* |
3281 | * Account user cpu time to a process. | |
3282 | * @p: the process that the cpu time gets accounted to | |
3283 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
3284 | * @cputime: the cpu time spent in user space since the last update | |
3285 | */ | |
3286 | void account_user_time(struct task_struct *p, cputime_t cputime) | |
3287 | { | |
3288 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
3289 | cputime64_t tmp; | |
3290 | ||
3291 | p->utime = cputime_add(p->utime, cputime); | |
3292 | ||
3293 | /* Add user time to cpustat. */ | |
3294 | tmp = cputime_to_cputime64(cputime); | |
3295 | if (TASK_NICE(p) > 0) | |
3296 | cpustat->nice = cputime64_add(cpustat->nice, tmp); | |
3297 | else | |
3298 | cpustat->user = cputime64_add(cpustat->user, tmp); | |
3299 | } | |
3300 | ||
3301 | /* | |
3302 | * Account system cpu time to a process. | |
3303 | * @p: the process that the cpu time gets accounted to | |
3304 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
3305 | * @cputime: the cpu time spent in kernel space since the last update | |
3306 | */ | |
3307 | void account_system_time(struct task_struct *p, int hardirq_offset, | |
3308 | cputime_t cputime) | |
3309 | { | |
3310 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
70b97a7f | 3311 | struct rq *rq = this_rq(); |
1da177e4 LT |
3312 | cputime64_t tmp; |
3313 | ||
3314 | p->stime = cputime_add(p->stime, cputime); | |
3315 | ||
3316 | /* Add system time to cpustat. */ | |
3317 | tmp = cputime_to_cputime64(cputime); | |
3318 | if (hardirq_count() - hardirq_offset) | |
3319 | cpustat->irq = cputime64_add(cpustat->irq, tmp); | |
3320 | else if (softirq_count()) | |
3321 | cpustat->softirq = cputime64_add(cpustat->softirq, tmp); | |
3322 | else if (p != rq->idle) | |
3323 | cpustat->system = cputime64_add(cpustat->system, tmp); | |
3324 | else if (atomic_read(&rq->nr_iowait) > 0) | |
3325 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); | |
3326 | else | |
3327 | cpustat->idle = cputime64_add(cpustat->idle, tmp); | |
3328 | /* Account for system time used */ | |
3329 | acct_update_integrals(p); | |
1da177e4 LT |
3330 | } |
3331 | ||
3332 | /* | |
3333 | * Account for involuntary wait time. | |
3334 | * @p: the process from which the cpu time has been stolen | |
3335 | * @steal: the cpu time spent in involuntary wait | |
3336 | */ | |
3337 | void account_steal_time(struct task_struct *p, cputime_t steal) | |
3338 | { | |
3339 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
3340 | cputime64_t tmp = cputime_to_cputime64(steal); | |
70b97a7f | 3341 | struct rq *rq = this_rq(); |
1da177e4 LT |
3342 | |
3343 | if (p == rq->idle) { | |
3344 | p->stime = cputime_add(p->stime, steal); | |
3345 | if (atomic_read(&rq->nr_iowait) > 0) | |
3346 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); | |
3347 | else | |
3348 | cpustat->idle = cputime64_add(cpustat->idle, tmp); | |
3349 | } else | |
3350 | cpustat->steal = cputime64_add(cpustat->steal, tmp); | |
3351 | } | |
3352 | ||
7835b98b | 3353 | static void task_running_tick(struct rq *rq, struct task_struct *p) |
1da177e4 | 3354 | { |
1da177e4 | 3355 | if (p->array != rq->active) { |
7835b98b | 3356 | /* Task has expired but was not scheduled yet */ |
1da177e4 | 3357 | set_tsk_need_resched(p); |
7835b98b | 3358 | return; |
1da177e4 LT |
3359 | } |
3360 | spin_lock(&rq->lock); | |
3361 | /* | |
3362 | * The task was running during this tick - update the | |
3363 | * time slice counter. Note: we do not update a thread's | |
3364 | * priority until it either goes to sleep or uses up its | |
3365 | * timeslice. This makes it possible for interactive tasks | |
3366 | * to use up their timeslices at their highest priority levels. | |
3367 | */ | |
3368 | if (rt_task(p)) { | |
3369 | /* | |
3370 | * RR tasks need a special form of timeslice management. | |
3371 | * FIFO tasks have no timeslices. | |
3372 | */ | |
3373 | if ((p->policy == SCHED_RR) && !--p->time_slice) { | |
3374 | p->time_slice = task_timeslice(p); | |
3375 | p->first_time_slice = 0; | |
3376 | set_tsk_need_resched(p); | |
3377 | ||
3378 | /* put it at the end of the queue: */ | |
3379 | requeue_task(p, rq->active); | |
3380 | } | |
3381 | goto out_unlock; | |
3382 | } | |
3383 | if (!--p->time_slice) { | |
3384 | dequeue_task(p, rq->active); | |
3385 | set_tsk_need_resched(p); | |
3386 | p->prio = effective_prio(p); | |
3387 | p->time_slice = task_timeslice(p); | |
3388 | p->first_time_slice = 0; | |
3389 | ||
3390 | if (!rq->expired_timestamp) | |
3391 | rq->expired_timestamp = jiffies; | |
48f24c4d | 3392 | if (!TASK_INTERACTIVE(p) || expired_starving(rq)) { |
1da177e4 LT |
3393 | enqueue_task(p, rq->expired); |
3394 | if (p->static_prio < rq->best_expired_prio) | |
3395 | rq->best_expired_prio = p->static_prio; | |
3396 | } else | |
3397 | enqueue_task(p, rq->active); | |
3398 | } else { | |
3399 | /* | |
3400 | * Prevent a too long timeslice allowing a task to monopolize | |
3401 | * the CPU. We do this by splitting up the timeslice into | |
3402 | * smaller pieces. | |
3403 | * | |
3404 | * Note: this does not mean the task's timeslices expire or | |
3405 | * get lost in any way, they just might be preempted by | |
3406 | * another task of equal priority. (one with higher | |
3407 | * priority would have preempted this task already.) We | |
3408 | * requeue this task to the end of the list on this priority | |
3409 | * level, which is in essence a round-robin of tasks with | |
3410 | * equal priority. | |
3411 | * | |
3412 | * This only applies to tasks in the interactive | |
3413 | * delta range with at least TIMESLICE_GRANULARITY to requeue. | |
3414 | */ | |
3415 | if (TASK_INTERACTIVE(p) && !((task_timeslice(p) - | |
3416 | p->time_slice) % TIMESLICE_GRANULARITY(p)) && | |
3417 | (p->time_slice >= TIMESLICE_GRANULARITY(p)) && | |
3418 | (p->array == rq->active)) { | |
3419 | ||
3420 | requeue_task(p, rq->active); | |
3421 | set_tsk_need_resched(p); | |
3422 | } | |
3423 | } | |
3424 | out_unlock: | |
3425 | spin_unlock(&rq->lock); | |
7835b98b CL |
3426 | } |
3427 | ||
3428 | /* | |
3429 | * This function gets called by the timer code, with HZ frequency. | |
3430 | * We call it with interrupts disabled. | |
3431 | * | |
3432 | * It also gets called by the fork code, when changing the parent's | |
3433 | * timeslices. | |
3434 | */ | |
3435 | void scheduler_tick(void) | |
3436 | { | |
3437 | unsigned long long now = sched_clock(); | |
3438 | struct task_struct *p = current; | |
3439 | int cpu = smp_processor_id(); | |
bdecea3a | 3440 | int idle_at_tick = idle_cpu(cpu); |
7835b98b | 3441 | struct rq *rq = cpu_rq(cpu); |
7835b98b CL |
3442 | |
3443 | update_cpu_clock(p, rq, now); | |
3444 | ||
bdecea3a | 3445 | if (!idle_at_tick) |
7835b98b | 3446 | task_running_tick(rq, p); |
e418e1c2 | 3447 | #ifdef CONFIG_SMP |
7835b98b | 3448 | update_load(rq); |
bdecea3a | 3449 | rq->idle_at_tick = idle_at_tick; |
46cb4b7c | 3450 | trigger_load_balance(cpu); |
e418e1c2 | 3451 | #endif |
1da177e4 LT |
3452 | } |
3453 | ||
1da177e4 LT |
3454 | #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT) |
3455 | ||
3456 | void fastcall add_preempt_count(int val) | |
3457 | { | |
3458 | /* | |
3459 | * Underflow? | |
3460 | */ | |
9a11b49a IM |
3461 | if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) |
3462 | return; | |
1da177e4 LT |
3463 | preempt_count() += val; |
3464 | /* | |
3465 | * Spinlock count overflowing soon? | |
3466 | */ | |
33859f7f MOS |
3467 | DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= |
3468 | PREEMPT_MASK - 10); | |
1da177e4 LT |
3469 | } |
3470 | EXPORT_SYMBOL(add_preempt_count); | |
3471 | ||
3472 | void fastcall sub_preempt_count(int val) | |
3473 | { | |
3474 | /* | |
3475 | * Underflow? | |
3476 | */ | |
9a11b49a IM |
3477 | if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) |
3478 | return; | |
1da177e4 LT |
3479 | /* |
3480 | * Is the spinlock portion underflowing? | |
3481 | */ | |
9a11b49a IM |
3482 | if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && |
3483 | !(preempt_count() & PREEMPT_MASK))) | |
3484 | return; | |
3485 | ||
1da177e4 LT |
3486 | preempt_count() -= val; |
3487 | } | |
3488 | EXPORT_SYMBOL(sub_preempt_count); | |
3489 | ||
3490 | #endif | |
3491 | ||
3dee386e CK |
3492 | static inline int interactive_sleep(enum sleep_type sleep_type) |
3493 | { | |
3494 | return (sleep_type == SLEEP_INTERACTIVE || | |
3495 | sleep_type == SLEEP_INTERRUPTED); | |
3496 | } | |
3497 | ||
1da177e4 LT |
3498 | /* |
3499 | * schedule() is the main scheduler function. | |
3500 | */ | |
3501 | asmlinkage void __sched schedule(void) | |
3502 | { | |
36c8b586 | 3503 | struct task_struct *prev, *next; |
70b97a7f | 3504 | struct prio_array *array; |
1da177e4 LT |
3505 | struct list_head *queue; |
3506 | unsigned long long now; | |
3507 | unsigned long run_time; | |
a3464a10 | 3508 | int cpu, idx, new_prio; |
48f24c4d | 3509 | long *switch_count; |
70b97a7f | 3510 | struct rq *rq; |
1da177e4 LT |
3511 | |
3512 | /* | |
3513 | * Test if we are atomic. Since do_exit() needs to call into | |
3514 | * schedule() atomically, we ignore that path for now. | |
3515 | * Otherwise, whine if we are scheduling when we should not be. | |
3516 | */ | |
77e4bfbc AM |
3517 | if (unlikely(in_atomic() && !current->exit_state)) { |
3518 | printk(KERN_ERR "BUG: scheduling while atomic: " | |
3519 | "%s/0x%08x/%d\n", | |
3520 | current->comm, preempt_count(), current->pid); | |
a4c410f0 | 3521 | debug_show_held_locks(current); |
3117df04 IM |
3522 | if (irqs_disabled()) |
3523 | print_irqtrace_events(current); | |
77e4bfbc | 3524 | dump_stack(); |
1da177e4 LT |
3525 | } |
3526 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); | |
3527 | ||
3528 | need_resched: | |
3529 | preempt_disable(); | |
3530 | prev = current; | |
3531 | release_kernel_lock(prev); | |
3532 | need_resched_nonpreemptible: | |
3533 | rq = this_rq(); | |
3534 | ||
3535 | /* | |
3536 | * The idle thread is not allowed to schedule! | |
3537 | * Remove this check after it has been exercised a bit. | |
3538 | */ | |
3539 | if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) { | |
3540 | printk(KERN_ERR "bad: scheduling from the idle thread!\n"); | |
3541 | dump_stack(); | |
3542 | } | |
3543 | ||
3544 | schedstat_inc(rq, sched_cnt); | |
3545 | now = sched_clock(); | |
238628ed | 3546 | if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) { |
1da177e4 | 3547 | run_time = now - prev->timestamp; |
238628ed | 3548 | if (unlikely((long long)(now - prev->timestamp) < 0)) |
1da177e4 LT |
3549 | run_time = 0; |
3550 | } else | |
3551 | run_time = NS_MAX_SLEEP_AVG; | |
3552 | ||
3553 | /* | |
3554 | * Tasks charged proportionately less run_time at high sleep_avg to | |
3555 | * delay them losing their interactive status | |
3556 | */ | |
3557 | run_time /= (CURRENT_BONUS(prev) ? : 1); | |
3558 | ||
3559 | spin_lock_irq(&rq->lock); | |
3560 | ||
1da177e4 LT |
3561 | switch_count = &prev->nivcsw; |
3562 | if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { | |
3563 | switch_count = &prev->nvcsw; | |
3564 | if (unlikely((prev->state & TASK_INTERRUPTIBLE) && | |
3565 | unlikely(signal_pending(prev)))) | |
3566 | prev->state = TASK_RUNNING; | |
3567 | else { | |
3568 | if (prev->state == TASK_UNINTERRUPTIBLE) | |
3569 | rq->nr_uninterruptible++; | |
3570 | deactivate_task(prev, rq); | |
3571 | } | |
3572 | } | |
3573 | ||
3574 | cpu = smp_processor_id(); | |
3575 | if (unlikely(!rq->nr_running)) { | |
1da177e4 LT |
3576 | idle_balance(cpu, rq); |
3577 | if (!rq->nr_running) { | |
3578 | next = rq->idle; | |
3579 | rq->expired_timestamp = 0; | |
1da177e4 LT |
3580 | goto switch_tasks; |
3581 | } | |
1da177e4 LT |
3582 | } |
3583 | ||
3584 | array = rq->active; | |
3585 | if (unlikely(!array->nr_active)) { | |
3586 | /* | |
3587 | * Switch the active and expired arrays. | |
3588 | */ | |
3589 | schedstat_inc(rq, sched_switch); | |
3590 | rq->active = rq->expired; | |
3591 | rq->expired = array; | |
3592 | array = rq->active; | |
3593 | rq->expired_timestamp = 0; | |
3594 | rq->best_expired_prio = MAX_PRIO; | |
3595 | } | |
3596 | ||
3597 | idx = sched_find_first_bit(array->bitmap); | |
3598 | queue = array->queue + idx; | |
36c8b586 | 3599 | next = list_entry(queue->next, struct task_struct, run_list); |
1da177e4 | 3600 | |
3dee386e | 3601 | if (!rt_task(next) && interactive_sleep(next->sleep_type)) { |
1da177e4 | 3602 | unsigned long long delta = now - next->timestamp; |
238628ed | 3603 | if (unlikely((long long)(now - next->timestamp) < 0)) |
1da177e4 LT |
3604 | delta = 0; |
3605 | ||
3dee386e | 3606 | if (next->sleep_type == SLEEP_INTERACTIVE) |
1da177e4 LT |
3607 | delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128; |
3608 | ||
3609 | array = next->array; | |
a3464a10 CS |
3610 | new_prio = recalc_task_prio(next, next->timestamp + delta); |
3611 | ||
3612 | if (unlikely(next->prio != new_prio)) { | |
3613 | dequeue_task(next, array); | |
3614 | next->prio = new_prio; | |
3615 | enqueue_task(next, array); | |
7c4bb1f9 | 3616 | } |
1da177e4 | 3617 | } |
3dee386e | 3618 | next->sleep_type = SLEEP_NORMAL; |
1da177e4 LT |
3619 | switch_tasks: |
3620 | if (next == rq->idle) | |
3621 | schedstat_inc(rq, sched_goidle); | |
3622 | prefetch(next); | |
383f2835 | 3623 | prefetch_stack(next); |
1da177e4 LT |
3624 | clear_tsk_need_resched(prev); |
3625 | rcu_qsctr_inc(task_cpu(prev)); | |
3626 | ||
3627 | update_cpu_clock(prev, rq, now); | |
3628 | ||
3629 | prev->sleep_avg -= run_time; | |
3630 | if ((long)prev->sleep_avg <= 0) | |
3631 | prev->sleep_avg = 0; | |
3632 | prev->timestamp = prev->last_ran = now; | |
3633 | ||
3634 | sched_info_switch(prev, next); | |
3635 | if (likely(prev != next)) { | |
c1e16aa2 | 3636 | next->timestamp = next->last_ran = now; |
1da177e4 LT |
3637 | rq->nr_switches++; |
3638 | rq->curr = next; | |
3639 | ++*switch_count; | |
3640 | ||
4866cde0 | 3641 | prepare_task_switch(rq, next); |
1da177e4 LT |
3642 | prev = context_switch(rq, prev, next); |
3643 | barrier(); | |
4866cde0 NP |
3644 | /* |
3645 | * this_rq must be evaluated again because prev may have moved | |
3646 | * CPUs since it called schedule(), thus the 'rq' on its stack | |
3647 | * frame will be invalid. | |
3648 | */ | |
3649 | finish_task_switch(this_rq(), prev); | |
1da177e4 LT |
3650 | } else |
3651 | spin_unlock_irq(&rq->lock); | |
3652 | ||
3653 | prev = current; | |
3654 | if (unlikely(reacquire_kernel_lock(prev) < 0)) | |
3655 | goto need_resched_nonpreemptible; | |
3656 | preempt_enable_no_resched(); | |
3657 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
3658 | goto need_resched; | |
3659 | } | |
1da177e4 LT |
3660 | EXPORT_SYMBOL(schedule); |
3661 | ||
3662 | #ifdef CONFIG_PREEMPT | |
3663 | /* | |
2ed6e34f | 3664 | * this is the entry point to schedule() from in-kernel preemption |
1da177e4 LT |
3665 | * off of preempt_enable. Kernel preemptions off return from interrupt |
3666 | * occur there and call schedule directly. | |
3667 | */ | |
3668 | asmlinkage void __sched preempt_schedule(void) | |
3669 | { | |
3670 | struct thread_info *ti = current_thread_info(); | |
3671 | #ifdef CONFIG_PREEMPT_BKL | |
3672 | struct task_struct *task = current; | |
3673 | int saved_lock_depth; | |
3674 | #endif | |
3675 | /* | |
3676 | * If there is a non-zero preempt_count or interrupts are disabled, | |
3677 | * we do not want to preempt the current task. Just return.. | |
3678 | */ | |
beed33a8 | 3679 | if (likely(ti->preempt_count || irqs_disabled())) |
1da177e4 LT |
3680 | return; |
3681 | ||
3682 | need_resched: | |
3683 | add_preempt_count(PREEMPT_ACTIVE); | |
3684 | /* | |
3685 | * We keep the big kernel semaphore locked, but we | |
3686 | * clear ->lock_depth so that schedule() doesnt | |
3687 | * auto-release the semaphore: | |
3688 | */ | |
3689 | #ifdef CONFIG_PREEMPT_BKL | |
3690 | saved_lock_depth = task->lock_depth; | |
3691 | task->lock_depth = -1; | |
3692 | #endif | |
3693 | schedule(); | |
3694 | #ifdef CONFIG_PREEMPT_BKL | |
3695 | task->lock_depth = saved_lock_depth; | |
3696 | #endif | |
3697 | sub_preempt_count(PREEMPT_ACTIVE); | |
3698 | ||
3699 | /* we could miss a preemption opportunity between schedule and now */ | |
3700 | barrier(); | |
3701 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
3702 | goto need_resched; | |
3703 | } | |
1da177e4 LT |
3704 | EXPORT_SYMBOL(preempt_schedule); |
3705 | ||
3706 | /* | |
2ed6e34f | 3707 | * this is the entry point to schedule() from kernel preemption |
1da177e4 LT |
3708 | * off of irq context. |
3709 | * Note, that this is called and return with irqs disabled. This will | |
3710 | * protect us against recursive calling from irq. | |
3711 | */ | |
3712 | asmlinkage void __sched preempt_schedule_irq(void) | |
3713 | { | |
3714 | struct thread_info *ti = current_thread_info(); | |
3715 | #ifdef CONFIG_PREEMPT_BKL | |
3716 | struct task_struct *task = current; | |
3717 | int saved_lock_depth; | |
3718 | #endif | |
2ed6e34f | 3719 | /* Catch callers which need to be fixed */ |
1da177e4 LT |
3720 | BUG_ON(ti->preempt_count || !irqs_disabled()); |
3721 | ||
3722 | need_resched: | |
3723 | add_preempt_count(PREEMPT_ACTIVE); | |
3724 | /* | |
3725 | * We keep the big kernel semaphore locked, but we | |
3726 | * clear ->lock_depth so that schedule() doesnt | |
3727 | * auto-release the semaphore: | |
3728 | */ | |
3729 | #ifdef CONFIG_PREEMPT_BKL | |
3730 | saved_lock_depth = task->lock_depth; | |
3731 | task->lock_depth = -1; | |
3732 | #endif | |
3733 | local_irq_enable(); | |
3734 | schedule(); | |
3735 | local_irq_disable(); | |
3736 | #ifdef CONFIG_PREEMPT_BKL | |
3737 | task->lock_depth = saved_lock_depth; | |
3738 | #endif | |
3739 | sub_preempt_count(PREEMPT_ACTIVE); | |
3740 | ||
3741 | /* we could miss a preemption opportunity between schedule and now */ | |
3742 | barrier(); | |
3743 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
3744 | goto need_resched; | |
3745 | } | |
3746 | ||
3747 | #endif /* CONFIG_PREEMPT */ | |
3748 | ||
95cdf3b7 IM |
3749 | int default_wake_function(wait_queue_t *curr, unsigned mode, int sync, |
3750 | void *key) | |
1da177e4 | 3751 | { |
48f24c4d | 3752 | return try_to_wake_up(curr->private, mode, sync); |
1da177e4 | 3753 | } |
1da177e4 LT |
3754 | EXPORT_SYMBOL(default_wake_function); |
3755 | ||
3756 | /* | |
3757 | * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just | |
3758 | * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve | |
3759 | * number) then we wake all the non-exclusive tasks and one exclusive task. | |
3760 | * | |
3761 | * There are circumstances in which we can try to wake a task which has already | |
3762 | * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns | |
3763 | * zero in this (rare) case, and we handle it by continuing to scan the queue. | |
3764 | */ | |
3765 | static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, | |
3766 | int nr_exclusive, int sync, void *key) | |
3767 | { | |
3768 | struct list_head *tmp, *next; | |
3769 | ||
3770 | list_for_each_safe(tmp, next, &q->task_list) { | |
48f24c4d IM |
3771 | wait_queue_t *curr = list_entry(tmp, wait_queue_t, task_list); |
3772 | unsigned flags = curr->flags; | |
3773 | ||
1da177e4 | 3774 | if (curr->func(curr, mode, sync, key) && |
48f24c4d | 3775 | (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) |
1da177e4 LT |
3776 | break; |
3777 | } | |
3778 | } | |
3779 | ||
3780 | /** | |
3781 | * __wake_up - wake up threads blocked on a waitqueue. | |
3782 | * @q: the waitqueue | |
3783 | * @mode: which threads | |
3784 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
67be2dd1 | 3785 | * @key: is directly passed to the wakeup function |
1da177e4 LT |
3786 | */ |
3787 | void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode, | |
95cdf3b7 | 3788 | int nr_exclusive, void *key) |
1da177e4 LT |
3789 | { |
3790 | unsigned long flags; | |
3791 | ||
3792 | spin_lock_irqsave(&q->lock, flags); | |
3793 | __wake_up_common(q, mode, nr_exclusive, 0, key); | |
3794 | spin_unlock_irqrestore(&q->lock, flags); | |
3795 | } | |
1da177e4 LT |
3796 | EXPORT_SYMBOL(__wake_up); |
3797 | ||
3798 | /* | |
3799 | * Same as __wake_up but called with the spinlock in wait_queue_head_t held. | |
3800 | */ | |
3801 | void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode) | |
3802 | { | |
3803 | __wake_up_common(q, mode, 1, 0, NULL); | |
3804 | } | |
3805 | ||
3806 | /** | |
67be2dd1 | 3807 | * __wake_up_sync - wake up threads blocked on a waitqueue. |
1da177e4 LT |
3808 | * @q: the waitqueue |
3809 | * @mode: which threads | |
3810 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
3811 | * | |
3812 | * The sync wakeup differs that the waker knows that it will schedule | |
3813 | * away soon, so while the target thread will be woken up, it will not | |
3814 | * be migrated to another CPU - ie. the two threads are 'synchronized' | |
3815 | * with each other. This can prevent needless bouncing between CPUs. | |
3816 | * | |
3817 | * On UP it can prevent extra preemption. | |
3818 | */ | |
95cdf3b7 IM |
3819 | void fastcall |
3820 | __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) | |
1da177e4 LT |
3821 | { |
3822 | unsigned long flags; | |
3823 | int sync = 1; | |
3824 | ||
3825 | if (unlikely(!q)) | |
3826 | return; | |
3827 | ||
3828 | if (unlikely(!nr_exclusive)) | |
3829 | sync = 0; | |
3830 | ||
3831 | spin_lock_irqsave(&q->lock, flags); | |
3832 | __wake_up_common(q, mode, nr_exclusive, sync, NULL); | |
3833 | spin_unlock_irqrestore(&q->lock, flags); | |
3834 | } | |
3835 | EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ | |
3836 | ||
3837 | void fastcall complete(struct completion *x) | |
3838 | { | |
3839 | unsigned long flags; | |
3840 | ||
3841 | spin_lock_irqsave(&x->wait.lock, flags); | |
3842 | x->done++; | |
3843 | __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, | |
3844 | 1, 0, NULL); | |
3845 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
3846 | } | |
3847 | EXPORT_SYMBOL(complete); | |
3848 | ||
3849 | void fastcall complete_all(struct completion *x) | |
3850 | { | |
3851 | unsigned long flags; | |
3852 | ||
3853 | spin_lock_irqsave(&x->wait.lock, flags); | |
3854 | x->done += UINT_MAX/2; | |
3855 | __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, | |
3856 | 0, 0, NULL); | |
3857 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
3858 | } | |
3859 | EXPORT_SYMBOL(complete_all); | |
3860 | ||
3861 | void fastcall __sched wait_for_completion(struct completion *x) | |
3862 | { | |
3863 | might_sleep(); | |
48f24c4d | 3864 | |
1da177e4 LT |
3865 | spin_lock_irq(&x->wait.lock); |
3866 | if (!x->done) { | |
3867 | DECLARE_WAITQUEUE(wait, current); | |
3868 | ||
3869 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3870 | __add_wait_queue_tail(&x->wait, &wait); | |
3871 | do { | |
3872 | __set_current_state(TASK_UNINTERRUPTIBLE); | |
3873 | spin_unlock_irq(&x->wait.lock); | |
3874 | schedule(); | |
3875 | spin_lock_irq(&x->wait.lock); | |
3876 | } while (!x->done); | |
3877 | __remove_wait_queue(&x->wait, &wait); | |
3878 | } | |
3879 | x->done--; | |
3880 | spin_unlock_irq(&x->wait.lock); | |
3881 | } | |
3882 | EXPORT_SYMBOL(wait_for_completion); | |
3883 | ||
3884 | unsigned long fastcall __sched | |
3885 | wait_for_completion_timeout(struct completion *x, unsigned long timeout) | |
3886 | { | |
3887 | might_sleep(); | |
3888 | ||
3889 | spin_lock_irq(&x->wait.lock); | |
3890 | if (!x->done) { | |
3891 | DECLARE_WAITQUEUE(wait, current); | |
3892 | ||
3893 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3894 | __add_wait_queue_tail(&x->wait, &wait); | |
3895 | do { | |
3896 | __set_current_state(TASK_UNINTERRUPTIBLE); | |
3897 | spin_unlock_irq(&x->wait.lock); | |
3898 | timeout = schedule_timeout(timeout); | |
3899 | spin_lock_irq(&x->wait.lock); | |
3900 | if (!timeout) { | |
3901 | __remove_wait_queue(&x->wait, &wait); | |
3902 | goto out; | |
3903 | } | |
3904 | } while (!x->done); | |
3905 | __remove_wait_queue(&x->wait, &wait); | |
3906 | } | |
3907 | x->done--; | |
3908 | out: | |
3909 | spin_unlock_irq(&x->wait.lock); | |
3910 | return timeout; | |
3911 | } | |
3912 | EXPORT_SYMBOL(wait_for_completion_timeout); | |
3913 | ||
3914 | int fastcall __sched wait_for_completion_interruptible(struct completion *x) | |
3915 | { | |
3916 | int ret = 0; | |
3917 | ||
3918 | might_sleep(); | |
3919 | ||
3920 | spin_lock_irq(&x->wait.lock); | |
3921 | if (!x->done) { | |
3922 | DECLARE_WAITQUEUE(wait, current); | |
3923 | ||
3924 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3925 | __add_wait_queue_tail(&x->wait, &wait); | |
3926 | do { | |
3927 | if (signal_pending(current)) { | |
3928 | ret = -ERESTARTSYS; | |
3929 | __remove_wait_queue(&x->wait, &wait); | |
3930 | goto out; | |
3931 | } | |
3932 | __set_current_state(TASK_INTERRUPTIBLE); | |
3933 | spin_unlock_irq(&x->wait.lock); | |
3934 | schedule(); | |
3935 | spin_lock_irq(&x->wait.lock); | |
3936 | } while (!x->done); | |
3937 | __remove_wait_queue(&x->wait, &wait); | |
3938 | } | |
3939 | x->done--; | |
3940 | out: | |
3941 | spin_unlock_irq(&x->wait.lock); | |
3942 | ||
3943 | return ret; | |
3944 | } | |
3945 | EXPORT_SYMBOL(wait_for_completion_interruptible); | |
3946 | ||
3947 | unsigned long fastcall __sched | |
3948 | wait_for_completion_interruptible_timeout(struct completion *x, | |
3949 | unsigned long timeout) | |
3950 | { | |
3951 | might_sleep(); | |
3952 | ||
3953 | spin_lock_irq(&x->wait.lock); | |
3954 | if (!x->done) { | |
3955 | DECLARE_WAITQUEUE(wait, current); | |
3956 | ||
3957 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
3958 | __add_wait_queue_tail(&x->wait, &wait); | |
3959 | do { | |
3960 | if (signal_pending(current)) { | |
3961 | timeout = -ERESTARTSYS; | |
3962 | __remove_wait_queue(&x->wait, &wait); | |
3963 | goto out; | |
3964 | } | |
3965 | __set_current_state(TASK_INTERRUPTIBLE); | |
3966 | spin_unlock_irq(&x->wait.lock); | |
3967 | timeout = schedule_timeout(timeout); | |
3968 | spin_lock_irq(&x->wait.lock); | |
3969 | if (!timeout) { | |
3970 | __remove_wait_queue(&x->wait, &wait); | |
3971 | goto out; | |
3972 | } | |
3973 | } while (!x->done); | |
3974 | __remove_wait_queue(&x->wait, &wait); | |
3975 | } | |
3976 | x->done--; | |
3977 | out: | |
3978 | spin_unlock_irq(&x->wait.lock); | |
3979 | return timeout; | |
3980 | } | |
3981 | EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); | |
3982 | ||
3983 | ||
3984 | #define SLEEP_ON_VAR \ | |
3985 | unsigned long flags; \ | |
3986 | wait_queue_t wait; \ | |
3987 | init_waitqueue_entry(&wait, current); | |
3988 | ||
3989 | #define SLEEP_ON_HEAD \ | |
3990 | spin_lock_irqsave(&q->lock,flags); \ | |
3991 | __add_wait_queue(q, &wait); \ | |
3992 | spin_unlock(&q->lock); | |
3993 | ||
3994 | #define SLEEP_ON_TAIL \ | |
3995 | spin_lock_irq(&q->lock); \ | |
3996 | __remove_wait_queue(q, &wait); \ | |
3997 | spin_unlock_irqrestore(&q->lock, flags); | |
3998 | ||
3999 | void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q) | |
4000 | { | |
4001 | SLEEP_ON_VAR | |
4002 | ||
4003 | current->state = TASK_INTERRUPTIBLE; | |
4004 | ||
4005 | SLEEP_ON_HEAD | |
4006 | schedule(); | |
4007 | SLEEP_ON_TAIL | |
4008 | } | |
1da177e4 LT |
4009 | EXPORT_SYMBOL(interruptible_sleep_on); |
4010 | ||
95cdf3b7 IM |
4011 | long fastcall __sched |
4012 | interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) | |
1da177e4 LT |
4013 | { |
4014 | SLEEP_ON_VAR | |
4015 | ||
4016 | current->state = TASK_INTERRUPTIBLE; | |
4017 | ||
4018 | SLEEP_ON_HEAD | |
4019 | timeout = schedule_timeout(timeout); | |
4020 | SLEEP_ON_TAIL | |
4021 | ||
4022 | return timeout; | |
4023 | } | |
1da177e4 LT |
4024 | EXPORT_SYMBOL(interruptible_sleep_on_timeout); |
4025 | ||
4026 | void fastcall __sched sleep_on(wait_queue_head_t *q) | |
4027 | { | |
4028 | SLEEP_ON_VAR | |
4029 | ||
4030 | current->state = TASK_UNINTERRUPTIBLE; | |
4031 | ||
4032 | SLEEP_ON_HEAD | |
4033 | schedule(); | |
4034 | SLEEP_ON_TAIL | |
4035 | } | |
1da177e4 LT |
4036 | EXPORT_SYMBOL(sleep_on); |
4037 | ||
4038 | long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) | |
4039 | { | |
4040 | SLEEP_ON_VAR | |
4041 | ||
4042 | current->state = TASK_UNINTERRUPTIBLE; | |
4043 | ||
4044 | SLEEP_ON_HEAD | |
4045 | timeout = schedule_timeout(timeout); | |
4046 | SLEEP_ON_TAIL | |
4047 | ||
4048 | return timeout; | |
4049 | } | |
4050 | ||
4051 | EXPORT_SYMBOL(sleep_on_timeout); | |
4052 | ||
b29739f9 IM |
4053 | #ifdef CONFIG_RT_MUTEXES |
4054 | ||
4055 | /* | |
4056 | * rt_mutex_setprio - set the current priority of a task | |
4057 | * @p: task | |
4058 | * @prio: prio value (kernel-internal form) | |
4059 | * | |
4060 | * This function changes the 'effective' priority of a task. It does | |
4061 | * not touch ->normal_prio like __setscheduler(). | |
4062 | * | |
4063 | * Used by the rt_mutex code to implement priority inheritance logic. | |
4064 | */ | |
36c8b586 | 4065 | void rt_mutex_setprio(struct task_struct *p, int prio) |
b29739f9 | 4066 | { |
70b97a7f | 4067 | struct prio_array *array; |
b29739f9 | 4068 | unsigned long flags; |
70b97a7f | 4069 | struct rq *rq; |
b29739f9 IM |
4070 | int oldprio; |
4071 | ||
4072 | BUG_ON(prio < 0 || prio > MAX_PRIO); | |
4073 | ||
4074 | rq = task_rq_lock(p, &flags); | |
4075 | ||
4076 | oldprio = p->prio; | |
4077 | array = p->array; | |
4078 | if (array) | |
4079 | dequeue_task(p, array); | |
4080 | p->prio = prio; | |
4081 | ||
4082 | if (array) { | |
4083 | /* | |
4084 | * If changing to an RT priority then queue it | |
4085 | * in the active array! | |
4086 | */ | |
4087 | if (rt_task(p)) | |
4088 | array = rq->active; | |
4089 | enqueue_task(p, array); | |
4090 | /* | |
4091 | * Reschedule if we are currently running on this runqueue and | |
4092 | * our priority decreased, or if we are not currently running on | |
4093 | * this runqueue and our priority is higher than the current's | |
4094 | */ | |
4095 | if (task_running(rq, p)) { | |
4096 | if (p->prio > oldprio) | |
4097 | resched_task(rq->curr); | |
4098 | } else if (TASK_PREEMPTS_CURR(p, rq)) | |
4099 | resched_task(rq->curr); | |
4100 | } | |
4101 | task_rq_unlock(rq, &flags); | |
4102 | } | |
4103 | ||
4104 | #endif | |
4105 | ||
36c8b586 | 4106 | void set_user_nice(struct task_struct *p, long nice) |
1da177e4 | 4107 | { |
70b97a7f | 4108 | struct prio_array *array; |
48f24c4d | 4109 | int old_prio, delta; |
1da177e4 | 4110 | unsigned long flags; |
70b97a7f | 4111 | struct rq *rq; |
1da177e4 LT |
4112 | |
4113 | if (TASK_NICE(p) == nice || nice < -20 || nice > 19) | |
4114 | return; | |
4115 | /* | |
4116 | * We have to be careful, if called from sys_setpriority(), | |
4117 | * the task might be in the middle of scheduling on another CPU. | |
4118 | */ | |
4119 | rq = task_rq_lock(p, &flags); | |
4120 | /* | |
4121 | * The RT priorities are set via sched_setscheduler(), but we still | |
4122 | * allow the 'normal' nice value to be set - but as expected | |
4123 | * it wont have any effect on scheduling until the task is | |
b0a9499c | 4124 | * not SCHED_NORMAL/SCHED_BATCH: |
1da177e4 | 4125 | */ |
b29739f9 | 4126 | if (has_rt_policy(p)) { |
1da177e4 LT |
4127 | p->static_prio = NICE_TO_PRIO(nice); |
4128 | goto out_unlock; | |
4129 | } | |
4130 | array = p->array; | |
2dd73a4f | 4131 | if (array) { |
1da177e4 | 4132 | dequeue_task(p, array); |
2dd73a4f PW |
4133 | dec_raw_weighted_load(rq, p); |
4134 | } | |
1da177e4 | 4135 | |
1da177e4 | 4136 | p->static_prio = NICE_TO_PRIO(nice); |
2dd73a4f | 4137 | set_load_weight(p); |
b29739f9 IM |
4138 | old_prio = p->prio; |
4139 | p->prio = effective_prio(p); | |
4140 | delta = p->prio - old_prio; | |
1da177e4 LT |
4141 | |
4142 | if (array) { | |
4143 | enqueue_task(p, array); | |
2dd73a4f | 4144 | inc_raw_weighted_load(rq, p); |
1da177e4 LT |
4145 | /* |
4146 | * If the task increased its priority or is running and | |
4147 | * lowered its priority, then reschedule its CPU: | |
4148 | */ | |
4149 | if (delta < 0 || (delta > 0 && task_running(rq, p))) | |
4150 | resched_task(rq->curr); | |
4151 | } | |
4152 | out_unlock: | |
4153 | task_rq_unlock(rq, &flags); | |
4154 | } | |
1da177e4 LT |
4155 | EXPORT_SYMBOL(set_user_nice); |
4156 | ||
e43379f1 MM |
4157 | /* |
4158 | * can_nice - check if a task can reduce its nice value | |
4159 | * @p: task | |
4160 | * @nice: nice value | |
4161 | */ | |
36c8b586 | 4162 | int can_nice(const struct task_struct *p, const int nice) |
e43379f1 | 4163 | { |
024f4747 MM |
4164 | /* convert nice value [19,-20] to rlimit style value [1,40] */ |
4165 | int nice_rlim = 20 - nice; | |
48f24c4d | 4166 | |
e43379f1 MM |
4167 | return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur || |
4168 | capable(CAP_SYS_NICE)); | |
4169 | } | |
4170 | ||
1da177e4 LT |
4171 | #ifdef __ARCH_WANT_SYS_NICE |
4172 | ||
4173 | /* | |
4174 | * sys_nice - change the priority of the current process. | |
4175 | * @increment: priority increment | |
4176 | * | |
4177 | * sys_setpriority is a more generic, but much slower function that | |
4178 | * does similar things. | |
4179 | */ | |
4180 | asmlinkage long sys_nice(int increment) | |
4181 | { | |
48f24c4d | 4182 | long nice, retval; |
1da177e4 LT |
4183 | |
4184 | /* | |
4185 | * Setpriority might change our priority at the same moment. | |
4186 | * We don't have to worry. Conceptually one call occurs first | |
4187 | * and we have a single winner. | |
4188 | */ | |
e43379f1 MM |
4189 | if (increment < -40) |
4190 | increment = -40; | |
1da177e4 LT |
4191 | if (increment > 40) |
4192 | increment = 40; | |
4193 | ||
4194 | nice = PRIO_TO_NICE(current->static_prio) + increment; | |
4195 | if (nice < -20) | |
4196 | nice = -20; | |
4197 | if (nice > 19) | |
4198 | nice = 19; | |
4199 | ||
e43379f1 MM |
4200 | if (increment < 0 && !can_nice(current, nice)) |
4201 | return -EPERM; | |
4202 | ||
1da177e4 LT |
4203 | retval = security_task_setnice(current, nice); |
4204 | if (retval) | |
4205 | return retval; | |
4206 | ||
4207 | set_user_nice(current, nice); | |
4208 | return 0; | |
4209 | } | |
4210 | ||
4211 | #endif | |
4212 | ||
4213 | /** | |
4214 | * task_prio - return the priority value of a given task. | |
4215 | * @p: the task in question. | |
4216 | * | |
4217 | * This is the priority value as seen by users in /proc. | |
4218 | * RT tasks are offset by -200. Normal tasks are centered | |
4219 | * around 0, value goes from -16 to +15. | |
4220 | */ | |
36c8b586 | 4221 | int task_prio(const struct task_struct *p) |
1da177e4 LT |
4222 | { |
4223 | return p->prio - MAX_RT_PRIO; | |
4224 | } | |
4225 | ||
4226 | /** | |
4227 | * task_nice - return the nice value of a given task. | |
4228 | * @p: the task in question. | |
4229 | */ | |
36c8b586 | 4230 | int task_nice(const struct task_struct *p) |
1da177e4 LT |
4231 | { |
4232 | return TASK_NICE(p); | |
4233 | } | |
1da177e4 | 4234 | EXPORT_SYMBOL_GPL(task_nice); |
1da177e4 LT |
4235 | |
4236 | /** | |
4237 | * idle_cpu - is a given cpu idle currently? | |
4238 | * @cpu: the processor in question. | |
4239 | */ | |
4240 | int idle_cpu(int cpu) | |
4241 | { | |
4242 | return cpu_curr(cpu) == cpu_rq(cpu)->idle; | |
4243 | } | |
4244 | ||
1da177e4 LT |
4245 | /** |
4246 | * idle_task - return the idle task for a given cpu. | |
4247 | * @cpu: the processor in question. | |
4248 | */ | |
36c8b586 | 4249 | struct task_struct *idle_task(int cpu) |
1da177e4 LT |
4250 | { |
4251 | return cpu_rq(cpu)->idle; | |
4252 | } | |
4253 | ||
4254 | /** | |
4255 | * find_process_by_pid - find a process with a matching PID value. | |
4256 | * @pid: the pid in question. | |
4257 | */ | |
36c8b586 | 4258 | static inline struct task_struct *find_process_by_pid(pid_t pid) |
1da177e4 LT |
4259 | { |
4260 | return pid ? find_task_by_pid(pid) : current; | |
4261 | } | |
4262 | ||
4263 | /* Actually do priority change: must hold rq lock. */ | |
4264 | static void __setscheduler(struct task_struct *p, int policy, int prio) | |
4265 | { | |
4266 | BUG_ON(p->array); | |
48f24c4d | 4267 | |
1da177e4 LT |
4268 | p->policy = policy; |
4269 | p->rt_priority = prio; | |
b29739f9 IM |
4270 | p->normal_prio = normal_prio(p); |
4271 | /* we are holding p->pi_lock already */ | |
4272 | p->prio = rt_mutex_getprio(p); | |
4273 | /* | |
4274 | * SCHED_BATCH tasks are treated as perpetual CPU hogs: | |
4275 | */ | |
4276 | if (policy == SCHED_BATCH) | |
4277 | p->sleep_avg = 0; | |
2dd73a4f | 4278 | set_load_weight(p); |
1da177e4 LT |
4279 | } |
4280 | ||
4281 | /** | |
72fd4a35 | 4282 | * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. |
1da177e4 LT |
4283 | * @p: the task in question. |
4284 | * @policy: new policy. | |
4285 | * @param: structure containing the new RT priority. | |
5fe1d75f | 4286 | * |
72fd4a35 | 4287 | * NOTE that the task may be already dead. |
1da177e4 | 4288 | */ |
95cdf3b7 IM |
4289 | int sched_setscheduler(struct task_struct *p, int policy, |
4290 | struct sched_param *param) | |
1da177e4 | 4291 | { |
48f24c4d | 4292 | int retval, oldprio, oldpolicy = -1; |
70b97a7f | 4293 | struct prio_array *array; |
1da177e4 | 4294 | unsigned long flags; |
70b97a7f | 4295 | struct rq *rq; |
1da177e4 | 4296 | |
66e5393a SR |
4297 | /* may grab non-irq protected spin_locks */ |
4298 | BUG_ON(in_interrupt()); | |
1da177e4 LT |
4299 | recheck: |
4300 | /* double check policy once rq lock held */ | |
4301 | if (policy < 0) | |
4302 | policy = oldpolicy = p->policy; | |
4303 | else if (policy != SCHED_FIFO && policy != SCHED_RR && | |
b0a9499c IM |
4304 | policy != SCHED_NORMAL && policy != SCHED_BATCH) |
4305 | return -EINVAL; | |
1da177e4 LT |
4306 | /* |
4307 | * Valid priorities for SCHED_FIFO and SCHED_RR are | |
b0a9499c IM |
4308 | * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and |
4309 | * SCHED_BATCH is 0. | |
1da177e4 LT |
4310 | */ |
4311 | if (param->sched_priority < 0 || | |
95cdf3b7 | 4312 | (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || |
d46523ea | 4313 | (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) |
1da177e4 | 4314 | return -EINVAL; |
57a6f51c | 4315 | if (is_rt_policy(policy) != (param->sched_priority != 0)) |
1da177e4 LT |
4316 | return -EINVAL; |
4317 | ||
37e4ab3f OC |
4318 | /* |
4319 | * Allow unprivileged RT tasks to decrease priority: | |
4320 | */ | |
4321 | if (!capable(CAP_SYS_NICE)) { | |
8dc3e909 ON |
4322 | if (is_rt_policy(policy)) { |
4323 | unsigned long rlim_rtprio; | |
4324 | unsigned long flags; | |
4325 | ||
4326 | if (!lock_task_sighand(p, &flags)) | |
4327 | return -ESRCH; | |
4328 | rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur; | |
4329 | unlock_task_sighand(p, &flags); | |
4330 | ||
4331 | /* can't set/change the rt policy */ | |
4332 | if (policy != p->policy && !rlim_rtprio) | |
4333 | return -EPERM; | |
4334 | ||
4335 | /* can't increase priority */ | |
4336 | if (param->sched_priority > p->rt_priority && | |
4337 | param->sched_priority > rlim_rtprio) | |
4338 | return -EPERM; | |
4339 | } | |
5fe1d75f | 4340 | |
37e4ab3f OC |
4341 | /* can't change other user's priorities */ |
4342 | if ((current->euid != p->euid) && | |
4343 | (current->euid != p->uid)) | |
4344 | return -EPERM; | |
4345 | } | |
1da177e4 LT |
4346 | |
4347 | retval = security_task_setscheduler(p, policy, param); | |
4348 | if (retval) | |
4349 | return retval; | |
b29739f9 IM |
4350 | /* |
4351 | * make sure no PI-waiters arrive (or leave) while we are | |
4352 | * changing the priority of the task: | |
4353 | */ | |
4354 | spin_lock_irqsave(&p->pi_lock, flags); | |
1da177e4 LT |
4355 | /* |
4356 | * To be able to change p->policy safely, the apropriate | |
4357 | * runqueue lock must be held. | |
4358 | */ | |
b29739f9 | 4359 | rq = __task_rq_lock(p); |
1da177e4 LT |
4360 | /* recheck policy now with rq lock held */ |
4361 | if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { | |
4362 | policy = oldpolicy = -1; | |
b29739f9 IM |
4363 | __task_rq_unlock(rq); |
4364 | spin_unlock_irqrestore(&p->pi_lock, flags); | |
1da177e4 LT |
4365 | goto recheck; |
4366 | } | |
4367 | array = p->array; | |
4368 | if (array) | |
4369 | deactivate_task(p, rq); | |
4370 | oldprio = p->prio; | |
4371 | __setscheduler(p, policy, param->sched_priority); | |
4372 | if (array) { | |
4373 | __activate_task(p, rq); | |
4374 | /* | |
4375 | * Reschedule if we are currently running on this runqueue and | |
4376 | * our priority decreased, or if we are not currently running on | |
4377 | * this runqueue and our priority is higher than the current's | |
4378 | */ | |
4379 | if (task_running(rq, p)) { | |
4380 | if (p->prio > oldprio) | |
4381 | resched_task(rq->curr); | |
4382 | } else if (TASK_PREEMPTS_CURR(p, rq)) | |
4383 | resched_task(rq->curr); | |
4384 | } | |
b29739f9 IM |
4385 | __task_rq_unlock(rq); |
4386 | spin_unlock_irqrestore(&p->pi_lock, flags); | |
4387 | ||
95e02ca9 TG |
4388 | rt_mutex_adjust_pi(p); |
4389 | ||
1da177e4 LT |
4390 | return 0; |
4391 | } | |
4392 | EXPORT_SYMBOL_GPL(sched_setscheduler); | |
4393 | ||
95cdf3b7 IM |
4394 | static int |
4395 | do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) | |
1da177e4 | 4396 | { |
1da177e4 LT |
4397 | struct sched_param lparam; |
4398 | struct task_struct *p; | |
36c8b586 | 4399 | int retval; |
1da177e4 LT |
4400 | |
4401 | if (!param || pid < 0) | |
4402 | return -EINVAL; | |
4403 | if (copy_from_user(&lparam, param, sizeof(struct sched_param))) | |
4404 | return -EFAULT; | |
5fe1d75f ON |
4405 | |
4406 | rcu_read_lock(); | |
4407 | retval = -ESRCH; | |
1da177e4 | 4408 | p = find_process_by_pid(pid); |
5fe1d75f ON |
4409 | if (p != NULL) |
4410 | retval = sched_setscheduler(p, policy, &lparam); | |
4411 | rcu_read_unlock(); | |
36c8b586 | 4412 | |
1da177e4 LT |
4413 | return retval; |
4414 | } | |
4415 | ||
4416 | /** | |
4417 | * sys_sched_setscheduler - set/change the scheduler policy and RT priority | |
4418 | * @pid: the pid in question. | |
4419 | * @policy: new policy. | |
4420 | * @param: structure containing the new RT priority. | |
4421 | */ | |
4422 | asmlinkage long sys_sched_setscheduler(pid_t pid, int policy, | |
4423 | struct sched_param __user *param) | |
4424 | { | |
c21761f1 JB |
4425 | /* negative values for policy are not valid */ |
4426 | if (policy < 0) | |
4427 | return -EINVAL; | |
4428 | ||
1da177e4 LT |
4429 | return do_sched_setscheduler(pid, policy, param); |
4430 | } | |
4431 | ||
4432 | /** | |
4433 | * sys_sched_setparam - set/change the RT priority of a thread | |
4434 | * @pid: the pid in question. | |
4435 | * @param: structure containing the new RT priority. | |
4436 | */ | |
4437 | asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param) | |
4438 | { | |
4439 | return do_sched_setscheduler(pid, -1, param); | |
4440 | } | |
4441 | ||
4442 | /** | |
4443 | * sys_sched_getscheduler - get the policy (scheduling class) of a thread | |
4444 | * @pid: the pid in question. | |
4445 | */ | |
4446 | asmlinkage long sys_sched_getscheduler(pid_t pid) | |
4447 | { | |
36c8b586 | 4448 | struct task_struct *p; |
1da177e4 | 4449 | int retval = -EINVAL; |
1da177e4 LT |
4450 | |
4451 | if (pid < 0) | |
4452 | goto out_nounlock; | |
4453 | ||
4454 | retval = -ESRCH; | |
4455 | read_lock(&tasklist_lock); | |
4456 | p = find_process_by_pid(pid); | |
4457 | if (p) { | |
4458 | retval = security_task_getscheduler(p); | |
4459 | if (!retval) | |
4460 | retval = p->policy; | |
4461 | } | |
4462 | read_unlock(&tasklist_lock); | |
4463 | ||
4464 | out_nounlock: | |
4465 | return retval; | |
4466 | } | |
4467 | ||
4468 | /** | |
4469 | * sys_sched_getscheduler - get the RT priority of a thread | |
4470 | * @pid: the pid in question. | |
4471 | * @param: structure containing the RT priority. | |
4472 | */ | |
4473 | asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param) | |
4474 | { | |
4475 | struct sched_param lp; | |
36c8b586 | 4476 | struct task_struct *p; |
1da177e4 | 4477 | int retval = -EINVAL; |
1da177e4 LT |
4478 | |
4479 | if (!param || pid < 0) | |
4480 | goto out_nounlock; | |
4481 | ||
4482 | read_lock(&tasklist_lock); | |
4483 | p = find_process_by_pid(pid); | |
4484 | retval = -ESRCH; | |
4485 | if (!p) | |
4486 | goto out_unlock; | |
4487 | ||
4488 | retval = security_task_getscheduler(p); | |
4489 | if (retval) | |
4490 | goto out_unlock; | |
4491 | ||
4492 | lp.sched_priority = p->rt_priority; | |
4493 | read_unlock(&tasklist_lock); | |
4494 | ||
4495 | /* | |
4496 | * This one might sleep, we cannot do it with a spinlock held ... | |
4497 | */ | |
4498 | retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; | |
4499 | ||
4500 | out_nounlock: | |
4501 | return retval; | |
4502 | ||
4503 | out_unlock: | |
4504 | read_unlock(&tasklist_lock); | |
4505 | return retval; | |
4506 | } | |
4507 | ||
4508 | long sched_setaffinity(pid_t pid, cpumask_t new_mask) | |
4509 | { | |
1da177e4 | 4510 | cpumask_t cpus_allowed; |
36c8b586 IM |
4511 | struct task_struct *p; |
4512 | int retval; | |
1da177e4 LT |
4513 | |
4514 | lock_cpu_hotplug(); | |
4515 | read_lock(&tasklist_lock); | |
4516 | ||
4517 | p = find_process_by_pid(pid); | |
4518 | if (!p) { | |
4519 | read_unlock(&tasklist_lock); | |
4520 | unlock_cpu_hotplug(); | |
4521 | return -ESRCH; | |
4522 | } | |
4523 | ||
4524 | /* | |
4525 | * It is not safe to call set_cpus_allowed with the | |
4526 | * tasklist_lock held. We will bump the task_struct's | |
4527 | * usage count and then drop tasklist_lock. | |
4528 | */ | |
4529 | get_task_struct(p); | |
4530 | read_unlock(&tasklist_lock); | |
4531 | ||
4532 | retval = -EPERM; | |
4533 | if ((current->euid != p->euid) && (current->euid != p->uid) && | |
4534 | !capable(CAP_SYS_NICE)) | |
4535 | goto out_unlock; | |
4536 | ||
e7834f8f DQ |
4537 | retval = security_task_setscheduler(p, 0, NULL); |
4538 | if (retval) | |
4539 | goto out_unlock; | |
4540 | ||
1da177e4 LT |
4541 | cpus_allowed = cpuset_cpus_allowed(p); |
4542 | cpus_and(new_mask, new_mask, cpus_allowed); | |
4543 | retval = set_cpus_allowed(p, new_mask); | |
4544 | ||
4545 | out_unlock: | |
4546 | put_task_struct(p); | |
4547 | unlock_cpu_hotplug(); | |
4548 | return retval; | |
4549 | } | |
4550 | ||
4551 | static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, | |
4552 | cpumask_t *new_mask) | |
4553 | { | |
4554 | if (len < sizeof(cpumask_t)) { | |
4555 | memset(new_mask, 0, sizeof(cpumask_t)); | |
4556 | } else if (len > sizeof(cpumask_t)) { | |
4557 | len = sizeof(cpumask_t); | |
4558 | } | |
4559 | return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; | |
4560 | } | |
4561 | ||
4562 | /** | |
4563 | * sys_sched_setaffinity - set the cpu affinity of a process | |
4564 | * @pid: pid of the process | |
4565 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
4566 | * @user_mask_ptr: user-space pointer to the new cpu mask | |
4567 | */ | |
4568 | asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len, | |
4569 | unsigned long __user *user_mask_ptr) | |
4570 | { | |
4571 | cpumask_t new_mask; | |
4572 | int retval; | |
4573 | ||
4574 | retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask); | |
4575 | if (retval) | |
4576 | return retval; | |
4577 | ||
4578 | return sched_setaffinity(pid, new_mask); | |
4579 | } | |
4580 | ||
4581 | /* | |
4582 | * Represents all cpu's present in the system | |
4583 | * In systems capable of hotplug, this map could dynamically grow | |
4584 | * as new cpu's are detected in the system via any platform specific | |
4585 | * method, such as ACPI for e.g. | |
4586 | */ | |
4587 | ||
4cef0c61 | 4588 | cpumask_t cpu_present_map __read_mostly; |
1da177e4 LT |
4589 | EXPORT_SYMBOL(cpu_present_map); |
4590 | ||
4591 | #ifndef CONFIG_SMP | |
4cef0c61 | 4592 | cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL; |
e16b38f7 GB |
4593 | EXPORT_SYMBOL(cpu_online_map); |
4594 | ||
4cef0c61 | 4595 | cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL; |
e16b38f7 | 4596 | EXPORT_SYMBOL(cpu_possible_map); |
1da177e4 LT |
4597 | #endif |
4598 | ||
4599 | long sched_getaffinity(pid_t pid, cpumask_t *mask) | |
4600 | { | |
36c8b586 | 4601 | struct task_struct *p; |
1da177e4 | 4602 | int retval; |
1da177e4 LT |
4603 | |
4604 | lock_cpu_hotplug(); | |
4605 | read_lock(&tasklist_lock); | |
4606 | ||
4607 | retval = -ESRCH; | |
4608 | p = find_process_by_pid(pid); | |
4609 | if (!p) | |
4610 | goto out_unlock; | |
4611 | ||
e7834f8f DQ |
4612 | retval = security_task_getscheduler(p); |
4613 | if (retval) | |
4614 | goto out_unlock; | |
4615 | ||
2f7016d9 | 4616 | cpus_and(*mask, p->cpus_allowed, cpu_online_map); |
1da177e4 LT |
4617 | |
4618 | out_unlock: | |
4619 | read_unlock(&tasklist_lock); | |
4620 | unlock_cpu_hotplug(); | |
4621 | if (retval) | |
4622 | return retval; | |
4623 | ||
4624 | return 0; | |
4625 | } | |
4626 | ||
4627 | /** | |
4628 | * sys_sched_getaffinity - get the cpu affinity of a process | |
4629 | * @pid: pid of the process | |
4630 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
4631 | * @user_mask_ptr: user-space pointer to hold the current cpu mask | |
4632 | */ | |
4633 | asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len, | |
4634 | unsigned long __user *user_mask_ptr) | |
4635 | { | |
4636 | int ret; | |
4637 | cpumask_t mask; | |
4638 | ||
4639 | if (len < sizeof(cpumask_t)) | |
4640 | return -EINVAL; | |
4641 | ||
4642 | ret = sched_getaffinity(pid, &mask); | |
4643 | if (ret < 0) | |
4644 | return ret; | |
4645 | ||
4646 | if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t))) | |
4647 | return -EFAULT; | |
4648 | ||
4649 | return sizeof(cpumask_t); | |
4650 | } | |
4651 | ||
4652 | /** | |
4653 | * sys_sched_yield - yield the current processor to other threads. | |
4654 | * | |
72fd4a35 | 4655 | * This function yields the current CPU by moving the calling thread |
1da177e4 LT |
4656 | * to the expired array. If there are no other threads running on this |
4657 | * CPU then this function will return. | |
4658 | */ | |
4659 | asmlinkage long sys_sched_yield(void) | |
4660 | { | |
70b97a7f IM |
4661 | struct rq *rq = this_rq_lock(); |
4662 | struct prio_array *array = current->array, *target = rq->expired; | |
1da177e4 LT |
4663 | |
4664 | schedstat_inc(rq, yld_cnt); | |
4665 | /* | |
4666 | * We implement yielding by moving the task into the expired | |
4667 | * queue. | |
4668 | * | |
4669 | * (special rule: RT tasks will just roundrobin in the active | |
4670 | * array.) | |
4671 | */ | |
4672 | if (rt_task(current)) | |
4673 | target = rq->active; | |
4674 | ||
5927ad78 | 4675 | if (array->nr_active == 1) { |
1da177e4 LT |
4676 | schedstat_inc(rq, yld_act_empty); |
4677 | if (!rq->expired->nr_active) | |
4678 | schedstat_inc(rq, yld_both_empty); | |
4679 | } else if (!rq->expired->nr_active) | |
4680 | schedstat_inc(rq, yld_exp_empty); | |
4681 | ||
4682 | if (array != target) { | |
4683 | dequeue_task(current, array); | |
4684 | enqueue_task(current, target); | |
4685 | } else | |
4686 | /* | |
4687 | * requeue_task is cheaper so perform that if possible. | |
4688 | */ | |
4689 | requeue_task(current, array); | |
4690 | ||
4691 | /* | |
4692 | * Since we are going to call schedule() anyway, there's | |
4693 | * no need to preempt or enable interrupts: | |
4694 | */ | |
4695 | __release(rq->lock); | |
8a25d5de | 4696 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); |
1da177e4 LT |
4697 | _raw_spin_unlock(&rq->lock); |
4698 | preempt_enable_no_resched(); | |
4699 | ||
4700 | schedule(); | |
4701 | ||
4702 | return 0; | |
4703 | } | |
4704 | ||
e7b38404 | 4705 | static void __cond_resched(void) |
1da177e4 | 4706 | { |
8e0a43d8 IM |
4707 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP |
4708 | __might_sleep(__FILE__, __LINE__); | |
4709 | #endif | |
5bbcfd90 IM |
4710 | /* |
4711 | * The BKS might be reacquired before we have dropped | |
4712 | * PREEMPT_ACTIVE, which could trigger a second | |
4713 | * cond_resched() call. | |
4714 | */ | |
1da177e4 LT |
4715 | do { |
4716 | add_preempt_count(PREEMPT_ACTIVE); | |
4717 | schedule(); | |
4718 | sub_preempt_count(PREEMPT_ACTIVE); | |
4719 | } while (need_resched()); | |
4720 | } | |
4721 | ||
4722 | int __sched cond_resched(void) | |
4723 | { | |
9414232f IM |
4724 | if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) && |
4725 | system_state == SYSTEM_RUNNING) { | |
1da177e4 LT |
4726 | __cond_resched(); |
4727 | return 1; | |
4728 | } | |
4729 | return 0; | |
4730 | } | |
1da177e4 LT |
4731 | EXPORT_SYMBOL(cond_resched); |
4732 | ||
4733 | /* | |
4734 | * cond_resched_lock() - if a reschedule is pending, drop the given lock, | |
4735 | * call schedule, and on return reacquire the lock. | |
4736 | * | |
4737 | * This works OK both with and without CONFIG_PREEMPT. We do strange low-level | |
4738 | * operations here to prevent schedule() from being called twice (once via | |
4739 | * spin_unlock(), once by hand). | |
4740 | */ | |
95cdf3b7 | 4741 | int cond_resched_lock(spinlock_t *lock) |
1da177e4 | 4742 | { |
6df3cecb JK |
4743 | int ret = 0; |
4744 | ||
1da177e4 LT |
4745 | if (need_lockbreak(lock)) { |
4746 | spin_unlock(lock); | |
4747 | cpu_relax(); | |
6df3cecb | 4748 | ret = 1; |
1da177e4 LT |
4749 | spin_lock(lock); |
4750 | } | |
9414232f | 4751 | if (need_resched() && system_state == SYSTEM_RUNNING) { |
8a25d5de | 4752 | spin_release(&lock->dep_map, 1, _THIS_IP_); |
1da177e4 LT |
4753 | _raw_spin_unlock(lock); |
4754 | preempt_enable_no_resched(); | |
4755 | __cond_resched(); | |
6df3cecb | 4756 | ret = 1; |
1da177e4 | 4757 | spin_lock(lock); |
1da177e4 | 4758 | } |
6df3cecb | 4759 | return ret; |
1da177e4 | 4760 | } |
1da177e4 LT |
4761 | EXPORT_SYMBOL(cond_resched_lock); |
4762 | ||
4763 | int __sched cond_resched_softirq(void) | |
4764 | { | |
4765 | BUG_ON(!in_softirq()); | |
4766 | ||
9414232f | 4767 | if (need_resched() && system_state == SYSTEM_RUNNING) { |
de30a2b3 IM |
4768 | raw_local_irq_disable(); |
4769 | _local_bh_enable(); | |
4770 | raw_local_irq_enable(); | |
1da177e4 LT |
4771 | __cond_resched(); |
4772 | local_bh_disable(); | |
4773 | return 1; | |
4774 | } | |
4775 | return 0; | |
4776 | } | |
1da177e4 LT |
4777 | EXPORT_SYMBOL(cond_resched_softirq); |
4778 | ||
1da177e4 LT |
4779 | /** |
4780 | * yield - yield the current processor to other threads. | |
4781 | * | |
72fd4a35 | 4782 | * This is a shortcut for kernel-space yielding - it marks the |
1da177e4 LT |
4783 | * thread runnable and calls sys_sched_yield(). |
4784 | */ | |
4785 | void __sched yield(void) | |
4786 | { | |
4787 | set_current_state(TASK_RUNNING); | |
4788 | sys_sched_yield(); | |
4789 | } | |
1da177e4 LT |
4790 | EXPORT_SYMBOL(yield); |
4791 | ||
4792 | /* | |
4793 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so | |
4794 | * that process accounting knows that this is a task in IO wait state. | |
4795 | * | |
4796 | * But don't do that if it is a deliberate, throttling IO wait (this task | |
4797 | * has set its backing_dev_info: the queue against which it should throttle) | |
4798 | */ | |
4799 | void __sched io_schedule(void) | |
4800 | { | |
70b97a7f | 4801 | struct rq *rq = &__raw_get_cpu_var(runqueues); |
1da177e4 | 4802 | |
0ff92245 | 4803 | delayacct_blkio_start(); |
1da177e4 LT |
4804 | atomic_inc(&rq->nr_iowait); |
4805 | schedule(); | |
4806 | atomic_dec(&rq->nr_iowait); | |
0ff92245 | 4807 | delayacct_blkio_end(); |
1da177e4 | 4808 | } |
1da177e4 LT |
4809 | EXPORT_SYMBOL(io_schedule); |
4810 | ||
4811 | long __sched io_schedule_timeout(long timeout) | |
4812 | { | |
70b97a7f | 4813 | struct rq *rq = &__raw_get_cpu_var(runqueues); |
1da177e4 LT |
4814 | long ret; |
4815 | ||
0ff92245 | 4816 | delayacct_blkio_start(); |
1da177e4 LT |
4817 | atomic_inc(&rq->nr_iowait); |
4818 | ret = schedule_timeout(timeout); | |
4819 | atomic_dec(&rq->nr_iowait); | |
0ff92245 | 4820 | delayacct_blkio_end(); |
1da177e4 LT |
4821 | return ret; |
4822 | } | |
4823 | ||
4824 | /** | |
4825 | * sys_sched_get_priority_max - return maximum RT priority. | |
4826 | * @policy: scheduling class. | |
4827 | * | |
4828 | * this syscall returns the maximum rt_priority that can be used | |
4829 | * by a given scheduling class. | |
4830 | */ | |
4831 | asmlinkage long sys_sched_get_priority_max(int policy) | |
4832 | { | |
4833 | int ret = -EINVAL; | |
4834 | ||
4835 | switch (policy) { | |
4836 | case SCHED_FIFO: | |
4837 | case SCHED_RR: | |
4838 | ret = MAX_USER_RT_PRIO-1; | |
4839 | break; | |
4840 | case SCHED_NORMAL: | |
b0a9499c | 4841 | case SCHED_BATCH: |
1da177e4 LT |
4842 | ret = 0; |
4843 | break; | |
4844 | } | |
4845 | return ret; | |
4846 | } | |
4847 | ||
4848 | /** | |
4849 | * sys_sched_get_priority_min - return minimum RT priority. | |
4850 | * @policy: scheduling class. | |
4851 | * | |
4852 | * this syscall returns the minimum rt_priority that can be used | |
4853 | * by a given scheduling class. | |
4854 | */ | |
4855 | asmlinkage long sys_sched_get_priority_min(int policy) | |
4856 | { | |
4857 | int ret = -EINVAL; | |
4858 | ||
4859 | switch (policy) { | |
4860 | case SCHED_FIFO: | |
4861 | case SCHED_RR: | |
4862 | ret = 1; | |
4863 | break; | |
4864 | case SCHED_NORMAL: | |
b0a9499c | 4865 | case SCHED_BATCH: |
1da177e4 LT |
4866 | ret = 0; |
4867 | } | |
4868 | return ret; | |
4869 | } | |
4870 | ||
4871 | /** | |
4872 | * sys_sched_rr_get_interval - return the default timeslice of a process. | |
4873 | * @pid: pid of the process. | |
4874 | * @interval: userspace pointer to the timeslice value. | |
4875 | * | |
4876 | * this syscall writes the default timeslice value of a given process | |
4877 | * into the user-space timespec buffer. A value of '0' means infinity. | |
4878 | */ | |
4879 | asmlinkage | |
4880 | long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval) | |
4881 | { | |
36c8b586 | 4882 | struct task_struct *p; |
1da177e4 LT |
4883 | int retval = -EINVAL; |
4884 | struct timespec t; | |
1da177e4 LT |
4885 | |
4886 | if (pid < 0) | |
4887 | goto out_nounlock; | |
4888 | ||
4889 | retval = -ESRCH; | |
4890 | read_lock(&tasklist_lock); | |
4891 | p = find_process_by_pid(pid); | |
4892 | if (!p) | |
4893 | goto out_unlock; | |
4894 | ||
4895 | retval = security_task_getscheduler(p); | |
4896 | if (retval) | |
4897 | goto out_unlock; | |
4898 | ||
b78709cf | 4899 | jiffies_to_timespec(p->policy == SCHED_FIFO ? |
1da177e4 LT |
4900 | 0 : task_timeslice(p), &t); |
4901 | read_unlock(&tasklist_lock); | |
4902 | retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; | |
4903 | out_nounlock: | |
4904 | return retval; | |
4905 | out_unlock: | |
4906 | read_unlock(&tasklist_lock); | |
4907 | return retval; | |
4908 | } | |
4909 | ||
2ed6e34f | 4910 | static const char stat_nam[] = "RSDTtZX"; |
36c8b586 IM |
4911 | |
4912 | static void show_task(struct task_struct *p) | |
1da177e4 | 4913 | { |
1da177e4 | 4914 | unsigned long free = 0; |
36c8b586 | 4915 | unsigned state; |
1da177e4 | 4916 | |
1da177e4 | 4917 | state = p->state ? __ffs(p->state) + 1 : 0; |
2ed6e34f AM |
4918 | printk("%-13.13s %c", p->comm, |
4919 | state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); | |
1da177e4 LT |
4920 | #if (BITS_PER_LONG == 32) |
4921 | if (state == TASK_RUNNING) | |
4922 | printk(" running "); | |
4923 | else | |
4924 | printk(" %08lX ", thread_saved_pc(p)); | |
4925 | #else | |
4926 | if (state == TASK_RUNNING) | |
4927 | printk(" running task "); | |
4928 | else | |
4929 | printk(" %016lx ", thread_saved_pc(p)); | |
4930 | #endif | |
4931 | #ifdef CONFIG_DEBUG_STACK_USAGE | |
4932 | { | |
10ebffde | 4933 | unsigned long *n = end_of_stack(p); |
1da177e4 LT |
4934 | while (!*n) |
4935 | n++; | |
10ebffde | 4936 | free = (unsigned long)n - (unsigned long)end_of_stack(p); |
1da177e4 LT |
4937 | } |
4938 | #endif | |
35f6f753 | 4939 | printk("%5lu %5d %6d", free, p->pid, p->parent->pid); |
1da177e4 LT |
4940 | if (!p->mm) |
4941 | printk(" (L-TLB)\n"); | |
4942 | else | |
4943 | printk(" (NOTLB)\n"); | |
4944 | ||
4945 | if (state != TASK_RUNNING) | |
4946 | show_stack(p, NULL); | |
4947 | } | |
4948 | ||
e59e2ae2 | 4949 | void show_state_filter(unsigned long state_filter) |
1da177e4 | 4950 | { |
36c8b586 | 4951 | struct task_struct *g, *p; |
1da177e4 LT |
4952 | |
4953 | #if (BITS_PER_LONG == 32) | |
4954 | printk("\n" | |
301827ac CC |
4955 | " free sibling\n"); |
4956 | printk(" task PC stack pid father child younger older\n"); | |
1da177e4 LT |
4957 | #else |
4958 | printk("\n" | |
301827ac CC |
4959 | " free sibling\n"); |
4960 | printk(" task PC stack pid father child younger older\n"); | |
1da177e4 LT |
4961 | #endif |
4962 | read_lock(&tasklist_lock); | |
4963 | do_each_thread(g, p) { | |
4964 | /* | |
4965 | * reset the NMI-timeout, listing all files on a slow | |
4966 | * console might take alot of time: | |
4967 | */ | |
4968 | touch_nmi_watchdog(); | |
39bc89fd | 4969 | if (!state_filter || (p->state & state_filter)) |
e59e2ae2 | 4970 | show_task(p); |
1da177e4 LT |
4971 | } while_each_thread(g, p); |
4972 | ||
04c9167f JF |
4973 | touch_all_softlockup_watchdogs(); |
4974 | ||
1da177e4 | 4975 | read_unlock(&tasklist_lock); |
e59e2ae2 IM |
4976 | /* |
4977 | * Only show locks if all tasks are dumped: | |
4978 | */ | |
4979 | if (state_filter == -1) | |
4980 | debug_show_all_locks(); | |
1da177e4 LT |
4981 | } |
4982 | ||
f340c0d1 IM |
4983 | /** |
4984 | * init_idle - set up an idle thread for a given CPU | |
4985 | * @idle: task in question | |
4986 | * @cpu: cpu the idle task belongs to | |
4987 | * | |
4988 | * NOTE: this function does not set the idle thread's NEED_RESCHED | |
4989 | * flag, to make booting more robust. | |
4990 | */ | |
5c1e1767 | 4991 | void __cpuinit init_idle(struct task_struct *idle, int cpu) |
1da177e4 | 4992 | { |
70b97a7f | 4993 | struct rq *rq = cpu_rq(cpu); |
1da177e4 LT |
4994 | unsigned long flags; |
4995 | ||
81c29a85 | 4996 | idle->timestamp = sched_clock(); |
1da177e4 LT |
4997 | idle->sleep_avg = 0; |
4998 | idle->array = NULL; | |
b29739f9 | 4999 | idle->prio = idle->normal_prio = MAX_PRIO; |
1da177e4 LT |
5000 | idle->state = TASK_RUNNING; |
5001 | idle->cpus_allowed = cpumask_of_cpu(cpu); | |
5002 | set_task_cpu(idle, cpu); | |
5003 | ||
5004 | spin_lock_irqsave(&rq->lock, flags); | |
5005 | rq->curr = rq->idle = idle; | |
4866cde0 NP |
5006 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
5007 | idle->oncpu = 1; | |
5008 | #endif | |
1da177e4 LT |
5009 | spin_unlock_irqrestore(&rq->lock, flags); |
5010 | ||
5011 | /* Set the preempt count _outside_ the spinlocks! */ | |
5012 | #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL) | |
a1261f54 | 5013 | task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0); |
1da177e4 | 5014 | #else |
a1261f54 | 5015 | task_thread_info(idle)->preempt_count = 0; |
1da177e4 LT |
5016 | #endif |
5017 | } | |
5018 | ||
5019 | /* | |
5020 | * In a system that switches off the HZ timer nohz_cpu_mask | |
5021 | * indicates which cpus entered this state. This is used | |
5022 | * in the rcu update to wait only for active cpus. For system | |
5023 | * which do not switch off the HZ timer nohz_cpu_mask should | |
5024 | * always be CPU_MASK_NONE. | |
5025 | */ | |
5026 | cpumask_t nohz_cpu_mask = CPU_MASK_NONE; | |
5027 | ||
5028 | #ifdef CONFIG_SMP | |
5029 | /* | |
5030 | * This is how migration works: | |
5031 | * | |
70b97a7f | 5032 | * 1) we queue a struct migration_req structure in the source CPU's |
1da177e4 LT |
5033 | * runqueue and wake up that CPU's migration thread. |
5034 | * 2) we down() the locked semaphore => thread blocks. | |
5035 | * 3) migration thread wakes up (implicitly it forces the migrated | |
5036 | * thread off the CPU) | |
5037 | * 4) it gets the migration request and checks whether the migrated | |
5038 | * task is still in the wrong runqueue. | |
5039 | * 5) if it's in the wrong runqueue then the migration thread removes | |
5040 | * it and puts it into the right queue. | |
5041 | * 6) migration thread up()s the semaphore. | |
5042 | * 7) we wake up and the migration is done. | |
5043 | */ | |
5044 | ||
5045 | /* | |
5046 | * Change a given task's CPU affinity. Migrate the thread to a | |
5047 | * proper CPU and schedule it away if the CPU it's executing on | |
5048 | * is removed from the allowed bitmask. | |
5049 | * | |
5050 | * NOTE: the caller must have a valid reference to the task, the | |
5051 | * task must not exit() & deallocate itself prematurely. The | |
5052 | * call is not atomic; no spinlocks may be held. | |
5053 | */ | |
36c8b586 | 5054 | int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask) |
1da177e4 | 5055 | { |
70b97a7f | 5056 | struct migration_req req; |
1da177e4 | 5057 | unsigned long flags; |
70b97a7f | 5058 | struct rq *rq; |
48f24c4d | 5059 | int ret = 0; |
1da177e4 LT |
5060 | |
5061 | rq = task_rq_lock(p, &flags); | |
5062 | if (!cpus_intersects(new_mask, cpu_online_map)) { | |
5063 | ret = -EINVAL; | |
5064 | goto out; | |
5065 | } | |
5066 | ||
5067 | p->cpus_allowed = new_mask; | |
5068 | /* Can the task run on the task's current CPU? If so, we're done */ | |
5069 | if (cpu_isset(task_cpu(p), new_mask)) | |
5070 | goto out; | |
5071 | ||
5072 | if (migrate_task(p, any_online_cpu(new_mask), &req)) { | |
5073 | /* Need help from migration thread: drop lock and wait. */ | |
5074 | task_rq_unlock(rq, &flags); | |
5075 | wake_up_process(rq->migration_thread); | |
5076 | wait_for_completion(&req.done); | |
5077 | tlb_migrate_finish(p->mm); | |
5078 | return 0; | |
5079 | } | |
5080 | out: | |
5081 | task_rq_unlock(rq, &flags); | |
48f24c4d | 5082 | |
1da177e4 LT |
5083 | return ret; |
5084 | } | |
1da177e4 LT |
5085 | EXPORT_SYMBOL_GPL(set_cpus_allowed); |
5086 | ||
5087 | /* | |
5088 | * Move (not current) task off this cpu, onto dest cpu. We're doing | |
5089 | * this because either it can't run here any more (set_cpus_allowed() | |
5090 | * away from this CPU, or CPU going down), or because we're | |
5091 | * attempting to rebalance this task on exec (sched_exec). | |
5092 | * | |
5093 | * So we race with normal scheduler movements, but that's OK, as long | |
5094 | * as the task is no longer on this CPU. | |
efc30814 KK |
5095 | * |
5096 | * Returns non-zero if task was successfully migrated. | |
1da177e4 | 5097 | */ |
efc30814 | 5098 | static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) |
1da177e4 | 5099 | { |
70b97a7f | 5100 | struct rq *rq_dest, *rq_src; |
efc30814 | 5101 | int ret = 0; |
1da177e4 LT |
5102 | |
5103 | if (unlikely(cpu_is_offline(dest_cpu))) | |
efc30814 | 5104 | return ret; |
1da177e4 LT |
5105 | |
5106 | rq_src = cpu_rq(src_cpu); | |
5107 | rq_dest = cpu_rq(dest_cpu); | |
5108 | ||
5109 | double_rq_lock(rq_src, rq_dest); | |
5110 | /* Already moved. */ | |
5111 | if (task_cpu(p) != src_cpu) | |
5112 | goto out; | |
5113 | /* Affinity changed (again). */ | |
5114 | if (!cpu_isset(dest_cpu, p->cpus_allowed)) | |
5115 | goto out; | |
5116 | ||
5117 | set_task_cpu(p, dest_cpu); | |
5118 | if (p->array) { | |
5119 | /* | |
5120 | * Sync timestamp with rq_dest's before activating. | |
5121 | * The same thing could be achieved by doing this step | |
5122 | * afterwards, and pretending it was a local activate. | |
5123 | * This way is cleaner and logically correct. | |
5124 | */ | |
b18ec803 MG |
5125 | p->timestamp = p->timestamp - rq_src->most_recent_timestamp |
5126 | + rq_dest->most_recent_timestamp; | |
1da177e4 | 5127 | deactivate_task(p, rq_src); |
0a565f79 | 5128 | __activate_task(p, rq_dest); |
1da177e4 LT |
5129 | if (TASK_PREEMPTS_CURR(p, rq_dest)) |
5130 | resched_task(rq_dest->curr); | |
5131 | } | |
efc30814 | 5132 | ret = 1; |
1da177e4 LT |
5133 | out: |
5134 | double_rq_unlock(rq_src, rq_dest); | |
efc30814 | 5135 | return ret; |
1da177e4 LT |
5136 | } |
5137 | ||
5138 | /* | |
5139 | * migration_thread - this is a highprio system thread that performs | |
5140 | * thread migration by bumping thread off CPU then 'pushing' onto | |
5141 | * another runqueue. | |
5142 | */ | |
95cdf3b7 | 5143 | static int migration_thread(void *data) |
1da177e4 | 5144 | { |
1da177e4 | 5145 | int cpu = (long)data; |
70b97a7f | 5146 | struct rq *rq; |
1da177e4 LT |
5147 | |
5148 | rq = cpu_rq(cpu); | |
5149 | BUG_ON(rq->migration_thread != current); | |
5150 | ||
5151 | set_current_state(TASK_INTERRUPTIBLE); | |
5152 | while (!kthread_should_stop()) { | |
70b97a7f | 5153 | struct migration_req *req; |
1da177e4 | 5154 | struct list_head *head; |
1da177e4 | 5155 | |
3e1d1d28 | 5156 | try_to_freeze(); |
1da177e4 LT |
5157 | |
5158 | spin_lock_irq(&rq->lock); | |
5159 | ||
5160 | if (cpu_is_offline(cpu)) { | |
5161 | spin_unlock_irq(&rq->lock); | |
5162 | goto wait_to_die; | |
5163 | } | |
5164 | ||
5165 | if (rq->active_balance) { | |
5166 | active_load_balance(rq, cpu); | |
5167 | rq->active_balance = 0; | |
5168 | } | |
5169 | ||
5170 | head = &rq->migration_queue; | |
5171 | ||
5172 | if (list_empty(head)) { | |
5173 | spin_unlock_irq(&rq->lock); | |
5174 | schedule(); | |
5175 | set_current_state(TASK_INTERRUPTIBLE); | |
5176 | continue; | |
5177 | } | |
70b97a7f | 5178 | req = list_entry(head->next, struct migration_req, list); |
1da177e4 LT |
5179 | list_del_init(head->next); |
5180 | ||
674311d5 NP |
5181 | spin_unlock(&rq->lock); |
5182 | __migrate_task(req->task, cpu, req->dest_cpu); | |
5183 | local_irq_enable(); | |
1da177e4 LT |
5184 | |
5185 | complete(&req->done); | |
5186 | } | |
5187 | __set_current_state(TASK_RUNNING); | |
5188 | return 0; | |
5189 | ||
5190 | wait_to_die: | |
5191 | /* Wait for kthread_stop */ | |
5192 | set_current_state(TASK_INTERRUPTIBLE); | |
5193 | while (!kthread_should_stop()) { | |
5194 | schedule(); | |
5195 | set_current_state(TASK_INTERRUPTIBLE); | |
5196 | } | |
5197 | __set_current_state(TASK_RUNNING); | |
5198 | return 0; | |
5199 | } | |
5200 | ||
5201 | #ifdef CONFIG_HOTPLUG_CPU | |
054b9108 KK |
5202 | /* |
5203 | * Figure out where task on dead CPU should go, use force if neccessary. | |
5204 | * NOTE: interrupts should be disabled by the caller | |
5205 | */ | |
48f24c4d | 5206 | static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p) |
1da177e4 | 5207 | { |
efc30814 | 5208 | unsigned long flags; |
1da177e4 | 5209 | cpumask_t mask; |
70b97a7f IM |
5210 | struct rq *rq; |
5211 | int dest_cpu; | |
1da177e4 | 5212 | |
efc30814 | 5213 | restart: |
1da177e4 LT |
5214 | /* On same node? */ |
5215 | mask = node_to_cpumask(cpu_to_node(dead_cpu)); | |
48f24c4d | 5216 | cpus_and(mask, mask, p->cpus_allowed); |
1da177e4 LT |
5217 | dest_cpu = any_online_cpu(mask); |
5218 | ||
5219 | /* On any allowed CPU? */ | |
5220 | if (dest_cpu == NR_CPUS) | |
48f24c4d | 5221 | dest_cpu = any_online_cpu(p->cpus_allowed); |
1da177e4 LT |
5222 | |
5223 | /* No more Mr. Nice Guy. */ | |
5224 | if (dest_cpu == NR_CPUS) { | |
48f24c4d IM |
5225 | rq = task_rq_lock(p, &flags); |
5226 | cpus_setall(p->cpus_allowed); | |
5227 | dest_cpu = any_online_cpu(p->cpus_allowed); | |
efc30814 | 5228 | task_rq_unlock(rq, &flags); |
1da177e4 LT |
5229 | |
5230 | /* | |
5231 | * Don't tell them about moving exiting tasks or | |
5232 | * kernel threads (both mm NULL), since they never | |
5233 | * leave kernel. | |
5234 | */ | |
48f24c4d | 5235 | if (p->mm && printk_ratelimit()) |
1da177e4 LT |
5236 | printk(KERN_INFO "process %d (%s) no " |
5237 | "longer affine to cpu%d\n", | |
48f24c4d | 5238 | p->pid, p->comm, dead_cpu); |
1da177e4 | 5239 | } |
48f24c4d | 5240 | if (!__migrate_task(p, dead_cpu, dest_cpu)) |
efc30814 | 5241 | goto restart; |
1da177e4 LT |
5242 | } |
5243 | ||
5244 | /* | |
5245 | * While a dead CPU has no uninterruptible tasks queued at this point, | |
5246 | * it might still have a nonzero ->nr_uninterruptible counter, because | |
5247 | * for performance reasons the counter is not stricly tracking tasks to | |
5248 | * their home CPUs. So we just add the counter to another CPU's counter, | |
5249 | * to keep the global sum constant after CPU-down: | |
5250 | */ | |
70b97a7f | 5251 | static void migrate_nr_uninterruptible(struct rq *rq_src) |
1da177e4 | 5252 | { |
70b97a7f | 5253 | struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL)); |
1da177e4 LT |
5254 | unsigned long flags; |
5255 | ||
5256 | local_irq_save(flags); | |
5257 | double_rq_lock(rq_src, rq_dest); | |
5258 | rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible; | |
5259 | rq_src->nr_uninterruptible = 0; | |
5260 | double_rq_unlock(rq_src, rq_dest); | |
5261 | local_irq_restore(flags); | |
5262 | } | |
5263 | ||
5264 | /* Run through task list and migrate tasks from the dead cpu. */ | |
5265 | static void migrate_live_tasks(int src_cpu) | |
5266 | { | |
48f24c4d | 5267 | struct task_struct *p, *t; |
1da177e4 LT |
5268 | |
5269 | write_lock_irq(&tasklist_lock); | |
5270 | ||
48f24c4d IM |
5271 | do_each_thread(t, p) { |
5272 | if (p == current) | |
1da177e4 LT |
5273 | continue; |
5274 | ||
48f24c4d IM |
5275 | if (task_cpu(p) == src_cpu) |
5276 | move_task_off_dead_cpu(src_cpu, p); | |
5277 | } while_each_thread(t, p); | |
1da177e4 LT |
5278 | |
5279 | write_unlock_irq(&tasklist_lock); | |
5280 | } | |
5281 | ||
5282 | /* Schedules idle task to be the next runnable task on current CPU. | |
5283 | * It does so by boosting its priority to highest possible and adding it to | |
48f24c4d | 5284 | * the _front_ of the runqueue. Used by CPU offline code. |
1da177e4 LT |
5285 | */ |
5286 | void sched_idle_next(void) | |
5287 | { | |
48f24c4d | 5288 | int this_cpu = smp_processor_id(); |
70b97a7f | 5289 | struct rq *rq = cpu_rq(this_cpu); |
1da177e4 LT |
5290 | struct task_struct *p = rq->idle; |
5291 | unsigned long flags; | |
5292 | ||
5293 | /* cpu has to be offline */ | |
48f24c4d | 5294 | BUG_ON(cpu_online(this_cpu)); |
1da177e4 | 5295 | |
48f24c4d IM |
5296 | /* |
5297 | * Strictly not necessary since rest of the CPUs are stopped by now | |
5298 | * and interrupts disabled on the current cpu. | |
1da177e4 LT |
5299 | */ |
5300 | spin_lock_irqsave(&rq->lock, flags); | |
5301 | ||
5302 | __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1); | |
48f24c4d IM |
5303 | |
5304 | /* Add idle task to the _front_ of its priority queue: */ | |
1da177e4 LT |
5305 | __activate_idle_task(p, rq); |
5306 | ||
5307 | spin_unlock_irqrestore(&rq->lock, flags); | |
5308 | } | |
5309 | ||
48f24c4d IM |
5310 | /* |
5311 | * Ensures that the idle task is using init_mm right before its cpu goes | |
1da177e4 LT |
5312 | * offline. |
5313 | */ | |
5314 | void idle_task_exit(void) | |
5315 | { | |
5316 | struct mm_struct *mm = current->active_mm; | |
5317 | ||
5318 | BUG_ON(cpu_online(smp_processor_id())); | |
5319 | ||
5320 | if (mm != &init_mm) | |
5321 | switch_mm(mm, &init_mm, current); | |
5322 | mmdrop(mm); | |
5323 | } | |
5324 | ||
054b9108 | 5325 | /* called under rq->lock with disabled interrupts */ |
36c8b586 | 5326 | static void migrate_dead(unsigned int dead_cpu, struct task_struct *p) |
1da177e4 | 5327 | { |
70b97a7f | 5328 | struct rq *rq = cpu_rq(dead_cpu); |
1da177e4 LT |
5329 | |
5330 | /* Must be exiting, otherwise would be on tasklist. */ | |
48f24c4d | 5331 | BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD); |
1da177e4 LT |
5332 | |
5333 | /* Cannot have done final schedule yet: would have vanished. */ | |
c394cc9f | 5334 | BUG_ON(p->state == TASK_DEAD); |
1da177e4 | 5335 | |
48f24c4d | 5336 | get_task_struct(p); |
1da177e4 LT |
5337 | |
5338 | /* | |
5339 | * Drop lock around migration; if someone else moves it, | |
5340 | * that's OK. No task can be added to this CPU, so iteration is | |
5341 | * fine. | |
054b9108 | 5342 | * NOTE: interrupts should be left disabled --dev@ |
1da177e4 | 5343 | */ |
054b9108 | 5344 | spin_unlock(&rq->lock); |
48f24c4d | 5345 | move_task_off_dead_cpu(dead_cpu, p); |
054b9108 | 5346 | spin_lock(&rq->lock); |
1da177e4 | 5347 | |
48f24c4d | 5348 | put_task_struct(p); |
1da177e4 LT |
5349 | } |
5350 | ||
5351 | /* release_task() removes task from tasklist, so we won't find dead tasks. */ | |
5352 | static void migrate_dead_tasks(unsigned int dead_cpu) | |
5353 | { | |
70b97a7f | 5354 | struct rq *rq = cpu_rq(dead_cpu); |
48f24c4d | 5355 | unsigned int arr, i; |
1da177e4 LT |
5356 | |
5357 | for (arr = 0; arr < 2; arr++) { | |
5358 | for (i = 0; i < MAX_PRIO; i++) { | |
5359 | struct list_head *list = &rq->arrays[arr].queue[i]; | |
48f24c4d | 5360 | |
1da177e4 | 5361 | while (!list_empty(list)) |
36c8b586 IM |
5362 | migrate_dead(dead_cpu, list_entry(list->next, |
5363 | struct task_struct, run_list)); | |
1da177e4 LT |
5364 | } |
5365 | } | |
5366 | } | |
5367 | #endif /* CONFIG_HOTPLUG_CPU */ | |
5368 | ||
5369 | /* | |
5370 | * migration_call - callback that gets triggered when a CPU is added. | |
5371 | * Here we can start up the necessary migration thread for the new CPU. | |
5372 | */ | |
48f24c4d IM |
5373 | static int __cpuinit |
5374 | migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) | |
1da177e4 | 5375 | { |
1da177e4 | 5376 | struct task_struct *p; |
48f24c4d | 5377 | int cpu = (long)hcpu; |
1da177e4 | 5378 | unsigned long flags; |
70b97a7f | 5379 | struct rq *rq; |
1da177e4 LT |
5380 | |
5381 | switch (action) { | |
5382 | case CPU_UP_PREPARE: | |
5383 | p = kthread_create(migration_thread, hcpu, "migration/%d",cpu); | |
5384 | if (IS_ERR(p)) | |
5385 | return NOTIFY_BAD; | |
5386 | p->flags |= PF_NOFREEZE; | |
5387 | kthread_bind(p, cpu); | |
5388 | /* Must be high prio: stop_machine expects to yield to it. */ | |
5389 | rq = task_rq_lock(p, &flags); | |
5390 | __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1); | |
5391 | task_rq_unlock(rq, &flags); | |
5392 | cpu_rq(cpu)->migration_thread = p; | |
5393 | break; | |
48f24c4d | 5394 | |
1da177e4 LT |
5395 | case CPU_ONLINE: |
5396 | /* Strictly unneccessary, as first user will wake it. */ | |
5397 | wake_up_process(cpu_rq(cpu)->migration_thread); | |
5398 | break; | |
48f24c4d | 5399 | |
1da177e4 LT |
5400 | #ifdef CONFIG_HOTPLUG_CPU |
5401 | case CPU_UP_CANCELED: | |
fc75cdfa HC |
5402 | if (!cpu_rq(cpu)->migration_thread) |
5403 | break; | |
1da177e4 | 5404 | /* Unbind it from offline cpu so it can run. Fall thru. */ |
a4c4af7c HC |
5405 | kthread_bind(cpu_rq(cpu)->migration_thread, |
5406 | any_online_cpu(cpu_online_map)); | |
1da177e4 LT |
5407 | kthread_stop(cpu_rq(cpu)->migration_thread); |
5408 | cpu_rq(cpu)->migration_thread = NULL; | |
5409 | break; | |
48f24c4d | 5410 | |
1da177e4 LT |
5411 | case CPU_DEAD: |
5412 | migrate_live_tasks(cpu); | |
5413 | rq = cpu_rq(cpu); | |
5414 | kthread_stop(rq->migration_thread); | |
5415 | rq->migration_thread = NULL; | |
5416 | /* Idle task back to normal (off runqueue, low prio) */ | |
5417 | rq = task_rq_lock(rq->idle, &flags); | |
5418 | deactivate_task(rq->idle, rq); | |
5419 | rq->idle->static_prio = MAX_PRIO; | |
5420 | __setscheduler(rq->idle, SCHED_NORMAL, 0); | |
5421 | migrate_dead_tasks(cpu); | |
5422 | task_rq_unlock(rq, &flags); | |
5423 | migrate_nr_uninterruptible(rq); | |
5424 | BUG_ON(rq->nr_running != 0); | |
5425 | ||
5426 | /* No need to migrate the tasks: it was best-effort if | |
5427 | * they didn't do lock_cpu_hotplug(). Just wake up | |
5428 | * the requestors. */ | |
5429 | spin_lock_irq(&rq->lock); | |
5430 | while (!list_empty(&rq->migration_queue)) { | |
70b97a7f IM |
5431 | struct migration_req *req; |
5432 | ||
1da177e4 | 5433 | req = list_entry(rq->migration_queue.next, |
70b97a7f | 5434 | struct migration_req, list); |
1da177e4 LT |
5435 | list_del_init(&req->list); |
5436 | complete(&req->done); | |
5437 | } | |
5438 | spin_unlock_irq(&rq->lock); | |
5439 | break; | |
5440 | #endif | |
5441 | } | |
5442 | return NOTIFY_OK; | |
5443 | } | |
5444 | ||
5445 | /* Register at highest priority so that task migration (migrate_all_tasks) | |
5446 | * happens before everything else. | |
5447 | */ | |
26c2143b | 5448 | static struct notifier_block __cpuinitdata migration_notifier = { |
1da177e4 LT |
5449 | .notifier_call = migration_call, |
5450 | .priority = 10 | |
5451 | }; | |
5452 | ||
5453 | int __init migration_init(void) | |
5454 | { | |
5455 | void *cpu = (void *)(long)smp_processor_id(); | |
07dccf33 | 5456 | int err; |
48f24c4d IM |
5457 | |
5458 | /* Start one for the boot CPU: */ | |
07dccf33 AM |
5459 | err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); |
5460 | BUG_ON(err == NOTIFY_BAD); | |
1da177e4 LT |
5461 | migration_call(&migration_notifier, CPU_ONLINE, cpu); |
5462 | register_cpu_notifier(&migration_notifier); | |
48f24c4d | 5463 | |
1da177e4 LT |
5464 | return 0; |
5465 | } | |
5466 | #endif | |
5467 | ||
5468 | #ifdef CONFIG_SMP | |
476f3534 CL |
5469 | |
5470 | /* Number of possible processor ids */ | |
5471 | int nr_cpu_ids __read_mostly = NR_CPUS; | |
5472 | EXPORT_SYMBOL(nr_cpu_ids); | |
5473 | ||
1a20ff27 | 5474 | #undef SCHED_DOMAIN_DEBUG |
1da177e4 LT |
5475 | #ifdef SCHED_DOMAIN_DEBUG |
5476 | static void sched_domain_debug(struct sched_domain *sd, int cpu) | |
5477 | { | |
5478 | int level = 0; | |
5479 | ||
41c7ce9a NP |
5480 | if (!sd) { |
5481 | printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); | |
5482 | return; | |
5483 | } | |
5484 | ||
1da177e4 LT |
5485 | printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); |
5486 | ||
5487 | do { | |
5488 | int i; | |
5489 | char str[NR_CPUS]; | |
5490 | struct sched_group *group = sd->groups; | |
5491 | cpumask_t groupmask; | |
5492 | ||
5493 | cpumask_scnprintf(str, NR_CPUS, sd->span); | |
5494 | cpus_clear(groupmask); | |
5495 | ||
5496 | printk(KERN_DEBUG); | |
5497 | for (i = 0; i < level + 1; i++) | |
5498 | printk(" "); | |
5499 | printk("domain %d: ", level); | |
5500 | ||
5501 | if (!(sd->flags & SD_LOAD_BALANCE)) { | |
5502 | printk("does not load-balance\n"); | |
5503 | if (sd->parent) | |
33859f7f MOS |
5504 | printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" |
5505 | " has parent"); | |
1da177e4 LT |
5506 | break; |
5507 | } | |
5508 | ||
5509 | printk("span %s\n", str); | |
5510 | ||
5511 | if (!cpu_isset(cpu, sd->span)) | |
33859f7f MOS |
5512 | printk(KERN_ERR "ERROR: domain->span does not contain " |
5513 | "CPU%d\n", cpu); | |
1da177e4 | 5514 | if (!cpu_isset(cpu, group->cpumask)) |
33859f7f MOS |
5515 | printk(KERN_ERR "ERROR: domain->groups does not contain" |
5516 | " CPU%d\n", cpu); | |
1da177e4 LT |
5517 | |
5518 | printk(KERN_DEBUG); | |
5519 | for (i = 0; i < level + 2; i++) | |
5520 | printk(" "); | |
5521 | printk("groups:"); | |
5522 | do { | |
5523 | if (!group) { | |
5524 | printk("\n"); | |
5525 | printk(KERN_ERR "ERROR: group is NULL\n"); | |
5526 | break; | |
5527 | } | |
5528 | ||
5517d86b | 5529 | if (!group->__cpu_power) { |
1da177e4 | 5530 | printk("\n"); |
33859f7f MOS |
5531 | printk(KERN_ERR "ERROR: domain->cpu_power not " |
5532 | "set\n"); | |
1da177e4 LT |
5533 | } |
5534 | ||
5535 | if (!cpus_weight(group->cpumask)) { | |
5536 | printk("\n"); | |
5537 | printk(KERN_ERR "ERROR: empty group\n"); | |
5538 | } | |
5539 | ||
5540 | if (cpus_intersects(groupmask, group->cpumask)) { | |
5541 | printk("\n"); | |
5542 | printk(KERN_ERR "ERROR: repeated CPUs\n"); | |
5543 | } | |
5544 | ||
5545 | cpus_or(groupmask, groupmask, group->cpumask); | |
5546 | ||
5547 | cpumask_scnprintf(str, NR_CPUS, group->cpumask); | |
5548 | printk(" %s", str); | |
5549 | ||
5550 | group = group->next; | |
5551 | } while (group != sd->groups); | |
5552 | printk("\n"); | |
5553 | ||
5554 | if (!cpus_equal(sd->span, groupmask)) | |
33859f7f MOS |
5555 | printk(KERN_ERR "ERROR: groups don't span " |
5556 | "domain->span\n"); | |
1da177e4 LT |
5557 | |
5558 | level++; | |
5559 | sd = sd->parent; | |
33859f7f MOS |
5560 | if (!sd) |
5561 | continue; | |
1da177e4 | 5562 | |
33859f7f MOS |
5563 | if (!cpus_subset(groupmask, sd->span)) |
5564 | printk(KERN_ERR "ERROR: parent span is not a superset " | |
5565 | "of domain->span\n"); | |
1da177e4 LT |
5566 | |
5567 | } while (sd); | |
5568 | } | |
5569 | #else | |
48f24c4d | 5570 | # define sched_domain_debug(sd, cpu) do { } while (0) |
1da177e4 LT |
5571 | #endif |
5572 | ||
1a20ff27 | 5573 | static int sd_degenerate(struct sched_domain *sd) |
245af2c7 SS |
5574 | { |
5575 | if (cpus_weight(sd->span) == 1) | |
5576 | return 1; | |
5577 | ||
5578 | /* Following flags need at least 2 groups */ | |
5579 | if (sd->flags & (SD_LOAD_BALANCE | | |
5580 | SD_BALANCE_NEWIDLE | | |
5581 | SD_BALANCE_FORK | | |
89c4710e SS |
5582 | SD_BALANCE_EXEC | |
5583 | SD_SHARE_CPUPOWER | | |
5584 | SD_SHARE_PKG_RESOURCES)) { | |
245af2c7 SS |
5585 | if (sd->groups != sd->groups->next) |
5586 | return 0; | |
5587 | } | |
5588 | ||
5589 | /* Following flags don't use groups */ | |
5590 | if (sd->flags & (SD_WAKE_IDLE | | |
5591 | SD_WAKE_AFFINE | | |
5592 | SD_WAKE_BALANCE)) | |
5593 | return 0; | |
5594 | ||
5595 | return 1; | |
5596 | } | |
5597 | ||
48f24c4d IM |
5598 | static int |
5599 | sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) | |
245af2c7 SS |
5600 | { |
5601 | unsigned long cflags = sd->flags, pflags = parent->flags; | |
5602 | ||
5603 | if (sd_degenerate(parent)) | |
5604 | return 1; | |
5605 | ||
5606 | if (!cpus_equal(sd->span, parent->span)) | |
5607 | return 0; | |
5608 | ||
5609 | /* Does parent contain flags not in child? */ | |
5610 | /* WAKE_BALANCE is a subset of WAKE_AFFINE */ | |
5611 | if (cflags & SD_WAKE_AFFINE) | |
5612 | pflags &= ~SD_WAKE_BALANCE; | |
5613 | /* Flags needing groups don't count if only 1 group in parent */ | |
5614 | if (parent->groups == parent->groups->next) { | |
5615 | pflags &= ~(SD_LOAD_BALANCE | | |
5616 | SD_BALANCE_NEWIDLE | | |
5617 | SD_BALANCE_FORK | | |
89c4710e SS |
5618 | SD_BALANCE_EXEC | |
5619 | SD_SHARE_CPUPOWER | | |
5620 | SD_SHARE_PKG_RESOURCES); | |
245af2c7 SS |
5621 | } |
5622 | if (~cflags & pflags) | |
5623 | return 0; | |
5624 | ||
5625 | return 1; | |
5626 | } | |
5627 | ||
1da177e4 LT |
5628 | /* |
5629 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must | |
5630 | * hold the hotplug lock. | |
5631 | */ | |
9c1cfda2 | 5632 | static void cpu_attach_domain(struct sched_domain *sd, int cpu) |
1da177e4 | 5633 | { |
70b97a7f | 5634 | struct rq *rq = cpu_rq(cpu); |
245af2c7 SS |
5635 | struct sched_domain *tmp; |
5636 | ||
5637 | /* Remove the sched domains which do not contribute to scheduling. */ | |
5638 | for (tmp = sd; tmp; tmp = tmp->parent) { | |
5639 | struct sched_domain *parent = tmp->parent; | |
5640 | if (!parent) | |
5641 | break; | |
1a848870 | 5642 | if (sd_parent_degenerate(tmp, parent)) { |
245af2c7 | 5643 | tmp->parent = parent->parent; |
1a848870 SS |
5644 | if (parent->parent) |
5645 | parent->parent->child = tmp; | |
5646 | } | |
245af2c7 SS |
5647 | } |
5648 | ||
1a848870 | 5649 | if (sd && sd_degenerate(sd)) { |
245af2c7 | 5650 | sd = sd->parent; |
1a848870 SS |
5651 | if (sd) |
5652 | sd->child = NULL; | |
5653 | } | |
1da177e4 LT |
5654 | |
5655 | sched_domain_debug(sd, cpu); | |
5656 | ||
674311d5 | 5657 | rcu_assign_pointer(rq->sd, sd); |
1da177e4 LT |
5658 | } |
5659 | ||
5660 | /* cpus with isolated domains */ | |
67af63a6 | 5661 | static cpumask_t cpu_isolated_map = CPU_MASK_NONE; |
1da177e4 LT |
5662 | |
5663 | /* Setup the mask of cpus configured for isolated domains */ | |
5664 | static int __init isolated_cpu_setup(char *str) | |
5665 | { | |
5666 | int ints[NR_CPUS], i; | |
5667 | ||
5668 | str = get_options(str, ARRAY_SIZE(ints), ints); | |
5669 | cpus_clear(cpu_isolated_map); | |
5670 | for (i = 1; i <= ints[0]; i++) | |
5671 | if (ints[i] < NR_CPUS) | |
5672 | cpu_set(ints[i], cpu_isolated_map); | |
5673 | return 1; | |
5674 | } | |
5675 | ||
5676 | __setup ("isolcpus=", isolated_cpu_setup); | |
5677 | ||
5678 | /* | |
6711cab4 SS |
5679 | * init_sched_build_groups takes the cpumask we wish to span, and a pointer |
5680 | * to a function which identifies what group(along with sched group) a CPU | |
5681 | * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS | |
5682 | * (due to the fact that we keep track of groups covered with a cpumask_t). | |
1da177e4 LT |
5683 | * |
5684 | * init_sched_build_groups will build a circular linked list of the groups | |
5685 | * covered by the given span, and will set each group's ->cpumask correctly, | |
5686 | * and ->cpu_power to 0. | |
5687 | */ | |
a616058b | 5688 | static void |
6711cab4 SS |
5689 | init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map, |
5690 | int (*group_fn)(int cpu, const cpumask_t *cpu_map, | |
5691 | struct sched_group **sg)) | |
1da177e4 LT |
5692 | { |
5693 | struct sched_group *first = NULL, *last = NULL; | |
5694 | cpumask_t covered = CPU_MASK_NONE; | |
5695 | int i; | |
5696 | ||
5697 | for_each_cpu_mask(i, span) { | |
6711cab4 SS |
5698 | struct sched_group *sg; |
5699 | int group = group_fn(i, cpu_map, &sg); | |
1da177e4 LT |
5700 | int j; |
5701 | ||
5702 | if (cpu_isset(i, covered)) | |
5703 | continue; | |
5704 | ||
5705 | sg->cpumask = CPU_MASK_NONE; | |
5517d86b | 5706 | sg->__cpu_power = 0; |
1da177e4 LT |
5707 | |
5708 | for_each_cpu_mask(j, span) { | |
6711cab4 | 5709 | if (group_fn(j, cpu_map, NULL) != group) |
1da177e4 LT |
5710 | continue; |
5711 | ||
5712 | cpu_set(j, covered); | |
5713 | cpu_set(j, sg->cpumask); | |
5714 | } | |
5715 | if (!first) | |
5716 | first = sg; | |
5717 | if (last) | |
5718 | last->next = sg; | |
5719 | last = sg; | |
5720 | } | |
5721 | last->next = first; | |
5722 | } | |
5723 | ||
9c1cfda2 | 5724 | #define SD_NODES_PER_DOMAIN 16 |
1da177e4 | 5725 | |
198e2f18 | 5726 | /* |
5727 | * Self-tuning task migration cost measurement between source and target CPUs. | |
5728 | * | |
5729 | * This is done by measuring the cost of manipulating buffers of varying | |
5730 | * sizes. For a given buffer-size here are the steps that are taken: | |
5731 | * | |
5732 | * 1) the source CPU reads+dirties a shared buffer | |
5733 | * 2) the target CPU reads+dirties the same shared buffer | |
5734 | * | |
5735 | * We measure how long they take, in the following 4 scenarios: | |
5736 | * | |
5737 | * - source: CPU1, target: CPU2 | cost1 | |
5738 | * - source: CPU2, target: CPU1 | cost2 | |
5739 | * - source: CPU1, target: CPU1 | cost3 | |
5740 | * - source: CPU2, target: CPU2 | cost4 | |
5741 | * | |
5742 | * We then calculate the cost3+cost4-cost1-cost2 difference - this is | |
5743 | * the cost of migration. | |
5744 | * | |
5745 | * We then start off from a small buffer-size and iterate up to larger | |
5746 | * buffer sizes, in 5% steps - measuring each buffer-size separately, and | |
5747 | * doing a maximum search for the cost. (The maximum cost for a migration | |
5748 | * normally occurs when the working set size is around the effective cache | |
5749 | * size.) | |
5750 | */ | |
5751 | #define SEARCH_SCOPE 2 | |
5752 | #define MIN_CACHE_SIZE (64*1024U) | |
5753 | #define DEFAULT_CACHE_SIZE (5*1024*1024U) | |
70b4d63e | 5754 | #define ITERATIONS 1 |
198e2f18 | 5755 | #define SIZE_THRESH 130 |
5756 | #define COST_THRESH 130 | |
5757 | ||
5758 | /* | |
5759 | * The migration cost is a function of 'domain distance'. Domain | |
5760 | * distance is the number of steps a CPU has to iterate down its | |
5761 | * domain tree to share a domain with the other CPU. The farther | |
5762 | * two CPUs are from each other, the larger the distance gets. | |
5763 | * | |
5764 | * Note that we use the distance only to cache measurement results, | |
5765 | * the distance value is not used numerically otherwise. When two | |
5766 | * CPUs have the same distance it is assumed that the migration | |
5767 | * cost is the same. (this is a simplification but quite practical) | |
5768 | */ | |
5769 | #define MAX_DOMAIN_DISTANCE 32 | |
5770 | ||
5771 | static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] = | |
4bbf39c2 IM |
5772 | { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] = |
5773 | /* | |
5774 | * Architectures may override the migration cost and thus avoid | |
5775 | * boot-time calibration. Unit is nanoseconds. Mostly useful for | |
5776 | * virtualized hardware: | |
5777 | */ | |
5778 | #ifdef CONFIG_DEFAULT_MIGRATION_COST | |
5779 | CONFIG_DEFAULT_MIGRATION_COST | |
5780 | #else | |
5781 | -1LL | |
5782 | #endif | |
5783 | }; | |
198e2f18 | 5784 | |
5785 | /* | |
5786 | * Allow override of migration cost - in units of microseconds. | |
5787 | * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost | |
5788 | * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs: | |
5789 | */ | |
5790 | static int __init migration_cost_setup(char *str) | |
5791 | { | |
5792 | int ints[MAX_DOMAIN_DISTANCE+1], i; | |
5793 | ||
5794 | str = get_options(str, ARRAY_SIZE(ints), ints); | |
5795 | ||
5796 | printk("#ints: %d\n", ints[0]); | |
5797 | for (i = 1; i <= ints[0]; i++) { | |
5798 | migration_cost[i-1] = (unsigned long long)ints[i]*1000; | |
5799 | printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]); | |
5800 | } | |
5801 | return 1; | |
5802 | } | |
5803 | ||
5804 | __setup ("migration_cost=", migration_cost_setup); | |
5805 | ||
5806 | /* | |
5807 | * Global multiplier (divisor) for migration-cutoff values, | |
5808 | * in percentiles. E.g. use a value of 150 to get 1.5 times | |
5809 | * longer cache-hot cutoff times. | |
5810 | * | |
5811 | * (We scale it from 100 to 128 to long long handling easier.) | |
5812 | */ | |
5813 | ||
5814 | #define MIGRATION_FACTOR_SCALE 128 | |
5815 | ||
5816 | static unsigned int migration_factor = MIGRATION_FACTOR_SCALE; | |
5817 | ||
5818 | static int __init setup_migration_factor(char *str) | |
5819 | { | |
5820 | get_option(&str, &migration_factor); | |
5821 | migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100; | |
5822 | return 1; | |
5823 | } | |
5824 | ||
5825 | __setup("migration_factor=", setup_migration_factor); | |
5826 | ||
5827 | /* | |
5828 | * Estimated distance of two CPUs, measured via the number of domains | |
5829 | * we have to pass for the two CPUs to be in the same span: | |
5830 | */ | |
5831 | static unsigned long domain_distance(int cpu1, int cpu2) | |
5832 | { | |
5833 | unsigned long distance = 0; | |
5834 | struct sched_domain *sd; | |
5835 | ||
5836 | for_each_domain(cpu1, sd) { | |
5837 | WARN_ON(!cpu_isset(cpu1, sd->span)); | |
5838 | if (cpu_isset(cpu2, sd->span)) | |
5839 | return distance; | |
5840 | distance++; | |
5841 | } | |
5842 | if (distance >= MAX_DOMAIN_DISTANCE) { | |
5843 | WARN_ON(1); | |
5844 | distance = MAX_DOMAIN_DISTANCE-1; | |
5845 | } | |
5846 | ||
5847 | return distance; | |
5848 | } | |
5849 | ||
5850 | static unsigned int migration_debug; | |
5851 | ||
5852 | static int __init setup_migration_debug(char *str) | |
5853 | { | |
5854 | get_option(&str, &migration_debug); | |
5855 | return 1; | |
5856 | } | |
5857 | ||
5858 | __setup("migration_debug=", setup_migration_debug); | |
5859 | ||
5860 | /* | |
5861 | * Maximum cache-size that the scheduler should try to measure. | |
5862 | * Architectures with larger caches should tune this up during | |
5863 | * bootup. Gets used in the domain-setup code (i.e. during SMP | |
5864 | * bootup). | |
5865 | */ | |
5866 | unsigned int max_cache_size; | |
5867 | ||
5868 | static int __init setup_max_cache_size(char *str) | |
5869 | { | |
5870 | get_option(&str, &max_cache_size); | |
5871 | return 1; | |
5872 | } | |
5873 | ||
5874 | __setup("max_cache_size=", setup_max_cache_size); | |
5875 | ||
5876 | /* | |
5877 | * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This | |
5878 | * is the operation that is timed, so we try to generate unpredictable | |
5879 | * cachemisses that still end up filling the L2 cache: | |
5880 | */ | |
5881 | static void touch_cache(void *__cache, unsigned long __size) | |
5882 | { | |
33859f7f MOS |
5883 | unsigned long size = __size / sizeof(long); |
5884 | unsigned long chunk1 = size / 3; | |
5885 | unsigned long chunk2 = 2 * size / 3; | |
198e2f18 | 5886 | unsigned long *cache = __cache; |
5887 | int i; | |
5888 | ||
5889 | for (i = 0; i < size/6; i += 8) { | |
5890 | switch (i % 6) { | |
5891 | case 0: cache[i]++; | |
5892 | case 1: cache[size-1-i]++; | |
5893 | case 2: cache[chunk1-i]++; | |
5894 | case 3: cache[chunk1+i]++; | |
5895 | case 4: cache[chunk2-i]++; | |
5896 | case 5: cache[chunk2+i]++; | |
5897 | } | |
5898 | } | |
5899 | } | |
5900 | ||
5901 | /* | |
5902 | * Measure the cache-cost of one task migration. Returns in units of nsec. | |
5903 | */ | |
48f24c4d IM |
5904 | static unsigned long long |
5905 | measure_one(void *cache, unsigned long size, int source, int target) | |
198e2f18 | 5906 | { |
5907 | cpumask_t mask, saved_mask; | |
5908 | unsigned long long t0, t1, t2, t3, cost; | |
5909 | ||
5910 | saved_mask = current->cpus_allowed; | |
5911 | ||
5912 | /* | |
5913 | * Flush source caches to RAM and invalidate them: | |
5914 | */ | |
5915 | sched_cacheflush(); | |
5916 | ||
5917 | /* | |
5918 | * Migrate to the source CPU: | |
5919 | */ | |
5920 | mask = cpumask_of_cpu(source); | |
5921 | set_cpus_allowed(current, mask); | |
5922 | WARN_ON(smp_processor_id() != source); | |
5923 | ||
5924 | /* | |
5925 | * Dirty the working set: | |
5926 | */ | |
5927 | t0 = sched_clock(); | |
5928 | touch_cache(cache, size); | |
5929 | t1 = sched_clock(); | |
5930 | ||
5931 | /* | |
5932 | * Migrate to the target CPU, dirty the L2 cache and access | |
5933 | * the shared buffer. (which represents the working set | |
5934 | * of a migrated task.) | |
5935 | */ | |
5936 | mask = cpumask_of_cpu(target); | |
5937 | set_cpus_allowed(current, mask); | |
5938 | WARN_ON(smp_processor_id() != target); | |
5939 | ||
5940 | t2 = sched_clock(); | |
5941 | touch_cache(cache, size); | |
5942 | t3 = sched_clock(); | |
5943 | ||
5944 | cost = t1-t0 + t3-t2; | |
5945 | ||
5946 | if (migration_debug >= 2) | |
5947 | printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n", | |
5948 | source, target, t1-t0, t1-t0, t3-t2, cost); | |
5949 | /* | |
5950 | * Flush target caches to RAM and invalidate them: | |
5951 | */ | |
5952 | sched_cacheflush(); | |
5953 | ||
5954 | set_cpus_allowed(current, saved_mask); | |
5955 | ||
5956 | return cost; | |
5957 | } | |
5958 | ||
5959 | /* | |
5960 | * Measure a series of task migrations and return the average | |
5961 | * result. Since this code runs early during bootup the system | |
5962 | * is 'undisturbed' and the average latency makes sense. | |
5963 | * | |
5964 | * The algorithm in essence auto-detects the relevant cache-size, | |
5965 | * so it will properly detect different cachesizes for different | |
5966 | * cache-hierarchies, depending on how the CPUs are connected. | |
5967 | * | |
5968 | * Architectures can prime the upper limit of the search range via | |
5969 | * max_cache_size, otherwise the search range defaults to 20MB...64K. | |
5970 | */ | |
5971 | static unsigned long long | |
5972 | measure_cost(int cpu1, int cpu2, void *cache, unsigned int size) | |
5973 | { | |
5974 | unsigned long long cost1, cost2; | |
5975 | int i; | |
5976 | ||
5977 | /* | |
5978 | * Measure the migration cost of 'size' bytes, over an | |
5979 | * average of 10 runs: | |
5980 | * | |
5981 | * (We perturb the cache size by a small (0..4k) | |
5982 | * value to compensate size/alignment related artifacts. | |
5983 | * We also subtract the cost of the operation done on | |
5984 | * the same CPU.) | |
5985 | */ | |
5986 | cost1 = 0; | |
5987 | ||
5988 | /* | |
5989 | * dry run, to make sure we start off cache-cold on cpu1, | |
5990 | * and to get any vmalloc pagefaults in advance: | |
5991 | */ | |
5992 | measure_one(cache, size, cpu1, cpu2); | |
5993 | for (i = 0; i < ITERATIONS; i++) | |
33859f7f | 5994 | cost1 += measure_one(cache, size - i * 1024, cpu1, cpu2); |
198e2f18 | 5995 | |
5996 | measure_one(cache, size, cpu2, cpu1); | |
5997 | for (i = 0; i < ITERATIONS; i++) | |
33859f7f | 5998 | cost1 += measure_one(cache, size - i * 1024, cpu2, cpu1); |
198e2f18 | 5999 | |
6000 | /* | |
6001 | * (We measure the non-migrating [cached] cost on both | |
6002 | * cpu1 and cpu2, to handle CPUs with different speeds) | |
6003 | */ | |
6004 | cost2 = 0; | |
6005 | ||
6006 | measure_one(cache, size, cpu1, cpu1); | |
6007 | for (i = 0; i < ITERATIONS; i++) | |
33859f7f | 6008 | cost2 += measure_one(cache, size - i * 1024, cpu1, cpu1); |
198e2f18 | 6009 | |
6010 | measure_one(cache, size, cpu2, cpu2); | |
6011 | for (i = 0; i < ITERATIONS; i++) | |
33859f7f | 6012 | cost2 += measure_one(cache, size - i * 1024, cpu2, cpu2); |
198e2f18 | 6013 | |
6014 | /* | |
6015 | * Get the per-iteration migration cost: | |
6016 | */ | |
33859f7f MOS |
6017 | do_div(cost1, 2 * ITERATIONS); |
6018 | do_div(cost2, 2 * ITERATIONS); | |
198e2f18 | 6019 | |
6020 | return cost1 - cost2; | |
6021 | } | |
6022 | ||
6023 | static unsigned long long measure_migration_cost(int cpu1, int cpu2) | |
6024 | { | |
6025 | unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0; | |
6026 | unsigned int max_size, size, size_found = 0; | |
6027 | long long cost = 0, prev_cost; | |
6028 | void *cache; | |
6029 | ||
6030 | /* | |
6031 | * Search from max_cache_size*5 down to 64K - the real relevant | |
6032 | * cachesize has to lie somewhere inbetween. | |
6033 | */ | |
6034 | if (max_cache_size) { | |
6035 | max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE); | |
6036 | size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE); | |
6037 | } else { | |
6038 | /* | |
6039 | * Since we have no estimation about the relevant | |
6040 | * search range | |
6041 | */ | |
6042 | max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE; | |
6043 | size = MIN_CACHE_SIZE; | |
6044 | } | |
6045 | ||
6046 | if (!cpu_online(cpu1) || !cpu_online(cpu2)) { | |
6047 | printk("cpu %d and %d not both online!\n", cpu1, cpu2); | |
6048 | return 0; | |
6049 | } | |
6050 | ||
6051 | /* | |
6052 | * Allocate the working set: | |
6053 | */ | |
6054 | cache = vmalloc(max_size); | |
6055 | if (!cache) { | |
33859f7f | 6056 | printk("could not vmalloc %d bytes for cache!\n", 2 * max_size); |
2ed6e34f | 6057 | return 1000000; /* return 1 msec on very small boxen */ |
198e2f18 | 6058 | } |
6059 | ||
6060 | while (size <= max_size) { | |
6061 | prev_cost = cost; | |
6062 | cost = measure_cost(cpu1, cpu2, cache, size); | |
6063 | ||
6064 | /* | |
6065 | * Update the max: | |
6066 | */ | |
6067 | if (cost > 0) { | |
6068 | if (max_cost < cost) { | |
6069 | max_cost = cost; | |
6070 | size_found = size; | |
6071 | } | |
6072 | } | |
6073 | /* | |
6074 | * Calculate average fluctuation, we use this to prevent | |
6075 | * noise from triggering an early break out of the loop: | |
6076 | */ | |
6077 | fluct = abs(cost - prev_cost); | |
6078 | avg_fluct = (avg_fluct + fluct)/2; | |
6079 | ||
6080 | if (migration_debug) | |
33859f7f MOS |
6081 | printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): " |
6082 | "(%8Ld %8Ld)\n", | |
198e2f18 | 6083 | cpu1, cpu2, size, |
6084 | (long)cost / 1000000, | |
6085 | ((long)cost / 100000) % 10, | |
6086 | (long)max_cost / 1000000, | |
6087 | ((long)max_cost / 100000) % 10, | |
6088 | domain_distance(cpu1, cpu2), | |
6089 | cost, avg_fluct); | |
6090 | ||
6091 | /* | |
6092 | * If we iterated at least 20% past the previous maximum, | |
6093 | * and the cost has dropped by more than 20% already, | |
6094 | * (taking fluctuations into account) then we assume to | |
6095 | * have found the maximum and break out of the loop early: | |
6096 | */ | |
6097 | if (size_found && (size*100 > size_found*SIZE_THRESH)) | |
6098 | if (cost+avg_fluct <= 0 || | |
6099 | max_cost*100 > (cost+avg_fluct)*COST_THRESH) { | |
6100 | ||
6101 | if (migration_debug) | |
6102 | printk("-> found max.\n"); | |
6103 | break; | |
6104 | } | |
6105 | /* | |
70b4d63e | 6106 | * Increase the cachesize in 10% steps: |
198e2f18 | 6107 | */ |
70b4d63e | 6108 | size = size * 10 / 9; |
198e2f18 | 6109 | } |
6110 | ||
6111 | if (migration_debug) | |
6112 | printk("[%d][%d] working set size found: %d, cost: %Ld\n", | |
6113 | cpu1, cpu2, size_found, max_cost); | |
6114 | ||
6115 | vfree(cache); | |
6116 | ||
6117 | /* | |
6118 | * A task is considered 'cache cold' if at least 2 times | |
6119 | * the worst-case cost of migration has passed. | |
6120 | * | |
6121 | * (this limit is only listened to if the load-balancing | |
6122 | * situation is 'nice' - if there is a large imbalance we | |
6123 | * ignore it for the sake of CPU utilization and | |
6124 | * processing fairness.) | |
6125 | */ | |
6126 | return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE; | |
6127 | } | |
6128 | ||
6129 | static void calibrate_migration_costs(const cpumask_t *cpu_map) | |
6130 | { | |
6131 | int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id(); | |
6132 | unsigned long j0, j1, distance, max_distance = 0; | |
6133 | struct sched_domain *sd; | |
6134 | ||
6135 | j0 = jiffies; | |
6136 | ||
6137 | /* | |
6138 | * First pass - calculate the cacheflush times: | |
6139 | */ | |
6140 | for_each_cpu_mask(cpu1, *cpu_map) { | |
6141 | for_each_cpu_mask(cpu2, *cpu_map) { | |
6142 | if (cpu1 == cpu2) | |
6143 | continue; | |
6144 | distance = domain_distance(cpu1, cpu2); | |
6145 | max_distance = max(max_distance, distance); | |
6146 | /* | |
6147 | * No result cached yet? | |
6148 | */ | |
6149 | if (migration_cost[distance] == -1LL) | |
6150 | migration_cost[distance] = | |
6151 | measure_migration_cost(cpu1, cpu2); | |
6152 | } | |
6153 | } | |
6154 | /* | |
6155 | * Second pass - update the sched domain hierarchy with | |
6156 | * the new cache-hot-time estimations: | |
6157 | */ | |
6158 | for_each_cpu_mask(cpu, *cpu_map) { | |
6159 | distance = 0; | |
6160 | for_each_domain(cpu, sd) { | |
6161 | sd->cache_hot_time = migration_cost[distance]; | |
6162 | distance++; | |
6163 | } | |
6164 | } | |
6165 | /* | |
6166 | * Print the matrix: | |
6167 | */ | |
6168 | if (migration_debug) | |
6169 | printk("migration: max_cache_size: %d, cpu: %d MHz:\n", | |
6170 | max_cache_size, | |
6171 | #ifdef CONFIG_X86 | |
6172 | cpu_khz/1000 | |
6173 | #else | |
6174 | -1 | |
6175 | #endif | |
6176 | ); | |
33859f7f MOS |
6177 | if (system_state == SYSTEM_BOOTING && num_online_cpus() > 1) { |
6178 | printk("migration_cost="); | |
6179 | for (distance = 0; distance <= max_distance; distance++) { | |
6180 | if (distance) | |
6181 | printk(","); | |
6182 | printk("%ld", (long)migration_cost[distance] / 1000); | |
bd576c95 | 6183 | } |
33859f7f | 6184 | printk("\n"); |
198e2f18 | 6185 | } |
198e2f18 | 6186 | j1 = jiffies; |
6187 | if (migration_debug) | |
33859f7f | 6188 | printk("migration: %ld seconds\n", (j1-j0) / HZ); |
198e2f18 | 6189 | |
6190 | /* | |
6191 | * Move back to the original CPU. NUMA-Q gets confused | |
6192 | * if we migrate to another quad during bootup. | |
6193 | */ | |
6194 | if (raw_smp_processor_id() != orig_cpu) { | |
6195 | cpumask_t mask = cpumask_of_cpu(orig_cpu), | |
6196 | saved_mask = current->cpus_allowed; | |
6197 | ||
6198 | set_cpus_allowed(current, mask); | |
6199 | set_cpus_allowed(current, saved_mask); | |
6200 | } | |
6201 | } | |
6202 | ||
9c1cfda2 | 6203 | #ifdef CONFIG_NUMA |
198e2f18 | 6204 | |
9c1cfda2 JH |
6205 | /** |
6206 | * find_next_best_node - find the next node to include in a sched_domain | |
6207 | * @node: node whose sched_domain we're building | |
6208 | * @used_nodes: nodes already in the sched_domain | |
6209 | * | |
6210 | * Find the next node to include in a given scheduling domain. Simply | |
6211 | * finds the closest node not already in the @used_nodes map. | |
6212 | * | |
6213 | * Should use nodemask_t. | |
6214 | */ | |
6215 | static int find_next_best_node(int node, unsigned long *used_nodes) | |
6216 | { | |
6217 | int i, n, val, min_val, best_node = 0; | |
6218 | ||
6219 | min_val = INT_MAX; | |
6220 | ||
6221 | for (i = 0; i < MAX_NUMNODES; i++) { | |
6222 | /* Start at @node */ | |
6223 | n = (node + i) % MAX_NUMNODES; | |
6224 | ||
6225 | if (!nr_cpus_node(n)) | |
6226 | continue; | |
6227 | ||
6228 | /* Skip already used nodes */ | |
6229 | if (test_bit(n, used_nodes)) | |
6230 | continue; | |
6231 | ||
6232 | /* Simple min distance search */ | |
6233 | val = node_distance(node, n); | |
6234 | ||
6235 | if (val < min_val) { | |
6236 | min_val = val; | |
6237 | best_node = n; | |
6238 | } | |
6239 | } | |
6240 | ||
6241 | set_bit(best_node, used_nodes); | |
6242 | return best_node; | |
6243 | } | |
6244 | ||
6245 | /** | |
6246 | * sched_domain_node_span - get a cpumask for a node's sched_domain | |
6247 | * @node: node whose cpumask we're constructing | |
6248 | * @size: number of nodes to include in this span | |
6249 | * | |
6250 | * Given a node, construct a good cpumask for its sched_domain to span. It | |
6251 | * should be one that prevents unnecessary balancing, but also spreads tasks | |
6252 | * out optimally. | |
6253 | */ | |
6254 | static cpumask_t sched_domain_node_span(int node) | |
6255 | { | |
9c1cfda2 | 6256 | DECLARE_BITMAP(used_nodes, MAX_NUMNODES); |
48f24c4d IM |
6257 | cpumask_t span, nodemask; |
6258 | int i; | |
9c1cfda2 JH |
6259 | |
6260 | cpus_clear(span); | |
6261 | bitmap_zero(used_nodes, MAX_NUMNODES); | |
6262 | ||
6263 | nodemask = node_to_cpumask(node); | |
6264 | cpus_or(span, span, nodemask); | |
6265 | set_bit(node, used_nodes); | |
6266 | ||
6267 | for (i = 1; i < SD_NODES_PER_DOMAIN; i++) { | |
6268 | int next_node = find_next_best_node(node, used_nodes); | |
48f24c4d | 6269 | |
9c1cfda2 JH |
6270 | nodemask = node_to_cpumask(next_node); |
6271 | cpus_or(span, span, nodemask); | |
6272 | } | |
6273 | ||
6274 | return span; | |
6275 | } | |
6276 | #endif | |
6277 | ||
5c45bf27 | 6278 | int sched_smt_power_savings = 0, sched_mc_power_savings = 0; |
48f24c4d | 6279 | |
9c1cfda2 | 6280 | /* |
48f24c4d | 6281 | * SMT sched-domains: |
9c1cfda2 | 6282 | */ |
1da177e4 LT |
6283 | #ifdef CONFIG_SCHED_SMT |
6284 | static DEFINE_PER_CPU(struct sched_domain, cpu_domains); | |
6711cab4 | 6285 | static DEFINE_PER_CPU(struct sched_group, sched_group_cpus); |
48f24c4d | 6286 | |
6711cab4 SS |
6287 | static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, |
6288 | struct sched_group **sg) | |
1da177e4 | 6289 | { |
6711cab4 SS |
6290 | if (sg) |
6291 | *sg = &per_cpu(sched_group_cpus, cpu); | |
1da177e4 LT |
6292 | return cpu; |
6293 | } | |
6294 | #endif | |
6295 | ||
48f24c4d IM |
6296 | /* |
6297 | * multi-core sched-domains: | |
6298 | */ | |
1e9f28fa SS |
6299 | #ifdef CONFIG_SCHED_MC |
6300 | static DEFINE_PER_CPU(struct sched_domain, core_domains); | |
6711cab4 | 6301 | static DEFINE_PER_CPU(struct sched_group, sched_group_core); |
1e9f28fa SS |
6302 | #endif |
6303 | ||
6304 | #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT) | |
6711cab4 SS |
6305 | static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map, |
6306 | struct sched_group **sg) | |
1e9f28fa | 6307 | { |
6711cab4 | 6308 | int group; |
a616058b SS |
6309 | cpumask_t mask = cpu_sibling_map[cpu]; |
6310 | cpus_and(mask, mask, *cpu_map); | |
6711cab4 SS |
6311 | group = first_cpu(mask); |
6312 | if (sg) | |
6313 | *sg = &per_cpu(sched_group_core, group); | |
6314 | return group; | |
1e9f28fa SS |
6315 | } |
6316 | #elif defined(CONFIG_SCHED_MC) | |
6711cab4 SS |
6317 | static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map, |
6318 | struct sched_group **sg) | |
1e9f28fa | 6319 | { |
6711cab4 SS |
6320 | if (sg) |
6321 | *sg = &per_cpu(sched_group_core, cpu); | |
1e9f28fa SS |
6322 | return cpu; |
6323 | } | |
6324 | #endif | |
6325 | ||
1da177e4 | 6326 | static DEFINE_PER_CPU(struct sched_domain, phys_domains); |
6711cab4 | 6327 | static DEFINE_PER_CPU(struct sched_group, sched_group_phys); |
48f24c4d | 6328 | |
6711cab4 SS |
6329 | static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, |
6330 | struct sched_group **sg) | |
1da177e4 | 6331 | { |
6711cab4 | 6332 | int group; |
48f24c4d | 6333 | #ifdef CONFIG_SCHED_MC |
1e9f28fa | 6334 | cpumask_t mask = cpu_coregroup_map(cpu); |
a616058b | 6335 | cpus_and(mask, mask, *cpu_map); |
6711cab4 | 6336 | group = first_cpu(mask); |
1e9f28fa | 6337 | #elif defined(CONFIG_SCHED_SMT) |
a616058b SS |
6338 | cpumask_t mask = cpu_sibling_map[cpu]; |
6339 | cpus_and(mask, mask, *cpu_map); | |
6711cab4 | 6340 | group = first_cpu(mask); |
1da177e4 | 6341 | #else |
6711cab4 | 6342 | group = cpu; |
1da177e4 | 6343 | #endif |
6711cab4 SS |
6344 | if (sg) |
6345 | *sg = &per_cpu(sched_group_phys, group); | |
6346 | return group; | |
1da177e4 LT |
6347 | } |
6348 | ||
6349 | #ifdef CONFIG_NUMA | |
1da177e4 | 6350 | /* |
9c1cfda2 JH |
6351 | * The init_sched_build_groups can't handle what we want to do with node |
6352 | * groups, so roll our own. Now each node has its own list of groups which | |
6353 | * gets dynamically allocated. | |
1da177e4 | 6354 | */ |
9c1cfda2 | 6355 | static DEFINE_PER_CPU(struct sched_domain, node_domains); |
d1b55138 | 6356 | static struct sched_group **sched_group_nodes_bycpu[NR_CPUS]; |
1da177e4 | 6357 | |
9c1cfda2 | 6358 | static DEFINE_PER_CPU(struct sched_domain, allnodes_domains); |
6711cab4 | 6359 | static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes); |
9c1cfda2 | 6360 | |
6711cab4 SS |
6361 | static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map, |
6362 | struct sched_group **sg) | |
9c1cfda2 | 6363 | { |
6711cab4 SS |
6364 | cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu)); |
6365 | int group; | |
6366 | ||
6367 | cpus_and(nodemask, nodemask, *cpu_map); | |
6368 | group = first_cpu(nodemask); | |
6369 | ||
6370 | if (sg) | |
6371 | *sg = &per_cpu(sched_group_allnodes, group); | |
6372 | return group; | |
1da177e4 | 6373 | } |
6711cab4 | 6374 | |
08069033 SS |
6375 | static void init_numa_sched_groups_power(struct sched_group *group_head) |
6376 | { | |
6377 | struct sched_group *sg = group_head; | |
6378 | int j; | |
6379 | ||
6380 | if (!sg) | |
6381 | return; | |
6382 | next_sg: | |
6383 | for_each_cpu_mask(j, sg->cpumask) { | |
6384 | struct sched_domain *sd; | |
6385 | ||
6386 | sd = &per_cpu(phys_domains, j); | |
6387 | if (j != first_cpu(sd->groups->cpumask)) { | |
6388 | /* | |
6389 | * Only add "power" once for each | |
6390 | * physical package. | |
6391 | */ | |
6392 | continue; | |
6393 | } | |
6394 | ||
5517d86b | 6395 | sg_inc_cpu_power(sg, sd->groups->__cpu_power); |
08069033 SS |
6396 | } |
6397 | sg = sg->next; | |
6398 | if (sg != group_head) | |
6399 | goto next_sg; | |
6400 | } | |
1da177e4 LT |
6401 | #endif |
6402 | ||
a616058b | 6403 | #ifdef CONFIG_NUMA |
51888ca2 SV |
6404 | /* Free memory allocated for various sched_group structures */ |
6405 | static void free_sched_groups(const cpumask_t *cpu_map) | |
6406 | { | |
a616058b | 6407 | int cpu, i; |
51888ca2 SV |
6408 | |
6409 | for_each_cpu_mask(cpu, *cpu_map) { | |
51888ca2 SV |
6410 | struct sched_group **sched_group_nodes |
6411 | = sched_group_nodes_bycpu[cpu]; | |
6412 | ||
51888ca2 SV |
6413 | if (!sched_group_nodes) |
6414 | continue; | |
6415 | ||
6416 | for (i = 0; i < MAX_NUMNODES; i++) { | |
6417 | cpumask_t nodemask = node_to_cpumask(i); | |
6418 | struct sched_group *oldsg, *sg = sched_group_nodes[i]; | |
6419 | ||
6420 | cpus_and(nodemask, nodemask, *cpu_map); | |
6421 | if (cpus_empty(nodemask)) | |
6422 | continue; | |
6423 | ||
6424 | if (sg == NULL) | |
6425 | continue; | |
6426 | sg = sg->next; | |
6427 | next_sg: | |
6428 | oldsg = sg; | |
6429 | sg = sg->next; | |
6430 | kfree(oldsg); | |
6431 | if (oldsg != sched_group_nodes[i]) | |
6432 | goto next_sg; | |
6433 | } | |
6434 | kfree(sched_group_nodes); | |
6435 | sched_group_nodes_bycpu[cpu] = NULL; | |
6436 | } | |
51888ca2 | 6437 | } |
a616058b SS |
6438 | #else |
6439 | static void free_sched_groups(const cpumask_t *cpu_map) | |
6440 | { | |
6441 | } | |
6442 | #endif | |
51888ca2 | 6443 | |
89c4710e SS |
6444 | /* |
6445 | * Initialize sched groups cpu_power. | |
6446 | * | |
6447 | * cpu_power indicates the capacity of sched group, which is used while | |
6448 | * distributing the load between different sched groups in a sched domain. | |
6449 | * Typically cpu_power for all the groups in a sched domain will be same unless | |
6450 | * there are asymmetries in the topology. If there are asymmetries, group | |
6451 | * having more cpu_power will pickup more load compared to the group having | |
6452 | * less cpu_power. | |
6453 | * | |
6454 | * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents | |
6455 | * the maximum number of tasks a group can handle in the presence of other idle | |
6456 | * or lightly loaded groups in the same sched domain. | |
6457 | */ | |
6458 | static void init_sched_groups_power(int cpu, struct sched_domain *sd) | |
6459 | { | |
6460 | struct sched_domain *child; | |
6461 | struct sched_group *group; | |
6462 | ||
6463 | WARN_ON(!sd || !sd->groups); | |
6464 | ||
6465 | if (cpu != first_cpu(sd->groups->cpumask)) | |
6466 | return; | |
6467 | ||
6468 | child = sd->child; | |
6469 | ||
5517d86b ED |
6470 | sd->groups->__cpu_power = 0; |
6471 | ||
89c4710e SS |
6472 | /* |
6473 | * For perf policy, if the groups in child domain share resources | |
6474 | * (for example cores sharing some portions of the cache hierarchy | |
6475 | * or SMT), then set this domain groups cpu_power such that each group | |
6476 | * can handle only one task, when there are other idle groups in the | |
6477 | * same sched domain. | |
6478 | */ | |
6479 | if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) && | |
6480 | (child->flags & | |
6481 | (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) { | |
5517d86b | 6482 | sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE); |
89c4710e SS |
6483 | return; |
6484 | } | |
6485 | ||
89c4710e SS |
6486 | /* |
6487 | * add cpu_power of each child group to this groups cpu_power | |
6488 | */ | |
6489 | group = child->groups; | |
6490 | do { | |
5517d86b | 6491 | sg_inc_cpu_power(sd->groups, group->__cpu_power); |
89c4710e SS |
6492 | group = group->next; |
6493 | } while (group != child->groups); | |
6494 | } | |
6495 | ||
1da177e4 | 6496 | /* |
1a20ff27 DG |
6497 | * Build sched domains for a given set of cpus and attach the sched domains |
6498 | * to the individual cpus | |
1da177e4 | 6499 | */ |
51888ca2 | 6500 | static int build_sched_domains(const cpumask_t *cpu_map) |
1da177e4 LT |
6501 | { |
6502 | int i; | |
89c4710e | 6503 | struct sched_domain *sd; |
d1b55138 JH |
6504 | #ifdef CONFIG_NUMA |
6505 | struct sched_group **sched_group_nodes = NULL; | |
6711cab4 | 6506 | int sd_allnodes = 0; |
d1b55138 JH |
6507 | |
6508 | /* | |
6509 | * Allocate the per-node list of sched groups | |
6510 | */ | |
51888ca2 | 6511 | sched_group_nodes = kzalloc(sizeof(struct sched_group*)*MAX_NUMNODES, |
d3a5aa98 | 6512 | GFP_KERNEL); |
d1b55138 JH |
6513 | if (!sched_group_nodes) { |
6514 | printk(KERN_WARNING "Can not alloc sched group node list\n"); | |
51888ca2 | 6515 | return -ENOMEM; |
d1b55138 JH |
6516 | } |
6517 | sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes; | |
6518 | #endif | |
1da177e4 LT |
6519 | |
6520 | /* | |
1a20ff27 | 6521 | * Set up domains for cpus specified by the cpu_map. |
1da177e4 | 6522 | */ |
1a20ff27 | 6523 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 LT |
6524 | struct sched_domain *sd = NULL, *p; |
6525 | cpumask_t nodemask = node_to_cpumask(cpu_to_node(i)); | |
6526 | ||
1a20ff27 | 6527 | cpus_and(nodemask, nodemask, *cpu_map); |
1da177e4 LT |
6528 | |
6529 | #ifdef CONFIG_NUMA | |
d1b55138 | 6530 | if (cpus_weight(*cpu_map) |
9c1cfda2 JH |
6531 | > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) { |
6532 | sd = &per_cpu(allnodes_domains, i); | |
6533 | *sd = SD_ALLNODES_INIT; | |
6534 | sd->span = *cpu_map; | |
6711cab4 | 6535 | cpu_to_allnodes_group(i, cpu_map, &sd->groups); |
9c1cfda2 | 6536 | p = sd; |
6711cab4 | 6537 | sd_allnodes = 1; |
9c1cfda2 JH |
6538 | } else |
6539 | p = NULL; | |
6540 | ||
1da177e4 | 6541 | sd = &per_cpu(node_domains, i); |
1da177e4 | 6542 | *sd = SD_NODE_INIT; |
9c1cfda2 JH |
6543 | sd->span = sched_domain_node_span(cpu_to_node(i)); |
6544 | sd->parent = p; | |
1a848870 SS |
6545 | if (p) |
6546 | p->child = sd; | |
9c1cfda2 | 6547 | cpus_and(sd->span, sd->span, *cpu_map); |
1da177e4 LT |
6548 | #endif |
6549 | ||
6550 | p = sd; | |
6551 | sd = &per_cpu(phys_domains, i); | |
1da177e4 LT |
6552 | *sd = SD_CPU_INIT; |
6553 | sd->span = nodemask; | |
6554 | sd->parent = p; | |
1a848870 SS |
6555 | if (p) |
6556 | p->child = sd; | |
6711cab4 | 6557 | cpu_to_phys_group(i, cpu_map, &sd->groups); |
1da177e4 | 6558 | |
1e9f28fa SS |
6559 | #ifdef CONFIG_SCHED_MC |
6560 | p = sd; | |
6561 | sd = &per_cpu(core_domains, i); | |
1e9f28fa SS |
6562 | *sd = SD_MC_INIT; |
6563 | sd->span = cpu_coregroup_map(i); | |
6564 | cpus_and(sd->span, sd->span, *cpu_map); | |
6565 | sd->parent = p; | |
1a848870 | 6566 | p->child = sd; |
6711cab4 | 6567 | cpu_to_core_group(i, cpu_map, &sd->groups); |
1e9f28fa SS |
6568 | #endif |
6569 | ||
1da177e4 LT |
6570 | #ifdef CONFIG_SCHED_SMT |
6571 | p = sd; | |
6572 | sd = &per_cpu(cpu_domains, i); | |
1da177e4 LT |
6573 | *sd = SD_SIBLING_INIT; |
6574 | sd->span = cpu_sibling_map[i]; | |
1a20ff27 | 6575 | cpus_and(sd->span, sd->span, *cpu_map); |
1da177e4 | 6576 | sd->parent = p; |
1a848870 | 6577 | p->child = sd; |
6711cab4 | 6578 | cpu_to_cpu_group(i, cpu_map, &sd->groups); |
1da177e4 LT |
6579 | #endif |
6580 | } | |
6581 | ||
6582 | #ifdef CONFIG_SCHED_SMT | |
6583 | /* Set up CPU (sibling) groups */ | |
9c1cfda2 | 6584 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 | 6585 | cpumask_t this_sibling_map = cpu_sibling_map[i]; |
1a20ff27 | 6586 | cpus_and(this_sibling_map, this_sibling_map, *cpu_map); |
1da177e4 LT |
6587 | if (i != first_cpu(this_sibling_map)) |
6588 | continue; | |
6589 | ||
6711cab4 | 6590 | init_sched_build_groups(this_sibling_map, cpu_map, &cpu_to_cpu_group); |
1da177e4 LT |
6591 | } |
6592 | #endif | |
6593 | ||
1e9f28fa SS |
6594 | #ifdef CONFIG_SCHED_MC |
6595 | /* Set up multi-core groups */ | |
6596 | for_each_cpu_mask(i, *cpu_map) { | |
6597 | cpumask_t this_core_map = cpu_coregroup_map(i); | |
6598 | cpus_and(this_core_map, this_core_map, *cpu_map); | |
6599 | if (i != first_cpu(this_core_map)) | |
6600 | continue; | |
6711cab4 | 6601 | init_sched_build_groups(this_core_map, cpu_map, &cpu_to_core_group); |
1e9f28fa SS |
6602 | } |
6603 | #endif | |
6604 | ||
6605 | ||
1da177e4 LT |
6606 | /* Set up physical groups */ |
6607 | for (i = 0; i < MAX_NUMNODES; i++) { | |
6608 | cpumask_t nodemask = node_to_cpumask(i); | |
6609 | ||
1a20ff27 | 6610 | cpus_and(nodemask, nodemask, *cpu_map); |
1da177e4 LT |
6611 | if (cpus_empty(nodemask)) |
6612 | continue; | |
6613 | ||
6711cab4 | 6614 | init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group); |
1da177e4 LT |
6615 | } |
6616 | ||
6617 | #ifdef CONFIG_NUMA | |
6618 | /* Set up node groups */ | |
6711cab4 SS |
6619 | if (sd_allnodes) |
6620 | init_sched_build_groups(*cpu_map, cpu_map, &cpu_to_allnodes_group); | |
9c1cfda2 JH |
6621 | |
6622 | for (i = 0; i < MAX_NUMNODES; i++) { | |
6623 | /* Set up node groups */ | |
6624 | struct sched_group *sg, *prev; | |
6625 | cpumask_t nodemask = node_to_cpumask(i); | |
6626 | cpumask_t domainspan; | |
6627 | cpumask_t covered = CPU_MASK_NONE; | |
6628 | int j; | |
6629 | ||
6630 | cpus_and(nodemask, nodemask, *cpu_map); | |
d1b55138 JH |
6631 | if (cpus_empty(nodemask)) { |
6632 | sched_group_nodes[i] = NULL; | |
9c1cfda2 | 6633 | continue; |
d1b55138 | 6634 | } |
9c1cfda2 JH |
6635 | |
6636 | domainspan = sched_domain_node_span(i); | |
6637 | cpus_and(domainspan, domainspan, *cpu_map); | |
6638 | ||
15f0b676 | 6639 | sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i); |
51888ca2 SV |
6640 | if (!sg) { |
6641 | printk(KERN_WARNING "Can not alloc domain group for " | |
6642 | "node %d\n", i); | |
6643 | goto error; | |
6644 | } | |
9c1cfda2 JH |
6645 | sched_group_nodes[i] = sg; |
6646 | for_each_cpu_mask(j, nodemask) { | |
6647 | struct sched_domain *sd; | |
6648 | sd = &per_cpu(node_domains, j); | |
6649 | sd->groups = sg; | |
9c1cfda2 | 6650 | } |
5517d86b | 6651 | sg->__cpu_power = 0; |
9c1cfda2 | 6652 | sg->cpumask = nodemask; |
51888ca2 | 6653 | sg->next = sg; |
9c1cfda2 JH |
6654 | cpus_or(covered, covered, nodemask); |
6655 | prev = sg; | |
6656 | ||
6657 | for (j = 0; j < MAX_NUMNODES; j++) { | |
6658 | cpumask_t tmp, notcovered; | |
6659 | int n = (i + j) % MAX_NUMNODES; | |
6660 | ||
6661 | cpus_complement(notcovered, covered); | |
6662 | cpus_and(tmp, notcovered, *cpu_map); | |
6663 | cpus_and(tmp, tmp, domainspan); | |
6664 | if (cpus_empty(tmp)) | |
6665 | break; | |
6666 | ||
6667 | nodemask = node_to_cpumask(n); | |
6668 | cpus_and(tmp, tmp, nodemask); | |
6669 | if (cpus_empty(tmp)) | |
6670 | continue; | |
6671 | ||
15f0b676 SV |
6672 | sg = kmalloc_node(sizeof(struct sched_group), |
6673 | GFP_KERNEL, i); | |
9c1cfda2 JH |
6674 | if (!sg) { |
6675 | printk(KERN_WARNING | |
6676 | "Can not alloc domain group for node %d\n", j); | |
51888ca2 | 6677 | goto error; |
9c1cfda2 | 6678 | } |
5517d86b | 6679 | sg->__cpu_power = 0; |
9c1cfda2 | 6680 | sg->cpumask = tmp; |
51888ca2 | 6681 | sg->next = prev->next; |
9c1cfda2 JH |
6682 | cpus_or(covered, covered, tmp); |
6683 | prev->next = sg; | |
6684 | prev = sg; | |
6685 | } | |
9c1cfda2 | 6686 | } |
1da177e4 LT |
6687 | #endif |
6688 | ||
6689 | /* Calculate CPU power for physical packages and nodes */ | |
5c45bf27 | 6690 | #ifdef CONFIG_SCHED_SMT |
1a20ff27 | 6691 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 | 6692 | sd = &per_cpu(cpu_domains, i); |
89c4710e | 6693 | init_sched_groups_power(i, sd); |
5c45bf27 | 6694 | } |
1da177e4 | 6695 | #endif |
1e9f28fa | 6696 | #ifdef CONFIG_SCHED_MC |
5c45bf27 | 6697 | for_each_cpu_mask(i, *cpu_map) { |
1e9f28fa | 6698 | sd = &per_cpu(core_domains, i); |
89c4710e | 6699 | init_sched_groups_power(i, sd); |
5c45bf27 SS |
6700 | } |
6701 | #endif | |
1e9f28fa | 6702 | |
5c45bf27 | 6703 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 | 6704 | sd = &per_cpu(phys_domains, i); |
89c4710e | 6705 | init_sched_groups_power(i, sd); |
1da177e4 LT |
6706 | } |
6707 | ||
9c1cfda2 | 6708 | #ifdef CONFIG_NUMA |
08069033 SS |
6709 | for (i = 0; i < MAX_NUMNODES; i++) |
6710 | init_numa_sched_groups_power(sched_group_nodes[i]); | |
9c1cfda2 | 6711 | |
6711cab4 SS |
6712 | if (sd_allnodes) { |
6713 | struct sched_group *sg; | |
f712c0c7 | 6714 | |
6711cab4 | 6715 | cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg); |
f712c0c7 SS |
6716 | init_numa_sched_groups_power(sg); |
6717 | } | |
9c1cfda2 JH |
6718 | #endif |
6719 | ||
1da177e4 | 6720 | /* Attach the domains */ |
1a20ff27 | 6721 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 LT |
6722 | struct sched_domain *sd; |
6723 | #ifdef CONFIG_SCHED_SMT | |
6724 | sd = &per_cpu(cpu_domains, i); | |
1e9f28fa SS |
6725 | #elif defined(CONFIG_SCHED_MC) |
6726 | sd = &per_cpu(core_domains, i); | |
1da177e4 LT |
6727 | #else |
6728 | sd = &per_cpu(phys_domains, i); | |
6729 | #endif | |
6730 | cpu_attach_domain(sd, i); | |
6731 | } | |
198e2f18 | 6732 | /* |
6733 | * Tune cache-hot values: | |
6734 | */ | |
6735 | calibrate_migration_costs(cpu_map); | |
51888ca2 SV |
6736 | |
6737 | return 0; | |
6738 | ||
a616058b | 6739 | #ifdef CONFIG_NUMA |
51888ca2 SV |
6740 | error: |
6741 | free_sched_groups(cpu_map); | |
6742 | return -ENOMEM; | |
a616058b | 6743 | #endif |
1da177e4 | 6744 | } |
1a20ff27 DG |
6745 | /* |
6746 | * Set up scheduler domains and groups. Callers must hold the hotplug lock. | |
6747 | */ | |
51888ca2 | 6748 | static int arch_init_sched_domains(const cpumask_t *cpu_map) |
1a20ff27 DG |
6749 | { |
6750 | cpumask_t cpu_default_map; | |
51888ca2 | 6751 | int err; |
1da177e4 | 6752 | |
1a20ff27 DG |
6753 | /* |
6754 | * Setup mask for cpus without special case scheduling requirements. | |
6755 | * For now this just excludes isolated cpus, but could be used to | |
6756 | * exclude other special cases in the future. | |
6757 | */ | |
6758 | cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map); | |
6759 | ||
51888ca2 SV |
6760 | err = build_sched_domains(&cpu_default_map); |
6761 | ||
6762 | return err; | |
1a20ff27 DG |
6763 | } |
6764 | ||
6765 | static void arch_destroy_sched_domains(const cpumask_t *cpu_map) | |
1da177e4 | 6766 | { |
51888ca2 | 6767 | free_sched_groups(cpu_map); |
9c1cfda2 | 6768 | } |
1da177e4 | 6769 | |
1a20ff27 DG |
6770 | /* |
6771 | * Detach sched domains from a group of cpus specified in cpu_map | |
6772 | * These cpus will now be attached to the NULL domain | |
6773 | */ | |
858119e1 | 6774 | static void detach_destroy_domains(const cpumask_t *cpu_map) |
1a20ff27 DG |
6775 | { |
6776 | int i; | |
6777 | ||
6778 | for_each_cpu_mask(i, *cpu_map) | |
6779 | cpu_attach_domain(NULL, i); | |
6780 | synchronize_sched(); | |
6781 | arch_destroy_sched_domains(cpu_map); | |
6782 | } | |
6783 | ||
6784 | /* | |
6785 | * Partition sched domains as specified by the cpumasks below. | |
6786 | * This attaches all cpus from the cpumasks to the NULL domain, | |
6787 | * waits for a RCU quiescent period, recalculates sched | |
6788 | * domain information and then attaches them back to the | |
6789 | * correct sched domains | |
6790 | * Call with hotplug lock held | |
6791 | */ | |
51888ca2 | 6792 | int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2) |
1a20ff27 DG |
6793 | { |
6794 | cpumask_t change_map; | |
51888ca2 | 6795 | int err = 0; |
1a20ff27 DG |
6796 | |
6797 | cpus_and(*partition1, *partition1, cpu_online_map); | |
6798 | cpus_and(*partition2, *partition2, cpu_online_map); | |
6799 | cpus_or(change_map, *partition1, *partition2); | |
6800 | ||
6801 | /* Detach sched domains from all of the affected cpus */ | |
6802 | detach_destroy_domains(&change_map); | |
6803 | if (!cpus_empty(*partition1)) | |
51888ca2 SV |
6804 | err = build_sched_domains(partition1); |
6805 | if (!err && !cpus_empty(*partition2)) | |
6806 | err = build_sched_domains(partition2); | |
6807 | ||
6808 | return err; | |
1a20ff27 DG |
6809 | } |
6810 | ||
5c45bf27 SS |
6811 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
6812 | int arch_reinit_sched_domains(void) | |
6813 | { | |
6814 | int err; | |
6815 | ||
6816 | lock_cpu_hotplug(); | |
6817 | detach_destroy_domains(&cpu_online_map); | |
6818 | err = arch_init_sched_domains(&cpu_online_map); | |
6819 | unlock_cpu_hotplug(); | |
6820 | ||
6821 | return err; | |
6822 | } | |
6823 | ||
6824 | static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt) | |
6825 | { | |
6826 | int ret; | |
6827 | ||
6828 | if (buf[0] != '0' && buf[0] != '1') | |
6829 | return -EINVAL; | |
6830 | ||
6831 | if (smt) | |
6832 | sched_smt_power_savings = (buf[0] == '1'); | |
6833 | else | |
6834 | sched_mc_power_savings = (buf[0] == '1'); | |
6835 | ||
6836 | ret = arch_reinit_sched_domains(); | |
6837 | ||
6838 | return ret ? ret : count; | |
6839 | } | |
6840 | ||
6841 | int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls) | |
6842 | { | |
6843 | int err = 0; | |
48f24c4d | 6844 | |
5c45bf27 SS |
6845 | #ifdef CONFIG_SCHED_SMT |
6846 | if (smt_capable()) | |
6847 | err = sysfs_create_file(&cls->kset.kobj, | |
6848 | &attr_sched_smt_power_savings.attr); | |
6849 | #endif | |
6850 | #ifdef CONFIG_SCHED_MC | |
6851 | if (!err && mc_capable()) | |
6852 | err = sysfs_create_file(&cls->kset.kobj, | |
6853 | &attr_sched_mc_power_savings.attr); | |
6854 | #endif | |
6855 | return err; | |
6856 | } | |
6857 | #endif | |
6858 | ||
6859 | #ifdef CONFIG_SCHED_MC | |
6860 | static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page) | |
6861 | { | |
6862 | return sprintf(page, "%u\n", sched_mc_power_savings); | |
6863 | } | |
48f24c4d IM |
6864 | static ssize_t sched_mc_power_savings_store(struct sys_device *dev, |
6865 | const char *buf, size_t count) | |
5c45bf27 SS |
6866 | { |
6867 | return sched_power_savings_store(buf, count, 0); | |
6868 | } | |
6869 | SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show, | |
6870 | sched_mc_power_savings_store); | |
6871 | #endif | |
6872 | ||
6873 | #ifdef CONFIG_SCHED_SMT | |
6874 | static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page) | |
6875 | { | |
6876 | return sprintf(page, "%u\n", sched_smt_power_savings); | |
6877 | } | |
48f24c4d IM |
6878 | static ssize_t sched_smt_power_savings_store(struct sys_device *dev, |
6879 | const char *buf, size_t count) | |
5c45bf27 SS |
6880 | { |
6881 | return sched_power_savings_store(buf, count, 1); | |
6882 | } | |
6883 | SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show, | |
6884 | sched_smt_power_savings_store); | |
6885 | #endif | |
6886 | ||
1da177e4 LT |
6887 | /* |
6888 | * Force a reinitialization of the sched domains hierarchy. The domains | |
6889 | * and groups cannot be updated in place without racing with the balancing | |
41c7ce9a | 6890 | * code, so we temporarily attach all running cpus to the NULL domain |
1da177e4 LT |
6891 | * which will prevent rebalancing while the sched domains are recalculated. |
6892 | */ | |
6893 | static int update_sched_domains(struct notifier_block *nfb, | |
6894 | unsigned long action, void *hcpu) | |
6895 | { | |
1da177e4 LT |
6896 | switch (action) { |
6897 | case CPU_UP_PREPARE: | |
6898 | case CPU_DOWN_PREPARE: | |
1a20ff27 | 6899 | detach_destroy_domains(&cpu_online_map); |
1da177e4 LT |
6900 | return NOTIFY_OK; |
6901 | ||
6902 | case CPU_UP_CANCELED: | |
6903 | case CPU_DOWN_FAILED: | |
6904 | case CPU_ONLINE: | |
6905 | case CPU_DEAD: | |
6906 | /* | |
6907 | * Fall through and re-initialise the domains. | |
6908 | */ | |
6909 | break; | |
6910 | default: | |
6911 | return NOTIFY_DONE; | |
6912 | } | |
6913 | ||
6914 | /* The hotplug lock is already held by cpu_up/cpu_down */ | |
1a20ff27 | 6915 | arch_init_sched_domains(&cpu_online_map); |
1da177e4 LT |
6916 | |
6917 | return NOTIFY_OK; | |
6918 | } | |
1da177e4 LT |
6919 | |
6920 | void __init sched_init_smp(void) | |
6921 | { | |
5c1e1767 NP |
6922 | cpumask_t non_isolated_cpus; |
6923 | ||
1da177e4 | 6924 | lock_cpu_hotplug(); |
1a20ff27 | 6925 | arch_init_sched_domains(&cpu_online_map); |
e5e5673f | 6926 | cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map); |
5c1e1767 NP |
6927 | if (cpus_empty(non_isolated_cpus)) |
6928 | cpu_set(smp_processor_id(), non_isolated_cpus); | |
1da177e4 LT |
6929 | unlock_cpu_hotplug(); |
6930 | /* XXX: Theoretical race here - CPU may be hotplugged now */ | |
6931 | hotcpu_notifier(update_sched_domains, 0); | |
5c1e1767 NP |
6932 | |
6933 | /* Move init over to a non-isolated CPU */ | |
6934 | if (set_cpus_allowed(current, non_isolated_cpus) < 0) | |
6935 | BUG(); | |
1da177e4 LT |
6936 | } |
6937 | #else | |
6938 | void __init sched_init_smp(void) | |
6939 | { | |
6940 | } | |
6941 | #endif /* CONFIG_SMP */ | |
6942 | ||
6943 | int in_sched_functions(unsigned long addr) | |
6944 | { | |
6945 | /* Linker adds these: start and end of __sched functions */ | |
6946 | extern char __sched_text_start[], __sched_text_end[]; | |
48f24c4d | 6947 | |
1da177e4 LT |
6948 | return in_lock_functions(addr) || |
6949 | (addr >= (unsigned long)__sched_text_start | |
6950 | && addr < (unsigned long)__sched_text_end); | |
6951 | } | |
6952 | ||
6953 | void __init sched_init(void) | |
6954 | { | |
1da177e4 | 6955 | int i, j, k; |
476f3534 | 6956 | int highest_cpu = 0; |
1da177e4 | 6957 | |
0a945022 | 6958 | for_each_possible_cpu(i) { |
70b97a7f IM |
6959 | struct prio_array *array; |
6960 | struct rq *rq; | |
1da177e4 LT |
6961 | |
6962 | rq = cpu_rq(i); | |
6963 | spin_lock_init(&rq->lock); | |
fcb99371 | 6964 | lockdep_set_class(&rq->lock, &rq->rq_lock_key); |
7897986b | 6965 | rq->nr_running = 0; |
1da177e4 LT |
6966 | rq->active = rq->arrays; |
6967 | rq->expired = rq->arrays + 1; | |
6968 | rq->best_expired_prio = MAX_PRIO; | |
6969 | ||
6970 | #ifdef CONFIG_SMP | |
41c7ce9a | 6971 | rq->sd = NULL; |
7897986b NP |
6972 | for (j = 1; j < 3; j++) |
6973 | rq->cpu_load[j] = 0; | |
1da177e4 LT |
6974 | rq->active_balance = 0; |
6975 | rq->push_cpu = 0; | |
0a2966b4 | 6976 | rq->cpu = i; |
1da177e4 LT |
6977 | rq->migration_thread = NULL; |
6978 | INIT_LIST_HEAD(&rq->migration_queue); | |
6979 | #endif | |
6980 | atomic_set(&rq->nr_iowait, 0); | |
6981 | ||
6982 | for (j = 0; j < 2; j++) { | |
6983 | array = rq->arrays + j; | |
6984 | for (k = 0; k < MAX_PRIO; k++) { | |
6985 | INIT_LIST_HEAD(array->queue + k); | |
6986 | __clear_bit(k, array->bitmap); | |
6987 | } | |
6988 | // delimiter for bitsearch | |
6989 | __set_bit(MAX_PRIO, array->bitmap); | |
6990 | } | |
476f3534 | 6991 | highest_cpu = i; |
1da177e4 LT |
6992 | } |
6993 | ||
2dd73a4f | 6994 | set_load_weight(&init_task); |
b50f60ce | 6995 | |
c9819f45 | 6996 | #ifdef CONFIG_SMP |
476f3534 | 6997 | nr_cpu_ids = highest_cpu + 1; |
c9819f45 CL |
6998 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL); |
6999 | #endif | |
7000 | ||
b50f60ce HC |
7001 | #ifdef CONFIG_RT_MUTEXES |
7002 | plist_head_init(&init_task.pi_waiters, &init_task.pi_lock); | |
7003 | #endif | |
7004 | ||
1da177e4 LT |
7005 | /* |
7006 | * The boot idle thread does lazy MMU switching as well: | |
7007 | */ | |
7008 | atomic_inc(&init_mm.mm_count); | |
7009 | enter_lazy_tlb(&init_mm, current); | |
7010 | ||
7011 | /* | |
7012 | * Make us the idle thread. Technically, schedule() should not be | |
7013 | * called from this thread, however somewhere below it might be, | |
7014 | * but because we are the idle thread, we just pick up running again | |
7015 | * when this runqueue becomes "idle". | |
7016 | */ | |
7017 | init_idle(current, smp_processor_id()); | |
7018 | } | |
7019 | ||
7020 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP | |
7021 | void __might_sleep(char *file, int line) | |
7022 | { | |
48f24c4d | 7023 | #ifdef in_atomic |
1da177e4 LT |
7024 | static unsigned long prev_jiffy; /* ratelimiting */ |
7025 | ||
7026 | if ((in_atomic() || irqs_disabled()) && | |
7027 | system_state == SYSTEM_RUNNING && !oops_in_progress) { | |
7028 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) | |
7029 | return; | |
7030 | prev_jiffy = jiffies; | |
91368d73 | 7031 | printk(KERN_ERR "BUG: sleeping function called from invalid" |
1da177e4 LT |
7032 | " context at %s:%d\n", file, line); |
7033 | printk("in_atomic():%d, irqs_disabled():%d\n", | |
7034 | in_atomic(), irqs_disabled()); | |
a4c410f0 | 7035 | debug_show_held_locks(current); |
3117df04 IM |
7036 | if (irqs_disabled()) |
7037 | print_irqtrace_events(current); | |
1da177e4 LT |
7038 | dump_stack(); |
7039 | } | |
7040 | #endif | |
7041 | } | |
7042 | EXPORT_SYMBOL(__might_sleep); | |
7043 | #endif | |
7044 | ||
7045 | #ifdef CONFIG_MAGIC_SYSRQ | |
7046 | void normalize_rt_tasks(void) | |
7047 | { | |
70b97a7f | 7048 | struct prio_array *array; |
1da177e4 | 7049 | struct task_struct *p; |
1da177e4 | 7050 | unsigned long flags; |
70b97a7f | 7051 | struct rq *rq; |
1da177e4 LT |
7052 | |
7053 | read_lock_irq(&tasklist_lock); | |
c96d145e | 7054 | for_each_process(p) { |
1da177e4 LT |
7055 | if (!rt_task(p)) |
7056 | continue; | |
7057 | ||
b29739f9 IM |
7058 | spin_lock_irqsave(&p->pi_lock, flags); |
7059 | rq = __task_rq_lock(p); | |
1da177e4 LT |
7060 | |
7061 | array = p->array; | |
7062 | if (array) | |
7063 | deactivate_task(p, task_rq(p)); | |
7064 | __setscheduler(p, SCHED_NORMAL, 0); | |
7065 | if (array) { | |
7066 | __activate_task(p, task_rq(p)); | |
7067 | resched_task(rq->curr); | |
7068 | } | |
7069 | ||
b29739f9 IM |
7070 | __task_rq_unlock(rq); |
7071 | spin_unlock_irqrestore(&p->pi_lock, flags); | |
1da177e4 LT |
7072 | } |
7073 | read_unlock_irq(&tasklist_lock); | |
7074 | } | |
7075 | ||
7076 | #endif /* CONFIG_MAGIC_SYSRQ */ | |
1df5c10a LT |
7077 | |
7078 | #ifdef CONFIG_IA64 | |
7079 | /* | |
7080 | * These functions are only useful for the IA64 MCA handling. | |
7081 | * | |
7082 | * They can only be called when the whole system has been | |
7083 | * stopped - every CPU needs to be quiescent, and no scheduling | |
7084 | * activity can take place. Using them for anything else would | |
7085 | * be a serious bug, and as a result, they aren't even visible | |
7086 | * under any other configuration. | |
7087 | */ | |
7088 | ||
7089 | /** | |
7090 | * curr_task - return the current task for a given cpu. | |
7091 | * @cpu: the processor in question. | |
7092 | * | |
7093 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
7094 | */ | |
36c8b586 | 7095 | struct task_struct *curr_task(int cpu) |
1df5c10a LT |
7096 | { |
7097 | return cpu_curr(cpu); | |
7098 | } | |
7099 | ||
7100 | /** | |
7101 | * set_curr_task - set the current task for a given cpu. | |
7102 | * @cpu: the processor in question. | |
7103 | * @p: the task pointer to set. | |
7104 | * | |
7105 | * Description: This function must only be used when non-maskable interrupts | |
7106 | * are serviced on a separate stack. It allows the architecture to switch the | |
7107 | * notion of the current task on a cpu in a non-blocking manner. This function | |
7108 | * must be called with all CPU's synchronized, and interrupts disabled, the | |
7109 | * and caller must save the original value of the current task (see | |
7110 | * curr_task() above) and restore that value before reenabling interrupts and | |
7111 | * re-starting the system. | |
7112 | * | |
7113 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
7114 | */ | |
36c8b586 | 7115 | void set_curr_task(int cpu, struct task_struct *p) |
1df5c10a LT |
7116 | { |
7117 | cpu_curr(cpu) = p; | |
7118 | } | |
7119 | ||
7120 | #endif |