]>
Commit | Line | Data |
---|---|---|
1da177e4 | 1 | /* |
391e43da | 2 | * kernel/sched/core.c |
1da177e4 LT |
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 | |
c31f2e8a IM |
19 | * 2007-04-15 Work begun on replacing all interactivity tuning with a |
20 | * fair scheduling design by Con Kolivas. | |
21 | * 2007-05-05 Load balancing (smp-nice) and other improvements | |
22 | * by Peter Williams | |
23 | * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith | |
24 | * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri | |
b9131769 IM |
25 | * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins, |
26 | * Thomas Gleixner, Mike Kravetz | |
1da177e4 LT |
27 | */ |
28 | ||
e1b77c92 | 29 | #include <linux/kasan.h> |
1da177e4 LT |
30 | #include <linux/mm.h> |
31 | #include <linux/module.h> | |
32 | #include <linux/nmi.h> | |
33 | #include <linux/init.h> | |
dff06c15 | 34 | #include <linux/uaccess.h> |
1da177e4 | 35 | #include <linux/highmem.h> |
f98db601 | 36 | #include <linux/mmu_context.h> |
1da177e4 | 37 | #include <linux/interrupt.h> |
c59ede7b | 38 | #include <linux/capability.h> |
1da177e4 LT |
39 | #include <linux/completion.h> |
40 | #include <linux/kernel_stat.h> | |
9a11b49a | 41 | #include <linux/debug_locks.h> |
cdd6c482 | 42 | #include <linux/perf_event.h> |
1da177e4 LT |
43 | #include <linux/security.h> |
44 | #include <linux/notifier.h> | |
45 | #include <linux/profile.h> | |
7dfb7103 | 46 | #include <linux/freezer.h> |
198e2f18 | 47 | #include <linux/vmalloc.h> |
1da177e4 LT |
48 | #include <linux/blkdev.h> |
49 | #include <linux/delay.h> | |
b488893a | 50 | #include <linux/pid_namespace.h> |
1da177e4 LT |
51 | #include <linux/smp.h> |
52 | #include <linux/threads.h> | |
53 | #include <linux/timer.h> | |
54 | #include <linux/rcupdate.h> | |
55 | #include <linux/cpu.h> | |
56 | #include <linux/cpuset.h> | |
57 | #include <linux/percpu.h> | |
b5aadf7f | 58 | #include <linux/proc_fs.h> |
1da177e4 | 59 | #include <linux/seq_file.h> |
e692ab53 | 60 | #include <linux/sysctl.h> |
1da177e4 LT |
61 | #include <linux/syscalls.h> |
62 | #include <linux/times.h> | |
8f0ab514 | 63 | #include <linux/tsacct_kern.h> |
c6fd91f0 | 64 | #include <linux/kprobes.h> |
0ff92245 | 65 | #include <linux/delayacct.h> |
dff06c15 | 66 | #include <linux/unistd.h> |
f5ff8422 | 67 | #include <linux/pagemap.h> |
8f4d37ec | 68 | #include <linux/hrtimer.h> |
30914a58 | 69 | #include <linux/tick.h> |
f00b45c1 | 70 | #include <linux/ctype.h> |
6cd8a4bb | 71 | #include <linux/ftrace.h> |
5a0e3ad6 | 72 | #include <linux/slab.h> |
f1c6f1a7 | 73 | #include <linux/init_task.h> |
91d1aa43 | 74 | #include <linux/context_tracking.h> |
52f5684c | 75 | #include <linux/compiler.h> |
8e05e96a | 76 | #include <linux/frame.h> |
1da177e4 | 77 | |
96f951ed | 78 | #include <asm/switch_to.h> |
5517d86b | 79 | #include <asm/tlb.h> |
838225b4 | 80 | #include <asm/irq_regs.h> |
db7e527d | 81 | #include <asm/mutex.h> |
e6e6685a GC |
82 | #ifdef CONFIG_PARAVIRT |
83 | #include <asm/paravirt.h> | |
84 | #endif | |
1da177e4 | 85 | |
029632fb | 86 | #include "sched.h" |
ea138446 | 87 | #include "../workqueue_internal.h" |
29d5e047 | 88 | #include "../smpboot.h" |
6e0534f2 | 89 | |
a8d154b0 | 90 | #define CREATE_TRACE_POINTS |
ad8d75ff | 91 | #include <trace/events/sched.h> |
a8d154b0 | 92 | |
029632fb PZ |
93 | DEFINE_MUTEX(sched_domains_mutex); |
94 | DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); | |
dc61b1d6 | 95 | |
fe44d621 | 96 | static void update_rq_clock_task(struct rq *rq, s64 delta); |
305e6835 | 97 | |
029632fb | 98 | void update_rq_clock(struct rq *rq) |
3e51f33f | 99 | { |
fe44d621 | 100 | s64 delta; |
305e6835 | 101 | |
9edfbfed PZ |
102 | lockdep_assert_held(&rq->lock); |
103 | ||
104 | if (rq->clock_skip_update & RQCF_ACT_SKIP) | |
f26f9aff | 105 | return; |
aa483808 | 106 | |
fe44d621 | 107 | delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; |
4036ac15 MG |
108 | if (delta < 0) |
109 | return; | |
fe44d621 PZ |
110 | rq->clock += delta; |
111 | update_rq_clock_task(rq, delta); | |
3e51f33f PZ |
112 | } |
113 | ||
bf5c91ba IM |
114 | /* |
115 | * Debugging: various feature bits | |
116 | */ | |
f00b45c1 | 117 | |
f00b45c1 PZ |
118 | #define SCHED_FEAT(name, enabled) \ |
119 | (1UL << __SCHED_FEAT_##name) * enabled | | |
120 | ||
bf5c91ba | 121 | const_debug unsigned int sysctl_sched_features = |
391e43da | 122 | #include "features.h" |
f00b45c1 PZ |
123 | 0; |
124 | ||
125 | #undef SCHED_FEAT | |
126 | ||
b82d9fdd PZ |
127 | /* |
128 | * Number of tasks to iterate in a single balance run. | |
129 | * Limited because this is done with IRQs disabled. | |
130 | */ | |
131 | const_debug unsigned int sysctl_sched_nr_migrate = 32; | |
132 | ||
e9e9250b PZ |
133 | /* |
134 | * period over which we average the RT time consumption, measured | |
135 | * in ms. | |
136 | * | |
137 | * default: 1s | |
138 | */ | |
139 | const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; | |
140 | ||
fa85ae24 | 141 | /* |
9f0c1e56 | 142 | * period over which we measure -rt task cpu usage in us. |
fa85ae24 PZ |
143 | * default: 1s |
144 | */ | |
9f0c1e56 | 145 | unsigned int sysctl_sched_rt_period = 1000000; |
fa85ae24 | 146 | |
029632fb | 147 | __read_mostly int scheduler_running; |
6892b75e | 148 | |
9f0c1e56 PZ |
149 | /* |
150 | * part of the period that we allow rt tasks to run in us. | |
151 | * default: 0.95s | |
152 | */ | |
153 | int sysctl_sched_rt_runtime = 950000; | |
fa85ae24 | 154 | |
3fa0818b RR |
155 | /* cpus with isolated domains */ |
156 | cpumask_var_t cpu_isolated_map; | |
157 | ||
1da177e4 | 158 | /* |
cc2a73b5 | 159 | * this_rq_lock - lock this runqueue and disable interrupts. |
1da177e4 | 160 | */ |
a9957449 | 161 | static struct rq *this_rq_lock(void) |
1da177e4 LT |
162 | __acquires(rq->lock) |
163 | { | |
70b97a7f | 164 | struct rq *rq; |
1da177e4 LT |
165 | |
166 | local_irq_disable(); | |
167 | rq = this_rq(); | |
05fa785c | 168 | raw_spin_lock(&rq->lock); |
1da177e4 LT |
169 | |
170 | return rq; | |
171 | } | |
172 | ||
3e71a462 PZ |
173 | /* |
174 | * __task_rq_lock - lock the rq @p resides on. | |
175 | */ | |
eb580751 | 176 | struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf) |
3e71a462 PZ |
177 | __acquires(rq->lock) |
178 | { | |
179 | struct rq *rq; | |
180 | ||
181 | lockdep_assert_held(&p->pi_lock); | |
182 | ||
183 | for (;;) { | |
184 | rq = task_rq(p); | |
185 | raw_spin_lock(&rq->lock); | |
186 | if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) { | |
e7904a28 | 187 | rf->cookie = lockdep_pin_lock(&rq->lock); |
3e71a462 PZ |
188 | return rq; |
189 | } | |
190 | raw_spin_unlock(&rq->lock); | |
191 | ||
192 | while (unlikely(task_on_rq_migrating(p))) | |
193 | cpu_relax(); | |
194 | } | |
195 | } | |
196 | ||
197 | /* | |
198 | * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. | |
199 | */ | |
eb580751 | 200 | struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf) |
3e71a462 PZ |
201 | __acquires(p->pi_lock) |
202 | __acquires(rq->lock) | |
203 | { | |
204 | struct rq *rq; | |
205 | ||
206 | for (;;) { | |
eb580751 | 207 | raw_spin_lock_irqsave(&p->pi_lock, rf->flags); |
3e71a462 PZ |
208 | rq = task_rq(p); |
209 | raw_spin_lock(&rq->lock); | |
210 | /* | |
211 | * move_queued_task() task_rq_lock() | |
212 | * | |
213 | * ACQUIRE (rq->lock) | |
214 | * [S] ->on_rq = MIGRATING [L] rq = task_rq() | |
215 | * WMB (__set_task_cpu()) ACQUIRE (rq->lock); | |
216 | * [S] ->cpu = new_cpu [L] task_rq() | |
217 | * [L] ->on_rq | |
218 | * RELEASE (rq->lock) | |
219 | * | |
220 | * If we observe the old cpu in task_rq_lock, the acquire of | |
221 | * the old rq->lock will fully serialize against the stores. | |
222 | * | |
223 | * If we observe the new cpu in task_rq_lock, the acquire will | |
224 | * pair with the WMB to ensure we must then also see migrating. | |
225 | */ | |
226 | if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) { | |
e7904a28 | 227 | rf->cookie = lockdep_pin_lock(&rq->lock); |
3e71a462 PZ |
228 | return rq; |
229 | } | |
230 | raw_spin_unlock(&rq->lock); | |
eb580751 | 231 | raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags); |
3e71a462 PZ |
232 | |
233 | while (unlikely(task_on_rq_migrating(p))) | |
234 | cpu_relax(); | |
235 | } | |
236 | } | |
237 | ||
8f4d37ec PZ |
238 | #ifdef CONFIG_SCHED_HRTICK |
239 | /* | |
240 | * Use HR-timers to deliver accurate preemption points. | |
8f4d37ec | 241 | */ |
8f4d37ec | 242 | |
8f4d37ec PZ |
243 | static void hrtick_clear(struct rq *rq) |
244 | { | |
245 | if (hrtimer_active(&rq->hrtick_timer)) | |
246 | hrtimer_cancel(&rq->hrtick_timer); | |
247 | } | |
248 | ||
8f4d37ec PZ |
249 | /* |
250 | * High-resolution timer tick. | |
251 | * Runs from hardirq context with interrupts disabled. | |
252 | */ | |
253 | static enum hrtimer_restart hrtick(struct hrtimer *timer) | |
254 | { | |
255 | struct rq *rq = container_of(timer, struct rq, hrtick_timer); | |
256 | ||
257 | WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); | |
258 | ||
05fa785c | 259 | raw_spin_lock(&rq->lock); |
3e51f33f | 260 | update_rq_clock(rq); |
8f4d37ec | 261 | rq->curr->sched_class->task_tick(rq, rq->curr, 1); |
05fa785c | 262 | raw_spin_unlock(&rq->lock); |
8f4d37ec PZ |
263 | |
264 | return HRTIMER_NORESTART; | |
265 | } | |
266 | ||
95e904c7 | 267 | #ifdef CONFIG_SMP |
971ee28c | 268 | |
4961b6e1 | 269 | static void __hrtick_restart(struct rq *rq) |
971ee28c PZ |
270 | { |
271 | struct hrtimer *timer = &rq->hrtick_timer; | |
971ee28c | 272 | |
4961b6e1 | 273 | hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED); |
971ee28c PZ |
274 | } |
275 | ||
31656519 PZ |
276 | /* |
277 | * called from hardirq (IPI) context | |
278 | */ | |
279 | static void __hrtick_start(void *arg) | |
b328ca18 | 280 | { |
31656519 | 281 | struct rq *rq = arg; |
b328ca18 | 282 | |
05fa785c | 283 | raw_spin_lock(&rq->lock); |
971ee28c | 284 | __hrtick_restart(rq); |
31656519 | 285 | rq->hrtick_csd_pending = 0; |
05fa785c | 286 | raw_spin_unlock(&rq->lock); |
b328ca18 PZ |
287 | } |
288 | ||
31656519 PZ |
289 | /* |
290 | * Called to set the hrtick timer state. | |
291 | * | |
292 | * called with rq->lock held and irqs disabled | |
293 | */ | |
029632fb | 294 | void hrtick_start(struct rq *rq, u64 delay) |
b328ca18 | 295 | { |
31656519 | 296 | struct hrtimer *timer = &rq->hrtick_timer; |
177ef2a6 | 297 | ktime_t time; |
298 | s64 delta; | |
299 | ||
300 | /* | |
301 | * Don't schedule slices shorter than 10000ns, that just | |
302 | * doesn't make sense and can cause timer DoS. | |
303 | */ | |
304 | delta = max_t(s64, delay, 10000LL); | |
305 | time = ktime_add_ns(timer->base->get_time(), delta); | |
b328ca18 | 306 | |
cc584b21 | 307 | hrtimer_set_expires(timer, time); |
31656519 PZ |
308 | |
309 | if (rq == this_rq()) { | |
971ee28c | 310 | __hrtick_restart(rq); |
31656519 | 311 | } else if (!rq->hrtick_csd_pending) { |
c46fff2a | 312 | smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd); |
31656519 PZ |
313 | rq->hrtick_csd_pending = 1; |
314 | } | |
b328ca18 PZ |
315 | } |
316 | ||
31656519 PZ |
317 | #else |
318 | /* | |
319 | * Called to set the hrtick timer state. | |
320 | * | |
321 | * called with rq->lock held and irqs disabled | |
322 | */ | |
029632fb | 323 | void hrtick_start(struct rq *rq, u64 delay) |
31656519 | 324 | { |
86893335 WL |
325 | /* |
326 | * Don't schedule slices shorter than 10000ns, that just | |
327 | * doesn't make sense. Rely on vruntime for fairness. | |
328 | */ | |
329 | delay = max_t(u64, delay, 10000LL); | |
4961b6e1 TG |
330 | hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), |
331 | HRTIMER_MODE_REL_PINNED); | |
31656519 | 332 | } |
31656519 | 333 | #endif /* CONFIG_SMP */ |
8f4d37ec | 334 | |
31656519 | 335 | static void init_rq_hrtick(struct rq *rq) |
8f4d37ec | 336 | { |
31656519 PZ |
337 | #ifdef CONFIG_SMP |
338 | rq->hrtick_csd_pending = 0; | |
8f4d37ec | 339 | |
31656519 PZ |
340 | rq->hrtick_csd.flags = 0; |
341 | rq->hrtick_csd.func = __hrtick_start; | |
342 | rq->hrtick_csd.info = rq; | |
343 | #endif | |
8f4d37ec | 344 | |
31656519 PZ |
345 | hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
346 | rq->hrtick_timer.function = hrtick; | |
8f4d37ec | 347 | } |
006c75f1 | 348 | #else /* CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
349 | static inline void hrtick_clear(struct rq *rq) |
350 | { | |
351 | } | |
352 | ||
8f4d37ec PZ |
353 | static inline void init_rq_hrtick(struct rq *rq) |
354 | { | |
355 | } | |
006c75f1 | 356 | #endif /* CONFIG_SCHED_HRTICK */ |
8f4d37ec | 357 | |
5529578a FW |
358 | /* |
359 | * cmpxchg based fetch_or, macro so it works for different integer types | |
360 | */ | |
361 | #define fetch_or(ptr, mask) \ | |
362 | ({ \ | |
363 | typeof(ptr) _ptr = (ptr); \ | |
364 | typeof(mask) _mask = (mask); \ | |
365 | typeof(*_ptr) _old, _val = *_ptr; \ | |
366 | \ | |
367 | for (;;) { \ | |
368 | _old = cmpxchg(_ptr, _val, _val | _mask); \ | |
369 | if (_old == _val) \ | |
370 | break; \ | |
371 | _val = _old; \ | |
372 | } \ | |
373 | _old; \ | |
374 | }) | |
375 | ||
e3baac47 | 376 | #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG) |
fd99f91a PZ |
377 | /* |
378 | * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG, | |
379 | * this avoids any races wrt polling state changes and thereby avoids | |
380 | * spurious IPIs. | |
381 | */ | |
382 | static bool set_nr_and_not_polling(struct task_struct *p) | |
383 | { | |
384 | struct thread_info *ti = task_thread_info(p); | |
385 | return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG); | |
386 | } | |
e3baac47 PZ |
387 | |
388 | /* | |
389 | * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set. | |
390 | * | |
391 | * If this returns true, then the idle task promises to call | |
392 | * sched_ttwu_pending() and reschedule soon. | |
393 | */ | |
394 | static bool set_nr_if_polling(struct task_struct *p) | |
395 | { | |
396 | struct thread_info *ti = task_thread_info(p); | |
316c1608 | 397 | typeof(ti->flags) old, val = READ_ONCE(ti->flags); |
e3baac47 PZ |
398 | |
399 | for (;;) { | |
400 | if (!(val & _TIF_POLLING_NRFLAG)) | |
401 | return false; | |
402 | if (val & _TIF_NEED_RESCHED) | |
403 | return true; | |
404 | old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED); | |
405 | if (old == val) | |
406 | break; | |
407 | val = old; | |
408 | } | |
409 | return true; | |
410 | } | |
411 | ||
fd99f91a PZ |
412 | #else |
413 | static bool set_nr_and_not_polling(struct task_struct *p) | |
414 | { | |
415 | set_tsk_need_resched(p); | |
416 | return true; | |
417 | } | |
e3baac47 PZ |
418 | |
419 | #ifdef CONFIG_SMP | |
420 | static bool set_nr_if_polling(struct task_struct *p) | |
421 | { | |
422 | return false; | |
423 | } | |
424 | #endif | |
fd99f91a PZ |
425 | #endif |
426 | ||
76751049 PZ |
427 | void wake_q_add(struct wake_q_head *head, struct task_struct *task) |
428 | { | |
429 | struct wake_q_node *node = &task->wake_q; | |
430 | ||
431 | /* | |
432 | * Atomically grab the task, if ->wake_q is !nil already it means | |
433 | * its already queued (either by us or someone else) and will get the | |
434 | * wakeup due to that. | |
435 | * | |
436 | * This cmpxchg() implies a full barrier, which pairs with the write | |
58fe9c46 | 437 | * barrier implied by the wakeup in wake_up_q(). |
76751049 PZ |
438 | */ |
439 | if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL)) | |
440 | return; | |
441 | ||
442 | get_task_struct(task); | |
443 | ||
444 | /* | |
445 | * The head is context local, there can be no concurrency. | |
446 | */ | |
447 | *head->lastp = node; | |
448 | head->lastp = &node->next; | |
449 | } | |
450 | ||
451 | void wake_up_q(struct wake_q_head *head) | |
452 | { | |
453 | struct wake_q_node *node = head->first; | |
454 | ||
455 | while (node != WAKE_Q_TAIL) { | |
456 | struct task_struct *task; | |
457 | ||
458 | task = container_of(node, struct task_struct, wake_q); | |
459 | BUG_ON(!task); | |
460 | /* task can safely be re-inserted now */ | |
461 | node = node->next; | |
462 | task->wake_q.next = NULL; | |
463 | ||
464 | /* | |
465 | * wake_up_process() implies a wmb() to pair with the queueing | |
466 | * in wake_q_add() so as not to miss wakeups. | |
467 | */ | |
468 | wake_up_process(task); | |
469 | put_task_struct(task); | |
470 | } | |
471 | } | |
472 | ||
c24d20db | 473 | /* |
8875125e | 474 | * resched_curr - mark rq's current task 'to be rescheduled now'. |
c24d20db IM |
475 | * |
476 | * On UP this means the setting of the need_resched flag, on SMP it | |
477 | * might also involve a cross-CPU call to trigger the scheduler on | |
478 | * the target CPU. | |
479 | */ | |
8875125e | 480 | void resched_curr(struct rq *rq) |
c24d20db | 481 | { |
8875125e | 482 | struct task_struct *curr = rq->curr; |
c24d20db IM |
483 | int cpu; |
484 | ||
8875125e | 485 | lockdep_assert_held(&rq->lock); |
c24d20db | 486 | |
8875125e | 487 | if (test_tsk_need_resched(curr)) |
c24d20db IM |
488 | return; |
489 | ||
8875125e | 490 | cpu = cpu_of(rq); |
fd99f91a | 491 | |
f27dde8d | 492 | if (cpu == smp_processor_id()) { |
8875125e | 493 | set_tsk_need_resched(curr); |
f27dde8d | 494 | set_preempt_need_resched(); |
c24d20db | 495 | return; |
f27dde8d | 496 | } |
c24d20db | 497 | |
8875125e | 498 | if (set_nr_and_not_polling(curr)) |
c24d20db | 499 | smp_send_reschedule(cpu); |
dfc68f29 AL |
500 | else |
501 | trace_sched_wake_idle_without_ipi(cpu); | |
c24d20db IM |
502 | } |
503 | ||
029632fb | 504 | void resched_cpu(int cpu) |
c24d20db IM |
505 | { |
506 | struct rq *rq = cpu_rq(cpu); | |
507 | unsigned long flags; | |
508 | ||
05fa785c | 509 | if (!raw_spin_trylock_irqsave(&rq->lock, flags)) |
c24d20db | 510 | return; |
8875125e | 511 | resched_curr(rq); |
05fa785c | 512 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
c24d20db | 513 | } |
06d8308c | 514 | |
b021fe3e | 515 | #ifdef CONFIG_SMP |
3451d024 | 516 | #ifdef CONFIG_NO_HZ_COMMON |
83cd4fe2 VP |
517 | /* |
518 | * In the semi idle case, use the nearest busy cpu for migrating timers | |
519 | * from an idle cpu. This is good for power-savings. | |
520 | * | |
521 | * We don't do similar optimization for completely idle system, as | |
522 | * selecting an idle cpu will add more delays to the timers than intended | |
523 | * (as that cpu's timer base may not be uptodate wrt jiffies etc). | |
524 | */ | |
bc7a34b8 | 525 | int get_nohz_timer_target(void) |
83cd4fe2 | 526 | { |
bc7a34b8 | 527 | int i, cpu = smp_processor_id(); |
83cd4fe2 VP |
528 | struct sched_domain *sd; |
529 | ||
9642d18e | 530 | if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu)) |
6201b4d6 VK |
531 | return cpu; |
532 | ||
057f3fad | 533 | rcu_read_lock(); |
83cd4fe2 | 534 | for_each_domain(cpu, sd) { |
057f3fad | 535 | for_each_cpu(i, sched_domain_span(sd)) { |
44496922 WL |
536 | if (cpu == i) |
537 | continue; | |
538 | ||
539 | if (!idle_cpu(i) && is_housekeeping_cpu(i)) { | |
057f3fad PZ |
540 | cpu = i; |
541 | goto unlock; | |
542 | } | |
543 | } | |
83cd4fe2 | 544 | } |
9642d18e VH |
545 | |
546 | if (!is_housekeeping_cpu(cpu)) | |
547 | cpu = housekeeping_any_cpu(); | |
057f3fad PZ |
548 | unlock: |
549 | rcu_read_unlock(); | |
83cd4fe2 VP |
550 | return cpu; |
551 | } | |
06d8308c TG |
552 | /* |
553 | * When add_timer_on() enqueues a timer into the timer wheel of an | |
554 | * idle CPU then this timer might expire before the next timer event | |
555 | * which is scheduled to wake up that CPU. In case of a completely | |
556 | * idle system the next event might even be infinite time into the | |
557 | * future. wake_up_idle_cpu() ensures that the CPU is woken up and | |
558 | * leaves the inner idle loop so the newly added timer is taken into | |
559 | * account when the CPU goes back to idle and evaluates the timer | |
560 | * wheel for the next timer event. | |
561 | */ | |
1c20091e | 562 | static void wake_up_idle_cpu(int cpu) |
06d8308c TG |
563 | { |
564 | struct rq *rq = cpu_rq(cpu); | |
565 | ||
566 | if (cpu == smp_processor_id()) | |
567 | return; | |
568 | ||
67b9ca70 | 569 | if (set_nr_and_not_polling(rq->idle)) |
06d8308c | 570 | smp_send_reschedule(cpu); |
dfc68f29 AL |
571 | else |
572 | trace_sched_wake_idle_without_ipi(cpu); | |
45bf76df IM |
573 | } |
574 | ||
c5bfece2 | 575 | static bool wake_up_full_nohz_cpu(int cpu) |
1c20091e | 576 | { |
53c5fa16 FW |
577 | /* |
578 | * We just need the target to call irq_exit() and re-evaluate | |
579 | * the next tick. The nohz full kick at least implies that. | |
580 | * If needed we can still optimize that later with an | |
581 | * empty IRQ. | |
582 | */ | |
c5bfece2 | 583 | if (tick_nohz_full_cpu(cpu)) { |
1c20091e FW |
584 | if (cpu != smp_processor_id() || |
585 | tick_nohz_tick_stopped()) | |
53c5fa16 | 586 | tick_nohz_full_kick_cpu(cpu); |
1c20091e FW |
587 | return true; |
588 | } | |
589 | ||
590 | return false; | |
591 | } | |
592 | ||
593 | void wake_up_nohz_cpu(int cpu) | |
594 | { | |
c5bfece2 | 595 | if (!wake_up_full_nohz_cpu(cpu)) |
1c20091e FW |
596 | wake_up_idle_cpu(cpu); |
597 | } | |
598 | ||
ca38062e | 599 | static inline bool got_nohz_idle_kick(void) |
45bf76df | 600 | { |
1c792db7 | 601 | int cpu = smp_processor_id(); |
873b4c65 VG |
602 | |
603 | if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu))) | |
604 | return false; | |
605 | ||
606 | if (idle_cpu(cpu) && !need_resched()) | |
607 | return true; | |
608 | ||
609 | /* | |
610 | * We can't run Idle Load Balance on this CPU for this time so we | |
611 | * cancel it and clear NOHZ_BALANCE_KICK | |
612 | */ | |
613 | clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); | |
614 | return false; | |
45bf76df IM |
615 | } |
616 | ||
3451d024 | 617 | #else /* CONFIG_NO_HZ_COMMON */ |
45bf76df | 618 | |
ca38062e | 619 | static inline bool got_nohz_idle_kick(void) |
2069dd75 | 620 | { |
ca38062e | 621 | return false; |
2069dd75 PZ |
622 | } |
623 | ||
3451d024 | 624 | #endif /* CONFIG_NO_HZ_COMMON */ |
d842de87 | 625 | |
ce831b38 | 626 | #ifdef CONFIG_NO_HZ_FULL |
76d92ac3 | 627 | bool sched_can_stop_tick(struct rq *rq) |
ce831b38 | 628 | { |
76d92ac3 FW |
629 | int fifo_nr_running; |
630 | ||
631 | /* Deadline tasks, even if single, need the tick */ | |
632 | if (rq->dl.dl_nr_running) | |
633 | return false; | |
634 | ||
1e78cdbd | 635 | /* |
2548d546 PZ |
636 | * If there are more than one RR tasks, we need the tick to effect the |
637 | * actual RR behaviour. | |
1e78cdbd | 638 | */ |
76d92ac3 FW |
639 | if (rq->rt.rr_nr_running) { |
640 | if (rq->rt.rr_nr_running == 1) | |
641 | return true; | |
642 | else | |
643 | return false; | |
1e78cdbd RR |
644 | } |
645 | ||
2548d546 PZ |
646 | /* |
647 | * If there's no RR tasks, but FIFO tasks, we can skip the tick, no | |
648 | * forced preemption between FIFO tasks. | |
649 | */ | |
650 | fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running; | |
651 | if (fifo_nr_running) | |
652 | return true; | |
653 | ||
654 | /* | |
655 | * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left; | |
656 | * if there's more than one we need the tick for involuntary | |
657 | * preemption. | |
658 | */ | |
659 | if (rq->nr_running > 1) | |
541b8264 | 660 | return false; |
ce831b38 | 661 | |
541b8264 | 662 | return true; |
ce831b38 FW |
663 | } |
664 | #endif /* CONFIG_NO_HZ_FULL */ | |
d842de87 | 665 | |
029632fb | 666 | void sched_avg_update(struct rq *rq) |
18d95a28 | 667 | { |
e9e9250b PZ |
668 | s64 period = sched_avg_period(); |
669 | ||
78becc27 | 670 | while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { |
0d98bb26 WD |
671 | /* |
672 | * Inline assembly required to prevent the compiler | |
673 | * optimising this loop into a divmod call. | |
674 | * See __iter_div_u64_rem() for another example of this. | |
675 | */ | |
676 | asm("" : "+rm" (rq->age_stamp)); | |
e9e9250b PZ |
677 | rq->age_stamp += period; |
678 | rq->rt_avg /= 2; | |
679 | } | |
18d95a28 PZ |
680 | } |
681 | ||
6d6bc0ad | 682 | #endif /* CONFIG_SMP */ |
18d95a28 | 683 | |
a790de99 PT |
684 | #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ |
685 | (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) | |
c09595f6 | 686 | /* |
8277434e PT |
687 | * Iterate task_group tree rooted at *from, calling @down when first entering a |
688 | * node and @up when leaving it for the final time. | |
689 | * | |
690 | * Caller must hold rcu_lock or sufficient equivalent. | |
c09595f6 | 691 | */ |
029632fb | 692 | int walk_tg_tree_from(struct task_group *from, |
8277434e | 693 | tg_visitor down, tg_visitor up, void *data) |
c09595f6 PZ |
694 | { |
695 | struct task_group *parent, *child; | |
eb755805 | 696 | int ret; |
c09595f6 | 697 | |
8277434e PT |
698 | parent = from; |
699 | ||
c09595f6 | 700 | down: |
eb755805 PZ |
701 | ret = (*down)(parent, data); |
702 | if (ret) | |
8277434e | 703 | goto out; |
c09595f6 PZ |
704 | list_for_each_entry_rcu(child, &parent->children, siblings) { |
705 | parent = child; | |
706 | goto down; | |
707 | ||
708 | up: | |
709 | continue; | |
710 | } | |
eb755805 | 711 | ret = (*up)(parent, data); |
8277434e PT |
712 | if (ret || parent == from) |
713 | goto out; | |
c09595f6 PZ |
714 | |
715 | child = parent; | |
716 | parent = parent->parent; | |
717 | if (parent) | |
718 | goto up; | |
8277434e | 719 | out: |
eb755805 | 720 | return ret; |
c09595f6 PZ |
721 | } |
722 | ||
029632fb | 723 | int tg_nop(struct task_group *tg, void *data) |
eb755805 | 724 | { |
e2b245f8 | 725 | return 0; |
eb755805 | 726 | } |
18d95a28 PZ |
727 | #endif |
728 | ||
45bf76df IM |
729 | static void set_load_weight(struct task_struct *p) |
730 | { | |
f05998d4 NR |
731 | int prio = p->static_prio - MAX_RT_PRIO; |
732 | struct load_weight *load = &p->se.load; | |
733 | ||
dd41f596 IM |
734 | /* |
735 | * SCHED_IDLE tasks get minimal weight: | |
736 | */ | |
20f9cd2a | 737 | if (idle_policy(p->policy)) { |
c8b28116 | 738 | load->weight = scale_load(WEIGHT_IDLEPRIO); |
f05998d4 | 739 | load->inv_weight = WMULT_IDLEPRIO; |
dd41f596 IM |
740 | return; |
741 | } | |
71f8bd46 | 742 | |
ed82b8a1 AK |
743 | load->weight = scale_load(sched_prio_to_weight[prio]); |
744 | load->inv_weight = sched_prio_to_wmult[prio]; | |
71f8bd46 IM |
745 | } |
746 | ||
1de64443 | 747 | static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags) |
2087a1ad | 748 | { |
a64692a3 | 749 | update_rq_clock(rq); |
1de64443 PZ |
750 | if (!(flags & ENQUEUE_RESTORE)) |
751 | sched_info_queued(rq, p); | |
371fd7e7 | 752 | p->sched_class->enqueue_task(rq, p, flags); |
71f8bd46 IM |
753 | } |
754 | ||
1de64443 | 755 | static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags) |
71f8bd46 | 756 | { |
a64692a3 | 757 | update_rq_clock(rq); |
1de64443 PZ |
758 | if (!(flags & DEQUEUE_SAVE)) |
759 | sched_info_dequeued(rq, p); | |
371fd7e7 | 760 | p->sched_class->dequeue_task(rq, p, flags); |
71f8bd46 IM |
761 | } |
762 | ||
029632fb | 763 | void activate_task(struct rq *rq, struct task_struct *p, int flags) |
1e3c88bd PZ |
764 | { |
765 | if (task_contributes_to_load(p)) | |
766 | rq->nr_uninterruptible--; | |
767 | ||
371fd7e7 | 768 | enqueue_task(rq, p, flags); |
1e3c88bd PZ |
769 | } |
770 | ||
029632fb | 771 | void deactivate_task(struct rq *rq, struct task_struct *p, int flags) |
1e3c88bd PZ |
772 | { |
773 | if (task_contributes_to_load(p)) | |
774 | rq->nr_uninterruptible++; | |
775 | ||
371fd7e7 | 776 | dequeue_task(rq, p, flags); |
1e3c88bd PZ |
777 | } |
778 | ||
fe44d621 | 779 | static void update_rq_clock_task(struct rq *rq, s64 delta) |
aa483808 | 780 | { |
095c0aa8 GC |
781 | /* |
782 | * In theory, the compile should just see 0 here, and optimize out the call | |
783 | * to sched_rt_avg_update. But I don't trust it... | |
784 | */ | |
785 | #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) | |
786 | s64 steal = 0, irq_delta = 0; | |
787 | #endif | |
788 | #ifdef CONFIG_IRQ_TIME_ACCOUNTING | |
8e92c201 | 789 | irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; |
fe44d621 PZ |
790 | |
791 | /* | |
792 | * Since irq_time is only updated on {soft,}irq_exit, we might run into | |
793 | * this case when a previous update_rq_clock() happened inside a | |
794 | * {soft,}irq region. | |
795 | * | |
796 | * When this happens, we stop ->clock_task and only update the | |
797 | * prev_irq_time stamp to account for the part that fit, so that a next | |
798 | * update will consume the rest. This ensures ->clock_task is | |
799 | * monotonic. | |
800 | * | |
801 | * It does however cause some slight miss-attribution of {soft,}irq | |
802 | * time, a more accurate solution would be to update the irq_time using | |
803 | * the current rq->clock timestamp, except that would require using | |
804 | * atomic ops. | |
805 | */ | |
806 | if (irq_delta > delta) | |
807 | irq_delta = delta; | |
808 | ||
809 | rq->prev_irq_time += irq_delta; | |
810 | delta -= irq_delta; | |
095c0aa8 GC |
811 | #endif |
812 | #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING | |
c5905afb | 813 | if (static_key_false((¶virt_steal_rq_enabled))) { |
095c0aa8 GC |
814 | steal = paravirt_steal_clock(cpu_of(rq)); |
815 | steal -= rq->prev_steal_time_rq; | |
816 | ||
817 | if (unlikely(steal > delta)) | |
818 | steal = delta; | |
819 | ||
095c0aa8 | 820 | rq->prev_steal_time_rq += steal; |
095c0aa8 GC |
821 | delta -= steal; |
822 | } | |
823 | #endif | |
824 | ||
fe44d621 PZ |
825 | rq->clock_task += delta; |
826 | ||
095c0aa8 | 827 | #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) |
5d4dfddd | 828 | if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY)) |
095c0aa8 GC |
829 | sched_rt_avg_update(rq, irq_delta + steal); |
830 | #endif | |
aa483808 VP |
831 | } |
832 | ||
34f971f6 PZ |
833 | void sched_set_stop_task(int cpu, struct task_struct *stop) |
834 | { | |
835 | struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; | |
836 | struct task_struct *old_stop = cpu_rq(cpu)->stop; | |
837 | ||
838 | if (stop) { | |
839 | /* | |
840 | * Make it appear like a SCHED_FIFO task, its something | |
841 | * userspace knows about and won't get confused about. | |
842 | * | |
843 | * Also, it will make PI more or less work without too | |
844 | * much confusion -- but then, stop work should not | |
845 | * rely on PI working anyway. | |
846 | */ | |
847 | sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); | |
848 | ||
849 | stop->sched_class = &stop_sched_class; | |
850 | } | |
851 | ||
852 | cpu_rq(cpu)->stop = stop; | |
853 | ||
854 | if (old_stop) { | |
855 | /* | |
856 | * Reset it back to a normal scheduling class so that | |
857 | * it can die in pieces. | |
858 | */ | |
859 | old_stop->sched_class = &rt_sched_class; | |
860 | } | |
861 | } | |
862 | ||
14531189 | 863 | /* |
dd41f596 | 864 | * __normal_prio - return the priority that is based on the static prio |
14531189 | 865 | */ |
14531189 IM |
866 | static inline int __normal_prio(struct task_struct *p) |
867 | { | |
dd41f596 | 868 | return p->static_prio; |
14531189 IM |
869 | } |
870 | ||
b29739f9 IM |
871 | /* |
872 | * Calculate the expected normal priority: i.e. priority | |
873 | * without taking RT-inheritance into account. Might be | |
874 | * boosted by interactivity modifiers. Changes upon fork, | |
875 | * setprio syscalls, and whenever the interactivity | |
876 | * estimator recalculates. | |
877 | */ | |
36c8b586 | 878 | static inline int normal_prio(struct task_struct *p) |
b29739f9 IM |
879 | { |
880 | int prio; | |
881 | ||
aab03e05 DF |
882 | if (task_has_dl_policy(p)) |
883 | prio = MAX_DL_PRIO-1; | |
884 | else if (task_has_rt_policy(p)) | |
b29739f9 IM |
885 | prio = MAX_RT_PRIO-1 - p->rt_priority; |
886 | else | |
887 | prio = __normal_prio(p); | |
888 | return prio; | |
889 | } | |
890 | ||
891 | /* | |
892 | * Calculate the current priority, i.e. the priority | |
893 | * taken into account by the scheduler. This value might | |
894 | * be boosted by RT tasks, or might be boosted by | |
895 | * interactivity modifiers. Will be RT if the task got | |
896 | * RT-boosted. If not then it returns p->normal_prio. | |
897 | */ | |
36c8b586 | 898 | static int effective_prio(struct task_struct *p) |
b29739f9 IM |
899 | { |
900 | p->normal_prio = normal_prio(p); | |
901 | /* | |
902 | * If we are RT tasks or we were boosted to RT priority, | |
903 | * keep the priority unchanged. Otherwise, update priority | |
904 | * to the normal priority: | |
905 | */ | |
906 | if (!rt_prio(p->prio)) | |
907 | return p->normal_prio; | |
908 | return p->prio; | |
909 | } | |
910 | ||
1da177e4 LT |
911 | /** |
912 | * task_curr - is this task currently executing on a CPU? | |
913 | * @p: the task in question. | |
e69f6186 YB |
914 | * |
915 | * Return: 1 if the task is currently executing. 0 otherwise. | |
1da177e4 | 916 | */ |
36c8b586 | 917 | inline int task_curr(const struct task_struct *p) |
1da177e4 LT |
918 | { |
919 | return cpu_curr(task_cpu(p)) == p; | |
920 | } | |
921 | ||
67dfa1b7 | 922 | /* |
4c9a4bc8 PZ |
923 | * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock, |
924 | * use the balance_callback list if you want balancing. | |
925 | * | |
926 | * this means any call to check_class_changed() must be followed by a call to | |
927 | * balance_callback(). | |
67dfa1b7 | 928 | */ |
cb469845 SR |
929 | static inline void check_class_changed(struct rq *rq, struct task_struct *p, |
930 | const struct sched_class *prev_class, | |
da7a735e | 931 | int oldprio) |
cb469845 SR |
932 | { |
933 | if (prev_class != p->sched_class) { | |
934 | if (prev_class->switched_from) | |
da7a735e | 935 | prev_class->switched_from(rq, p); |
4c9a4bc8 | 936 | |
da7a735e | 937 | p->sched_class->switched_to(rq, p); |
2d3d891d | 938 | } else if (oldprio != p->prio || dl_task(p)) |
da7a735e | 939 | p->sched_class->prio_changed(rq, p, oldprio); |
cb469845 SR |
940 | } |
941 | ||
029632fb | 942 | void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) |
1e5a7405 PZ |
943 | { |
944 | const struct sched_class *class; | |
945 | ||
946 | if (p->sched_class == rq->curr->sched_class) { | |
947 | rq->curr->sched_class->check_preempt_curr(rq, p, flags); | |
948 | } else { | |
949 | for_each_class(class) { | |
950 | if (class == rq->curr->sched_class) | |
951 | break; | |
952 | if (class == p->sched_class) { | |
8875125e | 953 | resched_curr(rq); |
1e5a7405 PZ |
954 | break; |
955 | } | |
956 | } | |
957 | } | |
958 | ||
959 | /* | |
960 | * A queue event has occurred, and we're going to schedule. In | |
961 | * this case, we can save a useless back to back clock update. | |
962 | */ | |
da0c1e65 | 963 | if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr)) |
9edfbfed | 964 | rq_clock_skip_update(rq, true); |
1e5a7405 PZ |
965 | } |
966 | ||
1da177e4 | 967 | #ifdef CONFIG_SMP |
5cc389bc PZ |
968 | /* |
969 | * This is how migration works: | |
970 | * | |
971 | * 1) we invoke migration_cpu_stop() on the target CPU using | |
972 | * stop_one_cpu(). | |
973 | * 2) stopper starts to run (implicitly forcing the migrated thread | |
974 | * off the CPU) | |
975 | * 3) it checks whether the migrated task is still in the wrong runqueue. | |
976 | * 4) if it's in the wrong runqueue then the migration thread removes | |
977 | * it and puts it into the right queue. | |
978 | * 5) stopper completes and stop_one_cpu() returns and the migration | |
979 | * is done. | |
980 | */ | |
981 | ||
982 | /* | |
983 | * move_queued_task - move a queued task to new rq. | |
984 | * | |
985 | * Returns (locked) new rq. Old rq's lock is released. | |
986 | */ | |
5e16bbc2 | 987 | static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu) |
5cc389bc | 988 | { |
5cc389bc PZ |
989 | lockdep_assert_held(&rq->lock); |
990 | ||
5cc389bc | 991 | p->on_rq = TASK_ON_RQ_MIGRATING; |
3ea94de1 | 992 | dequeue_task(rq, p, 0); |
5cc389bc PZ |
993 | set_task_cpu(p, new_cpu); |
994 | raw_spin_unlock(&rq->lock); | |
995 | ||
996 | rq = cpu_rq(new_cpu); | |
997 | ||
998 | raw_spin_lock(&rq->lock); | |
999 | BUG_ON(task_cpu(p) != new_cpu); | |
5cc389bc | 1000 | enqueue_task(rq, p, 0); |
3ea94de1 | 1001 | p->on_rq = TASK_ON_RQ_QUEUED; |
5cc389bc PZ |
1002 | check_preempt_curr(rq, p, 0); |
1003 | ||
1004 | return rq; | |
1005 | } | |
1006 | ||
1007 | struct migration_arg { | |
1008 | struct task_struct *task; | |
1009 | int dest_cpu; | |
1010 | }; | |
1011 | ||
1012 | /* | |
1013 | * Move (not current) task off this cpu, onto dest cpu. We're doing | |
1014 | * this because either it can't run here any more (set_cpus_allowed() | |
1015 | * away from this CPU, or CPU going down), or because we're | |
1016 | * attempting to rebalance this task on exec (sched_exec). | |
1017 | * | |
1018 | * So we race with normal scheduler movements, but that's OK, as long | |
1019 | * as the task is no longer on this CPU. | |
5cc389bc | 1020 | */ |
5e16bbc2 | 1021 | static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu) |
5cc389bc | 1022 | { |
5cc389bc | 1023 | if (unlikely(!cpu_active(dest_cpu))) |
5e16bbc2 | 1024 | return rq; |
5cc389bc PZ |
1025 | |
1026 | /* Affinity changed (again). */ | |
1027 | if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) | |
5e16bbc2 | 1028 | return rq; |
5cc389bc | 1029 | |
5e16bbc2 PZ |
1030 | rq = move_queued_task(rq, p, dest_cpu); |
1031 | ||
1032 | return rq; | |
5cc389bc PZ |
1033 | } |
1034 | ||
1035 | /* | |
1036 | * migration_cpu_stop - this will be executed by a highprio stopper thread | |
1037 | * and performs thread migration by bumping thread off CPU then | |
1038 | * 'pushing' onto another runqueue. | |
1039 | */ | |
1040 | static int migration_cpu_stop(void *data) | |
1041 | { | |
1042 | struct migration_arg *arg = data; | |
5e16bbc2 PZ |
1043 | struct task_struct *p = arg->task; |
1044 | struct rq *rq = this_rq(); | |
5cc389bc PZ |
1045 | |
1046 | /* | |
1047 | * The original target cpu might have gone down and we might | |
1048 | * be on another cpu but it doesn't matter. | |
1049 | */ | |
1050 | local_irq_disable(); | |
1051 | /* | |
1052 | * We need to explicitly wake pending tasks before running | |
1053 | * __migrate_task() such that we will not miss enforcing cpus_allowed | |
1054 | * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test. | |
1055 | */ | |
1056 | sched_ttwu_pending(); | |
5e16bbc2 PZ |
1057 | |
1058 | raw_spin_lock(&p->pi_lock); | |
1059 | raw_spin_lock(&rq->lock); | |
1060 | /* | |
1061 | * If task_rq(p) != rq, it cannot be migrated here, because we're | |
1062 | * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because | |
1063 | * we're holding p->pi_lock. | |
1064 | */ | |
1065 | if (task_rq(p) == rq && task_on_rq_queued(p)) | |
1066 | rq = __migrate_task(rq, p, arg->dest_cpu); | |
1067 | raw_spin_unlock(&rq->lock); | |
1068 | raw_spin_unlock(&p->pi_lock); | |
1069 | ||
5cc389bc PZ |
1070 | local_irq_enable(); |
1071 | return 0; | |
1072 | } | |
1073 | ||
c5b28038 PZ |
1074 | /* |
1075 | * sched_class::set_cpus_allowed must do the below, but is not required to | |
1076 | * actually call this function. | |
1077 | */ | |
1078 | void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask) | |
5cc389bc | 1079 | { |
5cc389bc PZ |
1080 | cpumask_copy(&p->cpus_allowed, new_mask); |
1081 | p->nr_cpus_allowed = cpumask_weight(new_mask); | |
1082 | } | |
1083 | ||
c5b28038 PZ |
1084 | void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) |
1085 | { | |
6c37067e PZ |
1086 | struct rq *rq = task_rq(p); |
1087 | bool queued, running; | |
1088 | ||
c5b28038 | 1089 | lockdep_assert_held(&p->pi_lock); |
6c37067e PZ |
1090 | |
1091 | queued = task_on_rq_queued(p); | |
1092 | running = task_current(rq, p); | |
1093 | ||
1094 | if (queued) { | |
1095 | /* | |
1096 | * Because __kthread_bind() calls this on blocked tasks without | |
1097 | * holding rq->lock. | |
1098 | */ | |
1099 | lockdep_assert_held(&rq->lock); | |
1de64443 | 1100 | dequeue_task(rq, p, DEQUEUE_SAVE); |
6c37067e PZ |
1101 | } |
1102 | if (running) | |
1103 | put_prev_task(rq, p); | |
1104 | ||
c5b28038 | 1105 | p->sched_class->set_cpus_allowed(p, new_mask); |
6c37067e PZ |
1106 | |
1107 | if (running) | |
1108 | p->sched_class->set_curr_task(rq); | |
1109 | if (queued) | |
1de64443 | 1110 | enqueue_task(rq, p, ENQUEUE_RESTORE); |
c5b28038 PZ |
1111 | } |
1112 | ||
5cc389bc PZ |
1113 | /* |
1114 | * Change a given task's CPU affinity. Migrate the thread to a | |
1115 | * proper CPU and schedule it away if the CPU it's executing on | |
1116 | * is removed from the allowed bitmask. | |
1117 | * | |
1118 | * NOTE: the caller must have a valid reference to the task, the | |
1119 | * task must not exit() & deallocate itself prematurely. The | |
1120 | * call is not atomic; no spinlocks may be held. | |
1121 | */ | |
25834c73 PZ |
1122 | static int __set_cpus_allowed_ptr(struct task_struct *p, |
1123 | const struct cpumask *new_mask, bool check) | |
5cc389bc | 1124 | { |
e9d867a6 | 1125 | const struct cpumask *cpu_valid_mask = cpu_active_mask; |
5cc389bc | 1126 | unsigned int dest_cpu; |
eb580751 PZ |
1127 | struct rq_flags rf; |
1128 | struct rq *rq; | |
5cc389bc PZ |
1129 | int ret = 0; |
1130 | ||
eb580751 | 1131 | rq = task_rq_lock(p, &rf); |
5cc389bc | 1132 | |
e9d867a6 PZI |
1133 | if (p->flags & PF_KTHREAD) { |
1134 | /* | |
1135 | * Kernel threads are allowed on online && !active CPUs | |
1136 | */ | |
1137 | cpu_valid_mask = cpu_online_mask; | |
1138 | } | |
1139 | ||
25834c73 PZ |
1140 | /* |
1141 | * Must re-check here, to close a race against __kthread_bind(), | |
1142 | * sched_setaffinity() is not guaranteed to observe the flag. | |
1143 | */ | |
1144 | if (check && (p->flags & PF_NO_SETAFFINITY)) { | |
1145 | ret = -EINVAL; | |
1146 | goto out; | |
1147 | } | |
1148 | ||
5cc389bc PZ |
1149 | if (cpumask_equal(&p->cpus_allowed, new_mask)) |
1150 | goto out; | |
1151 | ||
e9d867a6 | 1152 | if (!cpumask_intersects(new_mask, cpu_valid_mask)) { |
5cc389bc PZ |
1153 | ret = -EINVAL; |
1154 | goto out; | |
1155 | } | |
1156 | ||
1157 | do_set_cpus_allowed(p, new_mask); | |
1158 | ||
e9d867a6 PZI |
1159 | if (p->flags & PF_KTHREAD) { |
1160 | /* | |
1161 | * For kernel threads that do indeed end up on online && | |
1162 | * !active we want to ensure they are strict per-cpu threads. | |
1163 | */ | |
1164 | WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) && | |
1165 | !cpumask_intersects(new_mask, cpu_active_mask) && | |
1166 | p->nr_cpus_allowed != 1); | |
1167 | } | |
1168 | ||
5cc389bc PZ |
1169 | /* Can the task run on the task's current CPU? If so, we're done */ |
1170 | if (cpumask_test_cpu(task_cpu(p), new_mask)) | |
1171 | goto out; | |
1172 | ||
e9d867a6 | 1173 | dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask); |
5cc389bc PZ |
1174 | if (task_running(rq, p) || p->state == TASK_WAKING) { |
1175 | struct migration_arg arg = { p, dest_cpu }; | |
1176 | /* Need help from migration thread: drop lock and wait. */ | |
eb580751 | 1177 | task_rq_unlock(rq, p, &rf); |
5cc389bc PZ |
1178 | stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); |
1179 | tlb_migrate_finish(p->mm); | |
1180 | return 0; | |
cbce1a68 PZ |
1181 | } else if (task_on_rq_queued(p)) { |
1182 | /* | |
1183 | * OK, since we're going to drop the lock immediately | |
1184 | * afterwards anyway. | |
1185 | */ | |
e7904a28 | 1186 | lockdep_unpin_lock(&rq->lock, rf.cookie); |
5e16bbc2 | 1187 | rq = move_queued_task(rq, p, dest_cpu); |
e7904a28 | 1188 | lockdep_repin_lock(&rq->lock, rf.cookie); |
cbce1a68 | 1189 | } |
5cc389bc | 1190 | out: |
eb580751 | 1191 | task_rq_unlock(rq, p, &rf); |
5cc389bc PZ |
1192 | |
1193 | return ret; | |
1194 | } | |
25834c73 PZ |
1195 | |
1196 | int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) | |
1197 | { | |
1198 | return __set_cpus_allowed_ptr(p, new_mask, false); | |
1199 | } | |
5cc389bc PZ |
1200 | EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); |
1201 | ||
dd41f596 | 1202 | void set_task_cpu(struct task_struct *p, unsigned int new_cpu) |
c65cc870 | 1203 | { |
e2912009 PZ |
1204 | #ifdef CONFIG_SCHED_DEBUG |
1205 | /* | |
1206 | * We should never call set_task_cpu() on a blocked task, | |
1207 | * ttwu() will sort out the placement. | |
1208 | */ | |
077614ee | 1209 | WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && |
e2336f6e | 1210 | !p->on_rq); |
0122ec5b | 1211 | |
3ea94de1 JP |
1212 | /* |
1213 | * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING, | |
1214 | * because schedstat_wait_{start,end} rebase migrating task's wait_start | |
1215 | * time relying on p->on_rq. | |
1216 | */ | |
1217 | WARN_ON_ONCE(p->state == TASK_RUNNING && | |
1218 | p->sched_class == &fair_sched_class && | |
1219 | (p->on_rq && !task_on_rq_migrating(p))); | |
1220 | ||
0122ec5b | 1221 | #ifdef CONFIG_LOCKDEP |
6c6c54e1 PZ |
1222 | /* |
1223 | * The caller should hold either p->pi_lock or rq->lock, when changing | |
1224 | * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. | |
1225 | * | |
1226 | * sched_move_task() holds both and thus holding either pins the cgroup, | |
8323f26c | 1227 | * see task_group(). |
6c6c54e1 PZ |
1228 | * |
1229 | * Furthermore, all task_rq users should acquire both locks, see | |
1230 | * task_rq_lock(). | |
1231 | */ | |
0122ec5b PZ |
1232 | WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || |
1233 | lockdep_is_held(&task_rq(p)->lock))); | |
1234 | #endif | |
e2912009 PZ |
1235 | #endif |
1236 | ||
de1d7286 | 1237 | trace_sched_migrate_task(p, new_cpu); |
cbc34ed1 | 1238 | |
0c69774e | 1239 | if (task_cpu(p) != new_cpu) { |
0a74bef8 | 1240 | if (p->sched_class->migrate_task_rq) |
5a4fd036 | 1241 | p->sched_class->migrate_task_rq(p); |
0c69774e | 1242 | p->se.nr_migrations++; |
ff303e66 | 1243 | perf_event_task_migrate(p); |
0c69774e | 1244 | } |
dd41f596 IM |
1245 | |
1246 | __set_task_cpu(p, new_cpu); | |
c65cc870 IM |
1247 | } |
1248 | ||
ac66f547 PZ |
1249 | static void __migrate_swap_task(struct task_struct *p, int cpu) |
1250 | { | |
da0c1e65 | 1251 | if (task_on_rq_queued(p)) { |
ac66f547 PZ |
1252 | struct rq *src_rq, *dst_rq; |
1253 | ||
1254 | src_rq = task_rq(p); | |
1255 | dst_rq = cpu_rq(cpu); | |
1256 | ||
3ea94de1 | 1257 | p->on_rq = TASK_ON_RQ_MIGRATING; |
ac66f547 PZ |
1258 | deactivate_task(src_rq, p, 0); |
1259 | set_task_cpu(p, cpu); | |
1260 | activate_task(dst_rq, p, 0); | |
3ea94de1 | 1261 | p->on_rq = TASK_ON_RQ_QUEUED; |
ac66f547 PZ |
1262 | check_preempt_curr(dst_rq, p, 0); |
1263 | } else { | |
1264 | /* | |
1265 | * Task isn't running anymore; make it appear like we migrated | |
1266 | * it before it went to sleep. This means on wakeup we make the | |
1267 | * previous cpu our targer instead of where it really is. | |
1268 | */ | |
1269 | p->wake_cpu = cpu; | |
1270 | } | |
1271 | } | |
1272 | ||
1273 | struct migration_swap_arg { | |
1274 | struct task_struct *src_task, *dst_task; | |
1275 | int src_cpu, dst_cpu; | |
1276 | }; | |
1277 | ||
1278 | static int migrate_swap_stop(void *data) | |
1279 | { | |
1280 | struct migration_swap_arg *arg = data; | |
1281 | struct rq *src_rq, *dst_rq; | |
1282 | int ret = -EAGAIN; | |
1283 | ||
62694cd5 PZ |
1284 | if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu)) |
1285 | return -EAGAIN; | |
1286 | ||
ac66f547 PZ |
1287 | src_rq = cpu_rq(arg->src_cpu); |
1288 | dst_rq = cpu_rq(arg->dst_cpu); | |
1289 | ||
74602315 PZ |
1290 | double_raw_lock(&arg->src_task->pi_lock, |
1291 | &arg->dst_task->pi_lock); | |
ac66f547 | 1292 | double_rq_lock(src_rq, dst_rq); |
62694cd5 | 1293 | |
ac66f547 PZ |
1294 | if (task_cpu(arg->dst_task) != arg->dst_cpu) |
1295 | goto unlock; | |
1296 | ||
1297 | if (task_cpu(arg->src_task) != arg->src_cpu) | |
1298 | goto unlock; | |
1299 | ||
1300 | if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task))) | |
1301 | goto unlock; | |
1302 | ||
1303 | if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task))) | |
1304 | goto unlock; | |
1305 | ||
1306 | __migrate_swap_task(arg->src_task, arg->dst_cpu); | |
1307 | __migrate_swap_task(arg->dst_task, arg->src_cpu); | |
1308 | ||
1309 | ret = 0; | |
1310 | ||
1311 | unlock: | |
1312 | double_rq_unlock(src_rq, dst_rq); | |
74602315 PZ |
1313 | raw_spin_unlock(&arg->dst_task->pi_lock); |
1314 | raw_spin_unlock(&arg->src_task->pi_lock); | |
ac66f547 PZ |
1315 | |
1316 | return ret; | |
1317 | } | |
1318 | ||
1319 | /* | |
1320 | * Cross migrate two tasks | |
1321 | */ | |
1322 | int migrate_swap(struct task_struct *cur, struct task_struct *p) | |
1323 | { | |
1324 | struct migration_swap_arg arg; | |
1325 | int ret = -EINVAL; | |
1326 | ||
ac66f547 PZ |
1327 | arg = (struct migration_swap_arg){ |
1328 | .src_task = cur, | |
1329 | .src_cpu = task_cpu(cur), | |
1330 | .dst_task = p, | |
1331 | .dst_cpu = task_cpu(p), | |
1332 | }; | |
1333 | ||
1334 | if (arg.src_cpu == arg.dst_cpu) | |
1335 | goto out; | |
1336 | ||
6acce3ef PZ |
1337 | /* |
1338 | * These three tests are all lockless; this is OK since all of them | |
1339 | * will be re-checked with proper locks held further down the line. | |
1340 | */ | |
ac66f547 PZ |
1341 | if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) |
1342 | goto out; | |
1343 | ||
1344 | if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task))) | |
1345 | goto out; | |
1346 | ||
1347 | if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task))) | |
1348 | goto out; | |
1349 | ||
286549dc | 1350 | trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu); |
ac66f547 PZ |
1351 | ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); |
1352 | ||
1353 | out: | |
ac66f547 PZ |
1354 | return ret; |
1355 | } | |
1356 | ||
1da177e4 LT |
1357 | /* |
1358 | * wait_task_inactive - wait for a thread to unschedule. | |
1359 | * | |
85ba2d86 RM |
1360 | * If @match_state is nonzero, it's the @p->state value just checked and |
1361 | * not expected to change. If it changes, i.e. @p might have woken up, | |
1362 | * then return zero. When we succeed in waiting for @p to be off its CPU, | |
1363 | * we return a positive number (its total switch count). If a second call | |
1364 | * a short while later returns the same number, the caller can be sure that | |
1365 | * @p has remained unscheduled the whole time. | |
1366 | * | |
1da177e4 LT |
1367 | * The caller must ensure that the task *will* unschedule sometime soon, |
1368 | * else this function might spin for a *long* time. This function can't | |
1369 | * be called with interrupts off, or it may introduce deadlock with | |
1370 | * smp_call_function() if an IPI is sent by the same process we are | |
1371 | * waiting to become inactive. | |
1372 | */ | |
85ba2d86 | 1373 | unsigned long wait_task_inactive(struct task_struct *p, long match_state) |
1da177e4 | 1374 | { |
da0c1e65 | 1375 | int running, queued; |
eb580751 | 1376 | struct rq_flags rf; |
85ba2d86 | 1377 | unsigned long ncsw; |
70b97a7f | 1378 | struct rq *rq; |
1da177e4 | 1379 | |
3a5c359a AK |
1380 | for (;;) { |
1381 | /* | |
1382 | * We do the initial early heuristics without holding | |
1383 | * any task-queue locks at all. We'll only try to get | |
1384 | * the runqueue lock when things look like they will | |
1385 | * work out! | |
1386 | */ | |
1387 | rq = task_rq(p); | |
fa490cfd | 1388 | |
3a5c359a AK |
1389 | /* |
1390 | * If the task is actively running on another CPU | |
1391 | * still, just relax and busy-wait without holding | |
1392 | * any locks. | |
1393 | * | |
1394 | * NOTE! Since we don't hold any locks, it's not | |
1395 | * even sure that "rq" stays as the right runqueue! | |
1396 | * But we don't care, since "task_running()" will | |
1397 | * return false if the runqueue has changed and p | |
1398 | * is actually now running somewhere else! | |
1399 | */ | |
85ba2d86 RM |
1400 | while (task_running(rq, p)) { |
1401 | if (match_state && unlikely(p->state != match_state)) | |
1402 | return 0; | |
3a5c359a | 1403 | cpu_relax(); |
85ba2d86 | 1404 | } |
fa490cfd | 1405 | |
3a5c359a AK |
1406 | /* |
1407 | * Ok, time to look more closely! We need the rq | |
1408 | * lock now, to be *sure*. If we're wrong, we'll | |
1409 | * just go back and repeat. | |
1410 | */ | |
eb580751 | 1411 | rq = task_rq_lock(p, &rf); |
27a9da65 | 1412 | trace_sched_wait_task(p); |
3a5c359a | 1413 | running = task_running(rq, p); |
da0c1e65 | 1414 | queued = task_on_rq_queued(p); |
85ba2d86 | 1415 | ncsw = 0; |
f31e11d8 | 1416 | if (!match_state || p->state == match_state) |
93dcf55f | 1417 | ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ |
eb580751 | 1418 | task_rq_unlock(rq, p, &rf); |
fa490cfd | 1419 | |
85ba2d86 RM |
1420 | /* |
1421 | * If it changed from the expected state, bail out now. | |
1422 | */ | |
1423 | if (unlikely(!ncsw)) | |
1424 | break; | |
1425 | ||
3a5c359a AK |
1426 | /* |
1427 | * Was it really running after all now that we | |
1428 | * checked with the proper locks actually held? | |
1429 | * | |
1430 | * Oops. Go back and try again.. | |
1431 | */ | |
1432 | if (unlikely(running)) { | |
1433 | cpu_relax(); | |
1434 | continue; | |
1435 | } | |
fa490cfd | 1436 | |
3a5c359a AK |
1437 | /* |
1438 | * It's not enough that it's not actively running, | |
1439 | * it must be off the runqueue _entirely_, and not | |
1440 | * preempted! | |
1441 | * | |
80dd99b3 | 1442 | * So if it was still runnable (but just not actively |
3a5c359a AK |
1443 | * running right now), it's preempted, and we should |
1444 | * yield - it could be a while. | |
1445 | */ | |
da0c1e65 | 1446 | if (unlikely(queued)) { |
8eb90c30 TG |
1447 | ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); |
1448 | ||
1449 | set_current_state(TASK_UNINTERRUPTIBLE); | |
1450 | schedule_hrtimeout(&to, HRTIMER_MODE_REL); | |
3a5c359a AK |
1451 | continue; |
1452 | } | |
fa490cfd | 1453 | |
3a5c359a AK |
1454 | /* |
1455 | * Ahh, all good. It wasn't running, and it wasn't | |
1456 | * runnable, which means that it will never become | |
1457 | * running in the future either. We're all done! | |
1458 | */ | |
1459 | break; | |
1460 | } | |
85ba2d86 RM |
1461 | |
1462 | return ncsw; | |
1da177e4 LT |
1463 | } |
1464 | ||
1465 | /*** | |
1466 | * kick_process - kick a running thread to enter/exit the kernel | |
1467 | * @p: the to-be-kicked thread | |
1468 | * | |
1469 | * Cause a process which is running on another CPU to enter | |
1470 | * kernel-mode, without any delay. (to get signals handled.) | |
1471 | * | |
25985edc | 1472 | * NOTE: this function doesn't have to take the runqueue lock, |
1da177e4 LT |
1473 | * because all it wants to ensure is that the remote task enters |
1474 | * the kernel. If the IPI races and the task has been migrated | |
1475 | * to another CPU then no harm is done and the purpose has been | |
1476 | * achieved as well. | |
1477 | */ | |
36c8b586 | 1478 | void kick_process(struct task_struct *p) |
1da177e4 LT |
1479 | { |
1480 | int cpu; | |
1481 | ||
1482 | preempt_disable(); | |
1483 | cpu = task_cpu(p); | |
1484 | if ((cpu != smp_processor_id()) && task_curr(p)) | |
1485 | smp_send_reschedule(cpu); | |
1486 | preempt_enable(); | |
1487 | } | |
b43e3521 | 1488 | EXPORT_SYMBOL_GPL(kick_process); |
1da177e4 | 1489 | |
30da688e | 1490 | /* |
013fdb80 | 1491 | * ->cpus_allowed is protected by both rq->lock and p->pi_lock |
e9d867a6 PZI |
1492 | * |
1493 | * A few notes on cpu_active vs cpu_online: | |
1494 | * | |
1495 | * - cpu_active must be a subset of cpu_online | |
1496 | * | |
1497 | * - on cpu-up we allow per-cpu kthreads on the online && !active cpu, | |
1498 | * see __set_cpus_allowed_ptr(). At this point the newly online | |
1499 | * cpu isn't yet part of the sched domains, and balancing will not | |
1500 | * see it. | |
1501 | * | |
1502 | * - on cpu-down we clear cpu_active() to mask the sched domains and | |
1503 | * avoid the load balancer to place new tasks on the to be removed | |
1504 | * cpu. Existing tasks will remain running there and will be taken | |
1505 | * off. | |
1506 | * | |
1507 | * This means that fallback selection must not select !active CPUs. | |
1508 | * And can assume that any active CPU must be online. Conversely | |
1509 | * select_task_rq() below may allow selection of !active CPUs in order | |
1510 | * to satisfy the above rules. | |
30da688e | 1511 | */ |
5da9a0fb PZ |
1512 | static int select_fallback_rq(int cpu, struct task_struct *p) |
1513 | { | |
aa00d89c TC |
1514 | int nid = cpu_to_node(cpu); |
1515 | const struct cpumask *nodemask = NULL; | |
2baab4e9 PZ |
1516 | enum { cpuset, possible, fail } state = cpuset; |
1517 | int dest_cpu; | |
5da9a0fb | 1518 | |
aa00d89c TC |
1519 | /* |
1520 | * If the node that the cpu is on has been offlined, cpu_to_node() | |
1521 | * will return -1. There is no cpu on the node, and we should | |
1522 | * select the cpu on the other node. | |
1523 | */ | |
1524 | if (nid != -1) { | |
1525 | nodemask = cpumask_of_node(nid); | |
1526 | ||
1527 | /* Look for allowed, online CPU in same node. */ | |
1528 | for_each_cpu(dest_cpu, nodemask) { | |
aa00d89c TC |
1529 | if (!cpu_active(dest_cpu)) |
1530 | continue; | |
1531 | if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) | |
1532 | return dest_cpu; | |
1533 | } | |
2baab4e9 | 1534 | } |
5da9a0fb | 1535 | |
2baab4e9 PZ |
1536 | for (;;) { |
1537 | /* Any allowed, online CPU? */ | |
e3831edd | 1538 | for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { |
2baab4e9 PZ |
1539 | if (!cpu_active(dest_cpu)) |
1540 | continue; | |
1541 | goto out; | |
1542 | } | |
5da9a0fb | 1543 | |
e73e85f0 | 1544 | /* No more Mr. Nice Guy. */ |
2baab4e9 PZ |
1545 | switch (state) { |
1546 | case cpuset: | |
e73e85f0 ON |
1547 | if (IS_ENABLED(CONFIG_CPUSETS)) { |
1548 | cpuset_cpus_allowed_fallback(p); | |
1549 | state = possible; | |
1550 | break; | |
1551 | } | |
1552 | /* fall-through */ | |
2baab4e9 PZ |
1553 | case possible: |
1554 | do_set_cpus_allowed(p, cpu_possible_mask); | |
1555 | state = fail; | |
1556 | break; | |
1557 | ||
1558 | case fail: | |
1559 | BUG(); | |
1560 | break; | |
1561 | } | |
1562 | } | |
1563 | ||
1564 | out: | |
1565 | if (state != cpuset) { | |
1566 | /* | |
1567 | * Don't tell them about moving exiting tasks or | |
1568 | * kernel threads (both mm NULL), since they never | |
1569 | * leave kernel. | |
1570 | */ | |
1571 | if (p->mm && printk_ratelimit()) { | |
aac74dc4 | 1572 | printk_deferred("process %d (%s) no longer affine to cpu%d\n", |
2baab4e9 PZ |
1573 | task_pid_nr(p), p->comm, cpu); |
1574 | } | |
5da9a0fb PZ |
1575 | } |
1576 | ||
1577 | return dest_cpu; | |
1578 | } | |
1579 | ||
e2912009 | 1580 | /* |
013fdb80 | 1581 | * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. |
e2912009 | 1582 | */ |
970b13ba | 1583 | static inline |
ac66f547 | 1584 | int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) |
970b13ba | 1585 | { |
cbce1a68 PZ |
1586 | lockdep_assert_held(&p->pi_lock); |
1587 | ||
50605ffb | 1588 | if (tsk_nr_cpus_allowed(p) > 1) |
6c1d9410 | 1589 | cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); |
e9d867a6 PZI |
1590 | else |
1591 | cpu = cpumask_any(tsk_cpus_allowed(p)); | |
e2912009 PZ |
1592 | |
1593 | /* | |
1594 | * In order not to call set_task_cpu() on a blocking task we need | |
1595 | * to rely on ttwu() to place the task on a valid ->cpus_allowed | |
1596 | * cpu. | |
1597 | * | |
1598 | * Since this is common to all placement strategies, this lives here. | |
1599 | * | |
1600 | * [ this allows ->select_task() to simply return task_cpu(p) and | |
1601 | * not worry about this generic constraint ] | |
1602 | */ | |
fa17b507 | 1603 | if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || |
70f11205 | 1604 | !cpu_online(cpu))) |
5da9a0fb | 1605 | cpu = select_fallback_rq(task_cpu(p), p); |
e2912009 PZ |
1606 | |
1607 | return cpu; | |
970b13ba | 1608 | } |
09a40af5 MG |
1609 | |
1610 | static void update_avg(u64 *avg, u64 sample) | |
1611 | { | |
1612 | s64 diff = sample - *avg; | |
1613 | *avg += diff >> 3; | |
1614 | } | |
25834c73 PZ |
1615 | |
1616 | #else | |
1617 | ||
1618 | static inline int __set_cpus_allowed_ptr(struct task_struct *p, | |
1619 | const struct cpumask *new_mask, bool check) | |
1620 | { | |
1621 | return set_cpus_allowed_ptr(p, new_mask); | |
1622 | } | |
1623 | ||
5cc389bc | 1624 | #endif /* CONFIG_SMP */ |
970b13ba | 1625 | |
d7c01d27 | 1626 | static void |
b84cb5df | 1627 | ttwu_stat(struct task_struct *p, int cpu, int wake_flags) |
9ed3811a | 1628 | { |
d7c01d27 | 1629 | #ifdef CONFIG_SCHEDSTATS |
b84cb5df PZ |
1630 | struct rq *rq = this_rq(); |
1631 | ||
d7c01d27 PZ |
1632 | #ifdef CONFIG_SMP |
1633 | int this_cpu = smp_processor_id(); | |
1634 | ||
1635 | if (cpu == this_cpu) { | |
1636 | schedstat_inc(rq, ttwu_local); | |
1637 | schedstat_inc(p, se.statistics.nr_wakeups_local); | |
1638 | } else { | |
1639 | struct sched_domain *sd; | |
1640 | ||
1641 | schedstat_inc(p, se.statistics.nr_wakeups_remote); | |
057f3fad | 1642 | rcu_read_lock(); |
d7c01d27 PZ |
1643 | for_each_domain(this_cpu, sd) { |
1644 | if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { | |
1645 | schedstat_inc(sd, ttwu_wake_remote); | |
1646 | break; | |
1647 | } | |
1648 | } | |
057f3fad | 1649 | rcu_read_unlock(); |
d7c01d27 | 1650 | } |
f339b9dc PZ |
1651 | |
1652 | if (wake_flags & WF_MIGRATED) | |
1653 | schedstat_inc(p, se.statistics.nr_wakeups_migrate); | |
1654 | ||
d7c01d27 PZ |
1655 | #endif /* CONFIG_SMP */ |
1656 | ||
1657 | schedstat_inc(rq, ttwu_count); | |
9ed3811a | 1658 | schedstat_inc(p, se.statistics.nr_wakeups); |
d7c01d27 PZ |
1659 | |
1660 | if (wake_flags & WF_SYNC) | |
9ed3811a | 1661 | schedstat_inc(p, se.statistics.nr_wakeups_sync); |
d7c01d27 | 1662 | |
d7c01d27 PZ |
1663 | #endif /* CONFIG_SCHEDSTATS */ |
1664 | } | |
1665 | ||
1de64443 | 1666 | static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) |
d7c01d27 | 1667 | { |
9ed3811a | 1668 | activate_task(rq, p, en_flags); |
da0c1e65 | 1669 | p->on_rq = TASK_ON_RQ_QUEUED; |
c2f7115e PZ |
1670 | |
1671 | /* if a worker is waking up, notify workqueue */ | |
1672 | if (p->flags & PF_WQ_WORKER) | |
1673 | wq_worker_waking_up(p, cpu_of(rq)); | |
9ed3811a TH |
1674 | } |
1675 | ||
23f41eeb PZ |
1676 | /* |
1677 | * Mark the task runnable and perform wakeup-preemption. | |
1678 | */ | |
e7904a28 PZ |
1679 | static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags, |
1680 | struct pin_cookie cookie) | |
9ed3811a | 1681 | { |
9ed3811a | 1682 | check_preempt_curr(rq, p, wake_flags); |
9ed3811a | 1683 | p->state = TASK_RUNNING; |
fbd705a0 PZ |
1684 | trace_sched_wakeup(p); |
1685 | ||
9ed3811a | 1686 | #ifdef CONFIG_SMP |
4c9a4bc8 PZ |
1687 | if (p->sched_class->task_woken) { |
1688 | /* | |
cbce1a68 PZ |
1689 | * Our task @p is fully woken up and running; so its safe to |
1690 | * drop the rq->lock, hereafter rq is only used for statistics. | |
4c9a4bc8 | 1691 | */ |
e7904a28 | 1692 | lockdep_unpin_lock(&rq->lock, cookie); |
9ed3811a | 1693 | p->sched_class->task_woken(rq, p); |
e7904a28 | 1694 | lockdep_repin_lock(&rq->lock, cookie); |
4c9a4bc8 | 1695 | } |
9ed3811a | 1696 | |
e69c6341 | 1697 | if (rq->idle_stamp) { |
78becc27 | 1698 | u64 delta = rq_clock(rq) - rq->idle_stamp; |
9bd721c5 | 1699 | u64 max = 2*rq->max_idle_balance_cost; |
9ed3811a | 1700 | |
abfafa54 JL |
1701 | update_avg(&rq->avg_idle, delta); |
1702 | ||
1703 | if (rq->avg_idle > max) | |
9ed3811a | 1704 | rq->avg_idle = max; |
abfafa54 | 1705 | |
9ed3811a TH |
1706 | rq->idle_stamp = 0; |
1707 | } | |
1708 | #endif | |
1709 | } | |
1710 | ||
c05fbafb | 1711 | static void |
e7904a28 PZ |
1712 | ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags, |
1713 | struct pin_cookie cookie) | |
c05fbafb | 1714 | { |
b5179ac7 PZ |
1715 | int en_flags = ENQUEUE_WAKEUP; |
1716 | ||
cbce1a68 PZ |
1717 | lockdep_assert_held(&rq->lock); |
1718 | ||
c05fbafb PZ |
1719 | #ifdef CONFIG_SMP |
1720 | if (p->sched_contributes_to_load) | |
1721 | rq->nr_uninterruptible--; | |
b5179ac7 | 1722 | |
b5179ac7 | 1723 | if (wake_flags & WF_MIGRATED) |
59efa0ba | 1724 | en_flags |= ENQUEUE_MIGRATED; |
c05fbafb PZ |
1725 | #endif |
1726 | ||
b5179ac7 | 1727 | ttwu_activate(rq, p, en_flags); |
e7904a28 | 1728 | ttwu_do_wakeup(rq, p, wake_flags, cookie); |
c05fbafb PZ |
1729 | } |
1730 | ||
1731 | /* | |
1732 | * Called in case the task @p isn't fully descheduled from its runqueue, | |
1733 | * in this case we must do a remote wakeup. Its a 'light' wakeup though, | |
1734 | * since all we need to do is flip p->state to TASK_RUNNING, since | |
1735 | * the task is still ->on_rq. | |
1736 | */ | |
1737 | static int ttwu_remote(struct task_struct *p, int wake_flags) | |
1738 | { | |
eb580751 | 1739 | struct rq_flags rf; |
c05fbafb PZ |
1740 | struct rq *rq; |
1741 | int ret = 0; | |
1742 | ||
eb580751 | 1743 | rq = __task_rq_lock(p, &rf); |
da0c1e65 | 1744 | if (task_on_rq_queued(p)) { |
1ad4ec0d FW |
1745 | /* check_preempt_curr() may use rq clock */ |
1746 | update_rq_clock(rq); | |
e7904a28 | 1747 | ttwu_do_wakeup(rq, p, wake_flags, rf.cookie); |
c05fbafb PZ |
1748 | ret = 1; |
1749 | } | |
eb580751 | 1750 | __task_rq_unlock(rq, &rf); |
c05fbafb PZ |
1751 | |
1752 | return ret; | |
1753 | } | |
1754 | ||
317f3941 | 1755 | #ifdef CONFIG_SMP |
e3baac47 | 1756 | void sched_ttwu_pending(void) |
317f3941 PZ |
1757 | { |
1758 | struct rq *rq = this_rq(); | |
fa14ff4a | 1759 | struct llist_node *llist = llist_del_all(&rq->wake_list); |
e7904a28 | 1760 | struct pin_cookie cookie; |
fa14ff4a | 1761 | struct task_struct *p; |
e3baac47 | 1762 | unsigned long flags; |
317f3941 | 1763 | |
e3baac47 PZ |
1764 | if (!llist) |
1765 | return; | |
1766 | ||
1767 | raw_spin_lock_irqsave(&rq->lock, flags); | |
e7904a28 | 1768 | cookie = lockdep_pin_lock(&rq->lock); |
317f3941 | 1769 | |
fa14ff4a | 1770 | while (llist) { |
b7e7ade3 PZ |
1771 | int wake_flags = 0; |
1772 | ||
fa14ff4a PZ |
1773 | p = llist_entry(llist, struct task_struct, wake_entry); |
1774 | llist = llist_next(llist); | |
b7e7ade3 PZ |
1775 | |
1776 | if (p->sched_remote_wakeup) | |
1777 | wake_flags = WF_MIGRATED; | |
1778 | ||
1779 | ttwu_do_activate(rq, p, wake_flags, cookie); | |
317f3941 PZ |
1780 | } |
1781 | ||
e7904a28 | 1782 | lockdep_unpin_lock(&rq->lock, cookie); |
e3baac47 | 1783 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
317f3941 PZ |
1784 | } |
1785 | ||
1786 | void scheduler_ipi(void) | |
1787 | { | |
f27dde8d PZ |
1788 | /* |
1789 | * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting | |
1790 | * TIF_NEED_RESCHED remotely (for the first time) will also send | |
1791 | * this IPI. | |
1792 | */ | |
8cb75e0c | 1793 | preempt_fold_need_resched(); |
f27dde8d | 1794 | |
fd2ac4f4 | 1795 | if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick()) |
c5d753a5 PZ |
1796 | return; |
1797 | ||
1798 | /* | |
1799 | * Not all reschedule IPI handlers call irq_enter/irq_exit, since | |
1800 | * traditionally all their work was done from the interrupt return | |
1801 | * path. Now that we actually do some work, we need to make sure | |
1802 | * we do call them. | |
1803 | * | |
1804 | * Some archs already do call them, luckily irq_enter/exit nest | |
1805 | * properly. | |
1806 | * | |
1807 | * Arguably we should visit all archs and update all handlers, | |
1808 | * however a fair share of IPIs are still resched only so this would | |
1809 | * somewhat pessimize the simple resched case. | |
1810 | */ | |
1811 | irq_enter(); | |
fa14ff4a | 1812 | sched_ttwu_pending(); |
ca38062e SS |
1813 | |
1814 | /* | |
1815 | * Check if someone kicked us for doing the nohz idle load balance. | |
1816 | */ | |
873b4c65 | 1817 | if (unlikely(got_nohz_idle_kick())) { |
6eb57e0d | 1818 | this_rq()->idle_balance = 1; |
ca38062e | 1819 | raise_softirq_irqoff(SCHED_SOFTIRQ); |
6eb57e0d | 1820 | } |
c5d753a5 | 1821 | irq_exit(); |
317f3941 PZ |
1822 | } |
1823 | ||
b7e7ade3 | 1824 | static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags) |
317f3941 | 1825 | { |
e3baac47 PZ |
1826 | struct rq *rq = cpu_rq(cpu); |
1827 | ||
b7e7ade3 PZ |
1828 | p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED); |
1829 | ||
e3baac47 PZ |
1830 | if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) { |
1831 | if (!set_nr_if_polling(rq->idle)) | |
1832 | smp_send_reschedule(cpu); | |
1833 | else | |
1834 | trace_sched_wake_idle_without_ipi(cpu); | |
1835 | } | |
317f3941 | 1836 | } |
d6aa8f85 | 1837 | |
f6be8af1 CL |
1838 | void wake_up_if_idle(int cpu) |
1839 | { | |
1840 | struct rq *rq = cpu_rq(cpu); | |
1841 | unsigned long flags; | |
1842 | ||
fd7de1e8 AL |
1843 | rcu_read_lock(); |
1844 | ||
1845 | if (!is_idle_task(rcu_dereference(rq->curr))) | |
1846 | goto out; | |
f6be8af1 CL |
1847 | |
1848 | if (set_nr_if_polling(rq->idle)) { | |
1849 | trace_sched_wake_idle_without_ipi(cpu); | |
1850 | } else { | |
1851 | raw_spin_lock_irqsave(&rq->lock, flags); | |
1852 | if (is_idle_task(rq->curr)) | |
1853 | smp_send_reschedule(cpu); | |
1854 | /* Else cpu is not in idle, do nothing here */ | |
1855 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
1856 | } | |
fd7de1e8 AL |
1857 | |
1858 | out: | |
1859 | rcu_read_unlock(); | |
f6be8af1 CL |
1860 | } |
1861 | ||
39be3501 | 1862 | bool cpus_share_cache(int this_cpu, int that_cpu) |
518cd623 PZ |
1863 | { |
1864 | return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); | |
1865 | } | |
d6aa8f85 | 1866 | #endif /* CONFIG_SMP */ |
317f3941 | 1867 | |
b5179ac7 | 1868 | static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags) |
c05fbafb PZ |
1869 | { |
1870 | struct rq *rq = cpu_rq(cpu); | |
e7904a28 | 1871 | struct pin_cookie cookie; |
c05fbafb | 1872 | |
17d9f311 | 1873 | #if defined(CONFIG_SMP) |
39be3501 | 1874 | if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { |
f01114cb | 1875 | sched_clock_cpu(cpu); /* sync clocks x-cpu */ |
b7e7ade3 | 1876 | ttwu_queue_remote(p, cpu, wake_flags); |
317f3941 PZ |
1877 | return; |
1878 | } | |
1879 | #endif | |
1880 | ||
c05fbafb | 1881 | raw_spin_lock(&rq->lock); |
e7904a28 | 1882 | cookie = lockdep_pin_lock(&rq->lock); |
b5179ac7 | 1883 | ttwu_do_activate(rq, p, wake_flags, cookie); |
e7904a28 | 1884 | lockdep_unpin_lock(&rq->lock, cookie); |
c05fbafb | 1885 | raw_spin_unlock(&rq->lock); |
9ed3811a TH |
1886 | } |
1887 | ||
8643cda5 PZ |
1888 | /* |
1889 | * Notes on Program-Order guarantees on SMP systems. | |
1890 | * | |
1891 | * MIGRATION | |
1892 | * | |
1893 | * The basic program-order guarantee on SMP systems is that when a task [t] | |
1894 | * migrates, all its activity on its old cpu [c0] happens-before any subsequent | |
1895 | * execution on its new cpu [c1]. | |
1896 | * | |
1897 | * For migration (of runnable tasks) this is provided by the following means: | |
1898 | * | |
1899 | * A) UNLOCK of the rq(c0)->lock scheduling out task t | |
1900 | * B) migration for t is required to synchronize *both* rq(c0)->lock and | |
1901 | * rq(c1)->lock (if not at the same time, then in that order). | |
1902 | * C) LOCK of the rq(c1)->lock scheduling in task | |
1903 | * | |
1904 | * Transitivity guarantees that B happens after A and C after B. | |
1905 | * Note: we only require RCpc transitivity. | |
1906 | * Note: the cpu doing B need not be c0 or c1 | |
1907 | * | |
1908 | * Example: | |
1909 | * | |
1910 | * CPU0 CPU1 CPU2 | |
1911 | * | |
1912 | * LOCK rq(0)->lock | |
1913 | * sched-out X | |
1914 | * sched-in Y | |
1915 | * UNLOCK rq(0)->lock | |
1916 | * | |
1917 | * LOCK rq(0)->lock // orders against CPU0 | |
1918 | * dequeue X | |
1919 | * UNLOCK rq(0)->lock | |
1920 | * | |
1921 | * LOCK rq(1)->lock | |
1922 | * enqueue X | |
1923 | * UNLOCK rq(1)->lock | |
1924 | * | |
1925 | * LOCK rq(1)->lock // orders against CPU2 | |
1926 | * sched-out Z | |
1927 | * sched-in X | |
1928 | * UNLOCK rq(1)->lock | |
1929 | * | |
1930 | * | |
1931 | * BLOCKING -- aka. SLEEP + WAKEUP | |
1932 | * | |
1933 | * For blocking we (obviously) need to provide the same guarantee as for | |
1934 | * migration. However the means are completely different as there is no lock | |
1935 | * chain to provide order. Instead we do: | |
1936 | * | |
1937 | * 1) smp_store_release(X->on_cpu, 0) | |
1938 | * 2) smp_cond_acquire(!X->on_cpu) | |
1939 | * | |
1940 | * Example: | |
1941 | * | |
1942 | * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule) | |
1943 | * | |
1944 | * LOCK rq(0)->lock LOCK X->pi_lock | |
1945 | * dequeue X | |
1946 | * sched-out X | |
1947 | * smp_store_release(X->on_cpu, 0); | |
1948 | * | |
1949 | * smp_cond_acquire(!X->on_cpu); | |
1950 | * X->state = WAKING | |
1951 | * set_task_cpu(X,2) | |
1952 | * | |
1953 | * LOCK rq(2)->lock | |
1954 | * enqueue X | |
1955 | * X->state = RUNNING | |
1956 | * UNLOCK rq(2)->lock | |
1957 | * | |
1958 | * LOCK rq(2)->lock // orders against CPU1 | |
1959 | * sched-out Z | |
1960 | * sched-in X | |
1961 | * UNLOCK rq(2)->lock | |
1962 | * | |
1963 | * UNLOCK X->pi_lock | |
1964 | * UNLOCK rq(0)->lock | |
1965 | * | |
1966 | * | |
1967 | * However; for wakeups there is a second guarantee we must provide, namely we | |
1968 | * must observe the state that lead to our wakeup. That is, not only must our | |
1969 | * task observe its own prior state, it must also observe the stores prior to | |
1970 | * its wakeup. | |
1971 | * | |
1972 | * This means that any means of doing remote wakeups must order the CPU doing | |
1973 | * the wakeup against the CPU the task is going to end up running on. This, | |
1974 | * however, is already required for the regular Program-Order guarantee above, | |
1975 | * since the waking CPU is the one issueing the ACQUIRE (smp_cond_acquire). | |
1976 | * | |
1977 | */ | |
1978 | ||
9ed3811a | 1979 | /** |
1da177e4 | 1980 | * try_to_wake_up - wake up a thread |
9ed3811a | 1981 | * @p: the thread to be awakened |
1da177e4 | 1982 | * @state: the mask of task states that can be woken |
9ed3811a | 1983 | * @wake_flags: wake modifier flags (WF_*) |
1da177e4 LT |
1984 | * |
1985 | * Put it on the run-queue if it's not already there. The "current" | |
1986 | * thread is always on the run-queue (except when the actual | |
1987 | * re-schedule is in progress), and as such you're allowed to do | |
1988 | * the simpler "current->state = TASK_RUNNING" to mark yourself | |
1989 | * runnable without the overhead of this. | |
1990 | * | |
e69f6186 | 1991 | * Return: %true if @p was woken up, %false if it was already running. |
9ed3811a | 1992 | * or @state didn't match @p's state. |
1da177e4 | 1993 | */ |
e4a52bcb PZ |
1994 | static int |
1995 | try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) | |
1da177e4 | 1996 | { |
1da177e4 | 1997 | unsigned long flags; |
c05fbafb | 1998 | int cpu, success = 0; |
2398f2c6 | 1999 | |
e0acd0a6 ON |
2000 | /* |
2001 | * If we are going to wake up a thread waiting for CONDITION we | |
2002 | * need to ensure that CONDITION=1 done by the caller can not be | |
2003 | * reordered with p->state check below. This pairs with mb() in | |
2004 | * set_current_state() the waiting thread does. | |
2005 | */ | |
2006 | smp_mb__before_spinlock(); | |
013fdb80 | 2007 | raw_spin_lock_irqsave(&p->pi_lock, flags); |
e9c84311 | 2008 | if (!(p->state & state)) |
1da177e4 LT |
2009 | goto out; |
2010 | ||
fbd705a0 PZ |
2011 | trace_sched_waking(p); |
2012 | ||
c05fbafb | 2013 | success = 1; /* we're going to change ->state */ |
1da177e4 | 2014 | cpu = task_cpu(p); |
1da177e4 | 2015 | |
c05fbafb PZ |
2016 | if (p->on_rq && ttwu_remote(p, wake_flags)) |
2017 | goto stat; | |
1da177e4 | 2018 | |
1da177e4 | 2019 | #ifdef CONFIG_SMP |
ecf7d01c PZ |
2020 | /* |
2021 | * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be | |
2022 | * possible to, falsely, observe p->on_cpu == 0. | |
2023 | * | |
2024 | * One must be running (->on_cpu == 1) in order to remove oneself | |
2025 | * from the runqueue. | |
2026 | * | |
2027 | * [S] ->on_cpu = 1; [L] ->on_rq | |
2028 | * UNLOCK rq->lock | |
2029 | * RMB | |
2030 | * LOCK rq->lock | |
2031 | * [S] ->on_rq = 0; [L] ->on_cpu | |
2032 | * | |
2033 | * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock | |
2034 | * from the consecutive calls to schedule(); the first switching to our | |
2035 | * task, the second putting it to sleep. | |
2036 | */ | |
2037 | smp_rmb(); | |
2038 | ||
e9c84311 | 2039 | /* |
c05fbafb PZ |
2040 | * If the owning (remote) cpu is still in the middle of schedule() with |
2041 | * this task as prev, wait until its done referencing the task. | |
b75a2253 PZ |
2042 | * |
2043 | * Pairs with the smp_store_release() in finish_lock_switch(). | |
2044 | * | |
2045 | * This ensures that tasks getting woken will be fully ordered against | |
2046 | * their previous state and preserve Program Order. | |
0970d299 | 2047 | */ |
b3e0b1b6 | 2048 | smp_cond_acquire(!p->on_cpu); |
1da177e4 | 2049 | |
a8e4f2ea | 2050 | p->sched_contributes_to_load = !!task_contributes_to_load(p); |
e9c84311 | 2051 | p->state = TASK_WAKING; |
e7693a36 | 2052 | |
ac66f547 | 2053 | cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); |
f339b9dc PZ |
2054 | if (task_cpu(p) != cpu) { |
2055 | wake_flags |= WF_MIGRATED; | |
e4a52bcb | 2056 | set_task_cpu(p, cpu); |
f339b9dc | 2057 | } |
1da177e4 | 2058 | #endif /* CONFIG_SMP */ |
1da177e4 | 2059 | |
b5179ac7 | 2060 | ttwu_queue(p, cpu, wake_flags); |
c05fbafb | 2061 | stat: |
cb251765 MG |
2062 | if (schedstat_enabled()) |
2063 | ttwu_stat(p, cpu, wake_flags); | |
1da177e4 | 2064 | out: |
013fdb80 | 2065 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); |
1da177e4 LT |
2066 | |
2067 | return success; | |
2068 | } | |
2069 | ||
21aa9af0 TH |
2070 | /** |
2071 | * try_to_wake_up_local - try to wake up a local task with rq lock held | |
2072 | * @p: the thread to be awakened | |
2073 | * | |
2acca55e | 2074 | * Put @p on the run-queue if it's not already there. The caller must |
21aa9af0 | 2075 | * ensure that this_rq() is locked, @p is bound to this_rq() and not |
2acca55e | 2076 | * the current task. |
21aa9af0 | 2077 | */ |
e7904a28 | 2078 | static void try_to_wake_up_local(struct task_struct *p, struct pin_cookie cookie) |
21aa9af0 TH |
2079 | { |
2080 | struct rq *rq = task_rq(p); | |
21aa9af0 | 2081 | |
383efcd0 TH |
2082 | if (WARN_ON_ONCE(rq != this_rq()) || |
2083 | WARN_ON_ONCE(p == current)) | |
2084 | return; | |
2085 | ||
21aa9af0 TH |
2086 | lockdep_assert_held(&rq->lock); |
2087 | ||
2acca55e | 2088 | if (!raw_spin_trylock(&p->pi_lock)) { |
cbce1a68 PZ |
2089 | /* |
2090 | * This is OK, because current is on_cpu, which avoids it being | |
2091 | * picked for load-balance and preemption/IRQs are still | |
2092 | * disabled avoiding further scheduler activity on it and we've | |
2093 | * not yet picked a replacement task. | |
2094 | */ | |
e7904a28 | 2095 | lockdep_unpin_lock(&rq->lock, cookie); |
2acca55e PZ |
2096 | raw_spin_unlock(&rq->lock); |
2097 | raw_spin_lock(&p->pi_lock); | |
2098 | raw_spin_lock(&rq->lock); | |
e7904a28 | 2099 | lockdep_repin_lock(&rq->lock, cookie); |
2acca55e PZ |
2100 | } |
2101 | ||
21aa9af0 | 2102 | if (!(p->state & TASK_NORMAL)) |
2acca55e | 2103 | goto out; |
21aa9af0 | 2104 | |
fbd705a0 PZ |
2105 | trace_sched_waking(p); |
2106 | ||
da0c1e65 | 2107 | if (!task_on_rq_queued(p)) |
d7c01d27 PZ |
2108 | ttwu_activate(rq, p, ENQUEUE_WAKEUP); |
2109 | ||
e7904a28 | 2110 | ttwu_do_wakeup(rq, p, 0, cookie); |
cb251765 MG |
2111 | if (schedstat_enabled()) |
2112 | ttwu_stat(p, smp_processor_id(), 0); | |
2acca55e PZ |
2113 | out: |
2114 | raw_spin_unlock(&p->pi_lock); | |
21aa9af0 TH |
2115 | } |
2116 | ||
50fa610a DH |
2117 | /** |
2118 | * wake_up_process - Wake up a specific process | |
2119 | * @p: The process to be woken up. | |
2120 | * | |
2121 | * Attempt to wake up the nominated process and move it to the set of runnable | |
e69f6186 YB |
2122 | * processes. |
2123 | * | |
2124 | * Return: 1 if the process was woken up, 0 if it was already running. | |
50fa610a DH |
2125 | * |
2126 | * It may be assumed that this function implies a write memory barrier before | |
2127 | * changing the task state if and only if any tasks are woken up. | |
2128 | */ | |
7ad5b3a5 | 2129 | int wake_up_process(struct task_struct *p) |
1da177e4 | 2130 | { |
9067ac85 | 2131 | return try_to_wake_up(p, TASK_NORMAL, 0); |
1da177e4 | 2132 | } |
1da177e4 LT |
2133 | EXPORT_SYMBOL(wake_up_process); |
2134 | ||
7ad5b3a5 | 2135 | int wake_up_state(struct task_struct *p, unsigned int state) |
1da177e4 LT |
2136 | { |
2137 | return try_to_wake_up(p, state, 0); | |
2138 | } | |
2139 | ||
a5e7be3b JL |
2140 | /* |
2141 | * This function clears the sched_dl_entity static params. | |
2142 | */ | |
2143 | void __dl_clear_params(struct task_struct *p) | |
2144 | { | |
2145 | struct sched_dl_entity *dl_se = &p->dl; | |
2146 | ||
2147 | dl_se->dl_runtime = 0; | |
2148 | dl_se->dl_deadline = 0; | |
2149 | dl_se->dl_period = 0; | |
2150 | dl_se->flags = 0; | |
2151 | dl_se->dl_bw = 0; | |
40767b0d PZ |
2152 | |
2153 | dl_se->dl_throttled = 0; | |
40767b0d | 2154 | dl_se->dl_yielded = 0; |
a5e7be3b JL |
2155 | } |
2156 | ||
1da177e4 LT |
2157 | /* |
2158 | * Perform scheduler related setup for a newly forked process p. | |
2159 | * p is forked by current. | |
dd41f596 IM |
2160 | * |
2161 | * __sched_fork() is basic setup used by init_idle() too: | |
2162 | */ | |
5e1576ed | 2163 | static void __sched_fork(unsigned long clone_flags, struct task_struct *p) |
dd41f596 | 2164 | { |
fd2f4419 PZ |
2165 | p->on_rq = 0; |
2166 | ||
2167 | p->se.on_rq = 0; | |
dd41f596 IM |
2168 | p->se.exec_start = 0; |
2169 | p->se.sum_exec_runtime = 0; | |
f6cf891c | 2170 | p->se.prev_sum_exec_runtime = 0; |
6c594c21 | 2171 | p->se.nr_migrations = 0; |
da7a735e | 2172 | p->se.vruntime = 0; |
fd2f4419 | 2173 | INIT_LIST_HEAD(&p->se.group_node); |
6cfb0d5d | 2174 | |
ad936d86 BP |
2175 | #ifdef CONFIG_FAIR_GROUP_SCHED |
2176 | p->se.cfs_rq = NULL; | |
2177 | #endif | |
2178 | ||
6cfb0d5d | 2179 | #ifdef CONFIG_SCHEDSTATS |
cb251765 | 2180 | /* Even if schedstat is disabled, there should not be garbage */ |
41acab88 | 2181 | memset(&p->se.statistics, 0, sizeof(p->se.statistics)); |
6cfb0d5d | 2182 | #endif |
476d139c | 2183 | |
aab03e05 | 2184 | RB_CLEAR_NODE(&p->dl.rb_node); |
40767b0d | 2185 | init_dl_task_timer(&p->dl); |
a5e7be3b | 2186 | __dl_clear_params(p); |
aab03e05 | 2187 | |
fa717060 | 2188 | INIT_LIST_HEAD(&p->rt.run_list); |
ff77e468 PZ |
2189 | p->rt.timeout = 0; |
2190 | p->rt.time_slice = sched_rr_timeslice; | |
2191 | p->rt.on_rq = 0; | |
2192 | p->rt.on_list = 0; | |
476d139c | 2193 | |
e107be36 AK |
2194 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
2195 | INIT_HLIST_HEAD(&p->preempt_notifiers); | |
2196 | #endif | |
cbee9f88 PZ |
2197 | |
2198 | #ifdef CONFIG_NUMA_BALANCING | |
2199 | if (p->mm && atomic_read(&p->mm->mm_users) == 1) { | |
7e8d16b6 | 2200 | p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); |
cbee9f88 PZ |
2201 | p->mm->numa_scan_seq = 0; |
2202 | } | |
2203 | ||
5e1576ed RR |
2204 | if (clone_flags & CLONE_VM) |
2205 | p->numa_preferred_nid = current->numa_preferred_nid; | |
2206 | else | |
2207 | p->numa_preferred_nid = -1; | |
2208 | ||
cbee9f88 PZ |
2209 | p->node_stamp = 0ULL; |
2210 | p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; | |
4b96a29b | 2211 | p->numa_scan_period = sysctl_numa_balancing_scan_delay; |
cbee9f88 | 2212 | p->numa_work.next = &p->numa_work; |
44dba3d5 | 2213 | p->numa_faults = NULL; |
7e2703e6 RR |
2214 | p->last_task_numa_placement = 0; |
2215 | p->last_sum_exec_runtime = 0; | |
8c8a743c | 2216 | |
8c8a743c | 2217 | p->numa_group = NULL; |
cbee9f88 | 2218 | #endif /* CONFIG_NUMA_BALANCING */ |
dd41f596 IM |
2219 | } |
2220 | ||
2a595721 SD |
2221 | DEFINE_STATIC_KEY_FALSE(sched_numa_balancing); |
2222 | ||
1a687c2e | 2223 | #ifdef CONFIG_NUMA_BALANCING |
c3b9bc5b | 2224 | |
1a687c2e MG |
2225 | void set_numabalancing_state(bool enabled) |
2226 | { | |
2227 | if (enabled) | |
2a595721 | 2228 | static_branch_enable(&sched_numa_balancing); |
1a687c2e | 2229 | else |
2a595721 | 2230 | static_branch_disable(&sched_numa_balancing); |
1a687c2e | 2231 | } |
54a43d54 AK |
2232 | |
2233 | #ifdef CONFIG_PROC_SYSCTL | |
2234 | int sysctl_numa_balancing(struct ctl_table *table, int write, | |
2235 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
2236 | { | |
2237 | struct ctl_table t; | |
2238 | int err; | |
2a595721 | 2239 | int state = static_branch_likely(&sched_numa_balancing); |
54a43d54 AK |
2240 | |
2241 | if (write && !capable(CAP_SYS_ADMIN)) | |
2242 | return -EPERM; | |
2243 | ||
2244 | t = *table; | |
2245 | t.data = &state; | |
2246 | err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); | |
2247 | if (err < 0) | |
2248 | return err; | |
2249 | if (write) | |
2250 | set_numabalancing_state(state); | |
2251 | return err; | |
2252 | } | |
2253 | #endif | |
2254 | #endif | |
dd41f596 | 2255 | |
4698f88c JP |
2256 | #ifdef CONFIG_SCHEDSTATS |
2257 | ||
cb251765 | 2258 | DEFINE_STATIC_KEY_FALSE(sched_schedstats); |
4698f88c | 2259 | static bool __initdata __sched_schedstats = false; |
cb251765 | 2260 | |
cb251765 MG |
2261 | static void set_schedstats(bool enabled) |
2262 | { | |
2263 | if (enabled) | |
2264 | static_branch_enable(&sched_schedstats); | |
2265 | else | |
2266 | static_branch_disable(&sched_schedstats); | |
2267 | } | |
2268 | ||
2269 | void force_schedstat_enabled(void) | |
2270 | { | |
2271 | if (!schedstat_enabled()) { | |
2272 | pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n"); | |
2273 | static_branch_enable(&sched_schedstats); | |
2274 | } | |
2275 | } | |
2276 | ||
2277 | static int __init setup_schedstats(char *str) | |
2278 | { | |
2279 | int ret = 0; | |
2280 | if (!str) | |
2281 | goto out; | |
2282 | ||
4698f88c JP |
2283 | /* |
2284 | * This code is called before jump labels have been set up, so we can't | |
2285 | * change the static branch directly just yet. Instead set a temporary | |
2286 | * variable so init_schedstats() can do it later. | |
2287 | */ | |
cb251765 | 2288 | if (!strcmp(str, "enable")) { |
4698f88c | 2289 | __sched_schedstats = true; |
cb251765 MG |
2290 | ret = 1; |
2291 | } else if (!strcmp(str, "disable")) { | |
4698f88c | 2292 | __sched_schedstats = false; |
cb251765 MG |
2293 | ret = 1; |
2294 | } | |
2295 | out: | |
2296 | if (!ret) | |
2297 | pr_warn("Unable to parse schedstats=\n"); | |
2298 | ||
2299 | return ret; | |
2300 | } | |
2301 | __setup("schedstats=", setup_schedstats); | |
2302 | ||
4698f88c JP |
2303 | static void __init init_schedstats(void) |
2304 | { | |
2305 | set_schedstats(__sched_schedstats); | |
2306 | } | |
2307 | ||
cb251765 MG |
2308 | #ifdef CONFIG_PROC_SYSCTL |
2309 | int sysctl_schedstats(struct ctl_table *table, int write, | |
2310 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
2311 | { | |
2312 | struct ctl_table t; | |
2313 | int err; | |
2314 | int state = static_branch_likely(&sched_schedstats); | |
2315 | ||
2316 | if (write && !capable(CAP_SYS_ADMIN)) | |
2317 | return -EPERM; | |
2318 | ||
2319 | t = *table; | |
2320 | t.data = &state; | |
2321 | err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); | |
2322 | if (err < 0) | |
2323 | return err; | |
2324 | if (write) | |
2325 | set_schedstats(state); | |
2326 | return err; | |
2327 | } | |
4698f88c JP |
2328 | #endif /* CONFIG_PROC_SYSCTL */ |
2329 | #else /* !CONFIG_SCHEDSTATS */ | |
2330 | static inline void init_schedstats(void) {} | |
2331 | #endif /* CONFIG_SCHEDSTATS */ | |
dd41f596 IM |
2332 | |
2333 | /* | |
2334 | * fork()/clone()-time setup: | |
2335 | */ | |
aab03e05 | 2336 | int sched_fork(unsigned long clone_flags, struct task_struct *p) |
dd41f596 | 2337 | { |
0122ec5b | 2338 | unsigned long flags; |
dd41f596 IM |
2339 | int cpu = get_cpu(); |
2340 | ||
5e1576ed | 2341 | __sched_fork(clone_flags, p); |
06b83b5f | 2342 | /* |
0017d735 | 2343 | * We mark the process as running here. This guarantees that |
06b83b5f PZ |
2344 | * nobody will actually run it, and a signal or other external |
2345 | * event cannot wake it up and insert it on the runqueue either. | |
2346 | */ | |
0017d735 | 2347 | p->state = TASK_RUNNING; |
dd41f596 | 2348 | |
c350a04e MG |
2349 | /* |
2350 | * Make sure we do not leak PI boosting priority to the child. | |
2351 | */ | |
2352 | p->prio = current->normal_prio; | |
2353 | ||
b9dc29e7 MG |
2354 | /* |
2355 | * Revert to default priority/policy on fork if requested. | |
2356 | */ | |
2357 | if (unlikely(p->sched_reset_on_fork)) { | |
aab03e05 | 2358 | if (task_has_dl_policy(p) || task_has_rt_policy(p)) { |
b9dc29e7 | 2359 | p->policy = SCHED_NORMAL; |
6c697bdf | 2360 | p->static_prio = NICE_TO_PRIO(0); |
c350a04e MG |
2361 | p->rt_priority = 0; |
2362 | } else if (PRIO_TO_NICE(p->static_prio) < 0) | |
2363 | p->static_prio = NICE_TO_PRIO(0); | |
2364 | ||
2365 | p->prio = p->normal_prio = __normal_prio(p); | |
2366 | set_load_weight(p); | |
6c697bdf | 2367 | |
b9dc29e7 MG |
2368 | /* |
2369 | * We don't need the reset flag anymore after the fork. It has | |
2370 | * fulfilled its duty: | |
2371 | */ | |
2372 | p->sched_reset_on_fork = 0; | |
2373 | } | |
ca94c442 | 2374 | |
aab03e05 DF |
2375 | if (dl_prio(p->prio)) { |
2376 | put_cpu(); | |
2377 | return -EAGAIN; | |
2378 | } else if (rt_prio(p->prio)) { | |
2379 | p->sched_class = &rt_sched_class; | |
2380 | } else { | |
2ddbf952 | 2381 | p->sched_class = &fair_sched_class; |
aab03e05 | 2382 | } |
b29739f9 | 2383 | |
cd29fe6f PZ |
2384 | if (p->sched_class->task_fork) |
2385 | p->sched_class->task_fork(p); | |
2386 | ||
86951599 PZ |
2387 | /* |
2388 | * The child is not yet in the pid-hash so no cgroup attach races, | |
2389 | * and the cgroup is pinned to this child due to cgroup_fork() | |
2390 | * is ran before sched_fork(). | |
2391 | * | |
2392 | * Silence PROVE_RCU. | |
2393 | */ | |
0122ec5b | 2394 | raw_spin_lock_irqsave(&p->pi_lock, flags); |
5f3edc1b | 2395 | set_task_cpu(p, cpu); |
0122ec5b | 2396 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); |
5f3edc1b | 2397 | |
f6db8347 | 2398 | #ifdef CONFIG_SCHED_INFO |
dd41f596 | 2399 | if (likely(sched_info_on())) |
52f17b6c | 2400 | memset(&p->sched_info, 0, sizeof(p->sched_info)); |
1da177e4 | 2401 | #endif |
3ca7a440 PZ |
2402 | #if defined(CONFIG_SMP) |
2403 | p->on_cpu = 0; | |
4866cde0 | 2404 | #endif |
01028747 | 2405 | init_task_preempt_count(p); |
806c09a7 | 2406 | #ifdef CONFIG_SMP |
917b627d | 2407 | plist_node_init(&p->pushable_tasks, MAX_PRIO); |
1baca4ce | 2408 | RB_CLEAR_NODE(&p->pushable_dl_tasks); |
806c09a7 | 2409 | #endif |
917b627d | 2410 | |
476d139c | 2411 | put_cpu(); |
aab03e05 | 2412 | return 0; |
1da177e4 LT |
2413 | } |
2414 | ||
332ac17e DF |
2415 | unsigned long to_ratio(u64 period, u64 runtime) |
2416 | { | |
2417 | if (runtime == RUNTIME_INF) | |
2418 | return 1ULL << 20; | |
2419 | ||
2420 | /* | |
2421 | * Doing this here saves a lot of checks in all | |
2422 | * the calling paths, and returning zero seems | |
2423 | * safe for them anyway. | |
2424 | */ | |
2425 | if (period == 0) | |
2426 | return 0; | |
2427 | ||
2428 | return div64_u64(runtime << 20, period); | |
2429 | } | |
2430 | ||
2431 | #ifdef CONFIG_SMP | |
2432 | inline struct dl_bw *dl_bw_of(int i) | |
2433 | { | |
f78f5b90 PM |
2434 | RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(), |
2435 | "sched RCU must be held"); | |
332ac17e DF |
2436 | return &cpu_rq(i)->rd->dl_bw; |
2437 | } | |
2438 | ||
de212f18 | 2439 | static inline int dl_bw_cpus(int i) |
332ac17e | 2440 | { |
de212f18 PZ |
2441 | struct root_domain *rd = cpu_rq(i)->rd; |
2442 | int cpus = 0; | |
2443 | ||
f78f5b90 PM |
2444 | RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(), |
2445 | "sched RCU must be held"); | |
de212f18 PZ |
2446 | for_each_cpu_and(i, rd->span, cpu_active_mask) |
2447 | cpus++; | |
2448 | ||
2449 | return cpus; | |
332ac17e DF |
2450 | } |
2451 | #else | |
2452 | inline struct dl_bw *dl_bw_of(int i) | |
2453 | { | |
2454 | return &cpu_rq(i)->dl.dl_bw; | |
2455 | } | |
2456 | ||
de212f18 | 2457 | static inline int dl_bw_cpus(int i) |
332ac17e DF |
2458 | { |
2459 | return 1; | |
2460 | } | |
2461 | #endif | |
2462 | ||
332ac17e DF |
2463 | /* |
2464 | * We must be sure that accepting a new task (or allowing changing the | |
2465 | * parameters of an existing one) is consistent with the bandwidth | |
2466 | * constraints. If yes, this function also accordingly updates the currently | |
2467 | * allocated bandwidth to reflect the new situation. | |
2468 | * | |
2469 | * This function is called while holding p's rq->lock. | |
40767b0d PZ |
2470 | * |
2471 | * XXX we should delay bw change until the task's 0-lag point, see | |
2472 | * __setparam_dl(). | |
332ac17e DF |
2473 | */ |
2474 | static int dl_overflow(struct task_struct *p, int policy, | |
2475 | const struct sched_attr *attr) | |
2476 | { | |
2477 | ||
2478 | struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); | |
4df1638c | 2479 | u64 period = attr->sched_period ?: attr->sched_deadline; |
332ac17e DF |
2480 | u64 runtime = attr->sched_runtime; |
2481 | u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; | |
de212f18 | 2482 | int cpus, err = -1; |
332ac17e | 2483 | |
fec148c0 XP |
2484 | /* !deadline task may carry old deadline bandwidth */ |
2485 | if (new_bw == p->dl.dl_bw && task_has_dl_policy(p)) | |
332ac17e DF |
2486 | return 0; |
2487 | ||
2488 | /* | |
2489 | * Either if a task, enters, leave, or stays -deadline but changes | |
2490 | * its parameters, we may need to update accordingly the total | |
2491 | * allocated bandwidth of the container. | |
2492 | */ | |
2493 | raw_spin_lock(&dl_b->lock); | |
de212f18 | 2494 | cpus = dl_bw_cpus(task_cpu(p)); |
332ac17e DF |
2495 | if (dl_policy(policy) && !task_has_dl_policy(p) && |
2496 | !__dl_overflow(dl_b, cpus, 0, new_bw)) { | |
2497 | __dl_add(dl_b, new_bw); | |
2498 | err = 0; | |
2499 | } else if (dl_policy(policy) && task_has_dl_policy(p) && | |
2500 | !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { | |
2501 | __dl_clear(dl_b, p->dl.dl_bw); | |
2502 | __dl_add(dl_b, new_bw); | |
2503 | err = 0; | |
2504 | } else if (!dl_policy(policy) && task_has_dl_policy(p)) { | |
2505 | __dl_clear(dl_b, p->dl.dl_bw); | |
2506 | err = 0; | |
2507 | } | |
2508 | raw_spin_unlock(&dl_b->lock); | |
2509 | ||
2510 | return err; | |
2511 | } | |
2512 | ||
2513 | extern void init_dl_bw(struct dl_bw *dl_b); | |
2514 | ||
1da177e4 LT |
2515 | /* |
2516 | * wake_up_new_task - wake up a newly created task for the first time. | |
2517 | * | |
2518 | * This function will do some initial scheduler statistics housekeeping | |
2519 | * that must be done for every newly created context, then puts the task | |
2520 | * on the runqueue and wakes it. | |
2521 | */ | |
3e51e3ed | 2522 | void wake_up_new_task(struct task_struct *p) |
1da177e4 | 2523 | { |
eb580751 | 2524 | struct rq_flags rf; |
dd41f596 | 2525 | struct rq *rq; |
fabf318e | 2526 | |
98d8fd81 MR |
2527 | /* Initialize new task's runnable average */ |
2528 | init_entity_runnable_average(&p->se); | |
eb580751 | 2529 | raw_spin_lock_irqsave(&p->pi_lock, rf.flags); |
fabf318e PZ |
2530 | #ifdef CONFIG_SMP |
2531 | /* | |
2532 | * Fork balancing, do it here and not earlier because: | |
2533 | * - cpus_allowed can change in the fork path | |
2534 | * - any previously selected cpu might disappear through hotplug | |
fabf318e | 2535 | */ |
ac66f547 | 2536 | set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); |
0017d735 | 2537 | #endif |
b7fa30c9 | 2538 | rq = __task_rq_lock(p, &rf); |
2b8c41da | 2539 | post_init_entity_util_avg(&p->se); |
0017d735 | 2540 | |
cd29fe6f | 2541 | activate_task(rq, p, 0); |
da0c1e65 | 2542 | p->on_rq = TASK_ON_RQ_QUEUED; |
fbd705a0 | 2543 | trace_sched_wakeup_new(p); |
a7558e01 | 2544 | check_preempt_curr(rq, p, WF_FORK); |
9a897c5a | 2545 | #ifdef CONFIG_SMP |
0aaafaab PZ |
2546 | if (p->sched_class->task_woken) { |
2547 | /* | |
2548 | * Nothing relies on rq->lock after this, so its fine to | |
2549 | * drop it. | |
2550 | */ | |
e7904a28 | 2551 | lockdep_unpin_lock(&rq->lock, rf.cookie); |
efbbd05a | 2552 | p->sched_class->task_woken(rq, p); |
e7904a28 | 2553 | lockdep_repin_lock(&rq->lock, rf.cookie); |
0aaafaab | 2554 | } |
9a897c5a | 2555 | #endif |
eb580751 | 2556 | task_rq_unlock(rq, p, &rf); |
1da177e4 LT |
2557 | } |
2558 | ||
e107be36 AK |
2559 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
2560 | ||
1cde2930 PZ |
2561 | static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE; |
2562 | ||
2ecd9d29 PZ |
2563 | void preempt_notifier_inc(void) |
2564 | { | |
2565 | static_key_slow_inc(&preempt_notifier_key); | |
2566 | } | |
2567 | EXPORT_SYMBOL_GPL(preempt_notifier_inc); | |
2568 | ||
2569 | void preempt_notifier_dec(void) | |
2570 | { | |
2571 | static_key_slow_dec(&preempt_notifier_key); | |
2572 | } | |
2573 | EXPORT_SYMBOL_GPL(preempt_notifier_dec); | |
2574 | ||
e107be36 | 2575 | /** |
80dd99b3 | 2576 | * preempt_notifier_register - tell me when current is being preempted & rescheduled |
421cee29 | 2577 | * @notifier: notifier struct to register |
e107be36 AK |
2578 | */ |
2579 | void preempt_notifier_register(struct preempt_notifier *notifier) | |
2580 | { | |
2ecd9d29 PZ |
2581 | if (!static_key_false(&preempt_notifier_key)) |
2582 | WARN(1, "registering preempt_notifier while notifiers disabled\n"); | |
2583 | ||
e107be36 AK |
2584 | hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); |
2585 | } | |
2586 | EXPORT_SYMBOL_GPL(preempt_notifier_register); | |
2587 | ||
2588 | /** | |
2589 | * preempt_notifier_unregister - no longer interested in preemption notifications | |
421cee29 | 2590 | * @notifier: notifier struct to unregister |
e107be36 | 2591 | * |
d84525a8 | 2592 | * This is *not* safe to call from within a preemption notifier. |
e107be36 AK |
2593 | */ |
2594 | void preempt_notifier_unregister(struct preempt_notifier *notifier) | |
2595 | { | |
2596 | hlist_del(¬ifier->link); | |
2597 | } | |
2598 | EXPORT_SYMBOL_GPL(preempt_notifier_unregister); | |
2599 | ||
1cde2930 | 2600 | static void __fire_sched_in_preempt_notifiers(struct task_struct *curr) |
e107be36 AK |
2601 | { |
2602 | struct preempt_notifier *notifier; | |
e107be36 | 2603 | |
b67bfe0d | 2604 | hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) |
e107be36 AK |
2605 | notifier->ops->sched_in(notifier, raw_smp_processor_id()); |
2606 | } | |
2607 | ||
1cde2930 PZ |
2608 | static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr) |
2609 | { | |
2610 | if (static_key_false(&preempt_notifier_key)) | |
2611 | __fire_sched_in_preempt_notifiers(curr); | |
2612 | } | |
2613 | ||
e107be36 | 2614 | static void |
1cde2930 PZ |
2615 | __fire_sched_out_preempt_notifiers(struct task_struct *curr, |
2616 | struct task_struct *next) | |
e107be36 AK |
2617 | { |
2618 | struct preempt_notifier *notifier; | |
e107be36 | 2619 | |
b67bfe0d | 2620 | hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) |
e107be36 AK |
2621 | notifier->ops->sched_out(notifier, next); |
2622 | } | |
2623 | ||
1cde2930 PZ |
2624 | static __always_inline void |
2625 | fire_sched_out_preempt_notifiers(struct task_struct *curr, | |
2626 | struct task_struct *next) | |
2627 | { | |
2628 | if (static_key_false(&preempt_notifier_key)) | |
2629 | __fire_sched_out_preempt_notifiers(curr, next); | |
2630 | } | |
2631 | ||
6d6bc0ad | 2632 | #else /* !CONFIG_PREEMPT_NOTIFIERS */ |
e107be36 | 2633 | |
1cde2930 | 2634 | static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr) |
e107be36 AK |
2635 | { |
2636 | } | |
2637 | ||
1cde2930 | 2638 | static inline void |
e107be36 AK |
2639 | fire_sched_out_preempt_notifiers(struct task_struct *curr, |
2640 | struct task_struct *next) | |
2641 | { | |
2642 | } | |
2643 | ||
6d6bc0ad | 2644 | #endif /* CONFIG_PREEMPT_NOTIFIERS */ |
e107be36 | 2645 | |
4866cde0 NP |
2646 | /** |
2647 | * prepare_task_switch - prepare to switch tasks | |
2648 | * @rq: the runqueue preparing to switch | |
421cee29 | 2649 | * @prev: the current task that is being switched out |
4866cde0 NP |
2650 | * @next: the task we are going to switch to. |
2651 | * | |
2652 | * This is called with the rq lock held and interrupts off. It must | |
2653 | * be paired with a subsequent finish_task_switch after the context | |
2654 | * switch. | |
2655 | * | |
2656 | * prepare_task_switch sets up locking and calls architecture specific | |
2657 | * hooks. | |
2658 | */ | |
e107be36 AK |
2659 | static inline void |
2660 | prepare_task_switch(struct rq *rq, struct task_struct *prev, | |
2661 | struct task_struct *next) | |
4866cde0 | 2662 | { |
43148951 | 2663 | sched_info_switch(rq, prev, next); |
fe4b04fa | 2664 | perf_event_task_sched_out(prev, next); |
e107be36 | 2665 | fire_sched_out_preempt_notifiers(prev, next); |
4866cde0 NP |
2666 | prepare_lock_switch(rq, next); |
2667 | prepare_arch_switch(next); | |
2668 | } | |
2669 | ||
1da177e4 LT |
2670 | /** |
2671 | * finish_task_switch - clean up after a task-switch | |
2672 | * @prev: the thread we just switched away from. | |
2673 | * | |
4866cde0 NP |
2674 | * finish_task_switch must be called after the context switch, paired |
2675 | * with a prepare_task_switch call before the context switch. | |
2676 | * finish_task_switch will reconcile locking set up by prepare_task_switch, | |
2677 | * and do any other architecture-specific cleanup actions. | |
1da177e4 LT |
2678 | * |
2679 | * Note that we may have delayed dropping an mm in context_switch(). If | |
41a2d6cf | 2680 | * so, we finish that here outside of the runqueue lock. (Doing it |
1da177e4 LT |
2681 | * with the lock held can cause deadlocks; see schedule() for |
2682 | * details.) | |
dfa50b60 ON |
2683 | * |
2684 | * The context switch have flipped the stack from under us and restored the | |
2685 | * local variables which were saved when this task called schedule() in the | |
2686 | * past. prev == current is still correct but we need to recalculate this_rq | |
2687 | * because prev may have moved to another CPU. | |
1da177e4 | 2688 | */ |
dfa50b60 | 2689 | static struct rq *finish_task_switch(struct task_struct *prev) |
1da177e4 LT |
2690 | __releases(rq->lock) |
2691 | { | |
dfa50b60 | 2692 | struct rq *rq = this_rq(); |
1da177e4 | 2693 | struct mm_struct *mm = rq->prev_mm; |
55a101f8 | 2694 | long prev_state; |
1da177e4 | 2695 | |
609ca066 PZ |
2696 | /* |
2697 | * The previous task will have left us with a preempt_count of 2 | |
2698 | * because it left us after: | |
2699 | * | |
2700 | * schedule() | |
2701 | * preempt_disable(); // 1 | |
2702 | * __schedule() | |
2703 | * raw_spin_lock_irq(&rq->lock) // 2 | |
2704 | * | |
2705 | * Also, see FORK_PREEMPT_COUNT. | |
2706 | */ | |
e2bf1c4b PZ |
2707 | if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET, |
2708 | "corrupted preempt_count: %s/%d/0x%x\n", | |
2709 | current->comm, current->pid, preempt_count())) | |
2710 | preempt_count_set(FORK_PREEMPT_COUNT); | |
609ca066 | 2711 | |
1da177e4 LT |
2712 | rq->prev_mm = NULL; |
2713 | ||
2714 | /* | |
2715 | * A task struct has one reference for the use as "current". | |
c394cc9f | 2716 | * If a task dies, then it sets TASK_DEAD in tsk->state and calls |
55a101f8 ON |
2717 | * schedule one last time. The schedule call will never return, and |
2718 | * the scheduled task must drop that reference. | |
95913d97 PZ |
2719 | * |
2720 | * We must observe prev->state before clearing prev->on_cpu (in | |
2721 | * finish_lock_switch), otherwise a concurrent wakeup can get prev | |
2722 | * running on another CPU and we could rave with its RUNNING -> DEAD | |
2723 | * transition, resulting in a double drop. | |
1da177e4 | 2724 | */ |
55a101f8 | 2725 | prev_state = prev->state; |
bf9fae9f | 2726 | vtime_task_switch(prev); |
a8d757ef | 2727 | perf_event_task_sched_in(prev, current); |
4866cde0 | 2728 | finish_lock_switch(rq, prev); |
01f23e16 | 2729 | finish_arch_post_lock_switch(); |
e8fa1362 | 2730 | |
e107be36 | 2731 | fire_sched_in_preempt_notifiers(current); |
1da177e4 LT |
2732 | if (mm) |
2733 | mmdrop(mm); | |
c394cc9f | 2734 | if (unlikely(prev_state == TASK_DEAD)) { |
e6c390f2 DF |
2735 | if (prev->sched_class->task_dead) |
2736 | prev->sched_class->task_dead(prev); | |
2737 | ||
c6fd91f0 | 2738 | /* |
2739 | * Remove function-return probe instances associated with this | |
2740 | * task and put them back on the free list. | |
9761eea8 | 2741 | */ |
c6fd91f0 | 2742 | kprobe_flush_task(prev); |
1da177e4 | 2743 | put_task_struct(prev); |
c6fd91f0 | 2744 | } |
99e5ada9 | 2745 | |
de734f89 | 2746 | tick_nohz_task_switch(); |
dfa50b60 | 2747 | return rq; |
1da177e4 LT |
2748 | } |
2749 | ||
3f029d3c GH |
2750 | #ifdef CONFIG_SMP |
2751 | ||
3f029d3c | 2752 | /* rq->lock is NOT held, but preemption is disabled */ |
e3fca9e7 | 2753 | static void __balance_callback(struct rq *rq) |
3f029d3c | 2754 | { |
e3fca9e7 PZ |
2755 | struct callback_head *head, *next; |
2756 | void (*func)(struct rq *rq); | |
2757 | unsigned long flags; | |
3f029d3c | 2758 | |
e3fca9e7 PZ |
2759 | raw_spin_lock_irqsave(&rq->lock, flags); |
2760 | head = rq->balance_callback; | |
2761 | rq->balance_callback = NULL; | |
2762 | while (head) { | |
2763 | func = (void (*)(struct rq *))head->func; | |
2764 | next = head->next; | |
2765 | head->next = NULL; | |
2766 | head = next; | |
3f029d3c | 2767 | |
e3fca9e7 | 2768 | func(rq); |
3f029d3c | 2769 | } |
e3fca9e7 PZ |
2770 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
2771 | } | |
2772 | ||
2773 | static inline void balance_callback(struct rq *rq) | |
2774 | { | |
2775 | if (unlikely(rq->balance_callback)) | |
2776 | __balance_callback(rq); | |
3f029d3c GH |
2777 | } |
2778 | ||
2779 | #else | |
da19ab51 | 2780 | |
e3fca9e7 | 2781 | static inline void balance_callback(struct rq *rq) |
3f029d3c | 2782 | { |
1da177e4 LT |
2783 | } |
2784 | ||
3f029d3c GH |
2785 | #endif |
2786 | ||
1da177e4 LT |
2787 | /** |
2788 | * schedule_tail - first thing a freshly forked thread must call. | |
2789 | * @prev: the thread we just switched away from. | |
2790 | */ | |
722a9f92 | 2791 | asmlinkage __visible void schedule_tail(struct task_struct *prev) |
1da177e4 LT |
2792 | __releases(rq->lock) |
2793 | { | |
1a43a14a | 2794 | struct rq *rq; |
da19ab51 | 2795 | |
609ca066 PZ |
2796 | /* |
2797 | * New tasks start with FORK_PREEMPT_COUNT, see there and | |
2798 | * finish_task_switch() for details. | |
2799 | * | |
2800 | * finish_task_switch() will drop rq->lock() and lower preempt_count | |
2801 | * and the preempt_enable() will end up enabling preemption (on | |
2802 | * PREEMPT_COUNT kernels). | |
2803 | */ | |
2804 | ||
dfa50b60 | 2805 | rq = finish_task_switch(prev); |
e3fca9e7 | 2806 | balance_callback(rq); |
1a43a14a | 2807 | preempt_enable(); |
70b97a7f | 2808 | |
1da177e4 | 2809 | if (current->set_child_tid) |
b488893a | 2810 | put_user(task_pid_vnr(current), current->set_child_tid); |
1da177e4 LT |
2811 | } |
2812 | ||
2813 | /* | |
dfa50b60 | 2814 | * context_switch - switch to the new MM and the new thread's register state. |
1da177e4 | 2815 | */ |
04936948 | 2816 | static __always_inline struct rq * |
70b97a7f | 2817 | context_switch(struct rq *rq, struct task_struct *prev, |
e7904a28 | 2818 | struct task_struct *next, struct pin_cookie cookie) |
1da177e4 | 2819 | { |
dd41f596 | 2820 | struct mm_struct *mm, *oldmm; |
1da177e4 | 2821 | |
e107be36 | 2822 | prepare_task_switch(rq, prev, next); |
fe4b04fa | 2823 | |
dd41f596 IM |
2824 | mm = next->mm; |
2825 | oldmm = prev->active_mm; | |
9226d125 ZA |
2826 | /* |
2827 | * For paravirt, this is coupled with an exit in switch_to to | |
2828 | * combine the page table reload and the switch backend into | |
2829 | * one hypercall. | |
2830 | */ | |
224101ed | 2831 | arch_start_context_switch(prev); |
9226d125 | 2832 | |
31915ab4 | 2833 | if (!mm) { |
1da177e4 LT |
2834 | next->active_mm = oldmm; |
2835 | atomic_inc(&oldmm->mm_count); | |
2836 | enter_lazy_tlb(oldmm, next); | |
2837 | } else | |
f98db601 | 2838 | switch_mm_irqs_off(oldmm, mm, next); |
1da177e4 | 2839 | |
31915ab4 | 2840 | if (!prev->mm) { |
1da177e4 | 2841 | prev->active_mm = NULL; |
1da177e4 LT |
2842 | rq->prev_mm = oldmm; |
2843 | } | |
3a5f5e48 IM |
2844 | /* |
2845 | * Since the runqueue lock will be released by the next | |
2846 | * task (which is an invalid locking op but in the case | |
2847 | * of the scheduler it's an obvious special-case), so we | |
2848 | * do an early lockdep release here: | |
2849 | */ | |
e7904a28 | 2850 | lockdep_unpin_lock(&rq->lock, cookie); |
8a25d5de | 2851 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); |
1da177e4 LT |
2852 | |
2853 | /* Here we just switch the register state and the stack. */ | |
2854 | switch_to(prev, next, prev); | |
dd41f596 | 2855 | barrier(); |
dfa50b60 ON |
2856 | |
2857 | return finish_task_switch(prev); | |
1da177e4 LT |
2858 | } |
2859 | ||
2860 | /* | |
1c3e8264 | 2861 | * nr_running and nr_context_switches: |
1da177e4 LT |
2862 | * |
2863 | * externally visible scheduler statistics: current number of runnable | |
1c3e8264 | 2864 | * threads, total number of context switches performed since bootup. |
1da177e4 LT |
2865 | */ |
2866 | unsigned long nr_running(void) | |
2867 | { | |
2868 | unsigned long i, sum = 0; | |
2869 | ||
2870 | for_each_online_cpu(i) | |
2871 | sum += cpu_rq(i)->nr_running; | |
2872 | ||
2873 | return sum; | |
f711f609 | 2874 | } |
1da177e4 | 2875 | |
2ee507c4 TC |
2876 | /* |
2877 | * Check if only the current task is running on the cpu. | |
00cc1633 DD |
2878 | * |
2879 | * Caution: this function does not check that the caller has disabled | |
2880 | * preemption, thus the result might have a time-of-check-to-time-of-use | |
2881 | * race. The caller is responsible to use it correctly, for example: | |
2882 | * | |
2883 | * - from a non-preemptable section (of course) | |
2884 | * | |
2885 | * - from a thread that is bound to a single CPU | |
2886 | * | |
2887 | * - in a loop with very short iterations (e.g. a polling loop) | |
2ee507c4 TC |
2888 | */ |
2889 | bool single_task_running(void) | |
2890 | { | |
00cc1633 | 2891 | return raw_rq()->nr_running == 1; |
2ee507c4 TC |
2892 | } |
2893 | EXPORT_SYMBOL(single_task_running); | |
2894 | ||
1da177e4 | 2895 | unsigned long long nr_context_switches(void) |
46cb4b7c | 2896 | { |
cc94abfc SR |
2897 | int i; |
2898 | unsigned long long sum = 0; | |
46cb4b7c | 2899 | |
0a945022 | 2900 | for_each_possible_cpu(i) |
1da177e4 | 2901 | sum += cpu_rq(i)->nr_switches; |
46cb4b7c | 2902 | |
1da177e4 LT |
2903 | return sum; |
2904 | } | |
483b4ee6 | 2905 | |
1da177e4 LT |
2906 | unsigned long nr_iowait(void) |
2907 | { | |
2908 | unsigned long i, sum = 0; | |
483b4ee6 | 2909 | |
0a945022 | 2910 | for_each_possible_cpu(i) |
1da177e4 | 2911 | sum += atomic_read(&cpu_rq(i)->nr_iowait); |
46cb4b7c | 2912 | |
1da177e4 LT |
2913 | return sum; |
2914 | } | |
483b4ee6 | 2915 | |
8c215bd3 | 2916 | unsigned long nr_iowait_cpu(int cpu) |
69d25870 | 2917 | { |
8c215bd3 | 2918 | struct rq *this = cpu_rq(cpu); |
69d25870 AV |
2919 | return atomic_read(&this->nr_iowait); |
2920 | } | |
46cb4b7c | 2921 | |
372ba8cb MG |
2922 | void get_iowait_load(unsigned long *nr_waiters, unsigned long *load) |
2923 | { | |
3289bdb4 PZ |
2924 | struct rq *rq = this_rq(); |
2925 | *nr_waiters = atomic_read(&rq->nr_iowait); | |
2926 | *load = rq->load.weight; | |
372ba8cb MG |
2927 | } |
2928 | ||
dd41f596 | 2929 | #ifdef CONFIG_SMP |
8a0be9ef | 2930 | |
46cb4b7c | 2931 | /* |
38022906 PZ |
2932 | * sched_exec - execve() is a valuable balancing opportunity, because at |
2933 | * this point the task has the smallest effective memory and cache footprint. | |
46cb4b7c | 2934 | */ |
38022906 | 2935 | void sched_exec(void) |
46cb4b7c | 2936 | { |
38022906 | 2937 | struct task_struct *p = current; |
1da177e4 | 2938 | unsigned long flags; |
0017d735 | 2939 | int dest_cpu; |
46cb4b7c | 2940 | |
8f42ced9 | 2941 | raw_spin_lock_irqsave(&p->pi_lock, flags); |
ac66f547 | 2942 | dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); |
0017d735 PZ |
2943 | if (dest_cpu == smp_processor_id()) |
2944 | goto unlock; | |
38022906 | 2945 | |
8f42ced9 | 2946 | if (likely(cpu_active(dest_cpu))) { |
969c7921 | 2947 | struct migration_arg arg = { p, dest_cpu }; |
46cb4b7c | 2948 | |
8f42ced9 PZ |
2949 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); |
2950 | stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); | |
1da177e4 LT |
2951 | return; |
2952 | } | |
0017d735 | 2953 | unlock: |
8f42ced9 | 2954 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); |
1da177e4 | 2955 | } |
dd41f596 | 2956 | |
1da177e4 LT |
2957 | #endif |
2958 | ||
1da177e4 | 2959 | DEFINE_PER_CPU(struct kernel_stat, kstat); |
3292beb3 | 2960 | DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); |
1da177e4 LT |
2961 | |
2962 | EXPORT_PER_CPU_SYMBOL(kstat); | |
3292beb3 | 2963 | EXPORT_PER_CPU_SYMBOL(kernel_cpustat); |
1da177e4 | 2964 | |
c5f8d995 HS |
2965 | /* |
2966 | * Return accounted runtime for the task. | |
2967 | * In case the task is currently running, return the runtime plus current's | |
2968 | * pending runtime that have not been accounted yet. | |
2969 | */ | |
2970 | unsigned long long task_sched_runtime(struct task_struct *p) | |
2971 | { | |
eb580751 | 2972 | struct rq_flags rf; |
c5f8d995 | 2973 | struct rq *rq; |
6e998916 | 2974 | u64 ns; |
c5f8d995 | 2975 | |
911b2898 PZ |
2976 | #if defined(CONFIG_64BIT) && defined(CONFIG_SMP) |
2977 | /* | |
2978 | * 64-bit doesn't need locks to atomically read a 64bit value. | |
2979 | * So we have a optimization chance when the task's delta_exec is 0. | |
2980 | * Reading ->on_cpu is racy, but this is ok. | |
2981 | * | |
2982 | * If we race with it leaving cpu, we'll take a lock. So we're correct. | |
2983 | * If we race with it entering cpu, unaccounted time is 0. This is | |
2984 | * indistinguishable from the read occurring a few cycles earlier. | |
4036ac15 MG |
2985 | * If we see ->on_cpu without ->on_rq, the task is leaving, and has |
2986 | * been accounted, so we're correct here as well. | |
911b2898 | 2987 | */ |
da0c1e65 | 2988 | if (!p->on_cpu || !task_on_rq_queued(p)) |
911b2898 PZ |
2989 | return p->se.sum_exec_runtime; |
2990 | #endif | |
2991 | ||
eb580751 | 2992 | rq = task_rq_lock(p, &rf); |
6e998916 SG |
2993 | /* |
2994 | * Must be ->curr _and_ ->on_rq. If dequeued, we would | |
2995 | * project cycles that may never be accounted to this | |
2996 | * thread, breaking clock_gettime(). | |
2997 | */ | |
2998 | if (task_current(rq, p) && task_on_rq_queued(p)) { | |
2999 | update_rq_clock(rq); | |
3000 | p->sched_class->update_curr(rq); | |
3001 | } | |
3002 | ns = p->se.sum_exec_runtime; | |
eb580751 | 3003 | task_rq_unlock(rq, p, &rf); |
c5f8d995 HS |
3004 | |
3005 | return ns; | |
3006 | } | |
48f24c4d | 3007 | |
7835b98b CL |
3008 | /* |
3009 | * This function gets called by the timer code, with HZ frequency. | |
3010 | * We call it with interrupts disabled. | |
7835b98b CL |
3011 | */ |
3012 | void scheduler_tick(void) | |
3013 | { | |
7835b98b CL |
3014 | int cpu = smp_processor_id(); |
3015 | struct rq *rq = cpu_rq(cpu); | |
dd41f596 | 3016 | struct task_struct *curr = rq->curr; |
3e51f33f PZ |
3017 | |
3018 | sched_clock_tick(); | |
dd41f596 | 3019 | |
05fa785c | 3020 | raw_spin_lock(&rq->lock); |
3e51f33f | 3021 | update_rq_clock(rq); |
fa85ae24 | 3022 | curr->sched_class->task_tick(rq, curr, 0); |
cee1afce | 3023 | cpu_load_update_active(rq); |
3289bdb4 | 3024 | calc_global_load_tick(rq); |
05fa785c | 3025 | raw_spin_unlock(&rq->lock); |
7835b98b | 3026 | |
e9d2b064 | 3027 | perf_event_task_tick(); |
e220d2dc | 3028 | |
e418e1c2 | 3029 | #ifdef CONFIG_SMP |
6eb57e0d | 3030 | rq->idle_balance = idle_cpu(cpu); |
7caff66f | 3031 | trigger_load_balance(rq); |
e418e1c2 | 3032 | #endif |
265f22a9 | 3033 | rq_last_tick_reset(rq); |
1da177e4 LT |
3034 | } |
3035 | ||
265f22a9 FW |
3036 | #ifdef CONFIG_NO_HZ_FULL |
3037 | /** | |
3038 | * scheduler_tick_max_deferment | |
3039 | * | |
3040 | * Keep at least one tick per second when a single | |
3041 | * active task is running because the scheduler doesn't | |
3042 | * yet completely support full dynticks environment. | |
3043 | * | |
3044 | * This makes sure that uptime, CFS vruntime, load | |
3045 | * balancing, etc... continue to move forward, even | |
3046 | * with a very low granularity. | |
e69f6186 YB |
3047 | * |
3048 | * Return: Maximum deferment in nanoseconds. | |
265f22a9 FW |
3049 | */ |
3050 | u64 scheduler_tick_max_deferment(void) | |
3051 | { | |
3052 | struct rq *rq = this_rq(); | |
316c1608 | 3053 | unsigned long next, now = READ_ONCE(jiffies); |
265f22a9 FW |
3054 | |
3055 | next = rq->last_sched_tick + HZ; | |
3056 | ||
3057 | if (time_before_eq(next, now)) | |
3058 | return 0; | |
3059 | ||
8fe8ff09 | 3060 | return jiffies_to_nsecs(next - now); |
1da177e4 | 3061 | } |
265f22a9 | 3062 | #endif |
1da177e4 | 3063 | |
7e49fcce SR |
3064 | #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ |
3065 | defined(CONFIG_PREEMPT_TRACER)) | |
47252cfb SR |
3066 | /* |
3067 | * If the value passed in is equal to the current preempt count | |
3068 | * then we just disabled preemption. Start timing the latency. | |
3069 | */ | |
3070 | static inline void preempt_latency_start(int val) | |
3071 | { | |
3072 | if (preempt_count() == val) { | |
3073 | unsigned long ip = get_lock_parent_ip(); | |
3074 | #ifdef CONFIG_DEBUG_PREEMPT | |
3075 | current->preempt_disable_ip = ip; | |
3076 | #endif | |
3077 | trace_preempt_off(CALLER_ADDR0, ip); | |
3078 | } | |
3079 | } | |
7e49fcce | 3080 | |
edafe3a5 | 3081 | void preempt_count_add(int val) |
1da177e4 | 3082 | { |
6cd8a4bb | 3083 | #ifdef CONFIG_DEBUG_PREEMPT |
1da177e4 LT |
3084 | /* |
3085 | * Underflow? | |
3086 | */ | |
9a11b49a IM |
3087 | if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) |
3088 | return; | |
6cd8a4bb | 3089 | #endif |
bdb43806 | 3090 | __preempt_count_add(val); |
6cd8a4bb | 3091 | #ifdef CONFIG_DEBUG_PREEMPT |
1da177e4 LT |
3092 | /* |
3093 | * Spinlock count overflowing soon? | |
3094 | */ | |
33859f7f MOS |
3095 | DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= |
3096 | PREEMPT_MASK - 10); | |
6cd8a4bb | 3097 | #endif |
47252cfb | 3098 | preempt_latency_start(val); |
1da177e4 | 3099 | } |
bdb43806 | 3100 | EXPORT_SYMBOL(preempt_count_add); |
edafe3a5 | 3101 | NOKPROBE_SYMBOL(preempt_count_add); |
1da177e4 | 3102 | |
47252cfb SR |
3103 | /* |
3104 | * If the value passed in equals to the current preempt count | |
3105 | * then we just enabled preemption. Stop timing the latency. | |
3106 | */ | |
3107 | static inline void preempt_latency_stop(int val) | |
3108 | { | |
3109 | if (preempt_count() == val) | |
3110 | trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip()); | |
3111 | } | |
3112 | ||
edafe3a5 | 3113 | void preempt_count_sub(int val) |
1da177e4 | 3114 | { |
6cd8a4bb | 3115 | #ifdef CONFIG_DEBUG_PREEMPT |
1da177e4 LT |
3116 | /* |
3117 | * Underflow? | |
3118 | */ | |
01e3eb82 | 3119 | if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) |
9a11b49a | 3120 | return; |
1da177e4 LT |
3121 | /* |
3122 | * Is the spinlock portion underflowing? | |
3123 | */ | |
9a11b49a IM |
3124 | if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && |
3125 | !(preempt_count() & PREEMPT_MASK))) | |
3126 | return; | |
6cd8a4bb | 3127 | #endif |
9a11b49a | 3128 | |
47252cfb | 3129 | preempt_latency_stop(val); |
bdb43806 | 3130 | __preempt_count_sub(val); |
1da177e4 | 3131 | } |
bdb43806 | 3132 | EXPORT_SYMBOL(preempt_count_sub); |
edafe3a5 | 3133 | NOKPROBE_SYMBOL(preempt_count_sub); |
1da177e4 | 3134 | |
47252cfb SR |
3135 | #else |
3136 | static inline void preempt_latency_start(int val) { } | |
3137 | static inline void preempt_latency_stop(int val) { } | |
1da177e4 LT |
3138 | #endif |
3139 | ||
3140 | /* | |
dd41f596 | 3141 | * Print scheduling while atomic bug: |
1da177e4 | 3142 | */ |
dd41f596 | 3143 | static noinline void __schedule_bug(struct task_struct *prev) |
1da177e4 | 3144 | { |
664dfa65 DJ |
3145 | if (oops_in_progress) |
3146 | return; | |
3147 | ||
3df0fc5b PZ |
3148 | printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", |
3149 | prev->comm, prev->pid, preempt_count()); | |
838225b4 | 3150 | |
dd41f596 | 3151 | debug_show_held_locks(prev); |
e21f5b15 | 3152 | print_modules(); |
dd41f596 IM |
3153 | if (irqs_disabled()) |
3154 | print_irqtrace_events(prev); | |
8f47b187 TG |
3155 | #ifdef CONFIG_DEBUG_PREEMPT |
3156 | if (in_atomic_preempt_off()) { | |
3157 | pr_err("Preemption disabled at:"); | |
3158 | print_ip_sym(current->preempt_disable_ip); | |
3159 | pr_cont("\n"); | |
3160 | } | |
3161 | #endif | |
6135fc1e | 3162 | dump_stack(); |
373d4d09 | 3163 | add_taint(TAINT_WARN, LOCKDEP_STILL_OK); |
dd41f596 | 3164 | } |
1da177e4 | 3165 | |
dd41f596 IM |
3166 | /* |
3167 | * Various schedule()-time debugging checks and statistics: | |
3168 | */ | |
3169 | static inline void schedule_debug(struct task_struct *prev) | |
3170 | { | |
0d9e2632 | 3171 | #ifdef CONFIG_SCHED_STACK_END_CHECK |
29d64551 JH |
3172 | if (task_stack_end_corrupted(prev)) |
3173 | panic("corrupted stack end detected inside scheduler\n"); | |
0d9e2632 | 3174 | #endif |
b99def8b | 3175 | |
1dc0fffc | 3176 | if (unlikely(in_atomic_preempt_off())) { |
dd41f596 | 3177 | __schedule_bug(prev); |
1dc0fffc PZ |
3178 | preempt_count_set(PREEMPT_DISABLED); |
3179 | } | |
b3fbab05 | 3180 | rcu_sleep_check(); |
dd41f596 | 3181 | |
1da177e4 LT |
3182 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); |
3183 | ||
2d72376b | 3184 | schedstat_inc(this_rq(), sched_count); |
dd41f596 IM |
3185 | } |
3186 | ||
3187 | /* | |
3188 | * Pick up the highest-prio task: | |
3189 | */ | |
3190 | static inline struct task_struct * | |
e7904a28 | 3191 | pick_next_task(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie) |
dd41f596 | 3192 | { |
37e117c0 | 3193 | const struct sched_class *class = &fair_sched_class; |
dd41f596 | 3194 | struct task_struct *p; |
1da177e4 LT |
3195 | |
3196 | /* | |
dd41f596 IM |
3197 | * Optimization: we know that if all tasks are in |
3198 | * the fair class we can call that function directly: | |
1da177e4 | 3199 | */ |
37e117c0 | 3200 | if (likely(prev->sched_class == class && |
38033c37 | 3201 | rq->nr_running == rq->cfs.h_nr_running)) { |
e7904a28 | 3202 | p = fair_sched_class.pick_next_task(rq, prev, cookie); |
6ccdc84b PZ |
3203 | if (unlikely(p == RETRY_TASK)) |
3204 | goto again; | |
3205 | ||
3206 | /* assumes fair_sched_class->next == idle_sched_class */ | |
3207 | if (unlikely(!p)) | |
e7904a28 | 3208 | p = idle_sched_class.pick_next_task(rq, prev, cookie); |
6ccdc84b PZ |
3209 | |
3210 | return p; | |
1da177e4 LT |
3211 | } |
3212 | ||
37e117c0 | 3213 | again: |
34f971f6 | 3214 | for_each_class(class) { |
e7904a28 | 3215 | p = class->pick_next_task(rq, prev, cookie); |
37e117c0 PZ |
3216 | if (p) { |
3217 | if (unlikely(p == RETRY_TASK)) | |
3218 | goto again; | |
dd41f596 | 3219 | return p; |
37e117c0 | 3220 | } |
dd41f596 | 3221 | } |
34f971f6 PZ |
3222 | |
3223 | BUG(); /* the idle class will always have a runnable task */ | |
dd41f596 | 3224 | } |
1da177e4 | 3225 | |
dd41f596 | 3226 | /* |
c259e01a | 3227 | * __schedule() is the main scheduler function. |
edde96ea PE |
3228 | * |
3229 | * The main means of driving the scheduler and thus entering this function are: | |
3230 | * | |
3231 | * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. | |
3232 | * | |
3233 | * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return | |
3234 | * paths. For example, see arch/x86/entry_64.S. | |
3235 | * | |
3236 | * To drive preemption between tasks, the scheduler sets the flag in timer | |
3237 | * interrupt handler scheduler_tick(). | |
3238 | * | |
3239 | * 3. Wakeups don't really cause entry into schedule(). They add a | |
3240 | * task to the run-queue and that's it. | |
3241 | * | |
3242 | * Now, if the new task added to the run-queue preempts the current | |
3243 | * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets | |
3244 | * called on the nearest possible occasion: | |
3245 | * | |
3246 | * - If the kernel is preemptible (CONFIG_PREEMPT=y): | |
3247 | * | |
3248 | * - in syscall or exception context, at the next outmost | |
3249 | * preempt_enable(). (this might be as soon as the wake_up()'s | |
3250 | * spin_unlock()!) | |
3251 | * | |
3252 | * - in IRQ context, return from interrupt-handler to | |
3253 | * preemptible context | |
3254 | * | |
3255 | * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) | |
3256 | * then at the next: | |
3257 | * | |
3258 | * - cond_resched() call | |
3259 | * - explicit schedule() call | |
3260 | * - return from syscall or exception to user-space | |
3261 | * - return from interrupt-handler to user-space | |
bfd9b2b5 | 3262 | * |
b30f0e3f | 3263 | * WARNING: must be called with preemption disabled! |
dd41f596 | 3264 | */ |
499d7955 | 3265 | static void __sched notrace __schedule(bool preempt) |
dd41f596 IM |
3266 | { |
3267 | struct task_struct *prev, *next; | |
67ca7bde | 3268 | unsigned long *switch_count; |
e7904a28 | 3269 | struct pin_cookie cookie; |
dd41f596 | 3270 | struct rq *rq; |
31656519 | 3271 | int cpu; |
dd41f596 | 3272 | |
dd41f596 IM |
3273 | cpu = smp_processor_id(); |
3274 | rq = cpu_rq(cpu); | |
dd41f596 | 3275 | prev = rq->curr; |
dd41f596 | 3276 | |
b99def8b PZ |
3277 | /* |
3278 | * do_exit() calls schedule() with preemption disabled as an exception; | |
3279 | * however we must fix that up, otherwise the next task will see an | |
3280 | * inconsistent (higher) preempt count. | |
3281 | * | |
3282 | * It also avoids the below schedule_debug() test from complaining | |
3283 | * about this. | |
3284 | */ | |
3285 | if (unlikely(prev->state == TASK_DEAD)) | |
3286 | preempt_enable_no_resched_notrace(); | |
3287 | ||
dd41f596 | 3288 | schedule_debug(prev); |
1da177e4 | 3289 | |
31656519 | 3290 | if (sched_feat(HRTICK)) |
f333fdc9 | 3291 | hrtick_clear(rq); |
8f4d37ec | 3292 | |
46a5d164 PM |
3293 | local_irq_disable(); |
3294 | rcu_note_context_switch(); | |
3295 | ||
e0acd0a6 ON |
3296 | /* |
3297 | * Make sure that signal_pending_state()->signal_pending() below | |
3298 | * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) | |
3299 | * done by the caller to avoid the race with signal_wake_up(). | |
3300 | */ | |
3301 | smp_mb__before_spinlock(); | |
46a5d164 | 3302 | raw_spin_lock(&rq->lock); |
e7904a28 | 3303 | cookie = lockdep_pin_lock(&rq->lock); |
1da177e4 | 3304 | |
9edfbfed PZ |
3305 | rq->clock_skip_update <<= 1; /* promote REQ to ACT */ |
3306 | ||
246d86b5 | 3307 | switch_count = &prev->nivcsw; |
fc13aeba | 3308 | if (!preempt && prev->state) { |
21aa9af0 | 3309 | if (unlikely(signal_pending_state(prev->state, prev))) { |
1da177e4 | 3310 | prev->state = TASK_RUNNING; |
21aa9af0 | 3311 | } else { |
2acca55e PZ |
3312 | deactivate_task(rq, prev, DEQUEUE_SLEEP); |
3313 | prev->on_rq = 0; | |
3314 | ||
21aa9af0 | 3315 | /* |
2acca55e PZ |
3316 | * If a worker went to sleep, notify and ask workqueue |
3317 | * whether it wants to wake up a task to maintain | |
3318 | * concurrency. | |
21aa9af0 TH |
3319 | */ |
3320 | if (prev->flags & PF_WQ_WORKER) { | |
3321 | struct task_struct *to_wakeup; | |
3322 | ||
9b7f6597 | 3323 | to_wakeup = wq_worker_sleeping(prev); |
21aa9af0 | 3324 | if (to_wakeup) |
e7904a28 | 3325 | try_to_wake_up_local(to_wakeup, cookie); |
21aa9af0 | 3326 | } |
21aa9af0 | 3327 | } |
dd41f596 | 3328 | switch_count = &prev->nvcsw; |
1da177e4 LT |
3329 | } |
3330 | ||
9edfbfed | 3331 | if (task_on_rq_queued(prev)) |
606dba2e PZ |
3332 | update_rq_clock(rq); |
3333 | ||
e7904a28 | 3334 | next = pick_next_task(rq, prev, cookie); |
f26f9aff | 3335 | clear_tsk_need_resched(prev); |
f27dde8d | 3336 | clear_preempt_need_resched(); |
9edfbfed | 3337 | rq->clock_skip_update = 0; |
1da177e4 | 3338 | |
1da177e4 | 3339 | if (likely(prev != next)) { |
1da177e4 LT |
3340 | rq->nr_switches++; |
3341 | rq->curr = next; | |
3342 | ++*switch_count; | |
3343 | ||
c73464b1 | 3344 | trace_sched_switch(preempt, prev, next); |
e7904a28 | 3345 | rq = context_switch(rq, prev, next, cookie); /* unlocks the rq */ |
cbce1a68 | 3346 | } else { |
e7904a28 | 3347 | lockdep_unpin_lock(&rq->lock, cookie); |
05fa785c | 3348 | raw_spin_unlock_irq(&rq->lock); |
cbce1a68 | 3349 | } |
1da177e4 | 3350 | |
e3fca9e7 | 3351 | balance_callback(rq); |
1da177e4 | 3352 | } |
8e05e96a | 3353 | STACK_FRAME_NON_STANDARD(__schedule); /* switch_to() */ |
c259e01a | 3354 | |
9c40cef2 TG |
3355 | static inline void sched_submit_work(struct task_struct *tsk) |
3356 | { | |
3c7d5184 | 3357 | if (!tsk->state || tsk_is_pi_blocked(tsk)) |
9c40cef2 TG |
3358 | return; |
3359 | /* | |
3360 | * If we are going to sleep and we have plugged IO queued, | |
3361 | * make sure to submit it to avoid deadlocks. | |
3362 | */ | |
3363 | if (blk_needs_flush_plug(tsk)) | |
3364 | blk_schedule_flush_plug(tsk); | |
3365 | } | |
3366 | ||
722a9f92 | 3367 | asmlinkage __visible void __sched schedule(void) |
c259e01a | 3368 | { |
9c40cef2 TG |
3369 | struct task_struct *tsk = current; |
3370 | ||
3371 | sched_submit_work(tsk); | |
bfd9b2b5 | 3372 | do { |
b30f0e3f | 3373 | preempt_disable(); |
fc13aeba | 3374 | __schedule(false); |
b30f0e3f | 3375 | sched_preempt_enable_no_resched(); |
bfd9b2b5 | 3376 | } while (need_resched()); |
c259e01a | 3377 | } |
1da177e4 LT |
3378 | EXPORT_SYMBOL(schedule); |
3379 | ||
91d1aa43 | 3380 | #ifdef CONFIG_CONTEXT_TRACKING |
722a9f92 | 3381 | asmlinkage __visible void __sched schedule_user(void) |
20ab65e3 FW |
3382 | { |
3383 | /* | |
3384 | * If we come here after a random call to set_need_resched(), | |
3385 | * or we have been woken up remotely but the IPI has not yet arrived, | |
3386 | * we haven't yet exited the RCU idle mode. Do it here manually until | |
3387 | * we find a better solution. | |
7cc78f8f AL |
3388 | * |
3389 | * NB: There are buggy callers of this function. Ideally we | |
c467ea76 | 3390 | * should warn if prev_state != CONTEXT_USER, but that will trigger |
7cc78f8f | 3391 | * too frequently to make sense yet. |
20ab65e3 | 3392 | */ |
7cc78f8f | 3393 | enum ctx_state prev_state = exception_enter(); |
20ab65e3 | 3394 | schedule(); |
7cc78f8f | 3395 | exception_exit(prev_state); |
20ab65e3 FW |
3396 | } |
3397 | #endif | |
3398 | ||
c5491ea7 TG |
3399 | /** |
3400 | * schedule_preempt_disabled - called with preemption disabled | |
3401 | * | |
3402 | * Returns with preemption disabled. Note: preempt_count must be 1 | |
3403 | */ | |
3404 | void __sched schedule_preempt_disabled(void) | |
3405 | { | |
ba74c144 | 3406 | sched_preempt_enable_no_resched(); |
c5491ea7 TG |
3407 | schedule(); |
3408 | preempt_disable(); | |
3409 | } | |
3410 | ||
06b1f808 | 3411 | static void __sched notrace preempt_schedule_common(void) |
a18b5d01 FW |
3412 | { |
3413 | do { | |
47252cfb SR |
3414 | /* |
3415 | * Because the function tracer can trace preempt_count_sub() | |
3416 | * and it also uses preempt_enable/disable_notrace(), if | |
3417 | * NEED_RESCHED is set, the preempt_enable_notrace() called | |
3418 | * by the function tracer will call this function again and | |
3419 | * cause infinite recursion. | |
3420 | * | |
3421 | * Preemption must be disabled here before the function | |
3422 | * tracer can trace. Break up preempt_disable() into two | |
3423 | * calls. One to disable preemption without fear of being | |
3424 | * traced. The other to still record the preemption latency, | |
3425 | * which can also be traced by the function tracer. | |
3426 | */ | |
499d7955 | 3427 | preempt_disable_notrace(); |
47252cfb | 3428 | preempt_latency_start(1); |
fc13aeba | 3429 | __schedule(true); |
47252cfb | 3430 | preempt_latency_stop(1); |
499d7955 | 3431 | preempt_enable_no_resched_notrace(); |
a18b5d01 FW |
3432 | |
3433 | /* | |
3434 | * Check again in case we missed a preemption opportunity | |
3435 | * between schedule and now. | |
3436 | */ | |
a18b5d01 FW |
3437 | } while (need_resched()); |
3438 | } | |
3439 | ||
1da177e4 LT |
3440 | #ifdef CONFIG_PREEMPT |
3441 | /* | |
2ed6e34f | 3442 | * this is the entry point to schedule() from in-kernel preemption |
41a2d6cf | 3443 | * off of preempt_enable. Kernel preemptions off return from interrupt |
1da177e4 LT |
3444 | * occur there and call schedule directly. |
3445 | */ | |
722a9f92 | 3446 | asmlinkage __visible void __sched notrace preempt_schedule(void) |
1da177e4 | 3447 | { |
1da177e4 LT |
3448 | /* |
3449 | * If there is a non-zero preempt_count or interrupts are disabled, | |
41a2d6cf | 3450 | * we do not want to preempt the current task. Just return.. |
1da177e4 | 3451 | */ |
fbb00b56 | 3452 | if (likely(!preemptible())) |
1da177e4 LT |
3453 | return; |
3454 | ||
a18b5d01 | 3455 | preempt_schedule_common(); |
1da177e4 | 3456 | } |
376e2424 | 3457 | NOKPROBE_SYMBOL(preempt_schedule); |
1da177e4 | 3458 | EXPORT_SYMBOL(preempt_schedule); |
009f60e2 | 3459 | |
009f60e2 | 3460 | /** |
4eaca0a8 | 3461 | * preempt_schedule_notrace - preempt_schedule called by tracing |
009f60e2 ON |
3462 | * |
3463 | * The tracing infrastructure uses preempt_enable_notrace to prevent | |
3464 | * recursion and tracing preempt enabling caused by the tracing | |
3465 | * infrastructure itself. But as tracing can happen in areas coming | |
3466 | * from userspace or just about to enter userspace, a preempt enable | |
3467 | * can occur before user_exit() is called. This will cause the scheduler | |
3468 | * to be called when the system is still in usermode. | |
3469 | * | |
3470 | * To prevent this, the preempt_enable_notrace will use this function | |
3471 | * instead of preempt_schedule() to exit user context if needed before | |
3472 | * calling the scheduler. | |
3473 | */ | |
4eaca0a8 | 3474 | asmlinkage __visible void __sched notrace preempt_schedule_notrace(void) |
009f60e2 ON |
3475 | { |
3476 | enum ctx_state prev_ctx; | |
3477 | ||
3478 | if (likely(!preemptible())) | |
3479 | return; | |
3480 | ||
3481 | do { | |
47252cfb SR |
3482 | /* |
3483 | * Because the function tracer can trace preempt_count_sub() | |
3484 | * and it also uses preempt_enable/disable_notrace(), if | |
3485 | * NEED_RESCHED is set, the preempt_enable_notrace() called | |
3486 | * by the function tracer will call this function again and | |
3487 | * cause infinite recursion. | |
3488 | * | |
3489 | * Preemption must be disabled here before the function | |
3490 | * tracer can trace. Break up preempt_disable() into two | |
3491 | * calls. One to disable preemption without fear of being | |
3492 | * traced. The other to still record the preemption latency, | |
3493 | * which can also be traced by the function tracer. | |
3494 | */ | |
3d8f74dd | 3495 | preempt_disable_notrace(); |
47252cfb | 3496 | preempt_latency_start(1); |
009f60e2 ON |
3497 | /* |
3498 | * Needs preempt disabled in case user_exit() is traced | |
3499 | * and the tracer calls preempt_enable_notrace() causing | |
3500 | * an infinite recursion. | |
3501 | */ | |
3502 | prev_ctx = exception_enter(); | |
fc13aeba | 3503 | __schedule(true); |
009f60e2 ON |
3504 | exception_exit(prev_ctx); |
3505 | ||
47252cfb | 3506 | preempt_latency_stop(1); |
3d8f74dd | 3507 | preempt_enable_no_resched_notrace(); |
009f60e2 ON |
3508 | } while (need_resched()); |
3509 | } | |
4eaca0a8 | 3510 | EXPORT_SYMBOL_GPL(preempt_schedule_notrace); |
009f60e2 | 3511 | |
32e475d7 | 3512 | #endif /* CONFIG_PREEMPT */ |
1da177e4 LT |
3513 | |
3514 | /* | |
2ed6e34f | 3515 | * this is the entry point to schedule() from kernel preemption |
1da177e4 LT |
3516 | * off of irq context. |
3517 | * Note, that this is called and return with irqs disabled. This will | |
3518 | * protect us against recursive calling from irq. | |
3519 | */ | |
722a9f92 | 3520 | asmlinkage __visible void __sched preempt_schedule_irq(void) |
1da177e4 | 3521 | { |
b22366cd | 3522 | enum ctx_state prev_state; |
6478d880 | 3523 | |
2ed6e34f | 3524 | /* Catch callers which need to be fixed */ |
f27dde8d | 3525 | BUG_ON(preempt_count() || !irqs_disabled()); |
1da177e4 | 3526 | |
b22366cd FW |
3527 | prev_state = exception_enter(); |
3528 | ||
3a5c359a | 3529 | do { |
3d8f74dd | 3530 | preempt_disable(); |
3a5c359a | 3531 | local_irq_enable(); |
fc13aeba | 3532 | __schedule(true); |
3a5c359a | 3533 | local_irq_disable(); |
3d8f74dd | 3534 | sched_preempt_enable_no_resched(); |
5ed0cec0 | 3535 | } while (need_resched()); |
b22366cd FW |
3536 | |
3537 | exception_exit(prev_state); | |
1da177e4 LT |
3538 | } |
3539 | ||
63859d4f | 3540 | int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, |
95cdf3b7 | 3541 | void *key) |
1da177e4 | 3542 | { |
63859d4f | 3543 | return try_to_wake_up(curr->private, mode, wake_flags); |
1da177e4 | 3544 | } |
1da177e4 LT |
3545 | EXPORT_SYMBOL(default_wake_function); |
3546 | ||
b29739f9 IM |
3547 | #ifdef CONFIG_RT_MUTEXES |
3548 | ||
3549 | /* | |
3550 | * rt_mutex_setprio - set the current priority of a task | |
3551 | * @p: task | |
3552 | * @prio: prio value (kernel-internal form) | |
3553 | * | |
3554 | * This function changes the 'effective' priority of a task. It does | |
3555 | * not touch ->normal_prio like __setscheduler(). | |
3556 | * | |
c365c292 TG |
3557 | * Used by the rt_mutex code to implement priority inheritance |
3558 | * logic. Call site only calls if the priority of the task changed. | |
b29739f9 | 3559 | */ |
36c8b586 | 3560 | void rt_mutex_setprio(struct task_struct *p, int prio) |
b29739f9 | 3561 | { |
ff77e468 | 3562 | int oldprio, queued, running, queue_flag = DEQUEUE_SAVE | DEQUEUE_MOVE; |
83ab0aa0 | 3563 | const struct sched_class *prev_class; |
eb580751 PZ |
3564 | struct rq_flags rf; |
3565 | struct rq *rq; | |
b29739f9 | 3566 | |
aab03e05 | 3567 | BUG_ON(prio > MAX_PRIO); |
b29739f9 | 3568 | |
eb580751 | 3569 | rq = __task_rq_lock(p, &rf); |
b29739f9 | 3570 | |
1c4dd99b TG |
3571 | /* |
3572 | * Idle task boosting is a nono in general. There is one | |
3573 | * exception, when PREEMPT_RT and NOHZ is active: | |
3574 | * | |
3575 | * The idle task calls get_next_timer_interrupt() and holds | |
3576 | * the timer wheel base->lock on the CPU and another CPU wants | |
3577 | * to access the timer (probably to cancel it). We can safely | |
3578 | * ignore the boosting request, as the idle CPU runs this code | |
3579 | * with interrupts disabled and will complete the lock | |
3580 | * protected section without being interrupted. So there is no | |
3581 | * real need to boost. | |
3582 | */ | |
3583 | if (unlikely(p == rq->idle)) { | |
3584 | WARN_ON(p != rq->curr); | |
3585 | WARN_ON(p->pi_blocked_on); | |
3586 | goto out_unlock; | |
3587 | } | |
3588 | ||
a8027073 | 3589 | trace_sched_pi_setprio(p, prio); |
d5f9f942 | 3590 | oldprio = p->prio; |
ff77e468 PZ |
3591 | |
3592 | if (oldprio == prio) | |
3593 | queue_flag &= ~DEQUEUE_MOVE; | |
3594 | ||
83ab0aa0 | 3595 | prev_class = p->sched_class; |
da0c1e65 | 3596 | queued = task_on_rq_queued(p); |
051a1d1a | 3597 | running = task_current(rq, p); |
da0c1e65 | 3598 | if (queued) |
ff77e468 | 3599 | dequeue_task(rq, p, queue_flag); |
0e1f3483 | 3600 | if (running) |
f3cd1c4e | 3601 | put_prev_task(rq, p); |
dd41f596 | 3602 | |
2d3d891d DF |
3603 | /* |
3604 | * Boosting condition are: | |
3605 | * 1. -rt task is running and holds mutex A | |
3606 | * --> -dl task blocks on mutex A | |
3607 | * | |
3608 | * 2. -dl task is running and holds mutex A | |
3609 | * --> -dl task blocks on mutex A and could preempt the | |
3610 | * running task | |
3611 | */ | |
3612 | if (dl_prio(prio)) { | |
466af29b ON |
3613 | struct task_struct *pi_task = rt_mutex_get_top_task(p); |
3614 | if (!dl_prio(p->normal_prio) || | |
3615 | (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) { | |
2d3d891d | 3616 | p->dl.dl_boosted = 1; |
ff77e468 | 3617 | queue_flag |= ENQUEUE_REPLENISH; |
2d3d891d DF |
3618 | } else |
3619 | p->dl.dl_boosted = 0; | |
aab03e05 | 3620 | p->sched_class = &dl_sched_class; |
2d3d891d DF |
3621 | } else if (rt_prio(prio)) { |
3622 | if (dl_prio(oldprio)) | |
3623 | p->dl.dl_boosted = 0; | |
3624 | if (oldprio < prio) | |
ff77e468 | 3625 | queue_flag |= ENQUEUE_HEAD; |
dd41f596 | 3626 | p->sched_class = &rt_sched_class; |
2d3d891d DF |
3627 | } else { |
3628 | if (dl_prio(oldprio)) | |
3629 | p->dl.dl_boosted = 0; | |
746db944 BS |
3630 | if (rt_prio(oldprio)) |
3631 | p->rt.timeout = 0; | |
dd41f596 | 3632 | p->sched_class = &fair_sched_class; |
2d3d891d | 3633 | } |
dd41f596 | 3634 | |
b29739f9 IM |
3635 | p->prio = prio; |
3636 | ||
0e1f3483 HS |
3637 | if (running) |
3638 | p->sched_class->set_curr_task(rq); | |
da0c1e65 | 3639 | if (queued) |
ff77e468 | 3640 | enqueue_task(rq, p, queue_flag); |
cb469845 | 3641 | |
da7a735e | 3642 | check_class_changed(rq, p, prev_class, oldprio); |
1c4dd99b | 3643 | out_unlock: |
4c9a4bc8 | 3644 | preempt_disable(); /* avoid rq from going away on us */ |
eb580751 | 3645 | __task_rq_unlock(rq, &rf); |
4c9a4bc8 PZ |
3646 | |
3647 | balance_callback(rq); | |
3648 | preempt_enable(); | |
b29739f9 | 3649 | } |
b29739f9 | 3650 | #endif |
d50dde5a | 3651 | |
36c8b586 | 3652 | void set_user_nice(struct task_struct *p, long nice) |
1da177e4 | 3653 | { |
da0c1e65 | 3654 | int old_prio, delta, queued; |
eb580751 | 3655 | struct rq_flags rf; |
70b97a7f | 3656 | struct rq *rq; |
1da177e4 | 3657 | |
75e45d51 | 3658 | if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE) |
1da177e4 LT |
3659 | return; |
3660 | /* | |
3661 | * We have to be careful, if called from sys_setpriority(), | |
3662 | * the task might be in the middle of scheduling on another CPU. | |
3663 | */ | |
eb580751 | 3664 | rq = task_rq_lock(p, &rf); |
1da177e4 LT |
3665 | /* |
3666 | * The RT priorities are set via sched_setscheduler(), but we still | |
3667 | * allow the 'normal' nice value to be set - but as expected | |
3668 | * it wont have any effect on scheduling until the task is | |
aab03e05 | 3669 | * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: |
1da177e4 | 3670 | */ |
aab03e05 | 3671 | if (task_has_dl_policy(p) || task_has_rt_policy(p)) { |
1da177e4 LT |
3672 | p->static_prio = NICE_TO_PRIO(nice); |
3673 | goto out_unlock; | |
3674 | } | |
da0c1e65 KT |
3675 | queued = task_on_rq_queued(p); |
3676 | if (queued) | |
1de64443 | 3677 | dequeue_task(rq, p, DEQUEUE_SAVE); |
1da177e4 | 3678 | |
1da177e4 | 3679 | p->static_prio = NICE_TO_PRIO(nice); |
2dd73a4f | 3680 | set_load_weight(p); |
b29739f9 IM |
3681 | old_prio = p->prio; |
3682 | p->prio = effective_prio(p); | |
3683 | delta = p->prio - old_prio; | |
1da177e4 | 3684 | |
da0c1e65 | 3685 | if (queued) { |
1de64443 | 3686 | enqueue_task(rq, p, ENQUEUE_RESTORE); |
1da177e4 | 3687 | /* |
d5f9f942 AM |
3688 | * If the task increased its priority or is running and |
3689 | * lowered its priority, then reschedule its CPU: | |
1da177e4 | 3690 | */ |
d5f9f942 | 3691 | if (delta < 0 || (delta > 0 && task_running(rq, p))) |
8875125e | 3692 | resched_curr(rq); |
1da177e4 LT |
3693 | } |
3694 | out_unlock: | |
eb580751 | 3695 | task_rq_unlock(rq, p, &rf); |
1da177e4 | 3696 | } |
1da177e4 LT |
3697 | EXPORT_SYMBOL(set_user_nice); |
3698 | ||
e43379f1 MM |
3699 | /* |
3700 | * can_nice - check if a task can reduce its nice value | |
3701 | * @p: task | |
3702 | * @nice: nice value | |
3703 | */ | |
36c8b586 | 3704 | int can_nice(const struct task_struct *p, const int nice) |
e43379f1 | 3705 | { |
024f4747 | 3706 | /* convert nice value [19,-20] to rlimit style value [1,40] */ |
7aa2c016 | 3707 | int nice_rlim = nice_to_rlimit(nice); |
48f24c4d | 3708 | |
78d7d407 | 3709 | return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || |
e43379f1 MM |
3710 | capable(CAP_SYS_NICE)); |
3711 | } | |
3712 | ||
1da177e4 LT |
3713 | #ifdef __ARCH_WANT_SYS_NICE |
3714 | ||
3715 | /* | |
3716 | * sys_nice - change the priority of the current process. | |
3717 | * @increment: priority increment | |
3718 | * | |
3719 | * sys_setpriority is a more generic, but much slower function that | |
3720 | * does similar things. | |
3721 | */ | |
5add95d4 | 3722 | SYSCALL_DEFINE1(nice, int, increment) |
1da177e4 | 3723 | { |
48f24c4d | 3724 | long nice, retval; |
1da177e4 LT |
3725 | |
3726 | /* | |
3727 | * Setpriority might change our priority at the same moment. | |
3728 | * We don't have to worry. Conceptually one call occurs first | |
3729 | * and we have a single winner. | |
3730 | */ | |
a9467fa3 | 3731 | increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH); |
d0ea0268 | 3732 | nice = task_nice(current) + increment; |
1da177e4 | 3733 | |
a9467fa3 | 3734 | nice = clamp_val(nice, MIN_NICE, MAX_NICE); |
e43379f1 MM |
3735 | if (increment < 0 && !can_nice(current, nice)) |
3736 | return -EPERM; | |
3737 | ||
1da177e4 LT |
3738 | retval = security_task_setnice(current, nice); |
3739 | if (retval) | |
3740 | return retval; | |
3741 | ||
3742 | set_user_nice(current, nice); | |
3743 | return 0; | |
3744 | } | |
3745 | ||
3746 | #endif | |
3747 | ||
3748 | /** | |
3749 | * task_prio - return the priority value of a given task. | |
3750 | * @p: the task in question. | |
3751 | * | |
e69f6186 | 3752 | * Return: The priority value as seen by users in /proc. |
1da177e4 LT |
3753 | * RT tasks are offset by -200. Normal tasks are centered |
3754 | * around 0, value goes from -16 to +15. | |
3755 | */ | |
36c8b586 | 3756 | int task_prio(const struct task_struct *p) |
1da177e4 LT |
3757 | { |
3758 | return p->prio - MAX_RT_PRIO; | |
3759 | } | |
3760 | ||
1da177e4 LT |
3761 | /** |
3762 | * idle_cpu - is a given cpu idle currently? | |
3763 | * @cpu: the processor in question. | |
e69f6186 YB |
3764 | * |
3765 | * Return: 1 if the CPU is currently idle. 0 otherwise. | |
1da177e4 LT |
3766 | */ |
3767 | int idle_cpu(int cpu) | |
3768 | { | |
908a3283 TG |
3769 | struct rq *rq = cpu_rq(cpu); |
3770 | ||
3771 | if (rq->curr != rq->idle) | |
3772 | return 0; | |
3773 | ||
3774 | if (rq->nr_running) | |
3775 | return 0; | |
3776 | ||
3777 | #ifdef CONFIG_SMP | |
3778 | if (!llist_empty(&rq->wake_list)) | |
3779 | return 0; | |
3780 | #endif | |
3781 | ||
3782 | return 1; | |
1da177e4 LT |
3783 | } |
3784 | ||
1da177e4 LT |
3785 | /** |
3786 | * idle_task - return the idle task for a given cpu. | |
3787 | * @cpu: the processor in question. | |
e69f6186 YB |
3788 | * |
3789 | * Return: The idle task for the cpu @cpu. | |
1da177e4 | 3790 | */ |
36c8b586 | 3791 | struct task_struct *idle_task(int cpu) |
1da177e4 LT |
3792 | { |
3793 | return cpu_rq(cpu)->idle; | |
3794 | } | |
3795 | ||
3796 | /** | |
3797 | * find_process_by_pid - find a process with a matching PID value. | |
3798 | * @pid: the pid in question. | |
e69f6186 YB |
3799 | * |
3800 | * The task of @pid, if found. %NULL otherwise. | |
1da177e4 | 3801 | */ |
a9957449 | 3802 | static struct task_struct *find_process_by_pid(pid_t pid) |
1da177e4 | 3803 | { |
228ebcbe | 3804 | return pid ? find_task_by_vpid(pid) : current; |
1da177e4 LT |
3805 | } |
3806 | ||
aab03e05 DF |
3807 | /* |
3808 | * This function initializes the sched_dl_entity of a newly becoming | |
3809 | * SCHED_DEADLINE task. | |
3810 | * | |
3811 | * Only the static values are considered here, the actual runtime and the | |
3812 | * absolute deadline will be properly calculated when the task is enqueued | |
3813 | * for the first time with its new policy. | |
3814 | */ | |
3815 | static void | |
3816 | __setparam_dl(struct task_struct *p, const struct sched_attr *attr) | |
3817 | { | |
3818 | struct sched_dl_entity *dl_se = &p->dl; | |
3819 | ||
aab03e05 DF |
3820 | dl_se->dl_runtime = attr->sched_runtime; |
3821 | dl_se->dl_deadline = attr->sched_deadline; | |
755378a4 | 3822 | dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; |
aab03e05 | 3823 | dl_se->flags = attr->sched_flags; |
332ac17e | 3824 | dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); |
40767b0d PZ |
3825 | |
3826 | /* | |
3827 | * Changing the parameters of a task is 'tricky' and we're not doing | |
3828 | * the correct thing -- also see task_dead_dl() and switched_from_dl(). | |
3829 | * | |
3830 | * What we SHOULD do is delay the bandwidth release until the 0-lag | |
3831 | * point. This would include retaining the task_struct until that time | |
3832 | * and change dl_overflow() to not immediately decrement the current | |
3833 | * amount. | |
3834 | * | |
3835 | * Instead we retain the current runtime/deadline and let the new | |
3836 | * parameters take effect after the current reservation period lapses. | |
3837 | * This is safe (albeit pessimistic) because the 0-lag point is always | |
3838 | * before the current scheduling deadline. | |
3839 | * | |
3840 | * We can still have temporary overloads because we do not delay the | |
3841 | * change in bandwidth until that time; so admission control is | |
3842 | * not on the safe side. It does however guarantee tasks will never | |
3843 | * consume more than promised. | |
3844 | */ | |
aab03e05 DF |
3845 | } |
3846 | ||
c13db6b1 SR |
3847 | /* |
3848 | * sched_setparam() passes in -1 for its policy, to let the functions | |
3849 | * it calls know not to change it. | |
3850 | */ | |
3851 | #define SETPARAM_POLICY -1 | |
3852 | ||
c365c292 TG |
3853 | static void __setscheduler_params(struct task_struct *p, |
3854 | const struct sched_attr *attr) | |
1da177e4 | 3855 | { |
d50dde5a DF |
3856 | int policy = attr->sched_policy; |
3857 | ||
c13db6b1 | 3858 | if (policy == SETPARAM_POLICY) |
39fd8fd2 PZ |
3859 | policy = p->policy; |
3860 | ||
1da177e4 | 3861 | p->policy = policy; |
d50dde5a | 3862 | |
aab03e05 DF |
3863 | if (dl_policy(policy)) |
3864 | __setparam_dl(p, attr); | |
39fd8fd2 | 3865 | else if (fair_policy(policy)) |
d50dde5a DF |
3866 | p->static_prio = NICE_TO_PRIO(attr->sched_nice); |
3867 | ||
39fd8fd2 PZ |
3868 | /* |
3869 | * __sched_setscheduler() ensures attr->sched_priority == 0 when | |
3870 | * !rt_policy. Always setting this ensures that things like | |
3871 | * getparam()/getattr() don't report silly values for !rt tasks. | |
3872 | */ | |
3873 | p->rt_priority = attr->sched_priority; | |
383afd09 | 3874 | p->normal_prio = normal_prio(p); |
c365c292 TG |
3875 | set_load_weight(p); |
3876 | } | |
39fd8fd2 | 3877 | |
c365c292 TG |
3878 | /* Actually do priority change: must hold pi & rq lock. */ |
3879 | static void __setscheduler(struct rq *rq, struct task_struct *p, | |
0782e63b | 3880 | const struct sched_attr *attr, bool keep_boost) |
c365c292 TG |
3881 | { |
3882 | __setscheduler_params(p, attr); | |
d50dde5a | 3883 | |
383afd09 | 3884 | /* |
0782e63b TG |
3885 | * Keep a potential priority boosting if called from |
3886 | * sched_setscheduler(). | |
383afd09 | 3887 | */ |
0782e63b TG |
3888 | if (keep_boost) |
3889 | p->prio = rt_mutex_get_effective_prio(p, normal_prio(p)); | |
3890 | else | |
3891 | p->prio = normal_prio(p); | |
383afd09 | 3892 | |
aab03e05 DF |
3893 | if (dl_prio(p->prio)) |
3894 | p->sched_class = &dl_sched_class; | |
3895 | else if (rt_prio(p->prio)) | |
ffd44db5 PZ |
3896 | p->sched_class = &rt_sched_class; |
3897 | else | |
3898 | p->sched_class = &fair_sched_class; | |
1da177e4 | 3899 | } |
aab03e05 DF |
3900 | |
3901 | static void | |
3902 | __getparam_dl(struct task_struct *p, struct sched_attr *attr) | |
3903 | { | |
3904 | struct sched_dl_entity *dl_se = &p->dl; | |
3905 | ||
3906 | attr->sched_priority = p->rt_priority; | |
3907 | attr->sched_runtime = dl_se->dl_runtime; | |
3908 | attr->sched_deadline = dl_se->dl_deadline; | |
755378a4 | 3909 | attr->sched_period = dl_se->dl_period; |
aab03e05 DF |
3910 | attr->sched_flags = dl_se->flags; |
3911 | } | |
3912 | ||
3913 | /* | |
3914 | * This function validates the new parameters of a -deadline task. | |
3915 | * We ask for the deadline not being zero, and greater or equal | |
755378a4 | 3916 | * than the runtime, as well as the period of being zero or |
332ac17e | 3917 | * greater than deadline. Furthermore, we have to be sure that |
b0827819 JL |
3918 | * user parameters are above the internal resolution of 1us (we |
3919 | * check sched_runtime only since it is always the smaller one) and | |
3920 | * below 2^63 ns (we have to check both sched_deadline and | |
3921 | * sched_period, as the latter can be zero). | |
aab03e05 DF |
3922 | */ |
3923 | static bool | |
3924 | __checkparam_dl(const struct sched_attr *attr) | |
3925 | { | |
b0827819 JL |
3926 | /* deadline != 0 */ |
3927 | if (attr->sched_deadline == 0) | |
3928 | return false; | |
3929 | ||
3930 | /* | |
3931 | * Since we truncate DL_SCALE bits, make sure we're at least | |
3932 | * that big. | |
3933 | */ | |
3934 | if (attr->sched_runtime < (1ULL << DL_SCALE)) | |
3935 | return false; | |
3936 | ||
3937 | /* | |
3938 | * Since we use the MSB for wrap-around and sign issues, make | |
3939 | * sure it's not set (mind that period can be equal to zero). | |
3940 | */ | |
3941 | if (attr->sched_deadline & (1ULL << 63) || | |
3942 | attr->sched_period & (1ULL << 63)) | |
3943 | return false; | |
3944 | ||
3945 | /* runtime <= deadline <= period (if period != 0) */ | |
3946 | if ((attr->sched_period != 0 && | |
3947 | attr->sched_period < attr->sched_deadline) || | |
3948 | attr->sched_deadline < attr->sched_runtime) | |
3949 | return false; | |
3950 | ||
3951 | return true; | |
aab03e05 DF |
3952 | } |
3953 | ||
c69e8d9c DH |
3954 | /* |
3955 | * check the target process has a UID that matches the current process's | |
3956 | */ | |
3957 | static bool check_same_owner(struct task_struct *p) | |
3958 | { | |
3959 | const struct cred *cred = current_cred(), *pcred; | |
3960 | bool match; | |
3961 | ||
3962 | rcu_read_lock(); | |
3963 | pcred = __task_cred(p); | |
9c806aa0 EB |
3964 | match = (uid_eq(cred->euid, pcred->euid) || |
3965 | uid_eq(cred->euid, pcred->uid)); | |
c69e8d9c DH |
3966 | rcu_read_unlock(); |
3967 | return match; | |
3968 | } | |
3969 | ||
75381608 WL |
3970 | static bool dl_param_changed(struct task_struct *p, |
3971 | const struct sched_attr *attr) | |
3972 | { | |
3973 | struct sched_dl_entity *dl_se = &p->dl; | |
3974 | ||
3975 | if (dl_se->dl_runtime != attr->sched_runtime || | |
3976 | dl_se->dl_deadline != attr->sched_deadline || | |
3977 | dl_se->dl_period != attr->sched_period || | |
3978 | dl_se->flags != attr->sched_flags) | |
3979 | return true; | |
3980 | ||
3981 | return false; | |
3982 | } | |
3983 | ||
d50dde5a DF |
3984 | static int __sched_setscheduler(struct task_struct *p, |
3985 | const struct sched_attr *attr, | |
dbc7f069 | 3986 | bool user, bool pi) |
1da177e4 | 3987 | { |
383afd09 SR |
3988 | int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 : |
3989 | MAX_RT_PRIO - 1 - attr->sched_priority; | |
da0c1e65 | 3990 | int retval, oldprio, oldpolicy = -1, queued, running; |
0782e63b | 3991 | int new_effective_prio, policy = attr->sched_policy; |
83ab0aa0 | 3992 | const struct sched_class *prev_class; |
eb580751 | 3993 | struct rq_flags rf; |
ca94c442 | 3994 | int reset_on_fork; |
ff77e468 | 3995 | int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE; |
eb580751 | 3996 | struct rq *rq; |
1da177e4 | 3997 | |
66e5393a SR |
3998 | /* may grab non-irq protected spin_locks */ |
3999 | BUG_ON(in_interrupt()); | |
1da177e4 LT |
4000 | recheck: |
4001 | /* double check policy once rq lock held */ | |
ca94c442 LP |
4002 | if (policy < 0) { |
4003 | reset_on_fork = p->sched_reset_on_fork; | |
1da177e4 | 4004 | policy = oldpolicy = p->policy; |
ca94c442 | 4005 | } else { |
7479f3c9 | 4006 | reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); |
ca94c442 | 4007 | |
20f9cd2a | 4008 | if (!valid_policy(policy)) |
ca94c442 LP |
4009 | return -EINVAL; |
4010 | } | |
4011 | ||
7479f3c9 PZ |
4012 | if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK)) |
4013 | return -EINVAL; | |
4014 | ||
1da177e4 LT |
4015 | /* |
4016 | * Valid priorities for SCHED_FIFO and SCHED_RR are | |
dd41f596 IM |
4017 | * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, |
4018 | * SCHED_BATCH and SCHED_IDLE is 0. | |
1da177e4 | 4019 | */ |
0bb040a4 | 4020 | if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) || |
d50dde5a | 4021 | (!p->mm && attr->sched_priority > MAX_RT_PRIO-1)) |
1da177e4 | 4022 | return -EINVAL; |
aab03e05 DF |
4023 | if ((dl_policy(policy) && !__checkparam_dl(attr)) || |
4024 | (rt_policy(policy) != (attr->sched_priority != 0))) | |
1da177e4 LT |
4025 | return -EINVAL; |
4026 | ||
37e4ab3f OC |
4027 | /* |
4028 | * Allow unprivileged RT tasks to decrease priority: | |
4029 | */ | |
961ccddd | 4030 | if (user && !capable(CAP_SYS_NICE)) { |
d50dde5a | 4031 | if (fair_policy(policy)) { |
d0ea0268 | 4032 | if (attr->sched_nice < task_nice(p) && |
eaad4513 | 4033 | !can_nice(p, attr->sched_nice)) |
d50dde5a DF |
4034 | return -EPERM; |
4035 | } | |
4036 | ||
e05606d3 | 4037 | if (rt_policy(policy)) { |
a44702e8 ON |
4038 | unsigned long rlim_rtprio = |
4039 | task_rlimit(p, RLIMIT_RTPRIO); | |
8dc3e909 ON |
4040 | |
4041 | /* can't set/change the rt policy */ | |
4042 | if (policy != p->policy && !rlim_rtprio) | |
4043 | return -EPERM; | |
4044 | ||
4045 | /* can't increase priority */ | |
d50dde5a DF |
4046 | if (attr->sched_priority > p->rt_priority && |
4047 | attr->sched_priority > rlim_rtprio) | |
8dc3e909 ON |
4048 | return -EPERM; |
4049 | } | |
c02aa73b | 4050 | |
d44753b8 JL |
4051 | /* |
4052 | * Can't set/change SCHED_DEADLINE policy at all for now | |
4053 | * (safest behavior); in the future we would like to allow | |
4054 | * unprivileged DL tasks to increase their relative deadline | |
4055 | * or reduce their runtime (both ways reducing utilization) | |
4056 | */ | |
4057 | if (dl_policy(policy)) | |
4058 | return -EPERM; | |
4059 | ||
dd41f596 | 4060 | /* |
c02aa73b DH |
4061 | * Treat SCHED_IDLE as nice 20. Only allow a switch to |
4062 | * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. | |
dd41f596 | 4063 | */ |
20f9cd2a | 4064 | if (idle_policy(p->policy) && !idle_policy(policy)) { |
d0ea0268 | 4065 | if (!can_nice(p, task_nice(p))) |
c02aa73b DH |
4066 | return -EPERM; |
4067 | } | |
5fe1d75f | 4068 | |
37e4ab3f | 4069 | /* can't change other user's priorities */ |
c69e8d9c | 4070 | if (!check_same_owner(p)) |
37e4ab3f | 4071 | return -EPERM; |
ca94c442 LP |
4072 | |
4073 | /* Normal users shall not reset the sched_reset_on_fork flag */ | |
4074 | if (p->sched_reset_on_fork && !reset_on_fork) | |
4075 | return -EPERM; | |
37e4ab3f | 4076 | } |
1da177e4 | 4077 | |
725aad24 | 4078 | if (user) { |
b0ae1981 | 4079 | retval = security_task_setscheduler(p); |
725aad24 JF |
4080 | if (retval) |
4081 | return retval; | |
4082 | } | |
4083 | ||
b29739f9 IM |
4084 | /* |
4085 | * make sure no PI-waiters arrive (or leave) while we are | |
4086 | * changing the priority of the task: | |
0122ec5b | 4087 | * |
25985edc | 4088 | * To be able to change p->policy safely, the appropriate |
1da177e4 LT |
4089 | * runqueue lock must be held. |
4090 | */ | |
eb580751 | 4091 | rq = task_rq_lock(p, &rf); |
dc61b1d6 | 4092 | |
34f971f6 PZ |
4093 | /* |
4094 | * Changing the policy of the stop threads its a very bad idea | |
4095 | */ | |
4096 | if (p == rq->stop) { | |
eb580751 | 4097 | task_rq_unlock(rq, p, &rf); |
34f971f6 PZ |
4098 | return -EINVAL; |
4099 | } | |
4100 | ||
a51e9198 | 4101 | /* |
d6b1e911 TG |
4102 | * If not changing anything there's no need to proceed further, |
4103 | * but store a possible modification of reset_on_fork. | |
a51e9198 | 4104 | */ |
d50dde5a | 4105 | if (unlikely(policy == p->policy)) { |
d0ea0268 | 4106 | if (fair_policy(policy) && attr->sched_nice != task_nice(p)) |
d50dde5a DF |
4107 | goto change; |
4108 | if (rt_policy(policy) && attr->sched_priority != p->rt_priority) | |
4109 | goto change; | |
75381608 | 4110 | if (dl_policy(policy) && dl_param_changed(p, attr)) |
aab03e05 | 4111 | goto change; |
d50dde5a | 4112 | |
d6b1e911 | 4113 | p->sched_reset_on_fork = reset_on_fork; |
eb580751 | 4114 | task_rq_unlock(rq, p, &rf); |
a51e9198 DF |
4115 | return 0; |
4116 | } | |
d50dde5a | 4117 | change: |
a51e9198 | 4118 | |
dc61b1d6 | 4119 | if (user) { |
332ac17e | 4120 | #ifdef CONFIG_RT_GROUP_SCHED |
dc61b1d6 PZ |
4121 | /* |
4122 | * Do not allow realtime tasks into groups that have no runtime | |
4123 | * assigned. | |
4124 | */ | |
4125 | if (rt_bandwidth_enabled() && rt_policy(policy) && | |
f4493771 MG |
4126 | task_group(p)->rt_bandwidth.rt_runtime == 0 && |
4127 | !task_group_is_autogroup(task_group(p))) { | |
eb580751 | 4128 | task_rq_unlock(rq, p, &rf); |
dc61b1d6 PZ |
4129 | return -EPERM; |
4130 | } | |
dc61b1d6 | 4131 | #endif |
332ac17e DF |
4132 | #ifdef CONFIG_SMP |
4133 | if (dl_bandwidth_enabled() && dl_policy(policy)) { | |
4134 | cpumask_t *span = rq->rd->span; | |
332ac17e DF |
4135 | |
4136 | /* | |
4137 | * Don't allow tasks with an affinity mask smaller than | |
4138 | * the entire root_domain to become SCHED_DEADLINE. We | |
4139 | * will also fail if there's no bandwidth available. | |
4140 | */ | |
e4099a5e PZ |
4141 | if (!cpumask_subset(span, &p->cpus_allowed) || |
4142 | rq->rd->dl_bw.bw == 0) { | |
eb580751 | 4143 | task_rq_unlock(rq, p, &rf); |
332ac17e DF |
4144 | return -EPERM; |
4145 | } | |
4146 | } | |
4147 | #endif | |
4148 | } | |
dc61b1d6 | 4149 | |
1da177e4 LT |
4150 | /* recheck policy now with rq lock held */ |
4151 | if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { | |
4152 | policy = oldpolicy = -1; | |
eb580751 | 4153 | task_rq_unlock(rq, p, &rf); |
1da177e4 LT |
4154 | goto recheck; |
4155 | } | |
332ac17e DF |
4156 | |
4157 | /* | |
4158 | * If setscheduling to SCHED_DEADLINE (or changing the parameters | |
4159 | * of a SCHED_DEADLINE task) we need to check if enough bandwidth | |
4160 | * is available. | |
4161 | */ | |
e4099a5e | 4162 | if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) { |
eb580751 | 4163 | task_rq_unlock(rq, p, &rf); |
332ac17e DF |
4164 | return -EBUSY; |
4165 | } | |
4166 | ||
c365c292 TG |
4167 | p->sched_reset_on_fork = reset_on_fork; |
4168 | oldprio = p->prio; | |
4169 | ||
dbc7f069 PZ |
4170 | if (pi) { |
4171 | /* | |
4172 | * Take priority boosted tasks into account. If the new | |
4173 | * effective priority is unchanged, we just store the new | |
4174 | * normal parameters and do not touch the scheduler class and | |
4175 | * the runqueue. This will be done when the task deboost | |
4176 | * itself. | |
4177 | */ | |
4178 | new_effective_prio = rt_mutex_get_effective_prio(p, newprio); | |
ff77e468 PZ |
4179 | if (new_effective_prio == oldprio) |
4180 | queue_flags &= ~DEQUEUE_MOVE; | |
c365c292 TG |
4181 | } |
4182 | ||
da0c1e65 | 4183 | queued = task_on_rq_queued(p); |
051a1d1a | 4184 | running = task_current(rq, p); |
da0c1e65 | 4185 | if (queued) |
ff77e468 | 4186 | dequeue_task(rq, p, queue_flags); |
0e1f3483 | 4187 | if (running) |
f3cd1c4e | 4188 | put_prev_task(rq, p); |
f6b53205 | 4189 | |
83ab0aa0 | 4190 | prev_class = p->sched_class; |
dbc7f069 | 4191 | __setscheduler(rq, p, attr, pi); |
f6b53205 | 4192 | |
0e1f3483 HS |
4193 | if (running) |
4194 | p->sched_class->set_curr_task(rq); | |
da0c1e65 | 4195 | if (queued) { |
81a44c54 TG |
4196 | /* |
4197 | * We enqueue to tail when the priority of a task is | |
4198 | * increased (user space view). | |
4199 | */ | |
ff77e468 PZ |
4200 | if (oldprio < p->prio) |
4201 | queue_flags |= ENQUEUE_HEAD; | |
1de64443 | 4202 | |
ff77e468 | 4203 | enqueue_task(rq, p, queue_flags); |
81a44c54 | 4204 | } |
cb469845 | 4205 | |
da7a735e | 4206 | check_class_changed(rq, p, prev_class, oldprio); |
4c9a4bc8 | 4207 | preempt_disable(); /* avoid rq from going away on us */ |
eb580751 | 4208 | task_rq_unlock(rq, p, &rf); |
b29739f9 | 4209 | |
dbc7f069 PZ |
4210 | if (pi) |
4211 | rt_mutex_adjust_pi(p); | |
95e02ca9 | 4212 | |
4c9a4bc8 PZ |
4213 | /* |
4214 | * Run balance callbacks after we've adjusted the PI chain. | |
4215 | */ | |
4216 | balance_callback(rq); | |
4217 | preempt_enable(); | |
95e02ca9 | 4218 | |
1da177e4 LT |
4219 | return 0; |
4220 | } | |
961ccddd | 4221 | |
7479f3c9 PZ |
4222 | static int _sched_setscheduler(struct task_struct *p, int policy, |
4223 | const struct sched_param *param, bool check) | |
4224 | { | |
4225 | struct sched_attr attr = { | |
4226 | .sched_policy = policy, | |
4227 | .sched_priority = param->sched_priority, | |
4228 | .sched_nice = PRIO_TO_NICE(p->static_prio), | |
4229 | }; | |
4230 | ||
c13db6b1 SR |
4231 | /* Fixup the legacy SCHED_RESET_ON_FORK hack. */ |
4232 | if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) { | |
7479f3c9 PZ |
4233 | attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; |
4234 | policy &= ~SCHED_RESET_ON_FORK; | |
4235 | attr.sched_policy = policy; | |
4236 | } | |
4237 | ||
dbc7f069 | 4238 | return __sched_setscheduler(p, &attr, check, true); |
7479f3c9 | 4239 | } |
961ccddd RR |
4240 | /** |
4241 | * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. | |
4242 | * @p: the task in question. | |
4243 | * @policy: new policy. | |
4244 | * @param: structure containing the new RT priority. | |
4245 | * | |
e69f6186 YB |
4246 | * Return: 0 on success. An error code otherwise. |
4247 | * | |
961ccddd RR |
4248 | * NOTE that the task may be already dead. |
4249 | */ | |
4250 | int sched_setscheduler(struct task_struct *p, int policy, | |
fe7de49f | 4251 | const struct sched_param *param) |
961ccddd | 4252 | { |
7479f3c9 | 4253 | return _sched_setscheduler(p, policy, param, true); |
961ccddd | 4254 | } |
1da177e4 LT |
4255 | EXPORT_SYMBOL_GPL(sched_setscheduler); |
4256 | ||
d50dde5a DF |
4257 | int sched_setattr(struct task_struct *p, const struct sched_attr *attr) |
4258 | { | |
dbc7f069 | 4259 | return __sched_setscheduler(p, attr, true, true); |
d50dde5a DF |
4260 | } |
4261 | EXPORT_SYMBOL_GPL(sched_setattr); | |
4262 | ||
961ccddd RR |
4263 | /** |
4264 | * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. | |
4265 | * @p: the task in question. | |
4266 | * @policy: new policy. | |
4267 | * @param: structure containing the new RT priority. | |
4268 | * | |
4269 | * Just like sched_setscheduler, only don't bother checking if the | |
4270 | * current context has permission. For example, this is needed in | |
4271 | * stop_machine(): we create temporary high priority worker threads, | |
4272 | * but our caller might not have that capability. | |
e69f6186 YB |
4273 | * |
4274 | * Return: 0 on success. An error code otherwise. | |
961ccddd RR |
4275 | */ |
4276 | int sched_setscheduler_nocheck(struct task_struct *p, int policy, | |
fe7de49f | 4277 | const struct sched_param *param) |
961ccddd | 4278 | { |
7479f3c9 | 4279 | return _sched_setscheduler(p, policy, param, false); |
961ccddd | 4280 | } |
84778472 | 4281 | EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck); |
961ccddd | 4282 | |
95cdf3b7 IM |
4283 | static int |
4284 | do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) | |
1da177e4 | 4285 | { |
1da177e4 LT |
4286 | struct sched_param lparam; |
4287 | struct task_struct *p; | |
36c8b586 | 4288 | int retval; |
1da177e4 LT |
4289 | |
4290 | if (!param || pid < 0) | |
4291 | return -EINVAL; | |
4292 | if (copy_from_user(&lparam, param, sizeof(struct sched_param))) | |
4293 | return -EFAULT; | |
5fe1d75f ON |
4294 | |
4295 | rcu_read_lock(); | |
4296 | retval = -ESRCH; | |
1da177e4 | 4297 | p = find_process_by_pid(pid); |
5fe1d75f ON |
4298 | if (p != NULL) |
4299 | retval = sched_setscheduler(p, policy, &lparam); | |
4300 | rcu_read_unlock(); | |
36c8b586 | 4301 | |
1da177e4 LT |
4302 | return retval; |
4303 | } | |
4304 | ||
d50dde5a DF |
4305 | /* |
4306 | * Mimics kernel/events/core.c perf_copy_attr(). | |
4307 | */ | |
4308 | static int sched_copy_attr(struct sched_attr __user *uattr, | |
4309 | struct sched_attr *attr) | |
4310 | { | |
4311 | u32 size; | |
4312 | int ret; | |
4313 | ||
4314 | if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0)) | |
4315 | return -EFAULT; | |
4316 | ||
4317 | /* | |
4318 | * zero the full structure, so that a short copy will be nice. | |
4319 | */ | |
4320 | memset(attr, 0, sizeof(*attr)); | |
4321 | ||
4322 | ret = get_user(size, &uattr->size); | |
4323 | if (ret) | |
4324 | return ret; | |
4325 | ||
4326 | if (size > PAGE_SIZE) /* silly large */ | |
4327 | goto err_size; | |
4328 | ||
4329 | if (!size) /* abi compat */ | |
4330 | size = SCHED_ATTR_SIZE_VER0; | |
4331 | ||
4332 | if (size < SCHED_ATTR_SIZE_VER0) | |
4333 | goto err_size; | |
4334 | ||
4335 | /* | |
4336 | * If we're handed a bigger struct than we know of, | |
4337 | * ensure all the unknown bits are 0 - i.e. new | |
4338 | * user-space does not rely on any kernel feature | |
4339 | * extensions we dont know about yet. | |
4340 | */ | |
4341 | if (size > sizeof(*attr)) { | |
4342 | unsigned char __user *addr; | |
4343 | unsigned char __user *end; | |
4344 | unsigned char val; | |
4345 | ||
4346 | addr = (void __user *)uattr + sizeof(*attr); | |
4347 | end = (void __user *)uattr + size; | |
4348 | ||
4349 | for (; addr < end; addr++) { | |
4350 | ret = get_user(val, addr); | |
4351 | if (ret) | |
4352 | return ret; | |
4353 | if (val) | |
4354 | goto err_size; | |
4355 | } | |
4356 | size = sizeof(*attr); | |
4357 | } | |
4358 | ||
4359 | ret = copy_from_user(attr, uattr, size); | |
4360 | if (ret) | |
4361 | return -EFAULT; | |
4362 | ||
4363 | /* | |
4364 | * XXX: do we want to be lenient like existing syscalls; or do we want | |
4365 | * to be strict and return an error on out-of-bounds values? | |
4366 | */ | |
75e45d51 | 4367 | attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE); |
d50dde5a | 4368 | |
e78c7bca | 4369 | return 0; |
d50dde5a DF |
4370 | |
4371 | err_size: | |
4372 | put_user(sizeof(*attr), &uattr->size); | |
e78c7bca | 4373 | return -E2BIG; |
d50dde5a DF |
4374 | } |
4375 | ||
1da177e4 LT |
4376 | /** |
4377 | * sys_sched_setscheduler - set/change the scheduler policy and RT priority | |
4378 | * @pid: the pid in question. | |
4379 | * @policy: new policy. | |
4380 | * @param: structure containing the new RT priority. | |
e69f6186 YB |
4381 | * |
4382 | * Return: 0 on success. An error code otherwise. | |
1da177e4 | 4383 | */ |
5add95d4 HC |
4384 | SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, |
4385 | struct sched_param __user *, param) | |
1da177e4 | 4386 | { |
c21761f1 JB |
4387 | /* negative values for policy are not valid */ |
4388 | if (policy < 0) | |
4389 | return -EINVAL; | |
4390 | ||
1da177e4 LT |
4391 | return do_sched_setscheduler(pid, policy, param); |
4392 | } | |
4393 | ||
4394 | /** | |
4395 | * sys_sched_setparam - set/change the RT priority of a thread | |
4396 | * @pid: the pid in question. | |
4397 | * @param: structure containing the new RT priority. | |
e69f6186 YB |
4398 | * |
4399 | * Return: 0 on success. An error code otherwise. | |
1da177e4 | 4400 | */ |
5add95d4 | 4401 | SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) |
1da177e4 | 4402 | { |
c13db6b1 | 4403 | return do_sched_setscheduler(pid, SETPARAM_POLICY, param); |
1da177e4 LT |
4404 | } |
4405 | ||
d50dde5a DF |
4406 | /** |
4407 | * sys_sched_setattr - same as above, but with extended sched_attr | |
4408 | * @pid: the pid in question. | |
5778fccf | 4409 | * @uattr: structure containing the extended parameters. |
db66d756 | 4410 | * @flags: for future extension. |
d50dde5a | 4411 | */ |
6d35ab48 PZ |
4412 | SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr, |
4413 | unsigned int, flags) | |
d50dde5a DF |
4414 | { |
4415 | struct sched_attr attr; | |
4416 | struct task_struct *p; | |
4417 | int retval; | |
4418 | ||
6d35ab48 | 4419 | if (!uattr || pid < 0 || flags) |
d50dde5a DF |
4420 | return -EINVAL; |
4421 | ||
143cf23d MK |
4422 | retval = sched_copy_attr(uattr, &attr); |
4423 | if (retval) | |
4424 | return retval; | |
d50dde5a | 4425 | |
b14ed2c2 | 4426 | if ((int)attr.sched_policy < 0) |
dbdb2275 | 4427 | return -EINVAL; |
d50dde5a DF |
4428 | |
4429 | rcu_read_lock(); | |
4430 | retval = -ESRCH; | |
4431 | p = find_process_by_pid(pid); | |
4432 | if (p != NULL) | |
4433 | retval = sched_setattr(p, &attr); | |
4434 | rcu_read_unlock(); | |
4435 | ||
4436 | return retval; | |
4437 | } | |
4438 | ||
1da177e4 LT |
4439 | /** |
4440 | * sys_sched_getscheduler - get the policy (scheduling class) of a thread | |
4441 | * @pid: the pid in question. | |
e69f6186 YB |
4442 | * |
4443 | * Return: On success, the policy of the thread. Otherwise, a negative error | |
4444 | * code. | |
1da177e4 | 4445 | */ |
5add95d4 | 4446 | SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) |
1da177e4 | 4447 | { |
36c8b586 | 4448 | struct task_struct *p; |
3a5c359a | 4449 | int retval; |
1da177e4 LT |
4450 | |
4451 | if (pid < 0) | |
3a5c359a | 4452 | return -EINVAL; |
1da177e4 LT |
4453 | |
4454 | retval = -ESRCH; | |
5fe85be0 | 4455 | rcu_read_lock(); |
1da177e4 LT |
4456 | p = find_process_by_pid(pid); |
4457 | if (p) { | |
4458 | retval = security_task_getscheduler(p); | |
4459 | if (!retval) | |
ca94c442 LP |
4460 | retval = p->policy |
4461 | | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); | |
1da177e4 | 4462 | } |
5fe85be0 | 4463 | rcu_read_unlock(); |
1da177e4 LT |
4464 | return retval; |
4465 | } | |
4466 | ||
4467 | /** | |
ca94c442 | 4468 | * sys_sched_getparam - get the RT priority of a thread |
1da177e4 LT |
4469 | * @pid: the pid in question. |
4470 | * @param: structure containing the RT priority. | |
e69f6186 YB |
4471 | * |
4472 | * Return: On success, 0 and the RT priority is in @param. Otherwise, an error | |
4473 | * code. | |
1da177e4 | 4474 | */ |
5add95d4 | 4475 | SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) |
1da177e4 | 4476 | { |
ce5f7f82 | 4477 | struct sched_param lp = { .sched_priority = 0 }; |
36c8b586 | 4478 | struct task_struct *p; |
3a5c359a | 4479 | int retval; |
1da177e4 LT |
4480 | |
4481 | if (!param || pid < 0) | |
3a5c359a | 4482 | return -EINVAL; |
1da177e4 | 4483 | |
5fe85be0 | 4484 | rcu_read_lock(); |
1da177e4 LT |
4485 | p = find_process_by_pid(pid); |
4486 | retval = -ESRCH; | |
4487 | if (!p) | |
4488 | goto out_unlock; | |
4489 | ||
4490 | retval = security_task_getscheduler(p); | |
4491 | if (retval) | |
4492 | goto out_unlock; | |
4493 | ||
ce5f7f82 PZ |
4494 | if (task_has_rt_policy(p)) |
4495 | lp.sched_priority = p->rt_priority; | |
5fe85be0 | 4496 | rcu_read_unlock(); |
1da177e4 LT |
4497 | |
4498 | /* | |
4499 | * This one might sleep, we cannot do it with a spinlock held ... | |
4500 | */ | |
4501 | retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; | |
4502 | ||
1da177e4 LT |
4503 | return retval; |
4504 | ||
4505 | out_unlock: | |
5fe85be0 | 4506 | rcu_read_unlock(); |
1da177e4 LT |
4507 | return retval; |
4508 | } | |
4509 | ||
d50dde5a DF |
4510 | static int sched_read_attr(struct sched_attr __user *uattr, |
4511 | struct sched_attr *attr, | |
4512 | unsigned int usize) | |
4513 | { | |
4514 | int ret; | |
4515 | ||
4516 | if (!access_ok(VERIFY_WRITE, uattr, usize)) | |
4517 | return -EFAULT; | |
4518 | ||
4519 | /* | |
4520 | * If we're handed a smaller struct than we know of, | |
4521 | * ensure all the unknown bits are 0 - i.e. old | |
4522 | * user-space does not get uncomplete information. | |
4523 | */ | |
4524 | if (usize < sizeof(*attr)) { | |
4525 | unsigned char *addr; | |
4526 | unsigned char *end; | |
4527 | ||
4528 | addr = (void *)attr + usize; | |
4529 | end = (void *)attr + sizeof(*attr); | |
4530 | ||
4531 | for (; addr < end; addr++) { | |
4532 | if (*addr) | |
22400674 | 4533 | return -EFBIG; |
d50dde5a DF |
4534 | } |
4535 | ||
4536 | attr->size = usize; | |
4537 | } | |
4538 | ||
4efbc454 | 4539 | ret = copy_to_user(uattr, attr, attr->size); |
d50dde5a DF |
4540 | if (ret) |
4541 | return -EFAULT; | |
4542 | ||
22400674 | 4543 | return 0; |
d50dde5a DF |
4544 | } |
4545 | ||
4546 | /** | |
aab03e05 | 4547 | * sys_sched_getattr - similar to sched_getparam, but with sched_attr |
d50dde5a | 4548 | * @pid: the pid in question. |
5778fccf | 4549 | * @uattr: structure containing the extended parameters. |
d50dde5a | 4550 | * @size: sizeof(attr) for fwd/bwd comp. |
db66d756 | 4551 | * @flags: for future extension. |
d50dde5a | 4552 | */ |
6d35ab48 PZ |
4553 | SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, |
4554 | unsigned int, size, unsigned int, flags) | |
d50dde5a DF |
4555 | { |
4556 | struct sched_attr attr = { | |
4557 | .size = sizeof(struct sched_attr), | |
4558 | }; | |
4559 | struct task_struct *p; | |
4560 | int retval; | |
4561 | ||
4562 | if (!uattr || pid < 0 || size > PAGE_SIZE || | |
6d35ab48 | 4563 | size < SCHED_ATTR_SIZE_VER0 || flags) |
d50dde5a DF |
4564 | return -EINVAL; |
4565 | ||
4566 | rcu_read_lock(); | |
4567 | p = find_process_by_pid(pid); | |
4568 | retval = -ESRCH; | |
4569 | if (!p) | |
4570 | goto out_unlock; | |
4571 | ||
4572 | retval = security_task_getscheduler(p); | |
4573 | if (retval) | |
4574 | goto out_unlock; | |
4575 | ||
4576 | attr.sched_policy = p->policy; | |
7479f3c9 PZ |
4577 | if (p->sched_reset_on_fork) |
4578 | attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; | |
aab03e05 DF |
4579 | if (task_has_dl_policy(p)) |
4580 | __getparam_dl(p, &attr); | |
4581 | else if (task_has_rt_policy(p)) | |
d50dde5a DF |
4582 | attr.sched_priority = p->rt_priority; |
4583 | else | |
d0ea0268 | 4584 | attr.sched_nice = task_nice(p); |
d50dde5a DF |
4585 | |
4586 | rcu_read_unlock(); | |
4587 | ||
4588 | retval = sched_read_attr(uattr, &attr, size); | |
4589 | return retval; | |
4590 | ||
4591 | out_unlock: | |
4592 | rcu_read_unlock(); | |
4593 | return retval; | |
4594 | } | |
4595 | ||
96f874e2 | 4596 | long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) |
1da177e4 | 4597 | { |
5a16f3d3 | 4598 | cpumask_var_t cpus_allowed, new_mask; |
36c8b586 IM |
4599 | struct task_struct *p; |
4600 | int retval; | |
1da177e4 | 4601 | |
23f5d142 | 4602 | rcu_read_lock(); |
1da177e4 LT |
4603 | |
4604 | p = find_process_by_pid(pid); | |
4605 | if (!p) { | |
23f5d142 | 4606 | rcu_read_unlock(); |
1da177e4 LT |
4607 | return -ESRCH; |
4608 | } | |
4609 | ||
23f5d142 | 4610 | /* Prevent p going away */ |
1da177e4 | 4611 | get_task_struct(p); |
23f5d142 | 4612 | rcu_read_unlock(); |
1da177e4 | 4613 | |
14a40ffc TH |
4614 | if (p->flags & PF_NO_SETAFFINITY) { |
4615 | retval = -EINVAL; | |
4616 | goto out_put_task; | |
4617 | } | |
5a16f3d3 RR |
4618 | if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { |
4619 | retval = -ENOMEM; | |
4620 | goto out_put_task; | |
4621 | } | |
4622 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { | |
4623 | retval = -ENOMEM; | |
4624 | goto out_free_cpus_allowed; | |
4625 | } | |
1da177e4 | 4626 | retval = -EPERM; |
4c44aaaf EB |
4627 | if (!check_same_owner(p)) { |
4628 | rcu_read_lock(); | |
4629 | if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { | |
4630 | rcu_read_unlock(); | |
16303ab2 | 4631 | goto out_free_new_mask; |
4c44aaaf EB |
4632 | } |
4633 | rcu_read_unlock(); | |
4634 | } | |
1da177e4 | 4635 | |
b0ae1981 | 4636 | retval = security_task_setscheduler(p); |
e7834f8f | 4637 | if (retval) |
16303ab2 | 4638 | goto out_free_new_mask; |
e7834f8f | 4639 | |
e4099a5e PZ |
4640 | |
4641 | cpuset_cpus_allowed(p, cpus_allowed); | |
4642 | cpumask_and(new_mask, in_mask, cpus_allowed); | |
4643 | ||
332ac17e DF |
4644 | /* |
4645 | * Since bandwidth control happens on root_domain basis, | |
4646 | * if admission test is enabled, we only admit -deadline | |
4647 | * tasks allowed to run on all the CPUs in the task's | |
4648 | * root_domain. | |
4649 | */ | |
4650 | #ifdef CONFIG_SMP | |
f1e3a093 KT |
4651 | if (task_has_dl_policy(p) && dl_bandwidth_enabled()) { |
4652 | rcu_read_lock(); | |
4653 | if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) { | |
332ac17e | 4654 | retval = -EBUSY; |
f1e3a093 | 4655 | rcu_read_unlock(); |
16303ab2 | 4656 | goto out_free_new_mask; |
332ac17e | 4657 | } |
f1e3a093 | 4658 | rcu_read_unlock(); |
332ac17e DF |
4659 | } |
4660 | #endif | |
49246274 | 4661 | again: |
25834c73 | 4662 | retval = __set_cpus_allowed_ptr(p, new_mask, true); |
1da177e4 | 4663 | |
8707d8b8 | 4664 | if (!retval) { |
5a16f3d3 RR |
4665 | cpuset_cpus_allowed(p, cpus_allowed); |
4666 | if (!cpumask_subset(new_mask, cpus_allowed)) { | |
8707d8b8 PM |
4667 | /* |
4668 | * We must have raced with a concurrent cpuset | |
4669 | * update. Just reset the cpus_allowed to the | |
4670 | * cpuset's cpus_allowed | |
4671 | */ | |
5a16f3d3 | 4672 | cpumask_copy(new_mask, cpus_allowed); |
8707d8b8 PM |
4673 | goto again; |
4674 | } | |
4675 | } | |
16303ab2 | 4676 | out_free_new_mask: |
5a16f3d3 RR |
4677 | free_cpumask_var(new_mask); |
4678 | out_free_cpus_allowed: | |
4679 | free_cpumask_var(cpus_allowed); | |
4680 | out_put_task: | |
1da177e4 | 4681 | put_task_struct(p); |
1da177e4 LT |
4682 | return retval; |
4683 | } | |
4684 | ||
4685 | static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, | |
96f874e2 | 4686 | struct cpumask *new_mask) |
1da177e4 | 4687 | { |
96f874e2 RR |
4688 | if (len < cpumask_size()) |
4689 | cpumask_clear(new_mask); | |
4690 | else if (len > cpumask_size()) | |
4691 | len = cpumask_size(); | |
4692 | ||
1da177e4 LT |
4693 | return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; |
4694 | } | |
4695 | ||
4696 | /** | |
4697 | * sys_sched_setaffinity - set the cpu affinity of a process | |
4698 | * @pid: pid of the process | |
4699 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
4700 | * @user_mask_ptr: user-space pointer to the new cpu mask | |
e69f6186 YB |
4701 | * |
4702 | * Return: 0 on success. An error code otherwise. | |
1da177e4 | 4703 | */ |
5add95d4 HC |
4704 | SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, |
4705 | unsigned long __user *, user_mask_ptr) | |
1da177e4 | 4706 | { |
5a16f3d3 | 4707 | cpumask_var_t new_mask; |
1da177e4 LT |
4708 | int retval; |
4709 | ||
5a16f3d3 RR |
4710 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) |
4711 | return -ENOMEM; | |
1da177e4 | 4712 | |
5a16f3d3 RR |
4713 | retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); |
4714 | if (retval == 0) | |
4715 | retval = sched_setaffinity(pid, new_mask); | |
4716 | free_cpumask_var(new_mask); | |
4717 | return retval; | |
1da177e4 LT |
4718 | } |
4719 | ||
96f874e2 | 4720 | long sched_getaffinity(pid_t pid, struct cpumask *mask) |
1da177e4 | 4721 | { |
36c8b586 | 4722 | struct task_struct *p; |
31605683 | 4723 | unsigned long flags; |
1da177e4 | 4724 | int retval; |
1da177e4 | 4725 | |
23f5d142 | 4726 | rcu_read_lock(); |
1da177e4 LT |
4727 | |
4728 | retval = -ESRCH; | |
4729 | p = find_process_by_pid(pid); | |
4730 | if (!p) | |
4731 | goto out_unlock; | |
4732 | ||
e7834f8f DQ |
4733 | retval = security_task_getscheduler(p); |
4734 | if (retval) | |
4735 | goto out_unlock; | |
4736 | ||
013fdb80 | 4737 | raw_spin_lock_irqsave(&p->pi_lock, flags); |
6acce3ef | 4738 | cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); |
013fdb80 | 4739 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); |
1da177e4 LT |
4740 | |
4741 | out_unlock: | |
23f5d142 | 4742 | rcu_read_unlock(); |
1da177e4 | 4743 | |
9531b62f | 4744 | return retval; |
1da177e4 LT |
4745 | } |
4746 | ||
4747 | /** | |
4748 | * sys_sched_getaffinity - get the cpu affinity of a process | |
4749 | * @pid: pid of the process | |
4750 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
4751 | * @user_mask_ptr: user-space pointer to hold the current cpu mask | |
e69f6186 YB |
4752 | * |
4753 | * Return: 0 on success. An error code otherwise. | |
1da177e4 | 4754 | */ |
5add95d4 HC |
4755 | SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, |
4756 | unsigned long __user *, user_mask_ptr) | |
1da177e4 LT |
4757 | { |
4758 | int ret; | |
f17c8607 | 4759 | cpumask_var_t mask; |
1da177e4 | 4760 | |
84fba5ec | 4761 | if ((len * BITS_PER_BYTE) < nr_cpu_ids) |
cd3d8031 KM |
4762 | return -EINVAL; |
4763 | if (len & (sizeof(unsigned long)-1)) | |
1da177e4 LT |
4764 | return -EINVAL; |
4765 | ||
f17c8607 RR |
4766 | if (!alloc_cpumask_var(&mask, GFP_KERNEL)) |
4767 | return -ENOMEM; | |
1da177e4 | 4768 | |
f17c8607 RR |
4769 | ret = sched_getaffinity(pid, mask); |
4770 | if (ret == 0) { | |
8bc037fb | 4771 | size_t retlen = min_t(size_t, len, cpumask_size()); |
cd3d8031 KM |
4772 | |
4773 | if (copy_to_user(user_mask_ptr, mask, retlen)) | |
f17c8607 RR |
4774 | ret = -EFAULT; |
4775 | else | |
cd3d8031 | 4776 | ret = retlen; |
f17c8607 RR |
4777 | } |
4778 | free_cpumask_var(mask); | |
1da177e4 | 4779 | |
f17c8607 | 4780 | return ret; |
1da177e4 LT |
4781 | } |
4782 | ||
4783 | /** | |
4784 | * sys_sched_yield - yield the current processor to other threads. | |
4785 | * | |
dd41f596 IM |
4786 | * This function yields the current CPU to other tasks. If there are no |
4787 | * other threads running on this CPU then this function will return. | |
e69f6186 YB |
4788 | * |
4789 | * Return: 0. | |
1da177e4 | 4790 | */ |
5add95d4 | 4791 | SYSCALL_DEFINE0(sched_yield) |
1da177e4 | 4792 | { |
70b97a7f | 4793 | struct rq *rq = this_rq_lock(); |
1da177e4 | 4794 | |
2d72376b | 4795 | schedstat_inc(rq, yld_count); |
4530d7ab | 4796 | current->sched_class->yield_task(rq); |
1da177e4 LT |
4797 | |
4798 | /* | |
4799 | * Since we are going to call schedule() anyway, there's | |
4800 | * no need to preempt or enable interrupts: | |
4801 | */ | |
4802 | __release(rq->lock); | |
8a25d5de | 4803 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); |
9828ea9d | 4804 | do_raw_spin_unlock(&rq->lock); |
ba74c144 | 4805 | sched_preempt_enable_no_resched(); |
1da177e4 LT |
4806 | |
4807 | schedule(); | |
4808 | ||
4809 | return 0; | |
4810 | } | |
4811 | ||
02b67cc3 | 4812 | int __sched _cond_resched(void) |
1da177e4 | 4813 | { |
fe32d3cd | 4814 | if (should_resched(0)) { |
a18b5d01 | 4815 | preempt_schedule_common(); |
1da177e4 LT |
4816 | return 1; |
4817 | } | |
4818 | return 0; | |
4819 | } | |
02b67cc3 | 4820 | EXPORT_SYMBOL(_cond_resched); |
1da177e4 LT |
4821 | |
4822 | /* | |
613afbf8 | 4823 | * __cond_resched_lock() - if a reschedule is pending, drop the given lock, |
1da177e4 LT |
4824 | * call schedule, and on return reacquire the lock. |
4825 | * | |
41a2d6cf | 4826 | * This works OK both with and without CONFIG_PREEMPT. We do strange low-level |
1da177e4 LT |
4827 | * operations here to prevent schedule() from being called twice (once via |
4828 | * spin_unlock(), once by hand). | |
4829 | */ | |
613afbf8 | 4830 | int __cond_resched_lock(spinlock_t *lock) |
1da177e4 | 4831 | { |
fe32d3cd | 4832 | int resched = should_resched(PREEMPT_LOCK_OFFSET); |
6df3cecb JK |
4833 | int ret = 0; |
4834 | ||
f607c668 PZ |
4835 | lockdep_assert_held(lock); |
4836 | ||
4a81e832 | 4837 | if (spin_needbreak(lock) || resched) { |
1da177e4 | 4838 | spin_unlock(lock); |
d86ee480 | 4839 | if (resched) |
a18b5d01 | 4840 | preempt_schedule_common(); |
95c354fe NP |
4841 | else |
4842 | cpu_relax(); | |
6df3cecb | 4843 | ret = 1; |
1da177e4 | 4844 | spin_lock(lock); |
1da177e4 | 4845 | } |
6df3cecb | 4846 | return ret; |
1da177e4 | 4847 | } |
613afbf8 | 4848 | EXPORT_SYMBOL(__cond_resched_lock); |
1da177e4 | 4849 | |
613afbf8 | 4850 | int __sched __cond_resched_softirq(void) |
1da177e4 LT |
4851 | { |
4852 | BUG_ON(!in_softirq()); | |
4853 | ||
fe32d3cd | 4854 | if (should_resched(SOFTIRQ_DISABLE_OFFSET)) { |
98d82567 | 4855 | local_bh_enable(); |
a18b5d01 | 4856 | preempt_schedule_common(); |
1da177e4 LT |
4857 | local_bh_disable(); |
4858 | return 1; | |
4859 | } | |
4860 | return 0; | |
4861 | } | |
613afbf8 | 4862 | EXPORT_SYMBOL(__cond_resched_softirq); |
1da177e4 | 4863 | |
1da177e4 LT |
4864 | /** |
4865 | * yield - yield the current processor to other threads. | |
4866 | * | |
8e3fabfd PZ |
4867 | * Do not ever use this function, there's a 99% chance you're doing it wrong. |
4868 | * | |
4869 | * The scheduler is at all times free to pick the calling task as the most | |
4870 | * eligible task to run, if removing the yield() call from your code breaks | |
4871 | * it, its already broken. | |
4872 | * | |
4873 | * Typical broken usage is: | |
4874 | * | |
4875 | * while (!event) | |
4876 | * yield(); | |
4877 | * | |
4878 | * where one assumes that yield() will let 'the other' process run that will | |
4879 | * make event true. If the current task is a SCHED_FIFO task that will never | |
4880 | * happen. Never use yield() as a progress guarantee!! | |
4881 | * | |
4882 | * If you want to use yield() to wait for something, use wait_event(). | |
4883 | * If you want to use yield() to be 'nice' for others, use cond_resched(). | |
4884 | * If you still want to use yield(), do not! | |
1da177e4 LT |
4885 | */ |
4886 | void __sched yield(void) | |
4887 | { | |
4888 | set_current_state(TASK_RUNNING); | |
4889 | sys_sched_yield(); | |
4890 | } | |
1da177e4 LT |
4891 | EXPORT_SYMBOL(yield); |
4892 | ||
d95f4122 MG |
4893 | /** |
4894 | * yield_to - yield the current processor to another thread in | |
4895 | * your thread group, or accelerate that thread toward the | |
4896 | * processor it's on. | |
16addf95 RD |
4897 | * @p: target task |
4898 | * @preempt: whether task preemption is allowed or not | |
d95f4122 MG |
4899 | * |
4900 | * It's the caller's job to ensure that the target task struct | |
4901 | * can't go away on us before we can do any checks. | |
4902 | * | |
e69f6186 | 4903 | * Return: |
7b270f60 PZ |
4904 | * true (>0) if we indeed boosted the target task. |
4905 | * false (0) if we failed to boost the target. | |
4906 | * -ESRCH if there's no task to yield to. | |
d95f4122 | 4907 | */ |
fa93384f | 4908 | int __sched yield_to(struct task_struct *p, bool preempt) |
d95f4122 MG |
4909 | { |
4910 | struct task_struct *curr = current; | |
4911 | struct rq *rq, *p_rq; | |
4912 | unsigned long flags; | |
c3c18640 | 4913 | int yielded = 0; |
d95f4122 MG |
4914 | |
4915 | local_irq_save(flags); | |
4916 | rq = this_rq(); | |
4917 | ||
4918 | again: | |
4919 | p_rq = task_rq(p); | |
7b270f60 PZ |
4920 | /* |
4921 | * If we're the only runnable task on the rq and target rq also | |
4922 | * has only one task, there's absolutely no point in yielding. | |
4923 | */ | |
4924 | if (rq->nr_running == 1 && p_rq->nr_running == 1) { | |
4925 | yielded = -ESRCH; | |
4926 | goto out_irq; | |
4927 | } | |
4928 | ||
d95f4122 | 4929 | double_rq_lock(rq, p_rq); |
39e24d8f | 4930 | if (task_rq(p) != p_rq) { |
d95f4122 MG |
4931 | double_rq_unlock(rq, p_rq); |
4932 | goto again; | |
4933 | } | |
4934 | ||
4935 | if (!curr->sched_class->yield_to_task) | |
7b270f60 | 4936 | goto out_unlock; |
d95f4122 MG |
4937 | |
4938 | if (curr->sched_class != p->sched_class) | |
7b270f60 | 4939 | goto out_unlock; |
d95f4122 MG |
4940 | |
4941 | if (task_running(p_rq, p) || p->state) | |
7b270f60 | 4942 | goto out_unlock; |
d95f4122 MG |
4943 | |
4944 | yielded = curr->sched_class->yield_to_task(rq, p, preempt); | |
6d1cafd8 | 4945 | if (yielded) { |
d95f4122 | 4946 | schedstat_inc(rq, yld_count); |
6d1cafd8 VP |
4947 | /* |
4948 | * Make p's CPU reschedule; pick_next_entity takes care of | |
4949 | * fairness. | |
4950 | */ | |
4951 | if (preempt && rq != p_rq) | |
8875125e | 4952 | resched_curr(p_rq); |
6d1cafd8 | 4953 | } |
d95f4122 | 4954 | |
7b270f60 | 4955 | out_unlock: |
d95f4122 | 4956 | double_rq_unlock(rq, p_rq); |
7b270f60 | 4957 | out_irq: |
d95f4122 MG |
4958 | local_irq_restore(flags); |
4959 | ||
7b270f60 | 4960 | if (yielded > 0) |
d95f4122 MG |
4961 | schedule(); |
4962 | ||
4963 | return yielded; | |
4964 | } | |
4965 | EXPORT_SYMBOL_GPL(yield_to); | |
4966 | ||
1da177e4 | 4967 | /* |
41a2d6cf | 4968 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so |
1da177e4 | 4969 | * that process accounting knows that this is a task in IO wait state. |
1da177e4 | 4970 | */ |
1da177e4 LT |
4971 | long __sched io_schedule_timeout(long timeout) |
4972 | { | |
9cff8ade N |
4973 | int old_iowait = current->in_iowait; |
4974 | struct rq *rq; | |
1da177e4 LT |
4975 | long ret; |
4976 | ||
9cff8ade | 4977 | current->in_iowait = 1; |
10d784ea | 4978 | blk_schedule_flush_plug(current); |
9cff8ade | 4979 | |
0ff92245 | 4980 | delayacct_blkio_start(); |
9cff8ade | 4981 | rq = raw_rq(); |
1da177e4 LT |
4982 | atomic_inc(&rq->nr_iowait); |
4983 | ret = schedule_timeout(timeout); | |
9cff8ade | 4984 | current->in_iowait = old_iowait; |
1da177e4 | 4985 | atomic_dec(&rq->nr_iowait); |
0ff92245 | 4986 | delayacct_blkio_end(); |
9cff8ade | 4987 | |
1da177e4 LT |
4988 | return ret; |
4989 | } | |
9cff8ade | 4990 | EXPORT_SYMBOL(io_schedule_timeout); |
1da177e4 LT |
4991 | |
4992 | /** | |
4993 | * sys_sched_get_priority_max - return maximum RT priority. | |
4994 | * @policy: scheduling class. | |
4995 | * | |
e69f6186 YB |
4996 | * Return: On success, this syscall returns the maximum |
4997 | * rt_priority that can be used by a given scheduling class. | |
4998 | * On failure, a negative error code is returned. | |
1da177e4 | 4999 | */ |
5add95d4 | 5000 | SYSCALL_DEFINE1(sched_get_priority_max, int, policy) |
1da177e4 LT |
5001 | { |
5002 | int ret = -EINVAL; | |
5003 | ||
5004 | switch (policy) { | |
5005 | case SCHED_FIFO: | |
5006 | case SCHED_RR: | |
5007 | ret = MAX_USER_RT_PRIO-1; | |
5008 | break; | |
aab03e05 | 5009 | case SCHED_DEADLINE: |
1da177e4 | 5010 | case SCHED_NORMAL: |
b0a9499c | 5011 | case SCHED_BATCH: |
dd41f596 | 5012 | case SCHED_IDLE: |
1da177e4 LT |
5013 | ret = 0; |
5014 | break; | |
5015 | } | |
5016 | return ret; | |
5017 | } | |
5018 | ||
5019 | /** | |
5020 | * sys_sched_get_priority_min - return minimum RT priority. | |
5021 | * @policy: scheduling class. | |
5022 | * | |
e69f6186 YB |
5023 | * Return: On success, this syscall returns the minimum |
5024 | * rt_priority that can be used by a given scheduling class. | |
5025 | * On failure, a negative error code is returned. | |
1da177e4 | 5026 | */ |
5add95d4 | 5027 | SYSCALL_DEFINE1(sched_get_priority_min, int, policy) |
1da177e4 LT |
5028 | { |
5029 | int ret = -EINVAL; | |
5030 | ||
5031 | switch (policy) { | |
5032 | case SCHED_FIFO: | |
5033 | case SCHED_RR: | |
5034 | ret = 1; | |
5035 | break; | |
aab03e05 | 5036 | case SCHED_DEADLINE: |
1da177e4 | 5037 | case SCHED_NORMAL: |
b0a9499c | 5038 | case SCHED_BATCH: |
dd41f596 | 5039 | case SCHED_IDLE: |
1da177e4 LT |
5040 | ret = 0; |
5041 | } | |
5042 | return ret; | |
5043 | } | |
5044 | ||
5045 | /** | |
5046 | * sys_sched_rr_get_interval - return the default timeslice of a process. | |
5047 | * @pid: pid of the process. | |
5048 | * @interval: userspace pointer to the timeslice value. | |
5049 | * | |
5050 | * this syscall writes the default timeslice value of a given process | |
5051 | * into the user-space timespec buffer. A value of '0' means infinity. | |
e69f6186 YB |
5052 | * |
5053 | * Return: On success, 0 and the timeslice is in @interval. Otherwise, | |
5054 | * an error code. | |
1da177e4 | 5055 | */ |
17da2bd9 | 5056 | SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, |
754fe8d2 | 5057 | struct timespec __user *, interval) |
1da177e4 | 5058 | { |
36c8b586 | 5059 | struct task_struct *p; |
a4ec24b4 | 5060 | unsigned int time_slice; |
eb580751 PZ |
5061 | struct rq_flags rf; |
5062 | struct timespec t; | |
dba091b9 | 5063 | struct rq *rq; |
3a5c359a | 5064 | int retval; |
1da177e4 LT |
5065 | |
5066 | if (pid < 0) | |
3a5c359a | 5067 | return -EINVAL; |
1da177e4 LT |
5068 | |
5069 | retval = -ESRCH; | |
1a551ae7 | 5070 | rcu_read_lock(); |
1da177e4 LT |
5071 | p = find_process_by_pid(pid); |
5072 | if (!p) | |
5073 | goto out_unlock; | |
5074 | ||
5075 | retval = security_task_getscheduler(p); | |
5076 | if (retval) | |
5077 | goto out_unlock; | |
5078 | ||
eb580751 | 5079 | rq = task_rq_lock(p, &rf); |
a57beec5 PZ |
5080 | time_slice = 0; |
5081 | if (p->sched_class->get_rr_interval) | |
5082 | time_slice = p->sched_class->get_rr_interval(rq, p); | |
eb580751 | 5083 | task_rq_unlock(rq, p, &rf); |
a4ec24b4 | 5084 | |
1a551ae7 | 5085 | rcu_read_unlock(); |
a4ec24b4 | 5086 | jiffies_to_timespec(time_slice, &t); |
1da177e4 | 5087 | retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; |
1da177e4 | 5088 | return retval; |
3a5c359a | 5089 | |
1da177e4 | 5090 | out_unlock: |
1a551ae7 | 5091 | rcu_read_unlock(); |
1da177e4 LT |
5092 | return retval; |
5093 | } | |
5094 | ||
7c731e0a | 5095 | static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; |
36c8b586 | 5096 | |
82a1fcb9 | 5097 | void sched_show_task(struct task_struct *p) |
1da177e4 | 5098 | { |
1da177e4 | 5099 | unsigned long free = 0; |
4e79752c | 5100 | int ppid; |
1f8a7633 | 5101 | unsigned long state = p->state; |
1da177e4 | 5102 | |
1f8a7633 TH |
5103 | if (state) |
5104 | state = __ffs(state) + 1; | |
28d0686c | 5105 | printk(KERN_INFO "%-15.15s %c", p->comm, |
2ed6e34f | 5106 | state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); |
4bd77321 | 5107 | #if BITS_PER_LONG == 32 |
1da177e4 | 5108 | if (state == TASK_RUNNING) |
3df0fc5b | 5109 | printk(KERN_CONT " running "); |
1da177e4 | 5110 | else |
3df0fc5b | 5111 | printk(KERN_CONT " %08lx ", thread_saved_pc(p)); |
1da177e4 LT |
5112 | #else |
5113 | if (state == TASK_RUNNING) | |
3df0fc5b | 5114 | printk(KERN_CONT " running task "); |
1da177e4 | 5115 | else |
3df0fc5b | 5116 | printk(KERN_CONT " %016lx ", thread_saved_pc(p)); |
1da177e4 LT |
5117 | #endif |
5118 | #ifdef CONFIG_DEBUG_STACK_USAGE | |
7c9f8861 | 5119 | free = stack_not_used(p); |
1da177e4 | 5120 | #endif |
a90e984c | 5121 | ppid = 0; |
4e79752c | 5122 | rcu_read_lock(); |
a90e984c ON |
5123 | if (pid_alive(p)) |
5124 | ppid = task_pid_nr(rcu_dereference(p->real_parent)); | |
4e79752c | 5125 | rcu_read_unlock(); |
3df0fc5b | 5126 | printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, |
4e79752c | 5127 | task_pid_nr(p), ppid, |
aa47b7e0 | 5128 | (unsigned long)task_thread_info(p)->flags); |
1da177e4 | 5129 | |
3d1cb205 | 5130 | print_worker_info(KERN_INFO, p); |
5fb5e6de | 5131 | show_stack(p, NULL); |
1da177e4 LT |
5132 | } |
5133 | ||
e59e2ae2 | 5134 | void show_state_filter(unsigned long state_filter) |
1da177e4 | 5135 | { |
36c8b586 | 5136 | struct task_struct *g, *p; |
1da177e4 | 5137 | |
4bd77321 | 5138 | #if BITS_PER_LONG == 32 |
3df0fc5b PZ |
5139 | printk(KERN_INFO |
5140 | " task PC stack pid father\n"); | |
1da177e4 | 5141 | #else |
3df0fc5b PZ |
5142 | printk(KERN_INFO |
5143 | " task PC stack pid father\n"); | |
1da177e4 | 5144 | #endif |
510f5acc | 5145 | rcu_read_lock(); |
5d07f420 | 5146 | for_each_process_thread(g, p) { |
1da177e4 LT |
5147 | /* |
5148 | * reset the NMI-timeout, listing all files on a slow | |
25985edc | 5149 | * console might take a lot of time: |
57675cb9 AR |
5150 | * Also, reset softlockup watchdogs on all CPUs, because |
5151 | * another CPU might be blocked waiting for us to process | |
5152 | * an IPI. | |
1da177e4 LT |
5153 | */ |
5154 | touch_nmi_watchdog(); | |
57675cb9 | 5155 | touch_all_softlockup_watchdogs(); |
39bc89fd | 5156 | if (!state_filter || (p->state & state_filter)) |
82a1fcb9 | 5157 | sched_show_task(p); |
5d07f420 | 5158 | } |
1da177e4 | 5159 | |
dd41f596 | 5160 | #ifdef CONFIG_SCHED_DEBUG |
fb90a6e9 RV |
5161 | if (!state_filter) |
5162 | sysrq_sched_debug_show(); | |
dd41f596 | 5163 | #endif |
510f5acc | 5164 | rcu_read_unlock(); |
e59e2ae2 IM |
5165 | /* |
5166 | * Only show locks if all tasks are dumped: | |
5167 | */ | |
93335a21 | 5168 | if (!state_filter) |
e59e2ae2 | 5169 | debug_show_all_locks(); |
1da177e4 LT |
5170 | } |
5171 | ||
0db0628d | 5172 | void init_idle_bootup_task(struct task_struct *idle) |
1df21055 | 5173 | { |
dd41f596 | 5174 | idle->sched_class = &idle_sched_class; |
1df21055 IM |
5175 | } |
5176 | ||
f340c0d1 IM |
5177 | /** |
5178 | * init_idle - set up an idle thread for a given CPU | |
5179 | * @idle: task in question | |
5180 | * @cpu: cpu the idle task belongs to | |
5181 | * | |
5182 | * NOTE: this function does not set the idle thread's NEED_RESCHED | |
5183 | * flag, to make booting more robust. | |
5184 | */ | |
0db0628d | 5185 | void init_idle(struct task_struct *idle, int cpu) |
1da177e4 | 5186 | { |
70b97a7f | 5187 | struct rq *rq = cpu_rq(cpu); |
1da177e4 LT |
5188 | unsigned long flags; |
5189 | ||
25834c73 PZ |
5190 | raw_spin_lock_irqsave(&idle->pi_lock, flags); |
5191 | raw_spin_lock(&rq->lock); | |
5cbd54ef | 5192 | |
5e1576ed | 5193 | __sched_fork(0, idle); |
06b83b5f | 5194 | idle->state = TASK_RUNNING; |
dd41f596 IM |
5195 | idle->se.exec_start = sched_clock(); |
5196 | ||
e1b77c92 MR |
5197 | kasan_unpoison_task_stack(idle); |
5198 | ||
de9b8f5d PZ |
5199 | #ifdef CONFIG_SMP |
5200 | /* | |
5201 | * Its possible that init_idle() gets called multiple times on a task, | |
5202 | * in that case do_set_cpus_allowed() will not do the right thing. | |
5203 | * | |
5204 | * And since this is boot we can forgo the serialization. | |
5205 | */ | |
5206 | set_cpus_allowed_common(idle, cpumask_of(cpu)); | |
5207 | #endif | |
6506cf6c PZ |
5208 | /* |
5209 | * We're having a chicken and egg problem, even though we are | |
5210 | * holding rq->lock, the cpu isn't yet set to this cpu so the | |
5211 | * lockdep check in task_group() will fail. | |
5212 | * | |
5213 | * Similar case to sched_fork(). / Alternatively we could | |
5214 | * use task_rq_lock() here and obtain the other rq->lock. | |
5215 | * | |
5216 | * Silence PROVE_RCU | |
5217 | */ | |
5218 | rcu_read_lock(); | |
dd41f596 | 5219 | __set_task_cpu(idle, cpu); |
6506cf6c | 5220 | rcu_read_unlock(); |
1da177e4 | 5221 | |
1da177e4 | 5222 | rq->curr = rq->idle = idle; |
da0c1e65 | 5223 | idle->on_rq = TASK_ON_RQ_QUEUED; |
de9b8f5d | 5224 | #ifdef CONFIG_SMP |
3ca7a440 | 5225 | idle->on_cpu = 1; |
4866cde0 | 5226 | #endif |
25834c73 PZ |
5227 | raw_spin_unlock(&rq->lock); |
5228 | raw_spin_unlock_irqrestore(&idle->pi_lock, flags); | |
1da177e4 LT |
5229 | |
5230 | /* Set the preempt count _outside_ the spinlocks! */ | |
01028747 | 5231 | init_idle_preempt_count(idle, cpu); |
55cd5340 | 5232 | |
dd41f596 IM |
5233 | /* |
5234 | * The idle tasks have their own, simple scheduling class: | |
5235 | */ | |
5236 | idle->sched_class = &idle_sched_class; | |
868baf07 | 5237 | ftrace_graph_init_idle_task(idle, cpu); |
45eacc69 | 5238 | vtime_init_idle(idle, cpu); |
de9b8f5d | 5239 | #ifdef CONFIG_SMP |
f1c6f1a7 CE |
5240 | sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); |
5241 | #endif | |
19978ca6 IM |
5242 | } |
5243 | ||
f82f8042 JL |
5244 | int cpuset_cpumask_can_shrink(const struct cpumask *cur, |
5245 | const struct cpumask *trial) | |
5246 | { | |
5247 | int ret = 1, trial_cpus; | |
5248 | struct dl_bw *cur_dl_b; | |
5249 | unsigned long flags; | |
5250 | ||
bb2bc55a MG |
5251 | if (!cpumask_weight(cur)) |
5252 | return ret; | |
5253 | ||
75e23e49 | 5254 | rcu_read_lock_sched(); |
f82f8042 JL |
5255 | cur_dl_b = dl_bw_of(cpumask_any(cur)); |
5256 | trial_cpus = cpumask_weight(trial); | |
5257 | ||
5258 | raw_spin_lock_irqsave(&cur_dl_b->lock, flags); | |
5259 | if (cur_dl_b->bw != -1 && | |
5260 | cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw) | |
5261 | ret = 0; | |
5262 | raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags); | |
75e23e49 | 5263 | rcu_read_unlock_sched(); |
f82f8042 JL |
5264 | |
5265 | return ret; | |
5266 | } | |
5267 | ||
7f51412a JL |
5268 | int task_can_attach(struct task_struct *p, |
5269 | const struct cpumask *cs_cpus_allowed) | |
5270 | { | |
5271 | int ret = 0; | |
5272 | ||
5273 | /* | |
5274 | * Kthreads which disallow setaffinity shouldn't be moved | |
5275 | * to a new cpuset; we don't want to change their cpu | |
5276 | * affinity and isolating such threads by their set of | |
5277 | * allowed nodes is unnecessary. Thus, cpusets are not | |
5278 | * applicable for such threads. This prevents checking for | |
5279 | * success of set_cpus_allowed_ptr() on all attached tasks | |
5280 | * before cpus_allowed may be changed. | |
5281 | */ | |
5282 | if (p->flags & PF_NO_SETAFFINITY) { | |
5283 | ret = -EINVAL; | |
5284 | goto out; | |
5285 | } | |
5286 | ||
5287 | #ifdef CONFIG_SMP | |
5288 | if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span, | |
5289 | cs_cpus_allowed)) { | |
5290 | unsigned int dest_cpu = cpumask_any_and(cpu_active_mask, | |
5291 | cs_cpus_allowed); | |
75e23e49 | 5292 | struct dl_bw *dl_b; |
7f51412a JL |
5293 | bool overflow; |
5294 | int cpus; | |
5295 | unsigned long flags; | |
5296 | ||
75e23e49 JL |
5297 | rcu_read_lock_sched(); |
5298 | dl_b = dl_bw_of(dest_cpu); | |
7f51412a JL |
5299 | raw_spin_lock_irqsave(&dl_b->lock, flags); |
5300 | cpus = dl_bw_cpus(dest_cpu); | |
5301 | overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw); | |
5302 | if (overflow) | |
5303 | ret = -EBUSY; | |
5304 | else { | |
5305 | /* | |
5306 | * We reserve space for this task in the destination | |
5307 | * root_domain, as we can't fail after this point. | |
5308 | * We will free resources in the source root_domain | |
5309 | * later on (see set_cpus_allowed_dl()). | |
5310 | */ | |
5311 | __dl_add(dl_b, p->dl.dl_bw); | |
5312 | } | |
5313 | raw_spin_unlock_irqrestore(&dl_b->lock, flags); | |
75e23e49 | 5314 | rcu_read_unlock_sched(); |
7f51412a JL |
5315 | |
5316 | } | |
5317 | #endif | |
5318 | out: | |
5319 | return ret; | |
5320 | } | |
5321 | ||
1da177e4 | 5322 | #ifdef CONFIG_SMP |
1da177e4 | 5323 | |
e26fbffd TG |
5324 | static bool sched_smp_initialized __read_mostly; |
5325 | ||
e6628d5b MG |
5326 | #ifdef CONFIG_NUMA_BALANCING |
5327 | /* Migrate current task p to target_cpu */ | |
5328 | int migrate_task_to(struct task_struct *p, int target_cpu) | |
5329 | { | |
5330 | struct migration_arg arg = { p, target_cpu }; | |
5331 | int curr_cpu = task_cpu(p); | |
5332 | ||
5333 | if (curr_cpu == target_cpu) | |
5334 | return 0; | |
5335 | ||
5336 | if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) | |
5337 | return -EINVAL; | |
5338 | ||
5339 | /* TODO: This is not properly updating schedstats */ | |
5340 | ||
286549dc | 5341 | trace_sched_move_numa(p, curr_cpu, target_cpu); |
e6628d5b MG |
5342 | return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); |
5343 | } | |
0ec8aa00 PZ |
5344 | |
5345 | /* | |
5346 | * Requeue a task on a given node and accurately track the number of NUMA | |
5347 | * tasks on the runqueues | |
5348 | */ | |
5349 | void sched_setnuma(struct task_struct *p, int nid) | |
5350 | { | |
da0c1e65 | 5351 | bool queued, running; |
eb580751 PZ |
5352 | struct rq_flags rf; |
5353 | struct rq *rq; | |
0ec8aa00 | 5354 | |
eb580751 | 5355 | rq = task_rq_lock(p, &rf); |
da0c1e65 | 5356 | queued = task_on_rq_queued(p); |
0ec8aa00 PZ |
5357 | running = task_current(rq, p); |
5358 | ||
da0c1e65 | 5359 | if (queued) |
1de64443 | 5360 | dequeue_task(rq, p, DEQUEUE_SAVE); |
0ec8aa00 | 5361 | if (running) |
f3cd1c4e | 5362 | put_prev_task(rq, p); |
0ec8aa00 PZ |
5363 | |
5364 | p->numa_preferred_nid = nid; | |
0ec8aa00 PZ |
5365 | |
5366 | if (running) | |
5367 | p->sched_class->set_curr_task(rq); | |
da0c1e65 | 5368 | if (queued) |
1de64443 | 5369 | enqueue_task(rq, p, ENQUEUE_RESTORE); |
eb580751 | 5370 | task_rq_unlock(rq, p, &rf); |
0ec8aa00 | 5371 | } |
5cc389bc | 5372 | #endif /* CONFIG_NUMA_BALANCING */ |
f7b4cddc | 5373 | |
1da177e4 | 5374 | #ifdef CONFIG_HOTPLUG_CPU |
054b9108 | 5375 | /* |
48c5ccae PZ |
5376 | * Ensures that the idle task is using init_mm right before its cpu goes |
5377 | * offline. | |
054b9108 | 5378 | */ |
48c5ccae | 5379 | void idle_task_exit(void) |
1da177e4 | 5380 | { |
48c5ccae | 5381 | struct mm_struct *mm = current->active_mm; |
e76bd8d9 | 5382 | |
48c5ccae | 5383 | BUG_ON(cpu_online(smp_processor_id())); |
e76bd8d9 | 5384 | |
a53efe5f | 5385 | if (mm != &init_mm) { |
f98db601 | 5386 | switch_mm_irqs_off(mm, &init_mm, current); |
a53efe5f MS |
5387 | finish_arch_post_lock_switch(); |
5388 | } | |
48c5ccae | 5389 | mmdrop(mm); |
1da177e4 LT |
5390 | } |
5391 | ||
5392 | /* | |
5d180232 PZ |
5393 | * Since this CPU is going 'away' for a while, fold any nr_active delta |
5394 | * we might have. Assumes we're called after migrate_tasks() so that the | |
5395 | * nr_active count is stable. | |
5396 | * | |
5397 | * Also see the comment "Global load-average calculations". | |
1da177e4 | 5398 | */ |
5d180232 | 5399 | static void calc_load_migrate(struct rq *rq) |
1da177e4 | 5400 | { |
5d180232 PZ |
5401 | long delta = calc_load_fold_active(rq); |
5402 | if (delta) | |
5403 | atomic_long_add(delta, &calc_load_tasks); | |
1da177e4 LT |
5404 | } |
5405 | ||
3f1d2a31 PZ |
5406 | static void put_prev_task_fake(struct rq *rq, struct task_struct *prev) |
5407 | { | |
5408 | } | |
5409 | ||
5410 | static const struct sched_class fake_sched_class = { | |
5411 | .put_prev_task = put_prev_task_fake, | |
5412 | }; | |
5413 | ||
5414 | static struct task_struct fake_task = { | |
5415 | /* | |
5416 | * Avoid pull_{rt,dl}_task() | |
5417 | */ | |
5418 | .prio = MAX_PRIO + 1, | |
5419 | .sched_class = &fake_sched_class, | |
5420 | }; | |
5421 | ||
48f24c4d | 5422 | /* |
48c5ccae PZ |
5423 | * Migrate all tasks from the rq, sleeping tasks will be migrated by |
5424 | * try_to_wake_up()->select_task_rq(). | |
5425 | * | |
5426 | * Called with rq->lock held even though we'er in stop_machine() and | |
5427 | * there's no concurrency possible, we hold the required locks anyway | |
5428 | * because of lock validation efforts. | |
1da177e4 | 5429 | */ |
5e16bbc2 | 5430 | static void migrate_tasks(struct rq *dead_rq) |
1da177e4 | 5431 | { |
5e16bbc2 | 5432 | struct rq *rq = dead_rq; |
48c5ccae | 5433 | struct task_struct *next, *stop = rq->stop; |
e7904a28 | 5434 | struct pin_cookie cookie; |
48c5ccae | 5435 | int dest_cpu; |
1da177e4 LT |
5436 | |
5437 | /* | |
48c5ccae PZ |
5438 | * Fudge the rq selection such that the below task selection loop |
5439 | * doesn't get stuck on the currently eligible stop task. | |
5440 | * | |
5441 | * We're currently inside stop_machine() and the rq is either stuck | |
5442 | * in the stop_machine_cpu_stop() loop, or we're executing this code, | |
5443 | * either way we should never end up calling schedule() until we're | |
5444 | * done here. | |
1da177e4 | 5445 | */ |
48c5ccae | 5446 | rq->stop = NULL; |
48f24c4d | 5447 | |
77bd3970 FW |
5448 | /* |
5449 | * put_prev_task() and pick_next_task() sched | |
5450 | * class method both need to have an up-to-date | |
5451 | * value of rq->clock[_task] | |
5452 | */ | |
5453 | update_rq_clock(rq); | |
5454 | ||
5e16bbc2 | 5455 | for (;;) { |
48c5ccae PZ |
5456 | /* |
5457 | * There's this thread running, bail when that's the only | |
5458 | * remaining thread. | |
5459 | */ | |
5460 | if (rq->nr_running == 1) | |
dd41f596 | 5461 | break; |
48c5ccae | 5462 | |
cbce1a68 | 5463 | /* |
5473e0cc | 5464 | * pick_next_task assumes pinned rq->lock. |
cbce1a68 | 5465 | */ |
e7904a28 PZ |
5466 | cookie = lockdep_pin_lock(&rq->lock); |
5467 | next = pick_next_task(rq, &fake_task, cookie); | |
48c5ccae | 5468 | BUG_ON(!next); |
79c53799 | 5469 | next->sched_class->put_prev_task(rq, next); |
e692ab53 | 5470 | |
5473e0cc WL |
5471 | /* |
5472 | * Rules for changing task_struct::cpus_allowed are holding | |
5473 | * both pi_lock and rq->lock, such that holding either | |
5474 | * stabilizes the mask. | |
5475 | * | |
5476 | * Drop rq->lock is not quite as disastrous as it usually is | |
5477 | * because !cpu_active at this point, which means load-balance | |
5478 | * will not interfere. Also, stop-machine. | |
5479 | */ | |
e7904a28 | 5480 | lockdep_unpin_lock(&rq->lock, cookie); |
5473e0cc WL |
5481 | raw_spin_unlock(&rq->lock); |
5482 | raw_spin_lock(&next->pi_lock); | |
5483 | raw_spin_lock(&rq->lock); | |
5484 | ||
5485 | /* | |
5486 | * Since we're inside stop-machine, _nothing_ should have | |
5487 | * changed the task, WARN if weird stuff happened, because in | |
5488 | * that case the above rq->lock drop is a fail too. | |
5489 | */ | |
5490 | if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) { | |
5491 | raw_spin_unlock(&next->pi_lock); | |
5492 | continue; | |
5493 | } | |
5494 | ||
48c5ccae | 5495 | /* Find suitable destination for @next, with force if needed. */ |
5e16bbc2 | 5496 | dest_cpu = select_fallback_rq(dead_rq->cpu, next); |
48c5ccae | 5497 | |
5e16bbc2 PZ |
5498 | rq = __migrate_task(rq, next, dest_cpu); |
5499 | if (rq != dead_rq) { | |
5500 | raw_spin_unlock(&rq->lock); | |
5501 | rq = dead_rq; | |
5502 | raw_spin_lock(&rq->lock); | |
5503 | } | |
5473e0cc | 5504 | raw_spin_unlock(&next->pi_lock); |
1da177e4 | 5505 | } |
dce48a84 | 5506 | |
48c5ccae | 5507 | rq->stop = stop; |
dce48a84 | 5508 | } |
1da177e4 LT |
5509 | #endif /* CONFIG_HOTPLUG_CPU */ |
5510 | ||
1f11eb6a GH |
5511 | static void set_rq_online(struct rq *rq) |
5512 | { | |
5513 | if (!rq->online) { | |
5514 | const struct sched_class *class; | |
5515 | ||
c6c4927b | 5516 | cpumask_set_cpu(rq->cpu, rq->rd->online); |
1f11eb6a GH |
5517 | rq->online = 1; |
5518 | ||
5519 | for_each_class(class) { | |
5520 | if (class->rq_online) | |
5521 | class->rq_online(rq); | |
5522 | } | |
5523 | } | |
5524 | } | |
5525 | ||
5526 | static void set_rq_offline(struct rq *rq) | |
5527 | { | |
5528 | if (rq->online) { | |
5529 | const struct sched_class *class; | |
5530 | ||
5531 | for_each_class(class) { | |
5532 | if (class->rq_offline) | |
5533 | class->rq_offline(rq); | |
5534 | } | |
5535 | ||
c6c4927b | 5536 | cpumask_clear_cpu(rq->cpu, rq->rd->online); |
1f11eb6a GH |
5537 | rq->online = 0; |
5538 | } | |
5539 | } | |
5540 | ||
9cf7243d | 5541 | static void set_cpu_rq_start_time(unsigned int cpu) |
1da177e4 | 5542 | { |
969c7921 | 5543 | struct rq *rq = cpu_rq(cpu); |
1da177e4 | 5544 | |
a803f026 CM |
5545 | rq->age_stamp = sched_clock_cpu(cpu); |
5546 | } | |
5547 | ||
4cb98839 PZ |
5548 | static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ |
5549 | ||
3e9830dc | 5550 | #ifdef CONFIG_SCHED_DEBUG |
4dcf6aff | 5551 | |
d039ac60 | 5552 | static __read_mostly int sched_debug_enabled; |
f6630114 | 5553 | |
d039ac60 | 5554 | static int __init sched_debug_setup(char *str) |
f6630114 | 5555 | { |
d039ac60 | 5556 | sched_debug_enabled = 1; |
f6630114 MT |
5557 | |
5558 | return 0; | |
5559 | } | |
d039ac60 PZ |
5560 | early_param("sched_debug", sched_debug_setup); |
5561 | ||
5562 | static inline bool sched_debug(void) | |
5563 | { | |
5564 | return sched_debug_enabled; | |
5565 | } | |
f6630114 | 5566 | |
7c16ec58 | 5567 | static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, |
96f874e2 | 5568 | struct cpumask *groupmask) |
1da177e4 | 5569 | { |
4dcf6aff | 5570 | struct sched_group *group = sd->groups; |
1da177e4 | 5571 | |
96f874e2 | 5572 | cpumask_clear(groupmask); |
4dcf6aff IM |
5573 | |
5574 | printk(KERN_DEBUG "%*s domain %d: ", level, "", level); | |
5575 | ||
5576 | if (!(sd->flags & SD_LOAD_BALANCE)) { | |
3df0fc5b | 5577 | printk("does not load-balance\n"); |
4dcf6aff | 5578 | if (sd->parent) |
3df0fc5b PZ |
5579 | printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" |
5580 | " has parent"); | |
4dcf6aff | 5581 | return -1; |
41c7ce9a NP |
5582 | } |
5583 | ||
333470ee TH |
5584 | printk(KERN_CONT "span %*pbl level %s\n", |
5585 | cpumask_pr_args(sched_domain_span(sd)), sd->name); | |
4dcf6aff | 5586 | |
758b2cdc | 5587 | if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { |
3df0fc5b PZ |
5588 | printk(KERN_ERR "ERROR: domain->span does not contain " |
5589 | "CPU%d\n", cpu); | |
4dcf6aff | 5590 | } |
758b2cdc | 5591 | if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { |
3df0fc5b PZ |
5592 | printk(KERN_ERR "ERROR: domain->groups does not contain" |
5593 | " CPU%d\n", cpu); | |
4dcf6aff | 5594 | } |
1da177e4 | 5595 | |
4dcf6aff | 5596 | printk(KERN_DEBUG "%*s groups:", level + 1, ""); |
1da177e4 | 5597 | do { |
4dcf6aff | 5598 | if (!group) { |
3df0fc5b PZ |
5599 | printk("\n"); |
5600 | printk(KERN_ERR "ERROR: group is NULL\n"); | |
1da177e4 LT |
5601 | break; |
5602 | } | |
5603 | ||
758b2cdc | 5604 | if (!cpumask_weight(sched_group_cpus(group))) { |
3df0fc5b PZ |
5605 | printk(KERN_CONT "\n"); |
5606 | printk(KERN_ERR "ERROR: empty group\n"); | |
4dcf6aff IM |
5607 | break; |
5608 | } | |
1da177e4 | 5609 | |
cb83b629 PZ |
5610 | if (!(sd->flags & SD_OVERLAP) && |
5611 | cpumask_intersects(groupmask, sched_group_cpus(group))) { | |
3df0fc5b PZ |
5612 | printk(KERN_CONT "\n"); |
5613 | printk(KERN_ERR "ERROR: repeated CPUs\n"); | |
4dcf6aff IM |
5614 | break; |
5615 | } | |
1da177e4 | 5616 | |
758b2cdc | 5617 | cpumask_or(groupmask, groupmask, sched_group_cpus(group)); |
1da177e4 | 5618 | |
333470ee TH |
5619 | printk(KERN_CONT " %*pbl", |
5620 | cpumask_pr_args(sched_group_cpus(group))); | |
ca8ce3d0 | 5621 | if (group->sgc->capacity != SCHED_CAPACITY_SCALE) { |
63b2ca30 NP |
5622 | printk(KERN_CONT " (cpu_capacity = %d)", |
5623 | group->sgc->capacity); | |
381512cf | 5624 | } |
1da177e4 | 5625 | |
4dcf6aff IM |
5626 | group = group->next; |
5627 | } while (group != sd->groups); | |
3df0fc5b | 5628 | printk(KERN_CONT "\n"); |
1da177e4 | 5629 | |
758b2cdc | 5630 | if (!cpumask_equal(sched_domain_span(sd), groupmask)) |
3df0fc5b | 5631 | printk(KERN_ERR "ERROR: groups don't span domain->span\n"); |
1da177e4 | 5632 | |
758b2cdc RR |
5633 | if (sd->parent && |
5634 | !cpumask_subset(groupmask, sched_domain_span(sd->parent))) | |
3df0fc5b PZ |
5635 | printk(KERN_ERR "ERROR: parent span is not a superset " |
5636 | "of domain->span\n"); | |
4dcf6aff IM |
5637 | return 0; |
5638 | } | |
1da177e4 | 5639 | |
4dcf6aff IM |
5640 | static void sched_domain_debug(struct sched_domain *sd, int cpu) |
5641 | { | |
5642 | int level = 0; | |
1da177e4 | 5643 | |
d039ac60 | 5644 | if (!sched_debug_enabled) |
f6630114 MT |
5645 | return; |
5646 | ||
4dcf6aff IM |
5647 | if (!sd) { |
5648 | printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); | |
5649 | return; | |
5650 | } | |
1da177e4 | 5651 | |
4dcf6aff IM |
5652 | printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); |
5653 | ||
5654 | for (;;) { | |
4cb98839 | 5655 | if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) |
4dcf6aff | 5656 | break; |
1da177e4 LT |
5657 | level++; |
5658 | sd = sd->parent; | |
33859f7f | 5659 | if (!sd) |
4dcf6aff IM |
5660 | break; |
5661 | } | |
1da177e4 | 5662 | } |
6d6bc0ad | 5663 | #else /* !CONFIG_SCHED_DEBUG */ |
48f24c4d | 5664 | # define sched_domain_debug(sd, cpu) do { } while (0) |
d039ac60 PZ |
5665 | static inline bool sched_debug(void) |
5666 | { | |
5667 | return false; | |
5668 | } | |
6d6bc0ad | 5669 | #endif /* CONFIG_SCHED_DEBUG */ |
1da177e4 | 5670 | |
1a20ff27 | 5671 | static int sd_degenerate(struct sched_domain *sd) |
245af2c7 | 5672 | { |
758b2cdc | 5673 | if (cpumask_weight(sched_domain_span(sd)) == 1) |
245af2c7 SS |
5674 | return 1; |
5675 | ||
5676 | /* Following flags need at least 2 groups */ | |
5677 | if (sd->flags & (SD_LOAD_BALANCE | | |
5678 | SD_BALANCE_NEWIDLE | | |
5679 | SD_BALANCE_FORK | | |
89c4710e | 5680 | SD_BALANCE_EXEC | |
5d4dfddd | 5681 | SD_SHARE_CPUCAPACITY | |
d77b3ed5 VG |
5682 | SD_SHARE_PKG_RESOURCES | |
5683 | SD_SHARE_POWERDOMAIN)) { | |
245af2c7 SS |
5684 | if (sd->groups != sd->groups->next) |
5685 | return 0; | |
5686 | } | |
5687 | ||
5688 | /* Following flags don't use groups */ | |
c88d5910 | 5689 | if (sd->flags & (SD_WAKE_AFFINE)) |
245af2c7 SS |
5690 | return 0; |
5691 | ||
5692 | return 1; | |
5693 | } | |
5694 | ||
48f24c4d IM |
5695 | static int |
5696 | sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) | |
245af2c7 SS |
5697 | { |
5698 | unsigned long cflags = sd->flags, pflags = parent->flags; | |
5699 | ||
5700 | if (sd_degenerate(parent)) | |
5701 | return 1; | |
5702 | ||
758b2cdc | 5703 | if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) |
245af2c7 SS |
5704 | return 0; |
5705 | ||
245af2c7 SS |
5706 | /* Flags needing groups don't count if only 1 group in parent */ |
5707 | if (parent->groups == parent->groups->next) { | |
5708 | pflags &= ~(SD_LOAD_BALANCE | | |
5709 | SD_BALANCE_NEWIDLE | | |
5710 | SD_BALANCE_FORK | | |
89c4710e | 5711 | SD_BALANCE_EXEC | |
5d4dfddd | 5712 | SD_SHARE_CPUCAPACITY | |
10866e62 | 5713 | SD_SHARE_PKG_RESOURCES | |
d77b3ed5 VG |
5714 | SD_PREFER_SIBLING | |
5715 | SD_SHARE_POWERDOMAIN); | |
5436499e KC |
5716 | if (nr_node_ids == 1) |
5717 | pflags &= ~SD_SERIALIZE; | |
245af2c7 SS |
5718 | } |
5719 | if (~cflags & pflags) | |
5720 | return 0; | |
5721 | ||
5722 | return 1; | |
5723 | } | |
5724 | ||
dce840a0 | 5725 | static void free_rootdomain(struct rcu_head *rcu) |
c6c4927b | 5726 | { |
dce840a0 | 5727 | struct root_domain *rd = container_of(rcu, struct root_domain, rcu); |
047106ad | 5728 | |
68e74568 | 5729 | cpupri_cleanup(&rd->cpupri); |
6bfd6d72 | 5730 | cpudl_cleanup(&rd->cpudl); |
1baca4ce | 5731 | free_cpumask_var(rd->dlo_mask); |
c6c4927b RR |
5732 | free_cpumask_var(rd->rto_mask); |
5733 | free_cpumask_var(rd->online); | |
5734 | free_cpumask_var(rd->span); | |
5735 | kfree(rd); | |
5736 | } | |
5737 | ||
57d885fe GH |
5738 | static void rq_attach_root(struct rq *rq, struct root_domain *rd) |
5739 | { | |
a0490fa3 | 5740 | struct root_domain *old_rd = NULL; |
57d885fe | 5741 | unsigned long flags; |
57d885fe | 5742 | |
05fa785c | 5743 | raw_spin_lock_irqsave(&rq->lock, flags); |
57d885fe GH |
5744 | |
5745 | if (rq->rd) { | |
a0490fa3 | 5746 | old_rd = rq->rd; |
57d885fe | 5747 | |
c6c4927b | 5748 | if (cpumask_test_cpu(rq->cpu, old_rd->online)) |
1f11eb6a | 5749 | set_rq_offline(rq); |
57d885fe | 5750 | |
c6c4927b | 5751 | cpumask_clear_cpu(rq->cpu, old_rd->span); |
dc938520 | 5752 | |
a0490fa3 | 5753 | /* |
0515973f | 5754 | * If we dont want to free the old_rd yet then |
a0490fa3 IM |
5755 | * set old_rd to NULL to skip the freeing later |
5756 | * in this function: | |
5757 | */ | |
5758 | if (!atomic_dec_and_test(&old_rd->refcount)) | |
5759 | old_rd = NULL; | |
57d885fe GH |
5760 | } |
5761 | ||
5762 | atomic_inc(&rd->refcount); | |
5763 | rq->rd = rd; | |
5764 | ||
c6c4927b | 5765 | cpumask_set_cpu(rq->cpu, rd->span); |
00aec93d | 5766 | if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) |
1f11eb6a | 5767 | set_rq_online(rq); |
57d885fe | 5768 | |
05fa785c | 5769 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
a0490fa3 IM |
5770 | |
5771 | if (old_rd) | |
dce840a0 | 5772 | call_rcu_sched(&old_rd->rcu, free_rootdomain); |
57d885fe GH |
5773 | } |
5774 | ||
68c38fc3 | 5775 | static int init_rootdomain(struct root_domain *rd) |
57d885fe GH |
5776 | { |
5777 | memset(rd, 0, sizeof(*rd)); | |
5778 | ||
8295c699 | 5779 | if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL)) |
0c910d28 | 5780 | goto out; |
8295c699 | 5781 | if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL)) |
c6c4927b | 5782 | goto free_span; |
8295c699 | 5783 | if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) |
c6c4927b | 5784 | goto free_online; |
8295c699 | 5785 | if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) |
1baca4ce | 5786 | goto free_dlo_mask; |
6e0534f2 | 5787 | |
332ac17e | 5788 | init_dl_bw(&rd->dl_bw); |
6bfd6d72 JL |
5789 | if (cpudl_init(&rd->cpudl) != 0) |
5790 | goto free_dlo_mask; | |
332ac17e | 5791 | |
68c38fc3 | 5792 | if (cpupri_init(&rd->cpupri) != 0) |
68e74568 | 5793 | goto free_rto_mask; |
c6c4927b | 5794 | return 0; |
6e0534f2 | 5795 | |
68e74568 RR |
5796 | free_rto_mask: |
5797 | free_cpumask_var(rd->rto_mask); | |
1baca4ce JL |
5798 | free_dlo_mask: |
5799 | free_cpumask_var(rd->dlo_mask); | |
c6c4927b RR |
5800 | free_online: |
5801 | free_cpumask_var(rd->online); | |
5802 | free_span: | |
5803 | free_cpumask_var(rd->span); | |
0c910d28 | 5804 | out: |
c6c4927b | 5805 | return -ENOMEM; |
57d885fe GH |
5806 | } |
5807 | ||
029632fb PZ |
5808 | /* |
5809 | * By default the system creates a single root-domain with all cpus as | |
5810 | * members (mimicking the global state we have today). | |
5811 | */ | |
5812 | struct root_domain def_root_domain; | |
5813 | ||
57d885fe GH |
5814 | static void init_defrootdomain(void) |
5815 | { | |
68c38fc3 | 5816 | init_rootdomain(&def_root_domain); |
c6c4927b | 5817 | |
57d885fe GH |
5818 | atomic_set(&def_root_domain.refcount, 1); |
5819 | } | |
5820 | ||
dc938520 | 5821 | static struct root_domain *alloc_rootdomain(void) |
57d885fe GH |
5822 | { |
5823 | struct root_domain *rd; | |
5824 | ||
5825 | rd = kmalloc(sizeof(*rd), GFP_KERNEL); | |
5826 | if (!rd) | |
5827 | return NULL; | |
5828 | ||
68c38fc3 | 5829 | if (init_rootdomain(rd) != 0) { |
c6c4927b RR |
5830 | kfree(rd); |
5831 | return NULL; | |
5832 | } | |
57d885fe GH |
5833 | |
5834 | return rd; | |
5835 | } | |
5836 | ||
63b2ca30 | 5837 | static void free_sched_groups(struct sched_group *sg, int free_sgc) |
e3589f6c PZ |
5838 | { |
5839 | struct sched_group *tmp, *first; | |
5840 | ||
5841 | if (!sg) | |
5842 | return; | |
5843 | ||
5844 | first = sg; | |
5845 | do { | |
5846 | tmp = sg->next; | |
5847 | ||
63b2ca30 NP |
5848 | if (free_sgc && atomic_dec_and_test(&sg->sgc->ref)) |
5849 | kfree(sg->sgc); | |
e3589f6c PZ |
5850 | |
5851 | kfree(sg); | |
5852 | sg = tmp; | |
5853 | } while (sg != first); | |
5854 | } | |
5855 | ||
dce840a0 PZ |
5856 | static void free_sched_domain(struct rcu_head *rcu) |
5857 | { | |
5858 | struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); | |
e3589f6c PZ |
5859 | |
5860 | /* | |
5861 | * If its an overlapping domain it has private groups, iterate and | |
5862 | * nuke them all. | |
5863 | */ | |
5864 | if (sd->flags & SD_OVERLAP) { | |
5865 | free_sched_groups(sd->groups, 1); | |
5866 | } else if (atomic_dec_and_test(&sd->groups->ref)) { | |
63b2ca30 | 5867 | kfree(sd->groups->sgc); |
dce840a0 | 5868 | kfree(sd->groups); |
9c3f75cb | 5869 | } |
dce840a0 PZ |
5870 | kfree(sd); |
5871 | } | |
5872 | ||
5873 | static void destroy_sched_domain(struct sched_domain *sd, int cpu) | |
5874 | { | |
5875 | call_rcu(&sd->rcu, free_sched_domain); | |
5876 | } | |
5877 | ||
5878 | static void destroy_sched_domains(struct sched_domain *sd, int cpu) | |
5879 | { | |
5880 | for (; sd; sd = sd->parent) | |
5881 | destroy_sched_domain(sd, cpu); | |
5882 | } | |
5883 | ||
518cd623 PZ |
5884 | /* |
5885 | * Keep a special pointer to the highest sched_domain that has | |
5886 | * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this | |
5887 | * allows us to avoid some pointer chasing select_idle_sibling(). | |
5888 | * | |
5889 | * Also keep a unique ID per domain (we use the first cpu number in | |
5890 | * the cpumask of the domain), this allows us to quickly tell if | |
39be3501 | 5891 | * two cpus are in the same cache domain, see cpus_share_cache(). |
518cd623 PZ |
5892 | */ |
5893 | DEFINE_PER_CPU(struct sched_domain *, sd_llc); | |
7d9ffa89 | 5894 | DEFINE_PER_CPU(int, sd_llc_size); |
518cd623 | 5895 | DEFINE_PER_CPU(int, sd_llc_id); |
fb13c7ee | 5896 | DEFINE_PER_CPU(struct sched_domain *, sd_numa); |
37dc6b50 PM |
5897 | DEFINE_PER_CPU(struct sched_domain *, sd_busy); |
5898 | DEFINE_PER_CPU(struct sched_domain *, sd_asym); | |
518cd623 PZ |
5899 | |
5900 | static void update_top_cache_domain(int cpu) | |
5901 | { | |
5902 | struct sched_domain *sd; | |
5d4cf996 | 5903 | struct sched_domain *busy_sd = NULL; |
518cd623 | 5904 | int id = cpu; |
7d9ffa89 | 5905 | int size = 1; |
518cd623 PZ |
5906 | |
5907 | sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); | |
7d9ffa89 | 5908 | if (sd) { |
518cd623 | 5909 | id = cpumask_first(sched_domain_span(sd)); |
7d9ffa89 | 5910 | size = cpumask_weight(sched_domain_span(sd)); |
5d4cf996 | 5911 | busy_sd = sd->parent; /* sd_busy */ |
7d9ffa89 | 5912 | } |
5d4cf996 | 5913 | rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd); |
518cd623 PZ |
5914 | |
5915 | rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); | |
7d9ffa89 | 5916 | per_cpu(sd_llc_size, cpu) = size; |
518cd623 | 5917 | per_cpu(sd_llc_id, cpu) = id; |
fb13c7ee MG |
5918 | |
5919 | sd = lowest_flag_domain(cpu, SD_NUMA); | |
5920 | rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); | |
37dc6b50 PM |
5921 | |
5922 | sd = highest_flag_domain(cpu, SD_ASYM_PACKING); | |
5923 | rcu_assign_pointer(per_cpu(sd_asym, cpu), sd); | |
518cd623 PZ |
5924 | } |
5925 | ||
1da177e4 | 5926 | /* |
0eab9146 | 5927 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must |
1da177e4 LT |
5928 | * hold the hotplug lock. |
5929 | */ | |
0eab9146 IM |
5930 | static void |
5931 | cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) | |
1da177e4 | 5932 | { |
70b97a7f | 5933 | struct rq *rq = cpu_rq(cpu); |
245af2c7 SS |
5934 | struct sched_domain *tmp; |
5935 | ||
5936 | /* Remove the sched domains which do not contribute to scheduling. */ | |
f29c9b1c | 5937 | for (tmp = sd; tmp; ) { |
245af2c7 SS |
5938 | struct sched_domain *parent = tmp->parent; |
5939 | if (!parent) | |
5940 | break; | |
f29c9b1c | 5941 | |
1a848870 | 5942 | if (sd_parent_degenerate(tmp, parent)) { |
245af2c7 | 5943 | tmp->parent = parent->parent; |
1a848870 SS |
5944 | if (parent->parent) |
5945 | parent->parent->child = tmp; | |
10866e62 PZ |
5946 | /* |
5947 | * Transfer SD_PREFER_SIBLING down in case of a | |
5948 | * degenerate parent; the spans match for this | |
5949 | * so the property transfers. | |
5950 | */ | |
5951 | if (parent->flags & SD_PREFER_SIBLING) | |
5952 | tmp->flags |= SD_PREFER_SIBLING; | |
dce840a0 | 5953 | destroy_sched_domain(parent, cpu); |
f29c9b1c LZ |
5954 | } else |
5955 | tmp = tmp->parent; | |
245af2c7 SS |
5956 | } |
5957 | ||
1a848870 | 5958 | if (sd && sd_degenerate(sd)) { |
dce840a0 | 5959 | tmp = sd; |
245af2c7 | 5960 | sd = sd->parent; |
dce840a0 | 5961 | destroy_sched_domain(tmp, cpu); |
1a848870 SS |
5962 | if (sd) |
5963 | sd->child = NULL; | |
5964 | } | |
1da177e4 | 5965 | |
4cb98839 | 5966 | sched_domain_debug(sd, cpu); |
1da177e4 | 5967 | |
57d885fe | 5968 | rq_attach_root(rq, rd); |
dce840a0 | 5969 | tmp = rq->sd; |
674311d5 | 5970 | rcu_assign_pointer(rq->sd, sd); |
dce840a0 | 5971 | destroy_sched_domains(tmp, cpu); |
518cd623 PZ |
5972 | |
5973 | update_top_cache_domain(cpu); | |
1da177e4 LT |
5974 | } |
5975 | ||
1da177e4 LT |
5976 | /* Setup the mask of cpus configured for isolated domains */ |
5977 | static int __init isolated_cpu_setup(char *str) | |
5978 | { | |
a6e4491c PB |
5979 | int ret; |
5980 | ||
bdddd296 | 5981 | alloc_bootmem_cpumask_var(&cpu_isolated_map); |
a6e4491c PB |
5982 | ret = cpulist_parse(str, cpu_isolated_map); |
5983 | if (ret) { | |
5984 | pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids); | |
5985 | return 0; | |
5986 | } | |
1da177e4 LT |
5987 | return 1; |
5988 | } | |
8927f494 | 5989 | __setup("isolcpus=", isolated_cpu_setup); |
1da177e4 | 5990 | |
49a02c51 | 5991 | struct s_data { |
21d42ccf | 5992 | struct sched_domain ** __percpu sd; |
49a02c51 AH |
5993 | struct root_domain *rd; |
5994 | }; | |
5995 | ||
2109b99e | 5996 | enum s_alloc { |
2109b99e | 5997 | sa_rootdomain, |
21d42ccf | 5998 | sa_sd, |
dce840a0 | 5999 | sa_sd_storage, |
2109b99e AH |
6000 | sa_none, |
6001 | }; | |
6002 | ||
c1174876 PZ |
6003 | /* |
6004 | * Build an iteration mask that can exclude certain CPUs from the upwards | |
6005 | * domain traversal. | |
6006 | * | |
6007 | * Asymmetric node setups can result in situations where the domain tree is of | |
6008 | * unequal depth, make sure to skip domains that already cover the entire | |
6009 | * range. | |
6010 | * | |
6011 | * In that case build_sched_domains() will have terminated the iteration early | |
6012 | * and our sibling sd spans will be empty. Domains should always include the | |
6013 | * cpu they're built on, so check that. | |
6014 | * | |
6015 | */ | |
6016 | static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) | |
6017 | { | |
6018 | const struct cpumask *span = sched_domain_span(sd); | |
6019 | struct sd_data *sdd = sd->private; | |
6020 | struct sched_domain *sibling; | |
6021 | int i; | |
6022 | ||
6023 | for_each_cpu(i, span) { | |
6024 | sibling = *per_cpu_ptr(sdd->sd, i); | |
6025 | if (!cpumask_test_cpu(i, sched_domain_span(sibling))) | |
6026 | continue; | |
6027 | ||
6028 | cpumask_set_cpu(i, sched_group_mask(sg)); | |
6029 | } | |
6030 | } | |
6031 | ||
6032 | /* | |
6033 | * Return the canonical balance cpu for this group, this is the first cpu | |
6034 | * of this group that's also in the iteration mask. | |
6035 | */ | |
6036 | int group_balance_cpu(struct sched_group *sg) | |
6037 | { | |
6038 | return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); | |
6039 | } | |
6040 | ||
e3589f6c PZ |
6041 | static int |
6042 | build_overlap_sched_groups(struct sched_domain *sd, int cpu) | |
6043 | { | |
6044 | struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; | |
6045 | const struct cpumask *span = sched_domain_span(sd); | |
6046 | struct cpumask *covered = sched_domains_tmpmask; | |
6047 | struct sd_data *sdd = sd->private; | |
aaecac4a | 6048 | struct sched_domain *sibling; |
e3589f6c PZ |
6049 | int i; |
6050 | ||
6051 | cpumask_clear(covered); | |
6052 | ||
6053 | for_each_cpu(i, span) { | |
6054 | struct cpumask *sg_span; | |
6055 | ||
6056 | if (cpumask_test_cpu(i, covered)) | |
6057 | continue; | |
6058 | ||
aaecac4a | 6059 | sibling = *per_cpu_ptr(sdd->sd, i); |
c1174876 PZ |
6060 | |
6061 | /* See the comment near build_group_mask(). */ | |
aaecac4a | 6062 | if (!cpumask_test_cpu(i, sched_domain_span(sibling))) |
c1174876 PZ |
6063 | continue; |
6064 | ||
e3589f6c | 6065 | sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), |
4d78a223 | 6066 | GFP_KERNEL, cpu_to_node(cpu)); |
e3589f6c PZ |
6067 | |
6068 | if (!sg) | |
6069 | goto fail; | |
6070 | ||
6071 | sg_span = sched_group_cpus(sg); | |
aaecac4a ZZ |
6072 | if (sibling->child) |
6073 | cpumask_copy(sg_span, sched_domain_span(sibling->child)); | |
6074 | else | |
e3589f6c PZ |
6075 | cpumask_set_cpu(i, sg_span); |
6076 | ||
6077 | cpumask_or(covered, covered, sg_span); | |
6078 | ||
63b2ca30 NP |
6079 | sg->sgc = *per_cpu_ptr(sdd->sgc, i); |
6080 | if (atomic_inc_return(&sg->sgc->ref) == 1) | |
c1174876 PZ |
6081 | build_group_mask(sd, sg); |
6082 | ||
c3decf0d | 6083 | /* |
63b2ca30 | 6084 | * Initialize sgc->capacity such that even if we mess up the |
c3decf0d PZ |
6085 | * domains and no possible iteration will get us here, we won't |
6086 | * die on a /0 trap. | |
6087 | */ | |
ca8ce3d0 | 6088 | sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span); |
e3589f6c | 6089 | |
c1174876 PZ |
6090 | /* |
6091 | * Make sure the first group of this domain contains the | |
6092 | * canonical balance cpu. Otherwise the sched_domain iteration | |
6093 | * breaks. See update_sg_lb_stats(). | |
6094 | */ | |
74a5ce20 | 6095 | if ((!groups && cpumask_test_cpu(cpu, sg_span)) || |
c1174876 | 6096 | group_balance_cpu(sg) == cpu) |
e3589f6c PZ |
6097 | groups = sg; |
6098 | ||
6099 | if (!first) | |
6100 | first = sg; | |
6101 | if (last) | |
6102 | last->next = sg; | |
6103 | last = sg; | |
6104 | last->next = first; | |
6105 | } | |
6106 | sd->groups = groups; | |
6107 | ||
6108 | return 0; | |
6109 | ||
6110 | fail: | |
6111 | free_sched_groups(first, 0); | |
6112 | ||
6113 | return -ENOMEM; | |
6114 | } | |
6115 | ||
dce840a0 | 6116 | static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) |
1da177e4 | 6117 | { |
dce840a0 PZ |
6118 | struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); |
6119 | struct sched_domain *child = sd->child; | |
1da177e4 | 6120 | |
dce840a0 PZ |
6121 | if (child) |
6122 | cpu = cpumask_first(sched_domain_span(child)); | |
1e9f28fa | 6123 | |
9c3f75cb | 6124 | if (sg) { |
dce840a0 | 6125 | *sg = *per_cpu_ptr(sdd->sg, cpu); |
63b2ca30 NP |
6126 | (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu); |
6127 | atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */ | |
9c3f75cb | 6128 | } |
dce840a0 PZ |
6129 | |
6130 | return cpu; | |
1e9f28fa | 6131 | } |
1e9f28fa | 6132 | |
01a08546 | 6133 | /* |
dce840a0 PZ |
6134 | * build_sched_groups will build a circular linked list of the groups |
6135 | * covered by the given span, and will set each group's ->cpumask correctly, | |
ced549fa | 6136 | * and ->cpu_capacity to 0. |
e3589f6c PZ |
6137 | * |
6138 | * Assumes the sched_domain tree is fully constructed | |
01a08546 | 6139 | */ |
e3589f6c PZ |
6140 | static int |
6141 | build_sched_groups(struct sched_domain *sd, int cpu) | |
1da177e4 | 6142 | { |
dce840a0 PZ |
6143 | struct sched_group *first = NULL, *last = NULL; |
6144 | struct sd_data *sdd = sd->private; | |
6145 | const struct cpumask *span = sched_domain_span(sd); | |
f96225fd | 6146 | struct cpumask *covered; |
dce840a0 | 6147 | int i; |
9c1cfda2 | 6148 | |
e3589f6c PZ |
6149 | get_group(cpu, sdd, &sd->groups); |
6150 | atomic_inc(&sd->groups->ref); | |
6151 | ||
0936629f | 6152 | if (cpu != cpumask_first(span)) |
e3589f6c PZ |
6153 | return 0; |
6154 | ||
f96225fd PZ |
6155 | lockdep_assert_held(&sched_domains_mutex); |
6156 | covered = sched_domains_tmpmask; | |
6157 | ||
dce840a0 | 6158 | cpumask_clear(covered); |
6711cab4 | 6159 | |
dce840a0 PZ |
6160 | for_each_cpu(i, span) { |
6161 | struct sched_group *sg; | |
cd08e923 | 6162 | int group, j; |
6711cab4 | 6163 | |
dce840a0 PZ |
6164 | if (cpumask_test_cpu(i, covered)) |
6165 | continue; | |
6711cab4 | 6166 | |
cd08e923 | 6167 | group = get_group(i, sdd, &sg); |
c1174876 | 6168 | cpumask_setall(sched_group_mask(sg)); |
0601a88d | 6169 | |
dce840a0 PZ |
6170 | for_each_cpu(j, span) { |
6171 | if (get_group(j, sdd, NULL) != group) | |
6172 | continue; | |
0601a88d | 6173 | |
dce840a0 PZ |
6174 | cpumask_set_cpu(j, covered); |
6175 | cpumask_set_cpu(j, sched_group_cpus(sg)); | |
6176 | } | |
0601a88d | 6177 | |
dce840a0 PZ |
6178 | if (!first) |
6179 | first = sg; | |
6180 | if (last) | |
6181 | last->next = sg; | |
6182 | last = sg; | |
6183 | } | |
6184 | last->next = first; | |
e3589f6c PZ |
6185 | |
6186 | return 0; | |
0601a88d | 6187 | } |
51888ca2 | 6188 | |
89c4710e | 6189 | /* |
63b2ca30 | 6190 | * Initialize sched groups cpu_capacity. |
89c4710e | 6191 | * |
63b2ca30 | 6192 | * cpu_capacity indicates the capacity of sched group, which is used while |
89c4710e | 6193 | * distributing the load between different sched groups in a sched domain. |
63b2ca30 NP |
6194 | * Typically cpu_capacity for all the groups in a sched domain will be same |
6195 | * unless there are asymmetries in the topology. If there are asymmetries, | |
6196 | * group having more cpu_capacity will pickup more load compared to the | |
6197 | * group having less cpu_capacity. | |
89c4710e | 6198 | */ |
63b2ca30 | 6199 | static void init_sched_groups_capacity(int cpu, struct sched_domain *sd) |
89c4710e | 6200 | { |
e3589f6c | 6201 | struct sched_group *sg = sd->groups; |
89c4710e | 6202 | |
94c95ba6 | 6203 | WARN_ON(!sg); |
e3589f6c PZ |
6204 | |
6205 | do { | |
6206 | sg->group_weight = cpumask_weight(sched_group_cpus(sg)); | |
6207 | sg = sg->next; | |
6208 | } while (sg != sd->groups); | |
89c4710e | 6209 | |
c1174876 | 6210 | if (cpu != group_balance_cpu(sg)) |
e3589f6c | 6211 | return; |
aae6d3dd | 6212 | |
63b2ca30 NP |
6213 | update_group_capacity(sd, cpu); |
6214 | atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight); | |
89c4710e SS |
6215 | } |
6216 | ||
7c16ec58 MT |
6217 | /* |
6218 | * Initializers for schedule domains | |
6219 | * Non-inlined to reduce accumulated stack pressure in build_sched_domains() | |
6220 | */ | |
6221 | ||
1d3504fc | 6222 | static int default_relax_domain_level = -1; |
60495e77 | 6223 | int sched_domain_level_max; |
1d3504fc HS |
6224 | |
6225 | static int __init setup_relax_domain_level(char *str) | |
6226 | { | |
a841f8ce DS |
6227 | if (kstrtoint(str, 0, &default_relax_domain_level)) |
6228 | pr_warn("Unable to set relax_domain_level\n"); | |
30e0e178 | 6229 | |
1d3504fc HS |
6230 | return 1; |
6231 | } | |
6232 | __setup("relax_domain_level=", setup_relax_domain_level); | |
6233 | ||
6234 | static void set_domain_attribute(struct sched_domain *sd, | |
6235 | struct sched_domain_attr *attr) | |
6236 | { | |
6237 | int request; | |
6238 | ||
6239 | if (!attr || attr->relax_domain_level < 0) { | |
6240 | if (default_relax_domain_level < 0) | |
6241 | return; | |
6242 | else | |
6243 | request = default_relax_domain_level; | |
6244 | } else | |
6245 | request = attr->relax_domain_level; | |
6246 | if (request < sd->level) { | |
6247 | /* turn off idle balance on this domain */ | |
c88d5910 | 6248 | sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); |
1d3504fc HS |
6249 | } else { |
6250 | /* turn on idle balance on this domain */ | |
c88d5910 | 6251 | sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); |
1d3504fc HS |
6252 | } |
6253 | } | |
6254 | ||
54ab4ff4 PZ |
6255 | static void __sdt_free(const struct cpumask *cpu_map); |
6256 | static int __sdt_alloc(const struct cpumask *cpu_map); | |
6257 | ||
2109b99e AH |
6258 | static void __free_domain_allocs(struct s_data *d, enum s_alloc what, |
6259 | const struct cpumask *cpu_map) | |
6260 | { | |
6261 | switch (what) { | |
2109b99e | 6262 | case sa_rootdomain: |
822ff793 PZ |
6263 | if (!atomic_read(&d->rd->refcount)) |
6264 | free_rootdomain(&d->rd->rcu); /* fall through */ | |
21d42ccf PZ |
6265 | case sa_sd: |
6266 | free_percpu(d->sd); /* fall through */ | |
dce840a0 | 6267 | case sa_sd_storage: |
54ab4ff4 | 6268 | __sdt_free(cpu_map); /* fall through */ |
2109b99e AH |
6269 | case sa_none: |
6270 | break; | |
6271 | } | |
6272 | } | |
3404c8d9 | 6273 | |
2109b99e AH |
6274 | static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, |
6275 | const struct cpumask *cpu_map) | |
6276 | { | |
dce840a0 PZ |
6277 | memset(d, 0, sizeof(*d)); |
6278 | ||
54ab4ff4 PZ |
6279 | if (__sdt_alloc(cpu_map)) |
6280 | return sa_sd_storage; | |
dce840a0 PZ |
6281 | d->sd = alloc_percpu(struct sched_domain *); |
6282 | if (!d->sd) | |
6283 | return sa_sd_storage; | |
2109b99e | 6284 | d->rd = alloc_rootdomain(); |
dce840a0 | 6285 | if (!d->rd) |
21d42ccf | 6286 | return sa_sd; |
2109b99e AH |
6287 | return sa_rootdomain; |
6288 | } | |
57d885fe | 6289 | |
dce840a0 PZ |
6290 | /* |
6291 | * NULL the sd_data elements we've used to build the sched_domain and | |
6292 | * sched_group structure so that the subsequent __free_domain_allocs() | |
6293 | * will not free the data we're using. | |
6294 | */ | |
6295 | static void claim_allocations(int cpu, struct sched_domain *sd) | |
6296 | { | |
6297 | struct sd_data *sdd = sd->private; | |
dce840a0 PZ |
6298 | |
6299 | WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); | |
6300 | *per_cpu_ptr(sdd->sd, cpu) = NULL; | |
6301 | ||
e3589f6c | 6302 | if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) |
dce840a0 | 6303 | *per_cpu_ptr(sdd->sg, cpu) = NULL; |
e3589f6c | 6304 | |
63b2ca30 NP |
6305 | if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref)) |
6306 | *per_cpu_ptr(sdd->sgc, cpu) = NULL; | |
dce840a0 PZ |
6307 | } |
6308 | ||
cb83b629 | 6309 | #ifdef CONFIG_NUMA |
cb83b629 | 6310 | static int sched_domains_numa_levels; |
e3fe70b1 | 6311 | enum numa_topology_type sched_numa_topology_type; |
cb83b629 | 6312 | static int *sched_domains_numa_distance; |
9942f79b | 6313 | int sched_max_numa_distance; |
cb83b629 PZ |
6314 | static struct cpumask ***sched_domains_numa_masks; |
6315 | static int sched_domains_curr_level; | |
143e1e28 | 6316 | #endif |
cb83b629 | 6317 | |
143e1e28 VG |
6318 | /* |
6319 | * SD_flags allowed in topology descriptions. | |
6320 | * | |
5d4dfddd | 6321 | * SD_SHARE_CPUCAPACITY - describes SMT topologies |
143e1e28 VG |
6322 | * SD_SHARE_PKG_RESOURCES - describes shared caches |
6323 | * SD_NUMA - describes NUMA topologies | |
d77b3ed5 | 6324 | * SD_SHARE_POWERDOMAIN - describes shared power domain |
143e1e28 VG |
6325 | * |
6326 | * Odd one out: | |
6327 | * SD_ASYM_PACKING - describes SMT quirks | |
6328 | */ | |
6329 | #define TOPOLOGY_SD_FLAGS \ | |
5d4dfddd | 6330 | (SD_SHARE_CPUCAPACITY | \ |
143e1e28 VG |
6331 | SD_SHARE_PKG_RESOURCES | \ |
6332 | SD_NUMA | \ | |
d77b3ed5 VG |
6333 | SD_ASYM_PACKING | \ |
6334 | SD_SHARE_POWERDOMAIN) | |
cb83b629 PZ |
6335 | |
6336 | static struct sched_domain * | |
143e1e28 | 6337 | sd_init(struct sched_domain_topology_level *tl, int cpu) |
cb83b629 PZ |
6338 | { |
6339 | struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); | |
143e1e28 VG |
6340 | int sd_weight, sd_flags = 0; |
6341 | ||
6342 | #ifdef CONFIG_NUMA | |
6343 | /* | |
6344 | * Ugly hack to pass state to sd_numa_mask()... | |
6345 | */ | |
6346 | sched_domains_curr_level = tl->numa_level; | |
6347 | #endif | |
6348 | ||
6349 | sd_weight = cpumask_weight(tl->mask(cpu)); | |
6350 | ||
6351 | if (tl->sd_flags) | |
6352 | sd_flags = (*tl->sd_flags)(); | |
6353 | if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS, | |
6354 | "wrong sd_flags in topology description\n")) | |
6355 | sd_flags &= ~TOPOLOGY_SD_FLAGS; | |
cb83b629 PZ |
6356 | |
6357 | *sd = (struct sched_domain){ | |
6358 | .min_interval = sd_weight, | |
6359 | .max_interval = 2*sd_weight, | |
6360 | .busy_factor = 32, | |
870a0bb5 | 6361 | .imbalance_pct = 125, |
143e1e28 VG |
6362 | |
6363 | .cache_nice_tries = 0, | |
6364 | .busy_idx = 0, | |
6365 | .idle_idx = 0, | |
cb83b629 PZ |
6366 | .newidle_idx = 0, |
6367 | .wake_idx = 0, | |
6368 | .forkexec_idx = 0, | |
6369 | ||
6370 | .flags = 1*SD_LOAD_BALANCE | |
6371 | | 1*SD_BALANCE_NEWIDLE | |
143e1e28 VG |
6372 | | 1*SD_BALANCE_EXEC |
6373 | | 1*SD_BALANCE_FORK | |
cb83b629 | 6374 | | 0*SD_BALANCE_WAKE |
143e1e28 | 6375 | | 1*SD_WAKE_AFFINE |
5d4dfddd | 6376 | | 0*SD_SHARE_CPUCAPACITY |
cb83b629 | 6377 | | 0*SD_SHARE_PKG_RESOURCES |
143e1e28 | 6378 | | 0*SD_SERIALIZE |
cb83b629 | 6379 | | 0*SD_PREFER_SIBLING |
143e1e28 VG |
6380 | | 0*SD_NUMA |
6381 | | sd_flags | |
cb83b629 | 6382 | , |
143e1e28 | 6383 | |
cb83b629 PZ |
6384 | .last_balance = jiffies, |
6385 | .balance_interval = sd_weight, | |
143e1e28 | 6386 | .smt_gain = 0, |
2b4cfe64 JL |
6387 | .max_newidle_lb_cost = 0, |
6388 | .next_decay_max_lb_cost = jiffies, | |
143e1e28 VG |
6389 | #ifdef CONFIG_SCHED_DEBUG |
6390 | .name = tl->name, | |
6391 | #endif | |
cb83b629 | 6392 | }; |
cb83b629 PZ |
6393 | |
6394 | /* | |
143e1e28 | 6395 | * Convert topological properties into behaviour. |
cb83b629 | 6396 | */ |
143e1e28 | 6397 | |
5d4dfddd | 6398 | if (sd->flags & SD_SHARE_CPUCAPACITY) { |
caff37ef | 6399 | sd->flags |= SD_PREFER_SIBLING; |
143e1e28 VG |
6400 | sd->imbalance_pct = 110; |
6401 | sd->smt_gain = 1178; /* ~15% */ | |
143e1e28 VG |
6402 | |
6403 | } else if (sd->flags & SD_SHARE_PKG_RESOURCES) { | |
6404 | sd->imbalance_pct = 117; | |
6405 | sd->cache_nice_tries = 1; | |
6406 | sd->busy_idx = 2; | |
6407 | ||
6408 | #ifdef CONFIG_NUMA | |
6409 | } else if (sd->flags & SD_NUMA) { | |
6410 | sd->cache_nice_tries = 2; | |
6411 | sd->busy_idx = 3; | |
6412 | sd->idle_idx = 2; | |
6413 | ||
6414 | sd->flags |= SD_SERIALIZE; | |
6415 | if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) { | |
6416 | sd->flags &= ~(SD_BALANCE_EXEC | | |
6417 | SD_BALANCE_FORK | | |
6418 | SD_WAKE_AFFINE); | |
6419 | } | |
6420 | ||
6421 | #endif | |
6422 | } else { | |
6423 | sd->flags |= SD_PREFER_SIBLING; | |
6424 | sd->cache_nice_tries = 1; | |
6425 | sd->busy_idx = 2; | |
6426 | sd->idle_idx = 1; | |
6427 | } | |
6428 | ||
6429 | sd->private = &tl->data; | |
cb83b629 PZ |
6430 | |
6431 | return sd; | |
6432 | } | |
6433 | ||
143e1e28 VG |
6434 | /* |
6435 | * Topology list, bottom-up. | |
6436 | */ | |
6437 | static struct sched_domain_topology_level default_topology[] = { | |
6438 | #ifdef CONFIG_SCHED_SMT | |
6439 | { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) }, | |
6440 | #endif | |
6441 | #ifdef CONFIG_SCHED_MC | |
6442 | { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) }, | |
143e1e28 VG |
6443 | #endif |
6444 | { cpu_cpu_mask, SD_INIT_NAME(DIE) }, | |
6445 | { NULL, }, | |
6446 | }; | |
6447 | ||
c6e1e7b5 JG |
6448 | static struct sched_domain_topology_level *sched_domain_topology = |
6449 | default_topology; | |
143e1e28 VG |
6450 | |
6451 | #define for_each_sd_topology(tl) \ | |
6452 | for (tl = sched_domain_topology; tl->mask; tl++) | |
6453 | ||
6454 | void set_sched_topology(struct sched_domain_topology_level *tl) | |
6455 | { | |
6456 | sched_domain_topology = tl; | |
6457 | } | |
6458 | ||
6459 | #ifdef CONFIG_NUMA | |
6460 | ||
cb83b629 PZ |
6461 | static const struct cpumask *sd_numa_mask(int cpu) |
6462 | { | |
6463 | return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; | |
6464 | } | |
6465 | ||
d039ac60 PZ |
6466 | static void sched_numa_warn(const char *str) |
6467 | { | |
6468 | static int done = false; | |
6469 | int i,j; | |
6470 | ||
6471 | if (done) | |
6472 | return; | |
6473 | ||
6474 | done = true; | |
6475 | ||
6476 | printk(KERN_WARNING "ERROR: %s\n\n", str); | |
6477 | ||
6478 | for (i = 0; i < nr_node_ids; i++) { | |
6479 | printk(KERN_WARNING " "); | |
6480 | for (j = 0; j < nr_node_ids; j++) | |
6481 | printk(KERN_CONT "%02d ", node_distance(i,j)); | |
6482 | printk(KERN_CONT "\n"); | |
6483 | } | |
6484 | printk(KERN_WARNING "\n"); | |
6485 | } | |
6486 | ||
9942f79b | 6487 | bool find_numa_distance(int distance) |
d039ac60 PZ |
6488 | { |
6489 | int i; | |
6490 | ||
6491 | if (distance == node_distance(0, 0)) | |
6492 | return true; | |
6493 | ||
6494 | for (i = 0; i < sched_domains_numa_levels; i++) { | |
6495 | if (sched_domains_numa_distance[i] == distance) | |
6496 | return true; | |
6497 | } | |
6498 | ||
6499 | return false; | |
6500 | } | |
6501 | ||
e3fe70b1 RR |
6502 | /* |
6503 | * A system can have three types of NUMA topology: | |
6504 | * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system | |
6505 | * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes | |
6506 | * NUMA_BACKPLANE: nodes can reach other nodes through a backplane | |
6507 | * | |
6508 | * The difference between a glueless mesh topology and a backplane | |
6509 | * topology lies in whether communication between not directly | |
6510 | * connected nodes goes through intermediary nodes (where programs | |
6511 | * could run), or through backplane controllers. This affects | |
6512 | * placement of programs. | |
6513 | * | |
6514 | * The type of topology can be discerned with the following tests: | |
6515 | * - If the maximum distance between any nodes is 1 hop, the system | |
6516 | * is directly connected. | |
6517 | * - If for two nodes A and B, located N > 1 hops away from each other, | |
6518 | * there is an intermediary node C, which is < N hops away from both | |
6519 | * nodes A and B, the system is a glueless mesh. | |
6520 | */ | |
6521 | static void init_numa_topology_type(void) | |
6522 | { | |
6523 | int a, b, c, n; | |
6524 | ||
6525 | n = sched_max_numa_distance; | |
6526 | ||
e237882b | 6527 | if (sched_domains_numa_levels <= 1) { |
e3fe70b1 | 6528 | sched_numa_topology_type = NUMA_DIRECT; |
e237882b AG |
6529 | return; |
6530 | } | |
e3fe70b1 RR |
6531 | |
6532 | for_each_online_node(a) { | |
6533 | for_each_online_node(b) { | |
6534 | /* Find two nodes furthest removed from each other. */ | |
6535 | if (node_distance(a, b) < n) | |
6536 | continue; | |
6537 | ||
6538 | /* Is there an intermediary node between a and b? */ | |
6539 | for_each_online_node(c) { | |
6540 | if (node_distance(a, c) < n && | |
6541 | node_distance(b, c) < n) { | |
6542 | sched_numa_topology_type = | |
6543 | NUMA_GLUELESS_MESH; | |
6544 | return; | |
6545 | } | |
6546 | } | |
6547 | ||
6548 | sched_numa_topology_type = NUMA_BACKPLANE; | |
6549 | return; | |
6550 | } | |
6551 | } | |
6552 | } | |
6553 | ||
cb83b629 PZ |
6554 | static void sched_init_numa(void) |
6555 | { | |
6556 | int next_distance, curr_distance = node_distance(0, 0); | |
6557 | struct sched_domain_topology_level *tl; | |
6558 | int level = 0; | |
6559 | int i, j, k; | |
6560 | ||
cb83b629 PZ |
6561 | sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); |
6562 | if (!sched_domains_numa_distance) | |
6563 | return; | |
6564 | ||
6565 | /* | |
6566 | * O(nr_nodes^2) deduplicating selection sort -- in order to find the | |
6567 | * unique distances in the node_distance() table. | |
6568 | * | |
6569 | * Assumes node_distance(0,j) includes all distances in | |
6570 | * node_distance(i,j) in order to avoid cubic time. | |
cb83b629 PZ |
6571 | */ |
6572 | next_distance = curr_distance; | |
6573 | for (i = 0; i < nr_node_ids; i++) { | |
6574 | for (j = 0; j < nr_node_ids; j++) { | |
d039ac60 PZ |
6575 | for (k = 0; k < nr_node_ids; k++) { |
6576 | int distance = node_distance(i, k); | |
6577 | ||
6578 | if (distance > curr_distance && | |
6579 | (distance < next_distance || | |
6580 | next_distance == curr_distance)) | |
6581 | next_distance = distance; | |
6582 | ||
6583 | /* | |
6584 | * While not a strong assumption it would be nice to know | |
6585 | * about cases where if node A is connected to B, B is not | |
6586 | * equally connected to A. | |
6587 | */ | |
6588 | if (sched_debug() && node_distance(k, i) != distance) | |
6589 | sched_numa_warn("Node-distance not symmetric"); | |
6590 | ||
6591 | if (sched_debug() && i && !find_numa_distance(distance)) | |
6592 | sched_numa_warn("Node-0 not representative"); | |
6593 | } | |
6594 | if (next_distance != curr_distance) { | |
6595 | sched_domains_numa_distance[level++] = next_distance; | |
6596 | sched_domains_numa_levels = level; | |
6597 | curr_distance = next_distance; | |
6598 | } else break; | |
cb83b629 | 6599 | } |
d039ac60 PZ |
6600 | |
6601 | /* | |
6602 | * In case of sched_debug() we verify the above assumption. | |
6603 | */ | |
6604 | if (!sched_debug()) | |
6605 | break; | |
cb83b629 | 6606 | } |
c123588b AR |
6607 | |
6608 | if (!level) | |
6609 | return; | |
6610 | ||
cb83b629 PZ |
6611 | /* |
6612 | * 'level' contains the number of unique distances, excluding the | |
6613 | * identity distance node_distance(i,i). | |
6614 | * | |
28b4a521 | 6615 | * The sched_domains_numa_distance[] array includes the actual distance |
cb83b629 PZ |
6616 | * numbers. |
6617 | */ | |
6618 | ||
5f7865f3 TC |
6619 | /* |
6620 | * Here, we should temporarily reset sched_domains_numa_levels to 0. | |
6621 | * If it fails to allocate memory for array sched_domains_numa_masks[][], | |
6622 | * the array will contain less then 'level' members. This could be | |
6623 | * dangerous when we use it to iterate array sched_domains_numa_masks[][] | |
6624 | * in other functions. | |
6625 | * | |
6626 | * We reset it to 'level' at the end of this function. | |
6627 | */ | |
6628 | sched_domains_numa_levels = 0; | |
6629 | ||
cb83b629 PZ |
6630 | sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); |
6631 | if (!sched_domains_numa_masks) | |
6632 | return; | |
6633 | ||
6634 | /* | |
6635 | * Now for each level, construct a mask per node which contains all | |
6636 | * cpus of nodes that are that many hops away from us. | |
6637 | */ | |
6638 | for (i = 0; i < level; i++) { | |
6639 | sched_domains_numa_masks[i] = | |
6640 | kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); | |
6641 | if (!sched_domains_numa_masks[i]) | |
6642 | return; | |
6643 | ||
6644 | for (j = 0; j < nr_node_ids; j++) { | |
2ea45800 | 6645 | struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); |
cb83b629 PZ |
6646 | if (!mask) |
6647 | return; | |
6648 | ||
6649 | sched_domains_numa_masks[i][j] = mask; | |
6650 | ||
9c03ee14 | 6651 | for_each_node(k) { |
dd7d8634 | 6652 | if (node_distance(j, k) > sched_domains_numa_distance[i]) |
cb83b629 PZ |
6653 | continue; |
6654 | ||
6655 | cpumask_or(mask, mask, cpumask_of_node(k)); | |
6656 | } | |
6657 | } | |
6658 | } | |
6659 | ||
143e1e28 VG |
6660 | /* Compute default topology size */ |
6661 | for (i = 0; sched_domain_topology[i].mask; i++); | |
6662 | ||
c515db8c | 6663 | tl = kzalloc((i + level + 1) * |
cb83b629 PZ |
6664 | sizeof(struct sched_domain_topology_level), GFP_KERNEL); |
6665 | if (!tl) | |
6666 | return; | |
6667 | ||
6668 | /* | |
6669 | * Copy the default topology bits.. | |
6670 | */ | |
143e1e28 VG |
6671 | for (i = 0; sched_domain_topology[i].mask; i++) |
6672 | tl[i] = sched_domain_topology[i]; | |
cb83b629 PZ |
6673 | |
6674 | /* | |
6675 | * .. and append 'j' levels of NUMA goodness. | |
6676 | */ | |
6677 | for (j = 0; j < level; i++, j++) { | |
6678 | tl[i] = (struct sched_domain_topology_level){ | |
cb83b629 | 6679 | .mask = sd_numa_mask, |
143e1e28 | 6680 | .sd_flags = cpu_numa_flags, |
cb83b629 PZ |
6681 | .flags = SDTL_OVERLAP, |
6682 | .numa_level = j, | |
143e1e28 | 6683 | SD_INIT_NAME(NUMA) |
cb83b629 PZ |
6684 | }; |
6685 | } | |
6686 | ||
6687 | sched_domain_topology = tl; | |
5f7865f3 TC |
6688 | |
6689 | sched_domains_numa_levels = level; | |
9942f79b | 6690 | sched_max_numa_distance = sched_domains_numa_distance[level - 1]; |
e3fe70b1 RR |
6691 | |
6692 | init_numa_topology_type(); | |
cb83b629 | 6693 | } |
301a5cba | 6694 | |
135fb3e1 | 6695 | static void sched_domains_numa_masks_set(unsigned int cpu) |
301a5cba | 6696 | { |
301a5cba | 6697 | int node = cpu_to_node(cpu); |
135fb3e1 | 6698 | int i, j; |
301a5cba TC |
6699 | |
6700 | for (i = 0; i < sched_domains_numa_levels; i++) { | |
6701 | for (j = 0; j < nr_node_ids; j++) { | |
6702 | if (node_distance(j, node) <= sched_domains_numa_distance[i]) | |
6703 | cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); | |
6704 | } | |
6705 | } | |
6706 | } | |
6707 | ||
135fb3e1 | 6708 | static void sched_domains_numa_masks_clear(unsigned int cpu) |
301a5cba TC |
6709 | { |
6710 | int i, j; | |
135fb3e1 | 6711 | |
301a5cba TC |
6712 | for (i = 0; i < sched_domains_numa_levels; i++) { |
6713 | for (j = 0; j < nr_node_ids; j++) | |
6714 | cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); | |
6715 | } | |
6716 | } | |
6717 | ||
cb83b629 | 6718 | #else |
135fb3e1 TG |
6719 | static inline void sched_init_numa(void) { } |
6720 | static void sched_domains_numa_masks_set(unsigned int cpu) { } | |
6721 | static void sched_domains_numa_masks_clear(unsigned int cpu) { } | |
cb83b629 PZ |
6722 | #endif /* CONFIG_NUMA */ |
6723 | ||
54ab4ff4 PZ |
6724 | static int __sdt_alloc(const struct cpumask *cpu_map) |
6725 | { | |
6726 | struct sched_domain_topology_level *tl; | |
6727 | int j; | |
6728 | ||
27723a68 | 6729 | for_each_sd_topology(tl) { |
54ab4ff4 PZ |
6730 | struct sd_data *sdd = &tl->data; |
6731 | ||
6732 | sdd->sd = alloc_percpu(struct sched_domain *); | |
6733 | if (!sdd->sd) | |
6734 | return -ENOMEM; | |
6735 | ||
6736 | sdd->sg = alloc_percpu(struct sched_group *); | |
6737 | if (!sdd->sg) | |
6738 | return -ENOMEM; | |
6739 | ||
63b2ca30 NP |
6740 | sdd->sgc = alloc_percpu(struct sched_group_capacity *); |
6741 | if (!sdd->sgc) | |
9c3f75cb PZ |
6742 | return -ENOMEM; |
6743 | ||
54ab4ff4 PZ |
6744 | for_each_cpu(j, cpu_map) { |
6745 | struct sched_domain *sd; | |
6746 | struct sched_group *sg; | |
63b2ca30 | 6747 | struct sched_group_capacity *sgc; |
54ab4ff4 | 6748 | |
5cc389bc | 6749 | sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), |
54ab4ff4 PZ |
6750 | GFP_KERNEL, cpu_to_node(j)); |
6751 | if (!sd) | |
6752 | return -ENOMEM; | |
6753 | ||
6754 | *per_cpu_ptr(sdd->sd, j) = sd; | |
6755 | ||
6756 | sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), | |
6757 | GFP_KERNEL, cpu_to_node(j)); | |
6758 | if (!sg) | |
6759 | return -ENOMEM; | |
6760 | ||
30b4e9eb IM |
6761 | sg->next = sg; |
6762 | ||
54ab4ff4 | 6763 | *per_cpu_ptr(sdd->sg, j) = sg; |
9c3f75cb | 6764 | |
63b2ca30 | 6765 | sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(), |
9c3f75cb | 6766 | GFP_KERNEL, cpu_to_node(j)); |
63b2ca30 | 6767 | if (!sgc) |
9c3f75cb PZ |
6768 | return -ENOMEM; |
6769 | ||
63b2ca30 | 6770 | *per_cpu_ptr(sdd->sgc, j) = sgc; |
54ab4ff4 PZ |
6771 | } |
6772 | } | |
6773 | ||
6774 | return 0; | |
6775 | } | |
6776 | ||
6777 | static void __sdt_free(const struct cpumask *cpu_map) | |
6778 | { | |
6779 | struct sched_domain_topology_level *tl; | |
6780 | int j; | |
6781 | ||
27723a68 | 6782 | for_each_sd_topology(tl) { |
54ab4ff4 PZ |
6783 | struct sd_data *sdd = &tl->data; |
6784 | ||
6785 | for_each_cpu(j, cpu_map) { | |
fb2cf2c6 | 6786 | struct sched_domain *sd; |
6787 | ||
6788 | if (sdd->sd) { | |
6789 | sd = *per_cpu_ptr(sdd->sd, j); | |
6790 | if (sd && (sd->flags & SD_OVERLAP)) | |
6791 | free_sched_groups(sd->groups, 0); | |
6792 | kfree(*per_cpu_ptr(sdd->sd, j)); | |
6793 | } | |
6794 | ||
6795 | if (sdd->sg) | |
6796 | kfree(*per_cpu_ptr(sdd->sg, j)); | |
63b2ca30 NP |
6797 | if (sdd->sgc) |
6798 | kfree(*per_cpu_ptr(sdd->sgc, j)); | |
54ab4ff4 PZ |
6799 | } |
6800 | free_percpu(sdd->sd); | |
fb2cf2c6 | 6801 | sdd->sd = NULL; |
54ab4ff4 | 6802 | free_percpu(sdd->sg); |
fb2cf2c6 | 6803 | sdd->sg = NULL; |
63b2ca30 NP |
6804 | free_percpu(sdd->sgc); |
6805 | sdd->sgc = NULL; | |
54ab4ff4 PZ |
6806 | } |
6807 | } | |
6808 | ||
2c402dc3 | 6809 | struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, |
4a850cbe VK |
6810 | const struct cpumask *cpu_map, struct sched_domain_attr *attr, |
6811 | struct sched_domain *child, int cpu) | |
2c402dc3 | 6812 | { |
143e1e28 | 6813 | struct sched_domain *sd = sd_init(tl, cpu); |
2c402dc3 | 6814 | if (!sd) |
d069b916 | 6815 | return child; |
2c402dc3 | 6816 | |
2c402dc3 | 6817 | cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); |
60495e77 PZ |
6818 | if (child) { |
6819 | sd->level = child->level + 1; | |
6820 | sched_domain_level_max = max(sched_domain_level_max, sd->level); | |
d069b916 | 6821 | child->parent = sd; |
c75e0128 | 6822 | sd->child = child; |
6ae72dff PZ |
6823 | |
6824 | if (!cpumask_subset(sched_domain_span(child), | |
6825 | sched_domain_span(sd))) { | |
6826 | pr_err("BUG: arch topology borken\n"); | |
6827 | #ifdef CONFIG_SCHED_DEBUG | |
6828 | pr_err(" the %s domain not a subset of the %s domain\n", | |
6829 | child->name, sd->name); | |
6830 | #endif | |
6831 | /* Fixup, ensure @sd has at least @child cpus. */ | |
6832 | cpumask_or(sched_domain_span(sd), | |
6833 | sched_domain_span(sd), | |
6834 | sched_domain_span(child)); | |
6835 | } | |
6836 | ||
60495e77 | 6837 | } |
a841f8ce | 6838 | set_domain_attribute(sd, attr); |
2c402dc3 PZ |
6839 | |
6840 | return sd; | |
6841 | } | |
6842 | ||
2109b99e AH |
6843 | /* |
6844 | * Build sched domains for a given set of cpus and attach the sched domains | |
6845 | * to the individual cpus | |
6846 | */ | |
dce840a0 PZ |
6847 | static int build_sched_domains(const struct cpumask *cpu_map, |
6848 | struct sched_domain_attr *attr) | |
2109b99e | 6849 | { |
1c632169 | 6850 | enum s_alloc alloc_state; |
dce840a0 | 6851 | struct sched_domain *sd; |
2109b99e | 6852 | struct s_data d; |
822ff793 | 6853 | int i, ret = -ENOMEM; |
9c1cfda2 | 6854 | |
2109b99e AH |
6855 | alloc_state = __visit_domain_allocation_hell(&d, cpu_map); |
6856 | if (alloc_state != sa_rootdomain) | |
6857 | goto error; | |
9c1cfda2 | 6858 | |
dce840a0 | 6859 | /* Set up domains for cpus specified by the cpu_map. */ |
abcd083a | 6860 | for_each_cpu(i, cpu_map) { |
eb7a74e6 PZ |
6861 | struct sched_domain_topology_level *tl; |
6862 | ||
3bd65a80 | 6863 | sd = NULL; |
27723a68 | 6864 | for_each_sd_topology(tl) { |
4a850cbe | 6865 | sd = build_sched_domain(tl, cpu_map, attr, sd, i); |
22da9569 VK |
6866 | if (tl == sched_domain_topology) |
6867 | *per_cpu_ptr(d.sd, i) = sd; | |
e3589f6c PZ |
6868 | if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) |
6869 | sd->flags |= SD_OVERLAP; | |
d110235d PZ |
6870 | if (cpumask_equal(cpu_map, sched_domain_span(sd))) |
6871 | break; | |
e3589f6c | 6872 | } |
dce840a0 PZ |
6873 | } |
6874 | ||
6875 | /* Build the groups for the domains */ | |
6876 | for_each_cpu(i, cpu_map) { | |
6877 | for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { | |
6878 | sd->span_weight = cpumask_weight(sched_domain_span(sd)); | |
e3589f6c PZ |
6879 | if (sd->flags & SD_OVERLAP) { |
6880 | if (build_overlap_sched_groups(sd, i)) | |
6881 | goto error; | |
6882 | } else { | |
6883 | if (build_sched_groups(sd, i)) | |
6884 | goto error; | |
6885 | } | |
1cf51902 | 6886 | } |
a06dadbe | 6887 | } |
9c1cfda2 | 6888 | |
ced549fa | 6889 | /* Calculate CPU capacity for physical packages and nodes */ |
a9c9a9b6 PZ |
6890 | for (i = nr_cpumask_bits-1; i >= 0; i--) { |
6891 | if (!cpumask_test_cpu(i, cpu_map)) | |
6892 | continue; | |
9c1cfda2 | 6893 | |
dce840a0 PZ |
6894 | for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { |
6895 | claim_allocations(i, sd); | |
63b2ca30 | 6896 | init_sched_groups_capacity(i, sd); |
dce840a0 | 6897 | } |
f712c0c7 | 6898 | } |
9c1cfda2 | 6899 | |
1da177e4 | 6900 | /* Attach the domains */ |
dce840a0 | 6901 | rcu_read_lock(); |
abcd083a | 6902 | for_each_cpu(i, cpu_map) { |
21d42ccf | 6903 | sd = *per_cpu_ptr(d.sd, i); |
49a02c51 | 6904 | cpu_attach_domain(sd, d.rd, i); |
1da177e4 | 6905 | } |
dce840a0 | 6906 | rcu_read_unlock(); |
51888ca2 | 6907 | |
822ff793 | 6908 | ret = 0; |
51888ca2 | 6909 | error: |
2109b99e | 6910 | __free_domain_allocs(&d, alloc_state, cpu_map); |
822ff793 | 6911 | return ret; |
1da177e4 | 6912 | } |
029190c5 | 6913 | |
acc3f5d7 | 6914 | static cpumask_var_t *doms_cur; /* current sched domains */ |
029190c5 | 6915 | static int ndoms_cur; /* number of sched domains in 'doms_cur' */ |
4285f594 IM |
6916 | static struct sched_domain_attr *dattr_cur; |
6917 | /* attribues of custom domains in 'doms_cur' */ | |
029190c5 PJ |
6918 | |
6919 | /* | |
6920 | * Special case: If a kmalloc of a doms_cur partition (array of | |
4212823f RR |
6921 | * cpumask) fails, then fallback to a single sched domain, |
6922 | * as determined by the single cpumask fallback_doms. | |
029190c5 | 6923 | */ |
4212823f | 6924 | static cpumask_var_t fallback_doms; |
029190c5 | 6925 | |
ee79d1bd HC |
6926 | /* |
6927 | * arch_update_cpu_topology lets virtualized architectures update the | |
6928 | * cpu core maps. It is supposed to return 1 if the topology changed | |
6929 | * or 0 if it stayed the same. | |
6930 | */ | |
52f5684c | 6931 | int __weak arch_update_cpu_topology(void) |
22e52b07 | 6932 | { |
ee79d1bd | 6933 | return 0; |
22e52b07 HC |
6934 | } |
6935 | ||
acc3f5d7 RR |
6936 | cpumask_var_t *alloc_sched_domains(unsigned int ndoms) |
6937 | { | |
6938 | int i; | |
6939 | cpumask_var_t *doms; | |
6940 | ||
6941 | doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); | |
6942 | if (!doms) | |
6943 | return NULL; | |
6944 | for (i = 0; i < ndoms; i++) { | |
6945 | if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { | |
6946 | free_sched_domains(doms, i); | |
6947 | return NULL; | |
6948 | } | |
6949 | } | |
6950 | return doms; | |
6951 | } | |
6952 | ||
6953 | void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) | |
6954 | { | |
6955 | unsigned int i; | |
6956 | for (i = 0; i < ndoms; i++) | |
6957 | free_cpumask_var(doms[i]); | |
6958 | kfree(doms); | |
6959 | } | |
6960 | ||
1a20ff27 | 6961 | /* |
41a2d6cf | 6962 | * Set up scheduler domains and groups. Callers must hold the hotplug lock. |
029190c5 PJ |
6963 | * For now this just excludes isolated cpus, but could be used to |
6964 | * exclude other special cases in the future. | |
1a20ff27 | 6965 | */ |
c4a8849a | 6966 | static int init_sched_domains(const struct cpumask *cpu_map) |
1a20ff27 | 6967 | { |
7378547f MM |
6968 | int err; |
6969 | ||
22e52b07 | 6970 | arch_update_cpu_topology(); |
029190c5 | 6971 | ndoms_cur = 1; |
acc3f5d7 | 6972 | doms_cur = alloc_sched_domains(ndoms_cur); |
029190c5 | 6973 | if (!doms_cur) |
acc3f5d7 RR |
6974 | doms_cur = &fallback_doms; |
6975 | cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); | |
dce840a0 | 6976 | err = build_sched_domains(doms_cur[0], NULL); |
6382bc90 | 6977 | register_sched_domain_sysctl(); |
7378547f MM |
6978 | |
6979 | return err; | |
1a20ff27 DG |
6980 | } |
6981 | ||
1a20ff27 DG |
6982 | /* |
6983 | * Detach sched domains from a group of cpus specified in cpu_map | |
6984 | * These cpus will now be attached to the NULL domain | |
6985 | */ | |
96f874e2 | 6986 | static void detach_destroy_domains(const struct cpumask *cpu_map) |
1a20ff27 DG |
6987 | { |
6988 | int i; | |
6989 | ||
dce840a0 | 6990 | rcu_read_lock(); |
abcd083a | 6991 | for_each_cpu(i, cpu_map) |
57d885fe | 6992 | cpu_attach_domain(NULL, &def_root_domain, i); |
dce840a0 | 6993 | rcu_read_unlock(); |
1a20ff27 DG |
6994 | } |
6995 | ||
1d3504fc HS |
6996 | /* handle null as "default" */ |
6997 | static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, | |
6998 | struct sched_domain_attr *new, int idx_new) | |
6999 | { | |
7000 | struct sched_domain_attr tmp; | |
7001 | ||
7002 | /* fast path */ | |
7003 | if (!new && !cur) | |
7004 | return 1; | |
7005 | ||
7006 | tmp = SD_ATTR_INIT; | |
7007 | return !memcmp(cur ? (cur + idx_cur) : &tmp, | |
7008 | new ? (new + idx_new) : &tmp, | |
7009 | sizeof(struct sched_domain_attr)); | |
7010 | } | |
7011 | ||
029190c5 PJ |
7012 | /* |
7013 | * Partition sched domains as specified by the 'ndoms_new' | |
41a2d6cf | 7014 | * cpumasks in the array doms_new[] of cpumasks. This compares |
029190c5 PJ |
7015 | * doms_new[] to the current sched domain partitioning, doms_cur[]. |
7016 | * It destroys each deleted domain and builds each new domain. | |
7017 | * | |
acc3f5d7 | 7018 | * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. |
41a2d6cf IM |
7019 | * The masks don't intersect (don't overlap.) We should setup one |
7020 | * sched domain for each mask. CPUs not in any of the cpumasks will | |
7021 | * not be load balanced. If the same cpumask appears both in the | |
029190c5 PJ |
7022 | * current 'doms_cur' domains and in the new 'doms_new', we can leave |
7023 | * it as it is. | |
7024 | * | |
acc3f5d7 RR |
7025 | * The passed in 'doms_new' should be allocated using |
7026 | * alloc_sched_domains. This routine takes ownership of it and will | |
7027 | * free_sched_domains it when done with it. If the caller failed the | |
7028 | * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, | |
7029 | * and partition_sched_domains() will fallback to the single partition | |
7030 | * 'fallback_doms', it also forces the domains to be rebuilt. | |
029190c5 | 7031 | * |
96f874e2 | 7032 | * If doms_new == NULL it will be replaced with cpu_online_mask. |
700018e0 LZ |
7033 | * ndoms_new == 0 is a special case for destroying existing domains, |
7034 | * and it will not create the default domain. | |
dfb512ec | 7035 | * |
029190c5 PJ |
7036 | * Call with hotplug lock held |
7037 | */ | |
acc3f5d7 | 7038 | void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], |
1d3504fc | 7039 | struct sched_domain_attr *dattr_new) |
029190c5 | 7040 | { |
dfb512ec | 7041 | int i, j, n; |
d65bd5ec | 7042 | int new_topology; |
029190c5 | 7043 | |
712555ee | 7044 | mutex_lock(&sched_domains_mutex); |
a1835615 | 7045 | |
7378547f MM |
7046 | /* always unregister in case we don't destroy any domains */ |
7047 | unregister_sched_domain_sysctl(); | |
7048 | ||
d65bd5ec HC |
7049 | /* Let architecture update cpu core mappings. */ |
7050 | new_topology = arch_update_cpu_topology(); | |
7051 | ||
dfb512ec | 7052 | n = doms_new ? ndoms_new : 0; |
029190c5 PJ |
7053 | |
7054 | /* Destroy deleted domains */ | |
7055 | for (i = 0; i < ndoms_cur; i++) { | |
d65bd5ec | 7056 | for (j = 0; j < n && !new_topology; j++) { |
acc3f5d7 | 7057 | if (cpumask_equal(doms_cur[i], doms_new[j]) |
1d3504fc | 7058 | && dattrs_equal(dattr_cur, i, dattr_new, j)) |
029190c5 PJ |
7059 | goto match1; |
7060 | } | |
7061 | /* no match - a current sched domain not in new doms_new[] */ | |
acc3f5d7 | 7062 | detach_destroy_domains(doms_cur[i]); |
029190c5 PJ |
7063 | match1: |
7064 | ; | |
7065 | } | |
7066 | ||
c8d2d47a | 7067 | n = ndoms_cur; |
e761b772 | 7068 | if (doms_new == NULL) { |
c8d2d47a | 7069 | n = 0; |
acc3f5d7 | 7070 | doms_new = &fallback_doms; |
6ad4c188 | 7071 | cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); |
faa2f98f | 7072 | WARN_ON_ONCE(dattr_new); |
e761b772 MK |
7073 | } |
7074 | ||
029190c5 PJ |
7075 | /* Build new domains */ |
7076 | for (i = 0; i < ndoms_new; i++) { | |
c8d2d47a | 7077 | for (j = 0; j < n && !new_topology; j++) { |
acc3f5d7 | 7078 | if (cpumask_equal(doms_new[i], doms_cur[j]) |
1d3504fc | 7079 | && dattrs_equal(dattr_new, i, dattr_cur, j)) |
029190c5 PJ |
7080 | goto match2; |
7081 | } | |
7082 | /* no match - add a new doms_new */ | |
dce840a0 | 7083 | build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); |
029190c5 PJ |
7084 | match2: |
7085 | ; | |
7086 | } | |
7087 | ||
7088 | /* Remember the new sched domains */ | |
acc3f5d7 RR |
7089 | if (doms_cur != &fallback_doms) |
7090 | free_sched_domains(doms_cur, ndoms_cur); | |
1d3504fc | 7091 | kfree(dattr_cur); /* kfree(NULL) is safe */ |
029190c5 | 7092 | doms_cur = doms_new; |
1d3504fc | 7093 | dattr_cur = dattr_new; |
029190c5 | 7094 | ndoms_cur = ndoms_new; |
7378547f MM |
7095 | |
7096 | register_sched_domain_sysctl(); | |
a1835615 | 7097 | |
712555ee | 7098 | mutex_unlock(&sched_domains_mutex); |
029190c5 PJ |
7099 | } |
7100 | ||
d35be8ba SB |
7101 | static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ |
7102 | ||
1da177e4 | 7103 | /* |
3a101d05 TH |
7104 | * Update cpusets according to cpu_active mask. If cpusets are |
7105 | * disabled, cpuset_update_active_cpus() becomes a simple wrapper | |
7106 | * around partition_sched_domains(). | |
d35be8ba SB |
7107 | * |
7108 | * If we come here as part of a suspend/resume, don't touch cpusets because we | |
7109 | * want to restore it back to its original state upon resume anyway. | |
1da177e4 | 7110 | */ |
40190a78 | 7111 | static void cpuset_cpu_active(void) |
e761b772 | 7112 | { |
40190a78 | 7113 | if (cpuhp_tasks_frozen) { |
d35be8ba SB |
7114 | /* |
7115 | * num_cpus_frozen tracks how many CPUs are involved in suspend | |
7116 | * resume sequence. As long as this is not the last online | |
7117 | * operation in the resume sequence, just build a single sched | |
7118 | * domain, ignoring cpusets. | |
7119 | */ | |
7120 | num_cpus_frozen--; | |
7121 | if (likely(num_cpus_frozen)) { | |
7122 | partition_sched_domains(1, NULL, NULL); | |
135fb3e1 | 7123 | return; |
d35be8ba | 7124 | } |
d35be8ba SB |
7125 | /* |
7126 | * This is the last CPU online operation. So fall through and | |
7127 | * restore the original sched domains by considering the | |
7128 | * cpuset configurations. | |
7129 | */ | |
3a101d05 | 7130 | } |
135fb3e1 | 7131 | cpuset_update_active_cpus(true); |
3a101d05 | 7132 | } |
e761b772 | 7133 | |
40190a78 | 7134 | static int cpuset_cpu_inactive(unsigned int cpu) |
3a101d05 | 7135 | { |
3c18d447 | 7136 | unsigned long flags; |
3c18d447 | 7137 | struct dl_bw *dl_b; |
533445c6 OS |
7138 | bool overflow; |
7139 | int cpus; | |
3c18d447 | 7140 | |
40190a78 | 7141 | if (!cpuhp_tasks_frozen) { |
533445c6 OS |
7142 | rcu_read_lock_sched(); |
7143 | dl_b = dl_bw_of(cpu); | |
3c18d447 | 7144 | |
533445c6 OS |
7145 | raw_spin_lock_irqsave(&dl_b->lock, flags); |
7146 | cpus = dl_bw_cpus(cpu); | |
7147 | overflow = __dl_overflow(dl_b, cpus, 0, 0); | |
7148 | raw_spin_unlock_irqrestore(&dl_b->lock, flags); | |
3c18d447 | 7149 | |
533445c6 | 7150 | rcu_read_unlock_sched(); |
3c18d447 | 7151 | |
533445c6 | 7152 | if (overflow) |
135fb3e1 | 7153 | return -EBUSY; |
7ddf96b0 | 7154 | cpuset_update_active_cpus(false); |
135fb3e1 | 7155 | } else { |
d35be8ba SB |
7156 | num_cpus_frozen++; |
7157 | partition_sched_domains(1, NULL, NULL); | |
e761b772 | 7158 | } |
135fb3e1 | 7159 | return 0; |
e761b772 | 7160 | } |
e761b772 | 7161 | |
40190a78 | 7162 | int sched_cpu_activate(unsigned int cpu) |
135fb3e1 | 7163 | { |
7d976699 TG |
7164 | struct rq *rq = cpu_rq(cpu); |
7165 | unsigned long flags; | |
7166 | ||
40190a78 | 7167 | set_cpu_active(cpu, true); |
135fb3e1 | 7168 | |
40190a78 | 7169 | if (sched_smp_initialized) { |
135fb3e1 | 7170 | sched_domains_numa_masks_set(cpu); |
40190a78 | 7171 | cpuset_cpu_active(); |
e761b772 | 7172 | } |
7d976699 TG |
7173 | |
7174 | /* | |
7175 | * Put the rq online, if not already. This happens: | |
7176 | * | |
7177 | * 1) In the early boot process, because we build the real domains | |
7178 | * after all cpus have been brought up. | |
7179 | * | |
7180 | * 2) At runtime, if cpuset_cpu_active() fails to rebuild the | |
7181 | * domains. | |
7182 | */ | |
7183 | raw_spin_lock_irqsave(&rq->lock, flags); | |
7184 | if (rq->rd) { | |
7185 | BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); | |
7186 | set_rq_online(rq); | |
7187 | } | |
7188 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
7189 | ||
7190 | update_max_interval(); | |
7191 | ||
40190a78 | 7192 | return 0; |
135fb3e1 TG |
7193 | } |
7194 | ||
40190a78 | 7195 | int sched_cpu_deactivate(unsigned int cpu) |
135fb3e1 | 7196 | { |
135fb3e1 TG |
7197 | int ret; |
7198 | ||
40190a78 | 7199 | set_cpu_active(cpu, false); |
b2454caa PZ |
7200 | /* |
7201 | * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU | |
7202 | * users of this state to go away such that all new such users will | |
7203 | * observe it. | |
7204 | * | |
7205 | * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might | |
7206 | * not imply sync_sched(), so wait for both. | |
7207 | * | |
7208 | * Do sync before park smpboot threads to take care the rcu boost case. | |
7209 | */ | |
7210 | if (IS_ENABLED(CONFIG_PREEMPT)) | |
7211 | synchronize_rcu_mult(call_rcu, call_rcu_sched); | |
7212 | else | |
7213 | synchronize_rcu(); | |
40190a78 TG |
7214 | |
7215 | if (!sched_smp_initialized) | |
7216 | return 0; | |
7217 | ||
7218 | ret = cpuset_cpu_inactive(cpu); | |
7219 | if (ret) { | |
7220 | set_cpu_active(cpu, true); | |
7221 | return ret; | |
135fb3e1 | 7222 | } |
40190a78 TG |
7223 | sched_domains_numa_masks_clear(cpu); |
7224 | return 0; | |
135fb3e1 TG |
7225 | } |
7226 | ||
94baf7a5 TG |
7227 | static void sched_rq_cpu_starting(unsigned int cpu) |
7228 | { | |
7229 | struct rq *rq = cpu_rq(cpu); | |
7230 | ||
7231 | rq->calc_load_update = calc_load_update; | |
7232 | account_reset_rq(rq); | |
7233 | update_max_interval(); | |
7234 | } | |
7235 | ||
135fb3e1 TG |
7236 | int sched_cpu_starting(unsigned int cpu) |
7237 | { | |
7238 | set_cpu_rq_start_time(cpu); | |
94baf7a5 | 7239 | sched_rq_cpu_starting(cpu); |
135fb3e1 | 7240 | return 0; |
e761b772 | 7241 | } |
e761b772 | 7242 | |
f2785ddb TG |
7243 | #ifdef CONFIG_HOTPLUG_CPU |
7244 | int sched_cpu_dying(unsigned int cpu) | |
7245 | { | |
7246 | struct rq *rq = cpu_rq(cpu); | |
7247 | unsigned long flags; | |
7248 | ||
7249 | /* Handle pending wakeups and then migrate everything off */ | |
7250 | sched_ttwu_pending(); | |
7251 | raw_spin_lock_irqsave(&rq->lock, flags); | |
7252 | if (rq->rd) { | |
7253 | BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); | |
7254 | set_rq_offline(rq); | |
7255 | } | |
7256 | migrate_tasks(rq); | |
7257 | BUG_ON(rq->nr_running != 1); | |
7258 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
7259 | calc_load_migrate(rq); | |
7260 | update_max_interval(); | |
20a5c8cc | 7261 | nohz_balance_exit_idle(cpu); |
e5ef27d0 | 7262 | hrtick_clear(rq); |
f2785ddb TG |
7263 | return 0; |
7264 | } | |
7265 | #endif | |
7266 | ||
1da177e4 LT |
7267 | void __init sched_init_smp(void) |
7268 | { | |
dcc30a35 RR |
7269 | cpumask_var_t non_isolated_cpus; |
7270 | ||
7271 | alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); | |
cb5fd13f | 7272 | alloc_cpumask_var(&fallback_doms, GFP_KERNEL); |
5c1e1767 | 7273 | |
cb83b629 PZ |
7274 | sched_init_numa(); |
7275 | ||
6acce3ef PZ |
7276 | /* |
7277 | * There's no userspace yet to cause hotplug operations; hence all the | |
7278 | * cpu masks are stable and all blatant races in the below code cannot | |
7279 | * happen. | |
7280 | */ | |
712555ee | 7281 | mutex_lock(&sched_domains_mutex); |
c4a8849a | 7282 | init_sched_domains(cpu_active_mask); |
dcc30a35 RR |
7283 | cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); |
7284 | if (cpumask_empty(non_isolated_cpus)) | |
7285 | cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); | |
712555ee | 7286 | mutex_unlock(&sched_domains_mutex); |
e761b772 | 7287 | |
5c1e1767 | 7288 | /* Move init over to a non-isolated CPU */ |
dcc30a35 | 7289 | if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) |
5c1e1767 | 7290 | BUG(); |
19978ca6 | 7291 | sched_init_granularity(); |
dcc30a35 | 7292 | free_cpumask_var(non_isolated_cpus); |
4212823f | 7293 | |
0e3900e6 | 7294 | init_sched_rt_class(); |
1baca4ce | 7295 | init_sched_dl_class(); |
e26fbffd | 7296 | sched_smp_initialized = true; |
1da177e4 | 7297 | } |
e26fbffd TG |
7298 | |
7299 | static int __init migration_init(void) | |
7300 | { | |
94baf7a5 | 7301 | sched_rq_cpu_starting(smp_processor_id()); |
e26fbffd | 7302 | return 0; |
1da177e4 | 7303 | } |
e26fbffd TG |
7304 | early_initcall(migration_init); |
7305 | ||
1da177e4 LT |
7306 | #else |
7307 | void __init sched_init_smp(void) | |
7308 | { | |
19978ca6 | 7309 | sched_init_granularity(); |
1da177e4 LT |
7310 | } |
7311 | #endif /* CONFIG_SMP */ | |
7312 | ||
7313 | int in_sched_functions(unsigned long addr) | |
7314 | { | |
1da177e4 LT |
7315 | return in_lock_functions(addr) || |
7316 | (addr >= (unsigned long)__sched_text_start | |
7317 | && addr < (unsigned long)__sched_text_end); | |
7318 | } | |
7319 | ||
029632fb | 7320 | #ifdef CONFIG_CGROUP_SCHED |
27b4b931 LZ |
7321 | /* |
7322 | * Default task group. | |
7323 | * Every task in system belongs to this group at bootup. | |
7324 | */ | |
029632fb | 7325 | struct task_group root_task_group; |
35cf4e50 | 7326 | LIST_HEAD(task_groups); |
b0367629 WL |
7327 | |
7328 | /* Cacheline aligned slab cache for task_group */ | |
7329 | static struct kmem_cache *task_group_cache __read_mostly; | |
052f1dc7 | 7330 | #endif |
6f505b16 | 7331 | |
e6252c3e | 7332 | DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); |
6f505b16 | 7333 | |
1da177e4 LT |
7334 | void __init sched_init(void) |
7335 | { | |
dd41f596 | 7336 | int i, j; |
434d53b0 MT |
7337 | unsigned long alloc_size = 0, ptr; |
7338 | ||
7339 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
7340 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); | |
7341 | #endif | |
7342 | #ifdef CONFIG_RT_GROUP_SCHED | |
7343 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); | |
7344 | #endif | |
434d53b0 | 7345 | if (alloc_size) { |
36b7b6d4 | 7346 | ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); |
434d53b0 MT |
7347 | |
7348 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
07e06b01 | 7349 | root_task_group.se = (struct sched_entity **)ptr; |
434d53b0 MT |
7350 | ptr += nr_cpu_ids * sizeof(void **); |
7351 | ||
07e06b01 | 7352 | root_task_group.cfs_rq = (struct cfs_rq **)ptr; |
434d53b0 | 7353 | ptr += nr_cpu_ids * sizeof(void **); |
eff766a6 | 7354 | |
6d6bc0ad | 7355 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
434d53b0 | 7356 | #ifdef CONFIG_RT_GROUP_SCHED |
07e06b01 | 7357 | root_task_group.rt_se = (struct sched_rt_entity **)ptr; |
434d53b0 MT |
7358 | ptr += nr_cpu_ids * sizeof(void **); |
7359 | ||
07e06b01 | 7360 | root_task_group.rt_rq = (struct rt_rq **)ptr; |
eff766a6 PZ |
7361 | ptr += nr_cpu_ids * sizeof(void **); |
7362 | ||
6d6bc0ad | 7363 | #endif /* CONFIG_RT_GROUP_SCHED */ |
b74e6278 | 7364 | } |
df7c8e84 | 7365 | #ifdef CONFIG_CPUMASK_OFFSTACK |
b74e6278 AT |
7366 | for_each_possible_cpu(i) { |
7367 | per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node( | |
7368 | cpumask_size(), GFP_KERNEL, cpu_to_node(i)); | |
434d53b0 | 7369 | } |
b74e6278 | 7370 | #endif /* CONFIG_CPUMASK_OFFSTACK */ |
dd41f596 | 7371 | |
332ac17e DF |
7372 | init_rt_bandwidth(&def_rt_bandwidth, |
7373 | global_rt_period(), global_rt_runtime()); | |
7374 | init_dl_bandwidth(&def_dl_bandwidth, | |
1724813d | 7375 | global_rt_period(), global_rt_runtime()); |
332ac17e | 7376 | |
57d885fe GH |
7377 | #ifdef CONFIG_SMP |
7378 | init_defrootdomain(); | |
7379 | #endif | |
7380 | ||
d0b27fa7 | 7381 | #ifdef CONFIG_RT_GROUP_SCHED |
07e06b01 | 7382 | init_rt_bandwidth(&root_task_group.rt_bandwidth, |
d0b27fa7 | 7383 | global_rt_period(), global_rt_runtime()); |
6d6bc0ad | 7384 | #endif /* CONFIG_RT_GROUP_SCHED */ |
d0b27fa7 | 7385 | |
7c941438 | 7386 | #ifdef CONFIG_CGROUP_SCHED |
b0367629 WL |
7387 | task_group_cache = KMEM_CACHE(task_group, 0); |
7388 | ||
07e06b01 YZ |
7389 | list_add(&root_task_group.list, &task_groups); |
7390 | INIT_LIST_HEAD(&root_task_group.children); | |
f4d6f6c2 | 7391 | INIT_LIST_HEAD(&root_task_group.siblings); |
5091faa4 | 7392 | autogroup_init(&init_task); |
7c941438 | 7393 | #endif /* CONFIG_CGROUP_SCHED */ |
6f505b16 | 7394 | |
0a945022 | 7395 | for_each_possible_cpu(i) { |
70b97a7f | 7396 | struct rq *rq; |
1da177e4 LT |
7397 | |
7398 | rq = cpu_rq(i); | |
05fa785c | 7399 | raw_spin_lock_init(&rq->lock); |
7897986b | 7400 | rq->nr_running = 0; |
dce48a84 TG |
7401 | rq->calc_load_active = 0; |
7402 | rq->calc_load_update = jiffies + LOAD_FREQ; | |
acb5a9ba | 7403 | init_cfs_rq(&rq->cfs); |
07c54f7a AV |
7404 | init_rt_rq(&rq->rt); |
7405 | init_dl_rq(&rq->dl); | |
dd41f596 | 7406 | #ifdef CONFIG_FAIR_GROUP_SCHED |
029632fb | 7407 | root_task_group.shares = ROOT_TASK_GROUP_LOAD; |
6f505b16 | 7408 | INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); |
354d60c2 | 7409 | /* |
07e06b01 | 7410 | * How much cpu bandwidth does root_task_group get? |
354d60c2 DG |
7411 | * |
7412 | * In case of task-groups formed thr' the cgroup filesystem, it | |
7413 | * gets 100% of the cpu resources in the system. This overall | |
7414 | * system cpu resource is divided among the tasks of | |
07e06b01 | 7415 | * root_task_group and its child task-groups in a fair manner, |
354d60c2 DG |
7416 | * based on each entity's (task or task-group's) weight |
7417 | * (se->load.weight). | |
7418 | * | |
07e06b01 | 7419 | * In other words, if root_task_group has 10 tasks of weight |
354d60c2 DG |
7420 | * 1024) and two child groups A0 and A1 (of weight 1024 each), |
7421 | * then A0's share of the cpu resource is: | |
7422 | * | |
0d905bca | 7423 | * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% |
354d60c2 | 7424 | * |
07e06b01 YZ |
7425 | * We achieve this by letting root_task_group's tasks sit |
7426 | * directly in rq->cfs (i.e root_task_group->se[] = NULL). | |
354d60c2 | 7427 | */ |
ab84d31e | 7428 | init_cfs_bandwidth(&root_task_group.cfs_bandwidth); |
07e06b01 | 7429 | init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); |
354d60c2 DG |
7430 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
7431 | ||
7432 | rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; | |
052f1dc7 | 7433 | #ifdef CONFIG_RT_GROUP_SCHED |
07e06b01 | 7434 | init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); |
dd41f596 | 7435 | #endif |
1da177e4 | 7436 | |
dd41f596 IM |
7437 | for (j = 0; j < CPU_LOAD_IDX_MAX; j++) |
7438 | rq->cpu_load[j] = 0; | |
fdf3e95d | 7439 | |
1da177e4 | 7440 | #ifdef CONFIG_SMP |
41c7ce9a | 7441 | rq->sd = NULL; |
57d885fe | 7442 | rq->rd = NULL; |
ca6d75e6 | 7443 | rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE; |
e3fca9e7 | 7444 | rq->balance_callback = NULL; |
1da177e4 | 7445 | rq->active_balance = 0; |
dd41f596 | 7446 | rq->next_balance = jiffies; |
1da177e4 | 7447 | rq->push_cpu = 0; |
0a2966b4 | 7448 | rq->cpu = i; |
1f11eb6a | 7449 | rq->online = 0; |
eae0c9df MG |
7450 | rq->idle_stamp = 0; |
7451 | rq->avg_idle = 2*sysctl_sched_migration_cost; | |
9bd721c5 | 7452 | rq->max_idle_balance_cost = sysctl_sched_migration_cost; |
367456c7 PZ |
7453 | |
7454 | INIT_LIST_HEAD(&rq->cfs_tasks); | |
7455 | ||
dc938520 | 7456 | rq_attach_root(rq, &def_root_domain); |
3451d024 | 7457 | #ifdef CONFIG_NO_HZ_COMMON |
9fd81dd5 | 7458 | rq->last_load_update_tick = jiffies; |
1c792db7 | 7459 | rq->nohz_flags = 0; |
83cd4fe2 | 7460 | #endif |
265f22a9 FW |
7461 | #ifdef CONFIG_NO_HZ_FULL |
7462 | rq->last_sched_tick = 0; | |
7463 | #endif | |
9fd81dd5 | 7464 | #endif /* CONFIG_SMP */ |
8f4d37ec | 7465 | init_rq_hrtick(rq); |
1da177e4 | 7466 | atomic_set(&rq->nr_iowait, 0); |
1da177e4 LT |
7467 | } |
7468 | ||
2dd73a4f | 7469 | set_load_weight(&init_task); |
b50f60ce | 7470 | |
e107be36 AK |
7471 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
7472 | INIT_HLIST_HEAD(&init_task.preempt_notifiers); | |
7473 | #endif | |
7474 | ||
1da177e4 LT |
7475 | /* |
7476 | * The boot idle thread does lazy MMU switching as well: | |
7477 | */ | |
7478 | atomic_inc(&init_mm.mm_count); | |
7479 | enter_lazy_tlb(&init_mm, current); | |
7480 | ||
1b537c7d YD |
7481 | /* |
7482 | * During early bootup we pretend to be a normal task: | |
7483 | */ | |
7484 | current->sched_class = &fair_sched_class; | |
7485 | ||
1da177e4 LT |
7486 | /* |
7487 | * Make us the idle thread. Technically, schedule() should not be | |
7488 | * called from this thread, however somewhere below it might be, | |
7489 | * but because we are the idle thread, we just pick up running again | |
7490 | * when this runqueue becomes "idle". | |
7491 | */ | |
7492 | init_idle(current, smp_processor_id()); | |
dce48a84 TG |
7493 | |
7494 | calc_load_update = jiffies + LOAD_FREQ; | |
7495 | ||
bf4d83f6 | 7496 | #ifdef CONFIG_SMP |
4cb98839 | 7497 | zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); |
bdddd296 RR |
7498 | /* May be allocated at isolcpus cmdline parse time */ |
7499 | if (cpu_isolated_map == NULL) | |
7500 | zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); | |
29d5e047 | 7501 | idle_thread_set_boot_cpu(); |
9cf7243d | 7502 | set_cpu_rq_start_time(smp_processor_id()); |
029632fb PZ |
7503 | #endif |
7504 | init_sched_fair_class(); | |
6a7b3dc3 | 7505 | |
4698f88c JP |
7506 | init_schedstats(); |
7507 | ||
6892b75e | 7508 | scheduler_running = 1; |
1da177e4 LT |
7509 | } |
7510 | ||
d902db1e | 7511 | #ifdef CONFIG_DEBUG_ATOMIC_SLEEP |
e4aafea2 FW |
7512 | static inline int preempt_count_equals(int preempt_offset) |
7513 | { | |
da7142e2 | 7514 | int nested = preempt_count() + rcu_preempt_depth(); |
e4aafea2 | 7515 | |
4ba8216c | 7516 | return (nested == preempt_offset); |
e4aafea2 FW |
7517 | } |
7518 | ||
d894837f | 7519 | void __might_sleep(const char *file, int line, int preempt_offset) |
1da177e4 | 7520 | { |
8eb23b9f PZ |
7521 | /* |
7522 | * Blocking primitives will set (and therefore destroy) current->state, | |
7523 | * since we will exit with TASK_RUNNING make sure we enter with it, | |
7524 | * otherwise we will destroy state. | |
7525 | */ | |
00845eb9 | 7526 | WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change, |
8eb23b9f PZ |
7527 | "do not call blocking ops when !TASK_RUNNING; " |
7528 | "state=%lx set at [<%p>] %pS\n", | |
7529 | current->state, | |
7530 | (void *)current->task_state_change, | |
00845eb9 | 7531 | (void *)current->task_state_change); |
8eb23b9f | 7532 | |
3427445a PZ |
7533 | ___might_sleep(file, line, preempt_offset); |
7534 | } | |
7535 | EXPORT_SYMBOL(__might_sleep); | |
7536 | ||
7537 | void ___might_sleep(const char *file, int line, int preempt_offset) | |
1da177e4 | 7538 | { |
1da177e4 LT |
7539 | static unsigned long prev_jiffy; /* ratelimiting */ |
7540 | ||
b3fbab05 | 7541 | rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ |
db273be2 TG |
7542 | if ((preempt_count_equals(preempt_offset) && !irqs_disabled() && |
7543 | !is_idle_task(current)) || | |
e4aafea2 | 7544 | system_state != SYSTEM_RUNNING || oops_in_progress) |
aef745fc IM |
7545 | return; |
7546 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) | |
7547 | return; | |
7548 | prev_jiffy = jiffies; | |
7549 | ||
3df0fc5b PZ |
7550 | printk(KERN_ERR |
7551 | "BUG: sleeping function called from invalid context at %s:%d\n", | |
7552 | file, line); | |
7553 | printk(KERN_ERR | |
7554 | "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", | |
7555 | in_atomic(), irqs_disabled(), | |
7556 | current->pid, current->comm); | |
aef745fc | 7557 | |
a8b686b3 ES |
7558 | if (task_stack_end_corrupted(current)) |
7559 | printk(KERN_EMERG "Thread overran stack, or stack corrupted\n"); | |
7560 | ||
aef745fc IM |
7561 | debug_show_held_locks(current); |
7562 | if (irqs_disabled()) | |
7563 | print_irqtrace_events(current); | |
8f47b187 TG |
7564 | #ifdef CONFIG_DEBUG_PREEMPT |
7565 | if (!preempt_count_equals(preempt_offset)) { | |
7566 | pr_err("Preemption disabled at:"); | |
7567 | print_ip_sym(current->preempt_disable_ip); | |
7568 | pr_cont("\n"); | |
7569 | } | |
7570 | #endif | |
aef745fc | 7571 | dump_stack(); |
1da177e4 | 7572 | } |
3427445a | 7573 | EXPORT_SYMBOL(___might_sleep); |
1da177e4 LT |
7574 | #endif |
7575 | ||
7576 | #ifdef CONFIG_MAGIC_SYSRQ | |
dbc7f069 | 7577 | void normalize_rt_tasks(void) |
3a5e4dc1 | 7578 | { |
dbc7f069 | 7579 | struct task_struct *g, *p; |
d50dde5a DF |
7580 | struct sched_attr attr = { |
7581 | .sched_policy = SCHED_NORMAL, | |
7582 | }; | |
1da177e4 | 7583 | |
3472eaa1 | 7584 | read_lock(&tasklist_lock); |
5d07f420 | 7585 | for_each_process_thread(g, p) { |
178be793 IM |
7586 | /* |
7587 | * Only normalize user tasks: | |
7588 | */ | |
3472eaa1 | 7589 | if (p->flags & PF_KTHREAD) |
178be793 IM |
7590 | continue; |
7591 | ||
6cfb0d5d | 7592 | p->se.exec_start = 0; |
6cfb0d5d | 7593 | #ifdef CONFIG_SCHEDSTATS |
41acab88 LDM |
7594 | p->se.statistics.wait_start = 0; |
7595 | p->se.statistics.sleep_start = 0; | |
7596 | p->se.statistics.block_start = 0; | |
6cfb0d5d | 7597 | #endif |
dd41f596 | 7598 | |
aab03e05 | 7599 | if (!dl_task(p) && !rt_task(p)) { |
dd41f596 IM |
7600 | /* |
7601 | * Renice negative nice level userspace | |
7602 | * tasks back to 0: | |
7603 | */ | |
3472eaa1 | 7604 | if (task_nice(p) < 0) |
dd41f596 | 7605 | set_user_nice(p, 0); |
1da177e4 | 7606 | continue; |
dd41f596 | 7607 | } |
1da177e4 | 7608 | |
dbc7f069 | 7609 | __sched_setscheduler(p, &attr, false, false); |
5d07f420 | 7610 | } |
3472eaa1 | 7611 | read_unlock(&tasklist_lock); |
1da177e4 LT |
7612 | } |
7613 | ||
7614 | #endif /* CONFIG_MAGIC_SYSRQ */ | |
1df5c10a | 7615 | |
67fc4e0c | 7616 | #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) |
1df5c10a | 7617 | /* |
67fc4e0c | 7618 | * These functions are only useful for the IA64 MCA handling, or kdb. |
1df5c10a LT |
7619 | * |
7620 | * They can only be called when the whole system has been | |
7621 | * stopped - every CPU needs to be quiescent, and no scheduling | |
7622 | * activity can take place. Using them for anything else would | |
7623 | * be a serious bug, and as a result, they aren't even visible | |
7624 | * under any other configuration. | |
7625 | */ | |
7626 | ||
7627 | /** | |
7628 | * curr_task - return the current task for a given cpu. | |
7629 | * @cpu: the processor in question. | |
7630 | * | |
7631 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
e69f6186 YB |
7632 | * |
7633 | * Return: The current task for @cpu. | |
1df5c10a | 7634 | */ |
36c8b586 | 7635 | struct task_struct *curr_task(int cpu) |
1df5c10a LT |
7636 | { |
7637 | return cpu_curr(cpu); | |
7638 | } | |
7639 | ||
67fc4e0c JW |
7640 | #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ |
7641 | ||
7642 | #ifdef CONFIG_IA64 | |
1df5c10a LT |
7643 | /** |
7644 | * set_curr_task - set the current task for a given cpu. | |
7645 | * @cpu: the processor in question. | |
7646 | * @p: the task pointer to set. | |
7647 | * | |
7648 | * Description: This function must only be used when non-maskable interrupts | |
41a2d6cf IM |
7649 | * are serviced on a separate stack. It allows the architecture to switch the |
7650 | * notion of the current task on a cpu in a non-blocking manner. This function | |
1df5c10a LT |
7651 | * must be called with all CPU's synchronized, and interrupts disabled, the |
7652 | * and caller must save the original value of the current task (see | |
7653 | * curr_task() above) and restore that value before reenabling interrupts and | |
7654 | * re-starting the system. | |
7655 | * | |
7656 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
7657 | */ | |
36c8b586 | 7658 | void set_curr_task(int cpu, struct task_struct *p) |
1df5c10a LT |
7659 | { |
7660 | cpu_curr(cpu) = p; | |
7661 | } | |
7662 | ||
7663 | #endif | |
29f59db3 | 7664 | |
7c941438 | 7665 | #ifdef CONFIG_CGROUP_SCHED |
029632fb PZ |
7666 | /* task_group_lock serializes the addition/removal of task groups */ |
7667 | static DEFINE_SPINLOCK(task_group_lock); | |
7668 | ||
2f5177f0 | 7669 | static void sched_free_group(struct task_group *tg) |
bccbe08a PZ |
7670 | { |
7671 | free_fair_sched_group(tg); | |
7672 | free_rt_sched_group(tg); | |
e9aa1dd1 | 7673 | autogroup_free(tg); |
b0367629 | 7674 | kmem_cache_free(task_group_cache, tg); |
bccbe08a PZ |
7675 | } |
7676 | ||
7677 | /* allocate runqueue etc for a new task group */ | |
ec7dc8ac | 7678 | struct task_group *sched_create_group(struct task_group *parent) |
bccbe08a PZ |
7679 | { |
7680 | struct task_group *tg; | |
bccbe08a | 7681 | |
b0367629 | 7682 | tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO); |
bccbe08a PZ |
7683 | if (!tg) |
7684 | return ERR_PTR(-ENOMEM); | |
7685 | ||
ec7dc8ac | 7686 | if (!alloc_fair_sched_group(tg, parent)) |
bccbe08a PZ |
7687 | goto err; |
7688 | ||
ec7dc8ac | 7689 | if (!alloc_rt_sched_group(tg, parent)) |
bccbe08a PZ |
7690 | goto err; |
7691 | ||
ace783b9 LZ |
7692 | return tg; |
7693 | ||
7694 | err: | |
2f5177f0 | 7695 | sched_free_group(tg); |
ace783b9 LZ |
7696 | return ERR_PTR(-ENOMEM); |
7697 | } | |
7698 | ||
7699 | void sched_online_group(struct task_group *tg, struct task_group *parent) | |
7700 | { | |
7701 | unsigned long flags; | |
7702 | ||
8ed36996 | 7703 | spin_lock_irqsave(&task_group_lock, flags); |
6f505b16 | 7704 | list_add_rcu(&tg->list, &task_groups); |
f473aa5e PZ |
7705 | |
7706 | WARN_ON(!parent); /* root should already exist */ | |
7707 | ||
7708 | tg->parent = parent; | |
f473aa5e | 7709 | INIT_LIST_HEAD(&tg->children); |
09f2724a | 7710 | list_add_rcu(&tg->siblings, &parent->children); |
8ed36996 | 7711 | spin_unlock_irqrestore(&task_group_lock, flags); |
29f59db3 SV |
7712 | } |
7713 | ||
9b5b7751 | 7714 | /* rcu callback to free various structures associated with a task group */ |
2f5177f0 | 7715 | static void sched_free_group_rcu(struct rcu_head *rhp) |
29f59db3 | 7716 | { |
29f59db3 | 7717 | /* now it should be safe to free those cfs_rqs */ |
2f5177f0 | 7718 | sched_free_group(container_of(rhp, struct task_group, rcu)); |
29f59db3 SV |
7719 | } |
7720 | ||
4cf86d77 | 7721 | void sched_destroy_group(struct task_group *tg) |
ace783b9 LZ |
7722 | { |
7723 | /* wait for possible concurrent references to cfs_rqs complete */ | |
2f5177f0 | 7724 | call_rcu(&tg->rcu, sched_free_group_rcu); |
ace783b9 LZ |
7725 | } |
7726 | ||
7727 | void sched_offline_group(struct task_group *tg) | |
29f59db3 | 7728 | { |
8ed36996 | 7729 | unsigned long flags; |
29f59db3 | 7730 | |
3d4b47b4 | 7731 | /* end participation in shares distribution */ |
6fe1f348 | 7732 | unregister_fair_sched_group(tg); |
3d4b47b4 PZ |
7733 | |
7734 | spin_lock_irqsave(&task_group_lock, flags); | |
6f505b16 | 7735 | list_del_rcu(&tg->list); |
f473aa5e | 7736 | list_del_rcu(&tg->siblings); |
8ed36996 | 7737 | spin_unlock_irqrestore(&task_group_lock, flags); |
29f59db3 SV |
7738 | } |
7739 | ||
9b5b7751 | 7740 | /* change task's runqueue when it moves between groups. |
3a252015 IM |
7741 | * The caller of this function should have put the task in its new group |
7742 | * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to | |
7743 | * reflect its new group. | |
9b5b7751 SV |
7744 | */ |
7745 | void sched_move_task(struct task_struct *tsk) | |
29f59db3 | 7746 | { |
8323f26c | 7747 | struct task_group *tg; |
da0c1e65 | 7748 | int queued, running; |
eb580751 | 7749 | struct rq_flags rf; |
29f59db3 SV |
7750 | struct rq *rq; |
7751 | ||
eb580751 | 7752 | rq = task_rq_lock(tsk, &rf); |
29f59db3 | 7753 | |
051a1d1a | 7754 | running = task_current(rq, tsk); |
da0c1e65 | 7755 | queued = task_on_rq_queued(tsk); |
29f59db3 | 7756 | |
da0c1e65 | 7757 | if (queued) |
ff77e468 | 7758 | dequeue_task(rq, tsk, DEQUEUE_SAVE | DEQUEUE_MOVE); |
0e1f3483 | 7759 | if (unlikely(running)) |
f3cd1c4e | 7760 | put_prev_task(rq, tsk); |
29f59db3 | 7761 | |
f7b8a47d KT |
7762 | /* |
7763 | * All callers are synchronized by task_rq_lock(); we do not use RCU | |
7764 | * which is pointless here. Thus, we pass "true" to task_css_check() | |
7765 | * to prevent lockdep warnings. | |
7766 | */ | |
7767 | tg = container_of(task_css_check(tsk, cpu_cgrp_id, true), | |
8323f26c PZ |
7768 | struct task_group, css); |
7769 | tg = autogroup_task_group(tsk, tg); | |
7770 | tsk->sched_task_group = tg; | |
7771 | ||
810b3817 | 7772 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 7773 | if (tsk->sched_class->task_move_group) |
bc54da21 | 7774 | tsk->sched_class->task_move_group(tsk); |
b2b5ce02 | 7775 | else |
810b3817 | 7776 | #endif |
b2b5ce02 | 7777 | set_task_rq(tsk, task_cpu(tsk)); |
810b3817 | 7778 | |
0e1f3483 HS |
7779 | if (unlikely(running)) |
7780 | tsk->sched_class->set_curr_task(rq); | |
da0c1e65 | 7781 | if (queued) |
ff77e468 | 7782 | enqueue_task(rq, tsk, ENQUEUE_RESTORE | ENQUEUE_MOVE); |
29f59db3 | 7783 | |
eb580751 | 7784 | task_rq_unlock(rq, tsk, &rf); |
29f59db3 | 7785 | } |
7c941438 | 7786 | #endif /* CONFIG_CGROUP_SCHED */ |
29f59db3 | 7787 | |
a790de99 PT |
7788 | #ifdef CONFIG_RT_GROUP_SCHED |
7789 | /* | |
7790 | * Ensure that the real time constraints are schedulable. | |
7791 | */ | |
7792 | static DEFINE_MUTEX(rt_constraints_mutex); | |
9f0c1e56 | 7793 | |
9a7e0b18 PZ |
7794 | /* Must be called with tasklist_lock held */ |
7795 | static inline int tg_has_rt_tasks(struct task_group *tg) | |
b40b2e8e | 7796 | { |
9a7e0b18 | 7797 | struct task_struct *g, *p; |
b40b2e8e | 7798 | |
1fe89e1b PZ |
7799 | /* |
7800 | * Autogroups do not have RT tasks; see autogroup_create(). | |
7801 | */ | |
7802 | if (task_group_is_autogroup(tg)) | |
7803 | return 0; | |
7804 | ||
5d07f420 | 7805 | for_each_process_thread(g, p) { |
8651c658 | 7806 | if (rt_task(p) && task_group(p) == tg) |
9a7e0b18 | 7807 | return 1; |
5d07f420 | 7808 | } |
b40b2e8e | 7809 | |
9a7e0b18 PZ |
7810 | return 0; |
7811 | } | |
b40b2e8e | 7812 | |
9a7e0b18 PZ |
7813 | struct rt_schedulable_data { |
7814 | struct task_group *tg; | |
7815 | u64 rt_period; | |
7816 | u64 rt_runtime; | |
7817 | }; | |
b40b2e8e | 7818 | |
a790de99 | 7819 | static int tg_rt_schedulable(struct task_group *tg, void *data) |
9a7e0b18 PZ |
7820 | { |
7821 | struct rt_schedulable_data *d = data; | |
7822 | struct task_group *child; | |
7823 | unsigned long total, sum = 0; | |
7824 | u64 period, runtime; | |
b40b2e8e | 7825 | |
9a7e0b18 PZ |
7826 | period = ktime_to_ns(tg->rt_bandwidth.rt_period); |
7827 | runtime = tg->rt_bandwidth.rt_runtime; | |
b40b2e8e | 7828 | |
9a7e0b18 PZ |
7829 | if (tg == d->tg) { |
7830 | period = d->rt_period; | |
7831 | runtime = d->rt_runtime; | |
b40b2e8e | 7832 | } |
b40b2e8e | 7833 | |
4653f803 PZ |
7834 | /* |
7835 | * Cannot have more runtime than the period. | |
7836 | */ | |
7837 | if (runtime > period && runtime != RUNTIME_INF) | |
7838 | return -EINVAL; | |
6f505b16 | 7839 | |
4653f803 PZ |
7840 | /* |
7841 | * Ensure we don't starve existing RT tasks. | |
7842 | */ | |
9a7e0b18 PZ |
7843 | if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) |
7844 | return -EBUSY; | |
6f505b16 | 7845 | |
9a7e0b18 | 7846 | total = to_ratio(period, runtime); |
6f505b16 | 7847 | |
4653f803 PZ |
7848 | /* |
7849 | * Nobody can have more than the global setting allows. | |
7850 | */ | |
7851 | if (total > to_ratio(global_rt_period(), global_rt_runtime())) | |
7852 | return -EINVAL; | |
6f505b16 | 7853 | |
4653f803 PZ |
7854 | /* |
7855 | * The sum of our children's runtime should not exceed our own. | |
7856 | */ | |
9a7e0b18 PZ |
7857 | list_for_each_entry_rcu(child, &tg->children, siblings) { |
7858 | period = ktime_to_ns(child->rt_bandwidth.rt_period); | |
7859 | runtime = child->rt_bandwidth.rt_runtime; | |
6f505b16 | 7860 | |
9a7e0b18 PZ |
7861 | if (child == d->tg) { |
7862 | period = d->rt_period; | |
7863 | runtime = d->rt_runtime; | |
7864 | } | |
6f505b16 | 7865 | |
9a7e0b18 | 7866 | sum += to_ratio(period, runtime); |
9f0c1e56 | 7867 | } |
6f505b16 | 7868 | |
9a7e0b18 PZ |
7869 | if (sum > total) |
7870 | return -EINVAL; | |
7871 | ||
7872 | return 0; | |
6f505b16 PZ |
7873 | } |
7874 | ||
9a7e0b18 | 7875 | static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) |
521f1a24 | 7876 | { |
8277434e PT |
7877 | int ret; |
7878 | ||
9a7e0b18 PZ |
7879 | struct rt_schedulable_data data = { |
7880 | .tg = tg, | |
7881 | .rt_period = period, | |
7882 | .rt_runtime = runtime, | |
7883 | }; | |
7884 | ||
8277434e PT |
7885 | rcu_read_lock(); |
7886 | ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); | |
7887 | rcu_read_unlock(); | |
7888 | ||
7889 | return ret; | |
521f1a24 DG |
7890 | } |
7891 | ||
ab84d31e | 7892 | static int tg_set_rt_bandwidth(struct task_group *tg, |
d0b27fa7 | 7893 | u64 rt_period, u64 rt_runtime) |
6f505b16 | 7894 | { |
ac086bc2 | 7895 | int i, err = 0; |
9f0c1e56 | 7896 | |
2636ed5f PZ |
7897 | /* |
7898 | * Disallowing the root group RT runtime is BAD, it would disallow the | |
7899 | * kernel creating (and or operating) RT threads. | |
7900 | */ | |
7901 | if (tg == &root_task_group && rt_runtime == 0) | |
7902 | return -EINVAL; | |
7903 | ||
7904 | /* No period doesn't make any sense. */ | |
7905 | if (rt_period == 0) | |
7906 | return -EINVAL; | |
7907 | ||
9f0c1e56 | 7908 | mutex_lock(&rt_constraints_mutex); |
521f1a24 | 7909 | read_lock(&tasklist_lock); |
9a7e0b18 PZ |
7910 | err = __rt_schedulable(tg, rt_period, rt_runtime); |
7911 | if (err) | |
9f0c1e56 | 7912 | goto unlock; |
ac086bc2 | 7913 | |
0986b11b | 7914 | raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); |
d0b27fa7 PZ |
7915 | tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); |
7916 | tg->rt_bandwidth.rt_runtime = rt_runtime; | |
ac086bc2 PZ |
7917 | |
7918 | for_each_possible_cpu(i) { | |
7919 | struct rt_rq *rt_rq = tg->rt_rq[i]; | |
7920 | ||
0986b11b | 7921 | raw_spin_lock(&rt_rq->rt_runtime_lock); |
ac086bc2 | 7922 | rt_rq->rt_runtime = rt_runtime; |
0986b11b | 7923 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
ac086bc2 | 7924 | } |
0986b11b | 7925 | raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); |
49246274 | 7926 | unlock: |
521f1a24 | 7927 | read_unlock(&tasklist_lock); |
9f0c1e56 PZ |
7928 | mutex_unlock(&rt_constraints_mutex); |
7929 | ||
7930 | return err; | |
6f505b16 PZ |
7931 | } |
7932 | ||
25cc7da7 | 7933 | static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) |
d0b27fa7 PZ |
7934 | { |
7935 | u64 rt_runtime, rt_period; | |
7936 | ||
7937 | rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); | |
7938 | rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; | |
7939 | if (rt_runtime_us < 0) | |
7940 | rt_runtime = RUNTIME_INF; | |
7941 | ||
ab84d31e | 7942 | return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); |
d0b27fa7 PZ |
7943 | } |
7944 | ||
25cc7da7 | 7945 | static long sched_group_rt_runtime(struct task_group *tg) |
9f0c1e56 PZ |
7946 | { |
7947 | u64 rt_runtime_us; | |
7948 | ||
d0b27fa7 | 7949 | if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) |
9f0c1e56 PZ |
7950 | return -1; |
7951 | ||
d0b27fa7 | 7952 | rt_runtime_us = tg->rt_bandwidth.rt_runtime; |
9f0c1e56 PZ |
7953 | do_div(rt_runtime_us, NSEC_PER_USEC); |
7954 | return rt_runtime_us; | |
7955 | } | |
d0b27fa7 | 7956 | |
ce2f5fe4 | 7957 | static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us) |
d0b27fa7 PZ |
7958 | { |
7959 | u64 rt_runtime, rt_period; | |
7960 | ||
ce2f5fe4 | 7961 | rt_period = rt_period_us * NSEC_PER_USEC; |
d0b27fa7 PZ |
7962 | rt_runtime = tg->rt_bandwidth.rt_runtime; |
7963 | ||
ab84d31e | 7964 | return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); |
d0b27fa7 PZ |
7965 | } |
7966 | ||
25cc7da7 | 7967 | static long sched_group_rt_period(struct task_group *tg) |
d0b27fa7 PZ |
7968 | { |
7969 | u64 rt_period_us; | |
7970 | ||
7971 | rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); | |
7972 | do_div(rt_period_us, NSEC_PER_USEC); | |
7973 | return rt_period_us; | |
7974 | } | |
332ac17e | 7975 | #endif /* CONFIG_RT_GROUP_SCHED */ |
d0b27fa7 | 7976 | |
332ac17e | 7977 | #ifdef CONFIG_RT_GROUP_SCHED |
d0b27fa7 PZ |
7978 | static int sched_rt_global_constraints(void) |
7979 | { | |
7980 | int ret = 0; | |
7981 | ||
7982 | mutex_lock(&rt_constraints_mutex); | |
9a7e0b18 | 7983 | read_lock(&tasklist_lock); |
4653f803 | 7984 | ret = __rt_schedulable(NULL, 0, 0); |
9a7e0b18 | 7985 | read_unlock(&tasklist_lock); |
d0b27fa7 PZ |
7986 | mutex_unlock(&rt_constraints_mutex); |
7987 | ||
7988 | return ret; | |
7989 | } | |
54e99124 | 7990 | |
25cc7da7 | 7991 | static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) |
54e99124 DG |
7992 | { |
7993 | /* Don't accept realtime tasks when there is no way for them to run */ | |
7994 | if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) | |
7995 | return 0; | |
7996 | ||
7997 | return 1; | |
7998 | } | |
7999 | ||
6d6bc0ad | 8000 | #else /* !CONFIG_RT_GROUP_SCHED */ |
d0b27fa7 PZ |
8001 | static int sched_rt_global_constraints(void) |
8002 | { | |
ac086bc2 | 8003 | unsigned long flags; |
8c5e9554 | 8004 | int i; |
ec5d4989 | 8005 | |
0986b11b | 8006 | raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); |
ac086bc2 PZ |
8007 | for_each_possible_cpu(i) { |
8008 | struct rt_rq *rt_rq = &cpu_rq(i)->rt; | |
8009 | ||
0986b11b | 8010 | raw_spin_lock(&rt_rq->rt_runtime_lock); |
ac086bc2 | 8011 | rt_rq->rt_runtime = global_rt_runtime(); |
0986b11b | 8012 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
ac086bc2 | 8013 | } |
0986b11b | 8014 | raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); |
ac086bc2 | 8015 | |
8c5e9554 | 8016 | return 0; |
d0b27fa7 | 8017 | } |
6d6bc0ad | 8018 | #endif /* CONFIG_RT_GROUP_SCHED */ |
d0b27fa7 | 8019 | |
a1963b81 | 8020 | static int sched_dl_global_validate(void) |
332ac17e | 8021 | { |
1724813d PZ |
8022 | u64 runtime = global_rt_runtime(); |
8023 | u64 period = global_rt_period(); | |
332ac17e | 8024 | u64 new_bw = to_ratio(period, runtime); |
f10e00f4 | 8025 | struct dl_bw *dl_b; |
1724813d | 8026 | int cpu, ret = 0; |
49516342 | 8027 | unsigned long flags; |
332ac17e DF |
8028 | |
8029 | /* | |
8030 | * Here we want to check the bandwidth not being set to some | |
8031 | * value smaller than the currently allocated bandwidth in | |
8032 | * any of the root_domains. | |
8033 | * | |
8034 | * FIXME: Cycling on all the CPUs is overdoing, but simpler than | |
8035 | * cycling on root_domains... Discussion on different/better | |
8036 | * solutions is welcome! | |
8037 | */ | |
1724813d | 8038 | for_each_possible_cpu(cpu) { |
f10e00f4 KT |
8039 | rcu_read_lock_sched(); |
8040 | dl_b = dl_bw_of(cpu); | |
332ac17e | 8041 | |
49516342 | 8042 | raw_spin_lock_irqsave(&dl_b->lock, flags); |
1724813d PZ |
8043 | if (new_bw < dl_b->total_bw) |
8044 | ret = -EBUSY; | |
49516342 | 8045 | raw_spin_unlock_irqrestore(&dl_b->lock, flags); |
1724813d | 8046 | |
f10e00f4 KT |
8047 | rcu_read_unlock_sched(); |
8048 | ||
1724813d PZ |
8049 | if (ret) |
8050 | break; | |
332ac17e DF |
8051 | } |
8052 | ||
1724813d | 8053 | return ret; |
332ac17e DF |
8054 | } |
8055 | ||
1724813d | 8056 | static void sched_dl_do_global(void) |
ce0dbbbb | 8057 | { |
1724813d | 8058 | u64 new_bw = -1; |
f10e00f4 | 8059 | struct dl_bw *dl_b; |
1724813d | 8060 | int cpu; |
49516342 | 8061 | unsigned long flags; |
ce0dbbbb | 8062 | |
1724813d PZ |
8063 | def_dl_bandwidth.dl_period = global_rt_period(); |
8064 | def_dl_bandwidth.dl_runtime = global_rt_runtime(); | |
8065 | ||
8066 | if (global_rt_runtime() != RUNTIME_INF) | |
8067 | new_bw = to_ratio(global_rt_period(), global_rt_runtime()); | |
8068 | ||
8069 | /* | |
8070 | * FIXME: As above... | |
8071 | */ | |
8072 | for_each_possible_cpu(cpu) { | |
f10e00f4 KT |
8073 | rcu_read_lock_sched(); |
8074 | dl_b = dl_bw_of(cpu); | |
1724813d | 8075 | |
49516342 | 8076 | raw_spin_lock_irqsave(&dl_b->lock, flags); |
1724813d | 8077 | dl_b->bw = new_bw; |
49516342 | 8078 | raw_spin_unlock_irqrestore(&dl_b->lock, flags); |
f10e00f4 KT |
8079 | |
8080 | rcu_read_unlock_sched(); | |
ce0dbbbb | 8081 | } |
1724813d PZ |
8082 | } |
8083 | ||
8084 | static int sched_rt_global_validate(void) | |
8085 | { | |
8086 | if (sysctl_sched_rt_period <= 0) | |
8087 | return -EINVAL; | |
8088 | ||
e9e7cb38 JL |
8089 | if ((sysctl_sched_rt_runtime != RUNTIME_INF) && |
8090 | (sysctl_sched_rt_runtime > sysctl_sched_rt_period)) | |
1724813d PZ |
8091 | return -EINVAL; |
8092 | ||
8093 | return 0; | |
8094 | } | |
8095 | ||
8096 | static void sched_rt_do_global(void) | |
8097 | { | |
8098 | def_rt_bandwidth.rt_runtime = global_rt_runtime(); | |
8099 | def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); | |
ce0dbbbb CW |
8100 | } |
8101 | ||
d0b27fa7 | 8102 | int sched_rt_handler(struct ctl_table *table, int write, |
8d65af78 | 8103 | void __user *buffer, size_t *lenp, |
d0b27fa7 PZ |
8104 | loff_t *ppos) |
8105 | { | |
d0b27fa7 PZ |
8106 | int old_period, old_runtime; |
8107 | static DEFINE_MUTEX(mutex); | |
1724813d | 8108 | int ret; |
d0b27fa7 PZ |
8109 | |
8110 | mutex_lock(&mutex); | |
8111 | old_period = sysctl_sched_rt_period; | |
8112 | old_runtime = sysctl_sched_rt_runtime; | |
8113 | ||
8d65af78 | 8114 | ret = proc_dointvec(table, write, buffer, lenp, ppos); |
d0b27fa7 PZ |
8115 | |
8116 | if (!ret && write) { | |
1724813d PZ |
8117 | ret = sched_rt_global_validate(); |
8118 | if (ret) | |
8119 | goto undo; | |
8120 | ||
a1963b81 | 8121 | ret = sched_dl_global_validate(); |
1724813d PZ |
8122 | if (ret) |
8123 | goto undo; | |
8124 | ||
a1963b81 | 8125 | ret = sched_rt_global_constraints(); |
1724813d PZ |
8126 | if (ret) |
8127 | goto undo; | |
8128 | ||
8129 | sched_rt_do_global(); | |
8130 | sched_dl_do_global(); | |
8131 | } | |
8132 | if (0) { | |
8133 | undo: | |
8134 | sysctl_sched_rt_period = old_period; | |
8135 | sysctl_sched_rt_runtime = old_runtime; | |
d0b27fa7 PZ |
8136 | } |
8137 | mutex_unlock(&mutex); | |
8138 | ||
8139 | return ret; | |
8140 | } | |
68318b8e | 8141 | |
1724813d | 8142 | int sched_rr_handler(struct ctl_table *table, int write, |
332ac17e DF |
8143 | void __user *buffer, size_t *lenp, |
8144 | loff_t *ppos) | |
8145 | { | |
8146 | int ret; | |
332ac17e | 8147 | static DEFINE_MUTEX(mutex); |
332ac17e DF |
8148 | |
8149 | mutex_lock(&mutex); | |
332ac17e | 8150 | ret = proc_dointvec(table, write, buffer, lenp, ppos); |
1724813d PZ |
8151 | /* make sure that internally we keep jiffies */ |
8152 | /* also, writing zero resets timeslice to default */ | |
332ac17e | 8153 | if (!ret && write) { |
1724813d PZ |
8154 | sched_rr_timeslice = sched_rr_timeslice <= 0 ? |
8155 | RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); | |
332ac17e DF |
8156 | } |
8157 | mutex_unlock(&mutex); | |
332ac17e DF |
8158 | return ret; |
8159 | } | |
8160 | ||
052f1dc7 | 8161 | #ifdef CONFIG_CGROUP_SCHED |
68318b8e | 8162 | |
a7c6d554 | 8163 | static inline struct task_group *css_tg(struct cgroup_subsys_state *css) |
68318b8e | 8164 | { |
a7c6d554 | 8165 | return css ? container_of(css, struct task_group, css) : NULL; |
68318b8e SV |
8166 | } |
8167 | ||
eb95419b TH |
8168 | static struct cgroup_subsys_state * |
8169 | cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) | |
68318b8e | 8170 | { |
eb95419b TH |
8171 | struct task_group *parent = css_tg(parent_css); |
8172 | struct task_group *tg; | |
68318b8e | 8173 | |
eb95419b | 8174 | if (!parent) { |
68318b8e | 8175 | /* This is early initialization for the top cgroup */ |
07e06b01 | 8176 | return &root_task_group.css; |
68318b8e SV |
8177 | } |
8178 | ||
ec7dc8ac | 8179 | tg = sched_create_group(parent); |
68318b8e SV |
8180 | if (IS_ERR(tg)) |
8181 | return ERR_PTR(-ENOMEM); | |
8182 | ||
2f5177f0 PZ |
8183 | sched_online_group(tg, parent); |
8184 | ||
68318b8e SV |
8185 | return &tg->css; |
8186 | } | |
8187 | ||
2f5177f0 | 8188 | static void cpu_cgroup_css_released(struct cgroup_subsys_state *css) |
ace783b9 | 8189 | { |
eb95419b | 8190 | struct task_group *tg = css_tg(css); |
ace783b9 | 8191 | |
2f5177f0 | 8192 | sched_offline_group(tg); |
ace783b9 LZ |
8193 | } |
8194 | ||
eb95419b | 8195 | static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) |
68318b8e | 8196 | { |
eb95419b | 8197 | struct task_group *tg = css_tg(css); |
68318b8e | 8198 | |
2f5177f0 PZ |
8199 | /* |
8200 | * Relies on the RCU grace period between css_released() and this. | |
8201 | */ | |
8202 | sched_free_group(tg); | |
ace783b9 LZ |
8203 | } |
8204 | ||
b53202e6 | 8205 | static void cpu_cgroup_fork(struct task_struct *task) |
eeb61e53 KT |
8206 | { |
8207 | sched_move_task(task); | |
8208 | } | |
8209 | ||
1f7dd3e5 | 8210 | static int cpu_cgroup_can_attach(struct cgroup_taskset *tset) |
68318b8e | 8211 | { |
bb9d97b6 | 8212 | struct task_struct *task; |
1f7dd3e5 | 8213 | struct cgroup_subsys_state *css; |
bb9d97b6 | 8214 | |
1f7dd3e5 | 8215 | cgroup_taskset_for_each(task, css, tset) { |
b68aa230 | 8216 | #ifdef CONFIG_RT_GROUP_SCHED |
eb95419b | 8217 | if (!sched_rt_can_attach(css_tg(css), task)) |
bb9d97b6 | 8218 | return -EINVAL; |
b68aa230 | 8219 | #else |
bb9d97b6 TH |
8220 | /* We don't support RT-tasks being in separate groups */ |
8221 | if (task->sched_class != &fair_sched_class) | |
8222 | return -EINVAL; | |
b68aa230 | 8223 | #endif |
bb9d97b6 | 8224 | } |
be367d09 BB |
8225 | return 0; |
8226 | } | |
68318b8e | 8227 | |
1f7dd3e5 | 8228 | static void cpu_cgroup_attach(struct cgroup_taskset *tset) |
68318b8e | 8229 | { |
bb9d97b6 | 8230 | struct task_struct *task; |
1f7dd3e5 | 8231 | struct cgroup_subsys_state *css; |
bb9d97b6 | 8232 | |
1f7dd3e5 | 8233 | cgroup_taskset_for_each(task, css, tset) |
bb9d97b6 | 8234 | sched_move_task(task); |
68318b8e SV |
8235 | } |
8236 | ||
052f1dc7 | 8237 | #ifdef CONFIG_FAIR_GROUP_SCHED |
182446d0 TH |
8238 | static int cpu_shares_write_u64(struct cgroup_subsys_state *css, |
8239 | struct cftype *cftype, u64 shareval) | |
68318b8e | 8240 | { |
182446d0 | 8241 | return sched_group_set_shares(css_tg(css), scale_load(shareval)); |
68318b8e SV |
8242 | } |
8243 | ||
182446d0 TH |
8244 | static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, |
8245 | struct cftype *cft) | |
68318b8e | 8246 | { |
182446d0 | 8247 | struct task_group *tg = css_tg(css); |
68318b8e | 8248 | |
c8b28116 | 8249 | return (u64) scale_load_down(tg->shares); |
68318b8e | 8250 | } |
ab84d31e PT |
8251 | |
8252 | #ifdef CONFIG_CFS_BANDWIDTH | |
a790de99 PT |
8253 | static DEFINE_MUTEX(cfs_constraints_mutex); |
8254 | ||
ab84d31e PT |
8255 | const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ |
8256 | const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ | |
8257 | ||
a790de99 PT |
8258 | static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); |
8259 | ||
ab84d31e PT |
8260 | static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) |
8261 | { | |
56f570e5 | 8262 | int i, ret = 0, runtime_enabled, runtime_was_enabled; |
029632fb | 8263 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; |
ab84d31e PT |
8264 | |
8265 | if (tg == &root_task_group) | |
8266 | return -EINVAL; | |
8267 | ||
8268 | /* | |
8269 | * Ensure we have at some amount of bandwidth every period. This is | |
8270 | * to prevent reaching a state of large arrears when throttled via | |
8271 | * entity_tick() resulting in prolonged exit starvation. | |
8272 | */ | |
8273 | if (quota < min_cfs_quota_period || period < min_cfs_quota_period) | |
8274 | return -EINVAL; | |
8275 | ||
8276 | /* | |
8277 | * Likewise, bound things on the otherside by preventing insane quota | |
8278 | * periods. This also allows us to normalize in computing quota | |
8279 | * feasibility. | |
8280 | */ | |
8281 | if (period > max_cfs_quota_period) | |
8282 | return -EINVAL; | |
8283 | ||
0e59bdae KT |
8284 | /* |
8285 | * Prevent race between setting of cfs_rq->runtime_enabled and | |
8286 | * unthrottle_offline_cfs_rqs(). | |
8287 | */ | |
8288 | get_online_cpus(); | |
a790de99 PT |
8289 | mutex_lock(&cfs_constraints_mutex); |
8290 | ret = __cfs_schedulable(tg, period, quota); | |
8291 | if (ret) | |
8292 | goto out_unlock; | |
8293 | ||
58088ad0 | 8294 | runtime_enabled = quota != RUNTIME_INF; |
56f570e5 | 8295 | runtime_was_enabled = cfs_b->quota != RUNTIME_INF; |
1ee14e6c BS |
8296 | /* |
8297 | * If we need to toggle cfs_bandwidth_used, off->on must occur | |
8298 | * before making related changes, and on->off must occur afterwards | |
8299 | */ | |
8300 | if (runtime_enabled && !runtime_was_enabled) | |
8301 | cfs_bandwidth_usage_inc(); | |
ab84d31e PT |
8302 | raw_spin_lock_irq(&cfs_b->lock); |
8303 | cfs_b->period = ns_to_ktime(period); | |
8304 | cfs_b->quota = quota; | |
58088ad0 | 8305 | |
a9cf55b2 | 8306 | __refill_cfs_bandwidth_runtime(cfs_b); |
58088ad0 | 8307 | /* restart the period timer (if active) to handle new period expiry */ |
77a4d1a1 PZ |
8308 | if (runtime_enabled) |
8309 | start_cfs_bandwidth(cfs_b); | |
ab84d31e PT |
8310 | raw_spin_unlock_irq(&cfs_b->lock); |
8311 | ||
0e59bdae | 8312 | for_each_online_cpu(i) { |
ab84d31e | 8313 | struct cfs_rq *cfs_rq = tg->cfs_rq[i]; |
029632fb | 8314 | struct rq *rq = cfs_rq->rq; |
ab84d31e PT |
8315 | |
8316 | raw_spin_lock_irq(&rq->lock); | |
58088ad0 | 8317 | cfs_rq->runtime_enabled = runtime_enabled; |
ab84d31e | 8318 | cfs_rq->runtime_remaining = 0; |
671fd9da | 8319 | |
029632fb | 8320 | if (cfs_rq->throttled) |
671fd9da | 8321 | unthrottle_cfs_rq(cfs_rq); |
ab84d31e PT |
8322 | raw_spin_unlock_irq(&rq->lock); |
8323 | } | |
1ee14e6c BS |
8324 | if (runtime_was_enabled && !runtime_enabled) |
8325 | cfs_bandwidth_usage_dec(); | |
a790de99 PT |
8326 | out_unlock: |
8327 | mutex_unlock(&cfs_constraints_mutex); | |
0e59bdae | 8328 | put_online_cpus(); |
ab84d31e | 8329 | |
a790de99 | 8330 | return ret; |
ab84d31e PT |
8331 | } |
8332 | ||
8333 | int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) | |
8334 | { | |
8335 | u64 quota, period; | |
8336 | ||
029632fb | 8337 | period = ktime_to_ns(tg->cfs_bandwidth.period); |
ab84d31e PT |
8338 | if (cfs_quota_us < 0) |
8339 | quota = RUNTIME_INF; | |
8340 | else | |
8341 | quota = (u64)cfs_quota_us * NSEC_PER_USEC; | |
8342 | ||
8343 | return tg_set_cfs_bandwidth(tg, period, quota); | |
8344 | } | |
8345 | ||
8346 | long tg_get_cfs_quota(struct task_group *tg) | |
8347 | { | |
8348 | u64 quota_us; | |
8349 | ||
029632fb | 8350 | if (tg->cfs_bandwidth.quota == RUNTIME_INF) |
ab84d31e PT |
8351 | return -1; |
8352 | ||
029632fb | 8353 | quota_us = tg->cfs_bandwidth.quota; |
ab84d31e PT |
8354 | do_div(quota_us, NSEC_PER_USEC); |
8355 | ||
8356 | return quota_us; | |
8357 | } | |
8358 | ||
8359 | int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) | |
8360 | { | |
8361 | u64 quota, period; | |
8362 | ||
8363 | period = (u64)cfs_period_us * NSEC_PER_USEC; | |
029632fb | 8364 | quota = tg->cfs_bandwidth.quota; |
ab84d31e | 8365 | |
ab84d31e PT |
8366 | return tg_set_cfs_bandwidth(tg, period, quota); |
8367 | } | |
8368 | ||
8369 | long tg_get_cfs_period(struct task_group *tg) | |
8370 | { | |
8371 | u64 cfs_period_us; | |
8372 | ||
029632fb | 8373 | cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); |
ab84d31e PT |
8374 | do_div(cfs_period_us, NSEC_PER_USEC); |
8375 | ||
8376 | return cfs_period_us; | |
8377 | } | |
8378 | ||
182446d0 TH |
8379 | static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, |
8380 | struct cftype *cft) | |
ab84d31e | 8381 | { |
182446d0 | 8382 | return tg_get_cfs_quota(css_tg(css)); |
ab84d31e PT |
8383 | } |
8384 | ||
182446d0 TH |
8385 | static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, |
8386 | struct cftype *cftype, s64 cfs_quota_us) | |
ab84d31e | 8387 | { |
182446d0 | 8388 | return tg_set_cfs_quota(css_tg(css), cfs_quota_us); |
ab84d31e PT |
8389 | } |
8390 | ||
182446d0 TH |
8391 | static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, |
8392 | struct cftype *cft) | |
ab84d31e | 8393 | { |
182446d0 | 8394 | return tg_get_cfs_period(css_tg(css)); |
ab84d31e PT |
8395 | } |
8396 | ||
182446d0 TH |
8397 | static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, |
8398 | struct cftype *cftype, u64 cfs_period_us) | |
ab84d31e | 8399 | { |
182446d0 | 8400 | return tg_set_cfs_period(css_tg(css), cfs_period_us); |
ab84d31e PT |
8401 | } |
8402 | ||
a790de99 PT |
8403 | struct cfs_schedulable_data { |
8404 | struct task_group *tg; | |
8405 | u64 period, quota; | |
8406 | }; | |
8407 | ||
8408 | /* | |
8409 | * normalize group quota/period to be quota/max_period | |
8410 | * note: units are usecs | |
8411 | */ | |
8412 | static u64 normalize_cfs_quota(struct task_group *tg, | |
8413 | struct cfs_schedulable_data *d) | |
8414 | { | |
8415 | u64 quota, period; | |
8416 | ||
8417 | if (tg == d->tg) { | |
8418 | period = d->period; | |
8419 | quota = d->quota; | |
8420 | } else { | |
8421 | period = tg_get_cfs_period(tg); | |
8422 | quota = tg_get_cfs_quota(tg); | |
8423 | } | |
8424 | ||
8425 | /* note: these should typically be equivalent */ | |
8426 | if (quota == RUNTIME_INF || quota == -1) | |
8427 | return RUNTIME_INF; | |
8428 | ||
8429 | return to_ratio(period, quota); | |
8430 | } | |
8431 | ||
8432 | static int tg_cfs_schedulable_down(struct task_group *tg, void *data) | |
8433 | { | |
8434 | struct cfs_schedulable_data *d = data; | |
029632fb | 8435 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; |
a790de99 PT |
8436 | s64 quota = 0, parent_quota = -1; |
8437 | ||
8438 | if (!tg->parent) { | |
8439 | quota = RUNTIME_INF; | |
8440 | } else { | |
029632fb | 8441 | struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; |
a790de99 PT |
8442 | |
8443 | quota = normalize_cfs_quota(tg, d); | |
9c58c79a | 8444 | parent_quota = parent_b->hierarchical_quota; |
a790de99 PT |
8445 | |
8446 | /* | |
8447 | * ensure max(child_quota) <= parent_quota, inherit when no | |
8448 | * limit is set | |
8449 | */ | |
8450 | if (quota == RUNTIME_INF) | |
8451 | quota = parent_quota; | |
8452 | else if (parent_quota != RUNTIME_INF && quota > parent_quota) | |
8453 | return -EINVAL; | |
8454 | } | |
9c58c79a | 8455 | cfs_b->hierarchical_quota = quota; |
a790de99 PT |
8456 | |
8457 | return 0; | |
8458 | } | |
8459 | ||
8460 | static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) | |
8461 | { | |
8277434e | 8462 | int ret; |
a790de99 PT |
8463 | struct cfs_schedulable_data data = { |
8464 | .tg = tg, | |
8465 | .period = period, | |
8466 | .quota = quota, | |
8467 | }; | |
8468 | ||
8469 | if (quota != RUNTIME_INF) { | |
8470 | do_div(data.period, NSEC_PER_USEC); | |
8471 | do_div(data.quota, NSEC_PER_USEC); | |
8472 | } | |
8473 | ||
8277434e PT |
8474 | rcu_read_lock(); |
8475 | ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); | |
8476 | rcu_read_unlock(); | |
8477 | ||
8478 | return ret; | |
a790de99 | 8479 | } |
e8da1b18 | 8480 | |
2da8ca82 | 8481 | static int cpu_stats_show(struct seq_file *sf, void *v) |
e8da1b18 | 8482 | { |
2da8ca82 | 8483 | struct task_group *tg = css_tg(seq_css(sf)); |
029632fb | 8484 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; |
e8da1b18 | 8485 | |
44ffc75b TH |
8486 | seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); |
8487 | seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); | |
8488 | seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); | |
e8da1b18 NR |
8489 | |
8490 | return 0; | |
8491 | } | |
ab84d31e | 8492 | #endif /* CONFIG_CFS_BANDWIDTH */ |
6d6bc0ad | 8493 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
68318b8e | 8494 | |
052f1dc7 | 8495 | #ifdef CONFIG_RT_GROUP_SCHED |
182446d0 TH |
8496 | static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, |
8497 | struct cftype *cft, s64 val) | |
6f505b16 | 8498 | { |
182446d0 | 8499 | return sched_group_set_rt_runtime(css_tg(css), val); |
6f505b16 PZ |
8500 | } |
8501 | ||
182446d0 TH |
8502 | static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, |
8503 | struct cftype *cft) | |
6f505b16 | 8504 | { |
182446d0 | 8505 | return sched_group_rt_runtime(css_tg(css)); |
6f505b16 | 8506 | } |
d0b27fa7 | 8507 | |
182446d0 TH |
8508 | static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, |
8509 | struct cftype *cftype, u64 rt_period_us) | |
d0b27fa7 | 8510 | { |
182446d0 | 8511 | return sched_group_set_rt_period(css_tg(css), rt_period_us); |
d0b27fa7 PZ |
8512 | } |
8513 | ||
182446d0 TH |
8514 | static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, |
8515 | struct cftype *cft) | |
d0b27fa7 | 8516 | { |
182446d0 | 8517 | return sched_group_rt_period(css_tg(css)); |
d0b27fa7 | 8518 | } |
6d6bc0ad | 8519 | #endif /* CONFIG_RT_GROUP_SCHED */ |
6f505b16 | 8520 | |
fe5c7cc2 | 8521 | static struct cftype cpu_files[] = { |
052f1dc7 | 8522 | #ifdef CONFIG_FAIR_GROUP_SCHED |
fe5c7cc2 PM |
8523 | { |
8524 | .name = "shares", | |
f4c753b7 PM |
8525 | .read_u64 = cpu_shares_read_u64, |
8526 | .write_u64 = cpu_shares_write_u64, | |
fe5c7cc2 | 8527 | }, |
052f1dc7 | 8528 | #endif |
ab84d31e PT |
8529 | #ifdef CONFIG_CFS_BANDWIDTH |
8530 | { | |
8531 | .name = "cfs_quota_us", | |
8532 | .read_s64 = cpu_cfs_quota_read_s64, | |
8533 | .write_s64 = cpu_cfs_quota_write_s64, | |
8534 | }, | |
8535 | { | |
8536 | .name = "cfs_period_us", | |
8537 | .read_u64 = cpu_cfs_period_read_u64, | |
8538 | .write_u64 = cpu_cfs_period_write_u64, | |
8539 | }, | |
e8da1b18 NR |
8540 | { |
8541 | .name = "stat", | |
2da8ca82 | 8542 | .seq_show = cpu_stats_show, |
e8da1b18 | 8543 | }, |
ab84d31e | 8544 | #endif |
052f1dc7 | 8545 | #ifdef CONFIG_RT_GROUP_SCHED |
6f505b16 | 8546 | { |
9f0c1e56 | 8547 | .name = "rt_runtime_us", |
06ecb27c PM |
8548 | .read_s64 = cpu_rt_runtime_read, |
8549 | .write_s64 = cpu_rt_runtime_write, | |
6f505b16 | 8550 | }, |
d0b27fa7 PZ |
8551 | { |
8552 | .name = "rt_period_us", | |
f4c753b7 PM |
8553 | .read_u64 = cpu_rt_period_read_uint, |
8554 | .write_u64 = cpu_rt_period_write_uint, | |
d0b27fa7 | 8555 | }, |
052f1dc7 | 8556 | #endif |
4baf6e33 | 8557 | { } /* terminate */ |
68318b8e SV |
8558 | }; |
8559 | ||
073219e9 | 8560 | struct cgroup_subsys cpu_cgrp_subsys = { |
92fb9748 | 8561 | .css_alloc = cpu_cgroup_css_alloc, |
2f5177f0 | 8562 | .css_released = cpu_cgroup_css_released, |
92fb9748 | 8563 | .css_free = cpu_cgroup_css_free, |
eeb61e53 | 8564 | .fork = cpu_cgroup_fork, |
bb9d97b6 TH |
8565 | .can_attach = cpu_cgroup_can_attach, |
8566 | .attach = cpu_cgroup_attach, | |
5577964e | 8567 | .legacy_cftypes = cpu_files, |
b38e42e9 | 8568 | .early_init = true, |
68318b8e SV |
8569 | }; |
8570 | ||
052f1dc7 | 8571 | #endif /* CONFIG_CGROUP_SCHED */ |
d842de87 | 8572 | |
b637a328 PM |
8573 | void dump_cpu_task(int cpu) |
8574 | { | |
8575 | pr_info("Task dump for CPU %d:\n", cpu); | |
8576 | sched_show_task(cpu_curr(cpu)); | |
8577 | } | |
ed82b8a1 AK |
8578 | |
8579 | /* | |
8580 | * Nice levels are multiplicative, with a gentle 10% change for every | |
8581 | * nice level changed. I.e. when a CPU-bound task goes from nice 0 to | |
8582 | * nice 1, it will get ~10% less CPU time than another CPU-bound task | |
8583 | * that remained on nice 0. | |
8584 | * | |
8585 | * The "10% effect" is relative and cumulative: from _any_ nice level, | |
8586 | * if you go up 1 level, it's -10% CPU usage, if you go down 1 level | |
8587 | * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25. | |
8588 | * If a task goes up by ~10% and another task goes down by ~10% then | |
8589 | * the relative distance between them is ~25%.) | |
8590 | */ | |
8591 | const int sched_prio_to_weight[40] = { | |
8592 | /* -20 */ 88761, 71755, 56483, 46273, 36291, | |
8593 | /* -15 */ 29154, 23254, 18705, 14949, 11916, | |
8594 | /* -10 */ 9548, 7620, 6100, 4904, 3906, | |
8595 | /* -5 */ 3121, 2501, 1991, 1586, 1277, | |
8596 | /* 0 */ 1024, 820, 655, 526, 423, | |
8597 | /* 5 */ 335, 272, 215, 172, 137, | |
8598 | /* 10 */ 110, 87, 70, 56, 45, | |
8599 | /* 15 */ 36, 29, 23, 18, 15, | |
8600 | }; | |
8601 | ||
8602 | /* | |
8603 | * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated. | |
8604 | * | |
8605 | * In cases where the weight does not change often, we can use the | |
8606 | * precalculated inverse to speed up arithmetics by turning divisions | |
8607 | * into multiplications: | |
8608 | */ | |
8609 | const u32 sched_prio_to_wmult[40] = { | |
8610 | /* -20 */ 48388, 59856, 76040, 92818, 118348, | |
8611 | /* -15 */ 147320, 184698, 229616, 287308, 360437, | |
8612 | /* -10 */ 449829, 563644, 704093, 875809, 1099582, | |
8613 | /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326, | |
8614 | /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587, | |
8615 | /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126, | |
8616 | /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717, | |
8617 | /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153, | |
8618 | }; |