]>
Commit | Line | Data |
---|---|---|
79bf2bb3 TG |
1 | /* |
2 | * linux/kernel/time/tick-sched.c | |
3 | * | |
4 | * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de> | |
5 | * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar | |
6 | * Copyright(C) 2006-2007 Timesys Corp., Thomas Gleixner | |
7 | * | |
8 | * No idle tick implementation for low and high resolution timers | |
9 | * | |
10 | * Started by: Thomas Gleixner and Ingo Molnar | |
11 | * | |
12 | * For licencing details see kernel-base/COPYING | |
13 | */ | |
14 | #include <linux/cpu.h> | |
15 | #include <linux/err.h> | |
16 | #include <linux/hrtimer.h> | |
17 | #include <linux/interrupt.h> | |
18 | #include <linux/kernel_stat.h> | |
19 | #include <linux/percpu.h> | |
20 | #include <linux/profile.h> | |
21 | #include <linux/sched.h> | |
22 | #include <linux/tick.h> | |
23 | ||
24 | #include "tick-internal.h" | |
25 | ||
26 | /* | |
27 | * Per cpu nohz control structure | |
28 | */ | |
29 | static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched); | |
30 | ||
31 | /* | |
32 | * The time, when the last jiffy update happened. Protected by xtime_lock. | |
33 | */ | |
34 | static ktime_t last_jiffies_update; | |
35 | ||
289f480a IM |
36 | struct tick_sched *tick_get_tick_sched(int cpu) |
37 | { | |
38 | return &per_cpu(tick_cpu_sched, cpu); | |
39 | } | |
40 | ||
79bf2bb3 TG |
41 | /* |
42 | * Must be called with interrupts disabled ! | |
43 | */ | |
44 | static void tick_do_update_jiffies64(ktime_t now) | |
45 | { | |
46 | unsigned long ticks = 0; | |
47 | ktime_t delta; | |
48 | ||
49 | /* Reevalute with xtime_lock held */ | |
50 | write_seqlock(&xtime_lock); | |
51 | ||
52 | delta = ktime_sub(now, last_jiffies_update); | |
53 | if (delta.tv64 >= tick_period.tv64) { | |
54 | ||
55 | delta = ktime_sub(delta, tick_period); | |
56 | last_jiffies_update = ktime_add(last_jiffies_update, | |
57 | tick_period); | |
58 | ||
59 | /* Slow path for long timeouts */ | |
60 | if (unlikely(delta.tv64 >= tick_period.tv64)) { | |
61 | s64 incr = ktime_to_ns(tick_period); | |
62 | ||
63 | ticks = ktime_divns(delta, incr); | |
64 | ||
65 | last_jiffies_update = ktime_add_ns(last_jiffies_update, | |
66 | incr * ticks); | |
67 | } | |
68 | do_timer(++ticks); | |
69 | } | |
70 | write_sequnlock(&xtime_lock); | |
71 | } | |
72 | ||
73 | /* | |
74 | * Initialize and return retrieve the jiffies update. | |
75 | */ | |
76 | static ktime_t tick_init_jiffy_update(void) | |
77 | { | |
78 | ktime_t period; | |
79 | ||
80 | write_seqlock(&xtime_lock); | |
81 | /* Did we start the jiffies update yet ? */ | |
82 | if (last_jiffies_update.tv64 == 0) | |
83 | last_jiffies_update = tick_next_period; | |
84 | period = last_jiffies_update; | |
85 | write_sequnlock(&xtime_lock); | |
86 | return period; | |
87 | } | |
88 | ||
89 | /* | |
90 | * NOHZ - aka dynamic tick functionality | |
91 | */ | |
92 | #ifdef CONFIG_NO_HZ | |
93 | /* | |
94 | * NO HZ enabled ? | |
95 | */ | |
96 | static int tick_nohz_enabled __read_mostly = 1; | |
97 | ||
98 | /* | |
99 | * Enable / Disable tickless mode | |
100 | */ | |
101 | static int __init setup_tick_nohz(char *str) | |
102 | { | |
103 | if (!strcmp(str, "off")) | |
104 | tick_nohz_enabled = 0; | |
105 | else if (!strcmp(str, "on")) | |
106 | tick_nohz_enabled = 1; | |
107 | else | |
108 | return 0; | |
109 | return 1; | |
110 | } | |
111 | ||
112 | __setup("nohz=", setup_tick_nohz); | |
113 | ||
114 | /** | |
115 | * tick_nohz_update_jiffies - update jiffies when idle was interrupted | |
116 | * | |
117 | * Called from interrupt entry when the CPU was idle | |
118 | * | |
119 | * In case the sched_tick was stopped on this CPU, we have to check if jiffies | |
120 | * must be updated. Otherwise an interrupt handler could use a stale jiffy | |
121 | * value. We do this unconditionally on any cpu, as we don't know whether the | |
122 | * cpu, which has the update task assigned is in a long sleep. | |
123 | */ | |
124 | void tick_nohz_update_jiffies(void) | |
125 | { | |
126 | int cpu = smp_processor_id(); | |
127 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
128 | unsigned long flags; | |
129 | ktime_t now; | |
130 | ||
131 | if (!ts->tick_stopped) | |
132 | return; | |
133 | ||
134 | cpu_clear(cpu, nohz_cpu_mask); | |
135 | now = ktime_get(); | |
136 | ||
137 | local_irq_save(flags); | |
138 | tick_do_update_jiffies64(now); | |
139 | local_irq_restore(flags); | |
140 | } | |
141 | ||
142 | /** | |
143 | * tick_nohz_stop_sched_tick - stop the idle tick from the idle task | |
144 | * | |
145 | * When the next event is more than a tick into the future, stop the idle tick | |
146 | * Called either from the idle loop or from irq_exit() when an idle period was | |
147 | * just interrupted by an interrupt which did not cause a reschedule. | |
148 | */ | |
149 | void tick_nohz_stop_sched_tick(void) | |
150 | { | |
151 | unsigned long seq, last_jiffies, next_jiffies, delta_jiffies, flags; | |
152 | struct tick_sched *ts; | |
153 | ktime_t last_update, expires, now, delta; | |
154 | int cpu; | |
155 | ||
156 | local_irq_save(flags); | |
157 | ||
158 | cpu = smp_processor_id(); | |
159 | ts = &per_cpu(tick_cpu_sched, cpu); | |
160 | ||
161 | if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) | |
162 | goto end; | |
163 | ||
164 | if (need_resched()) | |
165 | goto end; | |
166 | ||
167 | cpu = smp_processor_id(); | |
168 | BUG_ON(local_softirq_pending()); | |
169 | ||
170 | now = ktime_get(); | |
171 | /* | |
172 | * When called from irq_exit we need to account the idle sleep time | |
173 | * correctly. | |
174 | */ | |
175 | if (ts->tick_stopped) { | |
176 | delta = ktime_sub(now, ts->idle_entrytime); | |
177 | ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); | |
178 | } | |
179 | ||
180 | ts->idle_entrytime = now; | |
181 | ts->idle_calls++; | |
182 | ||
183 | /* Read jiffies and the time when jiffies were updated last */ | |
184 | do { | |
185 | seq = read_seqbegin(&xtime_lock); | |
186 | last_update = last_jiffies_update; | |
187 | last_jiffies = jiffies; | |
188 | } while (read_seqretry(&xtime_lock, seq)); | |
189 | ||
190 | /* Get the next timer wheel timer */ | |
191 | next_jiffies = get_next_timer_interrupt(last_jiffies); | |
192 | delta_jiffies = next_jiffies - last_jiffies; | |
193 | ||
6ba9b346 IM |
194 | if (rcu_needs_cpu(cpu)) |
195 | delta_jiffies = 1; | |
79bf2bb3 TG |
196 | /* |
197 | * Do not stop the tick, if we are only one off | |
198 | * or if the cpu is required for rcu | |
199 | */ | |
6ba9b346 | 200 | if (!ts->tick_stopped && delta_jiffies == 1) |
79bf2bb3 TG |
201 | goto out; |
202 | ||
203 | /* Schedule the tick, if we are at least one jiffie off */ | |
204 | if ((long)delta_jiffies >= 1) { | |
205 | ||
6ba9b346 | 206 | if (delta_jiffies > 1) |
79bf2bb3 TG |
207 | cpu_set(cpu, nohz_cpu_mask); |
208 | /* | |
209 | * nohz_stop_sched_tick can be called several times before | |
210 | * the nohz_restart_sched_tick is called. This happens when | |
211 | * interrupts arrive which do not cause a reschedule. In the | |
212 | * first call we save the current tick time, so we can restart | |
213 | * the scheduler tick in nohz_restart_sched_tick. | |
214 | */ | |
215 | if (!ts->tick_stopped) { | |
216 | ts->idle_tick = ts->sched_timer.expires; | |
217 | ts->tick_stopped = 1; | |
218 | ts->idle_jiffies = last_jiffies; | |
219 | } | |
220 | /* | |
221 | * calculate the expiry time for the next timer wheel | |
222 | * timer | |
223 | */ | |
224 | expires = ktime_add_ns(last_update, tick_period.tv64 * | |
225 | delta_jiffies); | |
226 | ts->idle_expires = expires; | |
227 | ts->idle_sleeps++; | |
228 | ||
229 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { | |
230 | hrtimer_start(&ts->sched_timer, expires, | |
231 | HRTIMER_MODE_ABS); | |
232 | /* Check, if the timer was already in the past */ | |
233 | if (hrtimer_active(&ts->sched_timer)) | |
234 | goto out; | |
235 | } else if(!tick_program_event(expires, 0)) | |
236 | goto out; | |
237 | /* | |
238 | * We are past the event already. So we crossed a | |
239 | * jiffie boundary. Update jiffies and raise the | |
240 | * softirq. | |
241 | */ | |
242 | tick_do_update_jiffies64(ktime_get()); | |
243 | cpu_clear(cpu, nohz_cpu_mask); | |
244 | } | |
245 | raise_softirq_irqoff(TIMER_SOFTIRQ); | |
246 | out: | |
247 | ts->next_jiffies = next_jiffies; | |
248 | ts->last_jiffies = last_jiffies; | |
249 | end: | |
250 | local_irq_restore(flags); | |
251 | } | |
252 | ||
253 | /** | |
254 | * nohz_restart_sched_tick - restart the idle tick from the idle task | |
255 | * | |
256 | * Restart the idle tick when the CPU is woken up from idle | |
257 | */ | |
258 | void tick_nohz_restart_sched_tick(void) | |
259 | { | |
260 | int cpu = smp_processor_id(); | |
261 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
262 | unsigned long ticks; | |
263 | ktime_t now, delta; | |
264 | ||
265 | if (!ts->tick_stopped) | |
266 | return; | |
267 | ||
268 | /* Update jiffies first */ | |
269 | now = ktime_get(); | |
270 | ||
271 | local_irq_disable(); | |
272 | tick_do_update_jiffies64(now); | |
273 | cpu_clear(cpu, nohz_cpu_mask); | |
274 | ||
275 | /* Account the idle time */ | |
276 | delta = ktime_sub(now, ts->idle_entrytime); | |
277 | ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); | |
278 | ||
279 | /* | |
280 | * We stopped the tick in idle. Update process times would miss the | |
281 | * time we slept as update_process_times does only a 1 tick | |
282 | * accounting. Enforce that this is accounted to idle ! | |
283 | */ | |
284 | ticks = jiffies - ts->idle_jiffies; | |
285 | /* | |
286 | * We might be one off. Do not randomly account a huge number of ticks! | |
287 | */ | |
288 | if (ticks && ticks < LONG_MAX) { | |
289 | add_preempt_count(HARDIRQ_OFFSET); | |
290 | account_system_time(current, HARDIRQ_OFFSET, | |
291 | jiffies_to_cputime(ticks)); | |
292 | sub_preempt_count(HARDIRQ_OFFSET); | |
293 | } | |
294 | ||
295 | /* | |
296 | * Cancel the scheduled timer and restore the tick | |
297 | */ | |
298 | ts->tick_stopped = 0; | |
299 | hrtimer_cancel(&ts->sched_timer); | |
300 | ts->sched_timer.expires = ts->idle_tick; | |
301 | ||
302 | while (1) { | |
303 | /* Forward the time to expire in the future */ | |
304 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
305 | ||
306 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { | |
307 | hrtimer_start(&ts->sched_timer, | |
308 | ts->sched_timer.expires, | |
309 | HRTIMER_MODE_ABS); | |
310 | /* Check, if the timer was already in the past */ | |
311 | if (hrtimer_active(&ts->sched_timer)) | |
312 | break; | |
313 | } else { | |
314 | if (!tick_program_event(ts->sched_timer.expires, 0)) | |
315 | break; | |
316 | } | |
317 | /* Update jiffies and reread time */ | |
318 | tick_do_update_jiffies64(now); | |
319 | now = ktime_get(); | |
320 | } | |
321 | local_irq_enable(); | |
322 | } | |
323 | ||
324 | static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now) | |
325 | { | |
326 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
327 | return tick_program_event(ts->sched_timer.expires, 0); | |
328 | } | |
329 | ||
330 | /* | |
331 | * The nohz low res interrupt handler | |
332 | */ | |
333 | static void tick_nohz_handler(struct clock_event_device *dev) | |
334 | { | |
335 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
336 | struct pt_regs *regs = get_irq_regs(); | |
337 | ktime_t now = ktime_get(); | |
338 | ||
339 | dev->next_event.tv64 = KTIME_MAX; | |
340 | ||
341 | /* Check, if the jiffies need an update */ | |
342 | tick_do_update_jiffies64(now); | |
343 | ||
344 | /* | |
345 | * When we are idle and the tick is stopped, we have to touch | |
346 | * the watchdog as we might not schedule for a really long | |
347 | * time. This happens on complete idle SMP systems while | |
348 | * waiting on the login prompt. We also increment the "start | |
349 | * of idle" jiffy stamp so the idle accounting adjustment we | |
350 | * do when we go busy again does not account too much ticks. | |
351 | */ | |
352 | if (ts->tick_stopped) { | |
353 | touch_softlockup_watchdog(); | |
354 | ts->idle_jiffies++; | |
355 | } | |
356 | ||
357 | update_process_times(user_mode(regs)); | |
358 | profile_tick(CPU_PROFILING); | |
359 | ||
360 | /* Do not restart, when we are in the idle loop */ | |
361 | if (ts->tick_stopped) | |
362 | return; | |
363 | ||
364 | while (tick_nohz_reprogram(ts, now)) { | |
365 | now = ktime_get(); | |
366 | tick_do_update_jiffies64(now); | |
367 | } | |
368 | } | |
369 | ||
370 | /** | |
371 | * tick_nohz_switch_to_nohz - switch to nohz mode | |
372 | */ | |
373 | static void tick_nohz_switch_to_nohz(void) | |
374 | { | |
375 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
376 | ktime_t next; | |
377 | ||
378 | if (!tick_nohz_enabled) | |
379 | return; | |
380 | ||
381 | local_irq_disable(); | |
382 | if (tick_switch_to_oneshot(tick_nohz_handler)) { | |
383 | local_irq_enable(); | |
384 | return; | |
385 | } | |
386 | ||
387 | ts->nohz_mode = NOHZ_MODE_LOWRES; | |
388 | ||
389 | /* | |
390 | * Recycle the hrtimer in ts, so we can share the | |
391 | * hrtimer_forward with the highres code. | |
392 | */ | |
393 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); | |
394 | /* Get the next period */ | |
395 | next = tick_init_jiffy_update(); | |
396 | ||
397 | for (;;) { | |
398 | ts->sched_timer.expires = next; | |
399 | if (!tick_program_event(next, 0)) | |
400 | break; | |
401 | next = ktime_add(next, tick_period); | |
402 | } | |
403 | local_irq_enable(); | |
404 | ||
405 | printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n", | |
406 | smp_processor_id()); | |
407 | } | |
408 | ||
409 | #else | |
410 | ||
411 | static inline void tick_nohz_switch_to_nohz(void) { } | |
412 | ||
413 | #endif /* NO_HZ */ | |
414 | ||
415 | /* | |
416 | * High resolution timer specific code | |
417 | */ | |
418 | #ifdef CONFIG_HIGH_RES_TIMERS | |
419 | /* | |
420 | * We rearm the timer until we get disabled by the idle code | |
421 | * Called with interrupts disabled and timer->base->cpu_base->lock held. | |
422 | */ | |
423 | static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) | |
424 | { | |
425 | struct tick_sched *ts = | |
426 | container_of(timer, struct tick_sched, sched_timer); | |
427 | struct hrtimer_cpu_base *base = timer->base->cpu_base; | |
428 | struct pt_regs *regs = get_irq_regs(); | |
429 | ktime_t now = ktime_get(); | |
430 | ||
431 | /* Check, if the jiffies need an update */ | |
432 | tick_do_update_jiffies64(now); | |
433 | ||
434 | /* | |
435 | * Do not call, when we are not in irq context and have | |
436 | * no valid regs pointer | |
437 | */ | |
438 | if (regs) { | |
439 | /* | |
440 | * When we are idle and the tick is stopped, we have to touch | |
441 | * the watchdog as we might not schedule for a really long | |
442 | * time. This happens on complete idle SMP systems while | |
443 | * waiting on the login prompt. We also increment the "start of | |
444 | * idle" jiffy stamp so the idle accounting adjustment we do | |
445 | * when we go busy again does not account too much ticks. | |
446 | */ | |
447 | if (ts->tick_stopped) { | |
448 | touch_softlockup_watchdog(); | |
449 | ts->idle_jiffies++; | |
450 | } | |
451 | /* | |
452 | * update_process_times() might take tasklist_lock, hence | |
453 | * drop the base lock. sched-tick hrtimers are per-CPU and | |
454 | * never accessible by userspace APIs, so this is safe to do. | |
455 | */ | |
456 | spin_unlock(&base->lock); | |
457 | update_process_times(user_mode(regs)); | |
458 | profile_tick(CPU_PROFILING); | |
459 | spin_lock(&base->lock); | |
460 | } | |
461 | ||
462 | /* Do not restart, when we are in the idle loop */ | |
463 | if (ts->tick_stopped) | |
464 | return HRTIMER_NORESTART; | |
465 | ||
466 | hrtimer_forward(timer, now, tick_period); | |
467 | ||
468 | return HRTIMER_RESTART; | |
469 | } | |
470 | ||
471 | /** | |
472 | * tick_setup_sched_timer - setup the tick emulation timer | |
473 | */ | |
474 | void tick_setup_sched_timer(void) | |
475 | { | |
476 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
477 | ktime_t now = ktime_get(); | |
478 | ||
479 | /* | |
480 | * Emulate tick processing via per-CPU hrtimers: | |
481 | */ | |
482 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); | |
483 | ts->sched_timer.function = tick_sched_timer; | |
484 | ts->sched_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ; | |
485 | ||
486 | /* Get the next period */ | |
487 | ts->sched_timer.expires = tick_init_jiffy_update(); | |
488 | ||
489 | for (;;) { | |
490 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
491 | hrtimer_start(&ts->sched_timer, ts->sched_timer.expires, | |
492 | HRTIMER_MODE_ABS); | |
493 | /* Check, if the timer was already in the past */ | |
494 | if (hrtimer_active(&ts->sched_timer)) | |
495 | break; | |
496 | now = ktime_get(); | |
497 | } | |
498 | ||
499 | #ifdef CONFIG_NO_HZ | |
500 | if (tick_nohz_enabled) | |
501 | ts->nohz_mode = NOHZ_MODE_HIGHRES; | |
502 | #endif | |
503 | } | |
504 | ||
505 | void tick_cancel_sched_timer(int cpu) | |
506 | { | |
507 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
508 | ||
509 | if (ts->sched_timer.base) | |
510 | hrtimer_cancel(&ts->sched_timer); | |
511 | ts->tick_stopped = 0; | |
512 | ts->nohz_mode = NOHZ_MODE_INACTIVE; | |
513 | } | |
514 | #endif /* HIGH_RES_TIMERS */ | |
515 | ||
516 | /** | |
517 | * Async notification about clocksource changes | |
518 | */ | |
519 | void tick_clock_notify(void) | |
520 | { | |
521 | int cpu; | |
522 | ||
523 | for_each_possible_cpu(cpu) | |
524 | set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); | |
525 | } | |
526 | ||
527 | /* | |
528 | * Async notification about clock event changes | |
529 | */ | |
530 | void tick_oneshot_notify(void) | |
531 | { | |
532 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
533 | ||
534 | set_bit(0, &ts->check_clocks); | |
535 | } | |
536 | ||
537 | /** | |
538 | * Check, if a change happened, which makes oneshot possible. | |
539 | * | |
540 | * Called cyclic from the hrtimer softirq (driven by the timer | |
541 | * softirq) allow_nohz signals, that we can switch into low-res nohz | |
542 | * mode, because high resolution timers are disabled (either compile | |
543 | * or runtime). | |
544 | */ | |
545 | int tick_check_oneshot_change(int allow_nohz) | |
546 | { | |
547 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
548 | ||
549 | if (!test_and_clear_bit(0, &ts->check_clocks)) | |
550 | return 0; | |
551 | ||
552 | if (ts->nohz_mode != NOHZ_MODE_INACTIVE) | |
553 | return 0; | |
554 | ||
555 | if (!timekeeping_is_continuous() || !tick_is_oneshot_available()) | |
556 | return 0; | |
557 | ||
558 | if (!allow_nohz) | |
559 | return 1; | |
560 | ||
561 | tick_nohz_switch_to_nohz(); | |
562 | return 0; | |
563 | } |