]> git.proxmox.com Git - mirror_ubuntu-kernels.git/blob - kernel/time/tick-sched.c
Merge branches 'release' and 'menlo' into release
[mirror_ubuntu-kernels.git] / kernel / time / tick-sched.c
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 * Distribute under GPLv2.
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 <asm/irq_regs.h>
25
26 #include "tick-internal.h"
27
28 /*
29 * Per cpu nohz control structure
30 */
31 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
32
33 /*
34 * The time, when the last jiffy update happened. Protected by xtime_lock.
35 */
36 static ktime_t last_jiffies_update;
37
38 struct tick_sched *tick_get_tick_sched(int cpu)
39 {
40 return &per_cpu(tick_cpu_sched, cpu);
41 }
42
43 /*
44 * Must be called with interrupts disabled !
45 */
46 static void tick_do_update_jiffies64(ktime_t now)
47 {
48 unsigned long ticks = 0;
49 ktime_t delta;
50
51 /* Reevalute with xtime_lock held */
52 write_seqlock(&xtime_lock);
53
54 delta = ktime_sub(now, last_jiffies_update);
55 if (delta.tv64 >= tick_period.tv64) {
56
57 delta = ktime_sub(delta, tick_period);
58 last_jiffies_update = ktime_add(last_jiffies_update,
59 tick_period);
60
61 /* Slow path for long timeouts */
62 if (unlikely(delta.tv64 >= tick_period.tv64)) {
63 s64 incr = ktime_to_ns(tick_period);
64
65 ticks = ktime_divns(delta, incr);
66
67 last_jiffies_update = ktime_add_ns(last_jiffies_update,
68 incr * ticks);
69 }
70 do_timer(++ticks);
71 }
72 write_sequnlock(&xtime_lock);
73 }
74
75 /*
76 * Initialize and return retrieve the jiffies update.
77 */
78 static ktime_t tick_init_jiffy_update(void)
79 {
80 ktime_t period;
81
82 write_seqlock(&xtime_lock);
83 /* Did we start the jiffies update yet ? */
84 if (last_jiffies_update.tv64 == 0)
85 last_jiffies_update = tick_next_period;
86 period = last_jiffies_update;
87 write_sequnlock(&xtime_lock);
88 return period;
89 }
90
91 /*
92 * NOHZ - aka dynamic tick functionality
93 */
94 #ifdef CONFIG_NO_HZ
95 /*
96 * NO HZ enabled ?
97 */
98 static int tick_nohz_enabled __read_mostly = 1;
99
100 /*
101 * Enable / Disable tickless mode
102 */
103 static int __init setup_tick_nohz(char *str)
104 {
105 if (!strcmp(str, "off"))
106 tick_nohz_enabled = 0;
107 else if (!strcmp(str, "on"))
108 tick_nohz_enabled = 1;
109 else
110 return 0;
111 return 1;
112 }
113
114 __setup("nohz=", setup_tick_nohz);
115
116 /**
117 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
118 *
119 * Called from interrupt entry when the CPU was idle
120 *
121 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
122 * must be updated. Otherwise an interrupt handler could use a stale jiffy
123 * value. We do this unconditionally on any cpu, as we don't know whether the
124 * cpu, which has the update task assigned is in a long sleep.
125 */
126 void tick_nohz_update_jiffies(void)
127 {
128 int cpu = smp_processor_id();
129 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
130 unsigned long flags;
131 ktime_t now;
132
133 if (!ts->tick_stopped)
134 return;
135
136 touch_softlockup_watchdog();
137
138 cpu_clear(cpu, nohz_cpu_mask);
139 now = ktime_get();
140 ts->idle_waketime = now;
141
142 local_irq_save(flags);
143 tick_do_update_jiffies64(now);
144 local_irq_restore(flags);
145 }
146
147 void tick_nohz_stop_idle(int cpu)
148 {
149 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
150
151 if (ts->idle_active) {
152 ktime_t now, delta;
153 now = ktime_get();
154 delta = ktime_sub(now, ts->idle_entrytime);
155 ts->idle_lastupdate = now;
156 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
157 ts->idle_active = 0;
158 }
159 }
160
161 static ktime_t tick_nohz_start_idle(int cpu)
162 {
163 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
164 ktime_t now, delta;
165
166 now = ktime_get();
167 if (ts->idle_active) {
168 delta = ktime_sub(now, ts->idle_entrytime);
169 ts->idle_lastupdate = now;
170 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
171 }
172 ts->idle_entrytime = now;
173 ts->idle_active = 1;
174 return now;
175 }
176
177 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
178 {
179 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
180
181 *last_update_time = ktime_to_us(ts->idle_lastupdate);
182 return ktime_to_us(ts->idle_sleeptime);
183 }
184
185 /**
186 * tick_nohz_stop_sched_tick - stop the idle tick from the idle task
187 *
188 * When the next event is more than a tick into the future, stop the idle tick
189 * Called either from the idle loop or from irq_exit() when an idle period was
190 * just interrupted by an interrupt which did not cause a reschedule.
191 */
192 void tick_nohz_stop_sched_tick(void)
193 {
194 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies, flags;
195 unsigned long rt_jiffies;
196 struct tick_sched *ts;
197 ktime_t last_update, expires, now;
198 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
199 int cpu;
200
201 local_irq_save(flags);
202
203 cpu = smp_processor_id();
204 now = tick_nohz_start_idle(cpu);
205 ts = &per_cpu(tick_cpu_sched, cpu);
206
207 /*
208 * If this cpu is offline and it is the one which updates
209 * jiffies, then give up the assignment and let it be taken by
210 * the cpu which runs the tick timer next. If we don't drop
211 * this here the jiffies might be stale and do_timer() never
212 * invoked.
213 */
214 if (unlikely(!cpu_online(cpu))) {
215 if (cpu == tick_do_timer_cpu)
216 tick_do_timer_cpu = -1;
217 }
218
219 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
220 goto end;
221
222 if (need_resched())
223 goto end;
224
225 cpu = smp_processor_id();
226 if (unlikely(local_softirq_pending())) {
227 static int ratelimit;
228
229 if (ratelimit < 10) {
230 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n",
231 local_softirq_pending());
232 ratelimit++;
233 }
234 }
235
236 ts->idle_calls++;
237 /* Read jiffies and the time when jiffies were updated last */
238 do {
239 seq = read_seqbegin(&xtime_lock);
240 last_update = last_jiffies_update;
241 last_jiffies = jiffies;
242 } while (read_seqretry(&xtime_lock, seq));
243
244 /* Get the next timer wheel timer */
245 next_jiffies = get_next_timer_interrupt(last_jiffies);
246 delta_jiffies = next_jiffies - last_jiffies;
247
248 rt_jiffies = rt_needs_cpu(cpu);
249 if (rt_jiffies && rt_jiffies < delta_jiffies)
250 delta_jiffies = rt_jiffies;
251
252 if (rcu_needs_cpu(cpu))
253 delta_jiffies = 1;
254 /*
255 * Do not stop the tick, if we are only one off
256 * or if the cpu is required for rcu
257 */
258 if (!ts->tick_stopped && delta_jiffies == 1)
259 goto out;
260
261 /* Schedule the tick, if we are at least one jiffie off */
262 if ((long)delta_jiffies >= 1) {
263
264 if (delta_jiffies > 1)
265 cpu_set(cpu, nohz_cpu_mask);
266 /*
267 * nohz_stop_sched_tick can be called several times before
268 * the nohz_restart_sched_tick is called. This happens when
269 * interrupts arrive which do not cause a reschedule. In the
270 * first call we save the current tick time, so we can restart
271 * the scheduler tick in nohz_restart_sched_tick.
272 */
273 if (!ts->tick_stopped) {
274 if (select_nohz_load_balancer(1)) {
275 /*
276 * sched tick not stopped!
277 */
278 cpu_clear(cpu, nohz_cpu_mask);
279 goto out;
280 }
281
282 ts->idle_tick = ts->sched_timer.expires;
283 ts->tick_stopped = 1;
284 ts->idle_jiffies = last_jiffies;
285 }
286
287 /*
288 * If this cpu is the one which updates jiffies, then
289 * give up the assignment and let it be taken by the
290 * cpu which runs the tick timer next, which might be
291 * this cpu as well. If we don't drop this here the
292 * jiffies might be stale and do_timer() never
293 * invoked.
294 */
295 if (cpu == tick_do_timer_cpu)
296 tick_do_timer_cpu = -1;
297
298 ts->idle_sleeps++;
299
300 /*
301 * delta_jiffies >= NEXT_TIMER_MAX_DELTA signals that
302 * there is no timer pending or at least extremly far
303 * into the future (12 days for HZ=1000). In this case
304 * we simply stop the tick timer:
305 */
306 if (unlikely(delta_jiffies >= NEXT_TIMER_MAX_DELTA)) {
307 ts->idle_expires.tv64 = KTIME_MAX;
308 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
309 hrtimer_cancel(&ts->sched_timer);
310 goto out;
311 }
312
313 /*
314 * calculate the expiry time for the next timer wheel
315 * timer
316 */
317 expires = ktime_add_ns(last_update, tick_period.tv64 *
318 delta_jiffies);
319 ts->idle_expires = expires;
320
321 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
322 hrtimer_start(&ts->sched_timer, expires,
323 HRTIMER_MODE_ABS);
324 /* Check, if the timer was already in the past */
325 if (hrtimer_active(&ts->sched_timer))
326 goto out;
327 } else if (!tick_program_event(expires, 0))
328 goto out;
329 /*
330 * We are past the event already. So we crossed a
331 * jiffie boundary. Update jiffies and raise the
332 * softirq.
333 */
334 tick_do_update_jiffies64(ktime_get());
335 cpu_clear(cpu, nohz_cpu_mask);
336 }
337 raise_softirq_irqoff(TIMER_SOFTIRQ);
338 out:
339 ts->next_jiffies = next_jiffies;
340 ts->last_jiffies = last_jiffies;
341 ts->sleep_length = ktime_sub(dev->next_event, now);
342 end:
343 local_irq_restore(flags);
344 }
345
346 /**
347 * tick_nohz_get_sleep_length - return the length of the current sleep
348 *
349 * Called from power state control code with interrupts disabled
350 */
351 ktime_t tick_nohz_get_sleep_length(void)
352 {
353 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
354
355 return ts->sleep_length;
356 }
357
358 /**
359 * tick_nohz_restart_sched_tick - restart the idle tick from the idle task
360 *
361 * Restart the idle tick when the CPU is woken up from idle
362 */
363 void tick_nohz_restart_sched_tick(void)
364 {
365 int cpu = smp_processor_id();
366 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
367 unsigned long ticks;
368 ktime_t now;
369
370 local_irq_disable();
371 tick_nohz_stop_idle(cpu);
372
373 if (!ts->tick_stopped) {
374 local_irq_enable();
375 return;
376 }
377
378 /* Update jiffies first */
379 select_nohz_load_balancer(0);
380 now = ktime_get();
381 tick_do_update_jiffies64(now);
382 cpu_clear(cpu, nohz_cpu_mask);
383
384 /*
385 * We stopped the tick in idle. Update process times would miss the
386 * time we slept as update_process_times does only a 1 tick
387 * accounting. Enforce that this is accounted to idle !
388 */
389 ticks = jiffies - ts->idle_jiffies;
390 /*
391 * We might be one off. Do not randomly account a huge number of ticks!
392 */
393 if (ticks && ticks < LONG_MAX) {
394 add_preempt_count(HARDIRQ_OFFSET);
395 account_system_time(current, HARDIRQ_OFFSET,
396 jiffies_to_cputime(ticks));
397 sub_preempt_count(HARDIRQ_OFFSET);
398 }
399
400 /*
401 * Cancel the scheduled timer and restore the tick
402 */
403 ts->tick_stopped = 0;
404 ts->idle_exittime = now;
405 hrtimer_cancel(&ts->sched_timer);
406 ts->sched_timer.expires = ts->idle_tick;
407
408 while (1) {
409 /* Forward the time to expire in the future */
410 hrtimer_forward(&ts->sched_timer, now, tick_period);
411
412 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
413 hrtimer_start(&ts->sched_timer,
414 ts->sched_timer.expires,
415 HRTIMER_MODE_ABS);
416 /* Check, if the timer was already in the past */
417 if (hrtimer_active(&ts->sched_timer))
418 break;
419 } else {
420 if (!tick_program_event(ts->sched_timer.expires, 0))
421 break;
422 }
423 /* Update jiffies and reread time */
424 tick_do_update_jiffies64(now);
425 now = ktime_get();
426 }
427 local_irq_enable();
428 }
429
430 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
431 {
432 hrtimer_forward(&ts->sched_timer, now, tick_period);
433 return tick_program_event(ts->sched_timer.expires, 0);
434 }
435
436 /*
437 * The nohz low res interrupt handler
438 */
439 static void tick_nohz_handler(struct clock_event_device *dev)
440 {
441 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
442 struct pt_regs *regs = get_irq_regs();
443 int cpu = smp_processor_id();
444 ktime_t now = ktime_get();
445
446 dev->next_event.tv64 = KTIME_MAX;
447
448 /*
449 * Check if the do_timer duty was dropped. We don't care about
450 * concurrency: This happens only when the cpu in charge went
451 * into a long sleep. If two cpus happen to assign themself to
452 * this duty, then the jiffies update is still serialized by
453 * xtime_lock.
454 */
455 if (unlikely(tick_do_timer_cpu == -1))
456 tick_do_timer_cpu = cpu;
457
458 /* Check, if the jiffies need an update */
459 if (tick_do_timer_cpu == cpu)
460 tick_do_update_jiffies64(now);
461
462 /*
463 * When we are idle and the tick is stopped, we have to touch
464 * the watchdog as we might not schedule for a really long
465 * time. This happens on complete idle SMP systems while
466 * waiting on the login prompt. We also increment the "start
467 * of idle" jiffy stamp so the idle accounting adjustment we
468 * do when we go busy again does not account too much ticks.
469 */
470 if (ts->tick_stopped) {
471 touch_softlockup_watchdog();
472 ts->idle_jiffies++;
473 }
474
475 update_process_times(user_mode(regs));
476 profile_tick(CPU_PROFILING);
477
478 /* Do not restart, when we are in the idle loop */
479 if (ts->tick_stopped)
480 return;
481
482 while (tick_nohz_reprogram(ts, now)) {
483 now = ktime_get();
484 tick_do_update_jiffies64(now);
485 }
486 }
487
488 /**
489 * tick_nohz_switch_to_nohz - switch to nohz mode
490 */
491 static void tick_nohz_switch_to_nohz(void)
492 {
493 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
494 ktime_t next;
495
496 if (!tick_nohz_enabled)
497 return;
498
499 local_irq_disable();
500 if (tick_switch_to_oneshot(tick_nohz_handler)) {
501 local_irq_enable();
502 return;
503 }
504
505 ts->nohz_mode = NOHZ_MODE_LOWRES;
506
507 /*
508 * Recycle the hrtimer in ts, so we can share the
509 * hrtimer_forward with the highres code.
510 */
511 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
512 /* Get the next period */
513 next = tick_init_jiffy_update();
514
515 for (;;) {
516 ts->sched_timer.expires = next;
517 if (!tick_program_event(next, 0))
518 break;
519 next = ktime_add(next, tick_period);
520 }
521 local_irq_enable();
522
523 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n",
524 smp_processor_id());
525 }
526
527 #else
528
529 static inline void tick_nohz_switch_to_nohz(void) { }
530
531 #endif /* NO_HZ */
532
533 /*
534 * High resolution timer specific code
535 */
536 #ifdef CONFIG_HIGH_RES_TIMERS
537 /*
538 * We rearm the timer until we get disabled by the idle code.
539 * Called with interrupts disabled and timer->base->cpu_base->lock held.
540 */
541 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
542 {
543 struct tick_sched *ts =
544 container_of(timer, struct tick_sched, sched_timer);
545 struct pt_regs *regs = get_irq_regs();
546 ktime_t now = ktime_get();
547 int cpu = smp_processor_id();
548
549 #ifdef CONFIG_NO_HZ
550 /*
551 * Check if the do_timer duty was dropped. We don't care about
552 * concurrency: This happens only when the cpu in charge went
553 * into a long sleep. If two cpus happen to assign themself to
554 * this duty, then the jiffies update is still serialized by
555 * xtime_lock.
556 */
557 if (unlikely(tick_do_timer_cpu == -1))
558 tick_do_timer_cpu = cpu;
559 #endif
560
561 /* Check, if the jiffies need an update */
562 if (tick_do_timer_cpu == cpu)
563 tick_do_update_jiffies64(now);
564
565 /*
566 * Do not call, when we are not in irq context and have
567 * no valid regs pointer
568 */
569 if (regs) {
570 /*
571 * When we are idle and the tick is stopped, we have to touch
572 * the watchdog as we might not schedule for a really long
573 * time. This happens on complete idle SMP systems while
574 * waiting on the login prompt. We also increment the "start of
575 * idle" jiffy stamp so the idle accounting adjustment we do
576 * when we go busy again does not account too much ticks.
577 */
578 if (ts->tick_stopped) {
579 touch_softlockup_watchdog();
580 ts->idle_jiffies++;
581 }
582 update_process_times(user_mode(regs));
583 profile_tick(CPU_PROFILING);
584 }
585
586 /* Do not restart, when we are in the idle loop */
587 if (ts->tick_stopped)
588 return HRTIMER_NORESTART;
589
590 hrtimer_forward(timer, now, tick_period);
591
592 return HRTIMER_RESTART;
593 }
594
595 /**
596 * tick_setup_sched_timer - setup the tick emulation timer
597 */
598 void tick_setup_sched_timer(void)
599 {
600 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
601 ktime_t now = ktime_get();
602 u64 offset;
603
604 /*
605 * Emulate tick processing via per-CPU hrtimers:
606 */
607 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
608 ts->sched_timer.function = tick_sched_timer;
609 ts->sched_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
610
611 /* Get the next period (per cpu) */
612 ts->sched_timer.expires = tick_init_jiffy_update();
613 offset = ktime_to_ns(tick_period) >> 1;
614 do_div(offset, num_possible_cpus());
615 offset *= smp_processor_id();
616 ts->sched_timer.expires = ktime_add_ns(ts->sched_timer.expires, offset);
617
618 for (;;) {
619 hrtimer_forward(&ts->sched_timer, now, tick_period);
620 hrtimer_start(&ts->sched_timer, ts->sched_timer.expires,
621 HRTIMER_MODE_ABS);
622 /* Check, if the timer was already in the past */
623 if (hrtimer_active(&ts->sched_timer))
624 break;
625 now = ktime_get();
626 }
627
628 #ifdef CONFIG_NO_HZ
629 if (tick_nohz_enabled)
630 ts->nohz_mode = NOHZ_MODE_HIGHRES;
631 #endif
632 }
633
634 void tick_cancel_sched_timer(int cpu)
635 {
636 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
637
638 if (ts->sched_timer.base)
639 hrtimer_cancel(&ts->sched_timer);
640 ts->tick_stopped = 0;
641 ts->nohz_mode = NOHZ_MODE_INACTIVE;
642 }
643 #endif /* HIGH_RES_TIMERS */
644
645 /**
646 * Async notification about clocksource changes
647 */
648 void tick_clock_notify(void)
649 {
650 int cpu;
651
652 for_each_possible_cpu(cpu)
653 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
654 }
655
656 /*
657 * Async notification about clock event changes
658 */
659 void tick_oneshot_notify(void)
660 {
661 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
662
663 set_bit(0, &ts->check_clocks);
664 }
665
666 /**
667 * Check, if a change happened, which makes oneshot possible.
668 *
669 * Called cyclic from the hrtimer softirq (driven by the timer
670 * softirq) allow_nohz signals, that we can switch into low-res nohz
671 * mode, because high resolution timers are disabled (either compile
672 * or runtime).
673 */
674 int tick_check_oneshot_change(int allow_nohz)
675 {
676 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
677
678 if (!test_and_clear_bit(0, &ts->check_clocks))
679 return 0;
680
681 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
682 return 0;
683
684 if (!timekeeping_is_continuous() || !tick_is_oneshot_available())
685 return 0;
686
687 if (!allow_nohz)
688 return 1;
689
690 tick_nohz_switch_to_nohz();
691 return 0;
692 }