]>
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 | ||
9e203bcc DM |
24 | #include <asm/irq_regs.h> |
25 | ||
79bf2bb3 TG |
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 | ||
289f480a IM |
38 | struct tick_sched *tick_get_tick_sched(int cpu) |
39 | { | |
40 | return &per_cpu(tick_cpu_sched, cpu); | |
41 | } | |
42 | ||
79bf2bb3 TG |
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 | cpu_clear(cpu, nohz_cpu_mask); | |
137 | now = ktime_get(); | |
138 | ||
139 | local_irq_save(flags); | |
140 | tick_do_update_jiffies64(now); | |
141 | local_irq_restore(flags); | |
142 | } | |
143 | ||
144 | /** | |
145 | * tick_nohz_stop_sched_tick - stop the idle tick from the idle task | |
146 | * | |
147 | * When the next event is more than a tick into the future, stop the idle tick | |
148 | * Called either from the idle loop or from irq_exit() when an idle period was | |
149 | * just interrupted by an interrupt which did not cause a reschedule. | |
150 | */ | |
151 | void tick_nohz_stop_sched_tick(void) | |
152 | { | |
153 | unsigned long seq, last_jiffies, next_jiffies, delta_jiffies, flags; | |
154 | struct tick_sched *ts; | |
155 | ktime_t last_update, expires, now, delta; | |
156 | int cpu; | |
157 | ||
158 | local_irq_save(flags); | |
159 | ||
160 | cpu = smp_processor_id(); | |
161 | ts = &per_cpu(tick_cpu_sched, cpu); | |
162 | ||
163 | if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) | |
164 | goto end; | |
165 | ||
166 | if (need_resched()) | |
167 | goto end; | |
168 | ||
169 | cpu = smp_processor_id(); | |
bc5393a6 TG |
170 | if (unlikely(local_softirq_pending())) |
171 | printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n", | |
172 | local_softirq_pending()); | |
79bf2bb3 TG |
173 | |
174 | now = ktime_get(); | |
175 | /* | |
176 | * When called from irq_exit we need to account the idle sleep time | |
177 | * correctly. | |
178 | */ | |
179 | if (ts->tick_stopped) { | |
180 | delta = ktime_sub(now, ts->idle_entrytime); | |
181 | ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); | |
182 | } | |
183 | ||
184 | ts->idle_entrytime = now; | |
185 | ts->idle_calls++; | |
186 | ||
187 | /* Read jiffies and the time when jiffies were updated last */ | |
188 | do { | |
189 | seq = read_seqbegin(&xtime_lock); | |
190 | last_update = last_jiffies_update; | |
191 | last_jiffies = jiffies; | |
192 | } while (read_seqretry(&xtime_lock, seq)); | |
193 | ||
194 | /* Get the next timer wheel timer */ | |
195 | next_jiffies = get_next_timer_interrupt(last_jiffies); | |
196 | delta_jiffies = next_jiffies - last_jiffies; | |
197 | ||
6ba9b346 IM |
198 | if (rcu_needs_cpu(cpu)) |
199 | delta_jiffies = 1; | |
79bf2bb3 TG |
200 | /* |
201 | * Do not stop the tick, if we are only one off | |
202 | * or if the cpu is required for rcu | |
203 | */ | |
6ba9b346 | 204 | if (!ts->tick_stopped && delta_jiffies == 1) |
79bf2bb3 TG |
205 | goto out; |
206 | ||
207 | /* Schedule the tick, if we are at least one jiffie off */ | |
208 | if ((long)delta_jiffies >= 1) { | |
209 | ||
6ba9b346 | 210 | if (delta_jiffies > 1) |
79bf2bb3 TG |
211 | cpu_set(cpu, nohz_cpu_mask); |
212 | /* | |
213 | * nohz_stop_sched_tick can be called several times before | |
214 | * the nohz_restart_sched_tick is called. This happens when | |
215 | * interrupts arrive which do not cause a reschedule. In the | |
216 | * first call we save the current tick time, so we can restart | |
217 | * the scheduler tick in nohz_restart_sched_tick. | |
218 | */ | |
219 | if (!ts->tick_stopped) { | |
220 | ts->idle_tick = ts->sched_timer.expires; | |
221 | ts->tick_stopped = 1; | |
222 | ts->idle_jiffies = last_jiffies; | |
223 | } | |
224 | /* | |
225 | * calculate the expiry time for the next timer wheel | |
226 | * timer | |
227 | */ | |
228 | expires = ktime_add_ns(last_update, tick_period.tv64 * | |
229 | delta_jiffies); | |
230 | ts->idle_expires = expires; | |
231 | ts->idle_sleeps++; | |
232 | ||
233 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { | |
234 | hrtimer_start(&ts->sched_timer, expires, | |
235 | HRTIMER_MODE_ABS); | |
236 | /* Check, if the timer was already in the past */ | |
237 | if (hrtimer_active(&ts->sched_timer)) | |
238 | goto out; | |
239 | } else if(!tick_program_event(expires, 0)) | |
240 | goto out; | |
241 | /* | |
242 | * We are past the event already. So we crossed a | |
243 | * jiffie boundary. Update jiffies and raise the | |
244 | * softirq. | |
245 | */ | |
246 | tick_do_update_jiffies64(ktime_get()); | |
247 | cpu_clear(cpu, nohz_cpu_mask); | |
248 | } | |
249 | raise_softirq_irqoff(TIMER_SOFTIRQ); | |
250 | out: | |
251 | ts->next_jiffies = next_jiffies; | |
252 | ts->last_jiffies = last_jiffies; | |
253 | end: | |
254 | local_irq_restore(flags); | |
255 | } | |
256 | ||
257 | /** | |
258 | * nohz_restart_sched_tick - restart the idle tick from the idle task | |
259 | * | |
260 | * Restart the idle tick when the CPU is woken up from idle | |
261 | */ | |
262 | void tick_nohz_restart_sched_tick(void) | |
263 | { | |
264 | int cpu = smp_processor_id(); | |
265 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
266 | unsigned long ticks; | |
267 | ktime_t now, delta; | |
268 | ||
269 | if (!ts->tick_stopped) | |
270 | return; | |
271 | ||
272 | /* Update jiffies first */ | |
273 | now = ktime_get(); | |
274 | ||
275 | local_irq_disable(); | |
276 | tick_do_update_jiffies64(now); | |
277 | cpu_clear(cpu, nohz_cpu_mask); | |
278 | ||
279 | /* Account the idle time */ | |
280 | delta = ktime_sub(now, ts->idle_entrytime); | |
281 | ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); | |
282 | ||
283 | /* | |
284 | * We stopped the tick in idle. Update process times would miss the | |
285 | * time we slept as update_process_times does only a 1 tick | |
286 | * accounting. Enforce that this is accounted to idle ! | |
287 | */ | |
288 | ticks = jiffies - ts->idle_jiffies; | |
289 | /* | |
290 | * We might be one off. Do not randomly account a huge number of ticks! | |
291 | */ | |
292 | if (ticks && ticks < LONG_MAX) { | |
293 | add_preempt_count(HARDIRQ_OFFSET); | |
294 | account_system_time(current, HARDIRQ_OFFSET, | |
295 | jiffies_to_cputime(ticks)); | |
296 | sub_preempt_count(HARDIRQ_OFFSET); | |
297 | } | |
298 | ||
299 | /* | |
300 | * Cancel the scheduled timer and restore the tick | |
301 | */ | |
302 | ts->tick_stopped = 0; | |
303 | hrtimer_cancel(&ts->sched_timer); | |
304 | ts->sched_timer.expires = ts->idle_tick; | |
305 | ||
306 | while (1) { | |
307 | /* Forward the time to expire in the future */ | |
308 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
309 | ||
310 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { | |
311 | hrtimer_start(&ts->sched_timer, | |
312 | ts->sched_timer.expires, | |
313 | HRTIMER_MODE_ABS); | |
314 | /* Check, if the timer was already in the past */ | |
315 | if (hrtimer_active(&ts->sched_timer)) | |
316 | break; | |
317 | } else { | |
318 | if (!tick_program_event(ts->sched_timer.expires, 0)) | |
319 | break; | |
320 | } | |
321 | /* Update jiffies and reread time */ | |
322 | tick_do_update_jiffies64(now); | |
323 | now = ktime_get(); | |
324 | } | |
325 | local_irq_enable(); | |
326 | } | |
327 | ||
328 | static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now) | |
329 | { | |
330 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
331 | return tick_program_event(ts->sched_timer.expires, 0); | |
332 | } | |
333 | ||
334 | /* | |
335 | * The nohz low res interrupt handler | |
336 | */ | |
337 | static void tick_nohz_handler(struct clock_event_device *dev) | |
338 | { | |
339 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
340 | struct pt_regs *regs = get_irq_regs(); | |
341 | ktime_t now = ktime_get(); | |
342 | ||
343 | dev->next_event.tv64 = KTIME_MAX; | |
344 | ||
345 | /* Check, if the jiffies need an update */ | |
346 | tick_do_update_jiffies64(now); | |
347 | ||
348 | /* | |
349 | * When we are idle and the tick is stopped, we have to touch | |
350 | * the watchdog as we might not schedule for a really long | |
351 | * time. This happens on complete idle SMP systems while | |
352 | * waiting on the login prompt. We also increment the "start | |
353 | * of idle" jiffy stamp so the idle accounting adjustment we | |
354 | * do when we go busy again does not account too much ticks. | |
355 | */ | |
356 | if (ts->tick_stopped) { | |
357 | touch_softlockup_watchdog(); | |
358 | ts->idle_jiffies++; | |
359 | } | |
360 | ||
361 | update_process_times(user_mode(regs)); | |
362 | profile_tick(CPU_PROFILING); | |
363 | ||
364 | /* Do not restart, when we are in the idle loop */ | |
365 | if (ts->tick_stopped) | |
366 | return; | |
367 | ||
368 | while (tick_nohz_reprogram(ts, now)) { | |
369 | now = ktime_get(); | |
370 | tick_do_update_jiffies64(now); | |
371 | } | |
372 | } | |
373 | ||
374 | /** | |
375 | * tick_nohz_switch_to_nohz - switch to nohz mode | |
376 | */ | |
377 | static void tick_nohz_switch_to_nohz(void) | |
378 | { | |
379 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
380 | ktime_t next; | |
381 | ||
382 | if (!tick_nohz_enabled) | |
383 | return; | |
384 | ||
385 | local_irq_disable(); | |
386 | if (tick_switch_to_oneshot(tick_nohz_handler)) { | |
387 | local_irq_enable(); | |
388 | return; | |
389 | } | |
390 | ||
391 | ts->nohz_mode = NOHZ_MODE_LOWRES; | |
392 | ||
393 | /* | |
394 | * Recycle the hrtimer in ts, so we can share the | |
395 | * hrtimer_forward with the highres code. | |
396 | */ | |
397 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); | |
398 | /* Get the next period */ | |
399 | next = tick_init_jiffy_update(); | |
400 | ||
401 | for (;;) { | |
402 | ts->sched_timer.expires = next; | |
403 | if (!tick_program_event(next, 0)) | |
404 | break; | |
405 | next = ktime_add(next, tick_period); | |
406 | } | |
407 | local_irq_enable(); | |
408 | ||
409 | printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n", | |
410 | smp_processor_id()); | |
411 | } | |
412 | ||
413 | #else | |
414 | ||
415 | static inline void tick_nohz_switch_to_nohz(void) { } | |
416 | ||
417 | #endif /* NO_HZ */ | |
418 | ||
419 | /* | |
420 | * High resolution timer specific code | |
421 | */ | |
422 | #ifdef CONFIG_HIGH_RES_TIMERS | |
423 | /* | |
424 | * We rearm the timer until we get disabled by the idle code | |
425 | * Called with interrupts disabled and timer->base->cpu_base->lock held. | |
426 | */ | |
427 | static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) | |
428 | { | |
429 | struct tick_sched *ts = | |
430 | container_of(timer, struct tick_sched, sched_timer); | |
431 | struct hrtimer_cpu_base *base = timer->base->cpu_base; | |
432 | struct pt_regs *regs = get_irq_regs(); | |
433 | ktime_t now = ktime_get(); | |
434 | ||
435 | /* Check, if the jiffies need an update */ | |
436 | tick_do_update_jiffies64(now); | |
437 | ||
438 | /* | |
439 | * Do not call, when we are not in irq context and have | |
440 | * no valid regs pointer | |
441 | */ | |
442 | if (regs) { | |
443 | /* | |
444 | * When we are idle and the tick is stopped, we have to touch | |
445 | * the watchdog as we might not schedule for a really long | |
446 | * time. This happens on complete idle SMP systems while | |
447 | * waiting on the login prompt. We also increment the "start of | |
448 | * idle" jiffy stamp so the idle accounting adjustment we do | |
449 | * when we go busy again does not account too much ticks. | |
450 | */ | |
451 | if (ts->tick_stopped) { | |
452 | touch_softlockup_watchdog(); | |
453 | ts->idle_jiffies++; | |
454 | } | |
455 | /* | |
456 | * update_process_times() might take tasklist_lock, hence | |
457 | * drop the base lock. sched-tick hrtimers are per-CPU and | |
458 | * never accessible by userspace APIs, so this is safe to do. | |
459 | */ | |
460 | spin_unlock(&base->lock); | |
461 | update_process_times(user_mode(regs)); | |
462 | profile_tick(CPU_PROFILING); | |
463 | spin_lock(&base->lock); | |
464 | } | |
465 | ||
466 | /* Do not restart, when we are in the idle loop */ | |
467 | if (ts->tick_stopped) | |
468 | return HRTIMER_NORESTART; | |
469 | ||
470 | hrtimer_forward(timer, now, tick_period); | |
471 | ||
472 | return HRTIMER_RESTART; | |
473 | } | |
474 | ||
475 | /** | |
476 | * tick_setup_sched_timer - setup the tick emulation timer | |
477 | */ | |
478 | void tick_setup_sched_timer(void) | |
479 | { | |
480 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
481 | ktime_t now = ktime_get(); | |
482 | ||
483 | /* | |
484 | * Emulate tick processing via per-CPU hrtimers: | |
485 | */ | |
486 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); | |
487 | ts->sched_timer.function = tick_sched_timer; | |
488 | ts->sched_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ; | |
489 | ||
490 | /* Get the next period */ | |
491 | ts->sched_timer.expires = tick_init_jiffy_update(); | |
492 | ||
493 | for (;;) { | |
494 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
495 | hrtimer_start(&ts->sched_timer, ts->sched_timer.expires, | |
496 | HRTIMER_MODE_ABS); | |
497 | /* Check, if the timer was already in the past */ | |
498 | if (hrtimer_active(&ts->sched_timer)) | |
499 | break; | |
500 | now = ktime_get(); | |
501 | } | |
502 | ||
503 | #ifdef CONFIG_NO_HZ | |
504 | if (tick_nohz_enabled) | |
505 | ts->nohz_mode = NOHZ_MODE_HIGHRES; | |
506 | #endif | |
507 | } | |
508 | ||
509 | void tick_cancel_sched_timer(int cpu) | |
510 | { | |
511 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
512 | ||
513 | if (ts->sched_timer.base) | |
514 | hrtimer_cancel(&ts->sched_timer); | |
515 | ts->tick_stopped = 0; | |
516 | ts->nohz_mode = NOHZ_MODE_INACTIVE; | |
517 | } | |
518 | #endif /* HIGH_RES_TIMERS */ | |
519 | ||
520 | /** | |
521 | * Async notification about clocksource changes | |
522 | */ | |
523 | void tick_clock_notify(void) | |
524 | { | |
525 | int cpu; | |
526 | ||
527 | for_each_possible_cpu(cpu) | |
528 | set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); | |
529 | } | |
530 | ||
531 | /* | |
532 | * Async notification about clock event changes | |
533 | */ | |
534 | void tick_oneshot_notify(void) | |
535 | { | |
536 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
537 | ||
538 | set_bit(0, &ts->check_clocks); | |
539 | } | |
540 | ||
541 | /** | |
542 | * Check, if a change happened, which makes oneshot possible. | |
543 | * | |
544 | * Called cyclic from the hrtimer softirq (driven by the timer | |
545 | * softirq) allow_nohz signals, that we can switch into low-res nohz | |
546 | * mode, because high resolution timers are disabled (either compile | |
547 | * or runtime). | |
548 | */ | |
549 | int tick_check_oneshot_change(int allow_nohz) | |
550 | { | |
551 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
552 | ||
553 | if (!test_and_clear_bit(0, &ts->check_clocks)) | |
554 | return 0; | |
555 | ||
556 | if (ts->nohz_mode != NOHZ_MODE_INACTIVE) | |
557 | return 0; | |
558 | ||
559 | if (!timekeeping_is_continuous() || !tick_is_oneshot_available()) | |
560 | return 0; | |
561 | ||
562 | if (!allow_nohz) | |
563 | return 1; | |
564 | ||
565 | tick_nohz_switch_to_nohz(); | |
566 | return 0; | |
567 | } |