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 | * | |
b10db7f0 | 12 | * Distribute under GPLv2. |
79bf2bb3 TG |
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> | |
8083e4ad | 22 | #include <linux/module.h> |
79bf2bb3 | 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 | ||
7a14ce1d IM |
51 | /* |
52 | * Do a quick check without holding xtime_lock: | |
53 | */ | |
54 | delta = ktime_sub(now, last_jiffies_update); | |
55 | if (delta.tv64 < tick_period.tv64) | |
56 | return; | |
57 | ||
79bf2bb3 TG |
58 | /* Reevalute with xtime_lock held */ |
59 | write_seqlock(&xtime_lock); | |
60 | ||
61 | delta = ktime_sub(now, last_jiffies_update); | |
62 | if (delta.tv64 >= tick_period.tv64) { | |
63 | ||
64 | delta = ktime_sub(delta, tick_period); | |
65 | last_jiffies_update = ktime_add(last_jiffies_update, | |
66 | tick_period); | |
67 | ||
68 | /* Slow path for long timeouts */ | |
69 | if (unlikely(delta.tv64 >= tick_period.tv64)) { | |
70 | s64 incr = ktime_to_ns(tick_period); | |
71 | ||
72 | ticks = ktime_divns(delta, incr); | |
73 | ||
74 | last_jiffies_update = ktime_add_ns(last_jiffies_update, | |
75 | incr * ticks); | |
76 | } | |
77 | do_timer(++ticks); | |
49d670fb TG |
78 | |
79 | /* Keep the tick_next_period variable up to date */ | |
80 | tick_next_period = ktime_add(last_jiffies_update, tick_period); | |
79bf2bb3 TG |
81 | } |
82 | write_sequnlock(&xtime_lock); | |
83 | } | |
84 | ||
85 | /* | |
86 | * Initialize and return retrieve the jiffies update. | |
87 | */ | |
88 | static ktime_t tick_init_jiffy_update(void) | |
89 | { | |
90 | ktime_t period; | |
91 | ||
92 | write_seqlock(&xtime_lock); | |
93 | /* Did we start the jiffies update yet ? */ | |
94 | if (last_jiffies_update.tv64 == 0) | |
95 | last_jiffies_update = tick_next_period; | |
96 | period = last_jiffies_update; | |
97 | write_sequnlock(&xtime_lock); | |
98 | return period; | |
99 | } | |
100 | ||
101 | /* | |
102 | * NOHZ - aka dynamic tick functionality | |
103 | */ | |
104 | #ifdef CONFIG_NO_HZ | |
105 | /* | |
106 | * NO HZ enabled ? | |
107 | */ | |
108 | static int tick_nohz_enabled __read_mostly = 1; | |
109 | ||
110 | /* | |
111 | * Enable / Disable tickless mode | |
112 | */ | |
113 | static int __init setup_tick_nohz(char *str) | |
114 | { | |
115 | if (!strcmp(str, "off")) | |
116 | tick_nohz_enabled = 0; | |
117 | else if (!strcmp(str, "on")) | |
118 | tick_nohz_enabled = 1; | |
119 | else | |
120 | return 0; | |
121 | return 1; | |
122 | } | |
123 | ||
124 | __setup("nohz=", setup_tick_nohz); | |
125 | ||
126 | /** | |
127 | * tick_nohz_update_jiffies - update jiffies when idle was interrupted | |
128 | * | |
129 | * Called from interrupt entry when the CPU was idle | |
130 | * | |
131 | * In case the sched_tick was stopped on this CPU, we have to check if jiffies | |
132 | * must be updated. Otherwise an interrupt handler could use a stale jiffy | |
133 | * value. We do this unconditionally on any cpu, as we don't know whether the | |
134 | * cpu, which has the update task assigned is in a long sleep. | |
135 | */ | |
eed3b9cf | 136 | static void tick_nohz_update_jiffies(ktime_t now) |
79bf2bb3 TG |
137 | { |
138 | int cpu = smp_processor_id(); | |
139 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
140 | unsigned long flags; | |
79bf2bb3 | 141 | |
5df7fa1c | 142 | ts->idle_waketime = now; |
79bf2bb3 TG |
143 | |
144 | local_irq_save(flags); | |
145 | tick_do_update_jiffies64(now); | |
146 | local_irq_restore(flags); | |
02ff3755 IM |
147 | |
148 | touch_softlockup_watchdog(); | |
79bf2bb3 TG |
149 | } |
150 | ||
595aac48 AV |
151 | /* |
152 | * Updates the per cpu time idle statistics counters | |
153 | */ | |
8d63bf94 | 154 | static void |
8c215bd3 | 155 | update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time) |
6378ddb5 | 156 | { |
eed3b9cf | 157 | ktime_t delta; |
6378ddb5 | 158 | |
595aac48 AV |
159 | if (ts->idle_active) { |
160 | delta = ktime_sub(now, ts->idle_entrytime); | |
8c215bd3 | 161 | if (nr_iowait_cpu(cpu) > 0) |
0224cf4c | 162 | ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta); |
6beea0cd MH |
163 | else |
164 | ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); | |
8c7b09f4 | 165 | ts->idle_entrytime = now; |
595aac48 | 166 | } |
8d63bf94 | 167 | |
e0e37c20 | 168 | if (last_update_time) |
8d63bf94 AV |
169 | *last_update_time = ktime_to_us(now); |
170 | ||
595aac48 AV |
171 | } |
172 | ||
173 | static void tick_nohz_stop_idle(int cpu, ktime_t now) | |
174 | { | |
175 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
176 | ||
8c215bd3 | 177 | update_ts_time_stats(cpu, ts, now, NULL); |
eed3b9cf | 178 | ts->idle_active = 0; |
56c7426b | 179 | |
eed3b9cf | 180 | sched_clock_idle_wakeup_event(0); |
6378ddb5 VP |
181 | } |
182 | ||
8c215bd3 | 183 | static ktime_t tick_nohz_start_idle(int cpu, struct tick_sched *ts) |
6378ddb5 | 184 | { |
595aac48 | 185 | ktime_t now; |
6378ddb5 VP |
186 | |
187 | now = ktime_get(); | |
595aac48 | 188 | |
8c215bd3 | 189 | update_ts_time_stats(cpu, ts, now, NULL); |
595aac48 | 190 | |
6378ddb5 VP |
191 | ts->idle_entrytime = now; |
192 | ts->idle_active = 1; | |
56c7426b | 193 | sched_clock_idle_sleep_event(); |
6378ddb5 VP |
194 | return now; |
195 | } | |
196 | ||
b1f724c3 AV |
197 | /** |
198 | * get_cpu_idle_time_us - get the total idle time of a cpu | |
199 | * @cpu: CPU number to query | |
09a1d34f MH |
200 | * @last_update_time: variable to store update time in. Do not update |
201 | * counters if NULL. | |
b1f724c3 AV |
202 | * |
203 | * Return the cummulative idle time (since boot) for a given | |
6beea0cd | 204 | * CPU, in microseconds. |
b1f724c3 AV |
205 | * |
206 | * This time is measured via accounting rather than sampling, | |
207 | * and is as accurate as ktime_get() is. | |
208 | * | |
209 | * This function returns -1 if NOHZ is not enabled. | |
210 | */ | |
6378ddb5 VP |
211 | u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) |
212 | { | |
213 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
09a1d34f | 214 | ktime_t now, idle; |
6378ddb5 | 215 | |
8083e4ad | 216 | if (!tick_nohz_enabled) |
217 | return -1; | |
218 | ||
09a1d34f MH |
219 | now = ktime_get(); |
220 | if (last_update_time) { | |
221 | update_ts_time_stats(cpu, ts, now, last_update_time); | |
222 | idle = ts->idle_sleeptime; | |
223 | } else { | |
224 | if (ts->idle_active && !nr_iowait_cpu(cpu)) { | |
225 | ktime_t delta = ktime_sub(now, ts->idle_entrytime); | |
226 | ||
227 | idle = ktime_add(ts->idle_sleeptime, delta); | |
228 | } else { | |
229 | idle = ts->idle_sleeptime; | |
230 | } | |
231 | } | |
232 | ||
233 | return ktime_to_us(idle); | |
8083e4ad | 234 | |
6378ddb5 | 235 | } |
8083e4ad | 236 | EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); |
6378ddb5 | 237 | |
6beea0cd | 238 | /** |
0224cf4c AV |
239 | * get_cpu_iowait_time_us - get the total iowait time of a cpu |
240 | * @cpu: CPU number to query | |
09a1d34f MH |
241 | * @last_update_time: variable to store update time in. Do not update |
242 | * counters if NULL. | |
0224cf4c AV |
243 | * |
244 | * Return the cummulative iowait time (since boot) for a given | |
245 | * CPU, in microseconds. | |
246 | * | |
247 | * This time is measured via accounting rather than sampling, | |
248 | * and is as accurate as ktime_get() is. | |
249 | * | |
250 | * This function returns -1 if NOHZ is not enabled. | |
251 | */ | |
252 | u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time) | |
253 | { | |
254 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
09a1d34f | 255 | ktime_t now, iowait; |
0224cf4c AV |
256 | |
257 | if (!tick_nohz_enabled) | |
258 | return -1; | |
259 | ||
09a1d34f MH |
260 | now = ktime_get(); |
261 | if (last_update_time) { | |
262 | update_ts_time_stats(cpu, ts, now, last_update_time); | |
263 | iowait = ts->iowait_sleeptime; | |
264 | } else { | |
265 | if (ts->idle_active && nr_iowait_cpu(cpu) > 0) { | |
266 | ktime_t delta = ktime_sub(now, ts->idle_entrytime); | |
0224cf4c | 267 | |
09a1d34f MH |
268 | iowait = ktime_add(ts->iowait_sleeptime, delta); |
269 | } else { | |
270 | iowait = ts->iowait_sleeptime; | |
271 | } | |
272 | } | |
0224cf4c | 273 | |
09a1d34f | 274 | return ktime_to_us(iowait); |
0224cf4c AV |
275 | } |
276 | EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us); | |
277 | ||
280f0677 | 278 | static void tick_nohz_stop_sched_tick(struct tick_sched *ts) |
79bf2bb3 | 279 | { |
280f0677 | 280 | unsigned long seq, last_jiffies, next_jiffies, delta_jiffies; |
6378ddb5 | 281 | ktime_t last_update, expires, now; |
4f86d3a8 | 282 | struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev; |
98962465 | 283 | u64 time_delta; |
79bf2bb3 TG |
284 | int cpu; |
285 | ||
79bf2bb3 TG |
286 | cpu = smp_processor_id(); |
287 | ts = &per_cpu(tick_cpu_sched, cpu); | |
f2e21c96 | 288 | |
8c215bd3 | 289 | now = tick_nohz_start_idle(cpu, ts); |
79bf2bb3 | 290 | |
5e41d0d6 TG |
291 | /* |
292 | * If this cpu is offline and it is the one which updates | |
293 | * jiffies, then give up the assignment and let it be taken by | |
294 | * the cpu which runs the tick timer next. If we don't drop | |
295 | * this here the jiffies might be stale and do_timer() never | |
296 | * invoked. | |
297 | */ | |
298 | if (unlikely(!cpu_online(cpu))) { | |
299 | if (cpu == tick_do_timer_cpu) | |
6441402b | 300 | tick_do_timer_cpu = TICK_DO_TIMER_NONE; |
5e41d0d6 TG |
301 | } |
302 | ||
79bf2bb3 | 303 | if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) |
280f0677 | 304 | return; |
79bf2bb3 TG |
305 | |
306 | if (need_resched()) | |
280f0677 | 307 | return; |
79bf2bb3 | 308 | |
fa116ea3 | 309 | if (unlikely(local_softirq_pending() && cpu_online(cpu))) { |
35282316 TG |
310 | static int ratelimit; |
311 | ||
312 | if (ratelimit < 10) { | |
313 | printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n", | |
529eaccd | 314 | (unsigned int) local_softirq_pending()); |
35282316 TG |
315 | ratelimit++; |
316 | } | |
280f0677 | 317 | return; |
35282316 | 318 | } |
79bf2bb3 | 319 | |
79bf2bb3 | 320 | ts->idle_calls++; |
79bf2bb3 TG |
321 | /* Read jiffies and the time when jiffies were updated last */ |
322 | do { | |
323 | seq = read_seqbegin(&xtime_lock); | |
324 | last_update = last_jiffies_update; | |
325 | last_jiffies = jiffies; | |
27185016 | 326 | time_delta = timekeeping_max_deferment(); |
79bf2bb3 TG |
327 | } while (read_seqretry(&xtime_lock, seq)); |
328 | ||
3c5d92a0 | 329 | if (rcu_needs_cpu(cpu) || printk_needs_cpu(cpu) || |
396e894d | 330 | arch_needs_cpu(cpu)) { |
3c5d92a0 | 331 | next_jiffies = last_jiffies + 1; |
6ba9b346 | 332 | delta_jiffies = 1; |
3c5d92a0 MS |
333 | } else { |
334 | /* Get the next timer wheel timer */ | |
335 | next_jiffies = get_next_timer_interrupt(last_jiffies); | |
336 | delta_jiffies = next_jiffies - last_jiffies; | |
337 | } | |
79bf2bb3 TG |
338 | /* |
339 | * Do not stop the tick, if we are only one off | |
340 | * or if the cpu is required for rcu | |
341 | */ | |
6ba9b346 | 342 | if (!ts->tick_stopped && delta_jiffies == 1) |
79bf2bb3 TG |
343 | goto out; |
344 | ||
345 | /* Schedule the tick, if we are at least one jiffie off */ | |
346 | if ((long)delta_jiffies >= 1) { | |
347 | ||
00147449 WR |
348 | /* |
349 | * If this cpu is the one which updates jiffies, then | |
350 | * give up the assignment and let it be taken by the | |
351 | * cpu which runs the tick timer next, which might be | |
352 | * this cpu as well. If we don't drop this here the | |
353 | * jiffies might be stale and do_timer() never | |
27185016 TG |
354 | * invoked. Keep track of the fact that it was the one |
355 | * which had the do_timer() duty last. If this cpu is | |
356 | * the one which had the do_timer() duty last, we | |
357 | * limit the sleep time to the timekeeping | |
358 | * max_deferement value which we retrieved | |
359 | * above. Otherwise we can sleep as long as we want. | |
00147449 | 360 | */ |
27185016 | 361 | if (cpu == tick_do_timer_cpu) { |
00147449 | 362 | tick_do_timer_cpu = TICK_DO_TIMER_NONE; |
27185016 TG |
363 | ts->do_timer_last = 1; |
364 | } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { | |
365 | time_delta = KTIME_MAX; | |
366 | ts->do_timer_last = 0; | |
367 | } else if (!ts->do_timer_last) { | |
368 | time_delta = KTIME_MAX; | |
369 | } | |
370 | ||
00147449 | 371 | /* |
98962465 JH |
372 | * calculate the expiry time for the next timer wheel |
373 | * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals | |
374 | * that there is no timer pending or at least extremely | |
375 | * far into the future (12 days for HZ=1000). In this | |
376 | * case we set the expiry to the end of time. | |
377 | */ | |
378 | if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) { | |
379 | /* | |
380 | * Calculate the time delta for the next timer event. | |
381 | * If the time delta exceeds the maximum time delta | |
382 | * permitted by the current clocksource then adjust | |
383 | * the time delta accordingly to ensure the | |
384 | * clocksource does not wrap. | |
385 | */ | |
386 | time_delta = min_t(u64, time_delta, | |
387 | tick_period.tv64 * delta_jiffies); | |
98962465 | 388 | } |
00147449 | 389 | |
27185016 TG |
390 | if (time_delta < KTIME_MAX) |
391 | expires = ktime_add_ns(last_update, time_delta); | |
392 | else | |
393 | expires.tv64 = KTIME_MAX; | |
00147449 | 394 | |
00147449 WR |
395 | /* Skip reprogram of event if its not changed */ |
396 | if (ts->tick_stopped && ktime_equal(expires, dev->next_event)) | |
397 | goto out; | |
398 | ||
79bf2bb3 TG |
399 | /* |
400 | * nohz_stop_sched_tick can be called several times before | |
401 | * the nohz_restart_sched_tick is called. This happens when | |
402 | * interrupts arrive which do not cause a reschedule. In the | |
403 | * first call we save the current tick time, so we can restart | |
404 | * the scheduler tick in nohz_restart_sched_tick. | |
405 | */ | |
406 | if (!ts->tick_stopped) { | |
83cd4fe2 | 407 | select_nohz_load_balancer(1); |
46cb4b7c | 408 | |
cc584b21 | 409 | ts->idle_tick = hrtimer_get_expires(&ts->sched_timer); |
79bf2bb3 TG |
410 | ts->tick_stopped = 1; |
411 | ts->idle_jiffies = last_jiffies; | |
412 | } | |
d3ed7824 | 413 | |
eaad084b TG |
414 | ts->idle_sleeps++; |
415 | ||
98962465 JH |
416 | /* Mark expires */ |
417 | ts->idle_expires = expires; | |
418 | ||
eaad084b | 419 | /* |
98962465 JH |
420 | * If the expiration time == KTIME_MAX, then |
421 | * in this case we simply stop the tick timer. | |
eaad084b | 422 | */ |
98962465 | 423 | if (unlikely(expires.tv64 == KTIME_MAX)) { |
eaad084b TG |
424 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) |
425 | hrtimer_cancel(&ts->sched_timer); | |
426 | goto out; | |
427 | } | |
428 | ||
79bf2bb3 TG |
429 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { |
430 | hrtimer_start(&ts->sched_timer, expires, | |
5c333864 | 431 | HRTIMER_MODE_ABS_PINNED); |
79bf2bb3 TG |
432 | /* Check, if the timer was already in the past */ |
433 | if (hrtimer_active(&ts->sched_timer)) | |
434 | goto out; | |
4c9dc641 | 435 | } else if (!tick_program_event(expires, 0)) |
79bf2bb3 TG |
436 | goto out; |
437 | /* | |
438 | * We are past the event already. So we crossed a | |
439 | * jiffie boundary. Update jiffies and raise the | |
440 | * softirq. | |
441 | */ | |
442 | tick_do_update_jiffies64(ktime_get()); | |
79bf2bb3 TG |
443 | } |
444 | raise_softirq_irqoff(TIMER_SOFTIRQ); | |
445 | out: | |
446 | ts->next_jiffies = next_jiffies; | |
447 | ts->last_jiffies = last_jiffies; | |
4f86d3a8 | 448 | ts->sleep_length = ktime_sub(dev->next_event, now); |
280f0677 FW |
449 | } |
450 | ||
451 | /** | |
452 | * tick_nohz_idle_enter - stop the idle tick from the idle task | |
453 | * | |
454 | * When the next event is more than a tick into the future, stop the idle tick | |
455 | * Called when we start the idle loop. | |
456 | * This also enters into RCU extended quiescent state so that this CPU doesn't | |
457 | * need anymore to be part of any global grace period completion. This way | |
458 | * the tick can be stopped safely as we don't need to report quiescent states. | |
459 | */ | |
460 | void tick_nohz_idle_enter(void) | |
461 | { | |
462 | struct tick_sched *ts; | |
463 | ||
464 | WARN_ON_ONCE(irqs_disabled()); | |
465 | ||
466 | local_irq_disable(); | |
467 | ||
468 | ts = &__get_cpu_var(tick_cpu_sched); | |
469 | /* | |
470 | * set ts->inidle unconditionally. even if the system did not | |
471 | * switch to nohz mode the cpu frequency governers rely on the | |
472 | * update of the idle time accounting in tick_nohz_start_idle(). | |
473 | */ | |
474 | ts->inidle = 1; | |
475 | tick_nohz_stop_sched_tick(ts); | |
476 | rcu_idle_enter(); | |
477 | ||
478 | local_irq_enable(); | |
479 | } | |
480 | ||
481 | /** | |
482 | * tick_nohz_irq_exit - update next tick event from interrupt exit | |
483 | * | |
484 | * When an interrupt fires while we are idle and it doesn't cause | |
485 | * a reschedule, it may still add, modify or delete a timer, enqueue | |
486 | * an RCU callback, etc... | |
487 | * So we need to re-calculate and reprogram the next tick event. | |
488 | */ | |
489 | void tick_nohz_irq_exit(void) | |
490 | { | |
491 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
492 | ||
493 | if (!ts->inidle) | |
494 | return; | |
495 | ||
496 | tick_nohz_stop_sched_tick(ts); | |
79bf2bb3 TG |
497 | } |
498 | ||
4f86d3a8 LB |
499 | /** |
500 | * tick_nohz_get_sleep_length - return the length of the current sleep | |
501 | * | |
502 | * Called from power state control code with interrupts disabled | |
503 | */ | |
504 | ktime_t tick_nohz_get_sleep_length(void) | |
505 | { | |
506 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
507 | ||
508 | return ts->sleep_length; | |
509 | } | |
510 | ||
c34bec5a TG |
511 | static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) |
512 | { | |
513 | hrtimer_cancel(&ts->sched_timer); | |
268a3dcf | 514 | hrtimer_set_expires(&ts->sched_timer, ts->idle_tick); |
c34bec5a TG |
515 | |
516 | while (1) { | |
517 | /* Forward the time to expire in the future */ | |
518 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
519 | ||
520 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { | |
268a3dcf | 521 | hrtimer_start_expires(&ts->sched_timer, |
5c333864 | 522 | HRTIMER_MODE_ABS_PINNED); |
c34bec5a TG |
523 | /* Check, if the timer was already in the past */ |
524 | if (hrtimer_active(&ts->sched_timer)) | |
525 | break; | |
526 | } else { | |
268a3dcf TG |
527 | if (!tick_program_event( |
528 | hrtimer_get_expires(&ts->sched_timer), 0)) | |
c34bec5a TG |
529 | break; |
530 | } | |
531 | /* Update jiffies and reread time */ | |
532 | tick_do_update_jiffies64(now); | |
533 | now = ktime_get(); | |
534 | } | |
535 | } | |
536 | ||
79bf2bb3 | 537 | /** |
280f0677 | 538 | * tick_nohz_idle_exit - restart the idle tick from the idle task |
79bf2bb3 TG |
539 | * |
540 | * Restart the idle tick when the CPU is woken up from idle | |
280f0677 FW |
541 | * This also exit the RCU extended quiescent state. The CPU |
542 | * can use RCU again after this function is called. | |
79bf2bb3 | 543 | */ |
280f0677 | 544 | void tick_nohz_idle_exit(void) |
79bf2bb3 TG |
545 | { |
546 | int cpu = smp_processor_id(); | |
547 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
79741dd3 | 548 | #ifndef CONFIG_VIRT_CPU_ACCOUNTING |
79bf2bb3 | 549 | unsigned long ticks; |
79741dd3 | 550 | #endif |
6378ddb5 | 551 | ktime_t now; |
79bf2bb3 | 552 | |
6378ddb5 | 553 | local_irq_disable(); |
9b2e4f18 | 554 | rcu_idle_exit(); |
eed3b9cf MS |
555 | if (ts->idle_active || (ts->inidle && ts->tick_stopped)) |
556 | now = ktime_get(); | |
557 | ||
558 | if (ts->idle_active) | |
559 | tick_nohz_stop_idle(cpu, now); | |
6378ddb5 | 560 | |
b8f8c3cf TG |
561 | if (!ts->inidle || !ts->tick_stopped) { |
562 | ts->inidle = 0; | |
6378ddb5 | 563 | local_irq_enable(); |
79bf2bb3 | 564 | return; |
6378ddb5 | 565 | } |
79bf2bb3 | 566 | |
b8f8c3cf TG |
567 | ts->inidle = 0; |
568 | ||
79bf2bb3 | 569 | /* Update jiffies first */ |
46cb4b7c | 570 | select_nohz_load_balancer(0); |
79bf2bb3 | 571 | tick_do_update_jiffies64(now); |
79bf2bb3 | 572 | |
79741dd3 | 573 | #ifndef CONFIG_VIRT_CPU_ACCOUNTING |
79bf2bb3 TG |
574 | /* |
575 | * We stopped the tick in idle. Update process times would miss the | |
576 | * time we slept as update_process_times does only a 1 tick | |
577 | * accounting. Enforce that this is accounted to idle ! | |
578 | */ | |
579 | ticks = jiffies - ts->idle_jiffies; | |
580 | /* | |
581 | * We might be one off. Do not randomly account a huge number of ticks! | |
582 | */ | |
79741dd3 MS |
583 | if (ticks && ticks < LONG_MAX) |
584 | account_idle_ticks(ticks); | |
585 | #endif | |
79bf2bb3 | 586 | |
126e01bf | 587 | touch_softlockup_watchdog(); |
79bf2bb3 TG |
588 | /* |
589 | * Cancel the scheduled timer and restore the tick | |
590 | */ | |
591 | ts->tick_stopped = 0; | |
5df7fa1c | 592 | ts->idle_exittime = now; |
79bf2bb3 | 593 | |
c34bec5a | 594 | tick_nohz_restart(ts, now); |
79bf2bb3 | 595 | |
79bf2bb3 TG |
596 | local_irq_enable(); |
597 | } | |
598 | ||
599 | static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now) | |
600 | { | |
601 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
cc584b21 | 602 | return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0); |
79bf2bb3 TG |
603 | } |
604 | ||
605 | /* | |
606 | * The nohz low res interrupt handler | |
607 | */ | |
608 | static void tick_nohz_handler(struct clock_event_device *dev) | |
609 | { | |
610 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
611 | struct pt_regs *regs = get_irq_regs(); | |
d3ed7824 | 612 | int cpu = smp_processor_id(); |
79bf2bb3 TG |
613 | ktime_t now = ktime_get(); |
614 | ||
615 | dev->next_event.tv64 = KTIME_MAX; | |
616 | ||
d3ed7824 TG |
617 | /* |
618 | * Check if the do_timer duty was dropped. We don't care about | |
619 | * concurrency: This happens only when the cpu in charge went | |
620 | * into a long sleep. If two cpus happen to assign themself to | |
621 | * this duty, then the jiffies update is still serialized by | |
622 | * xtime_lock. | |
623 | */ | |
6441402b | 624 | if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) |
d3ed7824 TG |
625 | tick_do_timer_cpu = cpu; |
626 | ||
79bf2bb3 | 627 | /* Check, if the jiffies need an update */ |
d3ed7824 TG |
628 | if (tick_do_timer_cpu == cpu) |
629 | tick_do_update_jiffies64(now); | |
79bf2bb3 TG |
630 | |
631 | /* | |
632 | * When we are idle and the tick is stopped, we have to touch | |
633 | * the watchdog as we might not schedule for a really long | |
634 | * time. This happens on complete idle SMP systems while | |
635 | * waiting on the login prompt. We also increment the "start | |
636 | * of idle" jiffy stamp so the idle accounting adjustment we | |
637 | * do when we go busy again does not account too much ticks. | |
638 | */ | |
639 | if (ts->tick_stopped) { | |
640 | touch_softlockup_watchdog(); | |
641 | ts->idle_jiffies++; | |
642 | } | |
643 | ||
644 | update_process_times(user_mode(regs)); | |
645 | profile_tick(CPU_PROFILING); | |
646 | ||
79bf2bb3 TG |
647 | while (tick_nohz_reprogram(ts, now)) { |
648 | now = ktime_get(); | |
649 | tick_do_update_jiffies64(now); | |
650 | } | |
651 | } | |
652 | ||
653 | /** | |
654 | * tick_nohz_switch_to_nohz - switch to nohz mode | |
655 | */ | |
656 | static void tick_nohz_switch_to_nohz(void) | |
657 | { | |
658 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
659 | ktime_t next; | |
660 | ||
661 | if (!tick_nohz_enabled) | |
662 | return; | |
663 | ||
664 | local_irq_disable(); | |
665 | if (tick_switch_to_oneshot(tick_nohz_handler)) { | |
666 | local_irq_enable(); | |
667 | return; | |
668 | } | |
669 | ||
670 | ts->nohz_mode = NOHZ_MODE_LOWRES; | |
671 | ||
672 | /* | |
673 | * Recycle the hrtimer in ts, so we can share the | |
674 | * hrtimer_forward with the highres code. | |
675 | */ | |
676 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); | |
677 | /* Get the next period */ | |
678 | next = tick_init_jiffy_update(); | |
679 | ||
680 | for (;;) { | |
cc584b21 | 681 | hrtimer_set_expires(&ts->sched_timer, next); |
79bf2bb3 TG |
682 | if (!tick_program_event(next, 0)) |
683 | break; | |
684 | next = ktime_add(next, tick_period); | |
685 | } | |
686 | local_irq_enable(); | |
79bf2bb3 TG |
687 | } |
688 | ||
fb02fbc1 TG |
689 | /* |
690 | * When NOHZ is enabled and the tick is stopped, we need to kick the | |
691 | * tick timer from irq_enter() so that the jiffies update is kept | |
692 | * alive during long running softirqs. That's ugly as hell, but | |
693 | * correctness is key even if we need to fix the offending softirq in | |
694 | * the first place. | |
695 | * | |
696 | * Note, this is different to tick_nohz_restart. We just kick the | |
697 | * timer and do not touch the other magic bits which need to be done | |
698 | * when idle is left. | |
699 | */ | |
eed3b9cf | 700 | static void tick_nohz_kick_tick(int cpu, ktime_t now) |
fb02fbc1 | 701 | { |
ae99286b TG |
702 | #if 0 |
703 | /* Switch back to 2.6.27 behaviour */ | |
704 | ||
fb02fbc1 | 705 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); |
eed3b9cf | 706 | ktime_t delta; |
fb02fbc1 | 707 | |
c4bd822e TG |
708 | /* |
709 | * Do not touch the tick device, when the next expiry is either | |
710 | * already reached or less/equal than the tick period. | |
711 | */ | |
268a3dcf | 712 | delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now); |
c4bd822e TG |
713 | if (delta.tv64 <= tick_period.tv64) |
714 | return; | |
715 | ||
716 | tick_nohz_restart(ts, now); | |
ae99286b | 717 | #endif |
fb02fbc1 TG |
718 | } |
719 | ||
eed3b9cf MS |
720 | static inline void tick_check_nohz(int cpu) |
721 | { | |
722 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
723 | ktime_t now; | |
724 | ||
725 | if (!ts->idle_active && !ts->tick_stopped) | |
726 | return; | |
727 | now = ktime_get(); | |
728 | if (ts->idle_active) | |
729 | tick_nohz_stop_idle(cpu, now); | |
730 | if (ts->tick_stopped) { | |
731 | tick_nohz_update_jiffies(now); | |
732 | tick_nohz_kick_tick(cpu, now); | |
733 | } | |
734 | } | |
735 | ||
79bf2bb3 TG |
736 | #else |
737 | ||
738 | static inline void tick_nohz_switch_to_nohz(void) { } | |
eed3b9cf | 739 | static inline void tick_check_nohz(int cpu) { } |
79bf2bb3 TG |
740 | |
741 | #endif /* NO_HZ */ | |
742 | ||
719254fa TG |
743 | /* |
744 | * Called from irq_enter to notify about the possible interruption of idle() | |
745 | */ | |
746 | void tick_check_idle(int cpu) | |
747 | { | |
fb02fbc1 | 748 | tick_check_oneshot_broadcast(cpu); |
eed3b9cf | 749 | tick_check_nohz(cpu); |
719254fa TG |
750 | } |
751 | ||
79bf2bb3 TG |
752 | /* |
753 | * High resolution timer specific code | |
754 | */ | |
755 | #ifdef CONFIG_HIGH_RES_TIMERS | |
756 | /* | |
4c9dc641 | 757 | * We rearm the timer until we get disabled by the idle code. |
79bf2bb3 TG |
758 | * Called with interrupts disabled and timer->base->cpu_base->lock held. |
759 | */ | |
760 | static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) | |
761 | { | |
762 | struct tick_sched *ts = | |
763 | container_of(timer, struct tick_sched, sched_timer); | |
79bf2bb3 TG |
764 | struct pt_regs *regs = get_irq_regs(); |
765 | ktime_t now = ktime_get(); | |
d3ed7824 TG |
766 | int cpu = smp_processor_id(); |
767 | ||
768 | #ifdef CONFIG_NO_HZ | |
769 | /* | |
770 | * Check if the do_timer duty was dropped. We don't care about | |
771 | * concurrency: This happens only when the cpu in charge went | |
772 | * into a long sleep. If two cpus happen to assign themself to | |
773 | * this duty, then the jiffies update is still serialized by | |
774 | * xtime_lock. | |
775 | */ | |
6441402b | 776 | if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) |
d3ed7824 TG |
777 | tick_do_timer_cpu = cpu; |
778 | #endif | |
79bf2bb3 TG |
779 | |
780 | /* Check, if the jiffies need an update */ | |
d3ed7824 TG |
781 | if (tick_do_timer_cpu == cpu) |
782 | tick_do_update_jiffies64(now); | |
79bf2bb3 TG |
783 | |
784 | /* | |
785 | * Do not call, when we are not in irq context and have | |
786 | * no valid regs pointer | |
787 | */ | |
788 | if (regs) { | |
789 | /* | |
790 | * When we are idle and the tick is stopped, we have to touch | |
791 | * the watchdog as we might not schedule for a really long | |
792 | * time. This happens on complete idle SMP systems while | |
793 | * waiting on the login prompt. We also increment the "start of | |
794 | * idle" jiffy stamp so the idle accounting adjustment we do | |
795 | * when we go busy again does not account too much ticks. | |
796 | */ | |
797 | if (ts->tick_stopped) { | |
798 | touch_softlockup_watchdog(); | |
799 | ts->idle_jiffies++; | |
800 | } | |
79bf2bb3 TG |
801 | update_process_times(user_mode(regs)); |
802 | profile_tick(CPU_PROFILING); | |
79bf2bb3 TG |
803 | } |
804 | ||
79bf2bb3 TG |
805 | hrtimer_forward(timer, now, tick_period); |
806 | ||
807 | return HRTIMER_RESTART; | |
808 | } | |
809 | ||
810 | /** | |
811 | * tick_setup_sched_timer - setup the tick emulation timer | |
812 | */ | |
813 | void tick_setup_sched_timer(void) | |
814 | { | |
815 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
816 | ktime_t now = ktime_get(); | |
817 | ||
818 | /* | |
819 | * Emulate tick processing via per-CPU hrtimers: | |
820 | */ | |
821 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); | |
822 | ts->sched_timer.function = tick_sched_timer; | |
79bf2bb3 | 823 | |
3704540b | 824 | /* Get the next period (per cpu) */ |
cc584b21 | 825 | hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); |
79bf2bb3 TG |
826 | |
827 | for (;;) { | |
828 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
5c333864 AB |
829 | hrtimer_start_expires(&ts->sched_timer, |
830 | HRTIMER_MODE_ABS_PINNED); | |
79bf2bb3 TG |
831 | /* Check, if the timer was already in the past */ |
832 | if (hrtimer_active(&ts->sched_timer)) | |
833 | break; | |
834 | now = ktime_get(); | |
835 | } | |
836 | ||
837 | #ifdef CONFIG_NO_HZ | |
29c158e8 | 838 | if (tick_nohz_enabled) |
79bf2bb3 TG |
839 | ts->nohz_mode = NOHZ_MODE_HIGHRES; |
840 | #endif | |
841 | } | |
3c4fbe5e | 842 | #endif /* HIGH_RES_TIMERS */ |
79bf2bb3 | 843 | |
3c4fbe5e | 844 | #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS |
79bf2bb3 TG |
845 | void tick_cancel_sched_timer(int cpu) |
846 | { | |
847 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
848 | ||
3c4fbe5e | 849 | # ifdef CONFIG_HIGH_RES_TIMERS |
79bf2bb3 TG |
850 | if (ts->sched_timer.base) |
851 | hrtimer_cancel(&ts->sched_timer); | |
3c4fbe5e | 852 | # endif |
a7901766 | 853 | |
79bf2bb3 TG |
854 | ts->nohz_mode = NOHZ_MODE_INACTIVE; |
855 | } | |
3c4fbe5e | 856 | #endif |
79bf2bb3 TG |
857 | |
858 | /** | |
859 | * Async notification about clocksource changes | |
860 | */ | |
861 | void tick_clock_notify(void) | |
862 | { | |
863 | int cpu; | |
864 | ||
865 | for_each_possible_cpu(cpu) | |
866 | set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); | |
867 | } | |
868 | ||
869 | /* | |
870 | * Async notification about clock event changes | |
871 | */ | |
872 | void tick_oneshot_notify(void) | |
873 | { | |
874 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
875 | ||
876 | set_bit(0, &ts->check_clocks); | |
877 | } | |
878 | ||
879 | /** | |
880 | * Check, if a change happened, which makes oneshot possible. | |
881 | * | |
882 | * Called cyclic from the hrtimer softirq (driven by the timer | |
883 | * softirq) allow_nohz signals, that we can switch into low-res nohz | |
884 | * mode, because high resolution timers are disabled (either compile | |
885 | * or runtime). | |
886 | */ | |
887 | int tick_check_oneshot_change(int allow_nohz) | |
888 | { | |
889 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
890 | ||
891 | if (!test_and_clear_bit(0, &ts->check_clocks)) | |
892 | return 0; | |
893 | ||
894 | if (ts->nohz_mode != NOHZ_MODE_INACTIVE) | |
895 | return 0; | |
896 | ||
cf4fc6cb | 897 | if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) |
79bf2bb3 TG |
898 | return 0; |
899 | ||
900 | if (!allow_nohz) | |
901 | return 1; | |
902 | ||
903 | tick_nohz_switch_to_nohz(); | |
904 | return 0; | |
905 | } |