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 | ||
5bb96226 FW |
101 | |
102 | static void tick_sched_do_timer(ktime_t now) | |
103 | { | |
104 | int cpu = smp_processor_id(); | |
105 | ||
106 | #ifdef CONFIG_NO_HZ | |
107 | /* | |
108 | * Check if the do_timer duty was dropped. We don't care about | |
109 | * concurrency: This happens only when the cpu in charge went | |
110 | * into a long sleep. If two cpus happen to assign themself to | |
111 | * this duty, then the jiffies update is still serialized by | |
112 | * xtime_lock. | |
113 | */ | |
114 | if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) | |
115 | tick_do_timer_cpu = cpu; | |
116 | #endif | |
117 | ||
118 | /* Check, if the jiffies need an update */ | |
119 | if (tick_do_timer_cpu == cpu) | |
120 | tick_do_update_jiffies64(now); | |
121 | } | |
122 | ||
79bf2bb3 TG |
123 | /* |
124 | * NOHZ - aka dynamic tick functionality | |
125 | */ | |
126 | #ifdef CONFIG_NO_HZ | |
127 | /* | |
128 | * NO HZ enabled ? | |
129 | */ | |
9d2ad243 | 130 | int tick_nohz_enabled __read_mostly = 1; |
79bf2bb3 TG |
131 | |
132 | /* | |
133 | * Enable / Disable tickless mode | |
134 | */ | |
135 | static int __init setup_tick_nohz(char *str) | |
136 | { | |
137 | if (!strcmp(str, "off")) | |
138 | tick_nohz_enabled = 0; | |
139 | else if (!strcmp(str, "on")) | |
140 | tick_nohz_enabled = 1; | |
141 | else | |
142 | return 0; | |
143 | return 1; | |
144 | } | |
145 | ||
146 | __setup("nohz=", setup_tick_nohz); | |
147 | ||
148 | /** | |
149 | * tick_nohz_update_jiffies - update jiffies when idle was interrupted | |
150 | * | |
151 | * Called from interrupt entry when the CPU was idle | |
152 | * | |
153 | * In case the sched_tick was stopped on this CPU, we have to check if jiffies | |
154 | * must be updated. Otherwise an interrupt handler could use a stale jiffy | |
155 | * value. We do this unconditionally on any cpu, as we don't know whether the | |
156 | * cpu, which has the update task assigned is in a long sleep. | |
157 | */ | |
eed3b9cf | 158 | static void tick_nohz_update_jiffies(ktime_t now) |
79bf2bb3 TG |
159 | { |
160 | int cpu = smp_processor_id(); | |
161 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
162 | unsigned long flags; | |
79bf2bb3 | 163 | |
5df7fa1c | 164 | ts->idle_waketime = now; |
79bf2bb3 TG |
165 | |
166 | local_irq_save(flags); | |
167 | tick_do_update_jiffies64(now); | |
168 | local_irq_restore(flags); | |
02ff3755 IM |
169 | |
170 | touch_softlockup_watchdog(); | |
79bf2bb3 TG |
171 | } |
172 | ||
595aac48 AV |
173 | /* |
174 | * Updates the per cpu time idle statistics counters | |
175 | */ | |
8d63bf94 | 176 | static void |
8c215bd3 | 177 | update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time) |
6378ddb5 | 178 | { |
eed3b9cf | 179 | ktime_t delta; |
6378ddb5 | 180 | |
595aac48 AV |
181 | if (ts->idle_active) { |
182 | delta = ktime_sub(now, ts->idle_entrytime); | |
8c215bd3 | 183 | if (nr_iowait_cpu(cpu) > 0) |
0224cf4c | 184 | ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta); |
6beea0cd MH |
185 | else |
186 | ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); | |
8c7b09f4 | 187 | ts->idle_entrytime = now; |
595aac48 | 188 | } |
8d63bf94 | 189 | |
e0e37c20 | 190 | if (last_update_time) |
8d63bf94 AV |
191 | *last_update_time = ktime_to_us(now); |
192 | ||
595aac48 AV |
193 | } |
194 | ||
195 | static void tick_nohz_stop_idle(int cpu, ktime_t now) | |
196 | { | |
197 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
198 | ||
8c215bd3 | 199 | update_ts_time_stats(cpu, ts, now, NULL); |
eed3b9cf | 200 | ts->idle_active = 0; |
56c7426b | 201 | |
eed3b9cf | 202 | sched_clock_idle_wakeup_event(0); |
6378ddb5 VP |
203 | } |
204 | ||
8c215bd3 | 205 | static ktime_t tick_nohz_start_idle(int cpu, struct tick_sched *ts) |
6378ddb5 | 206 | { |
430ee881 | 207 | ktime_t now = ktime_get(); |
595aac48 | 208 | |
6378ddb5 VP |
209 | ts->idle_entrytime = now; |
210 | ts->idle_active = 1; | |
56c7426b | 211 | sched_clock_idle_sleep_event(); |
6378ddb5 VP |
212 | return now; |
213 | } | |
214 | ||
b1f724c3 AV |
215 | /** |
216 | * get_cpu_idle_time_us - get the total idle time of a cpu | |
217 | * @cpu: CPU number to query | |
09a1d34f MH |
218 | * @last_update_time: variable to store update time in. Do not update |
219 | * counters if NULL. | |
b1f724c3 AV |
220 | * |
221 | * Return the cummulative idle time (since boot) for a given | |
6beea0cd | 222 | * CPU, in microseconds. |
b1f724c3 AV |
223 | * |
224 | * This time is measured via accounting rather than sampling, | |
225 | * and is as accurate as ktime_get() is. | |
226 | * | |
227 | * This function returns -1 if NOHZ is not enabled. | |
228 | */ | |
6378ddb5 VP |
229 | u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) |
230 | { | |
231 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
09a1d34f | 232 | ktime_t now, idle; |
6378ddb5 | 233 | |
8083e4ad | 234 | if (!tick_nohz_enabled) |
235 | return -1; | |
236 | ||
09a1d34f MH |
237 | now = ktime_get(); |
238 | if (last_update_time) { | |
239 | update_ts_time_stats(cpu, ts, now, last_update_time); | |
240 | idle = ts->idle_sleeptime; | |
241 | } else { | |
242 | if (ts->idle_active && !nr_iowait_cpu(cpu)) { | |
243 | ktime_t delta = ktime_sub(now, ts->idle_entrytime); | |
244 | ||
245 | idle = ktime_add(ts->idle_sleeptime, delta); | |
246 | } else { | |
247 | idle = ts->idle_sleeptime; | |
248 | } | |
249 | } | |
250 | ||
251 | return ktime_to_us(idle); | |
8083e4ad | 252 | |
6378ddb5 | 253 | } |
8083e4ad | 254 | EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); |
6378ddb5 | 255 | |
6beea0cd | 256 | /** |
0224cf4c AV |
257 | * get_cpu_iowait_time_us - get the total iowait time of a cpu |
258 | * @cpu: CPU number to query | |
09a1d34f MH |
259 | * @last_update_time: variable to store update time in. Do not update |
260 | * counters if NULL. | |
0224cf4c AV |
261 | * |
262 | * Return the cummulative iowait time (since boot) for a given | |
263 | * CPU, in microseconds. | |
264 | * | |
265 | * This time is measured via accounting rather than sampling, | |
266 | * and is as accurate as ktime_get() is. | |
267 | * | |
268 | * This function returns -1 if NOHZ is not enabled. | |
269 | */ | |
270 | u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time) | |
271 | { | |
272 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
09a1d34f | 273 | ktime_t now, iowait; |
0224cf4c AV |
274 | |
275 | if (!tick_nohz_enabled) | |
276 | return -1; | |
277 | ||
09a1d34f MH |
278 | now = ktime_get(); |
279 | if (last_update_time) { | |
280 | update_ts_time_stats(cpu, ts, now, last_update_time); | |
281 | iowait = ts->iowait_sleeptime; | |
282 | } else { | |
283 | if (ts->idle_active && nr_iowait_cpu(cpu) > 0) { | |
284 | ktime_t delta = ktime_sub(now, ts->idle_entrytime); | |
0224cf4c | 285 | |
09a1d34f MH |
286 | iowait = ktime_add(ts->iowait_sleeptime, delta); |
287 | } else { | |
288 | iowait = ts->iowait_sleeptime; | |
289 | } | |
290 | } | |
0224cf4c | 291 | |
09a1d34f | 292 | return ktime_to_us(iowait); |
0224cf4c AV |
293 | } |
294 | EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us); | |
295 | ||
84bf1bcc FW |
296 | static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts, |
297 | ktime_t now, int cpu) | |
79bf2bb3 | 298 | { |
280f0677 | 299 | unsigned long seq, last_jiffies, next_jiffies, delta_jiffies; |
84bf1bcc | 300 | ktime_t last_update, expires, ret = { .tv64 = 0 }; |
aa9b1630 | 301 | unsigned long rcu_delta_jiffies; |
4f86d3a8 | 302 | struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev; |
98962465 | 303 | u64 time_delta; |
79bf2bb3 | 304 | |
79bf2bb3 TG |
305 | /* Read jiffies and the time when jiffies were updated last */ |
306 | do { | |
307 | seq = read_seqbegin(&xtime_lock); | |
308 | last_update = last_jiffies_update; | |
309 | last_jiffies = jiffies; | |
27185016 | 310 | time_delta = timekeeping_max_deferment(); |
79bf2bb3 TG |
311 | } while (read_seqretry(&xtime_lock, seq)); |
312 | ||
aa9b1630 | 313 | if (rcu_needs_cpu(cpu, &rcu_delta_jiffies) || printk_needs_cpu(cpu) || |
396e894d | 314 | arch_needs_cpu(cpu)) { |
3c5d92a0 | 315 | next_jiffies = last_jiffies + 1; |
6ba9b346 | 316 | delta_jiffies = 1; |
3c5d92a0 MS |
317 | } else { |
318 | /* Get the next timer wheel timer */ | |
319 | next_jiffies = get_next_timer_interrupt(last_jiffies); | |
320 | delta_jiffies = next_jiffies - last_jiffies; | |
aa9b1630 PM |
321 | if (rcu_delta_jiffies < delta_jiffies) { |
322 | next_jiffies = last_jiffies + rcu_delta_jiffies; | |
323 | delta_jiffies = rcu_delta_jiffies; | |
324 | } | |
3c5d92a0 | 325 | } |
79bf2bb3 TG |
326 | /* |
327 | * Do not stop the tick, if we are only one off | |
328 | * or if the cpu is required for rcu | |
329 | */ | |
6ba9b346 | 330 | if (!ts->tick_stopped && delta_jiffies == 1) |
79bf2bb3 TG |
331 | goto out; |
332 | ||
333 | /* Schedule the tick, if we are at least one jiffie off */ | |
334 | if ((long)delta_jiffies >= 1) { | |
335 | ||
00147449 WR |
336 | /* |
337 | * If this cpu is the one which updates jiffies, then | |
338 | * give up the assignment and let it be taken by the | |
339 | * cpu which runs the tick timer next, which might be | |
340 | * this cpu as well. If we don't drop this here the | |
341 | * jiffies might be stale and do_timer() never | |
27185016 TG |
342 | * invoked. Keep track of the fact that it was the one |
343 | * which had the do_timer() duty last. If this cpu is | |
344 | * the one which had the do_timer() duty last, we | |
345 | * limit the sleep time to the timekeeping | |
346 | * max_deferement value which we retrieved | |
347 | * above. Otherwise we can sleep as long as we want. | |
00147449 | 348 | */ |
27185016 | 349 | if (cpu == tick_do_timer_cpu) { |
00147449 | 350 | tick_do_timer_cpu = TICK_DO_TIMER_NONE; |
27185016 TG |
351 | ts->do_timer_last = 1; |
352 | } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { | |
353 | time_delta = KTIME_MAX; | |
354 | ts->do_timer_last = 0; | |
355 | } else if (!ts->do_timer_last) { | |
356 | time_delta = KTIME_MAX; | |
357 | } | |
358 | ||
00147449 | 359 | /* |
98962465 JH |
360 | * calculate the expiry time for the next timer wheel |
361 | * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals | |
362 | * that there is no timer pending or at least extremely | |
363 | * far into the future (12 days for HZ=1000). In this | |
364 | * case we set the expiry to the end of time. | |
365 | */ | |
366 | if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) { | |
367 | /* | |
368 | * Calculate the time delta for the next timer event. | |
369 | * If the time delta exceeds the maximum time delta | |
370 | * permitted by the current clocksource then adjust | |
371 | * the time delta accordingly to ensure the | |
372 | * clocksource does not wrap. | |
373 | */ | |
374 | time_delta = min_t(u64, time_delta, | |
375 | tick_period.tv64 * delta_jiffies); | |
98962465 | 376 | } |
00147449 | 377 | |
27185016 TG |
378 | if (time_delta < KTIME_MAX) |
379 | expires = ktime_add_ns(last_update, time_delta); | |
380 | else | |
381 | expires.tv64 = KTIME_MAX; | |
00147449 | 382 | |
00147449 WR |
383 | /* Skip reprogram of event if its not changed */ |
384 | if (ts->tick_stopped && ktime_equal(expires, dev->next_event)) | |
385 | goto out; | |
386 | ||
84bf1bcc FW |
387 | ret = expires; |
388 | ||
79bf2bb3 TG |
389 | /* |
390 | * nohz_stop_sched_tick can be called several times before | |
391 | * the nohz_restart_sched_tick is called. This happens when | |
392 | * interrupts arrive which do not cause a reschedule. In the | |
393 | * first call we save the current tick time, so we can restart | |
394 | * the scheduler tick in nohz_restart_sched_tick. | |
395 | */ | |
396 | if (!ts->tick_stopped) { | |
c1cc017c | 397 | nohz_balance_enter_idle(cpu); |
5167e8d5 | 398 | calc_load_enter_idle(); |
46cb4b7c | 399 | |
f5d411c9 | 400 | ts->last_tick = hrtimer_get_expires(&ts->sched_timer); |
79bf2bb3 | 401 | ts->tick_stopped = 1; |
79bf2bb3 | 402 | } |
d3ed7824 | 403 | |
eaad084b | 404 | /* |
98962465 JH |
405 | * If the expiration time == KTIME_MAX, then |
406 | * in this case we simply stop the tick timer. | |
eaad084b | 407 | */ |
98962465 | 408 | if (unlikely(expires.tv64 == KTIME_MAX)) { |
eaad084b TG |
409 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) |
410 | hrtimer_cancel(&ts->sched_timer); | |
411 | goto out; | |
412 | } | |
413 | ||
79bf2bb3 TG |
414 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { |
415 | hrtimer_start(&ts->sched_timer, expires, | |
5c333864 | 416 | HRTIMER_MODE_ABS_PINNED); |
79bf2bb3 TG |
417 | /* Check, if the timer was already in the past */ |
418 | if (hrtimer_active(&ts->sched_timer)) | |
419 | goto out; | |
4c9dc641 | 420 | } else if (!tick_program_event(expires, 0)) |
79bf2bb3 TG |
421 | goto out; |
422 | /* | |
423 | * We are past the event already. So we crossed a | |
424 | * jiffie boundary. Update jiffies and raise the | |
425 | * softirq. | |
426 | */ | |
427 | tick_do_update_jiffies64(ktime_get()); | |
79bf2bb3 TG |
428 | } |
429 | raise_softirq_irqoff(TIMER_SOFTIRQ); | |
430 | out: | |
431 | ts->next_jiffies = next_jiffies; | |
432 | ts->last_jiffies = last_jiffies; | |
4f86d3a8 | 433 | ts->sleep_length = ktime_sub(dev->next_event, now); |
84bf1bcc FW |
434 | |
435 | return ret; | |
280f0677 FW |
436 | } |
437 | ||
5b39939a FW |
438 | static bool can_stop_idle_tick(int cpu, struct tick_sched *ts) |
439 | { | |
440 | /* | |
441 | * If this cpu is offline and it is the one which updates | |
442 | * jiffies, then give up the assignment and let it be taken by | |
443 | * the cpu which runs the tick timer next. If we don't drop | |
444 | * this here the jiffies might be stale and do_timer() never | |
445 | * invoked. | |
446 | */ | |
447 | if (unlikely(!cpu_online(cpu))) { | |
448 | if (cpu == tick_do_timer_cpu) | |
449 | tick_do_timer_cpu = TICK_DO_TIMER_NONE; | |
450 | } | |
451 | ||
452 | if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) | |
453 | return false; | |
454 | ||
455 | if (need_resched()) | |
456 | return false; | |
457 | ||
458 | if (unlikely(local_softirq_pending() && cpu_online(cpu))) { | |
459 | static int ratelimit; | |
460 | ||
803b0eba PM |
461 | if (ratelimit < 10 && |
462 | (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) { | |
5b39939a FW |
463 | printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n", |
464 | (unsigned int) local_softirq_pending()); | |
465 | ratelimit++; | |
466 | } | |
467 | return false; | |
468 | } | |
469 | ||
470 | return true; | |
471 | } | |
472 | ||
19f5f736 FW |
473 | static void __tick_nohz_idle_enter(struct tick_sched *ts) |
474 | { | |
84bf1bcc | 475 | ktime_t now, expires; |
5b39939a | 476 | int cpu = smp_processor_id(); |
19f5f736 | 477 | |
5b39939a | 478 | now = tick_nohz_start_idle(cpu, ts); |
2ac0d98f | 479 | |
5b39939a FW |
480 | if (can_stop_idle_tick(cpu, ts)) { |
481 | int was_stopped = ts->tick_stopped; | |
482 | ||
483 | ts->idle_calls++; | |
84bf1bcc FW |
484 | |
485 | expires = tick_nohz_stop_sched_tick(ts, now, cpu); | |
486 | if (expires.tv64 > 0LL) { | |
487 | ts->idle_sleeps++; | |
488 | ts->idle_expires = expires; | |
489 | } | |
5b39939a FW |
490 | |
491 | if (!was_stopped && ts->tick_stopped) | |
492 | ts->idle_jiffies = ts->last_jiffies; | |
493 | } | |
280f0677 FW |
494 | } |
495 | ||
496 | /** | |
497 | * tick_nohz_idle_enter - stop the idle tick from the idle task | |
498 | * | |
499 | * When the next event is more than a tick into the future, stop the idle tick | |
500 | * Called when we start the idle loop. | |
2bbb6817 | 501 | * |
1268fbc7 | 502 | * The arch is responsible of calling: |
2bbb6817 FW |
503 | * |
504 | * - rcu_idle_enter() after its last use of RCU before the CPU is put | |
505 | * to sleep. | |
506 | * - rcu_idle_exit() before the first use of RCU after the CPU is woken up. | |
280f0677 | 507 | */ |
1268fbc7 | 508 | void tick_nohz_idle_enter(void) |
280f0677 FW |
509 | { |
510 | struct tick_sched *ts; | |
511 | ||
1268fbc7 FW |
512 | WARN_ON_ONCE(irqs_disabled()); |
513 | ||
0db49b72 LT |
514 | /* |
515 | * Update the idle state in the scheduler domain hierarchy | |
516 | * when tick_nohz_stop_sched_tick() is called from the idle loop. | |
517 | * State will be updated to busy during the first busy tick after | |
518 | * exiting idle. | |
519 | */ | |
520 | set_cpu_sd_state_idle(); | |
521 | ||
1268fbc7 FW |
522 | local_irq_disable(); |
523 | ||
280f0677 FW |
524 | ts = &__get_cpu_var(tick_cpu_sched); |
525 | /* | |
526 | * set ts->inidle unconditionally. even if the system did not | |
527 | * switch to nohz mode the cpu frequency governers rely on the | |
528 | * update of the idle time accounting in tick_nohz_start_idle(). | |
529 | */ | |
530 | ts->inidle = 1; | |
19f5f736 | 531 | __tick_nohz_idle_enter(ts); |
1268fbc7 FW |
532 | |
533 | local_irq_enable(); | |
280f0677 FW |
534 | } |
535 | ||
536 | /** | |
537 | * tick_nohz_irq_exit - update next tick event from interrupt exit | |
538 | * | |
539 | * When an interrupt fires while we are idle and it doesn't cause | |
540 | * a reschedule, it may still add, modify or delete a timer, enqueue | |
541 | * an RCU callback, etc... | |
542 | * So we need to re-calculate and reprogram the next tick event. | |
543 | */ | |
544 | void tick_nohz_irq_exit(void) | |
545 | { | |
546 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
547 | ||
548 | if (!ts->inidle) | |
549 | return; | |
550 | ||
19f5f736 | 551 | __tick_nohz_idle_enter(ts); |
79bf2bb3 TG |
552 | } |
553 | ||
4f86d3a8 LB |
554 | /** |
555 | * tick_nohz_get_sleep_length - return the length of the current sleep | |
556 | * | |
557 | * Called from power state control code with interrupts disabled | |
558 | */ | |
559 | ktime_t tick_nohz_get_sleep_length(void) | |
560 | { | |
561 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
562 | ||
563 | return ts->sleep_length; | |
564 | } | |
565 | ||
c34bec5a TG |
566 | static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) |
567 | { | |
568 | hrtimer_cancel(&ts->sched_timer); | |
f5d411c9 | 569 | hrtimer_set_expires(&ts->sched_timer, ts->last_tick); |
c34bec5a TG |
570 | |
571 | while (1) { | |
572 | /* Forward the time to expire in the future */ | |
573 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
574 | ||
575 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { | |
268a3dcf | 576 | hrtimer_start_expires(&ts->sched_timer, |
5c333864 | 577 | HRTIMER_MODE_ABS_PINNED); |
c34bec5a TG |
578 | /* Check, if the timer was already in the past */ |
579 | if (hrtimer_active(&ts->sched_timer)) | |
580 | break; | |
581 | } else { | |
268a3dcf TG |
582 | if (!tick_program_event( |
583 | hrtimer_get_expires(&ts->sched_timer), 0)) | |
c34bec5a TG |
584 | break; |
585 | } | |
6f103929 | 586 | /* Reread time and update jiffies */ |
c34bec5a | 587 | now = ktime_get(); |
6f103929 | 588 | tick_do_update_jiffies64(now); |
c34bec5a TG |
589 | } |
590 | } | |
591 | ||
19f5f736 | 592 | static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now) |
79bf2bb3 | 593 | { |
79bf2bb3 | 594 | /* Update jiffies first */ |
79bf2bb3 | 595 | tick_do_update_jiffies64(now); |
5aaa0b7a | 596 | update_cpu_load_nohz(); |
79bf2bb3 | 597 | |
749c8814 | 598 | calc_load_exit_idle(); |
2ac0d98f FW |
599 | touch_softlockup_watchdog(); |
600 | /* | |
601 | * Cancel the scheduled timer and restore the tick | |
602 | */ | |
603 | ts->tick_stopped = 0; | |
604 | ts->idle_exittime = now; | |
605 | ||
606 | tick_nohz_restart(ts, now); | |
607 | } | |
608 | ||
609 | static void tick_nohz_account_idle_ticks(struct tick_sched *ts) | |
610 | { | |
79741dd3 | 611 | #ifndef CONFIG_VIRT_CPU_ACCOUNTING |
2ac0d98f | 612 | unsigned long ticks; |
79bf2bb3 TG |
613 | /* |
614 | * We stopped the tick in idle. Update process times would miss the | |
615 | * time we slept as update_process_times does only a 1 tick | |
616 | * accounting. Enforce that this is accounted to idle ! | |
617 | */ | |
618 | ticks = jiffies - ts->idle_jiffies; | |
619 | /* | |
620 | * We might be one off. Do not randomly account a huge number of ticks! | |
621 | */ | |
79741dd3 MS |
622 | if (ticks && ticks < LONG_MAX) |
623 | account_idle_ticks(ticks); | |
624 | #endif | |
19f5f736 FW |
625 | } |
626 | ||
79bf2bb3 | 627 | /** |
280f0677 | 628 | * tick_nohz_idle_exit - restart the idle tick from the idle task |
79bf2bb3 TG |
629 | * |
630 | * Restart the idle tick when the CPU is woken up from idle | |
280f0677 FW |
631 | * This also exit the RCU extended quiescent state. The CPU |
632 | * can use RCU again after this function is called. | |
79bf2bb3 | 633 | */ |
280f0677 | 634 | void tick_nohz_idle_exit(void) |
79bf2bb3 TG |
635 | { |
636 | int cpu = smp_processor_id(); | |
637 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
6378ddb5 | 638 | ktime_t now; |
79bf2bb3 | 639 | |
6378ddb5 | 640 | local_irq_disable(); |
2bbb6817 | 641 | |
15f827be FW |
642 | WARN_ON_ONCE(!ts->inidle); |
643 | ||
644 | ts->inidle = 0; | |
645 | ||
646 | if (ts->idle_active || ts->tick_stopped) | |
eed3b9cf MS |
647 | now = ktime_get(); |
648 | ||
649 | if (ts->idle_active) | |
650 | tick_nohz_stop_idle(cpu, now); | |
6378ddb5 | 651 | |
2ac0d98f | 652 | if (ts->tick_stopped) { |
19f5f736 | 653 | tick_nohz_restart_sched_tick(ts, now); |
2ac0d98f | 654 | tick_nohz_account_idle_ticks(ts); |
6378ddb5 | 655 | } |
79bf2bb3 | 656 | |
79bf2bb3 TG |
657 | local_irq_enable(); |
658 | } | |
659 | ||
660 | static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now) | |
661 | { | |
662 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
cc584b21 | 663 | return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0); |
79bf2bb3 TG |
664 | } |
665 | ||
666 | /* | |
667 | * The nohz low res interrupt handler | |
668 | */ | |
669 | static void tick_nohz_handler(struct clock_event_device *dev) | |
670 | { | |
671 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
672 | struct pt_regs *regs = get_irq_regs(); | |
673 | ktime_t now = ktime_get(); | |
674 | ||
675 | dev->next_event.tv64 = KTIME_MAX; | |
676 | ||
5bb96226 | 677 | tick_sched_do_timer(now); |
79bf2bb3 TG |
678 | |
679 | /* | |
680 | * When we are idle and the tick is stopped, we have to touch | |
681 | * the watchdog as we might not schedule for a really long | |
682 | * time. This happens on complete idle SMP systems while | |
683 | * waiting on the login prompt. We also increment the "start | |
684 | * of idle" jiffy stamp so the idle accounting adjustment we | |
685 | * do when we go busy again does not account too much ticks. | |
686 | */ | |
687 | if (ts->tick_stopped) { | |
688 | touch_softlockup_watchdog(); | |
689 | ts->idle_jiffies++; | |
690 | } | |
691 | ||
692 | update_process_times(user_mode(regs)); | |
693 | profile_tick(CPU_PROFILING); | |
694 | ||
79bf2bb3 TG |
695 | while (tick_nohz_reprogram(ts, now)) { |
696 | now = ktime_get(); | |
697 | tick_do_update_jiffies64(now); | |
698 | } | |
699 | } | |
700 | ||
701 | /** | |
702 | * tick_nohz_switch_to_nohz - switch to nohz mode | |
703 | */ | |
704 | static void tick_nohz_switch_to_nohz(void) | |
705 | { | |
706 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
707 | ktime_t next; | |
708 | ||
709 | if (!tick_nohz_enabled) | |
710 | return; | |
711 | ||
712 | local_irq_disable(); | |
713 | if (tick_switch_to_oneshot(tick_nohz_handler)) { | |
714 | local_irq_enable(); | |
715 | return; | |
716 | } | |
717 | ||
718 | ts->nohz_mode = NOHZ_MODE_LOWRES; | |
719 | ||
720 | /* | |
721 | * Recycle the hrtimer in ts, so we can share the | |
722 | * hrtimer_forward with the highres code. | |
723 | */ | |
724 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); | |
725 | /* Get the next period */ | |
726 | next = tick_init_jiffy_update(); | |
727 | ||
728 | for (;;) { | |
cc584b21 | 729 | hrtimer_set_expires(&ts->sched_timer, next); |
79bf2bb3 TG |
730 | if (!tick_program_event(next, 0)) |
731 | break; | |
732 | next = ktime_add(next, tick_period); | |
733 | } | |
734 | local_irq_enable(); | |
79bf2bb3 TG |
735 | } |
736 | ||
fb02fbc1 TG |
737 | /* |
738 | * When NOHZ is enabled and the tick is stopped, we need to kick the | |
739 | * tick timer from irq_enter() so that the jiffies update is kept | |
740 | * alive during long running softirqs. That's ugly as hell, but | |
741 | * correctness is key even if we need to fix the offending softirq in | |
742 | * the first place. | |
743 | * | |
744 | * Note, this is different to tick_nohz_restart. We just kick the | |
745 | * timer and do not touch the other magic bits which need to be done | |
746 | * when idle is left. | |
747 | */ | |
eed3b9cf | 748 | static void tick_nohz_kick_tick(int cpu, ktime_t now) |
fb02fbc1 | 749 | { |
ae99286b TG |
750 | #if 0 |
751 | /* Switch back to 2.6.27 behaviour */ | |
752 | ||
fb02fbc1 | 753 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); |
eed3b9cf | 754 | ktime_t delta; |
fb02fbc1 | 755 | |
c4bd822e TG |
756 | /* |
757 | * Do not touch the tick device, when the next expiry is either | |
758 | * already reached or less/equal than the tick period. | |
759 | */ | |
268a3dcf | 760 | delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now); |
c4bd822e TG |
761 | if (delta.tv64 <= tick_period.tv64) |
762 | return; | |
763 | ||
764 | tick_nohz_restart(ts, now); | |
ae99286b | 765 | #endif |
fb02fbc1 TG |
766 | } |
767 | ||
eed3b9cf MS |
768 | static inline void tick_check_nohz(int cpu) |
769 | { | |
770 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
771 | ktime_t now; | |
772 | ||
773 | if (!ts->idle_active && !ts->tick_stopped) | |
774 | return; | |
775 | now = ktime_get(); | |
776 | if (ts->idle_active) | |
777 | tick_nohz_stop_idle(cpu, now); | |
778 | if (ts->tick_stopped) { | |
779 | tick_nohz_update_jiffies(now); | |
780 | tick_nohz_kick_tick(cpu, now); | |
781 | } | |
782 | } | |
783 | ||
79bf2bb3 TG |
784 | #else |
785 | ||
786 | static inline void tick_nohz_switch_to_nohz(void) { } | |
eed3b9cf | 787 | static inline void tick_check_nohz(int cpu) { } |
79bf2bb3 TG |
788 | |
789 | #endif /* NO_HZ */ | |
790 | ||
719254fa TG |
791 | /* |
792 | * Called from irq_enter to notify about the possible interruption of idle() | |
793 | */ | |
794 | void tick_check_idle(int cpu) | |
795 | { | |
fb02fbc1 | 796 | tick_check_oneshot_broadcast(cpu); |
eed3b9cf | 797 | tick_check_nohz(cpu); |
719254fa TG |
798 | } |
799 | ||
79bf2bb3 TG |
800 | /* |
801 | * High resolution timer specific code | |
802 | */ | |
803 | #ifdef CONFIG_HIGH_RES_TIMERS | |
804 | /* | |
4c9dc641 | 805 | * We rearm the timer until we get disabled by the idle code. |
79bf2bb3 TG |
806 | * Called with interrupts disabled and timer->base->cpu_base->lock held. |
807 | */ | |
808 | static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) | |
809 | { | |
810 | struct tick_sched *ts = | |
811 | container_of(timer, struct tick_sched, sched_timer); | |
79bf2bb3 TG |
812 | struct pt_regs *regs = get_irq_regs(); |
813 | ktime_t now = ktime_get(); | |
d3ed7824 | 814 | |
5bb96226 | 815 | tick_sched_do_timer(now); |
79bf2bb3 TG |
816 | |
817 | /* | |
818 | * Do not call, when we are not in irq context and have | |
819 | * no valid regs pointer | |
820 | */ | |
821 | if (regs) { | |
822 | /* | |
823 | * When we are idle and the tick is stopped, we have to touch | |
824 | * the watchdog as we might not schedule for a really long | |
825 | * time. This happens on complete idle SMP systems while | |
826 | * waiting on the login prompt. We also increment the "start of | |
827 | * idle" jiffy stamp so the idle accounting adjustment we do | |
828 | * when we go busy again does not account too much ticks. | |
829 | */ | |
830 | if (ts->tick_stopped) { | |
831 | touch_softlockup_watchdog(); | |
2b17c545 | 832 | if (is_idle_task(current)) |
2ac0d98f | 833 | ts->idle_jiffies++; |
79bf2bb3 | 834 | } |
79bf2bb3 TG |
835 | update_process_times(user_mode(regs)); |
836 | profile_tick(CPU_PROFILING); | |
79bf2bb3 TG |
837 | } |
838 | ||
79bf2bb3 TG |
839 | hrtimer_forward(timer, now, tick_period); |
840 | ||
841 | return HRTIMER_RESTART; | |
842 | } | |
843 | ||
5307c955 MG |
844 | static int sched_skew_tick; |
845 | ||
62cf20b3 TG |
846 | static int __init skew_tick(char *str) |
847 | { | |
848 | get_option(&str, &sched_skew_tick); | |
849 | ||
850 | return 0; | |
851 | } | |
852 | early_param("skew_tick", skew_tick); | |
853 | ||
79bf2bb3 TG |
854 | /** |
855 | * tick_setup_sched_timer - setup the tick emulation timer | |
856 | */ | |
857 | void tick_setup_sched_timer(void) | |
858 | { | |
859 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
860 | ktime_t now = ktime_get(); | |
861 | ||
862 | /* | |
863 | * Emulate tick processing via per-CPU hrtimers: | |
864 | */ | |
865 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); | |
866 | ts->sched_timer.function = tick_sched_timer; | |
79bf2bb3 | 867 | |
3704540b | 868 | /* Get the next period (per cpu) */ |
cc584b21 | 869 | hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); |
79bf2bb3 | 870 | |
5307c955 MG |
871 | /* Offset the tick to avert xtime_lock contention. */ |
872 | if (sched_skew_tick) { | |
873 | u64 offset = ktime_to_ns(tick_period) >> 1; | |
874 | do_div(offset, num_possible_cpus()); | |
875 | offset *= smp_processor_id(); | |
876 | hrtimer_add_expires_ns(&ts->sched_timer, offset); | |
877 | } | |
878 | ||
79bf2bb3 TG |
879 | for (;;) { |
880 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
5c333864 AB |
881 | hrtimer_start_expires(&ts->sched_timer, |
882 | HRTIMER_MODE_ABS_PINNED); | |
79bf2bb3 TG |
883 | /* Check, if the timer was already in the past */ |
884 | if (hrtimer_active(&ts->sched_timer)) | |
885 | break; | |
886 | now = ktime_get(); | |
887 | } | |
888 | ||
889 | #ifdef CONFIG_NO_HZ | |
29c158e8 | 890 | if (tick_nohz_enabled) |
79bf2bb3 TG |
891 | ts->nohz_mode = NOHZ_MODE_HIGHRES; |
892 | #endif | |
893 | } | |
3c4fbe5e | 894 | #endif /* HIGH_RES_TIMERS */ |
79bf2bb3 | 895 | |
3c4fbe5e | 896 | #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS |
79bf2bb3 TG |
897 | void tick_cancel_sched_timer(int cpu) |
898 | { | |
899 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
900 | ||
3c4fbe5e | 901 | # ifdef CONFIG_HIGH_RES_TIMERS |
79bf2bb3 TG |
902 | if (ts->sched_timer.base) |
903 | hrtimer_cancel(&ts->sched_timer); | |
3c4fbe5e | 904 | # endif |
a7901766 | 905 | |
79bf2bb3 TG |
906 | ts->nohz_mode = NOHZ_MODE_INACTIVE; |
907 | } | |
3c4fbe5e | 908 | #endif |
79bf2bb3 TG |
909 | |
910 | /** | |
911 | * Async notification about clocksource changes | |
912 | */ | |
913 | void tick_clock_notify(void) | |
914 | { | |
915 | int cpu; | |
916 | ||
917 | for_each_possible_cpu(cpu) | |
918 | set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); | |
919 | } | |
920 | ||
921 | /* | |
922 | * Async notification about clock event changes | |
923 | */ | |
924 | void tick_oneshot_notify(void) | |
925 | { | |
926 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
927 | ||
928 | set_bit(0, &ts->check_clocks); | |
929 | } | |
930 | ||
931 | /** | |
932 | * Check, if a change happened, which makes oneshot possible. | |
933 | * | |
934 | * Called cyclic from the hrtimer softirq (driven by the timer | |
935 | * softirq) allow_nohz signals, that we can switch into low-res nohz | |
936 | * mode, because high resolution timers are disabled (either compile | |
937 | * or runtime). | |
938 | */ | |
939 | int tick_check_oneshot_change(int allow_nohz) | |
940 | { | |
941 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
942 | ||
943 | if (!test_and_clear_bit(0, &ts->check_clocks)) | |
944 | return 0; | |
945 | ||
946 | if (ts->nohz_mode != NOHZ_MODE_INACTIVE) | |
947 | return 0; | |
948 | ||
cf4fc6cb | 949 | if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) |
79bf2bb3 TG |
950 | return 0; |
951 | ||
952 | if (!allow_nohz) | |
953 | return 1; | |
954 | ||
955 | tick_nohz_switch_to_nohz(); | |
956 | return 0; | |
957 | } |