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> | |
22 | #include <linux/tick.h> | |
8083e4ad | 23 | #include <linux/module.h> |
79bf2bb3 | 24 | |
9e203bcc DM |
25 | #include <asm/irq_regs.h> |
26 | ||
79bf2bb3 TG |
27 | #include "tick-internal.h" |
28 | ||
29 | /* | |
30 | * Per cpu nohz control structure | |
31 | */ | |
32 | static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched); | |
33 | ||
34 | /* | |
35 | * The time, when the last jiffy update happened. Protected by xtime_lock. | |
36 | */ | |
37 | static ktime_t last_jiffies_update; | |
38 | ||
289f480a IM |
39 | struct tick_sched *tick_get_tick_sched(int cpu) |
40 | { | |
41 | return &per_cpu(tick_cpu_sched, cpu); | |
42 | } | |
43 | ||
79bf2bb3 TG |
44 | /* |
45 | * Must be called with interrupts disabled ! | |
46 | */ | |
47 | static void tick_do_update_jiffies64(ktime_t now) | |
48 | { | |
49 | unsigned long ticks = 0; | |
50 | ktime_t delta; | |
51 | ||
7a14ce1d IM |
52 | /* |
53 | * Do a quick check without holding xtime_lock: | |
54 | */ | |
55 | delta = ktime_sub(now, last_jiffies_update); | |
56 | if (delta.tv64 < tick_period.tv64) | |
57 | return; | |
58 | ||
79bf2bb3 TG |
59 | /* Reevalute with xtime_lock held */ |
60 | write_seqlock(&xtime_lock); | |
61 | ||
62 | delta = ktime_sub(now, last_jiffies_update); | |
63 | if (delta.tv64 >= tick_period.tv64) { | |
64 | ||
65 | delta = ktime_sub(delta, tick_period); | |
66 | last_jiffies_update = ktime_add(last_jiffies_update, | |
67 | tick_period); | |
68 | ||
69 | /* Slow path for long timeouts */ | |
70 | if (unlikely(delta.tv64 >= tick_period.tv64)) { | |
71 | s64 incr = ktime_to_ns(tick_period); | |
72 | ||
73 | ticks = ktime_divns(delta, incr); | |
74 | ||
75 | last_jiffies_update = ktime_add_ns(last_jiffies_update, | |
76 | incr * ticks); | |
77 | } | |
78 | do_timer(++ticks); | |
49d670fb TG |
79 | |
80 | /* Keep the tick_next_period variable up to date */ | |
81 | tick_next_period = ktime_add(last_jiffies_update, tick_period); | |
79bf2bb3 TG |
82 | } |
83 | write_sequnlock(&xtime_lock); | |
84 | } | |
85 | ||
86 | /* | |
87 | * Initialize and return retrieve the jiffies update. | |
88 | */ | |
89 | static ktime_t tick_init_jiffy_update(void) | |
90 | { | |
91 | ktime_t period; | |
92 | ||
93 | write_seqlock(&xtime_lock); | |
94 | /* Did we start the jiffies update yet ? */ | |
95 | if (last_jiffies_update.tv64 == 0) | |
96 | last_jiffies_update = tick_next_period; | |
97 | period = last_jiffies_update; | |
98 | write_sequnlock(&xtime_lock); | |
99 | return period; | |
100 | } | |
101 | ||
102 | /* | |
103 | * NOHZ - aka dynamic tick functionality | |
104 | */ | |
105 | #ifdef CONFIG_NO_HZ | |
106 | /* | |
107 | * NO HZ enabled ? | |
108 | */ | |
109 | static int tick_nohz_enabled __read_mostly = 1; | |
110 | ||
111 | /* | |
112 | * Enable / Disable tickless mode | |
113 | */ | |
114 | static int __init setup_tick_nohz(char *str) | |
115 | { | |
116 | if (!strcmp(str, "off")) | |
117 | tick_nohz_enabled = 0; | |
118 | else if (!strcmp(str, "on")) | |
119 | tick_nohz_enabled = 1; | |
120 | else | |
121 | return 0; | |
122 | return 1; | |
123 | } | |
124 | ||
125 | __setup("nohz=", setup_tick_nohz); | |
126 | ||
127 | /** | |
128 | * tick_nohz_update_jiffies - update jiffies when idle was interrupted | |
129 | * | |
130 | * Called from interrupt entry when the CPU was idle | |
131 | * | |
132 | * In case the sched_tick was stopped on this CPU, we have to check if jiffies | |
133 | * must be updated. Otherwise an interrupt handler could use a stale jiffy | |
134 | * value. We do this unconditionally on any cpu, as we don't know whether the | |
135 | * cpu, which has the update task assigned is in a long sleep. | |
136 | */ | |
eed3b9cf | 137 | static void tick_nohz_update_jiffies(ktime_t now) |
79bf2bb3 TG |
138 | { |
139 | int cpu = smp_processor_id(); | |
140 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
141 | unsigned long flags; | |
79bf2bb3 | 142 | |
6a7b3dc3 | 143 | cpumask_clear_cpu(cpu, nohz_cpu_mask); |
5df7fa1c | 144 | ts->idle_waketime = now; |
79bf2bb3 TG |
145 | |
146 | local_irq_save(flags); | |
147 | tick_do_update_jiffies64(now); | |
148 | local_irq_restore(flags); | |
02ff3755 IM |
149 | |
150 | touch_softlockup_watchdog(); | |
79bf2bb3 TG |
151 | } |
152 | ||
595aac48 AV |
153 | /* |
154 | * Updates the per cpu time idle statistics counters | |
155 | */ | |
8d63bf94 | 156 | static void |
8c215bd3 | 157 | update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time) |
6378ddb5 | 158 | { |
eed3b9cf | 159 | ktime_t delta; |
6378ddb5 | 160 | |
595aac48 AV |
161 | if (ts->idle_active) { |
162 | delta = ktime_sub(now, ts->idle_entrytime); | |
163 | ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); | |
8c215bd3 | 164 | if (nr_iowait_cpu(cpu) > 0) |
0224cf4c | 165 | ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta); |
8c7b09f4 | 166 | ts->idle_entrytime = now; |
595aac48 | 167 | } |
8d63bf94 | 168 | |
e0e37c20 | 169 | if (last_update_time) |
8d63bf94 AV |
170 | *last_update_time = ktime_to_us(now); |
171 | ||
595aac48 AV |
172 | } |
173 | ||
174 | static void tick_nohz_stop_idle(int cpu, ktime_t now) | |
175 | { | |
176 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
177 | ||
8c215bd3 | 178 | update_ts_time_stats(cpu, ts, now, NULL); |
eed3b9cf | 179 | ts->idle_active = 0; |
56c7426b | 180 | |
eed3b9cf | 181 | sched_clock_idle_wakeup_event(0); |
6378ddb5 VP |
182 | } |
183 | ||
8c215bd3 | 184 | static ktime_t tick_nohz_start_idle(int cpu, struct tick_sched *ts) |
6378ddb5 | 185 | { |
595aac48 | 186 | ktime_t now; |
6378ddb5 VP |
187 | |
188 | now = ktime_get(); | |
595aac48 | 189 | |
8c215bd3 | 190 | update_ts_time_stats(cpu, ts, now, NULL); |
595aac48 | 191 | |
6378ddb5 VP |
192 | ts->idle_entrytime = now; |
193 | ts->idle_active = 1; | |
56c7426b | 194 | sched_clock_idle_sleep_event(); |
6378ddb5 VP |
195 | return now; |
196 | } | |
197 | ||
b1f724c3 AV |
198 | /** |
199 | * get_cpu_idle_time_us - get the total idle time of a cpu | |
200 | * @cpu: CPU number to query | |
201 | * @last_update_time: variable to store update time in | |
202 | * | |
203 | * Return the cummulative idle time (since boot) for a given | |
204 | * CPU, in microseconds. The idle time returned includes | |
205 | * the iowait time (unlike what "top" and co report). | |
206 | * | |
207 | * This time is measured via accounting rather than sampling, | |
208 | * and is as accurate as ktime_get() is. | |
209 | * | |
210 | * This function returns -1 if NOHZ is not enabled. | |
211 | */ | |
6378ddb5 VP |
212 | u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) |
213 | { | |
214 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
215 | ||
8083e4ad | 216 | if (!tick_nohz_enabled) |
217 | return -1; | |
218 | ||
8c215bd3 | 219 | update_ts_time_stats(cpu, ts, ktime_get(), last_update_time); |
8083e4ad | 220 | |
6378ddb5 VP |
221 | return ktime_to_us(ts->idle_sleeptime); |
222 | } | |
8083e4ad | 223 | EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); |
6378ddb5 | 224 | |
0224cf4c AV |
225 | /* |
226 | * get_cpu_iowait_time_us - get the total iowait time of a cpu | |
227 | * @cpu: CPU number to query | |
228 | * @last_update_time: variable to store update time in | |
229 | * | |
230 | * Return the cummulative iowait time (since boot) for a given | |
231 | * CPU, in microseconds. | |
232 | * | |
233 | * This time is measured via accounting rather than sampling, | |
234 | * and is as accurate as ktime_get() is. | |
235 | * | |
236 | * This function returns -1 if NOHZ is not enabled. | |
237 | */ | |
238 | u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time) | |
239 | { | |
240 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
241 | ||
242 | if (!tick_nohz_enabled) | |
243 | return -1; | |
244 | ||
8c215bd3 | 245 | update_ts_time_stats(cpu, ts, ktime_get(), last_update_time); |
0224cf4c AV |
246 | |
247 | return ktime_to_us(ts->iowait_sleeptime); | |
248 | } | |
249 | EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us); | |
250 | ||
79bf2bb3 TG |
251 | /** |
252 | * tick_nohz_stop_sched_tick - stop the idle tick from the idle task | |
253 | * | |
254 | * When the next event is more than a tick into the future, stop the idle tick | |
255 | * Called either from the idle loop or from irq_exit() when an idle period was | |
256 | * just interrupted by an interrupt which did not cause a reschedule. | |
257 | */ | |
b8f8c3cf | 258 | void tick_nohz_stop_sched_tick(int inidle) |
79bf2bb3 TG |
259 | { |
260 | unsigned long seq, last_jiffies, next_jiffies, delta_jiffies, flags; | |
261 | struct tick_sched *ts; | |
6378ddb5 | 262 | ktime_t last_update, expires, now; |
4f86d3a8 | 263 | struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev; |
98962465 | 264 | u64 time_delta; |
79bf2bb3 TG |
265 | int cpu; |
266 | ||
267 | local_irq_save(flags); | |
268 | ||
269 | cpu = smp_processor_id(); | |
270 | ts = &per_cpu(tick_cpu_sched, cpu); | |
f2e21c96 EN |
271 | |
272 | /* | |
273 | * Call to tick_nohz_start_idle stops the last_update_time from being | |
274 | * updated. Thus, it must not be called in the event we are called from | |
275 | * irq_exit() with the prior state different than idle. | |
276 | */ | |
277 | if (!inidle && !ts->inidle) | |
278 | goto end; | |
279 | ||
fdc6f192 EN |
280 | /* |
281 | * Set ts->inidle unconditionally. Even if the system did not | |
282 | * switch to NOHZ mode the cpu frequency governers rely on the | |
283 | * update of the idle time accounting in tick_nohz_start_idle(). | |
284 | */ | |
285 | ts->inidle = 1; | |
286 | ||
8c215bd3 | 287 | now = tick_nohz_start_idle(cpu, ts); |
79bf2bb3 | 288 | |
5e41d0d6 TG |
289 | /* |
290 | * If this cpu is offline and it is the one which updates | |
291 | * jiffies, then give up the assignment and let it be taken by | |
292 | * the cpu which runs the tick timer next. If we don't drop | |
293 | * this here the jiffies might be stale and do_timer() never | |
294 | * invoked. | |
295 | */ | |
296 | if (unlikely(!cpu_online(cpu))) { | |
297 | if (cpu == tick_do_timer_cpu) | |
6441402b | 298 | tick_do_timer_cpu = TICK_DO_TIMER_NONE; |
5e41d0d6 TG |
299 | } |
300 | ||
79bf2bb3 TG |
301 | if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) |
302 | goto end; | |
303 | ||
304 | if (need_resched()) | |
305 | goto end; | |
306 | ||
fa116ea3 | 307 | if (unlikely(local_softirq_pending() && cpu_online(cpu))) { |
35282316 TG |
308 | static int ratelimit; |
309 | ||
310 | if (ratelimit < 10) { | |
311 | printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n", | |
529eaccd | 312 | (unsigned int) local_softirq_pending()); |
35282316 TG |
313 | ratelimit++; |
314 | } | |
857f3fd7 | 315 | goto end; |
35282316 | 316 | } |
79bf2bb3 | 317 | |
39c0cbe2 MG |
318 | if (nohz_ratelimit(cpu)) |
319 | goto end; | |
320 | ||
79bf2bb3 | 321 | ts->idle_calls++; |
79bf2bb3 TG |
322 | /* Read jiffies and the time when jiffies were updated last */ |
323 | do { | |
324 | seq = read_seqbegin(&xtime_lock); | |
325 | last_update = last_jiffies_update; | |
326 | last_jiffies = jiffies; | |
27185016 | 327 | time_delta = timekeeping_max_deferment(); |
79bf2bb3 TG |
328 | } while (read_seqretry(&xtime_lock, seq)); |
329 | ||
3c5d92a0 MS |
330 | if (rcu_needs_cpu(cpu) || printk_needs_cpu(cpu) || |
331 | arch_needs_cpu(cpu)) { | |
332 | next_jiffies = last_jiffies + 1; | |
6ba9b346 | 333 | delta_jiffies = 1; |
3c5d92a0 MS |
334 | } else { |
335 | /* Get the next timer wheel timer */ | |
336 | next_jiffies = get_next_timer_interrupt(last_jiffies); | |
337 | delta_jiffies = next_jiffies - last_jiffies; | |
338 | } | |
79bf2bb3 TG |
339 | /* |
340 | * Do not stop the tick, if we are only one off | |
341 | * or if the cpu is required for rcu | |
342 | */ | |
6ba9b346 | 343 | if (!ts->tick_stopped && delta_jiffies == 1) |
79bf2bb3 TG |
344 | goto out; |
345 | ||
346 | /* Schedule the tick, if we are at least one jiffie off */ | |
347 | if ((long)delta_jiffies >= 1) { | |
348 | ||
00147449 WR |
349 | /* |
350 | * If this cpu is the one which updates jiffies, then | |
351 | * give up the assignment and let it be taken by the | |
352 | * cpu which runs the tick timer next, which might be | |
353 | * this cpu as well. If we don't drop this here the | |
354 | * jiffies might be stale and do_timer() never | |
27185016 TG |
355 | * invoked. Keep track of the fact that it was the one |
356 | * which had the do_timer() duty last. If this cpu is | |
357 | * the one which had the do_timer() duty last, we | |
358 | * limit the sleep time to the timekeeping | |
359 | * max_deferement value which we retrieved | |
360 | * above. Otherwise we can sleep as long as we want. | |
00147449 | 361 | */ |
27185016 | 362 | if (cpu == tick_do_timer_cpu) { |
00147449 | 363 | tick_do_timer_cpu = TICK_DO_TIMER_NONE; |
27185016 TG |
364 | ts->do_timer_last = 1; |
365 | } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { | |
366 | time_delta = KTIME_MAX; | |
367 | ts->do_timer_last = 0; | |
368 | } else if (!ts->do_timer_last) { | |
369 | time_delta = KTIME_MAX; | |
370 | } | |
371 | ||
00147449 | 372 | /* |
98962465 JH |
373 | * calculate the expiry time for the next timer wheel |
374 | * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals | |
375 | * that there is no timer pending or at least extremely | |
376 | * far into the future (12 days for HZ=1000). In this | |
377 | * case we set the expiry to the end of time. | |
378 | */ | |
379 | if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) { | |
380 | /* | |
381 | * Calculate the time delta for the next timer event. | |
382 | * If the time delta exceeds the maximum time delta | |
383 | * permitted by the current clocksource then adjust | |
384 | * the time delta accordingly to ensure the | |
385 | * clocksource does not wrap. | |
386 | */ | |
387 | time_delta = min_t(u64, time_delta, | |
388 | tick_period.tv64 * delta_jiffies); | |
98962465 | 389 | } |
00147449 | 390 | |
27185016 TG |
391 | if (time_delta < KTIME_MAX) |
392 | expires = ktime_add_ns(last_update, time_delta); | |
393 | else | |
394 | expires.tv64 = KTIME_MAX; | |
00147449 | 395 | |
6ba9b346 | 396 | if (delta_jiffies > 1) |
6a7b3dc3 | 397 | cpumask_set_cpu(cpu, nohz_cpu_mask); |
00147449 WR |
398 | |
399 | /* Skip reprogram of event if its not changed */ | |
400 | if (ts->tick_stopped && ktime_equal(expires, dev->next_event)) | |
401 | goto out; | |
402 | ||
79bf2bb3 TG |
403 | /* |
404 | * nohz_stop_sched_tick can be called several times before | |
405 | * the nohz_restart_sched_tick is called. This happens when | |
406 | * interrupts arrive which do not cause a reschedule. In the | |
407 | * first call we save the current tick time, so we can restart | |
408 | * the scheduler tick in nohz_restart_sched_tick. | |
409 | */ | |
410 | if (!ts->tick_stopped) { | |
46cb4b7c SS |
411 | if (select_nohz_load_balancer(1)) { |
412 | /* | |
413 | * sched tick not stopped! | |
414 | */ | |
6a7b3dc3 | 415 | cpumask_clear_cpu(cpu, nohz_cpu_mask); |
46cb4b7c SS |
416 | goto out; |
417 | } | |
418 | ||
cc584b21 | 419 | ts->idle_tick = hrtimer_get_expires(&ts->sched_timer); |
79bf2bb3 TG |
420 | ts->tick_stopped = 1; |
421 | ts->idle_jiffies = last_jiffies; | |
2232c2d8 | 422 | rcu_enter_nohz(); |
79bf2bb3 | 423 | } |
d3ed7824 | 424 | |
eaad084b TG |
425 | ts->idle_sleeps++; |
426 | ||
98962465 JH |
427 | /* Mark expires */ |
428 | ts->idle_expires = expires; | |
429 | ||
eaad084b | 430 | /* |
98962465 JH |
431 | * If the expiration time == KTIME_MAX, then |
432 | * in this case we simply stop the tick timer. | |
eaad084b | 433 | */ |
98962465 | 434 | if (unlikely(expires.tv64 == KTIME_MAX)) { |
eaad084b TG |
435 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) |
436 | hrtimer_cancel(&ts->sched_timer); | |
437 | goto out; | |
438 | } | |
439 | ||
79bf2bb3 TG |
440 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { |
441 | hrtimer_start(&ts->sched_timer, expires, | |
5c333864 | 442 | HRTIMER_MODE_ABS_PINNED); |
79bf2bb3 TG |
443 | /* Check, if the timer was already in the past */ |
444 | if (hrtimer_active(&ts->sched_timer)) | |
445 | goto out; | |
4c9dc641 | 446 | } else if (!tick_program_event(expires, 0)) |
79bf2bb3 TG |
447 | goto out; |
448 | /* | |
449 | * We are past the event already. So we crossed a | |
450 | * jiffie boundary. Update jiffies and raise the | |
451 | * softirq. | |
452 | */ | |
453 | tick_do_update_jiffies64(ktime_get()); | |
6a7b3dc3 | 454 | cpumask_clear_cpu(cpu, nohz_cpu_mask); |
79bf2bb3 TG |
455 | } |
456 | raise_softirq_irqoff(TIMER_SOFTIRQ); | |
457 | out: | |
458 | ts->next_jiffies = next_jiffies; | |
459 | ts->last_jiffies = last_jiffies; | |
4f86d3a8 | 460 | ts->sleep_length = ktime_sub(dev->next_event, now); |
79bf2bb3 TG |
461 | end: |
462 | local_irq_restore(flags); | |
463 | } | |
464 | ||
4f86d3a8 LB |
465 | /** |
466 | * tick_nohz_get_sleep_length - return the length of the current sleep | |
467 | * | |
468 | * Called from power state control code with interrupts disabled | |
469 | */ | |
470 | ktime_t tick_nohz_get_sleep_length(void) | |
471 | { | |
472 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
473 | ||
474 | return ts->sleep_length; | |
475 | } | |
476 | ||
c34bec5a TG |
477 | static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) |
478 | { | |
479 | hrtimer_cancel(&ts->sched_timer); | |
268a3dcf | 480 | hrtimer_set_expires(&ts->sched_timer, ts->idle_tick); |
c34bec5a TG |
481 | |
482 | while (1) { | |
483 | /* Forward the time to expire in the future */ | |
484 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
485 | ||
486 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { | |
268a3dcf | 487 | hrtimer_start_expires(&ts->sched_timer, |
5c333864 | 488 | HRTIMER_MODE_ABS_PINNED); |
c34bec5a TG |
489 | /* Check, if the timer was already in the past */ |
490 | if (hrtimer_active(&ts->sched_timer)) | |
491 | break; | |
492 | } else { | |
268a3dcf TG |
493 | if (!tick_program_event( |
494 | hrtimer_get_expires(&ts->sched_timer), 0)) | |
c34bec5a TG |
495 | break; |
496 | } | |
497 | /* Update jiffies and reread time */ | |
498 | tick_do_update_jiffies64(now); | |
499 | now = ktime_get(); | |
500 | } | |
501 | } | |
502 | ||
79bf2bb3 | 503 | /** |
8dce39c2 | 504 | * tick_nohz_restart_sched_tick - restart the idle tick from the idle task |
79bf2bb3 TG |
505 | * |
506 | * Restart the idle tick when the CPU is woken up from idle | |
507 | */ | |
508 | void tick_nohz_restart_sched_tick(void) | |
509 | { | |
510 | int cpu = smp_processor_id(); | |
511 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
79741dd3 | 512 | #ifndef CONFIG_VIRT_CPU_ACCOUNTING |
79bf2bb3 | 513 | unsigned long ticks; |
79741dd3 | 514 | #endif |
6378ddb5 | 515 | ktime_t now; |
79bf2bb3 | 516 | |
6378ddb5 | 517 | local_irq_disable(); |
eed3b9cf MS |
518 | if (ts->idle_active || (ts->inidle && ts->tick_stopped)) |
519 | now = ktime_get(); | |
520 | ||
521 | if (ts->idle_active) | |
522 | tick_nohz_stop_idle(cpu, now); | |
6378ddb5 | 523 | |
b8f8c3cf TG |
524 | if (!ts->inidle || !ts->tick_stopped) { |
525 | ts->inidle = 0; | |
6378ddb5 | 526 | local_irq_enable(); |
79bf2bb3 | 527 | return; |
6378ddb5 | 528 | } |
79bf2bb3 | 529 | |
b8f8c3cf TG |
530 | ts->inidle = 0; |
531 | ||
2232c2d8 SR |
532 | rcu_exit_nohz(); |
533 | ||
79bf2bb3 | 534 | /* Update jiffies first */ |
46cb4b7c | 535 | select_nohz_load_balancer(0); |
79bf2bb3 | 536 | tick_do_update_jiffies64(now); |
6a7b3dc3 | 537 | cpumask_clear_cpu(cpu, nohz_cpu_mask); |
79bf2bb3 | 538 | |
79741dd3 | 539 | #ifndef CONFIG_VIRT_CPU_ACCOUNTING |
79bf2bb3 TG |
540 | /* |
541 | * We stopped the tick in idle. Update process times would miss the | |
542 | * time we slept as update_process_times does only a 1 tick | |
543 | * accounting. Enforce that this is accounted to idle ! | |
544 | */ | |
545 | ticks = jiffies - ts->idle_jiffies; | |
546 | /* | |
547 | * We might be one off. Do not randomly account a huge number of ticks! | |
548 | */ | |
79741dd3 MS |
549 | if (ticks && ticks < LONG_MAX) |
550 | account_idle_ticks(ticks); | |
551 | #endif | |
79bf2bb3 | 552 | |
126e01bf | 553 | touch_softlockup_watchdog(); |
79bf2bb3 TG |
554 | /* |
555 | * Cancel the scheduled timer and restore the tick | |
556 | */ | |
557 | ts->tick_stopped = 0; | |
5df7fa1c | 558 | ts->idle_exittime = now; |
79bf2bb3 | 559 | |
c34bec5a | 560 | tick_nohz_restart(ts, now); |
79bf2bb3 | 561 | |
79bf2bb3 TG |
562 | local_irq_enable(); |
563 | } | |
564 | ||
565 | static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now) | |
566 | { | |
567 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
cc584b21 | 568 | return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0); |
79bf2bb3 TG |
569 | } |
570 | ||
571 | /* | |
572 | * The nohz low res interrupt handler | |
573 | */ | |
574 | static void tick_nohz_handler(struct clock_event_device *dev) | |
575 | { | |
576 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
577 | struct pt_regs *regs = get_irq_regs(); | |
d3ed7824 | 578 | int cpu = smp_processor_id(); |
79bf2bb3 TG |
579 | ktime_t now = ktime_get(); |
580 | ||
581 | dev->next_event.tv64 = KTIME_MAX; | |
582 | ||
d3ed7824 TG |
583 | /* |
584 | * Check if the do_timer duty was dropped. We don't care about | |
585 | * concurrency: This happens only when the cpu in charge went | |
586 | * into a long sleep. If two cpus happen to assign themself to | |
587 | * this duty, then the jiffies update is still serialized by | |
588 | * xtime_lock. | |
589 | */ | |
6441402b | 590 | if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) |
d3ed7824 TG |
591 | tick_do_timer_cpu = cpu; |
592 | ||
79bf2bb3 | 593 | /* Check, if the jiffies need an update */ |
d3ed7824 TG |
594 | if (tick_do_timer_cpu == cpu) |
595 | tick_do_update_jiffies64(now); | |
79bf2bb3 TG |
596 | |
597 | /* | |
598 | * When we are idle and the tick is stopped, we have to touch | |
599 | * the watchdog as we might not schedule for a really long | |
600 | * time. This happens on complete idle SMP systems while | |
601 | * waiting on the login prompt. We also increment the "start | |
602 | * of idle" jiffy stamp so the idle accounting adjustment we | |
603 | * do when we go busy again does not account too much ticks. | |
604 | */ | |
605 | if (ts->tick_stopped) { | |
606 | touch_softlockup_watchdog(); | |
607 | ts->idle_jiffies++; | |
608 | } | |
609 | ||
610 | update_process_times(user_mode(regs)); | |
611 | profile_tick(CPU_PROFILING); | |
612 | ||
79bf2bb3 TG |
613 | while (tick_nohz_reprogram(ts, now)) { |
614 | now = ktime_get(); | |
615 | tick_do_update_jiffies64(now); | |
616 | } | |
617 | } | |
618 | ||
619 | /** | |
620 | * tick_nohz_switch_to_nohz - switch to nohz mode | |
621 | */ | |
622 | static void tick_nohz_switch_to_nohz(void) | |
623 | { | |
624 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
625 | ktime_t next; | |
626 | ||
627 | if (!tick_nohz_enabled) | |
628 | return; | |
629 | ||
630 | local_irq_disable(); | |
631 | if (tick_switch_to_oneshot(tick_nohz_handler)) { | |
632 | local_irq_enable(); | |
633 | return; | |
634 | } | |
635 | ||
636 | ts->nohz_mode = NOHZ_MODE_LOWRES; | |
637 | ||
638 | /* | |
639 | * Recycle the hrtimer in ts, so we can share the | |
640 | * hrtimer_forward with the highres code. | |
641 | */ | |
642 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); | |
643 | /* Get the next period */ | |
644 | next = tick_init_jiffy_update(); | |
645 | ||
646 | for (;;) { | |
cc584b21 | 647 | hrtimer_set_expires(&ts->sched_timer, next); |
79bf2bb3 TG |
648 | if (!tick_program_event(next, 0)) |
649 | break; | |
650 | next = ktime_add(next, tick_period); | |
651 | } | |
652 | local_irq_enable(); | |
653 | ||
654 | printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n", | |
655 | smp_processor_id()); | |
656 | } | |
657 | ||
fb02fbc1 TG |
658 | /* |
659 | * When NOHZ is enabled and the tick is stopped, we need to kick the | |
660 | * tick timer from irq_enter() so that the jiffies update is kept | |
661 | * alive during long running softirqs. That's ugly as hell, but | |
662 | * correctness is key even if we need to fix the offending softirq in | |
663 | * the first place. | |
664 | * | |
665 | * Note, this is different to tick_nohz_restart. We just kick the | |
666 | * timer and do not touch the other magic bits which need to be done | |
667 | * when idle is left. | |
668 | */ | |
eed3b9cf | 669 | static void tick_nohz_kick_tick(int cpu, ktime_t now) |
fb02fbc1 | 670 | { |
ae99286b TG |
671 | #if 0 |
672 | /* Switch back to 2.6.27 behaviour */ | |
673 | ||
fb02fbc1 | 674 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); |
eed3b9cf | 675 | ktime_t delta; |
fb02fbc1 | 676 | |
c4bd822e TG |
677 | /* |
678 | * Do not touch the tick device, when the next expiry is either | |
679 | * already reached or less/equal than the tick period. | |
680 | */ | |
268a3dcf | 681 | delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now); |
c4bd822e TG |
682 | if (delta.tv64 <= tick_period.tv64) |
683 | return; | |
684 | ||
685 | tick_nohz_restart(ts, now); | |
ae99286b | 686 | #endif |
fb02fbc1 TG |
687 | } |
688 | ||
eed3b9cf MS |
689 | static inline void tick_check_nohz(int cpu) |
690 | { | |
691 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
692 | ktime_t now; | |
693 | ||
694 | if (!ts->idle_active && !ts->tick_stopped) | |
695 | return; | |
696 | now = ktime_get(); | |
697 | if (ts->idle_active) | |
698 | tick_nohz_stop_idle(cpu, now); | |
699 | if (ts->tick_stopped) { | |
700 | tick_nohz_update_jiffies(now); | |
701 | tick_nohz_kick_tick(cpu, now); | |
702 | } | |
703 | } | |
704 | ||
79bf2bb3 TG |
705 | #else |
706 | ||
707 | static inline void tick_nohz_switch_to_nohz(void) { } | |
eed3b9cf | 708 | static inline void tick_check_nohz(int cpu) { } |
79bf2bb3 TG |
709 | |
710 | #endif /* NO_HZ */ | |
711 | ||
719254fa TG |
712 | /* |
713 | * Called from irq_enter to notify about the possible interruption of idle() | |
714 | */ | |
715 | void tick_check_idle(int cpu) | |
716 | { | |
fb02fbc1 | 717 | tick_check_oneshot_broadcast(cpu); |
eed3b9cf | 718 | tick_check_nohz(cpu); |
719254fa TG |
719 | } |
720 | ||
79bf2bb3 TG |
721 | /* |
722 | * High resolution timer specific code | |
723 | */ | |
724 | #ifdef CONFIG_HIGH_RES_TIMERS | |
725 | /* | |
4c9dc641 | 726 | * We rearm the timer until we get disabled by the idle code. |
79bf2bb3 TG |
727 | * Called with interrupts disabled and timer->base->cpu_base->lock held. |
728 | */ | |
729 | static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) | |
730 | { | |
731 | struct tick_sched *ts = | |
732 | container_of(timer, struct tick_sched, sched_timer); | |
79bf2bb3 TG |
733 | struct pt_regs *regs = get_irq_regs(); |
734 | ktime_t now = ktime_get(); | |
d3ed7824 TG |
735 | int cpu = smp_processor_id(); |
736 | ||
737 | #ifdef CONFIG_NO_HZ | |
738 | /* | |
739 | * Check if the do_timer duty was dropped. We don't care about | |
740 | * concurrency: This happens only when the cpu in charge went | |
741 | * into a long sleep. If two cpus happen to assign themself to | |
742 | * this duty, then the jiffies update is still serialized by | |
743 | * xtime_lock. | |
744 | */ | |
6441402b | 745 | if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) |
d3ed7824 TG |
746 | tick_do_timer_cpu = cpu; |
747 | #endif | |
79bf2bb3 TG |
748 | |
749 | /* Check, if the jiffies need an update */ | |
d3ed7824 TG |
750 | if (tick_do_timer_cpu == cpu) |
751 | tick_do_update_jiffies64(now); | |
79bf2bb3 TG |
752 | |
753 | /* | |
754 | * Do not call, when we are not in irq context and have | |
755 | * no valid regs pointer | |
756 | */ | |
757 | if (regs) { | |
758 | /* | |
759 | * When we are idle and the tick is stopped, we have to touch | |
760 | * the watchdog as we might not schedule for a really long | |
761 | * time. This happens on complete idle SMP systems while | |
762 | * waiting on the login prompt. We also increment the "start of | |
763 | * idle" jiffy stamp so the idle accounting adjustment we do | |
764 | * when we go busy again does not account too much ticks. | |
765 | */ | |
766 | if (ts->tick_stopped) { | |
767 | touch_softlockup_watchdog(); | |
768 | ts->idle_jiffies++; | |
769 | } | |
79bf2bb3 TG |
770 | update_process_times(user_mode(regs)); |
771 | profile_tick(CPU_PROFILING); | |
79bf2bb3 TG |
772 | } |
773 | ||
79bf2bb3 TG |
774 | hrtimer_forward(timer, now, tick_period); |
775 | ||
776 | return HRTIMER_RESTART; | |
777 | } | |
778 | ||
779 | /** | |
780 | * tick_setup_sched_timer - setup the tick emulation timer | |
781 | */ | |
782 | void tick_setup_sched_timer(void) | |
783 | { | |
784 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
785 | ktime_t now = ktime_get(); | |
3704540b | 786 | u64 offset; |
79bf2bb3 TG |
787 | |
788 | /* | |
789 | * Emulate tick processing via per-CPU hrtimers: | |
790 | */ | |
791 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); | |
792 | ts->sched_timer.function = tick_sched_timer; | |
79bf2bb3 | 793 | |
3704540b | 794 | /* Get the next period (per cpu) */ |
cc584b21 | 795 | hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); |
3704540b | 796 | offset = ktime_to_ns(tick_period) >> 1; |
b2d9323d | 797 | do_div(offset, num_possible_cpus()); |
3704540b | 798 | offset *= smp_processor_id(); |
cc584b21 | 799 | hrtimer_add_expires_ns(&ts->sched_timer, offset); |
79bf2bb3 TG |
800 | |
801 | for (;;) { | |
802 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
5c333864 AB |
803 | hrtimer_start_expires(&ts->sched_timer, |
804 | HRTIMER_MODE_ABS_PINNED); | |
79bf2bb3 TG |
805 | /* Check, if the timer was already in the past */ |
806 | if (hrtimer_active(&ts->sched_timer)) | |
807 | break; | |
808 | now = ktime_get(); | |
809 | } | |
810 | ||
811 | #ifdef CONFIG_NO_HZ | |
812 | if (tick_nohz_enabled) | |
813 | ts->nohz_mode = NOHZ_MODE_HIGHRES; | |
814 | #endif | |
815 | } | |
3c4fbe5e | 816 | #endif /* HIGH_RES_TIMERS */ |
79bf2bb3 | 817 | |
3c4fbe5e | 818 | #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS |
79bf2bb3 TG |
819 | void tick_cancel_sched_timer(int cpu) |
820 | { | |
821 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
822 | ||
3c4fbe5e | 823 | # ifdef CONFIG_HIGH_RES_TIMERS |
79bf2bb3 TG |
824 | if (ts->sched_timer.base) |
825 | hrtimer_cancel(&ts->sched_timer); | |
3c4fbe5e | 826 | # endif |
a7901766 | 827 | |
79bf2bb3 TG |
828 | ts->nohz_mode = NOHZ_MODE_INACTIVE; |
829 | } | |
3c4fbe5e | 830 | #endif |
79bf2bb3 TG |
831 | |
832 | /** | |
833 | * Async notification about clocksource changes | |
834 | */ | |
835 | void tick_clock_notify(void) | |
836 | { | |
837 | int cpu; | |
838 | ||
839 | for_each_possible_cpu(cpu) | |
840 | set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); | |
841 | } | |
842 | ||
843 | /* | |
844 | * Async notification about clock event changes | |
845 | */ | |
846 | void tick_oneshot_notify(void) | |
847 | { | |
848 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
849 | ||
850 | set_bit(0, &ts->check_clocks); | |
851 | } | |
852 | ||
853 | /** | |
854 | * Check, if a change happened, which makes oneshot possible. | |
855 | * | |
856 | * Called cyclic from the hrtimer softirq (driven by the timer | |
857 | * softirq) allow_nohz signals, that we can switch into low-res nohz | |
858 | * mode, because high resolution timers are disabled (either compile | |
859 | * or runtime). | |
860 | */ | |
861 | int tick_check_oneshot_change(int allow_nohz) | |
862 | { | |
863 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
864 | ||
865 | if (!test_and_clear_bit(0, &ts->check_clocks)) | |
866 | return 0; | |
867 | ||
868 | if (ts->nohz_mode != NOHZ_MODE_INACTIVE) | |
869 | return 0; | |
870 | ||
cf4fc6cb | 871 | if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) |
79bf2bb3 TG |
872 | return 0; |
873 | ||
874 | if (!allow_nohz) | |
875 | return 1; | |
876 | ||
877 | tick_nohz_switch_to_nohz(); | |
878 | return 0; | |
879 | } |