]>
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 AV |
156 | static void |
157 | update_ts_time_stats(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); | |
8c7b09f4 | 164 | ts->idle_entrytime = now; |
595aac48 | 165 | } |
8d63bf94 | 166 | |
e0e37c20 | 167 | if (last_update_time) |
8d63bf94 AV |
168 | *last_update_time = ktime_to_us(now); |
169 | ||
595aac48 AV |
170 | } |
171 | ||
172 | static void tick_nohz_stop_idle(int cpu, ktime_t now) | |
173 | { | |
174 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
175 | ||
8d63bf94 | 176 | update_ts_time_stats(ts, now, NULL); |
eed3b9cf | 177 | ts->idle_active = 0; |
56c7426b | 178 | |
eed3b9cf | 179 | sched_clock_idle_wakeup_event(0); |
6378ddb5 VP |
180 | } |
181 | ||
903b8a8d | 182 | static ktime_t tick_nohz_start_idle(struct tick_sched *ts) |
6378ddb5 | 183 | { |
595aac48 | 184 | ktime_t now; |
6378ddb5 VP |
185 | |
186 | now = ktime_get(); | |
595aac48 | 187 | |
8d63bf94 | 188 | update_ts_time_stats(ts, now, NULL); |
595aac48 | 189 | |
6378ddb5 VP |
190 | ts->idle_entrytime = now; |
191 | ts->idle_active = 1; | |
56c7426b | 192 | sched_clock_idle_sleep_event(); |
6378ddb5 VP |
193 | return now; |
194 | } | |
195 | ||
b1f724c3 AV |
196 | /** |
197 | * get_cpu_idle_time_us - get the total idle time of a cpu | |
198 | * @cpu: CPU number to query | |
199 | * @last_update_time: variable to store update time in | |
200 | * | |
201 | * Return the cummulative idle time (since boot) for a given | |
202 | * CPU, in microseconds. The idle time returned includes | |
203 | * the iowait time (unlike what "top" and co report). | |
204 | * | |
205 | * This time is measured via accounting rather than sampling, | |
206 | * and is as accurate as ktime_get() is. | |
207 | * | |
208 | * This function returns -1 if NOHZ is not enabled. | |
209 | */ | |
6378ddb5 VP |
210 | u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) |
211 | { | |
212 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
213 | ||
8083e4ad | 214 | if (!tick_nohz_enabled) |
215 | return -1; | |
216 | ||
8d63bf94 | 217 | update_ts_time_stats(ts, ktime_get(), last_update_time); |
8083e4ad | 218 | |
6378ddb5 VP |
219 | return ktime_to_us(ts->idle_sleeptime); |
220 | } | |
8083e4ad | 221 | EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); |
6378ddb5 | 222 | |
79bf2bb3 TG |
223 | /** |
224 | * tick_nohz_stop_sched_tick - stop the idle tick from the idle task | |
225 | * | |
226 | * When the next event is more than a tick into the future, stop the idle tick | |
227 | * Called either from the idle loop or from irq_exit() when an idle period was | |
228 | * just interrupted by an interrupt which did not cause a reschedule. | |
229 | */ | |
b8f8c3cf | 230 | void tick_nohz_stop_sched_tick(int inidle) |
79bf2bb3 TG |
231 | { |
232 | unsigned long seq, last_jiffies, next_jiffies, delta_jiffies, flags; | |
233 | struct tick_sched *ts; | |
6378ddb5 | 234 | ktime_t last_update, expires, now; |
4f86d3a8 | 235 | struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev; |
98962465 | 236 | u64 time_delta; |
79bf2bb3 TG |
237 | int cpu; |
238 | ||
239 | local_irq_save(flags); | |
240 | ||
241 | cpu = smp_processor_id(); | |
242 | ts = &per_cpu(tick_cpu_sched, cpu); | |
f2e21c96 EN |
243 | |
244 | /* | |
245 | * Call to tick_nohz_start_idle stops the last_update_time from being | |
246 | * updated. Thus, it must not be called in the event we are called from | |
247 | * irq_exit() with the prior state different than idle. | |
248 | */ | |
249 | if (!inidle && !ts->inidle) | |
250 | goto end; | |
251 | ||
fdc6f192 EN |
252 | /* |
253 | * Set ts->inidle unconditionally. Even if the system did not | |
254 | * switch to NOHZ mode the cpu frequency governers rely on the | |
255 | * update of the idle time accounting in tick_nohz_start_idle(). | |
256 | */ | |
257 | ts->inidle = 1; | |
258 | ||
903b8a8d | 259 | now = tick_nohz_start_idle(ts); |
79bf2bb3 | 260 | |
5e41d0d6 TG |
261 | /* |
262 | * If this cpu is offline and it is the one which updates | |
263 | * jiffies, then give up the assignment and let it be taken by | |
264 | * the cpu which runs the tick timer next. If we don't drop | |
265 | * this here the jiffies might be stale and do_timer() never | |
266 | * invoked. | |
267 | */ | |
268 | if (unlikely(!cpu_online(cpu))) { | |
269 | if (cpu == tick_do_timer_cpu) | |
6441402b | 270 | tick_do_timer_cpu = TICK_DO_TIMER_NONE; |
5e41d0d6 TG |
271 | } |
272 | ||
79bf2bb3 TG |
273 | if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) |
274 | goto end; | |
275 | ||
276 | if (need_resched()) | |
277 | goto end; | |
278 | ||
fa116ea3 | 279 | if (unlikely(local_softirq_pending() && cpu_online(cpu))) { |
35282316 TG |
280 | static int ratelimit; |
281 | ||
282 | if (ratelimit < 10) { | |
283 | printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n", | |
529eaccd | 284 | (unsigned int) local_softirq_pending()); |
35282316 TG |
285 | ratelimit++; |
286 | } | |
857f3fd7 | 287 | goto end; |
35282316 | 288 | } |
79bf2bb3 | 289 | |
39c0cbe2 MG |
290 | if (nohz_ratelimit(cpu)) |
291 | goto end; | |
292 | ||
79bf2bb3 | 293 | ts->idle_calls++; |
79bf2bb3 TG |
294 | /* Read jiffies and the time when jiffies were updated last */ |
295 | do { | |
296 | seq = read_seqbegin(&xtime_lock); | |
297 | last_update = last_jiffies_update; | |
298 | last_jiffies = jiffies; | |
27185016 | 299 | time_delta = timekeeping_max_deferment(); |
79bf2bb3 TG |
300 | } while (read_seqretry(&xtime_lock, seq)); |
301 | ||
3c5d92a0 MS |
302 | if (rcu_needs_cpu(cpu) || printk_needs_cpu(cpu) || |
303 | arch_needs_cpu(cpu)) { | |
304 | next_jiffies = last_jiffies + 1; | |
6ba9b346 | 305 | delta_jiffies = 1; |
3c5d92a0 MS |
306 | } else { |
307 | /* Get the next timer wheel timer */ | |
308 | next_jiffies = get_next_timer_interrupt(last_jiffies); | |
309 | delta_jiffies = next_jiffies - last_jiffies; | |
310 | } | |
79bf2bb3 TG |
311 | /* |
312 | * Do not stop the tick, if we are only one off | |
313 | * or if the cpu is required for rcu | |
314 | */ | |
6ba9b346 | 315 | if (!ts->tick_stopped && delta_jiffies == 1) |
79bf2bb3 TG |
316 | goto out; |
317 | ||
318 | /* Schedule the tick, if we are at least one jiffie off */ | |
319 | if ((long)delta_jiffies >= 1) { | |
320 | ||
00147449 WR |
321 | /* |
322 | * If this cpu is the one which updates jiffies, then | |
323 | * give up the assignment and let it be taken by the | |
324 | * cpu which runs the tick timer next, which might be | |
325 | * this cpu as well. If we don't drop this here the | |
326 | * jiffies might be stale and do_timer() never | |
27185016 TG |
327 | * invoked. Keep track of the fact that it was the one |
328 | * which had the do_timer() duty last. If this cpu is | |
329 | * the one which had the do_timer() duty last, we | |
330 | * limit the sleep time to the timekeeping | |
331 | * max_deferement value which we retrieved | |
332 | * above. Otherwise we can sleep as long as we want. | |
00147449 | 333 | */ |
27185016 | 334 | if (cpu == tick_do_timer_cpu) { |
00147449 | 335 | tick_do_timer_cpu = TICK_DO_TIMER_NONE; |
27185016 TG |
336 | ts->do_timer_last = 1; |
337 | } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { | |
338 | time_delta = KTIME_MAX; | |
339 | ts->do_timer_last = 0; | |
340 | } else if (!ts->do_timer_last) { | |
341 | time_delta = KTIME_MAX; | |
342 | } | |
343 | ||
00147449 | 344 | /* |
98962465 JH |
345 | * calculate the expiry time for the next timer wheel |
346 | * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals | |
347 | * that there is no timer pending or at least extremely | |
348 | * far into the future (12 days for HZ=1000). In this | |
349 | * case we set the expiry to the end of time. | |
350 | */ | |
351 | if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) { | |
352 | /* | |
353 | * Calculate the time delta for the next timer event. | |
354 | * If the time delta exceeds the maximum time delta | |
355 | * permitted by the current clocksource then adjust | |
356 | * the time delta accordingly to ensure the | |
357 | * clocksource does not wrap. | |
358 | */ | |
359 | time_delta = min_t(u64, time_delta, | |
360 | tick_period.tv64 * delta_jiffies); | |
98962465 | 361 | } |
00147449 | 362 | |
27185016 TG |
363 | if (time_delta < KTIME_MAX) |
364 | expires = ktime_add_ns(last_update, time_delta); | |
365 | else | |
366 | expires.tv64 = KTIME_MAX; | |
00147449 | 367 | |
6ba9b346 | 368 | if (delta_jiffies > 1) |
6a7b3dc3 | 369 | cpumask_set_cpu(cpu, nohz_cpu_mask); |
00147449 WR |
370 | |
371 | /* Skip reprogram of event if its not changed */ | |
372 | if (ts->tick_stopped && ktime_equal(expires, dev->next_event)) | |
373 | goto out; | |
374 | ||
79bf2bb3 TG |
375 | /* |
376 | * nohz_stop_sched_tick can be called several times before | |
377 | * the nohz_restart_sched_tick is called. This happens when | |
378 | * interrupts arrive which do not cause a reschedule. In the | |
379 | * first call we save the current tick time, so we can restart | |
380 | * the scheduler tick in nohz_restart_sched_tick. | |
381 | */ | |
382 | if (!ts->tick_stopped) { | |
46cb4b7c SS |
383 | if (select_nohz_load_balancer(1)) { |
384 | /* | |
385 | * sched tick not stopped! | |
386 | */ | |
6a7b3dc3 | 387 | cpumask_clear_cpu(cpu, nohz_cpu_mask); |
46cb4b7c SS |
388 | goto out; |
389 | } | |
390 | ||
cc584b21 | 391 | ts->idle_tick = hrtimer_get_expires(&ts->sched_timer); |
79bf2bb3 TG |
392 | ts->tick_stopped = 1; |
393 | ts->idle_jiffies = last_jiffies; | |
2232c2d8 | 394 | rcu_enter_nohz(); |
79bf2bb3 | 395 | } |
d3ed7824 | 396 | |
eaad084b TG |
397 | ts->idle_sleeps++; |
398 | ||
98962465 JH |
399 | /* Mark expires */ |
400 | ts->idle_expires = expires; | |
401 | ||
eaad084b | 402 | /* |
98962465 JH |
403 | * If the expiration time == KTIME_MAX, then |
404 | * in this case we simply stop the tick timer. | |
eaad084b | 405 | */ |
98962465 | 406 | if (unlikely(expires.tv64 == KTIME_MAX)) { |
eaad084b TG |
407 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) |
408 | hrtimer_cancel(&ts->sched_timer); | |
409 | goto out; | |
410 | } | |
411 | ||
79bf2bb3 TG |
412 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { |
413 | hrtimer_start(&ts->sched_timer, expires, | |
5c333864 | 414 | HRTIMER_MODE_ABS_PINNED); |
79bf2bb3 TG |
415 | /* Check, if the timer was already in the past */ |
416 | if (hrtimer_active(&ts->sched_timer)) | |
417 | goto out; | |
4c9dc641 | 418 | } else if (!tick_program_event(expires, 0)) |
79bf2bb3 TG |
419 | goto out; |
420 | /* | |
421 | * We are past the event already. So we crossed a | |
422 | * jiffie boundary. Update jiffies and raise the | |
423 | * softirq. | |
424 | */ | |
425 | tick_do_update_jiffies64(ktime_get()); | |
6a7b3dc3 | 426 | cpumask_clear_cpu(cpu, nohz_cpu_mask); |
79bf2bb3 TG |
427 | } |
428 | raise_softirq_irqoff(TIMER_SOFTIRQ); | |
429 | out: | |
430 | ts->next_jiffies = next_jiffies; | |
431 | ts->last_jiffies = last_jiffies; | |
4f86d3a8 | 432 | ts->sleep_length = ktime_sub(dev->next_event, now); |
79bf2bb3 TG |
433 | end: |
434 | local_irq_restore(flags); | |
435 | } | |
436 | ||
4f86d3a8 LB |
437 | /** |
438 | * tick_nohz_get_sleep_length - return the length of the current sleep | |
439 | * | |
440 | * Called from power state control code with interrupts disabled | |
441 | */ | |
442 | ktime_t tick_nohz_get_sleep_length(void) | |
443 | { | |
444 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
445 | ||
446 | return ts->sleep_length; | |
447 | } | |
448 | ||
c34bec5a TG |
449 | static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) |
450 | { | |
451 | hrtimer_cancel(&ts->sched_timer); | |
268a3dcf | 452 | hrtimer_set_expires(&ts->sched_timer, ts->idle_tick); |
c34bec5a TG |
453 | |
454 | while (1) { | |
455 | /* Forward the time to expire in the future */ | |
456 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
457 | ||
458 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { | |
268a3dcf | 459 | hrtimer_start_expires(&ts->sched_timer, |
5c333864 | 460 | HRTIMER_MODE_ABS_PINNED); |
c34bec5a TG |
461 | /* Check, if the timer was already in the past */ |
462 | if (hrtimer_active(&ts->sched_timer)) | |
463 | break; | |
464 | } else { | |
268a3dcf TG |
465 | if (!tick_program_event( |
466 | hrtimer_get_expires(&ts->sched_timer), 0)) | |
c34bec5a TG |
467 | break; |
468 | } | |
469 | /* Update jiffies and reread time */ | |
470 | tick_do_update_jiffies64(now); | |
471 | now = ktime_get(); | |
472 | } | |
473 | } | |
474 | ||
79bf2bb3 | 475 | /** |
8dce39c2 | 476 | * tick_nohz_restart_sched_tick - restart the idle tick from the idle task |
79bf2bb3 TG |
477 | * |
478 | * Restart the idle tick when the CPU is woken up from idle | |
479 | */ | |
480 | void tick_nohz_restart_sched_tick(void) | |
481 | { | |
482 | int cpu = smp_processor_id(); | |
483 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
79741dd3 | 484 | #ifndef CONFIG_VIRT_CPU_ACCOUNTING |
79bf2bb3 | 485 | unsigned long ticks; |
79741dd3 | 486 | #endif |
6378ddb5 | 487 | ktime_t now; |
79bf2bb3 | 488 | |
6378ddb5 | 489 | local_irq_disable(); |
eed3b9cf MS |
490 | if (ts->idle_active || (ts->inidle && ts->tick_stopped)) |
491 | now = ktime_get(); | |
492 | ||
493 | if (ts->idle_active) | |
494 | tick_nohz_stop_idle(cpu, now); | |
6378ddb5 | 495 | |
b8f8c3cf TG |
496 | if (!ts->inidle || !ts->tick_stopped) { |
497 | ts->inidle = 0; | |
6378ddb5 | 498 | local_irq_enable(); |
79bf2bb3 | 499 | return; |
6378ddb5 | 500 | } |
79bf2bb3 | 501 | |
b8f8c3cf TG |
502 | ts->inidle = 0; |
503 | ||
2232c2d8 SR |
504 | rcu_exit_nohz(); |
505 | ||
79bf2bb3 | 506 | /* Update jiffies first */ |
46cb4b7c | 507 | select_nohz_load_balancer(0); |
79bf2bb3 | 508 | tick_do_update_jiffies64(now); |
6a7b3dc3 | 509 | cpumask_clear_cpu(cpu, nohz_cpu_mask); |
79bf2bb3 | 510 | |
79741dd3 | 511 | #ifndef CONFIG_VIRT_CPU_ACCOUNTING |
79bf2bb3 TG |
512 | /* |
513 | * We stopped the tick in idle. Update process times would miss the | |
514 | * time we slept as update_process_times does only a 1 tick | |
515 | * accounting. Enforce that this is accounted to idle ! | |
516 | */ | |
517 | ticks = jiffies - ts->idle_jiffies; | |
518 | /* | |
519 | * We might be one off. Do not randomly account a huge number of ticks! | |
520 | */ | |
79741dd3 MS |
521 | if (ticks && ticks < LONG_MAX) |
522 | account_idle_ticks(ticks); | |
523 | #endif | |
79bf2bb3 | 524 | |
126e01bf | 525 | touch_softlockup_watchdog(); |
79bf2bb3 TG |
526 | /* |
527 | * Cancel the scheduled timer and restore the tick | |
528 | */ | |
529 | ts->tick_stopped = 0; | |
5df7fa1c | 530 | ts->idle_exittime = now; |
79bf2bb3 | 531 | |
c34bec5a | 532 | tick_nohz_restart(ts, now); |
79bf2bb3 | 533 | |
79bf2bb3 TG |
534 | local_irq_enable(); |
535 | } | |
536 | ||
537 | static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now) | |
538 | { | |
539 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
cc584b21 | 540 | return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0); |
79bf2bb3 TG |
541 | } |
542 | ||
543 | /* | |
544 | * The nohz low res interrupt handler | |
545 | */ | |
546 | static void tick_nohz_handler(struct clock_event_device *dev) | |
547 | { | |
548 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
549 | struct pt_regs *regs = get_irq_regs(); | |
d3ed7824 | 550 | int cpu = smp_processor_id(); |
79bf2bb3 TG |
551 | ktime_t now = ktime_get(); |
552 | ||
553 | dev->next_event.tv64 = KTIME_MAX; | |
554 | ||
d3ed7824 TG |
555 | /* |
556 | * Check if the do_timer duty was dropped. We don't care about | |
557 | * concurrency: This happens only when the cpu in charge went | |
558 | * into a long sleep. If two cpus happen to assign themself to | |
559 | * this duty, then the jiffies update is still serialized by | |
560 | * xtime_lock. | |
561 | */ | |
6441402b | 562 | if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) |
d3ed7824 TG |
563 | tick_do_timer_cpu = cpu; |
564 | ||
79bf2bb3 | 565 | /* Check, if the jiffies need an update */ |
d3ed7824 TG |
566 | if (tick_do_timer_cpu == cpu) |
567 | tick_do_update_jiffies64(now); | |
79bf2bb3 TG |
568 | |
569 | /* | |
570 | * When we are idle and the tick is stopped, we have to touch | |
571 | * the watchdog as we might not schedule for a really long | |
572 | * time. This happens on complete idle SMP systems while | |
573 | * waiting on the login prompt. We also increment the "start | |
574 | * of idle" jiffy stamp so the idle accounting adjustment we | |
575 | * do when we go busy again does not account too much ticks. | |
576 | */ | |
577 | if (ts->tick_stopped) { | |
578 | touch_softlockup_watchdog(); | |
579 | ts->idle_jiffies++; | |
580 | } | |
581 | ||
582 | update_process_times(user_mode(regs)); | |
583 | profile_tick(CPU_PROFILING); | |
584 | ||
79bf2bb3 TG |
585 | while (tick_nohz_reprogram(ts, now)) { |
586 | now = ktime_get(); | |
587 | tick_do_update_jiffies64(now); | |
588 | } | |
589 | } | |
590 | ||
591 | /** | |
592 | * tick_nohz_switch_to_nohz - switch to nohz mode | |
593 | */ | |
594 | static void tick_nohz_switch_to_nohz(void) | |
595 | { | |
596 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
597 | ktime_t next; | |
598 | ||
599 | if (!tick_nohz_enabled) | |
600 | return; | |
601 | ||
602 | local_irq_disable(); | |
603 | if (tick_switch_to_oneshot(tick_nohz_handler)) { | |
604 | local_irq_enable(); | |
605 | return; | |
606 | } | |
607 | ||
608 | ts->nohz_mode = NOHZ_MODE_LOWRES; | |
609 | ||
610 | /* | |
611 | * Recycle the hrtimer in ts, so we can share the | |
612 | * hrtimer_forward with the highres code. | |
613 | */ | |
614 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); | |
615 | /* Get the next period */ | |
616 | next = tick_init_jiffy_update(); | |
617 | ||
618 | for (;;) { | |
cc584b21 | 619 | hrtimer_set_expires(&ts->sched_timer, next); |
79bf2bb3 TG |
620 | if (!tick_program_event(next, 0)) |
621 | break; | |
622 | next = ktime_add(next, tick_period); | |
623 | } | |
624 | local_irq_enable(); | |
625 | ||
626 | printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n", | |
627 | smp_processor_id()); | |
628 | } | |
629 | ||
fb02fbc1 TG |
630 | /* |
631 | * When NOHZ is enabled and the tick is stopped, we need to kick the | |
632 | * tick timer from irq_enter() so that the jiffies update is kept | |
633 | * alive during long running softirqs. That's ugly as hell, but | |
634 | * correctness is key even if we need to fix the offending softirq in | |
635 | * the first place. | |
636 | * | |
637 | * Note, this is different to tick_nohz_restart. We just kick the | |
638 | * timer and do not touch the other magic bits which need to be done | |
639 | * when idle is left. | |
640 | */ | |
eed3b9cf | 641 | static void tick_nohz_kick_tick(int cpu, ktime_t now) |
fb02fbc1 | 642 | { |
ae99286b TG |
643 | #if 0 |
644 | /* Switch back to 2.6.27 behaviour */ | |
645 | ||
fb02fbc1 | 646 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); |
eed3b9cf | 647 | ktime_t delta; |
fb02fbc1 | 648 | |
c4bd822e TG |
649 | /* |
650 | * Do not touch the tick device, when the next expiry is either | |
651 | * already reached or less/equal than the tick period. | |
652 | */ | |
268a3dcf | 653 | delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now); |
c4bd822e TG |
654 | if (delta.tv64 <= tick_period.tv64) |
655 | return; | |
656 | ||
657 | tick_nohz_restart(ts, now); | |
ae99286b | 658 | #endif |
fb02fbc1 TG |
659 | } |
660 | ||
eed3b9cf MS |
661 | static inline void tick_check_nohz(int cpu) |
662 | { | |
663 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
664 | ktime_t now; | |
665 | ||
666 | if (!ts->idle_active && !ts->tick_stopped) | |
667 | return; | |
668 | now = ktime_get(); | |
669 | if (ts->idle_active) | |
670 | tick_nohz_stop_idle(cpu, now); | |
671 | if (ts->tick_stopped) { | |
672 | tick_nohz_update_jiffies(now); | |
673 | tick_nohz_kick_tick(cpu, now); | |
674 | } | |
675 | } | |
676 | ||
79bf2bb3 TG |
677 | #else |
678 | ||
679 | static inline void tick_nohz_switch_to_nohz(void) { } | |
eed3b9cf | 680 | static inline void tick_check_nohz(int cpu) { } |
79bf2bb3 TG |
681 | |
682 | #endif /* NO_HZ */ | |
683 | ||
719254fa TG |
684 | /* |
685 | * Called from irq_enter to notify about the possible interruption of idle() | |
686 | */ | |
687 | void tick_check_idle(int cpu) | |
688 | { | |
fb02fbc1 | 689 | tick_check_oneshot_broadcast(cpu); |
eed3b9cf | 690 | tick_check_nohz(cpu); |
719254fa TG |
691 | } |
692 | ||
79bf2bb3 TG |
693 | /* |
694 | * High resolution timer specific code | |
695 | */ | |
696 | #ifdef CONFIG_HIGH_RES_TIMERS | |
697 | /* | |
4c9dc641 | 698 | * We rearm the timer until we get disabled by the idle code. |
79bf2bb3 TG |
699 | * Called with interrupts disabled and timer->base->cpu_base->lock held. |
700 | */ | |
701 | static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) | |
702 | { | |
703 | struct tick_sched *ts = | |
704 | container_of(timer, struct tick_sched, sched_timer); | |
79bf2bb3 TG |
705 | struct pt_regs *regs = get_irq_regs(); |
706 | ktime_t now = ktime_get(); | |
d3ed7824 TG |
707 | int cpu = smp_processor_id(); |
708 | ||
709 | #ifdef CONFIG_NO_HZ | |
710 | /* | |
711 | * Check if the do_timer duty was dropped. We don't care about | |
712 | * concurrency: This happens only when the cpu in charge went | |
713 | * into a long sleep. If two cpus happen to assign themself to | |
714 | * this duty, then the jiffies update is still serialized by | |
715 | * xtime_lock. | |
716 | */ | |
6441402b | 717 | if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) |
d3ed7824 TG |
718 | tick_do_timer_cpu = cpu; |
719 | #endif | |
79bf2bb3 TG |
720 | |
721 | /* Check, if the jiffies need an update */ | |
d3ed7824 TG |
722 | if (tick_do_timer_cpu == cpu) |
723 | tick_do_update_jiffies64(now); | |
79bf2bb3 TG |
724 | |
725 | /* | |
726 | * Do not call, when we are not in irq context and have | |
727 | * no valid regs pointer | |
728 | */ | |
729 | if (regs) { | |
730 | /* | |
731 | * When we are idle and the tick is stopped, we have to touch | |
732 | * the watchdog as we might not schedule for a really long | |
733 | * time. This happens on complete idle SMP systems while | |
734 | * waiting on the login prompt. We also increment the "start of | |
735 | * idle" jiffy stamp so the idle accounting adjustment we do | |
736 | * when we go busy again does not account too much ticks. | |
737 | */ | |
738 | if (ts->tick_stopped) { | |
739 | touch_softlockup_watchdog(); | |
740 | ts->idle_jiffies++; | |
741 | } | |
79bf2bb3 TG |
742 | update_process_times(user_mode(regs)); |
743 | profile_tick(CPU_PROFILING); | |
79bf2bb3 TG |
744 | } |
745 | ||
79bf2bb3 TG |
746 | hrtimer_forward(timer, now, tick_period); |
747 | ||
748 | return HRTIMER_RESTART; | |
749 | } | |
750 | ||
751 | /** | |
752 | * tick_setup_sched_timer - setup the tick emulation timer | |
753 | */ | |
754 | void tick_setup_sched_timer(void) | |
755 | { | |
756 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
757 | ktime_t now = ktime_get(); | |
3704540b | 758 | u64 offset; |
79bf2bb3 TG |
759 | |
760 | /* | |
761 | * Emulate tick processing via per-CPU hrtimers: | |
762 | */ | |
763 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); | |
764 | ts->sched_timer.function = tick_sched_timer; | |
79bf2bb3 | 765 | |
3704540b | 766 | /* Get the next period (per cpu) */ |
cc584b21 | 767 | hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); |
3704540b | 768 | offset = ktime_to_ns(tick_period) >> 1; |
b2d9323d | 769 | do_div(offset, num_possible_cpus()); |
3704540b | 770 | offset *= smp_processor_id(); |
cc584b21 | 771 | hrtimer_add_expires_ns(&ts->sched_timer, offset); |
79bf2bb3 TG |
772 | |
773 | for (;;) { | |
774 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
5c333864 AB |
775 | hrtimer_start_expires(&ts->sched_timer, |
776 | HRTIMER_MODE_ABS_PINNED); | |
79bf2bb3 TG |
777 | /* Check, if the timer was already in the past */ |
778 | if (hrtimer_active(&ts->sched_timer)) | |
779 | break; | |
780 | now = ktime_get(); | |
781 | } | |
782 | ||
783 | #ifdef CONFIG_NO_HZ | |
784 | if (tick_nohz_enabled) | |
785 | ts->nohz_mode = NOHZ_MODE_HIGHRES; | |
786 | #endif | |
787 | } | |
3c4fbe5e | 788 | #endif /* HIGH_RES_TIMERS */ |
79bf2bb3 | 789 | |
3c4fbe5e | 790 | #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS |
79bf2bb3 TG |
791 | void tick_cancel_sched_timer(int cpu) |
792 | { | |
793 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
794 | ||
3c4fbe5e | 795 | # ifdef CONFIG_HIGH_RES_TIMERS |
79bf2bb3 TG |
796 | if (ts->sched_timer.base) |
797 | hrtimer_cancel(&ts->sched_timer); | |
3c4fbe5e | 798 | # endif |
a7901766 | 799 | |
79bf2bb3 TG |
800 | ts->nohz_mode = NOHZ_MODE_INACTIVE; |
801 | } | |
3c4fbe5e | 802 | #endif |
79bf2bb3 TG |
803 | |
804 | /** | |
805 | * Async notification about clocksource changes | |
806 | */ | |
807 | void tick_clock_notify(void) | |
808 | { | |
809 | int cpu; | |
810 | ||
811 | for_each_possible_cpu(cpu) | |
812 | set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); | |
813 | } | |
814 | ||
815 | /* | |
816 | * Async notification about clock event changes | |
817 | */ | |
818 | void tick_oneshot_notify(void) | |
819 | { | |
820 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
821 | ||
822 | set_bit(0, &ts->check_clocks); | |
823 | } | |
824 | ||
825 | /** | |
826 | * Check, if a change happened, which makes oneshot possible. | |
827 | * | |
828 | * Called cyclic from the hrtimer softirq (driven by the timer | |
829 | * softirq) allow_nohz signals, that we can switch into low-res nohz | |
830 | * mode, because high resolution timers are disabled (either compile | |
831 | * or runtime). | |
832 | */ | |
833 | int tick_check_oneshot_change(int allow_nohz) | |
834 | { | |
835 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
836 | ||
837 | if (!test_and_clear_bit(0, &ts->check_clocks)) | |
838 | return 0; | |
839 | ||
840 | if (ts->nohz_mode != NOHZ_MODE_INACTIVE) | |
841 | return 0; | |
842 | ||
cf4fc6cb | 843 | if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) |
79bf2bb3 TG |
844 | return 0; |
845 | ||
846 | if (!allow_nohz) | |
847 | return 1; | |
848 | ||
849 | tick_nohz_switch_to_nohz(); | |
850 | return 0; | |
851 | } |