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> |
00b42959 | 23 | #include <linux/irq_work.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 | */ | |
33a5f626 | 32 | DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched); |
79bf2bb3 TG |
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 | */ | |
9d2ad243 | 109 | int tick_nohz_enabled __read_mostly = 1; |
79bf2bb3 TG |
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 | |
5df7fa1c | 143 | ts->idle_waketime = now; |
79bf2bb3 TG |
144 | |
145 | local_irq_save(flags); | |
146 | tick_do_update_jiffies64(now); | |
147 | local_irq_restore(flags); | |
02ff3755 IM |
148 | |
149 | touch_softlockup_watchdog(); | |
79bf2bb3 TG |
150 | } |
151 | ||
595aac48 AV |
152 | /* |
153 | * Updates the per cpu time idle statistics counters | |
154 | */ | |
8d63bf94 | 155 | static void |
8c215bd3 | 156 | update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time) |
6378ddb5 | 157 | { |
eed3b9cf | 158 | ktime_t delta; |
6378ddb5 | 159 | |
595aac48 AV |
160 | if (ts->idle_active) { |
161 | delta = ktime_sub(now, ts->idle_entrytime); | |
8c215bd3 | 162 | if (nr_iowait_cpu(cpu) > 0) |
0224cf4c | 163 | ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta); |
6beea0cd MH |
164 | else |
165 | ts->idle_sleeptime = ktime_add(ts->idle_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 | { |
430ee881 | 186 | ktime_t now = ktime_get(); |
595aac48 | 187 | |
6378ddb5 VP |
188 | ts->idle_entrytime = now; |
189 | ts->idle_active = 1; | |
56c7426b | 190 | sched_clock_idle_sleep_event(); |
6378ddb5 VP |
191 | return now; |
192 | } | |
193 | ||
b1f724c3 AV |
194 | /** |
195 | * get_cpu_idle_time_us - get the total idle time of a cpu | |
196 | * @cpu: CPU number to query | |
09a1d34f MH |
197 | * @last_update_time: variable to store update time in. Do not update |
198 | * counters if NULL. | |
b1f724c3 AV |
199 | * |
200 | * Return the cummulative idle time (since boot) for a given | |
6beea0cd | 201 | * CPU, in microseconds. |
b1f724c3 AV |
202 | * |
203 | * This time is measured via accounting rather than sampling, | |
204 | * and is as accurate as ktime_get() is. | |
205 | * | |
206 | * This function returns -1 if NOHZ is not enabled. | |
207 | */ | |
6378ddb5 VP |
208 | u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) |
209 | { | |
210 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
09a1d34f | 211 | ktime_t now, idle; |
6378ddb5 | 212 | |
8083e4ad | 213 | if (!tick_nohz_enabled) |
214 | return -1; | |
215 | ||
09a1d34f MH |
216 | now = ktime_get(); |
217 | if (last_update_time) { | |
218 | update_ts_time_stats(cpu, ts, now, last_update_time); | |
219 | idle = ts->idle_sleeptime; | |
220 | } else { | |
221 | if (ts->idle_active && !nr_iowait_cpu(cpu)) { | |
222 | ktime_t delta = ktime_sub(now, ts->idle_entrytime); | |
223 | ||
224 | idle = ktime_add(ts->idle_sleeptime, delta); | |
225 | } else { | |
226 | idle = ts->idle_sleeptime; | |
227 | } | |
228 | } | |
229 | ||
230 | return ktime_to_us(idle); | |
8083e4ad | 231 | |
6378ddb5 | 232 | } |
8083e4ad | 233 | EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); |
6378ddb5 | 234 | |
6beea0cd | 235 | /** |
0224cf4c AV |
236 | * get_cpu_iowait_time_us - get the total iowait time of a cpu |
237 | * @cpu: CPU number to query | |
09a1d34f MH |
238 | * @last_update_time: variable to store update time in. Do not update |
239 | * counters if NULL. | |
0224cf4c AV |
240 | * |
241 | * Return the cummulative iowait time (since boot) for a given | |
242 | * CPU, in microseconds. | |
243 | * | |
244 | * This time is measured via accounting rather than sampling, | |
245 | * and is as accurate as ktime_get() is. | |
246 | * | |
247 | * This function returns -1 if NOHZ is not enabled. | |
248 | */ | |
249 | u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time) | |
250 | { | |
251 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
09a1d34f | 252 | ktime_t now, iowait; |
0224cf4c AV |
253 | |
254 | if (!tick_nohz_enabled) | |
255 | return -1; | |
256 | ||
09a1d34f MH |
257 | now = ktime_get(); |
258 | if (last_update_time) { | |
259 | update_ts_time_stats(cpu, ts, now, last_update_time); | |
260 | iowait = ts->iowait_sleeptime; | |
261 | } else { | |
262 | if (ts->idle_active && nr_iowait_cpu(cpu) > 0) { | |
263 | ktime_t delta = ktime_sub(now, ts->idle_entrytime); | |
0224cf4c | 264 | |
09a1d34f MH |
265 | iowait = ktime_add(ts->iowait_sleeptime, delta); |
266 | } else { | |
267 | iowait = ts->iowait_sleeptime; | |
268 | } | |
269 | } | |
0224cf4c | 270 | |
09a1d34f | 271 | return ktime_to_us(iowait); |
0224cf4c AV |
272 | } |
273 | EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us); | |
274 | ||
84bf1bcc FW |
275 | static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts, |
276 | ktime_t now, int cpu) | |
79bf2bb3 | 277 | { |
280f0677 | 278 | unsigned long seq, last_jiffies, next_jiffies, delta_jiffies; |
84bf1bcc | 279 | ktime_t last_update, expires, ret = { .tv64 = 0 }; |
aa9b1630 | 280 | unsigned long rcu_delta_jiffies; |
4f86d3a8 | 281 | struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev; |
98962465 | 282 | u64 time_delta; |
79bf2bb3 | 283 | |
79bf2bb3 TG |
284 | /* Read jiffies and the time when jiffies were updated last */ |
285 | do { | |
286 | seq = read_seqbegin(&xtime_lock); | |
287 | last_update = last_jiffies_update; | |
288 | last_jiffies = jiffies; | |
27185016 | 289 | time_delta = timekeeping_max_deferment(); |
79bf2bb3 TG |
290 | } while (read_seqretry(&xtime_lock, seq)); |
291 | ||
aa9b1630 | 292 | if (rcu_needs_cpu(cpu, &rcu_delta_jiffies) || printk_needs_cpu(cpu) || |
00b42959 | 293 | arch_needs_cpu(cpu) || irq_work_needs_cpu()) { |
3c5d92a0 | 294 | next_jiffies = last_jiffies + 1; |
6ba9b346 | 295 | delta_jiffies = 1; |
3c5d92a0 MS |
296 | } else { |
297 | /* Get the next timer wheel timer */ | |
298 | next_jiffies = get_next_timer_interrupt(last_jiffies); | |
299 | delta_jiffies = next_jiffies - last_jiffies; | |
aa9b1630 PM |
300 | if (rcu_delta_jiffies < delta_jiffies) { |
301 | next_jiffies = last_jiffies + rcu_delta_jiffies; | |
302 | delta_jiffies = rcu_delta_jiffies; | |
303 | } | |
3c5d92a0 | 304 | } |
79bf2bb3 TG |
305 | /* |
306 | * Do not stop the tick, if we are only one off | |
307 | * or if the cpu is required for rcu | |
308 | */ | |
6ba9b346 | 309 | if (!ts->tick_stopped && delta_jiffies == 1) |
79bf2bb3 TG |
310 | goto out; |
311 | ||
312 | /* Schedule the tick, if we are at least one jiffie off */ | |
313 | if ((long)delta_jiffies >= 1) { | |
314 | ||
00147449 WR |
315 | /* |
316 | * If this cpu is the one which updates jiffies, then | |
317 | * give up the assignment and let it be taken by the | |
318 | * cpu which runs the tick timer next, which might be | |
319 | * this cpu as well. If we don't drop this here the | |
320 | * jiffies might be stale and do_timer() never | |
27185016 TG |
321 | * invoked. Keep track of the fact that it was the one |
322 | * which had the do_timer() duty last. If this cpu is | |
323 | * the one which had the do_timer() duty last, we | |
324 | * limit the sleep time to the timekeeping | |
325 | * max_deferement value which we retrieved | |
326 | * above. Otherwise we can sleep as long as we want. | |
00147449 | 327 | */ |
27185016 | 328 | if (cpu == tick_do_timer_cpu) { |
00147449 | 329 | tick_do_timer_cpu = TICK_DO_TIMER_NONE; |
27185016 TG |
330 | ts->do_timer_last = 1; |
331 | } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { | |
332 | time_delta = KTIME_MAX; | |
333 | ts->do_timer_last = 0; | |
334 | } else if (!ts->do_timer_last) { | |
335 | time_delta = KTIME_MAX; | |
336 | } | |
337 | ||
00147449 | 338 | /* |
98962465 JH |
339 | * calculate the expiry time for the next timer wheel |
340 | * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals | |
341 | * that there is no timer pending or at least extremely | |
342 | * far into the future (12 days for HZ=1000). In this | |
343 | * case we set the expiry to the end of time. | |
344 | */ | |
345 | if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) { | |
346 | /* | |
347 | * Calculate the time delta for the next timer event. | |
348 | * If the time delta exceeds the maximum time delta | |
349 | * permitted by the current clocksource then adjust | |
350 | * the time delta accordingly to ensure the | |
351 | * clocksource does not wrap. | |
352 | */ | |
353 | time_delta = min_t(u64, time_delta, | |
354 | tick_period.tv64 * delta_jiffies); | |
98962465 | 355 | } |
00147449 | 356 | |
27185016 TG |
357 | if (time_delta < KTIME_MAX) |
358 | expires = ktime_add_ns(last_update, time_delta); | |
359 | else | |
360 | expires.tv64 = KTIME_MAX; | |
00147449 | 361 | |
00147449 WR |
362 | /* Skip reprogram of event if its not changed */ |
363 | if (ts->tick_stopped && ktime_equal(expires, dev->next_event)) | |
364 | goto out; | |
365 | ||
84bf1bcc FW |
366 | ret = expires; |
367 | ||
79bf2bb3 TG |
368 | /* |
369 | * nohz_stop_sched_tick can be called several times before | |
370 | * the nohz_restart_sched_tick is called. This happens when | |
371 | * interrupts arrive which do not cause a reschedule. In the | |
372 | * first call we save the current tick time, so we can restart | |
373 | * the scheduler tick in nohz_restart_sched_tick. | |
374 | */ | |
375 | if (!ts->tick_stopped) { | |
c1cc017c | 376 | nohz_balance_enter_idle(cpu); |
5167e8d5 | 377 | calc_load_enter_idle(); |
46cb4b7c | 378 | |
f5d411c9 | 379 | ts->last_tick = hrtimer_get_expires(&ts->sched_timer); |
79bf2bb3 | 380 | ts->tick_stopped = 1; |
79bf2bb3 | 381 | } |
d3ed7824 | 382 | |
eaad084b | 383 | /* |
98962465 JH |
384 | * If the expiration time == KTIME_MAX, then |
385 | * in this case we simply stop the tick timer. | |
eaad084b | 386 | */ |
98962465 | 387 | if (unlikely(expires.tv64 == KTIME_MAX)) { |
eaad084b TG |
388 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) |
389 | hrtimer_cancel(&ts->sched_timer); | |
390 | goto out; | |
391 | } | |
392 | ||
79bf2bb3 TG |
393 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { |
394 | hrtimer_start(&ts->sched_timer, expires, | |
5c333864 | 395 | HRTIMER_MODE_ABS_PINNED); |
79bf2bb3 TG |
396 | /* Check, if the timer was already in the past */ |
397 | if (hrtimer_active(&ts->sched_timer)) | |
398 | goto out; | |
4c9dc641 | 399 | } else if (!tick_program_event(expires, 0)) |
79bf2bb3 TG |
400 | goto out; |
401 | /* | |
402 | * We are past the event already. So we crossed a | |
403 | * jiffie boundary. Update jiffies and raise the | |
404 | * softirq. | |
405 | */ | |
406 | tick_do_update_jiffies64(ktime_get()); | |
79bf2bb3 TG |
407 | } |
408 | raise_softirq_irqoff(TIMER_SOFTIRQ); | |
409 | out: | |
410 | ts->next_jiffies = next_jiffies; | |
411 | ts->last_jiffies = last_jiffies; | |
4f86d3a8 | 412 | ts->sleep_length = ktime_sub(dev->next_event, now); |
84bf1bcc FW |
413 | |
414 | return ret; | |
280f0677 FW |
415 | } |
416 | ||
5b39939a FW |
417 | static bool can_stop_idle_tick(int cpu, struct tick_sched *ts) |
418 | { | |
419 | /* | |
420 | * If this cpu is offline and it is the one which updates | |
421 | * jiffies, then give up the assignment and let it be taken by | |
422 | * the cpu which runs the tick timer next. If we don't drop | |
423 | * this here the jiffies might be stale and do_timer() never | |
424 | * invoked. | |
425 | */ | |
426 | if (unlikely(!cpu_online(cpu))) { | |
427 | if (cpu == tick_do_timer_cpu) | |
428 | tick_do_timer_cpu = TICK_DO_TIMER_NONE; | |
429 | } | |
430 | ||
431 | if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) | |
432 | return false; | |
433 | ||
434 | if (need_resched()) | |
435 | return false; | |
436 | ||
437 | if (unlikely(local_softirq_pending() && cpu_online(cpu))) { | |
438 | static int ratelimit; | |
439 | ||
803b0eba PM |
440 | if (ratelimit < 10 && |
441 | (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) { | |
5b39939a FW |
442 | printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n", |
443 | (unsigned int) local_softirq_pending()); | |
444 | ratelimit++; | |
445 | } | |
446 | return false; | |
447 | } | |
448 | ||
449 | return true; | |
450 | } | |
451 | ||
19f5f736 FW |
452 | static void __tick_nohz_idle_enter(struct tick_sched *ts) |
453 | { | |
84bf1bcc | 454 | ktime_t now, expires; |
5b39939a | 455 | int cpu = smp_processor_id(); |
19f5f736 | 456 | |
5b39939a | 457 | now = tick_nohz_start_idle(cpu, ts); |
2ac0d98f | 458 | |
5b39939a FW |
459 | if (can_stop_idle_tick(cpu, ts)) { |
460 | int was_stopped = ts->tick_stopped; | |
461 | ||
462 | ts->idle_calls++; | |
84bf1bcc FW |
463 | |
464 | expires = tick_nohz_stop_sched_tick(ts, now, cpu); | |
465 | if (expires.tv64 > 0LL) { | |
466 | ts->idle_sleeps++; | |
467 | ts->idle_expires = expires; | |
468 | } | |
5b39939a FW |
469 | |
470 | if (!was_stopped && ts->tick_stopped) | |
471 | ts->idle_jiffies = ts->last_jiffies; | |
472 | } | |
280f0677 FW |
473 | } |
474 | ||
475 | /** | |
476 | * tick_nohz_idle_enter - stop the idle tick from the idle task | |
477 | * | |
478 | * When the next event is more than a tick into the future, stop the idle tick | |
479 | * Called when we start the idle loop. | |
2bbb6817 | 480 | * |
1268fbc7 | 481 | * The arch is responsible of calling: |
2bbb6817 FW |
482 | * |
483 | * - rcu_idle_enter() after its last use of RCU before the CPU is put | |
484 | * to sleep. | |
485 | * - rcu_idle_exit() before the first use of RCU after the CPU is woken up. | |
280f0677 | 486 | */ |
1268fbc7 | 487 | void tick_nohz_idle_enter(void) |
280f0677 FW |
488 | { |
489 | struct tick_sched *ts; | |
490 | ||
1268fbc7 FW |
491 | WARN_ON_ONCE(irqs_disabled()); |
492 | ||
0db49b72 LT |
493 | /* |
494 | * Update the idle state in the scheduler domain hierarchy | |
495 | * when tick_nohz_stop_sched_tick() is called from the idle loop. | |
496 | * State will be updated to busy during the first busy tick after | |
497 | * exiting idle. | |
498 | */ | |
499 | set_cpu_sd_state_idle(); | |
500 | ||
1268fbc7 FW |
501 | local_irq_disable(); |
502 | ||
280f0677 FW |
503 | ts = &__get_cpu_var(tick_cpu_sched); |
504 | /* | |
505 | * set ts->inidle unconditionally. even if the system did not | |
506 | * switch to nohz mode the cpu frequency governers rely on the | |
507 | * update of the idle time accounting in tick_nohz_start_idle(). | |
508 | */ | |
509 | ts->inidle = 1; | |
19f5f736 | 510 | __tick_nohz_idle_enter(ts); |
1268fbc7 FW |
511 | |
512 | local_irq_enable(); | |
280f0677 FW |
513 | } |
514 | ||
515 | /** | |
516 | * tick_nohz_irq_exit - update next tick event from interrupt exit | |
517 | * | |
518 | * When an interrupt fires while we are idle and it doesn't cause | |
519 | * a reschedule, it may still add, modify or delete a timer, enqueue | |
520 | * an RCU callback, etc... | |
521 | * So we need to re-calculate and reprogram the next tick event. | |
522 | */ | |
523 | void tick_nohz_irq_exit(void) | |
524 | { | |
525 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
526 | ||
527 | if (!ts->inidle) | |
528 | return; | |
529 | ||
19f5f736 | 530 | __tick_nohz_idle_enter(ts); |
79bf2bb3 TG |
531 | } |
532 | ||
4f86d3a8 LB |
533 | /** |
534 | * tick_nohz_get_sleep_length - return the length of the current sleep | |
535 | * | |
536 | * Called from power state control code with interrupts disabled | |
537 | */ | |
538 | ktime_t tick_nohz_get_sleep_length(void) | |
539 | { | |
540 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
541 | ||
542 | return ts->sleep_length; | |
543 | } | |
544 | ||
c34bec5a TG |
545 | static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) |
546 | { | |
547 | hrtimer_cancel(&ts->sched_timer); | |
f5d411c9 | 548 | hrtimer_set_expires(&ts->sched_timer, ts->last_tick); |
c34bec5a TG |
549 | |
550 | while (1) { | |
551 | /* Forward the time to expire in the future */ | |
552 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
553 | ||
554 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { | |
268a3dcf | 555 | hrtimer_start_expires(&ts->sched_timer, |
5c333864 | 556 | HRTIMER_MODE_ABS_PINNED); |
c34bec5a TG |
557 | /* Check, if the timer was already in the past */ |
558 | if (hrtimer_active(&ts->sched_timer)) | |
559 | break; | |
560 | } else { | |
268a3dcf TG |
561 | if (!tick_program_event( |
562 | hrtimer_get_expires(&ts->sched_timer), 0)) | |
c34bec5a TG |
563 | break; |
564 | } | |
6f103929 | 565 | /* Reread time and update jiffies */ |
c34bec5a | 566 | now = ktime_get(); |
6f103929 | 567 | tick_do_update_jiffies64(now); |
c34bec5a TG |
568 | } |
569 | } | |
570 | ||
19f5f736 | 571 | static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now) |
79bf2bb3 | 572 | { |
79bf2bb3 | 573 | /* Update jiffies first */ |
79bf2bb3 | 574 | tick_do_update_jiffies64(now); |
5aaa0b7a | 575 | update_cpu_load_nohz(); |
79bf2bb3 | 576 | |
749c8814 | 577 | calc_load_exit_idle(); |
2ac0d98f FW |
578 | touch_softlockup_watchdog(); |
579 | /* | |
580 | * Cancel the scheduled timer and restore the tick | |
581 | */ | |
582 | ts->tick_stopped = 0; | |
583 | ts->idle_exittime = now; | |
584 | ||
585 | tick_nohz_restart(ts, now); | |
586 | } | |
587 | ||
588 | static void tick_nohz_account_idle_ticks(struct tick_sched *ts) | |
589 | { | |
79741dd3 | 590 | #ifndef CONFIG_VIRT_CPU_ACCOUNTING |
2ac0d98f | 591 | unsigned long ticks; |
79bf2bb3 TG |
592 | /* |
593 | * We stopped the tick in idle. Update process times would miss the | |
594 | * time we slept as update_process_times does only a 1 tick | |
595 | * accounting. Enforce that this is accounted to idle ! | |
596 | */ | |
597 | ticks = jiffies - ts->idle_jiffies; | |
598 | /* | |
599 | * We might be one off. Do not randomly account a huge number of ticks! | |
600 | */ | |
79741dd3 MS |
601 | if (ticks && ticks < LONG_MAX) |
602 | account_idle_ticks(ticks); | |
603 | #endif | |
19f5f736 FW |
604 | } |
605 | ||
79bf2bb3 | 606 | /** |
280f0677 | 607 | * tick_nohz_idle_exit - restart the idle tick from the idle task |
79bf2bb3 TG |
608 | * |
609 | * Restart the idle tick when the CPU is woken up from idle | |
280f0677 FW |
610 | * This also exit the RCU extended quiescent state. The CPU |
611 | * can use RCU again after this function is called. | |
79bf2bb3 | 612 | */ |
280f0677 | 613 | void tick_nohz_idle_exit(void) |
79bf2bb3 TG |
614 | { |
615 | int cpu = smp_processor_id(); | |
616 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
6378ddb5 | 617 | ktime_t now; |
79bf2bb3 | 618 | |
6378ddb5 | 619 | local_irq_disable(); |
2bbb6817 | 620 | |
15f827be FW |
621 | WARN_ON_ONCE(!ts->inidle); |
622 | ||
623 | ts->inidle = 0; | |
624 | ||
625 | if (ts->idle_active || ts->tick_stopped) | |
eed3b9cf MS |
626 | now = ktime_get(); |
627 | ||
628 | if (ts->idle_active) | |
629 | tick_nohz_stop_idle(cpu, now); | |
6378ddb5 | 630 | |
2ac0d98f | 631 | if (ts->tick_stopped) { |
19f5f736 | 632 | tick_nohz_restart_sched_tick(ts, now); |
2ac0d98f | 633 | tick_nohz_account_idle_ticks(ts); |
6378ddb5 | 634 | } |
79bf2bb3 | 635 | |
79bf2bb3 TG |
636 | local_irq_enable(); |
637 | } | |
638 | ||
639 | static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now) | |
640 | { | |
641 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
cc584b21 | 642 | return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0); |
79bf2bb3 TG |
643 | } |
644 | ||
645 | /* | |
646 | * The nohz low res interrupt handler | |
647 | */ | |
648 | static void tick_nohz_handler(struct clock_event_device *dev) | |
649 | { | |
650 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
651 | struct pt_regs *regs = get_irq_regs(); | |
d3ed7824 | 652 | int cpu = smp_processor_id(); |
79bf2bb3 TG |
653 | ktime_t now = ktime_get(); |
654 | ||
655 | dev->next_event.tv64 = KTIME_MAX; | |
656 | ||
d3ed7824 TG |
657 | /* |
658 | * Check if the do_timer duty was dropped. We don't care about | |
659 | * concurrency: This happens only when the cpu in charge went | |
660 | * into a long sleep. If two cpus happen to assign themself to | |
661 | * this duty, then the jiffies update is still serialized by | |
662 | * xtime_lock. | |
663 | */ | |
6441402b | 664 | if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) |
d3ed7824 TG |
665 | tick_do_timer_cpu = cpu; |
666 | ||
79bf2bb3 | 667 | /* Check, if the jiffies need an update */ |
d3ed7824 TG |
668 | if (tick_do_timer_cpu == cpu) |
669 | tick_do_update_jiffies64(now); | |
79bf2bb3 TG |
670 | |
671 | /* | |
672 | * When we are idle and the tick is stopped, we have to touch | |
673 | * the watchdog as we might not schedule for a really long | |
674 | * time. This happens on complete idle SMP systems while | |
675 | * waiting on the login prompt. We also increment the "start | |
676 | * of idle" jiffy stamp so the idle accounting adjustment we | |
677 | * do when we go busy again does not account too much ticks. | |
678 | */ | |
679 | if (ts->tick_stopped) { | |
680 | touch_softlockup_watchdog(); | |
681 | ts->idle_jiffies++; | |
682 | } | |
683 | ||
684 | update_process_times(user_mode(regs)); | |
685 | profile_tick(CPU_PROFILING); | |
686 | ||
79bf2bb3 TG |
687 | while (tick_nohz_reprogram(ts, now)) { |
688 | now = ktime_get(); | |
689 | tick_do_update_jiffies64(now); | |
690 | } | |
691 | } | |
692 | ||
693 | /** | |
694 | * tick_nohz_switch_to_nohz - switch to nohz mode | |
695 | */ | |
696 | static void tick_nohz_switch_to_nohz(void) | |
697 | { | |
698 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
699 | ktime_t next; | |
700 | ||
701 | if (!tick_nohz_enabled) | |
702 | return; | |
703 | ||
704 | local_irq_disable(); | |
705 | if (tick_switch_to_oneshot(tick_nohz_handler)) { | |
706 | local_irq_enable(); | |
707 | return; | |
708 | } | |
709 | ||
710 | ts->nohz_mode = NOHZ_MODE_LOWRES; | |
711 | ||
712 | /* | |
713 | * Recycle the hrtimer in ts, so we can share the | |
714 | * hrtimer_forward with the highres code. | |
715 | */ | |
716 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); | |
717 | /* Get the next period */ | |
718 | next = tick_init_jiffy_update(); | |
719 | ||
720 | for (;;) { | |
cc584b21 | 721 | hrtimer_set_expires(&ts->sched_timer, next); |
79bf2bb3 TG |
722 | if (!tick_program_event(next, 0)) |
723 | break; | |
724 | next = ktime_add(next, tick_period); | |
725 | } | |
726 | local_irq_enable(); | |
79bf2bb3 TG |
727 | } |
728 | ||
fb02fbc1 TG |
729 | /* |
730 | * When NOHZ is enabled and the tick is stopped, we need to kick the | |
731 | * tick timer from irq_enter() so that the jiffies update is kept | |
732 | * alive during long running softirqs. That's ugly as hell, but | |
733 | * correctness is key even if we need to fix the offending softirq in | |
734 | * the first place. | |
735 | * | |
736 | * Note, this is different to tick_nohz_restart. We just kick the | |
737 | * timer and do not touch the other magic bits which need to be done | |
738 | * when idle is left. | |
739 | */ | |
eed3b9cf | 740 | static void tick_nohz_kick_tick(int cpu, ktime_t now) |
fb02fbc1 | 741 | { |
ae99286b TG |
742 | #if 0 |
743 | /* Switch back to 2.6.27 behaviour */ | |
744 | ||
fb02fbc1 | 745 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); |
eed3b9cf | 746 | ktime_t delta; |
fb02fbc1 | 747 | |
c4bd822e TG |
748 | /* |
749 | * Do not touch the tick device, when the next expiry is either | |
750 | * already reached or less/equal than the tick period. | |
751 | */ | |
268a3dcf | 752 | delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now); |
c4bd822e TG |
753 | if (delta.tv64 <= tick_period.tv64) |
754 | return; | |
755 | ||
756 | tick_nohz_restart(ts, now); | |
ae99286b | 757 | #endif |
fb02fbc1 TG |
758 | } |
759 | ||
eed3b9cf MS |
760 | static inline void tick_check_nohz(int cpu) |
761 | { | |
762 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
763 | ktime_t now; | |
764 | ||
765 | if (!ts->idle_active && !ts->tick_stopped) | |
766 | return; | |
767 | now = ktime_get(); | |
768 | if (ts->idle_active) | |
769 | tick_nohz_stop_idle(cpu, now); | |
770 | if (ts->tick_stopped) { | |
771 | tick_nohz_update_jiffies(now); | |
772 | tick_nohz_kick_tick(cpu, now); | |
773 | } | |
774 | } | |
775 | ||
79bf2bb3 TG |
776 | #else |
777 | ||
778 | static inline void tick_nohz_switch_to_nohz(void) { } | |
eed3b9cf | 779 | static inline void tick_check_nohz(int cpu) { } |
79bf2bb3 TG |
780 | |
781 | #endif /* NO_HZ */ | |
782 | ||
719254fa TG |
783 | /* |
784 | * Called from irq_enter to notify about the possible interruption of idle() | |
785 | */ | |
786 | void tick_check_idle(int cpu) | |
787 | { | |
fb02fbc1 | 788 | tick_check_oneshot_broadcast(cpu); |
eed3b9cf | 789 | tick_check_nohz(cpu); |
719254fa TG |
790 | } |
791 | ||
79bf2bb3 TG |
792 | /* |
793 | * High resolution timer specific code | |
794 | */ | |
795 | #ifdef CONFIG_HIGH_RES_TIMERS | |
796 | /* | |
4c9dc641 | 797 | * We rearm the timer until we get disabled by the idle code. |
79bf2bb3 TG |
798 | * Called with interrupts disabled and timer->base->cpu_base->lock held. |
799 | */ | |
800 | static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) | |
801 | { | |
802 | struct tick_sched *ts = | |
803 | container_of(timer, struct tick_sched, sched_timer); | |
79bf2bb3 TG |
804 | struct pt_regs *regs = get_irq_regs(); |
805 | ktime_t now = ktime_get(); | |
d3ed7824 TG |
806 | int cpu = smp_processor_id(); |
807 | ||
808 | #ifdef CONFIG_NO_HZ | |
809 | /* | |
810 | * Check if the do_timer duty was dropped. We don't care about | |
811 | * concurrency: This happens only when the cpu in charge went | |
812 | * into a long sleep. If two cpus happen to assign themself to | |
813 | * this duty, then the jiffies update is still serialized by | |
814 | * xtime_lock. | |
815 | */ | |
6441402b | 816 | if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) |
d3ed7824 TG |
817 | tick_do_timer_cpu = cpu; |
818 | #endif | |
79bf2bb3 TG |
819 | |
820 | /* Check, if the jiffies need an update */ | |
d3ed7824 TG |
821 | if (tick_do_timer_cpu == cpu) |
822 | tick_do_update_jiffies64(now); | |
79bf2bb3 TG |
823 | |
824 | /* | |
825 | * Do not call, when we are not in irq context and have | |
826 | * no valid regs pointer | |
827 | */ | |
828 | if (regs) { | |
829 | /* | |
830 | * When we are idle and the tick is stopped, we have to touch | |
831 | * the watchdog as we might not schedule for a really long | |
832 | * time. This happens on complete idle SMP systems while | |
833 | * waiting on the login prompt. We also increment the "start of | |
834 | * idle" jiffy stamp so the idle accounting adjustment we do | |
835 | * when we go busy again does not account too much ticks. | |
836 | */ | |
837 | if (ts->tick_stopped) { | |
838 | touch_softlockup_watchdog(); | |
2b17c545 | 839 | if (is_idle_task(current)) |
2ac0d98f | 840 | ts->idle_jiffies++; |
79bf2bb3 | 841 | } |
79bf2bb3 TG |
842 | update_process_times(user_mode(regs)); |
843 | profile_tick(CPU_PROFILING); | |
79bf2bb3 TG |
844 | } |
845 | ||
79bf2bb3 TG |
846 | hrtimer_forward(timer, now, tick_period); |
847 | ||
848 | return HRTIMER_RESTART; | |
849 | } | |
850 | ||
5307c955 MG |
851 | static int sched_skew_tick; |
852 | ||
62cf20b3 TG |
853 | static int __init skew_tick(char *str) |
854 | { | |
855 | get_option(&str, &sched_skew_tick); | |
856 | ||
857 | return 0; | |
858 | } | |
859 | early_param("skew_tick", skew_tick); | |
860 | ||
79bf2bb3 TG |
861 | /** |
862 | * tick_setup_sched_timer - setup the tick emulation timer | |
863 | */ | |
864 | void tick_setup_sched_timer(void) | |
865 | { | |
866 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
867 | ktime_t now = ktime_get(); | |
868 | ||
869 | /* | |
870 | * Emulate tick processing via per-CPU hrtimers: | |
871 | */ | |
872 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); | |
873 | ts->sched_timer.function = tick_sched_timer; | |
79bf2bb3 | 874 | |
3704540b | 875 | /* Get the next period (per cpu) */ |
cc584b21 | 876 | hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); |
79bf2bb3 | 877 | |
5307c955 MG |
878 | /* Offset the tick to avert xtime_lock contention. */ |
879 | if (sched_skew_tick) { | |
880 | u64 offset = ktime_to_ns(tick_period) >> 1; | |
881 | do_div(offset, num_possible_cpus()); | |
882 | offset *= smp_processor_id(); | |
883 | hrtimer_add_expires_ns(&ts->sched_timer, offset); | |
884 | } | |
885 | ||
79bf2bb3 TG |
886 | for (;;) { |
887 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
5c333864 AB |
888 | hrtimer_start_expires(&ts->sched_timer, |
889 | HRTIMER_MODE_ABS_PINNED); | |
79bf2bb3 TG |
890 | /* Check, if the timer was already in the past */ |
891 | if (hrtimer_active(&ts->sched_timer)) | |
892 | break; | |
893 | now = ktime_get(); | |
894 | } | |
895 | ||
896 | #ifdef CONFIG_NO_HZ | |
29c158e8 | 897 | if (tick_nohz_enabled) |
79bf2bb3 TG |
898 | ts->nohz_mode = NOHZ_MODE_HIGHRES; |
899 | #endif | |
900 | } | |
3c4fbe5e | 901 | #endif /* HIGH_RES_TIMERS */ |
79bf2bb3 | 902 | |
3c4fbe5e | 903 | #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS |
79bf2bb3 TG |
904 | void tick_cancel_sched_timer(int cpu) |
905 | { | |
906 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
907 | ||
3c4fbe5e | 908 | # ifdef CONFIG_HIGH_RES_TIMERS |
79bf2bb3 TG |
909 | if (ts->sched_timer.base) |
910 | hrtimer_cancel(&ts->sched_timer); | |
3c4fbe5e | 911 | # endif |
a7901766 | 912 | |
79bf2bb3 TG |
913 | ts->nohz_mode = NOHZ_MODE_INACTIVE; |
914 | } | |
3c4fbe5e | 915 | #endif |
79bf2bb3 TG |
916 | |
917 | /** | |
918 | * Async notification about clocksource changes | |
919 | */ | |
920 | void tick_clock_notify(void) | |
921 | { | |
922 | int cpu; | |
923 | ||
924 | for_each_possible_cpu(cpu) | |
925 | set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); | |
926 | } | |
927 | ||
928 | /* | |
929 | * Async notification about clock event changes | |
930 | */ | |
931 | void tick_oneshot_notify(void) | |
932 | { | |
933 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
934 | ||
935 | set_bit(0, &ts->check_clocks); | |
936 | } | |
937 | ||
938 | /** | |
939 | * Check, if a change happened, which makes oneshot possible. | |
940 | * | |
941 | * Called cyclic from the hrtimer softirq (driven by the timer | |
942 | * softirq) allow_nohz signals, that we can switch into low-res nohz | |
943 | * mode, because high resolution timers are disabled (either compile | |
944 | * or runtime). | |
945 | */ | |
946 | int tick_check_oneshot_change(int allow_nohz) | |
947 | { | |
948 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
949 | ||
950 | if (!test_and_clear_bit(0, &ts->check_clocks)) | |
951 | return 0; | |
952 | ||
953 | if (ts->nohz_mode != NOHZ_MODE_INACTIVE) | |
954 | return 0; | |
955 | ||
cf4fc6cb | 956 | if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) |
79bf2bb3 TG |
957 | return 0; |
958 | ||
959 | if (!allow_nohz) | |
960 | return 1; | |
961 | ||
962 | tick_nohz_switch_to_nohz(); | |
963 | return 0; | |
964 | } |