~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/kernel/time/tick-sched.c

Version: ~ [ linux-5.3 ] ~ [ linux-5.2.15 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.73 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.144 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.193 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.193 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.73 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.102 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  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  *
 12  *  Distribute under GPLv2.
 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/module.h>
 23 #include <linux/irq_work.h>
 24 #include <linux/posix-timers.h>
 25 #include <linux/perf_event.h>
 26 
 27 #include <asm/irq_regs.h>
 28 
 29 #include "tick-internal.h"
 30 
 31 #include <trace/events/timer.h>
 32 
 33 /*
 34  * Per cpu nohz control structure
 35  */
 36 DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
 37 
 38 /*
 39  * The time, when the last jiffy update happened. Protected by jiffies_lock.
 40  */
 41 static ktime_t last_jiffies_update;
 42 
 43 struct tick_sched *tick_get_tick_sched(int cpu)
 44 {
 45         return &per_cpu(tick_cpu_sched, cpu);
 46 }
 47 
 48 /*
 49  * Must be called with interrupts disabled !
 50  */
 51 static void tick_do_update_jiffies64(ktime_t now)
 52 {
 53         unsigned long ticks = 0;
 54         ktime_t delta;
 55 
 56         /*
 57          * Do a quick check without holding jiffies_lock:
 58          */
 59         delta = ktime_sub(now, last_jiffies_update);
 60         if (delta.tv64 < tick_period.tv64)
 61                 return;
 62 
 63         /* Reevalute with jiffies_lock held */
 64         write_seqlock(&jiffies_lock);
 65 
 66         delta = ktime_sub(now, last_jiffies_update);
 67         if (delta.tv64 >= tick_period.tv64) {
 68 
 69                 delta = ktime_sub(delta, tick_period);
 70                 last_jiffies_update = ktime_add(last_jiffies_update,
 71                                                 tick_period);
 72 
 73                 /* Slow path for long timeouts */
 74                 if (unlikely(delta.tv64 >= tick_period.tv64)) {
 75                         s64 incr = ktime_to_ns(tick_period);
 76 
 77                         ticks = ktime_divns(delta, incr);
 78 
 79                         last_jiffies_update = ktime_add_ns(last_jiffies_update,
 80                                                            incr * ticks);
 81                 }
 82                 do_timer(++ticks);
 83 
 84                 /* Keep the tick_next_period variable up to date */
 85                 tick_next_period = ktime_add(last_jiffies_update, tick_period);
 86         }
 87         write_sequnlock(&jiffies_lock);
 88 }
 89 
 90 /*
 91  * Initialize and return retrieve the jiffies update.
 92  */
 93 static ktime_t tick_init_jiffy_update(void)
 94 {
 95         ktime_t period;
 96 
 97         write_seqlock(&jiffies_lock);
 98         /* Did we start the jiffies update yet ? */
 99         if (last_jiffies_update.tv64 == 0)
100                 last_jiffies_update = tick_next_period;
101         period = last_jiffies_update;
102         write_sequnlock(&jiffies_lock);
103         return period;
104 }
105 
106 
107 static void tick_sched_do_timer(ktime_t now)
108 {
109         int cpu = smp_processor_id();
110 
111 #ifdef CONFIG_NO_HZ_COMMON
112         /*
113          * Check if the do_timer duty was dropped. We don't care about
114          * concurrency: This happens only when the cpu in charge went
115          * into a long sleep. If two cpus happen to assign themself to
116          * this duty, then the jiffies update is still serialized by
117          * jiffies_lock.
118          */
119         if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)
120             && !tick_nohz_full_cpu(cpu))
121                 tick_do_timer_cpu = cpu;
122 #endif
123 
124         /* Check, if the jiffies need an update */
125         if (tick_do_timer_cpu == cpu)
126                 tick_do_update_jiffies64(now);
127 }
128 
129 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
130 {
131 #ifdef CONFIG_NO_HZ_COMMON
132         /*
133          * When we are idle and the tick is stopped, we have to touch
134          * the watchdog as we might not schedule for a really long
135          * time. This happens on complete idle SMP systems while
136          * waiting on the login prompt. We also increment the "start of
137          * idle" jiffy stamp so the idle accounting adjustment we do
138          * when we go busy again does not account too much ticks.
139          */
140         if (ts->tick_stopped) {
141                 touch_softlockup_watchdog();
142                 if (is_idle_task(current))
143                         ts->idle_jiffies++;
144         }
145 #endif
146         update_process_times(user_mode(regs));
147         profile_tick(CPU_PROFILING);
148 }
149 
150 #ifdef CONFIG_NO_HZ_FULL
151 static cpumask_var_t nohz_full_mask;
152 bool have_nohz_full_mask;
153 
154 static bool can_stop_full_tick(void)
155 {
156         WARN_ON_ONCE(!irqs_disabled());
157 
158         if (!sched_can_stop_tick()) {
159                 trace_tick_stop(0, "more than 1 task in runqueue\n");
160                 return false;
161         }
162 
163         if (!posix_cpu_timers_can_stop_tick(current)) {
164                 trace_tick_stop(0, "posix timers running\n");
165                 return false;
166         }
167 
168         if (!perf_event_can_stop_tick()) {
169                 trace_tick_stop(0, "perf events running\n");
170                 return false;
171         }
172 
173         /* sched_clock_tick() needs us? */
174 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
175         /*
176          * TODO: kick full dynticks CPUs when
177          * sched_clock_stable is set.
178          */
179         if (!sched_clock_stable) {
180                 trace_tick_stop(0, "unstable sched clock\n");
181                 /*
182                  * Don't allow the user to think they can get
183                  * full NO_HZ with this machine.
184                  */
185                 WARN_ONCE(have_nohz_full_mask,
186                           "NO_HZ FULL will not work with unstable sched clock");
187                 return false;
188         }
189 #endif
190 
191         return true;
192 }
193 
194 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now);
195 
196 /*
197  * Re-evaluate the need for the tick on the current CPU
198  * and restart it if necessary.
199  */
200 void tick_nohz_full_check(void)
201 {
202         struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
203 
204         if (tick_nohz_full_cpu(smp_processor_id())) {
205                 if (ts->tick_stopped && !is_idle_task(current)) {
206                         if (!can_stop_full_tick())
207                                 tick_nohz_restart_sched_tick(ts, ktime_get());
208                 }
209         }
210 }
211 
212 static void nohz_full_kick_work_func(struct irq_work *work)
213 {
214         tick_nohz_full_check();
215 }
216 
217 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
218         .func = nohz_full_kick_work_func,
219 };
220 
221 /*
222  * Kick the current CPU if it's full dynticks in order to force it to
223  * re-evaluate its dependency on the tick and restart it if necessary.
224  */
225 void tick_nohz_full_kick(void)
226 {
227         if (tick_nohz_full_cpu(smp_processor_id()))
228                 irq_work_queue(&__get_cpu_var(nohz_full_kick_work));
229 }
230 
231 static void nohz_full_kick_ipi(void *info)
232 {
233         tick_nohz_full_check();
234 }
235 
236 /*
237  * Kick all full dynticks CPUs in order to force these to re-evaluate
238  * their dependency on the tick and restart it if necessary.
239  */
240 void tick_nohz_full_kick_all(void)
241 {
242         if (!have_nohz_full_mask)
243                 return;
244 
245         preempt_disable();
246         smp_call_function_many(nohz_full_mask,
247                                nohz_full_kick_ipi, NULL, false);
248         preempt_enable();
249 }
250 
251 /*
252  * Re-evaluate the need for the tick as we switch the current task.
253  * It might need the tick due to per task/process properties:
254  * perf events, posix cpu timers, ...
255  */
256 void tick_nohz_task_switch(struct task_struct *tsk)
257 {
258         unsigned long flags;
259 
260         local_irq_save(flags);
261 
262         if (!tick_nohz_full_cpu(smp_processor_id()))
263                 goto out;
264 
265         if (tick_nohz_tick_stopped() && !can_stop_full_tick())
266                 tick_nohz_full_kick();
267 
268 out:
269         local_irq_restore(flags);
270 }
271 
272 int tick_nohz_full_cpu(int cpu)
273 {
274         if (!have_nohz_full_mask)
275                 return 0;
276 
277         return cpumask_test_cpu(cpu, nohz_full_mask);
278 }
279 
280 /* Parse the boot-time nohz CPU list from the kernel parameters. */
281 static int __init tick_nohz_full_setup(char *str)
282 {
283         int cpu;
284 
285         alloc_bootmem_cpumask_var(&nohz_full_mask);
286         if (cpulist_parse(str, nohz_full_mask) < 0) {
287                 pr_warning("NOHZ: Incorrect nohz_full cpumask\n");
288                 return 1;
289         }
290 
291         cpu = smp_processor_id();
292         if (cpumask_test_cpu(cpu, nohz_full_mask)) {
293                 pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", cpu);
294                 cpumask_clear_cpu(cpu, nohz_full_mask);
295         }
296         have_nohz_full_mask = true;
297 
298         return 1;
299 }
300 __setup("nohz_full=", tick_nohz_full_setup);
301 
302 static int tick_nohz_cpu_down_callback(struct notifier_block *nfb,
303                                                  unsigned long action,
304                                                  void *hcpu)
305 {
306         unsigned int cpu = (unsigned long)hcpu;
307 
308         switch (action & ~CPU_TASKS_FROZEN) {
309         case CPU_DOWN_PREPARE:
310                 /*
311                  * If we handle the timekeeping duty for full dynticks CPUs,
312                  * we can't safely shutdown that CPU.
313                  */
314                 if (have_nohz_full_mask && tick_do_timer_cpu == cpu)
315                         return NOTIFY_BAD;
316                 break;
317         }
318         return NOTIFY_OK;
319 }
320 
321 /*
322  * Worst case string length in chunks of CPU range seems 2 steps
323  * separations: 0,2,4,6,...
324  * This is NR_CPUS + sizeof('\0')
325  */
326 static char __initdata nohz_full_buf[NR_CPUS + 1];
327 
328 static int tick_nohz_init_all(void)
329 {
330         int err = -1;
331 
332 #ifdef CONFIG_NO_HZ_FULL_ALL
333         if (!alloc_cpumask_var(&nohz_full_mask, GFP_KERNEL)) {
334                 pr_err("NO_HZ: Can't allocate full dynticks cpumask\n");
335                 return err;
336         }
337         err = 0;
338         cpumask_setall(nohz_full_mask);
339         cpumask_clear_cpu(smp_processor_id(), nohz_full_mask);
340         have_nohz_full_mask = true;
341 #endif
342         return err;
343 }
344 
345 void __init tick_nohz_init(void)
346 {
347         if (!have_nohz_full_mask) {
348                 if (tick_nohz_init_all() < 0)
349                         return;
350         }
351 
352         cpu_notifier(tick_nohz_cpu_down_callback, 0);
353         cpulist_scnprintf(nohz_full_buf, sizeof(nohz_full_buf), nohz_full_mask);
354         pr_info("NO_HZ: Full dynticks CPUs: %s.\n", nohz_full_buf);
355 }
356 #else
357 #define have_nohz_full_mask (0)
358 #endif
359 
360 /*
361  * NOHZ - aka dynamic tick functionality
362  */
363 #ifdef CONFIG_NO_HZ_COMMON
364 /*
365  * NO HZ enabled ?
366  */
367 int tick_nohz_enabled __read_mostly  = 1;
368 
369 /*
370  * Enable / Disable tickless mode
371  */
372 static int __init setup_tick_nohz(char *str)
373 {
374         if (!strcmp(str, "off"))
375                 tick_nohz_enabled = 0;
376         else if (!strcmp(str, "on"))
377                 tick_nohz_enabled = 1;
378         else
379                 return 0;
380         return 1;
381 }
382 
383 __setup("nohz=", setup_tick_nohz);
384 
385 /**
386  * tick_nohz_update_jiffies - update jiffies when idle was interrupted
387  *
388  * Called from interrupt entry when the CPU was idle
389  *
390  * In case the sched_tick was stopped on this CPU, we have to check if jiffies
391  * must be updated. Otherwise an interrupt handler could use a stale jiffy
392  * value. We do this unconditionally on any cpu, as we don't know whether the
393  * cpu, which has the update task assigned is in a long sleep.
394  */
395 static void tick_nohz_update_jiffies(ktime_t now)
396 {
397         int cpu = smp_processor_id();
398         struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
399         unsigned long flags;
400 
401         ts->idle_waketime = now;
402 
403         local_irq_save(flags);
404         tick_do_update_jiffies64(now);
405         local_irq_restore(flags);
406 
407         touch_softlockup_watchdog();
408 }
409 
410 /*
411  * Updates the per cpu time idle statistics counters
412  */
413 static void
414 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
415 {
416         ktime_t delta;
417 
418         if (ts->idle_active) {
419                 delta = ktime_sub(now, ts->idle_entrytime);
420                 if (nr_iowait_cpu(cpu) > 0)
421                         ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
422                 else
423                         ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
424                 ts->idle_entrytime = now;
425         }
426 
427         if (last_update_time)
428                 *last_update_time = ktime_to_us(now);
429 
430 }
431 
432 static void tick_nohz_stop_idle(int cpu, ktime_t now)
433 {
434         struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
435 
436         update_ts_time_stats(cpu, ts, now, NULL);
437         ts->idle_active = 0;
438 
439         sched_clock_idle_wakeup_event(0);
440 }
441 
442 static ktime_t tick_nohz_start_idle(int cpu, struct tick_sched *ts)
443 {
444         ktime_t now = ktime_get();
445 
446         ts->idle_entrytime = now;
447         ts->idle_active = 1;
448         sched_clock_idle_sleep_event();
449         return now;
450 }
451 
452 /**
453  * get_cpu_idle_time_us - get the total idle time of a cpu
454  * @cpu: CPU number to query
455  * @last_update_time: variable to store update time in. Do not update
456  * counters if NULL.
457  *
458  * Return the cummulative idle time (since boot) for a given
459  * CPU, in microseconds.
460  *
461  * This time is measured via accounting rather than sampling,
462  * and is as accurate as ktime_get() is.
463  *
464  * This function returns -1 if NOHZ is not enabled.
465  */
466 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
467 {
468         struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
469         ktime_t now, idle;
470 
471         if (!tick_nohz_enabled)
472                 return -1;
473 
474         now = ktime_get();
475         if (last_update_time) {
476                 update_ts_time_stats(cpu, ts, now, last_update_time);
477                 idle = ts->idle_sleeptime;
478         } else {
479                 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
480                         ktime_t delta = ktime_sub(now, ts->idle_entrytime);
481 
482                         idle = ktime_add(ts->idle_sleeptime, delta);
483                 } else {
484                         idle = ts->idle_sleeptime;
485                 }
486         }
487 
488         return ktime_to_us(idle);
489 
490 }
491 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
492 
493 /**
494  * get_cpu_iowait_time_us - get the total iowait time of a cpu
495  * @cpu: CPU number to query
496  * @last_update_time: variable to store update time in. Do not update
497  * counters if NULL.
498  *
499  * Return the cummulative iowait time (since boot) for a given
500  * CPU, in microseconds.
501  *
502  * This time is measured via accounting rather than sampling,
503  * and is as accurate as ktime_get() is.
504  *
505  * This function returns -1 if NOHZ is not enabled.
506  */
507 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
508 {
509         struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
510         ktime_t now, iowait;
511 
512         if (!tick_nohz_enabled)
513                 return -1;
514 
515         now = ktime_get();
516         if (last_update_time) {
517                 update_ts_time_stats(cpu, ts, now, last_update_time);
518                 iowait = ts->iowait_sleeptime;
519         } else {
520                 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
521                         ktime_t delta = ktime_sub(now, ts->idle_entrytime);
522 
523                         iowait = ktime_add(ts->iowait_sleeptime, delta);
524                 } else {
525                         iowait = ts->iowait_sleeptime;
526                 }
527         }
528 
529         return ktime_to_us(iowait);
530 }
531 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
532 
533 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
534                                          ktime_t now, int cpu)
535 {
536         unsigned long seq, last_jiffies, next_jiffies, delta_jiffies;
537         ktime_t last_update, expires, ret = { .tv64 = 0 };
538         unsigned long rcu_delta_jiffies;
539         struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
540         u64 time_delta;
541 
542         /* Read jiffies and the time when jiffies were updated last */
543         do {
544                 seq = read_seqbegin(&jiffies_lock);
545                 last_update = last_jiffies_update;
546                 last_jiffies = jiffies;
547                 time_delta = timekeeping_max_deferment();
548         } while (read_seqretry(&jiffies_lock, seq));
549 
550         if (rcu_needs_cpu(cpu, &rcu_delta_jiffies) ||
551             arch_needs_cpu(cpu) || irq_work_needs_cpu()) {
552                 next_jiffies = last_jiffies + 1;
553                 delta_jiffies = 1;
554         } else {
555                 /* Get the next timer wheel timer */
556                 next_jiffies = get_next_timer_interrupt(last_jiffies);
557                 delta_jiffies = next_jiffies - last_jiffies;
558                 if (rcu_delta_jiffies < delta_jiffies) {
559                         next_jiffies = last_jiffies + rcu_delta_jiffies;
560                         delta_jiffies = rcu_delta_jiffies;
561                 }
562         }
563 
564         /*
565          * Do not stop the tick, if we are only one off (or less)
566          * or if the cpu is required for RCU:
567          */
568         if (!ts->tick_stopped && delta_jiffies <= 1)
569                 goto out;
570 
571         /* Schedule the tick, if we are at least one jiffie off */
572         if ((long)delta_jiffies >= 1) {
573 
574                 /*
575                  * If this cpu is the one which updates jiffies, then
576                  * give up the assignment and let it be taken by the
577                  * cpu which runs the tick timer next, which might be
578                  * this cpu as well. If we don't drop this here the
579                  * jiffies might be stale and do_timer() never
580                  * invoked. Keep track of the fact that it was the one
581                  * which had the do_timer() duty last. If this cpu is
582                  * the one which had the do_timer() duty last, we
583                  * limit the sleep time to the timekeeping
584                  * max_deferement value which we retrieved
585                  * above. Otherwise we can sleep as long as we want.
586                  */
587                 if (cpu == tick_do_timer_cpu) {
588                         tick_do_timer_cpu = TICK_DO_TIMER_NONE;
589                         ts->do_timer_last = 1;
590                 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
591                         time_delta = KTIME_MAX;
592                         ts->do_timer_last = 0;
593                 } else if (!ts->do_timer_last) {
594                         time_delta = KTIME_MAX;
595                 }
596 
597 #ifdef CONFIG_NO_HZ_FULL
598                 if (!ts->inidle) {
599                         time_delta = min(time_delta,
600                                          scheduler_tick_max_deferment());
601                 }
602 #endif
603 
604                 /*
605                  * calculate the expiry time for the next timer wheel
606                  * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
607                  * that there is no timer pending or at least extremely
608                  * far into the future (12 days for HZ=1000). In this
609                  * case we set the expiry to the end of time.
610                  */
611                 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) {
612                         /*
613                          * Calculate the time delta for the next timer event.
614                          * If the time delta exceeds the maximum time delta
615                          * permitted by the current clocksource then adjust
616                          * the time delta accordingly to ensure the
617                          * clocksource does not wrap.
618                          */
619                         time_delta = min_t(u64, time_delta,
620                                            tick_period.tv64 * delta_jiffies);
621                 }
622 
623                 if (time_delta < KTIME_MAX)
624                         expires = ktime_add_ns(last_update, time_delta);
625                 else
626                         expires.tv64 = KTIME_MAX;
627 
628                 /* Skip reprogram of event if its not changed */
629                 if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
630                         goto out;
631 
632                 ret = expires;
633 
634                 /*
635                  * nohz_stop_sched_tick can be called several times before
636                  * the nohz_restart_sched_tick is called. This happens when
637                  * interrupts arrive which do not cause a reschedule. In the
638                  * first call we save the current tick time, so we can restart
639                  * the scheduler tick in nohz_restart_sched_tick.
640                  */
641                 if (!ts->tick_stopped) {
642                         nohz_balance_enter_idle(cpu);
643                         calc_load_enter_idle();
644 
645                         ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
646                         ts->tick_stopped = 1;
647                         trace_tick_stop(1, " ");
648                 }
649 
650                 /*
651                  * If the expiration time == KTIME_MAX, then
652                  * in this case we simply stop the tick timer.
653                  */
654                  if (unlikely(expires.tv64 == KTIME_MAX)) {
655                         if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
656                                 hrtimer_cancel(&ts->sched_timer);
657                         goto out;
658                 }
659 
660                 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
661                         hrtimer_start(&ts->sched_timer, expires,
662                                       HRTIMER_MODE_ABS_PINNED);
663                         /* Check, if the timer was already in the past */
664                         if (hrtimer_active(&ts->sched_timer))
665                                 goto out;
666                 } else if (!tick_program_event(expires, 0))
667                                 goto out;
668                 /*
669                  * We are past the event already. So we crossed a
670                  * jiffie boundary. Update jiffies and raise the
671                  * softirq.
672                  */
673                 tick_do_update_jiffies64(ktime_get());
674         }
675         raise_softirq_irqoff(TIMER_SOFTIRQ);
676 out:
677         ts->next_jiffies = next_jiffies;
678         ts->last_jiffies = last_jiffies;
679         ts->sleep_length = ktime_sub(dev->next_event, now);
680 
681         return ret;
682 }
683 
684 static void tick_nohz_full_stop_tick(struct tick_sched *ts)
685 {
686 #ifdef CONFIG_NO_HZ_FULL
687        int cpu = smp_processor_id();
688 
689        if (!tick_nohz_full_cpu(cpu) || is_idle_task(current))
690                return;
691 
692        if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
693                return;
694 
695        if (!can_stop_full_tick())
696                return;
697 
698        tick_nohz_stop_sched_tick(ts, ktime_get(), cpu);
699 #endif
700 }
701 
702 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
703 {
704         /*
705          * If this cpu is offline and it is the one which updates
706          * jiffies, then give up the assignment and let it be taken by
707          * the cpu which runs the tick timer next. If we don't drop
708          * this here the jiffies might be stale and do_timer() never
709          * invoked.
710          */
711         if (unlikely(!cpu_online(cpu))) {
712                 if (cpu == tick_do_timer_cpu)
713                         tick_do_timer_cpu = TICK_DO_TIMER_NONE;
714                 return false;
715         }
716 
717         if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
718                 return false;
719 
720         if (need_resched())
721                 return false;
722 
723         if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
724                 static int ratelimit;
725 
726                 if (ratelimit < 10 &&
727                     (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
728                         pr_warn("NOHZ: local_softirq_pending %02x\n",
729                                 (unsigned int) local_softirq_pending());
730                         ratelimit++;
731                 }
732                 return false;
733         }
734 
735         if (have_nohz_full_mask) {
736                 /*
737                  * Keep the tick alive to guarantee timekeeping progression
738                  * if there are full dynticks CPUs around
739                  */
740                 if (tick_do_timer_cpu == cpu)
741                         return false;
742                 /*
743                  * Boot safety: make sure the timekeeping duty has been
744                  * assigned before entering dyntick-idle mode,
745                  */
746                 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
747                         return false;
748         }
749 
750         return true;
751 }
752 
753 static void __tick_nohz_idle_enter(struct tick_sched *ts)
754 {
755         ktime_t now, expires;
756         int cpu = smp_processor_id();
757 
758         now = tick_nohz_start_idle(cpu, ts);
759 
760         if (can_stop_idle_tick(cpu, ts)) {
761                 int was_stopped = ts->tick_stopped;
762 
763                 ts->idle_calls++;
764 
765                 expires = tick_nohz_stop_sched_tick(ts, now, cpu);
766                 if (expires.tv64 > 0LL) {
767                         ts->idle_sleeps++;
768                         ts->idle_expires = expires;
769                 }
770 
771                 if (!was_stopped && ts->tick_stopped)
772                         ts->idle_jiffies = ts->last_jiffies;
773         }
774 }
775 
776 /**
777  * tick_nohz_idle_enter - stop the idle tick from the idle task
778  *
779  * When the next event is more than a tick into the future, stop the idle tick
780  * Called when we start the idle loop.
781  *
782  * The arch is responsible of calling:
783  *
784  * - rcu_idle_enter() after its last use of RCU before the CPU is put
785  *  to sleep.
786  * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
787  */
788 void tick_nohz_idle_enter(void)
789 {
790         struct tick_sched *ts;
791 
792         WARN_ON_ONCE(irqs_disabled());
793 
794         /*
795          * Update the idle state in the scheduler domain hierarchy
796          * when tick_nohz_stop_sched_tick() is called from the idle loop.
797          * State will be updated to busy during the first busy tick after
798          * exiting idle.
799          */
800         set_cpu_sd_state_idle();
801 
802         local_irq_disable();
803 
804         ts = &__get_cpu_var(tick_cpu_sched);
805         /*
806          * set ts->inidle unconditionally. even if the system did not
807          * switch to nohz mode the cpu frequency governers rely on the
808          * update of the idle time accounting in tick_nohz_start_idle().
809          */
810         ts->inidle = 1;
811         __tick_nohz_idle_enter(ts);
812 
813         local_irq_enable();
814 }
815 EXPORT_SYMBOL_GPL(tick_nohz_idle_enter);
816 
817 /**
818  * tick_nohz_irq_exit - update next tick event from interrupt exit
819  *
820  * When an interrupt fires while we are idle and it doesn't cause
821  * a reschedule, it may still add, modify or delete a timer, enqueue
822  * an RCU callback, etc...
823  * So we need to re-calculate and reprogram the next tick event.
824  */
825 void tick_nohz_irq_exit(void)
826 {
827         struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
828 
829         if (ts->inidle)
830                 __tick_nohz_idle_enter(ts);
831         else
832                 tick_nohz_full_stop_tick(ts);
833 }
834 
835 /**
836  * tick_nohz_get_sleep_length - return the length of the current sleep
837  *
838  * Called from power state control code with interrupts disabled
839  */
840 ktime_t tick_nohz_get_sleep_length(void)
841 {
842         struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
843 
844         return ts->sleep_length;
845 }
846 
847 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
848 {
849         hrtimer_cancel(&ts->sched_timer);
850         hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
851 
852         while (1) {
853                 /* Forward the time to expire in the future */
854                 hrtimer_forward(&ts->sched_timer, now, tick_period);
855 
856                 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
857                         hrtimer_start_expires(&ts->sched_timer,
858                                               HRTIMER_MODE_ABS_PINNED);
859                         /* Check, if the timer was already in the past */
860                         if (hrtimer_active(&ts->sched_timer))
861                                 break;
862                 } else {
863                         if (!tick_program_event(
864                                 hrtimer_get_expires(&ts->sched_timer), 0))
865                                 break;
866                 }
867                 /* Reread time and update jiffies */
868                 now = ktime_get();
869                 tick_do_update_jiffies64(now);
870         }
871 }
872 
873 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
874 {
875         /* Update jiffies first */
876         tick_do_update_jiffies64(now);
877         update_cpu_load_nohz();
878 
879         calc_load_exit_idle();
880         touch_softlockup_watchdog();
881         /*
882          * Cancel the scheduled timer and restore the tick
883          */
884         ts->tick_stopped  = 0;
885         ts->idle_exittime = now;
886 
887         tick_nohz_restart(ts, now);
888 }
889 
890 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
891 {
892 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
893         unsigned long ticks;
894 
895         if (vtime_accounting_enabled())
896                 return;
897         /*
898          * We stopped the tick in idle. Update process times would miss the
899          * time we slept as update_process_times does only a 1 tick
900          * accounting. Enforce that this is accounted to idle !
901          */
902         ticks = jiffies - ts->idle_jiffies;
903         /*
904          * We might be one off. Do not randomly account a huge number of ticks!
905          */
906         if (ticks && ticks < LONG_MAX)
907                 account_idle_ticks(ticks);
908 #endif
909 }
910 
911 /**
912  * tick_nohz_idle_exit - restart the idle tick from the idle task
913  *
914  * Restart the idle tick when the CPU is woken up from idle
915  * This also exit the RCU extended quiescent state. The CPU
916  * can use RCU again after this function is called.
917  */
918 void tick_nohz_idle_exit(void)
919 {
920         int cpu = smp_processor_id();
921         struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
922         ktime_t now;
923 
924         local_irq_disable();
925 
926         WARN_ON_ONCE(!ts->inidle);
927 
928         ts->inidle = 0;
929 
930         if (ts->idle_active || ts->tick_stopped)
931                 now = ktime_get();
932 
933         if (ts->idle_active)
934                 tick_nohz_stop_idle(cpu, now);
935 
936         if (ts->tick_stopped) {
937                 tick_nohz_restart_sched_tick(ts, now);
938                 tick_nohz_account_idle_ticks(ts);
939         }
940 
941         local_irq_enable();
942 }
943 EXPORT_SYMBOL_GPL(tick_nohz_idle_exit);
944 
945 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
946 {
947         hrtimer_forward(&ts->sched_timer, now, tick_period);
948         return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
949 }
950 
951 /*
952  * The nohz low res interrupt handler
953  */
954 static void tick_nohz_handler(struct clock_event_device *dev)
955 {
956         struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
957         struct pt_regs *regs = get_irq_regs();
958         ktime_t now = ktime_get();
959 
960         dev->next_event.tv64 = KTIME_MAX;
961 
962         tick_sched_do_timer(now);
963         tick_sched_handle(ts, regs);
964 
965         while (tick_nohz_reprogram(ts, now)) {
966                 now = ktime_get();
967                 tick_do_update_jiffies64(now);
968         }
969 }
970 
971 /**
972  * tick_nohz_switch_to_nohz - switch to nohz mode
973  */
974 static void tick_nohz_switch_to_nohz(void)
975 {
976         struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
977         ktime_t next;
978 
979         if (!tick_nohz_enabled)
980                 return;
981 
982         local_irq_disable();
983         if (tick_switch_to_oneshot(tick_nohz_handler)) {
984                 local_irq_enable();
985                 return;
986         }
987 
988         ts->nohz_mode = NOHZ_MODE_LOWRES;
989 
990         /*
991          * Recycle the hrtimer in ts, so we can share the
992          * hrtimer_forward with the highres code.
993          */
994         hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
995         /* Get the next period */
996         next = tick_init_jiffy_update();
997 
998         for (;;) {
999                 hrtimer_set_expires(&ts->sched_timer, next);
1000                 if (!tick_program_event(next, 0))
1001                         break;
1002                 next = ktime_add(next, tick_period);
1003         }
1004         local_irq_enable();
1005 }
1006 
1007 /*
1008  * When NOHZ is enabled and the tick is stopped, we need to kick the
1009  * tick timer from irq_enter() so that the jiffies update is kept
1010  * alive during long running softirqs. That's ugly as hell, but
1011  * correctness is key even if we need to fix the offending softirq in
1012  * the first place.
1013  *
1014  * Note, this is different to tick_nohz_restart. We just kick the
1015  * timer and do not touch the other magic bits which need to be done
1016  * when idle is left.
1017  */
1018 static void tick_nohz_kick_tick(int cpu, ktime_t now)
1019 {
1020 #if 0
1021         /* Switch back to 2.6.27 behaviour */
1022 
1023         struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1024         ktime_t delta;
1025 
1026         /*
1027          * Do not touch the tick device, when the next expiry is either
1028          * already reached or less/equal than the tick period.
1029          */
1030         delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
1031         if (delta.tv64 <= tick_period.tv64)
1032                 return;
1033 
1034         tick_nohz_restart(ts, now);
1035 #endif
1036 }
1037 
1038 static inline void tick_check_nohz(int cpu)
1039 {
1040         struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1041         ktime_t now;
1042 
1043         if (!ts->idle_active && !ts->tick_stopped)
1044                 return;
1045         now = ktime_get();
1046         if (ts->idle_active)
1047                 tick_nohz_stop_idle(cpu, now);
1048         if (ts->tick_stopped) {
1049                 tick_nohz_update_jiffies(now);
1050                 tick_nohz_kick_tick(cpu, now);
1051         }
1052 }
1053 
1054 #else
1055 
1056 static inline void tick_nohz_switch_to_nohz(void) { }
1057 static inline void tick_check_nohz(int cpu) { }
1058 
1059 #endif /* CONFIG_NO_HZ_COMMON */
1060 
1061 /*
1062  * Called from irq_enter to notify about the possible interruption of idle()
1063  */
1064 void tick_check_idle(int cpu)
1065 {
1066         tick_check_oneshot_broadcast(cpu);
1067         tick_check_nohz(cpu);
1068 }
1069 
1070 /*
1071  * High resolution timer specific code
1072  */
1073 #ifdef CONFIG_HIGH_RES_TIMERS
1074 /*
1075  * We rearm the timer until we get disabled by the idle code.
1076  * Called with interrupts disabled.
1077  */
1078 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1079 {
1080         struct tick_sched *ts =
1081                 container_of(timer, struct tick_sched, sched_timer);
1082         struct pt_regs *regs = get_irq_regs();
1083         ktime_t now = ktime_get();
1084 
1085         tick_sched_do_timer(now);
1086 
1087         /*
1088          * Do not call, when we are not in irq context and have
1089          * no valid regs pointer
1090          */
1091         if (regs)
1092                 tick_sched_handle(ts, regs);
1093 
1094         hrtimer_forward(timer, now, tick_period);
1095 
1096         return HRTIMER_RESTART;
1097 }
1098 
1099 static int sched_skew_tick;
1100 
1101 static int __init skew_tick(char *str)
1102 {
1103         get_option(&str, &sched_skew_tick);
1104 
1105         return 0;
1106 }
1107 early_param("skew_tick", skew_tick);
1108 
1109 /**
1110  * tick_setup_sched_timer - setup the tick emulation timer
1111  */
1112 void tick_setup_sched_timer(void)
1113 {
1114         struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1115         ktime_t now = ktime_get();
1116 
1117         /*
1118          * Emulate tick processing via per-CPU hrtimers:
1119          */
1120         hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1121         ts->sched_timer.function = tick_sched_timer;
1122 
1123         /* Get the next period (per cpu) */
1124         hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1125 
1126         /* Offset the tick to avert jiffies_lock contention. */
1127         if (sched_skew_tick) {
1128                 u64 offset = ktime_to_ns(tick_period) >> 1;
1129                 do_div(offset, num_possible_cpus());
1130                 offset *= smp_processor_id();
1131                 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1132         }
1133 
1134         for (;;) {
1135                 hrtimer_forward(&ts->sched_timer, now, tick_period);
1136                 hrtimer_start_expires(&ts->sched_timer,
1137                                       HRTIMER_MODE_ABS_PINNED);
1138                 /* Check, if the timer was already in the past */
1139                 if (hrtimer_active(&ts->sched_timer))
1140                         break;
1141                 now = ktime_get();
1142         }
1143 
1144 #ifdef CONFIG_NO_HZ_COMMON
1145         if (tick_nohz_enabled)
1146                 ts->nohz_mode = NOHZ_MODE_HIGHRES;
1147 #endif
1148 }
1149 #endif /* HIGH_RES_TIMERS */
1150 
1151 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1152 void tick_cancel_sched_timer(int cpu)
1153 {
1154         struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1155 
1156 # ifdef CONFIG_HIGH_RES_TIMERS
1157         if (ts->sched_timer.base)
1158                 hrtimer_cancel(&ts->sched_timer);
1159 # endif
1160 
1161         memset(ts, 0, sizeof(*ts));
1162 }
1163 #endif
1164 
1165 /**
1166  * Async notification about clocksource changes
1167  */
1168 void tick_clock_notify(void)
1169 {
1170         int cpu;
1171 
1172         for_each_possible_cpu(cpu)
1173                 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1174 }
1175 
1176 /*
1177  * Async notification about clock event changes
1178  */
1179 void tick_oneshot_notify(void)
1180 {
1181         struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1182 
1183         set_bit(0, &ts->check_clocks);
1184 }
1185 
1186 /**
1187  * Check, if a change happened, which makes oneshot possible.
1188  *
1189  * Called cyclic from the hrtimer softirq (driven by the timer
1190  * softirq) allow_nohz signals, that we can switch into low-res nohz
1191  * mode, because high resolution timers are disabled (either compile
1192  * or runtime).
1193  */
1194 int tick_check_oneshot_change(int allow_nohz)
1195 {
1196         struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1197 
1198         if (!test_and_clear_bit(0, &ts->check_clocks))
1199                 return 0;
1200 
1201         if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1202                 return 0;
1203 
1204         if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1205                 return 0;
1206 
1207         if (!allow_nohz)
1208                 return 1;
1209 
1210         tick_nohz_switch_to_nohz();
1211         return 0;
1212 }
1213 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp