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

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

Version: ~ [ linux-5.3-rc5 ] ~ [ linux-5.2.9 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.67 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.139 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.189 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.189 ] ~ [ 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-3.16.72 ] ~ [ 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/nmi.h>
 21 #include <linux/profile.h>
 22 #include <linux/sched/signal.h>
 23 #include <linux/sched/clock.h>
 24 #include <linux/sched/stat.h>
 25 #include <linux/sched/nohz.h>
 26 #include <linux/module.h>
 27 #include <linux/irq_work.h>
 28 #include <linux/posix-timers.h>
 29 #include <linux/context_tracking.h>
 30 
 31 #include <asm/irq_regs.h>
 32 
 33 #include "tick-internal.h"
 34 
 35 #include <trace/events/timer.h>
 36 
 37 /*
 38  * Per-CPU nohz control structure
 39  */
 40 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
 41 
 42 struct tick_sched *tick_get_tick_sched(int cpu)
 43 {
 44         return &per_cpu(tick_cpu_sched, cpu);
 45 }
 46 
 47 #if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS)
 48 /*
 49  * The time, when the last jiffy update happened. Protected by jiffies_lock.
 50  */
 51 static ktime_t last_jiffies_update;
 52 
 53 /*
 54  * Must be called with interrupts disabled !
 55  */
 56 static void tick_do_update_jiffies64(ktime_t now)
 57 {
 58         unsigned long ticks = 0;
 59         ktime_t delta;
 60 
 61         /*
 62          * Do a quick check without holding jiffies_lock:
 63          */
 64         delta = ktime_sub(now, last_jiffies_update);
 65         if (delta < tick_period)
 66                 return;
 67 
 68         /* Reevaluate with jiffies_lock held */
 69         write_seqlock(&jiffies_lock);
 70 
 71         delta = ktime_sub(now, last_jiffies_update);
 72         if (delta >= tick_period) {
 73 
 74                 delta = ktime_sub(delta, tick_period);
 75                 last_jiffies_update = ktime_add(last_jiffies_update,
 76                                                 tick_period);
 77 
 78                 /* Slow path for long timeouts */
 79                 if (unlikely(delta >= tick_period)) {
 80                         s64 incr = ktime_to_ns(tick_period);
 81 
 82                         ticks = ktime_divns(delta, incr);
 83 
 84                         last_jiffies_update = ktime_add_ns(last_jiffies_update,
 85                                                            incr * ticks);
 86                 }
 87                 do_timer(++ticks);
 88 
 89                 /* Keep the tick_next_period variable up to date */
 90                 tick_next_period = ktime_add(last_jiffies_update, tick_period);
 91         } else {
 92                 write_sequnlock(&jiffies_lock);
 93                 return;
 94         }
 95         write_sequnlock(&jiffies_lock);
 96         update_wall_time();
 97 }
 98 
 99 /*
100  * Initialize and return retrieve the jiffies update.
101  */
102 static ktime_t tick_init_jiffy_update(void)
103 {
104         ktime_t period;
105 
106         write_seqlock(&jiffies_lock);
107         /* Did we start the jiffies update yet ? */
108         if (last_jiffies_update == 0)
109                 last_jiffies_update = tick_next_period;
110         period = last_jiffies_update;
111         write_sequnlock(&jiffies_lock);
112         return period;
113 }
114 
115 
116 static void tick_sched_do_timer(ktime_t now)
117 {
118         int cpu = smp_processor_id();
119 
120 #ifdef CONFIG_NO_HZ_COMMON
121         /*
122          * Check if the do_timer duty was dropped. We don't care about
123          * concurrency: This happens only when the CPU in charge went
124          * into a long sleep. If two CPUs happen to assign themselves to
125          * this duty, then the jiffies update is still serialized by
126          * jiffies_lock.
127          */
128         if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)
129             && !tick_nohz_full_cpu(cpu))
130                 tick_do_timer_cpu = cpu;
131 #endif
132 
133         /* Check, if the jiffies need an update */
134         if (tick_do_timer_cpu == cpu)
135                 tick_do_update_jiffies64(now);
136 }
137 
138 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
139 {
140 #ifdef CONFIG_NO_HZ_COMMON
141         /*
142          * When we are idle and the tick is stopped, we have to touch
143          * the watchdog as we might not schedule for a really long
144          * time. This happens on complete idle SMP systems while
145          * waiting on the login prompt. We also increment the "start of
146          * idle" jiffy stamp so the idle accounting adjustment we do
147          * when we go busy again does not account too much ticks.
148          */
149         if (ts->tick_stopped) {
150                 touch_softlockup_watchdog_sched();
151                 if (is_idle_task(current))
152                         ts->idle_jiffies++;
153         }
154 #endif
155         update_process_times(user_mode(regs));
156         profile_tick(CPU_PROFILING);
157 }
158 #endif
159 
160 #ifdef CONFIG_NO_HZ_FULL
161 cpumask_var_t tick_nohz_full_mask;
162 cpumask_var_t housekeeping_mask;
163 bool tick_nohz_full_running;
164 static atomic_t tick_dep_mask;
165 
166 static bool check_tick_dependency(atomic_t *dep)
167 {
168         int val = atomic_read(dep);
169 
170         if (val & TICK_DEP_MASK_POSIX_TIMER) {
171                 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
172                 return true;
173         }
174 
175         if (val & TICK_DEP_MASK_PERF_EVENTS) {
176                 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
177                 return true;
178         }
179 
180         if (val & TICK_DEP_MASK_SCHED) {
181                 trace_tick_stop(0, TICK_DEP_MASK_SCHED);
182                 return true;
183         }
184 
185         if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
186                 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
187                 return true;
188         }
189 
190         return false;
191 }
192 
193 static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
194 {
195         WARN_ON_ONCE(!irqs_disabled());
196 
197         if (unlikely(!cpu_online(cpu)))
198                 return false;
199 
200         if (check_tick_dependency(&tick_dep_mask))
201                 return false;
202 
203         if (check_tick_dependency(&ts->tick_dep_mask))
204                 return false;
205 
206         if (check_tick_dependency(&current->tick_dep_mask))
207                 return false;
208 
209         if (check_tick_dependency(&current->signal->tick_dep_mask))
210                 return false;
211 
212         return true;
213 }
214 
215 static void nohz_full_kick_func(struct irq_work *work)
216 {
217         /* Empty, the tick restart happens on tick_nohz_irq_exit() */
218 }
219 
220 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
221         .func = nohz_full_kick_func,
222 };
223 
224 /*
225  * Kick this CPU if it's full dynticks in order to force it to
226  * re-evaluate its dependency on the tick and restart it if necessary.
227  * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
228  * is NMI safe.
229  */
230 static void tick_nohz_full_kick(void)
231 {
232         if (!tick_nohz_full_cpu(smp_processor_id()))
233                 return;
234 
235         irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
236 }
237 
238 /*
239  * Kick the CPU if it's full dynticks in order to force it to
240  * re-evaluate its dependency on the tick and restart it if necessary.
241  */
242 void tick_nohz_full_kick_cpu(int cpu)
243 {
244         if (!tick_nohz_full_cpu(cpu))
245                 return;
246 
247         irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
248 }
249 
250 /*
251  * Kick all full dynticks CPUs in order to force these to re-evaluate
252  * their dependency on the tick and restart it if necessary.
253  */
254 static void tick_nohz_full_kick_all(void)
255 {
256         int cpu;
257 
258         if (!tick_nohz_full_running)
259                 return;
260 
261         preempt_disable();
262         for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
263                 tick_nohz_full_kick_cpu(cpu);
264         preempt_enable();
265 }
266 
267 static void tick_nohz_dep_set_all(atomic_t *dep,
268                                   enum tick_dep_bits bit)
269 {
270         int prev;
271 
272         prev = atomic_fetch_or(BIT(bit), dep);
273         if (!prev)
274                 tick_nohz_full_kick_all();
275 }
276 
277 /*
278  * Set a global tick dependency. Used by perf events that rely on freq and
279  * by unstable clock.
280  */
281 void tick_nohz_dep_set(enum tick_dep_bits bit)
282 {
283         tick_nohz_dep_set_all(&tick_dep_mask, bit);
284 }
285 
286 void tick_nohz_dep_clear(enum tick_dep_bits bit)
287 {
288         atomic_andnot(BIT(bit), &tick_dep_mask);
289 }
290 
291 /*
292  * Set per-CPU tick dependency. Used by scheduler and perf events in order to
293  * manage events throttling.
294  */
295 void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
296 {
297         int prev;
298         struct tick_sched *ts;
299 
300         ts = per_cpu_ptr(&tick_cpu_sched, cpu);
301 
302         prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
303         if (!prev) {
304                 preempt_disable();
305                 /* Perf needs local kick that is NMI safe */
306                 if (cpu == smp_processor_id()) {
307                         tick_nohz_full_kick();
308                 } else {
309                         /* Remote irq work not NMI-safe */
310                         if (!WARN_ON_ONCE(in_nmi()))
311                                 tick_nohz_full_kick_cpu(cpu);
312                 }
313                 preempt_enable();
314         }
315 }
316 
317 void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
318 {
319         struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
320 
321         atomic_andnot(BIT(bit), &ts->tick_dep_mask);
322 }
323 
324 /*
325  * Set a per-task tick dependency. Posix CPU timers need this in order to elapse
326  * per task timers.
327  */
328 void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
329 {
330         /*
331          * We could optimize this with just kicking the target running the task
332          * if that noise matters for nohz full users.
333          */
334         tick_nohz_dep_set_all(&tsk->tick_dep_mask, bit);
335 }
336 
337 void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
338 {
339         atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
340 }
341 
342 /*
343  * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
344  * per process timers.
345  */
346 void tick_nohz_dep_set_signal(struct signal_struct *sig, enum tick_dep_bits bit)
347 {
348         tick_nohz_dep_set_all(&sig->tick_dep_mask, bit);
349 }
350 
351 void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
352 {
353         atomic_andnot(BIT(bit), &sig->tick_dep_mask);
354 }
355 
356 /*
357  * Re-evaluate the need for the tick as we switch the current task.
358  * It might need the tick due to per task/process properties:
359  * perf events, posix CPU timers, ...
360  */
361 void __tick_nohz_task_switch(void)
362 {
363         unsigned long flags;
364         struct tick_sched *ts;
365 
366         local_irq_save(flags);
367 
368         if (!tick_nohz_full_cpu(smp_processor_id()))
369                 goto out;
370 
371         ts = this_cpu_ptr(&tick_cpu_sched);
372 
373         if (ts->tick_stopped) {
374                 if (atomic_read(&current->tick_dep_mask) ||
375                     atomic_read(&current->signal->tick_dep_mask))
376                         tick_nohz_full_kick();
377         }
378 out:
379         local_irq_restore(flags);
380 }
381 
382 /* Parse the boot-time nohz CPU list from the kernel parameters. */
383 static int __init tick_nohz_full_setup(char *str)
384 {
385         alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
386         if (cpulist_parse(str, tick_nohz_full_mask) < 0) {
387                 pr_warn("NO_HZ: Incorrect nohz_full cpumask\n");
388                 free_bootmem_cpumask_var(tick_nohz_full_mask);
389                 return 1;
390         }
391         tick_nohz_full_running = true;
392 
393         return 1;
394 }
395 __setup("nohz_full=", tick_nohz_full_setup);
396 
397 static int tick_nohz_cpu_down(unsigned int cpu)
398 {
399         /*
400          * The boot CPU handles housekeeping duty (unbound timers,
401          * workqueues, timekeeping, ...) on behalf of full dynticks
402          * CPUs. It must remain online when nohz full is enabled.
403          */
404         if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
405                 return -EBUSY;
406         return 0;
407 }
408 
409 static int tick_nohz_init_all(void)
410 {
411         int err = -1;
412 
413 #ifdef CONFIG_NO_HZ_FULL_ALL
414         if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) {
415                 WARN(1, "NO_HZ: Can't allocate full dynticks cpumask\n");
416                 return err;
417         }
418         err = 0;
419         cpumask_setall(tick_nohz_full_mask);
420         tick_nohz_full_running = true;
421 #endif
422         return err;
423 }
424 
425 void __init tick_nohz_init(void)
426 {
427         int cpu, ret;
428 
429         if (!tick_nohz_full_running) {
430                 if (tick_nohz_init_all() < 0)
431                         return;
432         }
433 
434         if (!alloc_cpumask_var(&housekeeping_mask, GFP_KERNEL)) {
435                 WARN(1, "NO_HZ: Can't allocate not-full dynticks cpumask\n");
436                 cpumask_clear(tick_nohz_full_mask);
437                 tick_nohz_full_running = false;
438                 return;
439         }
440 
441         /*
442          * Full dynticks uses irq work to drive the tick rescheduling on safe
443          * locking contexts. But then we need irq work to raise its own
444          * interrupts to avoid circular dependency on the tick
445          */
446         if (!arch_irq_work_has_interrupt()) {
447                 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
448                 cpumask_clear(tick_nohz_full_mask);
449                 cpumask_copy(housekeeping_mask, cpu_possible_mask);
450                 tick_nohz_full_running = false;
451                 return;
452         }
453 
454         cpu = smp_processor_id();
455 
456         if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
457                 pr_warn("NO_HZ: Clearing %d from nohz_full range for timekeeping\n",
458                         cpu);
459                 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
460         }
461 
462         cpumask_andnot(housekeeping_mask,
463                        cpu_possible_mask, tick_nohz_full_mask);
464 
465         for_each_cpu(cpu, tick_nohz_full_mask)
466                 context_tracking_cpu_set(cpu);
467 
468         ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
469                                         "kernel/nohz:predown", NULL,
470                                         tick_nohz_cpu_down);
471         WARN_ON(ret < 0);
472         pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
473                 cpumask_pr_args(tick_nohz_full_mask));
474 
475         /*
476          * We need at least one CPU to handle housekeeping work such
477          * as timekeeping, unbound timers, workqueues, ...
478          */
479         WARN_ON_ONCE(cpumask_empty(housekeeping_mask));
480 }
481 #endif
482 
483 /*
484  * NOHZ - aka dynamic tick functionality
485  */
486 #ifdef CONFIG_NO_HZ_COMMON
487 /*
488  * NO HZ enabled ?
489  */
490 bool tick_nohz_enabled __read_mostly  = true;
491 unsigned long tick_nohz_active  __read_mostly;
492 /*
493  * Enable / Disable tickless mode
494  */
495 static int __init setup_tick_nohz(char *str)
496 {
497         return (kstrtobool(str, &tick_nohz_enabled) == 0);
498 }
499 
500 __setup("nohz=", setup_tick_nohz);
501 
502 int tick_nohz_tick_stopped(void)
503 {
504         return __this_cpu_read(tick_cpu_sched.tick_stopped);
505 }
506 
507 /**
508  * tick_nohz_update_jiffies - update jiffies when idle was interrupted
509  *
510  * Called from interrupt entry when the CPU was idle
511  *
512  * In case the sched_tick was stopped on this CPU, we have to check if jiffies
513  * must be updated. Otherwise an interrupt handler could use a stale jiffy
514  * value. We do this unconditionally on any CPU, as we don't know whether the
515  * CPU, which has the update task assigned is in a long sleep.
516  */
517 static void tick_nohz_update_jiffies(ktime_t now)
518 {
519         unsigned long flags;
520 
521         __this_cpu_write(tick_cpu_sched.idle_waketime, now);
522 
523         local_irq_save(flags);
524         tick_do_update_jiffies64(now);
525         local_irq_restore(flags);
526 
527         touch_softlockup_watchdog_sched();
528 }
529 
530 /*
531  * Updates the per-CPU time idle statistics counters
532  */
533 static void
534 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
535 {
536         ktime_t delta;
537 
538         if (ts->idle_active) {
539                 delta = ktime_sub(now, ts->idle_entrytime);
540                 if (nr_iowait_cpu(cpu) > 0)
541                         ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
542                 else
543                         ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
544                 ts->idle_entrytime = now;
545         }
546 
547         if (last_update_time)
548                 *last_update_time = ktime_to_us(now);
549 
550 }
551 
552 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
553 {
554         update_ts_time_stats(smp_processor_id(), ts, now, NULL);
555         ts->idle_active = 0;
556 
557         sched_clock_idle_wakeup_event(0);
558 }
559 
560 static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
561 {
562         ktime_t now = ktime_get();
563 
564         ts->idle_entrytime = now;
565         ts->idle_active = 1;
566         sched_clock_idle_sleep_event();
567         return now;
568 }
569 
570 /**
571  * get_cpu_idle_time_us - get the total idle time of a CPU
572  * @cpu: CPU number to query
573  * @last_update_time: variable to store update time in. Do not update
574  * counters if NULL.
575  *
576  * Return the cumulative idle time (since boot) for a given
577  * CPU, in microseconds.
578  *
579  * This time is measured via accounting rather than sampling,
580  * and is as accurate as ktime_get() is.
581  *
582  * This function returns -1 if NOHZ is not enabled.
583  */
584 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
585 {
586         struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
587         ktime_t now, idle;
588 
589         if (!tick_nohz_active)
590                 return -1;
591 
592         now = ktime_get();
593         if (last_update_time) {
594                 update_ts_time_stats(cpu, ts, now, last_update_time);
595                 idle = ts->idle_sleeptime;
596         } else {
597                 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
598                         ktime_t delta = ktime_sub(now, ts->idle_entrytime);
599 
600                         idle = ktime_add(ts->idle_sleeptime, delta);
601                 } else {
602                         idle = ts->idle_sleeptime;
603                 }
604         }
605 
606         return ktime_to_us(idle);
607 
608 }
609 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
610 
611 /**
612  * get_cpu_iowait_time_us - get the total iowait time of a CPU
613  * @cpu: CPU number to query
614  * @last_update_time: variable to store update time in. Do not update
615  * counters if NULL.
616  *
617  * Return the cumulative iowait time (since boot) for a given
618  * CPU, in microseconds.
619  *
620  * This time is measured via accounting rather than sampling,
621  * and is as accurate as ktime_get() is.
622  *
623  * This function returns -1 if NOHZ is not enabled.
624  */
625 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
626 {
627         struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
628         ktime_t now, iowait;
629 
630         if (!tick_nohz_active)
631                 return -1;
632 
633         now = ktime_get();
634         if (last_update_time) {
635                 update_ts_time_stats(cpu, ts, now, last_update_time);
636                 iowait = ts->iowait_sleeptime;
637         } else {
638                 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
639                         ktime_t delta = ktime_sub(now, ts->idle_entrytime);
640 
641                         iowait = ktime_add(ts->iowait_sleeptime, delta);
642                 } else {
643                         iowait = ts->iowait_sleeptime;
644                 }
645         }
646 
647         return ktime_to_us(iowait);
648 }
649 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
650 
651 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
652 {
653         hrtimer_cancel(&ts->sched_timer);
654         hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
655 
656         /* Forward the time to expire in the future */
657         hrtimer_forward(&ts->sched_timer, now, tick_period);
658 
659         if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
660                 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
661         else
662                 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
663 }
664 
665 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
666                                          ktime_t now, int cpu)
667 {
668         struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
669         u64 basemono, next_tick, next_tmr, next_rcu, delta, expires;
670         unsigned long seq, basejiff;
671         ktime_t tick;
672 
673         /* Read jiffies and the time when jiffies were updated last */
674         do {
675                 seq = read_seqbegin(&jiffies_lock);
676                 basemono = last_jiffies_update;
677                 basejiff = jiffies;
678         } while (read_seqretry(&jiffies_lock, seq));
679         ts->last_jiffies = basejiff;
680 
681         if (rcu_needs_cpu(basemono, &next_rcu) ||
682             arch_needs_cpu() || irq_work_needs_cpu()) {
683                 next_tick = basemono + TICK_NSEC;
684         } else {
685                 /*
686                  * Get the next pending timer. If high resolution
687                  * timers are enabled this only takes the timer wheel
688                  * timers into account. If high resolution timers are
689                  * disabled this also looks at the next expiring
690                  * hrtimer.
691                  */
692                 next_tmr = get_next_timer_interrupt(basejiff, basemono);
693                 ts->next_timer = next_tmr;
694                 /* Take the next rcu event into account */
695                 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr;
696         }
697 
698         /*
699          * If the tick is due in the next period, keep it ticking or
700          * force prod the timer.
701          */
702         delta = next_tick - basemono;
703         if (delta <= (u64)TICK_NSEC) {
704                 tick = 0;
705 
706                 /*
707                  * Tell the timer code that the base is not idle, i.e. undo
708                  * the effect of get_next_timer_interrupt():
709                  */
710                 timer_clear_idle();
711                 /*
712                  * We've not stopped the tick yet, and there's a timer in the
713                  * next period, so no point in stopping it either, bail.
714                  */
715                 if (!ts->tick_stopped)
716                         goto out;
717 
718                 /*
719                  * If, OTOH, we did stop it, but there's a pending (expired)
720                  * timer reprogram the timer hardware to fire now.
721                  *
722                  * We will not restart the tick proper, just prod the timer
723                  * hardware into firing an interrupt to process the pending
724                  * timers. Just like tick_irq_exit() will not restart the tick
725                  * for 'normal' interrupts.
726                  *
727                  * Only once we exit the idle loop will we re-enable the tick,
728                  * see tick_nohz_idle_exit().
729                  */
730                 if (delta == 0) {
731                         tick_nohz_restart(ts, now);
732                         goto out;
733                 }
734         }
735 
736         /*
737          * If this CPU is the one which updates jiffies, then give up
738          * the assignment and let it be taken by the CPU which runs
739          * the tick timer next, which might be this CPU as well. If we
740          * don't drop this here the jiffies might be stale and
741          * do_timer() never invoked. Keep track of the fact that it
742          * was the one which had the do_timer() duty last. If this CPU
743          * is the one which had the do_timer() duty last, we limit the
744          * sleep time to the timekeeping max_deferment value.
745          * Otherwise we can sleep as long as we want.
746          */
747         delta = timekeeping_max_deferment();
748         if (cpu == tick_do_timer_cpu) {
749                 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
750                 ts->do_timer_last = 1;
751         } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
752                 delta = KTIME_MAX;
753                 ts->do_timer_last = 0;
754         } else if (!ts->do_timer_last) {
755                 delta = KTIME_MAX;
756         }
757 
758 #ifdef CONFIG_NO_HZ_FULL
759         /* Limit the tick delta to the maximum scheduler deferment */
760         if (!ts->inidle)
761                 delta = min(delta, scheduler_tick_max_deferment());
762 #endif
763 
764         /* Calculate the next expiry time */
765         if (delta < (KTIME_MAX - basemono))
766                 expires = basemono + delta;
767         else
768                 expires = KTIME_MAX;
769 
770         expires = min_t(u64, expires, next_tick);
771         tick = expires;
772 
773         /* Skip reprogram of event if its not changed */
774         if (ts->tick_stopped && (expires == dev->next_event))
775                 goto out;
776 
777         /*
778          * nohz_stop_sched_tick can be called several times before
779          * the nohz_restart_sched_tick is called. This happens when
780          * interrupts arrive which do not cause a reschedule. In the
781          * first call we save the current tick time, so we can restart
782          * the scheduler tick in nohz_restart_sched_tick.
783          */
784         if (!ts->tick_stopped) {
785                 nohz_balance_enter_idle(cpu);
786                 calc_load_enter_idle();
787                 cpu_load_update_nohz_start();
788 
789                 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
790                 ts->tick_stopped = 1;
791                 trace_tick_stop(1, TICK_DEP_MASK_NONE);
792         }
793 
794         /*
795          * If the expiration time == KTIME_MAX, then we simply stop
796          * the tick timer.
797          */
798         if (unlikely(expires == KTIME_MAX)) {
799                 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
800                         hrtimer_cancel(&ts->sched_timer);
801                 goto out;
802         }
803 
804         if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
805                 hrtimer_start(&ts->sched_timer, tick, HRTIMER_MODE_ABS_PINNED);
806         else
807                 tick_program_event(tick, 1);
808 out:
809         /* Update the estimated sleep length */
810         ts->sleep_length = ktime_sub(dev->next_event, now);
811         return tick;
812 }
813 
814 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
815 {
816         /* Update jiffies first */
817         tick_do_update_jiffies64(now);
818         cpu_load_update_nohz_stop();
819 
820         /*
821          * Clear the timer idle flag, so we avoid IPIs on remote queueing and
822          * the clock forward checks in the enqueue path:
823          */
824         timer_clear_idle();
825 
826         calc_load_exit_idle();
827         touch_softlockup_watchdog_sched();
828         /*
829          * Cancel the scheduled timer and restore the tick
830          */
831         ts->tick_stopped  = 0;
832         ts->idle_exittime = now;
833 
834         tick_nohz_restart(ts, now);
835 }
836 
837 static void tick_nohz_full_update_tick(struct tick_sched *ts)
838 {
839 #ifdef CONFIG_NO_HZ_FULL
840         int cpu = smp_processor_id();
841 
842         if (!tick_nohz_full_cpu(cpu))
843                 return;
844 
845         if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
846                 return;
847 
848         if (can_stop_full_tick(cpu, ts))
849                 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu);
850         else if (ts->tick_stopped)
851                 tick_nohz_restart_sched_tick(ts, ktime_get());
852 #endif
853 }
854 
855 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
856 {
857         /*
858          * If this CPU is offline and it is the one which updates
859          * jiffies, then give up the assignment and let it be taken by
860          * the CPU which runs the tick timer next. If we don't drop
861          * this here the jiffies might be stale and do_timer() never
862          * invoked.
863          */
864         if (unlikely(!cpu_online(cpu))) {
865                 if (cpu == tick_do_timer_cpu)
866                         tick_do_timer_cpu = TICK_DO_TIMER_NONE;
867                 return false;
868         }
869 
870         if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) {
871                 ts->sleep_length = NSEC_PER_SEC / HZ;
872                 return false;
873         }
874 
875         if (need_resched())
876                 return false;
877 
878         if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
879                 static int ratelimit;
880 
881                 if (ratelimit < 10 &&
882                     (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
883                         pr_warn("NOHZ: local_softirq_pending %02x\n",
884                                 (unsigned int) local_softirq_pending());
885                         ratelimit++;
886                 }
887                 return false;
888         }
889 
890         if (tick_nohz_full_enabled()) {
891                 /*
892                  * Keep the tick alive to guarantee timekeeping progression
893                  * if there are full dynticks CPUs around
894                  */
895                 if (tick_do_timer_cpu == cpu)
896                         return false;
897                 /*
898                  * Boot safety: make sure the timekeeping duty has been
899                  * assigned before entering dyntick-idle mode,
900                  */
901                 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
902                         return false;
903         }
904 
905         return true;
906 }
907 
908 static void __tick_nohz_idle_enter(struct tick_sched *ts)
909 {
910         ktime_t now, expires;
911         int cpu = smp_processor_id();
912 
913         now = tick_nohz_start_idle(ts);
914 
915         if (can_stop_idle_tick(cpu, ts)) {
916                 int was_stopped = ts->tick_stopped;
917 
918                 ts->idle_calls++;
919 
920                 expires = tick_nohz_stop_sched_tick(ts, now, cpu);
921                 if (expires > 0LL) {
922                         ts->idle_sleeps++;
923                         ts->idle_expires = expires;
924                 }
925 
926                 if (!was_stopped && ts->tick_stopped)
927                         ts->idle_jiffies = ts->last_jiffies;
928         }
929 }
930 
931 /**
932  * tick_nohz_idle_enter - stop the idle tick from the idle task
933  *
934  * When the next event is more than a tick into the future, stop the idle tick
935  * Called when we start the idle loop.
936  *
937  * The arch is responsible of calling:
938  *
939  * - rcu_idle_enter() after its last use of RCU before the CPU is put
940  *  to sleep.
941  * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
942  */
943 void tick_nohz_idle_enter(void)
944 {
945         struct tick_sched *ts;
946 
947         WARN_ON_ONCE(irqs_disabled());
948 
949         /*
950          * Update the idle state in the scheduler domain hierarchy
951          * when tick_nohz_stop_sched_tick() is called from the idle loop.
952          * State will be updated to busy during the first busy tick after
953          * exiting idle.
954          */
955         set_cpu_sd_state_idle();
956 
957         local_irq_disable();
958 
959         ts = this_cpu_ptr(&tick_cpu_sched);
960         ts->inidle = 1;
961         __tick_nohz_idle_enter(ts);
962 
963         local_irq_enable();
964 }
965 
966 /**
967  * tick_nohz_irq_exit - update next tick event from interrupt exit
968  *
969  * When an interrupt fires while we are idle and it doesn't cause
970  * a reschedule, it may still add, modify or delete a timer, enqueue
971  * an RCU callback, etc...
972  * So we need to re-calculate and reprogram the next tick event.
973  */
974 void tick_nohz_irq_exit(void)
975 {
976         struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
977 
978         if (ts->inidle)
979                 __tick_nohz_idle_enter(ts);
980         else
981                 tick_nohz_full_update_tick(ts);
982 }
983 
984 /**
985  * tick_nohz_get_sleep_length - return the length of the current sleep
986  *
987  * Called from power state control code with interrupts disabled
988  */
989 ktime_t tick_nohz_get_sleep_length(void)
990 {
991         struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
992 
993         return ts->sleep_length;
994 }
995 
996 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
997 {
998 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
999         unsigned long ticks;
1000 
1001         if (vtime_accounting_cpu_enabled())
1002                 return;
1003         /*
1004          * We stopped the tick in idle. Update process times would miss the
1005          * time we slept as update_process_times does only a 1 tick
1006          * accounting. Enforce that this is accounted to idle !
1007          */
1008         ticks = jiffies - ts->idle_jiffies;
1009         /*
1010          * We might be one off. Do not randomly account a huge number of ticks!
1011          */
1012         if (ticks && ticks < LONG_MAX)
1013                 account_idle_ticks(ticks);
1014 #endif
1015 }
1016 
1017 /**
1018  * tick_nohz_idle_exit - restart the idle tick from the idle task
1019  *
1020  * Restart the idle tick when the CPU is woken up from idle
1021  * This also exit the RCU extended quiescent state. The CPU
1022  * can use RCU again after this function is called.
1023  */
1024 void tick_nohz_idle_exit(void)
1025 {
1026         struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1027         ktime_t now;
1028 
1029         local_irq_disable();
1030 
1031         WARN_ON_ONCE(!ts->inidle);
1032 
1033         ts->inidle = 0;
1034 
1035         if (ts->idle_active || ts->tick_stopped)
1036                 now = ktime_get();
1037 
1038         if (ts->idle_active)
1039                 tick_nohz_stop_idle(ts, now);
1040 
1041         if (ts->tick_stopped) {
1042                 tick_nohz_restart_sched_tick(ts, now);
1043                 tick_nohz_account_idle_ticks(ts);
1044         }
1045 
1046         local_irq_enable();
1047 }
1048 
1049 /*
1050  * The nohz low res interrupt handler
1051  */
1052 static void tick_nohz_handler(struct clock_event_device *dev)
1053 {
1054         struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1055         struct pt_regs *regs = get_irq_regs();
1056         ktime_t now = ktime_get();
1057 
1058         dev->next_event = KTIME_MAX;
1059 
1060         tick_sched_do_timer(now);
1061         tick_sched_handle(ts, regs);
1062 
1063         /* No need to reprogram if we are running tickless  */
1064         if (unlikely(ts->tick_stopped))
1065                 return;
1066 
1067         hrtimer_forward(&ts->sched_timer, now, tick_period);
1068         tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1069 }
1070 
1071 static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1072 {
1073         if (!tick_nohz_enabled)
1074                 return;
1075         ts->nohz_mode = mode;
1076         /* One update is enough */
1077         if (!test_and_set_bit(0, &tick_nohz_active))
1078                 timers_update_migration(true);
1079 }
1080 
1081 /**
1082  * tick_nohz_switch_to_nohz - switch to nohz mode
1083  */
1084 static void tick_nohz_switch_to_nohz(void)
1085 {
1086         struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1087         ktime_t next;
1088 
1089         if (!tick_nohz_enabled)
1090                 return;
1091 
1092         if (tick_switch_to_oneshot(tick_nohz_handler))
1093                 return;
1094 
1095         /*
1096          * Recycle the hrtimer in ts, so we can share the
1097          * hrtimer_forward with the highres code.
1098          */
1099         hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1100         /* Get the next period */
1101         next = tick_init_jiffy_update();
1102 
1103         hrtimer_set_expires(&ts->sched_timer, next);
1104         hrtimer_forward_now(&ts->sched_timer, tick_period);
1105         tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1106         tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1107 }
1108 
1109 static inline void tick_nohz_irq_enter(void)
1110 {
1111         struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1112         ktime_t now;
1113 
1114         if (!ts->idle_active && !ts->tick_stopped)
1115                 return;
1116         now = ktime_get();
1117         if (ts->idle_active)
1118                 tick_nohz_stop_idle(ts, now);
1119         if (ts->tick_stopped)
1120                 tick_nohz_update_jiffies(now);
1121 }
1122 
1123 #else
1124 
1125 static inline void tick_nohz_switch_to_nohz(void) { }
1126 static inline void tick_nohz_irq_enter(void) { }
1127 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1128 
1129 #endif /* CONFIG_NO_HZ_COMMON */
1130 
1131 /*
1132  * Called from irq_enter to notify about the possible interruption of idle()
1133  */
1134 void tick_irq_enter(void)
1135 {
1136         tick_check_oneshot_broadcast_this_cpu();
1137         tick_nohz_irq_enter();
1138 }
1139 
1140 /*
1141  * High resolution timer specific code
1142  */
1143 #ifdef CONFIG_HIGH_RES_TIMERS
1144 /*
1145  * We rearm the timer until we get disabled by the idle code.
1146  * Called with interrupts disabled.
1147  */
1148 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1149 {
1150         struct tick_sched *ts =
1151                 container_of(timer, struct tick_sched, sched_timer);
1152         struct pt_regs *regs = get_irq_regs();
1153         ktime_t now = ktime_get();
1154 
1155         tick_sched_do_timer(now);
1156 
1157         /*
1158          * Do not call, when we are not in irq context and have
1159          * no valid regs pointer
1160          */
1161         if (regs)
1162                 tick_sched_handle(ts, regs);
1163 
1164         /* No need to reprogram if we are in idle or full dynticks mode */
1165         if (unlikely(ts->tick_stopped))
1166                 return HRTIMER_NORESTART;
1167 
1168         hrtimer_forward(timer, now, tick_period);
1169 
1170         return HRTIMER_RESTART;
1171 }
1172 
1173 static int sched_skew_tick;
1174 
1175 static int __init skew_tick(char *str)
1176 {
1177         get_option(&str, &sched_skew_tick);
1178 
1179         return 0;
1180 }
1181 early_param("skew_tick", skew_tick);
1182 
1183 /**
1184  * tick_setup_sched_timer - setup the tick emulation timer
1185  */
1186 void tick_setup_sched_timer(void)
1187 {
1188         struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1189         ktime_t now = ktime_get();
1190 
1191         /*
1192          * Emulate tick processing via per-CPU hrtimers:
1193          */
1194         hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1195         ts->sched_timer.function = tick_sched_timer;
1196 
1197         /* Get the next period (per-CPU) */
1198         hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1199 
1200         /* Offset the tick to avert jiffies_lock contention. */
1201         if (sched_skew_tick) {
1202                 u64 offset = ktime_to_ns(tick_period) >> 1;
1203                 do_div(offset, num_possible_cpus());
1204                 offset *= smp_processor_id();
1205                 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1206         }
1207 
1208         hrtimer_forward(&ts->sched_timer, now, tick_period);
1209         hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
1210         tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1211 }
1212 #endif /* HIGH_RES_TIMERS */
1213 
1214 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1215 void tick_cancel_sched_timer(int cpu)
1216 {
1217         struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1218 
1219 # ifdef CONFIG_HIGH_RES_TIMERS
1220         if (ts->sched_timer.base)
1221                 hrtimer_cancel(&ts->sched_timer);
1222 # endif
1223 
1224         memset(ts, 0, sizeof(*ts));
1225 }
1226 #endif
1227 
1228 /**
1229  * Async notification about clocksource changes
1230  */
1231 void tick_clock_notify(void)
1232 {
1233         int cpu;
1234 
1235         for_each_possible_cpu(cpu)
1236                 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1237 }
1238 
1239 /*
1240  * Async notification about clock event changes
1241  */
1242 void tick_oneshot_notify(void)
1243 {
1244         struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1245 
1246         set_bit(0, &ts->check_clocks);
1247 }
1248 
1249 /**
1250  * Check, if a change happened, which makes oneshot possible.
1251  *
1252  * Called cyclic from the hrtimer softirq (driven by the timer
1253  * softirq) allow_nohz signals, that we can switch into low-res nohz
1254  * mode, because high resolution timers are disabled (either compile
1255  * or runtime). Called with interrupts disabled.
1256  */
1257 int tick_check_oneshot_change(int allow_nohz)
1258 {
1259         struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1260 
1261         if (!test_and_clear_bit(0, &ts->check_clocks))
1262                 return 0;
1263 
1264         if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1265                 return 0;
1266 
1267         if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1268                 return 0;
1269 
1270         if (!allow_nohz)
1271                 return 1;
1272 
1273         tick_nohz_switch_to_nohz();
1274         return 0;
1275 }
1276 

~ [ 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