Line data Source code
1 : // SPDX-License-Identifier: GPL-2.0
2 : /*
3 : * This file contains the base functions to manage periodic tick
4 : * related events.
5 : *
6 : * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
7 : * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
8 : * Copyright(C) 2006-2007, Timesys Corp., Thomas Gleixner
9 : */
10 : #include <linux/cpu.h>
11 : #include <linux/err.h>
12 : #include <linux/hrtimer.h>
13 : #include <linux/interrupt.h>
14 : #include <linux/nmi.h>
15 : #include <linux/percpu.h>
16 : #include <linux/profile.h>
17 : #include <linux/sched.h>
18 : #include <linux/module.h>
19 : #include <trace/events/power.h>
20 :
21 : #include <asm/irq_regs.h>
22 :
23 : #include "tick-internal.h"
24 :
25 : /*
26 : * Tick devices
27 : */
28 : DEFINE_PER_CPU(struct tick_device, tick_cpu_device);
29 : /*
30 : * Tick next event: keeps track of the tick time. It's updated by the
31 : * CPU which handles the tick and protected by jiffies_lock. There is
32 : * no requirement to write hold the jiffies seqcount for it.
33 : */
34 : ktime_t tick_next_period;
35 :
36 : /*
37 : * tick_do_timer_cpu is a timer core internal variable which holds the CPU NR
38 : * which is responsible for calling do_timer(), i.e. the timekeeping stuff. This
39 : * variable has two functions:
40 : *
41 : * 1) Prevent a thundering herd issue of a gazillion of CPUs trying to grab the
42 : * timekeeping lock all at once. Only the CPU which is assigned to do the
43 : * update is handling it.
44 : *
45 : * 2) Hand off the duty in the NOHZ idle case by setting the value to
46 : * TICK_DO_TIMER_NONE, i.e. a non existing CPU. So the next cpu which looks
47 : * at it will take over and keep the time keeping alive. The handover
48 : * procedure also covers cpu hotplug.
49 : */
50 : int tick_do_timer_cpu __read_mostly = TICK_DO_TIMER_BOOT;
51 : #ifdef CONFIG_NO_HZ_FULL
52 : /*
53 : * tick_do_timer_boot_cpu indicates the boot CPU temporarily owns
54 : * tick_do_timer_cpu and it should be taken over by an eligible secondary
55 : * when one comes online.
56 : */
57 : static int tick_do_timer_boot_cpu __read_mostly = -1;
58 : #endif
59 :
60 : /*
61 : * Debugging: see timer_list.c
62 : */
63 0 : struct tick_device *tick_get_device(int cpu)
64 : {
65 0 : return &per_cpu(tick_cpu_device, cpu);
66 : }
67 :
68 : /**
69 : * tick_is_oneshot_available - check for a oneshot capable event device
70 : */
71 0 : int tick_is_oneshot_available(void)
72 : {
73 0 : struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
74 :
75 0 : if (!dev || !(dev->features & CLOCK_EVT_FEAT_ONESHOT))
76 : return 0;
77 0 : if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
78 : return 1;
79 0 : return tick_broadcast_oneshot_available();
80 : }
81 :
82 : /*
83 : * Periodic tick
84 : */
85 2943 : static void tick_periodic(int cpu)
86 : {
87 2943 : if (tick_do_timer_cpu == cpu) {
88 2943 : raw_spin_lock(&jiffies_lock);
89 5886 : write_seqcount_begin(&jiffies_seq);
90 :
91 : /* Keep track of the next tick event */
92 2943 : tick_next_period = ktime_add_ns(tick_next_period, TICK_NSEC);
93 :
94 2943 : do_timer(1);
95 5886 : write_seqcount_end(&jiffies_seq);
96 2943 : raw_spin_unlock(&jiffies_lock);
97 2943 : update_wall_time();
98 : }
99 :
100 2943 : update_process_times(user_mode(get_irq_regs()));
101 2943 : profile_tick(CPU_PROFILING);
102 2943 : }
103 :
104 : /*
105 : * Event handler for periodic ticks
106 : */
107 2943 : void tick_handle_periodic(struct clock_event_device *dev)
108 : {
109 2943 : int cpu = smp_processor_id();
110 2943 : ktime_t next = dev->next_event;
111 :
112 2943 : tick_periodic(cpu);
113 :
114 : #if defined(CONFIG_HIGH_RES_TIMERS) || defined(CONFIG_NO_HZ_COMMON)
115 : /*
116 : * The cpu might have transitioned to HIGHRES or NOHZ mode via
117 : * update_process_times() -> run_local_timers() ->
118 : * hrtimer_run_queues().
119 : */
120 : if (dev->event_handler != tick_handle_periodic)
121 : return;
122 : #endif
123 :
124 2943 : if (!clockevent_state_oneshot(dev))
125 : return;
126 : for (;;) {
127 : /*
128 : * Setup the next period for devices, which do not have
129 : * periodic mode:
130 : */
131 0 : next = ktime_add_ns(next, TICK_NSEC);
132 :
133 0 : if (!clockevents_program_event(dev, next, false))
134 : return;
135 : /*
136 : * Have to be careful here. If we're in oneshot mode,
137 : * before we call tick_periodic() in a loop, we need
138 : * to be sure we're using a real hardware clocksource.
139 : * Otherwise we could get trapped in an infinite
140 : * loop, as the tick_periodic() increments jiffies,
141 : * which then will increment time, possibly causing
142 : * the loop to trigger again and again.
143 : */
144 0 : if (timekeeping_valid_for_hres())
145 0 : tick_periodic(cpu);
146 : }
147 : }
148 :
149 : /*
150 : * Setup the device for a periodic tick
151 : */
152 1 : void tick_setup_periodic(struct clock_event_device *dev, int broadcast)
153 : {
154 1 : tick_set_periodic_handler(dev, broadcast);
155 :
156 : /* Broadcast setup ? */
157 2 : if (!tick_device_is_functional(dev))
158 : return;
159 :
160 1 : if ((dev->features & CLOCK_EVT_FEAT_PERIODIC) &&
161 : !tick_broadcast_oneshot_active()) {
162 1 : clockevents_switch_state(dev, CLOCK_EVT_STATE_PERIODIC);
163 : } else {
164 : unsigned int seq;
165 : ktime_t next;
166 :
167 : do {
168 0 : seq = read_seqcount_begin(&jiffies_seq);
169 0 : next = tick_next_period;
170 0 : } while (read_seqcount_retry(&jiffies_seq, seq));
171 :
172 0 : clockevents_switch_state(dev, CLOCK_EVT_STATE_ONESHOT);
173 :
174 : for (;;) {
175 0 : if (!clockevents_program_event(dev, next, false))
176 : return;
177 0 : next = ktime_add_ns(next, TICK_NSEC);
178 : }
179 : }
180 : }
181 :
182 : #ifdef CONFIG_NO_HZ_FULL
183 : static void giveup_do_timer(void *info)
184 : {
185 : int cpu = *(unsigned int *)info;
186 :
187 : WARN_ON(tick_do_timer_cpu != smp_processor_id());
188 :
189 : tick_do_timer_cpu = cpu;
190 : }
191 :
192 : static void tick_take_do_timer_from_boot(void)
193 : {
194 : int cpu = smp_processor_id();
195 : int from = tick_do_timer_boot_cpu;
196 :
197 : if (from >= 0 && from != cpu)
198 : smp_call_function_single(from, giveup_do_timer, &cpu, 1);
199 : }
200 : #endif
201 :
202 : /*
203 : * Setup the tick device
204 : */
205 1 : static void tick_setup_device(struct tick_device *td,
206 : struct clock_event_device *newdev, int cpu,
207 : const struct cpumask *cpumask)
208 : {
209 1 : void (*handler)(struct clock_event_device *) = NULL;
210 1 : ktime_t next_event = 0;
211 :
212 : /*
213 : * First device setup ?
214 : */
215 1 : if (!td->evtdev) {
216 : /*
217 : * If no cpu took the do_timer update, assign it to
218 : * this cpu:
219 : */
220 1 : if (tick_do_timer_cpu == TICK_DO_TIMER_BOOT) {
221 : ktime_t next_p;
222 : u32 rem;
223 :
224 1 : tick_do_timer_cpu = cpu;
225 :
226 1 : next_p = ktime_get();
227 2 : div_u64_rem(next_p, TICK_NSEC, &rem);
228 1 : if (rem) {
229 0 : next_p -= rem;
230 0 : next_p += TICK_NSEC;
231 : }
232 :
233 1 : tick_next_period = next_p;
234 : #ifdef CONFIG_NO_HZ_FULL
235 : /*
236 : * The boot CPU may be nohz_full, in which case set
237 : * tick_do_timer_boot_cpu so the first housekeeping
238 : * secondary that comes up will take do_timer from
239 : * us.
240 : */
241 : if (tick_nohz_full_cpu(cpu))
242 : tick_do_timer_boot_cpu = cpu;
243 :
244 : } else if (tick_do_timer_boot_cpu != -1 &&
245 : !tick_nohz_full_cpu(cpu)) {
246 : tick_take_do_timer_from_boot();
247 : tick_do_timer_boot_cpu = -1;
248 : WARN_ON(tick_do_timer_cpu != cpu);
249 : #endif
250 : }
251 :
252 : /*
253 : * Startup in periodic mode first.
254 : */
255 1 : td->mode = TICKDEV_MODE_PERIODIC;
256 : } else {
257 0 : handler = td->evtdev->event_handler;
258 0 : next_event = td->evtdev->next_event;
259 0 : td->evtdev->event_handler = clockevents_handle_noop;
260 : }
261 :
262 1 : td->evtdev = newdev;
263 :
264 : /*
265 : * When the device is not per cpu, pin the interrupt to the
266 : * current cpu:
267 : */
268 2 : if (!cpumask_equal(newdev->cpumask, cpumask))
269 : irq_set_affinity(newdev->irq, cpumask);
270 :
271 : /*
272 : * When global broadcasting is active, check if the current
273 : * device is registered as a placeholder for broadcast mode.
274 : * This allows us to handle this x86 misfeature in a generic
275 : * way. This function also returns !=0 when we keep the
276 : * current active broadcast state for this CPU.
277 : */
278 1 : if (tick_device_uses_broadcast(newdev, cpu))
279 : return;
280 :
281 1 : if (td->mode == TICKDEV_MODE_PERIODIC)
282 1 : tick_setup_periodic(newdev, 0);
283 : else
284 0 : tick_setup_oneshot(newdev, handler, next_event);
285 : }
286 :
287 0 : void tick_install_replacement(struct clock_event_device *newdev)
288 : {
289 0 : struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
290 0 : int cpu = smp_processor_id();
291 :
292 0 : clockevents_exchange_device(td->evtdev, newdev);
293 0 : tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
294 : if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
295 : tick_oneshot_notify();
296 0 : }
297 :
298 1 : static bool tick_check_percpu(struct clock_event_device *curdev,
299 : struct clock_event_device *newdev, int cpu)
300 : {
301 2 : if (!cpumask_test_cpu(cpu, newdev->cpumask))
302 : return false;
303 3 : if (cpumask_equal(newdev->cpumask, cpumask_of(cpu)))
304 : return true;
305 : /* Check if irq affinity can be set */
306 0 : if (newdev->irq >= 0 && !irq_can_set_affinity(newdev->irq))
307 : return false;
308 : /* Prefer an existing cpu local device */
309 0 : if (curdev && cpumask_equal(curdev->cpumask, cpumask_of(cpu)))
310 : return false;
311 : return true;
312 : }
313 :
314 1 : static bool tick_check_preferred(struct clock_event_device *curdev,
315 : struct clock_event_device *newdev)
316 : {
317 : /* Prefer oneshot capable device */
318 1 : if (!(newdev->features & CLOCK_EVT_FEAT_ONESHOT)) {
319 0 : if (curdev && (curdev->features & CLOCK_EVT_FEAT_ONESHOT))
320 : return false;
321 : if (tick_oneshot_mode_active())
322 : return false;
323 : }
324 :
325 : /*
326 : * Use the higher rated one, but prefer a CPU local device with a lower
327 : * rating than a non-CPU local device
328 : */
329 0 : return !curdev ||
330 1 : newdev->rating > curdev->rating ||
331 0 : !cpumask_equal(curdev->cpumask, newdev->cpumask);
332 : }
333 :
334 : /*
335 : * Check whether the new device is a better fit than curdev. curdev
336 : * can be NULL !
337 : */
338 1 : bool tick_check_replacement(struct clock_event_device *curdev,
339 : struct clock_event_device *newdev)
340 : {
341 1 : if (!tick_check_percpu(curdev, newdev, smp_processor_id()))
342 : return false;
343 :
344 1 : return tick_check_preferred(curdev, newdev);
345 : }
346 :
347 : /*
348 : * Check, if the new registered device should be used. Called with
349 : * clockevents_lock held and interrupts disabled.
350 : */
351 1 : void tick_check_new_device(struct clock_event_device *newdev)
352 : {
353 : struct clock_event_device *curdev;
354 : struct tick_device *td;
355 : int cpu;
356 :
357 1 : cpu = smp_processor_id();
358 1 : td = &per_cpu(tick_cpu_device, cpu);
359 1 : curdev = td->evtdev;
360 :
361 1 : if (!tick_check_replacement(curdev, newdev))
362 : goto out_bc;
363 :
364 1 : if (!try_module_get(newdev->owner))
365 : return;
366 :
367 : /*
368 : * Replace the eventually existing device by the new
369 : * device. If the current device is the broadcast device, do
370 : * not give it back to the clockevents layer !
371 : */
372 1 : if (tick_is_broadcast_device(curdev)) {
373 : clockevents_shutdown(curdev);
374 : curdev = NULL;
375 : }
376 1 : clockevents_exchange_device(curdev, newdev);
377 2 : tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
378 : if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
379 : tick_oneshot_notify();
380 : return;
381 :
382 : out_bc:
383 : /*
384 : * Can the new device be used as a broadcast device ?
385 : */
386 : tick_install_broadcast_device(newdev, cpu);
387 : }
388 :
389 : /**
390 : * tick_broadcast_oneshot_control - Enter/exit broadcast oneshot mode
391 : * @state: The target state (enter/exit)
392 : *
393 : * The system enters/leaves a state, where affected devices might stop
394 : * Returns 0 on success, -EBUSY if the cpu is used to broadcast wakeups.
395 : *
396 : * Called with interrupts disabled, so clockevents_lock is not
397 : * required here because the local clock event device cannot go away
398 : * under us.
399 : */
400 0 : int tick_broadcast_oneshot_control(enum tick_broadcast_state state)
401 : {
402 0 : struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
403 :
404 0 : if (!(td->evtdev->features & CLOCK_EVT_FEAT_C3STOP))
405 : return 0;
406 :
407 0 : return __tick_broadcast_oneshot_control(state);
408 : }
409 : EXPORT_SYMBOL_GPL(tick_broadcast_oneshot_control);
410 :
411 : #ifdef CONFIG_HOTPLUG_CPU
412 : /*
413 : * Transfer the do_timer job away from a dying cpu.
414 : *
415 : * Called with interrupts disabled. No locking required. If
416 : * tick_do_timer_cpu is owned by this cpu, nothing can change it.
417 : */
418 : void tick_handover_do_timer(void)
419 : {
420 : if (tick_do_timer_cpu == smp_processor_id())
421 : tick_do_timer_cpu = cpumask_first(cpu_online_mask);
422 : }
423 :
424 : /*
425 : * Shutdown an event device on a given cpu:
426 : *
427 : * This is called on a life CPU, when a CPU is dead. So we cannot
428 : * access the hardware device itself.
429 : * We just set the mode and remove it from the lists.
430 : */
431 : void tick_shutdown(unsigned int cpu)
432 : {
433 : struct tick_device *td = &per_cpu(tick_cpu_device, cpu);
434 : struct clock_event_device *dev = td->evtdev;
435 :
436 : td->mode = TICKDEV_MODE_PERIODIC;
437 : if (dev) {
438 : /*
439 : * Prevent that the clock events layer tries to call
440 : * the set mode function!
441 : */
442 : clockevent_set_state(dev, CLOCK_EVT_STATE_DETACHED);
443 : clockevents_exchange_device(dev, NULL);
444 : dev->event_handler = clockevents_handle_noop;
445 : td->evtdev = NULL;
446 : }
447 : }
448 : #endif
449 :
450 : /**
451 : * tick_suspend_local - Suspend the local tick device
452 : *
453 : * Called from the local cpu for freeze with interrupts disabled.
454 : *
455 : * No locks required. Nothing can change the per cpu device.
456 : */
457 0 : void tick_suspend_local(void)
458 : {
459 0 : struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
460 :
461 0 : clockevents_shutdown(td->evtdev);
462 0 : }
463 :
464 : /**
465 : * tick_resume_local - Resume the local tick device
466 : *
467 : * Called from the local CPU for unfreeze or XEN resume magic.
468 : *
469 : * No locks required. Nothing can change the per cpu device.
470 : */
471 0 : void tick_resume_local(void)
472 : {
473 0 : struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
474 0 : bool broadcast = tick_resume_check_broadcast();
475 :
476 0 : clockevents_tick_resume(td->evtdev);
477 : if (!broadcast) {
478 0 : if (td->mode == TICKDEV_MODE_PERIODIC)
479 0 : tick_setup_periodic(td->evtdev, 0);
480 : else
481 0 : tick_resume_oneshot();
482 : }
483 :
484 : /*
485 : * Ensure that hrtimers are up to date and the clockevents device
486 : * is reprogrammed correctly when high resolution timers are
487 : * enabled.
488 : */
489 0 : hrtimers_resume_local();
490 0 : }
491 :
492 : /**
493 : * tick_suspend - Suspend the tick and the broadcast device
494 : *
495 : * Called from syscore_suspend() via timekeeping_suspend with only one
496 : * CPU online and interrupts disabled or from tick_unfreeze() under
497 : * tick_freeze_lock.
498 : *
499 : * No locks required. Nothing can change the per cpu device.
500 : */
501 0 : void tick_suspend(void)
502 : {
503 : tick_suspend_local();
504 : tick_suspend_broadcast();
505 0 : }
506 :
507 : /**
508 : * tick_resume - Resume the tick and the broadcast device
509 : *
510 : * Called from syscore_resume() via timekeeping_resume with only one
511 : * CPU online and interrupts disabled.
512 : *
513 : * No locks required. Nothing can change the per cpu device.
514 : */
515 0 : void tick_resume(void)
516 : {
517 : tick_resume_broadcast();
518 0 : tick_resume_local();
519 0 : }
520 :
521 : #ifdef CONFIG_SUSPEND
522 : static DEFINE_RAW_SPINLOCK(tick_freeze_lock);
523 : static unsigned int tick_freeze_depth;
524 :
525 : /**
526 : * tick_freeze - Suspend the local tick and (possibly) timekeeping.
527 : *
528 : * Check if this is the last online CPU executing the function and if so,
529 : * suspend timekeeping. Otherwise suspend the local tick.
530 : *
531 : * Call with interrupts disabled. Must be balanced with %tick_unfreeze().
532 : * Interrupts must not be enabled before the subsequent %tick_unfreeze().
533 : */
534 0 : void tick_freeze(void)
535 : {
536 0 : raw_spin_lock(&tick_freeze_lock);
537 :
538 0 : tick_freeze_depth++;
539 0 : if (tick_freeze_depth == num_online_cpus()) {
540 0 : trace_suspend_resume(TPS("timekeeping_freeze"),
541 : smp_processor_id(), true);
542 0 : system_state = SYSTEM_SUSPEND;
543 : sched_clock_suspend();
544 0 : timekeeping_suspend();
545 : } else {
546 : tick_suspend_local();
547 : }
548 :
549 0 : raw_spin_unlock(&tick_freeze_lock);
550 0 : }
551 :
552 : /**
553 : * tick_unfreeze - Resume the local tick and (possibly) timekeeping.
554 : *
555 : * Check if this is the first CPU executing the function and if so, resume
556 : * timekeeping. Otherwise resume the local tick.
557 : *
558 : * Call with interrupts disabled. Must be balanced with %tick_freeze().
559 : * Interrupts must not be enabled after the preceding %tick_freeze().
560 : */
561 0 : void tick_unfreeze(void)
562 : {
563 0 : raw_spin_lock(&tick_freeze_lock);
564 :
565 0 : if (tick_freeze_depth == num_online_cpus()) {
566 0 : timekeeping_resume();
567 : sched_clock_resume();
568 0 : system_state = SYSTEM_RUNNING;
569 : trace_suspend_resume(TPS("timekeeping_freeze"),
570 : smp_processor_id(), false);
571 : } else {
572 : touch_softlockup_watchdog();
573 0 : tick_resume_local();
574 : }
575 :
576 0 : tick_freeze_depth--;
577 :
578 0 : raw_spin_unlock(&tick_freeze_lock);
579 0 : }
580 : #endif /* CONFIG_SUSPEND */
581 :
582 : /**
583 : * tick_init - initialize the tick control
584 : */
585 1 : void __init tick_init(void)
586 : {
587 : tick_broadcast_init();
588 : tick_nohz_init();
589 1 : }
|
