projects/linux·blob master · refs

595 lines · 15.7 KiB

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, Linutronix GmbH, Thomas Gleixner <[email protected]>
7 * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
8 * Copyright(C) 2006-2007, Timesys Corp., Thomas Gleixner
9 */
10#include <linux/compiler.h>
11#include <linux/cpu.h>
12#include <linux/err.h>
13#include <linux/hrtimer.h>
14#include <linux/interrupt.h>
15#include <linux/nmi.h>
16#include <linux/percpu.h>
17#include <linux/profile.h>
18#include <linux/sched.h>
19#include <linux/module.h>
20#include <trace/events/power.h>
21
22#include <asm/irq_regs.h>
23
24#include "tick-internal.h"
25
26/*
27 * Tick devices
28 */
29DEFINE_PER_CPU(struct tick_device, tick_cpu_device);
30/*
31 * Tick next event: keeps track of the tick time. It's updated by the
32 * CPU which handles the tick and protected by jiffies_lock. There is
33 * no requirement to write hold the jiffies seqcount for it.
34 */
35ktime_t tick_next_period;
36
37/*
38 * tick_do_timer_cpu is a timer core internal variable which holds the CPU NR
39 * which is responsible for calling do_timer(), i.e. the timekeeping stuff. This
40 * variable has two functions:
41 *
42 * 1) Prevent a thundering herd issue of a gazillion of CPUs trying to grab the
43 * timekeeping lock all at once. Only the CPU which is assigned to do the
44 * update is handling it.
45 *
46 * 2) Hand off the duty in the NOHZ idle case by setting the value to
47 * TICK_DO_TIMER_NONE, i.e. a non existing CPU. So the next cpu which looks
48 * at it will take over and keep the time keeping alive. The handover
49 * procedure also covers cpu hotplug.
50 */
51int tick_do_timer_cpu __read_mostly = TICK_DO_TIMER_BOOT;
52#ifdef CONFIG_NO_HZ_FULL
53/*
54 * tick_do_timer_boot_cpu indicates the boot CPU temporarily owns
55 * tick_do_timer_cpu and it should be taken over by an eligible secondary
56 * when one comes online.
57 */
58static int tick_do_timer_boot_cpu __read_mostly = -1;
59#endif
60
61/*
62 * Debugging: see timer_list.c
63 */
64struct tick_device *tick_get_device(int cpu)
65{
66 return &per_cpu(tick_cpu_device, cpu);
67}
68
69/**
70 * tick_is_oneshot_available - check for a oneshot capable event device
71 */
72int tick_is_oneshot_available(void)
73{
74 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
75
76 if (!dev || !(dev->features & CLOCK_EVT_FEAT_ONESHOT))
77 return 0;
78 if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
79 return 1;
80 return tick_broadcast_oneshot_available();
81}
82
83/*
84 * Periodic tick
85 */
86static void tick_periodic(int cpu)
87{
88 if (READ_ONCE(tick_do_timer_cpu) == cpu) {
89 raw_spin_lock(&jiffies_lock);
90 write_seqcount_begin(&jiffies_seq);
91
92 /* Keep track of the next tick event */
93 tick_next_period = ktime_add_ns(tick_next_period, TICK_NSEC);
94
95 do_timer(1);
96 write_seqcount_end(&jiffies_seq);
97 raw_spin_unlock(&jiffies_lock);
98 update_wall_time();
99 }
100
101 update_process_times(user_mode(get_irq_regs()));
102 profile_tick(CPU_PROFILING);
103}
104
105/*
106 * Event handler for periodic ticks
107 */
108void tick_handle_periodic(struct clock_event_device *dev)
109{
110 int cpu = smp_processor_id();
111 ktime_t next = dev->next_event;
112
113 dev->next_event_forced = 0;
114 tick_periodic(cpu);
115
116 /*
117 * The cpu might have transitioned to HIGHRES or NOHZ mode via
118 * update_process_times() -> run_local_timers() ->
119 * hrtimer_run_queues().
120 */
121 if (IS_ENABLED(CONFIG_TICK_ONESHOT) && dev->event_handler != tick_handle_periodic)
122 return;
123
124 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 next = ktime_add_ns(next, TICK_NSEC);
132
133 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 if (timekeeping_valid_for_hres())
145 tick_periodic(cpu);
146 }
147}
148
149/*
150 * Setup the device for a periodic tick
151 */
152void tick_setup_periodic(struct clock_event_device *dev, int broadcast)
153{
154 tick_set_periodic_handler(dev, broadcast);
155
156 /* Broadcast setup ? */
157 if (!tick_device_is_functional(dev))
158 return;
159
160 if ((dev->features & CLOCK_EVT_FEAT_PERIODIC) &&
161 !tick_broadcast_oneshot_active()) {
162 clockevents_switch_state(dev, CLOCK_EVT_STATE_PERIODIC);
163 } else {
164 unsigned int seq;
165 ktime_t next;
166
167 do {
168 seq = read_seqcount_begin(&jiffies_seq);
169 next = tick_next_period;
170 } while (read_seqcount_retry(&jiffies_seq, seq));
171
172 clockevents_switch_state(dev, CLOCK_EVT_STATE_ONESHOT);
173
174 for (;;) {
175 if (!clockevents_program_event(dev, next, false))
176 return;
177 next = ktime_add_ns(next, TICK_NSEC);
178 }
179 }
180}
181
182/*
183 * Setup the tick device
184 */
185static void tick_setup_device(struct tick_device *td,
186 struct clock_event_device *newdev, int cpu,
187 const struct cpumask *cpumask)
188{
189 void (*handler)(struct clock_event_device *) = NULL;
190 ktime_t next_event = 0;
191
192 /*
193 * First device setup ?
194 */
195 if (!td->evtdev) {
196 /*
197 * If no cpu took the do_timer update, assign it to
198 * this cpu:
199 */
200 if (READ_ONCE(tick_do_timer_cpu) == TICK_DO_TIMER_BOOT) {
201 WRITE_ONCE(tick_do_timer_cpu, cpu);
202 tick_next_period = ktime_get();
203#ifdef CONFIG_NO_HZ_FULL
204 /*
205 * The boot CPU may be nohz_full, in which case the
206 * first housekeeping secondary will take do_timer()
207 * from it.
208 */
209 if (tick_nohz_full_cpu(cpu))
210 tick_do_timer_boot_cpu = cpu;
211
212 } else if (tick_do_timer_boot_cpu != -1 && !tick_nohz_full_cpu(cpu)) {
213 tick_do_timer_boot_cpu = -1;
214 /*
215 * The boot CPU will stay in periodic (NOHZ disabled)
216 * mode until clocksource_done_booting() called after
217 * smp_init() selects a high resolution clocksource and
218 * timekeeping_notify() kicks the NOHZ stuff alive.
219 *
220 * So this WRITE_ONCE can only race with the READ_ONCE
221 * check in tick_periodic() but this race is harmless.
222 */
223 WRITE_ONCE(tick_do_timer_cpu, cpu);
224#endif
225 }
226
227 /*
228 * Startup in periodic mode first.
229 */
230 td->mode = TICKDEV_MODE_PERIODIC;
231 } else {
232 handler = td->evtdev->event_handler;
233 next_event = td->evtdev->next_event;
234 td->evtdev->event_handler = clockevents_handle_noop;
235 }
236
237 td->evtdev = newdev;
238
239 /*
240 * When the device is not per cpu, pin the interrupt to the
241 * current cpu:
242 */
243 if (!cpumask_equal(newdev->cpumask, cpumask))
244 irq_set_affinity(newdev->irq, cpumask);
245
246 /*
247 * When global broadcasting is active, check if the current
248 * device is registered as a placeholder for broadcast mode.
249 * This allows us to handle this x86 misfeature in a generic
250 * way. This function also returns !=0 when we keep the
251 * current active broadcast state for this CPU.
252 */
253 if (tick_device_uses_broadcast(newdev, cpu))
254 return;
255
256 if (td->mode == TICKDEV_MODE_PERIODIC)
257 tick_setup_periodic(newdev, 0);
258 else
259 tick_setup_oneshot(newdev, handler, next_event);
260}
261
262void tick_install_replacement(struct clock_event_device *newdev)
263{
264 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
265 int cpu = smp_processor_id();
266
267 clockevents_exchange_device(td->evtdev, newdev);
268 tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
269 if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
270 tick_oneshot_notify();
271}
272
273static bool tick_check_percpu(struct clock_event_device *curdev,
274 struct clock_event_device *newdev, int cpu)
275{
276 if (!cpumask_test_cpu(cpu, newdev->cpumask))
277 return false;
278 if (cpumask_equal(newdev->cpumask, cpumask_of(cpu)))
279 return true;
280 /* Check if irq affinity can be set */
281 if (newdev->irq >= 0 && !irq_can_set_affinity(newdev->irq))
282 return false;
283 /* Prefer an existing cpu local device */
284 if (curdev && cpumask_equal(curdev->cpumask, cpumask_of(cpu)))
285 return false;
286 return true;
287}
288
289static bool tick_check_preferred(struct clock_event_device *curdev,
290 struct clock_event_device *newdev)
291{
292 /* Prefer oneshot capable device */
293 if (!(newdev->features & CLOCK_EVT_FEAT_ONESHOT)) {
294 if (curdev && (curdev->features & CLOCK_EVT_FEAT_ONESHOT))
295 return false;
296 if (tick_oneshot_mode_active())
297 return false;
298 }
299
300 /*
301 * Use the higher rated one, but prefer a CPU local device with a lower
302 * rating than a non-CPU local device
303 */
304 return !curdev ||
305 newdev->rating > curdev->rating ||
306 !cpumask_equal(curdev->cpumask, newdev->cpumask);
307}
308
309/*
310 * Check whether the new device is a better fit than curdev. curdev
311 * can be NULL !
312 */
313bool tick_check_replacement(struct clock_event_device *curdev,
314 struct clock_event_device *newdev)
315{
316 if (!tick_check_percpu(curdev, newdev, smp_processor_id()))
317 return false;
318
319 return tick_check_preferred(curdev, newdev);
320}
321
322/*
323 * Check, if the new registered device should be used. Called with
324 * clockevents_lock held and interrupts disabled.
325 */
326void tick_check_new_device(struct clock_event_device *newdev)
327{
328 struct clock_event_device *curdev;
329 struct tick_device *td;
330 int cpu;
331
332 cpu = smp_processor_id();
333 td = &per_cpu(tick_cpu_device, cpu);
334 curdev = td->evtdev;
335
336 if (!tick_check_replacement(curdev, newdev))
337 goto out_bc;
338
339 if (!try_module_get(newdev->owner))
340 return;
341
342 /*
343 * Replace the eventually existing device by the new
344 * device. If the current device is the broadcast device, do
345 * not give it back to the clockevents layer !
346 */
347 if (tick_is_broadcast_device(curdev)) {
348 clockevents_shutdown(curdev);
349 curdev = NULL;
350 }
351 clockevents_exchange_device(curdev, newdev);
352 tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
353 if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
354 tick_oneshot_notify();
355 return;
356
357out_bc:
358 /*
359 * Can the new device be used as a broadcast device ?
360 */
361 tick_install_broadcast_device(newdev, cpu);
362}
363
364/**
365 * tick_broadcast_oneshot_control - Enter/exit broadcast oneshot mode
366 * @state: The target state (enter/exit)
367 *
368 * The system enters/leaves a state, where affected devices might stop
369 * Returns 0 on success, -EBUSY if the cpu is used to broadcast wakeups.
370 *
371 * Called with interrupts disabled, so clockevents_lock is not
372 * required here because the local clock event device cannot go away
373 * under us.
374 */
375int tick_broadcast_oneshot_control(enum tick_broadcast_state state)
376{
377 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
378
379 if (!(td->evtdev->features & CLOCK_EVT_FEAT_C3STOP))
380 return 0;
381
382 return __tick_broadcast_oneshot_control(state);
383}
384EXPORT_SYMBOL_GPL(tick_broadcast_oneshot_control);
385
386#ifdef CONFIG_HOTPLUG_CPU
387void tick_assert_timekeeping_handover(void)
388{
389 WARN_ON_ONCE(tick_do_timer_cpu == smp_processor_id());
390}
391/*
392 * Stop the tick and transfer the timekeeping job away from a dying cpu.
393 */
394int tick_cpu_dying(unsigned int dying_cpu)
395{
396 /*
397 * If the current CPU is the timekeeper, it's the only one that can
398 * safely hand over its duty. Also all online CPUs are in stop
399 * machine, guaranteed not to be idle, therefore there is no
400 * concurrency and it's safe to pick any online successor.
401 */
402 if (tick_do_timer_cpu == dying_cpu)
403 tick_do_timer_cpu = cpumask_first(cpu_online_mask);
404
405 /* Make sure the CPU won't try to retake the timekeeping duty */
406 tick_sched_timer_dying(dying_cpu);
407
408 /* Remove CPU from timer broadcasting */
409 tick_offline_cpu(dying_cpu);
410
411 return 0;
412}
413
414/*
415 * Shutdown an event device on the outgoing CPU:
416 *
417 * Called by the dying CPU during teardown, with clockevents_lock held
418 * and interrupts disabled.
419 */
420void tick_shutdown(void)
421{
422 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
423 struct clock_event_device *dev = td->evtdev;
424
425 td->mode = TICKDEV_MODE_PERIODIC;
426 if (dev) {
427 clockevents_exchange_device(dev, NULL);
428 dev->event_handler = clockevents_handle_noop;
429 td->evtdev = NULL;
430 }
431}
432#endif
433
434/**
435 * tick_suspend_local - Suspend the local tick device
436 *
437 * Called from the local cpu for freeze with interrupts disabled.
438 *
439 * No locks required. Nothing can change the per cpu device.
440 */
441void tick_suspend_local(void)
442{
443 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
444
445 clockevents_shutdown(td->evtdev);
446}
447
448/**
449 * tick_resume_local - Resume the local tick device
450 *
451 * Called from the local CPU for unfreeze or XEN resume magic.
452 *
453 * No locks required. Nothing can change the per cpu device.
454 */
455void tick_resume_local(void)
456{
457 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
458 bool broadcast = tick_resume_check_broadcast();
459
460 clockevents_tick_resume(td->evtdev);
461 if (!broadcast) {
462 if (td->mode == TICKDEV_MODE_PERIODIC)
463 tick_setup_periodic(td->evtdev, 0);
464 else
465 tick_resume_oneshot();
466 }
467
468 /*
469 * Ensure that hrtimers are up to date and the clockevents device
470 * is reprogrammed correctly when high resolution timers are
471 * enabled.
472 */
473 hrtimers_resume_local();
474}
475
476/**
477 * tick_suspend - Suspend the tick and the broadcast device
478 *
479 * Called from syscore_suspend() via timekeeping_suspend with only one
480 * CPU online and interrupts disabled or from tick_unfreeze() under
481 * tick_freeze_lock.
482 *
483 * No locks required. Nothing can change the per cpu device.
484 */
485void tick_suspend(void)
486{
487 tick_suspend_local();
488 tick_suspend_broadcast();
489}
490
491/**
492 * tick_resume - Resume the tick and the broadcast device
493 *
494 * Called from syscore_resume() via timekeeping_resume with only one
495 * CPU online and interrupts disabled.
496 *
497 * No locks required. Nothing can change the per cpu device.
498 */
499void tick_resume(void)
500{
501 tick_resume_broadcast();
502 tick_resume_local();
503}
504
505#ifdef CONFIG_SUSPEND
506static DEFINE_RAW_SPINLOCK(tick_freeze_lock);
507static DEFINE_WAIT_OVERRIDE_MAP(tick_freeze_map, LD_WAIT_SLEEP);
508static unsigned int tick_freeze_depth;
509
510/**
511 * tick_freeze - Suspend the local tick and (possibly) timekeeping.
512 *
513 * Check if this is the last online CPU executing the function and if so,
514 * suspend timekeeping. Otherwise suspend the local tick.
515 *
516 * Call with interrupts disabled. Must be balanced with %tick_unfreeze().
517 * Interrupts must not be enabled before the subsequent %tick_unfreeze().
518 */
519void tick_freeze(void)
520{
521 raw_spin_lock(&tick_freeze_lock);
522
523 tick_freeze_depth++;
524 if (tick_freeze_depth == num_online_cpus()) {
525 trace_suspend_resume(TPS("timekeeping_freeze"),
526 smp_processor_id(), true);
527 /*
528 * All other CPUs have their interrupts disabled and are
529 * suspended to idle. Other tasks have been frozen so there
530 * is no scheduling happening. This means that there is no
531 * concurrency in the system at this point. Therefore it is
532 * okay to acquire a sleeping lock on PREEMPT_RT, such as a
533 * spinlock, because the lock cannot be held by other CPUs
534 * or threads and acquiring it cannot block.
535 *
536 * Inform lockdep about the situation.
537 */
538 lock_map_acquire_try(&tick_freeze_map);
539 system_state = SYSTEM_SUSPEND;
540 sched_clock_suspend();
541 timekeeping_suspend();
542 lock_map_release(&tick_freeze_map);
543 } else {
544 tick_suspend_local();
545 }
546
547 raw_spin_unlock(&tick_freeze_lock);
548}
549
550/**
551 * tick_unfreeze - Resume the local tick and (possibly) timekeeping.
552 *
553 * Check if this is the first CPU executing the function and if so, resume
554 * timekeeping. Otherwise resume the local tick.
555 *
556 * Call with interrupts disabled. Must be balanced with %tick_freeze().
557 * Interrupts must not be enabled after the preceding %tick_freeze().
558 */
559void tick_unfreeze(void)
560{
561 raw_spin_lock(&tick_freeze_lock);
562
563 if (tick_freeze_depth == num_online_cpus()) {
564 /*
565 * Similar to tick_freeze(). On resumption the first CPU may
566 * acquire uncontended sleeping locks while other CPUs block on
567 * tick_freeze_lock.
568 */
569 lock_map_acquire_try(&tick_freeze_map);
570 timekeeping_resume();
571 sched_clock_resume();
572 lock_map_release(&tick_freeze_map);
573
574 system_state = SYSTEM_RUNNING;
575 trace_suspend_resume(TPS("timekeeping_freeze"),
576 smp_processor_id(), false);
577 } else {
578 touch_softlockup_watchdog();
579 tick_resume_local();
580 }
581
582 tick_freeze_depth--;
583
584 raw_spin_unlock(&tick_freeze_lock);
585}
586#endif /* CONFIG_SUSPEND */
587
588/**
589 * tick_init - initialize the tick control
590 */
591void __init tick_init(void)
592{
593 tick_broadcast_init();
594 tick_nohz_init();
595}