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/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(ticks: 1);
96 write_seqcount_end(&jiffies_seq);
97 raw_spin_unlock(&jiffies_lock);
98 update_wall_time();
99 }
100
101 update_process_times(user: user_mode(regs: 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 tick_periodic(cpu);
114
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 (IS_ENABLED(CONFIG_TICK_ONESHOT) && dev->event_handler != tick_handle_periodic)
121 return;
122
123 if (!clockevent_state_oneshot(dev))
124 return;
125 for (;;) {
126 /*
127 * Setup the next period for devices, which do not have
128 * periodic mode:
129 */
130 next = ktime_add_ns(next, TICK_NSEC);
131
132 if (!clockevents_program_event(dev, expires: next, force: false))
133 return;
134 /*
135 * Have to be careful here. If we're in oneshot mode,
136 * before we call tick_periodic() in a loop, we need
137 * to be sure we're using a real hardware clocksource.
138 * Otherwise we could get trapped in an infinite
139 * loop, as the tick_periodic() increments jiffies,
140 * which then will increment time, possibly causing
141 * the loop to trigger again and again.
142 */
143 if (timekeeping_valid_for_hres())
144 tick_periodic(cpu);
145 }
146}
147
148/*
149 * Setup the device for a periodic tick
150 */
151void tick_setup_periodic(struct clock_event_device *dev, int broadcast)
152{
153 tick_set_periodic_handler(dev, broadcast);
154
155 /* Broadcast setup ? */
156 if (!tick_device_is_functional(dev))
157 return;
158
159 if ((dev->features & CLOCK_EVT_FEAT_PERIODIC) &&
160 !tick_broadcast_oneshot_active()) {
161 clockevents_switch_state(dev, state: CLOCK_EVT_STATE_PERIODIC);
162 } else {
163 unsigned int seq;
164 ktime_t next;
165
166 do {
167 seq = read_seqcount_begin(&jiffies_seq);
168 next = tick_next_period;
169 } while (read_seqcount_retry(&jiffies_seq, seq));
170
171 clockevents_switch_state(dev, state: CLOCK_EVT_STATE_ONESHOT);
172
173 for (;;) {
174 if (!clockevents_program_event(dev, expires: next, force: false))
175 return;
176 next = ktime_add_ns(next, TICK_NSEC);
177 }
178 }
179}
180
181/*
182 * Setup the tick device
183 */
184static void tick_setup_device(struct tick_device *td,
185 struct clock_event_device *newdev, int cpu,
186 const struct cpumask *cpumask)
187{
188 void (*handler)(struct clock_event_device *) = NULL;
189 ktime_t next_event = 0;
190
191 /*
192 * First device setup ?
193 */
194 if (!td->evtdev) {
195 /*
196 * If no cpu took the do_timer update, assign it to
197 * this cpu:
198 */
199 if (READ_ONCE(tick_do_timer_cpu) == TICK_DO_TIMER_BOOT) {
200 WRITE_ONCE(tick_do_timer_cpu, cpu);
201 tick_next_period = ktime_get();
202#ifdef CONFIG_NO_HZ_FULL
203 /*
204 * The boot CPU may be nohz_full, in which case the
205 * first housekeeping secondary will take do_timer()
206 * from it.
207 */
208 if (tick_nohz_full_cpu(cpu))
209 tick_do_timer_boot_cpu = cpu;
210
211 } else if (tick_do_timer_boot_cpu != -1 && !tick_nohz_full_cpu(cpu)) {
212 tick_do_timer_boot_cpu = -1;
213 /*
214 * The boot CPU will stay in periodic (NOHZ disabled)
215 * mode until clocksource_done_booting() called after
216 * smp_init() selects a high resolution clocksource and
217 * timekeeping_notify() kicks the NOHZ stuff alive.
218 *
219 * So this WRITE_ONCE can only race with the READ_ONCE
220 * check in tick_periodic() but this race is harmless.
221 */
222 WRITE_ONCE(tick_do_timer_cpu, cpu);
223#endif
224 }
225
226 /*
227 * Startup in periodic mode first.
228 */
229 td->mode = TICKDEV_MODE_PERIODIC;
230 } else {
231 handler = td->evtdev->event_handler;
232 next_event = td->evtdev->next_event;
233 td->evtdev->event_handler = clockevents_handle_noop;
234 }
235
236 td->evtdev = newdev;
237
238 /*
239 * When the device is not per cpu, pin the interrupt to the
240 * current cpu:
241 */
242 if (!cpumask_equal(src1p: newdev->cpumask, src2p: cpumask))
243 irq_set_affinity(irq: newdev->irq, cpumask);
244
245 /*
246 * When global broadcasting is active, check if the current
247 * device is registered as a placeholder for broadcast mode.
248 * This allows us to handle this x86 misfeature in a generic
249 * way. This function also returns !=0 when we keep the
250 * current active broadcast state for this CPU.
251 */
252 if (tick_device_uses_broadcast(dev: newdev, cpu))
253 return;
254
255 if (td->mode == TICKDEV_MODE_PERIODIC)
256 tick_setup_periodic(dev: newdev, broadcast: 0);
257 else
258 tick_setup_oneshot(newdev, handler, nextevt: next_event);
259}
260
261void tick_install_replacement(struct clock_event_device *newdev)
262{
263 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
264 int cpu = smp_processor_id();
265
266 clockevents_exchange_device(old: td->evtdev, new: newdev);
267 tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
268 if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
269 tick_oneshot_notify();
270}
271
272static bool tick_check_percpu(struct clock_event_device *curdev,
273 struct clock_event_device *newdev, int cpu)
274{
275 if (!cpumask_test_cpu(cpu, cpumask: newdev->cpumask))
276 return false;
277 if (cpumask_equal(src1p: newdev->cpumask, cpumask_of(cpu)))
278 return true;
279 /* Check if irq affinity can be set */
280 if (newdev->irq >= 0 && !irq_can_set_affinity(irq: newdev->irq))
281 return false;
282 /* Prefer an existing cpu local device */
283 if (curdev && cpumask_equal(src1p: curdev->cpumask, cpumask_of(cpu)))
284 return false;
285 return true;
286}
287
288static bool tick_check_preferred(struct clock_event_device *curdev,
289 struct clock_event_device *newdev)
290{
291 /* Prefer oneshot capable device */
292 if (!(newdev->features & CLOCK_EVT_FEAT_ONESHOT)) {
293 if (curdev && (curdev->features & CLOCK_EVT_FEAT_ONESHOT))
294 return false;
295 if (tick_oneshot_mode_active())
296 return false;
297 }
298
299 /*
300 * Use the higher rated one, but prefer a CPU local device with a lower
301 * rating than a non-CPU local device
302 */
303 return !curdev ||
304 newdev->rating > curdev->rating ||
305 !cpumask_equal(src1p: curdev->cpumask, src2p: newdev->cpumask);
306}
307
308/*
309 * Check whether the new device is a better fit than curdev. curdev
310 * can be NULL !
311 */
312bool tick_check_replacement(struct clock_event_device *curdev,
313 struct clock_event_device *newdev)
314{
315 if (!tick_check_percpu(curdev, newdev, smp_processor_id()))
316 return false;
317
318 return tick_check_preferred(curdev, newdev);
319}
320
321/*
322 * Check, if the new registered device should be used. Called with
323 * clockevents_lock held and interrupts disabled.
324 */
325void tick_check_new_device(struct clock_event_device *newdev)
326{
327 struct clock_event_device *curdev;
328 struct tick_device *td;
329 int cpu;
330
331 cpu = smp_processor_id();
332 td = &per_cpu(tick_cpu_device, cpu);
333 curdev = td->evtdev;
334
335 if (!tick_check_replacement(curdev, newdev))
336 goto out_bc;
337
338 if (!try_module_get(module: newdev->owner))
339 return;
340
341 /*
342 * Replace the eventually existing device by the new
343 * device. If the current device is the broadcast device, do
344 * not give it back to the clockevents layer !
345 */
346 if (tick_is_broadcast_device(dev: curdev)) {
347 clockevents_shutdown(dev: curdev);
348 curdev = NULL;
349 }
350 clockevents_exchange_device(old: curdev, new: newdev);
351 tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
352 if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
353 tick_oneshot_notify();
354 return;
355
356out_bc:
357 /*
358 * Can the new device be used as a broadcast device ?
359 */
360 tick_install_broadcast_device(dev: newdev, cpu);
361}
362
363/**
364 * tick_broadcast_oneshot_control - Enter/exit broadcast oneshot mode
365 * @state: The target state (enter/exit)
366 *
367 * The system enters/leaves a state, where affected devices might stop
368 * Returns 0 on success, -EBUSY if the cpu is used to broadcast wakeups.
369 *
370 * Called with interrupts disabled, so clockevents_lock is not
371 * required here because the local clock event device cannot go away
372 * under us.
373 */
374int tick_broadcast_oneshot_control(enum tick_broadcast_state state)
375{
376 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
377
378 if (!(td->evtdev->features & CLOCK_EVT_FEAT_C3STOP))
379 return 0;
380
381 return __tick_broadcast_oneshot_control(state);
382}
383EXPORT_SYMBOL_GPL(tick_broadcast_oneshot_control);
384
385#ifdef CONFIG_HOTPLUG_CPU
386void tick_assert_timekeeping_handover(void)
387{
388 WARN_ON_ONCE(tick_do_timer_cpu == smp_processor_id());
389}
390/*
391 * Stop the tick and transfer the timekeeping job away from a dying cpu.
392 */
393int tick_cpu_dying(unsigned int dying_cpu)
394{
395 /*
396 * If the current CPU is the timekeeper, it's the only one that can
397 * safely hand over its duty. Also all online CPUs are in stop
398 * machine, guaranteed not to be idle, therefore there is no
399 * concurrency and it's safe to pick any online successor.
400 */
401 if (tick_do_timer_cpu == dying_cpu)
402 tick_do_timer_cpu = cpumask_first(cpu_online_mask);
403
404 /* Make sure the CPU won't try to retake the timekeeping duty */
405 tick_sched_timer_dying(cpu: dying_cpu);
406
407 /* Remove CPU from timer broadcasting */
408 tick_offline_cpu(cpu: dying_cpu);
409
410 return 0;
411}
412
413/*
414 * Shutdown an event device on the outgoing CPU:
415 *
416 * Called by the dying CPU during teardown, with clockevents_lock held
417 * and interrupts disabled.
418 */
419void tick_shutdown(void)
420{
421 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
422 struct clock_event_device *dev = td->evtdev;
423
424 td->mode = TICKDEV_MODE_PERIODIC;
425 if (dev) {
426 clockevents_exchange_device(old: dev, NULL);
427 dev->event_handler = clockevents_handle_noop;
428 td->evtdev = NULL;
429 }
430}
431#endif
432
433/**
434 * tick_suspend_local - Suspend the local tick device
435 *
436 * Called from the local cpu for freeze with interrupts disabled.
437 *
438 * No locks required. Nothing can change the per cpu device.
439 */
440void tick_suspend_local(void)
441{
442 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
443
444 clockevents_shutdown(dev: td->evtdev);
445}
446
447/**
448 * tick_resume_local - Resume the local tick device
449 *
450 * Called from the local CPU for unfreeze or XEN resume magic.
451 *
452 * No locks required. Nothing can change the per cpu device.
453 */
454void tick_resume_local(void)
455{
456 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
457 bool broadcast = tick_resume_check_broadcast();
458
459 clockevents_tick_resume(dev: td->evtdev);
460 if (!broadcast) {
461 if (td->mode == TICKDEV_MODE_PERIODIC)
462 tick_setup_periodic(dev: td->evtdev, broadcast: 0);
463 else
464 tick_resume_oneshot();
465 }
466
467 /*
468 * Ensure that hrtimers are up to date and the clockevents device
469 * is reprogrammed correctly when high resolution timers are
470 * enabled.
471 */
472 hrtimers_resume_local();
473}
474
475/**
476 * tick_suspend - Suspend the tick and the broadcast device
477 *
478 * Called from syscore_suspend() via timekeeping_suspend with only one
479 * CPU online and interrupts disabled or from tick_unfreeze() under
480 * tick_freeze_lock.
481 *
482 * No locks required. Nothing can change the per cpu device.
483 */
484void tick_suspend(void)
485{
486 tick_suspend_local();
487 tick_suspend_broadcast();
488}
489
490/**
491 * tick_resume - Resume the tick and the broadcast device
492 *
493 * Called from syscore_resume() via timekeeping_resume with only one
494 * CPU online and interrupts disabled.
495 *
496 * No locks required. Nothing can change the per cpu device.
497 */
498void tick_resume(void)
499{
500 tick_resume_broadcast();
501 tick_resume_local();
502}
503
504#ifdef CONFIG_SUSPEND
505static DEFINE_RAW_SPINLOCK(tick_freeze_lock);
506static DEFINE_WAIT_OVERRIDE_MAP(tick_freeze_map, LD_WAIT_SLEEP);
507static unsigned int tick_freeze_depth;
508
509/**
510 * tick_freeze - Suspend the local tick and (possibly) timekeeping.
511 *
512 * Check if this is the last online CPU executing the function and if so,
513 * suspend timekeeping. Otherwise suspend the local tick.
514 *
515 * Call with interrupts disabled. Must be balanced with %tick_unfreeze().
516 * Interrupts must not be enabled before the subsequent %tick_unfreeze().
517 */
518void tick_freeze(void)
519{
520 raw_spin_lock(&tick_freeze_lock);
521
522 tick_freeze_depth++;
523 if (tick_freeze_depth == num_online_cpus()) {
524 trace_suspend_resume(TPS("timekeeping_freeze"),
525 smp_processor_id(), start: true);
526 /*
527 * All other CPUs have their interrupts disabled and are
528 * suspended to idle. Other tasks have been frozen so there
529 * is no scheduling happening. This means that there is no
530 * concurrency in the system at this point. Therefore it is
531 * okay to acquire a sleeping lock on PREEMPT_RT, such as a
532 * spinlock, because the lock cannot be held by other CPUs
533 * or threads and acquiring it cannot block.
534 *
535 * Inform lockdep about the situation.
536 */
537 lock_map_acquire_try(&tick_freeze_map);
538 system_state = SYSTEM_SUSPEND;
539 sched_clock_suspend();
540 timekeeping_suspend();
541 lock_map_release(&tick_freeze_map);
542 } else {
543 tick_suspend_local();
544 }
545
546 raw_spin_unlock(&tick_freeze_lock);
547}
548
549/**
550 * tick_unfreeze - Resume the local tick and (possibly) timekeeping.
551 *
552 * Check if this is the first CPU executing the function and if so, resume
553 * timekeeping. Otherwise resume the local tick.
554 *
555 * Call with interrupts disabled. Must be balanced with %tick_freeze().
556 * Interrupts must not be enabled after the preceding %tick_freeze().
557 */
558void tick_unfreeze(void)
559{
560 raw_spin_lock(&tick_freeze_lock);
561
562 if (tick_freeze_depth == num_online_cpus()) {
563 /*
564 * Similar to tick_freeze(). On resumption the first CPU may
565 * acquire uncontended sleeping locks while other CPUs block on
566 * tick_freeze_lock.
567 */
568 lock_map_acquire_try(&tick_freeze_map);
569 timekeeping_resume();
570 sched_clock_resume();
571 lock_map_release(&tick_freeze_map);
572
573 system_state = SYSTEM_RUNNING;
574 trace_suspend_resume(TPS("timekeeping_freeze"),
575 smp_processor_id(), start: false);
576 } else {
577 touch_softlockup_watchdog();
578 tick_resume_local();
579 }
580
581 tick_freeze_depth--;
582
583 raw_spin_unlock(&tick_freeze_lock);
584}
585#endif /* CONFIG_SUSPEND */
586
587/**
588 * tick_init - initialize the tick control
589 */
590void __init tick_init(void)
591{
592 tick_broadcast_init();
593 tick_nohz_init();
594}
595