1 // SPDX-License-Identifier: GPL-2.0
3 * This file contains the base functions to manage periodic tick
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
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>
21 #include <asm/irq_regs.h>
23 #include "tick-internal.h"
28 DEFINE_PER_CPU(struct tick_device, tick_cpu_device);
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.
34 ktime_t tick_next_period;
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:
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.
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.
50 int tick_do_timer_cpu __read_mostly = TICK_DO_TIMER_BOOT;
51 #ifdef CONFIG_NO_HZ_FULL
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.
57 static int tick_do_timer_boot_cpu __read_mostly = -1;
61 * Debugging: see timer_list.c
63 struct tick_device *tick_get_device(int cpu)
65 return &per_cpu(tick_cpu_device, cpu);
69 * tick_is_oneshot_available - check for a oneshot capable event device
71 int tick_is_oneshot_available(void)
73 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
75 if (!dev || !(dev->features & CLOCK_EVT_FEAT_ONESHOT))
77 if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
79 return tick_broadcast_oneshot_available();
85 static void tick_periodic(int cpu)
87 if (tick_do_timer_cpu == cpu) {
88 raw_spin_lock(&jiffies_lock);
89 write_seqcount_begin(&jiffies_seq);
91 /* Keep track of the next tick event */
92 tick_next_period = ktime_add_ns(tick_next_period, TICK_NSEC);
95 write_seqcount_end(&jiffies_seq);
96 raw_spin_unlock(&jiffies_lock);
100 update_process_times(user_mode(get_irq_regs()));
101 profile_tick(CPU_PROFILING);
105 * Event handler for periodic ticks
107 void tick_handle_periodic(struct clock_event_device *dev)
109 int cpu = smp_processor_id();
110 ktime_t next = dev->next_event;
114 #if defined(CONFIG_HIGH_RES_TIMERS) || defined(CONFIG_NO_HZ_COMMON)
116 * The cpu might have transitioned to HIGHRES or NOHZ mode via
117 * update_process_times() -> run_local_timers() ->
118 * hrtimer_run_queues().
120 if (dev->event_handler != tick_handle_periodic)
124 if (!clockevent_state_oneshot(dev))
128 * Setup the next period for devices, which do not have
131 next = ktime_add_ns(next, TICK_NSEC);
133 if (!clockevents_program_event(dev, next, false))
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.
144 if (timekeeping_valid_for_hres())
150 * Setup the device for a periodic tick
152 void tick_setup_periodic(struct clock_event_device *dev, int broadcast)
154 tick_set_periodic_handler(dev, broadcast);
156 /* Broadcast setup ? */
157 if (!tick_device_is_functional(dev))
160 if ((dev->features & CLOCK_EVT_FEAT_PERIODIC) &&
161 !tick_broadcast_oneshot_active()) {
162 clockevents_switch_state(dev, CLOCK_EVT_STATE_PERIODIC);
168 seq = read_seqcount_begin(&jiffies_seq);
169 next = tick_next_period;
170 } while (read_seqcount_retry(&jiffies_seq, seq));
172 clockevents_switch_state(dev, CLOCK_EVT_STATE_ONESHOT);
175 if (!clockevents_program_event(dev, next, false))
177 next = ktime_add_ns(next, TICK_NSEC);
182 #ifdef CONFIG_NO_HZ_FULL
183 static void giveup_do_timer(void *info)
185 int cpu = *(unsigned int *)info;
187 WARN_ON(tick_do_timer_cpu != smp_processor_id());
189 tick_do_timer_cpu = cpu;
192 static void tick_take_do_timer_from_boot(void)
194 int cpu = smp_processor_id();
195 int from = tick_do_timer_boot_cpu;
197 if (from >= 0 && from != cpu)
198 smp_call_function_single(from, giveup_do_timer, &cpu, 1);
203 * Setup the tick device
205 static void tick_setup_device(struct tick_device *td,
206 struct clock_event_device *newdev, int cpu,
207 const struct cpumask *cpumask)
209 void (*handler)(struct clock_event_device *) = NULL;
210 ktime_t next_event = 0;
213 * First device setup ?
217 * If no cpu took the do_timer update, assign it to
220 if (tick_do_timer_cpu == TICK_DO_TIMER_BOOT) {
221 tick_do_timer_cpu = cpu;
222 tick_next_period = ktime_get();
223 #ifdef CONFIG_NO_HZ_FULL
225 * The boot CPU may be nohz_full, in which case set
226 * tick_do_timer_boot_cpu so the first housekeeping
227 * secondary that comes up will take do_timer from
230 if (tick_nohz_full_cpu(cpu))
231 tick_do_timer_boot_cpu = cpu;
233 } else if (tick_do_timer_boot_cpu != -1 &&
234 !tick_nohz_full_cpu(cpu)) {
235 tick_take_do_timer_from_boot();
236 tick_do_timer_boot_cpu = -1;
237 WARN_ON(tick_do_timer_cpu != cpu);
242 * Startup in periodic mode first.
244 td->mode = TICKDEV_MODE_PERIODIC;
246 handler = td->evtdev->event_handler;
247 next_event = td->evtdev->next_event;
248 td->evtdev->event_handler = clockevents_handle_noop;
254 * When the device is not per cpu, pin the interrupt to the
257 if (!cpumask_equal(newdev->cpumask, cpumask))
258 irq_set_affinity(newdev->irq, cpumask);
261 * When global broadcasting is active, check if the current
262 * device is registered as a placeholder for broadcast mode.
263 * This allows us to handle this x86 misfeature in a generic
264 * way. This function also returns !=0 when we keep the
265 * current active broadcast state for this CPU.
267 if (tick_device_uses_broadcast(newdev, cpu))
270 if (td->mode == TICKDEV_MODE_PERIODIC)
271 tick_setup_periodic(newdev, 0);
273 tick_setup_oneshot(newdev, handler, next_event);
276 void tick_install_replacement(struct clock_event_device *newdev)
278 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
279 int cpu = smp_processor_id();
281 clockevents_exchange_device(td->evtdev, newdev);
282 tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
283 if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
284 tick_oneshot_notify();
287 static bool tick_check_percpu(struct clock_event_device *curdev,
288 struct clock_event_device *newdev, int cpu)
290 if (!cpumask_test_cpu(cpu, newdev->cpumask))
292 if (cpumask_equal(newdev->cpumask, cpumask_of(cpu)))
294 /* Check if irq affinity can be set */
295 if (newdev->irq >= 0 && !irq_can_set_affinity(newdev->irq))
297 /* Prefer an existing cpu local device */
298 if (curdev && cpumask_equal(curdev->cpumask, cpumask_of(cpu)))
303 static bool tick_check_preferred(struct clock_event_device *curdev,
304 struct clock_event_device *newdev)
306 /* Prefer oneshot capable device */
307 if (!(newdev->features & CLOCK_EVT_FEAT_ONESHOT)) {
308 if (curdev && (curdev->features & CLOCK_EVT_FEAT_ONESHOT))
310 if (tick_oneshot_mode_active())
315 * Use the higher rated one, but prefer a CPU local device with a lower
316 * rating than a non-CPU local device
319 newdev->rating > curdev->rating ||
320 !cpumask_equal(curdev->cpumask, newdev->cpumask);
324 * Check whether the new device is a better fit than curdev. curdev
327 bool tick_check_replacement(struct clock_event_device *curdev,
328 struct clock_event_device *newdev)
330 if (!tick_check_percpu(curdev, newdev, smp_processor_id()))
333 return tick_check_preferred(curdev, newdev);
337 * Check, if the new registered device should be used. Called with
338 * clockevents_lock held and interrupts disabled.
340 void tick_check_new_device(struct clock_event_device *newdev)
342 struct clock_event_device *curdev;
343 struct tick_device *td;
346 cpu = smp_processor_id();
347 td = &per_cpu(tick_cpu_device, cpu);
350 if (!tick_check_replacement(curdev, newdev))
353 if (!try_module_get(newdev->owner))
357 * Replace the eventually existing device by the new
358 * device. If the current device is the broadcast device, do
359 * not give it back to the clockevents layer !
361 if (tick_is_broadcast_device(curdev)) {
362 clockevents_shutdown(curdev);
365 clockevents_exchange_device(curdev, newdev);
366 tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
367 if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
368 tick_oneshot_notify();
373 * Can the new device be used as a broadcast device ?
375 tick_install_broadcast_device(newdev, cpu);
379 * tick_broadcast_oneshot_control - Enter/exit broadcast oneshot mode
380 * @state: The target state (enter/exit)
382 * The system enters/leaves a state, where affected devices might stop
383 * Returns 0 on success, -EBUSY if the cpu is used to broadcast wakeups.
385 * Called with interrupts disabled, so clockevents_lock is not
386 * required here because the local clock event device cannot go away
389 int tick_broadcast_oneshot_control(enum tick_broadcast_state state)
391 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
393 if (!(td->evtdev->features & CLOCK_EVT_FEAT_C3STOP))
396 return __tick_broadcast_oneshot_control(state);
398 EXPORT_SYMBOL_GPL(tick_broadcast_oneshot_control);
400 #ifdef CONFIG_HOTPLUG_CPU
402 * Transfer the do_timer job away from a dying cpu.
404 * Called with interrupts disabled. No locking required. If
405 * tick_do_timer_cpu is owned by this cpu, nothing can change it.
407 void tick_handover_do_timer(void)
409 if (tick_do_timer_cpu == smp_processor_id())
410 tick_do_timer_cpu = cpumask_first(cpu_online_mask);
414 * Shutdown an event device on a given cpu:
416 * This is called on a life CPU, when a CPU is dead. So we cannot
417 * access the hardware device itself.
418 * We just set the mode and remove it from the lists.
420 void tick_shutdown(unsigned int cpu)
422 struct tick_device *td = &per_cpu(tick_cpu_device, cpu);
423 struct clock_event_device *dev = td->evtdev;
425 td->mode = TICKDEV_MODE_PERIODIC;
428 * Prevent that the clock events layer tries to call
429 * the set mode function!
431 clockevent_set_state(dev, CLOCK_EVT_STATE_DETACHED);
432 clockevents_exchange_device(dev, NULL);
433 dev->event_handler = clockevents_handle_noop;
440 * tick_suspend_local - Suspend the local tick device
442 * Called from the local cpu for freeze with interrupts disabled.
444 * No locks required. Nothing can change the per cpu device.
446 void tick_suspend_local(void)
448 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
450 clockevents_shutdown(td->evtdev);
454 * tick_resume_local - Resume the local tick device
456 * Called from the local CPU for unfreeze or XEN resume magic.
458 * No locks required. Nothing can change the per cpu device.
460 void tick_resume_local(void)
462 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
463 bool broadcast = tick_resume_check_broadcast();
465 clockevents_tick_resume(td->evtdev);
467 if (td->mode == TICKDEV_MODE_PERIODIC)
468 tick_setup_periodic(td->evtdev, 0);
470 tick_resume_oneshot();
474 * Ensure that hrtimers are up to date and the clockevents device
475 * is reprogrammed correctly when high resolution timers are
478 hrtimers_resume_local();
482 * tick_suspend - Suspend the tick and the broadcast device
484 * Called from syscore_suspend() via timekeeping_suspend with only one
485 * CPU online and interrupts disabled or from tick_unfreeze() under
488 * No locks required. Nothing can change the per cpu device.
490 void tick_suspend(void)
492 tick_suspend_local();
493 tick_suspend_broadcast();
497 * tick_resume - Resume the tick and the broadcast device
499 * Called from syscore_resume() via timekeeping_resume with only one
500 * CPU online and interrupts disabled.
502 * No locks required. Nothing can change the per cpu device.
504 void tick_resume(void)
506 tick_resume_broadcast();
510 #ifdef CONFIG_SUSPEND
511 static DEFINE_RAW_SPINLOCK(tick_freeze_lock);
512 static unsigned int tick_freeze_depth;
515 * tick_freeze - Suspend the local tick and (possibly) timekeeping.
517 * Check if this is the last online CPU executing the function and if so,
518 * suspend timekeeping. Otherwise suspend the local tick.
520 * Call with interrupts disabled. Must be balanced with %tick_unfreeze().
521 * Interrupts must not be enabled before the subsequent %tick_unfreeze().
523 void tick_freeze(void)
525 raw_spin_lock(&tick_freeze_lock);
528 if (tick_freeze_depth == num_online_cpus()) {
529 trace_suspend_resume(TPS("timekeeping_freeze"),
530 smp_processor_id(), true);
531 system_state = SYSTEM_SUSPEND;
532 sched_clock_suspend();
533 timekeeping_suspend();
535 tick_suspend_local();
538 raw_spin_unlock(&tick_freeze_lock);
542 * tick_unfreeze - Resume the local tick and (possibly) timekeeping.
544 * Check if this is the first CPU executing the function and if so, resume
545 * timekeeping. Otherwise resume the local tick.
547 * Call with interrupts disabled. Must be balanced with %tick_freeze().
548 * Interrupts must not be enabled after the preceding %tick_freeze().
550 void tick_unfreeze(void)
552 raw_spin_lock(&tick_freeze_lock);
554 if (tick_freeze_depth == num_online_cpus()) {
555 timekeeping_resume();
556 sched_clock_resume();
557 system_state = SYSTEM_RUNNING;
558 trace_suspend_resume(TPS("timekeeping_freeze"),
559 smp_processor_id(), false);
561 touch_softlockup_watchdog();
567 raw_spin_unlock(&tick_freeze_lock);
569 #endif /* CONFIG_SUSPEND */
572 * tick_init - initialize the tick control
574 void __init tick_init(void)
576 tick_broadcast_init();