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/percpu.h>
15 #include <linux/profile.h>
16 #include <linux/sched.h>
17 #include <linux/module.h>
18 #include <trace/events/power.h>
20 #include <asm/irq_regs.h>
22 #include "tick-internal.h"
27 DEFINE_PER_CPU(struct tick_device
, tick_cpu_device
);
29 * Tick next event: keeps track of the tick time
31 ktime_t tick_next_period
;
35 * tick_do_timer_cpu is a timer core internal variable which holds the CPU NR
36 * which is responsible for calling do_timer(), i.e. the timekeeping stuff. This
37 * variable has two functions:
39 * 1) Prevent a thundering herd issue of a gazillion of CPUs trying to grab the
40 * timekeeping lock all at once. Only the CPU which is assigned to do the
41 * update is handling it.
43 * 2) Hand off the duty in the NOHZ idle case by setting the value to
44 * TICK_DO_TIMER_NONE, i.e. a non existing CPU. So the next cpu which looks
45 * at it will take over and keep the time keeping alive. The handover
46 * procedure also covers cpu hotplug.
48 int tick_do_timer_cpu __read_mostly
= TICK_DO_TIMER_BOOT
;
49 #ifdef CONFIG_NO_HZ_FULL
51 * tick_do_timer_boot_cpu indicates the boot CPU temporarily owns
52 * tick_do_timer_cpu and it should be taken over by an eligible secondary
53 * when one comes online.
55 static int tick_do_timer_boot_cpu __read_mostly
= -1;
59 * Debugging: see timer_list.c
61 struct tick_device
*tick_get_device(int cpu
)
63 return &per_cpu(tick_cpu_device
, cpu
);
67 * tick_is_oneshot_available - check for a oneshot capable event device
69 int tick_is_oneshot_available(void)
71 struct clock_event_device
*dev
= __this_cpu_read(tick_cpu_device
.evtdev
);
73 if (!dev
|| !(dev
->features
& CLOCK_EVT_FEAT_ONESHOT
))
75 if (!(dev
->features
& CLOCK_EVT_FEAT_C3STOP
))
77 return tick_broadcast_oneshot_available();
83 static void tick_periodic(int cpu
)
85 if (tick_do_timer_cpu
== cpu
) {
86 write_seqlock(&jiffies_lock
);
88 /* Keep track of the next tick event */
89 tick_next_period
= ktime_add(tick_next_period
, tick_period
);
92 write_sequnlock(&jiffies_lock
);
96 update_process_times(user_mode(get_irq_regs()));
97 profile_tick(CPU_PROFILING
);
101 * Event handler for periodic ticks
103 void tick_handle_periodic(struct clock_event_device
*dev
)
105 int cpu
= smp_processor_id();
106 ktime_t next
= dev
->next_event
;
110 #if defined(CONFIG_HIGH_RES_TIMERS) || defined(CONFIG_NO_HZ_COMMON)
112 * The cpu might have transitioned to HIGHRES or NOHZ mode via
113 * update_process_times() -> run_local_timers() ->
114 * hrtimer_run_queues().
116 if (dev
->event_handler
!= tick_handle_periodic
)
120 if (!clockevent_state_oneshot(dev
))
124 * Setup the next period for devices, which do not have
127 next
= ktime_add(next
, tick_period
);
129 if (!clockevents_program_event(dev
, next
, false))
132 * Have to be careful here. If we're in oneshot mode,
133 * before we call tick_periodic() in a loop, we need
134 * to be sure we're using a real hardware clocksource.
135 * Otherwise we could get trapped in an infinite
136 * loop, as the tick_periodic() increments jiffies,
137 * which then will increment time, possibly causing
138 * the loop to trigger again and again.
140 if (timekeeping_valid_for_hres())
146 * Setup the device for a periodic tick
148 void tick_setup_periodic(struct clock_event_device
*dev
, int broadcast
)
150 tick_set_periodic_handler(dev
, broadcast
);
152 /* Broadcast setup ? */
153 if (!tick_device_is_functional(dev
))
156 if ((dev
->features
& CLOCK_EVT_FEAT_PERIODIC
) &&
157 !tick_broadcast_oneshot_active()) {
158 clockevents_switch_state(dev
, CLOCK_EVT_STATE_PERIODIC
);
164 seq
= read_seqbegin(&jiffies_lock
);
165 next
= tick_next_period
;
166 } while (read_seqretry(&jiffies_lock
, seq
));
168 clockevents_switch_state(dev
, CLOCK_EVT_STATE_ONESHOT
);
171 if (!clockevents_program_event(dev
, next
, false))
173 next
= ktime_add(next
, tick_period
);
178 #ifdef CONFIG_NO_HZ_FULL
179 static void giveup_do_timer(void *info
)
181 int cpu
= *(unsigned int *)info
;
183 WARN_ON(tick_do_timer_cpu
!= smp_processor_id());
185 tick_do_timer_cpu
= cpu
;
188 static void tick_take_do_timer_from_boot(void)
190 int cpu
= smp_processor_id();
191 int from
= tick_do_timer_boot_cpu
;
193 if (from
>= 0 && from
!= cpu
)
194 smp_call_function_single(from
, giveup_do_timer
, &cpu
, 1);
199 * Setup the tick device
201 static void tick_setup_device(struct tick_device
*td
,
202 struct clock_event_device
*newdev
, int cpu
,
203 const struct cpumask
*cpumask
)
205 void (*handler
)(struct clock_event_device
*) = NULL
;
206 ktime_t next_event
= 0;
209 * First device setup ?
213 * If no cpu took the do_timer update, assign it to
216 if (tick_do_timer_cpu
== TICK_DO_TIMER_BOOT
) {
217 tick_do_timer_cpu
= cpu
;
219 tick_next_period
= ktime_get();
220 tick_period
= NSEC_PER_SEC
/ HZ
;
221 #ifdef CONFIG_NO_HZ_FULL
223 * The boot CPU may be nohz_full, in which case set
224 * tick_do_timer_boot_cpu so the first housekeeping
225 * secondary that comes up will take do_timer from
228 if (tick_nohz_full_cpu(cpu
))
229 tick_do_timer_boot_cpu
= cpu
;
231 } else if (tick_do_timer_boot_cpu
!= -1 &&
232 !tick_nohz_full_cpu(cpu
)) {
233 tick_take_do_timer_from_boot();
234 tick_do_timer_boot_cpu
= -1;
235 WARN_ON(tick_do_timer_cpu
!= cpu
);
240 * Startup in periodic mode first.
242 td
->mode
= TICKDEV_MODE_PERIODIC
;
244 handler
= td
->evtdev
->event_handler
;
245 next_event
= td
->evtdev
->next_event
;
246 td
->evtdev
->event_handler
= clockevents_handle_noop
;
252 * When the device is not per cpu, pin the interrupt to the
255 if (!cpumask_equal(newdev
->cpumask
, cpumask
))
256 irq_set_affinity(newdev
->irq
, cpumask
);
259 * When global broadcasting is active, check if the current
260 * device is registered as a placeholder for broadcast mode.
261 * This allows us to handle this x86 misfeature in a generic
262 * way. This function also returns !=0 when we keep the
263 * current active broadcast state for this CPU.
265 if (tick_device_uses_broadcast(newdev
, cpu
))
268 if (td
->mode
== TICKDEV_MODE_PERIODIC
)
269 tick_setup_periodic(newdev
, 0);
271 tick_setup_oneshot(newdev
, handler
, next_event
);
274 void tick_install_replacement(struct clock_event_device
*newdev
)
276 struct tick_device
*td
= this_cpu_ptr(&tick_cpu_device
);
277 int cpu
= smp_processor_id();
279 clockevents_exchange_device(td
->evtdev
, newdev
);
280 tick_setup_device(td
, newdev
, cpu
, cpumask_of(cpu
));
281 if (newdev
->features
& CLOCK_EVT_FEAT_ONESHOT
)
282 tick_oneshot_notify();
285 static bool tick_check_percpu(struct clock_event_device
*curdev
,
286 struct clock_event_device
*newdev
, int cpu
)
288 if (!cpumask_test_cpu(cpu
, newdev
->cpumask
))
290 if (cpumask_equal(newdev
->cpumask
, cpumask_of(cpu
)))
292 /* Check if irq affinity can be set */
293 if (newdev
->irq
>= 0 && !irq_can_set_affinity(newdev
->irq
))
295 /* Prefer an existing cpu local device */
296 if (curdev
&& cpumask_equal(curdev
->cpumask
, cpumask_of(cpu
)))
301 static bool tick_check_preferred(struct clock_event_device
*curdev
,
302 struct clock_event_device
*newdev
)
304 /* Prefer oneshot capable device */
305 if (!(newdev
->features
& CLOCK_EVT_FEAT_ONESHOT
)) {
306 if (curdev
&& (curdev
->features
& CLOCK_EVT_FEAT_ONESHOT
))
308 if (tick_oneshot_mode_active())
313 * Use the higher rated one, but prefer a CPU local device with a lower
314 * rating than a non-CPU local device
317 newdev
->rating
> curdev
->rating
||
318 !cpumask_equal(curdev
->cpumask
, newdev
->cpumask
);
322 * Check whether the new device is a better fit than curdev. curdev
325 bool tick_check_replacement(struct clock_event_device
*curdev
,
326 struct clock_event_device
*newdev
)
328 if (!tick_check_percpu(curdev
, newdev
, smp_processor_id()))
331 return tick_check_preferred(curdev
, newdev
);
335 * Check, if the new registered device should be used. Called with
336 * clockevents_lock held and interrupts disabled.
338 void tick_check_new_device(struct clock_event_device
*newdev
)
340 struct clock_event_device
*curdev
;
341 struct tick_device
*td
;
344 cpu
= smp_processor_id();
345 td
= &per_cpu(tick_cpu_device
, cpu
);
348 /* cpu local device ? */
349 if (!tick_check_percpu(curdev
, newdev
, cpu
))
352 /* Preference decision */
353 if (!tick_check_preferred(curdev
, newdev
))
356 if (!try_module_get(newdev
->owner
))
360 * Replace the eventually existing device by the new
361 * device. If the current device is the broadcast device, do
362 * not give it back to the clockevents layer !
364 if (tick_is_broadcast_device(curdev
)) {
365 clockevents_shutdown(curdev
);
368 clockevents_exchange_device(curdev
, newdev
);
369 tick_setup_device(td
, newdev
, cpu
, cpumask_of(cpu
));
370 if (newdev
->features
& CLOCK_EVT_FEAT_ONESHOT
)
371 tick_oneshot_notify();
376 * Can the new device be used as a broadcast device ?
378 tick_install_broadcast_device(newdev
);
382 * tick_broadcast_oneshot_control - Enter/exit broadcast oneshot mode
383 * @state: The target state (enter/exit)
385 * The system enters/leaves a state, where affected devices might stop
386 * Returns 0 on success, -EBUSY if the cpu is used to broadcast wakeups.
388 * Called with interrupts disabled, so clockevents_lock is not
389 * required here because the local clock event device cannot go away
392 int tick_broadcast_oneshot_control(enum tick_broadcast_state state
)
394 struct tick_device
*td
= this_cpu_ptr(&tick_cpu_device
);
396 if (!(td
->evtdev
->features
& CLOCK_EVT_FEAT_C3STOP
))
399 return __tick_broadcast_oneshot_control(state
);
401 EXPORT_SYMBOL_GPL(tick_broadcast_oneshot_control
);
403 #ifdef CONFIG_HOTPLUG_CPU
405 * Transfer the do_timer job away from a dying cpu.
407 * Called with interrupts disabled. Not locking required. If
408 * tick_do_timer_cpu is owned by this cpu, nothing can change it.
410 void tick_handover_do_timer(void)
412 if (tick_do_timer_cpu
== smp_processor_id()) {
413 int cpu
= cpumask_first(cpu_online_mask
);
415 tick_do_timer_cpu
= (cpu
< nr_cpu_ids
) ? cpu
:
421 * Shutdown an event device on a given cpu:
423 * This is called on a life CPU, when a CPU is dead. So we cannot
424 * access the hardware device itself.
425 * We just set the mode and remove it from the lists.
427 void tick_shutdown(unsigned int cpu
)
429 struct tick_device
*td
= &per_cpu(tick_cpu_device
, cpu
);
430 struct clock_event_device
*dev
= td
->evtdev
;
432 td
->mode
= TICKDEV_MODE_PERIODIC
;
435 * Prevent that the clock events layer tries to call
436 * the set mode function!
438 clockevent_set_state(dev
, CLOCK_EVT_STATE_DETACHED
);
439 clockevents_exchange_device(dev
, NULL
);
440 dev
->event_handler
= clockevents_handle_noop
;
447 * tick_suspend_local - Suspend the local tick device
449 * Called from the local cpu for freeze with interrupts disabled.
451 * No locks required. Nothing can change the per cpu device.
453 void tick_suspend_local(void)
455 struct tick_device
*td
= this_cpu_ptr(&tick_cpu_device
);
457 clockevents_shutdown(td
->evtdev
);
461 * tick_resume_local - Resume the local tick device
463 * Called from the local CPU for unfreeze or XEN resume magic.
465 * No locks required. Nothing can change the per cpu device.
467 void tick_resume_local(void)
469 struct tick_device
*td
= this_cpu_ptr(&tick_cpu_device
);
470 bool broadcast
= tick_resume_check_broadcast();
472 clockevents_tick_resume(td
->evtdev
);
474 if (td
->mode
== TICKDEV_MODE_PERIODIC
)
475 tick_setup_periodic(td
->evtdev
, 0);
477 tick_resume_oneshot();
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);
566 raw_spin_unlock(&tick_freeze_lock
);
568 #endif /* CONFIG_SUSPEND */
571 * tick_init - initialize the tick control
573 void __init
tick_init(void)
575 tick_broadcast_init();