Linux Audio

Check our new training course

Yocto distribution development and maintenance

Need a Yocto distribution for your embedded project?
Loading...
v5.4
  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/percpu.h>
 15#include <linux/profile.h>
 16#include <linux/sched.h>
 17#include <linux/module.h>
 18#include <trace/events/power.h>
 19
 20#include <asm/irq_regs.h>
 21
 22#include "tick-internal.h"
 23
 24/*
 25 * Tick devices
 26 */
 27DEFINE_PER_CPU(struct tick_device, tick_cpu_device);
 28/*
 29 * Tick next event: keeps track of the tick time
 30 */
 31ktime_t tick_next_period;
 32ktime_t tick_period;
 33
 34/*
 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:
 38 *
 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.
 42 *
 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.
 47 */
 48int tick_do_timer_cpu __read_mostly = TICK_DO_TIMER_BOOT;
 49#ifdef CONFIG_NO_HZ_FULL
 50/*
 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.
 54 */
 55static int tick_do_timer_boot_cpu __read_mostly = -1;
 56#endif
 57
 58/*
 59 * Debugging: see timer_list.c
 60 */
 61struct tick_device *tick_get_device(int cpu)
 62{
 63	return &per_cpu(tick_cpu_device, cpu);
 64}
 65
 66/**
 67 * tick_is_oneshot_available - check for a oneshot capable event device
 68 */
 69int tick_is_oneshot_available(void)
 70{
 71	struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
 72
 73	if (!dev || !(dev->features & CLOCK_EVT_FEAT_ONESHOT))
 74		return 0;
 75	if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
 76		return 1;
 77	return tick_broadcast_oneshot_available();
 78}
 79
 80/*
 81 * Periodic tick
 82 */
 83static void tick_periodic(int cpu)
 84{
 85	if (tick_do_timer_cpu == cpu) {
 86		write_seqlock(&jiffies_lock);
 
 87
 88		/* Keep track of the next tick event */
 89		tick_next_period = ktime_add(tick_next_period, tick_period);
 90
 91		do_timer(1);
 92		write_sequnlock(&jiffies_lock);
 
 93		update_wall_time();
 94	}
 95
 96	update_process_times(user_mode(get_irq_regs()));
 97	profile_tick(CPU_PROFILING);
 98}
 99
100/*
101 * Event handler for periodic ticks
102 */
103void tick_handle_periodic(struct clock_event_device *dev)
104{
105	int cpu = smp_processor_id();
106	ktime_t next = dev->next_event;
107
108	tick_periodic(cpu);
109
110#if defined(CONFIG_HIGH_RES_TIMERS) || defined(CONFIG_NO_HZ_COMMON)
111	/*
112	 * The cpu might have transitioned to HIGHRES or NOHZ mode via
113	 * update_process_times() -> run_local_timers() ->
114	 * hrtimer_run_queues().
115	 */
116	if (dev->event_handler != tick_handle_periodic)
117		return;
118#endif
119
120	if (!clockevent_state_oneshot(dev))
121		return;
122	for (;;) {
123		/*
124		 * Setup the next period for devices, which do not have
125		 * periodic mode:
126		 */
127		next = ktime_add(next, tick_period);
128
129		if (!clockevents_program_event(dev, next, false))
130			return;
131		/*
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.
139		 */
140		if (timekeeping_valid_for_hres())
141			tick_periodic(cpu);
142	}
143}
144
145/*
146 * Setup the device for a periodic tick
147 */
148void tick_setup_periodic(struct clock_event_device *dev, int broadcast)
149{
150	tick_set_periodic_handler(dev, broadcast);
151
152	/* Broadcast setup ? */
153	if (!tick_device_is_functional(dev))
154		return;
155
156	if ((dev->features & CLOCK_EVT_FEAT_PERIODIC) &&
157	    !tick_broadcast_oneshot_active()) {
158		clockevents_switch_state(dev, CLOCK_EVT_STATE_PERIODIC);
159	} else {
160		unsigned int seq;
161		ktime_t next;
162
163		do {
164			seq = read_seqbegin(&jiffies_lock);
165			next = tick_next_period;
166		} while (read_seqretry(&jiffies_lock, seq));
167
168		clockevents_switch_state(dev, CLOCK_EVT_STATE_ONESHOT);
169
170		for (;;) {
171			if (!clockevents_program_event(dev, next, false))
172				return;
173			next = ktime_add(next, tick_period);
174		}
175	}
176}
177
178#ifdef CONFIG_NO_HZ_FULL
179static void giveup_do_timer(void *info)
180{
181	int cpu = *(unsigned int *)info;
182
183	WARN_ON(tick_do_timer_cpu != smp_processor_id());
184
185	tick_do_timer_cpu = cpu;
186}
187
188static void tick_take_do_timer_from_boot(void)
189{
190	int cpu = smp_processor_id();
191	int from = tick_do_timer_boot_cpu;
192
193	if (from >= 0 && from != cpu)
194		smp_call_function_single(from, giveup_do_timer, &cpu, 1);
195}
196#endif
197
198/*
199 * Setup the tick device
200 */
201static void tick_setup_device(struct tick_device *td,
202			      struct clock_event_device *newdev, int cpu,
203			      const struct cpumask *cpumask)
204{
205	void (*handler)(struct clock_event_device *) = NULL;
206	ktime_t next_event = 0;
207
208	/*
209	 * First device setup ?
210	 */
211	if (!td->evtdev) {
212		/*
213		 * If no cpu took the do_timer update, assign it to
214		 * this cpu:
215		 */
216		if (tick_do_timer_cpu == TICK_DO_TIMER_BOOT) {
217			tick_do_timer_cpu = cpu;
218
219			tick_next_period = ktime_get();
220			tick_period = NSEC_PER_SEC / HZ;
221#ifdef CONFIG_NO_HZ_FULL
222			/*
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
226			 * us.
227			 */
228			if (tick_nohz_full_cpu(cpu))
229				tick_do_timer_boot_cpu = cpu;
230
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);
236#endif
237		}
238
239		/*
240		 * Startup in periodic mode first.
241		 */
242		td->mode = TICKDEV_MODE_PERIODIC;
243	} else {
244		handler = td->evtdev->event_handler;
245		next_event = td->evtdev->next_event;
246		td->evtdev->event_handler = clockevents_handle_noop;
247	}
248
249	td->evtdev = newdev;
250
251	/*
252	 * When the device is not per cpu, pin the interrupt to the
253	 * current cpu:
254	 */
255	if (!cpumask_equal(newdev->cpumask, cpumask))
256		irq_set_affinity(newdev->irq, cpumask);
257
258	/*
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.
264	 */
265	if (tick_device_uses_broadcast(newdev, cpu))
266		return;
267
268	if (td->mode == TICKDEV_MODE_PERIODIC)
269		tick_setup_periodic(newdev, 0);
270	else
271		tick_setup_oneshot(newdev, handler, next_event);
272}
273
274void tick_install_replacement(struct clock_event_device *newdev)
275{
276	struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
277	int cpu = smp_processor_id();
278
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();
283}
284
285static bool tick_check_percpu(struct clock_event_device *curdev,
286			      struct clock_event_device *newdev, int cpu)
287{
288	if (!cpumask_test_cpu(cpu, newdev->cpumask))
289		return false;
290	if (cpumask_equal(newdev->cpumask, cpumask_of(cpu)))
291		return true;
292	/* Check if irq affinity can be set */
293	if (newdev->irq >= 0 && !irq_can_set_affinity(newdev->irq))
294		return false;
295	/* Prefer an existing cpu local device */
296	if (curdev && cpumask_equal(curdev->cpumask, cpumask_of(cpu)))
297		return false;
298	return true;
299}
300
301static bool tick_check_preferred(struct clock_event_device *curdev,
302				 struct clock_event_device *newdev)
303{
304	/* Prefer oneshot capable device */
305	if (!(newdev->features & CLOCK_EVT_FEAT_ONESHOT)) {
306		if (curdev && (curdev->features & CLOCK_EVT_FEAT_ONESHOT))
307			return false;
308		if (tick_oneshot_mode_active())
309			return false;
310	}
311
312	/*
313	 * Use the higher rated one, but prefer a CPU local device with a lower
314	 * rating than a non-CPU local device
315	 */
316	return !curdev ||
317		newdev->rating > curdev->rating ||
318	       !cpumask_equal(curdev->cpumask, newdev->cpumask);
319}
320
321/*
322 * Check whether the new device is a better fit than curdev. curdev
323 * can be NULL !
324 */
325bool tick_check_replacement(struct clock_event_device *curdev,
326			    struct clock_event_device *newdev)
327{
328	if (!tick_check_percpu(curdev, newdev, smp_processor_id()))
329		return false;
330
331	return tick_check_preferred(curdev, newdev);
332}
333
334/*
335 * Check, if the new registered device should be used. Called with
336 * clockevents_lock held and interrupts disabled.
337 */
338void tick_check_new_device(struct clock_event_device *newdev)
339{
340	struct clock_event_device *curdev;
341	struct tick_device *td;
342	int cpu;
343
344	cpu = smp_processor_id();
345	td = &per_cpu(tick_cpu_device, cpu);
346	curdev = td->evtdev;
347
348	/* cpu local device ? */
349	if (!tick_check_percpu(curdev, newdev, cpu))
350		goto out_bc;
351
352	/* Preference decision */
353	if (!tick_check_preferred(curdev, newdev))
354		goto out_bc;
355
356	if (!try_module_get(newdev->owner))
357		return;
358
359	/*
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 !
363	 */
364	if (tick_is_broadcast_device(curdev)) {
365		clockevents_shutdown(curdev);
366		curdev = NULL;
367	}
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();
372	return;
373
374out_bc:
375	/*
376	 * Can the new device be used as a broadcast device ?
377	 */
378	tick_install_broadcast_device(newdev);
379}
380
381/**
382 * tick_broadcast_oneshot_control - Enter/exit broadcast oneshot mode
383 * @state:	The target state (enter/exit)
384 *
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.
387 *
388 * Called with interrupts disabled, so clockevents_lock is not
389 * required here because the local clock event device cannot go away
390 * under us.
391 */
392int tick_broadcast_oneshot_control(enum tick_broadcast_state state)
393{
394	struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
395
396	if (!(td->evtdev->features & CLOCK_EVT_FEAT_C3STOP))
397		return 0;
398
399	return __tick_broadcast_oneshot_control(state);
400}
401EXPORT_SYMBOL_GPL(tick_broadcast_oneshot_control);
402
403#ifdef CONFIG_HOTPLUG_CPU
404/*
405 * Transfer the do_timer job away from a dying cpu.
406 *
407 * Called with interrupts disabled. Not locking required. If
408 * tick_do_timer_cpu is owned by this cpu, nothing can change it.
409 */
410void tick_handover_do_timer(void)
411{
412	if (tick_do_timer_cpu == smp_processor_id()) {
413		int cpu = cpumask_first(cpu_online_mask);
414
415		tick_do_timer_cpu = (cpu < nr_cpu_ids) ? cpu :
416			TICK_DO_TIMER_NONE;
417	}
418}
419
420/*
421 * Shutdown an event device on a given cpu:
422 *
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.
426 */
427void tick_shutdown(unsigned int cpu)
428{
429	struct tick_device *td = &per_cpu(tick_cpu_device, cpu);
430	struct clock_event_device *dev = td->evtdev;
431
432	td->mode = TICKDEV_MODE_PERIODIC;
433	if (dev) {
434		/*
435		 * Prevent that the clock events layer tries to call
436		 * the set mode function!
437		 */
438		clockevent_set_state(dev, CLOCK_EVT_STATE_DETACHED);
439		clockevents_exchange_device(dev, NULL);
440		dev->event_handler = clockevents_handle_noop;
441		td->evtdev = NULL;
442	}
443}
444#endif
445
446/**
447 * tick_suspend_local - Suspend the local tick device
448 *
449 * Called from the local cpu for freeze with interrupts disabled.
450 *
451 * No locks required. Nothing can change the per cpu device.
452 */
453void tick_suspend_local(void)
454{
455	struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
456
457	clockevents_shutdown(td->evtdev);
458}
459
460/**
461 * tick_resume_local - Resume the local tick device
462 *
463 * Called from the local CPU for unfreeze or XEN resume magic.
464 *
465 * No locks required. Nothing can change the per cpu device.
466 */
467void tick_resume_local(void)
468{
469	struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
470	bool broadcast = tick_resume_check_broadcast();
471
472	clockevents_tick_resume(td->evtdev);
473	if (!broadcast) {
474		if (td->mode == TICKDEV_MODE_PERIODIC)
475			tick_setup_periodic(td->evtdev, 0);
476		else
477			tick_resume_oneshot();
478	}
479}
480
481/**
482 * tick_suspend - Suspend the tick and the broadcast device
483 *
484 * Called from syscore_suspend() via timekeeping_suspend with only one
485 * CPU online and interrupts disabled or from tick_unfreeze() under
486 * tick_freeze_lock.
487 *
488 * No locks required. Nothing can change the per cpu device.
489 */
490void tick_suspend(void)
491{
492	tick_suspend_local();
493	tick_suspend_broadcast();
494}
495
496/**
497 * tick_resume - Resume the tick and the broadcast device
498 *
499 * Called from syscore_resume() via timekeeping_resume with only one
500 * CPU online and interrupts disabled.
501 *
502 * No locks required. Nothing can change the per cpu device.
503 */
504void tick_resume(void)
505{
506	tick_resume_broadcast();
507	tick_resume_local();
508}
509
510#ifdef CONFIG_SUSPEND
511static DEFINE_RAW_SPINLOCK(tick_freeze_lock);
512static unsigned int tick_freeze_depth;
513
514/**
515 * tick_freeze - Suspend the local tick and (possibly) timekeeping.
516 *
517 * Check if this is the last online CPU executing the function and if so,
518 * suspend timekeeping.  Otherwise suspend the local tick.
519 *
520 * Call with interrupts disabled.  Must be balanced with %tick_unfreeze().
521 * Interrupts must not be enabled before the subsequent %tick_unfreeze().
522 */
523void tick_freeze(void)
524{
525	raw_spin_lock(&tick_freeze_lock);
526
527	tick_freeze_depth++;
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();
534	} else {
535		tick_suspend_local();
536	}
537
538	raw_spin_unlock(&tick_freeze_lock);
539}
540
541/**
542 * tick_unfreeze - Resume the local tick and (possibly) timekeeping.
543 *
544 * Check if this is the first CPU executing the function and if so, resume
545 * timekeeping.  Otherwise resume the local tick.
546 *
547 * Call with interrupts disabled.  Must be balanced with %tick_freeze().
548 * Interrupts must not be enabled after the preceding %tick_freeze().
549 */
550void tick_unfreeze(void)
551{
552	raw_spin_lock(&tick_freeze_lock);
553
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);
560	} else {
 
561		tick_resume_local();
562	}
563
564	tick_freeze_depth--;
565
566	raw_spin_unlock(&tick_freeze_lock);
567}
568#endif /* CONFIG_SUSPEND */
569
570/**
571 * tick_init - initialize the tick control
572 */
573void __init tick_init(void)
574{
575	tick_broadcast_init();
576	tick_nohz_init();
577}
v5.9
  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 */
 28DEFINE_PER_CPU(struct tick_device, tick_cpu_device);
 29/*
 30 * Tick next event: keeps track of the tick time
 31 */
 32ktime_t tick_next_period;
 33ktime_t tick_period;
 34
 35/*
 36 * tick_do_timer_cpu is a timer core internal variable which holds the CPU NR
 37 * which is responsible for calling do_timer(), i.e. the timekeeping stuff. This
 38 * variable has two functions:
 39 *
 40 * 1) Prevent a thundering herd issue of a gazillion of CPUs trying to grab the
 41 *    timekeeping lock all at once. Only the CPU which is assigned to do the
 42 *    update is handling it.
 43 *
 44 * 2) Hand off the duty in the NOHZ idle case by setting the value to
 45 *    TICK_DO_TIMER_NONE, i.e. a non existing CPU. So the next cpu which looks
 46 *    at it will take over and keep the time keeping alive.  The handover
 47 *    procedure also covers cpu hotplug.
 48 */
 49int tick_do_timer_cpu __read_mostly = TICK_DO_TIMER_BOOT;
 50#ifdef CONFIG_NO_HZ_FULL
 51/*
 52 * tick_do_timer_boot_cpu indicates the boot CPU temporarily owns
 53 * tick_do_timer_cpu and it should be taken over by an eligible secondary
 54 * when one comes online.
 55 */
 56static int tick_do_timer_boot_cpu __read_mostly = -1;
 57#endif
 58
 59/*
 60 * Debugging: see timer_list.c
 61 */
 62struct tick_device *tick_get_device(int cpu)
 63{
 64	return &per_cpu(tick_cpu_device, cpu);
 65}
 66
 67/**
 68 * tick_is_oneshot_available - check for a oneshot capable event device
 69 */
 70int tick_is_oneshot_available(void)
 71{
 72	struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
 73
 74	if (!dev || !(dev->features & CLOCK_EVT_FEAT_ONESHOT))
 75		return 0;
 76	if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
 77		return 1;
 78	return tick_broadcast_oneshot_available();
 79}
 80
 81/*
 82 * Periodic tick
 83 */
 84static void tick_periodic(int cpu)
 85{
 86	if (tick_do_timer_cpu == cpu) {
 87		raw_spin_lock(&jiffies_lock);
 88		write_seqcount_begin(&jiffies_seq);
 89
 90		/* Keep track of the next tick event */
 91		tick_next_period = ktime_add(tick_next_period, tick_period);
 92
 93		do_timer(1);
 94		write_seqcount_end(&jiffies_seq);
 95		raw_spin_unlock(&jiffies_lock);
 96		update_wall_time();
 97	}
 98
 99	update_process_times(user_mode(get_irq_regs()));
100	profile_tick(CPU_PROFILING);
101}
102
103/*
104 * Event handler for periodic ticks
105 */
106void tick_handle_periodic(struct clock_event_device *dev)
107{
108	int cpu = smp_processor_id();
109	ktime_t next = dev->next_event;
110
111	tick_periodic(cpu);
112
113#if defined(CONFIG_HIGH_RES_TIMERS) || defined(CONFIG_NO_HZ_COMMON)
114	/*
115	 * The cpu might have transitioned to HIGHRES or NOHZ mode via
116	 * update_process_times() -> run_local_timers() ->
117	 * hrtimer_run_queues().
118	 */
119	if (dev->event_handler != tick_handle_periodic)
120		return;
121#endif
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(next, tick_period);
131
132		if (!clockevents_program_event(dev, next, 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, 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, CLOCK_EVT_STATE_ONESHOT);
172
173		for (;;) {
174			if (!clockevents_program_event(dev, next, false))
175				return;
176			next = ktime_add(next, tick_period);
177		}
178	}
179}
180
181#ifdef CONFIG_NO_HZ_FULL
182static void giveup_do_timer(void *info)
183{
184	int cpu = *(unsigned int *)info;
185
186	WARN_ON(tick_do_timer_cpu != smp_processor_id());
187
188	tick_do_timer_cpu = cpu;
189}
190
191static void tick_take_do_timer_from_boot(void)
192{
193	int cpu = smp_processor_id();
194	int from = tick_do_timer_boot_cpu;
195
196	if (from >= 0 && from != cpu)
197		smp_call_function_single(from, giveup_do_timer, &cpu, 1);
198}
199#endif
200
201/*
202 * Setup the tick device
203 */
204static void tick_setup_device(struct tick_device *td,
205			      struct clock_event_device *newdev, int cpu,
206			      const struct cpumask *cpumask)
207{
208	void (*handler)(struct clock_event_device *) = NULL;
209	ktime_t next_event = 0;
210
211	/*
212	 * First device setup ?
213	 */
214	if (!td->evtdev) {
215		/*
216		 * If no cpu took the do_timer update, assign it to
217		 * this cpu:
218		 */
219		if (tick_do_timer_cpu == TICK_DO_TIMER_BOOT) {
220			tick_do_timer_cpu = cpu;
221
222			tick_next_period = ktime_get();
223			tick_period = NSEC_PER_SEC / HZ;
224#ifdef CONFIG_NO_HZ_FULL
225			/*
226			 * The boot CPU may be nohz_full, in which case set
227			 * tick_do_timer_boot_cpu so the first housekeeping
228			 * secondary that comes up will take do_timer from
229			 * us.
230			 */
231			if (tick_nohz_full_cpu(cpu))
232				tick_do_timer_boot_cpu = cpu;
233
234		} else if (tick_do_timer_boot_cpu != -1 &&
235						!tick_nohz_full_cpu(cpu)) {
236			tick_take_do_timer_from_boot();
237			tick_do_timer_boot_cpu = -1;
238			WARN_ON(tick_do_timer_cpu != cpu);
239#endif
240		}
241
242		/*
243		 * Startup in periodic mode first.
244		 */
245		td->mode = TICKDEV_MODE_PERIODIC;
246	} else {
247		handler = td->evtdev->event_handler;
248		next_event = td->evtdev->next_event;
249		td->evtdev->event_handler = clockevents_handle_noop;
250	}
251
252	td->evtdev = newdev;
253
254	/*
255	 * When the device is not per cpu, pin the interrupt to the
256	 * current cpu:
257	 */
258	if (!cpumask_equal(newdev->cpumask, cpumask))
259		irq_set_affinity(newdev->irq, cpumask);
260
261	/*
262	 * When global broadcasting is active, check if the current
263	 * device is registered as a placeholder for broadcast mode.
264	 * This allows us to handle this x86 misfeature in a generic
265	 * way. This function also returns !=0 when we keep the
266	 * current active broadcast state for this CPU.
267	 */
268	if (tick_device_uses_broadcast(newdev, cpu))
269		return;
270
271	if (td->mode == TICKDEV_MODE_PERIODIC)
272		tick_setup_periodic(newdev, 0);
273	else
274		tick_setup_oneshot(newdev, handler, next_event);
275}
276
277void tick_install_replacement(struct clock_event_device *newdev)
278{
279	struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
280	int cpu = smp_processor_id();
281
282	clockevents_exchange_device(td->evtdev, newdev);
283	tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
284	if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
285		tick_oneshot_notify();
286}
287
288static bool tick_check_percpu(struct clock_event_device *curdev,
289			      struct clock_event_device *newdev, int cpu)
290{
291	if (!cpumask_test_cpu(cpu, newdev->cpumask))
292		return false;
293	if (cpumask_equal(newdev->cpumask, cpumask_of(cpu)))
294		return true;
295	/* Check if irq affinity can be set */
296	if (newdev->irq >= 0 && !irq_can_set_affinity(newdev->irq))
297		return false;
298	/* Prefer an existing cpu local device */
299	if (curdev && cpumask_equal(curdev->cpumask, cpumask_of(cpu)))
300		return false;
301	return true;
302}
303
304static bool tick_check_preferred(struct clock_event_device *curdev,
305				 struct clock_event_device *newdev)
306{
307	/* Prefer oneshot capable device */
308	if (!(newdev->features & CLOCK_EVT_FEAT_ONESHOT)) {
309		if (curdev && (curdev->features & CLOCK_EVT_FEAT_ONESHOT))
310			return false;
311		if (tick_oneshot_mode_active())
312			return false;
313	}
314
315	/*
316	 * Use the higher rated one, but prefer a CPU local device with a lower
317	 * rating than a non-CPU local device
318	 */
319	return !curdev ||
320		newdev->rating > curdev->rating ||
321	       !cpumask_equal(curdev->cpumask, newdev->cpumask);
322}
323
324/*
325 * Check whether the new device is a better fit than curdev. curdev
326 * can be NULL !
327 */
328bool tick_check_replacement(struct clock_event_device *curdev,
329			    struct clock_event_device *newdev)
330{
331	if (!tick_check_percpu(curdev, newdev, smp_processor_id()))
332		return false;
333
334	return tick_check_preferred(curdev, newdev);
335}
336
337/*
338 * Check, if the new registered device should be used. Called with
339 * clockevents_lock held and interrupts disabled.
340 */
341void tick_check_new_device(struct clock_event_device *newdev)
342{
343	struct clock_event_device *curdev;
344	struct tick_device *td;
345	int cpu;
346
347	cpu = smp_processor_id();
348	td = &per_cpu(tick_cpu_device, cpu);
349	curdev = td->evtdev;
350
351	/* cpu local device ? */
352	if (!tick_check_percpu(curdev, newdev, cpu))
353		goto out_bc;
354
355	/* Preference decision */
356	if (!tick_check_preferred(curdev, newdev))
357		goto out_bc;
358
359	if (!try_module_get(newdev->owner))
360		return;
361
362	/*
363	 * Replace the eventually existing device by the new
364	 * device. If the current device is the broadcast device, do
365	 * not give it back to the clockevents layer !
366	 */
367	if (tick_is_broadcast_device(curdev)) {
368		clockevents_shutdown(curdev);
369		curdev = NULL;
370	}
371	clockevents_exchange_device(curdev, newdev);
372	tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
373	if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
374		tick_oneshot_notify();
375	return;
376
377out_bc:
378	/*
379	 * Can the new device be used as a broadcast device ?
380	 */
381	tick_install_broadcast_device(newdev);
382}
383
384/**
385 * tick_broadcast_oneshot_control - Enter/exit broadcast oneshot mode
386 * @state:	The target state (enter/exit)
387 *
388 * The system enters/leaves a state, where affected devices might stop
389 * Returns 0 on success, -EBUSY if the cpu is used to broadcast wakeups.
390 *
391 * Called with interrupts disabled, so clockevents_lock is not
392 * required here because the local clock event device cannot go away
393 * under us.
394 */
395int tick_broadcast_oneshot_control(enum tick_broadcast_state state)
396{
397	struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
398
399	if (!(td->evtdev->features & CLOCK_EVT_FEAT_C3STOP))
400		return 0;
401
402	return __tick_broadcast_oneshot_control(state);
403}
404EXPORT_SYMBOL_GPL(tick_broadcast_oneshot_control);
405
406#ifdef CONFIG_HOTPLUG_CPU
407/*
408 * Transfer the do_timer job away from a dying cpu.
409 *
410 * Called with interrupts disabled. Not locking required. If
411 * tick_do_timer_cpu is owned by this cpu, nothing can change it.
412 */
413void tick_handover_do_timer(void)
414{
415	if (tick_do_timer_cpu == smp_processor_id()) {
416		int cpu = cpumask_first(cpu_online_mask);
417
418		tick_do_timer_cpu = (cpu < nr_cpu_ids) ? cpu :
419			TICK_DO_TIMER_NONE;
420	}
421}
422
423/*
424 * Shutdown an event device on a given cpu:
425 *
426 * This is called on a life CPU, when a CPU is dead. So we cannot
427 * access the hardware device itself.
428 * We just set the mode and remove it from the lists.
429 */
430void tick_shutdown(unsigned int cpu)
431{
432	struct tick_device *td = &per_cpu(tick_cpu_device, cpu);
433	struct clock_event_device *dev = td->evtdev;
434
435	td->mode = TICKDEV_MODE_PERIODIC;
436	if (dev) {
437		/*
438		 * Prevent that the clock events layer tries to call
439		 * the set mode function!
440		 */
441		clockevent_set_state(dev, CLOCK_EVT_STATE_DETACHED);
442		clockevents_exchange_device(dev, NULL);
443		dev->event_handler = clockevents_handle_noop;
444		td->evtdev = NULL;
445	}
446}
447#endif
448
449/**
450 * tick_suspend_local - Suspend the local tick device
451 *
452 * Called from the local cpu for freeze with interrupts disabled.
453 *
454 * No locks required. Nothing can change the per cpu device.
455 */
456void tick_suspend_local(void)
457{
458	struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
459
460	clockevents_shutdown(td->evtdev);
461}
462
463/**
464 * tick_resume_local - Resume the local tick device
465 *
466 * Called from the local CPU for unfreeze or XEN resume magic.
467 *
468 * No locks required. Nothing can change the per cpu device.
469 */
470void tick_resume_local(void)
471{
472	struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
473	bool broadcast = tick_resume_check_broadcast();
474
475	clockevents_tick_resume(td->evtdev);
476	if (!broadcast) {
477		if (td->mode == TICKDEV_MODE_PERIODIC)
478			tick_setup_periodic(td->evtdev, 0);
479		else
480			tick_resume_oneshot();
481	}
482}
483
484/**
485 * tick_suspend - Suspend the tick and the broadcast device
486 *
487 * Called from syscore_suspend() via timekeeping_suspend with only one
488 * CPU online and interrupts disabled or from tick_unfreeze() under
489 * tick_freeze_lock.
490 *
491 * No locks required. Nothing can change the per cpu device.
492 */
493void tick_suspend(void)
494{
495	tick_suspend_local();
496	tick_suspend_broadcast();
497}
498
499/**
500 * tick_resume - Resume the tick and the broadcast device
501 *
502 * Called from syscore_resume() via timekeeping_resume with only one
503 * CPU online and interrupts disabled.
504 *
505 * No locks required. Nothing can change the per cpu device.
506 */
507void tick_resume(void)
508{
509	tick_resume_broadcast();
510	tick_resume_local();
511}
512
513#ifdef CONFIG_SUSPEND
514static DEFINE_RAW_SPINLOCK(tick_freeze_lock);
515static unsigned int tick_freeze_depth;
516
517/**
518 * tick_freeze - Suspend the local tick and (possibly) timekeeping.
519 *
520 * Check if this is the last online CPU executing the function and if so,
521 * suspend timekeeping.  Otherwise suspend the local tick.
522 *
523 * Call with interrupts disabled.  Must be balanced with %tick_unfreeze().
524 * Interrupts must not be enabled before the subsequent %tick_unfreeze().
525 */
526void tick_freeze(void)
527{
528	raw_spin_lock(&tick_freeze_lock);
529
530	tick_freeze_depth++;
531	if (tick_freeze_depth == num_online_cpus()) {
532		trace_suspend_resume(TPS("timekeeping_freeze"),
533				     smp_processor_id(), true);
534		system_state = SYSTEM_SUSPEND;
535		sched_clock_suspend();
536		timekeeping_suspend();
537	} else {
538		tick_suspend_local();
539	}
540
541	raw_spin_unlock(&tick_freeze_lock);
542}
543
544/**
545 * tick_unfreeze - Resume the local tick and (possibly) timekeeping.
546 *
547 * Check if this is the first CPU executing the function and if so, resume
548 * timekeeping.  Otherwise resume the local tick.
549 *
550 * Call with interrupts disabled.  Must be balanced with %tick_freeze().
551 * Interrupts must not be enabled after the preceding %tick_freeze().
552 */
553void tick_unfreeze(void)
554{
555	raw_spin_lock(&tick_freeze_lock);
556
557	if (tick_freeze_depth == num_online_cpus()) {
558		timekeeping_resume();
559		sched_clock_resume();
560		system_state = SYSTEM_RUNNING;
561		trace_suspend_resume(TPS("timekeeping_freeze"),
562				     smp_processor_id(), false);
563	} else {
564		touch_softlockup_watchdog();
565		tick_resume_local();
566	}
567
568	tick_freeze_depth--;
569
570	raw_spin_unlock(&tick_freeze_lock);
571}
572#endif /* CONFIG_SUSPEND */
573
574/**
575 * tick_init - initialize the tick control
576 */
577void __init tick_init(void)
578{
579	tick_broadcast_init();
580	tick_nohz_init();
581}