tick-broadcast.c revision 1f73a9806bdd07a5106409bbcab3884078bd34fe
1/*
2 * linux/kernel/time/tick-broadcast.c
3 *
4 * This file contains functions which emulate a local clock-event
5 * device via a broadcast event source.
6 *
7 * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
8 * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
9 * Copyright(C) 2006-2007, Timesys Corp., Thomas Gleixner
10 *
11 * This code is licenced under the GPL version 2. For details see
12 * kernel-base/COPYING.
13 */
14#include <linux/cpu.h>
15#include <linux/err.h>
16#include <linux/hrtimer.h>
17#include <linux/interrupt.h>
18#include <linux/percpu.h>
19#include <linux/profile.h>
20#include <linux/sched.h>
21#include <linux/smp.h>
22#include <linux/module.h>
23
24#include "tick-internal.h"
25
26/*
27 * Broadcast support for broken x86 hardware, where the local apic
28 * timer stops in C3 state.
29 */
30
31static struct tick_device tick_broadcast_device;
32static cpumask_var_t tick_broadcast_mask;
33static cpumask_var_t tmpmask;
34static DEFINE_RAW_SPINLOCK(tick_broadcast_lock);
35static int tick_broadcast_force;
36
37#ifdef CONFIG_TICK_ONESHOT
38static void tick_broadcast_clear_oneshot(int cpu);
39#else
40static inline void tick_broadcast_clear_oneshot(int cpu) { }
41#endif
42
43/*
44 * Debugging: see timer_list.c
45 */
46struct tick_device *tick_get_broadcast_device(void)
47{
48	return &tick_broadcast_device;
49}
50
51struct cpumask *tick_get_broadcast_mask(void)
52{
53	return tick_broadcast_mask;
54}
55
56/*
57 * Start the device in periodic mode
58 */
59static void tick_broadcast_start_periodic(struct clock_event_device *bc)
60{
61	if (bc)
62		tick_setup_periodic(bc, 1);
63}
64
65/*
66 * Check, if the device can be utilized as broadcast device:
67 */
68static bool tick_check_broadcast_device(struct clock_event_device *curdev,
69					struct clock_event_device *newdev)
70{
71	if ((newdev->features & CLOCK_EVT_FEAT_DUMMY) ||
72	    (newdev->features & CLOCK_EVT_FEAT_C3STOP))
73		return false;
74
75	if (tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT &&
76	    !(newdev->features & CLOCK_EVT_FEAT_ONESHOT))
77		return false;
78
79	return !curdev || newdev->rating > curdev->rating;
80}
81
82/*
83 * Conditionally install/replace broadcast device
84 */
85void tick_install_broadcast_device(struct clock_event_device *dev)
86{
87	struct clock_event_device *cur = tick_broadcast_device.evtdev;
88
89	if (!tick_check_broadcast_device(cur, dev))
90		return;
91
92	if (!try_module_get(dev->owner))
93		return;
94
95	clockevents_exchange_device(cur, dev);
96	if (cur)
97		cur->event_handler = clockevents_handle_noop;
98	tick_broadcast_device.evtdev = dev;
99	if (!cpumask_empty(tick_broadcast_mask))
100		tick_broadcast_start_periodic(dev);
101	/*
102	 * Inform all cpus about this. We might be in a situation
103	 * where we did not switch to oneshot mode because the per cpu
104	 * devices are affected by CLOCK_EVT_FEAT_C3STOP and the lack
105	 * of a oneshot capable broadcast device. Without that
106	 * notification the systems stays stuck in periodic mode
107	 * forever.
108	 */
109	if (dev->features & CLOCK_EVT_FEAT_ONESHOT)
110		tick_clock_notify();
111}
112
113/*
114 * Check, if the device is the broadcast device
115 */
116int tick_is_broadcast_device(struct clock_event_device *dev)
117{
118	return (dev && tick_broadcast_device.evtdev == dev);
119}
120
121static void err_broadcast(const struct cpumask *mask)
122{
123	pr_crit_once("Failed to broadcast timer tick. Some CPUs may be unresponsive.\n");
124}
125
126static void tick_device_setup_broadcast_func(struct clock_event_device *dev)
127{
128	if (!dev->broadcast)
129		dev->broadcast = tick_broadcast;
130	if (!dev->broadcast) {
131		pr_warn_once("%s depends on broadcast, but no broadcast function available\n",
132			     dev->name);
133		dev->broadcast = err_broadcast;
134	}
135}
136
137/*
138 * Check, if the device is disfunctional and a place holder, which
139 * needs to be handled by the broadcast device.
140 */
141int tick_device_uses_broadcast(struct clock_event_device *dev, int cpu)
142{
143	unsigned long flags;
144	int ret = 0;
145
146	raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
147
148	/*
149	 * Devices might be registered with both periodic and oneshot
150	 * mode disabled. This signals, that the device needs to be
151	 * operated from the broadcast device and is a placeholder for
152	 * the cpu local device.
153	 */
154	if (!tick_device_is_functional(dev)) {
155		dev->event_handler = tick_handle_periodic;
156		tick_device_setup_broadcast_func(dev);
157		cpumask_set_cpu(cpu, tick_broadcast_mask);
158		tick_broadcast_start_periodic(tick_broadcast_device.evtdev);
159		ret = 1;
160	} else {
161		/*
162		 * When the new device is not affected by the stop
163		 * feature and the cpu is marked in the broadcast mask
164		 * then clear the broadcast bit.
165		 */
166		if (!(dev->features & CLOCK_EVT_FEAT_C3STOP)) {
167			int cpu = smp_processor_id();
168			cpumask_clear_cpu(cpu, tick_broadcast_mask);
169			tick_broadcast_clear_oneshot(cpu);
170		} else {
171			tick_device_setup_broadcast_func(dev);
172		}
173	}
174	raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
175	return ret;
176}
177
178#ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
179int tick_receive_broadcast(void)
180{
181	struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
182	struct clock_event_device *evt = td->evtdev;
183
184	if (!evt)
185		return -ENODEV;
186
187	if (!evt->event_handler)
188		return -EINVAL;
189
190	evt->event_handler(evt);
191	return 0;
192}
193#endif
194
195/*
196 * Broadcast the event to the cpus, which are set in the mask (mangled).
197 */
198static void tick_do_broadcast(struct cpumask *mask)
199{
200	int cpu = smp_processor_id();
201	struct tick_device *td;
202
203	/*
204	 * Check, if the current cpu is in the mask
205	 */
206	if (cpumask_test_cpu(cpu, mask)) {
207		cpumask_clear_cpu(cpu, mask);
208		td = &per_cpu(tick_cpu_device, cpu);
209		td->evtdev->event_handler(td->evtdev);
210	}
211
212	if (!cpumask_empty(mask)) {
213		/*
214		 * It might be necessary to actually check whether the devices
215		 * have different broadcast functions. For now, just use the
216		 * one of the first device. This works as long as we have this
217		 * misfeature only on x86 (lapic)
218		 */
219		td = &per_cpu(tick_cpu_device, cpumask_first(mask));
220		td->evtdev->broadcast(mask);
221	}
222}
223
224/*
225 * Periodic broadcast:
226 * - invoke the broadcast handlers
227 */
228static void tick_do_periodic_broadcast(void)
229{
230	raw_spin_lock(&tick_broadcast_lock);
231
232	cpumask_and(tmpmask, cpu_online_mask, tick_broadcast_mask);
233	tick_do_broadcast(tmpmask);
234
235	raw_spin_unlock(&tick_broadcast_lock);
236}
237
238/*
239 * Event handler for periodic broadcast ticks
240 */
241static void tick_handle_periodic_broadcast(struct clock_event_device *dev)
242{
243	ktime_t next;
244
245	tick_do_periodic_broadcast();
246
247	/*
248	 * The device is in periodic mode. No reprogramming necessary:
249	 */
250	if (dev->mode == CLOCK_EVT_MODE_PERIODIC)
251		return;
252
253	/*
254	 * Setup the next period for devices, which do not have
255	 * periodic mode. We read dev->next_event first and add to it
256	 * when the event already expired. clockevents_program_event()
257	 * sets dev->next_event only when the event is really
258	 * programmed to the device.
259	 */
260	for (next = dev->next_event; ;) {
261		next = ktime_add(next, tick_period);
262
263		if (!clockevents_program_event(dev, next, false))
264			return;
265		tick_do_periodic_broadcast();
266	}
267}
268
269/*
270 * Powerstate information: The system enters/leaves a state, where
271 * affected devices might stop
272 */
273static void tick_do_broadcast_on_off(unsigned long *reason)
274{
275	struct clock_event_device *bc, *dev;
276	struct tick_device *td;
277	unsigned long flags;
278	int cpu, bc_stopped;
279
280	raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
281
282	cpu = smp_processor_id();
283	td = &per_cpu(tick_cpu_device, cpu);
284	dev = td->evtdev;
285	bc = tick_broadcast_device.evtdev;
286
287	/*
288	 * Is the device not affected by the powerstate ?
289	 */
290	if (!dev || !(dev->features & CLOCK_EVT_FEAT_C3STOP))
291		goto out;
292
293	if (!tick_device_is_functional(dev))
294		goto out;
295
296	bc_stopped = cpumask_empty(tick_broadcast_mask);
297
298	switch (*reason) {
299	case CLOCK_EVT_NOTIFY_BROADCAST_ON:
300	case CLOCK_EVT_NOTIFY_BROADCAST_FORCE:
301		if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_mask)) {
302			if (tick_broadcast_device.mode ==
303			    TICKDEV_MODE_PERIODIC)
304				clockevents_shutdown(dev);
305		}
306		if (*reason == CLOCK_EVT_NOTIFY_BROADCAST_FORCE)
307			tick_broadcast_force = 1;
308		break;
309	case CLOCK_EVT_NOTIFY_BROADCAST_OFF:
310		if (!tick_broadcast_force &&
311		    cpumask_test_and_clear_cpu(cpu, tick_broadcast_mask)) {
312			if (tick_broadcast_device.mode ==
313			    TICKDEV_MODE_PERIODIC)
314				tick_setup_periodic(dev, 0);
315		}
316		break;
317	}
318
319	if (cpumask_empty(tick_broadcast_mask)) {
320		if (!bc_stopped)
321			clockevents_shutdown(bc);
322	} else if (bc_stopped) {
323		if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
324			tick_broadcast_start_periodic(bc);
325		else
326			tick_broadcast_setup_oneshot(bc);
327	}
328out:
329	raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
330}
331
332/*
333 * Powerstate information: The system enters/leaves a state, where
334 * affected devices might stop.
335 */
336void tick_broadcast_on_off(unsigned long reason, int *oncpu)
337{
338	if (!cpumask_test_cpu(*oncpu, cpu_online_mask))
339		printk(KERN_ERR "tick-broadcast: ignoring broadcast for "
340		       "offline CPU #%d\n", *oncpu);
341	else
342		tick_do_broadcast_on_off(&reason);
343}
344
345/*
346 * Set the periodic handler depending on broadcast on/off
347 */
348void tick_set_periodic_handler(struct clock_event_device *dev, int broadcast)
349{
350	if (!broadcast)
351		dev->event_handler = tick_handle_periodic;
352	else
353		dev->event_handler = tick_handle_periodic_broadcast;
354}
355
356/*
357 * Remove a CPU from broadcasting
358 */
359void tick_shutdown_broadcast(unsigned int *cpup)
360{
361	struct clock_event_device *bc;
362	unsigned long flags;
363	unsigned int cpu = *cpup;
364
365	raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
366
367	bc = tick_broadcast_device.evtdev;
368	cpumask_clear_cpu(cpu, tick_broadcast_mask);
369
370	if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
371		if (bc && cpumask_empty(tick_broadcast_mask))
372			clockevents_shutdown(bc);
373	}
374
375	raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
376}
377
378void tick_suspend_broadcast(void)
379{
380	struct clock_event_device *bc;
381	unsigned long flags;
382
383	raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
384
385	bc = tick_broadcast_device.evtdev;
386	if (bc)
387		clockevents_shutdown(bc);
388
389	raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
390}
391
392int tick_resume_broadcast(void)
393{
394	struct clock_event_device *bc;
395	unsigned long flags;
396	int broadcast = 0;
397
398	raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
399
400	bc = tick_broadcast_device.evtdev;
401
402	if (bc) {
403		clockevents_set_mode(bc, CLOCK_EVT_MODE_RESUME);
404
405		switch (tick_broadcast_device.mode) {
406		case TICKDEV_MODE_PERIODIC:
407			if (!cpumask_empty(tick_broadcast_mask))
408				tick_broadcast_start_periodic(bc);
409			broadcast = cpumask_test_cpu(smp_processor_id(),
410						     tick_broadcast_mask);
411			break;
412		case TICKDEV_MODE_ONESHOT:
413			if (!cpumask_empty(tick_broadcast_mask))
414				broadcast = tick_resume_broadcast_oneshot(bc);
415			break;
416		}
417	}
418	raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
419
420	return broadcast;
421}
422
423
424#ifdef CONFIG_TICK_ONESHOT
425
426static cpumask_var_t tick_broadcast_oneshot_mask;
427static cpumask_var_t tick_broadcast_pending_mask;
428static cpumask_var_t tick_broadcast_force_mask;
429
430/*
431 * Exposed for debugging: see timer_list.c
432 */
433struct cpumask *tick_get_broadcast_oneshot_mask(void)
434{
435	return tick_broadcast_oneshot_mask;
436}
437
438/*
439 * Called before going idle with interrupts disabled. Checks whether a
440 * broadcast event from the other core is about to happen. We detected
441 * that in tick_broadcast_oneshot_control(). The callsite can use this
442 * to avoid a deep idle transition as we are about to get the
443 * broadcast IPI right away.
444 */
445int tick_check_broadcast_expired(void)
446{
447	return cpumask_test_cpu(smp_processor_id(), tick_broadcast_force_mask);
448}
449
450/*
451 * Set broadcast interrupt affinity
452 */
453static void tick_broadcast_set_affinity(struct clock_event_device *bc,
454					const struct cpumask *cpumask)
455{
456	if (!(bc->features & CLOCK_EVT_FEAT_DYNIRQ))
457		return;
458
459	if (cpumask_equal(bc->cpumask, cpumask))
460		return;
461
462	bc->cpumask = cpumask;
463	irq_set_affinity(bc->irq, bc->cpumask);
464}
465
466static int tick_broadcast_set_event(struct clock_event_device *bc, int cpu,
467				    ktime_t expires, int force)
468{
469	int ret;
470
471	if (bc->mode != CLOCK_EVT_MODE_ONESHOT)
472		clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
473
474	ret = clockevents_program_event(bc, expires, force);
475	if (!ret)
476		tick_broadcast_set_affinity(bc, cpumask_of(cpu));
477	return ret;
478}
479
480int tick_resume_broadcast_oneshot(struct clock_event_device *bc)
481{
482	clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
483	return 0;
484}
485
486/*
487 * Called from irq_enter() when idle was interrupted to reenable the
488 * per cpu device.
489 */
490void tick_check_oneshot_broadcast(int cpu)
491{
492	if (cpumask_test_cpu(cpu, tick_broadcast_oneshot_mask)) {
493		struct tick_device *td = &per_cpu(tick_cpu_device, cpu);
494
495		/*
496		 * We might be in the middle of switching over from
497		 * periodic to oneshot. If the CPU has not yet
498		 * switched over, leave the device alone.
499		 */
500		if (td->mode == TICKDEV_MODE_ONESHOT) {
501			clockevents_set_mode(td->evtdev,
502					     CLOCK_EVT_MODE_ONESHOT);
503		}
504	}
505}
506
507/*
508 * Handle oneshot mode broadcasting
509 */
510static void tick_handle_oneshot_broadcast(struct clock_event_device *dev)
511{
512	struct tick_device *td;
513	ktime_t now, next_event;
514	int cpu, next_cpu = 0;
515
516	raw_spin_lock(&tick_broadcast_lock);
517again:
518	dev->next_event.tv64 = KTIME_MAX;
519	next_event.tv64 = KTIME_MAX;
520	cpumask_clear(tmpmask);
521	now = ktime_get();
522	/* Find all expired events */
523	for_each_cpu(cpu, tick_broadcast_oneshot_mask) {
524		td = &per_cpu(tick_cpu_device, cpu);
525		if (td->evtdev->next_event.tv64 <= now.tv64) {
526			cpumask_set_cpu(cpu, tmpmask);
527			/*
528			 * Mark the remote cpu in the pending mask, so
529			 * it can avoid reprogramming the cpu local
530			 * timer in tick_broadcast_oneshot_control().
531			 */
532			cpumask_set_cpu(cpu, tick_broadcast_pending_mask);
533		} else if (td->evtdev->next_event.tv64 < next_event.tv64) {
534			next_event.tv64 = td->evtdev->next_event.tv64;
535			next_cpu = cpu;
536		}
537	}
538
539	/* Take care of enforced broadcast requests */
540	cpumask_or(tmpmask, tmpmask, tick_broadcast_force_mask);
541	cpumask_clear(tick_broadcast_force_mask);
542
543	/*
544	 * Sanity check. Catch the case where we try to broadcast to
545	 * offline cpus.
546	 */
547	if (WARN_ON_ONCE(!cpumask_subset(tmpmask, cpu_online_mask)))
548		cpumask_and(tmpmask, tmpmask, cpu_online_mask);
549
550	/*
551	 * Wakeup the cpus which have an expired event.
552	 */
553	tick_do_broadcast(tmpmask);
554
555	/*
556	 * Two reasons for reprogram:
557	 *
558	 * - The global event did not expire any CPU local
559	 * events. This happens in dyntick mode, as the maximum PIT
560	 * delta is quite small.
561	 *
562	 * - There are pending events on sleeping CPUs which were not
563	 * in the event mask
564	 */
565	if (next_event.tv64 != KTIME_MAX) {
566		/*
567		 * Rearm the broadcast device. If event expired,
568		 * repeat the above
569		 */
570		if (tick_broadcast_set_event(dev, next_cpu, next_event, 0))
571			goto again;
572	}
573	raw_spin_unlock(&tick_broadcast_lock);
574}
575
576/*
577 * Powerstate information: The system enters/leaves a state, where
578 * affected devices might stop
579 */
580void tick_broadcast_oneshot_control(unsigned long reason)
581{
582	struct clock_event_device *bc, *dev;
583	struct tick_device *td;
584	unsigned long flags;
585	ktime_t now;
586	int cpu;
587
588	/*
589	 * Periodic mode does not care about the enter/exit of power
590	 * states
591	 */
592	if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
593		return;
594
595	/*
596	 * We are called with preemtion disabled from the depth of the
597	 * idle code, so we can't be moved away.
598	 */
599	cpu = smp_processor_id();
600	td = &per_cpu(tick_cpu_device, cpu);
601	dev = td->evtdev;
602
603	if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
604		return;
605
606	bc = tick_broadcast_device.evtdev;
607
608	raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
609	if (reason == CLOCK_EVT_NOTIFY_BROADCAST_ENTER) {
610		WARN_ON_ONCE(cpumask_test_cpu(cpu, tick_broadcast_pending_mask));
611		if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_oneshot_mask)) {
612			clockevents_set_mode(dev, CLOCK_EVT_MODE_SHUTDOWN);
613			/*
614			 * We only reprogram the broadcast timer if we
615			 * did not mark ourself in the force mask and
616			 * if the cpu local event is earlier than the
617			 * broadcast event. If the current CPU is in
618			 * the force mask, then we are going to be
619			 * woken by the IPI right away.
620			 */
621			if (!cpumask_test_cpu(cpu, tick_broadcast_force_mask) &&
622			    dev->next_event.tv64 < bc->next_event.tv64)
623				tick_broadcast_set_event(bc, cpu, dev->next_event, 1);
624		}
625	} else {
626		if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_oneshot_mask)) {
627			clockevents_set_mode(dev, CLOCK_EVT_MODE_ONESHOT);
628			if (dev->next_event.tv64 == KTIME_MAX)
629				goto out;
630			/*
631			 * The cpu which was handling the broadcast
632			 * timer marked this cpu in the broadcast
633			 * pending mask and fired the broadcast
634			 * IPI. So we are going to handle the expired
635			 * event anyway via the broadcast IPI
636			 * handler. No need to reprogram the timer
637			 * with an already expired event.
638			 */
639			if (cpumask_test_and_clear_cpu(cpu,
640				       tick_broadcast_pending_mask))
641				goto out;
642
643			/*
644			 * If the pending bit is not set, then we are
645			 * either the CPU handling the broadcast
646			 * interrupt or we got woken by something else.
647			 *
648			 * We are not longer in the broadcast mask, so
649			 * if the cpu local expiry time is already
650			 * reached, we would reprogram the cpu local
651			 * timer with an already expired event.
652			 *
653			 * This can lead to a ping-pong when we return
654			 * to idle and therefor rearm the broadcast
655			 * timer before the cpu local timer was able
656			 * to fire. This happens because the forced
657			 * reprogramming makes sure that the event
658			 * will happen in the future and depending on
659			 * the min_delta setting this might be far
660			 * enough out that the ping-pong starts.
661			 *
662			 * If the cpu local next_event has expired
663			 * then we know that the broadcast timer
664			 * next_event has expired as well and
665			 * broadcast is about to be handled. So we
666			 * avoid reprogramming and enforce that the
667			 * broadcast handler, which did not run yet,
668			 * will invoke the cpu local handler.
669			 *
670			 * We cannot call the handler directly from
671			 * here, because we might be in a NOHZ phase
672			 * and we did not go through the irq_enter()
673			 * nohz fixups.
674			 */
675			now = ktime_get();
676			if (dev->next_event.tv64 <= now.tv64) {
677				cpumask_set_cpu(cpu, tick_broadcast_force_mask);
678				goto out;
679			}
680			/*
681			 * We got woken by something else. Reprogram
682			 * the cpu local timer device.
683			 */
684			tick_program_event(dev->next_event, 1);
685		}
686	}
687out:
688	raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
689}
690
691/*
692 * Reset the one shot broadcast for a cpu
693 *
694 * Called with tick_broadcast_lock held
695 */
696static void tick_broadcast_clear_oneshot(int cpu)
697{
698	cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
699}
700
701static void tick_broadcast_init_next_event(struct cpumask *mask,
702					   ktime_t expires)
703{
704	struct tick_device *td;
705	int cpu;
706
707	for_each_cpu(cpu, mask) {
708		td = &per_cpu(tick_cpu_device, cpu);
709		if (td->evtdev)
710			td->evtdev->next_event = expires;
711	}
712}
713
714/**
715 * tick_broadcast_setup_oneshot - setup the broadcast device
716 */
717void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
718{
719	int cpu = smp_processor_id();
720
721	/* Set it up only once ! */
722	if (bc->event_handler != tick_handle_oneshot_broadcast) {
723		int was_periodic = bc->mode == CLOCK_EVT_MODE_PERIODIC;
724
725		bc->event_handler = tick_handle_oneshot_broadcast;
726
727		/* Take the do_timer update */
728		if (!tick_nohz_full_cpu(cpu))
729			tick_do_timer_cpu = cpu;
730
731		/*
732		 * We must be careful here. There might be other CPUs
733		 * waiting for periodic broadcast. We need to set the
734		 * oneshot_mask bits for those and program the
735		 * broadcast device to fire.
736		 */
737		cpumask_copy(tmpmask, tick_broadcast_mask);
738		cpumask_clear_cpu(cpu, tmpmask);
739		cpumask_or(tick_broadcast_oneshot_mask,
740			   tick_broadcast_oneshot_mask, tmpmask);
741
742		if (was_periodic && !cpumask_empty(tmpmask)) {
743			clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
744			tick_broadcast_init_next_event(tmpmask,
745						       tick_next_period);
746			tick_broadcast_set_event(bc, cpu, tick_next_period, 1);
747		} else
748			bc->next_event.tv64 = KTIME_MAX;
749	} else {
750		/*
751		 * The first cpu which switches to oneshot mode sets
752		 * the bit for all other cpus which are in the general
753		 * (periodic) broadcast mask. So the bit is set and
754		 * would prevent the first broadcast enter after this
755		 * to program the bc device.
756		 */
757		tick_broadcast_clear_oneshot(cpu);
758	}
759}
760
761/*
762 * Select oneshot operating mode for the broadcast device
763 */
764void tick_broadcast_switch_to_oneshot(void)
765{
766	struct clock_event_device *bc;
767	unsigned long flags;
768
769	raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
770
771	tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
772	bc = tick_broadcast_device.evtdev;
773	if (bc)
774		tick_broadcast_setup_oneshot(bc);
775
776	raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
777}
778
779
780/*
781 * Remove a dead CPU from broadcasting
782 */
783void tick_shutdown_broadcast_oneshot(unsigned int *cpup)
784{
785	unsigned long flags;
786	unsigned int cpu = *cpup;
787
788	raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
789
790	/*
791	 * Clear the broadcast masks for the dead cpu, but do not stop
792	 * the broadcast device!
793	 */
794	cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
795	cpumask_clear_cpu(cpu, tick_broadcast_pending_mask);
796	cpumask_clear_cpu(cpu, tick_broadcast_force_mask);
797
798	raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
799}
800
801/*
802 * Check, whether the broadcast device is in one shot mode
803 */
804int tick_broadcast_oneshot_active(void)
805{
806	return tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT;
807}
808
809/*
810 * Check whether the broadcast device supports oneshot.
811 */
812bool tick_broadcast_oneshot_available(void)
813{
814	struct clock_event_device *bc = tick_broadcast_device.evtdev;
815
816	return bc ? bc->features & CLOCK_EVT_FEAT_ONESHOT : false;
817}
818
819#endif
820
821void __init tick_broadcast_init(void)
822{
823	zalloc_cpumask_var(&tick_broadcast_mask, GFP_NOWAIT);
824	zalloc_cpumask_var(&tmpmask, GFP_NOWAIT);
825#ifdef CONFIG_TICK_ONESHOT
826	zalloc_cpumask_var(&tick_broadcast_oneshot_mask, GFP_NOWAIT);
827	zalloc_cpumask_var(&tick_broadcast_pending_mask, GFP_NOWAIT);
828	zalloc_cpumask_var(&tick_broadcast_force_mask, GFP_NOWAIT);
829#endif
830}
831