2 * linux/kernel/time/tick-broadcast.c
4 * This file contains functions which emulate a local clock-event
5 * device via a broadcast event source.
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
11 * This code is licenced under the GPL version 2. For details see
12 * kernel-base/COPYING.
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>
23 #include "tick-internal.h"
26 * Broadcast support for broken x86 hardware, where the local apic
27 * timer stops in C3 state.
30 static struct tick_device tick_broadcast_device;
31 static cpumask_var_t tick_broadcast_mask;
32 static cpumask_var_t tmpmask;
33 static DEFINE_RAW_SPINLOCK(tick_broadcast_lock);
34 static int tick_broadcast_force;
36 #ifdef CONFIG_TICK_ONESHOT
37 static void tick_broadcast_clear_oneshot(int cpu);
39 static inline void tick_broadcast_clear_oneshot(int cpu) { }
43 * Debugging: see timer_list.c
45 struct tick_device *tick_get_broadcast_device(void)
47 return &tick_broadcast_device;
50 struct cpumask *tick_get_broadcast_mask(void)
52 return tick_broadcast_mask;
56 * Start the device in periodic mode
58 static void tick_broadcast_start_periodic(struct clock_event_device *bc)
61 tick_setup_periodic(bc, 1);
65 * Check, if the device can be utilized as broadcast device:
67 void tick_install_broadcast_device(struct clock_event_device *dev)
69 struct clock_event_device *cur = tick_broadcast_device.evtdev;
71 if ((dev->features & CLOCK_EVT_FEAT_DUMMY) ||
72 (tick_broadcast_device.evtdev &&
73 tick_broadcast_device.evtdev->rating >= dev->rating) ||
74 (dev->features & CLOCK_EVT_FEAT_C3STOP))
77 clockevents_exchange_device(tick_broadcast_device.evtdev, dev);
79 cur->event_handler = clockevents_handle_noop;
80 tick_broadcast_device.evtdev = dev;
81 if (!cpumask_empty(tick_broadcast_mask))
82 tick_broadcast_start_periodic(dev);
84 * Inform all cpus about this. We might be in a situation
85 * where we did not switch to oneshot mode because the per cpu
86 * devices are affected by CLOCK_EVT_FEAT_C3STOP and the lack
87 * of a oneshot capable broadcast device. Without that
88 * notification the systems stays stuck in periodic mode
91 if (dev->features & CLOCK_EVT_FEAT_ONESHOT)
96 * Check, if the device is the broadcast device
98 int tick_is_broadcast_device(struct clock_event_device *dev)
100 return (dev && tick_broadcast_device.evtdev == dev);
103 static void err_broadcast(const struct cpumask *mask)
105 pr_crit_once("Failed to broadcast timer tick. Some CPUs may be unresponsive.\n");
108 static void tick_device_setup_broadcast_func(struct clock_event_device *dev)
111 dev->broadcast = tick_broadcast;
112 if (!dev->broadcast) {
113 pr_warn_once("%s depends on broadcast, but no broadcast function available\n",
115 dev->broadcast = err_broadcast;
120 * Check, if the device is disfunctional and a place holder, which
121 * needs to be handled by the broadcast device.
123 int tick_device_uses_broadcast(struct clock_event_device *dev, int cpu)
128 raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
131 * Devices might be registered with both periodic and oneshot
132 * mode disabled. This signals, that the device needs to be
133 * operated from the broadcast device and is a placeholder for
134 * the cpu local device.
136 if (!tick_device_is_functional(dev)) {
137 dev->event_handler = tick_handle_periodic;
138 tick_device_setup_broadcast_func(dev);
139 cpumask_set_cpu(cpu, tick_broadcast_mask);
140 tick_broadcast_start_periodic(tick_broadcast_device.evtdev);
144 * When the new device is not affected by the stop
145 * feature and the cpu is marked in the broadcast mask
146 * then clear the broadcast bit.
148 if (!(dev->features & CLOCK_EVT_FEAT_C3STOP)) {
149 int cpu = smp_processor_id();
150 cpumask_clear_cpu(cpu, tick_broadcast_mask);
151 tick_broadcast_clear_oneshot(cpu);
153 tick_device_setup_broadcast_func(dev);
156 raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
160 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
161 int tick_receive_broadcast(void)
163 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
164 struct clock_event_device *evt = td->evtdev;
169 if (!evt->event_handler)
172 evt->event_handler(evt);
178 * Broadcast the event to the cpus, which are set in the mask (mangled).
180 static void tick_do_broadcast(struct cpumask *mask)
182 int cpu = smp_processor_id();
183 struct tick_device *td;
186 * Check, if the current cpu is in the mask
188 if (cpumask_test_cpu(cpu, mask)) {
189 cpumask_clear_cpu(cpu, mask);
190 td = &per_cpu(tick_cpu_device, cpu);
191 td->evtdev->event_handler(td->evtdev);
194 if (!cpumask_empty(mask)) {
196 * It might be necessary to actually check whether the devices
197 * have different broadcast functions. For now, just use the
198 * one of the first device. This works as long as we have this
199 * misfeature only on x86 (lapic)
201 td = &per_cpu(tick_cpu_device, cpumask_first(mask));
202 td->evtdev->broadcast(mask);
207 * Periodic broadcast:
208 * - invoke the broadcast handlers
210 static void tick_do_periodic_broadcast(void)
212 raw_spin_lock(&tick_broadcast_lock);
214 cpumask_and(tmpmask, cpu_online_mask, tick_broadcast_mask);
215 tick_do_broadcast(tmpmask);
217 raw_spin_unlock(&tick_broadcast_lock);
221 * Event handler for periodic broadcast ticks
223 static void tick_handle_periodic_broadcast(struct clock_event_device *dev)
227 tick_do_periodic_broadcast();
230 * The device is in periodic mode. No reprogramming necessary:
232 if (dev->mode == CLOCK_EVT_MODE_PERIODIC)
236 * Setup the next period for devices, which do not have
237 * periodic mode. We read dev->next_event first and add to it
238 * when the event already expired. clockevents_program_event()
239 * sets dev->next_event only when the event is really
240 * programmed to the device.
242 for (next = dev->next_event; ;) {
243 next = ktime_add(next, tick_period);
245 if (!clockevents_program_event(dev, next, false))
247 tick_do_periodic_broadcast();
252 * Powerstate information: The system enters/leaves a state, where
253 * affected devices might stop
255 static void tick_do_broadcast_on_off(unsigned long *reason)
257 struct clock_event_device *bc, *dev;
258 struct tick_device *td;
262 raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
264 cpu = smp_processor_id();
265 td = &per_cpu(tick_cpu_device, cpu);
267 bc = tick_broadcast_device.evtdev;
270 * Is the device not affected by the powerstate ?
272 if (!dev || !(dev->features & CLOCK_EVT_FEAT_C3STOP))
275 if (!tick_device_is_functional(dev))
278 bc_stopped = cpumask_empty(tick_broadcast_mask);
281 case CLOCK_EVT_NOTIFY_BROADCAST_ON:
282 case CLOCK_EVT_NOTIFY_BROADCAST_FORCE:
283 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_mask)) {
284 if (tick_broadcast_device.mode ==
285 TICKDEV_MODE_PERIODIC)
286 clockevents_shutdown(dev);
288 if (*reason == CLOCK_EVT_NOTIFY_BROADCAST_FORCE)
289 tick_broadcast_force = 1;
291 case CLOCK_EVT_NOTIFY_BROADCAST_OFF:
292 if (!tick_broadcast_force &&
293 cpumask_test_and_clear_cpu(cpu, tick_broadcast_mask)) {
294 if (tick_broadcast_device.mode ==
295 TICKDEV_MODE_PERIODIC)
296 tick_setup_periodic(dev, 0);
301 if (cpumask_empty(tick_broadcast_mask)) {
303 clockevents_shutdown(bc);
304 } else if (bc_stopped) {
305 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
306 tick_broadcast_start_periodic(bc);
308 tick_broadcast_setup_oneshot(bc);
311 raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
315 * Powerstate information: The system enters/leaves a state, where
316 * affected devices might stop.
318 void tick_broadcast_on_off(unsigned long reason, int *oncpu)
320 if (!cpumask_test_cpu(*oncpu, cpu_online_mask))
321 printk(KERN_ERR "tick-broadcast: ignoring broadcast for "
322 "offline CPU #%d\n", *oncpu);
324 tick_do_broadcast_on_off(&reason);
328 * Set the periodic handler depending on broadcast on/off
330 void tick_set_periodic_handler(struct clock_event_device *dev, int broadcast)
333 dev->event_handler = tick_handle_periodic;
335 dev->event_handler = tick_handle_periodic_broadcast;
339 * Remove a CPU from broadcasting
341 void tick_shutdown_broadcast(unsigned int *cpup)
343 struct clock_event_device *bc;
345 unsigned int cpu = *cpup;
347 raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
349 bc = tick_broadcast_device.evtdev;
350 cpumask_clear_cpu(cpu, tick_broadcast_mask);
352 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
353 if (bc && cpumask_empty(tick_broadcast_mask))
354 clockevents_shutdown(bc);
357 raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
360 void tick_suspend_broadcast(void)
362 struct clock_event_device *bc;
365 raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
367 bc = tick_broadcast_device.evtdev;
369 clockevents_shutdown(bc);
371 raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
374 int tick_resume_broadcast(void)
376 struct clock_event_device *bc;
380 raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
382 bc = tick_broadcast_device.evtdev;
385 clockevents_set_mode(bc, CLOCK_EVT_MODE_RESUME);
387 switch (tick_broadcast_device.mode) {
388 case TICKDEV_MODE_PERIODIC:
389 if (!cpumask_empty(tick_broadcast_mask))
390 tick_broadcast_start_periodic(bc);
391 broadcast = cpumask_test_cpu(smp_processor_id(),
392 tick_broadcast_mask);
394 case TICKDEV_MODE_ONESHOT:
395 if (!cpumask_empty(tick_broadcast_mask))
396 broadcast = tick_resume_broadcast_oneshot(bc);
400 raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
406 #ifdef CONFIG_TICK_ONESHOT
408 static cpumask_var_t tick_broadcast_oneshot_mask;
409 static cpumask_var_t tick_broadcast_pending_mask;
410 static cpumask_var_t tick_broadcast_force_mask;
413 * Exposed for debugging: see timer_list.c
415 struct cpumask *tick_get_broadcast_oneshot_mask(void)
417 return tick_broadcast_oneshot_mask;
421 * Called before going idle with interrupts disabled. Checks whether a
422 * broadcast event from the other core is about to happen. We detected
423 * that in tick_broadcast_oneshot_control(). The callsite can use this
424 * to avoid a deep idle transition as we are about to get the
425 * broadcast IPI right away.
427 int tick_check_broadcast_expired(void)
429 return cpumask_test_cpu(smp_processor_id(), tick_broadcast_force_mask);
433 * Set broadcast interrupt affinity
435 static void tick_broadcast_set_affinity(struct clock_event_device *bc,
436 const struct cpumask *cpumask)
438 if (!(bc->features & CLOCK_EVT_FEAT_DYNIRQ))
441 if (cpumask_equal(bc->cpumask, cpumask))
444 bc->cpumask = cpumask;
445 irq_set_affinity(bc->irq, bc->cpumask);
448 static int tick_broadcast_set_event(struct clock_event_device *bc, int cpu,
449 ktime_t expires, int force)
453 if (bc->mode != CLOCK_EVT_MODE_ONESHOT)
454 clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
456 ret = clockevents_program_event(bc, expires, force);
458 tick_broadcast_set_affinity(bc, cpumask_of(cpu));
462 int tick_resume_broadcast_oneshot(struct clock_event_device *bc)
464 clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
469 * Called from irq_enter() when idle was interrupted to reenable the
472 void tick_check_oneshot_broadcast(int cpu)
474 if (cpumask_test_cpu(cpu, tick_broadcast_oneshot_mask)) {
475 struct tick_device *td = &per_cpu(tick_cpu_device, cpu);
477 clockevents_set_mode(td->evtdev, CLOCK_EVT_MODE_ONESHOT);
482 * Handle oneshot mode broadcasting
484 static void tick_handle_oneshot_broadcast(struct clock_event_device *dev)
486 struct tick_device *td;
487 ktime_t now, next_event;
488 int cpu, next_cpu = 0;
490 raw_spin_lock(&tick_broadcast_lock);
492 dev->next_event.tv64 = KTIME_MAX;
493 next_event.tv64 = KTIME_MAX;
494 cpumask_clear(tmpmask);
496 /* Find all expired events */
497 for_each_cpu(cpu, tick_broadcast_oneshot_mask) {
498 td = &per_cpu(tick_cpu_device, cpu);
499 if (td->evtdev->next_event.tv64 <= now.tv64) {
500 cpumask_set_cpu(cpu, tmpmask);
502 * Mark the remote cpu in the pending mask, so
503 * it can avoid reprogramming the cpu local
504 * timer in tick_broadcast_oneshot_control().
506 cpumask_set_cpu(cpu, tick_broadcast_pending_mask);
507 } else if (td->evtdev->next_event.tv64 < next_event.tv64) {
508 next_event.tv64 = td->evtdev->next_event.tv64;
514 * Remove the current cpu from the pending mask. The event is
515 * delivered immediately in tick_do_broadcast() !
517 cpumask_clear_cpu(smp_processor_id(), tick_broadcast_pending_mask);
519 /* Take care of enforced broadcast requests */
520 cpumask_or(tmpmask, tmpmask, tick_broadcast_force_mask);
521 cpumask_clear(tick_broadcast_force_mask);
524 * Wakeup the cpus which have an expired event.
526 tick_do_broadcast(tmpmask);
529 * Two reasons for reprogram:
531 * - The global event did not expire any CPU local
532 * events. This happens in dyntick mode, as the maximum PIT
533 * delta is quite small.
535 * - There are pending events on sleeping CPUs which were not
538 if (next_event.tv64 != KTIME_MAX) {
540 * Rearm the broadcast device. If event expired,
543 if (tick_broadcast_set_event(dev, next_cpu, next_event, 0))
546 raw_spin_unlock(&tick_broadcast_lock);
550 * Powerstate information: The system enters/leaves a state, where
551 * affected devices might stop
552 * Returns 0 on success, -EBUSY if the cpu is used to broadcast wakeups.
554 int tick_broadcast_oneshot_control(unsigned long reason)
556 struct clock_event_device *bc, *dev;
557 struct tick_device *td;
563 * Periodic mode does not care about the enter/exit of power
566 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
570 * We are called with preemtion disabled from the depth of the
571 * idle code, so we can't be moved away.
573 cpu = smp_processor_id();
574 td = &per_cpu(tick_cpu_device, cpu);
577 if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
580 bc = tick_broadcast_device.evtdev;
582 raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
583 if (reason == CLOCK_EVT_NOTIFY_BROADCAST_ENTER) {
584 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_oneshot_mask)) {
585 WARN_ON_ONCE(cpumask_test_cpu(cpu, tick_broadcast_pending_mask));
586 clockevents_set_mode(dev, CLOCK_EVT_MODE_SHUTDOWN);
588 * We only reprogram the broadcast timer if we
589 * did not mark ourself in the force mask and
590 * if the cpu local event is earlier than the
591 * broadcast event. If the current CPU is in
592 * the force mask, then we are going to be
593 * woken by the IPI right away.
595 if (!cpumask_test_cpu(cpu, tick_broadcast_force_mask) &&
596 dev->next_event.tv64 < bc->next_event.tv64)
597 tick_broadcast_set_event(bc, cpu, dev->next_event, 1);
600 if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_oneshot_mask)) {
601 clockevents_set_mode(dev, CLOCK_EVT_MODE_ONESHOT);
603 * The cpu which was handling the broadcast
604 * timer marked this cpu in the broadcast
605 * pending mask and fired the broadcast
606 * IPI. So we are going to handle the expired
607 * event anyway via the broadcast IPI
608 * handler. No need to reprogram the timer
609 * with an already expired event.
611 if (cpumask_test_and_clear_cpu(cpu,
612 tick_broadcast_pending_mask))
616 * Bail out if there is no next event.
618 if (dev->next_event.tv64 == KTIME_MAX)
621 * If the pending bit is not set, then we are
622 * either the CPU handling the broadcast
623 * interrupt or we got woken by something else.
625 * We are not longer in the broadcast mask, so
626 * if the cpu local expiry time is already
627 * reached, we would reprogram the cpu local
628 * timer with an already expired event.
630 * This can lead to a ping-pong when we return
631 * to idle and therefor rearm the broadcast
632 * timer before the cpu local timer was able
633 * to fire. This happens because the forced
634 * reprogramming makes sure that the event
635 * will happen in the future and depending on
636 * the min_delta setting this might be far
637 * enough out that the ping-pong starts.
639 * If the cpu local next_event has expired
640 * then we know that the broadcast timer
641 * next_event has expired as well and
642 * broadcast is about to be handled. So we
643 * avoid reprogramming and enforce that the
644 * broadcast handler, which did not run yet,
645 * will invoke the cpu local handler.
647 * We cannot call the handler directly from
648 * here, because we might be in a NOHZ phase
649 * and we did not go through the irq_enter()
653 if (dev->next_event.tv64 <= now.tv64) {
654 cpumask_set_cpu(cpu, tick_broadcast_force_mask);
658 * We got woken by something else. Reprogram
659 * the cpu local timer device.
661 tick_program_event(dev->next_event, 1);
665 raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
670 * Reset the one shot broadcast for a cpu
672 * Called with tick_broadcast_lock held
674 static void tick_broadcast_clear_oneshot(int cpu)
676 cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
679 static void tick_broadcast_init_next_event(struct cpumask *mask,
682 struct tick_device *td;
685 for_each_cpu(cpu, mask) {
686 td = &per_cpu(tick_cpu_device, cpu);
688 td->evtdev->next_event = expires;
693 * tick_broadcast_setup_oneshot - setup the broadcast device
695 void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
697 int cpu = smp_processor_id();
699 /* Set it up only once ! */
700 if (bc->event_handler != tick_handle_oneshot_broadcast) {
701 int was_periodic = bc->mode == CLOCK_EVT_MODE_PERIODIC;
703 bc->event_handler = tick_handle_oneshot_broadcast;
706 * We must be careful here. There might be other CPUs
707 * waiting for periodic broadcast. We need to set the
708 * oneshot_mask bits for those and program the
709 * broadcast device to fire.
711 cpumask_copy(tmpmask, tick_broadcast_mask);
712 cpumask_clear_cpu(cpu, tmpmask);
713 cpumask_or(tick_broadcast_oneshot_mask,
714 tick_broadcast_oneshot_mask, tmpmask);
716 if (was_periodic && !cpumask_empty(tmpmask)) {
717 clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
718 tick_broadcast_init_next_event(tmpmask,
720 tick_broadcast_set_event(bc, cpu, tick_next_period, 1);
722 bc->next_event.tv64 = KTIME_MAX;
725 * The first cpu which switches to oneshot mode sets
726 * the bit for all other cpus which are in the general
727 * (periodic) broadcast mask. So the bit is set and
728 * would prevent the first broadcast enter after this
729 * to program the bc device.
731 tick_broadcast_clear_oneshot(cpu);
736 * Select oneshot operating mode for the broadcast device
738 void tick_broadcast_switch_to_oneshot(void)
740 struct clock_event_device *bc;
743 raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
745 tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
746 bc = tick_broadcast_device.evtdev;
748 tick_broadcast_setup_oneshot(bc);
750 raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
755 * Remove a dead CPU from broadcasting
757 void tick_shutdown_broadcast_oneshot(unsigned int *cpup)
760 unsigned int cpu = *cpup;
762 raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
765 * Clear the broadcast mask flag for the dead cpu, but do not
766 * stop the broadcast device!
768 cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
770 raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
774 * Check, whether the broadcast device is in one shot mode
776 int tick_broadcast_oneshot_active(void)
778 return tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT;
782 * Check whether the broadcast device supports oneshot.
784 bool tick_broadcast_oneshot_available(void)
786 struct clock_event_device *bc = tick_broadcast_device.evtdev;
788 return bc ? bc->features & CLOCK_EVT_FEAT_ONESHOT : false;
793 void __init tick_broadcast_init(void)
795 zalloc_cpumask_var(&tick_broadcast_mask, GFP_NOWAIT);
796 zalloc_cpumask_var(&tmpmask, GFP_NOWAIT);
797 #ifdef CONFIG_TICK_ONESHOT
798 zalloc_cpumask_var(&tick_broadcast_oneshot_mask, GFP_NOWAIT);
799 zalloc_cpumask_var(&tick_broadcast_pending_mask, GFP_NOWAIT);
800 zalloc_cpumask_var(&tick_broadcast_force_mask, GFP_NOWAIT);