1// SPDX-License-Identifier: GPL-2.0
2/*
3 * This file contains functions which emulate a local clock-event
4 * device via a broadcast event source.
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/smp.h>
18#include <linux/module.h>
19
20#include "tick-internal.h"
21
22/*
23 * Broadcast support for broken x86 hardware, where the local apic
24 * timer stops in C3 state.
25 */
26
27static struct tick_device tick_broadcast_device;
28static cpumask_var_t tick_broadcast_mask __cpumask_var_read_mostly;
29static cpumask_var_t tick_broadcast_on __cpumask_var_read_mostly;
30static cpumask_var_t tmpmask __cpumask_var_read_mostly;
31static int tick_broadcast_forced;
32
33static __cacheline_aligned_in_smp DEFINE_RAW_SPINLOCK(tick_broadcast_lock);
34
35#ifdef CONFIG_TICK_ONESHOT
36static DEFINE_PER_CPU(struct clock_event_device *, tick_oneshot_wakeup_device);
37
38static void tick_broadcast_setup_oneshot(struct clock_event_device *bc, bool from_periodic);
39static void tick_broadcast_clear_oneshot(int cpu);
40static void tick_resume_broadcast_oneshot(struct clock_event_device *bc);
41# ifdef CONFIG_HOTPLUG_CPU
42static void tick_broadcast_oneshot_offline(unsigned int cpu);
43# endif
44#else
45static inline void
46tick_broadcast_setup_oneshot(struct clock_event_device *bc, bool from_periodic) { BUG(); }
47static inline void tick_broadcast_clear_oneshot(int cpu) { }
48static inline void tick_resume_broadcast_oneshot(struct clock_event_device *bc) { }
49# ifdef CONFIG_HOTPLUG_CPU
50static inline void tick_broadcast_oneshot_offline(unsigned int cpu) { }
51# endif
52#endif
53
54/*
55 * Debugging: see timer_list.c
56 */
57struct tick_device *tick_get_broadcast_device(void)
58{
59 return &tick_broadcast_device;
60}
61
62struct cpumask *tick_get_broadcast_mask(void)
63{
64 return tick_broadcast_mask;
65}
66
67static struct clock_event_device *tick_get_oneshot_wakeup_device(int cpu);
68
69const struct clock_event_device *tick_get_wakeup_device(int cpu)
70{
71 return tick_get_oneshot_wakeup_device(cpu);
72}
73
74/*
75 * Start the device in periodic mode
76 */
77static void tick_broadcast_start_periodic(struct clock_event_device *bc)
78{
79 if (bc)
80 tick_setup_periodic(dev: bc, broadcast: 1);
81}
82
83/*
84 * Check, if the device can be utilized as broadcast device:
85 */
86static bool tick_check_broadcast_device(struct clock_event_device *curdev,
87 struct clock_event_device *newdev)
88{
89 if ((newdev->features & CLOCK_EVT_FEAT_DUMMY) ||
90 (newdev->features & CLOCK_EVT_FEAT_PERCPU) ||
91 (newdev->features & CLOCK_EVT_FEAT_C3STOP))
92 return false;
93
94 if (tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT &&
95 !(newdev->features & CLOCK_EVT_FEAT_ONESHOT))
96 return false;
97
98 return !curdev || newdev->rating > curdev->rating;
99}
100
101#ifdef CONFIG_TICK_ONESHOT
102static struct clock_event_device *tick_get_oneshot_wakeup_device(int cpu)
103{
104 return per_cpu(tick_oneshot_wakeup_device, cpu);
105}
106
107static void tick_oneshot_wakeup_handler(struct clock_event_device *wd)
108{
109 /*
110 * If we woke up early and the tick was reprogrammed in the
111 * meantime then this may be spurious but harmless.
112 */
113 tick_receive_broadcast();
114}
115
116static bool tick_set_oneshot_wakeup_device(struct clock_event_device *newdev,
117 int cpu)
118{
119 struct clock_event_device *curdev = tick_get_oneshot_wakeup_device(cpu);
120
121 if (!newdev)
122 goto set_device;
123
124 if ((newdev->features & CLOCK_EVT_FEAT_DUMMY) ||
125 (newdev->features & CLOCK_EVT_FEAT_C3STOP))
126 return false;
127
128 if (!(newdev->features & CLOCK_EVT_FEAT_PERCPU) ||
129 !(newdev->features & CLOCK_EVT_FEAT_ONESHOT))
130 return false;
131
132 if (!cpumask_equal(src1p: newdev->cpumask, cpumask_of(cpu)))
133 return false;
134
135 if (curdev && newdev->rating <= curdev->rating)
136 return false;
137
138 if (!try_module_get(module: newdev->owner))
139 return false;
140
141 newdev->event_handler = tick_oneshot_wakeup_handler;
142set_device:
143 clockevents_exchange_device(old: curdev, new: newdev);
144 per_cpu(tick_oneshot_wakeup_device, cpu) = newdev;
145 return true;
146}
147#else
148static struct clock_event_device *tick_get_oneshot_wakeup_device(int cpu)
149{
150 return NULL;
151}
152
153static bool tick_set_oneshot_wakeup_device(struct clock_event_device *newdev,
154 int cpu)
155{
156 return false;
157}
158#endif
159
160/*
161 * Conditionally install/replace broadcast device
162 */
163void tick_install_broadcast_device(struct clock_event_device *dev, int cpu)
164{
165 struct clock_event_device *cur = tick_broadcast_device.evtdev;
166
167 if (tick_set_oneshot_wakeup_device(newdev: dev, cpu))
168 return;
169
170 if (!tick_check_broadcast_device(curdev: cur, newdev: dev))
171 return;
172
173 if (!try_module_get(module: dev->owner))
174 return;
175
176 clockevents_exchange_device(old: cur, new: dev);
177 if (cur)
178 cur->event_handler = clockevents_handle_noop;
179 tick_broadcast_device.evtdev = dev;
180 if (!cpumask_empty(srcp: tick_broadcast_mask))
181 tick_broadcast_start_periodic(bc: dev);
182
183 if (!(dev->features & CLOCK_EVT_FEAT_ONESHOT))
184 return;
185
186 /*
187 * If the system already runs in oneshot mode, switch the newly
188 * registered broadcast device to oneshot mode explicitly.
189 */
190 if (tick_broadcast_oneshot_active()) {
191 tick_broadcast_switch_to_oneshot();
192 return;
193 }
194
195 /*
196 * Inform all cpus about this. We might be in a situation
197 * where we did not switch to oneshot mode because the per cpu
198 * devices are affected by CLOCK_EVT_FEAT_C3STOP and the lack
199 * of a oneshot capable broadcast device. Without that
200 * notification the systems stays stuck in periodic mode
201 * forever.
202 */
203 tick_clock_notify();
204}
205
206/*
207 * Check, if the device is the broadcast device
208 */
209int tick_is_broadcast_device(struct clock_event_device *dev)
210{
211 return (dev && tick_broadcast_device.evtdev == dev);
212}
213
214int tick_broadcast_update_freq(struct clock_event_device *dev, u32 freq)
215{
216 int ret = -ENODEV;
217
218 if (tick_is_broadcast_device(dev)) {
219 raw_spin_lock(&tick_broadcast_lock);
220 ret = __clockevents_update_freq(dev, freq);
221 raw_spin_unlock(&tick_broadcast_lock);
222 }
223 return ret;
224}
225
226
227static void err_broadcast(const struct cpumask *mask)
228{
229 pr_crit_once("Failed to broadcast timer tick. Some CPUs may be unresponsive.\n");
230}
231
232static void tick_device_setup_broadcast_func(struct clock_event_device *dev)
233{
234 if (!dev->broadcast)
235 dev->broadcast = tick_broadcast;
236 if (!dev->broadcast) {
237 pr_warn_once("%s depends on broadcast, but no broadcast function available\n",
238 dev->name);
239 dev->broadcast = err_broadcast;
240 }
241}
242
243/*
244 * Check, if the device is dysfunctional and a placeholder, which
245 * needs to be handled by the broadcast device.
246 */
247int tick_device_uses_broadcast(struct clock_event_device *dev, int cpu)
248{
249 struct clock_event_device *bc = tick_broadcast_device.evtdev;
250 unsigned long flags;
251 int ret = 0;
252
253 raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
254
255 /*
256 * Devices might be registered with both periodic and oneshot
257 * mode disabled. This signals, that the device needs to be
258 * operated from the broadcast device and is a placeholder for
259 * the cpu local device.
260 */
261 if (!tick_device_is_functional(dev)) {
262 dev->event_handler = tick_handle_periodic;
263 tick_device_setup_broadcast_func(dev);
264 cpumask_set_cpu(cpu, dstp: tick_broadcast_mask);
265 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
266 tick_broadcast_start_periodic(bc);
267 else
268 tick_broadcast_setup_oneshot(bc, from_periodic: false);
269 ret = 1;
270 } else {
271 /*
272 * Clear the broadcast bit for this cpu if the
273 * device is not power state affected.
274 */
275 if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
276 cpumask_clear_cpu(cpu, dstp: tick_broadcast_mask);
277 else
278 tick_device_setup_broadcast_func(dev);
279
280 /*
281 * Clear the broadcast bit if the CPU is not in
282 * periodic broadcast on state.
283 */
284 if (!cpumask_test_cpu(cpu, cpumask: tick_broadcast_on))
285 cpumask_clear_cpu(cpu, dstp: tick_broadcast_mask);
286
287 switch (tick_broadcast_device.mode) {
288 case TICKDEV_MODE_ONESHOT:
289 /*
290 * If the system is in oneshot mode we can
291 * unconditionally clear the oneshot mask bit,
292 * because the CPU is running and therefore
293 * not in an idle state which causes the power
294 * state affected device to stop. Let the
295 * caller initialize the device.
296 */
297 tick_broadcast_clear_oneshot(cpu);
298 ret = 0;
299 break;
300
301 case TICKDEV_MODE_PERIODIC:
302 /*
303 * If the system is in periodic mode, check
304 * whether the broadcast device can be
305 * switched off now.
306 */
307 if (cpumask_empty(srcp: tick_broadcast_mask) && bc)
308 clockevents_shutdown(dev: bc);
309 /*
310 * If we kept the cpu in the broadcast mask,
311 * tell the caller to leave the per cpu device
312 * in shutdown state. The periodic interrupt
313 * is delivered by the broadcast device, if
314 * the broadcast device exists and is not
315 * hrtimer based.
316 */
317 if (bc && !(bc->features & CLOCK_EVT_FEAT_HRTIMER))
318 ret = cpumask_test_cpu(cpu, cpumask: tick_broadcast_mask);
319 break;
320 default:
321 break;
322 }
323 }
324 raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
325 return ret;
326}
327
328int tick_receive_broadcast(void)
329{
330 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
331 struct clock_event_device *evt = td->evtdev;
332
333 if (!evt)
334 return -ENODEV;
335
336 if (!evt->event_handler)
337 return -EINVAL;
338
339 evt->event_handler(evt);
340 return 0;
341}
342
343/*
344 * Broadcast the event to the cpus, which are set in the mask (mangled).
345 */
346static bool tick_do_broadcast(struct cpumask *mask)
347{
348 int cpu = smp_processor_id();
349 struct tick_device *td;
350 bool local = false;
351
352 /*
353 * Check, if the current cpu is in the mask
354 */
355 if (cpumask_test_cpu(cpu, cpumask: mask)) {
356 struct clock_event_device *bc = tick_broadcast_device.evtdev;
357
358 cpumask_clear_cpu(cpu, dstp: mask);
359 /*
360 * We only run the local handler, if the broadcast
361 * device is not hrtimer based. Otherwise we run into
362 * a hrtimer recursion.
363 *
364 * local timer_interrupt()
365 * local_handler()
366 * expire_hrtimers()
367 * bc_handler()
368 * local_handler()
369 * expire_hrtimers()
370 */
371 local = !(bc->features & CLOCK_EVT_FEAT_HRTIMER);
372 }
373
374 if (!cpumask_empty(srcp: mask)) {
375 /*
376 * It might be necessary to actually check whether the devices
377 * have different broadcast functions. For now, just use the
378 * one of the first device. This works as long as we have this
379 * misfeature only on x86 (lapic)
380 */
381 td = &per_cpu(tick_cpu_device, cpumask_first(mask));
382 td->evtdev->broadcast(mask);
383 }
384 return local;
385}
386
387/*
388 * Periodic broadcast:
389 * - invoke the broadcast handlers
390 */
391static bool tick_do_periodic_broadcast(void)
392{
393 cpumask_and(dstp: tmpmask, cpu_online_mask, src2p: tick_broadcast_mask);
394 return tick_do_broadcast(mask: tmpmask);
395}
396
397/*
398 * Event handler for periodic broadcast ticks
399 */
400static void tick_handle_periodic_broadcast(struct clock_event_device *dev)
401{
402 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
403 bool bc_local;
404
405 raw_spin_lock(&tick_broadcast_lock);
406
407 /* Handle spurious interrupts gracefully */
408 if (clockevent_state_shutdown(dev: tick_broadcast_device.evtdev)) {
409 raw_spin_unlock(&tick_broadcast_lock);
410 return;
411 }
412
413 bc_local = tick_do_periodic_broadcast();
414
415 if (clockevent_state_oneshot(dev)) {
416 ktime_t next = ktime_add_ns(dev->next_event, TICK_NSEC);
417
418 clockevents_program_event(dev, expires: next, force: true);
419 }
420 raw_spin_unlock(&tick_broadcast_lock);
421
422 /*
423 * We run the handler of the local cpu after dropping
424 * tick_broadcast_lock because the handler might deadlock when
425 * trying to switch to oneshot mode.
426 */
427 if (bc_local)
428 td->evtdev->event_handler(td->evtdev);
429}
430
431/**
432 * tick_broadcast_control - Enable/disable or force broadcast mode
433 * @mode: The selected broadcast mode
434 *
435 * Called when the system enters a state where affected tick devices
436 * might stop. Note: TICK_BROADCAST_FORCE cannot be undone.
437 */
438void tick_broadcast_control(enum tick_broadcast_mode mode)
439{
440 struct clock_event_device *bc, *dev;
441 struct tick_device *td;
442 int cpu, bc_stopped;
443 unsigned long flags;
444
445 /* Protects also the local clockevent device. */
446 raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
447 td = this_cpu_ptr(&tick_cpu_device);
448 dev = td->evtdev;
449
450 /*
451 * Is the device not affected by the powerstate ?
452 */
453 if (!dev || !(dev->features & CLOCK_EVT_FEAT_C3STOP))
454 goto out;
455
456 if (!tick_device_is_functional(dev))
457 goto out;
458
459 cpu = smp_processor_id();
460 bc = tick_broadcast_device.evtdev;
461 bc_stopped = cpumask_empty(srcp: tick_broadcast_mask);
462
463 switch (mode) {
464 case TICK_BROADCAST_FORCE:
465 tick_broadcast_forced = 1;
466 fallthrough;
467 case TICK_BROADCAST_ON:
468 cpumask_set_cpu(cpu, dstp: tick_broadcast_on);
469 if (!cpumask_test_and_set_cpu(cpu, cpumask: tick_broadcast_mask)) {
470 /*
471 * Only shutdown the cpu local device, if:
472 *
473 * - the broadcast device exists
474 * - the broadcast device is not a hrtimer based one
475 * - the broadcast device is in periodic mode to
476 * avoid a hiccup during switch to oneshot mode
477 */
478 if (bc && !(bc->features & CLOCK_EVT_FEAT_HRTIMER) &&
479 tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
480 clockevents_shutdown(dev);
481 }
482 break;
483
484 case TICK_BROADCAST_OFF:
485 if (tick_broadcast_forced)
486 break;
487 cpumask_clear_cpu(cpu, dstp: tick_broadcast_on);
488 if (cpumask_test_and_clear_cpu(cpu, cpumask: tick_broadcast_mask)) {
489 if (tick_broadcast_device.mode ==
490 TICKDEV_MODE_PERIODIC)
491 tick_setup_periodic(dev, broadcast: 0);
492 }
493 break;
494 }
495
496 if (bc) {
497 if (cpumask_empty(srcp: tick_broadcast_mask)) {
498 if (!bc_stopped)
499 clockevents_shutdown(dev: bc);
500 } else if (bc_stopped) {
501 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
502 tick_broadcast_start_periodic(bc);
503 else
504 tick_broadcast_setup_oneshot(bc, from_periodic: false);
505 }
506 }
507out:
508 raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
509}
510EXPORT_SYMBOL_GPL(tick_broadcast_control);
511
512/*
513 * Set the periodic handler depending on broadcast on/off
514 */
515void tick_set_periodic_handler(struct clock_event_device *dev, int broadcast)
516{
517 if (!broadcast)
518 dev->event_handler = tick_handle_periodic;
519 else
520 dev->event_handler = tick_handle_periodic_broadcast;
521}
522
523#ifdef CONFIG_HOTPLUG_CPU
524static void tick_shutdown_broadcast(void)
525{
526 struct clock_event_device *bc = tick_broadcast_device.evtdev;
527
528 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
529 if (bc && cpumask_empty(srcp: tick_broadcast_mask))
530 clockevents_shutdown(dev: bc);
531 }
532}
533
534/*
535 * Remove a CPU from broadcasting
536 */
537void tick_broadcast_offline(unsigned int cpu)
538{
539 raw_spin_lock(&tick_broadcast_lock);
540 cpumask_clear_cpu(cpu, dstp: tick_broadcast_mask);
541 cpumask_clear_cpu(cpu, dstp: tick_broadcast_on);
542 tick_broadcast_oneshot_offline(cpu);
543 tick_shutdown_broadcast();
544 raw_spin_unlock(&tick_broadcast_lock);
545}
546
547#endif
548
549void tick_suspend_broadcast(void)
550{
551 struct clock_event_device *bc;
552 unsigned long flags;
553
554 raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
555
556 bc = tick_broadcast_device.evtdev;
557 if (bc)
558 clockevents_shutdown(dev: bc);
559
560 raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
561}
562
563/*
564 * This is called from tick_resume_local() on a resuming CPU. That's
565 * called from the core resume function, tick_unfreeze() and the magic XEN
566 * resume hackery.
567 *
568 * In none of these cases the broadcast device mode can change and the
569 * bit of the resuming CPU in the broadcast mask is safe as well.
570 */
571bool tick_resume_check_broadcast(void)
572{
573 if (tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT)
574 return false;
575 else
576 return cpumask_test_cpu(smp_processor_id(), cpumask: tick_broadcast_mask);
577}
578
579void tick_resume_broadcast(void)
580{
581 struct clock_event_device *bc;
582 unsigned long flags;
583
584 raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
585
586 bc = tick_broadcast_device.evtdev;
587
588 if (bc) {
589 clockevents_tick_resume(dev: bc);
590
591 switch (tick_broadcast_device.mode) {
592 case TICKDEV_MODE_PERIODIC:
593 if (!cpumask_empty(srcp: tick_broadcast_mask))
594 tick_broadcast_start_periodic(bc);
595 break;
596 case TICKDEV_MODE_ONESHOT:
597 if (!cpumask_empty(srcp: tick_broadcast_mask))
598 tick_resume_broadcast_oneshot(bc);
599 break;
600 }
601 }
602 raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
603}
604
605#ifdef CONFIG_TICK_ONESHOT
606
607static cpumask_var_t tick_broadcast_oneshot_mask __cpumask_var_read_mostly;
608static cpumask_var_t tick_broadcast_pending_mask __cpumask_var_read_mostly;
609static cpumask_var_t tick_broadcast_force_mask __cpumask_var_read_mostly;
610
611/*
612 * Exposed for debugging: see timer_list.c
613 */
614struct cpumask *tick_get_broadcast_oneshot_mask(void)
615{
616 return tick_broadcast_oneshot_mask;
617}
618
619/*
620 * Called before going idle with interrupts disabled. Checks whether a
621 * broadcast event from the other core is about to happen. We detected
622 * that in tick_broadcast_oneshot_control(). The callsite can use this
623 * to avoid a deep idle transition as we are about to get the
624 * broadcast IPI right away.
625 */
626noinstr int tick_check_broadcast_expired(void)
627{
628#ifdef _ASM_GENERIC_BITOPS_INSTRUMENTED_NON_ATOMIC_H
629 return arch_test_bit(smp_processor_id(), cpumask_bits(tick_broadcast_force_mask));
630#else
631 return cpumask_test_cpu(smp_processor_id(), tick_broadcast_force_mask);
632#endif
633}
634
635/*
636 * Set broadcast interrupt affinity
637 */
638static void tick_broadcast_set_affinity(struct clock_event_device *bc,
639 const struct cpumask *cpumask)
640{
641 if (!(bc->features & CLOCK_EVT_FEAT_DYNIRQ))
642 return;
643
644 if (cpumask_equal(src1p: bc->cpumask, src2p: cpumask))
645 return;
646
647 bc->cpumask = cpumask;
648 irq_set_affinity(irq: bc->irq, cpumask: bc->cpumask);
649}
650
651static void tick_broadcast_set_event(struct clock_event_device *bc, int cpu,
652 ktime_t expires)
653{
654 if (!clockevent_state_oneshot(dev: bc))
655 clockevents_switch_state(dev: bc, state: CLOCK_EVT_STATE_ONESHOT);
656
657 clockevents_program_event(dev: bc, expires, force: 1);
658 tick_broadcast_set_affinity(bc, cpumask_of(cpu));
659}
660
661static void tick_resume_broadcast_oneshot(struct clock_event_device *bc)
662{
663 clockevents_switch_state(dev: bc, state: CLOCK_EVT_STATE_ONESHOT);
664}
665
666/*
667 * Called from irq_enter() when idle was interrupted to reenable the
668 * per cpu device.
669 */
670void tick_check_oneshot_broadcast_this_cpu(void)
671{
672 if (cpumask_test_cpu(smp_processor_id(), cpumask: tick_broadcast_oneshot_mask)) {
673 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
674
675 /*
676 * We might be in the middle of switching over from
677 * periodic to oneshot. If the CPU has not yet
678 * switched over, leave the device alone.
679 */
680 if (td->mode == TICKDEV_MODE_ONESHOT) {
681 clockevents_switch_state(dev: td->evtdev,
682 state: CLOCK_EVT_STATE_ONESHOT);
683 }
684 }
685}
686
687/*
688 * Handle oneshot mode broadcasting
689 */
690static void tick_handle_oneshot_broadcast(struct clock_event_device *dev)
691{
692 struct tick_device *td;
693 ktime_t now, next_event;
694 int cpu, next_cpu = 0;
695 bool bc_local;
696
697 raw_spin_lock(&tick_broadcast_lock);
698 dev->next_event = KTIME_MAX;
699 next_event = KTIME_MAX;
700 cpumask_clear(dstp: tmpmask);
701 now = ktime_get();
702 /* Find all expired events */
703 for_each_cpu(cpu, tick_broadcast_oneshot_mask) {
704 /*
705 * Required for !SMP because for_each_cpu() reports
706 * unconditionally CPU0 as set on UP kernels.
707 */
708 if (!IS_ENABLED(CONFIG_SMP) &&
709 cpumask_empty(srcp: tick_broadcast_oneshot_mask))
710 break;
711
712 td = &per_cpu(tick_cpu_device, cpu);
713 if (td->evtdev->next_event <= now) {
714 cpumask_set_cpu(cpu, dstp: tmpmask);
715 /*
716 * Mark the remote cpu in the pending mask, so
717 * it can avoid reprogramming the cpu local
718 * timer in tick_broadcast_oneshot_control().
719 */
720 cpumask_set_cpu(cpu, dstp: tick_broadcast_pending_mask);
721 } else if (td->evtdev->next_event < next_event) {
722 next_event = td->evtdev->next_event;
723 next_cpu = cpu;
724 }
725 }
726
727 /*
728 * Remove the current cpu from the pending mask. The event is
729 * delivered immediately in tick_do_broadcast() !
730 */
731 cpumask_clear_cpu(smp_processor_id(), dstp: tick_broadcast_pending_mask);
732
733 /* Take care of enforced broadcast requests */
734 cpumask_or(dstp: tmpmask, src1p: tmpmask, src2p: tick_broadcast_force_mask);
735 cpumask_clear(dstp: tick_broadcast_force_mask);
736
737 /*
738 * Sanity check. Catch the case where we try to broadcast to
739 * offline cpus.
740 */
741 if (WARN_ON_ONCE(!cpumask_subset(tmpmask, cpu_online_mask)))
742 cpumask_and(dstp: tmpmask, src1p: tmpmask, cpu_online_mask);
743
744 /*
745 * Wakeup the cpus which have an expired event.
746 */
747 bc_local = tick_do_broadcast(mask: tmpmask);
748
749 /*
750 * Two reasons for reprogram:
751 *
752 * - The global event did not expire any CPU local
753 * events. This happens in dyntick mode, as the maximum PIT
754 * delta is quite small.
755 *
756 * - There are pending events on sleeping CPUs which were not
757 * in the event mask
758 */
759 if (next_event != KTIME_MAX)
760 tick_broadcast_set_event(bc: dev, cpu: next_cpu, expires: next_event);
761
762 raw_spin_unlock(&tick_broadcast_lock);
763
764 if (bc_local) {
765 td = this_cpu_ptr(&tick_cpu_device);
766 td->evtdev->event_handler(td->evtdev);
767 }
768}
769
770static int broadcast_needs_cpu(struct clock_event_device *bc, int cpu)
771{
772 if (!(bc->features & CLOCK_EVT_FEAT_HRTIMER))
773 return 0;
774 if (bc->next_event == KTIME_MAX)
775 return 0;
776 return bc->bound_on == cpu ? -EBUSY : 0;
777}
778
779static void broadcast_shutdown_local(struct clock_event_device *bc,
780 struct clock_event_device *dev)
781{
782 /*
783 * For hrtimer based broadcasting we cannot shutdown the cpu
784 * local device if our own event is the first one to expire or
785 * if we own the broadcast timer.
786 */
787 if (bc->features & CLOCK_EVT_FEAT_HRTIMER) {
788 if (broadcast_needs_cpu(bc, smp_processor_id()))
789 return;
790 if (dev->next_event < bc->next_event)
791 return;
792 }
793 clockevents_switch_state(dev, state: CLOCK_EVT_STATE_SHUTDOWN);
794}
795
796static int ___tick_broadcast_oneshot_control(enum tick_broadcast_state state,
797 struct tick_device *td,
798 int cpu)
799{
800 struct clock_event_device *bc, *dev = td->evtdev;
801 int ret = 0;
802 ktime_t now;
803
804 raw_spin_lock(&tick_broadcast_lock);
805 bc = tick_broadcast_device.evtdev;
806
807 if (state == TICK_BROADCAST_ENTER) {
808 /*
809 * If the current CPU owns the hrtimer broadcast
810 * mechanism, it cannot go deep idle and we do not add
811 * the CPU to the broadcast mask. We don't have to go
812 * through the EXIT path as the local timer is not
813 * shutdown.
814 */
815 ret = broadcast_needs_cpu(bc, cpu);
816 if (ret)
817 goto out;
818
819 /*
820 * If the broadcast device is in periodic mode, we
821 * return.
822 */
823 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
824 /* If it is a hrtimer based broadcast, return busy */
825 if (bc->features & CLOCK_EVT_FEAT_HRTIMER)
826 ret = -EBUSY;
827 goto out;
828 }
829
830 if (!cpumask_test_and_set_cpu(cpu, cpumask: tick_broadcast_oneshot_mask)) {
831 WARN_ON_ONCE(cpumask_test_cpu(cpu, tick_broadcast_pending_mask));
832
833 /* Conditionally shut down the local timer. */
834 broadcast_shutdown_local(bc, dev);
835
836 /*
837 * We only reprogram the broadcast timer if we
838 * did not mark ourself in the force mask and
839 * if the cpu local event is earlier than the
840 * broadcast event. If the current CPU is in
841 * the force mask, then we are going to be
842 * woken by the IPI right away; we return
843 * busy, so the CPU does not try to go deep
844 * idle.
845 */
846 if (cpumask_test_cpu(cpu, cpumask: tick_broadcast_force_mask)) {
847 ret = -EBUSY;
848 } else if (dev->next_event < bc->next_event) {
849 tick_broadcast_set_event(bc, cpu, expires: dev->next_event);
850 /*
851 * In case of hrtimer broadcasts the
852 * programming might have moved the
853 * timer to this cpu. If yes, remove
854 * us from the broadcast mask and
855 * return busy.
856 */
857 ret = broadcast_needs_cpu(bc, cpu);
858 if (ret) {
859 cpumask_clear_cpu(cpu,
860 dstp: tick_broadcast_oneshot_mask);
861 }
862 }
863 }
864 } else {
865 if (cpumask_test_and_clear_cpu(cpu, cpumask: tick_broadcast_oneshot_mask)) {
866 clockevents_switch_state(dev, state: CLOCK_EVT_STATE_ONESHOT);
867 /*
868 * The cpu which was handling the broadcast
869 * timer marked this cpu in the broadcast
870 * pending mask and fired the broadcast
871 * IPI. So we are going to handle the expired
872 * event anyway via the broadcast IPI
873 * handler. No need to reprogram the timer
874 * with an already expired event.
875 */
876 if (cpumask_test_and_clear_cpu(cpu,
877 cpumask: tick_broadcast_pending_mask))
878 goto out;
879
880 /*
881 * Bail out if there is no next event.
882 */
883 if (dev->next_event == KTIME_MAX)
884 goto out;
885 /*
886 * If the pending bit is not set, then we are
887 * either the CPU handling the broadcast
888 * interrupt or we got woken by something else.
889 *
890 * We are no longer in the broadcast mask, so
891 * if the cpu local expiry time is already
892 * reached, we would reprogram the cpu local
893 * timer with an already expired event.
894 *
895 * This can lead to a ping-pong when we return
896 * to idle and therefore rearm the broadcast
897 * timer before the cpu local timer was able
898 * to fire. This happens because the forced
899 * reprogramming makes sure that the event
900 * will happen in the future and depending on
901 * the min_delta setting this might be far
902 * enough out that the ping-pong starts.
903 *
904 * If the cpu local next_event has expired
905 * then we know that the broadcast timer
906 * next_event has expired as well and
907 * broadcast is about to be handled. So we
908 * avoid reprogramming and enforce that the
909 * broadcast handler, which did not run yet,
910 * will invoke the cpu local handler.
911 *
912 * We cannot call the handler directly from
913 * here, because we might be in a NOHZ phase
914 * and we did not go through the irq_enter()
915 * nohz fixups.
916 */
917 now = ktime_get();
918 if (dev->next_event <= now) {
919 cpumask_set_cpu(cpu, dstp: tick_broadcast_force_mask);
920 goto out;
921 }
922 /*
923 * We got woken by something else. Reprogram
924 * the cpu local timer device.
925 */
926 tick_program_event(expires: dev->next_event, force: 1);
927 }
928 }
929out:
930 raw_spin_unlock(&tick_broadcast_lock);
931 return ret;
932}
933
934static int tick_oneshot_wakeup_control(enum tick_broadcast_state state,
935 struct tick_device *td,
936 int cpu)
937{
938 struct clock_event_device *dev, *wd;
939
940 dev = td->evtdev;
941 if (td->mode != TICKDEV_MODE_ONESHOT)
942 return -EINVAL;
943
944 wd = tick_get_oneshot_wakeup_device(cpu);
945 if (!wd)
946 return -ENODEV;
947
948 switch (state) {
949 case TICK_BROADCAST_ENTER:
950 clockevents_switch_state(dev, state: CLOCK_EVT_STATE_ONESHOT_STOPPED);
951 clockevents_switch_state(dev: wd, state: CLOCK_EVT_STATE_ONESHOT);
952 clockevents_program_event(dev: wd, expires: dev->next_event, force: 1);
953 break;
954 case TICK_BROADCAST_EXIT:
955 /* We may have transitioned to oneshot mode while idle */
956 if (clockevent_get_state(dev: wd) != CLOCK_EVT_STATE_ONESHOT)
957 return -ENODEV;
958 }
959
960 return 0;
961}
962
963int __tick_broadcast_oneshot_control(enum tick_broadcast_state state)
964{
965 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
966 int cpu = smp_processor_id();
967
968 if (!tick_oneshot_wakeup_control(state, td, cpu))
969 return 0;
970
971 if (tick_broadcast_device.evtdev)
972 return ___tick_broadcast_oneshot_control(state, td, cpu);
973
974 /*
975 * If there is no broadcast or wakeup device, tell the caller not
976 * to go into deep idle.
977 */
978 return -EBUSY;
979}
980
981/*
982 * Reset the one shot broadcast for a cpu
983 *
984 * Called with tick_broadcast_lock held
985 */
986static void tick_broadcast_clear_oneshot(int cpu)
987{
988 cpumask_clear_cpu(cpu, dstp: tick_broadcast_oneshot_mask);
989 cpumask_clear_cpu(cpu, dstp: tick_broadcast_pending_mask);
990}
991
992static void tick_broadcast_init_next_event(struct cpumask *mask,
993 ktime_t expires)
994{
995 struct tick_device *td;
996 int cpu;
997
998 for_each_cpu(cpu, mask) {
999 td = &per_cpu(tick_cpu_device, cpu);
1000 if (td->evtdev)
1001 td->evtdev->next_event = expires;
1002 }
1003}
1004
1005static inline ktime_t tick_get_next_period(void)
1006{
1007 ktime_t next;
1008
1009 /*
1010 * Protect against concurrent updates (store /load tearing on
1011 * 32bit). It does not matter if the time is already in the
1012 * past. The broadcast device which is about to be programmed will
1013 * fire in any case.
1014 */
1015 raw_spin_lock(&jiffies_lock);
1016 next = tick_next_period;
1017 raw_spin_unlock(&jiffies_lock);
1018 return next;
1019}
1020
1021/**
1022 * tick_broadcast_setup_oneshot - setup the broadcast device
1023 * @bc: the broadcast device
1024 * @from_periodic: true if called from periodic mode
1025 */
1026static void tick_broadcast_setup_oneshot(struct clock_event_device *bc,
1027 bool from_periodic)
1028{
1029 int cpu = smp_processor_id();
1030 ktime_t nexttick = 0;
1031
1032 if (!bc)
1033 return;
1034
1035 /*
1036 * When the broadcast device was switched to oneshot by the first
1037 * CPU handling the NOHZ change, the other CPUs will reach this
1038 * code via hrtimer_run_queues() -> tick_check_oneshot_change()
1039 * too. Set up the broadcast device only once!
1040 */
1041 if (bc->event_handler == tick_handle_oneshot_broadcast) {
1042 /*
1043 * The CPU which switched from periodic to oneshot mode
1044 * set the broadcast oneshot bit for all other CPUs which
1045 * are in the general (periodic) broadcast mask to ensure
1046 * that CPUs which wait for the periodic broadcast are
1047 * woken up.
1048 *
1049 * Clear the bit for the local CPU as the set bit would
1050 * prevent the first tick_broadcast_enter() after this CPU
1051 * switched to oneshot state to program the broadcast
1052 * device.
1053 *
1054 * This code can also be reached via tick_broadcast_control(),
1055 * but this cannot avoid the tick_broadcast_clear_oneshot()
1056 * as that would break the periodic to oneshot transition of
1057 * secondary CPUs. But that's harmless as the below only
1058 * clears already cleared bits.
1059 */
1060 tick_broadcast_clear_oneshot(cpu);
1061 return;
1062 }
1063
1064
1065 bc->event_handler = tick_handle_oneshot_broadcast;
1066 bc->next_event = KTIME_MAX;
1067
1068 /*
1069 * When the tick mode is switched from periodic to oneshot it must
1070 * be ensured that CPUs which are waiting for periodic broadcast
1071 * get their wake-up at the next tick. This is achieved by ORing
1072 * tick_broadcast_mask into tick_broadcast_oneshot_mask.
1073 *
1074 * For other callers, e.g. broadcast device replacement,
1075 * tick_broadcast_oneshot_mask must not be touched as this would
1076 * set bits for CPUs which are already NOHZ, but not idle. Their
1077 * next tick_broadcast_enter() would observe the bit set and fail
1078 * to update the expiry time and the broadcast event device.
1079 */
1080 if (from_periodic) {
1081 cpumask_copy(dstp: tmpmask, srcp: tick_broadcast_mask);
1082 /* Remove the local CPU as it is obviously not idle */
1083 cpumask_clear_cpu(cpu, dstp: tmpmask);
1084 cpumask_or(dstp: tick_broadcast_oneshot_mask, src1p: tick_broadcast_oneshot_mask, src2p: tmpmask);
1085
1086 /*
1087 * Ensure that the oneshot broadcast handler will wake the
1088 * CPUs which are still waiting for periodic broadcast.
1089 */
1090 nexttick = tick_get_next_period();
1091 tick_broadcast_init_next_event(mask: tmpmask, expires: nexttick);
1092
1093 /*
1094 * If the underlying broadcast clock event device is
1095 * already in oneshot state, then there is nothing to do.
1096 * The device was already armed for the next tick
1097 * in tick_handle_broadcast_periodic()
1098 */
1099 if (clockevent_state_oneshot(dev: bc))
1100 return;
1101 }
1102
1103 /*
1104 * When switching from periodic to oneshot mode arm the broadcast
1105 * device for the next tick.
1106 *
1107 * If the broadcast device has been replaced in oneshot mode and
1108 * the oneshot broadcast mask is not empty, then arm it to expire
1109 * immediately in order to reevaluate the next expiring timer.
1110 * @nexttick is 0 and therefore in the past which will cause the
1111 * clockevent code to force an event.
1112 *
1113 * For both cases the programming can be avoided when the oneshot
1114 * broadcast mask is empty.
1115 *
1116 * tick_broadcast_set_event() implicitly switches the broadcast
1117 * device to oneshot state.
1118 */
1119 if (!cpumask_empty(srcp: tick_broadcast_oneshot_mask))
1120 tick_broadcast_set_event(bc, cpu, expires: nexttick);
1121}
1122
1123/*
1124 * Select oneshot operating mode for the broadcast device
1125 */
1126void tick_broadcast_switch_to_oneshot(void)
1127{
1128 struct clock_event_device *bc;
1129 enum tick_device_mode oldmode;
1130 unsigned long flags;
1131
1132 raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
1133
1134 oldmode = tick_broadcast_device.mode;
1135 tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
1136 bc = tick_broadcast_device.evtdev;
1137 if (bc)
1138 tick_broadcast_setup_oneshot(bc, from_periodic: oldmode == TICKDEV_MODE_PERIODIC);
1139
1140 raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
1141}
1142
1143#ifdef CONFIG_HOTPLUG_CPU
1144void hotplug_cpu__broadcast_tick_pull(int deadcpu)
1145{
1146 struct clock_event_device *bc;
1147 unsigned long flags;
1148
1149 raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
1150 bc = tick_broadcast_device.evtdev;
1151
1152 if (bc && broadcast_needs_cpu(bc, cpu: deadcpu)) {
1153 /*
1154 * If the broadcast force bit of the current CPU is set,
1155 * then the current CPU has not yet reprogrammed the local
1156 * timer device to avoid a ping-pong race. See
1157 * ___tick_broadcast_oneshot_control().
1158 *
1159 * If the broadcast device is hrtimer based then
1160 * programming the broadcast event below does not have any
1161 * effect because the local clockevent device is not
1162 * running and not programmed because the broadcast event
1163 * is not earlier than the pending event of the local clock
1164 * event device. As a consequence all CPUs waiting for a
1165 * broadcast event are stuck forever.
1166 *
1167 * Detect this condition and reprogram the cpu local timer
1168 * device to avoid the starvation.
1169 */
1170 if (tick_check_broadcast_expired()) {
1171 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
1172
1173 cpumask_clear_cpu(smp_processor_id(), dstp: tick_broadcast_force_mask);
1174 tick_program_event(expires: td->evtdev->next_event, force: 1);
1175 }
1176
1177 /* This moves the broadcast assignment to this CPU: */
1178 clockevents_program_event(dev: bc, expires: bc->next_event, force: 1);
1179 }
1180 raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
1181}
1182
1183/*
1184 * Remove a dying CPU from broadcasting
1185 */
1186static void tick_broadcast_oneshot_offline(unsigned int cpu)
1187{
1188 if (tick_get_oneshot_wakeup_device(cpu))
1189 tick_set_oneshot_wakeup_device(NULL, cpu);
1190
1191 /*
1192 * Clear the broadcast masks for the dead cpu, but do not stop
1193 * the broadcast device!
1194 */
1195 cpumask_clear_cpu(cpu, dstp: tick_broadcast_oneshot_mask);
1196 cpumask_clear_cpu(cpu, dstp: tick_broadcast_pending_mask);
1197 cpumask_clear_cpu(cpu, dstp: tick_broadcast_force_mask);
1198}
1199#endif
1200
1201/*
1202 * Check, whether the broadcast device is in one shot mode
1203 */
1204int tick_broadcast_oneshot_active(void)
1205{
1206 return tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT;
1207}
1208
1209/*
1210 * Check whether the broadcast device supports oneshot.
1211 */
1212bool tick_broadcast_oneshot_available(void)
1213{
1214 struct clock_event_device *bc = tick_broadcast_device.evtdev;
1215
1216 return bc ? bc->features & CLOCK_EVT_FEAT_ONESHOT : false;
1217}
1218
1219#else
1220int __tick_broadcast_oneshot_control(enum tick_broadcast_state state)
1221{
1222 struct clock_event_device *bc = tick_broadcast_device.evtdev;
1223
1224 if (!bc || (bc->features & CLOCK_EVT_FEAT_HRTIMER))
1225 return -EBUSY;
1226
1227 return 0;
1228}
1229#endif
1230
1231void __init tick_broadcast_init(void)
1232{
1233 zalloc_cpumask_var(mask: &tick_broadcast_mask, GFP_NOWAIT);
1234 zalloc_cpumask_var(mask: &tick_broadcast_on, GFP_NOWAIT);
1235 zalloc_cpumask_var(mask: &tmpmask, GFP_NOWAIT);
1236#ifdef CONFIG_TICK_ONESHOT
1237 zalloc_cpumask_var(mask: &tick_broadcast_oneshot_mask, GFP_NOWAIT);
1238 zalloc_cpumask_var(mask: &tick_broadcast_pending_mask, GFP_NOWAIT);
1239 zalloc_cpumask_var(mask: &tick_broadcast_force_mask, GFP_NOWAIT);
1240#endif
1241}
1242