mm/page_alloc: prevent merging between isolated and other pageblocks
[linux-2.6-block.git] / kernel / time / tick-broadcast.c
CommitLineData
f8381cba
TG
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>
d7b90689 17#include <linux/interrupt.h>
f8381cba
TG
18#include <linux/percpu.h>
19#include <linux/profile.h>
20#include <linux/sched.h>
12ad1000 21#include <linux/smp.h>
ccf33d68 22#include <linux/module.h>
f8381cba
TG
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
a52f5c56 31static struct tick_device tick_broadcast_device;
b352bc1c 32static cpumask_var_t tick_broadcast_mask;
07bd1172 33static cpumask_var_t tick_broadcast_on;
b352bc1c 34static cpumask_var_t tmpmask;
b5f91da0 35static DEFINE_RAW_SPINLOCK(tick_broadcast_lock);
592a438f 36static int tick_broadcast_forced;
f8381cba 37
5590a536
TG
38#ifdef CONFIG_TICK_ONESHOT
39static void tick_broadcast_clear_oneshot(int cpu);
080873ce 40static void tick_resume_broadcast_oneshot(struct clock_event_device *bc);
5590a536
TG
41#else
42static inline void tick_broadcast_clear_oneshot(int cpu) { }
080873ce 43static inline void tick_resume_broadcast_oneshot(struct clock_event_device *bc) { }
5590a536
TG
44#endif
45
289f480a
IM
46/*
47 * Debugging: see timer_list.c
48 */
49struct tick_device *tick_get_broadcast_device(void)
50{
51 return &tick_broadcast_device;
52}
53
6b954823 54struct cpumask *tick_get_broadcast_mask(void)
289f480a 55{
b352bc1c 56 return tick_broadcast_mask;
289f480a
IM
57}
58
f8381cba
TG
59/*
60 * Start the device in periodic mode
61 */
62static void tick_broadcast_start_periodic(struct clock_event_device *bc)
63{
18de5bc4 64 if (bc)
f8381cba
TG
65 tick_setup_periodic(bc, 1);
66}
67
68/*
69 * Check, if the device can be utilized as broadcast device:
70 */
45cb8e01
TG
71static bool tick_check_broadcast_device(struct clock_event_device *curdev,
72 struct clock_event_device *newdev)
73{
74 if ((newdev->features & CLOCK_EVT_FEAT_DUMMY) ||
245a3496 75 (newdev->features & CLOCK_EVT_FEAT_PERCPU) ||
45cb8e01
TG
76 (newdev->features & CLOCK_EVT_FEAT_C3STOP))
77 return false;
78
79 if (tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT &&
80 !(newdev->features & CLOCK_EVT_FEAT_ONESHOT))
81 return false;
82
83 return !curdev || newdev->rating > curdev->rating;
84}
85
86/*
87 * Conditionally install/replace broadcast device
88 */
7172a286 89void tick_install_broadcast_device(struct clock_event_device *dev)
f8381cba 90{
6f7a05d7
TG
91 struct clock_event_device *cur = tick_broadcast_device.evtdev;
92
45cb8e01 93 if (!tick_check_broadcast_device(cur, dev))
7172a286 94 return;
45cb8e01 95
ccf33d68
TG
96 if (!try_module_get(dev->owner))
97 return;
f8381cba 98
45cb8e01 99 clockevents_exchange_device(cur, dev);
6f7a05d7
TG
100 if (cur)
101 cur->event_handler = clockevents_handle_noop;
f8381cba 102 tick_broadcast_device.evtdev = dev;
b352bc1c 103 if (!cpumask_empty(tick_broadcast_mask))
f8381cba 104 tick_broadcast_start_periodic(dev);
c038c1c4
SB
105 /*
106 * Inform all cpus about this. We might be in a situation
107 * where we did not switch to oneshot mode because the per cpu
108 * devices are affected by CLOCK_EVT_FEAT_C3STOP and the lack
109 * of a oneshot capable broadcast device. Without that
110 * notification the systems stays stuck in periodic mode
111 * forever.
112 */
113 if (dev->features & CLOCK_EVT_FEAT_ONESHOT)
114 tick_clock_notify();
f8381cba
TG
115}
116
117/*
118 * Check, if the device is the broadcast device
119 */
120int tick_is_broadcast_device(struct clock_event_device *dev)
121{
122 return (dev && tick_broadcast_device.evtdev == dev);
123}
124
627ee794
TG
125int tick_broadcast_update_freq(struct clock_event_device *dev, u32 freq)
126{
127 int ret = -ENODEV;
128
129 if (tick_is_broadcast_device(dev)) {
130 raw_spin_lock(&tick_broadcast_lock);
131 ret = __clockevents_update_freq(dev, freq);
132 raw_spin_unlock(&tick_broadcast_lock);
133 }
134 return ret;
135}
136
137
12ad1000
MR
138static void err_broadcast(const struct cpumask *mask)
139{
140 pr_crit_once("Failed to broadcast timer tick. Some CPUs may be unresponsive.\n");
141}
142
5d1d9a29
MR
143static void tick_device_setup_broadcast_func(struct clock_event_device *dev)
144{
145 if (!dev->broadcast)
146 dev->broadcast = tick_broadcast;
147 if (!dev->broadcast) {
148 pr_warn_once("%s depends on broadcast, but no broadcast function available\n",
149 dev->name);
150 dev->broadcast = err_broadcast;
151 }
152}
153
f8381cba
TG
154/*
155 * Check, if the device is disfunctional and a place holder, which
156 * needs to be handled by the broadcast device.
157 */
158int tick_device_uses_broadcast(struct clock_event_device *dev, int cpu)
159{
07bd1172 160 struct clock_event_device *bc = tick_broadcast_device.evtdev;
f8381cba 161 unsigned long flags;
e0454311 162 int ret = 0;
f8381cba 163
b5f91da0 164 raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
f8381cba
TG
165
166 /*
167 * Devices might be registered with both periodic and oneshot
168 * mode disabled. This signals, that the device needs to be
169 * operated from the broadcast device and is a placeholder for
170 * the cpu local device.
171 */
172 if (!tick_device_is_functional(dev)) {
173 dev->event_handler = tick_handle_periodic;
5d1d9a29 174 tick_device_setup_broadcast_func(dev);
b352bc1c 175 cpumask_set_cpu(cpu, tick_broadcast_mask);
a272dcca
SB
176 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
177 tick_broadcast_start_periodic(bc);
178 else
179 tick_broadcast_setup_oneshot(bc);
f8381cba 180 ret = 1;
5590a536
TG
181 } else {
182 /*
07bd1172
TG
183 * Clear the broadcast bit for this cpu if the
184 * device is not power state affected.
5590a536 185 */
07bd1172 186 if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
b352bc1c 187 cpumask_clear_cpu(cpu, tick_broadcast_mask);
07bd1172 188 else
5d1d9a29 189 tick_device_setup_broadcast_func(dev);
07bd1172
TG
190
191 /*
192 * Clear the broadcast bit if the CPU is not in
193 * periodic broadcast on state.
194 */
195 if (!cpumask_test_cpu(cpu, tick_broadcast_on))
196 cpumask_clear_cpu(cpu, tick_broadcast_mask);
197
198 switch (tick_broadcast_device.mode) {
199 case TICKDEV_MODE_ONESHOT:
200 /*
201 * If the system is in oneshot mode we can
202 * unconditionally clear the oneshot mask bit,
203 * because the CPU is running and therefore
204 * not in an idle state which causes the power
205 * state affected device to stop. Let the
206 * caller initialize the device.
207 */
208 tick_broadcast_clear_oneshot(cpu);
209 ret = 0;
210 break;
211
212 case TICKDEV_MODE_PERIODIC:
213 /*
214 * If the system is in periodic mode, check
215 * whether the broadcast device can be
216 * switched off now.
217 */
218 if (cpumask_empty(tick_broadcast_mask) && bc)
219 clockevents_shutdown(bc);
220 /*
221 * If we kept the cpu in the broadcast mask,
222 * tell the caller to leave the per cpu device
223 * in shutdown state. The periodic interrupt
e0454311
TG
224 * is delivered by the broadcast device, if
225 * the broadcast device exists and is not
226 * hrtimer based.
07bd1172 227 */
e0454311
TG
228 if (bc && !(bc->features & CLOCK_EVT_FEAT_HRTIMER))
229 ret = cpumask_test_cpu(cpu, tick_broadcast_mask);
07bd1172
TG
230 break;
231 default:
07bd1172 232 break;
5590a536
TG
233 }
234 }
b5f91da0 235 raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
f8381cba
TG
236 return ret;
237}
238
12572dbb
MR
239#ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
240int tick_receive_broadcast(void)
241{
242 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
243 struct clock_event_device *evt = td->evtdev;
244
245 if (!evt)
246 return -ENODEV;
247
248 if (!evt->event_handler)
249 return -EINVAL;
250
251 evt->event_handler(evt);
252 return 0;
253}
254#endif
255
f8381cba 256/*
6b954823 257 * Broadcast the event to the cpus, which are set in the mask (mangled).
f8381cba 258 */
2951d5c0 259static bool tick_do_broadcast(struct cpumask *mask)
f8381cba 260{
186e3cb8 261 int cpu = smp_processor_id();
f8381cba 262 struct tick_device *td;
2951d5c0 263 bool local = false;
f8381cba
TG
264
265 /*
266 * Check, if the current cpu is in the mask
267 */
6b954823 268 if (cpumask_test_cpu(cpu, mask)) {
8eb23126
TG
269 struct clock_event_device *bc = tick_broadcast_device.evtdev;
270
6b954823 271 cpumask_clear_cpu(cpu, mask);
8eb23126
TG
272 /*
273 * We only run the local handler, if the broadcast
274 * device is not hrtimer based. Otherwise we run into
275 * a hrtimer recursion.
276 *
277 * local timer_interrupt()
278 * local_handler()
279 * expire_hrtimers()
280 * bc_handler()
281 * local_handler()
282 * expire_hrtimers()
283 */
284 local = !(bc->features & CLOCK_EVT_FEAT_HRTIMER);
f8381cba
TG
285 }
286
6b954823 287 if (!cpumask_empty(mask)) {
f8381cba
TG
288 /*
289 * It might be necessary to actually check whether the devices
290 * have different broadcast functions. For now, just use the
291 * one of the first device. This works as long as we have this
292 * misfeature only on x86 (lapic)
293 */
6b954823
RR
294 td = &per_cpu(tick_cpu_device, cpumask_first(mask));
295 td->evtdev->broadcast(mask);
f8381cba 296 }
2951d5c0 297 return local;
f8381cba
TG
298}
299
300/*
301 * Periodic broadcast:
302 * - invoke the broadcast handlers
303 */
2951d5c0 304static bool tick_do_periodic_broadcast(void)
f8381cba 305{
b352bc1c 306 cpumask_and(tmpmask, cpu_online_mask, tick_broadcast_mask);
2951d5c0 307 return tick_do_broadcast(tmpmask);
f8381cba
TG
308}
309
310/*
311 * Event handler for periodic broadcast ticks
312 */
313static void tick_handle_periodic_broadcast(struct clock_event_device *dev)
314{
2951d5c0
TG
315 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
316 bool bc_local;
d4496b39 317
627ee794 318 raw_spin_lock(&tick_broadcast_lock);
c4288334
TG
319
320 /* Handle spurious interrupts gracefully */
321 if (clockevent_state_shutdown(tick_broadcast_device.evtdev)) {
322 raw_spin_unlock(&tick_broadcast_lock);
323 return;
324 }
325
2951d5c0 326 bc_local = tick_do_periodic_broadcast();
627ee794 327
472c4a94 328 if (clockevent_state_oneshot(dev)) {
2951d5c0 329 ktime_t next = ktime_add(dev->next_event, tick_period);
f8381cba 330
2951d5c0
TG
331 clockevents_program_event(dev, next, true);
332 }
333 raw_spin_unlock(&tick_broadcast_lock);
f8381cba
TG
334
335 /*
2951d5c0
TG
336 * We run the handler of the local cpu after dropping
337 * tick_broadcast_lock because the handler might deadlock when
338 * trying to switch to oneshot mode.
f8381cba 339 */
2951d5c0
TG
340 if (bc_local)
341 td->evtdev->event_handler(td->evtdev);
f8381cba
TG
342}
343
592a438f
TG
344/**
345 * tick_broadcast_control - Enable/disable or force broadcast mode
346 * @mode: The selected broadcast mode
347 *
348 * Called when the system enters a state where affected tick devices
349 * might stop. Note: TICK_BROADCAST_FORCE cannot be undone.
350 *
351 * Called with interrupts disabled, so clockevents_lock is not
352 * required here because the local clock event device cannot go away
353 * under us.
f8381cba 354 */
592a438f 355void tick_broadcast_control(enum tick_broadcast_mode mode)
f8381cba
TG
356{
357 struct clock_event_device *bc, *dev;
358 struct tick_device *td;
9c17bcda 359 int cpu, bc_stopped;
f8381cba 360
592a438f 361 td = this_cpu_ptr(&tick_cpu_device);
f8381cba 362 dev = td->evtdev;
f8381cba
TG
363
364 /*
1595f452 365 * Is the device not affected by the powerstate ?
f8381cba 366 */
1595f452 367 if (!dev || !(dev->features & CLOCK_EVT_FEAT_C3STOP))
592a438f 368 return;
f8381cba 369
3dfbc884 370 if (!tick_device_is_functional(dev))
592a438f 371 return;
1595f452 372
592a438f
TG
373 raw_spin_lock(&tick_broadcast_lock);
374 cpu = smp_processor_id();
375 bc = tick_broadcast_device.evtdev;
b352bc1c 376 bc_stopped = cpumask_empty(tick_broadcast_mask);
9c17bcda 377
592a438f
TG
378 switch (mode) {
379 case TICK_BROADCAST_FORCE:
380 tick_broadcast_forced = 1;
381 case TICK_BROADCAST_ON:
07bd1172 382 cpumask_set_cpu(cpu, tick_broadcast_on);
b352bc1c 383 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_mask)) {
e0454311
TG
384 /*
385 * Only shutdown the cpu local device, if:
386 *
387 * - the broadcast device exists
388 * - the broadcast device is not a hrtimer based one
389 * - the broadcast device is in periodic mode to
390 * avoid a hickup during switch to oneshot mode
391 */
392 if (bc && !(bc->features & CLOCK_EVT_FEAT_HRTIMER) &&
393 tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
2344abbc 394 clockevents_shutdown(dev);
f8381cba 395 }
1595f452 396 break;
592a438f
TG
397
398 case TICK_BROADCAST_OFF:
399 if (tick_broadcast_forced)
07bd1172
TG
400 break;
401 cpumask_clear_cpu(cpu, tick_broadcast_on);
402 if (!tick_device_is_functional(dev))
403 break;
404 if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_mask)) {
07454bff
TG
405 if (tick_broadcast_device.mode ==
406 TICKDEV_MODE_PERIODIC)
f8381cba
TG
407 tick_setup_periodic(dev, 0);
408 }
1595f452 409 break;
f8381cba
TG
410 }
411
c4d029f2
TG
412 if (bc) {
413 if (cpumask_empty(tick_broadcast_mask)) {
414 if (!bc_stopped)
415 clockevents_shutdown(bc);
416 } else if (bc_stopped) {
417 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
418 tick_broadcast_start_periodic(bc);
419 else
420 tick_broadcast_setup_oneshot(bc);
421 }
f8381cba 422 }
592a438f 423 raw_spin_unlock(&tick_broadcast_lock);
f8381cba 424}
592a438f 425EXPORT_SYMBOL_GPL(tick_broadcast_control);
f8381cba
TG
426
427/*
428 * Set the periodic handler depending on broadcast on/off
429 */
430void tick_set_periodic_handler(struct clock_event_device *dev, int broadcast)
431{
432 if (!broadcast)
433 dev->event_handler = tick_handle_periodic;
434 else
435 dev->event_handler = tick_handle_periodic_broadcast;
436}
437
a49b116d 438#ifdef CONFIG_HOTPLUG_CPU
f8381cba
TG
439/*
440 * Remove a CPU from broadcasting
441 */
a49b116d 442void tick_shutdown_broadcast(unsigned int cpu)
f8381cba
TG
443{
444 struct clock_event_device *bc;
445 unsigned long flags;
f8381cba 446
b5f91da0 447 raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
f8381cba
TG
448
449 bc = tick_broadcast_device.evtdev;
b352bc1c 450 cpumask_clear_cpu(cpu, tick_broadcast_mask);
07bd1172 451 cpumask_clear_cpu(cpu, tick_broadcast_on);
f8381cba
TG
452
453 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
b352bc1c 454 if (bc && cpumask_empty(tick_broadcast_mask))
2344abbc 455 clockevents_shutdown(bc);
f8381cba
TG
456 }
457
b5f91da0 458 raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
f8381cba 459}
a49b116d 460#endif
79bf2bb3 461
6321dd60
TG
462void tick_suspend_broadcast(void)
463{
464 struct clock_event_device *bc;
465 unsigned long flags;
466
b5f91da0 467 raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
6321dd60
TG
468
469 bc = tick_broadcast_device.evtdev;
18de5bc4 470 if (bc)
2344abbc 471 clockevents_shutdown(bc);
6321dd60 472
b5f91da0 473 raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
6321dd60
TG
474}
475
f46481d0
TG
476/*
477 * This is called from tick_resume_local() on a resuming CPU. That's
478 * called from the core resume function, tick_unfreeze() and the magic XEN
479 * resume hackery.
480 *
481 * In none of these cases the broadcast device mode can change and the
482 * bit of the resuming CPU in the broadcast mask is safe as well.
483 */
484bool tick_resume_check_broadcast(void)
485{
486 if (tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT)
487 return false;
488 else
489 return cpumask_test_cpu(smp_processor_id(), tick_broadcast_mask);
490}
491
492void tick_resume_broadcast(void)
6321dd60
TG
493{
494 struct clock_event_device *bc;
495 unsigned long flags;
6321dd60 496
b5f91da0 497 raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
6321dd60
TG
498
499 bc = tick_broadcast_device.evtdev;
6321dd60 500
cd05a1f8 501 if (bc) {
554ef387 502 clockevents_tick_resume(bc);
18de5bc4 503
cd05a1f8
TG
504 switch (tick_broadcast_device.mode) {
505 case TICKDEV_MODE_PERIODIC:
b352bc1c 506 if (!cpumask_empty(tick_broadcast_mask))
cd05a1f8 507 tick_broadcast_start_periodic(bc);
cd05a1f8
TG
508 break;
509 case TICKDEV_MODE_ONESHOT:
b352bc1c 510 if (!cpumask_empty(tick_broadcast_mask))
080873ce 511 tick_resume_broadcast_oneshot(bc);
cd05a1f8
TG
512 break;
513 }
6321dd60 514 }
b5f91da0 515 raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
6321dd60
TG
516}
517
79bf2bb3
TG
518#ifdef CONFIG_TICK_ONESHOT
519
b352bc1c 520static cpumask_var_t tick_broadcast_oneshot_mask;
26517f3e 521static cpumask_var_t tick_broadcast_pending_mask;
989dcb64 522static cpumask_var_t tick_broadcast_force_mask;
79bf2bb3 523
289f480a 524/*
6b954823 525 * Exposed for debugging: see timer_list.c
289f480a 526 */
6b954823 527struct cpumask *tick_get_broadcast_oneshot_mask(void)
289f480a 528{
b352bc1c 529 return tick_broadcast_oneshot_mask;
289f480a
IM
530}
531
eaa907c5
TG
532/*
533 * Called before going idle with interrupts disabled. Checks whether a
534 * broadcast event from the other core is about to happen. We detected
535 * that in tick_broadcast_oneshot_control(). The callsite can use this
536 * to avoid a deep idle transition as we are about to get the
537 * broadcast IPI right away.
538 */
539int tick_check_broadcast_expired(void)
540{
541 return cpumask_test_cpu(smp_processor_id(), tick_broadcast_force_mask);
542}
543
d2348fb6
DL
544/*
545 * Set broadcast interrupt affinity
546 */
547static void tick_broadcast_set_affinity(struct clock_event_device *bc,
548 const struct cpumask *cpumask)
549{
550 if (!(bc->features & CLOCK_EVT_FEAT_DYNIRQ))
551 return;
552
553 if (cpumask_equal(bc->cpumask, cpumask))
554 return;
555
556 bc->cpumask = cpumask;
557 irq_set_affinity(bc->irq, bc->cpumask);
558}
559
298dbd1c
TG
560static void tick_broadcast_set_event(struct clock_event_device *bc, int cpu,
561 ktime_t expires)
79bf2bb3 562{
472c4a94 563 if (!clockevent_state_oneshot(bc))
d7eb231c 564 clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
b9a6a235 565
298dbd1c
TG
566 clockevents_program_event(bc, expires, 1);
567 tick_broadcast_set_affinity(bc, cpumask_of(cpu));
79bf2bb3
TG
568}
569
080873ce 570static void tick_resume_broadcast_oneshot(struct clock_event_device *bc)
cd05a1f8 571{
d7eb231c 572 clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
cd05a1f8
TG
573}
574
fb02fbc1
TG
575/*
576 * Called from irq_enter() when idle was interrupted to reenable the
577 * per cpu device.
578 */
e8fcaa5c 579void tick_check_oneshot_broadcast_this_cpu(void)
fb02fbc1 580{
e8fcaa5c 581 if (cpumask_test_cpu(smp_processor_id(), tick_broadcast_oneshot_mask)) {
22127e93 582 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
fb02fbc1 583
1f73a980
TG
584 /*
585 * We might be in the middle of switching over from
586 * periodic to oneshot. If the CPU has not yet
587 * switched over, leave the device alone.
588 */
589 if (td->mode == TICKDEV_MODE_ONESHOT) {
d7eb231c 590 clockevents_switch_state(td->evtdev,
77e32c89 591 CLOCK_EVT_STATE_ONESHOT);
1f73a980 592 }
fb02fbc1
TG
593 }
594}
595
79bf2bb3
TG
596/*
597 * Handle oneshot mode broadcasting
598 */
599static void tick_handle_oneshot_broadcast(struct clock_event_device *dev)
600{
601 struct tick_device *td;
cdc6f27d 602 ktime_t now, next_event;
d2348fb6 603 int cpu, next_cpu = 0;
298dbd1c 604 bool bc_local;
79bf2bb3 605
b5f91da0 606 raw_spin_lock(&tick_broadcast_lock);
79bf2bb3 607 dev->next_event.tv64 = KTIME_MAX;
cdc6f27d 608 next_event.tv64 = KTIME_MAX;
b352bc1c 609 cpumask_clear(tmpmask);
79bf2bb3
TG
610 now = ktime_get();
611 /* Find all expired events */
b352bc1c 612 for_each_cpu(cpu, tick_broadcast_oneshot_mask) {
79bf2bb3 613 td = &per_cpu(tick_cpu_device, cpu);
d2348fb6 614 if (td->evtdev->next_event.tv64 <= now.tv64) {
b352bc1c 615 cpumask_set_cpu(cpu, tmpmask);
26517f3e
TG
616 /*
617 * Mark the remote cpu in the pending mask, so
618 * it can avoid reprogramming the cpu local
619 * timer in tick_broadcast_oneshot_control().
620 */
621 cpumask_set_cpu(cpu, tick_broadcast_pending_mask);
d2348fb6 622 } else if (td->evtdev->next_event.tv64 < next_event.tv64) {
cdc6f27d 623 next_event.tv64 = td->evtdev->next_event.tv64;
d2348fb6
DL
624 next_cpu = cpu;
625 }
79bf2bb3
TG
626 }
627
2938d275
TG
628 /*
629 * Remove the current cpu from the pending mask. The event is
630 * delivered immediately in tick_do_broadcast() !
631 */
632 cpumask_clear_cpu(smp_processor_id(), tick_broadcast_pending_mask);
633
989dcb64
TG
634 /* Take care of enforced broadcast requests */
635 cpumask_or(tmpmask, tmpmask, tick_broadcast_force_mask);
636 cpumask_clear(tick_broadcast_force_mask);
637
c9b5a266
TG
638 /*
639 * Sanity check. Catch the case where we try to broadcast to
640 * offline cpus.
641 */
642 if (WARN_ON_ONCE(!cpumask_subset(tmpmask, cpu_online_mask)))
643 cpumask_and(tmpmask, tmpmask, cpu_online_mask);
644
79bf2bb3 645 /*
298dbd1c 646 * Wakeup the cpus which have an expired event.
cdc6f27d 647 */
298dbd1c 648 bc_local = tick_do_broadcast(tmpmask);
cdc6f27d
TG
649
650 /*
651 * Two reasons for reprogram:
652 *
653 * - The global event did not expire any CPU local
654 * events. This happens in dyntick mode, as the maximum PIT
655 * delta is quite small.
656 *
657 * - There are pending events on sleeping CPUs which were not
658 * in the event mask
79bf2bb3 659 */
298dbd1c
TG
660 if (next_event.tv64 != KTIME_MAX)
661 tick_broadcast_set_event(dev, next_cpu, next_event);
662
b5f91da0 663 raw_spin_unlock(&tick_broadcast_lock);
298dbd1c
TG
664
665 if (bc_local) {
666 td = this_cpu_ptr(&tick_cpu_device);
667 td->evtdev->event_handler(td->evtdev);
668 }
79bf2bb3
TG
669}
670
5d1638ac
PM
671static int broadcast_needs_cpu(struct clock_event_device *bc, int cpu)
672{
673 if (!(bc->features & CLOCK_EVT_FEAT_HRTIMER))
674 return 0;
675 if (bc->next_event.tv64 == KTIME_MAX)
676 return 0;
677 return bc->bound_on == cpu ? -EBUSY : 0;
678}
679
680static void broadcast_shutdown_local(struct clock_event_device *bc,
681 struct clock_event_device *dev)
682{
683 /*
684 * For hrtimer based broadcasting we cannot shutdown the cpu
685 * local device if our own event is the first one to expire or
686 * if we own the broadcast timer.
687 */
688 if (bc->features & CLOCK_EVT_FEAT_HRTIMER) {
689 if (broadcast_needs_cpu(bc, smp_processor_id()))
690 return;
691 if (dev->next_event.tv64 < bc->next_event.tv64)
692 return;
693 }
d7eb231c 694 clockevents_switch_state(dev, CLOCK_EVT_STATE_SHUTDOWN);
5d1638ac
PM
695}
696
f32dd117 697int __tick_broadcast_oneshot_control(enum tick_broadcast_state state)
79bf2bb3
TG
698{
699 struct clock_event_device *bc, *dev;
da7e6f45 700 int cpu, ret = 0;
1fe5d5c3 701 ktime_t now;
79bf2bb3 702
b78f3f3c
TG
703 /*
704 * If there is no broadcast device, tell the caller not to go
705 * into deep idle.
706 */
707 if (!tick_broadcast_device.evtdev)
708 return -EBUSY;
709
e3ac79e0 710 dev = this_cpu_ptr(&tick_cpu_device)->evtdev;
79bf2bb3 711
1fe5d5c3 712 raw_spin_lock(&tick_broadcast_lock);
7372b0b1 713 bc = tick_broadcast_device.evtdev;
1fe5d5c3 714 cpu = smp_processor_id();
79bf2bb3 715
1fe5d5c3 716 if (state == TICK_BROADCAST_ENTER) {
d5113e13
TG
717 /*
718 * If the current CPU owns the hrtimer broadcast
719 * mechanism, it cannot go deep idle and we do not add
720 * the CPU to the broadcast mask. We don't have to go
721 * through the EXIT path as the local timer is not
722 * shutdown.
723 */
724 ret = broadcast_needs_cpu(bc, cpu);
725 if (ret)
726 goto out;
727
e3ac79e0
TG
728 /*
729 * If the broadcast device is in periodic mode, we
730 * return.
731 */
d3325726
TG
732 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
733 /* If it is a hrtimer based broadcast, return busy */
734 if (bc->features & CLOCK_EVT_FEAT_HRTIMER)
735 ret = -EBUSY;
e3ac79e0 736 goto out;
d3325726 737 }
e3ac79e0 738
b352bc1c 739 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_oneshot_mask)) {
2938d275 740 WARN_ON_ONCE(cpumask_test_cpu(cpu, tick_broadcast_pending_mask));
d5113e13
TG
741
742 /* Conditionally shut down the local timer. */
5d1638ac 743 broadcast_shutdown_local(bc, dev);
d5113e13 744
989dcb64
TG
745 /*
746 * We only reprogram the broadcast timer if we
747 * did not mark ourself in the force mask and
748 * if the cpu local event is earlier than the
749 * broadcast event. If the current CPU is in
750 * the force mask, then we are going to be
0cc5281a
TG
751 * woken by the IPI right away; we return
752 * busy, so the CPU does not try to go deep
753 * idle.
989dcb64 754 */
0cc5281a
TG
755 if (cpumask_test_cpu(cpu, tick_broadcast_force_mask)) {
756 ret = -EBUSY;
757 } else if (dev->next_event.tv64 < bc->next_event.tv64) {
298dbd1c 758 tick_broadcast_set_event(bc, cpu, dev->next_event);
d5113e13
TG
759 /*
760 * In case of hrtimer broadcasts the
761 * programming might have moved the
762 * timer to this cpu. If yes, remove
763 * us from the broadcast mask and
764 * return busy.
765 */
766 ret = broadcast_needs_cpu(bc, cpu);
767 if (ret) {
768 cpumask_clear_cpu(cpu,
769 tick_broadcast_oneshot_mask);
770 }
0cc5281a 771 }
79bf2bb3
TG
772 }
773 } else {
b352bc1c 774 if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_oneshot_mask)) {
d7eb231c 775 clockevents_switch_state(dev, CLOCK_EVT_STATE_ONESHOT);
26517f3e
TG
776 /*
777 * The cpu which was handling the broadcast
778 * timer marked this cpu in the broadcast
779 * pending mask and fired the broadcast
780 * IPI. So we are going to handle the expired
781 * event anyway via the broadcast IPI
782 * handler. No need to reprogram the timer
783 * with an already expired event.
784 */
785 if (cpumask_test_and_clear_cpu(cpu,
786 tick_broadcast_pending_mask))
787 goto out;
788
ea8deb8d
DL
789 /*
790 * Bail out if there is no next event.
791 */
792 if (dev->next_event.tv64 == KTIME_MAX)
793 goto out;
989dcb64
TG
794 /*
795 * If the pending bit is not set, then we are
796 * either the CPU handling the broadcast
797 * interrupt or we got woken by something else.
798 *
799 * We are not longer in the broadcast mask, so
800 * if the cpu local expiry time is already
801 * reached, we would reprogram the cpu local
802 * timer with an already expired event.
803 *
804 * This can lead to a ping-pong when we return
805 * to idle and therefor rearm the broadcast
806 * timer before the cpu local timer was able
807 * to fire. This happens because the forced
808 * reprogramming makes sure that the event
809 * will happen in the future and depending on
810 * the min_delta setting this might be far
811 * enough out that the ping-pong starts.
812 *
813 * If the cpu local next_event has expired
814 * then we know that the broadcast timer
815 * next_event has expired as well and
816 * broadcast is about to be handled. So we
817 * avoid reprogramming and enforce that the
818 * broadcast handler, which did not run yet,
819 * will invoke the cpu local handler.
820 *
821 * We cannot call the handler directly from
822 * here, because we might be in a NOHZ phase
823 * and we did not go through the irq_enter()
824 * nohz fixups.
825 */
826 now = ktime_get();
827 if (dev->next_event.tv64 <= now.tv64) {
828 cpumask_set_cpu(cpu, tick_broadcast_force_mask);
829 goto out;
830 }
831 /*
832 * We got woken by something else. Reprogram
833 * the cpu local timer device.
834 */
26517f3e 835 tick_program_event(dev->next_event, 1);
79bf2bb3
TG
836 }
837 }
26517f3e 838out:
1fe5d5c3 839 raw_spin_unlock(&tick_broadcast_lock);
da7e6f45 840 return ret;
79bf2bb3
TG
841}
842
5590a536
TG
843/*
844 * Reset the one shot broadcast for a cpu
845 *
846 * Called with tick_broadcast_lock held
847 */
848static void tick_broadcast_clear_oneshot(int cpu)
849{
b352bc1c 850 cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
dd5fd9b9 851 cpumask_clear_cpu(cpu, tick_broadcast_pending_mask);
5590a536
TG
852}
853
6b954823
RR
854static void tick_broadcast_init_next_event(struct cpumask *mask,
855 ktime_t expires)
7300711e
TG
856{
857 struct tick_device *td;
858 int cpu;
859
5db0e1e9 860 for_each_cpu(cpu, mask) {
7300711e
TG
861 td = &per_cpu(tick_cpu_device, cpu);
862 if (td->evtdev)
863 td->evtdev->next_event = expires;
864 }
865}
866
79bf2bb3 867/**
8dce39c2 868 * tick_broadcast_setup_oneshot - setup the broadcast device
79bf2bb3
TG
869 */
870void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
871{
07f4beb0
TG
872 int cpu = smp_processor_id();
873
9c17bcda
TG
874 /* Set it up only once ! */
875 if (bc->event_handler != tick_handle_oneshot_broadcast) {
472c4a94 876 int was_periodic = clockevent_state_periodic(bc);
7300711e 877
9c17bcda 878 bc->event_handler = tick_handle_oneshot_broadcast;
7300711e 879
7300711e
TG
880 /*
881 * We must be careful here. There might be other CPUs
882 * waiting for periodic broadcast. We need to set the
883 * oneshot_mask bits for those and program the
884 * broadcast device to fire.
885 */
b352bc1c
TG
886 cpumask_copy(tmpmask, tick_broadcast_mask);
887 cpumask_clear_cpu(cpu, tmpmask);
888 cpumask_or(tick_broadcast_oneshot_mask,
889 tick_broadcast_oneshot_mask, tmpmask);
6b954823 890
b352bc1c 891 if (was_periodic && !cpumask_empty(tmpmask)) {
d7eb231c 892 clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
b352bc1c 893 tick_broadcast_init_next_event(tmpmask,
6b954823 894 tick_next_period);
298dbd1c 895 tick_broadcast_set_event(bc, cpu, tick_next_period);
7300711e
TG
896 } else
897 bc->next_event.tv64 = KTIME_MAX;
07f4beb0
TG
898 } else {
899 /*
900 * The first cpu which switches to oneshot mode sets
901 * the bit for all other cpus which are in the general
902 * (periodic) broadcast mask. So the bit is set and
903 * would prevent the first broadcast enter after this
904 * to program the bc device.
905 */
906 tick_broadcast_clear_oneshot(cpu);
9c17bcda 907 }
79bf2bb3
TG
908}
909
910/*
911 * Select oneshot operating mode for the broadcast device
912 */
913void tick_broadcast_switch_to_oneshot(void)
914{
915 struct clock_event_device *bc;
916 unsigned long flags;
917
b5f91da0 918 raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
fa4da365
SS
919
920 tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
79bf2bb3
TG
921 bc = tick_broadcast_device.evtdev;
922 if (bc)
923 tick_broadcast_setup_oneshot(bc);
77b0d60c 924
b5f91da0 925 raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
79bf2bb3
TG
926}
927
a49b116d
TG
928#ifdef CONFIG_HOTPLUG_CPU
929void hotplug_cpu__broadcast_tick_pull(int deadcpu)
930{
931 struct clock_event_device *bc;
932 unsigned long flags;
933
934 raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
935 bc = tick_broadcast_device.evtdev;
936
937 if (bc && broadcast_needs_cpu(bc, deadcpu)) {
938 /* This moves the broadcast assignment to this CPU: */
939 clockevents_program_event(bc, bc->next_event, 1);
940 }
941 raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
942}
79bf2bb3
TG
943
944/*
945 * Remove a dead CPU from broadcasting
946 */
a49b116d 947void tick_shutdown_broadcast_oneshot(unsigned int cpu)
79bf2bb3 948{
79bf2bb3 949 unsigned long flags;
79bf2bb3 950
b5f91da0 951 raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
79bf2bb3 952
31d9b393 953 /*
c9b5a266
TG
954 * Clear the broadcast masks for the dead cpu, but do not stop
955 * the broadcast device!
31d9b393 956 */
b352bc1c 957 cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
c9b5a266
TG
958 cpumask_clear_cpu(cpu, tick_broadcast_pending_mask);
959 cpumask_clear_cpu(cpu, tick_broadcast_force_mask);
79bf2bb3 960
b5f91da0 961 raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
79bf2bb3 962}
a49b116d 963#endif
79bf2bb3 964
27ce4cb4
TG
965/*
966 * Check, whether the broadcast device is in one shot mode
967 */
968int tick_broadcast_oneshot_active(void)
969{
970 return tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT;
971}
972
3a142a06
TG
973/*
974 * Check whether the broadcast device supports oneshot.
975 */
976bool tick_broadcast_oneshot_available(void)
977{
978 struct clock_event_device *bc = tick_broadcast_device.evtdev;
979
980 return bc ? bc->features & CLOCK_EVT_FEAT_ONESHOT : false;
981}
982
f32dd117
TG
983#else
984int __tick_broadcast_oneshot_control(enum tick_broadcast_state state)
985{
986 struct clock_event_device *bc = tick_broadcast_device.evtdev;
987
988 if (!bc || (bc->features & CLOCK_EVT_FEAT_HRTIMER))
989 return -EBUSY;
990
991 return 0;
992}
79bf2bb3 993#endif
b352bc1c
TG
994
995void __init tick_broadcast_init(void)
996{
fbd44a60 997 zalloc_cpumask_var(&tick_broadcast_mask, GFP_NOWAIT);
07bd1172 998 zalloc_cpumask_var(&tick_broadcast_on, GFP_NOWAIT);
fbd44a60 999 zalloc_cpumask_var(&tmpmask, GFP_NOWAIT);
b352bc1c 1000#ifdef CONFIG_TICK_ONESHOT
fbd44a60
TG
1001 zalloc_cpumask_var(&tick_broadcast_oneshot_mask, GFP_NOWAIT);
1002 zalloc_cpumask_var(&tick_broadcast_pending_mask, GFP_NOWAIT);
1003 zalloc_cpumask_var(&tick_broadcast_force_mask, GFP_NOWAIT);
b352bc1c
TG
1004#endif
1005}