1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * This file contains the base functions to manage periodic tick
4 * related events.
5 *
6 * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
7 * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
8 * Copyright(C) 2006-2007, Timesys Corp., Thomas Gleixner
9 */
10 #include <linux/cpu.h>
11 #include <linux/err.h>
12 #include <linux/hrtimer.h>
13 #include <linux/interrupt.h>
14 #include <linux/nmi.h>
15 #include <linux/percpu.h>
16 #include <linux/profile.h>
17 #include <linux/sched.h>
18 #include <linux/module.h>
19 #include <trace/events/power.h>
20
21 #include <asm/irq_regs.h>
22
23 #include "tick-internal.h"
24
25 /*
26 * Tick devices
27 */
28 DEFINE_PER_CPU(struct tick_device, tick_cpu_device);
29 /*
30 * Tick next event: keeps track of the tick time
31 */
32 ktime_t tick_next_period;
33 ktime_t tick_period;
34
35 /*
36 * tick_do_timer_cpu is a timer core internal variable which holds the CPU NR
37 * which is responsible for calling do_timer(), i.e. the timekeeping stuff. This
38 * variable has two functions:
39 *
40 * 1) Prevent a thundering herd issue of a gazillion of CPUs trying to grab the
41 * timekeeping lock all at once. Only the CPU which is assigned to do the
42 * update is handling it.
43 *
44 * 2) Hand off the duty in the NOHZ idle case by setting the value to
45 * TICK_DO_TIMER_NONE, i.e. a non existing CPU. So the next cpu which looks
46 * at it will take over and keep the time keeping alive. The handover
47 * procedure also covers cpu hotplug.
48 */
49 int tick_do_timer_cpu __read_mostly = TICK_DO_TIMER_BOOT;
50 #ifdef CONFIG_NO_HZ_FULL
51 /*
52 * tick_do_timer_boot_cpu indicates the boot CPU temporarily owns
53 * tick_do_timer_cpu and it should be taken over by an eligible secondary
54 * when one comes online.
55 */
56 static int tick_do_timer_boot_cpu __read_mostly = -1;
57 #endif
58
59 /*
60 * Debugging: see timer_list.c
61 */
tick_get_device(int cpu)62 struct tick_device *tick_get_device(int cpu)
63 {
64 return &per_cpu(tick_cpu_device, cpu);
65 }
66
67 /**
68 * tick_is_oneshot_available - check for a oneshot capable event device
69 */
tick_is_oneshot_available(void)70 int tick_is_oneshot_available(void)
71 {
72 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
73
74 if (!dev || !(dev->features & CLOCK_EVT_FEAT_ONESHOT))
75 return 0;
76 if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
77 return 1;
78 return tick_broadcast_oneshot_available();
79 }
80
81 /*
82 * Periodic tick
83 */
tick_periodic(int cpu)84 static void tick_periodic(int cpu)
85 {
86 if (tick_do_timer_cpu == cpu) {
87 raw_spin_lock(&jiffies_lock);
88 write_seqcount_begin(&jiffies_seq);
89
90 /* Keep track of the next tick event */
91 tick_next_period = ktime_add(tick_next_period, tick_period);
92
93 do_timer(1);
94 write_seqcount_end(&jiffies_seq);
95 raw_spin_unlock(&jiffies_lock);
96 update_wall_time();
97 }
98
99 update_process_times(user_mode(get_irq_regs()));
100 profile_tick(CPU_PROFILING);
101 }
102
103 /*
104 * Event handler for periodic ticks
105 */
tick_handle_periodic(struct clock_event_device * dev)106 void tick_handle_periodic(struct clock_event_device *dev)
107 {
108 int cpu = smp_processor_id();
109 ktime_t next = dev->next_event;
110
111 tick_periodic(cpu);
112
113 #if defined(CONFIG_HIGH_RES_TIMERS) || defined(CONFIG_NO_HZ_COMMON)
114 /*
115 * The cpu might have transitioned to HIGHRES or NOHZ mode via
116 * update_process_times() -> run_local_timers() ->
117 * hrtimer_run_queues().
118 */
119 if (dev->event_handler != tick_handle_periodic)
120 return;
121 #endif
122
123 if (!clockevent_state_oneshot(dev))
124 return;
125 for (;;) {
126 /*
127 * Setup the next period for devices, which do not have
128 * periodic mode:
129 */
130 next = ktime_add(next, tick_period);
131
132 if (!clockevents_program_event(dev, next, false))
133 return;
134 /*
135 * Have to be careful here. If we're in oneshot mode,
136 * before we call tick_periodic() in a loop, we need
137 * to be sure we're using a real hardware clocksource.
138 * Otherwise we could get trapped in an infinite
139 * loop, as the tick_periodic() increments jiffies,
140 * which then will increment time, possibly causing
141 * the loop to trigger again and again.
142 */
143 if (timekeeping_valid_for_hres())
144 tick_periodic(cpu);
145 }
146 }
147
148 /*
149 * Setup the device for a periodic tick
150 */
tick_setup_periodic(struct clock_event_device * dev,int broadcast)151 void tick_setup_periodic(struct clock_event_device *dev, int broadcast)
152 {
153 tick_set_periodic_handler(dev, broadcast);
154
155 /* Broadcast setup ? */
156 if (!tick_device_is_functional(dev))
157 return;
158
159 if ((dev->features & CLOCK_EVT_FEAT_PERIODIC) &&
160 !tick_broadcast_oneshot_active()) {
161 clockevents_switch_state(dev, CLOCK_EVT_STATE_PERIODIC);
162 } else {
163 unsigned int seq;
164 ktime_t next;
165
166 do {
167 seq = read_seqcount_begin(&jiffies_seq);
168 next = tick_next_period;
169 } while (read_seqcount_retry(&jiffies_seq, seq));
170
171 clockevents_switch_state(dev, CLOCK_EVT_STATE_ONESHOT);
172
173 for (;;) {
174 if (!clockevents_program_event(dev, next, false))
175 return;
176 next = ktime_add(next, tick_period);
177 }
178 }
179 }
180
181 #ifdef CONFIG_NO_HZ_FULL
giveup_do_timer(void * info)182 static void giveup_do_timer(void *info)
183 {
184 int cpu = *(unsigned int *)info;
185
186 WARN_ON(tick_do_timer_cpu != smp_processor_id());
187
188 tick_do_timer_cpu = cpu;
189 }
190
tick_take_do_timer_from_boot(void)191 static void tick_take_do_timer_from_boot(void)
192 {
193 int cpu = smp_processor_id();
194 int from = tick_do_timer_boot_cpu;
195
196 if (from >= 0 && from != cpu)
197 smp_call_function_single(from, giveup_do_timer, &cpu, 1);
198 }
199 #endif
200
201 /*
202 * Setup the tick device
203 */
tick_setup_device(struct tick_device * td,struct clock_event_device * newdev,int cpu,const struct cpumask * cpumask)204 static void tick_setup_device(struct tick_device *td,
205 struct clock_event_device *newdev, int cpu,
206 const struct cpumask *cpumask)
207 {
208 void (*handler)(struct clock_event_device *) = NULL;
209 ktime_t next_event = 0;
210
211 /*
212 * First device setup ?
213 */
214 if (!td->evtdev) {
215 /*
216 * If no cpu took the do_timer update, assign it to
217 * this cpu:
218 */
219 if (tick_do_timer_cpu == TICK_DO_TIMER_BOOT) {
220 tick_do_timer_cpu = cpu;
221
222 tick_next_period = ktime_get();
223 tick_period = NSEC_PER_SEC / HZ;
224 #ifdef CONFIG_NO_HZ_FULL
225 /*
226 * The boot CPU may be nohz_full, in which case set
227 * tick_do_timer_boot_cpu so the first housekeeping
228 * secondary that comes up will take do_timer from
229 * us.
230 */
231 if (tick_nohz_full_cpu(cpu))
232 tick_do_timer_boot_cpu = cpu;
233
234 } else if (tick_do_timer_boot_cpu != -1 &&
235 !tick_nohz_full_cpu(cpu)) {
236 tick_take_do_timer_from_boot();
237 tick_do_timer_boot_cpu = -1;
238 WARN_ON(tick_do_timer_cpu != cpu);
239 #endif
240 }
241
242 /*
243 * Startup in periodic mode first.
244 */
245 td->mode = TICKDEV_MODE_PERIODIC;
246 } else {
247 handler = td->evtdev->event_handler;
248 next_event = td->evtdev->next_event;
249 td->evtdev->event_handler = clockevents_handle_noop;
250 }
251
252 td->evtdev = newdev;
253
254 /*
255 * When the device is not per cpu, pin the interrupt to the
256 * current cpu:
257 */
258 if (!cpumask_equal(newdev->cpumask, cpumask))
259 irq_set_affinity(newdev->irq, cpumask);
260
261 /*
262 * When global broadcasting is active, check if the current
263 * device is registered as a placeholder for broadcast mode.
264 * This allows us to handle this x86 misfeature in a generic
265 * way. This function also returns !=0 when we keep the
266 * current active broadcast state for this CPU.
267 */
268 if (tick_device_uses_broadcast(newdev, cpu))
269 return;
270
271 if (td->mode == TICKDEV_MODE_PERIODIC)
272 tick_setup_periodic(newdev, 0);
273 else
274 tick_setup_oneshot(newdev, handler, next_event);
275 }
276
tick_install_replacement(struct clock_event_device * newdev)277 void tick_install_replacement(struct clock_event_device *newdev)
278 {
279 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
280 int cpu = smp_processor_id();
281
282 clockevents_exchange_device(td->evtdev, newdev);
283 tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
284 if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
285 tick_oneshot_notify();
286 }
287
tick_check_percpu(struct clock_event_device * curdev,struct clock_event_device * newdev,int cpu)288 static bool tick_check_percpu(struct clock_event_device *curdev,
289 struct clock_event_device *newdev, int cpu)
290 {
291 if (!cpumask_test_cpu(cpu, newdev->cpumask))
292 return false;
293 if (cpumask_equal(newdev->cpumask, cpumask_of(cpu)))
294 return true;
295 /* Check if irq affinity can be set */
296 if (newdev->irq >= 0 && !irq_can_set_affinity(newdev->irq))
297 return false;
298 /* Prefer an existing cpu local device */
299 if (curdev && cpumask_equal(curdev->cpumask, cpumask_of(cpu)))
300 return false;
301 return true;
302 }
303
tick_check_preferred(struct clock_event_device * curdev,struct clock_event_device * newdev)304 static bool tick_check_preferred(struct clock_event_device *curdev,
305 struct clock_event_device *newdev)
306 {
307 /* Prefer oneshot capable device */
308 if (!(newdev->features & CLOCK_EVT_FEAT_ONESHOT)) {
309 if (curdev && (curdev->features & CLOCK_EVT_FEAT_ONESHOT))
310 return false;
311 if (tick_oneshot_mode_active())
312 return false;
313 }
314
315 /*
316 * Use the higher rated one, but prefer a CPU local device with a lower
317 * rating than a non-CPU local device
318 */
319 return !curdev ||
320 newdev->rating > curdev->rating ||
321 !cpumask_equal(curdev->cpumask, newdev->cpumask);
322 }
323
324 /*
325 * Check whether the new device is a better fit than curdev. curdev
326 * can be NULL !
327 */
tick_check_replacement(struct clock_event_device * curdev,struct clock_event_device * newdev)328 bool tick_check_replacement(struct clock_event_device *curdev,
329 struct clock_event_device *newdev)
330 {
331 if (!tick_check_percpu(curdev, newdev, smp_processor_id()))
332 return false;
333
334 return tick_check_preferred(curdev, newdev);
335 }
336
337 /*
338 * Check, if the new registered device should be used. Called with
339 * clockevents_lock held and interrupts disabled.
340 */
tick_check_new_device(struct clock_event_device * newdev)341 void tick_check_new_device(struct clock_event_device *newdev)
342 {
343 struct clock_event_device *curdev;
344 struct tick_device *td;
345 int cpu;
346
347 cpu = smp_processor_id();
348 td = &per_cpu(tick_cpu_device, cpu);
349 curdev = td->evtdev;
350
351 /* cpu local device ? */
352 if (!tick_check_percpu(curdev, newdev, cpu))
353 goto out_bc;
354
355 /* Preference decision */
356 if (!tick_check_preferred(curdev, newdev))
357 goto out_bc;
358
359 if (!try_module_get(newdev->owner))
360 return;
361
362 /*
363 * Replace the eventually existing device by the new
364 * device. If the current device is the broadcast device, do
365 * not give it back to the clockevents layer !
366 */
367 if (tick_is_broadcast_device(curdev)) {
368 clockevents_shutdown(curdev);
369 curdev = NULL;
370 }
371 clockevents_exchange_device(curdev, newdev);
372 tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
373 if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
374 tick_oneshot_notify();
375 return;
376
377 out_bc:
378 /*
379 * Can the new device be used as a broadcast device ?
380 */
381 tick_install_broadcast_device(newdev);
382 }
383
384 /**
385 * tick_broadcast_oneshot_control - Enter/exit broadcast oneshot mode
386 * @state: The target state (enter/exit)
387 *
388 * The system enters/leaves a state, where affected devices might stop
389 * Returns 0 on success, -EBUSY if the cpu is used to broadcast wakeups.
390 *
391 * Called with interrupts disabled, so clockevents_lock is not
392 * required here because the local clock event device cannot go away
393 * under us.
394 */
tick_broadcast_oneshot_control(enum tick_broadcast_state state)395 int tick_broadcast_oneshot_control(enum tick_broadcast_state state)
396 {
397 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
398
399 if (!(td->evtdev->features & CLOCK_EVT_FEAT_C3STOP))
400 return 0;
401
402 return __tick_broadcast_oneshot_control(state);
403 }
404 EXPORT_SYMBOL_GPL(tick_broadcast_oneshot_control);
405
406 #ifdef CONFIG_HOTPLUG_CPU
407 /*
408 * Transfer the do_timer job away from a dying cpu.
409 *
410 * Called with interrupts disabled. Not locking required. If
411 * tick_do_timer_cpu is owned by this cpu, nothing can change it.
412 */
tick_handover_do_timer(void)413 void tick_handover_do_timer(void)
414 {
415 if (tick_do_timer_cpu == smp_processor_id()) {
416 int cpu = cpumask_first(cpu_online_mask);
417
418 tick_do_timer_cpu = (cpu < nr_cpu_ids) ? cpu :
419 TICK_DO_TIMER_NONE;
420 }
421 }
422
423 /*
424 * Shutdown an event device on a given cpu:
425 *
426 * This is called on a life CPU, when a CPU is dead. So we cannot
427 * access the hardware device itself.
428 * We just set the mode and remove it from the lists.
429 */
tick_shutdown(unsigned int cpu)430 void tick_shutdown(unsigned int cpu)
431 {
432 struct tick_device *td = &per_cpu(tick_cpu_device, cpu);
433 struct clock_event_device *dev = td->evtdev;
434
435 td->mode = TICKDEV_MODE_PERIODIC;
436 if (dev) {
437 /*
438 * Prevent that the clock events layer tries to call
439 * the set mode function!
440 */
441 clockevent_set_state(dev, CLOCK_EVT_STATE_DETACHED);
442 clockevents_exchange_device(dev, NULL);
443 dev->event_handler = clockevents_handle_noop;
444 td->evtdev = NULL;
445 }
446 }
447 #endif
448
449 /**
450 * tick_suspend_local - Suspend the local tick device
451 *
452 * Called from the local cpu for freeze with interrupts disabled.
453 *
454 * No locks required. Nothing can change the per cpu device.
455 */
tick_suspend_local(void)456 void tick_suspend_local(void)
457 {
458 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
459
460 clockevents_shutdown(td->evtdev);
461 }
462
463 /**
464 * tick_resume_local - Resume the local tick device
465 *
466 * Called from the local CPU for unfreeze or XEN resume magic.
467 *
468 * No locks required. Nothing can change the per cpu device.
469 */
tick_resume_local(void)470 void tick_resume_local(void)
471 {
472 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
473 bool broadcast = tick_resume_check_broadcast();
474
475 clockevents_tick_resume(td->evtdev);
476 if (!broadcast) {
477 if (td->mode == TICKDEV_MODE_PERIODIC)
478 tick_setup_periodic(td->evtdev, 0);
479 else
480 tick_resume_oneshot();
481 }
482 }
483
484 /**
485 * tick_suspend - Suspend the tick and the broadcast device
486 *
487 * Called from syscore_suspend() via timekeeping_suspend with only one
488 * CPU online and interrupts disabled or from tick_unfreeze() under
489 * tick_freeze_lock.
490 *
491 * No locks required. Nothing can change the per cpu device.
492 */
tick_suspend(void)493 void tick_suspend(void)
494 {
495 tick_suspend_local();
496 tick_suspend_broadcast();
497 }
498
499 /**
500 * tick_resume - Resume the tick and the broadcast device
501 *
502 * Called from syscore_resume() via timekeeping_resume with only one
503 * CPU online and interrupts disabled.
504 *
505 * No locks required. Nothing can change the per cpu device.
506 */
tick_resume(void)507 void tick_resume(void)
508 {
509 tick_resume_broadcast();
510 tick_resume_local();
511 }
512
513 #ifdef CONFIG_SUSPEND
514 static DEFINE_RAW_SPINLOCK(tick_freeze_lock);
515 static unsigned int tick_freeze_depth;
516
517 /**
518 * tick_freeze - Suspend the local tick and (possibly) timekeeping.
519 *
520 * Check if this is the last online CPU executing the function and if so,
521 * suspend timekeeping. Otherwise suspend the local tick.
522 *
523 * Call with interrupts disabled. Must be balanced with %tick_unfreeze().
524 * Interrupts must not be enabled before the subsequent %tick_unfreeze().
525 */
tick_freeze(void)526 void tick_freeze(void)
527 {
528 raw_spin_lock(&tick_freeze_lock);
529
530 tick_freeze_depth++;
531 if (tick_freeze_depth == num_online_cpus()) {
532 trace_suspend_resume(TPS("timekeeping_freeze"),
533 smp_processor_id(), true);
534 system_state = SYSTEM_SUSPEND;
535 sched_clock_suspend();
536 timekeeping_suspend();
537 } else {
538 tick_suspend_local();
539 }
540
541 raw_spin_unlock(&tick_freeze_lock);
542 }
543
544 /**
545 * tick_unfreeze - Resume the local tick and (possibly) timekeeping.
546 *
547 * Check if this is the first CPU executing the function and if so, resume
548 * timekeeping. Otherwise resume the local tick.
549 *
550 * Call with interrupts disabled. Must be balanced with %tick_freeze().
551 * Interrupts must not be enabled after the preceding %tick_freeze().
552 */
tick_unfreeze(void)553 void tick_unfreeze(void)
554 {
555 raw_spin_lock(&tick_freeze_lock);
556
557 if (tick_freeze_depth == num_online_cpus()) {
558 timekeeping_resume();
559 sched_clock_resume();
560 system_state = SYSTEM_RUNNING;
561 trace_suspend_resume(TPS("timekeeping_freeze"),
562 smp_processor_id(), false);
563 } else {
564 touch_softlockup_watchdog();
565 tick_resume_local();
566 }
567
568 tick_freeze_depth--;
569
570 raw_spin_unlock(&tick_freeze_lock);
571 }
572 #endif /* CONFIG_SUSPEND */
573
574 /**
575 * tick_init - initialize the tick control
576 */
tick_init(void)577 void __init tick_init(void)
578 {
579 tick_broadcast_init();
580 tick_nohz_init();
581 }
582