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 #include <trace/hooks/sched.h>
21
22 #include <asm/irq_regs.h>
23
24 #include "tick-internal.h"
25
26 /*
27 * Tick devices
28 */
29 DEFINE_PER_CPU(struct tick_device, tick_cpu_device);
30 /*
31 * Tick next event: keeps track of the tick time
32 */
33 ktime_t tick_next_period;
34 ktime_t tick_period;
35
36 /*
37 * tick_do_timer_cpu is a timer core internal variable which holds the CPU NR
38 * which is responsible for calling do_timer(), i.e. the timekeeping stuff. This
39 * variable has two functions:
40 *
41 * 1) Prevent a thundering herd issue of a gazillion of CPUs trying to grab the
42 * timekeeping lock all at once. Only the CPU which is assigned to do the
43 * update is handling it.
44 *
45 * 2) Hand off the duty in the NOHZ idle case by setting the value to
46 * TICK_DO_TIMER_NONE, i.e. a non existing CPU. So the next cpu which looks
47 * at it will take over and keep the time keeping alive. The handover
48 * procedure also covers cpu hotplug.
49 */
50 int tick_do_timer_cpu __read_mostly = TICK_DO_TIMER_BOOT;
51 #ifdef CONFIG_NO_HZ_FULL
52 /*
53 * tick_do_timer_boot_cpu indicates the boot CPU temporarily owns
54 * tick_do_timer_cpu and it should be taken over by an eligible secondary
55 * when one comes online.
56 */
57 static int tick_do_timer_boot_cpu __read_mostly = -1;
58 #endif
59
60 /*
61 * Debugging: see timer_list.c
62 */
tick_get_device(int cpu)63 struct tick_device *tick_get_device(int cpu)
64 {
65 return &per_cpu(tick_cpu_device, cpu);
66 }
67
68 /**
69 * tick_is_oneshot_available - check for a oneshot capable event device
70 */
tick_is_oneshot_available(void)71 int tick_is_oneshot_available(void)
72 {
73 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
74
75 if (!dev || !(dev->features & CLOCK_EVT_FEAT_ONESHOT))
76 return 0;
77 if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
78 return 1;
79 return tick_broadcast_oneshot_available();
80 }
81
82 /*
83 * Periodic tick
84 */
tick_periodic(int cpu)85 static void tick_periodic(int cpu)
86 {
87 if (tick_do_timer_cpu == cpu) {
88 raw_spin_lock(&jiffies_lock);
89 write_seqcount_begin(&jiffies_seq);
90
91 /* Keep track of the next tick event */
92 tick_next_period = ktime_add(tick_next_period, tick_period);
93
94 do_timer(1);
95 write_seqcount_end(&jiffies_seq);
96 raw_spin_unlock(&jiffies_lock);
97 update_wall_time();
98 trace_android_vh_jiffies_update(NULL);
99 }
100
101 update_process_times(user_mode(get_irq_regs()));
102 profile_tick(CPU_PROFILING);
103 }
104
105 /*
106 * Event handler for periodic ticks
107 */
tick_handle_periodic(struct clock_event_device * dev)108 void tick_handle_periodic(struct clock_event_device *dev)
109 {
110 int cpu = smp_processor_id();
111 ktime_t next = dev->next_event;
112
113 tick_periodic(cpu);
114
115 #if defined(CONFIG_HIGH_RES_TIMERS) || defined(CONFIG_NO_HZ_COMMON)
116 /*
117 * The cpu might have transitioned to HIGHRES or NOHZ mode via
118 * update_process_times() -> run_local_timers() ->
119 * hrtimer_run_queues().
120 */
121 if (dev->event_handler != tick_handle_periodic)
122 return;
123 #endif
124
125 if (!clockevent_state_oneshot(dev))
126 return;
127 for (;;) {
128 /*
129 * Setup the next period for devices, which do not have
130 * periodic mode:
131 */
132 next = ktime_add(next, tick_period);
133
134 if (!clockevents_program_event(dev, next, false))
135 return;
136 /*
137 * Have to be careful here. If we're in oneshot mode,
138 * before we call tick_periodic() in a loop, we need
139 * to be sure we're using a real hardware clocksource.
140 * Otherwise we could get trapped in an infinite
141 * loop, as the tick_periodic() increments jiffies,
142 * which then will increment time, possibly causing
143 * the loop to trigger again and again.
144 */
145 if (timekeeping_valid_for_hres())
146 tick_periodic(cpu);
147 }
148 }
149
150 /*
151 * Setup the device for a periodic tick
152 */
tick_setup_periodic(struct clock_event_device * dev,int broadcast)153 void tick_setup_periodic(struct clock_event_device *dev, int broadcast)
154 {
155 tick_set_periodic_handler(dev, broadcast);
156
157 /* Broadcast setup ? */
158 if (!tick_device_is_functional(dev))
159 return;
160
161 if ((dev->features & CLOCK_EVT_FEAT_PERIODIC) &&
162 !tick_broadcast_oneshot_active()) {
163 clockevents_switch_state(dev, CLOCK_EVT_STATE_PERIODIC);
164 } else {
165 unsigned int seq;
166 ktime_t next;
167
168 do {
169 seq = read_seqcount_begin(&jiffies_seq);
170 next = tick_next_period;
171 } while (read_seqcount_retry(&jiffies_seq, seq));
172
173 clockevents_switch_state(dev, CLOCK_EVT_STATE_ONESHOT);
174
175 for (;;) {
176 if (!clockevents_program_event(dev, next, false))
177 return;
178 next = ktime_add(next, tick_period);
179 }
180 }
181 }
182
183 #ifdef CONFIG_NO_HZ_FULL
giveup_do_timer(void * info)184 static void giveup_do_timer(void *info)
185 {
186 int cpu = *(unsigned int *)info;
187
188 WARN_ON(tick_do_timer_cpu != smp_processor_id());
189
190 tick_do_timer_cpu = cpu;
191 }
192
tick_take_do_timer_from_boot(void)193 static void tick_take_do_timer_from_boot(void)
194 {
195 int cpu = smp_processor_id();
196 int from = tick_do_timer_boot_cpu;
197
198 if (from >= 0 && from != cpu)
199 smp_call_function_single(from, giveup_do_timer, &cpu, 1);
200 }
201 #endif
202
203 /*
204 * Setup the tick device
205 */
tick_setup_device(struct tick_device * td,struct clock_event_device * newdev,int cpu,const struct cpumask * cpumask)206 static void tick_setup_device(struct tick_device *td,
207 struct clock_event_device *newdev, int cpu,
208 const struct cpumask *cpumask)
209 {
210 void (*handler)(struct clock_event_device *) = NULL;
211 ktime_t next_event = 0;
212
213 /*
214 * First device setup ?
215 */
216 if (!td->evtdev) {
217 /*
218 * If no cpu took the do_timer update, assign it to
219 * this cpu:
220 */
221 if (tick_do_timer_cpu == TICK_DO_TIMER_BOOT) {
222 tick_do_timer_cpu = cpu;
223
224 tick_next_period = ktime_get();
225 tick_period = NSEC_PER_SEC / HZ;
226 #ifdef CONFIG_NO_HZ_FULL
227 /*
228 * The boot CPU may be nohz_full, in which case set
229 * tick_do_timer_boot_cpu so the first housekeeping
230 * secondary that comes up will take do_timer from
231 * us.
232 */
233 if (tick_nohz_full_cpu(cpu))
234 tick_do_timer_boot_cpu = cpu;
235
236 } else if (tick_do_timer_boot_cpu != -1 &&
237 !tick_nohz_full_cpu(cpu)) {
238 tick_take_do_timer_from_boot();
239 tick_do_timer_boot_cpu = -1;
240 WARN_ON(tick_do_timer_cpu != cpu);
241 #endif
242 }
243
244 /*
245 * Startup in periodic mode first.
246 */
247 td->mode = TICKDEV_MODE_PERIODIC;
248 } else {
249 handler = td->evtdev->event_handler;
250 next_event = td->evtdev->next_event;
251 td->evtdev->event_handler = clockevents_handle_noop;
252 }
253
254 td->evtdev = newdev;
255
256 /*
257 * When the device is not per cpu, pin the interrupt to the
258 * current cpu:
259 */
260 if (!cpumask_equal(newdev->cpumask, cpumask))
261 irq_set_affinity(newdev->irq, cpumask);
262
263 /*
264 * When global broadcasting is active, check if the current
265 * device is registered as a placeholder for broadcast mode.
266 * This allows us to handle this x86 misfeature in a generic
267 * way. This function also returns !=0 when we keep the
268 * current active broadcast state for this CPU.
269 */
270 if (tick_device_uses_broadcast(newdev, cpu))
271 return;
272
273 if (td->mode == TICKDEV_MODE_PERIODIC)
274 tick_setup_periodic(newdev, 0);
275 else
276 tick_setup_oneshot(newdev, handler, next_event);
277 }
278
tick_install_replacement(struct clock_event_device * newdev)279 void tick_install_replacement(struct clock_event_device *newdev)
280 {
281 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
282 int cpu = smp_processor_id();
283
284 clockevents_exchange_device(td->evtdev, newdev);
285 tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
286 if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
287 tick_oneshot_notify();
288 }
289
tick_check_percpu(struct clock_event_device * curdev,struct clock_event_device * newdev,int cpu)290 static bool tick_check_percpu(struct clock_event_device *curdev,
291 struct clock_event_device *newdev, int cpu)
292 {
293 if (!cpumask_test_cpu(cpu, newdev->cpumask))
294 return false;
295 if (cpumask_equal(newdev->cpumask, cpumask_of(cpu)))
296 return true;
297 /* Check if irq affinity can be set */
298 if (newdev->irq >= 0 && !irq_can_set_affinity(newdev->irq))
299 return false;
300 /* Prefer an existing cpu local device */
301 if (curdev && cpumask_equal(curdev->cpumask, cpumask_of(cpu)))
302 return false;
303 return true;
304 }
305
tick_check_preferred(struct clock_event_device * curdev,struct clock_event_device * newdev)306 static bool tick_check_preferred(struct clock_event_device *curdev,
307 struct clock_event_device *newdev)
308 {
309 /* Prefer oneshot capable device */
310 if (!(newdev->features & CLOCK_EVT_FEAT_ONESHOT)) {
311 if (curdev && (curdev->features & CLOCK_EVT_FEAT_ONESHOT))
312 return false;
313 if (tick_oneshot_mode_active())
314 return false;
315 }
316
317 /*
318 * Use the higher rated one, but prefer a CPU local device with a lower
319 * rating than a non-CPU local device
320 */
321 return !curdev ||
322 newdev->rating > curdev->rating ||
323 !cpumask_equal(curdev->cpumask, newdev->cpumask);
324 }
325
326 /*
327 * Check whether the new device is a better fit than curdev. curdev
328 * can be NULL !
329 */
tick_check_replacement(struct clock_event_device * curdev,struct clock_event_device * newdev)330 bool tick_check_replacement(struct clock_event_device *curdev,
331 struct clock_event_device *newdev)
332 {
333 if (!tick_check_percpu(curdev, newdev, smp_processor_id()))
334 return false;
335
336 return tick_check_preferred(curdev, newdev);
337 }
338
339 /*
340 * Check, if the new registered device should be used. Called with
341 * clockevents_lock held and interrupts disabled.
342 */
tick_check_new_device(struct clock_event_device * newdev)343 void tick_check_new_device(struct clock_event_device *newdev)
344 {
345 struct clock_event_device *curdev;
346 struct tick_device *td;
347 int cpu;
348
349 cpu = smp_processor_id();
350 td = &per_cpu(tick_cpu_device, cpu);
351 curdev = td->evtdev;
352
353 /* cpu local device ? */
354 if (!tick_check_percpu(curdev, newdev, cpu))
355 goto out_bc;
356
357 /* Preference decision */
358 if (!tick_check_preferred(curdev, newdev))
359 goto out_bc;
360
361 if (!try_module_get(newdev->owner))
362 return;
363
364 /*
365 * Replace the eventually existing device by the new
366 * device. If the current device is the broadcast device, do
367 * not give it back to the clockevents layer !
368 */
369 if (tick_is_broadcast_device(curdev)) {
370 clockevents_shutdown(curdev);
371 curdev = NULL;
372 }
373 clockevents_exchange_device(curdev, newdev);
374 tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
375 if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
376 tick_oneshot_notify();
377 return;
378
379 out_bc:
380 /*
381 * Can the new device be used as a broadcast device ?
382 */
383 tick_install_broadcast_device(newdev, cpu);
384 }
385
386 /**
387 * tick_broadcast_oneshot_control - Enter/exit broadcast oneshot mode
388 * @state: The target state (enter/exit)
389 *
390 * The system enters/leaves a state, where affected devices might stop
391 * Returns 0 on success, -EBUSY if the cpu is used to broadcast wakeups.
392 *
393 * Called with interrupts disabled, so clockevents_lock is not
394 * required here because the local clock event device cannot go away
395 * under us.
396 */
tick_broadcast_oneshot_control(enum tick_broadcast_state state)397 int tick_broadcast_oneshot_control(enum tick_broadcast_state state)
398 {
399 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
400
401 if (!(td->evtdev->features & CLOCK_EVT_FEAT_C3STOP))
402 return 0;
403
404 return __tick_broadcast_oneshot_control(state);
405 }
406 EXPORT_SYMBOL_GPL(tick_broadcast_oneshot_control);
407
408 #ifdef CONFIG_HOTPLUG_CPU
409 /*
410 * Transfer the do_timer job away from a dying cpu.
411 *
412 * Called with interrupts disabled. Not locking required. If
413 * tick_do_timer_cpu is owned by this cpu, nothing can change it.
414 */
tick_handover_do_timer(void)415 void tick_handover_do_timer(void)
416 {
417 if (tick_do_timer_cpu == smp_processor_id()) {
418 int cpu = cpumask_first(cpu_online_mask);
419
420 tick_do_timer_cpu = (cpu < nr_cpu_ids) ? cpu :
421 TICK_DO_TIMER_NONE;
422 }
423 }
424
425 /*
426 * Shutdown an event device on a given cpu:
427 *
428 * This is called on a life CPU, when a CPU is dead. So we cannot
429 * access the hardware device itself.
430 * We just set the mode and remove it from the lists.
431 */
tick_shutdown(unsigned int cpu)432 void tick_shutdown(unsigned int cpu)
433 {
434 struct tick_device *td = &per_cpu(tick_cpu_device, cpu);
435 struct clock_event_device *dev = td->evtdev;
436
437 td->mode = TICKDEV_MODE_PERIODIC;
438 if (dev) {
439 /*
440 * Prevent that the clock events layer tries to call
441 * the set mode function!
442 */
443 clockevent_set_state(dev, CLOCK_EVT_STATE_DETACHED);
444 clockevents_exchange_device(dev, NULL);
445 dev->event_handler = clockevents_handle_noop;
446 td->evtdev = NULL;
447 }
448 }
449 #endif
450
451 /**
452 * tick_suspend_local - Suspend the local tick device
453 *
454 * Called from the local cpu for freeze with interrupts disabled.
455 *
456 * No locks required. Nothing can change the per cpu device.
457 */
tick_suspend_local(void)458 void tick_suspend_local(void)
459 {
460 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
461
462 clockevents_shutdown(td->evtdev);
463 }
464
465 /**
466 * tick_resume_local - Resume the local tick device
467 *
468 * Called from the local CPU for unfreeze or XEN resume magic.
469 *
470 * No locks required. Nothing can change the per cpu device.
471 */
tick_resume_local(void)472 void tick_resume_local(void)
473 {
474 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
475 bool broadcast = tick_resume_check_broadcast();
476
477 clockevents_tick_resume(td->evtdev);
478 if (!broadcast) {
479 if (td->mode == TICKDEV_MODE_PERIODIC)
480 tick_setup_periodic(td->evtdev, 0);
481 else
482 tick_resume_oneshot();
483 }
484 }
485
486 /**
487 * tick_suspend - Suspend the tick and the broadcast device
488 *
489 * Called from syscore_suspend() via timekeeping_suspend with only one
490 * CPU online and interrupts disabled or from tick_unfreeze() under
491 * tick_freeze_lock.
492 *
493 * No locks required. Nothing can change the per cpu device.
494 */
tick_suspend(void)495 void tick_suspend(void)
496 {
497 tick_suspend_local();
498 tick_suspend_broadcast();
499 }
500
501 /**
502 * tick_resume - Resume the tick and the broadcast device
503 *
504 * Called from syscore_resume() via timekeeping_resume with only one
505 * CPU online and interrupts disabled.
506 *
507 * No locks required. Nothing can change the per cpu device.
508 */
tick_resume(void)509 void tick_resume(void)
510 {
511 tick_resume_broadcast();
512 tick_resume_local();
513 }
514
515 #ifdef CONFIG_SUSPEND
516 static DEFINE_RAW_SPINLOCK(tick_freeze_lock);
517 static unsigned int tick_freeze_depth;
518
519 /**
520 * tick_freeze - Suspend the local tick and (possibly) timekeeping.
521 *
522 * Check if this is the last online CPU executing the function and if so,
523 * suspend timekeeping. Otherwise suspend the local tick.
524 *
525 * Call with interrupts disabled. Must be balanced with %tick_unfreeze().
526 * Interrupts must not be enabled before the subsequent %tick_unfreeze().
527 */
tick_freeze(void)528 void tick_freeze(void)
529 {
530 raw_spin_lock(&tick_freeze_lock);
531
532 tick_freeze_depth++;
533 if (tick_freeze_depth == num_online_cpus()) {
534 trace_suspend_resume(TPS("timekeeping_freeze"),
535 smp_processor_id(), true);
536 system_state = SYSTEM_SUSPEND;
537 sched_clock_suspend();
538 timekeeping_suspend();
539 } else {
540 tick_suspend_local();
541 }
542
543 raw_spin_unlock(&tick_freeze_lock);
544 }
545
546 /**
547 * tick_unfreeze - Resume the local tick and (possibly) timekeeping.
548 *
549 * Check if this is the first CPU executing the function and if so, resume
550 * timekeeping. Otherwise resume the local tick.
551 *
552 * Call with interrupts disabled. Must be balanced with %tick_freeze().
553 * Interrupts must not be enabled after the preceding %tick_freeze().
554 */
tick_unfreeze(void)555 void tick_unfreeze(void)
556 {
557 raw_spin_lock(&tick_freeze_lock);
558
559 if (tick_freeze_depth == num_online_cpus()) {
560 timekeeping_resume();
561 sched_clock_resume();
562 system_state = SYSTEM_RUNNING;
563 trace_suspend_resume(TPS("timekeeping_freeze"),
564 smp_processor_id(), false);
565 } else {
566 touch_softlockup_watchdog();
567 tick_resume_local();
568 }
569
570 tick_freeze_depth--;
571
572 raw_spin_unlock(&tick_freeze_lock);
573 }
574 #endif /* CONFIG_SUSPEND */
575
576 /**
577 * tick_init - initialize the tick control
578 */
tick_init(void)579 void __init tick_init(void)
580 {
581 tick_broadcast_init();
582 tick_nohz_init();
583 }
584