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