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/compiler.h> |
11 | #include <linux/cpu.h> |
12 | #include <linux/err.h> |
13 | #include <linux/hrtimer.h> |
14 | #include <linux/interrupt.h> |
15 | #include <linux/nmi.h> |
16 | #include <linux/percpu.h> |
17 | #include <linux/profile.h> |
18 | #include <linux/sched.h> |
19 | #include <linux/module.h> |
20 | #include <trace/events/power.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. It's updated by the |
32 | * CPU which handles the tick and protected by jiffies_lock. There is |
33 | * no requirement to write hold the jiffies seqcount for it. |
34 | */ |
35 | ktime_t tick_next_period; |
36 | |
37 | /* |
38 | * tick_do_timer_cpu is a timer core internal variable which holds the CPU NR |
39 | * which is responsible for calling do_timer(), i.e. the timekeeping stuff. This |
40 | * variable has two functions: |
41 | * |
42 | * 1) Prevent a thundering herd issue of a gazillion of CPUs trying to grab the |
43 | * timekeeping lock all at once. Only the CPU which is assigned to do the |
44 | * update is handling it. |
45 | * |
46 | * 2) Hand off the duty in the NOHZ idle case by setting the value to |
47 | * TICK_DO_TIMER_NONE, i.e. a non existing CPU. So the next cpu which looks |
48 | * at it will take over and keep the time keeping alive. The handover |
49 | * procedure also covers cpu hotplug. |
50 | */ |
51 | int tick_do_timer_cpu __read_mostly = TICK_DO_TIMER_BOOT; |
52 | #ifdef CONFIG_NO_HZ_FULL |
53 | /* |
54 | * tick_do_timer_boot_cpu indicates the boot CPU temporarily owns |
55 | * tick_do_timer_cpu and it should be taken over by an eligible secondary |
56 | * when one comes online. |
57 | */ |
58 | static int tick_do_timer_boot_cpu __read_mostly = -1; |
59 | #endif |
60 | |
61 | /* |
62 | * Debugging: see timer_list.c |
63 | */ |
64 | struct tick_device *tick_get_device(int cpu) |
65 | { |
66 | return &per_cpu(tick_cpu_device, cpu); |
67 | } |
68 | |
69 | /** |
70 | * tick_is_oneshot_available - check for a oneshot capable event device |
71 | */ |
72 | int tick_is_oneshot_available(void) |
73 | { |
74 | struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev); |
75 | |
76 | if (!dev || !(dev->features & CLOCK_EVT_FEAT_ONESHOT)) |
77 | return 0; |
78 | if (!(dev->features & CLOCK_EVT_FEAT_C3STOP)) |
79 | return 1; |
80 | return tick_broadcast_oneshot_available(); |
81 | } |
82 | |
83 | /* |
84 | * Periodic tick |
85 | */ |
86 | static void tick_periodic(int cpu) |
87 | { |
88 | if (READ_ONCE(tick_do_timer_cpu) == cpu) { |
89 | raw_spin_lock(&jiffies_lock); |
90 | write_seqcount_begin(&jiffies_seq); |
91 | |
92 | /* Keep track of the next tick event */ |
93 | tick_next_period = ktime_add_ns(tick_next_period, TICK_NSEC); |
94 | |
95 | do_timer(ticks: 1); |
96 | write_seqcount_end(&jiffies_seq); |
97 | raw_spin_unlock(&jiffies_lock); |
98 | update_wall_time(); |
99 | } |
100 | |
101 | update_process_times(user: user_mode(regs: get_irq_regs())); |
102 | profile_tick(CPU_PROFILING); |
103 | } |
104 | |
105 | /* |
106 | * Event handler for periodic ticks |
107 | */ |
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 | /* |
116 | * The cpu might have transitioned to HIGHRES or NOHZ mode via |
117 | * update_process_times() -> run_local_timers() -> |
118 | * hrtimer_run_queues(). |
119 | */ |
120 | if (IS_ENABLED(CONFIG_TICK_ONESHOT) && dev->event_handler != tick_handle_periodic) |
121 | return; |
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_ns(next, TICK_NSEC); |
131 | |
132 | if (!clockevents_program_event(dev, expires: next, force: 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 | */ |
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, state: 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, state: CLOCK_EVT_STATE_ONESHOT); |
172 | |
173 | for (;;) { |
174 | if (!clockevents_program_event(dev, expires: next, force: false)) |
175 | return; |
176 | next = ktime_add_ns(next, TICK_NSEC); |
177 | } |
178 | } |
179 | } |
180 | |
181 | #ifdef CONFIG_NO_HZ_FULL |
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 | |
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 | */ |
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 (READ_ONCE(tick_do_timer_cpu) == TICK_DO_TIMER_BOOT) { |
220 | WRITE_ONCE(tick_do_timer_cpu, cpu); |
221 | tick_next_period = ktime_get(); |
222 | #ifdef CONFIG_NO_HZ_FULL |
223 | /* |
224 | * The boot CPU may be nohz_full, in which case set |
225 | * tick_do_timer_boot_cpu so the first housekeeping |
226 | * secondary that comes up will take do_timer from |
227 | * us. |
228 | */ |
229 | if (tick_nohz_full_cpu(cpu)) |
230 | tick_do_timer_boot_cpu = cpu; |
231 | |
232 | } else if (tick_do_timer_boot_cpu != -1 && |
233 | !tick_nohz_full_cpu(cpu)) { |
234 | tick_take_do_timer_from_boot(); |
235 | tick_do_timer_boot_cpu = -1; |
236 | WARN_ON(READ_ONCE(tick_do_timer_cpu) != cpu); |
237 | #endif |
238 | } |
239 | |
240 | /* |
241 | * Startup in periodic mode first. |
242 | */ |
243 | td->mode = TICKDEV_MODE_PERIODIC; |
244 | } else { |
245 | handler = td->evtdev->event_handler; |
246 | next_event = td->evtdev->next_event; |
247 | td->evtdev->event_handler = clockevents_handle_noop; |
248 | } |
249 | |
250 | td->evtdev = newdev; |
251 | |
252 | /* |
253 | * When the device is not per cpu, pin the interrupt to the |
254 | * current cpu: |
255 | */ |
256 | if (!cpumask_equal(src1p: newdev->cpumask, src2p: cpumask)) |
257 | irq_set_affinity(irq: newdev->irq, cpumask); |
258 | |
259 | /* |
260 | * When global broadcasting is active, check if the current |
261 | * device is registered as a placeholder for broadcast mode. |
262 | * This allows us to handle this x86 misfeature in a generic |
263 | * way. This function also returns !=0 when we keep the |
264 | * current active broadcast state for this CPU. |
265 | */ |
266 | if (tick_device_uses_broadcast(dev: newdev, cpu)) |
267 | return; |
268 | |
269 | if (td->mode == TICKDEV_MODE_PERIODIC) |
270 | tick_setup_periodic(dev: newdev, broadcast: 0); |
271 | else |
272 | tick_setup_oneshot(newdev, handler, nextevt: next_event); |
273 | } |
274 | |
275 | void tick_install_replacement(struct clock_event_device *newdev) |
276 | { |
277 | struct tick_device *td = this_cpu_ptr(&tick_cpu_device); |
278 | int cpu = smp_processor_id(); |
279 | |
280 | clockevents_exchange_device(old: td->evtdev, new: newdev); |
281 | tick_setup_device(td, newdev, cpu, cpumask_of(cpu)); |
282 | if (newdev->features & CLOCK_EVT_FEAT_ONESHOT) |
283 | tick_oneshot_notify(); |
284 | } |
285 | |
286 | static bool tick_check_percpu(struct clock_event_device *curdev, |
287 | struct clock_event_device *newdev, int cpu) |
288 | { |
289 | if (!cpumask_test_cpu(cpu, cpumask: newdev->cpumask)) |
290 | return false; |
291 | if (cpumask_equal(src1p: newdev->cpumask, cpumask_of(cpu))) |
292 | return true; |
293 | /* Check if irq affinity can be set */ |
294 | if (newdev->irq >= 0 && !irq_can_set_affinity(irq: newdev->irq)) |
295 | return false; |
296 | /* Prefer an existing cpu local device */ |
297 | if (curdev && cpumask_equal(src1p: curdev->cpumask, cpumask_of(cpu))) |
298 | return false; |
299 | return true; |
300 | } |
301 | |
302 | static bool tick_check_preferred(struct clock_event_device *curdev, |
303 | struct clock_event_device *newdev) |
304 | { |
305 | /* Prefer oneshot capable device */ |
306 | if (!(newdev->features & CLOCK_EVT_FEAT_ONESHOT)) { |
307 | if (curdev && (curdev->features & CLOCK_EVT_FEAT_ONESHOT)) |
308 | return false; |
309 | if (tick_oneshot_mode_active()) |
310 | return false; |
311 | } |
312 | |
313 | /* |
314 | * Use the higher rated one, but prefer a CPU local device with a lower |
315 | * rating than a non-CPU local device |
316 | */ |
317 | return !curdev || |
318 | newdev->rating > curdev->rating || |
319 | !cpumask_equal(src1p: curdev->cpumask, src2p: newdev->cpumask); |
320 | } |
321 | |
322 | /* |
323 | * Check whether the new device is a better fit than curdev. curdev |
324 | * can be NULL ! |
325 | */ |
326 | bool tick_check_replacement(struct clock_event_device *curdev, |
327 | struct clock_event_device *newdev) |
328 | { |
329 | if (!tick_check_percpu(curdev, newdev, smp_processor_id())) |
330 | return false; |
331 | |
332 | return tick_check_preferred(curdev, newdev); |
333 | } |
334 | |
335 | /* |
336 | * Check, if the new registered device should be used. Called with |
337 | * clockevents_lock held and interrupts disabled. |
338 | */ |
339 | void tick_check_new_device(struct clock_event_device *newdev) |
340 | { |
341 | struct clock_event_device *curdev; |
342 | struct tick_device *td; |
343 | int cpu; |
344 | |
345 | cpu = smp_processor_id(); |
346 | td = &per_cpu(tick_cpu_device, cpu); |
347 | curdev = td->evtdev; |
348 | |
349 | if (!tick_check_replacement(curdev, newdev)) |
350 | goto out_bc; |
351 | |
352 | if (!try_module_get(module: newdev->owner)) |
353 | return; |
354 | |
355 | /* |
356 | * Replace the eventually existing device by the new |
357 | * device. If the current device is the broadcast device, do |
358 | * not give it back to the clockevents layer ! |
359 | */ |
360 | if (tick_is_broadcast_device(dev: curdev)) { |
361 | clockevents_shutdown(dev: curdev); |
362 | curdev = NULL; |
363 | } |
364 | clockevents_exchange_device(old: curdev, new: newdev); |
365 | tick_setup_device(td, newdev, cpu, cpumask_of(cpu)); |
366 | if (newdev->features & CLOCK_EVT_FEAT_ONESHOT) |
367 | tick_oneshot_notify(); |
368 | return; |
369 | |
370 | out_bc: |
371 | /* |
372 | * Can the new device be used as a broadcast device ? |
373 | */ |
374 | tick_install_broadcast_device(dev: newdev, cpu); |
375 | } |
376 | |
377 | /** |
378 | * tick_broadcast_oneshot_control - Enter/exit broadcast oneshot mode |
379 | * @state: The target state (enter/exit) |
380 | * |
381 | * The system enters/leaves a state, where affected devices might stop |
382 | * Returns 0 on success, -EBUSY if the cpu is used to broadcast wakeups. |
383 | * |
384 | * Called with interrupts disabled, so clockevents_lock is not |
385 | * required here because the local clock event device cannot go away |
386 | * under us. |
387 | */ |
388 | int tick_broadcast_oneshot_control(enum tick_broadcast_state state) |
389 | { |
390 | struct tick_device *td = this_cpu_ptr(&tick_cpu_device); |
391 | |
392 | if (!(td->evtdev->features & CLOCK_EVT_FEAT_C3STOP)) |
393 | return 0; |
394 | |
395 | return __tick_broadcast_oneshot_control(state); |
396 | } |
397 | EXPORT_SYMBOL_GPL(tick_broadcast_oneshot_control); |
398 | |
399 | #ifdef CONFIG_HOTPLUG_CPU |
400 | void tick_assert_timekeeping_handover(void) |
401 | { |
402 | WARN_ON_ONCE(tick_do_timer_cpu == smp_processor_id()); |
403 | } |
404 | /* |
405 | * Stop the tick and transfer the timekeeping job away from a dying cpu. |
406 | */ |
407 | int tick_cpu_dying(unsigned int dying_cpu) |
408 | { |
409 | /* |
410 | * If the current CPU is the timekeeper, it's the only one that can |
411 | * safely hand over its duty. Also all online CPUs are in stop |
412 | * machine, guaranteed not to be idle, therefore there is no |
413 | * concurrency and it's safe to pick any online successor. |
414 | */ |
415 | if (tick_do_timer_cpu == dying_cpu) |
416 | tick_do_timer_cpu = cpumask_first(cpu_online_mask); |
417 | |
418 | /* Make sure the CPU won't try to retake the timekeeping duty */ |
419 | tick_sched_timer_dying(cpu: dying_cpu); |
420 | |
421 | /* Remove CPU from timer broadcasting */ |
422 | tick_offline_cpu(cpu: dying_cpu); |
423 | |
424 | return 0; |
425 | } |
426 | |
427 | /* |
428 | * Shutdown an event device on a given cpu: |
429 | * |
430 | * This is called on a life CPU, when a CPU is dead. So we cannot |
431 | * access the hardware device itself. |
432 | * We just set the mode and remove it from the lists. |
433 | */ |
434 | void tick_shutdown(unsigned int cpu) |
435 | { |
436 | struct tick_device *td = &per_cpu(tick_cpu_device, cpu); |
437 | struct clock_event_device *dev = td->evtdev; |
438 | |
439 | td->mode = TICKDEV_MODE_PERIODIC; |
440 | if (dev) { |
441 | /* |
442 | * Prevent that the clock events layer tries to call |
443 | * the set mode function! |
444 | */ |
445 | clockevent_set_state(dev, state: CLOCK_EVT_STATE_DETACHED); |
446 | clockevents_exchange_device(old: dev, NULL); |
447 | dev->event_handler = clockevents_handle_noop; |
448 | td->evtdev = NULL; |
449 | } |
450 | } |
451 | #endif |
452 | |
453 | /** |
454 | * tick_suspend_local - Suspend the local tick device |
455 | * |
456 | * Called from the local cpu for freeze with interrupts disabled. |
457 | * |
458 | * No locks required. Nothing can change the per cpu device. |
459 | */ |
460 | void tick_suspend_local(void) |
461 | { |
462 | struct tick_device *td = this_cpu_ptr(&tick_cpu_device); |
463 | |
464 | clockevents_shutdown(dev: td->evtdev); |
465 | } |
466 | |
467 | /** |
468 | * tick_resume_local - Resume the local tick device |
469 | * |
470 | * Called from the local CPU for unfreeze or XEN resume magic. |
471 | * |
472 | * No locks required. Nothing can change the per cpu device. |
473 | */ |
474 | void tick_resume_local(void) |
475 | { |
476 | struct tick_device *td = this_cpu_ptr(&tick_cpu_device); |
477 | bool broadcast = tick_resume_check_broadcast(); |
478 | |
479 | clockevents_tick_resume(dev: td->evtdev); |
480 | if (!broadcast) { |
481 | if (td->mode == TICKDEV_MODE_PERIODIC) |
482 | tick_setup_periodic(dev: td->evtdev, broadcast: 0); |
483 | else |
484 | tick_resume_oneshot(); |
485 | } |
486 | |
487 | /* |
488 | * Ensure that hrtimers are up to date and the clockevents device |
489 | * is reprogrammed correctly when high resolution timers are |
490 | * enabled. |
491 | */ |
492 | hrtimers_resume_local(); |
493 | } |
494 | |
495 | /** |
496 | * tick_suspend - Suspend the tick and the broadcast device |
497 | * |
498 | * Called from syscore_suspend() via timekeeping_suspend with only one |
499 | * CPU online and interrupts disabled or from tick_unfreeze() under |
500 | * tick_freeze_lock. |
501 | * |
502 | * No locks required. Nothing can change the per cpu device. |
503 | */ |
504 | void tick_suspend(void) |
505 | { |
506 | tick_suspend_local(); |
507 | tick_suspend_broadcast(); |
508 | } |
509 | |
510 | /** |
511 | * tick_resume - Resume the tick and the broadcast device |
512 | * |
513 | * Called from syscore_resume() via timekeeping_resume with only one |
514 | * CPU online and interrupts disabled. |
515 | * |
516 | * No locks required. Nothing can change the per cpu device. |
517 | */ |
518 | void tick_resume(void) |
519 | { |
520 | tick_resume_broadcast(); |
521 | tick_resume_local(); |
522 | } |
523 | |
524 | #ifdef CONFIG_SUSPEND |
525 | static DEFINE_RAW_SPINLOCK(tick_freeze_lock); |
526 | static unsigned int tick_freeze_depth; |
527 | |
528 | /** |
529 | * tick_freeze - Suspend the local tick and (possibly) timekeeping. |
530 | * |
531 | * Check if this is the last online CPU executing the function and if so, |
532 | * suspend timekeeping. Otherwise suspend the local tick. |
533 | * |
534 | * Call with interrupts disabled. Must be balanced with %tick_unfreeze(). |
535 | * Interrupts must not be enabled before the subsequent %tick_unfreeze(). |
536 | */ |
537 | void tick_freeze(void) |
538 | { |
539 | raw_spin_lock(&tick_freeze_lock); |
540 | |
541 | tick_freeze_depth++; |
542 | if (tick_freeze_depth == num_online_cpus()) { |
543 | trace_suspend_resume(TPS("timekeeping_freeze" ), |
544 | smp_processor_id(), start: true); |
545 | system_state = SYSTEM_SUSPEND; |
546 | sched_clock_suspend(); |
547 | timekeeping_suspend(); |
548 | } else { |
549 | tick_suspend_local(); |
550 | } |
551 | |
552 | raw_spin_unlock(&tick_freeze_lock); |
553 | } |
554 | |
555 | /** |
556 | * tick_unfreeze - Resume the local tick and (possibly) timekeeping. |
557 | * |
558 | * Check if this is the first CPU executing the function and if so, resume |
559 | * timekeeping. Otherwise resume the local tick. |
560 | * |
561 | * Call with interrupts disabled. Must be balanced with %tick_freeze(). |
562 | * Interrupts must not be enabled after the preceding %tick_freeze(). |
563 | */ |
564 | void tick_unfreeze(void) |
565 | { |
566 | raw_spin_lock(&tick_freeze_lock); |
567 | |
568 | if (tick_freeze_depth == num_online_cpus()) { |
569 | timekeeping_resume(); |
570 | sched_clock_resume(); |
571 | system_state = SYSTEM_RUNNING; |
572 | trace_suspend_resume(TPS("timekeeping_freeze" ), |
573 | smp_processor_id(), start: false); |
574 | } else { |
575 | touch_softlockup_watchdog(); |
576 | tick_resume_local(); |
577 | } |
578 | |
579 | tick_freeze_depth--; |
580 | |
581 | raw_spin_unlock(&tick_freeze_lock); |
582 | } |
583 | #endif /* CONFIG_SUSPEND */ |
584 | |
585 | /** |
586 | * tick_init - initialize the tick control |
587 | */ |
588 | void __init tick_init(void) |
589 | { |
590 | tick_broadcast_init(); |
591 | tick_nohz_init(); |
592 | } |
593 | |