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. It's updated by the
31  * CPU which handles the tick and protected by jiffies_lock. There is
32  * no requirement to write hold the jiffies seqcount for it.
33  */
34 ktime_t tick_next_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_ns(tick_next_period, TICK_NSEC);
93 
94 		do_timer(1);
95 		write_seqcount_end(&jiffies_seq);
96 		raw_spin_unlock(&jiffies_lock);
97 		update_wall_time();
98 	}
99 
100 	update_process_times(user_mode(get_irq_regs()));
101 	profile_tick(CPU_PROFILING);
102 }
103 
104 /*
105  * Event handler for periodic ticks
106  */
tick_handle_periodic(struct clock_event_device * dev)107 void tick_handle_periodic(struct clock_event_device *dev)
108 {
109 	int cpu = smp_processor_id();
110 	ktime_t next = dev->next_event;
111 
112 	tick_periodic(cpu);
113 
114 #if defined(CONFIG_HIGH_RES_TIMERS) || defined(CONFIG_NO_HZ_COMMON)
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 (dev->event_handler != tick_handle_periodic)
121 		return;
122 #endif
123 
124 	if (!clockevent_state_oneshot(dev))
125 		return;
126 	for (;;) {
127 		/*
128 		 * Setup the next period for devices, which do not have
129 		 * periodic mode:
130 		 */
131 		next = ktime_add_ns(next, TICK_NSEC);
132 
133 		if (!clockevents_program_event(dev, next, false))
134 			return;
135 		/*
136 		 * Have to be careful here. If we're in oneshot mode,
137 		 * before we call tick_periodic() in a loop, we need
138 		 * to be sure we're using a real hardware clocksource.
139 		 * Otherwise we could get trapped in an infinite
140 		 * loop, as the tick_periodic() increments jiffies,
141 		 * which then will increment time, possibly causing
142 		 * the loop to trigger again and again.
143 		 */
144 		if (timekeeping_valid_for_hres())
145 			tick_periodic(cpu);
146 	}
147 }
148 
149 /*
150  * Setup the device for a periodic tick
151  */
tick_setup_periodic(struct clock_event_device * dev,int broadcast)152 void tick_setup_periodic(struct clock_event_device *dev, int broadcast)
153 {
154 	tick_set_periodic_handler(dev, broadcast);
155 
156 	/* Broadcast setup ? */
157 	if (!tick_device_is_functional(dev))
158 		return;
159 
160 	if ((dev->features & CLOCK_EVT_FEAT_PERIODIC) &&
161 	    !tick_broadcast_oneshot_active()) {
162 		clockevents_switch_state(dev, CLOCK_EVT_STATE_PERIODIC);
163 	} else {
164 		unsigned int seq;
165 		ktime_t next;
166 
167 		do {
168 			seq = read_seqcount_begin(&jiffies_seq);
169 			next = tick_next_period;
170 		} while (read_seqcount_retry(&jiffies_seq, seq));
171 
172 		clockevents_switch_state(dev, CLOCK_EVT_STATE_ONESHOT);
173 
174 		for (;;) {
175 			if (!clockevents_program_event(dev, next, false))
176 				return;
177 			next = ktime_add_ns(next, TICK_NSEC);
178 		}
179 	}
180 }
181 
182 #ifdef CONFIG_NO_HZ_FULL
giveup_do_timer(void * info)183 static void giveup_do_timer(void *info)
184 {
185 	int cpu = *(unsigned int *)info;
186 
187 	WARN_ON(tick_do_timer_cpu != smp_processor_id());
188 
189 	tick_do_timer_cpu = cpu;
190 }
191 
tick_take_do_timer_from_boot(void)192 static void tick_take_do_timer_from_boot(void)
193 {
194 	int cpu = smp_processor_id();
195 	int from = tick_do_timer_boot_cpu;
196 
197 	if (from >= 0 && from != cpu)
198 		smp_call_function_single(from, giveup_do_timer, &cpu, 1);
199 }
200 #endif
201 
202 /*
203  * Setup the tick device
204  */
tick_setup_device(struct tick_device * td,struct clock_event_device * newdev,int cpu,const struct cpumask * cpumask)205 static void tick_setup_device(struct tick_device *td,
206 			      struct clock_event_device *newdev, int cpu,
207 			      const struct cpumask *cpumask)
208 {
209 	void (*handler)(struct clock_event_device *) = NULL;
210 	ktime_t next_event = 0;
211 
212 	/*
213 	 * First device setup ?
214 	 */
215 	if (!td->evtdev) {
216 		/*
217 		 * If no cpu took the do_timer update, assign it to
218 		 * this cpu:
219 		 */
220 		if (tick_do_timer_cpu == TICK_DO_TIMER_BOOT) {
221 			tick_do_timer_cpu = cpu;
222 
223 			tick_next_period = ktime_get();
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 	if (!tick_check_replacement(curdev, newdev))
352 		goto out_bc;
353 
354 	if (!try_module_get(newdev->owner))
355 		return;
356 
357 	/*
358 	 * Replace the eventually existing device by the new
359 	 * device. If the current device is the broadcast device, do
360 	 * not give it back to the clockevents layer !
361 	 */
362 	if (tick_is_broadcast_device(curdev)) {
363 		clockevents_shutdown(curdev);
364 		curdev = NULL;
365 	}
366 	clockevents_exchange_device(curdev, newdev);
367 	tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
368 	if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
369 		tick_oneshot_notify();
370 	return;
371 
372 out_bc:
373 	/*
374 	 * Can the new device be used as a broadcast device ?
375 	 */
376 	tick_install_broadcast_device(newdev, cpu);
377 }
378 
379 /**
380  * tick_broadcast_oneshot_control - Enter/exit broadcast oneshot mode
381  * @state:	The target state (enter/exit)
382  *
383  * The system enters/leaves a state, where affected devices might stop
384  * Returns 0 on success, -EBUSY if the cpu is used to broadcast wakeups.
385  *
386  * Called with interrupts disabled, so clockevents_lock is not
387  * required here because the local clock event device cannot go away
388  * under us.
389  */
tick_broadcast_oneshot_control(enum tick_broadcast_state state)390 int tick_broadcast_oneshot_control(enum tick_broadcast_state state)
391 {
392 	struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
393 
394 	if (!(td->evtdev->features & CLOCK_EVT_FEAT_C3STOP))
395 		return 0;
396 
397 	return __tick_broadcast_oneshot_control(state);
398 }
399 EXPORT_SYMBOL_GPL(tick_broadcast_oneshot_control);
400 
401 #ifdef CONFIG_HOTPLUG_CPU
402 /*
403  * Transfer the do_timer job away from a dying cpu.
404  *
405  * Called with interrupts disabled. No locking required. If
406  * tick_do_timer_cpu is owned by this cpu, nothing can change it.
407  */
tick_handover_do_timer(void)408 void tick_handover_do_timer(void)
409 {
410 	if (tick_do_timer_cpu == smp_processor_id())
411 		tick_do_timer_cpu = cpumask_first(cpu_online_mask);
412 }
413 
414 /*
415  * Shutdown an event device on a given cpu:
416  *
417  * This is called on a life CPU, when a CPU is dead. So we cannot
418  * access the hardware device itself.
419  * We just set the mode and remove it from the lists.
420  */
tick_shutdown(unsigned int cpu)421 void tick_shutdown(unsigned int cpu)
422 {
423 	struct tick_device *td = &per_cpu(tick_cpu_device, cpu);
424 	struct clock_event_device *dev = td->evtdev;
425 
426 	td->mode = TICKDEV_MODE_PERIODIC;
427 	if (dev) {
428 		/*
429 		 * Prevent that the clock events layer tries to call
430 		 * the set mode function!
431 		 */
432 		clockevent_set_state(dev, CLOCK_EVT_STATE_DETACHED);
433 		clockevents_exchange_device(dev, NULL);
434 		dev->event_handler = clockevents_handle_noop;
435 		td->evtdev = NULL;
436 	}
437 }
438 #endif
439 
440 /**
441  * tick_suspend_local - Suspend the local tick device
442  *
443  * Called from the local cpu for freeze with interrupts disabled.
444  *
445  * No locks required. Nothing can change the per cpu device.
446  */
tick_suspend_local(void)447 void tick_suspend_local(void)
448 {
449 	struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
450 
451 	clockevents_shutdown(td->evtdev);
452 }
453 
454 /**
455  * tick_resume_local - Resume the local tick device
456  *
457  * Called from the local CPU for unfreeze or XEN resume magic.
458  *
459  * No locks required. Nothing can change the per cpu device.
460  */
tick_resume_local(void)461 void tick_resume_local(void)
462 {
463 	struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
464 	bool broadcast = tick_resume_check_broadcast();
465 
466 	clockevents_tick_resume(td->evtdev);
467 	if (!broadcast) {
468 		if (td->mode == TICKDEV_MODE_PERIODIC)
469 			tick_setup_periodic(td->evtdev, 0);
470 		else
471 			tick_resume_oneshot();
472 	}
473 
474 	/*
475 	 * Ensure that hrtimers are up to date and the clockevents device
476 	 * is reprogrammed correctly when high resolution timers are
477 	 * enabled.
478 	 */
479 	hrtimers_resume_local();
480 }
481 
482 /**
483  * tick_suspend - Suspend the tick and the broadcast device
484  *
485  * Called from syscore_suspend() via timekeeping_suspend with only one
486  * CPU online and interrupts disabled or from tick_unfreeze() under
487  * tick_freeze_lock.
488  *
489  * No locks required. Nothing can change the per cpu device.
490  */
tick_suspend(void)491 void tick_suspend(void)
492 {
493 	tick_suspend_local();
494 	tick_suspend_broadcast();
495 }
496 
497 /**
498  * tick_resume - Resume the tick and the broadcast device
499  *
500  * Called from syscore_resume() via timekeeping_resume with only one
501  * CPU online and interrupts disabled.
502  *
503  * No locks required. Nothing can change the per cpu device.
504  */
tick_resume(void)505 void tick_resume(void)
506 {
507 	tick_resume_broadcast();
508 	tick_resume_local();
509 }
510 
511 #ifdef CONFIG_SUSPEND
512 static DEFINE_RAW_SPINLOCK(tick_freeze_lock);
513 static unsigned int tick_freeze_depth;
514 
515 /**
516  * tick_freeze - Suspend the local tick and (possibly) timekeeping.
517  *
518  * Check if this is the last online CPU executing the function and if so,
519  * suspend timekeeping.  Otherwise suspend the local tick.
520  *
521  * Call with interrupts disabled.  Must be balanced with %tick_unfreeze().
522  * Interrupts must not be enabled before the subsequent %tick_unfreeze().
523  */
tick_freeze(void)524 void tick_freeze(void)
525 {
526 	raw_spin_lock(&tick_freeze_lock);
527 
528 	tick_freeze_depth++;
529 	if (tick_freeze_depth == num_online_cpus()) {
530 		trace_suspend_resume(TPS("timekeeping_freeze"),
531 				     smp_processor_id(), true);
532 		system_state = SYSTEM_SUSPEND;
533 		sched_clock_suspend();
534 		timekeeping_suspend();
535 	} else {
536 		tick_suspend_local();
537 	}
538 
539 	raw_spin_unlock(&tick_freeze_lock);
540 }
541 
542 /**
543  * tick_unfreeze - Resume the local tick and (possibly) timekeeping.
544  *
545  * Check if this is the first CPU executing the function and if so, resume
546  * timekeeping.  Otherwise resume the local tick.
547  *
548  * Call with interrupts disabled.  Must be balanced with %tick_freeze().
549  * Interrupts must not be enabled after the preceding %tick_freeze().
550  */
tick_unfreeze(void)551 void tick_unfreeze(void)
552 {
553 	raw_spin_lock(&tick_freeze_lock);
554 
555 	if (tick_freeze_depth == num_online_cpus()) {
556 		timekeeping_resume();
557 		sched_clock_resume();
558 		system_state = SYSTEM_RUNNING;
559 		trace_suspend_resume(TPS("timekeeping_freeze"),
560 				     smp_processor_id(), false);
561 	} else {
562 		touch_softlockup_watchdog();
563 		tick_resume_local();
564 	}
565 
566 	tick_freeze_depth--;
567 
568 	raw_spin_unlock(&tick_freeze_lock);
569 }
570 #endif /* CONFIG_SUSPEND */
571 
572 /**
573  * tick_init - initialize the tick control
574  */
tick_init(void)575 void __init tick_init(void)
576 {
577 	tick_broadcast_init();
578 	tick_nohz_init();
579 }
580