1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * sched_clock() for unstable CPU clocks
  4 *
  5 *  Copyright (C) 2008 Red Hat, Inc., Peter Zijlstra
  6 *
  7 *  Updates and enhancements:
  8 *    Copyright (C) 2008 Red Hat, Inc. Steven Rostedt <srostedt@redhat.com>
  9 *
 10 * Based on code by:
 11 *   Ingo Molnar <mingo@redhat.com>
 12 *   Guillaume Chazarain <guichaz@gmail.com>
 13 *
 14 *
 15 * What this file implements:
 16 *
 17 * cpu_clock(i) provides a fast (execution time) high resolution
 18 * clock with bounded drift between CPUs. The value of cpu_clock(i)
 19 * is monotonic for constant i. The timestamp returned is in nanoseconds.
 20 *
 21 * ######################### BIG FAT WARNING ##########################
 22 * # when comparing cpu_clock(i) to cpu_clock(j) for i != j, time can #
 23 * # go backwards !!                                                  #
 24 * ####################################################################
 25 *
 26 * There is no strict promise about the base, although it tends to start
 27 * at 0 on boot (but people really shouldn't rely on that).
 28 *
 29 * cpu_clock(i)       -- can be used from any context, including NMI.
 30 * local_clock()      -- is cpu_clock() on the current CPU.
 31 *
 32 * sched_clock_cpu(i)
 33 *
 34 * How it is implemented:
 35 *
 36 * The implementation either uses sched_clock() when
 37 * !CONFIG_HAVE_UNSTABLE_SCHED_CLOCK, which means in that case the
 38 * sched_clock() is assumed to provide these properties (mostly it means
 39 * the architecture provides a globally synchronized highres time source).
 40 *
 41 * Otherwise it tries to create a semi stable clock from a mixture of other
 42 * clocks, including:
 43 *
 44 *  - GTOD (clock monotonic)
 45 *  - sched_clock()
 46 *  - explicit idle events
 47 *
 48 * We use GTOD as base and use sched_clock() deltas to improve resolution. The
 49 * deltas are filtered to provide monotonicity and keeping it within an
 50 * expected window.
 51 *
 52 * Furthermore, explicit sleep and wakeup hooks allow us to account for time
 53 * that is otherwise invisible (TSC gets stopped).
 54 *
 55 */
 56
 57/*
 58 * Scheduler clock - returns current time in nanosec units.
 59 * This is default implementation.
 60 * Architectures and sub-architectures can override this.
 61 */
 62notrace unsigned long long __weak sched_clock(void)
 63{
 64	return (unsigned long long)(jiffies - INITIAL_JIFFIES)
 65					* (NSEC_PER_SEC / HZ);
 66}
 67EXPORT_SYMBOL_GPL(sched_clock);
 68
 69static DEFINE_STATIC_KEY_FALSE(sched_clock_running);
 70
 71#ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
 72/*
 73 * We must start with !__sched_clock_stable because the unstable -> stable
 74 * transition is accurate, while the stable -> unstable transition is not.
 75 *
 76 * Similarly we start with __sched_clock_stable_early, thereby assuming we
 77 * will become stable, such that there's only a single 1 -> 0 transition.
 78 */
 79static DEFINE_STATIC_KEY_FALSE(__sched_clock_stable);
 80static int __sched_clock_stable_early = 1;
 81
 82/*
 83 * We want: ktime_get_ns() + __gtod_offset == sched_clock() + __sched_clock_offset
 84 */
 85__read_mostly u64 __sched_clock_offset;
 86static __read_mostly u64 __gtod_offset;
 87
 88struct sched_clock_data {
 89	u64			tick_raw;
 90	u64			tick_gtod;
 91	u64			clock;
 92};
 93
 94static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
 95
 96notrace static inline struct sched_clock_data *this_scd(void)
 97{
 98	return this_cpu_ptr(&sched_clock_data);
 99}
100
101notrace static inline struct sched_clock_data *cpu_sdc(int cpu)
102{
103	return &per_cpu(sched_clock_data, cpu);
104}
105
106notrace int sched_clock_stable(void)
107{
108	return static_branch_likely(&__sched_clock_stable);
109}
110
111notrace static void __scd_stamp(struct sched_clock_data *scd)
112{
113	scd->tick_gtod = ktime_get_ns();
114	scd->tick_raw = sched_clock();
115}
116
117notrace static void __set_sched_clock_stable(void)
118{
119	struct sched_clock_data *scd;
120
121	/*
122	 * Since we're still unstable and the tick is already running, we have
123	 * to disable IRQs in order to get a consistent scd->tick* reading.
124	 */
125	local_irq_disable();
126	scd = this_scd();
127	/*
128	 * Attempt to make the (initial) unstable->stable transition continuous.
129	 */
130	__sched_clock_offset = (scd->tick_gtod + __gtod_offset) - (scd->tick_raw);
131	local_irq_enable();
132
133	printk(KERN_INFO "sched_clock: Marking stable (%lld, %lld)->(%lld, %lld)\n",
134			scd->tick_gtod, __gtod_offset,
135			scd->tick_raw,  __sched_clock_offset);
136
137	static_branch_enable(&__sched_clock_stable);
138	tick_dep_clear(TICK_DEP_BIT_CLOCK_UNSTABLE);
139}
140
141/*
142 * If we ever get here, we're screwed, because we found out -- typically after
143 * the fact -- that TSC wasn't good. This means all our clocksources (including
144 * ktime) could have reported wrong values.
145 *
146 * What we do here is an attempt to fix up and continue sort of where we left
147 * off in a coherent manner.
148 *
149 * The only way to fully avoid random clock jumps is to boot with:
150 * "tsc=unstable".
151 */
152notrace static void __sched_clock_work(struct work_struct *work)
153{
154	struct sched_clock_data *scd;
155	int cpu;
156
157	/* take a current timestamp and set 'now' */
158	preempt_disable();
159	scd = this_scd();
160	__scd_stamp(scd);
161	scd->clock = scd->tick_gtod + __gtod_offset;
162	preempt_enable();
163
164	/* clone to all CPUs */
165	for_each_possible_cpu(cpu)
166		per_cpu(sched_clock_data, cpu) = *scd;
167
168	printk(KERN_WARNING "TSC found unstable after boot, most likely due to broken BIOS. Use 'tsc=unstable'.\n");
169	printk(KERN_INFO "sched_clock: Marking unstable (%lld, %lld)<-(%lld, %lld)\n",
170			scd->tick_gtod, __gtod_offset,
171			scd->tick_raw,  __sched_clock_offset);
172
173	static_branch_disable(&__sched_clock_stable);
174}
175
176static DECLARE_WORK(sched_clock_work, __sched_clock_work);
177
178notrace static void __clear_sched_clock_stable(void)
179{
180	if (!sched_clock_stable())
181		return;
182
183	tick_dep_set(TICK_DEP_BIT_CLOCK_UNSTABLE);
184	schedule_work(&sched_clock_work);
185}
186
187notrace void clear_sched_clock_stable(void)
188{
189	__sched_clock_stable_early = 0;
190
191	smp_mb(); /* matches sched_clock_init_late() */
192
193	if (static_key_count(&sched_clock_running.key) == 2)
194		__clear_sched_clock_stable();
195}
196
197notrace static void __sched_clock_gtod_offset(void)
198{
199	struct sched_clock_data *scd = this_scd();
200
201	__scd_stamp(scd);
202	__gtod_offset = (scd->tick_raw + __sched_clock_offset) - scd->tick_gtod;
203}
204
205void __init sched_clock_init(void)
206{
207	/*
208	 * Set __gtod_offset such that once we mark sched_clock_running,
209	 * sched_clock_tick() continues where sched_clock() left off.
210	 *
211	 * Even if TSC is buggered, we're still UP at this point so it
212	 * can't really be out of sync.
213	 */
214	local_irq_disable();
215	__sched_clock_gtod_offset();
216	local_irq_enable();
217
218	static_branch_inc(&sched_clock_running);
219}
220/*
221 * We run this as late_initcall() such that it runs after all built-in drivers,
222 * notably: acpi_processor and intel_idle, which can mark the TSC as unstable.
223 */
224static int __init sched_clock_init_late(void)
225{
226	static_branch_inc(&sched_clock_running);
227	/*
228	 * Ensure that it is impossible to not do a static_key update.
229	 *
230	 * Either {set,clear}_sched_clock_stable() must see sched_clock_running
231	 * and do the update, or we must see their __sched_clock_stable_early
232	 * and do the update, or both.
233	 */
234	smp_mb(); /* matches {set,clear}_sched_clock_stable() */
235
236	if (__sched_clock_stable_early)
237		__set_sched_clock_stable();
238
239	return 0;
240}
241late_initcall(sched_clock_init_late);
242
243/*
244 * min, max except they take wrapping into account
245 */
246
247notrace static inline u64 wrap_min(u64 x, u64 y)
248{
249	return (s64)(x - y) < 0 ? x : y;
250}
251
252notrace static inline u64 wrap_max(u64 x, u64 y)
253{
254	return (s64)(x - y) > 0 ? x : y;
255}
256
257/*
258 * update the percpu scd from the raw @now value
259 *
260 *  - filter out backward motion
261 *  - use the GTOD tick value to create a window to filter crazy TSC values
262 */
263notrace static u64 sched_clock_local(struct sched_clock_data *scd)
264{
265	u64 now, clock, old_clock, min_clock, max_clock, gtod;
266	s64 delta;
267
268again:
269	now = sched_clock();
270	delta = now - scd->tick_raw;
271	if (unlikely(delta < 0))
272		delta = 0;
273
274	old_clock = scd->clock;
275
276	/*
277	 * scd->clock = clamp(scd->tick_gtod + delta,
278	 *		      max(scd->tick_gtod, scd->clock),
279	 *		      scd->tick_gtod + TICK_NSEC);
280	 */
281
282	gtod = scd->tick_gtod + __gtod_offset;
283	clock = gtod + delta;
284	min_clock = wrap_max(gtod, old_clock);
285	max_clock = wrap_max(old_clock, gtod + TICK_NSEC);
286
287	clock = wrap_max(clock, min_clock);
288	clock = wrap_min(clock, max_clock);
289
290	if (!try_cmpxchg64(&scd->clock, &old_clock, clock))
291		goto again;
292
293	return clock;
294}
295
296notrace static u64 sched_clock_remote(struct sched_clock_data *scd)
297{
298	struct sched_clock_data *my_scd = this_scd();
299	u64 this_clock, remote_clock;
300	u64 *ptr, old_val, val;
301
302#if BITS_PER_LONG != 64
303again:
304	/*
305	 * Careful here: The local and the remote clock values need to
306	 * be read out atomic as we need to compare the values and
307	 * then update either the local or the remote side. So the
308	 * cmpxchg64 below only protects one readout.
309	 *
310	 * We must reread via sched_clock_local() in the retry case on
311	 * 32-bit kernels as an NMI could use sched_clock_local() via the
312	 * tracer and hit between the readout of
313	 * the low 32-bit and the high 32-bit portion.
314	 */
315	this_clock = sched_clock_local(my_scd);
316	/*
317	 * We must enforce atomic readout on 32-bit, otherwise the
318	 * update on the remote CPU can hit inbetween the readout of
319	 * the low 32-bit and the high 32-bit portion.
320	 */
321	remote_clock = cmpxchg64(&scd->clock, 0, 0);
322#else
323	/*
324	 * On 64-bit kernels the read of [my]scd->clock is atomic versus the
325	 * update, so we can avoid the above 32-bit dance.
326	 */
327	sched_clock_local(my_scd);
328again:
329	this_clock = my_scd->clock;
330	remote_clock = scd->clock;
331#endif
332
333	/*
334	 * Use the opportunity that we have both locks
335	 * taken to couple the two clocks: we take the
336	 * larger time as the latest time for both
337	 * runqueues. (this creates monotonic movement)
338	 */
339	if (likely((s64)(remote_clock - this_clock) < 0)) {
340		ptr = &scd->clock;
341		old_val = remote_clock;
342		val = this_clock;
343	} else {
344		/*
345		 * Should be rare, but possible:
346		 */
347		ptr = &my_scd->clock;
348		old_val = this_clock;
349		val = remote_clock;
350	}
351
352	if (!try_cmpxchg64(ptr, &old_val, val))
353		goto again;
354
355	return val;
356}
357
358/*
359 * Similar to cpu_clock(), but requires local IRQs to be disabled.
360 *
361 * See cpu_clock().
362 */
363notrace u64 sched_clock_cpu(int cpu)
364{
365	struct sched_clock_data *scd;
366	u64 clock;
367
368	if (sched_clock_stable())
369		return sched_clock() + __sched_clock_offset;
370
371	if (!static_branch_likely(&sched_clock_running))
372		return sched_clock();
373
374	preempt_disable_notrace();
375	scd = cpu_sdc(cpu);
376
377	if (cpu != smp_processor_id())
378		clock = sched_clock_remote(scd);
379	else
380		clock = sched_clock_local(scd);
381	preempt_enable_notrace();
382
383	return clock;
384}
385EXPORT_SYMBOL_GPL(sched_clock_cpu);
386
387notrace void sched_clock_tick(void)
388{
389	struct sched_clock_data *scd;
390
391	if (sched_clock_stable())
392		return;
393
394	if (!static_branch_likely(&sched_clock_running))
395		return;
396
397	lockdep_assert_irqs_disabled();
398
399	scd = this_scd();
400	__scd_stamp(scd);
401	sched_clock_local(scd);
402}
403
404notrace void sched_clock_tick_stable(void)
405{
406	if (!sched_clock_stable())
407		return;
408
409	/*
410	 * Called under watchdog_lock.
411	 *
412	 * The watchdog just found this TSC to (still) be stable, so now is a
413	 * good moment to update our __gtod_offset. Because once we find the
414	 * TSC to be unstable, any computation will be computing crap.
415	 */
416	local_irq_disable();
417	__sched_clock_gtod_offset();
418	local_irq_enable();
419}
420
421/*
422 * We are going deep-idle (irqs are disabled):
423 */
424notrace void sched_clock_idle_sleep_event(void)
425{
426	sched_clock_cpu(smp_processor_id());
427}
428EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
429
430/*
431 * We just idled; resync with ktime.
432 */
433notrace void sched_clock_idle_wakeup_event(void)
434{
435	unsigned long flags;
436
437	if (sched_clock_stable())
438		return;
439
440	if (unlikely(timekeeping_suspended))
441		return;
442
443	local_irq_save(flags);
444	sched_clock_tick();
445	local_irq_restore(flags);
446}
447EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
448
449#else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
450
451void __init sched_clock_init(void)
452{
453	static_branch_inc(&sched_clock_running);
454	local_irq_disable();
455	generic_sched_clock_init();
456	local_irq_enable();
457}
458
459notrace u64 sched_clock_cpu(int cpu)
460{
461	if (!static_branch_likely(&sched_clock_running))
462		return 0;
463
464	return sched_clock();
465}
466
467#endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
468
469/*
470 * Running clock - returns the time that has elapsed while a guest has been
471 * running.
472 * On a guest this value should be local_clock minus the time the guest was
473 * suspended by the hypervisor (for any reason).
474 * On bare metal this function should return the same as local_clock.
475 * Architectures and sub-architectures can override this.
476 */
477notrace u64 __weak running_clock(void)
478{
479	return local_clock();
480}