clock.c

  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#include <linux/sched/clock.h>
 58#include "sched.h"
 59
 60/*
 61 * Scheduler clock - returns current time in nanosec units.
 62 * This is default implementation.
 63 * Architectures and sub-architectures can override this.
 64 */
 65notrace unsigned long long __weak sched_clock(void)
 66{
 67	return (unsigned long long)(jiffies - INITIAL_JIFFIES)
 68					* (NSEC_PER_SEC / HZ);
 69}
 70EXPORT_SYMBOL_GPL(sched_clock);
 71
 72static DEFINE_STATIC_KEY_FALSE(sched_clock_running);
 73
 74#ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
 75/*
 76 * We must start with !__sched_clock_stable because the unstable -> stable
 77 * transition is accurate, while the stable -> unstable transition is not.
 78 *
 79 * Similarly we start with __sched_clock_stable_early, thereby assuming we
 80 * will become stable, such that there's only a single 1 -> 0 transition.
 81 */
 82static DEFINE_STATIC_KEY_FALSE(__sched_clock_stable);
 83static int __sched_clock_stable_early = 1;
 84
 85/*
 86 * We want: ktime_get_ns() + __gtod_offset == sched_clock() + __sched_clock_offset
 87 */
 88__read_mostly u64 __sched_clock_offset;
 89static __read_mostly u64 __gtod_offset;
 90
 91struct sched_clock_data {
 92	u64			tick_raw;
 93	u64			tick_gtod;
 94	u64			clock;
 95};
 96
 97static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
 98
 99static __always_inline struct sched_clock_data *this_scd(void)
100{
101	return this_cpu_ptr(&sched_clock_data);
102}
103
104notrace static inline struct sched_clock_data *cpu_sdc(int cpu)
105{
106	return &per_cpu(sched_clock_data, cpu);
107}
108
109notrace int sched_clock_stable(void)
110{
111	return static_branch_likely(&__sched_clock_stable);
112}
113
114notrace static void __scd_stamp(struct sched_clock_data *scd)
115{
116	scd->tick_gtod = ktime_get_ns();
117	scd->tick_raw = sched_clock();
118}
119
120notrace static void __set_sched_clock_stable(void)
121{
122	struct sched_clock_data *scd;
123
124	/*
125	 * Since we're still unstable and the tick is already running, we have
126	 * to disable IRQs in order to get a consistent scd->tick* reading.
127	 */
128	local_irq_disable();
129	scd = this_scd();
130	/*
131	 * Attempt to make the (initial) unstable->stable transition continuous.
132	 */
133	__sched_clock_offset = (scd->tick_gtod + __gtod_offset) - (scd->tick_raw);
134	local_irq_enable();
135
136	printk(KERN_INFO "sched_clock: Marking stable (%lld, %lld)->(%lld, %lld)\n",
137			scd->tick_gtod, __gtod_offset,
138			scd->tick_raw,  __sched_clock_offset);
139
140	static_branch_enable(&__sched_clock_stable);
141	tick_dep_clear(TICK_DEP_BIT_CLOCK_UNSTABLE);
142}
143
144/*
145 * If we ever get here, we're screwed, because we found out -- typically after
146 * the fact -- that TSC wasn't good. This means all our clocksources (including
147 * ktime) could have reported wrong values.
148 *
149 * What we do here is an attempt to fix up and continue sort of where we left
150 * off in a coherent manner.
151 *
152 * The only way to fully avoid random clock jumps is to boot with:
153 * "tsc=unstable".
154 */
155notrace static void __sched_clock_work(struct work_struct *work)
156{
157	struct sched_clock_data *scd;
158	int cpu;
159
160	/* take a current timestamp and set 'now' */
161	preempt_disable();
162	scd = this_scd();
163	__scd_stamp(scd);
164	scd->clock = scd->tick_gtod + __gtod_offset;
165	preempt_enable();
166
167	/* clone to all CPUs */
168	for_each_possible_cpu(cpu)
169		per_cpu(sched_clock_data, cpu) = *scd;
170
171	printk(KERN_WARNING "TSC found unstable after boot, most likely due to broken BIOS. Use 'tsc=unstable'.\n");
172	printk(KERN_INFO "sched_clock: Marking unstable (%lld, %lld)<-(%lld, %lld)\n",
173			scd->tick_gtod, __gtod_offset,
174			scd->tick_raw,  __sched_clock_offset);
175
176	static_branch_disable(&__sched_clock_stable);
177}
178
179static DECLARE_WORK(sched_clock_work, __sched_clock_work);
180
181notrace static void __clear_sched_clock_stable(void)
182{
183	if (!sched_clock_stable())
184		return;
185
186	tick_dep_set(TICK_DEP_BIT_CLOCK_UNSTABLE);
187	schedule_work(&sched_clock_work);
188}
189
190notrace void clear_sched_clock_stable(void)
191{
192	__sched_clock_stable_early = 0;
193
194	smp_mb(); /* matches sched_clock_init_late() */
195
196	if (static_key_count(&sched_clock_running.key) == 2)
197		__clear_sched_clock_stable();
198}
199
200notrace static void __sched_clock_gtod_offset(void)
201{
202	struct sched_clock_data *scd = this_scd();
203
204	__scd_stamp(scd);
205	__gtod_offset = (scd->tick_raw + __sched_clock_offset) - scd->tick_gtod;
206}
207
208void __init sched_clock_init(void)
209{
210	/*
211	 * Set __gtod_offset such that once we mark sched_clock_running,
212	 * sched_clock_tick() continues where sched_clock() left off.
213	 *
214	 * Even if TSC is buggered, we're still UP at this point so it
215	 * can't really be out of sync.
216	 */
217	local_irq_disable();
218	__sched_clock_gtod_offset();
219	local_irq_enable();
220
221	static_branch_inc(&sched_clock_running);
222}
223/*
224 * We run this as late_initcall() such that it runs after all built-in drivers,
225 * notably: acpi_processor and intel_idle, which can mark the TSC as unstable.
226 */
227static int __init sched_clock_init_late(void)
228{
229	static_branch_inc(&sched_clock_running);
230	/*
231	 * Ensure that it is impossible to not do a static_key update.
232	 *
233	 * Either {set,clear}_sched_clock_stable() must see sched_clock_running
234	 * and do the update, or we must see their __sched_clock_stable_early
235	 * and do the update, or both.
236	 */
237	smp_mb(); /* matches {set,clear}_sched_clock_stable() */
238
239	if (__sched_clock_stable_early)
240		__set_sched_clock_stable();
241
242	return 0;
243}
244late_initcall(sched_clock_init_late);
245
246/*
247 * min, max except they take wrapping into account
248 */
249
250static __always_inline u64 wrap_min(u64 x, u64 y)
251{
252	return (s64)(x - y) < 0 ? x : y;
253}
254
255static __always_inline u64 wrap_max(u64 x, u64 y)
256{
257	return (s64)(x - y) > 0 ? x : y;
258}
259
260/*
261 * update the percpu scd from the raw @now value
262 *
263 *  - filter out backward motion
264 *  - use the GTOD tick value to create a window to filter crazy TSC values
265 */
266static __always_inline u64 sched_clock_local(struct sched_clock_data *scd)
267{
268	u64 now, clock, old_clock, min_clock, max_clock, gtod;
269	s64 delta;
270
271again:
272	now = sched_clock_noinstr();
273	delta = now - scd->tick_raw;
274	if (unlikely(delta < 0))
275		delta = 0;
276
277	old_clock = scd->clock;
278
279	/*
280	 * scd->clock = clamp(scd->tick_gtod + delta,
281	 *		      max(scd->tick_gtod, scd->clock),
282	 *		      scd->tick_gtod + TICK_NSEC);
283	 */
284
285	gtod = scd->tick_gtod + __gtod_offset;
286	clock = gtod + delta;
287	min_clock = wrap_max(gtod, old_clock);
288	max_clock = wrap_max(old_clock, gtod + TICK_NSEC);
289
290	clock = wrap_max(clock, min_clock);
291	clock = wrap_min(clock, max_clock);
292
293	if (!raw_try_cmpxchg64(&scd->clock, &old_clock, clock))
294		goto again;
295
296	return clock;
297}
298
299noinstr u64 local_clock_noinstr(void)
300{
301	u64 clock;
302
303	if (static_branch_likely(&__sched_clock_stable))
304		return sched_clock_noinstr() + __sched_clock_offset;
305
306	if (!static_branch_likely(&sched_clock_running))
307		return sched_clock_noinstr();
308
309	clock = sched_clock_local(this_scd());
310
311	return clock;
312}
313
314u64 local_clock(void)
315{
316	u64 now;
317	preempt_disable_notrace();
318	now = local_clock_noinstr();
319	preempt_enable_notrace();
320	return now;
321}
322EXPORT_SYMBOL_GPL(local_clock);
323
324static notrace u64 sched_clock_remote(struct sched_clock_data *scd)
325{
326	struct sched_clock_data *my_scd = this_scd();
327	u64 this_clock, remote_clock;
328	u64 *ptr, old_val, val;
329
330#if BITS_PER_LONG != 64
331again:
332	/*
333	 * Careful here: The local and the remote clock values need to
334	 * be read out atomic as we need to compare the values and
335	 * then update either the local or the remote side. So the
336	 * cmpxchg64 below only protects one readout.
337	 *
338	 * We must reread via sched_clock_local() in the retry case on
339	 * 32-bit kernels as an NMI could use sched_clock_local() via the
340	 * tracer and hit between the readout of
341	 * the low 32-bit and the high 32-bit portion.
342	 */
343	this_clock = sched_clock_local(my_scd);
344	/*
345	 * We must enforce atomic readout on 32-bit, otherwise the
346	 * update on the remote CPU can hit in between the readout of
347	 * the low 32-bit and the high 32-bit portion.
348	 */
349	remote_clock = cmpxchg64(&scd->clock, 0, 0);
350#else
351	/*
352	 * On 64-bit kernels the read of [my]scd->clock is atomic versus the
353	 * update, so we can avoid the above 32-bit dance.
354	 */
355	sched_clock_local(my_scd);
356again:
357	this_clock = my_scd->clock;
358	remote_clock = scd->clock;
359#endif
360
361	/*
362	 * Use the opportunity that we have both locks
363	 * taken to couple the two clocks: we take the
364	 * larger time as the latest time for both
365	 * runqueues. (this creates monotonic movement)
366	 */
367	if (likely((s64)(remote_clock - this_clock) < 0)) {
368		ptr = &scd->clock;
369		old_val = remote_clock;
370		val = this_clock;
371	} else {
372		/*
373		 * Should be rare, but possible:
374		 */
375		ptr = &my_scd->clock;
376		old_val = this_clock;
377		val = remote_clock;
378	}
379
380	if (!try_cmpxchg64(ptr, &old_val, val))
381		goto again;
382
383	return val;
384}
385
386/*
387 * Similar to cpu_clock(), but requires local IRQs to be disabled.
388 *
389 * See cpu_clock().
390 */
391notrace u64 sched_clock_cpu(int cpu)
392{
393	struct sched_clock_data *scd;
394	u64 clock;
395
396	if (sched_clock_stable())
397		return sched_clock() + __sched_clock_offset;
398
399	if (!static_branch_likely(&sched_clock_running))
400		return sched_clock();
401
402	preempt_disable_notrace();
403	scd = cpu_sdc(cpu);
404
405	if (cpu != smp_processor_id())
406		clock = sched_clock_remote(scd);
407	else
408		clock = sched_clock_local(scd);
409	preempt_enable_notrace();
410
411	return clock;
412}
413EXPORT_SYMBOL_GPL(sched_clock_cpu);
414
415notrace void sched_clock_tick(void)
416{
417	struct sched_clock_data *scd;
418
419	if (sched_clock_stable())
420		return;
421
422	if (!static_branch_likely(&sched_clock_running))
423		return;
424
425	lockdep_assert_irqs_disabled();
426
427	scd = this_scd();
428	__scd_stamp(scd);
429	sched_clock_local(scd);
430}
431
432notrace void sched_clock_tick_stable(void)
433{
434	if (!sched_clock_stable())
435		return;
436
437	/*
438	 * Called under watchdog_lock.
439	 *
440	 * The watchdog just found this TSC to (still) be stable, so now is a
441	 * good moment to update our __gtod_offset. Because once we find the
442	 * TSC to be unstable, any computation will be computing crap.
443	 */
444	local_irq_disable();
445	__sched_clock_gtod_offset();
446	local_irq_enable();
447}
448
449/*
450 * We are going deep-idle (IRQs are disabled):
451 */
452notrace void sched_clock_idle_sleep_event(void)
453{
454	sched_clock_cpu(smp_processor_id());
455}
456EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
457
458/*
459 * We just idled; resync with ktime.
460 */
461notrace void sched_clock_idle_wakeup_event(void)
462{
463	unsigned long flags;
464
465	if (sched_clock_stable())
466		return;
467
468	if (unlikely(timekeeping_suspended))
469		return;
470
471	local_irq_save(flags);
472	sched_clock_tick();
473	local_irq_restore(flags);
474}
475EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
476
477#else /* !CONFIG_HAVE_UNSTABLE_SCHED_CLOCK: */
478
479void __init sched_clock_init(void)
480{
481	static_branch_inc(&sched_clock_running);
482	local_irq_disable();
483	generic_sched_clock_init();
484	local_irq_enable();
485}
486
487notrace u64 sched_clock_cpu(int cpu)
488{
489	if (!static_branch_likely(&sched_clock_running))
490		return 0;
491
492	return sched_clock();
493}
494
495#endif /* !CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
496
497/*
498 * Running clock - returns the time that has elapsed while a guest has been
499 * running.
500 * On a guest this value should be local_clock minus the time the guest was
501 * suspended by the hypervisor (for any reason).
502 * On bare metal this function should return the same as local_clock.
503 * Architectures and sub-architectures can override this.
504 */
505notrace u64 __weak running_clock(void)
506{
507	return local_clock();
508}