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