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
96static __always_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(bit: 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(bit: TICK_DEP_BIT_CLOCK_UNSTABLE);
184 schedule_work(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(key: &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
247static __always_inline u64 wrap_min(u64 x, u64 y)
248{
249 return (s64)(x - y) < 0 ? x : y;
250}
251
252static __always_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 */
263static __always_inline 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_noinstr();
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(x: gtod, y: old_clock);
285 max_clock = wrap_max(x: old_clock, y: gtod + TICK_NSEC);
286
287 clock = wrap_max(x: clock, y: min_clock);
288 clock = wrap_min(x: clock, y: max_clock);
289
290 if (!raw_try_cmpxchg64(&scd->clock, &old_clock, clock))
291 goto again;
292
293 return clock;
294}
295
296noinstr u64 local_clock_noinstr(void)
297{
298 u64 clock;
299
300 if (static_branch_likely(&__sched_clock_stable))
301 return sched_clock_noinstr() + __sched_clock_offset;
302
303 if (!static_branch_likely(&sched_clock_running))
304 return sched_clock_noinstr();
305
306 clock = sched_clock_local(scd: this_scd());
307
308 return clock;
309}
310
311u64 local_clock(void)
312{
313 u64 now;
314 preempt_disable_notrace();
315 now = local_clock_noinstr();
316 preempt_enable_notrace();
317 return now;
318}
319EXPORT_SYMBOL_GPL(local_clock);
320
321static notrace u64 sched_clock_remote(struct sched_clock_data *scd)
322{
323 struct sched_clock_data *my_scd = this_scd();
324 u64 this_clock, remote_clock;
325 u64 *ptr, old_val, val;
326
327#if BITS_PER_LONG != 64
328again:
329 /*
330 * Careful here: The local and the remote clock values need to
331 * be read out atomic as we need to compare the values and
332 * then update either the local or the remote side. So the
333 * cmpxchg64 below only protects one readout.
334 *
335 * We must reread via sched_clock_local() in the retry case on
336 * 32-bit kernels as an NMI could use sched_clock_local() via the
337 * tracer and hit between the readout of
338 * the low 32-bit and the high 32-bit portion.
339 */
340 this_clock = sched_clock_local(my_scd);
341 /*
342 * We must enforce atomic readout on 32-bit, otherwise the
343 * update on the remote CPU can hit inbetween the readout of
344 * the low 32-bit and the high 32-bit portion.
345 */
346 remote_clock = cmpxchg64(&scd->clock, 0, 0);
347#else
348 /*
349 * On 64-bit kernels the read of [my]scd->clock is atomic versus the
350 * update, so we can avoid the above 32-bit dance.
351 */
352 sched_clock_local(scd: my_scd);
353again:
354 this_clock = my_scd->clock;
355 remote_clock = scd->clock;
356#endif
357
358 /*
359 * Use the opportunity that we have both locks
360 * taken to couple the two clocks: we take the
361 * larger time as the latest time for both
362 * runqueues. (this creates monotonic movement)
363 */
364 if (likely((s64)(remote_clock - this_clock) < 0)) {
365 ptr = &scd->clock;
366 old_val = remote_clock;
367 val = this_clock;
368 } else {
369 /*
370 * Should be rare, but possible:
371 */
372 ptr = &my_scd->clock;
373 old_val = this_clock;
374 val = remote_clock;
375 }
376
377 if (!try_cmpxchg64(ptr, &old_val, val))
378 goto again;
379
380 return val;
381}
382
383/*
384 * Similar to cpu_clock(), but requires local IRQs to be disabled.
385 *
386 * See cpu_clock().
387 */
388notrace u64 sched_clock_cpu(int cpu)
389{
390 struct sched_clock_data *scd;
391 u64 clock;
392
393 if (sched_clock_stable())
394 return sched_clock() + __sched_clock_offset;
395
396 if (!static_branch_likely(&sched_clock_running))
397 return sched_clock();
398
399 preempt_disable_notrace();
400 scd = cpu_sdc(cpu);
401
402 if (cpu != smp_processor_id())
403 clock = sched_clock_remote(scd);
404 else
405 clock = sched_clock_local(scd);
406 preempt_enable_notrace();
407
408 return clock;
409}
410EXPORT_SYMBOL_GPL(sched_clock_cpu);
411
412notrace void sched_clock_tick(void)
413{
414 struct sched_clock_data *scd;
415
416 if (sched_clock_stable())
417 return;
418
419 if (!static_branch_likely(&sched_clock_running))
420 return;
421
422 lockdep_assert_irqs_disabled();
423
424 scd = this_scd();
425 __scd_stamp(scd);
426 sched_clock_local(scd);
427}
428
429notrace void sched_clock_tick_stable(void)
430{
431 if (!sched_clock_stable())
432 return;
433
434 /*
435 * Called under watchdog_lock.
436 *
437 * The watchdog just found this TSC to (still) be stable, so now is a
438 * good moment to update our __gtod_offset. Because once we find the
439 * TSC to be unstable, any computation will be computing crap.
440 */
441 local_irq_disable();
442 __sched_clock_gtod_offset();
443 local_irq_enable();
444}
445
446/*
447 * We are going deep-idle (irqs are disabled):
448 */
449notrace void sched_clock_idle_sleep_event(void)
450{
451 sched_clock_cpu(smp_processor_id());
452}
453EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
454
455/*
456 * We just idled; resync with ktime.
457 */
458notrace void sched_clock_idle_wakeup_event(void)
459{
460 unsigned long flags;
461
462 if (sched_clock_stable())
463 return;
464
465 if (unlikely(timekeeping_suspended))
466 return;
467
468 local_irq_save(flags);
469 sched_clock_tick();
470 local_irq_restore(flags);
471}
472EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
473
474#else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
475
476void __init sched_clock_init(void)
477{
478 static_branch_inc(&sched_clock_running);
479 local_irq_disable();
480 generic_sched_clock_init();
481 local_irq_enable();
482}
483
484notrace u64 sched_clock_cpu(int cpu)
485{
486 if (!static_branch_likely(&sched_clock_running))
487 return 0;
488
489 return sched_clock();
490}
491
492#endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
493
494/*
495 * Running clock - returns the time that has elapsed while a guest has been
496 * running.
497 * On a guest this value should be local_clock minus the time the guest was
498 * suspended by the hypervisor (for any reason).
499 * On bare metal this function should return the same as local_clock.
500 * Architectures and sub-architectures can override this.
501 */
502notrace u64 __weak running_clock(void)
503{
504 return local_clock();
505}
506

source code of linux/kernel/sched/clock.c