1 // SPDX-License-Identifier: GPL-2.0-only
3 * sched_clock() for unstable CPU clocks
5 * Copyright (C) 2008 Red Hat, Inc., Peter Zijlstra
7 * Updates and enhancements:
8 * Copyright (C) 2008 Red Hat, Inc. Steven Rostedt <srostedt@redhat.com>
11 * Ingo Molnar <mingo@redhat.com>
12 * Guillaume Chazarain <guichaz@gmail.com>
15 * What this file implements:
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.
21 * ######################### BIG FAT WARNING ##########################
22 * # when comparing cpu_clock(i) to cpu_clock(j) for i != j, time can #
24 * ####################################################################
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).
29 * cpu_clock(i) -- can be used from any context, including NMI.
30 * local_clock() -- is cpu_clock() on the current CPU.
34 * How it is implemented:
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).
41 * Otherwise it tries to create a semi stable clock from a mixture of other
44 * - GTOD (clock monotonic)
46 * - explicit idle events
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
52 * Furthermore, explicit sleep and wakeup hooks allow us to account for time
53 * that is otherwise invisible (TSC gets stopped).
58 * Scheduler clock - returns current time in nanosec units.
59 * This is default implementation.
60 * Architectures and sub-architectures can override this.
62 notrace unsigned long long __weak sched_clock(void)
64 return (unsigned long long)(jiffies - INITIAL_JIFFIES)
65 * (NSEC_PER_SEC / HZ);
67 EXPORT_SYMBOL_GPL(sched_clock);
69 static DEFINE_STATIC_KEY_FALSE(sched_clock_running);
71 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
73 * We must start with !__sched_clock_stable because the unstable -> stable
74 * transition is accurate, while the stable -> unstable transition is not.
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.
79 static DEFINE_STATIC_KEY_FALSE(__sched_clock_stable);
80 static int __sched_clock_stable_early = 1;
83 * We want: ktime_get_ns() + __gtod_offset == sched_clock() + __sched_clock_offset
85 __read_mostly u64 __sched_clock_offset;
86 static __read_mostly u64 __gtod_offset;
88 struct sched_clock_data {
94 static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
96 static __always_inline struct sched_clock_data *this_scd(void)
98 return this_cpu_ptr(&sched_clock_data);
101 notrace static inline struct sched_clock_data *cpu_sdc(int cpu)
103 return &per_cpu(sched_clock_data, cpu);
106 notrace int sched_clock_stable(void)
108 return static_branch_likely(&__sched_clock_stable);
111 notrace static void __scd_stamp(struct sched_clock_data *scd)
113 scd->tick_gtod = ktime_get_ns();
114 scd->tick_raw = sched_clock();
117 notrace static void __set_sched_clock_stable(void)
119 struct sched_clock_data *scd;
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.
128 * Attempt to make the (initial) unstable->stable transition continuous.
130 __sched_clock_offset = (scd->tick_gtod + __gtod_offset) - (scd->tick_raw);
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);
137 static_branch_enable(&__sched_clock_stable);
138 tick_dep_clear(TICK_DEP_BIT_CLOCK_UNSTABLE);
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.
146 * What we do here is an attempt to fix up and continue sort of where we left
147 * off in a coherent manner.
149 * The only way to fully avoid random clock jumps is to boot with:
152 notrace static void __sched_clock_work(struct work_struct *work)
154 struct sched_clock_data *scd;
157 /* take a current timestamp and set 'now' */
161 scd->clock = scd->tick_gtod + __gtod_offset;
164 /* clone to all CPUs */
165 for_each_possible_cpu(cpu)
166 per_cpu(sched_clock_data, cpu) = *scd;
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);
173 static_branch_disable(&__sched_clock_stable);
176 static DECLARE_WORK(sched_clock_work, __sched_clock_work);
178 notrace static void __clear_sched_clock_stable(void)
180 if (!sched_clock_stable())
183 tick_dep_set(TICK_DEP_BIT_CLOCK_UNSTABLE);
184 schedule_work(&sched_clock_work);
187 notrace void clear_sched_clock_stable(void)
189 __sched_clock_stable_early = 0;
191 smp_mb(); /* matches sched_clock_init_late() */
193 if (static_key_count(&sched_clock_running.key) == 2)
194 __clear_sched_clock_stable();
197 notrace static void __sched_clock_gtod_offset(void)
199 struct sched_clock_data *scd = this_scd();
202 __gtod_offset = (scd->tick_raw + __sched_clock_offset) - scd->tick_gtod;
205 void __init sched_clock_init(void)
208 * Set __gtod_offset such that once we mark sched_clock_running,
209 * sched_clock_tick() continues where sched_clock() left off.
211 * Even if TSC is buggered, we're still UP at this point so it
212 * can't really be out of sync.
215 __sched_clock_gtod_offset();
218 static_branch_inc(&sched_clock_running);
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.
224 static int __init sched_clock_init_late(void)
226 static_branch_inc(&sched_clock_running);
228 * Ensure that it is impossible to not do a static_key update.
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.
234 smp_mb(); /* matches {set,clear}_sched_clock_stable() */
236 if (__sched_clock_stable_early)
237 __set_sched_clock_stable();
241 late_initcall(sched_clock_init_late);
244 * min, max except they take wrapping into account
247 static __always_inline u64 wrap_min(u64 x, u64 y)
249 return (s64)(x - y) < 0 ? x : y;
252 static __always_inline u64 wrap_max(u64 x, u64 y)
254 return (s64)(x - y) > 0 ? x : y;
258 * update the percpu scd from the raw @now value
260 * - filter out backward motion
261 * - use the GTOD tick value to create a window to filter crazy TSC values
263 static __always_inline u64 sched_clock_local(struct sched_clock_data *scd)
265 u64 now, clock, old_clock, min_clock, max_clock, gtod;
269 now = sched_clock_noinstr();
270 delta = now - scd->tick_raw;
271 if (unlikely(delta < 0))
274 old_clock = scd->clock;
277 * scd->clock = clamp(scd->tick_gtod + delta,
278 * max(scd->tick_gtod, scd->clock),
279 * scd->tick_gtod + TICK_NSEC);
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);
287 clock = wrap_max(clock, min_clock);
288 clock = wrap_min(clock, max_clock);
290 if (!raw_try_cmpxchg64(&scd->clock, &old_clock, clock))
296 noinstr u64 local_clock_noinstr(void)
300 if (static_branch_likely(&__sched_clock_stable))
301 return sched_clock_noinstr() + __sched_clock_offset;
303 if (!static_branch_likely(&sched_clock_running))
304 return sched_clock_noinstr();
306 clock = sched_clock_local(this_scd());
311 u64 local_clock(void)
314 preempt_disable_notrace();
315 now = local_clock_noinstr();
316 preempt_enable_notrace();
319 EXPORT_SYMBOL_GPL(local_clock);
321 static notrace u64 sched_clock_remote(struct sched_clock_data *scd)
323 struct sched_clock_data *my_scd = this_scd();
324 u64 this_clock, remote_clock;
325 u64 *ptr, old_val, val;
327 #if BITS_PER_LONG != 64
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.
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.
340 this_clock = sched_clock_local(my_scd);
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.
346 remote_clock = cmpxchg64(&scd->clock, 0, 0);
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.
352 sched_clock_local(my_scd);
354 this_clock = my_scd->clock;
355 remote_clock = scd->clock;
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)
364 if (likely((s64)(remote_clock - this_clock) < 0)) {
366 old_val = remote_clock;
370 * Should be rare, but possible:
372 ptr = &my_scd->clock;
373 old_val = this_clock;
377 if (!try_cmpxchg64(ptr, &old_val, val))
384 * Similar to cpu_clock(), but requires local IRQs to be disabled.
388 notrace u64 sched_clock_cpu(int cpu)
390 struct sched_clock_data *scd;
393 if (sched_clock_stable())
394 return sched_clock() + __sched_clock_offset;
396 if (!static_branch_likely(&sched_clock_running))
397 return sched_clock();
399 preempt_disable_notrace();
402 if (cpu != smp_processor_id())
403 clock = sched_clock_remote(scd);
405 clock = sched_clock_local(scd);
406 preempt_enable_notrace();
410 EXPORT_SYMBOL_GPL(sched_clock_cpu);
412 notrace void sched_clock_tick(void)
414 struct sched_clock_data *scd;
416 if (sched_clock_stable())
419 if (!static_branch_likely(&sched_clock_running))
422 lockdep_assert_irqs_disabled();
426 sched_clock_local(scd);
429 notrace void sched_clock_tick_stable(void)
431 if (!sched_clock_stable())
435 * Called under watchdog_lock.
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.
442 __sched_clock_gtod_offset();
447 * We are going deep-idle (irqs are disabled):
449 notrace void sched_clock_idle_sleep_event(void)
451 sched_clock_cpu(smp_processor_id());
453 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
456 * We just idled; resync with ktime.
458 notrace void sched_clock_idle_wakeup_event(void)
462 if (sched_clock_stable())
465 if (unlikely(timekeeping_suspended))
468 local_irq_save(flags);
470 local_irq_restore(flags);
472 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
474 #else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
476 void __init sched_clock_init(void)
478 static_branch_inc(&sched_clock_running);
480 generic_sched_clock_init();
484 notrace u64 sched_clock_cpu(int cpu)
486 if (!static_branch_likely(&sched_clock_running))
489 return sched_clock();
492 #endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
495 * Running clock - returns the time that has elapsed while a guest has been
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.
502 notrace u64 __weak running_clock(void)
504 return local_clock();