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 monotomic)
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).
57 #include <linux/sched_clock.h>
60 * Scheduler clock - returns current time in nanosec units.
61 * This is default implementation.
62 * Architectures and sub-architectures can override this.
64 unsigned long long __weak sched_clock(void)
66 return (unsigned long long)(jiffies - INITIAL_JIFFIES)
67 * (NSEC_PER_SEC / HZ);
69 EXPORT_SYMBOL_GPL(sched_clock);
71 static DEFINE_STATIC_KEY_FALSE(sched_clock_running);
73 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
75 * We must start with !__sched_clock_stable because the unstable -> stable
76 * transition is accurate, while the stable -> unstable transition is not.
78 * Similarly we start with __sched_clock_stable_early, thereby assuming we
79 * will become stable, such that there's only a single 1 -> 0 transition.
81 static DEFINE_STATIC_KEY_FALSE(__sched_clock_stable);
82 static int __sched_clock_stable_early = 1;
85 * We want: ktime_get_ns() + __gtod_offset == sched_clock() + __sched_clock_offset
87 __read_mostly u64 __sched_clock_offset;
88 static __read_mostly u64 __gtod_offset;
90 struct sched_clock_data {
96 static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
98 static inline struct sched_clock_data *this_scd(void)
100 return this_cpu_ptr(&sched_clock_data);
103 static inline struct sched_clock_data *cpu_sdc(int cpu)
105 return &per_cpu(sched_clock_data, cpu);
108 int sched_clock_stable(void)
110 return static_branch_likely(&__sched_clock_stable);
113 static void __scd_stamp(struct sched_clock_data *scd)
115 scd->tick_gtod = ktime_get_ns();
116 scd->tick_raw = sched_clock();
119 static void __set_sched_clock_stable(void)
121 struct sched_clock_data *scd;
124 * Since we're still unstable and the tick is already running, we have
125 * to disable IRQs in order to get a consistent scd->tick* reading.
130 * Attempt to make the (initial) unstable->stable transition continuous.
132 __sched_clock_offset = (scd->tick_gtod + __gtod_offset) - (scd->tick_raw);
135 printk(KERN_INFO "sched_clock: Marking stable (%lld, %lld)->(%lld, %lld)\n",
136 scd->tick_gtod, __gtod_offset,
137 scd->tick_raw, __sched_clock_offset);
139 static_branch_enable(&__sched_clock_stable);
140 tick_dep_clear(TICK_DEP_BIT_CLOCK_UNSTABLE);
144 * If we ever get here, we're screwed, because we found out -- typically after
145 * the fact -- that TSC wasn't good. This means all our clocksources (including
146 * ktime) could have reported wrong values.
148 * What we do here is an attempt to fix up and continue sort of where we left
149 * off in a coherent manner.
151 * The only way to fully avoid random clock jumps is to boot with:
154 static void __sched_clock_work(struct work_struct *work)
156 struct sched_clock_data *scd;
159 /* take a current timestamp and set 'now' */
163 scd->clock = scd->tick_gtod + __gtod_offset;
166 /* clone to all CPUs */
167 for_each_possible_cpu(cpu)
168 per_cpu(sched_clock_data, cpu) = *scd;
170 printk(KERN_WARNING "TSC found unstable after boot, most likely due to broken BIOS. Use 'tsc=unstable'.\n");
171 printk(KERN_INFO "sched_clock: Marking unstable (%lld, %lld)<-(%lld, %lld)\n",
172 scd->tick_gtod, __gtod_offset,
173 scd->tick_raw, __sched_clock_offset);
175 static_branch_disable(&__sched_clock_stable);
178 static DECLARE_WORK(sched_clock_work, __sched_clock_work);
180 static void __clear_sched_clock_stable(void)
182 if (!sched_clock_stable())
185 tick_dep_set(TICK_DEP_BIT_CLOCK_UNSTABLE);
186 schedule_work(&sched_clock_work);
189 void clear_sched_clock_stable(void)
191 __sched_clock_stable_early = 0;
193 smp_mb(); /* matches sched_clock_init_late() */
195 if (static_key_count(&sched_clock_running.key) == 2)
196 __clear_sched_clock_stable();
199 static void __sched_clock_gtod_offset(void)
201 struct sched_clock_data *scd = this_scd();
204 __gtod_offset = (scd->tick_raw + __sched_clock_offset) - scd->tick_gtod;
207 void __init sched_clock_init(void)
210 * Set __gtod_offset such that once we mark sched_clock_running,
211 * sched_clock_tick() continues where sched_clock() left off.
213 * Even if TSC is buggered, we're still UP at this point so it
214 * can't really be out of sync.
217 __sched_clock_gtod_offset();
220 static_branch_inc(&sched_clock_running);
223 * We run this as late_initcall() such that it runs after all built-in drivers,
224 * notably: acpi_processor and intel_idle, which can mark the TSC as unstable.
226 static int __init sched_clock_init_late(void)
228 static_branch_inc(&sched_clock_running);
230 * Ensure that it is impossible to not do a static_key update.
232 * Either {set,clear}_sched_clock_stable() must see sched_clock_running
233 * and do the update, or we must see their __sched_clock_stable_early
234 * and do the update, or both.
236 smp_mb(); /* matches {set,clear}_sched_clock_stable() */
238 if (__sched_clock_stable_early)
239 __set_sched_clock_stable();
243 late_initcall(sched_clock_init_late);
246 * min, max except they take wrapping into account
249 static inline u64 wrap_min(u64 x, u64 y)
251 return (s64)(x - y) < 0 ? x : y;
254 static inline u64 wrap_max(u64 x, u64 y)
256 return (s64)(x - y) > 0 ? x : y;
260 * update the percpu scd from the raw @now value
262 * - filter out backward motion
263 * - use the GTOD tick value to create a window to filter crazy TSC values
265 static u64 sched_clock_local(struct sched_clock_data *scd)
267 u64 now, clock, old_clock, min_clock, max_clock, gtod;
272 delta = now - scd->tick_raw;
273 if (unlikely(delta < 0))
276 old_clock = scd->clock;
279 * scd->clock = clamp(scd->tick_gtod + delta,
280 * max(scd->tick_gtod, scd->clock),
281 * scd->tick_gtod + TICK_NSEC);
284 gtod = scd->tick_gtod + __gtod_offset;
285 clock = gtod + delta;
286 min_clock = wrap_max(gtod, old_clock);
287 max_clock = wrap_max(old_clock, gtod + TICK_NSEC);
289 clock = wrap_max(clock, min_clock);
290 clock = wrap_min(clock, max_clock);
292 if (cmpxchg64(&scd->clock, old_clock, clock) != old_clock)
298 static u64 sched_clock_remote(struct sched_clock_data *scd)
300 struct sched_clock_data *my_scd = this_scd();
301 u64 this_clock, remote_clock;
302 u64 *ptr, old_val, val;
304 #if BITS_PER_LONG != 64
307 * Careful here: The local and the remote clock values need to
308 * be read out atomic as we need to compare the values and
309 * then update either the local or the remote side. So the
310 * cmpxchg64 below only protects one readout.
312 * We must reread via sched_clock_local() in the retry case on
313 * 32-bit kernels as an NMI could use sched_clock_local() via the
314 * tracer and hit between the readout of
315 * the low 32-bit and the high 32-bit portion.
317 this_clock = sched_clock_local(my_scd);
319 * We must enforce atomic readout on 32-bit, otherwise the
320 * update on the remote CPU can hit inbetween the readout of
321 * the low 32-bit and the high 32-bit portion.
323 remote_clock = cmpxchg64(&scd->clock, 0, 0);
326 * On 64-bit kernels the read of [my]scd->clock is atomic versus the
327 * update, so we can avoid the above 32-bit dance.
329 sched_clock_local(my_scd);
331 this_clock = my_scd->clock;
332 remote_clock = scd->clock;
336 * Use the opportunity that we have both locks
337 * taken to couple the two clocks: we take the
338 * larger time as the latest time for both
339 * runqueues. (this creates monotonic movement)
341 if (likely((s64)(remote_clock - this_clock) < 0)) {
343 old_val = remote_clock;
347 * Should be rare, but possible:
349 ptr = &my_scd->clock;
350 old_val = this_clock;
354 if (cmpxchg64(ptr, old_val, val) != old_val)
361 * Similar to cpu_clock(), but requires local IRQs to be disabled.
365 u64 sched_clock_cpu(int cpu)
367 struct sched_clock_data *scd;
370 if (sched_clock_stable())
371 return sched_clock() + __sched_clock_offset;
373 if (!static_branch_likely(&sched_clock_running))
374 return sched_clock();
376 preempt_disable_notrace();
379 if (cpu != smp_processor_id())
380 clock = sched_clock_remote(scd);
382 clock = sched_clock_local(scd);
383 preempt_enable_notrace();
387 EXPORT_SYMBOL_GPL(sched_clock_cpu);
389 void sched_clock_tick(void)
391 struct sched_clock_data *scd;
393 if (sched_clock_stable())
396 if (!static_branch_likely(&sched_clock_running))
399 lockdep_assert_irqs_disabled();
403 sched_clock_local(scd);
406 void sched_clock_tick_stable(void)
408 if (!sched_clock_stable())
412 * Called under watchdog_lock.
414 * The watchdog just found this TSC to (still) be stable, so now is a
415 * good moment to update our __gtod_offset. Because once we find the
416 * TSC to be unstable, any computation will be computing crap.
419 __sched_clock_gtod_offset();
424 * We are going deep-idle (irqs are disabled):
426 void sched_clock_idle_sleep_event(void)
428 sched_clock_cpu(smp_processor_id());
430 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
433 * We just idled; resync with ktime.
435 void sched_clock_idle_wakeup_event(void)
439 if (sched_clock_stable())
442 if (unlikely(timekeeping_suspended))
445 local_irq_save(flags);
447 local_irq_restore(flags);
449 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
451 #else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
453 void __init sched_clock_init(void)
455 static_branch_inc(&sched_clock_running);
457 generic_sched_clock_init();
461 u64 sched_clock_cpu(int cpu)
463 if (!static_branch_likely(&sched_clock_running))
466 return sched_clock();
469 #endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
472 * Running clock - returns the time that has elapsed while a guest has been
474 * On a guest this value should be local_clock minus the time the guest was
475 * suspended by the hypervisor (for any reason).
476 * On bare metal this function should return the same as local_clock.
477 * Architectures and sub-architectures can override this.
479 u64 __weak running_clock(void)
481 return local_clock();