2 * sched_clock for unstable cpu clocks
4 * Copyright (C) 2008 Red Hat, Inc., Peter Zijlstra
6 * Updates and enhancements:
7 * Copyright (C) 2008 Red Hat, Inc. Steven Rostedt <srostedt@redhat.com>
10 * Ingo Molnar <mingo@redhat.com>
11 * Guillaume Chazarain <guichaz@gmail.com>
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.
20 * ######################### BIG FAT WARNING ##########################
21 * # when comparing cpu_clock(i) to cpu_clock(j) for i != j, time can #
23 * ####################################################################
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).
28 * cpu_clock(i) -- can be used from any context, including NMI.
29 * local_clock() -- is cpu_clock() on the current cpu.
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).
40 * Otherwise it tries to create a semi stable clock from a mixture of other
43 * - GTOD (clock monotomic)
45 * - explicit idle events
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
51 * Furthermore, explicit sleep and wakeup hooks allow us to account for time
52 * that is otherwise invisible (TSC gets stopped).
55 #include <linux/spinlock.h>
56 #include <linux/hardirq.h>
57 #include <linux/export.h>
58 #include <linux/percpu.h>
59 #include <linux/ktime.h>
60 #include <linux/sched.h>
61 #include <linux/nmi.h>
62 #include <linux/sched/clock.h>
63 #include <linux/static_key.h>
64 #include <linux/workqueue.h>
65 #include <linux/compiler.h>
66 #include <linux/tick.h>
69 * Scheduler clock - returns current time in nanosec units.
70 * This is default implementation.
71 * Architectures and sub-architectures can override this.
73 unsigned long long __weak sched_clock(void)
75 return (unsigned long long)(jiffies - INITIAL_JIFFIES)
76 * (NSEC_PER_SEC / HZ);
78 EXPORT_SYMBOL_GPL(sched_clock);
80 __read_mostly int sched_clock_running;
82 void sched_clock_init(void)
84 sched_clock_running = 1;
87 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
89 * We must start with !__sched_clock_stable because the unstable -> stable
90 * transition is accurate, while the stable -> unstable transition is not.
92 * Similarly we start with __sched_clock_stable_early, thereby assuming we
93 * will become stable, such that there's only a single 1 -> 0 transition.
95 static DEFINE_STATIC_KEY_FALSE(__sched_clock_stable);
96 static int __sched_clock_stable_early = 1;
99 * We want: ktime_get_ns() + __gtod_offset == sched_clock() + __sched_clock_offset
101 __read_mostly u64 __sched_clock_offset;
102 static __read_mostly u64 __gtod_offset;
104 struct sched_clock_data {
110 static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
112 static inline struct sched_clock_data *this_scd(void)
114 return this_cpu_ptr(&sched_clock_data);
117 static inline struct sched_clock_data *cpu_sdc(int cpu)
119 return &per_cpu(sched_clock_data, cpu);
122 int sched_clock_stable(void)
124 return static_branch_likely(&__sched_clock_stable);
127 static void __scd_stamp(struct sched_clock_data *scd)
129 scd->tick_gtod = ktime_get_ns();
130 scd->tick_raw = sched_clock();
133 static void __set_sched_clock_stable(void)
135 struct sched_clock_data *scd = this_scd();
138 * Attempt to make the (initial) unstable->stable transition continuous.
140 __sched_clock_offset = (scd->tick_gtod + __gtod_offset) - (scd->tick_raw);
142 printk(KERN_INFO "sched_clock: Marking stable (%lld, %lld)->(%lld, %lld)\n",
143 scd->tick_gtod, __gtod_offset,
144 scd->tick_raw, __sched_clock_offset);
146 static_branch_enable(&__sched_clock_stable);
147 tick_dep_clear(TICK_DEP_BIT_CLOCK_UNSTABLE);
151 * If we ever get here, we're screwed, because we found out -- typically after
152 * the fact -- that TSC wasn't good. This means all our clocksources (including
153 * ktime) could have reported wrong values.
155 * What we do here is an attempt to fix up and continue sort of where we left
156 * off in a coherent manner.
158 * The only way to fully avoid random clock jumps is to boot with:
161 static void __sched_clock_work(struct work_struct *work)
163 struct sched_clock_data *scd;
166 /* take a current timestamp and set 'now' */
170 scd->clock = scd->tick_gtod + __gtod_offset;
173 /* clone to all CPUs */
174 for_each_possible_cpu(cpu)
175 per_cpu(sched_clock_data, cpu) = *scd;
177 printk(KERN_INFO "sched_clock: Marking unstable (%lld, %lld)<-(%lld, %lld)\n",
178 scd->tick_gtod, __gtod_offset,
179 scd->tick_raw, __sched_clock_offset);
181 static_branch_disable(&__sched_clock_stable);
184 static DECLARE_WORK(sched_clock_work, __sched_clock_work);
186 static void __clear_sched_clock_stable(void)
188 if (!sched_clock_stable())
191 tick_dep_set(TICK_DEP_BIT_CLOCK_UNSTABLE);
192 schedule_work(&sched_clock_work);
195 void clear_sched_clock_stable(void)
197 __sched_clock_stable_early = 0;
199 smp_mb(); /* matches sched_clock_init_late() */
201 if (sched_clock_running == 2)
202 __clear_sched_clock_stable();
205 void sched_clock_init_late(void)
207 sched_clock_running = 2;
209 * Ensure that it is impossible to not do a static_key update.
211 * Either {set,clear}_sched_clock_stable() must see sched_clock_running
212 * and do the update, or we must see their __sched_clock_stable_early
213 * and do the update, or both.
215 smp_mb(); /* matches {set,clear}_sched_clock_stable() */
217 if (__sched_clock_stable_early)
218 __set_sched_clock_stable();
222 * min, max except they take wrapping into account
225 static inline u64 wrap_min(u64 x, u64 y)
227 return (s64)(x - y) < 0 ? x : y;
230 static inline u64 wrap_max(u64 x, u64 y)
232 return (s64)(x - y) > 0 ? x : y;
236 * update the percpu scd from the raw @now value
238 * - filter out backward motion
239 * - use the GTOD tick value to create a window to filter crazy TSC values
241 static u64 sched_clock_local(struct sched_clock_data *scd)
243 u64 now, clock, old_clock, min_clock, max_clock, gtod;
248 delta = now - scd->tick_raw;
249 if (unlikely(delta < 0))
252 old_clock = scd->clock;
255 * scd->clock = clamp(scd->tick_gtod + delta,
256 * max(scd->tick_gtod, scd->clock),
257 * scd->tick_gtod + TICK_NSEC);
260 gtod = scd->tick_gtod + __gtod_offset;
261 clock = gtod + delta;
262 min_clock = wrap_max(gtod, old_clock);
263 max_clock = wrap_max(old_clock, gtod + TICK_NSEC);
265 clock = wrap_max(clock, min_clock);
266 clock = wrap_min(clock, max_clock);
268 if (cmpxchg64(&scd->clock, old_clock, clock) != old_clock)
274 static u64 sched_clock_remote(struct sched_clock_data *scd)
276 struct sched_clock_data *my_scd = this_scd();
277 u64 this_clock, remote_clock;
278 u64 *ptr, old_val, val;
280 #if BITS_PER_LONG != 64
283 * Careful here: The local and the remote clock values need to
284 * be read out atomic as we need to compare the values and
285 * then update either the local or the remote side. So the
286 * cmpxchg64 below only protects one readout.
288 * We must reread via sched_clock_local() in the retry case on
289 * 32bit as an NMI could use sched_clock_local() via the
290 * tracer and hit between the readout of
291 * the low32bit and the high 32bit portion.
293 this_clock = sched_clock_local(my_scd);
295 * We must enforce atomic readout on 32bit, otherwise the
296 * update on the remote cpu can hit inbetween the readout of
297 * the low32bit and the high 32bit portion.
299 remote_clock = cmpxchg64(&scd->clock, 0, 0);
302 * On 64bit the read of [my]scd->clock is atomic versus the
303 * update, so we can avoid the above 32bit dance.
305 sched_clock_local(my_scd);
307 this_clock = my_scd->clock;
308 remote_clock = scd->clock;
312 * Use the opportunity that we have both locks
313 * taken to couple the two clocks: we take the
314 * larger time as the latest time for both
315 * runqueues. (this creates monotonic movement)
317 if (likely((s64)(remote_clock - this_clock) < 0)) {
319 old_val = remote_clock;
323 * Should be rare, but possible:
325 ptr = &my_scd->clock;
326 old_val = this_clock;
330 if (cmpxchg64(ptr, old_val, val) != old_val)
337 * Similar to cpu_clock(), but requires local IRQs to be disabled.
341 u64 sched_clock_cpu(int cpu)
343 struct sched_clock_data *scd;
346 if (sched_clock_stable())
347 return sched_clock() + __sched_clock_offset;
349 if (unlikely(!sched_clock_running))
352 preempt_disable_notrace();
355 if (cpu != smp_processor_id())
356 clock = sched_clock_remote(scd);
358 clock = sched_clock_local(scd);
359 preempt_enable_notrace();
363 EXPORT_SYMBOL_GPL(sched_clock_cpu);
365 void sched_clock_tick(void)
367 struct sched_clock_data *scd;
369 if (sched_clock_stable())
372 if (unlikely(!sched_clock_running))
375 WARN_ON_ONCE(!irqs_disabled());
379 sched_clock_local(scd);
382 void sched_clock_tick_stable(void)
386 if (!sched_clock_stable())
390 * Called under watchdog_lock.
392 * The watchdog just found this TSC to (still) be stable, so now is a
393 * good moment to update our __gtod_offset. Because once we find the
394 * TSC to be unstable, any computation will be computing crap.
397 gtod = ktime_get_ns();
398 clock = sched_clock();
399 __gtod_offset = (clock + __sched_clock_offset) - gtod;
404 * We are going deep-idle (irqs are disabled):
406 void sched_clock_idle_sleep_event(void)
408 sched_clock_cpu(smp_processor_id());
410 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
413 * We just idled delta nanoseconds (called with irqs disabled):
415 void sched_clock_idle_wakeup_event(u64 delta_ns)
417 if (timekeeping_suspended)
421 touch_softlockup_watchdog_sched();
423 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
425 #else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
427 u64 sched_clock_cpu(int cpu)
429 if (unlikely(!sched_clock_running))
432 return sched_clock();
435 #endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
438 * Running clock - returns the time that has elapsed while a guest has been
440 * On a guest this value should be local_clock minus the time the guest was
441 * suspended by the hypervisor (for any reason).
442 * On bare metal this function should return the same as local_clock.
443 * Architectures and sub-architectures can override this.
445 u64 __weak running_clock(void)
447 return local_clock();