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() + raw_offset
101 static __read_mostly u64 raw_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 __set_sched_clock_stable(void)
129 struct sched_clock_data *scd = this_scd();
132 * Attempt to make the (initial) unstable->stable transition continuous.
134 raw_offset = (scd->tick_gtod + gtod_offset) - (scd->tick_raw);
136 printk(KERN_INFO "sched_clock: Marking stable (%lld, %lld)->(%lld, %lld)\n",
137 scd->tick_gtod, gtod_offset,
138 scd->tick_raw, raw_offset);
140 static_branch_enable(&__sched_clock_stable);
141 tick_dep_clear(TICK_DEP_BIT_CLOCK_UNSTABLE);
144 static void __clear_sched_clock_stable(struct work_struct *work)
146 struct sched_clock_data *scd = this_scd();
149 * Attempt to make the stable->unstable transition continuous.
151 * Trouble is, this is typically called from the TSC watchdog
152 * timer, which is late per definition. This means the tick
153 * values can already be screwy.
155 * Still do what we can.
157 gtod_offset = (scd->tick_raw + raw_offset) - (scd->tick_gtod);
159 printk(KERN_INFO "sched_clock: Marking unstable (%lld, %lld)<-(%lld, %lld)\n",
160 scd->tick_gtod, gtod_offset,
161 scd->tick_raw, raw_offset);
163 static_branch_disable(&__sched_clock_stable);
164 tick_dep_set(TICK_DEP_BIT_CLOCK_UNSTABLE);
167 static DECLARE_WORK(sched_clock_work, __clear_sched_clock_stable);
169 void clear_sched_clock_stable(void)
171 __sched_clock_stable_early = 0;
173 smp_mb(); /* matches sched_clock_init_late() */
175 if (sched_clock_running == 2)
176 schedule_work(&sched_clock_work);
179 void sched_clock_init_late(void)
181 sched_clock_running = 2;
183 * Ensure that it is impossible to not do a static_key update.
185 * Either {set,clear}_sched_clock_stable() must see sched_clock_running
186 * and do the update, or we must see their __sched_clock_stable_early
187 * and do the update, or both.
189 smp_mb(); /* matches {set,clear}_sched_clock_stable() */
191 if (__sched_clock_stable_early)
192 __set_sched_clock_stable();
196 * min, max except they take wrapping into account
199 static inline u64 wrap_min(u64 x, u64 y)
201 return (s64)(x - y) < 0 ? x : y;
204 static inline u64 wrap_max(u64 x, u64 y)
206 return (s64)(x - y) > 0 ? x : y;
210 * update the percpu scd from the raw @now value
212 * - filter out backward motion
213 * - use the GTOD tick value to create a window to filter crazy TSC values
215 static u64 sched_clock_local(struct sched_clock_data *scd)
217 u64 now, clock, old_clock, min_clock, max_clock;
222 delta = now - scd->tick_raw;
223 if (unlikely(delta < 0))
226 old_clock = scd->clock;
229 * scd->clock = clamp(scd->tick_gtod + delta,
230 * max(scd->tick_gtod, scd->clock),
231 * scd->tick_gtod + TICK_NSEC);
234 clock = scd->tick_gtod + gtod_offset + delta;
235 min_clock = wrap_max(scd->tick_gtod, old_clock);
236 max_clock = wrap_max(old_clock, scd->tick_gtod + TICK_NSEC);
238 clock = wrap_max(clock, min_clock);
239 clock = wrap_min(clock, max_clock);
241 if (cmpxchg64(&scd->clock, old_clock, clock) != old_clock)
247 static u64 sched_clock_remote(struct sched_clock_data *scd)
249 struct sched_clock_data *my_scd = this_scd();
250 u64 this_clock, remote_clock;
251 u64 *ptr, old_val, val;
253 #if BITS_PER_LONG != 64
256 * Careful here: The local and the remote clock values need to
257 * be read out atomic as we need to compare the values and
258 * then update either the local or the remote side. So the
259 * cmpxchg64 below only protects one readout.
261 * We must reread via sched_clock_local() in the retry case on
262 * 32bit as an NMI could use sched_clock_local() via the
263 * tracer and hit between the readout of
264 * the low32bit and the high 32bit portion.
266 this_clock = sched_clock_local(my_scd);
268 * We must enforce atomic readout on 32bit, otherwise the
269 * update on the remote cpu can hit inbetween the readout of
270 * the low32bit and the high 32bit portion.
272 remote_clock = cmpxchg64(&scd->clock, 0, 0);
275 * On 64bit the read of [my]scd->clock is atomic versus the
276 * update, so we can avoid the above 32bit dance.
278 sched_clock_local(my_scd);
280 this_clock = my_scd->clock;
281 remote_clock = scd->clock;
285 * Use the opportunity that we have both locks
286 * taken to couple the two clocks: we take the
287 * larger time as the latest time for both
288 * runqueues. (this creates monotonic movement)
290 if (likely((s64)(remote_clock - this_clock) < 0)) {
292 old_val = remote_clock;
296 * Should be rare, but possible:
298 ptr = &my_scd->clock;
299 old_val = this_clock;
303 if (cmpxchg64(ptr, old_val, val) != old_val)
310 * Similar to cpu_clock(), but requires local IRQs to be disabled.
314 u64 sched_clock_cpu(int cpu)
316 struct sched_clock_data *scd;
319 if (sched_clock_stable())
320 return sched_clock() + raw_offset;
322 if (unlikely(!sched_clock_running))
325 preempt_disable_notrace();
328 if (cpu != smp_processor_id())
329 clock = sched_clock_remote(scd);
331 clock = sched_clock_local(scd);
332 preempt_enable_notrace();
336 EXPORT_SYMBOL_GPL(sched_clock_cpu);
338 void sched_clock_tick(void)
340 struct sched_clock_data *scd;
342 WARN_ON_ONCE(!irqs_disabled());
345 * Update these values even if sched_clock_stable(), because it can
346 * become unstable at any point in time at which point we need some
347 * values to fall back on.
349 * XXX arguably we can skip this if we expose tsc_clocksource_reliable
352 scd->tick_raw = sched_clock();
353 scd->tick_gtod = ktime_get_ns();
355 if (!sched_clock_stable() && likely(sched_clock_running))
356 sched_clock_local(scd);
360 * We are going deep-idle (irqs are disabled):
362 void sched_clock_idle_sleep_event(void)
364 sched_clock_cpu(smp_processor_id());
366 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
369 * We just idled delta nanoseconds (called with irqs disabled):
371 void sched_clock_idle_wakeup_event(u64 delta_ns)
373 if (timekeeping_suspended)
377 touch_softlockup_watchdog_sched();
379 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
381 #else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
383 u64 sched_clock_cpu(int cpu)
385 if (unlikely(!sched_clock_running))
388 return sched_clock();
391 #endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
394 * Running clock - returns the time that has elapsed while a guest has been
396 * On a guest this value should be local_clock minus the time the guest was
397 * suspended by the hypervisor (for any reason).
398 * On bare metal this function should return the same as local_clock.
399 * Architectures and sub-architectures can override this.
401 u64 __weak running_clock(void)
403 return local_clock();