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/static_key.h>
62 #include <linux/workqueue.h>
63 #include <linux/compiler.h>
64 #include <linux/tick.h>
67 * Scheduler clock - returns current time in nanosec units.
68 * This is default implementation.
69 * Architectures and sub-architectures can override this.
71 unsigned long long __weak sched_clock(void)
73 return (unsigned long long)(jiffies - INITIAL_JIFFIES)
74 * (NSEC_PER_SEC / HZ);
76 EXPORT_SYMBOL_GPL(sched_clock);
78 __read_mostly int sched_clock_running;
80 void sched_clock_init(void)
82 sched_clock_running = 1;
85 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
87 * We must start with !__sched_clock_stable because the unstable -> stable
88 * transition is accurate, while the stable -> unstable transition is not.
90 * Similarly we start with __sched_clock_stable_early, thereby assuming we
91 * will become stable, such that there's only a single 1 -> 0 transition.
93 static DEFINE_STATIC_KEY_FALSE(__sched_clock_stable);
94 static int __sched_clock_stable_early = 1;
97 * We want: ktime_get_ns() + gtod_offset == sched_clock() + raw_offset
99 static __read_mostly u64 raw_offset;
100 static __read_mostly u64 gtod_offset;
102 struct sched_clock_data {
108 static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
110 static inline struct sched_clock_data *this_scd(void)
112 return this_cpu_ptr(&sched_clock_data);
115 static inline struct sched_clock_data *cpu_sdc(int cpu)
117 return &per_cpu(sched_clock_data, cpu);
120 int sched_clock_stable(void)
122 return static_branch_likely(&__sched_clock_stable);
125 static void __set_sched_clock_stable(void)
127 struct sched_clock_data *scd = this_scd();
130 * Attempt to make the (initial) unstable->stable transition continuous.
132 raw_offset = (scd->tick_gtod + gtod_offset) - (scd->tick_raw);
134 printk(KERN_INFO "sched_clock: Marking stable (%lld, %lld)->(%lld, %lld)\n",
135 scd->tick_gtod, gtod_offset,
136 scd->tick_raw, raw_offset);
138 static_branch_enable(&__sched_clock_stable);
139 tick_dep_clear(TICK_DEP_BIT_CLOCK_UNSTABLE);
142 static void __clear_sched_clock_stable(struct work_struct *work)
144 struct sched_clock_data *scd = this_scd();
147 * Attempt to make the stable->unstable transition continuous.
149 * Trouble is, this is typically called from the TSC watchdog
150 * timer, which is late per definition. This means the tick
151 * values can already be screwy.
153 * Still do what we can.
155 gtod_offset = (scd->tick_raw + raw_offset) - (scd->tick_gtod);
157 printk(KERN_INFO "sched_clock: Marking unstable (%lld, %lld)<-(%lld, %lld)\n",
158 scd->tick_gtod, gtod_offset,
159 scd->tick_raw, raw_offset);
161 static_branch_disable(&__sched_clock_stable);
162 tick_dep_set(TICK_DEP_BIT_CLOCK_UNSTABLE);
165 static DECLARE_WORK(sched_clock_work, __clear_sched_clock_stable);
167 void clear_sched_clock_stable(void)
169 __sched_clock_stable_early = 0;
171 smp_mb(); /* matches sched_clock_init_late() */
173 if (sched_clock_running == 2)
174 schedule_work(&sched_clock_work);
177 void sched_clock_init_late(void)
179 sched_clock_running = 2;
181 * Ensure that it is impossible to not do a static_key update.
183 * Either {set,clear}_sched_clock_stable() must see sched_clock_running
184 * and do the update, or we must see their __sched_clock_stable_early
185 * and do the update, or both.
187 smp_mb(); /* matches {set,clear}_sched_clock_stable() */
189 if (__sched_clock_stable_early)
190 __set_sched_clock_stable();
194 * min, max except they take wrapping into account
197 static inline u64 wrap_min(u64 x, u64 y)
199 return (s64)(x - y) < 0 ? x : y;
202 static inline u64 wrap_max(u64 x, u64 y)
204 return (s64)(x - y) > 0 ? x : y;
208 * update the percpu scd from the raw @now value
210 * - filter out backward motion
211 * - use the GTOD tick value to create a window to filter crazy TSC values
213 static u64 sched_clock_local(struct sched_clock_data *scd)
215 u64 now, clock, old_clock, min_clock, max_clock;
220 delta = now - scd->tick_raw;
221 if (unlikely(delta < 0))
224 old_clock = scd->clock;
227 * scd->clock = clamp(scd->tick_gtod + delta,
228 * max(scd->tick_gtod, scd->clock),
229 * scd->tick_gtod + TICK_NSEC);
232 clock = scd->tick_gtod + gtod_offset + delta;
233 min_clock = wrap_max(scd->tick_gtod, old_clock);
234 max_clock = wrap_max(old_clock, scd->tick_gtod + TICK_NSEC);
236 clock = wrap_max(clock, min_clock);
237 clock = wrap_min(clock, max_clock);
239 if (cmpxchg64(&scd->clock, old_clock, clock) != old_clock)
245 static u64 sched_clock_remote(struct sched_clock_data *scd)
247 struct sched_clock_data *my_scd = this_scd();
248 u64 this_clock, remote_clock;
249 u64 *ptr, old_val, val;
251 #if BITS_PER_LONG != 64
254 * Careful here: The local and the remote clock values need to
255 * be read out atomic as we need to compare the values and
256 * then update either the local or the remote side. So the
257 * cmpxchg64 below only protects one readout.
259 * We must reread via sched_clock_local() in the retry case on
260 * 32bit as an NMI could use sched_clock_local() via the
261 * tracer and hit between the readout of
262 * the low32bit and the high 32bit portion.
264 this_clock = sched_clock_local(my_scd);
266 * We must enforce atomic readout on 32bit, otherwise the
267 * update on the remote cpu can hit inbetween the readout of
268 * the low32bit and the high 32bit portion.
270 remote_clock = cmpxchg64(&scd->clock, 0, 0);
273 * On 64bit the read of [my]scd->clock is atomic versus the
274 * update, so we can avoid the above 32bit dance.
276 sched_clock_local(my_scd);
278 this_clock = my_scd->clock;
279 remote_clock = scd->clock;
283 * Use the opportunity that we have both locks
284 * taken to couple the two clocks: we take the
285 * larger time as the latest time for both
286 * runqueues. (this creates monotonic movement)
288 if (likely((s64)(remote_clock - this_clock) < 0)) {
290 old_val = remote_clock;
294 * Should be rare, but possible:
296 ptr = &my_scd->clock;
297 old_val = this_clock;
301 if (cmpxchg64(ptr, old_val, val) != old_val)
308 * Similar to cpu_clock(), but requires local IRQs to be disabled.
312 u64 sched_clock_cpu(int cpu)
314 struct sched_clock_data *scd;
317 if (sched_clock_stable())
318 return sched_clock() + raw_offset;
320 if (unlikely(!sched_clock_running))
323 preempt_disable_notrace();
326 if (cpu != smp_processor_id())
327 clock = sched_clock_remote(scd);
329 clock = sched_clock_local(scd);
330 preempt_enable_notrace();
334 EXPORT_SYMBOL_GPL(sched_clock_cpu);
336 void sched_clock_tick(void)
338 struct sched_clock_data *scd;
340 WARN_ON_ONCE(!irqs_disabled());
343 * Update these values even if sched_clock_stable(), because it can
344 * become unstable at any point in time at which point we need some
345 * values to fall back on.
347 * XXX arguably we can skip this if we expose tsc_clocksource_reliable
350 scd->tick_raw = sched_clock();
351 scd->tick_gtod = ktime_get_ns();
353 if (!sched_clock_stable() && likely(sched_clock_running))
354 sched_clock_local(scd);
358 * We are going deep-idle (irqs are disabled):
360 void sched_clock_idle_sleep_event(void)
362 sched_clock_cpu(smp_processor_id());
364 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
367 * We just idled delta nanoseconds (called with irqs disabled):
369 void sched_clock_idle_wakeup_event(u64 delta_ns)
371 if (timekeeping_suspended)
375 touch_softlockup_watchdog_sched();
377 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
379 #else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
381 u64 sched_clock_cpu(int cpu)
383 if (unlikely(!sched_clock_running))
386 return sched_clock();
389 #endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
392 * Running clock - returns the time that has elapsed while a guest has been
394 * On a guest this value should be local_clock minus the time the guest was
395 * suspended by the hypervisor (for any reason).
396 * On bare metal this function should return the same as local_clock.
397 * Architectures and sub-architectures can override this.
399 u64 __weak running_clock(void)
401 return local_clock();