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 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
81 static DEFINE_STATIC_KEY_FALSE(__sched_clock_stable);
82 static int __sched_clock_stable_early;
84 int sched_clock_stable(void)
86 return static_branch_likely(&__sched_clock_stable);
89 static void __set_sched_clock_stable(void)
91 static_branch_enable(&__sched_clock_stable);
92 tick_dep_clear(TICK_DEP_BIT_CLOCK_UNSTABLE);
95 void set_sched_clock_stable(void)
97 __sched_clock_stable_early = 1;
99 smp_mb(); /* matches sched_clock_init() */
101 if (!sched_clock_running)
104 __set_sched_clock_stable();
107 static void __clear_sched_clock_stable(struct work_struct *work)
109 /* XXX worry about clock continuity */
110 static_branch_disable(&__sched_clock_stable);
111 tick_dep_set(TICK_DEP_BIT_CLOCK_UNSTABLE);
114 static DECLARE_WORK(sched_clock_work, __clear_sched_clock_stable);
116 void clear_sched_clock_stable(void)
118 __sched_clock_stable_early = 0;
120 smp_mb(); /* matches sched_clock_init() */
122 if (!sched_clock_running)
125 schedule_work(&sched_clock_work);
128 struct sched_clock_data {
134 static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
136 static inline struct sched_clock_data *this_scd(void)
138 return this_cpu_ptr(&sched_clock_data);
141 static inline struct sched_clock_data *cpu_sdc(int cpu)
143 return &per_cpu(sched_clock_data, cpu);
146 void sched_clock_init(void)
148 u64 ktime_now = ktime_to_ns(ktime_get());
151 for_each_possible_cpu(cpu) {
152 struct sched_clock_data *scd = cpu_sdc(cpu);
155 scd->tick_gtod = ktime_now;
156 scd->clock = ktime_now;
159 sched_clock_running = 1;
162 * Ensure that it is impossible to not do a static_key update.
164 * Either {set,clear}_sched_clock_stable() must see sched_clock_running
165 * and do the update, or we must see their __sched_clock_stable_early
166 * and do the update, or both.
168 smp_mb(); /* matches {set,clear}_sched_clock_stable() */
170 if (__sched_clock_stable_early)
171 __set_sched_clock_stable();
173 __clear_sched_clock_stable(NULL);
177 * min, max except they take wrapping into account
180 static inline u64 wrap_min(u64 x, u64 y)
182 return (s64)(x - y) < 0 ? x : y;
185 static inline u64 wrap_max(u64 x, u64 y)
187 return (s64)(x - y) > 0 ? x : y;
191 * update the percpu scd from the raw @now value
193 * - filter out backward motion
194 * - use the GTOD tick value to create a window to filter crazy TSC values
196 static u64 sched_clock_local(struct sched_clock_data *scd)
198 u64 now, clock, old_clock, min_clock, max_clock;
203 delta = now - scd->tick_raw;
204 if (unlikely(delta < 0))
207 old_clock = scd->clock;
210 * scd->clock = clamp(scd->tick_gtod + delta,
211 * max(scd->tick_gtod, scd->clock),
212 * scd->tick_gtod + TICK_NSEC);
215 clock = scd->tick_gtod + delta;
216 min_clock = wrap_max(scd->tick_gtod, old_clock);
217 max_clock = wrap_max(old_clock, scd->tick_gtod + TICK_NSEC);
219 clock = wrap_max(clock, min_clock);
220 clock = wrap_min(clock, max_clock);
222 if (cmpxchg64(&scd->clock, old_clock, clock) != old_clock)
228 static u64 sched_clock_remote(struct sched_clock_data *scd)
230 struct sched_clock_data *my_scd = this_scd();
231 u64 this_clock, remote_clock;
232 u64 *ptr, old_val, val;
234 #if BITS_PER_LONG != 64
237 * Careful here: The local and the remote clock values need to
238 * be read out atomic as we need to compare the values and
239 * then update either the local or the remote side. So the
240 * cmpxchg64 below only protects one readout.
242 * We must reread via sched_clock_local() in the retry case on
243 * 32bit as an NMI could use sched_clock_local() via the
244 * tracer and hit between the readout of
245 * the low32bit and the high 32bit portion.
247 this_clock = sched_clock_local(my_scd);
249 * We must enforce atomic readout on 32bit, otherwise the
250 * update on the remote cpu can hit inbetween the readout of
251 * the low32bit and the high 32bit portion.
253 remote_clock = cmpxchg64(&scd->clock, 0, 0);
256 * On 64bit the read of [my]scd->clock is atomic versus the
257 * update, so we can avoid the above 32bit dance.
259 sched_clock_local(my_scd);
261 this_clock = my_scd->clock;
262 remote_clock = scd->clock;
266 * Use the opportunity that we have both locks
267 * taken to couple the two clocks: we take the
268 * larger time as the latest time for both
269 * runqueues. (this creates monotonic movement)
271 if (likely((s64)(remote_clock - this_clock) < 0)) {
273 old_val = remote_clock;
277 * Should be rare, but possible:
279 ptr = &my_scd->clock;
280 old_val = this_clock;
284 if (cmpxchg64(ptr, old_val, val) != old_val)
291 * Similar to cpu_clock(), but requires local IRQs to be disabled.
295 u64 sched_clock_cpu(int cpu)
297 struct sched_clock_data *scd;
300 if (sched_clock_stable())
301 return sched_clock();
303 if (unlikely(!sched_clock_running))
306 preempt_disable_notrace();
309 if (cpu != smp_processor_id())
310 clock = sched_clock_remote(scd);
312 clock = sched_clock_local(scd);
313 preempt_enable_notrace();
317 EXPORT_SYMBOL_GPL(sched_clock_cpu);
319 void sched_clock_tick(void)
321 struct sched_clock_data *scd;
324 if (sched_clock_stable())
327 if (unlikely(!sched_clock_running))
330 WARN_ON_ONCE(!irqs_disabled());
333 now_gtod = ktime_to_ns(ktime_get());
337 scd->tick_gtod = now_gtod;
338 sched_clock_local(scd);
342 * We are going deep-idle (irqs are disabled):
344 void sched_clock_idle_sleep_event(void)
346 sched_clock_cpu(smp_processor_id());
348 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
351 * We just idled delta nanoseconds (called with irqs disabled):
353 void sched_clock_idle_wakeup_event(u64 delta_ns)
355 if (timekeeping_suspended)
359 touch_softlockup_watchdog_sched();
361 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
363 #else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
365 void sched_clock_init(void)
367 sched_clock_running = 1;
370 u64 sched_clock_cpu(int cpu)
372 if (unlikely(!sched_clock_running))
375 return sched_clock();
377 #endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
380 * Running clock - returns the time that has elapsed while a guest has been
382 * On a guest this value should be local_clock minus the time the guest was
383 * suspended by the hypervisor (for any reason).
384 * On bare metal this function should return the same as local_clock.
385 * Architectures and sub-architectures can override this.
387 u64 __weak running_clock(void)
389 return local_clock();