2 * sched_clock for unstable cpu clocks
4 * Copyright (C) 2008 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
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>
13 * Create a semi stable clock from a mixture of other events, including:
16 * - explicit idle events
18 * We use gtod as base and the unstable clock deltas. The deltas are filtered,
19 * making it monotonic and keeping it within an expected window.
21 * Furthermore, explicit sleep and wakeup hooks allow us to account for time
22 * that is otherwise invisible (TSC gets stopped).
24 * The clock: sched_clock_cpu() is monotonic per cpu, and should be somewhat
25 * consistent between cpus (never more than 2 jiffies difference).
27 #include <linux/spinlock.h>
28 #include <linux/hardirq.h>
29 #include <linux/module.h>
30 #include <linux/percpu.h>
31 #include <linux/ktime.h>
32 #include <linux/sched.h>
35 * Scheduler clock - returns current time in nanosec units.
36 * This is default implementation.
37 * Architectures and sub-architectures can override this.
39 unsigned long long __attribute__((weak)) sched_clock(void)
41 return (unsigned long long)(jiffies - INITIAL_JIFFIES)
42 * (NSEC_PER_SEC / HZ);
44 EXPORT_SYMBOL_GPL(sched_clock);
46 static __read_mostly int sched_clock_running;
48 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
49 __read_mostly int sched_clock_stable;
51 struct sched_clock_data {
57 static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
59 static inline struct sched_clock_data *this_scd(void)
61 return &__get_cpu_var(sched_clock_data);
64 static inline struct sched_clock_data *cpu_sdc(int cpu)
66 return &per_cpu(sched_clock_data, cpu);
69 void sched_clock_init(void)
71 u64 ktime_now = ktime_to_ns(ktime_get());
74 for_each_possible_cpu(cpu) {
75 struct sched_clock_data *scd = cpu_sdc(cpu);
78 scd->tick_gtod = ktime_now;
79 scd->clock = ktime_now;
82 sched_clock_running = 1;
86 * min, max except they take wrapping into account
89 static inline u64 wrap_min(u64 x, u64 y)
91 return (s64)(x - y) < 0 ? x : y;
94 static inline u64 wrap_max(u64 x, u64 y)
96 return (s64)(x - y) > 0 ? x : y;
100 * update the percpu scd from the raw @now value
102 * - filter out backward motion
103 * - use the GTOD tick value to create a window to filter crazy TSC values
105 static u64 sched_clock_local(struct sched_clock_data *scd)
107 u64 now, clock, old_clock, min_clock, max_clock;
112 delta = now - scd->tick_raw;
113 if (unlikely(delta < 0))
116 old_clock = scd->clock;
119 * scd->clock = clamp(scd->tick_gtod + delta,
120 * max(scd->tick_gtod, scd->clock),
121 * scd->tick_gtod + TICK_NSEC);
124 clock = scd->tick_gtod + delta;
125 min_clock = wrap_max(scd->tick_gtod, old_clock);
126 max_clock = wrap_max(old_clock, scd->tick_gtod + TICK_NSEC);
128 clock = wrap_max(clock, min_clock);
129 clock = wrap_min(clock, max_clock);
131 if (cmpxchg64(&scd->clock, old_clock, clock) != old_clock)
137 static u64 sched_clock_remote(struct sched_clock_data *scd)
139 struct sched_clock_data *my_scd = this_scd();
140 u64 this_clock, remote_clock;
141 u64 *ptr, old_val, val;
143 sched_clock_local(my_scd);
145 this_clock = my_scd->clock;
146 remote_clock = scd->clock;
149 * Use the opportunity that we have both locks
150 * taken to couple the two clocks: we take the
151 * larger time as the latest time for both
152 * runqueues. (this creates monotonic movement)
154 if (likely((s64)(remote_clock - this_clock) < 0)) {
156 old_val = remote_clock;
160 * Should be rare, but possible:
162 ptr = &my_scd->clock;
163 old_val = this_clock;
167 if (cmpxchg64(ptr, old_val, val) != old_val)
173 u64 sched_clock_cpu(int cpu)
175 struct sched_clock_data *scd;
178 WARN_ON_ONCE(!irqs_disabled());
180 if (sched_clock_stable)
181 return sched_clock();
183 if (unlikely(!sched_clock_running))
188 if (cpu != smp_processor_id())
189 clock = sched_clock_remote(scd);
191 clock = sched_clock_local(scd);
196 void sched_clock_tick(void)
198 struct sched_clock_data *scd;
201 if (sched_clock_stable)
204 if (unlikely(!sched_clock_running))
207 WARN_ON_ONCE(!irqs_disabled());
210 now_gtod = ktime_to_ns(ktime_get());
214 scd->tick_gtod = now_gtod;
215 sched_clock_local(scd);
219 * We are going deep-idle (irqs are disabled):
221 void sched_clock_idle_sleep_event(void)
223 sched_clock_cpu(smp_processor_id());
225 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
228 * We just idled delta nanoseconds (called with irqs disabled):
230 void sched_clock_idle_wakeup_event(u64 delta_ns)
232 if (timekeeping_suspended)
236 touch_softlockup_watchdog();
238 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
240 unsigned long long cpu_clock(int cpu)
242 unsigned long long clock;
245 local_irq_save(flags);
246 clock = sched_clock_cpu(cpu);
247 local_irq_restore(flags);
252 #else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
254 void sched_clock_init(void)
256 sched_clock_running = 1;
259 u64 sched_clock_cpu(int cpu)
261 if (unlikely(!sched_clock_running))
264 return sched_clock();
268 unsigned long long cpu_clock(int cpu)
270 return sched_clock_cpu(cpu);
273 #endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
275 EXPORT_SYMBOL_GPL(cpu_clock);