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/module.h>
29 #include <linux/percpu.h>
30 #include <linux/ktime.h>
31 #include <linux/sched.h>
34 * Scheduler clock - returns current time in nanosec units.
35 * This is default implementation.
36 * Architectures and sub-architectures can override this.
38 unsigned long long __attribute__((weak)) sched_clock(void)
40 return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
43 static __read_mostly int sched_clock_running;
45 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
46 __read_mostly int sched_clock_stable;
48 struct sched_clock_data {
50 * Raw spinlock - this is a special case: this might be called
51 * from within instrumentation code so we dont want to do any
52 * instrumentation ourselves.
61 static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
63 static inline struct sched_clock_data *this_scd(void)
65 return &__get_cpu_var(sched_clock_data);
68 static inline struct sched_clock_data *cpu_sdc(int cpu)
70 return &per_cpu(sched_clock_data, cpu);
73 void sched_clock_init(void)
75 u64 ktime_now = ktime_to_ns(ktime_get());
78 for_each_possible_cpu(cpu) {
79 struct sched_clock_data *scd = cpu_sdc(cpu);
81 scd->lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
83 scd->tick_gtod = ktime_now;
84 scd->clock = ktime_now;
87 sched_clock_running = 1;
91 * min, max except they take wrapping into account
94 static inline u64 wrap_min(u64 x, u64 y)
96 return (s64)(x - y) < 0 ? x : y;
99 static inline u64 wrap_max(u64 x, u64 y)
101 return (s64)(x - y) > 0 ? x : y;
105 * update the percpu scd from the raw @now value
107 * - filter out backward motion
108 * - use the GTOD tick value to create a window to filter crazy TSC values
110 static u64 __update_sched_clock(struct sched_clock_data *scd, u64 now)
112 s64 delta = now - scd->tick_raw;
113 u64 clock, min_clock, max_clock;
115 if (unlikely(delta < 0))
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, scd->clock);
126 max_clock = wrap_max(scd->clock, scd->tick_gtod + TICK_NSEC);
128 clock = wrap_max(clock, min_clock);
129 clock = wrap_min(clock, max_clock);
136 static void lock_double_clock(struct sched_clock_data *data1,
137 struct sched_clock_data *data2)
140 __raw_spin_lock(&data1->lock);
141 __raw_spin_lock(&data2->lock);
143 __raw_spin_lock(&data2->lock);
144 __raw_spin_lock(&data1->lock);
148 u64 sched_clock_cpu(int cpu)
150 u64 now, clock, this_clock, remote_clock;
151 struct sched_clock_data *scd;
153 if (sched_clock_stable)
154 return sched_clock();
157 WARN_ON_ONCE(!irqs_disabled());
160 if (cpu != raw_smp_processor_id()) {
161 struct sched_clock_data *my_scd = this_scd();
163 lock_double_clock(scd, my_scd);
165 this_clock = __update_sched_clock(my_scd, now);
166 remote_clock = scd->clock;
169 * Use the opportunity that we have both locks
170 * taken to couple the two clocks: we take the
171 * larger time as the latest time for both
172 * runqueues. (this creates monotonic movement)
174 if (likely((s64)(remote_clock - this_clock) < 0)) {
179 * Should be rare, but possible:
181 clock = remote_clock;
182 my_scd->clock = remote_clock;
185 __raw_spin_unlock(&my_scd->lock);
187 __raw_spin_lock(&scd->lock);
188 clock = __update_sched_clock(scd, now);
191 __raw_spin_unlock(&scd->lock);
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());
213 __raw_spin_lock(&scd->lock);
215 scd->tick_gtod = now_gtod;
216 __update_sched_clock(scd, now);
217 __raw_spin_unlock(&scd->lock);
221 * We are going deep-idle (irqs are disabled):
223 void sched_clock_idle_sleep_event(void)
225 sched_clock_cpu(smp_processor_id());
227 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
230 * We just idled delta nanoseconds (called with irqs disabled):
232 void sched_clock_idle_wakeup_event(u64 delta_ns)
234 if (timekeeping_suspended)
238 touch_softlockup_watchdog();
240 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
242 #else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
244 void sched_clock_init(void)
246 sched_clock_running = 1;
249 u64 sched_clock_cpu(int cpu)
251 if (unlikely(!sched_clock_running))
254 return sched_clock();
257 #endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
259 unsigned long long cpu_clock(int cpu)
261 unsigned long long clock;
264 local_irq_save(flags);
265 clock = sched_clock_cpu(cpu);
266 local_irq_restore(flags);
270 EXPORT_SYMBOL_GPL(cpu_clock);