2 * sched_clock for unstable cpu clocks
4 * Copyright (C) 2008 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Ingo Molnar <mingo@redhat.com>
8 * Guillaume Chazarain <guichaz@gmail.com>
10 * Create a semi stable clock from a mixture of other events, including:
14 * - explicit idle events
16 * We use gtod as base and the unstable clock deltas. The deltas are filtered,
17 * making it monotonic and keeping it within an expected window. This window
18 * is set up using jiffies.
20 * Furthermore, explicit sleep and wakeup hooks allow us to account for time
21 * that is otherwise invisible (TSC gets stopped).
23 * The clock: sched_clock_cpu() is monotonic per cpu, and should be somewhat
24 * consistent between cpus (never more than 1 jiffies difference).
26 #include <linux/sched.h>
27 #include <linux/percpu.h>
28 #include <linux/spinlock.h>
29 #include <linux/ktime.h>
30 #include <linux/module.h>
33 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
35 struct sched_clock_data {
37 * Raw spinlock - this is a special case: this might be called
38 * from within instrumentation code so we dont want to do any
39 * instrumentation ourselves.
43 unsigned long prev_jiffies;
50 static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
52 static inline struct sched_clock_data *this_scd(void)
54 return &__get_cpu_var(sched_clock_data);
57 static inline struct sched_clock_data *cpu_sdc(int cpu)
59 return &per_cpu(sched_clock_data, cpu);
62 static __read_mostly int sched_clock_running;
64 void sched_clock_init(void)
66 u64 ktime_now = ktime_to_ns(ktime_get());
67 unsigned long now_jiffies = jiffies;
70 for_each_possible_cpu(cpu) {
71 struct sched_clock_data *scd = cpu_sdc(cpu);
73 scd->lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
74 scd->prev_jiffies = now_jiffies;
77 scd->tick_gtod = ktime_now;
78 scd->clock = ktime_now;
81 sched_clock_running = 1;
85 * update the percpu scd from the raw @now value
87 * - filter out backward motion
88 * - use jiffies to generate a min,max window to clip the raw values
90 static void __update_sched_clock(struct sched_clock_data *scd, u64 now)
92 unsigned long now_jiffies = jiffies;
93 long delta_jiffies = now_jiffies - scd->prev_jiffies;
94 u64 clock = scd->clock;
95 u64 min_clock, max_clock;
96 s64 delta = now - scd->prev_raw;
98 WARN_ON_ONCE(!irqs_disabled());
99 min_clock = scd->tick_gtod + delta_jiffies * TICK_NSEC;
101 if (unlikely(delta < 0)) {
106 max_clock = min_clock + TICK_NSEC;
108 if (unlikely(clock + delta > max_clock)) {
109 if (clock < max_clock)
118 if (unlikely(clock < min_clock))
122 scd->prev_jiffies = now_jiffies;
126 static void lock_double_clock(struct sched_clock_data *data1,
127 struct sched_clock_data *data2)
130 __raw_spin_lock(&data1->lock);
131 __raw_spin_lock(&data2->lock);
133 __raw_spin_lock(&data2->lock);
134 __raw_spin_lock(&data1->lock);
138 u64 sched_clock_cpu(int cpu)
140 struct sched_clock_data *scd = cpu_sdc(cpu);
143 if (unlikely(!sched_clock_running))
146 WARN_ON_ONCE(!irqs_disabled());
149 if (cpu != raw_smp_processor_id()) {
151 * in order to update a remote cpu's clock based on our
152 * unstable raw time rebase it against:
153 * tick_raw (offset between raw counters)
154 * tick_gotd (tick offset between cpus)
156 struct sched_clock_data *my_scd = this_scd();
158 lock_double_clock(scd, my_scd);
160 now -= my_scd->tick_raw;
161 now += scd->tick_raw;
163 now -= my_scd->tick_gtod;
164 now += scd->tick_gtod;
166 __raw_spin_unlock(&my_scd->lock);
168 __raw_spin_lock(&scd->lock);
171 __update_sched_clock(scd, now);
174 __raw_spin_unlock(&scd->lock);
179 void sched_clock_tick(void)
181 struct sched_clock_data *scd = this_scd();
184 if (unlikely(!sched_clock_running))
187 WARN_ON_ONCE(!irqs_disabled());
190 now_gtod = ktime_to_ns(ktime_get());
192 __raw_spin_lock(&scd->lock);
193 __update_sched_clock(scd, now);
195 * update tick_gtod after __update_sched_clock() because that will
196 * already observe 1 new jiffy; adding a new tick_gtod to that would
197 * increase the clock 2 jiffies.
200 scd->tick_gtod = now_gtod;
201 __raw_spin_unlock(&scd->lock);
205 * We are going deep-idle (irqs are disabled):
207 void sched_clock_idle_sleep_event(void)
209 sched_clock_cpu(smp_processor_id());
211 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
214 * We just idled delta nanoseconds (called with irqs disabled):
216 void sched_clock_idle_wakeup_event(u64 delta_ns)
218 struct sched_clock_data *scd = this_scd();
219 u64 now = sched_clock();
222 * Override the previous timestamp and ignore all
223 * sched_clock() deltas that occured while we idled,
224 * and use the PM-provided delta_ns to advance the
227 __raw_spin_lock(&scd->lock);
229 scd->clock += delta_ns;
230 __raw_spin_unlock(&scd->lock);
232 touch_softlockup_watchdog();
234 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
239 * Scheduler clock - returns current time in nanosec units.
240 * This is default implementation.
241 * Architectures and sub-architectures can override this.
243 unsigned long long __attribute__((weak)) sched_clock(void)
245 return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);