1 // SPDX-License-Identifier: GPL-2.0
5 * Copyright (C) 2009 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
7 * Implements 3 trace clock variants, with differing scalability/precision
10 * - local: CPU-local trace clock
11 * - medium: scalable global clock with some jitter
12 * - global: globally monotonic, serialized clock
14 * Tracer plugins will chose a default from these clocks.
16 #include <linux/spinlock.h>
17 #include <linux/irqflags.h>
18 #include <linux/hardirq.h>
19 #include <linux/module.h>
20 #include <linux/percpu.h>
21 #include <linux/sched.h>
22 #include <linux/sched/clock.h>
23 #include <linux/ktime.h>
24 #include <linux/trace_clock.h>
27 * trace_clock_local(): the simplest and least coherent tracing clock.
29 * Useful for tracing that does not cross to other CPUs nor
30 * does it go through idle events.
32 u64 notrace trace_clock_local(void)
37 * sched_clock() is an architecture implemented, fast, scalable,
38 * lockless clock. It is not guaranteed to be coherent across
39 * CPUs, nor across CPU idle events.
41 preempt_disable_notrace();
42 clock = sched_clock();
43 preempt_enable_notrace();
47 EXPORT_SYMBOL_GPL(trace_clock_local);
50 * trace_clock(): 'between' trace clock. Not completely serialized,
51 * but not completely incorrect when crossing CPUs either.
53 * This is based on cpu_clock(), which will allow at most ~1 jiffy of
54 * jitter between CPUs. So it's a pretty scalable clock, but there
55 * can be offsets in the trace data.
57 u64 notrace trace_clock(void)
61 EXPORT_SYMBOL_GPL(trace_clock);
64 * trace_jiffy_clock(): Simply use jiffies as a clock counter.
65 * Note that this use of jiffies_64 is not completely safe on
66 * 32-bit systems. But the window is tiny, and the effect if
67 * we are affected is that we will have an obviously bogus
68 * timestamp on a trace event - i.e. not life threatening.
70 u64 notrace trace_clock_jiffies(void)
72 return jiffies_64_to_clock_t(jiffies_64 - INITIAL_JIFFIES);
74 EXPORT_SYMBOL_GPL(trace_clock_jiffies);
77 * trace_clock_global(): special globally coherent trace clock
79 * It has higher overhead than the other trace clocks but is still
80 * an order of magnitude faster than GTOD derived hardware clocks.
82 * Used by plugins that need globally coherent timestamps.
85 /* keep prev_time and lock in the same cacheline. */
89 } trace_clock_struct ____cacheline_aligned_in_smp =
91 .lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED,
94 u64 notrace trace_clock_global(void)
100 raw_local_irq_save(flags);
102 this_cpu = raw_smp_processor_id();
105 * The global clock "guarantees" that the events are ordered
106 * between CPUs. But if two events on two different CPUS call
107 * trace_clock_global at roughly the same time, it really does
108 * not matter which one gets the earlier time. Just make sure
109 * that the same CPU will always show a monotonic clock.
111 * Use a read memory barrier to get the latest written
112 * time that was recorded.
115 prev_time = READ_ONCE(trace_clock_struct.prev_time);
116 now = sched_clock_cpu(this_cpu);
118 /* Make sure that now is always greater than or equal to prev_time */
119 if ((s64)(now - prev_time) < 0)
123 * If in an NMI context then dont risk lockups and simply return
126 if (unlikely(in_nmi()))
129 /* Tracing can cause strange recursion, always use a try lock */
130 if (arch_spin_trylock(&trace_clock_struct.lock)) {
131 /* Reread prev_time in case it was already updated */
132 prev_time = READ_ONCE(trace_clock_struct.prev_time);
133 if ((s64)(now - prev_time) < 0)
136 trace_clock_struct.prev_time = now;
138 /* The unlock acts as the wmb for the above rmb */
139 arch_spin_unlock(&trace_clock_struct.lock);
142 raw_local_irq_restore(flags);
146 EXPORT_SYMBOL_GPL(trace_clock_global);
148 static atomic64_t trace_counter;
151 * trace_clock_counter(): simply an atomic counter.
152 * Use the trace_counter "counter" for cases where you do not care
153 * about timings, but are interested in strict ordering.
155 u64 notrace trace_clock_counter(void)
157 return atomic64_add_return(1, &trace_counter);