kernel/trace/trace_clock.c

0001 // SPDX-License-Identifier: GPL-2.0
0002 /*
0003  * tracing clocks
0004  *
0005  *  Copyright (C) 2009 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
0006  *
0007  * Implements 3 trace clock variants, with differing scalability/precision
0008  * tradeoffs:
0009  *
0010  *  -   local: CPU-local trace clock
0011  *  -  medium: scalable global clock with some jitter
0012  *  -  global: globally monotonic, serialized clock
0013  *
0014  * Tracer plugins will chose a default from these clocks.
0015  */
0016 #include <linux/spinlock.h>
0017 #include <linux/irqflags.h>
0018 #include <linux/hardirq.h>
0019 #include <linux/module.h>
0020 #include <linux/percpu.h>
0021 #include <linux/sched.h>
0022 #include <linux/sched/clock.h>
0023 #include <linux/ktime.h>
0024 #include <linux/trace_clock.h>
0025
0026 /*
0027  * trace_clock_local(): the simplest and least coherent tracing clock.
0028  *
0029  * Useful for tracing that does not cross to other CPUs nor
0030  * does it go through idle events.
0031  */
0032 u64 notrace trace_clock_local(void)
0033 {
0034     u64 clock;
0035
0036     /*
0037      * sched_clock() is an architecture implemented, fast, scalable,
0038      * lockless clock. It is not guaranteed to be coherent across
0039      * CPUs, nor across CPU idle events.
0040      */
0041     preempt_disable_notrace();
0042     clock = sched_clock();
0043     preempt_enable_notrace();
0044
0045     return clock;
0046 }
0047 EXPORT_SYMBOL_GPL(trace_clock_local);
0048
0049 /*
0050  * trace_clock(): 'between' trace clock. Not completely serialized,
0051  * but not completely incorrect when crossing CPUs either.
0052  *
0053  * This is based on cpu_clock(), which will allow at most ~1 jiffy of
0054  * jitter between CPUs. So it's a pretty scalable clock, but there
0055  * can be offsets in the trace data.
0056  */
0057 u64 notrace trace_clock(void)
0058 {
0059     return local_clock();
0060 }
0061 EXPORT_SYMBOL_GPL(trace_clock);
0062
0063 /*
0064  * trace_jiffy_clock(): Simply use jiffies as a clock counter.
0065  * Note that this use of jiffies_64 is not completely safe on
0066  * 32-bit systems. But the window is tiny, and the effect if
0067  * we are affected is that we will have an obviously bogus
0068  * timestamp on a trace event - i.e. not life threatening.
0069  */
0070 u64 notrace trace_clock_jiffies(void)
0071 {
0072     return jiffies_64_to_clock_t(jiffies_64 - INITIAL_JIFFIES);
0073 }
0074 EXPORT_SYMBOL_GPL(trace_clock_jiffies);
0075
0076 /*
0077  * trace_clock_global(): special globally coherent trace clock
0078  *
0079  * It has higher overhead than the other trace clocks but is still
0080  * an order of magnitude faster than GTOD derived hardware clocks.
0081  *
0082  * Used by plugins that need globally coherent timestamps.
0083  */
0084
0085 /* keep prev_time and lock in the same cacheline. */
0086 static struct {
0087     u64 prev_time;
0088     arch_spinlock_t lock;
0089 } trace_clock_struct ____cacheline_aligned_in_smp =
0090     {
0091         .lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED,
0092     };
0093
0094 u64 notrace trace_clock_global(void)
0095 {
0096     unsigned long flags;
0097     int this_cpu;
0098     u64 now, prev_time;
0099
0100     raw_local_irq_save(flags);
0101
0102     this_cpu = raw_smp_processor_id();
0103
0104     /*
0105      * The global clock "guarantees" that the events are ordered
0106      * between CPUs. But if two events on two different CPUS call
0107      * trace_clock_global at roughly the same time, it really does
0108      * not matter which one gets the earlier time. Just make sure
0109      * that the same CPU will always show a monotonic clock.
0110      *
0111      * Use a read memory barrier to get the latest written
0112      * time that was recorded.
0113      */
0114     smp_rmb();
0115     prev_time = READ_ONCE(trace_clock_struct.prev_time);
0116     now = sched_clock_cpu(this_cpu);
0117
0118     /* Make sure that now is always greater than or equal to prev_time */
0119     if ((s64)(now - prev_time) < 0)
0120         now = prev_time;
0121
0122     /*
0123      * If in an NMI context then dont risk lockups and simply return
0124      * the current time.
0125      */
0126     if (unlikely(in_nmi()))
0127         goto out;
0128
0129     /* Tracing can cause strange recursion, always use a try lock */
0130     if (arch_spin_trylock(&trace_clock_struct.lock)) {
0131         /* Reread prev_time in case it was already updated */
0132         prev_time = READ_ONCE(trace_clock_struct.prev_time);
0133         if ((s64)(now - prev_time) < 0)
0134             now = prev_time;
0135
0136         trace_clock_struct.prev_time = now;
0137
0138         /* The unlock acts as the wmb for the above rmb */
0139         arch_spin_unlock(&trace_clock_struct.lock);
0140     }
0141  out:
0142     raw_local_irq_restore(flags);
0143
0144     return now;
0145 }
0146 EXPORT_SYMBOL_GPL(trace_clock_global);
0147
0148 static atomic64_t trace_counter;
0149
0150 /*
0151  * trace_clock_counter(): simply an atomic counter.
0152  * Use the trace_counter "counter" for cases where you do not care
0153  * about timings, but are interested in strict ordering.
0154  */
0155 u64 notrace trace_clock_counter(void)
0156 {
0157     return atomic64_add_return(1, &trace_counter);
0158 }