kernel/trace/trace_clock.c

   1 /*
   2  * tracing clocks
   3  *
   4  *  Copyright (C) 2009 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
   5  *
   6  * Implements 3 trace clock variants, with differing scalability/precision
   7  * tradeoffs:
   8  *
   9  *  -   local: CPU-local trace clock
  10  *  -  medium: scalable global clock with some jitter
  11  *  -  global: globally monotonic, serialized clock
  12  *
  13  * Tracer plugins will chose a default from these clocks.
  14  */
  15 #include <linux/spinlock.h>
  16 #include <linux/irqflags.h>
  17 #include <linux/hardirq.h>
  18 #include <linux/module.h>
  19 #include <linux/percpu.h>
  20 #include <linux/sched.h>
  21 #include <linux/ktime.h>
  22 #include <linux/trace_clock.h>
  23
  24 #include "trace.h"
  25
  26 /*
  27  * trace_clock_local(): the simplest and least coherent tracing clock.
  28  *
  29  * Useful for tracing that does not cross to other CPUs nor
  30  * does it go through idle events.
  31  */
  32 u64 notrace trace_clock_local(void)
  33 {
  34         u64 clock;
  35
  36         /*
  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.
  40          */
  41         preempt_disable_notrace();
  42         clock = sched_clock();
  43         preempt_enable_notrace();
  44
  45         return clock;
  46 }
  47
  48 /*
  49  * trace_clock(): 'between' trace clock. Not completely serialized,
  50  * but not completely incorrect when crossing CPUs either.
  51  *
  52  * This is based on cpu_clock(), which will allow at most ~1 jiffy of
  53  * jitter between CPUs. So it's a pretty scalable clock, but there
  54  * can be offsets in the trace data.
  55  */
  56 u64 notrace trace_clock(void)
  57 {
  58         return local_clock();
  59 }
  60
  61
  62 /*
  63  * trace_clock_global(): special globally coherent trace clock
  64  *
  65  * It has higher overhead than the other trace clocks but is still
  66  * an order of magnitude faster than GTOD derived hardware clocks.
  67  *
  68  * Used by plugins that need globally coherent timestamps.
  69  */
  70
  71 /* keep prev_time and lock in the same cacheline. */
  72 static struct {
  73         u64 prev_time;
  74         arch_spinlock_t lock;
  75 } trace_clock_struct ____cacheline_aligned_in_smp =
  76         {
  77                 .lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED,
  78         };
  79
  80 u64 notrace trace_clock_global(void)
  81 {
  82         unsigned long flags;
  83         int this_cpu;
  84         u64 now;
  85
  86         local_irq_save(flags);
  87
  88         this_cpu = raw_smp_processor_id();
  89         now = cpu_clock(this_cpu);
  90         /*
  91          * If in an NMI context then dont risk lockups and return the
  92          * cpu_clock() time:
  93          */
  94         if (unlikely(in_nmi()))
  95                 goto out;
  96
  97         arch_spin_lock(&trace_clock_struct.lock);
  98
  99         /*
 100          * TODO: if this happens often then maybe we should reset
 101          * my_scd->clock to prev_time+1, to make sure
 102          * we start ticking with the local clock from now on?
 103          */
 104         if ((s64)(now - trace_clock_struct.prev_time) < 0)
 105                 now = trace_clock_struct.prev_time + 1;
 106
 107         trace_clock_struct.prev_time = now;
 108
 109         arch_spin_unlock(&trace_clock_struct.lock);
 110
 111  out:
 112         local_irq_restore(flags);
 113
 114         return now;
 115 }
 116
 117 static atomic64_t trace_counter;
 118
 119 /*
 120  * trace_clock_counter(): simply an atomic counter.
 121  * Use the trace_counter "counter" for cases where you do not care
 122  * about timings, but are interested in strict ordering.
 123  */
 124 u64 notrace trace_clock_counter(void)
 125 {
 126         return atomic64_add_return(1, &trace_counter);
 127 }