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         int resched;
  36
  37         /*
  38          * sched_clock() is an architecture implemented, fast, scalable,
  39          * lockless clock. It is not guaranteed to be coherent across
  40          * CPUs, nor across CPU idle events.
  41          */
  42         resched = ftrace_preempt_disable();
  43         clock = sched_clock();
  44         ftrace_preempt_enable(resched);
  45
  46         return clock;
  47 }
  48
  49 /*
  50  * trace_clock(): 'inbetween' trace clock. Not completely serialized,
  51  * but not completely incorrect when crossing CPUs either.
  52  *
  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.
  56  */
  57 u64 notrace trace_clock(void)
  58 {
  59         return cpu_clock(raw_smp_processor_id());
  60 }
  61
  62
  63 /*
  64  * trace_clock_global(): special globally coherent trace clock
  65  *
  66  * It has higher overhead than the other trace clocks but is still
  67  * an order of magnitude faster than GTOD derived hardware clocks.
  68  *
  69  * Used by plugins that need globally coherent timestamps.
  70  */
  71
  72 /* keep prev_time and lock in the same cacheline. */
  73 static struct {
  74         u64 prev_time;
  75         arch_spinlock_t lock;
  76 } trace_clock_struct ____cacheline_aligned_in_smp =
  77         {
  78                 .lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED,
  79         };
  80
  81 u64 notrace trace_clock_global(void)
  82 {
  83         unsigned long flags;
  84         int this_cpu;
  85         u64 now;
  86
  87         local_irq_save(flags);
  88
  89         this_cpu = raw_smp_processor_id();
  90         now = cpu_clock(this_cpu);
  91         /*
  92          * If in an NMI context then dont risk lockups and return the
  93          * cpu_clock() time:
  94          */
  95         if (unlikely(in_nmi()))
  96                 goto out;
  97
  98         arch_spin_lock(&trace_clock_struct.lock);
  99
 100         /*
 101          * TODO: if this happens often then maybe we should reset
 102          * my_scd->clock to prev_time+1, to make sure
 103          * we start ticking with the local clock from now on?
 104          */
 105         if ((s64)(now - trace_clock_struct.prev_time) < 0)
 106                 now = trace_clock_struct.prev_time + 1;
 107
 108         trace_clock_struct.prev_time = now;
 109
 110         arch_spin_unlock(&trace_clock_struct.lock);
 111
 112  out:
 113         local_irq_restore(flags);
 114
 115         return now;
 116 }