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 /*
  25  * trace_clock_local(): the simplest and least coherent tracing clock.
  26  *
  27  * Useful for tracing that does not cross to other CPUs nor
  28  * does it go through idle events.
  29  */
  30 u64 notrace trace_clock_local(void)
  31 {
  32         u64 clock;
  33
  34         /*
  35          * sched_clock() is an architecture implemented, fast, scalable,
  36          * lockless clock. It is not guaranteed to be coherent across
  37          * CPUs, nor across CPU idle events.
  38          */
  39         preempt_disable_notrace();
  40         clock = sched_clock();
  41         preempt_enable_notrace();
  42
  43         return clock;
  44 }
  45 EXPORT_SYMBOL_GPL(trace_clock_local);
  46
  47 /*
  48  * trace_clock(): 'between' trace clock. Not completely serialized,
  49  * but not completely incorrect when crossing CPUs either.
  50  *
  51  * This is based on cpu_clock(), which will allow at most ~1 jiffy of
  52  * jitter between CPUs. So it's a pretty scalable clock, but there
  53  * can be offsets in the trace data.
  54  */
  55 u64 notrace trace_clock(void)
  56 {
  57         return local_clock();
  58 }
  59
  60
  61 /*
  62  * trace_clock_global(): special globally coherent trace clock
  63  *
  64  * It has higher overhead than the other trace clocks but is still
  65  * an order of magnitude faster than GTOD derived hardware clocks.
  66  *
  67  * Used by plugins that need globally coherent timestamps.
  68  */
  69
  70 /* keep prev_time and lock in the same cacheline. */
  71 static struct {
  72         u64 prev_time;
  73         arch_spinlock_t lock;
  74 } trace_clock_struct ____cacheline_aligned_in_smp =
  75         {
  76                 .lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED,
  77         };
  78
  79 u64 notrace trace_clock_global(void)
  80 {
  81         unsigned long flags;
  82         int this_cpu;
  83         u64 now;
  84
  85         local_irq_save(flags);
  86
  87         this_cpu = raw_smp_processor_id();
  88         now = sched_clock_cpu(this_cpu);
  89         /*
  90          * If in an NMI context then dont risk lockups and return the
  91          * cpu_clock() time:
  92          */
  93         if (unlikely(in_nmi()))
  94                 goto out;
  95
  96         arch_spin_lock(&trace_clock_struct.lock);
  97
  98         /*
  99          * TODO: if this happens often then maybe we should reset
 100          * my_scd->clock to prev_time+1, to make sure
 101          * we start ticking with the local clock from now on?
 102          */
 103         if ((s64)(now - trace_clock_struct.prev_time) < 0)
 104                 now = trace_clock_struct.prev_time + 1;
 105
 106         trace_clock_struct.prev_time = now;
 107
 108         arch_spin_unlock(&trace_clock_struct.lock);
 109
 110  out:
 111         local_irq_restore(flags);
 112
 113         return now;
 114 }
 115
 116 static atomic64_t trace_counter;
 117
 118 /*
 119  * trace_clock_counter(): simply an atomic counter.
 120  * Use the trace_counter "counter" for cases where you do not care
 121  * about timings, but are interested in strict ordering.
 122  */
 123 u64 notrace trace_clock_counter(void)
 124 {
 125         return atomic64_add_return(1, &trace_counter);
 126 }