trace_clock.c 2.7 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116
  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. #include "trace.h"
  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. * sched_clock() is an architecture implemented, fast, scalable,
  35. * lockless clock. It is not guaranteed to be coherent across
  36. * CPUs, nor across CPU idle events.
  37. */
  38. preempt_disable_notrace();
  39. clock = sched_clock();
  40. preempt_enable_notrace();
  41. return clock;
  42. }
  43. /*
  44. * trace_clock(): 'between' trace clock. Not completely serialized,
  45. * but not completely incorrect when crossing CPUs either.
  46. *
  47. * This is based on cpu_clock(), which will allow at most ~1 jiffy of
  48. * jitter between CPUs. So it's a pretty scalable clock, but there
  49. * can be offsets in the trace data.
  50. */
  51. u64 notrace trace_clock(void)
  52. {
  53. return local_clock();
  54. }
  55. /*
  56. * trace_clock_global(): special globally coherent trace clock
  57. *
  58. * It has higher overhead than the other trace clocks but is still
  59. * an order of magnitude faster than GTOD derived hardware clocks.
  60. *
  61. * Used by plugins that need globally coherent timestamps.
  62. */
  63. /* keep prev_time and lock in the same cacheline. */
  64. static struct {
  65. u64 prev_time;
  66. arch_spinlock_t lock;
  67. } trace_clock_struct ____cacheline_aligned_in_smp =
  68. {
  69. .lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED,
  70. };
  71. u64 notrace trace_clock_global(void)
  72. {
  73. unsigned long flags;
  74. int this_cpu;
  75. u64 now;
  76. local_irq_save(flags);
  77. this_cpu = raw_smp_processor_id();
  78. now = cpu_clock(this_cpu);
  79. /*
  80. * If in an NMI context then dont risk lockups and return the
  81. * cpu_clock() time:
  82. */
  83. if (unlikely(in_nmi()))
  84. goto out;
  85. arch_spin_lock(&trace_clock_struct.lock);
  86. /*
  87. * TODO: if this happens often then maybe we should reset
  88. * my_scd->clock to prev_time+1, to make sure
  89. * we start ticking with the local clock from now on?
  90. */
  91. if ((s64)(now - trace_clock_struct.prev_time) < 0)
  92. now = trace_clock_struct.prev_time + 1;
  93. trace_clock_struct.prev_time = now;
  94. arch_spin_unlock(&trace_clock_struct.lock);
  95. out:
  96. local_irq_restore(flags);
  97. return now;
  98. }