trace_clock.c 3.0 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126
  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. * trace_clock_local(): the simplest and least coherent tracing clock.
  25. *
  26. * Useful for tracing that does not cross to other CPUs nor
  27. * does it go through idle events.
  28. */
  29. u64 notrace trace_clock_local(void)
  30. {
  31. u64 clock;
  32. /*
  33. * sched_clock() is an architecture implemented, fast, scalable,
  34. * lockless clock. It is not guaranteed to be coherent across
  35. * CPUs, nor across CPU idle events.
  36. */
  37. preempt_disable_notrace();
  38. clock = sched_clock();
  39. preempt_enable_notrace();
  40. return clock;
  41. }
  42. /*
  43. * trace_clock(): 'between' trace clock. Not completely serialized,
  44. * but not completely incorrect when crossing CPUs either.
  45. *
  46. * This is based on cpu_clock(), which will allow at most ~1 jiffy of
  47. * jitter between CPUs. So it's a pretty scalable clock, but there
  48. * can be offsets in the trace data.
  49. */
  50. u64 notrace trace_clock(void)
  51. {
  52. return local_clock();
  53. }
  54. /*
  55. * trace_clock_global(): special globally coherent trace clock
  56. *
  57. * It has higher overhead than the other trace clocks but is still
  58. * an order of magnitude faster than GTOD derived hardware clocks.
  59. *
  60. * Used by plugins that need globally coherent timestamps.
  61. */
  62. /* keep prev_time and lock in the same cacheline. */
  63. static struct {
  64. u64 prev_time;
  65. arch_spinlock_t lock;
  66. } trace_clock_struct ____cacheline_aligned_in_smp =
  67. {
  68. .lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED,
  69. };
  70. u64 notrace trace_clock_global(void)
  71. {
  72. unsigned long flags;
  73. int this_cpu;
  74. u64 now;
  75. local_irq_save(flags);
  76. this_cpu = raw_smp_processor_id();
  77. now = cpu_clock(this_cpu);
  78. /*
  79. * If in an NMI context then dont risk lockups and return the
  80. * cpu_clock() time:
  81. */
  82. if (unlikely(in_nmi()))
  83. goto out;
  84. arch_spin_lock(&trace_clock_struct.lock);
  85. /*
  86. * TODO: if this happens often then maybe we should reset
  87. * my_scd->clock to prev_time+1, to make sure
  88. * we start ticking with the local clock from now on?
  89. */
  90. if ((s64)(now - trace_clock_struct.prev_time) < 0)
  91. now = trace_clock_struct.prev_time + 1;
  92. trace_clock_struct.prev_time = now;
  93. arch_spin_unlock(&trace_clock_struct.lock);
  94. out:
  95. local_irq_restore(flags);
  96. return now;
  97. }
  98. static atomic64_t trace_counter;
  99. /*
  100. * trace_clock_counter(): simply an atomic counter.
  101. * Use the trace_counter "counter" for cases where you do not care
  102. * about timings, but are interested in strict ordering.
  103. */
  104. u64 notrace trace_clock_counter(void)
  105. {
  106. return atomic64_add_return(1, &trace_counter);
  107. }