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. int resched;
  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. resched = ftrace_preempt_disable();
  40. clock = sched_clock();
  41. ftrace_preempt_enable(resched);
  42. return clock;
  43. }
  44. /*
  45. * trace_clock(): 'inbetween' trace clock. Not completely serialized,
  46. * but not completely incorrect when crossing CPUs either.
  47. *
  48. * This is based on cpu_clock(), which will allow at most ~1 jiffy of
  49. * jitter between CPUs. So it's a pretty scalable clock, but there
  50. * can be offsets in the trace data.
  51. */
  52. u64 notrace trace_clock(void)
  53. {
  54. return cpu_clock(raw_smp_processor_id());
  55. }
  56. /*
  57. * trace_clock_global(): special globally coherent trace clock
  58. *
  59. * It has higher overhead than the other trace clocks but is still
  60. * an order of magnitude faster than GTOD derived hardware clocks.
  61. *
  62. * Used by plugins that need globally coherent timestamps.
  63. */
  64. /* keep prev_time and lock in the same cacheline. */
  65. static struct {
  66. u64 prev_time;
  67. arch_spinlock_t lock;
  68. } trace_clock_struct ____cacheline_aligned_in_smp =
  69. {
  70. .lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED,
  71. };
  72. u64 notrace trace_clock_global(void)
  73. {
  74. unsigned long flags;
  75. int this_cpu;
  76. u64 now;
  77. local_irq_save(flags);
  78. this_cpu = raw_smp_processor_id();
  79. now = cpu_clock(this_cpu);
  80. /*
  81. * If in an NMI context then dont risk lockups and return the
  82. * cpu_clock() time:
  83. */
  84. if (unlikely(in_nmi()))
  85. goto out;
  86. arch_spin_lock(&trace_clock_struct.lock);
  87. /*
  88. * TODO: if this happens often then maybe we should reset
  89. * my_scd->clock to prev_time+1, to make sure
  90. * we start ticking with the local clock from now on?
  91. */
  92. if ((s64)(now - trace_clock_struct.prev_time) < 0)
  93. now = trace_clock_struct.prev_time + 1;
  94. trace_clock_struct.prev_time = now;
  95. arch_spin_unlock(&trace_clock_struct.lock);
  96. out:
  97. local_irq_restore(flags);
  98. return now;
  99. }