cnt32_to_63.h 3.5 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104
  1. /* SPDX-License-Identifier: GPL-2.0-only */
  2. /*
  3. * Extend a 32-bit counter to 63 bits
  4. *
  5. * Author: Nicolas Pitre
  6. * Created: December 3, 2006
  7. * Copyright: MontaVista Software, Inc.
  8. */
  9. #ifndef __LINUX_CNT32_TO_63_H__
  10. #define __LINUX_CNT32_TO_63_H__
  11. #include <linux/compiler.h>
  12. #include <linux/types.h>
  13. #include <asm/byteorder.h>
  14. /* this is used only to give gcc a clue about good code generation */
  15. union cnt32_to_63 {
  16. struct {
  17. #if defined(__LITTLE_ENDIAN)
  18. u32 lo, hi;
  19. #elif defined(__BIG_ENDIAN)
  20. u32 hi, lo;
  21. #endif
  22. };
  23. u64 val;
  24. };
  25. /**
  26. * cnt32_to_63 - Expand a 32-bit counter to a 63-bit counter
  27. * @cnt_lo: The low part of the counter
  28. *
  29. * Many hardware clock counters are only 32 bits wide and therefore have
  30. * a relatively short period making wrap-arounds rather frequent. This
  31. * is a problem when implementing sched_clock() for example, where a 64-bit
  32. * non-wrapping monotonic value is expected to be returned.
  33. *
  34. * To overcome that limitation, let's extend a 32-bit counter to 63 bits
  35. * in a completely lock free fashion. Bits 0 to 31 of the clock are provided
  36. * by the hardware while bits 32 to 62 are stored in memory. The top bit in
  37. * memory is used to synchronize with the hardware clock half-period. When
  38. * the top bit of both counters (hardware and in memory) differ then the
  39. * memory is updated with a new value, incrementing it when the hardware
  40. * counter wraps around.
  41. *
  42. * Because a word store in memory is atomic then the incremented value will
  43. * always be in synch with the top bit indicating to any potential concurrent
  44. * reader if the value in memory is up to date or not with regards to the
  45. * needed increment. And any race in updating the value in memory is harmless
  46. * as the same value would simply be stored more than once.
  47. *
  48. * The restrictions for the algorithm to work properly are:
  49. *
  50. * 1) this code must be called at least once per each half period of the
  51. * 32-bit counter;
  52. *
  53. * 2) this code must not be preempted for a duration longer than the
  54. * 32-bit counter half period minus the longest period between two
  55. * calls to this code;
  56. *
  57. * Those requirements ensure proper update to the state bit in memory.
  58. * This is usually not a problem in practice, but if it is then a kernel
  59. * timer should be scheduled to manage for this code to be executed often
  60. * enough.
  61. *
  62. * And finally:
  63. *
  64. * 3) the cnt_lo argument must be seen as a globally incrementing value,
  65. * meaning that it should be a direct reference to the counter data which
  66. * can be evaluated according to a specific ordering within the macro,
  67. * and not the result of a previous evaluation stored in a variable.
  68. *
  69. * For example, this is wrong:
  70. *
  71. * u32 partial = get_hw_count();
  72. * u64 full = cnt32_to_63(partial);
  73. * return full;
  74. *
  75. * This is fine:
  76. *
  77. * u64 full = cnt32_to_63(get_hw_count());
  78. * return full;
  79. *
  80. * Note that the top bit (bit 63) in the returned value should be considered
  81. * as garbage. It is not cleared here because callers are likely to use a
  82. * multiplier on the returned value which can get rid of the top bit
  83. * implicitly by making the multiplier even, therefore saving on a runtime
  84. * clear-bit instruction. Otherwise caller must remember to clear the top
  85. * bit explicitly.
  86. */
  87. #define cnt32_to_63(cnt_lo) \
  88. ({ \
  89. static u32 __m_cnt_hi; \
  90. union cnt32_to_63 __x; \
  91. __x.hi = __m_cnt_hi; \
  92. smp_rmb(); \
  93. __x.lo = (cnt_lo); \
  94. if (unlikely((s32)(__x.hi ^ __x.lo) < 0)) \
  95. __m_cnt_hi = __x.hi = (__x.hi ^ 0x80000000) + (__x.hi >> 31); \
  96. __x.val; \
  97. })
  98. #endif