include/linux/cnt32_to_63.h

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