include/internal/tsan_assist.h

   1 /*
   2  * Copyright 2018-2022 The OpenSSL Project Authors. All Rights Reserved.
   3  *
   4  * Licensed under the Apache License 2.0 (the "License").  You may not use
   5  * this file except in compliance with the License.  You can obtain a copy
   6  * in the file LICENSE in the source distribution or at
   7  * https://www.openssl.org/source/license.html
   8  */
   9
  10 /*
  11  * Contemporary compilers implement lock-free atomic memory access
  12  * primitives that facilitate writing "thread-opportunistic" or even real
  13  * multi-threading low-overhead code. "Thread-opportunistic" is when
  14  * exact result is not required, e.g. some statistics, or execution flow
  15  * doesn't have to be unambiguous. Simplest example is lazy "constant"
  16  * initialization when one can synchronize on variable itself, e.g.
  17  *
  18  * if (var == NOT_YET_INITIALIZED)
  19  *     var = function_returning_same_value();
  20  *
  21  * This does work provided that loads and stores are single-instruction
  22  * operations (and integer ones are on *all* supported platforms), but
  23  * it upsets Thread Sanitizer. Suggested solution is
  24  *
  25  * if (tsan_load(&var) == NOT_YET_INITIALIZED)
  26  *     tsan_store(&var, function_returning_same_value());
  27  *
  28  * Production machine code would be the same, so one can wonder why
  29  * bother. Having Thread Sanitizer accept "thread-opportunistic" code
  30  * allows to move on trouble-shooting real bugs.
  31  *
  32  * Resolving Thread Sanitizer nits was the initial purpose for this module,
  33  * but it was later extended with more nuanced primitives that are useful
  34  * even in "non-opportunistic" scenarios. Most notably verifying if a shared
  35  * structure is fully initialized and bypassing the initialization lock.
  36  * It's suggested to view macros defined in this module as "annotations" for
  37  * thread-safe lock-free code, "Thread-Safe ANnotations"...
  38  *
  39  * It's assumed that ATOMIC_{LONG|INT}_LOCK_FREE are assigned same value as
  40  * ATOMIC_POINTER_LOCK_FREE. And check for >= 2 ensures that corresponding
  41  * code is inlined. It should be noted that statistics counters become
  42  * accurate in such case.
  43  *
  44  * Special note about TSAN_QUALIFIER. It might be undesired to use it in
  45  * a shared header. Because whether operation on specific variable or member
  46  * is atomic or not might be irrelevant in other modules. In such case one
  47  * can use TSAN_QUALIFIER in cast specifically when it has to count.
  48  */
  49
  50 #ifndef OSSL_INTERNAL_TSAN_ASSIST_H
  51 # define OSSL_INTERNAL_TSAN_ASSIST_H
  52 # pragma once
  53
  54 # if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L \
  55     && !defined(__STDC_NO_ATOMICS__)
  56 #  include <stdatomic.h>
  57
  58 #  if defined(ATOMIC_POINTER_LOCK_FREE) \
  59           && ATOMIC_POINTER_LOCK_FREE >= 2
  60 #   define TSAN_QUALIFIER _Atomic
  61 #   define tsan_load(ptr) atomic_load_explicit((ptr), memory_order_relaxed)
  62 #   define tsan_store(ptr, val) atomic_store_explicit((ptr), (val), memory_order_relaxed)
  63 #   define tsan_add(ptr, n) atomic_fetch_add_explicit((ptr), (n), memory_order_relaxed)
  64 #   define tsan_ld_acq(ptr) atomic_load_explicit((ptr), memory_order_acquire)
  65 #   define tsan_st_rel(ptr, val) atomic_store_explicit((ptr), (val), memory_order_release)
  66 #  endif
  67
  68 # elif defined(__GNUC__) && defined(__ATOMIC_RELAXED)
  69
  70 #  if defined(__GCC_ATOMIC_POINTER_LOCK_FREE) \
  71           && __GCC_ATOMIC_POINTER_LOCK_FREE >= 2
  72 #   define TSAN_QUALIFIER volatile
  73 #   define tsan_load(ptr) __atomic_load_n((ptr), __ATOMIC_RELAXED)
  74 #   define tsan_store(ptr, val) __atomic_store_n((ptr), (val), __ATOMIC_RELAXED)
  75 #   define tsan_add(ptr, n) __atomic_fetch_add((ptr), (n), __ATOMIC_RELAXED)
  76 #   define tsan_ld_acq(ptr) __atomic_load_n((ptr), __ATOMIC_ACQUIRE)
  77 #   define tsan_st_rel(ptr, val) __atomic_store_n((ptr), (val), __ATOMIC_RELEASE)
  78 #  endif
  79
  80 # elif defined(_MSC_VER) && _MSC_VER>=1200 \
  81       && (defined(_M_IX86) || defined(_M_AMD64) || defined(_M_X64) || \
  82           defined(_M_ARM64) || (defined(_M_ARM) && _M_ARM >= 7 && !defined(_WIN32_WCE)))
  83 /*
  84  * There is subtle dependency on /volatile:<iso|ms> command-line option.
  85  * "ms" implies same semantic as memory_order_acquire for loads and
  86  * memory_order_release for stores, while "iso" - memory_order_relaxed for
  87  * either. Real complication is that defaults are different on x86 and ARM.
  88  * There is explanation for that, "ms" is backward compatible with earlier
  89  * compiler versions, while multi-processor ARM can be viewed as brand new
  90  * platform to MSC and its users, and with non-relaxed semantic taking toll
  91  * with additional instructions and penalties, it kind of makes sense to
  92  * default to "iso"...
  93  */
  94 #  define TSAN_QUALIFIER volatile
  95 #  if defined(_M_ARM) || defined(_M_ARM64)
  96 #   define _InterlockedExchangeAdd _InterlockedExchangeAdd_nf
  97 #   pragma intrinsic(_InterlockedExchangeAdd_nf)
  98 #   pragma intrinsic(__iso_volatile_load32, __iso_volatile_store32)
  99 #   ifdef _WIN64
 100 #    define _InterlockedExchangeAdd64 _InterlockedExchangeAdd64_nf
 101 #    pragma intrinsic(_InterlockedExchangeAdd64_nf)
 102 #    pragma intrinsic(__iso_volatile_load64, __iso_volatile_store64)
 103 #    define tsan_load(ptr) (sizeof(*(ptr)) == 8 ? __iso_volatile_load64(ptr) \
 104                                                : __iso_volatile_load32(ptr))
 105 #    define tsan_store(ptr, val) (sizeof(*(ptr)) == 8 ? __iso_volatile_store64((ptr), (val)) \
 106                                                      : __iso_volatile_store32((ptr), (val)))
 107 #   else
 108 #    define tsan_load(ptr) __iso_volatile_load32(ptr)
 109 #    define tsan_store(ptr, val) __iso_volatile_store32((ptr), (val))
 110 #   endif
 111 #  else
 112 #   define tsan_load(ptr) (*(ptr))
 113 #   define tsan_store(ptr, val) (*(ptr) = (val))
 114 #  endif
 115 #  pragma intrinsic(_InterlockedExchangeAdd)
 116 #  ifdef _WIN64
 117 #   pragma intrinsic(_InterlockedExchangeAdd64)
 118 #   define tsan_add(ptr, n) (sizeof(*(ptr)) == 8 ? _InterlockedExchangeAdd64((ptr), (n)) \
 119                                                 : _InterlockedExchangeAdd((ptr), (n)))
 120 #  else
 121 #   define tsan_add(ptr, n) _InterlockedExchangeAdd((ptr), (n))
 122 #  endif
 123 #  if !defined(_ISO_VOLATILE)
 124 #   define tsan_ld_acq(ptr) (*(ptr))
 125 #   define tsan_st_rel(ptr, val) (*(ptr) = (val))
 126 #  endif
 127
 128 # endif
 129
 130 # ifndef TSAN_QUALIFIER
 131
 132 #  ifdef OPENSSL_THREADS
 133 #   define TSAN_QUALIFIER volatile
 134 #   define TSAN_REQUIRES_LOCKING
 135 #  else  /* OPENSSL_THREADS */
 136 #   define TSAN_QUALIFIER
 137 #  endif /* OPENSSL_THREADS */
 138
 139 #  define tsan_load(ptr) (*(ptr))
 140 #  define tsan_store(ptr, val) (*(ptr) = (val))
 141 #  define tsan_add(ptr, n) (*(ptr) += (n))
 142 /*
 143  * Lack of tsan_ld_acq and tsan_ld_rel means that compiler support is not
 144  * sophisticated enough to support them. Code that relies on them should be
 145  * protected with #ifdef tsan_ld_acq with locked fallback.
 146  */
 147
 148 # endif
 149
 150 # define tsan_counter(ptr) tsan_add((ptr), 1)
 151 # define tsan_decr(ptr) tsan_add((ptr), -1)
 152
 153 #endif