WL#2595 kernel-independent atomic operations

Backport from 6.0.14 to 5.6.0

Original code from Sergei Golubchik

configure.in

@@ -1750,64 +1750,74 @@ then
 fi
 
 AC_ARG_WITH([atomic-ops],
-	    AC_HELP_STRING([--with-atomic-ops=rwlocks|smp|up],
-	    [Implement atomic operations using pthread rwlocks or atomic CPU
-             instructions for multi-processor (default) or uniprocessor
-             configuration]), , [with_atomic_ops=smp])
+            AS_HELP_STRING([--with-atomic-ops=rwlocks|smp|up],
+            [Implement atomic operations using pthread rwlocks or atomic CPU
+             instructions for multi-processor or uniprocessor
+             configuration. By default gcc built-in sync functions are used,
+             if available and 'smp' configuration otherwise.]))
 case "$with_atomic_ops" in
   "up") AC_DEFINE([MY_ATOMIC_MODE_DUMMY], [1],
                   [Assume single-CPU mode, no concurrency]) ;;
   "rwlocks") AC_DEFINE([MY_ATOMIC_MODE_RWLOCKS], [1],
                   [Use pthread rwlocks for atomic ops]) ;;
   "smp") ;;
+  "")
+   ;;
    *) AC_MSG_ERROR(["$with_atomic_ops" is not a valid value for --with-atomic-ops]) ;;
 esac
 
 AC_CACHE_CHECK([whether the compiler provides atomic builtins],
-               [mysql_cv_gcc_atomic_builtins], [AC_TRY_RUN([
-  int main()
-  {
-    int foo= -10; int bar= 10;
-    if (!__sync_fetch_and_add(&foo, bar) || foo)
-      return -1;
-    bar= __sync_lock_test_and_set(&foo, bar);
-    if (bar || foo != 10)
-      return -1;
-    bar= __sync_val_compare_and_swap(&bar, foo, 15);
-    if (bar)
-      return -1;
-    return 0;
-  }
-], [mysql_cv_gcc_atomic_builtins=yes],
+               [mysql_cv_gcc_atomic_builtins],
+  [AC_RUN_IFELSE(
+     [AC_LANG_PROGRAM(
+        [
+        ],
+     [[
+        int foo= -10; int bar= 10;
+        if (!__sync_fetch_and_add(&foo, bar) || foo)
+          return -1;
+        bar= __sync_lock_test_and_set(&foo, bar);
+        if (bar || foo != 10)
+          return -1;
+        bar= __sync_val_compare_and_swap(&bar, foo, 15);
+        if (bar)
+          return -1;
+        return 0;
+     ]]
+     )],
+   [mysql_cv_gcc_atomic_builtins=yes],
    [mysql_cv_gcc_atomic_builtins=no],
-   [mysql_cv_gcc_atomic_builtins=no])])
-
+   [mysql_cv_gcc_atomic_builtins=no]
+)])
 if test "x$mysql_cv_gcc_atomic_builtins" = xyes; then
   AC_DEFINE(HAVE_GCC_ATOMIC_BUILTINS, 1,
             [Define to 1 if compiler provides atomic builtins.])
 fi
 
 AC_CACHE_CHECK([whether the OS provides atomic_* functions like Solaris],
-               [mysql_cv_solaris_atomic], [AC_TRY_RUN([
-#include <atomic.h>
-int
-main()
-{
-	int foo = -10; int bar = 10;
-	if (atomic_add_int_nv((uint_t *)&foo, bar) || foo)
-		return -1;
-	bar = atomic_swap_uint((uint_t *)&foo, (uint_t)bar);
-	if (bar || foo != 10)
-		return -1;
-	bar = atomic_cas_uint((uint_t *)&bar, (uint_t)foo, 15);
-	if (bar)
-		return -1;
-	return 0;
-}
-], [mysql_cv_solaris_atomic=yes],
+               [mysql_cv_solaris_atomic],
+  [AC_RUN_IFELSE(
+     [AC_LANG_PROGRAM(
+        [
+        #include <atomic.h>
+        ]
+     [[
+        int foo = -10; int bar = 10;
+        if (atomic_add_int_nv((uint_t *)&foo, bar) || foo)
+          return -1;
+        bar = atomic_swap_uint((uint_t *)&foo, (uint_t)bar);
+        if (bar || foo != 10)
+          return -1;
+        bar = atomic_cas_uint((uint_t *)&bar, (uint_t)foo, 15);
+        if (bar)
+          return -1;
+        return 0;
+     ]]
+     )],
+   [mysql_cv_solaris_atomic=yes],
    [mysql_cv_solaris_atomic=no],
-   [mysql_cv_solaris_atomic=no])])
-
+   [mysql_cv_solaris_atomic=no]
+)])
 if test "x$mysql_cv_solaris_atomic" = xyes; then
   AC_DEFINE(HAVE_SOLARIS_ATOMIC, 1,
             [Define to 1 if OS provides atomic_* functions like Solaris.])

include/Makefile.am

@@ -39,9 +39,9 @@ noinst_HEADERS =	config-win.h config-netware.h my_bit.h \
 			thr_lock.h t_ctype.h violite.h my_md5.h base64.h \
 			my_handler.h my_time.h service_versions.h \
 			my_vle.h my_user.h my_atomic.h atomic/nolock.h \
-			atomic/rwlock.h atomic/x86-gcc.h atomic/x86-msvc.h \
-                        atomic/solaris.h \
-			atomic/gcc_builtins.h my_libwrap.h my_stacktrace.h
+			atomic/rwlock.h atomic/x86-gcc.h atomic/generic-msvc.h \
+			atomic/gcc_builtins.h my_libwrap.h my_stacktrace.h \
+                        atomic/solaris.h
 
 EXTRA_DIST =        mysql.h.pp mysql/plugin.h.pp probes_mysql.d.base 
 

include/atomic/gcc_builtins.h

@@ -18,7 +18,7 @@
 
 #define make_atomic_add_body(S)                     \
   v= __sync_fetch_and_add(a, v);
-#define make_atomic_swap_body(S)                    \
+#define make_atomic_fas_body(S)                     \
   v= __sync_lock_test_and_set(a, v);
 #define make_atomic_cas_body(S)                     \
   int ## S sav;                                     \
@@ -28,7 +28,10 @@
 #ifdef MY_ATOMIC_MODE_DUMMY
 #define make_atomic_load_body(S)   ret= *a
 #define make_atomic_store_body(S)  *a= v
+#define MY_ATOMIC_MODE "gcc-builtins-up"
+
 #else
+#define MY_ATOMIC_MODE "gcc-builtins-smp"
 #define make_atomic_load_body(S)                    \
   ret= __sync_fetch_and_or(a, 0);
 #define make_atomic_store_body(S)                   \

include/atomic/generic-msvc.h

+/* Copyright (C) 2006-2008 MySQL AB, 2008-2009 Sun Microsystems, Inc.
+
+   This program is free software; you can redistribute it and/or modify
+   it under the terms of the GNU General Public License as published by
+   the Free Software Foundation; version 2 of the License.
+
+   This program is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+   GNU General Public License for more details.
+
+   You should have received a copy of the GNU General Public License
+   along with this program; if not, write to the Free Software
+   Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA */
+
+#ifndef _atomic_h_cleanup_
+#define _atomic_h_cleanup_ "atomic/generic-msvc.h"
+
+/*
+  We don't implement anything specific for MY_ATOMIC_MODE_DUMMY, always use
+  intrinsics.
+  8 and 16-bit atomics are not implemented, but it can be done if necessary.
+*/
+#undef MY_ATOMIC_HAS_8_16
+
+/*
+  x86 compilers (both VS2003 or VS2005) never use instrinsics, but generate 
+  function calls to kernel32 instead, even in the optimized build. 
+  We force intrinsics as described in MSDN documentation for 
+  _InterlockedCompareExchange.
+*/
+#ifdef _M_IX86
+
+#if (_MSC_VER >= 1500)
+#include <intrin.h>
+#else
+C_MODE_START
+/*Visual Studio 2003 and earlier do not have prototypes for atomic intrinsics*/
+LONG _InterlockedExchange (LONG volatile *Target,LONG Value);
+LONG _InterlockedCompareExchange (LONG volatile *Target, LONG Value, LONG Comp);
+LONG _InterlockedExchangeAdd (LONG volatile *Addend, LONG Value);
+C_MODE_END
+
+#pragma intrinsic(_InterlockedExchangeAdd)
+#pragma intrinsic(_InterlockedCompareExchange)
+#pragma intrinsic(_InterlockedExchange)
+#endif
+
+#define InterlockedExchange        _InterlockedExchange
+#define InterlockedExchangeAdd     _InterlockedExchangeAdd
+#define InterlockedCompareExchange _InterlockedCompareExchange
+/*
+ No need to do something special for InterlockedCompareExchangePointer
+ as it is a #define to InterlockedCompareExchange. The same applies to
+ InterlockedExchangePointer. 
+*/
+#endif /*_M_IX86*/
+
+#define MY_ATOMIC_MODE "msvc-intrinsics"
+#define IL_EXCHG_ADD32(X,Y)     InterlockedExchangeAdd((volatile LONG *)(X),(Y))
+#define IL_COMP_EXCHG32(X,Y,Z)  InterlockedCompareExchange((volatile LONG *)(X),(Y),(Z))
+#define IL_COMP_EXCHGptr        InterlockedCompareExchangePointer
+#define IL_EXCHG32(X,Y)         InterlockedExchange((volatile LONG *)(X),(Y))
+#define IL_EXCHGptr             InterlockedExchangePointer
+#define make_atomic_add_body(S) \
+  v= IL_EXCHG_ADD ## S (a, v)
+#define make_atomic_cas_body(S)                                 \
+  int ## S initial_cmp= *cmp;                                   \
+  int ## S initial_a= IL_COMP_EXCHG ## S (a, set, initial_cmp); \
+  if (!(ret= (initial_a == initial_cmp))) *cmp= initial_a;
+#define make_atomic_swap_body(S) \
+  v= IL_EXCHG ## S (a, v)
+#define make_atomic_load_body(S)       \
+  ret= 0; /* avoid compiler warning */ \
+  ret= IL_COMP_EXCHG ## S (a, ret, ret);
+
+/*
+  my_yield_processor (equivalent of x86 PAUSE instruction) should be used
+  to improve performance on hyperthreaded CPUs. Intel recommends to use it in
+  spin loops also on non-HT machines to reduce power consumption (see e.g 
+  http://softwarecommunity.intel.com/articles/eng/2004.htm)
+
+  Running benchmarks for spinlocks implemented with InterlockedCompareExchange
+  and YieldProcessor shows that much better performance is achieved by calling
+  YieldProcessor in a loop - that is, yielding longer. On Intel boxes setting
+  loop count in the range 200-300 brought best results.
+ */
+#ifndef YIELD_LOOPS
+#define YIELD_LOOPS 200
+#endif
+
+static __inline int my_yield_processor()
+{
+  int i;
+  for(i=0; i<YIELD_LOOPS; i++)
+  {
+#if (_MSC_VER <= 1310)
+    /* On older compilers YieldProcessor is not available, use inline assembly*/
+    __asm { rep nop }
+#else
+    YieldProcessor();
+#endif
+  }
+  return 1;
+}
+
+#define LF_BACKOFF my_yield_processor()
+#else /* cleanup */
+
+#undef IL_EXCHG_ADD32
+#undef IL_COMP_EXCHG32
+#undef IL_COMP_EXCHGptr
+#undef IL_EXCHG32
+#undef IL_EXCHGptr
+
+#endif

include/atomic/nolock.h

@@ -16,43 +16,36 @@
    along with this program; if not, write to the Free Software
    Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA */
 
-#if defined(__i386__) || defined(_M_IX86) || defined(HAVE_GCC_ATOMIC_BUILTINS)
-
-#ifdef MY_ATOMIC_MODE_DUMMY
-#  define LOCK ""
-#else
-#  define LOCK "lock"
-#endif
-
-#ifdef HAVE_GCC_ATOMIC_BUILTINS
-#include "gcc_builtins.h"
-#elif __GNUC__
-#include "x86-gcc.h"
-#elif defined(_MSC_VER)
-#include "x86-msvc.h"
-#endif
-
+#if defined(__i386__) || defined(_MSC_VER) || defined(__x86_64__)   \
+    || defined(HAVE_GCC_ATOMIC_BUILTINS)
+
+#  ifdef MY_ATOMIC_MODE_DUMMY
+#    define LOCK_prefix ""
+#  else
+#    define LOCK_prefix "lock"
+#  endif
+
+#  ifdef HAVE_GCC_ATOMIC_BUILTINS
+#    include "gcc_builtins.h"
+#  elif __GNUC__
+#    include "x86-gcc.h"
+#  elif defined(_MSC_VER)
+#    include "generic-msvc.h"
+#  endif
 #elif defined(HAVE_SOLARIS_ATOMIC)
-
 #include "solaris.h"
-
-#endif /* __i386__ || _M_IX86 || HAVE_GCC_ATOMIC_BUILTINS */
+#endif
 
 #if defined(make_atomic_cas_body) || defined(MY_ATOMICS_MADE)
 /*
  * We have atomics that require no locking
  */
 #define	MY_ATOMIC_NOLOCK
-
-#ifdef __SUNPRO_C
 /*
- * Sun Studio 12 (and likely earlier) does not accept a typedef struct {}
- */
-typedef char my_atomic_rwlock_t;
-#else
-typedef struct { } my_atomic_rwlock_t;
-#endif
-
+  Type not used so minimal size (emptry struct has different size between C
+  and C++, zero-length array is gcc-specific).
+*/
+typedef char my_atomic_rwlock_t __attribute__ ((unused));
 #define my_atomic_rwlock_destroy(name)
 #define my_atomic_rwlock_init(name)
 #define my_atomic_rwlock_rdlock(name)

include/atomic/rwlock.h

 #ifndef ATOMIC_RWLOCK_INCLUDED
 #define ATOMIC_RWLOCK_INCLUDED
 
-/* Copyright (C) 2006 MySQL AB
+/* Copyright (C) 2006 MySQL AB, 2009 Sun Microsystems, Inc.
 
    This program is free software; you can redistribute it and/or modify
    it under the terms of the GNU General Public License as published by
@@ -16,7 +16,7 @@
    along with this program; if not, write to the Free Software
    Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA */
 
-typedef struct {pthread_rwlock_t rw;} my_atomic_rwlock_t;
+#define MY_ATOMIC_MODE_RWLOCKS 1
 
 #ifdef MY_ATOMIC_MODE_DUMMY
 /*
@@ -26,6 +26,9 @@ typedef struct {pthread_rwlock_t rw;} my_atomic_rwlock_t;
   implementations (another way is to run a UP build on an SMP box).
 */
 #warning MY_ATOMIC_MODE_DUMMY and MY_ATOMIC_MODE_RWLOCKS are incompatible
+
+typedef char my_atomic_rwlock_t;
+
 #define my_atomic_rwlock_destroy(name)
 #define my_atomic_rwlock_init(name)
 #define my_atomic_rwlock_rdlock(name)
@@ -33,18 +36,63 @@ typedef struct {pthread_rwlock_t rw;} my_atomic_rwlock_t;
 #define my_atomic_rwlock_rdunlock(name)
 #define my_atomic_rwlock_wrunlock(name)
 #define MY_ATOMIC_MODE "dummy (non-atomic)"
-#else
-#define my_atomic_rwlock_destroy(name)     pthread_rwlock_destroy(& (name)->rw)
-#define my_atomic_rwlock_init(name)        pthread_rwlock_init(& (name)->rw, 0)
-#define my_atomic_rwlock_rdlock(name)      pthread_rwlock_rdlock(& (name)->rw)
-#define my_atomic_rwlock_wrlock(name)      pthread_rwlock_wrlock(& (name)->rw)
-#define my_atomic_rwlock_rdunlock(name)    pthread_rwlock_unlock(& (name)->rw)
-#define my_atomic_rwlock_wrunlock(name)    pthread_rwlock_unlock(& (name)->rw)
-#define MY_ATOMIC_MODE "rwlocks"
+#else /* not MY_ATOMIC_MODE_DUMMY */
+
+typedef struct {pthread_mutex_t rw;} my_atomic_rwlock_t;
+
+#ifndef SAFE_MUTEX
+
+/*
+  we're using read-write lock macros but map them to mutex locks, and they're
+  faster. Still, having semantically rich API we can change the
+  underlying implementation, if necessary.
+*/
+#define my_atomic_rwlock_destroy(name)     pthread_mutex_destroy(& (name)->rw)
+#define my_atomic_rwlock_init(name)        pthread_mutex_init(& (name)->rw, 0)
+#define my_atomic_rwlock_rdlock(name)      pthread_mutex_lock(& (name)->rw)
+#define my_atomic_rwlock_wrlock(name)      pthread_mutex_lock(& (name)->rw)
+#define my_atomic_rwlock_rdunlock(name)    pthread_mutex_unlock(& (name)->rw)
+#define my_atomic_rwlock_wrunlock(name)    pthread_mutex_unlock(& (name)->rw)
+
+#else /* SAFE_MUTEX */
+
+/*
+  SAFE_MUTEX pollutes the compiling name space with macros
+  that alter pthread_mutex_t, pthread_mutex_init, etc.
+  Atomic operations should never use the safe mutex wrappers.
+  Unfortunately, there is no way to have both:
+  - safe mutex macros expanding pthread_mutex_lock to safe_mutex_lock
+  - my_atomic macros expanding to unmodified pthread_mutex_lock
+  inlined in the same compilation unit.
+  So, in case of SAFE_MUTEX, a function call is required.
+  Given that SAFE_MUTEX is a debugging facility,
+  this extra function call is not a performance concern for
+  production builds.
+*/
+C_MODE_START
+extern void plain_pthread_mutex_init(safe_mutex_t *);
+extern void plain_pthread_mutex_destroy(safe_mutex_t *);
+extern void plain_pthread_mutex_lock(safe_mutex_t *);
+extern void plain_pthread_mutex_unlock(safe_mutex_t *);
+C_MODE_END
+
+#define my_atomic_rwlock_destroy(name)     plain_pthread_mutex_destroy(&(name)->rw)
+#define my_atomic_rwlock_init(name)        plain_pthread_mutex_init(&(name)->rw)
+#define my_atomic_rwlock_rdlock(name)      plain_pthread_mutex_lock(&(name)->rw)
+#define my_atomic_rwlock_wrlock(name)      plain_pthread_mutex_lock(&(name)->rw)
+#define my_atomic_rwlock_rdunlock(name)    plain_pthread_mutex_unlock(&(name)->rw)
+#define my_atomic_rwlock_wrunlock(name)    plain_pthread_mutex_unlock(&(name)->rw)
+
+#endif /* SAFE_MUTEX */
+
+#define MY_ATOMIC_MODE "mutex"
+#ifndef MY_ATOMIC_MODE_RWLOCKS
+#define MY_ATOMIC_MODE_RWLOCKS 1
+#endif
 #endif
 
 #define make_atomic_add_body(S)     int ## S sav; sav= *a; *a+= v; v=sav;
-#define make_atomic_swap_body(S)    int ## S sav; sav= *a; *a= v; v=sav;
+#define make_atomic_fas_body(S)     int ## S sav; sav= *a; *a= v; v=sav;
 #define make_atomic_cas_body(S)     if ((ret= (*a == *cmp))) *a= set; else *cmp=*a;
 #define make_atomic_load_body(S)    ret= *a;
 #define make_atomic_store_body(S)   *a= v;

include/atomic/solaris.h

@@ -186,25 +186,25 @@ my_atomic_storeptr(void * volatile *a, void *v)
 /* ------------------------------------------------------------------------ */
 
 STATIC_INLINE int8
-my_atomic_swap8(int8 volatile *a, int8 v)
+my_atomic_fas8(int8 volatile *a, int8 v)
 {
 	return ((int8) atomic_swap_8((volatile uint8_t *)a, (uint8_t)v));
 }
 
 STATIC_INLINE int16
-my_atomic_swap16(int16 volatile *a, int16 v)
+my_atomic_fas16(int16 volatile *a, int16 v)
 {
 	return ((int16) atomic_swap_16((volatile uint16_t *)a, (uint16_t)v));
 }
 
 STATIC_INLINE int32
-my_atomic_swap32(int32 volatile *a, int32 v)
+my_atomic_fas32(int32 volatile *a, int32 v)
 {
 	return ((int32) atomic_swap_32((volatile uint32_t *)a, (uint32_t)v));
 }
 
 STATIC_INLINE void *
-my_atomic_swapptr(void * volatile *a, void *v)
+my_atomic_fasptr(void * volatile *a, void *v)
 {
 	return (atomic_swap_ptr(a, v));
 }

include/atomic/x86-gcc.h

@@ -22,10 +22,18 @@
   architectures support double-word (128-bit) cas.
 */
 
-#ifdef MY_ATOMIC_NO_XADD
-#define MY_ATOMIC_MODE "gcc-x86" LOCK "-no-xadd"
+#ifdef __x86_64__
+#  ifdef MY_ATOMIC_NO_XADD
+#    define MY_ATOMIC_MODE "gcc-amd64" LOCK_prefix "-no-xadd"
+#  else
+#    define MY_ATOMIC_MODE "gcc-amd64" LOCK_prefix
+#  endif
 #else
-#define MY_ATOMIC_MODE "gcc-x86" LOCK
+#  ifdef MY_ATOMIC_NO_XADD
+#    define MY_ATOMIC_MODE "gcc-x86" LOCK_prefix "-no-xadd"
+#  else
+#    define MY_ATOMIC_MODE "gcc-x86" LOCK_prefix
+#  endif
 #endif
 
 /* fix -ansi errors while maintaining readability */
@@ -35,12 +43,12 @@
 
 #ifndef MY_ATOMIC_NO_XADD
 #define make_atomic_add_body(S)					\
-  asm volatile (LOCK "; xadd %0, %1;" : "+r" (v) , "+m" (*a))
+  asm volatile (LOCK_prefix "; xadd %0, %1;" : "+r" (v) , "+m" (*a))
 #endif
-#define make_atomic_swap_body(S)				\
-  asm volatile ("; xchg %0, %1;" : "+q" (v) , "+m" (*a))
+#define make_atomic_fas_body(S)				\
+  asm volatile ("xchg %0, %1;" : "+q" (v) , "+m" (*a))
 #define make_atomic_cas_body(S)					\
-  asm volatile (LOCK "; cmpxchg %3, %0; setz %2;"		\
+  asm volatile (LOCK_prefix "; cmpxchg %3, %0; setz %2;"	\
                : "+m" (*a), "+a" (*cmp), "=q" (ret): "r" (set))
 
 #ifdef MY_ATOMIC_MODE_DUMMY
@@ -49,14 +57,14 @@
 #else
 /*
   Actually 32-bit reads/writes are always atomic on x86
-  But we add LOCK here anyway to force memory barriers
+  But we add LOCK_prefix here anyway to force memory barriers
 */
 #define make_atomic_load_body(S)				\
   ret=0;							\
-  asm volatile (LOCK "; cmpxchg %2, %0"				\
+  asm volatile (LOCK_prefix "; cmpxchg %2, %0"			\
                : "+m" (*a), "+a" (ret): "r" (ret))
 #define make_atomic_store_body(S)				\
-  asm volatile ("; xchg %0, %1;" : "+m" (*a) : "r" (v))
+  asm volatile ("; xchg %0, %1;" : "+m" (*a), "+r" (v))
 #endif
 
 #endif /* ATOMIC_X86_GCC_INCLUDED */

include/my_atomic.h

@@ -16,9 +16,51 @@
    along with this program; if not, write to the Free Software
    Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA */
 
+/*
+  This header defines five atomic operations:
+
+  my_atomic_add#(&var, what)
+    add 'what' to *var, and return the old value of *var
+
+  my_atomic_fas#(&var, what)
+    'Fetch And Store'
+    store 'what' in *var, and return the old value of *var
+
+  my_atomic_cas#(&var, &old, new)
+    'Compare And Swap'
+    if *var is equal to *old, then store 'new' in *var, and return TRUE
+    otherwise store *var in *old, and return FALSE
+
+  my_atomic_load#(&var)
+    return *var
+
+  my_atomic_store#(&var, what)
+    store 'what' in *var
+
+  '#' is substituted by a size suffix - 8, 16, 32, or ptr
+  (e.g. my_atomic_add8, my_atomic_fas32, my_atomic_casptr).
+
+  NOTE This operations are not always atomic, so they always must be
+  enclosed in my_atomic_rwlock_rdlock(lock)/my_atomic_rwlock_rdunlock(lock)
+  or my_atomic_rwlock_wrlock(lock)/my_atomic_rwlock_wrunlock(lock).
+  Hint: if a code block makes intensive use of atomic ops, it make sense
+  to take/release rwlock once for the whole block, not for every statement.
+
+  On architectures where these operations are really atomic, rwlocks will
+  be optimized away.
+  8- and 16-bit atomics aren't implemented for windows (see generic-msvc.h),
+  but can be added, if necessary.
+*/
+
 #ifndef my_atomic_rwlock_init
 
 #define intptr         void *
+/**
+  On most platforms we implement 8-bit, 16-bit, 32-bit and "pointer"
+  operations. Thus the symbol below is defined by default; platforms
+  where we leave out 8-bit or 16-bit operations should undefine it.
+*/
+#define MY_ATOMIC_HAS_8_16 1
 
 #ifndef MY_ATOMIC_MODE_RWLOCKS
 /*
@@ -27,124 +69,223 @@
 #include "atomic/nolock.h"
 #endif
 
-#ifndef MY_ATOMIC_NOLOCK
-/*
- * Have to use rw-locks for atomic ops
- */
+#ifndef make_atomic_cas_body
+/* nolock.h was not able to generate even a CAS function, fall back */
 #include "atomic/rwlock.h"
-#endif
-
-#ifndef MY_ATOMICS_MADE
-
+#else
+/* define missing functions by using the already generated ones */
 #ifndef make_atomic_add_body
-#define make_atomic_add_body(S)					\
+#define make_atomic_add_body(S)                                 \
   int ## S tmp=*a;                                              \
   while (!my_atomic_cas ## S(a, &tmp, tmp+v));                  \
   v=tmp;
 #endif
+#ifndef make_atomic_fas_body
+#define make_atomic_fas_body(S)                                 \
+  int ## S tmp=*a;                                              \
+  while (!my_atomic_cas ## S(a, &tmp, v));                      \
+  v=tmp;
+#endif
+#ifndef make_atomic_load_body
+#define make_atomic_load_body(S)                                \
+  ret= 0; /* avoid compiler warning */                          \
+  (void)(my_atomic_cas ## S(a, &ret, ret));
+#endif
+#ifndef make_atomic_store_body
+#define make_atomic_store_body(S)                               \
+  (void)(my_atomic_fas ## S (a, v));
+#endif
+#endif
+
+/*
+  transparent_union doesn't work in g++
+  Bug ?
+
+  Darwin's gcc doesn't want to put pointers in a transparent_union
+  when built with -arch ppc64. Complains:
+  warning: 'transparent_union' attribute ignored
+*/
+#if defined(__GNUC__) && !defined(__cplusplus) && \
+      ! (defined(__APPLE__) && defined(_ARCH_PPC64))
+/*
+  we want to be able to use my_atomic_xxx functions with
+  both signed and unsigned integers. But gcc will issue a warning
+  "passing arg N of `my_atomic_XXX' as [un]signed due to prototype"
+  if the signedness of the argument doesn't match the prototype, or
+  "pointer targets in passing argument N of my_atomic_XXX differ in signedness"
+  if int* is used where uint* is expected (or vice versa).
+  Let's shut these warnings up
+*/
+#define make_transparent_unions(S)                              \
+        typedef union {                                         \
+          int  ## S  i;                                         \
+          uint ## S  u;                                         \
+        } U_ ## S   __attribute__ ((transparent_union));        \
+        typedef union {                                         \
+          int  ## S volatile *i;                                \
+          uint ## S volatile *u;                                \
+        } Uv_ ## S   __attribute__ ((transparent_union));
+#define uintptr intptr
+make_transparent_unions(8)
+make_transparent_unions(16)
+make_transparent_unions(32)
+make_transparent_unions(ptr)
+#undef uintptr
+#undef make_transparent_unions
+#define a       U_a.i
+#define cmp     U_cmp.i
+#define v       U_v.i
+#define set     U_set.i
+#else
+#define U_8    int8
+#define U_16   int16
+#define U_32   int32
+#define U_ptr  intptr
+#define Uv_8   int8
+#define Uv_16  int16
+#define Uv_32  int32
+#define Uv_ptr intptr
+#define U_a    volatile *a
+#define U_cmp  *cmp
+#define U_v    v
+#define U_set  set
+#endif /* __GCC__ transparent_union magic */
 
 #ifdef HAVE_INLINE
 
-#define make_atomic_add(S)					\
-STATIC_INLINE int ## S my_atomic_add ## S(			\
-                        int ## S volatile *a, int ## S v)	\
-{								\
-  make_atomic_add_body(S);					\
-  return v;							\
+#define make_atomic_cas(S)                                      \
+STATIC_INLINE int my_atomic_cas ## S(Uv_ ## S U_a,              \
+                            Uv_ ## S U_cmp, U_ ## S U_set)      \
+{                                                               \
+  int8 ret;                                                     \
+  make_atomic_cas_body(S);                                      \
+  return ret;                                                   \
 }
 
-#define make_atomic_swap(S)					\
-STATIC_INLINE int ## S my_atomic_swap ## S(			\
-                         int ## S volatile *a, int ## S v)	\
-{								\
-  make_atomic_swap_body(S);					\
-  return v;							\
+#define make_atomic_add(S)                                      \
+STATIC_INLINE int ## S my_atomic_add ## S(                      \
+                        Uv_ ## S U_a, U_ ## S U_v)              \
+{                                                               \
+  make_atomic_add_body(S);                                      \
+  return v;                                                     \
 }
 
-#define make_atomic_cas(S)					\
-STATIC_INLINE int my_atomic_cas ## S(int ## S volatile *a,	\
-                            int ## S *cmp, int ## S set)	\
-{								\
-  int8 ret;							\
-  make_atomic_cas_body(S);					\
-  return ret;							\
+#define make_atomic_fas(S)                                      \
+STATIC_INLINE int ## S my_atomic_fas ## S(                      \
+                         Uv_ ## S U_a, U_ ## S U_v)             \
+{                                                               \
+  make_atomic_fas_body(S);                                      \
+  return v;                                                     \
 }
 
-#define make_atomic_load(S)					\
-STATIC_INLINE int ## S my_atomic_load ## S(int ## S volatile *a) \
-{								\
-  int ## S ret;						\
-  make_atomic_load_body(S);					\
-  return ret;							\
+#define make_atomic_load(S)                                     \
+STATIC_INLINE int ## S my_atomic_load ## S(Uv_ ## S U_a)        \
+{                                                               \
+  int ## S ret;                                                 \
+  make_atomic_load_body(S);                                     \
+  return ret;                                                   \
 }
 
-#define make_atomic_store(S)					\
-STATIC_INLINE void my_atomic_store ## S(			\
-                     int ## S volatile *a, int ## S v)	\
-{								\
-  make_atomic_store_body(S);					\
+#define make_atomic_store(S)                                    \
+STATIC_INLINE void my_atomic_store ## S(                        \
+                     Uv_ ## S U_a, U_ ## S U_v)                 \
+{                                                               \
+  make_atomic_store_body(S);                                    \
 }
 
 #else /* no inline functions */
 
-#define make_atomic_add(S)					\
-extern int ## S my_atomic_add ## S(int ## S volatile *a, int ## S v);
+#define make_atomic_add(S)                                      \
+extern int ## S my_atomic_add ## S(Uv_ ## S U_a, U_ ## S U_v);
 
-#define make_atomic_swap(S)					\
-extern int ## S my_atomic_swap ## S(int ## S volatile *a, int ## S v);
+#define make_atomic_fas(S)                                      \
+extern int ## S my_atomic_fas ## S(Uv_ ## S U_a, U_ ## S U_v);
 
-#define make_atomic_cas(S)					\
-extern int my_atomic_cas ## S(int ## S volatile *a, int ## S *cmp, int ## S set);
+#define make_atomic_cas(S)                                      \
+extern int my_atomic_cas ## S(Uv_ ## S U_a, Uv_ ## S U_cmp, U_ ## S U_set);
 
-#define make_atomic_load(S)					\
-extern int ## S my_atomic_load ## S(int ## S volatile *a);
+#define make_atomic_load(S)                                     \
+extern int ## S my_atomic_load ## S(Uv_ ## S U_a);
 
-#define make_atomic_store(S)					\
-extern void my_atomic_store ## S(int ## S volatile *a, int ## S v);
+#define make_atomic_store(S)                                    \
+extern void my_atomic_store ## S(Uv_ ## S U_a, U_ ## S U_v);
 
 #endif /* HAVE_INLINE */
 
-make_atomic_cas( 8)
+#ifdef MY_ATOMIC_HAS_8_16
+make_atomic_cas(8)
 make_atomic_cas(16)
+#endif
 make_atomic_cas(32)
 make_atomic_cas(ptr)
 
-make_atomic_add( 8)
+#ifdef MY_ATOMIC_HAS_8_16
+make_atomic_add(8)
 make_atomic_add(16)
+#endif
 make_atomic_add(32)
 
-make_atomic_load( 8)
+#ifdef MY_ATOMIC_HAS_8_16
+make_atomic_load(8)
 make_atomic_load(16)
+#endif
 make_atomic_load(32)
 make_atomic_load(ptr)
 
-make_atomic_store( 8)
+#ifdef MY_ATOMIC_HAS_8_16
+make_atomic_fas(8)
+make_atomic_fas(16)
+#endif
+make_atomic_fas(32)
+make_atomic_fas(ptr)
+
+#ifdef MY_ATOMIC_HAS_8_16
+make_atomic_store(8)
 make_atomic_store(16)
+#endif
 make_atomic_store(32)
 make_atomic_store(ptr)
 
-make_atomic_swap( 8)
-make_atomic_swap(16)
-make_atomic_swap(32)
-make_atomic_swap(ptr)
+#ifdef _atomic_h_cleanup_
+#include _atomic_h_cleanup_
+#undef _atomic_h_cleanup_
+#endif
 
+#undef U_8
+#undef U_16
+#undef U_32
+#undef U_ptr
+#undef Uv_8
+#undef Uv_16
+#undef Uv_32
+#undef Uv_ptr
+#undef a
+#undef cmp
+#undef v
+#undef set
+#undef U_a
+#undef U_cmp
+#undef U_v
+#undef U_set
 #undef make_atomic_add
 #undef make_atomic_cas
 #undef make_atomic_load
 #undef make_atomic_store
-#undef make_atomic_swap
+#undef make_atomic_fas
 #undef make_atomic_add_body
 #undef make_atomic_cas_body
 #undef make_atomic_load_body
 #undef make_atomic_store_body
-#undef make_atomic_swap_body
+#undef make_atomic_fas_body
 #undef intptr
 
-#endif /* MY_ATOMICS_MADE */
-
-#ifdef _atomic_h_cleanup_
-#include _atomic_h_cleanup_
-#undef _atomic_h_cleanup_
+/*
+  the macro below defines (as an expression) the code that
+  will be run in spin-loops. Intel manuals recummend to have PAUSE there.
+  It is expected to be defined in include/atomic/ *.h files
+*/
+#ifndef LF_BACKOFF
+#define LF_BACKOFF (1)
 #endif
 
 #define MY_ATOMIC_OK       0

mysys/my_atomic.c

-/* Copyright (C) 2006 MySQL AB
+/* Copyright (C) 2006 MySQL AB, 2008-2009 Sun Microsystems, Inc
 
    This program is free software; you can redistribute it and/or modify
    it under the terms of the GNU General Public License as published by
@@ -14,7 +14,7 @@
    Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA */
 
 #include <my_global.h>
-#include <my_pthread.h>
+#include <my_sys.h>
 
 #ifndef HAVE_INLINE
 /* the following will cause all inline functions to be instantiated */
@@ -43,3 +43,32 @@ int my_atomic_initialize()
 #endif
 }
 
+#ifdef SAFE_MUTEX
+#undef pthread_mutex_init
+#undef pthread_mutex_destroy
+#undef pthread_mutex_lock
+#undef pthread_mutex_unlock
+
+void plain_pthread_mutex_init(safe_mutex_t *m)
+{
+  pthread_mutex_init(& m->mutex, NULL);
+}
+
+void plain_pthread_mutex_destroy(safe_mutex_t *m)
+{
+  pthread_mutex_destroy(& m->mutex);
+}
+
+void plain_pthread_mutex_lock(safe_mutex_t *m)
+{
+  pthread_mutex_lock(& m->mutex);
+}
+
+void plain_pthread_mutex_unlock(safe_mutex_t *m)
+{
+  pthread_mutex_unlock(& m->mutex);
+}
+
+#endif
+
+

mysys/my_getncpus.c

-/* Copyright (C) 2006 MySQL AB
+/* Copyright (C) 2006 MySQL AB, 2008-2009 Sun Microsystems, Inc
 
    This program is free software; you can redistribute it and/or modify
    it under the terms of the GNU General Public License as published by
@@ -16,24 +16,34 @@
 /* get the number of (online) CPUs */
 
 #include "mysys_priv.h"
+#ifdef HAVE_UNISTD_H
 #include <unistd.h>
+#endif
 
 static int ncpus=0;
 
-#ifdef _SC_NPROCESSORS_ONLN
 int my_getncpus()
 {
   if (!ncpus)
+  {
+#ifdef _SC_NPROCESSORS_ONLN
     ncpus= sysconf(_SC_NPROCESSORS_ONLN);
-  return ncpus;
-}
-
+#elif defined(__WIN__)
+    SYSTEM_INFO sysinfo;
+
+    /*
+    * We are not calling GetNativeSystemInfo here because (1) we
+    * don't believe that they return different values for number
+    * of processors and (2) if WOW64 limits processors for Win32
+    * then we don't want to try to override that.
+    */
+    GetSystemInfo(&sysinfo);
+
+    ncpus= sysinfo.dwNumberOfProcessors;
 #else
-/* unknown */
-int my_getncpus()
-{
-  return 2;
-}
-
+/* unknown so play safe: assume SMP and forbid uniprocessor build */
+    ncpus= 2;
 #endif
-
+  }
+  return ncpus;
+}

unittest/mysys/Makefile.am

@@ -16,6 +16,8 @@
 AM_CPPFLAGS      = @ZLIB_INCLUDES@ -I$(top_builddir)/include 
 AM_CPPFLAGS     += -I$(top_srcdir)/include -I$(top_srcdir)/unittest/mytap
 
+noinst_HEADERS = thr_template.c
+
 LDADD 		= $(top_builddir)/unittest/mytap/libmytap.a \
 		  $(top_builddir)/mysys/libmysys.a \
 		  $(top_builddir)/dbug/libdbug.a \

unittest/mysys/my_atomic-t.c

-/* Copyright (C) 2006 MySQL AB
+/* Copyright (C) 2006-2008 MySQL AB, 2008 Sun Microsystems, Inc.
 
    This program is free software; you can redistribute it and/or modify
    it under the terms of the GNU General Public License as published by
@@ -13,10 +13,7 @@
    along with this program; if not, write to the Free Software
    Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA */
 
-#include <my_global.h>
-#include <my_sys.h>
-#include <my_atomic.h>
-#include <tap.h>
+#include "thr_template.c"
 
 /* at least gcc 3.4.5 and 3.4.6 (but not 3.2.3) on RHEL */
 #if __GNUC__ == 3 && __GNUC_MINOR__ == 4
@@ -25,181 +22,125 @@
 #define GCC_BUG_WORKAROUND
 #endif
 
-int32 a32,b32,c32;
+volatile uint32 b32;
+volatile int32  c32;
 my_atomic_rwlock_t rwl;
 
-pthread_attr_t thr_attr;
-pthread_mutex_t mutex;
-pthread_cond_t cond;
-int N;
-
 /* add and sub a random number in a loop. Must get 0 at the end */
-pthread_handler_t test_atomic_add_handler(void *arg)
+pthread_handler_t test_atomic_add(void *arg)
 {
-  int    m=*(int *)arg;
+  int    m= (*(int *)arg)/2;
   GCC_BUG_WORKAROUND int32 x;
-  for (x=((int)((long)(&m))); m ; m--)
+  for (x= ((int)(intptr)(&m)); m ; m--)
   {
-    x=x*m+0x87654321;
+    x= (x*m+0x87654321) & INT_MAX32;
     my_atomic_rwlock_wrlock(&rwl);
-    my_atomic_add32(&a32, x);
+    my_atomic_add32(&bad, x);
     my_atomic_rwlock_wrunlock(&rwl);
 
     my_atomic_rwlock_wrlock(&rwl);
-    my_atomic_add32(&a32, -x);
+    my_atomic_add32(&bad, -x);
     my_atomic_rwlock_wrunlock(&rwl);
   }
   pthread_mutex_lock(&mutex);
-  N--;
-  if (!N) pthread_cond_signal(&cond);
+  if (!--running_threads) pthread_cond_signal(&cond);
   pthread_mutex_unlock(&mutex);
   return 0;
 }
 
 /*
   1. generate thread number 0..N-1 from b32
-  2. add it to a32
+  2. add it to bad
   3. swap thread numbers in c32
   4. (optionally) one more swap to avoid 0 as a result
-  5. subtract result from a32
-  must get 0 in a32 at the end
+  5. subtract result from bad
+  must get 0 in bad at the end
 */
-pthread_handler_t test_atomic_swap_handler(void *arg)
+pthread_handler_t test_atomic_fas(void *arg)
 {
-  int    m=*(int *)arg;
-  int32 x;
+  int    m= *(int *)arg;
+  int32  x;
 
   my_atomic_rwlock_wrlock(&rwl);
-  x=my_atomic_add32(&b32, 1);
+  x= my_atomic_add32(&b32, 1);
   my_atomic_rwlock_wrunlock(&rwl);
 
   my_atomic_rwlock_wrlock(&rwl);
-  my_atomic_add32(&a32, x);
+  my_atomic_add32(&bad, x);
   my_atomic_rwlock_wrunlock(&rwl);
 
   for (; m ; m--)
   {
     my_atomic_rwlock_wrlock(&rwl);
-    x=my_atomic_swap32(&c32, x);
+    x= my_atomic_fas32(&c32, x);
     my_atomic_rwlock_wrunlock(&rwl);
   }
 
   if (!x)
   {
     my_atomic_rwlock_wrlock(&rwl);
-    x=my_atomic_swap32(&c32, x);
+    x= my_atomic_fas32(&c32, x);
     my_atomic_rwlock_wrunlock(&rwl);
   }
 
   my_atomic_rwlock_wrlock(&rwl);
-  my_atomic_add32(&a32, -x);
+  my_atomic_add32(&bad, -x);
   my_atomic_rwlock_wrunlock(&rwl);
 
   pthread_mutex_lock(&mutex);
-  N--;
-  if (!N) pthread_cond_signal(&cond);
+  if (!--running_threads) pthread_cond_signal(&cond);
   pthread_mutex_unlock(&mutex);
   return 0;
 }
 
 /*
-  same as test_atomic_add_handler, but my_atomic_add32 is emulated with
-  (slower) my_atomic_cas32
+  same as test_atomic_add, but my_atomic_add32 is emulated with
+  my_atomic_cas32 - notice that the slowdown is proportional to the
+  number of CPUs
 */
-pthread_handler_t test_atomic_cas_handler(void *arg)
+pthread_handler_t test_atomic_cas(void *arg)
 {
-  int    m=*(int *)arg, ok;
-  GCC_BUG_WORKAROUND int32 x,y;
-  for (x=((int)((long)(&m))); m ; m--)
+  int    m= (*(int *)arg)/2, ok= 0;
+  GCC_BUG_WORKAROUND int32 x, y;
+  for (x= ((int)(intptr)(&m)); m ; m--)
   {
     my_atomic_rwlock_wrlock(&rwl);
-    y=my_atomic_load32(&a32);
+    y= my_atomic_load32(&bad);
     my_atomic_rwlock_wrunlock(&rwl);
-
-    x=x*m+0x87654321;
+    x= (x*m+0x87654321) & INT_MAX32;
     do {
       my_atomic_rwlock_wrlock(&rwl);
-      ok=my_atomic_cas32(&a32, &y, y+x);
+      ok= my_atomic_cas32(&bad, &y, (uint32)y+x);
       my_atomic_rwlock_wrunlock(&rwl);
-    } while (!ok);
+    } while (!ok) ;
     do {
       my_atomic_rwlock_wrlock(&rwl);
-      ok=my_atomic_cas32(&a32, &y, y-x);
+      ok= my_atomic_cas32(&bad, &y, y-x);
       my_atomic_rwlock_wrunlock(&rwl);
-    } while (!ok);
+    } while (!ok) ;
   }
   pthread_mutex_lock(&mutex);
-  N--;
-  if (!N) pthread_cond_signal(&cond);
+  if (!--running_threads) pthread_cond_signal(&cond);
   pthread_mutex_unlock(&mutex);
   return 0;
 }
 
-void test_atomic(const char *test, pthread_handler handler, int n, int m)
-{
-  pthread_t t;
-  ulonglong now=my_getsystime();
-
-  a32= 0;
-  b32= 0;
-  c32= 0;
-
-  diag("Testing %s with %d threads, %d iterations... ", test, n, m);
-  for (N=n ; n ; n--)
-  {
-    if (pthread_create(&t, &thr_attr, handler, &m) != 0)
-    {
-      diag("Could not create thread");
-      a32= 1;
-      goto err;
-    }
-  }
-
-  pthread_mutex_lock(&mutex);
-  while (N)
-    pthread_cond_wait(&cond, &mutex);
-  pthread_mutex_unlock(&mutex);
-  now=my_getsystime()-now;
-err:
-  ok(a32 == 0, "tested %s in %g secs", test, ((double)now)/1e7);
-}
 
-int main()
+void do_tests()
 {
-  int err;
-  MY_INIT("my_atomic-t.c");
-
-  diag("N CPUs: %d", my_getncpus());
-  err= my_atomic_initialize();
-
   plan(4);
-  ok(err == 0, "my_atomic_initialize() returned %d", err);
 
-  pthread_attr_init(&thr_attr);
-  pthread_attr_setdetachstate(&thr_attr,PTHREAD_CREATE_DETACHED);
-  pthread_mutex_init(&mutex, 0);
-  pthread_cond_init(&cond, 0);
+  bad= my_atomic_initialize();
+  ok(!bad, "my_atomic_initialize() returned %d", bad);
+
   my_atomic_rwlock_init(&rwl);
 
-#ifdef HPUX11
-#define CYCLES 1000
-#else
-#define CYCLES 10000
-#endif
-#define THREADS 100
-  test_atomic("my_atomic_add32", test_atomic_add_handler, THREADS, CYCLES);
-  test_atomic("my_atomic_swap32", test_atomic_swap_handler, THREADS, CYCLES);
-  test_atomic("my_atomic_cas32", test_atomic_cas_handler, THREADS, CYCLES);
-  /*
-    workaround until we know why it crashes randomly on some machine
-    (BUG#22320).
-  */
-  sleep(2);
-
-  pthread_mutex_destroy(&mutex);
-  pthread_cond_destroy(&cond);
-  pthread_attr_destroy(&thr_attr);
+  b32= c32= 0;
+  test_concurrently("my_atomic_add32", test_atomic_add, THREADS, CYCLES);
+  b32= c32= 0;
+  test_concurrently("my_atomic_fas32", test_atomic_fas, THREADS, CYCLES);
+  b32= c32= 0;
+  test_concurrently("my_atomic_cas32", test_atomic_cas, THREADS, CYCLES);
+
   my_atomic_rwlock_destroy(&rwl);
-  return exit_status();
 }
-

unittest/mysys/thr_template.c

+/* Copyright (C) 2006-2008 MySQL AB, 2008 Sun Microsystems, Inc.
+
+   This program is free software; you can redistribute it and/or modify
+   it under the terms of the GNU General Public License as published by
+   the Free Software Foundation; version 2 of the License.
+
+   This program is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+   GNU General Public License for more details.
+
+   You should have received a copy of the GNU General Public License
+   along with this program; if not, write to the Free Software
+   Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA */
+
+#include <my_global.h>
+#include <my_sys.h>
+#include <my_atomic.h>
+#include <tap.h>
+
+volatile uint32 bad;
+pthread_attr_t thr_attr;
+pthread_mutex_t mutex;
+pthread_cond_t cond;
+uint running_threads;
+
+void do_tests();
+
+void test_concurrently(const char *test, pthread_handler handler, int n, int m)
+{
+  pthread_t t;
+  ulonglong now= my_getsystime();
+
+  bad= 0;
+
+  diag("Testing %s with %d threads, %d iterations... ", test, n, m);
+  for (running_threads= n ; n ; n--)
+  {
+    if (pthread_create(&t, &thr_attr, handler, &m) != 0)
+    {
+      diag("Could not create thread");
+      abort();
+    }
+  }
+  pthread_mutex_lock(&mutex);
+  while (running_threads)
+    pthread_cond_wait(&cond, &mutex);
+  pthread_mutex_unlock(&mutex);
+
+  now= my_getsystime()-now;
+  ok(!bad, "tested %s in %g secs (%d)", test, ((double)now)/1e7, bad);
+}
+
+int main(int argc __attribute__((unused)), char **argv)
+{
+  MY_INIT("thd_template");
+
+  if (argv[1] && *argv[1])
+    DBUG_SET_INITIAL(argv[1]);
+
+  pthread_mutex_init(&mutex, 0);
+  pthread_cond_init(&cond, 0);
+  pthread_attr_init(&thr_attr);
+  pthread_attr_setdetachstate(&thr_attr,PTHREAD_CREATE_DETACHED);
+
+#ifdef MY_ATOMIC_MODE_RWLOCKS
+#if defined(HPUX11) || defined(__POWERPC__) /* showed to be very slow (scheduler-related) */
+#define CYCLES 300
+#else
+#define CYCLES 3000
+#endif
+#else
+#define CYCLES 3000
+#endif
+#define THREADS 30
+
+  diag("N CPUs: %d, atomic ops: %s", my_getncpus(), MY_ATOMIC_MODE);
+
+  do_tests();
+
+  /*
+    workaround until we know why it crashes randomly on some machine
+    (BUG#22320).
+  */
+  sleep(2);
+  pthread_mutex_destroy(&mutex);
+  pthread_cond_destroy(&cond);
+  pthread_attr_destroy(&thr_attr);
+  my_end(0);
+  return exit_status();
+}
+