mariadb/innobase/include/sync0sync.ic
unknown a3dc40e24a Applied InnoDB snapshot innodb-5.0-ss2095
Fixes the following bugs:

- Bug #29560: InnoDB >= 5.0.30 hangs on adaptive hash rw-lock 'waiting for an X-lock'

  Fixed a race condition in the rw_lock where an os_event_reset()
  can overwrite an earlier os_event_set() triggering an indefinite
  wait.
  NOTE: This fix for windows is different from that for other platforms.
  NOTE2: This bug is introduced in the scalability fix to the
  sync0arr which was applied to 5.0 only. Therefore, it need not be
  applied to the 5.1 tree. If we decide to port the scalability fix
  to 5.1 then this fix should be ported as well.

- Bug #32125: Database crash due to ha_innodb.cc:3896: ulint convert_search_mode_to_innobase

  When unknown find_flag is encountered in convert_search_mode_to_innobase()
  do not call assert(0); instead queue a MySQL error using my_error() and
  return the error code PAGE_CUR_UNSUPP. Change the functions that call
  convert_search_mode_to_innobase() to handle that error code by "canceling"
  execution and returning appropriate error code further upstream.


innobase/include/db0err.h:
  Applied InnoDB snapshot innodb-5.0-ss2095
  
  Revision r2091:
  branches/5.0:
   
  Merge r2088 from trunk:
   
  log for r2088:
  
  Fix Bug#32125 (http://bugs.mysql.com/32125)
  "Database crash due to ha_innodb.cc:3896: ulint convert_search_mode_to_innobase":
  
  When unknown find_flag is encountered in convert_search_mode_to_innobase()
  do not call assert(0); instead queue a MySQL error using my_error() and
  return the error code PAGE_CUR_UNSUPP. Change the functions that call
  convert_search_mode_to_innobase() to handle that error code by "canceling"
  execution and returning appropriate error code further upstream.
  
  Approved by:	Heikki
innobase/include/os0sync.h:
  Applied InnoDB snapshot innodb-5.0-ss2095
  
  Revision r2082:
  branches/5.0:  bug#29560
  
  Fixed a race condition in the rw_lock where an os_event_reset()
  can overwrite an earlier os_event_set() triggering an indefinite
  wait.
  NOTE: This fix for windows is different from that for other platforms.
  NOTE2: This bug is introduced in the scalability fix to the
  sync0arr which was applied to 5.0 only. Therefore, it need not be
  applied to the 5.1 tree. If we decide to port the scalability fix
  to 5.1 then this fix should be ported as well.
  
  Reviewed by: Heikki
innobase/include/page0cur.h:
  Applied InnoDB snapshot innodb-5.0-ss2095
  
  Revision r2091:
  branches/5.0:
   
  Merge r2088 from trunk:
   
  log for r2088:
  
  Fix Bug#32125 (http://bugs.mysql.com/32125)
  "Database crash due to ha_innodb.cc:3896: ulint convert_search_mode_to_innobase":
  
  When unknown find_flag is encountered in convert_search_mode_to_innobase()
  do not call assert(0); instead queue a MySQL error using my_error() and
  return the error code PAGE_CUR_UNSUPP. Change the functions that call
  convert_search_mode_to_innobase() to handle that error code by "canceling"
  execution and returning appropriate error code further upstream.
  
  Approved by:	Heikki
innobase/include/sync0rw.h:
  Applied InnoDB snapshot innodb-5.0-ss2095
  
  Revision r2082:
  branches/5.0:  bug#29560
  
  Fixed a race condition in the rw_lock where an os_event_reset()
  can overwrite an earlier os_event_set() triggering an indefinite
  wait.
  NOTE: This fix for windows is different from that for other platforms.
  NOTE2: This bug is introduced in the scalability fix to the
  sync0arr which was applied to 5.0 only. Therefore, it need not be
  applied to the 5.1 tree. If we decide to port the scalability fix
  to 5.1 then this fix should be ported as well.
  
  Reviewed by: Heikki
innobase/include/sync0rw.ic:
  Applied InnoDB snapshot innodb-5.0-ss2095
  
  Revision r2082:
  branches/5.0:  bug#29560
  
  Fixed a race condition in the rw_lock where an os_event_reset()
  can overwrite an earlier os_event_set() triggering an indefinite
  wait.
  NOTE: This fix for windows is different from that for other platforms.
  NOTE2: This bug is introduced in the scalability fix to the
  sync0arr which was applied to 5.0 only. Therefore, it need not be
  applied to the 5.1 tree. If we decide to port the scalability fix
  to 5.1 then this fix should be ported as well.
  
  Reviewed by: Heikki
innobase/include/sync0sync.ic:
  Applied InnoDB snapshot innodb-5.0-ss2095
  
  Revision r2082:
  branches/5.0:  bug#29560
  
  Fixed a race condition in the rw_lock where an os_event_reset()
  can overwrite an earlier os_event_set() triggering an indefinite
  wait.
  NOTE: This fix for windows is different from that for other platforms.
  NOTE2: This bug is introduced in the scalability fix to the
  sync0arr which was applied to 5.0 only. Therefore, it need not be
  applied to the 5.1 tree. If we decide to port the scalability fix
  to 5.1 then this fix should be ported as well.
  
  Reviewed by: Heikki
innobase/os/os0sync.c:
  Applied InnoDB snapshot innodb-5.0-ss2095
  
  Revision r2082:
  branches/5.0:  bug#29560
  
  Fixed a race condition in the rw_lock where an os_event_reset()
  can overwrite an earlier os_event_set() triggering an indefinite
  wait.
  NOTE: This fix for windows is different from that for other platforms.
  NOTE2: This bug is introduced in the scalability fix to the
  sync0arr which was applied to 5.0 only. Therefore, it need not be
  applied to the 5.1 tree. If we decide to port the scalability fix
  to 5.1 then this fix should be ported as well.
  
  Reviewed by: Heikki
innobase/srv/srv0srv.c:
  Applied InnoDB snapshot innodb-5.0-ss2095
  
  Revision r2082:
  branches/5.0:  bug#29560
  
  Fixed a race condition in the rw_lock where an os_event_reset()
  can overwrite an earlier os_event_set() triggering an indefinite
  wait.
  NOTE: This fix for windows is different from that for other platforms.
  NOTE2: This bug is introduced in the scalability fix to the
  sync0arr which was applied to 5.0 only. Therefore, it need not be
  applied to the 5.1 tree. If we decide to port the scalability fix
  to 5.1 then this fix should be ported as well.
  
  Reviewed by: Heikki
innobase/sync/sync0arr.c:
  Applied InnoDB snapshot innodb-5.0-ss2095
  
  Revision r2082:
  branches/5.0:  bug#29560
  
  Fixed a race condition in the rw_lock where an os_event_reset()
  can overwrite an earlier os_event_set() triggering an indefinite
  wait.
  NOTE: This fix for windows is different from that for other platforms.
  NOTE2: This bug is introduced in the scalability fix to the
  sync0arr which was applied to 5.0 only. Therefore, it need not be
  applied to the 5.1 tree. If we decide to port the scalability fix
  to 5.1 then this fix should be ported as well.
  
  Reviewed by: Heikki
innobase/sync/sync0rw.c:
  Applied InnoDB snapshot innodb-5.0-ss2095
  
  Revision r2082:
  branches/5.0:  bug#29560
  
  Fixed a race condition in the rw_lock where an os_event_reset()
  can overwrite an earlier os_event_set() triggering an indefinite
  wait.
  NOTE: This fix for windows is different from that for other platforms.
  NOTE2: This bug is introduced in the scalability fix to the
  sync0arr which was applied to 5.0 only. Therefore, it need not be
  applied to the 5.1 tree. If we decide to port the scalability fix
  to 5.1 then this fix should be ported as well.
  
  Reviewed by: Heikki
innobase/sync/sync0sync.c:
  Applied InnoDB snapshot innodb-5.0-ss2095
  
  Revision r2082:
  branches/5.0:  bug#29560
  
  Fixed a race condition in the rw_lock where an os_event_reset()
  can overwrite an earlier os_event_set() triggering an indefinite
  wait.
  NOTE: This fix for windows is different from that for other platforms.
  NOTE2: This bug is introduced in the scalability fix to the
  sync0arr which was applied to 5.0 only. Therefore, it need not be
  applied to the 5.1 tree. If we decide to port the scalability fix
  to 5.1 then this fix should be ported as well.
  
  Reviewed by: Heikki
sql/ha_innodb.cc:
  Applied InnoDB snapshot innodb-5.0-ss2095
  
  Revision r2091:
  branches/5.0:
   
  Merge r2088 from trunk:
   
  log for r2088:
  
  Fix Bug#32125 (http://bugs.mysql.com/32125)
  "Database crash due to ha_innodb.cc:3896: ulint convert_search_mode_to_innobase":
  
  When unknown find_flag is encountered in convert_search_mode_to_innobase()
  do not call assert(0); instead queue a MySQL error using my_error() and
  return the error code PAGE_CUR_UNSUPP. Change the functions that call
  convert_search_mode_to_innobase() to handle that error code by "canceling"
  execution and returning appropriate error code further upstream.
  
  Approved by:	Heikki
  
  
  Revision r2095:
  branches/5.0: Merge r2093 from trunk:
  
  convert_search_mode_to_innobase(): Add the missing case label
  HA_READ_MBR_EQUAL that was forgotten in r2088.
2007-11-20 10:53:19 -07:00

256 lines
7.1 KiB
Text

/******************************************************
Mutex, the basic synchronization primitive
(c) 1995 Innobase Oy
Created 9/5/1995 Heikki Tuuri
*******************************************************/
#if defined(not_defined) && defined(__GNUC__) && defined(UNIV_INTEL_X86)
/* %z0: Use the size of operand %0 which in our case is *m to determine
instruction size, it should end up as xchgl. "1" in the input constraint,
says that "in" has to go in the same place as "out".*/
#define TAS(m, in, out) \
asm volatile ("xchg%z0 %2, %0" \
: "=g" (*(m)), "=r" (out) \
: "1" (in)) /* Note: "1" here refers to "=r" (out) */
#endif
/**********************************************************************
Sets the waiters field in a mutex. */
void
mutex_set_waiters(
/*==============*/
mutex_t* mutex, /* in: mutex */
ulint n); /* in: value to set */
/**********************************************************************
Reserves a mutex for the current thread. If the mutex is reserved, the
function spins a preset time (controlled by SYNC_SPIN_ROUNDS) waiting
for the mutex before suspending the thread. */
void
mutex_spin_wait(
/*============*/
mutex_t* mutex, /* in: pointer to mutex */
const char* file_name,/* in: file name where mutex requested */
ulint line); /* in: line where requested */
#ifdef UNIV_SYNC_DEBUG
/**********************************************************************
Sets the debug information for a reserved mutex. */
void
mutex_set_debug_info(
/*=================*/
mutex_t* mutex, /* in: mutex */
const char* file_name, /* in: file where requested */
ulint line); /* in: line where requested */
#endif /* UNIV_SYNC_DEBUG */
/**********************************************************************
Releases the threads waiting in the primary wait array for this mutex. */
void
mutex_signal_object(
/*================*/
mutex_t* mutex); /* in: mutex */
/**********************************************************************
Performs an atomic test-and-set instruction to the lock_word field of a
mutex. */
UNIV_INLINE
ulint
mutex_test_and_set(
/*===============*/
/* out: the previous value of lock_word: 0 or
1 */
mutex_t* mutex) /* in: mutex */
{
#if defined(_WIN32) && defined(UNIV_CAN_USE_X86_ASSEMBLER)
ulint res;
ulint* lw; /* assembler code is used to ensure that
lock_word is loaded from memory */
ut_ad(mutex);
ut_ad(sizeof(ulint) == 4);
lw = &(mutex->lock_word);
__asm MOV ECX, lw
__asm MOV EDX, 1
__asm XCHG EDX, DWORD PTR [ECX]
__asm MOV res, EDX
/* The fence below would prevent this thread from reading the data
structure protected by the mutex before the test-and-set operation is
committed, but the fence is apparently not needed:
In a posting to comp.arch newsgroup (August 10, 1997) Andy Glew said
that in P6 a LOCKed instruction like XCHG establishes a fence with
respect to memory reads and writes and thus an explicit fence is not
needed. In P5 he seemed to agree with a previous newsgroup poster that
LOCKed instructions serialize all instruction execution, and,
consequently, also memory operations. This is confirmed in Intel
Software Dev. Manual, Vol. 3. */
/* mutex_fence(); */
return(res);
#elif defined(not_defined) && defined(__GNUC__) && defined(UNIV_INTEL_X86)
ulint res;
TAS(&mutex->lock_word, 1, res);
return(res);
#else
ibool ret;
ret = os_fast_mutex_trylock(&(mutex->os_fast_mutex));
if (ret == 0) {
/* We check that os_fast_mutex_trylock does not leak
and allow race conditions */
ut_a(mutex->lock_word == 0);
mutex->lock_word = 1;
}
return(ret);
#endif
}
/**********************************************************************
Performs a reset instruction to the lock_word field of a mutex. This
instruction also serializes memory operations to the program order. */
UNIV_INLINE
void
mutex_reset_lock_word(
/*==================*/
mutex_t* mutex) /* in: mutex */
{
#if defined(_WIN32) && defined(UNIV_CAN_USE_X86_ASSEMBLER)
ulint* lw; /* assembler code is used to ensure that
lock_word is loaded from memory */
ut_ad(mutex);
lw = &(mutex->lock_word);
__asm MOV EDX, 0
__asm MOV ECX, lw
__asm XCHG EDX, DWORD PTR [ECX]
#elif defined(not_defined) && defined(__GNUC__) && defined(UNIV_INTEL_X86)
ulint res;
TAS(&mutex->lock_word, 0, res);
#else
mutex->lock_word = 0;
os_fast_mutex_unlock(&(mutex->os_fast_mutex));
#endif
}
/**********************************************************************
Gets the value of the lock word. */
UNIV_INLINE
ulint
mutex_get_lock_word(
/*================*/
mutex_t* mutex) /* in: mutex */
{
volatile ulint* ptr; /* declared volatile to ensure that
lock_word is loaded from memory */
ut_ad(mutex);
ptr = &(mutex->lock_word);
return(*ptr);
}
/**********************************************************************
Gets the waiters field in a mutex. */
UNIV_INLINE
ulint
mutex_get_waiters(
/*==============*/
/* out: value to set */
mutex_t* mutex) /* in: mutex */
{
volatile ulint* ptr; /* declared volatile to ensure that
the value is read from memory */
ut_ad(mutex);
ptr = &(mutex->waiters);
return(*ptr); /* Here we assume that the read of a single
word from memory is atomic */
}
/**********************************************************************
Unlocks a mutex owned by the current thread. */
UNIV_INLINE
void
mutex_exit(
/*=======*/
mutex_t* mutex) /* in: pointer to mutex */
{
#ifdef UNIV_SYNC_DEBUG
ut_ad(mutex_own(mutex));
mutex->thread_id = ULINT_UNDEFINED;
sync_thread_reset_level(mutex);
#endif
mutex_reset_lock_word(mutex);
/* A problem: we assume that mutex_reset_lock word
is a memory barrier, that is when we read the waiters
field next, the read must be serialized in memory
after the reset. A speculative processor might
perform the read first, which could leave a waiting
thread hanging indefinitely.
Our current solution call every second
sync_arr_wake_threads_if_sema_free()
to wake up possible hanging threads if
they are missed in mutex_signal_object. */
if (mutex_get_waiters(mutex) != 0) {
mutex_signal_object(mutex);
}
#ifdef UNIV_SYNC_PERF_STAT
mutex_exit_count++;
#endif
}
/**********************************************************************
Locks a mutex for the current thread. If the mutex is reserved, the function
spins a preset time (controlled by SYNC_SPIN_ROUNDS), waiting for the mutex
before suspending the thread. */
UNIV_INLINE
void
mutex_enter_func(
/*=============*/
mutex_t* mutex, /* in: pointer to mutex */
const char* file_name, /* in: file name where locked */
ulint line) /* in: line where locked */
{
ut_ad(mutex_validate(mutex));
/* Note that we do not peek at the value of lock_word before trying
the atomic test_and_set; we could peek, and possibly save time. */
#if defined UNIV_DEBUG && !defined UNIV_HOTBACKUP
mutex->count_using++;
#endif /* UNIV_DEBUG && !UNIV_HOTBACKUP */
if (!mutex_test_and_set(mutex))
{
#ifdef UNIV_SYNC_DEBUG
mutex_set_debug_info(mutex, file_name, line);
#endif
return; /* Succeeded! */
}
mutex_spin_wait(mutex, file_name, line);
}