mirror/mariadb
1
0
Fork 0
mirror of https://github.com/MariaDB/server.git synced 2025-01-26 00:34:18 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions

branches/5.1: Port of r2267

Browse source 
This is a combination of changes that forward port the scalability fix applied to 5.0
through r1001.
It reverts changes r149 and r122 (these were 5.1 specific changes made in lieu of
scalability fix of 5.0)
Then it applies r1001 to 5.0 which is the original scalability fix.
Finally it applies r2082 which fixes an issue with the original fix.

Reviewed by: Heikki
This commit is contained in:
inaam 2008-02-03 19:52:44 +00:00
parent fc40679f02
commit baf6b527ab
11 changed files with 393 additions and 289 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

38
include/os0sync.h
View file

@ -112,9 +112,13 @@ os_event_set(
os_event_t event); /* in: event to set */
/**************************************************************
Resets an event semaphore to the nonsignaled state. Waiting threads will
stop to wait for the event. */
stop to wait for the event.
The return value should be passed to os_even_wait_low() if it is desired
that this thread should not wait in case of an intervening call to
os_event_set() between this os_event_reset() and the
os_event_wait_low() call. See comments for os_event_wait_low(). */
void
ib_longlong
os_event_reset(
/*===========*/
os_event_t event); /* in: event to reset */
@ -125,16 +129,38 @@ void
os_event_free(
/*==========*/
os_event_t event); /* in: event to free */
/**************************************************************
Waits for an event object until it is in the signaled state. If
srv_shutdown_state == SRV_SHUTDOWN_EXIT_THREADS this also exits the
waiting thread when the event becomes signaled (or immediately if the
event is already in the signaled state). */
event is already in the signaled state).
Typically, if the event has been signalled after the os_event_reset()
we'll return immediately because event->is_set == TRUE.
There are, however, situations (e.g.: sync_array code) where we may
lose this information. For example:
thread A calls os_event_reset()
thread B calls os_event_set() [event->is_set == TRUE]
thread C calls os_event_reset() [event->is_set == FALSE]
thread A calls os_event_wait() [infinite wait!]
thread C calls os_event_wait() [infinite wait!]
Where such a scenario is possible, to avoid infinite wait, the
value returned by os_event_reset() should be passed in as
reset_sig_count. */
#define os_event_wait(event) os_event_wait_low((event), 0)
void
os_event_wait(
/*==========*/
os_event_t event); /* in: event to wait */
os_event_wait_low(
/*==============*/
os_event_t event, /* in: event to wait */
ib_longlong reset_sig_count);/* in: zero or the value
returned by previous call of
os_event_reset(). */
/**************************************************************
Waits for an event object until it is in the signaled state or
a timeout is exceeded. In Unix the timeout is always infinite. */

21
include/sync0arr.h
View file

@ -66,26 +66,21 @@ sync_array_wait_event(
sync_array_t* arr, /* in: wait array */
ulint index); /* in: index of the reserved cell */
/**********************************************************************
Frees the cell safely by reserving the sync array mutex and decrementing
n_reserved if necessary. Should only be called from mutex_spin_wait. */
Frees the cell. NOTE! sync_array_wait_event frees the cell
automatically! */
void
sync_array_free_cell_protected(
/*===========================*/
sync_array_free_cell(
/*=================*/
sync_array_t* arr, /* in: wait array */
ulint index); /* in: index of the cell in array */
/**************************************************************************
Looks for the cells in the wait array which refer
to the wait object specified,
and sets their corresponding events to the signaled state. In this
way releases the threads waiting for the object to contend for the object.
It is possible that no such cell is found, in which case does nothing. */
Note that one of the wait objects was signalled. */
void
sync_array_signal_object(
/*=====================*/
sync_array_t* arr, /* in: wait array */
void* object);/* in: wait object */
sync_array_object_signalled(
/*========================*/
sync_array_t* arr); /* in: wait array */
/**************************************************************************
If the wakeup algorithm does not work perfectly at semaphore relases,
this function will do the waking (see the comment in mutex_exit). This

12
include/sync0rw.h
View file

@ -421,6 +421,18 @@ blocked by readers, a writer may queue for the lock by setting the writer
field. Then no new readers are allowed in. */
struct rw_lock_struct {
os_event_t event; /* Used by sync0arr.c for thread queueing */
#ifdef __WIN__
os_event_t wait_ex_event; /* This windows specific event is
used by the thread which has set the
lock state to RW_LOCK_WAIT_EX. The
rw_lock design guarantees that this
thread will be the next one to proceed
once the current the event gets
signalled. See LEMMA 2 in sync0sync.c */
#endif
ulint reader_count; /* Number of readers who have locked this
lock in the shared mode */
ulint writer; /* This field is set to RW_LOCK_EX if there

12
include/sync0rw.ic
View file

@ -381,7 +381,11 @@ rw_lock_s_unlock_func(
mutex_exit(mutex);
if (UNIV_UNLIKELY(sg)) {
sync_array_signal_object(sync_primary_wait_array, lock);
#ifdef __WIN__
os_event_set(lock->wait_ex_event);
#endif
os_event_set(lock->event);
sync_array_object_signalled(sync_primary_wait_array);
}
ut_ad(rw_lock_validate(lock));
@ -461,7 +465,11 @@ rw_lock_x_unlock_func(
mutex_exit(&(lock->mutex));
if (UNIV_UNLIKELY(sg)) {
sync_array_signal_object(sync_primary_wait_array, lock);
#ifdef __WIN__
os_event_set(lock->wait_ex_event);
#endif
os_event_set(lock->event);
sync_array_object_signalled(sync_primary_wait_array);
}
ut_ad(rw_lock_validate(lock));

1
include/sync0sync.h
View file

@ -470,6 +470,7 @@ Do not use its fields directly! The structure used in the spin lock
implementation of a mutual exclusion semaphore. */
struct mutex_struct {
os_event_t event; /* Used by sync0arr.c for the wait queue */
ulint lock_word; /* This ulint is the target of the atomic
test-and-set instruction in Win32 */
#if !defined(_WIN32) || !defined(UNIV_CAN_USE_X86_ASSEMBLER)

2
include/sync0sync.ic
View file

@ -211,7 +211,7 @@ mutex_exit(
perform the read first, which could leave a waiting
thread hanging indefinitely.
Our current solution call every 10 seconds
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. */

111
os/os0sync.c
View file

@ -21,6 +21,7 @@ Created 9/6/1995 Heikki Tuuri
/* Type definition for an operating system mutex struct */
struct os_mutex_struct{
os_event_t event; /* Used by sync0arr.c for queing threads */
void* handle; /* OS handle to mutex */
ulint count; /* we use this counter to check
that the same thread does not
@ -35,6 +36,7 @@ struct os_mutex_struct{
/* Mutex protecting counts and the lists of OS mutexes and events */
os_mutex_t os_sync_mutex;
ibool os_sync_mutex_inited = FALSE;
ibool os_sync_free_called = FALSE;
/* This is incremented by 1 in os_thread_create and decremented by 1 in
os_thread_exit */
@ -50,6 +52,10 @@ ulint os_event_count = 0;
ulint os_mutex_count = 0;
ulint os_fast_mutex_count = 0;
/* Because a mutex is embedded inside an event and there is an
event embedded inside a mutex, on free, this generates a recursive call.
This version of the free event function doesn't acquire the global lock */
static void os_event_free_internal(os_event_t event);
/*************************************************************
Initializes global event and OS 'slow' mutex lists. */
@ -76,6 +82,7 @@ os_sync_free(void)
os_event_t event;
os_mutex_t mutex;
os_sync_free_called = TRUE;
event = UT_LIST_GET_FIRST(os_event_list);
while (event) {
@ -99,6 +106,7 @@ os_sync_free(void)
mutex = UT_LIST_GET_FIRST(os_mutex_list);
}
os_sync_free_called = FALSE;
}
/*************************************************************
@ -144,17 +152,31 @@ os_event_create(
ut_a(0 == pthread_cond_init(&(event->cond_var), NULL));
#endif
event->is_set = FALSE;
event->signal_count = 0;
/* We return this value in os_event_reset(), which can then be
be used to pass to the os_event_wait_low(). The value of zero
is reserved in os_event_wait_low() for the case when the
caller does not want to pass any signal_count value. To
distinguish between the two cases we initialize signal_count
to 1 here. */
event->signal_count = 1;
#endif /* __WIN__ */
/* Put to the list of events */
os_mutex_enter(os_sync_mutex);
/* The os_sync_mutex can be NULL because during startup an event
can be created [ because it's embedded in the mutex/rwlock ] before
this module has been initialized */
if (os_sync_mutex != NULL) {
os_mutex_enter(os_sync_mutex);
}
/* Put to the list of events */
UT_LIST_ADD_FIRST(os_event_list, os_event_list, event);
os_event_count++;
os_mutex_exit(os_sync_mutex);
if (os_sync_mutex != NULL) {
os_mutex_exit(os_sync_mutex);
}
return(event);
}
@ -231,13 +253,20 @@ os_event_set(
/**************************************************************
Resets an event semaphore to the nonsignaled state. Waiting threads will
stop to wait for the event. */
stop to wait for the event.
The return value should be passed to os_even_wait_low() if it is desired
that this thread should not wait in case of an intervening call to
os_event_set() between this os_event_reset() and the
os_event_wait_low() call. See comments for os_event_wait_low(). */
void
ib_longlong
os_event_reset(
/*===========*/
/* out: current signal_count. */
os_event_t event) /* in: event to reset */
{
ib_longlong ret = 0;
#ifdef __WIN__
ut_a(event);
@ -252,9 +281,40 @@ os_event_reset(
} else {
event->is_set = FALSE;
}
ret = event->signal_count;
os_fast_mutex_unlock(&(event->os_mutex));
#endif
return(ret);
}
/**************************************************************
Frees an event object, without acquiring the global lock. */
static
void
os_event_free_internal(
/*===================*/
os_event_t event) /* in: event to free */
{
#ifdef __WIN__
ut_a(event);
ut_a(CloseHandle(event->handle));
#else
ut_a(event);
/* This is to avoid freeing the mutex twice */
os_fast_mutex_free(&(event->os_mutex));
ut_a(0 == pthread_cond_destroy(&(event->cond_var)));
#endif
/* Remove from the list of events */
UT_LIST_REMOVE(os_event_list, os_event_list, event);
os_event_count--;
ut_free(event);
}
/**************************************************************
@ -293,18 +353,38 @@ os_event_free(
Waits for an event object until it is in the signaled state. If
srv_shutdown_state == SRV_SHUTDOWN_EXIT_THREADS this also exits the
waiting thread when the event becomes signaled (or immediately if the
event is already in the signaled state). */
event is already in the signaled state).
Typically, if the event has been signalled after the os_event_reset()
we'll return immediately because event->is_set == TRUE.
There are, however, situations (e.g.: sync_array code) where we may
lose this information. For example:
thread A calls os_event_reset()
thread B calls os_event_set() [event->is_set == TRUE]
thread C calls os_event_reset() [event->is_set == FALSE]
thread A calls os_event_wait() [infinite wait!]
thread C calls os_event_wait() [infinite wait!]
Where such a scenario is possible, to avoid infinite wait, the
value returned by os_event_reset() should be passed in as
reset_sig_count. */
void
os_event_wait(
/*==========*/
os_event_t event) /* in: event to wait */
os_event_wait_low(
/*==============*/
os_event_t event, /* in: event to wait */
ib_longlong reset_sig_count)/* in: zero or the value
returned by previous call of
os_event_reset(). */
{
#ifdef __WIN__
DWORD err;
ut_a(event);
UT_NOT_USED(reset_sig_count);
/* Specify an infinite time limit for waiting */
err = WaitForSingleObject(event->handle, INFINITE);
@ -318,7 +398,11 @@ os_event_wait(
os_fast_mutex_lock(&(event->os_mutex));
old_signal_count = event->signal_count;
if (reset_sig_count) {
old_signal_count = reset_sig_count;
} else {
old_signal_count = event->signal_count;
}
for (;;) {
if (event->is_set == TRUE
@ -458,6 +542,7 @@ os_mutex_create(
mutex_str->handle = mutex;
mutex_str->count = 0;
mutex_str->event = os_event_create(NULL);
if (os_sync_mutex_inited) {
/* When creating os_sync_mutex itself we cannot reserve it */
@ -534,6 +619,10 @@ os_mutex_free(
{
ut_a(mutex);
if (!os_sync_free_called) {
os_event_free_internal(mutex->event);
}
if (os_sync_mutex_inited) {
os_mutex_enter(os_sync_mutex);
}

16
srv/srv0srv.c
View file

@ -1904,12 +1904,6 @@ loop:
os_thread_sleep(1000000);
/* In case mutex_exit is not a memory barrier, it is
theoretically possible some threads are left waiting though
the semaphore is already released. Wake up those threads: */
sync_arr_wake_threads_if_sema_free();
current_time = time(NULL);
time_elapsed = difftime(current_time, last_monitor_time);
@ -2106,9 +2100,15 @@ loop:
srv_refresh_innodb_monitor_stats();
}
/* In case mutex_exit is not a memory barrier, it is
theoretically possible some threads are left waiting though
the semaphore is already released. Wake up those threads: */
sync_arr_wake_threads_if_sema_free();
if (sync_array_print_long_waits()) {
fatal_cnt++;
if (fatal_cnt > 5) {
if (fatal_cnt > 10) {
fprintf(stderr,
"InnoDB: Error: semaphore wait has lasted"
@ -2128,7 +2128,7 @@ loop:
fflush(stderr);
os_thread_sleep(2000000);
os_thread_sleep(1000000);
if (srv_shutdown_state < SRV_SHUTDOWN_CLEANUP) {

398
sync/sync0arr.c
View file

@ -40,24 +40,23 @@ because we can do with a very small number of OS events,
say 200. In NT 3.51, allocating events seems to be a quadratic
algorithm, because 10 000 events are created fast, but
100 000 events takes a couple of minutes to create.
*/
As of 5.0.30 the above mentioned design is changed. Since now
OS can handle millions of wait events efficiently, we no longer
have this concept of each cell of wait array having one event.
Instead, now the event that a thread wants to wait on is embedded
in the wait object (mutex or rw_lock). We still keep the global
wait array for the sake of diagnostics and also to avoid infinite
wait The error_monitor thread scans the global wait array to signal
any waiting threads who have missed the signal. */
/* A cell where an individual thread may wait suspended
until a resource is released. The suspending is implemented
using an operating system event semaphore. */
struct sync_cell_struct {
/* State of the cell. SC_WAKING_UP means
sync_array_struct->n_reserved has been decremented, but the thread
in this cell has not waken up yet. When it does, it will set the
state to SC_FREE. Note that this is done without the protection of
any mutex. */
enum { SC_FREE, SC_RESERVED, SC_WAKING_UP } state;
void* wait_object; /* pointer to the object the
thread is waiting for; this is not
reseted to NULL when a cell is
freed. */
thread is waiting for; if NULL
the cell is free for use */
mutex_t* old_wait_mutex; /* the latest wait mutex in cell */
rw_lock_t* old_wait_rw_lock;/* the latest wait rw-lock in cell */
ulint request_type; /* lock type requested on the
@ -71,13 +70,23 @@ struct sync_cell_struct {
ibool waiting; /* TRUE if the thread has already
called sync_array_event_wait
on this cell */
ibool event_set; /* TRUE if the event is set */
os_event_t event; /* operating system event
semaphore handle */
ib_longlong signal_count; /* We capture the signal_count
of the wait_object when we
reset the event. This value is
then passed on to os_event_wait
and we wait only if the event
has not been signalled in the
period between the reset and
wait call. */
time_t reservation_time;/* time when the thread reserved
the wait cell */
};
/* NOTE: It is allowed for a thread to wait
for an event allocated for the array without owning the
protecting mutex (depending on the case: OS or database mutex), but
all changes (set or reset) to the state of the event must be made
while owning the mutex. */
struct sync_array_struct {
ulint n_reserved; /* number of currently reserved
cells in the wait array */
@ -220,12 +229,9 @@ sync_array_create(
for (i = 0; i < n_cells; i++) {
cell = sync_array_get_nth_cell(arr, i);
cell->state = SC_FREE;
cell->wait_object = NULL;
/* Create an operating system event semaphore with no name */
cell->event = os_event_create(NULL);
cell->event_set = FALSE; /* it is created in reset state */
cell->wait_object = NULL;
cell->waiting = FALSE;
cell->signal_count = 0;
}
return(arr);
@ -239,19 +245,12 @@ sync_array_free(
/*============*/
sync_array_t* arr) /* in, own: sync wait array */
{
ulint i;
sync_cell_t* cell;
ulint protection;
ut_a(arr->n_reserved == 0);
sync_array_validate(arr);
for (i = 0; i < arr->n_cells; i++) {
cell = sync_array_get_nth_cell(arr, i);
os_event_free(cell->event);
}
protection = arr->protection;
/* Release the mutex protecting the wait array complex */
@ -285,8 +284,7 @@ sync_array_validate(
for (i = 0; i < arr->n_cells; i++) {
cell = sync_array_get_nth_cell(arr, i);
if (cell->state == SC_RESERVED) {
if (cell->wait_object != NULL) {
count++;
}
}
@ -296,6 +294,29 @@ sync_array_validate(
sync_array_exit(arr);
}
/***********************************************************************
Puts the cell event in reset state. */
static
ib_longlong
sync_cell_event_reset(
/*==================*/
/* out: value of signal_count
at the time of reset. */
ulint type, /* in: lock type mutex/rw_lock */
void* object) /* in: the rw_lock/mutex object */
{
if (type == SYNC_MUTEX) {
return(os_event_reset(((mutex_t *) object)->event));
#ifdef __WIN__
} else if (type == RW_LOCK_WAIT_EX) {
return(os_event_reset(
((rw_lock_t *) object)->wait_ex_event));
#endif
} else {
return(os_event_reset(((rw_lock_t *) object)->event));
}
}
/**********************************************************************
Reserves a wait array cell for waiting for an object.
The event of the cell is reset to nonsignalled state. */
@ -324,21 +345,9 @@ sync_array_reserve_cell(
for (i = 0; i < arr->n_cells; i++) {
cell = sync_array_get_nth_cell(arr, i);
if (cell->state == SC_FREE) {
/* We do not check cell->event_set because it is
set outside the protection of the sync array mutex
and we had a bug regarding it, and since resetting
an event when it is not needed does no harm it is
safer always to do it. */
cell->event_set = FALSE;
os_event_reset(cell->event);
cell->state = SC_RESERVED;
cell->reservation_time = time(NULL);
cell->thread = os_thread_get_curr_id();
if (cell->wait_object == NULL) {
cell->waiting = FALSE;
cell->wait_object = object;
if (type == SYNC_MUTEX) {
@ -348,7 +357,6 @@ sync_array_reserve_cell(
}
cell->request_type = type;
cell->waiting = FALSE;
cell->file = file;
cell->line = line;
@ -359,6 +367,16 @@ sync_array_reserve_cell(
sync_array_exit(arr);
/* Make sure the event is reset and also store
the value of signal_count at which the event
was reset. */
cell->signal_count = sync_cell_event_reset(type,
object);
cell->reservation_time = time(NULL);
cell->thread = os_thread_get_curr_id();
return;
}
}
@ -368,68 +386,6 @@ sync_array_reserve_cell(
return;
}
/**********************************************************************
Frees the cell. Note that we don't have any mutex reserved when calling
this. */
static
void
sync_array_free_cell(
/*=================*/
sync_array_t* arr, /* in: wait array */
ulint index) /* in: index of the cell in array */
{
sync_cell_t* cell;
cell = sync_array_get_nth_cell(arr, index);
ut_a(cell->state == SC_WAKING_UP);
ut_a(cell->wait_object != NULL);
cell->state = SC_FREE;
}
/**********************************************************************
Frees the cell safely by reserving the sync array mutex and decrementing
n_reserved if necessary. Should only be called from mutex_spin_wait. */
void
sync_array_free_cell_protected(
/*===========================*/
sync_array_t* arr, /* in: wait array */
ulint index) /* in: index of the cell in array */
{
sync_cell_t* cell;
sync_array_enter(arr);
cell = sync_array_get_nth_cell(arr, index);
ut_a(cell->state != SC_FREE);
ut_a(cell->wait_object != NULL);
/* We only need to decrement n_reserved if it has not already been
done by sync_array_signal_object. */
if (cell->state == SC_RESERVED) {
ut_a(arr->n_reserved > 0);
arr->n_reserved--;
} else if (cell->state == SC_WAKING_UP) {
/* This is tricky; if we don't wait for the event to be
signaled, signal_object can set the state of a cell to
SC_WAKING_UP, mutex_spin_wait can call this and set the
state to SC_FREE, and then signal_object gets around to
calling os_set_event for the cell but since it's already
been freed things break horribly. */
sync_array_exit(arr);
os_event_wait(cell->event);
sync_array_enter(arr);
}
cell->state = SC_FREE;
sync_array_exit(arr);
}
/**********************************************************************
This function should be called when a thread starts to wait on
a wait array cell. In the debug version this function checks
@ -447,15 +403,28 @@ sync_array_wait_event(
ut_a(arr);
sync_array_enter(arr);
cell = sync_array_get_nth_cell(arr, index);
ut_a((cell->state == SC_RESERVED) || (cell->state == SC_WAKING_UP));
ut_a(cell->wait_object);
ut_a(!cell->waiting);
ut_ad(os_thread_get_curr_id() == cell->thread);
event = cell->event;
cell->waiting = TRUE;
if (cell->request_type == SYNC_MUTEX) {
event = ((mutex_t*) cell->wait_object)->event;
#ifdef __WIN__
/* On windows if the thread about to wait is the one which
has set the state of the rw_lock to RW_LOCK_WAIT_EX, then
it waits on a special event i.e.: wait_ex_event. */
} else if (cell->request_type == RW_LOCK_WAIT_EX) {
event = ((rw_lock_t*) cell->wait_object)->wait_ex_event;
#endif
} else {
event = ((rw_lock_t*) cell->wait_object)->event;
}
cell->waiting = TRUE;
#ifdef UNIV_SYNC_DEBUG
@ -464,7 +433,6 @@ sync_array_wait_event(
recursively sync_array routines, leading to trouble.
rw_lock_debug_mutex freezes the debug lists. */
sync_array_enter(arr);
rw_lock_debug_mutex_enter();
if (TRUE == sync_array_detect_deadlock(arr, cell, cell, 0)) {
@ -474,16 +442,16 @@ sync_array_wait_event(
}
rw_lock_debug_mutex_exit();
sync_array_exit(arr);
#endif
os_event_wait(event);
sync_array_exit(arr);
os_event_wait_low(event, cell->signal_count);
sync_array_free_cell(arr, index);
}
/**********************************************************************
Reports info of a wait array cell. Note: sync_array_print_long_waits()
calls this without mutex protection. */
Reports info of a wait array cell. */
static
void
sync_array_cell_print(
@ -503,17 +471,8 @@ sync_array_cell_print(
(ulong) os_thread_pf(cell->thread), cell->file,
(ulong) cell->line,
difftime(time(NULL), cell->reservation_time));
fprintf(file, "Wait array cell state %lu\n", (ulong)cell->state);
/* If the memory area pointed to by old_wait_mutex /
old_wait_rw_lock has been freed, this can crash. */
if (cell->state != SC_RESERVED) {
/* If cell has this state, then even if we are holding the sync
array mutex, the wait object may get freed meanwhile. Do not
print the wait object then. */
} else if (type == SYNC_MUTEX) {
if (type == SYNC_MUTEX) {
/* We use old_wait_mutex in case the cell has already
been freed meanwhile */
mutex = cell->old_wait_mutex;
@ -531,7 +490,11 @@ sync_array_cell_print(
#endif /* UNIV_SYNC_DEBUG */
(ulong) mutex->waiters);
} else if (type == RW_LOCK_EX || type == RW_LOCK_SHARED) {
} else if (type == RW_LOCK_EX
#ifdef __WIN__
|| type == RW_LOCK_WAIT_EX
#endif
|| type == RW_LOCK_SHARED) {
fputs(type == RW_LOCK_EX ? "X-lock on" : "S-lock on", file);
@ -565,8 +528,8 @@ sync_array_cell_print(
ut_error;
}
if (cell->event_set) {
fputs("wait is ending\n", file);
if (!cell->waiting) {
fputs("wait has ended\n", file);
}
}
@ -589,7 +552,7 @@ sync_array_find_thread(
cell = sync_array_get_nth_cell(arr, i);
if ((cell->state == SC_RESERVED)
if (cell->wait_object != NULL
&& os_thread_eq(cell->thread, thread)) {
return(cell); /* Found */
@ -679,7 +642,7 @@ sync_array_detect_deadlock(
depth++;
if (cell->event_set || !cell->waiting) {
if (!cell->waiting) {
return(FALSE); /* No deadlock here */
}
@ -704,10 +667,8 @@ sync_array_detect_deadlock(
depth);
if (ret) {
fprintf(stderr,
"Mutex %p owned by thread %lu"
" file %s line %lu\n",
(void*) mutex,
(ulong) os_thread_pf(mutex->thread_id),
"Mutex %p owned by thread %lu file %s line %lu\n",
mutex, (ulong) os_thread_pf(mutex->thread_id),
mutex->file_name, (ulong) mutex->line);
sync_array_cell_print(stderr, cell);
@ -717,7 +678,8 @@ sync_array_detect_deadlock(
return(FALSE); /* No deadlock */
} else if (cell->request_type == RW_LOCK_EX) {
} else if (cell->request_type == RW_LOCK_EX
|| cell->request_type == RW_LOCK_WAIT_EX) {
lock = cell->wait_object;
@ -816,7 +778,8 @@ sync_arr_cell_can_wake_up(
return(TRUE);
}
} else if (cell->request_type == RW_LOCK_EX) {
} else if (cell->request_type == RW_LOCK_EX
|| cell->request_type == RW_LOCK_WAIT_EX) {
lock = cell->wait_object;
@ -845,101 +808,47 @@ sync_arr_cell_can_wake_up(
return(FALSE);
}
/**************************************************************************
Looks for the cells in the wait array which refer to the wait object
specified, and sets their corresponding events to the signaled state. In this
way releases the threads waiting for the object to contend for the object.
It is possible that no such cell is found, in which case does nothing. */
/**********************************************************************
Frees the cell. NOTE! sync_array_wait_event frees the cell
automatically! */
void
sync_array_signal_object(
/*=====================*/
sync_array_free_cell(
/*=================*/
sync_array_t* arr, /* in: wait array */
void* object) /* in: wait object */
ulint index) /* in: index of the cell in array */
{
sync_cell_t* cell;
ulint count;
ulint i;
ulint res_count;
/* We store the addresses of cells we need to signal and signal
them only after we have released the sync array's mutex (for
performance reasons). cell_count is the number of such cells, and
cell_ptr points to the first one. If there are less than
UT_ARR_SIZE(cells) of them, cell_ptr == &cells[0], otherwise
cell_ptr points to malloc'd memory that we must free. */
sync_array_enter(arr);
sync_cell_t* cells[100];
sync_cell_t** cell_ptr = &cells[0];
ulint cell_count = 0;
ulint cell_max_count = UT_ARR_SIZE(cells);
cell = sync_array_get_nth_cell(arr, index);
ut_a(100 == cell_max_count);
ut_a(cell->wait_object != NULL);
cell->waiting = FALSE;
cell->wait_object = NULL;
cell->signal_count = 0;
ut_a(arr->n_reserved > 0);
arr->n_reserved--;
sync_array_exit(arr);
}
/**************************************************************************
Increments the signalled count. */
void
sync_array_object_signalled(
/*========================*/
sync_array_t* arr) /* in: wait array */
{
sync_array_enter(arr);
arr->sg_count++;
i = 0;
count = 0;
/* We need to store this to a local variable because it is modified
inside the loop */
res_count = arr->n_reserved;
while (count < res_count) {
cell = sync_array_get_nth_cell(arr, i);
if (cell->state == SC_RESERVED) {
count++;
if (cell->wait_object == object) {
cell->state = SC_WAKING_UP;
ut_a(arr->n_reserved > 0);
arr->n_reserved--;
if (cell_count == cell_max_count) {
sync_cell_t** old_cell_ptr = cell_ptr;
size_t old_size, new_size;
old_size = cell_max_count
* sizeof(sync_cell_t*);
cell_max_count *= 2;
new_size = cell_max_count
* sizeof(sync_cell_t*);
cell_ptr = malloc(new_size);
ut_a(cell_ptr);
memcpy(cell_ptr, old_cell_ptr,
old_size);
if (old_cell_ptr != &cells[0]) {
free(old_cell_ptr);
}
}
cell_ptr[cell_count] = cell;
cell_count++;
}
}
i++;
}
sync_array_exit(arr);
for (i = 0; i < cell_count; i++) {
cell = cell_ptr[i];
cell->event_set = TRUE;
os_event_set(cell->event);
}
if (cell_ptr != &cells[0]) {
free(cell_ptr);
}
}
/**************************************************************************
@ -959,33 +868,41 @@ sync_arr_wake_threads_if_sema_free(void)
sync_cell_t* cell;
ulint count;
ulint i;
ulint res_count;
sync_array_enter(arr);
i = 0;
count = 0;
/* We need to store this to a local variable because it is modified
inside the loop */
res_count = arr->n_reserved;
while (count < res_count) {
while (count < arr->n_reserved) {
cell = sync_array_get_nth_cell(arr, i);
if (cell->state == SC_RESERVED) {
if (cell->wait_object != NULL) {
count++;
if (sync_arr_cell_can_wake_up(cell)) {
cell->state = SC_WAKING_UP;
cell->event_set = TRUE;
os_event_set(cell->event);
ut_a(arr->n_reserved > 0);
arr->n_reserved--;
if (cell->request_type == SYNC_MUTEX) {
mutex_t* mutex;
mutex = cell->wait_object;
os_event_set(mutex->event);
#ifdef __WIN__
} else if (cell->request_type
== RW_LOCK_WAIT_EX) {
rw_lock_t* lock;
lock = cell->wait_object;
os_event_set(lock->wait_ex_event);
#endif
} else {
rw_lock_t* lock;
lock = cell->wait_object;
os_event_set(lock->event);
}
}
}
@ -1015,7 +932,7 @@ sync_array_print_long_waits(void)
cell = sync_array_get_nth_cell(sync_primary_wait_array, i);
if ((cell->state != SC_FREE)
if (cell->wait_object != NULL && cell->waiting
&& difftime(time(NULL), cell->reservation_time) > 240) {
fputs("InnoDB: Warning: a long semaphore wait:\n",
stderr);
@ -1023,7 +940,7 @@ sync_array_print_long_waits(void)
noticed = TRUE;
}
if ((cell->state != SC_FREE)
if (cell->wait_object != NULL && cell->waiting
&& difftime(time(NULL), cell->reservation_time)
> fatal_timeout) {
fatal = TRUE;
@ -1072,20 +989,25 @@ sync_array_output_info(
mutex */
{
sync_cell_t* cell;
ulint count;
ulint i;
fprintf(file,
"OS WAIT ARRAY INFO: reservation count %ld,"
" signal count %ld\n",
(long) arr->res_count,
(long) arr->sg_count);
for (i = 0; i < arr->n_cells; i++) {
"OS WAIT ARRAY INFO: reservation count %ld, signal count %ld\n",
(long) arr->res_count, (long) arr->sg_count);
i = 0;
count = 0;
while (count < arr->n_reserved) {
cell = sync_array_get_nth_cell(arr, i);
if (cell->state != SC_FREE) {
if (cell->wait_object != NULL) {
count++;
sync_array_cell_print(file, cell);
}
i++;
}
}

20
sync/sync0rw.c
View file

@ -151,6 +151,11 @@ rw_lock_create_func(
lock->last_x_file_name = "not yet reserved";
lock->last_s_line = 0;
lock->last_x_line = 0;
lock->event = os_event_create(NULL);
#ifdef __WIN__
lock->wait_ex_event = os_event_create(NULL);
#endif
mutex_enter(&rw_lock_list_mutex);
@ -184,6 +189,11 @@ rw_lock_free(
mutex_free(rw_lock_get_mutex(lock));
mutex_enter(&rw_lock_list_mutex);
os_event_free(lock->event);
#ifdef __WIN__
os_event_free(lock->wait_ex_event);
#endif
if (UT_LIST_GET_PREV(list, lock)) {
ut_a(UT_LIST_GET_PREV(list, lock)->magic_n == RW_LOCK_MAGIC_N);
@ -544,7 +554,15 @@ lock_loop:
rw_x_system_call_count++;
sync_array_reserve_cell(sync_primary_wait_array,
lock, RW_LOCK_EX,
lock,
#ifdef __WIN__
/* On windows RW_LOCK_WAIT_EX signifies
that this thread should wait on the
special wait_ex_event. */
(state == RW_LOCK_WAIT_EX)
? RW_LOCK_WAIT_EX :
#endif
RW_LOCK_EX,
file_name, line,
&index);

51
sync/sync0sync.c
View file

@ -95,17 +95,47 @@ have happened that the thread which was holding the mutex has just released
it and did not see the waiters byte set to 1, a case which would lead the
other thread to an infinite wait.
LEMMA 1: After a thread resets the event of the cell it reserves for waiting
========
for a mutex, some thread will eventually call sync_array_signal_object with
the mutex as an argument. Thus no infinite wait is possible.
LEMMA 1: After a thread resets the event of a mutex (or rw_lock), some
=======
thread will eventually call os_event_set() on that particular event.
Thus no infinite wait is possible in this case.
Proof: After making the reservation the thread sets the waiters field in the
mutex to 1. Then it checks that the mutex is still reserved by some thread,
or it reserves the mutex for itself. In any case, some thread (which may be
also some earlier thread, not necessarily the one currently holding the mutex)
will set the waiters field to 0 in mutex_exit, and then call
sync_array_signal_object with the mutex as an argument.
os_event_set() with the mutex as an argument.
Q.E.D.
LEMMA 2: If an os_event_set() call is made after some thread has called
=======
the os_event_reset() and before it starts wait on that event, the call
will not be lost to the second thread. This is true even if there is an
intervening call to os_event_reset() by another thread.
Thus no infinite wait is possible in this case.
Proof (non-windows platforms): os_event_reset() returns a monotonically
increasing value of signal_count. This value is increased at every
call of os_event_set() If thread A has called os_event_reset() followed
by thread B calling os_event_set() and then some other thread C calling
os_event_reset(), the is_set flag of the event will be set to FALSE;
but now if thread A calls os_event_wait_low() with the signal_count
value returned from the earlier call of os_event_reset(), it will
return immediately without waiting.
Q.E.D.
Proof (windows): If there is a writer thread which is forced to wait for
the lock, it may be able to set the state of rw_lock to RW_LOCK_WAIT_EX
The design of rw_lock ensures that there is one and only one thread
that is able to change the state to RW_LOCK_WAIT_EX and this thread is
guaranteed to acquire the lock after it is released by the current
holders and before any other waiter gets the lock.
On windows this thread waits on a separate event i.e.: wait_ex_event.
Since only one thread can wait on this event there is no chance
of this event getting reset before the writer starts wait on it.
Therefore, this thread is guaranteed to catch the os_set_event()
signalled unconditionally at the release of the lock.
Q.E.D. */
/* The number of system calls made in this module. Intended for performance
@ -217,6 +247,7 @@ mutex_create_func(
os_fast_mutex_init(&(mutex->os_fast_mutex));
mutex->lock_word = 0;
#endif
mutex->event = os_event_create(NULL);
mutex_set_waiters(mutex, 0);
#ifdef UNIV_DEBUG
mutex->magic_n = MUTEX_MAGIC_N;
@ -300,6 +331,8 @@ mutex_free(
mutex_exit(&mutex_list_mutex);
}
os_event_free(mutex->event);
#if !defined(_WIN32) || !defined(UNIV_CAN_USE_X86_ASSEMBLER)
os_fast_mutex_free(&(mutex->os_fast_mutex));
#endif
@ -509,8 +542,7 @@ spin_loop:
if (mutex_test_and_set(mutex) == 0) {
/* Succeeded! Free the reserved wait cell */
sync_array_free_cell_protected(sync_primary_wait_array,
index);
sync_array_free_cell(sync_primary_wait_array, index);
ut_d(mutex->thread_id = os_thread_get_curr_id());
#ifdef UNIV_SYNC_DEBUG
@ -591,8 +623,8 @@ mutex_signal_object(
/* The memory order of resetting the waiters field and
signaling the object is important. See LEMMA 1 above. */
sync_array_signal_object(sync_primary_wait_array, mutex);
os_event_set(mutex->event);
sync_array_object_signalled(sync_primary_wait_array);
}
#ifdef UNIV_SYNC_DEBUG
@ -1130,6 +1162,7 @@ sync_thread_add_level(
break;
case SYNC_TREE_NODE:
ut_a(sync_thread_levels_contain(array, SYNC_INDEX_TREE)
|| sync_thread_levels_contain(array, SYNC_DICT_OPERATION)
|| sync_thread_levels_g(array, SYNC_TREE_NODE - 1));
break;
case SYNC_TREE_NODE_NEW: