Commit f475a7f9 authored by Sergey Vojtovich's avatar Sergey Vojtovich

MDEV-6089 - MySQL WL#7305 "Improve MDL scalability by using lock-free hash"

Replaced m_ref_usage, m_ref_release, m_is_destroyed and m_version with single
m_state, which is updated atomically. Needed to get rid of MDL_map::m_mutex.
parent ecd3ff9f
......@@ -410,10 +410,7 @@ public:
MDL_lock(const MDL_key *key_arg)
: key(key_arg),
m_hog_lock_count(0),
m_ref_usage(0),
m_ref_release(0),
m_is_destroyed(FALSE),
m_version(0)
m_state(0)
{
mysql_prlock_init(key_MDL_lock_rwlock, &m_rwlock);
}
......@@ -423,47 +420,12 @@ public:
mysql_prlock_destroy(&m_rwlock);
}
inline static void destroy(MDL_lock *lock);
public:
/**
These three members are used to make it possible to separate
the MDL_map::m_mutex mutex and MDL_lock::m_rwlock in
MDL_map::find_or_insert() for increased scalability.
The 'm_is_destroyed' member is only set by destroyers that
have both the MDL_map::m_mutex and MDL_lock::m_rwlock, thus
holding any of the mutexes is sufficient to read it.
The 'm_ref_usage; is incremented under protection by
MDL_map::m_mutex, but when 'm_is_destroyed' is set to TRUE, this
member is moved to be protected by the MDL_lock::m_rwlock.
This means that the MDL_map::find_or_insert() which only
holds the MDL_lock::m_rwlock can compare it to 'm_ref_release'
without acquiring MDL_map::m_mutex again and if equal
it can also destroy the lock object safely.
The 'm_ref_release' is incremented under protection by
MDL_lock::m_rwlock.
Note since we are only interested in equality of these two
counters we don't have to worry about overflows as long as
their size is big enough to hold maximum number of concurrent
threads on the system.
*/
uint m_ref_usage;
uint m_ref_release;
bool m_is_destroyed;
/**
We use the same idea and an additional version counter to support
caching of unused MDL_lock object for further re-use.
This counter is incremented while holding both MDL_map::m_mutex
and MDL_lock::m_rwlock locks each time when a MDL_lock is moved from
the hash to the unused objects list (or destroyed).
A thread, which has found a MDL_lock object for the key in the hash
and then released the MDL_map::m_mutex before acquiring the
MDL_lock::m_rwlock, can determine that this object was moved to the
unused objects list (or destroyed) while it held no locks by comparing
the version value which it read while holding the MDL_map::m_mutex
with the value read after acquiring the MDL_lock::m_rwlock.
Note that since it takes several years to overflow this counter such
theoretically possible overflows should not have any practical effects.
Lock state: first 31 bits are reference counter, 32-nd bit is deleted flag.
*/
ulonglong m_version;
static const int32 DELETED= 1 << 31;
int32 m_state;
};
......@@ -648,7 +610,7 @@ int mdl_iterate(int (*callback)(MDL_ticket *ticket, void *arg), void *arg)
for (i= 0; i < mdl_locks.m_locks.records; i++)
{
MDL_lock *lock= (MDL_lock*) my_hash_element(&mdl_locks.m_locks, i);
lock->m_ref_usage++;
my_atomic_add32_explicit(&lock->m_state, 1, MY_MEMORY_ORDER_RELAXED);
insert_dynamic(&locks, &lock);
}
mysql_mutex_unlock(&mdl_locks.m_mutex);
......@@ -657,19 +619,15 @@ int mdl_iterate(int (*callback)(MDL_ticket *ticket, void *arg), void *arg)
for (i= 0; i < locks.elements; i++)
{
MDL_lock *lock= (MDL_lock*) *dynamic_element(&locks, i, MDL_lock**);
res= mdl_iterate_lock(lock, callback, arg);
res|= mdl_iterate_lock(lock, callback, arg);
mysql_prlock_wrlock(&lock->m_rwlock);
uint ref_usage= lock->m_ref_usage;
uint ref_release= ++lock->m_ref_release;
bool is_destroyed= lock->m_is_destroyed;
int32 old_state= my_atomic_add32_explicit(&lock->m_state, -1,
MY_MEMORY_ORDER_RELAXED);
mysql_prlock_unlock(&lock->m_rwlock);
if (unlikely(is_destroyed && ref_usage == ref_release))
if (unlikely(old_state == MDL_lock::DELETED + 1))
MDL_lock::destroy(lock);
if (res)
break;
}
end:
delete_dynamic(&locks);
......@@ -759,10 +717,7 @@ retry:
mdl_key->ptr(),
mdl_key->length())))
{
/*
No lock object found so we need to create a new one
or reuse an existing unused object.
*/
/* No lock object found so we need to create a new one. */
lock= MDL_lock::create(mdl_key);
if (!lock || my_hash_insert(&m_locks, (uchar*)lock))
{
......@@ -791,61 +746,26 @@ retry:
bool MDL_map::move_from_hash_to_lock_mutex(MDL_lock *lock)
{
ulonglong version;
DBUG_ASSERT(! lock->m_is_destroyed);
DBUG_ASSERT(!(my_atomic_load32(&lock->m_state) & MDL_lock::DELETED));
mysql_mutex_assert_owner(&m_mutex);
/*
We increment m_ref_usage which is a reference counter protected by
MDL_map::m_mutex under the condition it is present in the hash
and m_is_destroyed is FALSE.
*/
lock->m_ref_usage++;
/* Read value of the version counter under protection of m_mutex lock. */
version= lock->m_version;
my_atomic_add32_explicit(&lock->m_state, 1, MY_MEMORY_ORDER_RELAXED);
mysql_mutex_unlock(&m_mutex);
mysql_prlock_wrlock(&lock->m_rwlock);
lock->m_ref_release++;
if (unlikely(lock->m_version != version))
int32 old_state= my_atomic_add32_explicit(&lock->m_state, -1,
MY_MEMORY_ORDER_RELAXED);
if (unlikely(old_state & MDL_lock::DELETED))
{
/*
If the current value of version differs from one that was read while
we held m_mutex mutex, this MDL_lock object was moved to the unused
objects list or destroyed while we held no locks.
We should retry our search. But first we should destroy the MDL_lock
object if necessary.
Object was released while we held no locks, we need to
release it if no others hold references to it, while our own
reference count ensured that the object as such haven't got
its memory released yet.
*/
if (unlikely(lock->m_is_destroyed))
{
/*
Object was released while we held no locks, we need to
release it if no others hold references to it, while our own
reference count ensured that the object as such haven't got
its memory released yet. We can also safely compare
m_ref_usage and m_ref_release since the object is no longer
present in the hash (or unused objects list) so no one will
be able to find it and increment m_ref_usage anymore.
*/
uint ref_usage= lock->m_ref_usage;
uint ref_release= lock->m_ref_release;
mysql_prlock_unlock(&lock->m_rwlock);
if (ref_usage == ref_release)
MDL_lock::destroy(lock);
}
else
{
/*
Object was not destroyed but its version has changed.
This means that it was moved to the unused objects list
(and even might be already re-used). So now it might
correspond to a different key, therefore we should simply
retry our search.
*/
mysql_prlock_unlock(&lock->m_rwlock);
}
mysql_prlock_unlock(&lock->m_rwlock);
if (old_state == MDL_lock::DELETED + 1)
MDL_lock::destroy(lock);
return TRUE;
}
return FALSE;
......@@ -907,40 +827,17 @@ void MDL_map::remove(MDL_lock *lock)
mysql_mutex_lock(&m_mutex);
my_hash_delete(&m_locks, (uchar*) lock);
/*
To let threads holding references to the MDL_lock object know that it was
moved to the list of unused objects or destroyed, we increment the version
counter under protection of both MDL_map::m_mutex and
MDL_lock::m_rwlock locks. This allows us to read the version value while
having either one of those locks.
*/
lock->m_version++;
/*
Destroy the MDL_lock object, but ensure that anyone that is
holding a reference to the object is not remaining, if so he
has the responsibility to release it.
Setting of m_is_destroyed to TRUE while holding _both_
MDL_map::m_mutex and MDL_lock::m_rwlock mutexes transfers
the protection of m_ref_usage from MDL_map::m_mutex to
MDL_lock::m_rwlock while removal of the object from the hash
(and cache of unused objects) makes it read-only. Therefore
whoever acquires MDL_lock::m_rwlock next will see the most up
to date version of m_ref_usage.
This means that when m_is_destroyed is TRUE and we hold the
MDL_lock::m_rwlock we can safely read the m_ref_usage
member.
*/
uint ref_usage, ref_release;
lock->m_is_destroyed= TRUE;
ref_usage= lock->m_ref_usage;
ref_release= lock->m_ref_release;
int32 old_state= my_atomic_add32_explicit(&lock->m_state, MDL_lock::DELETED,
MY_MEMORY_ORDER_RELAXED);
mysql_mutex_unlock(&m_mutex);
mysql_prlock_unlock(&lock->m_rwlock);
if (ref_usage == ref_release)
if (!old_state)
MDL_lock::destroy(lock);
}
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment