FS-10167: fixed a couple deadlock issues and a misconception about the locks on hash

This commit is contained in:
Shane Bryldt committed 2017-04-17 11:10:20 -06:00
1 parent 37f9b2afdc
commit 8d4eac7f69
3 files changed
+23 -15

No files matched your search

+1 -1
View File
@@ -319,8 +319,8 @@ KS_DECLARE(void) blade_session_state_set(blade_session_t *bs, blade_session_stat
blade_handle_session_state_callbacks_execute(bs, BLADE_SESSION_STATE_CONDITION_PRE);
ks_cond_try_signal(bs->cond);
ks_mutex_unlock(bs->mutex);
ks_cond_try_signal(bs->cond);
}
KS_DECLARE(void) blade_session_hangup(blade_session_t *bs)
+16 -8
View File
@@ -196,22 +196,23 @@ KS_DECLARE(ks_status_t) blade_handle_create(blade_handle_t **bhP, ks_pool_t *poo
bh->pool = pool;
bh->tpool = tpool;
ks_hash_create(&bh->transports, KS_HASH_MODE_CASE_INSENSITIVE, KS_HASH_FLAG_NOLOCK | KS_HASH_FLAG_DUP_CHECK, bh->pool);
// @todo this needs to be reviewed, NOLOCK is incorrect, but RWLOCK causes deadlock somewhere
ks_hash_create(&bh->transports, KS_HASH_MODE_CASE_INSENSITIVE, KS_HASH_FLAG_RWLOCK | KS_HASH_FLAG_DUP_CHECK, bh->pool);
ks_assert(bh->transports);
ks_hash_create(&bh->spaces, KS_HASH_MODE_CASE_INSENSITIVE, KS_HASH_FLAG_NOLOCK | KS_HASH_FLAG_DUP_CHECK, bh->pool);
ks_hash_create(&bh->spaces, KS_HASH_MODE_CASE_INSENSITIVE, KS_HASH_FLAG_RWLOCK | KS_HASH_FLAG_DUP_CHECK, bh->pool);
ks_assert(bh->spaces);
ks_hash_create(&bh->events, KS_HASH_MODE_CASE_INSENSITIVE, KS_HASH_FLAG_NOLOCK | KS_HASH_FLAG_DUP_CHECK, bh->pool);
ks_hash_create(&bh->events, KS_HASH_MODE_CASE_INSENSITIVE, KS_HASH_FLAG_RWLOCK | KS_HASH_FLAG_DUP_CHECK, bh->pool);
ks_assert(bh->events);
ks_hash_create(&bh->connections, KS_HASH_MODE_CASE_INSENSITIVE, KS_HASH_FLAG_NOLOCK | KS_HASH_FLAG_DUP_CHECK, bh->pool);
ks_hash_create(&bh->connections, KS_HASH_MODE_CASE_INSENSITIVE, KS_HASH_FLAG_RWLOCK | KS_HASH_FLAG_DUP_CHECK, bh->pool);
ks_assert(bh->connections);
ks_hash_create(&bh->sessions, KS_HASH_MODE_CASE_INSENSITIVE, KS_HASH_FLAG_NOLOCK | KS_HASH_FLAG_DUP_CHECK, bh->pool);
ks_hash_create(&bh->sessions, KS_HASH_MODE_CASE_INSENSITIVE, KS_HASH_FLAG_RWLOCK | KS_HASH_FLAG_DUP_CHECK, bh->pool);
ks_assert(bh->sessions);
ks_hash_create(&bh->session_state_callbacks, KS_HASH_MODE_CASE_INSENSITIVE, KS_HASH_FLAG_NOLOCK | KS_HASH_FLAG_DUP_CHECK, bh->pool);
ks_hash_create(&bh->session_state_callbacks, KS_HASH_MODE_CASE_INSENSITIVE, KS_HASH_FLAG_RWLOCK | KS_HASH_FLAG_DUP_CHECK, bh->pool);
ks_assert(bh->session_state_callbacks);
ks_hash_create(&bh->requests, KS_HASH_MODE_CASE_INSENSITIVE, KS_HASH_FLAG_NOLOCK | KS_HASH_FLAG_DUP_CHECK, bh->pool);
ks_hash_create(&bh->requests, KS_HASH_MODE_CASE_INSENSITIVE, KS_HASH_FLAG_RWLOCK | KS_HASH_FLAG_DUP_CHECK, bh->pool);
ks_assert(bh->requests);
*bhP = bh;
@@ -665,6 +666,13 @@ KS_DECLARE(blade_session_t *) blade_handle_sessions_get(blade_handle_t *bh, cons
ks_assert(bh);
ks_assert(sid);
// @todo consider using blade_session_t via reference counting, rather than locking a mutex to simulate a reference count to halt cleanups while in use
// using actual reference counting would mean that mutexes would not need to be held locked when looking up a session by id just to prevent cleanup,
// instead cleanup would automatically occur when the last reference is actually removed (which SHOULD be at the end of the state machine thread),
// which is safer than another thread potentially waiting on the write lock to release while it's being destroyed, or external code forgetting to unlock
// then use short lived mutex or rwl for accessing the content of the session while it is referenced
// this approach should also be used for blade_connection_t, which has a similar threaded state machine
ks_hash_read_lock(bh->sessions);
bs = ks_hash_search(bh->sessions, (void *)sid, KS_UNLOCKED);
if (bs && blade_session_read_lock(bs, KS_FALSE) != KS_STATUS_SUCCESS) bs = NULL;
@@ -766,7 +774,7 @@ KS_DECLARE(ks_status_t) blade_handle_session_state_callback_unregister(blade_han
ks_hash_write_lock(bh->session_state_callbacks);
bhsscr = (blade_handle_session_state_callback_registration_t *)ks_hash_remove(bh->session_state_callbacks, (void *)id);
if (!bhsscr) ret = KS_STATUS_FAIL;
ks_hash_write_lock(bh->session_state_callbacks);
ks_hash_write_unlock(bh->session_state_callbacks);
if (bhsscr) blade_handle_session_state_callback_registration_destroy(&bhsscr);