summaryrefslogtreecommitdiff
path: root/storage/innobase/handler/ha_innodb.cc
diff options
context:
space:
mode:
Diffstat (limited to 'storage/innobase/handler/ha_innodb.cc')
-rw-r--r--storage/innobase/handler/ha_innodb.cc189
1 files changed, 114 insertions, 75 deletions
diff --git a/storage/innobase/handler/ha_innodb.cc b/storage/innobase/handler/ha_innodb.cc
index 74c04ec1607..3ad4aa934f2 100644
--- a/storage/innobase/handler/ha_innodb.cc
+++ b/storage/innobase/handler/ha_innodb.cc
@@ -5167,17 +5167,21 @@ UNIV_INTERN void lock_cancel_waiting_and_release(lock_t* lock);
@sa THD::awake() @sa ha_kill_query() */
static void innobase_kill_query(handlerton*, THD* thd, enum thd_kill_levels)
{
- DBUG_ENTER("innobase_kill_query");
+ DBUG_ENTER("innobase_kill_query");
#ifdef WITH_WSREP
- if (wsrep_thd_get_conflict_state(thd) != NO_CONFLICT) {
- /* if victim has been signaled by BF thread and/or aborting
- is already progressing, following query aborting is not necessary
- any more.
- Also, BF thread should own trx mutex for the victim, which would
- conflict with trx_mutex_enter() below
- */
- DBUG_VOID_RETURN;
- }
+ if (wsrep_thd_get_conflict_state(thd) != NO_CONFLICT)
+ {
+ /* if victim has been signaled by BF thread and/or aborting
+ is already progressing, following query aborting is not necessary
+ any more. E.g. wsrep_innobase_kill_one_trx().
+ Also, BF thread should own trx mutex for the victim, which would
+ conflict with trx_mutex_enter() below
+ */
+ WSREP_DEBUG("Victim thread %ld bail out conflict_state %s query %s",
+ thd_get_thread_id(thd),
+ wsrep_thd_conflict_state_str(thd), wsrep_thd_query(thd));
+ DBUG_VOID_RETURN;
+ }
#endif /* WITH_WSREP */
if (trx_t* trx= thd_to_trx(thd))
@@ -18799,9 +18803,9 @@ static struct st_mysql_storage_engine innobase_storage_engine=
{ MYSQL_HANDLERTON_INTERFACE_VERSION };
#ifdef WITH_WSREP
+static
void
wsrep_abort_slave_trx(
-/*==================*/
wsrep_seqno_t bf_seqno,
wsrep_seqno_t victim_seqno)
{
@@ -18811,19 +18815,17 @@ wsrep_abort_slave_trx(
"2) a bug in the code.\n\t"
"3) a database corruption.\n Node consistency compromized, "
"need to abort. Restart the node to resync with cluster.",
- (long long)bf_seqno, (long long)victim_seqno);
+ bf_seqno, victim_seqno);
abort();
}
/*******************************************************************//**
This function is used to kill one transaction in BF. */
-UNIV_INTERN
void
wsrep_innobase_kill_one_trx(
-/*========================*/
MYSQL_THD const bf_thd,
const trx_t * const bf_trx,
trx_t *victim_trx,
- ibool signal)
+ my_bool signal)
{
ut_ad(bf_thd);
ut_ad(victim_trx);
@@ -18831,38 +18833,41 @@ wsrep_innobase_kill_one_trx(
ut_ad(trx_mutex_own(victim_trx));
DBUG_ENTER("wsrep_innobase_kill_one_trx");
- THD *thd = (THD *) victim_trx->mysql_thd;
- int64_t bf_seqno = wsrep_thd_trx_seqno(bf_thd);
+ THD *thd= (THD *) victim_trx->mysql_thd;
+ int64_t bf_seqno= wsrep_thd_trx_seqno(bf_thd);
if (!thd) {
- DBUG_PRINT("wsrep", ("no thd for conflicting lock"));
WSREP_WARN("no THD for trx: " TRX_ID_FMT, victim_trx->id);
DBUG_VOID_RETURN;
}
- WSREP_LOG_CONFLICT(bf_thd, thd, TRUE);
-
- WSREP_DEBUG("BF kill (" ULINTPF ", seqno: " INT64PF
- "), victim: (%lu) trx: " TRX_ID_FMT,
- signal, bf_seqno,
- thd_get_thread_id(thd),
- victim_trx->id);
-
- WSREP_DEBUG("Aborting query: %s conf %d trx: %" PRId64,
- (thd && wsrep_thd_query(thd)) ? wsrep_thd_query(thd) : "void",
- wsrep_thd_conflict_state(thd, FALSE),
- wsrep_thd_ws_handle(thd)->trx_id);
-
+ /* Here we need to lock THD::LOCK_thd_data to protect from
+ concurrent usage or disconnect or delete. */
+ DEBUG_SYNC(bf_thd, "wsrep_before_BF_victim_lock");
wsrep_thd_LOCK(thd);
- DBUG_EXECUTE_IF("sync.wsrep_after_BF_victim_lock",
- {
- const char act[]=
- "now "
- "wait_for signal.wsrep_after_BF_victim_lock";
- DBUG_ASSERT(!debug_sync_set_action(bf_thd,
- STRING_WITH_LEN(act)));
- };);
+ DEBUG_SYNC(bf_thd, "wsrep_after_BF_victim_lock");
+
+ WSREP_DEBUG("Aborter %s trx_id: " TRX_ID_FMT " thread: %ld "
+ "seqno: %lld query_state: %s conflict_state: %s query: %s",
+ wsrep_thd_is_BF(bf_thd, false) ? "BF" : "normal",
+ bf_trx ? bf_trx->id : TRX_ID_MAX,
+ thd_get_thread_id(bf_thd),
+ bf_seqno,
+ wsrep_thd_query_state_str(bf_thd),
+ wsrep_thd_conflict_state_str(bf_thd),
+ wsrep_thd_query(bf_thd));
+
+ WSREP_DEBUG("Victim %s trx_id: " TRX_ID_FMT " thread: %ld "
+ "seqno: %lld query_state: %s conflict_state: %s query: %s",
+ wsrep_thd_is_BF(thd, false) ? "BF" : "normal",
+ victim_trx->id,
+ thd_get_thread_id(thd),
+ wsrep_thd_trx_seqno(thd),
+ wsrep_thd_query_state_str(thd),
+ wsrep_thd_conflict_state_str(thd),
+ wsrep_thd_query(thd));
+ WSREP_LOG_CONFLICT(bf_thd, thd, TRUE);
if (wsrep_thd_query_state(thd) == QUERY_EXITING) {
WSREP_DEBUG("kill trx EXITING for " TRX_ID_FMT,
@@ -18872,27 +18877,32 @@ wsrep_innobase_kill_one_trx(
}
if (wsrep_thd_exec_mode(thd) != LOCAL_STATE) {
- WSREP_DEBUG("withdraw for BF trx: " TRX_ID_FMT ", state: %d",
+ WSREP_DEBUG("withdraw for BF trx: " TRX_ID_FMT
+ ", state: %s exec %s",
victim_trx->id,
- wsrep_thd_get_conflict_state(thd));
+ wsrep_thd_conflict_state_str(thd),
+ wsrep_thd_exec_mode_str(thd));
}
switch (wsrep_thd_get_conflict_state(thd)) {
case NO_CONFLICT:
+ /* This will cause any call to innobase_kill_query()
+ for this thd to bail out. */
wsrep_thd_set_conflict_state(thd, MUST_ABORT);
break;
case MUST_ABORT:
WSREP_DEBUG("victim " TRX_ID_FMT " in MUST ABORT state",
victim_trx->id);
- wsrep_thd_UNLOCK(thd);
wsrep_thd_awake(thd, signal);
+ wsrep_thd_UNLOCK(thd);
DBUG_VOID_RETURN;
break;
case ABORTED:
case ABORTING: // fall through
default:
- WSREP_DEBUG("victim " TRX_ID_FMT " in state %d",
- victim_trx->id, wsrep_thd_get_conflict_state(thd));
+ WSREP_DEBUG("victim " TRX_ID_FMT " in state %s",
+ victim_trx->id,
+ wsrep_thd_conflict_state_str(thd));
wsrep_thd_UNLOCK(thd);
DBUG_VOID_RETURN;
break;
@@ -18922,8 +18932,8 @@ wsrep_innobase_kill_one_trx(
WSREP_DEBUG("cancel commit warning: "
TRX_ID_FMT,
victim_trx->id);
- wsrep_thd_UNLOCK(thd);
wsrep_thd_awake(thd, signal);
+ wsrep_thd_UNLOCK(thd);
DBUG_VOID_RETURN;
break;
case WSREP_OK:
@@ -18941,8 +18951,8 @@ wsrep_innobase_kill_one_trx(
break;
}
}
- wsrep_thd_UNLOCK(thd);
wsrep_thd_awake(thd, signal);
+ wsrep_thd_UNLOCK(thd);
break;
case QUERY_EXEC:
/* it is possible that victim trx is itself waiting for some
@@ -18964,23 +18974,19 @@ wsrep_innobase_kill_one_trx(
lock_cancel_waiting_and_release(wait_lock);
}
- wsrep_thd_UNLOCK(thd);
wsrep_thd_awake(thd, signal);
+ wsrep_thd_UNLOCK(thd);
} else {
/* abort currently executing query */
- DBUG_PRINT("wsrep",("sending KILL_QUERY to: %lu",
- thd_get_thread_id(thd)));
WSREP_DEBUG("kill query for: %ld",
thd_get_thread_id(thd));
- /* Note that innobase_kill_query will take lock_mutex
- and trx_mutex */
- wsrep_thd_UNLOCK(thd);
wsrep_thd_awake(thd, signal);
+ wsrep_thd_UNLOCK(thd);
/* for BF thd, we need to prevent him from committing */
if (wsrep_thd_exec_mode(thd) == REPL_RECV) {
wsrep_abort_slave_trx(bf_seqno,
- wsrep_thd_trx_seqno(thd));
+ wsrep_thd_trx_seqno(thd));
}
}
break;
@@ -18990,29 +18996,27 @@ wsrep_innobase_kill_one_trx(
if (wsrep_thd_exec_mode(thd) == REPL_RECV) {
WSREP_DEBUG("kill BF IDLE, seqno: %lld",
- (long long)wsrep_thd_trx_seqno(thd));
+ wsrep_thd_trx_seqno(thd));
wsrep_thd_UNLOCK(thd);
wsrep_abort_slave_trx(bf_seqno,
wsrep_thd_trx_seqno(thd));
DBUG_VOID_RETURN;
}
- /* This will lock thd from proceeding after net_read() */
+ /* This will lock thd from proceeding after net_read()
+ and innobase_kill_query to bail out for this thd. */
wsrep_thd_set_conflict_state(thd, ABORTING);
wsrep_lock_rollback();
if (wsrep_aborting_thd_contains(thd)) {
WSREP_WARN("duplicate thd aborter %lu",
- (ulong) thd_get_thread_id(thd));
+ thd_get_thread_id(thd));
} else {
wsrep_aborting_thd_enqueue(thd);
- DBUG_PRINT("wsrep",("enqueuing trx abort for %lu",
- thd_get_thread_id(thd)));
WSREP_DEBUG("enqueuing trx abort for (%lu)",
- thd_get_thread_id(thd));
+ thd_get_thread_id(thd));
}
- DBUG_PRINT("wsrep",("signalling wsrep rollbacker"));
WSREP_DEBUG("signaling aborter");
wsrep_unlock_rollback();
wsrep_thd_UNLOCK(thd);
@@ -19020,10 +19024,7 @@ wsrep_innobase_kill_one_trx(
break;
}
default:
- WSREP_WARN("bad wsrep query state: %d",
- wsrep_thd_query_state(thd));
- wsrep_thd_UNLOCK(thd);
- break;
+ ut_error;
}
DBUG_VOID_RETURN;
@@ -19032,7 +19033,6 @@ wsrep_innobase_kill_one_trx(
static
void
wsrep_abort_transaction(
-/*====================*/
handlerton*,
THD *bf_thd,
THD *victim_thd,
@@ -19040,27 +19040,66 @@ wsrep_abort_transaction(
{
DBUG_ENTER("wsrep_abort_transaction");
- trx_t* victim_trx = thd_to_trx(victim_thd);
- trx_t* bf_trx = (bf_thd) ? thd_to_trx(bf_thd) : NULL;
-
- WSREP_DEBUG("abort transaction: BF: %s victim: %s victim conf: %d",
- wsrep_thd_query(bf_thd),
- wsrep_thd_query(victim_thd),
- wsrep_thd_conflict_state(victim_thd, FALSE));
+ ut_ad(bf_thd);
+ ut_ad(victim_thd);
+ trx_t* victim_trx= thd_to_trx(victim_thd);
+ trx_t* bf_trx= thd_to_trx(bf_thd);
+
+ /* Here we should hold THD::LOCK_thd_data to protect
+ victim from concurrent usage and THD::LOCK_thd_kill
+ to protect from disconnect or delete. */
+ WSREP_DEBUG("wsrep_abort_transaction: BF:"
+ " thread %ld query_state %s conflict_state %s"
+ " exec %s query %s trx " TRX_ID_FMT,
+ thd_get_thread_id(bf_thd),
+ wsrep_thd_query_state_str(bf_thd),
+ wsrep_thd_conflict_state_str(bf_thd),
+ wsrep_thd_exec_mode_str(bf_thd),
+ wsrep_thd_query(bf_thd),
+ bf_trx ? bf_trx->id : 0);
+
+ WSREP_DEBUG("wsrep_abort_transaction: victim:"
+ " thread %ld query_state %s conflict_state %s"
+ " exec %s query %s trx " TRX_ID_FMT,
+ thd_get_thread_id(victim_thd),
+ wsrep_thd_query_state_str(victim_thd),
+ wsrep_thd_conflict_state_str(victim_thd),
+ wsrep_thd_exec_mode_str(victim_thd),
+ wsrep_thd_query(victim_thd),
+ victim_trx ? victim_trx->id : 0);
if (victim_trx) {
+ WSREP_DEBUG("wsrep_abort_transaction: Victim thread %ld "
+ "transaction " TRX_ID_FMT " trx_state %d",
+ thd_get_thread_id(victim_thd),
+ victim_trx->id,
+ victim_trx->state);
+ /* This is necessary as correct mutexing order is
+ lock_sys -> trx -> THD::LOCK_thd_data and below
+ function assumes we have lock_sys and trx locked
+ and takes THD::LOCK_thd_data for THD state check. */
+ wsrep_thd_UNLOCK(victim_thd);
+ DEBUG_SYNC(bf_thd, "wsrep_abort_victim_unlocked");
+ DBUG_EXECUTE_IF("wsrep_abort_replicated_sleep",
+ WSREP_DEBUG("wsrep_abort_transaction: sleeping "
+ "for thread %ld ",
+ thd_get_thread_id(victim_thd));
+ my_sleep(100000););
lock_mutex_enter();
trx_mutex_enter(victim_trx);
wsrep_innobase_kill_one_trx(bf_thd, bf_trx, victim_trx, signal);
lock_mutex_exit();
trx_mutex_exit(victim_trx);
wsrep_srv_conc_cancel_wait(victim_trx);
+ wsrep_thd_LOCK(victim_thd);
DBUG_VOID_RETURN;
} else {
- WSREP_DEBUG("victim does not have transaction");
- wsrep_thd_LOCK(victim_thd);
+ WSREP_DEBUG("wsrep_abort_transaction: Victim thread %ld "
+ "no transaction",
+ thd_get_thread_id(victim_thd));
+ /* This will cause any call to innobase_kill_query()
+ for this thd to bail out. */
wsrep_thd_set_conflict_state(victim_thd, MUST_ABORT);
- wsrep_thd_UNLOCK(victim_thd);
wsrep_thd_awake(victim_thd, signal);
}