void WiredTigerSessionCache::waitUntilDurable(bool forceCheckpoint) {
const int shuttingDown = _shuttingDown.fetchAndAdd(1);
ON_BLOCK_EXIT([this] { _shuttingDown.fetchAndSubtract(1); });
uassert(ErrorCodes::ShutdownInProgress,
"Cannot wait for durability because a shutdown is in progress",
!(shuttingDown & kShuttingDownMask));
// When forcing a checkpoint with journaling enabled, don't synchronize with other
// waiters, as a log flush is much cheaper than a full checkpoint.
if (forceCheckpoint && _engine->isDurable()) {
UniqueWiredTigerSession session = getSession();
WT_SESSION* s = session->getSession();
{
stdx::unique_lock<stdx::mutex> lk(_journalListenerMutex);
JournalListener::Token token = _journalListener->getToken();
invariantWTOK(s->checkpoint(s, NULL));
_journalListener->onDurable(token);
}
LOG(4) << "created checkpoint (forced)";
return;
}
uint32_t start = _lastSyncTime.load();
// Do the remainder in a critical section that ensures only a single thread at a time
// will attempt to synchronize.
stdx::unique_lock<stdx::mutex> lk(_lastSyncMutex);
uint32_t current = _lastSyncTime.loadRelaxed(); // synchronized with writes through mutex
if (current != start) {
// Someone else synced already since we read lastSyncTime, so we're done!
return;
}
_lastSyncTime.store(current + 1);
// Nobody has synched yet, so we have to sync ourselves.
auto session = getSession();
WT_SESSION* s = session->getSession();
// This gets the token (OpTime) from the last write, before flushing (either the journal, or a
// checkpoint), and then reports that token (OpTime) as a durable write.
stdx::unique_lock<stdx::mutex> jlk(_journalListenerMutex);
JournalListener::Token token = _journalListener->getToken();
// Use the journal when available, or a checkpoint otherwise.
if (_engine->isDurable()) {
invariantWTOK(s->log_flush(s, "sync=on"));
LOG(4) << "flushed journal";
} else {
invariantWTOK(s->checkpoint(s, NULL));
LOG(4) << "created checkpoint";
}
_journalListener->onDurable(token);
}
void WiredTigerSessionCache::waitUntilDurable(bool forceCheckpoint, bool stableCheckpoint) {
// For inMemory storage engines, the data is "as durable as it's going to get".
// That is, a restart is equivalent to a complete node failure.
if (isEphemeral()) {
return;
}
const int shuttingDown = _shuttingDown.fetchAndAdd(1);
ON_BLOCK_EXIT([this] { _shuttingDown.fetchAndSubtract(1); });
uassert(ErrorCodes::ShutdownInProgress,
"Cannot wait for durability because a shutdown is in progress",
!(shuttingDown & kShuttingDownMask));
// Stable checkpoints are only meaningful in a replica set. Replication sets the "stable
// timestamp". If the stable timestamp is unset, WiredTiger takes a full checkpoint, which is
// incidentally what we want. A "true" stable checkpoint (a stable timestamp was set on the
// WT_CONNECTION, i.e: replication is on) requires `forceCheckpoint` to be true and journaling
// to be enabled.
if (stableCheckpoint && getGlobalReplSettings().usingReplSets()) {
invariant(forceCheckpoint && _engine->isDurable());
}
// When forcing a checkpoint with journaling enabled, don't synchronize with other
// waiters, as a log flush is much cheaper than a full checkpoint.
if (forceCheckpoint && _engine->isDurable()) {
UniqueWiredTigerSession session = getSession();
WT_SESSION* s = session->getSession();
{
stdx::unique_lock<stdx::mutex> lk(_journalListenerMutex);
JournalListener::Token token = _journalListener->getToken();
auto config = stableCheckpoint ? "use_timestamp=true" : "use_timestamp=false";
invariantWTOK(s->checkpoint(s, config));
_journalListener->onDurable(token);
}
LOG(4) << "created checkpoint (forced)";
return;
}
uint32_t start = _lastSyncTime.load();
// Do the remainder in a critical section that ensures only a single thread at a time
// will attempt to synchronize.
stdx::unique_lock<stdx::mutex> lk(_lastSyncMutex);
uint32_t current = _lastSyncTime.loadRelaxed(); // synchronized with writes through mutex
if (current != start) {
// Someone else synced already since we read lastSyncTime, so we're done!
return;
}
_lastSyncTime.store(current + 1);
// Nobody has synched yet, so we have to sync ourselves.
// This gets the token (OpTime) from the last write, before flushing (either the journal, or a
// checkpoint), and then reports that token (OpTime) as a durable write.
stdx::unique_lock<stdx::mutex> jlk(_journalListenerMutex);
JournalListener::Token token = _journalListener->getToken();
// Initialize on first use.
if (!_waitUntilDurableSession) {
invariantWTOK(
_conn->open_session(_conn, NULL, "isolation=snapshot", &_waitUntilDurableSession));
}
// Use the journal when available, or a checkpoint otherwise.
if (_engine && _engine->isDurable()) {
invariantWTOK(_waitUntilDurableSession->log_flush(_waitUntilDurableSession, "sync=on"));
LOG(4) << "flushed journal";
} else {
invariantWTOK(_waitUntilDurableSession->checkpoint(_waitUntilDurableSession, NULL));
LOG(4) << "created checkpoint";
}
_journalListener->onDurable(token);
}