ClientCursor::ClientCursor(ClientCursorParams params,
CursorManager* cursorManager,
CursorId cursorId,
OperationContext* operationUsingCursor,
Date_t now)
: _cursorid(cursorId),
_nss(std::move(params.nss)),
_authenticatedUsers(std::move(params.authenticatedUsers)),
_lsid(operationUsingCursor->getLogicalSessionId()),
_txnNumber(operationUsingCursor->getTxnNumber()),
_readConcernLevel(params.readConcernLevel),
_cursorManager(cursorManager),
_originatingCommand(params.originatingCommandObj),
_queryOptions(params.queryOptions),
_exec(std::move(params.exec)),
_operationUsingCursor(operationUsingCursor),
_lastUseDate(now) {
invariant(_cursorManager);
invariant(_exec);
invariant(_operationUsingCursor);
cursorStatsOpen.increment();
if (isNoTimeout()) {
// cursors normally timeout after an inactivity period to prevent excess memory use
// setting this prevents timeout of the cursor in question.
cursorStatsOpenNoTimeout.increment();
}
}
Status waitForWriteConcern(OperationContext* txn,
const OpTime& replOpTime,
const WriteConcernOptions& writeConcern,
WriteConcernResult* result) {
// We assume all options have been validated earlier, if not, programming error
dassert(validateWriteConcern(writeConcern).isOK());
// Next handle blocking on disk
Timer syncTimer;
switch (writeConcern.syncMode) {
case WriteConcernOptions::NONE:
break;
case WriteConcernOptions::FSYNC: {
StorageEngine* storageEngine = getGlobalServiceContext()->getGlobalStorageEngine();
if (!storageEngine->isDurable()) {
result->fsyncFiles = storageEngine->flushAllFiles(true);
} else {
// We only need to commit the journal if we're durable
txn->recoveryUnit()->waitUntilDurable();
}
break;
}
case WriteConcernOptions::JOURNAL:
txn->recoveryUnit()->waitUntilDurable();
break;
}
result->syncMillis = syncTimer.millis();
// Now wait for replication
if (replOpTime.isNull()) {
// no write happened for this client yet
return Status::OK();
}
// needed to avoid incrementing gleWtimeStats SERVER-9005
if (writeConcern.wNumNodes <= 1 && writeConcern.wMode.empty()) {
// no desired replication check
return Status::OK();
}
// Now we wait for replication
// Note that replica set stepdowns and gle mode changes are thrown as errors
repl::ReplicationCoordinator::StatusAndDuration replStatus =
repl::getGlobalReplicationCoordinator()->awaitReplication(txn, replOpTime, writeConcern);
if (replStatus.status == ErrorCodes::WriteConcernFailed) {
gleWtimeouts.increment();
result->err = "timeout";
result->wTimedOut = true;
}
// Add stats
result->writtenTo = repl::getGlobalReplicationCoordinator()->getHostsWrittenTo(replOpTime);
gleWtimeStats.recordMillis(durationCount<Milliseconds>(replStatus.duration));
result->wTime = durationCount<Milliseconds>(replStatus.duration);
return replStatus.status;
}
void CursorCache::storeRef(const std::string& server, long long id, const std::string& ns) {
LOG(_myLogLevel) << "CursorCache::storeRef server: " << server << " id: " << id << endl;
verify(id);
stdx::lock_guard<stdx::mutex> lk(_mutex);
_refs[id] = server;
_refsNS[id] = ns;
cursorStatsSingleTarget.increment();
}
Status Collection::recordStoreGoingToMove(OperationContext* txn,
const RecordId& oldLocation,
const char* oldBuffer,
size_t oldSize) {
moveCounter.increment();
_cursorManager.invalidateDocument(txn, oldLocation, INVALIDATION_DELETION);
_indexCatalog.unindexRecord(txn, BSONObj(oldBuffer), oldLocation, true);
return Status::OK();
}
OplogReader::OplogReader() {
_tailingQueryOptions = QueryOption_SlaveOk;
_tailingQueryOptions |= QueryOption_CursorTailable | QueryOption_OplogReplay;
/* TODO: slaveOk maybe shouldn't use? */
_tailingQueryOptions |= QueryOption_AwaitData;
readersCreatedStats.increment();
}
ClientCursorPin::ClientCursorPin(OperationContext* opCtx, ClientCursor* cursor)
: _opCtx(opCtx), _cursor(cursor) {
invariant(_cursor);
invariant(_cursor->_operationUsingCursor);
invariant(_cursor->_cursorManager);
invariant(!_cursor->_disposed);
// We keep track of the number of cursors currently pinned. The cursor can become unpinned
// either by being released back to the cursor manager or by being deleted. A cursor may be
// transferred to another pin object via move construction or move assignment, but in this case
// it is still considered pinned.
cursorStatsOpenPinned.increment();
}
void OpDebug::recordStats() {
if ( nreturned > 0 )
returnedCounter.increment( nreturned );
if ( ninserted > 0 )
insertedCounter.increment( ninserted );
if ( nMatched > 0 )
updatedCounter.increment( nMatched );
if ( ndeleted > 0 )
deletedCounter.increment( ndeleted );
if ( nscanned > 0 )
scannedCounter.increment( nscanned );
if ( nscannedObjects > 0 )
scannedObjectCounter.increment( nscannedObjects );
if ( idhack )
idhackCounter.increment();
if ( scanAndOrder )
scanAndOrderCounter.increment();
if ( fastmod )
fastmodCounter.increment();
if ( writeConflicts )
writeConflictsCounter.increment( writeConflicts );
}
void OpDebug::recordStats() {
if ( nreturned > 0 )
returnedCounter.increment( nreturned );
if ( ninserted > 0 )
insertedCounter.increment( ninserted );
if ( nupdated > 0 )
updatedCounter.increment( nupdated );
if ( ndeleted > 0 )
deletedCounter.increment( ndeleted );
if ( nscanned > 0 )
scannedCounter.increment( nscanned );
if ( idhack )
idhackCounter.increment();
if ( scanAndOrder )
scanAndOrderCounter.increment();
if ( fastmod )
fastmodCounter.increment();
}
ShardedClientCursor::ShardedClientCursor(QueryMessage& q, ParallelSortClusteredCursor* cursor) {
verify(cursor);
_cursor = cursor;
_skip = q.ntoskip;
_ntoreturn = q.ntoreturn;
_totalSent = 0;
_done = false;
_id = 0;
if (q.queryOptions & QueryOption_NoCursorTimeout) {
_lastAccessMillis = 0;
} else
_lastAccessMillis = Listener::getElapsedTimeMillis();
cursorStatsMultiTarget.increment();
}
virtual void run() {
Client::initThread( name().c_str() );
while ( ! inShutdown() ) {
sleepsecs( 60 );
LOG(3) << "TTLMonitor thread awake" << endl;
if ( lockedForWriting() ) {
// note: this is not perfect as you can go into fsync+lock between
// this and actually doing the delete later
LOG(3) << " locked for writing" << endl;
continue;
}
// if part of replSet but not in a readable state (e.g. during initial sync), skip.
if ( theReplSet && !theReplSet->state().readable() )
continue;
set<string> dbs;
{
Lock::DBRead lk( "local" );
dbHolder().getAllShortNames( dbs );
}
ttlPasses.increment();
for ( set<string>::const_iterator i=dbs.begin(); i!=dbs.end(); ++i ) {
string db = *i;
try {
doTTLForDB( db );
}
catch ( DBException& e ) {
error() << "error processing ttl for db: " << db << " " << e << endl;
}
}
}
}
/* apply the log op that is in param o
@return bool success (true) or failure (false)
*/
bool SyncTail::syncApply(
OperationContext* txn, const BSONObj &op, bool convertUpdateToUpsert) {
const char *ns = op.getStringField("ns");
verify(ns);
if ( (*ns == '\0') || (*ns == '.') ) {
// this is ugly
// this is often a no-op
// but can't be 100% sure
if( *op.getStringField("op") != 'n' ) {
error() << "replSet skipping bad op in oplog: " << op.toString() << rsLog;
}
return true;
}
bool isCommand(op["op"].valuestrsafe()[0] == 'c');
boost::scoped_ptr<Lock::ScopedLock> lk;
if(isCommand) {
// a command may need a global write lock. so we will conservatively go
// ahead and grab one here. suboptimal. :-(
lk.reset(new Lock::GlobalWrite(txn->lockState()));
} else {
// DB level lock for this operation
lk.reset(new Lock::DBWrite(txn->lockState(), ns));
}
Client::Context ctx(ns);
ctx.getClient()->curop()->reset();
// For non-initial-sync, we convert updates to upserts
// to suppress errors when replaying oplog entries.
bool ok = !applyOperation_inlock(txn, ctx.db(), op, true, convertUpdateToUpsert);
opsAppliedStats.increment();
txn->recoveryUnit()->commitIfNeeded();
return ok;
}
void doTTLForDB( const string& dbName ) {
//check isMaster before becoming god
bool isMaster = isMasterNs( dbName.c_str() );
Client::GodScope god;
vector<BSONObj> indexes;
{
auto_ptr<DBClientCursor> cursor =
db.query( dbName + ".system.indexes" ,
BSON( secondsExpireField << BSON( "$exists" << true ) ) ,
0 , /* default nToReturn */
0 , /* default nToSkip */
0 , /* default fieldsToReturn */
QueryOption_SlaveOk ); /* perform on secondaries too */
if ( cursor.get() ) {
while ( cursor->more() ) {
indexes.push_back( cursor->next().getOwned() );
}
}
}
for ( unsigned i=0; i<indexes.size(); i++ ) {
BSONObj idx = indexes[i];
BSONObj key = idx["key"].Obj();
if ( key.nFields() != 1 ) {
error() << "key for ttl index can only have 1 field" << endl;
continue;
}
BSONObj query;
{
BSONObjBuilder b;
b.appendDate( "$lt" , curTimeMillis64() - ( 1000 * idx[secondsExpireField].numberLong() ) );
query = BSON( key.firstElement().fieldName() << b.obj() );
}
LOG(1) << "TTL: " << key << " \t " << query << endl;
long long n = 0;
{
string ns = idx["ns"].String();
Client::WriteContext ctx( ns );
NamespaceDetails* nsd = nsdetails( ns );
if ( ! nsd ) {
// collection was dropped
continue;
}
if ( nsd->setUserFlag( NamespaceDetails::Flag_UsePowerOf2Sizes ) ) {
nsd->syncUserFlags( ns );
}
// only do deletes if on master
if ( ! isMaster ) {
continue;
}
n = deleteObjects( ns.c_str() , query , false , true );
ttlDeletedDocuments.increment( n );
}
LOG(1) << "\tTTL deleted: " << n << endl;
}
}
Status waitForWriteConcern(OperationContext* txn,
const OpTime& replOpTime,
const WriteConcernOptions& writeConcern,
WriteConcernResult* result) {
LOG(2) << "Waiting for write concern. OpTime: " << replOpTime
<< ", write concern: " << writeConcern.toBSON();
auto replCoord = repl::ReplicationCoordinator::get(txn);
MONGO_FAIL_POINT_PAUSE_WHILE_SET(hangBeforeWaitingForWriteConcern);
// Next handle blocking on disk
Timer syncTimer;
WriteConcernOptions writeConcernWithPopulatedSyncMode =
replCoord->populateUnsetWriteConcernOptionsSyncMode(writeConcern);
switch (writeConcernWithPopulatedSyncMode.syncMode) {
case WriteConcernOptions::SyncMode::UNSET:
severe() << "Attempting to wait on a WriteConcern with an unset sync option";
fassertFailed(34410);
case WriteConcernOptions::SyncMode::NONE:
break;
case WriteConcernOptions::SyncMode::FSYNC: {
StorageEngine* storageEngine = getGlobalServiceContext()->getGlobalStorageEngine();
if (!storageEngine->isDurable()) {
result->fsyncFiles = storageEngine->flushAllFiles(true);
} else {
// We only need to commit the journal if we're durable
txn->recoveryUnit()->waitUntilDurable();
}
break;
}
case WriteConcernOptions::SyncMode::JOURNAL:
if (replCoord->getReplicationMode() != repl::ReplicationCoordinator::Mode::modeNone) {
// Wait for ops to become durable then update replication system's
// knowledge of this.
OpTime appliedOpTime = replCoord->getMyLastAppliedOpTime();
txn->recoveryUnit()->waitUntilDurable();
replCoord->setMyLastDurableOpTimeForward(appliedOpTime);
} else {
txn->recoveryUnit()->waitUntilDurable();
}
break;
}
result->syncMillis = syncTimer.millis();
// Now wait for replication
if (replOpTime.isNull()) {
// no write happened for this client yet
return Status::OK();
}
// needed to avoid incrementing gleWtimeStats SERVER-9005
if (writeConcernWithPopulatedSyncMode.wNumNodes <= 1 &&
writeConcernWithPopulatedSyncMode.wMode.empty()) {
// no desired replication check
return Status::OK();
}
// Replica set stepdowns and gle mode changes are thrown as errors
repl::ReplicationCoordinator::StatusAndDuration replStatus =
replCoord->awaitReplication(txn, replOpTime, writeConcernWithPopulatedSyncMode);
if (replStatus.status == ErrorCodes::WriteConcernFailed) {
gleWtimeouts.increment();
result->err = "timeout";
result->wTimedOut = true;
}
// Add stats
result->writtenTo = repl::getGlobalReplicationCoordinator()->getHostsWrittenTo(
replOpTime,
writeConcernWithPopulatedSyncMode.syncMode == WriteConcernOptions::SyncMode::JOURNAL);
gleWtimeStats.recordMillis(durationCount<Milliseconds>(replStatus.duration));
result->wTime = durationCount<Milliseconds>(replStatus.duration);
return replStatus.status;
}
// returns number of seconds to sleep, if any
uint32_t BackgroundSync::produce() {
// normally msgCheckNewState gets called periodically, but in a single node repl set
// there are no heartbeat threads, so we do it here to be sure. this is relevant if the
// singleton member has done a stepDown() and needs to come back up.
if (theReplSet->config().members.size() == 1 &&
theReplSet->myConfig().potentiallyHot()) {
Manager* mgr = theReplSet->mgr;
// When would mgr be null? During replsettest'ing, in which case we should
// fall through and actually apply ops as if we were a real secondary.
if (mgr) {
mgr->send(boost::bind(&Manager::msgCheckNewState, theReplSet->mgr));
// There should never be ops to sync in a 1-member set, anyway
return 1;
}
}
OplogReader r(true /* doHandshake */);
// find a target to sync from the last op time written
getOplogReader(r);
// no server found
GTID lastGTIDFetched = theReplSet->gtidManager->getLiveState();
{
boost::unique_lock<boost::mutex> lock(_mutex);
if (_currentSyncTarget == NULL) {
// if there is no one to sync from
return 1; //sleep one second
}
}
r.tailingQueryGTE(rsoplog, lastGTIDFetched);
// if target cut connections between connecting and querying (for
// example, because it stepped down) we might not have a cursor
if (!r.haveCursor()) {
return 0;
}
try {
// this method may actually run rollback, yes, the name is bad
if (isRollbackRequired(r)) {
// sleep 2 seconds and try again. (The 2 is arbitrary).
// If we are not fatal, then we will keep trying to sync
// from another machine
return 2;
}
}
catch (RollbackOplogException& re){
// we attempted a rollback and failed, we must go fatal.
log() << "Caught a RollbackOplogException during rollback, going fatal" << rsLog;
theReplSet->fatal();
return 2; // 2 is arbitrary, if we are going fatal, we are done
}
while (!_opSyncShouldExit) {
while (!_opSyncShouldExit) {
{
// check if we should bail out
boost::unique_lock<boost::mutex> lck(_mutex);
if (!_opSyncShouldRun) {
return 0;
}
}
if (!r.moreInCurrentBatch()) {
// check to see if we have a request to sync
// from a specific target. If so, get out so that
// we can restart the act of syncing and
// do so from the correct target
if (theReplSet->gotForceSync()) {
return 0;
}
verify(!theReplSet->isPrimary());
if (shouldChangeSyncTarget()) {
return 0;
}
//record time for each getmore
{
TimerHolder batchTimer(&getmoreReplStats);
r.more();
}
//increment
networkByteStats.increment(r.currentBatchMessageSize());
}
if (!r.more()) {
break;
}
// This is the operation we have received from the target
// that we must put in our oplog with an applied field of false
BSONObj o = r.nextSafe().getOwned();
opsReadStats.increment();
LOG(3) << "replicating " << o.toString(false, true) << " from " << _currentSyncTarget->fullName() << endl;
uint64_t ts = o["ts"]._numberLong();
// now that we have the element in o, let's check
// if there a delay is required (via slaveDelay) before
// writing it to the oplog
if (theReplSet->myConfig().slaveDelay > 0) {
handleSlaveDelay(ts);
{
boost::unique_lock<boost::mutex> lck(_mutex);
if (!_opSyncShouldRun) {
break;
}
}
}
{
Timer timer;
bool bigTxn = false;
{
Client::Transaction transaction(DB_SERIALIZABLE);
replicateFullTransactionToOplog(o, r, &bigTxn);
// we are operating as a secondary. We don't have to fsync
transaction.commit(DB_TXN_NOSYNC);
}
{
GTID currEntry = getGTIDFromOplogEntry(o);
uint64_t lastHash = o["h"].numberLong();
boost::unique_lock<boost::mutex> lock(_mutex);
// update counters
theReplSet->gtidManager->noteGTIDAdded(currEntry, ts, lastHash);
// notify applier thread that data exists
if (_deque.size() == 0) {
_queueCond.notify_all();
}
_deque.push_back(o);
bufferCountGauge.increment();
bufferSizeGauge.increment(o.objsize());
// this is a flow control mechanism, with bad numbers
// hard coded for now just to get something going.
// If the opSync thread notices that we have over 20000
// transactions in the queue, it waits until we get below
// 10000. This is where we wait if we get too high
// Once we have spilling of transactions working, this
// logic will need to be redone
if (_deque.size() > 20000) {
_queueCond.wait(lock);
}
if (bigTxn) {
// if we have a large transaction, we don't want
// to let it pile up. We want to process it immedietely
// before processing anything else.
while (_deque.size() > 0) {
_queueDone.wait(lock);
}
}
}
}
} // end while
if (shouldChangeSyncTarget()) {
return 0;
}
r.tailCheck();
if( !r.haveCursor() ) {
LOG(1) << "replSet end opSync pass" << rsLog;
return 0;
}
// looping back is ok because this is a tailable cursor
}
return 0;
}
void BackgroundSync::applyOpsFromOplog() {
GTID lastLiveGTID;
GTID lastUnappliedGTID;
while (1) {
try {
BSONObj curr;
{
boost::unique_lock<boost::mutex> lck(_mutex);
// wait until we know an item has been produced
while (_deque.size() == 0 && !_applierShouldExit) {
_queueDone.notify_all();
_queueCond.wait(lck);
}
if (_deque.size() == 0 && _applierShouldExit) {
return;
}
curr = _deque.front();
}
GTID currEntry = getGTIDFromOplogEntry(curr);
theReplSet->gtidManager->noteApplyingGTID(currEntry);
// we must do applyTransactionFromOplog in a loop
// because once we have called noteApplyingGTID, we must
// continue until we are successful in applying the transaction.
for (uint32_t numTries = 0; numTries <= 100; numTries++) {
try {
numTries++;
TimerHolder timer(&applyBatchStats);
applyTransactionFromOplog(curr);
opsAppliedStats.increment();
break;
}
catch (std::exception &e) {
log() << "exception during applying transaction from oplog: " << e.what() << endl;
log() << "oplog entry: " << curr.str() << endl;
if (numTries == 100) {
// something is really wrong if we fail 100 times, let's abort
::abort();
}
sleepsecs(1);
}
}
LOG(3) << "applied " << curr.toString(false, true) << endl;
theReplSet->gtidManager->noteGTIDApplied(currEntry);
{
boost::unique_lock<boost::mutex> lck(_mutex);
dassert(_deque.size() > 0);
_deque.pop_front();
bufferCountGauge.increment(-1);
bufferSizeGauge.increment(-curr.objsize());
// this is a flow control mechanism, with bad numbers
// hard coded for now just to get something going.
// If the opSync thread notices that we have over 20000
// transactions in the queue, it waits until we get below
// 10000. This is where we signal that we have gotten there
// Once we have spilling of transactions working, this
// logic will need to be redone
if (_deque.size() == 10000) {
_queueCond.notify_all();
}
}
}
catch (DBException& e) {
sethbmsg(str::stream() << "db exception in producer on applier thread: " << e.toString());
sleepsecs(2);
}
catch (std::exception& e2) {
sethbmsg(str::stream() << "exception in producer on applier thread: " << e2.what());
sleepsecs(2);
}
}
}
StatusWith<DiskLoc> Collection::updateDocument( const DiskLoc& oldLocation,
const BSONObj& objNew,
bool enforceQuota,
OpDebug* debug ) {
Record* oldRecord = getExtentManager()->recordFor( oldLocation );
BSONObj objOld = BSONObj::make( oldRecord );
if ( objOld.hasElement( "_id" ) ) {
BSONElement oldId = objOld["_id"];
BSONElement newId = objNew["_id"];
if ( oldId != newId )
return StatusWith<DiskLoc>( ErrorCodes::InternalError,
"in Collection::updateDocument _id mismatch",
13596 );
}
if ( ns().coll() == "system.users" ) {
// XXX - andy and spencer think this should go away now
V2UserDocumentParser parser;
Status s = parser.checkValidUserDocument(objNew);
if ( !s.isOK() )
return StatusWith<DiskLoc>( s );
}
/* duplicate key check. we descend the btree twice - once for this check, and once for the actual inserts, further
below. that is suboptimal, but it's pretty complicated to do it the other way without rollbacks...
*/
OwnedPointerVector<UpdateTicket> updateTickets;
updateTickets.mutableVector().resize(_indexCatalog.numIndexesTotal());
for (int i = 0; i < _indexCatalog.numIndexesTotal(); ++i) {
IndexDescriptor* descriptor = _indexCatalog.getDescriptor( i );
IndexAccessMethod* iam = _indexCatalog.getIndex( descriptor );
InsertDeleteOptions options;
options.logIfError = false;
options.dupsAllowed =
!(KeyPattern::isIdKeyPattern(descriptor->keyPattern()) || descriptor->unique())
|| ignoreUniqueIndex(descriptor);
updateTickets.mutableVector()[i] = new UpdateTicket();
Status ret = iam->validateUpdate(objOld, objNew, oldLocation, options,
updateTickets.mutableVector()[i]);
if ( !ret.isOK() ) {
return StatusWith<DiskLoc>( ret );
}
}
if ( oldRecord->netLength() < objNew.objsize() ) {
// doesn't fit, have to move to new location
if ( _details->isCapped() )
return StatusWith<DiskLoc>( ErrorCodes::InternalError,
"failing update: objects in a capped ns cannot grow",
10003 );
moveCounter.increment();
_details->paddingTooSmall();
// unindex old record, don't delete
// this way, if inserting new doc fails, we can re-index this one
ClientCursor::aboutToDelete(_ns.ns(), _details, oldLocation);
_indexCatalog.unindexRecord( objOld, oldLocation, true );
if ( debug ) {
if (debug->nmoved == -1) // default of -1 rather than 0
debug->nmoved = 1;
else
debug->nmoved += 1;
}
StatusWith<DiskLoc> loc = insertDocument( objNew, enforceQuota );
if ( loc.isOK() ) {
// insert successful, now lets deallocate the old location
// remember its already unindexed
_recordStore.deallocRecord( oldLocation, oldRecord );
}
else {
// new doc insert failed, so lets re-index the old document and location
_indexCatalog.indexRecord( objOld, oldLocation );
}
return loc;
}
_infoCache.notifyOfWriteOp();
_details->paddingFits();
if ( debug )
debug->keyUpdates = 0;
for (int i = 0; i < _indexCatalog.numIndexesTotal(); ++i) {
IndexDescriptor* descriptor = _indexCatalog.getDescriptor( i );
IndexAccessMethod* iam = _indexCatalog.getIndex( descriptor );
int64_t updatedKeys;
Status ret = iam->update(*updateTickets.vector()[i], &updatedKeys);
if ( !ret.isOK() )
return StatusWith<DiskLoc>( ret );
if ( debug )
debug->keyUpdates += updatedKeys;
}
// update in place
int sz = objNew.objsize();
memcpy(getDur().writingPtr(oldRecord->data(), sz), objNew.objdata(), sz);
return StatusWith<DiskLoc>( oldLocation );
}
Status waitForWriteConcern( OperationContext* txn,
const WriteConcernOptions& writeConcern,
const OpTime& replOpTime,
WriteConcernResult* result ) {
// We assume all options have been validated earlier, if not, programming error
dassert( validateWriteConcern( writeConcern ).isOK() );
// Next handle blocking on disk
Timer syncTimer;
switch( writeConcern.syncMode ) {
case WriteConcernOptions::NONE:
break;
case WriteConcernOptions::FSYNC:
if ( !getDur().isDurable() ) {
result->fsyncFiles = MemoryMappedFile::flushAll( true );
}
else {
// We only need to commit the journal if we're durable
txn->recoveryUnit()->awaitCommit();
}
break;
case WriteConcernOptions::JOURNAL:
txn->recoveryUnit()->awaitCommit();
break;
}
result->syncMillis = syncTimer.millis();
// Now wait for replication
if ( replOpTime.isNull() ) {
// no write happened for this client yet
return Status::OK();
}
if ( writeConcern.wNumNodes <= 1 && writeConcern.wMode.empty() ) {
// no desired replication check
return Status::OK();
}
if (!replset::anyReplEnabled() || serverGlobalParams.configsvr) {
// no replication check needed (validated above)
return Status::OK();
}
const bool isMasterSlaveNode = replset::anyReplEnabled() && !replset::theReplSet;
if ( writeConcern.wMode == "majority" && isMasterSlaveNode ) {
// with master/slave, majority is equivalent to w=1
return Status::OK();
}
// We're sure that replication is enabled and that we have more than one node or a wMode
TimerHolder gleTimerHolder( &gleWtimeStats );
// Now we wait for replication
// Note that replica set stepdowns and gle mode changes are thrown as errors
// TODO: Make this cleaner
Status replStatus = Status::OK();
try {
while ( 1 ) {
if ( writeConcern.wNumNodes > 0 ) {
if (replset::opReplicatedEnough(replOpTime, writeConcern.wNumNodes)) {
break;
}
}
else if (replset::opReplicatedEnough(replOpTime, writeConcern.wMode)) {
break;
}
if ( writeConcern.wTimeout > 0 &&
gleTimerHolder.millis() >= writeConcern.wTimeout ) {
gleWtimeouts.increment();
result->err = "timeout";
result->wTimedOut = true;
replStatus = Status( ErrorCodes::WriteConcernFailed,
"waiting for replication timed out" );
break;
}
sleepmillis(1);
txn->checkForInterrupt();
}
}
catch( const AssertionException& ex ) {
// Our replication state changed while enforcing write concern
replStatus = ex.toStatus();
}
// Add stats
result->writtenTo = replset::getHostsWrittenTo(replOpTime);
result->wTime = gleTimerHolder.recordMillis();
return replStatus;
}
DiskLoc SimpleRecordStoreV1::_allocFromExistingExtents( OperationContext* txn,
int lenToAlloc ) {
// align size up to a multiple of 4
lenToAlloc = (lenToAlloc + (4-1)) & ~(4-1);
freelistAllocs.increment();
DiskLoc loc;
{
DiskLoc *prev = 0;
DiskLoc *bestprev = 0;
DiskLoc bestmatch;
int bestmatchlen = INT_MAX; // sentinel meaning we haven't found a record big enough
int b = bucket(lenToAlloc);
DiskLoc cur = _details->deletedListEntry(b);
int extra = 5; // look for a better fit, a little.
int chain = 0;
while ( 1 ) {
{ // defensive check
int fileNumber = cur.a();
int fileOffset = cur.getOfs();
if (fileNumber < -1 || fileNumber >= 100000 || fileOffset < 0) {
StringBuilder sb;
sb << "Deleted record list corrupted in collection " << _ns
<< ", bucket " << b
<< ", link number " << chain
<< ", invalid link is " << cur.toString()
<< ", throwing Fatal Assertion";
log() << sb.str() << endl;
fassertFailed(16469);
}
}
if ( cur.isNull() ) {
// move to next bucket. if we were doing "extra", just break
if ( bestmatchlen < INT_MAX )
break;
if ( chain > 0 ) {
// if we looked at things in the right bucket, but they were not suitable
freelistBucketExhausted.increment();
}
b++;
if ( b > MaxBucket ) {
// out of space. alloc a new extent.
freelistIterations.increment( 1 + chain );
return DiskLoc();
}
cur = _details->deletedListEntry(b);
prev = 0;
continue;
}
DeletedRecord *r = drec(cur);
if ( r->lengthWithHeaders() >= lenToAlloc &&
r->lengthWithHeaders() < bestmatchlen ) {
bestmatchlen = r->lengthWithHeaders();
bestmatch = cur;
bestprev = prev;
if (r->lengthWithHeaders() == lenToAlloc)
// exact match, stop searching
break;
}
if ( bestmatchlen < INT_MAX && --extra <= 0 )
break;
if ( ++chain > 30 && b <= MaxBucket ) {
// too slow, force move to next bucket to grab a big chunk
//b++;
freelistIterations.increment( chain );
chain = 0;
cur.Null();
}
else {
cur = r->nextDeleted();
prev = &r->nextDeleted();
}
}
// unlink ourself from the deleted list
DeletedRecord *bmr = drec(bestmatch);
if ( bestprev ) {
*txn->recoveryUnit()->writing(bestprev) = bmr->nextDeleted();
}
else {
// should be the front of a free-list
int myBucket = bucket(bmr->lengthWithHeaders());
invariant( _details->deletedListEntry(myBucket) == bestmatch );
_details->setDeletedListEntry(txn, myBucket, bmr->nextDeleted());
}
*txn->recoveryUnit()->writing(&bmr->nextDeleted()) = DiskLoc().setInvalid(); // defensive.
invariant(bmr->extentOfs() < bestmatch.getOfs());
freelistIterations.increment( 1 + chain );
loc = bestmatch;
}
if ( loc.isNull() )
return loc;
// determine if we should chop up
DeletedRecord *r = drec(loc);
/* note we want to grab from the front so our next pointers on disk tend
to go in a forward direction which is important for performance. */
int regionlen = r->lengthWithHeaders();
invariant( r->extentOfs() < loc.getOfs() );
int left = regionlen - lenToAlloc;
if ( left < 24 || left < (lenToAlloc / 8) ) {
// you get the whole thing.
return loc;
}
// don't quantize:
// - $ collections (indexes) as we already have those aligned the way we want SERVER-8425
if ( _normalCollection ) {
// we quantize here so that it only impacts newly sized records
// this prevents oddities with older records and space re-use SERVER-8435
lenToAlloc = std::min( r->lengthWithHeaders(),
quantizeAllocationSpace( lenToAlloc ) );
left = regionlen - lenToAlloc;
if ( left < 24 ) {
// you get the whole thing.
return loc;
}
}
/* split off some for further use. */
txn->recoveryUnit()->writingInt(r->lengthWithHeaders()) = lenToAlloc;
DiskLoc newDelLoc = loc;
newDelLoc.inc(lenToAlloc);
DeletedRecord* newDel = drec(newDelLoc);
DeletedRecord* newDelW = txn->recoveryUnit()->writing(newDel);
newDelW->extentOfs() = r->extentOfs();
newDelW->lengthWithHeaders() = left;
newDelW->nextDeleted().Null();
addDeletedRec( txn, newDelLoc );
return loc;
}
/** Note: if the object shrinks a lot, we don't free up space, we leave extra at end of the record.
*/
const DiskLoc DataFileMgr::updateRecord(
const char *ns,
Collection* collection,
Record *toupdate, const DiskLoc& dl,
const char *_buf, int _len, OpDebug& debug, bool god) {
dassert( toupdate == dl.rec() );
BSONObj objOld = BSONObj::make(toupdate);
BSONObj objNew(_buf);
DEV verify( objNew.objsize() == _len );
DEV verify( objNew.objdata() == _buf );
if( !objNew.hasElement("_id") && objOld.hasElement("_id") ) {
/* add back the old _id value if the update removes it. Note this implementation is slow
(copies entire object multiple times), but this shouldn't happen often, so going for simple
code, not speed.
*/
BSONObjBuilder b;
BSONElement e;
verify( objOld.getObjectID(e) );
b.append(e); // put _id first, for best performance
b.appendElements(objNew);
objNew = b.obj();
}
NamespaceString nsstring(ns);
if (nsstring.coll() == "system.users") {
V2UserDocumentParser parser;
uassertStatusOK(parser.checkValidUserDocument(objNew));
}
uassert( 13596 , str::stream() << "cannot change _id of a document old:" << objOld << " new:" << objNew,
objNew["_id"] == objOld["_id"]);
/* duplicate key check. we descend the btree twice - once for this check, and once for the actual inserts, further
below. that is suboptimal, but it's pretty complicated to do it the other way without rollbacks...
*/
OwnedPointerVector<UpdateTicket> updateTickets;
updateTickets.mutableVector().resize(collection->details()->getTotalIndexCount());
for (int i = 0; i < collection->details()->getTotalIndexCount(); ++i) {
auto_ptr<IndexDescriptor> descriptor(CatalogHack::getDescriptor(collection->details(), i));
auto_ptr<IndexAccessMethod> iam(CatalogHack::getIndex(descriptor.get()));
InsertDeleteOptions options;
options.logIfError = false;
options.dupsAllowed = !(KeyPattern::isIdKeyPattern(descriptor->keyPattern())
|| descriptor->unique())
|| ignoreUniqueIndex(descriptor->getOnDisk());
updateTickets.mutableVector()[i] = new UpdateTicket();
Status ret = iam->validateUpdate(objOld, objNew, dl, options,
updateTickets.mutableVector()[i]);
if (Status::OK() != ret) {
uasserted(ASSERT_ID_DUPKEY, "Update validation failed: " + ret.toString());
}
}
if ( toupdate->netLength() < objNew.objsize() ) {
// doesn't fit. reallocate -----------------------------------------------------
moveCounter.increment();
uassert( 10003,
"failing update: objects in a capped ns cannot grow",
!(collection && collection->details()->isCapped()));
collection->details()->paddingTooSmall();
deleteRecord(ns, toupdate, dl);
DiskLoc res = insert(ns, objNew.objdata(), objNew.objsize(), false, god);
if (debug.nmoved == -1) // default of -1 rather than 0
debug.nmoved = 1;
else
debug.nmoved += 1;
return res;
}
collection->infoCache()->notifyOfWriteOp();
collection->details()->paddingFits();
debug.keyUpdates = 0;
for (int i = 0; i < collection->details()->getTotalIndexCount(); ++i) {
auto_ptr<IndexDescriptor> descriptor(CatalogHack::getDescriptor(collection->details(), i));
auto_ptr<IndexAccessMethod> iam(CatalogHack::getIndex(descriptor.get()));
int64_t updatedKeys;
Status ret = iam->update(*updateTickets.vector()[i], &updatedKeys);
if (Status::OK() != ret) {
// This shouldn't happen unless something disastrous occurred.
massert(16799, "update failed: " + ret.toString(), false);
}
debug.keyUpdates += updatedKeys;
}
// update in place
int sz = objNew.objsize();
memcpy(getDur().writingPtr(toupdate->data(), sz), objNew.objdata(), sz);
return dl;
}
DiskLoc SimpleRecordStoreV1::_allocFromExistingExtents(OperationContext* txn, int lenToAllocRaw) {
// Slowly drain the deletedListLegacyGrabBag by popping one record off and putting it in the
// correct deleted list each time we try to allocate a new record. This ensures we won't
// orphan any data when upgrading from old versions, without needing a long upgrade phase.
// This is done before we try to allocate the new record so we can take advantage of the new
// space immediately.
{
const DiskLoc head = _details->deletedListLegacyGrabBag();
if (!head.isNull()) {
_details->setDeletedListLegacyGrabBag(txn, drec(head)->nextDeleted());
addDeletedRec(txn, head);
}
}
// align size up to a multiple of 4
const int lenToAlloc = (lenToAllocRaw + (4 - 1)) & ~(4 - 1);
freelistAllocs.increment();
DiskLoc loc;
DeletedRecord* dr = NULL;
{
int myBucket;
for (myBucket = bucket(lenToAlloc); myBucket < Buckets; myBucket++) {
// Only look at the first entry in each bucket. This works because we are either
// quantizing or allocating fixed-size blocks.
const DiskLoc head = _details->deletedListEntry(myBucket);
if (head.isNull())
continue;
DeletedRecord* const candidate = drec(head);
if (candidate->lengthWithHeaders() >= lenToAlloc) {
loc = head;
dr = candidate;
break;
}
}
if (!dr)
return DiskLoc(); // no space
// Unlink ourself from the deleted list
_details->setDeletedListEntry(txn, myBucket, dr->nextDeleted());
*txn->recoveryUnit()->writing(&dr->nextDeleted()) = DiskLoc().setInvalid(); // defensive
}
invariant(dr->extentOfs() < loc.getOfs());
// Split the deleted record if it has at least as much left over space as our smallest
// allocation size. Otherwise, just take the whole DeletedRecord.
const int remainingLength = dr->lengthWithHeaders() - lenToAlloc;
if (remainingLength >= bucketSizes[0]) {
txn->recoveryUnit()->writingInt(dr->lengthWithHeaders()) = lenToAlloc;
const DiskLoc newDelLoc = DiskLoc(loc.a(), loc.getOfs() + lenToAlloc);
DeletedRecord* newDel = txn->recoveryUnit()->writing(drec(newDelLoc));
newDel->extentOfs() = dr->extentOfs();
newDel->lengthWithHeaders() = remainingLength;
newDel->nextDeleted().Null();
addDeletedRec(txn, newDelLoc);
}
return loc;
}
Status waitForWriteConcern(OperationContext* txn,
const OpTime& replOpTime,
const WriteConcernOptions& writeConcern,
WriteConcernResult* result) {
// We assume all options have been validated earlier, if not, programming error.
// Passing localDB name is a hack to avoid more rigorous check that performed for non local DB.
dassert(validateWriteConcern(txn, writeConcern, kLocalDB).isOK());
// We should never be waiting for write concern while holding any sort of lock, because this may
// lead to situations where the replication heartbeats are stalled.
//
// This check does not hold for writes done through dbeval because it runs with a global X lock.
dassert(!txn->lockState()->isLocked() || txn->getClient()->isInDirectClient());
// Next handle blocking on disk
Timer syncTimer;
auto replCoord = repl::getGlobalReplicationCoordinator();
WriteConcernOptions writeConcernWithPopulatedSyncMode =
replCoord->populateUnsetWriteConcernOptionsSyncMode(writeConcern);
switch (writeConcernWithPopulatedSyncMode.syncMode) {
case WriteConcernOptions::SyncMode::UNSET:
severe() << "Attempting to wait on a WriteConcern with an unset sync option";
fassertFailed(34410);
case WriteConcernOptions::SyncMode::NONE:
break;
case WriteConcernOptions::SyncMode::FSYNC: {
StorageEngine* storageEngine = getGlobalServiceContext()->getGlobalStorageEngine();
if (!storageEngine->isDurable()) {
result->fsyncFiles = storageEngine->flushAllFiles(true);
} else {
// We only need to commit the journal if we're durable
txn->recoveryUnit()->waitUntilDurable();
}
break;
}
case WriteConcernOptions::SyncMode::JOURNAL:
if (replCoord->getReplicationMode() != repl::ReplicationCoordinator::Mode::modeNone) {
// Wait for ops to become durable then update replication system's
// knowledge of this.
OpTime appliedOpTime = replCoord->getMyLastAppliedOpTime();
txn->recoveryUnit()->waitUntilDurable();
replCoord->setMyLastDurableOpTimeForward(appliedOpTime);
} else {
txn->recoveryUnit()->waitUntilDurable();
}
break;
}
result->syncMillis = syncTimer.millis();
// Now wait for replication
if (replOpTime.isNull()) {
// no write happened for this client yet
return Status::OK();
}
// needed to avoid incrementing gleWtimeStats SERVER-9005
if (writeConcernWithPopulatedSyncMode.wNumNodes <= 1 &&
writeConcernWithPopulatedSyncMode.wMode.empty()) {
// no desired replication check
return Status::OK();
}
// Now we wait for replication
// Note that replica set stepdowns and gle mode changes are thrown as errors
repl::ReplicationCoordinator::StatusAndDuration replStatus =
repl::getGlobalReplicationCoordinator()->awaitReplication(
txn, replOpTime, writeConcernWithPopulatedSyncMode);
if (replStatus.status == ErrorCodes::WriteConcernFailed) {
gleWtimeouts.increment();
result->err = "timeout";
result->wTimedOut = true;
}
// Add stats
result->writtenTo = repl::getGlobalReplicationCoordinator()->getHostsWrittenTo(
replOpTime,
writeConcernWithPopulatedSyncMode.syncMode == WriteConcernOptions::SyncMode::JOURNAL);
gleWtimeStats.recordMillis(durationCount<Milliseconds>(replStatus.duration));
result->wTime = durationCount<Milliseconds>(replStatus.duration);
return replStatus.status;
}