// takes an entry that was written _logTransactionOps
// and applies them to collections
//
// TODO: possibly improve performance of this. We create and destroy a
// context for each operation. Find a way to amortize it out if necessary
//
void applyTransactionFromOplog(BSONObj entry) {
bool transactionAlreadyApplied = entry["a"].Bool();
if (!transactionAlreadyApplied) {
Client::Transaction transaction(DB_SERIALIZABLE);
if (entry.hasElement("ref")) {
applyRefOp(entry);
} else if (entry.hasElement("ops")) {
applyOps(entry["ops"].Array());
} else {
verify(0);
}
// set the applied bool to true, to let the oplog know that
// this entry has been applied to collections
BSONElementManipulator(entry["a"]).setBool(true);
{
LOCK_REASON(lockReason, "repl: setting oplog entry's applied bit");
Lock::DBRead lk1("local", lockReason);
writeEntryToOplog(entry, false);
}
// If this code fails, it is impossible to recover from
// because we don't know if the transaction successfully committed
// so we might as well crash
// There is currently no known way this code can throw an exception
try {
// we are operating as a secondary. We don't have to fsync
transaction.commit(DB_TXN_NOSYNC);
}
catch (std::exception &e) {
log() << "exception during commit of applyTransactionFromOplog, aborting system: " << e.what() << endl;
printStackTrace();
logflush();
::abort();
}
}
}
// find all oplog entries for a given OID in the oplog.refs collection and apply them
// TODO this should be a range query on oplog.refs where _id.oid == oid and applyOps to
// each entry found. The locking of the query interleaved with the locking in the applyOps
// did not work, so it a sequence of point queries.
// TODO verify that the query plan is a indexed lookup.
// TODO verify that the query plan does not fetch too many docs and then only process one of them.
void applyRefOp(BSONObj entry) {
OID oid = entry["ref"].OID();
LOG(3) << "apply ref " << entry << " oid " << oid << endl;
long long seq = 0; // note that 0 is smaller than any of the seq numbers
while (1) {
BSONObj entry;
{
LOCK_REASON(lockReason, "repl: finding oplog.refs entry to apply");
Client::ReadContext ctx(rsOplogRefs, lockReason);
// TODO: Should this be using rsOplogRefsDetails, verifying non-null?
Collection *cl = getCollection(rsOplogRefs);
if (cl == NULL || !cl->findOne(BSON("_id" << BSON("$gt" << BSON("oid" << oid << "seq" << seq))), entry, true)) {
break;
}
}
BSONElement e = entry.getFieldDotted("_id.seq");
seq = e.Long();
BSONElement eOID = entry.getFieldDotted("_id.oid");
if (oid != eOID.OID()) {
break;
}
LOG(3) << "apply " << entry << " seq=" << seq << endl;
applyOps(entry["ops"].Array());
}
}
Status applyCommitTransaction(OperationContext* opCtx,
const repl::OplogEntry& entry,
repl::OplogApplication::Mode mode) {
// Return error if run via applyOps command.
uassert(50987,
"commitTransaction is only used internally by secondaries.",
mode != repl::OplogApplication::Mode::kApplyOpsCmd);
IDLParserErrorContext ctx("commitTransaction");
auto commitCommand = CommitTransactionOplogObject::parse(ctx, entry.getObject());
if (mode == repl::OplogApplication::Mode::kRecovering) {
const auto replCoord = repl::ReplicationCoordinator::get(opCtx);
const auto recoveryTimestamp = replCoord->getRecoveryTimestamp();
invariant(recoveryTimestamp);
// If the commitTimestamp is before the recoveryTimestamp, then the data already
// reflects the operations from the transaction.
const auto& commitTimestamp = commitCommand.getCommitTimestamp();
if (recoveryTimestamp.get() > commitTimestamp) {
return Status::OK();
}
// Get the corresponding prepareTransaction oplog entry.
TransactionHistoryIterator iter(entry.getOpTime());
invariant(iter.hasNext());
const auto commitOplogEntry = iter.next(opCtx);
invariant(iter.hasNext());
const auto prepareOplogEntry = iter.next(opCtx);
// Transform prepare command into a normal applyOps command.
const auto prepareCmd = prepareOplogEntry.getOperationToApply().removeField("prepare");
BSONObjBuilder resultWeDontCareAbout;
return applyOps(
opCtx, entry.getNss().db().toString(), prepareCmd, mode, &resultWeDontCareAbout);
}
// TODO: SERVER-36492 Only run on secondary until we support initial sync.
invariant(mode == repl::OplogApplication::Mode::kSecondary);
// Transaction operations are in its own batch, so we can modify their opCtx.
invariant(entry.getSessionId());
invariant(entry.getTxnNumber());
opCtx->setLogicalSessionId(*entry.getSessionId());
opCtx->setTxnNumber(*entry.getTxnNumber());
// The write on transaction table may be applied concurrently, so refreshing state
// from disk may read that write, causing starting a new transaction on an existing
// txnNumber. Thus, we start a new transaction without refreshing state from disk.
MongoDOperationContextSessionWithoutRefresh sessionCheckout(opCtx);
auto transaction = TransactionParticipant::get(opCtx);
invariant(transaction);
transaction->unstashTransactionResources(opCtx, "commitTransaction");
transaction->commitPreparedTransaction(opCtx, commitCommand.getCommitTimestamp());
return Status::OK();
}
// Doles out all the work to the writer pool threads and waits for them to complete
void SyncTail::multiApply( std::deque<BSONObj>& ops, MultiSyncApplyFunc applyFunc ) {
// Use a ThreadPool to prefetch all the operations in a batch.
prefetchOps(ops);
std::vector< std::vector<BSONObj> > writerVectors(theReplSet->replWriterThreadCount);
fillWriterVectors(ops, &writerVectors);
LOG(2) << "replication batch size is " << ops.size() << endl;
// We must grab this because we're going to grab write locks later.
// We hold this mutex the entire time we're writing; it doesn't matter
// because all readers are blocked anyway.
SimpleMutex::scoped_lock fsynclk(filesLockedFsync);
// stop all readers until we're done
Lock::ParallelBatchWriterMode pbwm;
applyOps(writerVectors, applyFunc);
}
// Applies a batch of oplog entries, by using a set of threads to apply the operations and then
// writes the oplog entries to the local oplog.
OpTime SyncTail::multiApply(OperationContext* txn, const OpQueue& ops) {
invariant(_applyFunc);
if (getGlobalServiceContext()->getGlobalStorageEngine()->isMmapV1()) {
// Use a ThreadPool to prefetch all the operations in a batch.
prefetchOps(ops.getDeque(), &_prefetcherPool);
}
std::vector<std::vector<BSONObj>> writerVectors(replWriterThreadCount);
fillWriterVectors(txn, ops.getDeque(), &writerVectors);
LOG(2) << "replication batch size is " << ops.getDeque().size() << endl;
// We must grab this because we're going to grab write locks later.
// We hold this mutex the entire time we're writing; it doesn't matter
// because all readers are blocked anyway.
stdx::lock_guard<SimpleMutex> fsynclk(filesLockedFsync);
// stop all readers until we're done
Lock::ParallelBatchWriterMode pbwm(txn->lockState());
ReplicationCoordinator* replCoord = getGlobalReplicationCoordinator();
if (replCoord->getMemberState().primary() && !replCoord->isWaitingForApplierToDrain()) {
severe() << "attempting to replicate ops while primary";
fassertFailed(28527);
}
applyOps(writerVectors, &_writerPool, _applyFunc, this);
OpTime lastOpTime;
{
ON_BLOCK_EXIT([&] { _writerPool.join(); });
std::vector<BSONObj> raws;
raws.reserve(ops.getDeque().size());
for (auto&& op : ops.getDeque()) {
raws.emplace_back(op.raw);
}
lastOpTime = writeOpsToOplog(txn, raws);
if (inShutdown()) {
return OpTime();
}
}
// We have now written all database writes and updated the oplog to match.
return lastOpTime;
}
// find all oplog entries for a given OID in the oplog.refs collection and apply them
// TODO this should be a range query on oplog.refs where _id.oid == oid and applyOps to
// each entry found. The locking of the query interleaved with the locking in the applyOps
// did not work, so it a sequence of point queries.
// TODO verify that the query plan is a indexed lookup.
// TODO verify that the query plan does not fetch too many docs and then only process one of them.
void applyRefOp(BSONObj entry) {
OID oid = entry["ref"].OID();
LOG(3) << "apply ref " << entry << " oid " << oid << endl;
long long seq = 0; // note that 0 is smaller than any of the seq numbers
while (1) {
BSONObj entry;
{
Client::ReadContext ctx(rsOplogRefs);
// TODO: Should this be using rsOplogRefsDetails, verifying non-null?
NamespaceDetails *d = nsdetails(rsOplogRefs);
if (d == NULL || !d->findOne(BSON("_id" << BSON("$gt" << BSON("oid" << oid << "seq" << seq))), entry, true)) {
break;
}
}
BSONElement e = entry.getFieldDotted("_id.seq");
seq = e.Long();
BSONElement eOID = entry.getFieldDotted("_id.oid");
if (oid != eOID.OID()) {
break;
}
LOG(3) << "apply " << entry << " seq=" << seq << endl;
applyOps(entry["ops"].Array());
}
}
