static void runUpdateFromOplogWithLock(const char* ns, BSONObj op, bool isRollback) {
NamespaceDetails* nsd = nsdetails(ns);
const char *names[] = {
KEY_STR_PK,
KEY_STR_OLD_ROW,
KEY_STR_NEW_ROW
};
BSONElement fields[3];
op.getFields(3, names, fields);
BSONObj pk = fields[0].Obj();
BSONObj oldRow = fields[1].Obj();
BSONObj newRow = fields[2].Obj();
// note the only difference between these two cases is
// what is passed as the before image, and what is passed
// as after. In normal replication, we replace oldRow with newRow.
// In rollback, we replace newRow with oldRow
uint64_t flags = (NamespaceDetails::NO_UNIQUE_CHECKS | NamespaceDetails::NO_LOCKTREE);
if (isRollback) {
// if this is a rollback, then the newRow is what is in the
// collections, that we want to replace with oldRow
updateOneObject(nsd, pk, newRow, oldRow, NULL, flags);
}
else {
// normal replication case
updateOneObject(nsd, pk, oldRow, newRow, NULL, flags);
}
}
/** @param fromRepl false if from ApplyOpsCmd
@return true if was and update should have happened and the document DNE. see replset initial sync code.
*/
bool applyOperation_inlock(const BSONObj& op, bool fromRepl, bool convertUpdateToUpsert) {
LOG(6) << "applying op: " << op << endl;
bool failedUpdate = false;
OpCounters * opCounters = fromRepl ? &replOpCounters : &globalOpCounters;
const char *names[] = { "o", "ns", "op", "b" };
BSONElement fields[4];
op.getFields(4, names, fields);
BSONObj o;
if( fields[0].isABSONObj() )
o = fields[0].embeddedObject();
const char *ns = fields[1].valuestrsafe();
Lock::assertWriteLocked(ns);
NamespaceDetails *nsd = nsdetails(ns);
// operation type -- see logOp() comments for types
const char *opType = fields[2].valuestrsafe();
if ( *opType == 'i' ) {
opCounters->gotInsert();
const char *p = strchr(ns, '.');
if ( p && strcmp(p, ".system.indexes") == 0 ) {
// updates aren't allowed for indexes -- so we will do a regular insert. if index already
// exists, that is ok.
theDataFileMgr.insert(ns, (void*) o.objdata(), o.objsize());
}
else {
// do upserts for inserts as we might get replayed more than once
OpDebug debug;
BSONElement _id;
if( !o.getObjectID(_id) ) {
/* No _id. This will be very slow. */
Timer t;
updateObjectsForReplication(ns, o, o, true, false, false, debug, false,
QueryPlanSelectionPolicy::idElseNatural() );
if( t.millis() >= 2 ) {
RARELY OCCASIONALLY log() << "warning, repl doing slow updates (no _id field) for " << ns << endl;
}
}
else {
// probably don't need this since all replicated colls have _id indexes now
// but keep it just in case
RARELY if ( nsd && !nsd->isCapped() ) { ensureHaveIdIndex(ns); }
/* todo : it may be better to do an insert here, and then catch the dup key exception and do update
then. very few upserts will not be inserts...
*/
BSONObjBuilder b;
b.append(_id);
updateObjectsForReplication(ns, o, b.done(), true, false, false , debug, false,
QueryPlanSelectionPolicy::idElseNatural() );
}
}
}
/** @param fromRepl false if from ApplyOpsCmd
@return true if was and update should have happened and the document DNE. see replset initial sync code.
*/
bool applyOperation_inlock(const BSONObj& op , bool fromRepl ) {
assertInWriteLock();
LOG(6) << "applying op: " << op << endl;
bool failedUpdate = false;
OpCounters * opCounters = fromRepl ? &replOpCounters : &globalOpCounters;
const char *names[] = { "o", "ns", "op", "b" };
BSONElement fields[4];
op.getFields(4, names, fields);
BSONObj o;
if( fields[0].isABSONObj() )
o = fields[0].embeddedObject();
const char *ns = fields[1].valuestrsafe();
NamespaceDetails *nsd = nsdetails(ns);
// operation type -- see logOp() comments for types
const char *opType = fields[2].valuestrsafe();
if ( *opType == 'i' ) {
opCounters->gotInsert();
const char *p = strchr(ns, '.');
if ( p && strcmp(p, ".system.indexes") == 0 ) {
// updates aren't allowed for indexes -- so we will do a regular insert. if index already
// exists, that is ok.
theDataFileMgr.insert(ns, (void*) o.objdata(), o.objsize());
}
else {
// do upserts for inserts as we might get replayed more than once
OpDebug debug;
BSONElement _id;
if( !o.getObjectID(_id) ) {
/* No _id. This will be very slow. */
Timer t;
updateObjects(ns, o, o, true, false, false, debug );
if( t.millis() >= 2 ) {
RARELY OCCASIONALLY log() << "warning, repl doing slow updates (no _id field) for " << ns << endl;
}
}
else {
/* erh 10/16/2009 - this is probably not relevant any more since its auto-created, but not worth removing */
RARELY if (nsd && !nsd->capped) {
ensureHaveIdIndex(ns); // otherwise updates will be slow
}
/* todo : it may be better to do an insert here, and then catch the dup key exception and do update
then. very few upserts will not be inserts...
*/
BSONObjBuilder b;
b.append(_id);
updateObjects(ns, o, b.done(), true, false, false , debug );
}
}
}
void applyOperationFromOplog(const BSONObj& op) {
LOG(6) << "applying op: " << op << endl;
OpCounters* opCounters = &replOpCounters;
const char *names[] = {
KEY_STR_NS,
KEY_STR_OP_NAME
};
BSONElement fields[2];
op.getFields(2, names, fields);
const char* ns = fields[0].valuestrsafe();
const char* opType = fields[1].valuestrsafe();
if (strcmp(opType, OP_STR_INSERT) == 0) {
opCounters->gotInsert();
runInsertFromOplog(ns, op);
}
else if (strcmp(opType, OP_STR_UPDATE) == 0) {
opCounters->gotUpdate();
runUpdateFromOplog(ns, op, false);
}
else if (strcmp(opType, OP_STR_DELETE) == 0) {
opCounters->gotDelete();
runDeleteFromOplog(ns, op);
}
else if (strcmp(opType, OP_STR_COMMAND) == 0) {
opCounters->gotCommand();
runCommandFromOplog(ns, op);
}
else if (strcmp(opType, OP_STR_COMMENT) == 0) {
// no-op
}
else if (strcmp(opType, OP_STR_CAPPED_INSERT) == 0) {
opCounters->gotInsert();
runCappedInsertFromOplog(ns, op);
}
else if (strcmp(opType, OP_STR_CAPPED_DELETE) == 0) {
opCounters->gotDelete();
runCappedDeleteFromOplog(ns, op);
}
else {
throw MsgAssertionException( 14825 , ErrorMsg("error in applyOperation : unknown opType ", *opType) );
}
}
void rollbackOperationFromOplog(const BSONObj& op) {
LOG(6) << "rolling back op: " << op << endl;
const char *names[] = {
KEY_STR_NS,
KEY_STR_OP_NAME
};
BSONElement fields[2];
op.getFields(2, names, fields);
const char* ns = fields[0].valuestrsafe();
const char* opType = fields[1].valuestrsafe();
if (strcmp(opType, OP_STR_INSERT) == 0) {
runRollbackInsertFromOplog(ns, op);
}
else if (strcmp(opType, OP_STR_UPDATE) == 0) {
runUpdateFromOplog(ns, op, true);
}
else if (strcmp(opType, OP_STR_DELETE) == 0) {
// the rollback of a delete is to do the insert
runInsertFromOplog(ns, op);
}
else if (strcmp(opType, OP_STR_COMMAND) == 0) {
rollbackCommandFromOplog(ns, op);
}
else if (strcmp(opType, OP_STR_COMMENT) == 0) {
// no-op
}
else if (strcmp(opType, OP_STR_CAPPED_INSERT) == 0) {
runCappedDeleteFromOplog(ns, op);
}
else if (strcmp(opType, OP_STR_CAPPED_DELETE) == 0) {
runCappedInsertFromOplog(ns, op);
}
else {
throw MsgAssertionException( 16795 , ErrorMsg("error in applyOperation : unknown opType ", *opType) );
}
}
bool processObj(const BSONObj &obj) {
if (obj.hasField("$err")) {
log() << "error getting oplog: " << obj << endl;
return false;
}
static const char *names[] = {"ts", "op", "ns", "o", "b"};
BSONElement fields[5];
obj.getFields(5, names, fields);
BSONElement &tsElt = fields[0];
if (!tsElt.ok()) {
log() << "oplog format error: " << obj << " missing 'ts' field." << endl;
return false;
}
if (tsElt.type() != Date && tsElt.type() != Timestamp) {
log() << "oplog format error: " << obj << " wrong 'ts' field type." << endl;
return false;
}
_thisTime = OpTime(tsElt.date());
BSONElement &opElt = fields[1];
if (!opElt.ok()) {
log() << "oplog format error: " << obj << " missing 'op' field." << endl;
return false;
}
StringData op = opElt.Stringdata();
// nop
if (op == "n") {
return true;
}
// "presence of a database"
if (op == "db") {
return true;
}
if (op != "c" && op != "i" && op != "u" && op != "d") {
log() << "oplog format error: " << obj << " has an invalid 'op' field of '" << op << "'." << endl;
return false;
}
if (op != "i" && !_insertBuf.empty()) {
flushInserts();
}
BSONElement &nsElt = fields[2];
if (!nsElt.ok()) {
log() << "oplog format error: " << obj << " missing 'ns' field." << endl;
return false;
}
StringData ns = nsElt.Stringdata();
size_t i = ns.find('.');
if (i == string::npos) {
log() << "oplog format error: invalid namespace '" << ns << "' in op " << obj << "." << endl;
return false;
}
StringData dbname = ns.substr(0, i);
StringData collname = ns.substr(i + 1);
BSONElement &oElt = fields[3];
if (!oElt.ok()) {
log() << "oplog format error: " << obj << " missing 'o' field." << endl;
return false;
}
BSONObj o = obj["o"].Obj();
if (op == "c") {
if (collname != "$cmd") {
log() << "oplog format error: invalid namespace '" << ns << "' for command in op " << obj << "." << endl;
return false;
}
BSONObj info;
bool ok = _conn.runCommand(dbname.toString(), o, info);
if (!ok) {
StringData fieldName = o.firstElementFieldName();
BSONElement errmsgElt = info["errmsg"];
StringData errmsg = errmsgElt.type() == String ? errmsgElt.Stringdata() : "";
bool isDropIndexes = (fieldName == "dropIndexes" || fieldName == "deleteIndexes");
if (((fieldName == "drop" || isDropIndexes) && errmsg == "ns not found") ||
(isDropIndexes && (errmsg == "index not found" || errmsg.find("can't find index with key:") == 0))) {
// This is actually ok. We don't mind dropping something that's not there.
LOG(1) << "Tried to replay " << o << ", got " << info << ", ignoring." << endl;
}
else {
log() << "replay of command " << o << " failed: " << info << endl;
return false;
}
}
}
else {
string nsstr = ns.toString();
if (op == "i") {
if (collname == "system.indexes") {
// Can't ensure multiple indexes in the same batch.
flushInserts();
// For now, we need to strip out any background fields from
// ensureIndex. Once we do hot indexing we can do something more
// like what vanilla applyOperation_inlock does.
if (o["background"].trueValue()) {
BSONObjBuilder builder;
BSONObjIterator it(o);
while (it.more()) {
BSONElement e = it.next();
if (strncmp(e.fieldName(), "background", sizeof("background")) != 0) {
builder.append(e);
}
}
o = builder.obj();
}
// We need to warn very carefully about dropDups.
if (o["dropDups"].trueValue()) {
BSONObjBuilder builder;
BSONObjIterator it(o);
while (it.more()) {
BSONElement e = it.next();
if (strncmp(e.fieldName(), "dropDups", sizeof("dropDups")) != 0) {
builder.append(e);
}
}
warning() << "Detected an ensureIndex with dropDups: true in " << o << "." << endl;
warning() << "This option is not supported in TokuMX, because it deletes arbitrary data." << endl;
warning() << "If it were replayed, it could result in a completely different data set than the source database." << endl;
warning() << "We will attempt to replay it without dropDups, but if that fails, you must restart your migration process." << endl;
_conn.insert(nsstr, o);
string err = _conn.getLastError(dbname.toString(), false, false);
if (!err.empty()) {
log() << "replay of operation " << obj << " failed: " << err << endl;
warning() << "You cannot continue processing this replication stream. You need to restart the migration process." << endl;
_running = false;
_logAtExit = false;
return true;
}
}
}
pushInsert(nsstr, o);
// Don't call GLE or update _maxOpTimeSynced yet.
_thisTime = OpTime();
return true;
}
else if (op == "u") {
BSONElement o2Elt = obj["o2"];
if (!o2Elt.ok()) {
log() << "oplog format error: " << obj << " missing 'o2' field." << endl;
return false;
}
BSONElement &bElt = fields[4];
bool upsert = bElt.booleanSafe();
BSONObj o2 = o2Elt.Obj();
_conn.update(nsstr, o2, o, upsert, false);
}
else if (op == "d") {
BSONElement &bElt = fields[4];
bool justOne = bElt.booleanSafe();
_conn.remove(nsstr, o, justOne);
}
string err = _conn.getLastError(dbname.toString(), false, false);
if (!err.empty()) {
log() << "replay of operation " << obj << " failed: " << err << endl;
return false;
}
}
// If we got here, we completed the operation successfully.
_maxOpTimeSynced = _thisTime;
_thisTime = OpTime();
return true;
}
void applyOperation_inlock(const BSONObj& op , bool fromRepl ) {
assertInWriteLock();
LOG(6) << "applying op: " << op << endl;
OpCounters * opCounters = fromRepl ? &replOpCounters : &globalOpCounters;
const char *names[] = { "o", "ns", "op", "b" };
BSONElement fields[4];
op.getFields(4, names, fields);
BSONObj o;
if( fields[0].isABSONObj() )
o = fields[0].embeddedObject();
const char *ns = fields[1].valuestrsafe();
// operation type -- see logOp() comments for types
const char *opType = fields[2].valuestrsafe();
if ( *opType == 'i' ) {
opCounters->gotInsert();
const char *p = strchr(ns, '.');
if ( p && strcmp(p, ".system.indexes") == 0 ) {
// updates aren't allowed for indexes -- so we will do a regular insert. if index already
// exists, that is ok.
theDataFileMgr.insert(ns, (void*) o.objdata(), o.objsize());
}
else {
// do upserts for inserts as we might get replayed more than once
OpDebug debug;
BSONElement _id;
if( !o.getObjectID(_id) ) {
/* No _id. This will be very slow. */
Timer t;
updateObjects(ns, o, o, true, false, false, debug );
if( t.millis() >= 2 ) {
RARELY OCCASIONALLY log() << "warning, repl doing slow updates (no _id field) for " << ns << endl;
}
}
else {
/* erh 10/16/2009 - this is probably not relevant any more since its auto-created, but not worth removing */
RARELY ensureHaveIdIndex(ns); // otherwise updates will be slow
/* todo : it may be better to do an insert here, and then catch the dup key exception and do update
then. very few upserts will not be inserts...
*/
BSONObjBuilder b;
b.append(_id);
updateObjects(ns, o, b.done(), true, false, false , debug );
}
}
}
else if ( *opType == 'u' ) {
opCounters->gotUpdate();
RARELY ensureHaveIdIndex(ns); // otherwise updates will be super slow
OpDebug debug;
updateObjects(ns, o, op.getObjectField("o2"), /*upsert*/ fields[3].booleanSafe(), /*multi*/ false, /*logop*/ false , debug );
}
else if ( *opType == 'd' ) {
opCounters->gotDelete();
if ( opType[1] == 0 )
deleteObjects(ns, o, /*justOne*/ fields[3].booleanSafe());
else
assert( opType[1] == 'b' ); // "db" advertisement
}
else if ( *opType == 'c' ) {
opCounters->gotCommand();
BufBuilder bb;
BSONObjBuilder ob;
_runCommands(ns, o, bb, ob, true, 0);
}
else if ( *opType == 'n' ) {
// no op
}
else {
throw MsgAssertionException( 14825 , ErrorMsg("error in applyOperation : unknown opType ", *opType) );
}
}
/** @param fromRepl false if from ApplyOpsCmd
@return true if was and update should have happened and the document DNE. see replset initial sync code.
*/
bool applyOperation_inlock(OperationContext* txn,
Database* db,
const BSONObj& op,
bool fromRepl,
bool convertUpdateToUpsert) {
LOG(3) << "applying op: " << op << endl;
bool failedUpdate = false;
OpCounters * opCounters = fromRepl ? &replOpCounters : &globalOpCounters;
const char *names[] = { "o", "ns", "op", "b", "o2" };
BSONElement fields[5];
op.getFields(5, names, fields);
BSONElement& fieldO = fields[0];
BSONElement& fieldNs = fields[1];
BSONElement& fieldOp = fields[2];
BSONElement& fieldB = fields[3];
BSONElement& fieldO2 = fields[4];
BSONObj o;
if( fieldO.isABSONObj() )
o = fieldO.embeddedObject();
const char *ns = fieldNs.valuestrsafe();
BSONObj o2;
if (fieldO2.isABSONObj())
o2 = fieldO2.Obj();
bool valueB = fieldB.booleanSafe();
txn->lockState()->assertWriteLocked(ns);
Collection* collection = db->getCollection( txn, ns );
IndexCatalog* indexCatalog = collection == NULL ? NULL : collection->getIndexCatalog();
// operation type -- see logOp() comments for types
const char *opType = fieldOp.valuestrsafe();
if ( *opType == 'i' ) {
opCounters->gotInsert();
const char *p = strchr(ns, '.');
if ( p && nsToCollectionSubstring( p ) == "system.indexes" ) {
if (o["background"].trueValue()) {
IndexBuilder* builder = new IndexBuilder(o);
// This spawns a new thread and returns immediately.
builder->go();
}
else {
IndexBuilder builder(o);
Status status = builder.buildInForeground(txn, db);
if ( status.isOK() ) {
// yay
}
else if ( status.code() == ErrorCodes::IndexOptionsConflict ||
status.code() == ErrorCodes::IndexKeySpecsConflict ) {
// SERVER-13206, SERVER-13496
// 2.4 (and earlier) will add an ensureIndex to an oplog if its ok or not
// so in 2.6+ where we do stricter validation, it will fail
// but we shouldn't care as the primary is responsible
warning() << "index creation attempted on secondary that conflicts, "
<< "skipping: " << status;
}
else {
uassertStatusOK( status );
}
}
}
else {
// do upserts for inserts as we might get replayed more than once
OpDebug debug;
BSONElement _id;
if( !o.getObjectID(_id) ) {
/* No _id. This will be very slow. */
Timer t;
const NamespaceString requestNs(ns);
UpdateRequest request(txn, requestNs);
request.setQuery(o);
request.setUpdates(o);
request.setUpsert();
request.setFromReplication();
UpdateLifecycleImpl updateLifecycle(true, requestNs);
request.setLifecycle(&updateLifecycle);
update(db, request, &debug);
if( t.millis() >= 2 ) {
RARELY OCCASIONALLY log() << "warning, repl doing slow updates (no _id field) for " << ns << endl;
}
}
else {
// probably don't need this since all replicated colls have _id indexes now
// but keep it just in case
RARELY if ( indexCatalog
&& !collection->isCapped()
&& !indexCatalog->haveIdIndex(txn) ) {
try {
Helpers::ensureIndex(txn, collection, BSON("_id" << 1), true, "_id_");
}
catch (const DBException& e) {
warning() << "Ignoring error building id index on " << collection->ns()
<< ": " << e.toString();
}
}
/* todo : it may be better to do an insert here, and then catch the dup key exception and do update
then. very few upserts will not be inserts...
*/
BSONObjBuilder b;
b.append(_id);
const NamespaceString requestNs(ns);
UpdateRequest request(txn, requestNs);
request.setQuery(b.done());
request.setUpdates(o);
request.setUpsert();
request.setFromReplication();
UpdateLifecycleImpl updateLifecycle(true, requestNs);
request.setLifecycle(&updateLifecycle);
update(db, request, &debug);
}
}
}
/** @param fromRepl false if from ApplyOpsCmd
@return true if was and update should have happened and the document DNE. see replset initial sync code.
*/
bool applyOperation_inlock(const BSONObj& op, bool fromRepl, bool convertUpdateToUpsert) {
LOG(3) << "applying op: " << op << endl;
bool failedUpdate = false;
OpCounters * opCounters = fromRepl ? &replOpCounters : &globalOpCounters;
const char *names[] = { "o", "ns", "op", "b" };
BSONElement fields[4];
op.getFields(4, names, fields);
BSONObj o;
if( fields[0].isABSONObj() )
o = fields[0].embeddedObject();
const char *ns = fields[1].valuestrsafe();
Lock::assertWriteLocked(ns);
NamespaceDetails *nsd = nsdetails(ns);
// operation type -- see logOp() comments for types
const char *opType = fields[2].valuestrsafe();
if ( *opType == 'i' ) {
opCounters->gotInsert();
const char *p = strchr(ns, '.');
if ( p && strcmp(p, ".system.indexes") == 0 ) {
if (o["background"].trueValue()) {
IndexBuilder* builder = new IndexBuilder(ns, o);
// This spawns a new thread and returns immediately.
builder->go();
}
else {
IndexBuilder builder(ns, o);
// Finish the foreground build before returning
builder.build();
}
}
else {
// do upserts for inserts as we might get replayed more than once
OpDebug debug;
BSONElement _id;
if( !o.getObjectID(_id) ) {
/* No _id. This will be very slow. */
Timer t;
const NamespaceString requestNs(ns);
UpdateRequest request(
requestNs, debug,
QueryPlanSelectionPolicy::idElseNatural());
request.setQuery(o);
request.setUpdates(o);
request.setUpsert();
request.setFromReplication();
update(request);
if( t.millis() >= 2 ) {
RARELY OCCASIONALLY log() << "warning, repl doing slow updates (no _id field) for " << ns << endl;
}
}
else {
// probably don't need this since all replicated colls have _id indexes now
// but keep it just in case
RARELY if ( nsd && !nsd->isCapped() ) { ensureHaveIdIndex(ns, false); }
/* todo : it may be better to do an insert here, and then catch the dup key exception and do update
then. very few upserts will not be inserts...
*/
BSONObjBuilder b;
b.append(_id);
const NamespaceString requestNs(ns);
UpdateRequest request(
requestNs, debug,
QueryPlanSelectionPolicy::idElseNatural());
request.setQuery(b.done());
request.setUpdates(o);
request.setUpsert();
request.setFromReplication();
update(request);
}
}
}