void pretouchOperation(OperationContext* txn, const BSONObj& op) {
if (txn->lockState()->isWriteLocked()) {
// no point pretouching if write locked. not sure if this will ever fire, but just in case.
return;
}
const char* which = "o";
const char* opType = op.getStringField("op");
if (*opType == 'i')
;
else if (*opType == 'u')
which = "o2";
else
return;
/* todo : other operations */
try {
BSONObj o = op.getObjectField(which);
BSONElement _id;
if (o.getObjectID(_id)) {
const char* ns = op.getStringField("ns");
BSONObjBuilder b;
b.append(_id);
BSONObj result;
AutoGetCollectionForRead ctx(txn, ns);
if (Helpers::findById(txn, ctx.getDb(), ns, b.done(), result)) {
_dummy_z += result.objsize(); // touch
}
}
} catch (DBException&) {
log() << "ignoring assertion in pretouchOperation()" << endl;
}
}
/** @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() );
}
}
}
void prefetchRecordPages(const char* ns, const BSONObj& obj) {
BSONElement _id;
if( obj.getObjectID(_id) ) {
TimerHolder timer(&prefetchDocStats);
BSONObjBuilder builder;
builder.append(_id);
BSONObj result;
try {
// we can probably use Client::Context here instead of ReadContext as we
// have locked higher up the call stack already
Client::ReadContext ctx( ns );
if( Helpers::findById(cc(), ns, builder.done(), result) ) {
// do we want to use Record::touch() here? it's pretty similar.
volatile char _dummy_char = '\0';
// Touch the first word on every page in order to fault it into memory
for (int i = 0; i < result.objsize(); i += g_minOSPageSizeBytes) {
_dummy_char += *(result.objdata() + i);
}
// hit the last page, in case we missed it above
_dummy_char += *(result.objdata() + result.objsize() - 1);
}
}
catch(const DBException& e) {
LOG(2) << "ignoring exception in prefetchRecordPages(): " << e.what() << endl;
}
}
}
void pretouchOperation(const BSONObj& op) {
if( dbMutex.isWriteLocked() )
return; // no point pretouching if write locked. not sure if this will ever fire, but just in case.
const char *which = "o";
const char *opType = op.getStringField("op");
if ( *opType == 'i' )
;
else if( *opType == 'u' )
which = "o2";
else
return;
/* todo : other operations */
try {
BSONObj o = op.getObjectField(which);
BSONElement _id;
if( o.getObjectID(_id) ) {
const char *ns = op.getStringField("ns");
BSONObjBuilder b;
b.append(_id);
BSONObj result;
readlock lk(ns);
Client::Context ctx( ns );
if( Helpers::findById(cc(), ns, b.done(), result) )
_dummy_z += result.objsize(); // touch
}
}
catch( DBException& ) {
log() << "ignoring assertion in pretouchOperation()" << endl;
}
}
/** @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 );
}
}
}
/**********************************************************
*similar to writetodb in FileRec
*
***********************************************************/
void VersionRec::writeToDB(mongo::DBClientConnection& conn) {
BSONObjBuilder record;
if (this->versionid.empty()) { //if no id has been read in because it is a new version
record.genOID();//create one
} else {
mongo::OID theoid(this->versionid); //use current id
record.append("_id", theoid);
}
//convert to BSON
record.append("Tempname", this->tmpname);
record.append("filehash", this->filehash);
record.append("length", this->length);
record.append("Version", this->versionnumber);
/* love you */ long long time = this->modifytime.tv_nsec;
record.append("Mtnsec", time);
time = this->modifytime.tv_sec;
record.append("mtsec", time);
mongo::BSONArrayBuilder Version;
for (vector<VersionDiffBlock>::iterator it = changes.begin(); it != changes.end(); ++it) {
BSONObjBuilder version;
version.append("Blknum", (*it).blockNo);
version.append("hash", (*it).blockHash);
Version.append(version.obj());
}
record.append("Blktable", Version.arr());
BSONObj result = record.obj();
if (this->versionid.empty()) {
mongo::BSONElement thing;
result.getObjectID(thing);
mongo::OID anoid = thing.OID();
this->versionid = anoid.toString();
}
auto_ptr<mongo::DBClientCursor> cursor = conn.query("fileRecords.FileVersion", MONGO_QUERY("_id" << mongo::OID(this->versionid)));
if (cursor->more()) {//already a version with same id, update
conn.update("fileRecords.FileVersion", MONGO_QUERY("_id" << mongo::OID(this->versionid)), result);
} else { //new version
conn.insert("fileRecords.FileVersion", result);
}
string e = conn.getLastError();
if (!e.empty()) {
cout << "something failed failed: " << e << std::endl;
sleep(1);
exit(1);
}
else{
cout << "Version " << this->versionnumber << " successfully written to database" << endl;
}
}
BSONObj UpdateDriver::makeOplogEntryQuery(const BSONObj doc, bool multi) const {
BSONObjBuilder idPattern;
BSONElement id;
// NOTE: If the matching object lacks an id, we'll log
// with the original pattern. This isn't replay-safe.
// It might make sense to suppress the log instead
// if there's no id.
if ( doc.getObjectID( id ) ) {
idPattern.append( id );
return idPattern.obj();
}
else {
uassert( 16980, "multi-update requires all modified objects to have an _id" , ! multi );
return doc;
}
}
string SavePaymentJson(int amount)
{
//BSONObj paymentBSON = mongo::fromjson(newPyamentJson);
BSONObj paymentBSON = BSON(GENOID
<< "PayedToUserId" << 8888
<< "PayedDate" << "2015-01-25 12:00:00"
<< "PayedPeriodStartDate" << "2015-01-01 00:00:00"
<< "PayedPeriodEndDate" << "2015-01-29 23:59:59"
<< "Amount" << amount);
db.insert(PAYMENTS_COLLECTION_NAMESPASE, paymentBSON);
BSONElement oi;
paymentBSON.getObjectID(oi);
OID oid = oi.__oid();
return oid.toString();
}
BSONObj UpdateDriver::makeOplogEntryQuery(const BSONObj& doc, bool multi) const {
BSONObjBuilder idPattern;
BSONElement id;
// NOTE: If the matching object lacks an id, we'll log
// with the original pattern. This isn't replay-safe.
// It might make sense to suppress the log instead
// if there's no id.
if (doc.getObjectID(id)) {
idPattern.append(id);
return idPattern.obj();
} else {
uassert(16980,
str::stream() << "Multi-update operations require all documents to "
"have an '_id' field. " << doc.toString(false, false),
!multi);
return doc;
}
}
void pretouchN(vector<BSONObj>& v, unsigned a, unsigned b) {
Client* c = currentClient.get();
if (c == 0) {
Client::initThread("pretouchN");
c = &cc();
}
OperationContextImpl txn; // XXX
ScopedTransaction transaction(&txn, MODE_S);
Lock::GlobalRead lk(txn.lockState());
for (unsigned i = a; i <= b; i++) {
const BSONObj& op = v[i];
const char* which = "o";
const char* opType = op.getStringField("op");
if (*opType == 'i')
;
else if (*opType == 'u')
which = "o2";
else
continue;
/* todo : other operations */
try {
BSONObj o = op.getObjectField(which);
BSONElement _id;
if (o.getObjectID(_id)) {
const char* ns = op.getStringField("ns");
BSONObjBuilder b;
b.append(_id);
BSONObj result;
Client::Context ctx(&txn, ns);
if (Helpers::findById(&txn, ctx.db(), ns, b.done(), result))
_dummy_z += result.objsize(); // touch
}
} catch (DBException& e) {
log() << "ignoring assertion in pretouchN() " << a << ' ' << b << ' ' << i << ' '
<< e.toString() << endl;
}
}
}
void pretouchN(vector<BSONObj>& v, unsigned a, unsigned b) {
DEV assert( !dbMutex.isWriteLocked() );
Client *c = currentClient.get();
if( c == 0 ) {
Client::initThread("pretouchN");
c = &cc();
}
readlock lk("");
for( unsigned i = a; i <= b; i++ ) {
const BSONObj& op = v[i];
const char *which = "o";
const char *opType = op.getStringField("op");
if ( *opType == 'i' )
;
else if( *opType == 'u' )
which = "o2";
else
continue;
/* todo : other operations */
try {
BSONObj o = op.getObjectField(which);
BSONElement _id;
if( o.getObjectID(_id) ) {
const char *ns = op.getStringField("ns");
BSONObjBuilder b;
b.append(_id);
BSONObj result;
Client::Context ctx( ns );
if( Helpers::findById(cc(), ns, b.done(), result) )
_dummy_z += result.objsize(); // touch
}
}
catch( DBException& e ) {
log() << "ignoring assertion in pretouchN() " << a << ' ' << b << ' ' << i << ' ' << e.toString() << endl;
}
}
}
UpdateResult _updateObjects( const char* ns,
const BSONObj& updateobj,
const BSONObj& patternOrig,
bool upsert,
bool multi,
bool logop ,
OpDebug& debug,
bool fromMigrate,
const QueryPlanSelectionPolicy& planPolicy ) {
TOKULOG(2) << "update: " << ns
<< " update: " << updateobj
<< " query: " << patternOrig
<< " upsert: " << upsert << " multi: " << multi << endl;
debug.updateobj = updateobj;
NamespaceDetails *d = getAndMaybeCreateNS(ns, logop);
auto_ptr<ModSet> mods;
const bool isOperatorUpdate = updateobj.firstElementFieldName()[0] == '$';
bool modsAreIndexed = false;
if ( isOperatorUpdate ) {
if ( d->indexBuildInProgress() ) {
set<string> bgKeys;
d->inProgIdx().keyPattern().getFieldNames(bgKeys);
mods.reset( new ModSet(updateobj, d->indexKeys(), &bgKeys) );
}
else {
mods.reset( new ModSet(updateobj, d->indexKeys()) );
}
modsAreIndexed = mods->isIndexed();
}
int idIdxNo = -1;
if ( planPolicy.permitOptimalIdPlan() && !multi && !modsAreIndexed &&
(idIdxNo = d->findIdIndex()) >= 0 && mayUpdateById(d, patternOrig) ) {
debug.idhack = true;
IndexDetails &idx = d->idx(idIdxNo);
BSONObj pk = idx.getKeyFromQuery(patternOrig);
TOKULOG(3) << "_updateObjects using simple _id query, pattern " << patternOrig << ", pk " << pk << endl;
UpdateResult result = _updateById( pk,
isOperatorUpdate,
mods.get(),
d,
ns,
updateobj,
patternOrig,
logop,
debug,
fromMigrate);
if ( result.existing || ! upsert ) {
return result;
}
else if ( upsert && ! isOperatorUpdate && ! logop) {
debug.upsert = true;
BSONObj objModified = updateobj;
insertAndLog( ns, d, objModified, logop, fromMigrate );
return UpdateResult( 0 , 0 , 1 , updateobj );
}
}
int numModded = 0;
debug.nscanned = 0;
shared_ptr<Cursor> c = getOptimizedCursor( ns, patternOrig, BSONObj(), planPolicy );
if( c->ok() ) {
set<BSONObj> seenObjects;
MatchDetails details;
auto_ptr<ClientCursor> cc;
do {
debug.nscanned++;
if ( mods.get() && mods->hasDynamicArray() ) {
// The Cursor must have a Matcher to record an elemMatchKey. But currently
// a modifier on a dynamic array field may be applied even if there is no
// elemMatchKey, so a matcher cannot be required.
//verify( c->matcher() );
details.requestElemMatchKey();
}
if ( !c->currentMatches( &details ) ) {
c->advance();
continue;
}
BSONObj currPK = c->currPK();
if ( c->getsetdup( currPK ) ) {
c->advance();
continue;
}
BSONObj currentObj = c->current();
BSONObj pattern = patternOrig;
if ( logop ) {
BSONObjBuilder idPattern;
BSONElement id;
// NOTE: If the matching object lacks an id, we'll log
// with the original pattern. This isn't replay-safe.
// It might make sense to suppress the log instead
// if there's no id.
if ( currentObj.getObjectID( id ) ) {
idPattern.append( id );
pattern = idPattern.obj();
}
else {
uassert( 10157 , "multi-update requires all modified objects to have an _id" , ! multi );
}
}
/* look for $inc etc. note as listed here, all fields to inc must be this type, you can't set some
regular ones at the moment. */
struct LogOpUpdateDetails loud;
loud.logop = logop;
loud.ns = ns;
loud.fromMigrate = fromMigrate;
if ( isOperatorUpdate ) {
if ( multi ) {
// Make our own copies of the currPK and currentObj before we invalidate
// them by advancing the cursor.
currPK = currPK.copy();
currentObj = currentObj.copy();
// Advance past the document to be modified. This used to be because of SERVER-5198,
// but TokuMX does it because we want to avoid needing to do manual deduplication
// of this PK on the next iteration if the current update modifies the next
// entry in the index. For example, an index scan over a:1 with mod {$inc: {a:1}}
// would cause every other key read to be a duplicate if we didn't advance here.
while ( c->ok() && currPK == c->currPK() ) {
c->advance();
}
// Multi updates need to do their own deduplication because updates may modify the
// keys the cursor is in the process of scanning over.
if ( seenObjects.count( currPK ) ) {
continue;
} else {
seenObjects.insert( currPK );
}
}
ModSet* useMods = mods.get();
auto_ptr<ModSet> mymodset;
if ( details.hasElemMatchKey() && mods->hasDynamicArray() ) {
useMods = mods->fixDynamicArray( details.elemMatchKey() );
mymodset.reset( useMods );
}
auto_ptr<ModSetState> mss = useMods->prepare( currentObj );
updateUsingMods( d, currPK, currentObj, *mss, &loud );
numModded++;
if ( ! multi )
return UpdateResult( 1 , 1 , numModded , BSONObj() );
continue;
} // end if operator is update
uassert( 10158 , "multi update only works with $ operators" , ! multi );
updateNoMods( d, currPK, currentObj, updateobj, &loud );
return UpdateResult( 1 , 0 , 1 , BSONObj() );
} while ( c->ok() );
} // endif
if ( numModded )
return UpdateResult( 1 , 1 , numModded , BSONObj() );
if ( upsert ) {
BSONObj newObj = updateobj;
if ( updateobj.firstElementFieldName()[0] == '$' ) {
// upsert of an $operation. build a default object
BSONObj newObj = mods->createNewFromQuery( patternOrig );
debug.fastmodinsert = true;
insertAndLog( ns, d, newObj, logop, fromMigrate );
return UpdateResult( 0 , 1 , 1 , newObj );
}
uassert( 10159 , "multi update only works with $ operators" , ! multi );
debug.upsert = true;
insertAndLog( ns, d, newObj, logop, fromMigrate );
return UpdateResult( 0 , 0 , 1 , newObj );
}
return UpdateResult( 0 , isOperatorUpdate , 0 , BSONObj() );
}
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) );
}
}
/** 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;
}
UpdateResult _updateObjects( bool su,
const char* ns,
const BSONObj& updateobj,
const BSONObj& patternOrig,
bool upsert,
bool multi,
bool logop ,
OpDebug& debug,
RemoveSaver* rs,
bool fromMigrate,
const QueryPlanSelectionPolicy& planPolicy,
bool forReplication ) {
DEBUGUPDATE( "update: " << ns
<< " update: " << updateobj
<< " query: " << patternOrig
<< " upsert: " << upsert << " multi: " << multi );
Client& client = cc();
debug.updateobj = updateobj;
// The idea with these here it to make them loop invariant for
// multi updates, and thus be a bit faster for that case. The
// pointers may be left invalid on a failed or terminal yield
// recovery.
NamespaceDetails* d = nsdetails(ns); // can be null if an upsert...
NamespaceDetailsTransient* nsdt = &NamespaceDetailsTransient::get(ns);
auto_ptr<ModSet> mods;
bool isOperatorUpdate = updateobj.firstElementFieldName()[0] == '$';
int modsIsIndexed = false; // really the # of indexes
if ( isOperatorUpdate ) {
mods.reset( new ModSet(updateobj, nsdt->indexKeys(), forReplication) );
modsIsIndexed = mods->maxNumIndexUpdated();
}
if( planPolicy.permitOptimalIdPlan() && !multi && isSimpleIdQuery(patternOrig) && d &&
!modsIsIndexed ) {
int idxNo = d->findIdIndex();
if( idxNo >= 0 ) {
debug.idhack = true;
UpdateResult result = _updateById( isOperatorUpdate,
idxNo,
mods.get(),
d,
nsdt,
su,
ns,
updateobj,
patternOrig,
logop,
debug,
fromMigrate);
if ( result.existing || ! upsert ) {
return result;
}
else if ( upsert && ! isOperatorUpdate ) {
// this handles repl inserts
checkNoMods( updateobj );
debug.upsert = true;
BSONObj no = updateobj;
theDataFileMgr.insertWithObjMod(ns, no, false, su);
if ( logop )
logOp( "i", ns, no, 0, 0, fromMigrate, &no );
return UpdateResult( 0 , 0 , 1 , no );
}
}
}
int numModded = 0;
debug.nscanned = 0;
shared_ptr<Cursor> c = getOptimizedCursor( ns, patternOrig, BSONObj(), planPolicy );
d = nsdetails(ns);
nsdt = &NamespaceDetailsTransient::get(ns);
bool autoDedup = c->autoDedup();
if( c->ok() ) {
set<DiskLoc> seenObjects;
MatchDetails details;
auto_ptr<ClientCursor> cc;
do {
if ( cc.get() == 0 &&
client.allowedToThrowPageFaultException() &&
! c->currLoc().isNull() &&
! c->currLoc().rec()->likelyInPhysicalMemory() ) {
throw PageFaultException( c->currLoc().rec() );
}
bool atomic = c->matcher() && c->matcher()->docMatcher().atomic();
if ( ! atomic && debug.nscanned > 0 ) {
// we need to use a ClientCursor to yield
if ( cc.get() == 0 ) {
shared_ptr< Cursor > cPtr = c;
cc.reset( new ClientCursor( QueryOption_NoCursorTimeout , cPtr , ns ) );
}
bool didYield;
if ( ! cc->yieldSometimes( ClientCursor::WillNeed, &didYield ) ) {
cc.release();
break;
}
if ( !c->ok() ) {
break;
}
if ( didYield ) {
d = nsdetails(ns);
if ( ! d )
break;
nsdt = &NamespaceDetailsTransient::get(ns);
if ( mods.get() ) {
mods->setIndexedStatus( nsdt->indexKeys() );
modsIsIndexed = mods->maxNumIndexUpdated();
}
}
} // end yielding block
debug.nscanned++;
if ( mods.get() && mods->hasDynamicArray() ) {
details.requestElemMatchKey();
}
if ( !c->currentMatches( &details ) ) {
c->advance();
continue;
}
Record* r = c->_current();
DiskLoc loc = c->currLoc();
if ( c->getsetdup( loc ) && autoDedup ) {
c->advance();
continue;
}
BSONObj js = BSONObj::make(r);
BSONObj pattern = patternOrig;
if ( logop ) {
BSONObjBuilder idPattern;
BSONElement id;
// NOTE: If the matching object lacks an id, we'll log
// with the original pattern. This isn't replay-safe.
// It might make sense to suppress the log instead
// if there's no id.
if ( js.getObjectID( id ) ) {
idPattern.append( id );
pattern = idPattern.obj();
}
else {
uassert( 10157 , "multi-update requires all modified objects to have an _id" , ! multi );
}
}
/* look for $inc etc. note as listed here, all fields to inc must be this type, you can't set some
regular ones at the moment. */
if ( isOperatorUpdate ) {
if ( multi ) {
// go to next record in case this one moves
c->advance();
// Update operations are deduped for cursors that implement their own
// deduplication. In particular, some geo cursors are excluded.
if ( autoDedup ) {
if ( seenObjects.count( loc ) ) {
continue;
}
// SERVER-5198 Advance past the document to be modified, provided
// deduplication is enabled, but see SERVER-5725.
while( c->ok() && loc == c->currLoc() ) {
c->advance();
}
}
}
const BSONObj& onDisk = loc.obj();
ModSet* useMods = mods.get();
auto_ptr<ModSet> mymodset;
if ( details.hasElemMatchKey() && mods->hasDynamicArray() ) {
useMods = mods->fixDynamicArray( details.elemMatchKey() );
mymodset.reset( useMods );
}
auto_ptr<ModSetState> mss = useMods->prepare( onDisk,
false /* not an insertion */ );
bool willAdvanceCursor = multi && c->ok() && ( modsIsIndexed || ! mss->canApplyInPlace() );
if ( willAdvanceCursor ) {
if ( cc.get() ) {
cc->setDoingDeletes( true );
}
c->prepareToTouchEarlierIterate();
}
// If we've made it this far, "ns" must contain a valid collection name, and so
// is of the form "db.collection". Therefore, the following expression must
// always be valid. "system.users" updates must never be done in place, in
// order to ensure that they are validated inside DataFileMgr::updateRecord(.).
bool isSystemUsersMod = (NamespaceString(ns).coll == "system.users");
BSONObj newObj;
if ( !mss->isUpdateIndexed() && mss->canApplyInPlace() && !isSystemUsersMod ) {
mss->applyModsInPlace( true );// const_cast<BSONObj&>(onDisk) );
DEBUGUPDATE( "\t\t\t doing in place update" );
if ( !multi )
debug.fastmod = true;
if ( modsIsIndexed ) {
seenObjects.insert( loc );
}
newObj = loc.obj();
d->paddingFits();
}
else {
newObj = mss->createNewFromMods();
checkTooLarge(newObj);
DiskLoc newLoc = theDataFileMgr.updateRecord(ns,
d,
nsdt,
r,
loc,
newObj.objdata(),
newObj.objsize(),
debug);
if ( newLoc != loc || modsIsIndexed ){
// log() << "Moved obj " << newLoc.obj()["_id"] << " from " << loc << " to " << newLoc << endl;
// object moved, need to make sure we don' get again
seenObjects.insert( newLoc );
}
}
if ( logop ) {
DEV verify( mods->size() );
BSONObj logObj = mss->getOpLogRewrite();
DEBUGUPDATE( "\t rewrite update: " << logObj );
// It is possible that the entire mod set was a no-op over this
// document. We would have an empty log record in that case. If we
// call logOp, with an empty record, that would be replicated as "clear
// this record", which is not what we want. Therefore, to get a no-op
// in the replica, we simply don't log.
if ( logObj.nFields() ) {
logOp("u", ns, logObj , &pattern, 0, fromMigrate, &newObj );
}
}
numModded++;
if ( ! multi )
return UpdateResult( 1 , 1 , numModded , BSONObj() );
if ( willAdvanceCursor )
c->recoverFromTouchingEarlierIterate();
getDur().commitIfNeeded();
continue;
}
uassert( 10158 , "multi update only works with $ operators" , ! multi );
BSONElementManipulator::lookForTimestamps( updateobj );
checkNoMods( updateobj );
theDataFileMgr.updateRecord(ns, d, nsdt, r, loc , updateobj.objdata(), updateobj.objsize(), debug, su);
if ( logop ) {
DEV wassert( !su ); // super used doesn't get logged, this would be bad.
logOp("u", ns, updateobj, &pattern, 0, fromMigrate, &updateobj );
}
return UpdateResult( 1 , 0 , 1 , BSONObj() );
} while ( c->ok() );
} // endif
if ( numModded )
return UpdateResult( 1 , 1 , numModded , BSONObj() );
if ( upsert ) {
if ( updateobj.firstElementFieldName()[0] == '$' ) {
// upsert of an $operation. build a default object
BSONObj newObj = mods->createNewFromQuery( patternOrig );
checkNoMods( newObj );
debug.fastmodinsert = true;
theDataFileMgr.insertWithObjMod(ns, newObj, false, su);
if ( logop )
logOp( "i", ns, newObj, 0, 0, fromMigrate, &newObj );
return UpdateResult( 0 , 1 , 1 , newObj );
}
uassert( 10159 , "multi update only works with $ operators" , ! multi );
checkNoMods( updateobj );
debug.upsert = true;
BSONObj no = updateobj;
theDataFileMgr.insertWithObjMod(ns, no, false, su);
if ( logop )
logOp( "i", ns, no, 0, 0, fromMigrate, &no );
return UpdateResult( 0 , 0 , 1 , no );
}
return UpdateResult( 0 , isOperatorUpdate , 0 , BSONObj() );
}
/** @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);
}
}
}
void applyOperation_inlock(const BSONObj& op , bool fromRepl ) {
OpCounters * opCounters = fromRepl ? &replOpCounters : &globalOpCounters;
if( logLevel >= 6 )
log() << "applying op: " << op << endl;
assertInWriteLock();
OpDebug debug;
BSONObj o = op.getObjectField("o");
const char *ns = op.getStringField("ns");
// operation type -- see logOp() comments for types
const char *opType = op.getStringField("op");
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
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 {
BSONObjBuilder b;
b.append(_id);
/* 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...
*/
updateObjects(ns, o, b.done(), true, false, false , debug );
}
}
}
else if ( *opType == 'u' ) {
opCounters->gotUpdate();
RARELY ensureHaveIdIndex(ns); // otherwise updates will be super slow
updateObjects(ns, o, op.getObjectField("o2"), /*upsert*/ op.getBoolField("b"), /*multi*/ false, /*logop*/ false , debug );
}
else if ( *opType == 'd' ) {
opCounters->gotDelete();
if ( opType[1] == 0 )
deleteObjects(ns, o, op.getBoolField("b"));
else
assert( opType[1] == 'b' ); // "db" advertisement
}
else if ( *opType == 'n' ) {
// no op
}
else if ( *opType == 'c' ) {
opCounters->gotCommand();
BufBuilder bb;
BSONObjBuilder ob;
_runCommands(ns, o, bb, ob, true, 0);
}
else {
stringstream ss;
ss << "unknown opType [" << opType << "]";
throw MsgAssertionException( 13141 , ss.str() );
}
}
/** @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);
}
}
}
string StorageEngine::getoid(BSONObj post) {
BSONElement el;
post.getObjectID(el);
return el.__oid().str();
}