DiskLoc DataFileMgr::insert(const char* ns,
const void* obuf,
int32_t len,
bool mayInterrupt,
bool god,
bool mayAddIndex,
bool* addedID) {
Database* database = cc().database();
bool wouldAddIndex = false;
massert( 10093 , "cannot insert into reserved $ collection", god || NamespaceString::normal( ns ) );
uassert( 10094 , str::stream() << "invalid ns: " << ns , isValidNS( ns ) );
{
const char *sys = strstr(ns, "system.");
if ( sys && !insert_checkSys(sys, ns, wouldAddIndex, obuf, god) )
return DiskLoc();
}
bool addIndex = wouldAddIndex && mayAddIndex;
Collection* collection = database->getCollection( ns );
if ( collection == NULL ) {
collection = database->createCollection( ns, false, NULL );
int ies = Extent::initialSize(len);
if( str::contains(ns, '$') &&
len + Record::HeaderSize >= BtreeData_V1::BucketSize - 256 &&
len + Record::HeaderSize <= BtreeData_V1::BucketSize + 256 ) {
// probably an index. so we pick a value here for the first extent instead of using
// initialExtentSize() which is more for user collections.
// TODO: we could look at the # of records in the parent collection to be smarter here.
ies = (32+4) * 1024;
}
collection->increaseStorageSize( ies, false);
if ( !god )
ensureIdIndexForNewNs(ns);
}
NamespaceDetails* d = collection->details();
string tabletoidxns;
Collection* collectionToIndex = 0;
NamespaceDetails* tableToIndex = 0;
BSONObj fixedIndexObject;
if ( addIndex ) {
verify( obuf );
BSONObj io((const char *) obuf);
tabletoidxns = io.getStringField( "ns" );
uassert(10096, "invalid ns to index", tabletoidxns.find( '.' ) != string::npos);
massert(10097,
str::stream() << "trying to create index on wrong db "
<< " db: " << database->name() << " collection: " << tabletoidxns,
database->ownsNS( tabletoidxns ) );
collectionToIndex = database->getCollection( tabletoidxns );
if ( !collectionToIndex ) {
collectionToIndex = database->createCollection( tabletoidxns, false, NULL );
verify( collectionToIndex );
if ( !god )
ensureIdIndexForNewNs( tabletoidxns.c_str() );
}
tableToIndex = collectionToIndex->details();
Status status = collectionToIndex->getIndexCatalog()->okToAddIndex( io );
if ( status.code() == ErrorCodes::IndexAlreadyExists ) {
// dup index, we ignore
return DiskLoc();
}
uassert( 17199,
str::stream() << "cannot build index on " << tabletoidxns
<< " because of " << status.toString(),
status.isOK() );
if( !prepareToBuildIndex(io,
mayInterrupt,
god,
tabletoidxns ) ) {
// prepare creates _id itself, or this indicates to fail the build silently (such
// as if index already exists)
return DiskLoc();
}
fixedIndexObject = IndexCatalog::fixIndexSpec( io );
obuf = fixedIndexObject.objdata();
len = fixedIndexObject.objsize();
}
IDToInsert idToInsert; // only initialized if needed
if( !god ) {
/* Check if we have an _id field. If we don't, we'll add it.
Note that btree buckets which we insert aren't BSONObj's, but in that case god==true.
*/
BSONObj io((const char *) obuf);
BSONElement idField = io.getField( "_id" );
uassert( 10099 , "_id cannot be an array", idField.type() != Array );
// we don't add _id for capped collections in local as they don't have an _id index
if( idField.eoo() &&
!wouldAddIndex &&
nsToDatabase( ns ) != "local" &&
d->haveIdIndex() ) {
if( addedID )
*addedID = true;
idToInsert.init();
len += idToInsert.size();
}
BSONElementManipulator::lookForTimestamps( io );
}
int lenWHdr = d->getRecordAllocationSize( len + Record::HeaderSize );
fassert( 16440, lenWHdr >= ( len + Record::HeaderSize ) );
// If the collection is capped, check if the new object will violate a unique index
// constraint before allocating space.
if ( d->isCapped() && !god) {
BSONObj temp = BSONObj( reinterpret_cast<const char *>( obuf ) );
Status ret = collection->getIndexCatalog()->checkNoIndexConflicts( temp );
uassert(12582, "duplicate key insert for unique index of capped collection", ret.isOK() );
}
DiskLoc loc = allocateSpaceForANewRecord(ns, d, lenWHdr, god);
if ( loc.isNull() ) {
log() << "insert: couldn't alloc space for object ns:" << ns
<< " capped:" << d->isCapped() << endl;
verify(d->isCapped());
return DiskLoc();
}
Record *r = loc.rec();
{
verify( r->lengthWithHeaders() >= lenWHdr );
r = (Record*) getDur().writingPtr(r, lenWHdr);
if( idToInsert.needed() ) {
/* a little effort was made here to avoid a double copy when we add an ID */
int originalSize = *((int*) obuf);
((int&)*r->data()) = originalSize + idToInsert.size();
memcpy(r->data()+4, idToInsert.rawdata(), idToInsert.size());
memcpy(r->data()+4+idToInsert.size(), ((char*)obuf)+4, originalSize-4);
}
else {
if( obuf ) // obuf can be null from internal callers
memcpy(r->data(), obuf, len);
}
}
addRecordToRecListInExtent(r, loc);
d->incrementStats( r->netLength(), 1 );
// we don't bother resetting query optimizer stats for the god tables - also god is true when adding a btree bucket
if ( !god )
collection->infoCache()->notifyOfWriteOp();
if ( tableToIndex ) {
insert_makeIndex(collectionToIndex, loc, mayInterrupt);
}
/* add this record to our indexes */
if ( d->getTotalIndexCount() > 0 ) {
try {
BSONObj obj(r->data());
collection->getIndexCatalog()->indexRecord(obj, loc);
}
catch( AssertionException& e ) {
// should be a dup key error on _id index
if( tableToIndex || d->isCapped() ) {
massert( 12583, "unexpected index insertion failure on capped collection", !d->isCapped() );
string s = e.toString();
s += " : on addIndex/capped - collection and its index will not match";
setLastError(0, s.c_str());
error() << s << endl;
}
else {
// normal case -- we can roll back
_deleteRecord(d, ns, r, loc);
throw;
}
}
}
d->paddingFits();
return loc;
}
UpdateResult update(const UpdateRequest& request, OpDebug* opDebug, UpdateDriver* driver) {
LOG(3) << "processing update : " << request;
const NamespaceString& nsString = request.getNamespaceString();
validateUpdate( nsString.ns().c_str(), request.getUpdates(), request.getQuery() );
NamespaceDetails* nsDetails = nsdetails( nsString.ns() );
NamespaceDetailsTransient* nsDetailsTransient =
&NamespaceDetailsTransient::get( nsString.ns().c_str() );
// TODO: This seems a bit circuitious.
opDebug->updateobj = request.getUpdates();
driver->refreshIndexKeys( nsDetailsTransient->indexKeys() );
shared_ptr<Cursor> cursor = getOptimizedCursor(
nsString.ns(), request.getQuery(), BSONObj(), request.getQueryPlanSelectionPolicy() );
// If the update was marked with '$isolated' (a.k.a '$atomic'), we are not allowed to
// yield while evaluating the update loop below.
//
// TODO: Old code checks this repeatedly within the update loop. Is that necessary? It seems
// that once atomic should be always atomic.
const bool isolated =
cursor->ok() &&
cursor->matcher() &&
cursor->matcher()->docMatcher().atomic();
// The 'cursor' the optimizer gave us may contain query plans that generate duplicate
// diskloc's. We set up here the mechanims that will prevent us from processing those
// twice if we see them. We also set up a 'ClientCursor' so that we can support
// yielding.
//
// TODO: Is it valid to call this on a non-ok cursor?
const bool dedupHere = cursor->autoDedup();
//
// We'll start assuming we have one or more documents for this update. (Othwerwise,
// we'll fallback to upserting.)
//
// We record that this will not be an upsert, in case a mod doesn't want to be applied
// when in strict update mode.
driver->setContext( ModifierInterface::ExecInfo::UPDATE_CONTEXT );
// Let's fetch each of them and pipe them through the update expression, making sure to
// keep track of the necessary stats. Recall that we'll be pulling documents out of
// cursors and some of them do not deduplicate the entries they generate. We have
// deduping logic in here, too -- for now.
unordered_set<DiskLoc, DiskLoc::Hasher> seenLocs;
int numMatched = 0;
// Reset these counters on each call. We might re-enter this function to retry this
// update if we throw a page fault exception below, and we rely on these counters
// reflecting only the actions taken locally. In particlar, we must have the no-op
// counter reset so that we can meaningfully comapre it with numMatched above.
opDebug->nscanned = 0;
opDebug->nupdateNoops = 0;
Client& client = cc();
mutablebson::Document doc;
mutablebson::DamageVector damages;
// If we are going to be yielding, we will need a ClientCursor scoped to this loop. We
// only loop as long as the underlying cursor is OK.
for ( auto_ptr<ClientCursor> clientCursor; cursor->ok(); ) {
// If we haven't constructed a ClientCursor, and if the client allows us to throw
// page faults, and if we are referring to a location that is likely not in
// physical memory, then throw a PageFaultException. The entire operation will be
// restarted.
if ( clientCursor.get() == NULL &&
client.allowedToThrowPageFaultException() &&
!cursor->currLoc().isNull() &&
!cursor->currLoc().rec()->likelyInPhysicalMemory() ) {
// We should never throw a PFE if we have already updated items. The numMatched
// variable includes no-ops, which do not prevent us from raising a PFE, so if
// numMatched is non-zero, we are still OK to throw as long all matched items
// resulted in a no-op.
dassert((numMatched == 0) || (numMatched == opDebug->nupdateNoops));
throw PageFaultException( cursor->currLoc().rec() );
}
if ( !isolated && opDebug->nscanned != 0 ) {
// We are permitted to yield. To do so we need a ClientCursor, so create one
// now if we have not yet done so.
if ( !clientCursor.get() )
clientCursor.reset(
new ClientCursor( QueryOption_NoCursorTimeout, cursor, nsString.ns() ) );
// Ask the client cursor to yield. We get two bits of state back: whether or not
// we yielded, and whether or not we correctly recovered from yielding.
bool yielded = false;
const bool recovered = clientCursor->yieldSometimes(
ClientCursor::WillNeed, &yielded );
if ( !recovered ) {
// If we failed to recover from the yield, then the ClientCursor is already
// gone. Release it so we don't destroy it a second time.
clientCursor.release();
break;
}
if ( !cursor->ok() ) {
// If the cursor died while we were yielded, just get out of the update loop.
break;
}
if ( yielded ) {
// We yielded and recovered OK, and our cursor is still good. Details about
// our namespace may have changed while we were yielded, so we re-acquire
// them here. If we can't do so, escape the update loop. Otherwise, refresh
// the driver so that it knows about what is currently indexed.
nsDetails = nsdetails( nsString.ns() );
if ( !nsDetails )
break;
nsDetailsTransient = &NamespaceDetailsTransient::get( nsString.ns().c_str() );
// TODO: This copies the index keys, but it may not need to do so.
driver->refreshIndexKeys( nsDetailsTransient->indexKeys() );
}
}
// Let's fetch the next candidate object for this update.
Record* record = cursor->_current();
DiskLoc loc = cursor->currLoc();
const BSONObj oldObj = loc.obj();
// We count how many documents we scanned even though we may skip those that are
// deemed duplicated. The final 'numUpdated' and 'nscanned' numbers may differ for
// that reason.
opDebug->nscanned++;
// Skips this document if it:
// a) doesn't match the query portion of the update
// b) was deemed duplicate by the underlying cursor machinery
//
// Now, if we are going to update the document,
// c) we don't want to do so while the cursor is at it, as that may invalidate
// the cursor. So, we advance to next document, before issuing the update.
MatchDetails matchDetails;
matchDetails.requestElemMatchKey();
if ( !cursor->currentMatches( &matchDetails ) ) {
// a)
cursor->advance();
continue;
}
else if ( cursor->getsetdup( loc ) && dedupHere ) {
// b)
cursor->advance();
continue;
}
else if (!driver->isDocReplacement() && request.isMulti()) {
// c)
cursor->advance();
if ( dedupHere ) {
if ( seenLocs.count( loc ) ) {
continue;
}
}
// There are certain kind of cursors that hold multiple pointers to data
// underneath. $or cursors is one example. In a $or cursor, it may be the case
// that when we did the last advance(), we finished consuming documents from
// one of $or child and started consuming the next one. In that case, it is
// possible that the last document of the previous child is the same as the
// first document of the next (see SERVER-5198 and jstests/orp.js).
//
// So we advance the cursor here until we see a new diskloc.
//
// Note that we won't be yielding, and we may not do so for a while if we find
// a particularly duplicated sequence of loc's. That is highly unlikely,
// though. (See SERVER-5725, if curious, but "stage" based $or will make that
// ticket moot).
while( cursor->ok() && loc == cursor->currLoc() ) {
cursor->advance();
}
}
// For some (unfortunate) historical reasons, not all cursors would be valid after
// a write simply because we advanced them to a document not affected by the write.
// To protect in those cases, not only we engaged in the advance() logic above, but
// we also tell the cursor we're about to write a document that we've just seen.
// prepareToTouchEarlierIterate() requires calling later
// recoverFromTouchingEarlierIterate(), so we make a note here to do so.
bool touchPreviousDoc = request.isMulti() && cursor->ok();
if ( touchPreviousDoc ) {
if ( clientCursor.get() )
clientCursor->setDoingDeletes( true );
cursor->prepareToTouchEarlierIterate();
}
// Found a matching document
numMatched++;
// Ask the driver to apply the mods. It may be that the driver can apply those "in
// place", that is, some values of the old document just get adjusted without any
// change to the binary layout on the bson layer. It may be that a whole new
// document is needed to accomodate the new bson layout of the resulting document.
doc.reset( oldObj, mutablebson::Document::kInPlaceEnabled );
BSONObj logObj;
// If there was a matched field, obtain it.
string matchedField;
if (matchDetails.hasElemMatchKey())
matchedField = matchDetails.elemMatchKey();
Status status = driver->update( matchedField, &doc, &logObj );
if ( !status.isOK() ) {
uasserted( 16837, status.reason() );
}
// If the driver applied the mods in place, we can ask the mutable for what
// changed. We call those changes "damages". :) We use the damages to inform the
// journal what was changed, and then apply them to the original document
// ourselves. If, however, the driver applied the mods out of place, we ask it to
// generate a new, modified document for us. In that case, the file manager will
// take care of the journaling details for us.
//
// This code flow is admittedly odd. But, right now, journaling is baked in the file
// manager. And if we aren't using the file manager, we have to do jounaling
// ourselves.
bool objectWasChanged = false;
BSONObj newObj;
const char* source = NULL;
bool inPlace = doc.getInPlaceUpdates(&damages, &source);
if ( inPlace && !driver->modsAffectIndices() ) {
// If a set of modifiers were all no-ops, we are still 'in place', but there is
// no work to do, in which case we want to consider the object unchanged.
if (!damages.empty() ) {
nsDetails->paddingFits();
// All updates were in place. Apply them via durability and writing pointer.
mutablebson::DamageVector::const_iterator where = damages.begin();
const mutablebson::DamageVector::const_iterator end = damages.end();
for( ; where != end; ++where ) {
const char* sourcePtr = source + where->sourceOffset;
void* targetPtr = getDur().writingPtr(
const_cast<char*>(oldObj.objdata()) + where->targetOffset,
where->size);
std::memcpy(targetPtr, sourcePtr, where->size);
}
objectWasChanged = true;
opDebug->fastmod = true;
}
newObj = oldObj;
}
else {
// The updates were not in place. Apply them through the file manager.
newObj = doc.getObject();
DiskLoc newLoc = theDataFileMgr.updateRecord(nsString.ns().c_str(),
nsDetails,
nsDetailsTransient,
record,
loc,
newObj.objdata(),
newObj.objsize(),
*opDebug);
// If we've moved this object to a new location, make sure we don't apply
// that update again if our traversal picks the objecta again.
//
// We also take note that the diskloc if the updates are affecting indices.
// Chances are that we're traversing one of them and they may be multi key and
// therefore duplicate disklocs.
if ( newLoc != loc || driver->modsAffectIndices() ) {
seenLocs.insert( newLoc );
}
objectWasChanged = true;
}
// Log Obj
if ( request.shouldUpdateOpLog() ) {
if ( driver->isDocReplacement() || !logObj.isEmpty() ) {
BSONObj idQuery = driver->makeOplogEntryQuery(newObj, request.isMulti());
logOp("u", nsString.ns().c_str(), logObj , &idQuery,
NULL, request.isFromMigration(), &newObj);
}
}
// If it was noop since the document didn't change, record that.
if (!objectWasChanged)
opDebug->nupdateNoops++;
if (!request.isMulti()) {
break;
}
// If we used the cursor mechanism that prepares an earlier seen document for a
// write we need to tell such mechanisms that the write is over.
if ( touchPreviousDoc ) {
cursor->recoverFromTouchingEarlierIterate();
}
getDur().commitIfNeeded();
}
// TODO: Can this be simplified?
if ((numMatched > 0) || (numMatched == 0 && !request.isUpsert()) ) {
opDebug->nupdated = numMatched;
return UpdateResult( numMatched > 0 /* updated existing object(s) */,
!driver->isDocReplacement() /* $mod or obj replacement */,
numMatched /* # of docments update, even no-ops */,
BSONObj() );
}
//
// We haven't found any existing document so an insert is done
// (upsert is true).
//
opDebug->upsert = true;
// Since this is an insert (no docs found and upsert:true), we will be logging it
// as an insert in the oplog. We don't need the driver's help to build the
// oplog record, then. We also set the context of the update driver to the INSERT_CONTEXT.
// Some mods may only work in that context (e.g. $setOnInsert).
driver->setLogOp( false );
driver->setContext( ModifierInterface::ExecInfo::INSERT_CONTEXT );
BSONObj baseObj;
// Reset the document we will be writing to
doc.reset( baseObj, mutablebson::Document::kInPlaceDisabled );
if ( request.getQuery().hasElement("_id") ) {
uassertStatusOK(doc.root().appendElement(request.getQuery().getField("_id")));
}
// If this is a $mod base update, we need to generate a document by examining the
// query and the mods. Otherwise, we can use the object replacement sent by the user
// update command that was parsed by the driver before.
// In the following block we handle the query part, and then do the regular mods after.
if ( *request.getUpdates().firstElementFieldName() == '$' ) {
uassertStatusOK(UpdateDriver::createFromQuery(request.getQuery(), doc));
opDebug->fastmodinsert = true;
}
// Apply the update modifications and then log the update as an insert manually.
Status status = driver->update( StringData(), &doc, NULL /* no oplog record */);
if ( !status.isOK() ) {
uasserted( 16836, status.reason() );
}
BSONObj newObj = doc.getObject();
theDataFileMgr.insertWithObjMod( nsString.ns().c_str(), newObj, false, request.isGod() );
if ( request.shouldUpdateOpLog() ) {
logOp( "i", nsString.ns().c_str(), newObj,
NULL, NULL, request.isFromMigration(), &newObj );
}
opDebug->nupdated = 1;
return UpdateResult( false /* updated a non existing document */,
!driver->isDocReplacement() /* $mod or obj replacement? */,
1 /* count of updated documents */,
newObj /* object that was upserted */ );
}
UpdateResult _updateObjects( bool su,
const char* ns,
const BSONObj& updateobj,
const BSONObj& patternOrig,
bool upsert,
bool multi,
bool logop ,
OpDebug& debug,
RemoveSaver* rs,
bool fromMigrate,
const QueryPlanSelectionPolicy& planPolicy,
bool forReplication ) {
DEBUGUPDATE( "update: " << ns
<< " update: " << updateobj
<< " query: " << patternOrig
<< " upsert: " << upsert << " multi: " << multi );
Client& client = cc();
debug.updateobj = updateobj;
// The idea with these here it to make them loop invariant for
// multi updates, and thus be a bit faster for that case. The
// pointers may be left invalid on a failed or terminal yield
// recovery.
NamespaceDetails* d = nsdetails(ns); // can be null if an upsert...
NamespaceDetailsTransient* nsdt = &NamespaceDetailsTransient::get(ns);
auto_ptr<ModSet> mods;
bool isOperatorUpdate = updateobj.firstElementFieldName()[0] == '$';
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() );
}
UpdateResult _updateObjectsNEW( 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 ) {
// TODO
// + Separate UpdateParser from UpdateRunner (the latter should be "stage-y")
// + All the yield and deduplicate logic would move to the query stage
// portion of it
//
// + Replication related
// + fast path for update for query by _id
// + support for relaxing viable path constraint in replication
//
// + Field Management
// + Force all upsert to contain _id
// + Prevent changes to immutable fields (_id, and those mentioned by sharding)
//
// + Yiedling related
// + $atomic support (or better, support proper yielding if not)
// + page fault support
debug.updateobj = updateobj;
NamespaceDetails* d = nsdetails( ns );
NamespaceDetailsTransient* nsdt = &NamespaceDetailsTransient::get( ns );
// TODO: Put this logic someplace central and check based on constants (maybe using the
// list of actually excluded config collections, and not global for the config db).
NamespaceString nsStr( ns );
// Should the modifiers validdate their embedded docs via okForStorage
bool shouldValidate = true;
// Config db docs shouldn't get checked for valid field names since the shard key can have
// a dot (".") in it. Therefore we disable validation for storage.
if ( nsStr.db() == "config" ) {
LOG(0) << "disabling okForStorage on config db";
shouldValidate = false;
}
UpdateDriver::Options opts;
opts.multi = multi;
opts.upsert = upsert;
opts.logOp = logop;
opts.modOptions = ModifierInterface::Options( forReplication, shouldValidate );
UpdateDriver driver( opts );
// TODO: This copies the index keys, but we may not actually need to.
Status status = driver.parse( nsdt->indexKeys(), updateobj );
if ( !status.isOK() ) {
uasserted( 16840, status.reason() );
}
shared_ptr<Cursor> cursor = getOptimizedCursor( ns, patternOrig, BSONObj(), planPolicy );
// If the update was marked with '$isolated' (a.k.a '$atomic'), we are not allowed to
// yield while evaluating the update loop below.
//
// TODO: Old code checks this repeatedly within the update loop. Is that necessary? It seems
// that once atomic should be always atomic.
const bool canYield =
cursor->ok() &&
cursor->matcher() &&
cursor->matcher()->docMatcher().atomic();
// The 'cursor' the optimizer gave us may contain query plans that generate duplicate
// diskloc's. We set up here the mechanims that will prevent us from processing those
// twice if we see them. We also set up a 'ClientCursor' so that we can support
// yielding.
//
// TODO: Is it valid to call this on a non-ok cursor?
const bool dedupHere = cursor->autoDedup();
//
// We'll start assuming we have one or more documents for this update. (Othwerwise,
// we'll fallback to upserting.)
//
// We record that this will not be an upsert, in case a mod doesn't want to be applied
// when in strict update mode.
driver.setContext( ModifierInterface::ExecInfo::UPDATE_CONTEXT );
// Let's fetch each of them and pipe them through the update expression, making sure to
// keep track of the necessary stats. Recall that we'll be pulling documents out of
// cursors and some of them do not deduplicate the entries they generate. We have
// deduping logic in here, too -- for now.
unordered_set<DiskLoc, DiskLoc::Hasher> seenLocs;
int numUpdated = 0;
debug.nscanned = 0;
Client& client = cc();
mutablebson::Document doc;
// If we are going to be yielding, we will need a ClientCursor scoped to this loop. We
// only loop as long as the underlying cursor is OK.
for ( auto_ptr<ClientCursor> clientCursor; cursor->ok(); ) {
// If we haven't constructed a ClientCursor, and if the client allows us to throw
// page faults, and if we are referring to a location that is likely not in
// physical memory, then throw a PageFaultException. The entire operation will be
// restarted.
if ( clientCursor.get() == NULL &&
client.allowedToThrowPageFaultException() &&
!cursor->currLoc().isNull() &&
!cursor->currLoc().rec()->likelyInPhysicalMemory() ) {
// We should never throw a PFE if we have already updated items.
dassert(numUpdated == 0);
throw PageFaultException( cursor->currLoc().rec() );
}
if ( !canYield && debug.nscanned != 0 ) {
// We are permitted to yield. To do so we need a ClientCursor, so create one
// now if we have not yet done so.
if ( !clientCursor.get() )
clientCursor.reset(
new ClientCursor( QueryOption_NoCursorTimeout, cursor, ns ) );
// Ask the client cursor to yield. We get two bits of state back: whether or not
// we yielded, and whether or not we correctly recovered from yielding.
bool yielded = false;
const bool recovered = clientCursor->yieldSometimes(
ClientCursor::WillNeed, &yielded );
// If we couldn't recover from the yield, or if the cursor died while we were
// yielded, get out of the update loop right away. We don't need to reset
// 'clientCursor' since we are leaving the scope.
if ( !recovered || !cursor->ok() )
break;
if ( yielded ) {
// Details about our namespace may have changed while we were yielded, so
// we re-acquire them here. If we can't do so, escape the update
// loop. Otherwise, refresh the driver so that it knows about what is
// currently indexed.
d = nsdetails( ns );
if ( !d )
break;
nsdt = &NamespaceDetailsTransient::get( ns );
// TODO: This copies the index keys, but it may not need to do so.
driver.refreshIndexKeys( nsdt->indexKeys() );
}
}
// Let's fetch the next candidate object for this update.
Record* r = cursor->_current();
DiskLoc loc = cursor->currLoc();
const BSONObj oldObj = loc.obj();
// We count how many documents we scanned even though we may skip those that are
// deemed duplicated. The final 'numUpdated' and 'nscanned' numbers may differ for
// that reason.
debug.nscanned++;
// Skips this document if it:
// a) doesn't match the query portion of the update
// b) was deemed duplicate by the underlying cursor machinery
//
// Now, if we are going to update the document,
// c) we don't want to do so while the cursor is at it, as that may invalidate
// the cursor. So, we advance to next document, before issuing the update.
MatchDetails matchDetails;
matchDetails.requestElemMatchKey();
if ( !cursor->currentMatches( &matchDetails ) ) {
// a)
cursor->advance();
continue;
}
else if ( cursor->getsetdup( loc ) && dedupHere ) {
// b)
cursor->advance();
continue;
}
else if (driver.dollarModMode() && multi) {
// c)
cursor->advance();
if ( dedupHere ) {
if ( seenLocs.count( loc ) ) {
continue;
}
}
// There are certain kind of cursors that hold multiple pointers to data
// underneath. $or cursors is one example. In a $or cursor, it may be the case
// that when we did the last advance(), we finished consuming documents from
// one of $or child and started consuming the next one. In that case, it is
// possible that the last document of the previous child is the same as the
// first document of the next (see SERVER-5198 and jstests/orp.js).
//
// So we advance the cursor here until we see a new diskloc.
//
// Note that we won't be yielding, and we may not do so for a while if we find
// a particularly duplicated sequence of loc's. That is highly unlikely,
// though. (See SERVER-5725, if curious, but "stage" based $or will make that
// ticket moot).
while( cursor->ok() && loc == cursor->currLoc() ) {
cursor->advance();
}
}
// For some (unfortunate) historical reasons, not all cursors would be valid after
// a write simply because we advanced them to a document not affected by the write.
// To protect in those cases, not only we engaged in the advance() logic above, but
// we also tell the cursor we're about to write a document that we've just seen.
// prepareToTouchEarlierIterate() requires calling later
// recoverFromTouchingEarlierIterate(), so we make a note here to do so.
bool touchPreviousDoc = multi && cursor->ok();
if ( touchPreviousDoc ) {
if ( clientCursor.get() )
clientCursor->setDoingDeletes( true );
cursor->prepareToTouchEarlierIterate();
}
// Ask the driver to apply the mods. It may be that the driver can apply those "in
// place", that is, some values of the old document just get adjusted without any
// change to the binary layout on the bson layer. It may be that a whole new
// document is needed to accomodate the new bson layout of the resulting document.
doc.reset( oldObj, mutablebson::Document::kInPlaceEnabled );
BSONObj logObj;
StringData matchedField = matchDetails.hasElemMatchKey() ?
matchDetails.elemMatchKey():
StringData();
status = driver.update( matchedField, &doc, &logObj );
if ( !status.isOK() ) {
uasserted( 16837, status.reason() );
}
// If the driver applied the mods in place, we can ask the mutable for what
// changed. We call those changes "damages". :) We use the damages to inform the
// journal what was changed, and then apply them to the original document
// ourselves. If, however, the driver applied the mods out of place, we ask it to
// generate a new, modified document for us. In that case, the file manager will
// take care of the journaling details for us.
//
// This code flow is admittedly odd. But, right now, journaling is baked in the file
// manager. And if we aren't using the file manager, we have to do jounaling
// ourselves.
bool objectWasChanged = false;
BSONObj newObj;
const char* source = NULL;
mutablebson::DamageVector damages;
bool inPlace = doc.getInPlaceUpdates(&damages, &source);
if ( inPlace && !damages.empty() && !driver.modsAffectIndices() ) {
d->paddingFits();
// All updates were in place. Apply them via durability and writing pointer.
mutablebson::DamageVector::const_iterator where = damages.begin();
const mutablebson::DamageVector::const_iterator end = damages.end();
for( ; where != end; ++where ) {
const char* sourcePtr = source + where->sourceOffset;
void* targetPtr = getDur().writingPtr(
const_cast<char*>(oldObj.objdata()) + where->targetOffset,
where->size);
std::memcpy(targetPtr, sourcePtr, where->size);
}
newObj = oldObj;
debug.fastmod = true;
objectWasChanged = true;
}
else {
// The updates were not in place. Apply them through the file manager.
newObj = doc.getObject();
DiskLoc newLoc = theDataFileMgr.updateRecord(ns,
d,
nsdt,
r,
loc,
newObj.objdata(),
newObj.objsize(),
debug);
// If we've moved this object to a new location, make sure we don't apply
// that update again if our traversal picks the objecta again.
//
// We also take note that the diskloc if the updates are affecting indices.
// Chances are that we're traversing one of them and they may be multi key and
// therefore duplicate disklocs.
if ( newLoc != loc || driver.modsAffectIndices() ) {
seenLocs.insert( newLoc );
}
objectWasChanged = true;
}
// Log Obj
if ( logop ) {
if ( !logObj.isEmpty() ) {
BSONObj idQuery = driver.makeOplogEntryQuery(newObj, multi);
logOp("u", ns, logObj , &idQuery, 0, fromMigrate, &newObj);
}
}
// If we applied any in-place updates, or asked the DataFileMgr to write for us,
// then count this as an update.
if (objectWasChanged)
numUpdated++;
if (!multi) {
break;
}
// If we used the cursor mechanism that prepares an earlier seen document for a
// write we need to tell such mechanisms that the write is over.
if ( touchPreviousDoc ) {
cursor->recoverFromTouchingEarlierIterate();
}
getDur().commitIfNeeded();
}
if (numUpdated > 0) {
return UpdateResult( true /* updated existing object(s) */,
driver.dollarModMode() /* $mod or obj replacement */,
numUpdated /* # of docments update */,
BSONObj() );
}
else if (numUpdated == 0 && !upsert) {
return UpdateResult( false /* no object updated */,
driver.dollarModMode() /* $mod or obj replacement */,
0 /* no updates */,
BSONObj() );
}
//
// We haven't succeeded updating any existing document but upserts are allowed.
//
// If this is a $mod base update, we need to generate a document by examining the
// query and the mods. Otherwise, we can use the object replacement sent by the user
// update command that was parsed by the driver before.
BSONObj oldObj;
if ( *updateobj.firstElementFieldName() == '$' ) {
if ( !driver.createFromQuery( patternOrig, &oldObj ) ) {
uasserted( 16835, "cannot create object to update" );
}
debug.fastmodinsert = true;
}
else {
// Copy the _id
if (patternOrig.hasElement("_id")) {
oldObj = patternOrig.getField("_id").wrap();
}
debug.upsert = true;
}
// Since this is an upsert, we will be oplogging it as an insert. We don't
// need the driver's help to build the oplog record, then. We also set the
// context of the update driver to an "upsert". Some mods may only work in that
// context (e.g. $setOnInsert).
driver.setLogOp( false );
driver.setContext( ModifierInterface::ExecInfo::INSERT_CONTEXT );
doc.reset( oldObj, mutablebson::Document::kInPlaceDisabled );
status = driver.update( StringData(), &doc, NULL /* no oplog record */);
if ( !status.isOK() ) {
uasserted( 16836, status.reason() );
}
BSONObj newObj = doc.getObject();
theDataFileMgr.insertWithObjMod( ns, newObj, false, su );
if ( logop ) {
logOp( "i", ns, newObj, 0, 0, fromMigrate, &newObj );
}
return UpdateResult( false /* updated a non existing document */,
driver.dollarModMode() /* $mod or obj replacement? */,
1 /* count of updated documents */,
newObj /* object that was upserted */ );
}