void Strategy::queryOp(OperationContext* txn, Request& request) {
verify(!NamespaceString(request.getns()).isCommand());
Timer queryTimer;
globalOpCounters.gotQuery();
QueryMessage q(request.d());
NamespaceString ns(q.ns);
ClientBasic* client = txn->getClient();
AuthorizationSession* authSession = AuthorizationSession::get(client);
Status status = authSession->checkAuthForFind(ns, false);
audit::logQueryAuthzCheck(client, ns, q.query, status.code());
uassertStatusOK(status);
LOG(3) << "query: " << q.ns << " " << q.query << " ntoreturn: " << q.ntoreturn
<< " options: " << q.queryOptions;
if (q.ntoreturn == 1 && strstr(q.ns, ".$cmd"))
throw UserException(8010, "something is wrong, shouldn't see a command here");
if (q.queryOptions & QueryOption_Exhaust) {
uasserted(18526,
string("the 'exhaust' query option is invalid for mongos queries: ") + q.ns +
" " + q.query.toString());
}
// Spigot which controls whether OP_QUERY style find on mongos uses the new ClusterClientCursor
// code path.
// TODO: Delete the spigot and always use the new code.
if (useClusterClientCursor) {
ReadPreferenceSetting readPreference(ReadPreference::PrimaryOnly, TagSet::primaryOnly());
BSONElement rpElem;
auto readPrefExtractStatus = bsonExtractTypedField(
q.query, LiteParsedQuery::kWrappedReadPrefField, mongo::Object, &rpElem);
if (readPrefExtractStatus.isOK()) {
auto parsedRps = ReadPreferenceSetting::fromBSON(rpElem.Obj());
uassertStatusOK(parsedRps.getStatus());
readPreference = parsedRps.getValue();
} else if (readPrefExtractStatus != ErrorCodes::NoSuchKey) {
uassertStatusOK(readPrefExtractStatus);
}
auto canonicalQuery = CanonicalQuery::canonicalize(q, WhereCallbackNoop());
uassertStatusOK(canonicalQuery.getStatus());
// If the $explain flag was set, we must run the operation on the shards as an explain
// command rather than a find command.
if (canonicalQuery.getValue()->getParsed().isExplain()) {
const LiteParsedQuery& lpq = canonicalQuery.getValue()->getParsed();
BSONObj findCommand = lpq.asFindCommand();
// We default to allPlansExecution verbosity.
auto verbosity = ExplainCommon::EXEC_ALL_PLANS;
const bool secondaryOk = (readPreference.pref != ReadPreference::PrimaryOnly);
rpc::ServerSelectionMetadata metadata(secondaryOk, readPreference);
BSONObjBuilder explainBuilder;
uassertStatusOK(ClusterFind::runExplain(
txn, findCommand, lpq, verbosity, metadata, &explainBuilder));
BSONObj explainObj = explainBuilder.done();
replyToQuery(0, // query result flags
request.p(),
request.m(),
static_cast<const void*>(explainObj.objdata()),
explainObj.objsize(),
1, // numResults
0, // startingFrom
CursorId(0));
return;
}
// Do the work to generate the first batch of results. This blocks waiting to get responses
// from the shard(s).
std::vector<BSONObj> batch;
// 0 means the cursor is exhausted and
// otherwise we assume that a cursor with the returned id can be retrieved via the
// ClusterCursorManager
auto cursorId =
ClusterFind::runQuery(txn, *canonicalQuery.getValue(), readPreference, &batch);
uassertStatusOK(cursorId.getStatus());
// TODO: this constant should be shared between mongos and mongod, and should
// not be inside ShardedClientCursor.
BufBuilder buffer(ShardedClientCursor::INIT_REPLY_BUFFER_SIZE);
// Fill out the response buffer.
int numResults = 0;
for (const auto& obj : batch) {
buffer.appendBuf((void*)obj.objdata(), obj.objsize());
numResults++;
}
replyToQuery(0, // query result flags
request.p(),
request.m(),
buffer.buf(),
buffer.len(),
numResults,
0, // startingFrom
cursorId.getValue());
return;
}
QuerySpec qSpec((string)q.ns, q.query, q.fields, q.ntoskip, q.ntoreturn, q.queryOptions);
// Parse "$maxTimeMS".
StatusWith<int> maxTimeMS = LiteParsedQuery::parseMaxTimeMSQuery(q.query);
uassert(17233, maxTimeMS.getStatus().reason(), maxTimeMS.isOK());
if (_isSystemIndexes(q.ns) && doShardedIndexQuery(txn, request, qSpec)) {
return;
}
ParallelSortClusteredCursor* cursor = new ParallelSortClusteredCursor(qSpec, CommandInfo());
verify(cursor);
// TODO: Move out to Request itself, not strategy based
try {
cursor->init(txn);
if (qSpec.isExplain()) {
BSONObjBuilder explain_builder;
cursor->explain(explain_builder);
explain_builder.appendNumber("executionTimeMillis",
static_cast<long long>(queryTimer.millis()));
BSONObj b = explain_builder.obj();
replyToQuery(0, request.p(), request.m(), b);
delete (cursor);
return;
}
} catch (...) {
delete cursor;
throw;
}
// TODO: Revisit all of this when we revisit the sharded cursor cache
if (cursor->getNumQueryShards() != 1) {
// More than one shard (or zero), manage with a ShardedClientCursor
// NOTE: We may also have *zero* shards here when the returnPartial flag is set.
// Currently the code in ShardedClientCursor handles this.
ShardedClientCursorPtr cc(new ShardedClientCursor(q, cursor));
BufBuilder buffer(ShardedClientCursor::INIT_REPLY_BUFFER_SIZE);
int docCount = 0;
const int startFrom = cc->getTotalSent();
bool hasMore = cc->sendNextBatch(q.ntoreturn, buffer, docCount);
if (hasMore) {
LOG(5) << "storing cursor : " << cc->getId();
int cursorLeftoverMillis = maxTimeMS.getValue() - queryTimer.millis();
if (maxTimeMS.getValue() == 0) { // 0 represents "no limit".
cursorLeftoverMillis = kMaxTimeCursorNoTimeLimit;
} else if (cursorLeftoverMillis <= 0) {
cursorLeftoverMillis = kMaxTimeCursorTimeLimitExpired;
}
cursorCache.store(cc, cursorLeftoverMillis);
}
replyToQuery(0,
request.p(),
request.m(),
buffer.buf(),
buffer.len(),
docCount,
startFrom,
hasMore ? cc->getId() : 0);
} else {
// Only one shard is used
// Remote cursors are stored remotely, we shouldn't need this around.
unique_ptr<ParallelSortClusteredCursor> cursorDeleter(cursor);
ShardPtr shard = grid.shardRegistry()->getShard(txn, cursor->getQueryShardId());
verify(shard.get());
DBClientCursorPtr shardCursor = cursor->getShardCursor(shard->getId());
// Implicitly stores the cursor in the cache
request.reply(*(shardCursor->getMessage()), shardCursor->originalHost());
// We don't want to kill the cursor remotely if there's still data left
shardCursor->decouple();
}
}
/**
* This is called by db/ops/query.cpp. This is the entry point for answering a query.
*/
std::string newRunQuery(CanonicalQuery* cq, CurOp& curop, Message &result) {
// This is a read lock.
Client::ReadContext ctx(cq->ns(), storageGlobalParams.dbpath);
// Parse, canonicalize, plan, transcribe, and get a runner.
Runner* rawRunner;
// Takes ownership of cq.
Status status = getRunner(cq, &rawRunner);
if (!status.isOK()) {
uasserted(17007, "Couldn't process query " + cq->toString()
+ " why: " + status.reason());
}
verify(NULL != rawRunner);
auto_ptr<Runner> runner(rawRunner);
QLOG() << "Running query on new system: " << cq->toString();
// We freak out later if this changes before we're done with the query.
const ChunkVersion shardingVersionAtStart = shardingState.getVersion(cq->ns());
// We use this a lot below.
const LiteParsedQuery& pq = cq->getParsed();
// TODO: Remove when impl'd
if (pq.hasOption(QueryOption_OplogReplay)) {
warning() << "haven't implemented findingstartcursor yet\n";
}
// Handle query option $maxTimeMS (not used with commands).
curop.setMaxTimeMicros(static_cast<unsigned long long>(pq.getMaxTimeMS()) * 1000);
killCurrentOp.checkForInterrupt(); // May trigger maxTimeAlwaysTimeOut fail point.
// uassert if we are not on a primary, and not a secondary with SlaveOk query parameter set.
replVerifyReadsOk(&pq);
// If this exists, the collection is sharded.
// If it doesn't exist, we can assume we're not sharded.
// If we're sharded, we might encounter data that is not consistent with our sharding state.
// We must ignore this data.
CollectionMetadataPtr collMetadata;
if (!shardingState.needCollectionMetadata(pq.ns())) {
collMetadata = CollectionMetadataPtr();
}
else {
collMetadata = shardingState.getCollectionMetadata(pq.ns());
}
// Run the query.
// bb is used to hold query results
// this buffer should contain either requested documents per query or
// explain information, but not both
BufBuilder bb(32768);
bb.skip(sizeof(QueryResult));
// How many results have we obtained from the runner?
int numResults = 0;
// If we're replaying the oplog, we save the last time that we read.
OpTime slaveReadTill;
// Do we save the Runner in a ClientCursor for getMore calls later?
bool saveClientCursor = false;
// We turn on auto-yielding for the runner here. The runner registers itself with the
// active runners list in ClientCursor.
ClientCursor::registerRunner(runner.get());
runner->setYieldPolicy(Runner::YIELD_AUTO);
auto_ptr<DeregisterEvenIfUnderlyingCodeThrows> safety(
new DeregisterEvenIfUnderlyingCodeThrows(runner.get()));
BSONObj obj;
Runner::RunnerState state;
uint64_t numMisplacedDocs = 0;
// set this outside loop. we will need to use this both within loop and when deciding
// to fill in explain information
const bool isExplain = pq.isExplain();
while (Runner::RUNNER_ADVANCED == (state = runner->getNext(&obj, NULL))) {
// If we're sharded make sure that we don't return any data that hasn't been migrated
// off of our shared yet.
if (collMetadata) {
// This information can change if we yield and as such we must make sure to re-fetch
// it if we yield.
KeyPattern kp(collMetadata->getKeyPattern());
// This performs excessive BSONObj creation but that's OK for now.
if (!collMetadata->keyBelongsToMe(kp.extractSingleKey(obj))) {
++numMisplacedDocs;
continue;
}
}
// Add result to output buffer. This is unnecessary if explain info is requested
if (!isExplain) {
bb.appendBuf((void*)obj.objdata(), obj.objsize());
}
// Count the result.
++numResults;
// Possibly note slave's position in the oplog.
if (pq.hasOption(QueryOption_OplogReplay)) {
BSONElement e = obj["ts"];
if (Date == e.type() || Timestamp == e.type()) {
slaveReadTill = e._opTime();
}
}
// TODO: only one type of 2d search doesn't support this. We need a way to pull it out
// of CanonicalQuery. :(
const bool supportsGetMore = true;
if (isExplain) {
if (enoughForExplain(pq, numResults)) {
break;
}
}
else if (!supportsGetMore && (enough(pq, numResults)
|| bb.len() >= MaxBytesToReturnToClientAtOnce)) {
break;
}
else if (enoughForFirstBatch(pq, numResults, bb.len())) {
// If only one result requested assume it's a findOne() and don't save the cursor.
if (pq.wantMore() && 1 != pq.getNumToReturn()) {
saveClientCursor = true;
}
break;
}
}
// If we cache the runner later, we want to deregister it as it receives notifications
// anyway by virtue of being cached.
//
// If we don't cache the runner later, we are deleting it, so it must be deregistered.
//
// So, no matter what, deregister the runner.
safety.reset();
// Caller expects exceptions thrown in certain cases:
// * in-memory sort using too much RAM.
if (Runner::RUNNER_ERROR == state) {
uasserted(17144, "Runner error, memory limit for sort probably exceeded");
}
// Why save a dead runner?
if (Runner::RUNNER_DEAD == state) {
saveClientCursor = false;
}
else if (pq.hasOption(QueryOption_CursorTailable) && (1 != pq.getNumToReturn())) {
// If pq.hasOption(tailable) the only plan the planner will output is a collscan with
// tailable set.
saveClientCursor = true;
}
// TODO(greg): This will go away soon.
if (!shardingState.getVersion(pq.ns()).isWriteCompatibleWith(shardingVersionAtStart)) {
// if the version changed during the query we might be missing some data and its safe to
// send this as mongos can resend at this point
throw SendStaleConfigException(pq.ns(), "version changed during initial query",
shardingVersionAtStart,
shardingState.getVersion(pq.ns()));
}
// Append explain information to query results by asking the runner to produce them.
if (isExplain) {
TypeExplain* bareExplain;
Status res = runner->getExplainPlan(&bareExplain);
if (!res.isOK()) {
error() << "could not produce explain of query '" << pq.getFilter()
<< "', error: " << res.reason();
// If numResults and the data in bb don't correspond, we'll crash later when rooting
// through the reply msg.
BSONObj emptyObj;
bb.appendBuf((void*)emptyObj.objdata(), emptyObj.objsize());
// The explain output is actually a result.
numResults = 1;
// TODO: we can fill out millis etc. here just fine even if the plan screwed up.
}
else {
boost::scoped_ptr<TypeExplain> explain(bareExplain);
// Fill in the missing run-time fields in explain, starting with propeties of
// the process running the query.
std::string server = mongoutils::str::stream()
<< getHostNameCached() << ":" << serverGlobalParams.port;
explain->setServer(server);
// Fill in the number of documents consummed that were involved in an ongoing
// (or aborted) migration.
explain->setNChunkSkips(numMisplacedDocs);
// We might have skipped some results due to chunk migration etc. so our count is
// correct and explain's is not.
explain->setN(numResults);
// Clock the whole operation.
explain->setMillis(curop.elapsedMillis());
BSONObj explainObj = explain->toBSON();
bb.appendBuf((void*)explainObj.objdata(), explainObj.objsize());
// The explain output is actually a result.
numResults = 1;
}
}
long long ccId = 0;
if (saveClientCursor) {
// We won't use the runner until it's getMore'd.
runner->saveState();
// Allocate a new ClientCursor. We don't have to worry about leaking it as it's
// inserted into a global map by its ctor.
ClientCursor* cc = new ClientCursor(runner.get(), cq->getParsed().getOptions(),
cq->getParsed().getFilter());
ccId = cc->cursorid();
QLOG() << "caching runner with cursorid " << ccId
<< " after returning " << numResults << " results" << endl;
// ClientCursor takes ownership of runner. Release to make sure it's not deleted.
runner.release();
// TODO document
if (pq.hasOption(QueryOption_OplogReplay) && !slaveReadTill.isNull()) {
cc->slaveReadTill(slaveReadTill);
}
// TODO document
if (pq.hasOption(QueryOption_Exhaust)) {
curop.debug().exhaust = true;
}
// Set attributes for getMore.
cc->setCollMetadata(collMetadata);
cc->setPos(numResults);
// If the query had a time limit, remaining time is "rolled over" to the cursor (for
// use by future getmore ops).
cc->setLeftoverMaxTimeMicros(curop.getRemainingMaxTimeMicros());
}
// Add the results from the query into the output buffer.
result.appendData(bb.buf(), bb.len());
bb.decouple();
// Fill out the output buffer's header.
QueryResult* qr = static_cast<QueryResult*>(result.header());
qr->cursorId = ccId;
curop.debug().cursorid = (0 == ccId ? -1 : ccId);
qr->setResultFlagsToOk();
qr->setOperation(opReply);
qr->startingFrom = 0;
qr->nReturned = numResults;
// TODO: nscanned is bogus.
// curop.debug().nscanned = ( cursor ? cursor->nscanned() : 0LL );
curop.debug().ntoskip = pq.getSkip();
curop.debug().nreturned = numResults;
// curop.debug().exhaust is set above.
return curop.debug().exhaust ? pq.ns() : "";
}
static bool receivedQuery(Client& c, DbResponse& dbresponse, Message& m ) {
bool ok = true;
MSGID responseTo = m.header()->id;
DbMessage d(m);
QueryMessage q(d);
auto_ptr< Message > resp( new Message() );
CurOp& op = *(c.curop());
shared_ptr<AssertionException> ex;
try {
dbresponse.exhaust = runQuery(m, q, op, *resp);
assert( !resp->empty() );
}
catch ( SendStaleConfigException& e ){
ex.reset( new SendStaleConfigException( e.getns(), e.getInfo().msg ) );
ok = false;
}
catch ( AssertionException& e ) {
ex.reset( new AssertionException( e.getInfo().msg, e.getCode() ) );
ok = false;
}
if( ex ){
op.debug().exceptionInfo = ex->getInfo();
LOGWITHRATELIMIT {
log() << "assertion " << ex->toString() << " ns:" << q.ns << " query:" <<
(q.query.valid() ? q.query.toString() : "query object is corrupt") << endl;
if( q.ntoskip || q.ntoreturn )
log() << " ntoskip:" << q.ntoskip << " ntoreturn:" << q.ntoreturn << endl;
}
SendStaleConfigException* scex = NULL;
if ( ex->getCode() == SendStaleConfigCode ) scex = static_cast<SendStaleConfigException*>( ex.get() );
BSONObjBuilder err;
ex->getInfo().append( err );
if( scex ) err.append( "ns", scex->getns() );
BSONObj errObj = err.done();
log() << errObj << endl;
BufBuilder b;
b.skip(sizeof(QueryResult));
b.appendBuf((void*) errObj.objdata(), errObj.objsize());
// todo: call replyToQuery() from here instead of this!!! see dbmessage.h
QueryResult * msgdata = (QueryResult *) b.buf();
b.decouple();
QueryResult *qr = msgdata;
qr->_resultFlags() = ResultFlag_ErrSet;
if( scex ) qr->_resultFlags() |= ResultFlag_ShardConfigStale;
qr->len = b.len();
qr->setOperation(opReply);
qr->cursorId = 0;
qr->startingFrom = 0;
qr->nReturned = 1;
resp.reset( new Message() );
resp->setData( msgdata, true );
}
op.debug().responseLength = resp->header()->dataLen();
dbresponse.response = resp.release();
dbresponse.responseTo = responseTo;
return ok;
}
void operator()( DBClientCursorBatchIterator &i ) {
Lock::GlobalWrite lk;
context.relocked();
bool createdCollection = false;
Collection* collection = NULL;
while( i.moreInCurrentBatch() ) {
if ( numSeen % 128 == 127 /*yield some*/ ) {
collection = NULL;
time_t now = time(0);
if( now - lastLog >= 60 ) {
// report progress
if( lastLog )
log() << "clone " << to_collection << ' ' << numSeen << endl;
lastLog = now;
}
mayInterrupt( _mayBeInterrupted );
dbtempreleaseif t( _mayYield );
}
if ( isindex == false && collection == NULL ) {
collection = context.db()->getCollection( to_collection );
if ( !collection ) {
massert( 17321,
str::stream()
<< "collection dropped during clone ["
<< to_collection << "]",
!createdCollection );
createdCollection = true;
collection = context.db()->createCollection( txn, to_collection );
verify( collection );
}
}
BSONObj tmp = i.nextSafe();
/* assure object is valid. note this will slow us down a little. */
const Status status = validateBSON(tmp.objdata(), tmp.objsize());
if (!status.isOK()) {
out() << "Cloner: skipping corrupt object from " << from_collection
<< ": " << status.reason();
continue;
}
++numSeen;
BSONObj js = tmp;
if ( isindex ) {
verify(nsToCollectionSubstring(from_collection) == "system.indexes");
js = fixindex(context.db()->name(), tmp);
indexesToBuild->push_back( js.getOwned() );
continue;
}
verify(nsToCollectionSubstring(from_collection) != "system.indexes");
StatusWith<DiskLoc> loc = collection->insertDocument( txn, js, true );
if ( !loc.isOK() ) {
error() << "error: exception cloning object in " << from_collection
<< ' ' << loc.toString() << " obj:" << js;
}
uassertStatusOK( loc.getStatus() );
if ( logForRepl )
logOp(txn, "i", to_collection, js);
getDur().commitIfNeeded();
RARELY if ( time( 0 ) - saveLast > 60 ) {
log() << numSeen << " objects cloned so far from collection " << from_collection;
saveLast = time( 0 );
}
}
}
void Strategy::queryOp(OperationContext* txn, Request& request) {
verify(!NamespaceString(request.getns()).isCommand());
globalOpCounters.gotQuery();
QueryMessage q(request.d());
NamespaceString ns(q.ns);
ClientBasic* client = txn->getClient();
AuthorizationSession* authSession = AuthorizationSession::get(client);
Status status = authSession->checkAuthForFind(ns, false);
audit::logQueryAuthzCheck(client, ns, q.query, status.code());
uassertStatusOK(status);
LOG(3) << "query: " << q.ns << " " << q.query << " ntoreturn: " << q.ntoreturn
<< " options: " << q.queryOptions;
if (q.ntoreturn == 1 && strstr(q.ns, ".$cmd"))
throw UserException(8010, "something is wrong, shouldn't see a command here");
if (q.queryOptions & QueryOption_Exhaust) {
uasserted(18526,
string("the 'exhaust' query option is invalid for mongos queries: ") + q.ns +
" " + q.query.toString());
}
// Determine the default read preference mode based on the value of the slaveOk flag.
ReadPreference readPreferenceOption = (q.queryOptions & QueryOption_SlaveOk)
? ReadPreference::SecondaryPreferred
: ReadPreference::PrimaryOnly;
ReadPreferenceSetting readPreference(readPreferenceOption, TagSet());
BSONElement rpElem;
auto readPrefExtractStatus = bsonExtractTypedField(
q.query, LiteParsedQuery::kWrappedReadPrefField, mongo::Object, &rpElem);
if (readPrefExtractStatus.isOK()) {
auto parsedRps = ReadPreferenceSetting::fromBSON(rpElem.Obj());
uassertStatusOK(parsedRps.getStatus());
readPreference = parsedRps.getValue();
} else if (readPrefExtractStatus != ErrorCodes::NoSuchKey) {
uassertStatusOK(readPrefExtractStatus);
}
auto canonicalQuery = CanonicalQuery::canonicalize(q, ExtensionsCallbackNoop());
uassertStatusOK(canonicalQuery.getStatus());
// If the $explain flag was set, we must run the operation on the shards as an explain command
// rather than a find command.
if (canonicalQuery.getValue()->getParsed().isExplain()) {
const LiteParsedQuery& lpq = canonicalQuery.getValue()->getParsed();
BSONObj findCommand = lpq.asFindCommand();
// We default to allPlansExecution verbosity.
auto verbosity = ExplainCommon::EXEC_ALL_PLANS;
const bool secondaryOk = (readPreference.pref != ReadPreference::PrimaryOnly);
rpc::ServerSelectionMetadata metadata(secondaryOk, readPreference);
BSONObjBuilder explainBuilder;
uassertStatusOK(
Strategy::explainFind(txn, findCommand, lpq, verbosity, metadata, &explainBuilder));
BSONObj explainObj = explainBuilder.done();
replyToQuery(0, // query result flags
request.p(),
request.m(),
static_cast<const void*>(explainObj.objdata()),
explainObj.objsize(),
1, // numResults
0, // startingFrom
CursorId(0));
return;
}
// Do the work to generate the first batch of results. This blocks waiting to get responses from
// the shard(s).
std::vector<BSONObj> batch;
// 0 means the cursor is exhausted. Otherwise we assume that a cursor with the returned id can
// be retrieved via the ClusterCursorManager.
auto cursorId = ClusterFind::runQuery(txn, *canonicalQuery.getValue(), readPreference, &batch);
uassertStatusOK(cursorId.getStatus());
// Fill out the response buffer.
int numResults = 0;
OpQueryReplyBuilder reply;
for (auto&& obj : batch) {
obj.appendSelfToBufBuilder(reply.bufBuilderForResults());
numResults++;
}
reply.send(request.p(),
0, // query result flags
request.m(),
numResults,
0, // startingFrom
cursorId.getValue());
}
/* note: this is only (as-is) called for
- not multi
- not mods is indexed
- not upsert
*/
static UpdateResult _updateById(bool isOperatorUpdate,
int idIdxNo,
ModSet* mods,
NamespaceDetails* d,
NamespaceDetailsTransient *nsdt,
bool su,
const char* ns,
const BSONObj& updateobj,
BSONObj patternOrig,
bool logop,
OpDebug& debug,
bool fromMigrate = false) {
DiskLoc loc;
{
IndexDetails& i = d->idx(idIdxNo);
BSONObj key = i.getKeyFromQuery( patternOrig );
loc = QueryRunner::fastFindSingle(i, key);
if( loc.isNull() ) {
// no upsert support in _updateById yet, so we are done.
return UpdateResult( 0 , 0 , 0 , BSONObj() );
}
}
Record* r = loc.rec();
if ( cc().allowedToThrowPageFaultException() && ! r->likelyInPhysicalMemory() ) {
throw PageFaultException( r );
}
/* 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. */
BSONObj newObj;
if ( isOperatorUpdate ) {
const BSONObj& onDisk = loc.obj();
auto_ptr<ModSetState> mss = mods->prepare( onDisk, false /* not an insertion */ );
if( mss->canApplyInPlace() ) {
mss->applyModsInPlace(true);
debug.fastmod = true;
DEBUGUPDATE( "\t\t\t updateById doing in place update" );
newObj = onDisk;
}
else {
newObj = mss->createNewFromMods();
checkTooLarge(newObj);
verify(nsdt);
theDataFileMgr.updateRecord(ns, d, nsdt, r, loc , newObj.objdata(), newObj.objsize(), debug);
}
if ( logop ) {
DEV verify( mods->size() );
BSONObj pattern = patternOrig;
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 );
}
}
return UpdateResult( 1 , 1 , 1 , BSONObj() );
} // end $operator update
// regular update
BSONElementManipulator::lookForTimestamps( updateobj );
checkNoMods( updateobj );
verify(nsdt);
theDataFileMgr.updateRecord(ns, d, nsdt, r, loc , updateobj.objdata(), updateobj.objsize(), debug );
if ( logop ) {
logOp("u", ns, updateobj, &patternOrig, 0, fromMigrate, &updateobj );
}
return UpdateResult( 1 , 0 , 1 , BSONObj() );
}
// PD_TRACE_DECLARE_FUNCTION ( SDB__DMSROUNIT_INSRCD, "_dmsReorgUnit::insertRecord" )
INT32 _dmsReorgUnit::insertRecord ( BSONObj &obj,
_pmdEDUCB *cb, UINT32 attributes )
{
INT32 rc = SDB_OK ;
PD_TRACE_ENTRY ( SDB__DMSROUNIT_INSRCD );
UINT32 dmsrecordSize = 0 ;
ossValuePtr recordPtr = 0 ;
ossValuePtr prevPtr = 0 ;
dmsOffset offset = DMS_INVALID_OFFSET ;
dmsOffset recordOffset = DMS_INVALID_OFFSET ;
dmsExtent *currentExtent = (dmsExtent*)_pCurrentExtent ;
BOOLEAN isCompressed = FALSE ;
const CHAR *compressedData = NULL ;
INT32 compressedDataSize = 0 ;
if ( obj.objsize() + DMS_RECORD_METADATA_SZ >
DMS_RECORD_MAX_SZ )
{
rc = SDB_CORRUPTED_RECORD ;
goto error ;
}
if ( OSS_BIT_TEST ( attributes, DMS_MB_ATTR_COMPRESSED ) )
{
rc = dmsCompress ( cb, obj, NULL, 0, &compressedData,
&compressedDataSize ) ;
PD_RC_CHECK ( rc, PDERROR,
"Failed to compress record, rc = %d: %s",
rc, obj.toString().c_str() ) ;
dmsrecordSize = compressedDataSize + sizeof(INT32) ;
if ( dmsrecordSize > (UINT32)(obj.objsize()) )
{
dmsrecordSize = obj.objsize() ;
}
else
{
isCompressed = TRUE ;
}
}
else
{
dmsrecordSize = obj.objsize() ;
}
dmsrecordSize += DMS_RECORD_METADATA_SZ ;
dmsrecordSize *= DMS_RECORD_OVERFLOW_RATIO ;
dmsrecordSize = OSS_MIN(DMS_RECORD_MAX_SZ, ossAlignX(dmsrecordSize,4)) ;
alloc:
if ( !_pCurrentExtent )
{
rc = _allocateExtent ( dmsrecordSize <<
DMS_RECORDS_PER_EXTENT_SQUARE ) ;
if ( rc )
{
PD_LOG ( PDERROR, "Failed to allocate new extent in reorg file, "
"rc = %d", rc ) ;
goto error ;
}
currentExtent = (dmsExtent*)_pCurrentExtent ;
}
if ( dmsrecordSize > (UINT32)currentExtent->_freeSpace )
{
rc = _flushExtent () ;
if ( rc )
{
PD_LOG ( PDERROR, "Failed to flush extent, rc = %d", rc ) ;
goto error ;
}
goto alloc ;
}
recordOffset = _currentExtentSize - currentExtent->_freeSpace ;
recordPtr = ((ossValuePtr)currentExtent) + recordOffset ;
if ( currentExtent->_freeSpace - (INT32)dmsrecordSize <
(INT32)DMS_MIN_RECORD_SZ &&
currentExtent->_freeSpace <= (INT32)DMS_RECORD_MAX_SZ )
{
dmsrecordSize = (UINT32)currentExtent->_freeSpace ;
}
DMS_RECORD_SETSTATE ( recordPtr, DMS_RECORD_FLAG_NORMAL ) ;
DMS_RECORD_RESETATTR ( recordPtr ) ;
DMS_RECORD_SETMYOFFSET ( recordPtr, recordOffset ) ;
DMS_RECORD_SETSIZE ( recordPtr, dmsrecordSize ) ;
if ( isCompressed )
{
DMS_RECORD_SETATTR ( recordPtr, DMS_RECORD_FLAG_COMPRESSED ) ;
DMS_RECORD_SETDATA ( recordPtr, compressedData, compressedDataSize ) ;
}
else
{
DMS_RECORD_SETDATA ( recordPtr, obj.objdata(), obj.objsize() ) ;
}
DMS_RECORD_SETNEXTOFFSET ( recordPtr, DMS_INVALID_OFFSET ) ;
DMS_RECORD_SETPREVOFFSET ( recordPtr, DMS_INVALID_OFFSET ) ;
currentExtent->_recCount ++ ;
currentExtent->_freeSpace -= dmsrecordSize ;
offset = currentExtent->_lastRecordOffset ;
if ( DMS_INVALID_OFFSET != offset )
{
prevPtr = ((ossValuePtr)currentExtent) + offset ;
DMS_RECORD_SETNEXTOFFSET ( prevPtr, recordOffset ) ;
DMS_RECORD_SETPREVOFFSET ( recordPtr, offset ) ;
}
currentExtent->_lastRecordOffset = recordOffset ;
offset = currentExtent->_firstRecordOffset ;
if ( DMS_INVALID_OFFSET == offset )
{
currentExtent->_firstRecordOffset = recordOffset ;
}
done :
PD_TRACE_EXITRC ( SDB__DMSROUNIT_INSRCD, rc );
return rc ;
error :
goto done ;
}
INT32 catMainController::_processMsg( const NET_HANDLE &handle,
MsgHeader *pMsg )
{
INT32 rc = SDB_OK ;
switch ( pMsg->opCode )
{
case MSG_BS_QUERY_REQ:
{
rc = _processQueryMsg( handle, pMsg );
break;
}
case MSG_BS_GETMORE_REQ :
{
rc = _processGetMoreMsg( handle, pMsg ) ;
break ;
}
case MSG_BS_KILL_CONTEXT_REQ:
{
rc = _processKillContext( handle, pMsg ) ;
break;
}
case MSG_BS_INTERRUPTE :
{
rc = _processInterruptMsg( handle, pMsg ) ;
break ;
}
case MSG_BS_DISCONNECT :
{
rc = _processDisconnectMsg( handle, pMsg ) ;
break ;
}
case MSG_CAT_QUERY_DATA_GRP_REQ :
{
rc = _processQueryDataGrp( handle, pMsg ) ;
break ;
}
case MSG_CAT_QUERY_COLLECTIONS_REQ :
{
rc = _processQueryCollections( handle, pMsg ) ;
break ;
}
case MSG_CAT_QUERY_COLLECTIONSPACES_REQ :
{
rc = _processQueryCollectionSpaces ( handle, pMsg ) ;
break ;
}
case MSG_AUTH_VERIFY_REQ :
{
rc = _processAuthenticate( handle, pMsg ) ;
break ;
}
case MSG_AUTH_CRTUSR_REQ :
{
_pCatCB->getCatDCMgr()->setImageCommand( TRUE ) ;
rc = _processAuthCrt( handle, pMsg ) ;
break ;
}
case MSG_AUTH_DELUSR_REQ :
{
_pCatCB->getCatDCMgr()->setImageCommand( TRUE ) ;
rc = _processAuthDel( handle, pMsg ) ;
break ;
}
case MSG_COOR_CHECK_ROUTEID_REQ :
{
rc = _processCheckRouteID( handle, pMsg ) ;
break;
}
default :
{
PD_LOG( PDERROR, "Recieve unknow msg[opCode:(%d)%d, len: %d, "
"tid: %d, reqID: %lld, nodeID: %u.%u.%u]",
IS_REPLY_TYPE(pMsg->opCode), GET_REQUEST_TYPE(pMsg->opCode),
pMsg->messageLength, pMsg->TID, pMsg->requestID,
pMsg->routeID.columns.groupID, pMsg->routeID.columns.nodeID,
pMsg->routeID.columns.serviceID ) ;
rc = SDB_UNKNOWN_MESSAGE ;
BSONObj err = utilGetErrorBson( rc, _pEDUCB->getInfo(
EDU_INFO_ERROR ) ) ;
MsgOpReply reply ;
reply.header.opCode = MAKE_REPLY_TYPE( pMsg->opCode ) ;
reply.header.messageLength = sizeof( MsgOpReply ) + err.objsize() ;
reply.header.requestID = pMsg->requestID ;
reply.header.routeID.value = 0 ;
reply.header.TID = pMsg->TID ;
reply.flags = rc ;
reply.contextID = -1 ;
reply.numReturned = 1 ;
reply.startFrom = 0 ;
_pCatCB->netWork()->syncSend( handle, (MsgHeader*)&reply,
(void*)err.objdata(),
err.objsize() ) ;
break ;
}
}
if ( rc && SDB_UNKNOWN_MESSAGE != rc )
{
PD_LOG( PDWARNING, "Process msg[opCode:(%d)%d, len: %d, tid: %d, "
"reqID: %lld, nodeID: %u.%u.%u] failed, rc: %d",
IS_REPLY_TYPE(pMsg->opCode), GET_REQUEST_TYPE(pMsg->opCode),
pMsg->messageLength, pMsg->TID, pMsg->requestID,
pMsg->routeID.columns.groupID, pMsg->routeID.columns.nodeID,
pMsg->routeID.columns.serviceID, rc ) ;
}
return rc ;
}
std::string runQuery(OperationContext* opCtx,
QueryMessage& q,
const NamespaceString& nss,
Message& result) {
CurOp& curOp = *CurOp::get(opCtx);
curOp.ensureStarted();
uassert(ErrorCodes::InvalidNamespace,
str::stream() << "Invalid ns [" << nss.ns() << "]",
nss.isValid());
invariant(!nss.isCommand());
// Set CurOp information.
const auto upconvertedQuery = upconvertQueryEntry(q.query, nss, q.ntoreturn, q.ntoskip);
beginQueryOp(opCtx, nss, upconvertedQuery, q.ntoreturn, q.ntoskip);
// Parse the qm into a CanonicalQuery.
const boost::intrusive_ptr<ExpressionContext> expCtx;
auto cq = uassertStatusOKWithContext(
CanonicalQuery::canonicalize(opCtx,
q,
expCtx,
ExtensionsCallbackReal(opCtx, &nss),
MatchExpressionParser::kAllowAllSpecialFeatures),
"Can't canonicalize query");
invariant(cq.get());
LOG(5) << "Running query:\n" << redact(cq->toString());
LOG(2) << "Running query: " << redact(cq->toStringShort());
// Parse, canonicalize, plan, transcribe, and get a plan executor.
AutoGetCollectionForReadCommand ctx(opCtx, nss, AutoGetCollection::ViewMode::kViewsForbidden);
Collection* const collection = ctx.getCollection();
{
const QueryRequest& qr = cq->getQueryRequest();
// Allow the query to run on secondaries if the read preference permits it. If no read
// preference was specified, allow the query to run iff slaveOk has been set.
const bool slaveOK = qr.hasReadPref()
? uassertStatusOK(ReadPreferenceSetting::fromContainingBSON(q.query))
.canRunOnSecondary()
: qr.isSlaveOk();
uassertStatusOK(
repl::ReplicationCoordinator::get(opCtx)->checkCanServeReadsFor(opCtx, nss, slaveOK));
}
// We have a parsed query. Time to get the execution plan for it.
auto exec = uassertStatusOK(getExecutorLegacyFind(opCtx, collection, nss, std::move(cq)));
const QueryRequest& qr = exec->getCanonicalQuery()->getQueryRequest();
// If it's actually an explain, do the explain and return rather than falling through
// to the normal query execution loop.
if (qr.isExplain()) {
BufBuilder bb;
bb.skip(sizeof(QueryResult::Value));
BSONObjBuilder explainBob;
Explain::explainStages(
exec.get(), collection, ExplainOptions::Verbosity::kExecAllPlans, &explainBob);
// Add the resulting object to the return buffer.
BSONObj explainObj = explainBob.obj();
bb.appendBuf((void*)explainObj.objdata(), explainObj.objsize());
// Set query result fields.
QueryResult::View qr = bb.buf();
qr.setResultFlagsToOk();
qr.msgdata().setLen(bb.len());
curOp.debug().responseLength = bb.len();
qr.msgdata().setOperation(opReply);
qr.setCursorId(0);
qr.setStartingFrom(0);
qr.setNReturned(1);
result.setData(bb.release());
return "";
}
// Handle query option $maxTimeMS (not used with commands).
if (qr.getMaxTimeMS() > 0) {
uassert(40116,
"Illegal attempt to set operation deadline within DBDirectClient",
!opCtx->getClient()->isInDirectClient());
opCtx->setDeadlineAfterNowBy(Milliseconds{qr.getMaxTimeMS()});
}
opCtx->checkForInterrupt(); // May trigger maxTimeAlwaysTimeOut fail point.
// Run the query.
// bb is used to hold query results
// this buffer should contain either requested documents per query or
// explain information, but not both
BufBuilder bb(FindCommon::kInitReplyBufferSize);
bb.skip(sizeof(QueryResult::Value));
// How many results have we obtained from the executor?
int numResults = 0;
BSONObj obj;
PlanExecutor::ExecState state;
// Get summary info about which plan the executor is using.
{
stdx::lock_guard<Client> lk(*opCtx->getClient());
curOp.setPlanSummary_inlock(Explain::getPlanSummary(exec.get()));
}
while (PlanExecutor::ADVANCED == (state = exec->getNext(&obj, NULL))) {
// If we can't fit this result inside the current batch, then we stash it for later.
if (!FindCommon::haveSpaceForNext(obj, numResults, bb.len())) {
exec->enqueue(obj);
break;
}
// Add result to output buffer.
bb.appendBuf((void*)obj.objdata(), obj.objsize());
// Count the result.
++numResults;
if (FindCommon::enoughForFirstBatch(qr, numResults)) {
LOG(5) << "Enough for first batch, wantMore=" << qr.wantMore()
<< " ntoreturn=" << qr.getNToReturn().value_or(0)
<< " numResults=" << numResults;
break;
}
}
// Caller expects exceptions thrown in certain cases.
if (PlanExecutor::FAILURE == state || PlanExecutor::DEAD == state) {
error() << "Plan executor error during find: " << PlanExecutor::statestr(state)
<< ", stats: " << redact(Explain::getWinningPlanStats(exec.get()));
uassertStatusOKWithContext(WorkingSetCommon::getMemberObjectStatus(obj),
"Executor error during OP_QUERY find");
MONGO_UNREACHABLE;
}
// Before saving the cursor, ensure that whatever plan we established happened with the expected
// collection version
auto css = CollectionShardingState::get(opCtx, nss);
css->checkShardVersionOrThrow(opCtx);
// Fill out CurOp based on query results. If we have a cursorid, we will fill out CurOp with
// this cursorid later.
long long ccId = 0;
if (shouldSaveCursor(opCtx, collection, state, exec.get())) {
// We won't use the executor until it's getMore'd.
exec->saveState();
exec->detachFromOperationContext();
// Allocate a new ClientCursor and register it with the cursor manager.
ClientCursorPin pinnedCursor = collection->getCursorManager()->registerCursor(
opCtx,
{std::move(exec),
nss,
AuthorizationSession::get(opCtx->getClient())->getAuthenticatedUserNames(),
opCtx->recoveryUnit()->getReadConcernLevel(),
upconvertedQuery});
ccId = pinnedCursor.getCursor()->cursorid();
LOG(5) << "caching executor with cursorid " << ccId << " after returning " << numResults
<< " results";
// TODO document
if (qr.isExhaust()) {
curOp.debug().exhaust = true;
}
pinnedCursor.getCursor()->setPos(numResults);
// We assume that cursors created through a DBDirectClient are always used from their
// original OperationContext, so we do not need to move time to and from the cursor.
if (!opCtx->getClient()->isInDirectClient()) {
// If the query had a time limit, remaining time is "rolled over" to the cursor (for
// use by future getmore ops).
pinnedCursor.getCursor()->setLeftoverMaxTimeMicros(opCtx->getRemainingMaxTimeMicros());
}
endQueryOp(opCtx, collection, *pinnedCursor.getCursor()->getExecutor(), numResults, ccId);
} else {
LOG(5) << "Not caching executor but returning " << numResults << " results.";
endQueryOp(opCtx, collection, *exec, numResults, ccId);
}
// Fill out the output buffer's header.
QueryResult::View queryResultView = bb.buf();
queryResultView.setCursorId(ccId);
queryResultView.setResultFlagsToOk();
queryResultView.msgdata().setLen(bb.len());
queryResultView.msgdata().setOperation(opReply);
queryResultView.setStartingFrom(0);
queryResultView.setNReturned(numResults);
// Add the results from the query into the output buffer.
result.setData(bb.release());
// curOp.debug().exhaust is set above.
return curOp.debug().exhaust ? nss.ns() : "";
}
// PD_TRACE_DECLARE_FUNCTION ( SDB__DMSSTORAGELOADEXT__IMPRTBLOCK, "dmsStorageLoadOp::pushToTempDataBlock" )
INT32 dmsStorageLoadOp::pushToTempDataBlock ( dmsMBContext *mbContext,
pmdEDUCB *cb,
BSONObj &record,
BOOLEAN isLast,
BOOLEAN isAsynchr )
{
INT32 rc = SDB_OK ;
PD_TRACE_ENTRY ( SDB__DMSSTORAGELOADEXT__IMPRTBLOCK );
UINT32 dmsrecordSize = 0 ;
dmsRecord *pRecord = NULL ;
dmsRecord *pPreRecord = NULL ;
dmsOffset offset = DMS_INVALID_OFFSET ;
dmsOffset recordOffset = DMS_INVALID_OFFSET ;
_IDToInsert oid ;
idToInsertEle oidEle((CHAR*)(&oid)) ;
CHAR *pNewRecordData = NULL ;
dmsRecordData recordData ;
dmsCompressorEntry *compressorEntry = NULL ;
SDB_ASSERT( mbContext, "mb context can't be NULL" ) ;
compressorEntry = _su->data()->getCompressorEntry( mbContext->mbID() ) ;
/* For concurrency protection with drop CL and set compresor. */
dmsCompressorGuard compGuard( compressorEntry, SHARED ) ;
try
{
recordData.setData( record.objdata(), record.objsize(),
FALSE, TRUE ) ;
/* (0) */
BSONElement ele = record.getField ( DMS_ID_KEY_NAME ) ;
const CHAR *pCheckErr = "" ;
if ( !dmsIsRecordIDValid( ele, TRUE, &pCheckErr ) )
{
PD_LOG( PDERROR, "Record[%s] _id is error: %s",
record.toString().c_str(), pCheckErr ) ;
rc = SDB_INVALIDARG ;
goto error ;
}
if ( ele.eoo() )
{
oid._oid.init() ;
rc = cb->allocBuff( oidEle.size() + record.objsize(),
&pNewRecordData ) ;
if ( rc )
{
PD_LOG( PDERROR, "Alloc memory[size:%u] failed, rc: %d",
oidEle.size() + record.objsize(), rc ) ;
goto error ;
}
*(UINT32*)pNewRecordData = oidEle.size() + record.objsize() ;
ossMemcpy( pNewRecordData + sizeof(UINT32), oidEle.rawdata(),
oidEle.size() ) ;
ossMemcpy( pNewRecordData + sizeof(UINT32) + oidEle.size(),
record.objdata() + sizeof(UINT32),
record.objsize() - sizeof(UINT32) ) ;
recordData.setData( pNewRecordData,
oidEle.size() + record.objsize(),
FALSE, TRUE ) ;
record = BSONObj( pNewRecordData ) ;
}
dmsrecordSize = recordData.len() ;
if ( recordData.len() + DMS_RECORD_METADATA_SZ >
DMS_RECORD_USER_MAX_SZ )
{
rc = SDB_DMS_RECORD_TOO_BIG ;
goto error ;
}
if ( compressorEntry->ready() )
{
const CHAR *compressedData = NULL ;
INT32 compressedDataSize = 0 ;
UINT8 compressRatio = 0 ;
rc = dmsCompress( cb, compressorEntry,
recordData.data(), recordData.len(),
&compressedData, &compressedDataSize,
compressRatio ) ;
if ( SDB_OK == rc &&
compressedDataSize + sizeof(UINT32) < recordData.orgLen() &&
compressRatio < DMS_COMPRESS_RATIO_THRESHOLD )
{
dmsrecordSize = compressedDataSize + sizeof(UINT32) ;
recordData.setData( compressedData, compressedDataSize,
TRUE, FALSE ) ;
}
else if ( rc )
{
if ( SDB_UTIL_COMPRESS_ABORT == rc )
{
PD_LOG( PDINFO, "Record compression aborted. "
"Insert the original data. rc: %d", rc ) ;
}
else
{
PD_LOG( PDWARNING, "Record compression failed. "
"Insert the original data. rc: %d", rc ) ;
}
rc = SDB_OK ;
}
}
/*
* Release the guard to avoid deadlock with truncate/drop collection.
*/
compGuard.release() ;
dmsrecordSize *= DMS_RECORD_OVERFLOW_RATIO ;
dmsrecordSize += DMS_RECORD_METADATA_SZ ;
dmsrecordSize = OSS_MIN( DMS_RECORD_MAX_SZ,
ossAlignX ( dmsrecordSize, 4 ) ) ;
INT32 expandSize = dmsrecordSize << DMS_RECORDS_PER_EXTENT_SQUARE ;
if ( expandSize > DMS_BEST_UP_EXTENT_SZ )
{
expandSize = expandSize < DMS_BEST_UP_EXTENT_SZ ?
DMS_BEST_UP_EXTENT_SZ : expandSize ;
}
if ( !_pCurrentExtent )
{
rc = _allocateExtent ( expandSize ) ;
if ( rc )
{
PD_LOG ( PDERROR, "Failed to allocate new extent in reorg file, "
"rc = %d", rc ) ;
goto error ;
}
_currentExtent = (dmsExtent*)_pCurrentExtent ;
}
if ( dmsrecordSize > (UINT32)_currentExtent->_freeSpace || isLast )
{
rc = mbContext->mbLock( EXCLUSIVE ) ;
if ( rc )
{
PD_LOG ( PDERROR, "Failed to lock collection, rc=%d", rc ) ;
goto error ;
}
if ( !isAsynchr )
{
_currentExtent->_firstRecordOffset = DMS_INVALID_OFFSET ;
_currentExtent->_lastRecordOffset = DMS_INVALID_OFFSET ;
}
rc = _su->loadExtentA( mbContext, _pCurrentExtent,
_currentExtentSize / _pageSize,
TRUE ) ;
mbContext->mbUnlock() ;
if ( rc )
{
PD_LOG ( PDERROR, "Failed to load extent, rc = %d", rc ) ;
goto error ;
}
if ( isLast )
{
goto done ;
}
rc = _allocateExtent ( expandSize ) ;
if ( rc )
{
PD_LOG ( PDERROR, "Failed to allocate new extent in reorg file, "
"rc = %d", rc ) ;
goto error ;
}
}
recordOffset = _currentExtentSize - _currentExtent->_freeSpace ;
pRecord = ( dmsRecord* )( (const CHAR*)_currentExtent + recordOffset ) ;
if ( _currentExtent->_freeSpace - (INT32)dmsrecordSize <
(INT32)DMS_MIN_RECORD_SZ &&
_currentExtent->_freeSpace <= (INT32)DMS_RECORD_MAX_SZ )
{
dmsrecordSize = _currentExtent->_freeSpace ;
}
pRecord->setNormal() ;
pRecord->setMyOffset( recordOffset ) ;
pRecord->setSize( dmsrecordSize ) ;
pRecord->setData( recordData ) ;
pRecord->setNextOffset( DMS_INVALID_OFFSET ) ;
pRecord->setPrevOffset( DMS_INVALID_OFFSET ) ;
if ( isAsynchr )
{
_currentExtent->_recCount++ ;
}
_currentExtent->_freeSpace -= dmsrecordSize ;
offset = _currentExtent->_lastRecordOffset ;
if ( DMS_INVALID_OFFSET != offset )
{
pPreRecord = (dmsRecord*)( (const CHAR*)_currentExtent + offset ) ;
pPreRecord->setNextOffset( recordOffset ) ;
pRecord->setPrevOffset( offset ) ;
}
_currentExtent->_lastRecordOffset = recordOffset ;
offset = _currentExtent->_firstRecordOffset ;
if ( DMS_INVALID_OFFSET == offset )
{
_currentExtent->_firstRecordOffset = recordOffset ;
}
}
catch( std::exception &e )
{
PD_LOG( PDERROR, "Occur exception: %s", e.what() ) ;
rc = SDB_SYS ;
goto error ;
}
done:
PD_TRACE_EXITRC ( SDB__DMSSTORAGELOADEXT__IMPRTBLOCK, rc );
return rc ;
error:
goto done ;
}
TEST(RocksRecordStoreTest, OplogHack) {
RocksRecordStoreHarnessHelper harnessHelper;
scoped_ptr<RecordStore> rs(harnessHelper.newNonCappedRecordStore("local.oplog.foo"));
{
scoped_ptr<OperationContext> opCtx(harnessHelper.newOperationContext());
// always illegal
ASSERT_EQ(insertBSON(opCtx, rs, Timestamp(2,-1)).getStatus(),
ErrorCodes::BadValue);
{
BSONObj obj = BSON("not_ts" << Timestamp(2,1));
ASSERT_EQ(rs->insertRecord(opCtx.get(), obj.objdata(), obj.objsize(),
false ).getStatus(),
ErrorCodes::BadValue);
obj = BSON( "ts" << "not an Timestamp" );
ASSERT_EQ(rs->insertRecord(opCtx.get(), obj.objdata(), obj.objsize(),
false ).getStatus(),
ErrorCodes::BadValue);
}
// currently dasserts
// ASSERT_EQ(insertBSON(opCtx, rs, BSON("ts" << Timestamp(-2,1))).getStatus(),
// ErrorCodes::BadValue);
// success cases
ASSERT_EQ(insertBSON(opCtx, rs, Timestamp(1,1)).getValue(),
RecordId(1,1));
ASSERT_EQ(insertBSON(opCtx, rs, Timestamp(1,2)).getValue(),
RecordId(1,2));
ASSERT_EQ(insertBSON(opCtx, rs, Timestamp(2,2)).getValue(),
RecordId(2,2));
}
{
scoped_ptr<OperationContext> opCtx(harnessHelper.newOperationContext());
// find start
ASSERT_EQ(rs->oplogStartHack(opCtx.get(), RecordId(0,1)), RecordId()); // nothing <=
ASSERT_EQ(rs->oplogStartHack(opCtx.get(), RecordId(2,1)), RecordId(1,2)); // between
ASSERT_EQ(rs->oplogStartHack(opCtx.get(), RecordId(2,2)), RecordId(2,2)); // ==
ASSERT_EQ(rs->oplogStartHack(opCtx.get(), RecordId(2,3)), RecordId(2,2)); // > highest
}
{
scoped_ptr<OperationContext> opCtx(harnessHelper.newOperationContext());
rs->temp_cappedTruncateAfter(opCtx.get(), RecordId(2,2), false); // no-op
}
{
scoped_ptr<OperationContext> opCtx(harnessHelper.newOperationContext());
ASSERT_EQ(rs->oplogStartHack(opCtx.get(), RecordId(2,3)), RecordId(2,2));
}
{
scoped_ptr<OperationContext> opCtx(harnessHelper.newOperationContext());
rs->temp_cappedTruncateAfter(opCtx.get(), RecordId(1,2), false); // deletes 2,2
}
{
scoped_ptr<OperationContext> opCtx(harnessHelper.newOperationContext());
ASSERT_EQ(rs->oplogStartHack(opCtx.get(), RecordId(2,3)), RecordId(1,2));
}
{
scoped_ptr<OperationContext> opCtx(harnessHelper.newOperationContext());
rs->temp_cappedTruncateAfter(opCtx.get(), RecordId(1,2), true); // deletes 1,2
}
{
scoped_ptr<OperationContext> opCtx(harnessHelper.newOperationContext());
ASSERT_EQ(rs->oplogStartHack(opCtx.get(), RecordId(2,3)), RecordId(1,1));
}
{
scoped_ptr<OperationContext> opCtx(harnessHelper.newOperationContext());
WriteUnitOfWork wuow(opCtx.get());
ASSERT_OK(rs->truncate(opCtx.get())); // deletes 1,1 and leaves collection empty
wuow.commit();
}
{
scoped_ptr<OperationContext> opCtx(harnessHelper.newOperationContext());
ASSERT_EQ(rs->oplogStartHack(opCtx.get(), RecordId(2,3)), RecordId());
}
}
/**
* This is called by db/ops/query.cpp. This is the entry point for answering a query.
*/
string newRunQuery(Message& m, QueryMessage& q, CurOp& curop, Message &result) {
log() << "Running query on new system: " << q.query.toString() << endl;
// This is a read lock.
Client::ReadContext ctx(q.ns, dbpath);
// Parse, canonicalize, plan, transcribe, and get a runner.
Runner* rawRunner;
CanonicalQuery* cq;
Status status = getRunner(q, &rawRunner, &cq);
if (!status.isOK()) {
uasserted(17007, "Couldn't process query " + q.query.toString()
+ " why: " + status.reason());
}
verify(NULL != rawRunner);
auto_ptr<Runner> runner(rawRunner);
// We freak out later if this changes before we're done with the query.
const ChunkVersion shardingVersionAtStart = shardingState.getVersion(q.ns);
// We use this a lot below.
const LiteParsedQuery& pq = cq->getParsed();
// TODO: Document why we do this.
// TODO: do this when we can pass in our own parsed query
//replVerifyReadsOk(&pq);
// If this exists, the collection is sharded.
// If it doesn't exist, we can assume we're not sharded.
// If we're sharded, we might encounter data that is not consistent with our sharding state.
// We must ignore this data.
CollectionMetadataPtr collMetadata;
if (!shardingState.needCollectionMetadata(pq.ns())) {
collMetadata = CollectionMetadataPtr();
}
else {
collMetadata = shardingState.getCollectionMetadata(pq.ns());
}
// Run the query.
BufBuilder bb(32768);
bb.skip(sizeof(QueryResult));
// How many results have we obtained from the runner?
int numResults = 0;
// If we're replaying the oplog, we save the last time that we read.
OpTime slaveReadTill;
// Do we save the Runner in a ClientCursor for getMore calls later?
bool saveClientCursor = false;
// We turn on auto-yielding for the runner here, so we must register it with the active
// runners list in ClientCursor.
ClientCursor::registerRunner(runner.get());
runner->setYieldPolicy(Runner::YIELD_AUTO);
BSONObj obj;
Runner::RunnerState state;
while (Runner::RUNNER_ADVANCED == (state = runner->getNext(&obj, NULL))) {
// If we're sharded make sure that we don't return any data that hasn't been migrated
// off of our shared yet.
if (collMetadata) {
// This information can change if we yield and as such we must make sure to re-fetch
// it if we yield.
KeyPattern kp(collMetadata->getKeyPattern());
// This performs excessive BSONObj creation but that's OK for now.
if (!collMetadata->keyBelongsToMe(kp.extractSingleKey(obj))) { continue; }
}
// Add result to output buffer.
bb.appendBuf((void*)obj.objdata(), obj.objsize());
// Count the result.
++numResults;
// Possibly note slave's position in the oplog.
if (pq.hasOption(QueryOption_OplogReplay)) {
BSONElement e = obj["ts"];
if (Date == e.type() || Timestamp == e.type()) {
slaveReadTill = e._opTime();
}
}
// TODO: only one type of 2d search doesn't support this. We need a way to pull it out
// of CanonicalQuery. :(
const bool supportsGetMore = true;
const bool isExplain = pq.isExplain();
if (isExplain && enoughForExplain(pq, numResults)) {
break;
}
else if (!supportsGetMore && (enough(pq, numResults)
|| bb.len() >= MaxBytesToReturnToClientAtOnce)) {
break;
}
else if (enoughForFirstBatch(pq, numResults, bb.len())) {
// If only one result requested assume it's a findOne() and don't save the cursor.
if (pq.wantMore() && 1 != pq.getNumToReturn()) {
saveClientCursor = true;
}
break;
}
}
// If we cache the runner later, we want to deregister it as it receives notifications
// anyway by virtue of being cached.
//
// If we don't cache the runner later, we are deleting it, so it must be deregistered.
//
// So, no matter what, deregister the runner.
ClientCursor::deregisterRunner(runner.get());
// Why save a dead runner?
if (Runner::RUNNER_DEAD == state) { saveClientCursor = false; }
// TODO: Stage creation can set tailable depending on what's in the parsed query. We have
// the full parsed query available during planning...set it there.
//
// TODO: If we're tailable we want to save the client cursor. Make sure we do this later.
//if (pq.hasOption(QueryOption_CursorTailable) && pq.getNumToReturn() != 1) { ... }
// TODO(greg): This will go away soon.
if (!shardingState.getVersion(pq.ns()).isWriteCompatibleWith(shardingVersionAtStart)) {
// if the version changed during the query we might be missing some data and its safe to
// send this as mongos can resend at this point
throw SendStaleConfigException(pq.ns(), "version changed during initial query",
shardingVersionAtStart,
shardingState.getVersion(pq.ns()));
}
long long ccId = 0;
if (saveClientCursor) {
// We won't use the runner until it's getMore'd.
runner->saveState();
// Allocate a new ClientCursor. We don't have to worry about leaking it as it's
// inserted into a global map by its ctor.
ClientCursor* cc = new ClientCursor(runner.get(), cq->getParsed().getOptions(),
cq->getParsed().getFilter());
ccId = cc->cursorid();
log() << "caching runner with cursorid " << ccId << endl;
// ClientCursor takes ownership of runner. Release to make sure it's not deleted.
runner.release();
// TODO document
if (pq.hasOption(QueryOption_OplogReplay) && !slaveReadTill.isNull()) {
cc->slaveReadTill(slaveReadTill);
}
// TODO document
if (pq.hasOption(QueryOption_Exhaust)) {
curop.debug().exhaust = true;
}
// Set attributes for getMore.
cc->setCollMetadata(collMetadata);
cc->setPos(numResults);
// If the query had a time limit, remaining time is "rolled over" to the cursor (for
// use by future getmore ops).
cc->setLeftoverMaxTimeMicros(curop.getRemainingMaxTimeMicros());
}
// Add the results from the query into the output buffer.
result.appendData(bb.buf(), bb.len());
bb.decouple();
// Fill out the output buffer's header.
QueryResult* qr = static_cast<QueryResult*>(result.header());
qr->cursorId = ccId;
curop.debug().cursorid = (0 == ccId ? -1 : ccId);
qr->setResultFlagsToOk();
qr->setOperation(opReply);
qr->startingFrom = 0;
qr->nReturned = numResults;
// TODO: nscanned is bogus.
// curop.debug().nscanned = ( cursor ? cursor->nscanned() : 0LL );
curop.debug().ntoskip = pq.getSkip();
curop.debug().nreturned = numResults;
// curop.debug().exhaust is set above.
return curop.debug().exhaust ? pq.ns() : "";
}
/** @return number of skipped (invalid) documents */
unsigned compactExtent(const char *ns, NamespaceDetails *d, const DiskLoc diskloc, int n,
int nidx, bool validate, double pf, int pb, bool useDefaultPadding) {
log() << "compact begin extent #" << n << " for namespace " << ns << endl;
unsigned oldObjSize = 0; // we'll report what the old padding was
unsigned oldObjSizeWithPadding = 0;
Extent *e = diskloc.ext();
e->assertOk();
verify( e->validates(diskloc) );
unsigned skipped = 0;
Database* db = cc().database();
{
// the next/prev pointers within the extent might not be in order so we first
// page the whole thing in sequentially
log() << "compact paging in len=" << e->length/1000000.0 << "MB" << endl;
Timer t;
DataFile* mdf = db->getFile( diskloc.a() );
HANDLE fd = mdf->getFd();
int offset = diskloc.getOfs();
Extent* ext = diskloc.ext();
size_t length = ext->length;
touch_pages(fd, offset, length, ext);
int ms = t.millis();
if( ms > 1000 )
log() << "compact end paging in " << ms << "ms "
<< e->length/1000000.0/ms << "MB/sec" << endl;
}
{
log() << "compact copying records" << endl;
long long datasize = 0;
long long nrecords = 0;
DiskLoc L = e->firstRecord;
if( !L.isNull() ) {
while( 1 ) {
Record *recOld = L.rec();
L = db->getExtentManager().getNextRecordInExtent(L);
BSONObj objOld = BSONObj::make(recOld);
if( !validate || objOld.valid() ) {
nrecords++;
unsigned sz = objOld.objsize();
oldObjSize += sz;
oldObjSizeWithPadding += recOld->netLength();
unsigned lenWHdr = sz + Record::HeaderSize;
unsigned lenWPadding = lenWHdr;
// maintain UsePowerOf2Sizes if no padding values were passed in
if (d->isUserFlagSet(NamespaceDetails::Flag_UsePowerOf2Sizes)
&& useDefaultPadding) {
lenWPadding = d->quantizePowerOf2AllocationSpace(lenWPadding);
}
// otherwise use the padding values (pf and pb) that were passed in
else {
lenWPadding = static_cast<unsigned>(pf*lenWPadding);
lenWPadding += pb;
lenWPadding = lenWPadding & quantizeMask(lenWPadding);
}
if (lenWPadding < lenWHdr || lenWPadding > BSONObjMaxUserSize / 2 ) {
lenWPadding = lenWHdr;
}
DiskLoc loc = allocateSpaceForANewRecord(ns, d, lenWPadding, false);
uassert(14024, "compact error out of space during compaction", !loc.isNull());
Record *recNew = loc.rec();
datasize += recNew->netLength();
recNew = (Record *) getDur().writingPtr(recNew, lenWHdr);
addRecordToRecListInExtent(recNew, loc);
memcpy(recNew->data(), objOld.objdata(), sz);
}
else {
if( ++skipped <= 10 )
log() << "compact skipping invalid object" << endl;
}
if( L.isNull() ) {
// we just did the very last record from the old extent. it's still pointed to
// by the old extent ext, but that will be fixed below after this loop
break;
}
// remove the old records (orphan them) periodically so our commit block doesn't get too large
bool stopping = false;
RARELY stopping = *killCurrentOp.checkForInterruptNoAssert() != 0;
if( stopping || getDur().aCommitIsNeeded() ) {
e->firstRecord.writing() = L;
Record *r = L.rec();
getDur().writingInt(r->prevOfs()) = DiskLoc::NullOfs;
getDur().commitIfNeeded();
killCurrentOp.checkForInterrupt(false);
}
}
} // if !L.isNull()
verify( d->firstExtent() == diskloc );
verify( d->lastExtent() != diskloc );
DiskLoc newFirst = e->xnext;
d->firstExtent().writing() = newFirst;
newFirst.ext()->xprev.writing().Null();
getDur().writing(e)->markEmpty();
cc().database()->getExtentManager().freeExtents( diskloc, diskloc );
// update datasize/record count for this namespace's extent
d->incrementStats( datasize, nrecords );
getDur().commitIfNeeded();
{
double op = 1.0;
if( oldObjSize )
op = static_cast<double>(oldObjSizeWithPadding)/oldObjSize;
log() << "compact finished extent #" << n << " containing " << nrecords << " documents (" << datasize/1000000.0 << "MB)"
<< " oldPadding: " << op << ' ' << static_cast<unsigned>(op*100.0)/100
<< endl;
}
}
return skipped;
}
bool _compact(const char *ns, NamespaceDetails *d, string& errmsg, bool validate,
BSONObjBuilder& result, double pf, int pb, bool useDefaultPadding) {
// this is a big job, so might as well make things tidy before we start just to be nice.
getDur().commitIfNeeded();
list<DiskLoc> extents;
for( DiskLoc L = d->firstExtent(); !L.isNull(); L = L.ext()->xnext )
extents.push_back(L);
log() << "compact " << extents.size() << " extents" << endl;
ProgressMeterHolder pm(cc().curop()->setMessage("compact extent",
"Extent Compacting Progress",
extents.size()));
// same data, but might perform a little different after compact?
Collection* collection = cc().database()->getCollection( ns );
verify( collection );
collection->infoCache()->addedIndex();
verify( d->getCompletedIndexCount() == d->getTotalIndexCount() );
int nidx = d->getCompletedIndexCount();
scoped_array<BSONObj> indexSpecs( new BSONObj[nidx] );
{
NamespaceDetails::IndexIterator ii = d->ii();
// For each existing index...
for( int idxNo = 0; ii.more(); ++idxNo ) {
// Build a new index spec based on the old index spec.
BSONObjBuilder b;
BSONObj::iterator i(ii.next().info.obj());
while( i.more() ) {
BSONElement e = i.next();
if ( str::equals( e.fieldName(), "v" ) ) {
// Drop any preexisting index version spec. The default index version will
// be used instead for the new index.
continue;
}
if ( str::equals( e.fieldName(), "background" ) ) {
// Create the new index in the foreground.
continue;
}
// Pass the element through to the new index spec.
b.append(e);
}
indexSpecs[idxNo] = b.obj().getOwned();
}
}
log() << "compact orphan deleted lists" << endl;
d->orphanDeletedList();
// Start over from scratch with our extent sizing and growth
d->setLastExtentSize( 0 );
// before dropping indexes, at least make sure we can allocate one extent!
uassert(14025, "compact error no space available to allocate", !allocateSpaceForANewRecord(ns, d, Record::HeaderSize+1, false).isNull());
// note that the drop indexes call also invalidates all clientcursors for the namespace, which is important and wanted here
log() << "compact dropping indexes" << endl;
Status status = collection->getIndexCatalog()->dropAllIndexes( true );
if ( !status.isOK() ) {
errmsg = str::stream() << "compact drop indexes failed: " << status.toString();
log() << status.toString() << endl;
return false;
}
getDur().commitIfNeeded();
long long skipped = 0;
int n = 0;
// reset data size and record counts to 0 for this namespace
// as we're about to tally them up again for each new extent
d->setStats( 0, 0 );
for( list<DiskLoc>::iterator i = extents.begin(); i != extents.end(); i++ ) {
skipped += compactExtent(ns, d, *i, n++, nidx, validate, pf, pb, useDefaultPadding);
pm.hit();
}
if( skipped ) {
result.append("invalidObjects", skipped);
}
verify( d->firstExtent().ext()->xprev.isNull() );
// indexes will do their own progress meter?
pm.finished();
// build indexes
NamespaceString s(ns);
string si = s.db().toString() + ".system.indexes";
for( int i = 0; i < nidx; i++ ) {
killCurrentOp.checkForInterrupt(false);
BSONObj info = indexSpecs[i];
log() << "compact create index " << info["key"].Obj().toString() << endl;
theDataFileMgr.insert(si.c_str(), info.objdata(), info.objsize());
}
return true;
}
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 = nsToCollectionSubstring(ns) == "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() );
}
std::string DBHashCmd::hashCollection(OperationContext* opCtx,
Database* db,
const std::string& fullCollectionName,
bool* fromCache) {
stdx::unique_lock<stdx::mutex> cachedHashedLock(_cachedHashedMutex, stdx::defer_lock);
if (isCachable(fullCollectionName)) {
cachedHashedLock.lock();
string hash = _cachedHashed[fullCollectionName];
if (hash.size() > 0) {
*fromCache = true;
return hash;
}
}
*fromCache = false;
Collection* collection = db->getCollection(fullCollectionName);
if (!collection)
return "";
IndexDescriptor* desc = collection->getIndexCatalog()->findIdIndex(opCtx);
unique_ptr<PlanExecutor> exec;
if (desc) {
exec.reset(InternalPlanner::indexScan(opCtx,
collection,
desc,
BSONObj(),
BSONObj(),
false,
InternalPlanner::FORWARD,
InternalPlanner::IXSCAN_FETCH));
} else if (collection->isCapped()) {
exec.reset(InternalPlanner::collectionScan(opCtx, fullCollectionName, collection));
} else {
log() << "can't find _id index for: " << fullCollectionName << endl;
return "no _id _index";
}
md5_state_t st;
md5_init(&st);
long long n = 0;
PlanExecutor::ExecState state;
BSONObj c;
verify(NULL != exec.get());
while (PlanExecutor::ADVANCED == (state = exec->getNext(&c, NULL))) {
md5_append(&st, (const md5_byte_t*)c.objdata(), c.objsize());
n++;
}
if (PlanExecutor::IS_EOF != state) {
warning() << "error while hashing, db dropped? ns=" << fullCollectionName << endl;
}
md5digest d;
md5_finish(&st, d);
string hash = digestToString(d);
if (cachedHashedLock.owns_lock()) {
_cachedHashed[fullCollectionName] = hash;
}
return hash;
}
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 );
UpdateDriver::Options opts;
opts.multi = multi;
opts.upsert = upsert;
opts.logOp = logop;
UpdateDriver driver( opts );
Status status = driver.parse( nsdt->indexKeys(), updateobj );
if ( !status.isOK() ) {
uasserted( 16840, status.reason() );
}
shared_ptr<Cursor> cursor = getOptimizedCursor( ns, patternOrig, BSONObj(), planPolicy );
// 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.
const bool dedupHere = cursor->autoDedup();
shared_ptr<Cursor> cPtr = cursor;
auto_ptr<ClientCursor> clientCursor( new ClientCursor( QueryOption_NoCursorTimeout,
cPtr,
ns ) );
//
// Upsert Logic
//
// We may or may not have documents for this update. If we don't, then try to upsert,
// if allowed.
if ( !cursor->ok() && upsert ) {
// 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 {
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 );
mutablebson::Document doc( 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 */ );
}
//
// We have one or more documents for this update.
//
// 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;
while ( cursor->ok() ) {
// 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 ) {
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.
mutablebson::Document doc( 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.
BSONObj newObj;
const char* source = NULL;
mutablebson::DamageVector damages;
bool inPlace = doc.getInPlaceUpdates(&damages, &source);
if ( inPlace && !driver.modsAffectIndices() ) {
// 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;
}
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 );
}
}
// Log Obj
if ( logop ) {
if ( !logObj.isEmpty() ) {
BSONObj pattern = patternOrig;
logOp("u", ns, logObj , &pattern, 0, fromMigrate, &newObj );
}
}
// One more document updated.
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();
}
return UpdateResult( true /* updated existing object(s) */,
driver.dollarModMode() /* $mod or obj replacement */,
numUpdated /* # of docments update */,
BSONObj() );
}
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() );
}
}
bool ShardedClientCursor::sendNextBatch(int batchSize, BufBuilder& buffer, int& docCount) {
uassert(10191, "cursor already done", !_done);
int maxSize = 1024 * 1024;
if (_totalSent > 0)
maxSize *= 3;
docCount = 0;
// If batchSize is negative, it means that we should send up to -batchSize results
// back to the client, and that we should only send a *single batch*. An batchSize of
// 1 is also a special case which means "return up to 1 result in a single batch" (so
// that +1 actually has the same meaning of -1). For all other values of batchSize, we
// may have to return multiple batches.
const bool sendMoreBatches = batchSize == 0 || batchSize > 1;
batchSize = abs(batchSize);
// Set the initial batch size to 101, just like mongoD.
if (batchSize == 0 && _totalSent == 0)
batchSize = 101;
// Set batch size to batchSize requested by the current operation unconditionally. This is
// necessary because if the loop exited due to docCount == batchSize then setBatchSize(0) was
// called, so the next _cusor->more() will be called with a batch size of 0 if the cursor
// buffer was drained the previous run. Unconditionally setting the batch size ensures that
// we don't ask for a batch size of zero as a side effect.
_cursor->setBatchSize(batchSize);
bool cursorHasMore = true;
while ((cursorHasMore = _cursor->more())) {
BSONObj o = _cursor->next();
buffer.appendBuf((void*)o.objdata(), o.objsize());
++docCount;
// Ensure that the next batch will never wind up requesting more docs from the shard
// than are remaining to satisfy the initial batchSize.
if (batchSize != 0) {
if (docCount == batchSize)
break;
_cursor->setBatchSize(batchSize - docCount);
}
if (buffer.len() > maxSize) {
break;
}
}
// We need to request another batch if the following two conditions hold:
//
// 1. batchSize is positive and not equal to 1 (see the comment above). This condition
// is stored in 'sendMoreBatches'.
//
// 2. The last call to _cursor->more() was true (i.e. we never explicitly got a false
// value from _cursor->more()). This condition is stored in 'cursorHasMore'. If the server
// hits EOF while executing a query or a getmore, it will pass a cursorId of 0 in the
// query response to indicate that there are no more results. In this case, _cursor->more()
// will be explicitly false, and we know for sure that we do not have to send more batches.
//
// On the other hand, if _cursor->more() is true there may or may not be more results.
// Suppose that the mongod generates enough results to fill this batch. In this case it
// does not know whether not there are more, because doing so would require requesting an
// extra result and seeing whether we get EOF. The mongod sends a valid cursorId to
// indicate that there may be more. We do the same here: we indicate that there may be
// more results to retrieve by setting 'hasMoreBatches' to true.
bool hasMoreBatches = sendMoreBatches && cursorHasMore;
LOG(5) << "\t hasMoreBatches: " << hasMoreBatches << " sendMoreBatches: " << sendMoreBatches
<< " cursorHasMore: " << cursorHasMore << " batchSize: " << batchSize
<< " num: " << docCount << " id:" << getId() << " totalSent: " << _totalSent << endl;
_totalSent += docCount;
_done = !hasMoreBatches;
return hasMoreBatches;
}
// PD_TRACE_DECLARE_FUNCTION ( SDB_RTNINSERT2, "rtnInsert" )
INT32 rtnInsert ( const CHAR *pCollectionName, BSONObj &objs, INT32 objNum,
INT32 flags, pmdEDUCB *cb, SDB_DMSCB *dmsCB,
SDB_DPSCB *dpsCB, INT16 w )
{
INT32 rc = SDB_OK ;
PD_TRACE_ENTRY ( SDB_RTNINSERT2 ) ;
SDB_ASSERT ( pCollectionName, "collection name can't be NULL" ) ;
SDB_ASSERT ( cb, "educb can't be NULL" ) ;
SDB_ASSERT ( dmsCB, "dmsCB can't be NULL" ) ;
dmsStorageUnit *su = NULL ;
dmsStorageUnitID suID = DMS_INVALID_CS ;
const CHAR *pCollectionShortName = NULL ;
UINT32 insertCount = 0 ;
BOOLEAN writable = FALSE ;
ossValuePtr pDataPos = 0 ;
rc = dmsCB->writable( cb ) ;
if ( rc )
{
PD_LOG ( PDERROR, "Database is not writable, rc = %d", rc ) ;
goto error;
}
writable = TRUE;
rc = rtnResolveCollectionNameAndLock ( pCollectionName, dmsCB, &su,
&pCollectionShortName, suID ) ;
if ( rc )
{
PD_LOG ( PDERROR, "Failed to resolve collection name %s",
pCollectionName ) ;
goto error ;
}
if ( objs.isEmpty () )
{
PD_LOG ( PDERROR, "Insert record can't be empty" ) ;
rc = SDB_INVALIDARG ;
goto error ;
}
pDataPos = (ossValuePtr)objs.objdata() ;
for ( INT32 i = 0 ; i < objNum ; ++i )
{
if ( ++insertCount > RTN_INSERT_ONCE_NUM )
{
insertCount = 0 ;
if ( cb->isInterrupted() )
{
rc = SDB_APP_INTERRUPT ;
goto error ;
}
}
try
{
BSONObj record ( (const CHAR*)pDataPos ) ;
rc = su->insertRecord ( pCollectionShortName, record, cb, dpsCB ) ;
if ( rc )
{
if ( ( SDB_IXM_DUP_KEY == rc ) &&
( FLG_INSERT_CONTONDUP & flags ) )
{
rc = SDB_OK ;
}
else
{
PD_LOG ( PDERROR, "Failed to insert record %s into "
"collection: %s, rc: %d", record.toString().c_str(),
pCollectionName, rc ) ;
goto error ;
}
}
pDataPos += ossAlignX ( (ossValuePtr)record.objsize(), 4 ) ;
}
catch ( std::exception &e )
{
PD_LOG ( PDERROR, "Failed to convert to BSON and insert to "
"collection: %s", e.what() ) ;
rc = SDB_INVALIDARG ;
goto error ;
}
}
done :
if ( DMS_INVALID_CS != suID )
{
dmsCB->suUnlock ( suID ) ;
}
if ( writable )
{
dmsCB->writeDown( cb );
}
if ( cb )
{
if ( SDB_OK == rc && dpsCB )
{
rc = dpsCB->completeOpr( cb, w ) ;
}
}
PD_TRACE_EXITRC ( SDB_RTNINSERT2, rc ) ;
return rc ;
error :
goto done ;
}
bool _handlePossibleShardedMessage( Message &m, DbResponse* dbresponse ) {
DEV verify( shardingState.enabled() );
int op = m.operation();
if ( op < 2000
|| op >= 3000
|| op == dbGetMore // cursors are weird
)
return false;
DbMessage d(m);
const char *ns = d.getns();
string errmsg;
// We don't care about the version here, since we're returning it later in the writeback
ChunkVersion received, wanted;
if ( shardVersionOk( ns , errmsg, received, wanted ) ) {
return false;
}
bool getsAResponse = doesOpGetAResponse( op );
LOG(1) << "connection sharding metadata does not match for collection " << ns
<< ", will retry (wanted : " << wanted << ", received : " << received << ")"
<< ( getsAResponse ? "" : " (queuing writeback)" ) << endl;
if( getsAResponse ){
verify( dbresponse );
BufBuilder b( 32768 );
b.skip( sizeof( QueryResult ) );
{
BSONObjBuilder bob;
bob.append( "$err", errmsg );
bob.append( "ns", ns );
wanted.addToBSON( bob, "vWanted" );
received.addToBSON( bob, "vReceived" );
BSONObj obj = bob.obj();
b.appendBuf( obj.objdata() , obj.objsize() );
}
QueryResult *qr = (QueryResult*)b.buf();
qr->_resultFlags() = ResultFlag_ErrSet | ResultFlag_ShardConfigStale;
qr->len = b.len();
qr->setOperation( opReply );
qr->cursorId = 0;
qr->startingFrom = 0;
qr->nReturned = 1;
b.decouple();
Message * resp = new Message();
resp->setData( qr , true );
dbresponse->response = resp;
dbresponse->responseTo = m.header()->id;
return true;
}
uassert(9517, "cannot queue a writeback operation to the writeback queue",
(d.reservedField() & Reserved_FromWriteback) == 0);
const OID& clientID = ShardedConnectionInfo::get(false)->getID();
massert( 10422 , "write with bad shard config and no server id!" , clientID.isSet() );
// We need to check this here, since otherwise we'll get errors wrapping the writeback -
// not just here, but also when returning as a command result.
// We choose 1/2 the overhead of the internal maximum so that we can still handle ops of
// 16MB exactly.
massert( 16437, "data size of operation is too large to queue for writeback",
m.dataSize() < BSONObjMaxInternalSize - (8 * 1024));
LOG(1) << "writeback queued for " << m.toString() << endl;
BSONObjBuilder b;
b.appendBool( "writeBack" , true );
b.append( "ns" , ns );
b.append( "connectionId" , cc().getConnectionId() );
b.append( "instanceIdent" , prettyHostName() );
wanted.addToBSON( b );
received.addToBSON( b, "yourVersion" );
b.appendBinData( "msg" , m.header()->len , bdtCustom , (char*)(m.singleData()) );
LOG(2) << "writing back msg with len: " << m.header()->len << " op: " << m.operation() << endl;
// we pass the builder to queueWriteBack so that it can select the writebackId
// this is important so that the id is guaranteed to be ascending
// that is important since mongos assumes if its seen a greater writeback
// that all former have been processed
OID writebackID = writeBackManager.queueWriteBack( clientID.str() , b );
lastError.getSafe()->writeback( writebackID );
return true;
}
StatusWith<DiskLoc> Collection::updateDocument( OperationContext* txn,
const DiskLoc& oldLocation,
const BSONObj& objNew,
bool enforceQuota,
OpDebug* debug ) {
BSONObj objOld = _recordStore->dataFor( txn, oldLocation ).toBson();
if ( objOld.hasElement( "_id" ) ) {
BSONElement oldId = objOld["_id"];
BSONElement newId = objNew["_id"];
if ( oldId != newId )
return StatusWith<DiskLoc>( ErrorCodes::InternalError,
"in Collection::updateDocument _id mismatch",
13596 );
}
if ( ns().coll() == "system.users" ) {
// XXX - andy and spencer think this should go away now
V2UserDocumentParser parser;
Status s = parser.checkValidUserDocument(objNew);
if ( !s.isOK() )
return StatusWith<DiskLoc>( s );
}
/* duplicate key check. we descend the btree twice - once for this check, and once for the actual inserts, further
below. that is suboptimal, but it's pretty complicated to do it the other way without rollbacks...
*/
OwnedPointerMap<IndexDescriptor*,UpdateTicket> updateTickets;
IndexCatalog::IndexIterator ii = _indexCatalog.getIndexIterator( txn, true );
while ( ii.more() ) {
IndexDescriptor* descriptor = ii.next();
IndexAccessMethod* iam = _indexCatalog.getIndex( descriptor );
InsertDeleteOptions options;
options.logIfError = false;
options.dupsAllowed =
!(KeyPattern::isIdKeyPattern(descriptor->keyPattern()) || descriptor->unique())
|| repl::getGlobalReplicationCoordinator()->shouldIgnoreUniqueIndex(descriptor);
UpdateTicket* updateTicket = new UpdateTicket();
updateTickets.mutableMap()[descriptor] = updateTicket;
Status ret = iam->validateUpdate(txn, objOld, objNew, oldLocation, options, updateTicket );
if ( !ret.isOK() ) {
return StatusWith<DiskLoc>( ret );
}
}
// this can callback into Collection::recordStoreGoingToMove
StatusWith<DiskLoc> newLocation = _recordStore->updateRecord( txn,
oldLocation,
objNew.objdata(),
objNew.objsize(),
_enforceQuota( enforceQuota ),
this );
if ( !newLocation.isOK() ) {
return newLocation;
}
_infoCache.notifyOfWriteOp();
if ( newLocation.getValue() != oldLocation ) {
if ( debug ) {
if (debug->nmoved == -1) // default of -1 rather than 0
debug->nmoved = 1;
else
debug->nmoved += 1;
}
_indexCatalog.indexRecord(txn, objNew, newLocation.getValue());
return newLocation;
}
if ( debug )
debug->keyUpdates = 0;
ii = _indexCatalog.getIndexIterator( txn, true );
while ( ii.more() ) {
IndexDescriptor* descriptor = ii.next();
IndexAccessMethod* iam = _indexCatalog.getIndex( descriptor );
int64_t updatedKeys;
Status ret = iam->update(txn, *updateTickets.mutableMap()[descriptor], &updatedKeys);
if ( !ret.isOK() )
return StatusWith<DiskLoc>( ret );
if ( debug )
debug->keyUpdates += updatedKeys;
}
// Broadcast the mutation so that query results stay correct.
_cursorCache.invalidateDocument(oldLocation, INVALIDATION_MUTATION);
return newLocation;
}
bool _compact(const char *ns, NamespaceDetails *d, string& errmsg, bool validate, BSONObjBuilder& result) {
//int les = d->lastExtentSize;
// this is a big job, so might as well make things tidy before we start just to be nice.
getDur().commitNow();
list<DiskLoc> extents;
for( DiskLoc L = d->firstExtent; !L.isNull(); L = L.ext()->xnext )
extents.push_back(L);
log() << "compact " << extents.size() << " extents" << endl;
ProgressMeterHolder pm( cc().curop()->setMessage( "compact extent" , extents.size() ) );
// same data, but might perform a little different after compact?
NamespaceDetailsTransient::get_w(ns).clearQueryCache();
int nidx = d->nIndexes;
scoped_array<IndexSpec> indexSpecs( new IndexSpec[nidx] );
scoped_array<SortPhaseOne> phase1( new SortPhaseOne[nidx] );
{
NamespaceDetails::IndexIterator ii = d->ii();
int x = 0;
while( ii.more() ) {
BSONObjBuilder b;
BSONObj::iterator i(ii.next().info.obj());
while( i.more() ) {
BSONElement e = i.next();
if( !str::equals(e.fieldName(), "v") && !str::equals(e.fieldName(), "background") ) {
b.append(e);
}
}
BSONObj o = b.obj().getOwned();
phase1[x].sorter.reset( new BSONObjExternalSorter( o.getObjectField("key") ) );
phase1[x].sorter->hintNumObjects( d->stats.nrecords );
indexSpecs[x++].reset(o);
}
}
log() << "compact orphan deleted lists" << endl;
for( int i = 0; i < Buckets; i++ ) {
d->deletedList[i].writing().Null();
}
// before dropping indexes, at least make sure we can allocate one extent!
uassert(14025, "compact error no space available to allocate", !allocateSpaceForANewRecord(ns, d, Record::HeaderSize+1).isNull());
// note that the drop indexes call also invalidates all clientcursors for the namespace, which is important and wanted here
log() << "compact dropping indexes" << endl;
BSONObjBuilder b;
if( !dropIndexes(d, ns, "*", errmsg, b, true) ) {
errmsg = "compact drop indexes failed";
log() << errmsg << endl;
return false;
}
getDur().commitNow();
long long skipped = 0;
int n = 0;
for( list<DiskLoc>::iterator i = extents.begin(); i != extents.end(); i++ ) {
skipped += compactExtent(ns, d, *i, n++, indexSpecs, phase1, nidx, validate);
pm.hit();
}
if( skipped ) {
result.append("invalidObjects", skipped);
}
assert( d->firstExtent.ext()->xprev.isNull() );
// indexes will do their own progress meter?
pm.finished();
// build indexes
NamespaceString s(ns);
string si = s.db + ".system.indexes";
for( int i = 0; i < nidx; i++ ) {
killCurrentOp.checkForInterrupt(false);
BSONObj info = indexSpecs[i].info;
log() << "compact create index " << info["key"].Obj().toString() << endl;
try {
precalced = &phase1[i];
theDataFileMgr.insert(si.c_str(), info.objdata(), info.objsize());
}
catch(...) {
precalced = 0;
throw;
}
precalced = 0;
}
return true;
}
static void insert( const BSONObj &o, bool god = false ) {
dblock lk;
Client::Context ctx( ns() );
theDataFileMgr.insert( ns(), o.objdata(), o.objsize(), god );
}
/**
* Also called by db/ops/query.cpp. This is the new getMore entry point.
*/
QueryResult* newGetMore(const char* ns, int ntoreturn, long long cursorid, CurOp& curop,
int pass, bool& exhaust, bool* isCursorAuthorized) {
exhaust = false;
int bufSize = 512 + sizeof(QueryResult) + MaxBytesToReturnToClientAtOnce;
BufBuilder bb(bufSize);
bb.skip(sizeof(QueryResult));
// This is a read lock. TODO: There is a cursor flag for not needing this. Do we care?
Client::ReadContext ctx(ns);
QLOG() << "running getMore in new system, cursorid " << cursorid << endl;
// This checks to make sure the operation is allowed on a replicated node. Since we are not
// passing in a query object (necessary to check SlaveOK query option), the only state where
// reads are allowed is PRIMARY (or master in master/slave). This function uasserts if
// reads are not okay.
replVerifyReadsOk();
// A pin performs a CC lookup and if there is a CC, increments the CC's pin value so it
// doesn't time out. Also informs ClientCursor that there is somebody actively holding the
// CC, so don't delete it.
ClientCursorPin ccPin(cursorid);
ClientCursor* cc = ccPin.c();
// These are set in the QueryResult msg we return.
int resultFlags = ResultFlag_AwaitCapable;
int numResults = 0;
int startingResult = 0;
if (NULL == cc) {
cursorid = 0;
resultFlags = ResultFlag_CursorNotFound;
}
else {
// Quote: check for spoofing of the ns such that it does not match the one originally
// there for the cursor
uassert(17011, "auth error", str::equals(ns, cc->ns().c_str()));
*isCursorAuthorized = true;
// TODO: fail point?
// If the operation that spawned this cursor had a time limit set, apply leftover
// time to this getmore.
curop.setMaxTimeMicros(cc->getLeftoverMaxTimeMicros());
killCurrentOp.checkForInterrupt(); // May trigger maxTimeAlwaysTimeOut fail point.
// TODO:
// curop.debug().query = BSONForQuery
// curop.setQuery(curop.debug().query);
// TODO: What is pass?
if (0 == pass) { cc->updateSlaveLocation(curop); }
CollectionMetadataPtr collMetadata = cc->getCollMetadata();
// If we're replaying the oplog, we save the last time that we read.
OpTime slaveReadTill;
// What number result are we starting at? Used to fill out the reply.
startingResult = cc->pos();
// What gives us results.
Runner* runner = cc->getRunner();
const int queryOptions = cc->queryOptions();
// Get results out of the runner.
runner->restoreState();
BSONObj obj;
Runner::RunnerState state;
while (Runner::RUNNER_ADVANCED == (state = runner->getNext(&obj, NULL))) {
// If we're sharded make sure that we don't return any data that hasn't been
// migrated off of our shard yet.
if (collMetadata) {
KeyPattern kp(collMetadata->getKeyPattern());
if (!collMetadata->keyBelongsToMe(kp.extractSingleKey(obj))) { continue; }
}
// Add result to output buffer.
bb.appendBuf((void*)obj.objdata(), obj.objsize());
// Count the result.
++numResults;
// Possibly note slave's position in the oplog.
if (queryOptions & QueryOption_OplogReplay) {
BSONElement e = obj["ts"];
if (Date == e.type() || Timestamp == e.type()) {
slaveReadTill = e._opTime();
}
}
if ((numResults && numResults >= ntoreturn)
|| bb.len() > MaxBytesToReturnToClientAtOnce) {
break;
}
}
if (Runner::RUNNER_EOF == state
&& 0 == numResults
&& (queryOptions & QueryOption_CursorTailable)
&& (queryOptions & QueryOption_AwaitData)
&& (pass < 1000)) {
// If the cursor is tailable we don't kill it if it's eof. We let it try to get
// data some # of times first.
return 0;
}
else if (Runner::RUNNER_DEAD == state || Runner::RUNNER_EOF == state) {
ccPin.free();
// cc is now invalid, as is the runner
cursorid = 0;
cc = NULL;
}
else {
// Continue caching the ClientCursor.
cc->incPos(numResults);
runner->saveState();
// Possibly note slave's position in the oplog.
if ((queryOptions & QueryOption_OplogReplay) && !slaveReadTill.isNull()) {
cc->slaveReadTill(slaveReadTill);
}
exhaust = (queryOptions & QueryOption_Exhaust);
// If the getmore had a time limit, remaining time is "rolled over" back to the
// cursor (for use by future getmore ops).
cc->setLeftoverMaxTimeMicros( curop.getRemainingMaxTimeMicros() );
}
}
QueryResult* qr = reinterpret_cast<QueryResult*>(bb.buf());
qr->len = bb.len();
qr->setOperation(opReply);
qr->_resultFlags() = resultFlags;
qr->cursorId = cursorid;
qr->startingFrom = startingResult;
qr->nReturned = numResults;
bb.decouple();
return qr;
}
/** object cannot be represented in compact format. so store in traditional bson format
with a leading sentinel byte IsBSON to indicate it's in that format.
Given that the KeyV1Owned constructor already grabbed a bufbuilder, we reuse it here
so that we don't have to do an extra malloc.
*/
void KeyV1Owned::traditional(const BSONObj& obj) {
b.reset();
b.appendUChar(IsBSON);
b.appendBuf(obj.objdata(), obj.objsize());
_keyData = (const unsigned char *) b.buf();
}
std::string runQuery(OperationContext* txn,
QueryMessage& q,
const NamespaceString& nss,
Message& result) {
CurOp& curop = *CurOp::get(txn);
// Validate the namespace.
uassert(16256, str::stream() << "Invalid ns [" << nss.ns() << "]", nss.isValid());
invariant(!nss.isCommand());
// Set curop information.
beginQueryOp(txn, nss, q.query, q.ntoreturn, q.ntoskip);
// Parse the qm into a CanonicalQuery.
auto statusWithCQ = CanonicalQuery::canonicalize(q, ExtensionsCallbackReal(txn, &nss));
if (!statusWithCQ.isOK()) {
uasserted(
17287,
str::stream() << "Can't canonicalize query: " << statusWithCQ.getStatus().toString());
}
unique_ptr<CanonicalQuery> cq = std::move(statusWithCQ.getValue());
invariant(cq.get());
LOG(5) << "Running query:\n" << cq->toString();
LOG(2) << "Running query: " << cq->toStringShort();
// Parse, canonicalize, plan, transcribe, and get a plan executor.
AutoGetCollectionForRead ctx(txn, nss);
Collection* collection = ctx.getCollection();
const int dbProfilingLevel =
ctx.getDb() ? ctx.getDb()->getProfilingLevel() : serverGlobalParams.defaultProfile;
// We have a parsed query. Time to get the execution plan for it.
std::unique_ptr<PlanExecutor> exec = uassertStatusOK(
getExecutorFind(txn, collection, nss, std::move(cq), PlanExecutor::YIELD_AUTO));
const LiteParsedQuery& pq = exec->getCanonicalQuery()->getParsed();
// If it's actually an explain, do the explain and return rather than falling through
// to the normal query execution loop.
if (pq.isExplain()) {
BufBuilder bb;
bb.skip(sizeof(QueryResult::Value));
BSONObjBuilder explainBob;
Explain::explainStages(exec.get(), ExplainCommon::EXEC_ALL_PLANS, &explainBob);
// Add the resulting object to the return buffer.
BSONObj explainObj = explainBob.obj();
bb.appendBuf((void*)explainObj.objdata(), explainObj.objsize());
// TODO: Does this get overwritten/do we really need to set this twice?
curop.debug().query = q.query;
// Set query result fields.
QueryResult::View qr = bb.buf();
bb.decouple();
qr.setResultFlagsToOk();
qr.msgdata().setLen(bb.len());
curop.debug().responseLength = bb.len();
qr.msgdata().setOperation(opReply);
qr.setCursorId(0);
qr.setStartingFrom(0);
qr.setNReturned(1);
result.setData(qr.view2ptr(), true);
return "";
}
ShardingState* const shardingState = ShardingState::get(txn);
// We freak out later if this changes before we're done with the query.
const ChunkVersion shardingVersionAtStart = shardingState->getVersion(nss.ns());
// Handle query option $maxTimeMS (not used with commands).
curop.setMaxTimeMicros(static_cast<unsigned long long>(pq.getMaxTimeMS()) * 1000);
txn->checkForInterrupt(); // May trigger maxTimeAlwaysTimeOut fail point.
// uassert if we are not on a primary, and not a secondary with SlaveOk query parameter set.
bool slaveOK = pq.isSlaveOk() || pq.hasReadPref();
Status serveReadsStatus =
repl::getGlobalReplicationCoordinator()->checkCanServeReadsFor(txn, nss, slaveOK);
uassertStatusOK(serveReadsStatus);
// Run the query.
// bb is used to hold query results
// this buffer should contain either requested documents per query or
// explain information, but not both
BufBuilder bb(FindCommon::kInitReplyBufferSize);
bb.skip(sizeof(QueryResult::Value));
// How many results have we obtained from the executor?
int numResults = 0;
// If we're replaying the oplog, we save the last time that we read.
Timestamp slaveReadTill;
BSONObj obj;
PlanExecutor::ExecState state;
// uint64_t numMisplacedDocs = 0;
// Get summary info about which plan the executor is using.
{
stdx::lock_guard<Client> lk(*txn->getClient());
curop.setPlanSummary_inlock(Explain::getPlanSummary(exec.get()));
}
while (PlanExecutor::ADVANCED == (state = exec->getNext(&obj, NULL))) {
// If we can't fit this result inside the current batch, then we stash it for later.
if (!FindCommon::haveSpaceForNext(obj, numResults, bb.len())) {
exec->enqueue(obj);
break;
}
// Add result to output buffer.
bb.appendBuf((void*)obj.objdata(), obj.objsize());
// Count the result.
++numResults;
// Possibly note slave's position in the oplog.
if (pq.isOplogReplay()) {
BSONElement e = obj["ts"];
if (Date == e.type() || bsonTimestamp == e.type()) {
slaveReadTill = e.timestamp();
}
}
if (FindCommon::enoughForFirstBatch(pq, numResults)) {
LOG(5) << "Enough for first batch, wantMore=" << pq.wantMore()
<< " ntoreturn=" << pq.getNToReturn().value_or(0) << " numResults=" << numResults
<< endl;
break;
}
}
// If we cache the executor later, we want to deregister it as it receives notifications
// anyway by virtue of being cached.
//
// If we don't cache the executor later, we are deleting it, so it must be deregistered.
//
// So, no matter what, deregister the executor.
exec->deregisterExec();
// Caller expects exceptions thrown in certain cases.
if (PlanExecutor::FAILURE == state || PlanExecutor::DEAD == state) {
const unique_ptr<PlanStageStats> stats(exec->getStats());
error() << "Plan executor error during find: " << PlanExecutor::statestr(state)
<< ", stats: " << Explain::statsToBSON(*stats);
uasserted(17144, "Executor error: " + WorkingSetCommon::toStatusString(obj));
}
// TODO: Currently, chunk ranges are kept around until all ClientCursors created while the
// chunk belonged on this node are gone. Separating chunk lifetime management from
// ClientCursor should allow this check to go away.
if (!shardingState->getVersion(nss.ns()).isWriteCompatibleWith(shardingVersionAtStart)) {
// if the version changed during the query we might be missing some data and its safe to
// send this as mongos can resend at this point
throw SendStaleConfigException(nss.ns(),
"version changed during initial query",
shardingVersionAtStart,
shardingState->getVersion(nss.ns()));
}
// Fill out curop based on query results. If we have a cursorid, we will fill out curop with
// this cursorid later.
long long ccId = 0;
if (shouldSaveCursor(txn, collection, state, exec.get())) {
// We won't use the executor until it's getMore'd.
exec->saveState();
exec->detachFromOperationContext();
// Allocate a new ClientCursor. We don't have to worry about leaking it as it's
// inserted into a global map by its ctor.
ClientCursor* cc =
new ClientCursor(collection->getCursorManager(),
exec.release(),
nss.ns(),
txn->recoveryUnit()->isReadingFromMajorityCommittedSnapshot(),
pq.getOptions(),
pq.getFilter());
ccId = cc->cursorid();
LOG(5) << "caching executor with cursorid " << ccId << " after returning " << numResults
<< " results" << endl;
// TODO document
if (pq.isOplogReplay() && !slaveReadTill.isNull()) {
cc->slaveReadTill(slaveReadTill);
}
// TODO document
if (pq.isExhaust()) {
curop.debug().exhaust = true;
}
cc->setPos(numResults);
// If the query had a time limit, remaining time is "rolled over" to the cursor (for
// use by future getmore ops).
cc->setLeftoverMaxTimeMicros(curop.getRemainingMaxTimeMicros());
endQueryOp(txn, collection, *cc->getExecutor(), dbProfilingLevel, numResults, ccId);
} else {
LOG(5) << "Not caching executor but returning " << numResults << " results.\n";
endQueryOp(txn, collection, *exec, dbProfilingLevel, numResults, ccId);
}
// Add the results from the query into the output buffer.
result.appendData(bb.buf(), bb.len());
bb.decouple();
// Fill out the output buffer's header.
QueryResult::View qr = result.header().view2ptr();
qr.setCursorId(ccId);
qr.setResultFlagsToOk();
qr.msgdata().setOperation(opReply);
qr.setStartingFrom(0);
qr.setNReturned(numResults);
// curop.debug().exhaust is set above.
return curop.debug().exhaust ? nss.ns() : "";
}
StatusWith<RecordId> Collection::updateDocument(OperationContext* txn,
const RecordId& oldLocation,
const Snapshotted<BSONObj>& oldDoc,
const BSONObj& newDoc,
bool enforceQuota,
bool indexesAffected,
OpDebug* debug,
oplogUpdateEntryArgs& args) {
{
auto status = checkValidation(txn, newDoc);
if (!status.isOK()) {
if (_validationLevel == STRICT_V) {
return status;
}
// moderate means we have to check the old doc
auto oldDocStatus = checkValidation(txn, oldDoc.value());
if (oldDocStatus.isOK()) {
// transitioning from good -> bad is not ok
return status;
}
// bad -> bad is ok in moderate mode
}
}
dassert(txn->lockState()->isCollectionLockedForMode(ns().toString(), MODE_IX));
invariant(oldDoc.snapshotId() == txn->recoveryUnit()->getSnapshotId());
SnapshotId sid = txn->recoveryUnit()->getSnapshotId();
BSONElement oldId = oldDoc.value()["_id"];
if (!oldId.eoo() && (oldId != newDoc["_id"]))
return StatusWith<RecordId>(
ErrorCodes::InternalError, "in Collection::updateDocument _id mismatch", 13596);
// At the end of this step, we will have a map of UpdateTickets, one per index, which
// represent the index updates needed to be done, based on the changes between oldDoc and
// newDoc.
OwnedPointerMap<IndexDescriptor*, UpdateTicket> updateTickets;
if (indexesAffected) {
IndexCatalog::IndexIterator ii = _indexCatalog.getIndexIterator(txn, true);
while (ii.more()) {
IndexDescriptor* descriptor = ii.next();
IndexCatalogEntry* entry = ii.catalogEntry(descriptor);
IndexAccessMethod* iam = ii.accessMethod(descriptor);
InsertDeleteOptions options;
options.logIfError = false;
options.dupsAllowed =
!(KeyPattern::isIdKeyPattern(descriptor->keyPattern()) || descriptor->unique()) ||
repl::getGlobalReplicationCoordinator()->shouldIgnoreUniqueIndex(descriptor);
UpdateTicket* updateTicket = new UpdateTicket();
updateTickets.mutableMap()[descriptor] = updateTicket;
Status ret = iam->validateUpdate(txn,
oldDoc.value(),
newDoc,
oldLocation,
options,
updateTicket,
entry->getFilterExpression());
if (!ret.isOK()) {
return StatusWith<RecordId>(ret);
}
}
}
// This can call back into Collection::recordStoreGoingToMove. If that happens, the old
// object is removed from all indexes.
StatusWith<RecordId> newLocation = _recordStore->updateRecord(
txn, oldLocation, newDoc.objdata(), newDoc.objsize(), _enforceQuota(enforceQuota), this);
if (!newLocation.isOK()) {
return newLocation;
}
// At this point, the old object may or may not still be indexed, depending on if it was
// moved. If the object did move, we need to add the new location to all indexes.
if (newLocation.getValue() != oldLocation) {
if (debug) {
if (debug->nmoved == -1) // default of -1 rather than 0
debug->nmoved = 1;
else
debug->nmoved += 1;
}
Status s = _indexCatalog.indexRecord(txn, newDoc, newLocation.getValue());
if (!s.isOK())
return StatusWith<RecordId>(s);
invariant(sid == txn->recoveryUnit()->getSnapshotId());
args.ns = ns().ns();
getGlobalServiceContext()->getOpObserver()->onUpdate(txn, args);
return newLocation;
}
// Object did not move. We update each index with each respective UpdateTicket.
if (debug)
debug->keyUpdates = 0;
if (indexesAffected) {
IndexCatalog::IndexIterator ii = _indexCatalog.getIndexIterator(txn, true);
while (ii.more()) {
IndexDescriptor* descriptor = ii.next();
IndexAccessMethod* iam = ii.accessMethod(descriptor);
int64_t updatedKeys;
Status ret = iam->update(txn, *updateTickets.mutableMap()[descriptor], &updatedKeys);
if (!ret.isOK())
return StatusWith<RecordId>(ret);
if (debug)
debug->keyUpdates += updatedKeys;
}
}
invariant(sid == txn->recoveryUnit()->getSnapshotId());
args.ns = ns().ns();
getGlobalServiceContext()->getOpObserver()->onUpdate(txn, args);
return newLocation;
}
Status FTDCFileWriter::writeMetadata(const BSONObj& metadata, Date_t date) {
BSONObj wrapped = FTDCBSONUtil::createBSONMetadataDocument(metadata, date);
return writeArchiveFileBuffer({wrapped.objdata(), static_cast<size_t>(wrapped.objsize())});
}
bool ShardedClientCursor::sendNextBatch(int ntoreturn, BufBuilder& buffer, int& docCount) {
uassert( 10191 , "cursor already done" , ! _done );
int maxSize = 1024 * 1024;
if ( _totalSent > 0 )
maxSize *= 3;
docCount = 0;
// If ntoreturn is negative, it means that we should send up to -ntoreturn results
// back to the client, and that we should only send a *single batch*. An ntoreturn of
// 1 is also a special case which means "return up to 1 result in a single batch" (so
// that +1 actually has the same meaning of -1). For all other values of ntoreturn, we
// may have to return multiple batches.
const bool sendMoreBatches = ntoreturn == 0 || ntoreturn > 1;
ntoreturn = abs( ntoreturn );
bool cursorHasMore = true;
while ( ( cursorHasMore = _cursor->more() ) ) {
BSONObj o = _cursor->next();
buffer.appendBuf( (void*)o.objdata() , o.objsize() );
docCount++;
// Ensure that the next batch will never wind up requesting more docs from the shard
// than are remaining to satisfy the initial ntoreturn.
if (ntoreturn != 0) {
_cursor->setBatchSize(ntoreturn - docCount);
}
if ( buffer.len() > maxSize ) {
break;
}
if ( docCount == ntoreturn ) {
// soft limit aka batch size
break;
}
if ( ntoreturn == 0 && _totalSent == 0 && docCount >= 100 ) {
// first batch should be max 100 unless batch size specified
break;
}
}
// We need to request another batch if the following two conditions hold:
//
// 1. ntoreturn is positive and not equal to 1 (see the comment above). This condition
// is stored in 'sendMoreBatches'.
//
// 2. The last call to _cursor->more() was true (i.e. we never explicitly got a false
// value from _cursor->more()). This condition is stored in 'cursorHasMore'. If the server
// hits EOF while executing a query or a getmore, it will pass a cursorId of 0 in the
// query response to indicate that there are no more results. In this case, _cursor->more()
// will be explicitly false, and we know for sure that we do not have to send more batches.
//
// On the other hand, if _cursor->more() is true there may or may not be more results.
// Suppose that the mongod generates enough results to fill this batch. In this case it
// does not know whether not there are more, because doing so would require requesting an
// extra result and seeing whether we get EOF. The mongod sends a valid cursorId to
// indicate that there may be more. We do the same here: we indicate that there may be
// more results to retrieve by setting 'hasMoreBatches' to true.
bool hasMoreBatches = sendMoreBatches && cursorHasMore;
LOG(5) << "\t hasMoreBatches: " << hasMoreBatches
<< " sendMoreBatches: " << sendMoreBatches
<< " cursorHasMore: " << cursorHasMore
<< " ntoreturn: " << ntoreturn
<< " num: " << docCount
<< " id:" << getId()
<< " totalSent: " << _totalSent << endl;
_totalSent += docCount;
_done = ! hasMoreBatches;
return hasMoreBatches;
}
