PlanStage::StageState AndHashStage::readFirstChild(WorkingSetID* out) {
verify(_currentChild == 0);
WorkingSetID id = WorkingSet::INVALID_ID;
StageState childStatus = workChild(0, &id);
if (PlanStage::ADVANCED == childStatus) {
WorkingSetMember* member = _ws->get(id);
// The child must give us a WorkingSetMember with a record id, since we intersect index keys
// based on the record id. The planner ensures that the child stage can never produce an WSM
// with no record id.
invariant(member->hasRecordId());
if (!_dataMap.insert(std::make_pair(member->recordId, id)).second) {
// Didn't insert because we already had this RecordId inside the map. This should only
// happen if we're seeing a newer copy of the same doc in a more recent snapshot.
// Throw out the newer copy of the doc.
_ws->free(id);
return PlanStage::NEED_TIME;
}
// Ensure that the BSONObj underlying the WorkingSetMember is owned in case we yield.
member->makeObjOwnedIfNeeded();
// Update memory stats.
_memUsage += member->getMemUsage();
return PlanStage::NEED_TIME;
} else if (PlanStage::IS_EOF == childStatus) {
// Done reading child 0.
_currentChild = 1;
// If our first child was empty, don't scan any others, no possible results.
if (_dataMap.empty()) {
_hashingChildren = false;
return PlanStage::IS_EOF;
}
_specificStats.mapAfterChild.push_back(_dataMap.size());
return PlanStage::NEED_TIME;
} else if (PlanStage::FAILURE == childStatus || PlanStage::DEAD == childStatus) {
// The stage which produces a failure is responsible for allocating a working set member
// with error details.
invariant(WorkingSet::INVALID_ID != id);
*out = id;
return childStatus;
} else {
if (PlanStage::NEED_YIELD == childStatus) {
*out = id;
}
return childStatus;
}
}
BSONObj getObj() const {
if (!WorkingSetCommon::fetchIfUnfetched(_opCtx, _ws, _id, _recordCursor))
throw DocumentDeletedException();
WorkingSetMember* member = _ws->get(_id);
// Make it owned since we are buffering results.
member->makeObjOwnedIfNeeded();
return member->obj.value();
}
Status CachedPlanStage::pickBestPlan(PlanYieldPolicy* yieldPolicy) {
// Adds the amount of time taken by pickBestPlan() to executionTimeMillis. There's lots of
// execution work that happens here, so this is needed for the time accounting to
// make sense.
ScopedTimer timer(&_commonStats.executionTimeMillis);
// If we work this many times during the trial period, then we will replan the
// query from scratch.
size_t maxWorksBeforeReplan =
static_cast<size_t>(internalQueryCacheEvictionRatio * _decisionWorks);
// The trial period ends without replanning if the cached plan produces this many results.
size_t numResults = MultiPlanStage::getTrialPeriodNumToReturn(*_canonicalQuery);
for (size_t i = 0; i < maxWorksBeforeReplan; ++i) {
// Might need to yield between calls to work due to the timer elapsing.
Status yieldStatus = tryYield(yieldPolicy);
if (!yieldStatus.isOK()) {
return yieldStatus;
}
WorkingSetID id = WorkingSet::INVALID_ID;
PlanStage::StageState state = child()->work(&id);
if (PlanStage::ADVANCED == state) {
// Save result for later.
WorkingSetMember* member = _ws->get(id);
// Ensure that the BSONObj underlying the WorkingSetMember is owned in case we yield.
member->makeObjOwnedIfNeeded();
_results.push_back(id);
if (_results.size() >= numResults) {
// Once a plan returns enough results, stop working. Update cache with stats
// from this run and return.
updatePlanCache();
return Status::OK();
}
} else if (PlanStage::IS_EOF == state) {
// Cached plan hit EOF quickly enough. No need to replan. Update cache with stats
// from this run and return.
updatePlanCache();
return Status::OK();
} else if (PlanStage::NEED_YIELD == state) {
if (id == WorkingSet::INVALID_ID) {
if (!yieldPolicy->allowedToYield()) {
throw WriteConflictException();
}
} else {
WorkingSetMember* member = _ws->get(id);
invariant(member->hasFetcher());
// Transfer ownership of the fetcher and yield.
_fetcher.reset(member->releaseFetcher());
}
if (yieldPolicy->allowedToYield()) {
yieldPolicy->forceYield();
}
Status yieldStatus = tryYield(yieldPolicy);
if (!yieldStatus.isOK()) {
return yieldStatus;
}
} else if (PlanStage::FAILURE == state) {
// On failure, fall back to replanning the whole query. We neither evict the
// existing cache entry nor cache the result of replanning.
BSONObj statusObj;
WorkingSetCommon::getStatusMemberObject(*_ws, id, &statusObj);
LOG(1) << "Execution of cached plan failed, falling back to replan."
<< " query: " << _canonicalQuery->toStringShort()
<< " planSummary: " << Explain::getPlanSummary(child().get())
<< " status: " << statusObj;
const bool shouldCache = false;
return replan(yieldPolicy, shouldCache);
} else if (PlanStage::DEAD == state) {
BSONObj statusObj;
WorkingSetCommon::getStatusMemberObject(*_ws, id, &statusObj);
LOG(1) << "Execution of cached plan failed: PlanStage died"
<< ", query: " << _canonicalQuery->toStringShort()
<< " planSummary: " << Explain::getPlanSummary(child().get())
<< " status: " << statusObj;
return WorkingSetCommon::getMemberObjectStatus(statusObj);
} else {
invariant(PlanStage::NEED_TIME == state);
}
}
// If we're here, the trial period took more than 'maxWorksBeforeReplan' work cycles. This
// plan is taking too long, so we replan from scratch.
LOG(1) << "Execution of cached plan required " << maxWorksBeforeReplan
<< " works, but was originally cached with only " << _decisionWorks
<< " works. Evicting cache entry and replanning query: "
<< _canonicalQuery->toStringShort()
<< " plan summary before replan: " << Explain::getPlanSummary(child().get());
const bool shouldCache = true;
return replan(yieldPolicy, shouldCache);
}
/**
* addToBuffer() and sortBuffer() work differently based on the
* configured limit. addToBuffer() is also responsible for
* performing some accounting on the overall memory usage to
* make sure we're not using too much memory.
*
* limit == 0:
* addToBuffer() - Adds item to vector.
* sortBuffer() - Sorts vector.
* limit == 1:
* addToBuffer() - Replaces first item in vector with max of
* current and new item.
* Updates memory usage if item was replaced.
* sortBuffer() - Does nothing.
* limit > 1:
* addToBuffer() - Does not update vector. Adds item to set.
* If size of set exceeds limit, remove item from set
* with lowest key. Updates memory usage accordingly.
* sortBuffer() - Copies items from set to vectors.
*/
void SortStage::addToBuffer(const SortableDataItem& item) {
// Holds ID of working set member to be freed at end of this function.
WorkingSetID wsidToFree = WorkingSet::INVALID_ID;
WorkingSetMember* member = _ws->get(item.wsid);
if (_limit == 0) {
// Ensure that the BSONObj underlying the WorkingSetMember is owned in case we yield.
member->makeObjOwnedIfNeeded();
_data.push_back(item);
_memUsage += member->getMemUsage();
} else if (_limit == 1) {
if (_data.empty()) {
member->makeObjOwnedIfNeeded();
_data.push_back(item);
_memUsage = member->getMemUsage();
return;
}
wsidToFree = item.wsid;
const WorkingSetComparator& cmp = *_sortKeyComparator;
// Compare new item with existing item in vector.
if (cmp(item, _data[0])) {
wsidToFree = _data[0].wsid;
member->makeObjOwnedIfNeeded();
_data[0] = item;
_memUsage = member->getMemUsage();
}
} else {
// Update data item set instead of vector
// Limit not reached - insert and return
vector<SortableDataItem>::size_type limit(_limit);
if (_dataSet->size() < limit) {
member->makeObjOwnedIfNeeded();
_dataSet->insert(item);
_memUsage += member->getMemUsage();
return;
}
// Limit will be exceeded - compare with item with lowest key
// If new item does not have a lower key value than last item,
// do nothing.
wsidToFree = item.wsid;
SortableDataItemSet::const_iterator lastItemIt = --(_dataSet->end());
const SortableDataItem& lastItem = *lastItemIt;
const WorkingSetComparator& cmp = *_sortKeyComparator;
if (cmp(item, lastItem)) {
_memUsage -= _ws->get(lastItem.wsid)->getMemUsage();
_memUsage += member->getMemUsage();
wsidToFree = lastItem.wsid;
// According to std::set iterator validity rules,
// it does not matter which of erase()/insert() happens first.
// Here, we choose to erase first to release potential resources
// used by the last item and to keep the scope of the iterator to a minimum.
_dataSet->erase(lastItemIt);
member->makeObjOwnedIfNeeded();
_dataSet->insert(item);
}
}
// If the working set ID is valid, remove from
// RecordId invalidation map and free from working set.
if (wsidToFree != WorkingSet::INVALID_ID) {
WorkingSetMember* member = _ws->get(wsidToFree);
if (member->hasLoc()) {
_wsidByDiskLoc.erase(member->loc);
}
_ws->free(wsidToFree);
}
}
bool MultiPlanStage::workAllPlans(size_t numResults, PlanYieldPolicy* yieldPolicy) {
bool doneWorking = false;
for (size_t ix = 0; ix < _candidates.size(); ++ix) {
CandidatePlan& candidate = _candidates[ix];
if (candidate.failed) {
continue;
}
// Might need to yield between calls to work due to the timer elapsing.
if (!(tryYield(yieldPolicy)).isOK()) {
return false;
}
WorkingSetID id = WorkingSet::INVALID_ID;
PlanStage::StageState state = candidate.root->work(&id);
if (PlanStage::ADVANCED == state) {
// Save result for later.
WorkingSetMember* member = candidate.ws->get(id);
// Ensure that the BSONObj underlying the WorkingSetMember is owned in case we choose to
// return the results from the 'candidate' plan.
member->makeObjOwnedIfNeeded();
candidate.results.push(id);
// Once a plan returns enough results, stop working.
if (candidate.results.size() >= numResults) {
doneWorking = true;
}
} else if (PlanStage::IS_EOF == state) {
// First plan to hit EOF wins automatically. Stop evaluating other plans.
// Assumes that the ranking will pick this plan.
doneWorking = true;
} else if (PlanStage::NEED_YIELD == state) {
invariant(id == WorkingSet::INVALID_ID);
if (!yieldPolicy->canAutoYield()) {
throw WriteConflictException();
}
if (yieldPolicy->canAutoYield()) {
yieldPolicy->forceYield();
}
if (!(tryYield(yieldPolicy)).isOK()) {
return false;
}
} else if (PlanStage::NEED_TIME != state) {
// FAILURE or DEAD. Do we want to just tank that plan and try the rest? We
// probably want to fail globally as this shouldn't happen anyway.
candidate.failed = true;
++_failureCount;
// Propagate most recent seen failure to parent.
if (PlanStage::FAILURE == state) {
_statusMemberId = id;
}
if (_failureCount == _candidates.size()) {
_failure = true;
return false;
}
}
}
return !doneWorking;
}
// Set "toReturn" when NEED_YIELD.
PlanStage::StageState NearStage::bufferNext(WorkingSetID* toReturn, Status* error) {
//
// Try to retrieve the next covered member
//
if (!_nextInterval) {
StatusWith<CoveredInterval*> intervalStatus =
nextInterval(getOpCtx(), _workingSet, _collection);
if (!intervalStatus.isOK()) {
_searchState = SearchState_Finished;
*error = intervalStatus.getStatus();
return PlanStage::FAILURE;
}
if (NULL == intervalStatus.getValue()) {
_searchState = SearchState_Finished;
return PlanStage::IS_EOF;
}
// CoveredInterval and its child stage are owned by _childrenIntervals
_childrenIntervals.push_back(intervalStatus.getValue());
_nextInterval = _childrenIntervals.back();
_specificStats.intervalStats.emplace_back();
_nextIntervalStats = &_specificStats.intervalStats.back();
_nextIntervalStats->minDistanceAllowed = _nextInterval->minDistance;
_nextIntervalStats->maxDistanceAllowed = _nextInterval->maxDistance;
_nextIntervalStats->inclusiveMaxDistanceAllowed = _nextInterval->inclusiveMax;
}
WorkingSetID nextMemberID;
PlanStage::StageState intervalState = _nextInterval->covering->work(&nextMemberID);
if (PlanStage::IS_EOF == intervalState) {
_searchState = SearchState_Advancing;
return PlanStage::NEED_TIME;
} else if (PlanStage::FAILURE == intervalState) {
*error = WorkingSetCommon::getMemberStatus(*_workingSet->get(nextMemberID));
return intervalState;
} else if (PlanStage::NEED_YIELD == intervalState) {
*toReturn = nextMemberID;
return intervalState;
} else if (PlanStage::ADVANCED != intervalState) {
return intervalState;
}
//
// Try to buffer the next covered member
//
WorkingSetMember* nextMember = _workingSet->get(nextMemberID);
// The child stage may not dedup so we must dedup them ourselves.
if (_nextInterval->dedupCovering && nextMember->hasLoc()) {
if (_seenDocuments.end() != _seenDocuments.find(nextMember->loc)) {
_workingSet->free(nextMemberID);
return PlanStage::NEED_TIME;
}
}
++_nextIntervalStats->numResultsBuffered;
StatusWith<double> distanceStatus = computeDistance(nextMember);
if (!distanceStatus.isOK()) {
_searchState = SearchState_Finished;
*error = distanceStatus.getStatus();
return PlanStage::FAILURE;
}
// If the member's distance is in the current distance interval, add it to our buffered
// results.
double memberDistance = distanceStatus.getValue();
// Ensure that the BSONObj underlying the WorkingSetMember is owned in case we yield.
nextMember->makeObjOwnedIfNeeded();
_resultBuffer.push(SearchResult(nextMemberID, memberDistance));
// Store the member's RecordId, if available, for quick invalidation
if (nextMember->hasLoc()) {
_seenDocuments.insert(std::make_pair(nextMember->loc, nextMemberID));
}
return PlanStage::NEED_TIME;
}
PlanStage::StageState DeleteStage::work(WorkingSetID* out) {
++_commonStats.works;
// Adds the amount of time taken by work() to executionTimeMillis.
ScopedTimer timer(&_commonStats.executionTimeMillis);
if (isEOF()) {
return PlanStage::IS_EOF;
}
invariant(_collection); // If isEOF() returns false, we must have a collection.
// It is possible that after a delete was executed, a WriteConflictException occurred
// and prevented us from returning ADVANCED with the old version of the document.
if (_idReturning != WorkingSet::INVALID_ID) {
// We should only get here if we were trying to return something before.
invariant(_params.returnDeleted);
WorkingSetMember* member = _ws->get(_idReturning);
invariant(member->getState() == WorkingSetMember::OWNED_OBJ);
*out = _idReturning;
_idReturning = WorkingSet::INVALID_ID;
++_commonStats.advanced;
return PlanStage::ADVANCED;
}
// Either retry the last WSM we worked on or get a new one from our child.
WorkingSetID id;
StageState status;
if (_idRetrying == WorkingSet::INVALID_ID) {
status = child()->work(&id);
} else {
status = ADVANCED;
id = _idRetrying;
_idRetrying = WorkingSet::INVALID_ID;
}
if (PlanStage::ADVANCED == status) {
WorkingSetMember* member = _ws->get(id);
// We want to free this member when we return, unless we need to retry it.
ScopeGuard memberFreer = MakeGuard(&WorkingSet::free, _ws, id);
if (!member->hasLoc()) {
// We expect to be here because of an invalidation causing a force-fetch, and
// doc-locking storage engines do not issue invalidations.
++_specificStats.nInvalidateSkips;
++_commonStats.needTime;
return PlanStage::NEED_TIME;
}
RecordId rloc = member->loc;
// Deletes can't have projections. This means that covering analysis will always add
// a fetch. We should always get fetched data, and never just key data.
invariant(member->hasObj());
try {
// If the snapshot changed, then we have to make sure we have the latest copy of the
// doc and that it still matches.
std::unique_ptr<RecordCursor> cursor;
if (getOpCtx()->recoveryUnit()->getSnapshotId() != member->obj.snapshotId()) {
cursor = _collection->getCursor(getOpCtx());
if (!WorkingSetCommon::fetch(getOpCtx(), _ws, id, cursor)) {
// Doc is already deleted. Nothing more to do.
++_commonStats.needTime;
return PlanStage::NEED_TIME;
}
// Make sure the re-fetched doc still matches the predicate.
if (_params.canonicalQuery &&
!_params.canonicalQuery->root()->matchesBSON(member->obj.value(), NULL)) {
// Doesn't match.
++_commonStats.needTime;
return PlanStage::NEED_TIME;
}
}
// Ensure that the BSONObj underlying the WorkingSetMember is owned because saveState()
// is allowed to free the memory.
if (_params.returnDeleted) {
member->makeObjOwnedIfNeeded();
}
// TODO: Do we want to buffer docs and delete them in a group rather than
// saving/restoring state repeatedly?
try {
if (supportsDocLocking()) {
// Doc-locking engines require this before saveState() since they don't use
// invalidations.
WorkingSetCommon::prepareForSnapshotChange(_ws);
}
child()->saveState();
} catch (const WriteConflictException& wce) {
std::terminate();
}
if (_params.returnDeleted) {
// Save a copy of the document that is about to get deleted.
BSONObj deletedDoc = member->obj.value();
member->obj.setValue(deletedDoc.getOwned());
member->loc = RecordId();
member->transitionToOwnedObj();
}
// Do the write, unless this is an explain.
if (!_params.isExplain) {
WriteUnitOfWork wunit(getOpCtx());
_collection->deleteDocument(getOpCtx(), rloc);
wunit.commit();
}
++_specificStats.docsDeleted;
} catch (const WriteConflictException& wce) {
// Ensure that the BSONObj underlying the WorkingSetMember is owned because it may be
// freed when we yield.
member->makeObjOwnedIfNeeded();
_idRetrying = id;
memberFreer.Dismiss(); // Keep this member around so we can retry deleting it.
*out = WorkingSet::INVALID_ID;
_commonStats.needYield++;
return NEED_YIELD;
}
// As restoreState may restore (recreate) cursors, cursors are tied to the
// transaction in which they are created, and a WriteUnitOfWork is a
// transaction, make sure to restore the state outside of the WritUnitOfWork.
try {
child()->restoreState();
} catch (const WriteConflictException& wce) {
// Note we don't need to retry anything in this case since the delete already
// was committed. However, we still need to return the deleted document
// (if it was requested).
if (_params.returnDeleted) {
// member->obj should refer to the deleted document.
invariant(member->getState() == WorkingSetMember::OWNED_OBJ);
_idReturning = id;
// Keep this member around so that we can return it on the next work() call.
memberFreer.Dismiss();
}
*out = WorkingSet::INVALID_ID;
_commonStats.needYield++;
return NEED_YIELD;
}
if (_params.returnDeleted) {
// member->obj should refer to the deleted document.
invariant(member->getState() == WorkingSetMember::OWNED_OBJ);
memberFreer.Dismiss(); // Keep this member around so we can return it.
*out = id;
++_commonStats.advanced;
return PlanStage::ADVANCED;
}
++_commonStats.needTime;
return PlanStage::NEED_TIME;
} else if (PlanStage::FAILURE == status || PlanStage::DEAD == status) {
*out = id;
// If a stage fails, it may create a status WSM to indicate why it failed, in which case
// 'id' is valid. If ID is invalid, we create our own error message.
if (WorkingSet::INVALID_ID == id) {
const std::string errmsg = "delete stage failed to read in results from child";
*out = WorkingSetCommon::allocateStatusMember(
_ws, Status(ErrorCodes::InternalError, errmsg));
}
return status;
} else if (PlanStage::NEED_TIME == status) {
++_commonStats.needTime;
} else if (PlanStage::NEED_YIELD == status) {
*out = id;
++_commonStats.needYield;
}
return status;
}
PlanStage::StageState TextOrStage::addTerm(WorkingSetID wsid, WorkingSetID* out) {
WorkingSetMember* wsm = _ws->get(wsid);
invariant(wsm->getState() == WorkingSetMember::RID_AND_IDX);
invariant(1 == wsm->keyData.size());
const IndexKeyDatum newKeyData = wsm->keyData.back(); // copy to keep it around.
TextRecordData* textRecordData = &_scores[wsm->recordId];
if (textRecordData->score < 0) {
// We have already rejected this document for not matching the filter.
invariant(WorkingSet::INVALID_ID == textRecordData->wsid);
_ws->free(wsid);
return NEED_TIME;
}
if (WorkingSet::INVALID_ID == textRecordData->wsid) {
// We haven't seen this RecordId before.
invariant(textRecordData->score == 0);
bool shouldKeep = true;
if (_filter) {
// We have not seen this document before and need to apply a filter.
bool wasDeleted = false;
try {
TextMatchableDocument tdoc(getOpCtx(),
newKeyData.indexKeyPattern,
newKeyData.keyData,
_ws,
wsid,
_recordCursor);
shouldKeep = _filter->matches(&tdoc);
} catch (const WriteConflictException& wce) {
// Ensure that the BSONObj underlying the WorkingSetMember is owned because it may
// be freed when we yield.
wsm->makeObjOwnedIfNeeded();
_idRetrying = wsid;
*out = WorkingSet::INVALID_ID;
return NEED_YIELD;
} catch (const TextMatchableDocument::DocumentDeletedException&) {
// We attempted to fetch the document but decided it should be excluded from the
// result set.
shouldKeep = false;
wasDeleted = true;
}
if (wasDeleted || wsm->hasObj()) {
++_specificStats.fetches;
}
}
if (shouldKeep && !wsm->hasObj()) {
// Our parent expects RID_AND_OBJ members, so we fetch the document here if we haven't
// already.
try {
shouldKeep = WorkingSetCommon::fetch(getOpCtx(), _ws, wsid, _recordCursor);
++_specificStats.fetches;
} catch (const WriteConflictException& wce) {
wsm->makeObjOwnedIfNeeded();
_idRetrying = wsid;
*out = WorkingSet::INVALID_ID;
return NEED_YIELD;
}
}
if (!shouldKeep) {
_ws->free(wsid);
textRecordData->score = -1;
return NEED_TIME;
}
textRecordData->wsid = wsid;
// Ensure that the BSONObj underlying the WorkingSetMember is owned in case we yield.
wsm->makeObjOwnedIfNeeded();
} else {
// We already have a working set member for this RecordId. Free the new WSM and retrieve the
// old one. Note that since we don't keep all index keys, we could get a score that doesn't
// match the document, but this has always been a problem.
// TODO something to improve the situation.
invariant(wsid != textRecordData->wsid);
_ws->free(wsid);
wsm = _ws->get(textRecordData->wsid);
}
// Locate score within possibly compound key: {prefix,term,score,suffix}.
BSONObjIterator keyIt(newKeyData.keyData);
for (unsigned i = 0; i < _ftsSpec.numExtraBefore(); i++) {
keyIt.next();
}
keyIt.next(); // Skip past 'term'.
BSONElement scoreElement = keyIt.next();
double documentTermScore = scoreElement.number();
// Aggregate relevance score, term keys.
textRecordData->score += documentTermScore;
return NEED_TIME;
}
PlanStage::StageState UpdateStage::doWork(WorkingSetID* out) {
if (isEOF()) {
return PlanStage::IS_EOF;
}
if (doneUpdating()) {
// Even if we're done updating, we may have some inserting left to do.
if (needInsert()) {
// TODO we may want to handle WriteConflictException here. Currently we bounce it
// out to a higher level since if this WCEs it is likely that we raced with another
// upsert that may have matched our query, and therefore this may need to perform an
// update rather than an insert. Bouncing to the higher level allows restarting the
// query in this case.
doInsert();
invariant(isEOF());
if (_params.request->shouldReturnNewDocs()) {
// Want to return the document we just inserted, create it as a WorkingSetMember
// so that we can return it.
BSONObj newObj = _specificStats.objInserted;
*out = _ws->allocate();
WorkingSetMember* member = _ws->get(*out);
member->obj = Snapshotted<BSONObj>(getOpCtx()->recoveryUnit()->getSnapshotId(),
newObj.getOwned());
member->transitionToOwnedObj();
return PlanStage::ADVANCED;
}
}
// At this point either we're done updating and there was no insert to do,
// or we're done updating and we're done inserting. Either way, we're EOF.
invariant(isEOF());
return PlanStage::IS_EOF;
}
// If we're here, then we still have to ask for results from the child and apply
// updates to them. We should only get here if the collection exists.
invariant(_collection);
// It is possible that after an update was applied, a WriteConflictException
// occurred and prevented us from returning ADVANCED with the requested version
// of the document.
if (_idReturning != WorkingSet::INVALID_ID) {
// We should only get here if we were trying to return something before.
invariant(_params.request->shouldReturnAnyDocs());
WorkingSetMember* member = _ws->get(_idReturning);
invariant(member->getState() == WorkingSetMember::OWNED_OBJ);
*out = _idReturning;
_idReturning = WorkingSet::INVALID_ID;
return PlanStage::ADVANCED;
}
// Either retry the last WSM we worked on or get a new one from our child.
WorkingSetID id;
StageState status;
if (_idRetrying == WorkingSet::INVALID_ID) {
status = child()->work(&id);
} else {
status = ADVANCED;
id = _idRetrying;
_idRetrying = WorkingSet::INVALID_ID;
}
if (PlanStage::ADVANCED == status) {
// Need to get these things from the result returned by the child.
RecordId recordId;
WorkingSetMember* member = _ws->get(id);
// We want to free this member when we return, unless we need to retry updating or returning
// it.
ScopeGuard memberFreer = MakeGuard(&WorkingSet::free, _ws, id);
if (!member->hasRecordId()) {
// We expect to be here because of an invalidation causing a force-fetch.
++_specificStats.nInvalidateSkips;
return PlanStage::NEED_TIME;
}
recordId = member->recordId;
// Updates can't have projections. This means that covering analysis will always add
// a fetch. We should always get fetched data, and never just key data.
invariant(member->hasObj());
// We fill this with the new RecordIds of moved doc so we don't double-update.
if (_updatedRecordIds && _updatedRecordIds->count(recordId) > 0) {
// Found a RecordId that refers to a document we had already updated. Note that
// we can never remove from _updatedRecordIds because updates by other clients
// could cause us to encounter a document again later.
return PlanStage::NEED_TIME;
}
bool docStillMatches;
try {
docStillMatches = write_stage_common::ensureStillMatches(
_collection, getOpCtx(), _ws, id, _params.canonicalQuery);
} catch (const WriteConflictException&) {
// There was a problem trying to detect if the document still exists, so retry.
memberFreer.Dismiss();
return prepareToRetryWSM(id, out);
}
if (!docStillMatches) {
// Either the document has been deleted, or it has been updated such that it no longer
// matches the predicate.
if (shouldRestartUpdateIfNoLongerMatches(_params)) {
throw WriteConflictException();
}
return PlanStage::NEED_TIME;
}
// Ensure that the BSONObj underlying the WorkingSetMember is owned because saveState()
// is allowed to free the memory.
member->makeObjOwnedIfNeeded();
// Save state before making changes
WorkingSetCommon::prepareForSnapshotChange(_ws);
try {
child()->saveState();
} catch (const WriteConflictException&) {
std::terminate();
}
// If we care about the pre-updated version of the doc, save it out here.
BSONObj oldObj;
if (_params.request->shouldReturnOldDocs()) {
oldObj = member->obj.value().getOwned();
}
BSONObj newObj;
try {
// Do the update, get us the new version of the doc.
newObj = transformAndUpdate(member->obj, recordId);
} catch (const WriteConflictException&) {
memberFreer.Dismiss(); // Keep this member around so we can retry updating it.
return prepareToRetryWSM(id, out);
}
// Set member's obj to be the doc we want to return.
if (_params.request->shouldReturnAnyDocs()) {
if (_params.request->shouldReturnNewDocs()) {
member->obj = Snapshotted<BSONObj>(getOpCtx()->recoveryUnit()->getSnapshotId(),
newObj.getOwned());
} else {
invariant(_params.request->shouldReturnOldDocs());
member->obj.setValue(oldObj);
}
member->recordId = RecordId();
member->transitionToOwnedObj();
}
// This should be after transformAndUpdate to make sure we actually updated this doc.
++_specificStats.nMatched;
// Restore state after modification
// As restoreState may restore (recreate) cursors, make sure to restore the
// state outside of the WritUnitOfWork.
try {
child()->restoreState();
} catch (const WriteConflictException&) {
// Note we don't need to retry updating anything in this case since the update
// already was committed. However, we still need to return the updated document
// (if it was requested).
if (_params.request->shouldReturnAnyDocs()) {
// member->obj should refer to the document we want to return.
invariant(member->getState() == WorkingSetMember::OWNED_OBJ);
_idReturning = id;
// Keep this member around so that we can return it on the next work() call.
memberFreer.Dismiss();
}
*out = WorkingSet::INVALID_ID;
return NEED_YIELD;
}
if (_params.request->shouldReturnAnyDocs()) {
// member->obj should refer to the document we want to return.
invariant(member->getState() == WorkingSetMember::OWNED_OBJ);
memberFreer.Dismiss(); // Keep this member around so we can return it.
*out = id;
return PlanStage::ADVANCED;
}
return PlanStage::NEED_TIME;
} else if (PlanStage::IS_EOF == status) {
// The child is out of results, but we might not be done yet because we still might
// have to do an insert.
return PlanStage::NEED_TIME;
} else if (PlanStage::FAILURE == status) {
*out = id;
// If a stage fails, it may create a status WSM to indicate why it failed, in which case
// 'id' is valid. If ID is invalid, we create our own error message.
if (WorkingSet::INVALID_ID == id) {
const std::string errmsg = "update stage failed to read in results from child";
*out = WorkingSetCommon::allocateStatusMember(
_ws, Status(ErrorCodes::InternalError, errmsg));
return PlanStage::FAILURE;
}
return status;
} else if (PlanStage::NEED_YIELD == status) {
*out = id;
}
return status;
}
PlanStage::StageState FetchStage::doWork(WorkingSetID* out) {
if (isEOF()) {
return PlanStage::IS_EOF;
}
// Either retry the last WSM we worked on or get a new one from our child.
WorkingSetID id;
StageState status;
if (_idRetrying == WorkingSet::INVALID_ID) {
status = child()->work(&id);
} else {
status = ADVANCED;
id = _idRetrying;
_idRetrying = WorkingSet::INVALID_ID;
}
if (PlanStage::ADVANCED == status) {
WorkingSetMember* member = _ws->get(id);
// If there's an obj there, there is no fetching to perform.
if (member->hasObj()) {
++_specificStats.alreadyHasObj;
} else {
// We need a valid loc to fetch from and this is the only state that has one.
verify(WorkingSetMember::LOC_AND_IDX == member->getState());
verify(member->hasLoc());
try {
if (!_cursor)
_cursor = _collection->getCursor(getOpCtx());
if (auto fetcher = _cursor->fetcherForId(member->loc)) {
// There's something to fetch. Hand the fetcher off to the WSM, and pass up
// a fetch request.
_idRetrying = id;
member->setFetcher(fetcher.release());
*out = id;
return NEED_YIELD;
}
// The doc is already in memory, so go ahead and grab it. Now we have a RecordId
// as well as an unowned object
if (!WorkingSetCommon::fetch(getOpCtx(), _ws, id, _cursor)) {
_ws->free(id);
return NEED_TIME;
}
} catch (const WriteConflictException& wce) {
// Ensure that the BSONObj underlying the WorkingSetMember is owned because it may
// be freed when we yield.
member->makeObjOwnedIfNeeded();
_idRetrying = id;
*out = WorkingSet::INVALID_ID;
return NEED_YIELD;
}
}
return returnIfMatches(member, id, out);
} else if (PlanStage::FAILURE == status || PlanStage::DEAD == status) {
*out = id;
// If a stage fails, it may create a status WSM to indicate why it
// failed, in which case 'id' is valid. If ID is invalid, we
// create our own error message.
if (WorkingSet::INVALID_ID == id) {
mongoutils::str::stream ss;
ss << "fetch stage failed to read in results from child";
Status status(ErrorCodes::InternalError, ss);
*out = WorkingSetCommon::allocateStatusMember(_ws, status);
}
return status;
} else if (PlanStage::NEED_YIELD == status) {
*out = id;
}
return status;
}
PlanStage::StageState MergeSortStage::doWork(WorkingSetID* out) {
if (isEOF()) {
return PlanStage::IS_EOF;
}
if (!_noResultToMerge.empty()) {
// We have some child that we don't have a result from. Each child must have a result
// in order to pick the minimum result among all our children. Work a child.
PlanStage* child = _noResultToMerge.front();
WorkingSetID id = WorkingSet::INVALID_ID;
StageState code = child->work(&id);
if (PlanStage::ADVANCED == code) {
WorkingSetMember* member = _ws->get(id);
// If we're deduping...
if (_dedup) {
if (!member->hasRecordId()) {
// Can't dedup data unless there's a RecordId. We go ahead and use its
// result.
_noResultToMerge.pop();
} else {
++_specificStats.dupsTested;
// ...and there's a RecordId and and we've seen the RecordId before
if (_seen.end() != _seen.find(member->recordId)) {
// ...drop it.
_ws->free(id);
++_specificStats.dupsDropped;
return PlanStage::NEED_TIME;
} else {
// Otherwise, note that we've seen it.
_seen.insert(member->recordId);
// We're going to use the result from the child, so we remove it from
// the queue of children without a result.
_noResultToMerge.pop();
}
}
} else {
// Not deduping. We use any result we get from the child. Remove the child
// from the queue of things without a result.
_noResultToMerge.pop();
}
// Store the result in our list.
StageWithValue value;
value.id = id;
value.stage = child;
// Ensure that the BSONObj underlying the WorkingSetMember is owned in case we yield.
member->makeObjOwnedIfNeeded();
_mergingData.push_front(value);
// Insert the result (indirectly) into our priority queue.
_merging.push(_mergingData.begin());
return PlanStage::NEED_TIME;
} else if (PlanStage::IS_EOF == code) {
// There are no more results possible from this child. Don't bother with it
// anymore.
_noResultToMerge.pop();
return PlanStage::NEED_TIME;
} else if (PlanStage::FAILURE == code || PlanStage::DEAD == code) {
*out = id;
// If a stage fails, it may create a status WSM to indicate why it
// failed, in which case 'id' is valid. If ID is invalid, we
// create our own error message.
if (WorkingSet::INVALID_ID == id) {
mongoutils::str::stream ss;
ss << "merge sort stage failed to read in results from child";
Status status(ErrorCodes::InternalError, ss);
*out = WorkingSetCommon::allocateStatusMember(_ws, status);
}
return code;
} else {
if (PlanStage::NEED_YIELD == code) {
*out = id;
}
return code;
}
}
// If we're here, for each non-EOF child, we have a valid WSID.
verify(!_merging.empty());
// Get the 'min' WSID. _merging is a priority queue so its top is the smallest.
MergingRef top = _merging.top();
_merging.pop();
// Since we're returning the WSID that came from top->stage, we need to work(...) it again
// to get a new result.
_noResultToMerge.push(top->stage);
// Save the ID that we're returning and remove the returned result from our data.
WorkingSetID idToTest = top->id;
_mergingData.erase(top);
// Return the min.
*out = idToTest;
return PlanStage::ADVANCED;
}
Status CachedPlanStage::pickBestPlan(PlanYieldPolicy* yieldPolicy) {
// Adds the amount of time taken by pickBestPlan() to executionTimeMillis. There's lots of
// execution work that happens here, so this is needed for the time accounting to
// make sense.
ScopedTimer timer(getClock(), &_commonStats.executionTimeMillis);
// During plan selection, the list of indices we are using to plan must remain stable, so the
// query will die during yield recovery if any index has been dropped. However, once plan
// selection completes successfully, we no longer need all indices to stick around. The selected
// plan should safely die on yield recovery if it is using the dropped index.
//
// Dismiss the requirement that no indices can be dropped when this method returns.
ON_BLOCK_EXIT([this] { releaseAllIndicesRequirement(); });
// If we work this many times during the trial period, then we will replan the
// query from scratch.
size_t maxWorksBeforeReplan =
static_cast<size_t>(internalQueryCacheEvictionRatio * _decisionWorks);
// The trial period ends without replanning if the cached plan produces this many results.
size_t numResults = MultiPlanStage::getTrialPeriodNumToReturn(*_canonicalQuery);
for (size_t i = 0; i < maxWorksBeforeReplan; ++i) {
// Might need to yield between calls to work due to the timer elapsing.
Status yieldStatus = tryYield(yieldPolicy);
if (!yieldStatus.isOK()) {
return yieldStatus;
}
WorkingSetID id = WorkingSet::INVALID_ID;
PlanStage::StageState state = child()->work(&id);
if (PlanStage::ADVANCED == state) {
// Save result for later.
WorkingSetMember* member = _ws->get(id);
// Ensure that the BSONObj underlying the WorkingSetMember is owned in case we yield.
member->makeObjOwnedIfNeeded();
_results.push(id);
if (_results.size() >= numResults) {
// Once a plan returns enough results, stop working. Update cache with stats
// from this run and return.
updatePlanCache();
return Status::OK();
}
} else if (PlanStage::IS_EOF == state) {
// Cached plan hit EOF quickly enough. No need to replan. Update cache with stats
// from this run and return.
updatePlanCache();
return Status::OK();
} else if (PlanStage::NEED_YIELD == state) {
invariant(id == WorkingSet::INVALID_ID);
if (!yieldPolicy->canAutoYield()) {
throw WriteConflictException();
}
if (yieldPolicy->canAutoYield()) {
yieldPolicy->forceYield();
}
Status yieldStatus = tryYield(yieldPolicy);
if (!yieldStatus.isOK()) {
return yieldStatus;
}
} else if (PlanStage::FAILURE == state) {
// On failure, fall back to replanning the whole query. We neither evict the
// existing cache entry nor cache the result of replanning.
BSONObj statusObj;
WorkingSetCommon::getStatusMemberObject(*_ws, id, &statusObj);
LOG(1) << "Execution of cached plan failed, falling back to replan."
<< " query: " << redact(_canonicalQuery->toStringShort())
<< " planSummary: " << Explain::getPlanSummary(child().get())
<< " status: " << redact(statusObj);
const bool shouldCache = false;
return replan(yieldPolicy, shouldCache);
} else {
invariant(PlanStage::NEED_TIME == state);
}
}
// If we're here, the trial period took more than 'maxWorksBeforeReplan' work cycles. This
// plan is taking too long, so we replan from scratch.
LOG(1) << "Execution of cached plan required " << maxWorksBeforeReplan
<< " works, but was originally cached with only " << _decisionWorks
<< " works. Evicting cache entry and replanning query: "
<< redact(_canonicalQuery->toStringShort())
<< " plan summary before replan: " << Explain::getPlanSummary(child().get());
const bool shouldCache = true;
return replan(yieldPolicy, shouldCache);
}
PlanStage::StageState FetchStage::doWork(WorkingSetID* out) {
if (isEOF()) {
return PlanStage::IS_EOF;
}
// Either retry the last WSM we worked on or get a new one from our child.
WorkingSetID id;
StageState status;
if (_idRetrying == WorkingSet::INVALID_ID) {
status = child()->work(&id);
} else {
status = ADVANCED;
id = _idRetrying;
_idRetrying = WorkingSet::INVALID_ID;
}
if (PlanStage::ADVANCED == status) {
WorkingSetMember* member = _ws->get(id);
// If there's an obj there, there is no fetching to perform.
if (member->hasObj()) {
++_specificStats.alreadyHasObj;
} else {
// We need a valid RecordId to fetch from and this is the only state that has one.
verify(WorkingSetMember::RID_AND_IDX == member->getState());
verify(member->hasRecordId());
try {
if (!_cursor)
_cursor = _collection->getCursor(getOpCtx());
if (auto fetcher = _cursor->fetcherForId(member->recordId)) {
// There's something to fetch. Hand the fetcher off to the WSM, and pass up
// a fetch request.
_idRetrying = id;
member->setFetcher(fetcher.release());
*out = id;
return NEED_YIELD;
}
// The doc is already in memory, so go ahead and grab it. Now we have a RecordId
// as well as an unowned object
if (!WorkingSetCommon::fetch(getOpCtx(), _ws, id, _cursor)) {
_ws->free(id);
return NEED_TIME;
}
} catch (const WriteConflictException&) {
// Ensure that the BSONObj underlying the WorkingSetMember is owned because it may
// be freed when we yield.
member->makeObjOwnedIfNeeded();
_idRetrying = id;
*out = WorkingSet::INVALID_ID;
return NEED_YIELD;
}
}
return returnIfMatches(member, id, out);
} else if (PlanStage::FAILURE == status || PlanStage::DEAD == status) {
// The stage which produces a failure is responsible for allocating a working set member
// with error details.
invariant(WorkingSet::INVALID_ID != id);
*out = id;
return status;
} else if (PlanStage::NEED_YIELD == status) {
*out = id;
}
return status;
}