BSONObj getObj() const {
if (!WorkingSetCommon::fetchIfUnfetched(_txn, _ws, _id, _recordCursor))
throw DocumentDeletedException();
WorkingSetMember* member = _ws->get(_id);
// Make it owned since we are buffering results.
member->makeObjOwned();
return member->obj.value();
}
void WorkingSetCommon::prepareForSnapshotChange(WorkingSet* workingSet) {
dassert(supportsDocLocking());
for (auto id : workingSet->getAndClearYieldSensitiveIds()) {
if (workingSet->isFree(id)) {
continue;
}
// We may see the same member twice, so anything we do here should be idempotent.
WorkingSetMember* member = workingSet->get(id);
if (member->getState() == WorkingSetMember::LOC_AND_IDX) {
member->isSuspicious = true;
} else if (member->getState() == WorkingSetMember::LOC_AND_OBJ) {
// Need to make sure that the data is owned, as underlying storage can change during a
// yield.
member->makeObjOwned();
}
}
}
PlanStage::StageState FetchStage::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;
}
// 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;
_commonStats.needYield++;
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);
_commonStats.needTime++;
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->makeObjOwned();
_idRetrying = id;
*out = WorkingSet::INVALID_ID;
_commonStats.needYield++;
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_TIME == status) {
++_commonStats.needTime;
} else if (PlanStage::NEED_YIELD == status) {
++_commonStats.needYield;
*out = id;
}
return status;
}
PlanStage::StageState MergeSortStage::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;
}
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->hasLoc()) {
// 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 diskloc and and we've seen the RecordId before
if (_seen.end() != _seen.find(member->loc)) {
// ...drop it.
_ws->free(id);
++_commonStats.needTime;
++_specificStats.dupsDropped;
return PlanStage::NEED_TIME;
} else {
// Otherwise, note that we've seen it.
_seen.insert(member->loc);
// 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->makeObjOwned();
_mergingData.push_front(value);
// Insert the result (indirectly) into our priority queue.
_merging.push(_mergingData.begin());
++_commonStats.needTime;
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();
++_commonStats.needTime;
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_TIME == code) {
++_commonStats.needTime;
} else if (PlanStage::NEED_YIELD == code) {
*out = id;
++_commonStats.needYield;
}
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;
++_commonStats.advanced;
// But don't return it if it's flagged.
if (_ws->isFlagged(*out)) {
return PlanStage::NEED_TIME;
}
return PlanStage::ADVANCED;
}
PlanStage::StageState TextOrStage::addTerm(WorkingSetID wsid, WorkingSetID* out) {
WorkingSetMember* wsm = _ws->get(wsid);
invariant(wsm->getState() == WorkingSetMember::LOC_AND_IDX);
invariant(1 == wsm->keyData.size());
const IndexKeyDatum newKeyData = wsm->keyData.back(); // copy to keep it around.
TextRecordData* textRecordData = &_scores[wsm->loc];
double* documentAggregateScore = &textRecordData->score;
if (WorkingSet::INVALID_ID == textRecordData->wsid) {
// We haven't seen this RecordId before. Keep the working set member around (it may be
// force-fetched on saveState()).
textRecordData->wsid = wsid;
if (_filter) {
// We have not seen this document before and need to apply a filter.
bool shouldKeep;
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->makeObjOwned();
_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 (!shouldKeep) {
if (wasDeleted || wsm->hasObj()) {
// We had to fetch but we're not going to return it.
++_specificStats.fetches;
}
_ws->free(textRecordData->wsid);
textRecordData->wsid = WorkingSet::INVALID_ID;
*documentAggregateScore = -1;
return NEED_TIME;
}
}
} 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);
}
if (*documentAggregateScore < 0) {
// We have already rejected this document for not matching the filter.
return NEED_TIME;
}
// 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.
*documentAggregateScore += documentTermScore;
return NEED_TIME;
}
PlanStage::StageState AndSortedStage::getTargetLoc(WorkingSetID* out) {
verify(numeric_limits<size_t>::max() == _targetNode);
verify(WorkingSet::INVALID_ID == _targetId);
verify(RecordId() == _targetLoc);
// Pick one, and get a loc to work toward.
WorkingSetID id = WorkingSet::INVALID_ID;
StageState state = _children[0]->work(&id);
if (PlanStage::ADVANCED == state) {
WorkingSetMember* member = _ws->get(id);
// Maybe the child had an invalidation. We intersect RecordId(s) so we can't do anything
// with this WSM.
if (!member->hasLoc()) {
_ws->flagForReview(id);
return PlanStage::NEED_TIME;
}
verify(member->hasLoc());
// We have a value from one child to AND with.
_targetNode = 0;
_targetId = id;
_targetLoc = member->loc;
// Ensure that the BSONObj underlying the WorkingSetMember is owned in case we yield.
member->makeObjOwned();
// We have to AND with all other children.
for (size_t i = 1; i < _children.size(); ++i) {
_workingTowardRep.push(i);
}
++_commonStats.needTime;
return PlanStage::NEED_TIME;
} else if (PlanStage::IS_EOF == state) {
_isEOF = true;
return state;
} else if (PlanStage::FAILURE == state) {
*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 << "sorted AND stage failed to read in results from first child";
Status status(ErrorCodes::InternalError, ss);
*out = WorkingSetCommon::allocateStatusMember(_ws, status);
}
_isEOF = true;
return state;
} else {
if (PlanStage::NEED_TIME == state) {
++_commonStats.needTime;
} else if (PlanStage::NEED_YIELD == state) {
++_commonStats.needYield;
*out = id;
}
// NEED_TIME, NEED_YIELD.
return state;
}
}
PlanStage::StageState AndSortedStage::moveTowardTargetLoc(WorkingSetID* out) {
verify(numeric_limits<size_t>::max() != _targetNode);
verify(WorkingSet::INVALID_ID != _targetId);
// We have nodes that haven't hit _targetLoc yet.
size_t workingChildNumber = _workingTowardRep.front();
auto& next = _children[workingChildNumber];
WorkingSetID id = WorkingSet::INVALID_ID;
StageState state = next->work(&id);
if (PlanStage::ADVANCED == state) {
WorkingSetMember* member = _ws->get(id);
// Maybe the child had an invalidation. We intersect RecordId(s) so we can't do anything
// with this WSM.
if (!member->hasLoc()) {
_ws->flagForReview(id);
return PlanStage::NEED_TIME;
}
verify(member->hasLoc());
if (member->loc == _targetLoc) {
// The front element has hit _targetLoc. Don't move it forward anymore/work on
// another element.
_workingTowardRep.pop();
AndCommon::mergeFrom(_ws, _targetId, *member);
_ws->free(id);
if (0 == _workingTowardRep.size()) {
WorkingSetID toReturn = _targetId;
_targetNode = numeric_limits<size_t>::max();
_targetId = WorkingSet::INVALID_ID;
_targetLoc = RecordId();
*out = toReturn;
++_commonStats.advanced;
return PlanStage::ADVANCED;
}
// More children need to be advanced to _targetLoc.
++_commonStats.needTime;
return PlanStage::NEED_TIME;
} else if (member->loc < _targetLoc) {
// The front element of _workingTowardRep hasn't hit the thing we're AND-ing with
// yet. Try again later.
_ws->free(id);
++_commonStats.needTime;
return PlanStage::NEED_TIME;
} else {
// member->loc > _targetLoc.
// _targetLoc wasn't successfully AND-ed with the other sub-plans. We toss it and
// try AND-ing with the next value.
_specificStats.failedAnd[_targetNode]++;
_ws->free(_targetId);
_targetNode = workingChildNumber;
_targetLoc = member->loc;
_targetId = id;
// Ensure that the BSONObj underlying the WorkingSetMember is owned in case we yield.
member->makeObjOwned();
_workingTowardRep = std::queue<size_t>();
for (size_t i = 0; i < _children.size(); ++i) {
if (workingChildNumber != i) {
_workingTowardRep.push(i);
}
}
// Need time to chase after the new _targetLoc.
++_commonStats.needTime;
return PlanStage::NEED_TIME;
}
} else if (PlanStage::IS_EOF == state) {
_isEOF = true;
_ws->free(_targetId);
return state;
} else if (PlanStage::FAILURE == state || PlanStage::DEAD == state) {
*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 << "sorted AND stage failed to read in results from child " << workingChildNumber;
Status status(ErrorCodes::InternalError, ss);
*out = WorkingSetCommon::allocateStatusMember(_ws, status);
}
_isEOF = true;
_ws->free(_targetId);
return state;
} else {
if (PlanStage::NEED_TIME == state) {
++_commonStats.needTime;
} else if (PlanStage::NEED_YIELD == state) {
++_commonStats.needYield;
*out = id;
}
return state;
}
}