// static
bool WorkingSetCommon::fetch(OperationContext* txn,
WorkingSet* workingSet,
WorkingSetID id,
unowned_ptr<SeekableRecordCursor> cursor) {
WorkingSetMember* member = workingSet->get(id);
// The RecordFetcher should already have been transferred out of the WSM and used.
invariant(!member->hasFetcher());
// We should have a RecordId but need to retrieve the obj. Get the obj now and reset all WSM
// state appropriately.
invariant(member->hasRecordId());
member->obj.reset();
auto record = cursor->seekExact(member->recordId);
if (!record) {
return false;
}
member->obj = {txn->recoveryUnit()->getSnapshotId(), record->data.releaseToBson()};
if (member->isSuspicious) {
// Make sure that all of the keyData is still valid for this copy of the document.
// This ensures both that index-provided filters and sort orders still hold.
// TODO provide a way for the query planner to opt out of this checking if it is
// unneeded due to the structure of the plan.
invariant(!member->keyData.empty());
for (size_t i = 0; i < member->keyData.size(); i++) {
BSONObjSet keys;
// There's no need to compute the prefixes of the indexed fields that cause the index to
// be multikey when ensuring the keyData is still valid.
MultikeyPaths* multikeyPaths = nullptr;
member->keyData[i].index->getKeys(member->obj.value(), &keys, multikeyPaths);
if (!keys.count(member->keyData[i].keyData)) {
// document would no longer be at this position in the index.
return false;
}
}
member->isSuspicious = false;
}
member->keyData.clear();
workingSet->transitionToRecordIdAndObj(id);
return true;
}
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);
}
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.
candidate.results.push_back(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) {
if (id == WorkingSet::INVALID_ID) {
if (!yieldPolicy->allowedToYield())
throw WriteConflictException();
}
else {
WorkingSetMember* member = candidate.ws->get(id);
invariant(member->hasFetcher());
// Transfer ownership of the fetcher and yield.
_fetcher.reset(member->releaseFetcher());
}
if (yieldPolicy->allowedToYield()) {
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;
}
PlanExecutor::ExecState PlanExecutor::getNextImpl(Snapshotted<BSONObj>* objOut, RecordId* dlOut) {
if (MONGO_FAIL_POINT(planExecutorAlwaysFails)) {
Status status(ErrorCodes::OperationFailed,
str::stream() << "PlanExecutor hit planExecutorAlwaysFails fail point");
*objOut =
Snapshotted<BSONObj>(SnapshotId(), WorkingSetCommon::buildMemberStatusObject(status));
return PlanExecutor::FAILURE;
}
invariant(_currentState == kUsable);
if (isMarkedAsKilled()) {
if (NULL != objOut) {
Status status(ErrorCodes::OperationFailed,
str::stream() << "Operation aborted because: " << *_killReason);
*objOut = Snapshotted<BSONObj>(SnapshotId(),
WorkingSetCommon::buildMemberStatusObject(status));
}
return PlanExecutor::DEAD;
}
if (!_stash.empty()) {
invariant(objOut && !dlOut);
*objOut = {SnapshotId(), _stash.front()};
_stash.pop();
return PlanExecutor::ADVANCED;
}
// When a stage requests a yield for document fetch, it gives us back a RecordFetcher*
// to use to pull the record into memory. We take ownership of the RecordFetcher here,
// deleting it after we've had a chance to do the fetch. For timing-based yields, we
// just pass a NULL fetcher.
unique_ptr<RecordFetcher> fetcher;
// Incremented on every writeConflict, reset to 0 on any successful call to _root->work.
size_t writeConflictsInARow = 0;
for (;;) {
// These are the conditions which can cause us to yield:
// 1) The yield policy's timer elapsed, or
// 2) some stage requested a yield due to a document fetch, or
// 3) we need to yield and retry due to a WriteConflictException.
// In all cases, the actual yielding happens here.
if (_yieldPolicy->shouldYield()) {
if (!_yieldPolicy->yield(fetcher.get())) {
// A return of false from a yield should only happen if we've been killed during the
// yield.
invariant(isMarkedAsKilled());
if (NULL != objOut) {
Status status(ErrorCodes::OperationFailed,
str::stream() << "Operation aborted because: " << *_killReason);
*objOut = Snapshotted<BSONObj>(
SnapshotId(), WorkingSetCommon::buildMemberStatusObject(status));
}
return PlanExecutor::DEAD;
}
}
// We're done using the fetcher, so it should be freed. We don't want to
// use the same RecordFetcher twice.
fetcher.reset();
WorkingSetID id = WorkingSet::INVALID_ID;
PlanStage::StageState code = _root->work(&id);
if (code != PlanStage::NEED_YIELD)
writeConflictsInARow = 0;
if (PlanStage::ADVANCED == code) {
WorkingSetMember* member = _workingSet->get(id);
bool hasRequestedData = true;
if (NULL != objOut) {
if (WorkingSetMember::RID_AND_IDX == member->getState()) {
if (1 != member->keyData.size()) {
_workingSet->free(id);
hasRequestedData = false;
} else {
// TODO: currently snapshot ids are only associated with documents, and
// not with index keys.
*objOut = Snapshotted<BSONObj>(SnapshotId(), member->keyData[0].keyData);
}
} else if (member->hasObj()) {
*objOut = member->obj;
} else {
_workingSet->free(id);
hasRequestedData = false;
}
}
if (NULL != dlOut) {
if (member->hasRecordId()) {
*dlOut = member->recordId;
} else {
_workingSet->free(id);
hasRequestedData = false;
}
}
if (hasRequestedData) {
_workingSet->free(id);
return PlanExecutor::ADVANCED;
}
// This result didn't have the data the caller wanted, try again.
} else if (PlanStage::NEED_YIELD == code) {
if (id == WorkingSet::INVALID_ID) {
if (!_yieldPolicy->canAutoYield())
throw WriteConflictException();
CurOp::get(_opCtx)->debug().writeConflicts++;
writeConflictsInARow++;
WriteConflictException::logAndBackoff(
writeConflictsInARow, "plan execution", _nss.ns());
} else {
WorkingSetMember* member = _workingSet->get(id);
invariant(member->hasFetcher());
// Transfer ownership of the fetcher. Next time around the loop a yield will
// happen.
fetcher.reset(member->releaseFetcher());
}
// If we're allowed to, we will yield next time through the loop.
if (_yieldPolicy->canAutoYield())
_yieldPolicy->forceYield();
} else if (PlanStage::NEED_TIME == code) {
// Fall through to yield check at end of large conditional.
} else if (PlanStage::IS_EOF == code) {
if (shouldWaitForInserts()) {
const bool locksReacquiredAfterYield = waitForInserts();
if (locksReacquiredAfterYield) {
// There may be more results, try to get more data.
continue;
}
invariant(isMarkedAsKilled());
if (objOut) {
Status status(ErrorCodes::OperationFailed,
str::stream() << "Operation aborted because: " << *_killReason);
*objOut = Snapshotted<BSONObj>(
SnapshotId(), WorkingSetCommon::buildMemberStatusObject(status));
}
return PlanExecutor::DEAD;
} else {
return PlanExecutor::IS_EOF;
}
} else {
invariant(PlanStage::DEAD == code || PlanStage::FAILURE == code);
if (NULL != objOut) {
BSONObj statusObj;
WorkingSetCommon::getStatusMemberObject(*_workingSet, id, &statusObj);
*objOut = Snapshotted<BSONObj>(SnapshotId(), statusObj);
}
return (PlanStage::DEAD == code) ? PlanExecutor::DEAD : PlanExecutor::FAILURE;
}
}
}
PlanExecutor::ExecState PlanExecutor::getNext(BSONObj* objOut, DiskLoc* dlOut) {
if (_killed) { return PlanExecutor::DEAD; }
// When a stage requests a yield for document fetch, it gives us back a RecordFetcher*
// to use to pull the record into memory. We take ownership of the RecordFetcher here,
// deleting it after we've had a chance to do the fetch. For timing-based yields, we
// just pass a NULL fetcher.
boost::scoped_ptr<RecordFetcher> fetcher;
for (;;) {
// There are two conditions which cause us to yield if we have an YIELD_AUTO
// policy:
// 1) The yield policy's timer elapsed, or
// 2) some stage requested a yield due to a document fetch (NEED_FETCH).
// In both cases, the actual yielding happens here.
if (NULL != _yieldPolicy.get() && (_yieldPolicy->shouldYield()
|| NULL != fetcher.get())) {
// Here's where we yield.
_yieldPolicy->yield(fetcher.get());
if (_killed) {
return PlanExecutor::DEAD;
}
}
// We're done using the fetcher, so it should be freed. We don't want to
// use the same RecordFetcher twice.
fetcher.reset();
WorkingSetID id = WorkingSet::INVALID_ID;
PlanStage::StageState code = _root->work(&id);
if (PlanStage::ADVANCED == code) {
// Fast count.
if (WorkingSet::INVALID_ID == id) {
invariant(NULL == objOut);
invariant(NULL == dlOut);
return PlanExecutor::ADVANCED;
}
WorkingSetMember* member = _workingSet->get(id);
bool hasRequestedData = true;
if (NULL != objOut) {
if (WorkingSetMember::LOC_AND_IDX == member->state) {
if (1 != member->keyData.size()) {
_workingSet->free(id);
hasRequestedData = false;
}
else {
*objOut = member->keyData[0].keyData;
}
}
else if (member->hasObj()) {
*objOut = member->obj;
}
else {
_workingSet->free(id);
hasRequestedData = false;
}
}
if (NULL != dlOut) {
if (member->hasLoc()) {
*dlOut = member->loc;
}
else {
_workingSet->free(id);
hasRequestedData = false;
}
}
if (hasRequestedData) {
_workingSet->free(id);
return PlanExecutor::ADVANCED;
}
// This result didn't have the data the caller wanted, try again.
}
else if (PlanStage::NEED_FETCH == code) {
// Yielding on a NEED_FETCH is handled above, so there's not much to do here.
// Just verify that the NEED_FETCH gave us back a WSM that is actually fetchable.
WorkingSetMember* member = _workingSet->get(id);
invariant(member->hasFetcher());
// Transfer ownership of the fetcher. Next time around the loop a yield will happen.
fetcher.reset(member->releaseFetcher());
}
else if (PlanStage::NEED_TIME == code) {
// Fall through to yield check at end of large conditional.
}
else if (PlanStage::IS_EOF == code) {
return PlanExecutor::IS_EOF;
}
else if (PlanStage::DEAD == code) {
return PlanExecutor::DEAD;
}
else {
verify(PlanStage::FAILURE == code);
if (NULL != objOut) {
WorkingSetCommon::getStatusMemberObject(*_workingSet, id, objOut);
}
return PlanExecutor::EXEC_ERROR;
}
}
}
PlanExecutor::ExecState PlanExecutor::getNextSnapshotted(Snapshotted<BSONObj>* objOut,
RecordId* dlOut) {
if (_killed) { return PlanExecutor::DEAD; }
// When a stage requests a yield for document fetch, it gives us back a RecordFetcher*
// to use to pull the record into memory. We take ownership of the RecordFetcher here,
// deleting it after we've had a chance to do the fetch. For timing-based yields, we
// just pass a NULL fetcher.
boost::scoped_ptr<RecordFetcher> fetcher;
// Incremented on every writeConflict, reset to 0 on any successful call to _root->work.
size_t writeConflictsInARow = 0;
for (;;) {
// These are the conditions which can cause us to yield:
// 1) The yield policy's timer elapsed, or
// 2) some stage requested a yield due to a document fetch, or
// 3) we need to yield and retry due to a WriteConflictException.
// In all cases, the actual yielding happens here.
if (_yieldPolicy->shouldYield()) {
_yieldPolicy->yield(fetcher.get());
if (_killed) {
return PlanExecutor::DEAD;
}
}
// We're done using the fetcher, so it should be freed. We don't want to
// use the same RecordFetcher twice.
fetcher.reset();
WorkingSetID id = WorkingSet::INVALID_ID;
PlanStage::StageState code = _root->work(&id);
if (code != PlanStage::NEED_YIELD)
writeConflictsInARow = 0;
if (PlanStage::ADVANCED == code) {
// Fast count.
if (WorkingSet::INVALID_ID == id) {
invariant(NULL == objOut);
invariant(NULL == dlOut);
return PlanExecutor::ADVANCED;
}
WorkingSetMember* member = _workingSet->get(id);
bool hasRequestedData = true;
if (NULL != objOut) {
if (WorkingSetMember::LOC_AND_IDX == member->state) {
if (1 != member->keyData.size()) {
_workingSet->free(id);
hasRequestedData = false;
}
else {
// TODO: currently snapshot ids are only associated with documents, and
// not with index keys.
*objOut = Snapshotted<BSONObj>(SnapshotId(),
member->keyData[0].keyData);
}
}
else if (member->hasObj()) {
*objOut = member->obj;
}
else {
_workingSet->free(id);
hasRequestedData = false;
}
}
if (NULL != dlOut) {
if (member->hasLoc()) {
*dlOut = member->loc;
}
else {
_workingSet->free(id);
hasRequestedData = false;
}
}
if (hasRequestedData) {
_workingSet->free(id);
return PlanExecutor::ADVANCED;
}
// This result didn't have the data the caller wanted, try again.
}
else if (PlanStage::NEED_YIELD == code) {
if (id == WorkingSet::INVALID_ID) {
if (!_yieldPolicy->allowedToYield()) throw WriteConflictException();
_opCtx->getCurOp()->debug().writeConflicts++;
writeConflictsInARow++;
WriteConflictException::logAndBackoff(writeConflictsInARow,
"plan execution",
_collection->ns().ns());
}
else {
WorkingSetMember* member = _workingSet->get(id);
invariant(member->hasFetcher());
// Transfer ownership of the fetcher. Next time around the loop a yield will
// happen.
fetcher.reset(member->releaseFetcher());
}
// If we're allowed to, we will yield next time through the loop.
if (_yieldPolicy->allowedToYield()) _yieldPolicy->forceYield();
}
else if (PlanStage::NEED_TIME == code) {
// Fall through to yield check at end of large conditional.
}
else if (PlanStage::IS_EOF == code) {
return PlanExecutor::IS_EOF;
}
else if (PlanStage::DEAD == code) {
return PlanExecutor::DEAD;
}
else {
verify(PlanStage::FAILURE == code);
if (NULL != objOut) {
BSONObj statusObj;
WorkingSetCommon::getStatusMemberObject(*_workingSet, id, &statusObj);
*objOut = Snapshotted<BSONObj>(SnapshotId(), statusObj);
}
return PlanExecutor::FAILURE;
}
}
}
