bool LockManager::unlock(LockRequest* request) {
invariant(request->lock);
// Fast path for decrementing multiple references of the same lock. It is safe to do this
// without locking, because 1) all calls for the same lock request must be done on the same
// thread and 2) if there are lock requests hanging of a given LockHead, then this lock
// will never disappear.
request->recursiveCount--;
if ((request->status == LockRequest::STATUS_GRANTED) && (request->recursiveCount > 0)) {
return false;
}
LockHead* lock = request->lock;
LockBucket* bucket = _getBucket(lock->resourceId);
scoped_spinlock scopedLock(bucket->mutex);
invariant(lock->grantedQueue != NULL);
invariant(lock->grantedModes != 0);
if (request->status == LockRequest::STATUS_WAITING) {
// This cancels a pending lock request
invariant(request->recursiveCount == 0);
lock->removeFromConflictQueue(request);
lock->changeConflictModeCount(request->mode, LockHead::Decrement);
}
else if (request->status == LockRequest::STATUS_CONVERTING) {
// This cancels a pending convert request
invariant(request->recursiveCount > 0);
// Lock only goes from GRANTED to CONVERTING, so cancelling the conversion request
// brings it back to the previous granted mode.
request->status = LockRequest::STATUS_GRANTED;
lock->conversionsCount--;
lock->changeGrantedModeCount(request->convertMode, LockHead::Decrement);
request->convertMode = MODE_NONE;
_onLockModeChanged(lock, lock->grantedCounts[request->convertMode] == 0);
}
else if (request->status == LockRequest::STATUS_GRANTED) {
// This releases a currently held lock and is the most common path, so it should be
// as efficient as possible.
invariant(request->recursiveCount == 0);
// Remove from the granted list
lock->removeFromGrantedQueue(request);
lock->changeGrantedModeCount(request->mode, LockHead::Decrement);
_onLockModeChanged(lock, lock->grantedCounts[request->mode] == 0);
}
else {
// Invalid request status
invariant(false);
}
return (request->recursiveCount == 0);
}
void LockManager::downgrade(LockRequest* request, LockMode newMode) {
invariant(request->lock);
invariant(request->status == LockRequest::STATUS_GRANTED);
invariant(request->recursiveCount > 0);
// The conflict set of the newMode should be a subset of the conflict set of the old mode.
// Can't downgrade from S -> IX for example.
invariant((LockConflictsTable[request->mode] | LockConflictsTable[newMode])
== LockConflictsTable[request->mode]);
LockHead* lock = request->lock;
LockBucket* bucket = _getBucket(lock->resourceId);
SimpleMutex::scoped_lock scopedLock(bucket->mutex);
invariant(lock->grantedQueueBegin != NULL);
invariant(lock->grantedQueueEnd != NULL);
invariant(lock->grantedModes != 0);
lock->changeGrantedModeCount(newMode, LockHead::Increment);
lock->changeGrantedModeCount(request->mode, LockHead::Decrement);
request->mode = newMode;
_onLockModeChanged(lock, true);
}
void OperationLatencyHistogram::increment(uint64_t latency, Command::ReadWriteType type) {
int bucket = _getBucket(latency);
switch (type) {
case Command::ReadWriteType::kRead:
_incrementData(latency, bucket, &_reads);
break;
case Command::ReadWriteType::kWrite:
_incrementData(latency, bucket, &_writes);
break;
case Command::ReadWriteType::kCommand:
_incrementData(latency, bucket, &_commands);
break;
default:
MONGO_UNREACHABLE;
}
}
LockResult LockManager::lock(const ResourceId& resId, LockRequest* request, LockMode mode) {
dassert(mode > MODE_NONE);
// Fast path for acquiring the same lock multiple times in modes, which are already covered
// by the current mode. It is safe to do this without locking, because 1) all calls for the
// same lock request must be done on the same thread and 2) if there are lock requests
// hanging off a given LockHead, then this lock will never disappear.
if ((LockConflictsTable[request->mode] | LockConflictsTable[mode]) ==
LockConflictsTable[request->mode]) {
request->recursiveCount++;
return LOCK_OK;
}
// TODO: For the time being we do not need conversions between unrelated lock modes (i.e.,
// modes which both add and remove to the conflicts set), so these are not implemented yet
// (e.g., S -> IX).
invariant((LockConflictsTable[request->mode] | LockConflictsTable[mode]) ==
LockConflictsTable[mode]);
LockBucket* bucket = _getBucket(resId);
scoped_spinlock scopedLock(bucket->mutex);
LockHead* lock;
LockHeadMap::iterator it = bucket->data.find(resId);
if (it == bucket->data.end()) {
// Lock is free (not on the map)
invariant(request->status == LockRequest::STATUS_NEW);
lock = new LockHead(resId);
bucket->data.insert(LockHeadPair(resId, lock));
}
else {
// Lock is not free
lock = it->second;
}
// Sanity check if requests are being reused
invariant(request->lock == NULL || request->lock == lock);
request->lock = lock;
request->recursiveCount++;
if (request->status == LockRequest::STATUS_NEW) {
invariant(request->recursiveCount == 1);
// New lock request
if (conflicts(mode, lock->grantedModes)) {
request->status = LockRequest::STATUS_WAITING;
request->mode = mode;
request->convertMode = MODE_NONE;
// Put it on the conflict queue. This is the place where various policies could be
// applied for where in the wait queue does a request go.
lock->addToConflictQueue(request);
lock->changeConflictModeCount(mode, LockHead::Increment);
return LOCK_WAITING;
}
else { // No conflict, new request
request->status = LockRequest::STATUS_GRANTED;
request->mode = mode;
request->convertMode = MODE_NONE;
lock->addToGrantedQueue(request);
lock->changeGrantedModeCount(mode, LockHead::Increment);
return LOCK_OK;
}
}
else {
// If we are here, we already hold the lock in some mode. In order to keep it simple,
// we do not allow requesting a conversion while a lock is already waiting or pending
// conversion, hence the assertion below.
invariant(request->status == LockRequest::STATUS_GRANTED);
invariant(request->recursiveCount > 1);
invariant(request->mode != mode);
// Construct granted mask without our current mode, so that it is not counted as
// conflicting
uint32_t grantedModesWithoutCurrentRequest = 0;
// We start the counting at 1 below, because LockModesCount also includes MODE_NONE
// at position 0, which can never be acquired/granted.
for (uint32_t i = 1; i < LockModesCount; i++) {
const uint32_t currentRequestHolds =
(request->mode == static_cast<LockMode>(i) ? 1 : 0);
if (lock->grantedCounts[i] > currentRequestHolds) {
grantedModesWithoutCurrentRequest |= modeMask(static_cast<LockMode>(i));
}
}
// This check favours conversion requests over pending requests. For example:
//
// T1 requests lock L in IS
// T2 requests lock L in X
// T1 then upgrades L from IS -> S
//
// Because the check does not look into the conflict modes bitmap, it will grant L to
// T1 in S mode, instead of block, which would otherwise cause deadlock.
if (conflicts(mode, grantedModesWithoutCurrentRequest)) {
request->status = LockRequest::STATUS_CONVERTING;
request->convertMode = mode;
lock->conversionsCount++;
lock->changeGrantedModeCount(request->convertMode, LockHead::Increment);
return LOCK_WAITING;
}
else { // No conflict, existing request
lock->changeGrantedModeCount(mode, LockHead::Increment);
lock->changeGrantedModeCount(request->mode, LockHead::Decrement);
request->mode = mode;
return LOCK_OK;
}
}
}
void BtreeBuilder<V>::buildNextLevel(DiskLoc loc, bool mayInterrupt) {
int levels = 1;
while( 1 ) {
if( _getBucket(loc)->tempNext().isNull() ) {
// only 1 bucket at this level. we are done.
_btreeState->setHead( loc );
break;
}
levels++;
DiskLoc upLoc = BtreeBucket<V>::addBucket(_btreeState);
DiskLoc upStart = upLoc;
BtreeBucket<V> *up = _getModifiableBucket( upLoc );
DiskLoc xloc = loc;
while( !xloc.isNull() ) {
killCurrentOp.checkForInterrupt( !mayInterrupt );
if ( getDur().commitIfNeeded() ) {
b = _getModifiableBucket( cur );
up = _getModifiableBucket( upLoc );
}
BtreeBucket<V> *x = _getModifiableBucket( xloc );
Key k;
DiskLoc r;
x->popBack(r,k);
bool keepX = ( x->n != 0 );
DiskLoc keepLoc = keepX ? xloc : x->nextChild;
if ( ! up->_pushBack(r, k, _btreeState->ordering(), keepLoc) ) {
// current bucket full
DiskLoc n = BtreeBucket<V>::addBucket(_btreeState);
up->setTempNext(n);
upLoc = n;
up = _getModifiableBucket( upLoc );
up->pushBack(r, k, _btreeState->ordering(), keepLoc);
}
DiskLoc nextLoc = x->tempNext(); // get next in chain at current level
if ( keepX ) {
x->parent = upLoc;
}
else {
if ( !x->nextChild.isNull() ) {
DiskLoc ll = x->nextChild;
_getModifiableBucket(ll)->parent = upLoc;
//(x->nextChild.btreemod<V>())->parent = upLoc;
}
x->deallocBucket( _btreeState, xloc );
}
xloc = nextLoc;
}
loc = upStart;
mayCommitProgressDurably();
}
if( levels > 1 ) {
LOG(2) << "btree levels: " << levels << endl;
}
}