void MetadataManager::_removeRangeToClean_inlock(const ChunkRange& range) {
auto it = _rangesToClean.upper_bound(range.getMin());
// We want our iterator to point at the greatest value
// that is still less than or equal to range.
if (it != _rangesToClean.begin()) {
--it;
}
for (; it != _rangesToClean.end() && it->first < range.getMax();) {
if (it->second <= range.getMin()) {
++it;
continue;
}
// There's overlap between *it and range so we remove *it
// and then replace with new ranges.
BSONObj oldMin = it->first, oldMax = it->second;
_rangesToClean.erase(it++);
if (oldMin < range.getMin()) {
_addRangeToClean_inlock(ChunkRange(oldMin, range.getMin()));
}
if (oldMax > range.getMax()) {
_addRangeToClean_inlock(ChunkRange(range.getMax(), oldMax));
}
}
}
void MetadataManager::beginReceive(const ChunkRange& range) {
stdx::lock_guard<stdx::mutex> scopedLock(_managerLock);
// Collection is not known to be sharded if the active metadata tracker is null
invariant(_activeMetadataTracker);
// If range is contained within pending chunks, this means a previous migration must have failed
// and we need to clean all overlaps
RangeVector overlappedChunks;
getRangeMapOverlap(_receivingChunks, range.getMin(), range.getMax(), &overlappedChunks);
for (const auto& overlapChunkMin : overlappedChunks) {
auto itRecv = _receivingChunks.find(overlapChunkMin.first);
invariant(itRecv != _receivingChunks.end());
const ChunkRange receivingRange(itRecv->first, itRecv->second);
_receivingChunks.erase(itRecv);
// Make sure any potentially partially copied chunks are scheduled to be cleaned up
_addRangeToClean_inlock(receivingRange);
}
// Need to ensure that the background range deleter task won't delete the range we are about to
// receive
_removeRangeToClean_inlock(range);
_receivingChunks.insert(std::make_pair(range.getMin().getOwned(), range.getMax().getOwned()));
// For compatibility with the current range deleter, update the pending chunks on the collection
// metadata to include the chunk being received
ChunkType chunk;
chunk.setMin(range.getMin());
chunk.setMax(range.getMax());
_setActiveMetadata_inlock(_activeMetadataTracker->metadata->clonePlusPending(chunk));
}
void MetadataManager::forgetReceive(const ChunkRange& range) {
stdx::lock_guard<stdx::mutex> scopedLock(_managerLock);
{
auto it = _receivingChunks.find(range.getMin());
invariant(it != _receivingChunks.end());
// Verify entire ChunkRange is identical, not just the min key.
invariant(it->second == range.getMax());
_receivingChunks.erase(it);
}
// This is potentially a partially received data, which needs to be cleaned up
_addRangeToClean_inlock(range);
// For compatibility with the current range deleter, update the pending chunks on the collection
// metadata to exclude the chunk being received, which was added in beginReceive
ChunkType chunk;
chunk.setMin(range.getMin());
chunk.setMax(range.getMax());
_setActiveMetadata_inlock(_activeMetadataTracker->metadata->cloneMinusPending(chunk));
}
void MetadataManager::refreshActiveMetadata(std::unique_ptr<CollectionMetadata> remoteMetadata) {
LOG(1) << "Refreshing the active metadata from "
<< (_activeMetadataTracker->metadata ? _activeMetadataTracker->metadata->toStringBasic()
: "(empty)")
<< ", to " << (remoteMetadata ? remoteMetadata->toStringBasic() : "(empty)");
stdx::lock_guard<stdx::mutex> scopedLock(_managerLock);
// Collection is not sharded anymore
if (!remoteMetadata) {
log() << "Marking collection as not sharded.";
_receivingChunks.clear();
_rangesToClean.clear();
_setActiveMetadata_inlock(nullptr);
return;
}
invariant(!remoteMetadata->getCollVersion().isWriteCompatibleWith(ChunkVersion::UNSHARDED()));
invariant(!remoteMetadata->getShardVersion().isWriteCompatibleWith(ChunkVersion::UNSHARDED()));
// Collection is not sharded currently
if (!_activeMetadataTracker->metadata) {
log() << "Marking collection as sharded with version " << remoteMetadata->toStringBasic();
invariant(_receivingChunks.empty());
invariant(_rangesToClean.empty());
_setActiveMetadata_inlock(std::move(remoteMetadata));
return;
}
// If the metadata being installed has a different epoch from ours, this means the collection
// was dropped and recreated, so we must entirely reset the metadata state
if (_activeMetadataTracker->metadata->getCollVersion().epoch() !=
remoteMetadata->getCollVersion().epoch()) {
log() << "Overwriting collection metadata due to epoch change.";
_receivingChunks.clear();
_rangesToClean.clear();
_setActiveMetadata_inlock(std::move(remoteMetadata));
return;
}
// We already have newer version
if (_activeMetadataTracker->metadata->getCollVersion() >= remoteMetadata->getCollVersion()) {
LOG(1) << "Attempted to refresh active metadata "
<< _activeMetadataTracker->metadata->toStringBasic() << " with an older version "
<< remoteMetadata->toStringBasic();
return;
}
// Resolve any receiving chunks, which might have completed by now
for (auto it = _receivingChunks.begin(); it != _receivingChunks.end();) {
const BSONObj min = it->first;
const BSONObj max = it->second;
// Our pending range overlaps at least one chunk
if (rangeMapContains(remoteMetadata->getChunks(), min, max)) {
// The remote metadata contains a chunk we were earlier in the process of receiving, so
// we deem it successfully received.
LOG(2) << "Verified chunk " << ChunkRange(min, max).toString()
<< " was migrated earlier to this shard";
_receivingChunks.erase(it++);
continue;
} else if (!rangeMapOverlaps(remoteMetadata->getChunks(), min, max)) {
++it;
continue;
}
// Partial overlap indicates that the earlier migration has failed, but the chunk being
// migrated underwent some splits and other migrations and ended up here again. In this
// case, we will request full reload of the metadata. Currently this cannot happen, because
// all migrations are with the explicit knowledge of the recipient shard. However, we leave
// the option open so that chunk splits can do empty chunk move without having to notify the
// recipient.
RangeVector overlappedChunks;
getRangeMapOverlap(remoteMetadata->getChunks(), min, max, &overlappedChunks);
for (const auto& overlapChunkMin : overlappedChunks) {
auto itRecv = _receivingChunks.find(overlapChunkMin.first);
invariant(itRecv != _receivingChunks.end());
const ChunkRange receivingRange(itRecv->first, itRecv->second);
_receivingChunks.erase(itRecv);
// Make sure any potentially partially copied chunks are scheduled to be cleaned up
_addRangeToClean_inlock(receivingRange);
}
// Need to reset the iterator
it = _receivingChunks.begin();
}
// For compatibility with the current range deleter, which is driven entirely by the contents of
// the CollectionMetadata update the pending chunks
for (const auto& receivingChunk : _receivingChunks) {
ChunkType chunk;
chunk.setMin(receivingChunk.first);
chunk.setMax(receivingChunk.second);
remoteMetadata = remoteMetadata->clonePlusPending(chunk);
}
_setActiveMetadata_inlock(std::move(remoteMetadata));
}
void MetadataManager::addRangeToClean(const ChunkRange& range) {
stdx::lock_guard<stdx::mutex> scopedLock(_managerLock);
_addRangeToClean_inlock(range);
}
std::shared_ptr<Notification<Status>> MetadataManager::addRangeToClean(const ChunkRange& range) {
stdx::lock_guard<stdx::mutex> scopedLock(_managerLock);
return _addRangeToClean_inlock(range);
}