private List<long> DoPersist(TreeWrite sequence, int tryCount)
{
// NOTE: nodes are written in order of branches and then leaf nodes. All
// branch nodes and leafs are grouped together.
// The list of nodes to be allocated,
IList<ITreeNode> allBranches = sequence.BranchNodes;
IList<ITreeNode> allLeafs = sequence.LeafNodes;
List<ITreeNode> nodes = new List<ITreeNode>(allBranches.Count + allLeafs.Count);
nodes.AddRange(allBranches);
nodes.AddRange(allLeafs);
int sz = nodes.Count;
// The list of allocated referenced for the nodes,
DataAddress[] refs = new DataAddress[sz];
long[] outRefs = new long[sz];
MessageStream allocateMessage = new MessageStream(MessageType.Request);
// Make a connection with the manager server,
IMessageProcessor manager = connector.Connect(managerAddress, ServiceType.Manager);
// Allocate the space first,
for (int i = 0; i < sz; ++i) {
ITreeNode node = nodes[i];
RequestMessage request = new RequestMessage("allocateNode");
// Is it a branch node?
if (node is TreeBranch) {
// Branch nodes are 1K in size,
request.Arguments.Add(1024);
} else {
// Leaf nodes are 4k in size,
request.Arguments.Add(4096);
}
allocateMessage.AddMessage(request);
}
// The result of the set of allocations,
MessageStream resultStream = (MessageStream) manager.Process(allocateMessage);
//DEBUG: ++network_comm_count;
// The unique list of blocks,
List<long> uniqueBlocks = new List<long>();
// Parse the result stream one message at a time, the order will be the
// order of the allocation messages,
int n = 0;
foreach (ResponseMessage m in resultStream) {
if (m.HasError)
throw m.Error.AsException();
DataAddress addr = (DataAddress) m.Arguments[0].Value;
refs[n] = addr;
// Make a list of unique block identifiers,
if (!uniqueBlocks.Contains(addr.BlockId)) {
uniqueBlocks.Add(addr.BlockId);
}
++n;
}
// Get the block to server map for each of the blocks,
IDictionary<long, IList<BlockServerElement>> blockToServerMap = GetServersForBlock(uniqueBlocks);
// Make message streams for each unique block
int ubid_count = uniqueBlocks.Count;
MessageStream[] ubidStream = new MessageStream[ubid_count];
for (int i = 0; i < ubidStream.Length; ++i) {
ubidStream[i] = new MessageStream(MessageType.Request);
}
// Scan all the blocks and create the message streams,
for (int i = 0; i < sz; ++i) {
byte[] nodeBuf;
ITreeNode node = nodes[i];
// Is it a branch node?
if (node is TreeBranch) {
TreeBranch branch = (TreeBranch)node;
// Make a copy of the branch (NOTE; we Clone() the array here).
long[] curNodeData = (long[])branch.ChildPointers.Clone();
int curNdsz = branch.DataSize;
branch = new TreeBranch(refs[i].Value, curNodeData, curNdsz);
// The number of children
int chsz = branch.ChildCount;
// For each child, if it's a heap node, look up the child id and
// reference map in the sequence and set the reference accordingly,
for (int o = 0; o < chsz; ++o) {
long childRef = branch.GetChild(o);
if (childRef < 0) {
// The ref is currently on the heap, so adjust accordingly
int ref_id = sequence.LookupRef(i, o);
branch.SetChildOverride(o, refs[ref_id].Value);
}
}
// Turn the branch into a 'node_buf' byte[] array object for
// serialization.
long[] nodeData = branch.ChildPointers;
int ndsz = branch.DataSize;
MemoryStream bout = new MemoryStream(1024);
BinaryWriter dout = new BinaryWriter(bout, Encoding.Unicode);
dout.Write(BranchType);
dout.Write(ndsz);
for (int o = 0; o < ndsz; ++o) {
dout.Write(nodeData[o]);
}
dout.Flush();
// Turn it into a byte array,
nodeBuf = bout.ToArray();
// Put this branch into the local cache,
networkCache.SetNode(refs[i], branch);
} else {
// If it's a leaf node,
TreeLeaf leaf = (TreeLeaf)node;
int lfsz = leaf.Length;
nodeBuf = new byte[lfsz + 6];
// Technically, we could comment these next two lines out.
ByteBuffer.WriteInt2(LeafType, nodeBuf, 0);
ByteBuffer.WriteInt4(lfsz, nodeBuf, 2);
leaf.Read(0, nodeBuf, 6, lfsz);
// Put this leaf into the local cache,
leaf = new ByteArrayTreeLeaf(refs[i].Value, nodeBuf);
networkCache.SetNode(refs[i], leaf);
}
// The DataAddress this node is being written to,
DataAddress address = refs[i];
// Get the block id,
long blockId = address.BlockId;
int bid = uniqueBlocks.IndexOf(blockId);
RequestMessage request = new RequestMessage("writeToBlock");
request.Arguments.Add(address);
request.Arguments.Add(nodeBuf);
request.Arguments.Add(0);
request.Arguments.Add(nodeBuf.Length);
ubidStream[bid].AddMessage(request);
// Update 'outRefs' array,
outRefs[i] = refs[i].Value;
}
// A log of successfully processed operations,
List<object> successProcess = new List<object>(64);
// Now process the streams on the servers,
for (int i = 0; i < ubidStream.Length; ++i) {
// The output message,
MessageStream requestMessageStream = ubidStream[i];
// Get the servers this message needs to be sent to,
long block_id = uniqueBlocks[i];
IList<BlockServerElement> blockServers = blockToServerMap[block_id];
// Format a message for writing this node out,
int bssz = blockServers.Count;
IMessageProcessor[] blockServerProcs = new IMessageProcessor[bssz];
// Make the block server connections,
for (int o = 0; o < bssz; ++o) {
IServiceAddress address = blockServers[o].Address;
blockServerProcs[o] = connector.Connect(address, ServiceType.Block);
MessageStream responseMessageStream = (MessageStream) blockServerProcs[o].Process(requestMessageStream);
//DEBUG: ++network_comm_count;
if (responseMessageStream.HasError) {
// If this is an error, we need to report the failure to the
// manager server,
ReportBlockServerFailure(address);
// Remove the block id from the server list cache,
networkCache.RemoveServers(block_id);
// Rollback any server writes already successfully made,
for (int p = 0; p < successProcess.Count; p += 2) {
IServiceAddress blockAddress = (IServiceAddress) successProcess[p];
MessageStream toRollback = (MessageStream) successProcess[p + 1];
List<DataAddress> rollbackNodes = new List<DataAddress>(128);
foreach(Message rm in toRollback) {
DataAddress raddr = (DataAddress) rm.Arguments[0].Value;
rollbackNodes.Add(raddr);
}
// Create the rollback message,
RequestMessage rollbackRequest = new RequestMessage("rollbackNodes");
rollbackRequest.Arguments.Add(rollbackNodes.ToArray());
// Send it to the block server,
Message responseMessage = connector.Connect(blockAddress, ServiceType.Block).Process(rollbackRequest);
//DEBUG: ++network_comm_count;
// If rollback generated an error we throw the error now
// because this likely is a serious network error.
if (responseMessage.HasError)
throw new NetworkException("Rollback wrote failed: " + responseMessage.ErrorMessage);
}
// Retry,
if (tryCount > 0)
return DoPersist(sequence, tryCount - 1);
// Otherwise we fail the write
throw new NetworkException(responseMessageStream.ErrorMessage);
}
// If we succeeded without an error, add to the log
successProcess.Add(address);
successProcess.Add(requestMessageStream);
}
}
// Return the references,
return new List<long>(outRefs);
}
public IList<long> Persist(TreeWrite write)
{
return DoPersist(write, 1);
}
public DataAddress CreateEmptyDatabase()
{
TreeNodeHeap nodeHeap = new TreeNodeHeap(17, 4 * 1024 * 1024);
TreeLeaf headLeaf = nodeHeap.CreateLeaf(null, Key.Head, 256);
// Insert a tree identification pattern
headLeaf.Write(0, new byte[] { 1, 1, 1, 1 }, 0, 4);
// Create an empty tail node
TreeLeaf tailLeaf = nodeHeap.CreateLeaf(null, Key.Tail, 256);
// Insert a tree identification pattern
tailLeaf.Write(0, new byte[] { 1, 1, 1, 1 }, 0, 4);
// The write sequence,
TreeWrite seq = new TreeWrite();
seq.NodeWrite(headLeaf);
seq.NodeWrite(tailLeaf);
IList<long> refs = Persist(seq);
// Create a branch,
int maxBranchSize = MaxBranchSize;
TreeBranch rootBranch = nodeHeap.CreateBranch(null, maxBranchSize);
rootBranch.Set(refs[0], 4,
Key.Tail.GetEncoded((1)),
Key.Tail.GetEncoded((2)),
refs[1], 4);
seq = new TreeWrite();
seq.NodeWrite(rootBranch);
refs = Persist(seq);
// The written root node reference,
long root_id = refs[0];
// Delete the head and tail leaf, and the root branch
nodeHeap.Delete(headLeaf.Id);
nodeHeap.Delete(tailLeaf.Id);
nodeHeap.Delete(rootBranch.Id);
// Return the root,
return new DataAddress(root_id);
}
private NodeId[] InternalPersist(TreeWrite sequence, int tryCount)
{
// NOTE: nodes are written in order of branches and then leaf nodes. All
// branch nodes and leafs are grouped together.
// The list of nodes to be allocated,
IList<ITreeNode> allBranches = sequence.BranchNodes;
IList<ITreeNode> allLeafs = sequence.LeafNodes;
List<ITreeNode> nodes = new List<ITreeNode>(allBranches.Count + allLeafs.Count);
nodes.AddRange(allBranches);
nodes.AddRange(allLeafs);
int sz = nodes.Count;
// The list of allocated referenced for the nodes,
DataAddress[] refs = new DataAddress[sz];
NodeId[] outNodeIds = new NodeId[sz];
MessageStream allocateMessageStream = new MessageStream();
// Allocate the space first,
for (int i = 0; i < sz; ++i) {
ITreeNode node = nodes[i];
// Is it a branch node?
if (node is TreeBranch) {
// Branch nodes are 1K in size,
allocateMessageStream.AddMessage(new Message("allocateNode", 1024));
}
// Otherwise, it must be a leaf node,
else {
// Leaf nodes are 4k in size,
allocateMessageStream.AddMessage(new Message("allocateNode", 4096));
}
}
// Process a command on the manager,
IEnumerable<Message> resultStream = ProcessManager(allocateMessageStream);
// The unique list of blocks,
List<BlockId> uniqueBlocks = new List<BlockId>();
// Parse the result stream one message at a time, the order will be the
// order of the allocation messages,
int n = 0;
foreach (Message m in resultStream) {
if (m.HasError)
throw new ApplicationException(m.ErrorMessage);
DataAddress addr = (DataAddress) m.Arguments[0].Value;
refs[n] = addr;
// Make a list of unique block identifiers,
if (!uniqueBlocks.Contains(addr.BlockId)) {
uniqueBlocks.Add(addr.BlockId);
}
++n;
}
// Get the block to server map for each of the blocks,
IDictionary<BlockId, IList<BlockServerElement>> blockToServerMap =
GetServerListForBlocks(uniqueBlocks);
// Make message streams for each unique block
int ubidCount = uniqueBlocks.Count;
MessageStream[] ubidStream = new MessageStream[ubidCount];
for (int i = 0; i < ubidStream.Length; ++i) {
ubidStream[i] = new MessageStream();
}
// Scan all the blocks and create the message streams,
for (int i = 0; i < sz; ++i) {
byte[] nodeBuf;
ITreeNode node = nodes[i];
// Is it a branch node?
if (node is TreeBranch) {
TreeBranch branch = (TreeBranch) node;
// Make a copy of the branch (NOTE; we clone() the array here).
long[] curNodeData = (long[]) branch.NodeData.Clone();
int curNdsz = branch.NodeDataSize;
branch = new TreeBranch(refs[i].Value, curNodeData, curNdsz);
// The number of children
int chsz = branch.ChildCount;
// For each child, if it's a heap node, look up the child id and
// reference map in the sequence and set the reference accordingly,
for (int o = 0; o < chsz; ++o) {
NodeId childId = branch.GetChild(o);
if (childId.IsInMemory) {
// The ref is currently on the heap, so adjust accordingly
int refId = sequence.LookupRef(i, o);
branch.SetChildOverride(refs[refId].Value, o);
}
}
// Turn the branch into a 'node_buf' byte[] array object for
// serialization.
long[] nodeData = branch.NodeData;
int ndsz = branch.NodeDataSize;
MemoryStream bout = new MemoryStream(1024);
BinaryWriter dout = new BinaryWriter(bout);
dout.Write(StoreBranchType);
dout.Write((short) 0); // Reserved for future
dout.Write(0); // The crc32 checksum will be written here,
dout.Write(ndsz);
for (int o = 0; o < ndsz; ++o) {
dout.Write(nodeData[o]);
}
dout.Flush();
// Turn it into a byte array,
nodeBuf = bout.ToArray();
// Write the crc32 of the data,
Crc32 checksum = new Crc32();
checksum.ComputeHash(nodeBuf, 8, nodeBuf.Length - 8);
ByteBuffer.WriteInt4((int) checksum.CrcValue, nodeBuf, 4);
// Put this branch into the local cache,
networkCache.SetNode(refs[i], branch);
}
// If it's a leaf node,
else {
TreeLeaf leaf = (TreeLeaf) node;
int lfsz = leaf.Length;
nodeBuf = new byte[lfsz + 12];
// Format the data,
ByteBuffer.WriteInt2(StoreLeafType, nodeBuf, 0);
ByteBuffer.WriteInt2(0, nodeBuf, 2); // Reserved for future
ByteBuffer.WriteInt4(lfsz, nodeBuf, 8);
leaf.Read(0, nodeBuf, 12, lfsz);
// Calculate and set the checksum,
Crc32 checksum = new Crc32();
checksum.ComputeHash(nodeBuf, 8, nodeBuf.Length - 8);
ByteBuffer.WriteInt4((int) checksum.CrcValue, nodeBuf, 4);
// Put this leaf into the local cache,
leaf = new MemoryTreeLeaf(refs[i].Value, nodeBuf);
networkCache.SetNode(refs[i], leaf);
}
// The DataAddress this node is being written to,
DataAddress address = refs[i];
// Get the block id,
BlockId blockId = address.BlockId;
int bid = uniqueBlocks.IndexOf(blockId);
ubidStream[bid].AddMessage(new Message("writeToBlock", address, nodeBuf, 0, nodeBuf.Length));
// Update 'out_refs' array,
outNodeIds[i] = refs[i].Value;
}
// A log of successfully processed operations,
List<object> successProcess = new List<object>(64);
// Now process the streams on the servers,
for (int i = 0; i < ubidStream.Length; ++i) {
// The output message,
MessageStream outputStream = ubidStream[i];
// Get the servers this message needs to be sent to,
BlockId blockId = uniqueBlocks[i];
IList<BlockServerElement> blockServers = blockToServerMap[blockId];
// Format a message for writing this node out,
int bssz = blockServers.Count;
IMessageProcessor[] blockServerProcs = new IMessageProcessor[bssz];
// Make the block server connections,
for (int o = 0; o < bssz; ++o) {
IServiceAddress address = blockServers[o].Address;
blockServerProcs[o] = connector.Connect(address, ServiceType.Block);
IEnumerable<Message> inputStream = blockServerProcs[o].Process(outputStream);
++NetworkCommCount;
foreach (Message m in inputStream) {
if (m.HasError) {
// If this is an error, we need to report the failure to the
// manager server,
ReportBlockServerFailure(address);
// Remove the block id from the server list cache,
networkCache.RemoveServersWithBlock(blockId);
// Rollback any server writes already successfully made,
for (int p = 0; p < successProcess.Count; p += 2) {
IServiceAddress blocksAddr = (IServiceAddress) successProcess[p];
MessageStream toRollback = (MessageStream) successProcess[p + 1];
List<DataAddress> rollbackNodes = new List<DataAddress>(128);
foreach (Message rm in toRollback) {
DataAddress raddr = (DataAddress) rm.Arguments[0].Value;
rollbackNodes.Add(raddr);
}
// Create the rollback message,
MessageStream rollbackMsg = new MessageStream();
rollbackMsg.AddMessage(new Message("rollbackNodes", new object[] {rollbackNodes.ToArray()}));
// Send it to the block server,
IEnumerable<Message> responseStream = connector.Connect(blocksAddr, ServiceType.Block).Process(rollbackMsg);
++NetworkCommCount;
foreach (Message rbm in responseStream) {
// If rollback generated an error we throw the error now
// because this likely is a serious network error.
if (rbm.HasError) {
throw new NetworkWriteException("Write failed (rollback failed): " + rbm.ErrorMessage);
}
}
}
// Retry,
if (tryCount > 0)
return InternalPersist(sequence, tryCount - 1);
// Otherwise we fail the write
throw new NetworkWriteException(m.ErrorMessage);
}
}
// If we succeeded without an error, add to the log
successProcess.Add(address);
successProcess.Add(outputStream);
}
}
// Return the references,
return outNodeIds;
}
public IList<NodeId> Persist(TreeWrite write)
{
return InternalPersist(write, 3);
}
public DataAddress CreateDatabase()
{
// The child reference is a sparse node element
NodeId childId = NodeId.CreateSpecialSparseNode((byte) 1, 4);
// Create a branch,
TreeBranch rootBranch = new TreeBranch(NodeId.CreateInMemoryNode(0L), MaxBranchSize);
rootBranch.Set(childId, 4, Key.Tail, childId, 4);
TreeWrite seq = new TreeWrite();
seq.NodeWrite(rootBranch);
IList<NodeId> refs = Persist(seq);
// The written root node reference,
NodeId rootId = refs[0];
// Return the root,
return new DataAddress(rootId);
}