The following source was built using Visual Studio 6.0 SP5 and Visual Studio .NET. You need to have a version of the Microsoft Platform SDK installed.
Note that the debug builds of the code waste a lot of CPU cycles due to the debug trace output. It's only worth profiling the release builds.
"How do you handle the problem of multiple pending
WSARecv() calls?" is a common question on the Winsock news groups. It seems that everyone knows that it's often a good idea to have more than one outstanding read waiting on a socket and everyone's equally aware that sometimes code doesn't work right when you do that. This article explains the potential problems with multiple pending recvs and provides a solution within the reusable server framework that we've been developing over the last few articles.
That's Out of Order
There is a subtle issue when using IO completion ports with multiple threads. Although operations using the IO completion port will always complete in the order that they were submitted, thread scheduling issues may mean that the actual work associated with the completion is processed in an undefined order. For example, if we were to submit three
WSARecv requests on a socket, then they are guaranteed to complete in the order that we submitted them, however if we have 2 threads servicing the IO completion port two of the completions could be being processed simultaneously. If the thread processing the 'first'
WSARecv completion is interrupted, the second may be completely processed before the first. This is even more likely to occur on machines with multiple processors where the two threads may really be executing simultaneously, but it's possible on single processor boxes too. As always, this is the kind of subtle problem that probably won't show its face until you release the software to production...
The example above is easy to avoid, simply don't have multiple
WSARecv requests outstanding on a single socket. This is what we have done so far in the example servers developed in the previous articles. However, this reduces performance, it's always more performant to have a receive pending when the data actually arrives on the wire than it is to post a receive after the data has already arrived. Having multiple
WSARecv calls outstanding ensures that there's always a call pending. What's more, the problem isn't limited to having multiple
WSARecvs. In our server framework, we marshall all socket IO calls from the user's threads into our IO thread pool using the IO completion port. This means that there is the potential for a user thread to issue multiple consecutive writes to the socket and for them to be executed in an undefined order.
In our example servers so far, code like this:
pSocket->Write(pBuffer); pSocket->Write(pResponse1); pSocket->Write(pResponse2); }
is potentially unsafe because the writes don't occur synchronously on the user's thread, they are posted to the IO completion port and occur in the IO thread pool.
Preserving the order of IO completion operations is relatively straight forward. As you'll remember, the overlapped structure passed to all calls that use IO completion ports represents 'per call' data. We can, and do, extend the overlapped structure to include our own 'per call' data by using the
CIOBuffer class. If we add a sequence number to the
CIOBuffer, we can set the sequence number to the '
next' value in the user's thread and then make sure we process the buffers in order in the IO thread pool. This concept applies to any IO completion port operation and each distinct operation requires its own sequence number. For our server framework, that means that our
Socket class must now maintain independent sequence numbers for read and write requests.
The sequence number management code inside the
Write method could be something like this:
To ensure that the sequence numbers actually represent the order that the operations are submitted requires that the setting of the sequence number and the submission of the operation are an atomic operation. For our socket writes, this isn't a problem as we only guarantee the order of writes that are performed on a single thread, for socket reads, we need to ensure that the allocation of the sequence number and call to
WSARecv() occur without another thread having a chance to perform read at the same time. This involves using a critical section to lock access to the socket during the sequence number allocation and
WSARecv() call. Failure to lock in this area can lead to the actual order that the
WSARecv() calls are made failing to match the ordering of the sequence numbers allocated.
The code to ensure that the IO completions are handled in order is a little more complex. For each distinct IO operation, we need to keep track of the next sequence number that we can process. When a call to
GetQueuedCompletionStatus() returns, we need to compare the sequence number in the request with the next sequence number that we can process. If these numbers match, then we can process the request. If they don't, then the request cannot be processed at this time. If an IO operation cannot be processed, it should be stored for later processing. The storage of the out of sequence request needs to be keyed on the sequence number. When an IO thread finds that it can't process the current request, it should add the current request to the store and see if there's a request in the store that can be processed. When a request is processed, the last thing that the IO thread should do is atomically increment the value representing the next sequence number to process and check to see if there's an IO request in the store that can be processed.
The above strategy handles the situation where multiple IO requests complete concurrently. Only one thread can be processing an IO request that meets the criteria of being the next one to process, all other threads will simply add their requests to the store. When the thread that's processing a request finishes processing, it can check to see if there are other requests in the store that can now be processed. If a thread needs to store its IO request, then it can do so and then check for a request that can be processed in an atomic operation.
It's actually more complex to read about than it is to look at, the code to process an operation in order might look like this:
pBuffer = pSocket->m_outOfSequenceWrites.GetNext(pBuffer);
DWORD dwFlags = 0;
DWORD dwSendNumBytes = 0;
if (SOCKET_ERROR == ::WSASend(
pBuffer = pSocket->m_outOfSequenceWrites.ProcessAndGetNext();
The store itself needs to map sequence numbers to
CIOBuffers. The obvious choice of data structure is a
std::map<> though your performance requirements and profiling may dictate a different choice.
GetNext() takes a buffer, compares its sequence number with the next one we can process and either returns the buffer or adds the buffer to the map and then checks the map to see if the first buffer in the map is the one we can process. Remember that the map stores its elements in order of their keys and that we're using the sequence number as the key, so
m_list.begin() refers to the element in the map with the lowest sequence number. If this function returns
null, then we're still waiting for the 'next' buffer to arrive.
if (m_next == pBuffer->GetSequenceNumber())
BS::value_type value(pBuffer->GetSequenceNumber(), pBuffer));
std::pair<BS::iterator, bool> result = m_list.insert(value);
if (result.second == false)
CIOBuffer *pNext = 0;
it = m_list.begin();
if (it != m_list.end())
if (it->first == m_next)
pNext = it->second;
After processing a buffer, the thread can check to see if there's another buffer that it can handle. It needs to increase the last processed value and perform the check atomically, hence the locking.
CIOBuffer *pNext = 0;
it = m_list.begin();
if (it != m_list.end())
if (it->first == m_next)
pNext = it->second;
CIOBuffer used by every write that occurs contains a sequence number, then similar code could be used to ensure that completed read requests are processed in the correct order. However, there's little point in this code being placed in the server framework as different users of the framework may require different functionality. The
CSocketServer derived class could use the
CIOBuffer::InOrderBufferList class to maintain processing order or it could simply dispatch the read completions to another IO completion port to pass them across to a business logic thread pool. In this case, it's the code in the business logic thread pool that actually processes the data and the order should be maintained there. It may even need do both, ensuring packet order in the
CSocketServer class itself so that it can successfully break the byte stream into messages and then dispatching the messages to the business logic thread pool and ensuring that these complete messages are also processed in the correct order.
Socket must now keep track of independent read and write sequence numbers and maintain a map of out of sequence write requests. Manipulation of the map and associated next sequence number counter must be protected. We use a critical section to protect this code. Be aware that allocating a critical section for each
Socket connection is potentially resource intensive. Instead, we could choose to trade locking granularity for performance. The
CSocketServer class already has a critical section that it uses to protect its lists of
Sockets, we could pass a reference to this critical section to each
Socket rather than have them create their own critical section. The problem with doing this is that we serialise every
Socket's map access. This work performed inside the critical section is small, but a better solution might be to create a critical section for every X sockets where X is a value that is determined by profiling your application.
Only Paying For What You Use
Including sequence numbers in all buffers used for sending and receiving and ensuring the writes are processed in order adds a little overhead to the work of the IO threads. If you are sure that your server doesn't require this functionality, perhaps because you know that due to your protocol design there will only ever be a single read or write request pending, you can opt not to include this functionality by passing
false as the
useSequenceNumbers flag in the
CSocketServer's constructor. Enabling read or write sequence numbers independently is left as an exercise for the reader.
To demonstrate the concept of ensuring the ordering of multiple reads and writes, we've come up with a rather contrived example. The packet echo server that we developed in the previous article has been changed as follows:
- It now does its work in a business logic thread pool so that we can demonstrate maintaining receive order in both the socket server and the business logic thread pool when the socket server's worker thread isn't doing the processing itself.
- It works with larger packets; we use a two byte packet header rather than a one byte header. This two byte header represents the length of the packet using the following format:
packetLength = byte1 + (byte2 * 256). The length of the packet includes the two byte header.
- When the client initially connects, it posts a configurable number of reads. As each read completes, it posts a new read so that it maintains the number of outstanding reads.
- It processes the reads in order, and due to the fact that we now have multiple reads outstanding,
CSocketServer::ProcessDataStream() has changed so that when more data is required, we don't simply reissue a read to read more data into the same buffer.
- It echoes the packet back to the client in pieces by issuing multiple write requests.
The large packet echo server is available for download here in SocketServer6.zip. Testing using telnet is possible, though more complex, you may find it easier to use the test harness that we develop here to test it. As with the previous examples, the server runs until a named event is set and then shuts down. The very simple Server Shutdown program, available here, provides an off switch for the server.
Although both the server and thread pool classes are configurable as to whether they use sequence numbers to maintain packet order, these settings can only be set in one way for the server to actually work in the way that the test harness expects. All packet ordering flags must be set to
true. The purpose of the flags is so that you can turn off the various sequencing required and see the effect on the test. It's not intended that the server can run reliably in any other configuration.
5, 5, 10, 5, 5000, 100, 10000, 20, true, true);
INADDR_ANY, 5001, 10, 10, 1024, pool,
65536, true, true, true);
The configuration flags can be adjusted to witness the following effects:
- The configuration shown above ensures that packets into the read completion method are processed in sequence - this maintains the validity of the incoming packet data; Packets into the worker thread are maintained in sequence - this maintains the order of echoing the actual packets; and the sequence of write calls is maintained - this maintains the validity of the outgoing packet data.
- If the number of reads to post (1) is reduced to 1, then there is no need to maintain the read completion sequencing ((4) can be set to
false) as long as read completion method doesn't issue another read until it has completed the processing of the current one ((6) should be set to
- If the business logic thread pool doesn't attempt to maintain packet ordering ((2) set to
false) then the test will likely report sequence number mismatches - as the packets are echoed out of sequence, and response != message errors as multiple threads in the business logic thread pool attempt to write fragments of the message to the socket in an unsynchronised manner and thus interleave sections of different messages.
- If (2) is left set to
false but (3) is also set to
false, then the test will only fail with sequence number mismatches as the threads are now echoing their packets in a single write so the data in the individual packets cant be corrupted by being interleaved with sections of other packets.
- Unfortunately, it's currently impossible to simply turn off write packet ordering at the socket server (option (5)) as doing so also turns off read packet sequencing and so makes it impossible to get valid data to the business logic thread so that it can echo it and we can witness the writes being performed out of sequence. If you're interested in seeing this in action, then you can hack at the socket server code.
- 15th July 2002 - Initial revision
- 12th August 2002 - Removed the race condition in socket closure - Thanks to David McConnell for pointing this out. Derived class can receive connection reset and connection error notifications. Socket provides a means to determine if send/receive are connected. Dispatch to the thread pool now uses shared
enums rather than hard coded constant values. General code cleaning and lint issues. Adjusted the code and article so that each socket has its own critical section and the resource utilisation optimisation is suggested, rather than imposed. Fixed a bug whereby the critical section that is used to protect the per socket data was owned by the worker thread rather than the per socket data.
Other Articles in the Series
- A reusable socket server class
- Business logic processing in a socket server
- Speeding up socket server connections with AcceptEx
- Handling multiple pending socket read and write operations
- Testing socket servers with C# and .NET
- A high performance TCP/IP socket server COM component for VB