DocsWhenever the server is writing to a given connection, it is permissible for the server to stop reading from that connection (but if the writing would block, it must continue to service other connections). The server is not required to buffer more than a single request per connection at one time. For a given connection to the server, a client can block while reading from the connection but should undertake to read (events and errors) when writing would block. Failure on the part of a client to obey this rule could result in a deadlocked connection, although deadlock is probably unlikely unless either the transport layer has very little buffering or the client attempts to send large numbers of requests without ever reading replies or checking for errors and events.
Whether or not a server is implemented with internal concurrency, the overall effect must be as if individual requests are executed to completion in some serial order, and requests from a given connection must be executed in delivery order (that is, the total execution order is a shuffle of the individual streams). The execution of a request includes validating all arguments, collecting all data for any reply, and generating and queueing all required events. However, it does not include the actual transmission of the reply and the events. In addition, the effect of any other cause that can generate multiple events (for example, activation of a grab or pointer motion) must effectively generate and queue all required events indivisibly with respect to all other causes and requests. For a request from a given client, any events destined for that client that are caused by executing the request must be sent to the client before any reply or error is sent.