Concurrency problem in remote profiler communication mechanism

[asfernandes]

The communication mechanism used in remote profiling is susceptible to race condition if the client abandon the read of a message.

That may happen if a RDB$PROFILER query is cancelled.

In that case, it will unlock the shared memory mutex while the server is processing a command. As soon the server is read to write the response, it will write things to the buffer, but as the client already unlocked the mutex, another client may already had taken the mutex and have written a new command to the same buffer.

[asfernandes]

A solution that looks simple for me is to split buffer (and bufferSize) into clientBuffer and serverBuffer.

[asfernandes]

A solution that looks simple for me is to split buffer (and bufferSize) into clientBuffer and serverBuffer.

It's not safe too. Before even the server may have read the client data, another client may have take control of the client buffer.