[THRIFT-4283] TNamedPipeServer race condition in interrupt

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 0.10.0
• Fix Version/s: 0.11.0
• Component/s: C++ - Library
• Labels: c++

Environment

Windows

• Patch Info: Patch Available
• Assignee: James E. King III
• Reporter: Jean-Noël Goor
• Created: 10/Aug/17 08:49
• Updated: 14/Dec/17 13:55
• Resolved: 22/Sep/17 17:12

Description

In TNamedPipeServer, there exists a race condition between the 'serve' thread and another thread that would attempt to interrupt it.

While the 'serve' thread is waiting for clients to connect, it is blocked on a GetOverlappedResult call.

To stop the server, another thread calls the interrupt function that submits a CANCELIO operation to the TOverlappedSubmissionThread.

The ::CancelIo function is called, freeing the 'serve' thread with an unsuccessful OvelappedResult which then throws and terminate the server by deleting the serverTransport_.

All this can happen while the call to interrupt has not yet returned, and even the event signaling the execution of the CANCELIO request might not have been caught.

In our code, this result in unpredictable behavior (mostly detected as stalled).

It is very hard to reproduce in test environment as it consists in a running threads race condition.

Issue Links

• links to: GitHub Pull Request #1368