RFR 8067105: Socket returned by ServerSocket.accept() is inherited by child process on Windows
Alan.Bateman at oracle.com
Thu Dec 18 13:20:50 UTC 2014
On 18/12/2014 11:52, Chris Hegarty wrote:
> Thanks for looking at this Alan.
> Surprisingly this issue has existed for a long time. I tested with the FCS version of JRE 5, 6, 7, 8, and 9-ea, on Windows Server 2008 R2, and 2012, and the newly accepted socket inherit behaviour can be seen. NIO server socket channels also suffers from the same problem.
> Since networking and NIO already sets HANDLE_FLAG_INHERIT on newly created sockets, I’d like to keep it consistent, and set it on accepted sockets too. I also think that this fix should be backpored to previous releases, so a small localised change seems appropriate.
> Complete and updated webrev:
> I will file a separate bug to investigate why CreateProcess is being called with bInheritHandles set to TRUE, as I think any changes here will have a wider impact.
I'm very surprised by this because the issue of inheriting handles into
child processes has come up several times in the past. The patch you
propose is okay but it doesn't deal with the case of a process being
started at around the time that a connection is accepted. The only way
to deal with this type of race is to never inherit (Dmitry's point about
the standard in/out/err should already be dealt with it). So yes, I
think it is important to dig into some of the history and to understand
whether we regressed somewhere and why we have the flag set this way.
More information about the net-dev