8199329: Remove code that attempts to read bytes after connection reset reported

David Lloyd david.lloyd at redhat.com
Wed Mar 14 14:47:15 UTC 2018


+1 from me (non-reviewer).  Semantically, it is impossible to
distinguish the difference between the OS dropping bytes when it
receives an RST compared to Java doing the same thing, so there is no
observable behavior change here AFAICT.

On Wed, Mar 14, 2018 at 9:30 AM, Alan Bateman <Alan.Bateman at oracle.com> wrote:
>
> Classic networking has some curious code to deal with connection resets. I
> needed to dig into ancient history to find the issues and original
> motivations.
>
> When a connection reset is reported (usually ECONNRESET) then further
> attempts to read from the socket will typically return -1 (EOF). Classic
> networking papers over this so that further attempts with the socket APIs to
> read bytes will continue to throw SocketException "connection reset".
> Furthermore, it allows for cases where the platform can read bytes from the
> socket buffer after ECONNRESET has been reported. None of the main stream
> platforms do this and it's hard to imagine anything depending on such
> unpredictable behavior. I'm running into this odd code as part of cleanup to
> the read/write code paths and reducing them down to a single blocking call.
>
> I would like to remove the legacy/undocumented behavior that attempts to
> read beyond the connection reset. The code to consistently throw IOException
> once ECONNRESET is returned is retained as it's possible that code relies on
> this.
>
> The proposed changes are here:
>    http://cr.openjdk.java.net/~alanb/8199329/webrev/
>
> All existing tests pass with these changes.
>
> -Alan



-- 
- DML


More information about the net-dev mailing list