Hi Chris,
TwoStacksPlainDatagramSocketImpl.java:136 could be explicit about the
native bind0
changing the state.
And I agree with Alan that the calls to updateCleaner (:144) should
occur whether or not
the native bind0 throws an exception.
Thanks, Roger
On 3/28/2018 1:22 PM, Alan Bateman wrote:
On 28/03/2018 10:08, Chris Hegarty wrote:
8195059 [1] updated the socket implementation to use a cleaner,
rather than the previous finalizer implementation. Unfortunately the
TwoStacksPlainDatagramSocketImpl on Windows, that uses two
native file descriptors, may close one or other when binding, or it
may in fact close both sockets and create two completely new
ones. This is problematic since the cleaner has been registered
with the native file descriptors of the original `fd` and `fd1` when
the impl was created. As such, the cleaner may interfere with
other resources that happen to re-use the native file descriptor.
This area of code is ultimately destined for replacement with the
dual stack implementation, but unfortunately we’re not quite at
that point yet, so I would like to resolve this issue with a small
localised fix that checks the file descriptors before and after the
actual bind call, unregistering and re-registering the cleaner as
necessary.
http://cr.openjdk.java.net/~chegar/8200304/00/
It's great that this issue has been tracked down.
bind0 can try an exception so shouldn't there be a try-finally around
this? Also would it be saner to unregister before bind0 and register
afresh after it is called so that there is no point where a closed
socket is registered with the cleaner?
-Alan
