RFR: 8278326: Socket close is not thread safe and other cleanup
Alan Bateman
alanb at openjdk.org
Mon Jan 9 10:30:53 UTC 2023
On Mon, 9 Jan 2023 10:01:05 GMT, Michael McMahon <michaelm at openjdk.org> wrote:
>> java.net.Socket is not specified to be thread safe but it is required to support async close. If you create an unbound Socket and close it at around the time that another thread is binding, connecting, or anything else that creates the underlying socket then it can leak. The simplest thing to do is to synchronize all methods but the underlying SocketImpl implementation is thread safe, and all we really need is for Socket (and ServerSocket) to synchronize the creation of the underlying socket (SocketImpl.create) with close. As part of this change I've replaced the 6 flags with a bit mask. A new test is added to the Socket/asyncClose directory to test closing concurrently with another operation, the test will detect if the closed Socket is connected to a SocketImpl with an open socket.
>>
>> Related is that ServerSocket.implAccept can be overridden to provide the Socket to accept. Its behavior is unspecified when called with a Socket that isn't newly created/unbound and there are number of silly scenarios that can arise. I've changed implAccept to coordinate with close so that accept doesn't return a closed Socket that is connected to an underlying socket. A new test is added to exercise these scenarios.
>>
>> There are a couple of random cleanup/formatting nits in this patch.
>
> src/java.base/share/classes/java/net/ServerSocket.java line 98:
>
>> 96: this.impl = Objects.requireNonNull(impl);
>> 97: }
>> 98:
>
> What's the reason for adding this new private constructor? I don't see what the problem with the original single arg one is. `checkPermission` could be declared to return void instead of Void though.
Permission checks in constructors of non-final classes are problematic when the subclass has a finalizer. It's fixed here with the same idiom that we use in other places. The return is used by as parameter so it has to be Void.
-------------
PR: https://git.openjdk.org/jdk/pull/11863
More information about the net-dev
mailing list