Found different in behavior of SocketInputStream#read in JDK13

[Alan Bateman]

On 01/08/2019 02:47, Enrico Olivelli wrote:
> :
>
> Just to be clear. What is the plan now ? to leave JDK13 as it is 
> currently or to change the behavior ?
> If I understand correctly Alan's answer means that the new behavior in 
> JDK13 is now consistent across Windows, Linux and Mac and we want to 
> keep it
> so it is better to adapt application code.
The test that Chris created to demonstrate zookeeper scenario is great 
as it makes it clear that the issue is closing the socket with bytes in 
the receive queue [1]. The JDK's behavior in this scenario has always 
been platform specific. On Windows, with java.net.Socket, the JDK has 
always shutdown the connection (when linger is not enabled). On other 
platforms it didn't do the shutdown. There are several subtle 
differences in the Windows TCP implemetation that means we can never 
change this without breaking many deployments (and dozens of JDK tests 
too). To add more complexity, NIO SocketChannel has to allow for closing 
when the channel is registered with a Selector and that will determine 
whether the connection is shutdown before the final close. So I think 
your updated zookeeper test is correct as it caters for the different 
possible behaviors that have always been possible.

I don't know yet if we are going to attempt to do anything for JDK 13. 
JDK 13 is in rampdown phase two and there is an workaround via a system 
property to bring back the old behavior. If we do fix it then I think 
the only feasible solution is to limit the shutdown to Windows as 
before. So we'll reply back soon on this.

-Alan

[1] http://cs.baylor.edu/~donahoo/practical/CSockets/TCPRST.pdf

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.java.net/pipermail/nio-dev/attachments/20190801/64125def/attachment.html>