8203059: (so) Closing a socket channel registered with Selector and with SO_LINGER set to 0 does not reset connection

[Chris Hegarty]

> On 14 May 2018, at 14:09, Alan Bateman <Alan.Bateman at oracle.com> wrote:
> 
> This is a follow-up to the "Behaviour of SocketChannelImpl.close() in Java11 ..." thread where it was observed that closing a SocketChannel registered with a Selector does not obey the linger on close interval when enabled.
> 
> As discussed in that thread, the SO_LINGER socket option is not specified for channels configured non-blocking. At the same time, it can be useful to enable to a value of 0 to force a connection reset. As things stand, closing a socket channel with SO_LINGER enabled will obey the linger value when the channel is not registered with a Selector (it doesn't matter if the channel is configured blocking or non-blocking). It does not obey the linger value when the channel is registered with a Selector - this is because closing a SocketChannel registered with a Selector first shuts down the connection for writing (so that the peer reads EOF).
> 
> The following patches changes close so that SO_LINGER can be be used force a connection reset for the case that a channel is registered with a Selector and the channel is flushed from the Selector in timely manner. The linger interval > 0 case is too problematic and it can delay selection operations so the change will disable it and avoid selection operations blocking when they deregister closed channels. The test case exercises all the combinations and ensures that the behavior is consistent across all platforms.
>     http://cr.openjdk.java.net/~alanb/8203059/webrev/

I’ve studied the changes and the surrounding code. Looks good.

-Chris