RFR 8231187: SelectorProvider.inheritedChannel() returns TCP socket channel for Unix domain socket

[Michael McMahon]

Hi Alan,


On 20/09/2019, 11:42, Alan Bateman wrote:
> On 20/09/2019 10:53, Michael McMahon wrote:
>> Hi,
>>
>> This issue was reported through a recent discussion proposing an API 
>> for Unix domain
>> sockets. It fixes the current erroneous behavior where TCP type 
>> SocketChannels are returned
>> and instead returns a Channel type which handles connected AF_UNIX 
>> sockets specifically.
>>
>> The fix also enables the whole inherited channel mechanism for Mac, 
>> which up to now
>> has not been working.
>>
>> http://cr.openjdk.java.net/~michaelm/8231187/webrev.1/index.html
> Yes, I remember that bug came up in the proposal to add the API so 
> it's good to get it fixed.
>
> I skimmed through the webrev. The async close doesn't look right, 
> mostly missing preClose and the final close by the last thread doing 
> I/O. The Unix implementation of SourceChannelImpl or SinkChannelImpl 
> are the simplest examples to follow if you need it. The main issues are:
>
> - need a state field to track that the channel is closing
> - endRead need to close the channel when the state is_ST_CLOSING and 
> there is no writer (the opposite for reader)
> - implCloseChannel should only close if reader and writer are both 0, 
> otherwise it should preClose and signal the reader and/or writer.
>

Is the above not required only to allow draining of the socket when 
SO_LINGER is enabled,
which is not supported in this case?

> A minor nit is that getFDVal can be removed, and getFD can be private 
> as the channel doesn't implement SelChImpl.
>

OK
> As a precaultion, createChannel could return null if the address 
> family is AF_UKNOWN.
>
OK

Thanks,
Michael.