RFR: 8338569: HTTP/1.1 CleanupTrigger may be triggerred after the next exchange started [v2]

[Daniel Fuchs]

On Thu, 22 Aug 2024 13:43:28 GMT, Daniel Fuchs <dfuchs at openjdk.org> wrote:

>> The java/net/httpclient/DigestEchoClient.java fails intemittently with IOException: HTTP/1.1 header parser received no bytes.
>> 
>> Analysis shows that this is caused by the CleanupTrigger which receives data after the reused connection has been taken out of the HTTP/1.1 clear pool. This should not happen.
>> 
>> The issue is caused by deferred registration of read subscribers with the SocketTube. The subscribers are registered within the SelectorManager thread to ensure proper interaction with the read method.
>> 
>> Sometimes pending subscribers are pushed faster than they are actually subscribed, which may cause the wrong subscriber to be subscribed at the wrong time. 
>> 
>> This is a redo of [JDK-8336655](https://bugs.openjdk.org/browse/JDK-8336655) which was a failed fix.
>> The main difference is that the new fix uses a ConcurrentLinkedQueue to make sure all subscribers get subscribed and dropped in the right sequence. Marking previous subscription as stopped ensures that data can only be routed to the last subscriber in the queue.
>
> Daniel Fuchs has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Better exception message for ShutdownNow

Note: the changes to ShutdownNow.java are coincidental. I observed this test failing once in my many tests, and decided to make the exception it throws more explicit. FWIW - I had also observed this test failing before the change, it's intermittent and rare. I took the opportunity to raise the timeout from 2000 to 2500.

-------------

PR Comment: https://git.openjdk.org/jdk/pull/20623#issuecomment-2304753871