RFR: 8286194: ExecutorShutdown test fails intermittently

[Jaikiran Pai]

On Thu, 5 May 2022 19:03:13 GMT, Daniel Fuchs <dfuchs at openjdk.org> wrote:

> Hi, please find here a patch that solves a rare intermittent test failure observed in the test `java/net/httpclient/ExecutorShutdown.java`
> 
> A race condition coupled with some too eager synchronization was causing a deadlock between an Http2Connection close, a thread trying to shutdown the HttpClient due to a RejectedTaskException, and the SelectorManager thread trying to exit.
> 
> The fix comprises several cleanup - in particular:
> 
> - `Http2Connection`: no need to try to send a `GOAWAY` frame if the underlying TCP connection is already closed
> - `SSLFlowDelegate`/`SubscriberWrapper`: no need to trigger code that will request more data from upstream if the sequential scheduler that is supposed to handle that data once it arrives is already closed
> - `Http1Exchange`/`Http1Request`: proper cancellation of subscription if an exception is raised before `onSubscribe()` has been called
> - `HttpClientImpl`: avoid calling callbacks from within synchronized blocks when not necessary
> - `ReferenceTracker`: better thread dumps in case where the selector is still alive at the end of the test (remove the limit that limited the stack traces to 8 element max by no longer relying on `ThreadInfo::toString`)

src/java.net.http/share/classes/jdk/internal/net/http/common/SSLFlowDelegate.java line 784:

> 782: 
> 783:                 while (Utils.synchronizedRemaining(writeList) > 0 || hsTriggered() || needWrap()) {
> 784:                     if (scheduler.isStopped()) return;

If the `scheduler` is stopped then would this task instance be ever called again? If it won't be called again, then do you think we should perhaps drain the queued `writeList` to reduce any memory resource accumulation till the next GC cycle?

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

PR: https://git.openjdk.java.net/jdk/pull/8562