On Mon, 17 Oct 2022 19:09:11 GMT, Daniel Fuchs <[email protected]> wrote:
>> When [JDK-8277969](https://bugs.openjdk.org/browse/JDK-8277969) was >> implemented, a list of outstanding response subscribers was added to >> `HttpClientImpl`. A body subscriber is added to the list after being created >> and is removed from the list when it is completed, either successfully or >> exceptionally. >> >> It appears that in the case where the subscription is cancelled before the >> subscriber is completed, the subscriber might remain registered in the list >> forever, or at least until the HttpClient gets garbage collected. This can >> be easily reproduced using streaming subscribers, such as >> BodySubscriber::ofInputStream. In the case where the input stream is closed >> without having read all the bytes, Subscription::cancel will be called. >> Whether the subscriber gets unregistered or not at that point becomes racy. >> >> Indeed, the reactive stream specification doesn't guarantee whether >> onComplete or onError will be called or not after a subscriber cancels its >> subscription. Any cleanup that would have been performed by >> onComplete/onError might therefore need to be performed when the >> subscription is cancelled too. > > Daniel Fuchs has updated the pull request incrementally with one additional > commit since the last revision: > > Incorporated review feedback I have updated the PR. There's a somewhat larger change in how tests can now track HttpClients. I noticed that `ReferenceTracker::check(gracedelay)` wasn't actually checking that all client selector managers threads had exited (as I thought it did) so I changed it to do that, and the SmallTimeout test started failing, so I had to fix that test to null out references and actually wait until all tasks had been executed. ------------- PR: https://git.openjdk.org/jdk/pull/10659
