On Sat, 10 Oct 2026 13:12:40 GMT, Alan Bateman <[email protected]> wrote:
> [JDK-8334574](https://bugs.openjdk.org/browse/JDK-8334574) tracks the Windows > issue of events getting "lost" where there is more than one AFD_POLL > outstanding for a socket. This can cause impact in scenarios where one > virtual thread blocks reading from a socket and another virtual treads blocks > writing to the same socket. There is good discussion in > [wepoll/issues/35](https://github.com/piscisaureus/wepoll/issues/35) about > this. A future PR may propose a new "poller group" implementation, prototypes > so far have required bookkeeping that impacts performance. > > This PR proposes a change that works around the issue by configuring Windows > to not have any write poller threads. It would be rare for write to block, > esp. if SO_SNDBUF is set to a sufficiently large value. If a write op does > need to block, the write will timed-park and retry until the write succeeds > (or fails). The change has no impact on other platforms. > > --------- > - [x] I confirm that I make this contribution in accordance with the [OpenJDK > Interim AI Policy](https://openjdk.org/legal/ai). Looks reasonable. Definitely better than blocking forever. ------------- Marked as reviewed by djelinski (Reviewer). PR Review: https://git.openjdk.org/jdk/pull/33297#pullrequestreview-5493364004
