Prevent hanging on Windows when exiting within 10s of a HTTP request - #493
Merged
Merged
Conversation
hyper resolves host names on tokio's blocking pool, whose threads linger for ten seconds after their last job, and dropping a runtime waits for each of them to acknowledge the shutdown. This used to be a thread local, which is dropped when its thread exits. On Windows the main thread's thread locals are dropped during ExitProcess, after the OS has already terminated every other thread, so a program that had sent a request in its last ten seconds waited forever for a pool thread that no longer existed. Leaking the runtime at exit costs nothing, since the OS reclaims it either way.
The unit test could not fail for the right reason: with the thread local it replaced, `TOKIO_RUNTIME.spawn_blocking` does not compile, so a revert is a build error rather than a failing test, and an explicit shutdown added later would keep it green. Nothing else could catch it either. hyper-util's connector parses an IP literal before it reaches a resolver (`SocketAddrs::try_parse`), so every example that talks to 127.0.0.1 skips `GaiResolver`, and with it the blocking pool whose shutdown is what hangs. Measured against the test server, an IP literal starts zero blocking-pool threads and a host name starts one. So point examples/http-client.roc at `localhost`. Both of its run cases now reach the pool and then exit immediately, which is the bug, and the connection-refused case does it without a server at all. http-simple.roc stays on the IP literal so the fast path keeps its coverage. The test server binds ::1 as well as 127.0.0.1, because Windows answers `localhost` with ::1 first; ::1 goes up first since scripts/test.py polls 127.0.0.1 for readiness. Move the reasoning to the static it explains, naming the two ways to reintroduce the hang: give the runtime an owner, or shut it down alongside `process_service::shutdown`. The test comment now describes the test. Give listener.rs the same static, since a thread-local runtime is the trap itself, not a thing to document. Nothing there reaches the blocking pool today -- socket2 binds and crate::tcp resolves -- and a static keeps it that way by construction, while letting listeners register with one IO driver. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
hyper resolves host names on tokio's blocking pool, whose threads linger for ten seconds after their last job, and dropping a runtime waits for each of them to acknowledge the shutdown. This used to be a thread local, which is dropped when its thread exits. On Windows the main thread's thread locals are dropped during ExitProcess, after the OS has already terminated every other thread, so a program that had sent a request in its last ten seconds waited forever for a pool thread that no longer existed. Leaking the runtime at exit costs nothing, since the OS reclaims it either way.