chore: fix windows CI #2597
chore: fix windows CI #2597
Conversation
| @@ -601,6 +601,7 @@ rt_test! { | |||
| } | |||
|
|
|||
| #[test] | |||
| #[cfg(not(windows))] | |||
There was a problem hiding this comment.
This test can't pass on windows. For now, I'm going to disable it. I will document why before completing the PR. I mostly want to see if CI passes.
| @@ -221,7 +221,7 @@ impl OwnedWriteHalf { | |||
| impl Drop for OwnedWriteHalf { | |||
| fn drop(&mut self) { | |||
| if self.shutdown_on_drop { | |||
| let _ = self.inner.shutdown(Shutdown::Both); | |||
| let _ = self.inner.shutdown(Shutdown::Write); | |||
There was a problem hiding this comment.
Shutdown::Read has no defined portable behavior. This is why it doesn't pass on windows.
|
I think this is now no longer has an equivalent behavior. The idea behind shutting down the reading end was to emulate the same kind of behavior that e.g. socket APIs in Go and Java have: If you close the socket in those environments on one handle (e.g. a writing thread), it will unblock any operation blocked on receive. This helps if you have have one of those typical scenarios where there is a reader and a writer task and the reader is basically permanently blocked on the socket since it just reads in a loop. If it doesn't get woken by the close it will be stuck there until the peer also closes. Now doing this with shutdown instead of close is not exactly equivalent, but nevertheless was intended to do that job. If that behavior is removed then any application which already relies on that behavior would get broken. The wakeup could probably be implemented on top of the native socket by grabbing the read waker from somewhere when the writer is closed, and storing and checking an extra "closed", but it's certainly more work. This would now get balanced again breaking the API. |
|
Additionally, dropping a write half should only impact the write half of things, not the read. it is incorrect for Tokio to shutdown the read half when the write is dropped. This change is fixing a bug. |
|
@Darksonn IMO the API needs to be either fixed by making the change in this PR or deprecated w/ the reason that it cannot be implemented in a portable way and done again in 0.3. Which route do you want to take? |
|
Considering that we are changing behavior on some OS with either option, I think it's ok to just fix it. |
- docs: misc improvements (#2572, #2658, #2663, #2656, #2647, #2630, #2487, #2621, #2624, #2600, #2623, #2622, #2577, #2569, #2589, #2575, #2540, #2564, #2567, #2520, #2521, #2572, #2493) - rt: allow calls to `block_on` inside calls to `block_in_place` that are themselves inside `block_on` (#2645) - net: fix non-portable behavior when dropping `TcpStream` `OwnedWriteHalf` (#2597) - io: improve stack usage by allocating large buffers on directly on the heap (#2634) - io: fix unsound pin projection in `AsyncReadExt::read_buf` and `AsyncWriteExt::write_buf` (#2612) - io: fix unnecessary zeroing for `AsyncRead` implementors (#2525) - io: Fix `BufReader` not correctly forwarding `poll_write_buf` (#2654) - coop: returning `Poll::Pending` no longer decrements the task budget (#2549) - io: little-endian variants of `AsyncReadExt` and `AsyncWriteExt` methods (#1915) - io: fix panic in `AsyncReadExt::read_line` (#2541) - task: add [`tracing`] instrumentation to spawned tasks (#2655) - sync: allow unsized types in `Mutex` and `RwLock` (via `default` constructors) (#2615) - net: add `ToSocketAddrs` implementation for `&[SocketAddr]` (#2604) - fs: add `OpenOptionsExt` for `OpenOptions` (#2515) - fs: add `DirBuilder` (#2524) [`tracing`]: https://crates.io/crates/tracing Signed-off-by: Eliza Weisman <[email protected]>
# 0.2.22 (July 2!, 2020) ### Fixes - docs: misc improvements (#2572, #2658, #2663, #2656, #2647, #2630, #2487, #2621, #2624, #2600, #2623, #2622, #2577, #2569, #2589, #2575, #2540, #2564, #2567, #2520, #2521, #2493) - rt: allow calls to `block_on` inside calls to `block_in_place` that are themselves inside `block_on` (#2645) - net: fix non-portable behavior when dropping `TcpStream` `OwnedWriteHalf` (#2597) - io: improve stack usage by allocating large buffers on directly on the heap (#2634) - io: fix unsound pin projection in `AsyncReadExt::read_buf` and `AsyncWriteExt::write_buf` (#2612) - io: fix unnecessary zeroing for `AsyncRead` implementors (#2525) - io: Fix `BufReader` not correctly forwarding `poll_write_buf` (#2654) - io: fix panic in `AsyncReadExt::read_line` (#2541) ### Changes - coop: returning `Poll::Pending` no longer decrements the task budget (#2549) ### Added - io: little-endian variants of `AsyncReadExt` and `AsyncWriteExt` methods (#1915) - task: add [`tracing`] instrumentation to spawned tasks (#2655) - sync: allow unsized types in `Mutex` and `RwLock` (via `default` constructors) (#2615) - net: add `ToSocketAddrs` implementation for `&[SocketAddr]` (#2604) - fs: add `OpenOptionsExt` for `OpenOptions` (#2515) - fs: add `DirBuilder` (#2524) [`tracing`]: https://crates.io/crates/tracing Signed-off-by: Eliza Weisman <[email protected]>
Trying to see if windows CI can be fixed.