Crash on 0.30.1+ when creating immediate child viewports #3800
Labels
Comments
7 tasks
|
Update: It seems like this was a false alarm and caused by unsafe undefined behaviour in egui. Waiting a bit for the winit upgrade PR to be fully tested & landed, and will close this then. Apologies for the noise! |
emilk
added a commit
to emilk/egui
that referenced
this issue
Jul 31, 2024
* Closes #1918 * Closes #4437 * Closes #4709 * [x] I have followed the instructions in the PR template Hiya, I need new winit for a specific fix for a android_native_actvity. There are already two PRs, but both don't seem to have a lot of movement, or are entirely complete: #4466 Seems to have gone stale & is missing some bits. #4702 Also seems stale (if less so), and is missing a refactor to run_on_demand. I also *think* the accesskit integration has a mistake and can't be enabled. I've marked them as a co-author on this as I started from this branch. (I think! Haven't done that on git before...). Sorry for the wall of text but just dumping some details / thoughts here: - There's an issue with creating child windows in winit 0.30.1 and up on macOS. The multiple_viewports, "create immediate viewport" example crashes on anything later 0.30.1, with a stack overflow in unsafe code. I've create [a winit issue](rust-windowing/winit#3800), it *might* already be fixed in 0.31.0 but I can't test as 0.31 will likely require another refactoring. For now I have just pinned things to 0.30.0 exatly. - Winit has deprecated run_on_demand, instead requiring the ApplicationHandler interface. In 0.31.0 run_on_demand is removed. I've refactored both the integration and the WinitApp trait to follow this pattern. I've left user_events a bit more opaque, as it seems 0.31.0 is doing a rework of UserEvents too. - I've used the new lazy init approach for access kit from this branch https://github.com/mwcampbell/egui/tree/accesskit-new-lazy-init and marked Matt as co-author, thanks Matt! - There was very similair but not quite the same code for run_and_return and run_and_exit. I've merged them, but looking at the github issues graveyard it seems vey finnicky. I *hope* this is more robust than before but it's a bit scary. - when receiving new_events this also used to check the redraw timing dictionary. That doesn't seem necesarry so left this out, but that is a slight behaviour change? - I have reeneabled serial_windows on macOS. I wondered whether it was fixed after this PR and does seem to be! However, even before this PR it seems to work, so maybe winit has sorted things out before that... Windows also works fine now without the extra hack. - I've done a very basic test of AccessKit on Windows and screen reader seems ok but I'm really not knowleadgable enough to say whether it's all good or not. - I've tested cargo tests & all examples on Windows & macOS, and ran a basic Android app. Still, testing native platforms is wel... hard so if anyone can test linux / iOs / older mac versions / windows 10 would probably be a good idea! - For consistencys sake I've made all event like functions in WinitApp return a `Result<EventResult>`. There's quite a bit of Ok-wrapping now, maybe too annoying? Not sure. Thank you for having a look! # Tested on * [x] macOS * [x] Windows * [x] Wayland (thanks [SiebenCorgie](https://github.com/SiebenCorgie)) * [x] X11 (thanks [crumblingstatue](https://github.com/crumblingstatue)!, [SiebenCorgie](https://github.com/SiebenCorgie)) # TODO * [x] Fix "follow system theme" not working on initial startup (winit issue, pinning to 0.30.2 for now). * [x] Fix `request_repaint_after` --------- Co-authored-by: mwcampbell <mattcampbell@pobox.com> Co-authored-by: j-axa <josef.axa@gmail.com> Co-authored-by: DataTriny <datatriny@gmail.com> Co-authored-by: Emil Ernerfeldt <emil.ernerfeldt@gmail.com>
rib
pushed a commit
to EmbarkStudios/egui
that referenced
this issue
Sep 30, 2024
* Closes emilk#1918 * Closes emilk#4437 * Closes emilk#4709 * [x] I have followed the instructions in the PR template Hiya, I need new winit for a specific fix for a android_native_actvity. There are already two PRs, but both don't seem to have a lot of movement, or are entirely complete: emilk#4466 Seems to have gone stale & is missing some bits. emilk#4702 Also seems stale (if less so), and is missing a refactor to run_on_demand. I also *think* the accesskit integration has a mistake and can't be enabled. I've marked them as a co-author on this as I started from this branch. (I think! Haven't done that on git before...). Sorry for the wall of text but just dumping some details / thoughts here: - There's an issue with creating child windows in winit 0.30.1 and up on macOS. The multiple_viewports, "create immediate viewport" example crashes on anything later 0.30.1, with a stack overflow in unsafe code. I've create [a winit issue](rust-windowing/winit#3800), it *might* already be fixed in 0.31.0 but I can't test as 0.31 will likely require another refactoring. For now I have just pinned things to 0.30.0 exatly. - Winit has deprecated run_on_demand, instead requiring the ApplicationHandler interface. In 0.31.0 run_on_demand is removed. I've refactored both the integration and the WinitApp trait to follow this pattern. I've left user_events a bit more opaque, as it seems 0.31.0 is doing a rework of UserEvents too. - I've used the new lazy init approach for access kit from this branch https://github.com/mwcampbell/egui/tree/accesskit-new-lazy-init and marked Matt as co-author, thanks Matt! - There was very similair but not quite the same code for run_and_return and run_and_exit. I've merged them, but looking at the github issues graveyard it seems vey finnicky. I *hope* this is more robust than before but it's a bit scary. - when receiving new_events this also used to check the redraw timing dictionary. That doesn't seem necesarry so left this out, but that is a slight behaviour change? - I have reeneabled serial_windows on macOS. I wondered whether it was fixed after this PR and does seem to be! However, even before this PR it seems to work, so maybe winit has sorted things out before that... Windows also works fine now without the extra hack. - I've done a very basic test of AccessKit on Windows and screen reader seems ok but I'm really not knowleadgable enough to say whether it's all good or not. - I've tested cargo tests & all examples on Windows & macOS, and ran a basic Android app. Still, testing native platforms is wel... hard so if anyone can test linux / iOs / older mac versions / windows 10 would probably be a good idea! - For consistencys sake I've made all event like functions in WinitApp return a `Result<EventResult>`. There's quite a bit of Ok-wrapping now, maybe too annoying? Not sure. Thank you for having a look! * [x] macOS * [x] Windows * [x] Wayland (thanks [SiebenCorgie](https://github.com/SiebenCorgie)) * [x] X11 (thanks [crumblingstatue](https://github.com/crumblingstatue)!, [SiebenCorgie](https://github.com/SiebenCorgie)) * [x] Fix "follow system theme" not working on initial startup (winit issue, pinning to 0.30.2 for now). * [x] Fix `request_repaint_after` --------- Co-authored-by: mwcampbell <mattcampbell@pobox.com> Co-authored-by: j-axa <josef.axa@gmail.com> Co-authored-by: DataTriny <datatriny@gmail.com> Co-authored-by: Emil Ernerfeldt <emil.ernerfeldt@gmail.com>
486c
pushed a commit
to 486c/egui
that referenced
this issue
Oct 9, 2024
* Closes emilk#1918 * Closes emilk#4437 * Closes emilk#4709 * [x] I have followed the instructions in the PR template Hiya, I need new winit for a specific fix for a android_native_actvity. There are already two PRs, but both don't seem to have a lot of movement, or are entirely complete: emilk#4466 Seems to have gone stale & is missing some bits. emilk#4702 Also seems stale (if less so), and is missing a refactor to run_on_demand. I also *think* the accesskit integration has a mistake and can't be enabled. I've marked them as a co-author on this as I started from this branch. (I think! Haven't done that on git before...). Sorry for the wall of text but just dumping some details / thoughts here: - There's an issue with creating child windows in winit 0.30.1 and up on macOS. The multiple_viewports, "create immediate viewport" example crashes on anything later 0.30.1, with a stack overflow in unsafe code. I've create [a winit issue](rust-windowing/winit#3800), it *might* already be fixed in 0.31.0 but I can't test as 0.31 will likely require another refactoring. For now I have just pinned things to 0.30.0 exatly. - Winit has deprecated run_on_demand, instead requiring the ApplicationHandler interface. In 0.31.0 run_on_demand is removed. I've refactored both the integration and the WinitApp trait to follow this pattern. I've left user_events a bit more opaque, as it seems 0.31.0 is doing a rework of UserEvents too. - I've used the new lazy init approach for access kit from this branch https://github.com/mwcampbell/egui/tree/accesskit-new-lazy-init and marked Matt as co-author, thanks Matt! - There was very similair but not quite the same code for run_and_return and run_and_exit. I've merged them, but looking at the github issues graveyard it seems vey finnicky. I *hope* this is more robust than before but it's a bit scary. - when receiving new_events this also used to check the redraw timing dictionary. That doesn't seem necesarry so left this out, but that is a slight behaviour change? - I have reeneabled serial_windows on macOS. I wondered whether it was fixed after this PR and does seem to be! However, even before this PR it seems to work, so maybe winit has sorted things out before that... Windows also works fine now without the extra hack. - I've done a very basic test of AccessKit on Windows and screen reader seems ok but I'm really not knowleadgable enough to say whether it's all good or not. - I've tested cargo tests & all examples on Windows & macOS, and ran a basic Android app. Still, testing native platforms is wel... hard so if anyone can test linux / iOs / older mac versions / windows 10 would probably be a good idea! - For consistencys sake I've made all event like functions in WinitApp return a `Result<EventResult>`. There's quite a bit of Ok-wrapping now, maybe too annoying? Not sure. Thank you for having a look! * [x] macOS * [x] Windows * [x] Wayland (thanks [SiebenCorgie](https://github.com/SiebenCorgie)) * [x] X11 (thanks [crumblingstatue](https://github.com/crumblingstatue)!, [SiebenCorgie](https://github.com/SiebenCorgie)) * [x] Fix "follow system theme" not working on initial startup (winit issue, pinning to 0.30.2 for now). * [x] Fix `request_repaint_after` --------- Co-authored-by: mwcampbell <mattcampbell@pobox.com> Co-authored-by: j-axa <josef.axa@gmail.com> Co-authored-by: DataTriny <datatriny@gmail.com> Co-authored-by: Emil Ernerfeldt <emil.ernerfeldt@gmail.com>
|
@ArthurBrussee We're seeing the same crash in Servo on winit 0.30.5. Do you have more information about what egui was doing incorrectly? |
|
Oh strange! In egui the cause was some bad unsafe code. Specifically around these bits: emilk/egui@6f2f006#diff-31986c1b45795fe76fd65d559692c61cb9fac99582526da317a37598d1b97d22L321 I can't 100% remember what was wrong - it was trying to pass the event_loop to child windows, but the claimed lifetime guarantees there weren't right at all. That was solved with an "EventLoopGaurd" emilk/egui@6f2f006#diff-61057fd82c6983a59c1045857b33f20b6d20ed7836019c46f6f842ac3e974505R1 I find it hard to say what exactly changed in winit that triggered this and how the crash relates, as it all might just be lottery of undefined behaviour, but if you have some unsafe code around event loops, I guess check those! |
|
That sounds extremely similar to what we're doing in Servo, complete with lifetime shenanigans around event loops. Thanks for the pointers! |
hacknus
pushed a commit
to hacknus/egui
that referenced
this issue
Oct 30, 2024
* Closes emilk#1918 * Closes emilk#4437 * Closes emilk#4709 * [x] I have followed the instructions in the PR template Hiya, I need new winit for a specific fix for a android_native_actvity. There are already two PRs, but both don't seem to have a lot of movement, or are entirely complete: emilk#4466 Seems to have gone stale & is missing some bits. emilk#4702 Also seems stale (if less so), and is missing a refactor to run_on_demand. I also *think* the accesskit integration has a mistake and can't be enabled. I've marked them as a co-author on this as I started from this branch. (I think! Haven't done that on git before...). Sorry for the wall of text but just dumping some details / thoughts here: - There's an issue with creating child windows in winit 0.30.1 and up on macOS. The multiple_viewports, "create immediate viewport" example crashes on anything later 0.30.1, with a stack overflow in unsafe code. I've create [a winit issue](rust-windowing/winit#3800), it *might* already be fixed in 0.31.0 but I can't test as 0.31 will likely require another refactoring. For now I have just pinned things to 0.30.0 exatly. - Winit has deprecated run_on_demand, instead requiring the ApplicationHandler interface. In 0.31.0 run_on_demand is removed. I've refactored both the integration and the WinitApp trait to follow this pattern. I've left user_events a bit more opaque, as it seems 0.31.0 is doing a rework of UserEvents too. - I've used the new lazy init approach for access kit from this branch https://github.com/mwcampbell/egui/tree/accesskit-new-lazy-init and marked Matt as co-author, thanks Matt! - There was very similair but not quite the same code for run_and_return and run_and_exit. I've merged them, but looking at the github issues graveyard it seems vey finnicky. I *hope* this is more robust than before but it's a bit scary. - when receiving new_events this also used to check the redraw timing dictionary. That doesn't seem necesarry so left this out, but that is a slight behaviour change? - I have reeneabled serial_windows on macOS. I wondered whether it was fixed after this PR and does seem to be! However, even before this PR it seems to work, so maybe winit has sorted things out before that... Windows also works fine now without the extra hack. - I've done a very basic test of AccessKit on Windows and screen reader seems ok but I'm really not knowleadgable enough to say whether it's all good or not. - I've tested cargo tests & all examples on Windows & macOS, and ran a basic Android app. Still, testing native platforms is wel... hard so if anyone can test linux / iOs / older mac versions / windows 10 would probably be a good idea! - For consistencys sake I've made all event like functions in WinitApp return a `Result<EventResult>`. There's quite a bit of Ok-wrapping now, maybe too annoying? Not sure. Thank you for having a look! # Tested on * [x] macOS * [x] Windows * [x] Wayland (thanks [SiebenCorgie](https://github.com/SiebenCorgie)) * [x] X11 (thanks [crumblingstatue](https://github.com/crumblingstatue)!, [SiebenCorgie](https://github.com/SiebenCorgie)) # TODO * [x] Fix "follow system theme" not working on initial startup (winit issue, pinning to 0.30.2 for now). * [x] Fix `request_repaint_after` --------- Co-authored-by: mwcampbell <mattcampbell@pobox.com> Co-authored-by: j-axa <josef.axa@gmail.com> Co-authored-by: DataTriny <datatriny@gmail.com> Co-authored-by: Emil Ernerfeldt <emil.ernerfeldt@gmail.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Description
Hiya! I've been doing some work for a winit upgrade for egui (emilk/egui#4849), but, it seems that this commit has introduced a crash on macOS. The multiple_viewports example spawns a immediate child viewport, but on macOS this crashes with:
Segmentation fault: 11It seems some precondition of retain is violated (though it's a safe function? Is
objc2wrong? Really above my paygrade I'm afraid!)This might already be fixed in 0.31.0 as I think the code in that area has changed a bit, but either way wanted to let you know.
macOS version
Winit version
0.30.0 works, 0.30.1+ crashes
The text was updated successfully, but these errors were encountered: