You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
atspi-common: register_tree panics (then aborts) when adapters become active out of order: push_adapter doesn't keep the list sorted for adapter_index's binary search #820
AppContext in accesskit_atspi_common finds adapters with a binary search, but push_adapter appends to the end, so the list is only sorted if adapters are added in order of id:
With accesskit_unix, an adapter is pushed when it goes from Pending to Active, which happens in the application's next update_if_active for that window after AT-SPI is enabled. With several windows, that is whatever order the application updates them in, not the order they were created, so the list can end up out of order (for example ids [1, 2, 0]). A binary search then misses an adapter that is in the list, and register_tree panics on
let adapter_index = app_context.adapter_index(self.id).unwrap();
while holding the app context's write lock. Unwinding drops the adapter, whose Drop calls write_app_context() on the now poisoned lock and panics again, so the process aborts:
panicked at accesskit_atspi_common-0.18.1/src/adapter.rs:461
panicked at accesskit_atspi_common-0.18.1/src/context.rs:85
panic in a destructor during cleanup
thread caused non-unwinding panic. aborting.
remove_adapter has the same problem in a quieter form: on an unsorted list it can fail to find the adapter and leave it in the list.
What should happen
Adapters can become active in any order without the lookup failing.
How to reproduce
We hit it in an application with several windows, each with an accesskit_winit adapter, on Ubuntu 26.04 (GNOME 50, Wayland and XWayland) with AT-SPI enabled, when the windows were updated in a different order than they were created. It depends on that order, so we don't have a small program that hits it every time; the lookup itself shows it. With the adapters pushed in the order 1, 2, 0:
let adapters:Vec<(usize,())> = vec![(1,()),(2,()),(0,())];let found = adapters.binary_search_by(|adapter| adapter.0.cmp(&0));println!("{found:?}");// Err(0): adapter 0 is in the list but isn't found
We're running with exactly this change (on 0.18.1, which egui 0.36 uses) and haven't seen the crash since. The code is the same on main (context.rs lines 109–115, adapter.rs line 463).
Versions: accesskit_atspi_common 0.18.1 through accesskit_unix 0.21.1 and accesskit_winit; also present in 0.21.0 and on main.
Activity