fix(kms): join surface threads before exit to avoid SIGABRT on shutdown

When the kiosk child exits, calling process::exit() directly from the calloop
callback runs libc's atexit handlers (Mesa's util_queue_kill_threads) while a
KMS surface thread may still be mid-frame in eglCreateSync(). Concurrent Mesa
teardown and live EGL use corrupt the heap and glibc aborts the process
("free(): corrupted unsorted chunks", SIGABRT).

Stop the event loop cleanly via should_stop instead, stashing the child's exit
code in Common::kiosk_exit_code. After the loop returns, pause() each DRM device
(releasing master so the surface drop path skips its blocking clear_state(), and
the incoming compositor gets master immediately), then drop_and_join() every
surface so no thread is left in Mesa. Only then is process::exit() called, at
which point atexit is safe to run.

The join is done here rather than in Surface::Drop because an unconditional join
in Drop deadlocks against apply_config_for_outputs during normal operation; once
the event loop has stopped, the surface thread can process ThreadCommand::End
and exit without the main thread servicing any pending message.

https://github.com/pop-os/cosmic-comp/issues/2375

Authored by Claude Opus 4.8
Reviewed by Timo Strunk <Timo.Strunk@gmail.com>

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

src/lib.rs

@@ -16,7 +16,7 @@ use smithay::{
 };
 
 use anyhow::{Context, Result};
-use state::{LastRefresh, State};
+use state::{BackendData, LastRefresh, State};
 use std::{
     env,
     ffi::OsString,
@@ -222,19 +222,17 @@ pub fn run(hooks: crate::hooks::Hooks) -> Result<(), Box<dyn Error>> {
                 // Kiosk child exited with status
                 Ok(Some(exit_status)) => {
                     info!("Command exited with status {:?}", exit_status);
-                    match exit_status.code() {
-                        // Exiting with the same status as the kiosk child
-                        Some(code) => process::exit(code),
-                        // The kiosk child exited with signal, exiting with error
-                        None => process::exit(1),
-                    }
+                    // Stop cleanly so surface threads are joined before exit() (signal -> 1).
+                    state.common.kiosk_exit_code = Some(exit_status.code().unwrap_or(1));
+                    state.common.should_stop = true;
                 }
                 // Command still running
                 Ok(None) => {}
                 // Kiosk child disappeared, exiting with error
                 Err(err) => {
                     warn!(?err, "Failed to wait for command");
-                    process::exit(1);
+                    state.common.kiosk_exit_code = Some(1);
+                    state.common.should_stop = true;
                 }
             }
         }
@@ -245,10 +243,32 @@ pub fn run(hooks: crate::hooks::Hooks) -> Result<(), Box<dyn Error>> {
         let _ = child.kill();
     }
 
+    let kiosk_exit_code = state.common.kiosk_exit_code;
+
+    // Join surface threads before exit() so no thread is mid-eglCreateSync when
+    // Mesa's atexit handlers run and corrupt the heap (issue #2375). Safe here
+    // because the event loop has stopped; an unconditional join in Surface::Drop
+    // would instead deadlock against apply_config_for_outputs.
+    if let BackendData::Kms(kms) = &mut state.backend {
+        // Release master first so the surface drop path skips its blocking commit.
+        for device in kms.drm_devices.values_mut() {
+            device.drm.pause();
+        }
+        for device in kms.drm_devices.values_mut() {
+            for (_, surface) in device.inner.surfaces.drain() {
+                surface.drop_and_join();
+            }
+        }
+    }
+
     // drop eventloop & state before logger
     std::mem::drop(event_loop);
     std::mem::drop(state);
 
+    if let Some(code) = kiosk_exit_code {
+        process::exit(code);
+    }
+
     Ok(())
 }
 

src/state.rs

@@ -244,6 +244,7 @@ pub struct Common {
     pub clock: Clock<Monotonic>,
     pub startup_done: Arc<AtomicBool>,
     pub should_stop: bool,
+    pub kiosk_exit_code: Option<i32>,
 
     pub gesture_state: Option<GestureState>,
 
@@ -754,6 +755,7 @@ impl State {
                 clock,
                 startup_done: Arc::new(AtomicBool::new(false)),
                 should_stop: false,
+                kiosk_exit_code: None,
                 gesture_state: None,
 
                 kiosk_child: None,