Merge pull request #858 from Dstack-TEE/codex/fix-vmm-svstop-child-reaping

fix(vmm): reap VM launcher children on SvStop

Author: kvinwang

dstack/supervisor/src/process.rs

@@ -83,6 +83,17 @@ impl ProcessStateRT {
     pub(crate) fn is_started(&self) -> bool {
         self.started
     }
+
+    /// An explicit stop may arrive just after the process exited on its own
+    /// (VM launchers exit cleanly after reaping their children). Report such
+    /// a clean exit as the intended stop, but keep non-zero exit codes and
+    /// errors visible for diagnostics. `stopped_at` recorded by the wait task
+    /// is left untouched.
+    fn normalize_clean_exit(&mut self) {
+        if matches!(self.status, ProcessStatus::Exited(0)) {
+            self.status = ProcessStatus::Stopped;
+        }
+    }
 }
 
 impl ProcessStateRT {
@@ -289,8 +300,13 @@ impl Process {
             if is_running {
                 bail!("Missing kill tx for process");
             }
+            state.normalize_clean_exit();
             return Ok(());
         };
+        if !is_running {
+            state.normalize_clean_exit();
+            return Ok(());
+        }
         match stop_tx.send(()) {
             Ok(()) => Ok(()),
             Err(()) => match is_running {

dstack/vmm/src/app.rs

@@ -468,7 +468,7 @@ impl App {
         Ok(())
     }
 
-    async fn stop_vm_process(&self, id: &str) -> Result<()> {
+    pub(crate) async fn stop_vm_process(&self, id: &str) -> Result<()> {
         let Some(info) = self.supervisor.info(id).await? else {
             return Ok(());
         };

dstack/vmm/src/main_service.rs

@@ -841,8 +841,16 @@ impl VmmRpc for RpcHandler {
     }
 
     async fn sv_stop(self, request: Id) -> Result<()> {
-        self.app.supervisor.stop(&request.id).await?;
-        Ok(())
+        // VM launcher processes own QEMU and swtpm children. Route them through
+        // the VM-aware stop path so the launcher can reap those children; the
+        // same helper preserves generic Supervisor stop semantics for every
+        // other process type.
+        self.app
+            .supervisor
+            .info(&request.id)
+            .await?
+            .context("Supervisor process not found")?;
+        self.app.stop_vm_process(&request.id).await
     }
 
     async fn sv_remove(self, request: Id) -> Result<()> {