mirror of
https://gitea.com/gitea/act_runner.git
synced 2026-06-10 11:54:27 +02:00
fix(cleanup): kill Windows step process tree on cancel to avoid hang (#1011)
## Problem Cancelling a job on a Windows host runner can leave the spawned process tree running and hang the runner. When a step launches a shell that starts a child which in turn spawns further GUI/background processes, cancelling the job kills only the direct child (the default `exec.CommandContext` behaviour). The surviving descendants inherited the step's stdout/stderr pipe, so the read end never hit EOF and `cmd.Wait()` blocked forever. Because the step executor never returned: - the orphaned processes kept running (the cancelled work was not actually stopped), and - end-of-job cleanup (`Remove` → `terminateRunningProcesses`) was never reached, so the runner appeared to go offline / stop picking up jobs. `CREATE_NEW_PROCESS_GROUP` does not help here — it affects Ctrl-C signal delivery, not handle inheritance or tree termination. ## Fix - Assign each Windows step process to a **Job Object** immediately after `cmd.Start()`. Descendants created afterwards are automatically part of the job. - Override `cmd.Cancel` to `TerminateJobObject`, so cancellation kills the **entire descendant tree** atomically. This also closes the inherited pipe handles, so `cmd.Wait()` can return. - Set `cmd.WaitDelay` (10s) as a safety net: once the process has exited, Wait force-closes the pipes and returns rather than blocking forever — covering the case where the job-object setup fails (e.g. nested-job restrictions), in which we fall back to the previous single-process kill. - The Job Object is created **without** `JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE`, so closing the handle on normal completion does not kill legitimate background processes; the tree is only torn down on explicit cancel. Implemented behind `runtime.GOOS == "windows"` with a Windows-only `processKiller` (Job Object) and no-op stubs elsewhere, so non-Windows behaviour (default cancellation + `Setpgid`) is unchanged. ## Changes - `act/container/process_windows.go` — Job Object `processKiller` (create / assign / terminate). - `act/container/process_other.go` — no-op stubs (`//go:build !windows`). - `act/container/host_environment.go` — wire `cmd.Cancel` (tree kill) and `cmd.WaitDelay` into `exec()`. - `go.mod` / `go.sum` — promote `golang.org/x/sys` to a direct dependency. ## Testing I fully tested it already ## Notes Follow-up to the Windows leftover-process reaping in #996: that sweep now actually runs on cancellation because the step no longer hangs before reaching it. Reviewed-on: https://gitea.com/gitea/runner/pulls/1011 Reviewed-by: techknowlogick <9+techknowlogick@noreply.gitea.com>
This commit is contained in:
19
act/container/process_other.go
Normal file
19
act/container/process_other.go
Normal file
@@ -0,0 +1,19 @@
|
||||
// Copyright 2026 The Gitea Authors. All rights reserved.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
//go:build !windows
|
||||
|
||||
package container
|
||||
|
||||
import "os"
|
||||
|
||||
// processKiller is a no-op on non-Windows platforms. The Job Object based
|
||||
// tree-kill is only wired in on Windows (see exec()); elsewhere the default
|
||||
// exec.CommandContext cancellation and Setpgid handling apply.
|
||||
type processKiller struct{}
|
||||
|
||||
func newProcessKiller(_ *os.Process) (*processKiller, error) { return &processKiller{}, nil }
|
||||
|
||||
func (k *processKiller) Kill() error { return nil }
|
||||
|
||||
func (k *processKiller) Close() error { return nil }
|
||||
Reference in New Issue
Block a user