reaper

Overview ¶

Package reaper reaps orphaned child processes when this process is PID 1.

In a container, the kernel reparents every orphan (a daemon that double-forked, a background job whose shell exited, whatever escaped a killed process group) to PID 1. Only a wait call by PID 1 removes such a process once it exits; without one it stays a zombie for the life of the container. Zombies keep their pid, still answer kill -0, [ -e /proc/PID ] and pgrep, so a script waiting for a stopped daemon to disappear never finishes, and ps shows <defunct> entries.

Go's exec.Cmd waits on exactly the child it started, so a process started with exec.Cmd must not be reaped here: cmd.Wait would fail with "no child processes" and lose the exit status. Two mechanisms keep them apart:

• StartCmd tracks the pid under the same lock the reap loop takes, so a child cannot exit and be reaped between Start and being recorded.
• Children Go holds a pidfd for (every exec.Cmd on kernels with pidfd support) are recognised through /proc/self/fdinfo and skipped even when nobody called StartCmd for them.

A child started with plain exec.Cmd on a kernel without pidfd support has neither protection and can lose its status if it exits at the moment the reap loop runs; callers that must be safe there use StartCmd.