Fall back to signalfd+SIGCHLD when pidfd_open returns EPERM

[paddor]

Summary

• When pidfd_open fails with EPERM (e.g. inside snap confinement), the epoll and uring selectors now fall back to a signalfd watching SIGCHLD instead of raising
• Since SIGCHLD fires for any child, the fallback loops until the specific target pid exits
• Guarded by #ifdef HAVE_SYS_SIGNALFD_H; on systems without signalfd, behavior is unchanged (raises on EPERM as before)

Context

This is not urgent — snapd 2.75+ already allows pidfd_open in its default seccomp profile, so the confinement issue is resolved for most users. However, older snapd versions (and potentially other seccomp-confined environments) still block the syscall. Feel free to close this if you'd rather not carry the extra code path.

Test plan

• Existing test suite passes (201 passed, 2 skipped)
• New tests use seccomp-BPF (via Fiddle) in a forked child to block pidfd_open with EPERM, then exercise #process_wait through the signalfd path
• Tests cover: already-exited process, still-running process, two sequential waits

[ioquatix]

For the reasons you stated, I'm slightly against merging this. But if we could make the burden a little lighter the calculus might change.

Do you think we can extract the logic into a shared code file, so that we don't duplicate this between epoll and uring backends?

I know I deliberately avoided sharing a lot of code in the backends but it's mostly the mechanics of the scheduler, and I was aiming to have independent implementations where the logic was significantly tied to the OS interface. But for this, I think it's the same, right? So we could probably share the implementation and just have if (errno == EPERM) fallback_entry_point(...);

WDYT?

[paddor]

Good idea — done in b2098b9. I extracted the signalfd lifecycle into process_wait_signalfd.c (included like pidfd.c), which provides three helpers:

• process_wait_signalfd_open — blocks SIGCHLD, creates signalfd, handles early exit
• process_wait_signalfd_check — drains signalfd, calls waitpid for the target pid
• process_wait_signalfd_close — closes signalfd, restores signal mask

Each backend keeps only its selector-specific registration loop (epoll_ctl vs io_uring_prep_poll_add). The EPERM fallback in each process_wait is now ~15 lines.