race condition with `systemd` notification protocol

[aanderse]

when setting notify to systemd there can be a problem if a daemon starts too early - the sd_notify socket isn't bound until after the fork

[troglobit]

Could you give an example to help me understand? Because often the call to sd_notify(), or pidfile creation, is actually malplaced in the daemon itself.

[aanderse]

imagine a daemon tries to access NOTIFY_SOCKET before it does anything else... since finit forks the process first it would be possible that the socket isn't bound yet...

i have a user running systemd-udev in their initramfs and they weren't getting a ready condition for the daemon... we ended up creating a script to test this out...

    #!/bin/sh
    if [ -n "''${NOTIFY_SOCKET:-}" ]; then
      found=
      tries=0
      while [ -z "$found" ] && [ "$tries" -lt 100000 ]; do
        while read -r _ _ _ _ _ _ _ name; do
          if [ "$name" = "$NOTIFY_SOCKET" ]; then
            found=1
            break
          fi
        done < /proc/net/unix
        tries=$((tries + 1))
      done
    fi
    exec "$@"

the command was that script followed by calling udev and this fixed the issue for them

i ended up creating a vm test to reproduce and was able to... in some cases i got up to 1158 tries before the name appeared

note that we have only been able to reproduce this in an initramfs so far

[troglobit]

Aha, got it! I'll see what my friend can make of it while we await review on #500 (stalling my pipeline atm :)