Environment: ApisCP v3.2.49, AlmaLinux 8.10, kernel 4.18.0-553.186.1 (cgroup v1), dnf-automatic.
What happened: dnf-automatic upgraded ImageMagick-libs 6.9.13.58 → 6.9.13.60 (EPEL) at 08:58 UTC.
The post-transaction trigger Trigger\ImageMagickLibs ran misc_release_fsghost() on libMagickCore. For each
process mapping the old inode, it freezes the site’s freezer cgroup, then calls Process::killWait($pid, SIGKILL)
(30 s wait); the thaw is deferred until all processes are handled. A task in a frozen cgroup v1 does not exit
on SIGKILL until thawed, so each wait ran to its timeout: 33 php-fpm processes across 13 sites ≈ 16.5 min
(08:58:52 → 09:15:22). Every PHP site on the box was frozen for that time. Frozen masters also missed the
php-fpm units’ 30 s systemd watchdog, giving SIGABRT/SIGKILL, “Processes still around after SIGKILL”,
and 8 core dumps.
Contrast: on AlmaLinux 9 (cgroup v2) the same trigger ran the same day and every kill completed at once.
Related: t/766 (release-fsghost runtime), t/957 (release-fsghost on libcurl).
Suggestion: on cgroup v1, kill before freezing, or thaw before killWait(), or kill the whole cgroup
(cgroup.procs), so the wait is not spent against a frozen task.
Workaround we use meanwhile: exclude = ImageMagick* under [base] in /etc/dnf/automatic.conf on EL8,
with ImageMagick updated by hand off-hours.
Thanks,
Bogdan Susala (with Claude)