When forking, the code would close all unused file descriptors up
to maximum number of files. In the default case this is 1024. In
the container case this is 1048576. Huge delays in startup were
seen due to this.

Even in a default 1024 ulimit case this drastically reduces the
number of file descriptors to mark FD_CLOEXEC but in the container
default case this saves another order of magnitude of work.

This patch takes inspiration from systemd[1] and walks /proc/self/fd
if it is available to determine the list of open descriptors. It
falls back to the "close all fds we don't care about up to limit"
method.

https://bugzilla.redhat.com/show_bug.cgi?id=1656519

[1] https://github.com/systemd/systemd/blob/5238e9575906297608ff802a27e2ff9effa3b338/src/basic/fd-util.c#L217