Bug#324900: nscd: umount /var fails (unclean shutdowns)

[Gabor Gombas <gombasg@sztaki.hu>]

On Wed, Aug 31, 2005 at 01:36:24PM +0900, GOTO Masanori wrote:

> > root      1119     1  0 12:47 ?        00:00:00 /bin/sh /etc/init.d/rc 6
[...]
> > rc        1119 root  mem       REG    8,9  217016   228931 /var/db/nscd/passwd
> 
> It's very weird behavior.  Please disable nscd when your boot up time,
> and then run /etc/init.d/nscd.  You can see which processes have such
> nscd file descriptor (fd), and you can clear process inheritance with
> pstree easily.  If nothing has nscd fd, we can clear rc behaves oddly.

Well, if /bin/sh is bash, then it is not weird at all, it is the same
"bash vs. NSS" problem that came up several times in the past (last time
quite recently on debian-devel). Previously it only happened with NSS
modules that link to libraries under /usr, now it also affects nscd.

What I _suspect_ is happening here:

- by calling /etc/init.d/rc, bash is executed
- bash unconditionally does some NSS calls during startup (getpwuid
  etc.); this in turn
- loads all NSS modules that serve passwd maps -> if a module uses
  libraries from /usr, now you have a live memory mapping under /usr so
  you cannot unmount it during shutdown
- bash (libc) connects to nscd
- nscd sends a file descriptor of /var/db/nscd/passwd to bash, and bash
  does an mmap(2) on the received fd -> now you have a live memory
  mapping under /var thus you cannot unmount it during shutdown
- /etc/init.d/rc eventually kills nscd but that does not help, since the
  bash process executing /etc/init.d/rc still has the cache file mapped
  (deleting the cache file also doesn't help since unlink(2) only
  operates on the directory and does not invalidate the memory mapping)

Options you have to avoid the problem:

1. Stop using bash as /bin/sh
2. Stop using nscd
3. Convert /etc/init.d/rc to be an ELF executable instead of a shell
   script
4. Let /var (and /usr if you are running more exotic NSS modules) be the
   same filesystem as /
5. Redesign the init system so unmounting of local file systems is done
   _after_ /etc/init/rc has finished (and the program that does the
   unmounting must not be a shell script)

Gabor

-- 
     ---------------------------------------------------------
     MTA SZTAKI Computer and Automation Research Institute
                Hungarian Academy of Sciences
     ---------------------------------------------------------