[lxc-users] Cannot unlink/rename over host-level file that has been bind-mounted out of a container

Mon Dec 15 19:29:17 UTC 2014

On Mon, Dec 15, 2014 at 2:19 PM, Serge Hallyn <serge.hallyn at ubuntu.com>
wrote:
>
> Since you are using / as the container rootfs, /empty is the same
> inode in the container and on the host.

No, it isn't. Within the container, /empty is the inode of the host's
/other_file, because /other_file is bind-mounted over /empty:

lxc.mount.entry=/other_file empty none rw,bind 0 0

host# ls -li /empty /other_file
57876 -rw-r--r-- 1 root root 0 Dec 15 19:26 /empty
58108 -rw-r--r-- 1 root root 0 Dec 15 19:26 /other_file
host# lxc-execute -n test -f test.config -- ls -li /empty /other_file
58108 -rw-r--r-- 1 root root 0 Dec 15 19:26 /empty
58108 -rw-r--r-- 1 root root 0 Dec 15 19:26 /other_file

I would expect the host's /other_file to be busy, but not the host's /empty.

Barry

> It is over-mounted.  So it
> is busy.
>
> Quoting Barry Jaspan (barry.jaspan at acquia.com):
> > I have reduced the problem I described in my previous message to a much
> > smaller test case. Here is an LXC config file that bind-mounts a single
> > file, /other_file, from the host's root filesystem on top of another
> file,
> > /empty, also from the host's root filesystem:
> >
> > lxc.console = none
> > lxc.rootfs = /
> > lxc.mount.entry=/other_file empty none rw,bind 0 0
> >
> > While the container is running, the host's /empty cannot be unlinked:
> >
> > host# rm /empty /other_file
> > host# touch /empty /other_file
> > host# lxc-execute -n test -f test.config -- sleep 30 &
> > [1] 2419
> > host# strace -o rm.out rm /empty
> > rm: cannot remove `/empty': Device or resource busy
> > host# grep EBUSY rm.out
> > unlinkat(AT_FDCWD, "/empty", 0)         = -1 EBUSY (Device or resource
> busy)
> > host# fg
> > lxc-execute -n test -f test.config -- sleep 30
> > ^C
> > host# rm /empty
> > host#
> >
> > Can anyone explain why the host's /empty cannot be unlinked even though
> it
> > is bind-mounted out of the running container? Why is it "busy"?
> >
> > Thanks,
> > Barry
> >
> >
> >
> >
> > On Mon, Dec 8, 2014 at 5:53 PM, Barry Jaspan <barry.jaspan at acquia.com>
> > wrote:
> > >
> > > I'm using LXC on Ubuntu 12.04 (Precise) on EC2. I am creating a
> container
> > > that uses the host's root filesystem. In the lxc config, I am
> bind-mounting
> > > a different passwd file on top of /etc/passwd:
> > >
> > > lxc.mount.entry=/container/passwd etc/passwd none rw,bind 0 0
> > >
> > > As expected, /etc/passwd on the host and /etc/passwd in the container
> are
> > > different inodes:
> > >
> > > host# ls -i /etc/passwd
> > > 58046 /etc/passwd
> > > host# ls -i /container/passwd
> > > 287145 /container/passwd
> > >
> > > container# ls -i /etc/passwd
> > > 287145 /etc/passwd
> > >
> > > What I did not expect is that now it is not possible to unlink
> /etc/passwd
> > > at the host level (or inside the container, although I don't care about
> > > that):
> > >
> > > host# rm /etc/passwd
> > > rm: cannot remove `/etc/passwd': Device or resource busy
> > > host#  chsh -s /bin/false root
> > > chsh: failure while writing changes to /etc/passwd
> > >
> > > strace confirms that unlinkat() and rename() are returning EBUSY. (chsh
> > > creates a replacement for /etc/passwd and then rename()'s it into
> place.)
> > >
> > > So, somehow, the fact that the host's /etc/passwd is *not* present in
> the
> > > container marks its inode as busy. I do not think it is as simple as
> "an
> > > inode that has been bind-mounted out of a namespace is busy" because a
> > > simple bind-mount test case does not reproduce it:
> > >
> > > # echo foo > foo
> > > # touch bar
> > > # mount --bind ./foo ./bar
> > > # cat bar
> > > foo
> > > # touch new
> > > # mv new foo
> > > #
> > >
> > > This seems related to http://lwn.net/Articles/570338/. However, in
> that
> > > article, a file which is bind-mounted into another namespace causes
> EBUSY
> > > from unlink/rename(). In my case, a file which is *not* bind-mounted
> into
> > > another namespace is causing EBUSY.
> > >
> > > Can someone explain what is going on?
> > >
> > > Thanks,
> > >
> > > Barry
> > >
> > > --
> > > Barry Jaspan
> > > Senior Architect | Acquia <http://acquia.com>
> > > barry.jaspan at acquia.com | (c) 617.905.2208 | (w) 781-313-8298
> > >
> > > Acquia Dev Cloud: You build killer websites. We do the rest.
> > > <http://www.acquia.com/dev-cloud> <http://acquia.com/dev-cloud>
> > > Acquia ranked #1 Software Vendor on the 2012 Inc 500
> > > <
> http://www.acquia.com/about-us/newsroom/press-releases/inc-magazine-unveils-31st-annual-list-america-s-fastest-growing
> >
> > >
> > >
> >
> > --
> > Barry Jaspan
> > Senior Architect | Acquia <http://acquia.com>
> > barry.jaspan at acquia.com | (c) 617.905.2208 | (w) 781-313-8298
> >
> > Acquia Dev Cloud: You build killer websites. We do the rest.
> > <http://www.acquia.com/dev-cloud> <http://acquia.com/dev-cloud>
> > Acquia ranked #1 Software Vendor on the 2012 Inc 500
> > <
> http://www.acquia.com/about-us/newsroom/press-releases/inc-magazine-unveils-31st-annual-list-america-s-fastest-growing
> >
>
> > _______________________________________________
> > lxc-users mailing list
> > lxc-users at lists.linuxcontainers.org
> > http://lists.linuxcontainers.org/listinfo/lxc-users
>
> _______________________________________________
> lxc-users mailing list
> lxc-users at lists.linuxcontainers.org
> http://lists.linuxcontainers.org/listinfo/lxc-users

-- 
Barry Jaspan
Senior Architect | Acquia <http://acquia.com>
barry.jaspan at acquia.com | (c) 617.905.2208 | (w) 781-313-8298

Acquia Dev Cloud: You build killer websites. We do the rest.
<http://www.acquia.com/dev-cloud> <http://acquia.com/dev-cloud>
Acquia ranked #1 Software Vendor on the 2012 Inc 500
<http://www.acquia.com/about-us/newsroom/press-releases/inc-magazine-unveils-31st-annual-list-america-s-fastest-growing>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.linuxcontainers.org/pipermail/lxc-users/attachments/20141215/06016c1d/attachment-0001.html>