[Impact]
* df displays bind mounts instead of "real" mounts if the bind mount is mounted to a shorter directory.
* justification - When trusty moved to using /proc/mounts this changed behavior from precise. Additionally it doesn't make sense that a bind mount should show up in df over a real root mount.
* Explanation - These patches change behavior of df to rely on /proc/self/mountinfo which has more complete info than /proc/mounts. Such as what directory of the filesystem was used as the source of the mount. Additionally given this new information there is a patch on df itself to make use of this new information.
[Test Case]
* $ mount
<snip>
192.168.1.2:/raid on /raid type nfs
/dev/sdc5 on /data type ext4 (rw)
<snip>
/data/a on /a type none (rw,bind)
/raid/temp on /b type none (rw,bind)
$df
Filesystem 1K-blocks Used Available Use% Mounted on
<snip>
/data/a 449830616 229975284 196982196 54% /a
/raid/temp 7752072192 5581343744 1780023296 76% /b
</snip>
I'd expect to see the real mount prioritized over the bind mount. Like so.
$df
Filesystem 1K-blocks Used Available Use% Mounted on
<snip>
/dev/sdc5 449830616 229975284 196982196 54% /data
192.168.1.2:/raid 7752072192 5581438976 1779929088 76% /raid
<snip>
[Regression Potential]
* Patch is upstreamed.
* df command now relies on /proc/self/mountinfo
[Other Info]
* The behavior of df, mount and similar number of other commands has changed going forward. Previously these commands all processed /etc/mtab which was maintained by the mount command. Going forward they still process /etc/mtab, but this is simply a symlink to /proc/mounts now which is maintained by the kernel and slightly more accurate. Unlike the mount command, the kernel makes no distinction between bind mounts and normal mounts. This is evident by the fact that you can mount a device, bind mount from that device, and then unmount the original mount. The default behavior of df in this case is to simply pick the mounted directory for a device that is the shortest since it has no other information to go on from /proc/mounts. Moving to using /proc/self/mountinfo resolves this issue, and is what upstream is doing moving forward.
* In order to solve this issue, I have written patches and got them integrated upstream.
gnulib commit: http://git.savannah.gnu.org/gitweb/?p=gnulib.git;a=commit;h=c6148bca89e9465fd6ba3a10d273ec4cb58c2dbe and
coreutils commit: http://git.savannah.gnu.org/gitweb/?p=coreutils.git;a=commit;h=3babaf83875ceac896c8dd3a64248e955dfecef9
have been authored by me.
While attempting to push these patches into Ubuntu, it became apparent that our version of coreutils does not contain the requisite patches to correctly support /proc/self/mountinfo. /proc/self/mountinfo gives more complete information than /proc/self/mounts which /etc/mtab now points to. Using /proc/self/mountinfo is the upstream way of doing things, and it allows us to resolve this specific bug.
Patches required in order to support /proc/self/mountinfo are.
Origin: upstream, gnulib - http://git.savannah.gnu.org/gitweb/?p=gnulib.git
commit: http://git.savannah.gnu.org/gitweb/?p=gnulib.git;a=commit;h=3ea43e02541ece750ffc6cd1dfe34195421b4ef3
commit: http://git.savannah.gnu.org/gitweb/?p=gnulib.git;a=commit;h=2768ceb7994506e2cfba88be3b6bd13ef5440a90
commit: http://git.savannah.gnu.org/gitweb/?p=gnulib.git;a=commit;h=de1cbdd48244c66c51a3e2bc1594ac3ad32ce038
commit: http://git.savannah.gnu.org/gitweb/?p=gnulib.git;a=commit;h=3fb6e360363744462ce15c381f0b116c6fc4ce82
Origin: upstream, coreutils - http://git.savannah.gnu.org/gitweb/?p=coreutils.git
commit: http://git.savannah.gnu.org/gitweb/?p=coreutils.git;a=commit;h=1b1c40e1d6f8cf30b6c7c9d31bbddbc3d5cc72e6
Original bug.
________________________________________________________________________
Depending on mount path length df sometimes prioritizes showing bind mounts over real mounts
for example.
$ mount
<snip>
192.168.1.2:/raid on /raid type nfs (rw,nosuid,nodev,noexec,vers=4,addr=192.168.1.2,clientaddr=192.168.1.3)
/dev/sdc5 on /data type ext4 (rw)
<snip>
/data/a on /a type none (rw,bind)
/raid/temp on /b type none (rw,bind)
$df
Filesystem 1K-blocks Used Available Use% Mounted on
<snip>
/data/a 449830616 229975284 196982196 54% /a
/raid/temp 7752072192 5581343744 1780023296 76% /b
</snip>
I'd expect to see the real mount prioritized over the bind mount. Like so.
$df
Filesystem 1K-blocks Used Available Use% Mounted on
<snip>
/dev/sdc5 449830616 229975284 196982196 54% /data
192.168.1.2:/raid 7752072192 5581438976 1779929088 76% /raid
<snip>
Here is the fix that I will propose to upstream. Please do not do this upload until we have buy in from upstream.