umount(/proc) after CLONE_NEWNS in 2.6.25?

umount(/proc) after CLONE_NEWNS in 2.6.25?

[Michael Tokarev]

Since 2.6.25, it isn't possible anymore to umount
/proc after clone(CLONE_NEWNS).  It worked fine
before.  E.g:

# strace newnamespace umount -n /proc

2.6.24:

[pid 30121] clone(child_stack=0xff8c21f4, flags=CLONE_NEWNS|SIGCHLD) = 30122
[pid 30122] execve("/usr/sbin/umount", ["umount", "-n", "/proc"])
...
[pid 30122] oldumount("/proc")          = 0


2.6.25:

[pid 6308] clone(child_stack=0xbfc9de94, flags=CLONE_NEWNS|SIGCHLD) = 6309
[pid  6309] execve("/usr/sbin/umount", ["umount", "-n", "/proc"])
...
[pid  6309] oldumount("/proc")          = -1 EBUSY (Device or resource busy)
umount: /proc: device is busy

Yes, various NAMESPACEs are enabled (i'm trying to experiment with
those).

Is it intentional?

Thanks.

/mjt

Re: umount(/proc) after CLONE_NEWNS in 2.6.25?

[Al Viro]

On Mon, Apr 28, 2008 at 09:30:13PM +0400, Michael Tokarev wrote:
> [pid 6308] clone(child_stack=0xbfc9de94, flags=CLONE_NEWNS|SIGCHLD) = 6309
> [pid  6309] execve("/usr/sbin/umount", ["umount", "-n", "/proc"])
> ...
> [pid  6309] oldumount("/proc")          = -1 EBUSY (Device or resource busy)
> umount: /proc: device is busy
>
> Yes, various NAMESPACEs are enabled (i'm trying to experiment with
> those).
>
> Is it intentional?

It's very odd.  Could you bisect that down to offending changeset or
at least narrow the things down to -rc<something>?  I'm going down right
now, so won't be able to look into that until tonight...

Re: umount(/proc) after CLONE_NEWNS in 2.6.25?

[Al Viro]

On Mon, Apr 28, 2008 at 07:27:59PM +0100, Al Viro wrote:
> > [pid 6308] clone(child_stack=0xbfc9de94, flags=CLONE_NEWNS|SIGCHLD) = 6309
> > [pid  6309] execve("/usr/sbin/umount", ["umount", "-n", "/proc"])
> > ...
> > [pid  6309] oldumount("/proc")          = -1 EBUSY (Device or resource busy)
> > umount: /proc: device is busy
> >
> > Yes, various NAMESPACEs are enabled (i'm trying to experiment with
> > those).
> >
> > Is it intentional?
> 
> It's very odd.  Could you bisect that down to offending changeset or
> at least narrow the things down to -rc<something>?  I'm going down right
> now, so won't be able to look into that until tonight...

Check the version of umount(8) you've got.  And see if that strace happens
to have open of /proc/mounts, without matching close by the time it calls
umount().  util-linux-ng 2.13.1 is _that_ dumb...

Please, make sure that you are using the same userland for testing - with
aforementioned version of umount(8) behaviour is triggered with .24 as well
as with .25 and there's nothing kernel could do about that...

Re: umount(/proc) after CLONE_NEWNS in 2.6.25?

[Al Viro]

On Tue, Apr 29, 2008 at 02:47:18AM +0100, Al Viro wrote:
> On Mon, Apr 28, 2008 at 07:27:59PM +0100, Al Viro wrote:
> > > [pid 6308] clone(child_stack=0xbfc9de94, flags=CLONE_NEWNS|SIGCHLD) = 6309
> > > [pid  6309] execve("/usr/sbin/umount", ["umount", "-n", "/proc"])
> > > ...
> > > [pid  6309] oldumount("/proc")          = -1 EBUSY (Device or resource busy)
> > > umount: /proc: device is busy
> > >
> > > Yes, various NAMESPACEs are enabled (i'm trying to experiment with
> > > those).
> > >
> > > Is it intentional?
> > 
> > It's very odd.  Could you bisect that down to offending changeset or
> > at least narrow the things down to -rc<something>?  I'm going down right
> > now, so won't be able to look into that until tonight...
> 
> Check the version of umount(8) you've got.  And see if that strace happens
> to have open of /proc/mounts, without matching close by the time it calls
> umount().  util-linux-ng 2.13.1 is _that_ dumb...

Umm...  Looks like something more odd is going on...

Re: umount(/proc) after CLONE_NEWNS in 2.6.25?

[Al Viro]

On Tue, Apr 29, 2008 at 03:06:38AM +0100, Al Viro wrote:

> > Check the version of umount(8) you've got.  And see if that strace happens
> > to have open of /proc/mounts, without matching close by the time it calls
> > umount().  util-linux-ng 2.13.1 is _that_ dumb...
> 
> Umm...  Looks like something more odd is going on...

... or something much simpler:

; cat /proc/mounts |grep proc
proc /proc proc rw,nosuid,nodev,noexec 0 0
usbfs /proc/bus/usb usbfs rw 0 0
;

and unmounting the stuff mounted under /proc before doing umount /proc
works as expected.  Amazing how well the first cup of coffee helps...

Re: umount(/proc) after CLONE_NEWNS in 2.6.25?

[Michael Tokarev]

Al Viro wrote:
> On Tue, Apr 29, 2008 at 03:06:38AM +0100, Al Viro wrote:
> 
>>> Check the version of umount(8) you've got.  And see if that strace happens
>>> to have open of /proc/mounts, without matching close by the time it calls
>>> umount().  util-linux-ng 2.13.1 is _that_ dumb...
>> Umm...  Looks like something more odd is going on...
> 
> ... or something much simpler:
> 
> ; cat /proc/mounts |grep proc
> proc /proc proc rw,nosuid,nodev,noexec 0 0
> usbfs /proc/bus/usb usbfs rw 0 0
> ;
> 
> and unmounting the stuff mounted under /proc before doing umount /proc
> works as expected.  Amazing how well the first cup of coffee helps...

Hey.  Don't paint ME *that* dumb... ;)

It was indeed the case here.  I never tried those things on my
home machine before, and decided to experiment, but I forgot
that it has /proc/bus/usb mounted.  And for comparison I used
another machine which does NOT have that filesystem mounted.
How.. dum^Wodd... ;)

Please excuse me for the noise.

/mj