fallocate() on XFS clobbers S*ID-bits

fallocate() on XFS clobbers S*ID-bits

[Andreas Wiese]

Hello.

I (with support from Cc'ed Ciaran) just noticed some odd behaviour with
fallocate() on XFS.  After open()ing some file and setting it S*ID via
fchmod(), S*ID bits vanish after calling fallocate() — as supposed to
for non-root users, but it also happens for root.

Is this intended behaviour or did we spot a bug here?

At least on ext2 it works as expected, thus I guess it's the latter one.

I'm running v2.6.35.7 vanilla-kernel, but diffing fs/xfs to master
doesn't seem to address this issue.

HAND & LG -- aw
-- 
This signature is intentionally left blank.

Re: fallocate() on XFS clobbers S*ID-bits

[Dave Chinner]

[ added xfs list on cc ]

On Thu, Oct 07, 2010 at 08:34:19PM +0200, Andreas Wiese wrote:
> Hello.
> 
> I (with support from Cc'ed Ciaran) just noticed some odd behaviour with
> fallocate() on XFS.  After open()ing some file and setting it S*ID via
> fchmod(), S*ID bits vanish after calling fallocate() — as supposed to
> for non-root users, but it also happens for root.
> 
> Is this intended behaviour or did we spot a bug here?
> 
> At least on ext2 it works as expected, thus I guess it's the latter one.

ext2 does not support the fallocate syscall. How does ext4 behave?

> I'm running v2.6.35.7 vanilla-kernel, but diffing fs/xfs to master
> doesn't seem to address this issue.

XFS has always cleared SUID/SGID bits when doing preallocation (via
XFS_IOC_RESVSP). Given that that XFS ioctl formed the model for
fallocate, I'd argue that the XFS behaviour is the on that should be
followed.

It'd be a bit silly for two different preallocation interfaces to
have different behaviour w.r.t. SUID bits, but it's not clear to me
which behaviour is correct. I'm happy to defer to whoever can say
what the behaviour is supposed to be here...

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

Re: fallocate() on XFS clobbers S*ID-bits

[Andreas Wiese]

Dave Chinner wrote:
> [ added xfs list on cc ]
> 

Thanks. ;)

> On Thu, Oct 07, 2010 at 08:34:19PM +0200, Andreas Wiese wrote:
> > Hello.
> > 
> > I (with support from Cc'ed Ciaran) just noticed some odd behaviour with
> > fallocate() on XFS.  After open()ing some file and setting it S*ID via
> > fchmod(), S*ID bits vanish after calling fallocate() — as supposed to
> > for non-root users, but it also happens for root.
> > 
> > Is this intended behaviour or did we spot a bug here?
> > 
> > At least on ext2 it works as expected, thus I guess it's the latter one.
> 
> ext2 does not support the fallocate syscall. How does ext4 behave?
> 

Oh, crap.  I first tested on tmpfs (which doesn't support fallocate,
too) and had a » && errno != ENOTSUP« to avoid bailing out but forgot to
warn. :/

But I did now and ext4 does _not_ touch the permissions any further.

> > I'm running v2.6.35.7 vanilla-kernel, but diffing fs/xfs to master
> > doesn't seem to address this issue.
> 
> XFS has always cleared SUID/SGID bits when doing preallocation (via
> XFS_IOC_RESVSP). Given that that XFS ioctl formed the model for
> fallocate, I'd argue that the XFS behaviour is the on that should be
> followed.
> 

If that's consense, we at least found a bug in ext4…

> It'd be a bit silly for two different preallocation interfaces to
> have different behaviour w.r.t. SUID bits, but it's not clear to me
> which behaviour is correct. I'm happy to defer to whoever can say
> what the behaviour is supposed to be here...
> 

I would appreciate that, too.

And while on it… just tested BTRFS, which behaves like ext4 (not
altering permissions).  CIFS and OCFS are not testable here due to lack
of according servers.

[snip…]

HAND & LG -- aw
np: Johnossi (Johnossi) — 08. Happiness a La Mode