Re: linux-3.14 nfsd regression

["J. Bruce Fields" <bfields@fieldses.org>]

On Thu, Apr 03, 2014 at 02:55:04PM -0400, Jeff Layton wrote:
> On Thu, 03 Apr 2014 13:51:06 -0400
> Mark Lord <mlord@pobox.com> wrote:
> 
> > On 14-04-03 01:16 PM, J. Bruce Fields wrote:
> > > On Thu, Apr 03, 2014 at 12:33:55PM -0400, Mark Lord wrote:
> > >> This commit from linux-3.14 breaks our NFS-root clients here:
> > >>
> > >> https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=6e14b46b91fee8a049b0940333ce13a820beaaa5
> > >>
> > >>
> > >> - *p++ = htonl((u32) stat->mode);
> > >> + *p++ = htonl((u32) (stat->mode & S_IALLUGO));
> > >>
> > >>
> > >> Reverting the one-liner above (on the server) fixes it for us,
> > >> as does reverting back to linux-3.13.8 on the server.
> > >>
> > >> The NFS-root clients are on PowerPC (big-endian) architecture,
> > >> running linux-3.12.16. The NFS server is on an Intel PC running linux-3.14.
> > >>
> > >> ACL is completely disabled on server and client,
> > >> and we're using NFSv2/v3.  No support for v4.
> > >>
> > >> I instrumented the function to see what other bits were being cleared
> > >> by the (stat->mode & S_IALLUGO) masking.  The results are attached.
> > > 
> > > Hm, it sounds like a bug in the client if it's depending on those high
> > > bits.
> > 
> > But only for mounting / starting up from the nfsroot, it seems.
> > I wonder if there's an unusual code path for that in there?
> > The regular stuff looks mostly fine:
> > 
> >         p = xdr_decode_ftype3(p, &fmode);
> >         fattr->mode = (be32_to_cpup(p++) & ~S_IFMT) | fmode;
> > 
> > Except perhaps that second line ought to use the same mask
> > as the server side is using, just in case there are some other
> > stray high (higher than S_IFMT) bits in there now/someday.
> > 
> > > The original behavior was in practice harmless and changing it broke
> > > something, so I think we should definitely just revert this patch.
> > 
> > Yup.  Who?
> > 
> > > But the client may need fixing too.
> > 
> > Probably a good thing in the longer term, for better compatibility
> > with non-Linux servers.  But we'll still have to keep the revert
> > on the server (nfsd) code for backward compatibility, I think.
> > 
> > Cheers
> > 
> 
> It would be good to understand where this is broken in the client.
> 
> It's incorrect for the client to interpret those bits, as I think that
> there's no guarantee that other OS's implement the type bits in the
> same way that Linux does. So if you end up mounting a different OS,
> it's possible that the client will get that wrong...

It turns out these bits actually are defined in rfc 1094, so this is
just an odd NFSv2-specific wart, and the nfsd change was just flat-out
wrong.

--b.