On Fri, May 03 2019, J. Bruce Fields wrote:
> On Thu, May 02, 2019 at 12:02:33PM +1000, NeilBrown wrote:
>> On Tue, Dec 06 2016, J. Bruce Fields wrote:
>>
>> > On Tue, Dec 06, 2016 at 02:18:31PM +0100, Andreas Gruenbacher wrote:
>> >> On Tue, Dec 6, 2016 at 11:08 AM, Miklos Szeredi <miklos@xxxxxxxxxx> wrote:
>> >> > On Tue, Dec 6, 2016 at 12:24 AM, Andreas Grünbacher
>> >> > <andreas.gruenbacher@xxxxxxxxx> wrote:
>> >> >> 2016-12-06 0:19 GMT+01:00 Andreas Grünbacher <andreas.gruenbacher@xxxxxxxxx>:
>> >> >
>> >> >>> It's not hard to come up with a heuristic that determines if a
>> >> >>> system.nfs4_acl value is equivalent to a file mode, and to ignore the
>> >> >>> attribute in that case. (The file mode is transmitted in its own
>> >> >>> attribute already, so actually converting .) That way, overlayfs could
>> >> >>> still fail copying up files that have an actual ACL. It's still an
>> >> >>> ugly hack ...
>> >> >>
>> >> >> Actually, that kind of heuristic would make sense in the NFS client
>> >> >> which could then hide the "system.nfs4_acl" attribute.
>> >> >
>> >> > Even simpler would be if knfsd didn't send the attribute if not
>> >> > necessary. Looks like there's code actively creating the nfs4_acl on
>> >> > the wire even if the filesystem had none:
>> >> >
>> >> > pacl = get_acl(inode, ACL_TYPE_ACCESS);
>> >> > if (!pacl)
>> >> > pacl = posix_acl_from_mode(inode->i_mode, GFP_KERNEL);
>> >> >
>> >> > What's the point?
>> >>
>> >> That's how the protocol is specified.
>> >
>> > Yep, even if we could make that change to nfsd it wouldn't help the
>> > client with the large number of other servers that are out there
>> > (including older knfsd's).
>> >
>> > --b.
>> >
>> >> (I'm not saying that that's very helpful.)
>> >>
>> >> Andreas
>>
>> Hi everyone.....
>> I have a customer facing this problem, and so stumbled onto the email
>> thread.
>> Unfortunately it didn't resolve anything. Maybe I can help kick things
>> along???
>>
>> The core problem here is that NFSv4 and ext4 use different and largely
>> incompatible ACL implementations. There is no way to accurately
>> translate from one to the other in general (common specific examples
>> can be converted).
>>
>> This means that either:
>> 1/ overlayfs cannot use ext4 for upper and NFS for lower (or vice
>> versa) or
>> 2/ overlayfs need to accept that sometimes it cannot copy ACLs, and
>> that is OK.
>>
>> Silently not copying the ACLs is probably not a good idea as it might
>> result in inappropriate permissions being given away. So if the
>> sysadmin wants this (and some clearly do), they need a way to
>> explicitly say "I accept the risk".
>
> So, I feel like silently copying ACLs up *also* carries a risk, if that
> means switching from server-enforcement to client-enforcement of those
> permissions.
Interesting perspective .... though doesn't NFSv4 explicitly allow
client-side ACL enforcement in the case of delegations?
Not sure how relevant that is....
It seems to me we have two options:
1/ declare the NFSv4 doesn't work as a lower layer for overlayfs and
recommend people use NFSv3, or
2/ Modify overlayfs to work with NFSv4 by ignoring nfsv4 ACLs either
2a/ always - and ignore all other acls and probably all system. xattrs,
or
2b/ based on a mount option that might be
2bi/ general "noacl" or might be
2bii/ explicit "noxattr=system.nfs4acl"
I think that continuing to discuss the miniature of the options isn't
going to help. No solution is perfect - we just need to clearly
document the implications of whatever we come up with.
I lean towards 2a, but I be happy with with any '2' and '1' won't kill
me.
Do we have a vote? Or does someone make an executive decision??
NeilBrown
Attachment:
signature.asc
Description: PGP signature
