[nfsv4] Re: Comments for draft-rmacklem-nfsv4-posix-acls-02

[Rick Macklem <[e-mail]>]

On Tue, Aug 6, 2024 at 10:05 AM Pali Rohár <[e-mail]>
wrote:

> Hello Rick,
>
> As we were discussed off-list, I'm sending to this list my comments for
> your draft-rmacklem-nfsv4-posix-acls-02 document. Quoted text is
> verbatim copy from your document.
>
> Your document describe POSIX ACL support only for NFS v4.2. The most
> commonly used NFS4 version is 4.1 and it would stay for a longer time,
> lot of enterprise NFS SW is being written as NFS3 + NFSv4.1 (no v4.2
> support). So to make extension widely usable, it needs to be also for
> NFS v4.1. Is there any reason why v4.1 is not included in your document?
>
As I understand it, new attributes cannot be added to 4.1.
Also, adding support for 4.2 is not a lot of work (some would say
it's trivial):
- For the client--> replace 1 with 2 in the minorversion field of the
  compounds. That's all there is to it, since everything in 4.2 are
  optional extensions to 4.1.
  --> These optional extensions can be added if/when the implementor
      needs them.
- For the server--> accept minorversion set to 2 and make sure all
  operation #s that are in 4.2 get a reply of NSF4ERR_NOTSUPP.
  --> Again, an implementor can then add support for 4.2 optional
      operations, attributes.. whenever needed.


> > Therefore, a client MUST not set the low order 9 bits of mode via the
> > mode or mode_set_masked attributes in the same SETATTR operation as
> > one that sets the posix_access_acl and/or posix_default_acl proposed
> > in this document. This is required because [RFC8881] does not specify
> > an ordering for setting attributes in a SETATTR operation.
>
> RFC 8881 in section "6.4.1. Setting the Mode and/or ACL Attributes"
> describes the ordering of setting MODE, ACL attributes.
>
> In section "6.4.1.3. Setting Both ACL and Mode" is covered any ACL
> attribute with any MODE attribute:
> "When setting both the mode (includes use of either the mode attribute
> or the mode_set_masked attribute) and the acl or dacl attributes in the
> same operation, the attributes MUST be applied in this order: mode (or
> mode_set_masked), then ACL."
>
> I think it would be better if your extension says that the new
> POSIX_*_ACL attributes has to be handled as ACL attribute according to
> RFC 8881 6.4.1.3. section. It would simplify MODE and ACL interop in
> SETATTR operation.
>
Yea, I suppose it is more consistent.
I'll change it.

To be honest, at least for the FreeBSD and Linux clients, the
acl attribute is never changed in the same SETATTR as mode/mode_set_masked
so I am a little concerned that servers may not get this right.
(I think the FreeBSD server does get the ordering right.)


> > Therefore, to maintain compatible semantics with the POSIX draft, for
> > NFSv4 operations that create new file objects (OPEN/OPEN4_CREATE,
> > CREATE) in a directory that has a POSIX default ACL, the low order 9
> > bits of the mode SHOULD be specified by either mode or mode_set_masked
> > in the setable attributes for the operation.
>
> RFC 8881 in section 6.2.5. specifies:
> "The mode_set_masked attribute is only valid in a SETATTR operation. If
> it is used in a CREATE or OPEN operation, the server MUST return
> NFS4ERR_INVAL."
>
Ok, I never noticed that.
I'll fix this. Thanks for pointing it out.


> So when creating a new object "mode_set_masked", cannot be used. For
> inheritance you need to use "mode_umask" attribute which is defined in
> RFC 8275. This attribute already handles this POSIX inheritance issue
> and I think it is required for proper POSIX ACL implementation.
>
Hmm. I suppose the document should say something about the use of
this attribute, if it is supported. (Since it is yet another
optional 4.2 extension, I'd rather not have these attributes
depend on support for it.)
I suppose it can go as far as recommending support of it?


> > For a SETATTR, setting a zero length acl (for a file object with a
> > true form ACL of ACL_MODEL_NFS4) ... will delete the true form ACL(s)
> > from the file object and revert it to ACL_MODEL_NONE.
>
> Why is this needed? Empty NFS4 ACL is same valid NFS4 ACL. If somebody
> want to restrict access to some file then setting empty list of ACEs
> which is zero-length acl. I think that ACL with no access should still
> be in NFS4 model.
>
For the case of ACL_SCOPE_FILE_OBJECT, I believe there needs to be
a way to "delete" the NFSv4 ACL. For example, here is a snippet from
IBM's doc. on GPFS:
You can also change the type of ACL by the mmdelacl command (causing the
permissions to revert to the mode, which is in effect a POSIX ACL).
As you might have guessed, "mmdelacl" deletes the ACL.

As such, it seems that there must be a way to "delete" a NFSv4 ACL and
setting "acl" to a 0 length ACL seemed the convenient way to do it.
--> For GPFS, "delete" seems to actually be defined as switch to a
    minimal POSIX ACL.


> Comparing with Windows. On Windows (and in SMB protocol) there are two
> different operations which can be done on object: Erasing DACL (by
> setting it to NULL) and setting empty DACL (by setting empty list of
> ACEs). Erasing is really erasing of the DACL in the storage (similar
> like erasing POSIX ACL and reverting to mode_only) which has same effect
> as ALLOW-EVERYONE-FULL-ACCESS. And setting empty list has effect of the
> DENY-EVERYONE-FULL-ACCESS (as ACLs are default-DENY, there is no rule
> which can grant some access).
>
> As NFS4 does not have a way to set ACL to NULL, it has only the section
> option and matches Windows behavior of empty-ACE-list as NFS4 is also
> default-DENY.
>
Thanks for pointing this out. I did not realize a zero length SMB/Windows
ACL had a defined meaning. I think this draft should avoid assigning a
meaning to a zero length ACL.
--> I think that a "deleted ACL" case could be defined for both models
    of ACL. (Something like an ACL with a single ACE that is defined
    as "deleted ACL".)
    I discuss this a bit more later.

I cannot find anything in RFC8881 that defines what setting a zero
length ACL means (although you've already seen that I am not good at
reading such things;-).
It might be better if this draft avoids the "setting of a zero length ACL"
and let David's draft address the semantics of that?

What do others think?


> > A server that is configured for a acl_trueform_scope other that
> > ACL_SCOPE_FILE_OBJECT MUST not support the posix_default_acl and
> > posix_access_acl unless the true form for the file system is
> > ACL_MODEL_POSIX_DRAFT.
>
> This is a restriction which basically means that no existing NFS4
> server with native NFS4 ACL support can support "posix_access_acl".
> In multiprotocol environment where is NFSv4.0 and NFSv4.1, the NFS4
> server would have to provide server scope native NFS4 model (to support
> NFSv4.1 as as your document is only for NFSv4.2) which implies that
> server must not support POSIX ACL. And same applies in the multiprotocol
> environment with NFS4 and SMB.
>
Not exactly. posix_access_acl refers to the proposed optional
attribute for NFSv4.2 (same goes for acl_trueform_scope).
What the server does w.r.t. "true form" is up to the server.
Same goes for the scope of it.
(For example, Linux server always use POSIX draft ACLs.)

The difference is that NFSv4.0 and 4.1 clients cannot know
what the "true form" is and can only see NFSv4 ACLs (possibly
mapped to/from POSIX draft ACLs).

What the above restriction is meant to do is allow a NFSv4.2
client to know that the "true form" is ACL_MODEL_POSIX_DRAFT
when the scope is ACL_SCOPE_FILE_SYSTEM or ACL_SCOPE_SERVER
by seeing that posix_access_acl and posix_default_acl are
supported_attrs for the file system.
--> Note that acl_trueform_scope is "per server", so a client
    only needs to GETATTR this attribute once for the server.
    Then, if that is ACL_SCOPE_FILE_SYSTEM or ACL_SCOPE_SERVER,
    it can tell that the "true form" is ACL_MODEL_POSIX_DRAFT
    as soon as it sees posix_access_acl and posix_default_acl
    in supported_attrs.
    --> It does not have to bother to check acl_trueform,
        which would often result in an extra RPC.


> I think that this is too strict and what would happen is either that
> this extension would not be implemented OR section option (which I think
> everybody would do) is ignoring this restriction.
>
As above, this does not define what the server does for NFSv4.0
and NFSv4.1, it just simplifies what a NFSv4.2 client needs to
check, if it knows about and the server supports the extension
proposed by this draft.

>
> I guess that for compatibility with FreeBSD and Linux, servers would
> implement it even on NFS4 server scope model.
>
Not sure what you mean by this. Linux always uses a "true form"
of POSIX ACL and FreeBSD supports one of POSIX ACL or NFSv4 ACL
on a per-file system basic (at least for now).
- That is not changed by this extension. All this extension does
  is allow NFSv4.2 clients to use the server more effectively
  by choosing to get/set the "true form" and avoiding mapping,
  assuming the client and server both implement the extension.
Other file systems (IBM's GPFS seems to be the extant example
I am aware of) also supports a "true form" on a per-file basis.
(Again this is a server file system configuration choice and
is not affected by whether or not this extension is supported
by a NFSv4.2 server.)


>
> And I have there two additional points without quoting your document.
>
> 1)
>
> Your document does not describe what existing RFC 8881 acl and dacl
> attributes should return when user sets new "posix_default_acl" or
> "posix_access_acl" attribute. Also what should happen when another
> client tries to change acl or dacl attribute when POSIX ACL is already
> active.
>
The intent (if that is not clear, the words need to be improved)
is that the server file system shall have one "true form" for each
file object. If the attribute(s) for the other "true form" is used
on the object (which assumes the server has chosen to put attributes for
both in supported_attrs for the file syetem, the GETATTR/READIR/SETATTR
will use mapping to translate it to/from the "true form".
The intent is to discourage use of mapping. As such, the client using
these extensions can avoid that by determining what the "true form" is
for the object and using the appropriate attributes.

My thinking is that servers like FreeBSD (where the "true form" is per-file
system and mapping is not supported) would:
(A) - Support posix_access_acl and posix_default_acl, but not acl, for file
  systems that use POSIX ACLs as their "true form".
(B) - Support acl, but not posix_access_acl and posix_default_acl, for file
  systems that ise NFSv4 ACLs as their "true form".
For Linux servers, they might choose (A) above, or they might choose
to support all of acl, posix_access_acl and posix_default_acl, where acl
is supported via the mapping algorithms they already use.

It is the case of a per-file scope where things get difficult:
The server will probably want to support all of acl, posix_access_acl
and posix_default_acl, but since the "true form" is defined a one
model for each file, then how should the attribute(s) for the other
model be handled?
- I would like to avoid mapping, so I think being able to return
  "no ACL" would be preferred.
- I had thought a zero length ACL would work for this, but that no
  longer appears to be the case.
Maybe, defining an ACL with a single ACE that says "not valid" would
work?
This could be the reply for GETATTR/READDIR and could be used to
"erase" the ACL for SETATTR.
--> I think this is what the next draft will propose.


> If interop between mode, dacl and posix_default_acl, posix_access_acl
> attributes is not explicitly defined in specification then every server
> implementation will invent its own rules and it would just lead to the
> mess in multiprotocol environment where clients would not be sure what
> can happen. Like it is already with "mode" attribute in NFS v4.1 (see 2).
>
For posix_access_acl and posix_default_acl the intent is that the
POSIX draft specification defines the interaction between mode and
the POSIX ACL.

For NFSv4 ACLs, you are correct. It is a mess.
And I wish David N. the best of luck w.r.t. cleaning it up.
(OpenZFS already has a knob that specifies 4 different ways
of doing this. Another knob setting may be needed.)

Thanks for the comments, rick

>
> 2)
>
> Then there is a problem with "group" bits of "mode" attribute.
> In RFC 8881 section 6.2.4 is written:
>
> "Bits MODE4_RGRP, MODE4_WGRP, and MODE4_XGRP apply to principals
> identified in the owner_group attribute but who are not identified in
> the owner attribute."
>
> And in section 6.3.2.1 is written:
>
> "Some server implementations also add bits permitted to named users and
> groups to the group bits (MODE4_RGRP, MODE4_WGRP, and MODE4_XGRP).
> Implementations are discouraged from doing this."
> "The same user confusion seen when fetching the mode also results if
> setting the mode does not effectively control permissions for the owner,
> group, and other users; this motivates some of the requirements that
> follow."
>
> POSIX ACL with at least one named user or named group requires that
> mode group bits refers to POSIX ACL mask, and not to the POSIX ACL
> group_obj (which represents owner_group).
>
> Those RFC 8881 parts of NFS mode attribute are basically not compatible
> with POSIX ACL. Your extension does not mention this problem and does
> not handle it.
>
> Existing NFS4 servers which follows the RFC 8881 meaning of mode group
> bits, would not be able to implement this your extension for POSIX ACL,
> because existing servers cannot change behavior of mode attribute due
> backward compatibility. I think that this is a big problem.
> Servers in multiprotocol environment mostly follows the RFC 8881 advice
> to never return something like "POSIX mask" in group bits.
>
> How to solve this problem. One option could be including all mode bits
> into your new "posix_access_acl" attribute and let NFS4 clients to use
> only "posix_access_acl" on chown operation, and never use existing
> "mode" attribute at all (if the extension is supported by server).
>
> Another option could be to introduce a new "posix_mode" attribute which
> would be strictly POSIX and old "mode" attribute can stay as before for
> backward compatibility. Again clients would not use "mode" attribute
> anymore.
>
>
> Regards,
> Pali
>