Re: [nfsv4] Possibility of adding additional REQUIRED attributes

[Trond Myklebust <[e-mail]>]

On Mar 25, 2024, at 13:00, David Noveck <[e-mail]> wrote:

IIRC, at an earlier interim(?) meeting you raised the issue of the possible addition of other attributes to the set of REQUIRED attributes.   As I am now updating rfc5661bis to reflect the attribute categorization changes made in other documents including toe regarding mode, owner and owner_group, I'd like to follow up on this issue soon,  if that is possible.

Since I am looking to finalize a full reworking of attribute categorization to appear in -03, I'd like to close out this discussion with regard to the rfc5661bis effort, if it still has relevance to that effort.  To make progress on this, I would need a list of attributes that you believe need to be made REQUIRED as part of rfc5661bis. Right now, I don't anticipating any, although in rewriting the general description of OPTIONAL (formerly "RECOMMENDED") attributes, I am trying to provide a more balanced discussion, getting rid of the text which stressed the difficulty (for servers) of implementing attributes and underplayed the issues for the client in doing without attributes typically expected.  I'm also explicitly mentioning the need for OPTIONAL attributes in connection with OPTIONAL features.

Although I don't know of any attributes that need to be made REQUIRED in the context of fc5661bis, I feel that there are a number of attributes that need to be considered for promotion in  NFSv4.3, now that it is in the process of graduating from the snide remark stage to the initial discussion stage.  These would include the ones you alluded to, the attributes brought over from NFSv3, but I would also like us to include a substantial subset  of the per-fs OPTIONAL attributes.  Some examples:

  *   fs_charset is very easy for the server to implement and making it REQUIRED avoid the added complexity for the client of dealing with the case of this attribute not being supported.

What clients are asking for this? The Linux client doesn’t even bother to ask for fs_charset because there is nothing useful it can do with it.

  *   aclsupport raises similar issues even if it is associated with an OPTIONAL feature.  This is similar to other attributes which provide gateways to OPTIONAL features, with very simple implementations possible when the OPTIONAL feature is not supported..
  *   fs_layout_type could be made REQUIRED with servers not supporting pNFS simply returning a zero-length array.

Again, who is asking for this?

It really only makes sense to change from OPTIONAL to REQUIRED if the absence of a specific attribute causes buggy behaviour.

Otherwise, while you might be making it easier for new clients to implement only the newest minor version, please assume that clients that need to support the existing minor versions are not going to change working code.

_________________________________
Trond Myklebust
Linux NFS client maintainer, Hammerspace
[e-mail]