NFS Version 4 Working Group, IETF 68th, Prague March 18, 2007 Co-chairs: Spencer Shepler (absent) and Brian Pawlowski [Minutes from Trond Myklebust, minor adds by beepy] ====== Connectathon testing - beepy: need for more testing in June determined during wrap-up session at Connectathon - Trond: [censored vendors list] New company announced an alpha NFSv4 client Interoperability testing went fine - Andy: Global name service was tested. Linux + EMC Suggests this become a feature for future Cthons v4.0 only tested. Good enough for referrals. v4.1 adds features that are needed for migration/replication - Benny: kerberos testing? - Andy: Went well. still issues with privacy testing on Linux. - beepy: Austin should be the last testing event. Anything not tested in Austin, may need to be pared back. Audrey will work with Lisa on testing content. - Area directors: Proposed draft (v4.1) does not need to have 2 interoperating implementations. However we DO need implementation report for 4.0 beepy: how do we deal with features that are not implemented? AD: you may drop them. Andy: SPKM3 implemented and interop tested, but no interop of SPKM3/LIPKEY done yet (Hummingbird have LIPKEY, Linux does not) - beepy to move forward on pruning draft. ====== SPKM3 - A BOF was held in order to figure out what could fill that space. - 4 proposals SPKM3 DTLS (Nico) Mike Eisler's PKUTU - PKUTU is the chosen direction for various reasons (see SPKM alias) - Andy: GRID wants this - beepy: We're not reopening v4.0 draft in order to change it if we can avoid - AD: have an informal meeting with Sam about this in order to make progress. - Need to define follow up actions ====== pNFS - Overall progress made on pNFS, and it is stabilizing, but we can do better on testing of the specification - Marc: Did not go well because Sessions was thrown in at the last minute - Solaris had more implemented than Linux, and had to turn off features in order to interoperate. - Testing with sessions turned off, worked very well. ====== RDMA last call - No outstanding comments - 3 IANA issues pending, but not in way of last call - last call note will call out - Spencer to work with IANA on the outstanding registry issues - [post meeting - Spencer starting working group last call - sorry for delay!] ====== June interop testing - Lisa Weeks to organise it - Should be last testing for the v4.1 spec - Andy: many features are missing (eg. directory delegations) - beepy: we drop features that are not being considered - only issue is to ensure we don't screw up royally no interop implementation are theoretically needed - beepy: June meeting may be an official interim ietf meeting in order to pare functionality. ====== Review of changes to v4.1 (Mike Eisler) - Draft 10 was added in order to clean up v9 - functionality changes v8 -> v9 - Added a destroy_clientid operation for consistency (destroy_session already existed). Also helps server Must delete sessions before calling destroy_clientid mandatory to implement on servers - Server Owners and trunking needed explaining beefed up text - Long standing issue with SECINFO/_NO_NAME dealing with RESTORE_FH fixed. - Max slot id concept thrown out because confusing: replaced by "highest slotid" concept - ClientID can be created for specific pNFS and non-pNFS roles. - Files layout type is now ONLY v4.1. v4.0 or v3 data servers are not supported by the document. - Opaque types have been changed to a standard format: layouttype4 + opaque blob. - All new data types have '4' in their name (Benny asked could we have 4.1? Problem is that we're carrying this forward for 4.2, 4.3....) - Discretionary acls vs security acls added mode_set_masked attribute added. (Bruce's text thrown in) ACL issues all closed. - Cleared up stateid definition - State loss detection cleaned up - Changed callback path testing to CB_COMPOUND/SEQUENCE - Introduced pNFS as an OPTIONAL feature vs. a proposal NFS servers that don't want to implement pNFS, can still act as a data server. - Discourage EXCLUSIVE4 OPEN/creates if persistent sessions are used since it is unnecessary (SHOULD). - Ban EXCLUSIVE4 if the pNFS layout_hint attribute is supported (MUST) Clients should use GUARDED open/create - beepy: Zzzzzzzzzzzzzzzzzzzzz - Issue with mapping device ids and address mappings. They are _leased_ and subject to revocation. - MDS recovery clarifications - Client has to keep a copy of modified data in memory even after a COMMIT, but before a LAYOUTCOMMIT; or - Server cannot fail a LAYOUTCOMMIT in reclaim mode - Should not be an issue for NFSv4.1 storage devices - Storage device recovery clarifications - write data to the metadata server when recovering - Some feedback from reviewers that multipathing might permit client to write to secondary path for storage device - Recently defined pNFS terms: unit, pattern, stripe, stripe width - Renamed some data types to reflect their meaning nfsv4_file_layout_device4 -> nfsv4_1_file_layout_ds_addr4 - More detailed example of a device (data server) list - Clarified stripe4_sparse vs stripe4_dense - Added EXCHANGE_ID and DESTROY_CLIENTID to list of ops an NFSv4.1 data server must support. - DESTROY_SESSION can fail if there are outstanding requests on the callback channel - Added error code for STRIPE4_DENSE: NFS4ERR_PNFS_IO_HOLE - Change GET_DIR_DELEGATION results so that the operation can fail without stopping compound processing. Same done for WANT_DELEGATION - Many fields of new data types changed to include a suffix that abbreviates data type name - Added optional "previous entry" to notifications of an addition to the directory. - Re-clarified that delegation stateid from metadata server is appropriate for I/O to data server - More field nameing consistency issues - Added directory filehandle to CB_NOTIFY arguments - Put NFS filehandle in consistent places ====== What will change in 11 - Error codes - Some pNFS feedback from Garth G and Rahul I - Locking and delegations sections to be reworked by Dave - Sets of issues in Issues Trackers - Formal review issues ====== Formal review process - Want to ensure review on certain sections of spec - 3 areas targeted pNFS Sessions ACLs - Process and steps Kickoff meeting: editors act as moderators - reviewers give feedback on whether the sections make sense Volunteers sought for - Reader - scribe - reviewers - moderators - Scribe records 'defects' reported by reviewers - Initially trying 3 sections/chapters - Timeline is end of april - Focus is on new things & modified things - Audrey has sent out formal invitations for reviewers - Volunteers sought - if too many people, they should organise themselves into teams Audrey will not do. ====== beepy: What do we do about charter AD: charter will be pushed out after last call ====== New Milestones - Feb 2007 WG Last call for RPC and NFS RDMA drafts -> March - (this week?) - March 2007 WG last call for rfc1831bis (RPC version 2) status? - - April 2007 NFSv4.1 prototype, testing and interop report -> delayed to June ====== Questions, issues? - How to deal with quotas in pNFS? return error in LAYOUTGET vs LAYOUTCOMMIT for files, can still return error in WRITE - Benny wants to add a wishlist for 4.2 Benny gets to maintain it - Audrey will provide hitlist for June testing - Inspection will provide issues to test - Andy & beepy will confer on status of SPKM3 - Has anybody gone through v4.0 and looked at MUST vs implementation reality - Issue of can a compliant v4.1 client/server not support v4.0 Breaks v4.0 minor versioning rules - ...but minor version is allowed to change v4 minor versioning rules. AD: please update v4.0 minor versioning rules maybe separate small rfc motivating change etc. - beepy: What in v4.0 is - Ready for draft? - inconsistent with reality? - ready to be dropped? - For the record, Nico does not understand why we minor version. - Benny: multiple filehandles per file. Client implementations do not support it very well. Trond+Benny: Please throw out the text that describes an algorithm for how to deal with multiple filehandles per file. beepy: Take it to the mailing list and make a proposal ====== Miscellany - Nico promised to sign the blue sheet.