From Linux NFS

Revision as of 14:36, 28 October 2012

Basic server support for NFSv4.1 was merged in 2.6.30, but it deviates from rfc 5661 in a number of important ways, and is currently recommended for developers only.

This is a list of what must be done before considering the #4.0 and #4.1 linux server implementation to be minimally acceptable for general use.

For the implementation to be minimally acceptable:

• someone upgrading from the previous version should experience no loss in functionality;
• server behavior should be close enough to the spec that clients will not be forced into undocumented workarounds.

This may not be a simple as looking through the rfc's for "MUST"'s, since there are a few cases of features that are mandated by the rfc's, but that nevertheless are rarely (or never) implemented, and whose absence is easily worked around on the client.

NFSv4.1

Tasks are unclaimed except where stated otherwise.

Highest priority

These are items which must be completed before the 4.1 server can be considered minimally acceptable.

SEQ4_STATUS_RECALLABLE_STATE_REVOKED

Set SEQ4_STATUS_RECALLABLE_STATE_REVOKED when a client's failure to return a recallable object causes us to revoke the object, and be prepared to handle a FREE_STATEID from the client as acknowledgement.

This is probably needed for delegations (see http://tools.ietf.org/search/rfc5661#section-10.4.5), though that seems like a buggy-client case, so perhaps this isn't really first priority.

Also does the fault injection code need to do something here? (Lower priority.)

backchannel attribute negotiation

alloc_init_session() should sanity-check the values of the backchannel attributes that the client gives us, and fail the CREATE_SESSION if they don't meet our minimal requirements.

One starting point here: http://marc.info/?l=linux-nfs&m=128647408432218&w=2

Callback failure handling

The server is required to set SEQ4_STATUS_CB_PATH_DOWN as long as it lacks any usable backchannel for the client. (Also, CB_PATH_DOWN should be returned on DESTROY_SESSION when appropriate.)

SEQ4_STATUS_CB_PATH_DOWN_SESSION is required when unable to retry a callback due to lack of a callback for that particular session.

Set SEQ4_STATUS_BACKCHANNEL_FAULT on encountering "unrecoverable fault with the backchannel (e.g. it has lost track of the sequence ID for a slot in the backchannel)."

(pynfs should also be checking for these flags on sequence replies, and erroring out if they're set when not expected, or vice versa. Note if server and pynfs did this correctly then bugs such as http://marc.info/?l=linux-nfs&m=129841460711820&w=2 would be caught more quickly.)

Respect client-requested backchannel security

(bfields is working on this)

We ignore the csa_sec_parms and bca_sec_parms fields that specify the security to be used on the backchannel. Instead, we *always* use auth_sys credentials, because we happen to know is what the Linux client currently expects.

The client can provide a list of possible parameters. To start, we'll just take the first one that we support and ignore the rest.

GSS on the backchannel

(bfields is working on this)

Clients don't currently request gss on the backchannel, but it is mandatory to support this.

More details for gss backchannel support:

We must allow the client to pass the server gss contexts to use on the backchannel.

SEQ4_STATUS_CB_GSS_CONTEXTS_EXPIRING and SEQ4_STATUS_CB_GSS_CONTEXTS_EXPIRED should be set when required.

See the end of section 18.36.4 for more implementation details.

SP4_MACH_CRED

SP4_MACH_CRED (like SSV) is mandatory for servers to implement. It is less complicated than SSV and provides some (not all) of the advantages, and there's a better chance a client may implement it. So we should implement it now.

We want to minimize the number of optional features we don't implement; each such omission makes it harder for future clients, which will be forced to negotiate support for features that the protocol wasn't designed to negotiate support for.

destroy lockowners on unlock

(bfields is working on this)

Unlike in the 4.0 case, in the 4.1 case a stateid should become invalid on an unlock that leaves the caller with no bytes locked; a 4.1 client will no longer call release_lockowner.

Make DESTROY_SESSION wait on in-progress requests

(bfields is working on this.)

See discussion on the ietf list: [1]

(This is a slightly odd one as it's not really required till 4.2. However it appears we have some races here that could cause memory corruption, and that those races would be most easily fixed by delaying session and client destruction till in-progress requests are processed.)

Done, needs testing

Check 4.0/4.1 interactions

I think this is covered now, but it might be interesting to write python tests that send 4.0 compounds referencing 4.1 clients, or vice versa, especially for create_session, setclientid, and friends.

New open claim types

We must support CLAIM_FH and CLAIM_DELEG_CUR_FH. (We shouldn't need CLAIM_DELEG_PREV_FH.) Needs some simple pynfs tests.

DRC limit checking

We check for replies that are too big only *after* performing the operation in question. Depending on the operation, that may be too late to return NFS4ERR_REP_TOO_BIG_TO_CACHE. (For example, an irreversible filesystem operation may already have been performed.) We need to figure out how to estimate the size of the response before performing an operation, at least for operations that actually change the filesystem.

Possible fix: add to each nfsd4_ops[] a field with an upper bound on the size of a reply to that operation. Before calling the operation, check that there's room for the worst case, and return TOO_BIG_TO_CACHE if not.

Trunking

Both clientid (multiple sessions per client) and session (multiple connections per session) trunking are mandatory for a server to support. Therefore a client would be within its rights to simply refuse to interoperate with a server that didn't support either.

We could ask whether it is actually likely that a client will do that, and if there are instead obvious errors we could return that any client is likely to be able to handle gracefully.

On the first question: Supporting trunking really just means doing what the spec says when we receive multiple exchange_id's, create_sessions, or transport connections from the same client. These can arise in simple situations. For example, multi-homed servers need to know how to handle the former. Client recovery of various kinds (see BIND_CONN_TO_SESSION above) may also require that multiple connections be associated with a single session over time, even if only one is in use at a time.

On the second question: I don't see any reliable way to error out: neither BIND_CONN_TO_SESSION nor BACKCHANNEL_CTL have allowable errors that seem reasonable to me in this case. CREATE_SESSION does at least allow returning NOSPC in the case where we can't commit to the additional DRC memory, so maybe we could get away with using that in the case where we don't want to provide any more sessions.

So I think if we don't implement this soon we'll end up with idiosyncratic behavior that will be hard for clients to work around.

We'll also need some pynfs tests to make sure we're getting this right.

BIND_CONN_TO_SESSION

This is mandatory to implement on the server.

This is not just for exotic multi-connection setups.

If a client opts for #SP4_MACH_CRED protection, and if its tcp connection is broken for some reason, then it may choose to give up all its state and start from scratch. That may be good enough for very minimal first 4.1 implementations. However, clients that wish to reconnect without giving up their session state (e.g., the reply cache) will need to use BIND_CONN_TO_SESSION to associate the new connection to the old session.

Even with SP4_NONE, clients will want to be able to reconnect without losing the backchannel. Again, that will require BIND_CONN_TO_SESSION.

deferral fixes

Fix merged for 2.6.37.

The current code returns ERR_DELAY whenever an upcall is required instead of using the server's deferral mechanism, since that mechanism replays a request internally, causing SEQUENCE to fail on the second time through.

These DELAYS are hard on clients, and will cause unacceptable delays in some cases. Fix the deferral code to sleep at least a little before giving up and returning ERR_DELAY.

compound op ordering enforcement

DESTROY_SESSION must be the final operation in a compound request, nfs4err_not_only_op should be returned when appropriate. Make sure a session is defined whenever the code expects it.

The risk here is that there may be nasty DOS's (or worse) against a server that doesn't check this kind of thing carefully.

Keep client from expiring while in use by session

The session associated with a compound may be implicitly referred to by individual operations. For example, RECLAIM_COMPLETE implicitly applies to the client associated with the current session. However, we don't currently do anything to prevent the client from being freed partway through processing a compound.

Not needed immediately

This is stuff that is still a high priority, but that we can temporarily get away without doing on the grounds that they aren't absolutely required for minimal interoperability, and/or they don't introduce any new problems that don't already exist in the 4.0 implementation.

Referring triples

So, the particular requirement, from 2.10.6.3, is below (and in the below you can take the "client operation" to be an open, and the "associated object" to be a delegation created by that open):

       "For each client operation which might result in some sort of
       server callback, the server SHOULD "remember" the { session ID,
       slot ID, sequence ID } triple of the client request until the
       slot ID retirement rules allow the server to determine that the
       client has, in fact, seen the server's reply.  Until the time
       the { session ID, slot ID, sequence ID } request triple can be
       retired, any recalls of the associated object MUST carry an
       array of these referring identifiers (in the CB_SEQUENCE
       operation's arguments), for the benefit of the client."

If we ignore that "MUST", the result will be for the client to return a BADHANDLE or BADSTATEID error, as in v4.0. We have code to handle that case (by retrying) on the server. So if we ignore this requirement, the resulting behavior will be no worse than in 4.0. So I think we can get away with keeping this a *slightly* lower priority than the other stuff.

(I'd still like to see this done--if possible, at about the time it's done on the client. But it's a higher priority task on the client because there it really is mandatory: a server that lists the referring triples correctly does have a right not to have to handle those temporary BADHANDLE/BADSTATEID errors.)

Fix ERROR_RESOURCE and BADXDR returns

We shouldn't be return RESOURCE to 4.1 clients at all, and most of our BADXDR returns are probably also incorrect--instead we should be returning NFS4ERR_REP_TOO_BIG, NFS4ERR_REQ_TOO_BIG, NFS4ERR_TOO_MANY_OPS, etc.

SSV

This is still listed as mandatory in the spec, and while clients and other servers don't seem to be working on implementing this, it's not yet clear to me that there's a consensus to drop it.

Done

DESTROY_CLIENTID

DESTROY_CLIENTID is not currently used by clients, but will be, and clients should not be expected to know how to recover from the case where it is not supported. It should also be fairly easy to implement.

FREE_STATEID and TEST_STATEID

Implement SECINFO_NO_NAME

It's mandatory and not very difficult.

SECINFO should consume current filehandle

See [2]

Basic Server Reboot Recovery for 4.1

We need at least basic RECLAIM_COMPLETE support.

Question: do we need to set SEQ4_STATUS_RESTART_RECLAIM_NEEDED on any new session created by a preexisting client during the grace period? Seems like that should be necessary only if we implement persistent sessions, but I suppose it can't harm to set it otherwise.

The reboot recovery system common to 4.0 and 4.1 needs some work, but that's a preexisting 4.0 problem.

Clarify RDMA non-support

Nobody has stepped up to work on RDMA and 4.1, so while it's a violation of our principal that "someone upgrading from the previous version should experience no loss in functionality", we should probably declare the combination of RDMA and 4.1 unsupported, until someone has a chance to spend some time on it.

We've checked that create_session and bind_conn_to_session returns both indicate non-support, so that should be all we need to do for now.

See also 4.1 RDMA issues

ACL retention bits

From inspection of the code, it appears that these bits are ignored on set, cleared on return. We have no plans to really implement these, so for now that's probably adequate.

NFSv4.0

Highest priority

Required for the 4.0 server to be minimally acceptable.

We may accept new features into 4.1 without requiring these be fixed, but it will be a huge problem if they aren't somehow fixed soon.

Breaking delegations when required

(bfields is working on this.)

Our delegation implementation does not currently recall delegations on rename or unlink of a delegated file, leading to stale client caches in some cases.

This has been recently fixed for NFS-only access, making this somewhat of a lower priority, but the problem still exists for multi-protocol or local access. For example, if you ssh into the server and remove a file, or remove a file using Samba, then an NFSv4 delegation on that file will not be recalled.

We have CITI patches to address this problem in the VFS. They still have some bugs, and the design needs to be revisited.

See [[3]] for discussion.

Fix changeid

We're still relying on ctime for this, inadequate especially for ext3 (with 1-second resolution). Newer filesystems are fixing this, but some more work is needed to take advantage of improvements (for example to improve ext4's native changeid feature.)

Lower priority

Still important, but arguably may not be required for the server to be minimally acceptable.

Accepting more compounds

Out-of-spec compound restrictions: we don't, for example, currently allow the client to send more than one IO (read, write, readdir) operation in a single compound. Some day adventurous clients may run across these cases.

Stateowner DOS protection

We don't remove lockowners until close, release_lockowner, or client expiration, making it possible to DOS the server by opening a file and repeatedly locking it with a different lockowner each time, without closing the file.

(Also check treatment of open owners.)

(XXX: Actually, I think the 4.0 code is right; this is probably a 4.1 problem.)

Done, needs testing

Turn on reply cache for 4.0

The reply cache is currently off for 4.0 and 4.1. We want it to stay that way for 4.1 (sessions replaces it), but 4.0 still needs it. The slightly tricky part is identifying idempotent versus non-idempotent operations. We should be able to do that by tagging individual ops as one or the other, then noting in the xdr-decode phase whether we've encountered any non-idempotent ops.

"Lower priority" only because NFSv4 is only supported over TCP, and while the reply cache is still needed over TCP, the current reply cache design seems unlikely to help in that case. Therefore NFSv4 without a reply cache is unlikely to be any worse than NFSv3 over TCP already is. So there is still a high-priority bug here (to fix the reply cache), but it already exists in NFSv3.

Done

Fix reboot recovery

The existing reboot recovery mechanism for NFSv4.0 has some architectural problems, and the core kernel developers have asked us to replace it. The transition between the new and old system will be awkward, and the earlier it's done the better.

We have a basic design for nfsd4 server recovery.