Network File System (NFS) Upper-Layer Binding To RPC-Over-RDMA Version 2

3.1. DDP-Eligibility

Generally, storage protocols based on RDMA divide both read and write operations into two steps. This division enables the payload receiver to allocate the sink buffer for each I/O operation in advance of the network payload transfer. By allocating the sink buffer tactically, a good quality receiver implementation reduces the amount of data movement it must perform during and after the I/O operation.¶

During an NFS WRITE that involves explicit RDMA, first the NFS client sends a request that indicates where the NFS server can find the payload buffer, then the NFS server pulls the WRITE payload from that buffer. Likewise, during an NFS READ that involves explicit RDMA, the NFS client provides the location of the destination buffer, then the NFS server pushes the READ payload to that buffer.¶

Therefore, the following XDR data items in NFS versions 2 and 3 are DDP-eligible:¶

• The opaque file data argument in the NFS WRITE procedure¶
• The pathname argument in the NFS SYMLINK procedure¶
• The opaque file data result in the NFS READ procedure¶
• The pathname result in the NFS READLINK procedure¶

All other argument or result data items in NFS versions 2 and 3 are not DDP-eligible.¶

Regardless of whether an NFS operation is considered non-idempotent, a transport error might not indicate whether the server has processed the arguments of the RPC Call or whether the server has accessed or modified client memory associated with that RPC.¶

3.2. Reply Size Estimation

During the construction of each RPC Call message, a Requester is responsible for allocating appropriate RDMA resources to receive the corresponding Reply message. These resources must be capable of holding the entire Reply. Therefore the Requester needs to estimate the maximum possible size of the expected Reply message.¶

• Often, the expected Reply can fit in a limited number of RDMA Send messages. The Requester need not provision any RDMA resources for the Reply, relying instead on message continuation to handle the entire Reply message.¶
• In cases where the Upper Layer Binding permits direct data placement of the results (DDP), a Requester can provision Write chunks to receive those results. The Requester MUST reliably estimate the maximum size of each result receive via a Write chunk.¶
• A Requester that expects a large Reply message can provision a Reply chunk. The Requester MUST reliably estimate the maximum size of the payload received via the Reply chunk.¶
• If RDMA resources are not available to send a Reply, a Responder falls back to message continuation.¶

A correctly implemented Legacy NFS client thus avoids retransmission of non-idempotent NFS requests due to improperly estimated Reply resources.¶

3.3. RPC Binding Considerations

Legacy NFS servers typically listen for clients on UDP and TCP port 2049. Additionally, they register these ports with a local portmapper service [RFC1833].¶

A Legacy NFS server supporting RPC-over-RDMA version 2 and registering itself with the RPC portmapper MAY choose an arbitrary port or MAY use the alternative well-known port number for its RPC-over-RDMA service (see Section 9). The chosen port MAY be registered with the RPC portmapper using the netids assigned in Section 12 of [I-D.ietf-nfsv4-rpcrdma-version-two].¶

3.4. Transport Considerations

4.1. MOUNT, NLM, and NSM Protocols

Historically, NFS/RDMA implementations have conveyed the MOUNT, NLM, and NSM protocols via TCP. A Legacy NFS server implementation MUST provide support for these auxiliary protocols via TCP.¶

Moreover, there is little benefit from transporting these protocols via RDMA. Thus this document does not provide an Upper-Layer binding for them.¶

4.2. NFSACL Protocol

Legacy NFS clients and servers convey NFSACL procedures on the same transport connection and port as the NFS RPC program (100003). Utilizing the same port obviates the need for a separate rpcbind query to discover server support for this RPC program.¶

ACLs are typically small, but even large ACLs must be encoded and decoded to some degree before being being stored in local filesystems. Thus no data item in this Upper-Layer Protocol is DDP-eligible.¶

For procedures whose replies do not include an ACL object, the size of each Reply is determined directly from the NFSACL RPC program's XDR definition.¶

The NFSACL protocol does not provide a mechanism to determine the size of a received ACL in advance. When preparing for responses that include ACLs, Legacy NFS clients estimate a maximum reply size based on limits within their local file systems. If that estimation is inadequate, a Responder falls back to message continuation.¶

5.1. DDP-Eligibility

Only the following XDR data items in the COMPOUND procedure of all NFS version 4 minor versions are DDP-eligible:¶

• The opaque data field in the WRITE4args structure¶
• The linkdata field of the NF4LNK arm in the createtype4 union¶
• The opaque data field in the READ4resok structure¶
• The linkdata field in the READLINK4resok structure¶

  Write list:

     A --> B --> C

  NFS version 4 COMPOUND request:

     PUTFH LOOKUP READ PUTFH LOOKUP READLINK PUTFH LOOKUP READ
                   |                   |                   |
                   v                   v                   v
                   A                   B                   C

5.2. Reply Size Estimation

Within NFS version 4, there are certain variable-length result data items whose maximum size cannot be estimated by clients reliably because there is no protocol-specified size limit on these result arrays. These include:¶

• The attrlist4 field¶
• Fields containing ACLs such as fattr4_acl, fattr4_dacl, and fattr4_sacl¶
• Fields in the fs_locations4 and fs_locations_info4 data structures¶
• Fields which pertain to pNFS layout metadata, such as loc_body, loh_body, da_addr_body, lou_body, lrf_body, fattr_layout_types, and fs_layout_types¶

5.3. RPC Binding Considerations

NFS version 4 servers are required to listen on TCP port 2049 and are not required to register with an rpcbind service [RFC7530]. Therefore, an NFS version 4 server supporting RPC-over-RDMA version 2 MUST use the alternative well-known port number for its RPC-over-RDMA service defined in Section 9.¶

5.4. Transport Considerations

5.5. Session-Related Considerations

The presence of an NFS version 4 session (as defined in [RFC8881]) does not affect the operation of RPC-over-RDMA version 2. None of the operations introduced to support NFS sessions (e.g., the SEQUENCE operation) contain DDP-eligible data items. There is no need to match the number of session slots with the available RPC-over-RDMA version 2 credits.¶

However, there are a few new cases where an RPC transaction can fail. For example, a Requester might receive, in response to an RPC request, an RDMA2_ERROR message with a rdma_err value of RDMA2_ERR_BADXDR. These situations are not different from existing RPC errors, which an NFS session implementation can already handle for other transport types. Moreover, there might be no SEQUENCE result available to the Requester to distinguish whether failure occurred before or after the Responder executed the requested operations.¶

When a transport error occurs (e.g., an RDMA2_ERROR type message is received), the Requester proceeds, as usual, to match the incoming XID value to a waiting RPC Call. The Requester terminates the RPC transaction and reports the result status to the RPC consumer. The Requester's session implementation then determines the session ID and slot for the failed request and performs slot recovery to make that slot usable again. Otherwise, that slot is rendered permanently unavailable.¶

When an NFS session is not present (for example, when NFS version 4.0 is in use), a transport error does not indicate whether the server has processed the arguments of the RPC Call, or whether the server has accessed or modified client memory associated with that RPC.¶

6.1. NFS Version 4.0 Callback

An NFS version 4.0 client uses the SETCLIENTID operation for advertising the IP address, port, and netid of its NFS version 4.0 callback service. When an NFS version 4.0 server provides a backchannel service to an NFS version 4.0 client that uses RPC-over-RDMA version 2 for its forward channel, the server MUST advertise the backchannel service using either the "tcp" or "tcp6" netid.¶

Because the NFSv4.0 backchannel does not operate on RPC-over-RDMA, this document does not specify an Upper-Layer binding for the NFSv4.0 backchannel RPC program.¶

6.2. NFS Version 4.1 Callback

In NFS version 4.1 and newer minor versions, callback operations may appear on the same connection that is in use for NFS version 4 forward channel client requests. NFS version 4 clients and servers MUST use the mechanisms described in Section 4.5 of [I-D.ietf-nfsv4-rpcrdma-version-two] to convey backchannel operations on an RPC-over-RDMA version 2 transport.¶

The csa_back_chan_attrs argument of the CREATE_SESSION operation contains a ca_maxresponsesize field. The value in this field is the absolute maximum size of backchannel replies generated by a replying NFS version 4 client.¶

There are no DDP-eligible data items in callback procedures defined in NFS version 4.1 or NFS version 4.2. However, some callback operations, such as messages that convey device ID information, can be sizeable. A sender can use Message Continuation or a Special Payload message in this situation.¶

When an NFS version 4.1 client can support Special Payload Calls in its backchannel, it reports a backchannel ca_maxrequestsize that is larger than the connection's inline thresholds. Otherwise, an NFS version 4 server MUST use only Simple Payload or Continued Payload messages to convey backchannel operations.¶