RPC-over-RDMA Version 2 Protocol

Introduction Remote Direct Memory Access (RDMA) is a technique for moving data efficiently between network nodes. By placing transferred data directly into destination buffers using Direct Memory Access, RDMA delivers the reciprocal benefits of faster data transfer and reduced host CPU overhead. Open Network Computing Remote Procedure Call (ONC RPC, often shortened in NFSv4 documents to RPC) is a Remote Procedure Call protocol that runs over a variety of transports. Most RPC implementations today use UDP or TCP . On UDP, a datagram encapsulates each RPC message. Within a TCP byte stream, a record marking protocol delineates RPC messages. An RDMA transport, too, conveys RPC messages in a fashion that must be fully defined if RPC implementations are to interoperate when using RDMA to transport RPC transactions. Although RDMA transports encapsulate messages like UDP, they deliver them reliably and in order, like TCP. Further, they implement a bulk data transfer service not provided by traditional network transports. Therefore, we treat RDMA as a novel transport type for RPC.

Design Goals The general mission of RPC-over-RDMA transports is to leverage network hardware capabilities to reduce host CPU needs related to the transport of RPC messages. In particular, this includes mitigating host interrupt rates and limiting the necessity to copy RPC payload bytes on receivers. These hardware capabilities benefit both RPC clients and servers. On balance, however, the RPC-over-RDMA protocol design approach has been to bolster clients more than servers, as the client is typically where applications are most hungry for CPU resources. Additionally, RPC-over-RDMA transports are designed to support RPC applications transparently. However, such transports can also provide mechanisms that enable further optimization of data transfer when RPC applications are structured to exploit direct data placement. In this context, the Network File System (NFS) family of protocols (as described in , , , , , and subsequent NFSv4 minor versions) are all potential beneficiaries of RPC-over-RDMA. A complete problem statement appears in .

Motivation for a New Version Storage administrators have broadly deployed the RPC-over-RDMA version 1 protocol specified in . However, there are known shortcomings to this protocol:

The protocol's default size of Receive buffers forces the use of RDMA Read and Write transfers for small payloads, and limits the size of reverse-direction messages.
It is difficult to make optimizations or protocol fixes that require changes to on-the-wire behavior.
For some RPC procedures, the maximum reply size is difficult or impossible for an RPC client to estimate in advance.

To address these issues in a way that preserves interoperation with existing RPC-over-RDMA version 1 deployments, the current document presents an updated version of the RPC-over-RDMA transport protocol. This version of RPC-over-RDMA is extensible, enabling the introduction of OPTIONAL extensions without impacting existing implementations. See for further discussion. It introduces a mechanism to exchange implementation properties to automatically provide further optimization of data transfer. This version also contains incremental changes that relieve performance constraints and enable recovery from unusual corner cases. These changes are outlined in and include a larger default inline threshold, the ability to convey a single RPC message using multiple RDMA Send operations, support for authentication of connection peers, richer error reporting, improved credit-based flow control, and support for Remote Invalidation.

Requirements Language The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 when, and only when, they appear in all capitals, as shown here.

Terminology

Remote Procedure Calls This section highlights critical elements of the RPC protocol and the External Data Representation (XDR) it uses. RPC-over-RDMA version 2 enables the transmission of RPC messges built using XDR and also uses XDR internally to describe its header format.

Upper-Layer Protocols RPCs are an abstraction used to implement the operations of an Upper-Layer Protocol (ULP). For RPC-over-RDMA, "ULP" refers to an RPC Program and Version tuple, which is a versioned set of procedure calls that comprise a single well-defined API. One example of a ULP is the Network File System Version 4.0 . In the current document, the term "RPC consumer" refers to an implementation of a ULP running on an RPC client.

RPC Procedures Like a local procedure call, every RPC procedure has a set of "arguments" and a set of "results". A calling context invokes an RPC procedure, passing arguments to it, and the procedure subsequently returns a set of results. Unlike a local procedure call, an RPC procedure is executed remotely rather than in the local application's execution context.

RPC Transactions The RPC protocol as described in is fundamentally a message-passing protocol between one or more clients, where RPC consumers are running, and a server, where a remote execution context is available to process RPC transactions on behalf of these consumers. ONC RPC transactions consist of two types of messages:

A CALL message, or "Call", requests work. An RPC Call message is designated by the value zero (0) in the message's msg_type field.
A REPLY message, or "Reply", reports the results of work requested by an RPC Call message. An RPC Reply message is designated by the value one (1) in the message's msg_type field.

introduces the RPC transaction identifier, or "XID" for short. Each connection endpoint interprets the value of an XID in the context of the message's msg_type field.

The sender of a Call message generates an arbitrary XID value for each RPC that is unique among outstanding Calls from that sender.
The sender of a Reply message copies the XID of the initiating Call to the Reply containing the results of that procedure.

After receiving a Reply, a Requester then matches the XID value in that Reply with a Call it previously sent. The ratio of Call messages to Reply messages is typically but not always one-to-one. The most common operational paradigm is when a Requester sends a Call message to a Responder, who then sends a Reply message back to the Requester with the results of that procedure. One Call message elicits a single Reply message in response. A Responder never sends more than one Reply for each received Call message. A "retransmission" occurs when a Requester sends exactly the same Call message, with the same arguments and XID, more than once. A Requester can retransmit if it believes the network layer or Responder has dropped a Call message, or if the Responder's Reply has been likewise lost. To prevent unnecessary network traffic or the execution of non-idempotent procedures multiple times, Requesters avoid retransmitting needlessly. In rare cases, an RPC procedure may not require any results or even acknowledgement that the Responder has executed the procedure. In that case, the Requester sends a Call message but no Reply is returned. This document refers to that case as "Call-only".

Message Serialization RPC messages are always transmitted atomically. RPC peers may interleave messages, but the contents of individual messages cannot be broken up or interleaved without making the messages illegible. An RPC peer acting as a "Requester" serializes the procedure's arguments and conveys them to a "Responder" endpoint via an RPC Call message. A Call message contains an RPC protocol header with a unique XID, a header describing the requested upper-layer operation, and all arguments. An RPC peer acting as a "Responder" deserializes these arguments and processes the requested procedure. It then serializes the procedure's results into an RPC Reply message. An RPC Reply message contains an RPC protocol header with the same XID, a header describing the upper-layer reply, and all results. The Requester deserializes the results and allows the RPC consumer to proceed. At this point, the RPC transaction designated by the XID in the RPC Call message is complete, and the XID is retired.

RPC Transports The role of an "RPC transport" is to mediate the exchange of RPC messages between Requesters and Responders, bridging the gap between the RPC message abstraction and the native operations of a network transport (e.g., a socket). When an RPC transport type is connection-oriented, RPC client endpoints initiate transport connections, while RPC server endpoints wait passively to accept incoming connection requests. RPC messages may also be exchanged without a connection association. Because RPC-over-RDMA is a connection-oriented RPC transport, connectionless operation is not discussed further in the current document.

Transport Failure Recovery So that appropriate and timely recovery action can be taken, the transport implementation is responsible for notifying a Requester when an RPC Call or Reply was not able to be conveyed. Recovery can take the form of establishing a new connection, re-sending RPC Calls, or terminating RPC transactions pending on the Requester. For instance, a connection loss may occur after a Responder has received an RPC Call but before it can send the matching RPC Reply. Once the transport notifies the Requester of the connection loss, the Requester can re-send all pending RPC Calls on a fresh connection.

Forward Direction Traditionally, an RPC client acts as a Requester, while an RPC service acts as a Responder. The current document refers to this direction of RPC message passing as "forward-direction" operation.

Reverse-Direction The RPC specification does not forbid performing RPC transactions in the other direction. An RPC service endpoint can act as a Requester, in which case an RPC client endpoint acts as a Responder. This direction of RPC message passing is known as "reverse-direction" operation. During reverse-direction operation, an RPC client is responsible for establishing transport connections, even though the RPC server originates RPC Calls. RPC clients and servers are usually optimized to perform and scale well when handling traffic in the forward direction. They might not be prepared to handle operation in the reverse direction. Not until NFS version 4.1 has there been a strong need to handle reverse-direction operation.

Bi-directional Operation A pair of connected RPC endpoints may choose to use only forward-direction or only reverse-direction operation on a particular transport connection. Or, these endpoints may send Calls in both directions concurrently on the same transport connection. "Bi-directional operation" occurs when both transport endpoints act as a Requester and a Responder at the same time on a single connection. Bi-directionality is an extension of RPC transport connection sharing. Two RPC endpoints wish to exchange independent RPC messages over a shared connection but in opposite directions. These messages may or may not be related to the same workloads or RPC Programs. During bi-directional operation, forward- and reverse- direction XIDs are typically generated on distinct hosts by possibly different algorithms. There is no coordination between the generation of XIDs used in forward-direction and reverse-direction operation. Therefore, a forward-direction Requester MAY use the same XID value at the same time as a reverse-direction Requester on the same transport connection. Although such concurrent requests use the same XID value, they represent distinct RPC transactions.

External Data Representation One cannot assume that all Requesters and Responders represent data objects in the same way internally. RPC uses External Data Representation (XDR) to translate native data types and serialize arguments and results . XDR encodes data independently of the endianness or size of host-native data types, enabling unambiguous decoding of data by a receiver. XDR assumes only that the number of bits in a byte (octet) and their order are the same on both endpoints and the physical network. The smallest indivisible unit of XDR encoding is a group of four octets. XDR can also flatten lists, arrays, and other complex data types into a stream of bytes. We refer to a serialized stream of bytes that is the result of XDR encoding as an "XDR stream". A sender encodes native data into an XDR stream and then transmits that stream to a receiver. The receiver decodes incoming XDR byte streams into its native data representation format.

XDR Opaque Data Sometimes, a data item is to be transferred as-is, without encoding or decoding. We refer to the contents of such a data item as "opaque data". XDR encoding places the content of opaque data items directly into an XDR stream without altering it in any way. ULPs or applications perform any needed data translation in this case. Examples of opaque data items include the content of files or generic byte strings.

XDR Roundup The number of octets in a variable-length data item precedes that item in an XDR stream. If the size of an encoded data item is not a multiple of four octets, the sender appends octets containing zero after the end of the data item. These zero octets shift the next encoded data item in the XDR stream so that it always starts on a four-octet boundary. The addition of extra octets does not change the encoded size of the data item. Receivers do not expose the extra octets to ULPs. We refer to this technique as "XDR roundup", and the extra octets as "XDR roundup padding".

Remote Direct Memory Access When a third party transfers large RPC payloads, RPC Requesters and Responders can become more efficient. An example of such a third party might be an intelligent network interface (data movement offload), which places data in the receiver's memory so that no additional adjustment of data alignment is necessary (direct data placement or "DDP"). RDMA transports enable both of these optimizations. In the current document, the standalone term "RDMA" refers to the physical mechanism an RDMA transport utilizes when moving data.

Direct Data Placement Typically, RPC implementations copy the contents of RPC messages into a buffer before being sent. An efficient RPC implementation sends bulk data without first copying it into a separate send buffer. However, socket-based RPC implementations are often unable to receive data directly into its final place in memory. Receivers often need to copy incoming data to finish an RPC operation, if only to adjust data alignment. Although it may not be efficient, before an RDMA transfer, a sender may copy data into an intermediate buffer. After an RDMA transfer, a receiver may copy that data again to its final destination. In this document, the term "DDP" refers to any optimized data transfer where a receiving host's CPU does not move transferred data to another location after arrival. RPC-over-RDMA version 2 enables the use of RDMA Read and Write operations to achieve both data movement offload and DDP. However, note that not all RDMA-based data transfer qualifies as DDP, and some mechanisms that do not employ explicit RDMA can place data directly.

RDMA Transport Operation RDMA transports require that RDMA consumers provision resources in advance to achieve good performance during receive operations. An RDMA consumer might provide Receive buffers in advance by posting an RDMA Receive Work Request for every expected RDMA Send from a remote peer. These buffers are provided before the remote peer posts RDMA Send Work Requests. Thus this is often referred to as "pre-posting" buffers. An RDMA Receive Work Request remains outstanding until the RDMA provider matches it to an inbound Send operation. The resources associated with that Receive must be retained in host memory, or "pinned", until the Receive completes. Given these tenets of operation, the RPC-over-RDMA version 2 protocol assumes each transport provides the following abstract operations. A more complete discussion of these operations appears in .

Memory Registration Memory registration assigns a steering tag to a region of memory, permitting the RDMA provider to perform data-transfer operations. The RPC-over-RDMA version 2 protocol assumes that a steering tag of no more than 32 bits and memory addresses of up to 64 bits in length identifies each registered memory region.

RDMA Send The RDMA provider supports an RDMA Send operation, with completion signaled on the receiving peer after the RDMA provider has placed data in a pre-posted buffer. Sends complete at the receiver in the order they were posted at the sender. The size of the remote peer's pre-posted buffers limits the amount of data that can be transferred by a single RDMA Send operation.

RDMA Receive The RDMA provider supports an RDMA Receive operation to receive data conveyed by incoming RDMA Send operations. To reduce the amount of memory that must remain pinned awaiting incoming Sends, the amount of memory posted per Receive is limited. The RDMA consumer (in this case, the RPC-over-RDMA version 2 protocol) provides flow control to prevent overrunning receiver resources.

RDMA Write The RDMA provider supports an RDMA Write operation to place data directly into a remote memory region. The local host initiates an RDMA Write and the RDMA provider signals completion there. The remote RDMA provider does not signal completion on the remote peer. The local host provides the steering tag, the memory address, and the length of the remote peer's memory region. RDMA Writes are not ordered relative to one another, but are ordered relative to RDMA Sends. Thus, a subsequent RDMA Send completion signaled on the local peer guarantees that prior RDMA Write data has been successfully placed in the remote peer's memory.

RDMA Read The RDMA provider supports an RDMA Read operation to place remote source data directly into local memory. The local host initiates an RDMA Read and and the RDMA provider signals completion there. The remote RDMA provider does not signal completion on the remote peer. The local host provides the steering tags, the memory addresses, and the lengths for the remote source and local destination memory regions. The RDMA consumer (in this case, the RPC-over-RDMA version 2 protocol) signals Read completion to the remote peer as part of a subsequent RDMA Send message. The remote peer can then invalidate steering tags and subsequently free associated source memory regions.

RPC-over-RDMA Framework Before an RDMA data transfer can occur, an endpoint first exposes regions of its memory to a remote endpoint. The remote endpoint then initiates RDMA Read and Write operations against the exposed memory. A "transfer model" designates which endpoint exposes its memory and which is responsible for initiating the transfer of data. In RPC-over-RDMA version 2, only Requesters expose their memory to the Responder, and only Responders initiate RDMA Read and Write operations. Read access to memory regions enables the Responder to pull RPC arguments or whole RPC Calls from each Requester. The Responder pushes RPC results or whole RPC Replies to a Requester's memory regions to which it has write access.

Message Framing Each RPC-over-RDMA version 2 message consists of at most two XDR streams:

The "Transport stream" contains a header that describes and controls the transfer of the Payload stream in this RPC-over-RDMA message. Every RDMA Send on an RPC-over-RDMA version 2 connection MUST begin with a Transport stream.
The "Payload stream" contains part or all of a single RPC message. The sender MAY divide an RPC message at any convenient boundary but MUST send RPC message fragments in XDR stream order and MUST NOT interleave Payload streams from multiple RPC messages.

The RPC-over-RDMA framing mechanism described in this section replaces all other RPC framing mechanisms. Connection peers use RPC-over-RDMA framing even when the underlying RDMA protocol runs on a transport type with well-defined RPC framing, such as TCP. However, a ULP can negotiate the use of RDMA, dynamically enabling the use of RPC-over-RDMA on a connection established on some other transport type. Because RPC framing delimits an entire RPC request or reply, the resulting shift in framing must occur between distinct RPC messages, and in concert with the underlying transport.

Reliable Message Delivery RPC-over-RDMA provides a reliable and in-order data transport service for RPC Calls and Replies. RPC-over-RDMA transports MUST operate only on a reliable Queue Pair (QP) such as the RDMA RC (Reliable Connected) QP type as defined in Section 9.7.7 of . The Marker PDU Aligned (MPA) protocol , when deployed on a reliable transport such as TCP, provides similar functionality. Using a reliable QP type ensures in-transit data integrity and proper recovery from packet loss in the lower layers. If any pre-posted Receive buffer on the connection is not large enough to contain an incoming message, the receiving RDMA provider cannot deliver that message to the upper-layer consumer. Likewise, if no pre-posted Receive buffer is available to accept an incoming message, the receiving RDMA provide cannot pass that message to the consumer. Exceeding these limits results in a transition to a QP error state, the loss of an in-flight message, and the potential loss of the connection. Therefore, senders need to respect peer receiver resource limits to ensure that the transport service can deliver every message reliably. Two operational parameters communicate these limits between RPC-over-RDMA peers: credits and inline threshold.

Flow Control RPC-over-RDMA version 2 employs end-to-end credit-based flow control on each connection to prevent a sender from transmitting more messages than a receiver is prepared to accept . Credit-based flow control is relatively simple, providing automated management of receive buffer allocation and robust operation in the face of bursty traffic while enabling effective pipelining. The RPC-over-RDMA version 2 flow control mechanism relies on reliable and in-order message delivery guarantees provided by the underlying RDMA transport service. An RPC-over-RDMA version 2 credit represents the capability to convey exactly one RPC-over-RDMA version 2 message, regardless of its size, via an RDMA Send/Receive pair. Because an RPC-over-RDMA version 2 connection is full-duplex, each connection peer has its own set of credits. The two peers manage their credit limits independently, although they communicate these values by piggy-backing them on a message in the opposite direction. Each peer tracks four critical values for each connection. The peer uses these values to determine when it is safe to send a message on the connection.

Sent message count:: The total number of RDMA Send channel operations the peer has posted on the connection. The peer MUST set this value to zero (0) when a connection is first established.
Received message count:: The total number of RDMA Receive channel operations that have completed on the connection. The peer MUST set this value to zero (0) when a connection is first established.
Advertised credits:: The number of messages the peer is prepared to receive. This value is typically equal to or less than the number of RDMA Receive channel operations that are currently pending on the connection.
Received credits:: The value in the rdma_start.rdma_credit field in the most recently received message on the connection. The peer MUST set this value to one (1) when a connection is first established.

When constructing an RPC-over-RDMA header to be sent, a sender MUST set the header's rdma_start.rdma_credit field to the sum of the connection's "Sent message count" and "Advertised credits" values. The sender MUST NOT post this message if the connection's "Sent message counter" is greater than the connection's current "Received credits" value. To handle counter wrapping, the sender uses appropriate modulo arithmetic to perform this comparison. Because the rdma_start.rdma_credit field is 32 bits wide, credit limit values MUST be less than 2^31 - 1. For a given bandwidth-delay product, a peer selects an "Advertised credits" value that is large enough to maximize throughput while not overwhelming its local memory resources. A peer MAY adjust its "Advertised credits" to match the needs or policies in effect on either peer. For instance, a peer may reduce its "Advertised credits" to accommodate the available resources in a Shared Receive Queue. Certain RDMA implementations may impose additional flow-control restrictions, such as limits on RDMA Read operations in progress at the Responder. Accommodation of such checks is considered the responsibility of each RPC-over-RDMA version 2 implementation.

Asynchronous Credit Grants Credit accounting information is usually piggy-backed on payload-bearing messages. However, on occasion, a peer might need to reset its credit limit without sending an RPC payload. A receiving peer can send a message using a special header type when the sender's credit limit approaches exhaustion during a stream of unacknowledged messages. See for information about this header type. Unlike RPC-over-RDMA version 1, the credit grant on an RPC-over-RDMA version 2 connection MAY be zero. In that case, the sender waits until the receiver sends it an asynchronous credit refresh. To prevent a sender from ever having to wait for a credit refresh, a good receiver implementation provides a credit refresh before half its credit limit is exceeded. To prevent transport deadlock, receivers MUST always be in a position to receive one asynchronous credit update message, in addition to payload-bearing messages. A receiver can do this is by posting one more RDMA Receive than its "Advertised credits" value.

Inline Threshold An "inline threshold" value is the largest message size (in octets) that can be conveyed in one direction between peer implementations using RDMA Send and Receive channel operations. An inline threshold value is less than or equal to the largest number of octets the sender can post in a single RDMA Send operation. It is also less than or equal to the largest number of octets the receiver can reliably accept via a single RDMA Receive operation. Each connection has two inline threshold values. There is one for messages flowing from Requester-to-Responder (referred to as the "call inline threshold"), and one for messages flowing from Responder-to-Requester (referred to as the "reply inline threshold"). Peers can advertise their inline threshold values via RPC-over-RDMA version 2 Transport Properties (see ). In the absence of an exchange of Transport Properties, connection peers MUST assume both inline thresholds are 4096 octets.

Initial Connection State Immediately upon connection establishment, both peers MUST allow only one outstanding RPC-over-RDMA message on the connection at a time until both the transport protocol version is established and both peers have received an initial credit limit. Note that because RPC-over-RDMA versions 1 and 2 each use a different flow control mechanism, the meaning of the value in the rdma_start.rdma_credit field depends on the value in the rdma_start.vers field. The second word of each transport header conveys the transport protocol version. Immediately after the client establishes a connection, it sends a single valid RPC-over-RDMA message with the value two (2) in the rdma_start.rdma_vers field. Because the server might support only RPC-over-RDMA version 1, this initial message MUST NOT be larger than the version 1 default inline threshold of 1024 octets.

Server Supports RPC-over-RDMA Version 2 If the server supports RPC-over-RDMA version 2, it sends RPC-over-RDMA messages back to the client with the value two (2) in the rdma_start.rdma_vers field. Both peers may assume the default inline threshold value for RPC-over-RDMA version 2 connections (4096 octets).

Server Does Not Support RPC-over-RDMA Version 2 If the server does not support RPC-over-RDMA version 2, it MUST send an RPC-over-RDMA message to the client with an XID that matches the client's first message, RDMA2_ERROR in the rdma_start.rdma_htype field, and with the error code RDMA2_ERR_VERS. This message also reports the range of RPC-over-RDMA protocol versions that the server supports. To continue operation, the client selects a protocol version in that range for subsequent messages on this connection. If the connection is dropped immediately after an RDMA2_ERROR/RDMA2_ERR_VERS message is received, the client should try to avoid a version negotiation loop when re-establishing another connection. It can assume that the server does not support RPC-over-RDMA version 2. A client can assume the same situation (i.e., no server support for RPC-over-RDMA version 2) if the initial negotiation message is lost or dropped. Once the version negotiation exchange is complete, both peers may use the default inline threshold value for the negotiated transport protocol version.

Client Does Not Support RPC-over-RDMA Version 2 The server examines the RPC-over-RDMA protocol version used in the first RPC-over-RDMA message it receives. If it supports this protocol version, it MUST use it in all subsequent messages it sends on that connection. The client MUST NOT change the protocol version for the duration of the connection.

Using Direct Data Placement RPC-over-RDMA version 2 provides a mechanism for moving part of an RPC message via a data transfer distinct from RDMA Send and Receive. For example, a sender can remove one or more XDR data items from the Payload stream. These items are then conveyed via other mechanisms, such as one or more RDMA Read or Write operations.

Chunks and Segments A Requester records the location information for each registered memory region associated with an RPC payload in the transport header of an RPC-over-RDMA message. With this information, the Responder uses RDMA Read and Write operations to retrieve arguments contained in the specified region of the Requester's memory or place results in that region. A "segment" is a transport header data object that contains the precise coordinates of a contiguous registered memory region. Each segment contains the following information:

Handle:: A steering Tag (STag) or R_key generated by registering this memory with the RDMA provider.
Length:: The length of the segment's memory region, in octets. The length of a segment MAY be aligned to a single octet. An "empty segment" is defined as a segment with the value zero (0) in its length field.
Offset:: The offset or beginning memory address of the segment's memory region.

The meaning of the values contained in these fields is elaborated in . A "chunk" is simply a set of segments that have a related purpose. A Requester MAY divide a chunk into segments using any convenient boundaries. The length of a chunk is defined as the sum of the lengths of the segments that comprise it.

Reducing a Payload Stream We refer to a data item that a sender removes from a Payload stream to transmit separately as a "reduced" data item. After a sender has finished removing XDR data items from a Payload stream, we refer to it as a "reduced" Payload stream. A set of segments that describe memory regions containing a single reduced data item is categorized as a "data item chunk." Not all XDR data items benefit from Direct Data Placement. For example, small data items or data items that require XDR unmarshaling by the receiver do not benefit from DDP. Moreover, it is impractical for receivers to prepare for every possible XDR data item in a protocol to appear in a data item chunk. Specifying which data items are DDP-eligible is done in separate standards track documents known as "Upper Layer Bindings". A ULB identifies which XDR data items a peer MAY transfer using DDP. We refer to such data items as "DDP-eligible." Senders MUST NOT reduce any other XDR data items. Detailed requirements for ULB specifications appear in of the current document.

Moving Whole RPC Messages using Explicit RDMA RPC-over-RDMA version 2 also enables the movement of a whole RPC message via data transfer distinct from RDMA Send and Receive. A sender registers the memory containing a Payload stream without regard to data item boundaries or DDP-eligibility. The Payload stream is then conveyed via other mechanisms, such as one or more RDMA Read or Write operations. A set of segments that describe memory regions containing a Payload stream is categorized as a "body chunk". A sender may first reduce that Payload stream if it contains one or more DDP-eligible data items. The sender moves these data items using data items chunks, and the reduced Payload stream using a body chunk.

Encoding Chunks The RPC-over-RDMA version 2 transport protocol does not place a limit on chunk size. However, each ULP may cap the amount of data that can be transferred by a single RPC transaction. For example, NFS implementations typically have settings that restrict the payload size of NFS READ and WRITE operations. The Responder can use such limits to sanity check chunk sizes before using them in RDMA operations.

Read Chunks A "Read chunk" contains data that its receiver pulls from the sender. Each Read chunk is a set of one or more "Read segments" encoded as a list. A Read segment consists of a Position field followed by a segment, as defined in .

Position:: The byte offset in the unreduced Payload stream where the receiver reinserts the data item conveyed in the chunk. The sender MUST compute the Position value from the beginning of the unreduced Payload stream, which begins at Position zero. All segments in the same Read chunk share the same Position value, even if one or more of the segments have a non-four-byte-aligned length. The value in this field MUST be a multiple of four.

When constructing an RPC-over-RDMA message, the sender registers memory regions containing data intended for RDMA Read operations. It advertises the coordinates of these regions in Read chunks added to the transport header of an RPC-over-RDMA message. The receiver of this message then pulls the chunk's data from the sender using RDMA Read operations. When receiving a Read chunk, the receiver inserts the first Read segment in a Read chunk into the Payload stream at the byte offset indicated by its Position field. The receiver concatenates Read segments whose Position field value matches this offset until there are no more Read segments at that Position value.

The Read List Each RPC-over-RDMA message carries a list of Read segments that make up the set of Read chunks for that message. When no RDMA Read operations are needed to complete the transmission of the message's Payload stream, the message's Read list is empty. If a Responder receives a Read list whose segment position values do not appear in monotonically increasing order, it MUST discard the message without processing it and respond with an RDMA2_ERROR message with the rdma_xid field set to the XID of the malformed message and the rdma_err field set to RDMA2_ERR_BAD_XDR.

The Call Chunk The Call chunk is a Read chunk that acts as a body chunk containing an RPC Call message. A Requester can utilize a Call chunk at any time. However, using a Call chunk is less efficient than an RDMA Send. A Read chunk may act as either a data item chunk or a body chunk. When the chunk's position is zero, it acts as a body chunk. Otherwise, it is a data item chunk containing exactly one XDR data item.

Read Completion A Responder acknowledges that it is finished with the Requester's Read chunk memory regions when it sends the corresponding RPC Reply message. The Requester may then invalidate memory regions belonging to Read chunks associated with the associated RPC Call message.

Write Chunks Each "Write chunk" consists of a counted array of zero or more segments, as defined in . The function of a Write chunk depends on the direction of the containing RPC-over-RDMA message. In a Call message, a Write chunk advertises registered memory regions into which the Responder may push data. In a Reply message, a Write chunk reports how much data has been pushed. A Requester provisions Write chunks for an RPC transaction long before the Responder has constructed a corresponding Reply message. A Requester typically does not know the actual length of the result data items or Reply to be returned, since the Reply does not yet exist. Thus, a Requester MUST provision Write chunks large enough to accommodate the maximum possible size of each returned data item. An "empty Write chunk" is a Write chunk with a zero segment count. By definition, the length of an empty Write chunk is zero. An "unused Write chunk" has a non-zero segment count, but all of its segments are empty segments.

The Write List Each RPC-over-RDMA message carries a list of Write chunks. When no DDP-eligible data items are to appear in the Reply to an RPC transaction, the Requester provides an empty Write list in the RPC Call, and the Responder leaves the Write list empty in the matching RPC Reply. When a Write chunk appears in the Write list, it acts only as a data item chunk. For each Write chunk in the Write list, the Responder pushes one DDP-eligible data item to the Requester. It fills the chunk contiguously and in segment array order until the Responder has written that data item to the Requester in its entirety. The Responder MUST copy the segment count and all segments from the Requester-provided Write chunk into the RPC Reply message's transport header. As it does so, the Responder updates each segment length field to reflect the actual amount of data returned in that segment. The Responder then sends the RPC Reply message via an RDMA Send operation.

The Reply Chunk The Reply chunk is a single Write chunk that acts as a body chunk. that contains an RPC Reply message. When a Requester estimates that the Reply message can exceed the connection's ability to convey that Reply using RDMA Send operations, it should provision a Reply chunk.

Write Completion A Responder acknowledges that it is finished updating the Requester's Write chunk memory regions when it sends the corresponding RPC Reply message. The RDMA provider guarantees that the written data is at rest before the next Receive operation, which typically contains the corresponding RPC Reply, completes. The Requester may then invalidate memory regions belonging to Write chunks associated with the associated RPC Call message.

Write Chunk Roundup When provisioning a Write chunk for a variable-length result data item, the Requester MUST NOT include additional space for XDR roundup padding. A Responder MUST NOT write XDR roundup padding into a Write chunk, even if the result is shorter than the available space in the chunk.

Reducing Complex XDR Data Types XDR data items may appear in body chunks without regard to their DDP-eligibility. As body chunks contain a Payload stream, they MUST include all appropriate XDR roundup padding to maintain proper XDR alignment of their contents. However, a data item chunk MUST contain only one XDR data item, and the chunk MUST occupy a four-byte aligned length in the Payload stream so that subsequent data items remain properly aligned once the reduced data item is removed from the Payload stream.

Variable-Length Data Items When a sender reduces a variable-length XDR data item, the length of the item MUST remain in the Payload stream. The sender MUST omit the item's XDR roundup padding from the Payload stream and the chunk. The chunk's total length MUST be the same as the encoded length of the data item.

Counted Arrays When reducing a data item that is a counted array data type, the count of array elements MUST remain in the Payload stream. The sender MUST move the array elements into the chunk. For example, when encoding an opaque byte array as a chunk, the count of bytes stays in the Payload stream, and the sender places the bytes in the array in the chunk. Individual array elements appear in a chunk in their entirety. For example, when encoding an array of arrays as a chunk, the count of items in the enclosing array stays in the Payload stream. But each enclosed array, including its item count, is transferred as part of the chunk.

Optional-Data Similar to a counted array, when reducing an optional-data data type, the discriminator field MUST remain in the Payload stream. The sender MUST place the data, when present, in the chunk.

XDR Unions A union data type MUST NOT be made DDP-eligible. However, one or more of its arms MAY be made DDP-eligible, subject to the other requirements in this section.

Reverse-Direction Operation The terminology used in this section is introduced in .

Sending a Reverse-Direction RPC Call An RPC-over-RDMA server endpoint constructs the transport header for a reverse-direction RPC Call as follows:

The server generates a new XID value (see for full requirements) and places it in the rdma_xid field of the transport header and the xid field of the RPC Call message. The RPC Call header MUST start with the same XID value that is present in the transport header.
The rdma_vers field of each reverse-direction Call MUST contain the same value as forward-direction Calls on the same connection.
The server fills in the rdma_credit field with the credit values for the connection, as described in .
The server determines the Payload format for the RPC message and fills in the rdma_htype field as appropriate (see Sections and ). also covers the disposition of the chunk lists.

Sending a Reverse-Direction RPC Reply An RPC-over-RDMA client endpoint constructs the transport header for a reverse-direction RPC Reply as follows:

The client copies the XID value from the matching RPC Call and places it in the rdma_xid field of the transport header and the xid field of the RPC Reply message. The RPC Reply header MUST start with the same XID value that is present in the transport header.
The rdma_vers field of each reverse-direction Call MUST contain the same value as forward-direction Replies on the same connection.
The client fills in the rdma_credit field with the credit values for the connection, as described in .
The client determines the Payload format for the RPC message and fills in the rdma_htype field as appropriate (see Sections and ). also covers the disposition of the chunk lists.

When Reverse-Direction Operation is Not Supported An RPC-over-RDMA transport endpoint does not have to support reverse-direction operation. There might be no mechanism in the transport implementation to do so. Or, the transport implementation might support operation in the reverse direction, but the Upper-Layer Protocol might not configure the transport to handle reverse-direction traffic. If an endpoint is unprepared to receive a reverse-direction message, loss of the RDMA connection might result. Thus a denial of service can occur if an RPC server continues to send reverse-direction messages after a client that is not prepared to receive them reconnects to that server. Connection peers indicate their support for reverse-direction operation as part of the exchange of Transport Properties just after a connection is established (see ). When dealing with the possibility that the remote peer has no transport level support for reverse-direction operation, the Upper-Layer Protocol is responsible for informing peers when reverse-direction operation is supported. Otherwise, even a simple reverse-direction RPC NULL procedure from a peer could result in a lost connection. Therefore, an Upper-Layer Protocol MUST NOT perform reverse-direction RPC operations until the RPC client indicates support for them.

Using Chunks During Reverse-Direction Operation Reverse-direction operations can use chunks for DDP-eligible data items and Special payload formats the same way chunks are used in forward-direction operation. Connection peers indicate their support for using chunks in the reverse direction as part of the exchange of Transport Properties just after a connection is established (see ). However, an implementation might support only Upper-Layer Protocols that have no DDP-eligible data items. Such Upper-Layer Protocols can use only small messages, or they might have a native mechanism for restricting the size of reverse-direction RPC messages, obviating the need to handle chunks in the reverse direction. When there is no Upper-Layer Protocol need for chunks in the reverse direction, implementers MAY choose not to provide support for chunks in the reverse direction, thus avoiding the complexity of implementing support for RDMA Reads and Writes in the reverse direction. When an RPC-over-RDMA transport implementation does not support chunks in the reverse direction, RPC endpoints use only the Simple Payload format without data item chunks or the Continued Payload format without data item chunks to send RPC messages in the reverse direction. If a reverse-direction Requester provides a non-empty chunk list to a Responder that does not support chunks, the Responder MUST report its lack of support using one of the error values defined in .

Reverse-Direction Retransmission In rare cases, an RPC server cannot complete an RPC transaction or cannot send a Reply. In these cases, the Requester may send the RPC transaction again using the same RPC XID. In the forward direction, an RPC client is the Requester. The client is always responsible for ensuring a transport connection is in place before sending a dropped Call again. With reverse-direction operation, an RPC server is the Requester. Because an RPC server is not responsible for establishing transport connections with clients, the Requester is unable to retransmit a reverse-direction Call whenever there is no transport connection. In this case, the RPC server must wait for the RPC client to re-establish a transport connection before it can retransmit reverse-direction RPC Calls. If the forward-direction Requester has no work to do, it can be some time before the RPC client re-establishes a transport connection. An RPC server may need to abandon a pending reverse-direction RPC Call to avoid waiting indefinitely for the client to re-establish a transport connection. Therefore forward-direction Requesters SHOULD maintain a transport connection as long as the RPC server might send reverse-direction Calls. For example, while an NFS version 4.1 client has open delegated files or active pNFS layouts, it maintains one or more transport connections to enable the NFS server to perform callback operations.

Call-Only Operation There is no corresponding Reply to a Call-only procedure. Thus there is no opportunity for the Responder to indicate it has completed its use of Read or Call chunks that hold arguments or the whole Call. In addition, Write and Reply chunks are not necessary because there are no results and no Reply message. Therefore, Requesters MUST NOT use chunks when sending Call-only RPC procedures.

Transport Properties RPC-over-RDMA version 2 enables connection endpoints to exchange information about implementation properties. Compatible endpoints use this information to optimize data transfer. Initially, only a small set of transport properties are defined. The protocol provides header types to exchange transport properties (see and ). Both the set of transport properties and the operations used to communicate them may be extended. Within RPC-over-RDMA version 2, such extensions are OPTIONAL. A discussion of extending the set of transport properties appears in .

Transport Properties Model The current document specifies a basic set of receiver and sender properties. Such properties are specified using a code point that identifies the particular transport property and a nominally opaque array containing the XDR encoding of the property. The following XDR types handle transport properties: ; }; typedef rpcrdma2_propval rpcrdma2_propset<>; typedef uint32 rpcrdma2_propsubset<>; ]]> The rpcrdma2_propid type specifies a distinct transport property. The property code points are defined as const values rather than elements in an enum type to enable the extension by concatenating XDR definition files. The rpcrdma2_propval type carries the value of a transport property. The rdma_which field identifies the particular property, and the rdma_data field contains the associated value of that property. A zero-length rdma_data field represents the default value of the property specified by rdma_which. Although the rdma_data field is opaque, receivers interpret its contents using the XDR type associated with the property specified by rdma_which. When the contents of the rdma_data field do not conform to that XDR type, the receiver MUST return the error RDMA2_ERR_BAD_PROPVAL using the header type RDMA2_ERROR, as described in . For example, the receiver of a message containing a valid rpcrdma2_propval returns this error if the length of rdma_data is greater than the length of the transferred message. Also, when the receiver recognizes the rpcrdma2_propid contained in rdma_which, it MUST report the error RDMA2_ERR_BAD_PROPVAL if either of the following occurs:

The nominally opaque data within rdma_data is not valid when interpreted using the property-associated typedef.
The length of rdma_data is insufficient to contain the data represented by the property-associated typedef.

A receiver does not report an error if it does not recognize the value contained in rdma_which. In that case, the receiver does not process that rpcrdma2_propval. Processing continues with the next rpcrdma2_propval, if any. The rpcrdma2_propset type specifies a set of transport properties. The protocol does not impose a particular ordering of the rpcrdma2_propval items within it. The rpcrdma2_propsubset type identifies a subset of the properties in a rpcrdma2_propset. Each bit in the mask denotes a particular element in a previously specified rpcrdma2_propset. If a particular rpcrdma2_propval is at position N in the array, then bit number N mod 32 in word N div 32 specifies whether the defined subset includes that particular rpcrdma2_propval. Words beyond the last one specified are assumed to contain zero.

Current Transport Properties specifies a basic set of transport properties. The columns contain the following information:

The column labeled "Property" contains a name of the transport property described by the current row.
The column labeled "Code" specifies the code point that identifies this property.
The column labeled "XDR type" gives the XDR type of the data used to communicate the value of this property. This data type overlays the data portion of the nominally opaque rdma_data field.
The column labeled "Default" gives the default value for the property.
The column labeled "Section" indicates the section within the current document that explains the use of this property.

Property	Code	XDR type	Default
Maximum Send Size	1	uint32	4096
Receive Buffer Size	2	uint32	4096
Maximum Segment Size	3	uint32	1048576
Maximum Segment Count	4	uint32	16
Reverse-Direction Support	5	uint32	0
Host Auth Message	6	opaque<>	N/A

Maximum Send Size The value of this property specifies the maximum size, in octets, of Send payloads. The endpoint receiving this value can size its Receive buffers based on the value of this property.

Receive Buffer Size The value of this property specifies the minimum size, in octets, of pre-posted receive buffers. A sender can subsequently use this value to determine when a message to be sent fits in pre-posted receive buffers that the receiver has set up. In particular:

Requesters may use the value to determine when to use a Call chunk or Message Continuation when sending a Call.
Requesters may use the value to determine when to provide a Reply chunk when sending a Call, based on the maximum possible size of the Reply.
Responders may use the value to determine when to use a Reply chunk provided by the Requester, given the actual size of a Reply.

Maximum Segment Size The value of this property specifies the maximum size, in octets, of a segment this endpoint is prepared to send or receive.

Maximum Segment Count The value of this property specifies the maximum number of segments that can appear in a Requester's transport header.

Reverse-Direction Support The value of this property specifies a client implementation's readiness to process messages that are part of reverse-direction RPC requests. Multiple levels of support are distinguished:

The value RDMA2_RVRSDIR_NONE indicates that the sender does not support reverse-direction operation.
The value RDMA2_RVRSDIR_SIMPLE indicates that the sender supports using only Simple Format messages without data item chunks for reverse-direction messages.
The value RDMA2_RVRSDIR_CONT indicates that the sender supports using either Simple Format without data item chunks or Continued Format messages without data item chunks for reverse-direction messages.
The value RDMA2_RVRSDIR_GENL indicates that the sender supports reverse-direction messages in the same way as forward-direction messages.

When a peer does not provide this property, the default is the peer does not support reverse-direction operation.

Host Authentication Message The value of this transport property enables the exchange of host authentication material. This property can accommodate authentication handshakes that require multiple challenge-response interactions and potentially large amounts of material. ; ]]> When this property is not present, the peer(s) remain unauthenticated. Local security policy on each peer determines whether the connection is permitted to continue.

Transport Messages Each transport message consists of multiple sections.

A transport header prefix, as defined in . Among other things, this structure indicates the header type.
The transport header proper, as defined by one of the sub-sections below. See for the mapping between header types and the corresponding header structure.
Potentially, all or part of an RPC message payload.

This organization differs from that presented in the definition of RPC-over-RDMA version 1 , which defined the first and second of the items above as a single XDR data structure. The new organization is in keeping with RPC-over-RDMA version 2's extensibility model, which enables the definition of new header types without modifying the XDR definition of existing header types.

Transport Header Types lists the RPC-over-RDMA version 2 header types. The columns contain the following information:

The column labeled "Operation" names the particular operation.
The column labeled "Code" specifies the value of the header type for this operation.
The column labeled "XDR type" gives the XDR type of the data structure used to organize the information in this new header type. This data immediately follows the universal portion on the transport header present in every RPC-over-RDMA transport header.
The column labeled "Msg" indicates whether this operation is followed (or not) by an RPC message payload.
The column labeled "Section" refers to the section within the current document that explains the use of this header type.

Operation	Code	XDR type	Msg
Report Transport Error	4	rpcrdma2_hdr_error	No
Grant Credits	5	void	No
Specify Properties (Middle)	6	rpcrdma2_hdr_connprop	No
Specify Properties (Final)	7	rpcrdma2_hdr_connprop	No
Convey External RPC Call Message	8	rpcrdma2_hdr_call_external	No
Convey Continued RPC Call Message	9	rpcrdma2_hdr_call_middle	Yes
Convey Inline RPC Call Message	10	rpcrdma2_hdr_call_inline	Yes
Convey External RPC Reply Message	11	rpcrdma2_hdr_reply_external	No
Convey Continued RPC Reply Message	12	rpcrdma2_hdr_reply_middle	Yes
Convey Inline RPC Reply Message	13	rpcrdma2_hdr_reply_inline	Yes

RPC-over-RDMA version 2 peers are REQUIRED to support all message header types in . RPC-over-RDMA version 2 implementations that receive an unrecognized header type MUST respond with an RDMA2_ERROR message with an rdma_err field containing RDMA2_ERR_INVAL_HTYPE and drop the incoming message without processing it further.

Headers and Chunks Most RPC-over-RDMA version 2 data structures have antecedents in corresponding structures in RPC-over-RDMA version 1. As is typical for new versions of an existing protocol, the XDR data structures have new names, and there are a few small changes in content. In some cases, there have been structural re-organizations to enable protocol extensibility.

Common Transport Header Prefix The rpcrdma_common structure defines the initial part of each RPC-over-RDMA transport header for RPC-over-RDMA version 2 and subsequent versions. RPC-over-RDMA version 2's use of these first four words aligns with that of version 1 as required by . However, there are crucial structural differences in the XDR definition of RPC-over-RDMA version 2: in the way that these words are described by the respective XDR descriptions:

The header type is represented as a uint32 rather than as an enum type. An enum would need to be modified to reflect additions to the set of header types made by later extensions.
The header type field is part of an XDR structure devoted to representing the transport header prefix, rather than being part of a discriminated union, that includes the body of each transport header type.
There is now a prefix structure (see ) of which the rpcrdma_common structure is the initial segment. This prefix is a newly defined XDR object within the protocol description, which constrains the universal portion of all header types to the four words in rpcrdma_common.

These changes are part of a more considerable structural change in the XDR definition of RPC-over-RDMA version 2 that facilitates a cleaner treatment of protocol extension. The XDR appearing in reflects these changes, which discusses in further detail.

Header Types The header types defined and used in RPC-over-RDMA version 1 are not carried over into RPC-over-RDMA version 2, although there are easy equivalents to the version 1 procedures:

The RDMA2_ERROR header (defined in ) has an XDR definition that differs from that in RPC-over-RDMA version 1, and its modifications are all compatible extensions.
Senders use RDMA2_CALL_INLINE or RDMA2_REPLY_INLINE (defined in Sections and ) in place of RDMA_MSG. There are minor differences in the on-the-wire format between the version 1 procedure and the version 2 header types.
Senders use RDMA2_CALL_EXTERNAL or RDMA2_REPLY_EXTERNAL (defined in Sections and ) in place of RDMA_NOMSG. There are minor differences in the on-the-wire format between the version 1 procedure and the version 2 header types.
RDMA2_CONNPROP_MIDDLE and RDMA2_CONNPROP_FINAL (defined in Sections and ) are new header types devoted to enabling connection peers to exchange information about their transport properties.

RDMA2_ERROR: Report Transport Error RDMA2_ERROR reports a transport layer error on a previous transmission. See for details on the use of this header type.

RDMA2_GRANT: Grant Credits The RDMA2_GRANT header type enables a connection peer to update credit information without conveying a payload. This message carries no payload except for a struct rpcrdma2_hdr_prefix. The rdma_xid field is unused. Senders MUST set the rdma_xid field to zero and receivers MUST ignore the value in this field.

RDMA2_CONNPROP_MIDDLE: Exchange Transport Properties The RDMA2_CONNPROP_MIDDLE header type enables a connection peer to publish the properties of its implementation to its remote peer. A peer sends an RDMA2_CONNPROP_MIDDLE header type when it has one or more properties to send that do not fit within the default inline threshold for the RPC-over-RDMA version that is in effect. A peer may encounter properties that it does not recognize or support. In such cases, the receiver ignores unsupported properties without generating an error response. If a peer sends follows an RDMA2_CONNPROP_MIDDLE header type with anything other than another RDMA2_CONNPROP_MIDDLE message or an RDMA2_CONNPROP_FINAL message, the receiver MUST respond with an RDMA2_ERROR header type and set its rdma_err field to RDMA2_ERR_INVAL_CONT and drop the incoming message without processing it further.

RDMA2_CONNPROP_FINAL: Exchange Transport Properties The RDMA2_CONNPROP_FINAL header type enables a connection peer to publish the properties of its implementation to its remote peer. Each peer sends an RDMA2_CONNPROP_FINAL header type as the final CONNPROP-type message after the client has established a connection. The size of this message is limited to the default inline threshold for the RPC-over-RDMA version that is in effect. A peer may encounter properties that it does not recognize or support. In such cases, the receiver ignores unsupported properties without generating an error response. If a peer sends a CONNPROP-type message on a connection after it has sent an RDMA2_CONNPROP_FINAL message, the receiver MUST respond with an RDMA2_ERROR header type and set its rdma_err field to RDMA2_ERR_INVAL_CONT and drop the incoming message without processing it further.

RDMA2_CALL_EXTERNAL: Convey External RPC Call Message RDMA2_CALL_EXTERNAL conveys an RPC Call message payload using explicit RDMA operations. The Responder reads the Payload stream from a memory area specified by the Call chunk. The sender MUST set the rdma_xid field to the same value as the xid of the RPC Reply message payload.

rdma_inv_handle:: The rdma_inv_handle field contains a 32-bit RDMA handle that the Responder may use in a Send With Invalidation operation. See .
rdma_call:: The rdma_call field anchors a list of one or more Read segments that contain the RPC Call's Payload stream.
rdma_reads:: The rdma_reads field anchors a list of zero or more Read segments that contain data item chunks.
rdma_provisional_writes:: The rdma_writes field anchors a list of zero or more provisional Write chunks.
rdma_provisional_reply:: The rdma_reply field is a list containing zero or one provisional Reply chunk.

RDMA2_CALL_MIDDLE: Convey Continued RPC Call Message RDMA2_CALL_MIDDLE conveys a beginning or middle portion of an RPC Call message immediately following the transport header in the send buffer. The sender MUST set the rdma_xid field to the same value as the xid of the RPC Reply message payload. The sender sets the rdma_remaining field to the number of bytes in the RPC Call message payload that remain to be sent. The rdma_rpc_first_word field demarks the first word of the Payload stream. If a peer sends follows an RDMA2_CALL_MIDDLE header type with anything other than an RDMA2_CALL_MIDDLE message or an RDMA2_CALL_INLINE message, the receiver MUST respond with an RDMA2_ERROR header type and set its rdma_err field to RDMA2_ERR_INVAL_CONT and drop the incoming message without processing it further.

RDMA2_CALL_INLINE: Convey Inline RPC Call Message RDMA2_CALL_INLINE conveys the only or final portion of an RPC Call message. The rdma_rpc_first_word field demarks the first word of this Payload stream.

rdma_inv_handle:: The rdma_inv_handle field contains a 32-bit RDMA handle that the Responder may use in a Send With Invalidation operation. See .
rdma_reads:: The rdma_reads field anchors a list of zero or more Read segments that contain only data item chunks. A Requester MUST NOT insert Position-zero Read chunks in this list.
rdma_provisional_writes:: The rdma_writes field anchors a list of zero or more provisional Write chunks.
rdma_provisional_reply:: The rdma_reply field is a list containing zero or one provisional Reply chunk.

RDMA2_REPLY_EXTERNAL: Convey External RPC Reply Message RDMA2_REPLY_EXTERNAL conveys an RPC Reply message payload using explicit RDMA operations. In particular, it is referred to as a Special Format Reply when the Responder writes the RPC payload into a memory area specified by a Reply chunk. The sender MUST set the rdma_xid field to the same value as the xid of the RPC Reply message payload.

rdma_writes:: The rdma_writes field anchors a list of zero or more Write chunks that are either empty or contain reduced data items.
rdma_reply:: The rdma_reply field is a list that MUST contain exactly one Reply chunk.

RDMA2_REPLY_MIDDLE: Convey Continued RPC Reply Message RDMA2_REPLY_MIDDLE conveys a beginning or middle portion of an RPC Reply message immediately following the transport header in the send buffer. The sender MUST set the rdma_xid field to the same value as the xid of the RPC Reply message payload. The sender sets the rdma_remaining field to the number of bytes in the RPC Call message payload that remain to be sent. The rdma_rpc_first_word field demarks the first word of the Payload stream. If a peer sends follows an RDMA2_REPLY_MIDDLE header type with anything other than an RDMA2_REPLY_MIDDLE message or an RDMA2_REPLY_INLINE message, the receiver MUST respond with an RDMA2_ERROR header type and set its rdma_err field to RDMA2_ERR_INVAL_CONT and drop the incoming message without processing it further.

RDMA2_REPLY_INLINE: Convey RPC Reply Message Inline RDMA2_REPLY_INLINE conveys the only or final portion of an RPC Reply message immediately following the transport header in the send buffer. If the Reply message payload has been reduced, the rdma_chunks object carries the reduced data item chunks.

rdma_writes:: The rdma_writes field anchors a list of zero or more Write chunks that are either empty or contain reduced data items.

Transport Header Prefix The following prefix structure appears at the start of each RPC-over-RDMA version 2 transport header.

Remote Invalidation To solicit the use of Remote Invalidation, a Requester sets the value of the rdma_inv_handle field in an RPC Call's transport header to a non-zero value that matches one of the rdma_handle fields in that header. If the Responder may invalidate none of the rdma_handle values in the header conveying the Call, the Requester sets the RPC Call's rdma_inv_handle field to the value zero. If the Responder chooses not to use remote invalidation for this particular RPC Reply, or the RPC Call's rdma_inv_handle field contains the value zero, the Responder simply uses RDMA Send to transmit the matching RPC reply. However, if the Responder chooses to use Remote Invalidation, it uses RDMA Send With Invalidate to transmit the RPC Reply. It MUST use the value in the corresponding Call's rdma_inv_handle field to construct the Send With Invalidate Work Request. A Responder never uses a Send With Invalidate Work Request when sending a control plane header type. This includes the RDMA2_ERROR header type, the RDMA2_GRANT header type, the RDMA2_CONNPROP_MIDDLE header type, and the RDMA2_CONNPROP_FINAL header type.

Payload Formats RPC-over-RDMA version 2 provides several ways, known as "payload formats", to convey an RPC-over-RDMA message. A sender chooses the payload format for each message based on several factors:

The existence of DDP-eligible data items in the RPC message payload
The size of the RPC message payload
The direction of the RPC message (i.e., Call or Reply)
The available hardware resources
The arrangement of source and sink memory buffers

The following subsections describe in detail how Requesters and Responders format RPC-over-RDMA message payloads.

Simple Format All RPC messages conveyed via RPC-over-RDMA version 2 need at least one RDMA Send operation to convey. Thus, the most efficient way to send an RPC message that is smaller than the inline threshold is to append the Payload stream directly to the Transport stream and use an RDMA Send to convey both. When no chunks are present, senders construct Calls and Replies the same way, and no other operations are needed.

Simple Format with Data Item Chunks If DDP-eligible data items are present in a Payload stream, a sender MAY reduce some or all of these items, removing them from the Payload stream. The sender then uses a separate mechanism to transfer the reduced data items. The Transport stream immediately followed by the reduced Payload stream is then transferred using one RDMA Send operation. When data item chunks are present, senders construct Calls differently than Replies.

Simple Call: After receiving the Transport and Payload streams of an RPC Call message with Read chunks, the Responder uses RDMA Read operations to move the reduced data items contained in the Read chunks. RPC-over-RDMA Calls can carry Write chunks for the Responder to use when sending the matching Reply.
Simple Reply: The Responder uses RDMA Write operations to move reduced data items contained in Write chunks. Afterward, it sends the Transport and Payload streams of the RPC Reply message using one RDMA Send. RPC-over-RDMA Replies always carry an empty Read chunk list.

Simple Format Examples

A Simple Call with a Read chunk and a Simple Reply without data item chunks | | RDMA Read | | <---------------------------------- | | RDMA Response (arg data) | | ----------------------------------> | | | | | Processing | | | RDMA Send (RDMA2_REPLY_INLINE) | | <---------------------------------- | Reply ]]>

Continued Format For various reasons, a sender can choose to split a message payload over multiple RPC-over-RDMA messages. The Payload stream of each RPC-over-RDMA message contains a part of the RPC message. The receiver reconstructs the original RPC message by concatenating the Payload stream of each RPC-over-RDMA message in received order. A sender MAY split the Payload stream on any convenient boundary.

Continued Format with Data Item Chunks If DDP-eligible data items are present in the Payload stream, a sender MAY reduce some or all of these items, removing them from the Payload stream. The sender then uses a separate mechanism to transfer the reduced data items. The Transport stream immediately follwed by the reduced Payload stream is then transferred using one RDMA Send operation. As with Simple Format messages, when chunks are present, senders construct Calls differently than Replies.

Continued Call: After receiving the Transport and Payload streams of an RPC Call message with Read chunks, the Responder uses RDMA Read operations to move the reduced data items contained in Read chunks. RPC-over-RDMA Calls can carry Write chunks for the Responder to use when sending the matching Reply.
Continued Reply: The Responder uses RDMA Write operations to move reduced data items contained in Write chunks. Afterward, it sends the Transport and Payload streams of the RPC Reply message using multiple RDMA Sends. RPC-over-RDMA Replies always carry an empty Read chunk list.

Continued Format Examples

A Continued Call without data item chunks and a Continued Reply without data item chunks | | RDMA Send (RDMA2_CALL_MIDDLE) | | ----------------------------------> | | RDMA Send (RDMA2_CALL_INLINE) | | ----------------------------------> | | | | | Processing | | | RDMA Send (RDMA2_REPLY_MIDDLE) | | <---------------------------------- | Reply | RDMA Send (RDMA2_REPLY_MIDDLE) | | <---------------------------------- | | RDMA Send (RDMA2_REPLY_INLINE) | | <---------------------------------- | ]]>

A Continued Call with a Read chunk and a Simple Reply without data item chunks | | RDMA Send (RDMA2_CALL_MIDDLE) | | ----------------------------------> | | RDMA Send (RDMA2_CALL_INLINE) | | ----------------------------------> | | RDMA Read | | <---------------------------------- | | RDMA Response (arg data) | | ----------------------------------> | | | | | Processing | | | RDMA Send (RDMA2_REPLY_INLINE) | | <---------------------------------- | Reply ]]>

A Simple Call without data item chunks and a Continued Reply with a Write chunk | | | | | Processing | | | RDMA Write (result data) | | <---------------------------------- | | RDMA Send (RDMA2_REPLY_MIDDLE) | | <---------------------------------- | Reply | RDMA Send (RDMA2_REPLY_MIDDLE) | | <---------------------------------- | | RDMA Send (RDMA2_REPLY_INLINE) | | <---------------------------------- | ]]>

Special Format Even after DDP-eligible data items have been removed, a Payload stream can sometimes be too large to send using only RDMA Send operations. In those cases, the sender can use RDMA Read or Write operations to convey the entire RPC message. We refer to this as a "Special Format" message. To transmit a Special Format message, the sender transmits only the Transport stream with an RDMA Send operation. The sender does not include the Payload stream in the send buffer. Instead, the Requester provides a body chunk that the Responder uses to move the Payload stream. Because chunks are always present in Special Format messages, the sender always handles Calls and Replies differently.

Special Call: The Requester provides a Read chunk that contains the RPC Call message's Payload stream. Every Read segment in this chunk MUST contain zero (0) in its Position field. This type of Read chunk is a body chunk known as a Call chunk.
Special Reply: The Requester provisions a Reply chunk in advance. This body chunk is a Write chunk into which the Responder places the RPC Reply message's Payload stream. The Requester provisions the Reply chunk to accommodate the maximum expected reply size for that upper-layer operation.

One purpose of a Special Format message is to handle large RPC messages. However, Requesters MAY use a Special Format message at any time to convey an RPC Call message. When it has alternatives, a Responder chooses which Format to use based on the chunks provided by the Requester. If a Requester provided a Write chunk and the Responder has a DDP-eligible result, it first reduces the reply Payload stream. If a Requester provided a Reply chunk and the reduced Payload stream is larger than the reply inline threshold, the Responder MUST use the Requester-provided Reply chunk for the reply.

Special Format Examples

A Special Call and a Simple Reply without data item chunks | | RDMA Read | | <---------------------------------- | | RDMA Response (RPC call) | | ----------------------------------> | | | | | Processing | | | RDMA Send (RDMA2_REPLY_INLINE) | | <---------------------------------- | Reply ]]>

Choosing a Reply Payload Format A Requester provisions all necessary registered memory resources for both an RPC Call and its matching RPC Reply. A Requester constructs each RPC Call, thus it can compute the exact memory resources needed to send every Call. However, the Requester allocates memory resources to receive the corresponding Reply before the Responder has constructed it. Occasionally, it is challenging for the Requester to know in advance precisely what resources are needed to receive the Reply. In RPC-over-RDMA version 2, a Requester can provide a Reply chunk for any transaction. The Responder can use the provided Reply chunk or it can decide to use another means to convey the RPC Reply. If the combination of the provided Write chunk list and Reply chunk is not adequate to convey a Reply, the Responder SHOULD use Message Continuation to send that Reply. If even that is not possible, the Responder sends an RDMA2_ERROR message to the Requester, as described in :

If the Write chunk list cannot accommodate the ULP's DDP-eligible data payload, the Responder sends an RDMA2_ERR_WRITE_RESOURCE error.
If the Reply chunk cannot accommodate the parts of the Reply that are not DDP-eligible, the Responder sends an RDMA2_ERR_REPLY_RESOURCE error.

When receiving such errors, the Requester can retry the ULP call using more substantial reply resources. In cases where retrying the ULP request is not possible (e.g., the request is non-idempotent), the Requester terminates the RPC transaction and presents an error to the RPC consumer.

Error Handling A receiver performs validity checks on each ingress RPC-over-RDMA message before it assembles that message's Payload stream and passes it to the RPC layer. For example, if an ingress RPC-over-RDMA message is not as long as the size of struct rpcrdma2_hdr_prefix (20 octets), the receiver cannot trust the value of the rdma_xid field. In this case, the receiver MUST silently discard the ingress message without processing it further, and without a response to the sender. When a request (for instance, an RPC Call or a control plane operation) is made, typically an RPC consumer blocks while waiting for the response. Thus when an incoming message conveys a request and that request cannot be acted upon, the receiver of that request needs to report the problem to its sender in order to unblock waiters. Likewise, if, after processing a request, a sender is unable to transmit the response on an otherwise healthy connection, the sender needs to report that problem for the same reason. The RDMA2_ERROR header type is used for this purpose. To form an RDMA2_ERROR type header:

The rdma_xid field MUST contain the same XID that was in the rdma_xid field in the ingress request.
The rdma_vers field MUST contain the same version that was in the rdma_vers field in the ingress request.
The sender sets the rdma_credit field to the credit values in effect for this connection.
The rdma_htype field MUST contain the value RDMA2_ERROR.
The rdma_err field contains a value that reflects the type of error that occurred, as described in the subsections below.

When a peer receives an RDMA2_ERROR message type with an unrecognized or unsupported value in its rdma_err field, it MUST silently discard the message without processing it further.

Basic Transport Stream Parsing Errors

RDMA2_ERR_VERS When a Responder detects an RPC-over-RDMA header version that it does not support (the current document defines version 2), it MUST respond with an RDMA2_ERROR message type and set its rdma_err field to RDMA2_ERR_VERS. The Responder then fills in the rpcrdma2_err_vers structure with the RPC-over-RDMA versions it supports. The Responder MUST silently discard the ingress message without passing it to the RPC layer. When a Requester receives this error message, it uses the information in the rpcrdma2_err_vers structure to select an RPC-over-RDMA version that both peers support for subsequent operations on the connection. A Requester MUST NOT subsequently send a message that uses a version that the Responder has indicated it does not support. RDMA2_ERR_VERS indicates a permanent error. Receipt of this error completes the RPC transaction associated with XID in the rdma_xid field.

RDMA2_ERR_VERS_MISMATCH When a Responder receives a message with a transport protocol version that does not match the protocol version that was used in previous successful exchanges on the same connection, it MUST respond with an RDMA2_ERROR message type and set its rdma_err field to RDMA2_ERR_VERS_MISMATCH. The Responder MUST silently discard the ingress message without passing it to the RPC layer. A Requester MUST NOT subsequently send a message that uses a protocol version that the Responder has indicated it does not recognize on this connection. The Requester can recover by sending the message again using a corrected protocol version, or it can terminate the RPC transaction associated with the XID in the rdma_xid field with an error.

RDMA2_ERR_INVAL_HTYPE If a Responder recognizes the value in an ingress rdma_vers field, but it does not recognize the value in the rdma_htype field or does not support that header type, it MUST set the rdma_err field to RDMA2_ERR_INVAL_HTYPE. The Responder MUST silently discard the incoming message without passing it to the RPC layer. A Requester MUST NOT subsequently send a message on the connection that uses an htype that the Responder has indicated it does not support. RDMA2_ERR_INVAL_HTYPE indicates a permanent error. Receipt of this error completes the RPC transaction associated with XID in the rdma_xid field.

RDMA2_ERR_INVAL_CONT If a Responder detects a problem with an ingress RPC-over-RDMA message that is part of a Message Continuation sequence, the Responder MUST set the rdma_err field to RDMA2_ERR_INVAL_CONT. The Responder MUST silently discard all ingress messages with an rdma_xid field that matches the failing message without reassembling the payload. RDMA2_ERR_INVAL_CONT indicates a permanent error. Receipt of this error completes the RPC transaction associated with XID in the rdma_xid field.

XDR Errors A receiver might encounter an XDR parsing error that prevents it from processing an ingress Transport stream. Examples of such errors include:

The value of the rdma_xid field does not match the value of the XID field in the accompanying RPC message.
The receive buffer ends before the end of a data object contained in the Transport stream.

Moreover, when a Responder receives a valid RPC-over-RDMA header but the Responder's ULP implementation cannot parse the RPC arguments in the RPC Call, the Responder returns an RPC Reply with status GARBAGE_ARGS, using an RDMA2_REPLY_INLINE message type. This type of parsing failure might be due to mismatches between chunk sizes or offsets and the contents of the Payload stream, for example. In this case, the error is permanent, but the Requester has no way to know how much processing the Responder has completed for this RPC transaction.

RDMA2_ERR_BAD_XDR If a Responder recognizes the values in the rdma_vers field, but it cannot otherwise parse the ingress Transport stream, it MUST set the rdma_err field to RDMA2_ERR_BAD_XDR. The Responder MUST silently discard the ingress message without passing it to the RPC layer. RDMA2_ERR_BAD_XDR indicates a permanent error. Receipt of this error completes the RPC transaction associated with XID in the rdma_xid field.

RDMA2_ERR_BAD_PROPVAL If a receiver recognizes the value in an ingress rdma_which field, but it cannot parse the accompanying propval, it MUST set the rdma_err field to RDMA2_ERR_BAD_PROPVAL (see ). The receiver MUST silently discard the ingress message without applying any of its property settings.

Responder RDMA Operational Errors In RPC-over-RDMA version 2, the Responder initiates RDMA Read and Write operations that target the Requester's memory. Problems might arise as the Responder attempts to use Requester-provided resources for RDMA operations. For example:

Usually, chunks can be validated only by using their contents to perform data transfers. If chunk contents are invalid (e.g., a memory region is no longer registered or a chunk length exceeds the end of the registered memory region), a Remote Access Error occurs.
If a Requester's Receive buffer is too small, the Responder's Send operation completes with a Local Length Error.
If the Requester-provided Reply chunk is too small to accommodate a large RPC Reply message, a Remote Access Error occurs. A Responder might detect this problem before attempting to write past the end of the Reply chunk.

RDMA operational errors can be fatal to the connection. To avoid a retransmission loop and repeated connection loss that deadlocks the connection, once the Requester has re-established a connection, the Responder SHOULD send an RDMA2_ERROR response to indicate that no RPC-level reply is possible for that transaction.

RDMA2_ERR_READ_CHUNKS If a Requester presents more DDP-eligible arguments than a Responder is prepared to Read, the Responder MUST set the rdma_err field to RDMA2_ERR_READ_CHUNKS and set the rdma_max_chunks field to the maximum number of Read chunks the Responder can process. If the Responder implementation cannot handle any Read chunks for a request, it MUST set the rdma_max_chunks to zero in this response. The Responder MUST silently discard the ingress message without processing it further. The Requester can reconstruct the Call using Message Continuation or a Special Format payload and resend it. If the Requester chooses not to resend the Call, it MUST terminate this RPC transaction with an error.

RDMA2_ERR_WRITE_CHUNKS If a Requester has constructed an RPC Call with more DDP-eligible results than the Responder is prepared to Write, the Responder MUST set the rdma_err field to RDMA2_ERR_WRITE_CHUNKS and set the rdma_max_chunks field to the maximum number of Write chunks the Responder can return. The Requester can reconstruct the Call with no Write chunks and a Reply chunk of appropriate size. If the Requester does not resend the Call, it MUST terminate this RPC transaction with an error. If the Responder implementation cannot handle any Write chunks for a request and cannot send the Reply using Message Continuation, it MUST return a response of RDMA2_ERR_REPLY_RESOURCE instead (see below).

RDMA2_ERR_SEGMENTS If a Requester has constructed an RPC Call with a chunk that contains more segments than the Responder supports, the Responder MUST set the rdma_err field to RDMA2_ERR_SEGMENTS and set the rdma_max_segments field to the maximum number of segments the Responder can process. The Requester can reconstruct the Call and resend it. If the Requester does not resend the Call, it MUST terminate this RPC transaction with an error.

RDMA2_ERR_WRITE_RESOURCE If a Requester has provided a Write chunk that is not large enough to contain a DDP-eligible result, the Responder MUST set the rdma_err field to RDMA2_ERR_WRITE_RESOURCE. The Responder MUST set the rdma_chunk_index field to point to the first Write chunk in the transport header that is too short, or to zero to indicate that it was not possible to determine which chunk is too small. Indexing starts at one (1), which represents the first Write chunk. The Responder MUST set the rdma_length_needed to the number of bytes needed in that chunk to convey the result data item. The Requester can reconstruct the Call with more reply resources and resend it. If the Requester does not resend the Call (for instance, if the Responder set the index and length fields to zero), it MUST terminate this RPC transaction with an error.

RDMA2_ERR_REPLY_RESOURCE If a Responder cannot send an RPC Reply using Message Continuation and the Reply does not fit in the Reply chunk, the Responder MUST set the rdma_err field to RDMA2_ERR_REPLY_RESOURCE. The Responder MUST set the rdma_length_needed to the number of Reply chunk bytes needed to convey the reply. The Requester can reconstruct the Call with more reply resources and resend it. If the Requester does not resend the Call (for instance, if the Responder set the length field to zero), it MUST terminate this RPC transaction with an error.

Other Operational Errors While a Requester is constructing an RPC Call message, an unrecoverable problem might occur that prevents the Requester from posting further RDMA Work Requests on behalf of that message. As with other transports, if a Requester is unable to construct and transmit an RPC Call, the associated RPC transaction fails immediately. After a Requester has received a Reply, if it is unable to invalidate a memory region due to an unrecoverable problem, the Requester MUST close the connection to protect that memory from Responder access before the associated RPC transaction is complete. While a Responder is constructing an RPC Reply message or error message, an unrecoverable problem might occur that prevents the Responder from posting further RDMA Work Requests on behalf of that message. If a Responder is unable to construct and transmit an RPC Reply or RPC-over-RDMA error message, the Responder MUST close the connection to signal to the Requester that a reply was lost.

RDMA2_ERR_SYSTEM If some problem occurs on a Responder that does not fit into the above categories, the Responder MAY report it to the Requester by setting the rdma_err field to RDMA2_ERR_SYSTEM. The Responder MUST silently discard the message(s) associated with the failing transaction without further processing. RDMA2_ERR_SYSTEM is a permanent error. This error does not indicate how much of the transaction the Responder has processed, nor does it indicate a particular recovery action for the Requester. A Requester that receives this error MUST terminate the RPC transaction associated with the XID value in the RDMA2_ERROR message's rdma_xid field.

RDMA Transport Errors The RDMA connection and physical link provide some degree of error detection and retransmission. The Marker PDU Aligned Framing (MPA) protocol (as described in ) as well as the InfiniBand link layer provide Cyclic Redundancy Check (CRC) protection of RDMA payloads. CRC-class protection is a general attribute of such transports. Additionally, the RPC layer itself can accept errors from the transport and recover via retransmission. RPC recovery can typically handle complete loss and re-establishment of a transport connection. The details of reporting and recovery from RDMA link-layer errors are described in specific link-layer APIs and operational specifications and are outside the scope of this protocol specification. See for further discussion of RPC-level integrity schemes.

XDR Protocol Definition This section contains a description of the core features of the RPC-over-RDMA version 2 protocol expressed in the XDR language . It organizes the description to make it simple to extract into a form that is ready to compile or combine with similar descriptions published later as extensions to RPC-over-RDMA version 2.

Code Component License Code Components extracted from the current document must include the following license text. When combining the extracted XDR code with other XDR code which has an identical license, only a single copy of the license text needs to be retained.

Extraction of the XDR Definition Implementers can apply the following sed script to the current document to produce a machine-readable XDR description of the base RPC-over-RDMA version 2 protocol. That is, if this document is in a file called "spec.txt", then implementers can do the following to extract an XDR description file and store it in the file rpcrdma-v2.x. rpcrdma-v2.x ]]> Although this file is a usable description of the base protocol, when extensions are to be supported, it may be desirable to divide the description into multiple files. The following script achieves that purpose: temp.x"); while() { if (m/FILE ENDS: (.*)$/) { close(OUT); rename("temp.x", $1); open(OUT,">temp.x"); } else { print OUT $_; } } close(IN); close(OUT); ]]> Running the above script results in two files:

The file common.x, containing the license plus the shared XDR definitions that need to be made available to both the base protocol and any subsequent extensions.
The file baseops.x containing the XDR definitions for the base protocol defined in this document.

Extensions to RPC-over-RDMA version 2, published as Standards Track documents, should have similarly structured XDR definitions. Once an implementer has extracted the XDR for all desired extensions and the base XDR definition contained in the current document, she can concatenate them to produce a consolidated XDR definition that reflects the set of extensions selected for her RPC-over-RDMA version 2 implementation. Alternatively, the XDR descriptions can be compiled separately. In that case, the combination of common.x and baseops.x defines the base transport. The combination of common.x and the XDR description of each extension produces a full XDR definition of that extension.

XDR Definition for RPC-over-RDMA Version 2 Core Structures ; /// }; /// /// struct rpcrdma2_write_list { /// struct rpcrdma2_write_chunk rdma_entry; /// struct rpcrdma2_write_list *rdma_next; /// }; /// /// /*************************************************************** /// * Transport Properties /// ***************************************************************/ /// /// /* /// * Types for transport properties model /// */ /// typedef rpcrdma2_propid uint32; /// /// struct rpcrdma2_propval { /// rpcrdma2_propid rdma_which; /// opaque rdma_data<>; /// }; /// /// typedef rpcrdma2_propval rpcrdma2_propset<>; /// typedef uint32 rpcrdma2_propsubset<>; /// /// /* /// * Transport propid values for basic properties /// */ /// const RDMA2_PROPID_SBSIZ = 1; /// const RDMA2_PROPID_RBSIZ = 2; /// const RDMA2_PROPID_RSSIZ = 3; /// const RDMA2_PROPID_RCSIZ = 4; /// const RDMA2_PROPID_BRS = 5; /// const RDMA2_PROPID_HOSTAUTH = 6; /// /// /* /// * Types specific to particular properties /// */ /// typedef uint32 rpcrdma2_prop_sbsiz; /// typedef uint32 rpcrdma2_prop_rbsiz; /// typedef uint32 rpcrdma2_prop_rssiz; /// typedef uint32 rpcrdma2_prop_rcsiz; /// typedef uint32 rpcrdma2_prop_brs; /// typedef opaque rpcrdma2_prop_hostauth<>; /// /// const RDMA2_RVRSDIR_NONE = 0; /// const RDMA2_RVRSDIR_SIMPLE = 1; /// const RDMA2_RVRSDIR_CONT = 2; /// const RDMA2_RVRSDIR_GENL = 3; /// /// /* FILE ENDS: common.x; */ ]]>

XDR Definition for RPC-over-RDMA Version 2 Base Header Types

Use of the XDR Description The files common.x and baseops.x, when combined with the XDR descriptions for extension defined later, produce a human-readable and compilable description of the RPC-over-RDMA version 2 protocol with the included extensions. Although this XDR description can generate encoders and decoders for the Transport and Payload streams, there are elements of the operation of RPC-over-RDMA version 2 that cannot be expressed within the XDR language. Implementations that use the output of an automated XDR processor need to provide additional code to bridge these gaps.

The Transport stream is not a single XDR object. Instead, the header prefix is one XDR data item, and the rest of the header is a separate XDR data item. expresses the mapping between the header type in the header prefix and the XDR object representing the header type.
The relationship between the Transport stream and the Payload stream is not specified using XDR. Comments within the XDR text make clear where transported messages, described by their own XDR definitions, need to appear. Such data is opaque to the transport.
Continuation of RPC messages across transport message boundaries requires that message assembly facilities not specifiable within XDR are part of transport implementations.
Transport properties are constant integer values. expresses the mapping between each property's code point and the XDR typedef that represents the structure of the property's value. XDR does not possess the facility to express that mapping in an extensible way.

The role of XDR in RPC-over-RDMA specifications is more limited than for protocols where the totality of the protocol is expressible within XDR. XDR lacks the facility to represent the embedding of XDR-encoded payload material. Also, the need to cleanly accommodate extensions has meant that those using rpcgen in their applications need to take an active role to provide the facilities that cannot be expressed within XDR.

RPC Bind Parameters Before establishing a new connection, an RPC client obtains a transport address for the RPC server. The means used to obtain this address and to open an RDMA connection is dependent on the type of RDMA transport and is the responsibility of each RPC protocol binding and its local implementation. RPC services typically register with a portmap or rpcbind service , which associates an RPC Program number with a service address. This policy is no different with RDMA transports. However, a distinct service address (port number) is sometimes required for operation on RPC-over-RDMA. When mapped atop MPA , which uses IP port addressing due to its layering on TCP or SCTP, port mapping is trivial and consists merely of issuing the port in the connection process. The NFS/RDMA protocol service address has been assigned port 20049 by IANA for this deployment scenario . When mapped atop InfiniBand , which uses a service endpoint naming scheme based on a Group Identifier (GID), a translation MUST be employed. One such translation is described in Annexes A3 (Application Specific Identifiers), A4 (Sockets Direct Protocol (SDP)), and A11 (RDMA IP CM Service) of , which is appropriate for translating IP port addressing to the InfiniBand network. Therefore, in this case, IP port addressing may be readily employed by the upper layer. When a mapping standard or convention exists for IP ports on an RDMA interconnect, there are several possibilities for each upper layer to consider:

One possibility is to have the server register its mapped IP port with the rpcbind service under the netid (or netids) defined in . An RPC-over-RDMA-aware RPC client can then resolve its desired service to a mappable port and proceed to connect. This method is the most flexible and compatible approach for those upper layers that are defined to use the rpcbind service.
A second possibility is to have the RPC server's portmapper register itself on the RDMA interconnect at a "well-known" service address (on UDP or TCP, this corresponds to port 111). An RPC client can connect to this service address and use the portmap protocol to obtain a service address in response to a program number (e.g., a TCP port number or an InfiniBand GID).
Alternately, an RPC client can connect to the mapped well-known port for the service itself, if it is appropriately defined. By convention, the NFS/RDMA service, when operating atop an InfiniBand fabric, uses the same 20049 assignment as for MPA.

Historically, different RPC protocols have taken different approaches to their port assignments. The current document leaves the specific method for each RPC-over-RDMA-enabled ULB. defines two new netid values to be used for registration of upper layers atop MPA and (when a suitable port translation service is available) InfiniBand. Additional RDMA-capable networks MAY define their own netids, or if they provide a port translation, they MAY share the one defined in .

Implementation Status This section records the status of known implementations of the protocol defined by this specification at the time of posting of this Internet-Draft, and is based on a proposal described in . The description of implementations in this section is intended to assist the IETF in its decision processes in progressing drafts to RFCs. Please note that the listing of any individual implementation here does not imply endorsement by the IETF. Furthermore, no effort has been spent to verify the information presented here that was supplied by IETF contributors. This is not intended as, and must not be construed to be, a catalog of available implementations or their features. Readers are advised to note that other implementations may exist. At this time, no known implementations of the protocol described in the current document exist.

Security Considerations

Memory Protection A primary consideration is the protection of the integrity and confidentiality of host memory by an RPC-over-RDMA transport. The use of an RPC-over-RDMA transport protocol MUST NOT introduce vulnerabilities to system memory contents nor memory owned by user processes. Any RDMA provider used for RPC transport MUST conform to the requirements of to satisfy these protections.

Protection Domains The use of a Protection Domain to limit the exposure of memory regions to a single connection is critical. Any attempt by an endpoint not participating in that connection to reuse memory handles needs to result in immediate failure of that connection. Because ULP security mechanisms rely on this aspect of Reliable Connected behavior, implementations SHOULD cryptographically authenticate connection endpoints.

Handle (STag) Predictability Implementations should use unpredictable memory handles for any operation requiring exposed memory regions. Exposing a continuously registered memory region allows a remote host to read or write to that region even when an RPC involving that memory is not underway. Therefore, implementations should avoid the use of persistently registered memory.

Memory Protection Requesters should register memory regions for remote access only when they are about to be the target of an RPC transaction that involves an RDMA Read or Write. Requesters should invalidate memory regions as soon as related RPC operations are complete. Invalidation and DMA unmapping of memory regions should complete before the receiver checks message integrity, and before the RPC consumer can use or alter the contents of the exposed memory region. An RPC transaction on a Requester can terminate before a Reply arrives, for example, if the RPC consumer is signaled, or a segmentation fault occurs. When an RPC terminates abnormally, memory regions associated with that RPC should be invalidated before the Requester reuses those regions for other purposes.

Denial of Service A detailed discussion of denial-of-service exposures that can result from the use of an RDMA transport appears in . A Responder is not obliged to pull unreasonably large Read chunks. A Responder can use an RDMA2_ERROR response to terminate RPCs with unreadable Read chunks. If a Responder transmits more data than a Requester is prepared to receive in a Write or Reply chunk, the RDMA provider typically terminates the connection. For further discussion, see . Such repeated connection termination can deny service to other users sharing the connection from the errant Requester. An RPC-over-RDMA transport implementation is not responsible for throttling the RPC request rate, other than to keep the number of concurrent RPC transactions within the per connection credit limits (see ). A sender can trigger a self-denial of service by exceeding the credit limit repeatedly. When an RPC transaction terminates due to a signal or premature exit of an application process, a Requester should invalidate the RPC's Write and Reply chunks. Invalidation prevents the subsequent arrival of the Responder's Reply from altering the memory regions associated with those chunks after the Requester has released that memory. On the Requester, a malfunctioning application or a malicious user can create a situation where RPCs initiate and abort continuously, resulting in Responder replies that terminate the underlying RPC-over-RDMA connection repeatedly. Such situations can deny service to other users sharing the connection from that Requester.

RPC Message Security ONC RPC provides cryptographic security via the RPCSEC_GSS framework . RPCSEC_GSS implements message authentication (rpc_gss_svc_none), per-message integrity checking (rpc_gss_svc_integrity), and per-message confidentiality (rpc_gss_svc_privacy) in a layer above the RPC-over-RDMA transport. The integrity and privacy services require significant computation and movement of data on each endpoint host. Some performance benefits enabled by RDMA transports can be lost.

RPC-over-RDMA Protection at Other Layers For any RPC transport, utilizing RPCSEC_GSS integrity or privacy services has performance implications. Protection below the RPC implementation is often a better choice in performance-sensitive deployments, especially if it, too, can be offloaded. Certain implementations of IPsec can be co-located in RDMA hardware, for example, without change to RDMA consumers and with little loss of data movement efficiency. Such arrangements can also provide a higher degree of privacy by hiding endpoint identity or altering the frequency at which messages are exchanged, at a performance cost. Implementations MAY negotiate the use of protection in another layer through the use of an RPCSEC_GSS security flavor defined in in conjunction with the Channel Binding mechanism and IPsec Channel Connection Latching .

RPCSEC_GSS on RPC-over-RDMA Transports Not all RDMA devices and fabrics support the above protection mechanisms. Also, NFS clients, where multiple users can access NFS files, still require per-message authentication. In these cases, RPCSEC_GSS can protect NFS traffic conveyed on RPC-over-RDMA connections. RPCSEC_GSS extends the ONC RPC protocol without changing the format of RPC messages. By observing the conventions described in this section, an RPC-over-RDMA transport can convey RPCSEC_GSS-protected RPC messages interoperably. Senders MUST NOT reduce protocol elements of RPCSEC_GSS that appear in the Payload stream of an RPC-over-RDMA message. Such elements include control messages exchanged as part of establishing or destroying a security context, or data items that are part of RPCSEC_GSS authentication material.

RPCSEC_GSS Context Negotiation Some NFS client implementations use a separate connection to establish a Generic Security Service (GSS) context for NFS operation. Such clients use TCP and the standard NFS port (2049) for context establishment. Therefore, an NFS server MUST also provide a TCP-based NFS service on port 2049 to enable the use of RPCSEC_GSS with NFS/RDMA.

RPC-over-RDMA with RPCSEC_GSS Authentication The RPCSEC_GSS authentication service has no impact on the DDP-eligibility of data items in a ULP. However, RPCSEC_GSS authentication material appearing in an RPC message header can be larger than, say, an AUTH_SYS authenticator. In particular, when an RPCSEC_GSS pseudoflavor is in use, a Requester needs to accommodate a larger RPC credential when marshaling RPC Calls and needs to provide for a maximum size RPCSEC_GSS verifier when allocating reply buffers and Reply chunks. RPC messages, and thus Payload streams, are larger on average as a result. ULP operations that fit in a Simple Format message when a simpler form of authentication is in use might need to be reduced or conveyed via a Special Format message when RPCSEC_GSS authentication is in use. It is therefore more likely that a Requester provisions both a Read list and a Reply chunk in the same RPC-over-RDMA Transport header to convey a Special Format Call and provision a receptacle for a Special Format Reply. In addition to this cost, the XDR encoding and decoding of each RPC message using RPCSEC_GSS authentication requires per-message host compute resources to construct the GSS verifier.

RPC-over-RDMA with RPCSEC_GSS Integrity or Privacy The RPCSEC_GSS integrity service enables endpoints to detect the modification of RPC messages in flight. The RPCSEC_GSS privacy service prevents all but the intended recipient from viewing the cleartext content of RPC arguments and results. RPCSEC_GSS integrity and privacy services are end-to-end. They protect RPC arguments and results from application to server endpoint, and back. The RPCSEC_GSS integrity and encryption services operate on whole RPC messages after they have been XDR encoded, and before they have been XDR decoded after receipt. Connection endpoints use intermediate buffers to prevent exposure of encrypted or unverified cleartext data to RPC consumers. After a sender has verified, encrypted, and wrapped a message, the transport layer MAY use RDMA data transfer between these intermediate buffers. The process of reducing a DDP-eligible data item removes the data item and its XDR padding from an encoded Payload stream. In a non-protected RPC-over-RDMA message, a reduced data item does not include XDR padding. After reduction, the Payload stream contains fewer octets than the whole XDR stream did beforehand. XDR padding octets are often zero bytes, but they don't have to be. Thus, reducing DDP-eligible items affects the result of message integrity verification and encryption. Therefore, a sender MUST NOT reduce a Payload stream when RPCSEC_GSS integrity or encryption services are in use. Effectively, no data item is DDP-eligible in this situation. Senders can use only Simple and Continued Formats without data item chunks, or Special Format. In this mode, an RPC-over-RDMA transport operates in the same manner as a transport that does not support DDP.

Protecting RPC-over-RDMA Transport Headers Like the header fields in an RPC message (e.g., the xid and mtype fields), RPCSEC_GSS does not protect the RPC-over-RDMA Transport stream. XIDs, connection credit limits, and chunk lists (though not the content of the data items they refer to) are exposed to malicious behavior, which can redirect data that is transferred by the RPC-over-RDMA message, result in spurious retransmits, or trigger connection loss. In particular, if an attacker alters the information contained in the chunk lists of an RPC-over-RDMA Transport header, data contained in those chunks can be redirected to other registered memory regions on Requesters. An attacker might alter the arguments of RDMA Read and RDMA Write operations on the wire to gain a similar effect. If such alterations occur, the use of RPCSEC_GSS integrity or privacy services enables a Requester to detect unexpected material in a received RPC message. Encryption at other layers, as described in , protects the content of the Transport stream. RDMA transport implementations should conform to to address attacks on RDMA protocols themselves.

Transport Properties Like other fields that appear in the Transport stream, transport properties are sent in the clear with no integrity protection, making them vulnerable to man-in-the-middle attacks. For example, if a man-in-the-middle were to change the value of the Receive buffer size, it could reduce connection performance or trigger loss of connection. Repeated connection loss can impact performance or even prevent a new connection from being established. The recourse is to deploy on a private network or use transport layer encryption.

Host Authentication [ cel: This subsection is unfinished. ] Wherein we use the relevant sections of to analyze the addition of host authentication to this RPC-over-RDMA transport. The authors refer readers to Appendix C of for information on how to design and test a secure authentication handshake implementation.

IANA Considerations The RPC-over-RDMA family of transports have been assigned RPC netids by . A netid is an rpcbind string used to identify the underlying protocol in order for RPC to select appropriate transport framing and the format of the service addresses and ports. The following netid registry strings are already defined for this purpose: The "rdma" netid is to be used when IPv4 addressing is employed by the underlying transport, and "rdma6" when IPv6 addressing is employed. The netid assignment policy and registry are defined in . The current document does not alter these netid assignments. These netids MAY be used for any RDMA network that satisfies the requirements of and that is able to identify service endpoints using IP port addressing, possibly through use of a translation service as described in .

References Normative References Informative References InfiniBand Architecture Specification Volume 1 InfiniBand Trade Association Available from https://www.infinibandta.org/ Credit-Based Flow Control for ATM Networks: Credit Update Protocol, Adaptive Credit Allocation, and Statistical Multiplexing Division of Applied Sciences, Harvard University

29 Oxford Street Cambridge MA 02138 United States of America

Division of Applied Sciences, Harvard University

29 Oxford Street Cambridge MA 02138 United States of America

Bell-Northern Research

P.O.Box 3511, Station C Ottawa Ontario KIY 4H7 Canada

ULB Specifications Typically, an Upper-Layer Protocol (ULP) is defined without regard to a particular RPC transport. An Upper-Layer Binding (ULB) specification provides guidance that helps a ULP interoperate correctly and efficiently over a particular transport. For RPC-over-RDMA version 2, a ULB may provide:

A taxonomy of XDR data items that are eligible for DDP
Constraints on which upper-layer procedures a sender may reduce, and on how many chunks may appear in a single RPC message
A method enabling a Requester to determine the maximum size of the reply Payload stream for all procedures in the ULP
An rpcbind port assignment for the RPC Program and Version when operating on the particular transport

Each RPC Program and Version tuple that operates on RPC-over-RDMA version 2 needs to have a ULB specification.

DDP-Eligibility A ULB designates specific XDR data items as eligible for DDP. As a sender constructs an RPC-over-RDMA message, it can remove DDP-eligible data items from the Payload stream so that the RDMA provider can place them directly in the receiver's memory. An XDR data item should be considered for DDP-eligibility if there is a clear benefit to moving the contents of the item directly from the sender's memory to the receiver's memory. Criteria for DDP-eligibility include:

The XDR data item is frequently sent or received, and its size is often much larger than typical inline thresholds.
If the XDR data item is a result, its maximum size must be predictable in advance by the Requester.
Transport-level processing of the XDR data item is not needed. For example, the data item is an opaque byte array, which requires no XDR encoding and decoding of its content.
The content of the XDR data item is sensitive to address alignment. For example, a data copy operation would be required on the receiver to enable the message to be parsed correctly, or to enable the data item to be accessed.
The XDR data item itself does not contain DDP-eligible data items.

In addition to defining the set of data items that are DDP-eligible, a ULB may limit the use of chunks to particular upper-layer procedures. If more than one data item in a procedure is DDP-eligible, the ULB may limit the number of chunks that a Requester can provide for a particular upper-layer procedure. Senders never reduce data items that are not DDP-eligible. Such data items can, however, be part of a Special Format payload. The programming interface by which an upper-layer implementation indicates the DDP-eligibility of a data item to the RPC transport is not described by this specification. The only requirements are that the receiver can re-assemble the transmitted RPC-over-RDMA message into a valid XDR stream and that DDP-eligibility rules specified by the ULB are respected. There is no provision to express DDP-eligibility within the XDR language. The only definitive specification of DDP-eligibility is a ULB. In general, a DDP-eligibility violation occurs when:

A Requester reduces a non-DDP-eligible argument data item. The Responder reports the violation as described in .
A Responder reduces a non-DDP-eligible result data item. The Requester terminates the pending RPC transaction and reports an appropriate permanent error to the RPC consumer.
A Responder does not reduce a DDP-eligible result data item into an available Write chunk. The Requester terminates the pending RPC transaction and reports an appropriate permanent error to the RPC consumer.

Maximum Reply Size When expecting small and moderately-sized Replies, a Requester should rely on Message Continuation rather than provision a Reply chunk. For each ULP procedure where there is no clear Reply size maximum and the maximum can be substantial, the ULB should specify a dependable means for determining the maximum Reply size.

Reverse-Direction Operation The direction of operation does not preclude the need for DDP-eligibility statements. Reverse-direction operation occurs on an already-established connection. Specification of RPC binding parameters is usually not necessary in this case. Other considerations may apply when distinct RPC Programs share an RPC-over-RDMA transport connection concurrently.

Additional Considerations There may be other details provided in a ULB.

A ULB may recommend inline threshold values or other transport-related parameters for RPC-over-RDMA version 2 connections bearing that ULP.
A ULP may provide a means to communicate transport-related parameters between peers.
Multiple ULPs may share a single RPC-over-RDMA version 2 connection when their ULBs allow the use of RPC-over-RDMA version 2 and the rpcbind port assignments for those protocols permit connection sharing. In this case, the same transport parameters (such as inline threshold) apply to all ULPs using that connection.

Each ULB needs to be designed to allow correct interoperation without regard to the transport parameters actually in use. Furthermore, implementations of ULPs must be designed to interoperate correctly regardless of the connection parameters in effect on a connection.

ULP Extensions An RPC Program and Version tuple may be extensible. For instance, the RPC version number may not reflect a ULP minor versioning scheme, or the ULP may allow the specification of additional features after the publication of the original RPC Program specification. ULBs are provided for interoperable RPC Programs and Versions by extending existing ULBs to reflect the changes made necessary by each addition to the existing XDR. [ cel: The final sentence is unclear, and may be inaccurate. I believe I copied this section directly from RFC 8166. Is there more to be said, now that we have some experience? ]

Extending RPC-over-RDMA Version 2 This Appendix is not addressed to protocol implementers, but rather to authors of documents that extend the protocol specified in the current document. RPC-over-RDMA version 2 extensibility facilitates limited extensions to the base protocol presented in the current document so that new optional capabilities can be introduced without a protocol version change while maintaining robust interoperability with existing RPC-over-RDMA version 2 implementations. It allows extensions to be defined, including the definition of new protocol elements, without requiring modification or recompilation of the XDR for the base protocol. Standards Track documents may introduce extensions to the base RPC-over-RDMA version 2 protocol in two ways:

They may introduce new OPTIONAL transport header types. covers such transport header types.
They may define new OPTIONAL transport properties. describes such transport properties.

These documents may also add the following sorts of ancillary protocol elements to the protocol to support the addition of new transport properties and header types:

They may create new error codes, as described in .

New capabilities can be proposed and developed independently of each other. Implementers can choose among them, making it straightforward to create and document experimental features and then bring them through the standards process.

Documentation Requirements As described earlier, a Standards Track document introduces a set of new protocol elements. Together these elements are considered an OPTIONAL feature. Each implementation is either aware of all the protocol elements introduced by that feature or is aware of none of them. Documents specifying extensions to RPC-over-RDMA version 2 should contain:

An explanation of the purpose and use of each new protocol element.
An XDR description including all of the new protocol elements, and a script to extract it.
A discussion of interactions with other extensions. This discussion includes requirements for other OPTIONAL features to be present, or that a particular level of support for an OPTIONAL facility is required.

Implementers combine the XDR descriptions of the new features they intend to use with the XDR description of the base protocol in the current document. This combination is necessary to create a valid XDR input file because extensions are free to use XDR types defined in the base protocol, and later extensions may use types defined by earlier extensions. The XDR description for the RPC-over-RDMA version 2 base protocol combined with that for any selected extensions should provide a human-readable and compilable definition of the extended protocol.

Adding New Header Types to RPC-over-RDMA Version 2 New transport header types are defined similar to Sections through . In particular, what is needed is:

A description of the function and use of the new header type.
A complete XDR description of the new header type.
A description of how receivers report errors, including mechanisms for reporting errors outside the available choices already available in the base protocol or other extensions.
An indication of whether a Payload stream must be present, and a description of its contents and how receivers use such Payload streams to reconstruct RPC messages.
As appropriate, a statement of whether a Responder may use Remote Invalidation when sending messages that contain the new header type.

There needs to be additional documentation that is made necessary due to the OPTIONAL status of new transport header types:

The document should discuss constraints on support for the new header types. For example, if support for one header type is implied or foreclosed by another one, this needs to be documented.
The document should describe the preferred method by which a sender determines whether its peer supports a particular header type. It is always possible to send a test invocation of a particular header type to see if support is available. However, when more efficient means are available (e.g., the value of a transport property), this should be noted.

Adding New Transport properties to the Protocol A Standards Track document defining a new transport property should include the following information paralleling that provided in this document for the transport properties defined herein:

The rpcrdma2_propid value identifying the new property.
The XDR typedef specifying the structure of its property value.
A description of the new property.
An explanation of how the receiver can use this information.
The default value if a peer never receives the new property.

There is no requirement that propid assignments occur in a continuous range of values. Implementations should not rely on all such values being small integers. Before the defining Standards Track document is published, the nfsv4 Working Group should select a unique propid value, and ensure that:

rpcrdma2_propid values specified in the document do not conflict with those currently assigned or in use by other pending working group documents defining transport properties.
rpcrdma2_propid values specified in the document do not conflict with the range reserved for experimental use, as defined in Section 8.2. [ cel: There is no longer a section 8.2 or an experimental range of propid values. Should we request the creation of an IANA registry for propid values? ].

When a Standards Track document proposes additional transport properties, reviewers should deal with possible security issues exposed by those new transport properties.

Adding New Error Codes to the Protocol The same Standards Track document that defines a new header type may introduce new error codes used to support it. A Standards Track document may similarly define new error codes that an existing header type can return. For error codes that do not require the return of additional information, a peer can use the existing RDMA_ERR2 header type to report the new error. The sender sets the new error code as the value of rdma_err with the result that the default switch arm of the rpcrdma2_error (i.e., void) is selected. For error codes that do require the return of related information together with the error, a new header type should be defined that returns the error together with the related information. The sender of a new header type needs to be prepared to accept header types necessary to report associated errors.

Differences from RPC-over-RDMA Version 1 The primary goal of RPC-over-RDMA version 2 is to relieve constraints that have become evident in RPC-over-RDMA version 1 with deployment experience:

RPC-over-RDMA version 1 has been challenging to update to address shortcomings or improve data transfer efficiency.
The average size of NFSv4 COMPOUNDs is significantly greater than NFSv3 requests, requiring the use of Long messages for frequent operations.
Reply size estimation is difficult more often than first expected.

This section details specific changes in RPC-over-RDMA version 2 that address these constraints directly, in addition to other changes to make implementation easier.

Changes to the XDR Definition Several XDR structural changes enable within-version protocol extensibility. defines the RPC-over-RDMA version 1 transport header as a single XDR object, with an RPC message potentially following it. In RPC-over-RDMA version 2, there are separate XDR definitions of the transport header prefix (see ), which specifies the transport header type to be used, and the transport header itself (defined within one of the subsections of ). This construction is similar to an RPC message, which consists of an RPC header (defined in ) followed by a message defined by an Upper-Layer Protocol. As a new version of the RPC-over-RDMA transport protocol, RPC-over-RDMA version 2 exists within the versioning rules defined in . In particular, it maintains the first four words of the protocol header, as specified in , even though, as explained in of the current document, the XDR definition of those words is structured differently. Although each of the first four fields retains its semantic function, there are differences in interpretation:

The first word of the header, the rdma_xid field, retains the format and function that it had in RPC-over-RDMA version 1. Because RPC-over-RDMA version 2 messages can convey non-RPC messages, a receiver should not use the contents of this field without consideration of the protocol version and header type.
The second word of the header, the rdma_vers field, retains the format and function that it had in RPC-over-RDMA version 1. To clearly distinguish version 1 and version 2 messages, senders need to fill in the correct version (fixed after version negotiation). Receivers should check that the content of the rdma_vers is correct before using the content of any other header field.
The third word of the header, the rdma_credit field, retains the size and general purpose that it had in RPC-over-RDMA version 1. However, RPC-over-RDMA version 2 divides this field into two 16-bit subfields. See for further details.
The fourth word of the header, previously the union discriminator field rdma_proc, retains its format and general function even though the set of valid values has changed. Within RPC-over-RDMA version 2, this word is the rdma_htype field of the structure rdma_start. The value of this field is now an unsigned 32-bit integer rather than an enum type, to facilitate header type extension.

Beyond conforming to the restrictions specified in , RPC-over-RDMA version 2 attempts to limit the scope of the changes made to ensure interoperability. Although it introduces the Call chunk and splits the two version 1 workhorse procedure types RDMA_MSG and RDMA_NOMSG into several variants, RPC-over-RDMA version 2 otherwise expresses chunks in the same format and utilizes them the same way.

Transport Properties RPC-over-RDMA version 2 provides a mechanism for exchanging an implementation's operational properties. The purpose of this exchange is to help endpoints improve the efficiency of data transfer by exploiting the characteristics of both peers rather than falling back on the lowest common denominator default settings. A full discussion of transport properties appears in .

Credit Management Changes RPC-over-RDMA transports employ credit-based flow control to ensure that a Requester does not emit more RDMA Sends than the Responder is prepared to receive. explains the operation of RPC-over-RDMA version 1 credit management in detail. In that design, each RDMA Send from a Requester contains an RPC Call with a credit request, and each RDMA Send from a Responder contains an RPC Reply with a credit grant. The credit grant implies that enough Receives have been posted on the Responder to handle the credit grant minus the number of pending RPC transactions (the number of remaining Receive buffers might be zero). Each RPC Reply acts as an implicit ACK for a previous RPC Call from the Requester. Without an RPC Reply message, the Requester has no way to know that the Responder is ready for subsequent RPC Calls. Because version 1 embeds credit management in each message, there is a strict one-to-one ratio between RDMA Send and RPC message. There are interesting use cases that might be enabled if this relationship were more flexible:

RPC-over-RDMA operations that do not carry an RPC message, e.g., control plane operations.
A single RDMA Send that conveys more than one RPC message, e.g., for interrupt mitigation.
An RPC message that requires several sequential RDMA Sends, e.g., to reduce the use of explicit RDMA operations for moderate-sized RPC messages.
An RPC transaction that requires multiple exchanges or an odd number of RPC-over-RDMA operations to complete.

RPC-over-RDMA version 2 provides a more sophisticated credit accounting mechanism to address these shortcomings. explains the new mechanism in detail.

Inline Threshold Changes An "inline threshold" value is the largest message size (in octets) that can be conveyed on an RDMA connection using only RDMA Send and Receive. Each connection has two inline threshold values: one for messages flowing from client-to-server (referred to as the "client-to-server inline threshold") and one for messages flowing from server-to-client (referred to as the "server-to-client inline threshold"). A connection's inline thresholds determine, among other things, when RDMA Read or Write operations are required because an RPC message cannot be conveyed via a single RDMA Send and Receive pair. When an RPC message does not contain DDP-eligible data items, a Requester can prepare a Special Format Call or Reply to convey the whole RPC message using RDMA Read or Write operations. RDMA Read and Write operations require that data payloads reside in memory registered with the local RNIC. When an RPC completes, that memory is invalidated to fence it from the Responder. Memory registration and invalidation typically have a latency cost that is insignificant compared to data handling costs. When a data payload is small, however, the cost of registering and invalidating memory where the payload resides becomes a significant part of total RPC latency. Therefore the most efficient operation of an RPC-over-RDMA transport occurs when the peers use explicit RDMA Read and Write operations for large payloads but avoid those operations for small payloads. When the authors of first conceived RPC-over-RDMA version 1, the average size of RPC messages that did not involve a significant data payload was under 500 bytes. A 1024-byte inline threshold adequately minimized the frequency of inefficient Long messages. With NFS version 4 , the increased size of NFS COMPOUND operations resulted in RPC messages that are, on average, larger than previous versions of NFS. With a 1024-byte inline threshold, frequent operations such as GETATTR and LOOKUP require RDMA Read or Write operations, reducing the efficiency of data transport. To reduce the frequency of Special Format messages, RPC-over-RDMA version 2 increases the default size of inline thresholds. This change also increases the maximum size of reverse-direction RPC messages.

Message Continuation Changes In addition to a larger default inline threshold, RPC-over-RDMA version 2 introduces Message Continuation. Message Continuation is a mechanism that enables the transmission of a data payload using more than one RDMA Send. The purpose of Message Continuation is to provide relief in several essential cases:

If a Requester finds that it is inefficient to convey a moderately-sized data payload using Read chunks, the Requester can use Message Continuation to send the RPC Call.
If a Requester has provided insufficient Reply chunk space for a Responder to send an RPC Reply, the Responder can use Message Continuation to send the RPC Reply.
If a sender has to convey a sizeable non-RPC data payload (e.g., a large transport property), the sender can use Message Continuation to avoid having to register memory.

Host Authentication Changes For the general operation of NFS on open networks, we eventually intend to rely on RPC-on-TLS to provide cryptographic authentication of the two ends of each connection. In turn, this can improve the trustworthiness of AUTH_SYS-style user identities that flow on TCP, which are not cryptographically protected. We do not have a similar solution for RPC-over-RDMA, however. Here, the RDMA transport layer already provides a strong guarantee of message integrity. On some network fabrics, IPsec or TLS can protect the privacy of in-transit data. However, this is not the case for all fabrics (e.g., InfiniBand ). Thus, RPC-over-RDMA version 2 introduces a mechanism for authenticating connection peers (see ). And like GSS channel binding, there is also a way to determine when the use of host authentication is unnecessary.

Support for Remote Invalidation When an RDMA consumer uses FRWR or Memory Windows to register memory, that memory may be invalidated remotely [RFC5040]. These mechanisms are available when a Requester's RNIC supports MEM_MGT_EXTENSIONS. For this discussion, there are two classes of STags. Dynamically-registered STags appear in a single RPC, then are invalidated. Persistently-registered STags survive longer than one RPC. They may persist for the life of an RPC-over-RDMA connection or even longer. An RPC-over-RDMA Requester can provide more than one STag in a transport header. It may provide a combination of dynamically- and persistently-registered STags in one RPC message, or any combination of these in a series of RPCs on the same connection. Only dynamically-registered STags using Memory Windows or FRWR may be invalidated remotely. There is no transport-level mechanism by which a Responder can determine how a Requester-provided STag was registered, nor whether it is eligible to be invalidated remotely. A Requester that mixes persistently- and dynamically-registered STags in one RPC, or mixes them across RPCs on the same connection, must, therefore, indicate which STag the Responder may invalidate remotely via a mechanism provided in the Upper-Layer Protocol. RPC-over-RDMA version 2 provides such a mechanism. A sender uses the RDMA Send With Invalidate operation to invalidate an STag on the remote peer. It is available only when both peers support MEM_MGT_EXTENSIONS (can send and process an IETH). Existing RPC-over-RDMA transport protocol specifications do not forbid direct data placement in the reverse direction. Moreover, there is currently no Upper-Layer Protocol that makes data items in reverse-direction operations eligible for direct data placement. When chunks are present in a reverse-direction RPC request, Remote Invalidation enables the Responder to trigger invalidation of a Requester's STags as part of sending an RPC Reply, the same way as is done in the forward direction. However, in the reverse direction, the server acts as the Requester, and the client is the Responder. The server's RNIC, therefore, must support receiving an IETH, and the server must have registered its STags with an appropriate registration mechanism.

Integration of Reverse-Direction Operation Because did not include specification of reverse-direction operation, does not include it either. Reverse-direction operation in RPC-over-RDMA version 1 is specified by a separate standards track document . Reverse-direction operation in RPC-over-RDMA version 1 was constrained by the limited ability to extend that version of the protocol. The most awkward issue is that a receiver needs to peek at ingress RPC message payloads to determine whether it is a Call or Reply message. This is necessary because the meaning of several fields in the RPC-over-RDMA transport header is determined by the direction of the RPC message payload:

The meaning of the value in the rdma_xid field is determined by the direction of the message because the XID spaces in the forward and reverse directions are distinct.
The meaning of the value in the rdma_credit field is determined by the direction of the message because credits are granted separately for forward and reverse direction operation.
The purpose of Write chunks and the meaning of their length fields is determined by the direction of the message because in Call messages, they are provisional, but in Reply messages, they represent returned results.

The current document remedies this awkwardness by integrating reverse-direction operation into RPC-over-RDMA version 2 so that it can make use of all facilities that are available in the forward-direction, including body chunks, remote invalidation, and message continuation. To enable this integration, the direction of the RPC message payload is encoded in each RPC-over-RDMA version 2 transport header.

Error Reporting Changes RPC-over-RDMA version 2 expands the repertoire of errors that connection peers may report to each other. The goals of this expansion are:

To fill in details of peer recovery actions.
To enable retrying certain conditions caused by mis-estimation of the maximum reply size.
To minimize the likelihood of a Requester waiting forever for a Reply when there are communications problems that prevent the Responder from sending it.

Changes in Terminology The RPC-over-RDMA version 2 specification makes the following changes in terminology. These changes do not result in changes in the behavior or operation of the protocol.

The current document explicitly acknowledges the different semantics and purpose of Write chunks appearing in Call messages and those appearing in Reply messages.
The current document introduces the term "payload format" to describe the selection of a mechanism for reducing and conveying an RPC message payload. It replaces the terms "short message" and "long message" with the terms "simple format" and "special format" because this selection is not based only on the size of the payload.
The current document introduces the terms "data item chunk" and "body chunk" in order to distinguish the purpose and operation of these two categories of chunk.
For improved readability, the current document replaces the terms "RDMA segment" and "plain segment" with the term "segment", and the term "RDMA read segment" with the term "Read segment".
The current document refers specifically to the RDMAP, DDP, and MPA standards track protocols rather than using the nebulous term "iWARP".

Acknowledgments The authors gratefully acknowledge the work of and on the original RPC-over-RDMA version 1 specification . We are deeply indebted to for contributing the RPC-over-RDMA version 2 flow control mechanism described in . The authors also wish to thank , , and for their support of this work. The XDR extraction conventions were first described by the authors of the NFS version 4.1 XDR specification . suggested the replacement sed script used in this document. Special thanks go to Transport Area Director , NFSV4 Working Group Chairs , and , and NFSV4 Working Group Secretary for their support.