| Network File System Version 4 | C. Lever, Ed. |
| Internet-Draft | Oracle |
| Obsoletes: 5666 (if approved) | W. Simpson |
| Intended status: Standards Track | Red Hat |
| Expires: September 28, 2017 | T. Talpey |
| Microsoft | |
| March 27, 2017 |
Remote Direct Memory Access Transport for Remote Procedure Call, Version One
draft-ietf-nfsv4-rfc5666bis-11
This document specifies a protocol for conveying Remote Procedure Call (RPC) messages on physical transports capable of Remote Direct Memory Access (RDMA). It requires no revision to application RPC protocols or the RPC protocol itself. This document obsoletes RFC 5666.
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in [RFC2119].
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at http://datatracker.ietf.org/drafts/current/.
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."
This Internet-Draft will expire on September 28, 2017.
Copyright (c) 2017 IETF Trust and the persons identified as the document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (http://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.
This document may contain material from IETF Documents or IETF Contributions published or made publicly available before November 10, 2008. The person(s) controlling the copyright in some of this material may not have granted the IETF Trust the right to allow modifications of such material outside the IETF Standards Process. Without obtaining an adequate license from the person(s) controlling the copyright in such materials, this document may not be modified outside the IETF Standards Process, and derivative works of it may not be created outside the IETF Standards Process, except to format it for publication as an RFC or to translate it into languages other than English.
This document specifies the RPC-over-RDMA Version One protocol, based on existing implementations of RFC 5666 and experience gained through deployment. This document obsoletes RFC 5666.
The new specification clarifies text that is subject to multiple interpretations, and removes support for unimplemented RPC-over-RDMA Version One protocol elements. It clarifies the role of Upper Layer Bindings and describes what they are to contain.
In addition, this document describes current practice using RPCSEC_GSS [RFC7861] on RDMA transports.
The protocol version number has not been changed because the protocol specified in this document fully interoperates with implementations of the RPC-over-RDMA Version One protocol specified in [RFC5666].
Remote Direct Memory Access (RDMA) [RFC5040] [RFC5041] [IB] is a technique for moving data efficiently between end nodes. By directing data into destination buffers as it is sent on a network, and placing it via direct memory access by hardware, the benefits of faster transfers and reduced host overhead are obtained.
Open Network Computing Remote Procedure Call (ONC RPC, often shortened in NFSv4 documents to RPC) [RFC5531] is a remote procedure call protocol that runs over a variety of transports. Most RPC implementations today use UDP [RFC0768] or TCP [RFC0793]. On UDP, RPC messages are encapsulated inside datagrams, while on a TCP byte stream, RPC messages are delineated by a record marking protocol. An RDMA transport also conveys RPC messages in a specific fashion that must be fully described if RPC implementations are to interoperate.
RDMA transports present semantics different from either UDP or TCP. They retain message delineations like UDP, but provide reliable and sequenced data transfer like TCP. They also provide an offloaded bulk transfer service not provided by UDP or TCP. RDMA transports are therefore appropriately viewed as a new transport type by RPC.
In this context, the Network File System (NFS) protocols as described in [RFC1094], [RFC1813], [RFC7530], [RFC5661], and future NFSv4 minor versions are all obvious beneficiaries of RDMA transports. A complete problem statement is presented in [RFC5532]. Many other RPC-based protocols can also benefit.
Although the RDMA transport described herein can provide relatively transparent support for any RPC application, this document also describes mechanisms that can optimize data transfer even further, when RPC applications are willing to exploit awareness of RDMA as the transport.
This section highlights key elements of the Remote Procedure Call [RFC5531] and External Data Representation [RFC4506] protocols, upon which RPC-over-RDMA Version One is constructed. Strong grounding with these protocols is recommended before reading this document.
Remote Procedure Calls are an abstraction used to implement the operations of an "Upper Layer Protocol," or ULP. The term Upper Layer Protocol 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 an Upper Layer Protocol is the Network File System Version 4.0 [RFC7530].
In this document, the term "RPC consumer" refers to an implementation of an Upper Layer Protocol running on a client.
Like a local procedure call, every Remote Procedure Call (RPC) has a set of "arguments" and a set of "results". A calling context is not allowed to proceed until the procedure's results are available to it. Unlike a local procedure call, the called procedure is executed remotely rather than in the local application's context.
The RPC protocol as described in [RFC5531] 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 those consumers).
ONC RPC transactions are made up of two types of messages:
The RPC client endpoint acts as a "requester". It serializes an RPC Call's arguments and conveys them to a server endpoint via an RPC Call message. This message contains an RPC protocol header, a header describing the requested upper layer operation, and all arguments.
The RPC server endpoint acts as a "responder". It deserializes Call arguments, and processes the requested operation. It then serializes the operation's results into another byte stream. This byte stream is conveyed back to the requester via an RPC Reply message. This message contains an RPC protocol header, a header describing the upper layer reply, and all results.
The requester deserializes the results and allows the original caller to proceed. At this point the RPC transaction designated by the xid in the Call message is complete, and the xid is retired.
In summary, CALL messages are sent by requesters to responders to initiate RPC transactions. REPLY messages are sent by responders to requesters to complete the processing on an RPC transaction.
The role of an "RPC transport" is to mediate the exchange of RPC messages between requesters and responders. An RPC transport bridges the gap between the RPC message abstraction and the native operations of a particular network transport.
RPC-over-RDMA is a connection-oriented RPC transport. When a connection-oriented transport is used, clients initiate transport connections, while servers wait passively for incoming connection requests.
One cannot assume that all requesters and responders represent data objects the same way internally. RPC uses eXternal Data Representation, or XDR, to translate native data types and serialize arguments and results [RFC4506].
The XDR protocol encodes data independent of the endianness or size of host-native data types, allowing unambiguous decoding of data on the receiving end. RPC Programs are specified by writing an XDR definition of their procedures, argument data types, and result data types.
XDR assumes that the number of bits in a byte (octet) and their order are the same on both endpoints and on the physical network. The smallest indivisible unit of XDR encoding is a group of four octets in little-endian order. XDR also flattens lists, arrays, and other complex data types so they can be conveyed as a stream of bytes.
A serialized stream of bytes that is the result of XDR encoding is referred to as an "XDR stream." A sending endpoint encodes native data into an XDR stream and then transmits that stream to a receiver. A receiving endpoint decodes incoming XDR byte streams into its native data representation format.
Sometimes a data item must be transferred as-is, without encoding or decoding. The contents of such a data item are referred to as "opaque data." XDR encoding places the content of opaque data items directly into an XDR stream without altering it in any way. Upper Layer Protocols or applications perform any needed data translation in this case. Examples of opaque data items include the content of files, or generic byte strings.
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, octets containing zero are added after the end of the item as it is encoded so that the next encoded data item in the XDR stream starts on a four-octet boundary. The encoded size of the item is not changed by the addition of the extra octets. These extra octets are never exposed to Upper Layer Protocols.
This technique is referred to as "XDR round-up," and the extra octets are referred to as "XDR round-up padding".
RPC requesters and responders can be made more efficient if large RPC messages are transferred by a third party such as intelligent network interface hardware (data movement offload), and placed in the receiver's memory so that no additional adjustment of data alignment has to be made (direct data placement). Remote Direct Memory Access (RDMA) transports enable both optimizations.
Typically, RPC implementations copy the contents of RPC messages into a buffer before being sent. An efficient RPC implementation sends bulk data without copying it into a separate send buffer first.
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; sometimes, only to adjust data alignment.
In this document, "RDMA" refers to the physical mechanism an RDMA transport utilizes when moving data. Although this may not be efficient, before an RDMA transfer a sender may copy data into an intermediate buffer before an RDMA transfer. After an RDMA transfer, a receiver may copy that data again to its final destination.
This document uses the term "direct data placement" (or DDP) to refer to any optimized data transfer where it is unnecessary for a receiving host's CPU to copy transferred data to another location after it has been received.
Just as [RFC5666] did, this document focuses on the use of RDMA Read and Write operations to achieve both data movement offload and Direct Data Placement. However, not all RDMA-based data transfer qualifies as Direct Data Placement, and DDP can be achieved using non-RDMA mechanisms.
To achieve good performance during receive operations, RDMA transports require that RDMA consumers provision resources in advance to receive incoming messages.
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 hardware matches it to an in-bound Send operation. The resources associated with that Receive must be retained in host memory, or "pinned," until the Receive completes.
Given these basic tenets of RDMA transport operation, the RPC-over-RDMA Version One protocol assumes each transport provides the following abstract operations. A more complete discussion of these operations is found in [RFC5040].
The RPC-over-RDMA Version One protocol is designed to be carried over RDMA transports that support the above abstract operations. This protocol conveys information sufficient for an RPC peer to direct an RDMA provider to perform transfers containing RPC data and to communicate their result(s).
A "transfer model" designates which endpoint exposes its memory, and which is responsible for initiating transfer of data. To enable RDMA Read and Write operations, for example, an endpoint first exposes regions of its memory to a remote endpoint, which initiates these operations against the exposed memory. [RFC5666] specifies the use of both the Read-Read and the Read-Write Transfer Model. All current RPC-over-RDMA Version One implementations use only the Read-Write Transfer Model. Therefore, protocol elements that enable the Read-Read Transfer Model have been removed from the RPC-over-RDMA Version One specification in this document. Transfer Models other than the Read-Write model may be used in future versions of RPC-over-RDMA.
On an RPC-over-RDMA transport, each RPC message is encapsulated by an RPC-over-RDMA message. An RPC-over-RDMA message consists of two XDR streams.
In its simplest form, an RPC-over-RDMA message consists of a Transport stream followed immediately by a Payload stream conveyed together in a single RDMA Send. To transmit large RPC messages, a combination of one RDMA Send operation and one or more other RDMA operations is employed.
RPC-over-RDMA framing replaces all other RPC framing (such as TCP record marking) when used atop an RPC-over-RDMA association, even when the underlying RDMA protocol may itself be layered atop a transport with a defined RPC framing (such as TCP).
It is however possible for RPC-over-RDMA to be dynamically enabled in the course of negotiating the use of RDMA via an Upper Layer Protocol exchange. 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.
It is critical to provide RDMA Send flow control for an RDMA connection. If any pre-posted receive buffer on the connection is not large enough to accept an incoming RDMA Send, or if a pre-posted receive buffer is not available to accept an incoming RDMA Send, the RDMA connection can be terminated. This is different than conventional TCP/IP networking, in which buffers are allocated dynamically as messages are received.
The longevity of an RDMA connection mandates that sending endpoints respect the resource limits of peer receivers. To ensure messages can be sent and received reliably, there are two operational parameters for each connection.
Flow control for RDMA Send operations directed to the responder is implemented as a simple request/grant protocol in the RPC-over-RDMA header associated with each RPC message.
An RPC-over-RDMA Version One credit is the capability to handle one RPC-over-RDMA transaction. Each RPC-over-RDMA message sent from requester to responder requests a number of credits from the responder. Each RPC-over-RDMA message sent from responder to requester informs the requester how many credits the responder has granted. The requested and granted values are carried in each RPC-over-RDMA message's rdma_credit field (see Section 4.2.3).
Practically speaking, the critical value is the granted value. A requester MUST NOT send unacknowledged requests in excess of the responder's granted credit limit. If the granted value is exceeded, the RDMA layer may signal an error, possibly terminating the connection. The granted value MUST NOT be zero, since such a value would result in deadlock.
RPC calls complete in any order, but the current granted credit limit at the responder is known to the requester from RDMA Send ordering properties. The number of allowed new requests the requester may send is then the lower of the current requested and granted credit values, minus the number of requests in flight. Advertised credit values are not altered when individual RPCs are started or completed.
The requested and granted credit values MAY be adjusted to match the needs or policies in effect on either peer. For instance, a responder may reduce the granted credit value to accommodate the available resources in a Shared Receive Queue. The responder MUST ensure that an increase in receive resources is effected before the next reply message is sent.
A requester MUST maintain enough receive resources to accommodate expected replies. Responders have to be prepared for there to be no receive resources available on requesters with no pending RPC transactions.
Certain RDMA implementations may impose additional flow control restrictions, such as limits on RDMA Read operations in progress at the responder. Accommodation of such restrictions is considered the responsibility of each RPC-over-RDMA Version One implementation.
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. The inline threshold value is the minimum of how large a message the sender can post via an RDMA Send operation, and how large a message the receiver can accept via an RDMA Receive operation. Each connection has two inline threshold values: 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").
Unlike credit limits, inline threshold values are not advertised to peers via the RPC-over-RDMA Version One protocol, and there is no provision for inline threshold values to change during the lifetime of an RPC-over-RDMA Version One connection.
When a connection is first established, peers might not know how many receive resources the other has, nor how large the other peer's inline thresholds are.
As a basis for an initial exchange of RPC requests, each RPC-over-RDMA Version One connection provides the ability to exchange at least one RPC message at a time, whose Call and Reply messages are no more 1024 bytes in size. A responder MAY exceed this basic level of configuration, but a requester MUST NOT assume more than one credit is available, and MUST receive a valid reply from the responder carrying the actual number of available credits, prior to sending its next request.
Receiver implementations MUST support inline thresholds of 1024 bytes, but MAY support larger inline thresholds values. An indepedent mechanism for discovering a peer's inline thresholds before a connection is established may be used to optimize the use of RDMA Send and Receive operations. In the absense of such a mechanism, senders and receives MUST assume the inline thresholds are 1024 bytes.
When a Direct Data Placement capability is available, the transport places the contents of one or more XDR data items directly into the receiver's memory, separately from the transfer of other parts of the containing XDR stream.
RPC-over-RDMA Version One provides a mechanism for moving part of an RPC message via a data transfer distinct from an RDMA Send/Receive pair. The sender removes one or more XDR data items from the Payload stream. They are conveyed via other mechanisms, such as one or more RDMA Read or Write operations. As the receiver decodes an incoming message, it skips over directly placed data items.
The portion of an XDR stream that is split out and moved separately is referred to as a "chunk". In some contexts, data in an RPC-over-RDMA header that describes these split out regions of memory may also be referred to as a "chunk".
A Payload stream after chunks have been removed is referred to as a "reduced" Payload stream. Likewise, a data item that has been removed from a Payload stream to be transferred separately is referred to as a "reduced" data item.
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. In addition, it is impractical for receivers to prepare for every possible XDR data item in a protocol to be transferred in a chunk.
To maintain interoperability on an RPC-over-RDMA transport, a determination must be made of which few XDR data items in each Upper Layer Protocol are allowed to use Direct Data Placement.
This is done by additional specifications that describe how Upper Layer Protocols employ Direct Data Placement. An "Upper Layer Binding specification," or ULB, identifies which specific individual XDR data items in an Upper Layer Protocol MAY be transferred via Direct Data Placement. Such data items are referred to as "DDP-eligible." All other XDR data items MUST NOT be reduced.
Detailed requirements for Upper Layer Bindings are provided in Section 6.
When encoding a Payload stream that contains a DDP-eligible data item, a sender may choose to reduce that data item. When it chooses to do so, the sender does not place the item into the Payload stream. Instead, the sender records in the RPC-over-RDMA header the location and size of the memory region containing that data item.
The requester provides location information for DDP-eligible data items in both RPC Calls and Replies. The responder uses this information to retrieve arguments contained in the specified region of the requester's memory, or place results in that memory region.
An "RDMA segment," or "plain segment," is an RPC-over-RDMA Transport header data object that contains the precise co-ordinates of a contiguous memory region that is to be conveyed separately from the Payload stream. Plain segments contain the following information:
[RFC5040] for further discussion of the meaning and use of these fields.
See
In RPC-over-RDMA Version One, a "chunk" refers to a portion of the Payload stream that is moved independently of the RPC-over-RDMA Transport header and Payload stream. Chunk data is removed from the sender's Payload stream, transferred via separate operations, and then re-inserted into the receiver's Payload stream to form a complete RPC message.
Each chunk consists of one or more RDMA segments. Each RDMA segment represents a single contiguous piece of that chunk. A requester MAY divide a chunk into RDMA segments using any boundaries that are convenient. The length of a chunk is the sum of the lengths of the RDMA segments that comprise it.
The RPC-over-RDMA Version One transport protocol does not place a limit on chunk size. However, each Upper Layer Protocol may cap the amount of data that can be transferred by a single RPC (for example, NFS has "rsize" and "wsize", which 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.
If a chunk contains a counted array data type, the count of array elements MUST remain in the Payload stream, while the array elements MUST be moved to the chunk. For example, when encoding an opaque byte array as a chunk, the count of bytes stays in the Payload stream, while the bytes in the array are removed from the Payload stream and transferred within 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.
If a chunk contains an optional-data data type, the "is present" field MUST remain in the Payload stream, while the data, if present, MUST be moved to the chunk.
A union data type should never be made DDP-eligible, but one or more of its arms may be DDP-eligible.
Except in special cases (covered in Section 3.5.3), a chunk MUST contain exactly one XDR data item. This makes it straightforward to reduce variable-length data items without affecting the XDR alignment of data items in the Payload stream.
When a variable-length XDR data item is reduced, the sender MUST remove XDR round-up padding for that data item from the Payload stream, so that data items remaining in the Payload stream begin on four-byte alignment.
A "Read chunk" represents an XDR data item that is to be pulled from the requester to the responder.
A Read chunk is a list of one or more RDMA read segments. An RDMA read segment consists of a Position field followed by a plain segment. See Section 4.1.2 for details.
While constructing an RPC-over-RDMA Call message, a requester registers memory regions that contain data to be transferred via RDMA Read operations. It advertises the co-ordinates of these regions in the RPC-over-RDMA Transport header of the RPC Call.
After receiving an RPC Call sent via an RDMA Send operation, a responder transfers the chunk data from the requester using RDMA Read operations. The responder reconstructs the transferred chunk data by concatenating the contents of each RDMA segment, in list order, into the received Payload stream at the Position value recorded in that RDMA segment.
Put another way, the responder inserts the first RDMA segment in a Read chunk into the Payload stream at the byte offset indicated by its Position field. RDMA segments whose Position field value match this offset are concatenated afterwards, until there are no more RDMA segments at that Position value.
The Position field in a read segment indicates where the containing Read chunk starts in the Payload stream. The value in this field MUST be a multiple of four. All segments in the same Read chunk share the same Position value, even if one or more of the RDMA segments have a non-four-byte aligned length.
While decoding a received Payload stream, whenever the XDR offset in the Payload stream matches that of a Read chunk, the responder initiates an RDMA Read to pull the chunk's data content into registered local memory.
The responder acknowledges its completion of use of Read chunk source buffers when it sends an RPC Reply to the requester. The requester may then release Read chunks advertised in the request.
When reducing a variable-length argument data item, the requester SHOULD NOT include the data item's XDR round-up padding in the chunk. The length of a Read chunk is determined as follows:
While constructing an RPC Call message, a requester prepares memory regions in which to receive DDP-eligible result data items. A "Write chunk" represents an XDR data item that is to be pushed from a responder to a requester. It is made up of an array of one or more plain segments.
Write chunks are provisioned by a requester long before the responder has prepared the reply Payload stream. A requester often does not know the actual length of the result data items to be returned, since the result does not yet exist. Thus it MUST register Write chunks long enough to accommodate the maximum possible size of each returned data item.
In addition, the XDR position of DDP-eligible data items in the reply's Payload stream is not predictable when a requester constructs a Call message. Therefore RDMA segments in a Write chunk do not have a Position field.
For each Write chunk provided by a requester, the responder pushes data to the requester, contiguously and in segment array order, until the result data item has been completely written to the requester. The responder MUST copy the segment count and all segments from the requester-provided Write chunk into the Reply's Transport header. As it does so, the responder updates each segment length field to reflect the actual amount of data that is being returned in that segment. The responder then sends the RPC Reply via an RDMA Send operation.
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.
After receiving the RPC Reply, the requester reconstructs the transferred data by concatenating the contents of each segment, in array order, into RPC Reply XDR stream at the known XDR position of the associated DDP-eligible result data item.
When provisioning a Write chunk for a variable-length result data item, the requester SHOULD NOT include additional space for XDR round-up padding. A responder MUST NOT write XDR round-up padding into a Write chunk, even if the requester made space available for it. Therefore, when returning a single variable-length result data item, a returned Write chunk's total length MUST be the same as the encoded length of the result data item.
A receiver of RDMA Send operations is required by RDMA to have previously posted one or more adequately sized buffers. Memory savings are achieved on both requesters and responders by posting small Receive buffers. However, not all RPC messages are small.
RPC messages are frequently smaller than typical inline thresholds. For example, the NFS version 3 GETATTR operation is only 56 bytes: 20 bytes of RPC header, plus a 32-byte file handle argument and 4 bytes for its length. The reply to this common request is about 100 bytes.
Since all RPC messages conveyed via RPC-over-RDMA require an RDMA Send operation, 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. An RPC-over-RDMA header with a small RPC Call or Reply message immediately following is transferred using a single RDMA Send operation. No other operations are needed.
Requester Responder
| RDMA Send (RDMA_MSG) |
Call | ------------------------------> |
| |
| | Processing
| |
| RDMA Send (RDMA_MSG) |
| <------------------------------ | Reply
An RPC-over-RDMA transaction using Short Messages:
If DDP-eligible data items are present in a Payload stream, a sender MAY reduce some or all of these items by removing them from the Payload stream. The sender uses a separate mechanism to transfer the reduced data items. The Transport stream with the reduced Payload stream immediately following is then transferred using a single RDMA Send operation
After receiving the Transport and Payload streams of a Chunked RPC-over-RDMA Call message, the responder uses RDMA Read operations to move reduced data items in Read chunks. Before sending the Transport and Payload streams of a Chunked RPC-over-RDMA Reply message, the responder uses RDMA Write operations to move reduced data items in Write and Reply chunks.
Requester Responder
| RDMA Send (RDMA_MSG) |
Call | ------------------------------> |
| RDMA Read |
| <------------------------------ |
| RDMA Response (arg data) |
| ------------------------------> |
| |
| | Processing
| |
| RDMA Send (RDMA_MSG) |
| <------------------------------ | Reply
An RPC-over-RDMA transaction with a Read chunk:
Requester Responder
| RDMA Send (RDMA_MSG) |
Call | ------------------------------> |
| |
| | Processing
| |
| RDMA Write (result data) |
| <------------------------------ |
| RDMA Send (RDMA_MSG) |
| <------------------------------ | Reply
An RPC-over-RDMA transaction with a Write chunk:
When a Payload stream is larger than the receiver's inline threshold, the Payload stream is reduced by removing DDP-eligible data items and placing them in chunks to be moved separately. If there are no DDP-eligible data items in the Payload stream, or the Payload stream is still too large after it has been reduced, the RDMA transport MUST use RDMA Read or Write operations to convey the Payload stream itself. This mechanism is referred to as a "Long Message."
To transmit a Long Message, the sender conveys only the Transport stream with an RDMA Send operation. The Payload stream is not included in the Send buffer in this instance. Instead, the requester provides chunks that the responder uses to move the Payload stream.
Though the purpose of a Long Message is to handle large RPC messages, requesters MAY use a Long Message at any time to convey an RPC Call.
A responder chooses which form of reply to use based on the chunks provided by the requester. If Write chunks were provided and the responder has a DDP-eligible result, it first reduces the reply Payload stream. If a Reply chunk was provided and the reduced Payload stream is larger than the reply inline threshold, the responder MUST use the requester-provided Reply chunk for the reply.
XDR data items may appear in these special chunks without regard to their DDP-eligibility. As these chunks contain a Payload stream, such chunks MUST include appropriate XDR round-up padding to maintain proper XDR alignment of their contents.
Requester Responder
| RDMA Send (RDMA_NOMSG) |
Call | ------------------------------> |
| RDMA Read |
| <------------------------------ |
| RDMA Response (RPC call) |
| ------------------------------> |
| |
| | Processing
| |
| RDMA Send (RDMA_MSG) |
| <------------------------------ | Reply
An RPC-over-RDMA transaction using a Long Call:
Requester Responder
| RDMA Send (RDMA_MSG) |
Call | ------------------------------> |
| |
| | Processing
| |
| RDMA Write (RPC reply) |
| <------------------------------ |
| RDMA Send (RDMA_NOMSG) |
| <------------------------------ | Reply
An RPC-over-RDMA transaction using a Long Reply:
Every RPC-over-RDMA Version One message has a header that includes a copy of the message's transaction ID, data for managing RDMA flow control credits, and lists of RDMA segments describing chunks. All RPC-over-RDMA header content is contained in the Transport stream, and thus MUST be XDR encoded.
RPC message layout is unchanged from that described in [RFC5531] except for the possible reduction of data items that are moved by separate operations.
The RPC-over-RDMA protocol passes RPC messages without regard to their type (CALL or REPLY). Apart from restrictions imposed by upper-layer bindings, each endpoint of a connection MAY send RDMA_MSG or RDMA_NOMSG message header types at any time (subject to credit limits).
This section contains a description of the core features of the RPC-over-RDMA Version One protocol, expressed in the XDR language [RFC4506].
This description is provided in a way that makes it simple to extract into ready-to-compile form. The reader can apply the following shell script to this document to produce a machine-readable XDR description of the RPC-over-RDMA Version One protocol.
<CODE BEGINS> #!/bin/sh grep '^ *///' | sed 's?^ /// ??' | sed 's?^ *///$??' <CODE ENDS>
That is, if the above script is stored in a file called "extract.sh" and this document is in a file called "spec.txt" then the reader can do the following to extract an XDR description file:
<CODE BEGINS> sh extract.sh < spec.txt > rpcrdma_corev1.x <CODE ENDS>
<CODE BEGINS> /// /* /// * Copyright (c) 2010, 2016 IETF Trust and the persons /// * identified as authors of the code. All rights reserved. /// * /// * The authors of the code are: /// * B. Callaghan, T. Talpey, and C. Lever /// * /// * Redistribution and use in source and binary forms, with /// * or without modification, are permitted provided that the /// * following conditions are met: /// * /// * - Redistributions of source code must retain the above /// * copyright notice, this list of conditions and the /// * following disclaimer. /// * /// * - Redistributions in binary form must reproduce the above /// * copyright notice, this list of conditions and the /// * following disclaimer in the documentation and/or other /// * materials provided with the distribution. /// * /// * - Neither the name of Internet Society, IETF or IETF /// * Trust, nor the names of specific contributors, may be /// * used to endorse or promote products derived from this /// * software without specific prior written permission. /// * /// * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS /// * AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED /// * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE /// * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS /// * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO /// * EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE /// * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, /// * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT /// * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR /// * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS /// * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF /// * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, /// * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING /// * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF /// * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. /// */ /// <CODE ENDS>
Code components extracted from this document must include the following license text. When the extracted XDR code is combined with other complementary XDR code which itself has an identical license, only a single copy of the license text need be preserved.
<CODE BEGINS>
/// /*
/// * Plain RDMA segment (Section 3.4.3)
/// */
/// struct xdr_rdma_segment {
/// uint32 handle; /* Registered memory handle */
/// uint32 length; /* Length of the chunk in bytes */
/// uint64 offset; /* Chunk virtual address or offset */
/// };
///
/// /*
/// * RDMA read segment (Section 3.4.5)
/// */
/// struct xdr_read_chunk {
/// uint32 position; /* Position in XDR stream */
/// struct xdr_rdma_segment target;
/// };
///
/// /*
/// * Read list (Section 4.3.1)
/// */
/// struct xdr_read_list {
/// struct xdr_read_chunk entry;
/// struct xdr_read_list *next;
/// };
///
/// /*
/// * Write chunk (Section 3.4.6)
/// */
/// struct xdr_write_chunk {
/// struct xdr_rdma_segment target<>;
/// };
///
/// /*
/// * Write list (Section 4.3.2)
/// */
/// struct xdr_write_list {
/// struct xdr_write_chunk entry;
/// struct xdr_write_list *next;
/// };
///
/// /*
/// * Chunk lists (Section 4.3)
/// */
/// struct rpc_rdma_header {
/// struct xdr_read_list *rdma_reads;
/// struct xdr_write_list *rdma_writes;
/// struct xdr_write_chunk *rdma_reply;
/// /* rpc body follows */
/// };
///
/// struct rpc_rdma_header_nomsg {
/// struct xdr_read_list *rdma_reads;
/// struct xdr_write_list *rdma_writes;
/// struct xdr_write_chunk *rdma_reply;
/// };
///
/// /* Not to be used */
/// struct rpc_rdma_header_padded {
/// uint32 rdma_align;
/// uint32 rdma_thresh;
/// struct xdr_read_list *rdma_reads;
/// struct xdr_write_list *rdma_writes;
/// struct xdr_write_chunk *rdma_reply;
/// /* rpc body follows */
/// };
///
/// /*
/// * Error handling (Section 4.5)
/// */
/// enum rpc_rdma_errcode {
/// ERR_VERS = 1, /* Value fixed for all versions */
/// ERR_CHUNK = 2
/// };
///
/// /* Structure fixed for all versions */
/// struct rpc_rdma_errvers {
/// uint32 rdma_vers_low;
/// uint32 rdma_vers_high;
/// };
///
/// union rpc_rdma_error switch (rpc_rdma_errcode err) {
/// case ERR_VERS:
/// rpc_rdma_errvers range;
/// case ERR_CHUNK:
/// void;
/// };
///
/// /*
/// * Procedures (Section 4.2.4)
/// */
/// enum rdma_proc {
/// RDMA_MSG = 0, /* Value fixed for all versions */
/// RDMA_NOMSG = 1, /* Value fixed for all versions */
/// RDMA_MSGP = 2, /* Not to be used */
/// RDMA_DONE = 3, /* Not to be used */
/// RDMA_ERROR = 4 /* Value fixed for all versions */
/// };
///
/// /* The position of the proc discriminator field is
/// * fixed for all versions */
/// union rdma_body switch (rdma_proc proc) {
/// case RDMA_MSG:
/// rpc_rdma_header rdma_msg;
/// case RDMA_NOMSG:
/// rpc_rdma_header_nomsg rdma_nomsg;
/// case RDMA_MSGP: /* Not to be used */
/// rpc_rdma_header_padded rdma_msgp;
/// case RDMA_DONE: /* Not to be used */
/// void;
/// case RDMA_ERROR:
/// rpc_rdma_error rdma_error;
/// };
///
/// /*
/// * Fixed header fields (Section 4.2)
/// */
/// struct rdma_msg {
/// uint32 rdma_xid; /* Position fixed for all versions */
/// uint32 rdma_vers; /* Position fixed for all versions */
/// uint32 rdma_credit; /* Position fixed for all versions */
/// rdma_body rdma_body;
/// };
<CODE ENDS>
XDR data items defined in this section encodes the Transport Header Stream in each RPC-over-RDMA Version One message. Comments identify items that cannot be changed in subsequent versions.
The RPC-over-RDMA header begins with four fixed 32-bit fields that control the RDMA interaction.
The first three words are individual fields in the rdma_msg structure. The fourth word is the first word of the rdma_body union which acts as the discriminator for the switched union. The contents of this field are described in Section 4.2.4.
These four fields must remain with the same meanings and in the same positions in all subsequent versions of the RPC-over-RDMA protocol.
The XID generated for the RPC Call and Reply. Having the XID at a fixed location in the header makes it easy for the receiver to establish context as soon as each RPC-over-RDMA message arrives. This XID MUST be the same as the XID in the RPC message. The receiver MAY perform its processing based solely on the XID in the RPC-over-RDMA header, and thereby ignore the XID in the RPC message, if it so chooses.
For RPC-over-RDMA Version One, this field MUST contain the value one (1). Rules regarding changes to this transport protocol version number can be found in Section 7.
When sent with an RPC Call message, the requested credit value is provided. When sent with an RPC Reply message, the granted credit value is returned. Further discussion of how the credit value is determined can be found in Section 3.3.
An RDMA_MSG procedure conveys the Transport stream and the Payload stream via an RDMA Send operation. The Transport stream contains the four fixed fields, followed by the Read and Write lists and the Reply chunk, though any or all three MAY be marked as not present. The Payload stream then follows, beginning with its XID field. If a Read or Write chunk list is present, a portion of the Payload stream has been excised and is conveyed via separate operations.
An RDMA_NOMSG procedure conveys the Transport stream via an RDMA Send operation. The Transport stream contains the four fixed fields, followed by the Read and Write chunk lists and the Reply chunk. Though any of these MAY be marked as not present, one MUST be present and MUST hold the Payload stream for this RPC-over-RDMA message. If a Read or Write chunk list is present, a portion of the Payload stream has been excised and is conveyed via separate operations.
An RDMA_ERROR procedure conveys the Transport stream via an RDMA Send operation. The Transport stream contains the four fixed fields, followed by formatted error information. No Payload stream is conveyed in this type of RPC-over-RDMA message.
A requester MUST NOT send an RPC-over-RDMA header with the RDMA_ERROR procedure. A responder MUST silently discard RDMA_ERROR procedures.
A gather operation on each RDMA Send operation can be used to combine the Transport and Payload streams, which might have been constructed in separate buffers. However, the total length of the gathered send buffers MUST NOT exceed the inline threshold.
The chunk lists in an RPC-over-RDMA Version One header are three XDR optional-data fields that follow the fixed header fields in RDMA_MSG and RDMA_NOMSG procedures. Read Section 4.19 of [RFC4506] carefully to understand how optional-data fields work. Examples of XDR encoded chunk lists are provided in Section 4.7 as an aid to understanding.
Often, an RPC-over-RDMA message has no associated chunks. In this case, all three chunk lists are marked empty (not present).
Each RDMA_MSG or RDMA_NOMSG procedure has one "Read list." The Read list is a list of zero or more RDMA Read segments, provided by the requester, that are grouped by their Position fields into Read chunks. Each Read chunk advertises the location of argument data the responder is to pull from the requester. The requester has removed the data items in these chunks from the call's Payload stream.
A requester may transmit the Payload stream of an RPC Call message using a Position Zero Read chunk. If the RPC Call has no argument data that is DDP-eligible and the Position Zero Read chunk is not being used, the requester leaves the Read list empty.
Responders MUST leave the Read list empty in all replies.
When reducing a DDP-eligible argument data item, a requester records the XDR stream offset of that data item in the Read chunk's Position field. The responder can then tell unambiguously where that chunk is to be re-inserted into the received Payload stream to form a complete RPC Call.
Each RDMA_MSG or RDMA_NOMSG procedure has one "Write list." The Write list is a list of zero or more Write chunks, provided by the requester. Each Write chunk is an array of plain segments, thus the Write list is a list of counted arrays.
If an RPC Reply has no possible DDP-eligible result data items, the requester leaves the Write list empty. When a requester provides a Write list, the responder MUST push data corresponding to DDP-eligible result data items to requester memory referenced in the Write list. The responder removes these data items from the reply's Payload stream.
A requester constructs the Write list for an RPC transaction before the responder has formulated its reply. When there is only one DDP-eligible result data item, the requester inserts only a single Write chunk in the Write list. If the returned Write chunk is not an unused Write chunk, the requester knows with certainty which result data item is contained in it.
When a requester has provided multiple Write chunks, the responder fills in each Write chunk with one DDP-eligible result until either there are no more DDP-eligible results, or no more Write chunks.
The requester might not be able to predict in advance which DDP-eligible data item goes in which chunk. Thus the requester is responsible for allocating and registering Write chunks large enough to accommodate the largest result data item that might be associated with each chunk in the Write list.
As a requester decodes a reply Payload stream, it is clear from the contents of the Reply which Write chunk contains which result data item.
There are occasions when a requester provides a non-empty Write chunk but the responder is not able to use it. For example, an Upper Layer Protocol may define a union result where some arms of the union contain a DDP-eligible data item while other arms do not. The responder is required to use requester-provided Write chunks in this case, but if the responder returns a result that uses an arm of the union that has no DDP-eligible data item, that Write chunk remains unconsumed.
If there is a subsequent DDP-eligible result data item in the Reply, it MUST be placed in that unconsumed Write chunk. Therefore the requester MUST provision each Write chunk so it can be filled with the largest DDP-eligible data item that can be placed in it.
If this is the last or only Write chunk available and it remains unconsumed, the responder MUST return this Write chunk as an unused Write chunk (see Section 3.4.6). The responder sets the segment count to a value matching the requester-provided Write chunk, but returns only empty segments in that Write chunk.
Unused Write chunks, or unused bytes in Write chunk segments, are returned to the RPC consumer as part of RPC completion. Even if a responder indicates that a Write chunk is not consumed, the responder may have written data into one or more segments before choosing not to return that data item. The requester MUST NOT assume that the memory regions backing a Write chunk have not been modified.
To force a responder to return a DDP-eligible result inline, a requester employs the following mechanism:
As always, a requester and responder must prepare for a Long Reply to be used if the resulting RPC Reply might be too large to be conveyed in an RDMA Send.
Each RDMA_MSG or RDMA_NOMSG procedure has one "Reply chunk." The Reply chunk is a Write chunk, provided by the requester. The Reply chunk is a single counted array of plain segments.
A requester MUST provide a Reply chunk whenever the maximum possible size of the reply message is larger than the inline threshold for messages from responder to requester. The Reply chunk MUST be large enough to contain a Payload stream (RPC message) of this maximum size. If the Transport stream and reply Payload stream together are smaller than the reply inline threshold, the responder MAY return it as a Short message rather than using the requester-provided Reply chunk.
When a requester has provided a Reply chunk in a Call message, the responder MUST copy that chunk into the associated Reply. The copied Reply chunk in the Reply is modified to reflect the actual amount of data that is being returned in the Reply chunk.
RDMA requires that data is transferred between only registered memory regions at the source and destination. All protocol headers as well as separately transferred data chunks must reside in registered memory.
Since the cost of registering and de-registering memory can be a significant proportion of the cost of an RPC-over-RDMA transaction, it is important to minimize registration activity. For memory that is targeted by RDMA Send and Receive operations, a local-only registration is sufficient and can be left in place during the life of a connection without any risk of data exposure.
Data transferred via RDMA Read and Write can reside in a memory allocation not in the control of the RPC-over-RDMA transport. These memory allocations can persist outside the bounds of an RPC transaction. They are registered and invalidated as needed, as part of each RPC transaction.
The requester endpoint must ensure that memory regions associated with each RPC transaction are properly fenced from responders before allowing Upper Layer access to the data contained in them. Moreover, the requester must not access these memory regions while the responder has access to them.
This includes memory regions that are associated with canceled RPCs. A responder cannot know that the requester is no longer waiting for a reply, and might proceed to read or even update memory that the requester might have released for other use.
The interface by which an Upper Layer Protocol implementation communicates the eligibility of a data item locally to its local RPC-over-RDMA endpoint is not described by this specification.
Depending on the implementation and constraints imposed by Upper Layer Bindings, it is possible to implement reduction transparently to upper layers. Such implementations may lead to inefficiencies, either because they require the RPC layer to perform expensive registration and de-registration of memory "on the fly", or they may require using RDMA chunks in reply messages, along with the resulting additional handshaking with the RPC-over-RDMA peer.
However, these issues are internal and generally confined to the local interface between RPC and its upper layers, one in which implementations are free to innovate. The only requirement, beyond constraints imposed by the Upper Layer Binding, is that the resulting RPC-over-RDMA protocol sent to the peer is valid for the upper layer.
The choice of which memory registration strategies to employ is left to requester and responder implementers. To support the widest array of RDMA implementations, as well as the most general steering tag scheme, an Offset field is included in each RDMA segment.
While zero-based offset schemes are available in many RDMA implementations, their use by RPC requires individual registration of each memory region. For such implementations, this can be a significant overhead. By providing an offset in each chunk, many pre-registration or region-based registrations can be readily supported.
A receiver performs basic validity checks on the RPC-over-RDMA header and chunk contents before it passes the RPC message to the RPC layer. If an incoming RPC-over-RDMA message is not as long as a minimal size RPC-over-RDMA header (28 bytes), the receiver cannot trust the value of the XID field, and therefore MUST silently discard the message before performing any parsing. If other errors are detected in the RPC-over-RDMA header of a Call message, a responder MUST send an RDMA_ERROR message back to the requester. If errors are detected in the RPC-over-RDMA header of a Reply message, a requester MUST silently discard the message.
To form an RDMA_ERROR procedure: The rdma_xid field MUST contain the same XID that was in the rdma_xid field in the failing request; The rdma_vers field MUST contain the same version that was in the rdma_vers field in the failing request; The rdma_proc field MUST contain the value RDMA_ERROR; The rdma_err field contains a value that reflects the type of error that occurred, as described below.
An RDMA_ERROR procedure indicates a permanent error. Receipt of this procedure completes the RPC transaction associated with XID in the rdma_xid field. A receiver MUST silently discard an RDMA_ERROR procedure that it cannot decode.
When a responder detects an RPC-over-RDMA header version that it does not support (currently this document defines only Version One), it MUST reply with an RDMA_ERROR procedure and set the rdma_err value to ERR_VERS, also providing the low and high inclusive version numbers it does, in fact, support.
A receiver might encounter an XDR parsing error that prevents it from processing the incoming Transport stream. Examples of such errors include an invalid value in the rdma_proc field, an RDMA_NOMSG message that has no chunk lists, or the contents of the rdma_xid field might not match the contents of the XID field in the accompanying RPC message. If the rdma_vers field contains a recognized value, but an XDR parsing error occurs, the responder MUST reply with an RDMA_ERROR procedure and set the rdma_err value to ERR_CHUNK.
When a responder receives a valid RPC-over-RDMA header but the responder's Upper Layer Protocol implementation cannot parse the RPC arguments in the RPC Call message, the responder SHOULD return an RPC Reply with status GARBAGE_ARGS, using an RDMA_MSG procedure. 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 RPC-over-RDMA Version One, 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:
RDMA operational errors are typically 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 RDMA_ERROR reply with an rdma_err value of ERR_CHUNK to indicate that no RPC-level reply is possible for that XID.
While a requester is constructing a 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 a Call message, 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 fence that memory from the responder before the associated RPC transaction is complete.
While a responder is constructing a 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 a Reply or error message, the responder MUST close the connection to signal to the requester that a reply was lost.
The RDMA connection and physical link provide some degree of error detection and retransmission. iWARP's Marker PDU Aligned (MPA) layer (when used over TCP), Stream Control Transmission Protocol (SCTP), as well as the InfiniBand link layer all provide Cyclic Redundancy Check (CRC) protection of the RDMA payload, and 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 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 Section 8 for further discussion of the use of RPC-level integrity schemes to detect errors.
The following protocol elements are no longer supported in RPC-over-RDMA Version One. Related enum values and structure definitions remain in the RPC-over-RDMA Version One protocol for backwards compatibility.
The specification of RDMA_MSGP in Section 3.9 of [RFC5666] is incomplete. To fully specify RDMA_MSGP would require:
The RDMA_MSGP message type is beneficial only when the padded data payload is at the end of an RPC message's argument or result list. This is not typical for NFSv4 COMPOUND RPCs, which often include a GETATTR operation as the final element of the compound operation array.
Without a full specification of RDMA_MSGP, there has been no fully implemented prototype of it. Without a complete prototype of RDMA_MSGP support, it is difficult to assess whether this protocol element has benefit, or can even be made to work interoperably.
Therefore, senders MUST NOT send RDMA_MSGP procedures. When receiving an RDMA_MSGP procedure, responders SHOULD reply with an RDMA_ERROR procedure, setting the rdma_err field to ERR_CHUNK; requesters MUST silently discard the message.
Because no implementation of RPC-over-RDMA Version One uses the Read-Read transfer model, there is never a need to send an RDMA_DONE procedure.
Therefore, senders MUST NOT send RDMA_DONE messages. Receivers MUST silently discard RDMA_DONE messages.
RPC-over-RDMA chunk lists are complex data types. In this section, illustrations are provided to help readers grasp how chunk lists are represented inside an RPC-over-RDMA header.
HLOO
A plain segment is the simplest component, being made up of a 32-bit handle (H), a 32-bit length (L), and 64-bits of offset (OO). Once flattened into an XDR stream, plain segments appear as
PHLOO
An RDMA read segment has an additional 32-bit position field. RDMA read segments appear as
1 PHLOO 1 PHLOO 1 PHLOO 0
A Read chunk is a list of RDMA read segments. Each RDMA read segment is preceded by a 32-bit word containing a one if a segment follows, or a zero if there are no more segments in the list. In XDR form, this would look like
The Read List is also a list of RDMA read segments. In XDR form, this would look like a Read chunk, except that the P values could vary across the list. An empty Read List is encoded as a single 32-bit zero.
3 HLOO HLOO HLOO
One Write chunk is a counted array of plain segments. In XDR form, the count would appear as the first 32-bit word, followed by an HLOO for each element of the array. For instance, a Write chunk with three elements would look like
1 3 HLOO HLOO HLOO 1 2 HLOO HLOO 0
The Write List is a list of counted arrays. In XDR form, this is a combination of optional-data and counted arrays. To represent a Write List containing a Write chunk with three segments and a Write chunk with two segments, XDR would encode
1 2 HLOO HLOO
The Reply chunk is a Write chunk. Since it is an optional-data field, however, there is a 32-bit field in front of it that contains a one if the Reply chunk is present, or a zero if it is not. After encoding, a Reply chunk with 2 segments would look like
Frequently a requester does not provide any chunks. In that case, after the four fixed fields in the RPC-over-RDMA header, there are simply three 32-bit fields that contain zero.
In setting up a new RDMA connection, the first action by a requester is to obtain a transport address for the responder. 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 normally register with a portmap or rpcbind service [RFC1833], which associates an RPC Program number with a service address. This policy is no different with RDMA transports. However, a different and distinct service address (port number) might sometimes be required for Upper Layer Protocol operation with RPC-over-RDMA.
When mapped atop the iWARP transport [RFC5040] [RFC5041], which uses IP port addressing due to its layering on TCP and/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 both iWARP/TCP and iWARP/SCTP [RFC5667].
When mapped atop InfiniBand [IB], which uses a Group Identifier (GID)-based service endpoint naming scheme, a translation MUST be employed. One such translation is defined in the InfiniBand Port Addressing Annex [IBPORT], 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:
Historically, different RPC protocols have taken different approaches to their port assignment. Therefore, the specific method is left to each RPC-over-RDMA-enabled Upper Layer Binding, and not addressed in this document.
In Section 9, this specification defines two new "netid" values, to be used for registration of upper layers atop iWARP [RFC5040] [RFC5041] and (when a suitable port translation service is available) InfiniBand [IB]. Additional RDMA-capable networks MAY define their own netids, or if they provide a port translation, MAY share the one defined in this document.
An Upper Layer Protocol is typically defined independently of any particular RPC transport. An Upper Layer Binding specification (ULB) provides guidance that helps the Upper Layer Protocol interoperate correctly and efficiently over a particular transport. For RPC-over-RDMA Version One, an Upper Layer Binding may provide:
Each RPC Program and Version tuple that utilizes RPC-over-RDMA Version One needs to have an Upper Layer Binding specification.
An Upper Layer Binding designates some XDR data items as eligible for Direct Data Placement. As an RPC-over-RDMA message is formed, DDP-eligible data items can be removed from the Payload stream and placed 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:
In addition to defining the set of data items that are DDP-eligible, an Upper Layer Binding may also limit the use of chunks to particular Upper Layer procedures. If more than one data item in a procedure is DDP-eligible, the Upper Layer Binding may also limit the number of chunks that a requester can provide for a particular Upper Layer procedure.
Senders MUST NOT reduce data items that are not DDP-eligible. Such data items MAY, however, be moved as part of a Position Zero Read chunk or a Reply chunk.
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 Upper Layer Binding are respected.
There is no provision to express DDP-eligibility within the XDR language. The only definitive specification of DDP-eligibility is an Upper Layer Binding.
In general a DDP-eligibility violation occurs when:
A requester provides resources for both a Call message and its matching Reply message. A requester forms the Call message itself, thus can compute the exact resources needed for it.
A requester must allocate resources for the Reply message (an RPC-over-RDMA credit, a Receive buffer, and possibly a Write list and Reply chunk) before the responder has formed the actual reply. To accommodate all possible replies for the procedure in the Call message, a requester must allocate reply resources based on the maximum possible size of the expected Reply message.
If there are procedures in the Upper Layer Protocol for which there is no clear reply size maximum, the Upper Layer Binding needs to specify a dependable means for determining the maximum.
There may be other details provided in an Upper Layer Binding.
Each Upper Layer Binding needs to be designed to allow correct interoperation without regard to the transport parameters actually in use. Furthermore, implementations of Upper Layer Protocols must be designed to interoperate correctly regardless of the connection parameters in effect on a connection.
An RPC Program and Version tuple may be extensible. For instance, there may be a minor versioning scheme that is not reflected in the RPC version number. Or, the Upper Layer Protocol may allow additional features to be specified after the original RPC program specification was ratified.
Upper Layer Bindings are provided for interoperable RPC Programs and Versions by extending existing Upper Layer Bindings to reflect the changes made necessary by each addition to the existing XDR.
The RPC-over-RDMA header format is specified using XDR, unlike the message header used with RPC over TCP. To maintain a high degree of interoperability among implementations of RPC-over-RDMA, any change to this XDR requires a protocol version number change. New versions of RPC-over-RDMA may be published as separate protocol specifications without updating this document.
The first four fields in every RPC-over-RDMA header must remain aligned at the same fixed offsets for all versions of the RPC-over-RDMA protocol. The version number must be in a fixed place to enable implementations to detect protocol version mismatches.
For version mismatches to be reported in a fashion that all future version implementations can reliably decode, the rdma_proc field must remain in a fixed place, the value of ERR_VERS must always remain the same, and the field placement in struct rpc_rdma_errvers must always remain the same.
Introducing new capabilities to RPC-over-RDMA Version One is limited to the adoption of conventions that make use of existing XDR (defined in this document) and allowed abstract RDMA operations. Because no mechanism for detecting optional features exists in RPC-over-RDMA Version One, implementations must rely on Upper Layer Protocols to communicate the existence of such extensions.
Such extensions must be specified in a Standards Track document with appropriate review by the nfsv4 Working Group and the IESG. An example of a conventional extension to RPC-over-RDMA Version One is the specification of backward direction message support to enable NFSv4.1 callback operations, described in [I-D.ietf-nfsv4-rpcrdma-bidirection].
A primary consideration is the protection of the integrity and confidentiality of local 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 to memory owned by user processes.
It is REQUIRED that any RDMA provider used for RPC transport be conformant to the requirements of [RFC5042] in order to satisfy these protections. These protections are provided by the RDMA layer specifications, and in particular, their security models.
The use of Protection Domains to limit the exposure of memory regions to a single connection is critical. Any attempt by an endpoint not participating in that connection to re-use memory handles needs to result in immediate failure of that connection. Because Upper Layer Protocol security mechanisms rely on this aspect of Reliable Connection behavior, strong authentication of remote endpoints is recommended.
Unpredictable memory handles should be used for any operation requiring advertised memory regions. Advertising 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 under way. Therefore implementations should avoid advertising persistently registered memory.
Requesters should register memory regions for remote access only when they are about to be the target of an RPC operation that involves an RDMA Read or Write.
Registered memory regions should be invalidated as soon as related RPC operations are complete. Invalidation and DMA unmapping of memory regions should be complete before message integrity checking is done, and before the RPC consumer is allowed to continue execution and use or alter the contents of a memory region.
An RPC transaction on a requester might be terminated before a reply arrives if the RPC consumer exits unexpectedly (for example it is signaled or a segmentation fault occurs). When an RPC terminates abnormally, memory regions associated with that RPC should be invalidated appropriately before the regions are released to be reused for other purposes on the requester.
A detailed discussion of denial of service exposures that can result from the use of an RDMA transport is found in Section 6.4 of [RFC5042].
A responder is not obliged to pull Read chunks that are unreasonably large. The responder can use an RDMA_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 RNICs typically terminate the connection. For further discussion, see Section 4.5. Such repeated chunk errors can deny service to other users sharing the connection from the errant requester.
An RPC-over-RDMA transport implemention is not responsible for throttling the RPC request rate, other than to keep the number of concurrent RPC transactions at or under the number of credits granted per connection. This is explained in Section 3.3.1. A sender can trigger a self denial of service by exceeding the credit grant repeatedly.
When an RPC has been canceled due to a signal or premature exit of an application process, a requester may 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 memory has been reused.
On the requester, a malfunctioning application or a malicious user can create a situation where RPCs are continuously initiated and then aborted, 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.
ONC RPC provides cryptographic security via the RPCSEC_GSS framework [RFC7861]. 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 the layer above RPC-over-RDMA. The latter two services require significant computation and movement of data on each endpoint host. Some performance benefits enabled by RDMA transports can be lost.
Performance loss is expected when RPCSEC_GSS integrity or privacy services are in use on any RPC transport. Protection below the RPC transport is often more appropriate in performance-sensitive deployments, especially if it, too, can be offloaded. Certain configurations of IPsec can be co-located in RDMA hardware, for example, without change to RDMA consumers and 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.
The use of protection in a lower layer MAY be negotiated through the use of an RPCSEC_GSS security flavor defined in [RFC7861] in conjunction with the Channel Binding mechanism [RFC5056] and IPsec Channel Connection Latching [RFC5660]. Use of such mechanisms is REQUIRED where integrity or confidentiality is desired and where efficiency is required.
Not all RDMA devices and fabrics support the above protection mechanisms. Also, per-message authentication is still required on NFS clients where multiple users access NFS files. In these cases, RPCSEC_GSS can protect NFS traffic conveyed on RPC-over-RDMA connections.
RPCSEC_GSS extends the ONC RPC protocol [RFC5531] 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.
As part of the ONC RPC protocol, protocol elements of RPCSEC_GSS that appear in the Payload stream of an RPC-over-RDMA message (such as control messages exchanged as part of establishing or destroying a security context, or data items that are part of RPCSEC_GSS authentication material) MUST NOT be reduced.
Some NFS client implementations use a separate connection to establish a GSS context for NFS operation. These clients use TCP and the standard NFS port (2049) for context establishment. To enable the use of RPCSEC_GSS with NFS/RDMA, an NFS server MUST also provide a TCP-based NFS service on port 2049.
The RPCSEC_GSS authentication service has no impact on the DDP-eligibity of data items in an Upper Layer Protocol.
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 Call messages, and to provide for a maximum size RPCSEC_GSS verifier when allocating reply buffers and Reply chunks.
RPC messages, and thus Payload streams, are made larger as a result. Upper Layer Protocol operations that fit in a Short Message when a simpler form of authentication is in use might need to be reduced, or conveyed via a Long Message, when RPCSEC_GSS authentication is in use. It is more likely that a requester provides both a Read list and a Reply chunk in the same RPC-over-RDMA header to convey a Long call and provision a receptacle for a Long reply. More frequent use of Long messages can impact transport efficiency.
The RPCSEC_GSS integrity service enables endpoints to detect 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 for transmit, and before they have been XDR decoded after receipt. Both sender and receiver endpoints use intermediate buffers to prevent exposure of encrypted data or unverified cleartext data to RPC consumers. After verification, encryption, and message wrapping has been performed, 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 the encoded XDR stream. XDR padding of a reduced data item is not transferred in an RPC-over-RDMA message. After reduction, the Payload stream contains fewer octets then 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 or 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, and Chunked Messages cannot be used. In this mode, an RPC-over-RDMA transport operates in the same manner as a transport that does not support direct data placement.
When a RPCSEC_GSS integrity or privacy service is in use, a requester provides both a Read list and a Reply chunk in the same RPC-over-RDMA header to convey a Long call and provision a receptacle for a Long reply.
Like the base fields in an ONC RPC message (XID, call direction, and so on), the contents of an RPC-over-RDMA message's Transport stream are not protected by RPCSEC_GSS. This exposes XIDs, connection credit limits, and chunk lists (but not the content of the data items they refer to) to malicious behavior, which could 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 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 similar effect. If such alterations occurs, the use of RPCSEC_GSS integrity or privacy services enable a requester to detect unexpected material in a received RPC message.
Encryption at lower layers, as described in Section 8.2.1, protects the content of the Transport stream. To address attacks on RDMA protocols themselves, RDMA transport implementations should conform to [RFC5042].
A set of RPC "netids" for resolving RPC-over-RDMA services is specified by this document. This is unchanged from [RFC5666].
The RPC-over-RDMA transport has been assigned an RPC "netid", which is an rpcbind [RFC1833] string used to describe the underlying protocol in order for RPC to select the appropriate transport framing, as well as the format of the service addresses and ports.
NC_RDMA "rdma" NC_RDMA6 "rdma6"
The following "netid" registry strings are defined for this purpose:
The "rdma" netid is to be used when IPv4 addressing is employed by the underlying transport, and "rdma6" for IPv6 addressing. The netid assignment policy and registry are defined in [RFC5665].
These netids MAY be used for any RDMA network satisfying the requirements of Section 2.2.2, and able to identify service endpoints using IP port addressing, possibly through use of a translation service as described in Section 5.
The use of the RPC-over-RDMA protocol has no effect on RPC Program numbers or existing registered port numbers. However, new port numbers MAY be registered for use by RPC-over-RDMA-enabled services, as appropriate to the new networks over which the services will operate.
For example, the NFS/RDMA service defined in [RFC5667] has been assigned the port 20049 in the IANA registry. This is distinct from the port number defined for NFS on TCP, which is assigned the port 2049 in the IANA registry. NFS clients use the same RPC Program number for NFS (100003) when using either transport [RFC5531].
[RFC5666] was listed as the reference for the nfsrdma port assignments. This document updates [RFC5666], but neither this document nor [RFC5666] specifies these port assignments. Therefore this document should not be listed as the reference for the nfsrdma port assignments.
The following alterations have been made to the RPC-over-RDMA Version One specification. The section numbers below refer to [RFC5666].
Although the protocol described herein interoperates with existing implementations of [RFC5666], the following changes have been made relative to the protocol described in that document:
The editor gratefully acknowledges the work of Brent Callaghan and Tom Talpey on the original RPC-over-RDMA Version One specification [RFC5666].
Dave Noveck provided excellent review, constructive suggestions, and consistent navigational guidance throughout the process of drafting this document. Dave also contributed much of the organization and content of Section 7 and helped the authors understand the complexities of XDR extensibility.
The comments and contributions of Karen Deitke, Dai Ngo, Chunli Zhang, Dominique Martinet, and Mahesh Siddheshwar are accepted with great thanks. The editor also wishes to thank Bill Baker, Greg Marsden, and Matt Benjamin for their support of this work.
The extract.sh shell script and formatting conventions were first described by the authors of the NFSv4.1 XDR specification [RFC5662].
Special thanks go to Transport Area Director Spencer Dawkins, nfsv4 Working Group Chair and document shepherd Spencer Shepler, and nfsv4 Working Group Secretary Thomas Haynes for their support.