]> Hewlett Packard Enterprise

165 Dascomb Road Andover MA 01810 USA +1 781-572-8038 davenoveck@gmail.com

Transport Network File System Version 4 It is expected that the RPC-over-RDMA transport will, at some point, allow protocol extensions to be defined. This would provide for the specification of OPTIONAL features allowing participants who implement the OPTIONAL features to cooperate as specified by that extension, while still interoperating with participants who do not support that extension. A particular extension is described herein, whose purpose is to reduce the latency due to inter-node round-trips needed to effect operations which involve direct data placement or which transfer RPC messages longer than the fixed inline buffer size limit.

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 .

This document describes a potential extension to the RPC-over-RDMA protocol, which would allow participating implementations to have more flexibility in how they use RDMA sends and receives to effect necessary transmission of RPC requests and replies. In contrast to existing facilities defined in RPC-over-RDMA Version One in which the mapping between RPC messages and RPC-over-RDMA messages is strictly one-to-one and DDP is effected only through use of explicit RDMA operations, the following features are made available through this extension: The ability to effect Direct Data Placement in the context of a single RPC-over-RDMA transmission, rather than requiring explicit RDMA operations to effect the necessary placement. The ability to continue an RPC request or reply over multiple RPC-over-RDMA transmissions

This document is written assuming that certain underlying facilities will be made available to build upon, in the context of a future version of RPC-over-RDMA. It is most likely that such facilities will be first available in Version Two. A protocol extension mechanism is needed to enable the extensions to RPC-over-RDMA described here. This document is currently written to conform to the extension model for the proposed RPC-over-RDMA Version Two as described in . An existing means of communicating transport characteristics between the RPC-over-RDMA endpoints is assumed. This document is currently written assuming the transport characteristic model defined in will be available and can be extended to meet the needs of this extension. As the documents referred to above are currently personal Internet rafts, and subject to change, adjustments to this document are expected to be necessary when and if, the needed facilities are defined in working group documents.

A number of different terms are used regarding the roles of the two participants in an RPC-over-RMA connection. Some of these roles last for the duration of a connection while others vary from request to request or from message to message. The roles of the client and server are fixed for the lifetime of the connection, with the client defined as the endpoint which initiated the connection. The roles of requester and responder often parallel those of client and server, although this is not always the case. Most requests are made in the forward direction, in which the client is the requester and the server is the responder. However, backward direction requests are possible, in which case the server is the requester and the client is the responder. As a result clients and servers may both act as requesters and responders for different requests issued on the same connection. The roles of sender and receiver vary from message to messages. With regard to the messages described in this document, the sender may act as a requester by sending RPC requests or a responder by sending RPC requests or as both at the same time by sending a mix of the two.

This extension is intended to function as part of RPC-over-RDMA and implementations should successfully interoperate with existing RPC-over-RDMA Version One implementations. Nevertheless, this extension seeks to take a somewhat different approach to high-performance RPC operation than has been used previously in that it seeks to de-emphasize the use of explicit RDMA operations. It does this in two ways: By implementing a send-based form of Direct Data Placement (see ), use of explicit RDMA operations can be avoided in many common cases in which data is directly placed. Use of explicit RDMA to support reply chunks and position-zero read chunks can be avoided by allowing a single message to be split into multiple transmissions. This can be used to avoid many instances of the only existing use of explicit RDMA operations not associated with Direct Data Placement. While use of explicit RDMA operations allows the cost of the actual data transfer to be offloaded from the client and server CPUs to the RNIC, there are ancillary costs in setting up the transfer that cannot be ignored. As a result send-based functions are often preferable, since the RNIC also uses DMA to effect these operations. In addition, the cost of the additional inter-node round trips required by explicit RDMA operation can be an issue, which can becomes increasingly troublesome as internode distances increase. Once one moves from in-machine-room to campus-wide or metropolitan-area distances the additional round-trip delay of 16 microseconds per mile becomes an issue impeding use of explicit RDMA operations.

Although explicit RDMA operations are used in the existing RPC-over-RDMA protocol for purposes unrelated to Direct Data Placement, all DDP is effected using explicit RDMA operations. As a result, all operations involving Direct Data Placement involve multiple internode round trips.

Although this was not stated explicitly, it appears that RDMA_MSGP (defined in , removed from RPC-OVER-RDMA by ), was an early attempt to effect correct placement of bulk data within a single RPC-over-RDMA transmission. As things turned out, the fields within the RDMA_MSGP header were not described in in a way that allowed this message type to be implemented. In attempting to provide DDP functionality, we have to keep in mind and avoid the problems that led to failure of RDMA_MSGP. It appears that the problems go deeper than neglecting to write a few relevant sentences. It is helpful to note that: The inline message size limits eventually adopted were too small to allow RDMA_MSGP to be used effectively. This is true of both the 1K limit in Version One and the 4K limit specified in . On the other hand, there is text within that suggests that much longer messages were anticipated at some points during the evolution of RPC-over-RDMA. The fact that NFSv4 COMPOUNDs often have additional operations beyond the one including the bulk data means that the RDMA_MSGP model cannot be extended to NFSv4. As a result, the bulk data needs to be excised from the data stream just as chunks are, so that the payload stream can include non-bulk data both before and after the logical position of the excised bulk data. In order for the sender to determine the appropriate amount of padding necessary within a transmission to place the bulk data at the proper position within receive buffer, the server must know more about the structure of the receiver's buffers. Since the padding needs to bring the bulk data to a position within the buffer that is appropriate to receive the bulk data, the sender needs to know where within the receive buffers such DDP-eligible areas are located. While appropriate padding could place the bulk data within a large WRITE into an appropriately aligned buffer or set of buffer, there is no corresponding provision for the bulk data associated with a READ. In short, there is no way to indicate to the responder that it should use RDMA_MSGP to appropriately place bulk data in the response. There is no explicit discussion of the required padding's use in effecting proper data placement or connection with the ULB's specification of DDP-eligible XDR. To summarize, RDMA_MSGP was an attempt to properly place data which was thought of as a local optimization and insufficient attention was given to it to make it successful. As a result, as RPC-over-RDMA Version One was developed, Direct Data Placement was identified with the use of explicit RDMA operations, and the possibility of Data Placement within sends was not recognized.

In the sections below, we will attempt to provide send-based data placement in a more complete way. By defining the structure of receive buffers as a transport characteristic. By treating positioning of bulk data within a message as an instance of DDP, causing the bulk data to be excised from the payload XDR stream, as is the case with other forms of DDP. By defining new DDP control data structures that support both send-based DDP and the form of DDP using explicit RDMA operations that was specified in RPC-over-RDMA Version One. These new control structures, described in are organized differently from the chunk-based structures described in .

In order to support send-based DDP, new DDP-related data structures have been defined, as described in Sections and . These new data structures support both send-based and RDMA-operation-based DDP. In addition, because of the restructuring described in , a number of additional facilities are made available: The ability to restrict entries regarding DDP in response data to XDR data items generated in response to performing particular constituent operations within a given RPC request (e.g. operations within an NFSv4 COMPOUND). The ability to make use of DDP contingent on the actual length of a DDP-eligible data item in the response. The ability to specify whether use of DDP for a particular DDP-eligible data item is required or optional. These additional facilities will be available to implementations that do not support send-based DDP, as long as both parties support the OPTIONAL Header types that include these new structures. For more information about the relationships among, the new transport characteristics, operations, and features, see .

Within RPC-over-RDMA Version One , Each transmission of a request or reply involves sending a single RDMA send message and conversely each message-related transmission involves only a single RPC request or reply. This strict one-to-one model leads to some potential performance issues. Because of RDMA's use of fixed-size receives, some requests and replies will inevitably not fit in the limited space available, even if they do not contain any DDP-eligible bulk data. Such cases will raise performance issues because, to deal with them, the server is interrupted twice to receive a single request and all the necessary transfers are serialized. In particular, there are two server interrupt latencies involved before the server can process the actual request, in addition to the OTW round-trip latencies. In the case of replies, there may be cases in which reply chucks need to be allocated and registered even if the actual reply would fit within the fixed receive-size limit. Because the decision to create a reply chunk is made at the time the request is sent, even an extremely low probability of a longer reply will trigger allocation of a reply chunk. Because this decision is made in conformance with ULB rules, which, by their nature, may only reference a limited set of data, a reply chunk may be required even when the actual probability of a long reply is exactly zero. For example a GETATTR request can generate a long reply due to a long ACL, and thus COMPOUND with this operation might allocate a reply chunk, even if the specific file system being interrogated only supports ACLs of limited sizes, or the GETATTR in question does not interrogate one of the ACL attributes. Also, the OWNER attribute is a string and may be impossible to determine a priori that the owner of any particular file has no chance of requiring more than 4K bytes of space, for example. The fact there are no such user names, while it is probably is true, is not a fact that RPC-over-RDMA implementations can depend on.

Continuing a single RPC request or reply is addressed by defining separate optional header types to begin and to continue sending a single RPC message. This is rather than creating a header with a continuation bit. In this approach, all of the DDP-related fields, which include support for send-based DDP, appear in the starting header (of typed ROPT_XMTREQ and ROPT_XMTRESP) and apply to the RPC message as a whole. Later RPC-over-RDMA messages (of type ROPT_XMTCONT) may extend the payload stream and/or provide additional buffers to which bulk data can be directed. In this case, all of the RPC-over-RDMA messages used together are referred to as a transmission group and must be received in order without any intervening message. In implementations using this optional facility, those decoding RPC messages received using RPC-over-RDMA no longer have the assurance that that each RPC message is in a contiguous buffer. As most XDR implementations are built based on the assumption that input will not be contiguous, this will not affect performance in most cases.

Using multiple transmissions to send a single request or response can complicate credit management. In the case of the message continuation feature, deadlocks can be avoided because use of message continuation is not obligatory. The requester or responder can use explicit RDMA operations if sufficient credits to use message continuation are not available. A requester is well positioned to make this choice with regard to the sending of requests. The requester must know, before sending a request, how long it will be, and therefore, how many credits it would require to send the request using message continuation. If these are not available, it can avoid message continuation by either creating read chunks sufficient to make the payload stream fit in a single transmission or by creating a position-zero read chunk. With regard to the response, the requester is not in position to know exactly how long the response will be. However, the ULB will allow the maximum response length to be determined based on the request. This value can be used: To determine the maximum number of receive buffers that might be required to receive any response sent. To allocate and register a reply chunk to hold a possible large reply. The requester can avoid doing the second of these if the responder has indicated it can use message continuation to send the response. In this case, it makes sure that the buffers will be available and indicates to the responder how many additional buffers (in the form of pre-posted reads have been made available to accommodate continuation transmissions. When the responder processes the request, those additional receive buffers may be used or not, or used only in part. This may be because the response is shorter than the maximum possible response, or because a reply chunk was used to transmit the response. After the first or only transmission associated with the response is received by the requester, it can be determined how many of the additional buffers were used for the response. Any unused buffers can be made available for other uses such as expanding the pool of receive buffers available for the initial transmissions of response or for receiving opposite direction requests. Alternatively, they can be kept in reserve for future uses, such as being made available to future requests which have potentially long responses.

In using existing RPC-over-RDMA facilities for protocol extension, interoperability with existing implementations needs to be assured. Because this document describes support for multiple features, we need to clearly specify the various possible extensions and how peers can determine whether certain facilities are supported by both ends of the connection.

Note that most of the new operations defined in this extension are not tightly tied to a specific feature. XOPT_XMTREQ and XOPT_XMTRESP are designed to support implementations that support either or both Send-based DDP or message continuation. However, the converse is not the case and these header types can be implemented by those not supporting either of these features. For example, implementations may only need support for the facilities described in . Implementations may determine whether a peer implementation supports XOPT_XMTREQ, XOPT_XMTREQ, or XOPT_XMTCONT by attempting these operations. An alternative is to interrogate the Rtr Support Characteristic for information about which operations are supported.

Implementations may determine and act based on the level of peer implementation of support for message continuation as follows: To deal with issues relating to sending the peer multi-transmission requests, the requester can interrogate the peer's value of the Request Transmission Receive Limit (). In cases in which the characteristic is not provided or has the value one, the requester implementation can avoid sending multi-transmission requests, and use the equivalent of position-zero read chunks to convey a request larger than the receive buffer limit. Similarly, if the request is longer than can fit in a set of transmissions given that limit, the request can be conveyed in the same fashion, To deal with issues relating to sending the peer multi-transmission responses, responders will only send multi-transmission responses for requests conveyed using XOPT_XMTREQ where the number of response transmissions is less than or equal to buffer reservation count (in the field optxrq_rsbuf). The requester can avoid receiving a message consisting of too many transmissions by setting this field appropriately. This includes the case in which the requester cannot handle any multi-transmission responses. To avoid reserving receive buffers that the responder is not prepared to use, the requester can interrogate the peer's value of the Response Transmission Send Receive Limit (). In cases in which it is possible that a request might result in a response too large for this set of buffers, the requester, the requester can provide a reply chunk to receive the response, which the responder can use if the count of buffers provided is insufficient.

Implementations may determine and adapt to the level of peer implementation support for send-based DDP as described below. Note that an implementation may be able to send messages containing bulk data items placed using send-based DDP while not being prepared to receive them, or the reverse. The requester can interrogate the responder's Receive Buffer Structure Characteristic. In cases in which the characteristic is not provided or shows no DDP-targetable buffer segments, an implementation knows that messages containing bulk data may not be sent using send-based DDP. In such cases, when XOPT_XMTREQ is used to send a request, bulk items may be transferred by setting the associated DDP information to indicate that the bulk data is to be fetched using explicit RDMA operations. In cases in which a requester is unprepared to accept messages using send-based DDP, its Receive Buffer Structure Characteristic will make this clear to the responder. Nevertheless, the requester will generally indicate to the responder that bulk data items are to be returned using explicit RDMA operations. As a result, requesters may use XOPT_XMTREQ (and get the benefit of the DDP-related features discussed in even if they support neither message continuation nor send-based DDP. Since it is possible for a responder to generate responses containing bulk data using send-based DDP even if it is not prepared to send such message, a requester who is prepared to accept such messages should specify in the request that the responses are to contain (or may contain) bulk data placed in this way. In deciding whether this is to be done the requester can interrogate the responder's RTR Support characteristic for information about which whether the peer can send responses in this form. It can do this without regard to whether the responder can accept messages containing bulk data items placed using send-based DDP. In determining whether bulk data will be placed using send-based DDP or via explicit RDMA operations, the level of support for message continuation will have a role. This is because DDP using explicit RDMA will reduce message size while send-based DDP reduces the size of the payload stream by rearranging the message, leaving the message size the same. As a result, the considerations discussed in will have to be attended to by the sender in determining which form of DDP is to be used.

The more extensive transport layer functionality described in this document requires its own means of reporting errors, to deal with issues that are distinct from: Errors (including XDR errors) in the XDR stream as received by responder or requester. XDR errors detected in the XDR headers defined by the base protocol. XDR errors detected in the new operations defined in this document. Beyond the above, the following sorts of errors will have to be dealt with, depending on which of the features of the extension are implemented. Information associated with send-based DDP may be inconsistent or otherwise invalid, even though it conforms to the XDR definition. There may be problems with the organization of transmission groups in that there are missing or extraneous transmissions. In each of the above cases, the problem will be reported to the sender using the Transmit Error operation which needs to be supported by every endpoint that sends ROPT_XMITREQ, ROPT_XMITRESP, or ROPT_XMITCONT. This includes cases in which the problem is one with a reply. The function of the Transmit Error operation is to aid in diagnosing transport protocol errors and allowing the sender to recover or decide recovery is not possible. Reporting failure to the requesting process is dealt with indirectly. For example, When the transmissions used to send a request are ill-formed, the requestor can respond to the error indication by proceeding to send the request using existing (i.e. non-extended) facilities. If it chooses not to do so, the requestor can report an RPC request failure to the initiator of the RPC. When the transmissions used to send a response are ill-formed, the responder need to know about the problem since it will otherwise assume that the transmissions succeeded. It can proceed to resend the reply using existing (i.e. non-extended) facilities. If it chooses not to do so, the requester will not see a response and eventually an RPC timeout will occur.

In order to implement message continuation, we have occasion to refer to particular RPC-over-RDMA transmissions within a transmission group or to characteristics if a later transmission group.

<CODE BEGINS> typedef uint32 xms_grpxn; typedef uint32 xms_grpxc; struct xms_id { uint32 xmsi_xid; msg_type xmsi_dir; xms_grpxn xmsi_seq; } <CODE ENDS>

An xms_grpxn designates a particular RPC-over-RDMA transmission within a set of transmissions devoted to sending a single RPC message. An xms_grpxc specifies the number of RPC-over-RDMA transmissions in a potential group of transmissions devoted to sending a single RPC message.

Data structures related to data placement use a number of XDR typedefs to help clarify the meaning of fields in the data structures which use these typedefs.

<CODE BEGINS> typedef uint32 xmddp_itemlen; typedef uint32 xmddp_pldisp; typedef uint32 xmddp_vsdisp; typedef uint32 xmddp_tbsn; enum xmddp_type { XMDTYPE_EXRW = 1, XMDTYPE_TBSN = 2, XMDTYPE_CHOOSE = 3, XMDTYPE_BYSIZE = 4, XMDTYPE_TOOSHORT = 5, XMDTYPE_NOITEM = 6 }; <CODE ENDS>

An xmddp_itemlen specifies the length of XDR item. Because items excised from the XDR stream are XDR items, lengths of items excised from the XDR stream are denoted by xmddp_itemlens. An xmddp_pldisp specifies a specific displacement with the payload stream associated with a single RPC-over-RDNA transmission or a group of such transmissions. Note that when multiple transmissions are used for a single message, all of the payload streams within a transmission group are considered concatenated. An xmddp_vsdisp specifies a displacement within the virtual XDR stream associates with the set of RPC messages transferred by single RPC-over-RDNA transmission or a group of such transmissions. The virtual XDR stream includes bulk data excised from the payload stream and so displacements within it reflect those of the corresponding objects in the XDR stream that might be sent and received if no bulk data excision facilities were involved in the RPC transmission. An xmddp_tbsn designates a particular target buffer segment within a (trivial or non-trivial) RPC-over-RDMA transmission group. Each DDP-targetable buffer segment is assigned a number starting with zero and proceeding through all the buffer segments for all the RPC-over-RDMA transmissions in the group. This includes buffer segments not actually used because transmission are shorter than the maximum size and those in which a DDP-targetable buffer segment is used to hold part of the payload XDR stream rather than bulk data. An xmddp_type allows a selection between DDP using explicit RDMA operations and that using send-based DDP. It is used in a number of contexts. The specific context governs which subset of the types is valid: In request messages, they indicate where each of the directly-placed data items within the request has been placed. In this case, xmddp_type appears as the discriminator within an xmddp_loc which is part of an xmddp_mitem that is an element within a request's optxrq_ddp field. In request messages, they direct the responder as to where potential directly-placed items are to be placed. In this case, xmddp_type appears as the discriminator within an xmddp_rsdloc which is part of an xmddp_rsditem that is an element within a request's optxrq_rsd field. In response messages, they indicate how each of the potential directly-placed items has been dealt with. A subset of these are directly-placed data items and are presented in the same form as that used for directly-placed data items within a request. In this case, xmddp_type appears as the discriminator within an xmddp_loc which is part of an xmddp_mitem that is an element within a response's optxrs_ddp field. A number of these type are valid in all of these contexts, since they specify use of a specific mode of direct placement which is to be used or has been used. XMDTYPE_EXRW selects DDP using explicit RDMA reads and writes. XMDTYPE_TBSN selects use of send-based DDP in which DDP-eligible data is located in DDP-targetable buffer segments. Another set of types is used to direct the use of specific sets of types but cannot specify an actual choice that has been made. XMDTYPE_CHOICE indicates that the responder may use either send-based DDP or chunk-based DDP using explicit RDMA operations, with a place for the latter having been provided by the requester. XMDTYPE_BYSIZE indicates that the responder is to use either send-based DDP or chunk-based DDP using explicit RDMA operations, with the choice between the two governed by the actual size of the associated DDP-eligible XDR item. The following types are used when no actual direct placement has occurred. They are used in responses to indicate ways in which a direction to govern DDP in a reply was responded to without resulting in direct placement. XMDTYPE_TOOSHORT indicates that the corresponding entry in an xmddp_rsdset was matched with a DDP-eligible item which was too small to be handled using direct placement, resulting in the DDP-eligible item being placed inline. XMDTYPE_NOITEM indicates that the corresponding entry in an xmddp_rsdset was not matched with a DDP-eligible item in the reply. The following table indicates which of the above types is valid in each of the contexts in which these types may appear. For valid occurrences, it distinguishes those which give sender-generated information about the message, and those that direct reply construction, from those that indicate how those directions governed the construction of a reply. For invalid occurrences, we distinguish between those that result in XDR decode errors and those which are valid from the XDR point of view but are semantically invalid. Type xmddp_loc in request xmddp_rsdloc in request xmddp_loc in response XMDTYPE_EXRW Valid Info Valid Direction Valid Result XMDTYPE_TBSN Valid Info Valid Direction Valid Result XMDTYPE_BYSIZE XDR Invalid Valid Direction XDR Invalid XMDTYPE_CHOICE XDR Invalid Valid Direction XDR Invalid XMDTYPE_TOOSHORT Sem. Invalid XDR Invalid Valid Result XMDTYPE_NOITEM Sem. Invalid XDR Invalid Valid Result

To understand the new DDP structure defined here, it is necessary to review the existing DDP structures used in RPC-over-RDMA Version One and look at the corresponding structures in the new message transmission headers defined in this document. We look first at the existing structures. Read chunks are specified on requests to indicate data items to be excised from the payload stream and fetched from the requester's memory by the responder. As such, they serve as a means of supplying data excised from the payload XDR stream. Read chunks appear in replies but they have no clear function there. Write chunks are specified on requests to provide locations in requester memory to which DDP-eligible items in the corresponding reply are to be transferred. They do not describe data in the request but serve to direct reply construction. When write chunks appear in replies they serve to indicate the length of the data transferred. The addresses to which the bulk reply data has been transferred is available, but this information is already known to the requester. Reply chunks are specified to provide a location in the requester's memory to which the responder can transfer the response using RDMA Write. Like write chunks, they do not describe data in the request but serve to direct reply construction. When reply chunks appear in reply message headers, they serve mainly to indicate whether the reply chunk was actually used. Within the DDP structures defined here a different organization is used, even where DDP using explicit RDMA operations in supported. All messages that contain bulk data contain structures that indicate where the excised data is located. See for details. Requests that might generate replies containing bulk data contain structures that provide guidance as to where the bulk data is to be placed. See for details. Both sets of data structure are defined at the granularity of an RPC-over-RDMA transmission group. That is, they describe the placement of data within an RPC message and the scope of description is not limited to a single RPC-over-RDMA transmission.

Buffer structure definition information is used to allow the sender to know how receive buffers are constructed, to allow it to appropriately pad messages being sent so that bulk data will be received into a memory area with the appropriate characteristics. In this case, Direct Data Placement will not place data in a specific address, picked and registered in advance as is done to effect DDP using explicit RDMA operations. Instead, a message is sent so that when it is matched with one of the preposted receives, the bulk data will be received into a memory area of the appropriate characteristics, including: size alignment DDP-targetability and potentially other memory characteristics such as speed, persistence.

<CODE BEGINS> struct xmrbs_seg { uint32 xmrseg_length; uint32 xmrseg_align; uint32 xmrseg_flags; }; const uint32 XMRSFLAG_DDP = 0x01; struct xmrbs_group { uint32 xmrgrp_count; xmrbs_seg xmrgrp_info; }; struct xmrbs_buf { uint32 xmrbuf_length; xmrbs_group xmrbuf_groups<>; }; <CODE ENDS>

Buffers can be, and typically are, structured to contain multiple segments. Preposted receives that target a buffer uses a scatter list to place received messages in successive buffer segments. An xmrbs_seg defines a single buffer segment. The fields included are: xmrseg_length is the length of this contiguous buffer segment xmrseg_align specifies the guaranteed alignment for the corresponding buffer segment. xmrseg_flags which specify some noteworthy characteristics of the associated buffer segment. The following flag bit is the only one currently defined: XMRSFLAG_DDP indicates that the buffer segment in question is to be considered suitable as a target for direct data placement. An xmrgs_group designates a set of buffer segment all with the same buffer segment characteristics as indicated by xmr_grpinfo. The buffer segments are contiguous within the buffer although they are likely not to be physically contiguous. An xmrbs_buf defines a receiver's buffer structure and consists of multiple xmrbs_groups. This buffer structure, when made available as a transport characteristic, allows the sender to structure transmissions so as to place DDP-eligible data in appropriate target buffer segments.

These data structures show where in the virtual XDR stream for the set of messages, data is to be excised from that XDR stream and where that excised bulk data is to be found instead.

<CODE BEGINS> union xmddp_loc switch(xmddp_type type) case XMDTPE_EXRW: rpcrdma1_segment xmdl_ex<>; case XMDTYPE_TBSN: xmddp_itemlen xmdl_offset; xmddp_tbsn xmdl_bsnum<>; case XMDTYPE_TOOSHORT: case XMDTYPE_NOITEM: void; }; struct xmddp_mitem { xmddp_vsdisp xmdmi_disp; xmddp_itemlen xmdmi_length; xmddp_loc xmdmi_where; }; typedef xmddp_mitem xmddp_grpinfo<>; <CODE ENDS>

An xmddp_loc shows where a particular piece of bulk data is located. This information exists in multiple forms. The case for DDP using explicit RDMA operations, contains, in xmdl_ex an array of rpcrdma1_segments showing where bulk data is to be fetched from or has been transferred to. The case for send-based DDP contains, in xmdl_tbsn an array DDP-targetable buffer segments, indicating where bulk data, excised from the payload stream, is actually located. The bulk data starts xmdl_offset bytes into the buffer segment designated by xmdl_bsnum[0] and then proceeds through buffer segments denoted by successive xmdl_bsnum entries until the length of the data item is exhausted. The cases for XMDDP_TOOSHORT and XMDDP_NPITEM are only valid in responses An xmddp_mitem denotes a specific item of bulk data. It consists of: The XDR stream displacement of the bulk data excised from the payload stream, in xmdmi_disp. The length of the data item, in xmdmi_length. The actual location of the bulk data, in xmdmi_loc. An xmddp_grpinfo consists of an array of xmddp_mitems describing all of the bulk data excised from all RPC messages sent in a single RPC-over-RDMA transmission group. Some possible cases: The array is of length zero, indicating that there is no DDP-eligible data excised from the virtual XDR stream. In this case, the virtual XDR stream and the payload stream are identical. The array consists of one or more xmddp_mitems, each of whose xmdmi_where fields is of type XMDTPE_EXRW. In this case, the DDP data corresponds to read chunks in the case in which a request is being sent and to write chunks in the case in which a reply is being sent. The array consists of one or more xmddp_mitems, each of whose xmdmi_where fields is of type XMDTPE_TBSN. In this case, each entry, whether it applies to bulk data in a request or a reply, describes data logically part of the message being sent, which may be part of any RPC-over-RDMA transmissions in the same transmission group. The array consists of one or more xmddp_mitems, with xmdmi_where fields of a mixture of types, In this case, each entry, whether it applies to bulk data in a request or a reply, describes data logically part of the message being sent, although the method of getting access to that data may vary from entry to entry.

These data structures, when sent as part of the request, instruct the responder how to use Direct Data Placement to place response data subject to direct data placement.

<CODE BEGINS> union xmddp_rsdloc switch(xmddp_type type) case XMDTPE_EXRW: case XMDTPE_CHOICE: rpcrdma1_segment xmdrsdl_ex<>; case XMDTPE_BYSIZE: xmddp_itemlen xmdrsdl_dsdov; rpcrdma1_segment xmdrsdl_bsex<>; case XMDTYPE_TBSN: void; }; struct xmddp_rsdrange { xmddp_vsdisp xmdrsdr_begin; xmddp_vsdisp xmdrsdr_end; }; struct xmddp_rsditem { xmddp_itemlen xmdrsdi_minlen; xmddp_rsdloc xmdrsdi_loc; }; struct xmddp_rsdset { xmddp_rsdrange xmdrsds_range; xmddp_rsditem xmdrsds_items<>; }; typedef xmddp_rsdset xmddp_rsdgroup<>; <CODE ENDS>

An xmddp_rsdloc contains information specifying where bulk data generated as part of a reply is to be placed. This information is defined as a union with the following cases: The case for DDP using explicit RDMA operations, XMDTYPE_EXRW, contains, in xmrsdl_ex, an array of rpcrdma1_segments showing where bulk data generated by the corresponding reply is to be transferred to. The case allowing the responder to freely choose the DDP method, XMDTYPE_CHOICE, is identical. It also contains, in xmrsdl_ex, an array of rpcrdma1_segments showing where bulk data generated by the corresponding reply is to be transferred to if explicit RDMA requests are to be used. The case for send-based ddp, XMDTYPE_TBSN. is void, since the decisions as to where bulk data is to be placed are made by the responder. In the case directing the responder to choose the DDP method based on item size, XMDTYPE_BYSIZE, an array of rpcrdma1_segments is in xmrsdl_bsex. In all cases, each xmddp_rsdloc sent as part of a request has a corresponding xmddp_loc in the associated response. The xmddp_type specified in the request will affect the type in the response, but the types are not necessarily the same. The table below describes the valid combinations of request and response xmdp_types. In this table, rows correspond to types in requests directing, the responder as to the desired placement in the response while the columns correspond to types in the ensuing response. Invalid combinations are labelled "Inv" while valid combination are labelled either "NDR" denoting no need to deregister memory, or "DR" to indicate that memory previously registered will need to be deregistered. Type EXRW TBSN TOOSHORT NOITEM EXRW DR Inv. DR DR TBSN Inv. NDR NDR NDR CHOICE DR NDR DR DR BYSIZE DR NDR DR DR An xmddp_rsdrange denotes a range of positions in the XDR stream associated with a request. Particular directions regarding bulk data in the corresponding response are limited to such ranges, where response XDR stream positions and request XDR stream positions can be reliably tied together. When the ULP supports multiple individual operations per RPC request (e.g., COMPOUND and CB_COMPOUND in NFSv4), an xmd_rsdrange can isolate elements of the reply due to particular operations. An xmddp_rsditem specifies the handling of one potential item of bulk data. The handling specified is qualified by a length range. If the item is smaller than xmdrsdi_minlen, it is not treated as bulk data and the corresponding data item appears in the payload stream, while that particular xmddp_rsditem is considered used up, making the next xmddp_rsditem in the xmddp_rsdset the target of the next DDP-eligible data item in the reply. Note that in the case in which xmdrsdi_loc specifies use of explicit RDMA operations, the area specified is not used and the requester is responsible for deregistering it. For each xmddp_rsditem, there will be a corresponding xmddp_mitem An xmddp_rsdset contains a set of xmddp_rsditems applicable to a given xmddp_range in the request. An xmddp_rsdgroup designates a set of xmddp_rsdsets applicable to a particular RPC-over-RDMA transmission group. The xmdrsds_range fields of successive xmddp_rsdsets must be disjoint and in strictly increasing order.

In this document we take advantage of the fact that the set of transport characteristics defined in is subject to later extension. The additional transport characteristics are summarized below in . In that table the columns have the following values: The column labeled "characteristic" identifies the transport characteristic described by the current row. The column labeled "#" specifies the xcharid value used to identify this characteristic. The column labeled "XDR type" gives XDR type of the data used to communicate the value of this characteristic. This data overlays the nominally opaque field xchv_data in an xcharval. The column labeled "default" gives the default value for the characteristic which is to be assumed by those who do not receive, or are unable to interpret, information about the actual value of the characteristic. The column labeled "section" indicates the section (within this document) that explains the semantics and use of this transport characteristic. characteristic # XDR type default section RTR Support 4 uint32 0 Receive Buffer Structure 5 xmrbs_buf Note1 Request Transmission Receive Limit 6 xms_grpxc 1 Response Transmission Send Limit 7 xms_grpxc 1 The following notes apply to the above table: The default value for the Receive Buffer Structure always consists of a single buffer segment, without any alignment restrictions and not targetable for DDP. The length of that buffer segment derives from the Receive Buffer Size characteristic if available, and from the default receive buffer size otherwise.

<CODE BEGINS> const uint32 XCHAR_RTRSUPP = 4; typedef uint32 xchrrtrs; const uint32 RTRS_XREQ = 1; const uint32 RTRS_XRESP = 2; const uint32 RTRS_XCONT = 4; <CODE ENDS>

This characteristic defines the structure of the endpoint's receive buffers, in order to give a sender the ability to place bulk data in specific DDP-targetable buffer segments.

<CODE BEGINS> const uint32 XCHAR_RBSTRUCT = 5; typedef xmrbs_buf xchrrbs; <CODE ENDS>

Normally, this characteristic, if specified should be in agreement with Receive Buffer Size characteristic. However, the following rules apply. If the value of Receive Buffer Structure characteristic is not specified, it is derived from the Receive Buffer Size characteristic, if known and the default buffer size otherwise. The buffer is considered to consist of a single non-DDP-targettable segment whose size is the buffer size. If the value of Receive Buffer Size characteristic is not specified and the Receive Buffer Structure characteristic is specified, the value of the former is derived from the latter, by adding up the length of all buffer segments specified.

This characteristic specifies the length of the longest request messages (in terms of number of transmissions) that a responder will accept.

<CODE BEGINS> const uint32 XCHAR_REQRXLIM = 6; typedef uint32 xchrqrxl; <CODE ENDS>

A requester can use this characteristic to determine whether to send long requests by use of message continuation or by using a position-zero read chunk.

This characteristic specifies the length of the longest response message (in terms of number of transmissions) that a responder will generate.

<CODE BEGINS> const uint32 XCHAR_RESPSXLIM = 7; typedef uint32 xchrssxl; <CODE ENDS>

The proposed new operation are set for in below. In that table, the columns have the following values: The column labeled "operation" specifies the particular operation. The column labeled "#" specifies the value of opttype for this operation. The column labeled "XDR type" gives XDR type of the data structure used to describe the information in this new message type. This data overlays the nominally opaque field optinfo in an RDMA_OPTIONAL message. The column labeled "msg" indicates whether this operation is followed (or not) by an RPC message payload (or something else). The column labeled "section" indicates the section (within this document) that explains the semantics and use of this optional operation. operation # XDR type msg section Transmit Request 5 optxmt_req Note1 Transmit Response 6 optxmt_resp Note1 Transmit Continue 7 optxmt_cont Note2 Transmit Error 8 optxmt_err No. The following notes apply to the above table: Contains an initial segment of the message payload stream for an RPC message, or the entre payload stream. The optxr[qs]_pslen field, indicates the length of the section present May contain a part of a message payload stream for an RPC message, although not the entre payload stream. The optxc_pslen field, if non-zero, indicates that this portion is present, and the length of the section.

The message definition for this operation is as follows:

<CODE BEGINS> const uint32 ROPT_XMTREQ = 5; struct optxmt_req { xmddp_grpinfo optxrq_ddp; xmddp_rsdgroup optxrq_rsd; xms_grpxc optxrq_count; xms_grpxc optxrq_rsbuf; xmddp_pldisp optxrq_pslen; }; <CODE ENDS>

The field optxrq_ddp describes the fields in virtual XDR stream which have been excised in forming the payload stream, and information about where the corresponding bulk data is located. The field optxrq_rsd consists of information directing the responder as to how to construct the reply, in terms of DDP. of length zero. The field optrq_count specifies the count of transmissions in this group of transmissions used to send a request. The field optrq_repch serves as a way to transfer a reply chunk to the responder to serve as a way in which a reply longer than the inline size limit may be transferred. Although, not prohibited by the protocol, it is unlikely to be used in environments in which message continuation is supported. The field optrq_pslen gives the length of the payload stream for the RPC transmitted. The payload stream begins right after the end of the optxmt_msg and proceeds for optxm_pslen bytes. This can include crossing buffer segment boundaries.

The message definition for this operation is as follows:

<CODE BEGINS> const uint32 ROPT_XMTRESP = 6; struct optxmt_resp { xmddp_grpinfo optxrs_ddp; xms_grpxn optxrs_count; xmddp_pldisp optxrs_pslen; }; <CODE ENDS>

The field optxrs_ddp describes the fields in virtual XDR stream which have been excised in forming the payload stream, and information about where the corresponding bulk data is located. The field optrs_count specifies the count of transmissions in this group of transmissions used to send a reply. The field optrq_pslen gives the length of the payload stream for the RPC transmitted. The payload stream begins right after the end of the optxmt_msg and proceeds for optxm_pslen bytes. This can include crossing buffer segment boundaries.

RPC-over-RDMA headers of this type are used to continue RPC messages begun by RPC-over-RDMA message of type ROPT_XMTREQ or ROPT_XMTRESP. The xid field of this message must match that in the initial transmission. This operation needs to be supported for the message continuation feature to be used. The message definition for this operation is as follows:

<CODE BEGINS> const uint32 ROPT_XMTCONT = 7; struct optxmt_cont { xms_grpxn optxc_xnum; uint32 optxc_itype; xmddp_pldisp; optxc_pslen; }; <CODE ENDS>

The field optxc_xnum indicates the transmission number of this transmission within its transmission group. The field optxc_pslen gives the length of the section of the payload stream which is located in the current RPC-over-RDMA transmission. It is valid for this length to be zero, indicating that there is no portion of the payload stream in this transmission. Except when the length is zero, the payload stream begins right after the end of the optxmt_cont and proceeds for optxc_pslen bytes. This can include crossing buffer segment boundaries. In any case, the payload streams for all transmissions within the same group are considered concatenated.

This RPC-over-RDMA message type is used to signal the occurrence of errors that do not involve: The preliminary error-related definition is as follows:

<CODE BEGINS> enum optx_err { OPTXERR_BADHMT = 1, OPTXERR_BADOMT = 2, OPTXERR_BADCONT = 3, OPTXERR_BADSEQ = 4, OPTXERR_BADXID = 5, OPTXERR_BADOFF = 6, OPTXERR_BADTBSN = 7, OPTXERR_BADPL = 8 } union optx_info switch(optx_err optxe_which) { case OPTXERR_BADHMT: case OPTXERR_BADOMT: case OPTXERR_BADSEQ: case OPTXERR_BADXID: uint32 optxi_expect; uint32 optxi_current; case OPTXERR_BADCONT: void; case OPTXERR_BADTBSN: case OPTXERR_BADOFF: case OPTXERR_BADPL: uint32 optxi_value; uint32 optxi_min; uint32 optxi_max; }; <CODE ENDS>

optx_err enumerates the various error conditions that might be reported. OPTXERR_BADHMT indicates that a header message type other than the one expected was received. In this context, a particular message type can be considered "expected" only because of message or group continuation. OPTXERR_BADOMT indicates that an optional message type other than the one expected was received. In this context, a particular message type can be considered "expected" only because of message or group continuation. OPTXERR_BADCONT indicates that a continuation messages was received when there was no reason to expect one. OPTXERR_BADSEQ indicate that a transmission sequence number other than the one expected was received. OPTXERR_BADXID indicate that an xid other than the one expected in a continuation context. OPTXERR_BADTBSN indicate that an invalid target buffer sequence number was received. OPTXERR_BADOFF indicate that a bad offset was received as part of an xmddp_loc. This is typically because the offset is larger than the buffer segment size. OPTXERR_BADPL indicates that a bad offset was received for the payload length. This is typically because the length would make the area devoted to the payload stream not a subset of the actual transmission. The optx_info gives error about the specfific invalid field being reported. The additional information given depends on the specific error. For the errors OPTXERR_BADHMT, OPTXERR_BADOMT, OPTXERR_BADSEQ, andd OPTXERR_BADXID, the expected and actual values of the field are reported For the error OPTXERR_CONT, no additional information is provided. For the errors OPTXERR_BADTBSN, OPTXERR_BADOFF, and OPTXERR_BADPL, the actual value together with a range of valid values is provided. When the actual value is with the valid range, it can be inferred that the actual value is not propely aligned (e.g. not on a 32-bit boundary) The message definition for this operation is as follows:

<CODE BEGINS> const uint32 ROPT_XMTERR = 8; struct optxmt_err { xms_id optxe_bad; xms_id *optxe_lead; optx_info optxe_info; }; <CODE ENDS>

The field optxe_bad is a description of the transmission on which the error was actually detected. The optional field optxe_lead is a description of an earlier transmission that might have led to the error reported. The field optxe_info provides informtion about the

This section contains an XDR description of the proposed extension. This description is provided in a way that makes it simple to extract into ready-to-use form. The reader can apply the following shell script to this document to produce a machine-readable XDR description of extension which can be combined with XDR for the base protocol to produce an XDR that incude the base protocol together with the optional extensions.

<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 "ext.txt" then the reader can do the following to extract an XDR description file for this extension:

<CODE BEGINS> sh extract.sh < ext.txt > xmitext.x <CODE ENDS>

The XDR descriptuion for this extension can be combined with tht for other extension and that for the base protocol. While this is a complete description and can be processed by the XDR compiler, the result might not be usable to process the extended protocol, for a number of reasons: The RPC-over-RDMA transport headers do not constitute an RPC program and version negotation and message selection part of the XDR, rather than being external to it. Headers used for requests and replies are not ncessarily paired, as they would be in an RPC program. Header types defined as optional extensions ovrlay existing nominally opaque fields in the base protocol. While this overlay architecture allows code aware of the overlay relationships to have a more complete view of header structure, this overlay relationship cannot be expressed within the XDR language

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> /// /* /// * Copyright (c) 2010, 2016 IETF Trust and the persons /// * identified as authors of the code. All rights reserved. /// * /// * The author of the code is: D. Noveck. /// * /// * 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 BEGINS> /// /* /// * TBD soon /// * /// * In a future iteration, this will be assembled from existing XDR /// * code fragments already aoppearing in the document. /// */ <CODE ENDS>

The information transferred in the transport characteristics described in this document do not raise any security issues. If and when additional transport characteristics are proposed, the review of the associated standards track document should deal with possible security issues raised by those new transport characteristics

This document does not require any actions by IANA.

&RFC2119; &RFC4506; Oracle DayDreamer Oracle Work in progress. &RFC5662; &RFC5666; &RFC5667; Oracle Hewlett Packard Enterprise Work in progress. Hewlett Packard Enterprise Work in progress.

The author gratefully acknowledges the work of Brent Callaghan and Tom Talpey producing the original RPC-over-RDMA Version One specification and also Tom's work in helping to clarify that specification. The author also wishes to thank Chuck Lever for his work resurrecting NFS support for RDMA in , for clarifying the relationshp between RDMA and direct data placement, and for beginning the work on RPC-over-RDMA Version Two. The extract.sh shell script and formatting conventions were first described by the authors of the NFSv4.1 XDR specification .