RDMA Extensions for Enhanced Memory Placement

1.1. Glossary

This document is an extension of RFC 5040 and RFC 7306, and key words are additionally defined in the glossaries of the referenced documents.¶

The following additional terms are used in this document as defined.¶

Flush:

The submitting of previously written data from volatile intermediate locations for subsequent placement, in a persistent and/or globally visible fashion.¶

Persistent:

The property that data is present, readable and remains stable after recovery from a power failure or other fatal error in an upper layer or hardware. Durability, Cache control, [SCSI].¶

Globally Visible:

The property of data being available for reading consistently by all processing elements on a system. Global visibility and persistence are not necessarily causally related; either one may precede the other, or they may take effect simultaneously, depending on the architecture of the platform.¶

ULP:

Upper-Layer Protocol, as defined in RFC 5040 and RFC 7306.¶

1.2. Problem Statement

RDMA is widely deployed in support of storage and shared memory over increasingly low-latency and high-bandwidth networks. The state of the art today yields end-to-end network latencies on the order of one to two microseconds for message transfer, and bandwidths exceeding 100 gigabit/s. These bandwidths are expected to increase over time, with latencies decreasing as a direct result, constrained of course by certain laws of physics.¶

In storage, another trend is emerging - greatly reduced latency of persistently storing data. While best-of-class Hard Disk Drives (HDDs) have delivered average latencies of several milliseconds for many years, Solid State Disks (SSDs) have improved this by one to two orders of magnitude. Technologies such as NVM Express NVMe and Compute Express Link CXL yield even higher-performing results by eliminating the traditional storage interconnect. The latest technologies providing memory-based persistence, such as Nonvolatile Memory DIMM NVDIMM-N, places storage-like semantics directly on the memory bus, reducing latency to less than a microsecond and increasing bandwidth to potentially many tens of gigabyte/s.¶

RDMA protocols, in turn, are used for many storage protocols, including NFS/RDMA [RFC8881] [RFC8166] [RFC8267], SMB3 over SMBDirect [MS-SMB2] [MS-SMBD] and iSER [RFC7145], to name just a few. These protocols allow storage and computing peers to take full advantage of these highly performant networks and storage technologies to achieve remarkable throughput, while minimizing the CPU overhead needed to drive their workloads. This leaves more computing resources available for the applications, which in turn can scale to even greater levels. Within the context of Cloud-based environments, and through scale-out approaches, this can directly reduce the number of servers that need to be deployed, making such attributes highly compelling.¶

However, limiting factors come into play when deploying ultra-low latency storage in such environments:¶

• The latency of the fabric, and of the necessary message exchanges to ensure reliable transfer is now higher than that of the storage itself.¶
• The requirement that storage be resilient to failure requires that multiple copies be sent for processing in multiple locations, adding extra hops which increase the latency and computing demand placed on implementing the resiliency.¶
• Processing is required at the receiver in order to ensure that the storage data has reached a persistent state, and acknowledge the transfer so that the sender can proceed.¶
• Typical latency optimizations, such as polling a receive memory location for a key that determines when the data arrives, can create both correctness and security issues because this approach requires the memory remain open to writes and therefore the buffer may not remain stable after the application determines that the IO has completed. This is of particular concern in security conscious environments.¶

The first issue is fundamental, and due to the nature of serial, shared communication channels, presents challenges that are not easily bypassed. Communication cannot exceed the speed of light, for example, and serialization/deserialization plus packet processing adds further delay. Therefore, a solution which offloads and reduces the overhead of exchanges which encounter such latencies is highly desirable.¶

The second issue requires that outbound transfers be made as efficient as possible, so that replication of data can be done with minimal overhead and delay (latency). A reliable "push" transfer method is highly suited to this.¶

The third issue requires that the transfer be performed without an upper-layer exchange required. Within security contraints, RDMA transfers, arbitrated only by lower layers into well-defined and pre-advertised buffers, present an ideal solution.¶

The fourth issue requires significant CPU activity, consuming power and valuable resources, and may not be guaranteed by the RDMA protocols, which themselves make no requirement of the order in which certain received data is placed or becomes visible; such guarantees are made only after signaling a completion to upper layers.¶

The RDMAP and DDP protocols, together, provide data transfer semantics with certain consistency guarantees to both the sender and receiver. Delivery of data transferred by these protocols is said to have been Placed in destination buffers upon Completion of specific operations. In general, these guarantees are limited to the visibility of the transferred data within the hardware domain of the receiver (data sink). Significantly, the guarantees do not necessarily extend to the actual storage of the data in memory cells, nor do they convey any guarantee that the data integrity is intact, nor that it remains present after a catastrophic failure. These guarantees may be provided by upper layers, such as the ones mentioned, after processing the Completions, and performing the necessary operations.¶

The NFSv4.1 and SMB3 are file oriented, and the iSER protocol is block oriented; all are used extensively for providing access to hard disk and solid state flash drive media. Such devices incur certain latencies in their operation, from the millisecond-order rotational and seek delays of rotating disk hardware, or the 100-microsecond-order erase/write and translation layers of solid state flash. These file and block protocols have benefited from the increased bandwidth, lower latency, and markedly lower CPU overhead of RDMA to provide excellent performance for such media, approximately 30-50 microseconds for 4 kilobyte writes in leading implementations.¶

The above storage protocols employ a "pull" model for write: the client, or initiator, sends an upper layer write request which contains an RDMA reference to the data to be written. The upper layer protocols encode this as one or more memory regions. The server, or target, then prepares the request for local write execution, and "pulls" the data with one or more RDMA Read operations. After processing the write, a response is returned. There are therefore two or more roundtrips on the RDMA network in processing the request. This is desirable for several reasons, as described in the relevant specifications, but it incurs latency. However, since as mentioned the network latency has been so much less than the storage processing, this has been a sound approach.¶

Today, a new class of Storage Class Memory is emerging, in the form of Non-Volatile DIMM and NVM Express devices, among others. These devices are characterized by further reduced latencies, in the 10-microsecond-order range for NVMe, and sub-microsecond for NVDIMM. The 30-50 microsecond write latencies of the above file and block protocols are therefore from one to two orders of magnitude larger than the storage media! The client/server processing model of traditional storage protocols are no longer amortizable at an acceptable level into the overall latency of storage access, due to their requiring request/response communication, CPU processing by the both server and client (or target and initiator), and the interrupts to signal such requests.¶

Another important property of certain such devices is the requirement for explicitly requesting that the data written to them be made persistent. Because persistence requires that data be committed to memory cells, it is a relatively expensive operation in time (and power), and in order to maintain the highest device throughput and most efficient operation, the device persistence operation is explicit. When the data is written by an application on the local platform, this responsibility naturally falls to that application (and the CPU on which it runs). However, when data is written by current RDMA protocols, no such semantic is provided. As a result, upper layer stacks, and the target CPU, must be invoked to perform it, adding overhead and latency that is now highly undesirable.¶

When such devices are deployed as the remote server, or target, storage, and when such a persistence can be requested and guaranteed remotely, a new transfer model can be considered. Instead of relying on the server, or target, to perform requested processing and to reply after the data is persistently stored, it becomes desirable for the client, or initiator, to perform these operations itself. By altering the transfer models to support a "push mode", that is, by allowing the requestor to push data with RDMA Write and subsequently make it persistent, a full round trip can be eliminated from the operation. Additionally, the signaling, and processing overheads at the remote peer (server or target) can be eliminated. This becomes an extremely compelling latency advantage.¶

1.3. Memory Placement Semantics

In DDP (RFC 5041), data is considered "Placed" when it is submitted by the RDMA Network Interface Controller ("RNIC") to the system. This operation is commonly an i/o bus write, e.g. via PCI. The submission is ordered, but there is no confirmation or necessary guarantee that the data has yet reached its destination, nor become visible to other devices in the system. The data will eventually become so, but possibly at a later time. The act of "Delivery", on the other hand, offers a stronger semantic, guaranteeing that not only have prior operations been executed, but also guaranteeing any data is in a consistent and visible state. Generally however, such "Delivery" requires raising a completion event, necessarily involving the host CPU. This is a relatively expensive, and latency-bound operation. Some systems perform "DMA snooping" to provide a somewhat higher guarantee of visibility after delivery and without CPU intervention, but others do not. The RDMA requirements remain the same, therefore, upper layers may make no broad assumption. Such platform behaviors, in any case, do not address persistence.¶

The extensions in this document primarily address a new "flush to persistence" RDMA operation. This operation, when invoked by a connected remote RDMA peer, can be used to request that previously-written data be moved into the persistent storage domain. This may be a simple flush to a memory cell, or it may require movement across one or more busses within the target platform, followed by an explicit persistence operation. Such matters are beyond the scope of this specification, which provides only the mechanism to request the operation, and to signal its successful completion.¶

In a similar vein, many applications desire to achieve visibility of remotely-provided data, and to do so with minimum latency. One example of such applications is "network shared memory", where publish-subscribe access to network-accessible buffers is shared by multiple peers, possibly from applications on the platform hosting the buffers, and others via network connection. There may therefore be multiple local devices accessing the buffer - for example, CPUs, and other RNICs. The topology of the hosting platform may be complex, with multiple i/o, memory, and interconnect busses, requiring multiple intervening steps to process arriving data.¶

To address this, the extensions additionally provide a "flush to global visibility", which requires the RNIC to perform platform-dependent processing in order to guarantee that the contents of a specific range are visible for all devices that access them. On certain highly-consistent platforms, this may be provided natively. On others, it may require platform-specific processing, to flush data from volatile caches, invalidate stale cached data from others, and to drain pending operations. Ideally, but not universally, this processing will take place without CPU intervention. With a global visibility guarantee, network shared memory and similar applications will be assured of broader compatibility and lower latency across all hardware platforms.¶

Subsequently, many applications will seek to obtain a guarantee that the integrity of the data has been preserved after it has been flushed to a persistent or globally visible state. This may be enforced at any time. Unlike traditional block-based storage, the data provided by RDMA is neither structured nor segmented, and is therefore not self-describing with respect to integrity. Only the originator of the data, or an upper layer, is able to determine that. Applications requiring such guarantees may include filesystems, database logwriters, replication agents, etc.¶

To provide an additional integrity guarantee, a new operation is provided by the extension, which will calculate, and optionally compare an integrity value for an arbitrary region. The operation is ordered with respect to preceding and subsequent operations, allowing for a request pipeline without "bubbles" - roundtrip delays to ascertain success or failure.¶

Finally, once data has been transmitted and directly placed by RDMA, flushed to its final state, and its integrity verified, applications will seek to commit the result with a transaction semantic. The previous application examples apply here, logwriters and replication are key, and both are highly latency- and integrity-sensitive. They desire a pipelined transaction marker which is placed atomically to indicate the validity of the preceding operations. They may require that the data be in a persistent and/or globally visible state, before placing this marker.¶

Together the above discussion argues for a new "one sided" transfer model supporting extended remote placement guarantees, provided by the RDMA transport, and used directly by upper layers on a data source, to control persistent storage of data on a remote data sink without requiring its remote interaction. Existing, or new, upper layers can use such a model in several ways, and evolutionary steps to support persistence guarantees without required protocol changes are explored in the remainder of this document.¶

Note that is intended that the requirements and concept of these extensions can be applied to any similar RDMA protocol, and that a compatible model can be applied broadly.¶

1.4. Requirements for RDMA Flush

The fundamental new requirement for extending RDMA protocols is to define the property of persistence. This new property is to be expressed by new operations to extend Placement as defined in existing RDMA protocols. The RFC 5040 protocols specify that Placement means that the data is visible consistently within a platform-defined domain on which the buffer resides, and to remote peers across the network via RDMA to an adapter within the domain. In modern hardware designs, this buffer can reside in memory, or also in cache, if that cache is part of the hardware consistency domain. Many designs use such caches extensively to improve performance of local access.¶

Persistence, by contrast, requires that the buffer contents be preserved across catastrophic failures. While it is possible for caches to be persistent, they are typically not, or they provide the persistence guarantee for a limited period of time, for example, while backup power is applied. Efficient designs, in fact, lead most implementations to simply make them volatile. In these designs, an explicit flush operation (writing dirty data from caches), often followed by an explicit operation (ensuring the data has reached its destination and is in a persistent state), is required to provide this guarantee. In some platforms, these operations may be combined.¶

For the RDMA protocol to remotely provide such guarantees, an extension is required. Note that this does not imply support for persistence or global visibility by the RDMA hardware implementation itself; it is entirely acceptable for the RDMA implementation to request these from another subsystem, for example, by requesting that the CPU perform the flush to persistence, or that the destination memory device do so. But, in an ideal implementation, the RDMA implementation will be able to act as a master and provide these services without further work requests local to the data sink. Note, it is possible that different buffers will require different processing, for example one buffer may reside in persistent memory, while another may place its blocks in a storage device. Many such memory-addressable designs are entering the market, from NVDIMM to NVMe and even to SSDs and hard drives.¶

Therefore, additionally any local memory registration primitive will be enhanced to specify new optional placement attributes, along with any local information required to achieve them. These attributes do not explicitly traverse the network - like existing local memory registration, the region is fully described by a { STag, Tagged offset, length } descriptor, and such aspects of the local physical address, memory type, protection (remote read, remote write, protection key), etc are not instantiated in the protocol. Indeed, each local RDMA implementation maintains these, and strictly performs processing based on them, and they are not exposed to the peer. Such considerations are discussed in the RDMAP security model [RFC5042].¶

Note, additionally, that by describing such attributes only through the presence of an optional property of each region, it is possible to describe regions referring to the same physical segment as a combination of attributes, in order to enable efficient processing. Processing of writes to regions marked as persistent, globally visible, or neither ("ordinary" memory) may be optimized appropriately. For example, such memory can be registered multiple times, yielding multiple different Steering Tags which nonetheless merge data in the underlying memory. This can be used by upper layers to enable bulk-type processing with low overhead, by assigning specific attributes through use of the Steering Tag.¶

When the underlying region is marked as persistent, the placement of data into persistence is guaranteed only after a successful RDMA Flush directed to the Steering Tag which holds the persistent attribute (i.e. any volatile buffering between the network and the underlying storage has been flushed, and all appropriate platform- and device-specific steps have been performed).¶

To enable the maximum generality, the RDMA Flush operation is specified to act on a set of bytes in a region, specified by a standard RDMA { STag, Tagged offset, length } descriptor. It is required that each byte of the specified segment be in the requested state before the response to the Flush is generated. However, depending on the implementation, other bytes in the region, or in other regions, may be acted upon as part of processing any RDMA Flush. In fact, any data in any buffer destined for persistent storage, may become persistent at any time, even if not requested explicitly. For example, the host system may flush cache entries due to cache pressure, or as part of platform housekeeping activities. Or, a simple and stateless approach to flushing a specific range might be for all data be flushed and made persistent, system-wide. A possibly more efficient implementation might track previously written bytes, or blocks with "dirty" bytes, and flush only those to persistence. Either result provides the required guarantee.¶

The RDMA Flush operation provides a response but does not return a status, or can result in an RDMA Terminate event upon failure. There are several possibilities for failure. A region permission check is performed first, prior to any attempt to process data. If the check is successful, the operation proceeds, but the RDMA Flush operation may fail to make the data persistent, perhaps due to a hardware failure, or a change in device capability (device read-only, device wear, etc). The device itself may support an integrity check, similar to modern error checking and correction (ECC) memory or media error detection on hard drive surfaces, which may signal failure. Or, the request may exceed device limits in size or even transient attribute such as temporary media failure. The behavior of the device itself is beyond the scope of this specification.¶

Because the RDMA Flush involves processing on the local platform and the actual storage device, in addition to being ordered with certain other RDMA operations, it is expected to take a certain time to be performed. For these reasons, the operation is required to be defined as a "queued" operation on the RDMA device, and therefore also the protocol. The RDMA protocol supports RDMA Read (RFC 5040) and Atomic (RFC 7306) in such a fashion. The iWARP family defines a "queue number" with queue-specific processing that is naturally suited for this. Queuing provides a convenient means for supporting ordering among other operations, and for flow control. Flow control for RDMA Reads and Atomics on any given Queue Pair share incoming and outgoing crediting depths ("IRD/ORD"); operations in this specification share these values and do not define their own separate values.¶

1.5. Requirements for Atomic Write

The persistence of data is a key property by which applications implement transactional behavior. Transactional applications, such as databases and log-based filesystems, among many others, implement a "two phase commit" wherein a write is made persistent, and only upon success, a validity indicator for the written data is set. Such semantics are challenging to provide over an RDMA fabric, as it exists today. The RDMA Write operation does not generate an acknowledgement at the RDMA layers. And, even when an RDMA Write is delivered, if the destination region is persistent, its data can be made persistent at any time, even before a Flush is requested. Out-of-order DDP processing, packet fragmentation, and other matters of scheduling transfers can introduce partial delivery and ordering differences. If a region is made persistent, or even globally visible, before such sequences are complete, significant application-layer inconsistencies can result. Therefore, applications may require fine-grained control over the placement of bytes. In current RDMA storage solutions, these semantics are implemented in upper layers, potentially with additional upper layer message signaling, and corresponding roundtrips and blocking behaviors.¶

In addition to controlling placement of bytes, the ordering of such placement can be important. By providing an ordered relationship among write and flush operations, a basic transaction scenario can be constructed, in a way which can function with equal semantics both locally and remotely. In a "log-based" scenario, for example, a relatively large segment (log "record") is placed, and made persistent. Once persistence of the segment is assured, a second small segment (log "pointer") is written, and optionally also made persistent. The visibility of the second segment is used to imply the validity, and persistence, of the first. Any sequence of such log-operation pairs can thereby always have a single valid state. In case of failure, the resulting sequence of transactions can therefore be recovered up to and including the final state.¶

Such semantics are typically a challenge to implement on general purpose hardware platforms, and a variety of application approaches have become common. Generally, they employ a small, well-aligned atom of storage for the second segment (the one used for validity). For example, an integer or pointer, aligned to natural memory address boundaries and CPU and other cache attributes, is stored using instructions which provide for atomic placement. Existing RDMA protocols, however, provide no such capability.¶

This document specifies an Atomic Write extension, which, appropriately constrained, can serve to provide similar semantics. A small (64 bit) payload, sent in a request which is ordered with respect to prior RDMA Flush operations on the same stream and targeted at a segment which is aligned such that it can be placed in a single hardware operation, can be used to satisfy the previously described scenario.¶

Note that the visibility of this atomically written payload can also serve as an indication that all prior operations have succeeded, enabling a highly efficient application-visible memory semaphore. This may benefit remote shared memory models, by providing stronger guarantees than previously obtainable with RDMA protocols.¶

1.6. Requirements for RDMA Verify

An additional matter remains with persistence - the integrity of the persistent data. Typically, storage stacks such as filesystems and media approches such as SCSI [T10DIF] or filesystem integrity checks provide for block- or file-level protection of data at rest on storage devices. With RDMA protocols and physical memory, no such stacks are present. And, to add such support would introduce CPU processing and its inherent latency, counter to the goals of the remote access approach. Requiring the peer to verify by remotely reading the data is prohibitive in both bandwidth and latency, and without additional mechanisms to ensure the actual stored data is read (and not a copy in some volatile cache), can not provide the necessary result.¶

To address this, an integrity operation is required. The integrity check is initiated by the upper layer or application, which requests computation of the hash of a given segment of arbitrary size via an RDMA Verify operation targeting the RDMA segment on the responder. The operation optionally specifies the expected hash value. The responder calculates the value from the contents of the targeted RDMA segment, bypassing any volatile copies remaining in caches. The responder responds with its computed hash value, or when the optional expected hash value was sent and does not match the computed hash value, terminates the connection with an appropriate error status. Specifying this optional termination behavior enables a transaction to be sent as WRITE-FLUSH-VERIFY-ATOMICWRITE, without any pipeline bubble. The result (carried by the subsequently ordered ATOMIC_WRITE) will not not be marked as valid if any prior operation is terminated, and in this case, recovery can be initiated by the requestor immediately from the point of failure. On the other hand, an errorless "scrub" can be implemented without the optional termination behavior, The responder will simply return the computed hash of the contents.¶

The hash algorithm is not specified by the RDMA protocol, instead it is left to the upper layer to select an appropriate choice based upon the strength, security, length, support by the RNIC, and other criteria. The size of the resulting hash is therefore also not specified by the RDMA protocol, but is dictated by the hash algorithm. The RDMA protocol becomes simply a transport for exchanging the values.¶

It should be noted that the design of the operation, passing of the hash value from requestor to responder (instead of, for example, computing it at the responder and simply returning it), allows both peers to determine immediately whether the segment is considered valid, permitting local processing by both peers if that is not the case. For example, a known-bad segment can be immediately marked as such ("poisoned") by the responder platform, requiring recovery before permitting access. [cf ACPI, JEDEC, SNIA NVMP specifications]¶

1.7. Local Semantics

The new operations imply new access methods ("verbs") to local persistent memory which backs registrations. Registrations of memory which support persistence will follow all existing practices to ensure permission-based remote access. The RDMA protocols do not expose these permissions on the wire, instead they are contained in local memory registration semantics. Existing attributes are Remote Read and Remote Write, which are granted individually through local registration on the machine. If an RDMA Read or RDMA Write operation arrives which targets a segment without the appropriate attribute, the connection is terminated.¶

In support of the new operations, new memory attributes are needed. For RDMA Flush, a new "Flushable" attribute provides permission to invoke the operation on memory in the region for persistence and/or global visibility. When registering, along with the attribute, additional local information can be provided to the RDMA layer such as the type of memory, the necessary processing to make its contents persistent, etc. If the attribute is requested for memory which cannot be persisted, it also allows the local provider to return an error to the upper layer, obviating the upper layer from providing the region to the remote peer.¶

For RDMA Verify, the new "Verifiable" attribute provides permission to compute the hash of the contents of the region. Again, along with the attribute, additional information such as the hash algorithm for the region is provided to the local operation. If the attribute is requested for non-verifiable memory, or if the hash algorithm is not available, the local provider can return an error to the upper layer. In the case of success, the upper layer can exchange the necessary information with the remote peer. Note that the verification algorithm is not identified by the on-the-wire operation as a result. Establishing the choice of hash for each region is done by the local consumer, and each hash result is merely transported by the RDMA protocol. Memory can be registered under multiple regions, if differing hashes are required, for example unique keys may be provisoned to implement secure hashing. Also note that, for certain "reversible" hash algorithms, this may allow peers to effectively read the memory, therefore, the local platform may require additional read permissions to be associated with the Verifiable permission, when such algorithms are selected.¶

The Atomic Write operation requires no new attributes, however it does require the "Remote Write" attribute on the target region, as is true for any remotely requested write. If the Atomic Write additionally targets a Flushable region, the RDMA Flush is performed separately. It is never generally possible to achieve persistence atomically with placement, even locally.¶

2.1. RDMA Flush

The RDMA Flush operation requests that all bytes in a specified region are to be made persistent and/or globally visible, under control of specified flags. As specified in section 3 its execution is ordered after the successful completion of any previous requested RDMA Write or certain other operations. The response is generated after the region has reached its specified state. The implementation MUST fail the operation and send a terminate message if the RDMA Flush cannot be performed, or has encountered an error.¶

The RDMA Flush operation MUST NOT be responded to until all data has attained the requested state. Achieving persistence may require programming and/or flushing of device buffers, while achieving global visibility may require flushing of cached buffers across the entire platform interconnect. In no event are persistence and global visibility achieved atomically, one may precede the other and either may complete at any time. A subsequent Atomic Write operation may be used by an upper layer consumer to indicate that the requested dispositions are available after completion of the RDMA Flush, in addition to other approaches.¶

+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                         Data Sink STag                        |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                        Data Sink Length                       |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                     Data Sink Tagged Offset                   |
+                                                               +
|                                                               |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                     Flush Disposition Flags             +R+G+P|
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

2.2. RDMA Verify

The RDMA Verify operation requests that all bytes in a specified region are to be read from the underlying storage and that an integrity hash be calculated on their value. As specified in section 3 its operation is ordered after the successful completion of any previous requested RDMA Write and RDMA Flush, or certain other operations.¶

+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                         Data Sink STag                        |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                        Data Sink Length                       |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                     Data Sink Tagged Offset                   |
+                                                               +
|                                                               |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                Hash Value (optional, variable)                |
|                              ...                              |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                    Hash Value (variable)                      |
|                              ...                              |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

2.3. Atomic Write

The Atomic Write operation provides a 64-bit block of data which is written to a specified region atomically, and as specified in section 4 its execution is ordered after the successful completion of all previous RDMA Flush and RDMA Verify. The target Buffer MUST be 64-bits at 64-bit alignment, and the implementation MUST fail the operation and send a terminate message if the data cannot be written to it atomically.¶

+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                         Data Sink STag                        |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                        Data Sink Length                       |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                     Data Sink Tagged Offset                   |
+                                                               +
|                                                               |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                              Data                             |
+                                                               +
|                                                               |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

2.4. Discovery of RDMAP Extensions

As for RFC 7306, explicit negotiation by the RDMAP peers of the extensions covered by this document is not specified. Instead, it is RECOMMENDED that RDMA applications and/or ULPs negotiate any use of these extensions at the application or ULP level. The definition of such application-specific mechanisms is outside the scope of this specification. For backward compatibility, existing applications and/or ULPs should not assume that these extensions are supported.¶

In the absence of application-specific negotiation of the features defined within this specification, the new operations can be attempted, and reported errors can be used to determine a remote peer's capabilities. In the case of RDMA Flush and Atomic Write, an operation to a previously Advertised buffer with remote write permission can be used to determine the peer's support. A Remote Operation Error or Unexpected OpCode error will be reported by the remote peer if the Operation is not supported by the remote peer. For RDMA Verify, such an operation may target a buffer with remote read permission.¶

4.1. Errors Detected at the Local Peer

The Local Peer MUST send a Terminate Message for each of the following cases:¶

1. For errors detected while creating an RDMA Flush, RDMA Verify or Atomic Write Request, or other reasons not directly associated with an incoming Message, the Terminate Message and Error code are sent instead of the Message. In this case, the Error Type and Error Code fields are included in the Terminate Message, but the Terminated DDP Header and Terminated RDMA Header fields are set to zero.¶
2. For errors detected on an incoming RDMA Flush, RDMA Verify or Atomic Write Request or Response, the Terminate Message is sent at the earliest possible opportunity, preferably in the next outgoing RDMA Message. In this case, the Error Type, Error Code, and Terminated DDP Header fields are included in the Terminate Message, but the Terminated RDMA Header field is set to zero.¶
3. For errors detected in the processing of the RDMA Flush or RDMA Verify itself, that is, the act of flushing or verifying the data, the Terminate Message is generated as per the referenced specifications. Even though data is not lost, the upper layer MUST be notified of the failure by informing the requester of the status, terminating any pending operations, and allow the requester to perform further action, for instance, recovery.¶

4.2. Errors Detected at the Remote Peer

On incoming RDMA Flush and RDMA Verify Requests, the following MUST be validated:¶

• The DDP layer MUST validate all DDP Segment fields.¶

The following additional validation MUST be performed:¶

• The RDMAP layer MUST validate the Data Sink STag, Data Sink Length and Data Sink Tagged Offset fields of the payload.¶
• If the RDMA Flush, RDMA Verify or Atomic Write operation cannot be satisfied, due to transient or permanent errors detected in the processing by the Responder, a Terminate message MUST be returned to the Requestor.¶

8.1. Normative References

[RFC2119]

Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, March 1997, <https://www.rfc-editor.org/info/rfc2119>.

[RFC5040]

Recio, R., Metzler, B., Culley, P., Hilland, J., and D. Garcia, "A Remote Direct Memory Access Protocol Specification", RFC 5040, DOI 10.17487/RFC5040, October 2007, <https://www.rfc-editor.org/info/rfc5040>.

[RFC5041]

Shah, H., Pinkerton, J., Recio, R., and P. Culley, "Direct Data Placement over Reliable Transports", RFC 5041, DOI 10.17487/RFC5041, October 2007, <https://www.rfc-editor.org/info/rfc5041>.

[RFC5042]

Pinkerton, J. and E. Deleganes, "Direct Data Placement Protocol (DDP) / Remote Direct Memory Access Protocol (RDMAP) Security", RFC 5042, DOI 10.17487/RFC5042, October 2007, <https://www.rfc-editor.org/info/rfc5042>.

[RFC6580]

Ko, M. and D. Black, "IANA Registries for the Remote Direct Data Placement (RDDP) Protocols", RFC 6580, DOI 10.17487/RFC6580, April 2012, <https://www.rfc-editor.org/info/rfc6580>.

[RFC7306]

Shah, H., Marti, F., Noureddine, W., Eiriksson, A., and R. Sharp, "Remote Direct Memory Access (RDMA) Protocol Extensions", RFC 7306, DOI 10.17487/RFC7306, June 2014, <https://www.rfc-editor.org/info/rfc7306>.

8.2. Informative References

[MS-SMB2]

Microsoft Corporation, "Server Message Block (SMB) Protocol Versions 2 and 3 (MS-SMB2)", 29 April 2022. https://docs.microsoft.com/en-us/openspecs/windows_protocols/ms-smb2/5606ad47-5ee0-437a-817e-70c366052962

[MS-SMBD]

Microsoft Corporation, "SMB2 Remote Direct Memory Access (RDMA) Transport Protocol (MS-SMBD)", 5 June 2021. https://docs.microsoft.com/en-us/openspecs/windows_protocols/ms-smbd/1ca5f4ae-e5b1-493d-b87d-f4464325e6e3

[RFC5045]

Bestler, C., Ed. and L. Coene, "Applicability of Remote Direct Memory Access Protocol (RDMA) and Direct Data Placement (DDP)", RFC 5045, DOI 10.17487/RFC5045, October 2007, <https://www.rfc-editor.org/info/rfc5045>.

[RFC7145]

Ko, M. and A. Nezhinsky, "Internet Small Computer System Interface (iSCSI) Extensions for the Remote Direct Memory Access (RDMA) Specification", RFC 7145, DOI 10.17487/RFC7145, April 2014, <https://www.rfc-editor.org/info/rfc7145>.

[RFC8166]

Lever, C., Ed., Simpson, W., and T. Talpey, "Remote Direct Memory Access Transport for Remote Procedure Call Version 1", RFC 8166, DOI 10.17487/RFC8166, June 2017, <https://www.rfc-editor.org/info/rfc8166>.

[RFC8267]

Lever, C., "Network File System (NFS) Upper-Layer Binding to RPC-over-RDMA Version 1", RFC 8267, DOI 10.17487/RFC8267, October 2017, <https://www.rfc-editor.org/info/rfc8267>.

[RFC8881]

Noveck, D., Ed. and C. Lever, "Network File System (NFS) Version 4 Minor Version 1 Protocol", RFC 8881, DOI 10.17487/RFC8881, August 2020, <https://www.rfc-editor.org/info/rfc8881>.

[SCSI]

ANSI, "SCSI Primary Commands - 3 (SPC-3) (INCITS 408-2005)", 4 May 2005.

[SNIANVMP]

SNIA NVM Programming TWG, "SNIA NVM Programming Model v1.2", 19 June 2017. https://www.snia.org/sites/default/files/technical_work/final/NVMProgrammingModel_v1.2.pdf

[T10DIF]

T10 technical subcommittee, "T10 Data Integrity Field (DIF) / Protection Information (PI)", date unknown.

A.1. DDP Segment for RDMA Flush Request

 0                   1                   2                   3
 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
                                +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
                                |   DDP Control | RDMA Control  |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                      Reserved (Not Used)                      |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|             DDP (Flush Request) Queue Number (1)              |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|           DDP (Flush Request) Message Sequence Number         |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|          DDP (Flush Request) Message Offset (Not Used)        |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                       Data Sink STag                          |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                      Data Sink Length                         |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                   Data Sink Tagged Offset                     |
+                                                               +
|                                                               |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                      Disposition Flags                  +R+G+P|
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

A.2. DDP Segment for RDMA Flush Response

 0                   1                   2                   3
 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
                                +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
                                |   DDP Control | RDMA Control  |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                      Reserved (Not Used)                      |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|              DDP (Flush Response) Queue Number (3)            |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|          DDP (Flush Response) Message Sequence Number         |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

A.3. DDP Segment for RDMA Verify Request

 0                   1                   2                   3
 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
                                +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
                                |   DDP Control | RDMA Control  |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                      Reserved (Not Used)                      |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|             DDP (Verify Request) Queue Number (1)             |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|           DDP (Verify Request) Message Sequence Number        |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|          DDP (Verify Request) Message Offset (Not Used)       |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                       Data Sink STag                          |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                      Data Sink Length                         |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                   Data Sink Tagged Offset                     |
+                                                               +
|                                                               |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                Hash Value (optional, variable)                |
|                              ...                              |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

A.4. DDP Segment for RDMA Verify Response

 0                   1                   2                   3
 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
                                +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
                                |   DDP Control | RDMA Control  |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                      Reserved (Not Used)                      |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|              DDP (Verify Response) Queue Number (3)           |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|          DDP (Verify Response) Message Sequence Number        |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                     Hash Value (variable)                     |
|                              ...                              |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

A.5. DDP Segment for Atomic Write Request

 0                   1                   2                   3
 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
                                +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
                                |   DDP Control | RDMA Control  |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                      Reserved (Not Used)                      |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|          DDP (Atomic Write Request) Queue Number (1)          |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|        DDP (Atomic Write Request) Message Sequence Number     |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|       DDP (Atomic Write Request) Message Offset (Not Used)    |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                       Data Sink STag                          |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                 Data Sink Length (value=8)                    |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                   Data Sink Tagged Offset                     |
+                                                               +
|                                                               |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                       Data (64 bits)                          |
+                                                               +
|                                                               |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

A.6. DDP Segment for Atomic Write Response

 0                   1                   2                   3
 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
                                +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
                                |   DDP Control | RDMA Control  |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                      Reserved (Not Used)                      |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|           DDP (Atomic Write Response) Queue Number (3)        |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|       DDP (Atomic Write Response) Message Sequence Number     |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+