| Network File System Version 4 | C. Lever |
| Internet-Draft | Oracle |
| Intended status: Informational | March 13, 2017 |
| Expires: September 14, 2017 |
Using Remote Invalidation With RPC-Over-RDMA Transports
draft-cel-nfsv4-reminv-design-05
Remote Invalidation relieves RDMA responders of some of the burden of preparing memory to be accessed remotely, thus reducing the latency of RDMA Read and Write operations. This document considers how to introduce generic support for Remote Invalidation to RPC-over-RDMA transport protocols.
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in [RFC2119].
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at http://datatracker.ietf.org/drafts/current/.
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."
This Internet-Draft will expire on September 14, 2017.
Copyright (c) 2017 IETF Trust and the persons identified as the document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (http://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.
Like other RDMA-enabled storage protocols, RPC-over-RDMA Version One employs a Read-Write transfer model when using explicit RDMA operations to transfer data [I-D.ietf-nfsv4-rfc5666bis]. This means an RPC-over-RDMA requester exposes regions of its memory to an RPC-over-RDMA responder. The responder then uses RDMA Read and Write operations to transfer bulk data payloads.
In preparation for a bulk data transfer, a requester asks its RNIC to assign a steering tag, or STag, to a region of memory containing the data to be moved. At this time, access rights are granted that allow the RNIC to access or update that memory on behalf of a remote peer. This act is referred to as "memory registration." The RNIC uses this STag to steer data to and from the registered memory region.
When data transfer is complete, each STag is dissociated from its memory region. This act is referred to as "memory invalidation." It prevents further responder access to that memory region by revoking its remote access rights. Invalidation should be done before RPC applications on the requester are allowed access to memory that was involved in an explicit RDMA operation.
Before an RPC transaction is terminated, the requester is responsible for fencing memory from the responder [I-D.ietf-nfsv4-rfc5666bis]. That makes the completion of RPC transactions synchronous with chunk invalidation. Therefore the latency of invalidation adds to the total execution time of the RPC transaction.
Remote Invalidation is a mechanism by which an RDMA peer can request that the remote peer RNIC invalidate an STag associated with memory on that remote peer [RFC5042]. An RDMA consumer requests Remote Invalidation by posting an RDMA Send With Invalidate Work Request in place of an RDMA Send Work Request. RDMA Send With Invalidate is similar to RDMA Send, but takes one additional argument: a single STag to be invalidated by the RNIC that receives the sent message. An RDMA Send message is transmitted with additional header information that conveys the STag that is to be invalidated [RFC5040].
The benefit of Remote Invalidation is that an extra Work Request, context switch, and interrupt to perform memory invalidation are not required by the requester as part of handling the completion of an RPC transaction. STag invalidation begins before the Receive completes, thus invalidation is started (and completes) sooner. The upshot is faster completion of RPC transactions that involve registered memory.
This mechanism has the most impact when explicit RDMA operations are needed to move moderate amounts of data. Invalidation latency is quite small compared to the time it takes to convey a large payload with an explicit RDMA operation. Small RPCs are already conveyed entirely via RDMA Send, thus Remote Invalidation is unnecessary for them. When the time it takes to invalidate a memory region is on the same order as the time it takes to move the contents of that region, Remote Invalidation has its greatest impact.
Remote Invalidation confers benefits similar to the benefits of increasing the size of Send and Receive buffers. However, Remote Invalidation does not incur the cost of maintaining a pool of large Receive buffers on either the requester or responder. Moderate-sized RPC payloads can be transferred without the usual costs of memory registration. Requesters can rely on RDMA Write to structure their Receive buffers without introducing additional latency.
There are some downsides, however. Remote Invalidation is not available on all RNIC devices. And, Remote Invalidation does not address the extra round trip latency incurred when using RDMA Read. This extra latency can be eliminated using a large inline threshold for transmitting RPC Calls.
The purpose of this document is to explore generally how Remote Invalidation can be introduced into the RPC-over-RDMA transport protocol. The primary design considerations for the transport protocol are to provide a mechanism to indicate when Remote Invalidation can be used by the transport, and to provide selection criteria for choosing which STag to invalidate remotely. Elements of the XDR definition of the RPC-over-RDMA protocol must be altered to some degree, depending on desired flexibility of operation, invasiveness of XDR changes, and broadness of hardware support.
Remote Invalidation was not available in the original RDMA Verbs API. New verbs API objects were specified that include operations that enable Remote Invalidation, now described in [IB]. The Verbs API provides a capabilities flag, MEM_MGT_EXTENSIONS, that indicates that an RNIC can provide the new APIs and objects.
An STag that is registered using the FRWR mechanism (in a privileged execution context), or is registered via a Memory Window (in user space), may be invalidated remotely [RFC5040]. These mechanisms are available when an RNIC supports MEM_MGT_EXTENSIONS.
RDMA Send With Invalidate is available only with MEM_MGT_EXTENSIONS.
For the purposes of this discussion, there are two classes of STags. Dynamically-registered STags are used in a single RPC, then invalidated. Persistently-registered STags live longer than one RPC. They may persist for the life of an RPC-over-RDMA connection, or longer.
In RPC-over-RDMA Version One, a requester may provide more than one STag in the chunk lists of an RPC. It may provide any combination of the following registration types in one RPC, any combination of these in a series of RPCs on the same connection, or it may use some other registration model.
Examples of persistently-registered STags include:
Examples of dynamically-registered STags include:
Among these examples, only dynamically-registered STags using Memory Windows or FRWR may be invalidated remotely.
Remote Invalidation protocol mechanisms come in different styles:
There is no RDMA layer mechanism by which a responder can determine how a requester-provided STag was registered. Thus a requester that mixes persistently- and dynamically-registered STags in one RPC, or mixes them across RPCs on the same connection, cannot tolerate Responder's Choice.
There are two related enhancements that further reduce the effort needed to invalidate STags associated with complex RPCs:
At this time, the first mechanism has been implemented in at least one RNIC on the market. The second is speculative (i.e., has not yet been implemented anywhere).
Given support for registering non-contiguous memory regions with one STag, when an RPC-over-RDMA requester constructs an RPC that has both a Read list and a Write list, the requester has a choice:
Having the ability to remotely invalidate multiple STags at once enables the combination of optimal performance and optimal security.
When requester memory is registered for remote access, an RPC-over-RDMA implementation can use Remote Invalidation by following these steps:
The requester no longer needs to invalidate the STag involved with this RPC. However, there are additional details that must be resolved before the use of Remote Invalidation can commence.
An RDMA consumer (an Upper Layer Protocol implementation) that does not support Remote Invalidation might not tolerate the use of RDMA Send With Invalidate by a responder. Such a requester performs Local Invalidation on STags that already happen to be invalid, and in some cases this can result in protection errors or other issues.
Thus, to avoid spurious connection termination, a responder must not post an RDMA Send With Invalidate Work Request unless it is sure the following three conditions are met:
When all three of these conditions are true, a requester can report positive Remote Invalidation support status to responders using an Upper Layer Protocol mechanism. When a responder does not know the requester's Remote Invalidation support status, it cannot use Remote Invalidation without endangering the connection.
The RDMA Send With Invalidate Work Request invalidates only one STag. RPC-over-RDMA requesters may register more than one STag to handle the movement of payloads for a single RPC. Either the client will have to specify which STag may be remotely invalidated, the protocol will have to specify a fixed way to select which STag to invalidate, or the responder will have to choose arbitrarily which STag to remotely invalidate.
In some circumstances, requesters may wish to utilize STags during transactions that are registered using a mechanism that does not tolerate Remote Invalidation. For example, an STag that is the requester's local DMA R_key should never be invalidated remotely. If a responder attempts to invalidate a such an STag, the result is undefined, but the connection may be terminated or other failures can occur.
Even with Remote Invalidation enabled, requesters remain responsible for ensuring all STags are invalid before RPC transactions complete. To avoid leaving STags registered, a requester must be prepared for the responder or the requester's own RNIC to have not invalidated any of an RPC's STags. When there are multiple STags associated with a single RPC, a requester must be prepared for any of the STags to have been remotely invalidated, or none of them.
As of this writing, no current RPC-over-RDMA implementation supports direct data placement in the reverse direction. However, existing protocol specifications do not forbid it [I-D.ietf-nfsv4-rfc5666bis] [I-D.ietf-nfsv4-rpcrdma-bidirection] [I-D.cel-nfsv4-rpcrdma-version-two].
When chunks are present in a reverse-direction RPC request, Remote Invalidation allows the responder to trigger invalidation of a requester's STags as part of sending a reply, the same as in the forward direction.
However, in the reverse direction, the server acts as the requester, and the client is the responder. The server's RNIC, therefore, must support receiving an IETH, and the server must have registered the STags with an appropriate registration mechanism. Thus the server must indicate its Remote Invalidation support status to the client (the opposite of forward direction Remote Invalidation).
In this section, a number of abstract protocol variations are considered. These vary in functionality and invasiveness. Some may be appropriate to use in combination.
When a higher protocol version number is negotiated, Remote Invalidation is always enabled. This new protocol version would then be usable only with RNICs that support Remote Invalidation. Both peers assume that Remote Invalidation may be used in either direction.
SMB Direct [MS-SMBD]
No XDR changes or protocol extensions are required.
Reverse direction use of Remote Invalidation is automatically supported.
The requester is not in control of which STags in an RPC may be invalidated. Thus, a requester must not advertise STags which must never be invalidated, or the protocol must specify a fixed choice of which STag(s) in each request are allowed to be invalidated remotely.
Other features and benefits of the new protocol version would not be available when an implementation employs an RNIC that does not support Remote Invalidation. In particular, RNICs that do not support MEM_MGT_EXTENTIONS could not use the new protocol version.
An extension or addition protocol version bump is required to indicate support for transport-level mechanisms that can invalidate multiple STags at once.
At connection initiation time, messages are exchanged that indicate each peer's Remote Invalidation support status. Without these messages, peers assume Remote Invalidation is not supported.
iSER [RFC7145]. Information is exchanged in RDMA-CM connection requests to report an implementation's Remote Invalidation support status.
No changes to the base protocol XDR are required.
Out-of-band messages are required to establish Remote Invalidation support status.
The requester is not in control of which STags in an RPC may be invalidated. Thus, a requester must not advertise STags which must never be invalidated.
To support reverse-direction operation, the server must separately indicate that it supports Remote Invalidation.
To enable support for multiple STag invalidation, this negotiation protocol would have to be extended again to indicate when mechanisms other than RDMA Send With Invalidate are supported by the requester's RNIC.
No new field is introduced to the transport header. Protocol specification determines how the responder chooses which STag is to be invalidated remotely. Some other means is used to determine whether Remote Invalidation can be used or not.
iSER [RFC7145]. Two STags fields appear in each request: one advertises Read data and one advertises Write data. When only one STag is used in the request, it may be invalidated remotely. One both STags are used, only the Read STag may be invalidated remotely.
SMB Direct [MS-SMBD]. The responder always chooses the first STag in each request to be invalidated remotely.
No changes to the base protocol XDR are required.
Out-of-band messages are required to establish support status.
The requester is not in control of which STags in an RPC may be invalidated. Thus, a requester must not advertise STags which must never be invalidated.
This mechanism may not work well for transport protocols that allow multiple read and write STags.
A field is added to the transport header that contains an STag which may be invalidated by the responder. A special value can be chosen to mean "no STag may be invalidated" for use by requesters that have no support for Remote Invalidation.
None.
A requester may advertise STags that cannot be invalidated remotely, as long as they are never marked as "may invalidate."
No out-of-band support status negotiation is needed.
Reverse-direction RPCs can each indicate whether a reverse-direction requester desires or does not support Remote Invalidation.
The responder needs no special logic or assumptions to choose the STag to invalidate remotely.
Either the base RPC-over-RDMA header XDR definition is altered, or a protocol extension is required.
Requesters transmit a little extra data per RPC, making RPC-over-RDMA messages slightly more costly to send and parse.
This mechanism cannot support the remote invalidation of multiple STags at once.
A new data structure is added to the transport header that indicate which STags which may be invalidated by the responder.
This information might appear as a new field in the RDMA segment data structure, as each segment has its own STag field. The field indicates whether or not that STag may be invalidated by the responder. Perhaps that field is a boolean, though in XDR, a boolean is a full 32 bits.
Or, this information could appear in the header as an array of STags, to reduce the amount of extra data contained in the RPC-over-RDMA header. Zero array elements means the requester does not support Remote Invalidation.
NVMe/Fabrics [NVME]. Each STag in a request has an associated bit flag that indicates whether the responder is allowed to invalidate it remotely.
A requester may advertise STags that cannot be invalidated remotely, as long as they are never marked as "may invalidate."
The mechanism allows a requester to request either invalidation of multiple STags at once, or to choose one STag to invalidate remotely.
No out-of-band support status negotiation is needed.
Each reverse-direction RPC can indicate whether a reverse-direction requester desires or does not support Remote Invalidation.
The responder needs no special logic or assumptions to choose the STag to invalidate remotely.
The RPC-over-RDMA header XDR definition is possibly extensively altered.
Requesters transmit extra data per RPC. However, it is limited to only one or two 32-bit words in most cases.
As a subfeature of support for Remote Invalidation, it is possible that a responder can remotely invalidate an STag (using RDMA Send With Invalidate) that refers to registered memory being used in the Read chunk of a different RPC. Such Remote Invalidation would be requested only after the RDMA Read has already been completed.
This can be useful when a responder is replying to an RPC via an inline message, but notices there are other RPC replies pending that have multiple STags, some of which are Read chunks.
None
This is one way to enable remote invalidation of multiple STags per RPC, using only RDMA Send With Invalidate.
Additional requester and responder complexity would be required to keep track of STags.
When constructing a protocol to support Remote Invalidation, one of the above designs, or some combination of them, may be chosen.
In no particular order, the author feels that the design priorities are:
An important question is whether the base RPC-over-RDMA protocol should support Remote Invalidation, whether Remote Invalidation support should be carried entirely on the shoulders of protocol extensions, or whether some combination of the two is best.
Upper Layer Protocols will likely always be responsible for some degree of signaling Remote Invalidation capabilities, as long as innovation continues at the transport layer (e.g., new RDMA operations that enable multi-STag Remote Invalidation). Predicting future hardware capabilities is challenging, limiting the ability to design long-lived protocol support for them.
Lastly, it is difficult to estimate how long the industry must continue to support less capable devices.
All things being equal, making no changes to the base XDR definition has great appeal. If the mechanism in Section 4.2 can be broadly effective at enabling Remote Invalidation in the current set of RPC-over-RDMA implementations, it would be the proper choice.
Unfortunately, among current RPC-over-RDMA client implementations, there is one client that can immediately use a per-connection style protocol, and one that can use only a per-RPC style protocol such as Section 4.4. A third known client resides in user space and uses FMR registration, thus it is incapable of immediately employing Remote Invalidation.
Because there is a wide latitude of implementation choice already allowed by the RPC-over-RDMA transport protocol, the author's preference is to implement Section 4.4. The target STag can be added to the RPC-over-RDMA transport as a single field in a new version of the RPC-over-RDMA transport protocol. No further changes or extensions are needed.
In the longer term, the requester appears to be in the better position to determine which STag may be invalidated remotely. With this mechanism, the requester can choose based on which STags may be invalidated remotely, or may use criteria based on the strengths of its RNIC. For instance, choosing the largest registered memory region might be beneficial in some cases.
Allowing the responder to select from among several choices does not seem to bring additional value, and burdens the responder with additional header parsing costs for each chunk-bearing RPC transaction.
Furthermore, the ability to request Remote Invalidation of multiple STags in a single Work Request appears to be somewhat distant. It would require additional Upper Layer Protocol mechanisms to distinguish the new mechanism from using RDMA Send With Invalidate, which we are not in a position to design today. Thus it does not seem worth the extra implementation and protocol complexity of having the requester provide a list of STags for the responder to choose from.
As an alternative to modifying the XDR definition for the RDMA_MSG and RDMA_NOMSG message types, a new RDMA message type could be introduced in a new version of RPC-over-RDMA that provides similar functionality to RDMA_MSG and RDMA_NOMSG but adds one or more new fields. This has the advantage of leaving the Version One-compatible parts of the the new XDR definition unchanged. It is an open question whether this introduces more complexity to existing implementations than adding new fields to RDMA_MSG and RDMA_NOMSG. However, this approach is similar to the introduction of READ_PLUS in the specification of NFSv4.2 [RFC7862].
Allowing the feature described in Section 4.6 is likely to increase the complexity of responder and especially requester implementations, as they would have to remember invalidated STags independently of RPC completions. Because it does not require any XDR changes, it could easily be enabled in a future protocol extension. The author's preference is to forbid this behavior in the initial specification, but allow for a future extension to introduce it.
<CODE BEGINS>
struct rpcrdma2_chunk_lists {
enum msg_type rdma_direction;
u32 rdma_inv_handle;
struct rpcrdma2_read_list *rdma_reads;
struct rpcrdma2_write_list *rdma_writes;
struct rpcrdma2_write_chunk *rdma_reply;
};
<CODE ENDS>
As an example of how to proceed, the simplest approach would replace struct rpcrdma2_chunk_lists (as defined in [I-D.cel-nfsv4-rpcrdma-version-two]) with the following:
The following language describes how to utilize the new field:
Remote Invalidation metadata is conveyed in the clear in RPC-over-RDMA headers. This does not expose any new information to attackers.
A man-in-the-middle can alter Remote Invalidation metadata while it is in transit. Requesters are prepared to handle the case where responders have not invalidated any STags associated with an RPC. An attacker can cause other STags in flight to be invalidated before the responder is finished with the associated memory. Or an attacker can replace the "to-be invalidated" STag with an STag in the same RPC that should not be invalidated remotely. Any of these might cause loss of connection, or other failures.
A connection relationship is required to exist between a requester and a responder. The requester's RNIC has associated a Protection Domain with that connection. The STag on the requester to be invalidated is associated with that Protection Domain. This protects against arbitrary invalidation of STags by network nodes not part of the connection.
Further discussion appears in [RFC5042].
There are no IANA considerations for this document.
| [IB] | InfiniBand Trade Association, "InfiniBand Architecture Specifications" |
| [MS-SMBD] | Microsoft Corporation, "SMB Remote Direct Memory Access (RDMA) Transport Protocol Specification", July 2016. |
| [NVME] | NVM Express, Inc., "NVM Express Revision 1.2.1", July 2016. |
| [RFC7862] | Haynes, T., "Network File System (NFS) Version 4 Minor Version 2 Protocol", RFC 7862, DOI 10.17487/RFC7862, November 2016. |
The author wishes to thank Sagi Grimberg, Christoph Hellwig, Karen Deitke, Dave Noveck, and Tom Talpey. The author also wishes to thank Bill Baker and Greg Marsden for their support of this work.
Special thanks go to Transport Area Director Spencer Dawkins, nfsv4 Working Group Chair Spencer Shepler, and nfsv4 Working Group Secretary Thomas Haynes for their support.