]> Hammerspace

loghyr@gmail.com

Transport Network File System Version 4 Internet-Draft The Parallel Network File System (pNFS) allows the metadata server to recall a layout from a client by file id, by file system id, or across all of a client's layouts. It also lets the server delete a deviceid via CB_NOTIFY_DEVICEID. It does not provide a mechanism for the metadata server to recall, in a single operation, all layouts that reference a specific deviceid. This document presents an extension to RFC7862 that adds a deviceid-scoped layout recall: a single CB_LAYOUTRECALL operation recalls every layout a client holds that references a given deviceid, leaving unrelated layouts in place. Without this capability, device unavailability can trigger large volumes of failed WRITE and LAYOUTERROR traffic, and administrators lack a surgical way to retire a deviceid without disturbing layouts that reference healthy devices. About This Document The latest revision of this draft can be found at . Status information for this document may be found at . Discussion of this document takes place on the Network File System Version 4 Working Group mailing list (), which is archived at . Subscribe at . Source for this draft and an issue tracker can be found at . Note to Readers This is an individual submission and does not reflect Working Group consensus. The "About This Document" section above has the current discussion venue, latest rendering, and source location.

Introduction In the Network File System version 4 (NFSv4) with a Parallel NFS (pNFS) metadata server (), there is no mechanism for the metadata server to recall layouts from a client based on the deviceid (see Section 3.3.14 of ) those layouts reference. The Flex Files layout type () is the primary motivating use case: a Flex Files layout describes data spread across multiple data servers, each identified by a distinct deviceid, often deliberately deployed across separate power or fault domains so that a single domain failure does not lose data. The absence of a deviceid-scoped recall has two distinct operational consequences. This section describes each, shows why the existing recall and notification mechanisms are the wrong shape to address it, and motivates the extension specified in the rest of the document.

Problem 1: Traffic amplification during device unavailability When a deviceid becomes unavailable -- because a power fault domain has been isolated, because a device has been taken down for maintenance, or because of a transient I/O failure -- clients holding layouts that reference the deviceid continue to issue I/O against it. Each such attempt produces a WRITE (see Section 18.32 of ) that fails with NFS4ERR_NXIO (see Section 15.1.16.3 of ), followed by a LAYOUTERROR (see Section 15.6 of ) returned to the metadata server. The metadata server learns about the failure once per (client, layout) pair; the network and the metadata server pay for the retries. Structurally, the wasted RPC count is approximately 2 x C x L x W, where:

• C is the number of clients holding layouts that reference the affected deviceid,
• L is the average number of such layouts per client, and
• W is the number of write attempts per layout during the unavailability window.

The factor of two captures the WRITE-then-LAYOUTERROR pair generated per attempt. For a representative deployment with C = 100, L = 50, and a several-minute unavailability window in which each layout sees W = 1000 write attempts, the wasted RPC count is on the order of 10 million. These numbers are illustrative order-of-magnitude estimates derived from the structural relationship; deployment-specific measurements are out of scope for this specification, but the structural multiplier is what the extension eliminates. A single deviceid-scoped recall delivered to each affected client replaces the (C x L x W) retry traffic with a single CB_LAYOUTRECALL per client plus the client-side processing to return the affected layouts. The existing recall mechanisms are the wrong shape for this problem. LAYOUTRECALL4_ALL recalls every layout a client holds, including layouts pointing at healthy deviceids; this fixes the traffic amplification at the cost of evicting unrelated layouts and forcing them to be re-issued. LAYOUTRECALL4_FILE recalls layouts one file at a time, which adds an RPC per affected file to the recovery cost; for a deployment where the number of affected files is large, the per-file recall traffic is itself a scaling concern.

Problem 2: Cleanly retiring a deviceid The metadata server can signal to clients that a deviceid no longer exists by setting NOTIFY_DEVICEID4_DELETE in the CB_NOTIFY_DEVICEID callback (see Section 20.12 of ). This flag cannot be set while any layout still references the deviceid: once a delete is announced, a client cannot have in-flight I/O against a deviceid that no longer exists. The result is that an administrator who wants to retire a deviceid -- for hardware refresh, decommission, capacity rebalancing, or removal of a failed device from the namespace -- has no surgical option. The available choices are:

• Wait for the affected layouts to expire naturally. The wait is bounded by a lease period and by the slowest client.
• Use LAYOUTRECALL4_ALL to recall every layout from every client. This evicts layouts pointing at healthy deviceids in the same operation, and may produce a thundering herd of re-LAYOUTGET requests once the recall completes.
• Use LAYOUTRECALL4_FILE for every affected file. This generates one recall RPC per file and is expensive at scale.
• Fence the deviceid (see Section 12.5.5 of ). This is disruptive: clients with active I/O see I/O failures rather than orderly recall, and the operation may invalidate concurrent legitimate I/O.

A deviceid-scoped recall provides the missing surgical option: the metadata server recalls exactly the layouts referencing the to-be-retired deviceid, waits for client acknowledgement, and then sends CB_NOTIFY_DEVICEID with NOTIFY_DEVICEID4_DELETE. The deviceid is retired without disturbing layouts that reference other deviceids.

What this document specifies Using the process detailed in , the revisions in this document become an extension of NFSv4.2 . They are built on top of the external data representation (XDR) generated from . The extension adds a deviceid-scoped recall to CB_LAYOUTRECALL: a single callback identifies a deviceid and the client returns every layout it holds that references that deviceid. The extension closes both operational gaps described above with a single new layout-recall scope.

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

Client Capability Advertisement Before the server may send a CB_LAYOUTRECALL with LAYOUTRECALL4_DEVICEID, it MUST know that the client supports the new union arm. Per Section 6, a server MUST NOT send a new callback operation or new union arm to a client that has not indicated support for it. A client that supports LAYOUTRECALL4_DEVICEID signals this by setting a new flag in the eia_flags field of the EXCHANGE_ID operation (see Section 18.35 of ): A client that sets EXCHGID4_FLAG_SUPP_RECALL_DEVICEID in its EXCHANGE_ID request advertises that it supports handling CB_LAYOUTRECALL with a LAYOUTRECALL4_DEVICEID recall type. A server MUST NOT send a CB_LAYOUTRECALL with LAYOUTRECALL4_DEVICEID to a client that did not set EXCHGID4_FLAG_SUPP_RECALL_DEVICEID in its EXCHANGE_ID request. Two kinds of server must be distinguished here. A server with no knowledge of EXCHGID4_FLAG_SUPP_RECALL_DEVICEID MUST reject the EXCHANGE_ID with NFS4ERR_INVAL, because Section 18.35.3 requires that of any undefined bit in eia_flags; it has no choice to make. A server that knows the flag from this document but does not implement deviceid-scoped recall accepts the EXCHANGE_ID and simply never sends LAYOUTRECALL4_DEVICEID. A client therefore cannot assume that setting the flag is free. It SHOULD probe as Section 4.4.3 describes: send the EXCHANGE_ID with the flag set, and on NFS4ERR_INVAL retry without it. The client cannot learn why the first attempt failed, since EXCHANGE_ID returns NFS4ERR_INVAL for several other reasons, so it MUST treat any NFS4ERR_INVAL as grounds to retry without the flag rather than as a permanent failure. A server MUST NOT set EXCHGID4_FLAG_SUPP_RECALL_DEVICEID in eir_flags unless the client set it in eia_flags. Section 18.35.3 permits a server to set a result bit the client did not request, but a client with no knowledge of this flag has no way to interpret it, and an implementation that validates eir_flags against the bits it knows will reject the reply. If the client does not set EXCHGID4_FLAG_SUPP_RECALL_DEVICEID, the server MAY fall back to recalling individual layouts via LAYOUTRECALL4_FILE (one CB_LAYOUTRECALL per layout file that references the unavailable deviceid), as described in Section 20.3 of . This is less efficient but correct, and preserves interoperability with clients that predate this extension.

Extension to CB_LAYOUTRECALL - Recall Layout from Client The original union layoutrecall4 (see Section 20.3.1 of ) is: The proposed extension is:

Completing a Deviceid-Scoped Recall A deviceid-scoped recall is a bulk recall, in the same sense that LAYOUTRECALL4_FSID is: it names a scope rather than a file, and Section 12.5.5.2.1.5 applies to it as written. This document defines no exception to that section. Section 18.44.3 requires a LAYOUTRETURN with a matching scope to complete a recall, even if all current layout ranges have been previously returned individually. The scope here is the deviceid, and the LAYOUT4RET_REC* constants are shared between layoutrecall_type4 (used in CB_LAYOUTRECALL) and layoutreturn_type4 (used in LAYOUTRETURN, see Section 18.44.1 of ), so adding LAYOUT4_RET_REC_DEVICEID = 4 gives layoutreturn_type4 a LAYOUTRETURN4_DEVICEID value as well. No new arm of layoutreturn4 is needed, and none may be added: that union carries a default case, which Section 4.2 forbids extending. None is needed because the aggregate returns carry no body. Section 18.44.3 identifies the file system of a LAYOUTRETURN4_FSID from the current filehandle, and uses no filehandle at all for LAYOUTRETURN4_ALL; LAYOUTRETURN4_DEVICEID is the same shape, and takes the default void arm. Adding the enum value is permitted on its own: Section 4.2 lists addition of new, previously unused values to existing enums among the allowed extensions. The recall being completed is identified by being the only one outstanding. Section 12.5.5.2.1.5 requires that a server not initiate a bulk recall while another recall is in progress or its LAYOUTRETURN is pending, so at most one deviceid-scoped recall is outstanding for a client at a time, and a void-bodied LAYOUTRETURN4_DEVICEID is unambiguous. An honoring client therefore completes the recall in one of two ways:

• it returns the affected layouts, individually or otherwise, and then sends a LAYOUTRETURN whose lr_returntype is LAYOUTRETURN4_DEVICEID; or
• it replies to the CB_LAYOUTRECALL with NFS4ERR_NOMATCHING_LAYOUT, when it holds no layout referencing the deviceid and no valid mapping for it ( Section 20.3.3).

Once that LAYOUTRETURN is sent, the layout stateids for the returned layouts are freed, and the client MUST send LAYOUTGET to obtain a new layout stateid before using a layout for any of those files again. A server with no knowledge of LAYOUTRETURN4_DEVICEID returns NFS4ERR_BADXDR, or an RPC XDR decode error, as Section 4.4.3 describes for an unrecognised switch arm. NFS4ERR_UNION_NOTSUPP is the answer of a responder that does know the arm, so it is not the error for this case. A client has no reason to probe for it: a server that sent LAYOUTRECALL4_DEVICEID has demonstrated knowledge of the value, and forbids sending that recall otherwise. The server MUST NOT send CB_LAYOUTRECALL with LAYOUTRECALL4_DEVICEID to a client that has not set EXCHGID4_FLAG_SUPP_RECALL_DEVICEID (see ). This satisfies the requirement in Section 6 of that a server establish client awareness before sending new callback extensions. The existing clora_iomode field in CB_LAYOUTRECALL4args (see Section 20.3.1 of ) applies normally: the client MUST return all layouts matching both the given deviceid and the given iomode. The server can determine that the client no longer has any layouts with the given deviceid and iomode once the client replies with NFS4ERR_NOMATCHING_LAYOUT.

Extraction of XDR This document contains the external data representation (XDR) description of the extension to CB_LAYOUTRECALL. The XDR description is presented in a manner that facilitates easy extraction into a ready-to-compile format. To extract the machine-readable XDR description, use the following shell script, which relies on the sh, grep, and sed utilities as specified by : #!/bin/sh grep '^ *///' $* | sed 's?^ */// ??' | sed 's?^ *///$??' ]]> For example, if the script is named 'extract.sh' and this document is named 'spec.txt', execute the following command: sh extract.sh < spec.txt > recalldevice.x ]]> This script removes leading blank spaces and the sentinel sequence '///' from each line. XDR descriptions with the sentinel sequence are embedded throughout the document. Note that the XDR code contained in this document depends on types from the NFSv4.2 nfs4_prot.x file (generated from ). This includes both nfs types that end with a 4, such as offset4, length4, etc., as well as more generic types such as uint32_t and uint64_t. The extracted XDR extends types that are already defined in rather than introducing new types. An implementer MUST integrate the extracted block into the base XDR by replacing the existing layoutrecall_type4 enum declaration and layoutrecall4 union declaration with the extended versions in this document, and by adding the LAYOUT4RET_REC_DEVICEID constant in the same file area as the existing LAYOUT4_RET_REC* constants. Appending the extracted block verbatim to the RFC 7863 XDR produces duplicate- type errors at compile time. The EXCHGID4_FLAG_SUPP_RECALL_DEVICEID constant is purely additive and is added to the EXCHGID4 flag constant block without replacing anything.

Security Considerations The extension introduced by this document inherits the Security Considerations of CB_LAYOUTRECALL in Section 20.3 of and of NFSv4.2 in . Three aspects of the deviceid-scoped recall warrant explicit mention.

Blast radius:

A single CB_LAYOUTRECALL with LAYOUTRECALL4_DEVICEID can cause a client to return many layouts in one callback exchange. This is not a new risk category -- LAYOUTRECALL4_FSID and LAYOUTRECALL4_ALL already allow wider-scoped recalls -- but administrators sizing callback-channel timeouts and client-side recall-handler resources SHOULD account for the possibility that a deviceid-scoped recall triggers per-layout cleanup work proportional to the number of layouts that reference the deviceid.

Cross-client scope:

In deployments where a single deviceid is shared across clients serving different tenants or applications, a deviceid-scoped recall is cross-client by construction: every client that holds a layout referencing the deviceid is recalled. This is the same property as LAYOUTRECALL4_FSID and does not introduce a new information-exposure surface. Deployments that prefer per-tenant failure-domain isolation MAY treat this extension as an operational benefit: by assigning distinct deviceids per tenant, deviceid-scoped recall becomes a per-tenant operation with no cross-tenant effect.

Capability negotiation as defence in depth:

The capability flag defined in ensures that a server cannot silently change the callback wire format for a client that did not opt in. A client that does not advertise EXCHGID4_FLAG_SUPP_RECALL_DEVICEID continues to receive only the layoutrecall_type4 values defined in . A server that sends LAYOUTRECALL4_DEVICEID to a client that did not advertise support is in violation of Section 6 of . Such a client does not recognise the union arm, so it responds as Section 4.4.3 describes for that case, with NFS4ERR_BADXDR or an RPC XDR decode error, and SHOULD log the server for operator attention.

IANA Considerations This document has no IANA actions.

References Normative References In many standards track documents several words are used to signify the requirements in the specification. These words are often capitalized. This document defines these words as they should be interpreted in IETF documents. This document specifies an Internet Best Current Practices for the Internet Community, and requests discussion and suggestions for improvements. This document describes the External Data Representation Standard (XDR) protocol as it is currently deployed and accepted. This document obsoletes RFC 1832. [STANDARDS-TRACK] This document describes NFS version 4 minor version 2; it describes the protocol extensions made from NFS version 4 minor version 1. Major extensions introduced in NFS version 4 minor version 2 include the following: Server-Side Copy, Application Input/Output (I/O) Advise, Space Reservations, Sparse Files, Application Data Blocks, and Labeled NFS. This document provides the External Data Representation (XDR) description for NFS version 4 minor version 2. RFC 2119 specifies common key words that may be used in protocol specifications. This document aims to reduce the ambiguity by clarifying that only UPPERCASE usage of the key words have the defined special meanings. This document describes the rules relating to the extension of the NFSv4 family of protocols. It covers the creation of minor versions, the addition of optional features to existing minor versions, and the correction of flaws in features already published as Proposed Standards. The rules relating to the construction of minor versions and the interaction of minor version implementations that appear in this document supersede the minor versioning rules in RFC 5661 and other RFCs defining minor versions. This document describes the Network File System (NFS) version 4 minor version 1, including features retained from the base protocol (NFS version 4 minor version 0, which is specified in RFC 7530) and protocol extensions made subsequently. The later minor version has no dependencies on NFS version 4 minor version 0, and is considered a separate protocol. This document obsoletes RFC 5661. It substantially revises the treatment of features relating to multi-server namespace, superseding the description of those features appearing in RFC 5661. Informative References IEEE The Open Group This paper describes the NFS version 3 protocol. This paper is provided so that people can write compatible implementations. This memo provides information for the Internet community. This memo does not specify an Internet standard of any kind. Parallel NFS (pNFS) allows a separation between the metadata (onto a metadata server) and data (onto a storage device) for a file. The flexible file layout type is defined in this document as an extension to pNFS that allows the use of storage devices that require only a limited degree of interaction with the metadata server and use already-existing protocols. Client-side mirroring is also added to provide replication of files.

Acknowledgments Trond Myklebust and Paul Saab were involved in the initial requirements for this functionality. Chuck Lever and Benjamin Coddington reviewed the document. Brian Pawlowski and Gorry Fairhurst helped guide this process.