NFSv4 S. Shepler Internet-Draft Editor Expires: June 15, 2006 December 12, 2005 NFSv4 Minor Version 1 draft-ietf-nfsv4-minorversion1-01.txt Status of this Memo By submitting this Internet-Draft, each author represents that any applicable patent or other IPR claims of which he or she is aware have been or will be disclosed, and any of which he or she becomes aware will be disclosed, in accordance with Section 6 of BCP 79. Internet-Drafts are working documents of the Internet Engineering Task Force (IETF), its areas, and its working groups. Note that other groups may also distribute working documents as Internet- Drafts. 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." The list of current Internet-Drafts can be accessed at http://www.ietf.org/ietf/1id-abstracts.txt. The list of Internet-Draft Shadow Directories can be accessed at http://www.ietf.org/shadow.html. This Internet-Draft will expire on June 15, 2006. Copyright Notice Copyright (C) The Internet Society (2005). Abstract This Internet-Draft describes the NFSv4 minor version 1 protocol extensions. These most significant of these extensions are commonly called: Sessions, Directory Delegations, and parallel NFS or pNFS Requirements Language 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 RFC 2119 [1]. Shepler Expires June 15, 2006 [Page 1] Internet-Draft NFSv4 Minior Version 1 December 2005 Table of Contents 1. Security Negotiation . . . . . . . . . . . . . . . . . . . . 6 2. Clarification of Security Negotiation in NFSv4.1 . . . . . . 6 2.1 PUTFH + LOOKUP . . . . . . . . . . . . . . . . . . . . . . 6 2.2 PUTFH + LOOKUPP . . . . . . . . . . . . . . . . . . . . . 7 2.3 PUTFH + SECINFO . . . . . . . . . . . . . . . . . . . . . 7 2.4 PUTFH + Anything Else . . . . . . . . . . . . . . . . . . 7 3. NFSv4.1 Sessions . . . . . . . . . . . . . . . . . . . . . . 8 3.1 Sessions Background . . . . . . . . . . . . . . . . . . . 8 3.1.1 Introduction to Sessions . . . . . . . . . . . . . . . 8 3.1.2 Motivation . . . . . . . . . . . . . . . . . . . . . . 9 3.1.3 Problem Statement . . . . . . . . . . . . . . . . . . 10 3.1.4 NFSv4 Session Extension Characteristics . . . . . . . 11 3.2 Transport Issues . . . . . . . . . . . . . . . . . . . . . 12 3.2.1 Session Model . . . . . . . . . . . . . . . . . . . . 12 3.2.2 Connection State . . . . . . . . . . . . . . . . . . . 13 3.2.3 NFSv4 Channels, Sessions and Connections . . . . . . . 14 3.2.4 Reconnection, Trunking and Failover . . . . . . . . . 16 3.2.5 Server Duplicate Request Cache . . . . . . . . . . . . 17 3.3 Session Initialization and Transfer Models . . . . . . . . 18 3.3.1 Session Negotiation . . . . . . . . . . . . . . . . . 18 3.3.2 RDMA Requirements . . . . . . . . . . . . . . . . . . 19 3.3.3 RDMA Connection Resources . . . . . . . . . . . . . . 20 3.3.4 TCP and RDMA Inline Transfer Model . . . . . . . . . . 21 3.3.5 RDMA Direct Transfer Model . . . . . . . . . . . . . . 23 3.4 Connection Models . . . . . . . . . . . . . . . . . . . . 26 3.4.1 TCP Connection Model . . . . . . . . . . . . . . . . . 27 3.4.2 Negotiated RDMA Connection Model . . . . . . . . . . . 28 3.4.3 Automatic RDMA Connection Model . . . . . . . . . . . 29 3.5 Buffer Management, Transfer, Flow Control . . . . . . . . 29 3.6 Retry and Replay . . . . . . . . . . . . . . . . . . . . . 32 3.7 The Back Channel . . . . . . . . . . . . . . . . . . . . . 33 3.8 COMPOUND Sizing Issues . . . . . . . . . . . . . . . . . . 34 3.9 Data Alignment . . . . . . . . . . . . . . . . . . . . . . 34 3.10 NFSv4 Integration . . . . . . . . . . . . . . . . . . . 36 3.10.1 Minor Versioning . . . . . . . . . . . . . . . . . . 36 3.10.2 Slot Identifiers and Server Duplicate Request Cache . . . . . . . . . . . . . . . . . . . . . . . 36 3.10.3 COMPOUND and CB_COMPOUND . . . . . . . . . . . . . . 40 3.10.4 eXternal Data Representation Efficiency . . . . . . 41 3.10.5 Effect of Sessions on Existing Operations . . . . . 41 3.10.6 Authentication Efficiencies . . . . . . . . . . . . 42 3.11 Sessions Security Considerations . . . . . . . . . . . . 43 3.11.1 Authentication . . . . . . . . . . . . . . . . . . . 44 4. Directory Delegations . . . . . . . . . . . . . . . . . . . 45 4.1 Introduction to Directory Delegations . . . . . . . . . . 45 4.2 Directory Delegation Design (in brief) . . . . . . . . . . 47 Shepler Expires June 15, 2006 [Page 2] Internet-Draft NFSv4 Minior Version 1 December 2005 4.3 Recommended Attributes in support of Directory Delegations . . . . . . . . . . . . . . . . . . . . . . . 48 4.4 Delegation Recall . . . . . . . . . . . . . . . . . . . . 48 4.5 Delegation Recovery . . . . . . . . . . . . . . . . . . . 49 5. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 49 6. General Definitions . . . . . . . . . . . . . . . . . . . . 51 6.1 Metadata Server . . . . . . . . . . . . . . . . . . . . . 52 6.2 Client . . . . . . . . . . . . . . . . . . . . . . . . . . 52 6.3 Storage Device . . . . . . . . . . . . . . . . . . . . . . 52 6.4 Storage Protocol . . . . . . . . . . . . . . . . . . . . . 52 6.5 Control Protocol . . . . . . . . . . . . . . . . . . . . . 53 6.6 Metadata . . . . . . . . . . . . . . . . . . . . . . . . . 53 6.7 Layout . . . . . . . . . . . . . . . . . . . . . . . . . . 53 7. pNFS protocol semantics . . . . . . . . . . . . . . . . . . 53 7.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . 54 7.1.1 Layout Types . . . . . . . . . . . . . . . . . . . . . 54 7.1.2 Layout Iomode . . . . . . . . . . . . . . . . . . . . 54 7.1.3 Layout Segments . . . . . . . . . . . . . . . . . . . 55 7.1.4 Device IDs . . . . . . . . . . . . . . . . . . . . . . 56 7.1.5 Aggregation Schemes . . . . . . . . . . . . . . . . . 56 7.2 Guarantees Provided by Layouts . . . . . . . . . . . . . . 56 7.3 Getting a Layout . . . . . . . . . . . . . . . . . . . . . 58 7.4 Committing a Layout . . . . . . . . . . . . . . . . . . . 58 7.4.1 LAYOUTCOMMIT and mtime/atime/change . . . . . . . . . 59 7.4.2 LAYOUTCOMMIT and size . . . . . . . . . . . . . . . . 60 7.4.3 LAYOUTCOMMIT and layoutupdate . . . . . . . . . . . . 61 7.5 Recalling a Layout . . . . . . . . . . . . . . . . . . . . 61 7.5.1 Basic Operation . . . . . . . . . . . . . . . . . . . 61 7.5.2 Recall Callback Robustness . . . . . . . . . . . . . . 62 7.5.3 Recall/Return Sequencing . . . . . . . . . . . . . . . 63 7.6 Metadata Server Write Propagation . . . . . . . . . . . . 65 7.7 Crash Recovery . . . . . . . . . . . . . . . . . . . . . . 66 7.7.1 Leases . . . . . . . . . . . . . . . . . . . . . . . . 66 7.7.2 Client Recovery . . . . . . . . . . . . . . . . . . . 67 7.7.3 Metadata Server Recovery . . . . . . . . . . . . . . . 68 7.7.4 Storage Device Recovery . . . . . . . . . . . . . . . 70 8. Security Considerations . . . . . . . . . . . . . . . . . . 71 8.1 File Layout Security . . . . . . . . . . . . . . . . . . . 72 8.2 Object Layout Security . . . . . . . . . . . . . . . . . . 72 8.3 Block/Volume Layout Security . . . . . . . . . . . . . . . 73 9. The NFSv4 File Layout Type . . . . . . . . . . . . . . . . . 74 9.1 File Striping and Data Access . . . . . . . . . . . . . . 74 9.1.1 Sparse and Dense Storage Device Data Layouts . . . . . 75 9.1.2 Metadata and Storage Device Roles . . . . . . . . . . 77 9.1.3 Device Multipathing . . . . . . . . . . . . . . . . . 78 9.1.4 Operations Issued to Storage Devices . . . . . . . . . 79 9.2 Global Stateid Requirements . . . . . . . . . . . . . . . 79 9.3 The Layout Iomode . . . . . . . . . . . . . . . . . . . . 80 Shepler Expires June 15, 2006 [Page 3] Internet-Draft NFSv4 Minior Version 1 December 2005 9.4 Storage Device State Propagation . . . . . . . . . . . . . 80 9.4.1 Lock State Propagation . . . . . . . . . . . . . . . . 80 9.4.2 Open-mode Validation . . . . . . . . . . . . . . . . . 81 9.4.3 File Attributes . . . . . . . . . . . . . . . . . . . 81 9.5 Storage Device Component File Size . . . . . . . . . . . . 82 9.6 Crash Recovery Considerations . . . . . . . . . . . . . . 83 9.7 Security Considerations . . . . . . . . . . . . . . . . . 83 9.8 Alternate Approaches . . . . . . . . . . . . . . . . . . . 84 10. pNFS Typed Data Structures . . . . . . . . . . . . . . . . . 85 10.1 pnfs_layouttype4 . . . . . . . . . . . . . . . . . . . . 85 10.2 pnfs_deviceid4 . . . . . . . . . . . . . . . . . . . . . 85 10.3 pnfs_deviceaddr4 . . . . . . . . . . . . . . . . . . . . 86 10.4 pnfs_devlist_item4 . . . . . . . . . . . . . . . . . . . 86 10.5 pnfs_layout4 . . . . . . . . . . . . . . . . . . . . . . 87 10.6 pnfs_layoutupdate4 . . . . . . . . . . . . . . . . . . . 87 10.7 pnfs_layouthint4 . . . . . . . . . . . . . . . . . . . . 88 10.8 pnfs_layoutiomode4 . . . . . . . . . . . . . . . . . . . 88 11. pNFS File Attributes . . . . . . . . . . . . . . . . . . . . 88 11.1 pnfs_layouttype4<> FS_LAYOUT_TYPES . . . . . . . . . . . 88 11.2 pnfs_layouttype4<> FILE_LAYOUT_TYPES . . . . . . . . . . 88 11.3 pnfs_layouthint4 FILE_LAYOUT_HINT . . . . . . . . . . . 89 11.4 uint32_t FS_LAYOUT_PREFERRED_BLOCKSIZE . . . . . . . . . 89 11.5 uint32_t FS_LAYOUT_PREFERRED_ALIGNMENT . . . . . . . . . 89 12. pNFS Error Definitions . . . . . . . . . . . . . . . . . . . 89 13. Layouts and Aggregation . . . . . . . . . . . . . . . . . . 90 13.1 Simple Map . . . . . . . . . . . . . . . . . . . . . . . 90 13.2 Block Extent Map . . . . . . . . . . . . . . . . . . . . 90 13.3 Striped Map (RAID 0) . . . . . . . . . . . . . . . . . . 90 13.4 Replicated Map . . . . . . . . . . . . . . . . . . . . . 91 13.5 Concatenated Map . . . . . . . . . . . . . . . . . . . . 91 13.6 Nested Map . . . . . . . . . . . . . . . . . . . . . . . 91 14. NFSv4.1 Operations . . . . . . . . . . . . . . . . . . . . . 91 14.1 LOOKUPP - Lookup Parent Directory . . . . . . . . . . . 91 14.2 SECINFO -- Obtain Available Security . . . . . . . . . . 93 14.3 SECINFO_NO_NAME - Get Security on Unnamed Object . . . . 96 14.4 CREATECLIENTID - Instantiate Clientid . . . . . . . . . 98 14.5 CREATESESSION - Create New Session and Confirm Clientid . . . . . . . . . . . . . . . . . . . . . . . . 104 14.6 BIND_BACKCHANNEL - Create a callback channel binding . . 109 14.7 DESTROYSESSION - Destroy existing session . . . . . . . 112 14.8 SEQUENCE - Supply per-procedure sequencing and control . 113 14.9 CB_RECALLCREDIT - change flow control limits . . . . . . 114 14.10 CB_SEQUENCE - Supply callback channel sequencing and control . . . . . . . . . . . . . . . . . . . . . . . . 115 14.11 GET_DIR_DELEGATION - Get a directory delegation . . . . 117 14.12 CB_NOTIFY - Notify directory changes . . . . . . . . . . 120 14.13 CB_RECALL_ANY - Keep any N delegations . . . . . . . . . 124 14.14 LAYOUTGET - Get Layout Information . . . . . . . . . . . 126 Shepler Expires June 15, 2006 [Page 4] Internet-Draft NFSv4 Minior Version 1 December 2005 14.15 LAYOUTCOMMIT - Commit writes made using a layout . . . . 128 14.16 LAYOUTRETURN - Release Layout Information . . . . . . . 131 14.17 GETDEVICEINFO - Get Device Information . . . . . . . . . 133 14.18 GETDEVICELIST - Get List of Devices . . . . . . . . . . 134 14.19 CB_LAYOUTRECALL . . . . . . . . . . . . . . . . . . . . 136 14.20 CB_SIZECHANGED . . . . . . . . . . . . . . . . . . . . . 138 15. References . . . . . . . . . . . . . . . . . . . . . . . . . 139 15.1 Normative References . . . . . . . . . . . . . . . . . . 139 15.2 Informative References . . . . . . . . . . . . . . . . . 139 Author's Address . . . . . . . . . . . . . . . . . . . . . . 139 A. Acknowledgments . . . . . . . . . . . . . . . . . . . . . . 139 Intellectual Property and Copyright Statements . . . . . . . 141 Shepler Expires June 15, 2006 [Page 5] Internet-Draft NFSv4 Minior Version 1 December 2005 1. Security Negotiation The NFSv4.0 specification contains three oversights and ambiguities with respect to the SECINFO operation. First, it is impossible for the client to use the SECINFO operation to determine the correct security triple for accessing a parent directory. This is because SECINFO takes as arguments the current file handle and a component name. However, NFSv4.0 uses the LOOKUPP operation to get the parent directory of the current file handle. If the client uses the wrong security when issuing the LOOKUPP, and gets back an NFS4ERR_WRONGSEC error, SECINFO is useless to the client. The client is left with guessing which security the server will accept. This defeats the purpose of SECINFO, which was to provide an efficient method of negotiating security. Second, there is ambiguity as to what the server should do when it is passed a LOOKUP operation such that the server restricts access to the current file handle with one security triple, and access to the component with a different triple, and remote procedure call uses one of the two security triples. Should the server allow the LOOKUP? Third, there is a problem as to what the client must do (or can do), whenever the server returns NFS4ERR_WRONGSEC in response to a PUTFH operation. The NFSv4.0 specification says that client should issue a SECINFO using the parent filehandle and the component name of the filehandle that PUTFH was issued with. This may not be convenient for the client. This document resolves the above three issues in the context of NFSv4.1. 2. Clarification of Security Negotiation in NFSv4.1 This section attempts to clarify NFSv4.1 security negotiation issues. Unless noted otherwise, for any mention of PUTFH in this section, the reader should interpret it as applying to PUTROOTFH and PUTPUBFH in addition to PUTFH. 2.1 PUTFH + LOOKUP The server implementation may decide whether to impose any restrictions on export security administration. There are at least three approaches (Sc is the flavor set of the child export, Sp that of the parent), Shepler Expires June 15, 2006 [Page 6] Internet-Draft NFSv4 Minior Version 1 December 2005 a) Sc <= Sp (<= for subset) b) Sc ^ Sp != {} (^ for intersection, {} for the empty set) c) free form To support b (when client chooses a flavor that is not a member of Sp) and c, PUTFH must NOT return NFS4ERR_WRONGSEC in case of security mismatch. Instead, it should be returned from the LOOKUP that follows. Since the above guideline does not contradict a, it should be followed in general. 2.2 PUTFH + LOOKUPP Since SECINFO only works its way down, there is no way LOOKUPP can return NFS4ERR_WRONGSEC without the server implementing SECINFO_NO_NAME. SECINFO_NO_NAME solves this issue because via style "parent", it works in the opposite direction as SECINFO (component name is implicit in this case). 2.3 PUTFH + SECINFO This case should be treated specially. A security sensitive client should be allowed to choose a strong flavor when querying a server to determine a file object's permitted security flavors. The security flavor chosen by the client does not have to be included in the flavor list of the export. Of course the server has to be configured for whatever flavor the client selects, otherwise the request will fail at RPC authentication. In theory, there is no connection between the security flavor used by SECINFO and those supported by the export. But in practice, the client may start looking for strong flavors from those supported by the export, followed by those in the mandatory set. 2.4 PUTFH + Anything Else PUTFH must return NFS4ERR_WRONGSEC in case of security mismatch. This is the most straightforward approach without having to add NFS4ERR_WRONGSEC to every other operations. PUTFH + SECINFO_NO_NAME (style "current_fh") is needed for the client to recover from NFS4ERR_WRONGSEC. Shepler Expires June 15, 2006 [Page 7] Internet-Draft NFSv4 Minior Version 1 December 2005 3. NFSv4.1 Sessions 3.1 Sessions Background 3.1.1 Introduction to Sessions This draft proposes extensions to NFS version 4 [RFC3530] enabling it to support sessions and endpoint management, and to support operation atop RDMA-capable RPC over transports such as iWARP. [RDMAP, DDP] These extensions enable support for exactly-once semantics by NFSv4 servers, multipathing and trunking of transport connections, and enhanced security. The ability to operate over RDMA enables greatly enhanced performance. Operation over existing TCP is enhanced as well. While discussed here with respect to IETF-chartered transports, the proposed protocol is intended to function over other standards, such as Infiniband. [IB] The following are the major aspects of this proposal: Changes are proposed within the framework of NFSv4 minor versioning. RPC, XDR, and the NFSv4 procedures and operations are preserved. The proposed extension functions equally well over existing transports and RDMA, and interoperates transparently with existing implementations, both at the local programmatic interface and over the wire. An explicit session is introduced to NFSv4, and new operations are added to support it. The session allows for enhanced trunking, failover and recovery, and authentication efficiency, along with necessary support for RDMA. The session is implemented as operations within NFSv4 COMPOUND and does not impact layering or interoperability with existing NFSv4 implementations. The NFSv4 callback channel is dynamically associated and is connected by the client and not the server, enhancing security and operation through firewalls. In fact, the callback channel will be enabled to share the same connection as the operations channel. An enhanced RPC layer enables NFSv4 operation atop RDMA. The session assists RDMA-mode connection, and additional facilities are provided for managing RDMA resources at both NFSv4 server and client. Existing NFSv4 operations continue to function as before, though certain size limits are negotiated. A companion draft to this document, "RDMA Transport for ONC RPC" [RPCRDMA] is to be referenced for details of RPC RDMA support. Shepler Expires June 15, 2006 [Page 8] Internet-Draft NFSv4 Minior Version 1 December 2005 Support for exactly-once semantics ("EOS") is enabled by the new session facilities, by providing to the server a way to bound the size of the duplicate request cache for a single client, and to manage its persistent storage. Block Diagram +-----------------+-------------------------------------+ | NFSv4 | NFSv4 + session extensions | +-----------------+------+----------------+-------------+ | Operations | Session | | +------------------------+----------------+ | | RPC/XDR | | +-------------------------------+---------+ | | Stream Transport | RDMA Transport | +-------------------------------+-----------------------+ 3.1.2 Motivation NFS version 4 [RFC3530] has been granted "Proposed Standard" status. The NFSv4 protocol was developed along several design points, important among them: effective operation over wide-area networks, including the Internet itself; strong security integrated into the protocol; extensive cross-platform interoperability including integrated locking semantics compatible with multiple operating systems; and protocol extensibility. The NFS version 4 protocol, however, does not provide support for certain important transport aspects. For example, the protocol does not address response caching, which is required to provide correctness for retried client requests across a network partition, nor does it provide an interoperable way to support trunking and multipathing of connections. This leads to inefficiencies, especially where trunking and multipathing are concerned, and presents additional difficulties in supporting RDMA fabrics, in which endpoints may require dedicated or specialized resources. Sessions can be employed to unify NFS-level constructs such as the clientid, with transport-level constructs such as transport endpoints. Each transport endpoint draws on resources via its membership in a session. Resource management can be more strictly maintained, leading to greater server efficiency in implementing the protocol. The enhanced operation over a session affords an opportunity to the server to implement a highly reliable duplicate request cache, and thereby export exactly-once semantics. NFSv4 advances the state of high-performance local sharing, by virtue Shepler Expires June 15, 2006 [Page 9] Internet-Draft NFSv4 Minior Version 1 December 2005 of its integrated security, locking, and delegation, and its excellent coverage of the sharing semantics of multiple operating systems. It is precisely this environment where exactly-once semantics become a fundamental requirement. Additionally, efforts to standardize a set of protocols for Remote Direct Memory Access, RDMA, over the Internet Protocol Suite have made significant progress. RDMA is a general solution to the problem of CPU overhead incurred due to data copies, primarily at the receiver. Substantial research has addressed this and has borne out the efficacy of the approach. An overview of this is the RDDP Problem Statement document, [RDDPPS]. Numerous upper layer protocols achieve extremely high bandwidth and low overhead through the use of RDMA. Products from a wide variety of vendors employ RDMA to advantage, and prototypes have demonstrated the effectiveness of many more. Here, we are concerned specifically with NFS and NFS-style upper layer protocols; examples from Network Appliance [DAFS, DCK+03], Fujitsu Prime Software Technologies [FJNFS, FJDAFS] and Harvard University [KM02] are all relevant. By layering a session binding for NFS version 4 directly atop a standard RDMA transport, a greatly enhanced level of performance and transparency can be supported on a wide variety of operating system platforms. These combined capabilities alter the landscape between local filesystems and network attached storage, enable a new level of performance, and lead new classes of application to take advantage of NFS. 3.1.3 Problem Statement Two issues drive the current proposal: correctness, and performance. Both are instances of "raising the bar" for NFS, whereby the desire to use NFS in new classes applications can be accommodated by providing the basic features to make such use feasible. Such applications include tightly coupled sharing environments such as cluster computing, high performance computing (HPC) and information processing such as databases. These trends are explored in depth in [NFSPS]. The first issue, correctness, exemplified among the attributes of local filesystems, is support for exactly-once semantics. Such semantics have not been reliably available with NFS. Server-based duplicate request caches [CJ89] help, but do not reliably provide strict correctness. For the type of application which is expected to make extensive use of the high-performance RDMA-enabled environment, the reliable provision of such semantics is a fundamental requirement. Shepler Expires June 15, 2006 [Page 10] Internet-Draft NFSv4 Minior Version 1 December 2005 Introduction of a session to NFSv4 will address these issues. With higher performance and enhanced semantics comes the problem of enabling advanced endpoint management, for example high-speed trunking, multipathing and failover. These characteristics enable availability and performance. RFC3530 presents some issues in permitting a single clientid to access a server over multiple connections. A second issue encountered in common by NFS implementations is the CPU overhead required to implement the protocol. Primary among the sources of this overhead is the movement of data from NFS protocol messages to its eventual destination in user buffers or aligned kernel buffers. The data copies consume system bus bandwidth and CPU time, reducing the available system capacity for applications. [RDDPPS] Achieving zero-copy with NFS has to date required sophisticated, "header cracking" hardware and/or extensive platform- specific virtual memory mapping tricks. Combined in this way, NFSv4, RDMA and the emerging high-speed network fabrics will enable delivery of performance which matches that of the fastest local filesystems, preserving the key existing local filesystem semantics, while enhancing them by providing network filesystem sharing semantics. RDMA implementations generally have other interesting properties, such as hardware assisted protocol access, and support for user space access to I/O. RDMA is compelling here for another reason; hardware offloaded networking support in itself does not avoid data copies, without resorting to implementing part of the NFS protocol in the NIC. Support of RDMA by NFS enables the highest performance at the architecture level rather than by implementation; this enables ubiquitous and interoperable solutions. By providing file access performance equivalent to that of local file systems, NFSv4 over RDMA will enable applications running on a set of client machines to interact through an NFSv4 file system, just as applications running on a single machine might interact through a local file system. This raises the issue of whether additional protocol enhancements to enable such interaction would be desirable and what such enhancements would be. This is a complicated issue which the working group needs to address and will not be further discussed in this document. 3.1.4 NFSv4 Session Extension Characteristics This draft will present a solution based upon minor versioning of NFSv4. It will introduce a session to collect transport endpoints Shepler Expires June 15, 2006 [Page 11] Internet-Draft NFSv4 Minior Version 1 December 2005 and resources such as reply caching, which in turn enables enhancements such as trunking, failover and recovery. It will describe use of RDMA by employing support within an underlying RPC layer [RPCRDMA]. Most importantly, it will focus on making the best possible use of an RDMA transport. These extensions are proposed as elements of a new minor revision of NFS version 4. In this draft, NFS version 4 will be referred to generically as "NFSv4", when describing properties common to all minor versions. When referring specifically to properties of the original, minor version 0 protocol, "NFSv4.0" will be used, and changes proposed here for minor version 1 will be referred to as "NFSv4.1". This draft proposes only changes which are strictly upward- compatible with existing RPC and NFS Application Programming Interfaces (APIs). 3.2 Transport Issues The Transport Issues section of the document explores the details of utilizing the various supported transports. 3.2.1 Session Model The first and most evident issue in supporting diverse transports is how to provide for their differences. This draft proposes introducing an explicit session. A session introduces minimal protocol requirements, and provides for a highly useful and convenient way to manage numerous endpoint- related issues. The session is a local construct; it represents a named, higher-layer object to which connections can refer, and encapsulates properties important to each associated client. A session is a dynamically created, long-lived server object created by a client, used over time from one or more transport connections. Its function is to maintain the server's state relative to the connection(s) belonging to a client instance. This state is entirely independent of the connection itself. The session in effect becomes the object representing an active client on a connection or set of connections. Clients may create multiple sessions for a single clientid, and may wish to do so for optimization of transport resources, buffers, or server behavior. A session could be created by the client to represent a single mount point, for separate read and write "channels", or for any number of other client-selected parameters. Shepler Expires June 15, 2006 [Page 12] Internet-Draft NFSv4 Minior Version 1 December 2005 The session enables several things immediately. Clients may disconnect and reconnect (voluntarily or not) without loss of context at the server. (Of course, locks, delegations and related associations require special handling, and generally expire in the extended absence of an open connection.) Clients may connect multiple transport endpoints to this common state. The endpoints may have all the same attributes, for instance when trunked on multiple physical network links for bandwidth aggregation or path failover. Or, the endpoints can have specific, special purpose attributes such as callback channels. The NFSv4 specification does not provide for any form of flow control; instead it relies on the windowing provided by TCP to throttle requests. This unfortunately does not work with RDMA, which in general provides no operation flow control and will terminate a connection in error when limits are exceeded. Limits are therefore exchanged when a session is created; These limits then provide maxima within which each session's connections must operate, they are managed within these limits as described in [RPCRDMA]. The limits may also be modified dynamically at the server's choosing by manipulating certain parameters present in each NFSv4.1 request. The presence of a maximum request limit on the session bounds the requirements of the duplicate request cache. This can be used to advantage by a server, which can accurately determine any storage needs and enable it to maintain duplicate request cache persistence and to provide reliable exactly-once semantics. Finally, given adequate connection-oriented transport security semantics, authentication and authorization may be cached on a per- session basis, enabling greater efficiency in the issuing and processing of requests on both client and server. A proposal for transparent, server-driven implementation of this in NFSv4 has been made. [CCM] The existence of the session greatly facilitates the implementation of this approach. This is discussed in detail in the Authentication Efficiencies section later in this draft. 3.2.2 Connection State In RFC3530, the combination of a connected transport endpoint and a clientid forms the basis of connection state. While has been made to be workable with certain limitations, there are difficulties in correct and robust implementation. The NFSv4.0 protocol must provide a server-initiated connection for the callback channel, and must carefully specify the persistence of client state at the server in the face of transport interruptions. The server has only the client's transport address binding (the IP 4-tuple) to identify the client RPC transaction stream and to use as a lookup tag on the Shepler Expires June 15, 2006 [Page 13] Internet-Draft NFSv4 Minior Version 1 December 2005 duplicate request cache. (A useful overview of this is in [RW96].) If the server listens on multiple adddresses, and the client connects to more than one, it must employ different clientid's on each, negating its ability to aggregate bandwidth and redundancy. In effect, each transport connection is used as the server's representation of client state. But, transport connections are potentially fragile and transitory. In this proposal, a session identifier is assigned by the server upon initial session negotiation on each connection. This identifier is used to associate additional connections, to renegotiate after a reconnect, to provide an abstraction for the various session properties, and to address the duplicate request cache. No transport-specific information is used in the duplicate request cache implementation of an NFSv4.1 server, nor in fact the RPC XID itself. The session identifier is unique within the server's scope and may be subject to certain server policies such as being bounded in time. It is envisioned that the primary transport model will be connection oriented. Connection orientation brings with it certain potential optimizations, such as caching of per-connection properties, which are easily leveraged through the generality of the session. However, it is possible that in future, other transport models could be accommodated below the session abstraction. 3.2.3 NFSv4 Channels, Sessions and Connections There are at least two types of NFSv4 channels: the "operations" channel used for ordinary requests from client to server, and the "back" channel, used for callback requests from server to client. As mentioned above, different NFSv4 operations on these channels can lead to different resource needs. For example, server callback operations (CB_RECALL) are specific, small messages which flow from server to client at arbitrary times, while data transfers such as read and write have very different sizes and asymmetric behaviors. It is sometimes impractical for the RDMA peers (NFSv4 client and NFSv4 server) to post buffers for these various operations on a single connection. Commingling of requests with responses at the client receive queue is particularly troublesome, due both to the need to manage both solicited and unsolicited completions, and to provision buffers for both purposes. Due to the lack of any ordering of callback requests versus response arrivals, without any other mechanisms, the client would be forced to allocate all buffers sized to the worst case. The callback requests are likely to be handled by a different task context from that handling the responses. Significant demultiplexing Shepler Expires June 15, 2006 [Page 14] Internet-Draft NFSv4 Minior Version 1 December 2005 and thread management may be required if both are received on the same queue. However, if callbacks are relatively rare (perhaps due to client access patterns), many of these difficulties can be minimized. Also, the client may wish to perform trunking of operations channel requests for performance reasons, or multipathing for availability. This proposal permits both, as well as many other session and connection possibilities, by permitting each operation to carry session membership information and to share session (and clientid) state in order to draw upon the appropriate resources. For example, reads and writes may be assigned to specific, optimized connections, or sorted and separated by any or all of size, idempotency, etc. To address the problems described above, this proposal allows multiple sessions to share a clientid, as well as for multiple connections to share a session. Single Connection model: NFSv4.1 Session / \ Operations_Channel [Back_Channel] \ / Connection | Multi-connection trunked model (2 operations channels shown): NFSv4.1 Session / \ Operations_Channels [Back_Channel] | | | Connection Connection [Connection] | | | Multi-connection split-use model (2 mounts shown): NFSv4.1 Session / \ (/home) (/usr/local - readonly) / \ | Operations_Channel [Back_Channel] | | | Operations_Channel Connection [Connection] | | | Connection Shepler Expires June 15, 2006 [Page 15] Internet-Draft NFSv4 Minior Version 1 December 2005 | In this way, implementation as well as resource management may be optimized. Each session will have its own response caching and buffering, and each connection or channel will have its own transport resources, as appropriate. Clients which do not require certain behaviors may optimize such resources away completely, by using specific sessions and not even creating the additional channels and connections. 3.2.4 Reconnection, Trunking and Failover Reconnection after failure references stored state on the server associated with lease recovery during the grace period. The session provides a convenient handle for storing and managing information regarding the client's previous state on a per- connection basis, e.g. to be used upon reconnection. Reconnection to a previously existing session, and its stored resources, are covered in the "Connection Models" section below. One important aspect of reconnection is that of RPC library support. Traditionally, an Upper Layer RPC-based Protocol such as NFS leaves all transport knowledge to the RPC layer implementation below it. This allows NFS to operate over a wide variety of transports and has proven to be a highly successful approach. The session, however, introduces an abstraction which is, in a way, "between" RPC and NFSv4.1. It is important that the session abstraction not have ramifications within the RPC layer. One such issue arises within the reconnection logic of RPC. Previously, an explicit session binding operation, which established session context for each new connection, was explored. This however required that the session binding also be performed during reconnect, which in turn required an RPC request. This additional request requires new RPC semantics, both in implementation and the fact that a new request is inserted into the RPC stream. Also, the binding of a connection to a session required the upper layer to become "aware" of connections, something the RPC layer abstraction architecturally abstracts away. Therefore the session binding is not handled in connection scope but instead explicitly carried in each request. For Reliability Availability and Serviceability (RAS) issues such as bandwidth aggregation and multipathing, clients frequently seek to make multiple connections through multiple logical or physical channels. The session is a convenient point to aggregate and manage these resources. Shepler Expires June 15, 2006 [Page 16] Internet-Draft NFSv4 Minior Version 1 December 2005 3.2.5 Server Duplicate Request Cache Server duplicate request caches, while not a part of an NFS protocol, have become a standard, even required, part of any NFS implementation. First described in [CJ89], the duplicate request cache was initially found to reduce work at the server by avoiding duplicate processing for retransmitted requests. A second, and in the long run more important benefit, was improved correctness, as the cache avoided certain destructive non-idempotent requests from being reinvoked. However, such caches do not provide correctness guarantees; they cannot be managed in a reliable, persistent fashion. The reason is understandable - their storage requirement is unbounded due to the lack of any such bound in the NFS protocol, and they are dependent on transport addresses for request matching. As proposed in this draft, the presence of maximum request count limits and negotiated maximum sizes allows the size and duration of the cache to be bounded, and coupled with a long-lived session identifier, enables its persistent storage on a per-session basis. This provides a single unified mechanism which provides the following guarantees required in the NFSv4 specification, while extending them to all requests, rather than limiting them only to a subset of state- related requests: "It is critical the server maintain the last response sent to the client to provide a more reliable cache of duplicate non- idempotent requests than that of the traditional cache described in [CJ89]..." [RFC3530] The maximum request count limit is the count of active operations, which bounds the number of entries in the cache. Constraining the size of operations additionally serves to limit the required storage to the product of the current maximum request count and the maximum response size. This storage requirement enables server- side efficiencies. Session negotiation allows the server to maintain other state. An NFSv4.1 client invoking the session destroy operation will cause the server to denegotiate (close) the session, allowing the server to deallocate cache entries. Clients can potentially specify that such caches not be kept for appropriate types of sessions (for example, read-only sessions). This can enable more efficient server operation resulting in improved response times, and more efficient sizing of buffers and response caches. Shepler Expires June 15, 2006 [Page 17] Internet-Draft NFSv4 Minior Version 1 December 2005 Similarly, it is important for the client to explicitly learn whether the server is able to implement reliable semantics. Knowledge of whether these semantics are in force is critical for a highly reliable client, one which must provide transactional integrity guarantees. When clients request that the semantics be enabled for a given session, the session reply must inform the client if the mode is in fact enabled. In this way the client can confidently proceed with operations without having to implement consistency facilities of its own. 3.3 Session Initialization and Transfer Models Session initialization issues, and data transfer models relevant to both TCP and RDMA are discussed in this section. 3.3.1 Session Negotiation The following parameters are exchanged between client and server at session creation time. Their values allow the server to properly size resources allocated in order to service the client's requests, and to provide the server with a way to communicate limits to the client for proper and optimal operation. They are exchanged prior to all session-related activity, over any transport type. Discussion of their use is found in their descriptions as well as throughout this section. Maximum Requests The client's desired maximum number of concurrent requests is passed, in order to allow the server to size its reply cache storage. The server may modify the client's requested limit downward (or upward) to match its local policy and/or resources. Over RDMA-capable RPC transports, the per-request management of low-level transport message credits is handled within the RPC layer. [RPCRDMA] Maximum Request/Response Sizes The maximum request and response sizes are exchanged in order to permit allocation of appropriately sized buffers and request cache entries. The size must allow for certain protocol minima, allowing the receipt of maximally sized operations (e.g. RENAME requests which contains two name strings). Note the maximum request/response sizes cover the entire request/response message and not simply the data payload as traditional NFS maximum read or write size. Also note the server implementation may not, in fact probably does not, require the reply cache entries to be sized as large as the maximum response. The server may reduce the client's Shepler Expires June 15, 2006 [Page 18] Internet-Draft NFSv4 Minior Version 1 December 2005 requested sizes. Inline Padding/Alignment The server can inform the client of any padding which can be used to deliver NFSv4 inline WRITE payloads into aligned buffers. Such alignment can be used to avoid data copy operations at the server for both TCP and inline RDMA transfers. For RDMA, the client informs the server in each operation when padding has been applied. [RPCRDMA] Transport Attributes A placeholder for transport-specific attributes is provided, with a format to be determined. Possible examples of information to be passed in this parameter include transport security attributes to be used on the connection, RDMA- specific attributes, legacy "private data" as used on existing RDMA fabrics, transport Quality of Service attributes, etc. This information is to be passed to the peer's transport layer by local means which is currently outside the scope of this draft, however one attribute is provided in the RDMA case: RDMA Read Resources RDMA implementations must explicitly provision resources to support RDMA Read requests from connected peers. These values must be explicitly specified, to provide adequate resources for matching the peer's expected needs and the connection's delay- bandwidth parameters. The client provides its chosen value to the server in the initial session creation, the value must be provided in each client RDMA endpoint. The values are asymmetric and should be set to zero at the server in order to conserve RDMA resources, since clients do not issue RDMA Read operations in this proposal. The result is communicated in the session response, to permit matching of values across the connection. The value may not be changed in the duration of the session, although a new value may be requested as part of a new session. 3.3.2 RDMA Requirements A complete discussion of the operation of RPC-based protocols atop RDMA transports is in [RPCRDMA]. Where RDMA is considered, this proposal assumes the use of such a layering; it addresses only the upper layer issues relevant to making best use of RPC/RDMA. A connection oriented (reliable sequenced) RDMA transport will be Shepler Expires June 15, 2006 [Page 19] Internet-Draft NFSv4 Minior Version 1 December 2005 required. There are several reasons for this. First, this model most closely reflects the general NFSv4 requirement of long-lived and congestion-controlled transports. Second, to operate correctly over either an unreliable or unsequenced RDMA transport, or both, would require significant complexity in the implementation and protocol not appropriate for a strict minor version. For example, retransmission on connected endpoints is explicitly disallowed in the current NFSv4 draft; it would again be required with these alternate transport characteristics. Third, the proposal assumes a specific RDMA ordering semantic, which presents the same set of ordering and reliability issues to the RDMA layer over such transports. The RDMA implementation provides for making connections to other RDMA-capable peers. In the case of the current proposals before the RDDP working group, these RDMA connections are preceded by a "streaming" phase, where ordinary TCP (or NFS) traffic might flow. However, this is not assumed here and sizes and other parameters are explicitly exchanged upon a session entering RDMA mode. 3.3.3 RDMA Connection Resources On transport endpoints which support automatic RDMA mode, that is, endpoints which are created in the RDMA-enabled state, a single, preposted buffer must initially be provided by both peers, and the client session negotiation must be the first exchange. On transport endpoints supporting dynamic negotiation, a more sophisticated negotiation is possible, but is not discussed in the current draft. RDMA imposes several requirements on upper layer consumers. Registration of memory and the need to post buffers of a specific size and number for receive operations are a primary consideration. Registration of memory can be a relatively high-overhead operation, since it requires pinning of buffers, assignment of attributes (e.g. readable/writable), and initialization of hardware translation. Preregistration is desirable to reduce overhead. These registrations are specific to hardware interfaces and even to RDMA connection endpoints, therefore negotiation of their limits is desirable to manage resources effectively. Following the basic registration, these buffers must be posted by the RPC layer to handle receives. These buffers remain in use by the RPC/NFSv4 implementation; the size and number of them must be known to the remote peer in order to avoid RDMA errors which would cause a fatal error on the RDMA connection. Shepler Expires June 15, 2006 [Page 20] Internet-Draft NFSv4 Minior Version 1 December 2005 The session provides a natural way for the server to manage resource allocation to each client rather than to each transport connection itself. This enables considerable flexibility in the administration of transport endpoints. 3.3.4 TCP and RDMA Inline Transfer Model The basic transfer model for both TCP and RDMA is referred to as "inline". For TCP, this is the only transfer model supported, since TCP carries both the RPC header and data together in the data stream. For RDMA, the RDMA Send transfer model is used for all NFS requests and replies, but data is optionally carried by RDMA Writes or RDMA Reads. Use of Sends is required to ensure consistency of data and to deliver completion notifications. The pure-Send method is typically used where the data payload is small, or where for whatever reason target memory for RDMA is not available. Inline message exchange Client Server : Request : Send : ------------------------------> : untagged : : buffer : Response : untagged : <------------------------------ : Send buffer : : Client Server : Read request : Send : ------------------------------> : untagged : : buffer : Read response with data : untagged : <------------------------------ : Send buffer : : Client Server : Write request with data : Send : ------------------------------> : untagged : : buffer : Write response : untagged : <------------------------------ : Send buffer : : Responses must be sent to the client on the same connection that the request was sent. It is important that the server does not assume Shepler Expires June 15, 2006 [Page 21] Internet-Draft NFSv4 Minior Version 1 December 2005 any specific client implementation, in particular whether connections within a session share any state at the client. This is also important to preserve ordering of RDMA operations, and especially RMDA consistency. Additionally, it ensures that the RPC RDMA layer makes no requirement of the RDMA provider to open its memory registration handles (Steering Tags) beyond the scope of a single RDMA connection. This is an important security consideration. Two values must be known to each peer prior to issuing Sends: the maximum number of sends which may be posted, and their maximum size. These values are referred to, respectively, as the message credits and the maximum message size. While the message credits might vary dynamically over the duration of the session, the maximum message size does not. The server must commit to preserving this number of duplicate request cache entires, and preparing a number of receive buffers equal to or greater than its currently advertised credit value, each of the advertised size. These ensure that transport resources are allocated sufficient to receive the full advertised limits. Note that the server must post the maximum number of session requests to each client operations channel. The client is not required to spread its requests in any particular fashion across connections within a session. If the client wishes, it may create multiple sessions, each with a single or small number of operations channels to provide the server with this resource advantage. Or, over RDMA the server may employ a "shared receive queue". The server can in any case protect its resources by restricting the client's request credits. While tempting to consider, it is not possible to use the TCP window as an RDMA operation flow control mechanism. First, to do so would violate layering, requiring both senders to be aware of the existing TCP outbound window at all times. Second, since requests are of variable size, the TCP window can hold a widely variable number of them, and since it cannot be reduced without actually receiving data, the receiver cannot limit the sender. Third, any middlebox interposing on the connection would wreck any possible scheme. [MIDTAX] In this proposal, maximum request count limits are exchanged at the session level to allow correct provisioning of receive buffers by transports. When operating over TCP or other similar transport, request limits and sizes are still employed in NFSv4.1, but instead of being required for correctness, they provide the basis for efficient server implementation of the duplicate request cache. The limits are chosen based upon the expected needs and capabilities of the client and server, and are in fact arbitrary. Sizes may be specified by the Shepler Expires June 15, 2006 [Page 22] Internet-Draft NFSv4 Minior Version 1 December 2005 client as zero (requesting the server's preferred or optimal value), and request limits may be chosen in proportion to the client's capabilities. For example, a limit of 1000 allows 1000 requests to be in progress, which may generally be far more than adequate to keep local networks and servers fully utilized. Both client and server have independent sizes and buffering, but over RDMA fabrics client credits are easily managed by posting a receive buffer prior to sending each request. Each such buffer may not be completed with the corresponding reply, since responses from NFSv4 servers arrive in arbitrary order. When an operations channel is also used for callbacks, the client must account for callback requests by posting additional buffers. Note that implementation- specific facilities such as a shared receive queue may also allow optimization of these allocations. When a session is created, the client requests a preferred buffer size, and the server provides its answer. The server posts all buffers of at least this size. The client must comply by not sending requests greater than this size. It is recommended that server implementations do all they can to accommodate a useful range of possible client requests. There is a provision in [RPCRDMA] to allow the sending of client requests which exceed the server's receive buffer size, but it requires the server to "pull" the client's request as a "read chunk" via RDMA Read. This introduces at least one additional network roundtrip, plus other overhead such as registering memory for RDMA Read at the client and additional RDMA operations at the server, and is to be avoided. An issue therefore arises when considering the NFSv4 COMPOUND procedures. Since an arbitrary number (total size) of operations can be specified in a single COMPOUND procedure, its size is effectively unbounded. This cannot be supported by RDMA Sends, and therefore this size negotiation places a restriction on the construction and maximum size of both COMPOUND requests and responses. If a COMPOUND results in a reply at the server that is larger than can be sent in an RDMA Send to the client, then the COMPOUND must terminate and the operation which causes the overflow will provide a TOOSMALL error status result. 3.3.5 RDMA Direct Transfer Model Placement of data by explicitly tagged RDMA operations is referred to as "direct" transfer. This method is typically used where the data payload is relatively large, that is, when RDMA setup has been performed prior to the operation, or when any overhead for setting up and performing the transfer is regained by avoiding the overhead of processing an ordinary receive. Shepler Expires June 15, 2006 [Page 23] Internet-Draft NFSv4 Minior Version 1 December 2005 The client advertises RDMA buffers in this proposed model, and not the server. This means the "XDR Decoding with Read Chunks" described in [RPCRDMA] is not employed by NFSv4.1 replies, and instead all results transferred via RDMA to the client employ "XDR Decoding with Write Chunks". There are several reasons for this. First, it allows for a correct and secure mode of transfer. The client may advertise specific memory buffers only during specific times, and may revoke access when it pleases. The server is not required to expose copies of local file buffers for individual clients, or to lock or copy them for each client access. Second, client credits based on fixed-size request buffers are easily managed on the server, but for the server additional management of buffers for client RDMA Reads is not well-bounded. For example, the client may not perform these RDMA Read operations in a timely fashion, therefore the server would have to protect itself against denial-of-service on these resources. Third, it reduces network traffic, since buffer exposure outside the scope and duration of a single request/response exchange necessitates additional memory management exchanges. There are costs associated with this decision. Primary among them is the need for the server to employ RDMA Read for operations such as large WRITE. The RDMA Read operation is a two-way exchange at the RDMA layer, which incurs additional overhead relative to RDMA Write. Additionally, RDMA Read requires resources at the data source (the client in this proposal) to maintain state and to generate replies. These costs are overcome through use of pipelining with credits, with sufficient RDMA Read resources negotiated at session initiation, and appropriate use of RDMA for writes by the client - for example only for transfers above a certain size. A description of which NFSv4 operation results are eligible for data transfer via RDMA Write is in [NFSDDP]. There are only two such operations: READ and READLINK. When XDR encoding these requests on an RDMA transport, the NFSv4.1 client must insert the appropriate xdr_write_list entries to indicate to the server whether the results should be transferred via RDMA or inline with a Send. As described in [NFSDDP], a zero-length write chunk is used to indicate an inline result. In this way, it is unnecessary to create new operations for RDMA-mode versions of READ and READLINK. Another tool to avoid creation of new, RDMA-mode operations is the Reply Chunk [RPCRDMA], which is used by RPC in RDMA mode to return large replies via RDMA as if they were inline. Reply chunks are used for operations such as READDIR, which returns large amounts of Shepler Expires June 15, 2006 [Page 24] Internet-Draft NFSv4 Minior Version 1 December 2005 information, but in many small XDR segments. Reply chunks are offered by the client and the server can use them in preference to inline. Reply chunks are transparent to upper layers such as NFSv4. In any very rare cases where another NFSv4.1 operation requires larger buffers than were negotiated when the session was created (for example extraordinarily large RENAMEs), the underlying RPC layer may support the use of "Message as an RDMA Read Chunk" and "RDMA Write of Long Replies" as described in [RPCRDMA]. No additional support is required in the NFSv4.1 client for this. The client should be certain that its requested buffer sizes are not so small as to make this a frequent occurrence, however. All operations are initiated by a Send, and are completed with a Send. This is exactly as in conventional NFSv4, but under RDMA has a significant purpose: RDMA operations are not complete, that is, guaranteed consistent, at the data sink until followed by a successful Send completion (i.e. a receive). These events provide a natural opportunity for the initiator (client) to enable and later disable RDMA access to the memory which is the target of each operation, in order to provide for consistent and secure operation. The RDMAP Send with Invalidate operation may be worth employing in this respect, as it relieves the client of certain overhead in this case. A "onetime" boolean advisory to each RDMA region might become a hint to the server that the client will use the three-tuple for only one NFSv4 operation. For a transport such as iWARP, the server can assist the client in invalidating the three-tuple by performing a Send with Solicited Event and Invalidate. The server may ignore this hint, in which case the client must perform a local invalidate after receiving the indication from the server that the NFSv4 operation is complete. This may be considered in a future version of this draft and [NFSDDP]. In a trusted environment, it may be desirable for the client to persistently enable RDMA access by the server. Such a model is desirable for the highest level of efficiency and lowest overhead. Shepler Expires June 15, 2006 [Page 25] Internet-Draft NFSv4 Minior Version 1 December 2005 RDMA message exchanges Client Server : Direct Read Request : Send : ------------------------------> : untagged : : buffer : Segment : tagged : <------------------------------ : RDMA Write buffer : : : : [Segment] : tagged : <------------------------------ : [RDMA Write] buffer : : : Direct Read Response : untagged : <------------------------------ : Send (w/Inv.) buffer : : Client Server : Direct Write Request : Send : ------------------------------> : untagged : : buffer : Segment : tagged : v------------------------------ : RDMA Read buffer : +-----------------------------> : : : : : [Segment] : tagged : v------------------------------ : [RDMA Read] buffer : +-----------------------------> : : : : Direct Write Response : untagged : <------------------------------ : Send (w/Inv.) buffer : : 3.4 Connection Models There are three scenarios in which to discuss the connection model. Each will be discussed individually, after describing the common case encountered at initial connection establishment. After a successful connection, the first request proceeds, in the case of a new client association, to initial session creation, and then optionally to session callback channel binding, prior to regular operation. Shepler Expires June 15, 2006 [Page 26] Internet-Draft NFSv4 Minior Version 1 December 2005 Commonly, each new client "mount" will be the action which drives creation of a new session. However there are any number of other approaches. Clients may choose to share a single connection and session among all their mount points. Or, clients may support trunking, where additional connections are created but all within a single session. Alternatively, the client may choose to create multiple sessions, each tuned to the buffering and reliability needs of the mount point. For example, a readonly mount can sharply reduce its write buffering and also makes no requirement for the server to support reliable duplicate request caching. Similarly, the client can choose among several strategies for clientid usage. Sessions can share a single clientid, or create new clientids as the client deems appropriate. For kernel-based clients which service multiple authenticated users, a single clientid shared across all mount points is generally the most appropriate and flexible approach. For example, all the client's file operations may wish to share locking state and the local client kernel takes the responsibility for arbitrating access locally. For clients choosing to support other authentication models, perhaps example userspace implementations, a new clientid is indicated. Through use of session create options, both models are supported at the client's choice. Since the session is explicitly created and destroyed by the client, and each client is uniquely identified, the server may be specifically instructed to discard unneeded presistent state. For this reason, it is possible that a server will retain any previous state indefinitely, and place its destruction under administrative control. Or, a server may choose to retain state for some configurable period, provided that the period meets other NFSv4 requirements such as lease reclamation time, etc. However, since discarding this state at the server may affect the correctness of the server as seen by the client across network partitioning, such discarding of state should be done only in a conservative manner. Each client request to the server carries a new SEQUENCE operation within each COMPOUND, which provides the session context. This session context then governs the request control, duplicate request caching, and other persistent parameters managed by the server for a session. 3.4.1 TCP Connection Model The following is a schematic diagram of the NFSv4.1 protocol exchanges leading up to normal operation on a TCP stream. Shepler Expires June 15, 2006 [Page 27] Internet-Draft NFSv4 Minior Version 1 December 2005 Client Server TCPmode : Create Clientid(nfs_client_id4) : TCPmode : ------------------------------> : : : : Clientid reply(clientid, ...) : : <------------------------------ : : : : Create Session(clientid, size S, : : maxreq N, STREAM, ...) : : ------------------------------> : : : : Session reply(sessionid, size S', : : maxreq N') : : <------------------------------ : : : : : : ------------------------------> : : <------------------------------ : : : : No net additional exchange is added to the initial negotiation by this proposal. In the NFSv4.1 exchange, the CREATECLIENTID replaces SETCLIENTID (eliding the callback "clientaddr4" addressing) and CREATESESSION subsumes the function of SETCLIENTID_CONFIRM, as described elsewhere in this document. Callback channel binding is optional, as in NFSv4.0. Note that the STREAM transport type is shown above, but since the transport mode remains unchanged and transport attributes are not necessarily exchanged, DEFAULT could also be passed. 3.4.2 Negotiated RDMA Connection Model One possible design which has been considered is to have a "negotiated" RDMA connection model, supported via use of a session bind operation as a required first step. However due to issues mentioned earlier, this proved problematic. This section remains as a reminder of that fact, and it is possible such a mode can be supported. It is not considered critical that this be supported for two reasons. One, the session persistence provides a way for the server to remember important session parameters, such as sizes and maximum request counts. These values can be used to restore the endpoint prior to making the first reply. Two, there are currently no critical RDMA parameters to set in the endpoint at the server side of the connection. RDMA Read resources, which are in general not settable after entering RDMA mode, are set only at the client - the originator of the connection. Therefore as long as the RDMA provider Shepler Expires June 15, 2006 [Page 28] Internet-Draft NFSv4 Minior Version 1 December 2005 supports an automatic RDMA connection mode, no further support is required from the NFSv4.1 protocol for reconnection. Note, the client must provide at least as many RDMA Read resources to its local queue for the benefit of the server when reconnecting, as it used when negotiating the session. If this value is no longer appropriate, the client should resynchronize its session state, destroy the existing session, and start over with the more appropriate values. 3.4.3 Automatic RDMA Connection Model The following is a schematic diagram of the NFSv4.1 protocol exchanges performed on an RDMA connection. Client Server RDMAmode : : : RDMAmode : : : Prepost : : : Prepost receive : : : receive : : : Create Clientid(nfs_client_id4) : : ------------------------------> : : : Prepost : Clientid reply(clientid, ...) : receive : <------------------------------ : Prepost : : receive : Create Session(clientid, size S, : : maxreq N, RDMA ...) : : ------------------------------> : : : Prepost <=N' : Session reply(sessionid, size S', : receives of : maxreq N') : size S' : <------------------------------ : : : : : : ------------------------------> : : <------------------------------ : : : : 3.5 Buffer Management, Transfer, Flow Control Inline operations in NFSv4.1 behave effectively the same as TCP sends. Procedure results are passed in a single message, and its completion at the client signal the receiving process to inspect the message. Shepler Expires June 15, 2006 [Page 29] Internet-Draft NFSv4 Minior Version 1 December 2005 RDMA operations are performed solely by the server in this proposal, as described in the previous "RDMA Direct Model" section. Since server RDMA operations do not result in a completion at the client, and due to ordering rules in RDMA transports, after all required RDMA operations are complete, a Send (Send with Solicited Event for iWARP) containing the procedure results is performed from server to client. This Send operation will result in a completion which will signal the client to inspect the message. In the case of client read-type NFSv4 operations, the server will have issued RDMA Writes to transfer the resulting data into client- advertised buffers. The subsequent Send operation performs two necessary functions: finalizing any active or pending DMA at the client, and signaling the client to inspect the message. In the case of client write-type NFSv4 operations, the server will have issued RDMA Reads to fetch the data from the client-advertised buffers. No data consistency issues arise at the client, but the completion of the transfer must be acknowledged, again by a Send from server to client. In either case, the client advertises buffers for direct (RDMA style) operations. The client may desire certain advertisement limits, and may wish the server to perform remote invalidation on its behalf when the server has completed its RDMA. This may be considered in a future version of this draft. In the absence of remote invalidation, the client may perform its own, local invalidation after the operation completes. This invalidation should occur prior to any RPCSEC GSS integrity checking, since a validly remotely accessible buffer can possibly be modified by the peer. However, after invalidation and the contents integrity checked, the contents are locally secure. Credit updates over RDMA transports are supported at the RPC layer as described in [RPCRDMA]. In each request, the client requests a desired number of credits to be made available to the connection on which it sends the request. The client must not send more requests than the number which the server has previously advertised, or in the case of the first request, only one. If the client exceeds its credit limit, the connection may close with a fatal RDMA error. The server then executes the request, and replies with an updated credit count accompanying its results. Since replies are sequenced by their RDMA Send order, the most recent results always reflect the server's limit. In this way the client will always know the maximum number of requests it may safely post. Shepler Expires June 15, 2006 [Page 30] Internet-Draft NFSv4 Minior Version 1 December 2005 Because the client requests an arbitrary credit count in each request, it is relatively easy for the client to request more, or fewer, credits to match its expected need. A client that discovered itself frequently queuing outgoing requests due to lack of server credits might increase its requested credits proportionately in response. Or, a client might have a simple, configurable number. The protocol also provides a per-operation "maxslot" exchange to assist in dynamic adjustment at the session level, described in a later section. Occasionally, a server may wish to reduce the total number of credits it offers a certain client on a connection. This could be encountered if a client were found to be consuming its credits slowly, or not at all. A client might notice this itself, and reduce its requested credits in advance, for instance requesting only the count of operations it currently has queued, plus a few as a base for starting up again. Such mechanisms can, however, be potentially complicated and are implementation-defined. The protocol does not require them. Because of the way in which RDMA fabrics function, it is not possible for the server (or client back channel) to cancel outstanding receive operations. Therefore, effectively only one credit can be withdrawn per receive completion. The server (or client back channel) would simply not replenish a receive operation when replying. The server can still reduce the available credit advertisement in its replies to the target value it desires, as a hint to the client that its credit target is lower and it should expect it to be reduced accordingly. Of course, even if the server could cancel outstanding receives, it cannot do so, since the client may have already sent requests in expectation of the previous limit. This brings out an interesting scenario similar to the client reconnect discussed earlier in "Connection Models". How does the server reduce the credits of an inactive client? One approach is for the server to simply close such a connection and require the client to reconnect at a new credit limit. This is acceptable, if inefficient, when the connection setup time is short and where the server supports persistent session semantics. A better approach is to provide a back channel request to return the operations channel credits. The server may request the client to return some number of credits, the client must comply by performing operations on the operations channel, provided of course that the request does not drop the client's credit count to zero (in which case the connection would deadlock). If the client finds that it has no requests with which to consume the credits it was previously Shepler Expires June 15, 2006 [Page 31] Internet-Draft NFSv4 Minior Version 1 December 2005 granted, it must send zero-length Send RDMA operations, or NULL NFSv4 operations in order to return the resources to the server. If the client fails to comply in a timely fashion, the server can recover the resources by breaking the connection. While in principle, the back channel credits could be subject to a similar resource adjustment, in practice this is not an issue, since the back channel is used purely for control and is expected to be statically provisioned. It is important to note that in addition to maximum request counts, the sizes of buffers are negotiated per-session. This permits the most efficient allocation of resources on both peers. There is an important requirement on reconnection: the sizes posted by the server at reconnect must be at least as large as previously used, to allow recovery. Any replies that are replayed from the server's duplicate request cache must be able to be received into client buffers. In the case where a client has received replies to all its retried requests (and therefore received all its expected responses), then the client may disconnect and reconnect with different buffers at will, since no cache replay will be required. 3.6 Retry and Replay NFSv4.0 forbids retransmission on active connections over reliable transports; this includes connected-mode RDMA. This restriction must be maintained in NFSv4.1. If one peer were to retransmit a request (or reply), it would consume an additional credit on the other. If the server retransmitted a reply, it would certainly result in an RDMA connection loss, since the client would typically only post a single receive buffer for each request. If the client retransmitted a request, the additional credit consumed on the server might lead to RDMA connection failure unless the client accounted for it and decreased its available credit, leading to wasted resources. RDMA credits present a new issue to the duplicate request cache in NFSv4.1. The request cache may be used when a connection within a session is lost, such as after the client reconnects. Credit information is a dynamic property of the connection, and stale values must not be replayed from the cache. This implies that the request cache contents must not be blindly used when replies are issued from it, and credit information appropriate to the channel must be refreshed by the RPC layer. Finally, RDMA fabrics do not guarantee that the memory handles (Steering Tags) within each rdma three-tuple are valid on a scope Shepler Expires June 15, 2006 [Page 32] Internet-Draft NFSv4 Minior Version 1 December 2005 outside that of a single connection. Therefore, handles used by the direct operations become invalid after connection loss. The server must ensure that any RDMA operations which must be replayed from the request cache use the newly provided handle(s) from the most recent request. 3.7 The Back Channel The NFSv4 callback operations present a significant resource problem for the RDMA enabled client. Clearly, callbacks must be negotiated in the way credits are for the ordinary operations channel for requests flowing from client to server. But, for callbacks to arrive on the same RDMA endpoint as operation replies would require dedicating additional resources, and specialized demultiplexing and event handling. Or, callbacks may not require RDMA sevice at all (they do not normally carry substantial data payloads). It is highly desirable to streamline this critical path via a second communications channel. The session callback channel binding facility is designed for exactly such a situation, by dynamically associating a new connected endpoint with the session, and separately negotiating sizes and counts for active callback channel operations. The binding operation is firewall-friendly since it does not require the server to initiate the connection. This same method serves as well for ordinary TCP connection mode. It is expected that all NFSv4.1 clients may make use of the session facility to streamline their design. The back channel functions exactly the same as the operations channel except that no RDMA operations are required to perform transfers, instead the sizes are required to be sufficiently large to carry all data inline, and of course the client and server reverse their roles with respect to which is in control of credit management. The same rules apply for all transfers, with the server being required to flow control its callback requests. The back channel is optional. If not bound on a given session, the server must not issue callback operations to the client. This in turn implies that such a client must never put itself in the situation where the server will need to do so, lest the client lose its connection by force, or its operation be incorrect. For the same reason, if a back channel is bound, the client is subject to revocation of its delegations if the back channel is lost. Any connection loss should be corrected by the client as soon as possible. Shepler Expires June 15, 2006 [Page 33] Internet-Draft NFSv4 Minior Version 1 December 2005 This can be convenient for the NFSv4.1 client; if the client expects to make no use of back channel facilities such as delegations, then there is no need to create it. This may save significant resources and complexity at the client. For these reasons, if the client wishes to use the back channel, that channel must be bound first, before using the operations channel. In this way, the server will not find itself in a position where it will send callbacks on the operations channel when the client is not prepared for them. There is one special case, that where the back channel is bound in fact to the operations channel's connection. This configuration would be used normally over a TCP stream connection to exactly implement the NFSv4.0 behavior, but over RDMA would require complex resource and event management at both sides of the connection. The server is not required to accept such a bind request on an RDMA connection for this reason, though it is recommended. 3.8 COMPOUND Sizing Issues Very large responses may pose duplicate request cache issues. Since servers will want to bound the storage required for such a cache, the unlimited size of response data in COMPOUND may be troublesome. If COMPOUND is used in all its generality, then the inclusion of certain non-idempotent operations within a single COMPOUND request may render the entire request non-idempotent. (For example, a single COMPOUND request which read a file or symbolic link, then removed it, would be obliged to cache the data in order to allow identical replay). Therefore, many requests might include operations that return any amount of data. It is not satisfactory for the server to reject COMPOUNDs at will with NFS4ERR_RESOURCE when they pose such difficulties for the server, as this results in serious interoperability problems. Instead, any such limits must be explicitly exposed as attributes of the session, ensuring that the server can explicitly support any duplicate request cache needs at all times. 3.9 Data Alignment A negotiated data alignment enables certain scatter/gather optimizations. A facility for this is supported by [RPCRDMA]. Where NFS file data is the payload, specific optimizations become highly attractive. Header padding is requested by each peer at session initiation, and may be zero (no padding). Padding leverages the useful property that Shepler Expires June 15, 2006 [Page 34] Internet-Draft NFSv4 Minior Version 1 December 2005 RDMA receives preserve alignment of data, even when they are placed into anonymous (untagged) buffers. If requested, client inline writes will insert appropriate pad bytes within the request header to align the data payload on the specified boundary. The client is encouraged to be optimistic and simply pad all WRITEs within the RPC layer to the negotiated size, in the expectation that the server can use them efficiently. It is highly recommended that clients offer to pad headers to an appropriate size. Most servers can make good use of such padding, which allows them to chain receive buffers in such a way that any data carried by client requests will be placed into appropriate buffers at the server, ready for filesystem processing. The receiver's RPC layer encounters no overhead from skipping over pad bytes, and the RDMA layer's high performance makes the insertion and transmission of padding on the sender a significant optimization. In this way, the need for servers to perform RDMA Read to satisfy all but the largest client writes is obviated. An added benefit is the reduction of message roundtrips on the network - a potentially good trade, where latency is present. The value to choose for padding is subject to a number of criteria. A primary source of variable-length data in the RPC header is the authentication information, the form of which is client-determined, possibly in response to server specification. The contents of COMPOUNDs, sizes of strings such as those passed to RENAME, etc. all go into the determination of a maximal NFSv4 request size and therefore minimal buffer size. The client must select its offered value carefully, so as not to overburden the server, and vice- versa. The payoff of an appropriate padding value is higher performance. Sender gather: |RPC Request|Pad bytes|Length| -> |User data...| \------+---------------------/ \ \ \ \ Receiver scatter: \-----------+- ... /-----+----------------\ \ \ |RPC Request|Pad|Length| -> |FS buffer|->|FS buffer|->... In the above case, the server may recycle unused buffers to the next posted receive if unused by the actual received request, or may pass the now-complete buffers by reference for normal write processing. For a server which can make use of it, this removes any need for data copies of incoming data, without resorting to complicated end-to-end buffer advertisement and management. This includes most kernel-based and integrated server designs, among many others. The client may perform similar optimizations, if desired. Shepler Expires June 15, 2006 [Page 35] Internet-Draft NFSv4 Minior Version 1 December 2005 Padding is negotiated by the session creation operation, and subsequently used by the RPC RDMA layer, as described in [RPCRDMA]. 3.10 NFSv4 Integration The following section discusses the integration of the proposed RDMA extensions with NFSv4.0. 3.10.1 Minor Versioning Minor versioning is the existing facility to extend the NFSv4 protocol, and this proposal takes that approach. Minor versioning of NFSv4 is relatively restrictive, and allows for tightly limited changes only. In particular, it does not permit adding new "procedures" (it permits adding only new "operations"). Interoperability concerns make it impossible to consider additional layering to be a minor revision. This somewhat limits the changes that can be proposed when considering extensions. To support the duplicate request cache integrated with sessions and request control, it is desirable to tag each request with an identifier to be called a Slotid. This identifier must be passed by NFSv4 when running atop any transport, including traditional TCP. Therefore it is not desirable to add the Slotid to a new RPC transport, even though such a transport is indicated for support of RDMA. This draft and [RPCRDMA] do not propose such an approach. Instead, this proposal conforms to the requirements of NFSv4 minor versioning, through the use of a new operation within NFSv4 COMPOUND procedures as detailed below. If sessions are in use for a given clientid, this same clientid cannot be used for non-session NFSv4 operation, including NFSv4.0. Because the server will have allocated session-specific state to the active clientid, it would be an unnecessary burden on the server implementor to support and account for additional, non- session traffic, in addition to being of no benefit. Therefore this proposal prohibits a single clientid from doing this. Nevertheless, employing a new clientid for such traffic is supported. 3.10.2 Slot Identifiers and Server Duplicate Request Cache The presence of deterministic maximum request limits on a session enables in-progress requests to be assigned unique values with useful properties. The RPC layer provides a transaction ID (xid), which, while required Shepler Expires June 15, 2006 [Page 36] Internet-Draft NFSv4 Minior Version 1 December 2005 to be unique, is not especially convenient for tracking requests. The transaction ID is only meaningful to the issuer (client), it cannot be interpreted at the server except to test for equality with previously issued requests. Because RPC operations may be completed by the server in any order, many transaction IDs may be outstanding at any time. The client may therefore perform a computationally expensive lookup operation in the process of demultiplexing each reply. In the proposal, there is a limit to the number of active requests. This immediately enables a convenient, computationally efficient index for each request which is designated as a Slot Identifier, or slotid. When the client issues a new request, it selects a slotid in the range 0..N-1, where N is the server's current "totalrequests" limit granted the client on the session over which the request is to be issued. The slotid must be unused by any of the requests which the client has already active on the session. "Unused" here means the client has no outstanding request for that slotid. Because the slot id is always an integer in the range 0..N-1, client implementations can use the slotid from a server response to efficiently match responses with outstanding requests, such as, for example, by using the slotid to index into a outstanding request array. This can be used to avoid expensive hashing and lookup functions in the performace-critical receive path. The sequenceid, which accompanies the slotid in each request, is important for a second, important check at the server: it must be able to be determined efficiently whether a request using a certain slotid is a retransmit or a new, never-before-seen request. It is not feasible for the client to assert that it is retransmitting to implement this, because for any given request the client cannot know the server has seen it unless the server actually replies. Of course, if the client has seen the server's reply, the client would not retransmit! The sequenceid must increase monotonically for each new transmit of a given slotid, and must remain unchanged for any retransmission. The server must in turn compare each newly received request's sequenceid with the last one previously received for that slotid, to see if the new request is: A new request, in which the sequenceid is greater than that previously seen in the slot (accounting for sequence wraparound). The server proceeds to execute the new request. Shepler Expires June 15, 2006 [Page 37] Internet-Draft NFSv4 Minior Version 1 December 2005 A retransmitted request, in which the sequenceid is equal to that last seen in the slot. Note that this request may be either complete, or in progress. The server performs replay processing in these cases. A misordered duplicate, in which the sequenceid is less than that previously seen in the slot. The server must drop the incoming request, which may imply dropping the connection if the transport is reliable, as dictated by section 3.1.1 of [RFC3530]. This last condition is possible on any connection, not just unreliable, unordered transports. Delayed behavior on abandoned TCP connections which are not yet closed at the server, or pathological client implementations can cause it, among other causes. Therefore, the server may wish to harden itself against certain repeated occurrences of this, as it would for retransmissions in [RFC3530]. It is recommended, though not necessary for protocol correctness, that the client simply increment the sequenceid by one for each new request on each slotid. This reduces the wraparound window to a minimum, and is useful for tracing and avoidance of possible implementation errors. The client may however, for implementation-specific reasons, choose a different algorithm. For example it might maintain a single sequence space for all slots in the session - e.g. employing the RPC XID itself. The sequenceid, in any case, is never interpreted by the server for anything but to test by comparison with previously seen values. The server may thereby use the slotid, in conjunction with the sessionid and sequenceid, within the SEQUENCE portion of the request to maintain its duplicate request cache (DRC) for the session, as opposed to the traditional approach of ONC RPC applications that use the XID along with certain transport information [RW96]. Unlike the XID, the slotid is always within a specific range; this has two implications. The first implication is that for a given session, the server need only cache the results of a limited number of COMPOUND requests. The second implication derives from the first, which is unlike XID-indexed DRCs, the slotid DRC by its nature cannot be overflowed. Through use of the sequenceid to identify retransmitted requests, it is notable that the server does not need to actually cache the request itself, reducing the storage requirements of the DRC further. These new facilities makes it practical to maintain all the required entries for an effective DRC. The slotid and sequenceid therefore take over the traditional role of Shepler Expires June 15, 2006 [Page 38] Internet-Draft NFSv4 Minior Version 1 December 2005 the port number in the server DRC implementation, and the session replaces the IP address. This approach is considerably more portable and completely robust - it is not subject to the frequent reassignment of ports as clients reconnect over IP networks. In addition, the RPC XID is not used in the reply cache, enhancing robustness of the cache in the face of any rapid reuse of XIDs by the client. It is required to encode the slotid information into each request in a way that does not violate the minor versioning rules of the NFSv4.0 specification. This is accomplished here by encoding it in a control operation within each NFSv4.1 COMPOUND and CB_COMPOUND procedure. The operation easily piggybacks within existing messages. The implementation section of this document describes the specific proposal. In general, the receipt of a new sequenced request arriving on any valid slot is an indication that the previous DRC contents of that slot may be discarded. In order to further assist the server in slot management, the client is required to use the lowest available slot when issuing a new request. In this way, the server may be able to retire additional entries. However, in the case where the server is actively adjusting its granted maximum request count to the client, it may not be able to use receipt of the slotid to retire cache entries. The slotid used in an incoming request may not reflect the server's current idea of the client's session limit, because the request may have been sent from the client before the update was received. Therefore, in the downward adjustment case, the server may have to retain a number of duplicate request cache entries at least as large as the old value, until operation sequencing rules allow it to infer that the client has seen its reply. The SEQUENCE (and CB_SEQUENCE) operation also carries a "maxslot" value which carries additional client slot usage information. The client must always provide its highest-numbered outstanding slot value in the maxslot argument, and the server may reply with a new recognized value. The client should in all cases provide the most conservative value possible, although it can be increased somewhat above the actual instantaneous usage to maintain some minimum or optimal level. This provides a way for the client to yield unused request slots back to the server, which in turn can use the information to reallocate resources. Obviously, maxslot can never be zero, or the session would deadlock. The server also provides a target maxslot value to the client, which is an indication to the client of the maxslot the server wishes the Shepler Expires June 15, 2006 [Page 39] Internet-Draft NFSv4 Minior Version 1 December 2005 client to be using. This permits the server to withdraw (or add) resources from a client that has been found to not be using them, in order to more fairly share resources among a varying level of demand from other clients. The client must always comply with the server's value updates, since they indicate newly established hard limits on the client's access to session resources. However, because of request pipelining, the client may have active requests in flight reflecting prior values, therefore the server must not immediately require the client to comply. It is worthwhile to note that Sprite RPC [BW87] defined a "channel" which in some ways is similar to the slotid proposed here. Sprite RPC used channels to implement parallel request processing and request/response cache retirement. 3.10.3 COMPOUND and CB_COMPOUND Support for per-operation control can be piggybacked onto NFSv4 COMPOUNDs with full transparency, by placing such facilities into their own, new operation, and placing this operation first in each COMPOUND under the new NFSv4 minor protocol revision. The contents of the operation would then apply to the entire COMPOUND. Recall that the NFSv4 minor revision is contained within the COMPOUND header, encoded prior to the COMPOUNDed operations. By simply requiring that the new operation always be contained in NFSv4 minor COMPOUNDs, the control protocol can piggyback perfectly with each request and response. In this way, the NFSv4 RDMA Extensions may stay in compliance with the minor versioning requirements specified in section 10 of [RFC3530]. Referring to section 13.1 of the same document, the proposed session- enabled COMPOUND and CB_COMPOUND have the form: +-----+--------------+-----------+------------+-----------+---- | tag | minorversion | numops | control op | op + args | ... | | (== 1) | (limited) | + args | | +-----+--------------+-----------+------------+-----------+---- and the reply's structure is: +------------+-----+--------+-------------------------------+--// |last status | tag | numres | status + control op + results | // +------------+-----+--------+-------------------------------+--// //-----------------------+---- // status + op + results | ... Shepler Expires June 15, 2006 [Page 40] Internet-Draft NFSv4 Minior Version 1 December 2005 //-----------------------+---- The single control operation within each NFSv4.1 COMPOUND defines the context and operational session parameters which govern that COMPOUND request and reply. Placing it first in the COMPOUND encoding is required in order to allow its processing before other operations in the COMPOUND. 3.10.4 eXternal Data Representation Efficiency RDMA is a copy avoidance technology, and it is important to maintain this efficiency when decoding received messages. Traditional XDR implementations frequently use generated unmarshaling code to convert objects to local form, incurring a data copy in the process (in addition to subjecting the caller to recursive calls, etc). Often, such conversions are carried out even when no size or byte order conversion is necessary. It is recommended that implementations pay close attention to the details of memory referencing in such code. It is far more efficient to inspect data in place, using native facilities to deal with word size and byte order conversion into registers or local variables, rather than formally (and blindly) performing the operation via fetch, reallocate and store. Of particular concern is the result of the READDIR operation, in which such encoding abounds. 3.10.5 Effect of Sessions on Existing Operations The use of a session replaces the use of the SETCLIENTID and SETCLIENTID_CONFIRM operations, and allows certain simplification of the RENEW and callback addressing mechanisms in the base protocol. The cb_program and cb_location which are obtained by the server in SETCLIENTID_CONFIRM must not be used by the server, because the NFSv4.1 client performs callback channel designation with BIND_BACKCHANNEL. Therefore the SETCLIENTID and SETCLIENTID_CONFIRM operations becomes obsolete when sessions are in use, and a server should return an error to NFSv4.1 clients which might issue either operation. Another favorable result of the session is that the server is able to avoid requiring the client to perform OPEN_CONFIRM operations. The existence of a reliable and effective DRC means that the server will be able to determine whether an OPEN request carrying a previously known open_owner from a client is or is not a retransmission. Because of this, the server no longer requires OPEN_CONFIRM to verify Shepler Expires June 15, 2006 [Page 41] Internet-Draft NFSv4 Minior Version 1 December 2005 whether the client is retransmitting an open request. This in turn eliminates the server's reason for requesting OPEN_CONFIRM - the server can simply replace any previous information on this open_owner. Client OPEN operations are therefore streamlined, reducing overhead and latency through avoiding the additional OPEN_CONFIRM exchange. Since the session carries the client liveness indication with it implicitly, any request on a session associated with a given client will renew that client's leases. Therefore the RENEW operation is made unnecessary when a session is present, as any request (including a SEQUENCE operation with or without additional NFSv4 operations) performs its function. It is possible (though this proposal does not make any recommendation) that the RENEW operation could be made obsolete. An interesting issue arises however if an error occurs on such a SEQUENCE operation. If the SEQUENCE operation fails, perhaps due to an invalid slotid or other non-renewal-based issue, the server may or may not have performed the RENEW. In this case, the state of any renewal is undefined, and the client should make no assumption that it has been performed. In practice, this should not occur but even if it did, it is expected the client would perform some sort of recovery which would result in a new, successful, SEQUENCE operation being run and the client assured that the renewal took place. 3.10.6 Authentication Efficiencies NFSv4 requires the use of the RPCSEC_GSS ONC RPC security flavor [RFC2203] to provide authentication, integrity, and privacy via cryptography. The server dictates to the client the use of RPCSEC_GSS, the service (authentication, integrity, or privacy), and the specific GSS-API security mechanism that each remote procedure call and result will use. If the connection's integrity is protected by an additional means than RPCSEC_GSS, such as via IPsec, then the use of RPCSEC_GSS's integrity service is nearly redundant (See the Security Considerations section for more explanation of why it is "nearly" and not completely redundant). Likewise, if the connection's privacy is protected by additional means, then the use of both RPCSEC_GSS's integrity and privacy services is nearly redundant. Connection protection schemes, such as IPsec, are more likely to be implemented in hardware than upper layer protocols like RPCSEC_GSS. Hardware-based cryptography at the IPsec layer will be more efficient than software-based cryptography at the RPCSEC_GSS layer. Shepler Expires June 15, 2006 [Page 42] Internet-Draft NFSv4 Minior Version 1 December 2005 When transport integrity can be obtained, it is possible for server and client to downgrade their per-operation authentication, after an appropriate exchange. This downgrade can in fact be as complete as to establish security mechanisms that have zero cryptographic overhead, effectively using the underlying integrity and privacy services provided by transport. Based on the above observations, a new GSS-API mechanism, called the Channel Conjunction Mechanism [CCM], is being defined. The CCM works by creating a GSS-API security context using as input a cookie that the initiator and target have previously agreed to be a handle for GSS-API context created previously over another GSS-API mechanism. NFSv4.1 clients and servers should support CCM and they must use as the cookie the handle from a successful RPCSEC_GSS context creation over a non-CCM mechanism (such as Kerberos V5). The value of the cookie will be equal to the handle field of the rpc_gss_init_res structure from the RPCSEC_GSS specification. The [CCM] Draft provides further discussion and examples. 3.11 Sessions Security Considerations The NFSv4 minor version 1 retains all of existing NFSv4 security; all security considerations present in NFSv4.0 apply to it equally. Security considerations of any underlying RDMA transport are additionally important, all the more so due to the emerging nature of such transports. Examining these issues is outside the scope of this draft. When protecting a connection with RPCSEC_GSS, all data in each request and response (whether transferred inline or via RDMA) continues to receive this protection over RDMA fabrics [RPCRDMA]. However when performing data transfers via RDMA, RPCSEC_GSS protection of the data transfer portion works against the efficiency which RDMA is typically employed to achieve. This is because such data is normally managed solely by the RDMA fabric, and intentionally is not touched by software. Therefore when employing RPCSEC_GSS under CCM, and where integrity protection has been "downgraded", the cooperation of the RDMA transport provider is critical to maintain any integrity and privacy otherwise in place for the session. The means by which the local RPCSEC_GSS implementation is integrated with the RDMA data protection facilities are outside the scope of this draft. It is logical to use the same GSS context on a session's callback channel as that used on its operations channel(s), particularly when Shepler Expires June 15, 2006 [Page 43] Internet-Draft NFSv4 Minior Version 1 December 2005 the connection is shared by both. The client must indicate to the server: - what security flavor(s) to use in the call back. A special callback flavor might be defined for this. - if the flavor is RPCSEC_GSS, then the client must have previously created an RPCSEC_GSS session with the server. The client offers to the server the the opaque handle<> value from the rpc_gss_init_res structure, the window size of RPCSEC_GSS sequence numbers, and an opaque gss_cb_handle. This exchange can be performed as part of session and clientid creation, and the issue warrants careful analysis before being specified. If the NFS client wishes to maintain full control over RPCSEC_GSS protection, it may still perform its transfer operations using either the inline or RDMA transfer model, or of course employ traditional TCP stream operation. In the RDMA inline case, header padding is recommended to optimize behavior at the server. At the client, close attention should be paid to the implementation of RPCSEC_GSS processing to minimize memory referencing and especially copying. These are well-advised in any case! The proposed session callback channel binding improves security over that provided by NFSv4 for the callback channel. The connection is client-initiated, and subject to the same firewall and routing checks as the operations channel. The connection cannot be hijacked by an attacker who connects to the client port prior to the intended server. The connection is set up by the client with its desired attributes, such as optionally securing with IPsec or similar. The binding is fully authenticated before being activated. 3.11.1 Authentication Proper authentication of the principal which issues any session and clientid in the proposed NFSv4.1 operations exactly follows the similar requirement on client identifiers in NFSv4.0. It must not be possible for a client to impersonate another by guessing its session identifiers for NFSv4.1 operations, nor to bind a callback channel to an existing session. To protect against this, NFSv4.0 requires appropriate authentication and matching of the principal used. This is discussed in Section 16, Security Considerations of [RFC3530]. The same requirement when using a session identifier applies to NFSv4.1 here. Going beyond NFSv4.0, the presence of a session associated with any Shepler Expires June 15, 2006 [Page 44] Internet-Draft NFSv4 Minior Version 1 December 2005 clientid may also be used to enhance NFSv4.1 security with respect to client impersonation. In NFSv4.0, there are many operations which carry no clientid, including in particular those which employ a stateid argument. A rogue client which wished to carry out a denial of service attack on another client could perform CLOSE, DELEGRETURN, etc operations with that client's current filehandle, sequenceid and stateid, after having obtained them from eavesdropping or other approach. Locking and open downgrade operations could be similarly attacked. When an NFSv4.1 session is in place for any clientid, countermeasures are easily applied through use of authentication by the server. Because the clientid and sessionid must be present in each request within a session, the server may verify that the clientid is in fact originating from a principal with the appropriate authenticated credentials, that the sessionid belongs to the clientid, and that the stateid is valid in these contexts. This is in general not possible with the affected operations in NFSv4.0 due to the fact that the clientid is not present in the requests. In the event that authentication information is not available in the incoming request, for example after a reconnection when the security was previously downgraded using CCM, the server must require the client re-establish the authentication in order that the server may validate the other client-provided context, prior to executing any operation. The sessionid, present in the newly retransmitted request, combined with the retransmission detection enabled by the NFSv4.1 duplicate request cache, are a convenient and reliable context for the server to use for this contingency. The server should take care to protect itself against denial of service attacks in the creation of sessions and clientids. Clients who connect and create sessions, only to disconnect and never use them may leave significant state behind. (The same issue applies to NFSv4.0 with clients who may perform SETCLIENTID, then never perform SETCLIENTID_CONFIRM.) Careful authentication coupled with resource checks is highly recommended. 4. Directory Delegations 4.1 Introduction to Directory Delegations The major addition to NFS version 4 in the area of caching is the ability of the server to delegate certain responsibilities to the client. When the server grants a delegation for a file to a client, the client receives certain semantics with respect to the sharing of that file with other clients. At OPEN, the server may provide the client either a read or write delegation for the file. If the client Shepler Expires June 15, 2006 [Page 45] Internet-Draft NFSv4 Minior Version 1 December 2005 is granted a read delegation, it is assured that no other client has the ability to write to the file for the duration of the delegation. If the client is granted a write delegation, the client is assured that no other client has read or write access to the file. This reduces network traffic and server load by allowing the client to perform certain operations on local file data and can also provide stronger consistency for the local data. Directory caching for the NFS version 4 protocol is similar to previous versions. Clients typically cache directory information for a duration determined by the client. At the end of a predefined timeout, the client will query the server to see if the directory has been updated. By caching attributes, clients reduce the number of GETATTR calls made to the server to validate attributes. Furthermore, frequently accessed files and directories, such as the current working directory, have their attributes cached on the client so that some NFS operations can be performed without having to make an RPC call. By caching name and inode information about most recently looked up entries in DNLC (Directory Name Lookup Cache), clients do not need to send LOOKUP calls to the server every time these files are accessed. This caching approach works reasonably well at reducing network traffic in many environments. However, it does not address environments where there are numerous queries for files that do not exist. In these cases of "misses", the client must make RPC calls to the server in order to provide reasonable application semantics and promptly detect the creation of new directory entries. Examples of high miss activity are compilation in software development environments. The current behavior of NFS limits its potential scalability and wide-area sharing effectiveness in these types of environments. Other distributed stateful filesystem architectures such as AFS and DFS have proven that adding state around directory contents can greatly reduce network traffic in high miss environments. Delegation of directory contents is proposed as an extension for NFSv4. Such an extension would provide similar traffic reduction benefits as with file delegations. By allowing clients to cache directory contents (in a read-only fashion) while being notified of changes, the client can avoid making frequent requests to interrogate the contents of slowly-changing directories, reducing network traffic and improving client performance. These extensions allow improved namespace cache consistency to be achieved through delegations and synchronous recalls alone without asking for notifications. In addition, if time-based consistency is sufficient, asynchronous notifications can provide performance Shepler Expires June 15, 2006 [Page 46] Internet-Draft NFSv4 Minior Version 1 December 2005 benefits for the client, and possibly the server, under some common operating conditions such as slowly-changing and/or very large directories. 4.2 Directory Delegation Design (in brief) A new operation GET_DIR_DELEGATION is used by the client to ask for a directory delegation. The delegation covers directory attributes and all entries in the directory. If either of these change the delegation will be recalled synchronously. The operation causing the recall will have to wait before the recall is complete. Any changes to directory entry attributes will not cause the delegation to be recalled. In addition to asking for delegations, a client can also ask for notifications for certain events. These events include changes to directory attributes and/or its contents. If a client asks for notification for a certain event, the server will notify the client when that event occurs. This will not result in the delegation being recalled for that client. The notifications are asynchronous and provide a way of avoiding recalls in situations where a directory is changing enough that the pure recall model may not be effective while trying to allow the client to get substantial benefit. In the absence of notifications, once the delegation is recalled the client has to refresh its directory cache which might not be very efficient for very large directories. The delegation is read only and the client may not make changes to the directory other than by performing NFSv4 operations that modify the directory or the associated file attributes so that the server has knowledge of these changes. In order to keep the client namespace in sync with the server, the server will notify the client holding the delegation of the changes made as a result. This is to avoid any subsequent GETATTR or READDIR calls to the server. If a client holding the delegation makes any changes to the directory, the delegation will not be recalled. Delegations can be recalled by the server at any time. Normally, the server will recall the delegation when the directory changes in a way that is not covered by the notification, or when the directory changes and notifications have not been requested. Also if the server notices that handing out a delegation for a directory is causing too many notifications to be sent out, it may decide not to hand out a delegation for that directory or recall existing delegations. If another client removes the directory for which a delegation has been granted, the server will recall the delegation. Shepler Expires June 15, 2006 [Page 47] Internet-Draft NFSv4 Minior Version 1 December 2005 Both the notification and recall operations need a callback path to exist between the client and server. If the callback path does not exist, then delegation can not be granted. Note that with the session extensions [talpey] that should not be an issue. In the absense of sessions, the server will have to establish a callback path to the client to send callbacks. 4.3 Recommended Attributes in support of Directory Delegations supp_dir_attr_notice - notification delays on directory attributes supp_child_attr_notice - notification delays on child attributes These attributes allow the client and server to negotiate the frequency of notifications sent due to changes in attributes. These attributes are returned as part of a GETATTR call on the directory. The supp_dir_attr_notice value covers all attribute changes to the directory and the supp_child_attr_notice covers all attribute changes to any child in the directory. These attributes are per directory. The client needs to get these values by doing a GETATTR on the directory for which it wants notifications. However these attributes are only required when the client is interested in getting attribute notifications. For all other types of notifications and delegation requests without notifications, these attributes are not required. When the client calls the GET_DIR_DELEGATION operation and asks for attribute change notifications, it will request a notification delay that is within the server's supported range. If the client violates what supp_attr_file_notice or supp_attr_dir_notice values are, the server should not commit to sending notifications for that change event. A value of zero for these attributes means the server will send the notification as soon as the change occurs. It is not recommended to set this value to zero since that can put a lot of burden on the server. A value of N means that the server will make a best effort guarentee that attribute notification are not delayed by more than that. nfstime4 values that compute to negative values are illegal. 4.4 Delegation Recall The server will recall the directory delegation by sending a callback to the client. It will use the same callback procedure as used for recalling file delegations. The server will recall the delegation when the directory changes in a way that is not covered by the notification. However the server will not recall the delegation if Shepler Expires June 15, 2006 [Page 48] Internet-Draft NFSv4 Minior Version 1 December 2005 attributes of an entry within the directory change. Also if the server notices that handing out a delegation for a directory is causing too many notifications to be sent out, it may decide not to hand out a delegation for that directory. If another client tries to remove the directory for which a delegation has been granted, the server will recall the delegation. The server will recall the delegation by sending a CB_RECALL callback to the client. If the recall is done because of a directory changing event, the request making that change will need to wait while the client returns the delegation. 4.5 Delegation Recovery Crash recovery has two main goals, avoiding the necessity of breaking application guarantees with respect to locked files and delivery of updates cached at the client. Neither of these applies to directories protected by read delegations and notifications. Thus, the client is required to establish a new delegation on a server or client reboot. 5. Introduction The NFSv4 protocol [2] specifies the interaction between a client that accesses files and a server that provides access to files and is responsible for coordinating access by multiple clients. As described in the pNFS problem statement, this requires that all access to a set of files exported by a single NFSv4 server be performed by that server; at high data rates the server may become a bottleneck. The parallel NFS (pNFS) extensions to NFSv4 allow data accesses to bypass this bottleneck by permitting direct client access to the storage devices containing the file data. When file data for a single NFSv4 server is stored on multiple and/or higher throughput storage devices (by comparison to the server's throughput capability), the result can be significantly better file access performance. The relationship among multiple clients, a single server, and multiple storage devices for pNFS (server and clients have access to all storage devices) is shown in this diagram: Shepler Expires June 15, 2006 [Page 49] Internet-Draft NFSv4 Minior Version 1 December 2005 +-----------+ |+-----------+ +-----------+ ||+-----------+ | | ||| | NFSv4 + pNFS | | +|| Clients |<------------------------------>| Server | +| | | | +-----------+ | | ||| +-----------+ ||| | ||| | ||| Storage +-----------+ | ||| Protocol |+-----------+ | ||+----------------||+-----------+ Control| |+-----------------||| | Protocol| +------------------+|| Storage |------------+ +| Devices | +-----------+ Figure 9 In this structure, the responsibility for coordination of file access by multiple clients is shared among the server, clients, and storage devices. This is in contrast to NFSv4 without pNFS extensions, in which this is primarily the server's responsibility, some of which can be delegated to clients under strictly specified conditions. The pNFS extension to NFSv4 takes the form of new operations that manage data location information called a "layout". The layout is managed in a similar fashion as NFSv4 data delegations (e.g., they are recallable and revocable). However, they are distinct abstractions and are manipulated with new operations. When a client holds a layout, it has rights to access the data directly using the location information in the layout. There are new attributes that describe general layout characteristics. However, much of the required information cannot be managed solely within the attribute framework, because it will need to have a strictly limited term of validity, subject to invalidation by the server. This requires the use of new operations to obtain, return, recall, and modify layouts, in addition to new attributes. This document specifies both the NFSv4 extensions required to distribute file access coordination between the server and its clients and a NFSv4 file storage protocol that may be used to access data stored on NFSv4 storage devices. Storage protocols used to access a variety of other storage devices are deliberately not specified here. These might include: Shepler Expires June 15, 2006 [Page 50] Internet-Draft NFSv4 Minior Version 1 December 2005 o Block/volume protocols such as iSCSI ([3]), and FCP ([4]). The block/volume protocol support can be independent of the addressing structure of the block/volume protocol used, allowing more than one protocol to access the same file data and enabling extensibility to other block/volume protocols. o Object protocols such as OSD over iSCSI or Fibre Channel [5]. o Other storage protocols, including PVFS and other file systems that are in use in HPC environments. pNFS is designed to accommodate these protocols and be extensible to new classes of storage protocols that may be of interest. The distribution of file access coordination between the server and its clients increases the level of responsibility placed on clients. Clients are already responsible for ensuring that suitable access checks are made to cached data and that attributes are suitably propagated to the server. Generally, a misbehaving client that hosts only a single-user can only impact files accessible to that single user. Misbehavior by a client hosting multiple users may impact files accessible to all of its users. NFSv4 delegations increase the level of client responsibility as a client that carries out actions requiring a delegation without obtaining that delegation will cause its user(s) to see unexpected and/or incorrect behavior. Some uses of pNFS extend the responsibility of clients beyond delegations. In some configurations, the storage devices cannot perform fine-grained access checks to ensure that clients are only performing accesses within the bounds permitted to them by the pNFS operations with the server (e.g., the checks may only be possible at file system granularity rather than file granularity). In situations where this added responsibility placed on clients creates unacceptable security risks, pNFS configurations in which storage devices cannot perform fine-grained access checks SHOULD NOT be used. All pNFS server implementations MUST support NFSv4 access to any file accessible via pNFS in order to provide an interoperable means of file access in such situations. See Section 8 on Security for further discussion. Finally, there are issues about how layouts interact with the existing NFSv4 abstractions of data delegations and byte range locking. These issues, and others, are also discussed here. 6. General Definitions This protocol extension partitions the NFSv4 file system protocol into two parts, the control path and the data path. The control path Shepler Expires June 15, 2006 [Page 51] Internet-Draft NFSv4 Minior Version 1 December 2005 is implemented by the extended (p)NFSv4 server. When the file system being exported by (p)NFSv4 uses storage devices that are visible to clients over the network, the data path may be implemented by direct communication between the extended (p)NFSv4 file system client and the storage devices. This leads to a few new terms used to describe the protocol extension and some clarifications of existing terms. 6.1 Metadata Server A pNFS "server" or "metadata server" is a server as defined by RFC3530 [2], which additionally provides support of the pNFS minor extension. When using the pNFS NFSv4 minor extension, the metadata server may hold only the metadata associated with a file, while the data can be stored on the storage devices. However, similar to NFSv4, data may also be written through the metadata server. Note: directory data is always accessed through the metadata server. 6.2 Client A pNFS "client" is a client as defined by RFC3530 [2], with the addition of supporting the pNFS minor extension server protocol and with the addition of supporting at least one storage protocol for performing I/O directly to storage devices. 6.3 Storage Device This is a device, or server, that controls the file's data, but leaves other metadata management up to the metadata server. A storage device could be another NFS server, or an Object Storage Device (OSD) or a block device accessed over a SAN (e.g., either FiberChannel or iSCSI SAN). The goal of this extension is to allow direct communication between clients and storage devices. 6.4 Storage Protocol This is the protocol between the pNFS client and the storage device used to access the file data. Three following types have been described: file protocols (e.g., NFSv4), object protocols (e.g., OSD), and block/volume protocols (e.g., based on SCSI-block commands). These protocols are in turn realizable over a variety of transport stacks. We anticipate there will be variations on these storage protocols, including new protocols that are unknown at this time or experimental in nature. The details of the storage protocols will be described in other documents so that pNFS clients can be written to use these storage protocols. Use of NFSv4 itself as a file-based storage protocol is described in Section 9. Shepler Expires June 15, 2006 [Page 52] Internet-Draft NFSv4 Minior Version 1 December 2005 6.5 Control Protocol This is a protocol used by the exported file system between the server and storage devices. Specification of such protocols is outside the scope of this draft. Such control protocols would be used to control such activities as the allocation and deallocation of storage and the management of state required by the storage devices to perform client access control. The control protocol should not be confused with protocols used to manage LUNs in a SAN and other sysadmin kinds of tasks. While the pNFS protocol allows for any control protocol, in practice the control protocol is closely related to the storage protocol. For example, if the storage devices are NFS servers, then the protocol between the pNFS metadata server and the storage devices is likely to involve NFS operations. Similarly, when object storage devices are used, the pNFS metadata server will likely use iSCSI/OSD commands to manipulate storage. However, this document does not mandate any particular control protocol. Instead, it just describes the requirements on the control protocol for maintaining attributes like modify time, the change attribute, and the end-of-file position. 6.6 Metadata This is information about a file, like its name, owner, where it stored, and so forth. The information is managed by the exported file system server (metadata server). Metadata also includes lower- level information like block addresses and indirect block pointers. Depending the storage protocol, block-level metadata may or may not be managed by the metadata server, but is instead managed by Object Storage Devices or other servers acting as a storage device. 6.7 Layout A layout defines how a file's data is organized on one or more storage devices. There are many possible layout types. They vary in the storage protocol used to access the data, and in the aggregation scheme that lays out the file data on the underlying storage devices. Layouts are described in more detail below. 7. pNFS protocol semantics This section describes the semantics of the pNFS protocol extension to NFSv4; this is the protocol between the client and the metadata server. Shepler Expires June 15, 2006 [Page 53] Internet-Draft NFSv4 Minior Version 1 December 2005 7.1 Definitions This sub-section defines a number of terms necessary for describing layouts and their semantics. In addition, it more precisely defines how layouts are identified and how they can be composed of smaller granularity layout segments. 7.1.1 Layout Types A layout describes the mapping of a file's data to the storage devices that hold the data. A layout is said to belong to a specific "layout type" (see Section 10.1 for its RPC definition). The layout type allows for variants to handle different storage protocols (e.g., block/volume [6], object [7], and file [Section 9] layout types). A metadata server, along with its control protocol, must support at least one layout type. A private sub-range of the layout type name space is also defined. Values from the private layout type range can be used for internal testing or experimentation. As an example, a file layout type could be an array of tuples (e.g., deviceID, file_handle), along with a definition of how the data is stored across the devices (e.g., striping). A block/volume layout might be an array of tuples that store along with information about block size and the file offset of the first block. An object layout might be an array of tuples and an additional structure (i.e., the aggregation map) that defines how the logical byte sequence of the file data is serialized into the different objects. Note, the actual layouts are more complex than these simple expository examples. This document defines a NFSv4 file layout type using a stripe-based aggregation scheme (see Section 9). Adjunct specifications are being drafted that precisely define other layout formats (e.g., block/ volume [6], and object [7] layouts) to allow interoperability among clients and metadata servers. 7.1.2 Layout Iomode The iomode indicates to the metadata server the client's intent to perform either READs (only) or a mixture of I/O possibly containing WRITEs as well as READs (i.e., READ/WRITE). For certain layout types, it is useful for a client to specify this intent at LAYOUTGET time. E.g., for block/volume based protocols, block allocation could occur when a READ/WRITE iomode is specified. A special LAYOUTIOMODE_ANY iomode is defined and can only be used for LAYOUTRETURN and LAYOUTRECALL, not for LAYOUTGET. It specifies that layouts pertaining to both READ and RW iomodes are being returned or recalled, respectively. Shepler Expires June 15, 2006 [Page 54] Internet-Draft NFSv4 Minior Version 1 December 2005 A storage device may validate I/O with regards to the iomode; this is dependent upon storage device implementation. Thus, if the client's layout iomode differs from the I/O being performed the storage device may reject the client's I/O with an error indicating a new layout with the correct I/O mode should be fetched. E.g., if a client gets a layout with a READ iomode and performs a WRITE to a storage device, the storage device is allowed to reject that WRITE. The iomode does not conflict with OPEN share modes or lock requests; open mode checks and lock enforcement are always enforced, and are logically separate from the pNFS layout level. As well, open modes and locks are the preferred method for restricting user access to data files. E.g., an OPEN of read, deny-write does not conflict with a LAYOUTGET containing an iomode of READ/WRITE performed by another client. Applications that depend on writing into the same file concurrently may use byte range locking to serialize their accesses. 7.1.3 Layout Segments Until this point, layouts have been defined in a fairly vague manner. A layout is more precisely identified by the following tuple: ; the FH refers to the FH of the file on the metadata server. Note, layouts describe a file, not a byte-range of a file. Since a layout that describes an entire file may be very large, there is a desire to manage layouts in smaller chunks that correspond to byte-ranges of the file. For example, the entire layout need not be returned, recalled, or committed. These chunks are called "layout segments" and are further identified by the byte-range they represent. Layout operations require the identification of the layout segment (i.e., clientID, FH, layout type, and byte-range), as well as the iomode. This structure allows clients and metadata servers to aggregate the results of layout operations into a singly maintained layout. It is important to define when layout segments overlap and/or conflict with each other. For a layout segment to overlap another layout segment both segments must be of the same layout type, correspond to the same filehandle, and have the same iomode; in addition, the byte-ranges of the segments must overlap. Layout segments conflict, when they overlap and differ in the content of the layout (i.e., the storage device/file mapping parameters differ). Note, differing iomodes do not lead to conflicting layouts. It is permissible for layout segments with different iomodes, pertaining to the same byte range, to be held by the same client. Shepler Expires June 15, 2006 [Page 55] Internet-Draft NFSv4 Minior Version 1 December 2005 7.1.4 Device IDs The "deviceID" is a short name for a storage device. In practice, a significant amount of information may be required to fully identify a storage device. Instead of embedding all that information in a layout, a level of indirection is used. Layouts embed device IDs, and a new operation (GETDEVICEINFO) is used to retrieve the complete identity information about the storage device according to its layout type. For example, the identity of a file server or object server could be an IP address and port. The identity of a block device could be a volume label. Due to multipath connectivity in a SAN environment, agreement on a volume label is considered the reliable way to locate a particular storage device. The device ID is qualified by the layout type and unique per file system (FSID). This allows different layout drivers to generate device IDs without the need for co-ordination. In addition to GETDEVICEINFO, another operation, GETDEVICELIST, has been added to allow clients to fetch the mappings of multiple storage devices attached to a metadata server. Clients cannot expect the mapping between device ID and storage device address to persist across server reboots, hence a client MUST fetch new mappings on startup or upon detection of a metadata server reboot unless it can revalidate its existing mappings. Not all layout types support such revalidation, and the means of doing so is layout specific. If data are reorganized from a storage device with a given device ID to a different storage device (i.e., if the mapping between storage device and data changes), the layout describing the data MUST be recalled rather than assigning the new storage device to the old device ID. 7.1.5 Aggregation Schemes Aggregation schemes can describe layouts like simple one-to-one mapping, concatenation, and striping. A general aggregation scheme allows nested maps so that more complex layouts can be compactly described. The canonical aggregation type for this extension is striping, which allows a client to access storage devices in parallel. Even a one-to-one mapping is useful for a file server that wishes to distribute its load among a set of other file servers. 7.2 Guarantees Provided by Layouts Layouts delegate to the client the ability to access data out of band. The layout guarantees the holder that the layout will be recalled when the state encapsulated by the layout becomes invalid (e.g., through some operation that directly or indirectly modifies Shepler Expires June 15, 2006 [Page 56] Internet-Draft NFSv4 Minior Version 1 December 2005 the layout) or, possibly, when a conflicting layout is requested, as determined by the layout's iomode. When a layout is recalled, and then returned by the client, the client retains the ability to access file data with normal NFSv4 I/O operations through the metadata server. Only the right to do I/O out-of-band is affected. Holding a layout does not guarantee that a user of the layout has the rights to access the data represented by the layout. All user access rights MUST be obtained through the appropriate open, lock, and access operations (i.e., those that would be used in the absence of pNFS). However, if a valid layout for a file is not held by the client, the storage device should reject all I/Os to that file's byte range that originate from that client. In summary, layouts and ordinary file access controls are independent. The act of modifying a file for which a layout is held, does not necessarily conflict with the holding of the layout that describes the file being modified. However, with certain layout types (e.g., block/volume layouts), the layout's iomode must agree with the type of I/O being performed. Depending upon the layout type and storage protocol in use, storage device access permissions may be granted by LAYOUTGET and may be encoded within the type specific layout. If access permissions are encoded within the layout, the metadata server must recall the layout when those permissions become invalid for any reason; for example when a file becomes unwritable or inaccessible to a client. Note, clients are still required to perform the appropriate access operations as described above (e.g., open and lock ops). The degree to which it is possible for the client to circumvent these access operations must be clearly addressed by the individual layout type documents, as well as the consequences of doing so. In addition, these documents must be clear about the requirements and non- requirements for the checking performed by the server. If the pNFS metadata server supports mandatory byte range locks then byte range locks must behave as specified by the NFSv4 protocol, as observed by users of files. If a storage device is unable to restrict access by a pNFS client who does not hold a required mandatory byte range lock then the metadata server must not grant layouts to a client, for that storage device, that permits any access that conflicts with a mandatory byte range lock held by another client. In this scenario, it is also necessary for the metadata server to ensure that byte range locks are not granted to a client if any other client holds a conflicting layout; in this case all conflicting layouts must be recalled and returned before the lock request can be granted. This requires the pNFS server to understand the capabilities of its storage devices. Shepler Expires June 15, 2006 [Page 57] Internet-Draft NFSv4 Minior Version 1 December 2005 7.3 Getting a Layout A client obtains a layout through a new operation, LAYOUTGET. The metadata server will give out layouts of a particular type (e.g., block/volume, object, or file) and aggregation as requested by the client. The client selects an appropriate layout type which the server supports and the client is prepared to use. The layout returned to the client may not line up exactly with the requested byte range. A field within the LAYOUTGET request, "minlength", specifies the minimum overlap that MUST exist between the requested layout and the layout returned by the metadata server. The "minlength" field should specify a size of at least one. A metadata server may give-out multiple overlapping, non-conflicting layout segments to the same client in response to a LAYOUTGET. There is no implied ordering between getting a layout and performing a file OPEN. For example, a layout may first be retrieved by placing a LAYOUTGET operation in the same compound as the initial file OPEN. Once the layout has been retrieved, it can be held across multiple OPEN and CLOSE sequences. The storage protocol used by the client to access the data on the storage device is determined by the layout's type. The client needs to select a "layout driver" that understands how to interpret and use that layout. The API used by the client to talk to its drivers is outside the scope of the pNFS extension. The storage protocol between the client's layout driver and the actual storage is covered by other protocols specifications such as iSCSI (block storage), OSD (object storage) or NFS (file storage). Although, the metadata server is in control of the layout for a file, the pNFS client can provide hints to the server when a file is opened or created about preferred layout type and aggregation scheme. The pNFS extension introduces a LAYOUT_HINT attribute that the client can set at creation time to provide a hint to the server for new files. It is suggested that this attribute be set as one of the initial attributes to OPEN when creating a new file. Setting this attribute separately, after the file has been created could make it difficult, or impossible, for the server implementation to comply. 7.4 Committing a Layout Due to the nature of the protocol, the file attributes, and data location mapping (e.g., which offsets store data vs. store holes) that exist on the metadata storage device may become inconsistent in relation to the data stored on the storage devices; e.g., when WRITEs occur before a layout has been committed (e.g., between a LAYOUTGET and a LAYOUTCOMMIT). Thus, it is necessary to occasionally re-sync Shepler Expires June 15, 2006 [Page 58] Internet-Draft NFSv4 Minior Version 1 December 2005 this state and make it visible to other clients through the metadata server. The LAYOUTCOMMIT operation is responsible for committing a modified layout segment to the metadata server. Note: the data should be written and committed to the appropriate storage devices before the LAYOUTCOMMIT occurs. Note, if the data is being written asynchronously through the metadata server a COMMIT to the metadata server is required to sync the data and make it visible on the storage devices (see Section 7.6 for more details). The scope of this operation depends on the storage protocol in use. For block/ volume-based layouts, it may require updating the block list that comprises the file and committing this layout to stable storage. While, for file-layouts it requires some synchronization of attributes between the metadata and storage devices (i.e., mainly the size attribute; EOF). It is important to note that the level of synchronization is from the point of view of the client who issued the LAYOUTCOMMIT. The updated state on the metadata server need only reflect the state as of the client's last operation previous to the LAYOUTCOMMIT, it need not reflect a globally synchronized state (e.g., other clients may be performing, or may have performed I/O since the client's last operation and the LAYOUTCOMMIT). The control protocol is free to synchronize the attributes before it receives a LAYOUTCOMMIT, however upon successful completion of a LAYOUTCOMMIT, state that exists on the metadata server that describes the file MUST be in sync with the state existing on the storage devices that comprise that file as of the issuing client's last operation. Thus, a client that queries the size of a file between a WRITE to a storage device and the LAYOUTCOMMIT may observe a size that does not reflects the actual data written. 7.4.1 LAYOUTCOMMIT and mtime/atime/change The change attribute and the modify/access times may be updated, by the server, at LAYOUTCOMMIT time; since for some layout types, the change attribute and atime/mtime can not be updated by the appropriate I/O operation performed at a storage device. The arguments to LAYOUTCOMMIT allow the client to provide suggested access and modify time values to the server. Again, depending upon the layout type, these client provided values may or may not be used. The server should sanity check the client provided values before they are used. For example, the server should ensure that time does not flow backwards. According to the NFSv4 specification, The client always has the option to set these attributes through an explicit SETATTR operation. As mentioned, for some layout protocols the change attribute and Shepler Expires June 15, 2006 [Page 59] Internet-Draft NFSv4 Minior Version 1 December 2005 mtime/atime may be updated at or after the time the I/O occurred (e.g., if the storage device is able to communicate these attributes to the metadata server). If, upon receiving a LAYOUTCOMMIT, the server implementation is able to determine that the file did not change since the last time the change attribute was updated (e.g., no WRITEs or over-writes occurred), the implementation need not update the change attribute; file-based protocols may have enough state to make this determination or may update the change attribute upon each file modification. This also applies for mtime and atime; if the server implementation is able to determine that the file has not been modified since the last mtime update, the server need not update mtime at LAYOUTCOMMIT time. Once LAYOUTCOMMIT completes, the new change attribute and mtime/atime should be visible if that file was modified since the latest previous LAYOUTCOMMIT or LAYOUTGET. 7.4.2 LAYOUTCOMMIT and size The file's size may be updated at LAYOUTCOMMIT time as well. The LAYOUTCOMMIT operation contains an argument that indicates the last byte offset to which the client wrote ("last_write_offset"). Note: for this offset to be viewed as a file size it must be incremented by one byte (e.g., a write to offset 0 would map into a file size of 1, but the last write offset is 0). The metadata server may do one of the following: 1. It may update the file's size based on the last write offset. However, to the extent possible, the metadata server should sanity check any value to which the file's size is going to be set. E.g., it must not truncate the file based on the client presenting a smaller last write offset than the file's current size. 2. If it has sufficient other knowledge of file size (e.g., by querying the storage devices through the control protocol), it may ignore the client provided argument and use the query-derived value. 3. It may use the last write offset as a hint, subject to correction when other information is available as above. The method chosen to update the file's size will depend on the storage device's and/or the control protocol's implementation. For example, if the storage devices are block devices with no knowledge of file size, the metadata server must rely on the client to set the size appropriately. A new size flag and length are also returned in the results of a LAYOUTCOMMIT. This union indicates whether a new size was set, and to what length it was set. If a new size is set as a result of LAYOUTCOMMIT, then the metadata server must reply with Shepler Expires June 15, 2006 [Page 60] Internet-Draft NFSv4 Minior Version 1 December 2005 the new size. As well, if the size is updated, the metadata server in conjunction with the control protocol SHOULD ensure that the new size is reflected by the storage devices immediately upon return of the LAYOUTCOMMIT operation; e.g., a READ up to the new file size should succeed on the storage devices (assuming no intervening truncations). Again, if the client wants to explicitly zero-extend or truncate a file, SETATTR must be used; it need not be used when simply writing past EOF. Since client layout holders may be unaware of changes made to the file's size, through LAYOUTCOMMIT or SETATTR, by other clients, an additional callback/notification has been added for pNFS. CB_SIZECHANGED is a notification that the metadata server sends to layout holders to notify them of a change in file size. This is preferred over issuing CB_LAYOUTRECALL to each of the layout holders. 7.4.3 LAYOUTCOMMIT and layoutupdate The LAYOUTCOMMIT operation contains a "layoutupdate" argument. This argument is a layout type specific structure. The structure can be used to pass arbitrary layout type specific information from the client to the metadata server at LAYOUTCOMMIT time. For example, if using a block/volume layout, the client can indicate to the metadata server which reserved or allocated blocks it used and which it did not. The "layoutupdate" structure need not be the same structure as the layout returned by LAYOUTGET. The structure is defined by the layout type and is opaque to LAYOUTCOMMIT. 7.5 Recalling a Layout 7.5.1 Basic Operation Since a layout protects a client's access to a file via a direct client-storage-device path, a layout need only be recalled when it is semantically unable to serve this function. Typically, this occurs when the layout no longer encapsulates the true location of the file over the byte range it represents. Any operation or action (e.g., server driven restriping or load balancing) that changes the layout will result in a recall of the layout. A layout is recalled by the CB_LAYOUTRECALL callback operation (see Section 14.19). This callback can either recall a layout segment identified by a byte range, or all the layouts associated with a file system (FSID). However, there is no single operation to return all layouts associated with an FSID; multiple layout segments may be returned in a single compound operation. Section 7.5.3 discusses sequencing issues surrounding the getting, returning, and recalling of layouts. The iomode is also specified when recalling a layout or layout Shepler Expires June 15, 2006 [Page 61] Internet-Draft NFSv4 Minior Version 1 December 2005 segment. Generally, the iomode in the recall request must match the layout, or segment, being returned; e.g., a recall with an iomode of RW should cause the client to only return RW layout segments (not R segments). However, a special LAYOUTIOMODE_ANY enumeration is defined to enable recalling a layout of any type (i.e., the client must return both read-only and read/write layouts). A REMOVE operation may cause the metadata server to recall the layout to prevent the client from accessing a non-existent file and to reclaim state stored on the client. Since a REMOVE may be delayed until the last close of the file has occurred, the recall may also be delayed until this time. As well, once the file has been removed, after the last reference, the client SHOULD no longer be able to perform I/O using the layout (e.g., with file-based layouts an error such as ESTALE could be returned). Although, the pNFS extension does not alter the caching capabilities of clients, or their semantics, it recognizes that some clients may perform more aggressive write-behind caching to optimize the benefits provided by pNFS. However, write-behind caching may impact the latency in returning a layout in response to a CB_LAYOUTRECALL; just as caching impacts DELEGRETURN with regards to data delegations. Client implementations should limit the amount of dirty data they have outstanding at any one time. Server implementations may fence clients from performing direct I/O to the storage devices if they perceive that the client is taking too long to return a layout once recalled. A server may be able to monitor client progress by watching client I/Os or by observing LAYOUTRETURNs of sub-portions of the recalled layout. The server can also limit the amount of dirty data to be flushed to storage devices by limiting the byte ranges covered in the layouts it gives out. Once a layout has been returned, the client MUST NOT issue I/Os to the storage devices for the file, byte range, and iomode represented by the returned layout. If a client does issue an I/O to a storage device for which it does not hold a layout, the storage device SHOULD reject the I/O. 7.5.2 Recall Callback Robustness For simplicity, the discussion thus far has assumed that pNFS client state for a file exactly matches the pNFS server state for that file and client regarding layout ranges and permissions. This assumption leads to the implicit assumption that any callback results in a LAYOUTRETURN or set of LAYOUTRETURNs that exactly match the range in the callback, since both client and server agree about the state being maintained. However, it can be useful if this assumption does not always hold. For example: Shepler Expires June 15, 2006 [Page 62] Internet-Draft NFSv4 Minior Version 1 December 2005 o It may be useful for clients to be able to discard layout information without calling LAYOUTRETURN. If conflicts that require callbacks are very rare, and a server can use a multi-file callback to recover per-client resources (e.g., via a FSID recall, or a multi-file recall within a single compound), the result may be significantly less client-server pNFS traffic. o It may be similarly useful for servers to enhance information about what layout ranges are held by a client beyond what a client actually holds. In the extreme, a server could manage conflicts on a per-file basis, only issuing whole-file callbacks even though clients may request and be granted sub-file ranges. o As well, the synchronized state assumption is not robust to minor errors. A more robust design would allow for divergence between client and server and the ability to recover. It is vital that a client not assign itself layout permissions beyond what the server has granted and that the server not forget layout permissions that have been granted in order to avoid errors. On the other hand, if a server believes that a client holds a layout segment that the client does not know about, it's useful for the client to be able to issue the LAYOUTRETURN that the server is expecting in response to a recall. Thus, in light of the above, it is useful for a server to be able to issue callbacks for layout ranges it has not granted to a client, and for a client to return ranges it does not hold. A pNFS client must always return layout segments that comprise the full range specified by the recall. Note, the full recalled layout range need not be returned as part of a single operation, but may be returned in segments. This allows the client to stage the flushing of dirty data, layout commits, and returns. Also, it indicates to the metadata server that the client is making progress. In order to ensure client/server convergence on the layout state, the final LAYOUTRETURN operation in a sequence of returns for a particular recall, SHOULD specify the entire range being recalled, even if layout segments pertaining to partial ranges were previously returned. In addition, if the client holds no layout segment that overlaps the range being recalled, the client should return the NFS4ERR_NOMATCHING_LAYOUT error code. This allows the server to update its view of the client's layout state. 7.5.3 Recall/Return Sequencing As with other stateful operations, pNFS requires the correct sequencing of layout operations. This proposal assumes that sessions will precede or accompany pNFS into NFSv4.x and thus, pNFS will Shepler Expires June 15, 2006 [Page 63] Internet-Draft NFSv4 Minior Version 1 December 2005 require the use of sessions. If the sessions proposal does not precede pNFS, then this proposal needs to be modified to provide for the correct sequencing of pNFS layout operations. Also, this specification is reliant on the sessions protocol to provide the correct sequencing between regular operations and callbacks. It is the server's responsibility to avoid inconsistencies regarding the layouts it hands out and the client's responsibility to properly serialize its layout requests. One critical issue with operation sequencing concerns callbacks. The protocol must defend against races between the reply to a LAYOUTGET operation and a subsequent CB_LAYOUTRECALL. It MUST NOT be possible for a client to process the CB_LAYOUTRECALL for a layout that it has not received in a reply message to a LAYOUTGET. 7.5.3.1 Client Side Considerations Consider a pNFS client that has issued a LAYOUTGET and then receives an overlapping recall callback for the same file. There are two possibilities, which the client cannot distinguish when the callback arrives: 1. The server processed the LAYOUTGET before issuing the recall, so the LAYOUTGET response is in flight, and must be waited for because it may be carrying layout info that will need to be returned to deal with the recall callback. 2. The server issued the callback before receiving the LAYOUTGET. The server will not respond to the LAYOUTGET until the recall callback is processed. This can cause deadlock, as the client must wait for the LAYOUTGET response before processing the recall in the first case, but that response will not arrive until after the recall is processed in the second case. This deadlock can be avoided by adhering to the following requirements: o A LAYOUTGET MUST be rejected with an error (i.e., NFS4ERR_RECALLCONFLICT) if there's an overlapping outstanding recall callback to the same client o When processing a recall, the client MUST wait for a response to all conflicting outstanding LAYOUTGETs before performing any RETURN that could be affected by any such response. o The client SHOULD wait for responses to all operations required to complete a recall before sending any LAYOUTGETs that would conflict with the recall because the server is likely to return Shepler Expires June 15, 2006 [Page 64] Internet-Draft NFSv4 Minior Version 1 December 2005 errors for them. Now the client can wait for the LAYOUTGET response, as it will be received in both cases. 7.5.3.2 Server Side Considerations Consider a related situation from the pNFS server's point of view. The server has issued a recall callback and receives an overlapping LAYOUTGET for the same file before the LAYOUTRETURN(s) that respond to the recall callback. Again, there are two cases: 1. The client issued the LAYOUTGET before processing the recall callback. 2. The client issued the LAYOUTGET after processing the recall callback, but it arrived before the LAYOUTRETURN that completed that processing. The simplest approach is to always reject the overlapping LAYOUTGET. The client has two ways to avoid this result - it can issue the LAYOUTGET as a subsequent element of a COMPOUND containing the LAYOUTRETURN that completes the recall callback, or it can wait for the response to that LAYOUTRETURN. This leads to a more general problem; in the absence of a callback if a client issues concurrent overlapping LAYOUTGET and LAYOUTRETURN operations, it is possible for the server to process them in either order. Again, a client must take the appropriate precautions in serializing its actions. [ASIDE: HighRoad forbids a client from doing this, as the per-file layout stateid will cause one of the two operations to be rejected with a stale layout stateid. This approach is simpler and produces better results by comparison to allowing concurrent operations, at least for this sort of conflict case, because server execution of operations in an order not anticipated by the client may produce results that are not useful to the client (e.g., if a LAYOUTRETURN is followed by a concurrent overlapping LAYOUTGET, but executed in the other order, the client will not retain layout extents for the overlapping range).] 7.6 Metadata Server Write Propagation Asynchronous writes written through the metadata server may be propagated lazily to the storage devices. For data written asynchronously through the metadata server, a client performing a read at the appropriate storage device is not guaranteed to see the Shepler Expires June 15, 2006 [Page 65] Internet-Draft NFSv4 Minior Version 1 December 2005 newly written data until a COMMIT occurs at the metadata server. While the write is pending, reads to the storage device can give out either the old data, the new data, or a mixture thereof. After either a synchronous write completes, or a COMMIT is received (for asynchronously written data), the metadata server must ensure that storage devices give out the new data and that the data has been written to stable storage. If the server implements its storage in any way such that it cannot obey these constraints, then it must recall the layouts to prevent reads being done that cannot be handled correctly. 7.7 Crash Recovery Crash recovery is complicated due to the distributed nature of the pNFS protocol. In general, crash recovery for layouts is similar to crash recovery for delegations in the base NFSv4 protocol. However, the client's ability to perform I/O without contacting the metadata server introduces subtleties that must be handled correctly if file system corruption is to be avoided. 7.7.1 Leases The layout lease period plays a critical role in crash recovery. Depending on the capabilities of the storage protocol, it is crucial that the client is able to maintain an accurate layout lease timer to ensure that I/Os are not issued to storage devices after expiration of the layout lease period. In order for the client to do so, it must know which operations renew a lease. 7.7.1.1 Lease Renewal The current NFSv4 specification allows for implicit lease renewals to occur upon receiving an I/O. However, due to the distributed pNFS architecture, implicit lease renewals are limited to operations performed at the metadata server; this includes I/O performed through the metadata server. So, a client must not assume that READ and WRITE I/O to storage devices implicitly renew lease state. If sessions are required for pNFS, as has been suggested, then the SEQUENCE operation is to be used to explicitly renew leases. It is proposed that the SEQUENCE operation be extended to return all the specific information that RENEW does, but not as an error as RENEW returns it. Since, when using session, beginning each compound with the SEQUENCE op allows renews to be performed without an additional operation and without an additional request. Again, the client must not rely on any operation to the storage devices to renew a lease. Using the SEQUENCE operation for renewals, simplifies the client's perception of lease renewal. Shepler Expires June 15, 2006 [Page 66] Internet-Draft NFSv4 Minior Version 1 December 2005 7.7.1.2 Client Lease Timer Depending on the storage protocol and layout type in use, it may be crucial that the client not issue I/Os to storage devices if the corresponding layout's lease has expired. Doing so may lead to file system corruption if the layout has been given out and used by another client. In order to prevent this, the client must maintain an accurate lease timer for all layouts held. RFC3530 has the following to say regarding the maintenance of a client lease timer: ...the client must track operations which will renew the lease period. Using the time that each such request was sent and the time that the corresponding reply was received, the client should bound the time that the corresponding renewal could have occurred on the server and thus determine if it is possible that a lease period expiration could have occurred. To be conservative, the client should start its lease timer based on the time that the it issued the operation to the metadata server, rather than based on the time of the response. It is also necessary to take propagation delay into account when requesting a renewal of the lease: ...the client should subtract it from lease times (e.g., if the client estimates the one-way propagation delay as 200 msec, then it can assume that the lease is already 200 msec old when it gets it). In addition, it will take another 200 msec to get a response back to the server. So the client must send a lock renewal or write data back to the server 400 msec before the lease would expire. Thus, the client must be aware of the one-way propagation delay and should issue renewals well in advance of lease expiration. Clients, to the extent possible, should try not to issue I/Os that may extend past the lease expiration time period. However, since this is not always possible, the storage protocol must be able to protect against the effects of inflight I/Os, as is discussed later. 7.7.2 Client Recovery Client recovery for layouts works in much the same way as NFSv4 client recovery works for other lock/delegation state. When an NFSv4 client reboots, it will lose all information about the layouts that it previously owned. There are two methods by which the server can reclaim these resources and allow otherwise conflicting layouts to be provided to other clients. Shepler Expires June 15, 2006 [Page 67] Internet-Draft NFSv4 Minior Version 1 December 2005 The first is through the expiry of the client's lease. If the client recovery time is longer than the lease period, the client's lease will expire and the server will know that state may be released. for layouts the server may release the state immediately upon lease expiry or it may allow the layout to persist awaiting possible lease revival, as long as there are no conflicting requests. On the other hand, the client may recover in less time than it takes for the lease period to expire. In such a case, the client will contact the server through the standard SETCLIENTID protocol. The server will find that the client's id matches the id of the previous client invocation, but that the verifier is different. The server uses this as a signal to release all the state associated with the client's previous invocation. 7.7.3 Metadata Server Recovery The server recovery case is slightly more complex. In general, the recovery process again follows the standard NFSv4 recovery model: the client will discover that the metadata server has rebooted when it receives an unexpected STALE_STATEID or STALE_CLIENTID reply from the server; it will then proceed to try to reclaim its previous delegations during the server's recovery grace period. However, layouts are not reclaimable in the same sense as data delegations; there is no reclaim bit, thus no guarantee of continuity between the previous and new layout. This is not necessarily required since a layout is not required to perform I/O; I/O can always be performed through the metadata server. [NOTE: there is no reclaim bit for getting a layout. Thus, in the case of reclaiming an old layout obtained through LAYOUTGET, there is no guarantee of continuity. If a reclaim bit existed a block/volume layout type might be happier knowing it got the layout back with the assurance of continuity. However, this would require the metadata server trusting the client in telling it the exact layout it had (i.e., the full block-list); however, divergence is avoided by having the server tell the client what is contained within the layout.] If the client has dirty data that it needs to write out, or an outstanding LAYOUTCOMMIT, the client should try to obtain a new layout segment covering the byte range covered by the previous layout segment. However, the client might not not get the same layout segment it had. The range might be different or it might get the same range but the content of the layout might be different. For example, if using a block/volume-based layout, the blocks provisionally assigned by the layout might be different, in which case the client will have to write the corresponding blocks again; in the interest of simplicity, the client might decide to always write Shepler Expires June 15, 2006 [Page 68] Internet-Draft NFSv4 Minior Version 1 December 2005 them again. Alternatively, the client might be unable to obtain a new layout and thus, must write the data using normal NFSv4 through the metadata server. There is an important safety concern associated with layouts that does not come into play in the standard NFSv4 case. If a standard NFSv4 client makes use of a stale delegation, while reading, the consequence could be to deliver stale data to an application. If writing, using a stale delegation or a stale state stateid for an open or lock would result in the rejection of the client's write with the appropriate stale stateid error. However, the pNFS layout enables the client to directly access the file system storage---if this access is not properly managed by the NFSv4 server the client can potentially corrupt the file system data or metadata. Thus, it is vitally important that the client discover that the metadata server has rebooted, and that the client stops using stale layouts before the metadata server gives them away to other clients. To ensure this, the client must be implemented so that layouts are never used to access the storage after the client's lease timer has expired. It is crucial that clients have precise knowledge of the lease periods of their layouts. For specific details on lease renewal and client lease timers, see Section 7.7.1. The prohibition on using stale layouts applies to all layout related accesses, especially the flushing of dirty data to the storage devices. If the client's lease timer expires because the client could not contact the server for any reason, the client MUST immediately stop using the layout until the server can be contacted and the layout can be officially recovered or reclaimed. However, this is only part of the solution. It is also necessary to deal with the consequences of I/Os already in flight. The issue of the effects of I/Os started before lease expiration and possibly continuing through lease expiration is the responsibility of the data storage protocol and as such is layout type specific. There are two approaches the data storage protocol can take. The protocol may adopt a global solution which prevents all I/Os from being executed after the lease expiration and thus is safe against a client who issues I/Os after lease expiration. This is the preferred solution and the solution used by NFSv4 file based layouts (see Section 9.6); as well, the object storage device protocol allows storage to fence clients after lease expiration. Alternatively, the storage protocol may rely on proper client operation and only deal with the effects of lingering I/Os. These solutions may impact the client layout-driver, the metadata server layout-driver, and the control protocol. Shepler Expires June 15, 2006 [Page 69] Internet-Draft NFSv4 Minior Version 1 December 2005 7.7.4 Storage Device Recovery Storage device crash recovery is mostly dependent upon the layout type in use. However, there are a few general techniques a client can use if it discovers a storage device has crashed while holding asynchronously written, non-committed, data. First and foremost, it is important to realize that the client is the only one who has the information necessary to recover asynchronously written data; since, it holds the dirty data and most probably nobody else does. Second, the best solution is for the client to err on the side or caution and attempt to re-write the dirty data through another path. The client, rather than hold the asynchronously written data indefinitely, is encouraged to, and can make sure that the data is written by using other paths to that data. The client may write the data to the metadata server, either synchronously or asynchronously with a subsequent COMMIT. Once it does this, there is no need to wait for the original storage device. In the event that the data range to be committed is transferred to a different storage device, as indicated in a new layout, the client may write to that storage device. Once the data has been committed at that storage device, either through a synchronous write or through a commit to that storage device (e.g., through the NFSv4 COMMIT operation for the NFSv4 file layout), the client should consider the transfer of responsibility for the data to the new server as strong evidence that this is the intended and most effective method for the client to get the data written. In either case, once the write is on stable storage (through either the storage device or metadata server), there is no need to continue either attempting to commit or attempting to synchronously write the data to the original storage device or wait for that storage device to become available. That storage device may never be visible to the client again. This approach does have a "lingering write" problem, similar to regular NFSv4. Suppose a WRITE is issued to a storage device for which no response is received. The client breaks the connection, trying to re-establish a new one, and gets a recall of the layout. The client issues the I/O for the dirty data through an alternative path, for example, through the metadata server and it succeeds. The client then goes on to perform additional writes that all succeed. If at some time later, the original write to the storage device succeeds, data inconsistency could result. The same problem can occur in regular NFSv4. For example, a WRITE is held in a switch for some period of time while other writes are issued and replied to, if the original WRITE finally succeeds, the same issues can occur. However, this is solved by sessions in NFSv4.x. Shepler Expires June 15, 2006 [Page 70] Internet-Draft NFSv4 Minior Version 1 December 2005 8. Security Considerations The pNFS extension partitions the NFSv4 file system protocol into two parts, the control path and the data path (i.e., storage protocol). The control path contains all the new operations described by this extension; all existing NFSv4 security mechanisms and features apply to the control path. The combination of components in a pNFS system (see Figure 9) is required to preserve the security properties of NFSv4 with respect to an entity accessing data via a client, including security countermeasures to defend against threats that NFSv4 provides defenses for in environments where these threats are considered significant. In some cases, the security countermeasures for connections to storage devices may take the form of physical isolation or a recommendation not to use pNFS in an environment. For example, it is currently infeasible to provide confidentiality protection for some storage device access protocols to protect against eavesdropping; in environments where eavesdropping on such protocols is of sufficient concern to require countermeasures, physical isolation of the communication channel (e.g., via direct connection from client(s) to storage device(s)) and/or a decision to forego use of pNFS (e.g., and fall back to NFSv4) may be appropriate courses of action. In full generality where communication with storage devices is subject to the same threats as client-server communication, the protocols used for that communication need to provide security mechanisms comparable to those available via RPSEC_GSS for NFSv4. Many situations in which pNFS is likely to be used will not be subject to the overall threat profile for which NFSv4 is required to provide countermeasures. pNFS implementations MUST NOT remove NFSv4's access controls. The combination of clients, storage devices, and the server are responsible for ensuring that all client to storage device file data access respects NFSv4 ACLs and file open modes. This entails performing both of these checks on every access in the client, the storage device, or both. If a pNFS configuration performs these checks only in the client, the risk of a misbehaving client obtaining unauthorized access is an important consideration in determining when it is appropriate to use such a pNFS configuration. Such configurations SHOULD NOT be used when client- only access checks do not provide sufficient assurance that NFSv4 access control is being applied correctly. The following subsections describe security considerations specifically applicable to each of the three major storage device protocol types supported for pNFS. Shepler Expires June 15, 2006 [Page 71] Internet-Draft NFSv4 Minior Version 1 December 2005 [Requiring strict equivalence to NFSv4 security mechanisms is the wrong approach. Will need to lay down a set of statements that each protocol has to make starting with access check location/properties.] 8.1 File Layout Security A NFSv4 file layout type is defined in Section 9; see Section 9.7 for additional security considerations and details. In summary, the NFSv4 file layout type requires that all I/O access checks MUST be performed by the storage devices, as defined by the NFSv4 specification. If another file layout type is being used, additional access checks may be required. But in all cases, the access control performed by the storage devices must be at least as strict as that specified by the NFSv4 protocol. 8.2 Object Layout Security The object storage protocol MUST implement the security aspects described in version 1 of the T10 OSD protocol definition [5]. The remainder of this section gives an overview of the security mechanism described in that standard. The goal is to give the reader a basic understanding of the object security model. Any discrepancies between this text and the actual standard are obviously to be resolved in favor of the OSD standard. The object storage protocol relies on a cryptographically secure capability to control accesses at the object storage devices. Capabilities are generated by the metadata server, returned to the client, and used by the client as described below to authenticate their requests to the Object Storage Device (OSD). Capabilities therefore achieve the required access and open mode checking. They allow the file server to define and check a policy (e.g., open mode) and the OSD to check and enforce that policy without knowing the details (e.g., user IDs and ACLs). Since capabilities are tied to layouts, and since they are used to enforce access control, the server should recall layouts and revoke capabilities when the file ACL or mode changes in order to signal the clients. Each capability is specific to a particular object, an operation on that object, a byte range w/in the object, and has an explicit expiration time. The capabilities are signed with a secret key that is shared by the object storage devices (OSD) and the metadata managers. clients do not have device keys so they are unable to forge capabilities. The the following sketch of the algorithm should help the reader understand the basic model. LAYOUTGET returns Shepler Expires June 15, 2006 [Page 72] Internet-Draft NFSv4 Minior Version 1 December 2005 {CapKey = MAC(CapArgs), CapArgs} The client uses CapKey to sign all the requests it issues for that object using the respective CapArgs. In other words, the CapArgs appears in the request to the storage device, and that request is signed with the CapKey as follows: ReqMAC = MAC(Req, Nonceln) The following is sent to the OSD: {CapArgs, Req, Nonceln, ReqMAC}. The OSD uses the SecretKey it shares with the metadata server to compare the ReqMAC the client sent with a locally computed MAC(CapArgs)>(Req, Nonceln) and if they match the OSD assumes that the capabilities came from an authentic metadata server and allows access to the object, as allowed by the CapArgs. Therefore, if the server LAYOUTGET reply, holding CapKey and CapArgs, is snooped by another client, it can be used to generate valid OSD requests (within the CapArgs access restriction). To provide the required privacy requirements for the capabilities returned by LAYOUTGET, the GSS-API can be used, e.g. by using a session key known to the file server and to the client to encrypt the whole layout or parts of it. Two general ways to provide privacy in the absence of GSS-API that are independent of NFSv4 are either an isolated network such as a VLAN or a secure channel provided by IPsec. 8.3 Block/Volume Layout Security As typically used, block/volume protocols rely on clients to enforce file access checks since the storage devices are generally unaware of the files they are storing and in particular are unaware of which blocks belongs to which file. In such environments, the physical addresses of blocks are exported to pNFS clients via layouts. An alternative method of block/volume protocol use is for the storage devices to export virtualized block addresses, which do reflect the files to which blocks belong. These virtual block addresses are exported to pNFS clients via layouts. This allows the storage device to make appropriate access checks, while mapping virtual block addresses to physical block addresses. In environments where access control is important and client-only access checks provide insufficient assurance of access control enforcement (e.g., there is concern about a malicious of malfunctioning client skipping the access checks) and where physical block addresses are exported to clients, the storage devices will Shepler Expires June 15, 2006 [Page 73] Internet-Draft NFSv4 Minior Version 1 December 2005 generally be unable to compensate for these client deficiencies. In such threat environments, block/volume protocols SHOULD NOT be used with pNFS, unless the storage device is able to implement the appropriate access checks, via use of virtualized block addresses, or other means. NFSv4 without pNFS or pNFS with a different type of storage protocol would be a more suitable means to access files in such environments. Storage-device/protocol-specific methods (e.g. LUN masking/mapping) may be available to prevent malicious or high- risk clients from directly accessing storage devices. 9. The NFSv4 File Layout Type This section describes the semantics and format of NFSv4 file-based layouts. 9.1 File Striping and Data Access The file layout type describes a method for striping data across multiple devices. The data for each stripe unit is stored within an NFSv4 file located on a particular storage device. The structures used to describe the stripe layout are as follows: enum stripetype4 { STRIPE_SPARSE = 1, STRIPE_DENSE = 2 }; struct nfsv4_file_layouthint { stripetype4 stripe_type; length4 stripe_unit; uint32_t stripe_width; }; struct nfsv4_file_layout { /* Per data stripe */ pnfs_deviceid4 dev_id<>; nfs_fh4 fh; }; struct nfsv4_file_layouttype4 { /* Per file */ stripetype4 stripe_type; length4 stripe_unit; length4 file_size; nfsv4_file_layout dev_list<>; }; The file layout specifies an ordered array of tuples, as well as the stripe size, type of stripe layout (discussed Shepler Expires June 15, 2006 [Page 74] Internet-Draft NFSv4 Minior Version 1 December 2005 a little later), and the file's current size as of LAYOUTGET time. The filehandle, "fh", identifies the file on a storage device identified by "dev_id", that holds a particular stripe of the file. The "dev_id" array can be used for multipathing and is discussed further in Section 9.1.3. The stripe width is determined by the stripe unit size multiplied by the number of devices in the dev_list. The stripe held by is determined by that tuples position within the device list, "dev_list". For example, consider a dev_list consisting of the following pairs: <(1,0x12), (2,0x13), (1,0x15)> and stripe_unit = 32KB The stripe width is 32KB * 3 devices = 96KB. The first entry specifies that on device 1 in the data file with filehandle 0x12 holds the first 32KB of data (and every 32KB stripe beginning where the file's offset % 96KB == 0). Devices may be repeated multiple times within the device list array; this is shown where storage device 1 holds both the first and third stripe of data. Filehandles can only be repeated if a sparse stripe type is used. Data is striped across the devices in the order listed in the device list array in increments of the stripe size. A data file stored on a storage device MUST map to a single file as defined by the metadata server; i.e., data from two files as viewed by the metadata server MUST NOT be stored within the same data file on any storage device. The "stripe_type" field specifies how the data is laid out within the data file on a storage device. It allows for two different data layouts: sparse and dense or packed. The stripe type determines the calculation that must be made to map the client visible file offset to the offset within the data file located on the storage device. The layout hint structure is described in more detail in Section 10.7. It is used, by the client, as by the FILE_LAYOUT_HINT attribute to specify the type of layout to be used for a newly created file. 9.1.1 Sparse and Dense Storage Device Data Layouts The stripe_type field allows for two storage device data file representations. Example sparse and dense storage device data layouts are illustrated below: Shepler Expires June 15, 2006 [Page 75] Internet-Draft NFSv4 Minior Version 1 December 2005 Sparse file-layout (stripe_unit = 4KB) ------------------ Is represented by the following file layout on the storage devices: Offset ID:0 ID:1 ID:2 0 +--+ +--+ +--+ +--+ indicates a |//| | | | | |//| stripe that 4KB +--+ +--+ +--+ +--+ contains data | | |//| | | 8KB +--+ +--+ +--+ | | | | |//| 12KB +--+ +--+ +--+ |//| | | | | 16KB +--+ +--+ +--+ | | |//| | | +--+ +--+ +--+ The sparse file-layout has holes for the byte ranges not exported by that storage device. This allows clients to access data using the real offset into the file, regardless of the storage device's position within the stripe. However, if a client writes to one of the holes (e.g., offset 4-12KB on device 1), then an error MUST be returned by the storage device. This requires that the storage device have knowledge of the layout for each file. When using a sparse layout, the offset into the storage device data file is the same as the offset into the main file. Dense/packed file-layout (stripe_unit = 4KB) ------------------------ Is represented by the following file layout on the storage devices: Offset ID:0 ID:1 ID:2 0 +--+ +--+ +--+ |//| |//| |//| 4KB +--+ +--+ +--+ |//| |//| |//| 8KB +--+ +--+ +--+ |//| |//| |//| 12KB +--+ +--+ +--+ |//| |//| |//| 16KB +--+ +--+ +--+ |//| |//| |//| +--+ +--+ +--+ The dense or packed file-layout does not leave holes on the storage Shepler Expires June 15, 2006 [Page 76] Internet-Draft NFSv4 Minior Version 1 December 2005 devices. Each stripe unit is spread across the storage devices. As such, the storage devices need not know the file's layout since the client is allowed to write to any offset. The calculation to determine the byte offset within the data file for dense storage device layouts is: stripe_width = stripe_unit * N; where N = |dev_list| dev_offset = floor(file_offset / stripe_width) * stripe_unit + file_offset % stripe_unit Regardless of the storage device data file layout, the calculation to determine the index into the device array is the same: dev_idx = floor(file_offset / stripe_unit) mod N Section 9.5 describe the semantics for dealing with reads to holes within the striped file. This is of particular concern, since each individual component stripe file (i.e., the component of the striped file that lives on a particular storage device) may be of different length. Thus, clients may experience 'short' reads when reading off the end of one of these component files. 9.1.2 Metadata and Storage Device Roles In many cases, the metadata server and the storage device will be separate pieces of physical hardware. The specification text is written as if that were always case. However, it can be the case that the same physical hardware is used to implement both a metadata and storage device and in this case, the specification text's references to these two entities are to be understood as referring to the same physical hardware implementing two distinct roles and it is important that it be clearly understood on behalf of which role the hardware is executing at any given time. Two sub-cases can be distinguished. In the first sub-case, the same physical hardware is used to implement both a metadata and data server in which each role is addressed through a distinct network interface (e.g., IP addresses for the metadata server and storage device are distinct). As long as the storage device address is obtained from the layout and is distinct from the metadata server's address, using the device ID therein to obtain the appropriate storage device address, it is always clear, for any given request, to what role it is directed, based on the destination IP address. However, it may also be the case that even though the metadata server and storage device are distinct from one client's point of view, the roles may be reversed according to another client's point of view. Shepler Expires June 15, 2006 [Page 77] Internet-Draft NFSv4 Minior Version 1 December 2005 For example, in the cluster file system model a metadata server to one client, may be a storage device to another client. Thus, it is safer to always mark the filehandle so that operations addressed to storage devices can be distinguished. The second sub-case is where both the metadata and storage device have the same network address. This requires us to make the distinction as to which role each request is directed, on a another basis. Since the network address is the same, the request is understood as being directed at one or the other, based on the filehandle of the first current filehandle value for the request. If the first current file handle is one derived from a layout (i.e., it is specified within the layout) (and it is recommended that these be distinguishable), then the request is to be considered as executed by a storage device. Otherwise, the operation is to be understood as executed by the metadata server. If a current filehandle is set that is inconsistent with the role to which it is directed, then the error NFS4ERR_BADHANDLE should result. For example, if a request is directed at the storage device, because the first current handle is from a layout, any attempt to set the current filehandle to be a value not from a layout should be rejected. Similarly, if the first current file handle was for a value not from a layout, a subsequent attempt to set the current file handle to a value obtained from a layout should be rejected. 9.1.3 Device Multipathing The NFSv4 file layout supports multipathing to 'equivalent' devices. Device-level multipathing is primarily of use in the case of a data server failure --- it allows the client to switch to another storage device that is exporting the same data stripe, without having to contact the metadata server for a new layout. To support device multipathing, an array of device IDs is encoded within the data stripe portion of the file's layout. This array represents an ordered list of devices where the first element has the highest priority. Each device in the list MUST be 'equivalent' to every other device in the list and each device must be attempted in the order specified. Equivalent devices MUST export the same system image (e.g., the stateids and filehandles that they use are the same) and must provide the same consistency guarantees. Two equivalent storage devices must also have sufficient connections to the storage, such that writing to one storage device is equivalent to writing to another, this also applies to reading. Also, if multiple copies of the same data exist, reading from one must provide access to all existing copies. As Shepler Expires June 15, 2006 [Page 78] Internet-Draft NFSv4 Minior Version 1 December 2005 such, it is unlikely that multipathing will provide additional benefit in the case of an I/O error. [NOTE: the error cases in which a client is expected to attempt an equivalent storage device should be specified.] 9.1.4 Operations Issued to Storage Devices Clients MUST use the filehandle described within the layout when accessing data on the storage devices. When using the layout's filehandle, the client MUST only issue READ, WRITE, PUTFH, COMMIT, and NULL operations to the storage device associated with that filehandle. If a client issues an operation other than those specified above, using the filehandle and storage device listed in the client's layout, that storage device SHOULD return an error to the client. The client MUST follow the instruction implied by the layout (i.e., which filehandles to use on which devices). As described in Section 7.2, a client MUST NOT issue I/Os to storage devices for which it does not hold a valid layout. The storage devices may reject such requests. GETATTR and SETATTR MUST be directed to the metadata server. In the case of a SETATTR of the size attribute, the control protocol is responsible for propagating size updates/truncations to the storage devices. In the case of extending WRITEs to the storage devices, the new size must be visible on the metadata server once a LAYOUTCOMMIT has completed (see Section 7.4.2). Section 9.5, describes the mechanism by which the client is to handle storage device file's that do not reflect the metadata server's size. 9.2 Global Stateid Requirements Note, there are no stateids returned embedded within the layout. The client MUST use the stateid representing open or lock state as returned by an earlier metadata operation (e.g., OPEN, LOCK), or a special stateid to perform I/O on the storage devices, as in regular NFSv4. Special stateid usage for I/O is subject to the NFSv4 protocol specification. The stateid used for I/O MUST have the same effect and be subject to the same validation on storage device as it would if the I/O was being performed on the metadata server itself in the absence of pNFS. This has the implication that stateids are globally valid on both the metadata and storage devices. This requires the metadata server to propagate changes in lock and open state to the storage devices, so that the storage devices can validate I/O accesses. This is discussed further in Section 9.4. Depending on when stateids are propagated, the existence of a valid stateid on the storage device may act as proof of a valid layout. Shepler Expires June 15, 2006 [Page 79] Internet-Draft NFSv4 Minior Version 1 December 2005 [NOTE: a number of proposals have been made that have the possibility of limiting the amount of validation performed by the storage device, if any of these proposals are accepted or obtain consensus, the global stateid requirement can be revisited.] 9.3 The Layout Iomode The layout iomode need not used by the metadata server when servicing NFSv4 file-based layouts, although in some circumstances it may be useful to use. For example, if the server implementation supports reading from read-only replicas or mirrors, it would be useful for the server to return a layout enabling the client to do so. As such, the client should set the iomode based on its intent to read or write the data. The client may default to an iomode of READ/WRITE (LAYOUTIOMODE_RW). The iomode need not be checked by the storage devices when clients perform I/O. However, the storage devices SHOULD still validate that the client holds a valid layout and return an error if the client does not. 9.4 Storage Device State Propagation Since the metadata server, which handles lock and open-mode state changes, as well as ACLs, may not be collocated with the storage devices where I/O access are validated, as such, the server implementation MUST take care of propagating changes of this state to the storage devices. Once the propagation to the storage devices is complete, the full effect of those changes must be in effect at the storage devices. However, some state changes need not be propagated immediately, although all changes SHOULD be propagated promptly. These state propagations have an impact on the design of the control protocol, even though the control protocol is outside of the scope of this specification. Immediate propagation refers to the synchronous propagation of state from the metadata server to the storage device(s); the propagation must be complete before returning to the client. 9.4.1 Lock State Propagation Mandatory locks MUST be made effective at the storage devices before the request that establishes them returns to the caller. Thus, mandatory lock state MUST be synchronously propagated to the storage devices. On the other hand, since advisory lock state is not used for checking I/O accesses at the storage devices, there is no semantic reason for propagating advisory lock state to the storage devices. However, since all lock, unlock, open downgrades and upgrades affect the sequence ID stored within the stateid, the stateid changes which may cause difficulty if this state is not propagated. Thus, when a client uses a stateid on a storage device Shepler Expires June 15, 2006 [Page 80] Internet-Draft NFSv4 Minior Version 1 December 2005 for I/O with a newer sequence number than the one the storage device has, the storage device should query the metadata server and get any pending updates to that stateid. This allows stateid sequence number changes to be propagated lazily, on-demand. [NOTE: With the reliance on the sessions protocol, there is no real need for sequence ID portion of the stateid to be validated on I/O accesses. It is proposed that the seq. ID checking is obsoleted.] Since updates to advisory locks neither confer nor remove privileges, these changes need not be propagated immediately, and may not need to be propagated promptly. The updates to advisory locks need only be propagated when the storage device needs to resolve a question about a stateid. In fact, if byte-range locking is not mandatory (i.e., is advisory) the clients are advised not to use the lock-based stateids for I/O at all. The stateids returned by open are sufficient and eliminate overhead for this kind of state propagation. 9.4.2 Open-mode Validation Open-mode validation MUST be performed against the open mode(s) held by the storage devices. However, the server implementation may not always require the immediate propagation of changes. Reduction in access because of CLOSEs or DOWNGRADEs do not have to be propagated immediately, but SHOULD be propagated promptly; whereas changes due to revocation MUST be propagated immediately. On the other hand, changes that expand access (e.g., new OPEN's and upgrades) don't have to be propagated immediately but the storage device SHOULD NOT reject a request because of mode issues without making sure that the upgrade is not in flight. 9.4.3 File Attributes Since the SETATTR operation has the ability to modify state that is visible on both the metadata and storage devices (e.g., the size), care must be taken to ensure that the resultant state across the set of storage devices is consistent; especially when truncating or growing the file. As described earlier, the LAYOUTCOMMIT operation is used to ensure that the metadata is synced with changes made to the storage devices. For the file-based protocol, it is necessary to re-sync state such as the size attribute, and the setting of mtime/atime. See Section 7.4 for a full description of the semantics regarding LAYOUTCOMMIT and attribute synchronization. It should be noted, that by using a file- based layout type, it is possible to synchronize this state before LAYOUTCOMMIT occurs. For example, the control protocol can be used to query the attributes present on the storage devices. Shepler Expires June 15, 2006 [Page 81] Internet-Draft NFSv4 Minior Version 1 December 2005 Any changes to file attributes that control authorization or access as reflected by ACCESS calls or READs and WRITEs on the metadata server, MUST be propagated to the storage devices for enforcement on READ and WRITE I/O calls. If the changes made on the metadata server result in more restrictive access permissions for any user, those changes MUST be propagated to the storage devices synchronously. Recall that the NFSv4 protocol [2] specifies that: ...since the NFS version 4 protocol does not impose any requirement that READs and WRITEs issued for an open file have the same credentials as the OPEN itself, the server still must do appropriate access checking on the READs and WRITEs themselves. This also includes changes to ACLs. The propagation of access right changes due to changes in ACLs may be asynchronous only if the server implementation is able to determine that the updated ACL is not more restrictive for any user specified in the old ACL. Due to the relative infrequency of ACL updates, it is suggested that all changes be propagated synchronously. [NOTE: it has been suggested that the NFSv4 specification is in error with regard to allowing principles other than those used for OPEN to be used for file I/O. If changes within a minor version alter the behavior of NFSv4 with regard to OPEN principals and stateids some access control checking at the storage device can be made less expensive. pNFS should be altered to take full advantage of these changes.] 9.5 Storage Device Component File Size A potential problem exists when a component data file on a particular storage device is grown past EOF; the problem exists for both dense and sparse layouts. Imagine the following scenario: a client creates a new file (size == 0) and writes to byte 128KB; the client then seeks to the beginning of the file and reads byte 100. The client should receive 0s back as a result of the read. However, if the read falls on a different storage device to the client's original write, the storage device servicing the READ may still believe that the file's size is at 0 and return no data with the EOF flag set. The storage device can only return 0s if it knows that the file's size has been extended. This would require the immediate propagation of the file's size to all storage devices, which is potentially very costly, instead, another approach as outlined below. First, the file's size is returned within the layout by LAYOUTGET. This size must reflect the latest size at the metadata server as set by the most recent of either the last LAYOUTCOMMIT or SETATTR; Shepler Expires June 15, 2006 [Page 82] Internet-Draft NFSv4 Minior Version 1 December 2005 however, it may be more recent. Second, if a client performs a read that is returned short (i.e., is fully within the file's size, but the storage device indicates EOF and returns partial or no data), the client must assume that it is a hole and substitute 0s for the data not read up until its known local file size. If a client extends the file, it must update its local file size. Third, if the metadata server receives a SETATTR of the size or a LAYOUTCOMMIT that alters the file's size, the metadata server must send out CB_SIZECHANGED messages with the new size to clients holding layouts; it need not send a notification to the client that performed the operation that resulted in the size changing). Upon reception of the CB_SIZECHANGED notification, clients must update their local size for that file. As well, if a new file size is returned as a result to LAYOUTCOMMIT, the client must update their local file size. 9.6 Crash Recovery Considerations As described in Section 7.7, the layout type specific storage protocol is responsible for handling the effects of I/Os started before lease expiration, extending through lease expiration. The NFSv4 file layout type prevents all I/Os from being executed after lease expiration, without relying on a precise client lease timer and without requiring storage devices to maintain lease timers. It works as follows. In the presence of sessions, each compound begins with a SEQUENCE operation that contains the "clientID". On the storage device, the clientID can be used to validate that the client has a valid layout for the I/O being performed, if it does not, the I/O is rejected. Before the metadata server takes any action to invalidate a layout given out by a previous instance, it must make sure that all layouts from that previous instance are invalidated at the storage devices. Note: it is sufficient to invalidate the stateids associated with the layout only if special stateids are not being used for I/O at the storage devices, otherwise the layout itself must be invalidated. This means that a metadata server may not restripe a file until it has contacted all of the storage devices to invalidate the layouts from the previous instance nor may it give out locks that conflict with locks embodied by the stateids associated with any layout from the previous instance without either doing a specific invalidation (as it would have to do anyway) or doing a global storage device invalidation. 9.7 Security Considerations The NFSv4 file layout type MUST adhere to the security considerations outlined in Section 8. More specifically, storage devices must make Shepler Expires June 15, 2006 [Page 83] Internet-Draft NFSv4 Minior Version 1 December 2005 all of the required access checks on each READ or WRITE I/O as determined by the NFSv4 protocol [2]. This impacts the control protocol and the propagation of state from the metadata server to the storage devices; see Section 9.4 for more details. 9.8 Alternate Approaches Two alternate approaches exist for file-based layouts and the method used by clients to obtain stateids used for I/O. Both approaches embed stateids within the layout. However, before examining these approaches it is important to understand the distinction between clients and owners. Delegations belong to clients, while locks (e.g., record and share reservations) are held by owners which in turn belong to a specific client. As such, delegations can only protect against inter-client conflicts, not intra-client conflicts. Layouts are held by clients and SHOULD NOT be associated with state held by owners. Therefore, if stateids used for data access are embedded within a layout, these stateids can only act as delegation stateids, protecting against inter-client conflicts; stateids pertaining to an owner can not be embedded within the layout. This has the implication that the client MUST arbitrate among all intra-client conflicts (e.g., arbitrating among lock requests by different processes) before issuing pNFS operations. Using the stateids stored within the layout, storage devices can only arbitrate between clients (not owners). The first alternate approach is to do away with global stateids, stateids returned by OPEN/LOCK that are valid on the metadata server and storage devices, and use only stateids embedded within the layout. This approach has the drawback that the stateids used for I/O access can not be validated against per owner state, since they are only associated with the client holding the layout. It breaks the semantics of tieing a stateid used for I/O to an open instance. This has the implication that clients must delegate per owner lock and open requests internally, rather than push the work onto the storage devices. The storage devices can still arbitrate and enforce inter-client lock and open state. The second approach is a hybrid approach. This approach allows for stateids to be embedded with the layout, but also allows for the possibility of global stateids. If the stateid embedded within the layout is a special stateid of all zeros, then the stateid referring to the last successful OPEN/LOCK should be used. This approach is recommended if it is decided that using NFSv4 as a control protocol is required. This proposal suggests the global stateid approach due to the cleaner Shepler Expires June 15, 2006 [Page 84] Internet-Draft NFSv4 Minior Version 1 December 2005 semantics it provides regarding the relationship between stateids used for I/O and their corresponding open instance or lock state. However, it does have a profound impact on the control protocol's implementation and the state propagation that is required (as described in Section 9.4). 10. pNFS Typed Data Structures 10.1 pnfs_layouttype4 enum pnfs_layouttype4 { LAYOUT_NFSV4_FILES = 1, LAYOUT_OSD2_OBJECTS = 2, LAYOUT_BLOCK_VOLUME = 3 }; A layout type specifies the layout being used. The implication is that clients have "layout drivers" that support one or more layout types. The file server advertises the layout types it supports through the LAYOUT_TYPES file system attribute. A client asks for layouts of a particular type in LAYOUTGET, and passes those layouts to its layout driver. The set of well known layout types must be defined. As well, a private range of layout types is to be defined by this document. This would allow custom installations to introduce new layout types. [OPEN ISSUE: Determine private range of layout types] New layout types must be specified in RFCs approved by the IESG before becoming part of the pNFS specification. The LAYOUT_NFSV4_FILES enumeration specifies that the NFSv4 file layout type is to be used. The LAYOUT_OSD2_OBJECTS enumeration specifies that the object layout, as defined in [7], is to be used. Similarly, the LAYOUT_BLOCK_VOLUME enumeration that the block/volume layout, as defined in [6], is to be used. 10.2 pnfs_deviceid4 typedef uint32_t pnfs_deviceid4; /* 32-bit device ID */ Layout information includes device IDs that specify a storage device through a compact handle. Addressing and type information is obtained with the GETDEVICEINFO operation. A client must not assume that device IDs are valid across metadata server reboots. The device ID is qualified by the layout type and are unique per file system (FSID). This allows different layout drivers to generate device IDs without the need for co-ordination. See Section 7.1.4 for more Shepler Expires June 15, 2006 [Page 85] Internet-Draft NFSv4 Minior Version 1 December 2005 details. 10.3 pnfs_deviceaddr4 struct pnfs_netaddr4 { string r_netid<>; /* network ID */ string r_addr<>; /* universal address */ }; struct pnfs_deviceaddr4 { pnfs_layouttype4 type; opaque device_addr<>; }; The device address is used to set up a communication channel with the storage device. Different layout types will require different types of structures to define how they communicate with storage devices. The opaque device_addr field must be interpreted based on the specified layout type. Currently, the only defined device address is that for the NFSv4 file layout (struct pnfs_netaddr4), which identifies a storage device by network IP address and port number. This is sufficient for the clients to communicate with the NFSv4 storage devices, and may also be sufficient for object-based storage drivers to communicate with OSDs. The other device address we expect to support is a SCSI volume identifier. The final protocol specification will detail the allowed values for device_type and the format of their associated location information. [NOTE: other device addresses will be added as the respective specifications mature. It has been suggested that a separate device_type enumeration is used as a switch to the pnfs_deviceaddr4 structure (e.g., if multiple types of addresses exist for the same layout type). Until such a time as a real case is made and the respective layout types have matured, the device address structure will be left as is.] 10.4 pnfs_devlist_item4 struct pnfs_devlist_item4 { pnfs_deviceid4 id; pnfs_deviceaddr4 addr; }; An array of these values is returned by the GETDEVICELIST operation. They define the set of devices associated with a file system. Shepler Expires June 15, 2006 [Page 86] Internet-Draft NFSv4 Minior Version 1 December 2005 10.5 pnfs_layout4 struct pnfs_layout4 { offset4 offset; length4 length; pnfs_layoutiomode4 iomode; pnfs_layouttype4 type; opaque layout<>; }; The pnfs_layout4 structure defines a layout for a file. The layout type specific data is opaque within this structure and must be interepreted based on the layout type. Currently, only the NFSv4 file layout type is defined; see Section 9.1 for its definition. Since layouts are sub-dividable, the offset and length together with the file's filehandle, the clientid, iomode, and layout type, identifies the layout. [OPEN ISSUE: there is a discussion of moving the striping information, or more generally the "aggregation scheme", up to the generic layout level. This creates a two-layer system where the top level is a switch on different data placement layouts, and the next level down is a switch on different data storage types. This lets different layouts (e.g., striping or mirroring or redundant servers) to be layered over different storage devices. This would move geometry information out of nfsv4_file_layouttype4 and up into a generic pnfs_striped_layout type that would specify a set of pnfs_deviceid4 and pnfs_devicetype4 to use for storage. Instead of nfsv4_file_layouttype4, there would be pnfs_nfsv4_devicetype4.] 10.6 pnfs_layoutupdate4 struct pnfs_layoutupdate4 { pnfs_layouttype4 type; opaque layoutupdate_data<>; }; The pnfs_layoutupdate4 structure is used by the client to return 'updated' layout information to the metadata server at LAYOUTCOMMIT time. This structure provides a channel to pass layout type specific information back to the metadata server. E.g., for block/volume layout types this could include the list of reserved blocks that were written. The contents of the opaque layoutupdate_data argument are determined by the layout type and are defined in their context. The NFSv4 file-based layout does not use this structure, thus the update_data field should have a zero length. Shepler Expires June 15, 2006 [Page 87] Internet-Draft NFSv4 Minior Version 1 December 2005 10.7 pnfs_layouthint4 struct pnfs_layouthint4 { pnfs_layouttype4 type; opaque layouthint_data<>; }; The pnfs_layouthint4 structure is used by the client to pass in a hint about the type of layout it would like created for a particular file. It is the structure specified by the FILE_LAYOUT_HINT attribute described below. The metadata server may ignore the hint, or may selectively ignore fields within the hint. This hint should be provided at create time as part of the initial attributes within OPEN. The NFSv4 file-based layout uses the "nfsv4_file_layouthint" structure as defined in Section 9.1. 10.8 pnfs_layoutiomode4 enum pnfs_layoutiomode4 { LAYOUTIOMODE_READ = 1, LAYOUTIOMODE_RW = 2, LAYOUTIOMODE_ANY = 3 }; The iomode specifies whether the client intends to read or write (with the possibility of reading) the data represented by the layout. The ANY iomode MUST NOT be used for LAYOUTGET, however, it can be used for LAYOUTRETURN and LAYOUTRECALL. The ANY iomode specifies that layouts pertaining to both READ and RW iomodes are being returned or recalled, respectively. The metadata server's use of the iomode may depend on the layout type being used. The storage devices may validate I/O accesses against the iomode and reject invalid accesses. 11. pNFS File Attributes 11.1 pnfs_layouttype4<> FS_LAYOUT_TYPES This attribute applies to a file system and indicates what layout types are supported by the file system. We expect this attribute to be queried when a client encounters a new fsid. This attribute is used by the client to determine if it has applicable layout drivers. 11.2 pnfs_layouttype4<> FILE_LAYOUT_TYPES This attribute indicates the particular layout type(s) used for a file. This is for informational purposes only. The client needs to use the LAYOUTGET operation in order to get enough information (e.g., Shepler Expires June 15, 2006 [Page 88] Internet-Draft NFSv4 Minior Version 1 December 2005 specific device information) in order to perform I/O. 11.3 pnfs_layouthint4 FILE_LAYOUT_HINT This attribute may be set on newly created files to influence the metadata server's choice for the file's layout. It is suggested that this attribute is set as one of the initial attributes within the OPEN call. The metadata server may ignore this attribute. This attribute is a sub-set of the layout structure returned by LAYOUTGET. For example, instead of specifying particular devices, this would be used to suggest the stripe width of a file. It is up to the server implementation to determine which fields within the layout it uses. [OPEN ISSUE: it has been suggested that the HINT is a well defined type other than pnfs_layoutdata4, similar to pnfs_layoutupdate4.] 11.4 uint32_t FS_LAYOUT_PREFERRED_BLOCKSIZE This attribute is a file system wide attribute and indicates the preferred block size for direct storage device access. 11.5 uint32_t FS_LAYOUT_PREFERRED_ALIGNMENT This attribute is a file system wide attribute and indicates the preferred alignment for direct storage device access. 12. pNFS Error Definitions NFS4ERR_BADLAYOUT Layout specified is invalid. NFS4ERR_BADIOMODE Layout iomode is invalid. NFS4ERR_LAYOUTUNAVAILABLE Layouts are not available for the file or its containing file system. NFS4ERR_LAYOUTTRYLATER Layouts are temporarily unavailable for the file, client should retry later. NFS4ERR_NOMATCHING_LAYOUT Client has no matching layout (segment) to return. NFS4ERR_RECALLCONFLICT Layout is unavailable due to a conflicting LAYOUTRECALL that is in progress. NFS4ERR_UNKNOWN_LAYOUTTYPE Layout type is unknown. Shepler Expires June 15, 2006 [Page 89] Internet-Draft NFSv4 Minior Version 1 December 2005 13. Layouts and Aggregation This section describes several aggregation schemes in a semi-formal way to provide context for layout formats. These definitions will be formalized in other protocols. However, the set of understood types is part of this protocol in order to provide for basic interoperability. The layout descriptions include (deviceID, objectID) tuples that identify some storage object on some storage device. The addressing formation associated with the deviceID is obtained with GETDEVICEINFO. The interpretation of the objectID depends on the storage protocol. The objectID could be a filehandle for an NFSv4 storage device. It could be a OSD object ID for an object server. The layout for a block device generally includes additional block map information to enumerate blocks or extents that are part of the layout. 13.1 Simple Map The data is located on a single storage device. In this case the file server can act as the front end for several storage devices and distribute files among them. Each file is limited in its size and performance characteristics by a single storage device. The simple map consists of (deviceID, objectID). 13.2 Block Extent Map The data is located on a LUN in the SAN. The layout consists of an array of (deviceID, blockID, offset, length) tuples. Each entry describes a block extent. 13.3 Striped Map (RAID 0) The data is striped across storage devices. The parameters of the stripe include the number of storage devices (N) and the size of each stripe unit (U). A full stripe of data is N * U bytes. The stripe map consists of an ordered list of (deviceID, objectID) tuples and the parameter value for U. The first stripe unit (the first U bytes) are stored on the first (deviceID, objectID), the second stripe unit on the second (deviceID, objectID) and so forth until the first complete stripe. The data layout then wraps around so that byte (N*U) of the file is stored on the first (deviceID, objectID) in the list, but starting at offset U within that object. The striped layout allows a client to read or write to the component objects in parallel to achieve high bandwidth. The striped map for a block device would be slightly different. The Shepler Expires June 15, 2006 [Page 90] Internet-Draft NFSv4 Minior Version 1 December 2005 map is an ordered list of (deviceID, blockID, blocksize), where the deviceID is rotated among a set of devices to achieve striping. 13.4 Replicated Map The file data is replicated on N storage devices. The map consists of N (deviceID, objectID) tuples. When data is written using this map, it should be written to N objects in parallel. When data is read, any component object can be used. This map type is controversial because it highlights the issues with error recovery. Those issues get interesting with any scheme that employs redundancy. The handling of errors (e.g., only a subset of replicas get updated) is outside the scope of this protocol extension. Instead, it is a function of the storage protocol and the metadata control protocol. 13.5 Concatenated Map The map consists of an ordered set of N (deviceID, objectID, size) tuples. Each successive tuple describes the next segment of the file. 13.6 Nested Map The nested map is used to compose more complex maps out of simpler ones. The map format is an ordered set of M sub-maps, each submap applies to a byte range within the file and has its own type such as the ones introduced above. Any level of nesting is allowed in order to build up complex aggregation schemes. 14. NFSv4.1 Operations 14.1 LOOKUPP - Lookup Parent Directory If the NFSv4 minor version is 1, then following replaces section 14.2.14 of the NFSv4.0 specification. The LOOKUPP operation's "over the wire" format is not altered, but the semantics are slightly modified to account for the addition of SECINFO_NO_NAME. SYNOPSIS (cfh) -> (cfh) Shepler Expires June 15, 2006 [Page 91] Internet-Draft NFSv4 Minior Version 1 December 2005 ARGUMENT /* CURRENT_FH: object */ void; RESULT struct LOOKUPP4res { /* CURRENT_FH: directory */ nfsstat4 status; }; DESCRIPTION The current filehandle is assumed to refer to a regular directory or a named attribute directory. LOOKUPP assigns the filehandle for its parent directory to be the current filehandle. If there is no parent directory an NFS4ERR_NOENT error must be returned. Therefore, NFS4ERR_NOENT will be returned by the server when the current filehandle is at the root or top of the server's file tree. As for LOOKUP, LOOKUPP will also cross mountpoints. If the current filehandle is not a directory or named attribute directory, the error NFS4ERR_NOTDIR is returned. If the requester's security flavor does not match that configured for the parent directory, then the server SHOULD return NFS4ERR_WRONGSEC (a future minor revision of NFSv4 may upgrade this to MUST) in the LOOKUPP response. However, if the server does so, it MUST support the new SECINFO_NO_NAME operation, so that the client can gracefully determine the correct security flavor. See the discussion of the SECINFO_NO_NAME operation for a description. ERRORS NFS4ERR_ACCESS NFS4ERR_BADHANDLE NFS4ERR_FHEXPIRED NFS4ERR_IO NFS4ERR_MOVED NFS4ERR_NOENT NFS4ERR_NOFILEHANDLE NFS4ERR_NOTDIR NFS4ERR_RESOURCE NFS4ERR_SERVERFAULT NFS4ERR_STALE NFS4ERR_WRONGSEC Shepler Expires June 15, 2006 [Page 92] Internet-Draft NFSv4 Minior Version 1 December 2005 14.2 SECINFO -- Obtain Available Security If the NFSv4 minor version is 1, then following replaces section 14.2.31 of the NFSv4.0 specification. The SECINFO operation's "over the wire" format is not altered, but the semantics are slightly modified to account for the addition of SECINFO_NO_NAME. SYNOPSIS (cfh), name -> { secinfo } ARGUMENT struct SECINFO4args { /* CURRENT_FH: directory */ component4 name; }; RESULT Shepler Expires June 15, 2006 [Page 93] Internet-Draft NFSv4 Minior Version 1 December 2005 enum rpc_gss_svc_t {/* From RFC 2203 */ RPC_GSS_SVC_NONE = 1, RPC_GSS_SVC_INTEGRITY = 2, RPC_GSS_SVC_PRIVACY = 3 }; struct rpcsec_gss_info { sec_oid4 oid; qop4 qop; rpc_gss_svc_t service; }; union secinfo4 switch (uint32_t flavor) { case RPCSEC_GSS: rpcsec_gss_info flavor_info; default: void; }; typedef secinfo4 SECINFO4resok<>; union SECINFO4res switch (nfsstat4 status) { case NFS4_OK: SECINFO4resok resok4; default: void; }; DESCRIPTION The SECINFO operation is used by the client to obtain a list of valid RPC authentication flavors for a specific directory filehandle, file name pair. SECINFO should apply the same access methodology used for LOOKUP when evaluating the name. Therefore, if the requester does not have the appropriate access to LOOKUP the name then SECINFO must behave the same way and return NFS4ERR_ACCESS. The result will contain an array which represents the security mechanisms available, with an order corresponding to the server's preferences, the most preferred being first in the array. The client is free to pick whatever security mechanism it both desires and supports, or to pick in the server's preference order the first one it supports. The array entries are represented by the secinfo4 structure. The field 'flavor' will contain a value of AUTH_NONE, AUTH_SYS (as defined in [RFC1831]), or RPCSEC_GSS (as defined in [RFC2203]). The field flavor can also any other security flavor registered with IANA. Shepler Expires June 15, 2006 [Page 94] Internet-Draft NFSv4 Minior Version 1 December 2005 For the flavors AUTH_NONE and AUTH_SYS, no additional security information is returned. The same is true of many (if not most) other security flavors, including AUTH_DH. For a return value of RPCSEC_GSS, a security triple is returned that contains the mechanism object id (as defined in [RFC2743]), the quality of protection (as defined in [RFC2743]) and the service type (as defined in [RFC2203]). It is possible for SECINFO to return multiple entries with flavor equal to RPCSEC_GSS with different security triple values. On success, the current filehandle retains its value. If the name has a length of 0 (zero), or if name does not obey the UTF-8 definition, the error NFS4ERR_INVAL will be returned. IMPLEMENTATION The SECINFO operation is expected to be used by the NFS client when the error value of NFS4ERR_WRONGSEC is returned from another NFS operation. This signifies to the client that the server's security policy is different from what the client is currently using. At this point, the client is expected to obtain a list of possible security flavors and choose what best suits its policies. As mentioned, the server's security policies will determine when a client request receives NFS4ERR_WRONGSEC. The operations which may receive this error are: LINK, LOOKUP, LOOKUPP, OPEN, PUTFH, PUTPUBFH, PUTROOTFH, RESTOREFH, RENAME, and indirectly READDIR. LINK and RENAME will only receive this error if the security used for the operation is inappropriate for saved filehandle. With the exception of READDIR, these operations represent the point at which the client can instantiate a filehandle into the "current filehandle" at the server. The filehandle is either provided by the client (PUTFH, PUTPUBFH, PUTROOTFH) or generated as a result of a name to filehandle translation (LOOKUP and OPEN). RESTOREFH is different because the filehandle is a result of a previous SAVEFH. Even though the filehandle, for RESTOREFH, might have previously passed the server's inspection for a security match, the server will check it again on RESTOREFH to ensure that the security policy has not changed. If the client wants to resolve an error return of NFS4ERR_WRONGSEC, the following will occur: * For LOOKUP and OPEN, the client will use SECINFO with the same current filehandle and name as provided in the original LOOKUP Shepler Expires June 15, 2006 [Page 95] Internet-Draft NFSv4 Minior Version 1 December 2005 or OPEN to enumerate the available security triples. * For LINK, PUTFH, PUTROOTFH, PUTPUBFH, RENAME, and RESTOREFH, the client will use SECINFO_NO_NAME { style = current_fh }. The client will prefix the SECINFO_NO_NAME operation with the appropriate PUTFH, PUTPUBFH, or PUTROOTFH operation that provides the file handled originally provided by the PUTFH, PUTPUBFH, PUTROOTFH, or RESTOREFH, or for the failed LINK or RENAME, the SAVEFH. * ========================================================= NOTE: In NFSv4.0, the client was required to use SECINFO, and had to reconstruct the parent of the original file handle, and the component name of the original filehandle. ======================================================== * For LOOKUPP, the client will use SECINFO_NO_NAME { style = parent } and provide the filehandle with equals the filehandle originally provided to LOOKUPP. The READDIR operation will not directly return the NFS4ERR_WRONGSEC error. However, if the READDIR request included a request for attributes, it is possible that the READDIR request's security triple did not match that of a directory entry. If this is the case and the client has requested the rdattr_error attribute, the server will return the NFS4ERR_WRONGSEC error in rdattr_error for the entry. See the section "Security Considerations" for a discussion on the recommendations for security flavor used by SECINFO and SECINFO_NO_NAME. ERRORS 14.3 SECINFO_NO_NAME - Get Security on Unnamed Object Obtain available security mechanisms with the use of the parent of an object or the current filehandle. SYNOPSIS (cfh), secinfo_style -> { secinfo } Shepler Expires June 15, 2006 [Page 96] Internet-Draft NFSv4 Minior Version 1 December 2005 ARGUMENT enum secinfo_style_4 { current_fh = 0, parent = 1 }; typedef secinfo_style_4 SECINFO_NO_NAME4args; RESULT typedef SECINFO4res SECINFO_NO_NAME4res; DESCRIPTION Like the SECINFO operation, SECINFO_NO_NAME is used by the client to obtain a list of valid RPC authentication flavors for a specific file object. Unlike SECINFO, SECINFO_NO_NAME only works with objects are accessed by file handle. There are two styles of SECINFO_NO_NAME, as determined by the value of the secinfo_style_4 enumeration. If "current_fh" is passed, then SECINFO_NO_NAME is querying for the required security for the current filehandle. If "parent" is passed, then SECINFO_NO_NAME is querying for the required security of the current filehandles's parent. If the style selected is "parent", then SECINFO should apply the same access methodology used for LOOKUPP when evaluating the traversal to the parent directory. Therefore, if the requester does not have the appropriate access to LOOKUPP the parent then SECINFO_NO_NAME must behave the same way and return NFS4ERR_ACCESS. Note that if PUTFH, PUTPUBFH, or PUTROOTFH return NFS4ERR_WRONGSEC, this is tantamount to the server asserting that the client will have to guess what the required security is, because there is no way to query. Therefore, the client must iterate through the security triples available at the client and reattempt the PUTFH, PUTROOTFH or PUTPUBFH operation. In the unfortunate event none of the MANDATORY security triples are supported by the client and server, the client SHOULD try using others that support integrity. Failing that, the client can try using other forms (e.g. AUTH_SYS and AUTH_NONE), but because such forms lack integrity checks, this puts the client at risk. Shepler Expires June 15, 2006 [Page 97] Internet-Draft NFSv4 Minior Version 1 December 2005 The server implementor should pay particular attention to the section "Clarification of Security Negotiation in NFSv4.1" for implementation suggestions for avoiding NFS4ERR_WRONGSEC error returns from PUTFH, PUTROOTFH or PUTPUBFH. Everything else about SECINFO_NO_NAME is the same as SECINFO. See the previous discussion on SECINFO. IMPLEMENTATION See the previous dicussion on SECINFO. ERRORS NFS4ERR_ACCESS NFS4ERR_BADCHAR NFS4ERR_BADHANDLE NFS4ERR_BADNAME NFS4ERR_BADXDR NFS4ERR_FHEXPIRED NFS4ERR_INVAL NFS4ERR_MOVED NFS4ERR_NAMETOOLONG NFS4ERR_NOENT NFS4ERR_NOFILEHANDLE NFS4ERR_NOTDIR NFS4ERR_RESOURCE NFS4ERR_SERVERFAULT NFS4ERR_STALE 14.4 CREATECLIENTID - Instantiate Clientid Create a clientid SYNOPSIS client -> clientid ARGUMENT struct CREATECLIENTID4args { nfs_client_id4 clientdesc; }; RESULT Shepler Expires June 15, 2006 [Page 98] Internet-Draft NFSv4 Minior Version 1 December 2005 struct CREATECLIENTID4resok { clientid4 clientid; verifier4 clientid_confirm; }; union SETCLIENTID4res switch (nfsstat4 status) { case NFS4_OK: CREATECLIENTID4resok resok4; case NFS4ERR_CLID_INUSE: void; default: void; }; DESCRIPTION The client uses the CREATECLIENTID operation to register a particular client identifier with the server. The clientid returned from this operation will be necessary for requests that create state on the server and will serve as a parent object to sessions created by the client. In order to verify the clientid it must first be used as an argument to CREATESESSION. IMPLEMENTATION A server's client record is a 5-tuple: 1. clientdesc.id: The long form client identifier, sent via the client.id subfield of the CREATECLIENTID4args structure 2. clientdesc.verifier: A client-specific value used to indicate reboots, sent via the clientdesc.verifier subfield of the CREATECLIENTID4args structure 3. principal: The RPCSEC_GSS principal sent via the RPC headers 4. clientid: The shorthand client identifier, generated by the server and returned via the clientid field in the Shepler Expires June 15, 2006 [Page 99] Internet-Draft NFSv4 Minior Version 1 December 2005 CREATECLIENTID4resok structure 5. confirmed: A private field on the server indicating whether or not a client record has been confirmed. A client record is confirmed if there has been a successful CREATESESSION operation to confirm it. Otherwise it is unconfirmed. An unconfirmed record is established by a CREATECLIENTID call. Any unconfirmed record that is not confirmed within a lease period may be removed. The following identifiers represent special values for the fields in the records. id_arg: The value of the clientdesc.id subfield of the CREATECLIENTID4args structure of the current request. verifier_arg: The value of the clientdesc.verifier subfield of the CREATECLIENTID4args structure of the current request. old_verifier_arg: A value of the clientdesc.verifier field of a client record received in a previous request; this is distinct from verifier_arg. principal_arg: The value of the RPCSEC_GSS principal for the current request. old_principal_arg: A value of the RPCSEC_GSS principal received for a previous request. This is distinct from principal_arg. clientid_ret: The value of the clientid field the server will return in the CREATECLIENTID4resok structure for the current request. Shepler Expires June 15, 2006 [Page 100] Internet-Draft NFSv4 Minior Version 1 December 2005 old_clientid_ret: The value of the clientid field the server returned in the CREATECLIENTID4resok structure for a previous request. This is distinct from clientid_ret. Since CREATECLIENTID is a non-idempotent operation, we must consider the possibility that replays may occur as a result of a client reboot, network partition, malfunctioning router, etc. Replays are identified by the value of the client field of CREATECLIENTID4args and the method for dealing with them is outlined in the scenarios below. The scenarios are described in terms of what client records whose clientdesc.id subfield have value equal to id_arg exist in the server's set of client records. Any cases in which there is more than one record with identical values for id_arg represent a server implementation error. Operation in the potential valid cases is summarized as follows. 1. Common case If no client records with clientdesc.id matching id_arg exist, a new shorthand client identifier clientid_ret is generated, and the following unconfirmed record is added to the server's state. { id_arg, verifier_arg, principal_arg, clientid_ret, FALSE } Subsequently, the server returns clientid_ret. 2. Router Replay If the server has the following confirmed record, then this request is likely the result of a replayed request due to a faulty router or lost connection. { id_arg, verifier_arg, principal_arg, clientid_ret, TRUE } Since the record has been confirmed, the client must have received the server's reply from the initial CREATECLIENTID request. Since this is simply a spurious request, there is no modification to the server's state, and the server makes no reply to the client. Shepler Expires June 15, 2006 [Page 101] Internet-Draft NFSv4 Minior Version 1 December 2005 3. Client Collision If the server has the following confirmed record, then this request is likely the result of a chance collision between the values of the clientdesc.id subfield of CREATECLIENTID4args for two different clients. { id_arg, *, old_principal_arg, clientid_ret, TRUE } Since the value of the clientdesc.id subfield of each client record must be unique, there is no modification of the server's state, and NFS4ERR_CLID_INUSE is returned to indicate the client should retry with a different value for the clientdesc.id subfield of CREATECLIENTID4args. This scenario may also represent a malicious attempt to destroy a client's state on the server. For security reasons, the server MUST NOT remove the client's state when there is a principal mismatch. 4. Replay If the server has the following unconfirmed record then this request is likely the result of a client replay due to a network partition or some other connection failure. { id_arg, verifier_arg, principal_arg, clientid_ret, FALSE } Since the response to the CREATECLIENTID request that created this record may have been lost, it is not acceptable to drop this duplicate request. However, rather than processing it normally, the existing record is left unchanged and clientid_ret, which was generated for the previous request, is returned. 5. Change of Principal If the server has the following unconfirmed record then this request is likely the result of a client which has for whatever reasons changed principals (possibly to change security flavor) after calling CREATECLIENTID, but before calling CREATESESSION. { id_arg, verifier_arg, old_principal_arg, clientid_ret, FALSE} Shepler Expires June 15, 2006 [Page 102] Internet-Draft NFSv4 Minior Version 1 December 2005 Since the client has not changed, the principal field of the unconfirmed record is updated to principal_arg and clientid_ret is again returned. There is a small possibility that this is merely a collision on the client field of CREATECLIENTID4args between unrelated clients, but since that is unlikely, and an unconfirmed record does not generally have any filesystem pertinent state, we can assume it is the same client without risking loss of any important state. After processing, the following record will exist on the server. { id_arg, verifier_arg, principal_arg, clientid_ret, FALSE} 6. Client Reboot If the server has the following confirmed client record, then this request is likely from a previously confirmed client which has rebooted. { id_arg, old_verifier_arg, principal_arg, clientid_ret, TRUE } Since the previous incarnation of the same client will no longer be making requests, lock and share reservations should be released immediately rather than forcing the new incarnation to wait for the lease time on the previous incarnation to expire. Furthermore, session state should be removed since if the client had maintained that information across reboot, this request would not have been issued. If the server does not support the CLAIM_DELEGATE_PREV claim type, associated delegations should be purged as well; otherwise, delegations are retained and recovery proceeds according to RFC3530. The client record is updated with the new verifier and its status is changed to unconfirmed. After processing, clientid_ret is returned to the client and the following record will exist on the server. { id_arg, verifier_arg, principal_arg, clientid_ret, FALSE } 7. Reboot before confirmation If the server has the following unconfirmed record, then this request is likely from a client which rebooted before Shepler Expires June 15, 2006 [Page 103] Internet-Draft NFSv4 Minior Version 1 December 2005 sending a CREATESESSION request. { id_arg, old_verifier_arg, *, clientid_ret, FALSE } Since this is believed to be a request from a new incarnation of the original client, the server updates the value of clientdesc.verifier and returns the original clientid_ret. After processing, the following state exists on the server. { id_arg, verifier_arg, *, clientid_ret, FALSE } ERRORS NFS4ERR_BADXDR NFS4ERR_CLID_INUSE NFS4ERR_INVAL NFS4ERR_RESOURCE NFS4ERR_SERVERFAULT 14.5 CREATESESSION - Create New Session and Confirm Clientid Start up session and confirm clientid. SYNOPSIS clientid, session_args -> sessionid, session_args ARGUMENT Shepler Expires June 15, 2006 [Page 104] Internet-Draft NFSv4 Minior Version 1 December 2005 struct CREATESESSION4args { clientid4 clientid; bool persist; count4 maxrequestsize; count4 maxresponsesize; count4 maxrequests; count4 headerpadsize; switch (bool clientid_confirm) { case TRUE: verifier4 setclientid_confirm; case FALSE: void; } switch (channelmode4 mode) { case DEFAULT: void; case STREAM: streamchannelattrs4 streamchanattrs; case RDMA: rdmachannelattrs4 rdmachanattrs; }; }; RESULT Shepler Expires June 15, 2006 [Page 105] Internet-Draft NFSv4 Minior Version 1 December 2005 typedef opaque sessionid4[16]; struct CREATESESSION4resok { sessionid4 sessionid; bool persist; count4 maxrequestsize; count4 maxresponsesize; count4 maxrequests; count4 headerpadsize; switch (channelmode4 mode) { case DEFAULT: void; case STREAM: streamchannelattrs4 streamchanattrs; case RDMA: rdmachannelattrs4 rdmachanattrs; }; }; union CREATESESSION4res switch (nfsstat4 status) { case NFS4_OK: CREATESESSION4resok resok4; default: void; }; DESCRIPTION This operation is used by the client to create new session objects on the server. Additionally the first session created with a new shorthand client identifier serves to confirm the creation of that client's state on the server. The server returns the parameter values for the new session. IMPLEMENTATION To describe the implementation, the same notation for client records introduced in the description of CREATECLIENTID is used with the following addition. clientid_arg: The value of the clientid field of the CREATESESSION4args structure of the current request. Since CREATESESSION is a non-idempotent operation, we must consider the possibility that replays may occur as a result of a client reboot, network partition, malfunctioning router, etc. Replays are identified by the value of the clientid and sessionid fields of CREATESESSION4args and the method for dealing with them Shepler Expires June 15, 2006 [Page 106] Internet-Draft NFSv4 Minior Version 1 December 2005 is outlined in the scenarios below. The processing of this operation is divided into two phases: clientid confirmation and session creation. In case the state for the provided clientid has not been verified, it is confirmed before the session is created. Otherwise the clientid confirmation phase is skipped and only the session creation phase occurs. Note that since only confirmed clients may create sessions, the clientid confirmation stage does not depend upon sessionid_arg. CLIENTID CONFIRMATION The operational cases are described in terms of what client records whose clientid field have value equal to clientid_arg exist in the server's set of client records. Any cases in which there is more than one record with identical values for clientid represent a server implementation error. Operation in the potential valid cases is summarized as follows. 1. Common Case If the server has the following unconfirmed record, then this is the expected confirmation of an unconfirmed record. { *, *, principal_arg, clientid_arg, FALSE } The confirmed field of the record is set to TRUE and processing of the operation continues normally. 2. Stale Clientid If the server contains no records with clientid equal to clientid_arg, then most likely the client's state has been purged during a period of inactivity, possibly due to a loss of connectivity. NFS4ERR_STALE_CLIENTID is returned, and no changes are made to any client records on the server. 3. Principal Change or Collision If the server has the following record, then the client has changed principals after the previous CREATECLIENTID request, or there has been a chance collision between shortand client identifiers. Shepler Expires June 15, 2006 [Page 107] Internet-Draft NFSv4 Minior Version 1 December 2005 { *, *, old_principal_arg, clientid_arg, * } Neither of these cases are permissible. Processing stops and NFS4ERR_CLID_INUSE is returned to the client. No changes are made to any client records on the server. SESSION CREATION To determine whether this request is a replay, the server examines the sessionid argument provided by the client. If the sessionid matches the identifier of a previously created session, then this request must be interpreted as a replay. No new state is created and a reply with the parameters of the existing session is returned to the client. If a session corresponding to the sessionid does not already exist, then the request is not a replay and is processed as follows. NOTE: It is the responsibility of the client to generate appropriate values for sessionid. Since the ordering of messages sent on different transport connections is not guaranteed, immediately reusing the sessionid of a previously destroyed session may yield unpredictable results. Client implementations should avoid recently used sessionids to ensure correct behavior. The server examines the persist, maxrequestsize, maxresponsesize, maxrequests and headerpadsize arguments. For each argument, if the value is acceptable to the server, it is recommended that the server use the provided value to create the new session. If it is not acceptable, the server may use a different value, but must return the value used to the client. These parameters have the following interpretation. persist: True if the client desires server support for "reliable" semantics. For sessions in which only idempotent operations will be used (e.g. a read-only session), clients should set this value to false. If the server does not or cannot provide "reliable" semantics this value must be set to false on return. maxrequestsize: The maximum size of a COMPOUND request that will be sent by the client including RPC headers. Shepler Expires June 15, 2006 [Page 108] Internet-Draft NFSv4 Minior Version 1 December 2005 maxresponsesize: The maximum size of a COMPOUND reply that the client will accept from the server including RPC headers. The server must not increase the value of this parameter. If a client sends a COMPOUND request for which the size of the reply would exceed this value, the server will return NFS4ERR_RESOURCE. maxrequests: The maximum number of concurrent COMPOUND requests that the client will issue on the session. Subsequent COMPOUND requests will each be assigned a slot identifier by the client on the range 0 to maxrequests - 1 inclusive. A slot id cannot be reused until the previous request on that slot has completed. headerpadsize: The maximum amount of padding the client is willing to apply to ensure that write payloads are aligned on some boundary at the server. The server should reply with its preferred value, or zero if padding is not in use. The server may decrease this value but must not increase it. The server creates the session by recording the parameter values used and if the persist parameter is true and has been accepted by the server, allocating space for the duplicate request cache (DRC). If the session state is created successfully, the server associates it with the session identifier provided by the client. This identifier must be unique among the client's active sessions but there is no need for it to be globally unique. Finally, the server returns the negotiated values used to create the session to the client. ERRORS NFS4ERR_BADXDR NFS4ERR_CLID_INUSE NFS4ERR_RESOURCE NFS4ERR_SERVERFAULT NFS4ERR_STALE_CLIENTID 14.6 BIND_BACKCHANNEL - Create a callback channel binding Establish a callback channel on the connection. Shepler Expires June 15, 2006 [Page 109] Internet-Draft NFSv4 Minior Version 1 December 2005 SYNOPSIS ARGUMENT struct BIND_BACKCHANNEL4args { clientid4 clientid; uint32_t callback_program; uint32_t callback_ident; count4 maxrequestsize; count4 maxresponsesize; count4 maxrequests; switch (channelmode4 mode) { case DEFAULT: void; case STREAM: streamchannelattrs4 streamchanattrs; case RDMA: rdmachannelattrs4 rdmachanattrs; }; }; RESULT Shepler Expires June 15, 2006 [Page 110] Internet-Draft NFSv4 Minior Version 1 December 2005 struct BIND_BACKCHANNEL4resok { count4 maxrequestsize; count4 maxresponsesize; count4 maxrequests; switch (channelmode4 mode) { case DEFAULT: void; case STREAM: streamchannelattrs4 streamchanattrs; case RDMA: rdmachannelattrs4 rdmachanattrs; }; }; union BIND_BACKCHANNEL4res switch (nfsstat4 status) { case NFS4_OK: BIND_BACKCHANNEL4resok resok4; default: void; }; DESCRIPTION The BIND_BACKCHANNEL operation serves to establish the current connection as a designated callback channel for the specified session. Normally, only one callback channel is bound, however if more than one are established, they are used at the server's prerogative, no affinity or preference is specified by the client. The arguments and results of the BIND_BACKCHANNEL call are a subset of the session parameters, and used identically to those values on the callback channel only. However, not all session operation channel parameters are relevant to the callback channel, for example header padding (since writes of bulk data are not performed in callbacks). IMPLEMENTATION No discussion at this time. ERRORS TBD Shepler Expires June 15, 2006 [Page 111] Internet-Draft NFSv4 Minior Version 1 December 2005 14.7 DESTROYSESSION - Destroy existing session Destroy existing session. SYNOPSIS void -> status ARGUMENT struct DESTROYSESSION4args { sessionid4 sessionid; }; RESULT struct SESSION_DESTROYres { nfsstat status; }; DESCRIPTION The SESSION_DESTROY operation closes the session and discards any active state such as locks, leases, and server duplicate request cache entries. Any remaining connections bound to the session are immediately unbound and may additionally be closed by the server. This operation must be the final, or only operation in any request. Because the operation results in destruction of the session, any duplicate request caching for this request, as well as previously completed requests, will be lost. For this reason, it is advisable to not place this operation in a request with other state-modifying operations. In addition, a SEQUENCE operation is not required in the request. Note that because the operation will never be replayed by the server, a client that retransmits the request may receive an error in response, even though the session may have been successfully destroyed. Shepler Expires June 15, 2006 [Page 112] Internet-Draft NFSv4 Minior Version 1 December 2005 IMPLEMENTATION No discussion at this time. ERRORS TBD 14.8 SEQUENCE - Supply per-procedure sequencing and control Supply per-procedure sequencing and control SYNOPSIS control -> control ARGUMENT typedef uint32_t sequenceid4; typedef uint32_t slotid4; struct SEQUENCE4args { clientid4 clientid; sessionid4 sessionid; sequenceid4 sequenceid; slotid4 slotid; slotid4 maxslot; }; RESULT Shepler Expires June 15, 2006 [Page 113] Internet-Draft NFSv4 Minior Version 1 December 2005 struct SEQUENCE4resok { clientid4 clientid; sessionid4 sessionid; sequenceid4 sequenceid; slotid4 slotid; slotid4 maxslot; slotid4 target_maxslot; }; union SEQUENCE4res switch (nfsstat4 status) { case NFS4_OK: SEQUENCE4resok resok4; default: void; }; DESCRIPTION The SEQUENCE operation is used to manage operational accounting for the session on which the operation is sent. The contents include the client and session to which this request belongs, slotid and sequenceid, used by the server to implement session request control and the duplicate reply cache semantics, and exchanged slot counts which are used to adjust these values. This operation must appear once as the first operation in each COMPOUND sent after the channel is successfully bound, or a protocol error must result. IMPLEMENTATION No discussion at this time. ERRORS NFS4ERR_BADSESSION NFS4ERR_BADSLOT 14.9 CB_RECALLCREDIT - change flow control limits Change flow control limits SYNOPSIS targetcount -> status Shepler Expires June 15, 2006 [Page 114] Internet-Draft NFSv4 Minior Version 1 December 2005 ARGUMENT struct CB_RECALLCREDIT4args { sessionid4 sessionid; uint32_t target; }; RESULT struct CB_RECALLCREDIT4res { nfsstat4 status; }; DESCRIPTION The CB_RECALLCREDIT operation requests the client to return session and transport credits to the server, by zero-length RDMA Sends or NULL NFSv4 operations. IMPLEMENTATION No discussion at this time. ERRORS NONE 14.10 CB_SEQUENCE - Supply callback channel sequencing and control Sequence and control SYNOPSIS control -> control ARGUMENT Shepler Expires June 15, 2006 [Page 115] Internet-Draft NFSv4 Minior Version 1 December 2005 typedef uint32_t sequenceid4; typedef uint32_t slotid4; struct CB_SEQUENCE4args { clientid4 clientid; sessionid4 sessionid; sequenceid4 sequenceid; slotid4 slotid; slotid4 maxslot; }; RESULT struct CB_SEQUENCE4resok { clientid4 clientid; sessionid4 sessionid; sequenceid4 sequenceid; slotid4 slotid; slotid4 maxslot; slotid4 target_maxslot; }; union CB_SEQUENCE4res switch (nfsstat4 status) { case NFS4_OK: CB_SEQUENCE4resok resok4; default: void; }; DESCRIPTION The CB_SEQUENCE operation is used to manage operational accounting for the callback channel of the session on which the operation is sent. The contents include the client and session to which this request belongs, slotid and sequenceid, used by the server to implement session request control and the duplicate reply cache semantics, and exchanged slot counts which are used to adjust these values. This operation must appear once as the first operation in each CB_COMPOUND sent after the callback channel is successfully bound, or a protocol error must result. IMPLEMENTATION Shepler Expires June 15, 2006 [Page 116] Internet-Draft NFSv4 Minior Version 1 December 2005 No discussion at this time. ERRORS NFS4ERR_BADSESSION NFS4ERR_BADSLOT 14.11 GET_DIR_DELEGATION - Get a directory delegation Obtain a directory delegation. SYNOPSIS (cfh), requested notification -> (cfh), cookieverf, stateid, supported notification ARGUMENT struct GET_DIR_DELEGATION4args { dir_notification_type4 notification_type; attr_notice4 child_attr_delay; attr_notice4 dir_attr_delay; }; /* * Notification types. */ const DIR_NOTIFICATION_NONE = 0x00000000; const DIR_NOTIFICATION_CHANGE_CHILD_ATTRIBUTES = 0x00000001; const DIR_NOTIFICATION_CHANGE_DIR_ATTRIBUTES = 0x00000002; const DIR_NOTIFICATION_REMOVE_ENTRY = 0x00000004; const DIR_NOTIFICATION_ADD_ENTRY = 0x00000008; const DIR_NOTIFICATION_RENAME_ENTRY = 0x00000010; const DIR_NOTIFICATION_CHANGE_COOKIE_VERIFIER = 0x00000020; typedef uint32_t dir_notification_type4; typedef nfstime4 attr_notice4; RESULT Shepler Expires June 15, 2006 [Page 117] Internet-Draft NFSv4 Minior Version 1 December 2005 struct GET_DIR_DELEGATION4resok { verifier4 cookieverf; /* Stateid for get_dir_delegation */ stateid4 stateid; /* Which notifications can the server support */ dir_notification_type4 supp_notification; bitmap4 child_attributes; bitmap4 dir_attributes; }; union GET_DIR_DELEGATION4res switch (nfsstat4 status) { case NFS4_OK: /* CURRENT_FH: delegated dir */ GET_DIR_DELEGATION4resok resok4; default: void; }; DESCRIPTION The GET_DIR_DELEGATION operation is used by a client to request a directory delegation. The directory is represented by the current filehandle. The client also specifies whether it wants the server to notify it when the directory changes in certain ways by setting one or more bits in a bitmap. The server may also choose not to grant the delegation. In that case the server will return NFS4ERR_DIRDELEG_UNAVAIL. If the server decides to hand out the delegation, it will return a cookie verifier for that directory. If the cookie verifier changes when the client is holding the delegation, the delegation will be recalled unless the client has asked for notification for this event. In that case a notification will be sent to the client. The server will also return a directory delegation stateid in addition to the cookie verifier as a result of the GET_DIR_DELEGATION operation. This stateid will appear in callback messages related to the delegation, such as notifications and delegation recalls. The client will use this stateid to return the delegation voluntarily or upon recall. A delegation is returned by calling the DELEGRETURN operation. The server may not be able to support notifications of certain events. If the client asks for such notifications, the server must inform the client of its inability to do so as part of the GET_DIR_DELEGATION reply by not setting the appropriate bits in the supported notifications bitmask contained in the reply. Shepler Expires June 15, 2006 [Page 118] Internet-Draft NFSv4 Minior Version 1 December 2005 The GET_DIR_DELEGATION operation can be used for both normal and named attribute directories. It covers all the entries in the directory except the ".." entry. That means if a directory and its parent both hold directory delegations, any changes to the parent will not cause a notification to be sent for the child even though the child's ".." entry points to the parent. IMPLEMENTATION Directory delegation provides the benefit of improving cache consistency of namespace information. This is done through synchronous callbacks. A server must support synchronous callbacks in order to support directory delegations. In addition to that, asynchronous notifications provide a way to reduce network traffic as well as improve client performance in certain conditions. Notifications would not be requested when the goal is just cache consitency. Notifications are specified in terms of potential changes to the directory. A client can ask to be notified whenever an entry is added to a directory by setting notification_type to DIR_NOTIFICATION_ADD_ENTRY. It can also ask for notifications on entry removal, renames, directory attribute changes and cookie verifier changes by setting notification_type flag appropriately. In addition to that, the client can also ask for notifications upon attribute changes to children in the directory to keep its attribute cache up to date. However any changes made to child attributes do not cause the delegation to be recalled. If a client is interested in directory entry caching, or negative name caching, it can set the notification_type appropriately and the server will notify it of all changes that would otherwise invalidate its name cache. The kind of notification a client asks for may depend on the directory size, its rate of change and the applications being used to access that directory. However, the conditions under which a client might ask for a notification, is out of the scope of this specification. The client will set one or more bits in a bitmap (notification_type) to let the server know what kind of notification(s) it is interested in. For attribute notifications it will set bits in another bitmap to indicate which attributes it wants to be notified of. If the server does not support notifications for changes to a certain attribute, it should not set that attribute in the supported attribute bitmap (supp_notification) specified in the reply. Shepler Expires June 15, 2006 [Page 119] Internet-Draft NFSv4 Minior Version 1 December 2005 In addition to that, the client will also let the server know if it wants to get the notification as soon as the attribute change occurs or after a certain delay by setting a delay factor, child_attr_delay for attribute changes to children and dir_attr_delay for attribute changes to the directory. If this delay factor is set to zero, that indicates to the server that the client wants to be notified of any attribute changes as soon as they occur. If the delay factor is set to N, the server will make a best effort guarantee that attribute updates are not out of sync by more than that. One value covers all attribute changes for the directory and another value covers all attribute changes for all children in the directory. If the client asks for a delay factor that the server does not support or that may cause significant resource consumption on the server by causing the server to send a lot of notifications, the server should not commit to sending out notifications for that attribute and therefore must not set the approprite bit in the child_attributes and dir_attributes bitmaps in the response. The server will let the client know about which notifications it can support by setting appropriate bits in a bitmap. If it agrees to send attribute notifications, it will also set two attribute masks indicating which attributes it will send change notifications for. One of the masks covers changes in directory attributes and the other covers atttribute changes to any files in the directory. The client should use a security flavor that the filesystem is exported with. If it uses a different flavor, the server should return NFS4ERR_WRONGSEC. ERRORS NFS4ERR_ACCESS NFS4ERR_BADHANDLE NFS4ERR_BADXDR NFS4ERR_FHEXPIRED NFS4ERR_INVAL NFS4ERR_MOVED NFS4ERR_NOFILEHANDLE NFS4ERR_NOTDIR NFS4ERR_RESOURCE NFS4ERR_SERVERFAULT NFS4ERR_STALE NFS4ERR_DIRDELEG_UNAVAIL NFS4ERR_WRONGSEC NFS4ERR_EIO NFS4ERR_NOTSUPP 14.12 CB_NOTIFY - Notify directory changes Tell the client of directory changes. Shepler Expires June 15, 2006 [Page 120] Internet-Draft NFSv4 Minior Version 1 December 2005 SYNOPSIS stateid, notification -> {} ARGUMENT struct CB_NOTIFY4args { stateid4 stateid; dir_notification4 changes<>; }; /* * Notification information sent to the client. */ union dir_notification4 switch (dir_notification_type4 notification_type) { case DIR_NOTIFICATION_CHANGE_CHILD_ATTRIBUTES: dir_notification_attribute4 change_child_attributes; case DIR_NOTIFICATION_CHANGE_DIR_ATTRIBUTES: fattr4 change_dir_attributes; case DIR_NOTIFICATION_REMOVE_ENTRY: dir_notification_remove4 remove_notification; case DIR_NOTIFICATION_ADD_ENTRY: dir_notification_add4 add_notification; case DIR_NOTIFICATION_RENAME_ENTRY: dir_notification_rename4 rename_notification; case DIR_NOTIFICATION_CHANGE_COOKIE_VERIFIER: dir_notification_verifier4 verf_notification; }; /* * Changed entry information. */ struct dir_entry { component4 file; fattr4 attrs; }; struct dir_notification_attribute4 { dir_entry changed_entry; }; struct dir_notification_remove4 { dir_entry old_entry; Shepler Expires June 15, 2006 [Page 121] Internet-Draft NFSv4 Minior Version 1 December 2005 nfs_cookie4 old_entry_cookie; }; struct dir_notification_rename4 { dir_entry old_entry; dir_notification_add4 new_entry; }; struct dir_notification_verifier4 { verifier4 old_cookieverf; verifier4 new_cookieverf; }; struct dir_notification_add4 { dir_entry new_entry; /* what READDIR would have returned for this entry */ nfs_cookie4 new_entry_cookie; bool last_entry; prev_entry_info4 prev_info; }; union prev_entry_info4 switch (bool isprev) { case TRUE: /* A previous entry exists */ prev_entry4 prev_entry_info; case FALSE: /* we are adding to an empty directory */ void; }; /* * Previous entry information */ struct prev_entry4 { dir_entry prev_entry; /* what READDIR returned for this entry */ nfs_cookie4 prev_entry_cookie; }; RESULT struct CB_NOTIFY4res { nfsstat4 status; }; Shepler Expires June 15, 2006 [Page 122] Internet-Draft NFSv4 Minior Version 1 December 2005 DESCRIPTION The CB_NOTIFY operation is used by the server to send notifications to clients about changes in a delegated directory. These notifications are sent over the callback path. The notification is sent once the original request has been processed on the server. The server will send an array of notifications for all changes that might have occurred in the directory. The dir_notification_type4 can only have one bit set for each notification in the array. If the client holding the delegation makes any changes in the directory that cause files or sub directories to be added or removed, the server will notify that client of the resulting change(s). If the client holding the delegation is making attribute or cookie verifier changes only, the server does not need to send notifications to that client. The server will send the following information for each operation: * ADDING A FILE: The server will send information about the new entry being created along with the cookie for that entry. The entry information contains the nfs name of the entry and attributes. If this entry is added to the end of the directory, the server will set a last_entry flag to true. If the file is added such that there is atleast one entry before it, the server will also return the previous entry information along with its cookie. This is to help clients find the right location in their DNLC or directory caches where this entry should be cached. * REMOVING A FILE: The server will send information about the directory entry being deleted. The server will also send the cookie value for the deleted entry so that clients can get to the cached information for this entry. * RENAMING A FILE: The server will send information about both the old entry and the new entry. This includes name and attributes for each entry. This notification is only sent if both entries are in the same directory. If the rename is across directories, the server will send a remove notification to one directory and an add notification to the other directory, assuming both have a directory delegation. * FILE/DIR ATTRIBUTE CHANGE: The client will use the attribute mask to inform the server of attributes for which it wants to receive notifications. This change notification can be requested for both changes to the attributes of the directory as well as changes to any file attributes in the directory by Shepler Expires June 15, 2006 [Page 123] Internet-Draft NFSv4 Minior Version 1 December 2005 using two separate attribute masks. The client can not ask for change attribute notification per file. One attribute mask covers all the files in the directory. Upon any attribute change, the server will send back the values of changed attributes. Notifications might not make sense for some filesystem wide attributes and it is up to the server to decide which subset it wants to support. The client can negotiate the frequency of attribute notifications by letting the server know how often it wants to be notified of an attribute change. The server will return supported notification frequencies or an indication that no notification is permitted for directory or child attributes by setting the supp_dir_attr_notice and supp_child_attr_notice attributes respectively. * COOKIE VERIFIER CHANGE: If the cookie verifier changes while a client is holding a delegation, the server will notify the client so that it can invalidate its cookies and reissue a READDIR to get the new set of cookies. IMPLEMENTATION ERRORS NFS4ERR_BAD_STATEID NFS4ERR_INVAL NFS4ERR_BADXDR NFS4ERR_SERVERFAULT 14.13 CB_RECALL_ANY - Keep any N delegations Notify client to return delegation and keep N of them. SYNOPSIS N -> {} ARGUMENT struct CB_RECALLANYY4args { uint4 dlgs_to_keep; } Shepler Expires June 15, 2006 [Page 124] Internet-Draft NFSv4 Minior Version 1 December 2005 RESULT struct CB_RECALLANY4res { nfsstat4 status; }; DESCRIPTION The server may decide that it can not hold all the delegation state without running out of resources. Since the server has no knowledge of which delegations are being used more than others, it can not implement an effective reclaim scheme that avoids reclaiming frequently used delegations. In that case the server may issue a CB_RECALL_ANY callback to the client asking it to keep N delegations and return the rest. The reason why CB_RECALL_ANY specifies a count of delegations the client may keep as opposed to a count of delegations the client must yield is as follows. Were it otherwise, there is a potential for a race between a CB_RECALL_ANY that had a count of delegations to free with a set of client originated operations to return delegations. As a result of the race the client and server would have differing ideas as to how many delegations to return. Hence the client could mistakenly free too many delegations. This operation applies to delegations for a regular file (read or write) as well as for a directory. The client can choose to return any type of delegation as a result of this callback i.e. read, write or directory delegation. The client can also choose to keep more delegations than what the server asked for and it is up to the server to handle this situation. The server must give the client enough time to return the delegations. This time should not be less than the lease period. IMPLEMENTATION ERRORS NFS4ERR_RESOURCE Shepler Expires June 15, 2006 [Page 125] Internet-Draft NFSv4 Minior Version 1 December 2005 14.14 LAYOUTGET - Get Layout Information SYNOPSIS (cfh), clientid, layout_type, iomode, offset, length, minlength, maxcount -> layout ARGUMENT struct LAYOUTGET4args { /* CURRENT_FH: file */ clientid4 clientid; pnfs_layouttype4 layout_type; pnfs_layoutiomode4 iomode; offset4 offset; length4 length; length4 minlength; count4 maxcount; }; RESULT struct LAYOUTGET4resok { pnfs_layout4 layout; }; union LAYOUTGET4res switch (nfsstat4 status) { case NFS4_OK: LAYOUTGET4resok resok4; default: void; }; DESCRIPTION Requests a layout for reading or writing (and reading) the file given by the filehandle at the byte range specified by offset and length. Layouts are identified by the clientid, filehandle, and layout type. The use of the iomode depends upon the layout type, but should reflect the client's data access intent. The LAYOUTGET operation returns layout information for the specified byte range, a layout segment. To get a layout segment from a specific offset through the end-of-file, regardless of the file's length, a length field with all bits set to 1 (one) should be used. If the length is zero, or if a length which is not all bits set to one is specified, and length when added to the offset exceeds the maximum 64-bit unsigned integer value, the error NFS4ERR_INVAL will Shepler Expires June 15, 2006 [Page 126] Internet-Draft NFSv4 Minior Version 1 December 2005 result. The "minlength" field specifies the minimum size overlap with the requested offset and length that is to be returned. If this requirement cannot be met, no layout must be returned; the error NFS4ERR_LAYOUTTRYLATER can be returned. The "maxcount" field specifies the maximum layout size (in bytes) that the client can handle. If the size of the layout structure exceeds the size specified by maxcount, the metadata server will return the NFS4ERR_TOOSMALL error. As well, the metadata server may adjust the range of the returned layout segment based on striping patterns and usage implied by the iomode. The client must be prepared to get a layout that does not line up exactly with their request; there MUST be at least an overlap of "minlength" between the layout returned by the server and the client's request, or the server SHOULD reject the request. See Section 7.3 for more details. The metadata server may also return a layout segment with an iomode other than that requested by the client. If it does so, it must ensure that the iomode is more permissive than the iomode requested. E.g., this allows an implementation to upgrade read-only requests to read/write requests at its discretion, within the limits of the layout type specific protocol. An iomode of either LAYOUTIOMODE_READ or LAYOUTIOMODE_RW must be returned. The format of the returned layout is specific to the underlying file system. Layout types other than the NFSv4 file layout type should be specified outside of this document. If layouts are not supported for the requested file or its containing file system the server SHOULD return NFS4ERR_LAYOUTUNAVAILABLE. If the layout type is not supported, the metadata server should return NFS4ERR_UNKNOWN_LAYOUTTYPE. If layouts are supported but no layout matches the client provided layout identification, the server should return NFS4ERR_BADLAYOUT. If an invalid iomode is specified, or an iomode of LAYOUTIOMODE_ANY is specified, the server should return NFS4ERR_BADIOMODE. If the layout for the file is unavailable due to transient conditions, e.g. file sharing prohibits layouts, the server must return NFS4ERR_LAYOUTTRYLATER. If the layout request is rejected due to an overlapping layout recall, the server must return NFS4ERR_RECALLCONFLICT. See Section 7.5.3 for details. Shepler Expires June 15, 2006 [Page 127] Internet-Draft NFSv4 Minior Version 1 December 2005 If the layout conflicts with a mandatory byte range lock held on the file, and if the storage devices have no method of enforcing mandatory locks, other than through the restriction of layouts, the metadata server should return NFS4ERR_LOCKED. On success, the current filehandle retains its value. IMPLEMENTATION Typically, LAYOUTGET will be called as part of a compound RPC after an OPEN operation and results in the client having location information for the file; a client may also hold a layout across multiple OPENs. The client specifies a layout type that limits what kind of layout the server will return. This prevents servers from issuing layouts that are unusable by the client. ERRORS NFS4ERR_BADLAYOUT NFS4ERR_BADIOMODE NFS4ERR_FHEXPIRED NFS4ERR_INVAL NFS4ERR_LAYOUTUNAVAILABLE NFS4ERR_LAYOUTTRYLATER NFS4ERR_LOCKED NFS4ERR_NOFILEHANDLE NFS4ERR_NOTSUPP NFS4ERR_RECALLCONFLICT NFS4ERR_STALE NFS4ERR_STALE_CLIENTID NFS4ERR_TOOSMALL NFS4ERR_UNKNOWN_LAYOUTTYPE 14.15 LAYOUTCOMMIT - Commit writes made using a layout Shepler Expires June 15, 2006 [Page 128] Internet-Draft NFSv4 Minior Version 1 December 2005 SYNOPSIS (cfh), clientid, offset, length, last_write_offset, time_modify, time_access, layoutupdate -> newsize ARGUMENT union newtime4 switch (bool timechanged) { case TRUE: nfstime4 time; case FALSE: void; }; union newsize4 switch (bool sizechanged) { case TRUE: length4 size; case FALSE: void; }; struct LAYOUTCOMMIT4args { /* CURRENT_FH: file */ clientid4 clientid; offset4 offset; length4 length; length4 last_write_offset; newtime4 time_modify; newtime4 time_access; pnfs_layoutupdate4 layoutupdate; }; RESULT struct LAYOUTCOMMIT4resok { newsize4 newsize; }; union LAYOUTCOMMIT4res switch (nfsstat4 status) { case NFS4_OK: LAYOUTCOMMIT4resok resok4; default: void; }; DESCRIPTION Shepler Expires June 15, 2006 [Page 129] Internet-Draft NFSv4 Minior Version 1 December 2005 Commits changes in the layout segment represented by the current filehandle, clientid, and byte range. Since layouts are sub- dividable, a smaller portion of a layout, retrieved via LAYOUTGET, may be committed. The region being committed is specified through the byte range (length and offset). Note: the "layoutupdate" structure does not include the length and offset, as they are already specified in the arguments. The LAYOUTCOMMIT operation indicates that the client has completed writes using a layout obtained by a previous LAYOUTGET. The client may have only written a subset of the data range it previously requested. LAYOUTCOMMIT allows it to commit or discard provisionally allocated space and to update the server with a new end of file. The layout referenced by LAYOUTCOMMIT is still valid after the operation completes and can be continued to be referenced by the clientid, filehandle, byte range, and layout type. The "last_write_offset" field specifies the offset of the last byte written by the client previous to the LAYOUTCOMMIT. Note: this value is never equal to the file's size (at most it is one byte less than the file's size). The metadata server may use this information to determine whether the file's size needs to be updated. If the metadata server updates the file's size as the result of the LAYOUTCOMMIT operation, it must return the new size as part of the results. The "time_modify" and "time_access" fields allow the client to suggest times it would like the metadata server to set. The metadata server may use these time values or it may use the time of the LAYOUTCOMMIT operation to set these time values. If the metadata server uses the client provided times, it should sanity check the values (e.g., to ensure time does not flow backwards). If the client wants to force the metadata server to set an exact time, the client should use a SETATTR operation in a compound right after LAYOUTCOMMIT. See Section 7.4 for more details. If the new client desires the resultant mtime or atime, it should issue a GETATTR following the LAYOUTCOMMIT; e.g., later in the same compound. The "layoutupdate" argument to LAYOUTCOMMIT provides a mechanism for a client to provide layout specific updates to the metadata server. For example, the layout update can describe what regions of the original layout have been used and what regions can be deallocated. There is no NFSv4 file layout specific layoutupdate structure. The layout information is more verbose for block devices than for objects and files because the latter hide the details of block allocation behind their storage protocols. At the minimum, the client needs to communicate changes to the end of file location back Shepler Expires June 15, 2006 [Page 130] Internet-Draft NFSv4 Minior Version 1 December 2005 to the server, and, if desired, its view of the file modify and access time. For block/volume layouts, it needs to specify precisely which blocks have been used. If the layout identified in the arguments does not exist, the error NFS4ERR_BADLAYOUT is returned. The layout being committed may also be rejected if it does not correspond to an existing layout with an iomode of RW. On success, the current filehandle retains its value. ERRORS NFS4ERR_BADLAYOUT NFS4ERR_BADIOMODE NFS4ERR_FHEXPIRED NFS4ERR_INVAL NFS4ERR_NOFILEHANDLE NFS4ERR_STALE NFS4ERR_STALE_CLIENTID NFS4ERR_UNKNOWN_LAYOUTTYPE 14.16 LAYOUTRETURN - Release Layout Information SYNOPSIS (cfh), clientid, offset, length, iomode, layout_type -> - ARGUMENT struct LAYOUTRETURN4args { /* CURRENT_FH: file */ clientid4 clientid; offset4 offset; length4 length; pnfs_layoutiomode4 iomode; pnfs_layouttype4 layout_type; }; RESULT struct LAYOUTRETURN4res { nfsstat4 status; }; DESCRIPTION Shepler Expires June 15, 2006 [Page 131] Internet-Draft NFSv4 Minior Version 1 December 2005 Returns the layout segment represented by the current filehandle, clientid, byte range, iomode, and layout type. After this call, the client MUST NOT use the layout and the associated storage protocol to access the file data. The layout being returned may be a subdivision of a layout previously fetched through LAYOUTGET. As well, it may be a subset or superset of a layout specified by CB_LAYOUTRECALL. However, if it is a subset, the recall is not complete until the full byte range has been returned. It is also permissible, and no error should result, for a client to return a byte range covering a layout it does not hold. If the length is all 1s, the layout covers the range from offset to EOF. An iomode of ANY specifies that all layouts that match the other arguments to LAYOUTRETURN (i.e., clientid, byte range, and type) are being returned. Layouts may be returned when recalled or voluntarily (i.e., before the server has recalled them). In either case the client must properly propagate state changed under the context of the layout to storage or to the server before returning the layout. If a client fails to return a layout in a timely manner, then the metadata server should use its control protocol with the storage devices to fence the client from accessing the data referenced by the layout. See Section 7.5 for more details. If the layout identified in the arguments does not exist, the error NFS4ERR_BADLAYOUT is returned. If a layout exists, but the iomode does not match, NFS4ERR_BADIOMODE is returned. On success, the current filehandle retains its value. [OPEN ISSUE: Should LAYOUTRETURN be modified to handle FSID callbacks?] ERRORS NFS4ERR_BADLAYOUT NFS4ERR_BADIOMODE NFS4ERR_FHEXPIRED NFS4ERR_INVAL NFS4ERR_NOFILEHANDLE NFS4ERR_STALE NFS4ERR_STALE_CLIENTID NFS4ERR_UNKNOWN_LAYOUTTYPE Shepler Expires June 15, 2006 [Page 132] Internet-Draft NFSv4 Minior Version 1 December 2005 14.17 GETDEVICEINFO - Get Device Information SYNOPSIS (cfh), device_id, layout_type, maxcount -> device_addr ARGUMENT struct GETDEVICEINFO4args { /* CURRENT_FH: file */ pnfs_deviceid4 device_id; pnfs_layouttype4 layout_type; count4 maxcount; }; RESULT struct GETDEVICEINFO4resok { pnfs_deviceaddr4 device_addr; }; union GETDEVICEINFO4res switch (nfsstat4 status) { case NFS4_OK: GETDEVICEINFO4resok resok4; default: void; }; DESCRIPTION Returns device type and device address information for a specified device. The returned device_addr includes a type that indicates how to interpret the addressing information for that device. The current filehandle (cfh) is used to identify the file system; device IDs are unique per file system (FSID) and are qualified by the layout type. See Section 7.1.4 for more details on device ID assignment. If the size of the device address exceeds maxcount bytes, the metadata server will return the error NFS4ERR_TOOSMALL. If an invalid device ID is given, the metadata server will respond with NFS4ERR_INVAL. ERRORS NFS4ERR_FHEXPIRED NFS4ERR_INVAL NFS4ERR_TOOSMALL Shepler Expires June 15, 2006 [Page 133] Internet-Draft NFSv4 Minior Version 1 December 2005 NFS4ERR_UNKNOWN_LAYOUTTYPE 14.18 GETDEVICELIST - Get List of Devices SYNOPSIS (cfh), layout_type, maxcount, cookie, cookieverf -> cookie, cookieverf, device_addrs<> ARGUMENT struct GETDEVICELIST4args { /* CURRENT_FH: file */ pnfs_layouttype4 layout_type; count4 maxcount; nfs_cookie4 cookie; verifier4 cookieverf; }; RESULT struct GETDEVICELIST4resok { nfs_cookie4 cookie; verifier4 cookieverf; pnfs_devlist_item4 device_addrs<>; }; union GETDEVICELIST4res switch (nfsstat4 status) { case NFS4_OK: GETDEVICELIST4resok resok4; default: void; }; DESCRIPTION In some applications, especially SAN environments, it is convenient to find out about all the devices associated with a file system. This lets a client determine if it has access to these devices, e.g., at mount time. This operation returns an array of items (pnfs_devlist_item4) that establish the association between the short pnfs_deviceid4 and the addressing information for that device, for a particular layout type. This operation may not be able to fetch all device information at once, thus it uses a cookie based approach, similar to READDIR, to fetch additional device information (see [2], section 14.2.24). As Shepler Expires June 15, 2006 [Page 134] Internet-Draft NFSv4 Minior Version 1 December 2005 in GETDEVICEINFO, the current filehandle (cfh) is used to identify the file system. As in GETDEVICEINFO, maxcount specifies the maximum number of bytes to return. If the metadata server is unable to return a single device address, it will return the error NFS4ERR_TOOSMALL. If an invalid device ID is given, the metadata server will respond with NFS4ERR_INVAL. ERRORS NFS4ERR_BAD_COOKIE NFS4ERR_FHEXPIRED NFS4ERR_INVAL NFS4ERR_TOOSMALL NFS4ERR_UNKNOWN_LAYOUTTYPE Shepler Expires June 15, 2006 [Page 135] Internet-Draft NFSv4 Minior Version 1 December 2005 14.19 CB_LAYOUTRECALL SYNOPSIS layout_type, iomode, layoutchanged, layoutrecall -> - ARGUMENT enum layoutrecall_type4 { RECALL_FILE = 1, RECALL_FSID = 2 }; struct layoutrecall_file4 { nfs_fh4 fh; offset4 offset; length4 length; }; union layoutrecall4 switch(layoutrecall_type4 recalltype) { case RECALL_FILE: layoutrecall_file4 layout; case RECALL_FSID: fsid4 fsid; }; struct CB_LAYOUTRECALLargs { pnfs_layouttype4 layout_type; pnfs_layoutiomode4 iomode; bool layoutchanged; layoutrecall4 layoutrecall; }; RESULT struct CB_LAYOUTRECALLres { nfsstat4 status; }; DESCRIPTION The CB_LAYOUTRECALL operation is used to begin the process of recalling a layout, a portion thereof, or all layouts pertaining to a particular file system (FSID). If RECALL_FILE is specified, the offset and length fields specify the portion of the layout to be returned. The iomode specifies the set of layouts to be returned. An iomode of ANY specifies that all matching layouts, regardless of iomode, must be returned; otherwise, only layouts that exactly match Shepler Expires June 15, 2006 [Page 136] Internet-Draft NFSv4 Minior Version 1 December 2005 the iomode must be returned. If the "layoutchanged" field is TRUE, then the client SHOULD not flush its dirty data to the devices specified by the layout being recalled. Instead, it is preferable for the client to flush the dirty data through the metadata server. Alternatively, the client may attempt to obtain a new layout. Note: in order to obtain a new layout the client must first return the old layout. Since obtaining a new layout is not guaranteed to succeed, the client must be ready to flush its dirty data through the metadata server. If RECALL_FSID is specified, the fsid specifies the file system for which any outstanding layouts must be returned. Layouts are returned through the LAYOUTRETURN operation. If the client does not hold any layout segment either matching or overlapping with the requested layout, it returns NFS4ERR_NOMATCHING_LAYOUT. If a length of all 1s is specified then the layout corresponding to the byte range from "offset" to the end- of-file MUST be returned. IMPLEMENTATION The client should reply to the callback immediately. Replying does not complete the recall except when an error is returned. The recall is not complete until the layout(s) are returned using a LAYOUTRETURN. The client should complete any in-flight I/O operations using the recalled layout(s) before returning it/them via LAYOUTRETURN. If the client has buffered dirty data there are a number of options for flushing that data. If "layoutchanged" is false, the client may choose to write dirty data directly to storage before calling LAYOUTRETURN. However, if "layoutchanged" is true, the client may either choose to write it later using normal NFSv4 WRITE operations to the metadata server or it may attempt to obtain a new layout, after first returning the recalled layout, using the new layout to flush the dirty data. Regardless of whether the client is holding a layout, it may always write data through the metadata server. If dirty data is flushed while the layout is held, the client must still issue LAYOUTCOMMIT operations at the appropriate time, especially before issuing the LAYOUTRETURN. If a large amount of dirty data is outstanding, the client may issue LAYOUTRETURNs for portions of the layout being recalled; this allows the server to monitor the client's progress and adherence to the callback. However, the last LAYOUTRETURN in a sequence of returns, SHOULD specify the full range being recalled (see Section 7.5.2 for Shepler Expires June 15, 2006 [Page 137] Internet-Draft NFSv4 Minior Version 1 December 2005 details). ERRORS NFS4ERR_NOMATCHING_LAYOUT 14.20 CB_SIZECHANGED SYNOPSIS fh, size -> - ARGUMENT struct CB_SIZECHANGEDargs { nfs_fh4 fh; length4 size; }; RESULT struct CB_SIZECHANGEDres { nfsstat4 status; }; DESCRIPTION The CB_SIZECHANGED operation is used to notify the client that the size pertaining to the filehandle associated with "fh", has changed. The new size is specified. Upon reception of this notification callback, the client should update its internal size for the file. If the layout being held for the file is of the NFSv4 file layout type, then the size field within that layout should be updated (see Section 9.5). For other layout types see Section 7.4.2 for more details. If the handle specified is not one for which the client holds a layout, an NFS4ERR_BADHANDLE error is returned. ERRORS NFS4ERR_BADHANDLE 15. References Shepler Expires June 15, 2006 [Page 138] Internet-Draft NFSv4 Minior Version 1 December 2005 15.1 Normative References [1] Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", March 1997. [2] Shepler, S., Callaghan, B., Robinson, D., Thurlow, R., Beame, C., Eisler, M., and D. Noveck, "Network File System (NFS) version 4 Protocol", RFC 3530, April 2003. 15.2 Informative References [3] Satran, J., Meth, K., Sapuntzakis, C., Chadalapaka, M., and E. Zeidner, "Internet Small Computer Systems Interface (iSCSI)", RFC 3720, April 2004. [4] Snively, R., "Fibre Channel Protocol for SCSI, 2nd Version (FCP-2)", ANSI/INCITS 350-2003, Oct 2003. [5] Weber, R., "Object-Based Storage Device Commands (OSD)", ANSI/ INCITS 400-2004, July 2004, . [6] Black, D., "pNFS Block/Volume Layout", July 2005, . [7] Zelenka, J., Welch, B., and B. Halevy, "Object-based pNFS Operations", July 2005, . Author's Address Spencer Shepler Sun Microsystems, Inc. 7808 Moonflower Drive Austin, TX 78750 USA Phone: +1-512-349-9376 Email: spencer.shepler@sun.com Appendix A. Acknowledgments The initial drafts for the SECINFO extensions were edited by Mike Eisler with contributions from Tom Talpey, Saadia Khan, and Jon Bauman. The initial drafts for the SESSIONS extensions were edited by Tom Shepler Expires June 15, 2006 [Page 139] Internet-Draft NFSv4 Minior Version 1 December 2005 Talpey, Spencer Shepler, Jon Bauman with contributions from Charles Antonelli, Brent Callaghan, Mike Eisler, John Howard, Chet Juszczak, Trond Myklebust, Dave Noveck, John Scott, Mike stolarchuk and Mark Wittle. The initial drafts for the Directory Delegations support were contributed by Saadia Khan with input from Dave Noveck, Mike Eisler, Carl Burnett, Ted Anderson and Tom Talpey. The initial drafts for the parellel NFS support were edited by Brent Welch and Garth Goodson. Additional authors for those documents were Benny Halevy, David Black, and Andy Adamson. Additional input came from the informal group which contributed to the construction of the initial pNFS drafts; specific acknowledgement goes to Gary Grider, Peter Corbett, Dave Noveck, and Peter Honeyman. The pNFS work was inspired by the NASD and OSD work done by Garth Gibson. Gary Grider of the national labs (LANL) has also been a champion of high- performance parallel I/O. Shepler Expires June 15, 2006 [Page 140] Internet-Draft NFSv4 Minior Version 1 December 2005 Intellectual Property Statement The IETF takes no position regarding the validity or scope of any Intellectual Property Rights or other rights that might be claimed to pertain to the implementation or use of the technology described in this document or the extent to which any license under such rights might or might not be available; nor does it represent that it has made any independent effort to identify any such rights. Information on the procedures with respect to rights in RFC documents can be found in BCP 78 and BCP 79. Copies of IPR disclosures made to the IETF Secretariat and any assurances of licenses to be made available, or the result of an attempt made to obtain a general license or permission for the use of such proprietary rights by implementers or users of this specification can be obtained from the IETF on-line IPR repository at http://www.ietf.org/ipr. The IETF invites any interested party to bring to its attention any copyrights, patents or patent applications, or other proprietary rights that may cover technology that may be required to implement this standard. Please address the information to the IETF at ietf-ipr@ietf.org. Disclaimer of Validity This document and the information contained herein are provided on an "AS IS" basis and THE CONTRIBUTOR, THE ORGANIZATION HE/SHE REPRESENTS OR IS SPONSORED BY (IF ANY), THE INTERNET SOCIETY AND THE INTERNET ENGINEERING TASK FORCE DISCLAIM ALL WARRANTIES, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO ANY WARRANTY THAT THE USE OF THE INFORMATION HEREIN WILL NOT INFRINGE ANY RIGHTS OR ANY IMPLIED WARRANTIES OF MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE. Copyright Statement Copyright (C) The Internet Society (2005). This document is subject to the rights, licenses and restrictions contained in BCP 78, and except as set forth therein, the authors retain all their rights. Acknowledgment Funding for the RFC Editor function is currently provided by the Internet Society. Shepler Expires June 15, 2006 [Page 141]