NFSv4 T. Haynes Internet-Draft Editor Intended status: Standards Track April 18, 2011 Expires: October 20, 2011 NFS Version 4 Minor Version 2 draft-ietf-nfsv4-minorversion2-00.txt Abstract This Internet-Draft describes NFS version 4 minor version two, focusing mainly on the protocol extensions made from NFS version 4 minor version 0 and NFS version 4 minor version 1. Major extensions introduced in NFS version 4 minor version two include: Server-side Copy, Space Reservations, and Support for Sparse Files. 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]. Status of this Memo This Internet-Draft is submitted to IETF in full conformance with the provisions of BCP 78 and 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 October 20, 2011. Copyright Notice Haynes Expires October 20, 2011 [Page 1] Internet-Draft NFSv4.2 April 2011 Copyright (c) 2011 IETF Trust and the persons identified as the document authors. All rights reserved. This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (http://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the BSD License. This document may contain material from IETF Documents or IETF Contributions published or made publicly available before November 10, 2008. The person(s) controlling the copyright in some of this material may not have granted the IETF Trust the right to allow modifications of such material outside the IETF Standards Process. Without obtaining an adequate license from the person(s) controlling the copyright in such materials, this document may not be modified outside the IETF Standards Process, and derivative works of it may not be created outside the IETF Standards Process, except to format it for publication as an RFC or to translate it into languages other than English. Haynes Expires October 20, 2011 [Page 2] Internet-Draft NFSv4.2 April 2011 Table of Contents 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 4 1.1. The NFS Version 4 Minor Version 2 Protocol . . . . . . . . 4 1.2. Scope of This Document . . . . . . . . . . . . . . . . . . 4 1.3. NFSv4.2 Goals . . . . . . . . . . . . . . . . . . . . . . 4 1.4. Overview of NFSv4.2 Features . . . . . . . . . . . . . . . 4 1.5. Differences from NFSv4.1 . . . . . . . . . . . . . . . . . 4 2. pNFS Access Permissions Check . . . . . . . . . . . . . . . . 4 2.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . 4 2.2. Changes to Operation 51: LAYOUTRETURN (RFC 5661) . . . . . 6 2.2.1. ARGUMENT (18.44.1) . . . . . . . . . . . . . . . . . . 7 2.2.2. RESULT (18.44.2) . . . . . . . . . . . . . . . . . . . 8 2.2.3. DESCRIPTION (18.44.3) . . . . . . . . . . . . . . . . 8 2.2.4. IMPLEMENTATION (18.44.4) . . . . . . . . . . . . . . . 9 2.3. Change to NFS4ERR_NXIO Usage . . . . . . . . . . . . . . . 11 2.4. Security Considerations . . . . . . . . . . . . . . . . . 11 2.5. IANA Considerations . . . . . . . . . . . . . . . . . . . 11 3. Sharing change attribute implementation details with NFSv4 clients . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 3.1. Abstract . . . . . . . . . . . . . . . . . . . . . . . . . 11 3.2. Introduction . . . . . . . . . . . . . . . . . . . . . . . 12 3.3. Definition of the 'change_attr_type' per-file system attribute . . . . . . . . . . . . . . . . . . . . . . . . 12 4. NFS Server-side Copy . . . . . . . . . . . . . . . . . . . . . 13 4.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . 14 4.2. Protocol Overview . . . . . . . . . . . . . . . . . . . . 14 4.2.1. Intra-Server Copy . . . . . . . . . . . . . . . . . . 16 4.2.2. Inter-Server Copy . . . . . . . . . . . . . . . . . . 17 4.2.3. Server-to-Server Copy Protocol . . . . . . . . . . . . 20 4.3. Operations . . . . . . . . . . . . . . . . . . . . . . . . 22 4.3.1. netloc4 - Network Locations . . . . . . . . . . . . . 22 4.3.2. Operation 61: COPY_NOTIFY - Notify a source server of a future copy . . . . . . . . . . . . . . . . . . . 23 4.3.3. Operation 62: COPY_REVOKE - Revoke a destination server's copy privileges . . . . . . . . . . . . . . . 25 4.3.4. Operation 59: COPY - Initiate a server-side copy . . . 26 4.3.5. Operation 60: COPY_ABORT - Cancel a server-side copy . . . . . . . . . . . . . . . . . . . . . . . . . 34 4.3.6. Operation 63: COPY_STATUS - Poll for status of a server-side copy . . . . . . . . . . . . . . . . . . . 35 4.3.7. Operation 15: CB_COPY - Report results of a server-side copy . . . . . . . . . . . . . . . . . . . 36 4.3.8. Copy Offload Stateids . . . . . . . . . . . . . . . . 37 4.4. Security Considerations . . . . . . . . . . . . . . . . . 38 4.4.1. Inter-Server Copy Security . . . . . . . . . . . . . . 38 4.5. IANA Considerations . . . . . . . . . . . . . . . . . . . 46 5. Space Reservation . . . . . . . . . . . . . . . . . . . . . . 46 Haynes Expires October 20, 2011 [Page 3] Internet-Draft NFSv4.2 April 2011 5.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . 46 5.2. Use Cases . . . . . . . . . . . . . . . . . . . . . . . . 47 5.2.1. Space Reservation . . . . . . . . . . . . . . . . . . 47 5.2.2. Space freed on deletes . . . . . . . . . . . . . . . . 48 5.2.3. Operations and attributes . . . . . . . . . . . . . . 49 5.2.4. Attribute 77: space_reserve . . . . . . . . . . . . . 49 5.2.5. Attribute 78: space_freed . . . . . . . . . . . . . . 49 5.2.6. Attribute 79: max_hole_punch . . . . . . . . . . . . . 49 5.2.7. Operation 64: HOLE_PUNCH - Zero and deallocate blocks backing the file in the specified range. . . . 50 5.3. Security Considerations . . . . . . . . . . . . . . . . . 51 5.4. IANA Considerations . . . . . . . . . . . . . . . . . . . 51 6. Simple and Efficient Read Support for Sparse Files . . . . . . 51 6.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . 51 6.2. Terminology . . . . . . . . . . . . . . . . . . . . . . . 52 6.3. Applications and Sparse Files . . . . . . . . . . . . . . 52 6.4. Overview of Sparse Files and NFSv4 . . . . . . . . . . . . 53 6.5. Operation 65: READPLUS . . . . . . . . . . . . . . . . . . 54 6.5.1. ARGUMENT . . . . . . . . . . . . . . . . . . . . . . . 55 6.5.2. RESULT . . . . . . . . . . . . . . . . . . . . . . . . 55 6.5.3. DESCRIPTION . . . . . . . . . . . . . . . . . . . . . 55 6.5.4. IMPLEMENTATION . . . . . . . . . . . . . . . . . . . . 57 6.5.5. READPLUS with Sparse Files Example . . . . . . . . . . 58 6.6. Related Work . . . . . . . . . . . . . . . . . . . . . . . 59 6.7. Security Considerations . . . . . . . . . . . . . . . . . 59 6.8. IANA Considerations . . . . . . . . . . . . . . . . . . . 59 7. Security Considerations . . . . . . . . . . . . . . . . . . . 60 8. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 60 9. References . . . . . . . . . . . . . . . . . . . . . . . . . . 60 9.1. Normative References . . . . . . . . . . . . . . . . . . . 60 9.2. Informative References . . . . . . . . . . . . . . . . . . 60 Appendix A. Acknowledgments . . . . . . . . . . . . . . . . . . . 62 Appendix B. RFC Editor Notes . . . . . . . . . . . . . . . . . . 62 Author's Address . . . . . . . . . . . . . . . . . . . . . . . . . 62 Haynes Expires October 20, 2011 [Page 4] Internet-Draft NFSv4.2 April 2011 1. Introduction 1.1. The NFS Version 4 Minor Version 2 Protocol The NFS version 4 minor version 2 (NFSv4.2) protocol is the third minor version of the NFS version 4 (NFSv4) protocol. The first minor version, NFSv4.0, is described in [10] and the second minor version, NFSv4.1, is described in [2]. It follows the guidelines for minor versioning that are listed in Section 11 of RFC 3530bis. As a minor version, NFSv4.2 is consistent with the overall goals for NFSv4, but extends the protocol so as to better meet those goals, based on experiences with NFSv4.1. In addition, NFSv4.2 has adopted some additional goals, which motivate some of the major extensions in NFSv4.2. 1.2. Scope of This Document This document describes the NFSv4.2 protocol. With respect to NFSv4.0 and NFSv4.1, this document does not: o describe the NFSv4.0 or NFSv4.1 protocols, except where needed to contrast with NFSv4.2. o modify the specification of the NFSv4.0 or NFSv4.1 protocols. o clarify the NFSv4.0 or NFSv4.1 protocols. 1.3. NFSv4.2 Goals 1.4. Overview of NFSv4.2 Features 1.5. Differences from NFSv4.1 2. pNFS Access Permissions Check 2.1. Introduction Figure 1 shows the overall architecture of a Parallel NFS (pNFS) system: Haynes Expires October 20, 2011 [Page 5] Internet-Draft NFSv4.2 April 2011 +-----------+ |+-----------+ +-----------+ ||+-----------+ | | ||| | NFSv4.1 + pNFS | | +|| Clients |<------------------------------>| MDS | +| | | | +-----------+ | | ||| +-----------+ ||| | ||| | ||| Storage +-----------+ | ||| Protocol |+-----------+ | ||+----------------||+-----------+ Control | |+-----------------||| | Protocol | +------------------+|| Storage |------------+ +| Devices | +-----------+ Figure 1: pNFS Architecture In this document, "storage device" is used as a general term for a data server and/or storage server for the file, block, or object pNFS layouts. The current pNFS protocol [2] assumes that a client can access every storage device (SD) included in a valid layout sent by the MDS server, and provides no means to communicate client access failures to the MDS. Access failures can impair pNFS performance scaling and allow significant errors to go unreported. If the MDS can access all the storage devices involved, but the client doesn't have sufficient access rights to some storage devices, the client may choose to fall back to accessing the file system using NFSV4.1 without pNFS support; there are environments in which this behavior is undesirable, especially if it occurs silently. An important example is addition of a new storage device to which a large population of pNFS clients (e.g., 1000s) lacks access permission. Layouts granted that use this new device, result in client errors, requiring that all I/Os to that new storage device be served by the MDS server. This creates a performance and scalability bottleneck that may be difficult to detect based on I/O behavior because the other storage devices are functioning correctly. The preferable approach to this scenario is to report the access failures before any client attempts to issue any I/Os that can only be serviced by the MDS server. This makes the problem explicit, rather than forcing the MDS, or a system administrator, to diagnose the performance problem caused by client I/O using NFS instead of pNFS. There are limits to this approach because complex mount Haynes Expires October 20, 2011 [Page 6] Internet-Draft NFSv4.2 April 2011 structures may prevent a client from detecting this situation at mount time, but at a minimum, access problems involving the root of the mount structure can be detected. The most suitable time for the client to report inability to access a storage device is at mount time, but this is not always possible. If the application uses a special tag or a switch to the mount command (e.g., -pnfs) and syscall to declare its intention to use pNFS, at the client, the client can check for both pNFS support and device accessibility. This document introduces an error reporting mechanism that is an extension to the return of a pNFS layout; a pNFS client MAY use this mechanism to inform the MDS that the layout is being returned because one or more data servers are not accessible to the client. Error reporting at I/O time is not affected because the result of an inaccessible data server may not be an I/O error if a subsequent retry of the operation via the MDS is successful. There is a related problem scenario involving an MDS that cannot access some storage devices and hence cannot perform I/Os on behalf of a client. In the case of the block layout [3] if the MDS lacks access to a storage device (e.g., LUN), MDS implementations generally do not export any filesystem using that storage device. In contrast to the block layout, MDSs for the file [2] and object [4] layouts may be unable to access the storage devices that store data for an exported filesystem. This enables a file or object layout MDS to provide layouts that contain client-inaccessible devices. For the specific case of adding a new storage device to a filesystem, MDS issuance of test I/Os to the newly added device before using it in layouts avoids this problem scenario, but does not cover loss of access to existing storage devices at a later time. In addition, [2] states that a client can write through or read from the MDS, even if it has a layout; this assumes that the MDS can access all the storage devices. This document makes that assumed access an explicit requirement. 2.2. Changes to Operation 51: LAYOUTRETURN (RFC 5661) The existing LAYOUTRETURN operation is extended by introducing three new layout return types that correspond to the existing types: o LAYOUT4_RET_REC_FILE_NO_ACCESS at file scope; o LAYOUT4_RET_REC_FSID_NO_ACCESS at fsid scope; and Haynes Expires October 20, 2011 [Page 7] Internet-Draft NFSv4.2 April 2011 o LAYOUT4_RET_REC_ALL_NO_ACCESS at client scope. The first return type returns the layout for an individual file and informs the server that the reason for the return is a storage device connectivity problem. The second return type performs that function for all layouts held by the client for the filesystem that corresponds to the current filehandle used for the LAYOUTRETURN operation. The third return type performs that function for all layouts held by the client; it is intended for situations in which a device is shared across all or most of the filesystems from a server for which the client has layouts. 2.2.1. ARGUMENT (18.44.1) The ARGUMENT specification of the LAYOUTRETURN operation in section 18.44.1 of [2] is replaced by the following XDR code [11]: /* Constants used for new LAYOUTRETURN and CB_LAYOUTRECALL */ const LAYOUT4_RET_REC_FILE = 1; const LAYOUT4_RET_REC_FSID = 2; const LAYOUT4_RET_REC_ALL = 3; const LAYOUT4_RET_REC_FILE_NO_ACCESS = 4; const LAYOUT4_RET_REC_FSID_NO_ACESSS = 5; const LAYOUT4_RET_REC_ALL_NO_ACCESS = 6; enum layoutreturn_type4 { LAYOUTRETURN4_FILE = LAYOUT4_RET_REC_FILE, LAYOUTRETURN4_FSID = LAYOUT4_RET_REC_FSID, LAYOUTRETURN4_ALL = LAYOUT4_RET_REC_ALL, LAYOUTRETURN4_FILE_NO_ACCESS = LAYOUT4_RET_REC_FILE_NO_ACCESS, LAYOUTRETURN4_FSID_NO_ACCESS = LAYOUT4_RET_REC_FSID_NO_ACCESS, LAYOUTRETURN4_ALL_NO_ACCESS = LAYOUT4_RET_REC_ALL_NO_ACCESS }; struct layoutreturn_file4 { offset4 lrf_offset; length4 lrf_length; stateid4 lrf_stateid; /* layouttype4 specific data */ opaque lrf_body<>; }; struct layoutreturn_device_no_access4 { deviceid4 lrdna_deviceid; nfsstat4 lrdna_status; }; struct layoutreturn_file_no_access4 { Haynes Expires October 20, 2011 [Page 8] Internet-Draft NFSv4.2 April 2011 offset4 lrfna_offset; length4 lrfna_length; stateid4 lrfna_stateid; deviceid4 lrfna_deviceid; nfsstat4 lrfna_status; /* layouttype4 specific data */ opaque lrfna_body<>; }; union layoutreturn4 switch(layoutreturn_type4 lr_returntype) { case LAYOUTRETURN4_FILE: layoutreturn_file4 lr_layout; case LAYOUTRETURN4_FILE_NO_ACCESS: layoutreturn_file_no_access4 lr_layout_na; case LAYOUTRETURN4_FSID_NO_ACCESS: case LAYOUTRETURN4_ALL_NO_ACCESS: layoutreturn_device_no_access4 lr_device<>; default: void; }; 2.2.2. RESULT (18.44.2) The RESULT of the LAYOUTRETURN operation is unchanged; see section 18.44.2 of [2] 2.2.3. DESCRIPTION (18.44.3) The following text is added to the end of the LAYOUTRETURN operation DESCRIPTION in section 18.44.3 of [2] There are three NO_ACCESS layoutreturn_type4 values that indicate a persistent lack of client ability to access storage device(s), LAYOUT4_RET_REC_FILE_NO_ACCESS, LAYOUT4_RET_REC_FSID_NO_ACCESS and LAYOUT4_RET_REC_ALL_NO_ACCESS. A client uses these return types to return a layout (or portion thereof) for a file, return all layouts for an FSID or all layouts from that server held by the client, and in all cases to inform the server that the reason for the return is the client's inability to access one or more storage devices. The same stateid may be used or the client MAY force use of a new stateid in order to report a new error. An NFS error value (nfsstat4) is included for each device for these three NO_ACCESS return types to provide additional information on the cause. The allowed NFS errors are those that are valid for an NFS READ or WRITE operation, and NFS4ERR_NXIO is also allowed to report an inaccessible device. The server SHOULD log the received NFS error value, but that error value does not affect server processing of the Haynes Expires October 20, 2011 [Page 9] Internet-Draft NFSv4.2 April 2011 LAYOUTRETURN operation. All uses of the NO_ACCESS layout return types that report NFS errors SHOULD be logged by the client. The client MAY use the new LAYOUT4_RET_REC_FILE_NO_ACCESS when only one file, or a small number of files are affected. If the access problem affects multiple devices, the client may use multiple file layout return operations; each return operation SHOULD return a layout extent obtained from the device for which an error is being reported. In contrast, both LAYOUT4_RET_REC_FSID_NO_ACCESS and LAYOUT4_RET_REC_ALL_NO_ACCESS include an array of pairs to enable a single operation to report errors for multiple devices in a single operation. 2.2.4. IMPLEMENTATION (18.44.4) The following text is added to the end of the LAYOUTRETURN operation IMPLEMENTATION in section 18.4.4 of [2] A client that expects to use pNFS for a mounted filesystem SHOULD check for pNFS support at mount time. This check SHOULD be performed by sending a GETDEVICELIST operation, followed by layout-type- specific checks for accessibility of each storage device returned by GETDEVICELIST. If the NFS server does not support pNFS, the GETDEVICELIST operation will be rejected with an NFS4ERR_NOTSUPP error; in this situation it is up to the client to determine whether it is acceptable to proceed with NFS-only access. Clients are expected to tolerate transient storage device errors, and hence clients SHOULD NOT use the NO_ACCESS layout return types for device access problems that may be transient. The methods by which a client decides whether an access problem is transient vs. persistent are implementation-specific, but may include retrying I/Os to a data server under appropriate conditions. When an I/O fails because a storage device is inaccessible, the client SHOULD retry the failed I/O via the MDS. In this situation, before retrying the I/O, the client SHOULD return the layout, or inaccessible portion thereof, and SHOULD indicate which storage device or devices was or were inaccessible. If the client does not do this, the MDS may issue a layout recall callback in order to perform the retried I/O. Backwards compatibility may require a client to perform two layout return operations to deal with servers that don't implement the NO_ACCESS layoutreturn_type4 values and hence respond to them with NFS4ERR_INVAL. In this situation, the client SHOULD perform an ordinary layout return operation and remember that the new layout NO_ACCESS return types are not to be used with that server. Haynes Expires October 20, 2011 [Page 10] Internet-Draft NFSv4.2 April 2011 The metadata server (MDS) SHOULD NOT use storage devices in pNFS layouts that are not accessible to the MDS. At a minimum, the server SHOULD check its own storage device accessibility before exporting a filesystem that supports pNFS and when the device configuration for such an exported filesystem is changed (e.g., to add a storage device). If an MDS is aware that a storage device is inaccessible to a client, the MDS SHOULD NOT include that storage device in any pNFS layouts sent to that client. An MDS SHOULD react to a client return of inaccessible layouts by not using the inaccessible storage devices in layouts for that client, but the MDS is not required to indefinitely retain per-client storage device inaccessibility information. An MDS is also not required to automatically reinstate use of a previously inaccessible storage device; administrative intervention may be required instead. A client MAY perform I/O via the MDS even when the client holds a layout that covers the I/O; servers MUST support this client behavior, and MAY recall layouts as needed to complete I/Os. 2.2.4.1. Storage Device Error Mapping (18.44.4.1, new) The following text is added as new subsection 18.44.4.1 of [2] An NFS error value is sent for each device that the client reports as inaccessible via a NO_ACCESS layout return type. In general: o If the client is unable to access the storage device, NFS4ERR_NXIO SHOULD be used. o If the client is able to access the storage device, but permission is denied, NFS4ERR_ACCESS SHOULD be used. Beyond these two rules, error code usage is layout-type specific: o For the pNFS file layout, an indicative NFS error from a failed read or write operation on the inaccessible device SHOULD be used. o For the pNFS block layout, other errors from the Storage Protocol SHOULD be mapped to NFS4ERR_IO. In addition, the client SHOULD log information about the actual storage protocol error (e.g., SCSI status and sense data), but that information is not sent to the pNFS server. o For the pNFS object layout, occurrences of the object error types specified in [4] SHOULD be mapped to the following NFS errors for use in LAYOUTRETURN: Haynes Expires October 20, 2011 [Page 11] Internet-Draft NFSv4.2 April 2011 * PNFS_OSD_ERR_EIO -> NFS4ERR_IO * PNFS_OSD_ERR_NOT_FOUND -> NFS4ERR_STALE * PNFS_OSD_ERR_NO_SPACE -> NFS4ERR_NOSPC * PNFS_OSD_ERR_BAD_CRED -> NFS4ERR_INVAL * PNFS_OSD_ERR_NO_ACCESS -> NFS4ERR_ACCESS * PNFS_OSD_ERR_UNREACHABLE -> NFS4ERR_NXIO * PNFS_OSD_ERR_RESOURCE -> NFS4ERR_SERVERFAULT The LAYOUTRETURN NO_ACCESS return types are used for persistent device errors; they do not replace other error reporting mechanisms that also apply to transient errors (e.g., as specified for the object layout in [4]). 2.3. Change to NFS4ERR_NXIO Usage This document specifies that the NFS4ERR_NXIO error SHOULD be used to report an inaccessible storage device. To enable that usage, this document updates [2] to allow use of the currently obsolete NFS4ERR_NXIO error in the ARGUMENT of LAYOUTRETURN; NFS4ERR_NXIO remains obsolete for all other uses of NFS errors. 2.4. Security Considerations This section adds a small extension to the NFSv4 LAYOUTRETURN operation. The NFS and pNFS security considerations in [2], [3], and [4] apply to the extended LAYOUTRETURN operation. 2.5. IANA Considerations There are no additional IANA considerations in this section beyond the IANA Considerations covered in [2] 3. Sharing change attribute implementation details with NFSv4 clients 3.1. Abstract This document describes an extension to the NFSv4 protocol that allows the server to share information about the implementation of its change attribute with the client. The aim is to improve the client's ability to determine the order in which parallel updates to the same file were processed. Haynes Expires October 20, 2011 [Page 12] Internet-Draft NFSv4.2 April 2011 3.2. Introduction Although both the NFSv4 [10] and NFSv4.1 protocol [2], define the change attribute as being mandatory to implement, there is little in the way of guidance. The only feature that is mandated by the spec is that the value must change whenever the file data or metadata change. While this allows for a wide range of implementations, it also leaves the client with a conundrum: how does it determine which is the most recent value for the change attribute in a case where several RPC calls have been issued in parallel? In other words if two COMPOUNDs, both containing WRITE and GETATTR requests for the same file, have been issued in parallel, how does the client determine which of the two change attribute values returned in the replies to the GETATTR requests corresponds to the most recent state of the file? In some cases, the only recourse may be to send another COMPOUND containing a third GETATTR that is fully serialised with the first two. In order to avoid this kind of inefficiency, we propose a method to allow the server to share details about how the change attribute is expected to evolve, so that the client may immediately determine which, out of the several change attribute values returned by the server, is the most recent. 3.3. Definition of the 'change_attr_type' per-file system attribute enum change_attr_typeinfo = { NFS4_CHANGE_TYPE_IS_MONOTONIC_INCR = 0, NFS4_CHANGE_TYPE_IS_VERSION_COUNTER = 1, NFS4_CHANGE_TYPE_IS_VERSION_COUNTER_NOPNFS = 2, NFS4_CHANGE_TYPE_IS_TIME_METADATA = 3, NFS4_CHANGE_TYPE_IS_UNDEFINED = 4 }; +------------------+----+---------------------------+-----+ | Name | Id | Data Type | Acc | +------------------+----+---------------------------+-----+ | change_attr_type | XX | enum change_attr_typeinfo | R | +------------------+----+---------------------------+-----+ The proposed solution is to enable the NFS server to provide additional information about how it expects the change attribute value to evolve after the file data or metadata has changed. To do so, we define a new recommended attribute, 'change_attr_type', which may take values from enum change_attr_typeinfo as follows: Haynes Expires October 20, 2011 [Page 13] Internet-Draft NFSv4.2 April 2011 NFS4_CHANGE_TYPE_IS_MONOTONIC_INCR: The change attribute value MUST monotonically increase for every atomic change to the file attributes, data or directory contents. NFS4_CHANGE_TYPE_IS_VERSION_COUNTER: The change attribute value MUST be incremented by one unit for every atomic change to the file attributes, data or directory contents. This property is preserved when writing to pNFS data servers. NFS4_CHANGE_TYPE_IS_VERSION_COUNTER_NOPNFS: The change attribute value MUST be incremented by one unit for every atomic change to the file attributes, data or directory contents. In the case where the client is writing to pNFS data servers, the number of increments is not guaranteed to exactly match the number of writes. NFS4_CHANGE_TYPE_IS_TIME_METADATA: The change attribute is implemented as suggested in the NFSv4 spec [10] in terms of the time_metadata attribute. NFS4_CHANGE_TYPE_IS_UNDEFINED: The change attribute does not take values that fit into any of these categories. If either NFS4_CHANGE_TYPE_IS_MONOTONIC_INCR, NFS4_CHANGE_TYPE_IS_VERSION_COUNTER, or NFS4_CHANGE_TYPE_IS_TIME_METADATA are set, then the client knows at the very least that the change attribute is monotonically increasing, which is sufficient to resolve the question of which value is the most recent. If the client sees the value NFS4_CHANGE_TYPE_IS_TIME_METADATA, then by inspecting the value of the 'time_delta' attribute it additionally has the option of detecting rogue server implementations that use time_metadata in violation of the spec. Finally, if the client sees NFS4_CHANGE_TYPE_IS_VERSION_COUNTER, it has the ability to predict what the resulting change attribute value should be after a COMPOUND containing a SETATTR, WRITE, or CREATE. This again allows it to detect changes made in parallel by another client. The value NFS4_CHANGE_TYPE_IS_VERSION_COUNTER_NOPNFS permits the same, but only if the client is not doing pNFS WRITEs. 4. NFS Server-side Copy Haynes Expires October 20, 2011 [Page 14] Internet-Draft NFSv4.2 April 2011 4.1. Introduction This document describes a server-side copy feature for the NFS protocol. The server-side copy feature provides a mechanism for the NFS client to perform a file copy on the server without the data being transmitted back and forth over the network. Without this feature, an NFS client copies data from one location to another by reading the data from the server over the network, and then writing the data back over the network to the server. Using this server-side copy operation, the client is able to instruct the server to copy the data locally without the data being sent back and forth over the network unnecessarily. In general, this feature is useful whenever data is copied from one location to another on the server. It is particularly useful when copying the contents of a file from a backup. Backup-versions of a file are copied for a number of reasons, including restoring and cloning data. If the source object and destination object are on different file servers, the file servers will communicate with one another to perform the copy operation. The server-to-server protocol by which this is accomplished is not defined in this document. 4.2. Protocol Overview The server-side copy offload operations support both intra-server and inter-server file copies. An intra-server copy is a copy in which the source file and destination file reside on the same server. In an inter-server copy, the source file and destination file are on different servers. In both cases, the copy may be performed synchronously or asynchronously. Throughout the rest of this document, we refer to the NFS server containing the source file as the "source server" and the NFS server to which the file is transferred as the "destination server". In the case of an intra-server copy, the source server and destination server are the same server. Therefore in the context of an intra- server copy, the terms source server and destination server refer to the single server performing the copy. The operations described below are designed to copy files. Other file system objects can be copied by building on these operations or using other techniques. For example if the user wishes to copy a directory, the client can synthesize a directory copy by first Haynes Expires October 20, 2011 [Page 15] Internet-Draft NFSv4.2 April 2011 creating the destination directory and then copying the source directory's files to the new destination directory. If the user wishes to copy a namespace junction [12] [13], the client can use the ONC RPC Federated Filesystem protocol [13] to perform the copy. Specifically the client can determine the source junction's attributes using the FEDFS_LOOKUP_FSN procedure and create a duplicate junction using the FEDFS_CREATE_JUNCTION procedure. For the inter-server copy protocol, the operations are defined to be compatible with a server-to-server copy protocol in which the destination server reads the file data from the source server. This model in which the file data is pulled from the source by the destination has a number of advantages over a model in which the source pushes the file data to the destination. The advantages of the pull model include: o The pull model only requires a remote server (i.e. the destination server) to be granted read access. A push model requires a remote server (i.e. the source server) to be granted write access, which is more privileged. o The pull model allows the destination server to stop reading if it has run out of space. In a push model, the destination server must flow control the source server in this situation. o The pull model allows the destination server to easily flow control the data stream by adjusting the size of its read operations. In a push model, the destination server does not have this ability. The source server in a push model is capable of writing chunks larger than the destination server has requested in attributes and session parameters. In theory, the destination server could perform a "short" write in this situation, but this approach is known to behave poorly in practice. The following operations are provided to support server-side copy: COPY_NOTIFY: For inter-server copies, the client sends this operation to the source server to notify it of a future file copy from a given destination server for the given user. COPY_REVOKE: Also for inter-server copies, the client sends this operation to the source server to revoke permission to copy a file for the given user. COPY: Used by the client to request a file copy. Haynes Expires October 20, 2011 [Page 16] Internet-Draft NFSv4.2 April 2011 COPY_ABORT: Used by the client to abort an asynchronous file copy. COPY_STATUS: Used by the client to poll the status of an asynchronous file copy. CB_COPY: Used by the destination server to report the results of an asynchronous file copy to the client. These operations are described in detail in Section 4.3. This section provides an overview of how these operations are used to perform server-side copies. 4.2.1. Intra-Server Copy To copy a file on a single server, the client uses a COPY operation. The server may respond to the copy operation with the final results of the copy or it may perform the copy asynchronously and deliver the results using a CB_COPY operation callback. If the copy is performed asynchronously, the client may poll the status of the copy using COPY_STATUS or cancel the copy using COPY_ABORT. A synchronous intra-server copy is shown in Figure 2. In this example, the NFS server chooses to perform the copy synchronously. The copy operation is completed, either successfully or unsuccessfully, before the server replies to the client's request. The server's reply contains the final result of the operation. Client Server + + | | |--- COPY ---------------------------->| Client requests |<------------------------------------/| a file copy | | | | Figure 2: A synchronous intra-server copy. An asynchronous intra-server copy is shown in Figure 3. In this example, the NFS server performs the copy asynchronously. The server's reply to the copy request indicates that the copy operation was initiated and the final result will be delivered at a later time. The server's reply also contains a copy stateid. The client may use this copy stateid to poll for status information (as shown) or to cancel the copy using a COPY_ABORT. When the server completes the copy, the server performs a callback to the client and reports the results. Haynes Expires October 20, 2011 [Page 17] Internet-Draft NFSv4.2 April 2011 Client Server + + | | |--- COPY ---------------------------->| Client requests |<------------------------------------/| a file copy | | | | |--- COPY_STATUS --------------------->| Client may poll |<------------------------------------/| for status | | | . | Multiple COPY_STATUS | . | operations may be sent. | . | | | |<-- CB_COPY --------------------------| Server reports results |\------------------------------------>| | | Figure 3: An asynchronous intra-server copy. 4.2.2. Inter-Server Copy A copy may also be performed between two servers. The copy protocol is designed to accommodate a variety of network topologies. As shown in Figure 4, the client and servers may be connected by multiple networks. In particular, the servers may be connected by a specialized, high speed network (network 192.168.33.0/24 in the diagram) that does not include the client. The protocol allows the client to setup the copy between the servers (over network 10.11.78.0/24 in the diagram) and for the servers to communicate on the high speed network if they choose to do so. Haynes Expires October 20, 2011 [Page 18] Internet-Draft NFSv4.2 April 2011 192.168.33.0/24 +-------------------------------------+ | | | | | 192.168.33.18 | 192.168.33.56 +-------+------+ +------+------+ | Source | | Destination | +-------+------+ +------+------+ | 10.11.78.18 | 10.11.78.56 | | | | | 10.11.78.0/24 | +------------------+------------------+ | | | 10.11.78.243 +-----+-----+ | Client | +-----------+ Figure 4: An example inter-server network topology. For an inter-server copy, the client notifies the source server that a file will be copied by the destination server using a COPY_NOTIFY operation. The client then initiates the copy by sending the COPY operation to the destination server. The destination server may perform the copy synchronously or asynchronously. A synchronous inter-server copy is shown in Figure 5. In this case, the destination server chooses to perform the copy before responding to the client's COPY request. An asynchronous copy is shown in Figure 6. In this case, the destination server chooses to respond to the client's COPY request immediately and then perform the copy asynchronously. Haynes Expires October 20, 2011 [Page 19] Internet-Draft NFSv4.2 April 2011 Client Source Destination + + + | | | |--- COPY_NOTIFY --->| | |<------------------/| | | | | | | | |--- COPY ---------------------------->| | | | | | | | |<----- read -----| | |\--------------->| | | | | | . | Multiple reads may | | . | be necessary | | . | | | | | | | |<------------------------------------/| Destination replies | | | to COPY Figure 5: A synchronous inter-server copy. Haynes Expires October 20, 2011 [Page 20] Internet-Draft NFSv4.2 April 2011 Client Source Destination + + + | | | |--- COPY_NOTIFY --->| | |<------------------/| | | | | | | | |--- COPY ---------------------------->| |<------------------------------------/| | | | | | | | |<----- read -----| | |\--------------->| | | | | | . | Multiple reads may | | . | be necessary | | . | | | | | | | |--- COPY_STATUS --------------------->| Client may poll |<------------------------------------/| for status | | | | | . | Multiple COPY_STATUS | | . | operations may be sent | | . | | | | | | | | | | |<-- CB_COPY --------------------------| Destination reports |\------------------------------------>| results | | | Figure 6: An asynchronous inter-server copy. 4.2.3. Server-to-Server Copy Protocol During an inter-server copy, the destination server reads the file data from the source server. The source server and destination server are not required to use a specific protocol to transfer the file data. The choice of what protocol to use is ultimately the destination server's decision. 4.2.3.1. Using NFSv4.x as a Server-to-Server Copy Protocol The destination server MAY use standard NFSv4.x (where x >= 1) to read the data from the source server. If NFSv4.x is used for the server-to-server copy protocol, the destination server can use the filehandle contained in the COPY request with standard NFSv4.x Haynes Expires October 20, 2011 [Page 21] Internet-Draft NFSv4.2 April 2011 operations to read data from the source server. Specifically, the destination server may use the NFSv4.x OPEN operation's CLAIM_FH facility to open the file being copied and obtain an open stateid. Using the stateid, the destination server may then use NFSv4.x READ operations to read the file. 4.2.3.2. Using an alternative Server-to-Server Copy Protocol In a homogeneous environment, the source and destination servers might be able to perform the file copy extremely efficiently using specialized protocols. For example the source and destination servers might be two nodes sharing a common file system format for the source and destination file systems. Thus the source and destination are in an ideal position to efficiently render the image of the source file to the destination file by replicating the file system formats at the block level. Another possibility is that the source and destination might be two nodes sharing a common storage area network, and thus there is no need to copy any data at all, and instead ownership of the file and its contents might simply be re- assigned to the destination. To allow for these possibilities, the destination server is allowed to use a server-to-server copy protocol of its choice. In a heterogeneous environment, using a protocol other than NFSv4.x (e.g. HTTP [14] or FTP [15]) presents some challenges. In particular, the destination server is presented with the challenge of accessing the source file given only an NFSv4.x filehandle. One option for protocols that identify source files with path names is to use an ASCII hexadecimal representation of the source filehandle as the file name. Another option for the source server is to use URLs to direct the destination server to a specialized service. For example, the response to COPY_NOTIFY could include the URL ftp://s1.example.com:9999/_FH/0x12345, where 0x12345 is the ASCII hexadecimal representation of the source filehandle. When the destination server receives the source server's URL, it would use "_FH/0x12345" as the file name to pass to the FTP server listening on port 9999 of s1.example.com. On port 9999 there would be a special instance of the FTP service that understands how to convert NFS filehandles to an open file descriptor (in many operating systems, this would require a new system call, one which is the inverse of the makefh() function that the pre-NFSv4 MOUNT service needs). Authenticating and identifying the destination server to the source server is also a challenge. Recommendations for how to accomplish this are given in Section 4.4.1.2.4 and Section 4.4.1.4. Haynes Expires October 20, 2011 [Page 22] Internet-Draft NFSv4.2 April 2011 4.3. Operations In the sections that follow, several operations are defined that together provide the server-side copy feature. These operations are intended to be OPTIONAL operations as defined in section 17 of [2]. The COPY_NOTIFY, COPY_REVOKE, COPY, COPY_ABORT, and COPY_STATUS operations are designed to be sent within an NFSv4 COMPOUND procedure. The CB_COPY operation is designed to be sent within an NFSv4 CB_COMPOUND procedure. Each operation is performed in the context of the user identified by the ONC RPC credential of its containing COMPOUND or CB_COMPOUND request. For example, a COPY_ABORT operation issued by a given user indicates that a specified COPY operation initiated by the same user be canceled. Therefore a COPY_ABORT MUST NOT interfere with a copy of the same file initiated by another user. An NFS server MAY allow an administrative user to monitor or cancel copy operations using an implementation specific interface. 4.3.1. netloc4 - Network Locations The server-side copy operations specify network locations using the netloc4 data type shown below: enum netloc_type4 { NL4_NAME = 0, NL4_URL = 1, NL4_NETADDR = 2 }; union netloc4 switch (netloc_type4 nl_type) { case NL4_NAME: utf8str_cis nl_name; case NL4_URL: utf8str_cis nl_url; case NL4_NETADDR: netaddr4 nl_addr; }; If the netloc4 is of type NL4_NAME, the nl_name field MUST be specified as a UTF-8 string. The nl_name is expected to be resolved to a network address via DNS, LDAP, NIS, /etc/hosts, or some other means. If the netloc4 is of type NL4_URL, a server URL [5] appropriate for the server-to-server copy operation is specified as a UTF-8 string. If the netloc4 is of type NL4_NETADDR, the nl_addr field MUST contain a valid netaddr4 as defined in Section 3.3.9 of [2]. When netloc4 values are used for an inter-server copy as shown in Figure 4, their values may be evaluated on the source server, destination server, and client. The network environment in which Haynes Expires October 20, 2011 [Page 23] Internet-Draft NFSv4.2 April 2011 these systems operate should be configured so that the netloc4 values are interpreted as intended on each system. 4.3.2. Operation 61: COPY_NOTIFY - Notify a source server of a future copy 4.3.2.1. ARGUMENT struct COPY_NOTIFY4args { /* CURRENT_FH: source file */ netloc4 cna_destination_server; }; 4.3.2.2. RESULT union COPY_NOTIFY4res switch (nfsstat4 cnr_status) { case NFS4_OK: nfstime4 cnr_lease_time; netloc4 cnr_source_server<>; default: void; }; 4.3.2.3. DESCRIPTION This operation is used for an inter-server copy. A client sends this operation in a COMPOUND request to the source server to authorize a destination server identified by cna_destination_server to read the file specified by CURRENT_FH on behalf of the given user. The cna_destination_server MUST be specified using the netloc4 network location format. The server is not required to resolve the cna_destination_server address before completing this operation. If this operation succeeds, the source server will allow the cna_destination_server to copy the specified file on behalf of the given user. If COPY_NOTIFY succeeds, the destination server is granted permission to read the file as long as both of the following conditions are met: o The destination server begins reading the source file before the cnr_lease_time expires. If the cnr_lease_time expires while the destination server is still reading the source file, the destination server is allowed to finish reading the file. Haynes Expires October 20, 2011 [Page 24] Internet-Draft NFSv4.2 April 2011 o The client has not issued a COPY_REVOKE for the same combination of user, filehandle, and destination server. The cnr_lease_time is chosen by the source server. A cnr_lease_time of 0 (zero) indicates an infinite lease. To renew the copy lease time the client should resend the same copy notification request to the source server. To avoid the need for synchronized clocks, copy lease times are granted by the server as a time delta. However, there is a requirement that the client and server clocks do not drift excessively over the duration of the lease. There is also the issue of propagation delay across the network which could easily be several hundred milliseconds as well as the possibility that requests will be lost and need to be retransmitted. To take propagation delay into account, the client should subtract it from copy lease times (e.g. if the client estimates the one-way propagation delay as 200 milliseconds, then it can assume that the lease is already 200 milliseconds old when it gets it). In addition, it will take another 200 milliseconds to get a response back to the server. So the client must send a lease renewal or send the copy offload request to the cna_destination_server at least 400 milliseconds before the copy lease would expire. If the propagation delay varies over the life of the lease (e.g. the client is on a mobile host), the client will need to continuously subtract the increase in propagation delay from the copy lease times. The server's copy lease period configuration should take into account the network distance of the clients that will be accessing the server's resources. It is expected that the lease period will take into account the network propagation delays and other network delay factors for the client population. Since the protocol does not allow for an automatic method to determine an appropriate copy lease period, the server's administrator may have to tune the copy lease period. A successful response will also contain a list of names, addresses, and URLs called cnr_source_server, on which the source is willing to accept connections from the destination. These might not be reachable from the client and might be located on networks to which the client has no connection. If the client wishes to perform an inter-server copy, the client MUST send a COPY_NOTIFY to the source server. Therefore, the source server MUST support COPY_NOTIFY. For a copy only involving one server (the source and destination are Haynes Expires October 20, 2011 [Page 25] Internet-Draft NFSv4.2 April 2011 on the same server), this operation is unnecessary. The COPY_NOTIFY operation may fail for the following reasons (this is a partial list): NFS4ERR_MOVED: The file system which contains the source file is not present on the source server. The client can determine the correct location and reissue the operation with the correct location. NFS4ERR_NOTSUPP: The copy offload operation is not supported by the NFS server receiving this request. NFS4ERR_WRONGSEC: The security mechanism being used by the client does not match the server's security policy. 4.3.3. Operation 62: COPY_REVOKE - Revoke a destination server's copy privileges 4.3.3.1. ARGUMENT struct COPY_REVOKE4args { /* CURRENT_FH: source file */ netloc4 cra_destination_server; }; 4.3.3.2. RESULT struct COPY_REVOKE4res { nfsstat4 crr_status; }; 4.3.3.3. DESCRIPTION This operation is used for an inter-server copy. A client sends this operation in a COMPOUND request to the source server to revoke the authorization of a destination server identified by cra_destination_server from reading the file specified by CURRENT_FH on behalf of given user. If the cra_destination_server has already begun copying the file, a successful return from this operation indicates that further access will be prevented. The cra_destination_server MUST be specified using the netloc4 network location format. The server is not required to resolve the cra_destination_server address before completing this operation. The COPY_REVOKE operation is useful in situations in which the source Haynes Expires October 20, 2011 [Page 26] Internet-Draft NFSv4.2 April 2011 server granted a very long or infinite lease on the destination server's ability to read the source file and all copy operations on the source file have been completed. For a copy only involving one server (the source and destination are on the same server), this operation is unnecessary. If the server supports COPY_NOTIFY, the server is REQUIRED to support the COPY_REVOKE operation. The COPY_REVOKE operation may fail for the following reasons (this is a partial list): NFS4ERR_MOVED: The file system which contains the source file is not present on the source server. The client can determine the correct location and reissue the operation with the correct location. NFS4ERR_NOTSUPP: The copy offload operation is not supported by the NFS server receiving this request. 4.3.4. Operation 59: COPY - Initiate a server-side copy 4.3.4.1. ARGUMENT const COPY4_GUARDED = 0x00000001; const COPY4_METADATA = 0x00000002; struct COPY4args { /* SAVED_FH: source file */ /* CURRENT_FH: destination file or */ /* directory */ offset4 ca_src_offset; offset4 ca_dst_offset; length4 ca_count; uint32_t ca_flags; component4 ca_destination; netloc4 ca_source_server<>; }; Haynes Expires October 20, 2011 [Page 27] Internet-Draft NFSv4.2 April 2011 4.3.4.2. RESULT union COPY4res switch (nfsstat4 cr_status) { /* CURRENT_FH: destination file */ case NFS4_OK: stateid4 cr_callback_id<1>; default: length4 cr_bytes_copied; }; 4.3.4.3. DESCRIPTION The COPY operation is used for both intra- and inter-server copies. In both cases, the COPY is always sent from the client to the destination server of the file copy. The COPY operation requests that a file be copied from the location specified by the SAVED_FH value to the location specified by the combination of CURRENT_FH and ca_destination. The SAVED_FH must be a regular file. If SAVED_FH is not a regular file, the operation MUST fail and return NFS4ERR_WRONG_TYPE. In order to set SAVED_FH to the source file handle, the compound procedure requesting the COPY will include a sub-sequence of operations such as PUTFH source-fh SAVEFH If the request is for a server-to-server copy, the source-fh is a filehandle from the source server and the compound procedure is being executed on the destination server. In this case, the source-fh is a foreign filehandle on the server receiving the COPY request. If either PUTFH or SAVEFH checked the validity of the filehandle, the operation would likely fail and return NFS4ERR_STALE. In order to avoid this problem, the minor version incorporating the COPY operations will need to make a few small changes in the handling of existing operations. If a server supports the server-to-server COPY feature, a PUTFH followed by a SAVEFH MUST NOT return NFS4ERR_STALE for either operation. These restrictions do not pose substantial difficulties for servers. The CURRENT_FH and SAVED_FH may be validated in the context of the operation referencing them and an NFS4ERR_STALE error returned for an invalid file handle at that point. Haynes Expires October 20, 2011 [Page 28] Internet-Draft NFSv4.2 April 2011 The CURRENT_FH and ca_destination together specify the destination of the copy operation. If ca_destination is of 0 (zero) length, then CURRENT_FH specifies the target file. In this case, CURRENT_FH MUST be a regular file and not a directory. If ca_destination is not of 0 (zero) length, the ca_destination argument specifies the file name to which the data will be copied within the directory identified by CURRENT_FH. In this case, CURRENT_FH MUST be a directory and not a regular file. If the file named by ca_destination does not exist and the operation completes successfully, the file will be visible in the file system namespace. If the file does not exist and the operation fails, the file MAY be visible in the file system namespace depending on when the failure occurs and on the implementation of the NFS server receiving the COPY operation. If the ca_destination name cannot be created in the destination file system (due to file name restrictions, such as case or length), the operation MUST fail. The ca_src_offset is the offset within the source file from which the data will be read, the ca_dst_offset is the offset within the destination file to which the data will be written, and the ca_count is the number of bytes that will be copied. An offset of 0 (zero) specifies the start of the file. A count of 0 (zero) requests that all bytes from ca_src_offset through EOF be copied to the destination. If concurrent modifications to the source file overlap with the source file region being copied, the data copied may include all, some, or none of the modifications. The client can use standard NFS operations (e.g. OPEN with OPEN4_SHARE_DENY_WRITE or mandatory byte range locks) to protect against concurrent modifications if the client is concerned about this. If the source file's end of file is being modified in parallel with a copy that specifies a count of 0 (zero) bytes, the amount of data copied is implementation dependent (clients may guard against this case by specifying a non-zero count value or preventing modification of the source file as mentioned above). If the source offset or the source offset plus count is greater than or equal to the size of the source file, the operation will fail with NFS4ERR_INVAL. The destination offset or destination offset plus count may be greater than the size of the destination file. This allows for the client to issue parallel copies to implement operations such as "cat file1 file2 file3 file4 > dest". If the destination file is created as a result of this command, the destination file's size will be equal to the number of bytes successfully copied. If the destination file already existed, the destination file's size may increase as a result of this operation (e.g. if ca_dst_offset plus ca_count is greater than the Haynes Expires October 20, 2011 [Page 29] Internet-Draft NFSv4.2 April 2011 destination's initial size). If the ca_source_server list is specified, then this is an inter- server copy operation and the source file is on a remote server. The client is expected to have previously issued a successful COPY_NOTIFY request to the remote source server. The ca_source_server list SHOULD be the same as the COPY_NOTIFY response's cnr_source_server list. If the client includes the entries from the COPY_NOTIFY response's cnr_source_server list in the ca_source_server list, the source server can indicate a specific copy protocol for the destination server to use by returning a URL, which specifies both a protocol service and server name. Server-to-server copy protocol considerations are described in Section 4.2.3 and Section 4.4.1. The ca_flags argument allows the copy operation to be customized in the following ways using the guarded flag (COPY4_GUARDED) and the metadata flag (COPY4_METADATA). [NOTE: Earlier versions of this document defined a COPY4_SPACE_RESERVED flag for controlling space reservations on the destination file. This flag has been removed with the expectation that the space_reserve attribute defined in XXX_TDH_XXX will be adopted.] If the guarded flag is set and the destination exists on the server, this operation will fail with NFS4ERR_EXIST. If the guarded flag is not set and the destination exists on the server, the behavior is implementation dependent. If the metadata flag is set and the client is requesting a whole file copy (i.e. ca_count is 0 (zero)), a subset of the destination file's attributes MUST be the same as the source file's corresponding attributes and a subset of the destination file's attributes SHOULD be the same as the source file's corresponding attributes. The attributes in the MUST and SHOULD copy subsets will be defined for each NFS version. For NFSv4.1, Table 1 and Table 2 list the REQUIRED and RECOMMENDED attributes respectively. A "MUST" in the "Copy to destination file?" column indicates that the attribute is part of the MUST copy set. A "SHOULD" in the "Copy to destination file?" column indicates that the attribute is part of the SHOULD copy set. Haynes Expires October 20, 2011 [Page 30] Internet-Draft NFSv4.2 April 2011 +--------------------+----+---------------------------+ | Name | Id | Copy to destination file? | +--------------------+----+---------------------------+ | supported_attrs | 0 | no | | type | 1 | MUST | | fh_expire_type | 2 | no | | change | 3 | SHOULD | | size | 4 | MUST | | link_support | 5 | no | | symlink_support | 6 | no | | named_attr | 7 | no | | fsid | 8 | no | | unique_handles | 9 | no | | lease_time | 10 | no | | rdattr_error | 11 | no | | filehandle | 19 | no | | suppattr_exclcreat | 75 | no | +--------------------+----+---------------------------+ Table 1 +--------------------+----+---------------------------+ | Name | Id | Copy to destination file? | +--------------------+----+---------------------------+ | acl | 12 | MUST | | aclsupport | 13 | no | | archive | 14 | no | | cansettime | 15 | no | | case_insensitive | 16 | no | | case_preserving | 17 | no | | change_policy | 60 | no | | chown_restricted | 18 | MUST | | dacl | 58 | MUST | | dir_notif_delay | 56 | no | | dirent_notif_delay | 57 | no | | fileid | 20 | no | | files_avail | 21 | no | | files_free | 22 | no | | files_total | 23 | no | | fs_charset_cap | 76 | no | | fs_layout_type | 62 | no | | fs_locations | 24 | no | | fs_locations_info | 67 | no | | fs_status | 61 | no | | hidden | 25 | MUST | | homogeneous | 26 | no | | layout_alignment | 66 | no | | layout_blksize | 65 | no | Haynes Expires October 20, 2011 [Page 31] Internet-Draft NFSv4.2 April 2011 | layout_hint | 63 | no | | layout_type | 64 | no | | maxfilesize | 27 | no | | maxlink | 28 | no | | maxname | 29 | no | | maxread | 30 | no | | maxwrite | 31 | no | | mdsthreshold | 68 | no | | mimetype | 32 | MUST | | mode | 33 | MUST | | mode_set_masked | 74 | no | | mounted_on_fileid | 55 | no | | no_trunc | 34 | no | | numlinks | 35 | no | | owner | 36 | MUST | | owner_group | 37 | MUST | | quota_avail_hard | 38 | no | | quota_avail_soft | 39 | no | | quota_used | 40 | no | | rawdev | 41 | no | | retentevt_get | 71 | MUST | | retentevt_set | 72 | no | | retention_get | 69 | MUST | | retention_hold | 73 | MUST | | retention_set | 70 | no | | sacl | 59 | MUST | | space_avail | 42 | no | | space_free | 43 | no | | space_total | 44 | no | | space_used | 45 | no | | system | 46 | MUST | | time_access | 47 | MUST | | time_access_set | 48 | no | | time_backup | 49 | no | | time_create | 50 | MUST | | time_delta | 51 | no | | time_metadata | 52 | SHOULD | | time_modify | 53 | MUST | | time_modify_set | 54 | no | +--------------------+----+---------------------------+ Table 2 [NOTE: The space_reserve attribute XXX_TDH_XXX will be in the MUST set.] [NOTE: The source file's attribute values will take precedence over any attribute values inherited by the destination file.] Haynes Expires October 20, 2011 [Page 32] Internet-Draft NFSv4.2 April 2011 In the case of an inter-server copy or an intra-server copy between file systems, the attributes supported for the source file and destination file could be different. By definition,the REQUIRED attributes will be supported in all cases. If the metadata flag is set and the source file has a RECOMMENDED attribute that is not supported for the destination file, the copy MUST fail with NFS4ERR_ATTRNOTSUPP. Any attribute supported by the destination server that is not set on the source file SHOULD be left unset. Metadata attributes not exposed via the NFS protocol SHOULD be copied to the destination file where appropriate. The destination file's named attributes are not duplicated from the source file. After the copy process completes, the client MAY attempt to duplicate named attributes using standard NFSv4 operations. However, the destination file's named attribute capabilities MAY be different from the source file's named attribute capabilities. If the metadata flag is not set and the client is requesting a whole file copy (i.e. ca_count is 0 (zero)), the destination file's metadata is implementation dependent. If the client is requesting a partial file copy (i.e. ca_count is not 0 (zero)), the client SHOULD NOT set the metadata flag and the server MUST ignore the metadata flag. If the operation does not result in an immediate failure, the server will return NFS4_OK, and the CURRENT_FH will remain the destination's filehandle. If an immediate failure does occur, cr_bytes_copied will be set to the number of bytes copied to the destination file before the error occurred. The cr_bytes_copied value indicates the number of bytes copied but not which specific bytes have been copied. A return of NFS4_OK indicates that either the operation is complete or the operation was initiated and a callback will be used to deliver the final status of the operation. If the cr_callback_id is returned, this indicates that the operation was initiated and a CB_COPY callback will deliver the final results of the operation. The cr_callback_id stateid is termed a copy stateid in this context. The server is given the option of returning the results in a callback because the data may require a relatively long period of time to copy. Haynes Expires October 20, 2011 [Page 33] Internet-Draft NFSv4.2 April 2011 If no cr_callback_id is returned, the operation completed synchronously and no callback will be issued by the server. The completion status of the operation is indicated by cr_status. If the copy completes successfully, either synchronously or asynchronously, the data copied from the source file to the destination file MUST appear identical to the NFS client. However, the NFS server's on disk representation of the data in the source file and destination file MAY differ. For example, the NFS server might encrypt, compress, deduplicate, or otherwise represent the on disk data in the source and destination file differently. In the event of a failure the state of the destination file is implementation dependent. The COPY operation may fail for the following reasons (this is a partial list). NFS4ERR_MOVED: The file system which contains the source file, or the destination file or directory is not present. The client can determine the correct location and reissue the operation with the correct location. NFS4ERR_NOTSUPP: The copy offload operation is not supported by the NFS server receiving this request. NFS4ERR_PARTNER_NOTSUPP: The remote server does not support the server-to-server copy offload protocol. NFS4ERR_PARTNER_NO_AUTH: The remote server does not authorize a server-to-server copy offload operation. This may be due to the client's failure to send the COPY_NOTIFY operation to the remote server, the remote server receiving a server-to-server copy offload request after the copy lease time expired, or for some other permission problem. NFS4ERR_FBIG: The copy operation would have caused the file to grow beyond the server's limit. NFS4ERR_NOTDIR: The CURRENT_FH is a file and ca_destination has non- zero length. NFS4ERR_WRONG_TYPE: The SAVED_FH is not a regular file. NFS4ERR_ISDIR: The CURRENT_FH is a directory and ca_destination has zero length. Haynes Expires October 20, 2011 [Page 34] Internet-Draft NFSv4.2 April 2011 NFS4ERR_INVAL: The source offset or offset plus count are greater than or equal to the size of the source file. NFS4ERR_DELAY: The server does not have the resources to perform the copy operation at the current time. The client should retry the operation sometime in the future. NFS4ERR_METADATA_NOTSUPP: The destination file cannot support the same metadata as the source file. NFS4ERR_WRONGSEC: The security mechanism being used by the client does not match the server's security policy. 4.3.5. Operation 60: COPY_ABORT - Cancel a server-side copy 4.3.5.1. ARGUMENT struct COPY_ABORT4args { /* CURRENT_FH: desination file */ stateid4 caa_stateid; }; 4.3.5.2. RESULT struct COPY_ABORT4res { nfsstat4 car_status; }; 4.3.5.3. DESCRIPTION COPY_ABORT is used for both intra- and inter-server asynchronous copies. The COPY_ABORT operation allows the client to cancel a server-side copy operation that it initiated. This operation is sent in a COMPOUND request from the client to the destination server. This operation may be used to cancel a copy when the application that requested the copy exits before the operation is completed or for some other reason. The request contains the filehandle and copy stateid cookies that act as the context for the previously initiated copy operation. The result's car_status field indicates whether the cancel was successful or not. A value of NFS4_OK indicates that the copy operation was canceled and no callback will be issued by the server. A copy operation that is successfully canceled may result in none, some, or all of the data copied. Haynes Expires October 20, 2011 [Page 35] Internet-Draft NFSv4.2 April 2011 If the server supports asynchronous copies, the server is REQUIRED to support the COPY_ABORT operation. The COPY_ABORT operation may fail for the following reasons (this is a partial list): NFS4ERR_NOTSUPP: The abort operation is not supported by the NFS server receiving this request. NFS4ERR_RETRY: The abort failed, but a retry at some time in the future MAY succeed. NFS4ERR_COMPLETE_ALREADY: The abort failed, and a callback will deliver the results of the copy operation. NFS4ERR_SERVERFAULT: An error occurred on the server that does not map to a specific error code. 4.3.6. Operation 63: COPY_STATUS - Poll for status of a server-side copy 4.3.6.1. ARGUMENT struct COPY_STATUS4args { /* CURRENT_FH: destination file */ stateid4 csa_stateid; }; 4.3.6.2. RESULT union COPY_STATUS4res switch (nfsstat4 csr_status) { case NFS4_OK: length4 csr_bytes_copied; nfsstat4 csr_complete<1>; default: void; }; 4.3.6.3. DESCRIPTION COPY_STATUS is used for both intra- and inter-server asynchronous copies. The COPY_STATUS operation allows the client to poll the server to determine the status of an asynchronous copy operation. This operation is sent by the client to the destination server. If this operation is successful, the number of bytes copied are Haynes Expires October 20, 2011 [Page 36] Internet-Draft NFSv4.2 April 2011 returned to the client in the csr_bytes_copied field. The csr_bytes_copied value indicates the number of bytes copied but not which specific bytes have been copied. If the optional csr_complete field is present, the copy has completed. In this case the status value indicates the result of the asynchronous copy operation. In all cases, the server will also deliver the final results of the asynchronous copy in a CB_COPY operation. The failure of this operation does not indicate the result of the asynchronous copy in any way. If the server supports asynchronous copies, the server is REQUIRED to support the COPY_STATUS operation. The COPY_STATUS operation may fail for the following reasons (this is a partial list): NFS4ERR_NOTSUPP: The copy status operation is not supported by the NFS server receiving this request. NFS4ERR_BAD_STATEID: The stateid is not valid (see Section 4.3.8 below). NFS4ERR_EXPIRED: The stateid has expired (see Copy Offload Stateid section below). 4.3.7. Operation 15: CB_COPY - Report results of a server-side copy 4.3.7.1. ARGUMENT union copy_info4 switch (nfsstat4 cca_status) { case NFS4_OK: void; default: length4 cca_bytes_copied; }; struct CB_COPY4args { nfs_fh4 cca_fh; stateid4 cca_stateid; copy_info4 cca_copy_info; }; Haynes Expires October 20, 2011 [Page 37] Internet-Draft NFSv4.2 April 2011 4.3.7.2. RESULT struct CB_COPY4res { nfsstat4 ccr_status; }; 4.3.7.3. DESCRIPTION CB_COPY is used for both intra- and inter-server asynchronous copies. The CB_COPY callback informs the client of the result of an asynchronous server-side copy. This operation is sent by the destination server to the client in a CB_COMPOUND request. The copy is identified by the filehandle and stateid arguments. The result is indicated by the status field. If the copy failed, cca_bytes_copied contains the number of bytes copied before the failure occurred. The cca_bytes_copied value indicates the number of bytes copied but not which specific bytes have been copied. In the absence of an established backchannel, the server cannot signal the completion of the COPY via a CB_COPY callback. The loss of a callback channel would be indicated by the server setting the SEQ4_STATUS_CB_PATH_DOWN flag in the sr_status_flags field of the SEQUENCE operation. The client must re-establish the callback channel to receive the status of the COPY operation. Prolonged loss of the callback channel could result in the server dropping the COPY operation state and invalidating the copy stateid. If the client supports the COPY operation, the client is REQUIRED to support the CB_COPY operation. The CB_COPY operation may fail for the following reasons (this is a partial list): NFS4ERR_NOTSUPP: The copy offload operation is not supported by the NFS client receiving this request. 4.3.8. Copy Offload Stateids A server may perform a copy offload operation asynchronously. An asynchronous copy is tracked using a copy offload stateid. Copy offload stateids are included in the COPY, COPY_ABORT, COPY_STATUS, and CB_COPY operations. Section 8.2.4 of [2] specifies that stateids are valid until either (A) the client or server restart or (B) the client returns the resource. A copy offload stateid will be valid until either (A) the client or Haynes Expires October 20, 2011 [Page 38] Internet-Draft NFSv4.2 April 2011 server restart or (B) the client returns the resource by issuing a COPY_ABORT operation or the client replies to a CB_COPY operation. A copy offload stateid's seqid MUST NOT be 0 (zero). In the context of a copy offload operation, it is ambiguous to indicate the most recent copy offload operation using a stateid with seqid of 0 (zero). Therefore a copy offload stateid with seqid of 0 (zero) MUST be considered invalid. 4.4. Security Considerations The security considerations pertaining to NFSv4 [10] apply to this document. The standard security mechanisms provide by NFSv4 [10] may be used to secure the protocol described in this document. NFSv4 clients and servers supporting the the inter-server copy operations described in this document are REQUIRED to implement [6], including the RPCSEC_GSSv3 privileges copy_from_auth and copy_to_auth. If the server-to-server copy protocol is ONC RPC based, the servers are also REQUIRED to implement the RPCSEC_GSSv3 privilege copy_confirm_auth. These requirements to implement are not requirements to use. NFSv4 clients and servers are RECOMMENDED to use [6] to secure server-side copy operations. 4.4.1. Inter-Server Copy Security 4.4.1.1. Requirements for Secure Inter-Server Copy Inter-server copy is driven by several requirements: o The specification MUST NOT mandate an inter-server copy protocol. There are many ways to copy data. Some will be more optimal than others depending on the identities of the source server and destination server. For example the source and destination servers might be two nodes sharing a common file system format for the source and destination file systems. Thus the source and destination are in an ideal position to efficiently render the image of the source file to the destination file by replicating the file system formats at the block level. In other cases, the source and destination might be two nodes sharing a common storage area network, and thus there is no need to copy any data at all, and instead ownership of the file and its contents simply gets re- assigned to the destination. o The specification MUST provide guidance for using NFSv4.x as a copy protocol. For those source and destination servers willing Haynes Expires October 20, 2011 [Page 39] Internet-Draft NFSv4.2 April 2011 to use NFSv4.x there are specific security considerations that this specification can and does address. o The specification MUST NOT mandate pre-configuration between the source and destination server. Requiring that the source and destination first have a "copying relationship" increases the administrative burden. However the specification MUST NOT preclude implementations that require pre-configuration. o The specification MUST NOT mandate a trust relationship between the source and destination server. The NFSv4 security model requires mutual authentication between a principal on an NFS client and a principal on an NFS server. This model MUST continue with the introduction of COPY. 4.4.1.2. Inter-Server Copy with RPCSEC_GSSv3 When the client sends a COPY_NOTIFY to the source server to expect the destination to attempt to copy data from the source server, it is expected that this copy is being done on behalf of the principal (called the "user principal") that sent the RPC request that encloses the COMPOUND procedure that contains the COPY_NOTIFY operation. The user principal is identified by the RPC credentials. A mechanism that allows the user principal to authorize the destination server to perform the copy in a manner that lets the source server properly authenticate the destination's copy, and without allowing the destination to exceed its authorization is necessary. An approach that sends delegated credentials of the client's user principal to the destination server is not used for the following reasons. If the client's user delegated its credentials, the destination would authenticate as the user principal. If the destination were using the NFSv4 protocol to perform the copy, then the source server would authenticate the destination server as the user principal, and the file copy would securely proceed. However, this approach would allow the destination server to copy other files. The user principal would have to trust the destination server to not do so. This is counter to the requirements, and therefore is not considered. Instead an approach using RPCSEC_GSSv3 [6] privileges is proposed. One of the stated applications of the proposed RPCSEC_GSSv3 protocol is compound client host and user authentication [+ privilege assertion]. For inter-server file copy, we require compound NFS server host and user authentication [+ privilege assertion]. The distinction between the two is one without meaning. RPCSEC_GSSv3 introduces the notion of privileges. We define three Haynes Expires October 20, 2011 [Page 40] Internet-Draft NFSv4.2 April 2011 privileges: copy_from_auth: A user principal is authorizing a source principal ("nfs@") to allow a destination principal ("nfs@ ") to copy a file from the source to the destination. This privilege is established on the source server before the user principal sends a COPY_NOTIFY operation to the source server. struct copy_from_auth_priv { secret4 cfap_shared_secret; netloc4 cfap_destination; /* the NFSv4 user name that the user principal maps to */ utf8str_mixed cfap_username; /* equal to seq_num of rpc_gss_cred_vers_3_t */ unsigned int cfap_seq_num; }; cap_shared_secret is a secret value the user principal generates. copy_to_auth: A user principal is authorizing a destination principal ("nfs@") to allow it to copy a file from the source to the destination. This privilege is established on the destination server before the user principal sends a COPY operation to the destination server. struct copy_to_auth_priv { /* equal to cfap_shared_secret */ secret4 ctap_shared_secret; netloc4 ctap_source; /* the NFSv4 user name that the user principal maps to */ utf8str_mixed ctap_username; /* equal to seq_num of rpc_gss_cred_vers_3_t */ unsigned int ctap_seq_num; }; ctap_shared_secret is a secret value the user principal generated and was used to establish the copy_from_auth privilege with the source principal. copy_confirm_auth: A destination principal is confirming with the source principal that it is authorized to copy data from the source on behalf of the user principal. When the inter-server copy protocol is NFSv4, or for that matter, any protocol capable of being secured via RPCSEC_GSSv3 (i.e. any ONC RPC protocol), Haynes Expires October 20, 2011 [Page 41] Internet-Draft NFSv4.2 April 2011 this privilege is established before the file is copied from the source to the destination. struct copy_confirm_auth_priv { /* equal to GSS_GetMIC() of cfap_shared_secret */ opaque ccap_shared_secret_mic<>; /* the NFSv4 user name that the user principal maps to */ utf8str_mixed ccap_username; /* equal to seq_num of rpc_gss_cred_vers_3_t */ unsigned int ccap_seq_num; }; 4.4.1.2.1. Establishing a Security Context When the user principal wants to COPY a file between two servers, if it has not established copy_from_auth and copy_to_auth privileges on the servers, it establishes them: o The user principal generates a secret it will share with the two servers. This shared secret will be placed in the cfap_shared_secret and ctap_shared_secret fields of the appropriate privilege data types, copy_from_auth_priv and copy_to_auth_priv. o An instance of copy_from_auth_priv is filled in with the shared secret, the destination server, and the NFSv4 user id of the user principal. It will be sent with an RPCSEC_GSS3_CREATE procedure, and so cfap_seq_num is set to the seq_num of the credential of the RPCSEC_GSS3_CREATE procedure. Because cfap_shared_secret is a secret, after XDR encoding copy_from_auth_priv, GSS_Wrap() (with privacy) is invoked on copy_from_auth_priv. The RPCSEC_GSS3_CREATE procedure's arguments are: struct { rpc_gss3_gss_binding *compound_binding; rpc_gss3_chan_binding *chan_binding_mic; rpc_gss3_assertion assertions<>; rpc_gss3_extension extensions<>; } rpc_gss3_create_args; The string "copy_from_auth" is placed in assertions[0].privs. The output of GSS_Wrap() is placed in extensions[0].data. The field extensions[0].critical is set to TRUE. The source server calls GSS_Unwrap() on the privilege, and verifies that the seq_num matches the credential. It then verifies that the NFSv4 user id Haynes Expires October 20, 2011 [Page 42] Internet-Draft NFSv4.2 April 2011 being asserted matches the source server's mapping of the user principal. If it does, the privilege is established on the source server as: <"copy_from_auth", user id, destination>. The successful reply to RPCSEC_GSS3_CREATE has: struct { opaque handle<>; rpc_gss3_chan_binding *chan_binding_mic; rpc_gss3_assertion granted_assertions<>; rpc_gss3_assertion server_assertions<>; rpc_gss3_extension extensions<>; } rpc_gss3_create_res; The field "handle" is the RPCSEC_GSSv3 handle that the client will use on COPY_NOTIFY requests involving the source and destination server. granted_assertions[0].privs will be equal to "copy_from_auth". The server will return a GSS_Wrap() of copy_to_auth_priv. o An instance of copy_to_auth_priv is filled in with the shared secret, the source server, and the NFSv4 user id. It will be sent with an RPCSEC_GSS3_CREATE procedure, and so ctap_seq_num is set to the seq_num of the credential of the RPCSEC_GSS3_CREATE procedure. Because ctap_shared_secret is a secret, after XDR encoding copy_to_auth_priv, GSS_Wrap() is invoked on copy_to_auth_priv. The RPCSEC_GSS3_CREATE procedure's arguments are: struct { rpc_gss3_gss_binding *compound_binding; rpc_gss3_chan_binding *chan_binding_mic; rpc_gss3_assertion assertions<>; rpc_gss3_extension extensions<>; } rpc_gss3_create_args; The string "copy_to_auth" is placed in assertions[0].privs. The output of GSS_Wrap() is placed in extensions[0].data. The field extensions[0].critical is set to TRUE. After unwrapping, verifying the seq_num, and the user principal to NFSv4 user ID mapping, the destination establishes a privilege of <"copy_to_auth", user id, source>. The successful reply to RPCSEC_GSS3_CREATE has: Haynes Expires October 20, 2011 [Page 43] Internet-Draft NFSv4.2 April 2011 struct { opaque handle<>; rpc_gss3_chan_binding *chan_binding_mic; rpc_gss3_assertion granted_assertions<>; rpc_gss3_assertion server_assertions<>; rpc_gss3_extension extensions<>; } rpc_gss3_create_res; The field "handle" is the RPCSEC_GSSv3 handle that the client will use on COPY requests involving the source and destination server. The field granted_assertions[0].privs will be equal to "copy_to_auth". The server will return a GSS_Wrap() of copy_to_auth_priv. 4.4.1.2.2. Starting a Secure Inter-Server Copy When the client sends a COPY_NOTIFY request to the source server, it uses the privileged "copy_from_auth" RPCSEC_GSSv3 handle. cna_destination_server in COPY_NOTIFY MUST be the same as the name of the destination server specified in copy_from_auth_priv. Otherwise, COPY_NOTIFY will fail with NFS4ERR_ACCESS. The source server verifies that the privilege <"copy_from_auth", user id, destination> exists, and annotates it with the source filehandle, if the user principal has read access to the source file, and if administrative policies give the user principal and the NFS client read access to the source file (i.e. if the ACCESS operation would grant read access). Otherwise, COPY_NOTIFY will fail with NFS4ERR_ACCESS. When the client sends a COPY request to the destination server, it uses the privileged "copy_to_auth" RPCSEC_GSSv3 handle. ca_source_server in COPY MUST be the same as the name of the source server specified in copy_to_auth_priv. Otherwise, COPY will fail with NFS4ERR_ACCESS. The destination server verifies that the privilege <"copy_to_auth", user id, source> exists, and annotates it with the source and destination filehandles. If the client has failed to establish the "copy_to_auth" policy it will reject the request with NFS4ERR_PARTNER_NO_AUTH. If the client sends a COPY_REVOKE to the source server to rescind the destination server's copy privilege, it uses the privileged "copy_from_auth" RPCSEC_GSSv3 handle and the cra_destination_server in COPY_REVOKE MUST be the same as the name of the destination server specified in copy_from_auth_priv. The source server will then delete the <"copy_from_auth", user id, destination> privilege and fail any subsequent copy requests sent under the auspices of this privilege from the destination server. Haynes Expires October 20, 2011 [Page 44] Internet-Draft NFSv4.2 April 2011 4.4.1.2.3. Securing ONC RPC Server-to-Server Copy Protocols After a destination server has a "copy_to_auth" privilege established on it, and it receives a COPY request, if it knows it will use an ONC RPC protocol to copy data, it will establish a "copy_confirm_auth" privilege on the source server, using nfs@ as the initiator principal, and nfs@ as the target principal. The value of the field ccap_shared_secret_mic is a GSS_VerifyMIC() of the shared secret passed in the copy_to_auth privilege. The field ccap_username is the mapping of the user principal to an NFSv4 user name ("user"@"domain" form), and MUST be the same as ctap_username and cfap_username. The field ccap_seq_num is the seq_num of the RPCSEC_GSSv3 credential used for the RPCSEC_GSS3_CREATE procedure the destination will send to the source server to establish the privilege. The source server verifies the privilege, and establishes a <"copy_confirm_auth", user id, destination> privilege. If the source server fails to verify the privilege, the COPY operation will be rejected with NFS4ERR_PARTNER_NO_AUTH. All subsequent ONC RPC requests sent from the destination to copy data from the source to the destination will use the RPCSEC_GSSv3 handle returned by the source's RPCSEC_GSS3_CREATE response. Note that the use of the "copy_confirm_auth" privilege accomplishes the following: o if a protocol like NFS is being used, with export policies, export policies can be overridden in case the destination server as-an- NFS-client is not authorized o manual configuration to allow a copy relationship between the source and destination is not needed. If the attempt to establish a "copy_confirm_auth" privilege fails, then when the user principal sends a COPY request to destination, the destination server will reject it with NFS4ERR_PARTNER_NO_AUTH. 4.4.1.2.4. Securing Non ONC RPC Server-to-Server Copy Protocols If the destination won't be using ONC RPC to copy the data, then the source and destination are using an unspecified copy protocol. The destination could use the shared secret and the NFSv4 user id to prove to the source server that the user principal has authorized the copy. For protocols that authenticate user names with passwords (e.g. HTTP Haynes Expires October 20, 2011 [Page 45] Internet-Draft NFSv4.2 April 2011 [14] and FTP [15]), the nfsv4 user id could be used as the user name, and an ASCII hexadecimal representation of the RPCSEC_GSSv3 shared secret could be used as the user password or as input into non- password authentication methods like CHAP [16]. 4.4.1.3. Inter-Server Copy via ONC RPC but without RPCSEC_GSSv3 ONC RPC security flavors other than RPCSEC_GSSv3 MAY be used with the server-side copy offload operations described in this document. In particular, host-based ONC RPC security flavors such as AUTH_NONE and AUTH_SYS MAY be used. If a host-based security flavor is used, a minimal level of protection for the server-to-server copy protocol is possible. In the absence of strong security mechanisms such as RPCSEC_GSSv3, the challenge is how the source server and destination server identify themselves to each other, especially in the presence of multi-homed source and destination servers. In a multi-homed environment, the destination server might not contact the source server from the same network address specified by the client in the COPY_NOTIFY. This can be overcome using the procedure described below. When the client sends the source server the COPY_NOTIFY operation, the source server may reply to the client with a list of target addresses, names, and/or URLs and assign them to the unique triple: . If the destination uses one of these target netlocs to contact the source server, the source server will be able to uniquely identify the destination server, even if the destination server does not connect from the address specified by the client in COPY_NOTIFY. For example, suppose the network topology is as shown in Figure 4. If the source filehandle is 0x12345, the source server may respond to a COPY_NOTIFY for destination 10.11.78.56 with the URLs: nfs://10.11.78.18//_COPY/10.11.78.56/_FH/0x12345 nfs://192.168.33.18//_COPY/10.11.78.56/_FH/0x12345 The client will then send these URLs to the destination server in the COPY operation. Suppose that the 192.168.33.0/24 network is a high speed network and the destination server decides to transfer the file over this network. If the destination contacts the source server from 192.168.33.56 over this network using NFSv4.1, it does the following: Haynes Expires October 20, 2011 [Page 46] Internet-Draft NFSv4.2 April 2011 COMPOUND { PUTROOTFH, LOOKUP "_COPY" ; LOOKUP "10.11.78.56"; LOOKUP "_FH" ; OPEN "0x12345" ; GETFH } The source server will therefore know that these NFSv4.1 operations are being issued by the destination server identified in the COPY_NOTIFY. 4.4.1.4. Inter-Server Copy without ONC RPC and RPCSEC_GSSv3 The same techniques as Section 4.4.1.3, using unique URLs for each destination server, can be used for other protocols (e.g. HTTP [14] and FTP [15]) as well. 4.5. IANA Considerations This section has no actions for IANA. 5. Space Reservation 5.1. Introduction This section describes a set of operations that allow applications such as hypervisors to reserve space for a file, report the amount of actual disk space a file occupies and freeup the backing space of a file when it is not required. In virtualized environments, virtual disk files are often stored on NFS mounted volumes. Since virtual disk files represent the hard disks of virtual machines, hypervisors often have to guarantee certain properties for the file. One such example is space reservation. When a hypervisor creates a virtual disk file, it often tries to preallocate the space for the file so that there are no future allocation related errors during the operation of the virtual machine. Such errors prevent a virtual machine from continuing execution and result in downtime. Another useful feature would be the ability to report the number of blocks that would be freed when a file is deleted. Currently, NFS reports two size attributes: size The logical file size of the file. space_used The size in bytes that the file occupies on disk While these attributes are sufficient for space accounting in traditional filesystems, they prove to be inadequate in modern Haynes Expires October 20, 2011 [Page 47] Internet-Draft NFSv4.2 April 2011 filesystems that support block sharing. Having a way to tell the number of blocks that would be freed if the file was deleted would be useful to applications that wish to migrate files when a volume is low on space. Since virtual disks represent a hard drive in a virtual machine, a virtual disk can be viewed as a filesystem within a file. Since not all blocks within a filesystem are in use, there is an opportunity to reclaim blocks that are no longer in use. A call to deallocate blocks could result in better space efficiency. Lesser space MAY be consumed for backups after block deallocation. We propose the following operations and attributes for the aforementioned use cases: space_reserve This attribute specifies whether the blocks backing the file have been preallocated. space_freed This attribute specifies the space freed when a file is deleted, taking block sharing into consideration. max_hole_punch This attribute specifies the maximum sized hole that can be punched on the filesystem. HOLE_PUNCH This operation zeroes and/or deallocates the blocks backing a region of the file. 5.2. Use Cases 5.2.1. Space Reservation Some applications require that once a file of a certain size is created, writes to that file never fail with an out of space condition. One such example is that of a hypervisor writing to a virtual disk. An out of space condition while writing to virtual disks would mean that the virtual machine would need to be frozen. Currently, in order to achieve such a guarantee, applications zero the entire file. The initial zeroing allocates the backing blocks and all subsequent writes are overwrites of already allocated blocks. This approach is not only inefficient in terms of the amount of I/O done, it is also not guaranteed to work on filesystems that are log structured or deduplicated. An efficient way of guaranteeing space reservation would be beneficial to such applications. If the space_reserved attribute is set on a file, it is guaranteed that writes that do not grow the file will not fail with NFSERR_NOSPC. Haynes Expires October 20, 2011 [Page 48] Internet-Draft NFSv4.2 April 2011 5.2.2. Space freed on deletes Currently, files in NFS have two size attributes: size The logical file size of the file. space_used The size in bytes that the file occupies on disk. While these attributes are sufficient for space accounting in traditional filesystems, they prove to be inadequate in modern filesystems that support block sharing. In such filesystems, multiple inodes can point to a single block with a block reference count to guard against premature freeing. If space_used of a file is interpreted to mean the size in bytes of all disk blocks pointed to by the inode of the file, then shared blocks get double counted, over-reporting the space utilization. This also has the adverse effect that the deletion of a file with shared blocks frees up less than space_used bytes. On the other hand, if space_used is interpreted to mean the size in bytes of those disk blocks unique to the inode of the file, then shared blocks are not counted in any file, resulting in under- reporting of the space utilization. For example, two files A and B have 10 blocks each. Let 6 of these blocks be shared between them. Thus, the combined space utilized by the two files is 14 * BLOCK_SIZE bytes. In the former case, the combined space utilization of the two files would be reported as 20 * BLOCK_SIZE. However, deleting either would only result in 4 * BLOCK_SIZE being freed. Conversely, the latter interpretation would report that the space utilization is only 8 * BLOCK_SIZE. Adding another size attribute, space_freed, is helpful in solving this problem. space_freed is the number of blocks that are allocated to the given file that would be freed on its deletion. In the example, both A and B would report space_freed as 4 * BLOCK_SIZE and space_used as 10 * BLOCK_SIZE. If A is deleted, B will report space_freed as 10 * BLOCK_SIZE as the deletion of B would result in the deallocation of all 10 blocks. The addition of this problem doesn't solve the problem of space being over-reported. However, over-reporting is better than under- reporting. Haynes Expires October 20, 2011 [Page 49] Internet-Draft NFSv4.2 April 2011 5.2.3. Operations and attributes In the sections that follow, one operation and three attributes are defined that together provide the space management facilities outlined earlier in the document. The operation is intended to be OPTIONAL and the attributes RECOMMENDED as defined in section 17 of [2]. 5.2.4. Attribute 77: space_reserve The space_reserve attribute is a read/write attribute of type boolean. It is a per file attribute. When the space_reserved attribute is set via SETATTR, the server must ensure that there is disk space to accommodate every byte in the file before it can return success. If the server cannot guarantee this, it must return NFS4ERR_NOSPC. If the client tries to grow a file which has the space_reserved attribute set, the server must guarantee that there is disk space to accommodate every byte in the file with the new size before it can return success. If the server cannot guarantee this, it must return NFS4ERR_NOSPC. It is not required that the server allocate the space to the file before returning success. The allocation can be deferred, however, it must be guaranteed that it will not fail for lack of space. The value of space_reserved can be obtained at any time through GETATTR. In order to avoid ambiguity, the space_reserve bit cannot be set along with the size bit in SETATTR. Increasing the size of a file with space_reserve set will fail if space reservation cannot be guaranteed for the new size. If the file size is decreased, space reservation is only guaranteed for the new size and the extra blocks backing the file can be released. 5.2.5. Attribute 78: space_freed space_freed gives the number of bytes freed if the file is deleted. This attribute is read only and is of type length4. It is a per file attribute. 5.2.6. Attribute 79: max_hole_punch max_hole_punch specifies the maximum size of a hole that the HOLE_PUNCH operation can handle. This attribute is read only and of type length4. It is a per filesystem attribute. This attribute MUST Haynes Expires October 20, 2011 [Page 50] Internet-Draft NFSv4.2 April 2011 be implemented if HOLE_PUNCH is implemented. 5.2.7. Operation 64: HOLE_PUNCH - Zero and deallocate blocks backing the file in the specified range. 5.2.7.1. ARGUMENT struct HOLE_PUNCH4args { /* CURRENT_FH: file */ offset4 hpa_offset; length4 hpa_count; }; 5.2.7.2. RESULT struct HOLEPUNCH4res { nfsstat4 hpr_status; }; 5.2.7.3. DESCRIPTION Whenever a client wishes to deallocate the blocks backing a particular region in the file, it calls the HOLE_PUNCH operation with the current filehandle set to the filehandle of the file in question, start offset and length in bytes of the region set in hpa_offset and hpa_count respectively. All further reads to this region MUST return zeros until overwritten. The filehandle specified must be that of a regular file. Situations may arise where hpa_offset and/or hpa_offset + hpa_count will not be aligned to a boundary that the server does allocations/ deallocations in. For most filesystems, this is the block size of the file system. In such a case, the server can deallocate as many bytes as it can in the region. The blocks that cannot be deallocated MUST be zeroed. Except for the block deallocation and maximum hole punching capability, a HOLE_PUNCH operation is to be treated similar to a write of zeroes. The server is not required to complete deallocating the blocks specified in the operation before returning. It is acceptable to have the deallocation be deferred. In fact, HOLE_PUNCH is merely a hint; it is valid for a server to return success without ever doing anything towards deallocating the blocks backing the region specified. However, any future reads to the region MUST return zeroes. HOLE_PUNCH will result in the space_used attribute being decreased by the number of bytes that were deallocated. The space_freed attribute Haynes Expires October 20, 2011 [Page 51] Internet-Draft NFSv4.2 April 2011 may or may not decrease, depending on the support and whether the blocks backing the specified range were shared or not. The size attribute will remain unchanged. The HOLE_PUNCH operation MUST NOT change the space reservation guarantee of the file. While the server can deallocate the blocks specified by hpa_offset and hpa_count, future writes to this region MUST NOT fail with NFSERR_NOSPC. The HOLE_PUNCH operation may fail for the following reasons (this is a partial list): NFS4ERR_NOTSUPP The Hole punch operations are not supported by the NFS server receiving this request. NFS4ERR_DIR The current filehandle is of type NF4DIR. NFS4ERR_SYMLINK The current filehandle is of type NF4LNK. NFS4ERR_WRONG_TYPE The current filehandle does not designate an ordinary file. 5.3. Security Considerations There are no security considerations for this section. 5.4. IANA Considerations This section has no actions for IANA. 6. Simple and Efficient Read Support for Sparse Files 6.1. Introduction NFS is now used in many data centers as the sole or primary method of data access. Consequently, more types of applications are using NFS than ever before, each with their own requirements and generated workloads. As part of this, sparse files are increasing in number while NFS continues to lack any specific knowledge of a sparse file's layout. This document puts forth a proposal for the NFSv4.2 protocol to support efficient reading of sparse files. A sparse file is a common way of representing a large file without having to reserve disk space for it. Consequently, a sparse file uses less physical space than its size indicates. This means the file contains 'holes', byte ranges within the file that contain no data. Most modern file systems support sparse files, including most Haynes Expires October 20, 2011 [Page 52] Internet-Draft NFSv4.2 April 2011 UNIX file systems and NTFS, but notably not Apple's HFS+. Common examples of sparse files include VM OS/disk images, database files, log files, and even checkpoint recovery files most commonly used by the HPC community. If an application reads a hole in a sparse file, the file system must returns all zeros to the application. For local data access there is little penalty, but with NFS these zeroes must be transferred back to the client. If an application uses the NFS client to read data into memory, this wastes time and bandwidth as the application waits for the zeroes to be transferred. Once the zeroes arrive, they then steal memory or cache space from real data. To make matters worse, if an application then proceeds to write data to another file system, the zeros are written into the file, expanding the sparse file into a full sized regular file. Beyond wasting disk space, this can actually prevent large sparse files from ever being copied to another storage location due to space limitations. This document adds a new READPLUS operation to efficiently read from sparse files by avoiding the transfer of all zero regions from the server to the client. READPLUS supports all the features of READ but includes a minimal extension to support sparse files. In addition, the return value of READPLUS is now compatible with NFSv4.1 minor versioning rules and could support other future extensions without requiring yet another operation. READPLUS is guaranteed to perform no worse than READ, and can dramatically improve performance with sparse files. READPLUS does not depend on pNFS protocol features, but can be used by pNFS to support sparse files. 6.2. Terminology Regular file Regular file: An object of file type NF4REG or NF4NAMEDATTR. Sparse file Sparse File. A Regular file that contains one or more Holes. Hole Hole. A byte range within a Sparse file that contains regions of all zeroes. For block-based file systems, this could also be an unallocated region of the file. 6.3. Applications and Sparse Files Applications may cause an NFS client to read holes in a file for several reasons. This section describes three different application workloads that cause the NFS client to transfer data unnecessarily. These workloads are simply examples, and there are probably many more workloads that are negatively impacted by sparse files. Haynes Expires October 20, 2011 [Page 53] Internet-Draft NFSv4.2 April 2011 The first workload that can cause holes to be read is sequential reads within a sparse file. When this happens, the NFS client may perform read requests ("readahead") into sections of the file not explicitly requested by the application. Since the NFS client cannot differentiate between holes and non-holes, the NFS client may prefetch empty sections of the file. This workload is exemplified by Virtual Machines and their associated file system images, e.g., VMware .vmdk files, which are large sparse files encapsulating an entire operating system. If a VM reads files within the file system image, this will translate to sequential NFS read requests into the much larger file system image file. Since NFS does not understand the internals of the file system image, it ends up performing readahead file holes. The second workload is generated by copying a file from a directory in NFS to either the same NFS server, to another file system, e.g., another NFS or Samba server, to a local ext3 file system, or even a network socket. In this case, bandwidth and server resources are wasted as the entire file is transferred from the NFS server to the NFS client. Once a byte range of the file has been transferred to the client, it is up to the client application, e.g., rsync, cp, scp, on how it writes the data to the target location. For example, cp supports sparse files and will not write all zero regions, whereas scp does not support sparse files and will transfer every byte of the file. The third workload is generated by applications that do not utilize the NFS client cache, but instead use direct I/O and manage cached data independently, e.g., databases. These applications may perform whole file caching with sparse files, which would mean that even the holes will be transferred to the clients and cached. 6.4. Overview of Sparse Files and NFSv4 This proposal seeks to provide sparse file support to the largest number of NFS client and server implementations, and as such proposes to add a new return code to the mandatory NFSv4.1 READPLUS operation instead of proposing additions or extensions of new or existing optional features (such as pNFS). As well, this document seeks to ensure that the proposed extensions are simple and do not transfer data between the client and server unnecessarily. For example, one possible way to implement sparse file read support would be to have the client, on the first hole encountered or at OPEN time, request a Data Region Map from the server. A Data Region Map would specify all zero and non-zero regions in a file. While this option seems simple, it is less useful Haynes Expires October 20, 2011 [Page 54] Internet-Draft NFSv4.2 April 2011 and can become inefficient and cumbersome for several reasons: o Data Region Maps can be large, and transferring them can reduce overall read performance. For example, VMware's .vmdk files can have a file size of over 100 GBs and have a map well over several MBs. o Data Region Maps can change frequently, and become invalidated on every write to the file. This can result the map being transferred multiple times with each update to the file. For example, a VM that updates a config file in its file system image would invalidate the Data Region Map not only for itself, but for all other clients accessing the same file system image. o Data Region Maps do not handle all zero-filled sections of the file, reducing the effectiveness of the solution. While it may be possible to modify the maps to handle zero-filled sections (at possibly great effort to the server), it is almost impossible with pNFS. With pNFS, the owner of the Data Region Map is the metadata server, which is not in the data path and has no knowledge of the contents of a data region. Another way to handle holes is compression, but this not ideal since it requires all implementations to agree on a single compression algorithm and requires a fair amount of computational overhead. Note that supporting writing to a sparse file does not require changes to the protocol. Applications and/or NFS implementations can choose to ignore WRITE requests of all zeroes to the NFS server without consequence. 6.5. Operation 65: READPLUS The section introduces a new read operation, named READPLUS, which allows NFS clients to avoid reading holes in a sparse file. READPLUS is guaranteed to perform no worse than READ, and can dramatically improve performance with sparse files. READPLUS supports all the features of the existing NFSv4.1 READ operation [2] and adds a simple yet significant extension to the format of its response. The change allows the client to avoid returning all zeroes from a file hole, wasting computational and network resources and reducing performance. READPLUS uses a new result structure that tells the client that the result is all zeroes AND the byte-range of the hole in which the request was made. Returning the hole's byte-range, and only upon request, avoids transferring large Data Region Maps that may be soon invalidated and contain information about a file that may not even be read in its Haynes Expires October 20, 2011 [Page 55] Internet-Draft NFSv4.2 April 2011 entirely. A new read operation is required due to NFSv4.1 minor versioning rules that do not allow modification of existing operation's arguments or results. READPLUS is designed in such a way to allow future extensions to the result structure. The same approach could be taken to extend the argument structure, but a good use case is first required to make such a change. 6.5.1. ARGUMENT struct COPY_NOTIFY4args { /* CURRENT_FH: source file */ netloc4 cna_destination_server; }; 6.5.2. RESULT union COPY_NOTIFY4res switch (nfsstat4 cnr_status) { case NFS4_OK: nfstime4 cnr_lease_time; netloc4 cnr_source_server<>; default: void; }; 6.5.3. DESCRIPTION The READPLUS operation is based upon the NFSv4.1 READ operation [2], and similarly reads data from the regular file identified by the current filehandle. The client provides an offset of where the READPLUS is to start and a count of how many bytes are to be read. An offset of zero means to read data starting at the beginning of the file. If offset is greater than or equal to the size of the file, the status NFS4_OK is returned with nfs_readplusrestype4 set to READ_OK, data length set to zero, and eof set to TRUE. The READPLUS is subject to access permissions checking. If the client specifies a count value of zero, the READPLUS succeeds and returns zero bytes of data, again subject to access permissions checking. In all situations, the server may choose to return fewer bytes than specified by the client. The client needs to check for this condition and handle the condition appropriately. Haynes Expires October 20, 2011 [Page 56] Internet-Draft NFSv4.2 April 2011 If the client specifies an offset and count value that is entirely contained within a hole of the file, the status NFS4_OK is returned with nfs_readplusresok4 set to READ_HOLE, and if information is available regarding the hole, a nfs_readplusreshole structure containing the offset and range of the entire hole. The nfs_readplusreshole structure is considered valid until the file is changed (detected via the change attribute). The server MUST provide the same semantics for nfs_readplusreshole as if the client read the region and received zeroes; the implied holes contents lifetime MUST be exactly the same as any other read data. If the client specifies an offset and count value that begins in a non-hole of the file but extends into hole the server should return a short read with status NFS4_OK, nfs_readplusresok4 set to READ_OK, and data length set to the number of bytes returned. The client will then issue another READPLUS for the remaining bytes, which the server will respond with information about the hole in the file. If the server knows that the requested byte range is into a hole of the file, but has no further information regarding the hole, it returns a nfs_readplusreshole structure with holeres4 set to HOLE_NOINFO. If hole information is available on the server and can be returned to the client, the server returns a nfs_readplusreshole structure with the value of holeres4 to HOLE_INFO. The values of hole_offset and hole_length define the byte-range for the current hole in the file. These values represent the information known to the server and may describe a byte-range smaller than the true size of the hole. Except when special stateids are used, the stateid value for a READPLUS request represents a value returned from a previous byte- range lock or share reservation request or the stateid associated with a delegation. The stateid identifies the associated owners if any and is used by the server to verify that the associated locks are still valid (e.g., have not been revoked). If the read ended at the end-of-file (formally, in a correctly formed READPLUS operation, if offset + count is equal to the size of the file), or the READPLUS operation extends beyond the size of the file (if offset + count is greater than the size of the file), eof is returned as TRUE; otherwise, it is FALSE. A successful READPLUS of an empty file will always return eof as TRUE. If the current filehandle is not an ordinary file, an error will be returned to the client. In the case that the current filehandle represents an object of type NF4DIR, NFS4ERR_ISDIR is returned. If the current filehandle designates a symbolic link, NFS4ERR_SYMLINK is Haynes Expires October 20, 2011 [Page 57] Internet-Draft NFSv4.2 April 2011 returned. In all other cases, NFS4ERR_WRONG_TYPE is returned. For a READPLUS with a stateid value of all bits equal to zero, the server MAY allow the READPLUS to be serviced subject to mandatory byte-range locks or the current share deny modes for the file. For a READPLUS with a stateid value of all bits equal to one, the server MAY allow READPLUS operations to bypass locking checks at the server. On success, the current filehandle retains its value. 6.5.4. IMPLEMENTATION If the server returns a "short read" (i.e., fewer data than requested and eof is set to FALSE), the client should send another READPLUS to get the remaining data. A server may return less data than requested under several circumstances. The file may have been truncated by another client or perhaps on the server itself, changing the file size from what the requesting client believes to be the case. This would reduce the actual amount of data available to the client. It is possible that the server reduce the transfer size and so return a short read result. Server resource exhaustion may also occur in a short read. If mandatory byte-range locking is in effect for the file, and if the byte-range corresponding to the data to be read from the file is WRITE_LT locked by an owner not associated with the stateid, the server will return the NFS4ERR_LOCKED error. The client should try to get the appropriate READ_LT via the LOCK operation before re- attempting the READPLUS. When the READPLUS completes, the client should release the byte-range lock via LOCKU. If another client has an OPEN_DELEGATE_WRITE delegation for the file being read, the delegation must be recalled, and the operation cannot proceed until that delegation is returned or revoked. Except where this happens very quickly, one or more NFS4ERR_DELAY errors will be returned to requests made while the delegation remains outstanding. Normally, delegations will not be recalled as a result of a READPLUS operation since the recall will occur as a result of an earlier OPEN. However, since it is possible for a READPLUS to be done with a special stateid, the server needs to check for this case even though the client should have done an OPEN previously. 6.5.4.1. Additional pNFS Implementation Information With pNFS, the semantics of using READPLUS remains the same. Any data server MAY return a READ_HOLE result for a READPLUS request that it receives. Haynes Expires October 20, 2011 [Page 58] Internet-Draft NFSv4.2 April 2011 When a data server chooses to return a READ_HOLE result, it has a certain level of flexibility in how it fills out the nfs_readplusreshole structure. 1. For a data server that cannot determine any hole information, the data server SHOULD return HOLE_NOINFO. 2. For a data server that can only obtain hole information for the parts of the file stored on that data server, the data server SHOULD return HOLE_INFO and the byte range of the hole stored on that data server. 3. For a data server that can obtain hole information for the entire file without severe performance impact, it MAY return HOLE_INFO nd the byte range of the entire file hole. In general, a data server should do its best to return as much information about a hole as is feasible. In general, pNFS server implementers should try ensure that data servers do not overload the metadata server with requests for information. Therefore, if supplying global sparse information for a file to data servers can overwhelm a metadata server, then data servers should use option 1 or 2 above. When a pNFS client receives a READ_HOLE result and a non-empty nfs_readplusreshole structure, it MAY use this information in conjunction with a valid layout for the file to determine the next data server for the next region of data that is not in a hole. 6.5.5. READPLUS with Sparse Files Example To see how the return value READ_HOLE will work, the following table describes a sparse file. For each byte range, the file contains either non-zero data or a hole. +-------------+----------+ | Byte-Range | Contents | +-------------+----------+ | 0-31999 | Non-Zero | | 32K-255999 | Hole | | 256K-287999 | Non-Zero | | 288K-353999 | Hole | | 354K-417999 | Non-Zero | +-------------+----------+ Table 3 Under the given circumstances, if a client was to read the file from Haynes Expires October 20, 2011 [Page 59] Internet-Draft NFSv4.2 April 2011 beginning to end with a max read size of 64K, the following will be the result. This assumes the client has already opened the file and acquired a valid stateid and just needs to issue READPLUS requests. 1. READPLUS(s, 0, 64K) --> NFS_OK, readplusrestype4 = READ_OK, eof = false, data<>[32K]. Return a short read, as the last half of the equest was all zeroes. 2. READPLUS(s, 32K, 64K) --> NFS_OK, readplusrestype4 = READ_HOLE, nfs_readplusreshole(HOLE_INFO)(32K, 224K). The requested range was all zeros, and the current hole begins at offset 32K and is 224K in length. 3. READPLUS(s, 256K, 64K) --> NFS_OK, readplusrestype4 = READ_OK, eof = false, data<>[32K]. Return a short read, as the last half of the request was all zeroes. 4. READPLUS(s, 288K, 64K) --> NFS_OK, readplusrestype4 = READ_HOLE, nfs_readplusreshole(HOLE_INFO)(288K, 66K). 5. READPLUS(s, 354K, 64K) --> NFS_OK, readplusrestype4 = READ_OK, eof = true, data<>[64K]. 6.6. Related Work Solaris and ZFS support an extension to lseek(2) that allows applications to discover holes in a file. The values, SEEK_HOLE and SEEK_DATA, allow clients to seek to the next hole or beginning of data, respectively. XFS supports the XFS_IOC_GETBMAP extended attribute, which returns the Data Region Map for a file. Clients can then use this information to avoid reading holes in a file. NTFS and CIFS support the FSCTL_SET_SPARSE attribute, which allows applications to control whether empty regions of the file are preallocated and filled in with zeros or simply left unallocated. 6.7. Security Considerations The additions to the NFS protocol for supporting sparse file reads does not alter the security considerations of the NFSv4.1 protocol [2]. 6.8. IANA Considerations There are no IANA considerations in this section. Haynes Expires October 20, 2011 [Page 60] Internet-Draft NFSv4.2 April 2011 7. Security Considerations 8. IANA Considerations This section uses terms that are defined in [17]. 9. References 9.1. Normative References [1] Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", March 1997. [2] Shepler, S., Eisler, M., and D. Noveck, "Network File System (NFS) Version 4 Minor Version 1 Protocol", RFC 5661, January 2010. [3] Black, D., Glasgow, J., and S. Fridella, "Parallel NFS (pNFS) Block/Volume Layout", RFC 5663, January 2010. [4] Halevy, B., Welch, B., and J. Zelenka, "Object-Based Parallel NFS (pNFS) Operations", RFC 5664, January 2010. [5] Berners-Lee, T., Fielding, R., and L. Masinter, "Uniform Resource Identifier (URI): Generic Syntax", STD 66, RFC 3986, January 2005. [6] Williams, N., "Remote Procedure Call (RPC) Security Version 3", draft-williams-rpcsecgssv3 (work in progress), 2008. [7] Shepler, S., Eisler, M., and D. Noveck, "Network File System (NFS) Version 4 Minor Version 1 External Data Representation Standard (XDR) Description", RFC 5662, January 2010. [8] Haynes, T., "Network File System (NFS) Version 4 Minor Version 2 External Data Representation Standard (XDR) Description", April 2011. [9] Eisler, M., Chiu, A., and L. Ling, "RPCSEC_GSS Protocol Specification", RFC 2203, September 1997. 9.2. Informative References [10] Haynes, T. and D. Noveck, "Network File System (NFS) version 4 Protocol", draft-ietf-nfsv4-rfc3530bis-09 (Work In Progress), April 2011. Haynes Expires October 20, 2011 [Page 61] Internet-Draft NFSv4.2 April 2011 [11] Eisler, M., "XDR: External Data Representation Standard", RFC 4506, May 2006. [12] Lentini, J., Everhart, C., Ellard, D., Tewari, R., and M. Naik, "NSDB Protocol for Federated Filesystems", draft-ietf-nfsv4-federated-fs-protocol (Work In Progress), 2010. [13] Lentini, J., Everhart, C., Ellard, D., Tewari, R., and M. Naik, "Administration Protocol for Federated Filesystems", draft-ietf-nfsv4-federated-fs-admin (Work In Progress), 2010. [14] Fielding, R., Gettys, J., Mogul, J., Frystyk, H., Masinter, L., Leach, P., and T. Berners-Lee, "Hypertext Transfer Protocol -- HTTP/1.1", RFC 2616, June 1999. [15] Postel, J. and J. Reynolds, "File Transfer Protocol", STD 9, RFC 959, October 1985. [16] Simpson, W., "PPP Challenge Handshake Authentication Protocol (CHAP)", RFC 1994, August 1996. [17] Narten, T. and H. Alvestrand, "Guidelines for Writing an IANA Considerations Section in RFCs", BCP 26, RFC 5226, May 2008. [18] Nowicki, B., "NFS: Network File System Protocol specification", RFC 1094, March 1989. [19] Callaghan, B., Pawlowski, B., and P. Staubach, "NFS Version 3 Protocol Specification", RFC 1813, June 1995. [20] Srinivasan, R., "Binding Protocols for ONC RPC Version 2", RFC 1833, August 1995. [21] Eisler, M., "NFS Version 2 and Version 3 Security Issues and the NFS Protocol's Use of RPCSEC_GSS and Kerberos V5", RFC 2623, June 1999. [22] Callaghan, B., "NFS URL Scheme", RFC 2224, October 1997. [23] Shepler, S., "NFS Version 4 Design Considerations", RFC 2624, June 1999. [24] Reynolds, J., "Assigned Numbers: RFC 1700 is Replaced by an On- line Database", RFC 3232, January 2002. [25] Linn, J., "The Kerberos Version 5 GSS-API Mechanism", RFC 1964, June 1996. Haynes Expires October 20, 2011 [Page 62] Internet-Draft NFSv4.2 April 2011 [26] 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. Appendix A. Acknowledgments For the pNFS Access Permissions Check, the original draft was by Sorin Faibish, David Black, Mike Eisler, and Jason Glasgow. The work was influenced by discussions with Benny Halevy and Bruce Fields. A review was done by Tom Haynes. For the Sharing change attribute implementation details with NFSv4 clients, the original draft was by Trond Myklebust. For the NFS Server-side Copy, the original draft was by James Lentini, Mike Eisler, Deepak Kenchammana, Anshul Madan, and Rahul Iyer. Talpey co-authored an unpublished version of that document. It was also was reviewed by a number of individuals: Pranoop Erasani, Tom Haynes, Arthur Lent, Trond Myklebust, Dave Noveck, Theresa Lingutla-Raj, Manjunath Shankararao, Satyam Vaghani, and Nico Williams. For the NFS space reservation operations, the original draft was by Mike Eisler, James Lentini, Manjunath Shankararao, and Rahul Iyer. For the sparse file support, the original draft was by Dean Hildebrand and Marc Eshel. Valuable input and advice was received from Sorin Faibish, Bruce Fields, Benny Halevy, Trond Myklebust, and Richard Scheffenegger. Appendix B. RFC Editor Notes [RFC Editor: please remove this section prior to publishing this document as an RFC] [RFC Editor: prior to publishing this document as an RFC, please replace all occurrences of RFCTBD10 with RFCxxxx where xxxx is the RFC number of this document] Haynes Expires October 20, 2011 [Page 63] Internet-Draft NFSv4.2 April 2011 Author's Address Thomas Haynes NetApp 9110 E 66th St Tulsa, OK 74133 USA Phone: +1 918 307 1415 Email: thomas@netapp.com URI: http://www.tulsalabs.com Haynes Expires October 20, 2011 [Page 64]