Security for the NFSv4 Protocols

1.1. Document Motivation

1.2. Document Annotation

In order to make progress on difficult issues which will require that changes be made in the existing handling of security issues, including many whose resolution would potentially involve compatibility issues with existing implementations, the author has tried his best to resolve these issues, even though there is no assurance that the resolution adopted by consensus will match the author's current best efforts. To provide possible resolutions that might be the basis of discussion while not foreclosing other possibilities, proposed changes are organized into a series of consensus items, which are listed in Appendix B and in a corresponding Appendix in the acl document.¶

For such pending issues, the following annotations will be used:¶

• A paragraph headed "[Author Aside]:", provides the author's comments about possible changes and will probably not appear in an eventual RFC.¶

  This paragraph can specify that certain changes within the current section are to be implicitly considered as part of a specific consensus item.¶

  The paragraph can indicate that all unannotated material in the current section is to be considered either the previous treatment or the proposed replacement text for a specific consensus item.¶

• A paragraph headed "[Consensus Needed (Item #NNx)]:", provides the author's preferred treatment of the matter and will only appear in the eventual RFC if working group consensus on the matter is obtained, allowing the necessary changes to be made permanent, without being conditional on a future consensus.¶

  The item id, represented above by "NNx" consists of a number identifying the specific consensus item and letter which is unique to appearance of that consensus item in a particular section. In cases in which a pending item is cited with no part of the discussion appearing in the current section, an item id of the form "#NN" is used.¶

• A paragraph headed "[Previous Treatment]:", indicates text that is provided for context but which the author believes, need not appear in the eventual RFC, because it is expected to be superseded by a corresponding consensus item.¶

  The corresponding consensus item is often easily inferred, but can be specified explicitly, as it is for items associated with the consensus item itself.¶

Each of the annotations above can be modified by addition of the phrase, "Including List" to indicate that it applies to a following bulleted list as well as the current paragraph or the phase "Entire Bulleted Item" to indicate it applies to all paragraphs within a specific bulleted item.¶

1.3. Compatibility and Compliance Issues

Changes that need to be adopted in this document might need to eventually change the behavior of clients and servers that were written to conform to earlier protocol specifications. There are two important classes of such changes discussed in Sections 1.3.1 and 1.3.2 below.¶

As [RFC2119] was originally conceived, compliance and compatibility issues were tightly bound together so that a change to compliance specifications would inevitably give rise to compatibility issues. However, over time, behaviors have come to be denigrated by use of the terms "SHOULD NOT" and "MUST NOT" to warn implementors of a harm deriving from from insecure operation rather than peer incompatibility.¶

When making changes in compliance requirements/recommendations we need to deal with the possibility that, in changing the specification, we might cause a previously compliant implementation to become non-compliant. Some implementers take the view that the compliance status of their implementations is of less importance than other considerations such as compatibility with local file system semantics. Others feel it is important to maintain the compliance status of existing implementations. In any case, the working group has been reluctant, when making such necessary changes, to make previously compliant implementations non-compliant, and will try not to do in cases in which it is known or can reasonably be expected that such implementations exist.¶

Although there is no way to be sure about the non-existence of such implementations, the working group has made judgments about what implementations are likely to exist. For example, when internationalization for NFSv4.0 was changed in the transition to [RFC7530], many previously non-compliant server implementations became officially compliant and there was a potential for conflict with implementations compliant with RFC3530 [RFC3530], if any such implementations existed.¶

In that particular case, it was decided that no such RFC3530-compliant server implementations existed and there was no need to accommodate servers whose internationalization was written to conform to [RFC3530] since no such servers existed and there were no client implementations expecting such behavior. As a result, no actual implementations became non-compliant as a result of this necessary shift.¶

In the corresponding cases dealt with in this document, the situation is more difficult since it may be harder to determine the actual behavior of all existing implementations, since the authors might no longer be actively involved with implementation issues. However, it will be necessary to warn implementors of the negative consequences of certain behaviors without going so far as to declare these practices non-compliant. For the most part, this document will use the terms "SHOULD" and "SHOULD NOT" to draw attention to the problems with troublesome behaviors that previous specifications mentioned with no indication of the problems with them. In such contexts, it is made clear that the need to maintain existing patterns of interoperation is a valid reason to bypass the normative term. The intention is to continue to allow previously acceptable implementations to be considered compliant while not placing the troublesome behaviors on the same levels as other alternatives as would happen if we used the terms "MAY" or "OPTIONAL" or continued to use some of the unfortunate practices discussed in Section 1.3.2.¶

This approach allows implementations to accept input from peers written in accord with previous specification while not obligating them to do so. The intent is to allow a transition to newer, better behaviors over time as client and server policies evolve. However, the specifics of this anticipated transition will vary:¶

• In the case of the issues dealt with in Section 1.3.1, the focus is on making implementers aware of the security problems with practices previously considered acceptable.¶

  It is specified that these practices SHOULD NOT be used by clients or allowed by servers in order to draw attention to the security problems with them. At the same time, the discussion of valid reasons to bypass these recommendations allows them to continue to be used while the infrastructure is developed to make their replacements easy to use.¶

  Servers are encouraged to adopt policies foreclosing client use but not obligated to do so.¶

• In the case of the issues dealt with in Section 1.3.2, the focus is different. In these cases, while we anticipate making changes in compliance specifications, there is no need to address a specific set of troublesome practices. Instead, the problem to be addressed is the vast range of allowable server behaviors previously defined as allowed, although not necessarily explicitly. This server-centered approach has made compatibility a hit-or-miss matter, requiring serious consideration of the question of what specific instances of multiple server behaviors need to be allowed and how clients can find out the choices that servers have made and possibly affect them, also making appropriate provision, as with other optional features, about how to adapt to the server's behavioral choices or to decide that they do not meet its needs.¶

  It appears that this approach was motivated, at least in large part, by the desire to fully support much of Windows ACL semantics while accommodating Unix servers incapable of providing much of that functionality. As a result, the working group will need to provide a way for the server to explicitly opt out of providing Windows functionality that it cannot provide and that Linux clients are not prepared to use.¶

  In addition, the working group will have to restrict, or at least better organize, sever behavioral choices related to the handling of ACLs.¶

2.1. Keyword Definitions

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 specified in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.¶

2.2. Special Considerations

Because this document needs to revise previous treatments of its subject, it will need to cite previous treatments of issues that now need to be dealt with in a different way.  This will take the form of quotations from documents whose treatment of the subject is being obsoleted, most often as direct quotation but sometimes as indirect ones as well.¶

Paragraphs headed "[Previous Treatment] or otherwise annotated as having that status, as described in Section 1, can be considered quotations in this context.¶

Such treatments in quotations will involve use of these BCP14-defined terms in two noteworthy ways:¶

• The term may have been used inappropriately (i.e not in accord with [RFC2119]), as has been the case for the "RECOMMENDED" attributes, which are in fact OPTIONAL.¶

  In such cases, the surrounding text will make clear that the quoted text does not have a normative effect.¶

  Some specific issues relating to this case are described below in Section 8.1.¶

• The term may been used in accord with [RFC2119], although the resulting normative statement is now felt to be inappropriate.¶

  In such cases, the surrounding text will need to make clear that the text quoted is no longer to be considered normative, often by providing new text that conflicts with the quoted, previously normative, text.¶

  An important instance of this situation is the description of AUTH_SYS as an "OPTIONAL" means of authentication". For detailed discussion of this case, see Sections 14 and 18.1.5¶

3.1. Per-connection Security Features

There are a number of security-related facilities that can be provided on a per-connection basis, eliminating the need to provide such support on a per-request basis, based on the RPC auth-flavor used.¶

These will initially be provided, in most cases, by RPC-with-TLS but similar facilities might be provided by new versions of existing transports or new RPC transports.¶

• The transport or a layer above it might provide encryption of requests and replies, eliminating the need for privacy and integrity services to be negotiated later and applied on a per-request basis.¶

  While clients might choose to establish connections that provide such encryption, servers can establish policies allowing access to certain pieces of the namespace using such security facilities, or limiting access to those providing privacy, allowing the use of either per-connection encryption or privacy services provided by RPCSEC_GSS.¶

• The transport or a layer above it might provide mutual authentication of the client and server peers as part of the establishment of the connection This authentication is distinct from the the mutual authentication of the client user and server peer, implemented within the RPCSEC_GSS framework.¶

  This form of authentication is of particular importance when the server allows the use of the auth-flavors AUTH_SYS and AUTH_NONE, which have no provision for the authentication of the user requesting the operation.¶

  While clients might choose, on their own,to establish connections without such peer authentication, servers can establish policies a limiting access to certain pieces of the namespace without such peer authentication or only allowing it when using RPCSEC_GSS.¶

To enable server policies to be effectively communicated to clients, the security negotiation framework now allows connection characteristics to be specified using pseudo-flavors returned as part of the response to SECINFO and SECINFO_NONAME. See Section 16 for details.¶

3.2. Handling of Multiple Minor Versions

In some cases, there are differences between minor versions in that there are security-related features, not present in all minor versions.¶

To deal with this issue, this document will focus on a few major areas listed below which are common to all minor versions.¶

• File access authorization (discussed in Section 8) is the same in all minor versions together with the identification/ authentication infrastructure supporting it (discussed in Section 14) provided by RPC and applying to all of NFS.¶

  An exception is made regarding labelled NFS, an optional feature within NFSv4.2, described in [RFC7862]. This is discussed as a version-specific feature in this document in Section 11.¶

• Features to secure data in-flight, all provided by RPC, together with the negotiation infrastructure to support them are common to all NFSv4 minor versions, are discussed in Section 16.¶

  However, the use of SECINFO_NONAME, together with changes needed for connection-based encryption, paralleling those proposed here for SECINFO, is treated as a version-specific feature and, while mentioned here, will be fully documented in new NFSv4.1 specification documents.¶

• The protection of state data from unauthorized modification is discussed in Section 12) is the same in all minor versions together with the identification/ authentication infrastructure supporting it (discussed in Section 14 by security services such as those provided by RPC-with-TLS.¶

  It needs to be noted that state protection based on RPCSEC_GSS is treated as a version-specific feature and will continue to be described by [RFC8881] or its successors. Also, it needs to be noted that the use of state protection was not discussed in [RFC7530].¶

3.3. Handling of Minor-version-specific features

There are a number of areas in which security features differ among minor versions, as discussed below. In some cases, a new feature requires specific security support while in others one version will have a new feature related to enhancing the security infrastructure.¶

How such features are dealt with in this document depends on the specific feature.¶

• In addition to SECINFO, whose enhanced description appears in this document, NFSv4.1 added a new SECINFO_NONAME operation, useful for pNFS file as well as having some non-pNFS uses.¶

  While the enhanced description of SECINFO mentions SECINFO_NONAME, this is handled as one of a number of cases in which the description has to indicate that different actions need to be taken for different minor versions.¶

  The definitive description of SECINFO_NONAME, now appearing in [RFC8881] needs to be modified to match the description of SECINFO appearing in this document. It is expected that this will be done as part of the rfc5661bis process.¶

  The security implications of the security negotiation facilities as a whole will be addressed in the security considerations section of this document.¶

• The OPTIONAL pNFS feature added in NFSv4.1 has its own security needs which parallel closely those of non-pNFS access but are distinct, especially when the storage access protocol used are not RPC protocols. As a result, these needs and the means to satisfy them are not discussed in this document.¶

  The definitive description of pNFS security will remain in [RFC8881] and its successors (i.e. the eventual rfc5661bis document). However, because pNFS security relies heavily on the infrastructure discussed here, it is anticipated that the new treatment of pNFS security will deal with many matters by referencing the overall NFS security document.¶

  The security considerations section of rfc5661bis will deal with pNFS security issues.¶

• In addition to the state protection facilities described in this document, NFS has another set of such facilities that are only implemented in NFSv4.1.¶

  While this document will discuss the security implications of protection against state modification, it will not discuss the details of the NFSv4.1-specific features to accomplish it.¶

• The additional NFSv4.1 acl attributes, sacl and dacl, are discussed in this document, together with the ACL inheritance features they enable.¶

  As a result, the responsibility for the definitive description of these attributes will move to overall NFS security document, with the fact that they are not available in NFSv4.0 duly noted. While these attributes will continue to be mentioned in NFSv4.1 specification documents, the detailed description appearing in [RFC8881] will be removed in successor documents.¶

• Both NFSv4.0 and NFSv4.1 specifications discussed the coordination of the values the mode and ACL-related attributes. While the treatment in [RFC8881] is more detailed, the differences in the approaches are quite minor.¶

  [Consensus Item #25a]: This document will provide a unified treatment of these issues, which will note any differences of treatment that apply to NFSv4.0. Changes applying to NFSv4.2 will also be noted.¶

  As a result, this document will override the treatment within [RFC7530] and [RFC8881]. This material will be removed in the rfc5661bis document suite and replaced by a reference to the treatment in the NFSv4 security RFC.¶

• The protocol extension defined in [RFC8257], now part of NFSv4.2, is also related to the issue of co-ordination of acl and mode attributes and will be discussed in that context.¶

  Nevertheless, the description in [RFC8257] will remain definitive.¶

• The NFSv4.1 attribute set-mode-masked attribute is mentioned together with the other attributes implementing the POSIX authorization model.¶

  Because this attribute. while related to security, does not substantively modify the security properties of the protocol, the full description of this attribute, will continue to be the province of the NFSv4.1 specification proper.¶

• There is a brief description of the v4.2 Labelled NFS feature in Section 11. Part of that description discusses the limitations in the description of that feature within [RFC7862].¶

  Because of some limitations in the description, it is not possible to provide an appropriate security considerations section for that feature in this document.¶

  As a result, the responsibility for providing an appropriate Security Considerations section remains, unrealized for now, with the NFSv4.2 specification document and its possible successors.¶

3.4. Features Needing Extensive Clarification

For a number of authorization-related features, the existing descriptions are inadequate for various reasons:¶

• In the description of the use of the mode attribute in implementing the POSIX-based authorization model, critical pieces of the semantics are not mentioned, while, ironically, the corresponding semantics for ACL-based authorization are discussed.¶

  This includes the authorization of file deletion and of modification of the mode, owner and owner-group attributes. For ACL-based authorization, there is an attempt to provide the corresponding description.¶

• The description of authorization for ACLs is more complete but it needs further work, because the previous specifications make extensive efforts, in my view misguided, to allow an enormous range of server behaviors, making it hard for a client to know what the effect of many actions, and the corresponding security-related consequences, might be.¶

  Troublesome in this connection were the discussion of ACE mask bits which essentially treats every mask bit, as its own OPTIONAL feature, the use of "SHOULD" and "SHOULD NOT" in situations which it is unclear what valid reasons to ignore the recommendation might be, and cases in which it is is simply stated that some servers do some particular thing, leaving the unfortunate implication that clients need to be prepared for a vast range of server behaviors.¶

  This approach essentially treated ACLs in a manner appropriate to an experimental feature even though in appeaed in a Proposed Standard.¶

• Similar issues apply to descriptions related to the need to co-ordinate the values of the mode attribute and the ACL-related attributes.¶

  Although the need for such coordination is recognized. There are multiple modes of mapping an ACL to a corresponding mode together with multiple sources of uncertainty about the reverse mapping.¶

  In addition, certain of the mapping algorithms have flaws in that their behavior under unusual circumstances providing results that appear erroneous.¶

Dealing with these issues is not straightforward, because the appropriate resolution will depend on:¶

• The actual existence of server implementations with non-preferred semantics.¶

  In some cases in which "SHOULD" was used, there may not have been any actual severs choosing to ignore the recommendation, eliminating the possibility of compatibility issues when changing the "SHOULD" to a formulation that restricts the server's choices.¶

• The difficulty of modifying server implementations to eliminate or narrow the effect of non-standard semantics.¶

  One aspect of that difficulty might be client or application expectations based on existing server implementations, even if the existing specifications give the client no assurance that that server's behavior is mandated by the standard.¶

• Whether the existing flaw in some existing recommended actions to be performed by the server is sufficiently troublesome to justify changing the specification at this point.¶

This sort of information will be used in deciding whether to:¶

• Narrow the scope of allowable server behavior to those actually used by existing severs.¶

• Limiting the negative effects of unmotivated SHOULDs by limiting valid reasons to ignore the recommendation to the difficulty of changing existing implementations.¶

  This would give significant guidance to future implementations, while forcing clients to live with the uncertainty about existing servers¶

• Tie a more restricted set of semantics to nominally unrelated OPTIONAL features such as implementation of dacl and sacl.¶

  This would provide a way to allow the development of newer servers to proceed on a firmer basis, without requiring changes on older servers that do not support these SMB-oriented attributes.¶

• Provide means that clients could use to determine, experimentally, what semantics are provided by the server.¶

  Would need to be supported by a requirement/assurance that a server behave uniformly, at least within the scope of a single file system.¶

• Allow the provision of other ways for the client to know the semantics choices made by the server or the file system.¶

Despite the difficulty of addressing these issues, if the protocol is to be secure and ACLs are to be widely available, these problems have to be addressed. While there has not been significant effort to provide client-side ACL APIs and there might not be for a while, we cannot have a situation in which the security specification makes that development essentially impossible.¶

3.5. Process Going Forward

Because of the scope of this document, and the fact that it is necessary to modify previous treatments of the subject previously published as Proposed Standards, it is necessary that the process of determining whether there is Working Group Consensus to submit it for publication be more structured than that used for the antecedent documents.¶

In order to facilitate this process, the necessary changes which need to be made, beyond those clearly editorial in nature, are listed in Appendix B. As working group review and discussion of this document and its successors proceeds, there will be occasion to discuss each of these changes, identified by the annotations described in Section 1.2.¶

Based on working group discussions, successive document versions will do one of the following for some set of consensus items:¶

• Deciding that the replacement text is now part of a new working group consensus.¶

  When this happens, future drafts of the document will be modified to remove the previous treatment, treat the proposed text as adopted, and remove Author Asides or replace them by new text explaining why a new treatment of the matter has been adopted or pointing the reader to an explanation in Appendix A.¶

  At this point, the consensus item will be removed from Appendix B and an explanation for the change will be added to Appendix A.¶

• Deciding that the general approach to the issue, if not necessarily the specific current text has reached the point of "general acceptance" as defined in Appendix B¶

  In this case, to facilitate discussion of remaining issues, the text of the document proper will remain as it is.¶

  At this point, the consensus item will be marked within the table in Appendix B as having reached general acceptance, indicating the need to prioritize discussion in the next document cycle, aimed at arriving at final text to address the issue.¶

  In addition, an explanation for the change will be added to Appendix A.¶

• Deciding that modification of the existing text is necessary to facilitate eventual consensus, based on the working group's input.¶

  In this case, there will be changes to the document proper in the next draft revision. In some cases, because of the need for a coherent description, text outside the consensus item may be affected.¶

  The table in Appendix B will be updated to reflect the new item status while Appendix A is not expected to change.¶

• Deciding that the item is best dropped in the next draft.¶

  In this case, the changes to the document proper will be the inverse of those when a change is accepted by consensus. The previous treatment will be restored as the current text while the proposed new text will vanish from the document at the next draft revision. The Author Aside will be the basis for an explanation of the consequences of not dealing with the issue.¶

  At this point, the consensus item will be removed from Appendix B.¶

The changes that the working group will need to reach consensus on, either to accept (as-is or with significant modifications) or reject can be divided into three groups.¶

• A large set of changes, all addressing issues mentioned in Section 1.1, were already present in the initial I-D so that there has been the opportunity for working group discussion of them, although that discussion has been quite limited so far.¶

  As a result, a small set of these changes is marked, in Appendix B, as having reached general acceptance.¶

  That subset of these changes changes, together with the organizational changes to support them are described in Appendix A.1.¶

• Another large set of changes were made in draft -02. These mostly concern the issues mentioned in Section 3.4 None of these changes is yet considered to have reached general acceptance.¶

  The issues that need to be addressed are described in Appendix A.2 while the possible approaches that might be taken to resolve these issues are described in Appendix A.3.¶

• There remain a set of potential changes for which a need is expected but for which no text is yet available.¶

  Such changes have associated Author Asides and are listed in Appendix B.¶

  The text for these changes is expected to be made available in future document revisions and they will be processed then, in the same way as other changes will be processed now.¶

  If and when such changes reach general acceptance, they will be explained in the appropriate subsection of Appendix A.¶

4.1. NFSv4 Security Terminology

In this section, we will define the security-related terminology used in this document. This is particularly important for NFSv4 because many of the terms terms related to security in previous specification may be hard to understand because their meanings have changed or have been used inconsistently, resulting in confusion.¶

The following terms are listed in alphabetical order:¶

• "Access Control" denotes any control implemented by a server peer to limit or regulate file system access to file system objects. It includes but is not limited to authorization decisions. Access control features can be divided into those which are "Discretionary" or "Mandatory" as described below.¶

• "ACL" or "Access Control List" denotes a structure used, like the mode (see below), to defines the privileges that individual users have with respect to a given file. These structures provide more options than modes with regard to the association of privileges with specific users or group and often provide a finer-graned privilege structure as well. This specification will have need to refer to two types of ACLs.¶

  The ACLs intended to be presented in the acl, sacl, and dacl attributes are called "NFSv4 ACLs". This ACL format, was modeled on the the semantics of the SMB ACL format which provide a privilege model substantially finer-grained than that provided by POSIX modes.¶

  [Consensus needed (Item #56a)]: Another ACL type derives from an attempt to define, within POSIX, a UNIX-oriented approach to ACLs which was published as a draft (POSIX 1003.1e draft 17), but subsequently withdrawn. Despite the withdrawal of this draft and the working group's decision to adopt a native NFsv4 ACL format based on SMB ACLs, this document will have to discuss these ACLs, which we will term "UNIX ACLs" because many server file systems do not support the finer-grained privilege model needed by the the NFSv4 ACL model and because many clients are built on systems whose only ACL-related API is based on the UNIX ACL model.¶

• "authentication" refers to a reliable determination that one making a request is in fact who he purports to be. Often this involves cryptographic means of demonstrating identity.¶

  This is to be distinguished from "identification" which simply provides a specified identity without any evidence to verify that the identification is accurate.¶

  In the past, these terms have been confused, most likely because of confusion engendered by th use of the term "authentication flavor" including flavors for which only identification is provided or which do not provide even identification.¶

• "authorization" refers to the process of determining whether a request is authorized, depending on the resources (e.g. files) to be accessed, the identity of the entity on whose behalf the request was issued, and the particular action to be performed.¶

  Depending on the type of request, the entity whose identity is referenced can be a user, a peer, or a combination of both.¶

  Authorization is distinct from authentication. However, performing authorization based on identities which have not been authenticated makes secure operation impossible since use of unauthenticated identities allows acceptance of requests that are not properly authorized if the sender has the ability, as it typically does, to pretend to be an authorized user/peer.¶

• "client" refers to the entity responsible for setting up a connection. In most cases the client and the requester reside on the same node but this not always the case for NFSv4 because of the possibility of callback requests in which the server makes some request of the client.¶

• "confidentiality" refers to the assurance provided, typically through encryption, that the contents of requests and responses are not inadvertently disclosed to unauthorized parties.¶

• "Discretionary Access Control" denotes forms of access control, that rely on a user, such as the owner, specifying the privileges that various users are to have.¶

• "Mandatory Access Control" denotes forms of access control that reflect choices made by the server peer and based on its policy and that are typically based on the identity of the client peer rather than the specific user making a request. While such access control is discussed in this document, it is important to note that many forms of mandatory access control are discussed by other NFsv4 documents and that there are forms that are not standardized.¶

• [Consensus Needed, Entire Bulleted Item (Items #21a, #57a)]: "Mode" designates a set of twelve flag bits used by POSIX-based systems to control access to the file with which it is associated. In NFSv4, there are represented by the REQUIRED attribute Mode.¶

  The three high-order flags are generally accessed only by the client while low-order bits are divided into three three-bit fields, which give, in order of decreasing numeric value, the privileges to be associated with, the owner of the file, other users in the group owning the file, and users not in the above two categories.¶

  In most cases, the privileges associated with each successive group are no greater than those for the previous group. Modes whose privileges are of this form are referred to as "forward-slope modes" because the privilege level proceeds downward as successive groups of users are specified. Cases in which the contrary possibility is realized are referred to as "reverse-slope modes".¶

• "peer" refer to the entity which is charged with requesting or performing a specified request as opposed to the entity on whose behalf the request is requested or performed, the principal;¶

• "principal" refers to the specific entity (e.g. user) on whose behalf a request is being made.¶

• "privacy", has in the past been used to refer, to what is now referred to as "confidentiality".¶

  over time, this usage has changed so that the word most often refers to applicability of data to a single individual and person's right to prevent its unauthorized disclosure¶

  As a result, many references to "privacy" in previous are no longer appropriate and really refer to confidentiality.¶

  The NFSv4 protocol has no way to determine whether particular data items raise privacy concerns (In the new sense). NFSv4 provides confidentiality whatever type of data is being accessed so that private data is kept private.¶

• "integrity" refers to the assurance that data in a request has not been modified in the process of transmission. Such an assurance is generally provided b means of a cryptographic hash of the requests or response.¶

• "requester" is the entity making a request, whether that entity is on the client-side, as it most often is (forward-direction request) or the server side, in th case of callback (reverse-direction requests)¶

• "responder" is the entity performing a request, whether that entity is on the server side, as it most often is (forward-direction request) or the client side, in the case of callbacks (reverse-direction requests.¶

• "server" refers to the entity to which the client connects. In most cases the client and the responder reside on the same node but this not always the case for NFSv4 because of the possibility of callback requests in which the server makes some request of the client.¶

4.2. NFSv4 Security Scope Limitations

This document describes the security features of the NFSv4 protocol and is unable to address security threats that are inherently outside the control of the protocol implementors. Such matters as out of this document's scope.¶

As a way of clarifying the threats that this document, and the threat analysis in Section 18.4 can and cannot deal with, we list below the potential threats discussed Section 3.1 of [nist-209] and review how, if at all, it is discussed in the current document. In cases in which the threat is dealt with in this document, distinctions are to be made between cases in which the issues have been dealt with directly or have been delegated to a lower layer on which the protocol is built and whether the issue has been addressed by the changes to NFSv4 security made by this document.¶

• Regarding the possibility of "Credential Theft or Compromise", this is not a matter that the NFSv4 protocols concern themselves with or can address directly, despite its importance for security. Depending on the auth flavor chosen, either the client (for AUTH_SYS) or a third-party (for RPCSEC_GSS), usually Kerberos, will be responsible for credential verification.¶

  Since experience has shown that credential compromise (e.g. through "phishing" attacks) is a common occurrence, this problem cannot be ignored, even though NFSv4's reliance on RPC facilities for authentication might be thought to make it out-of-scope as it would be RPC if had an effective solution to the issue. However, the urgency of the situation this issue is such that will be discussed in Section 18.4.2, even though no definitive solutions to this issue are likely before this document is completed and published.¶

  Regardless of such issues, the likelihood of such compromise has had a role in decisions made regarding the acceptance and use of "superuser" credentials. The possibility of such compromise is also relevant to implementation of means to synchronize credentials when they are managed by the client, as described in Section 18.4.6.1¶

• Regarding the possibility of "Cracking Encryption", prevention of this is responsibility of the NFSv4 protocols but it is one which has been delegated to RPC, so that its discussion in Security Considerations will rely on RPCSEC_GSS and RPC-with-TLS impelentations to manage the selectio and replacement of kets for encryption so as to limit the possibility of such unwanted encryption key discovery.¶

• Regarding the possibility of "Infection of Malware and Ransomware", NFSv4 has no direct role in preventing such infection, but does have an important role in limiting its consequences, by limiting the the ability of Malware to access or modify data, through the file access authorization model supported by NFSv4 to limit access to authorized users. Of course, malware will be able to execute on behalf of the user mistakenly invoking it but the authorization model will server to limit the potential damage.¶

  The possibility of vertical privilege escalation is of concern as regard the possible elevation to "superuser" privileges. For this reason, this document recommends that any such escalation not be effective on the server, even if it happens on local clients for which NFSv4 has no role.¶

  Execution of a ransomeware-based attack requires the attacker to have the ability to read existing data and replacing it with an encrypted version together with the ability to temporarily hide the encryption from ongoing operations by intercepting requests to read encrypted data and substitute the unencrypted data.¶

• Regarding the possibility of "Backdoors and Unpatched Vulnerabilities", it needs to be noted that the NFSv4 protocols do not specify any backdoors even though it is possible that might choose to provide such backdoors. Since it is not practical to specifically prohibit the existence of such backdoors nor would they be enforceable if written, this document will not attempt to do so. Instead, Section 18.2.3 will note the possibility of such backdoors and recommend against any such implementation, and include implementations containing backdoors in the category of insecure use that will not be dealt with in Section 18.4.¶

  Although it is expected that vulnerabilities will be due to incorrect implementations and thus outside the scope of this document, the possibility of a protocol design errors cannot be excluded. In dealing with such eventualities, it is likely that complete remediation would require co-ordinated changes on the client and server¶

• Regarding the possibility of "Privilege Escalation", NFSv4 has dealt with the possibility of vertical escalation by not allowing a client-local escalation to superuser privileges to be effective on the server.¶

  With regard to horizontal "escalation", NFSv4 provides for the use of various means RPC authentication of principals but relies on the client operating system to make sure that one user principal cannot masquerade as another.¶

• Regarding the possibility of "Human Error and Deliberate Misconfiguration", the approach taken is to limit the need for the server to make complicated decisions regarding the security requirements of each section of its namespace, with many opportunities for misconfiguration, if the chosen security requirements are insufficiently restrictive. This is in contrast to previous specifications which made such configuration the centerpiece of the security approach.¶

  Although it is possible to create configurations where certain data, generally publicly accessible, are to be made available without encryption, this is expected to be a rarely used option with the possibility of in-transit modification kept in mind before adopting such use.¶

• Regarding the possibility of "Physical Theft of Storage Media", this a matter which, while of concern to those deploying NFSv4 server, will be considered out-of-scope since there is nothing that the protocol could do to deal with this threat.¶

• Regarding the possibility of "Network Eavesdropping", when the protocol implementation follows the recommendations in this document, the protocol's use of RPC facilities is designed, through the consistent use of encryption to make it difficult for an attacker to have access to the data being transmitted, to modify it, or inject requests into an existing data stream.¶

  The possibility of an attacker with access to the network creating a new connection is best considered as a case of the attacker pretending to be a client and is addressed in Section 18.4.3.¶

• Regarding the possibility of "Insecure Images, Software and Firmware", while attention to such matters is important for those deploying NFSv4, it is important to note that these are matters outside the control the NFSv4, which has to assume that the infrastructure it is built is working properly. As a result, this document will not deal with the possibility of such threats.¶

5.1. Format of Id Strings in Authorization-related Attributes

Unlike the case in NFSv3, users and groups are represented in NFSv4 by UTF-8-encoded Unicode strings. These strings include:¶

• [Consensus needed (Item #57c)]: The values of the REQUIRED attribute Owner, as described in Section 5.3.3.¶
• [Consensus needed (Item #57c)]: The values of the REQUIRED attribute Owner_group, as described in Section 5.3.4¶
• Values within the "who" field within Access Control Entries as described in ACLSPEC(ACL-ace). These entries appear in the Acl, Dacl, and Sacl attributes.¶

[Author Aside, Including List (Item #59a)]: This section had to be extensively revised in order to be clearer and to allow more uses of string representations of numeric ids. The previous treatment, which is not reproduced here, can be found in Section 5.9 of [RFC8881]. In the author's opinion, major revisions were necessary because:¶

• The existing treatment, by leaving the methods of mapping between strings and local user ids unspecified, created an interoperability issue. Since providing this mapping in a secure way is difficult, the gap is likely to be filled in an a way vulnerable to attack.¶

  While the need for flexibility might make it impossible to specify mapping fully, the requirements for agreement between client and server needed to be specified more clearly and avoided where this can be done without impeding protool operation.¶

• There is inadequate attention to the additional implementation needs to support multiple domain values.¶

  The difficulties inherent in supporting all possible domains are obscured, with the very real security issues essentially ignored.¶

• The description of the domain value as "meant to be a DNS domain" is not clear and makes it harder to focus on the real difficulties involved in providing coherent secure means of providing necessary mappings for multiple domains.¶

• Some negarive characterizations of the use of numeric ids appear to be unjustified, leading to their non-use in situations in which artificially mapping to the name@domain format creates difficult issus while providing no benefit.¶

[Consensus Needed, Including Rest of Section (Items #52a, #59a)]:¶

These identifiers are represented by strings which have two possible formats:¶

• Strings of the format "name@domain" can be used with a number of important benefits. They are to be preferred in situations in which the mapping between names and the 32-bit numeric ids preferred by typical file systems is securely provided by means of the RPC auth flavor being used, as is the case when Kerberos services are made available through the use of RPCSECGSS.¶

  Servers would be able support multiple clients each with users from one or more domains, without requiring construction of a single list of users together with associated authentication info.¶

  The number of users would not be limited by existing hard-to-extend limits such as the POSIX/NFSv3 use of a 32-bit field to represent a user or group id.¶

• The string can be the ASCII representation o the 32-bit user or group id.¶

  This format is to preferred when the auth flavor AUTH_SYS is being used and in other cases in which the translation between user names and numeric ids is not available or creates security vulnerabilities.¶

These strings are assigned the xdr type utf8str_mixed because the name and domain portions are treated differently:¶

• The name portion is a utf8 string that is matched in a case- sensitive manner without any attempt to treat distinct canonically equivalent strings as the same.¶
• The handling of the domain portion, including recognition of equivalent string as the same is specified by the server. However, when multple domains are supported, case-insensitivity of ASCII characters is mandated by DNS, as well as translations to and from punycode-encoded forms.¶

It is expected that the client and server file systen will have their own local representation of users and user groups that is used for local storage or presentation to the end user. In addition, it is expected that when these attributes are transferred between the client and server and the format used is of the form name@domain, the local representation is translated to that orm¶

Although this allows for a client and server that both use the name@domain format and do not use the same local representation the ability to translate to a common format, there will often be cases in which both client and sever use the same internal format. This often happens with servers that only support a single domain. However, regrdless of whether these forms are, in fact, different, it is inecessary, when such translation is necessary, that client and server agree on the following:¶

• The set of valid strings that can appear as the domain portion of these identifiers. This set must have at least a single member but when this not the case, the set known to each client must be a subset of the set recognized by the server.¶

• For each valid domain string passed by a client and accepted by a server, the server and client must agree on a means by which the user designated by each user and group name can be mapped to a user or group so that each interprets all name@domain strings identically.¶

  The above applies whether the client and server use the same internal name repreentation or not.¶

When this translation is used, security principals and the groups they belong to need to be integrated within this arrangement since these identities need to be presented to the server in a form compatible with this new format. The specfics depend on the auth flavor providing the principal identfication:¶

• When auth flavors based on RPCSEC_GSS are used, the prinipal is generally identified in a form easily converted to the name@domain fornat. For example, with Kerberos, the kerberos user name can provide the name ortion while kerberos realm can serve as the domain portion.¶

• When the AUTH_SYS auth flavor is used, principals are identified by a 32-bit numeric identifier which can be generally teated in the same manner as numeric id's used locally within the file system to arive at a name with the domain being the one assigned that filesysten.¶

  The idenification of principals MAY be subject to filtering to eliminate users with a high level of privilege.¶

  In the case of servers with filesystems that support the use of multiple domains, clients for which use of AUTH_SYS is supported need to be assigned to a specific domain so that the principal identfications received are properly dealt with.¶

The translation used to interpret owner and group strings is not specified as part of the protocol. This allows various solutions to be employed, as long as the requirement mentioned above are satisfied and additionl security vulnerabilities are not created. For example, a local translation table may be consulted that maps a numeric identifier to the user@domain syntax. A name service may also be used to accomplish the translation. A server may provide a more general service, not limited by any particular translation (which would only translate a limited set of possible strings) by storing the owner and owner_group attributes in local storage without any translation or it may augment a translation method by storing the entire string for attributes for which no translation is available while using the local representation for those cases in which a translation is available.¶

[Author Aside:] The next sentence previously used the word "SHOULD. Since the author has been unable to determine valid reasons to do otherwise and has no reason believe that any exist, the new text uses "MUST". It is possible that the original author might have been thinking of "nobody" for this but that case needs to be addressed separately and will need to be clearly distinguished from "nobody@domain".¶

Servers that do not provide support for all possible values of the owner and owner_group attributes MUST return an error (NFS4ERR_BADOWNER) when a string is presented that has no translation, as the value to be set for a SETATTR of the owner, owner_group, or as the who value in an ACE within the acl, sacl, or dacl attributes. When a server does accept an owner or owner_group value as valid on a SETATTR (and similarly for the owner and group strings in an ACL), it is promising to return that same string when a corresponding GETATTR is done. Configuration changes (including changes from the mapping of the string to the local representation) and ill-constructed name translations (those that include aliasing) may make that promise impossible to honor. Servers need to make appropriate efforts to avoid a situation in which these attributes have their values changed when no real change to ownership has occurred.¶

The "domain" portion of the owner string will often be a DNS domain name, for example, user@example.org. Servers should accept as valid a set of users for at least one value of the domain portion. A server may treat other domains as having no valid translations. A more general service is provided when a server is capable of accepting users for multiple domains values but the following will need to be attended to:¶

• When use of AUTH_SYS is allowed, the translation of numeric ids to the form name@domain becomes problematic.¶

  The server will need to determine the domain value to be used based on the identity of the client peer.¶

• When numeric ids are used in the file system the handling of those ids needs to be modified to support multiple sets of users with one set for each of the various domain values supported.¶

  One option is to add a small (i.e. 8- or 16-bit) field to represent the chosen domain value will leaving the 32-bit id field as it is.¶

  It is also possible to map individual 32-bit id spaces into a single 32-bit id space. However, this involves eastablishing for each component id space a restricted mange, so that the mapping can be done without resulting in conflicting reverse mappings.¶

In cases in which translation is being used and there is no translation available to the client or server, the attribute value will be constructed without the "@". The absence of the @ from the user or group string signifies that no translation was available at the sender and that the receiver of the attribute should not use that string as a basis for translation into its own internal format. When the string consists of a numeric value with no leading zeroes it can be interprested as represeting the corresponding 32-bit numeric id. Even though such attribute values cannot be translated, they are still likely to be useful. In the case of a client, the attribute string may be used for local display of ownership. However, in the case in which the server receives such a string, it is less likely to be useful and might be harmful in that its use by the server might undercut the value of the translation to the form name@domain. In order to avoid these negative effects:¶

• When a numeric value is received in an attribute being set where an RPCSECGSS-based auth flavor is being used, the SETATTR MUST return an error (NFS4ERR_BADOWNER).¶
• When the auth flavor AUTH_SYS is being used, the seerver MAY accept user and groups identified by ids using the string numeric form and return them in that form. However, when it does so it still MUST return these values, in the name@domain form to those using other auth flavors.¶

[Author Aside, Including List (Items #52a, #60a)]: Something needs to be done about the failure of the corresponding text in RFC8881 to deal approriately with root@domain and nobody@domain, either to exclude these or include them with appropriate explanations. Specifically:¶

• With regard to the case of "nobody", the primary issue is that "nobody@domain" is not addressed, although "nobody", which should not occur, is.¶

  A further difficulty is the use of the term "anonymous" with the intention to include unidentified and unauthentcated users as well as those which are not approrpriately authenticated.¶

• With regard to the case of "root", the issue is not mentioned at all, leaving it unclear as to whether this sort of special handling is gone, or is dependent now on the name "root" or the mumeric id zero despite the geneneral deprecation of numeric ids.¶

  The possibility of support for multiple domains, creates additional complexity whih needs to be addressed.¶

[Consensus Needed (Item #60a)]: An owner string of the form "nobody@domain" may be used to designate an anonymous or unauthenticated user, which will be associated with a file created by a security principal whose identity is not determinable or cannot be mapped through normal means to the Owner attribute. Users and implementations of NFSv4.1 should avoid the use of the string "nobody" to identify an actual user.¶

[Consensus Needed (Item #52a)]: Certain id strings that map to known numeric values such as zero MAY be assigned special privileges with regard to operation authorization, allowing operations to be authoized that POSIX or ACL-based authentication might disallow. The granting of such privileges MUST NOT be based on the user name (e.g. "root"). Instead, the server MUST only use the principal id or information returned by a secure id mapping facility in assigning such privileges, which MAY be assigned differently based on the reliability of the authentication method used. When multiple domains are supported, the privilege assignments might be diffeent for diffeent domains, but the mapping of ids into a value used internally by the file systen is not to be considered in deciding about granting such privileges.¶

5.2. Table of Authorization-related Attributes

The list of authorization-related attributes appears in Table 1.¶

The meaning of the columns of the table are:¶

Name:

The name of the attribute.¶

Id:

The number assigned to the attribute. In the event of conflicts between the assigned number and minor version specification documents, the former is authoritative, but conflicts should be resolved with Errata to this document and/or the minor version specifcation document. See [errata] for the Errata process.¶

Data Type:

The XDR data type of the attribute.¶

Acc:

Access allowed to the attribute. R means read-only (GETATTR may retrieve, SETATTR may not set). W means write-only (SETATTR may set, GETATTR may not retrieve). R W means read/write (GETATTR may retrieve, SETATTR may set).¶

Defined in:

The section of this specification that describes the attribute.¶

5.3. POSIX-oriented Authorization-related Attributes

[Consensus Needed (Item #57d)]: The REQUIRED attributes Owner, Owning_group, and Mode enable use of a POSIX-based authorization model, as described in Section 8.3. Given that all of these attributes MUST be supported, this authorization model is always available.¶

These attributes are also of use when ACL-based authorization is in effect.¶

• The values of the Owner and Owner_group attributes affect the interpretation of the special "who" values "@OWNER" AND "@GROUP">¶

• Changes to the acl and dacl attributes will often result in corresponding changes to the mode attribute.¶

• Setting the mode attribute will often override the values in acl and dacl attributes, even if tere has been no changes to the authorzation-related bits.¶

  The set_mode_masked attribute can be set when changing bits in the mode, if the authorization-related bits are not to be changed and there is a need not to update authorization-related ACEs.¶

const MODE4_SUID = 0x800;  /* set user id on execution */
const MODE4_SGID = 0x400;  /* set group id on execution */
const MODE4_SVTX = 0x200;  /* save text even after use */
const MODE4_RUSR = 0x100;  /* read permission: owner */
const MODE4_WUSR = 0x080;  /* write permission: owner */
const MODE4_XUSR = 0x040;  /* execute permission: owner */
const MODE4_RGRP = 0x020;  /* read permission: group */
const MODE4_WGRP = 0x010;  /* write permission: group */
const MODE4_XGRP = 0x008;  /* execute permission: group */
const MODE4_ROTH = 0x004;  /* read permission: other */
const MODE4_WOTH = 0x002;  /* write permission: other */
const MODE4_XOTH = 0x001;  /* execute permission: other */

5.4. ACL-based Authorization-related Attributes

The following ACL-related attributes (all OPTIONAL are described in detail in the new document devoted to ACL handling [I-D.dnoveck-nfsv4-acls].¶

• The per-fs Aclsupport attribute.¶
• The per-fs Aclfeatures attribute being added as an extension to NFSv4.2..¶
• The per-object Acl attribute.¶
• The per-object Dacl attribute added in NFSv4.1..¶
• The per-object Sacl attribute added in NFSv4.1¶

5.5. Authorization-related Attributes for MAC

6.1. Previous Treatment of ACLs

The description of ACL-related attributes and ACEs that appeared in previous documents embodied a flawed appoach to protocol description that makes in unsuitable for use as a basis for a descrption intended to appear and in a standards-track document. As the goal of this effort is to produce a helpful standards-track document, these item will be described using a new approach, as described in Section 6.2.¶

This same approach was used in a number of documents published as Proposed Standards, despite its flaws. These included RFCs now obsoleted ([RFC3010], [RFC3530], and [RFC5661]) and a number RFCs to be obsoleted when this document is published as an RFC ([RFC7530] and [RFC8881]). Although some changes were made in the transition between minor versions, the essence of the approach remained the same.¶

The goal that led to this flawed approach remains as it was. There is need to support an extended ACL model, with a fine-grained permission model and other helpful extensions togther with providing support for a more limited ACL model with a more direct connection to POSIX semantics, very similar to that defined in the withdrawn POSIX ACL draft implemented by a number of file sytems implemented on UNIX systems.¶

The approach actually taken to this need had the folloing elements:¶

• The definition of the Acl attribute was based on the extended ACL model, rather than simpler, more POSIX-oriented core.¶

• To allow servers implement the UNIX ACL subset to be considered compliant, the specification was written providing an unacceptable degree of leeway for the server to implement many elements in pretty much whatever way it chose.¶

  This approach allowed UNIX ACls, NFSv4 ACLs and hybrids of the two to be considerd compliant, as well as many many unantcipated vaiants. In any case, the client had no way of determining the ACL semantics being implemented.¶

6.2. New Approach to Treatment of ACLs

The new approach, which will be explained in more detail in the companion document [I-D.dnoveck-nfsv4-acls] takes a very different approach, with the following chacteristics:¶

• The UNIX ACL subset is the basis of the canonical description which can asssume is available, if the Acl attribute is supported.¶

• Additions to this core are treated as OPTIONAL extensions. each of which might or might not be suported by a server implementing the Acl atribute.¶

• A new per-fs OPTIONAL attribute is defined to allow clients to determine which ACL model extensions are supported, if any.¶

This new approach is embodied in the companion document [I-D.dnoveck-nfsv4-acls]. As a result, that document has two functions:¶

• Together with this document it defines security for all minor versions, updating [RFC7530] and, together with the other rfc5661bis documents, onsoleteing [RFC8881]. This applies to all NFSv4 minor versions.¶

• Defining an extension that allows clients to determine the ACL extensions supported by any given file system. This applies to nFSv4.2 and subsequent minor versions.¶

8.1. Attributes for User-based File Access Authorization

NFSv4.1 provides for multiple authentication models, controlled by the support for particular recommended attributes implemented by the server, as discussed below:¶

• Consensus Needed (Item #57e)]: The REQUIRED attributes owner, owning_group, and mode enable use of a POSIX-based authorization model, as described in Section 8.3. Since these attributes are always supported, this authorization model is aways available.¶

• [Consensus Needed (Item #17a)]: The acl attribute (or the attribute dacl in NFSv4.1) can provide an ACL-based authorization model as described in Section 8.4 as long as support for ALLOW and DENY ACEs is provided.¶

  [Consensus Needed (Items #17a)]: When some of these ACE types are not supported attribute, this authorization model is unavailable, since there are some modes that cannot be represented as a corresponding NFSv4 ACL, when using only a single ACE type. See ACLSPEC(AUTHCOMB-attr) for details¶

8.2. Handling of Multiple Parallel File Access Authorization Models

NFSv4 ACLs and modes represent two well-established models for specifying user-based file access permissions. NFSv4 provides support for either or both depending on the attributes supported by the server and, in cases in which both NFSv4 ACLs and the mode attribute are supported, the actual attributes set for a particular object.¶

• [Consensus Needed (item #18b)]: When the attributes mode, owner, owner group are all supported, the posix-based authorization model, described in Section 8.3 can be used.¶

• [Consensus Needed (Items #17b, #18b)]: When the acl (or dacl) attribute is supported together with both of the ACE types ALLOW and DENY, the acl based authorization model, described in Section 8.4 can be used as long as the attributes owner and owner_group are also supported.¶

[Consensus Needed (item #18b)]: While formally recommended (essentially OPTIONAL) attributes, it appears that the owner and owner_group attributes need to be available to support any file access authorization model. As a result, this document will not discuss the possibility of servers that do not support both of these attributes and clients have no need to support such servers.¶

When both authorization models can be used, there are difficulties that can arise because the ACL-based model provides finer-grained access control than the POSIX model. The ways of dealing with these difficulties appear later in this section while more detail on the appropriate handling of this situation, which might depend on the minor version used, appears in Section 10.¶

The following describe NFSv4's handling in supporting multiple authorization models for file access.¶

• If a server supports the mode attribute, it needs to provide the appropriate POSIX semantics if no ACL-based attributes have ever been assigned to object. These semantics include the restriction of the ability to modify the mode, owner and owner-group to the current owner of the file.¶

• If a server supports ACL attributes, it needs to provide NFSv4 ACL semantics as described in this document for all objects for which the ACL attributes have actually been set. This includes the ACL-based restrictions on the authorization to modify the mode, owner and owner_group attributes.¶

• On servers that support the mode attribute, if ACL attributes have never been set on an object, via inheritance or explicitly, the behavior is to be the behavior mandated by POSIX, including tthose provisions that restrict the setting of authorization-related attributes.¶

• On servers that support the mode attribute, if the ACL attributes have been previously set on an object, either explicitly or via inheritance:¶

  • [Previous Treatment]: Setting only the mode attribute should effectively control the traditional UNIX-like permissions of read, write, and execute on owner, owner_group, and other.¶

    [Author Aside]: It isn't really clear what the above paragraph means, especially as it governs the handling of aces designating specific users and groups which are not the owner and have no overlap with the owning group¶

    {Consensus Needed (Item #19a)]: Setting only the mode attribute, will result in the access of the file being controlled just it would be if the existing acl did not exist, with file access decisions as to read made in accordance with the mode set. The ALLOW and DENY aces in the ACL will reflect the modified security although there is no need to modify AUDIT and ALARM aces or mask bits not affected by the mode bits, such as SYNCHRONIZE.¶

    [Author Aside]: the above may need to modified to reflect the resolution of Consensus Item #??.¶

  • [Previous Treatment]: Setting only the mode attribute should provide reasonable security. For example, setting a mode of 000 should be enough to ensure that future OPEN operations for OPEN4_SHARE_ACCESS_READ or OPEN4_SHARE_ACCESS_WRITE by any principal fail, regardless of a previously existing or inherited ACL.¶

    [Author Aside]: We need to get rid of or provide some some replacement for the subjective first sentence. While the specific example give is unexceptionable, it raises questions in other cases as to what would constitutes "reasonable semantics". While the resolution of such questions would be subject to dispute, the author believes that consensus item #19a deals with the matter adequately. As a result he proposes, that the that this bullet be removed and that the second-level list be collapsed to a single paragraph.¶

• Although [RFC7530] and [RFC8881] present different descriptions of the specific semantic requirements relating to the interaction of mode and ACL attributes, the difference are quite small, with the most important ones deriving from the absence of the set_mode_masked attribute. The unified treatment in Section 10 will indicate where version-specific differences exist.¶

8.3. Posix Authorization Model

Uses the REQUIRED attributes Mode, Owner, amd Owner_group to authorize actions applying to files and directories. Aside from the conversion of Owner and Owner_group to the forms of strings, this mirrors authorization in NFSv3, in being based on semantics defined by POSIX, originally used to control local finle access.¶

There are some potential differences to be notesd:¶

• The use of "owner-override" semanctics is allowed. This vaiant of autorization semantics was necessary in earlier NFS vesions, to allow continued access to open files even when the mode had been changed subequent to the open.¶

  Even though it is no longer necessary because NFSv4 is a stateful protocol, it is still allowed.¶

• [Consensus needed (Item #6b)]: The bit MODE_SVTX within the Mode attrbute on directories affects the authorization of REMOVE and RENAME.¶

8.4. ACL-based Authorization Model

While the relevant details appear in [I-D.dnoveck-nfsv4-acls], this model can be helpfully smmarize as follows:¶

The use of ACLs, via setting the OPTIONAL attributes Acl and Dacl, allows a more flexible approch to authorization in that the individual entries (ACEs) within the ACL can assign different privileges to different users or groups of users.¶

The ACEs within an ACL are examined in sequence with those designating the user requesting or a group to which that user belongs affecting the authorization decision to see if a needed authorzation for the requested operation is granted, or, depending the ACE type, denied.¶

Once all requested autnorization are granted, the operation is authorized and can proceed. If any requested authorzation is denied, the operation is not authotized. In either case, this terminates the processing of ACEs.¶

9.1. Handling of ACCESS and OPEN Operations

The primary means by which client-side users find out whether particular operations are authorized is to attempt those operations and have them executed successfully or rejected with an error. However, clients can use the ACCESS operation to determine in advance whether operations are permitted done without actually attempting those operations.¶

The use of the ACCESS operation is of particular importance when ACLs exist. This is due in part to the complexity of the ACLs but also derives from cases in which the client is incapable of determining whether a specific ACE applies to a particular request, as when some of the special who-field values are used in an ACE.¶

An additional difficulty precluding the client performing its own interpretation of an ACL concerns the existing specifications' lack of clear requirements for server support, making it difficult for the client to determine how a server might choose to behave in some cases. Even if this lack of specificity were remedied by clarifying the specification servers would still exist that reflected the previously valid scope of acceptable server behavior. As a result, it is impossible for the client to predict determine the choices a server might make without using ACCESS.¶

Similarly, the OPEN operation can determine acces rights in advance of actual access. There are two differences from ACCESS that clients need to be aware of, since they might require use of ACCESS in addition to OPEN or make the results of OPEN incompatible with the results of attempting IO operations.¶

• The permission model of OPEN is coarser-grained than that provided for by ACCESS or the ACE masks defined as part of the definition of NFSv4 ACLs.¶

  When reading files, the fact that OPEN treats requests to open a file to read it and to execute it together means that the client will need to do an ACCESS operation to determine whether particular reads are to be allowed.¶

  When writing files, the fact OPEN that does not distinguish between WRITEs which extend the file and those that modify bytes already written means that the client will need to get the permission information using ACCESS or be prepared to have some set of WRITES on a file open for Write rejected as unauthorized.¶

• [Consensus needed, Entire Bulleted Item (Item #23a)]: Just as with ACCESS, the granting of permission does not foreclose subsequent permission changes, there are good reasons for servers to provide ways of allowing IO allowed at OPEN time to continue even in the face of permission changes that would normally be expected to make the IO operation invalid. These reasons derive from the fact that local operations work this way and that it is is often necessary for requests over NFSv4 to behave just as they would if done locally. As a result, it is desirable for servers to provide this expected behavior in some way since application programs have come to depend on it.¶

  One way of handling this situation is described in Section 9.2. It deals with the most common of these situations in which the owner of the is both writing a file and seeking to make it subsequently unwritable, and while it does not deal with all such situations, it has proven satisfactory for many NFS protocols over a long time period¶

  Another way of dealing with this situation involves the server explicitly using the stateid to reference a particular open made by a user, and avoiding reverification of access for READs and WRITEs made by that user since that verification was made at OPEN time. To do this safely, the server needs to have authenticated principals and client peers and, in order to prevent man-in-middle attacks, it s necessary for all connections on which stateids are sent to provide encryption.¶

9.2. Server Considerations

The server uses the mode attribute or the acl attribute applying the algorithm described in ACLSPEC(AUTHFA-proc-ace) to determine whether an ACL allows access to an object.¶

[Author Aside, Including List]: The list previously in this section (now described as "Previous Treatment" combines two related issues in a way which obscures the very different security-related consequences of two distinct issues:¶

• In some cases an operation will be authorized but is not allowed for reasons unrelated to authorization.¶

  This has no negative effect on security.¶

• The converse case can have troubling effects on security which are mentioned in this section and discussed in more detail in Section 18¶

[Author Aside, Including List]: The items in that list have been dealt with as follows:¶

• The first and sixth items fit under the first (i.e. less troublesome) of these issues. They have have been transferred into an appropriate replacement list.¶

• The third item is to be deleted since it does not manifest either of these issues. In fact, it refers to the semantics already described in ACLSPEC(ACL-maskd).¶

• The second, fourth and fifth items need to be addressed in a new list dealing with the potentially troublesome issues arising from occasions in which the access semantics previously described are relaxed, for various reasons.¶

  Included are cases in which previous specifications explicitly allowed this by using the term "MAY" and others in which the existence of servers manifesting such behavior was reported, with the implication that clients need to be prepared for such behavior.¶

[Previous Treatment, Including List (Items #22a, #41a, #52b)]: However, these attributes might not be the sole determiner of access. For example:¶

• In the case of a file system exported as read-only, the server will deny write access even though an object's file access attributes would grant it.¶
• Server implementations MAY grant ACE4_WRITE_ACL and ACE4_READ_ACL permissions to prevent a situation from arising in which there is no valid way to ever modify the ACL.¶
• All servers will allow a user the ability to read the data of the file when only the execute permission is granted (e.g., if the ACL denies the user the ACE4_READ_DATA access and allows the user ACE4_EXECUTE, the server will allow the user to read the data of the file).¶
• Many servers implement "owner-override semantics" in which the owner of the object is allowed to perform accesses that are denied by the ACL or mode bits This may be helpful, for example, to allow users continued access to open files on which the permissions have changed.¶
• Many servers provide for the existence of a "superuser" that has privileges beyond an ordinary user. The superuser may be able to read or write data or metadata in ways that would not be permitted by the ACL or mode attributes.¶
• A retention attribute might also block access otherwise allowed by ACLs (see Section 5.13 of [RFC8881]).¶

[Consensus Needed, Including List (Item #22a)]: It needs to be noted that, even when an operation is authorized, it may be denied for reasons unrelated to authorization. For example:¶

• In the case of a file system exported as read-only, the server will deny write access even though an object's file access attributes would authorize it.¶
• A retention attribute might also block access otherwise allowed by ACLs (see Section 5.13 of [RFC8881]).¶

[Author aside, including List, (Item #22a)]: Unlike other cases in which previous specs have granted permission to the server to expand allowed behavior, in the case of "owner-override semantics", the need for supercession does not arise from security issues but instead derives from:¶

• The unacceptable implication in the superseded text that the fact that a server chooses to do something, means that the client needs to accept that behavior.¶

• The lack of any definition of the term "owner-override semantics. Subsequent investigation has led to the conclusion that the semantics being referred to derive from the following material in [RFC1813]:¶

  • Another problem arises due to the usually stateful open operation. Most operating systems check permission at open time, and then check that the file is open on each read and write request. With stateless servers, the server cannot detect that the file is open and must do permission checking on each read and write call. UNIX client semantics of access permission checking on open can be provided with the ACCESS procedure call in this revision, which allows a client to explicitly check access permissions without resorting to trying the operation. On a local file system, a user can open a file and then change the permissions so that no one is allowed to touch it, but will still be able to write to the file because it is open. On a remote file system, by contrast, the write would fail. To get around this problem, the server's permission checking algorithm should allow the owner of a file to access it regardless of the permission setting. This is needed in a practical NFS version 3 protocol server implementation, but it does depart from correct local file system semantics. This should not affect the return result of access permissions as returned by the ACCESS¶

  As a result, it appears that NFSv4 servers have implemented semantics different from that provided for by POSIX that were defined in order to deal with lack of an open operation in NFsv3 in the context of NFSv4, which does have an open operation.¶

• The need to make it clear that an alternate approach to the issue of permission change after open is to allow the OPEN's permission check to be dispositive as long as the appropriate security infrastructure is in place to allow client stateids to be trusted.¶

[Consensus Needed, (Item #22a)]: There are also cases in which the converse issue arises, so that an operation which is not authorized as specified by the mode and ACL attributes is, nevertheless, executed as if it were authorized. Because previous NFSv4 specifications have cited the cases listed below without reference to the security problems that they create, it is necessary to discuss them here to provide clarification of the security implications of following this guidance, which might now be superseded. These cases are listed below and discussed in more detail in Section 18.1.3.¶

[Consensus Needed, Including List (Item #22a, #41a, #52b)]: In the following list, the treatment used in [RFC8881] is quoted, while the corresponding text in [RFC7530] is essentially identical.¶

• [RFC8881] contains the following, which is now superseded:¶

  • Server implementations MAY grant ACE4_WRITE_ACL and ACE4_READ_ACL permissions to prevent a situation from arising in which there is no valid way to ever modify the ACL.¶

  While, as a practical matter, there do need to be provisions to deal with this issue, the "MAY" above is too broad,in that it describes the motivation without any limits providing appropriate restriction on the steps that might be taken to deal with the issue. See Section 18.1.3 for the updated treatment of this issue.¶

• [RFC8881] contains the following, which is now superseded:¶

  • Many servers implement owner-override semantics in which the owner of the object is allowed to override accesses that are denied by the ACL. This may be helpful, for example, to allow users continued access to open files on which the permissions have changed.¶

  The principal problem with the above statement is the lack of a clear definition of the term "owner-override semantics". Also, it needs to be made clear that the fact that a server manifests a particular behavior does not imply that it is valid according to the protocol specification.¶

  In this case, investigation has led to the conclusion that the semantics being referred to derive from the discussion of NFSv3 semantics appearing Section 4.4 of [RFC1813] and that this handling has been implemented in a number of NFSv4 servers, despite the fact that NFSv4, unlike NFSv3, does have an open operation.¶

  With regard to the second sentence of the quotation above, it is not clear whether it is helpful or hurtful to allow continued access to open files which have become inaccessible due to changes in security and it is not clear that the working group will make a decision on the matter in this document, despite the obvious security implications. In any case, the resolution is unlikely to depend on whether the owner is involved.¶

  Since this divergence from POSIX semantics is unlikely to result in security issues, we can clarify the above by saying that a server MAY diverge from POSIX semantics by always allowing READ or WRITE to be done by the file owner but that this divergence MUST NOT affect the handling of OPEN and ACCESS. It also needs to be noted that servers could address the issue of preventing permission change affecting that handling of READs and WRITEs of open files as described in Section 9.1.¶

• [RFC8881] contains the following, which is now superseded:¶

  • Many servers have the notion of a "superuser" that has privileges beyond an ordinary user. The superuser may be able to read or write data or metadata in ways that would not be permitted by the ACL or mode attributes.¶

  While many (or almost all) systems in which NFSv4 servers are embedded, have provisions for such privileged access to be provided, it does not follow that NFSv4 servers, as such, need to have provision for such access.¶

  Providing such access as part of the NFSv4 protocols, would necessitate a major revision of the semantics of ACL including such troublesome matters as the proper handling of AUDIT and ALARM ACEs in the face of such privileged access.¶

  Because of the effect such unrestricted access might have in facilitating and perpetuating attacks, Section 18.1.3 will the new approach to this issue, while Section 18.4.1, will explain how such access is addressed in the threat analysis.¶

9.3. Client Considerations

[Previous Treatment]: Clients SHOULD NOT do their own access checks based on their interpretation of the ACL, but rather use the OPEN and ACCESS operations to do access checks. This allows the client to act on the results of having the server determine whether or not access is to be granted based on its interpretation of the ACL.¶

[Author Aside]: With regard to the use of "SHOULD NOT" in the paragraph above, it is not clear what might be valid reasons to bypass this recommendation. Perhaps "MUST NOT" or "are not advised to" would be more appropriate.¶

[Consensus Needed (Item #23b)]: Clients are not expected to do their own access checks based on their interpretation of the ACL, but instead use the OPEN and ACCESS operations to do access checks, where this is possible. This allows the client to act on the results of having the server determine whether or not access is to be granted based on its interpretation of the ACL, rather then the client's which might be different. For a full discussion of limitations on the use of ACCESS and appropriate client approaches to deal these limitations, see Section 9.1.¶

[Previous Treatment]: Clients must be aware of situations in which an object's ACL will define a certain access even though the server will not enforce it. In general, but especially in these situations, the client needs to do its part in the enforcement of access as defined by the ACL.¶

[Author Aside]: Despite what is said later, the only such case I know of is the use of READ and EXECUTE where the client, but not the server, has any means of distinguishing these. I don't know of any others. If there were, how could ACCESS or OPEN be used to verify access?¶

[Consensus Needed (Item #23b)]; Clients need to be aware of situations in which an object's ACL will define a certain access even though the server is not in position to enforce it because the server does not have the relevant information, such as knowing whether a READ is for the purpose of executing a file. Because of such situations, the client needs to do be prepared to do its part in the enforcement of access as defined by the ACL.¶

To do this, the client will send the appropriate ACCESS operation prior to servicing the request of the user or application in order to determine whether the user or application is to be granted the access requested.¶

[Previous Treatment (Item #24a)]: For examples in which the ACL may define accesses that the server doesn't enforce, see Section 9.2.¶

[Author Aside]: The sentence above is clearly wrong since that section is about enforcement the server does do. The expectation is that it will be deleted as part of Consensus Item #24a.¶

14.1. Identification vs. Authentication

It is necessary to be clear about this distinction which has been obscured in the past, by the use of the term "RPC Authentication Flavor" in connection with situation in which identification without authentication occurred or in which there was neither identification nor authentication involved. As a result, we will use the term "Auth Flavors" instead¶

14.2. Items to be Identified

Some identifier are not security-relevant and can used be used without authentication, given that, in the authorization decision, the object acted upon needs only to be properly identified¶

• File names are of this type.¶

  Unlike the case for some other protocols, confusion of names that result from internationalization issues, while an annoyance, are not relevant to security. If the confusion between LATIN CAPITAL LETTER O and CYRILLIC CAPITAL LETTER O, results in the wrong file being accessed, the mechanisms described in Section 8 prevent inappropriate access being granted.¶

  Despite the above, it is desirable if file names together with similar are not transferred in the clear as the information exposed may give attackers useful information helpful in planning and executing attacks.¶

• The case of file handles is similar.¶

Identifiers that refer to state shared between client and server can be the basis of disruption attacks since clients and server necessarily assume that neither side will change the state corpus without appropriate notice.¶

While these identifiers do not need to be authenticated, they are associated with higher-level entities for which change of the state represented by those entities is subject to peer authentication.¶

• Unexpected closure of stateids or changes in state sequence values can disrupt client access as no clients have provision to deal with this source of interference.¶

  While encryption may make it more difficult to execute such attacks attackers can often guess stateid's since server generally not randomize them.¶

• Similarly, modification to NFSv4.1 session state information can result in confusion if an attacker changes the slot sequence by assuring spurious requests. Even if the request is rejected, the slot sequence is changed and clients may a difficult time getting back in sync with the server.¶

  While encryption may make it more difficult to execute such attacks attackers can often guess slot id's and obtain sessinid's since server generally do not randomize them.¶

• ¶

it is necessary that modification of the higher-levell entities be restricted to the client that created them.¶

• For NFSv4.0, the relevant entity is the clientid.¶

• for NFSv4.1, the relevant entity is the sessionid.¶

Identifiers describing the issuer of the request, whether in numeric or string form always require authentication.¶

14.3. Authentication Provided by specific RPC Auth Flavors

Different auth flavors differ quite considerably, as discussed below;¶

• When AUTH_NONE is used, the user making the request is neither authenticated nor identified to the server.¶

  Also, the server is not authenticated to the client and has no way to determine whether the server it is communicating with is an imposter.¶

• When AUTH_SYS is used, the user making is the request identified but there no authentication of that identification.¶

  As in the previous case, the server is not authenticated to the client and has no way to determine whether the server it is communicating with is an imposter.¶

• When RPCSEC_GSS is used, the user making the request is authenticated as is the server peer responding.¶

14.4. Authentication Provided by other RPC Security Services

Depending on the connection type used, RPC may provide additional means of authentication. In contrast with the case of RPC auth flavors, any authentication happens once, at connection establishment, rather than on each RPC request. As a result, it is the client and server peers, rather than individual users that are authenticated.¶

• For many types of connections such as those created TCP without RPC-with-TLS and RPC-over-RDMA version 1, there is no provision for peer authentication.¶

  As a result use of AUTH_SYS using such connections is inherently problematic.¶

• Some connection types provide for the possibility of mutual peer authentication. These currently include only those established by RPC-with-TLS. However, given the value of peer authentication, there is reason to believe further means of providing such services will be defined.¶

15.1. Data Security Provided by Services Associated with Auth Flavors

The only auth flavor providing these facilities is RPCSEC_GSS. When this auth flavor is used, data security can be negotiated between client and server as described in Section 16. However, when data security is provided for the connection level, as described in Section 15.2, the negotiation of privacy and integrity support, provided by the auth flavor, is unnecessary,¶

Other auth flavors, such as AUTH_SYS and AUTH_NONE have no such data security facilities. When these auth flavors are used, the only data security is provided on a per-connection basis.¶

15.2. Data Security Provided for a Connection by RPC

RPC, in many case, provide data security for all transactions performed on a connection, eliminating the need for that security to be provided or negotiated by the selection of particular auth flavors, mechanisms, or auth-flavor-associated services.¶

16.1. Dealing with Multiple Connections

[Author Aside]: All unannotated paragraphs in this section are considered part of Consensus Item #32b.¶

Because effective security will require both an appropriate auth flavor (and possibly services provide by the auth flavor) together with appropriate connection characteristics, it is often necessary that clients and server be aware of connection characteristics:¶

• When multiple connections with different security-related characteristics, are used to access a server, the clients needs to ensure that each request is issued on an appropriate connection.¶

• Similarly, in such situations, the server needs to be aware of the security-related characteristics for the connection on which each request is received, in order to enforce its security policy.¶

Depending on how the client and server implementations are structured, implementations may have to be changed to accomplish the above.¶

In the case of NFSv4.1 and above, the protocol requires that requests associated with a given session only be issued on connections bound to that session and accepted by the server only when that binding is present. This makes it likely that clients or servers will be able to correctly associate requests with the appropriate connections although additional work might be necessary to enable them to determine, for any given connection, its security characteristics.¶

In the case of NFSv4.0, no such binding is present in the protocol so that, depending on existing implementations' layering, channel binding functionality might have to be added.¶

17.1. Desirable Additional Security Facilities

[Author Aside]: All unannotated paragraphs in this section are to be considered part of Consensus Item #35a.¶

[Author Aside]: This section is basically an outline for now, to be filled out later based on Working Group input, particularly from Chuck Lever who suggested this section and has ideas about many of the items in it.¶

• Security for data-at-rest, most probably based on facilities defined within SAN.¶

• Support for content signing.¶

• Revision/extension of labelled NFS to provide true interoperability and server-based authorization.¶

• Work to provide more security for RDMA-based transports. This would include the peer authentication infrastructure now being developed as part of RPC-over-RDMA version 2. In addition, there is a need for an RDMA-based transport that provides for encryption, which might be provided in number of ways.¶

18.1. Changes in Security Considerations

Beyond the needed inclusion of a threat analysis as Section 18.4 and the fact that all minor versions are dealt with together, the Security Considerations in this section differ substantially from those in [RFC7530] and [RFC8881]. These differences derive from a number of substantive changes in the approach to NFSv4 security presented in [RFC7530] and [RFC8881] and that appearing in this document.¶

These changes were made in order to improve the security of the NFSv4 protocols because it had been concluded that the previous treatment of these matters was in error, leading to a situation in which NFSv4's security goals were not met. As a result, this document supersedes the treatment of security in earlier documents, now viewed as incorrect. However, it will, for the benefit of those familiar with the previous treatment of these matters, draw attention to the important changes listed here.¶

• There is a vastly expanded range of threats being considered as described in Section 18.1.1¶

• New facilities provided by RPC on a per-connection basis can be used to deal with security issues, as described in Section 18.1.2. These include the use encryption on a per-connection basis, and the use of peer mutual authentication, to mitigate the security problems that come with the use of AUTH_SYS.¶

• The handling of identities with superuser privileges is no longer part of NFSv4 semantics, even though many platforms on which NFSv4 servers are implemented continue to depend, for local operation, on the existence of such identities.¶

  NFSv4 servers SHOULD NOT provide for such unrestricted access since doing so would provide a means by which an escalation-of-privilege on a client could be used to compromise a server to which it was connected, affecting all clients of that server.¶

  In connection with the use of "SHOULD NOT" above, and similar uses elsewhere, it is to be understood that valid reasons to do other than recommended are limited to the difficulty of promptly changing existing server implementations and the need to accommodate clients that have become dependent upon the existing handling. Further, those maintaining or using such implementations need to be aware of the security consequences of such use as well as the fact that clients who become aware of this characteristic may not be inclined to store their data on such a system.¶

• The appropriate handling of ACL-based authorization and necessary interactions between ACLs and modes is now specified in this standards-track document rather it being assumed that the behavior of server implementations needs to be accepted and deferred to.¶

18.2. Security Considerations Scope

18.3. Major New Recommendations

18.4. Threat Analysis

19.1. New Authstat Values

[Author Aside]: All unannotated paragraphs in this section are to be considered part of Consensus Item #32d.¶

The following new authstat values are necessary to enable a server to indicate that the server's policy does not allows requests to be made on the current connection because of security issues associated with connection type used. In the event they are received, the client needs to establish a new connection.¶

• The value XP_CRYPT indicates that the server will not support access using unencrypted connections while the current connection is not encrypted.¶

• The value XP_CPAUTH indicates that the server will not support access using connections for which the client peer has not authenticated itself as part of connection while the current connection has not been set up in that way.¶

19.2. New Authentication Pseudo-Flavors

[Author Aside]: All unannotated paragraphs in this section are to be considered part of Consensus Item #32e.¶

The new pseudo-flavors described in this section are to be made available to allow their return as part of the response to the SECINFO and SECINFO_NONAME operations. How these operations are to used to negotiate the use of appropriate auth flavors and associated security-relevant connection characteristics is discussed in Section 16.¶

The following pseodo-flavors are to be defined:¶

• PFL_CRYPT is returned to indicate that subsequent secinfo entries are to considered limited to use on connections for which transport-level encryption is provided.¶

• PFL_CRYPTMPA is returned to indicate that subsequent secinfo entries are to be considered limited to use on connections on which mutual peer authentication has been provided at connection setup.¶

• PFL_ANY is returned to indicate that subsequent secinfo entries are not to be considered limited to any particular type of connection.¶

A.1. Motivating Security Changes

A.2. Need for Clarifying Changes

The need to make the major changes discussed in Appendix A.1 has meant that much text dealing with security has needed to be significantly revised or rewritten. As a result of that process, many issues involving unclear, inconsistent, or otherwise inappropriate text were uncovered and now need to be dealt with.¶

While the author believes that changes in this situation are necessary, the fact that we are changing a document adopted by consensus requires the working group to be clear about the acceptability of the changes. This applies to both the troublesome issues discussed in Section 3.4 and to the other changes included below.¶

The primary source of these not-clearly specified elements is the disparate goals of many participants in the specification process:¶

• For many participants, the ACL definition needed to include a large part of Windows-oriented semantics, including elements clearly distinct from the Unix semantics supported by earlier versions of NFS.¶

• For other participants, such semantic extensions were of little interest.¶

While the author believes such changes are necessary, the fact that we are changing a document adopted by consensus requires the working group to be clear about the acceptability of the changes. This applies to both the troublesome issues discussed in Section 3.4 and to the other changes included below.¶

Because of the need for re-organization of subjects relating to authorization, the ordering of the list follows the text of the current version which may differ considerably from that in earlier versions of the I-D.¶

The subjects related to authorization needed to be reorganized because of the following complexities:¶

• There are three kinds of authorization to deal with.¶

• The most important kind of authorization, file access authorization, is controlled by mode attributes and a set of ACL-related attributes, creating multiple authorization models and a need to coordinate them.¶

• ACLs, besides their prominent role with regard to file access authorization, provide control of alarm and audit facilities.¶

As a result, these matters are dealt with in the eight top-level sections within this document and the associated ACL document from 5 through 13. Of these, three sections, Sections 6, 9 and 13, have become brief overiews wjich introduce matters dealt with in more detail in the companion ACL document [I-D.dnoveck-nfsv4-acls].¶

• Section 5,a new top-level section, describes all the authorization-related attributes. Although this section cover all of the attributes, details relating to the ACL-related ones appear in the companion ACL document.¶

• Section 6, a new top-level section, provide an introduction to the structure of ACLs with the corresponding detail appearing in the companion ACL document [I-D.dnoveck-nfsv4-acls]. Later their uses are described later sections 8 through 10 and in Section 13 of which the last two serve as introductions to matters dealt with in the companion ACL document.¶

• In Section 7, there is a discussion of authorization in general. This includes a discussion of user-based file authentication, NFSv4's approach to mandatory access control, and a discussion of locking state authorization, each of which is described in later top-level sections.¶

• File access authorization is described primarily in Section 8. Because this authorization can be controlled by either the mode attribute or various ACL-related attributes, there is an introduction to these authorization- determining attributes in Section 8.1.¶

  Because file access authorization can be controlled by the posix-related authorization attributed decribed in Section 5.3 and by ACL-related attributes described in Section 5.4, there is an introduction to how these relate in Section 8.2 followed by discussion of the semantics of the two kinds of supported ACLs in Section 8.4.¶

  Issues related to the existence of two separate file access are dealt with in Sections 9, dealing with the similarities between these two, and 10 dealing with their co-existence and potential conflict.¶

• Sections 11 through 13 deal with, labelled NFS authorization, state modification authorization, and the uses of ACLs outside authorization.¶

Beyond the reorganization described above, there are numerous matters to be dealt with, reflecting the previous inability to clearly specify two sets of authorization behaviors and their potential interactions.¶

• Also in Section 8.2, there is added clarity in the discussion of support for multiple authorization approaches by eliminating use of the subjective term "reasonable semantics".¶

  In connection with this clarification, we have switched from describing the needed co-ordination between modes and acls as "goals" to describing them as "requirements" to give clients some basis for expecting interoperability in handling these issues.¶

  As a result of this shift to a prescriptive framework applying to all minor versions it becomes possible to treat all minor versions together. In existing RFCs and some earlier versions of this document, it had been assumed that NFSv4.0 was free to ignore the relevant prescriptions first put forth in RFC 5661 and only needed to address the "goals" for this co-ordination.¶

A.3. Addressing the Need for Clarifying Changes

Unlike the issues discussed in Appendix A.1, for which the path to make the necessary changes is laid out in that same section, the changes discussed in Appendix A.2, are discussed here, because resolving them is a more involved matter.¶

The difficulty derives principally from the fact that there are existing clients and server with very different approaches to these matters, making it difficult or impossible to reach a consensus that will make some existing clients non-compliant.¶

In such a situation, we need to allow a set of possibitlities, as has been done for internationalization of file names in [RFC7530] for NFSv4.0 and in [I-D.ietf-nfsv4-internationalization], for all minor versions.¶

As in the case of the internationalization of file names, we will have to accommodate a range of possible server implementations, but the task will, in this case, involve a lot more work. To see why, we will first explore the situation for internationalization of file names and later see how the situation for these issues poses greater difficulties and how they might be addressed. Some of these dificulties match other internationalization issues, such as the handling of internationalized domain names and the proper treatment of case-insensitive handling of file names that dealt with incorrectly or not at all in [RFC7530].¶

• The treatment of internationalization of file names in [RFC7530] and [RFC5661] was so divorced from the needs of implementations that it was necessary to be ignore it.¶

  On the other hand, the treatment of internationalization of file names in the obsoleted document [RFC3010] was available as a base and had been implemented by many clients and servers.¶

• The vast majority of clients and servers implemented the same approach to internationalization, even though it was not limited to the use of UTF-8.¶

  Despite the general aversion to this approach within the IESG, it was accepted, most likely because we were describing NFSv4 internationalization on "as-implemented" basis.¶

• Overall, we wound up allowing a total of three sorts of server behaviors, two of which were UTF-8-aware.¶

  One UTF-8-aware option that had been implemented, simply considred canonically equivalent names as deignating the same file with changing the name to a different canonically equivalent name. Although this approach diverged from IESG expectations regarding normalization, it was accepted.¶

  Another UTF-8-aware option was allowed, despite the absence of any implementations. It allowed servers to normalize file names while forbidding them from rejecting unormalized names. This was allowed because it would cause no difficulties for clients and because forbidding it might have caused problems for IESG acceptance.¶

• We were able, where necessary, to give the client the ability to detrmine which form of internationalization supported provided by the server for any particular file system, using an OPTIONAL attribute made available in NFSv4.1.¶

The situation with regard to the clarifying changes needed in this document is different for the reasons listed below.¶

• Rather than three valid choices for acceptable server behavior, with authorization we have a vast number.¶

  Given that there are nine ACE mask bits, each of whick might not be supported, there are at least 512 possible valid behaviors, even if none of the dubious instances of SHOULD creates additional valid behavioral patterns.¶

  Added to that are multiple (at least two) ways of mapping ACLs to modes and at leat three different behaviors in mdifying ACLs to reflect changes in mode.¶

  As a result, existing RFCs allow at least three thousand different server behaviors.¶

• There is little information available about the particular behaviors manifested by existing server-fs combinations.¶

  It may be difficult to get this informarion because the implementations were done by people no longer participating in the working or who were never involved in the working group¶

• There are no existing attributes that might be used to communicates server/fs choices to the client.¶

This creates a difficult situation. To fully resolve it will require substantial work as follows:¶

• Determining whether each of the dubious instances of "SHOULD" is in fact relied on by existing servers or might be needed by future servers.¶

• Determining which ACE mask bits are supported by all existing servers or not supported by any existing server to further limit the set of OPTIONAL features that clients need to be made aware of.¶

  There might also be cases where some set of mask bits are are always either supported or unsupported together.¶

• Determining the set of mappings from acls to modes that are actually and making each such OPTIONAL as opposed to the current situation in which an "intentional" "SHOULD" is used despite the mismatch between this use and [RFC2119].¶

• Making a similar determination of the actions actually applied to acl when the mode is changed.¶

Once we have the above information, it will be clear how we could create adress these issues inn either of the ways discused below:¶

• Have a way of determining whether support of NFSv4 ACL semaniics or UNIX ACL semantics is provided by any particular file system, such as that discussed in consensus item #61.¶

  This approach wiuld address the motive for the current under-specfication but would not address possible negative coonsequences arising from that under-specification.¶

  A further difficulty with it is that it does not address the possibility of hybrids of the two models.¶

• providing an appropriate extension as described in Section 6.2 to provide clients with an adequate description of expected server behavior. This document will be limited to providing client implementation guidance about dealing with the uncertinty in cases in which new attributes is not available. The situation will be similar to that described in Section 15 of [I-D.ietf-nfsv4-internationalization].¶