--- 1/draft-ietf-nfsv4-acl-mapping-04.txt 2006-08-23 22:12:26.000000000 +0200 +++ 2/draft-ietf-nfsv4-acl-mapping-05.txt 2006-08-23 22:12:26.000000000 +0200 @@ -1,18 +1,18 @@ Network Working Group M. Eriksen Internet-Draft J. Fields -Expires: November 16, 2006 CITI - May 15, 2006 +Expires: February 23, 2007 CITI + August 22, 2006 Mapping Between NFSv4 and Posix Draft ACLs - draft-ietf-nfsv4-acl-mapping-04 + draft-ietf-nfsv4-acl-mapping-05 Status of this Memo By submitting this Internet-Draft, each author represents that any applicable patent or other IPR claims of which he or she is aware have been or will be disclosed, and any of which he or she becomes aware will be disclosed, in accordance with Section 6 of BCP 79. Internet-Drafts are working documents of the Internet Engineering Task Force (IETF), its areas, and its working groups. Note that @@ -23,130 +23,138 @@ and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress." The list of current Internet-Drafts can be accessed at http://www.ietf.org/ietf/1id-abstracts.txt. The list of Internet-Draft Shadow Directories can be accessed at http://www.ietf.org/shadow.html. - This Internet-Draft will expire on November 16, 2006. + This Internet-Draft will expire on February 23, 2007. Copyright Notice Copyright (C) The Internet Society (2006). Abstract A number of filesystems and applications support ACLs based on a withdrawn POSIX draft [2]. Those ACLs differ significantly from NFS - version 4 (NFSv4) ACLs [1]. We describe how to translate between the - two types of ACLs. + version 4 ACLs [1]. We describe how to translate between the two + types of ACLs. 1. Introduction - Access Control Lists (ACLs) are used to specify fine-grained access - rights to file system objects. An ACL is a list of Access Control - Entries (ACEs), each specifying an entity (such as a user) and some - level of access for that entity. + Access Control Lists (ACLs) are used to specify access rights to file + system objects. An ACL is a list of Access Control Entries (ACEs), + each specifying an entity (such as a user) and some level of access + for that entity. - In the following sections we describe two ACL models: NFSv4 ACLs, and - ACLs based on a withdrawn POSIX draft. We will refer to the latter - as "POSIX ACLs". Since NFSv4 ACLs are more fine-grained than POSIX - ACLs, it is not possible in general to map an arbitrary NFSv4 ACL to - a POSIX ACL with the same semantics. However, it is possible to map - any POSIX ACL to a NFSv4 ACL with nearly identical semantics, and it - is possible to map any NFSv4 ACL to a POSIX ACL in a way that - preserves certain guarantees. We will explain how to do this, and - give guidelines for clients and servers performing such translation. + In the following sections describe NFSv4 ACLs and ACLs based on a + withdrawn POSIX draft. We will refer to the latter as "POSIX ACLs". + Since NFSv4 ACLs are more fine-grained than POSIX ACLs, it is not + possible in general to map an arbitrary NFSv4 ACL to a POSIX ACL with + the same semantics. However, it is possible to map any POSIX ACL to + a NFSv4 ACL with nearly identical semantics, and it is possible to + map any NFSv4 ACL to a POSIX ACL in a way that preserves certain + guarantees. We will explain how to do this, and give guidelines for + clients and servers performing such translation. 2. NFSv4 ACLs An NFSv4 ACL is an ordered sequence of ACEs, each having an entity, a type, some flags, and an access mask. The entity may be the name of a user or group, or may be one of a small set of special entities. Among the special entities are "OWNER@" (the current owner of the file), "GROUP@" (the group associated with the file), and "EVERYONE@". - The type may be ALLOW or DENY. (AUDIT or ALARM are also allowed, but - they are not relevant to our discussion). + An ACL may have a "type" of ALLOW or DENY. (AUDIT or ALARM are also + allowed, but they are not relevant to our discussion). The access mask has 14 separate bits, including bits to control read, write, execute, append, ACL modification, file owner modification, etc.; consult [1] for the full list. Of the flags, four are relevant here. The ACE4_IDENTIFIER_GROUP flag is used to indicate that the entity name is the name of a group. The other three concern inheritance: ACE4_DIRECTORY_INHERIT_ACE indicates that the ACE should be added to new subdirectories of the directory; ACE4_FILE_INHERIT_ACE does the same for new files; and ACE4_INHERIT_ONLY indicates that the ACE should be ignored when determining access to the directory itself. The NFSv4 ACL permission-checking algorithm is straightforward. - Assume a a requester asks for access, as specified by a single bit in - the access bitmask. We allow the access if the first ACE in the ACL - that matches the requester and that has that bit set is an ALLOW ACE, - and we deny the access if the first such ACE is a DENY ACE. If no - matching ACE has the bit in question set, behaviour is undefined. If - an access mask consisting of more than one bit is requested, it - succeeds if and only if each bit in the mask is allowed. + First, assume a requester asks for access specified by a single bit + in the access bitmask. We allow the access if the first ACE in the + ACL that matches the requester and that has that bit set is an ALLOW + ACE, and we deny the access if the first such ACE is a DENY ACE. If + no matching ACE has the bit in question set, access is normally + denied. + + If a requester asks for access requiring multiple bits from the + access bitmask simutaneously, then we allow the access if and only if + each bit in the requested bitmask would be allowed individually. We refer the reader to [1] for further details. 3. POSIX ACLs A number of operating systems implement ACLs based on the withdrawn POSIX 1003.1e/1003.2c Draft Standard 17 [2]. We will refer to such - ACLs as "POSIX ACLs". + ACLs as "POSIX ACLs", though they are not part of any published POSIX + standard. POSIX ACLs use access masks with only the traditional "read", "write", and "execute" bits. Each ACE in a POSIX ACL is one of five types: ACL_USER_OBJ, ACL_USER, ACL_GROUP_OBJ, ACL_GROUP, ACL_MASK, and ACL_OTHER. Each ACL_USER ACE has a uid associated with it, and each ACL_GROUP ACE has a gid associated with it. Every POSIX ACL must have exactly one ACL_USER_OBJ, ACL_GROUP_OBJ, and ACL_OTHER ACE, and at most one ACL_MASK ACE. The ACL_MASK ACE is required if the ACL has any ACL_USER or ACL_GROUP ACEs. There may not be two ACL_USER ACEs with the same uid, and there may not be two ACL_GROUP ACEs with the same gid. Given a POSIX ACL and a requester asking for access, permission is - determined as follows: + determined by consulting the ACEs in the order ACL_USER_OBJ, + ACL_USER, ACL_GROUP_OBJ, ACL_GROUP, ACL_OTHER, and allowing or + denying access based on the first ACE encountered that the requester + matches, except that we never allow the ACL_USER, ACL_OWNER_OBJ, or + ACL_GROUP objects to grant more than the ACL_MASK object does, and in + the case of ACL_GROUP_OBJ and ACL_GROUP ACEs, we allow access if any + one of those ACEs allows access. + + In more detail: 1. If the requester is the file owner, then allow or deny access depending on whether the ACL_USER_OBJ ACE allows or denies it. Otherwise, - 2. if the requester's uid matches the uid of one of the ACL_USER + 2. if the requester matches the file's group, and the ACL mask ACE + would deny the requested access, then skip to step 5. Otherwise, + + 3. if the requester's uid matches the uid of one of the ACL_USER ACEs, then allow or deny access depending on whether the ACL_USER_OBJ ACE allows or denies it. Otherwise, - 3. Consider the set of all ACL_GROUP ACEs whose gid the requester is + 4. Consider the set of all ACL_GROUP ACEs whose gid the requester is a member of. Add to that set the ACL_GROUP_OBJ ACE, if the requester is also a member of the file's group. Allow access if any ACE in the resulting set allows access. If the set of matching ACEs is nonempty, and none allow access, then deny access. Otherwise, if the set of matching ACEs is empty, - - 4. if the requester's access mask is allowed by the ACL_OTHER ACE, + 5. if the requester's access mask is allowed by the ACL_OTHER ACE, then grant access. Otherwise, deny access. - The above description omits one detail: in steps (2) and (3), the - requested bits must be granted both by the matching ACE and by the - ACL_MASK ACE. The ACL_MASK ACE thus limits the maximum permissions - which may be granted by any ACL_USER or ACL_GROUP ACE, or by the - ACL_GROUP_OBJ ACE. - Each file may have a single POSIX ACL associated with it, used to determine access to that file. Directories, however, may have two ACLs: one, the "access ACL", used to determine access to the directory, and one, the "default ACL", used only as the ACL to be inherited by newly created objects in the directory. 4. Ordering of NFSv4 and POSIX ACLs POSIX ACLs are unordered--the order in which the POSIX access- checking algorithm considers the entries is determined entirely by @@ -185,64 +193,62 @@ The NFSv4 ACL permission-checking algorithm has the property that it permits a group of bits whenever it would permit each bit individually, so it is impossible to mimic this behaviour with an NFSv4 ACL. 6. Mapping POSIX ACLs to NFSv4 ACLs 6.1. Requirements In the next section we give an example of a mapping of POSIX ACLs - into NFSv4 ACLs. We permit a server or client to use a different - mapping, provided the mapping meets the following requirements: + into NFSv4 ACLs. A server or client may use a different mapping, but + the mapping should meet the following requirements: It must map the POSIX ACL to an NFSv4 ACL with identical access semantics, ignoring the minor exception described in the previous section. It must map the read mode bit to ACE4_READ_DATA, the write bit to ACE4_WRITE_DATA and ACE4_APPEND_DATA (and ACE4_DELETE_CHILD for directories), and the EXECUTE bit to ACE4_EXECUTE. It should also allow ACE4_READ_ACL, ACE4_READ_ATTRIBUTES, and ACE4_SYNCHRONIZE unconditionally, and allow ACE4_WRITE_ACL and ACE4_WRITE_ATTRIBUTES to the owner. The handling of other NFSv4 mode bits may depend on the implementation, but it is preferable to leave them unused. It should avoid using DENY ACEs. If DENY ACEs are required, it should attempt to place them at the beginning. (This is not always possible.) - For simplicity's sake, the translator may choose to handle the mask - by first applying it to the USER, GROUP, and GROUP_OBJ ACEs, and then - mapping the resulting ACL. However, that will result in an ACL from - which it is impossible to determine the original value of the mask or - of the masked USER, GROUP, and GROUP_OBJ bitmasks. If the resulting - ACL is later translated back to a POSIX ACL, the translator will - assume that the value of the mask is the union of the bitmasks - permitted to any USER, GROUP, or GROUP_OBJ. If that would be - incorrect, the original translation should not modify the bitmasks of - the USER, GROUP, and GROUP_OBJ bitmasks, and should instead use - additional DENY ACEs as necessary to give the effect of the mask. It - should also arrange for the first GROUP@ ACE to be a DENY ACE whose - bitmask is determined by the mask, allowing that ACE to be used to - determine the original mask value. + The resulting NFSv4 ACL must take into account the mask ACE, by + ensuring that it does not give the group file owner or any users or + groups named in the ACL more permissions than permitted by the mask. + It would also be possible to specify a mapping that encoded the mask + in such a way that the original value of the mask could be recovered + by someone that knew the ACL was produced by our algorithm. However, + the added complexity and fragility of such a mapping is not worth the + small benefit of preserving the mask information, so we do not + attempt that here. 6.2. Example POSIX->NFSv4 Mapping We now describe an algorithm which maps any POSIX ACL to an NFSv4 ACL with the same semantics, meeting the above requirements. - First, translate the uid's and gid's on the ACL_USER and ACL_GROUP - ACEs into NFSv4 names, using directory services, etc., as - appropriate, and translate ACL_USER_OBJ, ACL_GROUP_OBJ, and ACL_OTHER - to the special NFSv4 names "OWNER@", "GROUP@", and "EVERYONE@", - respectively. + First, modify all ACL_USER, ACL_GROUP, and ACL_GROUP_OBJ ACEs by + removing any permissions not granted by the mask ACE. The mask ACE + may then be ignored for the rest of this process. + + Translate the uid's and gid's on the ACL_USER and ACL_GROUP ACEs into + NFSv4 names, using directory services, etc., as appropriate, and + translate ACL_USER_OBJ, ACL_GROUP_OBJ, and ACL_OTHER to the special + NFSv4 names "OWNER@", "GROUP@", and "EVERYONE@", respectively. Next, map each POSIX ACE (excepting any mask ACE) in the given POSIX ACL to an NFSv4 ALLOW ACE with an entity determined as above, and with a bitmask determined from the permission bits on the POSIX ACE as follows: 1. If the read bit is set in the POSIX ACE, then set ACE4_READ_DATA. 2. If the write bit is set in the POSIX ACE, then set ACE4_WRITE_DATA and ACE4_APPEND_DATA. If the object carrying the @@ -252,277 +258,371 @@ ACE4_EXECUTE. 4. Set ACE4_READ_ACL, ACE4_READ_ATTRIBUTES, and ACE4_SYNCHRONIZE unconditionally. 5. If the ACE is for the special "OWNER@" entity, set ACE4_WRITE_ACL and ACE4_WRITE_ATTRIBUTES. 6. Clear all other bits in the NFSv4 bitmask. - In addition, we set the GROUP flag in each ACE which corresponds to a + In addition, set the GROUP flag in each ACE which corresponds to a named group (but not in the GROUP@ ACE, or any of the other special entity ACEs). At this point, we've replaced the POSIX ACL by an NFSv4 ACL with the same number of ACEs (ignoring any mask ACE), all of them ALLOW ACEs. Order this NFSv4 ACL in the canonical order: OWNER@, users, GROUP@, groups, then EVERYONE@. If the bitmasks in the resulting ACEs are non-increasing (so no ACE allows a bit not allowed by a previous ACE), then we can skip the next step. Otherwise, we need to insert additional DENY ACE's to emulate the first-match semantics of the POSIX ACL permission-checking algorithm: 1. If an ACL_USER_OBJ, ACL_OTHER, or ACL_USER ACE fails to grant some permissions that are granted later in the ACL, then that ACE - must be prepended by a single DENY ACE. The DENY ACE should have + must be preceded by a single DENY ACE. The DENY ACE should have the same entity and flags as the corresponding ALLOW ACE, but the bitmask on the DENY ACE should be the bitwise NOT of the bitmask on the ALLOW ACE, except that the ACE4_WRITE_OWNER, ACE4_DELETE, - ACE4_READ_NAMED_ATTRIBUTES, ACE4_WRITE_NAMED_ATTRIBUTES bits + ACE4_READ_NAMED_ATTRIBUTES, and ACE4_WRITE_NAMED_ATTRIBUTES bits should be cleared, and the ACE4_DELETE_CHILD bit should be cleared on non-directories. (Also, in the xdr-encoded ACL that is transmitted, all bits not defined in the protocol should be cleared.) 2. All of the ACL_GROUP_OBJ and ACL_GROUP ACEs are consulted by the POSIX algorithm before determining permissions. To emulate this behaviour, instead of adding a single DENY before corresponding GROUP@ or named group ACEs, we insert a list of DENY ACEs after the list of GROUP@ and named group ACEs. Each DENY ACE is determined from its corresponding ALLOW ACE exactly as in the - previous step. As before, these DENY aces should only be added + previous step. As before, these DENY ACEs should only be added when they are necessitated by an ACE that is less permissive than - the final EVERYONE@ ace. - - Finally, we enforce the POSIX mask ACE by prepending each ALLOW ACE - for a named user, GROUP@, or named group, with a single DENY ACE - whose entity and flags are the same as those for the corresponding - ALLOW ACE, but whose bitmask is the inverse of the bitmask determined - from the mask ACE, with the inverse calculated as described above. - In the case of named users, these DENY aces may be coalesced with any - existing prepended DENY aces. The DENY aces are omitted entirely if - they would have no affect, or if the mask ACE has the same bitmask as - the maximum of the affected ACEs. (With the one exception that if - the POSIX ACL posesses exactly 4 ACEs, then a mask-derived DENY ace - should be inserted before the GROUP@ ace, even if it would not - otherwise be.) - - Regardless of what scheme is used to represent the mask, the receiver - will use the first GROUP@ DENY ace to determine the value of the mask - (if it is different from the union of the bitmasks on the affected - ACEs), and use the relevant ALLOWs to determine the pre-mask values - of user and group ACEs. - - The implementation may also choose to just mask out the bitmasks on - the relevant ALLOW ACEs. This will produce a simpler ACL (in - particular, an ACL that usually requires no DENY ACE's), at the - expense of losing some ACL information after a chmod. + the final EVERYONE@ ACE. On directories with default ACLs, we translate the default ACL as above, but set the ACE4_INHERIT_ONLY_ACE, ACE4_DIRECTORY_INHERIT_ACE, and ACE4_FILE_INHERIT_ACE flags on every ACE in the resulting ACL. On directories with both default and access ACLs, we translate the two ACLs and then concatenate them. The order of the concatenation is unimportant. 7. Mapping NFSv4 ACLs to POSIX ACLs 7.1. Requirements Any mapping of NFSv4 ACLs to POSIX ACLs must map any NFSv4 ACL that is semantically equivalent to a POSIX ACL (with the exception of the - "minor inaccuracy" mentioned above) to the equivalent POSIX ACL. It - should also extract the mask correctly; as the mask doesn't affect - the semantics of the NFSv4 ACL, and as there is more than one way the - mask might be encoded, we require a convention for this. - Specifically: we require that the mask be computed as the bitmask - used on the first GROUP@ DENY ACE which precedes any GROUP@ allow - ACE, unless no such DENY ACE exists, in which case the mask must be - computed as the union of the bitmasks allowed to all named users, - groups, and GROUP@ (where by the "bitmask allowed to" an entity we - mean the maximum bitmask that the ACL would permit to any user - matching the entity). + "minor inaccuracy" mentioned above) to an equivalent POSIX ACL. - Implementations may vary in how they deal with NFSv4 ACLs that are - not precisely semantically equivalent to any POSIX ACL. In - particular they may return errors for such ACLs instead of attempting - to map them. However, when possible without compromising security, - they should attempt to be forgiving. + However, a more difficult problem is presented by NFSv4 ACLs that are + not precisely equivalant to any POSIX ACL. + + The only way that the NFSv4 protocol gives servers to indicate that + they support only a subset of the ACL model is the "aclsupport" + attribute, which allows a server to advertise that it only supports + certain ACE types. This allows a server to report that it only + supports ALLOW ACEs, or that it does not support AUDIT or ALARM ACEs + (which will be the case for most servers with only POSIX ACLs). But + it does not give a way to claim support for more complex subsets of + the ACL model. + + While it is possible for a server to reject any ACLs that do not fit + its ACL model, this places a large burden on clients and users, since + the server has no way to explain why it rejected a particular ACL. + Therefore, it is preferable to be more forgiving, whenever that is + possible without compromising security, and to limit any restrictions + to those that are easily documented and verified by users. The language of [1] allows a server some flexibility in handling ACLs that it cannot enforce completely accurately, as long as it adheres to "the guiding principle... that the server must not accept ACLs that appear to make [a file] more secure than it really is." - Note that an NFSv4 ACL consisting entirely of ALLOW ACLs can always - be transformed into a POSIX-equivalent ACL by first sorting it into - the canonical order, and then inserting DENY ACEs as necessary to - ensure POSIX first-match semantics. Since inserting DENY ACEs can - only restrict access, it is safe for a server to do this. + ACLs with arbitrary sequences of ALLOWs and DENYs may be particularly + troublesome; but note that an NFSv4 ACL consisting entirely of ALLOW + ACLs can always be transformed into a POSIX-equivalent ACL by first + sorting it into the canonical order, then inserting DENY ACEs as + necessary to ensure POSIX first-match semantics. Since inserting + DENY ACEs can only restrict access, it is safe for a server to do + this. - We require any server to accept, at least, any NFSv4 ACL that + Therefore servers should accept, at least, any NFSv4 ACL that consists entirely of ALLOW ACLs. Clients should also be at least as forgiving, to promote interoperability when heterogeneous clients share files. 7.2. Example NFSv4->POSIX Mapping We now give an example of an algorithm that meets the above requirements. We assume it is to be used by a server mapping client- provided NFSv4 ACLs to POSIX ACLs it can store in its filesystem, so - the translation errs on the side of making the ACL less permissive. + the translation errs on the side of making the ACL more restrictive. - Given an NFSv4 ACL, first calculate the mask by taking the bitmask - from the first GROUP@ DENY ACE from the original NFSv4 ACL, if it - exists. After doing so, remove that DENY ACE, and clear the bits in - its bitmask from any DENY ACE for a named user, group, or GROUP@ - which precedes an ALLOW ACE for the same entity. + In fact, if we ignore some loss of information in the mask ACE, this + mapping takes an NFSv4 ACL to the unique most permissive POSIX ACL + that is no more permissive than the given NFSv4 ACL. - In the case where there is no such GROUP@ DENY ACE, continue through - the rest of the algorithm and then calculate the mask as the union of - the calculated permissions of all named users, group, and the - GROUP_OBJ ACE. + Before starting, if the ACL in question is for a directory, we split + it into two ACLs, one purely effective and one purely inherited, as + follows: - Given an NFSv4 ACL, sort it into canonical order (OWNER@ ACEs first, - then user ACEs, then GROUP@ ACEs, then group ACEs, then EVERYONE@ - ACEs.) Also, sort the GROUP@ and group ACEs that all ALLOW ACEs - precede all DENY ACEs. To do so, take advantage of the following - observations: + 1. ACEs with no inheritance flags are put in the purely effective + ACL. - 1. If two consecutive ACEs are either both ALLOW ACEs, or both DENY - ACEs, then we can swap their order without changing the effect of - the ACL. + 2. Aces with FILE_INHERIT and DIRECTORY_INHERIT both set are put in + both the effective and the inherited ACL - 2. If it would be impossible for a single user to match both of the - entities on two consecutive ACEs, then we can swap their order - without changing the effect of the ACL. + 3. Aces with FILE_INHERIT, DIRECTORY_INHERIT, and INHERIT_ONLY all + set are put only in the inherited ACL. - 3. If an ALLOW ACE is immediately followed by a DENY ACE, then - swapping the order of the two ACEs will not make the ACL any more - permissive. + Other combinations of ineritance flags may be rejected or silently + modified to one of the above. - 4. If a DENY ACE is immediately followed by an ALLOW ACE, then - swapping the order of the two ACEs will not make the ACL any more - permissive, *if* we modify the bitmask on the ALLOW ACE by - clearing any bits that are set in the DENY ACE. + The main algorithm that follows is then performed on each ACL, with + one used to set the effictive ACL, and one the default ACL. - The second observation is the trickiest: it may usually be safe to - assume that two distinct user names cannot match the same user. An - implementation with knowledge about group memberships or about the - current value of the file owner might also use that information, but - if it does so it will produce a translation that is no longer - accurate after owners or group memberships change. + First, we calculate the OTHER mode as follows: - Fortunately, observations 1, 3, and 4 are sufficient to sort any ACL - into canonical order, so a paranoid implementation can simply ignore - number 2 completely, while an implementation willing to sacrifice - some accuracy may choose to do something more complex. + 1. Initialize the bitmasks other_allow and other_deny both to zero. - Ensure that the resulting ACL posesses at least one each of OWNER@, - GROUP@, and EVERYONE@ ACEs, by inserting an ALLOW ACE with a zero - bitmask if necessary in the correct position. + 2. For each ACE in the ACL, starting from the top: - Next, for each entity, calculate a bitmask for that entity as - follows: Starting with the first ACE for that entity (ignoring all - previous ACEs), perform the NFSv4 ACL-checking algorithm for a user - that is assumed to match the entity on every DENY ACE that a user - matching the given entity might match, but is assumed to match only - those entities on ALLOW ACEs that *any* user matching the current - entity must match. + 1. If the ACE is not an EVERYONE@ ACE, ignore it and move to the + next ACE. - Finally, construct the POSIX ACL by translating NFSv4 entity names to - uid's and gid's (and handling special entities in the obvious way), - then assign a POSIX bitmask determined by the NFSv4 bitmask - calculated in the previous step; the bitmask calculation should use - the inverse of the mapping described previously in the POSIX-to-NFSv4 - mapping, erring on the side of denying bits if it cannot determine a - sensible mapping. However, if certain bits simply cannot be mapped - in a reasonable way to mode bits, the server may simply ignore them - rather than returning an error. (For example, the server should deny - write if either ACE4_WRITE_DATA or ACE4_APPEND_DATA are denied. But - it may choose to ignore ACE4_READ_ATTRIBUTES entirely.) + 2. If the ACE is an EVERYONE@ ALLOW ACE, then add to other_allow + any bits set in this ACE but not set in other_deny. + + 3. If the ACE is an EVERYONE@ DENY ACE, then add to other_deny + any bits set in this ACE but not set in other_allow. + + 3. Discard other_deny. Set the USER_OBJ mask from other_allow using + the inverse of the mapping described previously in the POSIX-to- + NFSv4 mapping, erring on the side of denying bits if it cannot + determine a sensible mapping. However, if certain bits simply + cannot be mapped in a reasonable way to mode bits, the server may + simply ignore them rather than returning an error. (For example, + the server should deny write if either ACE4_WRITE_DATA or + ACE4_APPEND_DATA are denied. But it may choose to ignore + ACE4_READ_ATTRIBUTES entirely; though in that case it may at + least want to treat specially the case where such bits are + explicitly denied by some DENY ACE.) + + Note that the bits determined above are exactly the maximum bits that + will always be permitted to a user that doesn't match the file owner + or group, or any of the named owners or groups. Thus this choice of + the OTHER mode is exactly the maximum choice we can safely make. + + Next we calculate the GROUP_OBJ and GROUP masks. + + 1. Initialize to zero an allow and deny bitmask for each GROUP_OBJ + and for each GROUP mask. + + 2. For each ACE in the ACL, starting from the top: + + 1. If the ACE is an OWNER@ or named user ACE, ignore it and move + to the next ACE. + + 2. If the ACE is an EVERYONE@ ALLOW ACE, then, for each GROUP or + GROUP_OBJ allow mask, set the bits allowed in the EVERYONE + ACE but not already in this GROUP or GROUP_OBJ's deny mask. + + 3. If the ACE is an EVERYONE@ DENY ACE, then, for each GROUP or + GROUP_OBJ deny mask, set the bits denied in the EVERYONE ACE + but not already allowed in this GROUP or GROUP_OBJ's deny + mask. + + 4. If the ACE is a GROUP or GROUP@ ALLOW ACE, then set the allow + bits in the corresponding GROUP or GROUP_OBJ allow mask that + are allowed by this ACE but not already denied by the + corresponding GROUP or GROUP_OBJ deny mask. + + 5. If the ACE is a GROUP or GROUP@ DENY ACE, then set the deny + bits in the corresponding GROUP or GROUP_OBJ deny mask that + are denied by this ACE but not already allowed by the + corresponding GROUP or GROUP_OBJ allow mask. Call the + resulting deny mask "m". In each GROUP or GROUP_OBJ deny + mask, set every bit that is in m and not already in that + GROUP or GROUP_OBJ allow mask. + + 3. Having calculated allow and deny masks for GROUP_OBJ and each + GROUP, we now set the corresponding modes from the allow masks as + we did in the last step of the USER_OBJ mask calculation above. + + Note that the bits thus determined for a group are exactly the + maximum bits that will always be permitted to a user that matches the + group in question, and that is denied any bits that could be denied + by matching other groups, without out being allowed bits by matching + any such groups. This is the most permissive mode we can choose that + will never permit more permissions than the original NFSv4 ACL, for + any possible choice of group memberships. + + An implementation with special knowledge about the current gowning + group or about group memberships may choose to use that knowledge to + calculate a more permissive mode. However, doing so may render + resulting POSIX ACL inaccurate after the owning group changes, or + after any group memberships change. + + Next, we calculate USER modes by first calculating allow and deny + masks for each USER as above, this time assuming we are a user that + does not match the file owner, that matches no user except for the + one user under consideration, and that matches groups only when they + would deny some permissions that they have not allowed yet. (To + ensure this last step it will also be necessary to maintain group + allow and deny ACEs, as we did in the previous calculation.) We omit + the detailed steps, which are similar. Again, the implementation may + choose to use special knowledge about group memberships at the risk + of increased complexity and of loss of some accuracy. + + Next, we calculate the USER_OBJ mode by calculating allow and deny + masks for a user that matches the file owner and any user or group + that denies bits that it does not first allow. + + Finally, if the resulting ACL has any named user or group ACEs, add a + mask ACE with bitmask equal to the union of the calculated + permissions of all named users, group, and the GROUP_OBJ ACE. The resulting mapping errs on the side of creating a more restrictive ACE. However it can be modified to produce a mapping that errs on the side of permissiveness, for the purposes of translating a server- provided NFSv4 ACL to a POSIX ACL to present to a user or application, as follows: - 1. When sorting ACEs, ALLOW ACEs can always be moved towards the - start of the ACL, but a DENY ACE can be moved towards the start - of the ACL only as long as we clear any of the DENY ACE's bitmask - bits that are set in the intervening ALLOW ACEs. + 1. When performing the final mapping from the allow bitmask to a + mode, we instead using a mapping that errs on the side of + permissiveness; for example, we allow write permissions even if + only one of WRITE_DATA, APPEND_DATA, or (in the case of + directories) DELETE_CHILD is allowed. - 2. When calculating the NFSv4 bitmask for each entity, err on the - side of assuming that ALLOW ACEs apply and that DENY ACEs don't, - with the one exception that when calculating the GROUP@ and named - group bitmasks, ALLOW ACEs for groups other than the one under - consideration should be ignored. + 2. Wherever in the above we pessimistically assume that a user will + match any entity that has permissions denied to it before they + are first allowed, we instead assume that the user will match any + entity that has permissions allowed to it before they are first + denied. - 3. When mapping the NFSv4 bitmask to POSIX mode bits, err on the - side of allowing access. + Once again, the resulting mapping may be seen to produce the unique + (up to choice of mask) POSIX ACL which is the most restrictive among + all POSIX ACLs no more restrictive than the given NFSv4 ACL. -8. Security Considerations + Note that the above algorithms may be optimized in a number of ways: + for example, although they are described in terms of multiple passes, + it will be simpler and more efficient to calculate the entire POSIX + ACL in a single pass. + +8. Backwards Compatibility + + Previous versions of this document recommended a different + POSIX->NFSv4 mapping, which enforces POSIX semantics by inserting + DENYs into the ACL even when those DENY's would have no effect, and + which represents the POSIX mask ACE using additional DENYs. The + resulting ACLs are overly complex and create problems for Windows + clients, because the default Windows ACL editor prefers to order + DENYs before ALLOWs. + + The NFSv4 to POSIX mapping we describe in this document can accept + the NFSv4 ACLs produced by the old mapping. + + However, previous versions of this document also recommended + accepting only NFSv4 ACLs that were precisely those produced by the + old POSIX->NFSv4 mapping; therefore, existing implementations of that + recommendation will reject the NFSv4 ACLs produced by the newer + mapping. + + We strongly recommend fixing implementations to accept a wider range + of NFSv4 ACLs. However, we briefly document the old mapping here in + case that is impossible: + + Names, bitmasks, and flags are determined as in the the current + mapping. + + Whenever the following instructions requiring taking "the complement" + of an NFSv4 bitmask, do so as follows: first, take the bitwise NOT of + the bitmask. Then clear the ACE4_WRITE_OWNER, ACE4_DELETE, + ACE4_READ_NAMED_ATTRIBUTES, and ACE4_WRITE_NAMED_ATTRIBUTES bits. + Also, clear the ACE4_DELETE_CHILD bit on non-directories, and clear + any bits not defined in the protocol. + + Create one ALLOW ACE for each entity (OWNER@, GROUP@, and EVERYONE@, + and each user and group named in the given POSIX ACL). After each + OWNER@, EVERYONE@, and named user ACE, append a DENY ACE with the + same entity and flags as the corresponding ALLOW ACE, but with + bitmask set to the complement (as defined above) of the ALLOW ACE. + + Do the same for each GROUP@ and named group ACE, but instead of + inserting each new DENY ACE after the corresponding ALLOW ACE, insert + all of the DENY ACEs at the end of the list of GROUP@ and named group + ACEs, in the same order that the GROUP@ and named group ALLOW ACEs + occur in. + + Finally, prepend each GROUP@, named user, and named group ACE by a + single DENY whose entity and flags are the same as the corresponding + ALLOW, but whose bitmask is the complement (as defined above) of the + bitmask determined from the mask ACE in the given POSIX ACL. Skip + this step if the given POSIX ACL has no mask ACE. + +9. Security Considerations Any automatic mapping from one ACL model to another must provide guarantees as to how the mapping affects the meaning of ACLs, or risk misleading users about the permissions set on filesystem objects. For this reason, caution is recommended when implementing this mapping. It is better to return errors than to break any such guarantees. That said, there may be cases where small losses in accuracy can avoid dramatic interoperability and usability problems; as long as the losses in accuracy are clearly documented, these tradeoffs may be found acceptable. For example, a server unable to support all of the NFSv4 mode bits does not have a way to communicate its exact limitations to clients, so clients (and users) may be unable to recover from such errors. For this reason we recommend ignoring bitmask bits that the server is - completely unable to map to mode bits, and advertising this fact - loudly in the server documentation. If this is considered - insufficient, we should add to the NFSv4 protocol additional - attributes necessary to advertise the server's limitations. + completely unable to map to mode bits, at least when no ACE + explicitly contradicts the server's default behavior. If this is + considered insufficient, we should add to the NFSv4 protocol + additional attributes necessary to advertise the server's + limitations. - Note also that this ACL mapping requires mapping between NFSv4 + Note also that any ACL mapping also requires mapping between NFSv4 usernames and local id's. When the mapping of id's depends on remote services, the method used for the mapping must be at least as secure as the method used to set or get ACLs. -9. References +10. References [1] Shepler, S., Callaghan, B., Robinson, D., Thurlow, R., Beame, C., Eisler, M., and D. Noveck, "Network File System (NFS) version 4 Protocol", RFC 3530, April 2003. [2] Institute of Electrical and Electronics Engineers, Inc., "IEEE Draft P1003.1e", October 1997, . Authors' Addresses Marius Aamodt Eriksen U. of Michigan Center for Information Technology Integration Email: marius@citi.umich.edu J. Bruce Fields U. of Michigan Center for Information Technology Integration - Email: marius@citi.umich.edu + Email: bfields@citi.umich.edu Intellectual Property Statement The IETF takes no position regarding the validity or scope of any Intellectual Property Rights or other rights that might be claimed to pertain to the implementation or use of the technology described in this document or the extent to which any license under such rights might or might not be available; nor does it represent that it has made any independent effort to identify any such rights. Information on the procedures with respect to rights in RFC documents can be