Copyright Andrew Bartlett 2005-2009 Copyright Donald T. Davis 2009 Released under the GPLv3 "Porting Samba4 to MIT-Krb" From Idmwiki IPA v3 will use a version of Samba4 built on top of MIT's Kerberos implementation, instead of Heimdal's version of Kerberos. Task list summary for porting changes needed, from Andrew Bartlett: * Rewrite or extend the LDAP driver that MIT-KDC will use. * MIT KDC changes: rewrite DAL, add TGS-KBAC, enable PACs,... * Full thread-safety for MIT's library code, * Many small changes Task list, without explanations (the list with explanations is in the later sections of this document): Porting Samba4 to MIT-krb comprises four main chunks of work: 1. Rewrite or extend the LDAP driver that MIT-KDC will use: a. Our LDAP driver for the KDB needs to know how to do Samba4's intricate canonicalization of server names, user-names, and realm names. b. AD-style aliases for HOST/ service names. c. Implicit names for Win2k accounts. d. Principal "types": client / server / krbtgs e. Most or all of this code is in 3 source files, ~1000 lines in all; 2. MIT KDC changes a. Rewrite the MIT KDC's Data-Abstraction Layer (DAL), mostly because he MIT KDC needs to see& manipulate more LDAP detail, on Samba4's behalf; b. Add HBAC to the KDC's TGT-issuance, so that Samba4 can refuse TGTs to kinit, based on time-of-day& IP-addr constraints; c. turn on MIT-krb 1.7's PAC handling d. add bad-password counts, for unified account-lockouts across all authT methods (Krb, NTLM, LDAP simple bind, etc) 3. Make sure MIT's library code is more fully thread-safe, by replacing all global and static variables with context parameters for the library routines. This may already be done. 4. Many small changes (~15) a. some extensions to MIT's libkrb5& GSSAPI libraries, including GSSAPI ticket-forwarding b. some refitting in Samba4's use of the MIT libraries; c. make sure Samba4's portable socket API works, including "packet too large" errors; d. MIT's GSSAPI code should support some legacy Samba3 clients that present incorrectly-calculated checksums; e. Samba4 app-server-host holds aUTF-16 PW, plus a key bitstring; f. in-memory-only credentials cache; g. in-memory-only keytab (nice to have); h. get OSS NTLM authT library (Likewise Software?); i. special Heimdal-specific functions; j. principal-manipulation functions; k. special check for misconfigured Samba4 hostnames; l. improved krb error-messages; m. improved krb logging n. MS GSSMonger test-suite o. testsuite for kpasswd daemon 0. Introduction: This document should be read alongside the Samba4 source code, as follows: * For DAL and KDC requirements, please see Samba4's source4/kdc/hdb-samba4.c in particular. This file is an implementation against Heimdal's HDB abstraction layer, and is the biggest part of the samba-to-krb glue layer, so the main part of the port to MIT is to replace hdb-samba4 with a similar glue layer that's designed for MIT's code. * Samba4's PAC requirements are implemented in source4/kdc/pac-glue.c * Both of the above two layers are Heimdal plugins, and both get loaded in source4/kdc/kdc.c * For GSSAPI requirements, see auth/gensec/gensec_gssapi.c (the consumer of GSSAPI in Samba4) * For Kerberos library requirements, see auth/kerberos/krb5_init_context.c * Samba has its own credentials system, wrapping GSS creds, just as GSS creds wrap around krb5 creds. For the interaction between Samba4 credential system and GSSAPI and Kerberos, see auth/credentials/credentials_krb5. 1. Rewrite or extend the LDAP driver that MIT-KDC will use. a. IPA'sLDAP driver for the KDB needs to know how to do Samba4's intricate canonicalization of server names, user-names, and realm names. For hostnames& usernames, alternate names appear in LDAP as extra values in the multivalued "principal name" attributes: * For a hostname, the alternate names (other than the short name, implied from the CN), are stored in the servicePrincipalName * For a username, the alternate names are stored in the userPrincipalName attribute, and can be long email-address-like names, such as joe@microsoft.com (see "Type 10 names," below). GSSAPI layer requirements: Welcome to the wonderful world of canonicalisation. The MIT Krb5 libs (including GSSAPI) do not enable the AS to send kinit a TGT containing a different realm-name than what the client asked for, even in U/L case differences. Heimdal has the same problem, and this applies to the krb5 layer too, not just GSSAPI. There are two kinds of name-canonicalization that can occur on Windows: * Lower-to-upper case conversion, because Windows domain names are usually in upper case; * An unrecognizable substitution of names, such as might happen when a user requests a ticket for a NetBIOS domain name, but gets back a ticket for the corresponding FQDN. As developers, we should test if the AD KDC's name-canonical- isation can be turned off with the KDCOption flags in the AS-REQ or TGS-REQ; Windows clients always send the Canonicalize flags as KDCOption values. Principal Names, long and short names: AD's KDC does not canonicalize servicePrincipalNames, except for the realm in the KDC reply. That is, the client gets back the principal it asked for, with the realm portion 'fixed' to uppercase, long form. Samba4 does some canonicalization, though Heimdal doesn't canonicalize names itself: For hostnames and usernames, Samba4 canonicalizes the requested name only for the LDAP principal-lookup, but then Samba4 returns the retrieved LDAP record with the request's original, uncanonicalized hostname replacing the canonicalized name that actually was found. Usernames: AndrewB says that Samba4 used to return the canonicalized username exactly as retrieved from LDAP. The reason Samba4 treated usernames differently was that the user needs to present his own canonicalized username to servers, for ACL-matching. For hostnames this isn't necessary. Realm-names: AD seems to accept a realm's short name in krb-requests, at least for AS_REQ operations, but the AD KDC always performs realm-canonicalisation, which converts the short realm-name to the canonical long form. So, this causes pain for current krb client libraries. Punchline: For bug-compatibility, we may need to selectively or optionally disable the MIT-KDC's name- canonicalization. Application-code: Name-canonicalisation matters not only for the KDC, but also for app-server-code that has to deal with keytabs. Further, with credential-caches, canonicalization can lead to cache-misses, but then the client just asks for new credentials for the variant server-name. This could happen, for example, if the user asks to access the server twice, using different variants of the server-name. Doubled realm-names: We also need to handle type 10 names (NT-ENTERPRISE), which are a full principal name in the principal field, unrelated to the realm. The principal field contains both principal& realm names, while the realm field contains a realm name, too, possibly different. For example, an NT-ENTERPRISE principal name might look like: joeblow@microsoft.com@NTDEV.MICROSOFT.COM , <--principal field-->|<----realm name--->| Where joe@microsoft.com is the leading portion, and NTDEV.MICROSOFT.COM is the realm. This is used for the 'email address-like login-name' feature of AD. b.AD-style aliases for HOST/ service names. AD keeps a list of service-prefixed aliases for the host's principal name. The AD KDC reads& parses this list, so as to allow the aliased services to share the HOST/ key. This means that every ticket-request for a service-alias gets a service-ticket encrypted in the HOST/ key. For example, this is how HTTP/ and CIFS/ can use the HOST/ AD-LDAP entry, without any explicitly CIFS-prefixed entry in the host's servicePrincipalName attribute. In the app-server host's AD record, the servicePrincipalName says only HOST/my.computer@MY.REALM , but the client asks for CIFS/my.omputer@MY.REALM tickets. So, AD looks in LDAP for both name-variants, and finds the HOST/ version, In AD's reply, AD replaces the HOST/ prefix with CIFS/ . We implement this in hdb-ldb. (TBD: Andrew, is this correct?:) List of HOST/ aliases: Samba4 currently uses only a small set of HOST/ aliases: sPNMappings: host=ldap,dns,cifs,http . Also, dns's presence in this list is a bug, somehow. AD's real list has 53 entries: sPNMappings: host=alerter,appmgmt,cisvc,clipsrv,browser, dhcp,dnscache,replicator,eventlog,eventsystem,policyagent, oakley,dmserver,dns,mcsvc,fax,msiserver,ias,messenger, netlogon,netman,netdde,netddedsm,nmagent,plugplay, protectedstorage,rasman,rpclocator,rpc,rpcss,remoteaccess, rsvp,samss,scardsvr,scesrv,seclogon,scm,dcom,cifs,spooler, snmp,schedule,tapisrv,trksvr,trkwks,ups,time,wins,www, http,w3svc,iisadmin,msdtc Domain members that expect the longer list will break in Samba4, as of 6/09. AB says he'll try to fix this right away. There is another post somewhere (ref lost for the moment) that details where in active directory the long list of stored aliases for HOST/ is. c.Implicit names for Win2000 Accounts: AD keys its server-records by CN or by servicePrincipalName, but a win2k box's server-entry in LDAP doesn't include the servicePrincipalName attribute, So, win2k server-accounts are keyed by the CN attribute instead. Because AD's LDAP doesn't have a servicePrincipalName for win2k servers' entries, Samba4 has to have an implicit mapping from host/computer.full.name and from host/computer, to the computer's CN-keyed entry in the AD LDAP database, so to be able to find the win2k server's host name in the KDB. d.Principal "types": We have modified Heimdal's 'hdb' interface to specify the 'class' of Principal being requested. This allows us to correctly behave with the different 'classes' of Principal name. This is necessary because of AD's LDAP structure, which uses very different record-structures for user-principals, trust principals& server-principals. We currently define 3 classes: * client (kinit) * server (tgt) * krbtgt the TGS's own ldap record Samba4 also now specifies the kerberos principal as an explicit parameter to LDB_fetch(), not an in/out value on the struct hdb_entry parameter itself. e. Most or all of this LDAP driver code is in three source files, ~1000 lines in all. These files are in samba4/kdc : * hdb-samba4.c (samba4-to-kdb glue-layer plugin) * pac-glue.c (samba4's pac glue-layer plugin) * kdc.c (loads the above two plugins). 2. MIT KDC changes a.Data-Abstraction Layer (DAL): It would be good to rewrite or circumvent the MIT KDC's DAL, mostly because the MIT KDC needs to see& manipulate more LDAP detail, on Samba4's behalf. AB says the MIT DAL may serve well- enough, though, mostly as is. AB says Samba4 will need the private pointer part of the KDC plugin API, though, or the PAC generation won't work (see sec.2.c, below). * MIT's DAL calls lack context parameters (as of 2006), so presumably they rely instead on global storage, and aren't fully thread-safe. * In Novell's pure DAL approach, the DAL only read in the principalName as the key, so it had trouble performing access-control decisions on things other than the user's name (like the addresses). * Here's why Samba4 needs more entry detail than the DAL provides: The AS needs to have ACL rules that will allow a TGT to a user only when the user logs in from the right desktop addresses, and at the right times of day. This coarse-granularity access-control could be enforced directly by the KDC's LDAP driver, without Samba having to see the entry's pertinent authZ attributes. But, there's a notable exception: a user whose TGT has expired, and who wants to change his password, should be allowed a restricted-use TGT that gives him access to the kpasswd service. This ACL-logic could be buried in the LDAP driver, in the same way as the TGS ACL could be enforced down there, but to do so would just be even uglier than it was to put the TGS's ACL-logic in the driver. * Yet another complaint is that the DAL always pulls an entire LDAP entry, non-selectively. The current DAL is OK for Samba4's purposes, because Samba4 only reads, and doesn't write, the KDB. But this all-or-nothing retrieval hurts the KDC's performance, and would do so even more, if Samba had to use the DAL to change KDB entries. b.Add HBAC to the KDC's TGT-issuance, so that Samba4 can refuse TGTs to kinit, based on time-of-day& IP-address constraints. AB asks, "Is a DAL the layer we need?" Looking at what we need to pass around, AB doesn't think the DAL is the right layer; what we really want instead is to create an account-authorization abstraction layer (e.g., is this account permitted to login to this computer, at this time?). Samba4 ended up doing account-authorization inside Heimdal, via a specialized KDC plugin. For a summary description of this plugin API, see Appendix 2. c. Turn on MIT-krb 1.7'sPAC handling. In addition, I have added a new interface hdb_fetch_ex(), which returns a structure including a private data-pointer, which may be used by the windc plugin interface functions. The windc plugin provides the hook for the PAC. d. Samba4 needsaccess control hooks in the Heimdal& MIT KDCs. We need to lockout accounts (eg, after 10 failed PW- attempts), and perform other controls. This is standard AD behavior, that Samba4 needs to get right, whether Heimdal or MIT-krb is doing the ticket work. - If PADL doesn't publish their patch for this, we'll need to write our own. - The windc plugin proivides a function for the main access control routines. A new windc plugin function should be added to increment the bad password counter on failure. - Samba4 doesn't yet handle bad password counts (or good password notification), so that a single policy can be applied against all means of checking a password (NTLM, Kerberos, LDAP Simple Bind, etc). Novell's original DAL did not provide a way to update the PW counts information. - Nevertheless, we know that this is very much required in AD, because Samba3 + eDirectory goes to great lengths to update this information. This may have been addressed in Simo's subsequent IPA-KDC design), * AllowedWorkstationNames and Krb5: Microsoft uses the clientAddresses *multiple value* field in the krb5 protocol (particularly the AS_REQ) to communicate the client's netbios name (legacy undotted name,<14 chars) AB guesses that this is to support the userWorkstations field (in user's AD record). The idea is to support client-address restrictions, as was standard in NT: The AD authentication server probably checks the netbios address against this userWorkstations value (BTW, the NetLogon server does this, too). 3. State Machine safety when using Kerberos and GSSAPI libraries * Samba's client-side& app-server-side libraries are built on a giant state machine, and as such have very different requirements to those traditionally expressed for kerberos and GSSAPI libraries. * Samba requires all of the libraries it uses to be "state machine safe" in their use of internal data. This does not necessarily mean "thread safe," and an application could be thread safe, but not state machine safe (if it instead used thread-local variables). so, if MIT's libraries were made thread-safe only by inserting spinlock() code, then the MIT libraries aren't yet "state machine safe." * So, what does it mean for a library to be state machine safe? This is mostly a question of context, and how the library manages whatever internal state machines it has. If the library uses a context variable, passed in by the caller, which contains all the information about the current state of the library, then it is safe. An example of this state is the sequence number and session keys for an ongoing encrypted session). * The other issue affecting state machines is 'blocking' (waiting for a read on a network socket). Samba's non-blocking I/O doesn't like waiting for libkrb5 to go away for awhile to talk to the KDC. * Samba4 provides a hook 'send_to_kdc', that allows Samba4 to take over the IO handling, and run other events in the meantime. This uses a 'nested event context' (which presents the challenges that the kerberos library might be called again, while still in the send_to_kdc hook). * Heimdal has this 'state machine safety' in parts, and we have modified Samba4's lorikeet branch to improve this behaviour, when using a new, non-standard API to tunnelling a ccache (containing a set of tickets) through the gssapi, by temporarily casting the ccache pointer to a gss credential pointer. This new API is Heimdal's samba4-requested gss_krb5_import_cred() fcn; this will have to be rewritten or ported in the MIT port. * This tunnelling trick replaces an older scheme using the KRB5_CCACHE environment variable to get the same job done. The tunnelling trick enables a command-line app-client to run kinit tacitly, before running GSSAPI for service-authentication. The tunnelling trick avoids the more usual approach of keeping the ccache pointer in a global variable. * [Heimdal uses a per-context variable for the 'krb5_auth_context', which controls the ongoing encrypted connection, but does use global variables for the ubiquitous krb5_context parameter. (No longer true, because the krb5_context global is gone now.)] * The modification that has added most to 'state machine safety' of GSSAPI is the addition of the gss_krb5_acquire_creds() function. This allows the caller to specify a keytab and ccache, for use by the GSSAPI code. Therefore there is no need to use global variables to communicate this information about keytab& ccache. * At a more theoretical level (simply counting static and global variables) Heimdal is not state machine safe for the GSSAPI layer. (But Heimdal is now (6/09) much more nearly free of globals.) The Krb5 layer alone is much closer, as far as I can tell, blocking excepted. . * As an alternate to fixing MIT Kerberos for better safety in this area, a new design might be implemented in Samba, where blocking read/write is made to the KDC in another (fork()ed) child process, and the results passed back to the parent process for use in other non-blocking operations. * To deal with blocking, we could have a fork()ed child per context, using the 'GSSAPI export context' function to transfer the GSSAPI state back into the main code for the wrap()/unwrap() part of the operation. This will still hit issues of static storage (one gss_krb5_context per process, and multiple GSSAPI encrypted sessions at a time) but these may not matter in practice. * This approach has long been controversial in the Samba team. An alternate way would be to be implement E_AGAIN in libkrb5: similar to the way to way read() works with incomplete operations. to do this in libkrb5 would be difficult, but valuable. * In the short-term, we deal with blocking by taking over the network send() and recv() functions, therefore making them 'semi-async'. This doesn't apply to DNS yet.These thread-safety context-variables will probably present porting problems, during the MIT port. This will probably be most of the work in the port to MIT. This may require more thorough thread-safe-ing work on the MIT libraries. 4. Many small changes (~15) a. Some extensions to MIT'slibkrb5& GSSAPI libraries, including GSSAPI ticket-forwarding: This is a general list of the other extensions Samba4 has made to / need from the kerberos libraries * DCE_STYLE : Microsoft's hard-coded 3-msg Challenge/Response handshake emulates DCE's preference for C/R. Microsoft calls this DCE_STYLE. MIT already has this nowadays (6/09). * gsskrb5_get_initiator_subkey() (return the exact key that Samba3 has always asked for. gsskrb5_get_subkey() might do what we need anyway). This routine is necessary, because in some spots, Microsoft uses raw Kerberos keys, outside the Kerberos protocols, as a direct input to MD5 and ARCFOUR, without using the make_priv() or make_safe() calls, and without GSSAPI wrappings etc. * gsskrb5_acquire_creds() (takes keytab and/or ccache as input parameters, see keytab and state machine discussion in prev section) * The new function to handle the PAC fully gsskrb5_extract_authz_data_from_sec_context() need to test that MIT's PAC-handling code checks the PAC's signature. * gsskrb5_wrap_size (Samba still needs this one, for finding out how big the wrapped packet will be, given input length). b. Some refitting in Samba4's use of the MIT libraries; c. Make sure Samba4'sportable socket API works: * An important detail in the use of libkdc is that we use samba4's own socket lib. This allows the KDC code to be as portable as the rest of samba, but more importantly it ensures consistency in the handling of requests, binding to sockets etc. * To handle TCP, we use of our socket layer in much the same way as we deal with TCP for CIFS. Tridge created a generic packet handling layer for this. * For the client, samba4 likewise must take over the socket functions, so that our single thread smbd will not lock up talking to itself. (We allow processing while waiting for packets in our socket routines). send_to_kdc() presents to its caller the samba-style socket interface, but the MIT port will reimplement send_to_kdc(), and this routine will use internally the same socket library that MIT-krb uses. * The interface we have defined for libkdc allows for packet injection into the post-socket layer, with a defined krb5_context and kdb5_kdc_configuration structure. These effectively redirect the kerberos warnings, logging and database calls as we require. * Samba4 socket-library's current TCP support does not send back 'too large' error messages if the high bit is set. This is needed for a proposed extension mechanism (SSL-armored kinit, by Leif Johansson), but is currently unsupported in both Heimdal and MIT. d. MIT's GSSAPI code should support some legacy Samba3 clients that presentincorrectly-calculated checksums. * Old Clients (samba3 and HPUX clients) use 'selfmade' gssapi/krb5 tokens for use in the CIFS session setup. These hand-crafted ASN.1 packets don't follow rfc1964 (GSSAPI) perfectly, so server-side krblib code has to be flexible enough to accept these bent tokens. * It turns out that Windows' GSSAPI server-side code is sloppy about checking some GSSAPI tokens' checksums. During initial work to implement an AD client, it was easier to make an acceptable solution (acceptable to Windows servers) than to correctly implement the GSSAPI specification, particularly on top of the (inflexible) MIT Kerberos API. It did not seem possible to write a correct, separate GSSAPI implementation on top of MIT Kerberos's public krb5lib API, and at the time, the effort did not need to extend beyond what Windows would require. * The upshot is that old Samba3 clients send GSSAPI tokens bearing incorrect checksums, which AD's GSSAPI library cheerfully accepts (but accepts the good checksums, too). Similarly, Samba4's Heimdal krb5lib accepts these incorrect checksums. Accordingly, if MIT's krb5lib wants to interoperate with the old Samba3 clients, then MIT's library will have to do the same. * Because these old clients use krb5_mk_req() the app-servers get a chksum field depending on the encryption type, but that's wrong for GSSAPI (see rfc 1964 section 1.1.1). The Checksum type 8003 should be used in the Authenticator of the AP-REQ! That (correct use of the 8003 type) would allow the channel bindings, the GCC_C_* req_flags and optional delegation tickets to be passed from the client to the server. However windows doesn't seem to care whether the checksum is of the wrong type, and for CIFS SessionSetups, it seems that the req_flags are just set to 0. This deviant checksum can't work for LDAP connections with sign or seal, or for any DCERPC connection, because those connections do not require the negotiation of GSS-Wrap paraemters (signing or sealing of whole payloads). Note: CIFS has an independent SMB signing mechanism, using the Kerberos key. * For the code that handles the incorrect& correct checksums, see heimdal/lib/gssapi/krb5/accept_sec_context.c, lines 390-450 or so. * This bug-compatibility is likely to be controversial in the kerberos community, but a similar need for bug- compatibility arose around MIT's& Heimdal's both failing to support TGS_SUBKEYs correctly, and there are numerous other cases. seehttps://lists.anl.gov/pipermail/ietf-krb-wg/2009-May/007630.html * So, MIT's krb5lib needs to also support old clients! e. Samba4 app-server-host holds aUTF-16 PW, plus a key bitstring; See Appendix 1, "Keytab Requirements." f.In-memory-only credentials cache for forwarded tickets Samba4 extracts forwarded tickets from the GSSAPI layer, and puts them into the memory-based credentials cache. We can then use them for proxy work. This needs to be ported, if the MIT library doesn't do it yet. g.In-memory-only keytab (nice to have): Heimdal used to offer "in-memory keytabs" for servers that use passwords. These server-side passwords were held in a Samba LDB database called secrets.ldb . The heimdal library would fetch the server's password from the ldb file and would construct an in-memory keytab struct containing the password, somewhat as if the library had read an MIT-style keytab file. Unfortunately, only later, at recv_auth() time, would the Heimdal library convert the server-PW into a salted-&-hashed AES key, by hashing 10,000 times with SHA-1. Naturally, this is really too slow for recv_auth(), which runs when an app-server authenticates a client's app-service- request. So, nowadays, this password-based in-memory keytab is falling into disuse. h. Get OSSNTLM authT library: AB says Likewise software probably will give us their freeware "NTLM for MIT-krb" implementation. i. Special Heimdal-specific functions; These functions didn't exist in the MIT code, years ago, when Samba started. AB will try to build a final list of these functions: * krb5_free_keyblock_contents() * j.Principal-manipulation functions: Samba makes extensive use of the principal manipulation functions in Heimdal, including the known structure behind krb_principal and krb5_realm (a char *). For example, * krb5_parse_name_flags(smb_krb5_context->krb5_context, name, KRB5_PRINCIPAL_PARSE_REQUIRE_REALM,&principal); * krb5_unparse_name_flags(smb_krb5_context->krb5_context, principal, KRB5_PRINCIPAL_UNPARSE_NO_REALM,&new_princ); * krb5_principal_get_realm() * krb5_principal_set_realm() These are needed for juggling the AD variant-structures for server names. k. SpecialShort name rules check for misconfigured Samba4 hostnames; Samba is highly likely to be misconfigured, in many weird and interesting ways. So, we have a patch for Heimdal that avoids DNS lookups on names without a "." in them. This should avoid some delay and root server load. (This errors need to be caught in MIT's library.) l.Improved krb error-messages; krb5_get_error_string(): This Heimdal-specific function does a lot to reduce the 'administrator pain' level, by providing specific, English text-string error messages instead of just error code translations. (This isn't necessary for the port, but it's more useful than MIT's default err-handling; Make sure this works for MIT-krb) m.Improved Kerberos logging support: krb5_log_facility(): Samba4 now uses this Heimdal function, which allows us to redirect the warnings and status from the KDC (and client/server Kerberos code) to Samba's DEBUG() system. Samba uses this logging routine optionally in the main code, but it's required for KDC errors. n. MSGSSMonger test-suite: Microsoft has released a krb-specific testsuite called gssmonger, which tests interoperability. We should compile it against lorikeet-heimdal& MIT and see if we can build a 'Samba4' server for it. GSSMonger wasn't intended to be Windows-specific. o.Testsuite for kpasswd daemon: I have a partial kpasswd server which needs finishing, and a Samba4 needs a client testsuite written, either via the krb5 API or directly against GENSEC and the ASN.1 routines. Samba4 likes to test failure-modes, not just successful behavior. Currently Samba4's kpasswd only works for Heimdal, not MIT clients. This may be due to call-ordering constraints. Appendix 1: Keytab Requirements Traditional 'MIT' keytab operation is very different from AD's account-handling for application-servers: a. Host PWs vs service-keys: * Traditional 'MIT' behaviour is for the app-server to use a keytab containing several named random-bitstring service-keys, created by the KDC. An MIT-style keytab holds a different service-key for every kerberized application-service that the server offers to clients. Heimdal also implements this behaviour. MIT's model doesn't use AD's UTF-16 'service password', and no salting is necessary for service-keys, because each service-key is random enough to withstand an exhaustive key-search attack. * In the Windows model, the server key's construction is very different: The app-server itself, not the KDC, generates a random UTF-16 pseudo-textual password, and sends this password to the KDC using SAMR, a DCE-RPC "domain-joining" protocol (but for windows 7, see below). Then, the KDC shares this server- password with every application service on the whole machine. * Only when the app-server uses kerberos does the password get salted by the member server (ie, an AD server-host). (That is, no salt information appears to be conveyed from the AD KDC to the member server, and the member server must use the rules described in Luke's mail, in Appendix 3, below). The salted- and-hashed version of the server-host's PW gets stored in the server-host's keytab. * Samba file-servers can have many server-names simultaneously (kind of like web servers' software-virtual-hosting), but since these servers are running in AD, these names can be set up to all share the same secret key. In AD, co-located server names almost always share a secret key like this. In samba3, this key-sharing was optional, so some samba3 hosts' keytabs did hold multiple keys. Samba4 abandons this traditional "old MIT" style of keytab, and only supports one key per keytab, and multiple server-names can use that keytab key in common. In dealing with this model, Samba4 uses both the traditional file keytab and an in-MEMORY keytabs. * Pre-Windows7 AD and samba3/4 both use SAMR, an older protocol, to jumpstart the member server's PW-sharing with AD (the "windows domain-join process"). This PW-sharing transfers only the PW's UTF-16 text, without any salting or hashing, so that non-krb security mechanisms can use the same utf-16 text PW. For Windows 7, this domain-joining uses LDAP for PW-setting. b. Flexible server-naming * The other big difference between AD's keytabs and MIT's is that Windows offers a lot more flexibility about service-principals' names. When the kerberos server-side library receives Windows-style tickets from an app-client, MIT's krb library (or GSSAPI) must accommodate Windows' flexibility about case-sensitivity and canonicalization. This means that an incoming application-request to a member server may use a wide variety of service-principal names. These include: machine$@REALM (samba clients) HOST/foo.bar@realm (win2k clients) cifs/foo.bar@realm (winxp clients) HOST/foo@realm (win2k clients, using netbios) cifs/foo@realm (winxp clients, using netbios), as well as all upper/lower-case variations on the above. c. Keytabs& Name-canonicalization * Heimdal's GSSAPI expects to to be called with a principal-name& a keytab, possibly containing multiple principals' different keys. However, AD has a different problem to solve, which is that the client may know the member- server by a non-canonicalized principal name, yet AD knows the keytab contains exactly one key, indexed by the canonical name. So, GSSAPI is unprepared to canonicalize the server-name that the cliet requested, and is also overprepared to do an unnecessary search through the keytab by principal-name. So Samba's server-side GSSAPI calls have to "game" the GSSAPI, by supplying the server's known canonical name, with the one-key keytab. This doesn't really affect IPA's port of Samba4 to MIT-krb. * Because the number of U/L case combinations got 'too hard' to put into a keytab in the traditional way (with the client to specify the name), we either pre-compute the keys into a traditional keytab or make an in-MEMORY keytab at run time. In both cases we specify the principal name to GSSAPI, which avoids the need to store duplicate principals. * We use a 'private' keytab in our private dir, referenced from the secrets.ldb by default. Appendix 2: KDC Plugin for Account-Authorization Here is how Samba4 ended up doing account-authorization in Heimdal, via a specialized KDC plugin. This plugin helps bridge an important gap: The user's AD record is much richer than the Heimdal HDB format allows, so we do AD-specific access-control checks in the plugin's AD-specific layer, not in the DB-agnostic KDC server: * We created a separate KDC plugin, with this API: typedef struct hdb_entry_ex { void *ctx; hdb_entry entry; void (*free_entry)(krb5_context, struct hdb_entry_ex *); } hdb_entry_ex; The void *ctx is a "private pointer," provided by the 'get' method's hdb_entry_ex retval. The APIs below use the void *ctx so as to find additional information about the user, not contained in the hdb_entry structure. Both the provider and the APIs below understand how to cast the private void *ctx pointer. typedef krb5_error_code (*krb5plugin_windc_pac_generate)(void * krb5_context, struct hdb_entry_ex *, krb5_pac*); typedef krb5_error_code (*krb5plugin_windc_pac_verify)(void * krb5_context, const krb5_principal, struct hdb_entry_ex *, struct hdb_entry_ex *, krb5_pac *); typedef krb5_error_code (*krb5plugin_windc_client_access)(void * krb5_context, struct hdb_entry_ex *, KDC_REQ *, krb5_data *); The krb5_data* here is critical, so that samba's KDC can return the right NTSTATUS code in the 'error string' returned to the client. Otherwise, the windows client won't get the right error message to the user (such as 'password expired' etc). The pure Kerberos error is not enough) typedef struct krb5plugin_windc_ftable { int minor_version; krb5_error_code (*init)(krb5_context, void **); void (*fini)(void *); krb5plugin_windc_pac_generate pac_generate; krb5plugin_windc_pac_verify pac_verify; krb5plugin_windc_client_access client_access; } krb5plugin_windc_ftable; This API has some Heimdal-specific stuff, that'll have to change when we port this KDC plugin to MIT krb. * 1st callback (pac_generate) creates an initial PAC from the user's AD record. * 2nd callback (pac_verify) checks that a PAC is correctly signed, adds additional groups (for cross-realm tickets) and re-signs with the key of the target kerberos service's account * 3rd callback (client_access) performs additional access checks, such as allowedWorkstations and account expiry. * For example, to register this plugin, use the kdc's standard plugin-system at Samba4's initialisation: /* first, setup the table of callback pointers */ /* Registar WinDC hooks */ ret = krb5_plugin_register(krb5_context, PLUGIN_TYPE_DATA, "windc",&windc_plugin_table); /* once registered, the KDC will invoke the callbacks */ /* while preparing each new ticket (TGT or app-tkt) */ * An alternative way to register the plugin is with a config-file that names a DSO (Dynamically Shared Object). Appendix 3: Samba4 stuff that doesn't need to get ported. Heimdal oddities * Heimdal is built such that it should be able to serve multiple realms at the same time. This isn't relevant for Samba's use, but it shows up in a lot of generalisations throughout the code. * Samba4's code originally tried internally to make it possible to use Heimdal's multi-realms-per-KDC ability, but this was ill-conceived, and AB has recently (6/09) ripped the last of that multi-realms stuff out of samba4. AB says that in AD, it's not really possible to make this work; several AD components structurally assume that there's one realm per KDC. However, we do use this to support canonicalization of realm-names: case variations, plus long-vs-short variants of realm-names. No MIT porting task here, as long as MIT kdc doesn't refuse to do some LDAP lookups (eg, alias' realm-name looks wrong). * Heimdal supports multiple passwords on a client account: Samba4 seems to call hdb_next_enctype2key() in the pre-authentication routines, to allow multiple passwords per account in krb5. (I think this was intended to allow multiple salts). AD doesn't support this, so the MIT port shouldn't bother with this. Not needed anymore, because MIT's code now handles PACs fully: * gss_krb5_copy_service_keyblock() (get the key used to actually encrypt the ticket to the server, because the same key is used for the PAC validation). * gsskrb5_extract_authtime_from_sec_context (get authtime from kerberos ticket) * gsskrb5_extract_authz_data_from_sec_context (get authdata from ticket, ie the PAC. Must unwrap the data if in an AD-IFRELEVANT)] Authz data extraction * We use krb5_ticket_get_authorization_data_type(), and expect it to return the correct authz data, even if wrapped in an AD-IFRELEVANT container. This doesn't need to be ported to MIT. This should be obsoleted by MIT's new PAC code. libkdc * Samba4 needs to be built as a single binary (design requirement), and this should include the KDC. Samba also (and perhaps more importantly) needs to control the configuration environment of the KDC. * But, libkdc doesn't matter for IPA; Samba invokes the Heimdal kdc as a library call, but this is just a convenience, and the MIT port can do otherwise w/o trouble.) Returned Salt for PreAuthentication When the AD-KDC replies to pre-authentication, it returns the salt, which may be in the form of a principalName that is in no way connected with the current names. (ie, even if the userPrincipalName and samAccountName are renamed, the old salt is returned). This is the kerberos standard salt, kept in the 'Key'. The AD generation rules are found in a Mail from Luke Howard dated 10 Nov 2004. The MIT glue layer doesn't really need to care about these salt-handling details; the samba4 code& the LDAP backend will conspire to make sure that MIT's KDC gets correct salts. > > From: Luke Howard > Organization: PADL Software Pty Ltd > To: lukeh@padl.com > Date: Wed, 10 Nov 2004 13:31:21 +1100 > Cc: huaraz@moeller.plus.com, samba-technical@lists.samba.org > Subject: Re: Samba-3.0.7-1.3E Active Directory Issues > ------- > > Did some more testing, it appears the behaviour has another > explanation. It appears that the standard Kerberos password salt > algorithm is applied in Windows 2003, just that the source principal > name is different. > > Here is what I've been able to deduce from creating a bunch of > different accounts: > [SAM name in this mail means the AD attribute samAccountName . > E.g., jbob for a user and jbcomputer$ for a computer.] > > [UPN is the AD userPrincipalName attribute. For example, jbob@mydomain.com] > Type of account Principal for Salting > ======================================================================== > Computer Account host/.realm@REALM > User Account Without UPN@REALM > User Account With UPN@REALM > > Note that if the computer account's SAM account name does not include > the trailing '$', then the entire SAM account name is used as input to > the salting principal. Setting a UPN for a computer account has no > effect. > > It seems to me odd that the RHS of the UPN is not used in the salting > principal. For example, a user with UPN foo@mydomain.com in the realm > MYREALM.COM would have a salt of MYREALM.COMfoo. Perhaps this is to > allow a user's UPN suffix to be changed without changing the salt. And > perhaps using the UPN for salting signifies a move away SAM names and > their associated constraints. > > For more information on how UPNs relate to the Kerberos protocol, > see: > > http://www.ietf.org/proceedings/01dec/I-D/draft-ietf-krb-wg-kerberos-referrals-02.txt > > -- Luke