$Id$ OpenSSL hacked to add support for the RFC 3779 X.509 v3 extensions. Some notes on OpenSSL internals. O'Reilly "Network Security with OpenSSL" is a bit dated (four years old, corresponds roughly to OpenSSL 0.9.7), but still appears to be the closest thing there is to coherent documentation. Some updates and machine readable copies of examples are available at http://www.opensslbook.com/. In spite of its age, the book is useful as it gives a readable overview of bit and pieces of OpenSSL's internal programming environment which one would otherwise have to absorb from the code via osmosis. Chapter 10 is particularly useful, as are the sections on error handling and abstract I/O (the ERR and BIO packages, respectively) in chapter 4. OpenSSL's own doc is very patchy, although fairly extensive in places. Most of it eventually comes down to "Use the Source, Luke" with pointers on which bit of source serves as an interesting example. For x509v3 extentions, the place to start is doc/openssl.txt, which, oddly, turns out to be mostly about certificate extensions. It gives an overview of the mechanisms, in particular of the method routine interface for certificate extensions. For something like the RFC 3779 extensions, it's pretty clear that we need to use a "raw" extension (which, as far as I can tell, just means that the RFC 3779 stuff is complicated enough that the extension handler has to do a lot of work to deal with a complex ASN.1 structure that the rest of the code doesn't know much about). openssl.txt refers the reader to the certificate policy extension handler as an example of how to write a raw extension. The relevant file appears to be crypto/x509v3/v3_cpols.c This is indeed a complicated mess, but it's a surprisingly good tutorial, on about the tenth reading. Note that this depends on extension-specific typedefs defined in crypto/x509v3/x509v3.h. Aside from the method interface and routines described in openssl.txt, v3_cpols.c also includes a bunch of template junk for the ASN.1 encoder and decoder. According to README.ASN1 (q.v.), the ASN.1 code is in transition. I haven't attempted to understand the "old" ASN.1 code; the new stuff is template driven. README.ASN.1 mumbles about this a bit, but the real doc on the template stuff appears to be the definitions in crypto/asn1/asn1t.h. Given that and a reading knowledge of ASN.1, the template definitions in v3_cpols.c are fairly obviously a direct hand translation from ASN.1 into C. General note on global symbols in OpenSSL: always look at the header file for any global symbol you're using. In fact, it's probably best to do a global search (m-x tags-search if you're an emacs user) for all instances of a global symbol before attempting to use it, as there are a lot of things that one just has to know about how all the global stuff hangs together. There are header files full of magic definitions that one just has to one need to be extended. There are magic pre-sorted lists of handlers that one just needs to know about. Little or none of this is documented. Use the Source, Luke. In some cases -portions- of files are automatically generated by Perl scripts (eg, the per-type stack definitions in safestack.h). Ouch. If you see a large block of very repetitive stuff, check for comments indicating that it's automatically generated. Oh, and the indentation style is demented. Header files you definitely need to read if you're going to touch this stuff: crypto/x509v3/x509v3.h crypto/x509v3/ext_dat.h crypto/stack/safestack.h crypto/asn1/asn1t.h crypto/asn1/asn1.h crypto/objects/objects.h Automatically generated header files you'll need to skim, then go read the input files listed in the header comments and perhaps the generating Perl code: crypto/objects/obj_mac.h crypto/objects/obj_dat.h Nice example of "new" style ASN.1 (if it compiles, which the README suggets it won't, sigh): demos/asn1/ocsp.c Much of the code shows a heavy Perl influence, presumably dating back clear to Eric Young. Some of the internal data structure operators have names that only make sense to a Perl programmer. Where code is automatically generated, it's done by Perl scripts. The configuration language for the whole package is a Perl script. Assembly code is all wrapped up inside perl scripts in a moderately clever attempt at being able to write the assembly language only once and use it with various assemblers with nontrivially different syntax. Much of the documentation markup (including manual pages) for the C code is .pod. At this point I think I understand enough that I could write the RFC 3779 extension I/O routines, although no doubt some of it would be take a while since I would be learning my way through some of the library routines as I go. I have not yet written up a syntax description for how we would encode the RFC 3779 extensions in an openssl.conf file. That should be fairly straightforward, if a bit tedious. I have not yet figured out where to hook in the extra goop that RFC 3779 will need for verification. Making extensions critical is easy enough, but the validation stuff in RFC 3779 2.3 and 3.3 needs to go somewhere, I'm not quite sure where yet. To the extent that I understand how OpenSSL does such things (not very well, yet), the "-purpose" flags to the CLI's x509 and verify commands (q.v.) look like the right hooks. I haven't (yet) investigated the code in back of the "-purpose" mechanism. At the moment the consensus appears to be that we should be using OpenSSL 0.9.8b, in order to get SHA-256 support. My research to date has been on OpenSSL 0.9.8a, but I doubt that any of the stuff I've been looking at has changed enough to matter. This is a first cut at what the openssl.conf syntax might look like for the RFC 3779 certificate extensions. Syntax is admittedly wretched, because it has to work with the existing OpenSSL code. Within that restriction, I've attempted to make this look as much as practical like the existing OpenSSL support for "multi-valued" extensions. RFC 3779 ASN.1 provided for easy reference. Notes: * Ranges are denoted with a hyphen, prefix lengths with a slash. I could tag ranges differently from the atomic types, but this seemed easier for the user to understand. * The "@" syntax indicating indirection through a separate section is lifted from the stock OpenSSL multi-valued extension support. * I didn't attempt to guess which addresses are IPv4 and which are IPv6 from the syntax, since the long form needed tags anyway. * For the moment I'm assuming that the only addresses we really care about making (relatively) easy are unicast v4 and v6 addresses; the "afisafi" syntax is there for completeness, but I'm not currently worrying about making it easy to do, eg, multicast addresses. Adding support for other specific AFI/SAFI pairs should be straightforward. * I'm particularly looking for feedback about things I've made impossible (syntax not general enough) or common things I've made harder than necessary. ### IPAddressFamily ::= SEQUENCE { -- AFI & optional SAFI -- addressFamily OCTET STRING (SIZE (2..3)), ipAddressChoice IPAddressChoice } IPAddressChoice ::= CHOICE { inherit NULL, -- inherit from issuer -- addressesOrRanges SEQUENCE OF IPAddressOrRange } IPAddressOrRange ::= CHOICE { addressPrefix IPAddress, addressRange IPAddressRange } IPAddressRange ::= SEQUENCE { min IPAddress, max IPAddress } IPAddress ::= BIT STRING # An address extension all specified on one line rfc3779-addr = critical, ipv4:10.1.1.1/32, ipv4:10.2.0.0-10.3.255.255 # An address extension all specified on one line, with inheritance rfc3779-addr = critical, ipv4:inherit, ipv6:2002::/16 # An address extension using raw AFI/SAFI syntax. # # This doesn't really work, we can't do prefixes without knowing how long # a particular kind of address is supposed to be. #rfc3779-addr = critical, afisafi:000101:001122334455-aabbccddee # An address extension using SAFIs rfc3779-addr = critical, ipv4-safi:1:10.1.1.1/32, ipv6-safi:1:2002::/16 # Address extension using an indirect section rfc3779-addr = critical, @addr-section [addr-section] ipv4.0 = 10.0.0.1 ipv4.1 = 10.0.1.0/24 ipv4.2 = 10.2.0.0 - 10.3.255.255 ipv6.0 = 2002:1::/64 ipv6.1 = 2002:2:: - 2002:8::ffff:ffff:ffff:ffff:ffff ### ASIdentifiers ::= SEQUENCE { asnum [0] EXPLICIT ASIdentifierChoice OPTIONAL, rdi [1] EXPLICIT ASIdentifierChoice OPTIONAL} ASIdentifierChoice ::= CHOICE { inherit NULL, -- inherit from issuer -- asIdsOrRanges SEQUENCE OF ASIdOrRange } ASIdOrRange ::= CHOICE { id ASId, range ASRange } ASRange ::= SEQUENCE { min ASId, max ASId } ASId ::= INTEGER # ASID extension all specified on one line: rfc3779-asid = critical, as:44, rdi:33-45 # ASID extension on one line using inheritance rfc3779-asid = critical, as:55, rdi:inherit # ASID extension using an indirect section rfc3779-asid = critical, @asid-section [asid-section] as.0 = 44 as.1 = 55 - 77 rdi.0 = 33 Problem: how to do the top level ASN.1 template thingie for a SEQUENCE (not SEQUENCE OF)? Possible example: crypto/krb5/krb5_asn.c, KRB5_TKTBODY and KRB5_TICKET. I -think- the flag combination I should be using is ASN1_TFLG_EXPLICIT (context-specific explicit tags, which, according to mtr and my vague recollection of ASN.1 tagging, is what RFC 3779 uses). No, that's not right, kerberos tags the SEQUENCE itself, which we don't. So either leave the flags field empty or use ASN1_TFLG_IMPLICIT. But this is all very spooky, I can't find a good example, and it seems wrong somehow that I'm ending up with template that points to a sequence instead of being a sequence. Hmm, maybe that means I need to call lower-level macros that do specific parts of the sequence constructor, then build a template with the SEQUENCE flags? Well, try what we have and see if it works, then figure out why not. Ok, one more try: crypto/x509v3/v3_{pci,pcia}.c implement RFC 3820 and show a SEQUENCE used as the top level template. Try it that way. Hmm, maybe openssl.txt really meant v3_{pci,pcia}.c as an example? Oooh, crypto/stack/ includes stack sorting functions. With the right structure definition and comparision function, this might solve a whole mess of our canonicalization problems. General plan: 1) Define an internal stack type for each of the extensions, with fields organized to allow us to handle everything with a single sort operation. Eg, major key is afi/safi, expand prefixes to ranges or some such, etc. 2) Set an appropiate comparison function (see sk_set_cmp_func()) 3) Parse config data into internal stack. 4) Call sk_sort(). 5) Loop calling sk_shift() (or sk_pop()) to pull items off the sorted stack so we can insert them into our real structure in the right canonical form. Hmm. What about other ways of building up the real structure? Would it be better to have a "canoicalize structure" function for each extension type? What about allowing a caller to say "no thanks it's already in canonical form please leave it alone"? util/mkstack.pl finds DECLARE_STACK_OF() declarations and generates safestack definitions automatically. Be afraid. Be very very afraid. Upon closer examination of RFC 3779, it's not possible to do a completely general handler for unknown AFIs, because it's the AFI that tells one that maximum length of an address (which in turn is required for certain operations, eg, canonization). So I guess we end up with: a) IPv4:foo, IPv4-SAFI:foo, IPv6:foo, IPv6-SAFI:foo (or something like that -- perhaps encode SAFI names too, eg IPv4-Unicast:foo), and b) A table of known AFIs telling us the length of an address for that AFI. For the moment I'm probably not going to bother filling that in for anything but IPv4 and IPv6, but having the mechanism in place for other AFIs seems harmless. Current plan for addr_canonize(): sort function for outer sequence should be straightforward. One could in theory write a sort function for the inner sequence, but doing so would require expanding addresses both in the sort function and in the subsequent cleanup loop in addr_canonize(). It would probably be simpler (and faster) to use a temporary stack containing pointers to the IPAddressOrRange elements and expanded versions of the addresses. Hmm, the alternative would be to write a comparision routine which expands a pair of bitstrings as addresses, and write the rest of the code in terms of that. Might be a useful function to have anyway. int addr_cmp(int afi, ASN1_BIT_STRING *a, unsigned char fill_a, ASN1_BIT_STRING *b, unsigned char fill_b); Problem is, this requires data not available in a sk_sort() handler. So I think we're stuck with the temporary stack approach. No, we can use a closure. addr_canonize() code is not right yet, mostly because merging prefixes and ranges is tricky. Logic that checks whether they need to be merged is probably ok, it's the code that does the merge that's whacked. Perhaps what we need is a function that takes two prefix-or-ranges (which the caller asserts should be merged), and generates the prefix or range that starts with the first one and ends with the second. Would probably be easier than what the code does now. Worry about optimizing this later, if it matters, get it right first. Current code in addr_canonize() is much more complicated than necessary, needs reorganization. All we really need to do is: 1) Expand begin and end addresses of A and B. 2) Check whether they overlap (or are adjacent). 3) If so, replace them with a merged prefix or range. Doh. A lot of the missing documentation is buried in ssleay.txt, which the other documentation says not to read because it's so old. But it's where Eric explains all the basic data structures and expected usage as of the dawn of time, so most of the stuff that's so old that it's undocumented is really documented there. Sigh. Anyway, it looks like the xxx_new() functions already set pointers of sub structures to NULL or allocate the substructures (have to check which for the things we care about, I guess), and the xxx_free() functions already clean up complex structures. So be sure to set unused pointers to NULL if one has been fiddling. foo->type fields may be initialized to -1 too, need to check that. So the destructor routines I wrote are unnecessary, and some of the memset()s may be actively harmful. Need to clean that up. Also need to check whether we need to call sk_xxx_pop_free() manually during cleanup. Seems unlikely, the template driven code has to be able to handle that one would think. Make sure that memory leak detection (MDEBUG in very old ssleay, figure out what it is now) is turned on. "make update" in the top level runs all the magic perl code that grovels through the code generating error codes, safe stacks, etc. My initial test configuration: ./Configure debug -DDEBUG_SAFESTACK This seems to be attempting to pull in devel/ElectricFence (-lefence), so I installed that. [Later...] Ok, the default debugging options are driving me nuts. ElectricFence is very slow, although no doubt I might want it anyway at some point. But it's the BN debugging printouts that finally pushed me into adding a "debug-sra" config of my own for the options I want. The identifiers we need are already in crypto/objects/objects.txt: id-pe 7 : sbgp-ipAddrBlock id-pe 8 : sbgp-autonomousSysNum Now I just have to figure out how to use them. :) crypto/objects/objects.h appears to be hand-maintained, not entirely surprising given that it appears to be the master NID assignment table. Oh good, if USE_OBJ_MAC is defined objects.h sucks in obj_mac.h which is automatically generated from objects.txt, so we're already there. NIDs have already been assigned and everything, just have to use them. -Werror is set in compilation flags, so can't use #warning. Remember to clean up the new stuff ni x509v3.h! I've seen occasional printouts of what look like empty address extensions, need to check that those really are empty (as opposed to just a printout bug that's aborting i2r...). asid v2i is losing on ranges because the ASN1_INTEGER reader code doesn't like our punctuation. strdup() is probably the right answer. addr v2i is having similar problems, inet_pton() doesn't like the slash of a prefix. Same answer. Might be able to get rid of inet_pton(): see ipv4_from_asc() and ipv6_from_asc() in v3_utl.c. Possible approach to memory leak: debugging code that prints out all pointers used in an extension just before our v2i function returns. Should be straightforward enough, and would also confirm this whackiness of being called twice. Heck, could use our i2r handler to do pretty output.