In message <589b86c1.10...@roumenpetrov.info> on Wed, 08 Feb 2017 22:59:45 +0200, Roumen Petrov <open...@roumenpetrov.info> said:
openssl> Hi Richard, openssl> openssl> Richard Levitte wrote: openssl> > Hi, openssl> > openssl> > I've some ponderings that I need to bounce a bit with you all. openssl> > openssl> > Some have talked about replace the X509_LOOKUP_METHOD openssl> X.509 lookup method could return certificate , revocation list or openssl> EVP_KEY (structure x509_object_st). openssl> openssl> Unfortunately functionality of EVP_KEY was never implemented. openssl> Another point is specific names of structures - x509_lookup_method_st openssl> , x509_lookup_st, x509_object_st. openssl> Third point is quite specific implementation - functions not just to openssl> retrieve objects( X.509 or CRL) but to fill them into "context of X509 openssl> store". openssl> openssl> Current lookup functionality look like "store" but implementation is openssl> specific to X.509 store. openssl> openssl> openssl> > bit with the openssl> > STORE module I'm building, and while STORE isn't ready for it yet openssl> openssl> I hope that you store functionality will fill gap between load of keys openssl> and load of certificates (+crl). openssl> openssl> Loadable module (engine) has interface to load key(private or public) openssl> but lack load of X.509 certificates or CRL. So far, key parameters, pkeys, certs and crls are covered... oh, and names, if the given URI would return a list of names (file:/foo/bar/ would typically do that, for example). openssl> > , I openssl> > have some thoughts on how the two can approach each other. This would openssl> > involve one or two hooks / callbacks, that a STORE user could specify openssl> > (details later) to pick and choose freely among the objects that the openssl> > STORE module finds (be it on file or whatever else that can be openssl> > represented as a URI). openssl> I think that functionality requires three phases : openssl> 1) instantiation : at this point store is created openssl> 2) specification (optional): set or check capability of store. For openssl> instance store could return only X.509 certificates or to request openssl> store to return only keys. openssl> 3) inquiry: fetch data based on specified criteria. Previous attempts at a STORE effort were along those lines of thinking. Unfortunately, they got stuck in the exact specification of attributes and how to combine them (i.e. forming a "language" of sorts), and then to figure out how to make that pratical, especially for the engines that would have to intrepret it. So this time around, I'm trying to make OpenSSL fairly agnostic and leave the exact spec up to be specified by the user or application in the URI and letting the diverse engines interpret the URI components at their leasure. Objects would then be returned according to those specs in form of one object at a time, be it parameters, keys, certs, crls or names, for the application or other parts of OpenSSL to do whatever they wish with. Given this, I determined yesterday that there would really just be one hook / callback that makes sense, and that's one that can massage or possibly discard returned objects before they make their way all the way to the calling application. Something like this. openssl> > The troublesome part would be to try to mimic by_dir... It highly openssl> > depends on the specified paths to really be directories, and that it openssl> > should find what it wants by adding very specific file names (a hash openssl> > of the subject name with a ".{n}" or ".r{n}" extension for X.509 certs openssl> > and for X.509 CRLs). And sure, that works, but will really only work openssl> > with regular files. openssl> I'm not sure what is issue. openssl> openssl> Lets see X.509 lookup method get_by_subject. openssl> openssl> - by_dir openssl> 2) specification : set directory(path), limit results to X.509 or CRL openssl> and may be to inform store that questions will be performed by openssl> subject. openssl> openssl> For instance URI scheme could befile://path?certificate="name" openssl> openssl> 3) query : from subject calculate hash and then process "{hash}.{n}" openssl> or "{hash}.r{n}" depending from URI openssl> openssl> - Ldap openssl> It is similar, URI is described in RFCs - at point 2) set host, port, openssl> base distinguished name, attribute (for instance cACertificate), openssl> construct filter from specified name. The issue is there, glaring back at you. How will by_dir know exactly how to massage the URI? Sure, we could program in the big known ones plus our own mangling of the file: scheme, but then what? What about the vendor who runs his own scheme foo:? How do we know how to encode their keystore specs into a URI with their scheme? There needs to be a solution where OpenSSL can be specification agnostic. So far, I've this idea of returned names that form a "directory listing", and it wouldn't be too hard to imagine a by_dir implementation that simply get a list of names, then loads each of them and caches *everything*, or possibly one that uses the hook / callback I mentioned further up to select the one with the subject it's looking for. It is workable, I'm just worried about the sudden huge "directory listing". openssl> > What if someone would specify a LDAP URI that can return a bunch of openssl> > objects? openssl> > openssl> > So... my ponderings are going along these lines: openssl> > openssl> > 1. Should the directory X509_LOOKUPs be restricted to on disk openssl> > directories, or should "directory" be redefined as "whatever URI openssl> > that returns a collection of objects"? The latter would mean that openssl> > all those objects get loaded and that a hook / callback would then openssl> > be called to check if it's an object that corresponds to what we openssl> > search for. openssl> I think that replacement of "by_dir" lookup has to be restricted to openssl> file system operation. That's workable but sub-optimal... we've had a recent complaint from someone who wants to create a X509_LOOKUP_METHOD for his own scheme (hw protected keys, I think). PKCS#11 has been mentioned quite a number of times as well. But sure, for the time being. openssl> Files could be located on network, memory not only on disk. Does that differ from what we have today? openssl> > 2. For on disk directories, should we preserve the rehash file form? openssl> > In other words, if we decide to load everything we can find, shall openssl> > we restrict the loading to files matching the regexp openssl> > [0-9a-f]{8}\.r?[0-9]+ ? If not, are we about to create a new form openssl> > of key store for ourselves and our users? Should we? openssl> openssl> For hash-dir please keep current file name format. Like I said, a changed keystore format isn't possible before 1.2.0 anyway, so at least for now, the hash-dir format stays. What I wonder, generally speaking, is that for a URI that's a directory spec, show the STORE file: scheme return the names of *all* the file names in that directory, or just those matching [0-9a-f]{8}\.r?[0-9]+ ? Either can be done, I just worry about the consequences... openssl> > Quite a lot also depends on what OpenSSL version we aim for. I would openssl> > very much like to see the STORE module itself become part of 1.1.1, openssl> > but a new key store to replace our current rehash links will obviously openssl> > have to wait 'til 1.2.0. -- Richard Levitte levi...@openssl.org OpenSSL Project http://www.openssl.org/~levitte/ -- openssl-dev mailing list To unsubscribe: https://mta.openssl.org/mailman/listinfo/openssl-dev