[Freeipa-devel] [PATCH 149] IPA KDB: allow case in-sensitive realm in AS request

Mon Aug 10 21:37:17 UTC 2015

On Mon, 2015-08-10 at 23:15 +0300, Alexander Bokovoy wrote:
> On Mon, 10 Aug 2015, Alexander Bokovoy wrote:
> >On Sun, 09 Aug 2015, Simo Sorce wrote:
> >>On Fri, 2015-08-07 at 23:56 +0300, Alexander Bokovoy wrote:
> >>>On Tue, 28 Jul 2015, Sumit Bose wrote:
> >>>>On Tue, Jul 28, 2015 at 01:42:29PM +0200, Sumit Bose wrote:
> >>>>> On Tue, Jul 28, 2015 at 02:26:34PM +0300, Alexander Bokovoy wrote:
> >>>>> > On Tue, 28 Jul 2015, Simo Sorce wrote:
> >>>>> > >On Tue, 2015-07-28 at 12:15 +0200, Sumit Bose wrote:
> >>>>> > >>On Wed, Jul 22, 2015 at 09:41:51AM -0400, Simo Sorce wrote:
> >>>>> > >>> ----- Original Message -----
> >>>>> > >>> > From: "Sumit Bose" <sbose at redhat.com>
> >>>>> > >>> > To: "freeipa-devel" <freeipa-devel at redhat.com>
> >>>>> > >>> > Sent: Tuesday, July 21, 2015 7:41:14 AM
> >>>>> > >>> > Subject: [Freeipa-devel] [PATCH 149] IPA KDB: allow case in-sensitive realm	in AS request
> >>>>> > >>> >
> >>>>> > >>> > Hi,
> >>>>> > >>> >
> >>>>> > >>> > this patch is my suggestion to solve
> >>>>> > >>> > https://fedorahosted.org/freeipa/ticket/4844 .
> >>>>> > >>> >
> >>>>> > >>> > The original issue in the ticket has two part. One is a loop in libkrb5
> >>>>> > >>> > which is already fixed. The other is to handle canonicalization better.
> >>>>> > >>>
> >>>>> > >>> Sorry Sumit,
> >>>>> > >>> I see several issues with this patck.
> >>>>> > >>>
> >>>>> > >>> first of all you should really not change ipadb_get_principal(), that's the
> >>>>> > >>> wrong place to apply your logic.
> >>>>> > >>>
> >>>>> > >>> To support searching for the realm name case-insensitively all we should do
> >>>>> > >>> is to always forcibly upper case the realm name at the same time we build the
> >>>>> > >>> filter (in ipadb_fetch_principals(), if canonicalization was requested.
> >>>>> > >>> Because we will never store (code to prevent that should probably be dded with
> >>>>> > >>> this patch) a realm name that is not all caps.
> >>>>> > >>> Then the post search matches should be done straight within ipadb_find_principal().
> >>>>> > >>>
> >>>>> > >>> > The general way to allow canonicalization on a principal is to add the
> >>>>> > >>> > attributes 'krbcanonicalname'[1] and 'ipakrbprincipalalias' together
> >>>>> > >>> > with the objectclass 'ipaKrbPrincipal' to the user object.
> >>>>> > >>>
> >>>>> > >>> We have already a ticket open since long to remove krbprincipalalias, it was
> >>>>> > >>> a mistake to add it and any patch that depends on it will be nacked by me.
> >>>>> > >>> We need to use krbPrincipalName and krbCanonicalName.
> >>>>> > >>>
> >>>>> > >>> > Then the IPA
> >>>>> > >>> > KDB backend will use 'ipakrbprincipalalias' for case in-sensitive
> >>>>> > >>> > matches and  the principal from 'krbcanonicalname' will be the canonical
> >>>>> > >>> > principal used further on. The 'krbPrincipalName' is not suitable for
> >>>>> > >>> > either because it has caseExact* matching rules and is a multivalue
> >>>>> > >>> > attribute [2].
> >>>>> > >>>
> >>>>> > >>> Case-exact match is a problem only if we do not canonicalize names when storing
> >>>>> > >>> them, otherwise all you need to do is store a "search form" in krbPrincipalName
> >>>>> > >>> and always change searches to that form (forcibly upper case realm, forcibly
> >>>>> > >>> lowercase components) when canonicalization is requested.
> >>>>> > >>>
> >>>>> > >>> Additionally in the patch you are using stcasecmp(), that function is not
> >>>>> > >>> acceptable, look at ipadb_find_principal() and you'll see we use ulc_casecmp()
> >>>>> > >>> there.
> >>>>> > >>> Also modyfing the principal before searching is done wrong (you use strchr()
> >>>>> > >>> to find the @ sign, but you could find an @ in the components this way, you
> >>>>> > >>> should use strrchr() at the very least), and is dangerous if done outside of
> >>>>> > >>> the inner functions because then we never have a way to know the original
> >>>>> > >>> form should it be needed. In any case as said above realm should be forcibly
> >>>>> > >>> uppercase, given a flag in the escape function instead.
> >>>>> > >>
> >>>>> > >>Thank for for the review and the comments.
> >>>>> > >>
> >>>>> > >>I changed the patch as you suggested to upper-case the realm in the
> >>>>> > >>escape function if the flag is set.
> >>>>> > >>
> >>>>> > >>I didn't add any checks to make sure that the realm of newly added
> >>>>> > >>principal attributes is always upper case. Since the attributes can be
> >>>>> > >>added via various ways I think the check should happen on the DS level
> >>>>> > >
> >>>>> > >We should indeed intercept add/modify operations and see if they try to
> >>>>> > >set krbPrincipalName/krbCanonicalName and then validate the name.
> >>>>> > >Return unwilling to perform if the case of the realm is different (or
> >>>>> > >fix it on the fly, up for discussion) from the default case as
> >>>>> > >configured in the server.
> >>>>> > Will break trusts -- ipasam does add these principals for krbtgt/IPA at AD.
> >>>>> >
> >>>>> > >>but I see this more in the context of full canonicalization fix covered
> >>>>> > >>by https://fedorahosted.org/freeipa/ticket/3864 . If you think this is a
> >>>>> > >>requirement for the patch attached I would suggest to drop
> >>>>> > >>https://fedorahosted.org/freeipa/ticket/4844 and solve it together with
> >>>>> > >>#3864.
> >>>>> > >
> >>>>> > >We should clsoe 4844 as fixed upstream (there *was* a bug in libkrb5).
> >>>>> > >I commented on #3864 about what we can do, and we can also avoid
> >>>>> > >changing the schema.
> >>>>> > Yep.
> >>>>> >
> >>>>> > >So on the new patches, what does "unify" means ? I do not get what it
> >>>>> > >means (so probably it is a poor name), I guess you may want to call it
> >>>>> > >"canonicalization" ? (or even 'canon' to shorten it a bit).
> >>>>> > I have same question. I tried to understand why it is called unify and
> >>>>> > failed.
> >>>>>
> >>>>> I didn't want to use 'canonical' because the result will not be the
> >>>>> canonical name in the general case but only a name we use for searching.
> >>>>> I was thinking about 'normalized' bit this has a special meaning with
> >>>>> unicode. So I came up with 'unify'. But if you prefer 'canon' I can
> >>>>> change it.
> >>>>>
> >>>>> >
> >>>>> > >I think the worst case for a utf8 string is more then length*2, probably
> >>>>> > >more like length*6, unless there is some guarantee around case changes
> >>>>> > >that I am not aware of, that said we could probably just allocate on the
> >>>>> > >stack a fixed size string of a KiB or so, the longest DNS name is 256
> >>>>> > >chars IIRC and a service name can't be that much longer, also usernames
> >>>>> > >can't be arbitrarily long. So 1/2 KiB should probably be fine for a full
> >>>>> > >principal name. (avoids a malloc too which is good).
> >>>>> > Yes, sounds good. A hostname label can be up to 63 characters and full
> >>>>> > domain name including dots would be 253 characters. At the same time, a
> >>>>> > a component of the principal may be of arbitrary length. From practical
> >>>>> > perspective it would probably be enough to go with a static buffer of
> >>>>> > 1/2 KiB for the quickest case and fall back to malloc() if the size is
> >>>>> > bigger than that one.
> >>>>>
> >>>>> ok, I will change this.
> >>>>
> >>>>new version with changed name and 1/2 KiB buffer attached. No changes to
> >>>>the 2nd patch.
> >>>Thanks.
> >>>
> >>>Patches look good to me. I, perhaps, would have added
> >>>  char *canon_princ = NULL;
> >>>
> >>>in the definition of canon_princ but as you never access it in case
> >>>asprintf() failed, that's fine.
> >>>
> >>>Simo?
> >>>
> >>
> >>LGTM.
> >Hold on. I think I've found a bug -- when krbPrincipalName values match
> >insensitively but krbCanonicalName value is missing, we do not set
> >principal to the matched value. This breaks canonicalization for case
> >when there is only one krbPrincipalName as you don't need to have
> >krbCanonicalName in this case.
> >
> >I have a prototype which still misses checks.
> ... and I think we miss checks in few other places. I'm getting
> canonicalization working randomly -- sometimes one or two times in a row
> I get 'Client principal is not found' for canonicalization case:
> 
> Aug 10 20:01:21 m1.example.com krb5kdc[18758](Error): searched for
> admin at example.com, found admin at EXAMPLE.COM, result is 1, index is 0,
> next val is (nil)
> 
> Aug 10 20:01:21 m1.example.com krb5kdc[18758](Error): searched for
> krbtgt/example.com at example.com, found krbtgt/EXAMPLE.COM at EXAMPLE.COM,
> result is 1, index is 0, next val is (nil)
> 
> Aug 10 20:01:21 m1.example.com krb5kdc[18758](info): AS_REQ (6 etypes
> {18 17 16 23 25 26}) 192.168.122.99: NEEDED_PREAUTH: admin at example.com
> for krbtgt/example.com at example.com, Additional pre-authentication
> required
> 
> Aug 10 20:01:24 m1.example.com krb5kdc[18758](Error): searched for
> admin at example.com, found admin at EXAMPLE.COM, result is 1, index is 0,
> next val is (nil)
> 
> Aug 10 20:01:24 m1.example.com krb5kdc[18758](Error): searched for
> krbtgt/example.com at example.com, found krbtgt/EXAMPLE.COM at EXAMPLE.COM,
> result is 1, index is 0, next val is (nil)
> 
> Aug 10 20:01:24 m1.example.com krb5kdc[18758](info): AS_REQ (6 etypes
> {18 17 16 23 25 26}) 192.168.122.99: ISSUE: authtime 1439236884, etypes
> {rep=18 tkt=18 ses=18}, admin at example.com for
> krbtgt/example.com at example.com
> 
> Aug 10 20:01:58 m1.example.com krb5kdc[18758](info): AS_REQ (6 etypes
> {18 17 16 23 25 26}) 192.168.122.99: CLIENT_NOT_FOUND: admin at example.com
> for krbtgt/example.com at example.com, Client not found in Kerberos
> database
> 
> These are logs with debugging I've added.
> -- 
> / Alexander Bokovoy
> 

After looking carefully at this with Alexander I think there are too
many things to fix and check, and given the looming deadline for Fedora
we should just postpone. These patches are not critical for us but it
would be bad if they'd go in and not work as expected.

I can take a better look at them when back from flock.

Simo.