Re: [AppArmor 38/45] AppArmor: Module and LSM hooks

["Serge E. Hallyn" <serue@us.ibm.com>]

Quoting Andreas Gruenbacher (agruen@suse.de):
> On Monday 11 June 2007 16:33, Stephen Smalley wrote:
> > On Mon, 2007-06-11 at 01:10 +0200, Andreas Gruenbacher wrote:
> > > On Wednesday 06 June 2007 15:09, Stephen Smalley wrote:
> > > > On Mon, 2007-06-04 at 16:30 +0200, Andreas Gruenbacher wrote:
> > > > > On Monday 04 June 2007 15:12, Pavel Machek wrote:
> > > > > > How will kernel work with very long paths? I'd suspect some
> > > > > > problems, if path is 1MB long and I attempt to print it in /proc
> > > > > > somewhere.
> > > > >
> > > > > Pathnames are only used for informational purposes in the kernel,
> > > > > except in AppArmor of course.
> > > >
> > > > I don't mean this as a flame, but isn't the above statement the very
> > > > crux of this discussion?
> > >
> > > I think the question at the core of it all is, shall a pathname based
> > > security mechanism be allowed. I was under the impression that this
> > > question had already been answered affirmatively. If the answer here was

That was the decision at ksummit last year, yes.

> > > no, then we could stop the entire discussion right there.
> >
> > There is a difference between using the pathname at the kernel/userland
> > interface as part of configuring a security mechanism and using it as
> > the basis for the runtime checking itself.
> 
> Yes, there is a difference. When I say pathname based security mechanism, I 
> literally mean a pathname based security mechanism, meaning the pathnames 
> determine the outcome o the decision. This includes designs that are based on 
> different abstractions internally, but the bahavior observable from 
> user-space must be the same (or else, it's a different model).
> 
> Unfortunately, translating pathnames to labels destroys this fundamental 
> abstraction. We explained why this is so in the following postings:
> 
> 	http://lkml.org/lkml/2007/6/9/94
> 	http://lkml.org/lkml/2007/6/10/141
> 
> > Further, there is a difference between generating and matching full
> > pathnames on each access vs. caching information in the parent dentry and
> > making decisions based on that cached information and the last
> > component-name.
> 
> Wait, you are mixing two issues here: access checks on existing files, and the 
> creation of new files. For AppArmor as it stands today the two are the same, 
> but when looking at emulating AppArmor using labels, they are not. Let's look 
> at things one at a time.
> 
> Generating and matching full pathnames on each access takes time, no question 
> about that. (In fact we are not checking on each access but only when 
> pathnames are involved, such as on open. Filehandle based operations do not 
> require access checks, but that's not a very important difference at this 
> level of discussion.) That's a quantitative statement though, not a 
> qualitative one: retrieving xattrs and checking labels also takes time; 
> additional checks are never for free. We find that doing the pathname checks 
> is easily fast enough. You may disagree, but then you don't have to use 
> AppArmor, and we are not standing in your way.
> 
> As far as new files are concerned, basing decisions on the parent dentry and 
> component name requires that you know where in the filesystem hierarchy this 
> dentry is located: with bind mounts, the same dentry shows up in multiple 
> locations in a process's namespace, corresponding to different pathnames. In 
> other words, to make the right decision, the dentry alone is not enough; it 
> takes a <dentry, vfsmount> pair. So there we are again.
> 
> >From the point on where you have a <dentry, vfsmount> pair of objects, you can 
> do two things: you can compute the full pathname and base your decision on 
> that, or you can do some caching to hopefully cut some of that work short 
> frequently enough to set off the additional cost. The difference between the 
> two approaches is quantitative -- if there is a difference in results, then 
> that's obviously a bug. I believe that caching could speed up things 
> measurably, but up to this point, neither I nor anybody else had the time to 
> look into it, and so we are not doing it -- not yet, any perhaps never at 
> all. It may be counter to your intuition, but doing those checks is not a big 
> issue.

My approach in DTE, which used pathnames to assign TE labels in the
kernel, was to use a 'shadow tree' to the vfsmnt+dentry tree, filled out
at policy load time to the depth of the deepest policy rule.  For the
behavior I was after, bind mounts were handled by storing pointers from
the new mount to the original, but the behavior AA wants would be
different.  Namespace clones are easily handled this way by copying the
cached pointers to the shadow tree.  And it was pretty fast.

When I talked about this with Tony last year, the biggest shortcoming to
use this for AA was the wildcards.  For instance, if there is a rule for
/var/log/HOSTS/*/messages, then when /var/log/HOSTS/sergelap/ is
created, a new rule would have to be created on the fly to tag
/var/log/HOSTS/sergelap/messages if it is created.  How to represent
that when only /var/log or /var/log/HOSTS exists at policy load time
would be, well, a fun problem to solve actually  :)

-serge

> > The only question I saw being answered was the abstract one of "can a
> > pathname play a role in the security mechanism" not "should the kernel be
> > regenerating a full path on each open and glob matching it against a
> > list of file globs."
> 
> No glob matching involved, and no list of patterns. There is a single DFA per 
> profile (and each process is in at most one profile at any one time), and 
> each access check involves exactly one DFA traversal.
> 
> [cut here for lack of time right now -- will take another look later]
> 
> > > Without the vfsmounts propagated down you won't know the pathnames.
> > > Whether or not a different problem can be solved without the vfsmounts is
> > > not really relevant.
> >
> > Well, that presumes that your mechanism has to generate full pathnames
> > on each access check.  But even if so, you could be doing your checking
> > in the higher level code then where you have a vfsmount available.
> 
> The LSM hooks are pretty high-level, meant for being able to plug in different 
> access checks. That's the right level of abstraction. We just need the 
> vfsmounts passed down as well. That's what the bulk of common code patches 
> are there for.
> 
> Andreas
> -
> To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html