Re: [RFC][PATCH 4/8] SLIM main patch

[Seth Arnold]

On Fri, Aug 18, 2006 at 10:24:02AM -0700, Kylene Jo Hall wrote:
> Thanks for your detailed review.  Responses inline below.

My pleasure; thanks for considering my comments.

> throughout (iac vs. sac).  However, on closer look I think you mean the
> ones which are actually defined with "#define" and those will not be a
> problem to move.

Yes, sorry for the confusion; I meant the #defines. :)

> > All of these booleans could be re-written to simply return the value of
> > the boolean check. I don't know if those are actually easier to read,
> > but someone should see them once and decide. :)
> > 
> Let me make sure I understand.  You think "return (isec && isec-
> >level.iacl_level != SLM_IAC_NOTDEFINED);" would be easier?

Yeah, I'm not a fan of if (foo) return 1 else return 0; perhaps I'm in
the minority, but I'd feel better if the choice of these two styles were
made consciously:

static int is_isec_defined(struct slm_isec_data *isec)
{
	if (isec && isec->level.iac_level != SLM_IAC_NOTDEFINED)
		return 1;
	return 0;
}

static int is_isec_defined(struct slm_isec_data *isec)
{
	return isec && (isec->level.iac_level != SLM_IAC_NOTDEFINED);
}

If you find the first easier to work with, then feel free to keep it.

> > > +static void revoke_mmap_wperm(struct slm_file_xattr *cur_level)
> > > +{
> > > +	struct vm_area_struct *mpnt;
> > > +	struct file *file;
> > > +	struct dentry *dentry;
> > > +	struct slm_isec_data *isec;
> > > +
> > > +	flush_cache_mm(current->mm);
> > 
> > Is it a good idea to flush the cache before making the modifications?
> > Feels like the wrong order to me.
> 
> Our thought was that we are going to revoke write access to the file at
> this point, bu all pending writes are still valid. So we flush to make
> sure they can still be written (since we are revoking permission to that
> operation).

Is there any danger of another task sharing the mm structure to
repopulate the cache before your changes are incorporated?

> > slm_get_xattr() seems remarkably subtle given its name: *status can be
> > updated at two points in the function, a positive 'rc' from
> > integrity_verify_data() is left to return at the end, but negative 'rc'
> > values (that aren't -EOPNOTSUPP) get returned immediately, and if an
> > error variable is negative, a specific value is returned..
> > 
> Yes that does look fishy.  I'll try to straighten it out better.

> > The complicated set of decision making in get_level() (which sets
> > levels, heh) might be simplified if slm_get_xattr() internals were
> > less complicated.
> 
> Yes I'll try to straighten out too.  The last review it was requested
> that INTEGRITY status be returned from the hooks in a *int and regular
> kernel errors returned from the function to get error propogation
> correct but seems like we still have it over complicated.

Ah, yes, separating integrity status from error returns makes sense; I
hope the functions can be made to read as clearly as your description.

> > > +	/* Derived from include/linux/sched.h:capable. */
> > > +	if (cap_raised(tsk->cap_effective, cap)) {
> > > +		spin_lock(&tsec->lock);
> > > +		if (tsec->iac_wx == SLM_IAC_UNTRUSTED &&
> > > +		    cap == CAP_SYS_ADMIN)
> > > +			rc = -EACCES;
> > 
> > Why is CAP_SYS_ADMIN handled specially?
> This function is here to add the ability to remove capabilities.
> CAP_SYS_ADMIN should definitely be removed even if you are running as
> root but have been demoted to UNTRUSTED.  We are testing others but some
> tend to break existing applications.

CAP_SYS_MODULE, CAP_SYS_PTRACE, CAP_SYS_RAWIO, come to mind immediately. :)

> > > +static int slm_ptrace(struct task_struct *parent, struct task_struct *child)
> > > +{
> > > +	struct slm_tsec_data *parent_tsec = parent->security,
> > > +	    *child_tsec = child->security;
> > > +	int rc = 0;
> > > +
> > > +	if (is_kernel_thread(parent) || is_kernel_thread(child))
> > > +		return 0;
> > 
> > Why was this added?
> > 
> Kernel threads are never demoted or restricted by SLIM

Makes sense; but should an UNTRUSTED process really have the ability to
ptrace a kernel thread? (Ok, on my tests I wasn't able to strace attach
to kernel threads, but I'm not positive that it can't be done.)


Thanks Kylene

Attachment: pgp00000.pgp
Description: PGP signature