Re: [PATCH 11/12] proc: make check_mem_permission() return anmm_struct on success

[Stephen Wilson]

On Sun, Mar 13, 2011 at 05:08:59PM -0700, Kees Cook wrote:
> On Sun, Mar 13, 2011 at 03:49:23PM -0400, Stephen Wilson wrote:
> >  	copied = -EIO;
> >  	if (file->private_data != (void *)((long)current->self_exec_id))
> > -		goto out;
> > +		goto out_mm;
> 
> The file->private_data test seems wrong to me. Is there a case were the mm
> returned from check_mem_permission(task) can refer to something that is no
> longer attached to task?
> 
> For example:
> - pid 100 ptraces pid 200
> - pid 100 opens /proc/200/mem
> - pid 200 execs into something else

If the _target_ task (pid 200) execs then we are OK -- we hold a
reference to the *old* mm and it is that to which we read and write via
access_remote_vm().

In the case of the file->private_data test we are looking at the
*ptracer* (pid 100).  Here we are guarding against the case where the
tracer exec's and accidentally leaks the fd (hence the test wrt
current).  IOW, /proc/pid/mem is implicitly close on exec.  This is just
a minor feature to protect against buggy user space reading/writing
mistakenly into the targets address space.

> only after passing check_mem_permission(task) again. This is stopped
> by the private_data test. But should it, since check_mem_permission()
> passed?

No.  I hope the above clears that up.

> Even if it does mean to block it, it's insufficient since pid 200
> could just exec u32 many times and align with the original private_data
> value. 

Just for clarity, in your example it would be pid 100 that would need to
exec many times.  And yes, I think it would be possible for pid 100 to
exec() N times before the next call to mem_read/mem_write and thus
subvert this check.

Perhaps we can improve things (I would need to look into how O_CLOEXEC
is implemented), however please note that the primary rationale here is
to protect against bugs:  the tracer already has the needed privilege,
and it would be silly for it to exec N times just to pass the fd out
across an exec().


> What is that test trying to do? And I'm curious for both mem_write
> as well as the existing mem_read use of the test, since I'd like to see
> a general solution to the "invalidate /proc fds across exec" so we can
> close CVE-2011-1020 for everything[1].

These patches certainly do not add to the problem -- but they do not try
to address the general issue either.

> Associated with this, the drop of check_mem_permission(task) during the
> mem_read loop implies that the mm is locked during that loop and seems to
> reflect what you're saying ("Holding a reference to the target mm_struct
> eliminates this vulnerability."), meaning there's no reason to recheck
> permissions. Is that accurate?

Yes, precisely.  Once we have a reference to the mm we do not need to
worry about things changing underneath our feet, so the second check in
mem_read() is redundant and can be dropped.


Take care,

> 
> Thanks,
> 
> -Kees
> 
> [1] https://lkml.org/lkml/2011/2/7/368
> 
> -- 
> Kees Cook
> Ubuntu Security Team


-- 
steve

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/