[lkml]   [2012]   [Aug]   [24]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
Patch in this message
SubjectRe: Search for patch for kernel stack data disclosure in binfmt_script during execve
Hash: SHA1

Kirill A. Shutemov wrote:
> On Wed, Aug 22, 2012 at 09:49:35PM +0000, halfdog wrote:
>> Got a hint via IRC, that I should not send patch idea for review
>> to "generic" list, but to maintainers and last (or relevant)
>> comitters of code.
>> halfdog wrote:
>>> halfdog wrote:
>>>> I'm searching for a patch for linux kernel stack disclosure
>>>> in binfmt_script with crafted interpreter names when
>>>> CONFIG_MODULES is active (see [1]).
>>> Please disregard my previous proposal [2], since it did not
>>> address the problem directly (referencing local stack frame
>>> data from bprm structure) but worked around it. I suspect,
>>> that this could increase probability to reintroduce similar
>>> bugs.
>>> Opinions on (untested sketch for) second solution: Could
>>> someone look on the source code comments and changes in patch
>>> to judge, if this is going in the right direction?
>>> Explanation of patch: Since load_script will start to
>>> irreversibly change bprm structures at some point (using stack
>>> local data was one of those changes), try to delay this point.
>>> Run checks if load_script could be the right handler, if not
>>> give other binfmt handlers the chance to do so.
>>> If binfmt_script is the right one, try to load the interpreter
>>> (causing bprm modification), if failing make sure that no
>>> other binfmt handler has the chance to continue on the now
>>> modified bprm data.
>>> CAVEAT: This assumes, that if binfmt_script could handle the
>>> load, that it would be the one and only binfmt with that
>>> ability, so no other one, e.g. binfmt_misc should have the
>>> chance to do so. If this assumption is wrong, leaving
>>> binfmt_script would have to rollback all bprm changes (e.g.
>>> restore old credentials).
>>> [1]
> What about (untested):
> diff --git a/fs/exec.c b/fs/exec.c index 574cf4d..ef13850 100644
> --- a/fs/exec.c +++ b/fs/exec.c @@ -1438,7 +1438,8 @@ int
> search_binary_handler(struct linux_binprm *bprm,struct pt_regs
> *regs) } read_unlock(&binfmt_lock); #ifdef CONFIG_MODULES - if
> (retval != -ENOEXEC || bprm->mm == NULL) { + if (retval !=
> -ENOEXEC || bprm->mm == NULL || + bprm->recursion_depth >
> BINPRM_MAX_RECURSION) { break; } else { #define printable(c)
> (((c)=='\t') || ((c)=='\n') || (0x20<=(c) && (c)<=0x7e))

- - From my understanding, this patch should not fix the problem, since
recursion depth is reset back to old value after call of binfmt handler.
This is done, so that fs/exec does not have to trust all binfmts to
reset the depth by themselfes when leaving.;a=blob;f=fs/exec.c;h=da27b91ff1e8cbe87d0fe42aa5d39513e6a9deeb;hb=HEAD
1408 read_unlock(&binfmt_lock);
1409 retval = fn(bprm, regs);
1410 /*
1411 * Restore the depth counter to its
starting value
1412 * in this call, so we don't have to
rely on every
1413 * load_binary function to restore it on
1414 */
1415 bprm->recursion_depth = depth;

I guess, the problem is, that recursion_depth usually not only limits
the depth, but also the maximal number of binfmt_xxx calls. That's why,
the use of local stack-frame data in bprm does not matter, there is no
going up the stack AND using bprm->interpreter, the last error is
terminates the search.

In the POC, search is not terminated because of ENOEXEC when max depth
reached and due to special filename, mod-loader triggers also (about 30
times? I do not known, if that could be a problem also, interfering with
other module loads. Usually non-root users cannot trigger rapid module
loads easily).

- --
PGP: 156A AE98 B91F 0114 FE88 2BD8 C459 9386 feed a bee
Version: GnuPG v1.4.11 (GNU/Linux)

 \ /
  Last update: 2012-08-24 12:41    [W:0.081 / U:2.568 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site