Re: [RFC] namei: prevent sgid-hardlinks for unmapped gids

From: Kees Cook
Date: Fri Nov 06 2015 - 19:16:49 EST


On Fri, Nov 6, 2015 at 4:11 PM, Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> On Fri, Nov 6, 2015 at 2:30 PM, Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote:
>> On Fri, Nov 6, 2015 at 1:59 PM, Kees Cook <keescook@xxxxxxxxxxxx> wrote:
>>> Adding Ted, who might know how this all hooks together. (The context
>>> is that a write() or truncate() on a setgid file clears the setgid,
>>> but mmap writes don't.)
>>>
>>> On Wed, Nov 4, 2015 at 9:59 AM, Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote:
>>>> On Tue, Nov 3, 2015 at 10:58 PM, Willy Tarreau <w@xxxxxx> wrote:
>>>>> On Tue, Nov 03, 2015 at 03:29:55PM -0800, Kees Cook wrote:
>>>>>> Using "write" does kill the set-gid bit. I haven't looked at
>>>>>> why.
>>>>>> Al or anyone else, is there a meaningful distinction here?
>>>>>
>>>>> I remember this one, I got caught once while trying to put a shell into
>>>>> a suid-writable file to get some privileges someone forgot to offer me :-)
>>>>>
>>>>> It's done by should_remove_suid() which is called upon write() and truncate().
>>>
>>> file_remove_privs() seems to be the right entry point.
>>> __generic_file_write_iter in mm/filemap.c calls it, though. Are these
>>> callbacks not used for mmap writes?
>>
>> They're certainly not used early enough -- we need to remove suid when
>> the page becomes writable via mmap (wp_page_shared), not when
>> writeback happens, or at least not only when writeback happens.
>
> Well, I'm shy about the change there. For example, we don't strip in
> on open(RDWR), just on write().

I take it back. Hooking wp_page_shared looks expensive. :) Maybe we do
need to hook the mmap?

-Kees

>
>> But IIRC mmaped writes go through a different path -- they go through
>> the address_space ops with names like writepages.
>
> Ah-ha.
>
>>>>>> Should the
>>>>>> mmap MAP_SHARED-write trigger the loss of the set-gid bit too? While
>>>>>> holding the file open with either open or mmap, I get a Text-in-use
>>>>>> error, so I would kind of expect the same behavior between either
>>>>>> close() and munmap(). I wonder if this is a bug, and if so, then your
>>>>>> link patch is indeed useful again. :)
>>>>>
>>>>> I don't see how this could be done with mmap(). Maybe we have a way to know
>>>>> when the first write is performed via this path, I have no idea.
>>>>
>>>> do_wp_page might be a decent bet.
>>>
>>> Or wp_page_shared? Can we get back to a file from the mm at that point?
>>
>> vma->vm_file, presumably (after checking whether it's null).
>> wp_page_shared AFAIK only happens from process context, and the vma
>> and its file should be valid.
>>
>> We could also get to an inode via page->address_space->mapping, but
>> I'm guessing that vma->vm_file would be more appropriate here.
>
> Yeah. Let me give it a try...
>
> -Kees
>
> --
> Kees Cook
> Chrome OS Security



--
Kees Cook
Chrome OS Security
--
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/