Date:	Wed, 30 Apr 2014 13:14:32 -0700
From:	Andy Lutomirski <luto@...capital.net>
To:	Kees Cook <keescook@...gle.com>
Cc:	Nathan Lynch <Nathan_Lynch@...tor.com>,
	"H. Peter Anvin" <hpa@...or.com>,
	"x86@...nel.org" <x86@...nel.org>,
	LKML <linux-kernel@...r.kernel.org>
Subject: Re: randomized placement of x86_64 vdso

On Wed, Apr 30, 2014 at 1:05 PM, Kees Cook <keescook@...gle.com> wrote:
> On Wed, Apr 30, 2014 at 11:30 AM, Andy Lutomirski <luto@...capital.net> wrote:
>> On Wed, Apr 30, 2014 at 11:13 AM, Kees Cook <keescook@...gle.com> wrote:
>>> On Wed, Apr 30, 2014 at 10:52 AM, Andy Lutomirski <luto@...capital.net> wrote:
>>>> On Wed, Apr 30, 2014 at 10:47 AM, Kees Cook <keescook@...gle.com> wrote:
>>>>> On Wed, Apr 23, 2014 at 10:06 AM, Nathan Lynch <Nathan_Lynch@...tor.com> wrote:
>>>>>> On 04/23/2014 11:30 AM, H. Peter Anvin wrote:
>>>>>>> On 04/21/2014 09:52 AM, Nathan Lynch wrote:
>>>>>>>> Hi x86/vdso people,
>>>>>>>>
>>>>>>>> I've been working on adding a vDSO to 32-bit ARM, and Kees suggested I
>>>>>>>> look at x86_64's algorithm for placing the vDSO at a randomized offset
>>>>>>>> above the stack VMA.  I found that when the stack top occupies the
>>>>>>>> last slot in the PTE (is that the right term?), the vdso_addr routine
>>>>>>>> returns an address below mm->start_stack, equivalent to
>>>>>>>> (mm->start_stack & PAGE_MASK).  For instance if mm->start_stack is
>>>>>>>> 0x7fff3ffffc96, vdso_addr returns 0x7fff3ffff000.
>>>>>>>>
>>>>>>>> Since the address returned is always already occupied by the stack,
>>>>>>>> get_unmapped_area detects the collision and falls back to
>>>>>>>> vm_unmapped_area.  This results in the vdso being placed in the
>>>>>>>> address space next to libraries etc.  While this is generally
>>>>>>>> unnoticeable and doesn't break anything, it does mean that the vdso is
>>>>>>>> placed below the stack when there is actually room above the stack.
>>>>>>>> To me it also seems uncomfortably close to placing the vdso in the way
>>>>>>>> of downward expansion of the stack.
>>>>>>>>
>>>>>>>> I don't have a patch because I'm not sure what the algorithm should
>>>>>>>> be, but thought I would bring it up as vdso_addr doesn't seem to be
>>>>>>>> behaving as intended in all cases.
>>>>>>>>
>>>>>>>
>>>>>>> If the stack occupies the last possible page, how can you say there is
>>>>>>> "space above the stack"?
>>>>>>
>>>>>> Sorry for being unclear.  I probably am getting terminology wrong.  What
>>>>>> I'm trying to express is that if the stack top is in the last page of
>>>>>> its last-level page table (which may be the last possible page, but
>>>>>> that's not really the interesting case), vdso_addr returns an address
>>>>>> below mm->start_stack.
>>>>>
>>>>> It seems like this is avoidable, then? From your example, it seems
>>>>> like we lose the separated randomization in this case, but we don't
>>>>> need to? Do you have a suggestion for what could be done to fix this?
>>>>
>>>> I don't understand why the vDSO should be special here.  Either the
>>>> standard logic for randomizing the placement of DSOs is good, in which
>>>> case it should be good for the vDSO too, or I think we should fix it
>>>> for everything.
>>>
>>> The issue is specific to the vdso randomizing-near-the-stack code;
>>> regular mmap randomization is operating correctly. The reason for
>>> randomizing stack, vdso, and mmap separately is to avoid correlation
>>> of leaked offsets in one to the other regions.
>>
>> I understand why the offset from the stack to the vDSO should be
>> randomized, and why the offset from the stack to, say, glibc should be
>> randomized.  What I don't get is why the offset from glibc to the vDSO
>> should be randomized but the offset from glibc to openssl should be
>> deterministic.  Or am I misunderstanding?
>
> Sure, I agree. In a perfect world, every DSO would have unrelated
> offsets. Ignoring that for the moment, my take on this is that stack,
> vDSO, and mmap have their locations tracked/exposed separately (SP
> register, AUXV, and syscall return respectively). As such, they should
> have separate randomization to avoid having any one method lead to the
> discovery of all the offsets (i.e. the vDSO location is communicated
> to the process in a different way than other DSOs that come into the
> process via the mmap syscall).
>
> As for per-mmap randomization, I would love this, but no one has
> proposed algorithms that aren't worse than single base offset
> randomization. This is especially true under 32-bit where attempting
> to balance the randomization against fragmentation just leads to
> non-random locations under certain situations/processes/DSOs, etc. It
> could be worth revisiting this for 64-bit, but I feel like it's not a
> high priority.

I think that the offending vDSO randomization code doesn't even run on 32-bit.

If anyone wants to fix this up (do something intelligent on 32-bit and
do something more intelligent on 64-bit), it might be easier if based
on my vdso cleanups
(https://git.kernel.org/cgit/linux/kernel/git/luto/linux.git/log/?h=vdso/cleanups).
 The code in question is a lot less twisted with those patches
applied.

--Andy

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

Powered by blists - more mailing lists