by Linus Torvalds

[permalink] [raw]

Subject: Re: [GIT PULL] x86/topology changes for v5.3

On Tue, Jul 9, 2019 at 4:00 PM Thomas Gleixner <[email protected]> wrote:
>
> That still does not explain the cr4/0 issue you have. Can you send me your
> .config please?

Gaah. Now I'm off my desktop and won't be at it again until tomorrow.
And that's the one that bisected to the cr0/cr4 bits (and I did all my
bisection on that because it's so much faster).

Anyway, the kernel config itself is pretty simple. It's basically a
F30 kernel config that has been through "make localmodconfig" and then
pared down some more to remove stuff I don't use (paravirt etc). I
would expect that if it's a confiuration difference, it's more likely
to be about the user mode side - since the hang does seem to happen in
user mode, not in the kernel (but a very early problem at cpu bringup
time could easily have been hidden).

On my laptop (which I am at right now), the hang is different, and
maybe it's similar to your ACPI hang issue. I will try that revert,
and at least see if that solves the laptop side.

Linus

2019-07-10 03:42:15

[permalink] [raw]

> On Jul 11, 2019, at 8:08 AM, Kees Cook <[email protected]> wrote:
>
> On Thu, Jul 11, 2019 at 10:01:34AM +0200, Peter Zijlstra wrote:
>> On Thu, Jul 11, 2019 at 07:11:19AM +0000, Nadav Amit wrote:
>>>> On Jul 10, 2019, at 7:22 AM, Jiri Kosina <[email protected]> wrote:
>>>>
>>>> On Wed, 10 Jul 2019, Peter Zijlstra wrote:
>>>>
>>>>> If we mark the key as RO after init, and then try and modify the key to
>>>>> link module usage sites, things might go bang as described.
>>>>>
>>>>> Thanks!
>>>>>
>>>>>
>>>>> diff --git a/arch/x86/kernel/cpu/common.c b/arch/x86/kernel/cpu/common.c
>>>>> index 27d7864e7252..5bf7a8354da2 100644
>>>>> --- a/arch/x86/kernel/cpu/common.c
>>>>> +++ b/arch/x86/kernel/cpu/common.c
>>>>> @@ -366,7 +366,7 @@ static __always_inline void setup_umip(struct cpuinfo_x86 *c)
>>>>> cr4_clear_bits(X86_CR4_UMIP);
>>>>> }
>>>>>
>>>>> -DEFINE_STATIC_KEY_FALSE_RO(cr_pinning);
>>>>> +DEFINE_STATIC_KEY_FALSE(cr_pinning);
>>>>
>>>> Good catch, I guess that is going to fix it.
>>>>
>>>> At the same time though, it sort of destroys the original intent of Kees'
>>>> patch, right? The exploits will just have to call static_key_disable()
>>>> prior to calling native_write_cr4() again, and the protection is gone.
>>>
>>> Even with DEFINE_STATIC_KEY_FALSE_RO(), I presume you can just call
>>> set_memory_rw(), make the page that holds the key writable, and then call
>>> static_key_disable(), followed by a call to native_write_cr4().
>>
>> Or call text_poke_bp() with the right set of arguments.
>
> Right -- the point is to make it defended against an arbitrary write,
> not arbitrary execution. Nothing is safe from arbitrary exec, but we can
> do our due diligence on making things read-only.

I don’t understand.

In the PoC that motivated this this patch [1], the attacker gained the
ability to call a function, control its first argument and used it to
disable SMEP/SMAP by calling native_write_cr4(), which he did before doing
an arbitrary write (another ability he gain).

After this patch, the attacker can instead call three functions, and by
controlling their first arguments (*) disable SMEP. I didn’t see something
in the motivating PoC that prevented calling 3 functions one at a time.

So it seems to me that it raised the bar for an attack by very little.

--

(*) set_memory_rw() has a second argument - the number of pages - but many
non-zero values may be fine (if not a warning from __cpa_process_fault()
might appear).

[1] https://googleprojectzero.blogspot.com/2017/05/exploiting-linux-kernel-via-packet.html