Re: [tip:x86/mm] x86/mm: Consider effective protection attributes in W+X check

From: Andrey Ryabinin
Date: Mon Feb 26 2018 - 07:45:45 EST




On 02/26/2018 01:54 PM, Jan Beulich wrote:
>>>> On 26.02.18 at 11:47, <aryabinin@xxxxxxxxxxxxx> wrote:
>
>>
>> On 02/26/2018 01:08 PM, Jan Beulich wrote:
>>>>>> On 26.02.18 at 11:00, <aryabinin@xxxxxxxxxxxxx> wrote:
>>>> On 02/26/2018 11:48 AM, tip-bot for Jan Beulich wrote:
>>>>> @@ -351,7 +362,7 @@ static inline bool kasan_page_table(struct seq_file *m,
>> struct pg_state *st,
>>>>> (pgtable_l5_enabled && __pa(pt) == __pa(kasan_zero_p4d)) ||
>>>>> __pa(pt) == __pa(kasan_zero_pud)) {
>>>>> pgprotval_t prot = pte_flags(kasan_zero_pte[0]);
>>>>> - note_page(m, st, __pgprot(prot), 5);
>>>>> + note_page(m, st, __pgprot(prot), 0, 5);
>>>>
>>>> Isn't this disables W+X check for kasan page table?
>>>> Methinks it should be 'prot' here.
>>>
>>> Might well be - I actually did ask the question before sending v3,
>>> but didn't get any answer (yet). The kasan_zero_p?d names
>>> suggested to me that this is a shortcut for mappings which
>>> otherwise would be non-present anyway, but that was merely a
>>> guess.
>>
>> kasan_zero_p?? are used to map kasan_zero_page. That's it.
>
> Ah, thanks for explaining.
>
>>> As to W+X checks - I can't see how the result could be
>>> any better if the protections of kasan_zero_pte[0] would be
>>> used: Those can't possibly be applicable independent of VA.
>>
>> I'm not sure I understand what do you mean.
>> If we somehow screw up and accidentally make kasan_zero_pte writable and executable,
>> note_page() should report this. With your patch, it won't work.
>
> If this is a case to care about, simply passing "prot" won't be right
> though - the callers accumulated effective protections would then
> need passing in here, and merging with prot.
>

Fine, but this won't change anything. Since kasan_zero_pte[] always ro+nx, the effective
protections should be always the same.

> Before I do this for a possible v4, I'd like to seek clarification
> though whether this really is a case to care about.

It may be not that important case, but one of the points of this code is to check for the absence
of W+X mappings. Passing known to be wrong value to bypass that check is certainly not the
right thing to do.


> Jan
>