Re: [PATCH v1] mm/page_alloc: clear pages in alloc_contig_pages() with init_on_alloc=1 or __GFP_ZERO

[David Hildenbrand <david@redhat.com>]

On 11.11.20 11:22, Michal Hocko wrote:
> On Wed 11-11-20 11:05:21, David Hildenbrand wrote:
>> On 11.11.20 10:58, Vlastimil Babka wrote:
>>> On 11/11/20 10:06 AM, David Hildenbrand wrote:
>>>> On 11.11.20 09:47, Michal Hocko wrote:
>>>>> On Tue 10-11-20 20:32:40, David Hildenbrand wrote:
>>>>>> commit 6471384af2a6 ("mm: security: introduce init_on_alloc=1 and
>>>>>> init_on_free=1 boot options") resulted with init_on_alloc=1 in all pages
>>>>>> leaving the buddy via alloc_pages() and friends to be
>>>>>> initialized/cleared/zeroed on allocation.
>>>>>>
>>>>>> However, the same logic is currently not applied to
>>>>>> alloc_contig_pages(): allocated pages leaving the buddy aren't cleared
>>>>>> with init_on_alloc=1 and init_on_free=0. Let's also properly clear
>>>>>> pages on that allocation path and add support for __GFP_ZERO.
>>>>>
>>>>> AFAIR we do not have any user for __GFP_ZERO right? Not that this is
>>>>
>>>> Sorry, I had extended information under "---" but accidentally
>>>> regenerated the patch before sending it out.
>>>>
>>>> __GFP_ZERO is not used yet. It's intended to be used in
>>>> https://lkml.kernel.org/r/20201029162718.29910-1-david@redhat.com
>>>> and I can move that change into a separate patch if desired.
> 
> OK, it would make sense to add it with its user.
> 
>>>>> harmful but it is better to call that explicitly because a missing
>>>>> implementation would be a real problem and as such a bug fix.
>>>>>
>>>>> I am also not sure handling init_on_free at the higher level is good.
>>>>> As we have discussed recently the primary point of this feature is to
>>>>> add clearing at very few well defined entry points rather than spill it over
>>>>> many places. In this case the entry point for the allocator is
>>>>> __isolate_free_page which removes pages from the page allocator. I
>>>>> haven't checked how much this is used elsewhere but I would expect
>>>>> init_on_alloc to be handled there.
>>>>
>>>> Well, this is the entry point to our range allocator, which lives in
>>>> page_alloc.c - used by actual high-level allocators (CMA, gigantic
>>>> pages, etc). It's just a matter of taste where we want to have that
>>>> handling exactly inside our allocator.
> 
> Yes I completely agree here. I just believe it should the lowest we can
> achieve.
> 
>>> I agree alloc_contig_range() is fine as an entry point.
>>
>> Thanks, let's see if Michal insists of having this somewhere inside
>> isolate_freepages_range() instead.
>   
> It's not that I would be insisting. I am just pointing out that changes
> like this one go against the idea of init_on_alloc because it is adding
> more special casing and long term more places to be really careful about
> when one has to be really careful to not undermine the security aspect
> of the feature. I haven't really checked why compaction is not the
> problem but I suspect it is the fact that it unconditionally copy the
> full page content to the isolated page so there is no way to sneak
> any data leak there. That is fine. We should however make that clear by

Exactly.

> using a special cased function which skips this particular
> initialization and make sure everybody else will just do the right thing
> without much thinking.

I totally agree, but I think we don't have many places where free pages 
actually leave the buddy besides alloc_pages() and friends (compaction 
is something special). I agree having a single place to handle that 
would be preferred. I'll have a look if that can be reworked without 
doing too much harm / affecting other hot paths.

-- 
Thanks,

David / dhildenb