2020-11-01 21:18:58

by John Hubbard

[permalink] [raw]
Subject: Re: [PATCH v5 05/15] mm/frame-vector: Use FOLL_LONGTERM

On 11/1/20 2:30 AM, Daniel Vetter wrote:
> On Sun, Nov 1, 2020 at 6:22 AM John Hubbard <[email protected]> wrote:
>>
>> On 10/31/20 7:45 AM, Daniel Vetter wrote:
>>> On Sat, Oct 31, 2020 at 3:55 AM John Hubbard <[email protected]> wrote:
>>>> On 10/30/20 3:08 AM, Daniel Vetter wrote:
>> ...
>>>> By removing this check from this location, and changing from
>>>> pin_user_pages_locked() to pin_user_pages_fast(), I *think* we end up
>>>> losing the check entirely. Is that intended? If so it could use a comment
>>>> somewhere to explain why.
>>>
>>> Yeah this wasn't intentional. I think I needed to drop the _locked
>>> version to prep for FOLL_LONGTERM, and figured _fast is always better.
>>> But I didn't realize that _fast doesn't have the vma checks, gup.c got
>>> me a bit confused.
>>
>> Actually, I thought that the change to _fast was a very nice touch, btw.
>>
>>>
>>> I'll remedy this in all the patches where this applies (because a
>>> VM_IO | VM_PFNMAP can point at struct page backed memory, and that
>>> exact use-case is what we want to stop with the unsafe_follow_pfn work
>>> since it wreaks things like cma or security).
>>>
>>> Aside: I do wonder whether the lack for that check isn't a problem.
>>> VM_IO | VM_PFNMAP generally means driver managed, which means the
>>> driver isn't going to consult the page pin count or anything like that
>>> (at least not necessarily) when revoking or moving that memory, since
>>> we're assuming it's totally under driver control. So if pup_fast can
>>> get into such a mapping, we might have a problem.
>>> -Daniel
>>>
>>
>> Yes. I don't know why that check is missing from the _fast path.
>> Probably just an oversight, seeing as how it's in the slow path. Maybe
>> the appropriate response here is to add a separate patch that adds the
>> check.
>>
>> I wonder if I'm overlooking something, but it certainly seems correct to
>> do that.
>
> You'll need the mmap_sem to get at the vma to be able to do this
> check. If you add that to _fast, you made it as fast as the slow one.

Arggh, yes of course. Strike that, please. :)

> Plus there's _fast_only due to locking recurion issues in fast-paths
> (I assume, I didn't check all the callers).
>
> I'm just wondering whether we have a bug somewhere with device
> drivers. For CMA regions we always check in try_grab_page, but for dax

OK, so here you're talking about a different bug than the VM_IO | VM_PFNMAP
pages, I think. This is about the "FOLL_LONGTERM + CMA + gup/pup _fast"
combination that is not allowed, right?

For that: try_grab_page() doesn't check anything, but try_grab_compound_head()
does, but only for pup_fast, not gup_fast. That was added by commit
df3a0a21b698d ("mm/gup: fix omission of check on FOLL_LONGTERM in gup fast
path") in April.

I recall that the patch was just plugging a very specific hole, as opposed
to locking down the API against mistakes or confused callers. And it does
seem that there are some holes.

> I'm not seeing where the checks in the _fast fastpaths are, and that
> all still leaves random device driver mappings behind which aren't
> backed by CMA but still point to something with a struct page behind
> it. I'm probably just missing something, but no idea what.
> -Daniel
>

Certainly we've established that we can't check VMA flags by that time,
so I'm not sure that there is much we can check by the time we get to
gup/pup _fast. Seems like the device drivers have to avoid calling _fast
with pages that live in VM_IO | VM_PFNMAP, by design, right? Or maybe
you're talking about CMA checks only?


thanks,
--
John Hubbard
NVIDIA