[PATCH v6 01/10] mm: add zone device coherent type memory support

[Felix Kuehling]

On 2022-02-15 07:15, David Hildenbrand wrote:
> On 11.02.22 17:56, Jason Gunthorpe wrote:
>> On Fri, Feb 11, 2022 at 05:49:08PM +0100, David Hildenbrand wrote:
>>> On 11.02.22 17:45, Jason Gunthorpe wrote:
>>>> On Fri, Feb 11, 2022 at 05:15:25PM +0100, David Hildenbrand wrote:
>>>>
>>>>> ... I'm pretty sure we cannot FOLL_PIN DEVICE_PRIVATE pages
>>>> Currently the only way to get a DEVICE_PRIVATE page out of the page
>>>> tables is via hmm_range_fault() and that doesn't manipulate any ref
>>>> counts.
>>> Thanks for clarifying Jason! ... and AFAIU, device exclusive entries are
>>> essentially just pointers at ordinary PageAnon() pages. So with DEVICE
>>> COHERENT we'll have the first PageAnon() ZONE_DEVICE pages mapped as
>>> present in the page tables where GUP could FOLL_PIN them.
>> This is my understanding
>>
>> Though you probably understand what PageAnon means alot better than I
>> do.. I wonder if it really makes sense to talk about that together
>> with ZONE_DEVICE which has alot in common with filesystem originated
>> pages too.
> For me, PageAnon() means that modifications are visible only to the
> modifying process. On actual CoW, the underlying page will get replaced
> -- in the world of DEVICE_COHERENT that would mean that once you write
> to a DEVICE_COHERENT you could suddenly have a !DEVICE_COHERENT page.
>
> PageAnon() pages don't have a mapping, thus they can only be found in
> MAP_ANON VMAs or in MAP_SHARED VMAs with MAP_PRIVATE. They can only be
> found via a page table, and not looked up via the page cache (excluding
> the swap cache).
>
> So if we have PageAnon() pages on ZONE_DEVICE, they generally have the
> exact same semantics as !ZONE_DEVICE pages, but the way they "appear" in
> the page tables the allocation/freeing path differs -- I guess :)
>
> ... and as we want pinning semantics to be different we have to touch GUP.
>
>> I'm not sure what AMDs plan is here, is there an expecation that a GPU
>> driver will somehow stuff these pages into an existing anonymous
>> memory VMA or do they always come from a driver originated VMA?
> My understanding is that a driver can just decide to replace "ordinary"
> PageAnon() pages e.g., in a MAP_ANON VMA by these pages. Hopefully AMD
> can clarify.

Yes. DEVICE_COHERENT pages are very similar to DEVICE_PRIVATE. They are 
in normal anonymous VMAs (e.g. the application called mmap(..., 
MAP_ANONYMOUS, ...)). You get DEVICE_PRIVATE/COHERENT pages as a result 
of the driver migrating normal anonymous pages into device memory. The 
main difference is, that DEVICE_PRIVATE pages aren't host accessible, 
while DEVICE_COHERENT pages are host-accessible and can be mapped in the 
CPU page table or DMA-mapped by peer devices. That's why GUP needs to 
deal with them.

Regards,
   Felix


>
>