[PATCH v1 04/12] mm/rmap: implement make_device_exclusive() using folio_walk instead of rmap walk

[David Hildenbrand]

On 30.01.25 23:31, Alistair Popple wrote:
> On Thu, Jan 30, 2025 at 10:24:37AM +0100, David Hildenbrand wrote:
>> On 30.01.25 10:01, David Hildenbrand wrote:
>>> On 30.01.25 07:11, Alistair Popple wrote:
>>>> On Wed, Jan 29, 2025 at 12:54:02PM +0100, David Hildenbrand wrote:
>>>>> We require a writable PTE and only support anonymous folio: we can only
>>>>> have exactly one PTE pointing at that page, which we can just lookup
>>>>> using a folio walk, avoiding the rmap walk and the anon VMA lock.
>>>>>
>>>>> So let's stop doing an rmap walk and perform a folio walk instead, so we
>>>>> can easily just modify a single PTE and avoid relying on rmap/mapcounts.
>>>>>
>>>>> We now effectively work on a single PTE instead of multiple PTEs of
>>>>> a large folio, allowing for conversion of individual PTEs from
>>>>> non-exclusive to device-exclusive -- note that the other way always
>>>>> worked on single PTEs.
>>>>>
>>>>> We can drop the MMU_NOTIFY_EXCLUSIVE MMU notifier call and document why
>>>>> that is not required: GUP will already take care of the
>>>>> MMU_NOTIFY_EXCLUSIVE call if required (there is already a device-exclusive
>>>>> entry) when not finding a present PTE and having to trigger a fault and
>>>>> ending up in remove_device_exclusive_entry().
>>>>
>>>> I will have to look at this a bit more closely tomorrow but this doesn't seem
>>>> right to me. We may be transitioning from a present PTE (ie. a writable
>>>> anonymous mapping) to a non-present PTE (ie. a device-exclusive entry) and
>>>> therefore any secondary processors (eg. other GPUs, iommus, etc.) will need to
>>>> update their copies of the PTE. So I think the notifier call is needed.
>>>
>>> Then it is all very confusing:
> 
> Can't argue with that in hindsight :-)
> 
>>> "MMU_NOTIFY_EXCLUSIVE: to signal a device driver that the device will no
>>> longer have exclusive access to the page."
>>
>> So the second sentence actually describes the other condition. Likely we
>> should make that clearer:
>>
>> --- a/include/linux/mmu_notifier.h
>> +++ b/include/linux/mmu_notifier.h
>> @@ -43,10 +43,11 @@ struct mmu_interval_notifier;
>>    * a device driver to possibly ignore the invalidation if the
>>    * owner field matches the driver's device private pgmap owner.
>>    *
>> - * @MMU_NOTIFY_EXCLUSIVE: to signal a device driver that the device will no
>> - * longer have exclusive access to the page. When sent during creation of an
>> - * exclusive range the owner will be initialised to the value provided by the
>> - * caller of make_device_exclusive(), otherwise the owner will be NULL.
>> + * @MMU_NOTIFY_EXCLUSIVE: (1) to signal a device driver that the device will no
>> + * longer have exclusive access to the page; and (2) to signal that a page will
>> + * be made exclusive to a device. During (1), the owner will be NULL, during
>> + * (2), the owner will be initialised to the value provided by the caller of
>> + * make_device_exclusive().
> 
> Yes, I think that makes things clearer. Logically these are really two different
> events though - I guess I didn't want to add another one at the time but I
> wonder if we should just make them separate events rather than overloading them?

I had the same thought and then I wondered: can't we simply use 
MMU_NOTIFY_CLEAR for the exclusive->ordinary path?

I mean, it's essentially a zap+flush followed by a re-insertion of the 
PFN swap entry. Similar to page migration ...

-- 
Cheers,

David / dhildenb