Re: [PATCH v1 2/4] mm/mmu_notifier: drop owner from MMU_NOTIFY_EXCLUSIVE

[David Hildenbrand <david@redhat.com>]

On 30.01.25 14:29, Simona Vetter wrote:
> On Thu, Jan 30, 2025 at 10:28:00AM +0100, David Hildenbrand wrote:
>> On 30.01.25 06:34, Alistair Popple wrote:
>>> Looking at hmm_test I see that doesn't use the sequence counter to ensure
>>> the PTE remains valid whilst it is mapped. I think that is probably wrong, so
>>> apologies if that lead you astray.
>>
>> Yes, the hmm_test does not completely follow the same model the nouveau
>> implementation does; so it might not be completely correct.
> 
> But unrelated but just crossed my mind:
> 
> I guess another crucial difference is that the hw (probably, not sure)
> will restart the fault if we don't repair it to its liking. So the
> hmm-test does need some kind of retry loop too somewhere to match that.

Yes. Especially for the folio lock spinning is a rather suboptimal 
approach. So we likely would want the option to just lock it instead of 
try-locking it. (or getting rid of it entirely :) )

> But might be good to also still land some of the other improvements
> discussed in these threads to make make_device_exclusive a bit more
> reliable instead of relying on busy-looping throug the hw fault handler
> for everything.

Right.

-- 
Cheers,

David / dhildenb