Re: [PATCH v2 2/3] userfaultfd: UFFDIO_REMAP uABI

[David Hildenbrand <david@redhat.com>]

On 04.10.23 01:39, Lokesh Gidra wrote:
> On Tue, Oct 3, 2023 at 11:26 PM Suren Baghdasaryan <surenb@google.com> wrote:
>>
>> On Tue, Oct 3, 2023 at 2:21 PM Peter Xu <peterx@redhat.com> wrote:
>>>
>>> On Tue, Oct 03, 2023 at 11:08:07PM +0200, David Hildenbrand wrote:
>>>> Sorry I have to ask: has this ever been discussed on the list? I don't see
>>>> any pointers. If not, then probably the number of people that know about the
>>>> history can be counted with my two hands and that shouldn't be the basis for
>>>> making decisions.
>>>
>>> For example:
>>>
>>> https://lore.kernel.org/all/1425575884-2574-21-git-send-email-aarcange@redhat.com/

Sorry, I had to process a family NMI the last couple of days.

>>
>> There was another submission in 2019:
>> https://lore.kernel.org/all/cover.1547251023.git.blake.caldwell@colorado.edu/

It would be good to link them in the cover letter and shortly explain 
why that wasn't merged back then (if there was any reason).

>>
>> Though both times it did not generate much discussion. I don't have a
>> strong preference though MOVE sounds more generic to me TBH (it
>> specifies the operation rather than REMAP which hints on how that
>> operation is carried out). But again, I'm fine either way.
> 
> That's a good point. IMHO, if in future we want to have the fallback
> implemented, then MOVE would be a more appropriate name than REMAP.
> 
>> As for UFFDIO_MOVE_ZERO_COPY_ONLY vs UFFDIO_MOVE_MODE_ALLOW_COPY, I
>> find it weird that the default (the most efficient/desired) mode of
>> operation needs a flag. I would prefer to have no flag initially and
>> add UFFDIO_MOVE_MODE_ALLOW_COPY or whatever name is more appropriate
>> when/if we ever need it. Makes sense?
> 
> Agreed!

I agree. One could have UFFDIO_MOVE that is best-effort and documented 
like that, and a to-be-named future extension that always works but 
might be more expensive.


Ideally we'd have an interface that does not expose and/or rely on such 
low-level information and simply always works, but getting that would 
mean that we'd have to implement the fallback immediately ... so I guess 
we'll have to expose a best-effort interface first.

-- 
Cheers,

David / dhildenb