Re: [PATCH 0/5] migration/multifd: Prerequisite cleanups for ongoing work

[Fabiano Rosas]

Peter Xu <peterx@redhat.com> writes:

> On Mon, Jan 29, 2024 at 09:51:06AM -0300, Fabiano Rosas wrote:
>> Peter Xu <peterx@redhat.com> writes:
>> 
>> > On Mon, Jan 29, 2024 at 01:41:01AM +0000, Liu, Yuan1 wrote:
>> >> Because this change has an impact on the previous live migration 
>> >> With IAA Patch, does the submission of the next version needs 
>> >> to be submitted based on this change?
>> >
>> > I'd say hold off a little while until we're more certain on the planned
>> > interface changes, to avoid you rebase your code back and forth; unless
>> > you're pretty confident that this will be the right approach.
>> >
>> > I apologize on not having looked at any of the QAT/IAA compression / zero
>> > detection series posted on the list; I do plan to read them very soon too
>> > after Fabiano.  So I may not have a complete full picture here yet, please
>> > bare with me.
>> >
>> > If this series is trying to provide a base ground for all the efforts,
>> > it'll be great if we can thoroughly discuss here and settle an approach
>> > soon that will satisfy everyone.
>> 
>> Just a summary if it helps:
>> 
>> For compression work (IAA/QPL, QAT) the discussion is around having a
>> new "compression acceleration" option that enables the accelerators and
>> is complementary to the existing zlib compression method. We'd choose
>> those automatically based on availability and we'd make HW accelerated
>> compression produce a stream that is compatible with QEMU's zlib stream
>> so we could migrate between solutions.
>> 
>> For zero page work and zero page acceleration (DSA), the question is how
>> to fit zero page detection into multifd and whether we need a new hook
>> multifd_ops->zero_page_detect() (or similar) to allow client code to
>> provide it's own zero page detection methods. My worry here is that
>> teaching multifd to recognize zero pages is one more coupling to the
>> "pages" data type. Ideallly we'd find a way to include that operation as
>> a prepare() responsibility and the client code would deal with it.
>
> Thanks Fabiano.
>
> Since I'm preparing the old series to post for some fundamental cleanups
> around multifd, and when I'm looking around the code, I noticed that
> _maybe_ it'll also be eaiser to apply such a series if we can cleanup more
> things then move towards a clean base to add more accelerators.
>
> I agree many ideas in your this series, but I may address it slightly
> different (e.g., I want to avoid send(), but you can consider that in the
> fixed-ram series instead), also it'll be after some other cleanup I plan to
> give a stab at which is not yet covered in this series.  I hope I can add
> your "Co-developed-by" in some of the patches there.  If you haven't spend
> more time on new version of this series, please wait 1-2 days so I can post
> my thoughts.

Sure, go ahead.