Re: [PATCH 1/5] migration/multifd: Separate compression ops from non-compression

[Peter Xu <peterx@redhat.com>]

On Fri, Jan 26, 2024 at 07:19:39PM -0300, Fabiano Rosas wrote:
> +static MultiFDMethods multifd_socket_ops = {
> +    .send_setup = multifd_socket_send_setup,
> +    .send_cleanup = multifd_socket_send_cleanup,
> +    .send_prepare = multifd_socket_send_prepare,

Here it's named with "socket", however not all socket-based multifd
migrations will go into this route, e.g., when zstd compression enabled it
will not go via this route, even if zstd also uses sockets as transport.
From that pov, this may be slightly confusing.  Maybe it suites more to be
called "socket_plain" / "socket_no_comp"?

One step back, I had a feeling that the current proposal tried to provide a
single ->ops to cover a model where we may need more than one layer of
abstraction.

Since it might be helpful to allow multifd send arbitrary data (e.g. for
VFIO?  Avihai might have an answer there..), I'll try to even consider that
into the picture.

Let's consider the ultimate goal of multifd, where the simplest model could
look like this in my mind (I'm only discussing sender side, but it'll be
similar on recv side):

               prepare()           send()
  Input   ----------------> IOVs ------------> iochannels

[I used prepare/send, but please think them as generic terms, not 100%
 aligned with what we have with existing multifd_ops, or what you proposed
 later]

Here what are sure, IMHO, is:

  - We always can have some input data to dump; I didn't use "guest pages"
    just to say we may allow arbitrary data.  For any multifd user that
    would like to dump arbitrary data, they can already provide IOVs, so
    here input can be either "MultiFDPages_t" or "IOVs".

  - We may always want to have IOVs to represent the buffers at some point,
    no matter what the input it

  - We always flush the IOVs to iochannels; basically I want to say we can
    always assume the last layer is connecting to QIOChannel APIs, while I
    don't think there's outliers here so far, even if the send() may differ.

Then _maybe_ it's clearer that we can have two layers of OPs?

  - prepare(): it tells how the "input" will be converted into a scatter
    gatter list of buffers.  All compression methods fall into this afaiu.
    This has _nothing_ to do on how the buffers will be sent.  For
    arbitrary-typed input, this can already be a no-op since the IOVs
    provided can already be passed over to send().

  - send(): how to dump the IOVs to the iochannels.  AFAIU this is motly
    only useful for fixed-ram migrations.

Would this be clearer, rather than keep using a single multifd_ops?

-- 
Peter Xu