Peter Xu <pet...@redhat.com> writes:
> On Mon, Jan 29, 2024 at 09:42:24AM -0300, Fabiano Rosas wrote:
>> Peter Xu <pet...@redhat.com> writes:
>>
>> > 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".
>>
>> Or anything else, since the client code also has control over send(),
>> no? So it could give multifd a pointer to some memory and then use
>> send() to do whatever it wants with it. Multifd is just providing worker
>> threads and "scheduling".
>
> IOVs contain the case of one single buffer, where n_iovs==1. Here I
> mentioned IOVs explicitly because I want to make it part of the protocol so
> that the interface might be clearer, on what is not changing, and what can
> change for a multifd client.
Got it. I agree.
>>
>> Also note that multifd clients currently _do not_ provide IOVs. They
>> merely provide data to multifd (p->pages) and then convert that data
>> into IOVs at prepare(). This is different, because multifd currently
>> holds that p->pages (and turns that into p->normal), which means the
>> client code does not need to store the data across iterations (in the
>> case of RAM which is iterative).
>
> They provide? AFAIU that's exactly MultiFDSendParams.iov as of now, while
> iov_nums is the length.
Before that, the ram code needs to pass in the p->pages->offset array
first. Then, that gets put into p->normal. Then, that gets put into
p->iov at prepare(). So it's not a simple "fill p->iov and pass it to
multifd".
Hmm, could we just replace multifd_send_state->pages with a
multifd_send_state->iov? I don't really understand why do we need to
carry that pages->offset around.
>>
>> >
>> > - 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?
>>
>> Sorry, I don't see how what you describe is any different than what we
>> have. And I don't see how any of this would mean more than one
>> multifd_ops. We already have multifd_ops->prepare() and
>> multifd_ops->send(). What am I missing?
>
> I meant instead of having a single MultiFDMethods, we can have
> MultifdPrepareOps and MultifdSendOps separately.
>
> Now with single MultiFDMethods, it must always provide e.g. both prepare()
> and send() in one set of OPs for one use case. What I wanted to say is
> maybe it is cleaner we split it into two OPs, then all the socket-based
> scenarios can already stick with the same send() method, even though they
> can prepare() differently.
Hmm, so zlib/zstd implement all ops except for the send one. And
socket_plain and file implement all prepare hooks plus the send. So we'd
have sort of a data handling layer and a transport layer. I'll see how
it looks.
>
> IOW, for this base patchset to pave way for compression accelerators, IIUC
> we don't need a send() yet so far? Should they still work pretty well with
> qio_channel_writev_full_all() with proper touchups on p->write_flags just
> for zero copy purposes?
Yes. The point here is to just give everyone a heads-up so we avoid
changing the code in incompatible ways.
>
> I'll have a read again to your previous multifd-packet-cleanups branch I
> guess. but this series definitely doesn't apply there already.
multifd-packet-cleanups attempts to replace MultiFDPages_t with a
generic data structure. That's a separate issue.
