Re: [RFC PATCH v8 00/19] virtio/vsock: introduce SOCK_SEQPACKET support

[Stefano Garzarella]

On Thu, Apr 22, 2021 at 01:29:54PM +0300, Arseny Krasnov wrote:
> On 22.04.2021 13:02, Stefano Garzarella wrote:
> > On Thu, Apr 22, 2021 at 12:40:17PM +0300, Arseny Krasnov wrote:
> > > On 22.04.2021 11:46, Stefano Garzarella wrote:
> > > > On Wed, Apr 21, 2021 at 06:06:28PM +0300, Arseny Krasnov wrote:
> > > > > Thank You, i'll prepare next version. Main question is: does this
> > > > > approach(no SEQ_BEGIN, SEQ_END, 'msg_len' and 'msg_id') considered
> > > > > good? In this case it will be easier to prepare final version, because
> > > > > is smaller and more simple than previous logic. Also patch to spec
> > > > > will be smaller.
> > > > Yes, it's definitely much better than before.
> > > >
> > > > The only problem I see is that we add some overhead per fragment
> > > > (header). We could solve that with the mergeable buffers that Jiang is
> > > > considering for DGRAM.
> > > If we are talking about receive, i think, i can reuse merge logic for
> > Yep, for TX the guest can potentially enqueue a big buffer.
> > Maybe it's still worth keeping a maximum size and fragmenting as we do
> > now.
> >
> > > stream sockets, the only difference is that buffers are mergeable
> > > until previous EOR(e.g. previous message) bit is found in rx queue.
> > I got a little lost.
> > Can you elaborate more?
>
> I'm talking about 'virtio_transport_recv_enqueue()': it tries to copy
>
> data of new packet to buffer of tail packet in rx queue. In case of
>
> SEQPACKET i can reuse it, just adding logic that check EOR bit of
>
> tail packet.

This might be a good idea.
It doesn't save us the transmitted header though, but at least it saves 
us from queuing it.
Even if with SEQPACKET I don't expect small packets, since it's the 
driver that divides them and I think it does everything to use the 
maximum available.

Instead the mergeable buffers I was referring to are based on the 
virito-net feature VIRTIO_NET_F_MRG_RXBUF.
Jiang is investigating whether we can reuse them for DGRAM.

Thanks,
Stefano