Re: [PATCH RFC net-next v4 6/8] virtio/vsock: support dgrams

[Bobby Eshleman <bobbyeshleman@gmail.com>]

On Mon, Jun 26, 2023 at 05:03:15PM +0200, Stefano Garzarella wrote:
> On Fri, Jun 23, 2023 at 04:37:55AM +0000, Bobby Eshleman wrote:
> > On Thu, Jun 22, 2023 at 06:09:12PM +0200, Stefano Garzarella wrote:
> > > On Sun, Jun 11, 2023 at 11:49:02PM +0300, Arseniy Krasnov wrote:
> > > > Hello Bobby!
> > > >
> > > > On 10.06.2023 03:58, Bobby Eshleman wrote:
> > > > > This commit adds support for datagrams over virtio/vsock.
> > > > >
> > > > > Message boundaries are preserved on a per-skb and per-vq entry basis.
> > > >
> > > > I'm a little bit confused about the following case: let vhost sends 4097 bytes
> > > > datagram to the guest. Guest uses 4096 RX buffers in it's virtio queue, each
> > > > buffer has attached empty skb to it. Vhost places first 4096 bytes to the first
> > > > buffer of guests RX queue, and 1 last byte to the second buffer. Now IIUC guest
> > > > has two skb in it rx queue, and user in guest wants to read data - does it read
> > > > 4097 bytes, while guest has two skb - 4096 bytes and 1 bytes? In seqpacket there is
> > > > special marker in header which shows where message ends, and how it works here?
> > > 
> > > I think the main difference is that DGRAM is not connection-oriented, so
> > > we don't have a stream and we can't split the packet into 2 (maybe we
> > > could, but we have no guarantee that the second one for example will be
> > > not discarded because there is no space).
> > > 
> > > So I think it is acceptable as a restriction to keep it simple.
> > > 
> > > My only doubt is, should we make the RX buffer size configurable,
> > > instead of always using 4k?
> > > 
> > I think that is a really good idea. What mechanism do you imagine?
> 
> Some parameter in sysfs?
> 

I comment more on this below.

> > 
> > For sendmsg() with buflen > VQ_BUF_SIZE, I think I'd like -ENOBUFS
> 
> For the guest it should be easy since it allocates the buffers, but for
> the host?
> 
> Maybe we should add a field in the configuration space that reports some
> sort of MTU.
> 
> Something in addition to what Laura had proposed here:
> https://markmail.org/message/ymhz7wllutdxji3e
> 

That sounds good to me.

IIUC vhost exposes the limit via the configuration space, and the guest
can configure the RX buffer size up to that limit via sysfs?

> > returned even though it is uncharacteristic of Linux sockets.
> > Alternatively, silently dropping is okay... but seems needlessly
> > unhelpful.
> 
> UDP takes advantage of IP fragmentation, right?
> But what happens if a fragment is lost?
> 
> We should try to behave in a similar way.
> 

AFAICT in UDP the sending socket will see EHOSTUNREACH on its error
queue and the packet will be dropped.

For more details:
- the IP defragmenter will emit an ICMP_TIME_EXCEEDED from ip_expire()
  if the fragment queue is not completed within time.
- Upon seeing ICMP_TIME_EXCEEDED, the sending stack will then add
  EHOSTUNREACH to the socket's error queue, as seen in __udp4_lib_err().

Given some updated man pages I think enqueuing EHOSTUNREACH is okay for
vsock too. This also reserves ENOBUFS/ENOMEM only for shortage on local
buffers / mem.

What do you think?

Thanks,
Bobby