Re: [PATCH v4 4/5] vhost/vsock: split packets to send using multiple buffers

[Stefano Garzarella]

On Fri, Jul 19, 2019 at 04:21:52PM +0800, Jason Wang wrote:
> 
> On 2019/7/19 äå4:08, Stefano Garzarella wrote:
> > On Thu, Jul 18, 2019 at 07:35:46AM -0400, Michael S. Tsirkin wrote:
> > > On Thu, Jul 18, 2019 at 11:37:30AM +0200, Stefano Garzarella wrote:
> > > > On Thu, Jul 18, 2019 at 10:13 AM Michael S. Tsirkin<mst@xxxxxxxxxx>  wrote:
> > > > > On Thu, Jul 18, 2019 at 09:50:14AM +0200, Stefano Garzarella wrote:
> > > > > > On Wed, Jul 17, 2019 at 4:55 PM Michael S. Tsirkin<mst@xxxxxxxxxx>  wrote:
> > > > > > > On Wed, Jul 17, 2019 at 01:30:29PM +0200, Stefano Garzarella wrote:
> > > > > > > > If the packets to sent to the guest are bigger than the buffer
> > > > > > > > available, we can split them, using multiple buffers and fixing
> > > > > > > > the length in the packet header.
> > > > > > > > This is safe since virtio-vsock supports only stream sockets.
> > > > > > > > 
> > > > > > > > Signed-off-by: Stefano Garzarella<sgarzare@xxxxxxxxxx>
> > > > > > > So how does it work right now? If an app
> > > > > > > does sendmsg with a 64K buffer and the other
> > > > > > > side publishes 4K buffers - does it just stall?
> > > > > > Before this series, the 64K (or bigger) user messages was split in 4K packets
> > > > > > (fixed in the code) and queued in an internal list for the TX worker.
> > > > > > 
> > > > > > After this series, we will queue up to 64K packets and then it will be split in
> > > > > > the TX worker, depending on the size of the buffers available in the
> > > > > > vring. (The idea was to allow EWMA or a configuration of the buffers size, but
> > > > > > for now we postponed it)
> > > > > Got it. Using workers for xmit is IMHO a bad idea btw.
> > > > > Why is it done like this?
> > > > Honestly, I don't know the exact reasons for this design, but I suppose
> > > > that the idea was to have only one worker that uses the vring, and
> > > > multiple user threads that enqueue packets in the list.
> > > > This can simplify the code and we can put the user threads to sleep if
> > > > we don't have "credit" available (this means that the receiver doesn't
> > > > have space to receive the packet).
> > > I think you mean the reverse: even without credits you can copy from
> > > user and queue up data, then process it without waking up the user
> > > thread.
> > I checked the code better, but it doesn't seem to do that.
> > The .sendmsg callback of af_vsock, check if the transport has space
> > (virtio-vsock transport returns the credit available). If there is no
> > space, it put the thread to sleep on the 'sk_sleep(sk)' wait_queue.
> > 
> > When the transport receives an update of credit available on the other
> > peer, it calls 'sk->sk_write_space(sk)' that wakes up the thread
> > sleeping, that will queue the new packet.
> > 
> > So, in the current implementation, the TX worker doesn't check the
> > credit available, it only sends the packets.
> > 
> > > Does it help though? It certainly adds up work outside of
> > > user thread context which means it's not accounted for
> > > correctly.
> > I can try to xmit the packet directly in the user thread context, to see
> > the improvements.
> 
> 
> It will then looks more like what virtio-net (and other networking device)
> did.

I'll try ASAP, the changes should not be too complicated... I hope :)

> 
> 
> > 
> > > Maybe we want more VQs. Would help improve parallelism. The question
> > > would then become how to map sockets to VQs. With a simple hash
> > > it's easy to create collisions ...
> > Yes, more VQs can help but the map question is not simple to answer.
> > Maybe we can do an hash on the (cid, port) or do some kind of estimation
> > of queue utilization and try to balance.
> > Should the mapping be unique?
> 
> 
> It sounds to me you want some kind of fair queuing? We've already had
> several qdiscs that do this.

Thanks for pointing it out!

> 
> So if we use the kernel networking xmit path, all those issues could be
> addressed.

One more point to AF_VSOCK + net-stack, but we have to evaluate possible
drawbacks in using the net-stack. (e.g. more latency due to the complexity
of the net-stack?)

Thanks,
Stefano