Re: [PATCH RFC 1/4] vsock/virtio: reduce credit update messages

[Stefano Garzarella]

On Thu, Apr 04, 2019 at 08:15:39PM +0100, Stefan Hajnoczi wrote:
> On Thu, Apr 04, 2019 at 12:58:35PM +0200, Stefano Garzarella wrote:
> > @@ -256,6 +257,7 @@ virtio_transport_stream_do_dequeue(struct vsock_sock *vsk,
> >  	struct virtio_vsock_sock *vvs = vsk->trans;
> >  	struct virtio_vsock_pkt *pkt;
> >  	size_t bytes, total = 0;
> > +	s64 free_space;
> 
> Why s64?  buf_alloc, fwd_cnt, and last_fwd_cnt are all u32.  fwd_cnt -
> last_fwd_cnt <= buf_alloc is always true.
> 

Right, I'll use a u32 for free_space!
Is is a leftover because initially I implemented something like
virtio_transport_has_space().

> >  	int err = -EFAULT;
> >  
> >  	spin_lock_bh(&vvs->rx_lock);
> > @@ -288,9 +290,15 @@ virtio_transport_stream_do_dequeue(struct vsock_sock *vsk,
> >  	}
> >  	spin_unlock_bh(&vvs->rx_lock);
> >  
> > -	/* Send a credit pkt to peer */
> > -	virtio_transport_send_credit_update(vsk, VIRTIO_VSOCK_TYPE_STREAM,
> > -					    NULL);
> > +	/* We send a credit update only when the space available seen
> > +	 * by the transmitter is less than VIRTIO_VSOCK_MAX_PKT_BUF_SIZE
> > +	 */
> > +	free_space = vvs->buf_alloc - (vvs->fwd_cnt - vvs->last_fwd_cnt);
> 
> Locking?  These fields should be accessed under tx_lock.
> 

Yes, we need a lock, but looking in the code, vvs->fwd_cnd is written
taking rx_lock (virtio_transport_dec_rx_pkt) and it is read with the
tx_lock (virtio_transport_inc_tx_pkt).

Maybe we should use another spin_lock shared between RX and TX for those
fields or use atomic variables.

What do you suggest?

Thanks,
Stefano