Re: [PATCH net-next] sunvnet: improve error handling when a remote crashes

[Sowmini Varadhan <sowmini.varadhan@oracle.com>]

> @@ -934,36 +933,36 @@ static struct sk_buff *vnet_clean_tx_ring(struct vnet_port *port,
>  
>  	*pending = 0;
>  
> -	txi = dr->prod-1;
> -	if (txi < 0)
> -		txi = VNET_TX_RING_SIZE-1;
> -
> +	txi = dr->prod;

As I understand it, this starts at dr->prod and goes through all
descriptors, cleaning up !READY descriptors as it goes around.

I think you'll have a higher reclaim rate for finding !READY if you
started at dr->cons instead: dr->cons is the one that was last ACK'ed,
and that ack would only have been sent after the peer had marked the 
descriptor as DONE. (consumer would have had a chance to read more
descriptors, by the time the tx-reclaim loop goes around) 

> +		if (port->tx_bufs[txi].skb) {
> +			if (d->hdr.state != VIO_DESC_DONE)
> +				pr_warn("invalid ring buffer state %d\n",
> +					d->hdr.state);

I would even suggest skipping the pr_warn (maybe make it a viodbg
instead) as it might alarm the end-user (who cannot really do 
anything about it other than call us anyway :-)).

>  			   dr->cookies, dr->ncookies);
> +	if (active_freed)
> +		pr_warn("%s: active transmit buffers freed for remote %pM\n",
> +			dev->name, port->raddr);

Same comment as above.

In general, I think we need some sysfs/ethtool  bean-counters/statistics
for sunvnet, to keep track of this sort of thing efficiently in 
a production env without triggering red-herrings calls.

--Sowmini