RE: [PATCH net-next] hyperv: Add support for Virtual Receive SideScaling (vRSS)

[Haiyang Zhang]

> -----Original Message-----
> From: Ben Hutchings [mailto:bhutchings@xxxxxxxxxxxxxx]
> Sent: Thursday, December 19, 2013 12:46 PM
> To: Haiyang Zhang
> Cc: davem@xxxxxxxxxxxxx; netdev@xxxxxxxxxxxxxxx; KY Srinivasan;
> olaf@xxxxxxxxx; jasowang@xxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx;
> driverdev-devel@xxxxxxxxxxxxxxxxxxxxxx
> Subject: Re: [PATCH net-next] hyperv: Add support for Virtual Receive Side
> Scaling (vRSS)
> 
> On Wed, 2013-12-18 at 14:21 -0800, Haiyang Zhang wrote:
> > This feature allows multiple channels to be used by each virtual NIC.
> > It is available on Hyper-V host 2012 R2.
> >  	} else if (ret == -EAGAIN) {
> > -		netif_stop_queue(ndev);
> > +		netif_tx_stop_all_queues(ndev);
> >  		if (atomic_read(&net_device->num_outstanding_sends) < 1) {
> > -			netif_wake_queue(ndev);
> > +			netif_tx_wake_all_queues(ndev);
> >  			ret = -ENOSPC;
> >  		}
> >  	} else {
> 
> This doesn't makes any sense to me.  How can you safely share the same
> channels between all TX queues?
> 
> I think you need to associate TX queues and channels 1-1.  If you are required to
> map packets to TX queues using the Toeplitz hash, you should implement
> ndo_select_queue and do the mapping there.  Then in
> netvsc_send() you would use the queue number from the skb to find which
> channel to use and which queue may need to be stopped/woken.
> 
> [...]
> > --- a/drivers/net/hyperv/netvsc_drv.c
> > +++ b/drivers/net/hyperv/netvsc_drv.c
> [...]
> > +/* Toeplitz hash function
> > + * data: network byte order
> > + * return: host byte order
> > + */
> 
> This looks incredibly slow.  I've seen software implementations that are likely to
> be more efficient, e.g.
> <http://thread.gmane.org/gmane.linux.network/284612/>
> 
> [...]
> > int netvsc_probe(struct hv_device *dev,
> >  		return ret;
> >  	}
> >  	memcpy(net->dev_addr, device_info.mac_adr, ETH_ALEN);
> > +	nvdev = hv_get_drvdata(dev);
> > +	rtnl_lock();
> > +	netif_set_real_num_tx_queues(net, nvdev->num_chn);
> > +	netif_set_real_num_rx_queues(net, nvdev->num_chn);
> [...]
> 
> These functions can fail if called after registering the net device, so you should
> either call them with the final values earlier or handle failure here.
> 
> Also, I notice that dev_addr is only set after registering; that should be fixed.
> 

Thank you for the suggestions! I will re-write the send queue selection, enhance
the hash calculation, also fix the initialization sequence.

Thanks,
- Haiyang

èº{.nÇ+‰·Ÿ®‰­†+%ŠËlzwm…ébëæìr¸›zX§»®w¥Š{ayºÊÚë,j­¢f£¢·hš‹àz¹®w¥¢¸¢·¦j:+v‰¨ŠwèjØm¶Ÿÿ¾«‘êçzZ+ƒùšŽŠÝj"ú!¶iO•æ¬z·švØ^¶m§ÿðÃnÆàþY&—