Re: [PATCH 3/4] Ethtool: convert get_tso/set_tso calls to hw_features flags

["Michał Mirosław" <mirq-linux@rere.qmqm.pl>]

On Mon, Nov 01, 2010 at 09:25:54PM +0000, Ben Hutchings wrote:
> On Sat, 2010-10-30 at 10:44 +0200, Michał Mirosław wrote:
> [...]
> > @@ -1670,7 +1668,7 @@ struct net_device *nes_netdev_init(struct nes_device *nesdev,
> >  	netdev->type = ARPHRD_ETHER;
> >  	netdev->features = NETIF_F_HIGHDMA;
> >  	netdev->netdev_ops = &nes_netdev_ops;
> > -	netdev->hw_features |= NETIF_F_SG;
> > +	netdev->hw_features |= NETIF_F_SG|NETIF_F_TSO;
> 
> There should be spaces on either side of the '|' operator.

The style varies among drivers' sources. I don't have any preference
whatsoever.

> [...]
> > diff --git a/drivers/net/atlx/atl1.c b/drivers/net/atlx/atl1.c
> > index 9e27bd6..814a06c 100644
> > --- a/drivers/net/atlx/atl1.c
> > +++ b/drivers/net/atlx/atl1.c
> > @@ -2992,6 +2992,7 @@ static int __devinit atl1_probe(struct pci_dev *pdev,
> >  	netdev->watchdog_timeo = 5 * HZ;
> >  
> >  	netdev->hw_features |= NETIF_F_SG;
> > +	netdev->hw_features |= NETIF_F_TSO;
> 
> Why not set both flags in the same statement?  You might as well make
> the drivers consistent in this regard.

Will fix. The compiler should combine those anyway.

> [...]
> > diff --git a/net/core/ethtool.c b/net/core/ethtool.c
> > index 017667c..9b0e598 100644
> > --- a/net/core/ethtool.c
> > +++ b/net/core/ethtool.c
> [...]
> > @@ -1065,10 +1048,13 @@ static int __ethtool_set_sg(struct net_device *dev, u32 data)
> >  {
> >  	int err;
> >  
> > -	if (!data && dev->ethtool_ops->set_tso) {
> > -		err = dev->ethtool_ops->set_tso(dev, 0);
> > -		if (err)
> > -			return err;
> > +	if (!data && (dev->hw_features & NETIF_F_ALL_TSO)) {
> > +		if (dev->ethtool_ops->hw_set_tso) {
> > +			err = dev->ethtool_ops->hw_set_tso(dev, 0);
> > +			if (err < 0)
> > +				return err;
> > +		}
> > +		dev->features &= dev->hw_features & NETIF_F_ALL_TSO;
> 
> Surely this should be:
> 		dev->features &= ~NETIF_F_ALL_TSO;

I originally thought of hw_features as a bitfield of features changable
by ethtool.  So if driver writer wanted to have TSO permanently on but
allow toggling of TSO6 for debug than that would work.  There is actually
a bug in original code: it doesn't clear NETIF_F_TSO if there is no set_tso
function, but clears NETIF_F_SG anyway (the bug is preserved).  This is
a material for successive patches.

BTW, why is scatter-gather needed for TSO and TX checksumming?  I see
no interdependence on s-g for those offloads.

> [...]
> > @@ -1158,7 +1149,18 @@ static int ethtool_set_tso(struct net_device *dev, char __user *useraddr)
> >  	if (edata.data && !(dev->features & NETIF_F_SG))
> >  		return -EINVAL;
> >  
> > -	return dev->ethtool_ops->set_tso(dev, edata.data);
> > +	if (dev->ethtool_ops->hw_set_tso) {
> > +		int err = dev->ethtool_ops->hw_set_tso(dev, edata.data);
> > +		if (err)
> > +			return min(err, 0);
> [...]
> 
> Again, the odd semantics of a positive value need to be documented.

Of course.

Best Regards,
Michał Mirosław