Re: [PATCH] net: sky2: convert to hw_features

[Stephen Hemminger <shemminger@linux-foundation.org>]

On Mon, 11 Apr 2011 19:58:23 +0200
Michał Mirosław <mirq-linux@rere.qmqm.pl> wrote:

> On Mon, Apr 11, 2011 at 07:58:29AM -0700, Stephen Hemminger wrote:
> > On Mon, 11 Apr 2011 02:51:00 +0200
> > Michał Mirosław <mirq-linux@rere.qmqm.pl> wrote:
> > 
> > > On Sun, Apr 10, 2011 at 11:53:02AM -0700, Stephen Hemminger wrote:
> > > > On Sun, 10 Apr 2011 15:13:21 +0200 (CEST)
> > > > Michał Mirosław <mirq-linux@rere.qmqm.pl> wrote:
> > > > > Caveats:
> > > > >  - driver modifies vlan_features on HW VLAN TX changes
> > > > >  - broken RX checksum will be reenabled on features change
> > > > To be more precise. This is acceptable if and only if all cases
> > > > where features are disabled in response to MTU and chip versions
> > > > are exactly the same. We don't want to let some user stumble upon
> > > > cases where hardware features don't work in their configuration.
> > > I was referring to the unlikely case detected by sky2_rx_checksum().
> > > Before this conversion, user could reenable the feature using ethtool.
> > > The change is that now, in this case, it's reenabled also when other
> > > features are changed (i.e. whenever netdev_update_features() gets called).
> > 
> > Ok.
> > 
> > It does expose a pre-existing issue. If this logic trips (and I have
> > gotten reports of it happening), then the GRO will not get disabled.
> > Probably need to mask of GRO as well, since GRO depends on RXCSUM.
> 
> This is not a problem, since the code checks skb->ip_summed for protocols
> that implement GRO. See tcp4_gro_receive() for example. This might be
> a slight hit in performance, though.

The performance hit is expected, just thought that GRO depended on rx checksum.