Re: [PATCH 0/3] mvneta: software TSO implementation

[Willy Tarreau <w@1wt.eu>]

On Fri, Apr 11, 2014 at 04:08:27PM -0300, Ezequiel Garcia wrote:
> On Apr 11, Willy Tarreau wrote:
> > On Thu, Apr 10, 2014 at 06:02:30PM -0700, Eric Dumazet wrote:
> > > On Thu, 2014-04-10 at 20:51 -0400, David Miller wrote:
> > > > This is not really driver specific code.
> > > > 
> > > > Please create an appropriate abstraction, and put this code in a
> > > > common place so that any driver can do this.
> > > > 
> > > > Thanks.
> > > 
> > > Note a prior implementation existed in
> > > 
> > > drivers/net/ethernet/tile/tilegx.c
> > > 
> > > Note sure how it can be properly factorized, but worth trying .
> > 
> > I also tried to find how to do this for mv643xx_eth a few years ago
> > 
> [..]
> 
> Can you share *any* code about this attempt on mv643xx_eth? It would be
> helpful to see other TSO cases, to have more examples in order to come
> up with a proper abstraction.

Unfortunately no. I remember that I picked the equivalent code from
mv_eth in the kirkwood's LSP and tried to adapt it to the mainline
driver, but unfortunately by then I didn't exactly understand well
what the driver was doing despite Eric explaining me a few things
about GSO and segmentation. Too many things were obscure to me and
I had to give up. The only vague memories I have are that this NIC
doesn't support Tx csum on VLANs nor IPv6. That's about all, I regret.

By the way, I could be wrong but I also have some memories about this
in the myri10ge driver. I recall something like it using software TSO
for VLANs or IPv6 or a combination of both, I don't exactly remember.
It's also possible that it depends on the firmware versions. Anyway the
driver is clean and well commented, it's easier to read than average.

Hoping this helps,
Willy