[PATCH] ucc_geth: Rework the TX logic.

[Joakim Tjernlund]

pku.leo at gmail.com wrote on 27/03/2009 11:39:07:
> On Fri, Mar 27, 2009 at 6:23 PM, Joakim Tjernlund
> >> > -               /* We freed a buffer, so now we can restart
> > transmission */
> >> > -               if (netif_queue_stopped(dev))
> >> > -                       netif_wake_queue(dev);
> >> > +               dev_kfree_skb(ugeth->tx_skbuff[txQ][tx_ind]);
> >> > +               ugeth->tx_skbuff[txQ][tx_ind] = NULL;
> >> > +               out_be32(bd2buf(bd), 0); /* Mark it free */
> >>
> >> Now I see why you depend on the buffer to judge if the BD ring is
> >> full.  If you want to do this, make sure it's well documented in 
code,
> >> or others will be confused.  And as Anton already commented, in/out
> >> access is slow.  I think the original way can be fixed if you think
> >> it's broken, and will have better performance.
> >
> > The code could use a better comment, but I really think this method
> > is superior and more robust. The problem is that in/out functions is
> > much slower than it has to be for QE IO memory.
> >
> > The original way is broken and I tired to fix it but it always broke
> > as soon one applied some load.
> 
> The only difference I see between your method and the original code is
> that you update the cleanup progress on every BD.  The original code
> updates progress only after all sent BDs are processed.  You can try
> move ugeth->confBd[txQ] = bd; into the loop then it will be the same
> as your proposed code.

Already tried that+lots of other combinatios. Doesn't work