Re: Regression in net-2.6.24?

[David Miller <davem@davemloft.net>]

From: "Michael Chan" <mchan@broadcom.com>
Date: Thu, 11 Oct 2007 20:17:16 -0700

> On Thu, 2007-10-11 at 18:14 -0700, David Miller wrote:
> > +       while (1) {
> > +               work_done = tg3_poll_work(tp, work_done, budget);
> > +
> > +               if (unlikely(tp->tg3_flags & TG3_FLAG_TX_RECOVERY_PENDING))
> > +                       goto tx_recovery;
> > +
> > +               if (unlikely(work_done >= budget))
> > +                       break;
> > +
> > +               if (likely(!tg3_has_work(tp))) {
> > +                       struct tg3_hw_status *sblk = tp->hw_status;
> > +
> 
> --> new status block DMA
> 
> > +                       if (tp->tg3_flags & TG3_FLAG_TAGGED_STATUS) {
> > +                               tp->last_tag = sblk->status_tag;
> > +                               rmb();
> > +                       } else
> > +                               sblk->status &= ~SD_STATUS_UPDATED;
> 
> We need to read the sblk->status_tag before calling tg3_has_work().  If
> a new status block DMA happens in between (shown above), tp->last_tag
> will get the new tag and we will end up acknowledging work that we
> haven't processed.

Hmmm, the old code didn't do that and seemingly has the same
problem.  Also, if you look at the before-patch code and think
about what it does if we ->poll() multiple times for a single
interrupt the side-effects are essentially the same.

What's the crucial difference?

> I'll go over this some more tonight and will send a patch to refine it.

Thanks Michael.