Re: [PATCH v6 0/2] Ethernet drivers for WIZnet chips

[Eric Dumazet <eric.dumazet@gmail.com>]

On Tue, 2012-04-03 at 18:15 +0600, Mike Sinkovsky wrote:
> 03.04.2012 17:39, Eric Dumazet написал:
> > On Tue, 2012-04-03 at 16:58 +0600, Mike Sinkovsky wrote:
> >> Based on original driver from chip manufacturer, but nearly full rewite.
> >> Tested and used in production with Blackfin BF531 embedded processor.
> >>
> >> Signed-off-by: Mike Sinkovsky<msink@permonline.ru>
> >> ---
> >> v6:
> >>   - remove (S0_TX_FSR<  skb->len) check in TX handler, it doesn't work
> >>     anyway. Chip can transmit only one frame with MTU 1500 at a time,
> >>     and tx buffer size is bigger.
> >
> > So what happens if XXX frames are given to start_xmit() in a flood ?
> >
> > You removed any flow control, how can this work ?
> >
> > Device has an infinite queue ?
> 
> As I understand from datasheet, device doesn't have tx queue at all.
> It have tx buffer, processor must save transmitted frame to it, and then 
> save SEND command to command register. When transmission completed, 
> SENDOK bit in interrupt register will be set to '1', and interrupt 
> handler will be called.
> 
> So, according to datasheet, driver must stop queue in start_xmit() 
> routine, and wake in interrupt handler of SENDOK bit.
> I tried this, and it basically works, but SOMETIMES, very rare, I see 
> tx_timeout().
> 
> And without any flow control - driver works perfectly.
> Weird, don't know why.

Really this should be fixed, since your driver makes qdisc flow control
impossible (unless adding a rate limiter like HTB/CBQ)

You probably had a race on your xmit routine, and interrupt routine.

You need proper synchronization : A spinlock to make this easy, or a
smart barrier game.