Re: [PATCH] Re: Q: natsemi.c spinlocks

[Manfred Spraul]

Donald Becker wrote:
> 
> > > However, natsemi.c's spinlock needs to be retained, and
> > > extended into start_tx(), because this driver has
> > > a race which has cropped up in a few others:
> > >         ...
> > >         if (np->cur_tx - np->dirty_tx >= TX_QUEUE_LEN - 1) {
> > >         /* WINDOW HERE */
> > >                 np->tx_full = 1;
> > >                 netif_stop_queue(dev);
> > >         }
> > > If the ring is currently full and an interrupt comes in
> > > at the indicated window and reaps ALL the packets in the
> > > ring, the driver ends up in state `tx_full = 1' and tramsmit
> > > disabled, but with no outstanding transmit interrupts.
> 
> The better solution, which I've been adding to the drivers, is to check
> again for a just-cleared Tx queue after setting tx_full.
> That trades an extra comparison on a rarely followed path for a spinlock
> that is taken for every transmit and interrupt.
>
Please do not forget the memory barrier(s):
	
	tx_full = 1;
	if(condition)
		...;

That's exactly the sequence that caused deadlocks with wait_queues -
even a Pentium cpu will evaluate the condition before the write to
tx_full is commited. I have a test program (userspace) that reliably
locks up on my P II. I can send you the details if you are interested.

I think you also need a memory barrier in the tx_interrupt codepath.

> 
> Remember: spinlocks are expensive!
> 

But memory barriers are extremely error prone.

What about

tx_interrupt()

	if(netif_queue_stopped(dev)) {
		spin_lock(&np->lock);
		if(np->cur_tx - np->dirty_tx <= TX_QUEUE_LEN/2)
			netif_wake_queue(dev));
		spin_unlock(&np->lock);
	}

hard_xmit()

	if(np->cur_tx - np->dirty_tx >= TX_QUEUE_LEN-1) {
		spin_lock_irq(&np->lock);
		if(np->cur_tx - np_dirty_tx >= TX_QUEUE_LEN-1)
			netif_stop_queue(dev);
		spin_unlock_irq(&np->lock);
	}

--
	Manfred