Re: [PATCH] try #2 even bigger natsemi patch

[Manfred Spraul]

Tim Hockin wrote:
> 
>  static void netdev_timer(unsigned long data)
>  {
>         struct net_device *dev = (struct net_device *)data;
>         struct netdev_private *np = dev->priv;
> -       int next_tick = 60*HZ;
> +       int next_tick = 5*HZ;
> +       long ioaddr = dev->base_addr;
> +       u16 dspcfg;
> 
>         if (debug > 3) {
>                 /* DO NOT read the IntrStatus register,
> @@ -933,10 +1107,27 @@
>                            dev->name);
>         }
>         spin_lock_irq(&np->lock);
> -       check_link(dev);
> +
> +       /* check for a nasty random phy-reset - use dspcfg as a flag */
> +       writew(1, ioaddr+PGSEL);
> +       dspcfg = readw(ioaddr+DSPCFG);
> +       writew(0, ioaddr+PGSEL);
> +       if (dspcfg != DSPCFG_VAL) {
> +               if (!netif_queue_stopped(dev)) {
> +                       printk(KERN_INFO
> +                               "%s: possible phy reset: re-initializing\n",
> +                               dev->name);
> +                       init_registers(dev);
That's not SMP safe:
The interrupt handler acquires np->lock only for packet tx and errors,
not for packet rx.
Thus an rx interrupt could be running during init_registers - and
init_registers modifies the rx ring.
I think you must first check for the phy reset, and if you find a reset
then lock rx interrupts out with disable_irq.

	disable_irq(dev->irq);
	spin_lock_irq(&np->lock);

Calling disable_irq() always is probably too slow.

--
	Manfred