Re: [PATCH] Fix SMP support on 3c527 net driver

[Manfred Spraul]

Felipe wrote:

> Also, the down/up function doesn't seem to be 
> used in interrupt context, so I think it will work.
>  
[snip]

> static int mc32_send_packet(struct sk_buff *skb, struct net_device *dev)
> {
> 	struct mc32_local *lp = (struct mc32_local *)dev->priv;
> -	unsigned long flags;
>
> 	volatile struct skb_header *p, *np;
>
> 	netif_stop_queue(dev);
>
> -	save_flags(flags);
> -	cli();
> +	down(&lp->mc32_sem);
>  
No, that's wrong. mc32_send_packet is the hard_start_xmit function, 
called from bottom half context, with the dev_xmit_lock spinlock held.
Additionally, you must replace the sleep_on calls with wait_event, or an 
open-coded wait queue: sleep_on is racy, it only works with cli().

IMHO the right way to fix cli() is
- add a single spinlock to the driver or the device structure. Do not 
forget the spin_lock_init().
- replace cli/sti with spin_lock_irqsave/spin_unlock_irqsave.
- Additionally acquire the spinlock in every interrupt handler (cli() 
stops all interrupts, spinlocks only stop interrupt on the current cpu).
- check if there were recursive cli() calls. Fix them.
- replace all sleep_on calls with wait queue calls.
- check if there are any kmalloc or schedule calls in the area now under 
the spinlock, and reorganize the code.

And please add a changelog entry that code was converted without testing.

--
   Manfred

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/