On Wed, 30 May 2007 00:20:41 +0200
Francois Romieu [off-list ref] wrote:

Stephen Hemminger [off-list ref] :
⅜...]

quoted

Better to just get rid of using the lock as a transmit lock and
use netif_tx_lock instead.

--- a/drivers/net/b44.c	2007-05-29 09:51:43.000000000 -0700
+++ b/drivers/net/b44.c	2007-05-29 14:26:03.000000000 -0700

@@ -607,6 +607,7 @@ static void b44_tx(struct b44 *bp)
 {
 	u32 cur, cons;
 
+	netif_tx_lock(bp->dev);
 	cur  = br32(bp, B44_DMATX_STAT) & DMATX_STAT_CDMASK;
 	cur /= sizeof(struct dma_desc);
 

(damn, you are quick)

I am not completely convinced.

1. netpoll_send_skb (calls netif_tx_trylock(dev))
   -> netpoll_poll(np)
      -> poll_napi(np)
         -> np->dev->poll(np->dev, &budget) ( == b44_poll)
            -> b44_tx
               -> netif_tx_lock(bp->dev) *deadlock*

Netpoll needs to be fixed. (or scrapped), as is it will break drivers
trying to use tx_lock in poll routine. I know sky2 would get borked.

Something like this:

--- a/net/core/netpoll.c	2007-05-08 14:19:32.000000000 -0700
+++ b/net/core/netpoll.c	2007-05-29 15:28:22.000000000 -0700

@@ -250,22 +250,23 @@ static void netpoll_send_skb(struct netp
 		unsigned long flags;
 
 		local_irq_save(flags);
-		if (netif_tx_trylock(dev)) {
-			/* try until next clock tick */
-			for (tries = jiffies_to_usecs(1)/USEC_PER_POLL;
-					tries > 0; --tries) {
+		/* try until next clock tick */
+		for (tries = jiffies_to_usecs(1)/USEC_PER_POLL;
+		     tries > 0; --tries) {
+			if (netif_tx_trylock(dev)) {
 				if (!netif_queue_stopped(dev))
 					status = dev->hard_start_xmit(skb, dev);
+				netif_tx_unlock(dev);
 
 				if (status == NETDEV_TX_OK)
 					break;
 
-				/* tickle device maybe there is some cleanup */
-				netpoll_poll(np);
-
-				udelay(USEC_PER_POLL);
 			}
-			netif_tx_unlock(dev);
+
+			/* tickle device maybe there is some cleanup */
+			netpoll_poll(np);
+
+			udelay(USEC_PER_POLL);
 		}
 		local_irq_restore(flags);
 	}