Skip to contentSkip to conversation

netpoll: prevent transmission stall

PATCH In review Stelian Pop 2 participants 2 messages Apr 8, 2004

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 netpoll: prevent transmission stall

While working on kgdb over netpoll I found out another bug in netpoll.

This one is in netpoll_send_skb(), and it is triggered by sending repeatedly more packets than the transmit buffers the ethernet card has, while being in the netpoll_trap. (this is rapidly triggered by kgdb when using a card with limited transmit buffers, like my pcmcia NE2000).

One should poll() the ethernet card interrupt function in order to get the transmit acks from the card and free the transmit buffers in order to be able to transmit more packets.

This is why in the original code, netpoll_send_skb() calls netpoll_poll() if dev->hard_start_xmit() fails. However, this does not work because netpoll_poll() is called only if netif_queue_stopped(). But the net queue functions are overridden if netpoll_trap() is on, causing netif_queue_stopped() to never return true.

The attached patch 'fixes' this, by forcing a netpoll_poll() every time dev->hard_start_xmit() fails. Maybe there is a cleaner way to do this by making the queue functions work even under netpoll_trap, but it shouldn't really be necessary, I cannot see how forcing the netpoll_poll() call could break anything.

Stelian.

core/netpoll.c renamed from core/netpoll.c.ORIG +3 -9
@@ -163,21 +163,15 @@
 	spin_lock(&np->dev->xmit_lock);
 	np->dev->xmit_lock_owner = smp_processor_id();
 
-	if (netif_queue_stopped(np->dev)) {
-		np->dev->xmit_lock_owner = -1;
-		spin_unlock(&np->dev->xmit_lock);
-
-		netpoll_poll(np);
-		goto repeat;
-	}
-
 	status = np->dev->hard_start_xmit(skb, np->dev);
 	np->dev->xmit_lock_owner = -1;
 	spin_unlock(&np->dev->xmit_lock);
 
 	/* transmit busy */
-	if(status)
+	if(status) {
+		netpoll_poll(np);
 		goto repeat;
+	}
 }
 
 void netpoll_send_udp(struct netpoll *np, const char *msg, int len)
Signature · 1 line
Stelian Pop <[email protected]>
#1 lore raw
Quoted 20 lines · Stelian Pop

While working on kgdb over netpoll I found out another bug in netpoll.

This one is in netpoll_send_skb(), and it is triggered by sending repeatedly more packets than the transmit buffers the ethernet card has, while being in the netpoll_trap. (this is rapidly triggered by kgdb when using a card with limited transmit buffers, like my pcmcia NE2000).

One should poll() the ethernet card interrupt function in order to get the transmit acks from the card and free the transmit buffers in order to be able to transmit more packets.

This is why in the original code, netpoll_send_skb() calls netpoll_poll() if dev->hard_start_xmit() fails. However, this does not work because netpoll_poll() is called only if netif_queue_stopped(). But the net queue functions are overridden if netpoll_trap() is on, causing netif_queue_stopped() to never return true.

Quoted 5 lines

The attached patch 'fixes' this, by forcing a netpoll_poll() every time dev->hard_start_xmit() fails. Maybe there is a cleaner way to do this by making the queue functions work even under netpoll_trap, but it shouldn't really be necessary, I cannot see how forcing the netpoll_poll() call could break anything.

Ok, makes sense. I'll queue this up.

Signature · 1 line
Matt Mackall : http://www.selenic.com : Linux development and consulting