Ben Greear wrote:

> On a related note, I am now working on a way to use a hook in the
> netif_wake_queue callback to wake up pktgen.  This should allow me to
> have a pktgen that does not need to spin in a tight loop like it does
> now.  So far, I was able to saturate two GigE ports using about 3% of
> the CPU (as reported by top), using 1514 byte pkts.  Still tweaking
> to fix some corner cases...

I think this is mostly working now.  The patch to netdevice.h looks like this::

--- linux-2.6.7/include/linux/netdevice.h	2004-06-15 22:20:04.000000000 -0700
+++ linux-2.6.7.p4s/include/linux/netdevice.h	2004-10-06 14:31:54.000000000 -0700

@@ -466,9 +474,17 @@
  	void                    (*poll_controller)(struct net_device *dev);
  #endif

+        /* Callback for when the queue is woken, used by pktgen currently */
+	int			(*notify_queue_woken)(struct net_device *dev);
+        void* nqw_data; /* To be used by the method above as needed */
+
  	/* bridge stuff */
  	struct net_bridge_port	*br_port;

+#if defined(CONFIG_MACVLAN) || defined(CONFIG_MACVLAN_MODULE)
+	struct macvlan_port *macvlan_priv;
+#endif
+
  #ifdef CONFIG_NET_FASTROUTE
  #define NETDEV_FASTROUTE_HMASK 0xF
  	/* Semi-private data. Keep it at the end of device struct. */
@@ -619,8 +635,13 @@
  	if (netpoll_trap())
  		return;
  #endif
-	if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state))
+	if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state)) {
  		__netif_schedule(dev);
+
+                if (dev->notify_queue_woken) {
+                   dev->notify_queue_woken(dev);
+                }
+        }
  }


I'm attaching my version of pktgen since the diff is bigger than
the actual file.  With this, I can send right at 600kpps to myself (60 byte + CRC pkts),
and only use about 35% cpu as reported by top.  I also tried maxing out 6
ports and I hit what I believe is the PCI-X limit on my machine
with an aggregate throughput of about 1.3Gbps tx + 1.3Gbps rx.  CPU load is about 21%
in this case.


It may be that other programs would like to use the notify_queue_woken hook, so
if this were ever to hit the kernel proper, might want to make this a linked list
of callbacks instead of a simple pointer.

Enjoy,
Ben

-- 
Ben Greear <greearb@candelatech.com>
Candela Technologies Inc  http://www.candelatech.com