David S. Miller a écrit :
No, all of your cpus are racing to get the transmit lock
of the tg3 driver. Whoever wins the race gets to queue
the packet, the others have to back off.
I believe the tg3_tx() holds the tx_lock for too long, in the case 200 or so
skb are delivered
Maybe adding a netif_stop_queue(dev) at the beginning of this function would
help ?
Or adding yet another bit in to tell that one qdisc_restart() is in flight..
int qdisc_restart(struct net_device *dev)
{
struct Qdisc *q = dev->qdisc;
struct sk_buff *skb;
+ if (dev->qrestart_active)
+ return 1;
/* Dequeue packet */
if ((skb = q->dequeue(q)) != NULL) {
...
{
+ dev->qrestart_active = 1;
/* And release queue */
spin_unlock(&dev->queue_lock);
if (!netif_queue_stopped(dev)) {
int ret;
if (netdev_nit)
dev_queue_xmit_nit(skb, dev);
ret = dev->hard_start_xmit(skb, dev);
if (ret == NETDEV_TX_OK) {
if (!nolock) {
dev->xmit_lock_owner = -1;
spin_unlock(&dev->xmit_lock);
}
spin_lock(&dev->queue_lock);
+ dev->qrestart_active = 0;
return -1;
}
if (ret == NETDEV_TX_LOCKED && nolock) {
spin_lock(&dev->queue_lock);
+ dev->qrestart_active = 0;
goto collision;
}
}
/* NETDEV_TX_BUSY - we need to requeue */
/* Release the driver */
if (!nolock) {
dev->xmit_lock_owner = -1;
spin_unlock(&dev->xmit_lock);
}
spin_lock(&dev->queue_lock);
+ dev->qrestart_active = 0;
q = dev->qdisc;
It does indeed look stupid that we resend the packet to
network taps repeatedly if we need to requeue. I wonder
what a clean way is to fix that. Probably the best idea
is to grab a reference to the SKB if netdev_nit, before
we send it off the the driver, and if the transmit succeeds
we actually call dev_queue_xmit_nit().
Actually, that won't work due to skb_header_cloned(). If
we pass the packet, or will, to the network taps, we have
to make sure skb_header_cloned() returns one else TSO mangling
of the TCP and IP headers by the driver will be seen by the
network taps.
So, this isn't easy to fix at all :)
Oh well ... :(
Given that only one skb (the first in the queue) could have the property
'already given to taps', I believe
we could use one bit at queue level : If set (by a previous requeue), then the
first packet *should not* be resent to taps.
-
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [EMAIL PROTECTED]
More majordomo info at http://vger.kernel.org/majordomo-info.html
