On Wed, 2008-05-07 at 19:41 +0800, Herbert Xu wrote:
quoted text
> On Wed, May 07, 2008 at 01:22:19PM +0200, Johannes Berg wrote:
> >
> > +	if (unlikely(netif_queue_stopped(local->mdev) ||
> > +		     __ieee80211_queue_stopped(local, control->queue)))
> > +		return NETDEV_TX_BUSY;
>=20
> Returning busy should be avoided because not everything expects it.

Hmm. Why does busy exist then? Historical accident? I really really
don't want to keep all that logic that does "if device enables queue
then first check if maybe we have pending packets and try to send them
and then go back to the regular queues and see if ..."

quoted text
> Besides, I'm not sure if this will even work if you're fiddling
> with skb->next.  Perhaps you can stash it in a pointer local to
> your device.

You're looking at an internal helper function that is never called with
skb->next assigned; if it returns BUSY for a fragmented skb then later
in __ieee80211_tx() I will store the fragmented frame away and return
OK.

But if I really need to avoid returning busy I need to hook into all the
queue management stuff which I'd like to avoid.

johannes