On Sat, Jul 26, 2008 at 02:18:46AM -0700, David Miller wrote:

From: Jarek Poplawski <redacted>
Date: Fri, 25 Jul 2008 22:01:37 +0200

quoted

On Fri, Jul 25, 2008 at 09:36:15PM +0200, Johannes Berg wrote:

quoted

On Fri, 2008-07-25 at 21:34 +0200, Jarek Poplawski wrote:

No, we need to be able to lock out multiple TX paths at once.

IMHO, it can do the same. We could e.g. insert a locked spinlock into
this noop_tx_handler(), to give everyone some waiting.

I think there might be an easier way, but we may have
to modify the state bits a little.

Yes, this looks definitely easier, but here is this one little bit
more, plus additional code to handle this in various places. Ingo's
proposal needs a (one?!) bit more thinking in one place, but it
shouldn't add even a bit to tx path (and it looks really cool!). Of
course, it could be re-considered in some other time too.

BTW, it seems with "Ingo's method" this netif_queue_stopped() check
could be removed too - the change of handlers could be done with
single qdiscs as well.

Jarek P.