On Sat, Jul 25, 2009 at 11:24:36AM +0800, Herbert Xu wrote:
...

However, I am definitely with you in that we should perform this
optimisation since it makes sense for the majority of people who
use multiqueue TX.

So the fact that our current architecture penalises the people
who actually need multiqueue TX in order to ensure correctness
for the people who cannot use multiqueue TX effectively (i.e.,
those who use non-default qdiscs) makes me uneasy.

It would be nice to establish what difference is made by each part of
this patch. The check for stopped queue before each skb xmit was done
earlier too, so it's not exactly multiqueue penalization.

Cheers,
Jarek P.