Le lundi 20 juin 2011 à 16:07 +0200, Maxime Bizon a écrit : > On Mon, 2011-06-20 at 09:48 +0200, Richard Cochran wrote: > > Hi Richard, > > > Because the socket buffer is freed in the completion interrupt, it is not > > safe to access it after submitting it to the hardware. > > I don't see why. > > skb is freed from txq_reclaim() which grabs the tx queue lock before, > (hence the lockless __skb_queue_xxx() in both functions) > > What am I missing ?
You are right, this driver still takes tx queue lock in its TX completion path. txq_reclaim() can currently run for a very long time, blocking other cpus to make progress if blocked on tx queue lock. So consider this patch as a cleanup : Its better to make all drivers behave the same way 1) Either increment tx stats in their start_xmit(), and making sure they dont access skb->len too late. 2) increment tx stats in their TX completion path. Then we can work on making TX completion path not taking tx queue lock in its fast path. _______________________________________________ stable mailing list [email protected] http://linux.kernel.org/mailman/listinfo/stable
