On Tue, Nov 15, 2016 at 08:42:03PM -0800, John Fastabend wrote: > On 16-11-14 03:06 PM, Michael S. Tsirkin wrote: > > On Thu, Nov 10, 2016 at 08:44:32PM -0800, John Fastabend wrote: > >> Signed-off-by: John Fastabend <john.r.fastab...@intel.com> > > > > This will naturally reduce the cache line bounce > > costs, but so will a _many API for ptr-ring, > > doing lock-add many-unlock. > > > > the number of atomics also scales better with the lock: > > one per push instead of one per queue. > > > > Also, when can qdisc use a _many operation? > > > > On dequeue we can pull off many skbs instead of one at a time and > then either (a) pass them down as an array to the driver (I started > to write this on top of ixgbe and it seems like a win) or (b) pass > them one by one down to the driver and set the xmit_more bit correctly. > > The pass one by one also seems like a win because we avoid the lock > per skb. > > On enqueue qdisc side its a bit more evasive to start doing this. > > > [...]
I see. So we could wrap __ptr_ring_consume and implement __skb_array_consume. You can call that in a loop under a lock. I would limit it to something small like 16 pointers, to make sure lock contention is not an issue. -- MST
