On Wed, 30 Nov 2016 18:27:45 +0200 Saeed Mahameed <sae...@dev.mellanox.co.il> wrote:
> >> All in all, this is risky business :), the right way to go is to > >> force the upper layer to use xmit-more and delay doorbells/use bulking > >> but from the same context (xmit routine). For example see > >> Achiad's suggestion (attached in Jesper's response), he used stop > >> queue to force the stack to queue up packets (TX bulking) > >> which would set xmit-more and will use the next completion to > >> release the "stopped" ring TXQ rather than hit the doorbell on > >> behalf of it. > > > > Well, you depend on having a higher level queue like a qdisc. > > > > Some users do not use a qdisc. > > If you stop the queue, they no longer can send anything -> drops. > > You do have a point that stopping the device might not be the best way to create a push-back (to allow stack queue packets). netif_tx_stop_queue() / __QUEUE_STATE_DRV_XOFF > In this case, i think they should implement their own bulking (pktgen > is not a good example) but XDP can predict if it has more packets to > xmit as long as all of them fall in the same NAPI cycle. > Others should try and do the same. I actually agree with Saeed here. Maybe we can come up with another __QUEUE_STATE_xxx that informs the upper layer what the driver is doing. Then users not using a qdisc can use this indication (like the qdisc could). (qdisc-bypass users already check the QUEUE_STATE flags e.g. via netif_xmit_frozen_or_drv_stopped). My main objection is that this is a driver local optimization. By not involving the upper layers, the netstack looses the ability to amortize it's cost, as it still does per packet handling. -- Best regards, Jesper Dangaard Brouer MSc.CS, Principal Kernel Engineer at Red Hat LinkedIn: http://www.linkedin.com/in/brouer
