> The way I understand the sequence of events w/o the > patch is: > > ib_req_notify_cq(IB_CQ_NEXT_COMP) > CQE 1 added to queue > callback scheduled via tasklet > future callbacks disarmed > callback function calls ib_req_notify_cq(IB_CQ_NEXT_COMP) > callback function calls ib_poll_cq() and gets CQE 1 > callback function calls ib_poll_cq() and gets none > CQE 2 added to queue via IRQ > callback scheduled via tasklet > future callbacks disarmed > callback function returns > some time later, tasklet runs and calls CQ callback function. > callback function calls ib_req_notify_cq(IB_CQ_NEXT_COMP) > callback function calls ib_poll_cq() and gets CQE 2 > > Since a tasklet or workqueue can be scheduled in the > callback function, the second CQE isn't "missed" but > there is a scheduling delay before the callback happens > and sees CQE 2.
Sure but a tasklet at least is likely to have pretty low latency. And maybe there's a better chance of batching completion processing by introducing a little delay there. So maybe it helps, maybe it doesn't. > I'm willing to withdraw the 1-3 patches. > > I still don't understand why the timing difference matters > to RDS. I'd like to have some data showing an improvement before making this change. And as I already said it would be good to know what's really going on with RDS before we change that. - R. -- Roland Dreier <[email protected]> || For corporate legal information go to: http://www.cisco.com/web/about/doing_business/legal/cri/index.html -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
