On Sun, Dec 31, 2017 at 02:30:09PM +0200, Sagi Grimberg wrote: > Not sure if stealing bios from requests is a better design. Note that > we do exactly this in other transport (nvme_[rdma|loop|fc]_is_ready). Well, we're currently failing requests that may succeed if we could back them out for re-entry. While such scenarios are uncommon, I think we can handle it better than ending them in failure. > I think it would be better to stick to a coherent behavior across > the nvme subsystem. If this condition statement is really something > that is buying us measurable performance gain, I think we should apply > it for other transports as well (although in fabrics we're a bit > different because we have a dedicated connect that enters .queue_rq) We should be able to remove all the state checks in the IO path because the handlers putting the controller in an IO-incapable state should be able to quiece the queues before transitioning to that state. This is not a significant gain compared to the other two patches in this series, but the sum of all the little things become meaningful. We'd need to remove this check anyway if we're considering exclusively polled queues since they wouldn't have a CQ vector.
