On Thu, Jan 11, 2024 at 10:18:31AM -0700, Jens Axboe wrote: > This also highlights a potential inefficiency in the patch, as now we're > grabbing+dropping references when we don't need to. May not be a big > deal, but it's one of the things that cached requests got rid of. Though > I'm not quite sure how to refactor to get rid of that, as we'd need to > shuffle the splitting and request get for that. > > Could you take another look at the series with that in mind? I thought about it, but it gets pretty ugly quickly. bio_queue_enter needs to move back into blk_mq_submit_bio, and then we'd skip it initially if bio_may_exceed_limits is false, and then we later need to add it back. (we'll probably also need to special case blk_queue_bounce as that setting could change to. I wish we could finally kill that)
