On Thu, Nov 21, 2019 at 9:37 PM Pankaj Gupta <pagupta@xxxxxxxxxx> wrote: > > > > > > > > > > > > > > > > I added that and was about to push this out, but what about the > > > > > > > fact > > > > > > > that now the guest will synchronously wait for flushing to occur. > > > > > > > The > > > > > > > goal of the child bio was to allow that to be an I/O wait with > > > > > > > overlapping I/O, or at least not blocking the submission thread. > > > > > > > Does > > > > > > > the block layer synchronously wait for PREFLUSH requests? If not I > > > > > > > think a synchronous wait is going to be a significant performance > > > > > > > regression. Are there any numbers to accompany this change? > > > > > > > > > > > > Why not just swap the parent child relationship in the PREFLUSH case? > > > > > > > > > > I we are already inside parent bio "make_request" function and we > > > > > create > > > > > child > > > > > bio. How we exactly will swap the parent/child relationship for > > > > > PREFLUSH > > > > > case? > > > > > > > > > > Child bio is queued after parent bio completes. > > > > > > > > Sorry, I didn't quite mean with bio_split, but issuing another request > > > > in front of the real bio. See md_flush_request() for inspiration. > > > > > > o.k. Thank you. Will try to post patch today to be considered for 5.4. > > > > > > > I think it is too late for v5.4-final, but we can get it in the > > -stable queue. Let's take the time to do it right and get some testing > > on it. > > Sure. > > Just sharing probable patch for early feedback, if I am doing it correctly? > I will test it thoroughly. > > Thanks, > Pankaj > > ======== > > diff --git a/drivers/nvdimm/nd_virtio.c b/drivers/nvdimm/nd_virtio.c > index 10351d5b49fa..c683e0e2515c 100644 > --- a/drivers/nvdimm/nd_virtio.c > +++ b/drivers/nvdimm/nd_virtio.c > @@ -112,6 +112,12 @@ int async_pmem_flush(struct nd_region *nd_region, struct bio *bio) > bio_copy_dev(child, bio); > child->bi_opf = REQ_PREFLUSH; > child->bi_iter.bi_sector = -1; > + > + if (unlikely(bio->bi_opf & REQ_PREFLUSH)) { > + struct request_queue *q = bio->bi_disk->queue; > + q->make_request_fn(q, child); > + return 0; > + } > bio_chain(child, bio); > submit_bio(child); In the md case there is a lower level device to submit to. In this case I expect you would - create a flush workqueue - queue the bio that workqueue and wait for any previous flush request to complete (md_flush_request does this) - run virtio_pmem_flush - complete the original bio Is there a way to make virtio_pmem_flush() get an interrupt when the flush is complete rather than synchronously waiting. That way if you get a storm of flush requests you can coalesce them like md_flush_request() does.