Skip to contentSkip to conversation

nvme-pci: fix single segment detection

PATCH In review Christoph Hellwig 3 participants 6 messages May 9, 2019 – May 13, 2019

Subsystems (2)
lore
Share & tools
lore

Christoph Hellwig

author maintainer
#0 lore raw

PATCH 1/1 nvme-pci: fix single segment detection

blk_rq_nr_phys_segments isn't exactly accurate at the request/bio level, so work around that fact with a few crude sanity checks. To be fixed for real in the block layer soon.

Fixes 297910571f08 ("nvme-pci: optimize mapping single segment requests using SGLs")
Signed-off-by Christoph Hellwig <hch at lst.de>
---
 drivers/nvme/host/pci.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)
drivers/nvme/host/pci.c +9 -1
index 2a8708c9ac18..9a4253be2723 100644
@@ -824,7 +824,15 @@ static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req,
 	blk_status_t ret = BLK_STS_RESOURCE;
 	int nr_mapped;
 
-	if (blk_rq_nr_phys_segments(req) == 1) {
+	/*
+	 * Crude hack to work around the block layer accounting for the fact
+	 * that the SG mapping can merge multiple physically contigous segments
+	 * in blk_rq_nr_phys_segments() despite the fact that not everyone is
+	 * mapping the request to a scatterlist.  To be fixed for real in the
+	 * block layer eventually..
+	 */
+	if (blk_rq_nr_phys_segments(req) == 1 &&
+	    !req->bio->bi_next && req->bio->bi_vcnt == 1) {
 		struct bio_vec bv = req_bvec(req);
 
 		if (!is_pci_p2pdma_page(bv.bv_page)) {
Signature · 1 line
2.20.1
#1 lore raw
Quoted 19 lines · Christoph Hellwig

blk_rq_nr_phys_segments isn't exactly accurate at the request/bio level, so work around that fact with a few crude sanity checks. To be fixed for real in the block layer soon.

Fixes 297910571f08 ("nvme-pci: optimize mapping single segment requests using SGLs")
Signed-off-by Christoph Hellwig <hch at lst.de>
---
 drivers/nvme/host/pci.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)
drivers/nvme/host/pci.c +0 -1
index 2a8708c9ac18..9a4253be2723 100644
@@ -824,7 +824,15 @@ static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req,
 	blk_status_t ret = BLK_STS_RESOURCE;
 	int nr_mapped;
 
-	if (blk_rq_nr_phys_segments(req) == 1) {

drivers/nvme/host/pci.c static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req, · line 836

Quoted 10 lines
drivers/nvme/host/pci.c +9 -0
+	/*
+	 * Crude hack to work around the block layer accounting for the fact
+	 * that the SG mapping can merge multiple physically contigous segments
+	 * in blk_rq_nr_phys_segments() despite the fact that not everyone is
+	 * mapping the request to a scatterlist.  To be fixed for real in the
+	 * block layer eventually..
+	 */
+	if (blk_rq_nr_phys_segments(req) == 1 &&
+	    !req->bio->bi_next && req->bio->bi_vcnt == 1) {
 		struct bio_vec bv = req_bvec(req);

I'd suggest to fix block layer instead of working around the issue here, then any driver may benefit from the fix.

Especially checking bio->bi_vcnt is just a hack, and drivers should never use .bi_vcnt.

Thanks,
Ming

author maintainer
#2 lore raw
Quoted 2 lines · Ming Lei

I'd suggest to fix block layer instead of working around the issue here, then any driver may benefit from the fix.

That is my plan, and I started on it. But the fix isn't trivial, and will probably take a while and be invasive.

Quoted 2 lines

Especially checking bio->bi_vcnt is just a hack, and drivers should never use .bi_vcnt.

That is why it is explicitly commented as a hack. But a good enough one to still speed up typical 4K I/Os - one a bio has been cloned chances are very high we don't care about the fast path any more.

Reviewed-by
#3 lore raw
Quoted 7 lines · Christoph Hellwig
Quoted 2 lines · Ming Lei

I'd suggest to fix block layer instead of working around the issue here, then any driver may benefit from the fix.

That is my plan, and I started on it. But the fix isn't trivial, and will probably take a while and be invasive.

Quoted 6 lines
Quoted 2 lines

Especially checking bio->bi_vcnt is just a hack, and drivers should never use .bi_vcnt.

That is why it is explicitly commented as a hack. But a good enough one to still speed up typical 4K I/Os - one a bio has been cloned chances are very high we don't care about the fast path any more.

NVMe hasn't max segment size limit, so fair enough for typical 4K IO workload:

Reviewed-by Ming Lei <ming.lei at redhat.com>

Thanks,
Ming

author maintainer
#4 lore raw

So actually, after running out ways to make any of my complex block layer fix ideas work I have a simple fix for the block layer now.

Lets see if is acceptable..