Skip to contentSkip to conversation

nvme: target: build bvec from sg directly

PATCH New Ming Lei 4 participants 10 messages Mar 25, 2019 – Mar 27, 2019

Subsystems (3)
lore
Share & tools
lore

Ming Lei

author
#0 lore raw

cover nvme: target: build bvec from sg directly

Hi,

The 1st patch fixes nvmet_file_init_bvec() for building bvec from sg list.

The 2nd patch converts to build bvec from sg directly, given multi-page bvec is ready.

Ming Lei (2):
nvme: target: fix nvmet_file_init_bvec()
nvme: target: build bvec from sg directly

 drivers/nvme/target/io-cmd-file.c | 26 ++++++++++++--------------
 1 file changed, 12 insertions(+), 14 deletions(-)
Cc Yi Zhang <yi.zhang at redhat.com>
Cc Sagi Grimberg <sagi at grimberg.me>
Cc Chaitanya Kulkarni <chaitanya.kulkarni at wdc.com>
Signature · 1 line
2.9.5
author
#1 lore raw

PATCH 1/2 nvme: target: fix nvmet_file_init_bvec()

There isn't sg iterator helper for building bvec, so invent one and fix the issue in nvmet_file_init_bvec().

The issue is that one sg may include multipge continuous pages, and only the 1st .bv_offset isn't zero, also the length for the last bvec has to consider the remained length.

Reported-by Yi Zhang <yi.zhang at redhat.com>
Fixes d5eff33ee6f8("nvmet: add simple file backed ns support")
Cc Yi Zhang <yi.zhang at redhat.com>
Cc Sagi Grimberg <sagi at grimberg.me>
Cc Chaitanya Kulkarni <chaitanya.kulkarni at wdc.com>
Cc <stable at vger.kernel.org>
Signed-off-by Ming Lei <ming.lei at redhat.com>
---
 drivers/nvme/target/io-cmd-file.c | 57 ++++++++++++++++++++-------------------
 1 file changed, 29 insertions(+), 28 deletions(-)
drivers/nvme/target/io-cmd-file.c +29 -28
index 3e43212d3c1c..1c76e9e2a474 100644
@@ -75,13 +75,6 @@ int nvmet_file_ns_enable(struct nvmet_ns *ns)
 	return ret;
 }
 
-static void nvmet_file_init_bvec(struct bio_vec *bv, struct sg_page_iter *iter)
-{
-	bv->bv_page = sg_page_iter_page(iter);
-	bv->bv_offset = iter->sg->offset;
-	bv->bv_len = PAGE_SIZE - iter->sg->offset;
-}
-
 static ssize_t nvmet_file_submit_bvec(struct nvmet_req *req, loff_t pos,
 		unsigned long nr_segs, size_t count, int ki_flags)
 {
@@ -129,13 +122,14 @@ static void nvmet_file_io_done(struct kiocb *iocb, long ret, long ret2)
 static bool nvmet_file_execute_io(struct nvmet_req *req, int ki_flags)
 {
 	ssize_t nr_bvec = DIV_ROUND_UP(req->data_len, PAGE_SIZE);
-	struct sg_page_iter sg_pg_iter;
+	struct scatterlist *sg;
+	struct bio_vec *bv;
 	unsigned long bv_cnt = 0;
 	bool is_sync = false;
 	size_t len = 0, total_len = 0;
 	ssize_t ret = 0;
 	loff_t pos;
-
+	int i, j, sg_done;
 
 	if (req->f.mpool_alloc && nr_bvec > NVMET_MAX_MPOOL_BVEC)
 		is_sync = true;
@@ -147,26 +141,33 @@ static bool nvmet_file_execute_io(struct nvmet_req *req, int ki_flags)
 	}
 
 	memset(&req->f.iocb, 0, sizeof(struct kiocb));
-	for_each_sg_page(req->sg, &sg_pg_iter, req->sg_cnt, 0) {
-		nvmet_file_init_bvec(&req->f.bvec[bv_cnt], &sg_pg_iter);
-		len += req->f.bvec[bv_cnt].bv_len;
-		total_len += req->f.bvec[bv_cnt].bv_len;
-		bv_cnt++;
-
-		WARN_ON_ONCE((nr_bvec - 1) < 0);
-
-		if (unlikely(is_sync) &&
-		    (nr_bvec - 1 == 0 || bv_cnt == NVMET_MAX_MPOOL_BVEC)) {
-			ret = nvmet_file_submit_bvec(req, pos, bv_cnt, len, 0);
-			if (ret < 0)
-				goto complete;
-
-			pos += len;
-			bv_cnt = 0;
-			len = 0;
+	for_each_sg(req->sg, sg, req->sg_cnt, i)
+		for (j = 0, sg_done = 0;
+		     (bv = &req->f.bvec[bv_cnt]) && sg_done < sg->length;
+		     j++, sg_done += bv->bv_len) {
+			bv->bv_offset = j ? 0 : sg->offset;
+			bv->bv_len = min_t(unsigned, PAGE_SIZE - bv->bv_offset,
+					   sg->length - sg_done);
+			bv->bv_page = nth_page(sg_page(sg), j);
+
+			len += bv->bv_len;
+			total_len += bv->bv_len;
+			bv_cnt++;
+
+			WARN_ON_ONCE((nr_bvec - 1) < 0);
+
+			if (unlikely(is_sync) &&
+			    (nr_bvec - 1 == 0 || bv_cnt == NVMET_MAX_MPOOL_BVEC)) {
+				ret = nvmet_file_submit_bvec(req, pos, bv_cnt, len, 0);
+				if (ret < 0)
+					goto complete;
+
+				pos += len;
+				bv_cnt = 0;
+				len = 0;
+			}
+			nr_bvec--;
 		}
-		nr_bvec--;
-	}
 
 	if (WARN_ON_ONCE(total_len != req->data_len)) {
 		ret = -EIO;
Signature · 1 line
2.9.5
#2 lore raw

PATCH 1/2 nvme: target: fix nvmet_file_init_bvec()

Quoted 6 lines · Ming Lei

There isn't sg iterator helper for building bvec, so invent one and fix the issue in nvmet_file_init_bvec().

The issue is that one sg may include multipge continuous pages, and only the 1st .bv_offset isn't zero, also the length for the last bvec has to consider the remained length.

The scatterlist in the nvme target is always allocated by the nvmet code itself an thus never contains multi-page sg list entries.

author
#3 lore raw

PATCH 1/2 nvme: target: fix nvmet_file_init_bvec()

Quoted 10 lines · Christoph Hellwig
Quoted 6 lines · Ming Lei

There isn't sg iterator helper for building bvec, so invent one and fix the issue in nvmet_file_init_bvec().

The issue is that one sg may include multipge continuous pages, and only the 1st .bv_offset isn't zero, also the length for the last bvec has to consider the remained length.

The scatterlist in the nvme target is always allocated by the nvmet code itself an thus never contains multi-page sg list entries.

I am wondering if it is true.

Not look at other target code yet, however seems it isn't true for loop, see the following code in nvme_loop_queue_rq():

	iod->req.sg = iod->sg_table.sgl;
	iod->req.sg_cnt = blk_rq_map_sg(req->q, req, iod->sg_table.sgl);
	iod->req.transfer_len = blk_rq_payload_bytes(req);

And it has been triggered by nvme/011 in Yi's test.

Thanks,
Ming

maintainer
#4 lore raw

PATCH 1/2 nvme: target: fix nvmet_file_init_bvec()

Quoted 6 lines
Quoted 3 lines

The scatterlist in the nvme target is always allocated by the nvmet code itself an thus never contains multi-page sg list entries.

I am wondering if it is true.

Quoted 8 lines

Not look at other target code yet, however seems it isn't true for loop, see the following code in nvme_loop_queue_rq():

	iod->req.sg = iod->sg_table.sgl;
	iod->req.sg_cnt = blk_rq_map_sg(req->q, req, iod->sg_table.sgl);
	iod->req.transfer_len = blk_rq_payload_bytes(req);

And it has been triggered by nvme/011 in Yi's test.

Yes, loop is an exception in this case.

author
#5 lore raw

PATCH 1/2 nvme: target: fix nvmet_file_init_bvec()

Quoted 9 lines · Sagi Grimberg
Quoted 9 lines
Quoted 3 lines

The scatterlist in the nvme target is always allocated by the nvmet code itself an thus never contains multi-page sg list entries.

I am wondering if it is true.

Not look at other target code yet, however seems it isn't true for loop, see the following code in nvme_loop_queue_rq():

Quoted 7 lines
Quoted 5 lines
	iod->req.sg = iod->sg_table.sgl;
	iod->req.sg_cnt = blk_rq_map_sg(req->q, req, iod->sg_table.sgl);
	iod->req.transfer_len = blk_rq_payload_bytes(req);

And it has been triggered by nvme/011 in Yi's test.

Yes, loop is an exception in this case.

Thanks for the clarification!

Another candidate fix is to set nvmet-loop's queue segment boundary mask as PAGE_SIZE - 1.

Also there is the same issue for block device backed target.

If no one objects, I'd like to take the approach of adjusting segment boundary mask.

Thanks,
Ming

#6 lore raw

PATCH 1/2 nvme: target: fix nvmet_file_init_bvec()

Quoted 4 lines · Ming Lei

Also there is the same issue for block device backed target.

If no one objects, I'd like to take the approach of adjusting segment boundary mask.

Just go for the merge of your two patches. That fixes the issue and actually makes the code simpler.

maintainer
#7 lore raw

PATCH 1/2 nvme: target: fix nvmet_file_init_bvec()

In reply to Ming Lei

Hi,

[This is an automated email]

This commit has been processed because it contains a "Fixes:" tag, fixing commit: .

The bot has tested the following trees: v5.0.4, v4.19.31, v4.14.108, v4.9.165, v4.4.177, v3.18.137.

v5.0.4: Build OK!
v4.19.31: Failed to apply! Possible dependencies:
50a909db36f2 ("nvmet: use IOCB_NOWAIT for file-ns buffered I/O")

v4.14.108: Failed to apply! Possible dependencies:
1b72b71facce ("nvmet: fix file discard return status") 50a909db36f2 ("nvmet: use IOCB_NOWAIT for file-ns buffered I/O") 543c09c89fdc ("nvmet: Fix nvmet_execute_write_zeroes sector count") 55eb942eda2c ("nvmet: add buffered I/O support for file backed ns") 618cff4285dc ("nvmet: remove duplicate NULL initialization for req->ns") 7756f72ccd43 ("nvmet: Change return code of discard command if not supported") 9c891c139894 ("nvmet: check fileio lba range access boundaries") d5eff33ee6f8 ("nvmet: add simple file backed ns support") e454d122e228 ("nvmet: kill nvmet_inline_bio_init")
e929f06d9eaa ("nvmet: constify struct nvmet_fabrics_ops") ea435e1b9392 ("block: add a poll_fn callback to struct request_queue")

v4.9.165: Failed to apply! Possible dependencies:
297e3d854784 ("blk-throttle: make throtl_slice tunable") 4151dd9a58c6 ("nvmet: Fixed avoided printing nvmet: twice in error logs.") 4e4cbee93d56 ("block: switch bios to blk_status_t")
778f067c185c ("nvme: add semicolon in nvme_command setting") 87760e5eef35 ("block: hook up writeback throttling") 986994a27587 ("nvme: Use CNS as 8-bit field and avoid endianness conversion") 9e234eeafbe1 ("blk-throttle: add a simple idle detection") cf43e6be865a ("block: add scalable completion tracking of requests") d5eff33ee6f8 ("nvmet: add simple file backed ns support") e806402130c9 ("block: split out request-only flags into a new namespace") fbbaf700e7b1 ("block: trace completion of all bios.")

v4.4.177: Failed to apply! Possible dependencies:
1c63dc66580d ("nvme: split a new struct nvme_ctrl out of struct nvme_dev") 21d34711e1b5 ("nvme: split command submission helpers out of pci.c") 2d79c7dc8fe5 ("admin-cmd: Added smart-log command support.") 3a85a5de29ea ("nvme-loop: add a NVMe loopback host driver") 4151dd9a58c6 ("nvmet: Fixed avoided printing nvmet: twice in error logs.") 540c801c65eb ("NVMe: Implement namespace list scanning") 6f3b0e8bcf3c ("blk-mq: add a flags parameter to blk_mq_alloc_request") 986994a27587 ("nvme: Use CNS as 8-bit field and avoid endianness conversion") a07b4970f464 ("nvmet: add a generic NVMe target")
d5eff33ee6f8 ("nvmet: add simple file backed ns support") fa60682677c5 ("nvme: use symbolic constants for CNS values")

v3.18.137: Failed to apply! Possible dependencies:
062261be4e39 ("nvme: Replace rcu_assign_pointer() with RCU_INIT_POINTER()") 1d0906246095 ("NVMe: Mismatched host/device page size support") 302c6727e5eb ("NVMe: Fix filesystem sync deadlock on removal") 57dacad5f228 ("nvme: move to a new drivers/nvme/host directory") 5905535610fc ("NVMe: Correctly handle IOCTL_SUBMIT_IO when cpus > online queues") 5940c8578fe7 ("NVMe: Clear QUEUE_FLAG_STACKABLE")
6fccf9383b28 ("NVMe: Async event request")
7963e521811e ("NVMe: Passthrough IOCTL for IO commands") 7be50e93fbc2 ("NVMe: Fix nvmeq waitqueue entry initialization") 9dbbfab7d541 ("NVMe: Do not over allocate for discard requests") a07b4970f464 ("nvmet: add a generic NVMe target")
a4aea5623d4a ("NVMe: Convert to blk-mq")
a96d4f5c2da6 ("NVMe: Reference count pci device")
aee4b9bd4505 ("nvmem: Add to MAINTAINERS for nvmem framework") b4ff9c8ddb6f ("NVMe: Translate NVMe status to errno")
d5eff33ee6f8 ("nvmet: add simple file backed ns support") dece45855a8b ("NFC: nxp-nci: Add support for NXP NCI chips") f435c2825b4c ("NVMe: Call nvme_free_queue directly")

How should we proceed with this patch?

--
Thanks,
Sasha

author
#8 lore raw

PATCH 2/2 nvme: target: build bvec from sg directly

In reply to Ming Lei

Now multi-page bvec is supported, and the whole IO stack is capable of handling multi-page bvec.

So build each bvec from the sg directly.

Cc Yi Zhang <yi.zhang at redhat.com>
Cc Sagi Grimberg <sagi at grimberg.me>
Cc Chaitanya Kulkarni <chaitanya.kulkarni at wdc.com>
Signed-off-by Ming Lei <ming.lei at redhat.com>
---
 drivers/nvme/target/io-cmd-file.c | 55 ++++++++++++++++++---------------------
 1 file changed, 26 insertions(+), 29 deletions(-)
drivers/nvme/target/io-cmd-file.c +26 -29
index 1c76e9e2a474..e2a92b85baa0 100644
@@ -121,15 +121,14 @@ static void nvmet_file_io_done(struct kiocb *iocb, long ret, long ret2)
 
 static bool nvmet_file_execute_io(struct nvmet_req *req, int ki_flags)
 {
-	ssize_t nr_bvec = DIV_ROUND_UP(req->data_len, PAGE_SIZE);
+	ssize_t nr_bvec = req->sg_cnt;
 	struct scatterlist *sg;
-	struct bio_vec *bv;
 	unsigned long bv_cnt = 0;
 	bool is_sync = false;
 	size_t len = 0, total_len = 0;
 	ssize_t ret = 0;
 	loff_t pos;
-	int i, j, sg_done;
+	int i;
 
 	if (req->f.mpool_alloc && nr_bvec > NVMET_MAX_MPOOL_BVEC)
 		is_sync = true;
@@ -141,33 +140,31 @@ static bool nvmet_file_execute_io(struct nvmet_req *req, int ki_flags)
 	}
 
 	memset(&req->f.iocb, 0, sizeof(struct kiocb));
-	for_each_sg(req->sg, sg, req->sg_cnt, i)
-		for (j = 0, sg_done = 0;
-		     (bv = &req->f.bvec[bv_cnt]) && sg_done < sg->length;
-		     j++, sg_done += bv->bv_len) {
-			bv->bv_offset = j ? 0 : sg->offset;
-			bv->bv_len = min_t(unsigned, PAGE_SIZE - bv->bv_offset,
-					   sg->length - sg_done);
-			bv->bv_page = nth_page(sg_page(sg), j);
-
-			len += bv->bv_len;
-			total_len += bv->bv_len;
-			bv_cnt++;
-
-			WARN_ON_ONCE((nr_bvec - 1) < 0);
-
-			if (unlikely(is_sync) &&
-			    (nr_bvec - 1 == 0 || bv_cnt == NVMET_MAX_MPOOL_BVEC)) {
-				ret = nvmet_file_submit_bvec(req, pos, bv_cnt, len, 0);
-				if (ret < 0)
-					goto complete;
-
-				pos += len;
-				bv_cnt = 0;
-				len = 0;
-			}
-			nr_bvec--;
+	for_each_sg(req->sg, sg, req->sg_cnt, i) {
+		struct bio_vec *bv = &req->f.bvec[bv_cnt];
+
+		bv->bv_offset = sg->offset;
+		bv->bv_len = sg->length;
+		bv->bv_page = sg_page(sg);
+
+		len += bv->bv_len;
+		total_len += bv->bv_len;
+		bv_cnt++;
+
+		WARN_ON_ONCE((nr_bvec - 1) < 0);
+
+		if (unlikely(is_sync) &&
+		    (nr_bvec - 1 == 0 || bv_cnt == NVMET_MAX_MPOOL_BVEC)) {
+			ret = nvmet_file_submit_bvec(req, pos, bv_cnt, len, 0);
+			if (ret < 0)
+				goto complete;
+
+			pos += len;
+			bv_cnt = 0;
+			len = 0;
 		}
+		nr_bvec--;
+	}
 
 	if (WARN_ON_ONCE(total_len != req->data_len)) {
 		ret = -EIO;
Signature · 1 line
2.9.5
#9 lore raw

PATCH 2/2 nvme: target: build bvec from sg directly

This looks pretty sensible even if we don't end up submitting multi-page vecs. But can you please just send it on top of mainline directly? The previos patch just creates pointless churn.