On 10/22/20 08:58, Logan Gunthorpe wrote:

On 2020-10-21 7:02 p.m., Chaitanya Kulkarni wrote:

quoted

In nvmet_passthru_execute_cmd() which is a high frequency function
it uses bio_alloc() which leads to memory allocation from the fs pool
for each I/O.

For NVMeoF nvmet_req we already have inline_bvec allocated as a part of
request allocation that can be used with preallocated bio when we
already know the size of request before bio allocation with bio_alloc(),
which we already do.

Introduce a bio member for the nvmet_req passthru anon union. In the
fast path, check if we can get away with inline bvec and bio from
nvmet_req with bio_init() call before actually allocating from the
bio_alloc().

This will be useful to avoid any new memory allocation under high
memory pressure situation and get rid of any extra work of
allocation (bio_alloc()) vs initialization (bio_init()) when
transfer len is < NVMET_MAX_INLINE_DATA_LEN that user can configure at
compile time.

Signed-off-by: Chaitanya Kulkarni <redacted>
---
 drivers/nvme/target/nvmet.h    |  1 +
 drivers/nvme/target/passthru.c | 20 ++++++++++++++++++--
 2 files changed, 19 insertions(+), 2 deletions(-)

diff --git a/drivers/nvme/target/nvmet.h b/drivers/nvme/target/nvmet.h
index 559a15ccc322..408a13084fb4 100644
--- a/drivers/nvme/target/nvmet.h
+++ b/drivers/nvme/target/nvmet.h

@@ -330,6 +330,7 @@ struct nvmet_req {
 			struct work_struct      work;
 		} f;
 		struct {
+			struct bio		inline_bio;
 			struct request		*rq;
 			struct work_struct      work;
 			bool			use_workqueue;

diff --git a/drivers/nvme/target/passthru.c b/drivers/nvme/target/passthru.c
index 496ffedb77dc..32498b4302cc 100644
--- a/drivers/nvme/target/passthru.c
+++ b/drivers/nvme/target/passthru.c

@@ -178,6 +178,14 @@ static void nvmet_passthru_req_done(struct request *rq,
 	blk_mq_free_request(rq);
 }
 
+static void nvmet_passthru_bio_done(struct bio *bio)
+{
+	struct nvmet_req *req = bio->bi_private;
+
+	if (bio != &req->p.inline_bio)
+		bio_put(bio);
+}
+
 static int nvmet_passthru_map_sg(struct nvmet_req *req, struct request *rq)
 {
 	int sg_cnt = req->sg_cnt;

@@ -186,13 +194,21 @@ static int nvmet_passthru_map_sg(struct nvmet_req *req, struct request *rq)
 	int i;
 
 	bio = bio_alloc(GFP_KERNEL, min(sg_cnt, BIO_MAX_PAGES));
-	bio->bi_end_io = bio_put;
+	if (req->transfer_len <= NVMET_MAX_INLINE_DATA_LEN) {
+		bio = &req->p.inline_bio;
+		bio_init(bio, req->inline_bvec, ARRAY_SIZE(req->inline_bvec));
+	} else {
+		bio = bio_alloc(GFP_KERNEL, min(sg_cnt, BIO_MAX_PAGES));
+	}
+
+	bio->bi_end_io = nvmet_passthru_bio_done;

I still think it's cleaner to change bi_endio for the inline/alloc'd
cases by simply setting bi_endi_io to bio_put() only in the bio_alloc
case. This should also be more efficient as it's one less indirect call
and condition for the inline case.

Besides that, the entire series looks good to me.

Reviewed-by: Logan Gunthorpe <logang@deltatee.com>

Logan

Sagi/Christoph, any comments on this one ?

This series been sitting out for a while now.