From a4fe2d3afe3ce77edeadb567c0d0a8d102c6b159 Mon Sep 17 00:00:00 2001 From: Chaitanya Kulkarni Date: Mon, 9 Nov 2020 18:24:04 -0800 Subject: [PATCH] nvmet: use blk_rq_bio_prep instead of blk_rq_append_bio The function blk_rq_append_bio() is a genereric API written for all types driver (having bounce buffers) and different context (where request is already having a bio i.e. rq->bio != NULL). It does mainly three things: calculating the segments, bounce queue and if req->bio == NULL call blk_rq_bio_prep() or handle low level merge() case. The NVMe PCIe and fabrics transports currently does not use queue bounce mechanism. In order to find this for each request processing in the passthru blk_rq_append_bio() does extra work in the fast path for each request. When I ran I/Os with different block sizes on the passthru controller I found that we can reuse the req->sg_cnt instead of iterating over the bvecs to find out nr_segs in blk_rq_append_bio(). This calculation in blk_rq_append_bio() is a duplication of work given that we have the value in req->sg_cnt. (correct me here if I'm wrong). With NVMe passthru request based driver we allocate fresh request each time, so every call to blk_rq_append_bio() rq->bio will be NULL i.e. we don't really need the second condition in the blk_rq_append_bio() and the resulting error condition in the caller of blk_rq_append_bio(). So for NVMeOF passthru driver recalculating the segments, bounce check and ll_back_merge code is not needed such that we can get away with the minimal version of the blk_rq_append_bio() which removes the error check in the fast path along with extra variable in nvmet_passthru_map_sg(). This patch updates the nvmet_passthru_map_sg() such that it does only appending the bio to the request in the context of the NVMeOF Passthru driver. Following are perf numbers :- With current implementation (blk_rq_append_bio()) :- ---------------------------------------------------- + 5.80% 0.02% kworker/0:2-mm_ [nvmet] [k] nvmet_passthru_execute_cmd + 5.44% 0.01% kworker/0:2-mm_ [nvmet] [k] nvmet_passthru_execute_cmd + 4.88% 0.00% kworker/0:2-mm_ [nvmet] [k] nvmet_passthru_execute_cmd + 5.44% 0.01% kworker/0:2-mm_ [nvmet] [k] nvmet_passthru_execute_cmd + 4.86% 0.01% kworker/0:2-mm_ [nvmet] [k] nvmet_passthru_execute_cmd + 5.17% 0.00% kworker/0:2-eve [nvmet] [k] nvmet_passthru_execute_cmd With this patch using blk_rq_bio_prep() :- ---------------------------------------------------- + 3.14% 0.02% kworker/0:2-eve [nvmet] [k] nvmet_passthru_execute_cmd + 3.26% 0.01% kworker/0:2-eve [nvmet] [k] nvmet_passthru_execute_cmd + 5.37% 0.01% kworker/0:2-mm_ [nvmet] [k] nvmet_passthru_execute_cmd + 5.18% 0.02% kworker/0:2-eve [nvmet] [k] nvmet_passthru_execute_cmd + 4.84% 0.02% kworker/0:2-mm_ [nvmet] [k] nvmet_passthru_execute_cmd + 4.87% 0.01% kworker/0:2-mm_ [nvmet] [k] nvmet_passthru_execute_cmd Signed-off-by: Chaitanya Kulkarni Reviewed-by: Logan Gunthorpe Signed-off-by: Christoph Hellwig --- drivers/nvme/target/passthru.c | 8 ++------ 1 file changed, 2 insertions(+), 6 deletions(-) diff --git a/drivers/nvme/target/passthru.c b/drivers/nvme/target/passthru.c index 1c84dad..2b24205 100644 --- a/drivers/nvme/target/passthru.c +++ b/drivers/nvme/target/passthru.c @@ -189,7 +189,7 @@ static int nvmet_passthru_map_sg(struct nvmet_req *req, struct request *rq) { struct scatterlist *sg; struct bio *bio; - int i, ret; + int i; if (req->sg_cnt > BIO_MAX_PAGES) return -EINVAL; @@ -206,11 +206,7 @@ static int nvmet_passthru_map_sg(struct nvmet_req *req, struct request *rq) } } - ret = blk_rq_append_bio(rq, &bio); - if (unlikely(ret)) { - bio_put(bio); - return ret; - } + blk_rq_bio_prep(rq, bio, req->sg_cnt); return 0; } -- 2.7.4