[PATCH V2 12/12] nvmet: use inline bio for passthru fast path

[Logan Gunthorpe]

On 2020-08-31 4:27 p.m., Chaitanya Kulkarni wrote:
> 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 <chaitanya.kulkarni at wdc.com>
> ---
>  drivers/nvme/target/nvmet.h    |  1 +
>  drivers/nvme/target/passthru.c | 21 ++++++++++++++++++---
>  2 files changed, 19 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/nvme/target/nvmet.h b/drivers/nvme/target/nvmet.h
> index 477439acb8e1..6b1430f8ac78 100644
> --- a/drivers/nvme/target/nvmet.h
> +++ b/drivers/nvme/target/nvmet.h
> @@ -331,6 +331,7 @@ 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 89d848006bd5..ff39f0635451 100644
> --- a/drivers/nvme/target/passthru.c
> +++ b/drivers/nvme/target/passthru.c
> @@ -179,6 +179,14 @@ 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)
>  {
>  	unsigned int op_flags = 0;
> @@ -190,14 +198,21 @@ static int nvmet_passthru_map_sg(struct nvmet_req *req, struct request *rq)
>  	if (req->cmd->common.opcode == nvme_cmd_flush)
>  		op_flags = REQ_FUA;
>  
> -	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;
>  	bio->bi_opf = req_op(rq) | op_flags;
> +	bio->bi_private = req;

I still think this would be cleaner to just do:

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 = bio_put;
}

>  	for_each_sg(req->sg, sg, req->sg_cnt, i) {
>  		if (bio_add_pc_page(rq->q, bio, sg_page(sg), sg->length,
>  				    sg->offset) < sg->length) {
> -			bio_put(bio);
> +			nvmet_passthru_bio_done(bio);

Then just open code this function here...

Logan