From: Ming Lei <tom.leiming@gmail.com>
To: Caleb Sander Mateos <csander@purestorage.com>
Cc: Jens Axboe <axboe@kernel.dk>,
linux-block@vger.kernel.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH] ublk: optimize ublk_queue_cmd_list() for length-1 lists
Date: Fri, 6 Nov 2026 09:04:48 -0600 [thread overview]
Message-ID: <au3skE6GvVt6SNx3@fedora-laptop> (raw)
In-Reply-To: <20261105021351.2295547-1-csander@purestorage.com>
On Wed, Nov 04, 2026 at 07:13:49PM -0700, Caleb Sander Mateos wrote:
> For a request list of length 1, ublk_cmd_list_tw_cb() is more expensive
> than ublk_cmd_tw_cb(). It loads rq_next, which is in a separate cache
> line from the other struct request fields accessed in the dispatch path.
> And it dereferences mq_hctx->driver_data to get the struct ublk_queue *,
> whereas ublk_cmd_tw_cb() gets it from struct ublk_uring_cmd_pdu
> alongside the request pointer. On the submission, ublk_queue_cmd_list()
> can cheaply check whether the request list just assembled has length 1.
> So pass ublk_cmd_tw_cb as the io_uring task work callback if the list
> has a single request and ublk_cmd_list_tw_cb if it contains multiple.
>
> Collapse the req/req_list union in ublk_uring_cmd_pdu into a single
> field so ublk_queue_cmd_list() can store the head request to it in
> either case.
>
> Signed-off-by: Caleb Sander Mateos <csander@purestorage.com>
> ---
> drivers/block/ublk_drv.c | 17 ++++++++---------
> 1 file changed, 8 insertions(+), 9 deletions(-)
>
> diff --git a/drivers/block/ublk_drv.c b/drivers/block/ublk_drv.c
> index 3d322fa1bc83..c397e354a7d6 100644
> --- a/drivers/block/ublk_drv.c
> +++ b/drivers/block/ublk_drv.c
> @@ -117,21 +117,18 @@ struct ublk_batch_fetch_cmd {
> unsigned short buf_group;
> };
>
> struct ublk_uring_cmd_pdu {
> /*
> - * Store requests in same batch temporarily for queuing them to
> - * daemon context.
> + * For ublk_cmd_list_tw_cb(), a list of batched requests queued to
> + * daemon context. For ublk_cmd_tw_cb(), a single request.
> *
> * It should have been stored to request payload, but we do want
> * to avoid extra pre-allocation, and uring_cmd payload is always
> * free for us
> */
> - union {
> - struct request *req;
> - struct request *req_list;
> - };
> + struct request *req;
>
> /*
> * The following two are valid in this cmd whole lifetime, and
> * setup in ublk uring_cmd handler
> */
> @@ -2104,11 +2101,11 @@ static void ublk_queue_cmd(struct ublk_queue *ubq, struct request *rq)
>
> static void ublk_cmd_list_tw_cb(struct io_tw_req tw_req, io_tw_token_t tw)
> {
> struct io_uring_cmd *cmd = io_uring_cmd_from_tw(tw_req);
> struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
> - struct request *rq = pdu->req_list;
> + struct request *rq = pdu->req;
> struct request *next;
>
> do {
> next = rq->rq_next;
> rq->rq_next = NULL;
> @@ -2119,14 +2116,16 @@ static void ublk_cmd_list_tw_cb(struct io_tw_req tw_req, io_tw_token_t tw)
>
> static void ublk_queue_cmd_list(struct ublk_io *io, struct rq_list *l)
> {
> struct io_uring_cmd *cmd = io->cmd;
> struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
> + io_req_tw_func_t tw_cb;
>
> - pdu->req_list = rq_list_peek(l);
> + tw_cb = l->head == l->tail ? ublk_cmd_tw_cb : ublk_cmd_list_tw_cb;
> + pdu->req = rq_list_peek(l);
Both callers guarantee the list is non-empty, so this is fine. It
would be a bit nicer as a rq_list_is_singular() helper in blk-mq.h
rather than open-coding rq_list internals in the driver, but not a
blocker, anyway:
Reviewed-by: Ming Lei <tom.leiming@gmail.com>
Thanks,
Ming
prev parent reply other threads:[~2026-11-06 15:04 UTC|newest]
Thread overview: 2+ messages / expand[flat|nested] mbox.gz Atom feed top
2026-11-05 2:13 Caleb Sander Mateos
2026-11-06 15:04 ` Ming Lei [this message]
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=au3skE6GvVt6SNx3@fedora-laptop \
--to=tom.leiming@gmail.com \
--cc=axboe@kernel.dk \
--cc=csander@purestorage.com \
--cc=linux-block@vger.kernel.org \
--cc=linux-kernel@vger.kernel.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox
all inboxes | Powered by JetHome®