| Message ID | 20230913151616.3164338-6-chengming.zhou@linux.dev (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | blk-mq: optimize queue_rqs() support | expand |
On Wed, Sep 13, 2023 at 03:16:16PM +0000, chengming.zhou@linux.dev wrote: > From: Chengming Zhou <zhouchengming@bytedance.com> > > Add batched mq_ops.queue_rqs() support in null_blk for testing. The > implementation is much easy since null_blk doesn't have commit_rqs(). > > We simply handle each request one by one, if errors are encountered, > leave them in the passed in list and return back. > > There is about 3.6% improvement in IOPS of fio/t/io_uring on null_blk > with hw_queue_depth=256 on my test VM, from 1.09M to 1.13M. I guess you pass 'shared_tags' to null_blk for the verification? > > Signed-off-by: Chengming Zhou <zhouchengming@bytedance.com> > --- > drivers/block/null_blk/main.c | 20 ++++++++++++++++++++ > 1 file changed, 20 insertions(+) > > diff --git a/drivers/block/null_blk/main.c b/drivers/block/null_blk/main.c > index 968090935eb2..79d6cd3c3d41 100644 > --- a/drivers/block/null_blk/main.c > +++ b/drivers/block/null_blk/main.c > @@ -1750,6 +1750,25 @@ static blk_status_t null_queue_rq(struct blk_mq_hw_ctx *hctx, > return null_handle_cmd(cmd, sector, nr_sectors, req_op(rq)); > } > > +static void null_queue_rqs(struct request **rqlist) > +{ > + struct request *requeue_list = NULL; > + struct request **requeue_lastp = &requeue_list; > + struct blk_mq_queue_data bd = { }; > + blk_status_t ret; > + > + do { > + struct request *rq = rq_list_pop(rqlist); > + > + bd.rq = rq; > + ret = null_queue_rq(rq->mq_hctx, &bd); > + if (ret != BLK_STS_OK) > + rq_list_add_tail(&requeue_lastp, rq); > + } while (!rq_list_empty(*rqlist)); > + > + *rqlist = requeue_list; > +} > + null_blk may not be one perfect example for showing queue_rqs() which is usually for handling request in batch, but for test or demo purpose, it is fine: Reviewed-by: Ming Lei <ming.lei@redhat.com> Thanks Ming
On 2023/9/22 16:54, Ming Lei wrote: > On Wed, Sep 13, 2023 at 03:16:16PM +0000, chengming.zhou@linux.dev wrote: >> From: Chengming Zhou <zhouchengming@bytedance.com> >> >> Add batched mq_ops.queue_rqs() support in null_blk for testing. The >> implementation is much easy since null_blk doesn't have commit_rqs(). >> >> We simply handle each request one by one, if errors are encountered, >> leave them in the passed in list and return back. >> >> There is about 3.6% improvement in IOPS of fio/t/io_uring on null_blk >> with hw_queue_depth=256 on my test VM, from 1.09M to 1.13M. > > I guess you pass 'shared_tags' to null_blk for the verification? IIRC it should be "modprobe null_blk hw_queue_depth=256 nr_devices=2 shared_tags=1". > >> >> Signed-off-by: Chengming Zhou <zhouchengming@bytedance.com> >> --- >> drivers/block/null_blk/main.c | 20 ++++++++++++++++++++ >> 1 file changed, 20 insertions(+) >> >> diff --git a/drivers/block/null_blk/main.c b/drivers/block/null_blk/main.c >> index 968090935eb2..79d6cd3c3d41 100644 >> --- a/drivers/block/null_blk/main.c >> +++ b/drivers/block/null_blk/main.c >> @@ -1750,6 +1750,25 @@ static blk_status_t null_queue_rq(struct blk_mq_hw_ctx *hctx, >> return null_handle_cmd(cmd, sector, nr_sectors, req_op(rq)); >> } >> >> +static void null_queue_rqs(struct request **rqlist) >> +{ >> + struct request *requeue_list = NULL; >> + struct request **requeue_lastp = &requeue_list; >> + struct blk_mq_queue_data bd = { }; >> + blk_status_t ret; >> + >> + do { >> + struct request *rq = rq_list_pop(rqlist); >> + >> + bd.rq = rq; >> + ret = null_queue_rq(rq->mq_hctx, &bd); >> + if (ret != BLK_STS_OK) >> + rq_list_add_tail(&requeue_lastp, rq); >> + } while (!rq_list_empty(*rqlist)); >> + >> + *rqlist = requeue_list; >> +} >> + > > null_blk may not be one perfect example for showing queue_rqs() > which is usually for handling request in batch, but for test or > demo purpose, it is fine: > > Reviewed-by: Ming Lei <ming.lei@redhat.com> > Yes, some other "real" drivers should be better choice that we can handle more things in batch to improve performance. Maybe ublk driver can benefit from this too. Thanks!
diff --git a/drivers/block/null_blk/main.c b/drivers/block/null_blk/main.c index 968090935eb2..79d6cd3c3d41 100644 --- a/drivers/block/null_blk/main.c +++ b/drivers/block/null_blk/main.c @@ -1750,6 +1750,25 @@ static blk_status_t null_queue_rq(struct blk_mq_hw_ctx *hctx, return null_handle_cmd(cmd, sector, nr_sectors, req_op(rq)); } +static void null_queue_rqs(struct request **rqlist) +{ + struct request *requeue_list = NULL; + struct request **requeue_lastp = &requeue_list; + struct blk_mq_queue_data bd = { }; + blk_status_t ret; + + do { + struct request *rq = rq_list_pop(rqlist); + + bd.rq = rq; + ret = null_queue_rq(rq->mq_hctx, &bd); + if (ret != BLK_STS_OK) + rq_list_add_tail(&requeue_lastp, rq); + } while (!rq_list_empty(*rqlist)); + + *rqlist = requeue_list; +} + static void cleanup_queue(struct nullb_queue *nq) { bitmap_free(nq->tag_map); @@ -1802,6 +1821,7 @@ static int null_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data, static const struct blk_mq_ops null_mq_ops = { .queue_rq = null_queue_rq, + .queue_rqs = null_queue_rqs, .complete = null_complete_rq, .timeout = null_timeout_rq, .poll = null_poll,