Re: [PATCH v1 1/1] nvme: complete directly for hctx with only one ctx mapping

From: Ed Tsai (蔡宗軒)
Date: Tue May 30 2023 - 21:14:59 EST


On Tue, 2023-05-30 at 11:45 -0600, Keith Busch wrote:
>
> External email : Please do not click links or open attachments until
> you have verified the sender or the content.
> On Tue, May 30, 2023 at 10:41:19AM +0800, Po-Wen Kao wrote:
> > ---
> > block/blk-mq.c | 8 +++-----
> > drivers/nvme/host/nvme.h | 4 ++++
> > 2 files changed, 7 insertions(+), 5 deletions(-)
> >
> > diff --git a/block/blk-mq.c b/block/blk-mq.c
> > index 1749f5890606..b60c78f5ad46 100644
> > --- a/block/blk-mq.c
> > +++ b/block/blk-mq.c
> > @@ -1181,12 +1181,10 @@ bool blk_mq_complete_request_remote(struct
> request *rq)
> > WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
> >
> > /*
> > - * For request which hctx has only one ctx mapping,
> > - * or a polled request, always complete locally,
> > - * it's pointless to redirect the completion.
> > + * For a polled request, always complete locally, it's pointless
> > + * to redirect the completion.
> > */
> > -if (rq->mq_hctx->nr_ctx == 1 ||
> > -rq->cmd_flags & REQ_POLLED)
> > +if (rq->cmd_flags & REQ_POLLED)
> > return false;
> >
> > if (blk_mq_complete_need_ipi(rq)) {
> > diff --git a/drivers/nvme/host/nvme.h b/drivers/nvme/host/nvme.h
> > index 7cf8e44d135e..acc9b1ce071d 100644
> > --- a/drivers/nvme/host/nvme.h
> > +++ b/drivers/nvme/host/nvme.h
> > @@ -702,6 +702,10 @@ static inline bool
> nvme_try_complete_req(struct request *req, __le16 status,
> > nvme_should_fail(req);
> > if (unlikely(blk_should_fake_timeout(req->q)))
> > return true;
> > +if (likely(req->mq_hctx->nr_ctx == 1)) {
> > +WRITE_ONCE(req->state, MQ_RQ_COMPLETE);
> > +return false;
> > +}
>
> I don't think we want low level drivers directly messing with blk-mq
> request state.
>
> Is the early nr_ctx check optimisation really worth it? Would the
> following work for your use case?

Ref to original discussion:

https://lore.kernel.org/lkml/1663432858-99743-1-git-send-email-liusong@xxxxxxxxxxxxxxxxx/

Seems it is what nvme hopes to optimize, so I put it back to nvme.
Otherwise, we can just remove the nr_ctx check from block, because the
submission and completion queues can be asymmetric in low level driver.