Message-ID: <aBBQDjLDkGWE63vT@fedora>
Date: Tue, 29 Apr 2025 12:05:34 +0800
From: Ming Lei <ming.lei@...hat.com>
To: Caleb Sander Mateos <csander@...estorage.com>
Cc: Jens Axboe <axboe@...nel.dk>, Uday Shankar <ushankar@...estorage.com>,
	linux-block@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH 5/8] ublk: factor out ublk_start_io() helper

On Mon, Apr 28, 2025 at 08:12:52AM -0700, Caleb Sander Mateos wrote:
> On Mon, Apr 28, 2025 at 7:28 AM Caleb Sander Mateos
> <csander@...estorage.com> wrote:
> >
> > On Sun, Apr 27, 2025 at 6:05 AM Ming Lei <ming.lei@...hat.com> wrote:
> > >
> > > On Sat, Apr 26, 2025 at 10:58:00PM -0600, Caleb Sander Mateos wrote:
> > > > In preparation for calling it from outside ublk_dispatch_req(), factor
> > > > out the code responsible for setting up an incoming ublk I/O request.
> > > >
> > > > Signed-off-by: Caleb Sander Mateos <csander@...estorage.com>
> > > > ---
> > > >  drivers/block/ublk_drv.c | 53 ++++++++++++++++++++++------------------
> > > >  1 file changed, 29 insertions(+), 24 deletions(-)
> > > >
> > > > diff --git a/drivers/block/ublk_drv.c b/drivers/block/ublk_drv.c
> > > > index 01fc92051754..90a38a82f8cc 100644
> > > > --- a/drivers/block/ublk_drv.c
> > > > +++ b/drivers/block/ublk_drv.c
> > > > @@ -1151,17 +1151,44 @@ static inline void __ublk_abort_rq(struct ublk_queue *ubq,
> > > >               blk_mq_requeue_request(rq, false);
> > > >       else
> > > >               blk_mq_end_request(rq, BLK_STS_IOERR);
> > > >  }
> > > >
> > > > +static void ublk_start_io(struct ublk_queue *ubq, struct request *req,
> > > > +                       struct ublk_io *io)
> > > > +{
> > > > +     unsigned mapped_bytes = ublk_map_io(ubq, req, io);
> > > > +
> > > > +     /* partially mapped, update io descriptor */
> > > > +     if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
> > > > +             /*
> > > > +              * Nothing mapped, retry until we succeed.
> > > > +              *
> > > > +              * We may never succeed in mapping any bytes here because
> > > > +              * of OOM. TODO: reserve one buffer with single page pinned
> > > > +              * for providing forward progress guarantee.
> > > > +              */
> > > > +             if (unlikely(!mapped_bytes)) {
> > > > +                     blk_mq_requeue_request(req, false);
> > > > +                     blk_mq_delay_kick_requeue_list(req->q,
> > > > +                                     UBLK_REQUEUE_DELAY_MS);
> > > > +                     return;
> > > > +             }
> > > > +
> > > > +             ublk_get_iod(ubq, req->tag)->nr_sectors =
> > > > +                     mapped_bytes >> 9;
> > > > +     }
> > > > +
> > > > +     ublk_init_req_ref(ubq, req);
> > > > +}
> > > > +
> > > >  static void ublk_dispatch_req(struct ublk_queue *ubq,
> > > >                             struct request *req,
> > > >                             unsigned int issue_flags)
> > > >  {
> > > >       int tag = req->tag;
> > > >       struct ublk_io *io = &ubq->ios[tag];
> > > > -     unsigned int mapped_bytes;
> > > >
> > > >       pr_devel("%s: complete: qid %d tag %d io_flags %x addr %llx\n",
> > > >                       __func__, ubq->q_id, req->tag, io->flags,
> > > >                       ublk_get_iod(ubq, req->tag)->addr);
> > > >
> > > > @@ -1204,33 +1231,11 @@ static void ublk_dispatch_req(struct ublk_queue *ubq,
> > > >               pr_devel("%s: update iod->addr: qid %d tag %d io_flags %x addr %llx\n",
> > > >                               __func__, ubq->q_id, req->tag, io->flags,
> > > >                               ublk_get_iod(ubq, req->tag)->addr);
> > > >       }
> > > >
> > > > -     mapped_bytes = ublk_map_io(ubq, req, io);
> > > > -
> > > > -     /* partially mapped, update io descriptor */
> > > > -     if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
> > > > -             /*
> > > > -              * Nothing mapped, retry until we succeed.
> > > > -              *
> > > > -              * We may never succeed in mapping any bytes here because
> > > > -              * of OOM. TODO: reserve one buffer with single page pinned
> > > > -              * for providing forward progress guarantee.
> > > > -              */
> > > > -             if (unlikely(!mapped_bytes)) {
> > > > -                     blk_mq_requeue_request(req, false);
> > > > -                     blk_mq_delay_kick_requeue_list(req->q,
> > > > -                                     UBLK_REQUEUE_DELAY_MS);
> > > > -                     return;
> > > > -             }
> > >
> > > Here it needs to break ublk_dispatch_req() for not completing the
> > > uring_cmd, however ublk_start_io() can't support it.
> >
> > Good catch. How about I change ublk_start_io() to return a bool
> > indicating whether the I/O was successfully started?

That is doable.

> 
> Thinking a bit more about this, is the existing behavior of returning
> early from ublk_dispatch_req() correct for UBLK_IO_NEED_GET_DATA? It

The requeue isn't related with UBLK_IO_NEED_GET_DATA actually, when
UBLK_IO_FLAG_NEED_GET_DATA is cleared.

It is usually caused by running out of pages, so we have to requeue until
ublk_map_io() can make progress.

> makes sense for the initial ublk_dispatch_req() because the req will
> be requeued without consuming the ublk fetch request, allowing it to
> be reused for a subsequent I/O. But for UBLK_IO_NEED_GET_DATA, doesn't
> it mean the io_uring_cmd will never complete? I would think it would
> be better to return an error code in this case.

The same request will be requeued and re-dispatched to ublk driver after
a short delay, so the uring_cmd won't be never complete.

Anyway, it isn't another story, which shouldn't be added into this
cleanup patch.

Thanks,
Ming

Powered by blists - more mailing lists