On Wed, Apr 23, 2025 at 07:52:19AM -0700, Caleb Sander Mateos wrote:
> On Wed, Apr 23, 2025 at 7:44 AM Caleb Sander Mateos
> <csander@xxxxxxxxxxxxxxx> wrote:
> >
> > On Wed, Apr 23, 2025 at 2:24 AM Ming Lei <ming.lei@xxxxxxxxxx> wrote:
> > >
> > > The in-tree code calls io_uring_cmd_complete_in_task() to schedule
> > > task_work for dispatching this request to handle
> > > UBLK_U_IO_NEED_GET_DATA.
> > >
> > > This ways is really not necessary because the current context is exactly
> > > the ublk queue context, so call ublk_dispatch_req() directly for handling
> > > UBLK_U_IO_NEED_GET_DATA.
> >
> > Indeed, I was planning to make the same change!
> >
> > >
> > > Signed-off-by: Ming Lei <ming.lei@xxxxxxxxxx>
> > > ---
> > > drivers/block/ublk_drv.c | 14 +++-----------
> > > 1 file changed, 3 insertions(+), 11 deletions(-)
> > >
> > > diff --git a/drivers/block/ublk_drv.c b/drivers/block/ublk_drv.c
> > > index 2de7b2bd409d..c4d4be4f6fbd 100644
> > > --- a/drivers/block/ublk_drv.c
> > > +++ b/drivers/block/ublk_drv.c
> > > @@ -1886,15 +1886,6 @@ static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
> > > }
> > > }
> > >
> > > -static void ublk_handle_need_get_data(struct ublk_device *ub, int q_id,
> > > - int tag)
> > > -{
> > > - struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
> > > - struct request *req = blk_mq_tag_to_rq(ub->tag_set.tags[q_id], tag);
> > > -
> > > - ublk_queue_cmd(ubq, req);
> > > -}
> >
> > Looks like this will conflict with Uday's patch:
> > https://lore.kernel.org/linux-block/20250421-ublk_constify-v1-3-3371f9e9f73c@xxxxxxxxxxxxxxx/
> > . Since that series already has reviews, I expect it will land first.
> >
> > > -
> > > static inline int ublk_check_cmd_op(u32 cmd_op)
> > > {
> > > u32 ioc_type = _IOC_TYPE(cmd_op);
> > > @@ -2103,8 +2094,9 @@ static int __ublk_ch_uring_cmd(struct io_uring_cmd *cmd,
> > > if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
> > > goto out;
> > > ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
> > > - ublk_handle_need_get_data(ub, ub_cmd->q_id, ub_cmd->tag);
> > > - break;
> > > + req = blk_mq_tag_to_rq(ub->tag_set.tags[ub_cmd->q_id], tag);
> > > + ublk_dispatch_req(ubq, req, issue_flags);
> >
> > Maybe it would make sense to factor the UBLK_IO_NEED_GET_DATA handling
> > out of ublk_dispatch_req()? Then ublk_dispatch_req() (called only for
> > incoming ublk requests) could assume the UBLK_IO_FLAG_NEED_GET_DATA
> > flag is not yet set, and this path wouldn't need to pay the cost of
> > re-checking current != ubq->ubq_daemon, ublk_need_get_data(ubq) &&
> > ublk_need_map_req(req), etc.
> >
> > > + return -EIOCBQUEUED;
> >
> > It's probably possible to return the result here synchronously to
> > avoid the small overhead of io_uring_cmd_done(). That may be easier to
> > do if the UBLK_IO_NEED_GET_DATA path is separated from
> > ublk_dispatch_req().
>
> And if we can avoid using io_uring_cmd_done(), calling
> ublk_fill_io_cmd() for UBLK_IO_NEED_GET_DATA would no longer be
> necessary. (This was my original motivation to handle
> UBLK_IO_NEED_GET_DATA synchronously; UBLK_IO_NEED_GET_DATA overwriting
> io->cmd is an obstacle to introducing a struct request * field that
> aliases io->cmd.)
All your comments are reasonable.
Here I just want to keep it simple for backport purpose, and we can clean
up them all by one followup.
Thanks,
Ming
