On Mon, Mar 18, 2024 at 11:50:32AM +0000, Pavel Begunkov wrote:
> On 3/18/24 08:10, Ming Lei wrote:
> > On Mon, Mar 18, 2024 at 12:41:48AM +0000, Pavel Begunkov wrote:
> > > io_uring_cmd_done() is called from the irq context and is considered to
> > > be irq safe, however that's not the case if the driver requires
> > > cancellations because io_uring_cmd_del_cancelable() could try to take
> > > the uring_lock mutex.
> > >
> > > Clean up the confusion, by deferring cancellation handling to
> > > locked task_work if we came into io_uring_cmd_done() from iowq
> > > or other IO_URING_F_UNLOCKED path.
> > >
> > > Signed-off-by: Pavel Begunkov <asml.silence@xxxxxxxxx>
> > > ---
> > > io_uring/uring_cmd.c | 24 +++++++++++++++++-------
> > > 1 file changed, 17 insertions(+), 7 deletions(-)
> > >
> > > diff --git a/io_uring/uring_cmd.c b/io_uring/uring_cmd.c
> > > index ec38a8d4836d..9590081feb2d 100644
> > > --- a/io_uring/uring_cmd.c
> > > +++ b/io_uring/uring_cmd.c
> > > @@ -14,19 +14,18 @@
> > > #include "rsrc.h"on #include "uring_cmd.h"
> > > -static void io_uring_cmd_del_cancelable(struct io_uring_cmd *cmd,
> > > - unsigned int issue_flags)
> > > +static void io_uring_cmd_del_cancelable(struct io_uring_cmd *cmd)
> > > {
> > > struct io_kiocb *req = cmd_to_io_kiocb(cmd);
> > > struct io_ring_ctx *ctx = req->ctx;
> > > + lockdep_assert_held(&ctx->uring_lock);
> > > +
> > > if (!(cmd->flags & IORING_URING_CMD_CANCELABLE))
> > > return;
> > > cmd->flags &= ~IORING_URING_CMD_CANCELABLE;
> > > - io_ring_submit_lock(ctx, issue_flags);
> > > hlist_del(&req->hash_node);
> > > - io_ring_submit_unlock(ctx, issue_flags);
> > > }
> > > /*
> > > @@ -44,6 +43,9 @@ void io_uring_cmd_mark_cancelable(struct io_uring_cmd *cmd,
> > > struct io_kiocb *req = cmd_to_io_kiocb(cmd);
> > > struct io_ring_ctx *ctx = req->ctx;
> > > + if (WARN_ON_ONCE(ctx->flags & IORING_SETUP_IOPOLL))
> > > + return;
> > > +
> >
> > This way limits cancelable command can't be used in iopoll context, and
> > it isn't reasonable, and supporting iopoll has been in ublk todo list,
> > especially single ring context is shared for both command and normal IO.
>
> That's something that can be solved when it's needed, and hence the
> warning so it's not missed. That would need del_cancelable on the
> ->iopoll side, but depends on the "leaving in cancel queue"
> problem resolution.
The current code is actually working with iopoll, so adding the warning
can be one regression. Maybe someone has been using ublk with iopoll.
Thanks,
Ming
