Re: [PATCH RFC] io_uring: Pass whole sqe to commands

[Ming Lei]

On Fri, Apr 14, 2023 at 02:12:10PM +0100, Pavel Begunkov wrote:
> On 4/14/23 03:12, Ming Lei wrote:
> > On Thu, Apr 13, 2023 at 09:47:56AM -0700, Breno Leitao wrote:
> > > Hello Ming,
> > > 
> > > On Thu, Apr 13, 2023 at 10:56:49AM +0800, Ming Lei wrote:
> > > > On Thu, Apr 06, 2023 at 09:57:05AM -0700, Breno Leitao wrote:
> > > > > Currently uring CMD operation relies on having large SQEs, but future
> > > > > operations might want to use normal SQE.
> > > > > 
> > > > > The io_uring_cmd currently only saves the payload (cmd) part of the SQE,
> > > > > but, for commands that use normal SQE size, it might be necessary to
> > > > > access the initial SQE fields outside of the payload/cmd block.  So,
> > > > > saves the whole SQE other than just the pdu.
> > > > > 
> > > > > This changes slighlty how the io_uring_cmd works, since the cmd
> > > > > structures and callbacks are not opaque to io_uring anymore. I.e, the
> > > > > callbacks can look at the SQE entries, not only, in the cmd structure.
> > > > > 
> > > > > The main advantage is that we don't need to create custom structures for
> > > > > simple commands.
> > > > > 
> > > > > Suggested-by: Pavel Begunkov <asml.silence@xxxxxxxxx>
> > > > > Signed-off-by: Breno Leitao <leitao@xxxxxxxxxx>
> > > > > ---
> > > > 
> > > > ...
> > > > 
> > > > > diff --git a/io_uring/uring_cmd.c b/io_uring/uring_cmd.c
> > > > > index 2e4c483075d3..9648134ccae1 100644
> > > > > --- a/io_uring/uring_cmd.c
> > > > > +++ b/io_uring/uring_cmd.c
> > > > > @@ -63,14 +63,15 @@ EXPORT_SYMBOL_GPL(io_uring_cmd_done);
> > > > >   int io_uring_cmd_prep_async(struct io_kiocb *req)
> > > > >   {
> > > > >   	struct io_uring_cmd *ioucmd = io_kiocb_to_cmd(req, struct io_uring_cmd);
> > > > > -	size_t cmd_size;
> > > > > +	size_t size = sizeof(struct io_uring_sqe);
> > > > >   	BUILD_BUG_ON(uring_cmd_pdu_size(0) != 16);
> > > > >   	BUILD_BUG_ON(uring_cmd_pdu_size(1) != 80);
> > > > > -	cmd_size = uring_cmd_pdu_size(req->ctx->flags & IORING_SETUP_SQE128);
> > > > > +	if (req->ctx->flags & IORING_SETUP_SQE128)
> > > > > +		size <<= 1;
> > > > > -	memcpy(req->async_data, ioucmd->cmd, cmd_size);
> > > > > +	memcpy(req->async_data, ioucmd->sqe, size);
> > > > 
> > > > The copy will make some fields of sqe become READ TWICE, and driver may see
> > > > different sqe field value compared with the one observed in io_init_req().
> > > 
> > > This copy only happens if the operation goes to the async path
> > > (calling io_uring_cmd_prep_async()).  This only happens if
> > > f_op->uring_cmd() returns -EAGAIN.
> > > 
> > >            ret = file->f_op->uring_cmd(ioucmd, issue_flags);
> > >            if (ret == -EAGAIN) {
> > >                    if (!req_has_async_data(req)) {
> > >                            if (io_alloc_async_data(req))
> > >                                    return -ENOMEM;
> > >                            io_uring_cmd_prep_async(req);
> > >                    }
> > >                    return -EAGAIN;
> > >            }
> > > 
> > > Are you saying that after this copy, the operation is still reading from
> > > sqe instead of req->async_data?
> > 
> > I meant that the 2nd read is on the sqe copy(req->aync_data), but same
> > fields can become different between the two READs(first is done on original
> > SQE during io_init_req(), and second is done on sqe copy in driver).
> > 
> > Will this kind of inconsistency cause trouble for driver? Cause READ
> > TWICE becomes possible with this patch.
> 
> Right it might happen, and I was keeping that in mind, but it's not
> specific to this patch. It won't reload core io_uring bits, and all

It depends if driver reloads core bits or not, anyway the patch exports
all fields and opens the window.

> fields cmds use already have this problem.

driver is supposed to load cmds field just once too, right?

> 
> Unless there is a better option, the direction we'll be moving in is
> adding a preparation step that should read and stash parts of SQE
> it cares about, which should also make full SQE copy not
> needed / optional.

Sounds good.


Thanks,
Ming