On Fri, May 02, 2025 at 02:21:05PM -0700, Caleb Sander Mateos wrote:
> On Fri, May 2, 2025 at 8:59 AM Ming Lei <ming.lei@xxxxxxxxxx> wrote:
> >
> > On Thu, May 01, 2025 at 06:31:03PM -0700, Caleb Sander Mateos wrote:
> > > On Wed, Apr 30, 2025 at 8:34 AM Ming Lei <ming.lei@xxxxxxxxxx> wrote:
> > > >
> > > > On Mon, Apr 28, 2025 at 05:43:12PM -0700, Caleb Sander Mateos wrote:
> > > > > On Mon, Apr 28, 2025 at 2:44 AM Ming Lei <ming.lei@xxxxxxxxxx> wrote:
> > > > > >
> > > > > > Extend io_buffer_register_bvec() and io_buffer_unregister_bvec() for
> > > > > > supporting to register/unregister bvec buffer to specified io_uring,
> > > > > > which FD is usually passed from userspace.
> > > > > >
> > > > > > Signed-off-by: Ming Lei <ming.lei@xxxxxxxxxx>
> > > > > > ---
> > > > > > include/linux/io_uring/cmd.h | 4 ++
> > > > > > io_uring/rsrc.c | 83 +++++++++++++++++++++++++++---------
> > > > > > 2 files changed, 67 insertions(+), 20 deletions(-)
> > > > > >
> > > > > > diff --git a/include/linux/io_uring/cmd.h b/include/linux/io_uring/cmd.h
> > > > > > index 78fa336a284b..7516fe5cd606 100644
> > > > > > --- a/include/linux/io_uring/cmd.h
> > > > > > +++ b/include/linux/io_uring/cmd.h
> > > > > > @@ -25,6 +25,10 @@ struct io_uring_cmd_data {
> > > > > >
> > > > > > struct io_buf_data {
> > > > > > unsigned short index;
> > > > > > + bool has_fd;
> > > > > > + bool registered_fd;
> > > > > > +
> > > > > > + int ring_fd;
> > > > > > struct request *rq;
> > > > > > void (*release)(void *);
> > > > > > };
> > > > > > diff --git a/io_uring/rsrc.c b/io_uring/rsrc.c
> > > > > > index 5f8ab130a573..701dd33fecf7 100644
> > > > > > --- a/io_uring/rsrc.c
> > > > > > +++ b/io_uring/rsrc.c
> > > > > > @@ -969,21 +969,6 @@ static int __io_buffer_register_bvec(struct io_ring_ctx *ctx,
> > > > > > return 0;
> > > > > > }
> > > > > >
> > > > > > -int io_buffer_register_bvec(struct io_uring_cmd *cmd,
> > > > > > - struct io_buf_data *buf,
> > > > > > - unsigned int issue_flags)
> > > > > > -{
> > > > > > - struct io_ring_ctx *ctx = cmd_to_io_kiocb(cmd)->ctx;
> > > > > > - int ret;
> > > > > > -
> > > > > > - io_ring_submit_lock(ctx, issue_flags);
> > > > > > - ret = __io_buffer_register_bvec(ctx, buf);
> > > > > > - io_ring_submit_unlock(ctx, issue_flags);
> > > > > > -
> > > > > > - return ret;
> > > > > > -}
> > > > > > -EXPORT_SYMBOL_GPL(io_buffer_register_bvec);
> > > > > > -
> > > > > > static int __io_buffer_unregister_bvec(struct io_ring_ctx *ctx,
> > > > > > struct io_buf_data *buf)
> > > > > > {
> > > > > > @@ -1006,19 +991,77 @@ static int __io_buffer_unregister_bvec(struct io_ring_ctx *ctx,
> > > > > > return 0;
> > > > > > }
> > > > > >
> > > > > > -int io_buffer_unregister_bvec(struct io_uring_cmd *cmd,
> > > > > > - struct io_buf_data *buf,
> > > > > > - unsigned int issue_flags)
> > > > > > +static inline int do_reg_unreg_bvec(struct io_ring_ctx *ctx,
> > > > > > + struct io_buf_data *buf,
> > > > > > + unsigned int issue_flags,
> > > > > > + bool reg)
> > > > > > {
> > > > > > - struct io_ring_ctx *ctx = cmd_to_io_kiocb(cmd)->ctx;
> > > > > > int ret;
> > > > > >
> > > > > > io_ring_submit_lock(ctx, issue_flags);
> > > > > > - ret = __io_buffer_unregister_bvec(ctx, buf);
> > > > > > + if (reg)
> > > > > > + ret = __io_buffer_register_bvec(ctx, buf);
> > > > > > + else
> > > > > > + ret = __io_buffer_unregister_bvec(ctx, buf);
> > > > >
> > > > > It feels like unifying __io_buffer_register_bvec() and
> > > > > __io_buffer_unregister_bvec() would belong better in the prior patch
> > > > > that changes their signatures.
> > > >
> > > > Can you share how to do above in previous patch?
> > >
> > > I was thinking you could define do_reg_unreg_bvec() in the previous
> > > patch. It's a logical step once you've extracted out all the
> > > differences between io_buffer_register_bvec() and
> > > io_buffer_unregister_bvec() into the helpers
> > > __io_buffer_register_bvec() and __io_buffer_unregister_bvec(). But
> > > either way is fine.
> >
> > 'has_fd' and 'ring_fd' fields isn't added yet, the defined do_reg_unreg_bvec()
> > could be quite simple, looks no big difference, I can do that...
> >
> > >
> > > >
> > > > >
> > > > > > io_ring_submit_unlock(ctx, issue_flags);
> > > > > >
> > > > > > return ret;
> > > > > > }
> > > > > > +
> > > > > > +static int io_buffer_reg_unreg_bvec(struct io_ring_ctx *ctx,
> > > > > > + struct io_buf_data *buf,
> > > > > > + unsigned int issue_flags,
> > > > > > + bool reg)
> > > > > > +{
> > > > > > + struct io_ring_ctx *remote_ctx = ctx;
> > > > > > + struct file *file = NULL;
> > > > > > + int ret;
> > > > > > +
> > > > > > + if (buf->has_fd) {
> > > > > > + file = io_uring_register_get_file(buf->ring_fd, buf->registered_fd);
> > > > > > + if (IS_ERR(file))
> > > > > > + return PTR_ERR(file);
> > > > >
> > > > > It would be good to avoid the overhead of this lookup and
> > > > > reference-counting in the I/O path. Would it be possible to move this
> > > > > lookup to when UBLK_IO_FETCH_REQ (and UBLK_IO_COMMIT_AND_FETCH_REQ, if
> > > > > it specifies a different ring_fd) is submitted? I guess that might
> > > > > require storing an extra io_ring_ctx pointer in struct ublk_io.
> > > >
> > > > Let's start from the flexible way & simple implementation.
> > > >
> > > > Any optimization & improvement can be done as follow-up.
> > >
> > > Sure, we can start with this as-is. But I suspect the extra
> > > reference-counting here will significantly decrease the benefit of the
> > > auto-register register feature.
> >
> > The reference-counting should only be needed for registering buffer to
> > external ring, which may have been slow because of the cross-ring thing...
>
> The current code is incrementing and decrementing the io_uring file
> reference count even if the remote_ctx == ctx, right? I agree it
Yes, but it can be changed to drop the inc/dec file reference easily since we
have a flag field.
> should definitely be possible to skip the reference count in that
> case, as this code is already running in task work context for a
> command on the io_uring.
The current 'uring_cmd' instance holds one reference of the
io_ring_ctx instance.
> It should also be possible to avoid atomic
> reference-counting in the UBLK_AUTO_BUF_REGISTERED_RING case too.
For registering buffer to external io_ring, it is hard to avoid to grag
the io_uring_ctx reference when specifying the io_uring_ctx via its FD.
>
> >
> > Maybe we can start automatic buffer register for ubq_daemon context only,
> > meantime allow to register buffer from external io_uring by adding per-io
> > spin_lock, which may help the per-io task Uday is working on too.
>
> I'm not sure I understand why a spinlock would be required? In Uday's
> patch set, each ublk_io still belongs to a single task. So no
> additional locking should be required.
I think it is very useful to allow to register io buffer in the
other(non-ubq_daemon) io_uring context by the offload style.
Especially the register/unregister io buffer uring_cmd is for handling
target IO, which should have been issued in same context of target io
handling.
Without one per-io spinlock, it is hard to avoid one race you mentioned:
https://lore.kernel.org/linux-block/aA2pNRkBhgKsofRP@fedora/#t
in case of bad ublk server implementation.
>
> >
> > And the interface still allow to support automatic buffer register to
> > external io_uring since `ublk_auto_buf_reg` includes 'flags' field, we can
> > enable it in future when efficient implementation is figured out.
>
> Sure, we can definitely start with support only for auto-registering
> the buffer with the ublk command's own io_uring. Implementing a flag
> in the future to specify a different io_uring seems like a good
> approach.
OK, I will send V2 by starting with auto-registering buffer to the ublk
uring_cmd io_uring first.
Thanks,
Ming
