[v1] io_uring/btrfs: remove struct io_uring_cmd_data

[PATCH 0/4] io_uring/btrfs: remove struct io_uring_cmd_data

Posted by Caleb Sander Mateos 3 months, 3 weeks ago

btrfs's ->uring_cmd() implementations are the only ones using io_uring_cmd_data
to store data that lasts for the lifetime of the uring_cmd. But all uring_cmds
have to pay the memory and CPU cost of initializing this field and freeing the
pointer if necessary when the uring_cmd ends. There is already a pdu field in
struct io_uring_cmd that ->uring_cmd() implementations can use for storage. The
only benefit of op_data seems to be that io_uring initializes it, so
->uring_cmd() can read it to tell if there was a previous call to ->uring_cmd().

Introduce a flag IORING_URING_CMD_REISSUE that ->uring_cmd() implementations can
use to tell if this is the first call to ->uring_cmd() or a reissue of the
uring_cmd. Switch btrfs to use the pdu storage for its btrfs_uring_encoded_data.
If IORING_URING_CMD_REISSUE is unset, allocate a new btrfs_uring_encoded_data.
If it's set, use the existing one in op_data. Free the btrfs_uring_encoded_data
in the btrfs layer instead of relying on io_uring to free op_data. Finally,
remove io_uring_cmd_data since it's now unused.

Caleb Sander Mateos (4):
  btrfs/ioctl: don't skip accounting in early ENOTTY return
  io_uring/cmd: introduce IORING_URING_CMD_REISSUE flag
  btrfs/ioctl: store btrfs_uring_encoded_data in io_btrfs_cmd
  io_uring/cmd: remove struct io_uring_cmd_data

 fs/btrfs/ioctl.c             | 41 +++++++++++++++++++++++++-----------
 include/linux/io_uring/cmd.h | 11 ++--------
 io_uring/uring_cmd.c         | 14 +++---------
 io_uring/uring_cmd.h         |  1 -
 4 files changed, 34 insertions(+), 33 deletions(-)

-- 
2.45.2

Re: [PATCH 0/4] io_uring/btrfs: remove struct io_uring_cmd_data

Posted by David Sterba 3 months, 2 weeks ago

On Thu, Jun 19, 2025 at 01:27:44PM -0600, Caleb Sander Mateos wrote:
> btrfs's ->uring_cmd() implementations are the only ones using io_uring_cmd_data
> to store data that lasts for the lifetime of the uring_cmd. But all uring_cmds
> have to pay the memory and CPU cost of initializing this field and freeing the
> pointer if necessary when the uring_cmd ends. There is already a pdu field in
> struct io_uring_cmd that ->uring_cmd() implementations can use for storage. The
> only benefit of op_data seems to be that io_uring initializes it, so
> ->uring_cmd() can read it to tell if there was a previous call to ->uring_cmd().
> 
> Introduce a flag IORING_URING_CMD_REISSUE that ->uring_cmd() implementations can
> use to tell if this is the first call to ->uring_cmd() or a reissue of the
> uring_cmd. Switch btrfs to use the pdu storage for its btrfs_uring_encoded_data.
> If IORING_URING_CMD_REISSUE is unset, allocate a new btrfs_uring_encoded_data.
> If it's set, use the existing one in op_data. Free the btrfs_uring_encoded_data
> in the btrfs layer instead of relying on io_uring to free op_data. Finally,
> remove io_uring_cmd_data since it's now unused.
> 
> Caleb Sander Mateos (4):
>   btrfs/ioctl: don't skip accounting in early ENOTTY return
>   io_uring/cmd: introduce IORING_URING_CMD_REISSUE flag
>   btrfs/ioctl: store btrfs_uring_encoded_data in io_btrfs_cmd
>   io_uring/cmd: remove struct io_uring_cmd_data

The first patch is a fix so it can be put to a -rc queue.

The rest change the io_uring logic, so it's not up to me, but regarding
how to merge them either via btrfs or io_uring tree work for me.

Re: [PATCH 0/4] io_uring/btrfs: remove struct io_uring_cmd_data

Posted by David Sterba 3 months, 1 week ago

On Fri, Jun 20, 2025 at 05:47:43PM +0200, David Sterba wrote:
> On Thu, Jun 19, 2025 at 01:27:44PM -0600, Caleb Sander Mateos wrote:
> > btrfs's ->uring_cmd() implementations are the only ones using io_uring_cmd_data
> > to store data that lasts for the lifetime of the uring_cmd. But all uring_cmds
> > have to pay the memory and CPU cost of initializing this field and freeing the
> > pointer if necessary when the uring_cmd ends. There is already a pdu field in
> > struct io_uring_cmd that ->uring_cmd() implementations can use for storage. The
> > only benefit of op_data seems to be that io_uring initializes it, so
> > ->uring_cmd() can read it to tell if there was a previous call to ->uring_cmd().
> > 
> > Introduce a flag IORING_URING_CMD_REISSUE that ->uring_cmd() implementations can
> > use to tell if this is the first call to ->uring_cmd() or a reissue of the
> > uring_cmd. Switch btrfs to use the pdu storage for its btrfs_uring_encoded_data.
> > If IORING_URING_CMD_REISSUE is unset, allocate a new btrfs_uring_encoded_data.
> > If it's set, use the existing one in op_data. Free the btrfs_uring_encoded_data
> > in the btrfs layer instead of relying on io_uring to free op_data. Finally,
> > remove io_uring_cmd_data since it's now unused.
> > 
> > Caleb Sander Mateos (4):
> >   btrfs/ioctl: don't skip accounting in early ENOTTY return
> >   io_uring/cmd: introduce IORING_URING_CMD_REISSUE flag
> >   btrfs/ioctl: store btrfs_uring_encoded_data in io_btrfs_cmd
> >   io_uring/cmd: remove struct io_uring_cmd_data
> 
> The first patch is a fix so it can be put to a -rc queue.

I've picked the first patch to for-next.

> The rest change the io_uring logic, so it's not up to me, but regarding
> how to merge them either via btrfs or io_uring tree work for me.

This is still pending. I can take it but need an ack.