[v8,00/11] io_uring: add mkdir and [sym]linkat support

Message

Dmitry Kadashev July 7, 2021, 12:27 p.m. UTC

This started out as an attempt to add mkdirat support to io_uring which
is heavily based on renameat() / unlinkat() support.

During the review process more operations were added (linkat, symlinkat,
mknodat) mainly to keep things uniform internally (in namei.c), and
with things changed in namei.c adding support for these operations to
io_uring is trivial, so that was done too (except for mknodat). See
https://lore.kernel.org/io-uring/20210514145259.wtl4xcsp52woi6ab@wittgenstein/

The first patch makes putname() ignore IS_ERR_OR_NULL names and converts
a couple of places where the check was already implemented in the
callers. No functional changes.

The second one splits filename_lookup() that used to consume the passed
struct filename * on error but not on the success (returning it) into
two: filename_lookup(), that always consumes the name and
__filename_lookup() that never does. This is a preparation change to
enable the subsequent changes to filename_create and filename_lookup. No
functional changes.

The third patch is preparation with no functional changes, makes
do_mkdirat accept struct filename pointer rather than the user string.

The fourth one leverages that to implement mkdirat in io_uring.

5-8 just convert other similar do_* functions in namei.c to accept
struct filename, for uniformity with do_mkdirat, do_renameat and
do_unlinkat. No functional changes there.

9 changes do_* helpers in namei.c to return ints rather than some of
them returning ints and some longs.

10-11 add symlinkat and linkat support to io_uring correspondingly.

Based on for-5.14/io_uring.

v8:
- update filename_parentat() calling conventions to be uniform with the
  ones followed by (changed in subsequent patches) filename_create()
  and filename_lookup()

v7:
- rebase
- make putname() ignore IS_ERR_OR_NULL names, remove conditional calls
  to it from the callers

v6:

- rebase
- add safety checks for IOPOLL mode
- add safety checks for unused sqe parts
- drop mknodat support from io_uring as requested by Jens
- add Christian's Acked-by

v5:
- rebase
- add symlinkat, linkat and mknodat support to io_uring

v4:
- update do_mknodat, do_symlinkat and do_linkat to accept struct
  filename for uniformity with do_mkdirat, do_renameat and do_unlinkat;

v3:
- rebase;

v2:
- do not mess with struct filename's refcount in do_mkdirat, instead add
  and use __filename_create() that does not drop the name on success;

Dmitry Kadashev (11):
  namei: ignore ERR/NULL names in putname()
  namei: change filename_parentat() calling conventions
  fs: make do_mkdirat() take struct filename
  io_uring: add support for IORING_OP_MKDIRAT
  fs: make do_mknodat() take struct filename
  fs: make do_symlinkat() take struct filename
  namei: add getname_uflags()
  fs: make do_linkat() take struct filename
  fs: update do_*() helpers to return ints
  io_uring: add support for IORING_OP_SYMLINKAT
  io_uring: add support for IORING_OP_LINKAT

 fs/exec.c                     |   8 +-
 fs/internal.h                 |   8 +-
 fs/io_uring.c                 | 196 ++++++++++++++++++++++++++++
 fs/namei.c                    | 239 +++++++++++++++++++---------------
 include/linux/fs.h            |   1 +
 include/uapi/linux/io_uring.h |   4 +
 6 files changed, 346 insertions(+), 110 deletions(-)

Comments

Dmitry Kadashev July 7, 2021, 12:32 p.m. UTC | #1

On Wed, Jul 7, 2021 at 7:27 PM Dmitry Kadashev <[e-mail]> wrote:
>
> This started out as an attempt to add mkdirat support to io_uring which
> is heavily based on renameat() / unlinkat() support.
>
> During the review process more operations were added (linkat, symlinkat,
> mknodat) mainly to keep things uniform internally (in namei.c), and
> with things changed in namei.c adding support for these operations to
> io_uring is trivial, so that was done too (except for mknodat). See
> https://lore.kernel.org/io-uring/20210514145259.wtl4xcsp52woi6ab@wittgenstein/

Christian, I've kept your Acked-by on the old commits, since changes
there are pretty minimal (conditional putname()s are unconditional now,
and with __filename_* functions not consuming their args now some goto
labels changed). Hope that was the right thing to do.

Linus Torvalds July 7, 2021, 7:26 p.m. UTC | #2

On Wed, Jul 7, 2021 at 5:28 AM Dmitry Kadashev <[e-mail]> wrote:
>
> This started out as an attempt to add mkdirat support to io_uring which
> is heavily based on renameat() / unlinkat() support.

Ok, sorry for having made you go through all the different versions,
but I like the new series and think it's a marked improvement.

I did send out a few comments to the individual patches that I think
it can all now be made to be even more legible by avoiding some of the
goto spaghetti, but I think that would be a series on top.

(And I'd like to note again that I based all that on just reading the
patches, so there may be something there that makes it not work well).

One final request: can you keep the fs/namei.c patches as one entirely
separate series, and then do the io_uring parts at the end, rather
than intermixing them?

But at least I am generally happy with this version.

Al - please holler now if you see any issues.

               Linus

Dmitry Kadashev July 8, 2021, 6:05 a.m. UTC | #3

On Thu, Jul 8, 2021 at 2:26 AM Linus Torvalds
<[e-mail]> wrote:
>
> On Wed, Jul 7, 2021 at 5:28 AM Dmitry Kadashev <[e-mail]> wrote:
> >
> > This started out as an attempt to add mkdirat support to io_uring which
> > is heavily based on renameat() / unlinkat() support.
>
> Ok, sorry for having made you go through all the different versions,
> but I like the new series and think it's a marked improvement.

No worries at all!

> I did send out a few comments to the individual patches that I think
> it can all now be made to be even more legible by avoiding some of the
> goto spaghetti, but I think that would be a series on top.
>
> (And I'd like to note again that I based all that on just reading the
> patches, so there may be something there that makes it not work well).

I'm happy to do that. I suppose it will have to go through Al's tree
rather than Jens' one?

> One final request: can you keep the fs/namei.c patches as one entirely
> separate series, and then do the io_uring parts at the end, rather
> than intermixing them?

Sure. I'm a bit confused if you mean splitting the series into two or
just moving io_uring bits to the end though. I'll send a v9 with
io_uring bits moved to the end of the series and if you prefer it to be
a completely separate series then let me know.

Thank you for the help again, Linus!