Re: [PATCH] do_open(): Fix O_DIRECTORY | O_CREAT behavior

[Christian Brauner]

On Mon, Mar 20, 2023 at 01:24:52PM -0700, Linus Torvalds wrote:
> On Mon, Mar 20, 2023 at 12:27 PM Pedro Falcato <pedro.falcato@xxxxxxxxx> wrote:
> >
> > 1) Pre v5.7 Linux did the open-dir-if-exists-else-create-regular-file
> > we all know and """love""".
> 
> So I think we should fall back to this as a last resort, as a "well,
> it's our historical behavior".
> 
> > 2) Post 5.7, we started returning this buggy -ENOTDIR error, even when
> > successfully creating a file.
> 
> Yeah, I think this is the worst of the bunch and has no excuse (unless
> some crazy program has started depending on it, which sounds really
> *really* unlikely).
> 
> > 3) NetBSD just straight up returns EINVAL on open(O_DIRECTORY | O_CREAT)
> > 4) FreeBSD's open(O_CREAT | O_DIRECTORY) succeeds if the file exists
> > and is a directory. Fails with -ENOENT if it falls onto the "O_CREAT"
> > path (i.e it doesn't try to create the file at all, just ENOENT's;
> > this changed relatively recently, in 2015)
> 
> Either of these sound sensible to me.
> 
> I suspect (3) is the clearest case.

Yeah, we should try that.

> 
> And (4) might be warranted just because it's closer to what we used to
> do, and it's *possible* that somebody happens to use O_DIRECTORY |
> O_CREAT on directories that exist, and never noticed how broken that
> was.

I really really hope that isn't the case because (4) is pretty nasty.
Having to put this on a manpage seems nightmarish.

> 
> And (4) has another special case: O_EXCL. Because I'm really hoping
> that O_DIRECTORY | O_EXCL will always fail.

I've detailed the semantics for that in the commit message below...

> 
> Is the proper patch something along the lines of this?

Yeah, I think that would do it and I think that's what we should try to
get away with. I just spent time and took a look who passes O_DIRECTORY
and O_CREAT and interestingly there are a number of comments roughly
along the lines of the following example:

/* Ideally we could use openat(O_DIRECTORY | O_CREAT | O_EXCL) here
 * to create and open the directory atomically

suggests that people who specify O_DIRECTORY | O_CREAT are interested in
creating a directory. But since this never did work they don't tend to
use that flag combination (I've collected a few samples in [1] to [4].).

(As a sidenote, posix made an interpretation change a long time ago to
at least allow for O_DIRECTORY | O_CREAT to create a directory (see [3]).

But that's a whole different can of worms and I haven't spent any
thoughts even on feasibility. And even if we should probably get through
a couple of kernels with O_DIRECTORY | O_CREAT failing with EINVAL first.)

> 
>    --- a/fs/open.c
>    +++ b/fs/open.c
>    @@ -1186,6 +1186,8 @@ inline int build_open_flags(const struct
> open_how *how, struct open_flags *op)
> 
>         /* Deal with the mode. */
>         if (WILL_CREATE(flags)) {
>    +            if (flags & O_DIRECTORY)
>    +                    return -EINVAL;

This will be problematic because for weird historical reasons O_TMPFILE
includes O_DIRECTORY so this would unfortunately break O_TMPFILE. :/
I'll try to have a patch ready in a bit.

I spent a long time digging through potential users of this nonsense.

Link: https://lore.kernel.org/lkml/20230320071442.172228-1-pedro.falcato@xxxxxxxxx
Link: https://sources.debian.org/src/flatpak/1.14.4-1/subprojects/libglnx/glnx-dirfd.c/?hl=324#L324 [1]
Link: https://sources.debian.org/src/flatpak-builder/1.2.3-1/subprojects/libglnx/glnx-shutil.c/?hl=251#L251 [2]
Link: https://sources.debian.org/src/ostree/2022.7-2/libglnx/glnx-dirfd.c/?hl=324#L324 [3]
Link: https://www.openwall.com/lists/oss-security/2014/11/26/14 [4]