On Sat, 22 Jun 2024 at 14:25, Mateusz Guzik [off-list ref] wrote:

+cc Linus

Thanks.

To sum up the problem: stat and statx met with "" + AT_EMPTY_PATH have
more work to do than fstat and its hypotethical statx counterpart:
- buf alloc/free for the path
- userspace access (very painful on x86_64 + SMAP)
- lockref acquire/release

Yes. That LOOKUP_EMPTY_NOCHECK is *not* the fix.

I do think that we should make AT_EMPTY_PATH with a NULL path
"JustWork(tm)", because the stupid "look if the pathname is empty" is
horrible.

But moving that check into getname() is *NOT* the right answer,
because by the time you get to getname(), you have already lost.

There's a very real reason why vfs_fstatat() catches this empty case
early, and never goes to filename lookup at all. You don't want to
generate a 'struct path' from the 'int fd', because you want to never
get anywhere close to that path, and instead only ever need a 'struct
fd' that can be looked up much more cheaply (particularly if not in a
threaded environment).

So the short-cut in vfs_fstatat() to never get a pathname is
disgusting - people should have used 'fstat()' - but it's _important_
disgusting.

This thing that tries to short-circuit things at the path level is too late.

              Linus