On Tue, Nov 11, 2025 at 07:41:24AM -0700, Jens Axboe wrote:
> > diff --git a/io_uring/openclose.c b/io_uring/openclose.c
> > index bfeb91b31bba..6bc14f626923 100644
> > --- a/io_uring/openclose.c
> > +++ b/io_uring/openclose.c
> > @@ -121,6 +118,7 @@ int io_openat2(struct io_kiocb *req, unsigned int
> > issue_flags)
> > struct file *file;
> > bool resolve_nonblock, nonblock_set;
> > bool fixed = !!open->file_slot;
> > + struct filename *name __free(putname) =
> > complete_getname(&open->filename);
> > int ret;
> >
> > ret = build_open_flags(&open->how, &op);
>
> I don't think this will work as-is - the prep has been done on the
> request, but we could be retrying io_openat2(). That will happen if this
> function returns -EAGAIN. That will then end up with a cleared out
> filename for the second (blocking) invocation.
If retry happens in a different thread, we do have a problem ;-/
This -EAGAIN might've come from ->open() itself (io_openat2() sets
O_NONBLOCK on the same calls), and by that point we have already
shoved that filename in direction of audit...
IMO the first 10 commits (up to and including audit_reusename() removal)
are useful on their own, but io_openat2() part does look broken.
Hmm... FWIW, we could do a primitive like
int putname_to_incomplete(struct incomplete_name *v, struct filename *name)
{
if (likely(name->refcnt == 1)) {
v->__incomplete_filename = name;
return 0;
}
v->__incomplete_filename = <duplicate name>;
putname(name);
if (unlikely(!v->__incomplete_filename))
return -ENOMEM;
return 0;
}
and have
if (ret == -EAGAIN &&
(!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
{
ret = putname_to_incomplete(&open->filename,
no_free_ptr(name));
if (unlikely(ret))
goto err;
return -EAGAIN;
}
in io_openat2() (in addition to what's already done in 11/13). Workable or
too disgusting?
