Re: [PATCH v2] fs: always return zero on success from replace_fd()

[Al Viro <viro@zeniv.linux.org.uk>]

On Mon, Aug 04, 2025 at 04:52:29PM +0100, Al Viro wrote:
> On Mon, Aug 04, 2025 at 02:33:13PM +0200, Christian Brauner wrote:
> 
> > +       guard(spinlock)(&files->file_lock);
> >         err = expand_files(files, fd);
> >         if (unlikely(err < 0))
> > -               goto out_unlock;
> > -       return do_dup2(files, file, fd, flags);
> > +               return err;
> > +       err = do_dup2(files, file, fd, flags);
> > +       if (err < 0)
> > +               return err;
> > 
> > -out_unlock:
> > -       spin_unlock(&files->file_lock);
> > -       return err;
> > +       return 0;
> >  }
> 
> NAK.  This is broken - do_dup2() drops ->file_lock.  And that's why I
> loathe the guard() - it's too easy to get confused *and* assume that
> it will DTRT, no need to check carefully.

Note, BTW, that in actual replacing case do_dup2() has blocking
operations (closing the replaced reference) after dropping ->file_lock,
so making it locking-neutral would not be easy; doable (have it
return the old reference in the replacing case and adjust the callers
accordingly), but it's seriously not pretty (NULL/address of old file/ERR_PTR()
for return value, boilerplate in callers, etc.).  Having do_dup2() called
without ->file_lock and taking it inside is not an option - we could pull
expand_files() in there, but lookup of oldfd in actual dup2(2)/dup3(2) has
to be done within the same ->file_lock scope where it is inserted into the
table.

Sure, all things equal it's better to have functions locking-neutral, but
it's not always the best approach.  And while __free() allows for "we'd
passed the object to somebody else, it's not ours to consume anymore",
guard() does not.