Re: [syzbot] WARNING in mntput_no_expire (2)

[Christian Brauner <christian.brauner@ubuntu.com>]

On Tue, Apr 06, 2021 at 01:13:13PM +0000, Al Viro wrote:
> On Tue, Apr 06, 2021 at 02:35:05PM +0200, Christian Brauner wrote:
> 
> > And while we're at it might I bring up the possibility of an additional
> > cleanup of how we currently call path_init().
> > Right now we pass the return value from path_init() directly into e.g.
> > link_path_walk() which as a first thing checks for error. Which feels
> > rather wrong and has always confused me when looking at these codepaths.
> 
> Why?

Why is a another function in charge of checking the return value of an
initialization function. If something like path_init() fails why is the
next caller responsible for rejecting it's return value and then we're
passing that failure value through the whole function with if (!err)
ladders but as I said it's mostly style preferences.

> 
> > I get that it might make sense for reasons unrelated to path_init() that
> > link_path_walk() checks its first argument for error but path_init()
> > should be checked for error right away especially now that we return
> > early when LOOKUP_CACHED is set without LOOKUP_RCU.
> 
> But you are making the _callers_ of path_init() do that, for no good
> reason.

I'm confused why having callers of functions responsible for checking
error values is such an out-of-band concept suddenly. I don't think it's
worth arguing over this though.

> 
> > thing especially in longer functions such as path_lookupat() where it
> > gets convoluted pretty quickly. I think it would be cleaner to have
> > something like [1]. The early exists make the code easier to reason
> > about imho. But I get that that's a style discussion.
> 
> Your variant is a lot more brittle, actually.
> 
> > @@ -2424,33 +2424,49 @@ static int path_lookupat(struct nameidata *nd, unsigned flags, struct path *path
> >         int err;
> > 
> >         s = path_init(nd, flags);
> > -       if (IS_ERR(s))
> > -               return PTR_ERR(s);
> 
> Where has that come from, BTW?  Currently path_lookupat() does no such thing.

Hm? Are you maybe overlooking path_init() which assigns straight into
the variable declaration? Or are you referring to sm else?

static int path_lookupat(struct nameidata *nd, unsigned flags, struct path *path)
{
	const char *s = path_init(nd, flags);
	int err;

	if (unlikely(flags & LOOKUP_DOWN) && !IS_ERR(s)) {
		err = handle_lookup_down(nd);
		if (unlikely(err < 0))
			s = ERR_PTR(err);
	}

	while (!(err = link_path_walk(s, nd)) &&
	       (s = lookup_last(nd)) != NULL)
		;