[PATCH 5/5] fs: Convert struct file::f_count to refcount_long_t

[Christian Brauner]

On Fri, May 03, 2024 at 12:36:14PM +0200, Peter Zijlstra wrote:
> On Fri, May 03, 2024 at 11:37:25AM +0200, Christian Brauner wrote:
> > On Thu, May 02, 2024 at 05:41:23PM -0700, Kees Cook wrote:
> > > On Fri, May 03, 2024 at 01:14:45AM +0100, Al Viro wrote:
> > > > On Thu, May 02, 2024 at 05:10:18PM -0700, Kees Cook wrote:
> > > > 
> > > > > But anyway, there needs to be a general "oops I hit 0"-aware form of
> > > > > get_file(), and it seems like it should just be get_file() itself...
> > > > 
> > > > ... which brings back the question of what's the sane damage mitigation
> > > > for that.  Adding arseloads of never-exercised failure exits is generally
> > > > a bad idea - it's asking for bitrot and making the thing harder to review
> > > > in future.
> > > 
> > > Linus seems to prefer best-effort error recovery to sprinkling BUG()s
> > > around.  But if that's really the solution, then how about get_file()
> > > switching to to use inc_not_zero and BUG on 0?
> > 
> > Making get_file() return an error is not an option. For all current
> > callers that's pointless churn for a condition that's not supposed to
> > happen at all.
> > 
> > Additionally, iirc *_inc_not_zero() variants are implemented with
> > try_cmpxchg() which scales poorly under contention for a condition
> > that's not supposed to happen.
> 
> 	unsigned long old = atomic_long_fetch_inc_relaxed(&f->f_count);
> 	WARN_ON(!old);
> 
> Or somesuch might be an option?

Yeah, I'd be fine with that. WARN_ON() (or WARN_ON_ONCE() even?) and
then people can do their panic_on_warn stuff to get the BUG_ON()
behavior if they want to.